Pith. sign in

REVIEW 4 major objections 4 minor 14 references

Joint Access Point Selection and Precoder Design under Statistical CSI

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A graph neural network that receives only channel covariance matrices can jointly select access points and design precoders, achieving higher sum-rate than an iterative stochastic WMMSE algorithm and generalizing to unseen numbers of users.

desk verdict A solid, incremental combination of SWMMSE and an attention-based Edge-GNN for joint AP selection under statistical CSI; the main empirical claim is plausible but slightly oversold and lacks baseline sensitivity and error bars. read the letter →

arxiv 2608.06251 v1 pith:FHPGHLJG submitted 2026-08-06 eess.SP

Tim N. Faisst , Franz Wei{ss}er , Wolfgang Utschick This is my paper · ORCID
classification eess.SP
keywords statisticalCSIaccesspointselectionprecodingsum-ratemaximizationstochasticWMMSEgraphneuralnetworkspermutationequivarianceattentionmechanism
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

Under statistical channel state information, where each channel is known only through its covariance matrix, this paper studies the joint problem of assigning users to access points and designing transmit precoders to maximize sum-rate. It proposes two solutions: an iterative algorithm that alternates stochastic WMMSE precoder updates with projected-gradient assignment updates, and a graph neural network that produces both the assignment and the precoders in a single forward pass. The central claim is that the GNN attains higher sum-rate than the iterative algorithm across the tested SNR range and generalizes to varying numbers of users with comparable performance. Both methods beat baselines that fix AP selection greedily or randomly, with the largest gains when users are spatially clustered. The appeal is practical: if true, a fast, permutation-equivariant learned policy can replace costly iterative stochastic optimization in dense indoor deployments.

What carries the argument

The load-bearing object is the GNN's layer update rule (Eq. 6), applied on a composite bipartite graph whose edges connect each user to each antenna of each AP and carry features extracted from the first row of the channel covariance matrix (valid because the APs are uniform linear arrays, giving Toeplitz covariances). The update has four aggregation terms: the edge's own features, same-AP antennas, attention-weighted interfering users on the same AP, and—new in this paper—a cross-AP aggregation that sums over antennas of other APs in a permutation-invariant way. This fourth term is what lets the network reason about whether another AP would serve the user better. After several layers, a precoder head outputs the real and imaginary parts of each precoding vector, and an AP selection head mean-pools over antennas and applies a shared linear layer followed by softmax to produce the relaxed assignment. The relaxation itself is justified by a vertex argument: for fixed precoders the objective is linear in the assignment variables, so the optimum over the simplex is at a standard basis vector.

What would settle it

Re-run the trained GNN and the iterative baseline on a dataset of measured (not simulated) 28 GHz indoor channel covariance matrices, and compare achieved sum-rates. If the GNN no longer outperforms the iterative algorithm, or its generalization to varying user counts degrades, the central claim is falsified.

Watch

Extended reading notes

Core claim

The core discovery is that joint AP selection and precoding under statistical CSI can be solved by an attention-based Edge-GNN whose assignment relaxation is provably lossless. Because the sum-rate objective is linear in the assignment variables for fixed precoders, its optimum over the simplex relaxation is attained at a vertex, which is a binary assignment; the paper uses this to justify training with softmax outputs. Empirically, on a simulated indoor 28 GHz environment with two 16-antenna APs, the GNN trained on scenarios with four users outperforms the iterative SWMMSE-based algorithm at every tested SNR, and it keeps comparable performance when evaluated on two to seven users despite never seeing those counts in training. The GNN also matches or beats the iterative method on clustered-user test sets, where greedy AP assignment is known to fail. The authors interpret the high-SNR advantage as the GNN having learned prior information about interference that the locally converged iterative algorithm cannot recover.

Load-bearing premise

The comparison is carried out entirely on covariance matrices produced by a conditional variational autoencoder trained on ray-traced indoor channel data, so the claimed GNN advantage may not transfer to real 28 GHz indoor environments.

Editorial extensions

If this is right

  • If the GNN's empirical advantage holds, joint AP selection and precoding under statistical CSI can be performed in one forward pass, eliminating the hundreds of iterations required by WMMSE-based methods and cutting inference latency.
  • The lossless relaxation means combinatorial search over AP assignments can be replaced by continuous optimization, making the problem amenable to gradient-based learning.
  • The demonstrated generalization to unseen user counts implies a single trained model can serve a network as the number of active users changes, without retraining per configuration.
  • The cross-AP aggregation term is a concrete architectural addition that enables multi-AP reasoning; the paper's results suggest it is load-bearing for outperforming greedy assignment.
  • Both proposed methods outperform greedy and random AP-selection baselines, indicating that joint optimization of assignment and precoding is worthwhile even when only statistical CSI is available.

Reading between the lines

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

  • The paper only evaluates two APs; because the cross-AP aggregation term in Eq. (6) sums over APs in an invariant way, the architecture is plausibly extendable to more than two APs, but the paper does not test this.
  • At high SNR the GNN's advantage suggests it implicitly learns interference-aware coordination that the iterative baseline's local updates miss; inspecting the learned attention weights could reveal whether the GNN effectively performs a softer, global assignment before committing.
  • A natural stress test would be to train on one simulated environment and evaluate on measured indoor covariance matrices, or on a mixture of environments, to see whether the learned policy transfers beyond the single ray-traced room.
  • The paper's convergence plots show the iterative method needs roughly 1000 iterations; at typical MAC scheduling timescales, the GNN's single-pass inference may be the only feasible option, but the paper does not report absolute latency numbers.
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 / 4 minor

Summary. The paper considers joint access-point (AP) selection and precoder design for sum-rate maximization in a multi-AP MISO downlink under statistical CSI, where only channel covariance matrices are available. Two methods are proposed: an iterative alternating-optimization algorithm (Joint SWMMSE) that combines stochastic WMMSE precoder updates with projected-gradient assignment updates, and a single-forward-pass graph neural network (Joint GNN) built on an attention-based Edge-GNN with a new cross-AP aggregation term and a softmax AP-selection head. The authors claim that the GNN outperforms the iterative algorithm across the tested SNR range and generalizes to varying numbers of users, while both proposed approaches beat fixed greedy/random AP-selection baselines. The evaluation is carried out on covariance matrices produced by a conditional VAE trained on Sionna ray-traced indoor channels, with simulations over SNR, number of users, and number of antennas.

Significance. If the empirical claims hold, the paper offers a practical way to replace iterative stochastic optimization with a low-latency, permutation-equivariant GNN for joint AP selection and precoding under statistical CSI. The architectural extension is plausible and the generalization to varying user counts is a useful property, consistent with prior GNN-based precoding work. The paper also gives a clean relaxation argument for the assignment variables, and it evaluates on a nontrivial clustered-user scenario where greedy assignment is known to be suboptimal. The central weakness is that the headline 'outperforms across SNR' claim is not uniformly supported by the presented results, and the lack of uncertainty quantification and baseline-convergence analysis makes the main comparison difficult to assess.

major comments (4)
  1. [Abstract and Section IV-B, Fig. 2] The abstract and conclusion state that the GNN outperforms the iterative algorithm across the tested SNR range, but Fig. 2b for spatially clustered users shows Joint GNN performing only as well as Joint SWMMSE, and the text itself says 'still performs as well as Joint SWMMSE' in that scenario. The 'outperforms across SNR' claim is therefore an overstatement of the presented evidence; it should be limited to the uniform-user setting or the claim should be revised to reflect comparable performance on clustered users.
  2. [Section IV-B, Figs. 2-5] All reported results are averages over 1000 test scenarios without error bars, confidence intervals, or statistical significance tests. The differences between Joint GNN and Joint SWMMSE in Fig. 2a, especially at high SNR, may be within the stochastic variation of the test set. The authors should report variance or confidence intervals and, ideally, run multiple seeds or a significance test to support the claimed superiority.
  3. [Section IV-A and Algorithm 1] The Joint SWMMSE baseline is evaluated with a single hyperparameter configuration (lambda=0.1, rho=0.95, L_o=40, L_i=25, 1000 iterations) and no convergence criterion or sensitivity analysis. The assignment updates in line 14 rely on rate estimates from only 25 Monte Carlo samples, and Fig. 3 shows convergence behavior only at 20 dB. Without evidence that the baseline is well-converged and its hyperparameters are not tuned unfavorably, the observed GNN advantage could be a baseline-tuning artifact rather than a genuine performance gain.
  4. [Sections II-B and IV-B] The entire training and evaluation is conducted on covariance matrices generated by the conditional VAE of reference [10], which is itself trained on Sionna ray-traced indoor channels. The paper provides no experiments with other channel statistics, different ray-tracing configurations, or covariance mismatch, so the claimed improvements may not transfer to other indoor environments. This limitation should be acknowledged, and a robustness study (e.g., evaluating on covariances with additive errors or from a different propagation model) would strengthen the central claim.
minor comments (4)
  1. [Eq. (6), Section III-B] The activation function f_act is not defined in the general update rule; the text later mentions ReLU in the simulation section, but the notation should be specified where Eq. (6) is introduced.
  2. [Algorithm 1, line 13] The rate estimate in line 13 uses h_{k,n}^{(ell_i)} but the earlier channel samples in line 6 are written as h_{k,n}; the indexing should be made consistent to avoid confusion about whether the same samples are reused.
  3. [Figure 3] The two panels of Fig. 3 appear identical in the caption and the text does not explain what differs between them; the caption should clarify the distinction or the duplicate panel should be removed.
  4. [Section IV-B] The statement that 'the GNN backbone comprises five hidden layers' is not fully consistent with the description of L-1 backbone layers plus a final precoding layer; specifying L explicitly would remove ambiguity.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the simplex-relaxation argument is an external result attributed to [1], and the GNN-vs-SWMMSE advantage is an in-paper held-out benchmark; self-citations [10], [6], and [7] supply data and architecture without forcing the claimed outcome.

full rationale

The paper's derivation chain is self-contained for its central claims. The claimed losslessness of the assignment relaxation (Section III) is a standard mathematical fact—a linear objective over the simplex attains its optimum at a vertex—and the paper explicitly attributes the argument to external work: 'This relaxation and the underlying vertex argument follow from [1]', where [1] is Sanjabi, Razaviyayn, and Luo (IEEE TSP 2014), not a work of the present authors. The SWMMSE inner loop is taken from external reference [2]; the simplex projection from [12]; the attention mechanism originates in external reference [5] (A2D-GNN). The central empirical claim (GNN outperforms Joint SWMMSE across SNR) is supported by the paper's own out-of-sample experiments: positions are split disjointly into 800 training, 100 validation, and 100 test positions, the GNN is evaluated on held-out test sets with 1000 scenarios each, and is compared against independent baselines (Greedy SWMMSE, Random SWMMSE, Greedy Eigenbeam) plus an external perfect-CSI upper bound (Joint WMMSE from [1]). The self-citations that exist—[10] (shared authors Weißer and Utschick), which supplies the conditional-VAE covariance statistics used to generate every training and test scenario, and [6]/[7] (shared author Utschick), which supply the Edge-GNN backbone extended here with a cross-AP aggregation term and a softmax assignment head—are real but do not force the result: the GNN-vs-SWMMSE comparison could have gone either way and is demonstrated by the paper's Figures 2–5, not imported from the cited works. The paper's own explanation of the GNN advantage ('the GNN learns prior information during training') is a legitimate algorithmic mechanism, not a definitional identity. Concerns that the evaluation depends entirely on the self-cited cVAE channel statistics, and that the fixed Joint SWMMSE hyperparameters (Lo=40, Li=25, rho=0.95, lambda=0.1, 1000 iterations) may make the baseline underconverged, are validity and transfer risks, not circularity: they do not correspond to any equation in the paper reducing to a fitted quantity or to a self-citation by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard linear-programming facts, the statistical CSI assumption, the no-inter-AP-interference assumption, the reliability of the cVAE-generated channel model, and the expressive power of the chosen GNN architecture. No new physical entities are introduced.

free parameters (4)
  • GNN hyperparameters beta, gamma, delta = beta=0.1/M, gamma=0.1, delta=0.1/K
    Chosen by hand in Section IV-B; these control the relative weighting of self, same-AP interference, and cross-AP aggregation in Eq. (6), and they affect the GNN's behavior.
  • SWMMSE hyperparameters rho, lambda, epsilon = rho=0.95, lambda=0.1, epsilon=1e-6
    Set in Section IV-A; the forgetting factor rho and step size lambda directly control the iterative algorithm's convergence and performance.
  • Iteration counts L_o, L_i = L_o=40, L_i=25
    Chosen in Section IV-A; the number of outer and inner iterations determines the quality of the SWMMSE solution and the rate estimate.
  • GNN training hyperparameters = learning rate 3e-4, batch size 200, 500 epochs, hidden dim 256, 5 hidden layers
    Selected via validation in Section IV-B; these are hand-tuned and affect the reported performance.
assumptions (5)
  • standard math For fixed precoders, a linear objective over the simplex attains its optimum at a vertex, so the binary assignment constraint can be relaxed to the simplex without loss of optimality.
    Invoked in Section III, following [1]. This is a standard linear programming fact and is mathematically sound.
  • domain assumption Each channel h_{k,n} is zero-mean circularly symmetric complex Gaussian with known covariance C_{k,n}.
    Stated in Section II-A; this is the definition of statistical CSI used throughout the paper and is the premise of the SWMMSE method.
  • domain assumption Orthogonal resource allocation across APs, so there is no inter-AP interference.
    Assumed in Section II-A, paragraph after Eq. (2). This restricts the problem to scenarios where APs use orthogonal resources, which is critical for the simple interference model.
  • domain assumption The conditional VAE of reference [10] accurately produces channel covariance matrices from user positions in the simulated room.
    Used in Section II-B; all training and test data for the GNN and SWMMSE are generated from these covariances, so the validity of the results depends on this model.
  • ad hoc to paper The specific GNN update rule of Eq. (6), including the attention mechanism and the cross-AP aggregation term, is expressive enough to approximate the optimal assignment and precoding policy.
    The architecture is a design choice made in Section III-B; no proof of approximation capability is given. Performance is validated only empirically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Joint Access Point Selection and Precoder Design under Statistical CSI." pith.science (2026). https://pith.science/paper/FHPGHLJG

@misc{pith2026260806251,
  author       = {Pith},
  title        = {Pith review of: Joint Access Point Selection and Precoder Design under Statistical CSI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FHPGHLJG}},
  note         = {Machine review of arXiv:2608.06251}
}
read the original abstract

This work addresses joint access point (AP) selection and precoding for sum-rate maximization under statistical channel state information (CSI) in multi-AP multi-user systems. To this end, we propose two approaches. The first method is an iterative alternating optimization algorithm that updates the precoding vectors via the stochastic WMMSE (SWMMSE) algorithm and the assignment variables via a projected gradient descent step. The second method is a graph neural network (GNN)-based framework that solves the same problem in a single forward pass during inference. Building on an attention-based Edge-GNN architecture, we extend it to a multi-AP scenario, enabling the joint learning of assignment variables and precoding vectors from statistical CSI alone. Results show that the GNN outperforms the iterative algorithm across the tested signal-to-noise ratio (SNR) range and generalizes to varying numbers of users with comparable performance. Both approaches are also compared to various baseline techniques.

Figures

Figures reproduced from arXiv: 2608.06251 by the authors.

Figure 1
Figure 1. Proposed GNN-based framework. A shared GNN backbone of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The average sum-rate over the SNR. The GNN is trained on [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. The average sum-rate over the number of users [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: The average sum-rate over the number of antennas [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages

  1. [10]

    Context-Aware CSI Prediction for Access Point Selection Utilizing Conditional VAEs

    F. Weißer, A. Kasibovic, and W. Utschick, “Context-aware CSI predic- tion for access point selection utilizing conditional V AEs,” 2026, preprint arXiv:2604.13720

    work page Pith review arXiv 2026

  2. [1]

    Optimal joint base station assignment and beamforming for heterogeneous networks,

    M. Sanjabi, M. Razaviyayn, and Z.-Q. Luo, “Optimal joint base station assignment and beamforming for heterogeneous networks,”IEEE Trans. Signal Process., vol. 62, no. 8, pp. 1950–1961, 2014

    work page 1950

  3. [2]

    A stochastic weighted MMSE approach to sum rate maximization for a MIMO interference channel,

    M. Razaviyayn, M. S. Boroujeni, and Z.-Q. Luo, “A stochastic weighted MMSE approach to sum rate maximization for a MIMO interference channel,” inProc. IEEE 14th Workshop Signal Process. Adv. Wireless Commun. (SPA WC), 2013, pp. 325–329

    work page 2013

  4. [3]

    Learning precoding policy: CNN or GNN?

    B. Zhao, J. Guo, and C. Yang, “Learning precoding policy: CNN or GNN?” inProc. IEEE Wireless Commun. Netw. Conf. (WCNC), 2022, pp. 1027–1032

    work page 2022

  5. [4]

    A bipartite graph neural network approach for scalable beamforming optimization,

    J. Kim, H. Lee, S.-E. Hong, and S.-H. Park, “A bipartite graph neural network approach for scalable beamforming optimization,”IEEE Trans. Wireless Commun., vol. 22, no. 1, pp. 333–347, 2023

    work page 2023

  6. [5]

    Multidimensional graph neural networks for wireless communications,

    S. Liu, J. Guo, and C. Yang, “Multidimensional graph neural networks for wireless communications,”IEEE Trans. Wireless Commun., vol. 23, no. 4, pp. 3057–3073, 2024

    work page 2024

  7. [6]

    Statistical precoder design in multi-user systems via graph neural networks and generative modeling,

    N. Turan, S. Allaparapu, D. B. Amor, B. B ¨ock, M. Joham, and W. Utschick, “Statistical precoder design in multi-user systems via graph neural networks and generative modeling,”IEEE Wireless Commun. Lett., vol. 14, no. 5, pp. 1491–1495, 2025

    work page 2025

  8. [7]

    Scalable multi- user precoding and pilot optimization with graph neural networks,

    V . Rizzello, D. B. Amor, M. Joham, and W. Utschick, “Scalable multi- user precoding and pilot optimization with graph neural networks,” in Proc. IEEE Int. Conf. Commun. (ICC), 2024, pp. 2956–2961

    work page 2024

Show all 14 references
  1. [8]

    Learning resource allocation policy: Vertex-GNN or Edge-GNN?

    Y . Peng, J. Guo, and C. Yang, “Learning resource allocation policy: Vertex-GNN or Edge-GNN?”IEEE Trans. Mach. Learn. Commun. Netw., vol. 2, pp. 190–209, 2024

    2024

  2. [9]

    Joint beamforming and integer user association using a GNN with Gumbel–softmax reparameterizations,

    Q. Lyu and M. Vu, “Joint beamforming and integer user association using a GNN with Gumbel–softmax reparameterizations,”IEEE Trans. V eh. Technol., vol. 74, no. 11, pp. 18 214–18 219, 2025

    2025

  3. [11]

    Sionna: An open-source library for next-generation physical layer research,

    J. Hoydis, S. Cammerer, F. A. Aoudia, A. Vem, N. Binder, G. Marcus, and A. Keller, “Sionna: An open-source library for next-generation physical layer research,” 2022, preprint arXiv:2203.11854

    2022 arXiv

  4. [12]

    Efficient projections onto the l1-ball for learning in high dimensions,

    J. Duchi, S. Shalev-Shwartz, Y . Singer, and T. Chandra, “Efficient projections onto the l1-ball for learning in high dimensions,” inProc. 25th Int. Conf. Mach. Learn. (ICML), 2008, pp. 272–279

    2008

  5. [13]

    Efficient use of fading correlations in MIMO systems,

    M. Ivrlac, T. Kurpjuhn, C. Brunner, and W. Utschick, “Efficient use of fading correlations in MIMO systems,” inProc. IEEE 54th V eh. Technol. Conf. (VTC Fall), 2001, pp. 2763–2767

    2001

  6. [14]

    Group normalization,

    Y . Wu and K. He, “Group normalization,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2018

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.