Pith. sign in

REVIEW 3 major objections 7 minor 14 references

Shared AO update structure lets a tiny neural precoder adapt across antenna systems without large models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 14:25 UTC pith:HR7YG3PW

load-bearing objection Clean unfolded design that reuses AO update skeletons across FP/HP and FAS/RIS; real few-shot gains on small setups, but the “structure not size” claim is only partly isolated. the 3 major comments →

arxiv 2607.23738 v1 pith:HR7YG3PW submitted 2026-07-26 eess.SP

Cross-System Neural Precoder: Exploiting Structural Consistency for Fast Adaptation

Jia Guo , Chenyang Yang This is my paper
classification eess.SP
keywords precodingcross-system adaptationstructural learningalternating optimizationhybrid precodingRISMU-MIMOmodel-driven neural networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Learning-based precoders usually have to be redesigned when the system changes—fully digital versus hybrid, fixed antennas versus RIS-aided, single- versus multi-antenna users—because the variables and constraints change. This paper argues that the redesign is unnecessary for a broad class of spectral-efficiency problems. When alternating optimization freezes all but one type of variable, the remaining subproblems share the same computational skeleton across those systems. The authors turn that skeleton into a layered network (XNP) whose layers are the AO updates, with only lightweight element-wise networks learning step sizes and nonlinear corrections. The resulting model has a few thousand parameters. Pre-training on one configuration and fine-tuning on another reaches high spectral-efficiency ratios with far fewer samples and epochs than training a comparable graph network from scratch or fine-tuning it. The practical claim is that cross-system adaptability can come from reusing shared update structure rather than from scaling model size.

Core claim

For spectral-efficiency precoding over fully digital/hybrid and fixed-antenna/RIS configurations, the AO subproblems for digital precoders, analog precoders, and RIS phases share common computational forms once the other variables are fixed. An XNP whose layers implement those forms, retaining matrix multiplies and learning only element-wise nonlinear maps, therefore transfers across systems with only several thousand trainable parameters and substantially lower sample and epoch cost than a GNN baseline.

What carries the argument

Cross-system neural precoder (XNP): each layer realizes AO-inspired updates for V, U, V_RF, and Θ, keeping the shared multiplications and interference terms while replacing scalar steps and element-wise nonlinearities with identical small FNNs, plus conditional projections for power and constant-modulus constraints.

Load-bearing premise

The particular gradient-style update skeletons chosen for each variable, together with element-wise networks, really do capture the transferable structure across systems so that only those light maps need retraining when dimensions or constraints change.

What would settle it

On the paper’s own target setups (HP RIS and FP MU-MIMO), check whether a pre-trained XNP still reaches high SE ratio with far fewer samples and epochs than a same-size GNN; if the gap disappears or the reused modules fail when antenna counts or channel statistics leave the tested regime, the structural-reuse claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single XNP template can be pre-trained once and adapted to new antenna architectures or user antenna counts by activating only the relevant update modules.
  • Cross-system precoding adaptation need not rely on large foundation-style models or millions of samples when shared AO structure is explicit.
  • Model-driven networks can be designed around reusable update equations rather than around a single fixed system formulation.
  • Training cost for new deployments drops to fine-tuning a few thousand parameters instead of full re-architecture and re-training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same structural-reuse idea could be tried on other multi-variable wireless problems that already use AO (power control, beam tracking, ISAC) without inventing new large architectures.
  • If the element-wise FNNs mainly learn effective step sizes, even simpler parameterized updates might retain most of the transfer benefit.
  • Failure modes would likely appear first when the equivalent-channel abstraction breaks (strongly correlated users, hardware impairments, or objectives far from weighted MSE).

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The manuscript observes that, for a class of sum-rate-maximizing precoding problems (fully-digital/hybrid precoding in fixed-antenna and RIS-aided systems), the alternating-optimization (AO) subproblems for each variable type — digital precoder V, combiner U, analog precoder V_RF, RIS coefficients Θ — can be written in common forms once system differences are absorbed into equivalent channels (Eqs. (2), (6), (8)) and conditional projections. Building on this, the authors propose the cross-system neural precoder (XNP), a deep-unfolded architecture in which each layer implements AO-inspired gradient updates (Eqs. (14)–(17)) with the matrix multiplications hard-coded and only lightweight element-wise FNNs (a few thousand parameters per module) learned. Unsupervised pretraining on one system followed by fine-tuning on another is shown, in two transfer pairs (HP-MISO-FAS → HP-MISO-RIS; FP-MISO-RIS → FP-MIMO-FAS), to reach a given SE ratio with substantially fewer samples and epochs than direct training, and to far outperform a GNN baseline.

Significance. If the results hold, the paper makes a useful and timely point against the current "scale up" trend in wireless foundation models: cross-system adaptability for precoding can come from reusing problem structure rather than from large models. Specific strengths: (i) the structural claim is not merely asserted — Eqs. (2), (6), (8) genuinely exhibit the shared subproblem forms across four system configurations, and the equivalent-channel factorization is derived, not fitted; (ii) the evaluation is non-tautological — performance is measured as SE ratio against external numerical benchmarks (WMMSE [8] for FAS, block gradient descent [13] for RIS), not against quantities defined by the learned modules; (iii) the parameter count is honestly tiny (2.2K–2.6K per update module), and the second transfer pair (MISO→MIMO) does change variable dimensions, so the size-generalization mechanism is exercised, not just asserted; (iv) training is unsupervised with the negative SE as loss, so no labeled optimal precoders are needed. The main limitation is that the experiments do not yet isolate "structural reuse" from "small networks fitted to similar channel statistics," since both transfer pairs keep

major comments (3)
  1. [§V, Figs. 2–3 (with §III-A, Eqs. (4)–(5))] The headline attribution — that fast adaptation comes from reusing the shared AO update skeleton rather than from large models — is not separated from a competing explanation by the current experiments. Both transfer pairs keep K=3, N_B=8, and the Rician K-factor (10) fixed, changing only architecture (FAS→RIS) or N_U/M. Moreover, §III-A adopts W_k ≈ I, so the retained gradient forms in (4)–(5) are steps on the unweighted sum-MSE, not on the SE objective (1a); the mismatch with the true SE-gradient direction depends on interference statistics and is therefore system-dependent, and the only components that can absorb it are the element-wise FNNs (which §IV-B2 says 'learn an effective update direction and step size'). What is reused on the target is thus a bundle of invariant skeleton plus source-fitted correction, and the two are never disentangled. A concrete fix: add one or both of (a)
  2. [§V, Figs. 2–3, GNN baseline] The GNN baseline appears to fail almost entirely: the text states all GNN curves (Direct and Tune) overlap and sit far below XNP, and in Fig. 3 the GNN results are omitted 'due to their poor performance.' A baseline that does not learn at all is uninformative, and the claim of 'significantly fewer training samples and epochs than a graph neural network-based baseline' (Abstract) rests on it. Matching layer and neuron counts to the XNP does not by itself make the comparison fair, since the GNN of [14] was designed for a different input/output structure and may need its own tuning (learning rate, normalization, input features) in this unfolded-update setting. Please either diagnose and fix the GNN's collapse (reporting the hyperparameters tried), or soften the comparative claim to rest on the internal XNP-Tune vs. XNP-Direct comparison, which is valid and already supports the fast-adaptati
  3. [§IV-C and §V (experimental protocol)] The accounting of which modules are reused, frozen, or trained from scratch in each transfer pair is incomplete, which limits reproducibility and bears on how much of the gain is 'reuse.' §V states that source 1 (HP-MISO-FAS) trains M_VRF and M_V, and source 2 (FP-MISO-RIS) trains M_V and M_Θ. But: (i) in target 2 (FP-MIMO-FAS), MIMO requires matrix combiners U_k, while the MISO source has scalar U_k — it is never stated whether M_U is learned at all in the sources, computed in closed form, or trained fresh on the target; (ii) it is not stated whether fine-tuning on the target updates all FNNs or only those for variables absent in the source (§IV-C suggests the latter, but the learning curves in Figs. 4–5 are ambiguous on this); (iii) target fine-tuning sample counts and the layer-0 initialization of V, U, V_RF, Θ are not given. Please add a per-pair table listing, for each module, its p
minor comments (7)
  1. [§III-B, Eq. (7)] Dimensional consistency of (7) is unclear: with V_RF of size N_B × N_RF, the first term Σ_j H̃_j^H U_j^H V_j^H appears to have dimension N_B × N_B. Either M = N_RF is being assumed or some terms are missing a factor; please check and clarify.
  2. [§III-A, Eqs. (4)–(5)] State explicitly whether the step size λ appears anywhere in the final XNP or is fully absorbed into the element-wise FNNs; as written, (14)–(17) drop λ without comment.
  3. [Figs. 2–5] The figure legends and axis labels are rendered as garbled unicode escape sequences in the compiled PDF, making the figures unreadable as shipped; this must be fixed. Also, Fig. 3's y-axis exceeds 100% SE ratio — a sentence noting that the numerical benchmarks ([8], [13]) are themselves suboptimal would preempt confusion.
  4. [Table I] J^(ℓ) is used but never given a numerical value; without it the reported parameter counts (2.6K/2.2K/2.6K) cannot be verified. Please show the parameter-count derivation.
  5. [Throughout] Terminology: 'alternative optimization' should be 'alternating optimization' (the standard term). Typo: 'learn a effective update direction' → 'an effective'. The notation H̃_k is redefined with different meanings in §III-A/B/C; a summary table of the equivalent-channel definitions would help.
  6. [§I and §V] An inference-complexity comparison (FLOPs or runtime) of the L-layer XNP against the numerical benchmarks would strengthen the efficiency story; currently only training-sample efficiency is quantified.
  7. [Footnote 1, §IV-B1] The argument that multiplication is hard to learn because it is non-monotonic is informal; either cite supporting evidence or soften to a motivation.

Circularity Check

0 steps flagged

No significant circularity: method paper with external SE benchmarks; shared AO structure is derived by inspection, not by fitting the evaluation metric.

full rationale

The load-bearing chain is: (i) write SE-max precoding under FP/HP × FAS/RIS in a common form (1); (ii) fix all but one variable type and observe that the resulting AO subproblem objectives (2), (6), (8) and example iterative updates (4)–(5), (7), (9) share the same algebraic skeleton once equivalent channels absorb system-specific factors; (iii) implement those skeletons as layer maps (14)–(17) with only element-wise FNNs learned; (iv) measure SE ratio against independent numerical solvers (WMMSE [8] for FAS, block gradient descent [13] for RIS). None of these steps defines the reported figure of merit in terms of the fitted FNN weights, nor does any step fit a parameter on a subset and then call a tautologically related quantity a prediction. Self-citations ([9], [10], [14]) supply the W_k≈I design choice, projection/activation motifs, and the GNN baseline; they do not underwrite a uniqueness claim that forces the adaptation result, and the adaptation claims remain externally falsifiable on held-out channels and configurations. Attribution debates (whether transfer is pure structure vs. small nets on similar Rician statistics) are correctness/experimental-design issues, not circularity. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central adaptation claim rests on standard multiuser MIMO rate and AO practice, a WMMSE-style gradient skeleton with an explicit weight approximation, hand-chosen network depth/width, and synthetic Rician geometry—not on new physical entities. Free parameters are architectural and training choices; invented entity is the XNP template itself.

free parameters (4)
  • XNP depth L = 6
    Number of AO-inspired layers set to 6 in experiments; controls capacity and iteration depth of the unfolded solver.
  • Element-wise FNN widths and layouts = [16,16,16,16,16,16]; M_V/M_Θ: [J×3,J]; M_VRF: [J×3,8,J]
    Hidden sizes [16]×6 and per-module I/O patterns in Table I are design choices that determine the few-thousand trainable weights and adaptation behavior.
  • Source/target training sample counts and epochs = pretrain 10000 samples; fine-tune sweeps as in Figs. 2–5
    Pretrain uses 10,000 channel samples; fine-tune sweeps (e.g., 10–1000 samples, 1–500 epochs) define the reported data-efficiency claims.
  • Effective step-size / nonlinear maps inside FNN_*_ele
    Classical λ in gradient updates is absorbed into learned element-wise FNNs; those weights are fit to SE loss and are essential to reported performance.
axioms (5)
  • domain assumption Sum-rate precoding under power and optional constant-modulus constraints admits useful AO decomposition into digital, analog, and RIS blocks.
    §II–III formulate the common problem and analyze per-variable subproblems; standard in MU-MIMO/RIS literature.
  • ad hoc to paper W_k ≈ I approximation in the WMMSE reformulation still yields useful gradient updates for V and U.
    §III-A cites [9] and adopts the approximation to avoid full weighted-MSE machinery; load-bearing for the shared digital/combiner modules.
  • ad hoc to paper Matrix multiplications implementing interference coupling should be hard-coded; only element-wise nonlinearities need learning for size-generalization and sample efficiency.
    §IV-B argues black-box FNNs struggle with multiplications and proposes retaining them; core design axiom of XNP.
  • domain assumption Synthetic Rician channels with given path-loss and fixed small system sizes are adequate to demonstrate cross-system adaptation.
    §V simulation setup; external validity of the adaptation claim depends on this.
  • domain assumption Power and unit-modulus constraints can be enforced by normalization/projection (or activation design) after each update without destroying learned directions.
    Stated in §III opening and §IV-B; standard but necessary for feasibility of outputs.
invented entities (1)
  • Cross-system neural precoder (XNP) no independent evidence
    purpose: Single AO-layered network template with activatable V/U/V_RF/Θ update modules and element-wise FNNs for cross-configuration precoding.
    Named architecture introduced in §IV; empirical object whose transfer performance is the paper’s main result.

pith-pipeline@v1.2.0-grok45-kimik3 · 18352 in / 3872 out tokens · 84039 ms · 2026-07-30T14:25:05.027144+00:00 · methodology

0 comments
read the original abstract

Adapting learning-based precoding across different system configurations is challenging due to multiple types of variables and constraints. While large-scale neural networks have been proposed for cross-task adaptation, whether such adaptability requires large models remains unclear. In this paper, we identify a structural property of a class of precoding problems: the subproblems associated with each type of variable in alternative optimization (AO) share a common computational structure across systems when other variables are fixed. This structural consistency enables the reuse of update rules across systems. Based on this observation, we propose a cross-system neural precoder (XNP), where each layer implements AO-inspired update equations, which define the layer-wise input-output mappings. By reusing common update structures and learning only lightweight nonlinear mappings, the XNP enables efficient adaptation across systems only with several thousand trainable parameters. Simulation results show that pre-trained XNPs achieve fast adaptation to new configurations with significantly fewer training samples and epochs than a graph neural network-based baseline. This demonstrates that cross-system adaptability can be achieved by exploiting shared computational structure, rather than relying on large models.

Figures

Figures reproduced from arXiv: 2607.23738 by Chenyang Yang, Jia Guo.

Figure 1
Figure 1. Figure 1: Architecture of XNP. B. Design of Update Equations 1) A straightforward design: We take M(ℓ) V (·) as an exam￾ple. From (4), M(ℓ) V should be a function of H˜ (ℓ) j , U (ℓ) j , j = 1, · · · , K and V (ℓ) k , where H˜ (ℓ) j is the equivalent channel for the j-th user in the ℓ-th layer. A straightforward way of designing M(ℓ) V is to use a neural network, say FNN, to learn the function, i.e., M(ℓ) V : V (ℓ+1… view at source ↗
Figure 5
Figure 5. Figure 5: Learning curve when the tar￾get system is FP in MU-MIMO FAS. pre-training the XNPs helps cross-system adaptation. VI. CONCLUSIONS This work studies cross-system adaptation for precoding from a structural viewpoint. Instead of treating different sys￾tem configurations as independent learning tasks, we exam￾ine the update procedures in alternative optimization (AO) and observe that they exhibit similar compu… view at source ↗
Figure 3
Figure 3. Figure 3: Performance of XNP for FP in MU-MIMO FAS system. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 4 linked inside Pith

  1. [1]

    WirelessGPT: A generative pre- trained multi-task learning framework for wireless communication,

    T. Yang, P. Zhang, M. Zhenget al., “WirelessGPT: A generative pre- trained multi-task learning framework for wireless communication,” IEEE Netw., vol. 39, no. 5, pp. 58–65, Sep. 2025

    2025

  2. [2]

    Large wireless model (LWM): A foundation model for wireless channels,

    S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (LWM): A foundation model for wireless channels,”arXiv:2411.08872, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Large language model-empowered channel prediction and predictive beamforming for leo satellite communications,

    Z. Chen, H. Shin, A. Nallanathan, and J. Chambers, “Large language model-empowered channel prediction and predictive beamforming for leo satellite communications,”arXiv:2510.10561, 2025

    arXiv 2025

  4. [4]

    Lvm4csi: Enabling direct application of pre-trained large vision models for wireless channel tasks,

    J. Guo, P. Jiang, C.-K. Wen, S. Jin, and J. Zhang, “Lvm4csi: Enabling direct application of pre-trained large vision models for wireless channel tasks,”arXiv preprint arXiv:2507.05121, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    A wireless foundation model for multi-task prediction,

    Y . Sheng, J. Wang, X. Zhou, L. Liang, H. Ye, S. Jin, and G. Y . Li, “A wireless foundation model for multi-task prediction,”arXiv:2507.05938, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    6G-oriented CSI-based multi-modal pre- training and downstream task adaptation paradigm,

    T. Jiao, C. Ye, Y . Huang, Y . Feng, Z. Xiao, Y . Xu, D. He, Y . Guan, B. Yang, J. Changet al., “6G-oriented CSI-based multi-modal pre- training and downstream task adaptation paradigm,”IEEE ICC Work- shops, 2024

    2024

  7. [7]

    Iterative algorithm induced deep-unfolding neural networks: Precoding design for multiuser MIMO systems,

    Q. Hu, Y . Cai, Q. Shi, K. Xu, G. Yu, and Z. Ding, “Iterative algorithm induced deep-unfolding neural networks: Precoding design for multiuser MIMO systems,”IEEE Trans. Wireless Commun., vol. 20, no. 2, pp. 1394–1410, Feb. 2021

    2021

  8. [8]

    An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,

    Q. Shi, M. Razaviyayn, Z. Luoet al., “An iteratively 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

    2011

  9. [9]

    A size-generalizable graph neural network for learning multi-user multi-stream MIMO precoding,

    J. Guo and C. Yang, “A size-generalizable graph neural network for learning multi-user multi-stream MIMO precoding,”IEEE MLSP, 2024

    2024

  10. [10]

    Recursive GNNs for learning precoding policies with size- generalizability,

    ——, “Recursive GNNs for learning precoding policies with size- generalizability,”IEEE Trans. Mach. Learn Commun. Netw., vol. 2, pp. 1558–1579, 2024

    2024

  11. [11]

    Low-complexity joint beamforming for RIS-assisted MU-MISO systems based on model-driven deep learn- ing,

    W. Jin, J. Zhang, C.-K. Wenet al., “Low-complexity joint beamforming for RIS-assisted MU-MISO systems based on model-driven deep learn- ing,”IEEE Trans. Wireless Commun., vol. 23, no. 7, pp. 6968–6982, July 2024

    2024

  12. [12]

    When attention is beneficial for learning wireless resource allocation efficiently?

    J. Guo and C. Yang, “When attention is beneficial for learning wireless resource allocation efficiently?”arXiv:2507.02427, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Weighted sum-rate maximization for reconfigurable intelligent surface aided wireless networks,

    H. Guo, Y .-C. Liang, J. Chenet al., “Weighted sum-rate maximization for reconfigurable intelligent surface aided wireless networks,”IEEE Trans. Wireless Commun., vol. 19, no. 5, pp. 3064–3076, May 2020

    2020

  14. [14]

    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, April 2024

    2024