Pith. sign in

REVIEW 2 major objections 5 minor 15 references

A shift-plus-experts neural codec reconstructs multi-user pinching-antenna CSI over noisy FDD feedback better than standard learning baselines.

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

2026-07-11 13:24 UTC pith:OBXLHMLL

load-bearing objection Competent letter-length engineering paper: Shift-MoE DJSCC for multi-user PASS CSI feedback works in synthetic sims, novelty is the application, perfect feedback CSI is the main soft spot. the 2 major comments →

arxiv 2607.04795 v1 pith:OBXLHMLL submitted 2026-07-06 eess.SP

Shift-MoE-Based DJSCC for CSI Feedback in Multi-User Pinching-Antenna Systems

Jian Zou , Yifan Lian , Yongsheng Liang , Fanyang Meng , Wenwu Xie , Liang Yang , Jian Xiao This is my paper
classification eess.SP
keywords pinching-antenna systemsCSI feedbackdeep joint source-channel codingmixture-of-expertsshift operationFDD multi-userNMSE
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.

Pinching-antenna systems can reshape wireless propagation cheaply, but in FDD networks the base station still needs accurate downlink CSI from each user over a noisy uplink feedback link. That CSI lives on a waveguide-by-antenna grid with clear local structure, yet the statistics differ strongly from user to user, so a single fixed compression map fails to generalize. The authors design an end-to-end deep joint source-channel codec called Shift-MoE: channel-grouped one-step shifts capture the grid correlations without quadratic attention, while a gated mixture of MLP experts adapts the mapping to each user's CSI distribution. Joint training over a Rician feedback channel yields lower normalized reconstruction error than CsiNet, plain MLP, Transformer, and shift-only variants across user counts, antenna sizes, and compression ratios. The result matters because it shows how structure-aware, user-adaptive feedback can unlock the promised multi-user gains of pinching antennas under realistic FDD constraints.

Core claim

An end-to-end DJSCC architecture that combines channel-grouped one-step feature shifts with a gated MLP mixture-of-experts module reconstructs multi-user PASS CSI more accurately over a noisy uplink than representative learning-based CSI feedback schemes, and the gains hold under changes in user number, waveguide-PA size, and feedback rate.

What carries the argument

Shift-MoE: deterministic one-step shifts of channel groups along the waveguide and pinching-antenna axes, followed by a token-wise gated mixture of MLP experts that routes each CSI realization to a weighted combination of specialists.

Load-bearing premise

The base station is given perfect knowledge of the instantaneous uplink feedback fading coefficient when it equalizes the received symbols, so residual feedback-channel estimation error never appears in the reported results.

What would settle it

Retrain and re-evaluate the same Shift-MoE pipeline when the equalizer uses only a noisy pilot-based estimate of the feedback fading coefficient; if the NMSE advantage over CsiNet, Transformer, and ShiftViT disappears or reverses, the central claim fails under realistic feedback CSI.

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

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

2 major / 5 minor

Summary. This letter studies CSI feedback for multi-user pinching-antenna systems (PASS) under FDD, where reciprocity is unavailable and the uplink feedback link is noisy. The authors formulate an end-to-end DJSCC problem that maps each user’s complex M×N PA-level CSI matrix to a compact latent code, transmits it over a quasi-static Rician feedback channel, and reconstructs the CSI at the BS. The proposed Shift-MoE encoder/decoder uses channel-grouped one-step shifts to capture waveguide–PA grid correlations without global attention, and a gated MLP mixture-of-experts module to adapt to heterogeneous multi-user CSI statistics. Numerical results on synthetic geometric multipath data report consistent NMSE gains over CsiNet, MLP, Transformer, and ShiftViT-without-MoE baselines across SNR, user count, compression ratio, and antenna configurations, with expert gating weights shown to vary across users.

Significance. PASS is an emerging reconfigurable architecture; CSI feedback under FDD is a practical bottleneck that has received little dedicated treatment. The paper’s contribution is a concrete, low-complexity architecture that couples structure-aware shift interactions with input-adaptive MoE under a DJSCC objective, together with systematic ablations (shift vs. MLP; MoE vs. ShiftViT) and scaling curves over K, η, and (M,N). If the gains hold under more realistic feedback-channel estimation and non-PASS controls, the design would be a useful template for grid-structured CSI feedback. Strengths include a clearly stated system model and training objective, explicit complexity comparison to MHSA, and multi-parameter numerical validation rather than a single operating point.

major comments (2)
  1. §II-B after Eq. (10): the simulations set ˆh_fb_k = h_fb_k (idealized perfect feedback CSI at the equalizer). Residual feedback-channel estimation error is never injected, so the reported NMSE curves are optimistic upper bounds on robustness. The paper itself lists imperfect feedback-channel knowledge as future work (§V). For a DJSCC claim whose central selling point is robustness over a noisy uplink, at least one imperfect-ˆh_fb curve (or a short sensitivity study) is load-bearing and should be added or the claim should be explicitly scoped to perfect equalization.
  2. §IV, Figs. 3–4 and the Abstract claim: all methods appear trained/evaluated under the same perfect-equalization DJSCC setup, and the baselines (CsiNet, TransNet-style Transformer, MLP) are generic CSI-feedback nets not re-designed for the waveguide–PA grid. There is no non-PASS (e.g., standard massive-MIMO) control and no separate source-coding-plus-channel-coding baseline. Without these, it remains unclear whether the gains are PASS-structure-specific or simply those of a generally stronger DJSCC backbone (shift + MoE). A short control experiment or a clearer attribution discussion is needed to support the PASS-specific framing of the strongest claim.
minor comments (5)
  1. Table I and §III: free hyperparameters (dm, N1, N2, ndiv, α1, α2) are listed but not ablated; a brief sensitivity note would strengthen reproducibility.
  2. §II-A, Eq. (6)–(8): the waveguide radiation vector and the definition of Hk are clear, but the relationship between the effective channel gk(Λ) and the matrix fed back could be restated once for readers less familiar with PASS.
  3. Fig. 3 legend and text: “Transformer-based schemes [7], [15]” should name the exact variants used (e.g., TransNet) so the comparison is reproducible.
  4. Minor notation: compression ratio η ≜ L/(2MN) is defined after Eq. (9); ensure L vs. Lc is used consistently in the text and figures.
  5. Fig. 5: expert gating weights for five users are informative; stating the SNR and (M,N,η) at which they were collected would help interpretation.

Circularity Check

0 steps flagged

No significant circularity: empirical DJSCC architecture trained and evaluated on simulated PASS CSI; NMSE gains are experimental, not definitional or forced by self-citation.

full rationale

The paper proposes a neural encoder–decoder (channel-grouped one-step shifts + gated MLP MoE) inside an end-to-end DJSCC loop that minimizes expected Frobenius reconstruction error (Eq. 12) over a generative multipath PASS channel model and a Rician feedback link. All reported claims are numerical NMSE comparisons against CsiNet, MLP, Transformer, and ShiftViT-w/o-MoE under identical training/evaluation conditions (Figs. 3–4, Table I). There is no first-principles derivation, no uniqueness theorem, no parameter fitted on a subset and then re-labeled as a prediction, and no load-bearing self-citation that forces the architecture or the NMSE curves. Minor author-overlap citations ([6], [13], [15]) supply related prior CSI-feedback or DJSCC context but are not used to import uniqueness or to define the target quantity. The idealized equalizer ˆh_fb = h_fb is an assumption about the simulation setting, not a circular reduction of a claimed result to its inputs. The work is therefore self-contained against its own generative benchmarks; circularity score is zero.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

Central performance claims rest on a geometric multipath PASS channel model, an idealized perfect-feedback-CSI equalizer, a suite of hand-chosen network hyperparameters, and synthetic data drawn from the same generative process used for evaluation. No new physical constants are fitted; the free parameters are architectural and simulation knobs.

free parameters (5)
  • embedding dimension dm = 48
    Set to 48 in Table I; controls feature width of every Shift-MoE block and therefore capacity and reported NMSE.
  • number of Shift-MoE blocks N1 and experts N2 = 6, 5
    Hand-chosen as 6 and 5; depth and expert count directly affect the claimed gains over ShiftViT-without-MoE.
  • channel group number ndiv and expansion factors alpha1, alpha2 = 12, 2, 2
    ndiv=12, alpha1=alpha2=2 fix the shift grouping and MLP widths; not derived from first principles.
  • number of scatterers P and geometric region sizes = P=3; 50 m x 6 m; height 5 m
    P=3, Dx=50 m, Dy=6 m, a=5 m define the synthetic CSI distribution on which all NMSE curves are measured.
  • learning rate, batch size, epochs = 1e-4, 32, 300
    1e-4, 32, 300; training schedule that produces the published curves.
axioms (4)
  • domain assumption PA-level channel is LoS free-space path plus P discrete NLoS scatterers with complex gains beta_k,p (Eqs. 1–5).
    All training and test CSI matrices are generated from this geometric multipath model; real PASS channels may differ.
  • ad hoc to paper Uplink feedback link is quasi-static flat Rician fading with perfect instantaneous CSI available at the BS equalizer (ˆh_fb = h_fb).
    Stated explicitly in §II-B to isolate the DJSCC mapping; removes a major practical impairment.
  • domain assumption Uniform coupling gains p_m,n = 1/sqrt(N) and fixed refractive index i_ref = 1.44.
    Used to form the waveguide radiation vector and CSI matrix entries (Eqs. 6–8).
  • domain assumption End-to-end MSE minimization over the joint source-channel mapping is a valid proxy for useful CSI feedback quality.
    Standard DJSCC training objective (Eq. 12); downstream beamforming utility is not measured.
invented entities (1)
  • Shift-MoE block (channel-grouped one-step shifts + gated MLP MoE) no independent evidence
    purpose: Capture waveguide-PA grid correlations cheaply and adapt to heterogeneous multi-user CSI statistics inside a DJSCC encoder/decoder.
    The composite block is the paper’s architectural contribution; its value is demonstrated only by the synthetic NMSE experiments in this work.

pith-pipeline@v1.1.0-grok45 · 13344 in / 3113 out tokens · 28516 ms · 2026-07-11T13:24:24.220505+00:00 · methodology

0 comments
read the original abstract

In frequency-division duplexing systems, the performance gains of pinching-antenna systems (PASS) critically depend on accurate channel state information (CSI) at the base station. However, PASS CSI exhibits structured correlations over the waveguide-antenna grid and pronounced heterogeneity across users, making conventional fixed feedback mappings difficult to generalize. To address this challenge, this letter proposes an end-to-end CSI feedback scheme over a noisy uplink feedback link based on deep joint source-channel coding, termed Shift-based Mixture-of-Experts (Shift-MoE). Specifically, Shift-MoE leverages channel-grouped one-step shift operations to capture grid dependencies without global attention, and employs a gated multilayer perceptron mixture-of-experts module to adapt to heterogeneous CSI statistics across users. Numerical results demonstrate that the proposed Shift-MoE consistently outperforms representative learning-based CSI feedback baselines in normalized mean squared error and remains effective under different system parameter settings.

Figures

Figures reproduced from arXiv: 2607.04795 by Fanyang Meng, Jian Xiao, Jian Zou, Liang Yang, Wenwu Xie, Yifan Lian, Yongsheng Liang.

Figure 1
Figure 1. Figure 1: Pinching antennas assisted multi-user systems. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Proposed Shift-MoE encoder architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: compares the NMSE performance versus SNR under the configuration of K = 5, M = 16, N = 8, and η = 1/2. The compared methods include CsiNet [4], an MLP baseline without the shift module, Transformer-based schemes [7], [15], as well as the proposed ShiftViT without MoE and Shift-MoE. As SNR increases, the NMSE of all methods decreases due to reduced feedback-link distortions. Compared with the MLP baseline, … view at source ↗
Figure 4
Figure 4. Figure 4: NMSE performance of the proposed Shift-MoE under various parameter settings. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The proposed Shift-MoE gating weights for different users. [PITH_FULL_IMAGE:figures/full_fig_p005_5.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

15 extracted references

  1. [1]

    Movable-antenna-assisted covert communications with reconfigurable intelligent surfaces,

    W. Xie, Z. Li, C. Yu, H. Xu, J. Wang, W. Wu, X. Li, and L. Yang, “Movable-antenna-assisted covert communications with reconfigurable intelligent surfaces,”IEEE Internet Things J., vol. 12, no. 9, pp. 12369- 12382, 1 May1, 2025

    2025

  2. [2]

    Multi- layer transmitting RIS-aided receiver for collaborative jamming and anti- jamming networks,

    J. Li, L. Yang, W. Hao, I. Ahmad, H. Liu, F. Shu, and D. Niyato, “Multi- layer transmitting RIS-aided receiver for collaborative jamming and anti- jamming networks,”IEEE Trans. Wireless Commun., vol. 24, no. 8, pp. 6518-6534, Aug. 2025

    2025

  3. [3]

    Pinching- antenna systems: Architecture designs, opportunities, and outlook,

    Y . Liu, Z. Wang, X. Mu, C. Ouyang, X. Xu, and Z. Ding, “Pinching- antenna systems: Architecture designs, opportunities, and outlook,”IEEE Commun. Mag., vol. 64, no. 1, pp. 190–196, Jan. 2025

    2025

  4. [4]

    Modeling and beamforming optimization for pinching-antenna systems,

    Z. Wang, C. Ouyang, X. Mu, Y . Liu and Z. Ding, “Modeling and beamforming optimization for pinching-antenna systems,”IEEE Trans. Commun., vol. 73, no. 12, pp. 13904-13919, Dec. 2025

    2025

  5. [5]

    Deep learning for massive MIMO CSI feedback,

    C. -K. Wen, W. -T. Shih and S. Jin, “Deep learning for massive MIMO CSI feedback,”IEEE Wireless Commun. Lett., vol. 7, no. 5, pp. 748-751, Oct. 2018

    2018

  6. [6]

    Distributed learning-based channel estimation and feedback for RIS-aided wireless communications,

    J. Zou, Q. Mao, J. Xiao, S. Liu and Y . Liang, “Distributed learning-based channel estimation and feedback for RIS-aided wireless communications,” IEEE Wireless Commun. Lett., vol. 14, no. 2, pp. 460-464, Feb. 2025

    2025

  7. [7]

    TransNet: Full attention network for CSI feedback in FDD massive MIMO system,

    Y . Cui, A. Guo, and C. Song, “TransNet: Full attention network for CSI feedback in FDD massive MIMO system,”IEEE Wireless Commun. Lett., vol. 11, no. 5, pp. 903–907, May 2022

    2022

  8. [8]

    Deep joint source- channel coding for wireless image transmission,

    E. Bourtsoulatze, D. Burth Kurka and D. G ¨und¨uz, “Deep joint source- channel coding for wireless image transmission,”IEEE Trans. Cogn. Commun. Netw., vol. 5, no. 3, pp. 567-579, Sept. 2019

    2019

  9. [9]

    Deep joint source-channel coding for CSI feedback: An end-to-end approach,

    J. Xu, B. Ai, N. Wang and W. Chen, “Deep joint source-channel coding for CSI feedback: An end-to-end approach,”IEEE J. Sel. Areas Commun.,, vol. 41, no. 1, pp. 260-273, Jan. 2023

    2023

  10. [10]

    Channel estimation for mmWave pinching-antenna systems,

    G. Zhou, V . K. Papanikolaou, Z. Ding and R. Schober, “Channel estimation for mmWave pinching-antenna systems,” inProc. IEEE 26th Int. Workshop Signal Process. Artif. Intell. Wireless Commun. (SPA WC), Jul. 2025, pp. 1–5

    2025

  11. [11]

    Channel estimation for pinching-antenna systems (PASS),

    J. Xiao, J. Wang and Y . Liu, “Channel estimation for pinching-antenna systems (PASS),”IEEE Commun. Lett., vol. 29, no. 8, pp. 1789-1793, Aug. 2025

    2025

  12. [12]

    When shift operation meets vision transformer: An extremely simple alternative to attention mechanism

    G. Wang, Y . Zhao, C. Tang, C. Luo, and W. Zeng, “When shift operation meets vision transformer: An extremely simple alternative to attention mechanism.” inProc. AAAI Conf. Artif. Intell., V ol. 36, 2022, pp. 7083- 7093

    2022

  13. [13]

    Superimposed pilot-based adaptive semantic communications for wire- less image transmission,

    J. Zou, J. Xiao, W. Xie, F. Meng, R. Feng, L. Yang, and Y . Liang, “Superimposed pilot-based adaptive semantic communications for wire- less image transmission,”IEEE Trans. Wireless Commun., vol. 25, pp. 17576-17590, 2026

    2026

  14. [14]

    A survey on mixture of experts in large language models,

    W. Cai, J. Jiang, F. Wang, J. Tang, S. Kim, and J. Huang, “A survey on mixture of experts in large language models,”IEEE Trans. Knowl. Data Eng., vol. 37, no. 7, pp. 3896–3915, Jul. 2025

    2025

  15. [15]

    Quan-transformer based channel feedback for RIS-aided wireless communication systems,

    W. Xie, J. Zou, J. Xiao, M. Li and X. Peng, “Quan-transformer based channel feedback for RIS-aided wireless communication systems,”IEEE Commun. Lett., vol. 26, no. 11, pp. 2631-2635, Nov. 2022

    2022