pith. sign in

arxiv: 2509.06820 · v1 · submitted 2025-09-08 · 📡 eess.SP · cs.AI· cs.IT· cs.LG· cs.NI· math.IT

Green Learning for STAR-RIS mmWave Systems with Implicit CSI

Yu-Hsiang Huang , Po-Heng Chou , Wan-Jen Huang , Walid Saad , C.-C. Jay Kuo This is my paper

Pith reviewed 2026-05-18 18:00 UTC · model grok-4.3

classification 📡 eess.SP cs.AIcs.ITcs.LGcs.NImath.IT
keywords STAR-RISmmWavegreen learningprecodingimplicit CSIbroadcastingspectral efficiencyXGBoost
0
0 comments X

The pith

Green learning predicts STAR-RIS coefficients and precoders directly from uplink pilots without explicit CSI estimation.

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

The paper proposes replacing both iterative optimization and deep neural networks with a green learning pipeline for designing transmit precoders and STAR-RIS transmission-reflection coefficients in mmWave MIMO broadcast systems. Training uses supervision generated by block coordinate descent under perfect CSI, yet inference runs entirely on received uplink pilot signals with no channel estimates required. The approach targets broadcasting scenarios in which many users receive identical information, aiming to cut redundant transmissions, lower power use, and enable real-time operation on resource-limited hardware. Simulations indicate that spectral efficiency stays competitive with both BCD and DL baselines while floating-point operations drop by more than four orders of magnitude.

Core claim

A green learning model that integrates Saab subspace approximation, relevant feature test selection, and XGBoost decision trees can jointly output the STAR-RIS coefficients and base-station precoder from uplink pilots alone. Although trained on labels produced by block coordinate descent assuming perfect CSI, the model performs fully CSI-free inference and reaches spectral efficiency values comparable to both the optimization baseline and deep-learning alternatives while requiring over 10,000 times fewer floating-point operations.

What carries the argument

Green learning pipeline combining Saab subspace approximation with adjusted bias, RFT-supervised feature selection, and XGBoost decision learning to map received pilots to STAR-RIS coefficients and precoder.

If this is right

  • Real-time precoding becomes feasible on hardware with severe compute and energy limits.
  • Broadcasting to multiple users sharing the same content can proceed without repeated CSI acquisition overhead.
  • Power consumption drops because redundant transmissions are avoided and iterative solvers are eliminated at runtime.

Where Pith is reading between the lines

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

  • The same lightweight pipeline could be retrained for other reconfigurable surface geometries or frequency bands without redesigning the core architecture.
  • Deployment on edge devices becomes practical because the inference cost is low enough for continuous operation.
  • Robustness checks under time-varying channels would clarify how often the model needs retraining when channel statistics drift.

Load-bearing premise

Labels generated by block coordinate descent under perfect CSI remain accurate enough training targets for a model that must later run without any channel state information.

What would settle it

A test set of mmWave channels with mobility or scattering statistics outside the training distribution in which the green learning spectral efficiency falls measurably below the BCD benchmark.

Figures

Figures reproduced from arXiv: 2509.06820 by C.-C. Jay Kuo, Po-Heng Chou, Walid Saad, Wan-Jen Huang, Yu-Hsiang Huang.

Figure 1
Figure 1. Figure 1: The STAR-RIS-aided MIMO system. runtime without relying on CSI estimation. This architecture eliminates the need for DNNs, yielding ultra-low floating￾point operations (FLOPs) and enabling real-time deployment on constrained hardware. In summary, our key contributions include: • We propose a GL-based precoding framework for STAR￾RIS-assisted mmWave broadcasting MIMO systems, which operates directly on rece… view at source ↗
Figure 2
Figure 2. Figure 2: Achievable rate versus transmission power under var [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Achievable rate versus number of RIS elements under d [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Achievable rate versus user distance to the STAR-RIS [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
read the original abstract

In this paper, a green learning (GL)-based precoding framework is proposed for simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS)-aided millimeter-wave (mmWave) MIMO broadcasting systems. Motivated by the growing emphasis on environmental sustainability in future 6G networks, this work adopts a broadcasting transmission architecture for scenarios where multiple users share identical information, improving spectral efficiency and reducing redundant transmissions and power consumption. Different from conventional optimization methods, such as block coordinate descent (BCD) that require perfect channel state information (CSI) and iterative computation, the proposed GL framework operates directly on received uplink pilot signals without explicit CSI estimation. Unlike deep learning (DL) approaches that require CSI-based labels for training, the proposed GL approach also avoids deep neural networks and backpropagation, leading to a more lightweight design. Although the proposed GL framework is trained with supervision generated by BCD under full CSI, inference is performed in a fully CSI-free manner. The proposed GL integrates subspace approximation with adjusted bias (Saab), relevant feature test (RFT)-based supervised feature selection, and eXtreme gradient boosting (XGBoost)-based decision learning to jointly predict the STAR-RIS coefficients and transmit precoder. Simulation results show that the proposed GL approach achieves competitive spectral efficiency compared to BCD and DL-based models, while reducing floating-point operations (FLOPs) by over four orders of magnitude. These advantages make the proposed GL approach highly suitable for real-time deployment in energy- and hardware-constrained broadcasting scenarios.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a green learning (GL) framework for precoding in STAR-RIS-aided mmWave MIMO broadcasting systems. It integrates Saab subspace approximation, RFT-based feature selection, and XGBoost to jointly predict STAR-RIS coefficients and transmit precoders directly from received uplink pilot signals, avoiding explicit CSI estimation and deep neural networks. The framework is trained using supervision from block coordinate descent (BCD) under perfect CSI but performs inference in a CSI-free manner. Simulations claim competitive spectral efficiency relative to BCD and DL baselines while reducing FLOPs by over four orders of magnitude, positioning the approach for energy-efficient real-time deployment in broadcasting scenarios.

Significance. If the performance and generalization claims hold, the work offers a lightweight, interpretable alternative to both iterative optimization and DNN-based methods for 6G wireless systems, with clear practical value in reducing computational overhead and power consumption for hardware-constrained broadcasting. The explicit avoidance of backpropagation and the reported FLOPs reduction are concrete strengths that could support reproducible low-complexity implementations.

major comments (2)
  1. [Abstract / Proposed GL Framework] Abstract and methodology description: The central claim that the GL model delivers competitive spectral efficiency via CSI-free inference rests on training labels generated by BCD under perfect CSI. No analysis is provided of how the Saab+RFT+XGBoost pipeline handles the distribution shift arising from noisy uplink pilots, pilot contamination, or channel estimation errors at test time; this directly affects whether the reported SE gap to BCD persists under realistic conditions.
  2. [Simulation Results] Simulation results: The abstract states competitive spectral efficiency but supplies no error bars, confidence intervals, dataset split details, or explicit verification that post-training generalization holds when inference inputs are limited to received pilots rather than perfect CSI. This omission makes it difficult to assess the reliability of the cross-method comparison.
minor comments (2)
  1. [System Model] Notation for the uplink pilot model and the exact mapping from received signals to input features for the RFT/XGBoost stages could be clarified to improve reproducibility.
  2. [Simulation Results] Figure legends should explicitly state the pilot SNR, number of antennas, and user count used for the reported SE curves to allow direct comparison with the BCD baseline.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and detailed review of our manuscript. We address each major comment below with clarifications and indicate the revisions we will implement.

read point-by-point responses

  1. Referee: [Abstract / Proposed GL Framework] Abstract and methodology description: The central claim that the GL model delivers competitive spectral efficiency via CSI-free inference rests on training labels generated by BCD under perfect CSI. No analysis is provided of how the Saab+RFT+XGBoost pipeline handles the distribution shift arising from noisy uplink pilots, pilot contamination, or channel estimation errors at test time; this directly affects whether the reported SE gap to BCD persists under realistic conditions.

    Authors: We thank the referee for this important observation. The GL framework is designed to take received uplink pilot signals directly as input; these signals inherently contain noise and other real-world impairments. Supervision is generated by BCD under perfect CSI to provide accurate target labels for the joint prediction of STAR-RIS coefficients and transmit precoders. While the current simulations employ realistic mmWave channel models that include noise, we acknowledge that an explicit robustness analysis against pilot contamination or varying noise levels is not present. In the revised manuscript we will add new simulation results evaluating spectral efficiency across a range of pilot SNRs and with imperfect CSI, thereby demonstrating whether the performance gap to BCD holds under realistic conditions. revision: yes

  2. Referee: [Simulation Results] Simulation results: The abstract states competitive spectral efficiency but supplies no error bars, confidence intervals, dataset split details, or explicit verification that post-training generalization holds when inference inputs are limited to received pilots rather than perfect CSI. This omission makes it difficult to assess the reliability of the cross-method comparison.

    Authors: We agree that additional statistical and experimental details would strengthen the presentation. In the revised manuscript we will include error bars (or confidence intervals) on all spectral-efficiency curves, obtained from multiple independent channel realizations. We will also document the train/validation/test split ratios and add an explicit statement confirming that inference operates exclusively on received pilot signals without access to perfect CSI. These changes will improve the transparency and reliability of the reported comparisons. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the GL framework derivation

full rationale

The paper describes a supervised green learning pipeline (Saab subspace approximation, RFT feature selection, and XGBoost regression) that maps received uplink pilot signals directly to STAR-RIS coefficients and transmit precoders. Training labels are generated offline by BCD under perfect CSI, but the inference stage uses only the pilot observations without explicit CSI or iteration. This is a standard supervised approximation setup rather than a self-definitional or fitted-input reduction; the learned mapping is not forced to reproduce its training inputs by construction. No load-bearing self-citations, uniqueness theorems, or ansatz smuggling appear in the provided description, and the central claim of CSI-free operation at inference rests on empirical generalization rather than definitional equivalence.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The framework depends on standard machine-learning components whose internal hyperparameters act as free parameters; the broadcasting architecture is treated as a domain assumption rather than derived.

free parameters (2)
  • XGBoost hyperparameters
    Tuning parameters for the decision learning stage that affect final predictions.
  • RFT relevance thresholds
    Thresholds used to select features from pilot signals.
axioms (1)
  • domain assumption Broadcasting scenario in which multiple users receive identical information
    Invoked to justify spectral-efficiency gains and reduced redundant transmissions.

pith-pipeline@v0.9.0 · 5835 in / 1305 out tokens · 44480 ms · 2026-05-18T18:00:22.525093+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

  1. [1]

    Next generation advanced transceiver tech nologies for 6G and beyond,

    C. Y ou, Y . Cai, Y . Liu, M. Di Renzo, T. M. Duman, A. Y ener, and A. Lee Swindlehurst, “Next generation advanced transceiver tech nologies for 6G and beyond,” IEEE J. Sel. Areas Commun. , vol. 43, no. 3, pp. 582–627, Mar. 2025

    work page 2025

  2. [2]

    Deep reinforcement learning-based precoding for m ulti-RIS- aided multiuser downlink systems with practical phase shif t,

    P .-H. Chou, B.-R. Zheng, W.-J. Huang, W. Saad, Y . Tsao, an d R. Y . Chang, “Deep reinforcement learning-based precoding for m ulti-RIS- aided multiuser downlink systems with practical phase shif t,” IEEE Wirel. Commun. Lett. , vol. 14, no. 1, pp. 23–27, Jan. 2025

    work page 2025

  3. [3]

    DNN-based precoding in RIS-aided mmWave MIMO syste ms with practical phase shift,

    P .-H. Chou, C.-W. Chen, W.-J. Huang, W. Saad, Y . Tsao, and R. Y . Chang, “DNN-based precoding in RIS-aided mmWave MIMO syste ms with practical phase shift,” in Proc. IEEE Global Commun. Conf. W orkshops (GLOBECOM Wkshps), Dec. 2024

    work page 2024

  4. [4]

    Capacity-net-based RIS precoding design without channel estimation for mmWave MIMO system,

    C.-Y . Huang, P .-H. Chou, W.-J. Huang, Y .-R. Chien, and Y . Tsao, “Capacity-net-based RIS precoding design without channel estimation for mmWave MIMO system,” in Proc. IEEE Int. Symp. Pers., Indoor Mobile Radio Commun. (PIMRC) , Sept. 2024

    work page 2024

  5. [5]

    Simultaneou sly transmitting and reflecting (STAR) RIS aided wireless commu nications,

    X. Mu, Y . Liu, L. Guo, J. Lin, and R. Schober, “Simultaneou sly transmitting and reflecting (STAR) RIS aided wireless commu nications,” IEEE Trans. Wireless Commun. , vol. 21, no. 5, pp. 3083–3098, May 2022

    work page 2022

  6. [6]

    Mul ticast and broadcast enablers for high-performing cellular V2X sy stems,

    M. Fallgren, B. Timus, M. Dillinger, M. F¨ arber, A. Hoglu nd, A. Kousari- das, C. Mannweiler, H. Nguyen, S. Pizzi, D. Sjoberg, D. Wieru ch, M. Boban, J. Alonso-Zarate, P . Popovski, and A. U. Quddus, “Mul ticast and broadcast enablers for high-performing cellular V2X sy stems,” IEEE Trans. Broadcast., vol. 65, no. 2, pp. 454–463, June 2019

    work page 2019

  7. [7]

    Capac ity characterization for reconfigurable intelligent surfaces assisted multiple- antenna multicast,

    L. Du, S. Shao, G. Y ang, J. Ma, Q. Liang, and Y . Tang, “Capac ity characterization for reconfigurable intelligent surfaces assisted multiple- antenna multicast,” IEEE Trans. Wireless Commun. , vol. 20, no. 10, pp. 6940–6953, Oct. 2021

    work page 2021

  8. [8]

    Weight ed sum rate optimization for STAR-RIS-assisted MIMO system,

    H. Niu, Z. Chu, F. Zhou, P . Xiao, and N. Al-Dhahir, “Weight ed sum rate optimization for STAR-RIS-assisted MIMO system,” IEEE Trans. V eh. Technol., vol. 71, no. 2, pp. 2122–2127, Feb. 2022

    work page 2022

  9. [9]

    Hyb rid reinforcement learning for STAR-RISs: A coupled phase-shi ft model based beamformer,

    R. Zhong, Y . Liu, X. Mu, Y . Chen, X. Wang, and L. Hanzo, “Hyb rid reinforcement learning for STAR-RISs: A coupled phase-shi ft model based beamformer,” IEEE J. Sel. Areas Commun. , vol. 40, no. 9, pp. 2556–2569, Sept. 2022

    work page 2022

  10. [10]

    Enhancing spectral efficiency of short-packet commu nications in STAR-RIS-assisted SWIPT MIMO-NOMA systems with deep learn ing,

    R. H. Y . Perdana, T.-V . Nguyen, Y . Pramitarini, D. H. N. N guyen, and B. An, “Enhancing spectral efficiency of short-packet commu nications in STAR-RIS-assisted SWIPT MIMO-NOMA systems with deep learn ing,” IEEE Trans. Wireless Commun. , vol. 24, no. 1, pp. 842–859, Jan. 2025

    work page 2025

  11. [11]

    STA R-RIS aided communication system with movable antennas exploiti ng statistical CSI,

    Y . Sun, H. Xu, B. Ning, L. Zhu, C. Ouyang, and H. Y ang, “STA R-RIS aided communication system with movable antennas exploiti ng statistical CSI,” IEEE Commun. Lett. , vol. 29, no. 4, pp. 824–828, Apr. 2025

    work page 2025

  12. [12]

    Green learning: Introduct ion, examples and outlook,

    C.-C. J. Kuo and A. M. Madni, “Green learning: Introduct ion, examples and outlook,” J. Vis. Commun. Image Represent. , vol. 90, p. 102187, Feb. 2023

    work page 2023

  13. [13]

    Inter pretable convolutional neural networks via feedforward design,

    C.-C. J. Kuo, M. Zhang, S. Li, J. Duan, and Y . Chen, “Inter pretable convolutional neural networks via feedforward design,” J. Vis. Commun. Image Represent., vol. 60, pp. 346–359, Jan. 2019

    work page 2019

  14. [14]

    On supervised f eature selection from high dimensional feature spaces,

    Y . Y ang, W. Wang, H. Fu, and C.-C. J. Kuo, “On supervised f eature selection from high dimensional feature spaces,” APSIPA Trans. Signal Inf. Process., vol. 11, no. 1, p. e31, Oct. 2022

    work page 2022

  15. [15]

    Green-learn ing based design of RIS-assisted MIMO systems based on implicit CSI,

    T.-C. Liao, W.-J. Huang, and C.-C. J. Kuo, “Green-learn ing based design of RIS-assisted MIMO systems based on implicit CSI,” in Proc. IEEE Int. Symp. Pers., Indoor Mobile Radio Commun. (PIMRC) , Sept. 2024

    work page 2024

  16. [16]

    XGBoost: A scalable tree boost ing system,

    T. Chen and C. Guestrin, “XGBoost: A scalable tree boost ing system,” in Proc. 22nd ACM SIGKDD Int. Conf. Knowledge Discovery and Dat a Mining (KDD) , Aug. 2016, pp. 785–794

    work page 2016

  17. [17]

    Learning to reflect and t o beamform for intelligent reflecting surface with implicit channel es timation,

    T. Jiang, H. V . Cheng, and W. Y u, “Learning to reflect and t o beamform for intelligent reflecting surface with implicit channel es timation,” IEEE J. Sel. Areas Commun. , vol. 39, no. 7, pp. 1931–1945, May 2021

    work page 1931

  18. [18]

    A unified converg ence analysis of block successive minimization methods for nons mooth opti- mization,

    M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified converg ence analysis of block successive minimization methods for nons mooth opti- mization,” SIAM J. Optim. , vol. 23, no. 2, pp. 1126–1153, June 2013

    work page 2013

  19. [19]

    DFT-based beamformi ng weight- vector codebook design for spatially correlated channels i n the unitary precoding aided multiuser downlink,

    D. Y ang, L.-L. Y ang, and L. Hanzo, “DFT-based beamformi ng weight- vector codebook design for spatially correlated channels i n the unitary precoding aided multiuser downlink,” in Proc. IEEE Int. Conf. Commun. (ICC), May 2010

    work page 2010

  20. [20]

    Attention is all you need,

    A. V aswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jone s, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , vol. 30, Dec. 2017, pp. 5998–6008

    work page 2017