References & Further Reading

References

1. C.-K. Wen, W.-T. Shih, S. Jin, Deep Learning for Massive MIMO CSI Feedback, 2018

   The original CsiNet paper. Introduces the encoder-decoder autoencoder with IFFT preprocessing that remains the reference for learned CSI feedback. Section 25.2 follows this architecture closely, and the 3GPP Rel-18 AI/ML study item uses CsiNet as the primary use case for CSI compression.

2. T. O'Shea, J. Hoydis, An Introduction to Deep Learning for the Physical Layer, 2017

   The foundational paper introducing the autoencoder paradigm for the physical layer. Historically important (defines the data-driven DL research direction) even though subsequent work has tempered the universality claims. The opening reference for the data-driven vs model-based debate in Section 25.5.

3. N. Samuel, T. Diamant, A. Wiesel, Learning to Detect, 2019

   DetNet: a deep-unfolded projected gradient descent for MIMO detection. One of the first model-based DL papers in wireless. A useful companion to the LAMP and OAMP-Net discussion of Section 25.5.

4. H. Ye, G. Y. Li, B.-H. Juang, Power of Deep Learning for Channel Estimation and Signal Detection in OFDM Systems, 2018

   Demonstrates that a DL receiver can beat MMSE-MF on an OFDM link under realistic impairments. Introduces the ChannelNet-style architecture used in Section 25.1. The bar-raising paper that started the data-driven channel estimation literature; later refined to show that the gains come from non-Gaussian impairments (Section 25.1).

5. H. Huang, Y. Song, J. Yang, G. Gui, F. Adachi, Deep Learning-Based Millimeter-Wave Massive MIMO for Hybrid Precoding, 2019

   Early paper on DL for mmWave beam prediction. Uses an MLP over RSRP history; the LSTM + Transformer generalizations came later (Alkhateeb group). Reference for the Markovian-approximation theorem in Section 25.3.

6. A. Alkhateeb, S. Alex, P. Varkey, Y. Li, Q. Qu, D. Tujkovic, Deep Learning Coordinated Beamforming for Highly-Mobile Millimeter Wave Systems, 2018

   The Alkhateeb beam prediction dataset and baseline methods — now the de facto benchmark for mmWave beam prediction research, reused in most subsequent papers. The LSTM training setup in Section 25.3 follows this reference.

7. 3GPP Technical Specification Group RAN, Study on Artificial Intelligence (AI) / Machine Learning (ML) for NR Air Interface (Release 18), 2024

   The Release-18 AI/ML study item document. Defines the three representative use cases (CSI compression, beam management, positioning) and summarizes the companies' performance simulations. Required reading for anyone deploying learned components into 3GPP NR. Section 25.2 references it for the CSI feedback standardization path; Section 25.3 references it for beam management.

8. A. Zappone, M. Di Renzo, M. Debbah, Wireless Networks Design in the Era of Deep Learning: Model-Based, AI-Based, or Both?, 2019

   The first widely-cited synthesis of the model-based vs data-driven debate. Develops the taxonomy used in Section 25.5 and argues for the hybrid model-based + classical fallback architecture that the CommIT / Huawei workshop later adopted as policy.

9. E. Bjornson, L. Sanguinetti, H. Wymeersch, J. Hoydis, T. L. Marzetta, Massive MIMO is a Reality - What is Next? Five Promising Research Directions for Antenna Arrays, 2019

   Survey by the three main schools of massive MIMO (Linkoping, Eurecom, MIT). Section 5 on 'machine learning' gives the researcher-level view of where DL is useful in massive MIMO. A useful counterpoint to the 6G@UT/Huawei workshop position: this paper is more optimistic about end-to-end autoencoders.

10. G. Caire, 6G Wireless Technologies: Advancing MIMO and AI Integration, 2024

    CommIT contribution. The 2024 edition of the TU Berlin / Huawei 6G workshop. Establishes the model-based DL as the recommended architecture for physical-layer AI/ML, with the three-pillar position statement (model-based DL default, physics-informed losses, hybrid deployment with safe fallback) that Section 25.5 develops. The definitional reference for the chapter.

11. Y. Yang, F. Gao, G. Y. Li, M. Jian, Deep Learning-Based Downlink Channel Prediction for FDD Massive MIMO System, 2019

    Early Transformer-based approach to CSI feedback and prediction. Beats CsiNet on NMSE at equal bit budget but at 3x the compute. Representative of the Transformer-CSI literature that Section 25.2 compares against.

12. V. Monga, Y. Li, Y. C. Eldar, Algorithm Unrolling: Interpretable, Efficient Deep Learning for Signal and Image Processing, 2021

    The definitive survey on algorithm unrolling / deep unfolding. Covers LISTA, LAMP, OAMP-Net, and DetNet in a unified framework. Foundation for Section 25.5 and the deep-unfolded ISTA algorithm. Required reading for anyone doing model-based DL in signal processing.

13. M. Borgerding, P. Schniter, S. Rangan, AMP-Inspired Deep Networks for Sparse Linear Inverse Problems, 2017

    LAMP: learned AMP. The canonical model-based DL algorithm for sparse recovery, directly applicable to sparse channel estimation in massive MIMO. Foundation for the LAMP / OAMP-Net discussion in Section 25.5.

14. J. Schulman, F. Wolski, P. Dhariwal, A. Radford, O. Klimov, Proximal Policy Optimization Algorithms, 2017. [Link]

    The PPO paper. Now the de facto standard deep RL algorithm for continuous control including wireless power control and scheduling. Section 25.4 follows this closely and references the monotonic-improvement argument.

15. R. S. Sutton, A. G. Barto, Reinforcement Learning: An Introduction, MIT Press, 2nd ed., 2018

    The standard RL textbook. Chapters 3 (MDPs), 6 (Temporal-Difference), and 13 (Policy Gradient) give the background needed for Section 25.4. The prerequisite if you have not seen RL before.

16. S. M. Kay, Fundamentals of Statistical Signal Processing, Volume I: Estimation Theory, Prentice Hall, 1993

    The standard graduate textbook for estimation theory. Chapter 11 on linear Bayesian estimation establishes the MMSE Bayes-optimality result used in Theorem 25.2. Reference for any derivation involving conditional Gaussian estimation.

17. T. M. Cover, J. A. Thomas, Elements of Information Theory, Wiley-Interscience, 2nd ed., 2006

    The standard graduate textbook for information theory. Chapter 10 on rate-distortion theory is the source for the Gaussian R(D) function used in Theorem 25.1. Referenced throughout the rate-distortion framing of Section 25.2.

18. 3GPP Technical Specification Group RAN, NR: Physical Layer Procedures for Data (Release 17), 2022

    The normative 3GPP specification for NR downlink data including the Type I and Type II codebooks. Section 5.2.2.2 is the definitive reference for Type II feedback structure. Any CSI feedback implementation that plans to be standards-compliant must match this document.

Further Reading

• Model-based deep learning for physical-layer signal processing

  V. Monga, Y. Li, Y. C. Eldar, 'Algorithm Unrolling,' IEEE SPM 2021, plus the Zappone-Di Renzo-Debbah 'Model-Based, AI-Based, or Both?' IEEE TComm 2019 position paper

  These two together are the definitive references for the model-based DL argument of Section 25.5. The Monga survey gives the technical content; the Zappone paper gives the community-level synthesis that led to the CommIT / Huawei workshop position.

• Deep RL for wireless with honest reporting

  P. Henderson et al., 'Deep Reinforcement Learning that Matters,' AAAI 2018; plus the recent surveys on RL for 5G/6G such as Y. Sun et al., IEEE CST 2022

  The Henderson paper explains the reproducibility crisis in deep RL (seed variance, hyperparameter sensitivity). The Sun survey covers the wireless application space. Reading them in sequence explains why the 1-5 % RL gains reported in wireless papers are often not reproducible and why Section 25.4 is so cautious about RL deployment.

• Learned AMP / OAMP for massive MIMO detection and estimation

  H. He, C.-K. Wen, S. Jin, G. Y. Li, 'Model-Driven Deep Learning for MIMO Detection,' IEEE TSP 2020

  Applies OAMP-Net to massive MIMO detection with strong empirical results under non-i.i.d. pilot matrices. Direct application of the deep-unfolding framework to a core massive MIMO problem, useful companion to Section 25.5.

• 3GPP Release-18 AI/ML study item and representative use cases

  3GPP TR 38.843 plus the companion contributions from Huawei, Nokia, Ericsson, and Samsung on CSI compression and beam management

  The standards-level view of what the industry actually plans to deploy. Reading TR 38.843 alongside the academic papers explains why the deployment choices are more conservative than the research papers suggest — and why model-based DL with classical fallback is winning the standardization battle.

• Open RAN RIC and xApp deployment for learned control

  O-RAN Alliance WG3 'Near-Real-Time RIC Architecture' specification, plus the OpenAirInterface xApp examples

  The realistic deployment vehicle for the RL power control of Section 25.4. Unlike pure simulation experiments, xApp deployments actually have to deal with telemetry latency, E2 interface latency, and the fallback semantics that Section 25.4 argues for.

• Physics-informed neural networks and physics-informed losses in wireless

  G. E. Karniadakis et al., 'Physics-Informed Machine Learning,' Nature Reviews Physics 2021

  The general framework for injecting physical constraints into neural network training. The 6G@UT / Huawei workshop's 'physics-informed losses' pillar for Section 25.5 is a wireless application of this framework; the Karniadakis survey is the cross-disciplinary source.