Chapter Summary

Chapter 25 Summary: AI/ML for Massive MIMO

Key Points

• 1.

  MMSE is Bayes-optimal; DL wins only where its assumptions break. Under a Gaussian channel prior with known spatial covariance $\boldsymbol{\Sigma}_{\ntn{ch}}$ and Gaussian noise, no neural network can beat the linear MMSE estimator. Learned channel estimators (ChannelNet, DnCNN denoisers) matter only where the assumptions fail: impulsive or non-Gaussian interference, unknown covariance, and spatial non-stationarity in XL-MIMO. The half of the 2018-2020 DL-for-estimation literature that claimed universal gains over MMSE did not survive a fair comparison.

• 2.

  CSI feedback is a rate-distortion problem with a hard-coded physical prior. CsiNet and Transformer-CSI approach the Gaussian $R(D)$ lower bound within 1-2 dB on their training distribution while 5G NR Type II gives up 2-4 dB to generalize across scenarios. The deployable architecture under 3GPP Release 18 is CsiNet-on-top-of-Type-II: learned encoder with a classical fallback. Per-scenario retraining is the real cost, and is what prevents pure data-driven codecs from shipping alone.

• 3.

  Beam prediction is the one physical-layer task where sequence learning decisively wins. An LSTM or Transformer trained on beam- index + RSRP histories cuts mmWave beam-search overhead by 8-16x with top-5 accuracy above 95 % on vehicular traces. 3GPP Rel-18 beam management is the leading AI/ML candidate for actual standardization, because the temporal structure of mobile channels is too rich for rule-based tracking and too non-stationary for blind search.

• 4.

  RL for power control and scheduling is an augmentation, not a replacement. The convex and log-utility methods of Chapter 5 remain the production workhorse because they have rigorous guarantees and zero retraining cost. PPO adds 1-5 % utility in narrow non-convex regimes (HARQ, nonlinear amplifiers, cross-layer latency) at the cost of GPU-hours and a painful simulation-to-reality gap. The honest verdict: use RL to tune classical algorithms, not to replace them.

• 5.

  Distribution shift is the central obstacle, not architecture choice. Every pure data-driven method in this chapter — CsiNet, learned channel estimators, LSTM beam predictors, PPO schedulers — loses 5-12 dB or $\\pm 30 \\%$ utility when the deployment differs from the training set. Model-based DL — deep-unfolded ISTA, LAMP, OAMP-Net — retains 80-90 % of its advantage under the same shift because the physics is hard-wired into the architecture rather than learned from scratch.

• 6.

  The CommIT / Huawei 6G workshop position is that model-based DL is the final answer. Deep unfolding of ISTA, AMP, and OAMP inherits classical convergence guarantees, needs only $O(K)$ trainable parameters per $K$ unrolled iterations, and generalizes across channel scenarios. Combined with physics-informed losses and a classical fallback, this architecture is the position the TU Berlin - Huawei 6G workshop is carrying into the 3GPP Release-19 standards conversation. The role of deep learning in the physical layer is to parameterize the iterative algorithms we already derived from information theory, not to replace them.