Semantic CSI Feedback for Beam Selection: When Task-Aware Embeddings from Sparse Pilots Outperform Full-Bandwidth Reconstruction

1. INTRODUCTION

• Classical feedback compresses CSI for faithful reconstruction before beam selection, leaving its encoder agnostic to the downstream task.The paper frames this two-stage separation as suboptimal because reconstruction can preserve signal-level fidelity that is not directly relevant to beam selection.Source: §1 · p. 1
• Semantic CSI feedback encodes pilot observations into a compact embedding trained end-to-end to represent optimal beam identity rather than channel coefficients.A decoder grounds the embedding in physical channel structure during training but is discarded at inference.Source: §1 · p. 1
• A task-aware embedding of d = 8 reals from 43 raw pilots surpasses every full-bandwidth method, while interpolation actively hurts.The evaluation also finds that the optimal input domain depends on observation density: SF wins with full bandwidth, while AD wins with sparse pilots.Source: §1 · p. 1

2.1. Downlink Channel and Beam Codebook

• The 8 × 8 UPA has 64 antennas and serves single-antenna UEs across 512 OFDM subcarriers at 3.5 GHz.The beam codebook uses an oversampled 2D DFT construction with 256 candidate beams; 88 active classes remain after removing beams never optimal for any UE.Source: §2.1 · pp. 1–2

2.2. What the UE Observes

• The UE observes noisy downlink channel measurements, with noise variance set according to the per-sample SNR.The three scenarios are full CSI on all 512 subcarriers, 43 raw pilots spaced every 12 subcarriers, or linear interpolation from those pilots to all 512 subcarriers.Source: §2.2 · p. 2

3. THE ROLE OF INPUT DOMAIN

• The representation choice matters before learning because it determines what information survives sparsification.The angular-delay transform can recover a meaningful representation from 43 pilots because the channel has fewer degrees of freedom than subcarriers.Source: §3 · p. 2
• Spatial-frequency inputs stack the channel’s real and imaginary parts, but 43 pilots leave 91.6% of the frequency axis empty.With full CSI, the representation includes all 512 subcarriers.Source: §3 · p. 2
• Angular-delay inputs apply an antenna FFT and frequency IFFT, producing 64 delay taps with full CSI or 43 delay samples from raw pilots.Multipath energy concentrates in the first few taps, and the transform acts as a parameter-free, physics-informed preprocessing step that reduces the downstream network’s burden.Source: §3 · p. 2

4.3. Task-Aware: Semantic Feedback

• Semantic feedback jointly trains the encoder, decoder, and interpreter, with beam prediction acting directly on the embedding and reconstruction providing physical-structure grounding.The semantic gradient shapes the embedding toward beam-discriminative structure, while decoder-based regularization is especially pronounced under pilot-only observations, where pure classification overfits.Source: §4 · p. 2, §4.3 · p. 3

4.1. Direct Classification (Upper Bound)

• Direct classification maps the full input to beam logits without compression, modeling an unconstrained feedback link or TDD reciprocity as an upper-bound benchmark.Source: §4.1 · p. 2

4.2. CsiNet-Style Autoencoder: Signal-Level Feedback

• CsiNet-style feedback trains an autoencoder for signal reconstruction, freezes it, then trains a classifier on the reconstructed channel rather than the embedding.The encoder therefore has no incentive to preserve beam-discriminative features orthogonal to the reconstruction objective.Source: §4.2 · p. 2

4.3. Task-Aware: Semantic Feedback

• The same compact semantic embedding can serve multiple gNB tasks, including beam selection and rank adaptation, without task-specific UE re-encoding.Source: §4.3 · p. 3

4.4. Training Protocol

• The training protocol uses Adam at 10^-3, batch size 128, up to 50 epochs with patience-15 early stopping, and d = 8.Training SNR is uniform from −5 to 30 dB; data are split 70/15/15, with evaluation at fixed SNRs from −5 to 30 dB and at infinite SNR.Source: §4.4 · p. 3

5.1. The Domain Reversal

• 89.1% for AD+Semantic with raw pilots surpasses every SF method and the best full-CSI result; with full CSI, SF+Semantic leads at 83.5%.Table 1 reports top-1 beam accuracy with d=8; the domain preference reverses with observation density.Source: §5.1 · p. 3, Table 1 · p. 4

5.2. Why Interpolation Hurts

• Interpolation degrades accuracy across evaluated methods, as linear interpolation introduces physically incorrect frequency responses that can lead networks to rely on non-generalizing patterns.The semantic encoder extracts beam-relevant structure directly from sparse pilots more effectively than reconstruct-then-classify approaches.Source: §5.2 · p. 4

5.3. Signal-Level vs. Semantic Feedback

• At the same d=8 bottleneck, CsiNet achieves 2–6% accuracy while semantic feedback reaches 89%, showing the advantage of shaping the embedding for beam prediction rather than channel reconstruction.Fig. 3 compares top-1 accuracy at infinite SNR; CsiNet is near-random across scenarios.Source: Figure 3 · p. 4, §5.3 · p. 4

• The proposed recipe transforms pilots to the angular-delay domain, avoids interpolation, and trains an 8-real embedding end-to-end for beam prediction.The paper reports a 32-bit payload and 89% accuracy from one CSI-RS transmission, and identifies quantization-aware training and cross-scenario generalization as future work.Source: §6 · p. 4

• This numerical simulation study required no ethical approval.Source: Compliance with Ethical Standards · p. 5

• The authors are Ericsson Research employees, received no external funding, and report no other relevant financial or nonfinancial interests.Source: Conflicts of Interest · p. 5