The Price of Geometry: Exact One-Shot Limits and Near-Optimal Codes for CSI Feedback in the AI Era
AI-based compression promises to cut the overhead of channel state information feedback in FDD massive MIMO, but the limit of any code, learned or not, is unknown. We answer this for one-shot feedback under the 3GPP subband squared generalized cosine similarity (SGCS). (i) With known geometry, optimal B-bit feedback of a wideband channel with r resolvable paths reduces exactly to Grassmannian quantization of an angular-central-Gaussian direction on CP^(r-1), for any number of subbands; a certified converse lies within about one bit of Lloyd codebooks. (ii) With unknown geometry, every code pays a geometry price, at high resolution exactly log2 of the number of unordered angle-delay patterns. (iii) A non-AI two-part code built on this structure needs 1.8-3.5x less payload than eType II at equal SGCS, at ranks 1 and 2. (iv) Autoencoders approach a covariance-aided limit on fixed-geometry CDL-C but stay 0.06-0.08 below the two-part code on random-drop UMa. (v) For a single-antenna UE, sending the geometry once per 50 reports cuts the payload by another 1.8-2.3x (3.6-4.7x below eType II); given this geometry, learned coders add at most 0.002 SGCS. AI gains on fixed-geometry datasets thus largely reflect learned geometry; the decisive lever is how geometry is conveyed. Code and results: https://doi.org/10.5281/zenodo.24294470 The simulation code was generated and run, and the article (including its analysis) was drafted, with Claude (Anthropic), a generative AI system. The author provided the computing servers, reviewed all content and takes full responsibility for it (see the Acknowledgment of the preprint).
Authors
- Yuan Shen
Institutions
- University of British Columbia, Okanagan Campus (CA)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24294597
- Primary Topic
- Advanced MIMO Systems Optimization
- Type
- preprint