How Far May a UE Encoder Deviate? Exact Interoperability Certificates for Standardized Two-Sided AI/ML CSI Compression

In the two-sided AI/ML channel state information (CSI) compression of 3GPP Release 20, a user-equipment (UE) encoder from one vendor must work with a network decoder from another. We certify how far a UE encoder may deviate from the reference before the paired squared generalized cosine similarity (SGCS) falls below a target. The scalar quantizer maps an ℓ∞ mismatch ball onto a product of level sets, so the worst-case SGCS over all encoders within mismatch ε is computed exactly by event-ordered enumeration. The problem is NP-hard for general decoders, the enumeration is query-optimal among black-box methods, and its cost, 2^N(ε) for N(ε) coordinates within ε of a quantization boundary, is known before the decoder is evaluated. The certificates yield class-uniform percentile bounds and tight decoder-free contracts for dataset exchange; single-shot bound propagation, including CROWN, is vacuous once one index changes. On proxies of the RAN4 reference model, the median certified radius is 0.021 for a 0.02 SGCS loss, yet a mismatch of 0.01 costs more than 0.5 SGCS on up to 0.15% of channels, mostly through a few switch-like coordinates; random (greedy) testing underestimates such losses by up to 0.8 (0.6). fp16, bf16 and post-training int8 encoder implementations stay within 0.02 SGCS of the reference pair with guaranteed probabilities 0.996, 0.977 and 0.60 (95% confidence each), and fp16 storage is immaterial. Margin-aware training enlarges the median radius by about 50% and makes certification an order of magnitude cheaper. We derive recommendations for RAN4. This is a preprint and has not been peer reviewed. Code, trained models and raw results: https://doi.org/10.5281/zenodo.24294815 AI use: The manuscript was drafted and the code was generated with an AI system (Claude, Anthropic). The author ran the code on cloud servers to obtain all results.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24294854
Primary Topic
Formal Methods in Verification
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

How Far May a UE Encoder Deviate? Exact Interoperability Certificates for Standardized Two-Sided AI/ML CSI Compression

Yuan Shen
2 links
Zenodo (CERN European Organization for Nuclear Research)
Formal Methods in Verification
preprint

How Far May a UE Encoder Deviate? Exact Interoperability Certificates for Standardized Two-Sided AI/ML CSI Compression

Yuan Shen
preprint en

Abstract

In the two-sided AI/ML channel state information (CSI) compression of 3GPP Release 20, a user-equipment (UE) encoder from one vendor must work with a network decoder from another. We certify how far a UE encoder may deviate from the reference before the paired squared generalized cosine similarity (SGCS) falls below a target. The scalar quantizer maps an ℓ∞ mismatch ball onto a product of level sets, so the worst-case SGCS over all encoders within mismatch ε is computed exactly by event-ordered enumeration. The problem is NP-hard for general decoders, the enumeration is query-optimal among black-box methods, and its cost, 2^N(ε) for N(ε) coordinates within ε of a quantization boundary, is known before the decoder is evaluated. The certificates yield class-uniform percentile bounds and tight decoder-free contracts for dataset exchange; single-shot bound propagation, including CROWN, is vacuous once one index changes. On proxies of the RAN4 reference model, the median certified radius is 0.021 for a 0.02 SGCS loss, yet a mismatch of 0.01 costs more than 0.5 SGCS on up to 0.15% of channels, mostly through a few switch-like coordinates; random (greedy) testing underestimates such losses by up to 0.8 (0.6). fp16, bf16 and post-training int8 encoder implementations stay within 0.02 SGCS of the reference pair with guaranteed probabilities 0.996, 0.977 and 0.60 (95% confidence each), and fp16 storage is immaterial. Margin-aware training enlarges the median radius by about 50% and makes certification an order of magnitude cheaper. We derive recommendations for RAN4. This is a preprint and has not been peer reviewed. Code, trained models and raw results: https://doi.org/10.5281/zenodo.24294815 AI use: The manuscript was drafted and the code was generated with an AI system (Claude, Anthropic). The author ran the code on cloud servers to obtain all results.

Zenodo (CERN European Organization for Nuclear Research)
University of British Columbia, Okanagan Campus (CA)
Formal Methods in Verification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon