Finite-Bit CSI Compression Under Feedback Constraints: Exact PCA Error Decomposition and Robust Dimension–Bit Allocation for MISO-OFDM
- 1School of Electrical and Telecommunications Engineering, University of New South Wales, Sydney, NSW 2052, Australia
- 2Department of Computer Science, Durham University, Stockton Road, Durham DH1 3LE, UK
Abstract
Finite-bit channel-state-information (CSI) feedback couples the number of retained coefficients with their encoding precision. This study combines representation-specific payload accounting, an exact orthogonal principal component analysis (PCA) error identity, and bounds on beamforming alignment and rate loss in a controlled single-user multiple-input single-output orthogonal frequency-division multiplexing system. Exact enumeration compares uniform and multi-length PCA, finite-bit singular value decomposition (SVD), direct quantization, three dense autoencoders and a CsiNet-style convolutional reference. The primary evaluation comprises 7150 run records across five scenarios, with inference based on five independent test datasets and neural training seeds averaged within each dataset. On nominal Test-A, the 256-bit multi-length PCA design attains −20.411 dB physical normalized mean square error and 7.272 bit/s/Hz, versus −14.729 dB and 7.239 bit/s/Hz for uncompressed estimated CSI. This comparison includes estimation-noise suppression by projection; the multi-length design also changes coefficient clipping ranges. The empirical minimax selection first switches to rank-2 6-bit SVD at 1164 bits. These results concern the evaluated candidates and ideal representation payloads, not universal optimality or deployment-ready feedback costs.
Related
- A JoLT for the KV cache: Near-Lossless KV Cache Compression via Joint Rank-bit Allocation
- DiffLyap-MEC: Diffusion-Guided Lyapunov Optimization for Joint Offloading, Semantic Compression, and Resource Allocation in ISAC-Enabled 6G Edge Computing
- Mixed Precision Compression of Tucker Decomposition
- Adaptive Local Proper Orthogonal Decomposition via Graph Partitioning and Provable Error-Bounded Lossy Compression for Parametric Series in Finite-Dimensional Function Spaces
- Existence of Kraus decomposition in infinite dimension via strongly-convergent direct process tomography
BibTeX
@article{wang2026finitebit,
title = {{Finite-Bit CSI Compression Under Feedback Constraints: Exact PCA Error Decomposition and Robust Dimension–Bit Allocation for MISO-OFDM}},
author = {Tengyu Wang and Qian Wang},
journal = {Mathematics},
year = {2026},
month = oct,
doi = {10.3390/math14193606},
url = {https://www.mdpi.com/2227-7390/14/19/3606},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Finite-Bit CSI Compression Under Feedback Constraints: Exact PCA Error Decomposition and Robust Dimension–Bit Allocation for MISO-OFDM
AU - Tengyu Wang
AU - Qian Wang
PY - 2026
DA - 2026/10/04
JO - Mathematics
DO - 10.3390/math14193606
UR - https://www.mdpi.com/2227-7390/14/19/3606
AB - Finite-bit channel-state-information (CSI) feedback couples the number of retained coefficients with their encoding precision. This study combines representation-specific payload accounting, an exact orthogonal principal component analysis (PCA) error identity, and bounds on beamforming alignment and rate loss in a controlled single-user multiple-input single-output orthogonal frequency-division multiplexing system. Exact enumeration compares uniform and multi-length PCA, finite-bit singular value decomposition (SVD), direct quantization, three dense autoencoders and a CsiNet-style convolutional reference. The primary evaluation comprises 7150 run records across five scenarios, with inference based on five independent test datasets and neural training seeds averaged within each dataset. On nominal Test-A, the 256-bit multi-length PCA design attains −20.411 dB physical normalized mean square error and 7.272 bit/s/Hz, versus −14.729 dB and 7.239 bit/s/Hz for uncompressed estimated CSI. This comparison includes estimation-noise suppression by projection; the multi-length design also changes coefficient clipping ranges. The empirical minimax selection first switches to rank-2 6-bit SVD at 1164 bits. These results concern the evaluated candidates and ideal representation payloads, not universal optimality or deployment-ready feedback costs.
SN - 2227-7390
ER -