Complex-Valued Differential Autoencoder for Temporal CSI Compression

Manoj Kumar CM, Lakshmi Narasimhan Theagarajan

Keywords: CSI compression, CSI feedback, Complex Valued Neural Networks, Autoencoder

Abstract:

This paper presents a complex-valued differential autoencoder architecture for efficient compression and reconstruction of channel state information (CSI) in cellular systems.
This minimizes the substantial feedback overhead associated with high-dimensional CSI in massive multi-antenna multiple-input multiple-output (MIMO) and broadband orthogonal frequency division modulation (OFDM).
The CSI is represented in the delay–angle domain, where it exhibits sparse and structured correlations across both spatial and temporal dimensions.
Unlike conventional feedback schemes that periodically transmit full CSI instances from the user equipment (UE) to the base station (BS), the proposed model leverages temporal correlation between consecutive CSI frames to reduce feedback bandwidth.
Instead of traditional models that split the real and imaginary values of the CSI into separate channels, the proposed architecture introduces a feedback driven mechanism that encodes and reconstructs only the information difference between CSI instances using complex-valued transformations.
Experiments are conducted on temporally correlated delay–angle CSI sequences, generated from standard 3GPP channel models at user speeds of 40 km/h, 100 km/h, and 360 km/h to emulate varying Doppler conditions.
The experimental results demonstrate that the proposed model achieves substantially higher compression efficiency and reconstruction fidelity than conventional frame-wise autoencoders.

Submission Number: 25