MIMO Detection via Gaussian Mixture Expectation Propagation: A Bayesian Machine Learning Approach for High-Order High-Dimensional MIMO Systems
Shachar Shayovitz, Doron Ezri, Yoav Levinbook
TL;DR
The paper tackles uplink MIMO detection in high-dimensional, high-order systems where optimal ML detection is intractable. It introduces Gaussian Mixture Messages within an Expectation Propagation framework (GMEP), replacing problematic Gaussian priors with a Gaussian Mixture to better approximate the discrete data prior. By selectively applying GMM approximations to negative-variance nodes and using MIL-based updates, GMEP achieves superior SER performance over EP with comparable or reduced complexity, especially when the mixture order is kept small. This Bayesian, soft-output approach offers a scalable path for detection in large MIMO systems and supports integration with channel coding for practical deployments.
Abstract
MIMO systems can simultaneously transmit multiple data streams within the same frequency band, thus exploiting the spatial dimension to enhance performance. MIMO detection poses considerable challenges due to the interference and noise introduced by the concurrent transmission of multiple streams. Efficient Uplink (UL) MIMO detection algorithms are crucial for decoding these signals accurately and ensuring robust communication. In this paper a MIMO detection algorithm is proposed which improves over the Expectation Propagation (EP) algorithm. The proposed algorithm is based on a Gaussian Mixture Model (GMM) approximation for Belief Propagation (BP) and EP messages. The GMM messages better approximate the data prior when EP fails to do so and thus improve detection. This algorithm outperforms state of the art detection algorithms while maintaining low computational complexity.