Spectral Graph Convolution Network-Aided Cross-Domain Expectation Propagation Detection for MIMO-OTFS Systems

Orthogonal time frequency space (OTFS) modulation is well suited to high-mobility communications, and its combination with multiple-input multiple-output (MIMO) transmission further improves spectral efficiency and link reliability. In practical MIMO-OTFS systems, the limited delay–Doppler resolution leads to doubly fractional channels and dense coupling among transmitted symbols, posing significant challenges for reliable signal detection. Expectation propagation (EP) can achieve accurate symbol detection in such channels but suffers from high computational complexity. This motivates cross-domain EP (Cross-EP), which performs linear Gaussian inference in the time domain and discrete symbol inference in the delay–Doppler domain. However, the diagonal covariance approximation in Cross-EP neglects symbol correlations and degrades cavity distribution estimation. In this paper, we propose a spectral graph convolution network-aided Cross-EP (SGCN Cross-EP) detector, which models the equivalent MIMO-OTFS channel as a graph and exploits graph features to refine cavity statistics during iterative detection. Simulation results show that SGCN Cross-EP consistently outperforms Cross-EP and learning-based detectors while approaching conventional EP under perfect channel state information (CSI). It achieves approximately 3 dB and up to 6 dB signal-to-noise ratio gains over Cross-EP at a bit error rate of 10−4 under perfect and imperfect CSI, respectively, demonstrating improved detection accuracy and robustness.