Dimensionality reduction method for gigantic MIMO channel matrices

Abstract This paper presents a novel and computationally efficient approach for dimensionality reduction in channel matrices in gigantic MIMO (GMIMO) systems, designed to address the extreme complexity of large-scale antenna arrays in future 6 G wireless networks. A high-fidelity unreal engine-based deterministic ray-tracing simulator is used to generate synthetic channel matrices under both line-of-sight (LoS) and non-line-of-sight (NLoS) conditions in realistic indoor and outdoor environments. The proposed method integrates principal component analysis (PCA) with k-means++ clustering to compress channel matrices by capturing dominant spatial features and grouping antenna elements with similar propagation characteristics. These hybrid techniques enable high compression ratios while preserving critical channel properties. To evaluate its effectiveness, the method is benchmarked against a baseline approach that combines singular value decomposition (SVD) with the same k-means++ clustering framework. The results demonstrate that the PCA-based method consistently outperforms SVD-based reduction in complex NLoS scenarios, achieving a matrix size reduction of up to 95% with significantly better retention of spatial and phase fidelity. In LoS environments, both methods perform comparably in terms of compression, although PCA maintains slightly higher accuracy under stronger dimensionality constraints. A clear trade-off is observed between achievable MIMO capacity and compression efficiency, whereby higher retained channel energy leads to increased capacity but requires a larger number of spatial components to be preserved. Additionally, the impact of the clustering factor on error convergence and spatial fidelity has been analysed, as determined by the number of clusters generated by the k-means algorithm. This dimensionality reduction strategy offers a scalable, interpretable, and lightweight solution for high-resolution 6G simulation platforms and Digital Twin environments, enabling real-time processing and efficient system-level evaluations of ultra-dense antenna systems.