A Deep Learning-based Low-Complexity SVD Approximation for Large-Scale MIMO Systems

Abstract

The cubic complexity of singular value decomposition (SVD) has become a major bottleneck for real-time processing in large-scale multiple-input multiple-output (MIMO) systems. To address this issue, in this paper, we propose a deep learning-based low-complexity SVD approximation method tailored for such systems. The proposed method employs discrete Fourier transform (DFT) dictionary-based physical priors for subspace initialization. It uses noise-aware embeddings to improve adaptability across varying signal-to-noise ratio (SNR) conditions. Additionally, the method includes residual refinement and canonicalization to encourage orthogonality and uniqueness of the singular vectors. Evaluations are conducted on channel data generated according to the 3GPP TR 38.901 standard, covering both Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) scenarios. The results demonstrate that the proposed framework achieves near-ideal reconstruction accuracy for antenna dimensions up to 128 × 128, while reducing computational latency by an order of magnitude compared to numerical SVD.