Sparse Prior-Guided Deep Learning for OTFS Channel Estimation

In this paper, we propose a deep residual shrinkage neural network (DRSNN) based on sparse prior for solving the orthogonal time frequency space (OTFS) channel estimation problem. Specifically, we formulate the problem as a denoising task and subsequently develop a residual learning-based denoiser to adeptly learn the residual noise. To harness the sparsity of the channel in the delay-Doppler (DD) domain, we insert proximal mapping as a new layer into the deep network, which directly and explicitly produces well-regularized outputs. In particular, the layer is implemented by a trainable soft shrinkage function with a threshold vector, where the thresholds can be adapted by a dedicated sub-network. Moreover, we derive a mathematical expression for the proposed network and conduct a thorough analysis exploiting Bayesian philosophy, which provides valuable insights into the interpretability of neural networks. Finally, simulation results demonstrate the superiority of the proposed method for sparse channel recovery in terms of both estimation performance and computational complexity.