Sum-Path-Gain Maximization for IRS-Aided MIMO Communication System via Riemannian Gradient Descent Network

Intelligent reflecting surface (IRS) is a key technique for enhancing the performance of wireless communications. In this letter, we focus on the sum-path-gain maximization (SPGM) problem in an IRS-aided MIMO communication system, which is non-convex due to the constant modulus constraint. The existing works mainly include the relaxation method with relaxation error and the non-relaxation methods with high complexity. Different from the existing methods, we notice that constant modulus constraint can naturally satisfy the Riemannian manifold, and the deep learning method has strong non-convex learning ability. By exploiting these characteristics, the Riemannian gradient descent network (RGD-Net) is proposed. In the proposed method, we first project the non-convex SPGM problem to the Riemannian manifold. Then, the Riemannian gradient descent iterations are unfolded as the network layers. Finally, the step sizes of each layer are learned in unsupervised manner to ensure converged performance. Compared with the existing methods, the proposed method achieves higher spectral efficiency with lower computational cost.