Statistical Multiport-Network Modeling and Efficient Discrete Optimization of RIS

Abstract: This Letter fills the research gap on physics-consistent optimization for reconfigurable intelligent surfaces (RISs) with mutual coupling (MC) and 1-bit-tunable elements, a common hardware constraint in existing RIS prototypes. We compare a model-based method (temperature-annealed back-propagation) and model-agnostic methods (coordinate descent, genetic algorithm), and evaluate potential benefits of intelligently initializing these methods. To facilitate our evaluation, we introduce a technique for generating statistical ensembles of multiport-network model parameters, wherein a single hyper-parameter adjusts the MC strength. The technique is a generalization of Rayleigh fading to radio environments with deterministic programmability, and it accounts for passivity constraints as well as the coherent-backscattering effect. We find that, except when MC is negligible, coordinate descent with random initialization yields the most favorable trade-off in terms of performance, execution time and memory usage. We expect our findings to extend to beyond-diagonal RIS, stacked intelligent metasurfaces, dynamic metasurface antennas, and wave-domain physical neural networks.