Deep Unfolding-Based SCA for ISAC Non-Convex Precoding Design: Learning Adaptive Schedules

Abstract

Real-time Integrated Sensing and Communication (ISAC) precoding requires solving a non-convex sensing–communication design under strict latency constraints. We propose Deep Unfolded SCA (DU-SCA), a fixed-depth unfolding of nested proximal Successive Convex Approximation (SCA) that replaces online Armijo searches and static schedules with learned layer-wise updates while preserving the optimization structure. Against tuned SCA, an Semidefinite Relaxation (SDR)-based reference, and a plain Deep Neural Network (DNN), DU-SCA matches nominal sensing quality, is markedly more reliable under tight computational budgets and overloaded regimes, and provides lower measured online latency. A paired equal-power waveform experiment with an Maximum Likelihood Estimator (MLE)/Generalized Likelihood Ratio Test (GLRT) receiver and a closed-form Zero-Forcing Isotropic Null-Space (ZF-ISO) reference confirms an estimation benefit from covariance optimization, while finding no statistically resolved detection difference. These results position DU-SCA as a practical structured solution for feasible low-latency ISAC precoding.