Latency-Aware Downlink Mutual Information Maximization for RIS-Assisted Federated Learning

Federated learning (FL) has emerged as an important framework for distributed edge intelligence over wireless networks, where the wall-clock training efficiency is strongly affected by the communication latency for global model dissemination and aggregation. In reconfigurable intelligent surface (RIS)-assisted wireless systems, improving the downlink transmission quality can support more timely model delivery and, consequently, faster FL progress under a fixed training-time budget. In this paper, we study a multi-user RIS-assisted FL downlink and develop an FL-oriented communication model that relates the global-model dissemination latency to the achievable mutual information (MI). To account for the bottleneck nature of FL downlink transmission, we clarify the relationship between sum-MI maximization and worst-user latency through a rate-dispersion factor, under which sum MI serves as a tractable communication-side surrogate for improving wall-clock FL efficiency. Based on this model, we formulate a joint active and passive beamforming problem under the base station transmit-power constraint and the RIS unit-modulus constraint. To handle the resulting non-convex coupling, we develop an efficient alternating optimization (AO) framework that combines weighted minimum mean-square error based active beamforming and block coordinate descent based passive beamforming, with monotonic objective improvement under the available channel state information (CSI). The impact of imperfect CSI is also discussed to reflect practical RIS-assisted deployment. Simulation results show that the proposed design achieves noticeable MI gains over representative communication-learning benchmarks, which are empirically associated with reduced downlink dissemination latency, improved communication–computation efficiency, and faster FL progress on standard image classification tasks.