A Distributed mMIMO Architecture for Over-the-Air Federated Learning

This paper compares distributed massive Multiple-Input Multiple-Output (mMIMO) and centralized MIMO in the context of over-the-air federated learning, taking into account a detailed wireless channel model that includes spatially uncorrelated Rician fading to represent the Line of Sight (LoS) component. Our simulation results provide valuable insight into how a distributed architecture can maintain a lower Mean Squared Error (MSE) in scenarios that require extensive coverage. Furthermore, our research highlights the importance of considering the impact of the LoS component and power limitations when analyzing over-the-air aggregation, as these characteristics can substantially influence aggregation performance. Adopting a distributed architecture allows devices to conserve energy while simultaneously achieving a lower MSE in the aggregation process compared to the centralized architecture. Specifically, when considering four Access Points (APs) and 200 devices, the average transmit power is reduced by approximately 31% while achieving 16% lower aggregation MSE compared to a centralized approach.