Resource-Element Energy Difference for Noncoherent Over-the-Air Federated Learning

Researchers propose REED (Resource-Element Energy Difference), a noncoherent aggregation method for over-the-air federated learning that eliminates the need for instantaneous channel state information. The technique uses energy differences across orthogonal resource elements to aggregate signed updates, achieving convergence rates comparable to conventional methods while reducing practical implementation complexity in wireless systems.

This research addresses a fundamental challenge in distributed machine learning over wireless networks: how to aggregate model updates efficiently without requiring perfect channel knowledge. Over-the-air federated learning traditionally relies on coherent phase alignment and instantaneous CSI—requirements that introduce significant overhead and complexity in real-world deployments. REED shifts the paradigm by mapping positive and negative update components to separate orthogonal resource elements with independent phase dithering, then recovering the signed aggregate from their energy difference. This approach requires only slow-timescale calibration rather than instantaneous channel feedback, substantially simplifying practical implementation. The theoretical contribution is substantial: the authors prove that REED remains unbiased for the desired aggregate and derive closed-form variance expressions under Rayleigh fading conditions. When integrated into FedAvg with full client participation, they demonstrate convergence to stationarity at the canonical (1/√T) rate under reasonable energy budgets. Experimental validation on MNIST and Fashion-MNIST shows REED performs nearly identically to clean FedAvg in IID data settings and matches coherent CSIT aggregation performance, while maintaining stable convergence even under significant data heterogeneity with only moderate degradation. This development has implications for edge computing and distributed AI training in bandwidth-constrained or rapidly-changing wireless environments. The elimination of instantaneous CSI requirements could enable faster deployment of federated learning in mobile and IoT networks where channel estimation overhead is prohibitive. Future research may explore how REED scales to larger networks and more complex learning tasks.

• →REED eliminates instantaneous CSI requirements by using energy differences on orthogonal resource elements with independent phase dithering
• →Theoretical analysis proves REED achieves canonical (1/√T) convergence rate under per-client energy budgets
• →Experimental results show REED matches coherent aggregation performance in IID settings with only moderate degradation under data heterogeneity
• →Slow-timescale calibration of average channel powers replaces complex instantaneous channel feedback mechanisms
• →The method significantly reduces practical implementation complexity for wireless federated learning systems