Noncoherent Aggregation for Over-the-Air Federated Learning
A novel technique known as resource-element energy difference (REED) facilitates noncoherent aggregation in over-the-air federated learning (OTA-FL), eliminating the necessity for real-time channel state information (CSI) and coherent phase alignment. By assigning positive and negative components of real-valued updates to transmit energies across two orthogonal resource elements with independent phase dithers, the server can derive the signed aggregate based on their energy difference. REED only requires slow-timescale calibration of average channel powers, maintains an unbiased approach for the intended signed sum, and provides an exact closed-form variance under Rayleigh fading. This method is integrated into the full-participation FedAvg framework.
Key facts
- REED is a noncoherent aggregation primitive for continuous signed updates.
- It avoids instantaneous CSI, channel inversion, and coherent phase alignment.
- REED maps positive and negative parts to transmit energies on two orthogonal resource elements.
- Independent phase dithers are used on each resource element.
- The server estimates the signed aggregate from the energy difference.
- Only slow-timescale calibration of average channel powers is needed.
- REED is unbiased for the desired signed sum.
- It has an exact closed-form variance under Rayleigh fading.
- The method is incorporated into full-participation FedAvg.
Entities
—