Deconstructing a TrainableWave Network: An Ablation Study of Analog Front Ends for Multi-Antenna Direction Finding

Trainable wave networks process radio-frequency (RF) signals in analog form before digitization, and have been proposed to reduce the cost of converting every antenna signal in a large array. They combine four physical mechanisms: memory, coupling, resonance, and nonlinearity. We performed a systematic simulation-driven ablation study across these mechanisms on a representative wave network for multi-antenna Direction of Arrival (DOA) estimation, and paired each removal with a variant designed to fail if our explanation were correct. On a task with a coherent reflection, removing memory raised the normalized score from at most 0.518 to 0.873, since the order of array snapshots carries no information. Coupling and resonance could be removed without loss of accuracy once a trained linear stage combined the antenna signals, whereas removing the nonlinearity reduced the scores to 0.15 and 0.03. The front end that remains, a trained linear combination of the antennas followed by square-law detection averaged over 16 snapshots, achieves scores of 0.985 and 0.940 on the two main tasks, at least as high as every resonator variant. At matched accuracy, it converts 50% to 75% fewer bits than the strongest digital estimators we tested. Based on our findings, we propose it as an effective new baseline for wave networks in direction finding.