We quantify distortion introduced by downsampling
and interpolation in a modular RF ensemble classifier. For the
spectral path, we compare power spectral densities (PSDs) at
varying FFT sizes against a 1024-bin baseline using KullbackLeibler (KL) divergence. For the temporal path, we study task
accuracy as a function of sequence length with zero-padding
and interpolation to 128 samples. Across six SNR bins from
−10 dB to 20 dB, we demonstrate critical thresholds: spectral
performance degrades sharply below 256 bins, while temporal
accuracy benefits monotonically from longer sequences up to
128 samples. Our reproducible measurement harness provides
policy guidance for latency-accuracy trade-offs in production RF
systems.
