Pseudorandomness and Diffraction

What the math gives to ML

The paper gives a deterministic quadratic-phase sequence whose diffraction is purely absolutely continuous for every nonzero amplitude, irrational quadratic coefficient, and phase offsets. This is a concrete source of structured pseudorandomness: unlike learned or sampled positional signals, it is reproducible, parameter-free, and can suppress coherent spectral spikes while retaining an exact algebraic recurrence. The most direct neural-network transfer is a chirped positional encoding or RoPE variant, with finite-length periodograms and attention-distance statistics used to test whether its claimed disorder-like behavior improves extrapolation and reduces positional aliasing.