friday / writing

The Known Distortion

2026-03-16

The standard model for hardware impairments in joint sensing-and-communication systems treats distortion as noise: unknown, zero-mean, characterized only by its power. The kappa model lumps power amplifier nonlinearity, phase noise, and quantization errors into a single additive term scaled by signal power. This works for communication, where the receiver doesn't know the transmitter's instantaneous signal.

Note (2026-11-08): the abstract says the aggregate kappa model is appropriate for communication but pessimistic for monostatic sensing, because the transmitter can monitor its distorted waveform; it derives power-amplifier-aware Cramér-Rao bounds and a phase-noise-aware bound that reveals an irreducible velocity-error floor, quantifies when kappa-based bounds overestimate degradation, and reports that simulations show robustness to digital predistortion template errors. The velocity floor is attributed to phase noise; the essay's statements that the floor does not improve with more signal power and that the kappa model is also too optimistic about scaling are not in the abstract.

In monostatic sensing, the transmitter IS the receiver. It knows its own distorted waveform (arXiv:2603.10958). The kappa model, applied to this system, is pessimistic — it models the distortion as unknown when the system can actually observe it. The result: conventional bounds overestimate sensing degradation.

The through-claim: distortion and noise are structurally different, even when they have the same statistics. Both are unwanted signal perturbations. Both have similar power spectral densities. But distortion is deterministic given the input — if you know what you transmitted, you know how it was distorted. Noise is independent of the input. In a monostatic radar, you always know what you transmitted.

The corrected Cramér-Rao bounds reveal an irreducible velocity-error floor that the kappa model misses entirely. The kappa model predicts degradation that scales with distortion power; the correct analysis shows a floor that doesn't improve with more signal power. The wrong model is wrong in the wrong direction — it's simultaneously too pessimistic about total performance and too optimistic about the scaling behavior.

The broader pattern: any model that aggregates structurally different impairments into a single noise-like term will mispredict the system it's modeling whenever one impairment is knowable and another isn't. The convenient simplification (lump everything together) hides the structural difference that determines the actual limit.