Abstract
The K-R Mathematical Framework is a universal signal processing decomposition principle that separates any estimation or equalization problem into two orthogonal components: a physics-based K step, which applies the best available analytical model — such as MMSE equalization, matched filtering, or Kalman prediction — and a data-driven R step, which corrects the structured residual that the physics model cannot eliminate. The fundamental insight is that residuals after any model-based estimator are not random noise but carry exploitable structure arising from nonlinearities, model mismatch, and unmodelled effects. The R step captures this structure using a nonlinear feature expansion — including quadratic cross-terms such as Re², Im², and Re·Im — and trains a two-layer neural network per observation slot from available pilot symbols via closed-form least squares, requiring no backpropagation and no offline dataset. The framework is governed by Theorem 1, which decomposes the MSE gain into two orthogonal terms: γ_Q, a provable quantisation-floor reduction proportional to SNR/(SNR+1), and γ_CE, an implicit channel-error correction term that accounts for the dominant observed gain. Applied to 5G millimetre-wave systems with 4-bit ADC, the NL-feat K-R receiver achieves +0.197 bps/Hz spectral efficiency gain over MMSE, +0.220 bps/Hz over OAMP-Net, and +1.387 bps/Hz over DetNet at SNR = 20 dB, with p < 0.001 across 5,000 Monte Carlo trials and full mobility robustness from 0 to 500 km/h. Beyond wireless communications, the K-R principle applies wherever physics models are approximate: LiFi optical systems (LED nonlinearity correction), radar signal processing (clutter residual tracking), medical imaging (MRI artefact removal), power systems (state estimation error correction), structural health monitoring (damage signature detection), and financial signal processing (alpha residual modelling). Across all domains, the K calculation employs linear methods — MMSE, ZF, Kalman filter, Wiener filter, or finite element models — while the R calculation employs nonlinear feature expansion combined with linear parameter estimation, positioning K-R uniquely at the boundary between classical signal processing and machine learning: nonlinear enough to capture complex distortion, linear enough to guarantee a closed-form solution every observation interval.