Nonlinear-Feature Learned K-R Receiver with MSE Decomposition for Low-Resolution ADC 5G mm Wave systems

We show that a simple nonlinear feature expansion combined with per-slot least-squares adaptation can explain and mitigate quantisation distortion in low-resolution ADC 5G mmWave receivers. The proposed NL-feat K-R receiver decomposes equalization into a physics-based MMSE K step and a data-driven R step trained per slot from 136 DMRS pilots using closed-form least squares—no backpropagation, no offline data. By appending quadratic terms [Re²,Im²,Re·Im] that capture the dominant second-order ADC distortion structure, the 9-feature model improves over the 4-feature baseline by + 0.070 bps/Hz at SNR = 20 dB. Proposition 1 provides an orthogonal decomposition of the residual error into a quantisation-floor component (γ_Q) and a channel-error projection component (γ_CE), offering interpretable insight into the gain mechanism. Simulations over 5,000 Monte Carlo trials (3GPP CDL-A, 4-bit ADC, 64-QAM, CR = 2/3) demonstrate statistically significant gains over all baselines: +0.197 bps/Hz over MMSE (p < 0.001), + 0.220 bps/Hz over OAMP-Net (p < 0.001)—which tracks near-MMSE as predicted by orthogonal AMP theory under non-Gaussian noise [14]—and + 1.387 bps/Hz over DetNet, whose gradient projection is not calibrated for continuous OFDM equalization with quantisation. SE gains are maintained at p < 0.001 for all UE speeds from 0 to 500 km/h at 28 GHz.