Noise-Normalized Shrinkage for OFDM Channel Estimation

We present a noise-normalized shrinkage framework for OFDM channel estimation that achieves out-of-distribution (OOD) generalization across channel profiles, bandwidths, SNR ranges, and interference conditions—all from a single lightweight model trained on two channel profiles. The key insight is that normalizing transform-domain coefficients by the estimated noise standard deviation converts the input into per-tap SNR, making the optimal shrinkage function deployment-invariant. We show that this σ-normalization outperforms AIC-based order selection by 1.8–7.8 dB on unseen channels and bandwidths, while enabling extreme architecture compression from 2929 to 13 parameters with less than 0.5 dB degradation. Experiments on CDL, TDL, and urban-macro channels demonstrate that alternative normalizations (RMS-norm, no-norm) degrade by 3–6 dB out-of-distribution, and that adding OOD training data cannot compensate for the wrong normalization. A gate visualization reveals that the learned shrinkage acts as an adaptive soft window that overcomes the structural limitation of single rectangular truncation.