Beyond GMP: Neural DPD for Mitigating Non-Separable Memory Effects in Wideband PAs

As wideband transmitters push power amplifiers (PAs) toward saturation, the common separability assumption between nonlinear distortion and memory effects may no longer hold, limiting the effectiveness of polynomial digital predistortion (DPD). While prior work on DPD mainly emphasizes modeling accuracy or computational efficiency, the conditions under which polynomial models fundamentally break down remain insufficiently understood. This Letter presents a controlled comparative study of generalized memory polynomial (GMP) and neural-network-based DPD under explicitly defined separable and non-separable PA operating regimes. By constructing PA behaviors that selectively violate the nonlinear–memory separability assumption, the impact of non-separable memory coupling is isolated. Experimental results show that GMP and neural DPD achieve comparable adjacent channel leakage ratio (ACLR) performance exceeding 30 dB in separable regimes. In contrast, under non-separable conditions, GMP performance degrades to approximately 25 dB, while neural DPD maintains stable ACLR around 30 dB. These results identify a concrete failure condition of polynomial DPD and provide a practical criterion for selecting DPD architectures in wideband PA linearization.