Mamba Based Digital Predistortion for Wideband Doherty PowerAmplifiers

This paper investigates a Mamba based digital predistortion (DPD) architecture for linearizing a wideband Doherty power amplifier (DPA). The DPA is excited by a 160 MHz complex baseband signal sampled at 491.52 MHz, representative of 5G/6G carrier aggregation scenarios with strong memory effects. The proposed DPD employs a selective state space (Mamba) model that learns the inverse PA behavior directly from measured input--output data in an indirect learning architecture, using both I/Q samples and engineered nonlinear features to capture high order, long memory nonlinearities with linear time complexity. Measurements on a GaN Doherty prototype show an adjacent channel power ratio (ACPR) of -47.1 dBc and a normalized mean squared error (NMSE) of -36.88, outperforming a conventional memory polynomial and a CNN based DPD with comparable model complexity, demonstrating the potential of Mamba style state space sequence models as compact and scalable DPD solutions for wideband Doherty transmitters.