Exact Inverse Models for Evaluating Nonlinear Digital Predistortion Techniques

Digital predistortion aims to negate distortion introduced by a power amplifier by estimating a preinverse system. Several algorithms for parameter estimation exist, but low-level conclusions are difficult to quantify as the true preinverse is not known and cannot serve as a ground truth comparison. We propose a novel method of constructing nonlinear models whose true inverse is known. This method utilizes the patterns that arise when the coefficients of arbitrary Volterra series kernels are transformed under composition with memory-storing linear filters and memoryless polynomial blocks. This method is used to construct and simulate power amplifier models whose known preinverses are employed to assess the performance of estimation algorithms using the previously unavailable metric of ground truth. The results of these simulations demonstrate that the secondary metric of spurious-free dynamic range (SFDR) does not always accurately indicate algorithm performance and may be highly misleading; specifically, we provide an instance for which SFDR indicates acceptable performance when the underlying model parameter estimates diverge.