Sample Pruning Based on Leverage for Digital Pre-Distortion

In this paper a method for identifying outlier data points during the training of a Digital Pre-Distortion (DPD) function is presented. In modern cellular networks, DPD functions are trained with complex modulated signals. The distribution of the magnitudes of the samples in these signals is heavily skewed. Through linear algebra manipulation, the leverage each sample exhibits on the calculated DPD coefficients can be obtained during the DPD training process. Data points in the training signal that are overly influential can be removed, and the function retrained.Experimental validation for this technique was performed using captured input and output power amplifier (PA) signals. Improvements to both PA modelling and DPD function modelling were observed. With regards to the improved DPD function, normalised mean squared error (NMSE) was improved by 5.4% and error vector magnitude (EVM) was improved by 30.3%.