Efficient Pruning Technique of Memory Polynomial Models Suitable for PA Behavioral Modeling and Digital Predistortion

This paper proposes an error variation ranking (EVR)-based pruning method to reduce the complexity of memory polynomials (MPs) for power amplifier behavioral modeling. During the EVR pruning, the variation of prediction error caused by removing each term is calculated and ranked as a quantification factor to show the term's importance. The dominant terms are then selected based on their ranking positions among all terms. This method is verified by comparing its results with all other possible selections under the same conditions. When it is used to prune digital predistorters, approximately 74% of the terms in the MP model and 78% of the terms in the 2-D digital-predistortion model can be removed with negligible deterioration of the prediction and linearization performance. Moreover, further discussion is presented to strategize the configuration of MP models based on the EVR pruning results.