On the Optimum Number of Coefficients of Sparse Digital Predistorters: A Bayesian Approach

This work presents insights on the application of the Bayesian information criterion (BIC) to fix the optimum number of coefficients in the Volterra series applied to the modeling and linearization of power amplifiers (PAs). The BIC is transformed from a rule to be applied after selection techniques to a stopping criterion, which enables the halting of the algorithm when a condition is reached. This study reveals that the BIC is equivalent to allow a certain identification normalized mean square error (NMSE) decrease after the inclusion of a model component. Experimental results of the digital predistortion of a class J PA are provided, demonstrating the proposal applicability in the attaining of the optimum number of coefficients. A comparison is made between the results obtained when the stopping rule is applied to the hill climbing (HC) and the doubly orthogonal matching pursuit (DOMP) algorithms.