This paper is concerned with digital predistortion for linearization of RF high power amplifiers (HPAs). It has two objectives. First, we establish a theoretical framework for a generic predistorter system, and show that if a postdistorter exists, then it is also a predistorter, and therefore, the predistorter and postdistorter are equivalent. This justifies the indirect learning methods for a large class of HPAs. Secondly, we establish a systematic and general structure for a predistorter that is capable of compensating nonlinearity for a large variety of HPAs. This systematic structure is derived using approximation by separable functions, and avoids selection of predistorters based on the assumption of HPA models traditionally done in the literature.