This thesis is focusing on developing a compact digital predistortion (DPD) system which costs less DPD added power consumptions. It explores a new theory and techniques to relieve the requirement of the number of training samples and the sampling-rate of feedback ADCs in DPD systems. A new theory about the information carried by training samples is introduced. It connects the generalized error of the DPD estimation algorithm with the statistical properties of modulated signals. Secondly, based on the proposed theory, this work introduces a compressed sample selection method to reduce the number of training samples by only selecting the minimal samples which satisfy the foreknown probability information. The number of training samples and complex multiplication operations required for coefficients estimation can be reduced by more than ten times without additional calculation resource. Thirdly, based on the proposed theory, this thesis proves that theoretically a DPD system using memory polynomial based behavioural modes and least-square (LS) based algorithms can be performed with any sampling-rate of feedback samples. The principle, implementation and practical concerns of the undersampling DPD which uses lower sampling-rate ADC are then introduced. Finally, the observation bandwidth of DPD systems can be extended by the proposed multi-rate track-and-hold circuits with the associated algorithm. By addressing several parameters of ADC and corresponding DPD algorithm, multi-GHz observation bandwidth using only a 61.44MHz ADC is achieved, and demonstrated the satisfactory linearization performance of multi-band and continued wideband RF transmitter applications via extensive experimental tests.