Abstract: In wideband applications, the number of coefficients in conventional digital predistortion (DPD) models increases prominently to compensate the high nonlinearity and deep memory effect of the PA. And the increase of the coefficients leads to very high computational complexity and ill-posed problems in the process of extracting the DPD parameters.This paper focuses on the parameter identification of complex DPD models and the implementation of low computational complexity DPD algorithms. A novel low complexity DPD method is proposed based on partial least squares (PLS). This method obtains new basis function matrices based on input/ output characteristics of a PA′s pre-inverse to reduce the number of coefficients. Compared with the existing dimension-reduction based DPD methods, this method further reduces both the number of coefficients in DPD models and the computational complexity with almost negligible performance loss. Experimental results demonstrate that the new method can greatly reduce complexity with good linearization.