Digital Predistortion (DPD) is a technique used to compensate for nonlinear distortion in RF Power Amplifiers (PAs). The Indirect Learning Architecture (ILA) is a commonly used method for identifying DPD coefficients. However, ILA produces biased estimates, leading to sub-optimal results in the presence of measurement noise. Additionally, the DPD coefficients obtained with ILA depend on the input signal used for identification, which requires a new identification process for different input signals. In this context, the question arises: is it possible to estimate DPD coefficients that fulfill some optimality criterion? To address this question, an approach is proposed that assumes knowledge of input signal statistics and a static quasi-memoryless polynomial model. The static nature of the model implies that the coefficients do not change with time, while the quasi-memoryless nature indicates that the polynomial coefficients are complex-valued, modeling both AMAM and AMPM. An analytical solution for the DPD coefficients that maximizes the Signal-to-Distortion-and-Noise Ratio (SDNR) is obtained. Simulation results shows that our approach outperforms the ILA. Furthermore, since our approach relies on knowledge of the PA models, it is possible to use it to obtain the optimal DPD coefficients for different input signals without the need for a new identification process.