By deriving the physical model of the system, which includes definitions of both state variables and measurement equations, we apply an extended Kalman filter to Stokes space based polarization demultiplexing for complex-modulated signals. The convergence ratio, tracking, computational complexity, and system performance of this method are investigated and compared with the geometrical approach previously proposed to adaptive computation of the best fit plane. An analysis of the tuning parameters of both methods reveals that the Kalman filtering provides a more robust and stable polarization demultiplexing of signals. Nevertheless, if properly tuned, the geometrical approach attains a similar performance, with a gain of 90% in terms of complexity reduction.