I am familiar with Kalman filtering given a linear (time-invariant) state space model. However, the state space parametrization is not unique. Given a controllable and observable state space model (A,B,C,D), I can find multiple equivalent forms, e.g. controllable canonical form, observable canonical form, Jordan normal form, etc.
Is there a state space parametrization which is beneficial to transform the model to before running the Kalman filter, especially in terms of numerical stability.
Are there papers you could point to? I am running a Kalman Filter on an 8th order SSM on a DSP and am running into numerical issues.
Thanks a lot!