Pykalman with non-square observation matrix

Question

In the documentation of Pykalman it says that it only accepts square matrices for the observation_matrices and transition_matrices arguments.

Is there a way around this? I have to estimate a state-space system with a non-square observation matrix. Using the Pykalman notation, its dimensions would be [n_dim_obs, n_dim_state].

I could not find any other kalman filter library that accepts non-square observation matrices and has a loglikelihood method.

Answer 1

I have some experience with Pykalman and its documentation seems to have a lot of mistakes. So the observation matrix does not have to be a square one. It just has to map the estimation state to the measurement matrix.

Please have a look at my answer here. The observation matrix looks like H = [0 0 1] and Pykalman works fine. The dimension is exactly how you expect it: [n_dim_obs, n_dim_state].

If you need additinal information I would be glad to help.