For LGSSMs with a high-dimensional observation space, inference can be much more efficient by assuming diagonal covariance R. Specifically, it should be possible to speed up the calculation of Kalman gain and observation log-likelihoods, and also to reduce memory requirements when the noise is dynamic by storing R as a (ntime x obs_dim) array rather than (ntime x obs_dim x obs_dim).
For LGSSMs with a high-dimensional observation space, inference can be much more efficient by assuming diagonal covariance
R. Specifically, it should be possible to speed up the calculation of Kalman gain and observation log-likelihoods, and also to reduce memory requirements when the noise is dynamic by storingRas a (ntime x obs_dim) array rather than (ntime x obs_dim x obs_dim).