I think that this must be possible using a forward and backward pass. At the very least, this must be possible when the predictive kernel is the same as the base kernel (by comparison with state-space models), but I feel like it must be possible for general semi-separable matrices.
I think that this must be possible using a forward and backward pass. At the very least, this must be possible when the predictive kernel is the same as the base kernel (by comparison with state-space models), but I feel like it must be possible for general semi-separable matrices.