To get more efficient than the KF, need to dispense with the big covariance matrix -- costs too much to compute and multiply.  Get away with a Òlow-rankÓ approximation.