A question about state covariance

[Richardson-Chong]

In the Forster's paper, the covariance of states relate to the random walk noise, but in practice the result is bad that the informative matrix base on the state is so large dominating a large part and I've found that in your code the covariance of the states if composed of measurement noise which might be the noise of gyro and acc? So does your method have a theoretical sustenance or just come from the experiment?