fips.Posterior.innovation_cov#

property Posterior.innovation_cov: Covariance[source]#

The covariance of the innovation, \(H S_0 H^T + S_z / \gamma\), dense.

The innovation is \(z - H x_0 - c\), the observations minus what the prior models; this is its covariance when the prior error and the model-data mismatch are right, and the m x m matrix the solve factors. Kalman filtering calls it the innovation covariance; Rodgers (2000) writes it \(S_y\). It includes the regularization factor: the mismatch is divided by \(\gamma\). Its diagonal, square-rooted, is the spread expected for each observation’s innovation, of which \(H S_0 H^T\) is the prior’s share.