We use a low-rank partial Cholesky plus diagonal observation noise as a preconditioner so we should make use of the matrix inversion lemma to save on compute when applying it.