So I think I have my model reparameterized in terms of the pm.gp class. Thanks for your help @junpenglao! But I am realizing now that well MarginalSparse has many features it doesn’t seem to accommodate easily a predefined metric by which to eliminate correlations between points in X. As I understand it I don’t want to use inducing points (whereby I reduce the size of the Covariance matrix), but instead use all of X, but only consider some of the correlations between elements of X e.g. Cov(xi, xj) for some i j. To be clear a covariance matrix with the original dimensions of the data n x n with a bunch of zeros in it. Is this easily accomplished in MarginalSparse?