Adaptive nugget gradient computation

[edaub]

In certain circumstances, the analytical derivative computation with an adaptive nugget can be somewhat off the value found by finite differences (particularly for the derivative with respect to the covariance scale). This is because when the nugget is just large enough to factorize the matrix, a small change in another parameter can cause the nugget found when computing the log posterior with finite differencing to change a bit, which gets mixed up in that derivative component.

I don't think there is anything to do about this -- the derivative is only used in finding the parameters that minimize the posterior, so it doesn't affect the correctness of any predictions. However, it's probably worth documenting that the derivative computation for adaptive nuggets treats the nugget as fixed in case anyone encounters this discrepancy.