Cookbook: lower bounds for kriging Maximum Likelihood Estimation (MLE)

• Y. Richet

This implementation is based on DiceKriging R package (CRAN version).

Considering maximization of log-likelihood in kriging models, we often face convergence issues when using gradient based optimization algorithms.
One reason stands for the fact that${dlogL \over d\theta_i} \left( u_j \right) = 0$ (if $i \neq j$), for all gaussians family kernels.
Another limitation of box optimization algorithm in $[0, +\infty[^d$ used on log-likekihood function is to ignore the extreme distances between conditional points (lowest or highest) which should affect the boundaries of the optimization domain.

For instance, we consider the following function to emulate with a
conditional gaussian process in $[0,1]$.

Proof of concept

We try to emulate such functions with some sparse sampling, to make the MLE optimization harder when $p$ increases:

Considering the range estimated by MLE for varying $p$ values, we observe that the convergence to $0$ value may have numerical causes:

We suggest to avoid $0$ range estimations using some heuristic lower bound related to worst (highest variation) low-distance data points:

Then, using a lower bound for range estimation related to this worst point may help to mitigate wrong $0$-range convergence issue:

where $\alpha$ is a tuning parameter, and

Robustness against Doe

Now we try to study the stability of this heuristic when the kriging conditional sampling varies:

Applying the heuristic for lower bound over range optimization gives:

Conclusion

The proposed heuristic seems to decrease bad convergence rate of MLE of kriging range.
It should be noted that the tuning parameter $\alpha$ may be related to the covariance kernel choosen, which was not studied here.