1: \begin{abstract}

2: Gaussian processes (GP) are widely used as a metamodel for emulating time-consuming computer codes.

3: We focus on problems involving categorical inputs, with a potentially large number $\level$ of levels (typically several tens),

4: partitioned in $\nbGroup \ll \level$ groups of various sizes. 

5: Parsimonious covariance functions, or kernels, can then be defined by block covariance matrices $\mb{\matCovCat}$

6: with constant covariances between pairs of blocks and within blocks.

7: %However, little is said about the positive definiteness of such matrices, which may limit their practical usage.\\ 

8: We study the positive definiteness of such matrices to encourage their practical use.

9: The hierarchical group/level structure, equivalent to a nested Bayesian linear model,

10: provides a parameterization of valid block matrices $\mb{\matCovCat}$. 

11: The same model can then be used when the assumption within blocks is relaxed,

12: giving a flexible parametric family of valid covariance matrices with constant covariances between pairs of blocks.

13: The positive definiteness of $\mb{\matCovCat}$ 

14: is equivalent to the positive definiteness of a smaller matrix of size $\nbGroup$,

15: obtained by averaging each block.

16: The model is applied to a problem in nuclear waste analysis,

17: %We illustrate with an application in nuclear engineering, 

18: where one of the categorical inputs is atomic number, %in Mendeleev's periodic table 

19: which has more than 90 levels.

20: \end{abstract}

21: