1: \begin{abstract}

2: In this work, we describe advanced numerical tools for working with multivariate functions and for the 

3: analysis of large data sets. These tools will drastically reduce the required computing time and the 

4: storage cost, and, therefore, will allow us to consider much larger data sets or finer meshes. 

5: Covariance matrices are crucial in spatio-temporal statistical tasks, but are often very expensive to 

6: compute and store, especially in 3D. Therefore, we approximate covariance functions by cheap surrogates 

7: in a low-rank tensor format. We apply the Tucker and canonical tensor decompositions to a family of 

8: Mat\'ern- and Slater-type functions with varying parameters and demonstrate numerically 

9: that their approximations exhibit exponentially fast convergence. 

10: We prove the exponential convergence of the Tucker and canonical approximations in tensor

11: rank parameters.

12: Several statistical operations are performed in this low-rank tensor format, including evaluating the 

13: conditional covariance matrix, spatially averaged

14: estimation variance, computing a quadratic form, determinant, trace, loglikelihood, inverse, 

15: and Cholesky decomposition of a large covariance matrix. 

16: Low-rank tensor approximations reduce the computing and storage costs essentially. 

17: For example, the storage cost 

18: is reduced from an exponential $\mathcal{O}(n^d)$ to a linear scaling $\mathcal{O}(drn)$, 

19: where $d$ is the spatial dimension, $n$ is the number of mesh points in one direction, 

20: and $r$ is the tensor rank. 

21: Prerequisites for applicability of the proposed techniques are the assumptions that the data, locations, 

22: and measurements lie on a tensor (axes-parallel) grid and that the covariance 

23: function depends on a distance, $\| x-y\|$. 

24: %

25: \end{abstract}

26: