1: \begin{abstract}

2: We analyze several Galerkin approximations 

3: of a Gaussian random field~$\mathcal{Z}\colon\mathcal{D}\times\Omega\to\mathbb{R}$ 

4: indexed by a Euclidean domain $\mathcal{D}\subset\mathbb{R}^d$

5: whose covariance structure 

6: is determined by a negative 

7: fractional power $L^{-2\beta}$   

8: of a second-order

9: elliptic differential operator 

10: $L:= -\nabla\cdot(A\nabla) + \kappa^2$.  

11: Under minimal assumptions 

12: on the domain $\mathcal{D}$,  

13: the coefficients 

14: $A\colon\mathcal{D}\to\mathbb{R}^{d\times d}$, 

15: $\kappa\colon\mathcal{D}\to\mathbb{R}$, 

16: and the fractional exponent $\beta>0$, 

17: we prove convergence in $L_q(\Omega; H^\sigma(\mathcal{D}))$ 

18: and in $L_q(\Omega; C^\delta(\overline{\mathcal{D}}))$ 

19: at (essentially) optimal rates 

20: for (i) spectral Galerkin methods and (ii) 

21: finite element approximations. 

22: Specifically, 

23: our analysis  

24: is solely based on  

25: $H^{1+\alpha}(\mathcal{D})$-regularity of 

26: the differential operator $L$, 

27: where $0<\alpha\leq 1$. 

28: % 

29: For this setting, 

30: we furthermore provide rigorous estimates 

31: for the error in the covariance function 

32: of these approximations 

33: in $L_{\infty}(\mathcal{D}\times\mathcal{D})$ 

34: and in the mixed Sobolev space 

35: $H^{\sigma,\sigma}(\mathcal{D}\times\mathcal{D})$, 

36: showing convergence which is 

37: more than twice as fast 

38: compared to the corresponding 

39: $L_q(\Omega; H^\sigma(\mathcal{D}))$-rate.  

40: 

41: For the well-known example of such 

42: Gaussian random fields, the 

43: original Whittle--Mat\'ern class, 

44: where $L=-\Delta + \kappa^2$ 

45: and $\kappa \equiv \operatorname{const.}$,  

46: we perform several numerical experiments 

47: which validate our theoretical results.  

48: \end{abstract}

49: