Discretisation Techniques
In the following, (Q, B, P) denotes a probability space, where Q is the set of elementary events, B is the а-algebra of events and P is the probability measure. The symbol ю always specifies an elementary event ю e Q.
The random field k(x, ю) needs to be discretised both in the stochastic and in the spatial dimensions. One of the main tools here is the Karhunen-Loeve expansion (KLE) [16]. By definition, KLE of a random field к(х, ю) is the following series [16] „
к(х, со) = к(х) + X уД^ф((х)^(со), (6)
t=
where are uncorrelated random variables and k(x) is the mean value of
к(х, ю), Xt and фе are the eigenvalues and the eigenvectors of problem
Tфе = Хф фе e L2(G),е e N, (7)
and operator T is defined like follows
T : L2(G) ^ L2(G), (Tф)(х) := / cov*(x, у)фШу,
where covK is a given covariance function. Throwing away all unimportant terms in KLE, one obtains the truncated KLE, which is a sparse representation of the random field к(х, ю). Each random variable %e can be approximated in a set of new independent Gaussian random variables (polynomial chaos expansions (PCE) of Wiener [7, 25]), e. g.
&(ю)= E $Р)ни(в(®))’
^J
where 9(a) = (0i (a), 92(a),…), %(в) are coefficients, H are multivariate Hermite polynomials, в є J is a multiindex, J := {fifi = (p1,…,Pj,…), f5j є N0} is a multi-index set [18].
For the purpose of actual computation, truncate the polynomial chaos expansion after finitely many terms, e. g.
в є Jm, p := {в є J y(p) < M, P<P}, r(p) := max{j є Nв > 0}.
Since Hermite polynomials are orthogonal, the coefficients <^e) can be computed by projection
ф = ^Івщ{ете)те).
This multidimensional integral over 0 can be computed approximately, for example, on a sparse Gauss-Hermite grid with nq grid points
1 nq
* ji%Hp(8i)ZiWwi, (8)
where weights wi and points 9 і are defined from sparse Gauss-Hermite integration rule. After a finite element discretisation (see [10] for more details) the discrete eigenvalue problem (7) looks like
МСМф = Cij = covK(x, yj). (9)
Here the mass matrix M is stored in a usual data sparse format and the dense matrix C є Rnxn (requires O(n2) units of memory) is approximated in the sparse H – matrix format [10] (requires only O(n log n) units of memory) or in the Kronecker low-rank tensor format [9]. To compute m eigenvalues (m C n) and corresponding eigenvectors we apply the Lanczos eigenvalue solver [11, 22].