Professional Documents
Culture Documents
problems
Saulo Pomponet Oliveiraa , Juarez dos Santos Azevedo.b
a
DMAT-UFPR, Centro Politécnico, 81531-980, Curitiba-PR, Brazil
b
CETEC-UFRB, Centro, 44380-000, Cruz das Almas-BA, Brazil.
Abstract
This work is concerned with the numerical approximation of homogeneous Fredholm integral
equations of second kind, with emphasis on computing the Karhunen-Loève expansion of
Gaussian processes. We employ the spectral element method with Gauss-Lobatto-Legendre
(GLL) collocation points. Similarly to piecewise-constant finite elements, this approach is
simple to implement and does not lead to generalized discrete eigenvalue problems, with
the additional advantage of providing high-order approximations of the eigenfunctions. Nu-
merical experiments confirm the expected convergence rates for some classical kernels and
illustrate how this approach can improve the finite element solution of partial differential
equations with random input data.
Keywords: Spectral element method, Fredholm integral of second kind, Karhunen-Loève
expansion.
1. Introduction
Eigenvalue problems involving the Fredholm integral equation of second kind are present
in many areas of sciences and engineering, such as acoustics [16], celestial mechanics [19], frac-
ture mechanics [9], image processing [24], information theory [8], laser theory [10], machine
learning [21], meteorology [11], and optics [34].
In the particular context of stochastic finite element methods, the discretization of ran-
dom fields must provide sufficient accuracy using as few random variables as possible. One
approach that is often employed is the Karhunen-Loève (KL) expansion [20]. The KL expan-
sion provides a parametrization of a random field in terms of uncorrelated random variables
and eigenfunctions of a homogeneous Fredholm integral equation of the second kind, whose
kernel is the covariance function of the random field.
Several techniques have been developed to approximate the solution to Fredholm equa-
tions of second kind, and three classes are particularly emphasized in the literature: degen-
erate kernel, projection, and Nyström methods [5]. The homogeneous case has been studied
within all three classes [1, 13, 23, 29] (see also [2, 14] for more recent approaches). As pointed
out in [5, 29], the discretization by degenerate kernel and projection methods demand the
evaluation of integrals; the conjunction of these methods with a quadrature scheme leads to
Nyström type methods (for instance, discrete Galerkin and discrete collocation methods [5]).
Wang [33] discusses pros and cons of collocation and discrete collocation methods.
In this work we propose the approximation of homogeneous Fredholm equations of sec-
ond kind by a discrete Galerkin method given by the spectral element method with Gauss-
Lobatto-Legendre (GLL) collocation points in conjunction with GLL quadrature. Spectral
element methods are high-order finite element methods that employ piecewise continuous
Lagrange shape functions whose collocation points are roots of orthogonal polynomials. This
class of methods has been successful on wave propagation problems mainly because they
2. Problem setting
Let D be a open and bounded domain in Rd (d = 1, 2, or 3) with closure D̄ and Lipschitz-
continuous boundary ∂D. We consider the Hilbert spaces L2 (D) and H k (D), k ≥ 1, equiped
with the usual inner products and induced norms. In particular, we consider the following
norms:
Z k
X X
2
kvkL2 (D) = |v(x)|2 dx, kvk2H k (D) = |v|2l , |v|2l = k∂ α vk2L2 (D) , (1)
D l=0 |α|=l
We refer to Y 0 = Y − hY i as its fluctuation around the mean, whereas hY 0 (x, ω)Y 0 (y, ω)i
is the covariance of Y . The process Y ∈ L2 (D × Ω) may be parameterized through of the
Karhunen-Loève expansion defined by
∞ p
X
Y (x; ω) = hY (x; ω)i + λk φk (x)ξk (ω), (2)
k=1
where ξk (ω) are independent Gaussian random variables with zero mean and unitary variance
and λk and φk (x) are respectively the eigenvalues and eigenfunctions associated the covariance
function K ∈ L2 (D × D). This eigenpairs of the covariance function K can be solved from
the following Fredholm equation:
Z
K(x, y)φk (y) dy = λk φk (x), k = 1, 2, . . . . (3)
D
In our examples we consider covariance functions, symmetric and positive definite, so that
all eigenvalues have real value.
which can be written as a generalized eigenvalue problem Kφk = λhk M φk , where the matrices
K and M are defined by the coefficients
The matrix K is typically dense, which demands a large computational effort, both in
memory and CPU time, to solve the discrete eigenvalue problem (efficient algorithms for this
purpose were proposed in [18, 22, 28]).
In the following we choose v1 , . . . vn as finite element basis functions. Let (Th )h>0 be a
family of regular and quasi-uniform affine partitions Th of D̄ into elements De (1 ≤ e ≤ Ne )
with maximum element length h > 0 such that Ḋe ∩ Ḋf = ∅ for e 6= f and
Ne
[
D̄ = D̄e . (8)
e=1
For each 1 ≤ e ≤ Ne , let x = F e (x̂) be the affine mapping from the element De to the
reference element D̂ = [−1, 1]d and let J e be the determinant of the Jacobian matrix of the
transformation F e .
By approximating the integral in (11) by the midpoint rule, we find following modified
eigenvalue problem
p
K P C φk = λhk φk , Ki,j
PC
= |Di ||Dj |K(x̄i , ȳ j ), (11)
where x̄i is the barycenter of the element Di . Note that this corresponds to the modified
Galerkin problem
Ne
X
aPh C (φhk , vh ) = λhk (φhk , vi ) ∀ vh ∈ V h , aPh C (u, v) = |Di ||Dj |K(x̄i , ȳ j )u(x̄i )v(ȳ j ). (12)
i,j=1
Finite elements of degree higher than zero lead to generalized eigenvalue problems in
which the matrix M is not necessarily diagonal. In analogy with finite element methods
for transient problems, mass-diagonal formulations are attractive, hence the GLL spectral
element method is a natural choice. Indeed, Andreev et al [3] have followed this approach
for an eigenvalue problem involving a second-order, elliptic differential operator.
Given a connectivity array IEN that assigns the local node i of the element De to the
global node I = IEN (i, e), we assemble the global basis functions vI ∈ Vh , in order that
Let NI be the total number of global nodes. By construction, the global functions vi ∈ Vh
also satisfy a Lagrange property
where the global nodes xJ satisfy xJ = F e (ξ i ) for any i, e such that J = IEN (i, e). On the
other hand, we perform numerical integration with a product GLL quadrature as follows:
d
Z Ne Z
X X X −1
Ne (N +1)
e
f (x) dx = e
f (F (x̂))J dx̂ ≈ wl f (F e (ξ l ))J e ,
D e=1 D̂ e=1 l=0
where wl > 0 are the product quadrature weights corresponding to the collocation points ξ l ,
1 ≤ l ≤ (N + 1)d − 1. In the case of a continuous integrand f , we can rewrite the expression
above as
d
X X −1
Ne (N +1) NI
X X
e e
wl f (F (ξ l ))J = w̃J f (xJ ), w̃J = wl J e , (14)
e=1 l=0 J=1 (l,e)∈IENJ
where IENJ = {(l, e) ; J = IEN (l, e)}. In particular for f = vi vi , and taking (13) into
account, we have that the entries of mass matrix M , as shown in eq. (7), are approximated
as follows:
NI
X
GLL
Mi,j ≈ Mi,j := w̃J vi (xJ )vj (xJ ) = w̃i δi,j . (15)
J=1
We can rewrite this system in the following manner (see, e.g., [11]):
p p
M̃ K̃ M̃ φ̃k = λhk φ̃k , M̃ = diag w̃1 . . . , w̃N I , K̃i,j = K(xi , xj ), φ̃k = M̃ φk . (18)
Analogoulsy to (12), the eigenvalue problem (18) corresponds to the modified Galerkin
problem
Ni
X
aGLL
h (φhk , vh ) = λhk (φhk , vi )h ∀ vh ∈ V h , (u, v)h = w̃i u(xi )v(y i ), (19)
i=1
XNi
aGLL
h (u, v) = w̃i w̃j K(xi , xj )u(xi )v(y j ).
i,j=1
Remark 3.1. The modified Galerkin problem (19) with N = 1 redusces to the Nyström
method with the trapezoidal rule [30]. Moreover, when N = 2 we recover the Nyström method
with the Simpson’s rule [33].
4. Error analysis
The error analysis of the finite element with piecewise constant and GLL basis functions
will be presented in two steps: a review of the error bounds of the Galerkin method, and a
complement to these results in order to take numerical integration into account.
In the following, V := {Vh }h∈H ⊂ L2 (D) denotes a finite element space family indexed
by the discretization parameter h, and Ph denotes the L2 (D) orthogonal projection onto Vh .
Moreover, K ∈ L2 (D × D) denotes a symmetric and nonegative covariance kernel and K
denotes the following compact operator associated with the kernel K,
Z
(Kφ)(x) = K(x, y)φ(y) dy, (20)
D
and (λm , φm )m≥1 be the eigenpair sequence of the associated integral operator K via (20),
such that kφm kL2 (D) = 1. In the results below, m is an arbitrary positive integer such that
λm > 0 and C denotes a generic positive constant that does not depend on the discretization
parameter h.
The following result concerns the regularity of the eigenvalues and eigenfunctions of this
operator.
Theorem 4.1. If K ∈ H k (D × D) with k > 0, then φm ∈ H k (D) and there exists C =
C(K) > 0 such that λm ≤ Cm−k/d .
Theorem 4.2. If K ∈ H k (D × D) with k > 3N + 4, then there exists C = C(K, N ) > 0 and
0 < s < 1/2 such that |φm |N +1 ≤ Cλ−s
m .
Proof: We have from [28, eq. (A.5)] that, for any l ∈ {0, 1, . . . , k − 1} there exists a
constant C > 0, which depends on k, D, and ε0 = max{|λm | ; m ≥ 1}, such that
Assumption 4.1. There exists 0 < s < 1/2 and C = C(K, V, s) > 0 and a function
Φ : H → R such that, for any h ∈ H,
Let λhm be the m-th eigenvalue (m ≥ 1) calculated with the Galerkin method (6). We
have the following result regarding the eigenvalue error λm − λhm .
Theorem 4.3. If K ∈ H k (D × D) with k > 0 and Assumption 4.1 holds, then there exists
a constant C = C(K, V) > 0 such that, for any h ∈ H,
Proof: We have from [28, Cor. 2.8] that, for any s > 0, there exists C = C(K, V, s) > 0
such that
0 ≤ λm − λhm ≤ CΦ(h)2 λ1/2−s
m
for any h ∈ H and m ≥ 1 with λm > 0. Thus, it follows from Theorem 4.1 that
k(1/2−s)
0 ≤ λm − λhm ≤ c1 c K,V,s Φ(h)2 m− d .
k(1/2−s)
For s < 1/2, we find m− d ≤ 1, and the result follows.
Remark 4.1. The above setting was presented by Schwab and Todor [28, 32]. Therein,
covariance kernels belong to the more general set of piecewise regular functions (see [28,
Def. 2.15]). Moreover, Schwab and Todor provided a stronger result when the operator K is
associated with a piecewise smooth covariance kernel, namely an error bound for the error of
approximating the trace of the operator K by the trace of the Galerkin operator Ph KPh [28,
Thm. 3.2].
Remark 4.2. Following Todor [32, Sec. 3.3], we have that the Galerkin method with piece-
wise constant polynomials satisfies Assumption 4.1 with Φ(h) := h and thus λm −λhm = O(h2 ).
In fact, Todor considers the more general spaces of discontinuous piecewise polynomials of
total degree at most p − 1, for which λm − λhm = O(hp ).
Let us verify that the space of continuous finite elements satisfy Assumption 4.1, in
analogy with [32, Sec. 3.3]. For each N ≥ 1 and h > 0, let Vh be the space of continuous
piecewise polynomials of total degree at most N on the regular affine mesh Th defined in Sec.
3.
Proof: By construction, the space Vh is associated with an affine family (K, PK , ΣK )K∈Th
of continuous finite elements with respect to the reference finite element (D̂, V̂N , Σ̂), where
Σ̂ denote C 0 interpolation degrees of freedom. We also have that PN (D̂) ⊂ V̂N ⊂ H 1 (D̂),
where PN (D̂) = span{x̂k11 . . . x̂kdd ; k1 + . . . kd ≤ N }. Moreover, H N +1 (D̂) is contained in
C 0 (D̂) with a compact injection, since N > d/2 − 1 [15, p. 114]. Therefore, all conditions of
[15, Thm. 3.2.1] are met and (23) holds.
Remark 4.3. The interpolation error estimate (23) holds in general for an affine family of
finite element methods (see [15, Thm. 3.2.1]). For general quadrilateral meshes, additional
assumptions on the finite element space are required (see [4, p. 915]).
By Céa’s lemma [15, Thm 2.4.1] we also have that the L2 (D) orthogonal projection
operator onto Vh satisfies
Proof: We have from Theorem 4.1 that φm ∈ H k (D) ⊂ H N (D) for any m ≥ 1. Following
[32, Prop. 3.14], we have from (24) and Theorem 4.2 that there exists 0 < s < 1/2 such that
Corollary 4.1. Let λm > 0 be the m-th eigenvalue of the operator K in (20), associated with
a symmetric and nonegative covariance kernel K ∈ H k (D × D) with k > 3N + 4, and let
λhm be the m-th eigenvalue calculated with the Galerkin method (6) with continuous piecewise
polynomials of degree N on a rectangular mesh Th of maximum mesh-size h. We have that
λm − λhm = O(hN +1 ).
Let us now proceed to the error analysis taking quadrature error into account. Let us
first consider the auxiliary problem
Proceeding as in [3], we find |λ̃hm − λhm | ≤ Ch2N (aqui podemos arriscar mandar desse
jeito ou justificar por alto as diferenças: o Teor. 5.2 e as eqs. (6.1), (6.3) e (6.5), no minimo).
Our next step is to find an upper bound for |λ̃hm − λh,GLL m |, where λh,GLL
m is the m-th
eigenvalue of the discrete problem (19). We have that the interpolant of K ∈ H k (D × D) in
the product space Vh × Vh satisfies, for k ≥ 2N + 1,
a(uh , vv ) − aGLL
h (uh , vv ) = e1 + e2 ,
Z Z
e1 = [K(x, y) − I h K(x, y)]vh (y)uh (x)dydy, (27)
D D
Z Z
e2 = I h K(x, y)vh (y)uh (x) dy dx − aGLL
h (uh , vv ).
D D
With the aid of (26) and the Cauchy-Schwartz inequality, we find the following upper
bound for the first term in (27):
|e1 | ≤ Ch2(N +1) |K|2(N +1) kuh kL2 (D) kvh kL2 (D) . (28)
where p̂e (x̂) = p(F e (x̂)) |De for any p ∈ Vh and ŵe (x̂, ŷ) = ûeh (x̂)v̂he (ŷ). Following [3], we note
that f e is a continuous linear functional that satisfies f e (q̂) = 0 for any q̂ ∈ PN −1 (D̂ × D̂),
and invoke Bramble-Hilbert’s lemma [15, Thm. 4.1.3] and [15, pp. 118-120] to find
|f e (ŵe )| ≤ C|ûeh |N |v̂he |N ≤ Ch2N (J e )−1 |uh |N |vh |N , (em Andreev eq (3.11), seria |ûeh |N/2 |v̂he |N/2 )
Next, following [3, Them. 5.2], the min-max characterization of the approximate eigen-
values λ̃hm of (25) and λh,GLL
m of (19) are respectively given as follows:
a(vh , vh ) aGLL
h (vh , vh )
λ̃hm = min maxm , λh,GLL
m = min maxm , (30)
mE ⊂Vh vh ∈E (vh , vh )h m E ⊂Vh vh ∈E (vh , vh )h
Theorem 4.6. Let φl be the exact eigenfunction of (3) corresponding to the simple eigenvalue
λl normalized em L2 (D). Assume that φl ∈ H k+1 (D), k ≥ 2, l = 1, 2, . . .. Let φ̃hl be the
approximate eigenfunction of (6) corresponding to the approximate eigenvalue λ̃hl , normalized
with respect to | · |. Then, for h sufficiently small
φ̃hi
Φ̃hi = , i = 1, . . . , N (h). (35)
|φ̃hi |0
Putting
|φk − φ̃hk |0 ≤ |φk − Φ̃hk |0 + |Φ̃hk − φ̃hk |0 . (36)
we estimate the two terms at the right-hand side.
Step I: Using (24) and the fact of norms | · |h and | · |0 are uniformly equivalent ( [3],
Lemma 3.2) we have
Following [3] we may show that bounded for the first term of the right-hand side from the
above equation is
N (h)
X
|(P φl , φ̃hi )h |2 ≤ Ch2(k−1) . (40)
i=1
i6=l
Note that
N (h)
X
|P φk − (P φk , φ̃hk )h φ̃hk |2h = |(P φl , φ̃hi )h |2 ≤ Ch2(k−1) . (41)
i=1
i6=l
From of this inequality, equivalence of norms, and (24) it follows that
|P φk − (P φk , φ̃hk )h φ̃hk |0 ≤ |φk |0 + |P φk − (P φk , φ̃hk )h φ̃hk |h ≤ Chk−1 (42)
which implies that
|(P φk , φ̃hk )h ||φ̃hk |0 ≤ |φk |0 + |P φk − (P φk , φ̃hk )h φ̃hk |h ≤ 1 + Chk−1 , (43)
and
|(P φk , φ̃hk )h ||φ̃hk |0 ≥ |φk |0 − |P φk − (P φk , φ̃hk )h φ̃hk |h ≤ 1 − Chk−1 . (44)
Consequently, as |φk |0 ≤ c2 by equivalence of norm and orthonormality of the eigenfunctions
(P φl , φ̃hl )h − 1 ≤ Chk−1 .
h
(45)
|φ̃ | i 0
Proof: We have
(λl − λ̃hl )(φl , φ̃hl ) = λ̃hl [(P φl , φ̃hl ) − (φl , φ̃hl )] = λ̃hl [(P φl − φl , φ̃hl ) − E(φl , φ̃hl )]. (54)
We remark that
At one side, from the eigenvalue equation (4) and considering the projection P , we have
|(φl − P φl , φl )| = λ−1 −1 2
l a(φl − P φl , φl − P φl )| ≤ Cλl |φl − P φl |1 , (58)
E(P φl , φ̃hl ) = E(P φl − (φl )l , φ̃hl − (φl )l ) + E(P φl − (φl )l (φl )l ) (61)
+ E(P φl − (φl )l , φ̃hl − (φl )l ) + E((φl )l , (φl )l ). (62)
We estimate each term separately. Invoking Corollary 3.6 ([2], pp. 297) with i = k − 1
and j = k we find and using (4.4), (24) and rever esta equacao and (33) we get
X
|E(P φl − (φl )l , φ̃hl − (φl )l )| ≤ |EK ((P φl − (φl )l )(φ̃hl − (φl )l ))|
K∈Th
X
≤ |ChK |P φl − (φl )l |0 |(φ̃hl − (φl )l |1
K∈Th
≤ Ch[kP φl − φl k1 + kφl − (φl )l k1 ]
× [|φ̃hl − φl |0 + |φl − (φl )l |0 ]
≤ Ch2k . (63)
Applying Theorem 3.7 ([2], pp. 297-298) we obtain similarly
X
|E(P φl − (φl )l , (φl )l )| ≤ |EK ((P φl − (φl )l )(φl )l ))|
K∈Th
X
≤ ChK |P φl − (φl )l |0 |(φl )l |k
K∈Th
≤ Ch2k .
and
X
|E((φl )l , (φl )l − φ̃hl )| ≤ |EK (((φl )l )((φl )l − φ̃hl ))|
K∈Th
X
≤ |ChK |(φl )l |0 |(φl )l − φ̃hl |0
K∈Th
!1/2
X
≤ Chk [ |φl − (φl )l |2k + |φl |k ]
K∈Th
h i
× |(φl )l − φl |0 + |φl − φ̃hl |0
≤ Ch2k−1 . (64)
To estimate the last term in (61) we invoke Theorem 3.5 ([2], pp. 297-298) and use xxx. We
find
X
|E((φl )l , (φl )l )| ≤ |EK ((φl )l )(φl )l ))|
K∈Th
X
≤ |Ch2k 2
K |(φl )l |k
K∈Th
!1/2 2
X
≤ Ch2k |φl − (φl )l |2k + |φl |k ≤ Ch2k . (65)
K∈Th
it follows that
Next, we derive from (56), equivalence of norms, orthonomality of eigenfunctions, (24) and
(34)
Invoking Corollary 3.6 ([2], pp. 297) with i = k and j = k we find and using (4.4), (24) and
(34) we get for first term
X
|E(φ̃hk − (φk )k , P φk − φ̃hk )| ≤ |EK ((φ̃hk − (φk )k )(P φk − φ̃hk )|
K∈Th
X
≤ |ChK |φ̃hk − (φk )k |0 |P φk − φ̃hk |1
K∈Th
The second term in (74) may be estimated by applying Theorem 3.7 ([2], pp. 297-298) and
by using again (4.4), (24) and (34),
X
|E((φk )k , P φk − φ̃hk )| ≤ |EK ((φk )k )(P φk − φ̃hk ))|
K∈Th
X
≤ |ChlK |(φk )k |0 |P φk − φ̃hk |0
K∈Th
!1/2
X
k
≤ Ch [ |φk − (φk )k |2k + |φk |k ]
K∈Th
h i
× |P φk − φk |0 + |φk − φ̃hk |0
≤ Ch2k−1 . (76)
Substituting these two inequalities in (74) we obtain
This combination of this inequality, (73) and (68) proves the theorem. Finally, the estimate
(67) leads to an improvement of estimate (53) when the bilinear form a(·, ·) on V × V is
regular in the sense of [[15], p. 138]. Then by the Aubin-Nitshe argument, (24) implies
Theorem 4.9. Retain the assumptions of Theorem 4.7. Moreover, assume the bilinear form
a(·, ·) to be ”regular”, then we have, for k ≥ 2
5. Numerical experiments
To illustrate the validity and merits of the spectral element method with GLL points for
solving the Fredholm integral equation and its importance to practical KL simulation, we
consider three numerical examples. In the first set of experiments we verify the error of eigen-
value approximation of piecewise constant and spectral elements considering three stationary
covariance kernels, where we can verify the convergence rates [32] of the proposed approach.
Afterwards, we consider the random linear advection equation for which the velocity and the
initial condition are independent random functions. Finally, we present simulations to head
covariance for transient flow in a two-dimensional statistically homogeneous porous medium
with a separable exponential covariance function of the log transmissivity.
In the following we refer to the solutions using piecewise-constant finite elements and
spectral elements as PC and GLL, respectively.
5.1. Example 1
In this example we aim to study the relative error ek = |1 − λhk /λk | of approximating
k-th eigenvalue of (3) with the methods described in the previous sections. Following [7], let
D =]0, 1[ and K(x, y) be defined as
where the variance σ is constant and η > 0 governs the decay rate of the spatial correlation.
The exact eigenvalues and eigenfunctions associated with K are
2ησ ηγi cos(γi x) + sin(γi x)
λi = , φi (x) = , (80)
η 2 γi2 + 1
p
(η 2 γi2 + 1)/2 + η
where the parameters γ1 , γ2 , . . . are roots of the equation (η 2 γ 2 − 1) sin(γ) = 2ηγ cos(γ). For
D =]0, 1[×]0, 1[, let K(x, y) be the separable exponential covariance function
The exact eigenvalues and eigenfunctions associated with K(x, y) are respectively λ2D k =
λi λj and φ2Dk (x 1 , x2 ) = φ (x
i 1 )φj (x 2 ), where {λi , φi } are defined in (80) and the index n=
n(i, j) is set to arrange the eigenvalues in decreasing order.
We consider the input parameters σ = 1 and η = 0.1. Figure 1 shows the relative error
of the 10th eigenvalue of the covariance kernel (79) and the 100th eigenvalue of the kernel
(81) using increasingly refined meshes and GLL elements of degree 1, 2, 3, and 4, along with
Figure 1: Relative GLL error the 10th (100th) eigenvalue versus number of elements for the 1D (left) and
2D (right) exponential covariance kernel.
piecewise-constant elements. The numerical convergence rates did not increase with the
polynomial degree. This is justified by the low regularity of the covariance kernels (79) and
(81) due to the evaluation of absolute values.
Let us now assess the convergence rates of the GLL spectral method for the smooth
Gaussian kernel
K(x, y) = σ 2 exp(−(x − y)2 /η 2 ), σ, η > 0 (82)
and the sinc kernel
sin c(x − y)
K(x, y) = , c > 0. (83)
π(x − y)
For the Gaussian kernel a reference solution was calculated with the GLL spectral element
method with N = 16 and Ne = 212 , whereas the eigenvalues of the sinc kernel were computed
according to the procedure presented in [12]. We employed the parameters σ = 1, η = 0.1,
and c = 15. Figure 2 shows the relative error (with respect to the reference solution) of
the 10th eigenvalue of the covariance kernel (82) for piecewise-constant and GLL elements
of degree 1, 2, 3 and 4. The convergence rate does increase with N , as expected. This shows
that the GLL has a positive effect on the discretization Fredholm integral equation and it is
relevant in the approximation of eigenvalues and eigenfunctions.
Figure 2: Relative GLL error the 10th eigenvalue versus number of elements for the Gaussian (left) and the
sinc (right) covariance kernel.
5.2. Example 2
This second experiment is a one-dimensional advection problem presented by Dorini and
Cunha [17]. In this problem, we seek the scalar field Q : R × Ω × [0, T ] → R such that,
µ-almost everywhere in Ω,
∂Q(x; ω; t) + A ∂Q(x; ω; t) = 0, T > 0, x ∈ R,
∂t ∂x (84)
Q(x; ω; 0) = Q0 (x; ω).
where the velocity A ∼ N (−0.5, 0.16) is a normal random variable and the random velocity
and Q0 (x; ω) is a Gaussian process that has the mean
1 , x ∈ (1.4, 2.2)
hQ0 (x; ω)i = −20(x−0.25)2 (85)
e , otherwise,
and the covariance given by equation (79) with correlation length η = 0.3. We also suppose
that the velocity and the initial condition are independent. Following [17], a reference solution
is computed with the Monte Carlo method using statistically 30000 independent suites of
realizations of A and Q0 (x; ω). The correlated realizations of Q0 (x; ω) are computed with
the MATLAB function mvnrnd(). Observe that each realization A(ω) and Q0 (x; ω) yields
analytical solution given by Q(x; t; ω) = Q0 (x − A(ω)t; ω).
Let Q0T (x; ω) = Q(x; ω; T ) − hQ(x; ω; T )i be the fluctuation of Q around the mean and let
T = 0.4. Fig. 3 represents the variance σQ 2
(x) = hQ0T (x; ω)2 i of the reference solution along
with PC and GLL approximations. For GLL we consider the degrees N = 1, 2, 4. For both
PC and GLL we retain all eigenpairs in the truncated KL expansion and use elements of size
h = 1/4. Note that PC accurately approximates the average of the reference solution at each
element, whereas GLL approaches the reference solution as N increases. On the other hand,
PC fails to approximate the reference solution elementwise when we consider the covariance
CQ0 (x) = hQ0T (x; ω)Q0T (0; ω)i, as shown in Fig. 4. The GLL approximation of CQ0 (x) was
consistent with the approximation of the variance, indicating the robustness of this method.
0
Figure 4: Covariance CQ (x) = hQ0T (x; ω)Q0T (0; ω)i of the scalar field Q(x; ω; T ) for T = 0.4.
5.3. Example 3
In this example, we consider a problem, presented by Zu and Zhang [25], of transient flow
in saturated two-dimensional bounded randomly heterogeneous porous media governed by
the following continuity equation and Darcy’s law
Here T(x; ω) is the random transmissivity defined by a log-normal process, ie, T(x; ω) =
exp(Y (x; ω)), where Y is a Gaussian process. S is storativity, H1 and H2 are prescribed
constant heads, H0 (x) is the initial head in the domain D, L1 = 10 and L2 = 10 are
the lengths of the flow domain in x1 and x2 directions, respectively. The flow domain is
uniformly discretized into square elements. The no-flow conditions are prescribed at two
lateral boundaries and constant heads are specified on the left and right boundaries. At time
t = 0, the constant heads on the left and right boundaries are suddenly changed to H1 = 11
and H2 = 10, respectively. The storativity is a deterministic constant S = 0.005. The mean
of the log transmissivity is given as hY (x; ω)i = 0.0. The variance and the correlation lengths
of the log transmissivity field are defined as σY2 = 1.0 and η = 1.0.
We conduct Monte Carlo simulations to verify the accuracy of the numerical solutions
for transient head and its related cross covariances. First, we generate 5000 two-dimensional
unconditional realizations of the log transmissivity with the separable covariance function
as given in (81), using the random field generator based on the Karhunen-Loève expansion
with stochastic dimension M = 400. The value of the number of grid points, was fixed at
20 × 20 for both GLL and PC methods. A reference solution also is computed by the Monte
Carlo method with 5000 samples on the Karhunen-Loève expansion, considering the number
of grid points 40 × 40 and as stochastic dimension M = 6400. In this target solution we use
the GLL with degree N = 1. The quality of the realizations has been discussed by Lu and
Zhang [25].
In the following figures we show the solutions only along the profile x2 = L2 /2.
Figure 5: Comparisons of the transient variance head computed at different times t = 0, 0.01, 0.05 and 0.4.
Target (dashed curves) and approach solution (solid curves) by PC and GLL which considers N = 1, 2 and
4.
Figure 6: Transient cross-covariance between the log transmissivity Y (x1 , L2 /2) and the hydraulic head
h(x1 , L2 /2, t) with t = 0, 0.01, 0.05 and 0.4. Target (dashed curves) and approach solution (solid curves) by
PC and GLL considering N = 1, 2 and 4.
Figure 5 compares the transient head variance obtained from PC, GLL solution on the
MC simulations at various times. It seems that the first-order GLL solutions with a grid
20 × 20 is close to the results from the first-order GLL on a grid 40 × 40. The orders N = 2
and 4 of the GLL are adequately accurate at σ 2 = 1.0, especially when the flow is at or
near steady state. For PC, the results deviate slightly from the target solution, and such
deviations will increase with the increase of the variance of the log transmissivity since the
dimension M is maximum in this problem.
Figure 6 illustrates the cross covariance between the log transmissivity,
where Y 0 (x; ω) is the fluctuation of the log transmissivity, obtained from Monte Carlo method
(solid curves) and (GLL or PC) methods (dashed curves) at four times t = 0, 0.01, 0.05 and
0.4. The results obtained from the GLL with degrees N = 1, 2 and 4, and PC methods,
are presented in distinct figures for comparison. Excellent agreements are achieved in all the
times and degree N . However, the PC solutions deviate significantly from their Monte Carlo
counterparts. In addition, the comparison tells us that the GLL is able to be competitive
with the PC when the number of relevant terms in the Karhunen-Loève expansion is not so
small to preserve computational efficiency.
Let x̄ = (L1 /2, L2 /2). Figure 7 depicts the restricted cross covariance
This experiment once again demonstrate good performance of the GLL despite the anti-
symmetry of the profile plot of CY h in steady state, and lack of symmetry in the transient
state.
6. Conclusions
The spectral element method was employed to numerically compute the eigenvalues and
eigenfunctions of the Fredholm integral equation (3). The use of Gauss-Lobatto-Legendre
(GLL) collocation points and reduced integration with GLL quadrature rendered the imple-
mentation similar to piecewise constant finite elements in the sense that the resulting discrete
equations are a standard eigenvalue problem rather than a generalized one. The accuracy
and the efficiency of the proposal are tested for random process with sinc, Gaussian and
exponential covariance functions.
The eigenpairs obtained by GLL through Fredholm integral equation are applied in the
Karhunen-Loéve expansion. This expansion is chosen due to its efficiency in the representa-
tion of random processes. We employ the GLL on one-dimensional linear advection equation
Figure 7: Transient cross-covariance between the log transmissivity Y (L1 /2, L2 /2) and the hydraulic head
h(x1 , L2 /2, t) with t = 0, 0.01, 0.05 and 0.4. MC (solid curves) and approach solution (dashed curves) by
PC and GLL which considers N = 1, 2 and 4.
subject to random velocity fields and random initial conditions. This numerical experiments
clearly show that the solutions obtained by GLL are adequately precise. Accurate results
also were obtained considering a model transient saturated flow in two-dimensional, bounded,
randomly heterogeneous porous media. The obtained results also indicate that, the solutions
for the moments and covariance converge quickly, for small degrees of GLL.
In future work we intend to extend this study to treat of convergence results for the GLL,
combined with upscaling techniques. This is a topic of significant importance to be explored.
Another subject of interest will include the analysis of second order wave equation with a
random wave speed.
Acknowledgements
This research is supported by Brazilian agency CNPq under grants 314553/2009-6 and
471182/2011-7.
References
[1] R. B. Adams. On the approximate solution of Fredholm’s homogeneous integral equa-
tion. Amer. J. Math., 51:139–148, 1929.
[3] A.B. Andreev, V.A. Kascieva, and M. Vanmaele. Some results in lumped mass finite-
element approximation of eigenvalue problems using numerical quadrature formulas. J.
Comput. Appl. Math., 43(3):291–311, 1992.
[5] K. Atkinson. The numerical solution of integral equations of the second kind. Cambridge
University Press, Cambridge, 1997.
[9] Z. Bažant and Y. Li. Stability of cohesive crack model: Part II – eigenvalue analysis of
size effect on strength and ductility of structures. J. Appl. Mech., 62:965–969, 1995.
[10] H. Brunner, A. Iserles, and S. P. Nørsett. The spectral problem for a class of highly
oscillatory Fredholm integral operators. IMA J. Numer. Anal., 30(1):108–130, 2010.
[11] C. Buell. Integral equation representation for factor analysis. J. Atmos. Sci., 28(8):1502–
1505, 1971.
[12] A. Van Buren and J. Boisvert. Accurate calculation of prolate spheroidal radial functions
of the first kind and their first derivatives. Q. Appl. Math., 60(3):589–599, 2002.
[14] Z. Chen, G. Nelakanti, Y. Xu, and Y. Zhang. A fast collocation method for eigen-
problems of weakly singular integral operators. J. Sci. Comput., 41(2):256–272, 2009.
[15] P. Ciarlet. The Finite Element Method for Elliptic Problems. SIAM, Philadelphia, 2002.
[17] F. A. Dorini and M. C. C. Cunha. Statistical moments of the random linear transport
equation. J. Comput. Phys., 227(19):8541–8550, 2008.
[18] M. Eiermann, O. Ernst, and E. Ullmann. Computational aspects of the stochastic finite
element method. Computing and Visualization in Science, 10(1):3–15, 2007.
[19] N. Evans and J. Read. Stability of power-law discs – I. The Fredholm integral equation.
Mon. Not. R. Astron. Soc., 300(1):83–105, 1998.
[20] R. Ghanem and P. Spanos. Stochastic Finite Element: A Spectral Aproach. Springer,
New York, 1991.
[21] M. Girolami. Orthogonal series density estimation and the kernel eigenvalue problem.
Neural Comput., 14(3):669–688, 2002.
[24] K. Khare. Sampling theorem, bandlimited integral kernels and inverse problems. Inverse
Probl., 23(4):1395–1416, 2007.
[25] Z. Lu and D. Zhang. Analytical solutions to statistical moments for transient flow in two-
dimensional, bounded, randomly heterogeneous media. Water Resour. Res., 41:W01016,
2005.
[26] S. P. Oliveira and G. Seriani. Effect of element distortion on the numerical dispersion
of spectral element methods. Commun. Comput. Phys., 9(4):937–958, 2011.
[29] I. Sloan. Iterated Galerkin method for eigenvalue problems. SIAM J. Numer. Anal.,
13(5):753–760, 1976.
[30] A. Spence. Error bounds and estimates for eigenvalues of integral equations. Numer.
Math., 29:133–147, 1978.
[31] T. Tang, X. Xu, and J. Cheng. On spectral methods for Volterra integral equations and
the convergence analysis. J. Comput. Math, 26:825–837, 2008.
[32] R. Todor. Robust eigenvalue computation for smoothing operators. SIAM J. Numer.
Anal., 44(2):865–878, 2006.
[34] E. Wolf. New theory of partial coherence in the space-frequency domain. part I: spectra
and cross spectra of steady-state sources. J. Opt. Soc. Am., 72(3):343–351, 1982.