Elsevier Editorial System(tm) for Journal of Computational and Applied Mathematics

Manuscript Draft

Manuscript Number:

Title: Spectral element approximation of Fredholm integral eigenvalue problems

Article Type: Research Paper

Section/Category: 65Rxx

Keywords: Spectral element method; Fredholm integral of second kind; Karhunen-Loeve expansion

Corresponding Author: Prof. Saulo P Oliveira,

Corresponding Author's Institution: Universidade Federal do Parana

First Author: Saulo P Oliveira

Order of Authors: Saulo P Oliveira; Juarez S Azevedo

Main file
Click here to view linked References

Spectral element approximation of Fredholm integral

eigenvalue problems

Saulo P. Oliveiraa,∗, Juarez S. Azevedob

a Departamento de Matemática, Universidade Federal do Paraná, 81531-980, Curitiba-PR,
Brazil.
b CETEC-UFRB, Centro, 44380-000, Cruz das Almas-BA, Brazil.

Abstract

We consider the numerical approximation of homogeneous Fredholm integral

equations of second kind, with emphasis on computing truncated Karhunen-
Loève expansions. We employ the spectral element method with Gauss-Lobatto-
Legendre (GLL) collocation points. Similarly to piecewise-constant finite ele-
ments, this approach is simple to implement and does not lead to generalized
discrete eigenvalue problems. Numerical experiments confirm the expected con-
vergence 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.
2000 MSC: 45C05, 65R20, 35R60, 65M70

1. Introduction

Eigenvalue problems involving the Fredholm integral equation of second kind

are present in many areas of sciences and engineering, such as acoustics [15],
celestial mechanics [18], fracture mechanics [8], image processing [24], informa-
tion theory [7], laser theory [10], machine learning [20], meteorology [11], and

∗ Corresponding
author.
Email addresses: [email protected] (Saulo P. Oliveira), [email protected] (Juarez S.
Azevedo)

Preprint submitted to Journal of Computational and Applied Mathematics January 11, 2013
optics [35].
In the context of stochastic finite element methods [19], the discretization
of random fields must provide sufficient accuracy using as few random variables
as possible. One approach that is often employed is the Karhunen-Loève (KL)
expansion, 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 of Fred-
holm equations of second kind, and three classes are particularly emphasized in
the literature: degenerate kernel, projection, and Nyström methods [4], which
have been employed in the homogeneous case (see, e.g., [1, 23, 30]). As pointed
out in [4, 30], the discretization by degenerate kernel and projection meth-
ods 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 [4]). Wang [34] discusses pros and
cons of collocation and discrete collocation methods.
In this work we propose the approximation of homogeneous Fredholm equa-
tions of second kind by a discrete Galerkin method given by the spectral el-
ement method with Gauss-Lobatto-Legendre (GLL) collocation points in con-
junction with GLL quadrature [22]. Spectral element methods are high-order
finite element methods that employ piecewise continuous Lagrange shape func-
tions whose collocation points are roots of orthogonal polynomials. This class
of methods has been successful on wave propagation problems mainly because
they are flexible to deal with complex geometries and produce low dispersion
error [27]. Spectral elements have been used in the contexts of eigenvalue prob-
lems for differential operators [2] and Volterra integral equations [32], but to the
authors’ knowledge they have not yet been considered for solving the Fredholm
integral eigenvalue problem.
The discrete eigenvalue system obtained from the GLL spectral element
method can be recast as a Nyström collocation method, for which a vast litera-
ture on error analysis is available [3, 9, 23, 31]. We rephrase the classical error

2
analysis to the proposed method as well as numerically evaluate the approxima-
tion error for some smooth and non-smooth kernels. We also perform numerical
comparisons between truncated KL expansions obtained form the spectral ele-
ment method with those computed with the piecewise-constant finite element
method [5, 29], or equivalently, the wavelet Galerkin method with Haar basis
functions [28].
The remainder of the paper is organized as follows: in the next section
we present the problem setting and introduce the notations that will be used
throughout this work. In Section 3 we review the piecewise-constant finite el-
ement method and present the spectral element method with GLL collocation
points along with its fundamental properties. Theoretical error bounds for these
methods are provided in Section 4. Section 5 is devoted to numerical examples
that illustrate the accuracy of the proposed method and its application to prob-
lems with random input data. The paper closes with concluding remarks in
Section 6.

2. Problem setting

Let D be a bounded domain in IRd (d = 1, 2, or 3) with closure D̄ and

Lipschitz-continuous boundary ∂D. We consider the Hilbert space L2 (D) equiped
with the usual inner product and induced norm
Z
(u, v) = u(x)v(x) dx, kvk2L2 (D) = (u, v)1/2 , (1)
D

along with the spaces H k (D), k ≥ 1 with the standard norms and semi-norms
k
X X
kvk2H k (D) = |v|2l , |v|2l = k∂ α vk2L2 (D) , (2)
l=0 |α|=l

where α = (α1 , . . . , αd ) ∈ IN d and |α| = α1 + . . . + αd .

Let (Ω, F, µ) be a complete probability space, where Ω represents the set of
outcomes, F ⊂ 2Ω is a σ-algebra of events, and µ : F → [0, 1] is a probability
measure. Let Y : D × Ω → IR be a random field with expectation
Z
hY (x, ω)i = Y (x, ω) dµ
Ω

3
such that hY ′ (x, ω)2 i < ∞ for any x ∈ D, where Y ′ = Y − hY i denotes the
fluctuation of Y around the mean. We have that the covariance function

K(x, y) = hY ′ (x, ω)Y ′ (y, ω)i

is of nonnegative-definite type, i.e., for any finite subset Dn ⊂ D and for any
function h : Dn → IR,
X
K(x, y)h(x)h(y) ≥ 0. (3)
x,y∈Dn

Moreover, if Y is continuous in quadratic mean, then K admits the spectral

decomposition
∞
X
K(x, y) = λk φk (x)φk (y), (4)
k=1
where the nonnegative eigenvalues λk and the orthonormal eigenfunctions φk
(k ≥ 0) are the solutions of the homogeneous Fredholm integral equation
Z
K(x, y)φ(y) dy = λφ(x), x ∈ D. (5)
D

The Gaussian field Y can be represented by the Karhunen-Loève expansion

∞ p
X
Y (x; ω) = hY (x; ω)i + λk φk (x)ξk (ω), (6)
k=1

−1/2
where ξk (ω) = λk (Y, φk ) satisfies hξi ξj i = δij . In particular, if Y is a Gaus-
sian field, then ξk (ω) are independent Gaussian random variables with zero
mean and unitary variance [25, Sec 37].
Motivated by (3)-(6), we will consider finite element approximations of (5)
for symmetric, nonnegative-definite kernels. From here on, 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.
We have the following result concerning the regularity of the eigenvalues and
eigenfunctions of (5):

Lemma 2.1. Let p be a positive integer. If K ∈ H k (D × D) with k > 2p + 1,

then φm ∈ H k (D) and there exists C = C(K, N ) > 0 and 0 < s < 1 such that
−s
|φm |l ≤ Cλm for any l ≤ p.

4
 Proof: The regularity of φm follows from Proposition A.2 in [29]. We have
from equation (A.5) in [29] 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
|φm |l ≤ C|λm |−l/(k−1−l) = Cλm
−l/(k−1−l)
.

The proof follows by choosing l ≤ p.

Remark 2.1. The spaces H k (D) and H k (D × D) are respectively contained in

C 0 (D) and C 0 (D × D) with a compact injection if k > d/2 and k > d [14, p.
114].

3. Finite element approximation

Let us consider the variational formulation of the Fredholm integral equation

(5): find λk ∈ R and φk (x) ∈ L2 (D) (k = 1, 2, . . .) such that

a(φk , v) = λk (φk , v) ∀ v ∈ L2 (D), (7)

Z Z
a(u, v) = K(x, y)u(y)v(x) dy dx. (8)
D D

Let Vh = span{v1 , . . . , vn } ⊂ L2 (D). The Galerkin approximation to (7) in

Vh consists of finding λhk ∈ IR and φhk (x) ∈ Vh (1 ≤ k ≤ n) such that

a(φhk , vh ) = λhk (φhk , vi ) ∀ vh ∈ Vh , (9)

which can be written as a generalized eigenvalue problem Kφk = λhk M φk , where

the matrices K and M are defined by the coefficients

Ki,j = a(vj , vi ), Mi,j = (vj , vi ), 1 ≤ i, j ≤ n. (10)

The matrix K is typically dense, which demands a large computational ef-

fort, both in memory and CPU time, to solve the discrete eigenvalue problem.
Efficient algorithms for this purpose were proposed in [17, 21, 29].
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

5
elements De (1 ≤ e ≤ Ne ) with maximum element length h > 0 such that
Ḋe ∩ Ḋf = ∅ for e 6= f and
Ne
[
D̄ = D̄e . (11)
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 .

3.1. Piecewise-constant finite element method

Let Vh = VhP C be the space of piecewise-constant functions defined in Th .
The shape functions

 |De |−1/2 , x ∈ De

ve (x) = (12)
 0, x 6∈ De , e = 1, . . . , Ne


constitute an orthonormal basis for Vh , hence the discrete eigenvalue problem

reduces to
s
1
Z Z
Kφk = λhk φk , Ki,j = K(x, y) dy dx. (13)
|D ||Dj |
i
Di Dj

By approximating the integral in (14) by the midpoint rule, we find following

modified eigenvalue problem

K P C φk = λhk φk , PC
|Di ||Dj |K(x̄i , ȳ j ),
p
Ki,j = (14)

where x̄e = F e (0), 1 ≤ e ≤ Ne .

3.2. Spectral element method

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.
Starting from the space ÛN of polynomials with degree N ≥ 1 in [−1, 1], we
build the product space V̂N = (ÛN )d of multivariate polynomials with degree
N ≥ 1 in the reference element D̂ = [−1, 1]d .

6
 We choose the 1D local basis functions ûi (0 ≤ i ≤ N ) for ÛN as the
Lagrangian polynomials of degree N satisfying the relation ûi (ξj ) = δi,j (0 ≤
i, j ≤ N ), where ξ0 , . . . , ξN are the Gauss-Lobatto-Legendre (GLL) collocation
points, which are found by numerically solving the equation (1 − ξ 2 )PN′ (ξ) = 0,
where PN′ denotes the derivative of the Legendre polynomial of degree N [13].
Let Nve = dim V̂N = (N + 1)d be the number of element nodes. We construct
the local basis for V̂N from products of the 1D local basis functions ûj (for
instance, v̂i (x̂) = v̂(N +1)i1 +i2 (x̂1 , x̂2 ) = ûi1 (x̂1 )ûi2 (x̂2 ) in 2D). In this manner,
the polynomial basis functions v̂i ∈ V̂N satisfy v̂i (ξ j ) = δi,j for 0 ≤ i, j ≤ Nve −1,
where ξ j ∈ D̂ are built from products of the GLL points.
We consider the space Vh = VhGLL defined as the following space of contin-
uous and piecewise functions:
n o
VhGLL = p ∈ C 0 (D) ; p |De = p̂e ◦ (F e )−1 and p̂e ∈ V̂N , 1 ≤ e ≤ Ne .

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

vI (F e (x̂)) |De = v̂i (x̂), e = 1, . . . , Ne .

Let Nv be the number of global vertices. By construction, the global func-

tions vi ∈ Vh also satisfy a Lagrange property

vI (xJ ) = δI,J , 1 ≤ I ≤ Nv , (15)

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:
e
Z Ne Z
X Ne NX
X v −1
e
f (x) dx = e
f (F (x̂))J dx̂ ≈ ŵl f (F e (ξ l ))J e , (16)
D e=1 D̂ e=1 l=0

where ŵl > 0 are the product quadrature weights corresponding to the colloca-
tion points ξ l , 1 ≤ l ≤ Nve − 1 [13]. In the case of a continuous integrand f , we

7
can rewrite (16) as
e
Z Ne NX
X v −1 Nv
X
f (x) dx ≈ ŵl f (F e (ξ l ))J e = wJ f (xJ ), (17)
D e=1 l=0 J=1

with the global weights wJ (J = 1, . . . , Nv ) defined as follows:

X
wJ = ŵl J e , IENJ = {(l, e) ; J = IEN (l, e)}. (18)
(l,e)∈IENJ

In particular for f = vi vi , and taking (15) into account, we have that the
entries of mass matrix M in (10) are approximated as follows:
Nv
X
GLL
Mi,j ≈ Mi,j := wJ vi (xJ )vj (xJ ) = wi δi,j . (19)
J=1

Analogously, for the entries of the matrix K,

Nv X
X Nv
GLL
Ki,j ≈ Ki,j := wI wI K(xJ , xI )vi (xI )vj (xJ ) = wi wj K(xi , xj ), (20)
I=1 J=1

hence the resulting discrete eigenvalue system is

K GLL φk = λhk M GLL φk , GLL

Mi,j GLL
= wi δi,j , Ki,j = wi wj K(xi , xj ). (21)

As pointed out in [23], this system can be easily transformed into a symmetric
eigenvalue problem, which is convenient for numerical calculations. Moreover,
system (21) can be reduced to

K̃ GLL φk = λhk φk , GLL

K̃i,j = wj K(xi , xj ), (22)

which is a straightforward extension of system (14) to high-order polynomials.

We can further recast this system in the form of the Nyström collocation method
Nv
X
wj K(xi , xj )φhk (xi ) = λhk φhk (xi ), 1 ≤ i ≤ Nv . (23)
j=1

Remark 3.1. In the one-dimensional case, we recover the collocation method

with the trapezoidal and Simpson’s rules [9] when N = 1 and N = 2, respectively.

8
4. Error analysis

In this section we present error bounds for the approximate eigenvalues com-
puted with the methods outlined in Section 3. Let us denote by λ̃hm the eigen-
values of the Galerkin eigenvalue problem (13). Schwab and Todor [29, 33] have
proved that
λm − λ̃hm = O(h2 ). (24)

We define the following bilinear form associated with the discrete Galerkin
eigenvalue problem (14):
Ne
X
aP C
h (uh , vh ) = |Di ||Dj |K(x̄i , ȳ j )uh (x̄i )vh (ȳ j ). (25)
i,j=1

Using the Cauchy-Schwarz inequality and proceeding as in [2, Thm. 3.5], we

find
|a(vh , vh ) − aP C 2 2 2
h (vh , vh )| ≤ Ch kvh kL2 (D) , (26)

where C depends on D and K. On the other hand, the min-max characterization

of the approximate eigenvalues λ̃hm of (13) and λhm of (14) are respectively given
as follows:

a(vh , vh ) aP C
h (vh , vh )
λ̃hm = min maxm , λhm = min maxm , (27)
m
E ⊂Vh vh ∈E (vh , vh ) m E ⊂Vh vh ∈E (vh , vh )

where E m (1 ≤ m ≤ Nv ) is an m−dimensional subspace of VhP C . We have from

(26) and (27) that

a(vh , vh ) aP C (vh , vh ) a(vh , vh )

− Ch2 ≤ h ≤ + Ch2 , (28)
(vh , vh ) (vh , vh ) (vh , vh )

which yields |λ̃hm − λhm | ≤ Ch2 (see also [2, Thm. 5.2]). Thus from (24) and the
triangular inequality we find

λm − λhm = O(h2 ). (29)

Let us now consider the spectral element method in the Nyström formulation
(23):

9
Theorem 4.1. Let K ∈ H k (D × D) with k > 4N + 1 be a symmetric and
nonnegative covariance kernel and let (λm , φm ) be an eigenpair of (5) with
kφm kL2 (D) = 1. There exists an approximate eigenvalue λhm of (23) such that
λm − λhm = O(h2N ).

Proof: The steps of this proof are essentially the same as in [23]. Let us
define the truncation error
Z Nv
X
ǫi = K(xi , y)φm (y) dy − wj K(xi , xj )φm (xj ). (30)
D j=1

Recalling (17), we have that

Ne Z Nve −1
X X
ǫi = EK̂ [ĝie ]J e , EK̂ [ĝ] = ĝ(ŷ) dŷ − ŵj ĝ(ξ j ), (31)
e=1 D̂ j=0

where ĝie (x̂) = gie (F e (x̂)), gi (x) = K(xi , x)φm (x), and the upper index e de-
notes restriction to the element De . Following [2, Thm. 3.5], we note that EK̂
is a continuous linear functional that satisfies f e (q̂) = 0 for any q̂ ∈ P2N −1 (D̂);
it follows from Bramble-Hilbert’s lemma [14, Thm. 4.1.3] and the inequality
(3.1.20) in [14] that

|EK̂ [ĝie ]| ≤ C|ĝie |2N ≤ Ch2N (J e )−1/2 |gie |2N . (32)

We have from (31)-(32) and Lemma 2.1 that

|ǫi | ≤ Ch2N , 1 ≤ i ≤ Nv , (33)

where C now depends also on K and λm . Let v = [φm (x1 ), . . . , φm (xNv )]T
and let h0 > 0 such that v 6= 0 for any h ≥ h0 . There exists an approximate
eigenvalue λhm of (23) such that

Nv
! Nv
!−1
X X
|λm − λhm |2 ≤ wi ǫ2i wi φ2m (xi ) . (34)
i=1 i=1

Indeed, we have from (30) that

Nv
X
wi ǫi = λm wi φm (xi ) − wi K(xi , xj )wj φm (xj ),
j=1

10
or in vector form ADv = −Dǫ, where

A = K̃ GLL − λm I, D = diag({w1 , . . . , wNv }).

If λm is an eigenvalue of K̃ GLL , then (34) trivially holds. Otherwise, the

matrix A is invertible and kD1/2 vk2 ≤ kD−1/2 A−1 D1/2 k2 kD1/2 ǫk2 . Since the
kernel K is symmetric, the matrix à = D−1/2 A−1 D1/2 is symmetric, hence

kÃk2 = ρ(Ã) = ρ(A−1 ),

and identity (34) follows by choosing λhm such that |λhm − λm | = ρ(A−1 ). On
the other hand, it follows from (33) that
Nv
X
kD1/2 ǫk22 = wi ǫ2i ≤ max |ǫi |2 |D| ≤ Ch4N .
1≤i≤Nv
i=1

Moreover, analogously to (31)-(32), we have

 Nv

  Z X 
1/2 2 2
1 − kD vk2  =  φm (x) dx − 2
wi φm (xi ) ≤ Ch2N |φ2m |2N .
  
 D 
i=1

In particular, we can find h1 > 0 such that kD1/2 vk2 > 1/2 for any h > h1 .
Thus,
|λm − λhm | ≤ Ch2N ∀ h > max{h0 , h1 }

5. Numerical experiments

In the following we numerically evaluate the performance of the piecewise-

constant finite element (PC) and the spectral element method (GLL) on solving
the homogeneous Fredholm integral equation and computing truncated KL ex-
pansions. For GLL, we consider the polynomial degree ranging from N = 1 to
N = 4.

5.1. Numerical convergence

Our first experiments concern the numerical convergence of these methods.
We study the relative error ek = |1 − λhk /λk | in the approximation of the k-th
eigenvalue of (5). Let D =]0, 1[ and K(x, y) be the sinc kernel
sin c(x − y)
K(x, y) = , c > 0. (35)
π(x − y)

11
 Figure 1 (left) shows the relative error of the 10th eigenvalue of the sinc kernel
(with c = 15) for PC and GLL. We evaluated λ10 according to the procedure
presented in [12].
A similar experiment is carried out with the Gaussian covariance kernel

K(x, y) = σ 2 exp(−(x − y)2 /η 2 ), (36)

where the positive constants σ and η represent the variance and the correlation
length, respectively. The relative errors are shown in Figure 1 (right). We
employed the parameters σ = 1 and η = 0.1, and a reference solution was
calculated with the GLL spectral element method with N = 16 and Ne = 212 .
The error curves agree with predicted convergence rates presented in Section 4.

0 0
10 10

−5 −5
10 10
e10

e10

−10 −10
10 10
PC PC
N =1 N =1
N =2 N =2
N =3 N =3
−15 N =4 −15 N =4
10 10
−2 −1 −2 −1
10 10 10 10
h h

Figure 1: Relative GLL error the 10th eigenvalue versus number of elements for the sinc (left)
and the Gaussian (right) covariance kernel.

The accuracy of GLL strongly depends on the regularity of the covariance

kernel K, as seen in Theorem 4.1. In order to illustrate this fact, let us consider
the non-smooth exponential kernel

K(x, y) = σ 2 exp(−|x − y|/η). (37)

The parameters σ and η have the same meaning as in the Gaussian kernel
(36). The exponential kernel (37) is often used on benchmark tests, since closed
formulas are available for its eigenvalues and eigenfunctions (see, e.g., [6, 19, 26]).

12
The exact eigenvalues are
2ησ
λi = , (38)
η 2 γi2 + 1
where the parameters γ1 , γ2 , . . . are roots of the equation (η 2 γ 2 − 1) sin(γ) =
2ηγ cos(γ). Moreover, for D =]0, 1[×]0, 1[, the exact eigenvalues of the separable
exponential covariance function

K(x, y) = σ 2 exp(−|x1 − y1 |/η − |x2 − y2 |/η) (39)

are λ2D
n = λi λj , where 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 2 shows the
relative error of the 10th eigenvalue of the 1D kernel (37) and the 100th eigen-
value of the 2D kernel (39). The error curves of both PC and GLL decay with
order O(h2 ), regardless of the polynomial degree. These results are consistent
with the experiments of Wang [34] on the Nyström method with Simpson’s rule.

0
0
10 10

−1 −1
10 10
e100
e10

−2 −2
10 PC 10 PC
N =1 N =1
N =2 N =2
N =3 N =3
N =4 N =4
−3 −3
10 −2 −1
10 −2 −1
10 10 10 10
h h

Figure 2: Relative GLL error the 10th (100th) eigenvalue versus number of elements for the
1D (left) and 2D (right) exponential covariance kernel.

5.2. Random linear advection

This second experiment is a one-dimensional advection problem presented

by Dorini and Cunha [16]. We seek the scalar field Q : IR × Ω × [0, T ] → IR such

13
that, µ-almost everywhere in Ω,
∂Q ∂Q

 (x; ω; t) + A(ω) (x; ω; t) = 0, x ∈ IR, t > 0,
∂t ∂t

(40)

Q(x; ω; 0) = Q0 (x; ω).


2
The velocity A(ω) is a normal random variable A ∼ N (µA , σA ) with µA =
−0.5 and σA = 0.6. The initial condition Q0 (x; ω) is a Gaussian process defined
by the expectation

 1 , x ∈ (1.4, 2.2),
hQ0 (x; ω)i = (41)
 e−20(x−0.25)2 , otherwise,

and the covariance K(x, y) given by equation (37) with correlation length η =
0.3 and variance σ 2 = 0.16. The velocity and the initial condition are assumed
independent.
Following [16], a reference solution is computed with the Monte Carlo method
using 30000 statistically independent suites of realizations of A and Q0 (x; ω).
Each realization A(ω) and Q0 (x; ω) yields the analytical solution Q(x; ω; t) =
Q0 (x − A(ω)t; ω). Similarly, we consider the approximate solutions
n q
X
Qh (x; ω; t) = Qh0 (x − A(ω)t; ω), Qh0 (x; ω) = hQ0 (x; ω)i + λhk φhk (x)ξk (ω),
k=1

where n = Ne for PC and n = Nv for GLL. We employ a homogeneous grid of

element size h = 1/4 on both methods.
2
Figure 3 represents the variance σQ (x) = hQ′T (x; ω; T )2 i for T = 4 of the
reference solution (MC) along with PC and GLL approximations. Note that PC
accurately approximates the average of the reference solution at each element
and GLL approaches the reference solution as N increases. However, PC is less
0
accurate when we consider the covariance CQ (x) = hQ′ (x; ω; T )Q′ (0; ω; T )i, as
0
shown in Figure 4. The GLL approximation of CQ (x) has a similar accuracy as
2
in the variance σQ (x), indicating the robustness of this method.

5.3. Transient flow in saturated heterogeneous media

In this example, we consider a problem, presented by Zu and Zhang [26], of
transient saturated flow in two-dimensional heterogeneous porous media gov-

14
 0.45 MC 0.45 MC
PC N =1
0.4 0.4
0.35 0.35
σQ

σQ
2

2
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15

−2 −1 0 1 2 3 −2 −1 0 1 2 3
x x

0.45 MC 0.45 MC
N =2 N =4
0.4 0.4
0.35 0.35
σQ

σQ
2

2
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15

−2 −1 0 1 2 3 −2 −1 0 1 2 3
x x

2 (x) = hQ′ (x; ω; T )2 i of the scalar field Q for T = 0.4. Dashed curves:
Figure 3: Variance σQ T
reference solution; solid curves: PC and GLL with N = 1, 2, and 4.

erned by the following continuity equation and Darcy’s law

q(x; ω; t) = −T(x; ω)∇h(x; ω; t), x ∈ D, t > 0,

(42)
∂h
S (x; ω; t) + ∇ · q(x; ω; t) = 0,
∂t
for D =]0, L1 [×]0, L2 [, where L1 = L2 = 10. The transmissivity T is a random
field T(x; ω) = exp(Y (x; ω)), where Y is a Gaussian field, and the storativity S
is constant. The boundary and initial conditions are




 h(x; ω; t) = H1 , x1 = 0, t > 0



 h(x; ω; t) = H2 , x1 = L1 , t > 0


∂h (43)

 (x; ω; t) = 0, x2 = 0, L2 , t > 0



 ∂x2


 h(x; ω; t) = H0 (x), x ∈ D, t = 0.


15
 0.25 0.25

0.2 0.2

0.15 0.15
C Q0

C Q0
0.1 0.1

0.05 0.05
MC MC
0 PC 0 N =1
−0.05 −0.05
−2 −1 0 1 2 3 −2 −1 0 1 2 3
x x

0.25 0.25

0.2 0.2

0.15 0.15
C Q0

C Q0
0.1 0.1

0.05 0.05
MC MC
0 N =2 0 N =4
−0.05 −0.05
−2 −1 0 1 2 3 −2 −1 0 1 2 3
x x

0 (x) = hQ′ (x; ω; T )Q′ (0; ω; T )i of the scalar field Q for T = 0.4.
Figure 4: Covariance CQ
Dashed curves: reference solution; solid curves: PC and GLL with N = 1, 2, and 4.

Initially, the flow is in steady state with constant heads H10 = 9.5 on the
left boundary and H20 = 9.0 on the right boundary. At time t = 0, the constant
heads on the left and right boundaries are suddenly changed to H1 = 11 and
H2 = 10, respectively [26]. The storativity is chosen as S = 0.005. The mean
of the log transmissivity is given as hY (x; ω)i = 0.0, and its covariance kernel is
the separable exponential kernel (39) with the parameters σ 2 = 1.0 and η = 1.0.
The flow domain is uniformly discretized into Ne × Ne square elements,
and the time integration is performed with the Crank-Nicolson method with
∆t = 10−3 . As in [26], the reference solution is computed by Monte Carlo
simulation with 5000 unconditional realizations, in which the log transmissivity
is approximated by a truncated KL expansion with Ne2 terms. We employed
Ne = 40 on the reference solution and Ne = 20 on the GLL and PC approx-
imations. Once the truncated KL expansion is found for either approach, we

16
compute realizations of the hydraulic head by numerically solving (42)-(43) with
piecewise bilinear finite elements.
We exhibit the solutions along the profile x2 = L2 /2 and at the times
t = 0, 0.01, 0.05, 0.4. The reference solution is shown in dashed lines, whereas
PC and GLL solutions are shown in solid lines. Figure 5 shows the transient
head variance σh2 (x1 , t) = hh(x1 , L2 /2; ω; t)i, and Figure 6 illustrates the cross
covariance between the log transmissivity and the hydraulic head,

CY h (x1 ; t) = hY ′ (x1 , L2 /2; ω)h(x1 , L2 /2; ω; t)i.

The performance of the piecewise constant method is similar to the previous

section, in the sense that the accuracy drops when we consider non-diagonal
second-order moments. The spectral element method has again provided robust
approximations, but it was less sensitive to the polynomial degree.

PC N =1
t= 0 t= 0
0.06 t= 0.01 0.06 t= 0.01
t= 0.05 t= 0.05
t= 0.4 t= 0.4
0.04 0.04
σh2

σh2

0.02 0.02

0 0

0 2 4 6 8 10 0 2 4 6 8 10
x1 x1
N =2 N =4
t= 0 t= 0
0.06 t= 0.01 0.06 t= 0.01
t= 0.05 t= 0.05
t= 0.4 t= 0.4
0.04 0.04
σh2

σh2

0.02 0.02

0 0

0 2 4 6 8 10 0 2 4 6 8 10
x1 x1

2 (x , t) = hh(x , L /2; ω; t)i. Dashed curves: reference solution;

Figure 5: Transient variance σh 1 1 2

solid curves: PC and GLL with N = 1, 2, and 4.

17
 PC N =1
0.1 0.1
t= 0 t= 0
t= 0.01 t= 0.01
0.08 t= 0.05 0.08 t= 0.05
t= 0.4 t= 0.4
0.06 0.06
CY h

CY h
0.04 0.04

0.02 0.02

0 0

−0.02 −0.02
0 2 4 6 8 10 0 2 4 6 8 10
x1 x1
N =2 N =4
0.1 0.1
t= 0 t= 0
t= 0.01 t= 0.01
0.08 t= 0.05 0.08 t= 0.05
t= 0.4 t= 0.4
0.06 0.06
CY h

0.04
CY h 0.04

0.02 0.02

0 0

−0.02 −0.02
0 2 4 6 8 10 0 2 4 6 8 10
x1 x1

Figure 6: Transient cross-covariance CY h (x1 ; t) = hY ′ (x1 , L2 /2; ω)h(x1 , L2 /2; ω; t)i. Dashed
curves: reference solution; solid curves: PC and GLL with N = 1, 2, and 4.

6. Conclusions

The Gauss-Lobatto-Legendre (GLL) spectral element method was employed

to numerically solve the homogeneous Fredholm integral equation (5). The
GLL quadrature rendered the matrix M in (10) diagonal, in order that the
resulting discrete eigenvalue problem (22) is analogous to the system (14) ob-
tained from the piecewise constant finite element discretization. As expected
from classical error estimates, spectral elements of degree N provide eigenvalue
approximations of order O(h2N ), assuming that the kernel function K is suffi-
ciently regular (Theorem 4.1). The numerical experiments on Section 5 indicate
that this method is an attractive choice for solving equation (5) and numerically
computing truncated Karhunen-Loéve expansions.
For practical purposes, the accuracy of the proposed method could also be
investigated, both theoretically and experimentally, for non-affine and simplicial

18
elements. Moreover, the knowledge of spectral element methods for determinis-
tic problems [22] could be further explored by providing the discretization space
for not only the eigenfunctions of the auxiliary Fredholm integral equation, but
also the unknowns of initial and boundary value problems with random input
data.

Acknowledgements

This research is supported by Brazilian agency CNPq under grants 314553/2009-

6 and 471182/2011-7.

References

[1] C. Alouch, P. Sablonnière, D. Sbibih, and M. Tahrichi. Superconvergent

Nyström and degenerate kernel methods for eigenvalue problems. Appl.
Math. Comput., 217(20):7851–7866, 2011.

[2] 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.

[3] K. Atkinson. Convergence rates for approximate eigenvalues of compact

integral operators. SIAM J. Numer. Anal., 12(2):213–222, 1975.

[4] K. Atkinson. The numerical solution of integral equations of the second

kind. Cambridge University Press, Cambridge, 1997.

[5] J. S. Azevedo, M. A. Murad, M. R. Borges, and S. P. Oliveira. A space–time

multiscale method for computing statistical moments in strongly heteroge-
neous poroelastic media of evolving scales. Int. J. Numer. Meth. Engrg.,
90(6):671–706, 2012.

[6] J. S. Azevedo and S. P. Oliveira. A numerical comparison between quasi-

Monte Carlo and sparse grid stochastic collocation methods. Commun.
Comput. Phys., 12:1051–1069, 2012.

19
 [7] A. Baggeroer. A state-variable approach to the solution of Fredholm inte-
gral equations. IEEE Trans. Inform. Theory, 15(5):557–570, 1969.

[8] 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.

[9] H. Brakhage. Zur Fehlerabschätzung für die numerische Eigenwertbestim-

mung bei Integralgleichungen. Numer. Math., 3:174–179, 1961.

[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.

[13] C. Canuto, M. Hussaini, A. Quarteroni, and T. Zang. Spectral methods:

Fundamentals in Single Domains. Springer-Verlag, New York, 2006.

[14] P. Ciarlet. The Finite Element Method for Elliptic Problems. SIAM,
Philadelphia, 2002.

[15] F. R. DiNapoli and F. H. Middleton. Surface-wave propagation in a con-

tinuously stratified medium. J. Acoust. Soc. Am., 39(5A):899–903, 1966.

[16] F. A. Dorini and M. C. C. Cunha. Statistical moments of the random linear

transport equation. J. Comput. Phys., 227(19):8541–8550, 2008.

[17] M. Eiermann, O. Ernst, and E. Ullmann. Computational aspects of the

stochastic finite element method. Comput. Vis. Sci., 10(1):3–15, 2007.

20
[18] 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.

[19] R. Ghanem and P. Spanos. Stochastic Finite Element: A Spectral Aproach.

Springer, New York, 1991.

[20] M. Girolami. Orthogonal series density estimation and the kernel eigenvalue
problem. Neural Comput., 14(3):669–688, 2002.

[21] I. G. Graham, F. Y. Kuo, D. Nuyens, R. Scheichl, and I. H. Sloan. Quasi-

Monte Carlo methods for elliptic PDEs with random coefficients and ap-
plications. J. Comput. Phys., 230(10):3668–3694, 2011.

[22] G. Karniadakis and S. Sherwin. Spectral/hp element methods for compu-

tational fluid dynamics. 2nd ed. Oxford University Press, Oxford, 2005.

[23] H. B. Keller. On the accuracy of finite difference approximations to the

eigenvalues of differential and integral operators. Numer. Math., 7:412–419,
1965.

[24] K. Khare. Sampling theorem, bandlimited integral kernels and inverse

problems. Inverse Probl., 23(4):1395–1416, 2007.

[25] M. Loève. Probability Theory II, 4th ed. Spring-Verlag, New York, 1978.

[26] Z. Lu and D. Zhang. Analytical solutions to statistical moments for tran-

sient flow in two-dimensional, bounded, randomly heterogeneous media.
Water Resour. Res., 41:W01016, 2005.

[27] S. P. Oliveira and G. Seriani. Effect of element distortion on the numer-

ical dispersion of spectral element methods. Commun. Comput. Phys.,
9(4):937–958, 2011.

[28] K. K. Phoon, S. P. Huang, and S. T. Quek. Implementation of Karhunen-

Loève expansion for simulation using a wavelet-Galerkin scheme. Prob.
Eng. Mech., 17(3):293–303, 2002.

21
[29] C. Schwab and R. Todor. Karhunen-Loeve approximation of random fields
by generalized fast multipole methods. J. Comput. Phys., 217(1):100–122,
2006.

[30] I. Sloan. Iterated Galerkin method for eigenvalue problems. SIAM J.

Numer. Anal., 13(5):753–760, 1976.

[31] A. Spence. Error bounds and estimates for eigenvalues of integral equations.
Numer. Math., 29:133–147, 1978.

[32] 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.

[33] R. Todor. Robust eigenvalue computation for smoothing operators. SIAM

J. Numer. Anal., 44(2):865–878, 2006.

[34] J. Wang. On the numerical computation of eigenvalues and eigenfunctions

of compact integral operators using spline functions. J. Inst. Math. Appl.,
18(2):177–188, 1976.

[35] 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.

22