AN EXTENDED ERROR ANALYSIS FOR A MESHFREE

DISCRETIZATION METHOD OF DARCY’S PROBLEM

DANIELA SCHRÄDER∗ AND HOLGER WENDLAND†

Abstract. Recently ([21]), a new meshfree approximation method for Darcy’s problem has been
introduced and analyzed. This method is based on a symmetric collocation approach using radial
basis functions producing solutions with an analytically divergence-free velocity part. However, the
error analysis provided in [21] works only for smooth solutions, where the smoothness is intrinsically
linked to the smoothness of the employed basis function. In this paper, we will extend the error
analysis to less smooth functions, showing that the approximation order for rougher solutions is
determined rather by the smoothness of the solution than the smoothness of the basis function.

Key words. partial differential equations, radial basis functions, high-order method, collocation,
error analysis

AMS subject classifications. 65N15, 65N35

1. Introduction. Darcy’s problem plays an important role in porous media flow

[3, 10]. It can be stated in the following way

u + K∇p = f in Ω, (1.1)
∇·u =0 in Ω, (1.2)
u·n =g·n on ∂Ω. (1.3)

Here, Ω ⊆ Rd is a bounded domain with boundary ∂Ω having a unique outer

normal vector n. The right-hand sides f and g · n and the permeability tensor K are
given. The boundary function g must satisfy the compatibility condition
Z
g · n dS = 0. (1.4)
∂Ω

The tensor K is supposed to be symmetric, K = K T , and strongly elliptic in the

sense that there is a constant α > 0 such that

ξT K(x)ξ ≥ αkξk22 , ξ ∈ Rd , x ∈ Ω. (1.5)

The solution consists of a velocity term u : Ω → Rd and a pressure term p : Ω → R.

Recently, a new discretization scheme for Darcy’s problem has been developed in
[21]. This scheme is based upon symmetric collocation employing matrix-valued “ra-
dial” basis functions. It produces analytically divergence-free approximations to the
velocity field and, as a meshfree method, is flexible about the shape of the underlying
domain.
The error analysis given in [21] applies, unfortunately, only if the solution (u, p)
comes from a specific Hilbert space of smooth functions, which is intrinsically con-
nected to the employed basis function. The Hilbert space is the reproducing kernel
Hilbert space associated to the basis function. This is in principle not a problem,
as long as the smoothness of the solution is known before-hand. Then, the strategy
∗ German Aerospace Center (DLR), Institute of Aerodynamics and Flow Technology, Bunsen-

strasse 10, D-37073 Göttingen, Germany ([email protected])

† University of Oxford, Mathematical Institute, 24-29 St Giles’, Oxford, OX1 3LB, United King-

dom ([email protected])
1
would be to choose the basis function accordingly to this smoothness such that the
error analysis of [21] applies. However, in the case that the smoothness of the solution
is unknown or can only be estimated, it is very likely that the basis of the discretiza-
tion space is chosen too smoothly. Hence, the natural question that arises is what
kind of error estimates remains valid if the solution (u, p) is less smooth than required
by the theory in [21].
In this paper, we will show that in such a situation, an error analysis still holds
and gives the expected order, determined by the smoothness of the solution rather
than the smoothness of the underlying discretization spaces. It turns out that as
natural this result is, its proof is rather complicated and requires deep techniques
from multivariate approximation theory.
So far, in the context of meshfree methods based on radial basis functions, such
results were only known for classical interpolation and approximation but not for
collocation methods for partial differential equations. Error estimates for “rougher”
target functions were first presented by Narcowich and Ward in [16] in the special
situation that the data sites are located on a sphere. Other work on Rd followed in
[5, 12, 13, 18, 17]. An overview can be found in [14]. Recently Fuselier presented
error estimates for interpolation problems with divergence-free or curl-free matrix-
valued kernels, where the target function is rougher than required by the classical
reproducing kernel Hilbert space theory, see [8].
In this paper, we will show that this concept can be extended to collocation
methods for solving partial differential equations. We will establish new Sobolev-
type approximation rates for the discretization scheme of Darcy’s problem, where the
velocity and the pressure are too rough for the error analysis provided in [21].
For this purpose, the paper is organized as follows. In the rest of this section
we collect necessary notation on vector-valued Sobolev spaces. The next section is
devoted to our discretization scheme, hence covering matrix-valued kernels, their re-
producing kernel Hilbert spaces and optimal recovery as well as stating the approxi-
mation scheme. In the third section, we introduce technical tools required for our error
analysis, comprising the interpolation and approximation properties of band-limited
functions. After this we derive our main result, error estimates for ‘rougher’ solutions
to Darcy’s problem. In the final section, we give numerical examples to corroborate
our theoretical estimates.

1.1. Sobolev Spaces. We will work with the usual scalar-valued Sobolev spaces.
For a domain Ω ⊆ Rd , a real number r ≥ 1 or r = ∞ and an integer k ∈ N0 , we denote
by Wrk (Ω) the space of all functions f ∈ Lr (Ω) having weak derivatives Dα f ∈ Lr (Ω)
for every multi-index α ∈ Nd0 with |α| = α1 + · · · + αd ≤ k. We will also work with
fractional order Sobolev spaces Wrτ (Ω), particularly with τ > d/2 so that we have
continuous functions. For the introduction of such fractional order Sobolev spaces we
refer, for example, to [1, 4, 23].
Since the pressure p in the solution of (1.1)–(1.3) is determined only up to a
constant we will work with the quotient spaces Wrτ (Ω)/R equipped with the norm

kpkWrτ (Ω)/R := inf kp + ckWrτ (Ω) .

c∈R

We define the vector-valued Sobolev space Wrτ (Ω) to consist of all vector-valued
functions u = (u1 , . . . , un )T : Ω → Rn , where each component uj belongs to Wrτ (Ω).
A norm on Wrτ (Ω) can be defined by taking the discrete ℓr norm of the Wrτ (Ω) norms
2
of the components, i.e. by
 1/r
 Pn
j=1 kuj krW τ (Ω) for 1 ≤ r < ∞,
kukWrτ (Ω) = r
max kuj kW∞
τ (Ω) for r = ∞.
1≤j≤n

Note that we do not use an index to indicate the dimension n since it will become
clear from the context. We only distinguish between scalar-valued function spaces
and vector-valued ones. Finally, in the case r = 2, we will also use the notation
Hτ (Ω) = W2τ (Ω).
2. The Discretization Scheme. In this section, we will review the necessary
material on matrix-valued kernels and the way we will use them for discretizing
Darcy’s problem (1.1)–(1.3).
First, we will discuss the kernels and their reproducing Hilbert spaces. For this,
we will mainly rely on material from [6, 7, 8, 25, 21]. In the last part of this section,
we will state the concrete discretization scheme from [21].
2.1. Positive Definite Matrix-valued Kernels. Definition 2.1. A function
φ : Rd → R is positive definite if for all N ∈ N, all pairwise distinct x1 , . . . , xN ∈ Rd
and all α ∈ RN \ {0}, the quadratic form
N
X
αj αk φ(xj − xk )
j,k=1

is positive. More generally, a matrix-valued function Φ : Rd → Rn×n is positive

definite if it is even Φ(−x) = Φ(x), symmetric Φ(x) = Φ(x)T and satisfies
N
X
αTj Φ(xj − xk )αk > 0
j,k=1

for all pairwise distinct xj ∈ Rd and all αj ∈ Rn such that not all αj are vanishing.
The theory of the associated function spaces can be formulated for positive definite
matrix-valued functions as it can be done for scalar-valued ones. Let Ω ⊆ Rd be non-
empty. The following definition is taken from [8].
Definition 2.2. Let H be a Hilbert space of vector-valued functions f : Ω → Rn .
The space H is called a reproducing kernel Hilbert space if there exists a continuous
n × n matrix-valued kernel Φ such that
(1) Φ(· − x)α ∈ H,
(2) αT f (x) = (f , Φ(· − x)α)H
for all x ∈ Ω and α ∈ Rn . The function Φ is called the reproducing kernel of H.
It is well known that the reproducing kernel of a reproducing kernel Hilbert space
is a positive definite kernel and that every positive definite kernel generates a Hilbert
space in a natural way, in which it is the reproducing kernel.
Here, we are interested in the following two reproducing kernel Hilbert spaces (see
[15, 6, 7, 25, 21]).
Theorem 2.3.
(1) Suppose φ ∈ W12 (Rd ) ∩ C 2 (Rd ) is positive definite. Let the matrix-valued
kernel Φdiv : Rd → Rd×d be defined by Φdiv := (−∆I + ∇∇T )φ, where ∆ is
the usual Laplace operator, ∇ denotes the gradient and I the identity matrix.
3
 Then the associated Hilbert space HΦdiv of Φdiv consists of all functions f ∈
L2 (Rd ) with
Z
kb
f (ω)k22
kf k2HΦ (Rd ) = (2π)−d/2 dω < ∞.
div
Rd b
kωk2 φ(ω)
2

(2) Let the e : Rd → R(d+1)×(d+1) be defined by Φe = Φdiv ⊗

 matrix-valued
 kernel Φ
Φdiv 0
ψ := with a positive definite function ψ. Then, the corresponding
0 ψ
reproducing kernel Hilbert space is given by
d d d
HΦ
e (R ) = HΦdiv (R ) × Hψ (R )

with norm for f = (fu , fp ) given by

kf k2H e (Rd ) = kfu k2HΦ + kfp k2Hψ (Rd )

(Rd )
Φ div
Z " #
kbfu (ω)k 2
|fbp (ω)|2
2
= (2π)−d/2 + dω.
2b b
Rd kωk2 φ(ω) ψ(ω)

We are particularly interested in reproducing kernel Hilbert spaces that are norm
equivalent to Sobolev spaces. A scalar-valued RKHS Hφ (Rd ) is norm equivalent to
the Sobolev space H τ (Rd ) if the kernel function φ : Rd → R has a Fourier transform
φb satisfying
b
c1 (1 + kωk22 )−τ ≤ φ(ω) ≤ c2 (1 + kωk22 )−τ , ω ∈ Rd ,

with two constants 0 < c1 ≤ c2 , see [24, Corollary 10.13]. This gives for the matrix-
valued kernels the following result.
Corollary 2.4. Assume φ generates H τ +1 (Rd ) and ψ generates H ρ (Rd ), i.e.,
Hφ (Rd ) = H τ +1 (Rd ) and Hψ (Rd ) = H ρ (Rd ). The associated reproducing kernel
Hilbert space of the combined kernel is given by

HΦ d eτ d ρ d
e (R ) = H (R ; div) × H (R ).

Here,
( Z )
e τ (Rd ; div) = kb
f (ω)k22
H f ∈ Hτ (Rd ; div) : (1 + kωk22 )τ +1 dω < ∞ ,
Rd kωk22

Hτ (Rd ; div) = f ∈ Hτ (Rd ) : ∇ · f = 0 .

In addition to the two Sobolev-like spaces above, we are interested in the subspace
of curl-free functions of
( Z )
kbf (ω)k 2
e (R ) = f ∈ L2 (R ) :
τ d d 2 2 τ +1
H 2 (1 + kωk2 ) <∞ .
Rd kωk2

A function f is curl-free if and only if we can find a function g such that f = ∇g.
Hence
n o
He τ (Rd , curl) = f ∈ H
e τ (Rd ) : There exists g ∈ H τ +1 (Rd )/R such that ∇g = f .
4
 e τ (Rd ), H
The norm in the spaces H e τ (Rd , div) and H
e τ (Rd , curl) will be denoted by
Z
−d/2 kb
f(ω)k22
kf kH
e τ (Rd ) = (2π) (1 + kωk22 )dω.
Rd kωk22

Finally, since we mainly work on bounded domains, we need for technical reasons
to extend our locally defined Sobolev functions to globally defined ones. We will use
the following result from [25].
Proposition 2.5. Let d = 2, 3. Let τ, σ ≥ 0 and let Ω ⊆ Rd be a simply-
connected domain with C k,1 boundary, where k ≥ τ is an integer. Then there exists a
continuous operator E = (E e div , ES ) : Hτ (Ω; div) × H σ (Ω) → H
e τ (Rd ; div) × H σ (Rd )
τ σ
such that Ev|Ω = v|Ω for all v = (u, p) ∈ H (Ω) × H (Ω) and


e div uk e τ d + kES pkH σ (Rd ) ≤ c kukHτ (Ω) + kpkH σ (Ω) .
kE H (R )

The extension operator for the pressure part is the standard Stein extension operator
ES , see [22].
2.2. The Discretization. In [21], the following discretization scheme for solving
Darcy’s problem numerically has been introduced. It is based upon discretization
points X = {x1 , . . . , xN } ⊆ Ω in the interior and Y = {y1 , . . . , yM } ⊆ ∂Ω on the
boundary. Associated with these discretization points are functionals acting on the
combined function v = (u, p). They are defined as
(i)
λj (v) = ui (xj ) + (K∇p)i (xj ) (2.1)
d
X
= ui (xj ) + Kik (xj )∂k p(xj ), 1 ≤ i ≤ d, 1 ≤ j ≤ N =: Ni (2.2)
k=1
d
X
(d+1)
λj (v) = uk (yj )nk (yj ), 1 ≤ j ≤ M =: Nd+1 , (2.3)
k=1

and represent on the one hand the partial differential equation and on the other hand
the boundary conditions point-wise.
The approximate solution is then given as a linear combination of the Riesz rep-
resenters of these functionals. Since it is known (see Lemma 3.4 below) that the Riesz
representer of a functional is given by the function that results if the functional is
applied to one argument of the reproducing kernel, our approximate solution takes
the form
Nk
d+1 X
X (k) (k)
e − y)
sv (x) := αj (λj )y Φ(x (2.4)
k=1 j=1

and the coefficients are determined by enforcing the collocation conditions



fk (xi ) for 1 ≤ k ≤ d, 1 ≤ i ≤ N =: Nk
(k) (k)
λi (sv ) = λi (v) = (2.5)


g · n(xi ) for k = d + 1, 1 ≤ i ≤ M =: Nd+1

The following result ensures a unique solution to this linear system. It is taken from
[21, Theorem 2.7]
5
 Theorem 2.6. Let Ω ⊆ Rd with a C 1,1 boundary. Assume that the build-
ing functions φ, ψ : Rd → R are positive definite and chosen such that HΦ d
e (R ) =
e τ d
H (R ; div)×H τ +1
(R ) with τ > d/2. Then, the interpolation function sv = (su , sp )T
d

from (2.4) is well-defined and uniquely determined by the interpolation conditions

(2.5). It satisfies therefore Lsv (xj ) = f (xj ) with Lv := u + K∇p and su (yj ) · n(yj ) =
g(yj )·n(yj ). Furthermore, the approximate solution su is analytically divergence free,
i.e. ∇ · su = 0 in Rd .
3. Band-limited Interpolation and Approximation. In this section, we in-
troduce band-limited functions and establish some results regarding the interpolation
and the approximation with band-limited functions. These results will be essential for
the extended error analysis of the collocation scheme introduced in the last section.
A band-limited function is a function f ∈ L2 (Rd ) with a compactly supported
Fourier transform fb. For σ > 0 we denote the d-variate ball with centre 0 and radius
σ > 0 by B(0, σ). Then, the set of all band-limited functions with band-width σ will
be denoted by
n o
B σ := f ∈ L2 (Rd ) : supp fb ⊆ B(0, σ) .

The concept of band-limited functions can be extended to vector-valued functions.

We are interested in two different kinds of vector-valued functions: Divergence-free
and curl-free band-limited functions. We define the following spaces:
n o
B σ := f ∈ L2 (Rd ) : supp b f ⊆ B(0, σ) ,
( Z )
kbf (ω)k 2
Be := f ∈ B :
σ σ 2
2 dω < ∞ ,
Rd kωk2
n o
Bediv
σ
:= f ∈ Beσ : ωT b
f (ω) = 0 .

In the first case, the requirement supp bf ⊆ B(0, σ) is meant component-wise. Since
band-limited functions have a compactly supported Fourier transform, they belong to
the reproducing kernel Hilbert spaces of all relevant (radial) basis functions.
We will use band-limited function to approximate a given function, which does
not belong to the reproducing kernel Hilbert space and then approximate the band-
limited function using our collocation scheme to derive error estimates for functions
outside the reproducing kernel Hilbert space.
We will use a general concept to derive approximation results for band-limited
functions, which can be stated in a quite general form. Its proof can be found in
[17]. It shows the existence of an abstract interpolant, which is comparable to a best
approximation. To state it, we use the distance between an element y of a linear space
Y and a subspace V of Y, which is defined by

distY (y, V) := min ky − vkY .

v∈V

Proposition 3.1. Let Y be a (possibly complex) Banach space, V be a subspace

of Y, and Z ∗ be a finite dimensional subspace of Y ∗ , the dual of Y. If for every
λ∗ ∈ Z ∗ and some β > 1, β independent of λ∗ ,

kλ∗ kY ∗ ≤ βkλ∗ |V kV ∗ ,
6
then for y ∈ Y there exists v ∈ V such that v interpolates y on Z ∗ ; that is, λ∗ (y) =
λ∗ (v) for all λ∗ ∈ Z ∗ . In addition, v approximates y in the sense that

ky − vkY ≤ (1 + 2β) distY (y, V).

The following lemma was proven by Fuselier in [8, Lemma 1]. It shows that every
f ∈H e τ (Rd ; div) can be approximated by a band-limited function fσ ∈ Beσ . We give
div
a slightly extended version, since we need it for functions in H e τ (Rd ; div) × H ρ (Rd ),
its proof can be done analogously.
Lemma 3.2. Let τ ≥ β ≥ 0 and σ > 0. For every f ∈ H e τ (Rd ; div) there exists a
eσ
function gσ ∈ Bdiv with
β−τ
kf − gσ kH
e β (Rd ) ≤ σ kf kH
e τ (Rd ) .

Moreover, for every f ∈ H τ (Rd ) there exists a function gσ ∈ B σ with

kf − gσ kH β (Rd ) ≤ σ β−τ kf kH τ (Rd ) .

Obviously, the first statement of the previous lemma would also hold for curl-free
functions.
e τ (Rd ) and its subspaces H
The Sobolev-like space H e τ (Rd , div) and H
e τ (Rd , curl)
are reproducing kernel Hilbert spaces with canonical reproducing kernels given by
Ke τ := −∆Kτ +1 , K
e τ := (−∆I + ∇∇T )Kτ +1 and K e τ := −∇∇T Kτ +1 respectively,
div curl
where K is the canonical reproducing kernel of H (Rd ) defined by its Fourier trans-
s s

form
cs (ω) := (1 + kωk2 )−s .
K 2

The following lemma is essentially lemma 2 in [8]. The case of the divergence-
free functions was proven there, the case of the curl-free functions can be done in an
analogous way. A detailed proof of this is given in [20].
Lemma 3.3. Let qX = 12 minj6=k kxj −xk k2 be the separation radius of the discrete
PN e τ PN e τ
set X = {x1 , . . . , xN }. Let g = j=1 K div (·−xj )αj or g = j=1 Kcurl (·−xj )αj , τ >
d/2, respectively and define gσ by gbσ = g
bχσ , where χσ is the characteristic function of

1/(d+1)
the ball B(0, σ). Then, there exists a constant κ ≥ 24 π(d+2)(d+3)d
4(d−1) Γ2 d+2 2 =:
e which is independent of X and the αj ’s, such that for σ = κ/qX the following
C,
inequality holds:
1
Iσ := kg − gσ kH
e τ (Rd ) ≤ kgkH
e τ (Rd ) .
2

The following theorem is an extension of [24, Theorem 16.7] for vector-valued

reproducing kernel Hilbert spaces. Its proof is exactly the same as the proof in the
scalar-valued case. It is the already mentioned fact on Riesz representers.
Lemma 3.4. Suppose that H is a real, vector-valued Hilbert space of functions with
reproducing matrix-valued kernel Φ : Rd → Rn×n and dual H∗ . For every λ ∈ H∗ ,
we have that λy (Φ(· − y)α) ∈ H and

λ(f ) = (f , λy (Φ(· − y)α))H (3.1)

7
for all f ∈ H and all α ∈ Rn . Moreover,

kλkH∗ = kλy (Φ(· − y)α)kH . (3.2)

We now state and prove the main result of this section, which is central for the
proof of the extended error estimates later on. It guarantees the existence of a band-
limited function, which approximates the true solution of Darcy’s problem and also
gives a bound for the error.
Theorem 3.5. Let τ , ρ, t, r ∈ R with τ > d/2, ρ > d/2 + 1 and t, r ≥ 0. Let
Ω ⊆ Rd be bounded with a C ⌈τ ⌉+1,1 boundary and outer unit normal vector n. Let
X = {x1 , . . . , xN } ⊆ Ω and Y = {y1 , . . . , yM } ⊆ ∂Ω be discrete sets with separation
radius q := qX∪Y . For a smooth, combined function v = (u, p) we define the operator
Lv := u + K∇p, where Kij ∈ H ρ (Rd ) for all 1 ≤ i, j ≤ d.
For a given v = (u, p) ∈ H e τ (Rd ; div) × H ρ (Rd ) there exists a function vσ =
(uσu , pσp ) ∈ Bediv × B p such that
σu σ

Lv|X = Lvσ |X , u · n|Y = uσ · n|Y

and
 1/2
−2t 2 −2r
kv − vσ kH
e τ (Rd )×H ρ (Rd ) ≤ 5 σu kukH
e τ +t (Rd ) + σp kpk2H ρ+r (Rd ) ,

e e is the constant from lemma 3.3.

whenever σu , σp > Cq , where C
Proof. To prove this result, we will use proposition 3.1 with
e τ (Rd ; div) × H ρ (Rd ),
Y := H V := Bediv
σu
× B σp

and Z ∗ := span{ZX
∗
∪ ZY∗ } with
∗
ZX := {λ(v) = ui (x) + (K∇p)i (x) : x ∈ X, v = (u, p) ∈ Y, 1 ≤ i ≤ d} ,

ZY := λ(v) = n(x)T u(x) : x ∈ Y, v = (u, p) ∈ Y .
∗

Obviously, Y is a Banach space, V is a subspace of Y, Z ∗ is finite dimensional and

every λ ∈ Z ∗ is linear. Since Y is a reproducing kernel Hilbert space, the point
evaluation functionals are in Y ∗ . Furthermore, the Sobolev embedding theorem, our
assumption on K and the boundary of Ω guarantee that the functions u, p, K and n
are sufficiently smooth. Thus, all functionals λ ∈ Z ∗ are indeed continuous on Y.
We will now show that for every λ ∈ Z ∗ we have

kλkY ∗ ≤ 2kλ|V kV ∗ . (3.3)

First of all, we will calculate the Riesz representer and express the norms of the dual
space in terms of the original space. Then we can bound kλkY ∗ and show that (3.3)
holds.
Let xN +j := yj for all 1 ≤ j ≤ M . Let f = (fu , fp ) ∈ Y. We pick an arbitrary
element λ ∈ Z ∗ , which can be written as
N
X NX
+M
λ(f ) = αTj [fu (xj ) + K(xj )∇fp (xj )] + αj n(xj )T fu (xj ),
j=1 j=N +1
8
for all f = (fu , fp ) ∈ Y, where αj ∈ Rd for 1 ≤ j ≤ N and αj ∈ R for N < j ≤ N +M .
If we define
(
αj , if 1 ≤ j ≤ N
γ j :=
αj n(xj ), if N < j ≤ N + M

and ζ Tj := αTj K(xj ) then we can write the functional as λ(f ) := λu (fu ) + λp (fp ),
PN +M PN
where λu (fu ) := j=1 γ Tj fu (xj ) and λp (fp ) := j=1 ζ Tj ∇fp (xj ).
The reproducing kernel of the space H e τ (Rd ; div) × H ρ (Rd ), equipped with the
inner product

(f , g)H
e τ (Rd )×H ρ (Rd ) = (fu , gu )H
e τ (Rd ) + (fp , gp )H ρ (Rd ) ,

is given by
 
Ke τ (x − y) 0
KH div
e τ (Rd ;div)×H ρ (Rd ) (x − y) := .
0 Kρ (x − y)

This means that we can work out the Riesz representer gu and gp for the functionals
λu ∈ He τ (Rd ; div) and λp ∈ H ρ (Rd ) separately to derive the Riesz representer gλ for
the functional λ.
In the first case, Lemma 3.4 gives

NX
+M
gu = e τ (· − xj )γ ,
K div j
j=1
NX
+M
kλu k2H e div
γ Tj K τ
e τ (Rd ;div)∗ = (xj − xk )γ k .
j,k=1

In the second case, the same lemma shows that the representer for λp is given by
P
gp = N T ρ T
j=1 ζ j ∇K (· − xj ). Since ζ k ∇p is a scalar, we have

N
X  T
kλp k2H ρ (Rd )∗ = kgp k2H ρ (Rd ) = ζ Tj ∇x ζ Tk ∇y Kρ (xj − xk )
j,k=1
N
X
= ζ Tj ∇x ∇Ty Kρ (xj − xk )ζ k .
j,k=1

Altogether we have derived that

NX
+M N
X
kλk2Y ∗ = kgλ k2Y = e τ (xj − xk )γ k +
γ Tj K ζ Tj ∇x ∇Ty Kρ (xj − xk )ζ k .
div
j,k=1 j,k=1

The next step is to show that kλ|V kV ∗ = kgσ kY , where gσ = (gu,σu , gp,σp ) is the
Riesz representer from the band-limited space. Since V is a subspace of Y the norms
are the same for every element in V. Again we first deal with λu . Let f ∈ Bediv σu
and
9
gu,σu be defined by g[ cu χσu . This gives
u,σu = g

Z
−d/2 cu (ω)∗b
g f (ω)
λu (f ) = (f , gu )H
e τ (Rd ) = (2π) (1 + kωk22 )τ +1 dω
Rd kωk22
Z
−d/2 gcu (ω)∗bf (ω)
= (2π) 2 (1 + kωk22 )τ +1 dω
2
kωk2 ≤σu kωk 2
Z ∗b
−d/2 g[u,σu (ω) f (ω)
= (2π) 2 (1 + kωk22 )τ +1 dω
kωk22 ≤σu kωk2
Z ∗b
−d/2 g[ u,σu (ω) f (ω)
= (2π) 2 (1 + kωk22 )τ +1 dω
Rd kωk2
= (f , gu,σu )H
e τ (Rd ) ,

where we used the fact that bf vanishes outside the ball B(0, σu ). This equality and
the same idea as in the proof of lemma 3.4 lead us to

kλu |Beσu k(Beσu )∗ = kλu |Beσu kH

e τ (Rd ;div)∗ = kgu,σu kH
e τ (Rd ) .
div div div

Using similar arguments and gp,σp defined by g[

p,σp = gbp χσp allows us to derive

(f, gp )H ρ (Rd ) = (f, gp,σp )H ρ (Rd )

and

kλp |Bσp kBσp ∗ = kλp |Bσp kH ρ (Rd )∗ = kgp,σp kH ρ (Rd ) .

This all together means that

kλ|V k2V ∗ = kλ|V k2Y ∗ = kλu |Beσu k2H 2

e τ (Rd )∗ + kλp |B p kH ρ (Rd )∗
σ
div

= kgu,σu k2H 2 2
e τ (Rd ) + kgp,σp kH ρ (Rd ) = kgλ,σ kY .

Later we want to apply lemma 3.3 to bound kgp − gp,σp kH ρ (Rd ) by kgp kH ρ (Rd ) . Before
e ρ−1 (Rd ) and kgp − gp,σp kH ρ (Rd ) =
doing so, we need to show that kgp kH ρ (Rd ) = kgcurl kH
PN e ρ−1
kgcurl − gcurl,σp k e ρ−1 d
H , where gcurl =
(R ;curl) K (· − xj )ζ and gp,σp , gcurl,σp
j=1 curl j
p,σp = gbp χσp and g\
are defined by g[ curl,σp = gd
curl χσp respectively. We have that
e ρ−1
Kcurl = −∇∇ K , where K is the reproducing kernel of H ρ (Rd ). Therefore
T ρ ρ

N
X N
X
gcurl = e ρ−1 (· − xj )ζ j = −
K ∇∇T Kρ (· − xj )ζ j = −∇gp .
curl
j=1 j=1

With the identity above, we can derive

Z 2
k(gdcurl − g\
curl,σp )(ω)k2
kgcurl − gcurl,σp k2H
e ρ−1 (Rd ) = (2π)
−d/2
(1 + kωk22 )ρ dω
Rd kωk22
Z \p (ω)k22
−d/2 k−∇g
= (2π) (1 + kωk22 )ρ dω
kωk2 ≥σp kωk22
= kgp − gp,σp k2H ρ (Rd ) . (3.4)
10
and
kgcurlkH
e ρ−1 (Rd ) = kgp kH ρ (Rd ) . (3.5)
We add the term g − g and apply the inverse triangle inequality to establish
kgσ kH
e τ (Rd )×H ρ (Rd ) = kg − (g − gσ )kH
e τ (Rd )×H ρ (Rd )

≥ kgkH
e τ (Rd )×H ρ (Rd ) − kg − gσ kH
e τ (Rd )×H ρ (Rd )

To bound the norm above, we apply (3.4) and (3.5) together with lemma 3.3,
 1/2
2 2
kg − gσ kHe τ (Rd )×H ρ (Rd ) = kg u − g k
u,σu H e τ (Rd ) + kg curl − g curl,σp Hk e ρ−1 (Rd )
 1/2
1 1
≤ kgu k2He τ (Rd ) + kgcurl kH
2
e ρ−1 (Rd )
4 4
1
= kgkH e τ (Rd )×H ρ (Rd ) .
2
Thus we have
1
kλ|V kV ∗ = kgσ kH e τ (Rd )×H ρ (Rd ) ≥ kgkH e τ (Rd )×H ρ (Rd ) − kgkH e τ (Rd )×H ρ (Rd )
2
1 1
= kgkH e τ (Rd )×H ρ (Rd ) = kλkY ∗ .
2 2
Hence we have proven that kλkY ∗ ≤ 2kλ|V kV ∗ for every λ ∈ Z ∗ , i.e., all assumptions
of theorem 3.1 are satisfied with β = 2. Thus for every v ∈ H e τ (Rd ; div) × H ρ (Rd )
e σu σp
there exists a vσ = (uσu , pσp ) ∈ Bdiv × B such that vσ interpolates v on Z ∗ , i.e.,
λ(v) = λ(vσ ) for all λ ∈ Z ∗ . In addition, vσ approximates v in the sense that
kv − vσ k2Y = ku − uσu k2H 2 2 eσu σp 2
e τ (Rd ) + kp − pσp kH ρ (Rd ) ≤ 5 distY (v, Bdiv × B ) .

Next, we will bound the distance between v and Bediv σu

× B σp . We will do this again
component-wise, since the definition of the distance gives that
n o
distY (v, Bediv
σu
× B σp )2 = min ku − uσu k2H
e τ (Rd ) + kp − p σ k 2
ρ d
p H (R )
vσ ∈V

= min ku − uσu k2H

e τ (Rd ) + min kp − pσp k2H ρ (Rd )
eσu
u σu ∈ B pσp ∈Bσp
div

= distH(R eσu 2 σp 2
e d ) (u, Bdiv ) + distH ρ (Rd ) (p, B ) .

Note that, to simplify notation, we have again used the notation uσu and pσp , though,
this time, they denoted arbitrary elements of the respective spaces of band-limited
functions. We proceed now by bounding both terms separately. We have by lemma
3.2 that
distH(R eσu 2 2(τ −(τ +t))
e d ) (u, Bdiv ) ≤ σu
−2t
e τ +t (Rd ) . ≤ σu kukH
kukH e τ +t (Rd ) .

and
distH ρ (Rd ) (p, B σp )2 ≤ σp2(ρ−(ρ+r)) kpkH ρ+r (Rd ) ≤ σp−2r kpkH ρ+r (Rd )
and thus
distY (v, Bediv
σu
× B σp )2 ≤ σu−2t kuk2H −2r
e τ +t (Rd ) + σp kpk2H ρ+r (Rd ) ,
which finishes the proof.
11
 4. Error Analysis. Our error analysis is mainly based on a ’shift’ type theorem
for the analytical solution of Darcy’s problem. It is taken from [21, Theorem 3.2].
Proposition 4.1. Let Ω be a bounded open subset of Rd with a C ⌈τ ⌉+1,1 boundary
τ +1−1/r
∂Ω. AssumeR that the data f ∈ Wrτ +1 (Ω) and g ∈ Wr (∂Ω) for 1 < r < ∞ and
satisfies ∂Ω g · n dS = 0. Assume that the permeability tensor K = (Kij ) satisfies
(1.5), K = K T and Kij ∈ Wrτ +1 (Ω). Then there exist a velocity u ∈ Wrτ +1 (Ω) and
a pressure p ∈ Wrτ +2 (Ω)/R, solutions to (1.1)–(1.3), which satisfy
 
kukWrτ +1 (Ω) + kpkWrτ +2 (Ω)/R ≤ c kf kWrτ +1 (Ω) + kg · nkW τ +1−1/r (∂Ω) .
r

Besides the shift type theorem, we will apply so called sampling inequalities. To
state them, we have to introduce a measure for the data density on Ω and ∂Ω. In the
first case we introduce the “fill distance”

hX,Ω := sup min kx − xj k2 .

x∈Ω xj ∈X

The following result has precursors in [2, 18, 19] and comes in its vector-valued
form from [25].
Lemma 4.2. Let 1 < r < ∞, and τ, η ∈ R with τ > d/2 and 0 ≤ η ≤ τ − d(1/2 −
1/r)+ . Suppose Ω ⊆ Rd is a bounded domain having a Lipschitz boundary. Let X ⊆ Ω
be a discrete set with fill distance hX,Ω sufficiently small. Assume that u ∈ Hτ (Ω)
satisfies u|X = 0. Then we also have
τ −η−d(1/2−1/r)+
kukWrη (Ω) ≤ chX,Ω kukHτ (Ω) .

To introduce a measure on the boundary, we follow ideas from [9, 25]. Let ∂Ω =
∪Jj=1 Vj , where Vj ⊆ ∂Ω are relatively open sets. Furthermore,

ϕj : B → Vj ,

where ϕj is a C k,s -diffeomorphism and B = B(0, 1) denotes the unit ball in Rd−1 .
We will measure the density of the points Y on ∂Ω by introducing

hY,∂Ω := max hTj ,B

1≤j≤J

with Tj = ϕ−1j (Y ∩ Vj ) ⊆ B analogously to the definition of the fill distance. We

assume that the atlas Vj is fixed, i.e., we do not have to worry about the dependence
of hY,∂Ω on the atlas.
To derive the estimate on the boundary, we need a similar result as lemma 4.2
on manifolds. This has been done in [11] for the special case of ∂Ω being the sphere
in Rd and in a more general context in [9]. We give an extended version which also
deals with non-integer orders η. Its proof can be found in [25].
Lemma 4.3. Let 1 < r < ∞ and τ = k + s > d/2, where k ∈ N0 and 0 < s ≤ 1.
Let Ω ⊆ Rd be a bounded domain having a C k,s smooth boundary. Assume that
Y ⊆ ∂Ω with hY,∂Ω sufficiently small. Then there is a constant c > 0 such that for
all u ∈ Hτ (Ω) with u|Y = 0 we have for 0 ≤ η ≤ τ − 1/2 − (d − 1)(1/2 − 1/r)+ that
τ −1/2−η−(d−1)(1/2−1/r)+
kukWrη (∂Ω) ≤ chY,∂Ω kukHτ (Ω) .

12
 The standard trace theorem establishes that if u ∈ H τ (Ω) then u ∈ H τ −1/2 (∂Ω),
cf. [26, Theorem 8.7]. If τ > d/2, then this guarantees in combination with the
Sobolev embedding theorem that u is continuous on the boundary ∂Ω.
The following result gives the error estimates for smooth target functions, i.e.,
target functions form the associated reproducing kernel Hilbert space. It was the
main result in [21].
Theorem 4.4. Let Ω be a bounded, simply connected, open subset of Rd , d = 2, 3,
with a C ⌈τ ⌉+1,1 boundary ∂Ω. Suppose that Φ is chosen such that its associated Hilbert
space is HΦ (Rd ) = H e τ (Rd ; div) × H τ +1 (Rd ) and the permeability tensor K = Kij
satisfies (1.5), K = K T and Kij ∈ H τ +1 (Ω). R Furthermore, assume that the data
satisfy f ∈ Hτ +1 (Ω), g ∈ Hτ +1/2 (∂Ω) and ∂Ω g · n dS = 0, where τ > d/2. Then,
the error between the true solution and the collocation approximation can be bounded
by

τ −η−1−d(1/2−1/r)
+
ku − su kWrη+1 (Ω) + kp − sp kWrη+2 (Ω)/R ≤ c hX,Ω +
τ −η−1−1/2+1/r−(d−1)(1/2−1/r)+


+ hY,∂Ω kf kHτ (Ω) + kg · nkH τ −1/2 (∂Ω)

for 1 < r < ∞ and 0 ≤ η ≤ τ − d(1/2 − 1/r)+ − 1. If r ≥ 2 and h = hX,Ω ≈ hY,∂Ω

this reduces to

ku − su kWrη+1 (Ω) + kp − sp kWrη+2 (Ω)/R

≤ chτ −η−1−d(1/2−1/r) kf kHτ (Ω) + kg · nkH τ −1/2 (∂Ω) .

We now state and prove the error estimates for rougher target functions. Besides a
similar approach as in the standard error analysis, the main idea is to find a band-
limited function vσ , which approximates the true solution. Then we can add the term
vσ − vσ to the difference between the true solution v and our approximating function
sv . With the triangle inequality the norm can be split into two. The difference
between the true solution and the band-limited function can be bounded with theorem
3.5. The difference between the band-limited function and the approximating function
can be bounded with standard error analysis, since sv also approximates vσ and both
functions are sufficiently smooth.
Theorem 4.5. Let Ω be a bounded, simply connected, open subset of Rd , d = 2, 3,
with a C ⌈β⌉+1,1 boundary ∂Ω. Suppose that Φ e is chosen such that its associated Hilbert
space is HΦ e (R d
) = e
H τ
(R d
; div) × H τ +1
(R d
) and the permeability tensor K = Kij
T β+1
satisfies (1.5), K = K and Kij ∈ H (Ω).R Furthermore, assume that the data
satisfy f ∈ Hβ+1 (Ω), g ∈ Hβ+1/2 (∂Ω) and ∂Ω g · n dS = 0, where τ ≥ β > d/2.
Then, the error between the true solution and the collocation approximation can be
bounded by

ku − su kWrη+1 (Ω) + kp − sp kWrη+2 (Ω)/R

−β  
hτX,Ω
−β
+ hτY,∂Ω β−η−1−d(1/2−1/r)+ β−η−1−1/2+1/r−(d−1)(1/2−1/r)+
≤c hX,Ω + hY,∂Ω ×
q τ −β


× kf kHβ (Ω) + kg · nkH β−1/2 (∂Ω)

for every 1 < r < ∞ and 0 ≤ η ≤ β − d(1/2 − 1/r)+ − 1 and separation radius
13
q := qX∪Y . If r ≥ 2 and h = hX,Ω ≈ hY,∂Ω this reduces to
ku − su kWrη+1 (Ω) + kp − sp kWrη+2 (Ω)/R
 τ −β
h

≤ chβ−η−1−d(1/2−1/r) kf kHβ (Ω) + kg · nkH β−1/2 (Ω) .
q

Proof. First of all, we pick a representer p of the pressure such that kpkWrβ+2 (Ω) =
kpkWrβ+2 (Ω)/R .
Let v = (u, p). Since all norms on Rd are equivalent, we have that ku −
su kWrη+1 (Ω) + kp − sp kWrη+2 (Ω) is equivalent to kv − sv kWrη+1 (Ω)×Wrη+2 (Ω)/R .
We now apply proposition 4.1 to the difference v − sv instead of v, i.e., we derive
ku − su kWrη+1 (Ω) + kp − sp kWrη+2 (Ω)
 
≤ c kLv − Lsv kWrη+1 (Ω) + k(u − su ) · nkW η+1−1/r (∂Ω) ,
r

for all 0 ≤ η ≤ β. To estimate the two terms on the right hand side of the last
equation, we first observe that we have
(Lv − Lsv )(xj ) = 0, 1 ≤ j ≤ N,
(u − su ) · n(yj ) = 0, 1 ≤ j ≤ M.
Hence, we are dealing with smooth functions that have a large number of zeros.
In the first case we have functions defined on a bounded region of Rd , while in the
second case we are dealing with functions on a manifold. For such functions, we
can apply the sampling inequalities. We will now bound kLv − Lsv kWrη+1 (Ω) and
k(u − su ) · nkW η+1−1/r (∂Ω) separately.
r
We will start with the estimate in the interior. The function Lv − Lsv has many
zeros, i.e., we can apply the sampling inequality lemma 4.2, such that
β−η−1−d(1/2−1/r)+
kLv − Lsv kWrη+1 (Ω) ≤ chX,Ω kLv − Lsv kHβ (Ω) .
Obviously,


kLv − Lsv kHβ (Ω) ≤ c ku − su kHβ (Ω) + kp − sp kH β+1 (Ω)
≤ ckv − sv kHβ (Ω)×H β+1 (Ω) .
To bound (4.1), we apply the extension operator E to v and extend K component-
wise with Stein’s extension operator ES , see proposition 2.5. Then there exists a
band-limited function vσ which approximates the extension of v, see theorem 3.5.
Adding vσ − vσ and using the triangle inequality leads to
kv − sv kHβ (Ω)×H β+1 (Ω) = kEv − sEv kHβ (Ω)×H β+1 (Ω)
≤ kEv − vσ kHβ (Ω)×H β+1 (Ω) + kvσ − sEv kHβ (Ω)×H β+1 (Ω) .
(4.1)
The first part of (4.1) can be bounded by theorem 3.5 with t, r = 0 and the
properties of the extension operator:
kEv − vσ kHβ (Ω)×H β+1 (Ω) ≤ kEv − vσ kH e β (Rd )×H β+1 (Rd )
 
e div uk e β d + kES pkH β+1 (Rd )
≤ c kE H (R )


≤ c kukHβ (Ω) + kpkH β+1 (Ω)
≤ ckvkHβ (Ω)×H β+1 (Ω) . (4.2)
14
 To bound the second part of (4.1) we can apply theorem 4.4 with f := Lvσ , g :=
uσ , r := 2 and η := β − 1, since all functions are sufficiently smooth. The definition of
vσ provides that Lvσ = Lv on X and vσ · n = v · n on Y . This means in particular
that svσ = sEv , i.e., the reconstruction of vσ is the same as the reconstruction of sEv .
Furthermore, vσ is a smooth target function such that we can indeed apply theorem
4.4. If we use also the trace theorem and kLvσ kHτ (Ω) ≤ ckvσ kHτ (Ω)×H τ +1 we can
derive

 

τ −β
kvσ − sEv kHβ (Ω)×H β+1 (Ω) ≤ c hX,Ω + hτY,∂Ω
−β
kLvσ kHτ (Ω) + kuσ · nkH τ −1/2 (∂Ω)
 
≤ c hτX,Ω
−β
+ hτY,∂Ω
−β
kvσ kHτ (Ω)×H τ +1 (Ω) .

Since vσ = (uσu , pσp ) is band-limited, we have Bernstein estimates of the form

kvσ kHτ (Ω)×H τ +1 (Ω) ≤ kvσ kHτ (Rd )×H τ +1 (Rd )

≤ c kuσu kHτ (Rd ) + kpσp kH τ +1 (Rd )


≤ c σuτ −β kuσu kH β (Rd ) + σpτ −β kpσp kH β +1 (Rd )
≤ cq β−τ kvσ kHβ (Rd )×H β+1 (Rd )
≤ cq β−τ kvσ kH
e β (Rd )×H β+1 (Rd ) ,

where we have also used the fact that σu and σp are proportional to 1/q.
Adding and subtracting Ev and applying the triangle inequality gives in combi-
nation with (4.2) and the properties of the extension operator that

kvσ kH
e β (Rd )×H β+1 (Rd ) ≤ kEv − vσ kH
e β (Rd )×H β+1 (Rd ) + kEvkH
e β (Rd )×H β+1 (Rd )

≤ ckvkHβ (Ω)×H β+1 (Ω) . (4.3)

With (4.3), we can bound the second part of (4.1) by

 
kvσ − sEv kHβ (Ω)×H β+1 (Ω) ≤ cq β−τ hτX,Ω
−β
+ hτY,∂Ω
−β
kvkHβ (Ω)×H β+1 (Ω) .

hτ −β +hτ −β
Since q ≤ qX ≤ hX,Ω and β ≤ τ , we have that X,Ωqτ −βY,∂Ω is greater than or equal to
one. Combining the above inequalities and applying proposition 4.1 gives

hτX,Ω
−β
+ hτY,∂Ω
−β


kLv − Lsv kHβ (Ω) ≤ c kukHβ (Ω) + kpkH β+1 (Ω)/R
q τ −β
hτX,Ω
−β
+ hτY,∂Ω
−β


≤c kf kHβ (Ω) + kg · nkH β−1/2 (Ω) . (4.4)
q τ −β

We now bound the boundary part. The proof of [21, proposition 3.6] establishes
e ∈ H⌈β⌉ (Ω) of the normals n to the interior of Ω with
that there exists an extension n
e |∂Ω = n|∂Ω .
n
The function u − su has many zeros, i.e., we can apply the sampling inequality
lemma 4.3. Therefore
β−1−1/2−η+1/r−(d−1)(1/2−1/r)+
k(u − su ) · nkW η+1−1/r (∂Ω) ≤ chY,∂Ω e kH β (Ω) .
k(u − su )· n
r

15
The proof of [21, proposition 3.6] also establishes
e kH β (Ω) ≤ ke
k(u − su ) · n nkHβ (Ω) ku − su kHβ (Ω) ≤ cku − su kHβ (Ω) .
With (4.1) and (4.4) we have
τ −β
hX,Ω + hτY,∂Ω
−β


ku − su kHβ (Ω) ≤ c kf kHβ (Ω) + kg · nkH β−1/2 (Ω) ,
q τ −β
which finishes the proof.
In the case that β = τ , the result above is identical to the one in theorem 4.4.
Note that the limitation of this result to the dimensions d = 2, 3 is only due to
the fact that the extension operator has not yet been proven for a general d > 3.
5. Numerical Examples. We will now illustrate our theory by considering a
numerical example for rougher target functions.
In all computations, the compactly supported Wendland functions φd,ℓ are chosen
e Thus they are reproducing kernels of
for the underlying functions φ and ψ of Φ.
d ℓ+1/2
HΦ
e (R ) = H (Rd ; div) × H ℓ+3/2 (Rd ).

We will focus on the L2 and L∞ errors. Let f ∈ Hβ+1 (Ω) and g ∈ Hβ+1/2 (∂Ω),
where d/2 < β ≤ τ := d/2 + ℓ + 1/2. Theorem 4.5 shows that we have to expect the
following behaviour of the error:
 τ −β
h
ku − su kHη (Ω) + kp − sp kH η+1 (Ω) ≤ cf ,g hβ−η ,
q
 τ −β
h
ku − su kW∞
η
(Ω) + kp − s p k η+1
W∞ (Ω) ≤ c f ,g hβ−η−d/2 .
q
Note that the convergence order does not depend on the smoothness τ of the employed
kernel if the separation radius is comparable to the fill distance, i.e., on quasi-uniform
data sets.
We choose f and g such that the true solution of the velocity and the pressure
are
 
−∂y
u(x, y) = φ2,1 (r), p(x, y) = x3 y 2 ,
∂x
p
where r := (x − x0 )2 + (y − y0 )2 /γ with x0 = y0 = γ = 0.5. Furthermore, we pick
K = I the identity matrix. Figure 5.1 shows the velocity field and the contour lines
of the pressure.
The Wendland function φd,ℓ is an element of all Sobolev spaces H α (Rd ) with
α < 2τ − d/2 = d + 2ℓ + 1 − d/2. Therefore the function φ2,1 is in H α (Ω) with α < 4.
Due to our choice of the velocity, we have u ∈ H β+1 (Ω) for all β < 2. Thus u is not
an element of the associated Hilbert space of φ2,3 , where τ = 3.5. We chose β = 2
which is the supremum of the smoothness, to work out the theoretical approximation
orders.
In all cases the notation eu = u − su and ep = p − sp is used. The numerical
tests were run on a sequence of equidistant grids. The computational approximation
orders are given by
log(En /E2n )
,
log(1/2)
16
 1 1

0.4
0.001

0.
8
0.9 0.9

0.1
0.01
0.0001

0.
6
0.8 0.8

0.0

0.
2
5
0.7 0.7

0.6 0.6
0.
1

0.00
0.5 0.5 0.2
y

y
0.

0.0
1
05

0.0001

1
0.4 0.4
0.1
0.3 0.3
0.05
0.2 0.2 0.01
0.0
01
0.1 0.1 0.01
0.00
01 0.001
0.0001 0.0001
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x

(a) The velocity field for the homogeneous exam- (b) The contour lines of the pressure field.
ple visualized with unit vectors.

Figure 5.1. The true solution of the homogeneous example.

n keu kL2 keu kL∞ keu kH1 k∇ep kL2 k∇ep kL∞
4 1.2836e+00 2.2024e+00 9.8131e+00 2.0342e-01 1.5629e+00
8 7.9192e-02 3.8168e-01 1.6632e+00 3.4692e-02 2.1824e-01
16 8.7058e-03 5.1376e-02 3.9872e-01 2.9099e-03 3.0136e-02
32 7.4455e-04 1.1950e-02 7.9047e-02 2.7432e-04 3.0789e-03
64 6.5388e-05 2.8183e-03 1.6406e-02 2.8833e-05 7.2737e-04
Table 5.1
Approximation errors for the example with rougher target functions with φ = ψ = φ2,3 .

where En is one of the errors for given n × n input grid.

All results are displayed in tables 5.1 and 5.2. Here, the values in the brackets
give the approximation orders if the target functions were in the reproducing kernel
Hilbert space. Figure 5.2 illustrates the numerical approximation errors. From table
5.2 it can be seen that the numerical approximation orders more than match the
theoretical ones. Moreover, some of them even match the approximation orders for
smoother target functions.

REFERENCES

[1] Robert A. Adams, Sobolev spaces, New York : Academic P., 1975.
[2] Rémi Arcangéli, Marı́a Cruz López de Silanes, and Juan José Torrens, An extension
of a bound for functions in Sobolev spaces, with applications to (m, s)-spline interpolation
and smoothing, Numer. Math., 107 (2007), pp. 181–211.
[3] Jacob Bear, Dynamics of fluids in porous media, American Elsevier Publishing Company,
inc., 1972.
[4] Susanne C. Brenner and L. Ridgway Scott, The Mathematical Theory of Finite Element
Methods, Springer, 1994.
[5] Robert A. Brownlee and Will Light, Approximation orders for interpolation by surface
splines to rough functions, IMA Journal of Numerical Analysis, 24 (2004), pp. 179–192.
[6] Edward J. Fuselier, Improved stability estimates and a characterization of the native space
for matrix-valued rbf ’s, Adv Comput Math, 29 (2007), pp. 269–290.
[7] , Erratum: Improved stability estimates and a characterization of the native space for
matrix-valued rbfs, Advances in Computational Mathematics, 29 (2008), pp. 311–313.
[8] , Sobolev-type approximation rates for divergence-free and curl-free rbf interpolants,
17
 keu kL2 keu kL∞ keu kH1 k∇ep kL2 k∇ep kL∞
computed 4.0187 2.5286 2.5607 2.5518 2.8403
3.1853 2.8932 2.0605 3.5756 2.8564
3.5475 2.1041 2.3346 3.4070 3.2910
3.5093 2.0841 2.2685 3.2501 2.0817
estimated 2 (3.5) 1 (2.5) 1 (2.5) 2 (3.5) 1 (2.5)
Table 5.2
Approximation orders for the example with rougher target functions with φ = ψ = φ2,3 , where
the values in the brackets give the approximation orders if the target function would be in the
reproducing kernel Hilbert space.

1
10

0
10

−1
10
Error

−2
10
|| e ||
u L
2

−3 ||eu||L
10 ∞

|| eu ||H1
−4 ||∇ ep||L
10 2

||∇ ep||L
∞

−2 −1
10 10
h

Figure 5.2. Approximation errors of example with rougher target functions, where φ = ψ = φ2,3 .

Mathematics of Computation, 77 (2008), pp. 1407–1423.

[9] Peter Giesl and Holger Wendland, Meshless collocation: Error estimates with application
to dynamical systems, SIAM J. Numer. Anal., 45 (2007), pp. 1723–1741.
[10] Georg M. Hornberger, Jeffrey P. Raffensperger, Patricia L. Wilberg, and Keith N.
Eshleman, Elements of Physical Hydrology, John Hopkins University Press, 1998.
[11] Quoc T. Le Gia, Francis J. Narcowich, Joseph D. Ward, and Holger Wendland, Con-
tinuous and discrete least-squares approximation by radial basis functions on spheres, J.
Approx. Theory, 143 (2006), pp. 124–133.
[12] Jeremy Levesley and Xingping Sun, Approximation in rough native spaces by shifts of
smooth kernels on spheres, J. Approx. Theory, 133 (2005), pp. 269–283.
[13] , Corrigendum to and two open questions arising from the article: “Approximation
in rough native spaces by shifts of smooth kernels on spheres” [ J. Approx. Theory,
133(2):269–283, 2005], J. Approx. Theory, 138 (2006), pp. 124–127.
[14] Francis J. Narcowich, Recent developments in error estimates for scattered-data interpola-
tion via radial basis functions, Numerical Algorithms, 39 (2005), pp. 307–315.
18
[15] Francis J. Narcowich and Joseph D. Ward, Generalized Hermite interpolation via matrix-
valued conditionally positive definite functions, Math. Comput., 63 (1994), pp. 661–687.
[16] , Scattered data interpolation on spheres: Error estimates and locally supported basis
functions, SIAM Journal on Mathematical Analysis, 33 (2002), pp. 1393–1410.
[17] , Scattered-data interpolation on Rn : Error estimates for radial basis and band-limited
functions, SIAM Journal on Mathematical Analysis, 36 (2004), pp. 284–300.
[18] Francis J. Narcowich, Joseph D. Ward, and Holger Wendland, Sobolev bounds on func-
tions with scattered zeros, with applications to radial basis function surface fitting, Math.
Comp., 74 (2005), pp. 743–763.
[19] , Sobolev error estimates and a Bernstein inequality for scattered data interpolation via
radial basis functions, Constructive Approximation, 24 (2006), pp. 175–186.
[20] Daniela Schräder, Analytically Divergence-free Discretisation Methods for Darcy’s P, PhD
thesis, 2009.
[21] Daniela Schräder and Holger Wendland, A high-order, analytically divergence-free dis-
cretization method for Darcy’s problem. Preprint Sussex, to appear in Mathematics of
Computation, 2008.
[22] Elias M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton
University Press, 1971.
[23] Hans Triebel, Interpolation theory, function spaces, differential operators, North-Holland
Publishing Company, 1978.
[24] Holger Wendland, Scattered Data Approximation, Cambridge Monographs on Applied and
Computational Mathematics, Cambridge University Press, 2005.
[25] , Divergence-free kernel methods for approximating Stoke’s problem, SIAM Journal on
Numerical Analysis, 47 (2009), pp. 3158–3179.
[26] Joseph Wloka, Partial Differential Equations, Cambridge University Press, 1987.

19