We study meshless collocation methods using radial basis functions to approximate reg-
ular solutions of systems of equations with linear differential or integral operators. Our
method can be interpreted as one of the emerging meshless methods, cf. T. Belytschko et
al. (1996). Its range of application is not confined to elliptic problems. However, the ap-
plication to the boundary value problem for an elliptic operator, connected with an integral
equation, is given as an example. Although the method has been used for special cases for
about ten years, cf. E.J. Kansa (1990), there are no error bounds known. We put the main
emphasis on detailed proofs of such error bounds, following the general outline described
in C. Franke and R. Schaback (preprint).
Keywords: convergence orders, meshless collocation, radial basis functions, partial differ-
ential equations
AMS subject classification: 35A40, 65N35, 41A25, 41A63, 65R20
1. Introduction
We will treat systems of linear equations, each of the form
Lu = f on , (1.1)
where is a domain in Rd and L is a linear operator, acting on complex valued
functions on . The function f is given (data) and each equation should have a
nonempty space of solutions.
where the j : C are certain linearly independent basis functions. The connection
between u and su shall be the pointwise equality
(Lsu )(x) = (Lu)(x) = f (x) for all x X ,
or, in short form,
(Lsu )(X ) = f (X ) (1.3)
on a finite, ordered set X of centers. Since X is ordered, f (X ) is a vector.
The set (1.3) of generalized interpolation conditions leads to the linear interpo-
lation system
(Lj )(x) xX , j=1,...,n = f (X ), (1.4)
using (1.2) for collocation. Now there are two main questions:
Does the inverse of the interpolation matrix ((Lj )(x))xX , j=1,...,n exist?
Are there any upper bounds for ksu uk or kLsu Luk?
Both questions will be answered here, with focus on the latter.
collocation conditions L u(x) = f (x) can be imposed at a single point. Now the
interpolation system (1.4) takes the form of the N simultaneous equation systems
(L j )(X ) j=1,...,n = f (X ) for = 1, . . . , N . (1.8)
For a fixed x, x denotes the linear functional mapping a function f to f (x). For
a given map T , the complex conjugate map T is defined by T (f ) := T (f ), and
the superscript y of T y denotes the application of T to (x, y) with respect to the
variable y.
To simplify the notation, and to connect the theory with HermiteBirkhoff inter-
polation, we define the functionals
j := xj L for xj X , = 1, . . . , N ,
Next we define the vector on the right hand side of the interpolation system (1.8) to
be z := (f (X ))=1,...,N . Now this system takes the simple form
y
x (x, y) = z. (1.10)
y
Since is semisymmetric, the interpolation matrix x (x, y) is Hermitian. In [19,
theorem 3.1], it is shown to be positive definite under reasonable conditions, and then
it is invertible. We will reuse parts of the proof of the above-mentioned theorem in an
extended context. To express the dependence of the approximating function su on
and , we subsequently denote it by su,,. It takes the form
X
N X y
su,,(x) = xj L (x, y) j ,
=1 xj X
384 C. Franke, R. Schaback / Collocation using radial basis functions
or briefly,
y
su,,(x) = (x, y) . (1.11)
y
Here, (x, y) is regarded as a 1 n matrix of functions in x.
Let L be a linear space of complex valued functionals, being defined for functions
on . For example,
L := span {x L}x {x B}x
is useful in the case of the boundary value problem (1.7).
Definition 2.1 (Feasible basis functions and native spaces). A feasible basis function
with respect to L is a function : C, being
semisymmetric: (x, y) = (y, x) for all x, y ,
positive definite: for every finite, nonempty, linearly independent ordered set L,
y
6= 0, the matrix x (x, y) is positive definite, and
sufficiently smooth to apply two functionals , of L to each variable and to let
the functionals commute: x y (x, y) = y x (x, y).
For any feasible basis function , the term
h | iF := x y (x, y) (2.1)
defines a scalar product on the dual native space
F := : h | iF < with L F . (2.2)
Subsequently it is assumed to be a Hilbert space. The dual F of F is identified
with the native space
F := f : kf kF < (2.3)
of (generalized) functions f which allow to be evaluated by all F . The norm is
defined by
|(f )|
kf kF := sup . (2.4)
F ,6=0 kkF
Since F is a Hilbert space, so is F .
Remark 2.2. The term x y (x, y) in equation (2.1) defines a scalar product, since
it is sesquilinear and is assumed to be positive definite. Methods for proving the
positive definiteness and examples are given in section 3.
C. Franke, R. Schaback / Collocation using radial basis functions 385
for the reconstruction error |(u) (su,,)| holds for every F , where
P, () := inf k kF (2.6)
span
Equations (2.5)(2.7) show how the influences of a test functional and of the
function u on the error are separable. This separation results in two approximation
errors: The power function is the error of approximation of in span , while the
y
second factor is the error concerning u and span (, y). This factor is independent
of .
To adapt the previous theorem to systems of linear equations, we need some
information about the behaviour of the power function concerning ordered subsets 0
of , which contain functionals of the same type.
for the infima. The second assertion is proved analogously, using equation (2.7).
The Splitting Theorem allows us to focus our attention on just one (denoted
by to save subscripts), its centers X := X and its operator L := L , as they are
given in section 1.2.
Theorem 2.7 (Transformation Theorem). Let (x, y) be a feasible basis function. Let
y
the linear functionals (z L)x and y applied to L (x, y) commute for all z
and F . If
v
L (x, y) := (x L)u y L (u, v) (2.8)
is also a feasible basis function, and F L , FL and PL , are defined for it according
to definition 2.1 and equation (2.6), then
kkF = k LkF (2.9)
L
Proof. Since y and (x L)u commute, equations (2.8) and (2.1) yield for all ,
F L
v
h | iF = x y (x L)u y L (u, v)
L
v
= ( L)u L (u, v) = h L | LiF .
Equation (2.9) follows immediately. Moreover, we see ( ) L F for
F L and span . Therefore PL , () is well-defined. Using equation (2.6), we
calculate
PL , () = inf k kF = inf k L LkF
span L span
= inf k L kF = P,L ( L).
span L
C. Franke, R. Schaback / Collocation using radial basis functions 387
To demonstrate the use of the theorems of this section, we give a typical appli-
cation. Regard and as given. The ordered subsets of should have the form
of (1.9), i.e., = X L with X := X finite. Then from theorem 2.5
and the Splitting Theorem 2.6, we get the error bound
(u su,,) 6 P, () ku su,,kF
Under certain circumstances one has to resort to the generalized distributional Fourier
transform, but we omit these technical details here. Further we will assume that is
nonnegative and positive almost everywhere. This ensures to be positive definite
with respect to x -functionals, cf. [2]. To take advantage of the representation of
by (3.1) and (3.2), we have to restrict the dual native space F to functionals
commuting with the Fourier transform. That means, we assume
Z Z
iT x
F : x
() e d = ()sym () d (3.3)
is called the symbol function of the functional . It may even be a generalized function,
defined by the equality in (3.3). In addition, we require any function f in F to have
a representation via the inverse Fourier transform:
for any f F there is a function fb such that f = FT1 fb . (3.5)
Where necessary, we assume fb = FT(f ), that is, the Fourier transform is bijective
on F . Apparently these are many assumptions, but we know several kinds of feasible
basis functions which allow our construction.
Remark 3.1. It is also common to define the symbol function by (ei x ), using
T
the negative sign Rin the exponent. If is a regular distribution, i.e., a functional of
the form l (g) = l()g() d with a representing function l, the so defined symbol
function coincides with the representer of the Fourier transform
Z
T
FT(l )(f ) := l FT(f ) = f (x)l ei x dx
C. Franke, R. Schaback / Collocation using radial basis functions 389
of l by the use of Fubinis theorem. Therefore, this definition of the symbol function
b But we will not use this notation, because it requires to
justifies to denote it by .
be regular or to use generalized Fourier transforms.
= (2)d
T
()x ei x y eiT y d
ZR
d
Theorem 3.2.
Equations (3.3) and (3.5) yield
Z
(f ) = (2) d
fb() sym () d (3.8)
Rd
for all F and f F .
If the linear operator L is translation invariant, i.e.,
Lx f (x z) = (Lf )(x z) (3.9)
for all z Rd and f in the domain of L, we have
Tx
symx L () = sym0 L () ei for all x, Rd . (3.10)
Lemma 3.3. Let F and F satisfy equations (3.3) and (3.5), respectively. Then for
every f F , the equation
fb() = () symf ()
Proof. Due to the theorem of Riesz, for any given f F , there is a unique f F
satisfying (f ) = h | f iF for all F . Using equations (3.6) and (3.8), we
find
Z
1 b
d
FT f = (f ) = h | f iF = (2) ()sym ()symf () d
Rd
= FT1 symf
Theorem 3.4. Let be a feasible basis function which has a representation via equa-
tions (3.1) and (3.2). Let F and F satisfy equations (3.3) and (3.5), respectively.
Then the equation
Z
hf | giF = (2)d fb()b
g ()/() d (3.11)
Rd
holds for every f , g F . Recall that can only vanish on a set of measure 0.
Equation (3.11) is the analog of equation (3.6) for the scalar product in F . More
about this representation is found in [20, section 4].
Proof. Due to lemma 3.3, we know symf () = fb()/() almost everywhere for
any given function f , where f F is the Riesz representer of f . We use its property
hg | f iF = hf | giF and equation (3.6) to calculate
Z
hf | giF = hg | f iF = (2)d symg ()symf () () d
Rd
Z
= (2)d
gb()fb()/() d.
Rd
As we saw above, the symbol function is an essential tool for proving positive
definiteness of the basis functions and L . Moreover, we will use it to establish
a connection between the norms of their native spaces F and FL and between
their power functions, respectively. For this reason, we assume the operator L to
be translation invariant. Since is translation invariant, so is L , and a function
C. Franke, R. Schaback / Collocation using radial basis functions 391
where
Tx Tz
Lx ei := xz Lz ei = symx L ()
denotes the symbol function for functionals of the form = x L with a linear
operator L. Using equation (3.10), we calculate
T 2 T
Lx ei x Ly ei y = sym0 L () ei (xy) .
T
Theorem 3.5 (Primal Transformation Theorem). Let L be translation invariant and let
it commute with the Fourier transform integral for f F . Let and L be feasible
basis functions. Let F and F satisfy equations (3.3) and (3.5), respectively. Then
the connection between k kFL and k kF is
The combination of these results proves (3.13). To prove the second assertion, we
use (3.13) to calculate
infy ku skF = infy kLu LskFL
sspan (,y) sspan (,y)
= inf y
kLu s0 kFL
s0 span L (,y)
= inf
y
kLu skFL .
sspan X L (,y)
The last assertion follows from the first line of this equation.
We give two simple, but typical examples for translation invariant linear opera-
tors L:
Example 3.6. Let L := p(D) be a partial differential operator with constant coeffi-
cients, i.e., p is a polynomial on i Rd and D := (/xj )j=1,...,d . Then we find
y
L (x, y) = p(D)x p(D) (x, y)
and
Tx
symx L () = p(i) ei (3.16)
for every x . With equation (3.7), this implies
Z
kx kF = kx LkF = (2)
2 2 d p(i)2 () d.
L
Rd
Remark 3.8. The paper [6] shows a general technique to apply the above abstract
results to general problems involving partial differential equations. However, [6] does
not produce explicit convergence orders, and thus we add a section to demonstrate
how the technique works for elliptic boundary value problems.
is a feasible basis function with respect to L := span{x }xRd . The native space F
is norm isomorphic to the Sobolev space H (Rd ), i.e., F = H (Rd ).
for L1 = FT(L1 ,0 ) and for all Rd . Since p is a polynomial of order m and does
not vanish anywhere, we get with new constants c2 > c1 > 0 the inequalities
2(m) 2(m)
c1 1 + kk2 6 L1 () 6 c2 1 + kk2 for all Rd , (4.3)
Theorem 4.2P(cf. [7, theorem 3.7]). Let the polynomial according to L1 have the
form p(x) = di,j=1 ai,j xi xj + bT x + c, where A := (ai,j )i,j=1,...,d , b and c 6 0 are real
valued and constant. The operator L1 shall be elliptic in the sense of A = AT and
xT x 6 xT Ax 6 M xT x
for every x Rd with constants M > > 0. The domain 1 shall be bounded. If
e C 2 (1 ) C 0 (1 ) and fe C 0 (1 ) satisfy L1 u
the functions u e = fe on 1 , then
c
sup ue(x) 6 sup u e(x) + sup fe(x) (4.5)
x1 x1 x1
holds, where the constant c depends only on diam and kbk2 /.
e := u su,, and fe := L1 (e
We apply this theorem to u u) to get
sup u(x) su,,(x)
x1
c
6 sup x (u su,,) + sup x L1 (u su,,). (4.6)
x3 x1
We can use the theorem, since > > m + d/2, u H (Rd ) and su,, F =
d
H (R ) implies u d 2 d e
e H (R ) C (R ) and f C (R ). Due to theorem 2.5, we
0 d
know
x (u su,,) 6 P, (x ) ku su,,kF for all x 3 ,
x L1 (u su,,) 6 P, (x L1 ) ku su,,kF for all x 1 , (4.7)
x L2 (u su,,) 6 P, (x L2 ) ku su,,kF for all x 2 .
for all x , = 1, 2. The first line allows the application of the classical theory
of interpolation with radial basis functions, while the second line still needs some
work. We use the Transformation Theorem 2.7 and see
P,X L (x L ) = PL ,X (x ) for all x , = 1, 2. (4.9)
The Primal Transformation Theorem 3.5 yields
ku su,X L , kF = kL u L su,X L , kFL
where we know that its hypotheses (3.3) and (3.5) are satisfied, since u e H (Rd ).
The second equality is due to theorem 2.8. It contains the approximation error of L u
instead of us. Note that L u = f on .
396 C. Franke, R. Schaback / Collocation using radial basis functions
for all x 1 using (4.6), (4.8), (4.9) and (4.10), and finally we saw
L2 (u su,,)(x) 6 P, L (x L2 ) kf2 sf , , kF
X2 2 2 X2 L2 L
2
for all x 2 from (4.8), and (4.10). Applying equation (2.11), we can bound the
last item by
L2 (u su,,)(x) 6 P, (x )kL2 k kf2 sf , , kF (4.13)
X2 2 X2 L2 L
2
for all x 2 .
If = X , then a finer distribution X 0 X of centers implies a decrease of
P, (x ) by means of the Splitting Theorem 2.6. This effect shall be used now to
establish convergence orders in terms of powers of hX, . We recall theorem 5 of [20].
It says that for and satisfying condition (4.2) there exist constants h0 , C > 0 with
d/2
P,X (x) 6 ChX, (4.14)
for any distribution X of centers with hX, < h0 and any x .
Remark 4.4. The L -norm can be replaced by the L2 -norm to gain an additional factor
d/2
hX , cf. [17, theorem 5]. Since every sf ,X ,L is the result of a minimization,
C. Franke, R. Schaback / Collocation using radial basis functions 397
the norms on the right hand side can be bounded by kukH (Rd ) , kf1 kHm (Rd ) and
kf2 kH (Rd ) , respectively.
5. Conclusion
The theory of finite element methods (FEM) yields the following bound of ap-
proximation error.
Theorem 5.1 (cf. [9, Satz 4.2]). Let the domain Rd have a polyhedral boundary
and a quasi-uniform decomposition T into finite elements, which are affine images of
a common reference element. Let the maximal diameter of all finite elements be 2h.
The order m of the given differential operator L shall be even. Let V be a
subspace of Hm/2 (). Partial integration is used to define the continuous and V -
elliptical bilinear form a(u, v) := hLu | viV on V V . Let f be a continuous linear
form on V . We assume there is a solution u V H() of the problem a(u, v) = f (v)
for all v V with the higher regularity = qFEM + 1 > m/2.
Let the span Vh V of basis and test functions contain the space of on T
piecewise polynomial functions of degree at most qFEM.
Then the conforming finite element problem a(uh , vh ) = f (vh ) for all vh Vh
has a unique solution uh which satisfies
ku uh kHm/2 () 6 C|u|H () hm/2 . (5.1)
Comparing the error bound (5.1) of the FEM with the error bounds of corollary 4.3
for collocation, we note several points:
The collocation method requires a very regular solution u H () with >
m + d/2. It constructs an approximation of smoothness order > m + d/2 and
approximation order md/2. The FEM needs only > m/2. Its approximating
function has smoothness order 1 > m/2 1 and approximation order m/2.
But our method yields an L -error bound, while FEM yields a Hm/2 one. There is
an additional hd/2 convergence factor, if our estimate is rewritten to an L2 -norm.
The additional regularity required by our method clearly limits its direct applicabil-
ity. However, current research along the directions of, e.g., [3] shows that there are
promising techniques to handle cases of low regularity in such a way that the core
solution method has to deal only with the regular part of the solution. Combined
with such techniques, the regularity requirements are much less serious.
Collocation as a meshless method needs no geometric information. Thus the main
impact of our approach will be towards high-dimensional problems with high reg-
ularity.
We recall that Wendlands functions produce sparse systems due to their compact
support. Therefore, the complexity of the collocation method can possibly be re-
duced to O(#centers), cf. [14]. However, the underlying theory is difficult and still
398 C. Franke, R. Schaback / Collocation using radial basis functions
Altogether we see: The collocation method is feasible for problems with very
regular solutions in high space dimensions or with many different operators. This
roughly complements the set of problems where the FEM has proven to be an extremely
effective tool.
Acknowledgements
References
[1] T. Belytschko, Y. Krongauz, D.J. Organ, M. Fleming and P. Krysl, Meshless methods: An overview
and recent developments, Comput. Methods Appl. Mech. Engrg. 139 (1996) 347. Preprint: http:
//tam6.mech.nwu.edu/mfleming/EFG/meshless.ps.
[2] S. Bochner, Monotone Funktionen, Stieltjessche Integrale und harmonische Analyse, Math. Ann.
108 (1933) 378410.
[3] S.C. Brenner, Multigrid methods for the computation of singular solutions and stress intensity
factors II: Crack singularities, BIT 37(3) (1997) 623643.
[4] G.E. Fasshauer, Solving partial differential equations by collocation with radial basis functions, in:
Surface Fitting and Multiresolution Methods, eds. A. Le Mehaute, C. Rabut and L.L. Schumaker
(Vanderbilt University Press, Nashville, TN, 1997) pp. 131138. Preprint: http://www.math.
nwu.edu/fass/collocate.ps.
[5] G.E. Fasshauer and J.W. Jerome, Multistep approximation algorithms: Improved convergence rates
through postconditioning with smoothing kernels (1995). Preprint: http://www.math.nwu.
edu/fass/smooth.dvi.
[6] C. Franke and R. Schaback, Solving partial differential equations by collocation using ra-
dial basis functions, Appl. Math. Comput., to appear. Preprint: http://www.num.math.
uni-goettingen.de/schaback/research/papers/SPDEbCuRBF.ps.
[7] D. Gilbarg and N.S. Trudinger, Elliptic Partial Differential Equations of Second Order, Grundlehren
der Mathematischen Wissenschaften, Vol. 224 (Springer, Berlin, Heidelberg, 2nd ed., 1983).
[8] M.A. Golberg, Recent developments in the numerical evaluation of particular solutions in the
boundary element method, Appl. Math. Comput. 75 (1996) 91101.
[9] C. Gromann and H.-G. Roos, Numerik Partieller Differentialgleichungen (Teubner, Stuttgart, 2nd
ed., 1994).
[10] A. Iske, Reconstruction of functions from generalized HermiteBirkhoff data, in: Approximation
Theory VIII, Vol. 1, Approximation and Interpolation, eds. C.K. Chui and L.L. Schumaker (World
Scientific, Singapore, 1995) pp. 257264.
C. Franke, R. Schaback / Collocation using radial basis functions 399
[11] E.J. Kansa, Multiquadrics a scattered data approximation scheme with applications to computa-
tional fluid-dynamics I: Surface approximations and partial derivative estimates, Comput. Math.
Appl. 19(8/9) (1990) 127145.
[12] F.J. Narcowich, R. Schaback and J.D. Ward, Multilevel interpolation and approximation
(1997). Preprint: http://www.num.math.uni-goettingen.de/schaback/research/
papers/MIaA.ps.
[13] R. Schaback, Multivariate interpolation and approximation by translates of a basis function, in:
Approximation Theory VIII, Vol. 1, Approximation and Interpolation, eds. C.K. Chui and L.L. Schu-
maker (World Scientific, Singapore, 1995) pp. 491514.
[14] R. Schaback, On the efficiency of interpolation by radial basis functions, in: Surface Fitting and
Multiresolution Methods, eds. A. Le Mehaute, C. Rabut and L.L. Schumaker (Vanderbilt University
Press, Nashville, TN, 1997) pp. 309328. Preprint: ftp://ftp.gwdg.de/pub/numerik/
schaback/efficiency1.dvi.Z.
[15] H. Wendland, Piecewise polynomial, positive definite and compactly supported radial functions of
minimal degree, Adv. Comput. Math. 4 (1995) 389396.
[16] H. Wendland, Error estimates for interpolation by compactly supported radial basis func-
tions of minimal degree, J. Approx. Theory, to appear. Preprint: http://www.num.math.
uni-goettingen.de/wendland/error.ps.gz.
[17] H. Wendland, Sobolev-type error estimates for interpolation by radial basis functions, in: Surface
Fitting and Multiresolution Methods, eds. A. Le Mehaute, C. Rabut and L.L. Schumaker (Vander-
bilt University Press, Nashville, TN, 1997) pp. 337344. Preprint: http://www.num.math.
uni-goettingen.de/wendland/sobtyp.ps.gz.
[18] J.T. Wloka, Partial Differential Equations (Cambridge University Press, Cambridge, 1987).
[19] Z. Wu, HermiteBirkhoff interpolation of scattered data by radial basis functions, Approx. Theory
Appl. 8(2) (1992) 110.
[20] Z. Wu and R. Schaback, Local error estimates for radial basis function interpolation of scattered
data, IMA J. Numer. Anal. 13 (1993) 1327. Preprint: ftp://ftp.gwdg.de/pub/numerik/
schaback/wu1.dvi.Z.