Meshfree Chapter 15

MATH 590: Meshfree Methods
Chapter 15: Refined and Improved Error Bounds
Greg Fasshauer
Department of Applied Mathematics

Illinois Institute of Technology
Fall 2010
fasshauer@iit.edu MATH 590 Chapter 15 1

Outline
1 Native Space Error Bounds for Specific Basis Functions
2 Improvements for Native Space Error Bounds
3 Error Bounds for Functions Outside the Native Space
4 Error Bounds for Stationary Approximation
5 Convergence With Respect to the Shape Parameter
6 Dimension-Independent Error Bounds
7 Polynomial Interpolation as the Limit of RBF Interpolation

Native Space Error Bounds for Specific Basis Functions
For the first part of this chapter we discuss the non-stationary

approximation setting.
The additional refinement of the error estimate of the
approximation theorem from Chapter 14 for specific kernels K is
rather technical (for details see, e.g., the book
[Wendland (2005a)]).
A large body of literature exists on this topic such as, e.g.,
[Buhmann and Dyn (1991), Light (1996), Light and Wayne (1995),
Light and Wayne (1998), Madych (1992),
Madych and Nelson (1992), Narcowich and Ward (2004),
Narcowich et al. (2003), Narcowich et al. (2005),
Schaback (1995a), Schaback (1996), Wendland (1998),
Wendland (1997), Wu and Schaback (1993), Yoon (2003)]).
We now list some of the results that can be obtained.

Native Space Error Bounds for Specific Basis Functions Infinitely Smooth Basis Functions
As mentioned before, an application of the estimate

k ||
p
|D f (x) D Pf (x)| ChX , CK (x)|f |NK () ,
for interpolation on bounded domains Rs from Chapter 14 to

infinitely smooth functions such as Gaussians or generalized (inverse)
multiquadrics immediately yields arbitrarily high algebraic convergence
rates, i.e., for every ` N and || ` we have
|D f (x) D Pf (x)| C` h`|| |f |NK () . (1)
whenever f NK ().
A considerable amount of work has gone into investigating the

dependence of the constant C` on ` (see, e.g., [Wendland (2001)] or
the recent work in [Luh (2009)]).
In general, the constant C` even depends on the space dimension s, in

essence allowing high-order approximation only for small values of s
(see, e.g., [Fasshauer et al. (2010), Luh (2009)]).
Using different proof techniques (see [Madych and Nelson (1988)] or

[Wendland (2005a)] for details) it is possible to derive more precise
error bounds for Gaussians and (inverse) multiquadrics (whose
constants still depend on s). We quote from [Wendland (2005a)]:
Theorem
Let be a cube in Rs . Suppose that = (k k) is a strictly

conditionally positive definite radial function such that = ( )
satisfies | (`) (r )| `!M ` for all integers ` `0 and all r 0, where M
is a fixed positive constant.
Then there exists a constant c such that for any f N ()
c
kf Pf kL () e hX , |f |N () , (2)
for all data sites X with sufficiently small fill distance hX , .

If satisfies even | (`) (r )| M ` , then for sufficiently small hX ,
c| log hX , |
hX ,
kf Pf kL () e kf kN () . (3)
Example
For Gaussians
2 kxk2
(x) = e , > 0 fixed,
2
we have (r ) = e r , so that
2r
(`) (r ) = (1)` 2` e for ` `0 = 0.
Thus, M = 2 , and the error bound corresponding to (3) applies.

This kind of exponential approximation order is usually referred to as
spectral (or even super-spectral) approximation order.
Remark
We emphasize that this nice property holds only in the non-stationary
setting and for data functions f that are in the native space of the
Gaussians such as band-limited functions.

Example
For generalized (inverse) multiquadrics
(x) = (1 + kxk2 ) , < 0, or 0 <

/ N,
we have (r ) = (1 + r ) . In this case one can show that
| ` (r )| `!M ` whenever ` de.
Here M = 1 + | + 1|. Therefore, the error estimate corresponding to

(2) applies, i.e., in the non-stationary setting generalized (inverse)
multiquadrics have spectral approximation order.

Example
For Laguerre-Gaussians
s/2 2 kxk2
(x) = Ln (kxk2 )e , > 0 fixed,
we have
s/2 2r
(r ) = Ln (2 r )e
and the derivatives (`) will be bounded by (`) (0) = pn (`)2` , where
pn is a polynomial of degree n.
Thus, the approximation power of Laguerre-Gaussians falls between
(3) and (2) and Laguerre-Gaussians have at least spectral
approximation power.

Native Space Error Bounds for Specific Basis Functions Basis Functions with Finite Smoothness
For functions with finite smoothness such as the

Matrn functions,
radial powers,
thin plate splines, and
Wendlands compactly supported functions
it is possible to bound the constant C (x) by some additional powers
of h, and thereby to improve the estimate
k ||
p
|D f (x) D Pf (x)| ChX , C (x)|f |N ()
discussed in Chapter 14.

In particular, for C k functions the factor C (x) can be expressed as
C (x) = max kD kL (B(0,2chX , ))

||=2k
independent of x (see [Wendland (2005a)]).
Therefore, this results in the following error estimates (see, e.g.,

[Wendland (2005a)], or the much earlier [Wu and Schaback (1993)]
where other proof techniques were used).

Example
For the Matrn functions

s
K s (kxk)kxk 2 s
2
(x) = , > ,
21 () 2
we get for bounded domains

s ||
|D f (x) D Pf (x)| ChX ,2 |f |N () (4)
provided || d s+1
2 e, hX , is sufficiently small, and f N ().

Example
For the compactly supported Wendland functions
s,k (x) = s,k (kxk)
this first refinement leads to

k + 1 ||
|D f (x) D Pf (x)| ChX ,2 kf kN () (5)
provided is a bounded domain, || k , hX , is sufficiently small,

and f N ().

Example
For the radial powers
(x) = (1)d/2e kxk , 0<

/ 2N,
we get

||
|D f (x) D Pf (x)| ChX2 , |f |N () (6)
de1
provided is a bounded domain, || 2 , hX , is sufficiently
small, and f N ().

Example
For thin plate splines
(x) = (1)k +1 kxk2k log kxk,
we get
k ||
|D f (x) D Pf (x)| ChX , |f |N () (7)
provided is a bounded domain, || k 1, hX , is sufficiently small,
and f N ().

Improvements for Native Space Error Bounds
Radial powers and thin plate splines can be interpreted as a

generalization of univariate natural splines.
Therefore, we know that the approximation order estimates obtained

via the native space approach are not optimal.
Example
For interpolation with univariate piecewise linear splines on a bounded
interval we know the approximation order to be O(h), whereas using
(x) = |x|, x R
the estimate (6) yields only approximation order O(h1/2 ).

Example
Similarly, for thin plate splines
(x) = kxk2 log kxk
one would expect order O(h2 ) in the case of pure, i.e., || = 0, function
approximation.
However, the estimate (7) yields only O(h).
Remark
These two examples suggest that it should be possible to double the
approximation orders obtained thus far.

One can improve the estimates for functions with finite smoothness
(i.e., Matrn functions, Wendland functions, radial powers, and thin
plate splines) by either (or both) of the following two ideas:
by requiring the data function f to be even smoother than what the
native space prescribes, i.e., by building certain boundary
conditions into the native space;
by using weaker norms to measure the error.

The idea to localize the data by adding boundary conditions was

introduced in the paper [Light and Wayne (1998)].
This trick allows us to square the approximation order, and thus

reach the expected approximation orders.
The second idea to weaken the norm is rather obvious and can
already be found in the early paper [Duchon (1978)].

Example
After applying both of these techniques the final approximation order
estimate for interpolation with the compactly supported functions s,k
on a bounded domain is (see [Wendland (1997)])
2k +s+1
kf Pf kL2 () ChX , kf kW 2k +s+1 (Rs ) , (8)
2
k + s+1
where we assume f W22k +s+1 (Rs ) W2 2
(Rs ) = N (Rs ).
For example, for the popular basic function
3,1 (r ) = (1 r )4+ (4r + 1)
we have with s = 3 and k = 1

6
kf Pf kL2 () ChX , kf kW 6 (R3 ) .
2
Note that the numerical experiments in Chapter 12 produced

RMS-convergence rates only as high as 4.5.
Example (cont.)
The L error for Wendland functions s,k (without the L2 relaxation) is
2k +1+ 2s
kf Pf kL () ChX , kf kW 2k +s+1 (Rs ) ,
2
k + s+1
where f needs to be such that it not only lies in N () = W2 2 (),
but also has an extension to the Sobolev space W22k +s+1 (Rs ) that can
be written as a convolution of an appropriate L2 function supported in
with the basic function s,k .
For the specific example of 3,1 (r ) = (1 r )4+ (4r + 1) on R3 we get
4.5
|f (x) Pf (x)| ChX , kf kW 6 (R3 ) 2
which is more in line with the numerical evidence obtained earlier (see
[Wendland (1997)] for details).

Example
For radial powers one obtains L2 -error estimates of order O(h+s ).
For thin plate splines one obtains L2 -error estimates of order
O(h2k +s ).
These estimates are optimal, i.e., exact approximation orders, as

shown in [Bejancu (1999)].
Remark
More work on improved error bounds can be found in, e.g.,
[Johnson (2004)] or [Schaback (1999)].

Remark
Since for the examples of piecewise linear splines and thin plate
splines discussed above our kernels (for interpolation on bounded
domains) were given by full space Greens functions, another
possibility to obtain the full approximation order on bounded domains
may be to instead use a Greens function for the bounded domain with
appropriate boundary conditions added.

Error Bounds for Functions Outside the Native Space
The error bounds mentioned so far were all valid under the assumption
that the function f providing the data came from (a subspace of) the
native space of the RBF employed in the interpolation.
We now mention a few recent results that provide error bounds for
interpolation of functions f not in the native space of the basic function.
In particular, the case when f lies in some Sobolev space that is larger
than the native space is of great interest.
A rather general theorem (sometimes referred to as a sampling
inequality) was recently given in [Narcowich et al. (2005)].
In this theorem Narcowich, Ward and Wendland provide Sobolev
bounds for functions with many zeros. However, since the interpolation
error function is just such a function, these bounds have a direct
application to our situation.
We point out that this theorem again applies to the non-stationary
setting.
Theorem
Let k be a positive integer, 0 < 1, 1 p < , 1 q and let
be a multi-index satisfying k > || + ps or, for p = 1, k || + s. Let
X be a discrete set with fill distance h = hX , where is a
compact set with Lipschitz boundary which satisfies an interior cone
condition. If u Wpk + () satisfies u|X = 0, then
k +||s( p1 q1 )+
|u|W || () ch |u|W k + () ,
q p
where c is a constant independent of u and h, and (x)+ is the cutoff

function.
Here | |W || () is actually a norm for = 0, but only a semi-norm in
q
general.

Suppose we have an interpolation process
P : Wpk + () V
that maps Sobolev functions to a finite-dimensional subspace V of

Wpk + () with the additional property
|Pf |W k + () |f |W k + () ,
p p
then the previous theorem immediately yields the error estimate

k +||s( p1 q1 )+
|f Pf |W || () ch |f |W k + () .
q p
The additional property |Pf |W k + () |f |W k + () is certainly satisfied

p p
provided the native space of the basic function is a Sobolev space.

Thus, the use of the sampling inequality provides an alternative to the

power function approach discussed in the previous chapter if we base
P on linear combinations of shifts of the basic function .
This new approach has the advantage that the term C (x) which may
depend on both and X no longer needs to be dealt with.

In particular, the authors of [Narcowich et al. (2005)] show that if the

Fourier transform of satisfies
c1 (1 + kk22 ) ()
c2 (1 + kk22 ) , kk , Rs , (9)
then the estimate

||s( 21 q1 )+
|u|W || () ch |u|W2 () ,
q
holds provided the fill distance is sufficiently small.

Example
Examples of basic functions with an appropriately decaying Fourier
transform are provided by the families of
Wendland, or
Matrn functions.
Analogous error bounds also hold for

radial powers and
thin plate splines
since their native spaces are Beppo-Levi spaces.

For functions f outside the native space of a basic function whose

Fourier transform satisfies (9) Narcowich, Ward and Wendland prove
Theorem
Let k and n be integers with 0 n < k and k > 2s , and let
f C k (). Also suppose that X = {x 1 , . . . , x N } satisfies
diam(X ) 1 with sufficiently small fill distance. Then for any
1 q we have
k ns( 21 q1 )+
|f Pf |Wqn () cXk h kf kC k () ,
h
where X = qX is the mesh ratio for X and qX is the separation
1
distance 2 mini6=j kx i x j k2 .

Recall that
the fill distance corresponds to the radius of the largest possible empty
ball that can be placed between the points in X , while
the separation distance (c.f. Chapter 16), on the other hand, can be
interpreted as the radius of the largest ball that can be
placed around every point in X such that no two balls
overlap. Thus,
the mesh ratio is a measure of the non-uniformity of the distribution of
the points in X .
Similar results were obtained earlier in [Brownlee and Light (2004)] (for
radial powers and thin plate splines only), and in [Yoon (2003)] (for
shifted surface splines, see below).

Example
If we consider polyharmonic splines, then the decay condition (9) for
the Fourier transform is satisfied with
= 2 for thin plate splines and with
= for radial powers.
If we take k = , n = 0, and q = in the error estimate
k ns( 12 q1 )+
|f Pf |Wqn () cXk h kf kC k () ,
then we arrive,
for thin plate splines (x) = kxk2 log(kxk), at the bound
s
|f Pf |L ch2 2 kf kC 2 ()
and, for radial powers (x) = kxk , at
s
|f Pf |L ch 2 kf kC () .
These bounds immediately correspond to the optimal native space

bounds obtained earlier only after the improvements discussed in the
previous subsection.
For data functions f with less smoothness the approximation order is

reduced accordingly.

Lower bounds on the approximation order for approximation by

polyharmonic splines were provided in [Maiorov (2005)].
Maiorov studies for any 1 p, q and s > ( p1 q1 )+ the error E of
Lq -approximation of Wp functions by polyharmonic splines.
More precisely,

E(Wp ([0, 1]s ), RN ( , ), Lq ([0, 1]s )) cN s ,
where RN ( , ) denotes the linear space formed by all possible linear
combinations of N polyharmonic (or thin-plate type) splines
(
r 2s if s is odd s
(r ) = 2s
> ,
r log r if s is even, 2
and multivariate polynomials of degree at most 1.

Note that these bounds are in terms of the number N of data sites
instead of the usual fill distance h.
For the special cases p = q = and p = 2, 1 q 2 the above
lower bound is shown to be asymptotically exact.
Recently (see [Rieger and Zwicknagl (2010)]), the use of sampling

inequalities on bounded domains was extended to the case of infinitely
smooth functions.
For Gaussians and (inverse) multiquadrics the authors recover the

exponential convergence rates discussed at the beginning of this
chapter.
However, the results in [Rieger and Zwicknagl (2010)] also cover

approximation of derivatives,
the case of ridge regression (see Chapter 19), and
provide guidance on how to choose the smoothing parameter so
that the approximation order of the interpolant is preserved under
smoothing.

Error Bounds for Stationary Approximation
The stationary setting is a natural approach for use with local

basis functions.
The main motivation comes from the computational point of view.
We are interested in maintaining sparse interpolation matrices as
the density of the data increases.
This can be achieved by scaling the basis functions proportional to
the data density.
In principle we can take any of our basic functions and apply a
scaling of the variable, i.e., we replace x by x, > 0.
As mentioned several times earlier, this scaling results in
peaked or narrow basis functions for large values of , and
flat basis functions for 0.

We will now discuss what happens if we choose inversely

proportional to the fill distance, i.e.,
0
h = (10)
hX ,
for some fixed base scale 0 and study the approximation error based
on the RBF interpolant
N
X
Pf (x) = cj h (kx x j k),
j=1
where
h = (h ).

Example
A rather disappointing fact is that Gaussians do not provide any
positive approximation order, i.e., the approximation process is
saturated.
This was studied by [Buhmann (1989)] on infinite lattices.
For quasi-interpolation the approximate approximation approach of
Mazya shows that it is possible to choose 0 in such a way that the
level at which the saturation occurs can be controlled (see, e.g.,
[Mazya and Schmidt (1996), Mazya and Schmidt (2007)]).
Therefore, Gaussians may very well be used for stationary
interpolation provided an appropriate initial shape parameter is
chosen.
Remark
We will illustrate this behavior in the next chapter.
The same kind of argument also applies to the Laguerre-Gaussians of
Chapter 4.
Example
Basis functions with compact support such as the Wendland functions
also do not provide any positive approximation order in the stationary
case.
This can be seen by looking at the power function for the scaled basic
function h = (h ) which is of the form
Ph ,X (x) = P,Xh (h x),
where Xh = {h x 1 , . . . , h x N } denotes the scaled data set.

Moreover, the fill distances of the sets Xh and X satisfy
hXh , = h hX , .
h

Example (cont.)
Therefore, the power function (which can be bounded in terms of the
fill distance) satisfies

Ph ,X (x) C h hX ,
h
for some > 0.

This, however, does not go to zero if h is chosen as in (10), i.e.,
h = hX0, .
Remark
If, on the other hand, we work in the approximate approximation
regime, then we can obtain good convergence in many cases (see the
next chapter for some numerical experiments).

Example
Stationary interpolation with (inverse) multiquadrics, radial powers and
thin plate splines presents no difficulties.
In fact, [Schaback (1995b)] shows that the native space error bound for
thin plate splines and radial powers is invariant under a stationary
scaling.
Therefore, the non-stationary bound of the theorem from

[Narcowich et al. (2005)] (see above) applies in the stationary case
also.
The advantage of scaling thin plate splines or radial powers comes

from the added stability one can gain by preventing the separation
distance from becoming too small (see Chapter 16 and the work of
Iske on local polyharmonic spline approximation, e.g., [Iske (2004)]).

Example (cont.)
Yoon provides error estimates for stationary approximation of rough
functions (i.e., functions that are not in the native space of the basic
function) by so-called shifted surface splines.
Shifted surface splines are of the form
(1)ds/2e (1 + kxk2 )s/2 ,

s odd,
(x) =
(1)s/2+1 (1 + kxk2 )s/2 log(1 + kxk2 )1/2 , s even,
where s/2 < N.
Remark
These functions include all of the (inverse) multiquadrics, radial powers
and thin plate splines.

Example (cont.)
Yoon has the following theorem (see [Yoon (2003)] for the Lp case, and
[Yoon (2001)] for L bounds only).
Theorem
Let be a shifted surface spline with shape parameter inversely
proportional to the fill distance hX , . Then there exists a positive
constant C (independent of X ) such that for every f in the Sobolev
space W2 () W
() we have
kf Pf kLp () Chp |f |W (Rs ) , 1 p ,

2
with
s
p = min{, 2 + ps }.
() with max{0, s s } < < , then
Furthermore, if f W2 () W 2 p
kf Pf kLp () = o(hp + ).
Example (cont.)
Yoons estimates address the question
How well do the infinitely smooth (inverse) multiquadrics
approximate functions that are less smooth than those in their
native space?
For example, Yoons theorem states that L2 -approximation

p to functions
in W22 (), Rs , by multiquadrics (x) = 1 + kxk2 is of the
order O(h2 ).
However, we emphasize once more that this refers to stationary

approximation of rough functions, i.e., is scaled inversely proportional
to the fill distance and f need not lie in the native space of , whereas
the spectral order given in (2) corresponds to approximation of
functions in the native space in the non-stationary case with fixed .

Remark
For thin plate splines and radial powers the approximation orders
in Yoons theorem are equivalent to those of the theorem from
[Narcowich et al. (2005)] and the results of Brownlee and Light
mentioned above.
This is to be expected due to the invariance of these basic functions
with respect to scaling.
The second part of Yoons result is a step toward exact
approximation orders as is the work of [Maiorov (2005)] and
[Bejancu (1999)] mentioned above.

Convergence With Respect to the Shape Parameter
None of the error bounds discussed thus far have taken into
account the possibility of varying the shape parameter for a fixed
data set X .
However, in the literature the infinitely smooth basic functions
such as the Gaussians and (inverse) multiquadrics are usually
formulated including the shape parameter (or another parameter
equivalent to it) and one may wonder how a change in this shape
parameter affects the convergence properties of the RBF
interpolant.
In fact, quite a bit of work has been spent on the quest for the
optimal shape parameter (see, e.g.,
[Carlson and Foley (1991), Fasshauer and Zhang (2007),
Foley (1994), Hagan and Kansa (1994), Luh (2010a),
Kansa and Carlson (1992), Rippa (1999), Tarwater (1985),
Wertz et al. (2006)]).

Convergence With Respect to the Shape Parameter
Convergence of the infinitely smooth Gaussians and (inverse)

multiquadrics with respect to the shape parameter was studied early
on in [Madych (1991)].
Madych showed that for these basic functions there exists a positive
constant < 1 such that
|f (x) Pf (x)| C1/(hX , ) (11)
provided f is in the native space of .
Remark
This estimate shows that taking either the shape parameter or
the fill distance hX , to zero results in exponential convergence.
However, numerical experiments as well as a more careful
theoretical analysis (see [Luh (2010b)]) show that the constant C
is not independent of , so that there may be a minimal error for a
positive value of .
Dimension-Independent Error Bounds
In this section we mention some new results (see

[Fasshauer et al. (2010)] for much more details) on the rates of
convergence of Gaussian kernel approximation.
We address weighted L2 approximation when the data is specified
either by function values of an unknown function f (from the native
space of the kernel) or with the help of arbitrary linear functionals.
The convergence results pay special attention to the dependence of
the estimates on the space dimension s.
We will see that the use of anisotropic Gaussian kernels instead of
isotropic ones provides improved convergence rates.

Dimension-Independent Error Bounds The current situation
The error bounds in the previous sections were formulated in terms of

the fill distance, while the results we mention below are in terms of N,
the number of data.
Therefore, we restate some of the earlier bounds in terms of N using
the fact that for quasi-uniformly distributed data sites we have
hX , = O(N 1/s ).
If f has derivatives up to total order `, then the first error bound for
Gaussians kernels K was of the form
kf Pf k C`,s N `/s kf kNK ()
where we now emphasize the possible dimension-dependence of the

constant C`,s .
This bound shows that infinitely smooth functions can be approximated
with order ` = .

Dimension-Independent Error Bounds The current situation
With extra effort we obtained the spectral estimate

c 1/s
kf Pf k e s N log N
kf kNK () .
Remark
Both of these bounds show that the rate of convergence
deteriorates as s increases.
Moreover, the dependence of the constants on s is not clear.
Therefore, these kinds of error bounds and in fact almost all
error bounds in the RBF literature suffer from the curse of
dimensionality.
We will now present some results from [Fasshauer et al. (2010)]
on dimension-independent convergence rates for Gaussian kernel
approximation.

Dimension-Independent Error Bounds New results on (minimal) worst-case weighted L2 error
We make several assumptions in order to be able to obtain

dimension-independent error bounds.
We define the worst-case weighted L2, error as
errwc
2, = sup kf Pf k2, ,
kf kN s 1
K (R )
where Pf is our kernel (minimum norm) approximation calculated in

the usual way.
Therefore
kf Pf k2, errwc
2, kf kNK (Rs ) for all f NK (Rs ).
The N th minimal worst case error errwc

2, (N) refers to the worst case
error that can be achieved with an optimal design, i.e., data generated
by N optimally chosen linear functionals.

For function approximation this means that the data sites have to
be chosen in an optimal way.
The results in [Fasshauer et al. (2010)] are non-constructive, i.e.,
no such optimal design is specified.
However, a Smolyak or sparse grid algorithm is a natural candidate
for such a design.
If we are allowed to choose arbitrary linear functionals, then the
optimal choice for weighted L2 approximation is known.
In either case we will need an eigenfunction expansion of the
Gaussian kernel.

Following [Rasmussen and Williams (2006)], the eigenvalues and

2 2
eigenfunctions of the univariate Gaussian kernel K (x, z) = e (xz)
are:
2(k 1)
k = k 1 , k = 1, 2, . . . ,
1
2
2 2
2 (1 + 1 + 4 ) +
and eigenfunctions
s !
(1 + 42 )1/4 2 x 2
2 1/4

k (x) = exp Hk 1 (1 + 4 ) x ,
2k 1 (k 1)! 1 2
2 (1 + 1 + 4 )
where Hk are the classical Hermite polynomials of degree k , i.e.,

2 dk x 2
Hk (x) = (1)k ex e for all x R, k = 0, 1, 2, . . .
dx k
so that

Z
2
Hk2 (x) ex dx = 2k k ! for k = 0, 1, 2, . . . .
R

Remark
The multivariate (and anisotropic) case can be handled using products
of univariate eigenvalues and eigenfunctions.
For details see [Fasshauer et al. (2010)] or [Fasshauer (2010)].

For weighted L2 approximation we use generalized Fourier

coefficients, i.e., the optimal linear functionals are
Lk = h, k iNK (Rs ) , where the k are the eigenfunctions of K .
We obtain the truncated generalized Fourier series approximation
N
X
Pf (x) = hf , j iNK (Rs ) j (x) for all f NK (Rs ),
j=1
where

X Z
K (x, y) = k k (x)k (y), K (x, y)k (y)(y)dy = k k (x).
k =1
It is then known [Novak and Wozniakowski (2008)] that

errwc
p
2, (N) = N+1 ,
the (N + 1)st largest eigenvalue, which is easy to identify in the
univariate case, but takes some care to specify in the multivariate
setting.
In [Fasshauer et al. (2010)] it is then proved that in the isotropic

case, i.e., with a truly radial Gaussian kernel of the form
2 kxyk2
K (x, y) = e
one can approximate

function data with an N th minimal error of the order O(N 1/4+ ),
and
Fourier data (i.e., arbitrary linear functional data) with an N th
minimal error of the order O(N 1/2+ ).
Here the constants in the O-notation do not depend on the
dimension s and is arbitrarily small.

With anisotropic kernels, i.e.,

2 2 ...2 (x z )2
K (x, y) = e1 (x1 z1 ) s s s
one can do much better.

If the shape parameters decay like ` = ` , then one can
approximate
function data with an N th minimal error of the order
2
O(N max( 2+ ,1/4)+ ), and
Fourier data (i.e., arbitrary linear functional data) with an N th
minimal error of the order O(N max(,1/2)+ ).
Again, the constants in the O-notation do not depend on the
dimension s.

Remark
Even if we do not have an eigenfunction expansion of a specific
kernel available, the work of [Fasshauer et al. (2010)] shows that
for any radial (isotropic) kernel one has a dimension-independent
Monte-Carlo type convergence rate of O(N 1/2+ ) provided
arbitrary linear functionals are allowed to generate the data.
For translation-invariant (stationary) kernels the situation is similar.
However, the constant in the O-notation depends in any case
on the sum of the eigenvalues of the kernel. For the radial case
this sum is simply (0) (independent of s), while for general
translation invariant kernels it is K
e (0), which may depend on s.

Remark
These results show that even though RBF methods are often
advertised as being dimension-blind their rates of
convergence are only excellent (i.e., spectral for infinitely smooth
kernels) if the dimension s is small.
For large dimensions the constants in the O-notation take over.
If one, however, permits an anisotropic scaling of the kernel (i.e.,
elliptical symmetry instead of strict radial symmetry) and if those
scale parameters decay rapidly with increasing dimension, then
excellent convergence rates for approximation of smooth functions
can be maintained independent of s.

Polynomial Interpolation as the Limit of RBF Interpolation Infinitely Smooth RBFs
Recently, a number of authors (see, e.g.,

[Driscoll and Fornberg (2002), Fornberg and Flyer (2005),
Fornberg and Wright (2004), Larsson and Fornberg (2005),
Lee et al. (2007), Schaback (2005), Schaback (2008)]) have studied
the limiting case as 0 of scaled radial basis function interpolation
with infinitely smooth basic functions such as Gaussians and
generalized (inverse) multiquadrics.
It turns out that there is an interesting connection to polynomial

interpolation.

In [Driscoll and Fornberg (2002)] univariate (s = 1) interpolation with

-scaled infinitely smooth radial basic functions is studied.
Driscoll and Fornberg show that the RBF interpolant

N
X
Pf (x) = cj (k(x xj )k), x [a, b] R,
j=1
to function values at N distinct data sites tends to the Lagrange

interpolating polynomial of f as 0.

The multivariate case is more complicated.
However, the limiting RBF interpolant is given by a low-degree

multivariate polynomial (see [Larsson and Fornberg (2005),
Lee et al. (2007), Schaback (2005), Schaback (2008)]).
For example, if the data sites are located such that they guarantee a
unique polynomial interpolant, then the limiting RBF interpolant is
given by this polynomial.
If polynomial interpolation is not unique, then the RBF limit is still a

polynomial whose form depends on the choice of basic function.

Theorem (Driscoll, Fornberg, Larsson, Schaback, Yoon [2002-08])

Assume the strictly positive definite radial kernel has an expansion

X
(r ) = aj r 2j
j=0
into even powers of r (i.e., is infinitely smooth), and that the data X
are unisolvent with respect to any set of N linearly independent
polynomials of degree at most m. Then
lim Pf (x) = pm,f (x), x Rs ,

0
where pm,f is determined as follows:

If interpolation with polynomials of degree at most m is unique,
then pm,f is that unique polynomial interpolant.
If interpolation with polynomials of degree at most m is not unique,
then pm,f is a polynomial interpolant whose form depends on the
choice of RBF.
Remark
These statements require the RBFs to satisfy a condition on
certain coefficient matrices Ap,J . This condition was left unproven
in [Larsson and Fornberg (2005)] and verified in
[Lee et al. (2007)].
In [Larsson and Fornberg (2005)] the authors also provide an

explanation for the error behavior for small values of the shape
parameter, and for the existence of an optimal (positive) value of
giving rise to a global minimum of the error function. Recent
theoretical work in [Luh (2010a), Luh (2010b)] also goes in this
direction.
For the special case of scaled Gaussians [Schaback (2005)]

shows that as 0 the RBF interpolant converges to the de Boor
and Ron least polynomial interpolant (see [de Boor (1992),
de Boor and Ron (1990), de Boor and Ron (1992)] and also
[de Boor (2006)]).
Remark
In [Fornberg and Wright (2004)] the authors describe a so-called
Contour-Pad algorithm that makes it possible (for data sets of
relatively modest size) to compute the RBF interpolant for all
values of the shape parameter including the limiting case 0.
We present some numerical result obtained with Grady Wrights

M ATLAB toolbox in Chapter 17.
Other recent work obtaining RBF interpolants close to the

polynomial limit, i.e., for small , is
[Fasshauer and McCourt (2010), Fornberg et al. (2009)].
We will later exploit the connection between RBF and polynomial

interpolants to design numerical solvers for partial differential
equations.

Polynomial Interpolation as the Limit of RBF Interpolation RBFs with Finite Smoothness
To our knowledge, the flat limit of RBFs with finite smoothness

was not studied until the recent paper [Song at al. (2009)] in which
interpolation on Rd was investigated.
Before we explain the results obtained in [Song at al. (2009)], we

look at a few finitely smooth radial kernels as full space Greens
functions.
The connection between (conditionally) positive definite kernels

and Greens functions is currently being studied by Qi Ye (see
[Fasshauer and Ye (2010a), Fasshauer and Ye (2010b),
Ye (2010)]).

Example (Radial kernels with finite smoothness)

.
The univariate C 0 Matrn kernel K (x, z) = e|xz| is the
full-space Greens function for the differential operator
d2
L= + 2 I.
dx 2
On the other hand, it is well-known that univariate C 0 piecewise
linear splines may be expressed in terms of kernels of the form
.
K (x, z) = |x z|. The corresponding differential operator is
d2
L= .
dx 2
Note that the differential operator for the Matrn kernel
converges to that of the piecewise linear splines as 0.

Example (cont.)
The univariate C 2 tension spline kernel [Renka (1987)]
.
K (x, z) = e|xz| + |x z| is the Greens kernel of
d4 2 d
2
L= + ,
dx 4 dx 2
.
while the univariate C 2 cubic spline kernel K (x, z) = |x z|3
corresponds to
d4
L = 4.
dx
Again, the differential operator for the tension spline converges
to that of the cubic spline as 0.

Example (cont.)
In [Berlinet and Thomas-Agnan (2004)] we find a so-called
univariate Sobolev kernel of the form
.
K (x, z) = e|xz| sin |x z| + 4 which is associated with

d4
L= 2 I.
dx 4
The operator for this kernel also converges to that of the cubic
spline kernel, but the effect of the scale parameter is different than
for the tension spline.
Remark
Note that this Sobolev kernel is different from the Sobolev splines
(Matrn functions) discussed earlier terminology . . .

Example (cont.)
The general multivariate Matrn kernels are of the form
. s
K (x, y) = Kms/2 (kx yk) (kx yk)ms/2 , m> ,
2
and can be obtained as Greens kernels of (see [Ye (2010)])
m s
L = 2 I , m> .
2
We contrast this with the polyharmonic spline kernels
(
. kx yk2ms , s odd,
K (x, y) = 2ms
kx yk log kx yk, s even,
and
s
L = (1)m m , m> .
2

All examples above show that the differential operators associated

with finitely smooth RBF kernels converge to those of a
piecewise polynomial or polyharmonic spline kernel as 0.
We therefore ask if RBF interpolants based on finitely smooth

kernels converge to (polyharmonic) spline interpolants for 0
as is the case for infinitely smooth radial kernels and polynomials.
As mentioned above, infinitely smooth radial kernels can be

expanded into an infinite series of even powers of r .
Finitely smooth radial kernels can also be expanded into an

infinite series of powers of r .
In this case there always exists some minimal odd power of r with
nonzero coefficient indicating the smoothness of the kernel.

Example
For univariate C 0 , C 2 and C 4 Matrn kernels, respectively, we have
.
(r ) = er
1 1
= 1 r + (r )2 (r )3 + ,
2 6
. r
(r ) = (1 + r )e
1 1 1
= 1 (r )2 + (r )3 (r )4 + ,
2 3 8
. 2 r
(r ) = 3 + 3r + (r ) e
1 1 1 1
= 3 (r )2 + (r )4 (r )5 + (r )6 + .
2 8 15 48

Theorem ([Song at al. (2009)])
Suppose is radial and conditionally positive definite of order m n

with an expansion of the form
(r ) = a0 + a2 r 2 + . . . + a2n r 2n + a2n+1 r 2n+1 + a2n+2 r 2n+2 + . . . ,
where 2n + 1 denotes the smallest odd power of r present in the

expansion (i.e., is finitely smooth). Also assume that the data X
contain a unisolvent set with respect to the space s2n of s-variate
polynomials of degree less than 2n. Then
N
X M
X
lim Pf (x) = cj kx x j k2n+1 + dk pk (x), x Rs ,
0
j=1 k =1
where {pk : k = 1, . . . , M} denotes a basis of sn .

Remark
The previous theorem does not cover Matrn kernels with odd-order
smoothness. However, all other examples listed above are covered.
1.6
1.6
1.4
1.4
1.2 1.2
1 1
0.8 0.8
0.6 0.6 data

data
=2
=2
0.4 0.4 =1
=1
=0.1
0.2 =0.1
0.2 cubic spline
piecewise linear
0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Figure: Convergence of C 0 (left) and C 2 (right) Matrn interpolants to

piecewise linear (left) and cubic (right) spline interpolants.

Appendix References
References I
Berlinet, A. and Thomas-Agnan, C. (2004).

Reproducing Kernel Hilbert Spaces in Probability and Statistics.
Kluwer, Dordrecht.
Buhmann, M. D. (2003).
Radial Basis Functions: Theory and Implementations.
Cambridge University Press.
Fasshauer, G. E. (2007).
Meshfree Approximation Methods with M ATLAB.
World Scientific Publishers.
Iske, A. (2004).
Multiresolution Methods in Scattered Data Modelling.
Lecture Notes in Computational Science and Engineering 37, Springer Verlag
(Berlin).

Appendix References
References II
Mazya, V. and Schmidt, G. (2007).

Approximate Approximations.
Mathematical Surveys and Monographs, vol. 141, Americal Mathematical Society
(Providence, RI).
Novak, E. and Wozniakowski, H. (2008).
Tractability of Multivariate Problems, Volume 1: Linear Information.
EMS Tracts in Mathematics, no. 6, European Mathematical Society.
Rasmussen, C. E. and Williams, C. (2006).
Gaussian Processes for Machine Learning.
MIT Press (online version at http://www.gaussianprocess.org/gpml/).
Wendland, H. (2005a).
Scattered Data Approximation.
Cambridge University Press (Cambridge).
Bejancu, A. (1999).
Local accuracy for radial basis function interpolation on finite uniform grids.
J. Approx. Theory 99, pp. 242257.

Appendix References
References III
de Boor, C. (1992).
On the error in multivariate polynomial interpolation.
Applied Numerical Mathematics 10, pp. 297305.
de Boor, C. (2006).
On interpolation by radial polynomials.
Adv. in Comput. Math. 24, pp. 143153.
de Boor, C. and Ron, A. (1990).
On multivariate polynomial interpolation.
Constr. Approx. 6, pp. 287302.
de Boor, C. and Ron, A. (1992).
The least solution for the polynomial interpolation problem.
Math. Z. 210, pp. 347378.
Brownlee, R. and Light, W. (2004).
Approximation orders for interpolation by surface splines to rough functions.
IMA J. Numer. Anal. 24, pp. 179192.

Appendix References
References IV
Buhmann, M. D. (1989).
Multivariate interpolation using radial basis functions.
Ph.D. Dissertation, University of Cambridge.
Buhmann, M. D. and Dyn, N. (1991).
Error estimates for multiquadric interpolation.
in Curves and Surfaces, P.-J. Laurent, A. Le Mhaut, and L. L. Schumaker
(eds.), Academic Press (New York), pp. 5158.
Carlson, R. E. and Foley, T. A. (1991).
The parameter R 2 in multiquadric interpolation.
Comput. Math. Appl. 21, pp. 2942.
Driscoll, T. A. and Fornberg, B. (2002).
Interpolation in the limit of increasingly flat radial basis functions.

Appendix References
References V
Duchon, J. (1978).
Sur lerreur dinterpolation des fonctions de plusieurs variables par les
D m -splines.
Rev. Francaise Automat. Informat. Rech. Opr., Anal. Numer. 12, pp. 325334.
Fasshauer, G. E. (2010).
Greens functions: taking another look at kernel approximation, radial basis
functions and splines.
Submitted.
Fasshauer, G. E., Hickernell, F. J. and Wozniakowski, H. (2010).
Rate of convergence and tractability of the radial function approximation problem.
Submitted.
Fasshauer, G. E. and McCourt, M. J. (2010).
Stable evaluation of Gaussian RBF interpolants.
In preparation.

Appendix References
References VI
Fasshauer, G. E. and Ye, Q. (2010).

Reproducing kernels of generalized Sobolev spaces via a Green function
approach with distributional operators.
Submitted.
Fasshauer, G.E. and Ye, Q. (2010).
Reproducing kernels of Sobolev spaces in bounded domains via a Greens
kernel approach.
In preparation.
Fasshauer, G. E. and Zhang, J. G. (2007).
On choosing optimal shape parameters for RBF approximation.
Numer. Algorithms 45, 345368.
Foley, T. A. (1994).
Near optimal parameter selection for multiquadric interpolation.
J. Appl. Sc. Comp. 1, pp. 5469.

Appendix References
References VII
Fornberg, B. and Flyer, N. (2005).

Accuracy of radial basis function interpolation and derivative approximations on
1-D infinite grids.
Adv. Comput. Math. 23 1-2, pp. 520.
Fornberg, B., Larsson, E. and Flyer, N. (2009).
Stable computations with Gaussian radial basis functions in 2-D.
Technical Report 2009-020, Uppsala University, Department of Information
Technology.
Fornberg, B. and Wright, G. (2004).
Stable computation of multiquadric interpolants for all values of the shape
parameter.
Hagan, R. E. and Kansa, E. J. (1994).
Studies of the R parameter in the multiquadric function applied to ground water
pumping.
J. Appl. Sci. Comput. 1, pp. 266281.

Appendix References
References VIII
Johnson, M. J. (2004).
An error analysis for radial basis function interpolation.
Numer. Math. 98 4, pp. 675694.
Kansa, E. J. and Carlson, R. E. (1992).
Improved accuracy of multiquadric interpolation using variable shape parameters.

Larsson, E. and Fornberg, B. (2005).
Theoretical and computational aspects of multivariate interpolation with
increasingly flat radial basis functions.
Lee, Y. J., Yoon, G. J. and Yoon, J. (2007).
Convergence of increasingly flat radial basis interpolants to polynomial
interpolants.
SIAM J. Math. Anal. 39, pp. 537553.

Appendix References
References IX
Light, W. A. (1996).
Variational error bounds for radial basis functions.
in Numerical Analysis 1995, D. F. Griffiths and G. A. Watson (eds.), Longman
(Harlow), pp. 94106.
Light, W. A. and Wayne, H. (1995).
Error estimates for approximation by radial basis functions.
in Approximation Theory, Wavelets and Applications, S. P. Singh (ed.), Kluwer
(Dordrecht), pp. 215246.
Light, W. A. and Wayne, H. (1998).
On power functions and error estimates for radial basis function interpolation.
Luh L.-T. (2009).
An improved error bound for multiquadric and inverse multiquadric interpolations.
Int. J. Numer. Meth. Applic. 1/2, pp. 101120.

Appendix References
References X
Madych, W. R. (1991).
Error estimates for interpolation by generalized splines.
in Curves and Surfaces, P.-J. Laurent, A. Le Mhaut, and L. L. Schumaker
(eds.), Academic Press (New York), pp. 297306.
Madych, W. R. (1992).
Miscellaneous error bounds for multiquadric and related interpolators.
Madych, W. R. and Nelson, S. A. (1988).
Multivariate interpolation and conditionally positive definite functions.
Approx. Theory Appl. 4, pp. 7789.
Madych, W. R. and Nelson, S. A. (1992).
Bounds on multivariate polynomials and exponential error estimates for
multiquadric interpolation.

Appendix References
References XI
Maiorov, V. (2005).
On lower bounds in radial basis approximation.
Adv. in Comp. Math. 22, pp. 103113.
Mazya, V. and Schmidt, G. (1996).
On approximate approximations using Gaussian kernels.
Narcowich, F. J. and Ward, J. D. (2004).
Scattered-data interpolation on Rn : error estimates for radial basis and
band-limited functions.
SIAM J. Math. Anal. 36 1, pp. 284300.
Narcowich, F. J., Ward, J. D. and Wendland, H. (2003).
Refined error estimates for radial basis function interpolation.
Constr. Approx. 19 4, pp. 541564.

Appendix References
References XII
Narcowich, F. J., Ward, J. D. and Wendland, H. (2005).

Sobolev bounds on functions with scattered zeros, with applications to radial
basis function surface fitting.
Math. Comp. 74, pp. 743763.
Renka, R. J. (1987).
Interpolatory tension splines with automatic selection of tension factors.
SIAM J. Sci. Stat. Comput. 8, pp. 393415.
Rieger, C. and Zwicknagl, B. (2010).
Sampling inequalities for infinitely smooth functions, with applications to
interpolation and machine learning.
Adv. in Comput. Math. 32/1, pp. 103129.
Rippa, S. (1999).
An algorithm for selecting a good value for the parameter c in radial basis
function interpolation.

Appendix References
References XIII
Schaback, R. (1995a).
Error estimates and condition numbers for radial basis function interpolation.
Schaback, R. (1995b).
Multivariate interpolation and approximation by translates of a basis function.
in Approximation Theory VIII, Vol. 1: Approximation and Interpolation, C. Chui,
and L. Schumaker (eds.), World Scientific Publishing (Singapore), pp. 491514.
Schaback, R. (1996).
Approximation by radial basis functions with finitely many centers.
Improved error bounds for scattered data interpolation by radial basis functions.
Math. Comp. 68 225, pp. 201216.
Multivariate interpolation by polynomials and radial basis functions.

Appendix References
References XIV
Limit problems for interpolation by analytic radial basis functions.
J. Comp. Appl. Math. 212(2), pp. 127149.
Song, G., Riddle, J., Fasshauer, G.E. and Hickernell, F.J. (2009).
Multivariate interpolation with increasingly flat radial basis functions of finite
smoothness.
Submitted.
Tarwater, A. E. (1985).
A parameter study of Hardys multiquadric method for scattered data
interpolation.
Lawrence Livermore National Laboratory, TR UCRL-563670.
Wendland, H. (1997).
Sobolev-type error estimates for interpolation by radial basis functions.
in Surface Fitting and Multiresolution Methods, A. Le Mhaut, C. Rabut, and L.
L. Schumaker (eds.), Vanderbilt University Press (Nashville, TN), pp. 337344.

Appendix References
References XV
Error estimates for interpolation by compactly supported radial basis functions of
minimal degree.
Gaussian interpolation revisited.
in Trends in Approximation Theory, K. Kopotun, T. Lyche, and M. Neamtu (eds.),
Vanderbilt University Press, pp. 417426.
Wertz, J., Kansa, E. J. and Ling, L. (2006).
The role of the multiquadric shape parameters in solving elliptic partial differential
equations.
Comput. Math. Appl. 51 8, pp. 13351348.
Wu, Z. and Schaback, R. (1993).
Local error estimates for radial basis function interpolation of scattered data.

Appendix References
References XVI
Ye, Q. (2010).
Reproducing kernels of generalized Sobolev spaces via a Green function
approach with differential operators.
Submitted.
Yoon, J. (2001).
Interpolation by radial basis functions on Sobolev space.
Yoon, J. (2003).
Lp -error estimates for shifted surface spline interpolation on Sobolev space.
Math. Comp. 72, pp. 13491367.
Luh L.-T. (2010).

The mystery of the shape parameter: Parts I-IV.
http://arxiv.org/abs/1001.5087, http://arxiv.org/abs/1002.2082,
http://arxiv.org/abs/1004.0759, http://arxiv.org/abs/1004.0761.

Appendix References
References XVII
Luh L.-T. (2010).

The shape parameter in the Gaussian function.
http://arxiv.org/abs/1006.2318.

Meshfree Chapter 15

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Meshfree Chapter 15

Hochgeladen von

Copyright:

Verfügbare Formate

MATH 590: Meshfree Methods

Chapter 15: Refined and Improved Error Bounds

Department of Applied Mathematics

fasshauer@iit.edu MATH 590 Chapter 15 1

1 Native Space Error Bounds for Specific Basis Functions

2 Improvements for Native Space Error Bounds

3 Error Bounds for Functions Outside the Native Space

4 Error Bounds for Stationary Approximation

5 Convergence With Respect to the Shape Parameter

6 Dimension-Independent Error Bounds

7 Polynomial Interpolation as the Limit of RBF Interpolation

fasshauer@iit.edu MATH 590 Chapter 15 2

For the first part of this chapter we discuss the non-stationary

fasshauer@iit.edu MATH 590 Chapter 15 4

As mentioned before, an application of the estimate

for interpolation on bounded domains Rs from Chapter 14 to

|D f (x) D Pf (x)| C` h`|| |f |NK () . (1)

A considerable amount of work has gone into investigating the

In general, the constant C` even depends on the space dimension s, in

Using different proof techniques (see [Madych and Nelson (1988)] or

for all data sites X with sufficiently small fill distance hX , .

Thus, M = 2 , and the error bound corresponding to (3) applies.

fasshauer@iit.edu MATH 590 Chapter 15 7

For generalized (inverse) multiquadrics

(x) = (1 + kxk2 ) , < 0, or 0 <

we have (r ) = (1 + r ) . In this case one can show that

| ` (r )| `!M ` whenever ` de.

Here M = 1 + | + 1|. Therefore, the error estimate corresponding to

fasshauer@iit.edu MATH 590 Chapter 15 8

fasshauer@iit.edu MATH 590 Chapter 15 9

For functions with finite smoothness such as the

discussed in Chapter 14.

fasshauer@iit.edu MATH 590 Chapter 15 10

In particular, for C k functions the factor C (x) can be expressed as

C (x) = max kD kL (B(0,2chX , ))

independent of x (see [Wendland (2005a)]).

Therefore, this results in the following error estimates (see, e.g.,

fasshauer@iit.edu MATH 590 Chapter 15 11

For the Matrn functions

we get for bounded domains

fasshauer@iit.edu MATH 590 Chapter 15 12

For the compactly supported Wendland functions

s,k (x) = s,k (kxk)

this first refinement leads to

provided is a bounded domain, || k , hX , is sufficiently small,

fasshauer@iit.edu MATH 590 Chapter 15 13

For the radial powers

(x) = (1)d/2e kxk , 0<

fasshauer@iit.edu MATH 590 Chapter 15 14

For thin plate splines

(x) = (1)k +1 kxk2k log kxk,

fasshauer@iit.edu MATH 590 Chapter 15 15

Radial powers and thin plate splines can be interpreted as a

Therefore, we know that the approximation order estimates obtained

the estimate (6) yields only approximation order O(h1/2 ).

fasshauer@iit.edu MATH 590 Chapter 15 17

(x) = kxk2 log kxk

fasshauer@iit.edu MATH 590 Chapter 15 18

fasshauer@iit.edu MATH 590 Chapter 15 19

The idea to localize the data by adding boundary conditions was

This trick allows us to square the approximation order, and thus

fasshauer@iit.edu MATH 590 Chapter 15 20

3,1 (r ) = (1 r )4+ (4r + 1)

we have with s = 3 and k = 1