Professional Documents
Culture Documents
Greg Fasshauer
Fall 2010
Outline
Our first set of error bounds will come rather naturally once we
associate with each (strictly positive definite) radial basic function a
certain space of functions called its native space.
Remark
The name reproducing kernel is inspired by the reproducing
property (2) above.
It is known that
the reproducing kernel of a Hilbert space is unique, and
It is known that
the reproducing kernel of a Hilbert space is unique, and
that existence of a reproducing kernel is equivalent to the fact that
the point evaluation functionals x are bounded linear functionals
on , i.e., there exists a positive constant M = Mx such that
|x f | = |f (x)| Mkf kH
It is known that
the reproducing kernel of a Hilbert space is unique, and
that existence of a reproducing kernel is equivalent to the fact that
the point evaluation functionals x are bounded linear functionals
on , i.e., there exists a positive constant M = Mx such that
|x f | = |f (x)| Mkf kH
x f = f (x) = hf , K (, x)iH
It is known that
the reproducing kernel of a Hilbert space is unique, and
that existence of a reproducing kernel is equivalent to the fact that
the point evaluation functionals x are bounded linear functionals
on , i.e., there exists a positive constant M = Mx such that
|x f | = |f (x)| Mkf kH
x f = f (x) = hf , K (, x)iH
It is known that
the reproducing kernel of a Hilbert space is unique, and
that existence of a reproducing kernel is equivalent to the fact that
the point evaluation functionals x are bounded linear functionals
on , i.e., there exists a positive constant M = Mx such that
|x f | = |f (x)| Mkf kH
x f = f (x) = hf , K (, x)iH
kf fn kH 0 for n
then
|f (x) fn (x)| 0 for all x .
Proof.
By Property (1) of the definition of a RKHS above K (, y) H for every
y .
Proof.
By Property (1) of the definition of a RKHS above K (, y) H for every
y .
Then the reproducing property (2) of the definition gives us
for all x, y .
Proof.
By Property (1) of the definition of a RKHS above K (, y) H for every
y .
Then the reproducing property (2) of the definition gives us
for all x, y .
This establishes (1).
Proof.
By Property (1) of the definition of a RKHS above K (, y) H for every
y .
Then the reproducing property (2) of the definition gives us
for all x, y .
This establishes (1).
Property (2) follows from (1) by the symmetry of the Hilbert space
inner product.
Proof.
By Property (1) of the definition of a RKHS above K (, y) H for every
y .
Then the reproducing property (2) of the definition gives us
for all x, y .
This establishes (1).
Property (2) follows from (1) by the symmetry of the Hilbert space
inner product.
For (3) we use the reproducing property of K along with the
Cauchy-Schwarz inequality:
Proof.
By Property (1) of the definition of a RKHS above K (, y) H for every
y .
Then the reproducing property (2) of the definition gives us
for all x, y .
This establishes (1).
Property (2) follows from (1) by the symmetry of the Hilbert space
inner product.
For (3) we use the reproducing property of K along with the
Cauchy-Schwarz inequality:
Remark
Property (1) of the previous theorem shows us what the bound Mx for
the point evaluation functional is:
Remark
Property (1) of the previous theorem shows us what the bound Mx for
the point evaluation functional is:
|f (x)| = |x f | kf kH kK (, x)kH
Remark
Property (1) of the previous theorem shows us what the bound Mx for
the point evaluation functional is:
|f (x)| = |x f | kf kH kK (, x)kH
p
= kf kH hK (, x), K (, x)iH
Remark
Property (1) of the previous theorem shows us what the bound Mx for
the point evaluation functional is:
|f (x)| = |x f | kf kH kK (, x)kH
p
= kf kH hK (, x), K (, x)iH
p
= kf kH K (x, x)
Remark
Property (1) of the previous theorem shows us what the bound Mx for
the point evaluation functional is:
|f (x)| = |x f | kf kH kK (, x)kH
p
= kf kH hK (, x), K (, x)iH
p
= kf kH K (x, x)
p
= Mx = K (x, x).
Remark
Property (1) of the previous theorem shows us what the bound Mx for
the point evaluation functional is:
|f (x)| = |x f | kf kH kK (, x)kH
p
= kf kH hK (, x), K (, x)iH
p
= kf kH K (x, x)
p
= Mx = K (x, x).
Theorem
Suppose H is a reproducing kernel Hilbert function space with
reproducing kernel K : R. Then K is positive definite.
Theorem
Suppose H is a reproducing kernel Hilbert function space with
reproducing kernel K : R. Then K is positive definite.
Theorem
Suppose H is a reproducing kernel Hilbert function space with
reproducing kernel K : R. Then K is positive definite.
Remark
Here the space of bounded linear functionals on H is known as its
dual, and denoted by H .
Proof.
Since the kernel is real-valued we can restrict ourselves to a quadratic
form with real coefficients.
Proof.
Since the kernel is real-valued we can restrict ourselves to a quadratic
form with real coefficients.
For distinct points x 1 , . . . , x N and arbitrary c RN we have
N X
X N
cj ck K (x j , x k )
j=1 k =1
Proof.
Since the kernel is real-valued we can restrict ourselves to a quadratic
form with real coefficients.
For distinct points x 1 , . . . , x N and arbitrary c RN we have
X N
N X N X
X N
cj ck K (x j , x k ) = cj ck hK (, x j ), K (, x k )iH
j=1 k =1 j=1 k =1
Proof.
Since the kernel is real-valued we can restrict ourselves to a quadratic
form with real coefficients.
For distinct points x 1 , . . . , x N and arbitrary c RN we have
X N
N X N X
X N
cj ck K (x j , x k ) = cj ck hK (, x j ), K (, x k )iH
j=1 k =1 j=1 k =1
XN N
X
= h cj K (, x j ), ck K (, x k )iH
j=1 k =1
Proof.
Since the kernel is real-valued we can restrict ourselves to a quadratic
form with real coefficients.
For distinct points x 1 , . . . , x N and arbitrary c RN we have
X N
N X N X
X N
cj ck K (x j , x k ) = cj ck hK (, x j ), K (, x k )iH
j=1 k =1 j=1 k =1
XN N
X
= h cj K (, x j ), ck K (, x k )iH
j=1 k =1
XN
= k cj K (, x j )k2H 0.
j=1
Proof.
Since the kernel is real-valued we can restrict ourselves to a quadratic
form with real coefficients.
For distinct points x 1 , . . . , x N and arbitrary c RN we have
X N
N X N X
X N
cj ck K (x j , x k ) = cj ck hK (, x j ), K (, x k )iH
j=1 k =1 j=1 k =1
XN N
X
= h cj K (, x j ), ck K (, x k )iH
j=1 k =1
XN
= k cj K (, x j )k2H 0.
j=1
Proof (cont.).
To establish the second claim we assume K is not strictly positive
definite and show that the point evaluation functionals must be linearly
dependent.
Proof (cont.).
To establish the second claim we assume K is not strictly positive
definite and show that the point evaluation functionals must be linearly
dependent.
If K is not strictly positive definite then there exist distinct points
x 1 , . . . , x N and nonzero coefficients cj such that
N X
X N
cj ck K (x j , x k ) = 0.
j=1 k =1
Proof (cont.).
To establish the second claim we assume K is not strictly positive
definite and show that the point evaluation functionals must be linearly
dependent.
If K is not strictly positive definite then there exist distinct points
x 1 , . . . , x N and nonzero coefficients cj such that
N X
X N
cj ck K (x j , x k ) = 0.
j=1 k =1
Proof (cont.).
Now we take the Hilbert space inner product with an arbitrary function
f H and use the reproducing property of K to obtain
N
X
0 = hf , cj K (, x j )iH
j=1
Proof (cont.).
Now we take the Hilbert space inner product with an arbitrary function
f H and use the reproducing property of K to obtain
N
X
0 = hf , cj K (, x j )iH
j=1
N
X
= cj hf , K (, x j )iH
j=1
Proof (cont.).
Now we take the Hilbert space inner product with an arbitrary function
f H and use the reproducing property of K to obtain
N
X
0 = hf , cj K (, x j )iH
j=1
N
X
= cj hf , K (, x j )iH
j=1
XN
= cj f (x j )
j=1
Proof (cont.).
Now we take the Hilbert space inner product with an arbitrary function
f H and use the reproducing property of K to obtain
N
X
0 = hf , cj K (, x j )iH
j=1
N
X
= cj hf , K (, x j )iH
j=1
XN N
X
= cj f (x j ) = cj x j (f ).
j=1 j=1
Proof (cont.).
Now we take the Hilbert space inner product with an arbitrary function
f H and use the reproducing property of K to obtain
N
X
0 = hf , cj K (, x j )iH
j=1
N
X
= cj hf , K (, x j )iH
j=1
XN N
X
= cj f (x j ) = cj x j (f ).
j=1 j=1
Proof (cont.).
Now we take the Hilbert space inner product with an arbitrary function
f H and use the reproducing property of K to obtain
N
X
0 = hf , cj K (, x j )iH
j=1
N
X
= cj hf , K (, x j )iH
j=1
XN N
X
= cj f (x j ) = cj x j (f ).
j=1 j=1
Remark
This theorem provides one direction of the connection between strictly
positive definite functions and reproducing kernels.
Remark
This theorem provides one direction of the connection between strictly
positive definite functions and reproducing kernels.
Remark
This theorem provides one direction of the connection between strictly
positive definite functions and reproducing kernels.
Since the RBFs we have built our interpolation methods from are
strictly positive definite functions, we want to know how to construct a
reproducing kernel Hilbert space associated with those strictly positive
definite basic functions.
Outline
In this section we will show that every strictly positive definite radial
basic function can indeed be associated with a reproducing kernel
Hilbert space
In this section we will show that every strictly positive definite radial
basic function can indeed be associated with a reproducing kernel
Hilbert space its native space.
In this section we will show that every strictly positive definite radial
basic function can indeed be associated with a reproducing kernel
Hilbert space its native space.
First, we note that the definition of an RKHS tells us that H contains all
functions of the form
XN
f = cj K (, x j )
j=1
provided x j .
Using the properties of RKHSs established earlier along with the form
of f just mentioned we have that
XN N
X
kf k2H = hf , f iH = h cj K (, x j ), ck K (, x k )iH
j=1 k =1
Using the properties of RKHSs established earlier along with the form
of f just mentioned we have that
XN N
X
kf k2H = hf , f iH = h cj K (, x j ), ck K (, x k )iH
j=1 k =1
N X
X N
= cj ck hK (, x j ), K (, x k )iH
j=1 k =1
Using the properties of RKHSs established earlier along with the form
of f just mentioned we have that
XN N
X
kf k2H = hf , f iH = h cj K (, x j ), ck K (, x k )iH
j=1 k =1
N X
X N
= cj ck hK (, x j ), K (, x k )iH
j=1 k =1
XN X N
= cj ck K (x j , x k ).
j=1 k =1
Using the properties of RKHSs established earlier along with the form
of f just mentioned we have that
XN N
X
kf k2H = hf , f iH = h cj K (, x j ), ck K (, x k )iH
j=1 k =1
N X
X N
= cj ck hK (, x j ), K (, x k )iH
j=1 k =1
XN X N
= cj ck K (x j , x k ).
j=1 k =1
Remark
Note that this definition implies that a general element in HK () has
the form
X N
f = cj K (, x j ).
j=1
However, not only the coefficients cj , but also the specific value of N
and choice of points x j will vary with f .
fasshauer@iit.edu MATH 590 Chapter 13 20
Native Spaces for Strictly Positive Definite Functions
Theorem
If K : R is a symmetric strictly positive definite kernel, then
the bilinear form h, iK defines an inner product on HK ().
Theorem
If K : R is a symmetric strictly positive definite kernel, then
the bilinear form h, iK defines an inner product on HK ().
Remark
A pre-Hilbert space is an inner product space whose completion is a
Hilbert space.
Proof.
h, iK is obviously bilinear and symmetric.
Proof.
h, iK is obviously bilinear and symmetric.
We just need to show that hf , f iK > 0 for nonzero f HK ().
Proof.
h, iK is obviously bilinear and symmetric.
We just need to show that hf , f iK > 0 for nonzero f HK ().
Any such f can be written in the form
N
X
f = cj K (, x j ), x j .
j=1
Proof.
h, iK is obviously bilinear and symmetric.
We just need to show that hf , f iK > 0 for nonzero f HK ().
Any such f can be written in the form
N
X
f = cj K (, x j ), x j .
j=1
Then
N X
X N
hf , f iK = cj ck K (x j , x k ) > 0
j=1 k =1
Proof.
h, iK is obviously bilinear and symmetric.
We just need to show that hf , f iK > 0 for nonzero f HK ().
Any such f can be written in the form
N
X
f = cj K (, x j ), x j .
j=1
Then
N X
X N
hf , f iK = cj ck K (x j , x k ) > 0
j=1 k =1
x (f ) = hf , K (, x)iK , f H
e K ().
x (f ) = hf , K (, x)iK , f H
e K ().
Remark
The technical details concerned with this construction are discussed in
[Wendland (2005a)].
fasshauer@iit.edu MATH 590 Chapter 13 23
Native Spaces for Strictly Positive Definite Functions
HK () = span{K (, y) : y }
HK () = span{K (, y) : y }
In the special case when we are dealing with strictly positive definite
(translation invariant) functions (x y) = K (x, y) and when = Rs
we get a characterization of native spaces in terms of Fourier
transforms.
HK () = span{K (, y) : y }
In the special case when we are dealing with strictly positive definite
(translation invariant) functions (x y) = K (x, y) and when = Rs
we get a characterization of native spaces in terms of Fourier
transforms.
Theorem
Suppose C(Rs ) L1 (Rs ) is a real-valued strictly positive definite
function. Define
f
G = {f L2 (Rs ) C(Rs ) : p L2 (Rs )}
and equip this space with the bilinear form
Z
1 f g 1 f ()g()
hf , giG = p h p , p iL2 (Rs ) = p d.
(2)s (2)s ()
Rs
Theorem
Suppose C(Rs ) L1 (Rs ) is a real-valued strictly positive definite
function. Define
f
G = {f L2 (Rs ) C(Rs ) : p L2 (Rs )}
and equip this space with the bilinear form
Z
1 f g 1 f ()g()
hf , giG = p h p , p iL2 (Rs ) = p d.
(2)s (2)s ()
Rs
Theorem
Suppose C(Rs ) L1 (Rs ) is a real-valued strictly positive definite
function. Define
f
G = {f L2 (Rs ) C(Rs ) : p L2 (Rs )}
and equip this space with the bilinear form
Z
1 f g 1 f ()g()
hf , giG = p h p , p iL2 (Rs ) = p d.
(2)s (2)s ()
Rs
Remark
The eigenvalues k , k = 1, 2, . . ., and eigenfunctions k of this
operator play the central role in Mercers theorem [Mercer (1909),
Riesz and Sz.-Nagy (1955), Rasmussen and Williams (2006)].
Theorem (Mercer)
Theorem (Mercer)
Theorem (Mercer)
Remark
We can interpret condition (2) as a type of integral positive
definiteness.
Remark
We can interpret condition (2) as a type of integral positive
definiteness.
As usual, the eigenvalues and eigenfunctions satisfy
TK k = k k
or Z
K (x, y)k (y)dy = k k (x), k = 1, 2, . . . .
X
X
hf , giH = h cj j , dk k iH
j=1 k =1
jk
hj , k iH = p
j k
of the eigenfunctions.
X
X
hf , K (, x)iH = h cj j , k k k (x)iH
j=1 k =1
X
X
hf , K (, x)iH = h cj j , k k k (x)iH
j=1 k =1
X ck k k (x)
=
k
k =1
X
X
hf , K (, x)iH = h cj j , k k k (x)iH
j=1 k =1
X ck k k (x)
=
k
k =1
X
= ck k (x)
k =1
X
X
hf , K (, x)iH = h cj j , k k k (x)iH
j=1 k =1
X ck k k (x)
=
k
k =1
X
= ck k (x) = f (x).
k =1
Finally, one has (c.f. [Wendland (2005a)]) that the native space NK ()
is given by
( )
X 1 2
NK () = f L2 () : |hf , k iL2 () | <
k
k =1
Finally, one has (c.f. [Wendland (2005a)]) that the native space NK ()
is given by
( )
X 1 2
NK () = f L2 () : |hf , k iL2 () | <
k
k =1
Finally, one has (c.f. [Wendland (2005a)]) that the native space NK ()
is given by
( )
X 1 2
NK () = f L2 () : |hf , k iL2 () | <
k
k =1
Remark
Since NK () is a subspace of L2 () this corresponds to the
identification ck = hf , k iL2 () of the generalized Fourier coefficients in
the discussion above.
Outline
W2m (Rs ) = {f L2 (Rs ) C(Rs ) : f ()(1 + k k22 )m/2 L2 (Rs )}. (4)
W2m (Rs ) = {f L2 (Rs ) C(Rs ) : f ()(1 + k k22 )m/2 L2 (Rs )}. (4)
Remark
One also frequently sees the definition
The former interpretation will make clear the connection between the
natives spaces of Sobolev splines (or Matrn functions) and those of
polyharmonic splines.
The former interpretation will make clear the connection between the
natives spaces of Sobolev splines (or Matrn functions) and those of
polyharmonic splines.
or Z 1/2
2
2 m
kf kW2m (Rs ) = f () (1 + kk2 ) d .
Rs
The former interpretation will make clear the connection between the
natives spaces of Sobolev splines (or Matrn functions) and those of
polyharmonic splines.
or Z 1/2
2
2 m
kf kW2m (Rs ) = f () (1 + kk2 ) d .
Rs
Example
The Matrn functions
s
K s (kxk)kxk 2 s
2
(x) = , > ,
21 () 2
can immediately (from our earlier native space characterization theorem ) be seen to have
native space
N (Rs ) = W2 (Rs ) with > s/2
which is why some people refer to the Matrn functions as Sobolev
splines.
Example
Wendlands compactly supported functions s,k = s,k (k k2 ) of
Chapter 11 can be shown to have native spaces
s/2+k +1/2
Ns,k (Rs ) = W2 (Rs )
Remark
Native spaces for strictly conditionally positive definite functions can
also be constructed.
Remark
Native spaces for strictly conditionally positive definite functions can
also be constructed.
Remark
In fact, the intersection of all Beppo-Levi spaces BLk () of order
k m yields the Sobolev space W2m ().
Remark
In fact, the intersection of all Beppo-Levi spaces BLk () of order
k m yields the Sobolev space W2m ().
and the formulas given in Chapter 8 for the Fourier transforms of radial
powers and thin plate splines show immediately that their native
spaces are Beppo-Levi spaces.
and the formulas given in Chapter 8 for the Fourier transforms of radial
powers and thin plate splines show immediately that their native
spaces are Beppo-Levi spaces.
The semi-norm on BLk is given by
1/2
X k!
|f |BLk = kD f k2L2 (Rs ) , (6)
1 ! . . . d !
||=k
and the formulas given in Chapter 8 for the Fourier transforms of radial
powers and thin plate splines show immediately that their native
spaces are Beppo-Levi spaces.
The semi-norm on BLk is given by
1/2
X k!
|f |BLk = kD f k2L2 (Rs ) , (6)
1 ! . . . d !
||=k
Remark
For more details see [Wendland (2005a)].
Beppo-Levi spaces were already studied in the early papers
[Duchon (1976), Duchon (1977), Duchon (1978), Duchon (1980)].
fasshauer@iit.edu MATH 590 Chapter 13 40
Examples of Native Spaces for Popular Radial Basic Functions
The native spaces for Gaussians and (inverse) multiquadrics are rather
small.
The native spaces for Gaussians and (inverse) multiquadrics are rather
small.
Example
According to the Fourier transform characterization of the native space,
for Gaussians the Fourier transform of f N () must decay faster
than the Fourier transform of the Gaussian (which is itself a Gaussian).
The native spaces for Gaussians and (inverse) multiquadrics are rather
small.
Example
According to the Fourier transform characterization of the native space,
for Gaussians the Fourier transform of f N () must decay faster
than the Fourier transform of the Gaussian (which is itself a Gaussian).
The native spaces for Gaussians and (inverse) multiquadrics are rather
small.
Example
According to the Fourier transform characterization of the native space,
for Gaussians the Fourier transform of f N () must decay faster
than the Fourier transform of the Gaussian (which is itself a Gaussian).
The native spaces for Gaussians and (inverse) multiquadrics are rather
small.
Example
According to the Fourier transform characterization of the native space,
for Gaussians the Fourier transform of f N () must decay faster
than the Fourier transform of the Gaussian (which is itself a Gaussian).
Remark
The content of this theorem was already known much earlier (see
[Whittaker (1915)]).
Remark
The content of this theorem was already known much earlier (see
[Whittaker (1915)]).
Remark
The content of this theorem was already known much earlier (see
[Whittaker (1915)]).
References I
Adams, R. (1975).
Sobolev Spaces. Academic Press (New York).
Buhmann, M. D. (2003).
Radial Basis Functions: Theory and Implementations.
Cambridge University Press.
Cheney, E. W. and Light, W. A. (1999).
A Course in Approximation Theory.
Brooks/Cole (Pacific Grove, CA).
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).
References II
References III
Duchon, J. (1977).
Splines minimizing rotation-invariant semi-norms in Sobolev spaces.
in Constructive Theory of Functions of Several Variables, Oberwolfach 1976, W.
Schempp and K. Zeller (eds.), Springer Lecture Notes in Math. 571,
Springer-Verlag (Berlin), pp. 85100.
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.
Duchon, J. (1980).
Fonctions splines homognes plusiers variables.
Universit de Grenoble.
Mercer, J. (1909).
Functions of positive and negative type, and their connection with the theory of
integral equations.
Phil. Trans. Royal Soc. London Series A 209, pp. 415446.
References IV
Schaback, R. (1999a).
Native Hilbert spaces for radial basis functions I.
in New Developments in Approximation Theory, M. W. Mller, M. D. Buhmann,
D. H. Mache and M. Felten (eds.), Birkhuser (Basel), pp. 255282.
Schaback, R. (2000a).
A unified theory of radial basis functions. Native Hilbert spaces for radial basis
functions II.
J. Comput. Appl. Math. 121, pp. 165177.
Shannon, C. (1949).
Communication in the presence of noise.
Proc. IRE 37, pp. 1021.
Unser, M. (2000).
Sampling 50 years after Shannon.
Proc. IEEE 88, pp. 569587.
References V
Whittaker, J. M. (1915).
On the functions which are represented by expansions of the interpolation theory.
Proc. Roy. Soc. Edinburgh 35, pp. 181194.
Ye, Q. (2010).
Reproducing kernels of generalized Sobolev spaces via a Green function
approach with differential operators.
Submitted.