Anda di halaman 1dari 22

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

R. K. BEATSON AND H.-Q. BUI

A BSTRACT. This paper develops some mollification formulas involving convolutions between popular radial basis function (RBF) basic functions , and suitable mollifiers. Polyharmonic splines, scaled Bessel kernels (Matern functions) and compactly supported basic
functions are considered. A typical result is that in Rd the convolution of | | and
(2 + c2 )(+2d)/2 is the generalized multiquadric (2 + c2 )/2 up to a multiplicative
constant. The constant depends on c > 0, where <() > d, and d. An application
which motivated the development of the formulas is a technique called implicit smoothing. This computationally efficient technique smooths a previously obtained RBF fit by
replacing the basic function with a smoother version during evaluation.

1. I NTRODUCTION
This paper develops some mollification formulas involving convolutions between popular radial basis function (RBF) basic functions , and suitable mollifiers k. Polyharmonic
splines, scaled Bessel kernels (Matern functions) and compactly supported basic functions
are considered. An application which motivated the development of the formulas is a
technique called implicit smoothing. This computationally efficient technique smooths a
previously obtained RBF fit by replacing the basic function with a smoother version
during evaluation. In the case of the polyharmonic spline basic functions the smoothed
basic function is a generalised multiquadric or shifted thin-plate spline (at least up to a
polynomial).
Special cases of one of the mollification formulas developed here were given in the
1D setting in [3]. That paper concerned error estimates for quasi interpolation with 1D
generalised multiquadrics, and showed by elementary methods
j

(2j 1)!! 2j
c | |(2j1) ? (2 + c2 )(2j+1)/2 = (2 + c2 )(2j1)/2 ,
(2j)!!

for c > 0and j N . A multivariate analog relating (x) = |x| and the multiquadric
(x) = x2 + c2 in R3 has been used to smooth implicit surface fits to lidar and laser
scanner data (see Figures 1 and 2, and Section 2). This application is detailed in [6]. That
paper presents the application but not the mathematics underlying it.
The purpose of the current paper is to present a mathematical treatment of general versions of these, and related, mollification formulas. An important case of the mollification
formulas proved in this paper, and used in the lidar and laser scanner application [6] mentioned above, is the following formula
(2d+)/2 
/2
(( + 2d)/2) d+ 
c
| | ? 2 + c2
(x) = |x|2 + c2
d/2
,
(( + d)/2)
holding whenever c > 0, <() > d, and x Rd .
2000 Mathematics Subject Classification. 41A30, 65D10.
Key words and phrases. Mollification formulas, radial basis functions, implicit smoothing, data smoothing.
1

R. K. BEATSON AND H.-Q. BUI

Initially our development for the polyharmonic spline case was based on viewing odd
powers of |x| as multiples of fundamental solutions of iterated versions of Laplaces equation. As such our treatment was restricted to polyharmonic splines in odd dimensions.
Changing to arguments based more directly on generalised functions has enabled many
restrictions to be dropped. For example in the case of the results for |x| (Theorem 4.1)
there is no longer any restriction on the parity of the dimension, nor any requirement that
the power of |x| be odd or integer. Furthermore we develop analogous formulas for
polyharmonic splines in even dimension. Related results for scaled Bessel kernels (Matern
functions) and compactly supported radial basis functions are also discussed.
Notation: In this paper the Fourier transform is defined as follows
Z
b
f () :=
eix f (x)dx,
f L1 (Rd ).
Rd

The Fourier transform is extended to the space of tempered distributions S 0 (Rd ) in the
usual manner.
2. A N APPLICATION IMPLICIT SMOOTHING
This section concerns an application of the mollification formulas to come to the smoothing of RBF fits. This particular application motivated the development of the formulas. The
process will be called implicit smoothing and can be viewed as smoothing an interpolant
to noisy data rather than smoothing the data itself.
The process starts with a noisy data set to be approximated. Figure 2 shows one example
of such a noisy data set, a noisy Lidar scan. Firstly an RBF is fitted to the noisy data
yielding an RBF approximation
(2.1)

s(x) = p(x) +

N
X

i (x xi ).

i=1

Then the initial RBF approximation is smoothed by convolution with the mollifier k yielding a smoother fit
(2.2)

se(x) = q(x) +

N
X

i (x xi ),

i=1

where q = p ? k and = ? k. Figure 1 shows zoomed in views of the isosurfaces arising


when this strategy applied to the noisy Lidar scan
of Figure 2. Here (x) = |x| is the
biharmonic spline basic function in R3 and (x) = x2 + c2 is the ordinary multiquadric.
In Figure 1 one can clearly see the amount of smoothing increase with the parameter c.
Figure 3 shows a thin-plate spline fit to data from the Mexican hat function f (x) = (1
x2 ) exp(x2 /2) at 400 scattered points in R2 . Uniform random noise of magnitude 0.7
2
has been added to the original Mexican hat
height data. Here (x) = x log |x| is the thin2
2
2
2
plate spline and (x) = (x + c ) log x + c is the shifted thin-plate spline. Implicit
smoothing has been employed to obtain an approximation to the noise free signal.
For important choices of , and suitable choices of k, the smoothed basic function
turns out to be a simple and easy to evaluate, function. Therefore in applying the technique
there is no need for any explicit, and computationally expensive, evaluation of convolutions. Rather one fits the initial radial basis function s, and then smooths it when evaluating
by substituting the smoothed basic function for the original basic function . Thus the

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

Original unsmoothed fit

Smoothed with c =5mm

Smoothed with c =10mm

Smoothed with c =20mm

F IGURE 1. Implicit smoothing applied to a noisy Lidar scan.


technique can be viewed as smoothing by basic function substitution. In some important
cases fast evaluators are available for the smoothed RBF se.
Advantages of the technique are
No explicit convolution to do.
There is no requirement that the data or evaluation points be gridded.
Well understood linear filtering. Fourier transform and absolute moments of the
smoothing kernel are known.
Smoothing can be chosen for appropriate frequency or length scale.
A posteriori parameter for user to play with noise level/frequency need not be
known a priori.
Disadvantages of the method are
There is a parameter for the user to play with.
The method is just linear filtering so will blur sharp features.

R. K. BEATSON AND H.-Q. BUI

F IGURE 2. A noisy lidar scan of a statue in Santa Barbara.

(a) The Mexican hat function

(b) Exact fit to Mexican hat plus noise data

(c) Smoothed with c = 0.2

(d) Smoothed with c = 0.6

F IGURE 3. Various fits to noisy data created from the Mexican hat function.

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

Implicit smoothing of globally supported RBFs should also have many other applications. [11] has used related basic function substitution techniques as part of a process for
the numerical solution of PDEs with compactly supported RBFs.
3. T ECHNICAL LEMMAS
The following technical lemmas which deal with distributions and special functions will
be needed in later sections.
Lemma 3.1. Identification of a convolution
(i) Let ,  > 0. Suppose g C(Rd ), g(x) = O(|x| ) as |x| , and k
L (Rd ), k(x) = O(|x|(d++) ) as |x| . Then g ? k is a C(Rd ) function
with (g ? k)(x) = O(|x| ) as |x| .
(ii) Let , , g and k be as in part (i). Let h C(Rd ) be such that h(x) = O(|x| ) as
|x| . Viewing gb and b
h as tempered
distributions suppose that there exist func

tions G and H in L1loc Rd \{0} such that for all test functions D Rd \{0}
Z
hb
g , i =
G()()d
Rd

and

Z
hb
h, i =

H()()d.
Rd

Further writing b
k for the classical Fourier transform of k suppose
G()b
k() = H(),

for almost all 6= 0.

Then
(g ? k)(x) = p(x) + h(x)
for all x Rd
where p is a polynomial of degree not exceeding the integer part of .
Proof of part (i).
Z
g(x y)k(y)dy
Z


=O
(1 + |x| + |y|) (1 + |y|)(d++) dy


Z

= O (1 + |x|) (1 + |y|) (1 + |y|)(d++) dy

(g ? k)(x) =

= O((1 + |x|) ).
Now fix x Rd and let {tn } be a sequence tending to zero in Rd with |tn | 1 for all
n. By an analogous argument to that above there is a constant C so that
|g(x + tn y)k(y)| C(2 + |x|) (1 + |y|)(d+)
for almost all y. Hence, applying the Lebesgue dominated convergence theorem,
lim (g ? k)(x + tn ) = (g ? k)(x).

It follows that g ? k is continuous.


Proof of part (ii)

Below we will view g ? k as a tempered distribution. Let D Rd \{0} . Hence b S
the space of rapidly decreasing functions. The growth conditions on g, k, and b combine to

R. K. BEATSON AND H.-Q. BUI

imply that all the iterated integrals below are absolutely convergent, so that the applications
of Fubinis Theorem that occur are justified.
b
h(g ? k)b, i = hg ? k, i
Z Z
b
=
g(x y)k(y)dy (x)dx
Z Z
b + z)dz k(y)dy
=
g(z)(y

Z Z


=
g(z) eiy () b(z)dz k(y)dy
Z

 iy

=
g, e
() b k(y)dy
Z

iy

gb, e
() k(y)dy
=

Z Z
=
G()eiy ()d k(y)dy
Z Z
=
G()eiy ()k(y)dy d
Z

Z
iy
= G()()
e
k(y)dy d
Z
= G()b
k() ()d
Z
= H()()d
= hb
h, i.
Hence g[
?kb
h is a distribution supported at the origin. Therefore g ? k = p + h where p
is a polynomial. The growth of g ? k and h implies that p is of degree at most the integer
part of .

In the following B is the Beta function
Z
dt
(z)(w)
tz
B(z, w) :=
=
,
z+w t
(1
+
t)
(z + w)
0
and is the Digamma function (z) := 0 (z)/(z) (see e.g. [1]). 1
Lemma 3.2. Let c, w, v R with c > 0 and v > w/2 > 0. Then
Z
rw
dr
cw2v  w
w
(3.1)
=
B
,
v

.
v
2
2
2
(c2 + r2 ) r
0
Further
Z w
r log r dr
I(w, v) :=
v
(c2 + r2 ) r
0


w

w cw2v
1 w 1 
w
(3.2)
=B
,v
log(c) +
v
.
2
2
2
2
2
2
2
1Care is needed in interpreting the literature as many authors use (z) to denote the function 0 (z+1)/(z+
1) instead. See for example [14, page 114].

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

and

rw log(r2 + c2 ) dr
v
r
(c2 + r2 )
0
w


w2v n
w o
w c
=B
2 log(c) + (v) v
.
,v
2
2
2
2
Z

M (w, v) :=
(3.3)

Proof of the first identity. The assumptions on v and w clearly imply the integral is convergent. Substituting c2 t = r2

Z
0

dr
1
rw
=
v
2
2
2
(c + r ) r

Z
0
w

w/2
c2 t
dt
v
2
2
(c + c t) t

Z w/2
t
dt
c
2c2v 0 (1 + t)v t
cw2v  w
w
=
B
,v
.
2
2
2
=

rw | log r| dr
< we use the Lebesgue Domi(c2 + r2 )v r
0
nated Convergence Theorem and differentiate under the integral sign to obtain
Z

Proof of the second identity. Since

d
I(w, v) =
dw

Z
0

dr
rw
.
v
2
2
(c + r ) r

Employing (3.1) this implies



cw2v  w
w
B
,v
2
2
2
(

)
w2v w v w
d
c
2
2
=
dw
2
(v)
n
w 
1
w
=
log(c)cw2v
v
2(v)
2
2
1  w   w  w2v 
w
+

c
v
2
2
2
2

1 w2v  w  
w 
w
c

v
v
2
2
2
2



w
w


B 2 , v 2 w2v
1
w
1 
w
=
c
log(c) +
v
.
2
2
2
2
2

I(w, v) =

d
dw

Proof of the third identity. Proceeding as in the proof of the second integral identity
d
M (w, v) =
dv

Z
0

rw
dr
.
v
(c2 + r2 ) r

R. K. BEATSON AND H.-Q. BUI

Employing (3.1) this implies


(

)
d cw2v w2 v w2
M (w, v) =
dv
2
(v)
(


w
w2
w2v v 2
=
2 log(c)c
2
(v)


 0 )
w
w
w
w2v v 2 v 2
w2v v 2 (v)
+c
c
2
(v)
((v))


B w2 , v w2 w2v n
w o
2 log(c) + (v) v
.

=
c
2
2
4. M OLLIFICATION FORMULAS FOR POWERS OF THE MODULUS
In this section mollification formulas will be developed for powers of the modulus. The
flavour of the main result, Theorem 4.1, is that the convolution of |x| against an appropri/2
ate inverse multiquadric is the generalised multiquadric |x|2 + c2
. Further, a quantitative Korovkin Theorem, Proposition 4.2, estimates the distance between the original
unsmoothed RBF s and the corresponding smoothed RBF se = s ? kd,,c .
Define the generalised multiquadric basic function ( the generalised Fourier transform
of a Bessel kernel) as
,c (x) = (|x|2 + c2 )/2 ,

(4.1)

x Rd .

where c > 0. These functions are most often considered in the case that is a positive odd
integer. Clearly ,c can be viewed as a smoothed out version of |x| . The results of this
section show that the smoothing out is actually given by a convolution.
More precisely Theorem 4.1 to come shows that in Rd and for all > d
,c (x) = ( ? kd,,c )(x),

(4.2)

where (x) = |x| and the convolution kernel kd,,c is the generalised multiquadric with
negative index 2d,c , normalised to have integral one. That is
kd,,c (x) = ad, cd+ 2d,c (x),

(4.3)

for some constant ad, . Write


d := 2 d/2 /(d/2),

(4.4)

for the surface area of the unit sphere in Rd . Then


Z
Z
1=
kd,,c (x)dx = d ad, cd+
Rd

rd
(r2

c2 )(+2d)/2

dr
,
r

implying from equation (3.1) that


(4.5)

ad, = d/2

(( + 2d)/2)
.
(( + d)/2)

Interpretation of convolution against the kernel as a low pass filter will be aided by the
expression (4.8) for its Fourier transform
21(d+)/2
(+d)/2
b
kd,,c () =
K +d (c||) (c||)
,
2
(( + d)/2)

> 0.

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

This Fourier transform is a positive function tending to zero exponentially fast with ||.
Considering the graph of b
kd,,c (||) against || it is clear that the width of the graph at
any fixed height is inversely proportional to c. This expresses precisely how the graph of
b
kd,,c () grows more and more peaked as c increases. Thus convolution with b
kd,,c will
attenuate high frequencies more and more as c increases.
Using equation (3.1) again one has
R
Z

d |x| kd,,c (x)dx

|x| kd,,c (x)dx = RR


k
(x)dx
Rd
Rd d,,c
R
d+
r
dr
=R

R (c2 +r 2 )(+2d)/2 r
rd
dr
R (c2 +r 2 )(+2d)/2 r

= c B(
(4.6)

d+ d
d d+
, )/B( ,
)
2
2
2
2

= c .

We are particularly interested in the polyharmonic splines. In odd dimension odd


powers of the modulus are multiples of fundamental solutions of iterated versions of
Laplaces equation. Radial basis functions based on sums of shifts of these fundamental
solutions supplemented by polynomials have many wonderful properties. Such polyharmonic splines arise naturally as smoothest interpolants (see [8]) and have performed extremely well in many practical applications (see e.g. [5]). A particularly important special
case is that of the basic function (x) = |x| in R3 . The corresponding RBFs, which take
the form of a linear polynomial plus sums of shifts of the modulus, are called biharmonic
splines in R3 .
We will now show the mollification formula
Theorem 4.1. For c > 0 and such that <() > d
(4.7)

( ? kd,,c )(x) = ,c (x),

for all x Rd .

Remark: This result includes the cases where the power and the dimension d are both
odd, and (x) = |x| is a polyharmonic spline basic function. It also includes many
other cases. In particular the range of validity of the formula includes those exceptional
s for which and ,c are polynomial.
Proof. We note the following generalized Fourier transforms for the relevant functions
viewed as distributions acting on D(Rd \{0}).
 (d+)/2
||
2 d/2
(F,c ) () =
K(d+)/2 (c||)
(/2)
2c
for 6 2N0 , where K(d+)/2 is a modified Bessel function (see [1, page 374] or [2, page
415]). For < d this is a classical Fourier transform on Rd . Also
(F ) () = 2d+ d/2

((d + )/2) d
||
(/2)

S
for
/ (d 2N0 ) (2N0 ) (see [12, page 363]).
Then using the normalising constant
ad, = d/2

(( + 2d)/2)
(( + d)/2)

10

R. K. BEATSON AND H.-Q. BUI

defined above and that K (z) = K (z) we find


b
b 2d,c
kd,,c = ad, cd+
=

(4.8)

21(d+)/2
(+d)/2
,
K +d (c||) (c||)
2
(( + d)/2)

for > d . It is then clear that


b ,c (),
(F ) ()b
kd,,c () =

Rd \{0}.

for a set of s including 0 < < 1.


Now fix c > 0, 0 < < 1, and apply Lemma 3.1. The Lemma implies that
? kd,,c = p,c + ,c

(4.9)

where p,c is a polynomial of degree 0 and the equality holds pointwise. Considering the
point x = 0 we find
Z
( ? kd,,c ) (0) =
|t| ? kd,,c (t)dt = c ,
Rd

where we have used (4.6). Then observing that ,c (0) = c we deduce that the polynomial p,c in (4.9) must be identically zero. We have therefore established that for all c > 0
and 0 < < 1
(4.10)

( ? kd,,c )(x) = ,c (x),

for all x Rd .

Now fix c > 0 and x Rd . The right-hand side of (4.7) is an entire function of .
The left-hand side is continuous on d := { : <() > d} by the Lesbesgue dominated
convergence theorem. A standard argument using Moreras theorem then shows that the
left-hand side is analytic on d . (4.10) shows that (4.7) holds for 0 < < 1. Hence, by
analytic continuation, it holds for all <() > d.

The remainder of this section will concern the application of the Theorem above to
implicit smoothing.
Recall from (4.6) that
Z
|x| kd,,c (x)dx = c .
Rd

A routine application of Holders inequality then shows that


Z
(4.11)
|x| kd,,c (x)dx < c ,
for all 0 < < .
Rd

These expressions for the -th absolute moment of kd,,c clearly quantify the manner
in which the kernel becomes peaked as c approaches zero. Loosely speaking they show
that the dominant part of convolution against the kernel is averaging on a length scale of
approximately c, at least for functions of sufficiently slow growth. Thus we can expect
convolution against kd,,c to lose, or smooth, detail at this length scale.
A more precise statement about the error between the original RBF s and its smoothed
version se is implied by the quantitative Korovkin theorem we are about to present. See the
last paragraph of this section for the details.
Given a uniformly continuous function g : Rd R define its uniform norm modulus
of continuity
(g, Rd , ) :=
sup
|g(x) g(y)|.
x,yRd :|xy|

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

11

Let f C ` (Rd ) be a function with all `-th order partials uniformly continuous on Rd .
(`)
Define the `-th order directional derivative of f at x in the direction of u, fu (x), as the
`-th derivative of the univariate function g(t) = f (x + tu), at t = 0. The joint modulus of
continuity of all directional derivatives of order ` of f is given by
(f (`) , Rd , ) := sup (fu(`) , Rd , ).
|u|=1

For a discussion of the properties of this modulus of continuity see [4]. Given ` N0 and
0 < 1 we will say the total derivative f (`) is in LipM () if there exists a positive a
constant M such that
(f (`) , Rd , ) M ,

for all 0 < < .

For each multiindex adopt the usual notation defining || = kk1 and the normalized
mononomial
V (x) =

1
1
x =
x1 x2 xdd
!
1 !2 ! d ! 1 2

With this notation in hand we can state the following folklore quantitative Korovkin Theorem. We include the simple proof for the sake of completeness, and also because we do
not know of a convenient reference.
Proposition 4.2. Let ` N0 , 0 < 1, and B > 0. For each c > 0 let kc : Rd R be
a bounded function such that
Z

t kc (t)dt = 0,|| ,

(4.12)

0 || `,

Rd

Z
(4.13)

|t|`+ |kc (t)|dt Bc`+ .

Rd

Then for all f C ` (Rd ) with `-th total derivative in LipM () and all c > 0
(4.14)

kf ? kc f k

BM `+
c .
`!

Proof. Taylors theorem with integral remainder for univariate functions implies




`
j
X

(t) |t|`
(j)
(`)
g(x t)
g
(x)

`! (g , |t|),
j!


j=0

x, t R.

Applying this along the line segment joining x and x t in Rd we find the multivariate
Taylor theorem in the form




X

|t|`  (`) d 


f (x t)
(D
f
)
(x)V
(t)
f , R , |t| .



`!


||`

12

R. K. BEATSON AND H.-Q. BUI

Hence using the hypotheses



Z


(f (x t) f (x)) kc (t)dt
|(f ? kc )(x) f (x) | =
Rd

Z

X

f (x t)
(D f ) (x)V (t) kc (t)dt
=
Rd

||`
Z
|t|`  (`) d 

f , R , |t| |kc (t)|dt


Rd `!
Z
M
BM `+
|t|`+ |kc (t)|dt
c .

`! Rd
`!

As a first application of Theorem 4.1 and Proposition 4.2 consider implicit smoothing
of an RBF of form (2.1) when (x) = |x| and p is of degree 1. Implicit smoothing of
a surface in R3 modelled with such a biharmonic spline is illustrated in Figures 1 and 2.
Then the unsmoothed function s LipM (1) where
M=

|s(x)|.

sup
xRd \{xi :1iN }

Convolving against kd,1,c we note that linear polynomials are preserved so that the smoothed
RBF (2.2) takes the special form
se(x) = p(x) +

N
X

i 1,c (x xi ).

i=1

Applying the Korovkin theorem Proposition 4.2, using that


(4.6), we see that ks sek M c.

R
Rd

|x|kd,1,c (x)dx = c by

5. R ADIAL FUNCTIONS
This section outlines some known fundamental properties of radial functions.
A function f : Rd R is radial if there is a univarate function g such that f (x) =
g(|x|) for all x.
Lemma 5.1. Let f, k : Rd R be such that the integral defining (f ? k)(x) is absolutely
convergent for all x. If f and k are radial then so is f ? k.
Proof. Given x Rd choose a rotation matrix Q so that Qx = |x|e1 where e1 is the vector
(1, 0, . . . , 0). Then
Z
(f ? k)(x) =
f (x t)k(t)dt
Rd
Z
=
f (Qx Qt)k(Qt)dt
d
ZR
=
f (|x|e1 s)k(s)ds
Rd

= (f ? k)(|x|e1 ).

p : Rd R write it in terms of the monomial basis as p(x) =


P Given a polynomial

d a x . Define the homogeneous part of degree j of p, Hj p by (Hj p) (x) =


PN0
||=j a x .
Lemma 5.2. Let p : Rd R be a polynomial which is also radial. Then

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

13

(a) All the homogeneous parts Hj p of p, j = 0, 1, . . ., are also radial.


(b) There is a univariate polynomial q such that
p(x) = q(|x|),

for all x.

Proof of (a). Suppose that p is radial yet at least one homogeneous part of p is not. Let
Hm p be the first non radial homogeneous part of p. Then there exist x, y with |x| = |y| = 1
but (Hm p) (x) 6= (Hm p) (y). Now let
e=p

m1
X

(Hk p) =

k=0

(Hk p) .

k=m

Then

e(rx) e(ry) = rm ((Hm p) (x) (Hm p) (y)) + O rm+1 ,
as r 0+ implying e(rx) 6= e(ry) for all sufficiently small r > 0. But e is radial by
choice of m. Contradiction.
Proof of (b) From part (a) if j is odd then Hj p is both odd and radial, therefore identically
zero. Hence writing e1 for the unit vector (1, 0, . . . , 0) and using part (a) again

X
p(x) =
(H2j p) (x)
j=0

X
j=0

(H2j p) (|x|e1 )
|x|2j (H2j p) (e1 )

j=0

b2j r2j ,

r = |x|,

j=0

where b2j = (H2j p) (e1 ).

6. M OLLIFICATION FORMULAS FOR FUNCTIONS OF THE FORM r2j log r


In this section mollification formulas are developed for the generalised thin-plate spline
basic function |x|2j log |x|, j N . The flavour of the main result Theorem 6.1 is that convolution of the generalised thin-plate basic function against a certain inverse multiquadric
p
j
yields the corresponding shifted thin-plate spline |x|2 + c2 log |x|2 + c2 , plus a polynomial of degree 2j 2.
In even dimension even powers of the modulus multiplied by log |x| are fundamental
solutions of iterated versions of Laplaces equation. In particular RBFs taking the form of
a linear polynomial plus a sum of shifts of x2 log |x| are the biharmonic RBFs in R2 . These
thin-plate splines can be shown to be the solutions of various smoothest interpolation and
penalized smoothing problems (see e.g. [8], [20]) and have proved very successful in many
scattered data fitting applications, see for example [13]. For such functions we will show
mollification formulas of the form
p
/2

(6.1)
|| log | | ? kd,,c = 2 + c2
log 2 + c2 + p ,
where 2N , kd,,c is as in(4.3), and p is a polynomial depending on d, and c. The
first function on the right above
1/2
(6.2)
,c (x) := (|x|2 + c2 )/2 log |x|2 + c2
,
x Rd .

14

R. K. BEATSON AND H.-Q. BUI

is the shifted thin-plate spline basic function of Dyn, Levin and Rippa. Some properties of
these functions can be found in [10] and [9].
In contrast to the case of smoothing |x| discussed in Section 4 a nonzero polynomial
part does arise in smoothing |x|2j log |x|. For example the convolution on the left of (6.1)
for d = 2, = 2, and c = 1 evaluated at zero can be rewritten using polar coordinates as
2

2r r2 log(r)(r2 + 1)3 dr.

This equals 1/2, while (r2 + 1) log r2 + 1 evaluated at r = 0 is 0. This direct calculation
for the special case d = 2 is in agreement with the general formula (6.5) which we are about
to prove.
Explicitly we will show
Theorem 6.1. For c > 0, j N and x Rd


(6.3)



| |2j log | | ? kd,2j,c (x) = 2j,c (x) + p2j (x)

where p2j is a radial polynomial of degree 2j 2. Writing p2j in the form


p2j (x) = b2j,0 + b2j,2 |x|2 + + b2j,2j2 |x|2j2 ,

(6.4)

b2j,0 = c2j

1
1
1
+
+ +
(2j 2) + d (2j 4) + d
d


,

and in particular
b2,0 =

(6.5)

c2
.
d

Proof. Start with the the mollification formula of Theorem 4.1


2


2 /2

|x| + c

d+

= ad, c

|y| |x y|2 + c2

(+2d)/2

dy

Rd

for all x Rd and all <() > d. Differentiating both sides with respect to yields
/2
1/2
|x|2 + c2
log |x|2 + c2
Z

(+2d)/2
d
d+
ad, c
|y| |x y|2 + c2
dy
=
d
d
R
Z
(+2d)/2
+ ad, cd+
|y| log |y| |x y|2 + c2
dy
Rd
Z
1/2
(+2d)/2
ad, cd+
|y| |x y|2 + c2
log |x y|2 + c2
dy
Rd

= A + B C .

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

15

Calculation of quantity A . From (4.5)





d
ad, cd+
ad, cd+ log c
d


d
(( + 2d)/2)
= cd+
d/2
d
(( + d)/2)



 

+2d

0 +2d
0 +d

2
2
2
1


= cd+ d/2


2

+d
+d
2
2



 
1
+d
+ 2d
= ad, cd+

.
2
2
2
Hence employing the mollification formula (4.7) we find






1
+ 2d
1
+d
2
2 /2
A = |x| + c
log c +

.
2
2
2
2
Quantity B2j is the convolution



| |2j log | | ? kd,2j,c (x)
we are interested in.
Calculation of quantity C2j .
Z
(d+j)

1
d+2j
C2j = ad,2j c
|x y|2j |y|2 + c2
log |y|2 + c2 dy
2
j
Expanding the term |x y|2j = |x|2 2 < x, y > +|y|2 that occurs above using the
Binomial Theorem it is clear that |xy|2j a polynomial of degree 2j in x and y. Collecting
terms in the expansion by degree in x
j
|x y|2j = |x|2 2 < x, y > +|y|2
= |x|2j 2j < x, y > |x|2j2
+ j|x|2j2 |y|2 + 2j(j 1) < x, y >2 |x|2j4
+ terms of lower degree in x .

Substituting this expansion into the expression for C2j reveals that C2j is a polynomial
of degree 2j in x. From Lemma 5.1 this convolution of radial functions yields a radial
function. Hence from Lemma 5.2 C2j is a polynomial of degree j in |x|2 .
The term involving |x|2j in the expression C2j is
Z
(d+j)

1
ad,2j cd+2j
|x|2j |y|2 + c2
log |y|2 + c2 dy
2
Rd
Z
(d+j)
 dr
1
rd r2 + c2
= |x|2j ad,2j d cd+2j
log r2 + c2
2
r
0
1
= |x|2j ad,2j d cd+2j M (d, d + j)
2


1
1
2j + d
= |x|2j log(c) + (d + j)
2
2
2
where we have employed (3.3).

16

R. K. BEATSON AND H.-Q. BUI

Combining the expressions for A2j and B2j , with the expression above for the coefficient of |x|2j in C2j we find that the terms in |x|2j cancel and therefore the convolution
has the form given in equation (6.3) of the Theorem.
We proceed to identify the constant part b2j,0 of the polynomial p2j . It follows from
formula (6.3) that


(6.6)
| |2j log | | ? kd,2j,c (0) = 2j,c (0) + b2j,0 .
Using the radial symmetry the left hand side can be rewritten as the univariate integral
Z
(2j+2d)/2 d1
d+2j
r2j log(r) r2 + c2
d ad,2j c
r
dr.
r=0

Applying the notation and results of Lemma 3.2 this becomes


2

 cd+2j I(d + 2j, j + d)
d 2j+d
B 2, 2



 
c2j
d
d
=
log(c2 ) + j +

.
2
2
2
Observing that 2j,c (0) = c2j log(c) it follows that
 

 
d
c2j
d
b2j,0 =
j+

.
2
2
2
Employing the recurrence [1, (6.3.5) and (6.3.6)]
1
1
1
+
+ + + (z)
(n + z) =
(n 1) + z
(n 2) + z
z
the expression for b2j,0 can be rewritten as


1
1
1
+
+ +
,
b2j,0 = c2j
(2j 2) + d (2j 4) + d
d
establishing formula (6.4). Substituting j = 1 gives the formula (6.5).

As an application of the mollification formula of Theorem 6.1 consider implicit smoothing of an ordinary thin-plate spline of form
s(x) = p1 (x) +

N
X

i |x xi |2 log |x xi |

i=1

in R2 . Implicit smoothing of such a fit is illustrated in Figure 3. If the RBF coefficients i


satisfy the usual side conditions
(6.7)

N
X

i = 0

and

i=1

N
X

i xi = 0,

i=1

then the sum of all the constant parts arising from smoothing,
the smoothed thin-plate spline is
se(x) = p1 (x) +

N
X


i c2 /2 , is zero. Thus

p

i |x xi |2 + c2 log |x xi |2 + c2 .

i=1

Further, when (6.7) holds, one can form a far field expansion
A log(|x|) + P1 (x) +

P2 (x) P3 (x)
+
+ ...,
|x|2
|x|4

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

17

of s(x) where Pj denotes a polynomial of degree j. This expansion converges to s(x) with
a geometric rate for all sufficiently large |x|. It follows that the gradient s(1) (x) = s(x)
is bounded and hence s LipM (1) with
M = sup |s(x)|.
xRd

Hence noting from (4.11) that Rd |x| k2,2,c (x)dx < c for all 0 < < 2 and applying
the Korovkin theorem Proposition 4.2 we find that in this case
R

ks sek < M c.
Consider now using the triharmonic spline in R2 based on sums of shifts of |x|4 log |x|
plus quadratics
N
X
4
s(x) = p2 (x) +
i |x xi | log |x xi |.
i=1

The triharmonic spline is C 3 . It is natural to use such a triharmonic spline, rather than the
usual thin-plate spline, if the function being approximated is smoother, or if a smoother
approximation is required.
The usual side conditions for the triharmonic spline are that the coefficients i are orthogonal to quadratics. That is that
(6.8)

N
X

i q(xi ) = 0,

for all quadratics q.

i=1

These conditions imply that the polynomial parts arising from smoothing of the weighted
shifts of | |4 log | | cancel to give the zero polynomial. Therefore the smoothed RBF has
the form
N
X
p
2
se(x) = q2 (x) +
i |x xi |2 + c2 log |x xi |2 + c2 ,
i=1

where q2 = p2 ? k2,4,c will usually differ from p2 .


Further, when the side conditions (6.8) hold, one can form a far field expansion
Q1 (x) log(|x|) + P2 (x) +

P3 (x) P4 (x)
+
+ ...,
|x|2
|x|4

of s(x) where Q1 (x) is a polynomial of degree 1, and for all j, Pj denotes a polynomial of
degree j. This expansion converges to s(x) with a geometric rate for all sufficiently large
|x|. It follows that all the second partials of s are bounded. Hence
R the first total derivative
s(1) LipM (1) for some constant M . Noting from (4.11) that Rd |x| k2,4,c (x)dx < c
for all 0 < < 4 and applying the Korovkin theorem Proposition 4.2 we find that in this
case
ks sek < M c2 .
7. B ESSEL KERNELS AND M ATERN FUNCTIONS
This section discusses mollification formulas for Bessel kernels and the scaled versions
known as Matern functions. These mollification formulas can be exploited in an implicit
smoothing technique for Matern RBFs.
Recall the Bessel kernels Gd, for Rd given for > 0 by

b d, () = 1 + ||2 /2 ,
G

18

R. K. BEATSON AND H.-Q. BUI

and

1
K(d)/2 (|x|)|x|(d)/2 .
d/2 2(d+2)/2 (/2)
Good references for the many wonderful properties of these functions are [2] and [7]. The
properties that we will need in the following are
Gd, L2 , > d/2.
Gd, is continuous when > d.
Gd, (x) = Dd, |x|(d1)/2 e|x| (1 + o(1)) as |x| . Here Dd, is a constant depending on d and .
Gd, is positive definite for > d.
Gd, ? Gd, = Gd,+ , , > 0.
The Bessel kernels are also basic functions corresponding to natural smoothest interpolation problems. [16] and [17] having shown that for k > d/2 RBF interpolation with basic
function Gd,2k yields the interpolant minimizing the Sobolev inner product for W2k (Rd )
Z
k
|b
g ()|2 1 + ||2 d
kgk2W k (Rd ) =
Gd, (x) =

Rd

over all sufficiently smooth interpolants.


More explicit forms for some of the Bessel kernels are
Gd,d+1 (x) =

(d1)/2 |x|
e
,
2d d+1
2

which is Lipschitz,
Gd,d+3 (x) =

(d1)/2
 (1 + |x|)e|x| ,
2d+1 d+3
2

which is twice continuously differentiable and


Gd,d+5 (x) =

(d1)/2
2d+2

d+5
2

 (3 + 3|x| + |x|2 )e|x| .

which is four times continuously differentiable. These formulas and others can be derived
using [1, (10.2.15) or (10.2.17)].
Introduce a parameter c by considering the dilated version scaled to have integral 1
Md,,c (x) = cd Gd, (x/c).

(7.1)

We have used the letter M for these functions as in the Statistics literature they are often
called Matern functions rather than scaled Bessel kernels (see e.g. [19]). From (7.1) and
its immediate consequence
Z
Z
(7.2)
|x|Md,,c (x)dx = c
|x|Gd, (x)dx
Rd

Rd

it is clear that c is a length scale associated with the kernel Md,,c .


The positive definiteness of the Gd, s implies that of the Md,,c s. Hence the scattered
data interpolation problem of finding a Matern RBF of the form
X
s() =
i Md,,c ( xi ).
i

taking given values fi at a finite number of given distinct points xi is guaranteed to have a
unique solution. [15] discuss an application of Matern RBFs to the numerical solution of
PDEs in which they significantly outperform multiquadrics.

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

19

The convolution property of the Gd, s implies


Md,,c ? Md,,c = Md,+,c ,

, > 0.

Therefore given an RBF built upon the basic function Md,,c we can carry out implicit
smoothing in the form
X
s=
i Md,,c ( xi )
i

(
X

)
i Md,,c ( xi )

? Md,,c

i Md,+,c ( xi ) =: se.

The RBF s based on the kernel Md,,c is Lipschitz whenever d + 1. Hence, the
Korovkin theorem and the moment estimate (7.2) imply that

Z

ks sek
|x|Gd, (x)dx
sup |s(x)| c,
Rd

xRd

whenever d + 1.
8. M OLLIFICATION FORMULAS FOR COMPACTLY SUPPORTED RBF S
This section discusses mollification formulas for compactly supported RBFs. These
mollification formulas can be exploited in implicit smoothing techniques for compactly
supported RBFs.
In this section we abuse notation and do not distinguish between radial functions in
terms of the dimension of their Euclidean domain. Thus we write f (x) for a radial function
which when written as g(|x|) we can consider as having any finite dimensional Euclidean
domain Rn . Consequently we need to identify the dimensionality of the domain differently
and do so by writing ?d for the convolution in Rd . Following [21] given an L compactly
supported radial function f form from it another radial function (If ) by defining (viewing
f as a function of the variable r = |x|)
Z

sf (s)ds, 0 t < ,
(8.1)
(If )(t) =
s=t

(If )(t),
t otherwise.
Also define for t > 0

1 d
f (t).
t dt
In the interior of intervals of continuity of f the fundamental theorem of calculus implies
(Df )(t) =

(DIf )(t) = f (t).


Then Wendland derives from the work of [18] formulas for the convolutions of compactly supported radial functions one of which should read
(8.2)

I(f ?d+2 g) = 2(If ) ?d (Ig).

Now let kc be the characteristic function of the ball with radius c in R3 , normalised to
have integral one. By Bochners theorem C3 (x) = (kc ?3 kc )(x) will necessarily be positive
definite. Statisticians call this function the spherical covariance, but it is also known as
the Euclid hat. The latter name derives from the analogy with the univariate piecewise
linear hat function, or linear B-spline, on a uniform mesh, which is the convolution of two

20

R. K. BEATSON AND H.-Q. BUI

characteristic functions. It follows easily from formula (8.2) that the spherical covariance,
normalised to have value 1 at zero, and to be supported on a ball radius 1, is

C3 (x) =


3
1
1 |x| + |x|3 .
2
2

Other frequently used positive definite compactly supported basic functions in R3 are
the Askey function A and the Wendland function W given by
(8.3)
(8.4)

A(r) = 2,0 (r) = (1 r)2+ ,


W (r) = 3,1 (r) = (1

r)4+ (4r

which is Lipschitz
+ 1),

which is C2 .

To use these functions at a length scale R one simply replaces r by |x|/R.


We consider implicit smoothing of these basic functions by convolving them with the
kernel kc defined above. The resulting function will clearly be supported on {x : |x| 1 +
c}. Looking in the Fourier domain we see that the smoothed function will not be positive
definite. Positive definiteness is important as it guarantees the existence of solutions to
interpolation problems. However, it does not matter for our implicit smoothing application,
in which one smooths a previous fitted RBF. If one wants the smoothed basic function to
be positive definite then clearly one should choose a compactly supported positive definite
function such as the Askey function or spherical covariance, normalised to have integral 1,
as the smoothing kernel rather than kc .
Maple code based on (8.2) yields piecewise definitions for the smooth approximations
= ?3 kc to these basic functions when 0 < c < 1/2. In particular (A ?3 kc ) (r)
equals





1 4
1
3
3 2

r + 1
r + 1 c+ c ,

10c3
c  
5


2

3 2
2c2 1

r 2r + 1 + c
,
5
5
r
2
1 5
1 4 1 3c 3 1 + c 2 1 + 6c2 + 16c3 + 9c4

r +
r +
r
r

80c3
20c3
16c3
2c
16c3

1 + 10c3 + 15c4 + 6c5

1 5c2 + 15c4 + +16c5 + 5c6 1

,
3
20c
80c3
r

0 r < c,
c r < 1 c,

1 c r < 1 + c.

Also (C3 ?3 kc ) (r) equals





3
1
1
6

r +
2 + 2 r4

140c
40c




 c

3c
1
9
1 3

+
1
r + 1 c+ c ,
0 r < c,

4  c2
8 4




1 3
3 2 3
3 4
3
1
r +
c
r+1+
c c2
,
c r < 1 c,

2
5
2
70
10
r

2
3
2

1
3(1 + 2c ) 4 1 + 4c 3 3(1 c ) 2 3(1 + 5c2 + 10c3 4c5 )

r +
r +
r
r

280c3
80c3
16c3
8c
40c3

3
4
6
2
4
5
7

+ 1 + 8c + 9c 2c 3(3 14c + 35c + 28c 4c ) 1 , 1 c r < 1 + c.


3
3
16c
560c
r

MOLLIFICATION FORMULAS AND IMPLICIT SMOOTHING

21

Finally, (W ?3 kc ) (r) equals





 2
6
6
2
1 8

6
4
2
3

42c3 r + 7c 7c3 r + 15 + 9c + c r + 10 + 30c 30c + 10c r




45 4 3 5

2
3

0 r < c,
+ 1 6c + 10c c + c ,

7
2




5
 3
 2
36 4

4
2
2
2

4r 15r + 20 + 12c r 10 + 30c r + 24c + 7 c r


 


45 4
12 4
4 6 1

+ 1 6c c +
c + c
,
c r < 1 c,

7
7
21
r

2
2
3
1 8
15 7 1 3c 6 1 15c 16c 5
r
r +
r
r
3
3

84c
224c
7c3
8c3

2
2
3
4

3(2 + 5c + 3c ) 4 1 + 30c + 160c + 225c + 96c5 3

r
r +

2c
16c


2
4
5

 2 3 1 + 7c 105c 224c 175c6 48c7

2
3

r
5 1 + 3c + 3c + c r

56c3

3
5
6
7
8

1 + 14c 84c 140c 90c 21c

28c3

2
4

36c
+
126c
420c6 576c7 64c9 1

,
1 c r < 1 + c.
672c3
r
These smoothing formulas can be usefully employed when the original unsmoothed
compactly supported RBF, s, contains many shifts of a single basic function, , with a
constant value of the radius R. Then the coefficients of the powers of r = |x|/R above can
all be precomputed, and the smoothed piecewise basic function, , evaluated reasonably
efficiently by Horners method applied to r or r2 as appropriate.
R EFERENCES
1. M. Abramowitz and I. A. Stegun (eds.), Handbook of mathematical functions, Dover, New York, 1965.
2. N. Aronszajn and K. T. Smith, Theory of Bessel Potentials. Part 1, Ann. Inst. Fourier., Grenoble 11 (1961),
385475.
3. R. K. Beatson and N. Dyn, Multiquadric B-splines, J. Approximation Theory 87 (1996), 124.
4. R. K. Beatson and W. A. Light, Quasi-interpolation by thin plate splines on the square, Constructive Approximation 9 (1993), 407433.
5. J. C. Carr, R. K. Beatson, J. B. Cherrie, T. J. Mitchell, W. R. Fright, B. C. McCallum, and T. R. Evans,
Reconstruction and representation of 3d objects with radial basis functions, Computer Graphics SIGGRAPH
2001 proceedings), 12-17 August 2001, pp. 6776.
6. J. C. Carr, R. K. Beatson, B. C. McCallum, W. R. Fright, T. J. McLennan, and T. J. Mitchell, Smooth surface
reconstruction from noisy range data, GRAPHITE 2003 (New York), ACM Press, 1114 February 2003,
pp. 119126.
7. W. F. Donoghue, Distributions and fourier transforms, Academic Press, New York, 1969.
8. J. Duchon, Splines minimizing rotation-invariant semi-norms in Sobolev spaces, Constructive Theory of
Functions of Several Variables (Berlin) (W. Schempp and K. Zeller, eds.), Lecture Notes in Mathematics, no.
571, Springer-Verlag, 1977, pp. 85100.
9. N. Dyn, Interpolation and approximation with radial and related functions, Approximation Theory VI (C. K.
Chui, L. L. Schumaker, and J. D. Ward, eds.), vol. 1, Academic Press, 1989, pp. 211234.
10. N. Dyn, D. Levin, and S. Rippa, Numerical procedures for surface fitting of scattered data by radial functions,
SIAM J. Sci. Stat. Comput. 7 (1986), no. 2, 639659.
11. G. E. Fasshauer, On smoothing for multilevel approximation with radial basis functions, Approximation
Theory IX (C. K. Chui and L. L. Schumaker, eds.), vol. 2, Vanderbilt University Press, 1999, pp. 5562.
12. I. M. Gelfand and G. E. Shilov, Generalized functions, vol. 1, Academic Press, New York, 1964.
13. M. F. Hutchinson and P. E. Gessler, Splines more than just a smooth interpolator, Geoderma 62 (1994),
4567.
14. D. S. Jones, The theory of generalised functions, Cambridge University Press, Cambridge, 1982.

22

R. K. BEATSON AND H.-Q. BUI

15. C. T. Mouat and R. K. Beatson, RBF collocation, Tech. Report UCDMS 2002/3, Department of Mathematics
and Statistics, University of Canterbury, 2002.
16. R. Schaback, Comparison of radial basis function interpolants, Multivariate approximation from CAGD to
wavelets (Singapore), World Scientific, 1993, pp. 293305.
17.
, Multivariate interpolation and approximation by translates of a basis function, Approximation
Theory VIII, Vol. I (Singapore), World Scientific, 1995, pp. 491514.
18. R. Schaback and Z. Wu, Operators on radial functions, J. Computational and Applied Mathematics 71
(1996), 257270.
19. M. L. Stein, Interpolation of spatial data: Some theory for kriging, Springer-Verlag, New York, 1999.
20. G. Wahba, Spline models for observational data, CBMS-NSF Regional Conference Series in Applied Math,
no. 59, SIAM, 1990.
21. H. Wendland, Piecewise polynomial, positive definite and compactly supported radial functions of minimal
degree, Advances in Computational Mathematics 4 (1995), 389396.
D EPARTMENT OF M ATHEMATICS AND S TATISTICS , U NIVERSITY OF C ANTERBURY, P RIVATE BAG 4800,
C HRISTCHURCH , NEW ZEALAND

Anda mungkin juga menyukai