A NOTE ON THE LIMITED STABILITY

OF SURFACE SPLINE INTERPOLATION

Michael J. Johnson
johnson@mcs.sci.kuniv.edu.kw
Kuwait University
Abstract. Given a nite subset R
d
and data f
|

, the surface spline interpolant to

the data f
|

is a function s which minimizes a certain seminorm subject to the interpolation

conditions s
|

= f
|

. It is known that surface spline interpolation is stable on the Sobolev

space W
m
in the sense that s
L

()
const f
W
m, where m is an integer parameter
which species the surface spline. In this note we show that surface spline interpolation is
not stable on W

whenever < m1/2.

1. Introduction
Let m, d be positive integers with m > d/2. The Beppo-Levi space H
m
is dened to be
the space of tempered distributions f for which D

f L
2
:= L
2
(R
d
) for all || = m, and
the seminorm ||
H
m
is dened by
|f|
H
m
:=
_
_
_||
m

f
_
_
_
L
2
, f H
m
,
where

f denotes the Fourier transform of f. Let
k
denote the space of polynomials
over R
d
having total degree at most k, and let be an arbitrary nonempty subset of R
d
.
Duchon [5] has shown that if is not contained in the zero-set of any nontrivial polynomial
Typeset by A
M
S-T
E
X
1
2 LIMITED STABILITY OF SURFACE SPLINE INTERPOLATION
in
m1
, then for all f H
m
there exists a unique s H
m
which minimizes |s|
H
m
subject
to the interpolation conditions
(1.1) s() = f() for all .
The function s is called the surface spline interpolant to f at , and will be denoted T

f.
In case is nite, Duchon has characterized T

f as the unique function which satises

(1.1) and has the form
T

f = q +

( ),
where
(x) :=
_
|x|
2md
if d is odd
|x|
2md
log |x| if d is even
,
q
m1
, and satises the auxiliary conditions

p() = 0 for all p

m1
.
With the above formulation, the coecients and the polynomial q
m1
can be
readily found by solving a system of linear equations, and this has made surface spline
interpolation an attractive method for interpolating scattered data. The function T

f
is called a radial basis function because its essential part,

( ), is a linear
combination of translates of a single radially symmetric function. For a more general
construction of radial basis function interpolants to scattered data, the reader is referred
to the work of Light and Wayne [10].
Duchon has estimated the error in surface spline interpolation in terms of the ll distance
from to , given by
h := h(, ) := sup
x
inf

|x | .
One formulation of Duchons [6] error estimate is the following:
M.J. JOHNSON 3
Theorem 1.2. Let R
d
have the cone property. There exists h
1
> 0 (depending only
on and m) such that if satises h := h(, ) h
1
, then
g
L
p
()
const(, m)h
md/2+d/p
|g|
H
m
for all g H
m
which vanish on and for all 2 p .
By choosing g = f T

f and noting that |f s|

H
m
2 |f|
H
m
one immediately obtains
the familiar error estimate:
f T

f
L
p
()
const(, m)h
md/2+d/p
|f|
H
m
.
We wish to draw the readers attention to the fact that it follows from this error estimate
that surface spline interpolation is stable on the Sobolev space W
m
(R
d
) in the sense that
if h h
1
, then
T

f
L

()
const(, m) f
W
m
(R
d
)
,
where W

(R
d
) ( 0) is dened to be the space of all f L
2
(R
d
) for which
f
W

(R
d
)
:=
_
_
_(1 + ||
2
)
/2

f
_
_
_
L
2
(R
d
)
< .
To see this, we recall that by the Sobolev imbedding theorem [1, p.97] (as m > d/2),
f
L

(R
d
)
const(d, m) f
W
m
(R
d
)
, f W
m
(R
d
).
Hence,
T

f
L

()
T

f f
L

()
+ f
L

(R
d
)
const(, m)h
md/2
|f|
H
m
+ const(d, m) f
W
m
(R
d
)
const(, m) f
W
m
(R
d
)
,
where we have used the imbedding |f|
H
m
f
W
m
(R
d
)
in the last inequality.
The purpose of this note is to show that surface spline interpolation is not stable on
W

(R
d
) whenever d/2 < < m 1/2. Specically we prove the following:
4 LIMITED STABILITY OF SURFACE SPLINE INTERPOLATION
Theorem 1.3. Let be the closed unit ball {x R
d
: |x| 1}, and assume d/2 < <
m 1/2. For every h
0
> 0 there exists f W

(R
d
) and a sequence of nite pointsets

n
, with h(
n
, ) h
0
, such that T

n
f
L
1
()
as n .
Note that this theorem leaves open the interesting possibility that surface spline interpo-
lation remains stable on W

(R
d
) when m
1
2
.
We mention that the question of whether univariate spline interpolation is stable on
C(R) has been addressed by de Boor [3, pp. 194197]. In contrast to Theorem 1.3,
Brownlee and Light [4] (see also [11] and [12]) show that if the interpolation points are
quasi-uniformly scattered, then in addition to being stable on W

(R
d
) ( Z(
d
2
, m1])
surface spline interpolation actually acheives the expected order of approximation.
2. Construction of a function in W

(R
d
)
Our rst task is to construct the function f which will be used in the proof of Theorem
1.3. Let C

c
([4, 4]) be such that = 1 on [3, 3]. Here C

c
(A) denotes the space
of compactly supported C

functions whose support is contained in A. For > 0 and

R, we dene f
,
C(R) by
f
,
(x) := (x ) |x |

.
Lemma 2.1. If > 0 and 0 < + 1/2, then f
,
W

(R) for all R.

Proof. Assume > 0 and 0 < + 1/2. If is an even integer, then f
,
C

c
(R)
W

(R), so assume is not an even integer. Since f

,
is simply a translate of f
,0
and
W

(R) is translation invariant, we may assume without loss of generality that = 0.

Dene C(R) by (x) := |x|

so that we can write f

,0
= . Note that since has
M.J. JOHNSON 5
at most polynomial growth, is a tempered distribution. It is known [7] that

can be
identied on R\{0} with c ||
1
for some constant c. Writing

+ (1 )c ||
1
=: g
1
+ g
2
,
we see that g
2
admits the estimate | g
2
(w)| const(, )(1 + |w|)
1
whence it readily
follows that g
2
W

(R); hence, g
2
W

(R). As for g
1
, we note that g
1
C

(R) since
g
1
is compactly supported; hence, g
1
C

c
(R) W

(R). Therefore, f
,0
= g
1
+g
2

W

(R).
Let C

c
([1/4, 4]) satisfy 0 1 and = 1 on [1/2, 2]. Let : R
d
[0, )
denote the modulus mapping
(x) := |x| .
We dene the linear operator M : C
c
(R) C
c
(R
d
) by Mf := (f) ; in other words
Mf(x) := (|x|)f(|x|), x R
d
.
Having dened M on C
c
(R), we note that by changing to spherical coordinates one sees
that the L
2
(R
d
)-norm of Mf is dominated by a constant multiple of the L
2
(R)-norm of f:
Mf
2
L
2
= const(d)
_

0
t
d1
|(t)f(t)|
2
dt const(d)
_
4
1/4
|f(t)|
2
dt const(d) f
2
L
2
(R)
.
Since M is linear and C
c
(R) is dense in L
2
(R) it follows that M can be uniquely extended
to a continuous linear operator from L
2
(R) into L
2
(R
d
). For brevity, let us denote this
extension also by M.
Proposition 2.2. For all 0 and f W

(R),
Mf
W

(R
d
)
const(d, , ) f
W

(R)
.
6 LIMITED STABILITY OF SURFACE SPLINE INTERPOLATION
Proof. Let us rst consider the case = 2n for an integer n 0. We will employ the
following equivalent norms for
W
2n
(R)
and
W
2n
(R
d
)
, respectively:
g
W
2n
(R)

2n

k=0
_
_
_g
(k)
_
_
_
L
2
(R)
, g W
2n
(R);
g
W
2n
(R
d
)
g
L
2
(R
d
)
+
n
g
L
2
(R
d
)
, g W
2n
(R
d
),
where :=

2
x
2
1
+

2
x
2
2
+ +

2
x
2
d
denotes the Laplacian operator. Since C

c
(R) is dense
in W
2n
(R), it suces to consider the case f C

c
(R). Put g := f. It is shown in the
proof of [8, Lem. 2] that there exist functions p
0
, p
1
, . . . , p
2m1
C

(0, ) such that

n
(g ) = g
(2n)
() +p
2n1
()g
(2n1)
() + +p
0
()g().
Hence,

n
Mf
2
L
2
(R
d
)
= const(d)
_

0
t
d1

g
(2n)
(t) +p
2n1
(t)g
(2n1)
(t) + +p
0
(t)g(t)

2
dt
const(d, n)
_
4
1/4
2n

k=0

g
(k)
(t)

2
dt const(d, n) g
2
W
2n
(R)
const(d, n, ) f
2
W
2n
(R)
.
Since we have already established
(2.3) Mf
L
2
(R
d
)
const(d) f
L
2
(R)
,
we obtain the desired estimate
(2.4) Mf
W
2n
(R
d
)
const(d, , n) f
W
2n
(R)
.
We consider now the general case 0. Let n be the smallest integer satisfying 2n.
Since M : L
2
(R) L
2
(R
d
) is a linear operator satisfying (2.3) and (2.4), we can interpolate
between these inequalities (see [2, p. 301,302] and [13, p. 39,40]) to obtain
Mf
W

(R
d
)
const(d, , ) f
W

(R)
.
Combining this proposition with Lemma 2.1 yields
Corollary 2.5. If > 0 and 0 < + 1/2, then Mf
,
W

(R
d
) for all R.
M.J. JOHNSON 7
3. Proof of Theorem 1.3
Let us invoke the hypothesis of Theorem 1.3; namely that d/2 < < m 1/2 and
h
0
> 0. We will assume, without loss of generality, that h
0
< min{1, h
1
}, where h
1
is as
described in Theorem 1.2. Let be a non-integer satisfying < + 1/2 < m 1/2. Put

0
:= 1 h
0
/3, and let us assume that (1 h
0
/2,
0
). Note that by Corollary 2.5,
f := Mf
,
0
W

(R
d
).
The construction of Mf
,
0
ensures that f is C

on the compliment of the sphere S

0
:=
{x R
d
: |x| =
0
} and that
f(x) = ||x|
0
|

for 1/2 < |x| < 2.

Let B := {x R
d
: |x| < 1} denote the open unit ball. Although we cannot claim that f
belongs to H
m
, we note that T
B
f is well-dened since there exists g H
m
which satises
g
|
B
= f
|
B
(eg. g = f C

c
(
0
B), with C

c
(
0
B) satisfying = 1 on B).
Lemma 3.1.
T
B
f
L
1
(B)
as
0
.
Proof. Since the seminorm ||
H
m
is rotationally invariant and f is radially symmetric, it
follows that T
B
f is radially symmetric. It is shown in [9, Th. 4.1] that there exists a
polynomial q
m1
and a distribution , with supp B such that T
B
f = q + .
Since is C

on R
d
\{0}, it follows that T
B
f is C

on R
d
\B. In [9, Lem. 5.9], it is
shown that
m
( ) = c for some constant c. In particular,
m
(T
B
f) = 0 on R
d
\B.
To reiterate, we have shown that T
B
f is radially symmetric, C

on R
d
\B, and satises
8 LIMITED STABILITY OF SURFACE SPLINE INTERPOLATION

m
(T
B
f) = 0 on R
d
\B. Consequently, we can invoke [8, Lem. 2] to conclude that
T
B
f can be identied on R
d
\B with an element v
f
of the nite dimensional space
V := span{1,
2
, . . . ,
2m2
}
+
_
span{
2d
,
4d
, . . . ,
2md
} if d is odd,
span{
2d
,
4d
, . . . ,
2
, log ,
2
log , . . . ,
2md
log } if d is even.
Since the functions in V are radially symmetric and C

on R
d
\{0}, for each v V , there
exists a unique v C

(0, ) such that v = v . Similarly, we can write T

B
f =

f
for some

f C([0, )). Note that

f = (
0
)

on [1/2, ] and

f = v
f
on (, ). Since

f H
m
, it follows from [1, Th. 7.55] that
v
f
(j)
() =
d
j
dt
j
(
0
t)

|
t=
for j = 0, 1, . . . , m 1.
In particular, v
f
(m1)
() = ( 1) ( m+ 2)(
0
)
m+1
, and since is not an
integer and m+ 1 < 0, it follows that
(3.2)

v
f
(m1)
()

as
0
.
For t [1/2, 1], let
t
denote the continuous linear functional on V dened by
v,
t
:= v
(m1)
(t),
and note that the family {
t
}
1/2t1
is equicontinuous; hence, there exists a constant c
such that
|v,
t
| c v
L
1
(B\
0
B)
for all 1/2 t 1, v V,
where we have used the fact that
L
1
(B\
0
B)
is a norm on V . It thus follows from (3.2)
that v
f

L
1
(B\
0
B)
as
0
. Since T
B
f = v
f
on B\
0
B, we obtain the desired
conclusion that T
B
f
L
1
(B)
as
0
.
M.J. JOHNSON 9
Proof of Theorem 1.3. It suces to show that for each N > 0 (large), there exists a nite
subset B, with h(, B) h
0
, such that T

f
L
1
(B)
N.
Let N > 0. By Lemma 3.1, there exists (1 h
0
/2,
0
) such that T
B
f
L
1
(B)
> N.
Let
n
be an increasing sequence of nite subsets of B which satisfy h(
n
, B) h
0
/2n,
n N, and note that h(
n
, B) h
0
for all n. We recall that there exists g H
m
such
that g
|
B
= f
|
B
, and hence T
B
f = T
B
g and T

n
f = T

n
g. Duchon [6] has shown that
|T
B
f T

n
f|
H
m
0 as n (see also [9, Th. 1.5]). We invoke Duchons inequality,
Theorem 1.2, to write
T
B
f T

n
f
L

(B)
const(d, m)h
md/2
0
|T
B
f T

n
f|
H
m
0 as n .
It follows that T

n
f
L
1
(B)
T
B
f
L
1
(B)
and hence T

n
f
L
1
(B)
> N for suciently
large n.
References
1. R.A. Adams, Sobolev Spaces, Academic Press, New York, 1975.
2. C. Bennett and R. Sharpley, Interpolation of Operators, Academic Press, New York, 1988.
3. C. de Boor, On bounding spline interpolation, J. Approx. Th. 14 (1975), 191203.
4. R. Brownlee and W. Light, Approximation orders for interpolation by surface splines to rough
functions, IMA J. Numer. Anal. 24 (2004), 179192.
5. J. Duchon, Splines minimizing rotation-invariant seminorms in Sobolev spaces, Constructive Theory
of Functions of Several Variables, Lecture Notes in Mathematics 571 (W. Schempp, K. Zeller, eds.),
Springer-Verlag, Berlin, 1977, pp. 85100.
6. J. Duchon, Sur lerreur dinterpolation des fonctions de plusieur variables par les D
m
-splines,
RAIRO Analyse Numerique 12 (1978), 325334.
7. I.M. Gelfand and G.E. Shilov, Generalized Functions, vol. 1, Academic Press, 1964.
8. M.J. Johnson, A bound on the approximation order of surface splines, Constr. Approx. 14 (1998),
429438.
9. M.J. Johnson, The L
2
-approximation order of surface spline interpolation, Math. Comp. 70 (2000),
719737.
10. W. Light and H. Wayne, Spaces of distributions, interpolation by translates of a basis function and
error estimates, Numer. Math. 81 (1999), 415450.
11. F.J. Narcowich and J.D. Ward, Scattered-Data Interpolation on Spheres:Error Estimates and Locally
supported Basis Functions, SIAM J. Math. Anal. 33 (2002), 1393-1410.
12. F.J. Narcowich and J.D. Ward, Scattered-Data interpolation on R
n
: Error Estimates for Radial
Basis and Band-limited Functions,, SIAM J. Math. Anal. 36 (2004), 284-300.
13. H. Triebel, Theory of function spaces II, Birkh auser, Berlin, 1992.