Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Keywords: This paper investigates the global errors which result when the method of approximate
Approximate approximation approximations is applied to a function defined on a compact interval. By extending the
Gaussian kernels functions to a wider interval, we are able to introduce modified forms of the quasi-inter-
Error estimates polant operators. Using these operators as approximation tools, we estimate upper bounds
Modulus of continuity
on the errors in terms of a uniform norm. We consider only continuous and differentiable
functions. A similar problem is solved for the two-dimensional case.
Crown Copyright Ó 2010 Published by Elsevier Inc. All rights reserved.
1. Introduction
The concept of approximate approximation and related results were first published by Mazya in [2,3].
This approximation method is based on generating functions representing an approximate partition of unity, only. As a
consequence, it does not converge as the grid size tends to zero. It is nonetheless possible to construct formulae of arbitrarily
high order that provide an approximation to any desired level of accuracy. In numerical computations, the lack of conver-
gence does not play an important role because the error chosen can always be less than the machine precision. The main
advantages of the method are related to properties of the generating functions themselves, i.e., simplicity, smoothness
and exponential decay behavior. Various applications of these approximations and aspects of a more general theory have
been investigated extensively by Mazýa and Schmidt in [4–9] and also by Müller and Varnhorn in [12]. We also refer the
reader to Mazýa’s monograph [10], which lays out a systematic theory of the method.
Let C m
½a;b ðm 2 N0 Þ be the space of functions having continuous derivatives up to order m on the interval [a, b]. For k 2 Z, we
define the Gaussian kernel
1 1 2
g k : R ! R; g k ðxÞ :¼ pffiffiffiffiffiffi edðxkÞ :
pd
For n 2 N, d < n, we consider the operators
X
n
k 1 X n
k 1ðnxkÞ2
Gn;d ðf ; xÞ :¼ f g k ðnxÞ ¼ pffiffiffiffiffiffi f e d ;
k¼n
n p d k¼n
n
* Corresponding author.
E-mail address: feilongcao@gmail.com (F. Cao).
0096-3003/$ - see front matter Crown Copyright Ó 2010 Published by Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2010.06.010
726 Z. Chen, F. Cao / Applied Mathematics and Computation 217 (2010) 725–734
The approximate approximation method has mainly been used on functions defined on the whole space. However, some
applications (such as the treatment of differential equations and boundary integral equations) require an efficient approx-
imation procedure for functions defined only on compact intervals. In [12], Müller and Varnhorn applied approximate
approximations to such functions. In contrast to the whole space, a truncation error arises in these cases that has to be con-
trolled. Pointwise estimates and L1-estimates for the resulting total error are also given in [12], where all the constants are
determined explicitly.
For f 2 C 2½1;1 , Müller and Varnhorn [12] obtained the following pointwise estimate (see Theorem 5 of [12]):
rffiffiffiffi !!
1 d 1 d 1 1
jf ðxÞ Gn;d ðf ; xÞj 6 kf k1 ðr d;n ðxÞ þ e0 ðdÞÞ þ kf 0 k1 e1 ðdÞ þ 2e2 þ pffiffiffiffiffiffiffi þ kf 00 k1 cðdÞ;
n 2 p 2p n2
where
2 2d
2ep d 4pep
e0 ðdÞ ¼ ; e1 ðdÞ ¼ ;
1 ep2 d ð1 ep2 d Þ2
2 2
8p2 ep d ð1 þ ep d Þ d d
e2 ðdÞ ¼ ; cðdÞ ¼ 1 þ e0 ðdÞ þ e2 ðdÞ
ð1 ep2 d Þ3 4 2
and
1 2 1 2
!
1 edðnðxþ1Þþ1Þ edðnð1xÞþ1Þ
rd;n ðxÞ ¼ pffiffiffiffiffiffi þ :
pd 1 e2dðnðxþ1Þþ1Þ 1 e2dðnð1xÞþ1Þ
The notation kk1 refers to the maximum norm.
It is obvious that rd, n(±1) ? 1 as d ? +1 for any n 2 N. Furthermore, it is well known that the aim of approximate
approximation is to choose d and n such that jf(x) Gn, d(f, x)j satisfies the expected bounds. In this paper, we modify the
operators Gn, d(f, x) to improve the error estimate for compact sets. In the next section, we will introduce approximate approx-
imations by means of the extending function method. Then the error of approximating continuous functions will be esti-
mated. We will study these errors for functions differentiable to the first and second order. In the third section, we will
consider the two-dimensional case.
In order to prove the main results of this paper we need several Lemmas.
Proof. Since
1 X
nþN 1 X nþN X1 1 X 1
1dðnxkÞ2 1dðnxkÞ2 1dðnxkÞ2 1dðnxkÞ2
pffiffiffiffiffiffi e 1 6 pffiffiffiffiffiffi e e þ pffiffiffiffiffiffi e 1 ¼: I1 þ I2 ;
pd k¼nN pd k¼nN k¼1
pd k¼1
we have
1 nN1
X 1 2 X1
1
2 1 X
1 1 2 1 2
I1 ¼ pffiffiffiffiffiffi edðnxkÞ þ edðnxkÞ ¼ pffiffiffiffiffiffi edðnxþkÞ þ edðnxkÞ
pd k¼1 k¼nþNþ1
pd k¼nþNþ1
!
1 2 2
1 X
1 1
1 2 1 2 1 edðnð1þxÞþNþ1Þ edðnð1xÞþNþ1Þ
¼ pffiffiffiffiffiffi edðnð1þxÞþNþkþ1Þ þ edðnð1xÞþNþkþ1Þ 6 pffiffiffiffiffiffi þ
pd k¼0 pd 1 edðnð1þxÞþNþ1Þ 1 edðnð1xÞþNþ1Þ2
2 2 2
ðN 2 1Þ
2 e d 4 ðNþ1Þ2
6 pffiffiffiffiffiffi 2ðNþ1Þ 6 pffiffiffiffiffiffi e d :
pd e d 1 pd
Setting
1 X 1
1 2
gðxÞ ¼ pffiffiffiffiffiffi edðxkÞ ;
pd k¼1
by the Poisson summation formula (see [15]) we obtain
X
1
2 k2 d
gðxÞ ¼ 1 þ 2 ep cosð2pkxÞ:
k¼1
Z. Chen, F. Cao / Applied Mathematics and Computation 217 (2010) 725–734 727
Hence
X
1
2 k2 d
X1
2
2
2ep d
jgðxÞ 1j 6 2 ep 62 ðep d Þk ¼ ; ð2Þ
k¼1 k¼1
1 e p2 d
which implies
1 X 1 2
2ep d
1dðnxkÞ2 2
I2 ¼ pffiffiffiffiffiffi e 1 6 6 4ep d :
pd k¼1 1 ep2 d
2ðNþ1Þ
;
e d 1
which will be used in Lemma 2.
From (2) we have for x 2 (1, +1)
1 X 1
1 2
pffiffiffiffiffiffi edðxkÞ 6 2: ð3Þ
pd k¼1
Now we define
nþN
1 X k 1
2
I3 ¼ pffiffiffiffiffiffi x edðnxkÞ
pd k¼nN n
It follows
nþN
X 1ðnxkÞ2 ¼ p1ffiffiffi PnþN
ðnxkÞ
1 k d n pd 1 2
pffiffiffiffiffiffi x e k¼nN edðnxkÞ
pd k¼nN n
d 0 X
1 nN1 1 2 1 X
1
1 2
¼ g ðnxÞ þ pffiffiffiffiffiffi ðnx kÞedðnxkÞ þ pffiffiffiffiffiffi ðnx kÞedðnxkÞ
2n n pd k¼1 n pd k¼nþNþ1
d 0 1 X
1
1 2 1 X
1
1 2
¼ g ðnxÞ þ pffiffiffiffiffiffi ðnx kÞedðnxkÞ þ pffiffiffiffiffiffi ðnx þ kÞedðnxþkÞ :
2n n pd k¼nþNþ1 n pd k¼nþNþ1
1 2
h qffiffii qffiffi
The function h defined by hðsÞ ¼ seds is strictly increasing on 0; 2d , has its maximum in the point 2d and decreases
hqffiffi
d
strictly on 2
; þ1 . When 1 6 x 6 1 and k P n + N + 1, we have
rffiffiffi
d d
nx þ k P N þ 1 P P :
2 2
728 Z. Chen, F. Cao / Applied Mathematics and Computation 217 (2010) 725–734
Thus
rffiffiffi Z þ1 ! rffiffiffi !
1 X
1
1 2 1 d 1 1 2 1 d 1 2 N 2 pffiffiffi N2
pffiffiffiffiffiffi ðnx þ kÞedðnxþkÞ 6 pffiffiffiffiffiffi e 2þ yedy dy 6 pffiffiffiffiffiffi e 2 þ e d 6 1 þ de d :
pd k¼nþNþ1 pd 2 N pd 2 d
Similarly
1 X
1
1 2 pffiffiffi N2
pffiffiffiffiffiffi ðk nxÞedðnxkÞ 6 1 þ de d :
pd k¼nþNþ1
and
1 X
1
1 2 pffiffiffi N2
pffiffiffiffiffiffi ðk nxÞedðknxÞ 6 de d ;
pd k¼nþNþ1
we thus have
2 p2 d
pffiffiffi N2
I3 6 16de þ de d :
n
From the equation
1 X 1
1 2 X1
2 2
gðxÞ ¼ pffiffiffiffiffiffi edðxkÞ ¼ 1 þ 2 ep k d cosð2pkxÞ;
pd k¼1 k¼1
it follows that
2 1 X 1
1 2 X1
p2 k2 d
g 0 ðxÞ ¼ pffiffiffiffiffiffi ðx kÞedðxkÞ ¼ 4p ke sinð2pkxÞ
d pd k¼1 k¼1
and
2
2 2 1 X 1
1 2 X1
2 2 2
g 00 ðxÞ ¼ gðxÞ þ pffiffiffiffiffiffi ðx kÞ2 edðxkÞ ¼ 8p2 k ep k d cosð2pkxÞ:
d d pd k¼1 k¼1
nþN
X 2 X
nþN X
1
1 k 1 2 1 1 2 1 1 2
pffiffiffiffiffiffi x edðnxkÞ ¼ pffiffiffiffiffiffi ðnx kÞ2 edðnxkÞ 6 pffiffiffiffiffiffi ðnx kÞ2 edðnxkÞ
pd k¼nN n n pd k¼nN
2 n pd k¼1
2
d d d d
¼ 2 gðnxÞ þ g 00 ðnxÞ 6 2 2 þ e3 ðdÞ :
2n 2 2n 2
The above discussion leads to the following estimation:
nþN
X 2
1 k 1 2 d d
pffiffiffiffiffiffi x edðnxkÞ 6 2 2 þ e3 ðdÞ :
pd k¼nN n 2n 2
We now prove another inequality:
pffiffiffi
1 X 1 2 d ða1Þ2
pffiffiffiffiffiffi edðnxkÞ 6 e d :
pd jnxkj>a a
X Z þ1
1 1 2 2 1 2
pffiffiffiffiffiffi edðnxkÞ 6 pffiffiffiffiffiffi edy dy
pd jnxkj>a pd a1
and applying the inequality (see (3.7.3) of [11])
Z 1 pffiffiffiffi
2 p 2 1 a2
ex dx < min ea ; e ; a > 0;
a 2 2a
a simple computation implies Lemma 4. h
Lemma 5 (see Lemma 2 of [14]). Let [a, b] [c, d]. For a positive integer r and a measurable function F(x) on [a, b] we define
8
>
> FðxÞ; x 2 ½a; b;
>
>P r
>
< a Fða þ 2i ba ða xÞÞ; x 2 ½c; aÞ;
i ac
F 0 ðxÞ ¼ i¼0
> r
>
>
> P
> i ba
: bi Fðb 2 db ðx bÞÞ; x 2 ðb; d;
i¼0
rffiffiffi pffi !
4 ðNþ1Þ2 2 d ð n1Þ2 1
kG0n;d;N ðf Þ f k1 6 kf k1 pffiffiffiffiffiffi e d þ 4ep d þ 2 e d þ 2x f ; pffiffiffi ;
pd n n
where x(f, d): = x(f, d,[1, 1])(d > 0) is the first-order modulus of continuity of f (see [1]).
730 Z. Chen, F. Cao / Applied Mathematics and Computation 217 (2010) 725–734
rffiffiffi pffi !
1 1 X
þ1
1 X 1 d ð n1Þ2
1dðnxkÞ2 1dðnxkÞ2
6 x f ; pffiffiffi pffiffiffiffiffiffi e þ 2kf k1 pffiffiffiffiffiffi e 6 2 x f ; pffiffiffi þ kf k1 e d :
n pd k¼1 pd jnxkj> np ffiffi n n
A main result of [12] was estimating the pointwise error of Gn, d(f, x) f(x) for f 2 C 2½1; 1 . We will now discuss the opera-
tor’s global error for f 2 C 2½1;1 . By Lemma 5 we first extend the function f 2 C 2½1;1 to f2(x) defined on 32 ; 32 as follows:
8
< f ðxÞ;
> x 2 ½1; 1;
f2 ðxÞ ¼ f ð4x 5Þ 5f ð2x 3Þ þ 5f ðx 2Þ; x 2 32 ; 1 ;
>
: 3
f ð4x þ 5Þ 5f ð2x 3Þ þ 5f ðx þ 2Þ; x 2 1; 2 :
2
k k 1 k
f2 ¼ f2 ðxÞ þ f20 ðxÞ x þ f200 ðnk Þ x ; x 2 ½1; 1;
n n 2 n
Z. Chen, F. Cao / Applied Mathematics and Computation 217 (2010) 725–734 731
Collecting Lemmas 1–3 and (6) yields a limit on the absolute value:
1 nþN
X
nþN
1dðnxkÞ2 1 X k
1dðnxkÞ2
jG2n;d;N ðf ; xÞ f ðxÞj 6 kf k1 pffiffiffiffiffiffi e 0
1 þ kf k1 pffiffiffiffiffiffi x e
pd k¼nN
pd k¼nN n
nþN
X 2
1 k 1 2
þ kf200 k1 pffiffiffiffiffiffi x edðnxkÞ
2 pd k¼nN n
pffiffiffi N2 12d
4 ðNþ1Þ2 2 kf 0 k1 p2 d
d
6 kf k1 pffiffiffiffiffiffi e d þ 4ep d þ 32de þ 2 1 þ de d þ 2 2 þ e3 ðdÞ kf 00 k1 :
pd n n 2
This finishes the proof of Theorem 2. h
For f 2 C 1½1;1 , we may build the extension by Lemma 5, but here we use the form:
8 0
< f ð1Þðx 1Þ þ f ð1Þ;
> x 2 1; 32 ;
f1 ðxÞ ¼ f ðxÞ; x 2 ½1; 1;
>
: 0
fþ ð1Þðx þ 1Þ þ f ð1Þ; x 2 32 ; 1 :
We then introduce the new operators for d < n; N 6 2n:
X
nþN
1 k 1ðnxkÞ2
G1n;d;N ðf ; xÞ ¼ pffiffiffiffiffiffi f1 e d ; x 2 ½1; 1:
pd k¼nN n
It is not difficult to obtain that if x 2 [1, 1], we have
"Z rffiffiffi #
nþN
X
X
nþN þ1 pffiffiffiffiffiffi
k xe1dðnxkÞ2 ¼ 1 2 d 1 1
2
1dðnxkÞ2 yd
n jnx kje 6 ye dy þ 2 e 2 6 ðd þ 8dÞ:
k¼nN
n k¼nN n 0 2 n
Following the same methods used to prove Theorem 2, we can show that.
In this section, we consider the two-dimensional case. The two-dimensional Gaussian kernels are
1 x21 þx22
Gðx1 ; x2 Þ ¼ e d ;
pd
we can compute their Fourier transforms,
b 1 ; x2 Þ ¼ ep2 dðx21 þx22 Þ :
Gðx
The Poisson summation formula leads to
1 X
þ1 X
þ1 ðx1 k1 Þ2 þðx2 k2 Þ2 X
þ1 X
þ1
2 2
ep dðk1 þk2 Þ e2piðk1 x1 þk2 x2 Þ :
2
e d ¼
pd k1 ¼1 k2 ¼1 k1 ¼1 k2 ¼1
732 Z. Chen, F. Cao / Applied Mathematics and Computation 217 (2010) 725–734
Hence
where
8
> f ðx1 ; x2 Þ; jx1 j 6 1; jx2 j 6 1;
>
>
< f ðsgnðx Þ; x Þ; 1 < jx1 j 6 32 ; jx2 j 6 1;
1 2
fe ðx1 ; x2 Þ ¼
>
> f ðx1 ; sgnðx2 ÞÞ; jx1 j 6 1; 1 < jx2 j 6 32 ;
>
:
f ðsgnðx1 Þ; sgnðx2 ÞÞ; 1 < jx1 j 6 32 ; 1 < jx2 j 6 32
and
1; x P 0;
sgnðxÞ ¼
1; x < 0:
The modulus of continuity, which will be used in the next theorem, is defined by (see [13])
xðf ; h; hÞ ¼ sup jf ðx1 ; x2 Þ f ðy1 ; y2 Þj; h > 0:
jxi yi j6h
Clearly,
xðfe ; h; hÞ ¼ xðf ; h; hÞ: ð9Þ
We now prove the main result of this section:
Theorem 4. Let f(x1, x2) 2 C[1, 1][1, 1], n > 9. Then for Gn, d, N(f;x1, x2) there holds
rffiffiffi pffi pffiffiffi !
1 1 d ð n1Þ2 d ðN1Þ2 p2 d
kGn;d;N ðf Þ f k1 6 4x f ; pffiffiffi ; pffiffiffi þ 10kf k1 e d þ e d þe :
n n n N
X X
1 nþN nþN
k1 k2 ðnx1 k1 Þ2 þðnx2 k2 Þ2 1 X 1 X1 ðnx1 k1 Þ2 þðnx2 k2 Þ2
Gn;d;N ðf ; x1 ; x2 Þ f ðx1 ; x2 Þ ¼ fe ; e d f ðx1 ; x2 Þe d
pd k1 ¼nN k2 ¼nN
n n pd k1 ¼1 k2 ¼1
X
nþN X
nþN
1 k1 k2 ðnx1 k1 Þ2 þðnx2 k2 Þ2
þ f ðx1 ; x2 ÞI6 ¼ fe ; fe ðx1 ; x2 Þ e d
pd k1 ¼nN k ¼nN 2
n n
1 X X ðnx1 k1 Þ2 þðnx2 k2 Þ2 1 X X ðnx1 k1 Þ2 þðnx2 k2 Þ2
f ðx1 ; x2 Þe d f ðx1 ; x2 Þe d
pd jk1 j>nþN jk2 j>nþN
pd jk1 j>nþN jk2 j6nþN
X þ1
X
ðnx k Þ2 þðnx k Þ2
1 fe k1 ; k2 fe ðx1 ; x2 Þe 1 1 d 2 2
þ
pd n n
j n1 x1 j>p1ffi k2 ¼1
k
n
X X
1 k1 k2 ðnx1 k1 Þ2 þðnx2 k2 Þ2
þ fe ; fe ðx1 ; x2 Þe d
pd n n
j n1 x1 j>p1ffi j n2 x2 j>p1ffi
k k
n n
rffiffiffi pffi
1 1 1 þ1 ðnx1 k1 Þ2 X
X þ1 ðnx2 k2 Þ2 d ð n1Þ2 d 2ðpffin1Þ2
6 x fe ; pffiffiffi ; pffiffiffi e d e d þ 8kfe k1 e d þ 2kfe k1 e d
n n pd k ¼1 k2 ¼1
n n
1
rffiffiffi pffi pffi
rffiffiffi pffi
1 1 d ð n1Þ2
d 2ð n1Þ 2
1 1 d ð n1Þ2
6 4x f ; pffiffiffi ; pffiffiffi þ 8kf k1 e d þ 2kf k1 e d 6 4x f ; pffiffiffi ; pffiffiffi þ 10kf k1 e d ;
n n n n n n n
!2 !2
1 X ðnxkÞ2 1 X ðnxkÞ2 dkf k1 2ðN1Þ2
jD2 j 6 kf k1 pffiffiffiffiffiffi e d 6 kf k1 pffiffiffiffiffiffi e d 6 2
e d ;
pd jkj>nþN pd jnxkj>N N
! ! pffiffiffi
1 X ðnx1 k1 Þ2 1 X
1 ðnx2 k2 Þ2 2 dkf k1 ðN1Þ2
jD3 j 6 kf k1 pffiffiffiffiffiffi e d pffiffiffiffiffiffi e d 6 e d ;
pd jk1 j>nþN pd k2 ¼1 N
pffiffiffi
2 dkf k1 ðN1Þ2
jD4 j 6 e d
N
and
2
jD5 j 6 9kf k1 ep d :
Therefore,
rffiffiffi pffi pffiffiffi !
1 1 d ð n1Þ2 d ðN1Þ2 p2d
kGn;d;N ðf Þ f k1 6 4x f ; pffiffiffi ; pffiffiffi þ 10kf k1 e d þ e d þe :
n n n N
Acknowledgement
This research was supported by the National Natural Science Foundation of China (No. 90818020).
References
[1] G.G. Lorentz, Approximation of Functions, Holt, Rinehart and Winston, New York, 1966.
[2] V. Maz’ya, A new approximation method and its applications to the calculation of volume potentials. Boundary point method, in: 3. DFG-Kolloqium des
DFC-Forschungsschwerpunktes ‘‘Randelementmethoden”, 30 September–5 October, 1991.
[3] V. Maz’ya, Approximate approximation, in: J.R. Whiteman (Ed.), The mathematics of finite elements and applications, Highlights 1993, Wiley & Sons,
Chichester, 1994, pp. 77–104.
[4] V. Mazýa, G. Schmidt, Approximate approximations and the cubature of potentials, Rend. Mat. Acc. Lincei 9 (6) (1995) 161–184.
[5] V. Mazýa, G. Schmidt, On approximate approximations using Gaussian kernels, IMA J. Num. Anal. 16 (1996) 13–29.
[6] V. Mazýa, G. Schmidt, Approximate wavelets and the approximation of pseudodifferential operators, Appl. Comp. Harm. Anal. 6 (1999) 287–313.
[7] V. Mazýa, G. Schmidt, On quasi-interpolation with non-uniformly distributed centers on domains and manifolds, J. Approx. Theory 110 (2001) 124–
145.
[8] V. Mazýa, G. Schmidt, W. Wendland, On the computation of multi-dimensional single layer potential via approximate approximations, Calcolo 40 (1)
(2003) 33–53.
[9] V. Mazýa, G. Schmidt, Potentials of Guassians and approximate wavelets, Math. Machrichten 280 (9–10) (2007).
[10] V. Mazýa, G. Schmidt, Approximate Approximations, American Mathematical Society, 2007.
[11] D.S. Mitrinovic, Analytic Inequalities, Springer-Verlag, 1987.
734 Z. Chen, F. Cao / Applied Mathematics and Computation 217 (2010) 725–734
[12] F. Müller, W. Varnhorn, Error estimates for approximate approximations with Gaussian kernels on compact intervals, J. Approx. Theory 145 (2007)
171–181.
[13] S.M. Nikolskii, Approximation of Functions of Several Variables and Imbedding Theorems, Springer-Verlag, 1975.
[14] A.-R.K. Ramazanov, On approximation by polynomials and rational functions in Orlicz spaces, Anal. Math. 10 (1984) 117–132.
[15] E.M. Stein, G. Weiss, Fourier Analysis on Euclidean spaces, Princeton University Press, Princeton, New Jersey, 1971.