Rev. Anal. Num´er. Th´eor. Approx., vol. 24 (1995) nos. 1–2, pp. 131–145 ictp.acad.ro/jnaat
USING WAVELETS FOR SZ ´ASZ-TYPE OPERATORS
HEINZ H. GONSKA and DING-XUAN ZHOU (Duisburg)
1. INTRODUCTION
The Sz´asz-Mirakjan operators are defined onC[0,∞) as
Sn(f, x) =
∞
X
k=0
f kn
Sn,k(x), (1.1)
Sn,k(x) =e−nx(nx)k!k.
There has been an extensive study on the approximation by these operators. In 1978, Beker [1] extended a result of Berens and Lorentz [2] to the interval [0,∞) and showed that forf ∈CB:=C[0,∞)∩L∞[0,∞), 0< α <2,
(1.2) ω2(f, t) =O(tα) ⇔ |Sn(f, x)−f(x)| ≤M xnα2 .
Here M is a constant independent of n ∈ N and x ∈ [0,∞), ω2(f, t) is the modulus of smoothness defined as
ω2(f, t) = sup
0<h<t
42hf(·) ∞, (1.3)
42hf(x) =
f(x+h)−2f(x) +f(x−h), when h≤x,
0, otherwise.
By (1.2), we know that the second-order Lipschitz functions (i.e., the Lips- chitz functions with respect to the second-order modulus of smoothness) can be
characterized by the rate of convergence of Sz´asz-Mirakjan operators. Another interesting result was given by Totik [10] in 1983 who proved the following equivalence:
(1.4) ω2ϕ(f, t)∞=O(tα) ⇔ kSn(f)−fk∞=O n−α2
, 0< α <2.
Here ωϕ2(f, t)∞ is the so-called Ditzian-Totik modulus of smoothness, which is defined for 1≤p≤ ∞as
ωϕ2(f, t)p= sup
0<h<t
42hϕ(x)f(x) p, (1.5)
ϕ(x) =√ x.
The Sz´asz-Mirakjan operators can not be used forLp(1≤p≤ ∞)-approximation.
For this purpose, we must modify these operators. Two versions of this type are Sz´asz-Kantorovich operators:
(1.6) Kn(f, x) :=
∞
X
k=0
n Z k+1
n k n
f(t)dt Sn,k(x) and Sz´asz-Durrmeyer operators:
(1.7) Dn(f, x) :=
∞
X
k=0
n Z ∞
0
f(t)Sn,k(t)dt Sn,k(x).
These operators can be used for Lp-approximation on [0,∞).In fact, for Ln = Kn orDn,0< α <2, 1≤p <∞, f ∈Lp[0,∞),we have (c.f., [5], [7])
(1.8) ω2ϕ(f, t)p =O(tα) ⇔ kLn(f)−fkp =O n−α2 .
Parallel to Sz´asz-Mirakjan operators, it is natural to consider characteriza- tions of Lipschitz functions by means of the above two versions of Sz´asz-type operators. Early in 1985, Mazhar and Totik [9] modified the Sz´asz-Durrmeyer operators and gave the same equivalence to (1.2). However, their modified op- erators have the disadvantage that they can not be used forLp-approximation.
In fact, for the original Sz´asz-Durrmeyer operators (1.7), Mazhar and Totik [9] posed the open problem to find an inverse theorem to the following direct estimate:
(1.9)
Dn(f, x)−f(x)
≤M ω1
f,
qx n+n12
. Here ω1(f, t) is the modulus of continuity:
(1.10) ω1(f, t) = sup
0<h<t
f(·+h)−f(·) ∞.
In 1991, Guo and the second author [6] solved this problem and showed that for 0< α <1, f ∈CB,
(1.11) ω1(f, t) =O(tα) ⇔
Dn(f, x)−f(x)
≤M xn+n12
α2 . This is the first characterization of Lipschitz functions by means of linear oper- ators which do not reproduce linear functions. Extensions to higher orders of Lipschitz functions and some other discussions can be found in a series of the second author’s (joint) papers [8], [11], [12], [14]. On the other hand, we also showed that for any 1< α <2,there exist no functions{Ψn,α(x)}n∈Nsuch that (1.12) ω2(f, t) =O(tα) ⇔
Dn(f, x)−f(x)
≤MΨn,α(x).
Thus, the second-order Lipschitz functions can not be characterized by means of Sz´asz-Durrmeyer operators when 1< α <2.This happens also for Kantorovich operators, see [14]. To overcome this difficulty, Ye and the second author [11]
introduced a technique of matrices and modified the Kantorovich operators so that they can be used for characterization of second-order Lipschitz functions as well as forLp-approximation.
The puropose of this paper is to introduce a class of Sz´asz-type operators by means of Daubechies’ compactly supported wavelets (see [3], [4]). These operators have the following advantages: Firstly, they have the same moments of finitely many orders as we arbitrarily choose as Sz´asz-Mirakjan operators, hence they can be used to characterize the second-order Lipschiz functions.
Secondly, they can be used for Lp-approximation (1< p≤ ∞) and a similar result to (1.8) holds.
In the following sections we discuss these two aspects. We shall denote byM a constant which may be different at each occurrence.
2. CONSTRUCTION OF SZ ´ASZ-TYPE OPERATORS BY WAVELETS
We recall some facts about Daubechies’ compactly supported wavelets (see [3], [4]).
GivenN ∈N, Daubechies’ compactly supported scaling waveletNφis defined by the following refinement equation
φ(·) = 2
∞
X
k=0
hkφ(2· −k), (2.1)
φ(0) = 1,
where{hk}k∈Z is the finitely sequence given by
∞
X
k=0
hke−ikω = 1+e2−iωN
LN(ω) with
LN(ω)
2 =PN sin2 ω2
N
X
n=0
N−1+n n
sin2 ω2n
.
This function is compactly supported with suppNφ = [0,2N−1]. Moreover, there exists a positive constantβ >0 such that forN ≥2, Nφ∈CβN(R) and for 1≤k≤βN,
(2.2)
Z
R
xk Nφ(x)dx= 0.
In particular, whenN = 1, 1φ=χ[0,1] is the classical Haar basis.
In what follows, we assume thatφ∈L∞(R) has the following properties:
(i) suppφ⊂[0, C] with 0< C <∞.
(ii) Z
R
φ(x)dx= 1,and, for 1≤k≤K,(2.2) is satisfied where K∈N. Then, our Sz´asz-type operators are defined as
(2.3) Ln(f, x) :=
∞
X
k=0
n Z
R
f(t)φ(nt−k)dtSn,k(x).
When K = 0, and φ is the Haar basis, these operators are exactly the Sz´asz- Kantorovich operators. Thus, we see that our operators are extensions of the Sz´asz-Kantorovich operators.
By the moment condition (ii),we have the following theorem.
Theorem2.1. Let Ln(f(t), x)be defined by(2.3). Then, for0≤k≤K,we have
(2.4) Ln tk, x
=Sn tk, x
, x∈[0,∞).
In particular,
(2.5) Ln(1, x) = 1;
(2.6) Ln(t, x) =x.
The moment condition (2.4) is the main improvement to Sz´asz-Kantorovich and Sz´asz-Durrmeyer operators.
3. CHARACTERIZATION OF SECOND-ORDER LIPSCHITZ FUNCTIONS
We need some preliminary results to state our main result in this section.
For our proof, we need Peetre’sK-functional given by (3.1) K2(f, t) := inf
g∈C2[0,∞)∩CB
kf−gk∞+t g00
∞ , t >0.
Since forg /∈L∞[0,∞), kf−gk∞ =∞, this K-functional is equivalent to the modulus of smoothness:
(3.2) M−1ω2(f, t)≤K2(f, t2)≤M ω2(f, t), f ∈CB, 0< t≤1.
Two types of Bernstein-Markov inequalities are necessary for our purpose, which we state as follows.
Lemma3.1. Let f ∈CB.Then we have
kLn(f)k∞≤Mkfk∞; (3.3)
L0n(f)
∞≤M nkfk∞; (3.4)
L00n(f)
∞≤M n2kfk∞; (3.5)
ϕ2L00n(f)
∞≤M nkfk∞. (3.6)
Proof of Lemma 3.1. We observe that
(3.7) Ln(f, x) =
∞
X
k=0
Z C 0
f t+kn
φ(t)dtSn,k(x). Therefore, forf ∈CB, x∈[0,∞),
Ln(f, x) ≤
∞
X
k=0
Z C 0
|φ(t)|dtSn,k(x)kfk∞≤Ckφk∞kfk∞. Hence (3.3) holds.
To prove (3.4) and (3.5), we observe that
(3.8) Sn,k0 (x) =n Sn,k−1(x)−Sn,k(x) . Here we setSn,−1(x)≡0.Then for f ∈CB, x∈[0,∞),
L0n(f, x) =
n
∞
X
k=0
Z C 0
f t+kn
φ(t)dt Sn,k−1(x)−Sn,k(x)
≤2nCkφk∞kfk∞,
and L00n(f, x)
=
n2
∞
X
k=0
Z C 0
f t+kn
φ(t)dt Sn,k−2(x)−2Sn,k−1(x) +Sn,k(x)
≤4n2Ckφk∞kfk∞. Hence (3.4) and (3.5) hold.
Finally, using another expression for the derivative, namely (3.9) Sn,k0 (x) = k−nxx Sn,k(x),
we obtain (3.6):
ϕ2(x)L00n(f, x) =
x
∞
X
k=0
Z C 0
f t+kn
φ(t)dtSn,k(x)
(k−nx)2
x2 −xk2
≤Ckφk∞kfk∞x1
∞
X
k=0
(k−nx)2+k
Sn,k(x)
≤2nCkφk∞kfk∞.
Here we used the moments of Sz´asz-Mirakjan operators:
Sn(t, x) =x, (3.10)
Sn (t−x)2, x
= xn. (3.11)
The proof of Lemma 3.1 is complete.
Lemma 3.2. Let f ∈C2[0,∞)∩CB.Then we have L0n(f)
∞≤M f0
∞; (3.12)
L00n(f)
∞≤M f00
∞; (3.13)
ϕ2L00n(f)
∞≤M ϕ2f00
∞. (3.14)
Proof of Lemma 3.2. By (3.7) and (3.8), for x∈[0,∞),we have L0n(f, x)
=
∞
X
k=0
n Z C
0
f t+k+1n
−f t+kn
φ(t)dtSn,k(x)
≤Ckφk∞ f0
∞,
and L00n(f, x)
=
∞
X
k=0
n2 Z C
0
f t+k+2n
−2f t+k+1n
+f t+kn
φ(t)dtSn,k(x)
≤Ckφk∞ f00
∞.
To prove (3.14), we need the following inequality derived from a result of Becker [1]:
(3.15)
Z h
0
Z h
0 1
x+u+vdu dv≤ (x+2h)(1−x−2h)6h2
with 0< h≤ 18, 0≤x≤1−2h.
Then, forn≥8, x∈(0,∞),using (3.15) with x= 0, h= 1n,we have ϕ2(x)L00n(f, x)
=
xn2
∞
X
k=0
Z C 0
Z 1
n
0
Z 1
n
0
f00 t+kn +u+v
dudv φ(t)dt Sn,k(x)
≤xn2
∞
X
k=0
Z C 0
Z 1
n
0
Z 1
n
0 1
t+k
n +u+vdu dv dt Sn,k(x)kφk∞ ϕ2f00
∞
≤xkφk∞ ϕ2f00
∞ (
C
∞
X
k=1 n
kSn,k(x) + 12CnSn,0(x) )
≤Ckφk∞ ϕ2f00
∞ (∞
X
k=1
2Sn,k+1(x) + 12Sn,1(x) )
≤12Ckφk∞ ϕ2f00
∞.
In the casen <8,(3.14) is easily derived from (3.6).
The proof of Lemma 3.2 is complete.
With the above preparations, we can state our characterization theorem as follows.
Theorem3.3. Let 0< α <2, f ∈CB, Ln(f, x) be given by (2.3). Then
(3.16) ω2(f, t) =O(tα) if and only if
(3.17)
Ln(f, x)−f(x)
≤M xn+n12
α2 .
Proof of Theorem 3.3. Necessity. Suppose that (3.16) holds. By (3.1) , we know that for 0< t <1,
K2(f, t)≤M tα2. Letg∈C2[0,∞)∩CB.We use the Taylor expansion (3.18) g(t) =g(x) +g0(x) (t−x) +
Z t x
(t−u)g00(u)du.
By Theorem 2.1 forx∈(0,∞),we have
|Ln(g, x)−g(x)|=
Ln Z t
x
(t−u)g00(u)du, x
=
∞
X
k=0
Z C 0
Z t+k
n
x
t+k n −u
g00(u)duφ(t)dtSn,k(x)
≤ g00
∞
∞
X
k=0
Z C 0
t+kn −x
2kφk∞dt Sn,k(x)
≤2Ckφk∞ g00
∞
∞
X
k=0
n k n−x2
+ Cn22
o
Sn,k(x)
≤2Ckφk∞ 1 +C2 x
n+n12 g00 ∞. Taking the infimum over g∈C2[0,∞)∩CB,by (3.3) we have
Ln(f, x)−f(x)
≤ inf
g∈C2[0,∞)∩CB
n
kLn(f−g)k∞+kf−gk∞+
Ln(g, x)−g(x) o
≤ inf
g∈C2[0,∞)∩CB
n
(1 +M)kf −gk∞+M xn+n12 g00 ∞
o
≤M K2 f,xn+n12
≤M xn+n12
α2 .
Hence (3.17) holds. The proof of necessity is complete.
Sufficiency. Suppose that (3.17) holds. For d > 0, by (3.2) we choose fd ∈ C2[0,∞)∩CB such that
kf −fdk∞≤M ω2(f, d),
f00
∞≤M d−2ω2(f, d).
Let 0< t≤ 18, x > t, n∈N.Then, by (3.15) whenx≤ 12,Lemmas 3.1 and 3.2, we have
∆2tf(x) ≤
≤
∆2tLn(f) (x) +
∆2t(f−Lnf) (x)
≤ Z t
2
−12
Z τ
2
−t2
L00n(f−fd, x+u+v)
dudv+ Z t
2
−2t
Z t
2
−t2
L00n(fd, x+u+v) dudv
+|f(x−t)−Ln(f, x−t)|+ 2|f(x)−Ln(f, x)|+|f(x+t)−Ln(f, x+t)|
≤min (
M n2kf−fdk∞t2, M nkf −fdk∞ Z t
2 t 2
Z t
2
−t
2
1
x+u+vdudv )
+M fd00
∞t2+ 4M x+tn +n12
α2
≤M nkf−fdk∞minn
nt2,12tx+t2o +M
fd00
∞t2+ 8M
maxx+t
n ,n12
α2
≤M ω2(f, d)t2
max1
n2,x+tn −1
+M d−2ω2(f, d)t2+M
maxx+t
n ,n12
α
2.
Letd= max
nqx+t n ,n1
o .Then
∆2tf(x) ≤
≤M
maxnq
x+t n ,1noα
+M t2
maxnq
x+t
n ,n1o−2
ω2
f,maxnq
x+t n ,n1o
. Now for anyδ∈ 0,18
,we choosen∈Nsuch that
δ
2 <max
qx+t n ,1n
≤δ.
Under this choice, we have ∆2tf(x)
≤Mn
δα+h2ωδ22(f,δ)
o ,
which implies
ω2(f, h)≤M n
δα+h2ωδ2(f,δ)2
o .
By the Berens-Lorentz Lemma (see [2], [5]), we have ω2(f, h) =O(hα) (h→0). Hence (3.16) holds and the sufficiency is true.
The proof of Theorem 3.3 is complete.
4. APPROXIMATION THEOREMS INLp
In this section, we give approximation theorems in Lp (1< p≤ ∞) for our Sz´asz-type operators.
To state the equivalence result, we need the Ditzian-Totik K-functional de- fined by
(4.1) Kϕ,2(f, t)p = inf
g,g0∈A.C.loc,ϕ2g00∈Lp
n
kf−gkp+t ϕ2g00
p
o
, t >0.
ThisK-functional is equivalent to the Ditzian-Totik modulus of smoothness:
(4.2) M−1ω2ϕ(f, t)p ≤Kϕ,2 f, t2
≤M ω2ϕ(f, t), where 1≤p≤ ∞, 0< t <1, f ∈Lp[0,∞).
Some Bernstein-Markov-type inequalities are also needed here.
Lemma 4.1. Let 1≤p≤ ∞, f ∈Lp[0,∞).Then we have kLn(f)kp ≤Mkfkp;
(4.3)
ϕ2L00n(f)
p ≤M nfkfkp. (4.4)
Proof of Lemma 4.1. By the Riesz-Thorin Theorem and Lemma 3.1, we need only to consider the casep= 1.
We note that for 0≤k <∞,
Z ∞ 0
Sn,k(x)dx= 1.
LetC≤C0 ∈N.For f ∈L1[0,∞),we have kLn(f)k1≤
Z ∞ 0
∞
X
k=0
Z C 0
f t+kn
kφk∞dt Sn,k(x)dx
≤ kφk∞
∞
X
k=0 1 n
Z C 0
f t+kn dt
≤ kφk∞C0kfk1. Hence (4.3) holds.
By (3.9), we also have
ϕ2L00n(f) 1 ≤
Z ∞ 0
∞
X
k=0
Z C 0
f t+kn
kφk∞dt x Sn,k(x)
(k−nx)2
x2 +xk2
dx
=
∞
X
k=0
Z C 0
f t+kn
kφk∞dt
k n Z ∞
0
Sn,k−1(x)dx−2k+
+n(k+ 1) Z ∞
0
Sn,k+1(x)dx+n Z ∞
0
Sn,k−1(x)dx
≤2kφk∞
∞
X
k=0
Z C0
0
f t+kn dt
≤2C0kφk∞nkfk1. Hence (4.4) also holds.
The proof of Lemma 4.1 is complete.
Lemma4.2. Let 1≤p≤ ∞, f, f0 ∈A.C.loc, ϕ2f00∈Lp. Then we have
(4.5)
ϕ2L00n(f)
p≤M ϕ2f00
p. Proof of Lemma 4.2. We prove only the casep= 1 again.
By (3.8) and (3.15), forn≥8,we have
ϕ2L00n(f) 1 ≤
≤
∞
X
k=0
Z C 0
Z 1
n
0
Z 1
n
0
f00 t+kn +u+v
dudvdt n(k+ 1) Z ∞
0
Sn,k+1(x)dxkφk∞
≤
∞
X
k=1 n
k(k+ 1) Z C
0
Z 1
n
0
Z 1
n
0
ϕ2f00 t+k2 +u+v
dudvdtkφk∞+
+ Z C
0
Z n
0
Z n
0 1 u+v
ϕ2f00 nt +u+v
du dv dtkφk∞
≤2nkφk∞ Z 1
n
0
Z 1
n
0
( ∞ X
k=1
Z C
0
ϕ2f00 t+kn +u+v dt
) du dv
+kφk∞ Z 1
n
0
Z 1
n
0 n u+v
Z ∞ 0
ϕ2f00(y)
dy du dv
≤2n2kφk∞ Z 1
n
0
Z 1
n
0
C0
ϕ2f00
1du dv+nkφk∞ ϕ2f00
1 12
n
≤M ϕ2f00
1.
The proof of Lemma 4.2 is complete.
We estimate the approximation order first for smooth functions.
Theorem 4.3. Let 1< p≤ ∞, f0∈A.C.loc, f0, ϕ2f00∈Lp.Then we have (4.6) kLn(f)−fkp≤ Mn
f0 p+
ϕ2f00 p
.
Proof of Theorem 4.3. We denote the Hardy-Littlewood maximal function of a locally integrable function g by
(4.7) M(g) (x) = sup
t6=x
Z t x
|g(u)|du
|t−x| . Letx > n1, n∈N. Then, by (3.18) we have
|Ln(f, x)−f(x)|=
Ln Z t
x
(t−u)f00(u)du, x
≤
∞
X
k=0
Z C 0
Z t+k
n
n
t+kn −u
f00(u) du
kφk∞dtSn,k(x)
≤
∞
X
k=0
Z C 0
|t+kn −x|2
x dt Sn,k(x)kφk∞M ϕ2f00 (x)
≤ (
2
∞
X
k=0
(nk−x)2
x Sn,k(x) + 2nC22x
)
Ckφk∞M ϕ2f00 (x)
≤ 2(1+C2)Ckφk∞
n M ϕ2f00 (x). Here we used the fact foru∈[x, t] or [t, x],
|t−u|
u ≤ |t−x|
x .
For 0≤x < n1,we have
|Ln(f, x)−f(x)|=
Ln Z t
x
f0(u)du, x
≤
∞
X
k=0
Z C 0
Z t+k
n
x
f0(u) du
kφk∞dtSn,k(x)
≤
∞
X
k=0
Z C 0
t+kn −x
dtkφk∞dtSn,k(x)M(f0) (x)
≤
C2
n kφk∞+Ckφk∞ v u u t
∞
X
k=0 k n−x2
Sn,k(x)
M f0 (x)
≤ (1+C)Ckφkn ∞M(f0) (x).
Combining the above two cases, we have forx∈[0,∞),
|Ln(f, x)−f(x)| ≤ Mn M ϕ2f00
(x) +M(f0) (x) , which implies for 1< p≤ ∞
kLn(f)−fkp≤ Mn
M ϕ2f00 p+
M f0 p
≤ Mn ϕ2f00
p+ f0
p
.
The proof of Theorem 4.3 is complete.
Finally, with all the above preliminary results, we state our main theorem in this section.
Theorem 4.4. Let f ∈ Lp[0,∞) for 1 < p < ∞ and f ∈ CB for p = ∞.
Then, for 0< α <1, we have
(4.8) kLn(f)−fkp =O n−α if and only if
(4.9) ωϕ2(f, t)p=O t2α
.
Proof of Theorem 4.4. When 1 < p < ∞, we use the following inequality (see [5])
(4.10)
f0
p ≤M
kfkp+ ϕ2f00
p
, f, f0 ∈A.C.loc.
Then the proof for this case can be completed by the standard method (c.f. [5]) using Lemmas 4.1, 4.2 and Theorem 4.3.
When p = ∞, we use the following K-functional introduced by the second author in [13]
K1,2(f, t) = inf
g∈C2[0,∞)∩CB
kf−gk∞+t g0
∞+ ϕ2f00
∞ .
By Lemmas 3.1, 3.2, 4.1, 4.2 and Theorem 4.3, using the equivalence derived in [11], we know that (4.8) holds if and only if
K1,2(f, t) =O(tα), which is equivalent to (4.9).
The proof of Theorem 4.3 is complete.
It can thus be seen our Sz´asz-type operators have the advantages of both Sz´asz-Mirakjan operators and Sz´asz-Kantorovich or Durrmeyer operators.
REFERENCES
[1] M. Becker,Global approximation theorems for Sz´asz-Mirakjan and Baskakov operators in polynomial weight spaces, Indiana Univ. Math. J.27(1978), 127–142.
[2] H. Berens and G. G. Lorentz,Inverse theorems for Bernstein polynomials, Indiana Univ.
Math. J.21(1972), 693–708.
[3] I. Daubechies, Orthonormal bases of compactly supported wavelets,Comm. Pure Appl.
Math.41(1988), 909–996.
[4] I. Daubechies, Ten Lectures on Wavelets, CBMS-NSF Series in Appl. Math. 61, SIAM Publ. Philadelphia, 1992.
[5] Z. Ditzian and V. Totik,Moduli of Smoothness, Springer Series in Computational Math- ematics, vol.9, Springer-Verlag, Berlin/Heidelberg/New York, 1987.
[6] Zhu-Rui Guo and Ding-Xuan Zhou,Approximation theorems for modified Sz´asz operators, Acta Sci. Math. (Szeged)56(1992), 311–321.
[7] M. Heilmann,Direct and converse results for operators of Baskakov-Durrmeyer type, Ap- prox. Theory Appl.5(1989), 105–127.
[8] D. Mache and Ding-Xuan Zhou, Approximation theorems for a class of Bernstein- Durrmeyer operators,preprint, 1993.
[9] S. M. Mazhar and V. Totik,Approximation by modified Sz´asz operators, Acta Sci. Math.
(Szeged)49(1985), 257–269.
[10] V. Totik,Uniform approximation by Sz´asz-Mirakjan type operators, Acta Math. Hungar.
41(1983), 291–307.
[11] Mao-Dong Ye and Ding-Xuan Zhou, A class of operators by means of three-diagonal matrices, J. Approx. Theory78(1994), 239–259.
[12] Ding-Xuan Zhou,On a paper of Mazhar and Totik, J. Approx. Theory72(1993), 290–300.
[13] Ding-Xuan Zhou,Uniform approximation by some Durrmeyer operators,Approx. Theory Appl. 6(1990), 87–100.
[14] Din-Xuan Zhou,On smoothness characterized by Bernstein type operators, J. Approx.
Theory, to appear.
Received May 10, 1994 Department of Mathematics University of Duisburg
D-47057 Duisburg Germany