Rev. Anal. Num´er. Th´eor. Approx., vol. 30 (2001) no. 1, pp. 47–54 ictp.acad.ro/jnaat
A VORONOVSKAYA–TYPE THEOREM
MIRCEA IVAN and IOAN RAS¸A
Dedicated to the memory of Acad. Tiberiu Popoviciu
Abstract. We give an asymptotic estimation for some sequences of divided differences. We use this estimation to obtain a Voronovskaya–type formula in- volving linear positive operators.
MSC 2000. 41A36.
Keywords. Divided difference, linear operator, approximation, Voronovskaya- type theorem.
1. INTRODUCTION AND NOTATIONS
Consider the pointsx0 < x1 < . . . < xnon the real axis and letf: [x0, xn]→ Rbe an arbitrary function. Denote by [x0, . . . , xn;f] the divided difference of the functionf on the knotsx0, . . . , xn,usually defined by
[x0, . . . , xn;f] :=
n
P
i=0
f(xi)
(xi−x0)...(xi−xi−1)(xi−xi+1)...(xi−xn).
Consider the polynomial functions ei:R → R, ei(x) = xi, i= 0,1, . . . . It is known that [x0, . . . , xn;ei] = 0, i = 0, . . . , n−1, [x0, . . . , xn;en] = 1. The problem was to calculate Ak:= [x0, . . . , xn;en+k], k= 1,2, . . . .In [6] Tiberiu Popoviciu uses the identity
h
x0, . . . , xn;x−·1 i
= (x−x 1
0)...(x−xn), to prove the following formula
Technical University of Cluj–Napoca, Dept. of Mathematics, Str. C. Daicovi- ciu 15, RO-3400, Cluj–Napoca, ROMANIA, e-mail: [email protected], [email protected].
Ak= P
0≤i0,...,in≤k i0+···+in=k
xi00· · ·xinn.
This formula was rediscovered in 1981 by E. Neuman [3]. It does not look much “friendlier” than the initial one,
Ak = Pn
i=0
xn+ki
(xi−x0)...(xi−xi−1)(xi−xi+1)...(xi−xn).
Therefore, in [7], it is suggested that a recurrence formula might be more useful. We shall use such a formula in order to give an asymptotic estimation forAk under some supplementary assumptions on the knots (see Theorem 1).
Consider now a triangular matrix of nodes (xn,k), n= 0,1, . . .;k= 0, . . . , n, (1) −1≤xn,0 < xn,1< . . . < xn,n≤1, n= 0,1, . . .
satisfying the conditions:
(2) xn,n−i =−xn,i, i= 0, . . . , n, n= 0,1, . . . .
Leta >0. Forn≥1 consider the operatorLn:C[−a−1, a+ 1]→C[−a, a], Lnf(x) :=n![x+xn,0, . . . , x+xn,n;Fn],
wheref ∈C[−a−1, a+ 1], x∈[−a, a], Fn∈Cn[−a−1, a+ 1], Fn(n)=f.
The Ln are positive linear operators of probabilistic type and Bernstein–
Schnabl type operators.
For particular choices of the matrix (xn,k) various inequalities involvingLnf have been studied in [4], [5], [8], [12]. Ifxn,i=−1 +2in,i= 0, . . . , n, we have also [10]
(3) Lnf(x) = 2−n Z x+1
x−1
· · · Z x+1
x−1
f t1+···+tn n
dt1. . . dtn.
Using theLn operator notation, [7] gives (4)
Lnf(0)−
k−1
P
i=0
Lne2i(0)
(2i)! f(2i)(0)
≤ Lne2k(0)
(2k)! kf(2k)k[−1,1].
for all f ∈C2k[−a−1, a+ 1], where k · k[−1,1] denotes the uniform norm on C[−1,1]. As positive operators,Ln have been studied in [9], [10].
They verify:
f convex =⇒Lnf ≥Ln+1f ≥f, kLnf−fk ≤2ω
f,√1
3n
.
We have: Le0 =e0, Le1 =e1, L(e1−x e0)2(x) = (n+1)(n+2)1 Pn i=0x2n,i. For equidistant knotsxn,i=−1 +2in,i= 0, . . . , n, we obtain
L(e1−x e0)2(x) = 3n1 , hence, using [1, Corollary 4.12], we can prove now that
kLnf−fk ≤2.25ω2 f,√1
3n
.
Our aim is to give a more refined analysis of the convergence behaviour of the operators Ln.This is accomplished in Theorem 2.
2. MAIN RESULTS
Theorem 1. If the triangular matrix (xn,k) satisfies the relations (1), (2) and
(5) lim
n→∞
1 n
n
P
i=0
x2n,i= 2λ∈R then, for all k∈N,the following equality is fulfilled
(6) lim
n→∞n−k[xn,0, . . . , xn,n;en+2k] = λk!k. The previous relation can be written in the form
(7) lim
n→∞ nkLne2k(0) = λk!k(2k)!.
Theorem 2. If f ∈C2k[−a−1, a+ 1], then, for every matrix (xn,k) satis- fying the conditions(1),(2)and(5), the following Voronovskaya–type relation holds true:
(8) lim
n→∞ nk
Lnf(x)−
k−1
P
i=0
Lne2i(0)
(2i)! f(2i)(x)
= λk!kf(2k)(x), uniformly for x∈[−a, a].
3. PROOF OF THEOREM 1.
Consider the polynomial function (e1−xn,0). . .(e1−xn,n),which we write as en+1−Cn,1en− · · · −Cn,ne1−Cn,n+1e0.Consider also the sumsSn,p:=
n
P
i=0
xpn,i, p= 1,2, . . . .Using (2) it is obvious that
(9) Sn,p= 0, for oddp.
Using (1) it can be easily shown that
(10) lim
n→∞
Sn,p
nk = 0, p= 1,2, . . .; k >1.
We write the relation (5) in the form
(11) lim
n→∞
Sn,2
n = 2λ.
The coefficients
Cn,p= (−1)p+1 P
0≤i1<...<ip≤n
xn,i1· · ·xn,ip, can be computed by using Newton’s formulas:
Cn,1 = Sn,1 Cn,p = 1p
Sn,p−
p−1
P
i=1
Sn,iCn,p−i
, p= 2, . . . , n+ 1.
By considering (9) it follows that:
Cn,p= 0, for odd p, and
(12)
Cn,2 = 12Sn,2
Cn,2k = 2k1 Sn,2k−
k−1
X
i=1
Sn,2iCn,2(k−i)
!
, k= 1, . . . ,b(n+ 1)/2c.
We define γk as
γk:= lim
n→∞
Cn,2k
nk , k= 1,2, . . . . We have
γ1= lim
n→∞
Cn,2
n = lim
n→∞
Sn,2
2n =λ, and, from (10) and (12), it follows that
γk=−λkγk−1, k≥2 hence
(13) γk= (−1)k+1λk
k! , k≥1.
Using the divided difference functional, define the numbers:
An,j := [xn,0, . . . , xn,n;en+j], j=−n,−n+ 1, . . . . It is well known that
(14) An,j =
0, ifj=−n, . . . ,−1, 1, ifj= 0.
In order to calculate An,j forj≥1,observe that
[xn,0, . . . , xn,n;ej−1(e1−xn,0). . .(e1−xn,n)] = 0, that is
[xn,0, . . . , xn,n;en+j−Cn,1en+j−1− · · · −Cn,n+1ej−1] = 0.
As a consequence we have An,j =
n+1P
i=1
Cn,iAn,j−i, j= 1,2, . . . and using (14),we find that
(15) An,j =
j
P
i=1
Cn,iAn,j−i, j= 1, . . . , n+ 1.
Using
An,0 = 1
An,1 = Cn,1An,0 = 0 in (15), it can be deduced that
(16) An,p= 0, for odd p,
and hence
(17) An,2k=
k
P
i=1
Cn,2iAn,2(k−i), 1≤k≤ n+12 . By defining
Bk:= lim
n→∞
An,2k
nk , k≥0, we have B0 = 1,and using (17) we find
Bk=
k
P
i=1
γiBk−i, k≥1 i.e.,
(18) Bk=
k
P
i=1
(−1)i+1λi
i! Bk−i, k≥1.
Using (18) we can prove by mathematical induction that
(19) Bk= λk
k!, k≥0 which completes the proof.
4. PROOF OF THEOREM 2
For arbitraryx∈[−a, a] consider the functiongx : [−a−1, a+ 1]→R, gx:=f −
2k
P
i=0
(e1−xe0)i
i! f(i)(x).
Taylor’s formula implies the existence of a point ξ∈(−a−1, a+ 1),|x−ξ| ≤
|x−t|,such that
gx(t) =(t−x)2k
(2k)! (f(2k)(ξ)−f(2k)(x)).
For anyε >0 there exists a number δ >0 such that
|gx(t)| ≤(t−x)2kε for all t∈[−a−1, a+ 1],|t−x|< δ.
Let C be a constant such that |gx(t)| ≤ C δ2k+2, for all x ∈ [−a, a], t ∈ [−a−1, a+ 1]. Consequently, we obtain
|gx(t)| ≤ε(t−x)2k+C(t−x)2k+2 for all x∈[−a, a], t∈[−a−1, a+ 1], that is,
|gx| ≤ε(e1−xe0)2k+C(e1−xe0)2k+2, and so,
|Lngx(x)| ≤ε Ln(e1−xe0)2k(x) +C Ln(e1−xe0)2k+2(x).
Using the equality
Ln(f)(x) =Ln(f ◦(e1+xe0))(0), we obtain
|Lngx(x)| ≤ε Lne2k(0) +C Lne2k+2(0).
Taking into account the fact that (20) Lnei(0) = n!i!
(n+i)!An,i, i= 1,2, . . . it follows
(21) lim
n→∞niLne2i(0) = λi
i! (2i)! i= 1,2, . . . . Consequently, we obtain
n→∞lim nkLngx(x) = 0, uniformly for x∈[−a, a],that is
n→∞lim nk
Lnf(x)−
2k
P
i=0 Lnei(0)
i! f(i)(x)
= 0.
Finally, using (16) the relation (8) is proved.
5. REMARKS
(a) Suppose that (1), (2) and (5) are satisfied and letf ∈C[−a−1, a+ 1]
be 2k–times differentiable atx∈[−a, a].By using the Lemma and [11, Corollary 2] we obtain
(22) lim
n→∞nk
Lnf(x)−
k−1
P
i=0
Lne2i(0)
(2i)! f(2i)(x)
= λk
k!f(2k)(x).
(b) If xn,i=−1 + 2i/n,i= 0, . . . , n, thenλ= 1/6; in this special case the formula (22) can be found in [2]. In particular, for k = 1 andk = 3, we have
(23) lim
n→∞n(Lnf(x)−f(x)) = 16f00(x), respectively
(24) lim
n→∞n
n
n Lnf(x)−f(x)
− f006(x)
−fIV72(x)
= fV I1296(x) −fIV180(x). (c) In the case of Chebyshev’s knots
xn,k= cos2 2k+12(n+1)π, k= 0, . . . , n, we obtain
1 n+1
n
P
k=0
cos2 2k+12(n+1)π= 12, ∀n≥1, henceλ= 1.
Acknowledgement. The authors gratefully acknowledge Heinz H. Gonska for his critical remarks on an earlier version.
REFERENCES
[1] H. H. GonskaandR. K. Kovacheva,The second order modulus revisited: remarks, applications, problems, Conf. Sem. Mat. Univ. Bari, 257, 1994.
[2] M. Ivan and I. Ras¸a, A sequence of positive linear operators, Rev. Anal. Num´er.
Th´eor. Approx.,24(1995), 159–164.
[3] E. Neuman,Problem E 2900, Amer. Math. Month.,88(1981), 538.
[4] E. Neuman and J. Peˇcari´c, Inequalities involving multivariate convex functions, J.
Math. Anal. Appl.,137(1989), 541–549.
[5] J. Peˇcari´c, An inequality for 3–convex functions, J. Math. Anal. Appl., 90 (1982), 213–218.
[6] T. Popoviciu,Introduction `a la th´eorie des diff´erences divis´ees, Bull. Math. de la Soc.
Roumaine des Sci.,42(1940), 65–78.
[7] T. Popoviciu, Remarques sur le reste de certaines formules d’approximation d’une diff´erence divis´ee par les d´eriv´ees, Buletinul Institutului Politehnic din Ia¸si, Serie nou˘a, 13 (17)(1967), 103–109.
[8] J. Peˇcari´c and I. Ras¸a, Inequalities for divided differences of n–convex functions, Studia Univ. Babe¸s–Bolyai, Math.,33(1990), 7–10.
[9] J. Peˇcari´c and I. Ras¸a, A linear operator preserving k–convex functions, Bul. S¸t.
IPCN,33(1990), 23–26.
[10] I. Ras¸a,Korovkin approximation and parabolic functions, Conf. Sem. Mat. Univ. Bari, 236(1991).
[11] P. C. Sikkema,On some linear positive operators, Indag. Math.,32(1970), 327–337.
[12] D. Zwick,A divided difference inequality forn–convex functions, J. Math. Anal. Appl., 104(1984), 435–436.
Received April 11, 2000.