Paulina Pych-Taberska, Harun Karsli
On the rates of convergence of
Chlodovsky-Kantorovich operators and their B´ezier variant
Abstract. In the present paper we consider the B´ezier variant of Chlodovsky- Kantorovich operators Kn−1,αf for functions f measurable and locally bounded on the interval [0,∞). By using the Chanturiya modulus of variation we estimate the rate of pointwise convergence of Kn−1,αf(x) at those x > 0 at which the one-sided limits f(x+) , f(x−) exist.
2000 Mathematics Subject Classification: 41A25, 41A35.
Key words and phrases: rate of convergence, Chlodovsky-Kantorovich operator, B´ezier basis, Chanturiya’s modulus of variation, pth power variation.
1. Introduction. Let f be a function defined on the interval [0, ∞) and let ℕ := {1, 2, ...} . The Bernstein-Chlodovsky polynomials Cnf applied to f are defined as
(1) Cnf (x) :=
Xn k=0
f
kbn
n
Pn,k
x bn
for x∈ [0, bn], n ∈ ℕ,
where Pn,k(t) =
n k
tk(1 − t)n−k for t ∈ [0, 1] and (bn) is a positive increasing sequence satisfying the properties
(2) lim
n→∞bn= ∞ and limn→∞bn
n = 0
(see [5,11,15]). They were introduced by I. Chlodovsky [5] in 1937 as a generalization of the classical Bernstein polynomials Bnf of a function f on the interval [0, 1] ( defined by (1) in which bn = 1 for all n ∈ ℕ).
Recently, Butzer and Karsli [3] considered the Kantorovich variant of the Chlodovsky- Bernstein polynomials as
(3) Knf (x) := n + 1 bn+1
Xn k=0
Pn,k
x bn+1
(k+1)bn+1Zn+1
kbn+1 n+1
f (u)du if 0 ¬ x ¬ bn+1,
where f ∈ Lloc[0, ∞) i.e. f is locally integrable on the interval [0, ∞) and (bn) satisfies conditions (2). Note that the classical Kantorovich polynomials Kn∗f of a function f integrable on the interval [0, 1] are defined as in (3) in which bn= 1 for all n ∈ ℕ. Several authors have studied approximation properties of those classical Kantorovich polynomials (see e.g. [1, 6, 9, 11, 12, 17, 18] etc.). It is easy to see that the operators Kn and Kn∗ are described in the following way. If F denotes an indefinite integral of f, then Kn∗f (x) = dxd Bn+1F (x) and Knf (x) = dxdCn+1F (x). As is well- known, if f is Lebesgue integrable on [0, 1] then
nlim→∞Kn∗f (x) = f (x) almost everywhere on [0, 1].
Analogous relation holds true for operators (3). Namely (see [14]) if f ∈ Lloc[0, ∞) and if
nlim→∞
bn
Z
0
|f(u)| du exp(−σn bn
) = 0 for each σ > 0,
then
nlim→∞Knf (x) = f (x) almost everywhere on [0,∞),
i.e. at every x > 0 at which F0(x) = f(x). Some modified Chlodovsky-Kantorovich operators are considered also in [8].
In this paper we introduce the B´ezier variant of the Chlodovsky-Kantorovich operators (3) of order (n − 1) for f ∈ Lloc[0, ∞) as
(4) Kn−1,αf (x) := n bn
nX−1 k=0
Q(α)n−1,k
x bn
(k+1)bnZn
kbnn
f (u)du if 0¬ x ¬ bn,
where α > 0, Q(α)n−1,k(t) = Jnα−1,k(t) − Jnα−1,k+1(t) and Jn−1,k(t) are the B´ezier basis functions defined for t ∈ [0, 1] as
Jn−1,k(t) =
n−1
X
j=k
Pn−1,j(t) if k = 0, 1, . . . , n − 1,
Jn−1,n(t) = 0 (see [2]). Clearly, if α = 1 then Kn−1,αf reduce to operators (3) with n replaced by n− 1. Our paper is concerned with the rate of pointwise convergence of operators (4) when f ∈ Mloc[0, ∞),i.e. f is measurable and locally bounded
on [0, ∞). By using the Chanturiya modulus of variation we present estimates for the rate of convergence of Kn−1,αf (x) at the points x of continuity of f and at the discontinuity points of the first kind of f. We will formulate our results for Kn−1,αf with α > 0. The corresponding estimates for the Chlodovsky-Kantorovich polynomials Kn−1f follow immediately as a special case α = 1.
At first, we give the following definition.
Definition 1.1 . Let g be a bounded function on a compact interval I = [a, b].
The modulus of variation νk(g; [a, b]) of the function g is defined for nonnegative integers k as follows:
ν0(g; [a, b]) := 0 and for k 1
νk(g; [a, b]) := sup
Πk
k−1
X
j=0
|g(x2j+1) − g(x2j)| ,
where Πk is an arbitrary system of k disjoint intervals (x2j, x2j+1), j = 0, 1, ..., k −1 i.e., a ¬ x0< x1¬ x2< x3...¬ x2k−2< x2k−1¬ b ( see [4,10] ).
Some properties of the modulus of variation and its applications can be found e.g. in [4] ( see [10,13] also). In particular, νk(g; I) is a non-decreasing function of k∈ ℕ,
νk(g; I) ¬ 2k sup
t∈I|g(t)| , νk(g; Z) ¬ νk(g; I) for any compact interval Z contained in I.
Let p 1. The pth power variation of a function g on a compact interval I = [a, b]
is denoted by Vp(g, I) and is defined as the upper bound of the set of numbers
X
j
|g(sj) − g(tj)|p
1/p
over all finite systems of non-overlapping intervals (sj, tj) ⊂ I. If f is of bounded pth power variation on I, then for every k∈ ℕ,
(5) νk(g; I) ¬ k1−1/pVp(g, I),
(see [4]). The class of all functions of bounded pth power variation on every compact interval contained in [0, ∞) will be denoted by BVlocp [0, ∞).
Throughout this paper we will assume that the sequence (bn) satisfies the funda- mental conditions (2). The symbols cγ and cγ,ρwill denote some positive constants, not necessarily the same at each occurrence, depending only on the indicated para- meter γ and parameters γ, ρ, respectively. The symbol [a] will denote the greatest integer not greater than a. Moreover, if x ∈ (0, ∞) the intervals Hx(u) = [x−u, x+u]
for 0 < u ¬ x will be used.
Theorem 1.2 Let f ∈ Mloc[0, ∞) and let the one-sided limits f(x+), f(x−) exist at a fixed point x ∈ (0, ∞). Then, for all integers n 2 such that bn> 2x, n/bn N (α, x) one has
Kn−1,αf (x)− 1
2αf (x+)− (1 − 1
2α)f(x−)
¬ 2v1(gx; Hx(xp bn/n))
+cαϕn(x) x2
m−1X
j=1
vj(gx; Hx(jxp bn/n))
j3 +vm(gx; Hx(x)) m2
+cα,qM (bn; f) x2q
bn
n
q
ϕqn(x) + 4κα
rbn
n
s bn
x (bn− x)|f(x+) − f(x−)| , where m := [p
n/bn], M(b; f) := sup
0¬t¬b|f(t)| , ϕn(x) = x 1 − bxn
+bnn,
(6) gx(t) :=
f (t)− f(x+) if t > x,
0 if t = x,
f (t)− f(x−) if 0 ¬ t < x,
κα = max {1, α} , q is an arbitrary positive integer, N(α, x) = 4 if α 1 and N (α, x) = max{4, 50/x} if 0 < α < 1. If α 1 then cα= α and cα,q= αcq. Retaining the notation used in Theorem 1.2 and using inequality (5), we get
Theorem 1.3 Let f ∈ BVlocp [0, ∞), p 1, and let x ∈ (0, ∞). Then, for all integers n 2 such that bn> 2x, n/bn N(α, x) one has
Kn−1,αf (x)− 1
2αf (x+)− (1 − 1
2α)f(x−) ¬
2 + cαϕn(x) x2
Vp(gx; Hx(xp bn/n))
+cαϕn(x) x2
rbn
n
!1+1/p [n/bn] X
k=1
Vp(gx; Hx(x/√
√ k)
k1−1/p +cα,qM (bn; f) x2q
bn
n
q
ϕqn(x)
+4κα
rbn
n
s bn
x (bn− x)|f(x+) − f(x−)| .
Here we note that in some subclasses of Mloc[0, ∞) the right-hand sides of the inequalities mentioned in the hypotheses of our theorems tend to zero as n → ∞.
In view of (3) we have m = hp n/bn
i → ∞ as n → ∞. So, in Theorem 1.2 it is enough to consider only the term
Λm(x) =
mX−1 j=1
vj(gx; Hx(jxdn))
j3 where dn =p
bn/n.
Clearly,
Λm(x) ¬
mX−1 j=1
v1(gx; Hx(jxdn)) j2 ¬ 4dn
mdZ n
dn
v1(gx; Hx(xt)) t2 dt
¬ 4dn
m+1Z
1
v1(gx; Hx(x
s))ds ¬ 4 m
Xm k=1
v1(gx; Hx(x k)).
Since the function gxis continuous at x and v1(gx; Hx(x/k)) denotes the oscillation of gxon the interval Hx(x/k), we have
klim→∞v1(gx; Hx(x k)) = 0 and consequently
mlim→∞Λm(x) = 0.
As regards Theorem 1.3, it is easy to verify that in view of continuity of gx at x,
mlim→∞
1 m1+1/p
(m+1)2
X
k=1
√ 1
k1−1/pVp(gx; Hx( x
√k)) = 0.
So, we get the following approximation theorem.
Corollary 1.4 Suppose that f ∈ Mloc[0, ∞) ( in particular, f ∈ BVlocp [0, ∞) , p 1) and that there exists a positive integer q such that
nlim→∞
bn
n
q
M (bn; f) = 0.
Then, at every point x ∈ (0, ∞) at which the limits f(x+), f(x−) exist, we have
n→∞limKn−1,αf (x) = 1
2αf (x+) + (1− 1
2α)f(x−).
Obviously, the above relations hold true for every measurable function f bounded on [0, ∞), in particular for every function f of bounded pth power variation (p 1) on the whole interval [0, ∞).
2. Auxiliary results. We now mention certain results which are necessary to prove our main theorem. For this, let us introduce the following notation. Given any x ∈ [0, bn] and any non-negative integer q we will write
ψqx(t) := (t − x)q f or t∈ [0, ∞),
µn−1,q(x) := Kn−1ψqx(x) ≡ n bn
n−1
X
k=0
Pn−1,k
x bn
(k+1)bnZn
kbnn
(t − x)qdt.
Lemma 2.1 If n ∈ ℕ (n 2) , x ∈ [0, bn], then
µn−1,0(x) = 1, µn−1,1(x) =bn− 2x 2n , µn−1,2(x) = bn
n
n− 2 n x
1 − x
bn
+ bn
3n
¬bn
nϕn(x) and, for q > 1,
(7) µn−1,2q(x) ¬ cq
bn
n
q
ϕqn(x), where
ϕn(x) = x
1 − x
bn
+bn
n.
Proof Formulas for µn−1,0(x), µn−1,1(x), µn−1,2(x) follow by simple calculation.
Now, we will prove (7). Suppose q > 1 and put y := x/bn. Then y∈ [0, 1] and
µn−1,2q(x) = n bn
nX−1 k=0
Pn−1,k(y)
(k+1)bn
Zn
kbnn
(t − ybn)2qdt
= nb2qn
nX−1 k=0
Pn−1,k(y)
k+1
Zn k n
(s − y)2qds = b2qnKn∗−1ψ2qy (y) , (8)
where Kn−1∗ is the classical Kantorovich operator of order n − 1. Clearly, Kn∗−1ψ2qy (y) = d
dyBnG(y), where G is an indefinite integral of ψy, i.e.
G(y) = Zy
0
(t − y)2qdt for y ∈ [0, 1]
and BnG is the Bernstein polynomial of order n of the function G.
Hence (see [6, formula 9.5.14]), Kn−1∗ ψy2q(y) = 1
2q + 1 d
dyBnψy2q+1(y) + Bnψy2q(y)
= 1
n2q
1
n (2q + 1) d
dyTn,2q+1(y) + Tn,2q(y)
(9) ,
where
Tn,r(y) = Xn k=0
Pn,k(y) (k − ny)r, y∈ [0, 1] .
As is known (see [6, Lemma 9.4.3] or [7, Lemma 3.6]), for Tn,r(y) there holds the recursion relation
Tn,0(y) = 1, Tn,1(y) = 0,
Tn,r+1(y) = nry(1 − y)Tn,r−1(y) + y(1 − y) d
dyTn,r(y) for r 2.
From this relation there follow by induction the following representation formulas
Tn,2q(y) = Xq j=1
βj,2q(n) (ny(1 − y))j,
Tn,2q+1(y) = (1 − 2y) Xq j=1
βj,2q+1(n) (ny(1 − y))j,
where βj,2q(n), βj,2q+1(n) are real numbers independent of y and bounded uniformly in n ( see [7, Corollary 3.7]).
Observing that
d
dyy(1− y) = 1 − 2y and that
(1 − 2y)2= 1 − 4y(1 − y) we easily get from (9) that
Kn−1∗ ψy2q(y) = 1 n2q
Xq j=0
ηj,q(n) (ny(1 − y))j,
where |ηj,q(n)| ¬ γj,q and γj,q are non-negative numbers depending only on j and q. Now, it is enough to use the following inequalities:
a ) if y ∈0,n1
or y ∈
1 − 1n, 1
, then ny(1− y) ¬ (n − 1)/n < 1 and
Xq j=0
ηj,q(n) (ny(1 − y))j ¬
Xq j=0
γj,q,
b ) if y ∈1
n, 1−n1
, then (ny(1− y))−1¬ n/(n − 1) ¬ 2 and
Xq j=0
ηj,q(n) (ny(1 − y))j
¬ (ny(1 − y))q Xq j=0
2q−jγj,q.
Consequently,
Kn−1∗ ψy2q(y) ¬ cq
n2q (1 + (ny(1 − y))q)
with
cq = Xq j=0
γj,q2q−j, and by (8)
µn−1,2q(x) ¬ cq
bn
n
2q 1 +nx
bn
1 − x
bn
q
.
This inequality is equivalent to the one to be proved. ■ Note that inequality (7) can be obtained also by using the representation formulas for Tn,2q(y) and Tn,2q+1(y) given in [6, Lemma 9.5.5].
Lemma 2.2 Let 0 < x < bn, α > 0and let
Φn−1,α(x, t) := n bn
nX−1 k=0
Q(α)n−1,k(x bn
)χn,k(t),
where χn,k(t)are the characteristic functions of the intervals In,k:= [kbnn,(k+1)bn n), k = 0, 1, ..., n − 1.
(i) If 0 < s < x then
(10)
Zs
0
Φn−1,α(x, t) dt ¬ καcq
(x − s)2q
bn
n
q ϕqn(x),
(ii) If x < s < bn then
(11)
bn
Z
s
Φn−1,α(x, t) dt ¬ cα,q
(s − x)2q
bn
n
q
ϕqn(x).
In the above inequalities
ϕn(x) = x
1 − x
bn
+bn
n,
qis an arbitrary positive integer, κα= max {1, α} . If α 1 then cα,q = αcq, c1= 1.
Proof (i) Let 0 < s < x < bn. If α 1 then
Q(α)n−1,k
x bn
= Jnα−1,k
x bn
− Jnα−1,k+1
x bn
¬ αPn−1,k
x bn
. Hence
Zs
0
Φn−1,α(x, t) dt ¬ αn bn
nX−1 k=0
Pn−1,k(x bn
) Zs
0
χn,k(t)dt ¬ α
(x − s)2qµn−1,2q(x).
Consider now the case 0 < α < 1. Let s ∈ [k∗bn/n, (k∗+ 1) bn/n). Then we can write s = (k∗+ ε) bn/n where 0¬ ε < 1 and
Zs
0
Φn−1,α(x, t) dt = n bn
nX−1 k=0
Q(α)n−1,k(x bn
) Zs
0
χn,k(t)dt
=
kX∗−1 k=0
Q(α)n−1,k(x bn
) + n bn
Q(α)n−1,k(x bn
)
(k∗+ε)bZ n/n
k∗bn/n
dt
=
kX∗−1 k=0
Q(α)n−1,k(x bn
) + εQ(α)n−1,k∗(x bn
)
= 1 −
n−1X
k=k∗
Q(α)n−1,k(x bn
) + εQ(α)n−1,k∗(x bn
)
= 1 − (1 − ε) Jnα−1,k∗(x
bn) − εJnα−1,k∗+1(x bn
)
¬ 1 − (1 − ε) Jn−1,k∗(x
bn) − εJn−1,k∗+1(x bn
)
=
kX∗−1 k=0
Q(1)n−1,k(x bn
) + εQ(1)n−1,k∗(x bn
)
= Zs
0
Φn−1,1(x, t) dt
¬ 1
(x − s)2qµn−1,2q(x)
(see [13,18]). Thus, estimate (10) follows by Lemma 2.1.
(ii) Let 0 < x < s < bn. If α 1, by using the same method in the proof of (i), one has the same bound, namely
bn
Z
s
Φn−1,α(x, t) dt ¬ α
(x − s)2qµn−1,2q(x) ¬ αcq
(x − s)2q
bn
n
q
ϕqn(x).
If 0 < α < 1 then retaining the symbol k∗ used in the proof of (i) we obtain
bn
Z
s
Φn−1,α(x, t) dt = n bn
n−1
X
k=0
Q(α)n−1,k(x bn
)
bn
Z
(k∗+ε)bn/n
χn,k(t)dt
= n
bnQ(α)n−1,k∗(x
bn)(k∗+ 1) bn
n − s
+
n−1
X
k=k∗+1
Q(α)n−1,k(x bn)
¬ n
bn
Pn−1,kα ∗(x bn
)(k∗+ 1) bn
n − s
+
n−1
X
k=k∗+1
Pn−1,k(x bn
)
!α
¬ 21−α n bn
Pn−1,k∗(x bn
)(k∗+ 1) bn
n − s
+
n−1
X
k=k∗+1
Pn−1,k(x bn
)
!α
¬ 21−α
(s − x)2q∗n
bn
α
Pn−1,k∗(
x bn)
(k∗+1)bn
Zn s
|t − x|2q/αdt
+
n−1
X
k=k∗+1
Pn−1,k(x bn
)
(k+1)bn
Zn
kbnn
|t − x|2q/αdt
α
¬ 21−α
(s − x)2q mn−1,2q/α(x)α
,
where
mn−1,γ(x) = n bn
n−1
X
k=0
Pn−1,k(x bn
)
(k+1)bn
Zn
kbnn
|t − x|γdt.
Write l = 2/α and denote by [l] the integral part of l. As in [18, Lemma 6] choose the numbers
p = 2[l]
2[l] + 2 − l, p0 = 2[l]
l− 2, r = 2
p, r0= 2v
p0, v = [l] + 1.
Clearly, l > 2, p > 1, p0 > 1, 1/p + 1/p0 = 1 and l = r + r0. Applying twice the
H¨older inequality (first for integrals, next for sums) we obtain
mn−1,2q/α(x) ¬ n bn
n−1X
k=0
Pn1/p−1,k(x bn
)Pn1/p−1,k0 (x bn
)
(k+1)bn
Zn
kbn n
|t − x|qr|t − x|qr0dt
¬
bnn
n−1
X
k=0
Pn−1,k(x bn
)
(k+1)bn
Zn
kbnn
|t − x|qrpdt
1/p
bnn
n−1
X
k=0
Pn−1,k(x bn)
(k+1)bn
Zn kbn
n
|t − x|qr0p0dt
1/p0
= (µn−1,2q(x))1/p(µn−1,2vq(x))1/p0.
Applying Lemma 2.1 and observing that 1/p + v/p0= 1/α we get
mn−1,2q/α(x) ¬
cq
bn
n
q
ϕqn(x)
1/p cvq
bn
n
vq
ϕvqn (x)
1/p0
= cα,q
bn
nϕn(x)
q/α
.
Thus inequality (11) follows. ■
In order to formulate the next lemma we introduce the following intervals. If x > 0 we will write
Ix(u) := [x + u, x] ∩ [0, ∞) if u < 0 Ix(u) := [x, x + u] if u > 0.
Lemma 2.3 Let f ∈ Mloc[0, ∞) and let at a fixed point x ∈ (0, ∞) the one-sided limits f(x+), f(x−) exist. Consider the function gx defined by (6) and write dn :=
pbn/n. If h = −x or h = x, then for all integers n such that n/bn 4 we have
Z
Ix(h)
gx(t)Φn−1,α(x, t) dt
¬ v1(gx; Ix(hdn))
+cαϕn(x) x2
m−1X
j=1
vj(gx; Ix(jhdn))
j3 +vm(gx; Ix(h)) m2
! ,
where m := [p
n/bn], ϕn(x) = x (1 − x/bn) + bn/n.If α 1 then cα= 8α.
Proof Restricting the proof to h = −x we define the points tj = x + jhdn for j = 0, 1, 2, ..., m and we denote tm+1= 0. Put Tj= [tj, x] for j = 1, 2, ..., m + 1 and write
Z
Ix(h)
gx(t)Φn−1,α(x, t) dt = Zx
t1
gx(t)Φn−1,α(x, t) dt + Xm j=1
gx(tj)
tj
Z
tj+1
Φn−1,α(x, t) dt
+ Xm j=1
tj
Z
tj+1
(gx(t) − gx(tj)) Φn−1,α(x, t) dt
= : I1(n, x) + I2(n, x) + I3(n, x), say.
Clearly,
|I1(n, x)| ¬ Zx
t1
|gx(t) − gx(x)| Φn−1,α(x, t) dt
¬ v1(gx; T1)
bn
Z
0
Φn−1,α(x, t) dt = v1(gx; T1).
By the Abel lemma on summation by parts and by (10) (with q = 1) we have
|I2(n, x)| ¬ |gx(t1)|
t1
Z
0
Φn−1,α(x, t) dt
+
mX−1 j=1
|gx(tj+1) − gx(tj)|
tj+1
Z
0
Φn−1,α(x, t) dt
¬ καϕn(x) x2
|gx(t1) − gx(x)| +
mX−1 j=1
|gx(tj+1) − gx(tj)| 1 (j + 1)2
= καϕn(x) x2
|gx(t1) − gx(x)| +
mX−2 j=1
Xj k=1
|gx(tk+1) − gx(tk)|
1
(j + 1)2 − 1 (j + 2)2
+ 1
m2 Xm k=1
|gx(tk+1) − gx(tk)|
!
¬ καϕn(x) x2
v1(gx; T1) + 2
m−2X
j=1
vj+1(gx; Tj+1)
(j + 1)3 +vm(gx; Tm) m2
¬ καϕn(x) x2
2
mX−1 j=1
vj(gx; Tj)
j3 +vm(gx; Tm+1) m2
.
Next, in view of (10) (with q = 1) and the Abel transformation,
|I3(n, x)| ¬ Xm j=1
v1(gx; [tj+1,tj])
tj
Z
tj+1
Φn−1,α(x, t) dt
¬ καϕn(x) x2
Xm j=1
v1(gx; [tj+1,tj]) j2
= καϕn(x) x2
Xm k=1
v1(gx; [tk+1,tk]) m2
+
mX−1 j=1
Xj k=1
v1(gx; [tk+1,tk])
1
j2− 1 (j + 1)2
¬ καϕn(x) x2
vm(gx; Tm+1)
m2 + 6
m−1X
j=1
vj(gx; Tj+1) (j + 1)3
¬ καϕn(x) x2
vm(gx; Tm+1)
m2 + 6
Xm j=2
vj(gx; Tj) j3
.
Collecting the results and observing that Tj = Ix(jhdn) we get the desired estimate for h = −x. In case h = x the proof runs analogously; we use inequality (11) instead
of (10). ■
Lemma 2.4 Let α > 0 and let
sgn(α)x (t) :=
2α− 1 if t > x,
0 if t = x,
−1 if 0 ¬ t < x.
Then
Kn−1,αsgn(α)x (x) ¬ 2α
n−1X
k=k0+1
Pn−1,k(x bn
)
!α
− 1 2α
+ 2αQ(α)n−1,k∗(x bn
),
where k0= [(n − 1)x/bn], k∗= [nx/bn].
Proof Since x ∈ [k∗bn/n, (k∗+ 1) bn/n) we have
Kn−1,αsgn(α)x (x) = (2α− 1)
n−1
X
k=k∗+1
Q(α)n−1,k(x bn) −
kX∗−1 k=0
Q(α)n−1,k(x bn
)
+n bn
Q(α)n−1,k∗(x bn
)
(2α− 1)
(k∗+ 1) bn
n − x
−
x−k∗bn
n
,
i.e.
Kn−1,αsgn(α)x (x) = 2α
nX−1 k=k∗+1
Q(α)n−1,k(x
bn) − 1 + 2αQ(α)n−1,k∗(x bn
)n bn ∗
∗
(k∗+ 1) bn
n − x
. The above identity can be rewritten in the form
Kn−1,αsgn(α)x (x) = 2α
n−1
X
k=k∗
Q(α)n−1,k(x
bn) − 1 − 2αQ(α)n−1,k∗(x bn
)
x−k∗bn
n
.
Observing that k∗= k0 or k∗= k0+ 1 and that
nX−1 k=v
Q(α)n−1,k(x bn
) =
nX−1 k=v
Pn−1,k(x bn
)
!α
we easily get the estimate forKn−1,αsgn(α)x (x) as desired. ■
Lemma 2.5 Let x > 0, α > 0 and let k0 = [(n − 1)x/bn]. Then
n−1X
k=k0+1
Pn−1,k
x bn
!α
− 1 2α
¬√
2κα
rbn
n
s bn
x (bn− x),
where κα = max {1, α} and n 2 is so large that bn > 2x, n/bn > n0(α, x) with n0(α, x) = 0if α 1, n0(α, x) = 21/x if 0 < α < 1.
Proof Let y = x/bn. Then k0 = [(n−1)y]. In view of the well-known Berry-Ess´een theorem
X
ny<k¬n
Pn,k(y) − 1 2
¬ 0.8 pny(1− y) (see e.g. [17, Lemma 2]). Consequently, assuming α 1 we have
n−1X
k=k0+1
Pn−1,k
x bn
!α
− 1 2α
¬ α
X
k>(n−1)y
Pn−1,k(y) − 1 2
¬ 0.8α
p(n − 1) y(1 − y)¬√ 2α
rbn
n
s bn
x (bn− x). If 0 < α < 1 then
X
ny<k¬n
Pn,k(y)
α
− 1 2α
¬ 1 pny(1− y)
for all n > 25y(1−y)256 (see [18, Lemma 1]). Hence
nX−1 k=k0+1
Pn−1,k(y)
!α
− 1 2α
=
X
k>(n−1)y
Pn−1,k(y)
α
− 1 2α
¬ 1
p(n − 1) y(1 − y)¬√ 2
rbn
n
s bn
x (bn− x) for all n such that n − 1 > 25y(1−y)256 , i.e. for all n such that bn > 2x and
n/bn > 21/x. ■
Lemma 2.6 Let x > 0, α > 0 and let k∗= [nx/bn].Then
Q(α)n−1,k∗
x bn
¬ κα rbn
n
s bn
x (bn− x),
where κα = max {1, α} and n 2 is so large that bn > 2x, n/bn > n1(α, x) with n1(α, x) = 0if α 1, n1(α, x) = 50/x if 0 < α < 1.
Proof If α 1 then Q(α)n−1,k∗(x/bn) ¬ αPn−1,k∗(x/bn) . The desired estimate follows with κα= α by the known inequality
(12) Pn,k(y) ¬ 1
√2ep
ny(1− y) f or y∈ (0, 1) , k ∈ ℕ (see [16, Theorem 1]). Indeed, putting y = x/bn we have
Q(α)n−1,k∗
x bn
¬ αPn−1,k∗(y) ¬ α
√2ep
(n − 1) y(1 − y) < α rbn
n
s bn
x (bn− x). Suppose now 0 < α < 1. Using the mean-value theorem we get
Q(α)n−1,k∗(y) = αγnα−1,k−1 ∗(y) (Jn−1,k∗(y) − Jn−1,k∗+1(y))
= αγnα−1,k−1 ∗(y)Pn−1,k∗(y)
where Jn−1,k∗+1(y) < γn−1,k∗(y) < Jn−1,k∗(y) . Denoting k0 = [(n − 1) y] we have k∗= k0 or k∗ = k0+ 1.
If k∗= k0 then
γn−1,k∗(y) > Jn−1,k0+1(y) = X
k>(n−1)y
Pn−1,k(y) > 1 4
for all n such that n − 1 > 25y(1−y)256 (see [18, inequality (12)]).