Statistical Approximation to Max-Product

Operators

Sevda Karakuş and Kamil Demirci

Sinop University, Faculty of Sciences and Arts, Department of Mathematics,
57000 Sinop, Turkey

Abstract

In this paper, using the concept of statistical convergence which is stronger than
the statistical convergence, we obtain a statistical approximation theorem for
a general sequence of max-product operators, including Shepard type operators,
although its classical limit fails. We also compute the corresponding statistical
rates of the approximation.

Key words: Statistical convergence, statistical convergence, max-product

operators, Shepard operators.

1 Introduction

Approximation theory, which has a close relationship with other branches of

mathematics, has been used in the theory of polynomial approximation and
various domains of functional analysis [1], and in numerical studies of di¤er-
ential and integral operators [9]. In the classical approximation theory, many
well-known approximating operators obey the linearity condition. In recent
years, Bede et al. [2] have shown that it is possible to …nd some approximat-
ing operators that are not linear, such as, max-product and max-min Shep-
ard type approximating operators. Actually, these operators are pseudo-linear
which is a quite e¤ective structure in solving the problems in many branches of
applied mathematics, such as, image processing [4], di¤erential equations [13],
[14], idempotent analysis [10] and approximation theory [2], [3]. However, so
far, almost all results regarding approximations by pseudo linear operators are
based on the validity of the classical limit of the operators. Using the notion
of statistical convergence Duman [5] obtain various statistical approximation
theorems for a general sequence of max-product approximating operators, in-
cluding Shepard type operators, although its classical limit fails. Recently
various kind of statistical convergence (statistical convergence) which is

Preprint submitted to Elsevier Science 18 August 2010

stronger than the statistical convergence has been introduced by Mursaleen
and Edely [11]. In this paper, using statistical convergence, we give the
Duman’s results in [5].

Let K be the subset of positive integers. Then the natural density of K is

given by (K) = lim n1 jKn j if it exists, where Kn := fk 2 K : k ng and the
n
vertical bars denote the cardinality of the set Kn . A sequence x = fxk g is said
to be statistical convergent to the number L if for each " > 0,

1
lim
n n
jfk n : jxk Lj "gj = 0;

i.e. if the set K = K (") := fk n : jxk Lj "g has natural density zero
[6], [7], [8]. In this case we write st lim x = L.

Recently various kinds of statistical convergence for sequences have been in-
troduced by Mursaleen and Edely [11]. Now we recall this concept.

Let be a mapping of the set of N into itself. A continuous linear functional '
de…ned on the space l1 of all bounded sequences is called an invariant mean
( or mean) [15] if and only if

(i) '(x) 0 when the sequence x = fxk g has xk 0 for all k;

(ii) '(e) = 1, where e = (1; 1; :::);

(iii) '(x) = '((x (n) )) for all x 2 l1 :

Thus, the mean extends the limit functional on c of all convergent sequences
in the sense that '(x) = lim x for all x 2 c [12]. Consequently, c V where V
is the set of bounded sequences all of whose means are equal. It is known
[16] that

V = x 2 l1 : lim
p
tpm (x) = L uniformly in m, L = lim x

where
xm + x
+ x 2 (m) + ::: + x p (m)
(m)
tpm (x) := :
p+1
We say that a bounded sequence x = fxk g is convergent if and only if
x 2 V . Let
( )
s
V = x 2 l1 : st lim tpm (x) = L uniformly in m, L = lim x :
p

A sequence x = fxk g is said to be statistically convergent to L if and only

s
if x 2 V (see, for details, [11]). In this case we write ( ) lim xk = L. That

2
is,
1
limjfp n : jtpm (x) Lj "gj = 0; uniformly in m:
n
n
Using the above de…nitions, the next result follows immediately.

Lemma 1 Statistical convergence implies statistical convergence.

However, one can construct an example which guarantees that the converse of
Lemma 1 is not always true. Such an example was given in [11] as follows:

Example 2 Consider the case (n) = n + 1 and the sequence u = fum g

de…ned as 8
>
< 1; if m is odd,
um = > (1)
: 1; if m is even,
is statistically convergence ( ( ) lim um = 0) but it is neither convergent
nor statistically convergent.

2 Statistical approximation of max-product operators

Let (X; d) be an arbitrary compact metric space. By C (X; [0; 1)) we denote
the space of all non-negative continuous functions on X. Then we consider the
following max-product operators:
n
_
Ln (f ; x) = Kn (x; xk ) f (xk ) , x 2 X and f 2 C (X; [0; 1)) ; (2)
k=0

where xk 2 X, k = 0; 1; :::; n; are the knots; and Kn (x; xk ) are non-negative

continuous functions on X having relatively simple expression (algebraic or
trigonometric polynomials, rational functions, wavelets, etc.) such that, for
any x 2 X,
08 p (m)
91
< _ =
@ p2N: K p (m) (x; xk ) = 1 A = 1 for every m 2 N: (3)
: ;
k=0

holds. Observe that the operators mapping C (X; [0; 1)) into C (X; [0; 1))
are pseudo-linear, i.e., for every f; g 2 C (X; [0; 1)) and for any non-negative
numbers ; ;
_ _
Ln f g; x = Ln (f ; x) Ln (g; x)

is satis…ed (see [2]).

We …rst recall the following lemma.

3
Lemma 3 [2]For any ak ; bk 2 [0; 1), k = 0; 1; :::; n, we have
n
_ n
_ n
_
ak bk jak bk j :
k=0 k=0 k=0

Now we get the following result for the max-product operators.

Theorem 4 Let (X; d) be an arbitrary compact metric space. If, for the op-
erators L := fLn g given by (2) and (3),
n_ o
( ) lim
n
fjLn ('x ; x)j : x 2 Xg = 0 with 'x (y) = d2 (y; x) (4)

then, for all f 2 C (X; [0; 1)), we have

n_ o
( ) lim
n
fjLn (f ; x) f (x)j : x 2 Xg = 0:

PROOF. Let f 2 C (X; [0; 1)) and x 2 X be …xed. Then, using the con-
tinuity of f and also considering the compactness of X, we immediately see
that, for a given " > 0, there exists a positive number such that

2Mf
jf (y) f (x)j "+ 2 'x (y) (5)

_
holds for all y 2 X, where Mf := fjf (y)j : y 2 Xg. Now put
8 p (m)
9
< _ =
K := p2N: K p (m) (x; xk ) = 1 (6)
: ;
k=0

for every m 2 N. Then, by (3) we may write that

jfp n : p 2 Kgj jfp n : p 2 N=Kgj

lim = 1; lim =0 (7)
n n n n

for every m 2 N. So, by (3), (5) and Lemma 3, we get for all p 2 K, that

jtpm (L (f ; x)) f (x)j

Lm (f ; x) + L (m) (f ; x) + ::: + L p (m) (f ; x)

= f (x)
p+1
(Lm (f ; x) f (x)) + L (m) (f ; x) f (x) + ::: + L p (m) (f ; x) f (x)
=
p+1

4
 m m
!
_ _
Km (x; xk ) f (xk ) Km (x; xk ) f (x)
k=0 k=0
=
p+1

0 1
(m) (m)
_ _
@ K (m) (x; xk ) f (xk ) K (m) (x; xk ) f (x)A
k=0 k=0
+ + :::
p+1
0 p (m) p (m)
1
_ _
@ K p (m) (x; xk ) f (xk ) K p (m) (x; xk ) f (x)A
k=0 k=0
+
p+1

m (m)
_ _
Km (x; xk ) jf (xk ) f (x)j + K (m) (x; xk ) jf (xk ) f (x)j
k=0 k=0
+ :::
p+1
p (m)
_
K p (m) (x; xk ) jf (xk ) f (x)j
k=0
+
p+1
m (m)
_ 2Mf _ 2Mf
Km (x; xk ) "+ 2 'x (xk ) + K (m) (x; xk ) "+ 2 'x (xk )
k=0 k=0
p+1
p (m)
_ 2Mf
K p (m) (x; xk ) "+ 2 'x (xk )
k=0
+::: +
p+1
! 0 1
m (m)
2Mf _ 2Mf _
"+ 2 Km (x; xk ) 'x (xk ) + @" + 2 K (m) (x; xk ) 'x (xk )A
k=0 k=0
+ :::
p+1
0 p (m)
1
2Mf _
@" + 2 K p (m) (x; xk ) 'x (xk )A
k=0
+
p+1
2
m (m)
_ _
6 Km (x; xk ) 'x (xk ) + K (x; xk ) 'x (xk )
6 (m)
2Mf 6
6 k=0 k=0
="+ 2 6
6 p+1
4

p (m)
3
_
K p (m) (x; xk ) 'x (xk ) 7
7
k=0 7
+::: + 7
7
p+1 7
5

5
 " #
2Mf Lm ('x ; x) + L (m) ('x ; x) + ::: + L p (m) ('x ; x)
=" + 2
p+1
2Mf
=" + 2 tpm (L ('x ; x)) :

Now, taking the maximum over x 2 X, the last inequality gives, for all p 2 K,
that
_ 2Mf _
fjtpm (L (f ; x)) f (x)j : x 2 Xg "+ 2 fjtpm (L ('x ; x))j : x 2 Xg :
(8)
For a given r > 0, choose an " > 0 such that " < r. Then, it follows from (8)
that

n _ o
p n: fjtpm (L (f ; x)) f (x)j : x 2 Xg r
n
n _ o
p n : p 2 K and fjtpm (L (f ; x)) f (x)j : x 2 Xg r
=
n
n _ o
p n : p 2 N=K and fjtpm (L (f ; x)) f (x)j : x 2 Xg r
+
n
n _ 2 o
(r ")
p n : p 2 K and fjtpm (L ('x ; x))j : x 2 Xg 2Mf
n
n _ o
p n : p 2 N=K and fjtpm (L (f ; x)) f (x)j : x 2 Xg r
+
n
n _ 2 o
(r ")
p n: fjtpm (L ('x ; x))j : x 2 Xg 2Mf
n
jfp n : p 2 N=Kgj
+ :
n
Then, using (7) and the hypothesis (4), we have
n _ o
p n: fjtpm (L (f ; x)) f (x)j : x 2 Xg r
lim ; uniformly in m
n n
for every r > 0. The proof is complete.

The following theorems give the classical and the statistical approximation
to a function f 2 C (X; [0; 1)) by means of the max-product operators Ln ,
respectively.

6
Theorem 5 [5]Let (X; d) be an arbitrary compact metric space. Assume that
the operators Ln given by (2) satisfy the condition
n
_
Kn (x; xk ) = 1 (for n 2 N and x 2 X) :
k=0

If the sequence fLn ('x ; x)gn2N converges uniformly to zero function with re-
spect to x 2 X, then, for all f 2 C (X; [0; 1)), fLn (f ; x)gn2N is also uniformly
convergent to f (x) with respect to x 2 X.

Theorem 6 [5]Let (X; d) be an arbitrary compact metric space. If, for the
operators Ln given by (2) and
( n
)!
_
n2N: Kn (x; xk ) = 1 = 1;
k=0
n_ o
st lim
n
fjLn ('x ; x)j : x 2 Xg = 0 with 'x (y) = d2 (y; x) ;
then, for all f 2 C (X; [0; 1)), we have
n_ o
st lim
n
fjLn (f ; x) f (x)j : x 2 Xg = 0:

Remark 7 We now show that our result Theorem 4 is stronger than its clas-
sical version (Theorem 5) and statistical version (Theorem 6).

Let (X; d) be an arbitrary compact metric space. Consider the Shepard-type

max-product operators (see [3]) as follows:
0 1 m
_ f (xk )
m B
B 1 C
_ C d (x;xk )
B d (x;xk ) C k=0
Sm (f ; x) = B m C f (xk ) = m ; (9)
B_ C _
k=0 @ 1 A 1
d (x;xj ) d (x;xj )
j=0 j=0

where x 2 X, , m 2 N and f 2nC (X; [0; 1)).

o We know from [3] that, for all
f 2 C (X; [0; 1)), the sequence Sm (f ; x) in (9) is uniformly convergent to
f on X. Now, consider the case (n) = n+1. Then, we de…ne the max-product
operators on C (X; [0; 1)) as

Tm (f ; x) = (1 + um ) Sm (f ; x) ; x 2 X and f 2 C (X; [0; 1)) (10)

where the operators Sm are given by (9) and u = fum g is given by (1). Since
( ) lim um = 0, we observe that the sequence of positive linear operators
Tm de…ned by (10) satisfy all hypotheses of Theorem 4. Therefore, for all
f 2 C (X; [0; 1)), we conclude that
n_ o
( ) lim fjTm (f ; x) f (x)j : x 2 Xg = 0:
m

7
However, since fum g is not convergent and statistical convergent, we conclude
that Theorem 5 and Theorem 6 do not work for the operators Tm in (10) while
our Theorem 4 still works.

3 Rate of Statistically Convergence

In this section we compute the rate of statistically convergence of Theorem

4.

De…nition 8 A bounded sequence x = fxn g is statistically -convergent to a

number L with the rate of 2 (0; 1) if for every " > 0,
jfp n : jtpm (x) Lj "gj
lim = 0; uniformly in m:
n n1
Then, this is denoted by

xn L = o(n ) ( ( )) :

Using this de…nition, we obtain the following auxiliary result.

Lemma 9 Let x = fxn g and y = fyn g be bounded sequences. Assume that

xn L1 = o(n 1 ) ( ( )) and yn L2 = o(n 2 ) ( ( )) : Then we have

(i) (xn L1 ) (yn L2 ) = o(n ) ( ( )), where := min f 1; 2g ;

(ii) (xn L1 ) = o(n ) ( ( )), for any real number .
1

PROOF. (i) Assume that xn L1 = o(n 1 ) ( ( )) and yn L2 = o(n 2 ) ( ( )) :

Then, for " > 0, observe that

jfp n : j(tpm (x) L1 ) (tpm (y) L2 )j "gj

(11)
n1
n o n o
" "
p n : jtpm (x) L1 j 2
+ p n : jtpm (y) L2 j 2
n1
n o n o
" "
p n : jtpm (x) L1 j 2
p n : jtpm (y) L2 j 2
+ :
n1 1 n1 2

Now by taking the limit as n ! 1 in (11) and using the hypotheses, we

conclude that
jfp n : j(tpm (x) L1 ) (tpm (y) L2 )j "gj
lim = 0; uniformly in m,
n n1

8
which completes the proof of (i). Since the proof of (ii) is similar, we omit it.

Also we need the following lemma.

Lemma 10 [5]For every ak ; bk 0 (k = 0; 1; :::; n) we have

v v
n u n u n
_ u_ u_
ak b k = t 2t
ak b2k :
k=0 k=0 k=0

Now we recall the concept of modulus of continuity. Let f 2 C (X; [0; 1)).
Then the function ! (f; :) : [0; 1) ! [0; 1) de…ned by
_
! (f; ) = fjf (x) f (y)j : x; y 2 X; d (x; y) g

is called the modulus of continuous of f . In order to obtain our result, we

will make use of the elementary inequality, for all f 2 C (X; [0; 1)) and for
; 2 [0; 1) ;
! (f; ) ( + 1) ! (f; ) :
Then we have the following result.

Theorem 11 Let (X; d) be an arbitrary compact metric space. If the operators

L := fLn g given by (2) and (3) satisfy that

! (f; pm ) = o(n ) ( ( )) on X (12)

r n q o
_
where pm := tpm L ('x ; x) : x 2 X with 'x (y) = d2 (y; x). Then
we have for all f 2 C (X; [0; 1)) ;
_
fjLn (f ; x) f (x)j : x 2 Xg = o(n ) ( ( )) on X:

PROOF. Let f 2 C (X; [0; 1)) and x 2 X be …xed. Consider the set K
given by (6), we can write for every p 2 K and for any > 0, that

jtpm (L (f ; x)) f (x)j

Lm (f ; x) + L (m) (f ; x) + ::: + L p (m) (f ; x)

= f (x)
p+1
(Lm (f ; x) f (x)) + L (m) (f ; x) f (x) + ::: + L p (m) (f ; x) f (x)
p+1

9
 m (m)
_ _
Km (x; xk ) jf (xk ) f (x)j + K (m) (x; xk ) jf (xk ) f (x)j
k=0 k=0
+ :::
p+1
p (m)
_
K p (m) (x; xk ) jf (xk ) f (x)j
k=0
+
p+1
m (m)
_ _
Km (x; xk ) ! (f; d (xk ; x)) + K (m) (x; xk ) ! (f; d (xk ; x))
k=0 k=0
+ :::
p+1
p (m)
_
K p (m) (x; xk ) ! (f; d (xk ; x))
k=0
+
p+1
m
_ d(xk ;x)
! (f; ) Km (x; xk ) 1+
k=0
p+1
(m)
_ d(xk ;x)
! (f; ) K (m) (x; xk ) 1+
k=0
+ + :::
p+1
p (m)
_ d(xk ;x)
! (f; ) K p (m) (x; xk ) 1+
k=0
+
p+1
8 p (m)
>
> m
_ (m)
_ _
>
>
>
>
<
Km (x; xk ) + K (m) (x; xk ) + ::: + K p (m) (x; xk )
k=0 k=0 k=0
= ! (f; ) >
> > p+1
>
>
>
:
0
m (m)
_ _
B Km (x; xk ) d (xk ; x) + K (x; xk ) d (xk ; x)
B (m)
1B
B k=0 k=0
+ B
B p+1
@

p (m)
19
_ >
>
>
K p (m) (x; xk ) d (xk ; x) C>
C>
>
k=0 C=
+::: + C
C>
p+1 C>
A>
>
>
>
;

10
 0_ 8
m h i h i
>
> 1=2 1=2
B >
>
<
Km (x; xk ) Km (x; xk ) d (xk ; x)
1 B k=0
= ! (f; ) >1 + BB
>
> @ p+1
>
:
(m) h i h i
_ 1=2 1=2
K (m) (x; xk ) K (m) (x; xk ) d (xk ; x)
+ k=0 + :::
p+1
p (m)
19
_ h i h i >
>
1=2 1=2 >
K p (m) (x; xk ) K p (m) (x; xk ) d (xk ; x) C>
C>
>
k=0 C=
+ C :
C>
p+1 C>
A>
>
>
>
;

Now, by using Lemma 6, we immediately see that

jtpm (L (f ; x)) f (x)j

8 q q q 9
>
< Lm (d2 (:; x) ; x) + L (d2 (:; x) ; x) + ::: + L (d2 (:; x) ; x) >
=
1 (m) p (m)
! (f; ) >1 + >
: p+1 ;

1 q
= ! (f; ) 1 + tpm L (d2 (:; x) ; x)

holds for every p 2 K and for any > 0. Now taking the maximum over
x 2 X, the last inequality gives for all p 2 K and > 0, that
_ 1 _ q
fjtpm (L (f ; x)) f (x)j : x 2 Xg ! (f; ) 1 + tpm L (d2 (:; x) ; x) : x 2 X :

So, we get _
fjtpm (L (f ; x)) f (x)j : x 2 Xg 2! (f; ) (13)
r n q o
_
where := pm := tpm L ('x ; x) : x 2 X : Hence, given " > 0, it
follows from (13) that

n _ o
p n: fjtpm (L (f ; x)) f (x)j : x 2 Xg "
n1
n _ o
p n : p 2 K and fjtpm (L (f ; x)) f (x)j : x 2 Xg "
=
n1
n _ o
p n : p 2 N=K and fjtpm (L (f ; x)) f (x)j : x 2 Xg "
+
n1

11
 n _ o
p n : p 2 K and fjtpm (L (f ; x)) f (x)j : x 2 Xg "
n1
jfp n : p 2 N=K gj
+
n1
n o
"
p n : p 2 K and ! (f; ) 2 jfp n : p 2 N=K gj
+
n1 n
n o
"
p n : ! (f; ) 2 jfp n : p 2 N=K gj
+ :
n1 n
Then using (7) and the hypothesis (12), we have
_
fj(L (f ; x)) f (x)j : x 2 Xg = o(n ) ( ( )) on X:

The proof is complete.

References

[1] Altomare, F., Campiti, M.: Korovkin-Type Approximation Theory and Its
Applications, Walter de Gruyter, Berlin, 1994.

[2] Bede, B., Nobuhara, H., Daµnková, M. and Nola, A.D.: Approximation by
pseudo-linear operators, Fuzzy Sets & Systems 159, 804-820 (2008).

[3] Bede, B., Nobuhara, H., Fodor, J. and Hirota, K.: Max-product Shepard
approximation operators, J. Adv. Comput. Intelligence Intelligent Informatics
10, 494-497 (2006).

[4] Bede, B., Schwab, E.D., Nobuhara, H. and Rudas, I.J.: Approximation by
Shepard type pseudo-linear operators and applications to image processing,
Internat. J. Approx. Reason. 50, 21-36 (2009).

[5] Duman, O.: Statistical convergence of max-product approximating operators,

Turk. J. Math. 33, 1-14 (2009).

[6] Fast, H.: Sur la convergence statistique, Colloq. Math. 2, 241-244 (1951).

[7] Fridy, J. A.: On statistical convergence, Analysis 5, 301-313 (1985).

[8] Kolk, E.: The statistical convergence in Banach spaces, Acta Et

Commentationes Tartuensis 928, 41-52 (1991).

[9] Korovkin, P.P.: Linear Operators and Approximation Theory, Hindustan Publ.
Co., Delhi, 1960.

[10] Maslov, V.P. and Samborskii, S.N.: Idempotent Analysis, Adv. Soviet Math.,
Vol. 13, Amer.Math. Soc., Providence, RI, 1992.

12
[11] Mursaleen, M., Edely, O.H.H.: On the invariant mean and statistical
convergence, Applied Mathematics Letters 22, 1700-1704 (2009).

[12] Mursaleen, M.: On some new invariant matrix methods of summability, Q. J.

Math. 34, 77–86 (1983).

[13] Pap, E.: Applications of the generated pseudo-analysis to nonlinear partial

di¤erential equations, Contemp. Math. 377, 239-259 (2005).

[14] Pap, E. and Vivona, D.: Noncommutative and nonassociative pseudo-analysis

and its applications on nonlinear partial di¤erential equations, J. Math. Anal.
Appl. 246, 390-408 (2000).

[15] Raimi, R.A.: Invariant means and invariant matrix methods of summability,
Duke Math. J. 30, 81-94 (1963).

[16] Schaefer, P.: In…nite matrices and invariant means, Proc. Amer. Math. Soc. 36,
104–110 (1972).

E-mail addresses: [email protected] (S. Karakuş)

[email protected] (K. Demirci)

13