Hacettepe Journal of Mathematics and Statistics Volume 39 (4) (2010), 497 – 514 STATISTICAL FUZZY APPROXIMATION TO FUZZY DIFFERENTIABLE FUNCTIONS BY FUZZY LINEAR OPERATORS Oktay Duman∗ Received 10 : 09 : 2009 : Accepted 18 : 03 : 2010 Abstract In this paper, we obtain fuzzy approximations to fuzzy differentiable functions by means of fuzzy linear operators whose positivity condition and classical limits fail. In order to get more powerful results than the classical approach we investigate the effects of matrix summability methods on the fuzzy approximation. So, we mainly use the notion of A-statistical convergence from summability theory instead of the usual convergence. Keywords: Korovkin theorem, Statistical convergence, Fuzzy approximation, Fuzzy continuity, Fuzzy differentiable function. 2000 AMS Classification: 26 E 50, 41 A 25, 41 A 36. 1. Introduction The classical Korovkin theory is mainly based on two conditions: positivity of linear operators and existence of their classical limits (see [1, 28]). So far, the former has been weakened by considering monotonicity and convexity of the functions being approximated (see, for instance, [2, 8, 9, 26]). Furthermore, by using the notion of statistical convergence, some Korovkin–type approximation theorems have been obtained when the classical limit of a sequence of positive linear operators fails (see [6, 10, 11, 12, 14, 15, 16, 17, 25]). Their fuzzy analogs have also been introduced in [5, 13, 21]. In this study, we obtain various fuzzy approximation results without these two conditions as mentioned above. We use the notions of statistical convergence of sequences of fuzzy numbers and fuzzy differentiability of fuzzy real valued functions to get our theorems. When proving our results, we use not only classical techniques from This research was supported by the Scientific and Technological Research Council of Turkey (TUBITAK); Project No: 108T229. ∗ TOBB Economics and Technology University, Faculty of Arts and Sciences, Department of Mathematics, 06530 Söğütözü, Ankara, Turkey. E-mail: [email protected] 498 O. Duman approximation theory but also new methods from summability theory and fuzzy logic theory. We first recall some basic concepts used in the present paper. The (asymptotic) density of a subset K of N, the set of all natural numbers, is defined by δ(K) := lim j # {n ≤ j : n ∈ K} , j provided the limit exists, where the symbol # {B} denotes the cardinality of the set B. Then, we say that a sequence (xn )n∈N is statistically convergent to a number L (see [18]) if, for every ε > 0, the set {n ∈ N : |xn − L| ≥ ε} has density zero, i.e., δ ({n ∈ N : |xn − L| ≥ ε}) = lim j # {n ≤ j : |xn − L| ≥ ε} = 0. j Now let A = (ajn ) be an infinite non-negative regular summability matrix. Recall that A is said to be regular if limj (Ax)j = L whenever P limj xj = L, where Ax := ((Ax)j ) denotes the A-transform of x given by (Ax)j = ∞ n=1 ajn xn , provided the series converges for each j (see, for instance, [7]). Then, the A-density of a subset K is defined by X δA (K) := lim ajn j n∈K provided the limit exists. Observe that if we take A = C1 = (cjn ), the Cesáro matrix of order one, defined by 1 , if 1 ≤ n ≤ j, cjn := j 0, otherwise, then C1 -density coincides with (asymptotic) density. With the help of A-density, Freedman and Sember [19] introduced the notion of A-statistical convergence, which is a more general method of statistical convergence. Recall that the sequence (xn )n∈N is said to be A-statistically convergent to L if, for every ε > 0, δA {n ∈ N : |xn − L| ≥ ε} = 0; or equivalently X lim ajn = 0. j n: |xn −L|≥ε This limit is denoted by stA -limn xn = L. It is not hard to see that if we take A = C1 , then C1 -statistical convergence coincides with the statistical convergence mentioned above. If A is replaced by the identity matrix, then we get the ordinary convergence of number sequences. We also note that if A = (ajn ) is any non-negative regular summability matrix for which limj maxn {ajn } = 0, then A-statistical convergence is stronger than convergence (see [27]). Actually, every convergent sequence is A-statistically convergent to the same value for any non-negative regular matrix A, but its converse is not always true. Some other results regarding statistical and A-statistical convergence may be found in the papers [20, 30]. As usual, a fuzzy number is a function µ : R → [0, 1], which is normal, convex, upper semi-continuous and the closure of the set supp(µ) is compact, where supp(µ) := {x ∈ R : µ(x) > 0}. We denote the set of all fuzzy numbers by RF . Let [µ]0 := {x ∈ R : µ(x) > 0} and [µ]r := {x ∈ R : µ(x) ≥ r}, (0 < r ≤ 1). Then, it is well-known [22] that, for each r ∈ [0, 1], the set [µ]r is a closed and bounded interval of R. For any u, v ∈ RF and λ ∈ R, it is possible to define uniquely the sum Statistical Fuzzy Approximation 499 u ⊕ v and the product λ ⊙ v as follows: [u ⊕ v]r = [u]r + [v]r and [λ ⊙ u]r = λ[u]r , (0 ≤ r ≤ 1). (r) (r) (r) Now denote the interval [u]r by [u(r) , u+ ], where u(r) ≤ u+ and u(r) , u+ r ∈ [0, 1]. Then, for u, v ∈ RF , define (r) ∈ R for (r) u v ⇐⇒ u(r) ≤ v (r) and u+ ≤ v+ for all 0 ≤ r ≤ 1. Define also the following metric D : RF × RF → R+ by o n (r) (r) D(u, v) = sup max u(r) − v (r) , u+ − v+ . r∈[0,1] In this case, it is known [32] that (RF , D) is a complete metric space on RF with the properties D(u ⊕ w, v ⊕ w) = D(u, v) for all u, v, w ∈ RF , D(λ ⊙ u, λ ⊙ v) = |λ| D(u, v) for all u, v ∈ RF and λ ∈ R, D(u ⊕ v, w ⊕ z) ≤ D(u, w) + D(v, z) for all u, v, w, z ∈ RF . Let f, g : [a, b] ⊂ R → RF be fuzzy valued functions. Then, the distance between f and g on [a, b] is given by n (r) o (r) D∗ (f, g) = sup sup max f (r) (x) − g (r) (x), f+ (x) − g+ (x) . x∈[a,b] r∈[0,1] Now let (µn )n∈N be a fuzzy number valued sequence. Then, Nuray and Savaş [31] introduced the fuzzy analog of statistical convergence by using the fuzzy metric D instead of the classical absolute value in the above definition. So, by a similar idea, one can obtain the following definition of A-statistical convergence of fuzzy valued sequences. We say that a sequence (µn )n∈N of fuzzy numbers is A-statistically convergent to µ ∈ RF , which is denoted by stA -limn D(µn , µ) = 0, if for every ε > 0, δA ({n ∈ N : D(µn , µ) ≥ ε}) = 0, i.e., X lim ajn = 0 j n:D(µn ,µ)≥ε holds. Of course, the case of A = C1 immediately reduces to the statistical convergence of fuzzy valued sequences. Also, replacing A with the identity matrix, we get the classical fuzzy convergence introduced by Matloka [29]. 2. Statistical Fuzzy Approximation A fuzzy valued function f : [a, b] → RF is said to be fuzzy continuous at x0 ∈ [a, b] provided that, whenever xn → x0 , then D (f (xn ), f (x0 )) → 0 as n → ∞. Also, we say that f is fuzzy continuous on [a, b] if it is fuzzy continuous at every point x ∈ [a, b]. The set of all fuzzy continuous functions on [a, b] is denoted by CF [a, b] (see, for instance, [3, 4]). Notice that CF [a, b] is only a cone, not a vector space. Following [24, 33], a fuzzy valued function f : [a, b] → RF is said to satisfy the condition (H) on [a, b] if, for any x, y ∈ [a, b] satisfying x ≤ y, there exists u ∈ RF such that f (y) = f (x) + u. In this case, we call u the H-difference (or, Henstock-difference) of f (y) and f (x), and denote it by f (y) − f (x). For brevity, throughout this paper, when we use the “−” operation of fuzzy numbers, we always assume that the condition (H) is satisfied. 500 O. Duman Assume that f : [a, b] → RF is a fuzzy valued function. Then f is called fuzzy differentiable at x ∈ (a, b) if there exists a f ′ (x) ∈ RF such that the following limits lim h→0+ f (x + h) − f (x) f (x) − f (x − h) , lim h h h→0+ exist and are equal to f ′ (x). Also, if x = a or x = b, then we use the following f ′ (a) := lim h→0+ f (a + h) − f (a) f (b) − f (b − h) and f ′ (b) := lim h h h→0+ provided that the above limits exist. Observe that the limits are taken in the metric space (RF , D). If f is (fuzzy) differentiable at every point x ∈ [a, b], then we say that f is (fuzzy) differentiable on [a, b] with derivative f ′ (see [32]). It follows from [23] that a function f : [a, b] → RF is fuzzy differentiable at x ∈ [a, b], and f ′ (x) is the fuzzy derivative, if and only if for every ε > 0 there exists a δ > 0 such that, for any interval [x1 , x2 ] ⊂ (x − δ, x + δ) we have f (x2 ) − f (x1 ) ′ , f (x) < ε. D x2 − x1 Similarly, we can define higher order fuzzy derivatives. Also, by CFm [a, b], (m ∈ N), we mean the set of all fuzzy valued functions from [a, b] into RF that are m-times continuously differentiable in the fuzzy sense. However, in this paper, we only discuss the cases m = 0, 1, 2. Using these definitions, Kaleva [24] proved the following result. 2.1. Lemma. [24] Let f : [a, b] ⊂ R → RF be fuzzy differentiable, and x ∈ [a, b], 0 ≤ r ≤ 1. Then, clearly [f (x)]r = [(f (x))(r) , (f (x))(r) + ] ⊂ R. Then (f (x))(r) ± is differentiable and (r) ′ [f (x)]r = ((f (x))(r) )′ , ((f (x))+ )′ , i.e., ′ (r) (r) (f ′ )± = f± for any r ∈ [0, 1]. Also, for higher order fuzzy derivatives, Anastassiou [4] gave the similar result: (r) 2.2. Lemma. [4] Let k ∈ N and f ∈ CFk [a, b]. Then, we have f± ∈ C k [a, b] (for any r ∈ [0, 1]) and h i (i) (r) (i) [f (i) (x)]r = (f (x))(r) , (f (x))+ for i = 0, 1, . . . , k, and, in particular, we get (r) (r) (i) f (i) ± = f± for any r ∈ [0, 1] and i = 0, 1, . . . , k. In this paper, for simplicity, we use the unit interval [0, 1] instead of [a, b]. Now let L : CF [0, 1] → CF [0, 1] be an operator. Then L is said to be fuzzy linear if, for every λ1 , λ2 ≥ 0, f1 , f2 ∈ CF [0, 1], and x ∈ [0, 1], L (λ1 ⊙ f1 ⊕ λ2 ⊙ f2 ; x) = λ1 ⊙ L(f1 ; x) ⊕ λ2 ⊙ L(f2 ; x) holds. Let k be a non-negative integer. As usual, by C k [0, 1] we denote the space of all k-times continuously differentiable functions (in the usual sense) on [0, 1], endowed Statistical Fuzzy Approximation 501 with the sup-norm k · k. Then, throughout the paper, we consider the following function spaces: A := f ∈ C 2 [0, 1] : f ≥ 0 , B := f ∈ C 2 [0, 1] : f ′′ ≥ 0 , C := f ∈ C 2 [0, 1] : f ′′ ≤ 0 , D := f ∈ C 1 [0, 1] : f ≥ 0 , E := f ∈ C 1 [0, 1] : f ′ ≥ 0 , F := {f ∈ C[0, 1] : f ≥ 0} . We also consider the test functions ei (y) = y i (i = 0, 1, 2, . . .) for every y ∈ [0, 1]. Then we have following results. 2.3. Theorem. Let A = (ajn ) be a non-negative regular summability matrix, and {Ln }n∈N a sequence of fuzzy linear operators from CF2 [0, 1] onto itself. Assume that there exists a corresponding sequence {L̃n }n∈N of linear operators from C 2 [0, 1] onto itself for which the following conditions hold: (r) (r) (2.1) Ln (f ; x)± = L̃n f± ; x (for x ∈ [0, 1], r ∈ [0, 1], n ∈ N) and (2.2) If (2.3) n o δA n ∈ N : L̃n (A ∩ B) ⊂ A = 1. stA - lim L̃n (ei ) − ei = 0 for i = 0, 1, 2, n then we have (2.4) stA - lim D∗ (Ln (f ), f ) = 0 for all f ∈ CF2 [0, 1]. n Proof. Let x ∈ [0, 1] be fixed, and f ∈ CF2 [0, 1] and r ∈ [0, 1]. Then we may write that, for every ε > 0 there exists a δ > 0 such that, for every y ∈ [0, 1], D (f (y), f (x)) < ε holds. So, for every y, r ∈ [0, 1] and for any β ≥ 1, we have (r) (2.5) (r) where M± and (r) 2M± β 2M± β (r) (r) ϕx (y) ≤ f± (y) − f± (x) ≤ ε + ϕx (y), δ2 δ2 (r) = f± and ϕx (y) = (y − x)2 . Then, by (2.5), we obtain that −ε − (r) gβ (r) ± (y) := 2M± β (r) (r) ϕx (y) + ε − f± (y) + f± (x) ≥ 0 δ2 (r) 2M± β (r) (r) ϕx (y) + ε + f± (y) − f± (x) ≥ 0 δ2 ± (r) (r) hold for all y ∈ [0, 1]. So, the functions gβ and hβ belong to A. On the other (r) hβ (y) := ± hand, it is clear that, for all y ∈ [0, 1], (r) 4M± β h (r) i′′ (r) gβ (y) = − f± (y) δ2 ± ± 502 O. Duman and (r) 4M± β h (r) i′′ + f± (y). δ2 ± If we choose the number β such that h i′′ 2 (r) f± δ (2.6) β ≥ max 1, , (r) 4M± (r) hβ ′′ (y) = (r) (r) we observe that (2.5) holds for such β’s and also the functions gβ and hβ ± ± ′′ ′′ (r) (r) belong to B because of gβ (y) ≥ 0 and hβ (y) ≥ 0 for all y ∈ [0, 1]. So, we ± ± (r) (r) have gβ , hβ ∈ A ∩ B under the condition (2.6). Let ± ± K := {n ∈ N : L̃n (A ∩ B) ⊂ A}. By (2.2), it is clear that δA {K} = 1, and so (2.7) δA {N \ K} = 0. Then, we may write that (r) (r) L̃n gβ ; x ≥ 0 and L̃n hβ ; x ≥ 0 for any n ∈ K. ± ± Now using the fact that ϕx ∈ A ∩ B, and considering the linearity of L̃n , we obtain, for every n ∈ K, that (r) and 2M± β (r) (r) L̃n (ϕx ; x) + εL̃n (e0 ; x) − L̃n f± ; x + f± (x)L̃n (e0 ; x) ≥ 0 δ2 (r) 2M± β (r) (r) L̃n (ϕx ; x) + εL̃n (e0 ; x) + L̃n f± ; x − f± (x)L̃n (e0 ; x) ≥ 0, δ2 or equivalently (r) − 2M± β (r) L̃n ϕx ; x − εL̃n (e0 ; x) + f± (x) L̃n (e0 ; x) − e0 δ2 (r) (r) ≤ L̃n f± ; x − f± (x) (r) ≤ Then, we have 2M± β (r) L̃n (ϕx ; x) + εL̃n (e0 ; x) + f± (x) L̃n (e0 ; x) − e0 . 2 δ (r) L̃n (f (r) ; x) − f (r) (x) ≤ ε + 2M± β L̃n ϕx ; x ± ± δ2 (r) + ε + f± (x) L̃n (e0 ; x) − e0 for every n ∈ K. The last inequality gives that, for every ε > 0 and n ∈ K, (r) (r) (r) L̃n (f± ) − f± ≤ ε + (ε + M± )L̃n (e0 ) − e0 + (r) (r) 2M± β L̃n (e2 ) − e2 + 4M± β L̃n (e1 ) − e1 2 2 δ δ (r) 2M± β L̃n (e0 ) − e0 . + δ2 Statistical Fuzzy Approximation 503 Hence, we get, for any n ∈ K, that (2.8) 2 X (r) (r) (r) L̃n (f± L̃n (ek ) − ek , ) − f± ≤ ε + C± (ε) k=0 (r) 4M± β . Now it follows from (2.1) that δ2 D∗ Ln (f ), f = sup D Ln (f ; x), f (x) (r) (r) where C± (ε) := max ε + M± + (r) 2M± β , δ2 x∈[0,1] = sup n sup max L̃n f (r) ; x − f (r) (x), x∈[0,1] r∈[0,1] n = sup max L̃n r∈[0,1] o (r) (r) L̃n f+ ; x − f+ (x) o (r) (r) f (r) − f (r) , L̃n f+ − f+ . Combining the above equality with (2.8), we have, for any n ∈ K, (2.9) D∗ (Ln (f ), f ) ≤ ε + C(ε) 2 X L̃n (ek ) − ek , k=0 where C(ε) = sup max r∈[0,1] ′ n (r) (r) C− (ε), C+ (ε) o , Now, for a given ε′ > 0, choose an ε > 0 such that ε < ε , and define the following sets: F := n ∈ N : D∗ (Ln (f ), f ) ≥ ε′ , n ε′ − ε o Fk := n ∈ N : L̃n (ek ) − ek ≥ , k = 0, 1, 2. 3C(ε) Then, it follows from (2.9) that F ∩K ⊂ 2 [ (Fk ∩ K), k=0 which yields, for every j ∈ N, that X 2 2 X X X X ajn . (2.10) ajn ≤ ajn ≤ n∈F ∩K k=0 n∈Fk ∩K k=0 n∈Fk Now, taking the limit as j → ∞ on both-sides of (2.10), and using (2.3), we immediately see that X (2.11) lim ajn = 0. j n∈F ∩K Furthermore, since X X ajn = ajn + n∈F ∩K n∈F ≤ X n∈F ∩K ajn + X ajn n∈F ∩(N\K) X ajn n∈(N\K) holds for every j ∈ N, taking again the limit as j → ∞ in the last inequality, and using (2.7), (2.11), we obtain X lim ajn = 0, j n∈F 504 O. Duman which means that stA - lim D∗ (Ln (f ), f ) = 0. n The theorem is proved. 2.4. Theorem. Let A = (ajn ) be a non-negative regular summability matrix, and {Ln }n∈N a sequence of fuzzy linear operators from CF2 [0, 1] onto itself. Assume that there exists a corresponding sequence {L̃n }n∈N of linear operators from C 2 [0, 1] onto itself for which the following conditions hold: h i′′ h i′′ (r) (2.12) {Ln (f )}(r) (x) = L̃n f± (x) for x ∈ [0, 1], r ∈ [0, 1], n ∈ N ± and (2.13) If (2.14) δA n ∈ N : L̃n (A ∩ C) ⊂ C = 1. stA - lim [L̃n (ei )]′′ − e′′i = 0 for i = 0, 1, 2, 3, 4, n then, we have (2.15) stA - lim D∗ [Ln (f )]′′ , f ′′ = 0 for all f ∈ CF2 [0, 1]. n Proof. We should remark that the derivatives in (2.12) and (2.14) are in the usual sense while the derivatives in (2.15) are in the fuzzy sense. Now let f ∈ CF2 [0, 1], x ∈ [0, 1] and r ∈ [0, 1] be fixed. By Lemmas 2.1 and 2.2, and as in the proof of Theorem 2.3, we can write that, for every ε > 0, there exists a δ > 0 such that (r) (r) (2.16) −ε + (r) ′′ (r) ′′ 2U± β ′′ 2U± β ′′ σx (y) ≤ f± (y) − f± (x) ≤ ε − σx (y) 2 δ δ2 (y − x)4 (r) holds for all y ∈ [0, 1] and for any β ≥ 1, where σx (y) = − + 1 and U± = 12 (r) ′′ f . Then, define the following functions on [0, 1]: ± (r) ′′ (r) f (x) 2 2U± β ε 2 (r) (r) uβ (y) := σx (y) + f± (y) − y − ± y , δ2 2 2 ± and (r) vβ (r) ′′ (r) 2U± β f± (x) 2 ε 2 (r) σx (y) − f± (y) − y + y . (y) := δ2 2 2 ± It follows from (2.16) that ′′ ′′ (r) (r) uβ (y) ≤ 0 and vβ (y) ≤ 0 for all y ∈ [0, 1], ± ± which implies that the functions 11 for all y ∈ [0, 1]. Then 12 (r) ±f± (y) − ε2 y 2 ± (r) h (r) uβ i (r) ′′ f± 2U± σx (y) 2 (x) ± (r) and vβ belong to C. Observe that σx (y) ≥ y2 δ2 ± ≤ (r) (r) M± + U± + ε δ 2 (r) U± Statistical Fuzzy Approximation (r) 505 h i (r) (r) ′′ = f± and U±(r) = f± as stated before. holds for all y ∈ [0, 1], where M± Now, if we choose β such that ( ) (r) (r) (M± + U± + ε)δ 2 (2.17) β ≥ max 1, , (r) U± then inequality (2.16) holds for such β’s and (r) (r) uβ (y) ≥ 0 and vβ (y) ≥ 0 for all y ∈ [0, 1]. ± ± (r) (r) (r) Hence, we also get uβ , vβ ∈ A, which gives that the functions uβ and ± ± ± (r) vβ belong to A ∩ C under the condition (2.17). Now let ± K := {n ∈ N : L̃n (A ∩ C) ⊂ C}. Then, by (2.13), we have (2.18) δA (N \ K) = 0. Also we get, for every n ∈ K, h i′′ h i′′ (r) (r) L̃n vβ ± ≤ 0 and L̃n vβ ± ≤ 0. Then, we obtain, for every n ∈ K, that (r) ′′ (r) i′′ h i′′ ε f (x) 2U± β h (r) ′′ L̃n (σx ) + L̃n (f± ) − [L̃n (e2 )] − ± [L̃n (e2 )]′′ ≤ 0 δ2 2 2 and (r) ′′ (r) h i′′ ε f± (x) 2U± β (r) ′′ ′′ [L̃n (σx )] − L̃n (f± ) − [L̃n (e2 )] + [L̃n (e2 )]′′ ≤ 0. δ2 2 2 These inequalities yield that (r) ′′ (r) (r) ′′ 2U± β f± (x) ε ′′ ′′ [ L̃ (σ )] (x) − [ L̃ (e )] (x) + [L̃n (e2 )]′′ (x) − f± (x) n x n 2 2 δ 2 2 h i′′ h i′′ (r) (r) ≤ L̃n (f± ) (x) − f± (x) (r) ≤− 2U± β ε [L̃n (σx )]′′ (x) + [L̃n (e2 )]′′ (x) δ2 2 (r) ′′ (r) ′′ f (x) + ± [L̃n (e2 )]′′ (x) − f± (x). 2 Observe now that [L̃n (σx )]′′ ≤ 0 on [0, 1] for every n ∈ K because of σx ∈ A ∩ C. Using this, the last inequality gives, for every n ∈ K, that (r) (r) ′′ ′′ ′′ 2U β ε (r) ′′ (x) ≤ − ±2 L̃n (σx ) (x) + L̃n (e2 ) (x) L̃n (f± ) (x) − f± δ 2 (r) ′′ f± ′′ (x) + L̃n (e2 ) (x) − 2 , 2 and hence (r) ′′ ′′ ε + f± (x) (r) ′′ ′′ ′′ L̃n (f± ) (x) − f (x) ≤ ε + L̃n (e2 ) (x) − e2 (x) 2 (2.19) (r) ′′ 2U± β + L̃n (−σx ) (x). δ2 506 O. Duman Now we compute the quantity [Ln (−σx )]′′ in inequality (2.19). Observe that h (y − x)4 i′′ ′′ L̃n (−σx ) (x) = L̃n −1 (x) 12 ′′ ′′ ′′ x x2 1 L̃n (e4 ) (x) − L̃n (e3 ) (x) + L̃n (e2 ) (x) = 12 3 2 x4 ′′ ′′ x3 − L̃n (e1 ) (x) + − 1 L̃n (e0 ) (x) 3 12 o x n o ′′ ′′ 1 n = L̃n (e4 ) (x) − e′′4 (x) − L̃n (e3 ) (x) − e′′3 (x) 12 3 o x3 n o ′′ ′′ x2 n + L̃n (e2 ) (x) − e′′2 (x) − L̃n (e1 ) (x) − e′′1 (x) 2 3 x4 n o ′′ + −1 L̃n (e0 ) (x) − e′′0 (x) . 12 Combining this with (2.19), for every ε > 0 and n ∈ K, we have (r) ′′ (r) (r) ′′ ε + f± (x) U β± x2 (r) ′′ + ± 2 (x) ≤ ε + L̃n (f± ) (x) − f± 2 δ ′′ × Ln (e2 ) (x) − e′′2 (x) (r) ′′ U± β ′′ L̃ (e ) (x) − e (x) n 4 4 6δ 2 (r) ′′ 2U± βx ′′ + L̃n (e3 ) (x) − e3 (x) 2 3δ (r) ′′ 2U± βx3 ′′ L̃ (e ) (x) − e (x) + n 1 1 3δ 2 (r) ′′ 2U± β x4 ′′ + 1− L̃n (e0 ) (x) − e0 (x) . 2 3δ 12 Therefore, we obtain, for every ε > 0 and n ∈ K, that 4 X (r) ′′ ′′ (r) ′′ (r) ′′ (2.20) L̃n (f± ) − f± ≤ ε + E± (ε) L̃n (ek ) − ek , + k=0 (r) where E± (ε) := ε+ (r) U± + (r) U± β δ2 (r) ′′ (r) and U± = f± , as stated before. On the other 2 hand, by (2.12), since ′′ ′′ D∗ Ln (f ) , f ′′ = sup D Ln (f ) (x), f ′′ (x) x∈[0,1] n ′′ (r) ′′ (x) − f (r) (x) , L̃n f x∈[0,1] r∈[0,1] (r) ′′ o (r) ′′ (x) − f+ (x) L̃n f+ n ′′ (r) ′′ = sup max L̃n f (r) − f , = sup sup max r∈[0,1] o (r) ′′ (r) ′′ − f+ L̃n f+ , we may write from (2.20) that (2.21) 4 X ′′ ′′ ′′ D∗ Ln (f ) , f ′′ ≤ ε + E(ε) L̃n (ek ) − ek k=0 Statistical Fuzzy Approximation 507 (r) (r) holds for all n ∈ K, where E(ε) = sup max E− (ε), E+ (ε) . Now, for a given ε′ > 0, r∈[0,1] choose an ε such that 0 < ε < ε′ , and consider the following sets: n o ′′ G := n ∈ N : D∗ Ln (f ) , f ′′ ≥ ε′ , ′′ ε′ − ε Gk := n ∈ N : L̃n (ek ) − e′′k ≥ , k = 0, 1, 2, 3, 4. 5E(ε) In this case, by (2.21), G∩K ⊂ 4 [ (Gk ∩ K), k=0 which yields, for every j ∈ N, that X 4 4 X X X X (2.22) ajn ≤ ajn ≤ ajn . n∈G∩K k=0 n∈Gk ∩K k=0 n∈Gk Letting j → ∞ on both sides of (2.22), and using (2.14), we immediately see that X (2.23) lim ajn = 0. j n∈G∩K Furthermore, if we use the inequality X X X ajn = ajn + n∈G n∈G∩K ≤ X n∈G∩K ajn n∈G∩(N\K) ajn + X ajn n∈(N\K) and take the limit as j → ∞, then it follows from (2.18) and (2.23) that X lim ajn = 0. j n∈G Thus, we get ′′ stA - lim D∗ Ln (f ) , f ′′ = 0. n The theorem is proved. 2.5. Theorem. Let A = (ajn ) be a non-negative regular summability matrix and {Ln }n∈N a sequence of fuzzy linear operators from CF1 [0, 1] onto itself. Assume that there exists a corresponding sequence {L̃n }n∈N of linear operators from C 1 [0, 1] onto itself for which the following conditions hold: (r) ′ (r) ′ (2.24) Ln (f ) ± (x) = L̃n f± (x) (for x ∈ [0, 1], r ∈ [0, 1], n ∈ N) and (2.25) If (2.26) δA n ∈ N : L̃n (D ∩ E) ⊂ E = 1. ′ stA - lim L̃n (ei ) − e′i = 0 for i = 0, 1, 2, 3, n then we have (2.27) ′ stA - lim D∗ Ln (f ) , f ′ = 0 for all f ∈ C 1 [0, 1]. n 508 O. Duman Proof. Let f ∈ CF1 [0, 1], x ∈ [0, 1] and r ∈ [0, 1] be fixed. Then, for every ε > 0, there exists a positive number δ such that (r) (2.28) (r) (r) ′ (r) ′ 2V± β ′ 2V β wx (y) ≤ f± (y) − f± (x) ≤ ε + ±2 wx′ (y) δ2 δ −ε − holds for all y ∈ [0, 1] and for any β ≥ 1, where wx (y) := (r) ′ f± , Now considering the functions defined by (y − x)3 (r) + 1 and V± := 3 (r) (r) θβ and (y) := ± (r) ′ 2V± β (r) wx (y) − f± (y) + εy + y f± (x) 2 δ (r) (r) ′ 2V± β (r) wx (y) + f± (y) + εy − y f± (x), 2 δ (r) (r) we can easily check that θβ ± and λβ ± belong to E for any β ≥ 1. Also, observe 2 that wx (y) ≥ for all y ∈ [0, 1]. Then 3 (r) ′ (r) (r) (r) ± f± (y) − εy ± f± (x)y δ 2 (M± + V± + ε)δ 2 ≤ (r) (r) 2V± wx (y) V± (r) λβ (y) := ± (r) (r) holds for all y ∈ [0, 1], where M± = kf± k, as stated before. Now, if we choose a number β such that ( ) (r) (r) (M± + V± + ε)δ 2 (2.29) β ≥ max 1, , (r) V± then inequality (2.28) holds for such β’s and (r) (r) θβ ± (y) ≥ 0 and λβ ± (y) ≥ 0 for all y ∈ [0, 1], (r) (r) (r) (r) which yields that θβ ± , λβ ± ∈ D. Thus, we get θβ ± , λβ ± ∈ D ∩ E for any β satisfying (2.29). Let K := {n ∈ N : L̃n (D ∩ E) ⊂ E}. Then, by (2.25), we have (2.30) δA {N \ K} = 0. Also we get, for every n ∈ K, (r) ′ (r) ′ L̃n θβ ± ≥ 0 and L̃n λβ ± ≥ 0. Hence we obtain, for every n ∈ K, that (r) and ′ (r) ′ ′ 2V± β (r) ′ [L̃n (wx )]′ − L̃n f± + ε L̃n (e1 ) + f± (x) L̃n (e1 ) ≥ 0 δ2 (r) ′ (r) ′ ′ 2V± β (r) ′ [L̃n (wx )]′ + L̃n (f± ) + ε L̃n (e1 ) − f± (x) L̃n (e1 ) ≥ 0. δ2 Statistical Fuzzy Approximation 509 Then, we may write that − (r) ′ (r) ′ ′ (r) ′ 2V± β [L̃n (wx )]′ (x) − ε L̃n (e1 ) (x) + f± (x) L̃n (e1 ) (x) − f± (x) 2 δ (r) ′ (r) ′ ≤ L̃n f± (x) − f± (x) ≤ (r) ′ ′ 2V± β L̃n (wx ) (x) + ε L̃n (e1 ) (x) δ2 (r) ′ ′ (r) ′ + f± (x) L̃n (e1 ) (x) − f± (x), and hence (r) ′ ′ (r) ′ (r) ′ ′ L̃n (f± ) (x) − f± (x) ≤ ε + ε + f± (x) L̃n (e1 ) (x) − e1 (x) (2.31) (r) ′ 2V β + ±2 L̃n (wx ) (x) δ holds for every n ∈ K because of the fact that the function wx belongs to D ∩ E. Since i′ h (y − x)3 ′ + 1 (x) L̃n (wx ) (x) = L̃n 3 ′ ′ ′ 1 = L̃n (e3 ) (x) − x L̃n (e2 ) (x) + x2 L̃n (e1 ) (x) 3 ′ x3 + 1− L̃n (e0 ) (x) 3 ′ ′ 1 = L̃n (e3 ) (x) − e′3 (x) − x L̃n (e2 ) (x) − e′2 (x) 3 ′ + x2 L̃n (e1 ) (x) − e′1 (x) ′ x3 + 1− L̃n (e0 ) (x) − e′0 (x) , 3 it follows from (2.31) that (r) ′ (r) ′ L̃n (f± ) (x) − f± (x) ≤ ε + ! (r) 2 (r) ′ 2V± β± x ε + f± (x) + δ2 ′ × L̃n (e1 ) (x) − e′1 (x) (r) ′ 2V± β ′ L̃ (e ) (x) − e (x) n 3 3 3δ 2 (r) ′ 2V βx ′ + ±2 L̃n (e2 ) (x) − e2 (x) δ (r) ′ 2V β x 3 ′ + ±2 1− L̃n (e0 ) (x) − e0 (x) . δ 3 + Thus, we deduce from the last inequality that (2.32) 3 X (r) ′ ′ (r) ′ (r) ′ L̃n (f± ) − f± ≤ ε + F± (ε) L̃n (ek ) − ek k=0 (r) (r) (r) holds for any n ∈ K, where F± (ε) := ε + V± + since 2V± β± . On the other hand, by (2.24), δ2 510 O. Duman ′ ′ D∗ Ln (f ) , f ′ = sup D Ln (f ) (x), f ′ (x) x∈[0,1] = sup sup max x∈[0,1] r∈[0,1] = sup max r∈[0,1] n ′ (r) ′ (x) − f (r) (x) , L̃n f n L̃n (r) ′ o (r) ′ (x) − f+ (x) L̃n f+ ′ ′ f (r) − f (r) , o (r) ′ (r) ′ − f+ , L̃n f+ we may write from (2.32) that (2.33) 3 X ′ ′ ′ D∗ Ln (f ) , f ′ ≤ ε + F (ε) L̃n (ek ) − ek k=0 (r) (r) holds for all n ∈ K, where F (ε) = sup max F− (ε), F+ (ε) . Now, for a given ε′ > 0, r∈[0,1] choose an ε such that 0 < ε < ε′ , and consider the following sets: ′ J := n ∈ N : D∗ Ln (f ) , f ′ ≥ ε′ , ′ ε′ − ε , k = 0, 1, 2, 3. Jk := n ∈ N : L̃n (ek ) − e′k ≥ 4G(ε) In this case, by (2.33), J ∩K ⊂ 3 [ (Jk ∩ K), k=0 which yields, for every j ∈ N, that (2.34) X ajn ≤ n∈J ∩K 3 X k=0 X ajn n∈Jk ∩K ! ≤ 3 X k=0 X n∈Jk ajn ! Letting j → ∞ on both sides of (2.34), and also using (2.26), we immediately see that X (2.35) lim ajn = 0. j n∈J ∩K Now, using the fact that X X ajn = ajn + n∈J ∩K n∈J ≤ X n∈J ∩K ajn + X ajn n∈J ∩(N\K) X ajn , n∈(N\K) and taking the limit as j → ∞, then it follows from (2.30) and (2.35) that X lim ajn = 0. j n∈J Thus, we get ′ stA - lim D∗ Ln (f ) , f ′ = 0, n whence the result. Statistical Fuzzy Approximation 511 2.6. Theorem. Let A = (ajn ) be a non-negative regular summability matrix and {Ln }n∈N a sequence of fuzzy linear operators from CF [0, 1] onto itself. Assume that there exists a corresponding sequence {L̃n }n∈N of linear operators from C[0, 1] onto itself for which the following conditions hold: (r) (r) (2.36) Ln (f ; x)± = L̃n f± ; x for x ∈ [0, 1], r ∈ [0, 1], n ∈ N and If δA n ∈ N : L̃n F ⊂ F = 1. (2.38) stA - lim kL̃n (ei ) − ei k = 0 for i = 0, 1, 2, (2.37) n then we have (2.39) ′ stA - lim D∗ Ln (f ) , f ′ for all f ∈ C[0, 1]. n Proof. If L̃n (F) ⊂ F holds for every n ∈ N instead of (2.37), i.e., the corresponding sequence {L̃n }n∈N consists of positive linear operators on C[0, 1], then the proof follows immediately from [5, Theorem 2.1]. However, our previous proofs show that even when the weaker condition (2.37) holds, the property (2.38) implies (2.39). 3. Concluding Remarks If we replace the non-negative regular matrix A by the identity matrix, then our Theorems 2.3-2.6 reduce to the following results which are also new in the literature, except for the last one. 3.1. Corollary. Let {Ln }n∈N be a sequence of fuzzy linear operators from CF2 [0, 1] onto itself. Assume that there exists a corresponding sequence {L̃n }n∈N of linear operators from C 2 [0, 1] onto itself for which the following conditions hold: (r) (r) Ln (f ; x)± = L̃n f± ; x for x ∈ [0, 1], r ∈ [0, 1], n ∈ N and L̃n (A ∩ B) ⊂ A for n ∈ N). If the sequence {L̃n (ei )}, (i = 0, 1, 2), is uniformly convergent to ei on [0, 1] then, for all f ∈ CF2 [0, 1], lim D∗ Ln (f ), f = 0. n 3.2. Corollary. Let {Ln }n∈N be a sequence of fuzzy linear operators from CF2 [0, 1] onto itself. Assume that there exists a corresponding sequence {L̃n }n∈N of linear operators from C 2 [0, 1] onto itself for which the following conditions hold: (r) ′′ (r) ′′ {Ln (f )}± (x) = L̃n f± (x) for x ∈ [0, 1], r ∈ [0, 1], n ∈ N and L̃n (A ∩ C) ⊂ C for n ∈ N . ′′ If the sequence L̃n (ei ) , (i = 0, 1, 2, 3, 4), is uniformly convergent to e′′i on [0, 1] then, 2 for all f ∈ CF [0, 1], ′′ lim D∗ Ln (f ) , f ′′ = 0. n 512 O. Duman 3.3. Corollary. Let {Ln }n∈N be a sequence of fuzzy linear operators from CF1 [0, 1] onto itself. Assume that there exists a corresponding sequence {L̃n }n∈N of linear operators from C 1 [0, 1] onto itself for which the following conditions hold: (r) ′ (r) ′ Ln (f ) ± (x) = L̃n f± (x) for x ∈ [0, 1], r ∈ [0, 1], n ∈ N and L̃n (D ∩ E) ⊂ E for n ∈ N . ′ If the sequence L̃n (ei ) , (i = 0, 1, 2, 3), is uniformly convergent to e′i on [0, 1] then, for all f ∈ CF1 [0, 1], ′ lim D∗ Ln (f ) , f ′ = 0. n 3.4. Corollary. [3] Let {Ln }n∈N be a sequence of fuzzy linear operators from CF [0, 1] onto itself. Assume that there exists a corresponding sequence {L̃n }n∈N of positive linear operators from C[0, 1] onto itself for which the following condition holds: (r) (r) Ln (f ; x)± = L̃n f± ; x for x ∈ [0, 1], r ∈ [0, 1], n ∈ N). If the sequence {L̃n (ei )} (i = 0, 1, 2) is uniformly convergent to ei on [0, 1] then, for all f ∈ CF [0, 1], lim D∗ Ln (f ), f = 0. n Finally, we give an application satisfying all the conditions of our Theorem 2.3. 3.5. Application. We first take A = C1 , the Cesáro matrix. Then, we know that C1 statistical convergence reduces to the concept of statistical convergence; in this case, we use the notation st- lim instead of stC1 - lim. Also, we have δC1 ≡ δ. Assume now that {Ln } is a sequence of linear operators on CF2 [0, 1] whose corresponding sequence {L̃n } of linear operators on C 2 [0, 1] is given by ( −x2 , if n = m2 m ∈ N , (r) (3.1) L̃n (f± ; x) = (r) Bn (f± ; x), if n 6= m2 , where x ∈ [0, 1], r ∈ [0, 1], n ∈ N, and {Bn } denotes the sequence of classical Bernstein polynomials on C[0, 1]. Then, observe that δ n ∈ N : L̃n A ∩ B ⊂ A = δ n 6= m2 : m ∈ N = 1. Also we have, for each i = 0, 1, 2, st- lim L̃n (ei ) − ei = 0. n Then, it follows from Theorem 2.3 that, for all f ∈ CF2 [0, 1], st- lim D∗ Ln (f ), f = 0. n However, for the function e0 = 1, since ( −x2 if n = m2 m ∈ N L̃n (e0 ; x) := 1 otherwise, we get, for all x ∈ [0, 1], that the sequence {L̃n (e0 ; x)} is non-convergent in the ordinary sense. Furthermore, we see that infinitely many terms of L̃n do not satisfy the inclusion L̃n A ∩ B ⊂ A. Hence, this example clearly shows that our statistical fuzzy approximation theorems obtained in this paper have a wider range of applications than the classical fuzzy approximation results. Statistical Fuzzy Approximation 513 3.6. Remark. In this paper, we focus on fuzzy approximation only for the one dimensional case. However, by simple manipulations, all our results are also valid for multi dimensional cases. However, we omit their details. Acknowledgement. The author would like to thank the referee for carefully reading the manuscript. References [1] Altomare, F. and Campiti, M. Korovkin Type Approximation Theory and Its Application (Walter de Gruyter Publ., Berlin, 1994). [2] Altomare, F. and Rasa, I. Approximation by positive operators in spaces C p ([a, b]), Anal. Numér. Théor. Approx. 18, 1–11, 1989. [3] Anastassiou, G. A. On basic fuzzy Korovkin theory, Studia Univ. Babeş-Bolyai Math. 50, 3–10, 2005. [4] Anastassiou, G. A. Higher order fuzzy Korovkin theory via inequalities, Commun. Appl. Anal. 10, 359–392, 2006. [5] Anastassiou, G. A. and Duman, O. Statistical fuzzy approximation by fuzzy positive linear operators, Comput. Math. Appl. 55, 573–580, 2008. [6] Anastassiou, G. A. and Duman, O. A Baskakov type generalization of statistical Korovkin theory, J. Math. Anal. Appl. 340, 476–486, 2008. [7] Boos, J. Classical and Modern Methods in Summability (Oxford University Press, UK, 2000). [8] Brosowksi, B. A Korovkin-type theorem for differentiable functions, Approximation Theory, III (Proc. Conf., Univ. Texas, Austin, Tex., 1980) (Academic Press, New York-London, 1980), 255–260. [9] Delgado, F. J. M., Gonzáles, V. R. and Morales, D. C. Qualitative Korovkin-type results on conservative approximation, J. Approx. Theory 94, 144–159, 1998. [10] Dirik, F., Duman, O. and Demirci, K. Statistical approximation to Bögel-type continuous and periodic functions, Cent. Eur. J. Math. 7, 539–549, 2009. [11] Doğru, O. and Duman, O. Statistical approximation of Meyer-König and Zeller operators based on q-integers, Publ. Math. Debrecen 68, 199–214, 2006. [12] Duman, O. Statistical approximation for periodic functions, Demonstratio Math. 36, 873– 878, 2003. [13] Duman, O. and Anastassiou, G. A. On statistical fuzzy trigonometric Korovkin theory, J. Comput. Anal. Appl. 10, 333–344, 2008. [14] Duman, O., Khan, M. K. and Orhan, C. A-Statistical convergence of approximating operators, Math. Inequal. Appl. 6, 689–699, 2003. [15] Duman, O., Erkuş, E. and Gupta, V. Statistical rates on the multivariate approximation theory, Math. Comput. Modelling 44, 763–770, 2006. [16] Erkuş, E. and Duman, O. A Korovkin type approximation theorem in statistical sense, Studia Sci. Math. Hungar. 43, 285–294, 2006. [17] Erkuş, E. and Duman, O. A-statistical extension of the Korovkin type approximation theorem, Proc. Indian Acad. Sci. (Math. Sci.) 115, 499–508, 2005. [18] Fast, H. Sur la convergence statistique, Colloq. Math. 2, 241–244, 1951. [19] Freedman, A. R. and Sember, J. J. Densities and summability, Pacific J. Math. 95, 293–305, 1981. [20] Fridy, J. A. On statistical convergence, Analysis 5, 301–313, 1985. [21] Gal, S. G. Approximation theory in fuzzy setting, Handbook of Analytic-Computational Methods in Applied Mathematics (Chapman & Hall/CRC, Boca Raton, FL, 2000), 617– 666. [22] Goetschel, R. J. and Voxman, W. Elementary fuzzy calculus, Fuzzy Sets and Systems 18, 31–43, 1986. [23] Gong, Z., Wu, C. and Li, B. On the problem of characterizing derivatives for the fuzzy-valued functions, Fuzzy Sets and Systems 127, 315–322, 2002. 514 O. Duman [24] Kaleva, O. Fuzzy differential equations, Fuzzy Sets and Systems 24, 301–317, 1987. [25] Karakuş, S., Demirci, K. and Duman, O. Equi-statistical convergence of positive linear operators, J. Math. Anal. Appl. 339, 1065–1072, 2008. [26] Knoop, H. B. and Pottinger, P. Ein satz von Korovkin-typ für C k -raüme, Math. Z. 148, 23–32, 1976. [27] Kolk, E. Matrix summability of statistically convergent sequences, Analysis 13, 77–83, 1993. [28] Korovkin, P. P. Linear Operators and Approximation Theory (Hindustan Publ. Corp., Delhi, 1960). [29] Matloka, M. Sequences of fuzzy numbers, BUSEFAL 28, 28–37, 1986. [30] Miller, H. I. A measure theoretical subsequence characterization of statistical convergence, Trans. Amer. Math. Soc. 347, 1811–1819, 1995. [31] Nuray, F. and Savaş, E. Statistical convergence of sequences of fuzzy numbers, Math. Slovaca 45, 269–273, 1995. [32] Wu, C. X. and Ma, M. Embedding problem of fuzzy number space I, Fuzzy Sets and Systems 44, 33–38, 1991. [33] Wu, C. and Gong, Z. On Henstock integral of fuzzy-number-valued functions I, Fuzzy Sets and Systems 120, 523–532, 2001.