Approximation capability of the convolution

methods for fuzzy numbers

arXiv:1408.1882v1 [math.GM] 6 Aug 2014

Huan Huang
Department of Mathematics, Jimei University, Xiamen 361021, China

Abstract

This paper shows that how to approximate general fuzzy number by using convo-
lution method.

Key words: Fuzzy numbers; Convolution; Supremum metric; Differentiable

1 Instructions

The approximation of fuzzy numbers attract many peoples attentions [1,2,5,6].

Since differentiable fuzzy numbers play an important role in the implementa-
tion of fuzzy intelligent systems and their applications (see [3,4]). Chalco-Cano
et al. [5,6] introduced a method based on the convolution to approximate non-
differentiable fuzzy numbers by differentiable fuzzy numbers.

A significant advantage of this method is that it can generate differentiable

fuzzy number such that the distance between which and the original fuzzy
number is less than or equal to arbitrary predetermine positive number.

However, in the previous work, not all the fuzzy numbers can be approximated
by this method. So it is natural for us to consider the question that can we
use the convolution method to approximate general fuzzy numbers?

In this paper, we want to answer this question.

Email address: [email protected] .

Preprint submitted to Elsevier

2 Preliminaries

2.1 Fuzzy numbers

Let N be the set of all natural numbers, R be the set of all real numbers. For
details, we refer the reader to references [7,8].

A fuzzy subsets u on R can be seen as a mapping from R to [0,1]. For α ∈ (0, 1],
let [u]α denote the α-cut of u; i.e., [u]α ≡ {x ∈ R : u(x) ≥ α} and [u]0 denotes
{x ∈ R : u(x) > 0}. We call u a fuzzy number if u has the following properties:
(i) [u]1 6= ∅; and
(ii) [u]α = [u− (α), u+ (α)] are compact intervals of R for all α ∈ [0, 1].
The set of all fuzzy numbers is denoted by F (R). In [5], a fuzzy number is
also called a fuzzy number.

The following is a widely used representation theorem of fuzzy numbers.

Proposition 2.1 (Goetschel and Voxman [10]) Given u ∈ F (R), then

(i) u− (·) is a left-continuous nondecreasing bounded function on (0, 1];

(ii) u+ (·) is a left-continuous nonincreasing bounded function on (0, 1];
(iii) u− (·) and u+ (·) are right continuous at α = 0;
(iv) u− (1) ≤ u+ (1).

Moreover, if the pair of functions a(λ) and b(λ) satisfy conditions (i) through
(iv), then there exists a unique u ∈ F (R) such that [u]α = [a(λ), b(λ)] for each
α ∈ (0, 1].

The algebraic operations on F (R) are defined as follows: given u, v ∈ F (R),

α ∈ [0, 1],

[u + v]α = [u]α + [v]α = [u− (α) + v − (α), u+ (α) + v + (α)],

[u − v]α = [u]α − [v]α = [u− (α) − v + (α), u+ (α) − v − (α)],
[u · v]α = [u]α · [v]α = [min{xy : x ∈ [u]α , y ∈ [v]α }, max{xy : x ∈ [u]α , y ∈ [v]α }].

The supremum metric on F (R) is defined by

d∞ (u, v) = sup max{|u− (α) − v − (α)|, |u+ (α) − v + (α)|},

α∈[0,1]

where u, v ∈ F (R).

2
2.2 Convolution of fuzzy numbers

In this paper, we want to discuss the properties of sup-min convolution u∇v

of fuzzy numbers u and v, which is defined by

(u∇v)(x) = sup{u(y) ∧ v(x − y)}.

y∈R

Remark 2.2 In fact u∇v = u + v for all u, v ∈ F (R). For details, see [7,8].

In the following, we list some symbols which are used to denote subsets of
F (R).

• FT (R) is denoted the family of all fuzzy numbers u such that u is strictly
increasing on [u− (0), u− (1)], strictly decreasing on [u+ (1), u+ (0)], and dif-
ferentiable on (u− (0), u− (1)) ∪ (u+ (1), u+ (0)).
• FN (R) is denoted the family of all fuzzy numbers u such that u is differen-
tiable on (u− (0), u−(1)) ∪ (u+ (1), u+ (0)).
• FC (R) is denoted the family of all continuous fuzzy numbers, i.e., the
family of all fuzzy numbers u such that u : R → [0, 1] is continuous on
(u− (0), u+ (0)).
• FD (R) is denoted the family of all differentiable fuzzy numbers, i.e., the
family of all fuzzy numbers u such that u : R → [0, 1] is differentiable on
(u− (0), u+ (0)).

Given a fuzzy number u in FN (R), u need not be strictly increasing on

(u− (0), u−(1)) and strictly decreasing on (u+ (1), u+ (0)). So

FT (R) ( FN (R).

Observe that p is differentiable on (p− (1), p+ (1)) for all p ∈ F (R). Thus we
know that, for each u ∈ FN (R), the possible non-differentiable points are u− (1)
and u+ (1).

It is easy to check that

FD (R) ( FC (R) ∩ FN (R).

Chalco-Cano et al. [5] constructed fuzzy numbers wp , p > 0, which are defined
by   2
1− x

p
, if x ∈ [−p, p],
wp (x) =  (1)
 0, if x ∈
/ [−p, p].
Obviously, wp ∈ FD (R) for all p > 0. They presented the following result.

3
Proposition 2.3 [5] If u ∈ FT (R), then u∇wp ∈ FD (R).

Notice that d∞ (u, u∇wp) → 0 as p → 0. Thus Proposition 2.3 indicates that

every fuzzy number in FT (R) can be approximated by fuzzy numbers se-
quences in FD (R).

We can see that the fuzzy numbers wp , p > 0, work as smoothers, which
transfer each fuzzy number u to a differentiable (smooth) fuzzy number u∇wp .
This sequence of smooth fuzzy numbers construct a approximation sequence
of the original fuzzy number u, i.e., u∇wp → u as p → 0.

Chalco-Cano et al. [6] further putted forward a method to define smoothers.

Suppose that p > 0 is a real number and that f : [0, 1] → [0, 1] is a continuous
and strictly decreasing function with f (0) = 1, f (1) = 0. A class of fuzzy
numbers Zpf is defined by

f −1 (kxk/p), kxk ≤ p,
Zpf (x) =
0, kxk > p.
√
It is easy to show that Zpf = wp when f = 1 − t. They gave the following
result.

Proposition 2.4 [6] If f is differentiable and limα→1− f ′ (α) = −∞, then

u∇Zpf ∈ FD (R) for each u ∈ FT (R).

Notice that d∞ (u∇Zpf , u) → 0 as p → 0. This means that given f satisfies

the above conditions, we obtain a smooth approximation {u∇Zpf : p > 0} of
the fuzzy number u. Different f corresponds to different sequence of smooth
approximation.

3 Approximation

In the previous work, only fuzzy numbers in FT (R) can be approximated. This
type of fuzzy numbers have at most two possible non-differentiable points:
the endpoints of the 1-cut. Whereas, an arbitrarily given fuzzy number may
have other non-differentiable points or non-continuous points. So it is natural
for us to consider the question that can we use the convolution method to
approximate general fuzzy numbers?

In this section, to answer this question, it discusses how to choose smoothers to

smooth an arbitrarily given fuzzy number. The following theorems are devoted
to this problem.

4
Theorem 3.1 Suppose that u ∈ FN (R) ∩ FC (R) and that w ∈ FD (R), then
u∇w ∈ FD (R) when w satisfies the following conditions.

(i) w(w − (0)) = u(u− (0)) and w(w + (0)) = u(u+ (0)).
(ii) If u− (1) is the inner point of [u]0 , then w−
′
(w − (1)) = 0.
(iii) If u (1) is the inner point of [u]0 , then w+ (w + (1)) = 0.
+ ′

Theorem 3.2 Suppose that u ∈ FC (R) and that w ∈ FD (R), then u∇w ∈
FD (R) when w satisfies the conditions (i), (ii), and (iii) in Theorem 3.1 and
the following condition (iv).

(iv) Given x ∈ (u− (0), u+ (0)) with u(x) = α < 1, if x is a non-differentiable

point of u, then
(iv) (1) if x < u− (1), then w ′(w − (α)) = 0;
(iv) (2) if x > u+ (1), then w ′(w + (α)) = 0.

Theorem 3.3 Suppose that u ∈ F (R) and that w ∈ FD (R), then u∇w ∈
FD (R) when w satisfies the conditions (i), (ii), and (iii) in Theorem 3.1, the
condition (iv) in Theorem 3.2, and the following condition (v).

(v) Given x ∈ (u− (0), u+ (0)) with u(x) = α < 1, if x is a non-continuous

point of u, then
(v) (1) if x < u− (1), then w ′(w − (β)) = 0, where β = limy→x− u(y);
(v) (2) if x > u+ (1), then w ′(w + (γ)) = 0, where γ = limz→x+ u(z).

By above theorems, given an arbitrary fuzzy number, if the number of non-

continuous points of u in [u]0 is finite, then we can construct a sequence of
smoothers vn , n ∈ N, such that u∇vn is a differentiable fuzzy number for all
n, and that {u∇vn } converges to u in the supremum metric d∞ .

References

[1] P. Grzegorzewski, K. Pasternak-Winiarska. Natural trapezoidal approximations

of fuzzy numbers. Fuzzy Sets and Systems, 250: 90C109, 2014.

[2] H. Román-Flores, Y. Chalco-Cano, M.A. Rojas- Medar. Convolution of fuzzy

sets and applications. Comput. Math. Appl., 46: 1245-1251, 2003.

[3] G. Castellano, A.M. Fanelli, C. Mencar. An empirical risk functional to improve

learning in a fuzzy-neuro classifier. IEEE Trans. Systems Man Cybernet., 34: 725-
731, 2004.

[4] F.G. Guimarães, F. Campelo, R.R. Saldanha, J.A. Ramı́ez. A hybrid

methodology for fuzzy optimization of electromagnetic devices. IEEE Trans.
Magnetics, 41: 1744-1747, 2005.

5
[5] Y. Chalco-Cano, H. Román-Flores, Fernando Gomide. A new type of
approximation for fuzzy intervals. Fuzzy Sets and Systems, 159: 1376-1383, 2008.

[6] Y. Chalco-Cano, A.D. Bez-Snchez, H. Romn-Flores, M.A. Rojas-Medar. On

the approximation of compact fuzzy sets. Computers and Mathematics with
Applications, Volume 61, Issue 2, January 2011, Pages 412-420.

[7] P. Diamond, P. Kloeden. Metric Spaces of Fuzzy Sets. World Scientific,

Singapore, 1994.

[8] C. Wu, M. Ma. The Basic of Fuzzy Analysis (in Chinese). National Defence
Industry press, Beijing, 1991.

[9] D. Dubois, H. Prade. Fundamentals of Fuzzy Sets, The Handbooks of Fuzzy Sets
Series, Vol. 7. Kluwer Academic Publishers, London, 2000.

[10] R. Goetschel, W. Voxman, Elementary Calculus, Fuzzy Sets and Systems, 18(1)
(1986) 31-43.