Está en la página 1de 6

468

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 3, AUGUST 1997

Correspondence
Approximation Capability of Fuzzy Systems
Using Translations and Dilations of One
Fixed Function as Membership Functions

Let us see the following examples:


1) in a single input and single output fuzzy system, the membership functions are in the form of
1

Zhi-Hong Mao, Yan-Da Li, and Xue-Feng Zhang

Abstract This paper reported on a related study on approximation


theory of fuzzy systems. First, some basic principles were presented to
construct membership functions. Then, an approach was proposed to
form membership functions by using translations and dilations of one
fixed function (we named it basis function), which was very similar to that
in wavelets analysis. The properties of this type of membership function
reflected the advantages of the given approach. Finally, it was proved
that fuzzy systems based on such membership functions are universal
approximators under certain mild conditions on the basis function (i.e.,
integrable with nonvanishing integral and almost everywhere (a.e.) continuous). This conclusion enlarged the family of fuzzy systems, which can
be universal approximators.
Index TermsApproximation theory, fuzzy systems, membership functions, universal approximators.

I. INTRODUCTION
The study on approximation theory of fuzzy systems is very
important and necessary. In most applications of fuzzy systems, the
main design objective can be transformed to find desired mappings
from the input space to the output space, which may also be denoted
as functions. Thus, the problems of designing fuzzy systems can be
considered as approximation problems of functions. Before a type of
fuzzy systems is put into application, it is helpful if we know clearly
the basic mechanism of how they approximate a desired function
and whether they are universal approximators (i.e., whether they can
approximate any continuous functions on a compact set to an arbitrary
degree of accuracy). All this will be significant for the appropriate
design of fuzzy systems.
The approximation capability of fuzzy systems has been studied in
detail in the past few years [1][5]. It is shown that the appropriate
selection or construction of membership functions plays an essential
role to ensure the system capability. In the recent studies of B. Kosko
[1], L. X. Wang et al. [2], X. L. Zeng et al. [3], [4], and J. L. Castro
[5], different types of membership functions are discussed. From
their work, we can see that under certain conditions, fuzzy systems
can be universal approximators. In these fuzzy systems, the types
of membership functions are constrained to Gaussian membership
functions [2] or membership functions with compact supports [1],
[3][5]. However, many fuzzy systems do not belong to these classes,
the main reason being that other membership functions are used.

Manuscript received August 30, 1996; revised December 12, 1996. This
work was supported in part by the Climbing Program, National Key Project
for Fundamental Research, Beijing, China, under Grant NSC92097 and the
National Science Foundation, Beijing, China, under Grant 69682010.
Z.-H. Mao and X.-F. Zhang are with the Department of Automation,
Tsinghua University, Beijing, 100084 PR China.
Y.-D. Li is with the School of Information Science and Technology,
Tsinghua University, Beijing, 100084 PR China.
Publisher Item Identifier S 1063-6706(97)04845-5.

1 + e(0x0a)=

0 1 + e 0x10a =
(

+15

(1)

where  > 0, a 2 R;
2) in a two input (x1 ; x2 ) and single output fuzzy system, each
membership function describing (x1 ; x2 ) is the product of a
Gaussian membership function and a triangular membership
function;
3) in a multi-input x 2 Rr and single output fuzzy system, the
membership functions describing the components of vector x
are Gaussian membership functions, but the T -norm is min.
In 1), the membership functions do not belong to Gaussian membership functions or membership functions with compact supports.
In 2), the membership functions are not identical to either Gaussian
membership functions or membership functions with compact supports, but the combination of them. In 3), although the membership
functions are Gaussian membership functions, the T -norm applied
is not a product as discussed in [2]. We can easily find that
the fuzzy systems in the above examples do not belong to any
classes of fuzzy systems that have been proved to be universal
approximators by the known results. Thus, the question, Are they
also universal approximators?, or in a more general form, What
other types of fuzzy systems are universal approximators?, still
remains unanswered.
Here, we will answer whether a fuzzy system with weaker constraints to its membership functions can be a universal approximator.
The rest of the paper is designed as follows. Section II introduces the
formula of fuzzy systems. Section III presents the basic principles
of choosing and constructing membership functions and proposes the
concept of membership functions with translation factors and dilation
factors. Section IV discusses the approximation capability of fuzzy
systems using translations and dilations of one fixed function (we
call it basis function) as their membership functions and proves that
a fuzzy system is a universal approximator under conditions that its
basis function is integrable with nonvanishing integral and almost
everywhere (a.e.) continuous. Finally, some conclusions are given.
II. FORMULA

OF

FUZZY SYSTEMS

In this section, we assume that fuzzy systems are multi-input


single-output systems y : U 7! V , where U = U1 2 U2 2 1 1 1 2 Ur 
Rr (r 2 Z + ) is the input space and V  R is the output space.
A multi-output system can always be separated into a group of
single-output systems.
Consider a fuzzy system which is comprised of four principal
components: fuzzifier, fuzzy rule base, fuzzy inference engine, and
defuzzifier. Assume that the fuzzifier is the most commonly used
singleton fuzzifier and that the fuzzy rule base consists of rules in
the following form:
Rk : If x1 is A1i and x2 is A2i and 1 1 1 and xr is Ari
then y is Ck
(2)
or

Rk : If x is Ak ; then y is Ck

10636706/97$10.00 1997 IEEE

(3)

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 3, AUGUST 1997

where x = (x1 ; x2 ; 1 1 1 ; xr )T 2 U is the input vector of the


fuzzy system, y is the output variable of the fuzzy system, and
fuzzy sets Aji (ij = 1; 1 1 1 ; nj ; j = 1; 1 1 1 ; r) in Uj and Ck (k =
1; 1 1 1 ; rj=1 nj  M ) in V are linguistic terms characterized
by fuzzy membership functions Aji (xj ) and Ck (y ), respectively;
Ak = A1i 2 A2i 21 1 12 Ari is fuzzy set describing vector x in which
k = 1; 1 1 1 ; M is one-to-one corresponding to the members in the
index set f(i1 ; i2 ; 1 1 1 ; ir )jij = 1; 1 1 1 ; nj ; j = 1; 1 1 1 ; rg. Each
Rk can be reviewed as a fuzzy implication (relation) Rk : Ak 7! Ck ,
which is a fuzzy set in U 2 V with membership function Rk (x; y ) =
Ak (x)?Ck (y) = A1i (x1 )?Ai2 (x2 )?1 1 1?Ari (xr )?Ck (y), where ? is
a T -norm (the definition of T -norms can be found in [5, Appendix]).
The most commonly used T -norms are product and min.
The fuzzy inference engine is a decision-making logic that employs
fuzzy rules from the fuzzy rule base to determine a mapping from the
fuzzy sets in the input space U to the fuzzy sets in the output space
V . Let A be an arbitrary fuzzy set in U ; then each Rk of (2) or (3)
determines a fuzzy set VAR in V based on the sup-star composition

VAR (y) = sup fA(x) ? Rk (x; y)g:


x2U

(4)

The defuzzifier performs a mapping from the fuzzy sets in V to


crisp points in V . Here, we choose the defuzzifier to be the centroid
defuzzifier (also called center-average defuzzifier), which maps the
fuzzy set VAR in V to a crisp point

y = k=1
M

yk VAR (yk )

k=1

(5)

VAR (yk )

where yk is the point in V at which Ck (y ) achieves its maximum


value (when Ck is a normal fuzzy set Ck (yk ) = 1; here, we always
assume that Ck is a normal fuzzy set).
When A = Ax is a fuzzy singleton [i.e., Ax (x) = 1 and
Ax (x0 ) = 0 for x0 6= x], (5) can be expressed as the following:

y = y(x)
M

= k=1M

yk Ak (x)

k=1

Ak (x)

(6)

Equation (6) is the most commonly used expression of fuzzy systems.


III. USING TRANSLATIONS AND DILATIONS OF ONE
FIXED FUNCTION AS MEMBERSHIP FUNCTIONS
A. Basic Principles of Constructing Membership Functions
The construction of membership functions is a fundamental work
in practical applications of fuzzy systems. From linear (including
piece wise linear) membership functions to spline-based membership
functions, there have been various types of membership functions we
can choose and there have been different approaches to constructing
membership functions. Some outlines on developing the membership
functions are listed in [6].
When facing a specific problem, we should consider many factors
to choose or construct appropriate membership functions. First of all,
three basic principles should be followed. That is, the appropriate
membership functions applied should be: 1) intuitively understandable; 2) easily realizable; and 3) able to solve relatively general
problems, e.g., to the approximation problems of functions, the

469

membership functions applied should enable the fuzzy system to be


universal approximator. Principle 1) and 2) need further explanations.
1) Intuitively Understandable: The membership functions should
be intuitively understandable, and should be capable of utilizing
linguistic information when such information is available. In another
point of view, they should be model-free because fuzzy systems are
usually applied to those practical systems whose models are unknown
[8]. Thus, the membership functions could be generalized to different
problems (models). Let us look at an example. If we do not consider
the principle of intuitively understandable, we will have Proposition
1.
Proposition 1 [7]: Let f : U 7! V be a bounded function where
U; V  R. Then there is a fuzzy subset A of U and a standard model
additive fuzzy system F that uses A (and possibly Ac ) in its rules
such that f (x) = F (x) for all x 2 U .
The proof of Proposition 1 is given in [7]. In the proof, the
membership function of A is constructed as

A(x) =
where = supU f + 1;
based on the two rules

1 f (x) 0
0

(7)

= inf U f . Then construct a fuzzy system

x is A; then f is B2
(8)
If x is not A; then f is B1
(9)
where the output fuzzy sets B1 and B2 are unit area fuzzy subsets of
R with centroid at and , respectively. We will find that the fuzzy
systems output is F (x) = f (x).
If

From Proposition 1, we see that if the membership function can


be constructed arbitrary, every bounded function f : R 7! R has an
exact representation as an additive fuzzy system defined by only one
fuzzy set and two rules. However, it can be implied immediately
from (7) that A(x) is model-dependent and it cannot be seen how
linguistic information can be utilized in this exact representation.
When the approximated function f (x) is complex, A(x) is as
complex as f (x) and it is difficult or impossible to be expressed
as an understandable linguistic term [8] so we say that the exact
representation is difficult to understand. Therefore, we do not think
the construction of membership function in the proof of Proposition
1 is a good approach.
2) Easily Realizable: The membership functions should be easily
realizable. Usually the membership functions are in the form of
A(x; ) where A is a function depending on the systems input
vector x 2 Rr and parameter  2 Rq (q 2 Z + ). For example, in [3], [4] pseudotrapezoid-shaped (PTS) membership functions
A(x; a; b; c; d; h) are discussed where a; b; c; d; and h are the
adjustable parameters. By choosing different parameters, we may get
different membership functions.
To make A(x; ) easily realizable, the dimensions of  should
be finite and should not be too great. Sometimes, to a fixed design
objective, the increase of the parameters dimensions will allow a
fuzzy system to decrease the amount of fuzzy sets (fuzzy membership
functions) applied. However, large dimensions will bring difficulty
to the parameters training of the system, so the dimensions of the
parameters should be constrained within an appropriate sizeneither
too small nor too great. The approach used in the proof of Proposition
1 does not satisfy Principle 2), either.
B. Translations and Dilations of One Fixed Function

Suppose function A: Rr 7!
dilations of A have the form

[0; 1];

then the translations and

A[3(x 0 )] = A x1 0 1 ; 1 1 1 ; xr 0 r
1
r

(10)

470

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 3, AUGUST 1997

where = ( 1 ; 1 1 1 ; r )T 2 Rr is called the translation factor;


3 is an r 2 r diagonal matrix: 3 = diag f1=1 ; 1 1 1 ; 1=r g, i >
0(i = 1; 1 1 1 ; r), which is called the dilation factor. In (10), we
name A(1) basis function. Note that the concept of the translation
factor and dilation factor is very similar to that in wavelets analysis.
When A(x) is the combination of some membership functions of
the components of x = (x1 ; 1 1 1 ; xr )T based on T -norm ?, i.e.,
A(x) = A1 (x1 ) ? A2 (x2 ) ? 1 1 1 ? Ar (xr ), we have

A[3(x 0 )] = A1 x1 0 1 ? A2 x2 0 2
1
2
x
r 0 r
r
?111 ?A
:
r
Especially in the case that 1 = 2 = 1 1 1 = r =  , (10)

simplified to

A x0 :


(11)
can be

(12)

Membership functions in the form of (10) have the following


properties.
1) Since all A[3(1 0 )] are the translations and dilations of
A(1), they are similar in the shape. If A(1) is intuitively
understandable in the shape, so are all A[3(1 0 )].
2) The adjustable parameters in A[3(1 0 )] are the translation
factor and the dilation factor. The dimensions of parameters are
not too great, so it is not difficult for system training, which
satisfies the principle easily realizable.
3) Because of the dilation factor, an A[3(1 0 )] can change its
shape to an arbitrary size whether large or small. This makes it
suitable for linguistic representationwe can use it for sketchy
descriptions as well as for detailed descriptionsand because
of the translation factor, an A[3(1 0 )] can move arbitrary
in the input space so that its center (the point or points at which
A[3(1 0 )] obtains its maximum value) can erode the input
space.
4) Even without strong constraints to A(1), i.e., when A(1) is
integrable with nonvanishing integral and continuous almost
everywhere, fuzzy systems based on membership functions in
the form of (10) can be universal approximators. (Detailed
discussion of this property is given in the next section.)
The above properties reflect the advantages of using translations
and dilations of some basis function as membership functions.
IV. UNIVERSAL APPROXIMATION CAPABILITY
In this section, we consider the approximation of a function
by some element of a specific family of fuzzy systems based on
membership functions of (10). By applying membership functions of
(10), (6) turns to
M

wi A[3(x 0 ai )]

y(x) = =1M
i

=1

A[3(x 0 ai )]

(13)

where !i 2 R; ai 2 Rr , each membership function has the same


A[3(x 0 ai )] 6= 0 for all
dilation factor and the denominator M
i=1
x 2 U . Now we have a family containing functions of (13). We call
this family SA .
Next, we will analyze the distribution of SA in Lp (p > 1) (space
p
L on a compact set E contains the space of continuous functions
on E ). Therefore, we can evaluate the capability of fuzzy systems in
(13) when approximating functions in Lp .

Fig. 1. A membership function of the form 1=(1+ e0x ) 1=(1+ e0x+15 ),


which is neither a Gaussian membership function nor a membership function
with compact support.

Through out this section, we use the following notations and


definitions. To a Lebesgue (L)-measurable set E  Rr , denote m(E )
as its measure. Lp (E )  ff j E jf jp dx < 1; f is L-measurable on
E g, p > 0. If f; g 2 Lp (E ), we denote kf kp; E  [ E jf jp dx]1=p ,
and p; E (f; g )  kf 0 g kp; E . When E = Rr , k 1 kp; E , p; E are
simplified to k 1 kp and p . Let S be a function family. Suppose
that for any f 2 Lp (E ) and  > 0 there exists g in S such that
p; E (f; g) < . Then we say that S contains a subset p; E dense
in Lp (E ). If, in addition, g 2 Lp (E ) for every g in S , we say that
S is p; E dense in Lp (E ).
The following theorem establishes that under certain mild conditions on the basis function A, fuzzy systems represented by (13)
are capable of approximating arbitrarily well any function in Lp (E )
where E  Rr is an arbitrarily chosen L-measurable bounded set
whose measure is nonzero.
Theorem 1: Let A: Rr 7! [0; 1] satisfy:
1) A 2 L1 (Rr ) and R A(x) dx 6= 0;
2) A is a.e. continuous.
Then to any L-measurable bounded E  Rr [m(E ) 6= 0] and any
p 2 [1; +1), SA is p; E dense in Lp (E ).
The proof of Theorem 1 is given in the Appendix of this paper.
Theorem 1 shows that a wide class of fuzzy systems are capable
of approximating any real continuous function on a compact set to
arbitrary accuracy. The types of membership functions discussed in
[1][5] all satisfy the condition of Theorem 1. Besides, many fuzzy
systems using other types of membership functions can be proved to
be universal approximators by Theorem 1.
Let us look back at the examples presented in Section I. By
the conclusion of Theorem 1, it can be tested that now we are
able to design universal approximators using fuzzy systems based
on membership functions as described in the three examples (see
Figs. 13). We give another example as shown in Fig. 4 which depicts
a somewhat irregular basis function allowed by Theorem 1.
It is easy to find that in (11), if each Aj satisfies a) Aj 2 L1 (R)
and R Aj (xj ) dxj 6= 0 and b) Aj is a.e. continuous and the T
norm is product, then the conditions of Theorem 1 are satisfied.
However, we should notice that sometimes even each Aj satisfies
the above two conditions and A(x) = A1 (x1 ) ? 1 1 1 ? Ar (xr ) is not
allowed by the theorem because of the selection of some other T
norm instead of product. For example, consider a two-dimensional

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 3, AUGUST 1997

Fig. 2. A two-dimensional membership function constructed by the product


of a Gaussian membership function and a triangular membership function.

471

Fig. 4. A somewhat irregular-shaped membership function allowed by Theorem 1.

Comparing with (13), the adjustable parameters in (14) are reduced


in dimension size, so in some situations, applying (14) instead of (13)
may make it more convenient to train parameters.
V. CONCLUSIONS
Here, we present an approach to construct membership functions
by using translations and dilations of one fixed function (called the
basis function). We find that this approach has many advantages.
We prove, especially, that without strong constrains to the basis
function (i.e., under conditions that the basis function is integrable
with nonvanishing integral and a.e. continuous), fuzzy systems can
be universal approximators. This conclusion enlarged the family of
fuzzy systems with universal approximation capability.
However, for a particular problem, we still have not discussed the
practical problem: how to choose the basis function with appropriate
shape that can construct the optimal fuzzy system to the given
problem. Further studies are needed to answer this question.
APPENDIX
Fig. 3. A two-dimensional membership function which is the combination
of two different scaled Gaussian membership functions by using T norm of
min: min fe0(x =4) ; 0:6e0(x =16) g.

membership function A(x1 ; x2 ) = min fA1 (x1 ); A2 (x2 )g, where


Ai (xi ) = [1=(jxi j +1)2 ](i = 1; 2). It can be tested that each Ai (xi )
is integrable with nonzero integral and continuous but A(x1 ; x2 ) 2
=
L1 (R2 ).
In SA , if each i (i = 1; 1 1 1 ; r) in 3 is selected the same value
0 containing functions of
of  > 0, we can get a family SA
M

y (x) = i=1
M

!i A

i=1

x 0 ai

x 0 ai


(14)

Repeating steps in the proof of Theorem 1, we can get the following


corollary immediately.
Corollary 1: Let A satisfy the conditions in Theorem 1, then
to any L-measurable bounded E  Rr [m(E ) 6= 0] and any
0 is p; E dense in Lp (E ).
p 2 [1; +1), SA

Proof of Theorem 1: Let us first see some notations and a lemma.


To a Lebesgue (L)-measurable set E  Rr , denote 1E (1) as
its characteristic function, denote C (E ) as the space of continuous
functions on E , and C0 (E ) as the space of continuous functions
with compact support containing in E . suppf  fxjf (x) 6= 0g is
defined as the support of f . The convolution operation is denoted by
3. Let : Rr 7! R be an integrable function and define 3 (x) =
j3j(3x) = [1=(12 1 1 1 r )](x1 =1; x2 =2; 1 1 1 ; xr =r ) where
3  diag f1=1 ; 1 1 1 ; 1=r g, i > 0(i = 1; 1 1 1 ; r), j3j  det 3 =
1=( ri=1 i ), x 2 Rr .
As a direct generalization of a lemma in [9, p. 249], we have the
following.
Lemma 1: Let f 2 Lp (Rr ), p 2 [1; 1), and let : Rr 7!
R satisfy R (x) dx = 1. Then 3 3 f 2 Lp (Rr ) and
limmax f g!0 k3 3 f 0 f kp = 0.
The following is the proof of Theorem 1.
Let E  Rr [m(E ) 6= 0] be an L-measurable and bounded set,
p 2 [1; +1); f 2 Lp (E ), and " > 0.
Denote fE = f 1 1E ; then obviously fE 2 Lp (Rr ). Since C0 (Rr )
is dense in Lp (Rr ) [10, Theorem 3.14], there exists an fC 2 C0 (Rr )
such that kfC 0 fE kp < "=4. Denote h  max fsup [jfC (x)j; x 2
Rr ]; 1g.

472

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 3, AUGUST 1997

Rr 7! R be defined by (x) = A(x)= R A( ) d ;


(x) dx = 1. Thus, by Lemma 2 there is a 3 such that
k3 3 fC 0 fC kp < "=4 where 3 3 fC 2 Lp (Rr ).
Since fC has compact support, there exists T1 > 0 such that
suppfC  [0T1 ; T1 ]r . Since E is bounded, there exists T2 > 0
such that E  [0T2 ; T2 ]r . Since A3 2 L1 (Rr ), there exists T3 > 0
Let :
then R

Using Lebesgue dominated convergence theorem again, we have

lim

n!1 E

jun ( ) 0 vn ( )jp d

[ lim jun ( ) 0 vn ( )jp ] d


E n!1

such that

A(x) dx
"
A3 (x) dx < R 1=p
8
m
(
E
) 1h
R nD

[0T3 ; T3 ]r . Since R

for any D 
from (15), we have

A(x) dx 0



(15)

A3 (x) dx < :

r
3 ( 0 ai )fC (ai ) 2T
n
i=1
n
r
wn ( ) =
A3 ( 0 ai ) 2T
n
i=1
n

(17)

(18)

i=1

A3 ( 0 ai )

[0T;T ]

Since 3 3 fC 2 Lp (Rr ) and


[from (17)], one has

lim
n!1

[0T; T ]

A3 ( 0 x) dx
A(x) dx

(20)

A3 ( 0 x) dx:

(21)

jvn( )j  [hj3j(2T )r =

j(3 3 fC )( ) 0 vn ( )jp d = 0

A(x) dx]
(22)

by Lebesgue dominated convergence theorem and so

lim k(vn 0 3 3 fC )1E kp = 0

n!1

E  [0T; T ]r . Thus, there is an N1 > 0


k(vn 0 3 3 fC )1E kp < "=4 for any n  N1 .
From (19) we have jun ( )j < h; then
because

jun ( ) 0 vn ( )j jun ( )j + jvn ( )j


r
h + hj3j(2T ) :
A(x) dx

(23)

01

[0T; T ] A3 ( 0 x) dx 0 1 d :
E
A(x) dx
R

= ( 1 ; 2 ; 1 1 1 ; r )T , i 2 R(i = 1; 1 1 1 ; r), denote


[0T; T ]r +  [0T + 1 ; T + 1 ] 2 [0T + 2 ; T + 2 ] 2 1 1 1 2
[0T + r ; T + r ]; then
For

A3 ( 0 x) dx =

A3 (x) dx
(25)
[0T; T ] +
when 2 E  [0T2 ; T2 ]r , one has [0T; T ]r +  [0T3 ; T3 ]r .
[0T; T ]

A(x) dx 0

2 E , we have

[0T; T ] +

A3 (x) dx < 

so

3 ( 0 x)fC (x) dx

wn ( ) 0 1
A(x) dx

lim ju ( )jp [0T; T ]


n!1 n
R

 hp 1

(19)

lim v ( ) =
n!1 n
[0T; T ]

Therefore, using (16) for any

i = 1; 1 1 1 ; nr g consists of all points in


where the set fai 2
r
[0T; T ] of the form [0T +(2i1 T=n); 0T +(2i2 T=n); 1 1 1 ; 0T +
(2ir T=n)], (ij = 1; 1 1 1 ; n; j = 1; 1 1 1 ; r). Note that vn ( ); wn ( )
are Riemann sums for
[0T; T ] 3 ( 0 x)fC (x) dx and
r
[0T; T ] A3 ( 0 x) dx, respectively. Thus, for any 2 R

lim w ( ) =
n!1 n

2 d

A3 ( 0 ai )fC (ai )

un ( ) = i=1 n
Rr :

(16)

lim un ( )

n!1

2 d

A3 (x) dx = R A(x) dx, and

Let T = max fT1 ; T2 + T3 g; then obviously suppfC  [0T; T ]r .


Then (3 3 fC )( ) = R 3 ( 0 x)fC (x) dx = [0T; T ] 3 ( 0
x)fC (x) dx.
Note that 3 ( 0 x)fC (x) is Riemann integrable on [0T; T ]r
because it is a.e. continuous and is bounded by hj3j=[ R A( ) d ].
Also A3 ( 0 x) is Riemann integrable on [0T; T ]r .
Define vn ; wn ; un by

vn ( ) =

(26)

lim

n!1 E

jun ( ) 0 vn( )jp d  hp 1


p
= 


m(E )
A(x) dx

i.e.,

lim k(un 0 vn )1E kp  8 :

n!1

(27)

(28)

Thus, there exists N2 > 0 such that k(un 0 vn )1E kp  =8+ =8 =
for any n  N2 .
Let N = max fN1 ; N2 g; then k(uN 0 f )1E kp = k(uN 0

=4

fE )1E kp  k(uN 0 vN )1E kp + k(vN 0 3 3 fC )1E kp + k3 3


fC 0 fC kp + kfC 0 fE kp < =4 + =4 + =4 + =4 = .
N
Note that uN (x) = N
i=1 A3 (x 0 ai )fC (ai )= i=1 A3 (x 0 ai ) =
N A[3(x 0 a )]f (a )= N A[3(x 0 a )] 2 S (A; 3; a) and
i C i
i
i=1 p
i=1 p
E jy (x)j dx  (maxiMp fjwi jg) m(E ) < 1 for any py (x) 2
S (A; 3; a), i.e., y(x) 2 L (E ), so SA is p; E dense in L (E ).
The proof is complete.

such that
ACKNOWLEDGMENT

(24)

The authors would like to thank Prof. F.-S. Yang, Department of


Electrical Engineering and Applied Electronic Technology, Tsinghua,
University, Beijing, China, for his enlightening course of wavelets
analysis and Y. Huang, Department of Computer Science, Sichun
Union University, for her constructive suggestions.

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 3, AUGUST 1997

REFERENCES
[1] B. Kosko, Fuzzy system as universal approximators, in Proc. IEEE
Int. Conf. Fuzzy Syst., San Diego, CA, Mar. 1992, pp. 11531162.
[2] L. X. Wang and J. M. Mendel, Fuzzy basis functions, universal approximation, and othogonal least-squares learning, IEEE Trans. Neural
Networks, vol. 3, pp. 807814, May 1992.
[3] X. J. Zeng and M. G. Singh, Approximation theory of fuzzy systemsSISO case, IEEE Trans. Fuzzy Syst., vol. 2, pp. 162176, May
1994.
, Approximation theory of fuzzy systemsMIMO case, IEEE
[4]
Trans. Fuzzy Syst., vol. 3, pp. 219235, May 1995.
[5] J. L. Castro, Fuzzy logic controllers are universal approximators, IEEE
Trans. Syst., Man, Cybern., vol. 25, pp. 629635, Apr. 1995.

473

[6] C. H. Wang, W. Y. Wang, T. T. Lee, and P. S. Tseng, Fuzzy Bspline membership function (BMF) and its applications in fuzzy-neural
control, IEEE Trans. Syst., Man, Cybern., vol. 25, pp. 841851, May
1995.
[7] F. A. Watkins, Comments on Singh and Zeng: Approximation theory
of fuzzy systemsSISO case, IEEE Trans. Fuzzy Syst., vol. 4, p. 80,
Feb. 1996.
[8] X. J. Zeng and M. G. Singh, Authors reply to Watkins comments,
IEEE Trans. Fuzzy Syst., vol. 4, p. 81, Feb. 1996.
[9] J. Park and I. W. Sandberg, Universal approximation using radial-basisfunction networks, Neural Computat., vol. 3, pp. 246257, 1991.
[10] W. Rudin, Real and Complex Analysis. New York: McGraw-Hill,
1974.