Fuzzy Sets and Systems 18 (1986) 31-43 31

North-Holland

ELEMENTARY FUZZY CALCULUS

Roy GOETSCHEL Jr. and WiLliam V O X M A N

Department of Mathematics and Applied Stat/st/cs, University of Idaho, Moscow, ID 83843,
USA

Received April 1984

Revised December 1984

In [2] Dubois and Prade introduced the notion of fuzzy numbers and defined the basic
operations of addition, subtraction, multiplication, and division. A slightly modified definition
of fuzzy numbers was presented in [4], and in that paper a metric was defined for this family of
fuzzy sets. Another less restrictive definition of fuzzy numbers can be found in [1].
In the present paper we consider fuzzy numbers from a somewhat different perspective.
Basically, we shall view fuzzy numbers in a topological vector space setting. Using the
customary vector space operations together with the metric given in [4] we ~ define
differentiation and integration of fuzzy-valued functions in ways that parallel closely the
corresponding definitions for real differentiation and integration.

Keywords: Fuzzy number, Fuzzy functions, Derivative, Integral, Differential.

1. Fuzzy numbers

In [4] a fuzzy n u m b e r is defined as follows.

Definition 1.1. A f u z z y number is a fuzzy set g : R 1 - - + I = [ O , 1] with the p r o p e r -

ties
(i) /~ is u p p e r semicontinuous,
(ii) ~ ( x ) = O, outside of s o m e interval [c, d],
(iii) there are real n u m b e r s a and b, c <~ a ~< b ~<d such that ~ is increasing o n
[c, a], decreasing o n [b, d ] a n d ~ ( x ) = 1 f o r each x ~ [ a , b].
W e let ~; d e n o t e the family o f fuzzy numbers.

Suppose that ~ :R1--*I is a fuzzy set and for O~<r~<l, define C,(~) by

f{(x,r)[l~(x)>~r} if O<r~<l,
~(~) !
Lcl(supp ~) if r = 0,
where cl(supp I~) d e n o t e s the closure of the s u p p o r t of ~. T h e n it is easily
established that IL is a fuzzy n u m b e r if and only if
(i) C , ( ~ ) is a closed and b o u n d e d interval for each r, O ~ r ~ l , and
(ii) c d ~ ) ÷ 8.
F r o m this characterization of fuzzy n u m b e r s we see that a fuzzy n u m b e r is
d e t e r m i n e d by the endpoints of the intervals C,. T h u s we can identify a fuzzy

0165-0114/86/$3.50 (~) 1986, Elsevier Science Publishers B.V. (North-Holland)

32 R. Goetschel, W. Voxman

number p with the parameterized triples

{(a(r), b(r), r ) ] 0 ~< r ~< 1} (1)
where a(r) denotes the left hand endpoint of ~(ix) and b(r) denotes the right
hand endpoint. This leads to the following characterization of a fuzzy number in
terms of the two 'endpoint' functions a and b. The proof of this theorem is
straightforward (and omitted).

Theorem 1.1. Suppose that a : I--->R 1 and b : I--->R 1 satisfy the conditions:
(i) a is a bounded increasing function,
(ii) b is a bounded decreasing function,
(iii) a ( 1 ) ~ b(1),
(iv) for 0 < k ~ 1, lim,__,k- a(r) = a(k ) and limr._,k- b(r) = b(k.),
(v) lim,._,o+ a(r) = a(0) and lim,__.o, b(r) = b(O).
Then Ix : R a--->I defined by
Ix(x) = sup{r [ a(r) ~ x ~ b(r)} (2)
is a fuzzy number with parameterization given by (1). Moreover, if Ix : R1---> I is a
fuzzy number with parameterization given by (1), then the functions a and b satisfy
conditions (i)-(v).

Suppose that Ix, v e ~: are fuzzy numbers represented by {(a(r), b(r), r)[0 ~<r ~<
1} and {(c(r), d(r), r)[0~<r~<l}, respectively. Then it is easily verified that the
Dubois-Prade definition [2] of the addition of fuzzy numbers
(Ix + v)(z) = X-Fy
max mZ
min(Ix(x), v(y)) (3)

is equivalent to the 'vector' addition of the parametric representations of Ix and v

{(a(r), b(r), r ) [ 0 ~< r ~ 1}+{(c(r), d(r), r ) [ 0 ~< r ~< 1}
= {Ca (r) + c (r), b (r) + d (r), r) I 0 ~< r ~< 1} (4)
in the sense that the parameterized triples on the right hand side of (4) define a
fuzzy number (as in (2)) that is equal to Ix + v as defined by (3).
To deal with subtraetion Dubois and Prade define the 'opposite' of a fuzzy
number Ix to be the fuzzy number v where
vCx) = IxC-x)
In parametric form, if Ix is represented by {(a(r),b(r),r)[O~r<~l}, then v
corresponds to the fuzzy number
{ ( - b ( r ) , - a ( r ) , r)[0 ~ r~< 1}. (5)
Rather than take this approach we use the vector 'opposite' of Ix,
{(-a (r), - b (r), r) I 0 ~ r ~<1}, (6)
to represent -/x. It should be noted that -Ix is not a fuzzy number.
 Elementary fuzzy calculus 33

For each fuzzy number t~ parameterized by

{(a(r), b(r), r) I 0<~ r<~ 11
and each real number c, we define the 'scalar' product ctz by
c~ = {(ca(r), cb(r), r) 10 ~< r ~< 11. (7)
With this definition it is clear that if ~ is a fuzzy number, then
-/~ = (-1)~.
The family of parametric representations of members of ~ and the parametric
representations of their 'opposites' form subsets of the vector space
F = {((a (r), b(r), r) I 0 ~< r ~< 1, a : I ~ R x and b : I---> R a are
bounded functions}
where addition and scalar multiplication in V are defined by (4) and (7). We
define a metric D on V by
D({(a(r), b(r), r) 10~< r ~< 11, {(c(r), d(r), r) 1 0 <~ r ~< 1})
= sup{max{la (r) - c(r)l, Ib(r) - d(r)lt I 0 ~ r ~< 1}. (8)
The reader may verify that D, when restricted to ~, the set of fuzzy numbers,
coincides with the m e t r i c / 9 defined in [4]. It is also clear that the vector space 7/"
together with the metric D form a topological vector space.

2. F u z z y differentiation

A function f:R~--->IT is said to be a fuzzy function. In [3] Dubois and Prade

have defined the notion of a derivative for these functions. In that paper they note
that under their definition of subtraction the 'natural' definition of a derivative
f(x + h)-f(x)
f'(x) = lira
h--~ h
is of no use, since, in general, this limit will not exist. We show, however, that if
we replace the Dubois definition of an 'opposite' with that defined by (6) in the
previous section, then this limit will exist for a wide variety of functions.

D e f i n i t i o n 2.1. Suppose that f:R~-->3 ~ (where ~ is considered a subset of the

topological vector space F ) and let Xo~ R ~. The derivative f'(xo) of f at the point
Xo is defined by
f(xo + h ) - f(xo)
f'(xo) = lim
h---#O h

provided this limit exists. The limit is taken with respect to the metric D defined
by (8) in the previous section.

In the next theorem we see that ff f'(xo) exists, then f'(xo)~ F.

34 R. Goetschel, W. Voxman
Theorem 2.2. Suppose that f : R1---> ~ ;, Xo ~ R ~ and that

lim
f(xo+ h)-f(x)
tt.--~ h
exists. Let
71 = {Ca(r),/3(r), r) I 0 ~< r ~<1}
be the parametric representation of this limit and for each x ~ R ~, let
{(a(r, x), b(r, x), r) l O<~ r<~l}
be the parametric representation of f(x). Then
(i) ~ ~ Y, and
(ii) a(r)=a~(r, xo) and /3(r)=b~(r, xo) for each r ~ I where a~ and b~ are the
partial derivatives of a and b with respect to x.

l~roof. To establish (i) it suffices to show that ot and/3 are bounded. Let e > 0 be
given. Then there is a 8 > 0 such that if 0<[h[~<~,

D([(xo + h ) - , rl) < ~.

By the definition of D, we have for each r, 0~<r<~ 1, that

[ a(r, xo + 8~) - a ( r , xo) a(r) I < ~ . (9)

Since (for fixed Xo+8 and xo) a(r, xo+8) and a(r, xo) are bounded, say by M, it
follows that ~,(r) is bounded by e +2M/8. A completely analogous argument
shows that/3(r) is bounded and hence ~ ~ V.
Part (ii) follows immediately from (9).

In the next theorem we establish a simple criterion for the existence of the
derivative.

Theorem 2.3. Suppose that f : Rl.--~ :~ is a fuzzy function and that for each x,
f(x) is represented parametrically by {(a(r, x), b(r, x), r) I 0 ~<r ~< 1}. If a~ and b~ are
continuous, then f'(x) exists for each x ~ R ~.

Proof. Let x o ~ R ~ and let [c,d] be any interval containing Xo. The partial
derivatives ax and bx are uniformly continuous on I x [c, d]; hence, given e > 0
there is a 8 > 0 such that if 1~[~<8, then for each r, 0~<r<~l,
la,,(r, xo+ O-a~(r, xo)l<e and Ibx(r,Xo+~)-bx(r, xo)l<e.
Therefore, by the mean value theorem we have that if 0<lhl<a, then

[a(r' x° + h-)h- a(r' x°) a~(r, xo)l = [a~(r, x*)_ a,,(r, xo)[ < e

where x* lies between Xo and xo+h. Similarly we find that if 0<lhl<8, then for
 Elementaryfuzzycalculus 35

each r, O~r~<l,

[b(r, xo + h_)h- b(r, xo)

b~(r,Xo)]<=~.
Hence
f(xo + h) + f(Xo)
lira = f ' (Xo)
h.-,o h
exists.
Equicontinuity of the endpoint functions with respect to x provides us with
another sut~ciency condition for the existence of the derivative. Note that in the
following theorem continuity of the functions a~ and bx with respect to r is not
required. The proof of this result is similar to the proof of Theorem 2.3 and is
omitted.

Theorem 2.4. Let ~ : R1---~ ~ r be a [uzzy [unction represented parametrically by

{(a(r, x), b(r, x), r) l r e I}
Suppose that in a neighborhood N o[ a point x o e R ~, the [amilies {a~(r, x) I r e / }
and {bx(r, x) I re I} exist and are equicontinuous with respect to x at xo. Then ['(Xo)
exists.

Next we define an ordering, ~ , for ~r--- {{(a(r), bCr), r) 10~<r~<l}e P'I a and b
are Lebesgue integrable}.
Defmition 2.5. Suppose that /z = {(a (r), b (r), r) [ 0 <~ r ~< 1} and v = {(c(r), d (r), r) [
0 ~<r ~< 1} are members of F. Then ft precedes v (/x < v) ff
1
r[a(r)+ b(r)] dr ~< r[c(r)+ d(r)] dr. (10)

This ordering is reflexive and transitive; moreover, any two elements in 9; are
comparable under ~ . Note that this ordering generalizes the usual order for
'crisp' numbers. As was pointed out by a referee this order coincides with the
order relation defined in [6] in the case that a(r) and b(r) are continuous. This
order also generalizes the order ~ ) defined in [5].
The definition of ~ given in (10) is motivated in part by the desire to give less
importance to the lower levels of fuzzy numbers. Thus, for instance, this definition
will ensure that in Figure 1, /~ ~ v, even though at the lower levels ~ would
appear to be 'greater' than v.

Notation. For each fuzzy function f: R a__,,~r we define Tf: R x_.~R x by

Tr(x) = Iox r[a(r, x) + b(r, x)] dr

where for each x, [(x) is represented parametrically by

{(a(r, x), b(r, x), r) I O<~r~
< 1}.
36 R. Goetschel, W. Voxman

_1
f"

Fig. 1.

Definition 2.6. A fuzzy function f: R ~ - - ~ : is monotonic increasing [monotonic

decreasing] if
xl<x2 ~ f(xl)<f(x2) [Xl<X2 ~ f(x2)<~f(xx)].

Observation. A fuzzy function f:R~-+~: is monotonic increasing [monotonic

decreasing] if and only if Tf:R~--->R x is monotonic increasing [monotonic de-
creasing].

In the next few theorems we see that a number of elementary results from
standard calculus can be easily generalized to fuzzy calculus.

Theorem 2.7. Suppose that f : R 1 - - ~ : and that for each x, f(x) is represented
parametrically by {(a(r, x), b(r, x), r) [ O<~r <~1}. Suppose further that {a~(r, x) l r
1'} and {bx(r, x) l r ~ I } are equicontinuous families of functions with respect to x.
Then f is monotonic increasing [monotonic decreasing] if and only if O<~f'(x)
[['(x)~< O] for each x, where
0 = {(0, O, r) {0 ~<r ~<1}. (11)

Proof. Note that Tf is differentiable; in fact,

T~(x) = r[a~(r, x) + bx (r, x)] dr. (12)

From the observation preceding this theorem it follows that f is monotonic

increasing [monotonic decreasing] if and only if T~(x)~ 0 [T~(x)~< 0] for each x.
But by Theorems 2.2 and 2.4, if(x) exists and
if(x) = {(a~(r, x), b~(r, x)r) [ 0 <~r ~ 1}.
Therefore, from (10), (11) and (12) we have O<~f' [f'~ 0].

Corollary 2.8. Under the hypothesis of Theorem 2.7, if Xo is a maximum or

minimum point with respect to ~, then f'(xo)= O.
 Elementaryfuzzy calculus 37
Lemma 2.9. If f : R 1._>~; is continuous, then Tf : R 1__>R 1 is also continuous.

Proof. Let Xoe R 1 and let e > 0 be given. Choose 8 > 0 such that ff Ix -xol < 8,
then D(f(x), f(Xo))< e. Therefore, ff IX-Xol < 8 we have
[a(r,x)-a(r, xo)l<e and IbCr,x)-b(r, xo)l<~
for each r, 0 ~<r ~<1. From this it follows that

IT,(x)-T,(xo)l IIolr[a(r,x)-a(r, xo)+b(r,x)-b(r, xo)]drl

2er dr =

and hence Tf is continuous at Xo.

As an immediate consequence of this lemma we obtain the following standard

analysis result.

Theorem 2.10. If f:[a,b]--->~: is continuous, then f attains a maximum

(minimum) value on [a, b].

Proof. Since f is continuous, it follows from Lemma 2.9 that Tf is also continu-
ous, and hence attains a maximum (minimum) value of [a, b]. To conclude the
proof simply observe that if Tf attains a maximum (minimum) at Xo~ [a, b], then
so does f.

Now we calculate the derivatives of a few fuzzy functions. In what follows we

let Sc denote the unit 'spike' at the point c, i.e.,
{~ if x = c,
So(x) = otherwise. (13)

Example 1. Translations. Let g :RI-->R 1 be a differentiable function such that

g(0) = 0, and let/~o be a fuzzy number with parametric representation
{(a(r, 0), b(r, 0), r) I r ~< 1}.
Let f:RX--->~ be the fuzzy map defined by
f(x) = ~o+ s~(~)
where Sg(~) is the 'spike' at g(x) defined by (13).
Since the parametric representation/¢ is
{(a(r, 0) + g(x), b(r, 0) + g(x), r) [ 0 ~<r ~<1},
it follows that f ' is defined by
f'(x) = {(g'(x), g'(x), r) I 0 ~< r ~< 1} = Sg,(~).
Thus, for each x, f'(x) is a fuzzy n u m b e r - t h e spike Sg,(x).
38 R. Goetschel, W, Voxman

Example 2. Let f:R1-->P be defined parametrically by

{((O,O,r)[O~r~l} if x ~ O ,
f(x)= ((_x l,/-f~_r,x~,r) l O~r~l} if x > O .
It is easily seen that f'(x) is defined parametrieally by
I((O,O, r) [O<~r~< 1} , if x < O ,
f'(x) = l{(- 14T- r, 14T- r, r) I 0 ~ r ~< 1} if x > 0.
Note that in this example the derivative of [ does not exist at x = 0, and f'(x) acts
like a simple fuzzy step function.

Example 3. Let [ : R 1.., ~: be defined parametrically by

~{(0, 0, r)[ 0~<r~< 1} if x~<0,
f(x) = [{(_x3 1,/~-__r+ x*, xS1,~-- r+ x4, r) [ 0 <~r ~< 1} if x > 0.
then (parametrically) we have
~{(0, 0, r) [0<~r~< 1} if x<~0,
['(x) = [{(_3x2 1,/~__ r+4x3 ' 3x2 1,fi--__r+4x3, r) 10~<r~< 1} if x > 0 ,
and
. . . . [{(0, 0, r) I 0<~r~< 1} if x~>0,
tx) = [{(-6x 1.,/]-- r+ 12x 2, 6x 1.,/T--r+ 12x 2, r) I 0 <~r ~< 1} if x > 0.
Note that for each x, f'(x) and ["(x) are fuzzy numbers.
Example 4. Let g :R1---> [0, w) be a differentiable function such that g(0)= 1, and
let /~o be a fuzzy number with parametric representation
((a(r, 0), b(r, 0), r) I 0~<r~< 1}.
Let f: R 1~ ~; be defined by
f(x) = g(x)tLo.
Then the parametric representation of [ is
{(g(x)a(r, 0), g(x)b(r, 0), r) [ 0~<r~<l}
and consequently
f'(x) = {(g'(x)a(r, 0), g'(x)b(r, 0), r) I 0~< r~< 1}.
Thus f'(x) is a fuzzy number if g ' ( x ) ~ 0 ; otherwise -f'(x) is a fuzzy number.

3. Integration o[ l~zzy numbers

Definition 3.1. Suppose that f:[c, d]--> ~;. For each partition P = {Xo, Xl . . . . . x.}
of [c, d] and for arbitrary points ~l, x~-l~l~<x~, let

i-1
 Elementaryfuzzy calculus 39

where zii = x~-x~_l. Then the definite integral Idf(r)dr of f over [c, d] is defined
by

Jedf(x) dx = lira Rp
IPI---*o
provided this limit exists (and is independent of the partition and the selected
points ~).

Theorem 3.2. If the fuzzy [uncaon [ :[c, d]--->ff; is continuous (with respect to the
metric D) and if for each x ~[c, d], [(x) has the parametric representation
{(a(r, x), b(r, x), r) I 0~<r~< 1},
then I d f ( x ) d x exists, belongs to ~:, and is parameterized by

{ (Ida(r, x) dx, l db(r,x)dx, r) l O<~r<~l }. (14)

Proof. That I d f(x) dx exists and is parameterized by (14) is an immediate conse-

quence of equicontinuiuty of the families of functions under the metric D. To see
that I~af(x) dx is a fuzzy number note that for each x, c~x<~d, and for
0~<rx~<r2~<l, we have
a(rx, x) ~ a(r2, x) <~b(r2, x) <~b(rl, x).
Therefore,

a(r,x)dx~ a(r2, x ) d x ~ b(r2,x)<~ b(rx, x)dx (15)

"C

and it follows from (15) and earlier remarks that I~[(x) dx satisfies conditions (i),
(ii), and (iii) of Theorem 1.1.
Observe also that for c~x<<-d and 0 < k ~ < l ,
lira a(r, x) = a(k, x) lira b(r, x) = b(k, x)
r--*k- r--~k -

and
lira a(r, x) = a(0, x), lira b(r, x) = b(O, x),
r--~O + r--~O +

It now follows from the monotone convergence theorem for integrable functions
that S~f(x) dx also satisfies conditions (iv) and (v) of Theorem 1.1. Consequently,
Ia¢f(x)dx ~ 3~, as was to be shown.

The proofs of the following theorems are completely straightforward.

Theorem 3.3. If f: RX-->~ is a continuous fuzzy function, then for each triple of
real numbers a, b, c,

I~ f(x) dx + Icbf(x) d x - I~ f(x) dx = O.

40 R. Goetschel, W. Voxman

Theorem 3.4. I[ [ : [c, d ]--->~; and g :[c, d]--* ~; are integrable fuzzy [unctions, and
i[ a and [3 are real numbers, then

(af(x)+ [3g(x)) dx = ot f(x) dx +[3 g(x) dx.

Next we obtain a Fundamental Theorem of Fuzzy Calculus.

Theorem 3.5. Suppose that [ : R1--->~; is a continuous fuzzy [unction and let
F(x) = S'~[(t) dr. Then F'(x) exists and F'(x) = [(x).

Proof. Observe that

= h~O
lira -ff
1 (\j~
( { f ~÷aa(r, t) dr, I f +hb(r, t) dr, r)

- ( ~ ' a ( r , t ) d t , ~b(r,t)dt, r) lO<~r<<-l}

=lira
h--~O -x
a(r, t) dt, -~ J. b(r, t) dt, r) I O~<r~<l ]
= lira {(a (r, ~h.,), b (r, n".,), r) [ 0 ~< r ~ Z}
h---~

where ~h,, and ~h,, lie between x and x + h. Again we make use of the metric D
and the resulting equicontinuity of the families of functions {a(r, x)} and {b(r, x)}
to conclude that the last limit in the above equalities converges to the parametric
representation of f.

4. The d/fferential and tangent lines

In this section we define the differential of a fuzzy function, and we then apply
this idea to obtain tangent line segments that consist entirely of fuzzy numbers.
We also show that these tangent line segments yield the best local 'linear'
approximation of a fuzzy function.

Definition 4.1. Suppose that [:RX-->~; is a fuzzy function and that for each x,
[(x) is represented parametrically by
{(a(r, x), b(r, x), r) I 0~<r~<l} (16)
If ['(x) exists we define the differential (if of f by
d[(x, A) = Aft(x)
where A is a real number.
 Elementary fuzzy calculus 41

Definition 4.2. Suppose that f:RX--*~: is a ditterentiable fuzzy function and that
x e R a. Then the tangent line T ~ to f at x is defined by
T~(A) = f(x) + dr(x, A).
Note that if f(x) has the parametric representation defined by (16) then T " ( A )
is represented parametrically by
T~(A ) = {(a(r, x) + a~(r, x)A, b(r, x) + b~(r, x)A, r) I 0 <~r <~1}. (17)
We wish to place a bound on A to ensure that T~(A) will be a fuzzy number. To
this end we establish the following criterion.

T h e o r e m 4.3. With the notation of (17) suppose

(i) there exists k > 0 such that
a~(r, xo) ~ k, b,(r, Xo) ~<-k, 0~<r~< 1,
(ii) there exists c > 0 such that
lab(r, Xo)l< c, Ib,.(r, xo)l<c, 0~<r~<l,
and either
(iii) a(1, Xo) = b(1, Xo) and a~(1, Xo) = bx(1, xo) or
(iv) b(1, xo) - a(1, xo) = d > 0.
If (iii) holds and Iz~l<~k/c, or if (iv) holds and Iz~l<~min(k/c,dl2~} where a =
max{lab(I, Xo)l, Ib~(1, xo)l}, then T~o(A ) is a fuzzy number.

Proof. It suffices to show that TXo(A) satisfies the conditions of Theorem 1.1.
Suppose that 0 <~ r ~ s ~ 1. Then two applications of the mean value theorem
yield real numbers 01, 02 such that
a(s, Xo) + a~(s, Xo)A - [ a(r, Xo) + a~(r, Xo)A ]
= a(s, Xo) - a(r, Xo) + [a~ (s, Xo)- a~(r, Xo)]A
= [o~(pl, Xo) + a~(o2, xo)a ](s - r)
> ~ ( k - c Izil)(s-r)~>0.
In addition, it follows from (ii) that {a(r, Xo)+ a~(r, Xo)A I r~ I} is bounded. Thus
(i) of Theorem 1.1 is established; Condition (ii) of this theorem is established in
an analogous fashion. Condition Ctii) of Theorem 1.1 obviously holds if part (iii) of
Theorem 4.3 is satisfied. Moreover, if (iv) of this theorem is satisfied, then to
see that (iii) of Theorem 1.1 holds suppose that b(1, X o ) - a ( 1 , X o ) = d > O and
that lal ~<dl2a, where a = max{lab(I, Xo)[, Ibx(1, Xo)[}. Then
b(1, Xo)+ b~(1, Xo)A - [ a ( 1 , Xo)+ a~(1, Xo)A]
= b(1, Xo)-a(1, Xo)+ (bx(1, Xo)-a~(1, Xo))A
>~d-2a lal 0.
Finally, since f(xo) is a fuzzy number, and since the functions o~(., Xo):I'-~R 1
and b~(., Xo):I'--~R 1 are continuous (a~ and b,~ exist by (ii) of Theorem 4.3), it
 42 R. Goetschel, W. Voxman
follows that TCXo~(A)satisfies the five conditions of T h e o r e m 1.1 and therefore is a
fuzzy number.

Definition 4.4. Suppose that ~, v e//'. The line in F passing through Ix in the
'direction' v is defined by
{g.+~ltER 1}
and is denoted by g, + tv.

Theorem 4.5. Suppose that f: R 1.._>~:, Xo ~ R 1 and ['(Xo) exists. Let f(xo) + tu be a
line in F passing through f(xo) in the direction v. Then there is a ~ > 0 such that if
Itl<8,
D(f(xo + t), f(xo) + W) >~D(f(xo + t), f(xo) + tf'(Xo) ).
Hence, the tangent line T xo provides the best local 'linear' approximation of f.

Proof. If no such 8 exists then u~['(xo) and there is a sequence {Ak} of real
numbers that converges to 0, such that for each k,

D ~(Xo + Ak), f(xo) + Aku) < D ff(Xo + Ak), f(Xo) + A•'(Xo)).

Since f'(xo) exists we have

klira
~ 1¼ D(f(xo+ Ak), [(Xo) + A~f'(xo))
I
k--~ ~ ak , ['(Xo) = 0
and hence
D ffCxo+ ), fCxo) + ak ) [ =0.

But this implies that the parametric representation of u is given by

{(a~(r, xo), bx(r, xo), r) [ 0 <~r <~1}
which by T h e o r e m 2.2 is also the parametric representation of ['(xo). Thus
v =f'(xo) which contradicts our assumption; therefore, the desired 8 must exist.

References

[1] D. Dubois, ModUles math6matiques de l'impr~cis et de rincertain en rue d'applications aux

techniques d'aide ~ la d6cision, Thee de doctorat d'Etat, Universit~ P. Sabatier, Toulouse (1983).
[2] D. Dubois and H. Prade, Operations on fuzzy numbers, Internat. J. of Systems Sci. 9 (1978)
613-626.
 Elementary fuzzy calculus 43

[3] D. Dubois and H. Prade, Towards fuzzy differential calculus, Part 3: Differentiation, Fuzzy Sets
and Systems 8 (1982) 225-235.
[4] R. Goetschel and W. Voxman, Topological properties of fuzzy numbers, Fuzzy Sets and Systems 9
(1983) 8"/-99.
[5] R. Goetschel and W. Voxman, Eigen fuzzy number sets, Fuzzy Sets and Systems 16 (1985) 75--85.
[6] M. Rizzi, S61ectionet classement d'actions en avenir incertaln, Document du LAMSADE No. 14
(1981).