Fuzziness and randomness

Irwin R. Goodman 1 and Hung T. Nguyen 2

1 Code D4223, SPAWARSYSCEN, San Diego, CA 92152-7446-USA

2 Department of Mathematical Sciences, New Mexico State University, Las
Cruces, NM 88003-800l-USA

Abstract. This paper presents a survey and some new results of the mathematical
investigation into formal connections between two types of uncertainty: fuzziness
and randomness.

1 Introduction
In his pioneering work on random elements in metric spaces, Fn3chet (1948)
pointed out that besides standard random objects (such as points, vectors,
functions), nature, science and technology offer other random elements, some
of which cannot be described mathematically. For example, to each population
of humans, chosen at random, one might be interested in its "morality," its
"political spiritj" to each town chosen at random, one might be interested in
its "form," its "beauty," .... It is clear that when we use natural language to
describe properties of things, we often run into such situations. A property p
on a collection of objects il defines a subset A of il, namely those elements of
il which possess the property p, provided that p is crisp, in the sense that each
element of il has the property or does not have it. If a property p stands for
"tall" in a human population, it is not clear how p determines a subset of il.
Concepts such as "tall" are called fuzzy concepts. Thus, examples of random
elements mentioned by Frechet are fuzzy concepts chosen at random. In 1965,
Zadeh (1965) proposed a mathematical theory for modeling offuzzy concepts,
in which fuzziness is a matter of degree and is described by membership
functions.
While it appears that fuzziness and randomness are two distinct types
of uncertainty, it is of interest to find out whether or not there exist for-
mal relations between them. This is similar to Potential Theory and Markov
Processes where a formal connection between them is beneficial : Potential
Theory provides powerful tools for studying Markov processes, Markov pro-
cesses provide probability interpretations for various concepts in Potential
Theory. This dual aspect is also reminiscent in many other areas of mathe-
matics : Fourier or Laplace tranforms are considered as appropriate according
to the need for a time domain analysis or a frequency domain analysisj or, as
an analogue with complex analyis : using complex variable z = x + iy such
as in residue representation theorems and complex Taylor series expansions,
but alternatively, at times, considering the real part (x) and the imaginary
part (y) such as in the Cauchy-Riemann criterion and in harmonic analysis.

C. Bertoluzza et al. (eds.), Statistical Modeling, Analysis and Management of Fuzzy Data
© Springer-Verlag Berlin Heidelberg 2002
4 I.R. Goodman & H.T. Nguyen

In subsequent sections we will investigate formal relationships between

fuzziness and randomness and explore their consequences.

2 Generalities on" fuzzy set: theory

A fuzzy (sub) set A of a set U is a map A : U ~ [0,1]. Note that we

use the same symbol A to denote the fuzzy concept and its mathematical
modeling by the membership function. For u E U, A(u) stands for the degree
to which u is compatible with the meaning of A. As generalized sets (where
membership functions generalize indicator functions of ordinary (crisp) sets),
operations on them form the so-called fuzzy logic. Basically these operations
are defined via operations called t-norms ( as copulas and co-copulas ), t-
conorms, negations, implications. When the set of membership values is the
unit interval [0,1], the associated fuzzy logic is termed first order fuzzy logic.
If [0,1] is replaced by a complete lattice L, then we talk about L- fuzzy
sets. For example, L could be the lattice of all [O,l]-valued functions on U
(i.e. first order fuzzy sets) with the natural partial order between functions,
and in this case, the associated logic is called second order fuzzy logic. As
an intermediate level, interval-valued fuzzy logic could be considered. For
background on mathematical foundations of fuzzy logic, see e.g. Nguye and
Walker (1996). In view of the popular interval computations, sometimes the
study of membership functions can be carried out by considering their a-level
sets, i.e.,

All! = {u E U: A(u) 2': a} for a E [0,1]

Perhaps, for philosophical or application reasons, the question of how

to obtain membership functions for fuzzy concepts is essential. Although
membership functions are assigned, in general, subjectively, say, by experts
in specific domains, it is useful to understand some factors used by experts
in this process. We mention here one explanation due to Orlowski (1994) in
the context of Decision Theory. In modeling a given fuzzy concept W, a set-
valued map S, defined on some space fl with values in the power set P(U)
of U is available. Also, a set-function J.l defined on P(fl) with values in [0,1]
is specified by experts. This set-function J.l satisfies simply the conditions
: J.l(0) = a and the monotonicity : if A, B are (crisp) subsets of fl such
that A ~ B then J.l(A) ~ J.l(B). Set-functions such as these are called fuzzy
measures by Sugeno (1974), see also Nguyen and Prasad (1998). Then the
membership function W is obtained as W(u) = J.l{w : u E S(w)}. Early in
1976, Goodman showed that any membership function W can be represented
as W (u) = P {w : u E S (w)}, where P is a probability measure on some a-field
of fl, and S is a random set in U.
 Fuzziness and randomness 5

3 Generalities on random sets

Roughly speaking, a random set is a set chosen at random. Formally, by a

random set we mean a set-valued random variable. Specifically, let E be a
subset of P(U) and £ be a O"-field of E, and (fl, A) a measurable space. Then
a random set Sin U is a map from fl -+ E such that S-l(£) ~ A. If P is a
probability on (fl, A), then the probability law of S is the probability mea-
sure P S-l on (E, £). In applications, we are mainly concerned with locally
compact (Hausdorff) topological spaces such as euclidean spaces U = JRk .
When E is the class of all closed set F(JR k ) of JRk and £ is the Borel O"-field
on F(JR k ) where F(JR k ) is topologized according to the so-called hit-and-
miss topology (see, e.g. Matheron, 1975), we talk about random closed sets
(see also Goutsias et al., 1997). Probability laws of random closed sets can
be characterized by their capacity functionals (counter-parts of cumulative
distribution functions ofrandom vectors), namely by set-functions T defined
on the class of compact sets 1C(JRk ) with values in [0,1] such that:
(i) If Kn is a decreasing sequence in 1C then T(Kn) '\. T(nKn ),
(ii) T is alternating of infinite order on 1C, i.e. T is monotone increasing
and for any K 1 , K 2 , ••• , Kn in 1C, n;::: 2,

i=l, ... ,n 0~IS;;{1,2, ... ,n} iEI

where III denotes the cardinality of the set I. This important result is referred
to as Choquet Theorem, see,e.g. Matheron (1975).
The capacity functional T of the random closed set S is defined as

T(K) = P{w: S(w) n K =I- 0 }.

T induces the covering function (or the one-point-coverage function) of

S:
7r: U -+ [0,1], 7r(u) =T({u}) = P{w: u E S(w)}

Obviously, the knowledge of the covering function 7r is not enough to

specify the capacity functional T. However, various quantities related to S
can be computed in terms of 7r alone. For example, let A denote the Lebesgue
measure on JRk, or a O"-finite measure. Then, A(S) is a non-negative random
variable whose expected value is given by

EA(S) = IIRk 7r(x)dA(x).

This formula is refered to as Robbins' Formula (1944).

6 I.R. Goodman & H.T. Nguyen

4 Random set representation of membership functions

Let A be a fuzzy set of U. The a-level sets of A can be obtained by ran-
domizing the level a. Suppose we do that by choosing a uniformly in [0, 1].
That is, we view a as a random variable, defined on some probability space
(S?, A, P) with values in [0,1] such that
P(w : a(w) :=:; x) = x for x E [0,1].
Then a induces a random set Sin U as S(w) = {u E U:A(u) ~ a(w)}
which is simply the randomized level set A,,(w)' Now the covering function of
this canonical random set, denoted as SA, is
n(u) = P(w : u E SA) = P(w : A(u) ~ a(w)) = A(u).
In other words, SA is a random set in U whose covering function coincides
with the membership function of the fuzzy concept A. It is important to
note that we start with the function A(.) and obtain a random set SA such
that A(u) = P(w : u E SA(W)). Thus the class S(A) of random sets in U
whose covering functions are identical to A(.) is not empty. Elements in S(A)
are considered as equivalent in the sense that they have the same covering
function. Of course, the covering function of any random set in U defines
formally a fuzzy subset of U.
In the case where U = JR and A(.) is upper semi-continuous (usc), i.e. all
a-level sets A" are closed (for example, when A is a fuzzy number), SA is a
random closed set in JR whose capacity functional is precisely
T(K) = sup{A(u) : u E K} for K E JC(JR) (1)
Indeed,
P(w : SA(W) n K =f. 0) = P(w : a(w) :=:; A(x) for some x E K )
= P(w : a(w) :=:; sup{A(x) : x E K}) = sup{A(x) : x E K}

This is a probabilistic proof of the fact that (1) is capacity functional of

some random set. A direct proof of this fact is interesting and of independent
interest.
(i) Let Kn E JC(JR k ) with Kn "" K E JC. We have T(K) :=:; inf T(Kn)
= a, say. Let E: > a and Bn = Aa-c n K n , n ~ 1. Since A(.) is u.s.c., the
Bn ' s are closed. Also the Bn 's are all contained in the compact Kl, and
hence (since each Bn =f. 0) nBn =f. 0. Now, Bn ~ K n , we have T(K) ~ sup
{A(x) : x E nBn } ~ a-E.
(ii) The fact that a set-function like T( K) = sup {A( x) : x E K} is
alternating of infinite order is a consequence of the following general result.
Let C be a class of subsets of some space S?, such that C contains the
empty set 0 and is stable under finite intersections and unions. Let T : C -;
[0, 00) be maxitive, i.e.
for any B, C in C, T(B U C) = max{T(B), T(C)}
 Fuzziness and randomness 7

Then, T is alternating of infinite order. The proof goes as follows.

Let Bi E C, i = 1,2, ... , n. Then T(Ur B i ) = max {T(Bi), i = 1,2, ... , n}.
Without loss of generality, assume

0::; bn = T(Bn) ::; ... ::; b2 = T(B2) ::; T(B 1 ) = b1 .

For k E {1,2, ... ,n}, let T(k) = {I <:;; {1,2, ... ,n}: and for III = k},
I E T(k), and i = 1,2, ... , n - k + 1, let 7i(k) = {I E T(k) : min I = i}.
Then T(UjEIBj) = bj for every I E 7i(k), and i = 1,2, ... , n - k + 1. Next
n-k+l
L T(UiEI) = L L T(UjEIBj)
IET(k) i=l IET(k)
n-k+l n-k+l
L L bi = L (~=D bi .
i=l IET(k) i=l
Thus,
L (-I)III+IT(U iE1 B i )
{I:0#~{1,2, ... ,n} }
n
= L( _1)k+l L T(UiEIB i )
k=l IET(k)

But
n-i
L (~-i) (_I)k = 0 for every i = 1,2, ... , n - 1
k=O
Hence
L (-I)III+IT(UiEI) = bn = T(Bn) ~ T(nk=lB k ).
{I:0#~{1,2, ... ,n} }

The range R(SA) of the canonical random set SA associated with the
fuzzy set A is nested in the sense that R( SA) = {At : 0 ::; t ::; I} is
totally ordered (by set-inclusion). It is clear that P(w : SA(W) E R(SA)) = 1.
Moreover, for any At E R(SA)'

{w : At <:;; SA(W)} = {w : A(At) <:;; [a(w), I]} = {w : a(w) ::; inf A(At )} E A
where, as usual, A(At) is the image set of the set At by the function A(.),
i.e. {A(x) : x E Ad. In fact,
VB <:;; U, {w : B <:;; SA(W)} = {w : a(w) ::; in! A(B)} E A.
This leads to the following formal definition:
8 I.R. Goodman & H.T. Nguyen

Definition 1. Let S : (st, A, P) --) (lE, £) be a random set in U. We say that

S is nested if

(i) R(S) is a totally ordered (by set-inclusion) family of subsets of U,

(ii) VB E R(S), {w : B ~ S(w)} EA.
It turns out that SA is the only nested random set representing A.

Theorem 1. Let S : (st, A, P) --) (lE, £) be a random set in U, with covering

Junction 7r. Is S is nested, then S = S" in distribution where S7r is the
canonical random set representing the Juzzy set 7r.

Proof. In view of Theorems 3 and 5 in Goodman and Nguyen (1985) (pp.

341-343), it suffices to show that

VB E R(S), P(w: B ~ S(w)) = P(w: B ~ S7r) (2)

Let B E R(S). If B = 0, then (2) is clearly satisfied since both sides are
equal to 1. Thus, consider the case where B =1= 0. Let W(B) = {D E R(S) :
0=1= DeB}.
If W(B) = 0, then for any x E B, we have, since R(S) is totally ordered

{w : x E S(w)} = {w : B ~ S(w)}. (3)

If W(B) =1= 0, then there exists some x E B n D' for any D E W(B). For
that x, (3) is clearly satisfied. Thus, whenever B =1= 0, there is an x E B such
that (3) holds.
Now

P(w : B ~ S7r(w)) = P(w : 7r(B) ~ [a(w), In

= P(w: inh(B) ~ a(w)} = inf7r(B).
Next,

P(w : B ~ S(w)) = P(w: y E S(w),Vy E B)

:::; P(w : z E S(w) for some z E B).

Thus,

P(w: B ~ S(w)) :::; inf{P(w: y E S(w)),y E B} = inf7r(B)

But, in view of (3), we have

inf7r(B) = inf{P(w: Y E S(w)) : y B}E

:::; P(w : x E S(w)) = P(w : B ~ S(w)).

Therefore (2) is proved.

•
 Fuzziness and randomness 9

Remark 1. Nested random set representations might not be appropriate for

multimodal membership functions. For example, the fuzzy concept "happy"
can have a multimodal membership function due to happiness periods in
human childhood, not adolescence, again in middle age, but not in old age,
etc ... On the other hand, membership functions of fuzzy concepts can arise
durectly from covering functions of arbitrary random sets, and as such, they
might have quite irregular shapes. As an extreme case, consider a discrete
random variable X, sat taking values in the set of integers IN with probability
density function f. X is a special case of a random set S = {X}, a singleton
random set. Its covering function n(n) = P{w : n E S(w)} = P{w : X(w) =
n} = f (n). Obviously, the random set S which represents the fuzzy set n
is not nested- quite the opposite, with all values (here, points in IN) being
mutually disjoint.

Remark 2. If A is a fuzzy subset of U, then according to Zadeh, A induces a

possibility measure on the power set P(U) : JL(B) = sup{ A(u) : u E B}. This
set-function JL is a fuzzy measure and can be represented as a Choquet integral
with respect to another fuzzy measure, namely v(B) = 0 or 1 according to
B empty or not. Indeed, we have

JL(B) = 1
00
v{u: (A(u) 2: t) n B}dt for any B s;;: U

By analogy with standard Measure Theory, we can write dJL/dv = f, and call
fthe Radon-Nikodym derivative of JL with respect to v.

Remark 3. The problem of finding all possible random sets S which represent
a given fuzzy set A of a (finite) set U will be studied in Section 6. In the next
section, we are going to establish isomorphic and homomorphic-like relations
between Zadeh fuzzy logic operators and ordinary set operators over canonical
random sets.

5 Connections with fuzzy logic

Since canonical nested random sets representing fuzzy sets via covering func-
tions are randomized level sets, the connections between operations on fuzzy
sets and operations on (random) sets are essentially based upon results con-
cerning a-level sets.
Let J be an arbitrary index set, and A (j), j E J, be fuzzy subsets of
U. Then !\jEJ A(j), VjEJ A(j) stand for fuzzy sets of U with membership
functions inf { A (j) (.) : j E J }, sup { A (j) (.) : j E J }, respectively. For any
a E [0,1]' we have:
10 I.R. Goodman & H.T. Nguyen

but in general,

U A~jl ~ ( V A(j))
jEJ jEJ a
with equality if "ix E U, sup {A(j)(x) : j E J} is attained on J. This
fact is somewhat related to Nguyen's Theorem (1978) on level-sets of fuzzy
sets defined via Zadeh's extension principle. See also Fuller and Keresztfalvi
(1990), Bertoluzza and Bodini (1998). When J is finite, the above condition
is satisfied and hence equality holds for V. Recall that the (fuzzy/as well as
the ordinary) negation of A, denoted as A', has membership function 1-A(.).
Obviously, (A')a -=I- (Aa)'.
The probabilistic versions of the above are as follows. Let A and B be
fuzzy sets of U. Viewing now a as a random variable uniformly distributed
on [0,1]. We have, "ix E U :

P(w : x E AaCw) n BaCw)) = A(x) A B(x)

= P(w : x E AaCw)) A P(w : x E BaCw))'
P(w : x E AaCw) U BaCw)) = A(x) V B(x)
= P(w : x E AaCw)) V P(w : x E BaCw))'
More generally, let J be an arbitrary index set and A(j),j E J, be fuzzy
subsets of U,then

"ix E U, P(w: x E njEJA~{w)) = inf{ACj) (x) : j E J}

-- inf {P(w .. x E A aCw)'

(j) ). J' E J}

Although only inclusion holds for U in general, we always have

Indeed, the above equation holds since

{w : x E A~iw)} = {w: A(i)(x) ~ a(w)}

~ {w : a(w) ::; sup(A(j) (x) : j E J)},

it follows that

P(w : x E UjEJA~{w)) = P(UjEJ{W : x E A~{w)})

::; P(w : x E (VjEJACj))aCw)) = sup{A(j) (x) : j E J}.

On the other hand, for each j,

 Fuzziness and randomness 11

thus

sup{A(j)(x) : j E J} = sup{P(w : x E A~{w)) : j E J}

~ P(UjEJ{W: x E A~{w)})'
As for negation, we note that, for each w,

(A')a(w) = {x E U: 1 - A(x) ?: a(w)}

= {x E U: A(x) > 1- a(w)}'

and
A~(w) = {x E U: A'(x) > 1- a(w)}.
Thus,

P{w : x E (A')a(w)} = 1 - A(x)

= P{w : x E (Aa(w))'}, \Ix E U
despite the fact that (A')a(.) =1= (A a(.))" although they are identically dis-
tributed. In summary, the map A --) SA,between fuzzy sets of U and their
nested canonical random sets, is not, in general, a "full" isomorphism with
respect to logical operations "and," "or," and "not" but it is an "and"-"or"
isomorphism. This is expected since the class of all fuzzy subsets of U equiped
with Zadeh's logical operations (max, min, 1-(.)) is not boolean, but rather
it is a browerian lattice or, equivalently, a (non-complemented) Stone lattice.
Other operations on fuzzy sets also have similar corresponding relations
in terms of covering functions of canonical random sets. For example, the
cartesian product of fuzzy sets A(j) of Uj , for j E J, denoted as TIjEJ A(j),
has membership function on TIjEJ Uj given by inf{A(j) (Uj) : j E J} which is
equal to

P {w: (Uj,j E J) E nA~{W)}

JEJ
= P {w: (Uj,j E J) E (IT
JEJ
A(j))
a(w)
}.

The cartesian sum EBjEJA(j) = (TIjEJA(j),)' has the following dual

relations:

P {w :(u.i,j E J) E EB
jEJ
A~{W)}

= P {w : (Uj,j E J) E (EB A(j)) } = sup{A(j)(uj) : j E J}.

JEJ a(w)
 12 I.R. Goodman & H.T. Nguyen

Also, multivariate versions of covering functions can be considered. Re-

placing the uniformly distributed random variable a(.) on [0,1] by a ran-
dom vector, still denoted as a = (al' a2, ... , an) : f2 -> [O,l]n where each
marginal component ai is uniformly distributed on [0,1]. The (joint) cumu-
lative distribution function of such a random vector is called an n-copula (
see e.g., Schweizer and Sklar, 1983), that is a function C : [0, l]n -> [0,1] such
that:
(i) C is n-increasing, i.e. for any box B in [0, l]n where B = [al' ell x ... x
[an, en], the quantity E sgnB(x)C(X) 2: 0, where the summation is taken
over all vertices x = (Xl"'" xn) of B (i.e. x E [0, l]n such that each Xi
is equal to either ai or ei), and sgnB(x) = 1 or -1 according to Xi = ai

°
for an even number or for an odd number of i's.
(ii) C is grounded, i.e. C(Xl,'" ,xn) = for all (Xl, ... ,xn) such that Xi =
for at least one i.
°
(iii) Any (one-dimensional) margin C i of C satisfies Ci(x) = X for X E [0,1],
where C i : [0,1] -> [0,1] is defined as Ci(x) = C(l, ... , 1, x, 1, ... ,1), X
being located at the ith position.
Copulas are closely related to t-norms: associative 2-copulas are t-norms,
2-increasing t-norms are 2-copulas. The (DeMorgan) dual (or co-) copula of
an n-copula C is defined as C* (Xl, ... ,Xn ) = 1 - C(l - Xl> ... , 1 - xn). The
basic Sklar Theorem is this. If F is the joint cumulative distribution function
ofthe random vector (Xl"'" Xn) with marginal Fj,j = 1, ... , n, then there
exists a unique n-copula C such that F(Xl"" ,xn) = C(Fl(Xl), ... ,Fn(xn))
for all Xj E R Conversely, any n-copula C together the one-dimensional Fj's
define a legitimate joint distribution function this way. See Schweizer and
Sklar (1983).
With the above notations, it can be verified that : Let A (j) be fuzzy sets
of Uj, then

= p(nf=l {w : Ui E (A(i))"'i(W)}) = C(A(1) (Ul), ... ,A(n) (un))

= P {W : (Ul,"" un) E E9(A(j))aj (W)}

jEJ

~ 1 - P {W , (Ub" Un) E (g ((A(j))aj(W)) }

= P(UjEJ{W : Uj E (A(j))",(w)})

L (( -l)IIIC(A(j)(uj),j E 1)).
0#~{1, ... ,n}
 Fuzziness and randomness 13

6 A solution for random set representation

Let A be a fuzzy subset of U. Let S(A) denote the class of all random sets
on U having A as their common covering function. We are going to specify
S(A) in the case where U is finite.
Without lost of generality, assume U = {I, 2, ... ,n}. A random set S on
U is characterized by its probability function f : P(U) -7 [0,1] where

f(B) = P(w : S(w) = B),

P(U) denoting the power set of U. If we let Vi(w) = Is(w)(i) (indicator

function of the set S(w)), i E U, then Vi is a random variable taking values
in {O, I} with

P(w : Vi(w) = 1) = P(w : i E S(w)) = 1 - P(w : Vi(w) = 0)

The distribution of the random vector Vs = (Vi, V2, ... , Vn) is completely
determined by that of S and vice versa. Indeed, for any x = (Xl, ... , Xn) E
[O,I]n,

P(w : Vs(w) = x) = P{w : (Vi (w), ... , Vn(w)) = x}

= P(w: S(w) = B) where B = {i E U: Xi = I}
For any S E S(A), the cumulative distribution function (cdf) of each Vi is

GA(i) (x)
0,
= { 1- A(i),
if X < °
ifO:S;x<1
1, if 1:S; X

If the fuzzy set A is given, then the marginal cdfs G A(i), i = 1, ... , n

where C is an n-copula.
Thus,

Theorem 2. Let A be a fuzzy subset of a finite set U. Then all possible

distributions for random sets S in S(A) are of the form C(GA(u),U E U)
with C being any lUI-copula. In particular, the canonical nested random set
A",(.) corresponds to the choice of the copula C(Xl"" ,xlUl) = min{xi : i =
1, ... , lUI}, Xi E [0,1].

Remark 4. The above result can be readily extended to the case of a finite
number of fuzzy sets on finite domains. Specifically, let A (j) be fuzzy sets of
Uj, j E J (a finite index set). Then the joint distribution of (Sj, j E J),
where each Sj E S(ACi)), is of the form C(GA(j)(x), X E Uj , j E J) where C
is a I UjEJ (Uj X {j}) I-copula
14 LR. Goodman & H.T. Nguyen

Moreover, we have the following homomorphic-like relations:

Theorem 3. For any Uj E Uj , j E J,

(i)

p(njEJ{W: Uj E Sj(w)}) = L ((-1)IKI+IC*(A(j)(uj),j E K))

0i-Kr;J

where each Sj has distribution C(GA(j)(u), U E Uj ).

(ii)

Proof. (i) Let Vj denote the indicator function of {w : Uj E Sj(w)}. We have

p(njEJ{W: Vj(w) = 1}) = L ((-1)IKlp(njEK{w: Vj = O}))

Kr;,J

with the convention that when K = 0, the corresponding term is 1. But for
K =f. 0,
p(njEK{W : Vj(w) = O}) = C(P(w: Vj(w) :::; 0) : j E K)
= C(P(w : Vj(w) = 0) : j E K) = C((1- A(j)(uj)) : j E K)
= 1 - C* (A(j) (Uj) :j E K).

Thus,

p(njEJ{W : Uj E Sj(w)}) = p(njEJ{W: Vj(w) = 1})

= 1+ L ((_1)IKI(1_ C* (A(j) (u)) : j E K)

= L ((-1)IKI+1C*(A(j)(uj): j E K).
0i-Kr;,J

(ii) By DeMorgan dual:

P(UjEJ{W: Uj E Sj(w)}) = 1- p(njEJ{W: Uj E (Sj(w))'})

= 1- p(njEJ{W : Vj(w) = o}) = C*(A(j) (Uj) : j E J). •

Remark 5. The expression on the right hand side of Theorem 3( (i)) in general
does not coincide with C(A(j)(uj),j E J). But, by choosing a restricted family
for C (and thus C*) it does coincide. This includes C = min or product.
This is also related to the issue of characterizing homomorphic-like relations
for arbitrary finite combinations of C and C* acting upon the A(j)'s. See
Goodman (1994).
 Fuzziness and randomness 15

7 Some applications

Random set representations of fuzzy concepts can be useful in interpretating

fuzzy concepts or in manipulating them using random set operations. On
the other hand, random sets can be used to define more appropriate fuzzy
conceptsa) First, here are some formal random set representations of other
fuzzy concepts. A discrete fuzzy partition of a set U is a countable collection
of fuzzy subsets An of U. As such we have, for any u E U, 'En An(u) = 1. A
clear connection with probability is this. For each u E U, we have a discrete
probability density function On IN, namely

fu : IN - t [0,1], fu(n) = An(u).

Extending the concept of projection of ordinary sets, the fuzzy projection

of a fuzzy subset A of a cartesian product space U x V onto the factor space
U has membership function given by

proju(A)(u) = sup{A(u,v): v E V},u E U

In view of results in Section 5, we have

proju(A)(u) = P(UvEv{w : u E A~(w)}) = P{w : u E proju(Aa(w))}

where AV(u) = A(u,v).
Operations of fuzzy numbers are defined via the well known extension
principle. For example addition of two fuzzy numbers A and B (fuzzy subsets
of the real line lR with appropriate properties) is defined as

(A + B)(x) = sup{min(A(u), B(v)) : u + v = x}

From Section 5, we get

(A + B)(x) = P{w : x E Aa(w) + Ba(w)}'

Next let us discuss some interpretations of relative attributes of popula-
tions involving fuzzy cardinality. Consider, for example, two attributes A =
"blond," and B = "tall," fuzzy subset of a (finite) population U. Recall the
the fuzzy cardinality of A, say, is 'E"EU A(x). Let C be a 2-copula, then a
fuzzy conjunction can be defined as H(t, s) = s + t - C*(t, s), and the total
fuzzy cardinality of A conjoined with B is given by 'E"Eu H(A(x), B(x)).
The average degree to which attribute B is implied by A can be defined as

av(AIB) = L H(A(x), B(x))j L B(x).

xEU
16 I.R. Goodman & H.T. Nguyen

It turns out that Robbins' formula can provide a probabilistic interpre-

tation for this quantity. Indeed, let Sand W be random subsets of U
corresponding to A and B for some choice of copula C. Expressing Robbins'
formula for the case of U being finite and the measure ji, being the counting
measure on U, using the fact that

\/X E U, A(x) = P(w : x E S(w)) = PS-l(Fx )

where Fx = {B ~ U : x E B}, we get

L
xEU
A(x) = 1aEP(U)
ji,(a)d(PS-l)(a) = E(ISI).

Applying this to the two components of av(AIB) yields

av(AIB) = E(IS n WI)/ E(IWI).

which can be interpreted as a sort of conditional expectation of cardinalities
of the two random set representations.
Note that for U ~ IRn and>' being the Lebesgue measure, a continuous
version of the above interpretation is also valid since Ju A(x)d>.(x) = E(ISI).
Another application of Robbins'formula is to a new random set interpre-
tation of the well-established concept of center of area (GOA) or center of
gravity of a membership function which is widely used in fuzzy inference,
e.g. in fuzzy control designs, see e.g. Nguyen and Sugeno, 1998. Specifically,
consider the situation where control laws of complex dynamical systems are
derived from rules of the form

Rj = "If Xl is A jl ,... , Xn is A jn , then y is B j ," j = 1,2, ... , r

where x = (Xl, ... , Xn) E IR n is the input vector, the Aji's and Bj's are
linguistic labels (fuzzy concepts in a natural language), y is the output scalar,
say. The approach to modeling and deriving control laws using Fuzzy Sets
Theory is called Fuzzy Control in the literature. First, it consists of modeling
the above linguistic labels by fuzzy sets of appropriate spaces. Next, decide
upon some inferential method to translate and combine information. For
example, the strength of each above rule R j is defined as

then the (typical) fuzzy output is obtained as a weighted average of the fuzzy
sets Bj's, via the extension principle, generalizing Minkowski's operations on
ordinary sets :
r

Bx(.) = L wj(x)Bj(.).
j=l
 Fuzziness and randomness 17

For a given input x, the typical value Bx is precisely the " expected
value" of the random fuzzy set which takes values B j with corresponding
probabilities Wj(x), j = 1,2, ... ,r. The numerical output y, when the input
is x, is obtained by "defuzzifying" the fuzzy set Bx :

y(x) = [fIR YBx(Y)dY] . [fIR Bx(Y)dy]-l

This value is called the center-of-area (or centroid) of the function B x (.),
which is in fact the expected value of a random variable with conditional
density
f(yjx) = [Bx(Y)]' [fIR Bx(Y)dy]-l
Now, let D be a Borel set of IR. Then, the centroid of Dis

C(D) = [l/>.(D)] 1 xd>.(x) ,

where>. denotes Lebesgue measure. Using again the fact that, if S is a random
set in U (~ IR)having the fuzzy set (on U) A as its covering function, then

we get
COA(A) = [E(C(S)>'(S))]· [E(>'(8))]-1
which is a sort of conditional expected centroid of 8 with respect to >'(8).
A multivariate version (i.e. for D ~ IRn) of the above exists using vector
means. Note that in many situations, such as when A(.) is strictly unimodal,
symmetric, then, with 8 being the canonical random set, COA(A) can be
taken as E(C(8)) in some approximate fashion.
Modeling and manipulating fuzzy rules are essential in applying fuzzy
technology to real world problems. A fuzzy rule is an uncertain conditional of
the form: "If X is A then Y is B" where X, Y are variables and A, B are fuzzy
subsets. Since such rules are in general uncertain, it is necessary to attach
to them their degrees of applicability, indicating their "strength". Two issues
arise. What is the logic behind using some form of implication modeling? How
the empirical "weights of evidence" of rules are obtained? The two issues are
related. In fact, from a practical point of view, the (empirically obtained)
strengths of the rules should dictate how the rules are going to be modelled.
This is typical in probabilistic systems in which conditional probabilities are
empirical strengths for "If... Then .... " rules. While in fuzzy systems, i.e.
systems involving fuzzy sets, there are several ways to model rules and to
extract their strengths consistently, e.g. in standard fuzzy control, it seems
that a satisfactory mathematical concept for " fuzzy conditionals" is lacking.
In the following, we will indicate a solution to this problem, based upon our
18 LR. Goodman & H.T. Nguyen

work on conditional events (see. e.g. Goodman et al., 1991, Goodman et al.,
1997) and on random set representations of fuzzy sets.
A rule can be written as a conditional A ==* B, where A and B are crisp
sets, elements of a (O")-algebra A of subsets of n. In the context of boolean
logic, it is natural to interprete such a rule as material implication, i.e. A'UB.
However, if the strength of A ==* B is quantified as conditional probability
P(BJA) (= P(A n B)/ P(A)), then obviously the material implication is not
appropriate, since

P(A' U B) = P(BJA) + P(A')P(B'JA) > P(BJA)

in general, unless P(A) = 1 or P(B' n A) = O. Thus, a basic question is : how
to model A ==* B mathematically so that the assignment of P(BJA) to it is
consistent? One thing is clear: (BJA) cannot be an ordinary event, i.e. an
element of A, in view of the so-called Lewis'Triviality Result (Goodman et
al., 1991). An approach to this problem is given in Goodman et al. (1991),
see also Hailperin (1996) and Milne (1997), where "intervals" in the boolean
ring A of the form

[AB, A -t B] = {C E A: An B ~ C ~ A' U B}

are taken to be conditional objects, denoted as (BJA). It is shown that,

with interval operations, the space of such conditional objects forms a Stone
algebra containing the sub Stone algebra of rough sets. See Pawlak (1992),
Nguyen (1992). Here, we present another approach which is boolean, called
the product space approach.

Theorem 4. Let (n, A) be a measurable space. There exist a measurable

space (n*, A*) and a map tJi : A x A - t A* which embeds (n, A) into
(n*, A*), such that for any probability measure P on (n, A), there is a prob-
ability measure P* on (n*,A*) such that P*(tJi(A,B)) = P(BJA).

Proof n* is taken to be the infinite countable cartesian product of n's, and

A* is the associated product O"-field. Define

tJi: A x A -t A*

w(A, B) = (AB) U (A' x AB) U (A' x A' x AB) U ...

where AB stands for An B, and terms like AB, A' x AB are short hand
notation for AB x n x n x .... i.e. representing the cylinder in n* with base
AB, and thus U among these terms is the union of subsets in n*. Next, for any
P on (n,A), let P* denote the infinite product measure on (n*,A*) whose
one- dimensional marginals are all identical to P. Note that the terms in the
 Fuzziness and randomness 19

definition of w(A, B) are disjoint elements of A* . We have, by construction

of P*
00

P*(w(A,B)) = LP*[(A,)n x AB]

n=O
00 00

=L P(AB) [P(A')t = P(AB) L[p(A,)]n

n=O n=O
= P(AB)/[l- P(A')] = P(AB)/P(A) = P(BIA). •

The object w(A, B) constructed above is called a conditional event, de-

noted as (BIA) (by itself, without the operator P), which is an event, not
in A, but in a bigger space A*. It is this type of mathematical objects that
we use to model conditionals A ==? B consistently with conditional proba-
bilities. When A and B are fuzzy subsets of U, we can use their random set
representations to define fuzzy conditionals. If U is finite, then such random
set representations can be chosen by lUI-copulas, otherwise, canonical nested
random sets (uniformily randomized level-sets) can be used. So let SA, SB
denote any random set representations of A, B, respectively. Then (SBISA)
is a random conditional event. The conditional fuzzy set, denoted as (BIA),
is a fuzzy subset of U* (countable product of U's) with membership function
given by

which can be explicitly evaluated. For details see Goodman et al. (1997).
Another situation where covering functions of random sets can be related
to fuzzy concepts is in the usual way fuzzy logic employs modifiers. In general,
the modifier of a fuzzy concept A of U is obtained by composition with a fuzzy
subset M of the unit interval [0,1] : M 0 A, for example, "very A" is (A)2, i.e.
take M(x) = x 2 . Now the canonical random set of such a modifier is clearly

Thus, for all u E U

M(A(u)) = P{w : u E (M 0 A)a(w)} = P{w : A(u) E Ma(w)}'

A more interesting way in representing fuzzy set modifiers is as follows.

Suppose 9 : [0,1] --t [0,1] is such that for any probability space (Q, A, P)
there always exists a larger probability space (Q*, A *, P*) and a correspond-
ing relational operator g* : A --t A*, not depending on P, such that g(P(A)) =
P*(g*(A)), for any A E A (g commutes with P). Here is a simple ex-
ample. Let m be a positive integer. Define g(s) = 8 m for 8 E [0,1], and
g*(A) = A x A x .,. x A (m times). In the same flavor, conditional events
correspond to arithmetic division of numbers. It is thus natural to look for
probabilistic counter-parts of other operations such as exponentiation, linear
20 I.R. Goodman & H.T. Nguyen

functions, polynomials, analytic functions, etc. Such probabilistic counter-

parts are called relational events, extending the concept of conditional events.
Here is an interesting example of constant-probability events. Let m,n be non
negative integer with m ~ n. Then there exists an event E(m.n) in some ap-
propriate space such that the probability of it is equal to min. Specifically,
let C E A such that P(C) E (0,1) for P on (n,A). Form

r(m,n,C) = C m - 1 X C' X c n- m

where C k= C X C x ... x C (k times), with the convention that Co xC' =

C', C' x Co = C'. Consider the corresponding n-fold product probability
space (nn,An,pn). Each r(m,n,C) is an element of An, so that we can
consider conditional events in the corresponding space ((nn)*, (An)*, (pn)*)
built from (nn, An, pn) as in the construction of conditional events presented
above. Let

E(m,n,C) = (u~lr(j,n,C)1 Uj'=l r(j,n, C)).

Since the r(j, n, C) 's are mutually disjoint, we have

(pn)*(E(m, n, C)) = [m(p(c))n-l P(C')JI[n(p(c))n-l P(C')] = min.

Note that we can in fact suppress C in the above because the construc-
tion of E(n, m, C) is valid for any such C. Applications of relational event
algebra, including constant-probability events, arise mainly in combination,
comparisons and testing for similarity of models. See again Goodman et al.
(1997) for details.

References
1. Bertoluzza, C. and Bodini, A. (1998). A new proof of Nguyen's compatibility
theorem in a more general context, Fuzzy Sets and Systems 95, 99-102.
2. Frechet, M. (1948). Les elements aleatoires de nature que1conque dans un espace
distancie, Ann. Inst. Henri Poincare X-IV, 215-310.
3. Fuller, R. and Keresztfalvi, T. (1990). On generalization of Nguyen's theorem,
Fuzzy Sets and Systems 41, 371-374.
4. Goodman, I.R. (1976). Some relations between fuzzy sets and random sets.
Preprint (cited in Nguyen, H.T. (1979) Some mathematical tools for linguistic
probabilities, Fuzzy Sets and Systems 2, 53-65.)
5. Goodman, I.R. (1994). A new characterization of fuzzy logic operators pro-
ducing homomorphic-like relations with one-point coverages of random sets.
In Advances in Fuzzy Theory and Technology. vol II ( P.P.Wang, Ed.). Duke
Univ., Durham, NC, 133-159.
6. Goodman, I.R. and Nguyen, H.T. (1985). Uncertainty Models for Knowledge-
Based Systems. North-Holland, Amsterdam.
7. Goodman, I.R., Nguyen, H.T. and Walker, E.A. (1991). Conditional Inference
and Logic for Intelligent Systems. North-Holland, Amsterdam.
 Fuzziness and randomness 21

8. Goodman, I.R., Mahler, R. and Nguyen, H.T. (1997). Mathematics of Data

Fusion. Kluwer Academic.
9. Goutsias, J., Mahler, R. and Nguyen, H.T. (1997). Random Sets: Theory and
Applications. Springer-Verlag, Heidelberg.
10. Hailperin, T.(1996). Sentential Probability Logic. Lehigh Univ. Press.
11. Matheron, G. (1975). Random Sets and Integral Geometry. J. Wiley & Sons,
New York.
12. Milne, P. (1997). Bruno de Finetti and the logic of conditional events, British
J. Philos. Sci. 48, 195-232.
13. Nguyen, H.T. (1978). A note on the extension principle for fuzzy sets, J. Math.
Anal. Appl. 64, 369-380.
14. Nguyen, H.T. (1992). Interval in boolean rings: approximation and logic,. J.
Foundations of Computing and Decision Sciences 17, 131-138.
15. Nguyen, H.T. and Sugeno, M.(1998). Fuzzy Systems: Modeling and Control.
Kluwer Academic.
16. Nguyen, H.T. and Prasad, N.R.(1998). Fuzzy Modeling and Control: Selected
Works of M.Sugeno. eRe Press, Boca Raton.
17. Nguyen, H.T. and Walker, E.A. (1996). A First Course in Fuzzy Logic. eRe
Press, Boca Raton.
18. Orlowski, S.A. (1994), Calculus of Decomposable Properties, Fuzzy Sets and
Decisions. Allerton Press.
19. Pawlak, Z. (1992). Rough Sets. Kluwer Academic.
20. Robbins, H.E. (1944). On the measure of a random set. Ann. Math. Statist. 15,
70-74.
21. Schweizer, B. and Sklar, A. (1983). Probabilistic Metric Spaces. North-Holland,
Amsterdam.
22. Sugeno, M. (1974). Theory of fuzzy integrals and its applications. PhD thesis.
Tokyo Institute of Technology, Japan.
23. Zadeh, L.A. (1965). Fuzzy sets, Inform. Control 8, 338-353.