KYBERNETIKA — VOLUME // (1975), NUMBER 5

Fuzzy Metrics and Statistical Metric Spaces

IVAN KRAMOSIL, JIŘÍ MICHÁLEK

The aim of this paper is to use the notion of fuzzy set and other notions derived from this one
in order to define, in a natural and intuitively justifiable way, the notion of fuzzy metric. The
notion is then compared with that of statistical metric space and both the conceptions are proved
to be equivalent in certain sense.

The adjective "fuzzy" seems to be a very popular and very frequent one in the
contemporary studies concerning the logical and set-theoretical foundations of
mathematics. The main reason of this quick development is, in our opinion, easy to
be understood. The surrounding us world is full of uncertainty, the information we
obtain from the environment, the notions we use and the data resulting from our
observation or measurement are, in general, vague and incorrect. So every formal
description of the real world or some of its aspects is, in every case, only an approxima­
tion and an idealization of the actual state. The notions like fuzzy sets, fuzzy orderings,
fuzzy languages etc. enable to handle and to study the degree of uncertainty mentioned
above in a purely mathematic and formal way. A very brief survey of the most interest­
ing results and applications concerning the notion of fuzzy set and the related ones
can be found in [l].
The aim of this paper is to apply the concept of fuzziness to the clasical notions
of metric and metric spaces and to compare the obtained notions with those resulting
from some other, namely probabilistic statistical, generalizations of metric spaces.
Our aim is to write this paper on a quite self-explanatory level the references being
necessary only for the reader wanting to study these matters in more details.

Definition 1. Let I be a non-empty set. A fuzzy set A in X is a pair <X,/X>

where fA is a function defined on X and taking its values in the set <0, 1> of reals.
Intuitively speaking a fuzzy set is defined, supposing a set X is given and to every
element of this set a real non-negative, not greater than 1, is ascribed expressing the
degree or the likelihood of the membership of this element in the considered fuzzy
set. The sets in the "classical" sense can be considered as a special case of fuzzy sets,
especially those for which fA takes only values 0 or 1, as in this case fA reduces to
a characteristic function defining a subset of the set X. Therefore fA is usually called
the (generalized) characteristic function of the fuzzy set A in X.
In the following the basic set X is considered to be fixed, so instead "fuzzy set
in X" only the term "fuzzy set" will be used, fA being an exhaustive characteristic
of the fuzzy set A.

Definition 2. The binary relations of equality ( = ) and inclusion ( c ) , unary

operation of forming the complement and binary operations of forming the union
and intersection ( u , n ) for fuzzy sets are defined as follows:

(1) A = B, if for all x eX fA(x) = fB(x),

(2) Li°M = * - LtW for all xeX,

(3) A c B, if fA(x) = fB(x) for all xeX ,

Max
(4) jxU*) = (Li(4jB(X>) ^ all xeX,

(5) f*nB(x) = Mm(fA(x),fB(x)) for all xeX.

Clearly, these relations and operations are a generalization of those set-theoretic

ones and reduce to them supposing fA, fB take only the values 0 or 1. As an illustra-
tion, the well-known de Morgan laws are valid even for fuzzy sets. Actually,

(A u B)c = Ac nBc,
as
1 - Max (fA,fB) = Min (1 - fA, 1 - fB)
and
(A n B)c = AcuBc
as
1 - Min (fA,fB) = Max (1 - fA, 1 - f B ) .

Lemma 1. Let A, B be fuzzy sets. Then A = B if and only if the systems of sets

STA = {{x: x e X,fA(x) < a}, a e (0, 1>} ,

9>B = {{x:xeX,fB(x) < a } , a e ( 0 , 1>}

are identical for every a e (0, 1>.

338 Proof. If A = B, then for all x e X,fA(a) = fB(x), so for a e (0, 1>

{x: x E I , / A ( X ) < a} = {x: xeX,fB(x) < a} ,

which implies the identity of SPA

and S"B. In the opposite direction equality y A = yB
assures fA(x) = / fl (x) for all xeX, hence A = B. Q. E. D.
This assertion enables to characterize a fuzzy set A up to relation of equality of
fuzzy sets by the system S"A of classical sets.
The notion of fuzzy set can be directly extended in such a way that the notions of
fuzzy relation and fuzzy function will be obtained. A "classical" n-ary relation R
defined on the set X is clearly defined by a subset of the Cartesian product X", namely
by the subset of those n-tuples of elements of X, for which relation R(xt, x2,..., x„)
holds.

Definition 3. An n-ary fuzzy relation R in the set X is a fuzzy set in the set X",
i.e. it is a pair < Z " , / R > , where fR maps the n-tuples of elements of X into <0, 1>.
Again, considering only the function on X", the values of which are only 0 or 1,
the notion of fuzzy relation reduces to that of "classical" relation.
Now, it is the very time for us to concern our attention to the notion of metric and
metric spaces. We shall start from the well-known definition of usual metric.

Definition 4. Metric Q on the set X is a function defined on the Cartesian product

X x X and taking its values in the set £ . of reals such that the following conditions
are valid:

(1) Q(X, y) 2: 0 for all x, y e X (positivity),

(2) Q(X, y) = 0 if and only if x = y, x, y e X (identity),

(3) Q(X, y) = g(y, x) for all x,yeX (symmetry),

(4) Q(X, Z) ^ Q(X, y) + g(y, z) for all x, y, z eX (triangle inequality).

It is well-known fact that in practice when measuring a distance we are not able, in
general, to measure it precisely. This can be explicitly seen from the fact that mea-
suring several times the same distance the results may differ. Usually the average
value is taken as an appropriate approximation in such a case. There are at least
two approaches enabling to describe and to handle somehow this situation. The
first, probabilistic and statistical approach has been developed already for many
years; a brief survey can be found in [2]. The other, fuzzy approach, will be explained
later, it seems to be, as far as the authors know, an original one.
The probabilistic approach is based on the idea that the distance d(x, y) of two
points x, y is an actually existing real number, however, it is, in general, beyond our
powers and abilities to obtain its precise value. Our attempts to measure this distance 339
are, from probabilistic point of view, nothing else than random experiment that can
be formally described by random variables. If some conditions are satisfied (e.g. if
these random experiments are independent and can be repeated potentially infinite
many times) we are able, for every positive e, 8, to obtain a real d0 such that the
sentence "with a probability at least 1 — e the distance d(x, y) differs from d0 by
less than 5 "will be valid. And the sentence of this type is the maximum we are able
to obtain, no other information concerning the value d(x, y) is obtainable.
From this intuitive explanation follows that any actually obtained value re-
presenting, in some measure, the value d(x, y) is, in fact, a value taken by a random
variable and can be, therefore, characterized by its distribution function. So the
following definition seems to be quite understandable.

Definition 5. A statistical metric space over the set X is a pair <X, J ^ ) where J5"
is a mapping ascribing to every pair x, y of elements from X a distribution function
!F(x, y) (.) (denoted also Fxy(.)) under the condition that the following is valid:

(1) Fxy(k) = 1 for all k > 0 if and only if x = y, x, y e X ,

(2) Fxy(0) = 0 for all x,yeX,

(3) Fxy(k) = Fyx(k) for all x, y e X, k e (- oo, oo),

(4) If Fxy(k) = 1 and Fyz(fi) = 1, then Fxz(k + p) = 1 .

Considering the value Fxy(k) as the probability that the obtained value of distance
is smaller than k the conditions (l) — (3) of the foregoing definition can be seen to
be direct generalizations of positivity, identity and symmetry conditions in the de-
finition of "usual" metric. As far as the triangle inequality is considered the situation
is not so simple. Condition (4) expresses the weakest request, namely: if we are sure
that d(x, y) is smaller than k and if we are sure, at the same time, that d(y, z) is
smaller than /., then we can be sure that d(x, z) is smaller than k + \i. Clearly, a
demand of such a type must be admitted as its ommiting could lead to a contradiction
with the usual definition of metric, which ought to be embeddable in our definition
of statistical metric spaces supposing the distributions Fxy are those concentrated
in one point. For more details concerning this case see [3].
However, from the other side there are many reasons supporting the opinion that
this extension of the triangle inequality is too weak to lead to some interesting results.
Some discussion concerning this problem can be found in [2]. Probably, the most
serious objection concerns the fact, that the triangle inequality generalized in such
a way does not bring any limitation for those values of A, for which Fxy(k) < 1, being
for this case vacuously satisfied. If, e.g., Fxy(k) < 1 for any x, y e X, k e ( — oo, oo)
(and this is, in general, the case), then triangle inequality is an empty, tautological,
condition. In the paper [3] it is proved that for every statistical metric space there
exists a function TF defined by the following equality:

TF(a, b) = M(Fxy(X + n) : FXZ(X) = a, Fyz(n) = b).

This function TF is defined on Cartesian product <0, 1> x <0, 1> and, taking its
values in the interval <0, 1>, satisfies the following conditions:

(1) TF(a, b) = TF(c, d) for a = c,b = d,

(2) TF(a, b) = TF(b, a) ,

(3) 2X1, 1) = 1,
(4) TF(Fxz(X),Fyz(»)) = Fxy(X + n).

This way of reasoning gave arise the concept of Menger space.

Definition 6. Menger space is a pair {X, J*> satisfying the same conditions as
statistical metric spaces the condition (4) being replaced by the following one:

(4M) Fxy(X + n) = T(FXZ(X), Fyz(fi)) for all x,y,zeX, X, n e ( - oo, + oo),

where Tis a binary real function satisfying for all a, b, c, de <0, 1>

(a) T(a, 1) = a , T(0, 0) = 0

(b) T(a, b) = T(c, d) for a = c, b = d

(c) T(a, b) = T(b, a)

(d) T(a,T(b,c)) = T(T(a,b),c).

The fuzzy approach to the notion of distance follows from the idea that the distance
between two points is not an actually existing real number which we have to find
or to approximate, but that it is a fuzzy notion, i.e. the only way in which the distance
in question can be described is to ascribe some values from <0, 1> to. various sentences
proclaiming something concerning this distance. Namely, in the following we shall
limit ourseselves to the assertions claiming the considering distance to be smaller
than an a priori given real. As a justification for this limitation can serve the
following assertion.

Lemma 2. A metric Q on the set X is uniquely determined by the following relation

Re <= X x X x Ex: for all x, y e X, X e Et relation Re(x, y, X) is valid if and only
if Q(X, y) < X.
 Proof. Let QL, Q2 be two different metrics on X. Then there exists at least one 34
pair <x, y} eX x X such that QL(X, y) 4= Q2(X, y), suppose QL(X, y) < Q2(X, y).
Then
(x, y, Q2(X, y)y 6 RQX , but

<x, y, Q2(x, y)y i RQ2 , i.e. RQL 4= RQ2 . Q. E. D.

This assertion leads directly to the following definition.

Definition 7. Fuzzy metric R on the set X is a fuzzy set in the Cartesian product
X x X x EL the characteristic function fR of which satisfies:

(1) fR(x, y, X) = 0 for all x,yeX and all 1^0,

(2) fR(x, y, X) = 1 for A> 0 if and only if x = y ,

(3) fR(x, y, X) = fR(y, x, X) for all x, y e X and all Xe EL,

(4) jii(x, z> A + /0 iS S(fR(x, y, X),fR(y, z, n)), where S is a measurable binary

real function defined on <0, 1> x <0, 1> taking its values in <0, 1> and such
that S(l, 1) = 1,

(5) fx(x, y, X) is for every pair <x, y> e X x X a left-continuous and non-de-
creasing function of X such that lim/j^x, y, X) = 1, if A —> oo.
All the conditions mentioned in the foregoing Definition 7 seem to have a quite
natural interpretation. Conditions (l) — (3) generalize the conditions of identity,
non-negativity and symmetry in the usual definition of metric. These conditions
express also the fact that the properties of identity, non-negativity and symmetry
are generalized, but are not subjected to some fuzziness or uncertainty, only the
values of distances are fuzzy notions. Clearly, replacing (2) by

(2') fR(x, x,X) = l for all xeX and X> 0

we would obtain the notion of fuzzy pseudo-metric.

Condition (4) expresses probably the most weak form of the triangle inequality
saying that the likelihood of the fact that the distance between x and z is smaller
than X + fi is a function of likelihoods of the two particular assertions under the
condition that if we are sure that Q(X, y) < X and g(y, z) < n we can be also sure
that Q(X, z) < X + fi. Of course, this condition may be subjected to the same criticism
as in the case of the statistical metric spaces and it is possible to modify our definition
of fuzzy metric supposing S satisfies some more conditions.
Finally, (5) expresses the fact that if we believe, in certain degree, that a distance
is beyond a limit, we believe also, in the same or greater degree, that this distance
 is beyond any larger limit. Written in a slightly precized form: if the r e a l j ^ x , y, X)
is understood as a degree of certainty that the distance Q(X, y) is smaller than X,
it seems to be quite natural to request that for any n 2; X the inequality fR(x, y, X) ^
= A(x> y> M) should be valid. Hence, for every pair <x, y} e X x X the function
fR(x, y, X) should be a non-decreasing function of X. The set of discontinuity points
of such a function is at most countable, in every point of this type we define the
function fR to be a left-continuous one. Left continuity is chosen to enable to under-
stand fR(x, y, X) as the degree of our belief that the distance between x and y is
smaller than X. Clearly, it is also possible to suppose fR to be right-continuous in X
and to interprete the value fR(x, y, X) as the degree of our belief that the distance in
question is smaller than or equal to X. The natural assumption of finiteness of any
distance justifies the condition concerning the limit value of jR(x, y, X).

Theorem 1. Any fuzzy metric R defined on X is equivalent to a statistical metric

space <Z, IF} in the sense that for all x, y e X and for all X e (— oo, oo)

fR(x, y, X) = Fxy(X) .

Proof. Let R be a fuzzy metric on X. The conditions imposed to fR imply that

fR(x, y, X), considered for any fixed pair <x, y> e l 2 as a function of X, posesses
all the properties which a distribution function is to posess. Hence, mapping J*
ascribing to <x, y} the function Fxy(X) = fR(x,. y, X) defines a requested statistical
metric space <Z, J^> on the set X. On the other side, considering a statistical metric
space <Z, #•> and ascribing to every triple <x, y, X) from X x X x Et the value
Fxy(X) € <0, 1> we obtain, as can be easily seen when looking at the condition (1) — (5)
of Definition 6, a fuzzy metric.

Corollary. Any special type of statistical metric spaces resulting from imposing
some more conditions on the function T is equivalent, in the same sense as above, to
the special type of fuzzy metrics resulting from imposing the same conditions on the
function S.

Theorem 2. Let R be a fuzzy metric on the set X satisfying the generalized triangle
inequality in the sense of Menger

fR(x, y,X + n) ^ T(fR(x, z, X),fR(y, z, fi)) ,

where T is continuous and enables, for every n 2: 2, to define a Lebesgue-Stieltjes

measure on the Cartesian product <0, 1>". Then it is possible to construct a random
function R(co, x, y) defined, for every <x, y>, on a probability space (Q, £f, P) and
such that
P({co : R(co, x, y) < X}) = fR(x, y, X),
 P({co : R(co, x, y) < X + /.}) ^ T(P({cw : R(co, x, z) < X}),

P({co : R(co, y, z) < n}))

for every x, y, z eX, X, fieEt.

Proof. Let xu yu x2, y2, ..., x„, yn eX. For every such 2n-tuple we define the
function
FXiyi...Xnyn(Xt, X2, ..., X„) = Tn(Fxiyi(Xt), ..., FXnyn(Xn))
where
Fxtyt(Xt} = fR(xt,yblt), .=-l,2,...,n,
and
Tn(at, a2, ..., an) = T(au T(a2, T(a3,..., T(a„_1; an)...))).

The role of the basic space Q is played by the space of all real functions defined on
X x X. The only thing which rests is to check the Kolmogorov consistency con-
ditions. Q. E. D.
The previous assertions deserve some more comment. It is a well-known fact,
see e.g. [2], that it is possible to generalize the usual notion of convergence in such
a way that the obtained generalization will be adequate for statistical metric spaces.
Moreover, if T is continuous and satisfies the conditons of Menger space, then the
covergence in the considered statistical metric space implies the Levy's convergence
of distribution functions, see e.g. [4]. Theorem 1 and its corollary enable, hence,
to "translate" the notion of convergence and other resulting from it topological
notions into the language of fuzzy sets, fuzzy relations and fuzzy metrics. The authors
feel that there is an intuitive difference between probability and fuzziness, even if
both of these notions wish to describe some aspects of uncertainty connected with
events and notions in surrounding us world. The further process of introducing
fuzziness into the topological spaces theory seems to be a way enabling to achieve
explicit results claiming this difference.
There are, as far as the authors know, two papers dealing with fuzzy topological
spaces. In [5] the author follows the pattern used in the process of abstract definition
of topological spaces and investigates in which measure this pattern can be followed
supposing that instead of "usual" sets the fuzzy sets are considered. Another idea
is explained in [6], where sets are again the "usual" ones, however, the property of
belonging to the closure of a set is subjected to a fuzzification, in other words, the
closures of sets are fuzzy sets. It is proved, in [6], that this approach leads to some
interesting results which are expressible in the terms of probability theory but which
have not been studied or proved when some attempts to apply probability theory
in topology were considered.
(Received April 8, 1974.)
 REFERENCES

[1] H. W. Gottinger: Toward a Fuzzy Reasoning in the Behavioral Sciences. Ekonomicko-

matematický obzor 9 (1973), 4, 404—422.
[2] B. Schweizer, A. Sklar. Statistical Metric Spaces. Pacific Journal of Mathematics 10 (1960),
313-334.
[3] E. Nishiura: Constructive Methods in Probabilistic Metric Spaces. Fundamenta Mathematicae
67(1970), 115-124.
[4] B. Schweizer: On the Uniform Continuity of the Probabilistic Distance. Zeitschrift fur
Wahrscheinlichkeitstheorie 5 (1966), 357-360.
[5] C. L. Chang: Fuzzy Topological Spaces. Journal of Mathematical Analysis and Applications
24(1968), 1, 182-190.
[6] J. Michálek: Fuzzy Topologies. Kybernetika 11 (1975), 5, 345-354.

Dr. Ivan Kramosil, CSc, Jiří Michálek, prom, mat.; Ústav teorie informace a automatizace
ČSAV (Institute of Information Theory and Automation — Czechoslovak Academy of Sciences),
Pod vodárenskou věží 4, 180 76 Praha 8. Czechoslovakia.