Thermodynamics
Thermodynamics
Thermodynamics
CONNER HANSEN
1. Introduction
L. A. Zadeh’s paper Fuzzy Sets* [1] introduces the concept of a fuzzy set, provides def-
initions for various fuzzy set operations, and proves several properties regarding these
operations, culminating in a theorem analogous to the hyperplane separation theorem for
traditional sets. The paper aims to introduce and establish a groundwork for these objects,
upon which future work may be built. We cover the material in the paper, condensing or
elaborating as appropriate, in an attempt to tailor it to students in the 33X series. Finally,
we provide a brief summary of the reception and current usage of fuzzy sets in the scientific
and mathematical community.
fuzzy set S, denoted S 0 , is the fuzzy set defined by the characteristic function:
(3) fS 0 = 1 − fS
As a set is entirely defined by its characteristic function, fS will sometimes be used to
denote set S.
The intersection and union operators on fuzzy sets are defined as follows:
(4) A ∪ B = max(fA , fB )
A ∩ B = min(fA , fB )
From the definitions of minimum and maximum functions, it follows that these operations
act as expected on crisp sets, and that they maintain the associativity for all fuzzy sets.
The characterization of A ∪ B as the smallest set containing both A and B (respectively,
A ∩ B as the largest contained set) holds for fuzzy sets. As a brief demonstration of this
in the union case, let C = A ∪ B. Thus fC = max(fA , fB ) is greater than or equal to both
fA and fB and C contains both sets. At the same time, ∀x, fC = fA or else fC = fB , so
any fD greater than both will also be greater than fC and thus C is contained by any set
containing A and B.
The standard identities of crisp set logic apply to fuzzy sets, most provable by employing
some casework. The equality:
(5) 1 − max(fA , fB ) = min(1 − fA , 1 − fB )
can be easily verified by examining the three cases fA (x) {>, <, =} fB (x) for a given x,
and gives the first of the two forms of DeMorgan’s Laws:
(6) (A ∪ B)0 = A0 ∩ B 0
(A ∩ B)0 = A0 ∪ B 0
The equality:
(7) max(fC , min(fA , fB )) = min(max(fC , fA ), max(fC , fB ))
which can be verified by examining all six possible weak orderings fX ≤ fY ≤ fZ of the
characteristic functions of A, B, and C for a given x, give the first of the two distributive
identities:
(8) C ∪ (A ∩ B) = (C ∪ A) ∩ (C ∪ B)
C ∩ (A ∪ B) = (C ∩ A) ∪ (C ∩ B)
Other identities such as the idempotent, domination, and commutative laws follow directly
from the definitions of the minimum and maximum functions.
2.2. Mappings & Algebraic Operators. We define as well several algebraic operations
making use of arithmetic on characteristic functions. The algebraic product of two sets is
a set with characteristic function equal to the product of their characteristic functions:
(9) AB = fAB = fA fB
REVIEW OF FUZZY SETS 3
that is, the maximal characteristic value among all such ordered pairs. The composition
is associative, as can be demonstrated using supremum and minimum properties, but the
proof is ugly and not of interest to us.
Now we turn to fuzzy sets induced by mappings. Consider T : X → Y for X and Y
spaces (or crisp sets) in Ω. If T is one-to-one and A is a fuzzy set in X then T defines a
fuzzy set B ∈ Y :
(14) fB (y) = fA (x), y = T (x) ∀x ∈ X
and likewise T −1 defines A given B. For T not one-to-one we resolve the ambiguity in the
same manner was we did with compositions of relations–by using the maximal value.
(15) fB (y) = sup fA (x), ∀x ∈ X
y∈T −1 (y)
A convex combination of vectors u and v has the form λu + (1 − λ)v, λ ∈ [0, 1]. For
fuzzy sets, the convex combination is a ternary operator, with the third argument standing
4 CONNER HANSEN
in it as well, and they form the boundary of that interval. We will leave off finding a
counterexample in this direction, since the result is not especially useful.
Second, intersections maintain both strict and strong convexity. We will walk through
the reasoning.
Proof. Let A and B be fuzzy sets in Rn . Let C = A ∩ B.
Suppose A and B are strongly convex, let x and y be two arbitrary, distinct points, and
let z be an arbitrary convex combination of x and y. Suppose fC (z) ≤ min(fA (x), fA (y)).
Then by the strong convexity of B and by the definition of intersection, either fB (x) <
fC (z) and fC (x) = fB (x) or fB (y) < fC (z) and fC (y) = fB (y). The same logic applies if
fC (z) ≤ min(fB (x), fB (y)). Thus x, y, and z still satisfy strong convexity properties for
C.
Now suppose A and B are strictly convex, x and y arbitrary points belonging to partition
of grade α for both A and B, and z is their midpoint .5x + .5y. Then by strict convexity
of both sets, for every point u near z, fA (u) ≥ α and fB (u) ≥ α, and thus fC (u) =
min(fA (u), fB (u)) ≥ α. Thus u ∈ Γα the partition by grade α of C, and since this is true
for every u near z, z ∈ Γint
α .
We note a few interesting properties of strongly convex sets that Zadeh skipped over.
These are: the characteristic function of a strongly convex set has no zeroes, attains its
supremum at no more than one point, and has at most one maximum on any line. To show
the first we assume z is a zero of fS and take x and y to be two points such that z is a
convex combination of them. These three points determine a line, with z lying between x
and y. If fS (x) 6= 0 6= fS (y) then we have violated strong convexity. Otherwise, if x has
nonzero grade then select a new point on the line, on the same side of z as x. Repeat until
fS (x) is nonzero, and follow the same procedure with y. If for either x or y, no such point
can be found, then fS is uniformly zero on that ray so strong convexity is violated. For
the second statement, assume fS assumes its maximum at two points and note that strong
convexity is violated for any convex combination of the two points. For the final statement
assume that the maximum value of fS on the line occurs twice, and again strong convexity
is violated for convex combinations of the two points.
Definition 7 (core of a fuzzy set). If S is a fuzzy set with maximum grade M , its core,
denoted C(S) is the crisp set of all points at which M is essentially attained.
Theorem 4. If S is a convex fuzzy set, C(S) is a convex set.
Proof. Let x and y be any two points in C(S), and let l be the line segment connecting them
(the set of all convex combinations of the two points). The for any given > 0 and any δ > 0
we can choose x0 and y0 such that |x0 − x|, |y0 − y| < δ and fS (x0 ), fS (y0 ) > M − . By
the convexity of S, all points on the line segment l0 connecting x0 to y0 are also contained
in ΓM − . If we let P be the cylinder of radius δ around l (extended slightly to contain the
balls around x and y at either end), then P contains x0 and y0 , and by the convexity of
cylinders, all points on l0 are contained within P . Thus for any convex combination of x
and y we can choose any ball Bδ and any and a point with the desired membership grade
lying on the resultant segment l0 will be contained in Bδ .
8 CONNER HANSEN
Corollary 1. A strongly convex fuzzy set S ⊂ R1 attains or essentially attains its maximal
grade at exactly one point.
Proof. By lemma 3 S attains or essentially attains M at a point. As previously stated, a
strongly convex set cannot attain M at two points. Suppose it attains M at one point and
essentially attains M at another point x. Then choose any point between them, and let its
membership grade be α < M . We can choose a point close to x with membership grade
arbitrarily close to M , so we may choose a point with grade higher than alpha too violate
strong convexity. Finally suppose M is essentially attained at two points x and y. Then
by theorem 4, M is essentially attained on the interval [x, y]. Then again we can choose
any point on the interval with grade alpha and find points on either side in the interval
with grades greater than α.
We now define the shadow of a fuzzy set, that is, its projection of a fuzzy set in Rn
onto a hyperplane H (dimension n − 1) in Rn . This can be viewed as a map from every
point in Rn to the nearest point on H (mapping each line normal to the plane onto the
point at which it intersects the plane). In accordance with equation 15, the shadow takes
on the supremum of all values mapped to it. Zadeh provides the following definition for
the shadow on a hyperplane aligned with the axis:
Definition 8 (shadow of a fuzzy set (axis version)). For fuzzy set A ⊂ Rn and hyperplane
H normal to the ith basis vector, the shadow of A on H is defined by:
(21) fSH (A) (x̂) = fSH (A) (x1 , . . . , xi−1 , xi+1 , . . . , xn ) = sup [fA (x1 , . . . .xn )]
xi ∈R1
However, using dot product notation for planes, this concept can easily be stated for an
arbitrary hyperplane:
Definition 9 (shadow of a fuzzy set). For fuzzy set A ⊂ Rn , u, v ∈ Rn , and hyperplane
H defined as {x ∈ Rn |(x − u) · v = 0}, the shadow of A on H is defined by:
(22) fSH (A) (x̂) = fSH (A) (x1 , . . . , xi−1 , xi+1 , . . . , xn ) = sup [fA (x1 , . . . .xn )]
xi ∈R1
ORNOT
Lemma 2. The shadow fSH (A) (x) of convex fuzzy set A on hyperplane H is convex.
To demonstrate this, let x and y be any two points on H. Assume fA has a maximum
on the normals lines intersecting H at x and y, and let x0 and y0 be points on the lines
at which fA takes on those values. Then by the convexity of A, fA is at least equal to
min(fA (x0 ), fA (y0 )) on the segment connecting those two points. The projection of that
segment onto H will be the line segment connecting x and y, and thus convex combinations
of x and y have the necessary minimum grade. Suppose instead either one or both of the
lines have no maximum. Then we can instead take x0 or y0 so that their grades are
arbitrarily close to the supremums to show that the connecting line segments have grade
at least equal to the supremum.
Theorem 5. Let A and B be convex fuzzy sets in Rn . If SH (A) = SH (B)∀H, then A = B.
REVIEW OF FUZZY SETS 9
Before we prove this, a note on convexity. Zadeh’s initial definition of convexity for
fuzzy sets was stated int terms of partitions using weak inequalities ({x|fS (x) ≥ α}). In
the latter half of the paper he begins making use of strict partitions ({x|fS (x) > α})
without acknowledging the switch. It is easy, but still worthwhile, to verify that convexity
for weak partitions implies convexity for strict partitions, and thus that this usage is valid.
We prove this by contradiction. Assume there exists some convex fuzzy set S and some
value α such that {x|fS (x) > α} is not convex. Then for some point x with fS (x) ≤ α,
there exist two points x1 and x2 , for which x = λx1 + (1 − λ)x2 for some λ ∈ (0, 1) and
fS (x1 ), fS (x2 ) > α. We define β = min[fS (x1 ), fS (x2 )] and consider the weak partition Γβ .
Then clearly x1 , x2 ∈ Γβ and x ∈ / Γβ , so we have a contradiction.
We have thus show that for a fuzzy set S to be convex implies that both its weak and
its strict partitions are convex. We now prove theorem 5
Proof. We show that the existence of a point x0 with fA (x0 ) 6= fB (x0 ) implies the existence
of a hyperplane H such that fSH (A) (x0 ∗) 6= fSH (B) (x0 ∗) where x0 ∗ is the projection of
x0 onto H. Suppose ∃x0 with fA (x0 ) = α > β = fB (x0 ). Define the partition of B,
Γβ = {x ∈ Rn |fB (x) > β}. By convexity of Γβ , and since x0 is not contained in the set,
we can find a hyperplane F such that Γβ lies entirely on one side of it (and not on F
itself). 2 Then if we let H be a hyperplane orthogonal to F and take the shadow of B
on H, fSH (B) (x∗ ), then fSH (B) (x∗0 ) = β since a line normal to H at x0 must lie in F , and
maxx∈H fB (x) = β. On the other hand fA (x0 ) = α > β so fSH (A) (x∗0 ) ≥ α > β, and the
two shadows are unequal.
We now derive a fuzzy set analog to the convex set separation theorem, stating that two
disjoint convex sets can be separated by a hyperplane in the ambient space. We start by
defining the degree of separation of two fuzzy sets.
Definition 10 (degree of separation of fuzzy sets). Let A and B be convex fuzzy sets in
Rn , and let H be a hypersurface in Rn . Suppose ∃KH ∈ R such that fA (x) ≤ KH on one
side of H, and fB (x) ≤ KH on the other, and moreover let it be the infimum of all such
values. Then the degree of separation of A and B by H is DH = 1 − KH .
The problem of minimizing the degree of separation, as stated above, of two fuzzy sets is
beyond the scope of Zadeh’s paper. We consider a special case of this situation, where H is
a hyperplane (rather than a hypersurface), and denote the minimum degree of separation
across all hyperplanes D = 1 − M̄ , M̄ = inf H (KH ). From here on, the degree of separation
will refer to the case of H limited to hyperplanes, and we now state the separation theorem’s
analog:
Theorem 6. Let A and B be bounded convex fuzzy sets in Rn with maximal grades MA
and MB . Let M = supx∈Rn [min(fA (x), fB (x))] be the maximal grade of their intersection.
Then the degree of separation of A and B is D = 1 − M .
2Here Zadeh states that we can find “supporting hyperplane,” implying that x must lie on the boundary
0
of Γβ . This is not quite trivial since it is a result of A’s convexity and the relationship α > β, but for the
proof, the weaker statement featured above is sufficient.
10 CONNER HANSEN
Proof. We divide the problem into two cases, M = min(MA , MB ) and M < min(MA , MB ).
Case 1: Assume M = MA < MB . Then by the boundedness of B there exists a
hyperplane H such that fB (x) ≤ M for all x on one side of H. On the other side,
fA (x) ≤ M by our assumptions on M . Now suppose there exists some other hyperplane
H 0 and constant M 0 < M such that fA (x) ≤ M 0 on one side and fB (x) ≤ M 0 on the
other. We denote the set of points on the first side of the plane H + and the set of points
on the second side of the plane H − (both sets including the boundary H). Then on H + ,
fA (x) ≤ M 0 < MA and thus the core of A must lie entirely on the second side. By theorem 3
A has a core, or else attains its maximal grade, so fA > M 0 on H − , and thus if fB (x) > M 0
for any x, it occurs on H + . We now have supx∈H + (min[fA (x), fB (x)]) limited by fA (x) and
supx∈H − (min[fA (x), fB (x)]) limited by fB (x), and thus supx∈H + ∪H − (min[fA (x), fB (x)]) =
M ≤ M 0 , which contradicts our assumptions on M 0 .
Case 2: Now assume M < min(MA , MB ). Let ΓA n
M = {x ∈ R |fA (x) > M }, and
B
respectively for ΓM . For either partition to be empty would violate our assumption for
case 2, and for the partitions to not be disjoint would violate our definition of M . Since
ΓA B
M and ΓM are convex disjoint crisp sets, we simply apply the regular separation theorem
to guarantee the existence of H. The same argument by contradiction used in case 1
guarantees that no higher degree of separation is possible.
References
[1] L. A. Zadeh. Fuzzy Sets*, Information and Control, Vol. 8 (1965), 338-353.
[2] Bart Kosko. Fuzziness vs. Probability, Int. J. General Systems, Vol. 17 (1990), 211-240.