You are on page 1of 10

SET THEORY

Herbert B. Enderton
University of California, Los Angeles
I. The Role of Set Theory in Mathematics
A. A Common Language and Framework
B. Foundational Role
II. Basic Ideas
A. Membership and Extensionality
B. Basic Operations on Sets
III. The Iterative Concept of Sets
A. Sets and Rank
B. Proper Classes
IV. Embedding Mathematics in Set Theory
A. Functions and Relations
B. Natural Numbers
C. Other Number Systems
V. Innite Sets
A. Cardinal Numbers
B. Ordinal Numbers
VI. Axiomatization
A. ZermeloFraenkel Axioms
B. Choice
C. Consistency and Incompleteness
D. Stronger Axioms
GLOSSARY
Cardinal number Measure of the size of a set (possibly innite).
Choice axiom Principle that a set can be formed by making
any number of arbitrary choices.
Continuum hypothesis Statement that the set of real numbers has the
smallest uncountable cardinality.
Equinumerous Property of being in a one-to-one correspondence.
Extensionality Principle that sets having the same members,
however described, are identical.
Ordinal number Measure of the length of a well-ordering.
Rank Ordinal number measuring the number of iterations
of the power-set operation needed to construct a set.
Well-ordering Ordering for which any nonempty set has a least
element.
Set theory is that part of mathematics concerned with the abstract concepts of sets (i.e., collections of
objects) and membership in sets. Thus the initial ideas from which set theory begins are extremely simple,
and indeed are ideas that occur throughout mathematics. It is remarkable that from simple beginnings, a
fascinating and useful theory emerges. And because set theorys concepts occur throughout mathematics,
it has been able to play a unifying role, providing both a common language and a foundational basis for
mathematics.
1
I. THE ROLE OF SET THEORY IN MATHEMATICS
A. A Common Language and Framework
Set theory provides a language for mathematics in the sense of providing concepts in terms of which
other mathematical ideas can be formulated. For example, in calculus one studies functions on the real
numbers. But what are functions? They are sets of ordered pairs with a certain property (discussed in detail
in section IV.A). And even ordered pairs can be taken to be certain sets. When in mathematics we want
to make a concept precise, we formulate its exact denitionin terms of other, more basic concepts. Those
more basic concepts in turn have denitions. But this cannot become an innite regress. At some point we
stop with primitive notions not further formally dened (but about which we hope to have clear informal
ideas). Those primitive notions are set and membership.
It is not that calculus problems become any easier to solve when viewed as problems about setsthey do
not. Rather it is that set theory provides an infrastructure that can remain unseen until it is needed, as when
one asks what exactly functions are. And it does this not simply for calculus, but for all of mathematics.
Thus it provides a universal language for mathematics, a language into which statements of calculus or other
areas couldin principlebe translated.
B. Foundational Role
The business of mathematicslike other branches of scienceis to establish what is true about its
subject. But it is not an experimental science; instead the basis of mathematical knowledge is proofs. But
proofs from what assumptions? In the early part of the twentieth century, the need to clarify exactly which
assumptions are allowable became clear. For example, the statement known as the axiom of choice (discussed
further in section VI.B) generated considerable controversy. The axiom of choice was very useful; it enabled
one to prove that every vector space has a basis, for example. But was it a correct assumption?
Here set theorys role as a universal language becomes important. If we can agree on the allowable
assumptions for set theory then we can argue as follows: The concepts from any area of mathematics can be
dened in terms of set theory. Statements about those concepts are thereby translatable into statement of
set theory. If the translated statement can be proved from the assumptions (the axioms) of set theory, then
the original statement is proved (from those axioms).
But can we agree on the axioms of set theory? The answer is partly yes and partly no. For the working
mathematician (a category generally understood to exclude set-theorists and logicians), the answer is yes
the axiomatization discussed in part VI has become the accepted standard. But one can still question
accepted standards; see section VI.C.
II. BASIC IDEAS
A. Membership and Extensionality
Set theory starts from two concepts that are primitive in the sense of not being dened in terms of
other concepts: the concept of set and the concept of membership. For these concepts we can adopt axioms
(as in part VI) but not formal denitions. Even in the axiomatic approach, however, our choice of axioms is
guided by our informal, intuitive ideas concerning sets and membership. That is, a set should be a collection
of objects (called its members), the collection being regarded as a single entity. In the simplest cases, such
as the set 2, 3, 5, 7 consisting of the four numbers listed, the concept seems clear enough. It becomes
deeperand more usefulwhen we start forming sets of sets (i.e., sets whose members are sets) and sets of
sets of sets.. . .
One aspect that needs to be claried is that set theory deals not with names of sets (which are linguistic
objects), but with sets themselves. For example, take the set consisting of the real roots of the polynomial
x
4
17x
3
+ 101x
2
247x + 210 which we can denote by
x [ x
4
17x
3
+ 101x
2
247x + 210 = 0.
Then doing a little algebra shows that this polynomial has exactly four roots, namely 2, 3, 5, and 7.
Consequently the above-named set is exactly the same set as 2, 3, 5, 7:
x [ x
4
17x
3
+ 101x
2
247x + 210 = 0 = 2, 3, 5, 7
2
That is, we have two names for the same set; that is what the displayed equation is telling us. Since
the two names denote one and the same set, any true statement about this set utilizing one name can be
automatically translated into an equally true statement utilizing the other name. The fact that sets A and
B having exactly the same members are identical (A = B) is called the extensionality principle.
B. Basic Operations on Sets
The smallest set is the empty set , having no members at all. A dierent set is , the set having one
member, namely . And is dierent from both. That is, , , is a set with three distinct
members.
The ordinary idea of union
A B = x [ x A or x B (or both)
can be generalized as follows:

/ = x [ x is a member of some set in /.


Thus x

/ if and only if x is a member of a member of /. For example,


A B =

A, B and

, , = , .
The dual concept of intersection
A B = x [ x A and x B
has the analogous generalization:

/ = x [ x is a member of every set in /


for nonempty /. (The requirement that / ,= is needed to avoid inadvertently having everything a member
of

/ for vacuous reasons.) For example,


A B =

A, B and

, = .
If all the members of a set A are also members of B, then we say that A is a subset of B (written
A B) or that A is included in B. Thus any set is a subset of itself. At the other extreme, is a subset of
every set. The fact that B is vacuously true, in the sense that the task of verifying that every member
of is also in B requires doing nothing at all.
Any set A will have one or more subsets. We can take all those subsets, and put them together into a
new set. This gives us the set of all subsets of A, called the power set TA of A. For example,
T0, 1 = , 0, 1, 0, 1.
Notice here that 0, 1 is a set of numbers, but its power set is a set of sets of numbers. The power set
operation moves us up one level in a certain ranking of sets, to be discussed further in the next section.
From sets A and B we can construct the relative complement A B consisting of those members of
A that are not in B. In contexts in which all the sets being considered are subsets of some large set U,
we can write U B simply as

B, and refer to

B as the complement of B (with the understanding that the
complement is relative to the universe U). For example, when working with sets of real numbers, we can
write

B for R B, where R is the set of all real numbers.
3
III. THE ITERATIVE CONCEPT OF SETS
A. Sets and Rank
It is possible to think of the universe of all sets as being built up by iteration of the power-set operation.
The idea is that at any stage in the process, we are allowed to construct any set all of whose members were
constructed at an earlier stage. Thus if V

is the collection of sets formed at stage , then


a V

all members of a are in V

for some less than


or equivalently
a V

a TV

for some less than .


Following this principle yields (after some simplication) for the rst few stages, V
0
= , V
1
= T, V
2
= TT,
and so forth.
But what numbers are to be used to index this construction? After any set of stages, there must be
a next stage. So even after constructing V

for every natural number = 0, 1, . . . there is still a next


stage. And a next one after that. The correct system of numbering uses the ordinal numbers, discussed
below in section V.B. The key feature of the ordinal numbers is that after any seteven an innite setof
such numbers, there is still a next one. We dene the rank of a set S to be the smallest ordinal number
for which S V

. Thus the rank of S is a measurement of how long it takes to construct all members of S.
It should be mentioned that some recent work has adopted a viewpoint towards sets that does not t
into this iterative concept. In particular, this work allows sets to have strange circularity properties, such
as being members of themselves. Although this viewpoint is not suitable for set theorys foundational role
discussed in Section I,B, these non-well-founded sets may nd applications in computer science.
B. Proper Classes
There is no set of all sets. In the construction described above, after any set of stages, there is a next
stage at which new sets emerge. We never get a set V to which every set belongs.
But we still want to talk of the collection of all sets, so we use the word class for this collection. The
class of all sets is too large to be a set; such a class is called a proper class. The class of all ordinal numbers
is also a proper class, because after any set of ordinal numbers, there are others. In general, we can say that
a class C is a set if eventually, i.e., at some stage , we have all the members of C. But if C is so large that
we never obtain all of it, no matter how large a set of stages we examine, then C is a proper class.
IV. EMBEDDING MATHEMATICS IN SET THEORY
A. Functions and Relations
The idea of a function, such as the squaring function f dened by the equation f(x) = x
2
, occurs
throughout mathematics. The graph of a function f is the set of all ordered pairs x, f(x)) for each x in
the domain of the function. For the squaring function, these ordered pairs are points in the Cartesian plane,
and the graph is a subset of the planethe familiar parabola.
To formalize the concept of a function in general, it is convenient to take a function simply to be its
graph, i.e., a certain set of ordered pairs. To begin, then, we need to incorporate into set theory the idea of
an ordered pair x, y).
As a consequence of the extensionality principle, x, y = y, x; the set x, y assigns no ordering to
x and y. By contrast, we need x, y) to encode the information that x is to precede y. The most convenient
device for doing this is to dene
x, y) = x, x, y.
Then from x, y) one can recover not only the objects x and y, but also the fact that x comes rst. The
rst component of the pair can be recovered by using the equation x =

x, y).
From sets A and B we can now form their Cartesian product
AB = x, y) [ x A and y B.
By a relation from A to B we mean a subset R of A B; the relation R tells us which members of A are
related to which members of B. More generally, a relation is simply any set of ordered pairs.
4
Any relation R has an inverse R
1
dened by the equation R
1
= x, y) [ y, x) R. If R is a
relation from A to B, then R
1
is a relation from B to A.
A graph of a function f (which is the function f) is a relation with a special property: we never have
both x, y) f and x, z) f with y ,= z (because y = f(x) = z). Call a relation single-valued if it has
this special property. Then the functions are exactly the single-valued relations. The domain of f is the set
of all rst components of ordered pairs that belong to f; the range of f is the set of all second components.
We will write f : A B to mean that f is a function with domain A and with range included in B; this
can also be stated by saying that f maps A into B.
A function f is said to be one-to-one if the inverse relation f
1
is also single-valued, i.e., if we never
have both f(y) = x and f(z) = x with y ,= z. In this case, f
1
is function whose domain is the range of f.
In addition to the functions, there are two other kinds of relations that deserve mention: the equivalence
relations and the ordering relations. Here it is convenient to write xRy to mean that the pair x, y) is a
member of R.
An equivalence relation on a set A is a relation R with R AA for which the following three conditions
are met: (1) xRx for each x in A; (2) whenever xRy then also yRx; and (3) whenever xRy and yRz then
also xRz. These three properties are called reexivity (on A), symmetry, and transitivity, respectively.
Equivalence relations are important because they correspond to ways of partitioning A. If x A and
R is an equivalence relation on A, then dene xs equivalence class [x]
R
to be the set of things to which x
is related, i.e., [x]
R
= t [ xRt. Then xs equivalence class is the same as ys equivalence class if and only
if xRy. Two dierent equivalence classes are always disjoint, and the union of all the equivalence classes is
the entire set A. Thus the set A is partitioned into the equivalence classes. We can form the quotient set
A/R of all equivalence classes. In section IV.C we will see applications of these concepts. Roughly speaking,
A/R is what A becomes when we convert R into equality by identifying related members of A.
A partial ordering is a relation R that is transitive (whenever xRy and yRz then xRz) and irreexive
(we never have xRx). The preferred symbol for partial orderings is not R but rather < or a similar symbol.
Since xRy means x, y) R, in particular x < y means x, y) <. Moreover we can write x y to mean
that either this happens or else x = y. For a partial ordering <, at most one of the three possibilities x < y,
x = y, and y < x can occur.
A linear ordering on a set A is a subset < of AA that is transitive (whenever x < y < z then x < z)
and for which exactly one of the three possibilities x < y, x = y, and y < x occurs for each x and y in A.
Such a relation is also a partial ordering, but it enjoys the additional property that for distinct x and y in
A, either x < y or else y < x, a property that fails for partial orderings in general. A subset D of A might
or might not have an element m that is less than every other member of D; if so then m is said to be the
least member of D.
A well-ordering is a special kind of linear ordering. The well-orderings correspond to constructions
built up from below. In such a construction, at any stage except the last, there is always a next step.
That is, as long as some steps remain undone, there is an earliest (least) undone step. A well-ordering of A
is a linear ordering on A with the property that every nonempty subset of A has a least element. We will
have more to say about well-orderings in section V.B.
B. Natural Numbers
The natural numbers are the nonnegative integers: 0, 1, 2, .... But what are these? They can be (which
is not to say that they must be) sets. It was proposed by Gottlob Frege that the number 2 be dened as
that property that is true of exactly those properties that, for some distinct x and y, are true of x and y and
nothing else. A set-theoretic version of this would be to dene 2 as the set containing all doubletons, i.e.,
all sets that, for some distinct x and y, have as members x and y and nothing else. But this is impossible
because the collection of all doubletons is a proper class. Instead we will dene the number 2 to be a
particular doubleton, namely the set 0, 1. We will build up the natural numbers from below, letting each
number be the set of all preceding numbers. This will have the eect of making 3 turn out to be a set with
exactly three members (which are 0, 1, and 2):
0 = ,
1 = 0 = ,
5
2 = 0, 1 = , ,
3 = 0, 1, 2 = , , , .
Here each number is obtained from the preceding number by the operation of adding exactly one new
member, namely the preceding number itself. Dene the successor a
+
of a set a to be the result of applying
the foregoing operation to a: a
+
= a a. Then 0
+
= 1, 1
+
= 2, and 2
+
= 3. Call a set S inductive if it
contains 0 and is closed under the successor operation. The set of all natural numbers can be dened by
the condition
x x belongs to every inductive set.
This has the desired consequence that itself is inductive, and is the smallest inductive set, i.e., S for
any inductive S.
One feature of constructing the natural numbers in this way is that the numbers are ordered by :
0 1 2 3 .... More precisely, the relation

= x, y) [ x y is a linear ordering, in fact a


well-ordering, of .
C. Other Number Systems
From the natural numbers we can construct the integers (positive and negative), the rational numbers,
and the real numbers. The guiding principle in each of the three cases is the same: take a suitable set of
names and divide out by the equivalence relation of naming the same number.
An integer can be named by giving a pair of natural numbers whose dierence is that integer. For
example, 3 = 2 5 = 6 9, so the pairs 2, 5) and 6, 9) can each serve as names of 3. We need an
equivalence relation R under which these two names are equivalent. More generally, we want
m, n)Rp, q) m+q = p +n.
The relation R dened in this way is an equivalence relation on . We dene the set Z of integers by
the equation
Z = ( )/R.
Thus Z is the quotient set formed from when we identify pairs that name the same integer.
For the rational numbers, similar ideas are used. A rational number can be named by a fraction, which
consists two integers, the numerator and the denominator (which is nonzero). For example, 1/3 = 2/6 =
3/(9), so the pairs 1, 3), 2, 6), and 3, 9) all name the same rational number. We dene the
appropriate equivalence relation S under which these pairs are equivalent by the equation
a, b)Sc, d) a d = c b.
Then S is an equivalence relation on Z Z

, where Z

is the set of nonzero integers. This relation enables


us to dene the set Q by the equation
Q = (Z Z

)/S.
Thus if we think of Z Z

as the set of fractions, then Q is the set of fractions modulo the equivalence
relation S.
Similar concepts apply to real numbers, except that now the names are innite. To specify a particular
real number, innitely many rational approximations are, in general, necessary. The idea is to name a real
number by means of a sequence of rational numbers that converge to it. We know that the convergent
sequences are exactly the Cauchy sequences, i.e., the sequences s such that for any positive rational ,
[s
m
s
n
[ < for all suciently large m and n. Let C be the set of Cauchy sequences of rational numbers.
Then we need to identify two Cauchy sequences that converge to the same limit. To accomplish this,
dene an equivalence relation E on C by specifying that two sequences r and s are equivalent if for every
positive rational , [r
n
s
n
[ < for all suciently large n. Then we dene the set R by the equation
R = C/E.
(This approach to constructing R is due to G. Cantor. Another approach using sets of rational numbers
instead of sequences is due to R. Dedekind.)
6
V. INFINITE SETS
A. Cardinal Numbers
All innite sets are big, but some are bigger than others. To make sense of this statement, we need to
clarify the concepts involved. For innite sets, what does it mean to say that they are or are not of the same
size?
The nite sets are easier to understand. So we want to take concepts that are understood in the nite
case, and generalize them to arbitrary sets. A set S is said to be nite if there is a natural number n telling
us how many members S has, i.e., some one-to-one function has domain n and range S (is a one-to-one
correspondence between n and S). This function counts the members of S: s
0
, s
1
, , s
n1
. And of course
S is said to be innite if there is no natural number with a one-to-one correspondence between it and S.
Dene A to be equinumerous to B (written A B) if there exists a one-to-one function f with domain
A and range B (i.e., if there is a one-to-one correspondence between A and B). (If this happens, then also
B A because f
1
has domain B and range A.) For example, the preceding paragraph species that a set
S is nite if some natural number is equinumerous to it. If 5 is equinumerous to S, then we know just how
big S is.
But the concept of equinumerosity applies to innite sets as well as nite. We will think of equinumerous
sets as having the same size, even if they are innite. For example, the set of natural numbers and the
set 0, 1, 4, 9, of squares of natural numbers have the same size (are equinumerous). (The latter set
appears smaller than the former only when we consider embedding the squares in the ordered sequence of
natural numberswhich is extraneous to considering the size of the two sets as abstract sets.) In fact the
set of natural numbers, the set of integers, and the set of rational numbers are all equinumerous. The set of
real numbers is equinumerous to the power set of the set of natural numbers.
But not all innite sets are equinumerous. Cantor proved in 1873 that the set of natural numbers is
not equinumerous to the set of real numbers. (His clever proof has become known as the Cantor diagonal
argument.) More generally, no set is equinumerous to its power set. (If f : S TS then x S [ x / f(x)
belongs to TS but not to the range of f; thus the range of f is not all of TS.) So innite sets come in
dierent sizes.
A set is said to be countable if it either is nite or is equinumerous to the set of natural numbers.
These are the sets that, although possibly innite, are not as innite as the uncountable sets. For example,
, Z, and Q are countable, whereas R and T are uncountable.
For a nite set S, we have a number to measure its size: the natural number equinumerous to S. These
measurements can be extended into the innite. To each set, we can assign a number known as the cardinal
number of that set. Two sets receive the same cardinal number if and only if they are equinumerous. In
other words, we assign the same number to sets of the same size, and dierent numbers to sets of dierent
size.
How? For a two-element set S, its cardinal number is 2. Now 2 is the set , , a canonically selected
set equinumerous to S. Analogously, the cardinal number of an innite set will be a certain canonically
selected equinumerous set. The details of the selection method will be discussed in the next section.
The cardinal number of is denoted
0
. This is the cardinal number of each of the countably innite
sets.
0
is the smallest innite cardinal number, under a natural concept of ordering.
B. Ordinal Numbers
We constructed the sequence 0, 1, 2, . . . of natural numbers from the bottom up, letting each natural
number be the set of preceding ones. We now want to generalize the ideas behind the construction of this
sequence, to obtain larger transnite numbers.
The numbers so obtained will be known as the ordinal numbers. They are constructed (together with
their ordering) according to the following three principles: (1) 0 is an ordinal number, and is the least ordinal
number. (2) For any ordinal number , the next ordinal number (i.e., the least ordinal number greater than
) is its successor
+
. (3) Any initial segment of the ordinal numbers, i.e., any set of ordinal numbers
containing all numbers smaller than each of its members (no gaps), is itself an ordinal number, and is the
least ordinal number greater than all of its members.
Lets see how these work. The rst two principles give us the natural numbers. Apply the third principle
to . Since is an initial segment of the ordinal numbers, is the least ordinal number greater than each
7
of the natural numbers. That is, is the least transnite ordinal. The next one is
+
and the next one
after that is
++
. Both of these are countable ordinal numbers; the set of all countable ordinal numbers is
the rst uncountable one. This sequence goes on forever, in a very strong sense of forever. Whenever you
have a set of ordinal numbers that seems at rst glance to contain all of them, then by principle (3) that
set itself is a new ordinal number, larger than any of its members. Consequently the collection of all ordinal
numbers will be a proper class, not a set.
(Actually principles (1) and (2) are redundant. For an ordinal , the set
+
is an initial segment of the
ordinals, and so is the next one by principle (3).)
There is a connection between ordinal numbers and the well-orderings mentioned in section IV.A. For
one thing, any ordinal number is well-ordered by in the following sense. Suppose is an ordinal number,
and let

be the relation x, y) [ x y . Then

is a well-ordering on .
Moreover, for any well-ordering <on a set A, there is a unique ordinal number that measures the length
of the well-ordering. That is, there is a unique ordinal number and a unique one-to-one correspondence f
between A and such that
x < y f(x)

f(y)
for all x and y in A. (Here f is said to be an isomorphism from A, <) onto ,

).) Any linear ordering


on a nite set must be a well-ordering. And any linear ordering on, say, a ve-element set has the ordinal
number 5 as its length. But there is more variety in the innite case. The set can be well-ordered in
dierent ways. There is the usual ordering of the natural numbers, whose length is the ordinal number
itself. Or we could, for example, put the even numbers rst (in their usual order) followed by the odd
numbers. This gives a well-ordering longer than the usual ordering.
Not only is each ordinal number well-ordered by , but the (proper) class of all ordinal numbers is, in a
sense, well-ordered by . The proper class of ordered pairs , ) [ and both are ordinal numbers
has all the properties of a well-ordering on the class of all ordinal numbers, except here we have proper
classes instead of sets. In particular, any nonempty set of ordinal numbers has a least element with respect
to the ordering .
Finally, let us return to the question of selecting appropriate cardinal numbers. For a nite set, say a
ve-element set, its cardinal number (which is 5) is the ordinal that measures the length of well-orderings
on the set. We can generalize this idea. For any set S, we dene the cardinal number of S to be the least
(with respect to ) ordinal number that measures the length of some well-ordering on S.
There is only one hitch. For a large set, such as R, how do we know there is any well-ordering on the
set? Certainly it is hard to think of any way to well-order R; the standard ordering on R fails to be a
well-ordering. But it can by done if we are very good at making choices. A well-ordering of R can be formed
from the bottom up, where at any stage we choose some real number (if any) not already incorporated into
the ordering, and make it the least number greater than those previously incorporated. This makes heavy
essential use of the axiom of choice; see section VI.B.
VI. AXIOMATIZATION
A. ZermeloFraenkel Axioms
The standard list of axioms for set theory is known as the ZermeloFraenkel (ZF) axiomatization.
(Zermelo gave the rst axiomatization for set theory, and later improvements were made by Fraenkel. Skolem
also contributed to the formulation of the axioms.) This article will describe the axioms informally, rather
than give exact statements of the axioms.
The extensionality axiom expresses the extensionality principle mentioned in section II.A. Several of the
other axioms say the certain basic set-theoretic operations are possible. The power-set axiom and the union
axiom assert the existence of the power set (of any set) and the union, respectively. These axioms are needed
to make the denitions of these two operation legitimate. The pairing axiom plays a similar role for x, y.
The innity axiom asserts the existence of ; without it one could not prove the existence of any innite
sets. The axiom of choice will be described in the next section. The axiom of regularity (or foundation) is
equivalent to the assertion that every set has some rank, i.e., that it appears eventually in some V

for some
ordinal number . Finally there are innitely many axioms saying that any meaningful expression written in
the language of set theory, under certain conditions, determines a set. Suppose (x, y) is such an expression
8
with the variables x and y (and possibly names for other sets). Then a replacement axiom states that if for
each x in some set S there is at most one value for y that makes (x, y) true, then y [ (x, y) for some
x in S is a legitimate set. (The idea is that the ranks of the various ys could not contain arbitrarily large
ordinal numbers, because S is merely a set whereas the ordinal numbers for a proper class.)
These axioms all are in accord with out informal conception of what sets ought to be like (although
objections have been raised about the axiom of choice). And they suce for the development of mathematics,
e.g. the theorems of algebra or analysis. In section VI.D we will consider the possibility of adding further
axioms.
B. Choice
The axiom of choice can be stated in a variety of equivalent ways. One such formulation is that for any
set o of nonempty sets, there exists a function f whose domain is o for which f(X) X for each X in o.
This axiom diers from such axioms as the union axiom or the power-set axiom in that it asserts the
existence of a set without specifying exactly what is in this set. For this reason, the axiom of choice had a
somewhat controversial history before ultimately being accepted.
Several of the standard results of mathematics require the axiom of choice for their proof. For example,
the proof that every vector space has a basis and the proof that every eld has an algebraic closure each
make essential use of the axiom of choice. Moreover, the axiom of choice is needed to prove that every set
has a well-ordering. In fact this property is equivalent to the axiom of choice.
The ZF axiomatization together with the axiom of choice is generally denoted ZFC. This explicit mention
of choice reects the controversial reputation the axiom has had.
C. Consistency and Incompleteness
It is reasonable to ask whether the ZFC axioms are strong enough. They suce for the development of
ordinary mathematics. But one might ask whether or not we have completeness: Is it the case that for
every sentence in the language of set theory, the axioms allow us either to prove or to refute ?
Here a theorem of logic, the Godel incompleteness theorem, comes into play. Applied to ZFC, this
theorem tells us that completeness cannot possibly hold for ZFC, unless ZFC is inconsistent. (An inconsis-
tent axiomatization would let us prove anything.) Moreover, Godels theorem tells us that the problem is
irreparable: adding new axioms in any reasonable way would not give us completeness.
And what about consistency? It would be reassuring to have a proof of the consistency of ZFC, at least
if that proof were carried out from assumptions that seemed safer than ZFC itself. Here Godels second
incompleteness theorem applies: there is no proof of the consistency of ZFC from ZFC itselflet alone from
weaker assumptionsunless in fact ZFC is inconsistent.
We have, nevertheless, some reasons for believing that ZFC is consistent. For one thing, ZFC seems to
state true facts from the viewpoint of our informal ideas about how sets ought to behave. In addition,
there is the empirical evidence that no contradiction has yet been found.
D. Stronger Axioms
The incompleteness of ZFC is an invitation to add further axioms. Because the universe of all sets is
constructed by iterating the power set axiom over the ordinal numbers, it is not surprising that two sorts of
additional possible axioms have been studied: those related to the power-set operation and those related to
the ordinal numbers.
An axiom that has been studied (and rejected) is the so-called axiom of constructibility, which has the
eect of restraining the power-set operation. It does have several interesting features. For one, the axiom of
choice follows from it. Moreover, the continuum hypothesis (which says that the cardinal number of the set
R of real numbers is the least uncountable cardinal number) follows from it. But the continuum hypothesis
is not generally accepted as a fact about R.
More interesting are the principles concerning the extent of the ordinal numbers. These are known as
large cardinal axioms. They have the form: there exists a cardinal number having a certain property,
where it is known that any such cardinal number must be very large. In fact the axiom of innity can be
regarded as the rst large cardinal axioms; it states that the sequence of ordinal numbers does not stop at
. The study of stronger large cardinal axioms is currently an area of active research.
9
BIBLIOGRAPHY
The books by Enderton, by Hrbacek and Jech, and by Vaught are basic textbooks on the subject.
The books by Jech and by Kunen are more advanced works. The book by Fraenkel et al. emphasizes the
philosophical and foundational aspects of set theory.
Aczel, Peter, (1988). Non-well-founded sets, Center for the Study of Language and Information, Stanford,
1988.
Enderton, Herbert B., Elements of Set Theory, Academic Press, New York, 1977.
Fraenkel, Abraham A., Yehoshua Bar-Hillel, and Azriel Levy, Foundations of Set Theory, 2nd ed., North-
Holland, Amsterdam, 1973.
Hrbacek, Karel, and Thomas Jech, Introduction to Set Theory, 2nd ed., Marcel Dekker, New York, 1984.
Jech, Thomas, Set Theory, Academic Press, New York, 1978.
Kunen, Kenneth, Set Theory, North-Holland, Amsterdam, 1980.
Vaught, Robert L., Set Theory, Birkh auser, Boston, 1985.
10

You might also like