v

To Monika, Agata, and Victoria

and to the memory of my Mother, who taught me mathematics

Contents
Preface ix
Part I Basics of set theory 1
1 Axiomatic set theory 3
1.1 Why axiomatic set theory? 3
1.2 The language and the basic axioms 6
2 Relations, functions, and Cartesian product 12
2.1 Relations and the axiom of choice 12
2.2 Functions and the replacement scheme axiom 16
2.3 Generalized union, intersection, and Cartesian product 19
2.4 Partial- and linear-order relations 21
3 Natural numbers, integers, and real numbers 25
3.1 Natural numbers 25
3.2 Integers and rational numbers 30
3.3 Real numbers 31
Part II Fundamental tools of set theory 35
4 Well orderings and transnite induction 37
4.1 Well-ordered sets and the axiom of foundation 37
4.2 Ordinal numbers 44
4.3 Denitions by transnite induction 49
4.4 Zorns lemma in algebra, analysis, and topology 54
5 Cardinal numbers 61
5.1 Cardinal numbers and the continuum hypothesis 61
5.2 Cardinal arithmetic 68
5.3 Conality 74
vii

viii Contents
Part III The power of recursive denitions 77
6 Subsets of R
n
79
6.1 Strange subsets of R
n
and the diagonalization argument 79
6.2 Closed sets and Borel sets 89
6.3 Lebesgue-measurable sets and sets with the Baire property 98
7 Strange real functions 104
7.1 Measurable and nonmeasurable functions 104
7.2 Darboux functions 106
7.3 Additive functions and Hamel bases 111
7.4 Symmetrically discontinuous functions 118
Part IV When induction is too short 127
8 Martins axiom 129
8.1 RasiowaSikorski lemma 129
8.2 Martins axiom 139
8.3 Suslin hypothesis and diamond principle 154
9 Forcing 164
9.1 Elements of logic and other forcing preliminaries 164
9.2 Forcing method and a model for CH 168
9.3 Model for CH and 182
9.4 Product lemma and Cohen model 189
9.5 Model for MA+CH 196
A Axioms of set theory 211
B Comments on the forcing method 215
C Notation 220
References 225
Index 229

Preface
The course presented in this text concentrates on the typical methods of
modern set theory: transnite induction, Zorns lemma, the continuum hy-
pothesis, Martins axiom, the diamond principle , and elements of forcing.
The choice of the topics and the way in which they are presented is subordi-
nate to one purpose to get the tools that are most useful in applications,
especially in abstract geometry, analysis, topology, and algebra. In par-
ticular, most of the methods presented in this course are accompanied by
many applications in abstract geometry, real analysis, and, in a few cases,
topology and algebra. Thus the text is dedicated to all readers that would
like to apply set-theoretic methods outside set theory.
The course is presented as a textbook that is appropriate for either a
lower-level graduate course or an advanced undergraduate course. How-
ever, the potential readership should also include mathematicians whose
expertise lies outside set theory but who would like to learn more about
modern set-theoretic techniques that might be applicable in their eld.
The reader of this text is assumed to have a good understanding of
abstract proving techniques, and of the basic geometric and topological
structure of the n-dimensional Euclidean space R
n
. In particular, a com-
fort in dealing with the continuous functions from R
n
into R is assumed. A
basic set-theoretic knowledge is also required. This includes a good under-
standing of the basic set operations (union, intersection, Cartesian product
of arbitrary families of sets, and dierence of two sets), abstract functions
(the operations of taking images and preimages of sets with respect to func-
tions), and elements of the theory of cardinal numbers (nite, countable,
and uncountable sets.) Most of this knowledge is included in any course in
analysis, topology, or algebra. These prerequisites are also discussed briey
in Part I of the text.
The book is organized as follows. Part I introduces the reader to ax-
iomatic set theory and uses it to develop basic set-theoretic concepts. In
particular, Chapter 1 contains the necessary background in logic, discusses
the most fundamental axioms of ZFC, and uses them to dene basic set-
ix

x Preface
theoretic operations. In Chapter 2 the notions of relation, function, and
Cartesian product are dened within the framework of ZFC theory. Related
notions are also introduced and their fundamental properties are discussed.
Chapter 3 describes the set-theoretic interpretation of the sets of natural
numbers, integers, and real numbers. Most of the facts presented in Part I
are left without proof.
Part II deals with the fundamental concepts of classical set theory.
The ordinal and cardinal numbers are introduced and their arithmetic is
developed. A theorem on denition by recursion is proved and used to
prove Zorns lemma. Section 4.4 contains some standard applications of
Zorns lemma in analysis, topology, and algebra.
Part III is designed to familiarize the reader with proofs by trans-
nite induction. In particular, Section 6.1 illustrates a typical transnite-
induction construction and the diagonalization argument by describing
several dierent constructions of subsets of R
n
with strange geometric prop-
erties. The two remaining sections of Chapter 6 introduce the basic ele-
ments of descriptive set theory and discuss Borel and Lebesgue-measurable
sets and sets with the Baire property. Chapter 7 is designed to help the
reader master the recursive-denitions technique. Most of the examples
presented there concern real functions, and in many cases consist of the
newest research results in this area.
Part IV is designed to introduce the tools of modern set theory: Mar-
tins axiom, the diamond principle , and the forcing method. The overall
idea behind their presentation is to introduce them as natural renements
of the method of transnite induction. Thus, based on the solid founda-
tion built in Part III, the forcing notions and forcing arguments presented
there are obtained as transformed transnite-induction arguments. In
particular, the more standard axiomatic approach to these methods is de-
scribed in Chapter 8, where Martins axiom and the diamond principle are
introduced and discussed. Chapter 9 is the most advanced part of this text
and describes the forcing method. Section 9.1 consists of some additional
prerequisites, mainly logical, necessary to follow the other sections. In Sec-
tion 9.2 the main theoretical basis for the forcing theory is introduced while
proving the consistency of ZFC and the negation of CH. In Section 9.3 is
constructed a generic model for ZFC+ (thus, also for ZFC+CH). Sec-
tion 9.4 discusses the product lemma and uses it to deduce a few more
properties of the Cohen model, that is, the model from Section 9.2. The
book nishes with Section 9.5 in which is proved the simultaneous con-
sistency of Martins axiom and the negation of the continuum hypothesis.
This proof, done by iterated forcing, shows that even in the world of the
sophisticated recursion method of forcing the classical recursion tech-
nique is still a fundamental method of set theory the desired model is
obtained by constructing forcing extensions by transnite induction.

Preface xi
It is also worthwhile to point out here that readers with dierent back-
grounds will certainly be interested in dierent parts of this text. Most
advanced graduate students as well as mathematical researchers using this
book will almost certainly just skim Part I. The same may be also true
for some of these readers for at least some portion of Part II. Part III and
the rst chapter of Part IV should be considered as the core of this text
and are written for the widest group of readers. Finally, the last chapter
(concerning forcing) is the most dicult and logic oriented, and will prob-
ably be of interest only to the most dedicated readers. It certainly can be
excluded from any undergraduate course based on this text.

Part I
Basics of set theory
1

Chapter 1
Axiomatic set theory
1.1 Why axiomatic set theory?
Essentially all mathematical theories deal with sets in one way or another.
In most cases, however, the use of set theory is limited to its basics: el-
ementary operations on sets, fundamental facts about functions, and, in
some cases, rudimentary elements of cardinal arithmetic. This basic part
of set theory is very intuitive and can be developed using only our good
intuition for what sets are. The theory of sets developed in that way is
called naive set theory, as opposed to axiomatic set theory, where all
properties of sets are deduced from a xed set of axioms.
Clearly the naive approach is very appealing. It allows us to prove a
lot of facts on sets in a quick and convincing way. Also, this was the way
the rst mathematicians studied sets, including Georg Cantor, a father
of set theory. However, modern set theory departed from the paradise
of this simple-minded approach, replacing it with axiomatic set theory,
the highly structured form of set theory. What was the reason for such a
replacement?
Intuitively, a set is any collection of all elements that satisfy a certain
given property. Thus, the following axiom schema of comprehension,
due to Frege (1893), seems to be very intuitive.
If is a property, then there exists a set Y = X: (X) of all
elements having property .
This principle, however, is false! It follows from the following theorem of
Russell (1903) known as Russells antinomy or Russells paradox.
3

4 1 Axiomatic set theory
Russells paradox There is no set S = X: X X.
The axiom schema of comprehension fails for the formula (X) dened
as X X. To see it, notice that if S had been a set we would have had
for every Y
Y S Y Y.
Substituting S for Y we obtain
S S S S,
which evidently is impossible.
This paradox, and other similar to it, convinced mathematicians that
we cannot rely on our intuition when dealing with abstract objects such as
arbitrary sets. To avoid this trouble, naive set theory has been replaced
with axiomatic set theory.
The task of nding one universal axiomatic system for set theory
that agrees with our intuition and is free of paradoxes was not easy, and
was not without some disagreement. Some of the disagreement still ex-
ists today. However, after almost a century of discussions, the set of ten
axioms/schemas, known as the ZermeloFraenkel axioms (abbreviated as
ZFC, where C stands for the axiom of choice), has been chosen as the most
natural. These axioms will be introduced and explained in the next chap-
ters. The full list of these axioms with some comments is also included in
Appendix A.
It should be pointed out here that the ZFC axioms are far from per-
fect. It could be expected that a perfect set of axioms should be com-
plete, that is, that for any statement expressed in the language of set
theory (which is described in the next section) either or its negation is a
consequence of the axioms. Also, a good set of axioms should certainly
be consistent, that is, should not lead to a contradiction. Unfortunately, we
cannot prove either of these properties for the ZFC axioms. More precisely,
we do believe that the ZFC axioms are consistent. However, if this belief
is correct, we cant prove it using the ZFC axioms alone. Does it mean
that we should search for a better system of set-theory axioms that would
be without such a deciency? Unfortunately, there is no use in searching
for it, since no reasonable set of axioms of set theory can prove its own
consistency. This follows from the following celebrated theorem of G odel.
Theorem 1.1.1 (G odels second incompleteness theorem) Let T be a set
of axioms expressed in a formal language L (such as the language of set
theory described in Section 1.2) and assume that T has the following rea-
sonable properties.
(1) T is consistent.

1.1 Why axiomatic set theory? 5
(2) There is an eective algorithm that decides for an arbitrary sentence
of the language L whether it is in T or not.
(3) T is complicated enough to encode simple arithmetic of the natural
numbers.
Then there is a sentence of the language L that encodes the statement
T is consistent. However, is not a consequence of the axioms T.
In other words, Theorem 1.1.1 shows us that for whatever reasonable
systems of axioms of set theory we choose, we will always have to rely on
our intuition for its consistency. Thus, the ZFC axioms are as good (or
bad) in this aspect as any other reasonable system of axioms.
So what about the completeness of the ZFC axioms? Can we prove
at least that much? The answer is again negative and once again it is a
common property for all reasonable systems of axioms, as follows from
another theorem of G odel.
Theorem 1.1.2 (G odels rst incompleteness theorem) Let T be a set of
axioms expressed in a formal language L (such as the language of set theory
described in Section 1.2) and assume that T has the following reasonable
properties.
(1) T is consistent.
(2) There is an eective algorithm that decides for an arbitrary sentence
of the language L whether it is in T or not.
(3) T is complicated enough to encode simple arithmetic of the natural
numbers.
Then there is a sentence of the language L such that neither nor its
negation can be deduced from the axioms T.
A sentence as in Theorem 1.1.2 is said to be independent of the
axioms T. It is not dicult to prove that a sentence is independent of
the consistent set of axioms T if and only if both T and T are
consistent too. Part of this course will be devoted to studying the sentences
of set theory that are independent of the ZFC axioms.
The preceding discussion shows that there is no way to nd a good
complete set of axioms for set theory. On the other hand, we can nd a
set of axioms that reach far enough to allow encoding of all set-theoretic
operations and all classical mathematical structures. Indeed, the ZFC ax-
ioms do satisfy this requirement, and the rest of Part I will be devoted to
describing such encodings of all structures of interest.

6 1 Axiomatic set theory
1.2 The language and the basic axioms
Any mathematical theory must begin with undened concepts. In the case
of set theory these concepts are the notion of a set and the relation is
an element of between the sets. In particular, we write x y for x is
an element of y.
The relation is primitive for set theory, that is, we do not dene it.
All other objects, including the notion of a set, are described by the axioms.
In particular, all objects considered in formal set theory are sets. (Thus,
the word set is superuous.)
In order to talk about any formal set theory it is necessary to specify
rst the language that we will use and in which we will express the axioms.
The correct expressions in our language are called formulas. The basic for-
mulas are x y and z = t, where x, y, z, and t (or some other variable
symbols) stand for the sets. We can combine these expressions using the
basic logical connectors of negation , conjunction &, disjunction , impli-
cation , and equivalence . Thus, for example, means not and
stands for implies . In addition, we will use two quantiers:
existential and universal . Thus, an expression x is interpreted as
for all x formula holds. In addition, the parentheses ( and ) are
used, when appropriate.
Formally, the formulas can be created only as just described. However,
for convenience, we will also often use some shortcuts. For example, an ex-
pression x A(x) will be used as an abbreviation for x(x A & (x)),
and we will write x A(x) to abbreviate the formula x(x A(x)).
Also we will use the shortcuts x = y, x y, x y, and x y, where, for
example, x y stands for z(z xz y).
Finally, only variables, the relations = and , and logical symbols al-
ready mentioned are allowed in formal formulas. However, we will often
use some other constants. For example, we will write x = (x is an empty
set) in place of y(y x).
We will discuss ZFC axioms throughout the next few sections as they
are needed. Also, in most cases, we wont write in the main text the
formulas representing the axioms. However, the full list of ZFC axioms
together with the formulas can be found in Appendix A.
Let us start with the two most basic axioms.
Set existence axiom There exists a set: x(x = x).
Extensionality axiom If x and y have the same elements, then x is equal
to y.
The set existence axiom follows from the others. However, it is the
most basic of all the axioms, since it ensures that set theory is not a trivial

1.2 The language and the basic axioms 7
theory. The extensionality axiom tells us that the sets can be distinguish
only by their elements.
Comprehension scheme (or schema of separation) For every for-
mula (s, t) with free variables s and t, set x, and parameter p there
exists a set y = u x: (u, p) that contains all those u x that
have the property .
Notice that the comprehension scheme is, in fact, the scheme for in-
nitely many axioms, one for each formula . It is a weaker version
of Freges axiom schema of comprehension. However, the contradiction
of Russells paradox can be avoided, since the elements of the new set
y = u x: (u, p) are chosen from a xed set x, rather than from an
undened object such as the class of all sets.
From the set existence axiom and the comprehension scheme used with
the formula u = u, we can conclude the following stronger version of the
set existence axiom.
Empty set axiom There exists the empty set .
To see the implication, simply dene = y x: y = y, where x is a
set from the set existence axiom. Notice that by the extensionality axiom
the empty set is unique.
An interesting consequence of the comprehension scheme axiom is the
following theorem.
Theorem 1.2.1 There is no set of all sets.
Proof If there were a set S of all sets then the following set
Z = X S: X X
would exist by the comprehension scheme axiom. However, with S being
the set of all sets, we would have that Z = X: X X, the set from
Russells paradox. This contradiction shows that the set S of all sets cannot
exist.
By the previous theorem all sets do not form a set. However, we some-
times like to talk about this object. In such a case we will talk about a
class of sets or the set-theoretic universe. We will talk about classes only
on an intuitive level. It is worth mentioning, however, that the theory of
classes can also be formalized similarly to the theory of sets. This, however,
is far beyond the scope of this course. Let us mention only that there are
other proper classes of sets (i.e., classes that are not sets) that are strictly
smaller than the class of all sets.

8 1 Axiomatic set theory
The comprehension scheme axiom is a conditional existence axiom, that
is, it describes how to obtain a set (subset) from another set. Other basic
conditional existence axioms are listed here.
Pairing axiom For any a and b there exists a set x that contains a and
b.
Union axiom For every family T there exists a set U containing the union

T of all elements of T.
Power set axiom For every set X there exists a set P containing the set
{(X) (the power set) of all subsets of X.
In particular, the pairing axiom states that for any a and b there exists
a set x such that a, b x. Although it does not state directly that there
exists a set a, b, the existence of this set can easily be concluded from
the existence of x and the comprehension scheme axiom:
a, b = u x: u = a u = b.
Similarly, we can conclude from the union and power set axioms that for
every sets T and X there exist the union of T

T = x: F T (x F) = x U : F T (x F)
and the power set of X
{(X) = z : z X = z P : z X.
It is also easy to see that these sets are dened uniquely. Notice also that
the existence of a set a, b implies the existence of a singleton set a,
since a = a, a.
The other basic operations on sets can be dened as follows: the union
of two sets x and y by
x y =

x, y;
the dierence of sets x and y by
x ` y = z x: z y;
the arbitrary intersections of a family T by

T =

T: F T (z F)

;
and the intersections of sets x and y by
x y =

x, y.

1.2 The language and the basic axioms 9
The existence of sets x`y and

T follows from the axiom of comprehension.

We will also sometimes use the operation of symmetric dierence of two
sets, dened by
xy = (x ` y) (y ` x).
Its basic properties are listed in the next theorem. Its proof is left as an
exercise.
Theorem 1.2.2 For every x, y, and z
(a) xy = yx,
(b) xy = (x y) ` (x y),
(c) (xy)z = x(yz).
We will dene an ordered pair 'a, b` for arbitrary a and b by
'a, b` = a, a, b. (1.1)
It is dicult to claim that this denition is natural. However, it is com-
monly accepted in modern set theory, and the next theorem justies it by
showing that it maintains the intuitive properties we usually associate with
the ordered pair.
Theorem 1.2.3 For arbitrary a, b, c, and d
'a, b` = 'c, d` if and only if a = c and b = d.
Proof Implication is obvious.
To see the other implication, assume that 'a, b` = 'c, d`. This means
that a, a, b = c, c, d. In particular, by the axiom of extension-
ality, a is equal to either c or c, d.
If a = c then a = c. If a = c, d, then c must belong to a and
we also conclude that a = c. In any case, a = c and we can deduce that
a, a, b = a, a, d. We wish to show that this implies b = d.
But a, b belongs to a, a, d. Thus we have two cases.
Case 1: a, b = a, d. Then b = a or b = d. If b = d we are done. If
b = a then a, b = a and so a, d = a. But d belongs then to a
and so d = a. Since we had also a = b we conclude b = d.
Case 2: a, b = a. Then b belongs to a and so b = a. Hence we
conclude that a, d = a, and as in case 1 we can conclude that b = d.

10 1 Axiomatic set theory
Now we can dene an ordered triple 'a, b, c` by identifying it with
''a, b`, c` and, in general, an ordered n-tuple by
'a
1
, a
2
, . . . , a
n1
, a
n
` = ''a
1
, a
2
, . . . , a
n1
`, a
n
`.
The agreement of this denition with our intuition is given by the following
theorem, presented without proof.
Theorem 1.2.4 'a
1
, a
2
, . . . , a
n1
, a
n
` = 'a
t
1
, a
t
2
, . . . , a
t
n1
, a
t
n
` if and only
if a
i
= a
t
i
for all i = 1, 2, . . . , n.
Next we will dene a Cartesian product X Y as the set of all ordered
pairs 'x, y` such that x X and y Y . To make this denition formal,
we have to use the comprehension axiom. For this, notice that for every
x X and y Y we have
'x, y` = x, x, y {({(X Y )).
Hence, we can dene
X Y = z {({(X Y )): x X y Y (z = 'x, y`). (1.2)
The basic properties of the Cartesian product and its relation to other
set-theoretic operations are described in the exercises.
The last axiom we would like to discuss in this section is the innity
axiom. It states that there exists at least one innite set. This is the
only axiom that implies the existence of an innite object. Without it,
the family T of all nite subsets of the set of natural numbers would be a
good model of set theory, that is, T satises all the axioms of set theory
except the innity axiom.
To make the statements of the innity axiom more readable we intro-
duce the following abbreviation. We say that y is a successor of x and write
y = S(x) if y = x x, that is, when
z[z y (z x z = x)].
Innity axiom (Zermelo 1908) There exists an innite set (of some spe-
cial form):
x[z(z = z x) & y xz(z = S(y)z x)] .
Notice that the innity axiom obviously implies the set existence axiom.

1.2 The language and the basic axioms 11
EXERCISES
1 Prove that if F T then

T F

T.
2 Show that for every family T and every set A
(a) if A F for every F T then A

T, and
(b) if F A for every F T then

T A.
3 Prove that if T ( = then

T

(

(T (). Give examples
showing that the inclusion cannot be replaced by equality and that the
assumption T ( = is essential.
4 Prove Theorem 1.2.2.
5 Show that ''a, b`, c` = ''a
t
, b
t
`, c
t
` if and only if 'a, 'b, c`` = 'a
t
, 'b
t
, c
t
``
if and only if a = a
t
, b = b
t
, and c = c
t
. Conclude that we could dene an
ordered triple 'a, b, c` as 'a, 'b, c`` instead of ''a, b`, c`.
6 Prove that X Y = if and only if X = or Y = .
7 Show that for arbitrary sets X, Y, and Z the following holds.
(a) (X Y ) Z = (X Z) (Y Z).
(b) (X Y ) Z = (X Z) (Y Z).
(c) (X ` Y ) Z = (X Z) ` (Y Z).
8 Prove that if X Z Y T and X Z = then X Y and Z T.
Give an example showing that the assumption X Z = is essential.

Chapter 2
Relations, functions, and
Cartesian product
2.1 Relations and the axiom of choice
A subset R of a Cartesian product X Y is called a (binary) relation.
For a relation R of a Cartesian product X Y , we usually write aRb
instead of 'a, b` R and read: a is in the relation R to b, or the relation
R between a and b holds.
The domain dom(R) of a relation R is dened as the set of all x such
that 'x, y` R for some y Y , that is,
dom(R) = x X: y Y ('x, y` R);
the range range(R) of a relation R is dened as the set of all y such that
'x, y` R for some x X, that is,
range(R) = y Y : x X ('x, y` R).
The set Z = dom(R) range(R) for relation R is called a eld of R.
Notice that R Z Z. In this case we often say that R is dened on Z.
Examples 1. The relation R
<
on the set of real numbers R dened as
'x, y` R
<
if and only if x < y is usually denoted by <.
1
Notice that
1
In the examples we will often use notions that you supposedly know from other courses,
even if we have not yet dened them within the framework of set theory (such as the set
of real numbers R in this example). This will be used only to help you develop the right
intuition. We will try to avoid this kind of situation in the main stream of the course.
12

2.1 Relations and the axiom of choice 13
R
<
is the subset of the plane consisting of those points that are above the
identity line y = x. The domain and the range of this relation are equal
to R.
2. Consider the relation R
div
on the set of natural numbers N dened as
'x, y` R
div
if and only if x and y are dierent natural numbers greater
than 1 and x divides y. Then dom(R
div
) = 2, 3, 4, . . . and range(R
div
) is
the set of all composite natural numbers.
Let R X Y and S Y Z. The relation
'y, x` Y X: xRy
is called the inverse of R and is denoted by R
1
. The relation
'x, z` X Z: y Y (xRy & ySz)
is called the composition of R and S and is denoted by S R.
Note that (R
1
)
1
= R, dom(R
1
) = range(R), range(R
1
) = dom(R),
dom(SR) dom(R), and range(SR) range(S). Moreover, (SR)
1
=
R
1
S
1
.
Examples 1. If and are dened on R in the natural way then
()
1
= .
2. () () is equal to the relation R R.
3. () () is equal to .
Let R be a binary relation on X X. We say that R is
reexive if xRx for every x X;
symmetric if xRy implies yRx for every x, y X; and
transitive if xRy and yRz imply xRz for every x, y, z X.
Examples 1. The relations < and > on R are transitive, but they are
neither reexive nor symmetric.
2. The relations and on R are transitive and reexive, but not sym-
metric.
3. The relation = on R is symmetric, but is neither reexive nor transitive.
4. The relation RR is symmetric and transitive, but is not reexive.
5. The relation R on R dened by xRy if and only if y = x
2
has neither of
the three properties.
6. The relation = on R is reexive, symmetric, and transitive.

14 2 Relations, functions, and Cartesian product
A binary relation R X X is an equivalence relation on X if it
is reexive, symmetric, and transitive. Equivalence relations are usually
denoted by symbols such as , , or .
Examples 1. The relation = on any set X is an equivalence relation.
2. If L is the family of all straight lines on the plane and R is the relation
on L of being parallel, then R is an equivalence relation.
3. Let C be the family of all Cauchy sequences 'a
n
` = 'a
1
, a
2
, . . .` of
rational numbers. Dene relation on C by 'a
n
` 'b
n
` if and only if
lim
n
(a
n
b
n
) = 0. Then is an equivalence relation.
4. A nonempty family 1 {(X) is said to be an ideal on X if for every
A, B {(X)
A B &B 1 A 1 and A, B 1 A B 1.
For every ideal 1 on a set {(X) the relation
ArB if and only if AB 1
is an equivalence relation.
Let X be a set. A family T of nonempty subsets of X is said to be a
partition of X if

T = X and sets belonging to T are pairwise disjoint,
that is, for every A, B T either A = B or A B = .
For an equivalence relation E on X and x X the set
[x] = y X: xEy
is called the equivalence class of x (with respect to E). The family
[x] {(X): x X
of all equivalence classes for E is denoted by X/E and is called the quotient
class of X with respect to E.
For a partition T of a set X let us dene a relation R
J
on X by
xR
J
y F T (x F & y F).
Theorem 2.1.1
(A) If E is an equivalence relation on X then the family X/E of all
equivalence classes forms a partition of X, that is, for every x, y X
either [x] = [y] or [x] [y] = .
(B) If T is a partition of X then R
J
is an equivalence relation on X.
Moreover, X/R
J
= T and R
X/E
= E.

2.1 Relations and the axiom of choice 15
The proof is left as an exercise.
Theorem 2.1.1 shows that there is a one-to-one correspondence between
the class of all partitions of X and the family of all equivalence relations
on X. In general, equivalence classes generalize the notion of equality.
Elements within one equivalence class are identied by the relation.
An element x of an equivalence class C is called a representative of C.
A set of representatives of an equivalence relation E is a set that contains
exactly one element in common with each equivalence class. Notice that,
by Theorem 2.1.1 and the axiom of choice, to be stated shortly, a set of
representatives exists for every equivalence relation.
Axiom of choice For every family T of nonempty disjoint sets there ex-
ists a selector, that is, a set S that intersects every F T in precisely
one point.
The axiom of choice (usually abbreviated as AC) has the conditional
existence character of the pairing, union, and power set axioms. However,
it has also a very dierent character, since a selector, which exists by this
axiom, does not have to be unique. This nonconstructive character of the
axiom of choice was, in the past, the reason that some mathematicians
(including Borel and Lebesgue) did not like to accept it. However, the
discussion on the validity of the axiom of choice has been for the most part
resolved today, in favor of accepting it.
The axiom of choice will be one of the most important tools in this
course.
EXERCISES
1 Let R XY and S Y Z be the relations. Prove that dom(R
1
) =
range(R), dom(S R) dom(R), and range(S R) range(S).
2 Show the formulas (RS)
1
= R
1
S
1
, (RS)
1
= R
1
S
1
, and
(S R)
1
= R
1
S
1
.
3 Prove the formulas
(R S) T = (R T) (S T), T (R S) = (T R) (T S),
(R S) T (R T) (S T), T (R S) (T R) (T S).
Find examples of relations R, S, and T that show that the inclusions in
the display cannot be replaced with equations.
4 Find examples of relations on R that are
(a) reexive, but neither symmetric nor transitive;
(b) reexive and symmetric but not transitive.
5 Prove Theorem 2.1.1.

16 2 Relations, functions, and Cartesian product
2.2 Functions and the replacement scheme
axiom
A relation R X Y is called a function if
(x, y
1
, y
2
)(xRy
1
& xRy
2
y
1
= y
2
).
Functions are usually denoted by the letters f, g, h, . . . .The domain dom(f)
and the range range(f) of a function f are dened as for relations. For a
function f if dom(f) = X and range(f) Y then f is called a function
(or map or transformation) from X into Y and it is denoted by f : X
Y . If, moreover, range(f) = Y then f is said to be a function (map or
transformation) from X onto Y , or a surjective function. The set of all
functions from X into Y is denoted by Y
X
.
If f Y
X
and x X then there exists precisely one y Y such that
xfy. The element y is called the value of f at x and is denoted by f(x).
Thus, the formula y = f(x) has the same meaning as xfy.
Notice that for f, g Y
X
f = g x X (f(x) = g(x)).
A function f : X Y is a one-to-one (or injective) function if
f(x) = f(y) x = y
for all x, y X. A function f : X Y is a bijection if it is one-to-one and
onto Y .
For f : X Y , A X, and B Y we dene
f[A] = f(x): x X = y Y : x X (y = f(x))
and
f
1
(B) = x X: f(x) B.
We use square brackets in f[A] rather than regular parentheses to avoid
a double meaning for the symbol f(A) when A is at the same time an
element of X and its subset. A similar double meaning may happen when
the symbol f
1
(B) is used. However, in this case it will be always clear
from the context which meaning of the symbol we have in mind. The sets
f[A] and f
1
(B) are called the image of A and the preimage of B with
respect to f, respectively.

2.2 Functions and the replacement scheme axiom 17
Theorem 2.2.1 If f Y
X
then f
1
is a function if and only if f is
one-to-one. Moreover, if f
1
is a function then f
1
X
Z
, where Z =
range(f), and f
1
is one-to-one.
Theorem 2.2.2 If f Y
X
and g Z
Y
then g f is also a function and
g f Z
X
. Moreover, (g f)(x) = g(f(x)) for every x X.
Theorem 2.2.3 Let f Y
X
and g Z
Y
. If f and g are one-to-one, then
so is g f. If f and g are onto, then so is g f.
The proofs are left as exercises.
For the proof of the next theorem we need one more axiom scheme.
Replacement scheme axiom (Fraenkel 1922; Skolem 1922) For every
formula (s, t, U, w) with free variables s, t, U, and w, set A, and
parameter p if (s, t, A, p) denes a function F on A by F(x) =
y (x, y, A, p), then there exists a set Y containing the range F[A]
of the function F, where F[A] = F(x): x A.
As with the comprehension scheme axiom, the replacement scheme
axiom is in fact a scheme for innitely many axioms, one for each for-
mula . In conjunction with the comprehension scheme axiom, the re-
placement scheme axiom implies that for a function dened by formula
F(x) = y (x, y, A, p) on a set A, the range of F exists, since it can be
dened:
F[A] = y Y : x A (y = F(x)).
Sometimes the replacement scheme axiom is formulated in a stronger ver-
sion, which states that the set Y existing by the axiom is equal to the range
of F. It is worth noticing that such a stronger version of the replacement
scheme axiom implies the comprehension scheme axiom.
Now we are ready for the next theorem.
Theorem 2.2.4 If T is a family of nonempty sets then there is a function
f : T

T such that f(X) X for every X T.

Proof Take a formula (X, Y ) that can be abbreviated to
Y = X X.
Notice that if (X, Y ) and (X, Y
t
) hold then Y = X X = Y
t
. In
particular, satises the assumptions of the replacement scheme axiom.
So there exists a family

18 2 Relations, functions, and Cartesian product
( = Y : X T ((X, Y )) = X X: X T.
Notice that elements of ( are nonempty and that they are pairwise disjoint,
since
(X X) (X
t
X
t
) = (X X
t
) (X X
t
).
Thus, by the axiom of choice, there is a selector f for (. Notice that f is
the desired function.
The function from Theorem 2.2.4 is called a choice function for the
family T. In fact, the statement that for every family of nonempty sets
there exists a choice function is equivalent to the axiom of choice. The
statement follows from the axiom of choice by Theorem 2.2.4. To see that
the axiom of choice follows from it dene a selector S for a family T of
nonempty pairwise-disjoint sets as S = range(f) when f is a choice function
for T.
For a function f Y
X
and A X the relation f (AY ) is a function
that belongs to Y
A
. It is called the restriction of f to the set A and is
denoted by f[
A
.
A function g is said to be an extension of a function f if f g. We
also say that f is a restriction of g.
Theorem 2.2.5 If f and g are functions then g extends f if and only if
dom(f) dom(g) and f(x) = g(x) for every x dom(f).
The proof is left as an exercise.
Functions for which the domain is the set N of natural numbers are
called (innite) sequences. A (nite) sequence is any function whose do-
main is any natural number n (= 0, 1, . . . , n 1). If a is any sequence,
then we usually write a
m
in place of a(m). We will sometimes represent
such a sequence by the symbol 'a
m
` or a
m
, indicating the domain of the
sequence if necessary.
EXERCISES
1 Prove Theorem 2.2.1.
2 Prove Theorem 2.2.2.
3 Prove Theorem 2.2.3.
4 Prove Theorem 2.2.5.

2.3 Generalized union, intersection, and Cartesian product 19
2.3 Generalized union, intersection, and Car-
tesian product
Let F be a function from a nonempty set T into a family of sets (. We will
usually write F
t
instead of F(t) and F
t
: t T or F
t

tT
for range(F).
The family T = F
t

tT
is called an indexed family, with T being the index
set.
The following notation will be used for T = F
t

tT
:

tT
F
t
=

T,

tT
F
t
=

T.
In the text these sets will also appear as

tT
F
t
and

tT
F
t
, respectively.
When the set T is xed, we will sometimes abbreviate this notation and
write F
t

t
or F
t
in place of F
t

tT
, and

t
F
t
and

t
F
t
for

tT
F
t
and

tT
F
t
. When the index set of the index family T is a Cartesian product
S T then we usually denote its elements by F
st
instead of F
s,t)
and we
say that T is a doubly indexed family. We denote it as F
st
: s S, t T
or simply F
st
. In such a case we will write

sS,tT
F
st
or

s,t
F
st
for

s,t)ST
F
st
and

sS,tT
F
st
or

s,t
F
st
for

s,t)ST
F
st
. If S = T
we will also write

s,tT
F
st
and

s,tT
F
st
.
The following properties are easy to verify:
x

tT
F
t
t T (x F
t
) and x

tT
F
t
t T (x F
t
);

tT
F
t
= F
p
=

tT
F
t
for T = p;

tT
F
t
= F
p
F
q
and

tT
F
t
= F
p
F
q
for T = p, q.
It is also easy to see that for any formula (t, x) and sets T, X

tT
x X: (t, x) = x X: t T ((t, x))

tT
x X: (t, x) = x X: t T ((t, x)).
Other properties of these operations are listed in the exercises.
The behavior of generalized union and intersection under the action of
image and preimage of a function is described in the next theorem.

20 2 Relations, functions, and Cartesian product
Theorem 2.3.1 If f Y
X
, F
t

tT
is an indexed family of subsets of X,
G
t

tT
is an indexed family of subsets of Y , A, B X, and C, D Y
then
(a) f

tT
F
t

tT
f[F
t
];
(b) f

tT
F
t

tT
f[F
t
];
(c) f[A] ` f[B] f[A` B];
(d) f
1

tT
G
t

tT
f
1
(G
t
);
(e) f
1

tT
G
t

tT
f
1
(G
t
);
(f) f
1
(A) ` f
1
(B) = f
1
(A` B).
The proof is left as an exercise.
For an indexed family F
t
: t T we dene its Cartesian product by

tT
F
t
= h Z
T
: t T (h(t) F
t
),
where Z =

tT
F
t
. If all the sets F
t
are identical, F
t
= Y , then

tT
F
t
=
Y
T
. The set Y
T
is called a Cartesian power of the set Y .
For t T the function p
t
:

tT
F
t
F
t
dened by p
t
(x) = x(t) is
called the projection of

tT
F
t
onto F
t
.
Remark For T a two-element set a, b the Cartesian products

tT
F
t
and F
a
F
b
are dierent. (The rst one is a set of functions on T, the second
one is a set of ordered pairs.) However, there is a natural identication of
every element 'a, x`, 'b, y` from

tT
F
t
with 'x, y` F
a
F
b
. Therefore
we will usually identify these products.
Theorem 2.3.2 A product

tT
F
t
of nonempty sets F
t
is nonempty.
Proof The choice function f for the family F
t
: t T, which exists by
Theorem 2.2.4, is an element of

tT
F
t
.
Remark Theorem 2.3.2 easily implies Theorem 2.2.4. Thus its statement
is equivalent to the axiom of choice. However, we do not need the axiom
of choice to prove Theorem 2.3.2 if either T is nite or

tT
F
t
= Y
T
.

2.4 Partial- and linear-order relations 21
EXERCISES
1 Prove that for every indexed families F
t

tT
and G
t

tT
(a)

tT
(F
t
G
t
) =

tT
F
t

tT
G
t
,
(b)

tT
(F
t
G
t
) =

tT
F
t

tT
G
t
,
(c)

tT
F
t

tT
G
t
=

s,tT
(F
s
G
t
)

tT
(F
t
G
t
), and
(d)

tT
F
t

tT
G
t
=

s,tT
(F
s
G
t
)

tT
(F
t
G
t
).
Give examples showing that the inclusions cannot be replaced by equalities.
2 Show that for every indexed family F
t

tT
and every set A
(a) A`

tT
F
t
=

tT
(A` F
t
) and A`

tT
F
t
=

tT
(A` F
t
),
(b)

tT
(A F
t
) = A

tT
F
t
and

tT
(A F
t
) = A

tT
F
t
.
3 Prove that for every indexed families F
t

tT
and G
t

tT
(a)

tT
F
t

tT
G
t

s,tT
(F
s
G
t
),
(b)

tT
F
t

tT
G
t

s,tT
(F
s
G
t
).
4 Show that

sS

tT
F
st


tT

sS
F
st
for every doubly indexed
family F
st
: s S, t T.
5 Prove Theorem 2.3.1(a), (b), and (c). Show, by giving examples, that
the inclusions in parts (b) and (c) cannot be replaced by equality.
6 Prove Theorem 2.3.1(d), (e), and (f).
7 For r, s R let A
r
= [r, r +1] and B
rs
= [r, s). Calculate

s0

rs
A
r
,

s0

rs
A
r
,

r0

s>r
B
rs
, and

r0

s>r
B
rs
.
2.4 Partial- and linear-order relations
A binary relation R on X is antisymmetric if
xRy & yRx x = y
for every x, y X.
A relation R on X is a (partial-)order relation if it is reexive, transitive,
and antisymmetric. Order relations are usually denoted by the symbols ,

22 2 Relations, functions, and Cartesian product
_, or <. If is a partial-order relation then the ordered pair 'X, ` is
called a partially ordered set (abbreviated also as poset).
Examples 1. The relations and on R are order relations.
2. For any set X the relation is an order relation on {(X).
3. The relation [ on the set 2, 3, 4, . . . dened by
n[m if and only if n divides m
is an order relation.
An element m X of an ordered set 'X, ` is minimal if for every
x X the condition x m implies x = m. Similarly, an element M X
is maximal if for every x X the condition x M implies x = M.
An element m X is the smallest element (least element or rst ele-
ment) in X if m x for every x X, and M X is the greatest element
(largest element or last element) in X if x M for every x X.
Theorem 2.4.1 Let 'X, ` be a partially ordered set.
(a) X can have at most one greatest and one smallest element.
(b) The smallest element of X, if it exists, is the only minimal element
of X.
(c) The greatest element of X, if it exists, is the only maximal element
of X.
Proof (a) If a and b are the smallest elements of X then a b and b a.
Hence a = b. The argument for the greatest element is the same.
(b) If a is the smallest element of X then it is minimal, since the condi-
tion x a combined with a x, which is true for every x, implies x = a.
Moreover, if m is minimal, then a m, since a is the smallest element,
and, by minimality of m, m = a.
(c) The argument is the same as in (b).
Examples 1. 'R, ` has neither minimal nor maximal elements.
2. '[0, 1], ` has 0 as the least element and 1 as the last element.
3. is the smallest element of '{(X), `. X is the greatest element of
'{(X), `.
4. Let '2, 3, 4, . . . , [` be dened as before. It does not have any maximal
element. A number m is minimal in this order if and only if m is a prime
number. (Thus '2, 3, 4, . . . , [` has innitely many minimal elements!)

2.4 Partial- and linear-order relations 23
A relation R on X is connected if
xRy yRx
for every x, y X. An order relation is called a linear-order relation if
it is connected. In this case we also say that 'X, ` (or just X) is linearly
ordered.
Examples 1. 'R, ` is linearly ordered.
2. '{(X), ` is linearly ordered if and only if X has at most one element.
3. '2, 3, 4, . . . , [` is not linearly ordered, since neither 2[3 nor 3[2.
4. The relation on N is a linear-order relation.
Theorem 2.4.2 If 'X, ` is linearly ordered then every minimal element
in X is the smallest element and every maximal element in X is the greatest
element. In particular, linearly ordered sets can have at most one maximal
element and at most one minimal element.
The proof is left as an exercise.
Notice that if is an order relation on a set X and Y X then the
relation (Y Y ) is an order relation on Y . It is called the restriction
of to Y . We often write 'Y, ` in place of 'Y, (Y Y )`.
Notice also that a subset of a partially ordered set is partially ordered
and a subset of a linearly ordered set is linearly ordered.
In general, for any partial order denoted by we will write for ()
1
and dene relations < and > by
x < y x y & x = y
and
x > y x y & x = y.
EXERCISES
1 Prove that the restriction (Y Y ) of an order relation on X is
an order relation, provided Y X. Show that 'Y, ` is linearly ordered
if 'X, ` is linearly ordered. Give an example such that 'Y, ` is linearly
ordered, while 'X, ` is not.
2 Find all minimal, maximal, greatest, and smallest elements of 'T, `,
where T = X N: X is nite.
3 Prove Theorem 2.4.2.

24 2 Relations, functions, and Cartesian product
4 A binary relation R X X is said to be a preorder relation if it is
transitive and reexive.
Let _ X X be a preorder relation.
(a) Show that the relation on X dened by
x y x _ y & y _ x
is an equivalence relation on X.
(b) Dene the relation on the family X/ of all equivalence classes for
by
[x] [y] x _ y.
Show that the relation is well dened and that it is a partial-order
relation.

Chapter 3
Natural numbers,
integers, and real numbers
From the results of Section 1.2 it is clear that sets such as , , , ,
, and so forth exist. Using the axiom of innity we can also conclude
that we can build similar innite sets. But how do we construct complicated
sets, such as the sets of natural and real numbers, if the only tools we have
to build them are the empty set and braces ? We will solve this
problem by identifying the aforementioned objects with some specic sets.
Briey, we will identify the number 0 with the empty set , and the
other natural numbers will be constructed by induction, identifying n with
0, . . . , n 1. The existence of the set N of all natural numbers is guar-
anteed by the innity axiom. The real numbers from the interval [0, 1] will
be identied with the set of functions 0, 1
N
, where an innite sequence
a: N 0, 1 is identied with the binary expansion of a number, that is,
with

nN
a(n)/2
n+1
. The details of these constructions are described in
the rest of this chapter.
3.1 Natural numbers
In this section we will nd a set that represents the set N of natural numbers
in our set-theoretic universe. For this, we need to nd for each natural
number n a set that represents it. Moreover, we will have to show that the
class of all such dened natural numbers forms a set.
When picking the natural numbers, we will have to pick also the order-
ing relation < between them. Essentially, the only relation that we have
available for this purpose is the relation . Thus, we will choose the natural
numbers to satisfy the following principle:
25

26 3 Natural numbers, integers, and real numbers
m < n if and only if m n.
Also, since natural numbers are going to be distinguished by the relation
of being an element, it seems to be natural to have the following intuitive
principle:
Each natural number n should have n elements.
These two principles give us no choice about our denition.
By the second principle we have 0 = .
Now suppose that we have already dened n and want to dene n + 1.
Since n < n + 1 we have n n + 1. Also, for every m n we have m < n,
so m < n+1 and m n+1. In particular, n n+1 and nn n+1.
But n has n elements, so n n has n + 1 elements, since n n, as
n < n. Therefore nn n+1 and both sets have n+1 elements. Thus,
n + 1 = n n.
By the foregoing discussion we see that 0 = , 1 = 0 = , 2 =
0, 1 = , , 3 = 0, 1, 2 = , , , , and so forth.
Now, showing that the class of all such numbers forms a set is another
problem. We will use for this the axiom of innity. We will also make
sure that the following three principles are satised, where S(n) = n + 1
is known as a successor of n and P stands for some property of natural
numbers.
P1 0 = S(n) for every n N.
P2 If S(n) = S(m) then n = m for every n, m N.
P3 If 0 has property P and for every n N
S(n) has property P provided n has property P
then n has property P for every n N.
Principles P1P3, known as the Peano axioms of arithmetic, are the
most commonly accepted axioms for the natural numbers. Axiom P3 is
the principle of mathematical induction.
Notice that the denition of the successor operator S(x) = xx from
the axiom of innity coincides with the denition given in this section, since
S(n) = n + 1 = n n.
Now we are ready to construct the set N.

3.1 Natural numbers 27
Theorem 3.1.1 There exists exactly one set N such that
(1) N,
(2) x N S(x) N for every x, and
(3) if K is any set that satises (1) and (2) then N K.
Proof By the axiom of innity there exists at least one set X satisfying
(1) and (2). Let
T = Y {(X): Y & x Y (S(x) Y )
and put N =

T. It is easy to see that the intersection of any nonempty

family of sets satisfying (1) and (2) still satises (1) and (2). To see that
(3) holds let K satisfy (1) and (2). Then X K T and N =

T
X K K. Theorem 3.1.1 is proved.
The set N from Theorem 3.1.1 will be called the set of natural numbers.
Natural numbers will usually be denoted by the letters m, n, p, . . . . The
operation S(n) is the counterpart to (n +1). Thus for a natural number n
we will henceforth write n + 1 instead of S(n).
The set of natural numbers will also be denoted by the symbol .
Notice that condition (3) of Theorem 3.1.1 can be rephrased as follows:
[0 K & x (x K S(x) K)] implies N K. (3.1)
If we additionally know that K N then (3.1) can be expressed as
[0 K & n (n K n + 1 K)] implies N = K. (3.2)
Property (3.2) is called the induction property of the set of natural numbers
and the proofs that use it are cited as proofs by induction.
It is easy to see that the induction property (3.2) implies axiom P3 if
we put K = n N: n has property P. Axiom P2 follows from The-
orem 3.1.2(f). To see that N satises P1 just notice that 0 = and
S(n) = n n is clearly nonempty.
Essentially all the known properties of the natural numbers can be
deduced from these denitions and the ZFC axioms.
1
However, we will
restrict ourselves only to those properties that are more connected with
our representation of the natural numbers and to those that will be of
more use in what follows.
1
In fact, the Peano axioms are strictly weaker than the ZFC axioms. More precisely,
everything that can be proved in the Peano arithmetic can be proved in ZFC. However,
there are properties of natural numbers that can be deduced from the ZFC axioms but
cannot be proved on the basis of the Peano axioms alone.

28 3 Natural numbers, integers, and real numbers
Theorem 3.1.2 For every natural numbers m, n, and p
(a) m m+ 1 and m m+ 1;
(b) if x m then x N;
(c) if m n then m+ 1 n;
(d) if m n and n p then m p;
(e) m m;
(f) if n + 1 = m+ 1 then n = m;
(g) m n if and only if m = n or m n;
(h) m n or n m.
Proof (a) This follows immediately from the denition m+1 = mm.
(b) The proof follows by induction. Let
K = n N: x (x n x N).
Notice that 0 = K, since there is no x 0. Now let n K and
x n + 1. Since n + 1 = n n we have two cases: If x n then x N,
since n K; if x n then x = n and also x N. Thus n K implies
n + 1 K. Hence, by (3.2), K = N and (b) is proved.
(c) The proof is left as an exercise.
(d) Since m n and, by (a), n n+1 we have m n+1. But, by (c),
n p implies n + 1 p. So m p.
(e) This can be easily proved by induction. It is left as an exercise.
(f) This is left as an exercise.
(g) : If m = n then m n. If m n then, by (a) and (c), m
m+ 1 n.
: Fix m N and consider the set
K = n N: m n [m n m = n].
It is enough to show that K = N.
First notice that 0 K since m 0 = implies m = = 0.
Now assume that n K. By induction it is enough to show that
n + 1 K. So assume that m n + 1 = n n. If m n then n m
and, by (c), n + 1 m n + 1, that is, m = n + 1. If m n then, since
n K, either m n or m = n. But in both cases m n n = n + 1; so
n + 1 K.
(h) Dene the set
K = m N: n N (n m m n).

3.1 Natural numbers 29
First we will show that K = N.
First notice that 0 K, since 0 = n for every n.
Now assume that m K. We will show that this implies m + 1 K.
So let n N be such that n m+ 1 = m m. Then n = m and n m.
But m K and n m imply that m n. This last condition and n = m
imply in turn, by (g), that m n and, by (c), m+1 n. Thus we conclude
that n K implies n + 1 K and, by induction, that K = N.
Now, to nish the proof, let n, m N. If n m then the conclusion is
correct. So assume that n m. Then, by (g), n m. Since m N = K
we conclude m n.
From Theorem 3.1.2 it follows immediately that every natural number
n = 0, 1, 2, . . . , n 1. However, we cannot express it in the formal lan-
guage of set theory, since the symbol . . . does not belong to its language.
We can only express it using innitely many sentences, one for each natural
number. On the other hand, this informal expression is more intuitive for
most mathematicians than the formal denition. We will often use it to
describe the set of natural numbers less than a xed number. In particular,
we will very often write 2 instead of 0, 1.
Let us also dene for natural numbers m and n
m < n if and only if m n
and
m n if and only if m n.
It is not dicult to see that these denitions imply the properties of < and
that we usually associate with them.
We will nish this section by dening on the set N the arithmetic opera-
tions of sum, product, and exponentiation. The denitions are by induction
and go as follows:
n + 0 = n; n + (m+ 1) = (n +m) + 1, where k + 1 stands for S(k).
n0 = 0; n(m+ 1) = (nm) +n.
n
0
= 1; n
m+1
= n
m
n.
It can be proved, by induction, that these operations satisfy all known
properties of the standard operations.
EXERCISES
1 Prove that if T is a nonempty family such that every F T satises (1)
and (2) from Theorem 3.1.1 then

T also satises these conditions.
2 Prove by induction Theorem 3.1.2(c), (e), and (f).

30 3 Natural numbers, integers, and real numbers
3.2 Integers and rational numbers
Intuitively, we will dene integers as dierences nm of two natural num-
bers. The formalization of this intuition has, however, two problems. First,
we do not have an operation dened as of yet. Thus we will talk about or-
dered pairs 'n, m` of natural numbers, thinking of them as nm. However,
representing the number n m as 'n, m` is not unique: 1 = 2 1 = 3 2.
So, instead, we will use equivalence classes of pairs of natural numbers,
where the equivalence relation E on N N is dened by the formula:
'n, m`E'n
t
, m
t
` n +m
t
= n
t
+m. (3.3)
Now the set of integers Z is dened as the quotient class (N N)/E.
We dene a relation on Z by
['n, m`] ['n
t
, m
t
`] n +m
t
n
t
+m. (3.4)
The arithmetic operations on Z are dened as follows:
['n, m`] + ['n
t
, m
t
`] = ['n +n
t
, m+m
t
`].
['n, m`] ['n
t
, m
t
`] = ['n, m`] + ['m
t
, n
t
`].
['n, m`] ['n
t
, m
t
`] = ['nn
t
+mm
t
, nm
t
+mn
t
`].
We will leave as an exercise the proof that E is an equivalence relation,
that the preceding operations are well dened (i.e., they do not depend on
the choice of representatives of equivalence classes), and that they have the
properties we know from algebra.
Following the same path, we dene the set of rational numbers as
ordered pairs 'a, b` of integers representing of a/b. More precisely, we
dene an equivalence relation E
t
on Z Z by
'a, b`E
t
'a
t
, b
t
` [a b
t
= a
t
b & b = 0 = b
t
] or [b = b
t
= 0].
Then
= ['a, b`] : a, b Z & b = 0,
where ['a, b`] is an equivalence class with respect to E
t
. We dene a relation
and arithmetic operations on Z as follows:
['a, b`] + ['a
t
, b
t
`] = ['a b
t
+a
t
b, b b
t
`];
['a, b`] ['a
t
, b
t
`] = ['a a
t
, b b
t
`];
['a, b`] ['a
t
, b
t
`] b 0 & b
t
0 & a b
t
a
t
b.
As before, we will leave as an exercise the proof that E
t
is an equivalence
relation, that the preceding denitions are well dened, and that they have
the desired properties.

3.3 Real numbers 31
EXERCISES
1 Prove that the relation E dened by (3.3) is an equivalence relation.
2 Show that on Z dened by (3.4) is well dened and is a linear-order
relation.
3.3 Real numbers
In the previous section we constructed integers from the natural numbers
and rational numbers from the integers. Both constructions were primarily
algebraic. This was possible since the rational numbers are the algebraic
completion of the set of natural numbers. The same approach, however,
cannot be used in a construction of the real numbers, since there are real
numbers of no simple algebraic relation to rational numbers.
The crucial geometric property that is often used to described real num-
bers is that the real line does not have holes. This intuition was used by
Dedekind to dene the set R of real numbers as follows.
A subset A of is said to be a Dedekind cut if A does not contain a
largest element and is a proper initial segment of , that is, = A =
and A contains every p q provided q A. Now, R can be dened as
the set of all Dedekind cuts, intuitively identifying A with a real number
supA. Then for Dedekind cuts A and B we can dene
A+B = p +q : p A & q B
and
A B A B.
Similarly, we can dene the product AB of two numbers (though the def-
inition must be done more carefully, since the product of two negative
numbers is positive).
Another approach for the denition of real numbers is to use their
numerical expansions. This approach is a little more messy, but it better
ts the methods of this course. So we will include it here too.
The construction of real numbers that follows will be done in two steps.
First, consider the set 2
N
of all sequences s: N 0, 1. It is usually
called the Cantor set. (The classical Cantor ternary set can be obtained
from this set by identifying a sequence a 2
N
with

nN
2a
n
/3
n+1
. See
Section 6.2 for more details.) We would like to identify a sequence a 2
N
with the real number

nN
a
n
/2
n+1
. However, this identication function
is not one-to-one. To correct it, we will dene [0, 1] as the quotient class

32 3 Natural numbers, integers, and real numbers
2
N
/E, where aEb if and only if a = b or there exists an n N such that
for every k N
[k < na
k
= b
k
] & [a
n
= 1 & b
n
= 0] &[k > n(a
k
= 0 &b
k
= 1)]. (3.5)
We will leave as an exercise the proof that E is indeed an equivalence
relation.
The linear-order relation on [0, 1] can be dened by
[a] [b] [a] = [b] n N [a
n
< b
n
& k n (a
k
= b
k
)]. (3.6)
Again, we will not prove the correctness of this denition.
Numbers 0 and 1 in [0, 1] are dened as equivalence classes of functions
constantly equal to 0 and 1, respectively. Then we dene
[0, 1) = [0, 1] ` 1, (0, 1] = [0, 1] ` 0, (0, 1) = [0, 1] ` 0, 1.
The set R of real numbers is dened as Z [0, 1), where intuitively we
identify a pair 'k, r` with k +r. In particular, we can dene on R by
'k, r` 'l, s` k < l (k = l & r s). (3.7)
In the remainder of this section we will recall some geometric and topo-
logical properties of the n-dimensional Euclidean space R
n
(n N). In
particular, the distance between two points p = 'p
1
, . . . , p
n
` and q =
'q
1
, . . . , q
n
` of R
n
is given by the formula
d(p, q) =

i=1
(p
i
q
i
)
2
.
The open ball in R
n
centered at p R
n
and with radius > 0 is dened as
B(p, ) = q R
n
: d(p, q) < .
A subset U of R
n
is open if U is a union of some family of open balls in
R
n
. The family of all open subsets of R
n
is called the (natural) topology
on R
n
and is closed under nite intersections and arbitrary unions.
A set F R
n
is closed if its complement R
n
` F is open. Notice that
nite unions of closed sets and arbitrary intersections of closed sets are also
closed.
For a subset S of R
n
its interior int(S) is dened as the largest open
subset of S, that is,
int(S) =

U R
n
: U is open in R
n
.

3.3 Real numbers 33
The closure of S is the smallest closed set containing S, that is,
cl(S) =

F S: F is closed in R
n
.
A subset D of R
n
is dense in R
n
if cl(D) = R
n
or, equivalently, when
D U = for every nonempty open set U R
n
. In particular,
n
is a
dense subset of R
n
. A subset N of R
n
is nowhere dense if int(cl(N)) = .
Now let us recall a few more specic properties of R
n
. A function
f : R
n
R
m
is continuous if f
1
(U) = x R
n
: f(x) U is open in
R
n
for every open set U R
m
. It is easy to see that for every dense
subset D of R
n
and continuous functions f : R
n
R
m
and g : R
n
R
m
if
f(d) = g(d) for every d D then f = g.
A subset B of R
n
is bounded if B B(p, ) for some p R
n
and
> 0. Closed, bounded subsets of R
n
are called compact. One of the most
important properties of compact subsets of R
n
is given in the next theorem.
Theorem 3.3.1 If K
0
K
1
K
2
is a decreasing sequence of com-
pact nonempty subsets of R
n
then their intersection

i=0
K
i
is nonempty.
A subset C of R
n
is connected if there do not exist two disjoint open
sets U, V R
n
such that U C = , V C = , and C U V . Recall
that any interval in R as well as any R
n
are connected.
A sequence p
k

k=0
of points in R
n
is a Cauchy sequence if for every
> 0 there is a number N such that d(p
i
, p
j
) < for every i, j > N. Every
Cauchy sequence in R
n
has a limit point. This fact serves as a basis for
the proof of the Baire category theorem, which follows.
Theorem 3.3.2 (Baire category theorem) If N
0
, N
1
, N
2
, . . . is a sequence
of nowhere-dense subsets of R
n
then its union

i=0
N
i
has an empty inte-
rior.
EXERCISES
1 Dene the product of two real numbers using Dedekinds denition of
R. Show that your product has the distributive property, that is, that
a(b +c) = ab +ac for every a, b, c R.
2 Prove that the relation E dened by (3.5) is an equivalence relation.
3 Show that on [0, 1] dened by (3.6) is well dened and is a linear-order
relation.
4 Prove that on R dened by (3.7) is a linear-order relation.

Part II
Fundamental tools of set
theory
35

Chapter 4
Well orderings and
transnite induction
4.1 Well-ordered sets and the axiom of foun-
dation
A binary relation R on a set X is said to be well founded if every nonempty
subset Y of X has an R-minimal element, that is, if
Y X [Y = m Y y Y (yRm)].
Examples 1. For every nite linearly ordered set 'X, ` the relation < is
well founded.
2. The relation < on the set N of all natural numbers is well founded. This
is known as the well-ordering principle for the natural numbers. It can be
easily deduced from the principle of mathematical induction P3. (See also
(3.2).) It is also a special case of the next example.
3. The relation is well founded on every nonempty set X. This, however,
does not follow from the axioms we have studied so far. For this we need
the following axiom known as the axiom of foundation or regularity. This
is the last of the axioms of ZFC.
Foundation axiom (Skolem 1922; von Neumann 1925) Every nonempty
set has an -minimal element:
x[y(y x)y[y x & z(z x & z y)]] .
This axiom is not essential for this course. However, it is good to know
some of its basic consequences.
37

38 4 Well orderings and transnite induction
Theorem 4.1.1 x x for every x.
Proof To obtain a contradiction assume that there exists a t such that
t t. Put x = t. We will see that the axiom of foundation fails for x.
Evidently, there exists a y x, since t x. Moreover, if y x then
y = t. But then, there exists a z, z = t, such that z = t x and
z = t t = y, contradicting the axiom of foundation.
Theorem 4.1.2 There is no innite -decreasing sequence, that is, there
is no sequence 'x
n
: n N` such that x
n+1
x
n
for all n N.
Proof Otherwise the set x
n
: n N would have no -minimal element.

A binary relation on X is a well-ordering relation if 'X, ` is linearly

ordered and the relation < is well founded. In other words, a linearly
ordered set 'X, ` is well ordered if every nonempty subset A of X has a
smallest element a A, usually denoted by minA.
It is pretty easy to see that a linearly ordered set 'X, ` is well ordered
if and only if it does not contain an innite strictly decreasing sequence,
that is, a sequence 'x
n
X: n N` such that x
n+1
< x
n
for every
n N. However, a formal proof of this fact requires a recursive denition
technique, which is still not available to us. Thus its proof will be postponed
until Section 4.3 (see Theorem 4.3.2).
Notice also that a subset of a well-ordered set is well ordered.
Examples 1. Every nite linearly ordered set is well ordered.
2. The set N of all natural numbers is well ordered by the usual order .
3. The set S = 1 1/n: n N, n > 0 1 is a well-ordered subset of
'R, ` since S
0
= 1 1/n: n N, n > 0 is ordered the same way as N
and 1 is greater than any number from S
0
.
4. The interval [0, 1] with the usual order is linearly ordered and has the
smallest element. However, it is not well ordered, since its subset (0, 1)
does not have a smallest element (0 is not an element of (0, 1)).
5. Any family of sets linearly ordered by the relation
xy x y x = y
is well ordered by . This follows from the axiom of foundation.
The partially ordered sets 'X, ` and 'Y, _` are (order) isomorphic (or
have the same order type) if there is a bijection f : X Y such that
a b if and only if f(a) _ f(b)

4.1 Well-ordered sets and the axiom of foundation 39
for every a, b X. The function f is called an order isomorphism (or just
isomorphism) between X and Y .
The next two propositions list some basic properties of order isomor-
phisms. Their proofs are left as exercises.
Proposition 4.1.3 Let 'X, `, 'Y, _`, and 'Z, <` be partially ordered
sets.
(1) If f : X Y is an order isomorphism between X and Y then the
inverse function f
1
: Y X is an order isomorphism between Y
and X.
(2) If f : X Y is an order isomorphism between X and Y and g : Y
Z is an order isomorphism between Y and Z then g f : X Z is
an order isomorphism between X and Z.
Proposition 4.1.4 Let f : X Y be an order isomorphism between par-
tially ordered sets 'X, ` and 'Y, _`. Then
(1) X is linearly ordered (well ordered) if and only if Y is linearly ordered
(well ordered);
(2) x X is the smallest (the greatest, a minimal, a maximal) element
of X if and only if f(x) is the smallest (the greatest, a minimal, a
maximal) element of Y .
Examples 1. The intervals [0, 1] and (0, 1) are not order isomorphic since
[0, 1] has a smallest element and (0, 1) does not have one. The set '{(R), `
is isomorphic to neither of these sets, since it is not linearly ordered.
2. The sets N of natural numbers and P of even natural numbers have the
same order type. The isomorphism is established by the function f(n) =
2n. The order type of this set is denoted by and represents an innite
strictly increasing sequence. (In Section 4.2 we will make the notion of
order type of well-ordered sets more formal.)
3. The order type of the set S = 1 1/n: n N, n > 0 1 is denoted
by +1 and it is not the same as , since S has a greatest element and N
does not.
4. The set n: n N ordered by is isomorphic to 'N, ()
1
` = 'N, `
by the isomorphism f(n) = n. The order type of this set is denoted by

. The sets with order type

are innite strictly decreasing sequences.

A subset S of a partially ordered set 'X, ` is an initial segment of X
if for every x, y X
x y & y S x S.

40 4 Well orderings and transnite induction
An initial segment S of X is proper if S = X. It is easy to see that if
f : X Y is an order isomorphism between 'X, ` and 'Y, _` then S X
is a (proper) initial segment of X if and only if f[S] is a (proper) initial
segment of Y .
Note that for every element x
0
of a partially ordered set 'X, ` the set
O(x
0
) = x X: x < x
0

is a proper initial segment of X. However, not every proper initial segment

of a partially (even linearly) ordered set is of the form O(x
0
), as is shown
by the subset (, 0] = x R: x 0 of R. On the other hand, such an
example cannot be found in a well-ordered set, as is asserted by the next
theorem.
Theorem 4.1.5 Every proper initial segment S of a well-ordered set W is
of the form O() for some W.
Proof Since S is a proper subset of W the set W ` S is nonempty. Let
= min(W ` S). Then S = O().
Theorem 4.1.6 (Principle of transnite induction) If a set A is well-
ordered, B A, and for every x A the set B satises the condition
O(x) B x B, (4.1)
then B = A.
Proof It is enough to show that A ` B = . To arrive at a contradiction
assume that A` B = and let x = min(A` B). Then, by minimality of x,
O(x) B. Hence, by (4.1), x B. This contradicts x A` B.
If we apply Theorem 4.1.6 to A = N then we obtain the following version
of the induction schema for the set of natural numbers.
Corollary 4.1.7 If B N and for every n N the set B satises the
condition
n B n B, (4.2)
then B = N.
The statement of Corollary 4.1.7 also follows immediately from the in-
duction property (3.2). The converse implication is also easy to prove and
can be concluded directly, without the use of Theorem 4.1.6.

4.1 Well-ordered sets and the axiom of foundation 41
Proofs using Theorem 4.1.6 are called proofs by transnite induction.
Such proofs are the basic techniques of set theory and will be used as a main
tool in this course. The proof of the next theorem is a classic example of
the use of this technique. To formulate it we need the following denition.
Let 'A, ` and 'B, ` be partially ordered sets. A function f : A B
is an increasing function if x y implies f(x) f(y) for every x, y A;
f is strictly increasing if x < y implies f(x) < f(y). Similarly we dene
decreasing and strictly decreasing functions.
Notice that if f : A B is strictly increasing then it establishes an
order isomorphism between A and f[A].
Theorem 4.1.8 If f : A A is a strictly increasing function dened on
a well-ordered set A then x f(x) for all x A.
Proof Let B = x A: x f(x). We will show that B = A. Let x A
be such that O(x) B. By Theorem 4.1.6 it is enough to show that x B.
So take y O(x). Then y f(y). Moreover, f(y) < f(x), since y < x.
Hence y < f(x) for every y O(x). In particular, f(x) A ` O(x). Since
x is the rst element of A` O(x) we conclude that x f(x). Thus x B.

Corollary 4.1.9 If well-ordered sets A and B are isomorphic, then the

isomorphism f : A B is unique.
Proof Let g : A B be another order isomorphism. We will show that
f = g. So let x A. The function g
1
f : A A is also an isomorphism.
Using Theorem 4.1.8 applied to g
1
f we conclude that x g
1
(f(x))
and, applying function g to both sides, g(x) f(x). Similarly we argue for
f(x) g(x). Thus f(x) = g(x). Since x was arbitrary, f = g.
Corollary 4.1.10 No well-ordered set is isomorphic to any of its proper
initial segments.
Proof If a well-ordered set A were isomorphic to an initial segment O(x)
for x A, and f : A O(x) were an isomorphism establishing this fact,
then we would have f(x) O(x), that is, f(x) < x, contradicting Theo-
rem 4.1.8.
Corollary 4.1.11 No two distinct initial segments of a well-ordered set
are isomorphic.

42 4 Well orderings and transnite induction
Theorem 4.1.12 If A and B are well-ordered sets then either
(i) A and B are isomorphic,
(ii) A is isomorphic to a proper initial segment of B, or
(iii) B is isomorphic to a proper initial segment of A.
Proof Let
Z = x A: y
x
B (O(x) and O(y
x
) are order isomorphic).
Notice that if x Z then, by Corollary 4.1.11, y
x
from the denition of
the set Z is unique. Dene f : Z B by putting f(x) = y
x
, that is,
f = 'x, y` AB: O(x) and O(y) are order isomorphic.
First we will show that Z is an initial segment of A. To see it, let
x < z and z Z. We have to show that x Z. Since z Z, there is an
isomorphism h between O(z) and O(f(z)). The restriction of h to O(x)
gives an isomorphism, whose range must be a proper initial segment S of
O(f(z)), and, by Theorem 4.1.5, S = O(y) for some y O(f(z)). Thus
x Z and we conclude that Z is an initial segment of A.
This argument shows also that f is strictly increasing, since x < z and
z Z imply that O(x) is isomorphic to O(y) for some y O(f(z)), that
is, f(x) = y < f(z).
Similarly as for Z we can also prove that f[Z] is an initial segment of B.
(Just note that f[Z] = y B: x A (O(y) and O(x) are isomorphic).)
So f : Z f[Z] is an order isomorphism.
If Z = A then f establishes an isomorphism between A and either B or
a proper initial segment of B. Similarly, if f[Z] = B then f
1
shows that
B is isomorphic to either A or a proper initial segment of A. Thus, it is
enough to prove that either Z = A or f[Z] = B.
To obtain a contradiction assume that Z = A and f[Z] = B. Then Z
is a proper initial segment of A and f[Z] is a proper initial segment of B.
Hence, by Theorem 4.1.5, Z = O(x) for some x A and f[Z] = O(y) for
some y B. But then, f is an isomorphism between Z = O(x) and O(y),
contradicting the fact that x Z.
We will nish this section with the following construction of an ordering
relation on a product space.
Theorem 4.1.13 Let F
t

tT
be an indexed family of nonempty linearly
ordered sets 'F
t
,
t
` and let T be well ordered by . If _ is a binary relation
on the Cartesian product P =

tT
F
t
dened by the formula
f _ g f = g s (s = mint T : f(t) = g(t) & f(s)
s
g(s))

4.1 Well-ordered sets and the axiom of foundation 43
for every f, g P, then _ is a linear order on P.
Moreover, if T is nite then P is well ordered by _ if and only if all
sets F
t
are well ordered.
The proof is left as an exercise.
The relation _ dened in Theorem 4.1.13 is called the lexicographic
order of

tT
F
t
and is sometimes denoted by
lex
.
EXERCISES
1 Prove Proposition 4.1.3.
2 Prove Proposition 4.1.4.
3 Let f : X Y be an isomorphism between partially ordered sets 'X, `
and 'Y, _`. Show that S X is a (proper) initial segment of X if and only
if f[S] is a (proper) initial segment of Y .
4 Complete the proof of Theorem 4.1.5 by showing that S = O().
5 Prove Theorem 4.1.13.
6 Show that in the additional part of Theorem 4.1.13 the assumption that
T is nite is essential by giving an example of well-ordered sets T and F
for which the lexicographic order on F
T
is not a well-ordering relation.
7 Let 'K, ` and 'X, ` be nonempty linearly ordered sets such that X has
a xed element 0 X. For f : K X let supp(f) = k K: f(k) = 0
and put
T(K, X) = f X
K
: supp(f) is nite.
Dene a binary relation _ on T(K, X) by putting
f _ g f = g or f(m) g(m), where m = maxk K: f(k) = g(k),
for every f, g T(K, X). Prove that
(a) the relation _ is a linear-order relation on T(K, X);
(b) if X has at least two elements and 0 is the minimal element of X
then 'T(K, X), _` is well ordered if and only if 'X, ` and 'K, `
are well ordered.
The relation _ is usually called an antilexicographic-order relation.

44 4 Well orderings and transnite induction
4.2 Ordinal numbers
In the previous section we were informally talking about the order type
of a given partially ordered set 'X, `. The natural way of thinking about
such an object is to consider it as a representative from the equivalence
class [X]

of all ordered sets order isomorphic to 'X, `. This equivalence

class, however, is not a set! (The argument is similar to that for the class
of all sets.) Moreover, if we want to talk about all order types, we nd also
that there are too many of them; that is, even if we discover a good way of
choosing a representative from each equivalence class [X]

, the class of all

such representatives will not be a set.
One way to avoid this problem is to consider the sentences 'Y, ` has
the order type of 'X, ` and 'Y, ` and 'X, ` have the same order
type just as abbreviations for the well-dened sentence 'Y, ` and 'X, `
are order isomorphic. That is the way we will think about the order types
of arbitrary partially ordered sets. For the well-ordered sets, however,
we can indeed follow the path of our intuition described in the previous
paragraph. This section is dedicated to fullling this goal by constructing
the canonical order types of well-ordered sets, called ordinal numbers. The
construction is due to von Neumann (1929).
The class of all ordinal numbers is still too big to be a set. Thus, instead
of constructing it as a set, we will dene a formula () that will represent
an intuitive sentence is an ordinal number. The denition is as follows.
A set is said to be an ordinal number (or just an ordinal ) if it has
the following properties:
(A

) If then ;
(B

) If , then = , , or ;
(C

) If = B then there exists a B such that B = .

It follows from Theorem 3.1.2 that if stands for either any natural
number or the set N of natural numbers then satises conditions (A

)
and (B

). Also, the well-ordering principle for the natural numbers implies

condition (C

) in this case. So every natural number is an ordinal number

and so is N. It is also easy to see that the set N N is an ordinal
number. The sets N and NN, when considered as ordinal numbers, are
traditionally denoted as and + 1, respectively.
We will start with the following basic property.
Theorem 4.2.1 If is an ordinal number and then is also an
ordinal number.
Proof Assume that satises (A

)(C

) and let . We have to show

that satises (A

)(C

).

4.2 Ordinal numbers 45
First notice that (A

) implies . To see (C

) let = B . Then
= B and (C

) follows from (C

). To see (B

) take , . Then
, and (B

) follows from (B

).
Finally, to see (A

) let . We have to show that . So choose

. The proof will be nished when we show that . But .
So, by (A

), and , since . We have proved that , .

Now, using (B

), we conclude that , = , or . If then we

are done. We will prove that the other two cases are impossible.
To obtain a contradiction, assume rst that . Then, in particular,
. Put B = , , . Then = B . But then (C

) fails
for this B, since B, B, and B, that is, the sets B,
B, and B are nonempty.
So assume = . Then = . Choosing B = , gives a
contradiction similar to that in the previous case.
Now notice that for an ordinal number and for arbitrary ,
if and only if = . (4.3)
The implication follows immediately from (A

), which holds by The-

orem 4.2.1. To see the other implication let . Since , , condition
(B

) implies that = , , or . If either of the rst two condi-

tions holds, we are done. But if then, by (A

), , and combining
this with we obtain = .
Theorem 4.2.2 If is an ordinal number then the relation is a well-
ordering relation on .
Proof Inclusion is clearly a partial-order relation on any family of sets,
so it is on . It is a linear order on by (B

) and (4.3).
To see that is well ordered by , let = B and let B be
as in (C

), that is, such that B = . We will show that is minimal

with respect to . So choose arbitrary B. Then, by (B

) and (4.3),
either or . But implies B = , which is impossible.
Thus . Since was an arbitrary element of B we conclude that is
a minimal element in B.
It is customary that the relation on an ordinal number is denoted
by . It is also easy to see that, according to our general agreement, the
relation < on an ordinal number is identical to .
Notice also that by Theorem 4.1.5 every proper initial segment of an
ordinal number is of the form O() for some and that O() = .
Thus every initial segment of an ordinal number is an ordinal number.
The basic properties of ordinal numbers are as follows.

46 4 Well orderings and transnite induction
Theorem 4.2.3 For every ordinal numbers and
(i) if and are order isomorphic then = ;
(ii) = , < , or < ;
(iii) if T is a nonempty family of ordinal numbers then

T is also an
ordinal number.
Proof (i) Let f : be an order isomorphism and let
Z = : f() = .
We will show, by transnite induction, that Z = . This will nish the
proof, since Z = implies that f is the identity function on , and thus
= f[] = .
To see that Z = let be such that O() Z. Then f[O()] =
O(f()) and f() = for all < . Hence
f() = O(f()) = f[O()] = f(): < = : < = .
Thus we have proved that for every condition O() Z implies that
Z. Hence Z = . Condition (i) has been proved.
(ii) By Theorem 4.1.12 either and are isomorphic, or is isomorphic
to a proper initial segment of , or is isomorphic to a proper initial
segment of . By (i), this implies (ii).
(iii) We have to check that =

T satises (A

)(C

).
(A

): If

T then there is an T such that . Then

T.
(B

): If ,

T then there are
1
,
2
T such that
1
and

2
. By (ii) either
1

2
or
2

1
. Let be the greater of these
two. Then , and, by (B

), = , , or .
(C

): Let = B

T and let B. Then there is an T such
that . Hence, by Theorem 4.2.1, is an ordinal. If B = put
= . Otherwise, put = min( B), which exists by (C

). It is easy to
see that B = .
It is not dicult to see that for an ordinal number the set is
also an ordinal number. We will denote it by +1 and call it the (ordinal
immediate) successor of . Number is also called the (ordinal immediate)
predecessor of +1. It is also easy to check that +1 is the smallest ordinal
number greater than . Thus every ordinal number has its successor. Not
every number, however, has an immediate predecessor; for example, 0 and
do not have one. The ordinal numbers that have immediate predecessors
(i.e., those in the form + 1) are called ordinal successors. Those that do
not have immediate predecessors are called limit ordinals.

4.2 Ordinal numbers 47
The next theorem justies our intuition of considering ordinal num-
bers as representatives of the abstract classes of all order-isomorphic well-
ordered sets.
Theorem 4.2.4 For every well-ordered set 'W, ` there exists precisely
one ordinal number that is order isomorphic to W.
Proof The uniqueness of follows immediately from Theorem 4.2.3(i).
For w W let O[w] = O(w) w and dene
Z = w W : O[w] is order isomorphic to some ordinal number
w
.
Notice that by Theorem 4.2.3(i) and Corollary 4.1.9 the ordinal number

w
and the isomorphism f
w
: O[w]
w
are unique.
Now, if v, w Z and v w then f
w
[O[v]] is an ordinal number ,
as an initial segment of
w
. In particular, f
w
[
O[v]
: O[v] is an order
isomorphism and, by the preceding uniqueness remark, f
w
[
O[v]
= f
v
. So
we have shown that
f
v
f
w
for every v, w Z, v w. (4.4)
We will prove, by transnite induction, that Z = W.
So let O(w) Z. Then, by (4.4), f =

vO(w)
f
v
is a strictly increasing
function from

vO(w)
O[v] onto

vO(w)

v
. But

vO(w)
O[v] = O(w)
and, by Theorem 4.2.3(iii), =

vO(w)

v
is an ordinal number. So
f : O(w) is an order isomorphism. Extend f to a function F : O[w]
+ 1 by putting F(w) = . Then the existence of F proves that w Z.
We have shown that Z = W.
Now, by (4.4), the function f =

wW
f
w
is an isomorphism between
W =

wW
O[w] and an ordinal number

wW

w
.
In what follows it will be convenient to use the following theorems.
Theorem 4.2.5 If and are ordinal numbers and f : is a strictly
increasing function then . Moreover, f() for every < .
Proof If < then and f is a strictly increasing function from
into . But then f() , that is, f() < , contradicting Theorem 4.1.8.
So, by Theorem 4.2.3(ii), .
The additional part follows from Theorem 4.1.8.
For a well-ordered set W let Otp(W) stand for the order type of W,
that is, the unique ordinal number that is order isomorphic to W.
Corollary 4.2.6 If W is well ordered and B W then Otp(B) Otp(W).

48 4 Well orderings and transnite induction
Proof Let f : Otp(B) B and g : W Otp(W) be order isomorphisms.
Then gf : Otp(B) Otp(W) is strictly increasing, and by Theorem 4.2.5
Otp(B) Otp(W).
We will also introduce the following arithmetic for ordinal numbers. For
ordinal numbers and we dene the sum + of and as the order
type of the well-ordered set (0 ) (1 ) ordered by
'i, ` 'j, ` i < j (i = j & ).
Thus, we append to . It is easy to see that the relation so dened is
indeed a well-ordering relation.
Similarly, we dene the product of ordinal numbers and as the
order type of the set ordered lexicographically. By Theorem 4.1.13
it is well ordered.
It is not dicult to see that, in general, + = + and = (see
Exercise 1). We have the following monotonic laws for these operations.
Theorem 4.2.7 For arbitrary ordinal numbers , ,
(a) < implies + < +;
(b) implies + +;
(c) < and > 0 imply < ;
(d) implies .
Moreover, the inequalities in (b) and (d) cannot be replaced by < even
if < .
Proof To see (a) let f : (0)(1) (0)(1) be the
identity map. Notice that f is an isomorphism between (0)(1)
and the initial segment O('1, `) of (0 ) (1 ). Thus + is
isomorphic to a proper initial segment of +, so + < +.
To see (b) let f : (0 ) (1 ) (0 ) (1 ) be the
identity map. Notice that f is strictly increasing. Thus, via an appropriate
isomorphism, f can be transformed into a strictly increasing function from
+ into +. But then, by Theorem 4.2.5, + +.
To see that equality can hold even if < , notice that 1+ = = 0+
(see Exercise 1).
Parts (c) and (d) are left as exercises.
For an ordinal number any function a on is called a transnite
sequence and is usually denoted by a

<
or 'a

: < `, where a

=
a(). If is a domain of a transnite sequence a then we also often say
that a is an -sequence.

4.3 Denitions by transnite induction 49
EXERCISES
1 Show that for ordinal numbers and we might have + = +
and = by proving that
(a) 1 + = , so 1 + = + 1;
(b) 2 = , 2 = +, and = +.
2 Prove parts (c) and (d) of Theorem 4.2.7.
4.3 Denitions by transnite induction
Most mathematicians take for granted that if we have a procedure asso-
ciating a number a
n+1
to a number a
n
and we have dened a
0
, then the
sequence a
n

nN
constructed in such a way exists. The proof of this fact
for a nite portion a
n

nm
, m N, of such a sequence does not require
any special axioms. The proof of existence of the innite sequence a
n

nN
,
however, requires some form of argument, and the axiom of choice is needed
to prove any procedure justifying such an argument. (For example, if we
stay only within the framework of Peano arithmetic, we have no way to
prove the existence of an innite object a
n

nN
, since we operate in a
world with nite objects only.) This argument is given in the next theo-
rem.
Theorem 4.3.1 (Recursion theorem) Let Z be a set, be an ordinal num-
ber, and T be the family of all -sequences for < with values in Z, that
is, T =

<
Z

. Then for each function h: T Z there exists precisely

one function f : Z such that
f() = h(f[

) for all < . (4.5)

The function h from the theorem tells us how to choose f() knowing
the values of the sequence 'f(): < ` chosen so far. Because of this, we
will sometimes call it an oracle function.
Proof of Theorem 4.3.1 First we prove the uniqueness of f. So let g
be another function satisfying (4.5) and let W = < : f() = g().
It is enough to show that W = . But if = O() W then f[

= g[

and f() = h(f[

) = h(g[

) = g(). Hence W, and by a transnite

induction argument W = . The uniqueness of f has been proved.
To prove the main body of the theorem assume, to obtain a contradic-
tion, that the theorem fails for some ordinal number . We may assume

50 4 Well orderings and transnite induction
that is the smallest ordinal number for which the theorem fails, since if
it is not, we can replace it by the minimum
min < : theorem fails for .
Let h be a function for which the theorem fails for ordinal . For every
< there is a function f

: Z such that
f

() = h(f

) for all < . (4.6)

But for any < the function f

also satises (4.6). Thus, by the

uniqueness part that we have already proved, f

= f

. In particular,
f

() = f

() for every < < .

Now, if is a successor ordinal, say = + 1, then dene f : Z
as an extension of f

by putting f() = h(f

). Then f satises (4.5) since

f[

= f

for every < . This contradicts the choice of .

Thus assume that is a limit ordinal. Then +1 < for every < .
Dene f : Z by
f() = f
+1
() for <
and notice that such a function satises (4.5), since
f() = f
+1
() = h(f
+1
[

) = h(f[

)
for every < . This contradicts the choice of .
Most commonly, the oracle function h is dened as a choice function C
on some family of sets. In such a case a particular value h() is chosen as a
value of C on some set, whose denition depends on the previously chosen
values of f, that is, it is of the form
h(p) = C(z Z: (z, p)),
where C: {(Z) Z is a choice function and (z, p) is some formula. Of
course, such a function h is not dened for the values of p T for which
the set z Z: (z, p) is empty. Formally, to use the recursion theorem
we have to dene it for these values of p as well; however, in most cases we
are not interested in such values anyway. Thus, we usually assume that h
is arbitrarily dened somehow for such values and we dont even bother to
mention it. (The same is true for the other values for the oracle function
in which we are not interested.) Also, it is customary in such situations to
simply write that we choose in the inductive step
f() z Z: (z, f[

)

4.3 Denitions by transnite induction 51
or simply
f() = z such that z satises (z, f[

)
without mentioning explicitly the oracle function at all.
In the applications using the recursion theorem we often separately
specify the value of the function h for the empty sequence by h(0) = A and
for successor ordinals + 1 by h() = F(()) for some function F Z
Z
,
where Z
+1
. Thus the value of h on successor ordinal + 1 depends
only on the value () of on its last element . The -sequence f obtained
that way satises the conditions
f(0) = A, f( + 1) = F(f()) for any ordinal < ,
f() = h(f[

) for any limit ordinal < .

Notice that to dene the function f only on = we dont need the
very last condition. Then the rst two conditions can be rewritten as
f(0) = A, f(n + 1) = F(f(n)) for any n < .
The existence of such a sequence is usually deduced from the standard (not
transnite) recursion theorem.
There is probably no need to explain to the reader the importance of
this last theorem. Without this theorem it is sometimes hard to prove even
very intuitive theorems. As an example we can state here the following
characterization of well orderings.
Theorem 4.3.2 A linearly ordered set 'X, ` is well ordered if and only
if it does not contain a subset of order type

(i.e., a strictly decreasing

-sequence).
Proof : If there exists a strictly decreasing sequence 'x
n
: n < ` then
the set x
n
: n < does not have a smallest element.
: Assume that X is not well ordered and construct, by induction on
n < , a strictly decreasing sequence 'x
n
: n < `. The details are left as
an exercise.
The next theorem is one of the most fundamental theorems in set theory.
Theorem 4.3.3 (Well-ordering or Zermelos theorem) Every nonempty
set X can be well ordered.
Proof Let C be a choice function on the family {(X) ` . Choose an
arbitrary p X, for example, p = X, and dene F : {(Xp) Xp
by
F(Z) =

C(Z) for Z {(X) ` ,

p otherwise.

52 4 Well orderings and transnite induction
Let ( be the family of all ordinal numbers such that there is a one-
to-one -sequence in X. Notice that ( is a set by the replacement axiom,
since
( = Otp('Y, `): Y {(X) & {(X X) & 'Y, ` is well ordered.
Let be an ordinal number that does not belong to (. For example,
take = (

() + 1. By the recursion theorem there exists a function

f : X p such that
f() = F(X ` f(): < ) (4.7)
for all < . (The function h is dened by h(s) = F(X`range(s)).) Notice
that for every
if f[] X then f[

is one-to-one. (4.8)
This is the case since, by (4.7), f() X ` f(): < for < . In
particular, f[] X since there is no one-to-one -sequence in X. So the
set S = : f() = p is nonempty. Let = minS. Then f[

is
one-to-one and it establishes a well ordering of X of type by
x _ y f
1
(x) f
1
(y).
Remark It is clear that we used the axiom of choice in the proof of The-
orem 4.3.3. It is also easy to prove in ZF that Theorem 4.3.3 implies the
axiom of choice.
1
To argue for this it is enough to show that Theorem 4.3.3
implies the existence of a choice function for every family T of nonempty
sets. But if is a well ordering of

T then the function f(F) = minF
for F T is a choice function dened only with help of the comprehension
schema.
The combination of the recursion and well-ordering theorems gives a
very strong proving technique. However, for its use usually some kind of
cardinal argument is needed, a tool that we still have not developed. Thus,
the direct use of this technique will be postponed until subsequent chapters.
On the other hand, the next theorem will give us a way to prove the results
that require some kind of transnite induction argument without doing it
explicitly. To state it we need the following denitions.
For a partially ordered set 'P, ` we say that a set S P is a chain in
P if S is linearly ordered by . An element b P is said to be an upper
bound of a set S P if s b for every s S. Similarly, b P is a lower
bound of S P if b s for every s S.
1
ZF stands for the ZFC axioms from which the axiom of choice AC has been removed.
See Appendix A.

4.3 Denitions by transnite induction 53
Notice that not all subsets of a partially ordered set must have upper
or lower bounds. For example, (0, ) R does not have an upper bound
and every number r 0 is a lower bound of (0, ).
Let us also recall that m P is a maximal element of 'P, ` if there is
no p P such that m < p.
Theorem 4.3.4 (Zorns lemma
2
) If 'P, ` is a partially ordered set such
that every chain in P has an upper bound then P has a maximal element m.
Proof The proof is similar to that of Theorem 4.3.3. Let C be a choice
function on the family {(P)`. Choose an arbitrary x P, for example,
x = P, and dene F : {(P x) P x by
F(Z) =

C(Z) for Z {(P) ` ,

x otherwise.
Let ( be the family of all ordinal numbers such that there is a strictly
increasing sequence 'p

: < ` in P. Let be an ordinal number that

does not belong to (. For example, take = (

() + 1.
Dene a function h on the family of all -sequences 'p

: < ` in
P x, < , by putting
h('p

: < `) = F(B ` p

: < ), (4.9)
where B = b P : b is an upper bound of p

: < . (Notice that

B = if p

: < P.) By the recursion theorem, there is a function

f such that
f() = h(f[

) for all < .

Notice that for every
if f[] P then f[

is strictly increasing. (4.10)

This is the case, since then, by (4.9), f() B ` p

: < for < . In

particular, f[] P since there is no strictly increasing -sequence in P.
So the set S = : f() = x is nonempty. Let = minS. Then f[

is strictly increasing, so B = b P : b is an upper bound of p

: <
is nonempty, p

: < being a chain. Let m B. It is enough to argue

that m is a maximal element of P. But if there is a y P with y > m,
then y > m p

for every < . In particular, y B ` p

: < , so
2
M. Zorn proved this lemma in 1935 and published it in Bulletin AMS. The same theo-
rem was proved in 1922 by K. Kuratowski and published in Fundamenta Mathematicae.
Thus priority for this theorem belongs without any doubt to Kuratowski. However, in
essentially all published sources the name Zorn is associated to this theorem and it seems
that the battle for historical justice has been lost.

54 4 Well orderings and transnite induction
B ` p

: < = . Thus, by (4.9), f() P, contradicting the choice

of .
Zorns lemma is used most often in the situation when P is a family of
subsets of a set A ordered by inclusion and such that for every chain S P
its union

S is also in P. (Notice that

S is an upper bound for S with
respect to .) In this particular case Zorns lemma has an especially nice
form, stated as the following corollary.
Corollary 4.3.5 (Hausdor maximal principle) If T is a nonempty family
of subsets of a set A such that for every chain S T its union

S belongs
to T then T has an -maximal element.
Remark Zorns lemma and the Hausdor maximal principle are equivalent
to the axiom of choice within the ZF theory. To see this rst notice that
Zorns lemma follows from the axiom of choice (Theorem 4.3.4) and that the
Hausdor maximal principle follows from Zorns lemma (Corollary 4.3.5).
To see that the Hausdor maximal principle implies the axiom of choice
let ( be a nonempty family of pairwise-disjoint nonempty sets. We have
to nd a selector S for (. So let T be the family of all T

( such that
T G has at most one element for every G (. Notice that T satises the
assumptions of the Hausdor maximal principle. Let S T be maximal.
We will show that S is a selector for (. But if S is not a selector for (
then G S = for some G (. Then there is a g G and S g is in
T, contradicting the maximality of S.
EXERCISE
1 Complete the details of the proof of Theorem 4.3.2.
4.4 Zorns lemma in algebra, analysis, and
topology
In this section we will see three standard applications of Zorns lemma in
three main branches of mathematics: algebra, analysis, and topology. Each
of these areas will be represented, respectively, by the theorem that every
linear space has a basis, the HahnBanach theorem, and the Tychono
theorem. Going through these proofs should help the reader to appreci-
ate the power of Zorns lemma and, implicitly, the axiom of choice. The
HahnBanach theorem and the Tychono theorem will not be used in the
remaining part of this text.
Algebraic application To state the algebraic example let us recall the
following denitions. A set G with a binary operation +: GG G is a

4.4 Zorns lemma in algebra, analysis, and topology 55
group if + is associative (i.e., (u+v) +w = u+(v +w) for all u, v, w G),
G has the identity element 0 (i.e., 0 +v = v + 0 = v for every v G), and
every element of G has an inverse element (i.e., for every v G there exists
v G such that (v) +v = v + (v) = 0). A group 'G, +` is Abelian if
a + b = b + a for every a, b G. A set K with two binary operations +
and on K is a eld if 'K, +` is an Abelian group with 0 as an identity
element, 'K`0, ` is an Abelian group with 1 as an identity element, and
a(b +c) = ab +ac for every a, b, c K. In what follows we will mainly be
concerned with the elds 'R, +, ` of real numbers and ', +, ` of rational
numbers.
An Abelian group 'V, +` is said to be a linear space (or vector space)
over a eld K if there is an operation from K V to V , 'k, v` kv, such
that (k + l)v = kv + lv, k(v + w) = kv + kw, k(lv) = (kl)v, 0v = 0, and
1v = v for every k, l K and v, w V . A subset S of a vector space V over
a eld K is a linear subspace of V if it is a linear space when considered
with the same operations. The main examples of vector spaces considered
in this text will be R
n
over either R or .
A subset S of a vector space V over a eld K is linearly independent
if for every nite number of distinct elements v
1
, . . . , v
n
of S and every
k
1
, . . . , k
n
K the condition k
1
v
1
+ +k
n
v
n
= 0 implies k
1
= = k
n
=
0. A subset S of a vector space V over a eld K spans V if every v V
can be represented as v = k
1
v
1
+ + k
n
v
n
for some v
1
, . . . , v
n
S and
k
1
, . . . , k
n
K. A basis of V is a linearly independent subset of V that
spans V .
It is easy to see that if B is a basis of V then every v V has a unique
representation v = k
1
v
1
+ + k
n
v
n
, where v
1
, . . . , v
n
B are dierent
and k
1
, . . . , k
n
K (and where we ignore the v
i
s for which k
i
= 0). This is
the case since if v = l
1
v
1
+ +l
n
v
n
is a dierent representation of v, then
(l
1
k
1
)v
1
+ + (l
n
k
n
)v
n
= v v = 0 so l
1
k
1
= = l
n
k
n
= 0.
Thus l
i
= k
i
for i 1, . . . , n.
Theorem 4.4.1 If S
0
is a linearly independent subset of a vector space V
over K then there exists a basis B of V that contains S
0
.
In particular, every vector space has a basis.
Proof The additional part follows from the main part, since the empty
set is linearly independent in any vector space.
To prove the main part of the theorem, let
T = S V : S
0
S and S is linearly independent in V .
Notice rst that T satises the assumptions of the Hausdor maximal
principle.

56 4 Well orderings and transnite induction
Clearly T is nonempty, since S
0
T. To check the main assumption,
let ( T be a chain in T with respect to . We will show that

( is
linearly independent in V . So let v
1
, . . . , v
n
be dierent elements of

(
and choose k
1
, . . . , k
n
K such that k
1
v
1
+ + k
n
v
n
= 0. For every
i 1, . . . , n let G
i
( be such that v
i
G
i
. Since G
1
, . . . , G
n
is a
nite subset of a linearly ordered set (, we can nd the largest element,
say G
j
, in this set. Then v
i
G
i
G
j
for all i 1, . . . , n. Hence all v
i
s
are in a linearly independent set G
j
. Thus k
1
v
1
+ + k
n
v
n
= 0 implies
k
1
= = k
n
= 0. So we can use the Hausdor maximal principle on T.
Let B be a maximal element in T. We will show that it is a basis of V .
It is linearly independent, since it belongs to T. So it is enough to prove
that B spans V .
To obtain a contradiction assume that there is a v V such that
v = k
1
v
1
+ +k
n
v
n
(4.11)
for every v
1
, . . . , v
n
V and k
1
, . . . , k
n
K. We will show that this
implies that B v is linearly independent in V , which contradicts the
maximality of B in T. So choose dierent elements v
1
, . . . , v
n
from B and
k
0
, . . . , k
n
K such that k
0
v +k
1
v
1
+ +k
n
v
n
= 0. There are two cases.
Case 1: k
0
= 0. Then k
1
v
1
+ +k
n
v
n
= 0 and v
1
, . . . , v
n
are from B.
Hence, also k
1
= = k
n
= 0.
Case 2: k
0
= 0. Then v = (k
1
/k
0
)v
1
+ (k
n
/k
0
)v
n
, contradicting
(4.11).
Analytic application For the next theorem the term vector space will be
used for the vector spaces over the eld R of real numbers. Recall that a
function f is said to be a linear functional on a vector space V if f : V R
is such that f(ax +by) = af(x) +bf(y) for every x, y V and a, b R.
Theorem 4.4.2 (HahnBanach theorem) Let V be a vector space and
p: V R be such that p(x + y) p(x) + p(y) and p(ax) = ap(x) for
all x, y V and a 0. If f is a linear functional on a linear subspace S of
V such that f(s) p(s) for all s S, then there exists a linear functional
F on V such that F extends f and F(x) p(x) for all x V .
Proof Let T be the family of all functionals g on a linear subspace T of
V such that g extends f and g(x) p(x) for all x T. First we will show
that T satises the assumptions of the Hausdor maximal principle.
So let ( T be a chain in T with respect to . We will show that

( T. Clearly f

(. To see that g =

( is a functional, notice rst

that g is a function since for 'x, s
0
`, 'x, s
1
` g we can nd g
0
, g
1
( such
that 'x, s
i
` g
i
for i < 2. Then, for g
0
g
1
, s
0
= g
0
(x) = g
1
(x) = s
1
.
Similarly, if g
1
g
0
then s
1
= g
1
(x) = g
0
(x) = s
0
.

4.4 Zorns lemma in algebra, analysis, and topology 57
In a similar way we show that the domain of g is a linear subspace and
that g is a linear functional on its domain. Since also g(x) p(x) for all
x dom(g), we conclude that g T, that is, T satises the assumptions
of the Hausdor maximal principle.
Now let F : T R be a maximal element of T. Then F is a linear
functional extending f such that F(x) p(x) for every x T. It is enough
to prove that T = V .
To obtain a contradiction assume that V = T and let w V `T. We will
show that this contradicts the maximality of F by constructing a G T
that extends F and is dened on a linear subspace V
0
= Span(T w) =
aw +t : a R, t T of V .
First notice that for every s, t T
F(s) +F(t) = F(s +t) p(s +t) = p(s w+t +w) p(s w) +p(t +w)
and so
p(s w) +F(s) p(t +w) F(t).
Hence, in particular,
sup
tT
[p(t w) +F(t)] inf
tT
[p(t +w) F(t)].
Choose such that sup[p(t w) +F(t)] inf[p(t +w) F(t)]. Then
inf[p(t +w) F(t)] & inf[p(t w) F(t)].
(4.12)
Now for a R and t T dene
G(aw +t) = a +F(t).
Clearly G is a functional on V
0
that extends F. We have to show only that
G(s) p(s) for all s V
0
. So let s = aw +t with t T, a R.
If a = 0 then G(aw +t) = F(t) p(t) = p(aw +t).
If a > 0 then, by the rst part of (4.12),
G(aw +t) = a +F(t)
= a[ +F(t/a)]
a[[p(t/a +w) F(t/a)] +F(t/a)]
= ap(t/a +w)
= p(t +aw).

58 4 Well orderings and transnite induction
If a = b < 0 then, by the second part of (4.12),
G(aw +t) = b +F(t)
= b[ +F(t/b)]
b[[p(t/b w) F(t/b)] +F(t/b)]
= bp(t/b w)
= p(t +aw).
This nishes the proof of Theorem 4.4.2.
Topological application Let us start by recalling some denitions. A
family T of subsets of a set X is said to be a cover of a set A X if
A

T. If T is a cover of A then a family ( T is said to be a subcover

of T if A

(. If 'X, ` is a topological space then T is said to be an
open cover of A X if A

T and T . A topological space 'X, ` is

compact if every open cover of X has a nite subcover.
For a topological space 'X, ` a family B is said to be a base for
X if for every U and every x U there exists a B B such that
x B U. A family o is said to be a subbase of topological space
'X, ` if the family
B(o) =

i<n
S
i
{(X): n N & i < n (S
i
o)

forms a base for X. It is easy to see that for any family o of subsets of a
set X the family
T (o) =

(: ( B(o)

forms a topology on X. This topology is said to be generated by o. The

family o is also a subbase for the topology T (o).
Finally, let us recall that for a family 'X

`
/
of topological
spaces their Tychono product is a topological space dened on the product
set X =

/
X

with topology generated by the subbase

p
1

(U): / & U

,
where p

: X X

is the projection of X onto its th coordinate X

.
Theorem 4.4.3 (Tychono theorem) The product of an arbitrary family
of compact spaces is compact.
The proof will be based on the following lemma.

4.4 Zorns lemma in algebra, analysis, and topology 59
Lemma 4.4.4 (Alexander subbase theorem) If there is a subbase o of a
topological space X such that every open cover 1 o of X has a nite
subcover then X is compact.
Proof Let o be a subbase of X such that any cover 1 o of X has a
nite subcover. Let B = B(o) be a base for X generated by o and, to
obtain a contradiction, assume that X is not compact. Then there is a
subfamily |
0
of B(o) that covers X but does not have a nite subcover.
(We use here an elementary fact that for any base B of a topological space
X, the space X is compact if and only if every cover | B of X has a
nite subcover.)
Dene T as

| B: |
0
|, X

|, and there is no nite subcover 1 | of X

.
To see that T satises the assumptions of the Hausdor maximal principle
let ( T be a chain. If T

( is nite then there is a | ( such that

T |. So T cannot cover X, since | T. Thus

( does not have a
nite subcover, that is,

( T.
So, by the Hausdor maximal principle, there is a maximal element
| T. Now let U |. Then U =

i<n
U
i
for some U
i
o. We claim
that
there is a j < n such that U
j
|. (4.13)
Otherwise, for every i < n we can nd a nite |
i
| such that |
i
U
i

covers X, that is, |

i
covers X`U
i
. But then

i<n
|
i
covers

i<n
(X`U
i
) =
X `

i<n
U
i
= X ` U. So

i<n
|
i
U | is a nite subcover of X,
contradicting the fact that | T.
But (4.13) implies that for every U | there is a V = U
j
| o such
that U V . So 1 = | o is also a cover of X. Hence, by our assumption,
1 has a nite subcover of X. This contradiction nishes the proof.
Proof of the Tychono theorem Let
o = p
1

(U): / & U

be a subbase of X =

/
X

and let | o be a cover of X. By the

Alexander subbase theorem it is enough to nd a nite subcover of |. For
/ let
|

= U X

: U is open in X

and p
1

(U) |.
If for some / the family |

covers X

then, by the compactness

of X

, there is a nite subcover 1 |

that covers X

. But then

60 4 Well orderings and transnite induction
|
t
= p
1

(U): U 1 | is a nite subcover of X. So, to obtain a

contradiction, assume that no |

covers X

and let x() X

for
every /. Then x X `

p
1

(U): U |

= X `

|. This
nishes the proof.
EXERCISES
1 Prove that every partial-order relation_ on a set X can be extended to
a linear-order relation on X. (Here extends _ if _.)
2 We say that a subset A R is algebraically independent if for every
nonzero polynomial p(x
1
, . . . , x
n
) of n variables with rational coecients
and any sequence a
1
, . . . , a
n
of dierent elements from A, p(a
1
, . . . , a
n
) =
0. Show that there exists an algebraically independent subset / of R such
that if (/) is a eld generated by and / then for every b R ` (/)
there exists a nonzero polynomial p(x) with coecients in (/) such that
p(b) = 0. (A family / with this property is called a transcendental basis of
R over .)
3 A lter on a set X is a nonempty family T of subsets of X such that
(1) A B T provided A, B T; and (2) if A B X and A T then
B T. A lter T on X is proper if T = {(X), and it is prime if for every
A X either A T or X ` A T. Show that every proper lter on a set
X can be extended to a proper prime lter.
4 A graph on a (nite or innite) set V is an ordered pair 'V, E` such that
E [V ]
2
, where [V ]
2
is the set of all two-element subsets of V . For E
0
E
a graph 'V, E
0
` is a forest in graph 'V, E` if it does not contain any cycle,
that is, if there is no sequence v
0
, v
1
, . . . , v
n
= v
0
with n > 2 such that
v
i
, v
i+1
E
0
for i n. For E
0
E a graph 'V, E
0
` spans 'V, E` if for
every v V there is a w V such that v, w E
0
.
Let 'V, E` be a graph that spans itself. Show that there exists a forest
'V, E
0
` that spans 'V, E`.

Chapter 5
Cardinal numbers
5.1 Cardinal numbers and the continuum hy-
pothesis
We say that the sets A and B have the same cardinality and write A B
if there exists a bijection f : A B.
It is easy to see that for every A, B, and C
A A;
if A B then B A; and
if A B and B C then A C.
Thus the relation of having the same cardinality is an equivalence re-
lation. We put the word relation in quotation marks since our denition
does not specify any eld for . This means that the eld of this rela-
tion is the class of all sets. However, the restriction of to any set X is
an equivalence relation on X. Thus we will use the term relation for in
this sense.
By Zermelos theorem (Theorem 4.3.3) and Theorem 4.2.4, for every set
A there exists an ordinal number such that A . The smallest ordinal
number with this property is called the cardinality of A and is denoted by
[A[. Thus
[A[ = min: is an ordinal number and A .
In particular, A [A[ for every A. It is not dicult to see that A B if
61

62 5 Cardinal numbers
and only if [A[ = [B[. Thus the terminology sets A and B have the same
cardinality can be used for A B as well as for [A[ = [B[.
Notice also that
[[ for every ordinal number ,
since and [[ is the smallest ordinal such that . This implies
also that
[[ [[ for all ordinal numbers .
An ordinal number is said to be a cardinal number (or just a cardinal )
provided = [A[ for some set A. We can distinguish the cardinal numbers
from all other ordinal numbers by using the following properties.
Proposition 5.1.1 For an ordinal number the following conditions are
equivalent:
(i) is a cardinal number;
(ii) [[ = ;
(iii) < [[ for every < ;
(iv) [[ < [[ for every < ;
(v) [[ = [[ for every < .
Proof (i)(ii): If is a cardinal number then there exists a set A such
that [A[ = . This means, in particular, that A , that is, [A[ = [[. So
[[ = [A[ = .
(ii)(iii): It is obvious.
(iii)(iv): It follows from the fact that [[ for every ordinal .
(iv)(v): It is obvious.
(v)(i): Condition (v) implies that for every < . Thus is
the smallest ordinal such that . But this means, by the denition
of cardinality, that is a cardinal number, as then = [[.
Property (v) of Proposition 5.1.1 explains why we often say that the
cardinal numbers are the initial ordinals, that is, ordinals that are the
smallest of a given cardinality.
Let us also note the following theorem.

5.1 Cardinal numbers and the continuum hypothesis 63
Theorem 5.1.2 Let A and B be arbitrary sets. The following conditions
are equivalent:
(i) [A[ [B[;
(ii) there exists a one-to-one function g : A B.
Moreover, if A = then these conditions are equivalent to the condition
that
(iii) there exists a function f : B A from B onto A.
Proof For A = the equivalence of (i) and (ii) is obvious. So assume that
A = .
(i)(iii): Let h: B [B[ and g : [A[ A be bijections. Pick a A.
Notice that [A[ [B[ and h
1
([A[) B.
Dene f : B A by
f(b) =

g(h(b)) for b h
1
([A[),
a otherwise.
Then
A f[B] f[h
1
([A[)] = (g h)[h
1
([A[)] = g[[A[] = A.
So f is onto A.
(iii)(ii): Let f : B A be onto. Let G be a choice function for the
family f
1
(a): a A. Then g(a) = G(f
1
(a)) is a one-to-one function
from A into B.
(ii)(i): Let g : A B be a one-to-one function and h: B [B[
be a bijection. Then (h g): A [B[ is a one-to-one function from A
onto (h g)[A] [B[. Hence [A[ = [(h g)[A][ and, by Corollary 4.2.6,
Otp((h g)[A]) [B[. So
[A[ = [(h g)[A][ = [ Otp((h g)[A])[ Otp((h g)[A]) [B[.
Corollary 5.1.3 If A B then [A[ [B[.
Proof Since the identity function from A into B is one-to-one, the corol-
lary follows immediately from Theorem 5.1.2.
Corollary 5.1.4 If A B C and [A[ = [C[ then [B[ = [A[.
Proof By Corollary 5.1.3 we have [A[ [B[ [C[ = [A[. So [B[ = [A[.
Remark It is often the case that the relation A _ B for a set A having
cardinality less than or equal to the cardinality of B, [A[ [B[, is dened

64 5 Cardinal numbers
as in Theorem 5.1.2(ii), that is, by saying that there exists a one-to-one
function f : A B. Then it is necessary to prove that A _ B and B _ A
imply that there exists a bijection from A onto B. This fact is known as the
Schr oderBernstein theorem and follows immediately from our denition
and Theorem 5.1.2.
Theorem 5.1.5 is a cardinal number and every n is a cardinal
number.
Proof First we show that every n is a cardinal number. So let n .
By Proposition 5.1.1(iv) it is enough to prove that [k[ < [n[ for every k < n.
Thus it is enough to show that the inequality [n[ [k[ is false, that is, by
Theorem 5.1.2, that
there is no one-to-one function f : n k for every k < n. (5.1)
We will prove (5.1) by induction on n < .
Notice that for n = 0 condition (5.1) is true, since there is no k < n.
So assume that it is true for some n < . We will show that this implies
(5.1) for n + 1.
To obtain a contradiction assume that for some k < n+1 there is a one-
to-one function f : (n+1) k. If k < n then f[
n
: n k contradicts (5.1).
So k = n, f : (n +1) n, and f[n +1] (n 1), that is, n 1 f[n +1].
(Notice that n1 exists, since the existence of f : (n+1) n implies that
n = 0.)
Let i < n+1 be such that f(i) = n1. Dene g : n n1 by putting
g(j) = f(j) for j < n, j = i, and g(i) = f(n). (If i = n this last condition is
redundant and can simply be ignored.) Notice that the values of g indeed
belong to n 1 and that g is one-to-one. Thus g contradicts (5.1) for n.
This contradiction nishes the proof of (5.1) and the fact that every n
is a cardinal number.
To see that is a cardinal number notice that n (n+1) for every
n < . Hence, by Corollary 5.1.3, [n[ [n+1[ [[. But by what we have
already proved, [n[ = n < n+1 = [n+1[. So [n[ < [n+1[ [[, that is, we
conclude that [n[ < [[ for every n < . Therefore, by Proposition 5.1.1,
we deduce that is a cardinal number.
The sets with cardinality less than , that is, equal to some n < , are
called nite sets. A set A is countable if [A[ = . A set is innite if it is
not nite. A set is uncountable if it is innite and not countable.
We have already established a convention that general innite ordinal
numbers are denoted by the greek letters , , , , , . The innite cardi-
nal numbers will usually be denoted by the letters , , .
The natural numbers and are cardinal numbers. Are there any other
(i.e., uncountable) cardinal numbers? The next theorem gives us a tool to
construct a lot of uncountable cardinal numbers.

5.1 Cardinal numbers and the continuum hypothesis 65
Theorem 5.1.6 (Cantors theorem) [X[ < [{(X)[ for every set X.
Proof Fix a set X. First notice that [X[ [{(X)[, since the function
f : X {(X) dened by f(x) = x for x X is one-to-one.
To nish the proof it is enough to show that [X[ = [{(X)[. So let
f : X {(X). We will prove that f is not a bijection by showing that f
is not onto {(X). To see it, put
Y = x X: x f(x).
Then Y {(X) is not in the range of f, since if there were a z X such
that f(z) = Y then we would have
z Y z x X: x f(x) z f(z) z Y,
a contradiction.
Cantors theorem tell us, in particular, that for any cardinal number
there is a cardinal number larger than , namely, = [{()[. We will
denote this cardinal number by 2

. That way, the symbol 2

will stand
for two objects: a cardinal number and the family of all functions from
into 2. Although these two object are denitely dierent, this notation is
consistent in the sense that
[{(X)[ =

2
X

for every set X. (5.2)

This last equation is established by a bijection

: {(X) 2
X
dened by

(A) =

A
, where

A
is the characteristic function of A, that is,

A
(x) = 1
for x A and

A
(x) = 0 for x X ` A.
The notation 2

is also consistent with the cardinal exponentiation op-

eration that will be introduced in the next section.
Notice also that for every cardinal number there is the smallest car-
dinal number greater than , namely, a cardinal equal to the minimum
of the set
K = 2
2

: is a cardinal and < .

The minimum exists, since K is a nonempty (2

K) subset of a well-
ordered set 2
2

. This unique cardinal number is denoted by

+
and is
called the cardinal successor of .
It is also worthwhile to note the following property of cardinal numbers.
Proposition 5.1.7 If T is a family of cardinal numbers then =

T is
also a cardinal number.
Proof By Theorem 4.2.3(iii) is an ordinal number. Moreover, if <
then and there exists a T such that . So < = [[ [[.
Thus, by Proposition 5.1.1(iii), is a cardinal number.

66 5 Cardinal numbers
With the use of Proposition 5.1.7 and the operation of cardinal successor
we can construct for every ordinal number a cardinal number

by
induction on in the following way:

0
= ;

+1
= (

)
+
for every < ;

<

for every limit .

It is easy to see that

is the th innite cardinal number. It is also

worthwhile to mention that very often

is denoted by

. ( is a Hebrew
letter pronounced aleph.)
Similarly, for every ordinal number we can construct a cardinal num-
ber Z

by induction on as follows:
Z
0
= ;
Z
+1
= 2

for every < ;

Z

<
Z

for every limit .

(Z is a Hebrew letter, which we read bet.) In particular, the cardinality of
{() is equal to Z
1
= 2

. This number is usually denoted by c, a Hebrew

letter c, which we read continuum.
Obviously,
0
= = Z
0
and, by Cantors theorem,
1
Z
1
= c.
It is also not dicult to prove that

for every ordinal number

. Two natural questions that arise in this context are the following: Is

1
= c? Is

= Z

for every ordinal number ? Surprisingly, both these

questions cannot be decided within ZFC set theory. The statement that

1
= c is called the continuum hypothesis (usually abbreviated by CH) and
is independent of ZFC set theory (see Section 1.1). Similarly, the statement

= Z

for every ordinal number is called the generalized continuum

hypothesis (usually abbreviated by GCH) and is also independent of ZFC
set theory.
The independence of CH from the ZFC axioms will be proved in Chap-
ter 9. Meanwhile, we will often use CH as an additional assumption to
deduce other properties of interest. That way, we will know that the de-
duced properties are consistent with ZFC set theory, that is, that we have
no way to prove their negations using only ZFC axioms.
The last theorem of this section gives some rules on comparing the
cardinalities of sets.

5.1 Cardinal numbers and the continuum hypothesis 67
Theorem 5.1.8 If [A[ = [B[ and [C[ = [D[ then
(a) [AC[ = [B D[;
(b) [A C[ = [B D[, provided A C = = B D;
(c)

A
C

B
D

.
Proof Let f : A B and g : C D be bijections. Dene the bijections
for (a)(c) as follows:
(a) F : AC B D, F(a, c) = 'f(a), g(c)`.
(b) F : A C B D, F = f g.
(c) F : A
C
B
D
, and F(h) B
D
for h A
C
is dened by the formula
F(h)(d) = f(h(g
1
(d))) for every d D.
It is left as an exercise to show that the functions F so dened are
indeed bijections.
We will nish this section with the following remark. The only cardinal
numbers that were dened in a natural way (i.e., without an essential
use of the axiom of choice) are the natural numbers and . The innite
ordinal numbers that we can easily construct, that is, construct with the
use of ordinal number operations, are of the form + 1, + 2, + ,
, and so forth, and are not cardinal numbers. (For example, notice that
[[ by dening the one-to-one function f : with the formula
f(n, m) = 2
n
3
m
, where we identify with the set from the denition
of .) In fact, there is no way to prove without the axiom of choice the
existence of uncountable ordinal numbers. This might seem strange, since
the proof of Cantors theorem evidently did not use the axiom of choice.
However, we in fact proved there only that [X[ = [{(X)[. The proof that
the set {(X) can indeed be well ordered requires the use of Zermelos
theorem, which is equivalent to the axiom of choice.
EXERCISES
1 Let n . Show that if is an ordinal number such that [[ = n then
= n.
2 Prove that

for every ordinal number .

3 Complete the proof of Theorem 5.1.8 by showing that all functions F
dened in this proof are bijections.

68 5 Cardinal numbers
5.2 Cardinal arithmetic
For cardinal numbers and dene their cardinal sum by
= [( 0) ( 1)[,
their cardinal product by
= [ [ ,
and a cardinal exponentiation operation by

.
Notice that the exponentiation operation for = 2 gives us another def-
inition of 2

. However, it is consistent with the previous denition, since

= 2

. Notice also that by the foregoing denitions and Theorem 5.1.8,

for all sets A and B,
[A B[ = [A[ [B[ for A and B disjoint (5.3)
and
[AB[ = [A[ [B[,

A
B

= [A[
]B]
. (5.4)
Proposition 5.2.1 For all cardinal numbers and
(i) = ,
(ii) = .
Proof From Theorem 5.1.8 we have
= [( 0) ( 1)[ = [( 0) ( 1)[ = ,
where the second equation follows from the fact that [ 0[ = [ 1[
and [ 1[ = [ 0[. Also,
= [ [ = [ [ = ,
since f(, ) = ', ` is a bijection between and .
Thus, unlike the ordinal operations + and , the cardinal operations
and are commutative.

5.2 Cardinal arithmetic 69
Proposition 5.2.2 For every m, n
(i) mn = m+n < ,
(ii) mn = mn < .
The proof is left as an exercise.
Lemma 5.2.3 Every innite cardinal number is a limit ordinal number.
Proof First notice that
[ x[ = [[.
If x then this is obvious. Otherwise this is justied by a function
f : x dened by f(x) = 0 and f(n) = n + 1 for n < . Hence,
by Theorem 5.1.8(ii), for every innite ordinal number
[ + 1[ = [ [ = [( ` ) ( )[ = [( ` ) [ = [[.
Therefore, by Proposition 5.1.1, an ordinal number + 1 is not a cardinal
number for any innite .
Theorem 5.2.4 If is an innite cardinal then = .
Proof Let be an innite cardinal number. We will prove by transnite
induction on that
[[ [[ = [[. (5.5)
So let be such that (5.5) holds for every < . We will
show that (5.5) holds for as well. This will nish the proof.
If there is an < such that [[ = [[ then
[[ [[ = [[ [[ = [[ = [[.
Hence we may assume that is a cardinal number.
Now notice that for every <
[ [ < . (5.6)
It follows from Proposition 5.2.2(ii) for a nite and from (5.5) for an
innite . Dene a well ordering _ on by putting ', ` _ ', ` if
and only if
max, < max,
or
max, = max, and ', `
lex
', `,

70 5 Cardinal numbers
where
lex
stands for the lexicographic order on (see Theorem 4.1.13).
The proof that _ is indeed a well ordering is left as an exercise.
Now let ', ` and let = max + 1, + 1. Then <
by Lemma 5.2.3. Moreover, the initial segment O
_
(', `) (generated by
', ` with respect to _) is a subset of , since ', ` _ ', ` implies
that max, max, < . Hence, by (5.6),
[ Otp(O
_
(', `))[ = [O
_
(', `)[ [ [ = [[ <
and so also
Otp(O
_
(', `)) < .
(Otherwise, we would have = [[ [ Otp(O
_
(', `))[.) Since every
ordinal number Otp( , _) is of the form Otp(O
_
(', `)) we can
conclude that Otp( , _) . So
[ [ = [ Otp( , _)[ Otp( , _) = [[.
Thus [ [ = [[, since the inequality [[ [ [ is obvious.
Corollary 5.2.5 If and are innite cardinals then
= = max, .
Proof We may assume that . Then
= [( 0) ( 1)[ [ 2[ [ [ =
and
= [ [ [ [ = .
Since evidently we obtain
.
Corollary 5.2.6 A countable union of countable sets is countable.
Proof Let T be a countable family of countable sets, T = F
n
: n < .
Then for every n < there is a bijection f
n
: F
n
. Dene function
f :

T by putting f(n, m) = f
n
(m). Notice that f is onto. So,
by Theorem 5.1.2, [

T[ [ [ = . Hence [

T[ = .
Similarly, we can prove the following.
Corollary 5.2.7 If is an innite cardinal and [X

[ for all <

then

<
X

.

5.2 Cardinal arithmetic 71
The proof is left as an exercise.
Corollary 5.2.8 A nite union of nite sets is nite.
The proof is similar to that of Corollary 5.2.6 (use Proposition 5.2.2).
For a set A let A
<
=

n<
A
n
. Thus A
<
is the set of all nite
sequences with values in A.
Corollary 5.2.9 If is an innite cardinal, then [
<
[ = .
Proof It is easy to prove, by induction on n < , that [
n
[ = for every
0 < n < . So, by Corollary 5.2.7,

<

n<

.
Corollary 5.2.10 [[ = [Z[ = .
Proof Clearly = [N[ [Z[ [[. To see that [[ it is enough to
show that [[ [2 [, since [2 [ = 2 = . So let
f : 2 be dened by f(i, m, n) = (1)
i m
n+1
. It is easy to see
that f is onto .
Theorem 5.2.11 [R[ = [2

[ = [{()[ = c.
Proof [2

[ = 2

= c by the denition of c. [2

[ = [{()[ was proved in

(5.2). The equation [R[ = [2

[ is proved by two inequalities.

The rst inequality [2

[ [R[ is established by a one-to-one function

f : 2

R dened by
f('a
0
, a
1
, . . .`) =

n<
2a
n
3
n+1
.
(The range of f is Cantors classical ternary set; see Section 6.2.)
To see that [R[ [2

[ dene a one-to-one function g : R {() by

g(r) = q : q < r (i.e., g(r) is the Dedekind cut associated with r).
So
[R[ [{()[ =

2
Q

= 2
]Q]
= 2

.
Theorem 5.2.12 If and are cardinal numbers such that and
2 then

= 2

.
In particular,

= 2

for every innite cardinal number .

72 5 Cardinal numbers
Proof Notice that
2

{( ).
Hence
2

[{( )[ =

= 2
]]
= 2

.
Theorem 5.2.13 If , , and are cardinals, then

and (

.
Proof It follows immediately from the following properties:
A
(BC)
(A
B
) (A
C
) for B C = , (5.7)
and
(A
B
)
C
A
BC
. (5.8)
Their proof is left as an exercise.
For a set X and a cardinal number dene
[X]

= A {(X): [A[ , [X]

<
= A {(X): [A[ < ,
[X]

= A {(X): [A[ = .
Proposition 5.2.14 For every innite set Xand nonzero cardinal [X[
[[X]

[ =

[X]

= [X[

.
Proof Dene : X

[X]

by (f) = f[]. Since is clearly onto,

and [X]

[X]

, we have
[[X]

[X]

[X[

.
To nish the proof it is enough to show that [X[

[[X]

[.
To prove it dene : X

[ X]

by (f) = f. Clearly is one-to-

one. Hence
[X[

= [X

[
[[ X]

[
= [[[ X[]

[ (5.9)
= [[[X[]

[
= [[X]

[ , (5.10)
where equations (5.9) and (5.10) follow from the fact that [Y [ = [Z[ implies
[[Y ]

[ = [[Z]

[, which is left as an exercise.

We will nish this section with the following important fact.

5.2 Cardinal arithmetic 73
Theorem 5.2.15 The family ((R) = f R
R
: f is continuous has car-
dinality continuum.
Proof Dene a function R: ((R) R
Q
by R(f) = f[
Q
. By the density of
in R and the continuity of our functions it follows that R is one-to-one
(see Section 3.3). Hence
[((R)[

R
Q

= [R[
]Q]
= [2

= (2

= 2

= 2

.
The other inequality is proved by a one-to-one function c: R ((R), where
c(a): R R is a constant function with value a.
EXERCISES
1 Use the denitions of cardinal arithmetic operations and Theorem 5.1.8
to prove (5.3) and (5.4).
2 Prove Proposition 5.2.2. Hint: Show by induction on n that m+n <
and mn < . Then use Exercise 1 from Section 5.1.
3 Prove that the relation _ dened in the proof of Theorem 5.2.4 is a
well-ordering relation on .
4 Prove Corollary 5.2.7.
5 Complete the proof of Theorem 5.2.13 by proving (5.7) and (5.8).
6 Complete the proof of Proposition 5.2.14 by showing that [Y [ = [Z[
implies [[Y ]

[ = [[Z]

[ for every sets Y and Z and any cardinal number .

7 A real number r is an algebraic number if there is a polynomial p(x)
with rational coecients such that p(r) = 0. Show that the set A R
of all algebraic numbers is countable. Conclude that the set R ` A of all
nonalgebraic real numbers has cardinality continuum.
8 A function f : R R is a Baire class-one function if there exists a
sequence f
n
: R R of continuous functions such that f
n
converges to f
pointwise, that is, such that lim
n
f
n
(x) = f(x) for every x R. Find
the cardinality of the family B
1
of all Baire class-one functions.

74 5 Cardinal numbers
5.3 Conality
We say that a subset A of an ordinal number is unbounded in if A has
no strict upper bound in , that is, when there is no with <
for all A. It is also easy to see that A is unbounded in when
=

+ 1: A or when
< A ( ).
For an ordinal number let cf() be the smallest ordinal number
such that there exists a function f : with the property that f[] is
unbounded in . Such a function f is called a conal map (in ). The
number cf() is called the conality of . Thus f : is conal in if
( f())
and
cf() = min: there is a conal map f : .
Notice that
cf() for every ordinal number ,
since the identity map is a conal map. Also,
cf( + 1) = 1 for every ordinal number ,
since f : 1 +1, f(0) = , is conal in +1. It is also easy to see that
if cf() = 1 then has a greatest element. Thus cf() = 1 if and only if
cf() < if and only if is a successor ordinal.
We will be interested mainly in conalities of limit ordinal numbers.
Lemma 5.3.1 For every ordinal number there is a strictly increasing
conal map f : cf() .
Proof For a successor ordinal the map just shown works.
If is a limit ordinal and g : cf() is a conal map, dene by
transnite induction on < cf()
f() = max

g(),

<
(f() + 1)

.
Clearly f() < f() + 1 f() for every < < cf(). Thus f is strictly
increasing.

5.3 Conality 75
The fact that
f() for every < cf() (5.11)
is proved by induction on < cf(). If for some < cf() condition (5.11)
holds for all < then f() = maxg(),

<
(f() + 1) . But f()
cant be equal to since otherwise the restriction of f to would be a
conal map in and this would contradict the minimality of cf(). Thus
< cf(). Condition (5.11) has been proved.
Finally, f is conal in since for every there is an < cf() with
g() f().
Corollary 5.3.2 cf(cf()) = cf() for every ordinal number .
Proof Let f : cf(cf()) cf() and g : cf() be strictly increasing
conal maps. Then g f : cf(cf()) is also conal, since for every
< there exist cf() with g() and cf(cf()) with f(),
so
g() g(f()) = (g f)().
Thus, by the minimality of cf(), we have cf(cf()) cf(). So cf(cf()) =
cf().
An ordinal number is regular if it is a limit ordinal and cf() = .
Theorem 5.3.3 If is a regular ordinal number then is a cardinal num-
ber.
Proof Put = [ cf()[, and choose a bijection f : cf() and a conal
function g : cf() . Then (gf): is conal in , since (gf)[] =
g[f[]] = g[cf()]. Hence, by minimality of cf(), cf() = [ cf()[.
Proposition 5.3.4 is regular.
Proof It follows easily from Corollary 5.2.8.
Theorem 5.3.5
+
is regular for every innite cardinal number .
Proof Let <
+
and f :
+
. Then [[ and [f()[ for every
< . Hence [

<
[f() +1][ since a union of sets of cardinality
has cardinality (see Corollary 5.2.7). Thus

<
[f() + 1] =
+
,
that is, cf(
+
) = for every <
+
. So cf(
+
) =
+
.
One of the most useful properties of regular cardinals is given in the
following proposition.

76 5 Cardinal numbers
Proposition 5.3.6 If is a regular cardinal, A , and [A[ < then
there is an < such that A .
Proof Let = Otp(A) and let f : A be an order isomorphism.
Notice that < cf(), since [[ = [A[ < = cf(). So the set A = f[]
cannot be unbounded in . This implies the existence of an < such
that < for all A. Thus A .
Another useful fact is the following.
Proposition 5.3.7 If <
1
is a limit ordinal, then cf() = .
The proof is left as an exercise.
We will nish this section with the following cardinal inequality.
Theorem 5.3.8 For every innite cardinal

cf()
> .
Proof Let f : cf() be a conal map and let G:
cf()
. We will
show that G is not onto
cf()
.
Dene h: cf() by
h() = min( ` G()(): < f())
for < cf(). Such a denition makes sense, since [G()(): < f()[
[f()[ < . But for every < there exists an < cf() such that
< f(), that is, such that h() = G()(). Hence h = G() for every
< and h G[].
Corollary 5.3.9 cf(c) > .
Proof Assume, to obtain a contradiction, that cf(c) = . Then c
cf(c)
=
(2

= 2

= c, contradicting Theorem 5.3.8.

EXERCISE
1 Prove Proposition 5.3.7.

Part III
The power of recursive
denitions
77

Chapter 6
Subsets of R
n
6.1 Strange subsets of R
n
and the diagonal-
ization argument
In this section we will illustrate some typical constructions by transnite
induction. We will do so by constructing recursively some subsets of R
n
with strange geometric properties. The choice of geometric descriptions of
these sets is not completely arbitrary here we still have not developed
the basic facts concerning nice subsets of R
n
that are necessary for most
of our applications. This will be done in the remaining sections of this
chapter.
To state the next theorem, we will need the following notation. For a
subset A of the plane R
2
the horizontal section of A generated by y R
(or, more precisely, its projection onto the rst coordinate) will be denoted
by A
y
and dened as A
y
= x R: 'x, y` A. The vertical section of A
generated by x R is dened by A
x
= y R: 'x, y` A.
We will start with the following example.
Theorem 6.1.1 There exists a subset A of the plane with every horizontal
section A
y
being dense in R and with every vertical section A
x
having
precisely one element.
Proof We will dene the desired set by induction. To do so, we will rst
reduce our problem to the form that is most appropriate for a recursive
construction.
The requirement that every horizontal section of A is dense in R tells
us that A is reasonably big. More precisely, the condition says that
A
y
(a, b) = for every a < b and y R. Therefore if we put T =
79

80 6 Subsets of R
n
(a, b) y: a, b, y R & a < b then we can restate it as
A J = for every J T. (6.1)
The requirement that every vertical section of A has precisely one point
tells us that A is reasonably small. It can be restated as
[A [x R][ = 1 for every x R. (6.2)
The idea of constructing A is to grow it bit by bit to obtain at the
end a set satisfying (6.1), while keeping approximations reasonably small.
More precisely, by transnite induction on < c, we will dene a sequence
''x

, y

`: < c` such that A

0
= 'x

, y

`: < c satises (6.1) and has the

property that [A
0
[x R][ 1 for every x R. It is enough, since then
the set A = A
0
'x, 0`: A
0
[x R] = satises the theorem.
To describe our construction of the sequence, notice rst that [T[ = c,
since c = [R[ [T[ [R
3
[ = c. Let J

: < c be an enumeration of
the family T. If for some < c the sequence ''x

, y

`: < ` is already
dened, choose
'x

, y

` J

<
x

. (6.3)
The choice can be made, since for every < c we have [J

[ = [R[ = c >
[[

<
x

. So, by the recursion theorem, the sequence

''x

, y

`: < c` exists. (More precisely, we dene 'x

, y

` = f() = h(f[

)
where the oracle function h is dened implicitly by formula (6.3). Explicitly,
h(''x

, y

`: < `) is dened as C(J

` 'x

, y

`: < ), where C is a
choice function from the family {(R
2
) ` .)
It is clear that A
0
= 'x

, y

`: < c satises (6.1), since for every

J T there exists a < c such that J = J

and 'x

, y

` A
0
J

= A
0
J.
It is also easy to see that the choice as in (6.3) implies that every vertical
section of A
0
has at most one point.
The proof of Theorem 6.1.1 is a typical example of a diagonalization
technique. In order to prove the theorem we had to nd for every J T
a point 'x
J
, y
J
` J that would belong to a future set A. The choice was
supposed to be done while preserving at every step the property that no
two points chosen so far belong to the same vertical line. In order to prove
the theorem we listed all elements of T and made our construction, taking
care of each of its elements one at a time, using the fact that we still had
enough room to make our choice. This technique of making inductive
constructions by listing all important objects for the desired property and

6.1 Strange subsets of R
n
and the diagonalization argument 81
then taking care of each of them one at a time is called the diagonalization
technique. It is typical for such constructions that the inductive step is
possible because of some cardinal argument similar to the one given earlier.
The preservation part, like that mentioned before, is usually the hard part
of the argument.
The next theorem, due to Mazurkiewicz (1914), is very similar in char-
acter to Theorem 6.1.1.
Theorem 6.1.2 There exists a subset A of the plane R
2
that intersects
every straight line in exactly two points.
Proof Let L

: < c be an enumeration of all straight lines in the

plane R
2
. By transnite induction on < c we will construct a sequence
A

: < c of subsets of R
2
such that for every < c
(I) A

has at most two points;

(P)

does not have three collinear points;

(D)

contains precisely two points of L

.
Then the set A =

<c
A

will have the desired property. This is the

case since the preservation condition (P) implies that every line contains
at most two points of A, while the diagonal condition (D) makes sure that
every line L contains at least two points from A. Thus, it is enough to show
that we can choose a set A

satisfying (I), (P), and (D) for every < c.

This will be proved by induction on < c.
So assume that for some < c the sequence A

: < is already
constructed. By condition (I) the set B =

<
A

has cardinality < c,

being a union of [[ < c many nite sets. Similarly, the family ( of all lines
containing two points from B has cardinality [B
2
[ < c. Notice that by
the inductive assumption (P) the set B L

has at most two points. If

it has precisely two points put A

= and notice that (P) and (D) are

satised. If B L

contains less than two points then L

intersects every
line from ( in at most one point. Thus L

( =

L(
L

L has at
most [([ < c many points.
Choose A

( to have two elements if B L

= and to have
one element if B L

has one point. It is easy to see that conditions (P)

and (D) are satised with either choice of A

. The construction has now

been completed.
The next theorem is yet another example of a geometric problem solved
with help of the recursion theorem. In this theorem a circle in R
3
is un-
derstood to be a set of points forming any nontrivial circle in any plane
in R
3
.

82 6 Subsets of R
n
Theorem 6.1.3 R
3
is a union of disjoint circles.
Proof We have to construct a family ( of subsets of R
3
such that (1) each
C ( is a circle, (2) the family ( covers R
3
, and (3) dierent elements of
( are disjoint. We will construct the family ( as C

: < c by induction
on < c. For this, the natural approach is the following. Choose an
enumeration p

: < c of R
3
, and for each p

choose a circle C

such that
p

. This certainly would take care of (2); however, then we could not
keep circles disjoint: For p

C
0
` p
0
the circles C
0
and C

would not
be disjoint. Thus we have to settle for a slightly weaker condition: We will
choose C

such that
(D) p

.
This is the diagonal condition that will imply that ( covers R
3
. However,
we have to make the choice in such a way that circles in ( are pairwise
disjoint. We will do this by assuming that for each < c the circles
constructed so far are pairwise disjoint and we will choose C

preserving
this property. More precisely, we will choose C

such that the following

preservation condition is satised:
(P) C

<
C

= .
Evidently, if we could construct ( = C

: < c such that conditions (D)

and (P) are satised for every < c, then ( would be the desired family of
circles.
On the other hand, the recursion theorem tells us that we can nd such
a sequence 'C

: < c` as long as the family of all circles C

satisfying (D)
and (P) is nonempty for every < c. (Since then the choice function will
work as an oracle function.)
So assume that for some < c the sequence C

: < is already
constructed satisfying (D) and (P). We will prove that there exists a circle
C

satisfying (D) and (P).

If p

<
C

dene p = p

. Otherwise, choose an arbitrary point

p R
3
`

<
C

. This can be done since for every straight line L in R

3
the
set L

<
C

<
(LC

) has cardinality less than continuum, being

a union of less than continuum many nite sets LC

. We will choose C

containing p and satisfying (P). To do this, take a plane P containing p that

does not contain any of the circles C

for < . This can be done since

there are continuum many planes passing through p and there are only
[[ < c many planes containing circles from C

: < . Now notice

that the plane P intersects each circle C

in at most two points. Thus the

set S = P

<
C

has cardinality less than continuum. Fix a line L in

P containing p and let (
0
be the family of all circles in P containing p and

6.1 Strange subsets of R
n
and the diagonalization argument 83
tangent to L. Notice that dierent circles from (
0
intersect only at the
point p. Thus there is a circle C

(
0
disjoint from S. This nishes the
inductive construction and the proof.
If in the previous three theorems the reader got the impression that we
can construct a subset or a partition of R
n
with almost arbitrary paradox-
ical geometric properties, the next theorem will prove that this impression
is wrong.
Theorem 6.1.4 The plane R
2
is not a union of disjoint circles.
Proof Assume, to obtain a contradiction, that there is a family T of
disjoint circles in R
2
such that R
2
=

T. Construct, by induction on
n < , a sequence C
n
: n < of circles from T in the following way:
Start with an arbitrary circle C
0
T and in step n + 1 choose a circle
C
n+1
T that contains the center c
n
of circle C
n
.
Notice that if r
n
is the radius of C
n
then [c
n+1
c
n
[ = r
n+1
< r
n
/2.
Thus 'c
n
: n < ` is a Cauchy sequence. Let p = lim
n
c
n
. Then p
belongs to the closed disk D
n
bounded by C
n
for every n < . So p cannot
belong to any circle C
n
since it belongs to D
n+1
, which is disjoint from C
n
.
Now let C T be such that p C. Then C = C
n
for every n < . But
if n < is such that r
n
is smaller than the radius of C then C C
n
= .
This contradicts the choice of T.
The next theorem shows that a simple diagonalization may lead to
messy multiple-case considerations. In the theorem R
2
will be considered
as a vector space. In particular, A + r will stand for the algebraic sum of
a set A R
2
and r R
2
, that is, A+r = a +r : a A.
Theorem 6.1.5 There is a subset A of R
2
that intersects every one of its
translations in a singleton, that is, such that the set (A+r)A has precisely
one element for every r R
2
, r = 0.
Proof Enumerate R
2
` 0 as r

: < c. We will dene, by induction

on < c, the sequences 'a

R
2
: < c`, 'b

R
2
: < c`, and 'A

: < c`
such that for every < c
(I) A

= a

: b

: ;
(D) b

= r

;
(P) if a, b, a
t
, b
t
A

are such that b a = b

t
a
t
= 0 then a = a
t
and
b = b
t
.
Before we describe the construction notice that A =

<c
A

has the de-

sired properties: (A + r) A has at least one element by the diagonal

84 6 Subsets of R
n
condition (D) for r = r

, since b

(A+r

)A; it has at most one element

by the preservation condition (P), where we take r = b a = b
t
a
t
.
To make an inductive step assume that for some < c the sequences
are already constructed. The diculty will be to nd points a

and b

satisfying (D) while preserving (P).

Let B =

<
A

. Notice that B satises (P).

If B (B + r

) = and b B (B + r

) then a = b r

B and we
can dene a

= a and b

= b. Condition (D) is satised by the choice of a

and b, and condition (P) holds, since A

= B.
So assume that B(B+r

) = and notice that if we nd a

= x then
b

= a

+r

= x+r

would be uniquely dened. Thus, it is enough to nd

x guaranteeing satisfaction of (P). In order to nd it we have to avoid the
situation
b a = b
t
a
t
= 0 and b = b
t
, (6.4)
where a, b, a
t
, b
t
A

= B x, x + r

. So take x, a, b, a
t
, b
t
that satisfy
(6.4).
There are several cases to consider. First notice that by our inductive
hypothesis a, b, a
t
, b
t
x, x +r

= .
In the rst group of cases we assume that all elements a, b, a
t
, b
t
are
dierent. Then there are the following possibilities:
a, b, a
t
, b
t
x, x + r

= x. Then x = c + d e for some c, d, e

B. To avoid this situation it is enough to take x outside the set S
1
=
c +d e: c, d, e B.
a, b, a
t
, b
t
x, x + r

= x + r

. Then x = c + d e r

for some
c, d, e B. To avoid this situation it is enough to take x outside the
set S
2
= c +d e r

: c, d, e B.
a, b, a
t
, b
t
x, x +r

= x, x +r

.
If a, b = x, x+r

then r

= b
t
a
t
. However, this would contradict
the fact that B (B + r

) = . Similarly x, x + r

cannot be equal
to a
t
, b
t
, a, a
t
, or b, b
t
.
If x, x + r

= a, b
t
or x, x + r

= a
t
, b then 2x = c + d r

for
some c, d B. To avoid this situation it is enough to take x outside the
set S
3
= (c +d r

)/2: c, d B.
The second group of cases consists of the situations in which some of
the elements a, b, a
t
, b
t
are equal to each other. It is easy to see that there
are only two possibilities for such equations: b = a
t
and a = b
t
. Moreover,
these equations cannot happen at the same time. Thus assume that a
t
= b

6.1 Strange subsets of R
n
and the diagonalization argument 85
and that the other elements are dierent (the case when a = b
t
is identical).
Then 2b = a +b
t
.
All the cases when b x, x + r

are identical to those previously

considered. So assume that b x, x + r

. Then there are the following

possibilities:
b = x. Then 2x = a +b
t
.
If x + r

a, b
t
then 2x = c + d for some c, d B. To avoid this
case it is enough to take x outside the set S
4
= (c +d)/2: c, d B.
If x + r

a, b
t
then 2x = c + x + r

for some c B. To avoid this

situation it is enough to take x outside the set S
5
= c +r

: c B.
b = x +r

. Then 2x = a +b
t
2r

.
If x a, b
t
then 2x = c + d 2r

for some points c, d B. To

avoid this particular situation it is enough to choose x outside the set
S
6
= (c +d 2r

)/2: c, d B.
If x a, b
t
then 2x = c + x 2r

for some c B. To avoid this case

it is enough to take x outside the set S
7
= c 2r

: c B.
It is easy to see that each of the sets S
i
for i = 1, . . ., 7 has cardinality
[[
3
< c. Thus we can take x from the set R
2
`

7
i=1
S
i
.
In all of the previous examples of this section we used preservation
conditions of nite character such as noncollinearity, nite intersections
of dierent circles, and so forth. This made obvious the fact that these
conditions were preserved when we took a union of previously constructed
sets. This approach, however, does not always work. For example, the next
theorem will state that R
2
is a union of countably many sets S
i
such that
no set S
i
contains two dierent points of rational distance apart. If we try
to prove this theorem starting with an arbitrary well ordering p

: < c
of R
2
and at step try to add the point p

to some set S
i
, then already at
step we might nd ourselves in trouble it might happen that for every
i < there is already an element p
n
S
i
, n < , of rational distance from
p

. If so, we could not assign p

to any S
i
. To solve these diculties we
will proceed in a dierent manner, to be described subsequently.
We will start with the following easy but very useful lemma that will
also be used very often in the rest of this text. To state it, we need the
following denitions. Let F : X
k
[X]

for some k < . We say that

a subset Y of X is closed under the action of F if F(y
1
, . . ., y
k
) Y for
every y
1
, . . ., y
k
Y . If f : X
k
X then Y X is closed under the action
of f if it is closed under the action of F, where F(x) = f(x). Finally, if
T is a family of functions from nite powers of X into either [X]

or X
then Y X is closed under the action of T if it is closed under the action
of F for every F T.

86 6 Subsets of R
n
Lemma 6.1.6 Let T be an at most countable family of functions from
nite powers of X into [X]

or X. Then
(a) for every Z X there exists a smallest subset Y of X closed under
the action of T and containing Z; this set, denoted by cl
J
(Z), has
cardinality less than or equal to [Z[ +; in particular, [ cl
J
(Z)[ = [Z[
for every innite Z;
(b) if [X[ = > then there exists an increasing sequence 'X

: < `
of subsets of X closed under the action of T such that X =

<
X

,
[X

[ < for all < , and X

=

<
X

for every limit ordinal

< .
Proof In the proof we will identify functions f : X X from T with
F : X [X]
1
, F(x) = f(x).
(a) Let = [Z[ + . Construct, by induction on n < , an increasing
sequence 'Z
n
[X]

: n < ` by putting Z
0
= Z and dening Z
n+1
as
Z
n

F(z
1
, . . . , z
m
): F T & domF = X
m
& 'z
1
, . . . , z
m
` Z
m
n
.
Notice that [Z
n+1
[ . This is the case since [Z
n
[ and the second
set is a union of a family of sets of cardinality indexed by a set of
cardinality [T[ [Z
n
[
<
.
Dene cl
J
(Z) =

n<
Z
n
. Obviously Z cl
J
(Z) X and [ cl
J
(Z)[
, as cl
J
(Z) is a union of countably many sets of cardinality at most .
To see that cl
J
(Z) is closed under the action of T take F T. If X
m
is the domain of F and z
1
, . . . , z
m
cl
J
(Z) then there is an n < such
that z
1
, . . . , z
m
Z
n
, and so F(z
1
, . . . , z
m
) Z
n+1
cl
J
(Z).
In order to prove that cl
J
(Z) is the smallest subset of X closed under the
action of T and containing Z let Y be another such set. Then Z
0
= Z Y
and by an easy induction we can prove that Z
n
Y for every n < . So
cl
J
(Z) =

n<
Z
n
Y .
(b) Enumerate X as x

: < and dene X

= cl
J
(x

: < ).
Then the equation X =

<
X

is obvious and [X

[ [[ + < for
< follows immediately from (a). The two other conditions are simple
consequences of the fact that cl
J
(Z) is the smallest subset of X containing
Z and being closed under the action of T.
The next theorem is due to Erd os and Hajnal (Erd os 1969).
Theorem 6.1.7 There is a countable partition S
i
: i < of R
2
such
that the distance between any two dierent points of the same set S
i
is
irrational.

6.1 Strange subsets of R
n
and the diagonalization argument 87
Proof By induction on = [X[ we are going to prove that every set
X R
2
can be decomposed as described in the theorem.
So let X R
2
be such that [X[ = and suppose that the foregoing
statement is true for every Y R
2
of cardinality < . We have to prove
the statement for X.
If then the statement is obvious, since we can put every element
of X in a dierent set S
i
. So assume > .
For every p, q dene F
pq
: (R
2
)
2
[R
2
]
<
by
F
pq
(x, y) = z R
2
: [x z[ = p & [y z[ = q.
Notice that indeed F
pq
(x, y) [R
2
]
<
since it has at most two points. Let
T = F
pq
: p, q and let X =

<
X

, where 'X

: < ` is a
sequence of subsets closed under the action of T as in Lemma 6.1.6(b).
We will dene a decomposition of X into sets S
i
by dening g : X
and S
i
= g
1
(i). (Thus the function g tells us that an element x X is
put into S
g(x)
.) The function g must have the property that
if g(x) = g(y) for dierent x, y X then [x y[ . (6.5)
We will dene the function g inductively on the sets X

for < .
Assume that for some < the function g is already dened on each
X

for all < . If is a limit ordinal, then g is already dened on

X

<
X

and it is easy to see that it satises property (6.5). Assume

that is a successor ordinal, say = + 1. Then g is dened on X

. We
have to extend our denition to the set Z = X
+1
` X

.
Since [Z[ [X
+1
[ < we can nd a function h: Z satisfying
(6.5). We might try to dene g on Z as h. However, this might not work,
since for z Z there might be an x X

such that [z x[ . But

X

is closed under the action of T. Thus for every z Z there is at most

one x
z
X

such that [z x
z
[ . So we will dene g on Z such that
for every z Z and h(z) = n we have g(z) 2n, 2n + 1. Then g[
Z
will
satisfy (6.5), as h did. Now it is enough to choose g(z) dierent from g(x
z
),
if x
z
exists, and arbitrarily otherwise.
It can be proved also that R
n
can be decomposed as in Theorem 6.1.7
for all n N. However, the proof for n > 2 is considerably more dicult.
The last theorem of this section is due to Sierpi nski (1919). Although
its proof is essentially easier that those of previous theorems it has its
own avor the set constructed in the theorem exists if and only if the
continuum hypothesis is assumed.

88 6 Subsets of R
n
Theorem 6.1.8 The continuum hypothesis is equivalent to the existence
of a subset A of R
2
with the property that
[A
y
[ and

R
2
` A

for every x, y R.
Proof : First assume the continuum hypothesis c =
1
and let _ be a
well ordering of R of type
1
. (To nd it, take a bijection f : R
1
and
dene x _ y if and only if f(x) f(y).) Dene A = 'x, y`: x _ y.
Now A
y
= x R: 'x, y` A = x R: x _ y is an initial segment
of a set with order type
1
, so it has cardinality <
1
, that is, .
Similarly,
(R
2
` A)
x
= ('x, y` R
2
: 'x, y` A)
x
= ('x, y` R
2
: y x)
x
= y R: y x
is an initial segment of a set with order type
1
, that is, has cardinality
.
: Assume that [R[ = c >
1
and let A R
2
be such that [A
y
[ for
every y R. We will show that the complement of A has an uncountable
vertical section.
Let Y R be such that [Y [ =
1
(if f : c R is a bijection, take
Y = f[
1
]). Let X =

yY
A
y
. Then [X[
1
as it is a union of
1
sets
of cardinality <
1
. Take x R ` X. Then 'x, y` A for every y Y ,
since x A
y
. Hence x Y R
2
` A and so Y (R
2
` A)
x
. Therefore

(R
2
` A)
x

1
> .
EXERCISES
1 Complete the proof of Lemma 6.1.6(b) by showing that for every limit
ordinal <
1
the set X

=

<
X

is closed under the action of T as

long as every X

is closed under the action of T for < .

2 Prove that R
3
`
3
is a union of disjoint lines.
3 Generalize Theorem 6.1.3 by proving that R
3
is a union of disjoint circles
of radius 1.
4 Modify Mazurkiewiczs theorem (Theorem 6.1.2) by proving that there
exists a subset A of the plane R
2
that intersects every circle in exactly three
points.

6.2 Closed sets and Borel sets 89
5 Modify Sierpi nskis theorem (Theorem 6.1.8) by proving that the contin-
uum hypothesis is equivalent to the following statement: There exist sets
A
0
, A
1
, A
2
R
3
such that R
3
= A
0
A
1
A
2
and
['x
0
, x
1
, x
2
` A
i
: x
i
= x[
for every i < 3 and x R.
Try to generalize this and Theorem 6.1.8 to R
n
for all 1 < n < . Prove
your claim.
6 (Challenging) Prove the following theorem of Ceder: There is a decom-
position S
i

i<
of R
2
with no S
i
spanning an equilateral triangle. Remark:
It is also possible to nd a decomposition S
i

i<
of R
2
with no S
i
spanning
an isosceles triangle. This has been proved recently by Schmerl.
6.2 Closed sets and Borel sets
In what follows we will concentrate on the topological structure of R
n
.
Recall that
n
is a countable dense subset of R
n
and that the family
B = B(p, ): p
n
& of open balls forms a countable base for
R
n
, that is, for every open set U in R
n
and every p U there is a B B
such that p B U.
A point p of a subset P of R
n
is an isolated point of P if there is an
open set U (in R
n
or in P) such that U P = p. A subset S of R
n
is
discrete if every point of S is isolated in S. A nonempty closed subset F
of R
n
is said to be perfect if it has no isolated points.
We will start this section with a study of the structure of closed subsets
of R
n
. For this we will need a few theorems that are of interest in their
own right.
Theorem 6.2.1
(i) Every family | of pairwise-disjoint open subsets of R
n
is at most
countable.
(ii) Every discrete subset S of R
n
is at most countable.
(iii) If is an ordinal number and S

: < is a strictly increasing

(decreasing) sequence of open (closed) sets in R
n
then is at most
countable.
Proof (i) Dene f : | `
n
by choosing f(U) U
n
for every
U |. Then f is one-to-one, and so [|[ [
n
[ + 1 = .

90 6 Subsets of R
n
(ii) For every s S let r
s
> 0 be such that S B(s, 2r
s
) = s and
let | = B(s, r
s
): s S. Notice that B(s, r
s
) B(t, r
t
) = for distinct
s, t S. Thus, by (i), | is at most countable. Hence [S[ [|[ .
(iii) We assume that | = S

: < is a strictly increasing sequence

of open sets. The argument for the other case is essentially the same.
To obtain a contradiction assume that is uncountable. Thus
1
.
For <
1
choose x

S
+1
` S

and let B

B be such that x

S
+1
. Then the set X = x

: <
1
is uncountable, since x

= x

for
< <
1
. But the function f : X B, f(x

) = B

, is one-to-one, since
for < <
1
, x

`B

, that is, B

= B

. However, this is impossible,

since it would imply
1
= [X[ [B[ = .
This contradiction nishes the proof.
Theorem 6.2.2 For an ordinal number the following conditions are
equivalent:
(i) <
1
;
(ii) there exists an S R well ordered by the standard relation of order
type .
Proof (ii)(i): Assume, to obtain a contradiction, that for some
1
there exists an S R with order type . Let h: S be an order iso-
morphism and dene U

= (h(), h( + 1)) for <

1
. Then U

: <
1

is an uncountable family of nonempty, open, and pairwise-disjoint subsets

of R. This contradicts Theorem 6.2.1(i).
(i)(ii): For the other implication we will prove the following fact by
induction on <
1
:
() for every a, bR with a<b there exists an S [a, b) with Otp(S)=.
For = 0 the condition is clearly true. So let 0 < <
1
be such that
() holds for all < . We will show that () holds for .
Take a, b R with a < b. If = + 1 is an ordinal successor, pick c
with a < c < b and S [a, c) with Otp(S) = . This can be done by the
inductive assumption. Then S c [a, b) and Otp(S c) = +1 = .
If is a limit ordinal, then cf() <
1
and cf() is innite. So
cf() = . Let
n
< : n < be a strictly increasing sequence that is
conal in with
0
= 0. Find a strictly increasing sequence a = a
0
< a
1
<
< a
n
< < b (for example, put a
n
= b [(b a)/(n + 1)]). For every
n < choose S
n
[a
n
, a
n+1
) with Otp(S
n
) = Otp(
n+1
`
n
). This can be
done by the inductive hypothesis, since Otp(
n+1
`
n
)
n+1
< . Put
S =

n<
S
n
. Then S [a, b). Moreover, if f
n
: (
n+1
`
n
) S
n
is an
order isomorphism then f =

n<
f
n
: S is also an order isomorphism,
that is, Otp(S) = . Condition () has been proved for every <
1
.
Now () clearly implies (ii).

6.2 Closed sets and Borel sets 91
Theorem 6.2.3 Every perfect subset P of R
n
has cardinality continuum.
Proof Clearly [P[ [R
n
[ = c. Thus it is enough to show that [P[ c.
Before we prove it we notice the following easy fact.
(A) If R R
n
is perfect, U R
n
is open, and R U = R cl(U) =
then R cl(U) is perfect.
Clearly R cl(U) is closed. Also, p R cl(U) cannot be isolated since
otherwise there would exist an open set W R
n
with the property that
p = (R cl(U)) W = (R U) W = R (U W) and the set U W
would show that p is an isolated point of R.
Now let p
i
: R
n
R be the projection onto the ith coordinate. If
there exists an i < n such that [p
i
[P][ = c, then [P[ c. So, to obtain a
contradiction, assume that
(B) [p
i
[P][ < c for every i < n.
Notice also that
(C) if R P is perfect, then there are disjoint perfect sets R
0
, R
1
R.
To see it, let i < n be such that p
i
[R] contains two numbers a < b. By (B)
we can nd r (a, b) `p
i
[R]. Then, by (A), the sets R
0
= Rp
1
i
((, r])
and R
1
= R p
1
i
([r, )) are disjoint and perfect.
Now we construct a family P
s
P : s 2
<
of perfect bounded sets
by induction on the length [s[ of a sequence s.
We put P

= P [a, b]
n
, where a < b are chosen in such a way that
P

= and a, b R`

i<n
p
i
[P]. This can be done by (B). So P (a, b)
n
=
P [a, b]
n
. In particular, P

is compact and, by (A), perfect. We will

continue the induction, maintaining the following inductive condition to be
satised for every s 2
<
.
(I) P
s0
and P
s1
are disjoint perfect subsets of P
s
,
where s0 and s1 stand for the sequences of length [s[ + 1 extending s by 0
and 1, respectively. We can nd such sets by (C).
Now notice that for f 2

the sets P
f]
k
for k < form a decreas-
ing sequence of nonempty compact sets. Thus, the sets P
f
=

k<
P
f]
k
are nonempty. Notice also that the sets P
f
and P
g
are disjoint for dif-
ferent f, g 2

. To see this, let k = mini : f(i) = g(i). Then

f(k), g(k) = 0, 1 and
P
f
P
g
P
f]
k+1
P
g]
k+1
= P
s0
P
s1
= ,
where s = f[
k
= g[
k
.

92 6 Subsets of R
n
Let : 2

P be such that (f) P

f
(this is a choice function). Then
is clearly one-to-one and so [P[ c.
The induction constructions of the preceding proof are often referred
to as tree constructions, since T = 2
<
is called a tree. (We imagine
every s T as a tree-branching point, where the tree is splitting into s0
and s1.)
For F R
n
we denote by F
t
the set of all limit points of F, that is, all
points of cl(F) that are not isolated. Clearly F ` F
t
is equal to the set of
all isolated points of F.
Now we are ready to prove the following theorem characterizing closed
subsets of R
n
.
Theorem 6.2.4 (CantorBendixson theorem) Every uncountable, closed
subset F of R
n
can be represented as a disjoint union of a perfect set P
and an at most countable set C.
Proof By induction on <
1
we construct a sequence F

: <
1
of
closed subsets of R
n
by dening:
F
0
= F, F

<
F

for <
1
, a limit ordinal,
and
F
+1
= (F

)
t
.
The sequence F

: <
1
is decreasing and formed of closed sets. It
cant be strictly decreasing by Theorem 6.2.1(iii). Thus there exists an <

1
such that F

= F
+1
. This means that F

does not have any isolated

points, that is, F

is either perfect or empty. But F`F

<
(F

`F
+1
)
and every set F

` F
+1
= F

` (F

)
t
is discrete, so, by Theorem 6.2.1(ii),
it is at most countable. Hence F ` F

is a union of [[ sets, each of

which is at most countable. Therefore C = F ` F

is at most countable,
and P = F

= F ` C cannot be empty. So P is perfect.

Corollary 6.2.5 The cardinality of a closed subset of R
n
is either equal
to continuum or is at most countable.
Theorems 6.2.3 and 6.2.4 tell us the properties that closed subsets of
R
n
must have. Can we tell anything more? Exercise 1 shows that we
cannot expect much improvement in the case of Theorem 6.2.4. But what
about Theorem 6.2.3? How can perfect sets look? They can certainly be
formed as a union of some number of closed intervals. But for such sets
Theorem 6.2.3 is obvious. Are there any other perfect sets? The armative

6.2 Closed sets and Borel sets 93
answer is given by a set C R to be constructed and known as the Cantor
set (or Cantor ternary set). It is constructed as follows.
First, we dene the family I
s
P : s 2
<
of closed intervals by
induction on the length [s[ of s. We put I

= [0, 1]. If the interval I

s
=
[a
s
, b
s
] is already constructed, we remove from it the middle third interval,
and dene I
s0
and I
s1
as the left interval and the right interval, respectively,
of the remaining two intervals, that is, I
s0
= [a
s
, a
s
+ (b
s
a
s
)/3] and
I
s1
= [b
s
(b
s
a
s
)/3, b
s
]. Put C
n
=

I
s
: [s[ = n. Thus C
n
is the
union of all 2
n
internals of length 1/3
n
constructed at step n. Notice that
[0, 1] = C
0
C
1
and that each C
n
is closed, being a nite union of
closed intervals. We dene C =

n<
C
n
.
The set C is clearly closed and nonempty. To see that it is perfect, let
p C and let J be an open interval containing p. Then there is an s 2
<
such that the interval I
s
(of length 1/3
]s]
) contains p and is contained in J.
Let j < 2 be such that p I
sj
. Since CI
sj
=

n<
(C
n
I
sj
) is nonempty,
being the intersection of a decreasing sequence of nonempty compact sets,
we conclude that C J = p. So C is perfect.
Notice also that C does not contain any nonempty interval. This is so
since for any interval J of length > 0 we can nd an n < such that
1/3
n
< . But C
n
is a union of intervals of length 1/3
n
< . So J C
n
,
and J C since C C
n
.
Finally, it is worthwhile to mention that C is the set of all numbers
from [0, 1] whose ternary representations do not contain the digit 1, that
is,
C =

n=0
d(n)
3
n+1
: d 0, 2

.
This will be left without a proof.
To dene Borel subsets of topological spaces we need the following def-
initions. A family / {(X) is a -algebra on X if / is nonempty and
closed under complements and under countable unions, that is, such that
(i) , X /;
(ii) if A / then X ` A /;
(iii) if A
k
/ for every k < then

k<
A
k
/.
Notice also that every -algebra is also closed under countable intersections,
since

k<
A
k
= X `

k<
(X ` A
k
).

94 6 Subsets of R
n
Examples For every nonempty set X the following families form-algebras:
(1) / = , X,
(2) / = {(X),
(3) / = A {(X): [A[ or [X ` A[ for any innite cardinal .
Notice that for every T {(X) there is a smallest -algebra [T] on
X containing T, namely,
[T] =

/ {(X): T / & / is a -algebra on X. (6.6)

In particular, for every topological space 'X, ` we dene the -algebra Bor
of Borel sets by Bor = [].
It is obvious from the denition that the following sets are Borel:
G

sets, that is, the countable intersections of open sets;

F

sets, that is, the countable unions of closed sets;

G

sets, that is, the countable unions of G

sets;
F

sets, that is, the countable intersections of F

sets;
and so on.
Is an arbitrary subset of a topological space 'X, ` Borel? The answer to
this question depends on the topological space. For example, in a discrete
space 'X, {(X)` every set is open and hence Borel. This is, however, quite
an exceptional example and more often {(X) = Bor. We will not approach
this question in general. However, we will address it in the case of R
n
.
To shed more light on the structure of Borel sets, dene inductively the
following hierarchy of sets for a topological space 'X, `:

0
1
= ,

k<
(X ` A
k
): A
k

0<<

for all k <

for 1 < <

1
.
Dene also

= X ` A: A
0

for 0 < <

1
.
Thus

k<
A
k
: A
k

0<<

for all k <

for 1 < <

1
.

6.2 Closed sets and Borel sets 95
It is very easy to prove, by induction on 0 < <
1
, that

,
0

Bor for every 0 < <

1
. (6.7)
Notice that
0
1
is equal to the family of all closed subsets of X,
0
2
= F

0
2
= G

,
0
3
= G

, and
0
3
= F

.
From now on we will assume that X = R
n
. In particular, any closed
subset F of R
n
is G

, as
F =

k<
x R
n
: y F (d(y, x) < 1/(k +1)) =

k<

yF
B(y, 1/(k +1)).
Similarly, every open set in R
n
is F

. Hence

0
1

0
2
and
0
1

0
2
. (6.8)
Using these inclusions it is not dicult to prove
Proposition 6.2.6 For every 0 < < <
1

,
0

,
0

,
0

.
Proof Inclusion
0

follows immediately from the denition of

0

.
Now, to see
0

notice that
A
0

R
n
` A
0

A
0

.
We prove

,
0

for every 0 < <

by joint induction on . So assume that for some 1 < <
1
this condition
holds for every < and let 0 < < . We have to prove that
0

and
0

.
But then

<

0

<

0

, and so

k<
A
k
: A
k

<

k<
A
k
: A
k

<

=
0

.
Now
0

follows from
A
0

R
n
` A
0

A
0

.
Proposition 6.2.6 has been proved.

96 6 Subsets of R
n
Theorem 6.2.7 The family Bor of Borel subsets of R
n
is equal to
Bor =

0<<1

0<<1

.
Proof First notice that Proposition 6.2.6 implies

0<<1

0<<1

.
Let T denote this set. Notice also that (6.7) implies that T Bor. Since
T contains the topology of R
n
it is enough to show that T is a -algebra,
since Bor = [] is the smallest -algebra containing . Thus we have to
prove that T is closed under complements and countable unions.
If A T then A
0

for some 0 < <

1
. Hence R
n
` A
0

T.
If A
k
T for k < then for every k < there exists 0 <
k
<
1
such
that A
k

0

k
. But
1
=
+
is regular, so there exists an <
1
such
that
k
< for every k < . Hence
A
k

0

0<<

for every k < , and so

k<
A
k

0

T.
Theorem 6.2.8 The family Bor of all Borel subsets of R
n
has cardinality
continuum.
Proof First we will prove by induction that
[
0

[ = [
0

[ = c for every 0 < <

1
. (6.9)
Equation [
0

[ = [
0

[ is established by a bijection f :
0

dened by
f(A) = X ` A. Next we will prove that [
0
1
[ = c. The inequality [
0
1
[ =
[[ c is justied by a surjective function f : {(B) , f(|) =

|,
where B is a countable base for R
n
. The inequality [[ c follows from the
facts that the base B contains an innite subfamily ( of pairwise-disjoint
sets and that f restricted to {(() is one-to-one, while [{(()[ = c.
Now assume that for some 1 < <
1
we have [
0

[ = [
0

[ = c for all
0 < < . We have to prove [
0

[ = c.
But, by the inductive hypothesis, [

<

0

[ = c since it is a union
of [[ <
1
c sets of cardinality c. Moreover, F :

<

0

dened by F('A
0
, A
1
, . . .`) =

n<
A
n
is onto
0

, so

<

<

= (2

= c.

6.2 Closed sets and Borel sets 97
Since c = [
0
1
[ [
0

[, condition (6.9) has been proved.

Now, by Theorem 6.2.7, Bor =

0<<1

0

is a union of
1
c sets of
cardinality c, that is, it has cardinality continuum.
Corollary 6.2.9 There is a non-Borel subset of R
n
.
Proof By Theorem 6.2.8 we have
[{(R
n
)[ = 2
]R
n
]
= 2
c
> c = [Bor[,
that is, {(R
n
) ` Bor = .
We will construct some explicit examples of non-Borel sets in the next
section.
EXERCISES
1 For a closed set F R let F

: <
1
be a sequence from the proof
of Theorem 6.2.4. Show that for every <
1
there exists a closed subset
F R for which F
+1
= F

.
2 Prove that the family [T] dened by (6.6) is indeed a -algebra on X.
3 Show that Theorem 6.2.7 remains true for an arbitrary topological space.
4 Prove the following generalization of Lemma 6.1.6.
Let X be a set and let T be a family of functions of the form f : X

X, where . If [T[ c then for every Z X there exists a Y X

such that
(i) Z Y ;
(ii) Y is closed under the action of T, that is, for every f : X

X from
T and S Y

we have f(S) Y ;
(iii) [Y [ [Z[

+ c.
Notice that this fact (used with X = {(R
n
), Z = , and T composed
of countable unions and complements) implies Theorem 6.2.8.
5 Prove that every uncountable Borel subset of R contains a perfect subset.

98 6 Subsets of R
n
6.3 Lebesgue-measurable sets and sets with
the Baire property
To construct the most useful -algebras we need the following denitions.
Recall, from Section 2.1, that a nonempty family 1 of subsets of a set
X is an ideal on X if it is closed under the subset operation and under
nite unions, that is, such that
(i) if A, B 1 then A B 1;
(ii) if A 1 and B A then B 1.
An ideal 1 is said to be a -ideal if it is closed under countable unions,
that is, if
(i
t
)

k
A
k
1 provided A
k
1 for all k < .
The elements of an ideal on X are usually considered as small in some
sense.
Examples 1. 1 = is a -ideal on every set X.
2. For every A X the family 1 = {(A) forms a -ideal on X. In the
case when A = X we obtain 1 = {(X). This ideal does not agree with our
intuitive notion of a family of small sets. Thus, usually we will work with
the proper ideals on X, that is, the ideals that are not equal to {(X).
3. For every innite cardinal number and every set X the family 1 =
[X]
<
is an ideal on X. For = this is the ideal of nite subsets of X.
For cf() > it is also a -ideal. Notice also that for =
1
this is the
ideal [X]

of at most countable subsets of X.

4. For every topological space 'X, ` the family ^T of all nowhere-dense
subsets of X (i.e., the subsets S of X such that int(cl(S)) = ) is an ideal
on X.
5. For a topological space 'X, ` the family
=

k<
A
k
: A
k
is nowhere dense in X

is a -ideal on X. If X = R
n
(or, more generally, X is a complete metric
space or a compact space) then, by the Baire category theorem, is
proper. If moreover X does not have any isolated points, then [X]

.
The ideal is usually called the ideal of Meager (or rst-category) subsets
of X.
6. For n-dimensional Euclidean space R
n
we say that X R
n
is a
(Lebesgue) measure-zero set or a null set in R
n
if for every > 0 there

6.3 Lebesgue-measurable sets and sets with the Baire property 99
is a family of open balls B(x
k
, r
k
): k < such that X

k<
B(x
k
, r
k
)
and

k<
r
n
k
< . The family
^ = X R
n
: X is a null set in R
n

forms a proper -ideal in R

n
called the -ideal of (Lebesgue) measure-zero
sets or null sets.
The ideals and ^ on R
n
are of main interest in real analysis. It is
easy to see that every countable set belongs to both and ^. Also, for
n = 1, the Cantor set C (see Section 6.2) is in ^. It belongs to since
it is closed and does not contain any nonempty open interval. It belongs
to ^ since the total length of intervals forming C
n
C is 2
n 1
3
n
=

2
3

n
.
In what follows we will study -ideals and ^ from the set-theoretic
point of view. It is worthwhile to notice, however, that both these ide-
als measure smallness in a very dierent sense. Ideal describes the
smallness in a topological sense whereas ^ does so in a measure sense. The
dierence between these two smallness notions is best captured by the
following example.
Proposition 6.3.1 There exists a dense G

set G R
n
such that G ^
and R
n
` G .
Proof Recall that
n
is a countable dense subset of R
n
. Let q
k
: k <
be an enumeration of
n
. For m < let
G
m
=

k<
B(q
k
, 2
(m+k)/n
)
and let G =

m<
G
m
.
G ^ since for every > 0 there exists an m < with 2
(m1)
<
and
G G
m
=

k<
B(q
k
, 2
(m+k)/n
),
while

k<

2
(m+k)/n

n
= 2
(m1)
< .
To see that R
n
` G notice that
R
n
` G = R
n
`

m<
G
m
=

m<
(R
n
` G
m
)
and each of the sets R
n
`G
m
is closed and nowhere dense, since G
m
is open
and dense.

100 6 Subsets of R
n
To continue our journey through the special subsets of R
n
we need the
following denitions and constructions. For a -ideal 1 on a set X dene
a binary relation
1
on {(X) by
A
1
B AB 1.
Notice the following easy facts.
Proposition 6.3.2 Let 1 be a -ideal on X.
(i) For every A, B X,
A
1
B if and only if A` D = B ` D for some D 1.
(ii)
1
is an equivalence relation on {(X).
(iii) If A
1
B then X ` A
1
X ` B.
(iv) If A
n

1
B
n
for every n < then

n<
A
n

1

n<
B
n
.
Proof (i) If A
1
B then D = AB 1 and A` D = B` D. Conversely,
if A` D = B ` D for some D 1 then AB D 1.
(ii) It is easy to see that A
1
A and that A
1
B implies B
1
A for
every A, B X. Now, if A
1
B and B
1
C then A ` D = B ` D and
B`E = C`E for some D, E 1. So A`(DE) = B`(DE) = C`(DE)
and D E 1. So A
1
C.
(iii) If A
1
B then A`D = B`D for some D 1 and so (X`A)`D =
(X ` B) ` D. Hence X ` A
1
X ` B.
(iv) Let A
n
, B
n
X be such that A
n

1
B
n
for every n < and let
D
n
1 be such that A
n
`D
n
= B
n
`D
n
for n < . Then D =

n<
D
n
1
and so

n<
A
n
` D =

n<
B
n
` D. Therefore

n<
A
n

1

n<
B
n
.
Theorem 6.3.3 Let 1 be a -ideal on X and let / {(X) be nonempty.
The family
/[1] = AD: A / & D 1
forms a -algebra on X if and only if
(1) X ` A /[1] for every A /, and
(2)

n<
A
n
/[1] provided A
n
/ for every n < .
In particular, if / is a -algebra on X then /[1] is a -algebra generated
by / 1.

6.3 Lebesgue-measurable sets and sets with the Baire property 101
Proof The implication as well as the additional part of the theorem
are obvious. To prove the other implication we have to prove that /[1] is
closed under complements and under countable unions. First notice that
C
1
B & B /[1] C /[1], (6.10)
since B /[1] implies that B
1
A for some A /, so C
1
A and
C = AD /[1], where D = AC 1.
Now, if B = AD for some A / and D 1 then BA 1 and, by
Proposition 6.3.1(iii), X`B
1
X`A. But, by our assumption, X`A /[1]
and so, by (6.10), X ` B /[1].
Similarly, if B
n
= A
n
D
n
for some A
n
/and D
n
1 then B
n
A
n

1 and, by Proposition 6.3.1(iv),

n<
B
n

1

n<
A
n
. Since

n<
A
n

/[1], condition (6.10) implies that

n<
B
n
/[1].
The most important -algebras on R
n
generated as in Theorem 6.3.3
are the -algebras L = Bor[^] of Lebesgue-measurable sets and the -
algebra Baire = Bor[] of sets with the Baire property, where Bor stands
for the -algebra of Borel subsets of R
n
. Both these families are very rich.
In fact, without use of the axiom of choice we cannot prove the existence
of a subset of R
n
that does not have the Baire property. A similar theorem
for the family L of all measurable sets can also be proved.
In what follows we prove that there is a subset of R
n
that is nonmea-
surable and does not have the Baire property. For this, however, we need
some structural theorems about Baire and L.
Theorem 6.3.4 Baire = [] = UM: U is open in R
n
and M .
Proof This follows immediately from Theorem 6.3.3 since a union of open
sets is open and since for every open set A its complement F = R
n
` A
is closed, and F = int(F) (F ` int(F)), where F ` int(F) is closed and
nowhere dense.
The similar characterization for L is more complicated and is given by
the next theorem.
Theorem 6.3.5 For every A L there exists an F

set F and a G

set
G such that F A G and G` F ^.
In particular,
L = GN: G is G

in R
n
and N ^
= FN: F is F

in R
n
and N ^.
The property described in this theorem is called the regularity of the
family L and is a basic fact about Lebesgue measure. We will leave it here
without proof (see, e.g., Royden 1988).

102 6 Subsets of R
n
Theorem 6.3.6 If A L`^ or A Baire` then there exists a perfect
set P such that P A.
Proof First assume that A L`^. Then, by Theorem 6.3.5, there exists
an F

set F such that F A and A ` F ^. Thus F L ` ^. Let

F =

n<
F
n
, where the sets F
n
are closed in R
n
. Notice that at least one
of the sets F
n
must be uncountable, since otherwise F would be at most
countable and then it would belong to ^. So let F
n
be uncountable. Then,
by the CantorBendixson theorem (Theorem 6.2.4), there is a perfect set
P F
n
F A.
Now assume that A Baire ` and let A = US for some nonempty
open set U and S . Let S =

n<
S
n
for some S
n
^T.
As in Theorem 6.2.3 we will use a tree-construction argument to
dene a family U
s
U : s 2
<
of nonempty open sets; that is, the
construction will be done by induction on the length [s[ of a sequence s.
We choose a bounded U

such that cl(U

) U ` S
0
and continue the
induction, maintaining the following condition to be satised for every
s 2
<
:
(I) U
s0
and U
s1
are open balls such that
cl(U
s0
) cl(U
s1
) = , cl(U
s0
) cl(U
s1
) U
s
` S
]s]
.
To make an inductive step assume that U
s
is already constructed. Since
cl(S
]s]
) ^T we have U
s
` cl(S
]s]
) = and we can nd two open disjoint
balls in U
s
` cl(S
]s]
). Decreasing their radii, if necessary, we can satisfy
condition (I). This completes the construction.
Now let F
n
=

cl(U
s
): s 2
<
, [s[ = n and F =

n<
F
n
. By
our construction, F U and F
n
S
n
= for every n < . Hence,
F U `S A. Also, each of the sets F
n
is compact, being a nite union of
closed bounded sets. Thus F is compact. To nish the proof it is enough to
show that F is uncountable, since then, by the CantorBendixson theorem,
there is a perfect set P F A.
To see this, for every f 2

consider a set F
f
=

n<
cl(U
f]n
)
F. Notice that, by Theorem 3.3.1, each set F
f
is nonempty. Note also
that the sets F
f
: f 2

are pairwise disjoint. Indeed, if f, g 2

are dierent, n = mini < : f(i) = g(i), and s = f[

n
= g[
n
then
F
f
F
g
cl(U
s0
) cl(U
s1
) = . Now let c: 2

F be a choice function
for the family F
f
: f 2

, that is, c(f) F

f
for every f 2

. Then c is
one-to-one and so [2

[ [F[, proving that F is indeed uncountable.

This nishes the proof.
We conclude this section with the following two theorems.

6.3 Lebesgue-measurable sets and sets with the Baire property 103
Theorem 6.3.7 There exists a set B R
n
such that neither B nor R
n
`B
contains any perfect subset.
Proof First notice that the family T of all perfect subsets of R
n
has
cardinality c, since T Bor and [Bor[ = c.
Let T = P

: < c. By transnite induction on < c dene the

sequences 'a

: < c` and 'b

: < c` by choosing in step < c points

a

= b

from R
n
such that
a

, b

` (a

: < b

: < ).
This can be done since
[a

: < b

: < [ = 2 [[ < c
and, by Theorem 6.2.3, [P

[ = c. This nishes the construction.

Now dene B = a

: < c. Then for every perfect set P there is a

< c such that P = P

. Hence a

P B and b

P (R
n
` B). Thus
P B and P R
n
` B.
The set B from Theorem 6.3.7 is called a Bernstein set.
Theorem 6.3.8 A Bernstein subset B of R
n
neither is measurable nor
has the Baire property.
Proof To obtain a contradiction assume that B L. Then, by Theo-
rem 6.3.6, B ^. Similarly, R
n
` B L and R
n
` B ^. But then
R
n
= B (R
n
` B) ^, which is false. So B L.
The proof that B Baire is exactly the same.
EXERCISES
1 Show that there exists a Bernstein set B such that B + B = R, where
B +B = b
0
+b
1
: b
0
, b
1
B.
2 Show that there exists a Bernstein set B such that B + B is also a
Bernstein set.

Chapter 7
Strange real functions
This chapter is designed to help the reader to master the technique of
recursive denitions. Thus, most of the examples presented will involve
constructions by transnite induction.
7.1 Measurable and nonmeasurable functions
Let B be a -algebra on R
n
. A function f : R
n
R is said to be a B-
measurable function if f
1
(U) B for every open set U R. Notice that
if f is B-measurable then f
1
(B) B for every Borel set B R. This is
the case since the family B R: f
1
(B) B is a -algebra containing
all open sets.
We will use this notion mainly for the -algebras of Borel, Lebesgue-
measurable, and Baire subsets of R
n
, respectively. In each of these cases
B-measurable functions will be termed, respectively, as Borel functions (or
Borel-measurable functions), measurable functions (or Lebesgue-measurable
functions), and Baire functions (or Baire-measurable functions). Clearly,
every continuous function is Borel-measurable and every Borel-measurable
function is measurable and Baire.
A function f : R
n
R is non-Borel (or nonBorel-measurable) if it
is not Borel. Similarly, we dene non-Baire(-measurable) functions and
non-(Lebesgue-)measurable functions.
Also recall that the characteristic function

A
of a subset A of a set X
is dened by putting

A
(x) = 1 if x A and

A
(x) = 0 for x X ` A.
The rst theorem is a corollary to Theorems 6.3.7 and 6.3.8.
Theorem 7.1.1 The characteristic function

B
of a Bernstein subset B
of R
n
is neither a measurable nor a Baire function. In particular, there
exists a nonmeasurable, non-Baire function from R
n
to R.
104

7.1 Measurable and nonmeasurable functions 105
Proof The set (

B
)
1
(1) = B neither is measurable nor has the Baire
property.
Let us recall that the FubiniTonelli theorem says that for a measurable
function f : R
2
R that is either integrable or nonnegative we have

f dm
2
=

f
x
dm
1

dm
1
=

f
y
dm
1

dm
1
,
where m
2
and m
1
stand for the Lebesgue measures on R
2
and R, respec-
tively, and f
x
, f
y
: R R are dened by f
x
(y) = f(x, y) and f
y
(x) =
f(x, y). The integrals

f
x
dm
1

dm
1
and

f
y
dm
1

dm
1
are called
iterated integrals. Thus, the FubiniTonelli theorem tells us for measurable
functions what for the continuous functions is taught in every multivariate
calculus course: A two-dimensional integral is equal to both iterated inte-
grals. But what if the function f is nonmeasurable? Then we cant talk
about the integral

f dm
2
. However, it is still possible that both iterated
integrals exist. Must they be equal? The next theorem, due to Sierpi nski
(1920), gives a negative answer to this question.
Theorem 7.1.2 If the continuum hypothesis holds then there exists a func-
tion f : [0, 1]
2
[0, 1] such that

f
x
dm
1
= 1 and

f
y
dm
1
= 0 for all
x, y [0, 1]. In particular,

f
x
dm
1

dm
1
= 1 and

f
y
dm
1

dm
1
= 0.
Proof Let A be as in Theorem 6.1.8, put B = A [0, 1]
2
, and dene
f =

B
. Notice that [B
y
[ and [([0, 1]
2
` B)
x
[ for all x, y [0, 1].
Thus f
y
(x) = 0 for all but countably many x R, and so

f
y
dm
1
=

0 dm
1
= 0. Similarly, for every x [0, 1] we have f
x
(y) = 1 for all but
countably many y [0, 1]. Thus

f
x
dm
1
=

[0,1]
1 dm
1
= 1.
A set B [0, 1]
2
for which the function f =

B
satises Theorem 7.1.2
is usually called a 0-1 set. Its existence is not equivalent to the continuum
hypothesis; that is, it might exist when the continuum hypothesis fails.
On the other hand, the existence of the function f from Theorem 7.1.2
cannot be proved in ZFC, which has been shown by H. Friedman (1980)
and, independently, by C. Freiling (1986).
EXERCISE
1 Prove that there exists a function f : R R such that f
1
(r) is a
Bernstein set for every r R.

106 7 Strange real functions
7.2 Darboux functions
To motivate what follows let us recall one of the most fundamental theo-
rems of calculus the intermediate value theorem. It tells us that every
continuous function f : R R has the following property:
(DP) For every a < b and every number y between f(a) and f(b) there is
an x (a, b) such that f(x) = y.
The functions satisfying property (DP) form an important class of functions
known as Darboux functions. In particular, we say that a function f : R
R has the Darboux property (or is a Darboux function) if it satises property
(DP). Thus, the intermediate value theorem says that every continuous
function f : R R has the Darboux property.
Not every Darboux function, however, is continuous. For example, the
function f(x) = sin(1/x) for x = 0 and f(0) = 0 is a discontinuous Dar-
boux function. The next theorem tells us that there are also everywhere-
discontinuous Darboux functions (that is, Darboux functions that are dis-
continuous at every point). For this we will need one more denition. We
say that a function f : R R is strongly Darboux if
(SD) for every a < b and every number y there is an x (a, b) such that
f(x) = y.
Thus f is strongly Darboux if f[(a, b)] = R for every a < b. Clearly, every
strongly Darboux function is Darboux and everywhere discontinuous.
Theorem 7.2.1 There exists a strongly Darboux function f : R R. In
particular, f is Darboux and everywhere discontinuous.
Proof The set A constructed in Theorem 6.1.1 is the graph of the desired
function f.
The next theorem is a generalization of Theorem 7.2.1.
Theorem 7.2.2 Let ( be a family of real functions, ( R
R
, with [([ c.
Then there exists a function f : R R such that f +g is strongly Darboux
for every g (.
Proof Let T = 'g, I, r`: g ( & r R & I = (a, b) for some a < b.
Then [T[ = c, since c = [R[ [T[ [([ [R
3
[ = c. Let 'g

, I

, r

`: < c
be an enumeration of family T. By transnite induction dene a sequence
x

: < c such that

x

` x

: < .
Such a choice can be made since [I

[ = [R[ = c > [[ [x

: < [ for
every < c.

7.2 Darboux functions 107
Dene f(x

) = r

(x

) and extend it to all of R arbitrarily. Notice

that it has the desired property, since for every nonempty interval (a, b),
every r R, and every g ( there exists a < c such that 'g

, I

, r

` =
'g, (a, b), r`, and so
r = r

= f(x

) +g

(x

) (f +g

)[I

] = (f +g)[(a, b)].
Notice that Theorem 7.2.2 generalizes Theorem 7.2.1, since we can as-
sume that the function that is identically zero belongs to (. In particular,
the function f in Theorem 7.2.2 can be chosen to be strongly Darboux.
Corollary 7.2.3 Let ( be a family of real functions such that [([ c.
Then there exists a (strongly) Darboux function f : R R such that f +g
is (strongly) Darboux for every g (.
Another way to look at Corollary 7.2.3 is to use the language of cardinal
functions as follows. Let T stand for the family of all Darboux functions
from R to R and let /(T) be the minimal cardinality of a family ( for
which Corollary 7.2.3 fails, that is,
/(T) = min[([ : f R
R
g ( (f +g T).
Notice that the set [([ : f R
R
g ( (f + g T) is not empty,
since [R
R
[ = c
c
= 2
c
belongs to this set: For every f R
R
there exists a
g ( = R
R
such that f + g is equal to any xed function h, and such
that h can be chosen to be not Darboux. So /(T) 2
c
. In this language
Corollary 7.2.3 can be stated as follows:
c < /(T) 2
c
. (7.1)
If 2
c
= c
+
, which follows from the generalized continuum hypothesis, then
/(T) = 2
c
. However, this equation cannot be proved in ZFC alone. It
has been proved by K. Ciesielski and A. W. Miller (19945) that (7.1) is
essentially everything that can be proved in ZFC about /(T).
In what follows we will consider some other generalizations of Theo-
rem 7.2.1. We start by noticing that the sum of a Darboux function and
a continuous function does not have to be Darboux. To see it, modify the
proof of Theorem 7.2.2 to get a strongly Darboux function f such that
g(x) = f(x) x is not Darboux, by choosing
x

` (x

: < r

)
and dening f(x

) = r

. Then f(x

) = x

and we can extend f to all of

R to have f(x) = x for all x R. So g(x) = f(x) x = 0 for every x R
and Proposition 7.2.4, stated next, implies that g is not Darboux.

108 7 Strange real functions
Proposition 7.2.4 If f is strongly Darboux, then for every continuous
function g,
f +g is strongly Darboux i f +g is Darboux.
Proof Implication is obvious. The other implication follows from the
fact that on any interval I = (c, d), with c < d, the continuous function
g is bounded, while f[I] = R. So f + g must be unbounded from below
and from above. In particular, for every y R there exist A < y < B and
a, b I such that (f + g)(a) = A and (f + g)(b) = B. Thus there exists
an x between a and b with (f +g)(x) = y, since f +g is Darboux. Hence
(f +g)[I] = R.
The next theorem, due to Kirchheim and Natkaniec (19901), gener-
alizes the preceding observation. It concerns continuous nowhere-constant
functions, that is, continuous functions that are not constant on any open
nonempty interval.
Theorem 7.2.5 If the continuum hypothesis holds then there is a strongly
Darboux function f such that f + h is not Darboux for every continuous
nowhere-constant function h.
Proof Let h

: < c be an enumeration of all continuous nowhere-

constant functions from R to R. Notice that for any such function h the
set h
1
(r) is closed and nowhere dense for every r R.
By Proposition 7.2.4 it is enough to nd f : R R such that
< c r

R x R (f(x) +h

(x) = r

). (7.2)
For this we will construct sequences x

R: < c, y

R: < c, and
r

R: < c such that the function f dened by

f(x

) = y

(7.3)
is strongly Darboux and satises condition (7.2). As we know from Theo-
rems 7.2.1 and 7.2.2, it is relatively easy to construct the sequences x

and y

such that f dened by (7.3) is strongly Darboux. In previous

constructions, however, we did not have to worry about x

: < c be-
ing equal to R, since any extension of the part of f given by (7.3) was
still strongly Darboux. But this time, we also have to take care of condi-
tion (7.2), and the extension may fail to have this property. To avoid this
problem we will arrange our construction to have
x

: < c = R. (7.4)
Now let S and T be subsets of c of cardinality c. Let s

: S be an
enumeration of R and let 'I

, t

`: T be an enumeration of the family

7.2 Darboux functions 109
T = 'I, t`: t R & I = (a, b) for some a < b. To satisfy (7.4) it is
enough to have
(S1) x

= s

for every S.
To make sure that f is strongly Darboux it is enough to proceed as in
Theorems 7.2.1 and 7.2.2; that is, for every T choose
x

` x

: < (7.5)
and put
(T1) y

= t

.
Notice, however, that for s

conditions (S1) and (7.5) cannot be

satised simultaneously. To avoid this problem we will assume that the
sets S and T are disjoint. We can also assume that S T = c.
Property (7.2) will hold if we construct the sequences in such a way
that for every < c
< < (y

+h

(x

) = r

). (7.6)
To preserve this condition while constructing x

, y

, and r

we will have
to choose r

such that y

+h

(x

) = r

for every < , that is, pick

() r

R ` y

+h

(x

): < ,
and choose x

and y

such that
y

+h

(x

) = r

for every . (7.7)

Now for T we have no freedom in choosing y

since, by (T1),
y

= t

. But to satisfy (7.7) it is enough to choose x

h
1

(r

).
Combining this with (7.5) we need only choose
(T2) x

: <

h
1

(r

):

for every T.
For S, by (S1), we have x

= s

. If x

= x

for some < then, by

(7.3), we have to dene
(S2) y

= y

.
But then conditions (7.6) and (7.7) will be preserved. On the other hand,
if x

= x

for every < , in order to maintain (7.7) it is enough to choose

(S3) y

R ` r

(x

): .

110 7 Strange real functions
The foregoing discussion shows that the function f can be constructed
if we can construct the sequences x

R: < c, y

R: < c, and
r

R: < c such that the following conditions hold for every < c:
() r

R ` y

+h

(x

): < ;
(T1) y

= t

for T;
(T2) x

: <

h
1

(r

):

for T;
(S1) x

= s

for S;
(S2) y

= y

for S if x

= x

for some < ;

(S3) y

R ` r

(x

): for S if x

= x

for all < .

By the recursion theorem it is enough just to argue that such a choice is
always possible.
The possibility of the choices in (), (T1), (S1), (S2), and (S3) is obvi-
ous. The choice of x

in (T2) is possible by the Baire category theorem,

since the set x

: <

h
1

(r

): is of rst category, being

an at most countable union of nowhere-dense sets (this is the place where
we used the continuum hypothesis!).
It is worthwhile to mention that the conclusion of Theorem 7.2.5 is not
equivalent to the continuum hypothesis. Essentially the same proof works
with a weaker additional set-theoretic axiom, known as the additivity of
category, that the union of less than continuum many meager sets is still
a meager set (see Theorem 8.2.6). However, it has been recently proved
by Stepr ans (1993) that the conclusion of Theorem 7.2.5 is independent of
ZFC set theory.
EXERCISES
1 Show that every function f : C R on a nowhere-dense subset C of R
can be extended to a Darboux function g : R R.
Since there are nowhere-dense subsets C of R of cardinality c, for ex-
ample, the Cantor set, conclude that the family T of Darboux functions
has cardinality 2
c
.
2 Generalize Theorem 7.2.5 to prove the following:
Let ( be the family of all continuous functions from R to R such
that for every h in ( there exists a nonempty open interval I
such that h is not constant on any nonempty open subinterval
of I. If the continuum hypothesis holds then there exists a
strongly Darboux function f such that f + h is not Darboux
for every h (.

7.3 Additive functions and Hamel bases 111
3 Show that if h is a continuous function with the property that every
nonempty open interval contains a nonempty open subinterval on which
h is constant then f + h is strongly Darboux for every strongly Darboux
function f.
Remark: There are continuous nonconstant functions with the property
just described. For example, if C is the Cantor set then we can dene a
function h: [0, 1] R by h(x) =

n=1
x
n
/2
n+1
, where

n=1
x
n
/3
n
=
max(C [0, x]) and x
n
0, 2.
4 Find uncountable sets X, Y R ` such that for every continuous
function f : R R if f[X] Y then f is constant.
5 Prove the following theorem of Sierpi nski and Zygmund: There exists a
function f : R R such that for every continuous function g : R R the
set x R: f(x) = g(x) has cardinality less than c. Hint: Use the fact
that every continuous function h: X R, with X R, has a continuous
extension

h: G R to a G

set G R. (This can be proved by noticing

that the set of all z R for which the limit lim
xz
h(x) does not exist is
an F

set.)
7.3 Additive functions and Hamel bases
A function F : R R is an additive function if F(x + y) = F(x) + F(y)
for every x, y R. Every linear function F(x) = ax is clearly additive and
continuous, and it is quite easy to see that these are the only continuous
additive functions. However, there exist discontinuous additive functions.
The rst example of such a function was constructed by Hamel, with the
use of a Hamel basis, that is, a linear basis of R over , which exists by
Theorem 4.4.1. For its construction we will need some easy facts.
For B R let LIN
Q
(B) denote the smallest linear subspace of R over
containing B. Notice that LIN
Q
(B) can be obtained by closing set B
under the action of the family T of operations 'x, y` x +y and x qx
for every q . Thus LIN
Q
(B) = cl
J
(B), where we are using the notation
of Lemma 6.1.6. In particular, Lemma 6.1.6(a) implies that
[LIN
Q
(B)[ = [B[ + for every B R. (7.8)
From this we immediately conclude that
every Hamel basis H has cardinality continuum (7.9)
since LIN
Q
(H) = R.

112 7 Strange real functions
Proposition 7.3.1 If F : R R is additive then it is linear over , that
is,
F(px +qy) = pF(x) +qF(y)
for every p, q and x, y R.
Proof Let F : R R be additive. Then F(0) +F(0) = F(0 + 0) = F(0),
so
F(0) = 0.
It is easy to prove by induction on n that F(nx) = nF(x) for every
x R. Since F(x) + F(x) = F(x + x) = F(0) = 0, we have also that
F(x) = F(x). Therefore we conclude that
F(nx) = nF(x) for every x R and n Z.
In particular, for 0 < m < we have mF

1
m
x

= F

m
1
m
x

= F(x), that
is, F

1
m
x

=
1
m
F(x). So
F(px) = pF(x) for every x R and p ,
and F(px +qy) = F(px) +F(qy) = pF(x) +qF(y) for every p, q and
x, y R.
Theorem 7.3.2 If H R is a Hamel basis then every function f : H R
can be extended uniquely to an additive function F : R R.
Proof Since every x R has a unique representation x = q
1
b
1
+ +q
m
b
m
in the basis H (that is, b
1
< < b
m
are from H and q
1
, . . . , q
m
`0;
see Section 4.4), we can dene
F(x) =
m

i=1
q
i
f(b
i
). (7.10)
This function is clearly additive and extends f. Its uniqueness follows from
Proposition 7.3.1.
Corollary 7.3.3 There are additive discontinuous functions F : R .
Proof Let H be a Hamel basis and let x H. Dene f : H R by
putting f(x) = 1 and f(y) = 0 for y H ` x. Extend f to the additive
function F. By (7.10) the range of f is equal to . Thus, F cannot be
continuous, since it does not have the Darboux property.
The function from Corollary 7.3.3 has a small (countable) range. It is
not dicult to modify the argument from Corollary 7.3.3 to get an additive
discontinuous function onto R. The next theorem shows that we can even
do a lot better than this.

7.3 Additive functions and Hamel bases 113
Theorem 7.3.4 There exists a Hamel basis H and a function f : H R
such that f
1
(r) is a Bernstein set for every r R.
Proof First notice that if f : H R is such that
f
1
(r) P = for every r R and perfect set P R
then f has the desired property, since the complement of each f
1
(r) con-
tains f
1
(r + 1) and so it must also intersect every perfect set.
Let 'P

, r

`: < c be an enumeration of P R: P is perfect R.

We will construct, by induction on < c, a one-to-one sequence x

: < c
that is linearly independent over and such that x

for every < c.

Notice that this will be enough to construct function f. To see it, extend
x

: < c to a Hamel basis (see Theorem 4.4.1). Dene f(x

) = r

for
< c and extend f arbitrarily to H. Then for every perfect set P R
and every r R there exists a < c such that 'P, r` = 'P

, r

`, so that
f(x

) = r

= r and x

= P. Therefore f
1
(r) P = .
To make the sequence x

: < c linearly independent over for every

< c we will choose
(i) x

` LIN
Q
(x

: < ).
The choice is possible since, by (7.8), [LIN
Q
(x

: < )[ = [[ + <
c. Thus, by the recursion theorem, we can nd a sequence satisfying (i).
To nish the proof it is enough to show that the choice of x

s from the
complement of LIN
Q
(x

: < ) makes the sequence x

: < c linearly
independent over . This part is left as an exercise.
Corollary 7.3.5 There exists a nonmeasurable non-Baire additive func-
tion F : R R that is strongly Darboux.
Proof Let F be an additive extension of f from Theorem 7.3.4. Then
F
1
(r) is a Bernstein set for every r R.
In fact it can be proved that every discontinuous additive function is
neither measurable nor Baire.
Corollary 7.3.6 There exists a Hamel basis H that is neither measurable
nor has the Baire property.
Proof It is enough to notice that the Hamel basis H from Theorem 7.3.4
is a Bernstein set.
It clearly intersects every perfect set. On the other hand, if a H, then
a+H = a+x: x H is disjoint from H. This is the case since otherwise
there would exist x, y H such that a +x = y and so a, x, y H would

114 7 Strange real functions
be linearly dependent. Thus a + H R ` H. But a + H intersects every
perfect set P, since H(a+P) = for every perfect set P, the set a+P
being perfect.
Notice that there are Hamel bases that are measurable and have the
Baire property (see Exercise 2).
Corollary 7.3.7 There exists a Hamel basis H such that [H P[ = c for
every perfect set P.
Proof It is clear that the basis H from Theorem 7.3.4 has the desired
property, since HP =

rR
(f
1
(r) P) is the union of continuum many
nonempty disjoint sets.
Next, we will construct an additive function F whose graph is connected
as a subset of the plane. The graph of such a function is called a Jones
space.
Theorem 7.3.8 There exists a discontinuous additive function F : R R
whose graph is connected.
Proof Let H be the Hamel basis from Corollary 7.3.7 and let T be
the family of all closed subsets P of the plane such that p[P] contains
a perfect set, where p is the projection of the plane onto the x-axis. Let
P

: < c be an enumeration of T. Dene, by induction on < c, a

sequence 'x

, y

` R
2
: < c such that
(i) x

(H p[P

]) ` x

: < ,
(ii) 'x

, y

` P

.
The choice as in (i) can be made, since [H p[P

][ = c.
Now dene f(x

) = y

and extend f to H in an arbitrary way. Let F be

the additive extension of f. We will show that the graph of F is connected.
To obtain a contradiction assume that the graph of F is not connected.
Then there exist disjoint open subsets U and V of R
2
such that U F =
= V F and F U V . Let P = R
2
` (U V ). We will show that
P F = , which will give us the desired contradiction.
If P T then there exists a < c such that P = P

, and so 'x

, y

`
P

F = P F. Thus it is enough to prove that P T. Clearly, P is

closed. So we have to show only that
p[P] contains a perfect subset.
In order to prove it we rst show that there exists an x R such that
U (x R) = = V (x R). (7.11)

7.3 Additive functions and Hamel bases 115
To obtain a contradiction, assume that (7.11) is false. Then the sets U
1
=
p[U] and V
1
= p[V ] are disjoint. But they are open in R, since the projection
of an open set is open. They are clearly nonempty, since U and V are
nonempty. Finally, U
1
V
1
= p[U V ] p[F] = R, that is, sets U
1
and V
1
violate the connectedness of R. Condition (7.11) has been proved.
Now let x
0
R be such that U (x
0
R) = = V (x
0
R)
and let y
1
, y
2
R be such that 'x
0
, y
1
` U and 'x
0
, y
2
` V . Then there
exists an open interval I = (a, b) containing x
0
such that I y
1
U
and I y
2
V . But for every x I the interval J
x
= x [y
1
, y
2
]
is a connected set intersecting both U and V . Thus J
x
must intersect P
as well, since it is connected. Therefore x = p[J
x
P] p[P] for every
x I, that is, p[P] I. This nishes the proof that P T.
We will nish this section with following theorem of Erd os and Kakutani
(1943).
Theorem 7.3.9 The continuum hypothesis is equivalent to the existence
of a countable partition H
n
: n < of R ` 0 such that every H
n
is a
Hamel basis.
In the proof of the theorem we will use the following two lemmas, the
rst of which is due to Erd os and Hajnal (Erd os 19789). They are also
very interesting in their own right.
Lemma 7.3.10 If f : A B , where [A[ =
2
and [B[ =
1
, then for
every n < there exist B
0
[B]
n
and A
0
[A]
2
such that f(a
0
, b
0
) =
f(a
1
, b
1
) for every a
0
, a
1
A
0
and b
0
, b
1
B
0
.
Proof Fix n < . First we will show that
a A B
a
[B]
n
m
a
< b B
a
(f(a, b) = m
a
). (7.12)
To see this, notice that for every a A the sets S
m
a
= b B: f(a, b) = m
form a countable partition of an uncountable set B. Thus there is an
m = m
a
for which the set S
m
a
is uncountable and any n-element subset B
a
of S
m
a
satises (7.12).
Now let F : A [B]
n
be such that for every a A the pair
F(a) = 'B
a
, m
a
` satises (7.12). The set [B]
n
has cardinality
1
, so
there exists 'B
0
, m` [B]
n
such that A
0
= F
1
(B
0
, m) has cardinality

2
. But then, if a
i
A
0
for i < 2 then F(a
i
) = 'B
0
, m` and
f(a
i
, b) = m
ai
= m
for every b B
ai
= B
0
. So the sets A
0
and B
0
have the desired property.

116 7 Strange real functions
Notice that Lemma 7.3.10 can also be expressed in the following graph-
theoretic language. Let A and B be disjoint sets of cardinality
2
and
1
,
respectively, and let G = 'V, E` be the bipartite graph between A and B,
that is, V = A B and E = a, b: a A & b B. Then for any n <
and any coloring f : E of the edges of graph G there are A
0
[A]
2
and B
0
[B]
n
such that the subgraph G
0
= 'A
0
B
0
, E {(A
0
B
0
)` of
G generated by A
0
B
0
is monochromatic, that is, the coloring function f
is constant on the edges of G
0
.
Lemma 7.3.11 Assume that c >
1
and let H [R]
2
. Then for every
partition H
n
: n < of R there exist n < and disjoint sets A
0
[H]
2
and B
0
[H]
2
such that
a +b H
n
for every a A
0
and b B
0
.
Proof Choose disjoint A [H]
2
and B [H]
1
. Dene f : A B
by
f(a, b) = m if and only if a +b H
m
.
Then the sets A
0
and B
0
from Lemma 7.3.10 used with f and n = 2 have
the desired properties.
Proof of Theorem 7.3.9 : Let H be any Hamel basis and suppose that
there exists a partition H
n
: n < of R ` 0 into Hamel bases.
To obtain a contradiction assume that c >
1
. Then [H[ = c
2
so
we can choose A
0
and B
0
as in Lemma 7.3.11. Take dierent a
0
, a
1
A
0
and b
0
, b
1
B
0
. Then the numbers x
ij
= a
i
+ b
j
for i, j 2 are dierent
and belong to the same H
n
. However,
x
00
x
10
= (a
0
+b
0
) (a
1
+b
0
) = (a
0
+b
1
) (a
1
+b
1
) = x
01
x
11
,
contradicting the fact that x
00
, x
10
, x
01
, x
11
H
n
are linearly independent
over .
: Let z

R: <
1
be a Hamel basis and for every
1
dene
L

= LIN
Q
(z

: < ). Notice that L

for <
1
, that
L

=

<
L

for limit
1
, and that L
1
= R (compare this with
Lemma 6.1.6). Thus the sets L
+1
L

: <
1
form a partition of
R ` L
0
= R ` 0. For every <
1
choose a bijection f

: L
+1
L

and let f =

<1
f

. Then f : R ` 0 . Let H
n
= f
1
(n).
The sets H
n
: n < form a partition of R ` 0. To nish the proof
it is enough to show that every H
n
is a Hamel basis.
So choose n < and let x

= (L
+1
L

) f
1
(n) for <
1
. It is
enough to prove that
L

= LIN
Q
(x

: < ) (7.13)

7.3 Additive functions and Hamel bases 117
for every
1
, since then H
n
= x

: < c spans R and x

: < c
is linearly independent by Exercise 1. Condition (7.13) can be proved by
induction on
1
. So let
1
be such that (7.13) holds for every
< .
If is a limit ordinal then
L

<
L

<
LIN
Q
(x

: < ) = LIN
Q
(x

: < ).
If = + 1 then L

= LIN
Q
(x

: < ) and
x

LIN
Q
(x

: < ) = x

L
+1
so LIN
Q
(x

: < + 1) L
+1
.
To prove the other inclusion, recall that x

L
+1
L

. Therefore
x

= qz

i<m
q
i
b
i
for some q, q
0
, . . . , q
m1
and b
0
, . . . , b
m1

x

: < . Moreover, q = 0 because x

= LIN
Q
(x

: < ). So
z

=
1
q

i<m
q
i
b
i

LIN
Q
(x

: < + 1).
Now z

LIN
Q
(x

: < + 1), and so we can conclude that

L
+1
LIN
Q
(x

: < + 1).
EXERCISES
1 Complete the proof of Theorem 7.3.4 by showing that if x

R: < c
is such that x

LIN
Q
(x

: < ) for every < c, then x

R: < c
is linearly independent over .
2 Let C R be the Cantor set. Show that there is a Hamel basis contained
in C. Since C is nowhere dense and has measure zero, it follows that Hamel
bases can be measurable and have the Baire property. Hint: Use the fact
that C +C = a +b: a, b C contains the unit interval [0, 1].
3 Let be an equivalence relation on R dened by x y if and only if
x y . If V is a selector from the family of all equivalence classes of
then V is called a Vitali set. It is known that any Vitali set is neither
measurable nor has the Baire property.
Construct a Vitali set V such that
(a) V +V = R,
(b) V +V is a Bernstein set.
4 Prove that the graph of any additive discontinuous function F : R R
is dense in R
2
.

118 7 Strange real functions
7.4 Symmetrically discontinuous functions
This section is motivated in part by the following generalization of the con-
tinuity of real functions. A function f : R R is said to be symmetrically
continuous at the point x R if
lim
h0
+
[f(x h) f(x +h)] = 0,
that is, if for every > 0 there exists a d > 0 such that
(0, d) S

x
,
where S

x
= h > 0: [f(x h) f(x + h)[ < . A function f : R R is
symmetrically continuous if it is symmetrically continuous at every point
x R.
Clearly, every continuous function is symmetrically continuous. The
converse implication is not true, since there are symmetrically continuous
functions that are not continuous in the usual sense. For example, the
characteristic function

0
of a singleton set 0 is symmetrically contin-
uous and discontinuous at 0. However, symmetrically continuous functions
cannot behave too badly in this respect: It can be proved that the set of
points of discontinuity of a symmetrically continuous function must be of
rst category and have measure zero.
The study of symmetrically continuous functions is an important sub-
ject in real analysis, motivated by Fourier analysis. In this section, how-
ever, this notion serves merely as a motivation, since we will study here the
functions that are not symmetrically continuous.
It is easy to nd a function that is not symmetrically continuous. For
example, a characteristic function of any nontrivial interval is symmetri-
cally discontinuous at the endpoints of that interval. In fact, it is also not
dicult to construct a function f : R R that is nowhere symmetrically
continuous. Such a function must have the property that for every x R
there exists an > 0 such that
d > 0 h (0, d) ([f(x h) f(x +h)[ ), (7.14)
or, equivalently, such that
(0, d) S

x
for every d > 0. (7.15)
To get such an example dene f as the characteristic function

H
of any
dense Hamel basis H. Notice that such a basis exists by Theorem 7.3.4. To
see that

H
is nowhere symmetrically continuous, take x R, (0, 1),
and an arbitrary d > 0. We have to nd h (0, d) with the property

7.4 Symmetrically discontinuous functions 119
that [

H
(x h)

H
(x +h)[ . So let q
1
b
1
+ + q
n
b
n
be a repre-
sentation of x in basis H. Then, by the density of H, there exists a
b H (x d, x) ` b
1
, . . . , b
n
. Pick h = x b. Then h (0, d) and

H
(x h) =

H
(b) = 1. On the other hand,

H
(x + h) = 0 as x + h =
2x b = 2q
1
b
1
+ +2q
n
b
n
b H. Therefore [

H
(xh)

H
(x+h)[ =
1 > .
Considerations of how badly nowheresymmetrically continuous func-
tions can behave led several people
1
to to ask whether there exists a func-
tion f satisfying (7.14) with the quantiers d > 0 h (0, d) replaced by
the reversed quantiers d > 0 h (0, d), that is, whether there exists a
function f : R R such that for every x R there exists an > 0 with
d > 0 0 < h < d ([f(x h) f(x +h)[ ),
or, equivalently, such that
(0, d) S

x
= for some d > 0.
Replacing and d by their minima, we can rephrase this problem by asking
whether there exists a function f : R R such that
() for every x R there exists a d > 0 with
(0, d) S
d
x
= , (7.16)
where S
d
x
= h > 0: [f(x h) f(x +h)[ < d.
A function f satisfying () will be called a uniformly antisymmetric func-
tion.
The existence of uniformly antisymmetric functions can be inferred from
the following theorem due to Ciesielski and Larson (19934).
Theorem 7.4.1 There exists a partition { = P
n
: n N of R such that
for every x R the set
S
x
=

nN
h > 0: x h, x +h P
n
(7.17)
is nite.
Before proving this theorem, we rst show how to use it to construct a
uniformly antisymmetric function.
Corollary 7.4.2 There exists a uniformly antisymmetric function
f : RN.
1
Evans and Larson in 1984; Kostyrko in 1991.

120 7 Strange real functions
Proof Let { = P
n
: n N be a partition satisfying (7.17) and for x R
dene
f(x) = n if and only if x P
n
.
Then f is uniformly antisymmetric, since for every x R the set
S
1
x
= h > 0: [f(xh)f(x+h)[ < 1 = h > 0: f(xh) = f(x+h) = S
x
is nite, so d = min(S
x
1) satises (), as S
d
x
= S
1
x
.
Proof of Theorem 7.4.1 Let H be a Hamel basis. Then every x
R has a unique representation x = q
1
b
1
+ + q
n
b
n
in the basis H
(that is, b
1
< < b
n
are from H and q
1
, . . . , q
n
` 0). Let B
x
=
b
1
, . . . , b
n
H and dene c
x
: B
x
by putting c
x
(b
i
) = q
i
for
1 i n. Moreover, extend each c
x
to c
x
: H by putting
c
x
(b) =

c
x
(b) for b B
x
,
0 otherwise.
Then, for every x R,
x =

bBx
c
x
(b) b =

bH
c
x
(b) b.
We will start the argument by proving that there exists a countable set
D = g
n
: n < of functions from H into such that
x R n < (c
x
g
n
). (7.18)
To see it, let | = (p, q) H: p, q and notice that | is countable.
Dene
D =

i<n
q
i

Ui

H
: n and 'q
i
, U
i
` | for every i < n

.
Clearly, D is countable, since it is indexed by a countable set

n<
(|)
n
.
To see that it satises (7.18), take x R and nd a family U
b
| : b B
x

of disjoint sets such that b U

b
for every b B
x
. Then g =

bBx
c
x
(b)

U
b
belongs to D and c
x
(b) = g(b) for every b B
x
. So c
x
g. Condition
(7.18) has been proved.
Now let g
n
: n < be as in (7.18). Dene f : R by
f(x) = minn < : c
x
g
n

and let P
n
= f
1
(n) for every n < . We will show that the partition
{ = P
n
: n N of R satises (7.17).

7.4 Symmetrically discontinuous functions 121
So x x R and let h S
x
. We will show that
c
x+h
(b) 0, c
x
(b), 2 c
x
(b) for every b H. (7.19)
This will nish the proof, since 0, c
x
(b), 2 c
x
(b) = 0 for b H ` B
x
and
so there are at most 3
]Bx]
< numbers x + h such that h S
x
. Thus S
x
is nite.
To see (7.19), notice that (x +h) + (x h) = 2x, so

bH
c
x+h
(b) b +

bH
c
xh
(b) b = 2

bH
c
x
(b) b.
In particular,
c
x+h
(b) + c
xh
(b) = 2 c
x
(b) (7.20)
for every b H.
Now, if x +h, x h P
n
then c
x+h
g
n
and c
xh
g
n
. Hence
c
x+h
(b) = g
n
(b) = c
xh
(b) for every b B
x+h
B
xh
.
Therefore, by (7.20), c
x+h
(b) = c
x
(b) for every b B
x+h
B
xh
. But for
b H ` B
x+h
we have c
x+h
(b) = 0, and for every b B
x+h
` B
xh
we have
c
x+h
(b) = c
x+h
(b) + c
xh
(b) = 2 c
x
(b). Thus c
x+h
(b) 0, c
x
(b), 2 c
x
(b) for
every b H. This nishes the proof of condition (7.20) and Theorem 7.4.1.

An interesting open problem is whether or not there exists a uniformly

antisymmetric function with nite or bounded range. Some partial results
are known in this direction. Komj ath and Shelah (19934) proved that
there is no function f : R R with nite range such that all the sets S
x
from Theorem 7.4.1 are nite. It has also been proved by Ciesielski (1995
6) that the range of any uniformly antisymmetric function must have at
least four elements. The next theorem, due to Ciesielski and Larson (1993
4), shows only that the range of a uniformly antisymmetric function must
have at least three elements.
Theorem 7.4.3 If f : R 0, 1 then f is not uniformly antisymmetric.
Proof To obtain a contradiction assume that there exists a uniformly
antisymmetric function f : R 0, 1 and for every x R let n
x
N be
such that for every h (0, 1/n
x
),
[f(x h) f(x +h)[ 1/n
x
.
Fix n N such that L = x R: n
x
= n is uncountable. Then
f(x h) = f(x +h) for every x L and h (0, 1/n). (7.21)

122 7 Strange real functions
Since (
k
n
,
k+1
n
] L: k Z forms a countable partition of L, at least one
of these sets must be uncountable. In particular, we can choose x, y, z L
such that x < y < z < x+1/n. Now we would like to nd points a, b, c R
such that x, y, and z are the midpoints of pairs 'a, b`, 'b, c`, and 'a, c`,
respectively, that is, such that a +b = 2x, b +c = 2y, and a +c = 2z. It is
easy to see that the points
a = x y +z, b = x +y z, c = x +y +z
have these properties. Moreover, for h = z y (0, 1/n) we have a = x+h
and b = x h. Hence, by (7.21), f(a) = f(b). Similarly, f(b) = f(c) and
f(c) = f(a). But this is impossible, since the points a, b, and c are distinct
and f attains only two values.
We proved Corollary 7.4.2, via Theorem 7.4.1, by showing that for every
x R the set S
1
x
= S
x
of exceptional points, being nite, is bounded away
from 0. What if we allow the sets S
x
to be countable? Certainly, such sets
do not have to be bounded away from 0. But we can replace the condition
of being bounded away from 0 by the weaker condition of being almost
bounded away from 0, in the sense that S
x
(0, d) is countable for some
d > 0. Can we prove then an analog of Theorem 7.4.1? Surprisingly,
the answer depends on the continuum hypothesis, as is proved in the next
theorem.
The general form of this theorem is due to Ciesielski and Larson (1993
4). However, implication (i)(ii) was rst proved by Sierpi nski (1936).
The equivalence of (i) and (ii) is also implicitly contained in a work of
Freiling (198990).
Theorem 7.4.4 The following conditions are equivalent.
(i) The continuum hypothesis.
(ii) There exists a partition { = A
0
, A
1
of R such that for every x R
the set S
x
=

i<2
h > 0: x h, x +h A
i
is at most countable.
(iii) There exists a function f : R 0, 1 such that for every x R there
is a d > 0 with the property that

S
d
x
(0, d)

.
Proof (i)(ii): Let H = b

: <
1
be a Hamel basis. For x R` 0
let q(x) = q
n
, where x = q
1
b
1
+ +q
n
b
n
is the unique representation of
x in the basis H, with
1
< <
n
and q
i
= 0. Put
x A
0
if and only if q(x) > 0,
and A
1
= R`A
0
. We will show that the partition { = A
0
, A
1
satises (ii).

7.4 Symmetrically discontinuous functions 123
For <
1
let K

= LIN
Q
(b

: ). Notice that R =

<1
K

and that every K

is countable. We will show that if x K

then S
x
K

.
This will nish the proof.
But it is easy to see that for h R ` K

and x K

,
h A
0
if and only if x +h A
0
,
since q(h) = q(x + h). However, h and h cannot belong to the same A
i
for h = 0, so S
x
K

.
(ii)(iii): For x R dene f(x) = i if x A
i
. Then S
d
x
= S
x
is
countable for every x R and d (0, 1).
(iii)(i): Let f be as in (iii), and for every x R let n
x
N be such
that the set
C
x
= S
1/nx
x
(0, 1/n
x
) = h (0, 1/n
x
): f(x h) = f(x +h)
is countable.
To obtain a contradiction, assume that the continuum hypothesis fails
and let B be a linearly independent subset of R over of cardinality
2
.
Choose K B of cardinality
2
such that for some n N we have n
x
= n
for all x K. Let U be an open interval of length less than 1/n such that
the set L = K U has cardinality
2
. Then, in particular,
f(x h) = f(x +h) for x L and h (0, 1/n) ` C
x
, (7.22)
and
[x y[ < 1/n for every x, y L. (7.23)
Dene, by transnite induction, a sequence 't

L: <
2
` such that
t

L ` T

for every <

2
, (7.24)
where T

is the smallest linear subspace of R containing t

: < and
such that
C
x
T

for every x T

. (7.25)
Such T

is obtained by applying Lemma 6.1.6(a) to the set Z = t

: <
and the family T of operations x C
x
, 'x, y` x + y, and x qx for
every q . Then T

has cardinality
1
, so the induction can be done
easily.
Now put x = t
0
, z = t
1
and, for 0 < <
1
, consider the numbers
[ x+t

[. All these numbers are dierent, so there is 0 < <

1
such that
[ x +t

[ C
z
. (7.26)

124 7 Strange real functions
Put y = t

and proceed as in Theorem 7.4.3. Dene

a = x y +z, b = x +y z, c = x +y +z. (7.27)
Then (a +b)/2 = x, (b +c)/2 = y, and (c +a)/2 = z. We will show that
f(a) = f(b), f(b) = f(c), f(c) = f(a), (7.28)
which will give us the desired contradiction, since the points a, b, and c are
distinct and f admits only two values.
To prove (7.28), notice rst that (7.28) follows from (7.22) as long as
[a x[ < 1/n, [b y[ < 1/n, [c z[ < 1/n (7.29)
and
[a x[ C
x
, [b y[ C
y
, [c z[ C
z
. (7.30)
But (7.29) follows easily from (7.27) and (7.23). Finally, (7.30) can be
proved as follows.
[a x[ C
x
, since otherwise we would have z y = a x T
1
and
z y +T
1
T
1
, contradicting z = t
1
T
1
.
[b y[ C
y
, since otherwise we would have z x = y b T
1
and
z x +T
1
T
1
, contradicting z = t
1
T
1
.
[c z[ C
z
, since otherwise we would have [ x + t

[ = [ x + y[ =
[c z[ C
z
, contradicting (7.26).
To motivate the last theorem of this section we need to reformulate
Theorem 7.4.1 in another language. For this we need the following deni-
tions. We say that partition { of R is sum free if for every x R the
equation a+b = x has less than solutions with a and b being in the same
element of the partition, that is, when for every x R the set
T
x
= 'a, b`: a +b = x and a, b P for some P {
is nite. Similarly, we say that partition { of R is dierence free if the
set
D
x
= 'a, b`: a b = x and a, b P for some P {
is nite for every x R, x = 0.
In this language Theorem 7.4.1 reads as follows.
Corollary 7.4.5 There exists a countable partition { of R that is sum
free.
Proof It is enough to notice that T
2x
= 'x +h, x h`: [h[ S
x
0,
where S
x
is from Theorem 7.4.1.
Can we prove the same results about -dierence-free partitions of R?
The answer is given by the next theorem, due to Ciesielski (1996).

7.4 Symmetrically discontinuous functions 125
Theorem 7.4.6 The continuum hypothesis is equivalent to the existence
of a countable partition { of R that is dierence free.
Proof The proof of the theorem is similar to that of Theorem 7.3.9.
: To obtain a contradiction assume that c
2
and that there exists
a countable partition { of R that is dierence free.
Let H be any Hamel basis. Since [H[ = c
2
by Lemma 7.3.11 we
can choose disjoint sets A
0
[H]
2
and B
0
[H]
2
and P { such that
a +b P for every a A
0
and b B
0
.
Take dierent b
0
, b
1
B
0
. Then
(a +b
0
) (a +b
1
) = b
0
b
1
for every a A
0
, that is, D
b0b1
contains all pairs 'a + b
0
, a + b
1
`. This
contradicts the assumption that D
x
is nite for x = b
0
b
1
= 0.
: Represent R as the union of an increasing sequence 'V

: <
1
`
of countable linear subspaces V

of R over such that V

<
V

for
every limit ordinal <
1
. Such a sequence exists by Lemma 6.1.6(b)
applied to the family T of operations 'x, y` x +y and x qx for every
q . For convenience we will also assume that V
0
= .
Thus V
+1
` V

: <
1
is a partition of R into countable sets. For
<
1
let p

n
: n < be an enumeration of V
+1
` V

. By induction on
<
1
we will dene one-to-one functions f : V
+1
` V

such that the

following inductive condition holds
2
f(p

n
) ` f(p): p V

& p = p

n
p

j
for some j n.
We will show that the partition { = f
1
(n): n < of R is dierence
free.
So choose an arbitrary x = p

n
= 0 and consider the pairs 'a, b` satisfying
a b = x with a and b being from the same element of {, that is, such
that f(a) = f(b). It is enough to show that a, b p

j
: j n = .
Let a = p

m
and b = p

k
. Then p

m
p

k
= p

n
. Notice that =
max, , must be equal to at least two of , , and , since otherwise
the number p with the index would belong to V

. Moreover, = , since
otherwise f(p

m
) = f(a) = f(b) = f(p

k
), contradicting the fact that f is
one-to-one on V
+1
` V

. We are left with two cases:

If = > then p

m
p

n
= a x = b = p

k
V

. So f(p

m
) = f(a) =
f(b) = f(p

m
p

n
) implies that m < n and a = p

m
p

j
: j n.
2
We dene here f separately on each set from the partition {V
+1
\ V: <
1
} of R.
Formally, we should be using a dierent symbol, such as f, for such a part of f, and
dene f as the union of all fs. However, this would obscure the clarity of this notation.

126 7 Strange real functions
If = > then p

k
+ p

n
= b + x = a = p

m
V

. So f(p

k
) = f(b) =
f(a) = f(p

k
+p

n
) implies that k < n and b = p

k
p

j
: j n.
EXERCISES
1 Prove that if f is uniformly antisymmetric then it does not have the
Baire property. Hint: First prove the following fact due to Kuratowski:
For every function f : R R with the Baire property there
exists a rst-category set S such that f[
R\S
: R ` S R is
continuous.
2 (Project) Consider the following classes of subsets of R:
B = B R: B is a Bernstein set;
H = H R: H is a Hamel basis;
1 = V R: V is a Vitali set;
T = T R: B is a transcendental basis of R over .
Moreover, if T is any of the preceding families let
T

= F R: F +F T.
Find the complete intersection/subset relations among the classes B, H, 1,
T , B

, H

, 1

, and T

.

Part IV
When induction is too
short
127

Chapter 8
Martins axiom
8.1 RasiowaSikorski lemma
The previous chapter was devoted to constructing objects by transnite
induction. A typical scheme for such constructions was a diagonalization
argument like the following. To nd a subset S of a set X concerning a
family { = P

: < we chose S = x

X: < by picking each

x

to take care of a set P

. But what can we do if the cardinal is too big

compared to the freedom of choice of the x

s; for example, if the set X has

cardinality less than ?
There is no absolute answer to this question. In some cases you can
do nothing. For example, if you try to construct a subset S of dierent
from every set from the family {() = B

: < c, then you are obviously

condemned to failure. The inductive construction does not work, since you
would have to take care of continuum many conditions, having the freedom
to choose only countably many points for S.
In some other cases you can reduce a family { to the appropriate size.
This was done, for example, in Theorem 6.3.7 (on the existence of Bernstein
sets) in which we constructed a nonmeasurable subset B of R
n
: The natural
family { = L of cardinality 2
c
was replaced by the family {
0
= P

: < c
of all perfect subsets of R
n
. The dicult part of the proof, that is, Theo-
rem 6.3.6, was to show that the family {
0
does the job.
Yet in other cases, such as Theorems 7.3.9, 7.4.4, and 7.4.6, the induc-
tion could be performed in only
1
steps while we had to take care of a
family { of cardinality c. These cases resulted in us assuming that c =
1
.
In other words, in order to rescue a diagonal argument, we had to assume
an additional set-theoretic assumption, the continuum hypothesis. In fact,
in Theorems 7.3.9, 7.4.4, and 7.4.6 we also proved that this assumption
was necessary.
129

130 8 Martins axiom
The continuum hypothesis was also assumed in the proof of Theo-
rem 7.2.5. In this theorem, however, we did not show that the continuum
hypothesis is a necessary assumption, though we stated that the theorem
cannot be proved in ZFC alone. The reason for this is that Theorem 7.2.5
can also be proved under weaker set-theoretic assumptions, which can be
true even if the continuum hypothesis fails.
In remainder of this text we will be mainly interested in studying dier-
ent kinds of set-theoretic axioms that will allow us to solve the problem of
having a too short induction. For this we will need some more denitions
and notations.
Consider a partially ordered set 'P, `. A subset D P is said to be
dense in P provided for every p P there exists a d D such that d p.
Examples 1. If X = and 'P, ` = '{(X), ` then D = is dense in
P. Notice that is the smallest element of P.
In fact, D = m is dense in 'P, ` if and only if m is the smallest
element of P. Also, if 'P, ` has the smallest element m, then D P is
dense in 'P, ` if and only if m D. Dense sets such as these are too easy
to describe to be of much interest. To avoid them, we will usually study
partially ordered sets without a smallest element.
2. If 'P, ` = 'R, ` then the set D = (, 0) is dense in 'P, `. Notice
that the word dense is used here in a dierent sense than it is usually
used for the linearly ordered sets (see (8.1)). However, there will be very
little chance to confuse these two notions of density, since we will usually
use our new denition of density for partially ordered sets that are not
linearly ordered.
3. Let 'X, ` be any topological space. If 'P, ` = ' ` , ` then any
base B in X is dense in P.
A subset F of a partially ordered set 'P, ` is a lter in P if
(F1) for every p, q F there is an r F such that r p and r q, and
(F2) if q F and p P are such that q p then p F.
Note that a simple induction argument shows that condition (F1) is equiv-
alent to the following stronger condition.
(F1
t
) For every nite subset F
0
of F there exists an r F such that r p
for every p F
0
.

8.1 RasiowaSikorski lemma 131
Examples 1. For any chain F in a partially ordered set 'P, ` the family
F

= p P: q F(q p)
is a lter in P.
2. Let ( be any family of subsets of a nonempty set X that is closed
under nite intersections and let 'P, ` = '(, `. If x X then the family
F
x
= Y (: x Y is a lter in P. A lter F in such 'P, ` is called a
principal lter if it is of this form, that is, if F = F
x
for some x X.
3. Let ( be a nonempty family of subsets of an innite set X and let
[X[. If ( is closed under nite unions then ( [X]
<
is a
lter in 'P, ` = '(, `. If ( is closed under nite intersections then
F

= Y (: [X ` Y [ < is a lter in 'P, ` = '(, `.

4. Let 'P, ` = '{(X), `. Then F P is a lter in P if (1) A B F
provided A, B F; and (2) if A B X and A F then B F.
(Compare this with Exercise 3 in Section 4.4.)
Before we formulate our next example we introduce the following nota-
tion. For nonempty sets X and Y we will use the symbol Func(X, Y ) to
denote the family of all partial functions from X into Y , that is,
Func(X, Y ) =

Y
D
: D X

.
Also, for an innite cardinal number we put
Func

(X, Y ) =

Y
D
: D [X]
<

= s Func(X, Y ): [s[ < .

5. Let T Func(X, Y ) be such that for every g
0
, g
1
T
if g
0
g
1
Func(X, Y ) then g
0
g
1
T.
(In particular, Func

(X, Y ) satises this condition.) Then for every f

Func(X, Y ) the set G
f
= g T: g f is a lter in 'P, ` = 'T, `. To
see (F1) let g
0
, g
1
G
f
. Then g
0
g
1
f, so h = g
0
g
1
Func(X, Y ).
Thus h T and, clearly, h g
0
and h g
1
. To see (F2) notice that
g G
f
, h T, and g h imply that h g f, so h G
f
.
In what follows we will very often use the following fact. Its main part
can be viewed as a kind of converse of the previous Example 5.

132 8 Martins axiom
Proposition 8.1.1 Let 'P, ` = 'T, ` for some T Func(X, Y ). If
F P is a lter in P then f =

F is a function and F g T: g f.
Moreover,
(a) if F intersects every set D
x
= s P: x dom(s) for x X then
dom(f) = X;
(b) if F intersects every set R
y
= s P: y range(s) for y Y then
range(f) = Y .
Proof Let F P be a lter in P and let g
0
, g
1
F. To prove that
f is a function it is enough to show that g
0
(x) = g
1
(x) for every x
dom(g
0
) dom(g
1
). So let x dom(g
0
) dom(g
1
) and let h F be
such that h g
0
and h g
1
. Then h g
0
and h g
1
. In particular,
x dom(h) and g
0
(x) = h(x) = g
1
(x).
The inclusion F g T: g f is obvious.
To see (a) let x X. Then there exists an s F D
x
. But s f, so
x dom(s) dom(f).
Condition (b) is proved similarly.
In what follows we will often use lters in partially ordered sets to con-
struct functions in a manner similar to that of Proposition 8.1.1. Usually we
will be interested in the entire functions, and part (a) of Proposition 8.1.1
suggests how to achieve this goal. This and the next theorem lead to the
following denition.
Let 'P, ` be a partially ordered set and let T be a family of dense
subsets of P. We say that a lter F in P is T-generic if
F D = for all D T.
The partial orders used in the context of T-generic lters will often be called
forcings. Also, if 'P, ` is a forcing then elements of P will sometimes be
referred to as conditions. For conditions p, q P we say that p is stronger
than q provided p q.
In this terminology, if all sets D
x
are dense in P and D
x
: x X T
then for every T-generic lter F in P the domain of f =

F is equal
to X, that is, f : X Y . Similarly, if all sets R
y
are dense in P and
R
y
: y Y T then range(

F) = Y for every T-generic lter F in P.

The next theorem, due to Rasiowa and Sikorski, shows that there are
some interesting T-generic lters.
Theorem 8.1.2 (RasiowaSikorski lemma) Let 'P, ` be a partially or-
dered set and p P. If T is a countable family of dense subsets of P then
there exists a T-generic lter F in P such that p F.

8.1 RasiowaSikorski lemma 133
Proof Let T = D
n
: n < . We dene a sequence 'p
n
: n < ` by
induction on n < . We start by picking p
0
D
0
such that p
0
p. We
continue by choosing p
n+1
D
n+1
such that p
n+1
p
n
.
Now let E = p
n
: n < and put F = E

= p P: q E(q p).
Then F is a lter in P intersecting every D T.
The RasiowaSikorski lemma is one of the most fundamental facts that
will be used in the remaining sections. Its importance, however, does not
come from its power. Its proof is too simple for this. It is the language
of generic lters it employs that makes it so useful. In particular, it mo-
tivates the dierent generalizations described in the next sections, which
are consistent with ZFC and can be used for our problem of a too short
induction.
In most of the applications of the RasiowaSikorski lemma and its gen-
eralizations the intuition behind the proofs comes from an attempt at prov-
ing the theorem by (transnite) induction. More precisely, a partial order
P used to construct an object will usually be built on the basis of an at-
tempted inductive construction of the object. That is, conditions (elements
of P) will be chosen as a description of the current stage of induction.
The inductive steps will be related to the dense subsets of P in the sense
that the density of a particular set D
x
= p P: (p, x) will be equivalent
to the fact that at an arbitrary stage q of the inductive construction we
can make the next inductive step by extending the condition q to p having
the property (p, x). In particular, the family T of dense subsets of P will
always represent the set of all inductive conditions of which we have to take
care, and a T-generic lter in P will be an oracle that takes care of all
our problems, and from which we will recover the desired object.
Evidently, if the number [T[ of conditions we have to take care of is
not more than the number of steps in our induction, then usually the
(transnite) induction will be powerful enough to construct the object,
and the language of forcing will be redundant. In particular, this will be
the case for all the applications of the RasiowaSikorski lemma presented
in the rest of this section. These applications, however, are presented here
to see the use of the generic-lters technique in the simplest situations.
Moreover, in the next sections the same theorems will be either generalized
or used for some motivation.
To state the next theorem let us recall that, for a linearly ordered set
'X, `, a subset D of X is dense in X if
for every x, y X with x < y there is a d D such that x < d < y. (8.1)
A linearly ordered set 'X, ` is said to be dense if it is dense in itself, that
is, if X is dense in X.

134 8 Martins axiom
Theorem 8.1.3 Any two countable dense linearly ordered sets, neither of
which has a rst or a last element, are order isomorphic.
Proof Let 'X, ` and 'Y, _` be two linearly ordered sets as in the theorem.
The inductive proof of the theorem may go as follows. Enumerate X and
Y as X = x
n
: n < and Y = y
n
: n < . Construct, by induction on
n < , a sequence h
0
h
1
h
2
of functions such that each h
n
is an
order isomorphism between X
n
[X]
<
and Y
n
[Y ]
<
, where x
n
X
n
and y
n
Y
n
for every n < . Then h =

n<
h
n
is an order isomorphism
between X and Y . The dicult part of the proof is to extend h
n
to h
n+1
while maintaining the condition x
n+1
X
n+1
and y
n+1
Y
n+1
.
To translate this proof into the language of partially ordered sets, let
P be the set of all possible functions h
n
: X
n
Y
n
from our inductive
construction, that is, the set of all nite partial isomorphisms from X
to Y :
P = h Func

(X, Y ): h is strictly increasing.

Consider P to be ordered by reverse inclusion, that is, 'P, ` = 'P, `. For
x X and y Y put
D
x
= h P: x dom(h) and R
y
= h P: y range(h).
These sets are the counterparts of the conditions x
n
X
n
and y
n
Y
n
.
We will show that
the sets D
x
and R
y
are dense in P, (8.2)
which is a translation of the fact that the inductive step of our inductive
construction can always be made.
Before we prove (8.2), let us show how it implies the theorem. Let
T = D
x
: x X R
y
: y Y . Then T is countable since the sets X
and Y are countable. Thus, by the RasiowaSikorski lemma, there exists a
T-generic lter F in P. Then, by Proposition 8.1.1, f =

F is a function
from X onto Y . It remains only to show that f is strictly increasing.
To see this, let x
0
, x
1
X with x
0
< x
1
. Then there are g
0
, g
1
F
such that x
0
dom(g
0
) and x
1
dom(g
1
). Let g F be such that g g
0
and g g
1
. Then g
0
g
1
g f. So f(x
0
) = g(x
0
) g(x
1
) = f(x
1
),
since g is strictly increasing. Thus f is strictly increasing. We have proved
that f : X Y is an order isomorphism.
To nish the proof it is enough to show (8.2). So let x X. To prove
that D
x
is dense, let g P. We have to nd h D
x
such that h g, that is,
a function h P such that h g and x dom(h). If x dom(g) then h = g
works. So assume that x dom(g). Let dom(g) = x
0
, x
1
, . . . , x
n
with
x
0
< x
1
< < x
n
and let y
i
= g(x
i
) for i n. Then y
0
y
1
y
n
,
since g is strictly increasing. We will dene h as g'x, y`, where y Y is

8.1 RasiowaSikorski lemma 135
chosen in such a way that h is strictly increasing. To do this, it is enough to
pick y Y ` y
0
, y
1
, . . . , y
n
such that the following holds for every i n:
x < x
i
y y
i
. (8.3)
If x < x
0
, pick y y
0
. Such y exists, since Y does not have a smallest
element. If x > x
n
, pick y ~ y
n
, which exists since Y does not have a
largest element. So assume that x
i
< x < x
i+1
for some i < n. Then
choose y Y such that y
i
y y
i+1
, which exists since Y is dense. It is
easy to see that such y satises (8.3).
We have proved that each D
x
is dense in P. The proof that every R
y
is dense in P is almost identical, and is left as an exercise.
Evidently ', ` is an example of a dense linearly ordered set with nei-
ther a rst nor a last element. Thus Theorem 8.1.3 says that any countable
linearly ordered dense set without a rst or a last element is isomorphic to
', `. The order type of this class is usually denoted by the letter .
In fact, it can be proved that ', ` is universal for the class of all
countable linearly ordered sets, in the sense that every countable linearly
ordered set is isomorphic to 'S, ` for some S (see Exercise 1).
Clearly, 'R, ` is not isomorphic to ', `. The order type of 'R, `
is usually denoted by . To give a characterization of the order type
similar to that of Theorem 8.1.3 we will need the following denition. A
linearly ordered set 'X, ` is said to be complete if every subset of X
that is bounded from above has a least upper bound. That is, if the set
B(S) = b X: x S (x b) is not empty for some S X then B(S)
has a least element, denoted by supS. It is a fundamental fact that 'R, `
is complete.
Theorem 8.1.4 Any two complete linearly ordered sets both having count-
able dense subsets, and having neither a least nor a largest element, are
order isomorphic.
Proof Let 'X, ` and 'Y, _` be linearly ordered sets as described in the
theorem. Let X
0
and Y
0
be countable dense subsets of X and Y , re-
spectively. It is easy to see that X
0
and Y
0
satisfy the assumptions of
Theorem 8.1.3.
Let f
0
: X
0
Y
0
be an order isomorphism between 'X
0
, ` and 'Y
0
, _`.
Dene f : X Y by f(x) = supf
0
(x
0
): x
0
X
0
& x
0
x. It is not
dicult to prove that f is an order isomorphism. The details are left as an
exercise.
The next theorem will show that we can also use the RasiowaSikorski
lemma in proofs by transnite induction. To formulate it we need the
following denition. For f, g : we dene
f <

g f(n) < g(n) for all but nitely many n < . (8.4)

136 8 Martins axiom
It is easy to see that the relation <

is transitive on

.
Theorem 8.1.5 If G

has cardinality
1
then there exists a <

-
increasing sequence 'f

: <
1
` such that for every g G there exists a
<
1
with g <

.
Before we prove it, notice rst that it immediately implies the following
corollary. For its formulation we need the following denition. A sequence
'f

: < ` is called a scale in

if it is <

-increasing and if for

every g

there exists a < such that g <

.
Corollary 8.1.6 If the continuum hypothesis holds then there exists a scale
'f

: < c`.
In the next section we will show that Corollary 8.1.6 can also be proved
when the continuum hypothesis is false.
The proof of Theorem 8.1.5 will be based on two lemmas. The rst
of these lemmas is not essential for the proof, but it serves as a good
approximation for the second one.
Lemma 8.1.7 If G

has cardinality then there exists an f

such that for every g G

g(n) < f(n) for innitely many n < .
Proof An inductive construction of such f can be done as follows. Let
G = 'g
n
, k
n
`: n < . Construct a sequence f
0
f
1
f
2

of functions such that each f
n
Func

(, ) and for every n < there

exists an m dom(f
n
) with m k
n
such that f
n
(m) > g
n
(m). Then

f =

n<
f
n
Func(, ) and any extension f of

f to has the desired
property, since for every g G and k < there exists an m k with
f(m) > g(m).
To translate this proof into the language of partially ordered sets, let P
be the set of all possible functions f
n
as before, that is, take
P = Func

(, ),
and order it by reverse inclusion . For g

and k < dene

D
k
g
= s P: m k (m dom(s) & s(m) > g(m)).
These sets are the counterparts of the inductive condition f
n
(m) > g
n
(m)
for some m k
n
.
Notice that the sets D
k
g
are dense in P, since for every t P there exists
an m ` (k dom(t)) and for s = t 'm, g(m) + 1` we have s D
k
g
and s t.

8.1 RasiowaSikorski lemma 137
Let T = D
k
g
: k < & g G and let F be a T-generic lter in
P. Then, by Proposition 8.1.1,

F is a function. Let f : be any
extension of

F. Then f has the desired property, since for every g G
and k < there are s F D
k
g
and m k such that f(m) = s(m) > g(m).

The lemma that we really need for the proof of Theorem 8.1.5 is the
following.
Lemma 8.1.8 If G

has cardinality then there exists an f

such that
g <

f for every g G.
Proof An inductive construction of such f can be done as follows. Let
G = g
n
: n < and = x
n
: n < . Construct a sequence f
0
f
1

f
2
of functions such that f
n
Func

(, ) and x
n
dom(f
n
) for
every n < . Moreover, at the inductive step n < choose f
n
such that
f
n
(x) > g
i
(x) for all functions g
i
looked at so far and all numbers x that
have not yet been considered up to this point of the induction, that is, such
that
f
n
(x) > g
i
(x) for all x dom(f
n
) ` dom(f
n1
) and i < n. (8.5)
Then f =

n<
f
n
: and g
i
<

f for every i < .

One problem in nding an appropriate forcing for this construction
is that the current stage of induction includes information not only on
functions f
n
constructed so far but also on all functions g
i
considered
so far. The rst part of this information is included in functions from
P = Func

(, ), as in Lemma 8.1.7. However, we need the second piece

of information as well, which is to be coded by nite subsets of G. Thus
we dene the partially ordered set P

= P [G]
<
. Moreover, the partial
order on P

, in order to describe the extension as in (8.5), must be dened

as follows. For 's, A`, 't, B` P

we dene 's, A` 't, B` provided s t,

A B, and
s(n) > g(n) for all n dom(s) ` dom(t) and g B. (8.6)
Now the inductive conditions x
n
dom(f
n
) and every element g of G
is taken care of at some stage of the induction are coded by the following
subsets of P

:
D
n
= 's, A` P

: n dom(s) and E
g
= 's, A` P

: g A,
where g G

and n < .

138 8 Martins axiom
The sets E
g
are dense in P

, since for every 's, A` P

we can pick
's, A g` E
g
and clearly 's, A g` 's, A`, as condition (8.6) is
then satised vacuously.
To see that a set D
n
is dense in P

take p = 't, A` P

. We have to
nd q D
n
such that q p. If n dom(t) then p D
n
and q = p works.
So assume that n dom(t). Dene q = 's, A` D
n
with s = t 'n, k`
and k = supg(n) + 1: g A. Then q p, since clearly s t, A A,
and (8.6) holds, as s(n) = k > g(n) for all g A.
Now let T = D
n
: n < E
g
: g G. Then T is a countable
family of dense subsets of P

so, by the RasiowaSikorski lemma, we can

nd a T-generic lter F in P

. Notice that F
0
= s: 's, A` F is a lter
in P. Thus, by Proposition 8.1.1, f =

F
0
is a function. Notice also that
D
n
F = implies n dom(f). Therefore f maps into itself.
To nish the proof it is enough to show that g <

f for every g G.
So let g G. Then there exists 't, B` F E
g
. We will prove that
f(n) > g(n) for every n ` dom(t).
This is the case since for every n ` dom(t) there is 't
t
, B
t
` F D
n
.
Take 's, A` F with 's, A` 't, B` and 's, A` 't
t
, B
t
`. Then n
dom(t
t
) dom(s) so n dom(s) ` dom(t) and g B. Therefore condition
(8.6) for 's, A` 't, B` implies that f(n) = s(n) > g(n).
Proof of Theorem 8.1.5 Let G = g

: <
1
. Dene 'f

: <
1
` by
induction on <
1
by choosing f

as the function f from Lemma 8.1.8

applied to G = f

: < g

: < . It is easy to see that this

sequence has the desired property.
EXERCISES
1 Prove that for every countable linearly ordered set 'X, _` there exists
a strictly increasing function f : X . (Such a function f is said to be
an order embedding of X into , and it establishes an order isomorphism
between 'X, _` and 'f[X], `.)
2 Prove that 'R, ` is complete as a linearly ordered set (use Dedekinds
denition of real numbers).
3 Complete the details of the proof of Theorem 8.1.4 by showing that
(a) the sets X
0
and Y
0
are dense in themselves and have neither least nor
largest elements;
(b) the function f : X Y is an order isomorphism.

8.2 Martins axiom 139
4 A subset D of a linearly ordered set 'X, ` is weakly dense if
for every x, y X with x < y there is a d D such that x d y.
Prove that for every linearly ordered set 'X, ` containing a countable
weakly dense subset there exists a strictly increasing function f : X R.
5 Prove that the relation dened on the set P

as in Lemma 8.1.8 is
indeed a partial-order relation.
6 For f, g : we dene
f

g f(n) g(n) for all but nitely many n < .

Show that

is a preorder relation on

(see Exercise 4 from Section 2.4).

The equivalence relation generated by

is usually denoted by =

. Show
also that the partial-order relation _

induced by

on the family of all

equivalence classes of =

is a linear-order relation.
8.2 Martins axiom
In this section we would like to introduce an axiom that says that in a large
number of situations the statement for countably many can be replaced
by for less than continuum many, even when the continuum hypothesis
fails.
In particular, we would like to be able to make such a replacement in
the RasiowaSikorski lemma, which leads us to the following statement.
() Let 'P, ` be a partially ordered set. If T is a family of dense subsets
of P such that [T[ < c, then there exists a T-generic lter F in P.
Clearly () is implied by the continuum hypothesis. However, it is false
under the negation of the continuum hypothesis. To see it, consider the
partially ordered set 'P, ` = 'Func

(,
1
), ` and, for <
1
, the sets
R

= s P: range(s). Each set R

is dense, since for every s P

we have t = s 'n, ` R

and t s for any n ` dom(s). But

T = R

: <
1
has cardinality
1
< c. So () implies the existence
of a T-generic lter F in P, and, by Proposition 8.1.1,

F is a function
from a subset of onto
1
. This is clearly impossible.
Thus, in order to nd a ()-like axiom that is consistent with the nega-
tion of the continuum hypothesis we will have to restrict the class of par-
tially ordered sets allowed in its statement. In particular, we will have to
exclude forcings such as 'Func

(,
1
), `. To dene such a class we need
some new denitions.

140 8 Martins axiom
Let 'P, ` be a partially ordered set.
x, y P are comparable if either x y or y x. Thus a chain in P is a
subset of P of pairwise-comparable elements.
x, y P are compatible (in P) if there exists a z P such that z x
and z y. In particular, condition (F1) from the denition of a lter
says that any two elements of a lter F are compatible in F.
x, y P are incompatible if they are not compatible.
A subset A of P is an antichain (in P) if every two distinct elements
of A are incompatible. An antichain is maximal if it is not a proper
subset of any other antichain. An elementary application of the Hausdor
maximal principle shows that every antichain in P is contained in some
maximal antichain.
To illustrate these notions consider 'P, ` = 'Func

(X, Y ), `. Then
s, t P are compatible if and only if s t Func

(X, Y ). Therefore
elements s, t P are incompatible if there exists an x dom(s) dom(t)
such that s(x) = t(x). For any nonempty D [X]
<
the set A = Y
D
is a
maximal antichain. On the other hand, if C P is a family of functions
with pairwise-disjoint domains, then any two elements of C are compatible.
The forcings that will be used in our ()-like axiom are dened in terms
of antichains in the following way.
A partially ordered set 'P, ` is ccc (or satises the countable chain
condition) if every antichain of P is at most countable.
1
Clearly, every countable partially ordered set is ccc. In particular, the
forcing 'P, ` = 'Func

(, ), ` from Lemma 8.1.7 is ccc. Notice also

that the forcing
P

= P [G]
<
from Lemma 8.1.8 is ccc (8.7)
for an arbitrary G

, including G =

. To see this, take an un-

countable subset A = p

: <
1
of P

with p

= 's

, A

` and
notice that there must be a < <
1
such that s

= s

. But then, for

p = 's

, A

` P

we have p p

and p p

, condition (8.6) being

satised vacuously. So A is not an antichain.
On the other hand, the forcing 'P, ` = 'Func

(,
1
), ` is not ccc,
since '0, ` P: <
1
is an uncountable antichain in P. Thus,
1
The name countable chain condition is certainly misleading. A more appropriate
name would be countable antichain condition or cac. However, the tradition of this
name is very strong and outweighs reason. This tradition can be explained by the fact
that every partially ordered set can be canonically embedded into a complete Boolean
algebra, and for such algebras the maximal sizes of chains and antichains are equal.

8.2 Martins axiom 141
restricting () to the ccc forcings removes the immediate threat of the pre-
viously described contradiction with CH, the negation of the continuum
hypothesis. In fact such a restriction also removes all possibility of any con-
tradiction with CH. More precisely, consider the following axiom, known
as Martins axiom and usually abbreviated by MA.
Martins axiom Let 'P, ` be a ccc partially ordered set. If T is a family
of dense subsets of P such that [T[ < c, then there exists a T-generic
lter F in P.
Clearly CH implies MA. But MA is also consistent with ZFC and CH,
as stated by the next theorem.
Theorem 8.2.1 Martins axiom plus the negation of the continuum hy-
pothesis MA+CH is consistent with ZFC set theory.
The proof of Theorem 8.2.1 will be postponed until Section 9.5 where
we will prove that MA is consistent with c =
2
. However, the same proof
can be used to prove the consistency of MA with c = for most regular
cardinals .
In the remainder of this section we will see several consequences of MA.
First we will see the following generalization of Corollary 8.1.6.
Theorem 8.2.2 If MA holds then there exists a scale 'f

: < c`.
Proof The argument is essentially identical to that for Theorem 8.1.5.
First notice that the following generalization of Lemma 8.1.8 is implied by
Martins axiom.
(I) If G

has cardinality < c then there exists an f

such that
g <

f for every g G.
To see why, consider the forcing P

= Func

(, ) [G]
<
dened as
in Lemma 8.1.8, that is, ordered by
's, A` 't, B` s t & A B
& s(n) > g(n) for all n dom(s) ` dom(t) and g B,
and its dense subsets
D
n
= 's, A` P

: n dom(s) and E
g
= 's, A` P

: g A
for g

and n < . Then T = D

n
: n < E
g
: g G has
cardinality [G[ + < c. Since by (8.7) P

is ccc, MA implies that there

exists a T-generic lter F in P

. Now we argue as in Lemma 8.1.8. We

dene F
0
= s: 's, A` F and notice that f =

F
0
is a function. Then

142 8 Martins axiom
D
n
F = implies that n dom(f), that is, that f maps into . Finally,
we notice that E
g
F = implies g <

f.
To prove the theorem from (I) enumerate

as g

: < c and dene

the scale 'f

: < c` by induction on < c by choosing g

as a function f
from (I) applied to G = f

: < g

: < .
Another application of Martins axiom is stated in the next theorem.
Theorem 8.2.3 Assume MA. If X [R]
<c
then every subset Y of X is a
G

subset of X, that is, there exists a G

set G R such that GX = Y .

Proof Let X [R]
<c
and x Y X. We will show that Y is G

in X.
Let B = B
n
: n < be a countable base for R. First notice that it is
enough to nd a set

A such that for every x X
x Y x B
n
for innitely many n from

A. (8.8)
To see why, dene for every k < an open set G
k
=

B
n
: n

A &n > k
and put G =

k<
G
k
. Then G is a G

set and, by (8.8), for every x X

we have
x Y x G
k
for all k < .
Thus G X = Y .
For countable X we could prove (8.8) in the following way. Let X`Y =
z
n
: n < and Y = 'y
n
, k
n
`: n < . Then construct

A as a union
of an increasing sequence 'A
n
: n < ` of nite subsets of such that for
every n < there is an m A
n
such that m > k
n
and y
n
B
m
, and such
that z
i
B
m
for all new ms from A
n
and all points z
i
considered so far,
that is, such that
z
i
B
m
for all m A
n
` A
n1
and i n. (8.9)
We will transform this idea into a forcing similarly as in Lemma 8.1.8.
We dene the partially ordered set 'P, ` by putting P = []
<
[X`Y ]
<
and dene to take care of (8.9). That is, for 'A
1
, C
1
`, 'A
0
, C
0
` P we
dene 'A
1
, C
1
` 'A
0
, C
0
` provided A
1
A
0
, C
1
C
0
, and
c B
m
for all m A
1
` A
0
and c C
0
. (8.10)
Now inductive conditions are coded by the following subsets of P:
D
k
y
= 'A, C` P: m A (m k & y B
m
)
and
E
z
= 'A, C` P: z C,

8.2 Martins axiom 143
where y Y , k < , and z X ` Y . We will use Martins axiom to nd a
T-generic lter for
T = D
k
y
: y Y & k < E
z
: z X ` Y .
To use Martins axiom, we have to check whether its assumptions are
satised.
Clearly [T[ [X[ + < c.
To see that P is ccc consider an uncountable subset 'A

, C

`: <
1

of P. Since []
<
is countable, there are A []
<
and < <
1
such
that A

= A

= A. Then 'A

, C

` = 'A, C

` and 'A

, C

` = 'A, C

` are
compatible, since 'A, C

` P extends them both, as condition (8.10)

is satised vacuously.
To see that each set E
z
is dense in P take 'A, C` P and notice that
'A, C z` E
z
extends 'A, C`.
Finally, to see that a set D
k
y
is dense in P, take 'A, C` P. Notice that
there exist innitely many basic open sets B
m
such that
y B
m
and C B
m
= . (8.11)
Take m > k satisfying (8.11), and notice that 'A m, C` D
k
y
extends
'A, C`.
Now apply Martins axiom to nd a T-generic lter F in P, and dene

A =

A: 'A, C` F. We will show that

A satises (8.8). So let x X.
If x Y then for every k < there exists 'A, C` F D
k
x
. In
particular, there exists an m A

A with m > k such that x B
m
. So
x B
m
for innitely many m from

A.
If x X` Y then there exists 'A
0
, C
0
` F E
x
. In particular, x C
0
.
It is enough to prove that x B
m
for every m

A`A
0
. So take m

A`A
0
.
By the denition of

A there exists 'A, C` F such that m A. But, by
the denition of a lter, there exists 'A
1
, C
1
` F extending 'A, C` and
'A
0
, C
0
`. Now 'A
1
, C
1
` 'A
0
, C
0
`, m A A
1
, m A
0
, and x C
0
.
Hence, by (8.10), x B
m
.
Corollary 8.2.4 If MA holds then 2

= 2

for every innite cardinal

< c.
Proof Let < c be an innite cardinal number. Then clearly 2

.
To see the other inequality take X [R]

. Then, by Theorem 8.2.3,

2

= [{(X)[ = [B X: B Bor[ [Bor[ = 2

.

144 8 Martins axiom
Corollary 8.2.5 If MA holds then c is a regular cardinal.
Proof Notice that by Theorem 5.3.8
2
cf(c)
= (2

)
cf(c)
= c
cf(c)
> c.
Hence, by Corollary 8.2.4, cf(c) c.
The next theorem tells us that MA implies the continuum additivity
of category, which has been mentioned in the remark after Theorem 7.2.5.
In particular, it implies that Theorem 7.2.5 can be proved when CH is
replaced with MA.
Theorem 8.2.6 If MA holds then a union of less than continuum many
meager subsets of R
n
is meager in R
n
, that is,

T for every T []
<c
.
Proof The idea of the proof is very similar to that of Theorem 8.2.3.
Let T []
<c
. Since every F from T is a countable union of nowhere-
dense sets,

T is a union of [T[ + < c nowhere-dense sets. Thus we may

assume that every set in T is nowhere dense.
Let B = B
j
= : j < be a countable base for R
n
. Notice that it is
enough to nd a set

A such that
m

A: B
m
F = is nite for every F T (8.12)
and
m

A: B
m
B
j
is innite for every j < . (8.13)
To see why, dene U
k
=

B
m
: m

A & m > k for every k < .
Notice that by (8.13) the sets U
k
are dense and open. Therefore the sets
R
n
` U
k
are nowhere dense. But condition (8.12) implies that

T

k<
(R
n
` U
k
) , since for every F T there exists a k < such that
U
k
F =

B
m
F : m

A & m > k = .
To prove the existence of such a set

A consider the partial order P =
[]
<
[T]
<
, where we put 'A
1
, (
1
` 'A
0
, (
0
` provided A
1
A
0
,
(
1
(
0
, and
B
m
F = for all m A
1
` A
0
and F (
0
. (8.14)
We will dene

A as the union of all A such that 'A, C` belongs to an
appropriate generic lter in P. In particular, in the condition 'A, (` P

8.2 Martins axiom 145
the set A approximates

A. The sets F from ( represent elements of T
looked at so far in our induction and (8.14) guarantees that the basic
open sets indexed by new elements of

A will not intersect these old
Fs. This will take care of (8.12).
The forcing P is ccc since []
<
is countable and any conditions 'A, (
0
`
and 'A, (
1
` are compatible, having 'A, C
0
C
1
` as a common extension.
Conditions (8.13) and (8.12) are related to the following dense subsets
of P:
D
k
j
= 'A, (` P: m A (m k & B
m
B
j
)
and
E
F
= 'A, (` P: F (,
where k, j < and F T. The sets E
F
are dense, since 'A, ( F` E
F
extends 'A, (` for every 'A, (` P. To see that the sets D
k
j
are dense take
'A, (` P. Since

( is nowhere dense, there exist innitely many basic
open sets B
m
such that
B
m
B
j
`

(. (8.15)
Take m > k satisfying (8.15) and notice that 'A m, (` D
k
j
extends
'A, (`.
Hence, by Martins axiom, there exists a T-generic lter

F in P, where
T = D
k
j
: k, j < E
F
: F T.
Dene

A =

A: 'A, (`

F. It is enough to show that

A satises
properties (8.13) and (8.12).
To see (8.13) take j < . It is enough to prove that for every k <
there exists an m

A with m > k and B
m
B
j
. So x k < and pick
'A, (`

F D
k
j
. Then, by the denition of D
k
j
, there exists an m A

A
with m > k such that B
m
B
j
.
To see (8.12) take F T. Then there exists 'A
0
, (
0
`

F E
F
. In
particular, F (
0
. It is enough to prove that B
m
F = for every
m

A ` A
0
. Take m

A ` A
0
. By the denition of

A there exists
'A, (`

F such that m A. So there exists 'A
1
, (
1
`

F extending 'A, (`
and 'A
0
, (
0
`. Now 'A
1
, (
1
` 'A
0
, (
0
`, m A A
1
, m A
0
, and F (
0
.
Hence, by (8.14), B
m
F = .
An analog of Theorem 8.2.6 for the ideal ^ of measure-zero subsets of
R
n
is also true.
Theorem 8.2.7 If MA holds then a union of less than continuum many
null subsets of R
n
is null in R
n
, that is,

T ^ for every T [^]

<c
.

146 8 Martins axiom
Proof To make the argument simpler, we will prove the theorem only for
R. The proof of the general case is essentially the same.
Let B
0
be the family of all open intervals in R and let l(I) stand for
the length of I B
0
. Recall that S ^ if for every > 0 there exists a
sequence 'I
k
B
0
: k < ` such that S

k<
I
k
and

k<
l(I
k
) < .
Now, if B = I
n
: n < is the family of all intervals with rational
endpoints, then the family B
0
can be replaced by B. To see it, take S ^
and let > 0. Then there exists a sequence 'J
k
B
0
: k < ` such
that S

k<
J
k
and

k<
l(J
k
) < /2. But for every k < there
exists an I
k
B
0
such that J
k
I
k
and l(I
k
) < l(J
k
) + /2
k+2
. So
S

k<
J
k

k<
I
k
and

k<
l(I
k
)

k<
[l(J
k
) +/2
k+2
] < .
Let T [^]
<c
and x > 0. We will nd an

A such that

A
I
n
and

A
l(I
n
) . (8.16)
Dene P = A :

nA
l(I
n
) < and order it by reverse inclusion:
A
1
A
0
A
1
A
0
. To see that P is ccc let / P be uncountable. We
have to nd dierent A, A
t
/ that are compatible, that is, such that

nAA
l(I
n
) < .
So for every A / let m
A
< be such that

nA
l(I
n
) + 1/m
A
< .
Since / is uncountable, there exists an uncountable subset /
t
of / and an
m < such that m
A
= m for every A /
t
, that is,

nA
l(I
n
) +
1
m
< for every A /
t
. (8.17)
Next, for every A /
t
choose k
A
< such that

nA\k
A
l(I
n
) < 1/m.
Since /
t
is uncountable, there exists an uncountable subset /
tt
of /
t
and
a k < such that k
A
= k for every A /
tt
, that is,

nA\k
l(I
n
) <
1
m
for every A /
tt
. (8.18)
Now we can nd two dierent A, A
t
/
tt
such that A k = A
t
k. For
such A and A
t
we have, in particular, A A
t
= A (A
t
` k). So, by (8.18)
and (8.17),

nAA

l(I
n
)

nA
l(I
n
) +

nA

\k
l(I
n
) <

nA
l(I
n
) +
1
m
< .
Thus A and A
t
are compatible and P is ccc.
Now for every F T let D
F
= A P: F

nA
I
n
. Notice that
every D
F
is dense in P. To see this, let A P and put =

nA
l(I
n
) >

8.2 Martins axiom 147
0. Since F ^ we can nd a B such that F

nB
I
n
and

nB
l(I
n
) < . But then A B D
F
since

nAB
l(I
n
)

nA
l(I
n
) +

nB
l(I
n
) <

nA
l(I
n
) + =
and F

nAB
I
n
. Since A B A, the set D
F
is dense.
Let T = D
F
: F T. Since P is ccc and [T[ [T[ < c, by Martins
axiom there exists a T-generic lter

F in P. Let

A =


F. We will show
that

A satises (8.16).
For every F T there exists an A

F D
F
. So F

nA
I
n

A
I
n
since A

A. Then

T

A
I
n
.
To see

n

A
l(I
n
) it is enough to show that

n

Ak
l(I
n
) < for
every k < . So x k < . Now for every n

Ak there exists an A
n


F
such that n A
n
. Using condition (F1
t
) of the denition of a lter we can
nd an A

F such that A A
n
for every n

A k. But then

A k A
and

n

Ak
l(I
n
)

nA
l(I
n
) < .
This nishes the proof.
All previous applications of Martins axiom could be deduced as well
from the continuum hypothesis, but in the remaining part of this section
we will show that this will not always be the case, by discussing the con-
sequences of MA+CH, which do not follow from CH. Similar results can
also be found in the next section.
In the next theorem we will use the following terminology. A subset A
of a partially ordered set 'P, ` is compatible if for every nite subset A
0
of A there exists a p P such that p q for all q A
0
(compare this with
condition (F1
t
) from the denition of a lter).
Theorem 8.2.8 Assume MA+CH and let 'P, ` be a ccc partially or-
dered set. If A P is uncountable, then there exists an uncountable com-
patible subset

A of A.
Proof Let A P be uncountable. Without loss of generality we may
assume that [A[ =
1
.
The inductive approach to the proof is to construct a strictly increasing
sequence 'A

: <
1
` of compatible subsets of A such that

<
A

is
an uncountable compatible subset of A. Although this idea is basically
correct, it has a fundamental aw. If you start with A
0
[A]

such that
A
0
is compatible with at most countably many elements of A, then this
construction must fail. Fortunately, there are only countably many sets

148 8 Martins axiom
A
0
[A]

that may cause such a problem, and after removing them from
A we will be able to follow the idea just described.
So for p P let A
p
= q A: q is compatible with p. We will show
that the set
A
t
= q A: p q ([A
p
[ ) is at most countable. (8.19)
Indeed, consider the family T = C B: C is an antichain in P,
where B = p P: [A
p
[ . It is easy to see that T satises the
assumptions of the Hausdor maximal principle. Thus we can choose a
maximal element C
0
of T. Now C
0
is clearly an antichain in P. So C
0
is
at most countable, since P is ccc. It is enough to prove that
A
t

rC0
A
r
,
since the set

rC0
A
r
is countable, being a countable union of countable
sets. To see the inclusion, take q A
t
. Then there exists a p q such
that p B. By the maximality of C
0
there exists an r C
0
such that r
is compatible with p, that is, s r and s p q for some s P. In
particular, r and q are compatible, so q A
r
. Condition (8.19) has been
proved.
Now the set A

= A ` A
t
is uncountable. Moreover, if p q for some
q A

then A
p
must be uncountable. In particular,
A

p
= q A

: q is compatible with p is uncountable (8.20)

for every p P such that p q for some q A

.
Coming back to the idea of an inductive proof, we can see that we
could have a hard time extending an innite set A

to a compatible set
A
+1
properly containing A

. This problem will be solved with the help

of MA.
For this, consider a partially ordered set
P

= F [A

]
<
: F is compatible in P
ordered by reverse inclusion . Notice that the forcing P

is ccc. To see
why, let F

: <
1
. We will nd < <
1
such that F

and F

are compatible, that is, that F

. But for every <

1
there exists
a p

P with p

p for every p F

. Moreover, the set p

: <
1

cannot be an antichain, since P is ccc. Thus there are compatible p

and
p

for some < <

1
. Now, if q P is such that q p

and q p

then
q p for every p F

. Therefore F

and P

is ccc.
Let 'q

: <
1
` be a one-to-one enumeration of A

. For <
1
dene
D

= F P

: > (q

F)

8.2 Martins axiom 149
and notice that the sets D

are dense in P

. This is so since for every

F P

there exists a p P such that p q for every q F, and, by

(8.20), p is compatible with uncountably many q

. In particular, there
exists a > such that q

is compatible with p, and it is easy to see that

F q

extends F.
Let T = D

: <
1
. Then [T[
1
< c so, by MA, there exists
a T-generic lter

F in P

. We will show that

A =

F is an uncountable
compatible subset of A.
Clearly,

A A. To see that

A is uncountable, notice that for every
<
1
there exists a > such that q

A, since

F D

= . To nish
the proof, it is enough to show that

A is compatible.
So let F = r
0
, . . . , r
n


A. For every i n there exists an F
i


F
such that r
i
F
i
. By (F1
t
) we can nd an E

F such that E F
i
for all
i n. In particular, E r
0
, . . . , r
n
= F, that is, F is compatible.
By denition, a forcing P is ccc if every uncountable set p

: A
contains two dierent compatible elements. Equivalently, P is ccc if for
every sequence 'p

: A` (not necessarily one-to-one) with uncountable

index set A there are dierent indices , A such that p

and p

are
compatible. In this language Theorem 8.2.8 can be restated as follows.
Corollary 8.2.9 Assume MA+CH and let 'P, ` be a ccc partially or-
dered set. If 'p

: A` is an uncountable sequence of elements of P

then there exists an uncountable subset

A of A such that p

:

A is
compatible in P.
To state the next corollary, we need the following important denition.
Let 'P
0
,
0
` and 'P
1
,
1
` be partially ordered sets. Their product 'P, ` is
dened by P = P
0
P
1
and
'p
0
, p
1
` 'q
0
, q
1
` p
0

0
q
0
& p
1

1
q
1
.
Corollary 8.2.10 If MA+CH holds then the product of two ccc forcings
is ccc.
Proof Let 'P, ` be a product of ccc forcings 'P
0
,
0
` and 'P
1
,
1
` and
let ''p

, q

`: A` be an uncountable sequence of elements of P. By

Corollary 8.2.9 used for the forcing P
0
and a sequence 'p

: A`, we
can nd an uncountable subset

A of A such that the set p

:

A is
compatible in P
0
. Then there are dierent ,

A such that q

and q

are compatible in P
1
, since P
1
is ccc. It is easy to see that 'p

, q

` and
'p

, q

` are compatible in P.
The last theorem of this section shows that Corollary 8.2.10 is false
under CH. In its proof we will use the following lemma, which is the main

150 8 Martins axiom
combinatorial tool needed to prove that dierent kinds of forcing (built
with nite sets) are ccc.
Lemma 8.2.11 (-system lemma) If / is an uncountable family of nite
sets then there exists an uncountable subfamily /
0
of / and a nite set A
such that X Y = A for every distinct X, Y /
0
.
Proof Since [/[ > we may assume that there is an n < such that
each element of / has exactly n elements. The proof is by induction on n.
By our assumption we must have n > 0. (Otherwise / has
cardinality 1 < .) If n = 1 then elements of / must be pairwise
disjoint and the theorem holds with A = and /
0
= /. So assume that
n > 1 and that the theorem holds for (n 1)-element sets. Consider two
cases.
Case 1: There exists an A
0
/ such that A
0
intersects uncountably
many A /. Then there exists an a A
0
such that a belongs to uncount-
ably many A /. In particular,
B = A` a: A / & a A
is an uncountable family of sets of size n1, and by the inductive hypothe-
sis, we can nd an uncountable B
0
B and a nite set B such that CD =
B for every distinct C, D B
0
. But then /
0
= C a: C B
0
/ is
uncountable and X Y = B a for every distinct X, Y /
0
.
Case 2: For every A / the set
S
A
= B /: A B =
is at most countable. Construct, by transnite induction, a one-to-one
sequence 'A

: <
1
` of pairwise-disjoint subsets of /. This can be done
since for every <
1
the set
B /: B A

= for some < =

<
S
A

is at most countable. Then /

0
= A

: <
1
and A = satisfy the
desired requirements.
In the next theorem, due to Galvin (1980), for sets A and B we will use
the notation
AB = a, b: a A & b B.
Theorem 8.2.12 If CH holds then there are two ccc forcings such that
their product is not ccc.

8.2 Martins axiom 151
Proof The forcings will be constructed as follows. We will construct a
coloring function f : [
1
]
2
2. Then, for i < 2, we put K
i
= f
1
(i) and
dene
P
i
= F [
1
]
<
: [F]
2
K
i

ordered by reverse inclusion .

First we notice that the forcing P
0
P
1
is not ccc, since the family
', ` P
0
P
1
: <
1
forms an antichain. This is the case because
compatibility of ', ` and ', ` implies ', , , ` P
0
P
1
,
and this is impossible since P
0
P
1
= [
1
]
1
is disjoint from [
1
]
2
.
Thus it is enough to construct an f such that the forcings P
i
are ccc.
We will rst translate the ccc property of P
i
into a condition that will be
appropriate for the inductive construction of f.
So x i < 2 and let 'F

P
i
: <
1
`. To prove that P
i
is ccc we will
have to nd < <
1
such that F

and F

are compatible, that is, that

[F

]
2
K
i
.
First notice that by the -system lemma we can assume, choosing a sub-
sequence, if necessary, that for some F [
1
]
<
F

= F for all < <

1
.
But [F

]
2
= [F

]
2
[F

]
2
(F

` F) (F

` F) and [F

]
2
[F

]
2
K
i
.
Thus we must nd < <
1
such that
(F

` F) (F

` F) K
i
.
Replacing F

with F

` F, we notice that it is enough to prove that for

every sequence 'F

P
i
: <
1
` of pairwise-disjoint nonempty sets there
are < <
1
such that
F

K
i
.
Moreover, for every <
1
such that

n<
F
n
there exists a <
1
with F = F

1
` . Thus we can reduce our task by showing that for
every sequence 'F
n
P
i
: n < ` of pairwise-disjoint nonempty sets there
exists an <
1
with

n<
F
n
such that for every F [
1
` ]
<
n < (F
n
F K
i
),
that is, that
n < F
n
F (f(, ) = i). (8.21)
Now, by induction on <
1
, we will construct an increasing sequence of
partial functions f : []
2
2 such that the entire function f : [
1
]
2
2 will

152 8 Martins axiom
satisfy condition (8.21).
2
For this, let ''F

n
`
n<
: <
1
` be an enumeration
of all sequences 'F
n
`
n<
of pairwise-disjoint nonempty nite subsets of
1
such that each sequence appears in the list
1
times. Such an enumeration
can be chosen by CH, since the family of all such sequences has cardinality

([
1
]
<
)

= c. The construction will be done while maintaining the

following inductive condition for every <
1
:
(I

) For every i < 2, < , and F [ ` ]

<
, if

n<
F

n
then the
set
c(i, , F) = F

n
: F

n
F (f(, ) = i) (8.22)
is innite.
Notice that this will nish the proof, since then f will satisfy (8.21)
for every i < 2 and every sequence 'F
n
`
n<
of pairwise-disjoint nonempty
nite subsets of
1
. To see this, choose <
1
such that

n<
F
n

and 'F

n
`
n<
= 'F
n
`
n<
, and for every F [
1
` ]
<
nd a <
1
such that F [ ` ]
<
. Then, by (I

), there exists an n < such that

F

n
c(i, , F). This n satises (8.21).
To make an inductive step, let <
1
be such that the construction is
already made for all < .
If is a limit ordinal, then f : []
2
2 is already constructed and it is
easy to see that f satises (I

). So assume that = + 1. We have to

extend f to , : < while maintaining (I

). So let
T =

'c(i, , F), i`: i < 2, , F [ ` ]

<
,

n<
F

.
Clearly T is countable, being indexed by a countable set. Note also that
if 'c(i, , F), i` T then c(i, , F) is innite. For < this follows
directly from the inductive assumption (I

). But if = then F = and

c(i, , F) = c(i, , ) = F

n
: n < is innite as well.
Let ''c
m
, i
m
`: m < ` be an enumeration of T with each pair appearing
innitely many times. Since each c
m
is an innite family of pairwise-
disjoint nite sets, we can construct by induction on m < a sequence
'E
m
: n < ` of pairwise-disjoint sets such that E
m
c
m
for every m <
. Dene f(, ) = i
m
for every E
m
and m < , and extend it
arbitrarily to []
2
. It is enough to show that f satises condition (I

).
So let i < 2, < = +1, and

F [`]
<
be such that

n<
F

n
.
If

F = then c(i, ,

F) = F

n
: n < is innite, and (8.22) holds. So
2
Formally we are dening an increasing sequence of functions f

: []
2
2, aiming for
f to be their union. But such an additional index would only obscure a clear idea.

8.2 Martins axiom 153
assume that

F = . Now, if

F then (8.22) holds for

F by (I

), since
=

F ` implies < . So assume that

F and let F =

F ` .
Then 'c(i, , F), i` T and 'c(i, , F), i` = 'c
m
, i
m
` for innitely many
m < . In particular, for every such m we have
f(, ) = i
m
= i (8.23)
for every E
m
. We claim that every such E
m
belongs to c(i, ,

F). This
will nish the proof, since all sets E
m
are dierent. But for every E
m
f(, ) = i
holds for F =

F ` , since E
m
c
m
= c(i, , F), and for = by
(8.23). This nishes the proof.
EXERCISES
1 Let X and Y be nonempty sets. Show that the forcing 'Func

(X, Y ), `
is ccc if and only if [Y [ . Hint: Use the -system lemma.
2 A subset Z of R has strong measure zero if for every sequence '
n
: n < `
of positive numbers there exists a sequence 'J
n
: n < ` of open intervals
such that each J
n
has length less than
n
and Z

n<
J
n
. Assuming MA
show that every Z [R]
<c
has strong measure zero. Hint: Let B be the
family of all open intervals with rational endpoints. For every '
n
: n < `
of positive numbers use forcing
P = J B
n
: n < & for every k < n the length of J(k) is less than
k

ordered by reverse inclusion .

3 (Due to Solovay) Innite sets A and B are said to be almost disjoint
provided A B is nite. Show that MA implies the following fact: If /
is a family of almost-disjoint subsets of , [/[ < c, and ( /, then there
exists a set S such that S C is nite for every C ( and S A is
innite for A / ` (. Hint: Use a forcing P = []
<
[(]
<
ordered by
's, c` 's
t
, c
t
` if and only if s s
t
, c c
t
, and (s ` s
t
)

c
t
= .
4 We say that a set A is almost contained in B and write A

B if A` B
is nite. Let B []

be such that

B
0
is innite for every nite B
0
B.
If MA holds and [B[ < c show that there exists an A []

such that
A

B for every B B.

154 8 Martins axiom
5 Assume MA and let B be a family of almost-disjoint subsets of such
that [B[ < c. Show that for every / [B]

there exists a d such that

a

d for every a / and d b is nite for every b B ` /. Hint: Let

/

= a`n: n < & a / and dene P = [/

]
<
[B`/]
<
, ordered by
'A, B` 'A
t
, B
t
` if and only if A A
t
, B B
t
, and

(A` A
t
)

B
t
= .
Use the forcing P to dene the set d as the union of all sets

A with 'A, B`
from an appropriate generic lter in P.
6 (Challenging) Assume MA and let < c be an uncountable regular
cardinal. Show that for every family / of countable subsets of such that
[/[ < c there exists a B []

such that A B is nite for every A /.

Hint: Use a forcing similar to that from Exercise 3. Use the -system
lemma to prove that it is ccc.
7 (Challenging) Generalize Exercise 6 as follows. Assume MA and let
< c be an uncountable regular cardinal. Show that for every family /
of countable subsets of such that [/[ < c there exists a countable cover
P
n
: n < of such that A P
n
is nite for every A / and n < .
Hint: If 'P, _` is the partially ordered set used in Exercise 6, use the forcing
P

= Func

(, P) ordered by s t if dom(s) dom(t) and s(n) _ t(n) for

every n dom(t).
8.3 Suslin hypothesis and diamond principle
Let 'X, ` be a linear order. For a, b X, an open interval with endpoints
a and b is dened in a natural way as (a, b) = x X: a < x < b. A
linearly ordered set X is ccc provided every family of pairwise-disjoint open
intervals in X is at most countable or, equivalently, when the partial order
' ` , ` is ccc, where is an order topology on X, that is, the topology
generated by the family of all open intervals in X.
In this terminology, Theorem 6.2.1(i) says that 'R, ` is ccc.
Now let us turn our attention to Theorem 8.1.4. It says that a linearly
ordered set 'X, ` is isomorphic to 'R, ` if and only if it has the following
properties.
(a) X has neither a rst nor a last element, and is dense in itself.
(b) X is complete.
(c) X contains a countable dense subset.
In 1920 Suslin asked whether, in the preceding characterization, condi-
tion (c) can be replaced by the weaker condition
(c
t
) X is ccc.

8.3 Suslin hypothesis and diamond principle 155
The Suslin hypothesis SH is the statement that such a replacement in-
deed can be made, that is, that any ccc linearly ordered set 'X, ` satisfying
(a) and (b) contains a countable dense subset.
In what follows we will use the term Suslin line for a ccc linearly ordered
set 'X, ` satisfying (a) that does not contain a countable dense subset.
Thus a complete Suslin line is a counterexample for the Suslin hypothesis.
In Exercise 1 we sketch the proof that the existence of a Suslin line im-
plies the existence of a complete Suslin line. Thus the Suslin hypothesis is
equivalent to the nonexistence of a Suslin line.
3
In what follows we will show that the Suslin hypothesis is independent
of the ZFC axioms. We start with the following theorem.
Theorem 8.3.1 MA+CH implies the Suslin hypothesis.
Proof Assume MA+CH and, to obtain a contradiction, that there ex-
ists a Suslin line 'X, `. Then the partially ordered set 'P, ` = '., ` is
ccc, where . is the family of all nonempty open intervals in 'X, `. We
will show that the product forcing P P is not ccc, contradicting Corol-
lary 8.2.10.
For this, we construct by induction a sequence ''a

, b

, c

` X
3
: <
1
`
such that for every <
1
(i) a

< b

< c

, and
(ii) (a

, c

) b

: < = .
Such a construction can be easily made, since for every <
1
the countable
set b

: < cannot be dense, so there must exist a

< c

in X such
that (a

, c

) b

: < = . Thus such an element b

can be chosen,
since X is dense in itself.
Now let U

= '(a

, b

), (b

, c

)` P P. It is enough to show that U

and U

are incompatible in P P for every < <

1
. This is the case
since by (ii) either b

, in which case (a

, b

) (a

, b

) = , or b

,
in which case (b

, c

) (b

, c

) = .
The remainder of this section is devoted to the proof that the existence
of a Suslin line is consistent with ZFC.
Notice that in the proof of Theorem 8.3.1 we used MA+CH only to
conclude that the product of ccc forcings is ccc, from which we deduced
3
In the literature it is probably more common to nd the term Suslin line used for a
ccc linearly ordered set that, considered as a topological space with the order topology,
does not have a countable dense subset. However, every such ordering can be extended
to a Suslin line in the sense dened here (see Exercise 2). Thus, independently of which
meaning of Suslin line is used, the Suslin hypothesis is equivalent to the nonexistence of
a Suslin line.

156 8 Martins axiom
that there is no Suslin line. Moreover, from Theorem 8.2.12 we know that
CH implies the existence of ccc forcings whose product is not ccc. Can we
generalize this argument to construct a Suslin line 'X, ` under CH?
In what follows we will describe some diculties that await any at-
tempted inductive construction of a Suslin line. Then we will show the
way to overcome these problems. In particular, the next lemma will be
used to achieve both of these goals. It will illustrate the aforementioned
diculties, and will be used in our construction of a Suslin line. To state
the lemma, we need the following notation, which will be used for the
remainder of this section.
For
1
the symbol

will denote the th limit ordinal number

in
1
+ 1. Thus

:
1
is a one-to-one increasing enumeration of
the set of all limit ordinals in
1
+ 1. In particular,
0
= 0,
1
= ,

2
= +, . . . , and

1
=
1
and

<

for every limit ordinal

1
.
Also, for <
1
let

+ n: n < and let

be a linear-order
relation on

giving it the order type of ', `. Notice that

+1
=

for every <

1
and that

: < for every

1
.
Lemma 8.3.2 Let
1
. If 'S

: < ` and '_

: `
are such that for every <
(1) _

is a linear-order relation on

;
(2) S

is a proper initial segment of '

, _

`;
(3) _

<
(_

) if is a limit ordinal;
(4) if = +1 then the relation _

on

is dened from _

by sticking the entire '

` between S

and

` S

; that is, _

extends _

and is dened for every 'x, q`

by: x _

q
if x S

, and q _

x if x

` S

;
then for every
(I) '

, _

` is a linearly ordered set that is dense in itself and has neither

a largest nor a smallest element.
Moreover, if =
1
then '
1
, _
1
` does not have a countable dense subset.

8.3 Suslin hypothesis and diamond principle 157
Proof Condition (I) is easily proved by induction on
1
. Its proof is
left as an exercise.
To see that '
1
, _
1
` does not have a countable dense subset take a
countable set D
1
. Then there is a <
1
such that D

. Choose
a, b

such that a

b. Then, by (4), there is no d D such that

a
1
d
1
b. So D is not dense.
The linear order '
1
, _
1
` constructed as in Lemma 8.3.2 has already
all the properties of a Suslin line, except possibly that of being ccc. Thus
the question is, Can we make it ccc?
Clearly, we have the freedom of choice of initial segments S

. Such a
choice can kill a potential uncountable antichain '(a

, b

): <
1
` by
making sure that, starting from some <
1
, all

for will be
placed inside some (a

, b

) for < . This would imply that there are only

countably many points of
1
outside

<
(a

, b

), making it impossible
for '(a

, b

): <
1
` to be an antichain. There is, however, a problem
in carrying out such a construction using only CH, since we have an in-
duction of only length
1
, and we have to kill 2
1
potential antichains
'(a

, b

): <
1
`. Thus once more we face the problem of having a too
short induction. Moreover, Theorem 8.3.1 shows that Martins axiom
cant help us this time. Someone may still have a hope that, through some
trick, we can reduce the number of steps necessary to rescue this construc-
tion. Indeed, this is what we will do. However, such a trick cant be
found with the help of CH alone, as it is known that the existence of a
Suslin line cannot be concluded just from the continuum hypothesis. Thus
we will need another magic axiom, which will show us the way out of our
dilemma. To formulate it, we need some new denitions.
A subset C of
1
is closed if for every S C its union

S is in C
1

or, equivalently, when every limit ordinal <

1
is in C provided C is
unbounded in . (In fact, being closed in
1
is equivalent to being closed in
the order topology of
1
.) For example, a nonzero ordinal number <
1
is closed in
1
(as a subset) if and only if it is a successor ordinal.
In what follows we will be primarily interested in closed unbounded
subsets of
1
(see Section 5.3). Notice that these sets are quite big, as
follows from the next proposition.
Proposition 8.3.3 If C and D are closed, unbounded subsets of
1
then
C D is also closed and unbounded.
Proof Clearly, C D is closed, since for every S C D its union

S
is both in C
1
and in D
1
.
To see that C D is unbounded in
1
, pick <
1
and dene by
induction the sequences '
n
C: n < ` and '
n
D: n < ` such that
<
n
<
n
<
n+1
for every n < . It can be done since sets C and

158 8 Martins axiom
D are unbounded in
1
. Then =

n<

n
=

n<

n
<
1
belongs to
C D and is greater than .
One of the most important facts we will use about the closed unbounded
sets is the following proposition.
Proposition 8.3.4 Let T be a countable family of functions of the form
f :
n
1

1
or f :
n
1
[
1
]

with n < . If
D = <
1
: is closed under the action of T
then D is closed and unbounded.
Proof It is easy to see that D is closed (compare the proof of Lemma 6.1.6).
To see that it is unbounded pick <
1
and let i :
1
[
1
]

be given
by i() = = : < [
1
]

. Then, by Lemma 6.1.6(a) used with

Z = , there is a countable subset Y of
1
that is closed under the action
of T i. In particular, closure under the action of i implies that Y is
an ordinal number. So Y is a countable ordinal number containing and
belonging to D.
A subset S of
1
is stationary if S C = for every closed unbounded
set C
1
. The theory of stationary sets is one of the most beautiful and
useful parts of set theory. However, due to lack of space, we will not be
able to develop it here in full. Some traces of its power can be found in the
rest of this section and in the exercises.
The additional axiom we will need is called the diamond principle and
is usually denoted by the symbol . It reads as follows.
Diamond principle There exists a sequence 'A

1
: <
1
`,
known as a -sequence, such that for every A
1
the set
<
1
: A = A

is stationary.
Notice that
Proposition 8.3.5 implies CH.
Proof Let A . Since C =
1
` is closed and unbounded, there exists
an C such that A

= A = A = A. Thus A

: <
1
=
{().
In the next chapter we will show that is consistent with the ZFC
axioms. It is also true that it does not follow from CH. Thus we have the
following implications, none of which can be reversed:
CH MA.

8.3 Suslin hypothesis and diamond principle 159
As we have just seen a -sequence lists all subsets of . But it also
captures a lot of information regarding uncountable subsets of
1
. This will
be enough to overcome the obstacles to the construction of a Suslin line
described earlier. For the construction we need the following easy lemma.
Lemma 8.3.6 If holds then there exists a sequence 'B

: <
1
`
such that for every B
1

1
the set
<
1
: B ( ) = B

is stationary.
Proof Let f
0
, f
1
:
1

1
be such that f :
1

2
1
given by f() =
'f
0
(), f
1
()` is a bijection and let B

= f[A

] ( ). Notice that,
by Proposition 8.3.4 used with the family T = f
0
, f
1
, f
1
, the set D =
: f[] = is closed and unbounded.
To see that the sequence 'B

: <
1
` has the desired properties, take
B
1

1
and let C
1
be closed and unbounded. We must nd an
C such that B

= B ( ).
Since C D is closed and unbounded and <
1
: f
1
(B) = A

is stationary, there is an CD such that A

= f
1
(B). But D
implies that
A

= f
1
(B) = f
1
(B) f
1
( ) = f
1
(B ( )).
Hence B

= f[A

] ( ) = B ( ).
Theorem 8.3.7 If holds then there exists a Suslin line 'X, _`. In par-
ticular, the Suslin hypothesis fails.
Proof We will construct 'X, _` = '
1
, _
1
` as in Lemma 8.3.2, following
the idea described after its proof. The sequence 'B

: <
1
`
from Lemma 8.3.6 will be used as an oracle that will tell us how to choose
initial segments S

so as to kill all potential uncountable antichains.

Thus we will construct inductively the sequences 'S

: <
1
`
and '_

:
1
` such that for every <
1
(1) _

is a linear-order relation on

;
(2) S

is a proper initial segment of '

, _

`;
(3) _

<
(_

) if is a limit ordinal;
(4) if = +1 then _

extends _

and for every 'x, q`

is dened by: x _

q if x S

, and q _

x if x

` S

.

160 8 Martins axiom
Note that by Lemma 8.3.2 we know that such a construction can be
made and that 'X, _` obtained this way is linearly ordered, is dense in
itself, has neither a largest nor a smallest element, and does not have a
countable dense subset. So it is enough to show that we can choose sets
S

such that 'X, _` will be ccc.

So assume that for some <
1
the sequences 'S

: < ` and
'_

: ` are already constructed. We must construct the

proper initial segment S

of

. It will be done as follows.

For a, b

let (a, b)

be an open interval in '

, _

` with the end-

points a and b, that is,
(a, b)

= x

: a

b,
and for let
(

= (a, b)

: 'a, b` B

.
Notice that B

, so (

is well dened. Dene the family

T

as
(

: and (

is a maximal family of disjoint intervals in '

, _

`
and let c
n
: n < = T

(, where

( is an arbitrary maximal family
of pairwise-disjoint intervals in '

, _`. (We add

( to this list just to avoid
a problem when T

= .)
The idea of the proof is that the elements of

<1
T

will be approxi-
mations of all the potential families of pairwise-disjoint intervals in 'X, _`.
We will make sure that every family ( T

remains maximal after step

by sticking every

for <
1
inside some interval I (, thus
adding no new elements of X outside

(.
So dene, by induction on n < , a sequence ''a
n
, b
n
`

: n < `
such that for every n <
(A) a

a
n

b
n

b for some (a, b)

c
n
; and
(B) a
n1

a
n

b
n

b
n1
for n > 0.
Such a construction can be made since, by Lemma 8.3.2, '

, _

` is dense
in itself and, by the maximality of c
n
, the interval (a
n1
, b
n1
)

must
intersect some (a, b)

c
n
. We dene S

= x

: n < (x _

a
n
).
This nishes the construction.
Now notice that if (

for some <

1
, then there are
n < and a, b

such that a

a
n

b
n

b, (a, b)

, and
a
n
S

but b
n
S

. In particular,

(a
n
, b
n
)
+1

(
+1

and (
+1

remains a maximal family of pairwise-disjoint intervals in '

+1
, _
+1
`. So
(
+1

T
+1
.

8.3 Suslin hypothesis and diamond principle 161
Using this, we can easily prove by induction on <
1
that for every
<
1
if (

and < <

1
then

<

and (

.
This, in particular, implies that

<1

(
1

provided (

,
that is, that
(
1

is a maximal family of disjoint intervals in 'X, _`, if (

. (8.24)
To see that 'X, _` is ccc assume, to obtain a contradiction, that there is
an uncountable family /
0
of nonempty pairwise-disjoint intervals in 'X, _`.
Using Zorns lemma we can extend /
0
to a maximal family / of nonempty
pairwise-disjoint intervals in 'X, _`. Put B = 'a, b`: (a, b)
1
/. The
maximality of / means that for every a, b X, with a b, there exists
a nonempty interval (l, r)
1
= (l(a, b), r(a, b))
1
/ intersecting (a, b)
1
.
Let l, r :
2
1

1
be functions having the foregoing property and let
:
1

1
be given by () =

+ 1. Put
D = <
1
: is closed under the action of , l, r.
By Proposition 8.3.4, D is closed and unbounded in
1
. Moreover, for
every D,

= , (8.25)
and
T

= (l, r)

: 'l, r` B ( ) (8.26)
is a maximal family of pairwise-disjoint intervals in '

, _

` = ', _

`.
Condition (8.25) follows from the fact that is closed under the action
of . To see why, notice rst that is a limit ordinal, since +1

+1 =
() < for every < . Now

<

.
Condition (8.26) follows from the closure of under the actions of l and
r, since for every a, b =

, with a

b, there are l = l(a, b) < and

r = r(a, b) < with (l, r)
1
/ intersecting (a, b)
1
. This clearly implies
that (l, r)

intersects (a, b)

, making T

maximal in '

, _

` =
', _

`.
By the choice of the sequence 'B

: <
1
` there exists an
D such that B ( ) = B

. Hence, by (8.25) and (8.26),

(

= (l, r)

: 'l, r` B

= (l, r)

: 'l, r` B ( ) = T

is a maximal family of pairwise-disjoint intervals in ', _

` = '

, _

`.
Hence, by the denition of T

, (

. So, by (8.24), (
1

is a maximal
antichain in 'X, _`, contradicting the fact that (
1

is a proper subset of
an uncountable antichain /. Thus we have proved that 'X, _` is ccc.

162 8 Martins axiom
EXERCISES
1 Let 'X, ` be a Suslin line and let X

be the family of all proper

nonempty initial segments of X without last elements. Prove that 'X

, `
is a complete Suslin line.
2 Let 'X, ` be a ccc linearly ordered set and assume that X, considered
as a topological space with the order topology, does not have a countable
dense subset. Show that there exists an X
0
X such that 'X
0
, ` is a
Suslin line and for every a, b X
0
, with a < b, there is no countable dense
subset of (a, b) X
0
. Hint: Dene an equivalence relation on X by
putting x y if and only if the interval between them contains a countable
subset that is dense in it with respect to the order topology. Choose X
0
to
be any selector from the family of all equivalence classes of , from which
the least and the greatest elements are removed, if they exist.
3 Complete the proof of Lemma 8.3.2 by proving condition (I).
4 Prove that if sets C
n

1
are closed and unbounded then

n<
C
n
is
closed and unbounded.
5 Let C

1
: <
1
be a family of closed unbounded sets. Prove
that the set
D = <
1
: C

for all < ,

known as the diagonal intersection of the sets C

, is closed and unbounded.

6 Prove the following theorem, known as the pressing-down lemma or
Fodors theorem:
Let S
1
be a stationary set and let f : S
1
be such that
f() < for every S (such a function is called a regressive
function). Then there exists an <
1
such that f
1
() is
stationary.
Hint: Otherwise for every there exists a closed and unbounded set C

disjoint from f
1
(). Consider the diagonal intersection D of the sets
C

and show that it is disjoint from S.

8.3 Suslin hypothesis and diamond principle 163
7 A tree is a partially ordered set 'T, ` with a smallest element such that
the relation < is well founded, that is, that every initial segment of T is
well ordered. A Suslin tree is an uncountable tree that, considered as a
forcing with the reversed order 'T, `, has neither uncountable chains nor
uncountable antichains. Show the existence of a Suslin line implies the
existence of a Suslin tree. Hint: Start with a Suslin line 'X
0
, ` as in
Exercise 2. Dene, by induction on <
1
, a sequence 'T

: <
1
` such
that T
0
= X
0
and each T

for > 0 is a maximal family of pairwise-disjoint

intervals in X
0
, none of which contain any endpoint of any interval from

<
T

. Then T =

<1
T

ordered by reverse inclusion is a Suslin

tree.
8 Show that the existence of a Suslin tree 'T, ` implies the existence of a
Suslin line. (Thus the existence of a Suslin line is equivalent to the existence
of a Suslin tree.) Hint: Let _ be any linear order on T and let X be the
family of all strictly increasing functions f from an ordinal number into
'T, ` such that f[] is a maximal chain in T (such an f is called a branch
of T). For dierent f, g X dene f < g if f() g(), where is the
smallest ordinal for which f() = g(). Show that 'X, ` is a ccc linearly
ordered set such that X, considered as a topological space with the order
topology, does not have a countable dense subset. Then use Exercise 2.
9 For a tree 'T, ` and an ordinal number , the -level of T is the set
of all t T for which the initial segment generated by t has order type .
(Thus any level of T is an antichain.) An Aronszajn tree is an uncountable
tree with no uncountable chains and all levels countable. (Thus every Suslin
tree is an Aronszajn tree.) Show, in ZFC, that there exists an Aronszajn
tree. Hint: For f, g : write f =

g if the set < : f() = g() is

nite. Dene, by induction on <
1
, a sequence of one-to-one functions
's

: <
1
` such that ` range(s

) is innite and s

for
every < . Dene T = s

: <
1
& s =

and relation as
reverse inclusion .

Chapter 9
Forcing
In this chapter we will describe a technique for proving that some set-
theoretic statements are independent of the ZFC axioms. This technique
is known as the forcing method. We will not prove here all the theorems
needed to justify this method. (Sketches of some of the missing proofs are
included in Appendix B. The complete proofs can be found, for example,
in Kunen (1980).) Instead, we will describe only the essentials for its use
and concentrate on its applications.
1
9.1 Elements of logic and other forcing pre-
liminaries
We will start here with some denitions, which will serve as technical tools
to develop the forcing method.
A set M is said to be transitive if x M for every x M, that is, if
a x and x M imply that a M.
Lemma 9.1.1 For every set x there exists a smallest transitive set trcl(x)
such that x trcl(x).
Proof Dene, by induction on n < ,
U
0
(x) = x and U
n+1
(x) =

U
n
(x) for n < .
1
The material included in Appendix B is not essential for the applicability of the forcing
method and can be completely skipped. However, those interested in reading it should
consider waiting at least until the end of this section, since the material included here
should make Appendix B easier to follow.
164

9.1 Elements of logic and other forcing preliminaries 165
Then trcl(x) =

n<
U
n
(x) is the desired set. It is transitive, since for
every y trcl(x) there is an n < such that y U
n
(x). Then y

U
n
(x) = U
n+1
(x) trcl(x).
To prove the minimality of trcl(x) it is enough to take a transitive set
M with x M and show, by induction on n < , that U
n
M for every
n < . The details are left as an exercise.
The set trcl(x) from Lemma 9.1.1 is called the transitive closure of x.
For an ordinal number dene inductively a sequence 'R(): `
by putting
(a) R(0) = ,
(b) R( + 1) = {(R()) for < , and
(c) R() =

<
R() for every limit ordinal .
Note that the sequence 'R(): ` is increasing, that is, that
R() R() for every ordinal numbers < . (9.1)
An easy inductive proof of this fact is left as an exercise.
Lemma 9.1.2 For every set x there exists an ordinal number such that
x R().
Proof First notice that for any set x,
() if for every y x there exists an ordinal number such that y R(),
then x R(
0
) for some ordinal
0
.
To see this, let r be a function such that r(y) = min: y R( + 1)
for every y x. Such a function exists by the axiom of replacement.
Let =

yx
r(y). Then is an ordinal number and x R(). So
x {(R()) = R( + 1).
Now, to obtain a contradiction, assume that there exists an x such that
x R() for every ordinal number . So, by (), there exists a y x such
that y R() for every ordinal number . In particular, the set
A = y trcl(x): y R() for every ordinal number
is not empty. So, by the axiom of regularity, A contains an -minimal
element, that is, there exists an x
0
A such that A x
0
= . But x
0

A trcl(x). Hence x
0
trcl(x), since trcl(x) is transitive. Therefore x
0
satises the assumption of () so x
0
R(
0
) for some ordinal number
0
.
But this contradicts x
0
A.

166 9 Forcing
For a set x dene the rank of x by
rank(x) = min: x R( + 1).
Notice that by Lemma 9.1.2 rank(x) is dened for every set x.
Lemma 9.1.3
(a) rank(x) < if and only if x R().
(b) If x y then rank(x) < rank(y).
Proof (a) This is obvious by (9.1) and the denition of rank(x).
(b) If = rank(y) then y R( + 1) = {(R()). So x y R().
Therefore rank(x) < = rank(y).
Before we describe the forcing method, we will reexamine some notions
discussed in Sections 1.1 and 1.2. This should give us a better under-
standing of what we mean when we say that a property is independent of
ZFC.
In what follows the term formula will always be understood as a
formula of the language of set theory, that is, the language described in
Section 1.2. In particular, the set of all formulas is dened by induction on
their length from only two basic kinds of formula: x y and x = y,
where symbols x and y represent variables. A more complicated formula
can be built from the less complicated formulas and only by connecting
them with logical connectors, &, , , , or by preceding
one of them with the negation, , or a quantier, x, x. For example,
x = y & x y (x y) is a correct formula, whereas expressions such
as x y z or z = t c are not. (Although x y z could be
interpreted as x y & y z, the latter being a correct formula.)
A variable in a formula is any symbol that represents a set. More
precisely, it is any symbol that is neither of the following: , =,
, &, , , , , , (, or ). In a given formula
not all variables must be alike. For example, in the formula dened as
x = y & z (z = z) the variables x and y are parameters and might
be either true or false, depending on what value we associate with x and
y. On the other hand, the variable z lies within the scope of the quantier
and is not a parameter.
The variables of a formula that are within the scope of a quantier
are said to be bound in . The parameters of a formula will be referred to
as free variables. Notice that the same symbol in the same formula may be
used both as a bound variable and as a free variable. For example, this is
the case in the formula x (x y) & x = y, where the rst x is bound,
while the second plays the role of a free variable. (Notice the importance
of the distribution of parentheses.)

9.1 Elements of logic and other forcing preliminaries 167
Formulas without free variables are called sentences. Since all variables
in a sentence are bound, no interpretation of free variables is necessary to
decide whether it is true or false. For this reason only sentences will be used
as axioms. On the other hand, if x
1
, . . . , x
n
are free variables of a formula
, we often write (x
1
, . . . , x
n
) instead of to emphasize its dependence
on x
1
, . . . , x
n
. In particular, we will usually write x(x) and x(x) in
place of x and x.
By a theory we will mean any set of sentences. The sentences belonging
to a theory T will be treated as its axioms.
2
In particular, ZFC is a theory,
and it consists of innitely many axioms since each of the scheme axioms
(comprehension or replacement) stands for innitely many axioms.
We say that a sentence is a consequence of a theory T (or that can
be proved in a theory T) and write it as T if there is a formal proof of
using as axioms only sentences from T. It is also equivalent to the fact
that there exists a nite set T
0
=
0
, . . . ,
n
of axioms from T such that
(
0
& &
n
) is a consequence of axioms of logic. (See Appendix B for
more details.)
A theory T is said to be inconsistent if there is a sentence such that
T (&), that is, if it leads to a contradiction. Equivalently, T is incon-
sistent if every sentence is a consequence of T. (The equivalence follows
from the fact that the implication (&) is true for all formulas
and .) Conversely, we say that a theory T is consistent and write Con(T)
if T is not inconsistent, that is, if T does not imply a contradiction. A
sentence is consistent with theory T if T+ is consistent.
Evidently, from the point of view that theories should carry useful in-
formation, only consistent ones are interesting. In particular, we will be
assuming here that ZFC is consistent. (Recall that by the G odels second
incompleteness theorem, Theorem 1.1.1, there is no hope of proving the
consistency of ZFC in itself as long as it is really consistent.)
The relation between the notions of being a consequence of a theory
and being consistent with a theory is best captured by the following fact,
whose easy proof is left as an exercise:
T - if and only if Con(T + ). (9.2)
Recall also that a sentence is independent of theory T if neither T
nor T or, equivalently, when both theories T + and T +
are consistent. Thus G odels rst incompleteness theorem, Theorem 1.1.2,
implies that there are sentences of set theory that are independent of ZFC.
2
We assume that theory contains only sentences, since it is more dicult to accept a
formula with some free variables as an axiom. (Intuitively whether it is true or false can
depend on the value of the free variables.) However, logically any formula is equivalent
to a sentence x
1
xn, where {x
1
, . . . , xn} is a list of all free variables in . So our
restriction is not essential.

168 9 Forcing
The forcing method will be a tool to show that a given sentence is
consistent with ZFC. So, in order to prove that a sentence is independent
of ZFC, we will be proving that the theories ZFC+ and ZFC+ are
both consistent.
EXERCISES
1 Complete the details of the proof of Lemma 9.1.1 by showing the mini-
mality of trcl(x).
2 Prove (9.1).
3 Prove (9.2). Hint: Use the fact that (T ) if and only if
T ().
9.2 Forcing method and a model for CH
Consider a transitive set M. For every formula of the language of set
theory (without any shortcuts) we dene a formula
M
, called its relativiza-
tion to M. It is obtained by replacing in each unbounded quantier x
or x with its bounded counterpart x M or x M. For example, if

0
is a sentence from the axiom of extensionality
x y [z (z xz y)x = y]
then
M
0
stands for
x M y M [z M (z xz y)x = y] .
In particular, if (x
1
, . . . , x
n
) is a formula with free variables x
1
, . . . , x
n
and t
1
, . . . , t
n
M then
M
(t
1
, . . . , t
n
) says that (t
1
, . . . , t
n
) is true
under the interpretation that all variables under quantiers are bound to
M. In other words,
M
(t
1
, . . . , t
n
) represents the formula (t
1
, . . . , t
n
) as
seen by a person living inside M, that is, thinking that M represents the
entire class of all sets. (This is the best way to think of
M
.)
For a transitive set M and a formula (with possible parameters from
M) we say that is true in M and write M [= if
M
is true. For a
theory T we say that T is true in M or that M is a model for T if

M
holds for every from T.
Note that a model for a theory T is just a set satisfying some properties.
3
Thus, if M is the empty set, then M is a model for the theory T consisting
3
For those readers who have been exposed previously to any kind of model theory we
give here a bit of an explanation. We do not need to interpret any constants of set
theory in M, since the language of set theory does not contain any constants. And the
only relations that we have to take care of are the relation symbols and =, which
are interpreted as the real relations of being an element of and being equal to.

9.2 Forcing method and a model for CH 169
of all the axioms of ZFC except the set existence and the innity axioms.
This is so since any other axiom starts with a general quantier x, and its
relativization to M = starts with x , that is, is satised vacuously.
Similarly, if M is any transitive set, then the axiom of extensionality is
satised in M, since for any set x M
z x if and only if z M & z x.
It is also relatively easy to see that M = R() is a model for ZFC minus
the innity axiom (see Exercise 1).
Forcing consistency proofs will be based on the following fundamental
principle.
Forcing principle In order to prove that the consistency of ZFC implies
the consistency of ZFC+ it is enough to show (in ZFC) that
(F) every countable transitive model M of ZFC can be extended to
a countable transitive model N of ZFC+.
We use here the word extend in the sense of inclusion, that is, N extends
M means that M N.
In what follows we will use the letters CTM as an abbreviation for
countable transitive model. Thus the forcing principle asserts that if we
can prove in ZFC the implication
[M(M is a CTM for ZFC)][N(M N & N is a CTM for ZFC+)]
then we can conclude from it that Con(ZFC) implies Con(ZFC + ).
We will not prove the forcing principle here, since its proof is not im-
portant for the applications of the forcing method. However, a sketch of
its proof can be found in Appendix B.
To use the forcing principle we will assume that we have a countable
transitive model M of ZFC and then will extend it to an appropriate
model N. To describe the extension method we need a few more de-
nitions. Let M be an arbitrary family of sets and let P be a partially
ordered set. A lter G in P is M-generic if D G = for every dense
subset D of P that belongs to M. Notice that the RasiowaSikorski lemma
(Theorem 8.1.2) immediately implies the following lemma since the family
D M: D P & D is dense in P is countable for any countable set M.
Lemma 9.2.1 For every partially ordered set P, p P, and countable set
M there exists an M-generic lter G in P such that p G.
The model N from (F) will be constructed by using the following the-
orem.

170 9 Forcing
Theorem 9.2.2 For any countable transitive model M of ZFC, partially
ordered set P = 'P
0
, ` M, and M-generic lter G in P there exists
a smallest countable transitive model N of ZFC such that M N and
G N.
The model N from Theorem 9.2.2 is usually denoted by M[G] and is
called a generic extension of M. We will also refer to M as the ground
model of M[G].
Before we present a construction of M[G] we will have a closer look at
Theorem 9.2.2. To make good use of it we must expect that a countable
transitive model M for some property (such as CH) can be extended to a
model M[G] of its negation . For this, M[G] must be a proper extension
of M. However, Theorem 9.2.2 does not rule out that M[G] = M. In fact,
if G M then clearly M[G] = M. Fortunately, the extension will be proper
for a large class of partially ordered sets described in Proposition 9.2.3.
Proposition 9.2.3 Let M be a transitive model of ZFC and let P M be
a partially ordered set such that
() for every p P there are two incompatible q, r P below p.
If G is an M-generic lter in P then G M.
Proof Notice that () implies that D = P`G is dense in P. Now if D M
then, by the denition of an M-generic lter, we would have G D = ,
which is impossible, since GD = G(P`G) = . So D M. But P M
and M is closed under taking the dierence of two sets, since it is a model
of ZFC. Therefore G M, since otherwise D = P ` G M.
Without any doubt the reader should notice a similarity between Mar-
tins axiom and the statement of Theorem 9.2.2. In fact, the generic lter
G from the theorem will be used in M[G] in a way similar to the way we
used generic lters in the proofs where Martins axiom was used. To see the
similarities as well as dierences between such uses in both cases consider
the following example.
Example 9.1 Let M be a countable transitive model of ZFC. For sets X
and Y from M consider the forcing P(X, Y ) = 'Func

(X, Y ), `. Then
P(X, Y ) belongs to M, being dened from X, Y M. It is easy to see that
if X is innite then the sets
D
x
= p P(X, Y ): x dom(p) for x X
and
R
y
= p P(X, Y ): y range(p) for y Y

9.2 Forcing method and a model for CH 171
are dense in P(X, Y ). They belong to M, since each of them is dened
from P(X, Y ) and either x X M or y Y M. In particular, if G
is an M-generic lter in P(X, Y ) then, by Proposition 8.1.1, g =

G is a
function from X onto Y .
In the proof that CH is consistent with ZFC we will use the forcing
P
0
= P(
2
, 2). Then, for an M-generic lter G
0
in P
0
, we see that
g =

G
0
:
2
2 belongs to M[G
0
]. Moreover, notice that for every
< <
2
the sets
E

= p P
0
: n < [', n`, ', n` dom(p) & p(, n) = p(, n)]
are dense in P
0
and belong to M. Thus an M-generic lter intersects every
such set, that is,
for every < <
2
there exists an n < such that g(, n) =g(, n). (9.3)
Now for every <
2
we can dene a function g

: 2 by g

(n) = g(, n).

The set g

: <
2
belongs to M[G], since it has been dened only
from g M[G]. Moreover, by (9.3), all functions g

are distinct and

g

: <
2
2

. So we have proved that (in M[G

0
]) the family 2

contains a subset g

: <
2
of cardinality
2
, so c = [2

[
2
>
1
.
That is, in M[G
0
] the continuum hypothesis is false.
Is this really all we need to prove? Is it that simple? Unfortunately the
preceding argument contains several imprecisions and gaps. To see this let
us consider another forcing P

= P(,
1
) = 'Func

(,
1
), ` and let G

be an M-generic lter in P

. Then, as we noticed before, g =

G

is a
function from onto
1
! This looks like a clear contradiction. What is
wrong? Is Theorem 9.2.2 false?
The answer is that there is nothing wrong with Theorem 9.2.2. What
is wrong with our contradiction is that there is a misconception of what
the theorem really says. It tells us that M[G] is a model for ZFC as long
as M is such a model. In particular,
M
and
M[G]
are true for every
sentence that is a consequence of ZFC axioms. However, the constants
are not a part of the formal language of set theory, and we use them only
as shortcuts for the formulas representing them. Thus we write x is equal
to
1
to express the fact that x is the unique set satisfying the formula
(x): x is the smallest uncountable ordinal number. Theorem 9.2.2
tells us that the sentence dened by !x (x) is true both in M and
in M[G], since the existence of a unique
1
can be proved in ZFC. Thus

N
holds for N = M or N = M[G], and there exists a unique
N
1
N
such that
N
(
N
1
) is true, that is, that
N
1
is an
1
with respect to N.
However, Theorem 9.2.2 does not tell us that
M
1
=
M[G]
1
! In fact, since
M is countable and transitive, every element of M is countable. Thus by

172 9 Forcing
manipulating M from outside we may be able to take any set x M that
is uncountable in M and nd a generic extension M[G] of M in which x
is countable. And this is precisely what happened in M[G

]! The ordinal
number
M
1
became a countable object in M[G

], and
M
1
is not equal to

M[G

]
1
. (The argument that indeed
M
1
=
M[G

]
1
will be completed after
Lemma 9.2.5.)
In fact, so far we are not even sure whether what is an ordinal number
in M remains so in M[G], or even whether what is 2 in M remains 2 in
M[G]. To address this issue we have to examine our set-theoretic vocab-
ulary more carefully. In particular, the general question we would like to
examine is, Which properties of elements of M are preserved in its exten-
sion M[G]? (This can be viewed as an analog of the preservation problems
for transnite-induction constructions discussed in Section 6.1.)
To address this issue recall once more that every set-theoretic term we
use can be expressed by a formula of set theory. So the question may be
rephrased as follows:
For which properties P does there exist a formula (x
1
, . . . , x
n
)
describing P such that if (t
1
, . . . , t
n
) is true in M for some
t
1
, . . . , t
n
M then (t
1
, . . . , t
n
) is also true in M[G]?
The properties that are always preserved in such extensions will be called
absolute properties. More precisely, a property is absolute provided it can
be expressed by a formula (x
1
, . . . , x
n
) such that for every transitive
models M and N of ZFC with M N and for every t
1
, . . . , t
n
M
formula
M
(t
1
, . . . , t
n
) holds if and only if
N
(t
1
, . . . , t
n
) holds.
The absoluteness of most properties we are interested in follows just
from the fact that the models we consider are transitive. To identify a
large class of such properties we need the following notion. A formula is
a
0
-formula if it can be written such that the only quantiers it contains
are bounded quantiers, that is, quantiers of the form x y or x y.
Lemma 9.2.4 Let M be a transitive set and let (x
1
, . . . , x
n
) be a
0
-
formula. If t
1
, . . . , t
n
M then
M
(t
1
, . . . , t
n
) is equivalent to (t
1
, . . . , t
n
)
in the sense that the sentence
M[M is transitive (t
1
, . . . , t
n
M)(
M
(t
1
, . . . , t
n
) (t
1
, . . . , t
n
))]
is provable in ZFC. In particular, the property expressed by is absolute.
Proof The proof goes by induction on the length of formula .
Clearly it is true for any basic formula x y or x = y, since for any
quantier-free formula its relativization
M
is identical to .
So assume that formula is built from the less complicated formulas

0
and
1
, that is, is in one of the following forms:
0
&
1
,
0

1
,

9.2 Forcing method and a model for CH 173

1
,
0

1
,
0
, x t
0
, or x t
0
. For example, if =
1

2
and the lemma is true for all formulas of length less than the length of ,
then it holds for
0
and
1
. In particular,

M
= (
0

1
)
M
=
M
0

M
1
=
0

1
= .
We argue similarly if is the negation of a formula, or is obtained from
two formulas by using the logical operations &, , or .
If is of the form x t
0
(x) for some t M then its formal represen-
tation is x [x t &
0
(x)]. Then
M
stands for x M [x t &
M
0
(x)],
which is equivalent to (x tM)
0
(x) since
0
(x) is equivalent to
M
0
(x)
by the inductive assumption. But t M and M is transitive. So t M
and t M = t. Therefore
M
is equivalent to x t
0
(x), that is, to .
The case when is of the form x t
0
(x) is similar.
For example, the property y is a union of x, y =

x, is absolute,
since it can be written as the following
0
-formula
1
(x, y):
w yz x(w z) & z xw z(w y).
Lemma 9.2.5 The following properties can be expressed by a
0
-formula.
In particular, they are absolute.
(0) x y.
(1) y =

x.
(2) y =

x.
(3) z = x y.
(4) z = x y.
(5) z = x ` y.
(6) z = x, y.
(7) z is an unordered pair.
(8) z = 'x, y` (i.e., z = x, x, y).
(9) z is an ordered pair.
(10) z = x y.
(11) r is a binary relation.
(12) d is the domain of binary relation r.

174 9 Forcing
(13) R is the range of binary relation r.
(14) f is a function.
(15) Function f is injective.
(16) is a partial-order relation on P (i.e., 'P, ` is a partially ordered
set).
(17) D is a dense subset of the partially ordered set 'P, `.
(18) A is an antichain in the partially ordered set 'P, `.
(19) Set x is transitive.
(20) is an ordinal number.
(21) is a limit ordinal number.
(22) is the rst nonzero limit ordinal number (i.e., = ).
(23) is a nite ordinal number.
(24) is a successor ordinal number.
Proof The
0
-formulas for these properties can be found in Appendix B.

Now, for a property P expressed by a formula (x), a transitive model

M of ZFC, and t M the statement t has the property P may be inter-
preted in two dierent ways: as either that t satises the formula (x), that
is, that (t) holds, or that (x) is satised by t in M, that is, that
M
(t)
is true. In the latter case we will say that t has the property P in M,
reserving the statement t has the property P for the former case. How-
ever, for the absolute properties, such as is an ordinal number, such
a distinction is redundant. Thus, we will say is an ordinal number
independently of whether we think of as an ordinal number in M or
outside it.
A similar distinction has to be made when we talk about constants.
When we write we think of it as is the rst innite ordinal number. In
principle, an object
M
that satises this property in M might be dierent
from the real . However, for objects represented by absolute formulas
this cannot happen. So in all these cases we will drop the superscript
M
from the symbol representing an absolute constant, even when we think of
it as being in M. In particular, we will use the same symbol regardless
of the context, since
M
= . For the cases of the constants that are not
absolute, such as
1
or c, we will keep using the superscript
M
to denote

9.2 Forcing method and a model for CH 175
their counterparts in M. For example, we will write
M
1
and c
M
to express

1
and c in M.
In fact, we have already used this convention when arguing that
M
1
=

M[G

]
1
, with G

being an M-generic lter in P

= 'Func

(,
1
), `. And
now we can indeed see that
M
1
=
M[G

]
1
. This is so since
M
1
and
M[G

]
1
are (real) ordinal numbers, and in M[G

] there is a (real) function f from

(true) onto
M
1
. So
M
1
is a countable ordinal in M[G

], whereas
M[G

]
1
is an uncountable ordinal in M[G

] by its denition. Thus they cannot be

equal.
Proof of Theorem 9.2.2 To construct the set M[G] x a countable tran-
sitive model M of ZFC, a partially ordered set P M, and an M-generic
lter G in P.
We say that is a P-name provided is a binary relation and
', p` ( is a P-name & p P).
Formally, we dene the property is a P-name recursively by induction
on rank(). Thus is a P-name if it satises the following recursive formula:
(): there exists an ordinal number and a function f such that dom(f) =
+1, f() = , and for every +1, t f(), and z t there
exist , f(), and p P such that z = ', p`.
It is not dicult to see that is absolute (see Exercise 2). So the property
is a P-name is absolute too.
The set of all P-names that belong to M will be denoted by M
P
, that is,
M
P
= M: is a P-name.
It can be proved that the class of all P-names is a proper class. Thus M
P
is a proper class in M. It particular, it does not belong to M.
For a P-name we dene the valuation of by
val
G
() = val
G
(): p G (', p` ).
Once again, the formal denition of val
G
() should be given by a recursive
formula. Now we dene M[G] by
M[G] = val
G
(): M
P
.
This is the model from the statement of Theorem 9.2.2.
We will not check here that all ZFC axioms are true in M[G], since it
is tedious work and has little to do with the forcing applications. However,
we will argue here for its other properties listed in Theorem 9.2.2.

176 9 Forcing
Transitivity of M[G] follows immediately from the denition of M[G].
Take y M[G]. Thus y = val
G
() = val
G
(): p G (', p` ) for
some M
P
. So any x y is in the form val
G
() for some ', p` M.
Thus is a P-name that belongs to M, and x = val
G
() M[G].
The minimality of M[G] is clear, since every element of M[G] is of
the form val
G
() and so must belong to any transitive model of ZFC that
contains G and .
To see that M[G] is countable notice rst that it is a set by the axiom
of replacement since val
G
is a function dened by a formula and M[G] =
val
G
[M
P
]. It is countable, since M
P
M is countable.
To argue for M M[G] and G M[G] we need a few more denitions.
For a set x dene, by recursion, its P-name x by
x = ' y, p`: y x & p P.
(Note that the denition of x depends on P.)
Lemma 9.2.6 If M is a countable transitive model of ZFC, P M is a
partially ordered set, and G is an M-generic lter over P then for every
x M
x M
P
and val
G
( x) = x.
In particular, M M[G].
Proof It is an easy inductive argument. First prove by induction on
rank(x) that x is a P-name. Then notice that the denition of x is absolute
and x M. So x M. Finally, another induction on rank(x) shows that
val
G
( x) = x. The details are left as an exercise.
To see that G M[G] let = ' p, p`: p P. It is clearly a P-name
dened in M from P. So M
P
. Moreover,
val
G
() = val
G
( p): p G (' p, p` ) = p P: p G = G.
So G M[G].
This nishes the proof of Theorem 9.2.2.
The last important basic problem we have to face is how to check that a
given property is true in M[G]. For this, we will introduce the following
denition. For a formula (x
1
, . . . , x
n
), a countable transitive model M
of ZFC, a partial order P M, and P-names
1
, . . . ,
n
M we say that
p P forces (
1
, . . . ,
n
) and write
p |
P,M
(
1
, . . . ,
n
)
provided (val
G
(
1
), . . . , val
G
(
n
)) is true in M[G] for every M-generic
lter G in P that contains p. We will usually just write | in place of

9.2 Forcing method and a model for CH 177
|
P,M
when P and M are xed. We will also write P | (
1
, . . . ,
n
) if
p | (
1
, . . . ,
n
) for every p P. Notice that
if p | (
1
, . . . ,
n
) and q p then q | (
1
, . . . ,
n
), (9.4)
since every M-generic lter G that contains q contains also p.
Clearly, the denition of the relation | depends on a model M and
a set P and involves the knowledge of all M-generic lters in P. Thus, it
is clearly dened outside M. One of the most important facts concerning
forcing is that an equivalent form of | can also be found inside M, in a
sense described in the following theorem.
Theorem 9.2.7 For every formula (x
1
, . . . , x
n
) of set theory there exists
another formula (, P, x
1
, . . . , x
n
), denoted by |

P
(x
1
, . . . , x
n
), such
that for every countable transitive model M of ZFC, partial order P M,
and P-names
1
, . . . ,
n
M
p |
P,M
(
1
, . . . ,
n
) M [= (p |

P
(
1
, . . . ,
n
)) (9.5)
for every p P. Moreover,
M[G] [= (val
G
(
1
), . . . , val
G
(
n
)) p G(p | (
1
, . . . ,
n
)) (9.6)
for every M-generic lter G in P.
This theorem will be left without proof.
The relation |

from Theorem 9.2.7 will justify the use of all forcing

arguments within a model M and will let us conclude all essential properties
of M[G] without knowledge of G. We will often write | where formally
the relation |

should be used.
This completes all general details concerning the forcing method. How-
ever, before we give a proof of Con(ZFC+CH) we still need some techni-
cal lemmas. The rst of them tells us that a P-name representing a function
with domain and range belonging to a ground model can be chosen to have
a particularly nice form.
Lemma 9.2.8 Let M be a countable transitive model of ZFC and P M
be a partially ordered set. If X, Y M, p
0
P, and M
P
are such that
p
0
| is a function from

X into

Y then there exists a M
P
such
that
(a) '

'x, y`, p`: x X & y Y & p P & p p

0
;
(b) p
0
| = ; and
(c) A
x
= p P: y Y ('

'x, y`, p` ) is an antichain for every

x X.

178 9 Forcing
Proof For x X let
D
x
= p P: p p
0
& y Y (p |

'x, y` ).
Notice that D
x
M since, by condition (9.5) of Theorem 9.2.7,
D
x
= p P: p p
0
& y Y (p |

'x, y` )
M
.
Let

A
x
be a maximal subset of D
x
of pairwise-incompatible elements. Thus
every

A
x
is an antichain. Dene
= '

'x, y`, p`: x X & y Y & p

A
x
& p |

'x, y` .
Once again M follows from (9.5), since
= '

'x, y`, p`: x X & y Y & p

A
x
& (p |

'x, y` )
M

and the family

A
x
: x X can be chosen from M.
Clearly is a P-name satisfying (a).
To see (c) it is enough to notice that A
x
=

A
x
for every x X. This is
the case since
p A
x
y Y '

'x, y`, p`
y Y

p

A
x
& p |

'x, y`

p

A
x
.
To prove that p
0
| = , let G be an M-generic lter in P such that
p
0
G. Then g = val
G
() is a function from X into Y , since p
0
forces it.
Also, clearly val
G
() X Y . So let x X and y Y . It is enough to
show that 'x, y` val
G
() if and only if 'x, y` val
G
().
But if 'x, y` val
G
() then there exists a p G such that '

'x, y`, p` .
In particular, p |

'x, y` . But p G. Therefore 'x, y` val
G
().
To prove the other implication rst note that the set
E
x
= q p
0
: p

A
x
(q p) q P: q is incompatible with p
0

is dense in P. To see it, let s P. We have to nd q E

x
with q s.
If s is incompatible with p
0
then q = s E
x
. So assume that s is
compatible with p
0
and choose p
1
P extending s and p
0
. Let H be an
M-generic lter in P containing p
1
. Then h = val
H
() is a function from X
into Y since the condition p
0
H forces it. Let y = h(x). Then, by (9.6),
there exists an r H forcing it, that is, such that r |

'x, y` . Choose
p
2
H extending r H and p
1
H. Then, by (9.4), p
2
|

'x, y` . In

9.2 Forcing method and a model for CH 179
particular, p
2
D
x
. But

A
x
is a maximal antichain in D
x
. So there exists
a q p
2
extending some p from

A
x
. Hence q p
2
p
0
, so q E
x
, and
q p
2
p
1
s. The density of E
x
has been proved.
Now let 'x, y` val
G
(). Then, by (9.6), there exists an r G that
forces it, that is, such that r |

'x, y` . Moreover, there exists a
q E
x
G, since G is M-generic and E
x
M is dense. But q G is
compatible with p
0
G. So there exists a p

A
x
D
x
with q p. Thus,
by the denition of D
x
, there exists a y
1
Y such that p |

'x, y
1
` .
In particular, '

'x, y
1
`, p` and 'x, y
1
` val
G
(), since p G. To nish
the proof it is enough to notice that y
1
= y. So let s G be a common
extension of r, p, and p
0
. Then, by (9.4),
s |

'x, y` &

'x, y
1
` & is a function

.
Therefore, if H is an M-generic lter in P containing s then h = val
H
()
is a function and y = h(x) = y
1
.
Corollary 9.2.9 Let M be a countable transitive model of ZFC and P M
be a partially ordered set that is ccc in M. Let G be an M-generic lter in
P, and let g M[G] be such that g : X Y and X, Y M. Then there
exists a function F M such that
M [= F : X [Y ]

and g(x) F(x) for all x X.

Proof Let M be a P-name such that val
G
() = g. Since
M[G] [= g is a function from X into Y ,
by Theorem 9.2.7(9.6) there exists a p
0
G such that
p
0
| is a function from

X into

Y .
Let M be a P-name from Lemma 9.2.8 and dene
F(x) = y Y : p
y
P ('

'x, y`, p
y
` ).
Then clearly F M and M [= F : X {(Y ). To see that for every
x X the set F(x) is countable in M note that the map g : F(x) A
x
,
g(y) = p
y
, is one-to-one.
Indeed, if y, z F(x) and p
y
= p
z
then p = p
y
p
0
and so
p |

'x, y` &

'x, z` & is a function

.

180 9 Forcing
Thus y = z. In particular,
M [= [F(x)[ [A
x
[ ,
as P is ccc in M. Finally, since g = val
G
() = val
G
()
y = g(x) p G

'

'x, y`, p`

y F(x).
Corollary 9.2.9 has been proved.
Next we will show that ccc forcings are very nice with respect to the
cardinal numbers. To formulate this more precisely we need a few deni-
tions.
Let M be a countable transitive model of ZFC, P M be a partial order,
and M be an ordinal number such that M [= is a cardinal number.
If M[G] [= is a cardinal number for every M-generic lter G in P then
we say that forcing P preserves the cardinal . On the other hand, if
M[G] [= is not a cardinal number for every M-generic lter G in P
then we say that forcing P collapses the cardinal . If P preserves all
cardinal numbers from M then we simply say that P preserves cardinals.
Lemma 9.2.10 Let M be a countable transitive model of ZFC and P M
be a partial order that preserves cardinals. If G is an M-generic in P then

=
M[G]

for every ordinal number M.

Proof This follows by an easy induction; it is left as an exercise.
Theorem 9.2.11 Let M be a countable transitive model of ZFC. If P M
is a partial order such that M [= P is ccc then P preserves cardinals.
Proof Let M and P be as in the theorem and assume, to obtain a contra-
diction, that there is an ordinal number and an M-generic lter G in P
such that is a cardinal in M but is not in M[G]. By the absoluteness of
nite cardinals and it must be that > .
Since is still an ordinal in M[G] it means that it is not initial, that
is, there exist an ordinal < and a function g M[G] such that g
maps onto . So, by Corollary 9.2.9, there exists a function F M
such that M [= F : []

and g() F() for all . Hence

= g[]

F().
But

F() M, since F M. So the following holds in M:

F()

[[ < ,

9.2 Forcing method and a model for CH 181
since all sets F() are countable in M. This contradiction nishes the
proof.
To apply this theorem to the proof of the consistency of ZFC+CH we
need one more fact (compare Exercise 1 from Section 8.2).
Lemma 9.2.12 Forcing P = 'Func

(X, Y ), ` is ccc for [Y [ .

Proof Let 'p

: <
1
` be a one-to-one sequence of elements of P. We
have to nd < <
1
such that p

and p

are compatible.
Consider the family / = dom(p

): <
1
of nite sets. Notice that
/ must be uncountable, since for every A / there is at most [Y
A
[
p

s with A = dom(p

). So, by the -system lemma (Lemma 8.2.11), there

exist an uncountable set S
1
and a nite A with the property that
dom(p

)dom(p

) = A for any distinct , S. But the set Y

A
is at most
countable. So there exist distinct and from S such that p

[
A
= p

[
A
.
This implies that p = p

is a function, that is, p P. Since p clearly

extends p

and p

we conclude that p

and p

are compatible.
Theorem 9.2.13 Theory ZFC+CH is consistent.
Proof We will follow the path described in Example 9.1. Let M be a
countable transitive model of ZFC. Dene (in M)
P = 'Func

(
M
2
, 2), `
and let G be an M-generic lter in P. We will show that c >
1
in M[G].
For x
M
2
let
D
x
= p P: x dom(p).
It is easy to see that all sets D
x
belong to M since they are dened using
x, P M. They are dense, since for every p P either x dom(p)
and p D
x
, or x dom(p) and p 'x, 0` D
x
extends p. Thus,
by Proposition 8.1.1, g =

G is a function from
M
2
to 2. Clearly
g M[G], since it is constructed from G M[G].
Similarly as for D
x
we argue that for every < <
M
2
the sets
E

= p P: n < [', n`, ', n` dom(p) & p(, n) = p(, n)]

are dense in P and belong to M. Thus an M-generic lter intersects every
such set, that is, for every < <
M
2
there exists an n < such that
g(, n) = g(, n).
Now for every <
M
2
we can dene (in M[G]) a function g

: 2
by g

(n) = g(, n). The set g

: <
M
2
belongs to M[G], since it has

182 9 Forcing
been constructed only from g M[G]. Moreover, g

: <
M
2
2

and
all the functions g

are dierent. So
M[G] [= c = [2

[
M
2
.
But, by Lemma 9.2.12, forcing P is ccc in M and so, by Theorem 9.2.7, it
preserves cardinals. In particular, by Lemma 9.2.10,
M
2
=
M[G]
2
, that is,
M[G] [= c
M[G]
2
>
1
.
This nishes the proof.
A model M[G] obtained as a generic extension of a ground model M via
the forcing P = 'Func

(X, 2), ` (with X M) is called a Cohen model.

The numbers g

from Theorem 9.2.13 are called Cohen (real) numbers.

EXERCISES
1 Prove that x R() if and only if [ trcl(x)[ < . Use this to show that
M = R() is a model for ZFC minus the innity axiom.
2 Let () be a formula expressing the property that is a P-name and
let M be a CTM of ZFC with P M. Show, by induction on rank(), that
for every M
() if and only if
M
().
Thus the property is a P-name is absolute.
3 Complete the details of the proof of Lemma 9.2.6.
4 Let M be a countable transitive model of ZFC, P M be a partial
order, and G be an M-generic lter in P. Show that
M

M[G]

for every
ordinal number M.
5 Prove Lemma 9.2.10.
9.3 Model for CH and
The main goal of this section is to prove the following theorem.
Theorem 9.3.1 Let M be a countable transitive model of ZFC, and let
P = (Func
1
(
1
, 2))
M
. If G is an M-generic lter in P then
M
1
=
M[G]
1
,
{() M = {() M[G], and holds in M[G].
Since implies CH the theorem immediately implies the following
corollary.

9.3 Model for CH and 183
Corollary 9.3.2 Theories ZFC+CH and ZFC+ are consistent.
Combining this with Theorem 9.2.13 we immediately obtain the follow-
ing.
Corollary 9.3.3 The continuum hypothesis is independent of ZFC.
Note that P = Func
1
(
1
, 2) stands for the set of all functions p: A 2,
where A is an at most countable subset of
1
. Since the terms
1
and
being countable are not absolute, the superscript
M
in the denition of
P in Theorem 9.3.1 is essential.
We will prove Theorem 9.3.1 with the help of several general facts that
are, in most cases, as important as the theorem itself.
Let P be a partially ordered set and let p P. A set D P is dense
below p if for every q p there exists a d D such that d q.
Proposition 9.3.4 Let M be a countable transitive model of ZFC, P M
be a partially ordered set, and p P. If D M is a subset of P that is
dense below p then G D = for every M-generic lter G in P such that
p G.
Proof Let R = r P: r is incompatible with p. Then D
0
= DR M
and D
0
is dense in P. So if G is an M-generic lter in P then G D
0
= .
But if p G then G R = , so G D = .
A partially ordered set P is countably closed if for every decreasing
sequence 'p
n
P: n < ` there is a p

P such that p

p
n
for every
n < . Notice that P

= Func
1
(
1
, 2) is countably closed since p

n<
p
n
P extends every p
n
. So P = (Func
1
(
1
, 2))
M
is countably
closed in M.
Theorem 9.3.5 Let M be a countable transitive model of ZFC and P M
be a partially ordered set that is countably closed in M. Let A, B M be
such that A is countable in M and let G be an M-generic lter in P. If
f M[G] is a function from A into B then f M.
Proof Let M be a P-name such that val
G
() = f. Then, by Theo-
rem 9.2.7, there exists a p G such that
p | is a function from

A into

B.
Dene in M
D = q P: g B
A
(q | = g).
(D belongs to M since the relation | can be replaced by the formula |

.)
By Proposition 9.3.4 it is enough to prove that D is dense below p since

184 9 Forcing
then there exists a q G D, that is, there is a g M such that g maps
A to B and
f = val
G
() = val
G
( g) = g M.
To prove that D is dense below p, x r P such that r p. We have
to nd a q D with q r. Let a
n
: n < be an enumeration of A in
M. Dene, in M, sequences 'p
n
P: n < ` and 'b
n
B: n < ` such
that for all n <
(1) p
0
= r,
(2) p
n+1
p
n
,
(3) p
n+1
| ( a
n
) =

b
n
.
To see that such a construction can be made assume that p
i
for i n
and b
j
for j < n have already been constructed. It is enough to show that
there exist p
n+1
and b
n
satisfying (2) and (3).
The proof of the existence of such p
n+1
and b
n
will be done outside M.
For this, let H be an M-generic lter in P with p
n
H. Then h = val
H
()
is a function in M[H] from A into B, since p
n
p forces it. Thus there is
a b B such that h(a
n
) = b and we can nd an r H which forces it, that
is, such that r | ( a
n
) =

b. Rening r, if necessary, we can also assume
that r p
n
. Then p
n+1
= r and b
n
= b satisfy (2) and (3).
Now let p

P be such that p

p
n
for every n < and let g : A B
be such that g(a
n
) = b
n
. Then g M and
p

| ( a
n
) =

b
n
for every n < . To nish the proof it is enough to show that
p

| = g
since then q = p

r and belongs to D.
So let H be an arbitrary M-generic lter in P with p

H. Then
h = val
H
() is a function in M[H] from A into B. Moreover, h(a
n
) = b
n
for every n < since p

p
n
forces it. In particular,
M[H] [= h = val
H
() = g
that is, p

| = g.
Theorem 9.3.5 tells us that in a generic extension of a model M obtained
from a countably closed forcing the extension will have the same countable
sequences with elements from a xed set from M. This implies the following
corollary.

9.3 Model for CH and 185
Corollary 9.3.6 Let M be a countable transitive model of ZFC and P M
be a partially ordered set that is countably closed in M. If G is an M-
generic lter in P then {() M = {() M[G] and
M
1
=
M[G]
1
.
Proof Let =
M
1
. If then is an ordinal number in M and in
M[G]. Moreover, if < then there exists an f M M[G] such that f
maps onto . Thus every < is at most countable (in M and in M[G]).
In particular, it is enough to show that is uncountable in M[G], since
then it is the rst uncountable ordinal in M[G], that is,
M
1
= =
M[G]
1
.
To see that is uncountable in M[G], take f : from M[G] and
notice that, by Theorem 9.3.5, f M. So f cannot be onto =
M
1
. Thus
is uncountable in M[G].
To see that {() M = {() M[G] it is enough to prove that
{() M[G] {()M. So let A {()M[G] and let f =

A
: 2.
Then f M[G] and, by Theorem 9.3.5, f M. Therefore A = f
1
(1)
{() M.
Notice that Corollary 9.3.6 implies the statements
M
1
=
M[G]
1
and
{() M = {() M[G] of Theorem 9.3.1 since the forcing P from the
theorem is countably closed in M. Notice also that Corollary 9.3.6 does
not imply that countably closed forcings preserve all cardinal numbers. In
fact, by proving Theorem 9.3.1 we will show that in M[G] there exists a
bijection between
M
1
=
M[G]
1
and {
M
() = {
M[G]
(), that is, there is
a bijection between
M
1
and c
M
. Thus, if
M
1
< c
M
(which can happen
by a result from the previous section) then the cardinal c
M
is collapsed in
M[G].
To prove that holds in the model M[G] from Theorem 9.3.1 the
following lemma will be useful.
Lemma 9.3.7 Let M be a countable transitive model of ZFC and let P, P
1

M be partially ordered sets such that there is an order isomorphism f M

between P and P
1
. If G is an M-generic lter in P then G
1
= f[G] is an
M-generic lter in P
1
and M[G] = M[G
1
].
Proof It is easy to see that G
1
= f[G] is a lter in P
1
. To see that G
1
is M-generic let D
1
M be a dense subset of P
1
. Then D = f
1
(D
1
)
belongs to M and is dense in P. So D G = and
D
1
G
1
= f[D] f[G] = f[D G] = .
To see that M[G] = M[G
1
] notice that M M[G] and M M[G
1
].
Moreover, G
1
= f[G] M[G]; thus, by the minimality of M[G
1
], we con-
clude M[G
1
] M[G]. Similarly, G = f
1
(G
1
) M[G
1
]. Thus M[G]
M[G
1
].

186 9 Forcing
Lemma 9.3.7 tells us that if we replace a forcing by its isomorphic copy
then the appropriate generic extensions will be identical. In particular, if
X M is such that M [= [X[ =
1
then forcings P = (Func
1
(
1
, 2))
M
and
P
1
= (Func
1
(X, 2))
M
are isomorphic in M (by an isomorphism F : P P
1
with F(p)(x) = p(f(x)), where f : X
M
1
is a bijection with f M).
Before we prove that holds in the model M[G] from Theorem 9.3.1
we will show that M[G] [= CH. This is not used in the proof of M[G] [= .
However, the proof of CH in M[G] is simpler than the one for , and it
will prepare the reader for the later one. Moreover, the reader interested
only in the consistency of ZFC+CH can skip the latter part of this section.
By Lemma 9.3.7 we can replace P with P
1
= (Func
1
(
1
, 2))
M
. Let
G
1
be an M-generic lter in P
1
. Thus M[G] = M[G
1
]. Clearly g =

G
1
belongs to M[G] and, by Proposition 8.1.1, it is a function from
M
1

to 2. Notice that for every f 2

M the set
D
f
= p P
1
: <
M
1
[ dom(p) & n < (p(, n) = f(n))]
belongs to M and is dense in P
1
. (To see its denseness, choose p P
1
,
nd <
M
1
with ( ) dom(p) = , and notice that the condition
p '', n`, f(n)`: n < D
f
extends p.) So if f

: 2 for <
M
1
is dened by f

(n) = g(, n) then, by Theorem 9.3.5,

2

M f

: <
M
1
2

M[G] = 2

M.
Thus, in M[G],

M[G]
1
=
M
1
= [
M
1
[ = [f

: <
M
1
[ = [2

M[G][ = c
M[G]
.
So CH holds in M[G].
Proof of Theorem 9.3.1 The proof is similar to that of Theorem 9.3.5.
Let X = ', `: < <
M
1
and replace P = (Func
1
(
1
, 2))
M
with P
2
= (Func
1
(X, 2))
M
. We can do this by Lemma 9.3.7 since these
forcings are isomorphic in M. Let G be an M-generic lter in P
2
and let
g =

G M[G]. Then g is a function from X to 2. For <
M
1
let
g

: 2 be dened by g

() = g(, ) and put A

= g
1

(1). It is easy
to see that the sequence 'A

: <
M
1
` belongs to M[G]. We will show
that it is a -sequence in M[G].
So let A M[G] be such that A
M[G]
1
=
M
1
and let f =

A
. We
have to show that the set
S = <
M
1
: A = A

= <
M
1
: f[

= g

is stationary in M[G] (remember that

M[G]
1
=
M
1
). For this, let C
M
1
,
C M[G], be such that
M[G] [= C is a closed unbounded subset of
1
.

9.3 Model for CH and 187
We will show that S C = , that is, that there exists an C such that
f[

= g

.
Let , M be P
2
-names such that val
G
() = f and val
G
() = C.
Then there exists a p G such that
p |

is a closed unbounded subset of

1
& :
1


2

.
Dene in M the set D by

q P
2
: <
M
1

q |

& <

() =

(, )

,
where is the standard P
2
-name for a generic lter. It is enough to prove
that D is dense below p since then there exists a q GD and so there is
an val
G
() = C such that
f() = g(, ) for all < ,
that is, f[

= g

.
To prove that D is dense below p x an r P such that r p. We have
to nd a q D with q r.
Dene in M the sequences 'p
n
P
2
: n < `, '
n
<
M
1
: n < `,
'
n
<
M
1
: n < `, and 'b
n
2
n
: n < ` such that for every n <
(1) p
0
= r and
0
= 0;
(2) dom(p
n
) ', `: < <
n
;
(3) p
n
p
n1
if n > 0;
(4)
n1
<
n
<
n
if n > 0;
(5) p
n
|

n
& [

n1
=

b
n

if n > 0.
To see that such a construction can be made assume that for some n <
the sequences are dened for all i < n. If n = 0 then p
0
and
0
are already
constructed and
0
and b
0
can be easily chosen, since the conditions (3),
(4), and (5) do not concern this case. So assume that n > 0. We have to
show that the inductive step can be made, that is, that appropriate p
n
,
n
,

n
, and b
n
exist.
So let G
0
be an M-generic lter in P
2
with p
n1
G
0
. Then f
0
=
val
G0
() M[G
0
] is a function from
M
1
to 2 and C
0
= val
G0
() is a
closed and unbounded subset of
1
in M[G
0
], since p
n1
p forces it. In
particular, there exists a d C
0
with d >
n1
. Also, b = h[
n1
M,
since P
2
is countably closed. So there exists a q G
0
, with q p
n1
, that
forces it. Then p
n
= q, b
n
= b, and
n
= d satisfy (3), (5), and the rst part
of (4). The choice of
n
satisfying (2) and (4) nishes the construction.

188 9 Forcing
Now let p

n<
p
n
P
2
. Then for all n <
p

n+1
& [

n
=

b
n+1

.
Thus the functions b
n+1
must be compatible, and b =

n<
b
n+1
M is a
function from =

n<

n
=

n<

n
into 2.
Put q = p

'', `, b()`: < . Notice that q is a function since,

by (2), dom(p

) is disjoint from ', `: < . Since q r it is enough

to show that q belongs to D.
So let G
1
be an M-generic lter in P
2
containing q and let C
1
= val
G1
()
and f
1
= val
G1
(). It is enough to show that C
1
and f
1
() =
(

) (, ) for every < . But C

1
is a closed unbounded subset of
M[G1]
1
and
n+1
C
1
for every n < , since q forces it. Thus =

n<

n
C
1
.
Also, for every < there exists an n < such that <
n
. So
f
1
() = val
G1
()[
n
() = b
n+1
() = b() = q(, ) =

(, ).
This nishes the proof.
The proof just presented can be easily generalized to the following the-
orem. Its proof is left as an exercise.
Theorem 9.3.8 Let M be a countable transitive model of ZFC, be a
regular cardinal in M, and P = (Func

+(
+
, 2))
M
. Then forcing P
preserves all cardinals
+
and 2

=
+
holds in M[G], where G is an
M-generic lter in P.
Theorem 9.3.8 implies immediately the following corollary, which will
be used in Section 9.5.
Corollary 9.3.9 It is relatively consistent with ZFC that 2
1
=
2
.
EXERCISES
1 Complete the details of Lemma 9.3.7 by showing that for every isomor-
phism f between partially ordered sets P and P
1
and every G, D P
(1) G is a lter in P if and only if f[G] is a lter in P
1
;
(2) D is dense in P if and only if f[D] is dense in P
1
.
2 Show that the functions b
n+1
from the proof of Theorem 9.3.1 are indeed
compatible and that b =

n<
b
n+1
is a function from to 2.

9.4 Product lemma and Cohen model 189
3 Prove Theorem 9.3.8 in the following steps.
(a) Generalize the proof of Theorem 9.3.5 to show that if f M[G] is a
function from into A M then f M.
(b) Use (a) and the ideas from the proof of Corollary 9.3.6 to show that
P preserves cardinals
+
and {() M = {() M[G].
(c) Show that the function f :
+
2

, f()() = (

G)(, ), is
onto 2

.
9.4 Product lemma and Cohen model
In the next section we will prove the consistency of MA+CH. The method
used in its proof is called iterated forcing. The idea behind this method is
to repeat the forcing extension process recursively, that is, to construct by
transnite induction of some length a sequence
M = M
0
M
1
M

of countable transitive models of ZFC such that M

+1
= M

[G

] for every
< , where G

is an M

-generic lter in some forcing P

.
This description is very specic and easy to handle at successor stages.
The limit stage, however, presents a problem. For a limit ordinal we
cannot simply take M

<
M

, since such an M

does not have to be

a model of ZFC. Intuitively, we can avoid this problem by dening M

as
a generic extension of M by a sequence 'G

: < `. But in what forcing

should the sequence 'G

: < ` be a generic lter?

To solve this problem, we will nd one forcing in the ground model that
will encode all forcings used in our intuitive inductive construction. In this
project, however, we will nd that the successor step creates more trouble
than the limit one. In fact, at the moment we dont even know whether the
model M
2
= M[G
0
][G
1
] obtained by two consecutive generic extensions can
also be obtained by a single generic extension, that is, whether M
2
= M[G]
for some P M and an M-generic lter G in P.
In this section we will concentrate on this problem only in the easier
case where both partially ordered sets P
0
and P
1
leading to the extensions
M
1
= M[G
1
] and M
2
= M
1
[G
1
], respectively, belong to the ground model
M. The general iteration, including that of two-stage iteration, will be
described in the next section.
In the simple case when P
0
, P
1
M the forcing P is just a product of
forcings 'P
0
,
0
` and 'P
1
,
1
`, that is, P = P
0
P
1
and
'p
0
, p
1
` 'q
0
, q
1
` p
0

0
q
0
& p
1

1
q
1
.

190 9 Forcing
The relation between the forcing extensions obtained by these partially
ordered sets is described in the next theorem. We will start, however, with
the following simple fact.
Proposition 9.4.1 Let P be a product of the forcings P
0
and P
1
. Then G
is a lter in P if and only if G = G
0
G
1
for some lters G
0
and G
1
in
P
0
and P
1
, respectively.
Proof First, assume that G is a lter in P. Let G
0
and G
1
be the pro-
jections of G onto the rst and the second coordinates, respectively. Then
G G
0
G
1
.
To see the other inclusion take p
0
G
0
and q
1
G
1
. Then there exist
q
0
G
0
and p
1
G
1
such that 'p
0
, p
1
`, 'q
0
, q
1
` G. Let 'r
0
, r
1
` G be an
extension of 'p
0
, p
1
` and 'q
0
, q
1
`. Then 'r
0
, r
1
` 'p
0
, q
1
`, so 'p
0
, q
1
` G.
The proof that G
0
and G
1
are lters is left as an exercise. It is also
easy to see that G = G
0
G
1
is a lter, provided G
0
and G
1
are.
Proposition 9.4.1 tells us, in particular, that any lter in P
0
P
1
is a
product of lters in P
0
and in P
1
. This gives a better perspective on the
next theorem.
Theorem 9.4.2 (Product lemma) Let M be a countable transitive model
of ZFC, let P
0
, P
1
M be partially ordered sets, and let G
0
P
0
and
G
1
P
1
be lters. Then the following conditions are equivalent.
(i) G
0
G
1
is an M-generic lter in P
0
P
1
.
(ii) G
0
is an M-generic lter in P
0
and G
1
is an M[G
0
]-generic lter
in P
1
.
(iii) G
1
is an M-generic lter in P
1
and G
0
is an M[G
1
]-generic lter
in P
0
.
Moreover, if any of these conditions hold then M[G
0
G
1
] = M[G
0
][G
1
] =
M[G
1
][G
0
].
Proof Note rst that the natural isomorphism between P
0
P
1
and P
1
P
0
maps G
0
G
1
onto G
1
G
0
. Thus, by Lemma 9.3.7, condition (i) is
equivalent to the condition
(i
t
) G
1
G
0
is an M-generic lter in P
1
P
0
.
But the equivalence (i
t
)(iii) is obtained from (i)(ii) by exchanging in-
dices. So all equivalences follow from (i)(ii).

9.4 Product lemma and Cohen model 191
To prove (i)(ii), assume (i). To see that G
0
is M-generic in P
0
take a
dense subset D
0
M of P
0
. Then D
0
P
1
M is dense in P
0
P
1
. Thus,
by (i), (D
0
P
1
) (G
0
G
1
) = , and we conclude that D
0
G
0
= .
To see that G
1
is an M[G
0
]-generic lter in P
1
take a dense subset
D
1
M[G
0
] of P
1
. Let M be a P
0
-name such that D
1
= val
G0
() and
let p
0
G
0
be such that
p
0
| is dense in

P
1
.
Choose p
1
G
1
, dene
D = 'q
0
, q
1
`: q
0
p
0
& q
0
| q
1
M,
and note that D is dense below 'p
0
, p
1
`.
Indeed, let 'r
0
, r
1
` 'p
0
, p
1
` and let G
t
0
be an M-generic lter in P
0
containing r
0
. Then D
t
1
= val
G

0
() is dense in P
1
since r
0
p
0
, and p
0
forces it. Take q
1
r
1
with q
1
D
t
1
and nd a q
0
G
t
0
that forces it:
q
0
| q
1
.
Since q
0
and r
0
belong to the same lter G
t
0
they are compatible, and
taking their common extension, if necessary, we can assume that q
0
r
0
.
But then 'q
0
, q
1
` 'r
0
, r
1
` and 'q
0
, q
1
` D. Thus D is dense below
'p
0
, p
1
` G
0
G
1
.
So we can nd 'q
0
, q
1
` D(G
0
G
1
). But q
0
G
0
and q
0
| q
1
.
Thus q
1
val
G0
() = D
1
, so we have found q
1
D
1
G
1
.
To prove that (ii) implies (i) take a dense subset D of P
0
P
1
that
belongs to M. We will show that D (G
0
G
1
) = .
For this, dene
D
1
= p
1
P
1
: p
0
G
0
('p
0
, p
1
` D) M[G
0
]
and note that D
1
is dense in P
1
. Indeed, if r
1
P
1
then the set
D
0
= p
0
P
0
: p
1
r
1
('p
0
, p
1
` D) M
is dense in P
0
and so there exists a p
0
D
0
G
0
. Thus there exists a
p
1
r
1
with 'p
0
, p
1
` D and p
1
D
1
.
Now, by genericity, there exists a p
1
G
1
D
1
. So, by the denition
of D
1
, we can nd a p
0
G
0
with 'p
0
, p
1
` D. But then 'p
0
, p
1
`
D (G
0
G
1
), that is, D (G
0
G
1
) = .
Finally, if conditions (i)(iii) hold then the equation M[G
0
G
1
] =
M[G
0
][G
1
] holds by the minimality of dierent generic extensions.
To see that M[G
0
G
1
] M[G
0
][G
1
] note that G
0
M[G
0
]
M[G
0
][G
1
] and G
1
M[G
0
][G
1
]. So G
0
G
1
M[G
0
][G
1
]. Moreover,

192 9 Forcing
M M[G
1
] M[G
0
][G
1
]. Thus, by the minimality of M[G
0
G
1
], the
inclusion holds.
To see the reverse inclusion note that M M[G
0
G
1
] and G
0

M[G
0
G
1
]. So, by the minimality of M[G
0
], we have M[G
0
] M[G
0
G
1
].
But we have also G
1
M[G
0
G
1
]. Thus M[G
0
][G
1
] M[G
0
G
1
] by
the minimality of M[G
0
][G
1
].
The equation M[G
0
G
1
] = M[G
1
][G
0
] is proved similarly.
Next, we will see an application of Theorem 9.4.2 to the Cohen model,
that is, the model from Section 9.2.
First note that if P(Z) = 'Func

(Z, 2), ` and A, B is a partition of

Z then P(Z) is isomorphic to P(A) P(B) via the isomorphism i : P(Z)
P(A) P(B), i(s) = 's[
A
, s[
B
`. Also, if G is a lter in P(Z) and C Z
then
G
C
= s[
C
: s G = G P(C)
and i[G] = G
A
G
B
= (G P(A)) (G P(B)).
Now, if M is a countable transitive model of ZFC, Z, A, B M are
such that A, B is a partition of Z, and G is an M-generic lter in P(Z)
then, by Theorem 9.4.2, G
A
= G P(A) is an M-generic lter in P(A),
G
B
= G P(B) is an M[G
A
]-generic lter in P(B), and
M[G] = M[G
A
G
B
] = M[G
A
][G
B
].
This fact and the next lemma are powerful tools for proving dierent facts
about the Cohen model.
Lemma 9.4.3 Let M be a countable transitive model of ZFC, X, Y, Z
M, and G an M-generic lter in P(Z). If f : X Y is in M[G] then there
exists an A {
M
(Z) such that
M [= [A[ [X[ +
and f M[G P(A)].
Proof By Lemma 9.2.8 we can nd a P(Z)-name M such that f =
val
G
(), '

'x, y`, p`: x X & y Y & p P(Z), and

() A
x
= p P(Z): y Y ('

'x, y`, p` ) is an antichain for every

x X.
So every A
x
is countable in M, since P(Z) is ccc in M.
Let
A =

xX

pAx
dom(p).

9.4 Product lemma and Cohen model 193
Then A M, A Z, and [A[ [X[+ in M, since every set

pAx
dom(p)
is countable in M. To nish the proof it is enough to show that f
M[G P(A)].
But for every s = '

'x, y`, p` we have p A

x
P(A). Thus

= '

'x, y`

, p`: '

'x, y`, p` M
is a P(A)-name, where

'x, y`

is a standard P(A)-name for 'x, y` M. So

f = val
G
() = 'x, y`: p G ('

'x, y`, p` )
= 'x, y`: (p G P(A))('

'x, y`

, p`

)
= val
GP(A)
(

) M[G P(A)].
This nishes the proof.
Since every real number is identied with a function r : 2, we
obtain immediately the following corollary.
Corollary 9.4.4 Let M be a countable transitive model of ZFC, Z M,
and G be an M-generic lter in P(Z). If r R
M[G]
then, in M, there
exists a countable subset A of Z such that r M[G P(A)].
Our next goal is to prove that Martins axiom is false in the Cohen
model obtained by the forcing P(Z) with [Z[ >
1
. For this, rst note that
the formula r is a real number, which is identied with r is a function
from into 2, is absolute. So for every countable transitive model M of
ZFC we have
R
M
= R M and [0, 1]
M
= [0, 1] M.
In particular, if M[G] is any generic extension of M then
[0, 1] M, R M M[G],
since R M = R
M
M M[G] and [0, 1] M = [0, 1]
M
M M[G].
Lemma 9.4.5 Let M be a countable transitive model of ZFC, let Z
M be countable in M, and let G be an M-generic lter in P(Z). Then
M [0, 1] M[G] has Lebesgue measure zero in M[G].
Proof Since the forcings P(Z) and P() are isomorphic in M we can
assume, by Lemma 9.3.7, that Z = . Also, we will identify numbers from
[0, 1] with their binary expansions, that is, functions from into 2.

194 9 Forcing
For s P() let [s] = t [0, 1] : s t. Notice that the Lebesgue
measure of [s] is equal to l([s]) = 2
n
, where n = [s[. Now let r =

G 2

and for k < put

S
k
=

m>k
[r[
[2
m
,2
m+1
)
].
Notice that
l(S
k
)

m>k
l([r[
[2
m
,2
m+1
)
]) =

m>k
2
2
m

m>k
2
m
= 2
k
.
Thus the set S =

k<
S
k
has measure zero. It is enough to prove that
M [0, 1] S
k
for every k < .
So x k < and t M [0, 1] and dene
D =

s P(): m > k

s[
[2
m
,2
m+1
)
= t[
[2
m
,2
m+1
)

M.
It is easy to see that D is dense in P(). Thus, GD = , that is, there is
an m > k such that r[
[2
m
,2
m+1
)
= t[
[2
m
,2
m+1
)
. So t [r[
[2
m
,2
m+1
)
] S
k
.
Theorem 9.4.6 Let M be a countable transitive model of ZFC, Z M
uncountable in M, and G an M-generic lter in P(Z). Then, in M[G], the
interval [0, 1] is a union of
1
sets of Lebesgue measure zero.
Proof First assume that Z has cardinality
1
in M. Then the forcings
P(Z) and P(
1
) are isomorphic. So, by Lemma 9.3.7, we can assume
that Z =
1
.
Now, for every <
1
,
M[G] = M[G P((
1
` ) )][G P( )],
so, by Lemma 9.4.5 used with M[GP((
1
` ) )] as a ground model,
there exists a set S

M[G] of Lebesgue measure zero such that

M[G P((
1
` ) )] [0, 1] S

.
But, by Corollary 9.4.4, for every r M[G] [0, 1] there exists in M a
countable subset A of Z with r M[G P(A)]. Hence there exists a
<
1
such that A (
1
` ) and
r M[G P(A)] M[G P((
1
` ) )].
Therefore M[G] [0, 1]

<1
S

. This nishes the proof of the case

when M [= [Z[ =
1
.
For the general case, take Z
0
{
M
(Z) of cardinality
1
in M. Then
M[G] = M
1
[G
1
], where M
1
= M[GP(Z ` Z
0
)] and G
1
= GP(Z
0
) is an
M
1
-generic lter in P(Z
0
). The conclusion of the theorem follows from the
rst case.

9.4 Product lemma and Cohen model 195
Corollary 9.4.7 Let M be a countable transitive model of ZFC and as-
sume that Z M has cardinality at least
2
in M. If G is an M-generic
lter in P(Z) then, in M[G], the interval [0, 1] is a union of less than
continuum many sets of Lebesgue measure zero.
In particular, Martins axiom is false in M[G].
Proof Theorem 9.2.13 implies that c >
1
in M[G]. Hence, by Theo-
rem 9.4.6, we can conclude the rst part of the corollary. The second part
follows from the rst part and Theorem 8.2.7.
We will nish this section with an estimation of the size of the contin-
uum in the models obtained by ccc forcing extensions.
Theorem 9.4.8 Let M be a countable transitive model of ZFC, P M
be an innite ccc forcing in M, and G be an M-generic lter in P. If
= [P

[
M
then
M[G] [= c .
Proof Since, by Theorem 9.2.11, the models M and M[G] have the same
cardinals, we will not distinguish between them. Let T be the family of all
P-names '

'n, i`, p`: n & i 2 & p P such that

() A
n
= p P: i 2 ('

'n, i`, p` ) = is an antichain for every

n .
Then T M and every A
n
is at most countable.
By Lemma 9.2.8 for every f 2

M[G] we can nd a T such that

f = val
G
(). In particular, the image val
G
[T] of T under val
G
contains
2

M[G]. So c [T[ in M[G], and it is enough to prove that [T[ .

We will show this in M.
To see it, dene in M a map h: (2 P)

T such that for every

t (2 P)

if t(n, j) = 'i
n,j
, p
n,j
` then
h(t) = '

'n, i
n,j
`, p
n,j
`: n, j .
Note that h is onto T, since every set A
n
is at most countable. Thus
[T[

(2 P)

= .
This nishes the proof.
Corollary 9.4.9 Let M be a countable transitive model of ZFC, an in-
nite cardinal in M, and G an M-generic lter in P(). If = [

[
M
then
M[G] [= c = .
In particular, if CH holds in M and =
M
2
then c
M[G]
=
M[G]
2
.

196 9 Forcing
Proof First note that [P()[ = , and so = [

[
M
= [P()

[
M
. Thus,
by Theorem 9.4.8, c
M[G]
. The other inequality for =
2
follows from
Theorem 9.2.13. The general case is left as an exercise.
To see the additional part note that under CH the regularity of
2
implies

2
=

<2

=
2

1
=
2
c

=
2
c =
2

1
=
2
.
So, under CH, =
2
satises the assumptions of the main part of the
corollary.
EXERCISES
1 Complete the proof of Proposition 9.4.1.
2 Show that the set D from the proof of Lemma 9.4.5 is indeed dense
in P().
3 Complete the proof of Corollary 9.4.9 for arbitrary by showing that
c
M[G]
.
4 Let M be a countable transitive model of ZFC, P M a partially ordered
set, and G an M-generic lter in P. Prove that M[G] and M have the same
ordinal numbers.
9.5 Model for MA+CH
The goal of this section is to prove the consistency of MA+c =
2
. For
this, we will follow the idea mentioned at the beginning of the previous sec-
tion. We will start with a countable transitive model M of ZFC+2
1
=
2
,
which exists by Corollary 9.3.9, and nd its generic extension N via a ccc
forcing P

M for which there exists a sequence

M = M
0
M
1
M

= N
with =
M
2
=
N
2
and the following properties:
P1 M
+1
= M

[G

] for every < , where G

is an M

-generic lter in
some forcing P

.
P2 For every Y M and S Y with N [= [S[ <
2
there exists an <
such that S M

.

9.5 Model for MA+CH 197
P3 [(P

[
M

M
2
.
P4 For every forcing ', <` N such that
N [= ( is an ordinal number,
1
, and ', <` is ccc)
and every < there is a < , with > , such that P

= ', <`.
To argue that P1P4 imply N [= MA+c =
2
we also need the
following denitions and lemma. For an uncountable cardinal number
and a partially ordered set P we will write MA

(P) to denote the statement

MA

(P) For every family T of dense subsets of P such that [T[ < there
exists a T-generic lter F in P.
We will also use the symbol MA

for the statement that MA

(P) holds for

every ccc forcing P. Thus MA is equivalent to MA
c
.
Lemma 9.5.1 For an uncountable cardinal number the following state-
ments are equivalent:
(i) MA

;
(ii) MA

(P) holds for every ccc forcing P = 'P, ` with [P[ < ;
(iii) MA

(P) holds for every ccc forcing P = ', <`, where is a cardinal
number less than .
Before we prove Lemma 9.5.1 let us see how it can be used to conclude
N [= MA+c =
2
from P1P4.
First, we will prove that
N [= MA
2
. (9.7)
By Lemma 9.5.1 it is enough to show in N that MA
2
(P) holds for every
ccc forcing P = ', <` with being an ordinal number less than or equal
to
N
1
. For this, let ', <` N be ccc in N with
N
1
, and let
T = D

: <
N
1
N be a family of dense subsets of ', <`. Since
P

is ccc we have
N
1
=
M
1
, and < as well as

<
M
1
D

are
subsets of
M
1

M
1
M. Hence, by P2, there exists an < such that
', `, T M

. Moreover, by P4 there exists < < with P

= ', <`.
In particular, G

M
+1
is an M

-generic lter in P

= ', <`. But

T M

. So G

is also T-generic in ', <`. Condition (9.7) has

been proved.
Now MA
2
and c =
2
clearly imply MA. Thus, it is enough to argue
that N [= c =
2
. But the inequality c
N

N
2
follows from P3 and
Theorem 9.4.8. The other inequality follows from the fact that

198 9 Forcing
MA
2
G [2

]
1
f 2

` G. (9.8)
To see it, put P() = 'Func

(, 2), ` and for every g 2

dene D
g
=
s P(): s g. Then the sets D
g
are dense in P(). Let F be a
D
g
: g G-generic lter in P(). Then any extension of

F to a function
f : 2 will have the desired property. The details are left as an exercise.
Before we move to the technical aspects of the construction, it is worth-
while to reect for a moment on the idea behind the foregoing argument.
It is clearly of the transnite-induction nature and the argument is of a
diagonal character. Condition P1 represents an inductive step in which we
take care of one problem at a time. Condition P2 represents a kind of
closure argument. It tells us that the small objects from N can already
be found in the earlier steps of our construction and thus we will have a
chance to take care of them in the later part of the induction.
The construction of the forcing P

leading to a model satisfying P1

and P2 is of a very general nature and can be compared to the recursion
theorem. The specic aspects of our model are addressed mainly by con-
dition P4, with Lemma 9.5.1 allowing us to reduce the induction to one of
reasonably short length.
This construction closes, in a way, a full circle that we have made in this
course. We started with simple recursion proofs. Next, when the diculties
with the length of induction mounted, we started to use rened recursive
arguments by introducing additional axioms such as CH, , and MA. This
led us all the way to the forcing arguments of the previous sections, and
nally to nish here by coming back to a relatively simple transnite-
induction argument.
The preceding argument shows that in order to prove the consistency
of MA+c =
2
it is enough to prove Lemma 9.5.1 and nd a generic
extension N of M satisfying P1P4. We will start with Lemma 9.5.1.
Proof of Lemma 9.5.1 Clearly (i)(iii). Thus it is enough to show
(iii)(ii) and (ii)(i).
(ii)(i): Let P = 'P, ` be a ccc partially ordered set and let T be
a family of dense subsets of P such that [T[ < . Choose a function
f : P P P such that f(p, q) extends p and q provided p and q are
compatible in P, and for every D T pick f
D
: P D such that f
D
(p) p
for every p P. Let p
0
P and let Q be the smallest subset of P containing
p
0
and being closed under the operations f and f
D
for every D T. Thus
Q =

n<
Q
n
,

9.5 Model for MA+CH 199
where Q
0
= p
0
and Q
n+1
= Q
n
f[Q
n
Q
n
]

DT
f
D
[Q
n
] (compare
with Lemma 6.1.6). Then [Q[ [T[ + < .
Note that any p, q Q that are compatible in P are also compatible
in P
0
= 'Q, `, since Q is closed under the action of f. Thus P
0
is ccc
and, by (ii), MA

(P
0
) holds. But, by the closure of Q under the action
of f
D
, for every D T the set D Q is dense in P
0
. So there exists a
DQ: D T-generic lter F
0
in P
0
. To nish the proof it is enough to
notice that F = p P : q F
0
(q p) is a T-generic lter in P. The
details are left as an exercise.
(iii)(ii): The proof is very similar to that of Lemma 9.3.7. To see it,
let P = 'P, ` be a ccc partially ordered set such that = [P[ < and let
T be a family of dense subsets of P such that [T[ < . Choose a bijection
f between and P and dene a partial-order relation < on by putting
< f() f().
Notice that f is an order isomorphism between P
0
= ', <` and P =
'P, `. In particular, f
1
(D) is dense in P
0
for every D T. So T
0
=
f
1
(D): D T is a family of dense subsets of P
0
and [T
0
[ = [T[ < .
Hence, by (iii), there exists a T
0
-generic lter F
0
in P
0
. To nish the proof
it is enough to notice that f[F
0
] is a T-generic lter in P. The details are
left as an exercise.
To nd a model N satisfying P1P4 we rst need to come back to
the problem of expressing a model obtained by two consecutive generic
extensions as a single generic extension. More precisely, if M
1
= M[G
0
],
where G
0
is an M-generic lter in P
0
M, and M
2
= M[G
0
][G
1
], where
G
1
is an M[G
0
]-generic lter in P
1
M[G
0
], we would like to nd a P M
and an M-generic lter G in P such that M
2
= M[G]. The product lemma
(Theorem 9.4.2) gives a solution to this problem when P
1
M. Thus, we
will concentrate here on the case when P
1
M[G
0
] ` M. To dene such P
we need the following denition.
A pair ', ` M of P-names satisfying the condition
P | is a partial-order relation on (9.9)
will be called a good P-name for a partially ordered set. (Note that, formally
speaking, a good P-name is not a P-name.) If ', ` is a good P-name for
a partially ordered set we will often write for ', `.
The following lemma tells us that any forcing P
1
in M[G
0
] has a good
P
0
-name representing it. So it explains the restriction of our attention to
good names only. However, the lemma will not be used in what follows, so
we will leave it without a proof.

200 9 Forcing
Lemma 9.5.2 Let M be a countable transitive model of ZFC, P M a
partially ordered set, and G
0
an M-generic lter in P. If P
1
= 'P, `
M[G
0
] is a partially ordered set in M[G
0
] then there are P-names , M
satisfying (9.9) and such that 'P, ` = 'val
G0
(), val
G0
()`.
Now let P be a partially ordered set and ',

` be a good P-name for

a partially ordered set. Then their iteration P is dened by
P = 'p, `: p P & dom() & p | .
Moreover, on P we dene a binary relation _ by
'p, ` _ 'q, ` p q & p |

.
Notice that the relation _ is reexive and transitive on P . It is
reexive since for any 'p, ` P condition p forces and

to be
a partial-order relation on (since ',

` is good). The proof that _ is

transitive is left as an exercise.
Unfortunately, _ does not have to be antisymmetric on P . To solve
this problem, we notice that the relation on P dened by
'p, ` 'q, ` 'p, ` _ 'q, ` & 'q, ` _ 'p, `
is an equivalence relation on P (see Exercise 4 from Section 2.4). Thus,
in order to consider 'P, _` as a partially ordered set, we will identify the
elements of P that are -equivalent. (Formally, we should replace P
by the quotient class P /

and the relation _ by the quotient relation

dened as in part (b) of Exercise 4 in Section 2.4. However, this would
obscure the clarity of the simple-identication approach.) Upon such an
identication we will consider P as a partially ordered set.
Let us also notice that if the partially ordered sets P
0
and P
1
are in the
ground model M and and

are the standard P

0
-names for P
1
and its
order relation then ',

` is a good name for a partially ordered set and

P
0
is isomorphic to P
0
P
1
. (9.10)
Thus the process of iteration is truly a generalization of the product of
forcings as described in the previous section. To formulate an iteration
analog of the product lemma, we need one more denition.
Let M be a countable transitive model of ZFC, P
0
M be a partially
ordered set, and G
0
be an M-generic lter in P
0
. If is a good P
0
-name
for a partially ordered set and G
1
val
G0
() then we dene
G
0
G
1
= 'p, ` P
0
: p G
0
& val
G0
() G
1
.

9.5 Model for MA+CH 201
Theorem 9.5.3 Let M be a countable transitive model of ZFC and P
0

be an iteration of a partially ordered set P
0
M and a good P
0
-name
M for a partially ordered set. If G is an M-generic lter in P
0
and
G
0
= p P
0
: 'q, ` G (q p)
then G
0
is an M-generic lter in P
0
. Moreover, if
G
1
= val
G0
(): 'q, ` G
then G
1
is an M[G
0
]-generic lter in P
1
= val
G0
(), G = G
0
G
1
, and
M[G] = M[G
0
][G
1
].
Proof It is easy to see that G
0
is a lter in P
0
. To show that G
0
is M-
generic take a D
0
M that is dense in P
0
.Then D='p, ` P
0
: p D
0

is dense in P
0
, since 'p, ` _ 'q, ` and 'p, ` D for every 'q, ` P
0

and p D
0
with p q. Take 'p, ` G D. Then p G
0
D
0
. Thus G
0
is an M-generic lter in P
0
.
Next, we will show that G
1
is a lter in P
1
. So take val
G0
() G
1
with
'q, ` G witnessing it and let dom() be such that val
G0
() P
1
and val
G0
() val
G0
(). We will show that val
G0
() G
1
. For this, take
p
0
G
0
with p
0
| and 'p, ` G such that p p
0
. Pick 'r, ` G
with 'r, ` _ 'q, ` and 'r, ` _ 'p, `. Then r | and r | ,
since r p
0
. So r | , and 'r, ` _ 'r, `. Hence 'r, ` G and
val
G0
() G
1
.
To see that any two elements of G
1
have a common extension in G
1
take val
G0
() G
1
with 'q, ` G and val
G0
() G
1
with 'p, ` G.
Then there exists 'r, ` G such that 'r, ` _ 'q, ` and 'r, ` _ 'p, `. So
r G
0
and r | ( & ). Thus val
G0
() G
1
extends val
G0
()
and val
G0
().
For the proof that G
1
is M[G
0
]-generic in P
1
take a D
1
M[G
0
] that is
dense in P
1
and let be a P
0
-name such that D
1
= val
G0
(). Pick p G
0
that forces it, that is,
p | is a dense subset of ,
and dene
D = 'q, ` P
0
: q | . (9.11)
Now, if 'q, ` G is such that q p then it is easy to see that D is dense
below 'q, ` (the proof is left as an exercise). So there exists 'r, ` DG,
and val
G0
() D
1
G
1
.

202 9 Forcing
To argue that G = G
0
G
1
take 'p, ` G. Then p G
0
and val
G0
()
G
1
by the denitions of G
0
and G
1
. So G G
0
G
1
. To see the other
inclusion, take 'p, ` G
0
G
1
P
0
. Then p G
0
and val
G0
() G
1
.
So there exist 'q, ` G, with q p, and 'r, ` G. Let 's, ` G be a
common extension of 'q, ` and 'r, `. Then s q p and s | . So
's, ` _ 'p, ` and 'p, ` G.
Finally, G
0
, G
1
M[G], as they are dened using G. So M[G
0
] M[G]
by the minimality of M[G
0
], and M[G
0
][G
1
] M[G] by the minimality of
M[G
0
][G
1
]. Also, G = G
0
G
1
M[G
0
][G
1
]. So M[G] M[G
0
][G
1
].
Theorem 9.5.3 is an analog of the implication (i)(ii) from the product
lemma. The iteration analog of the reverse implication, stated next, is also
true. However, since we will not use this fact, we will leave its proof as an
exercise.
Theorem 9.5.4 Let M be a countable transitive model of ZFC, G
0
an M-
generic lter in P
0
M, and G
1
an M[G
0
]-generic lter in P
1
M[G
0
].
If M is a good P
0
-name representing P
1
then G
0
G
1
is an M-generic
lter in P
0
and M[G
0
][G
1
] = M[G
0
G
1
].
In the proof of the consistency of MA+CH we will be interested only
in the ccc forcings. The next lemma tells us that we will remain within
this class if the iteration concerns two ccc forcings.
Lemma 9.5.5 Let M be a countable transitive model of ZFC, P
0
M be
a ccc forcing in M, and ',

` be a good P
0
-name for a forcing such that
P
0
| ',

` is ccc. (9.12)
Then P
0
is ccc in M.
Proof To obtain a contradiction assume that there exists an uncountable
antichain A = 'p

`: <
1
M in P
0
. Let
= '

, p

`: <
1
.
Then M is a P
0
-name. Dene, in M,
D = p P
0
: <
1
(p |

)
and notice that D is dense in P
0
.
To see this, let q P
0
and let H be an M-generic lter containing q.
Then
S = val
H
() = <
1
: p

H.
Notice that for distinct , S the conditions val
H
(

) and val
H
(

) are
incompatible in val
H
().

9.5 Model for MA+CH 203
Indeed, if , S, = , and val
H
(

) and val
H
(

) are compatible in
val
H
() then there are r H extending p

and p

and dom() such

that r |

and r |

. But then 'r, ` would be a common

extension of 'p

` and 'p

`, contradicting our assumption that A is

an antichain.
Thus S is countable in M[H] since, by (9.12), val
H
() is ccc in M[H].
So there exists a <
1
containing S. Now, if p H is such that p q
and p | S

then p D. So D is dense.
Next consider the set
T = <
1
: p

P
0
(p

| sup =

)
and notice that A
1
= p

: T is an antichain in P
0
. Thus A
1
and T
are countable in M. Let <
1
be such that T . Then the set
D
t
= p P
0
: p |
is dense in P
0
, since it contains D.
Now, to obtain a contradiction, take an M-generic lter H in P
0
con-
taining p

. Then, by the denseness of D

t
, val
H
() . But, straight from
the denition of , val
H
()! This contradiction nishes the proof.
Now we turn our attention to the -stage iteration for an arbitrary
ordinal number . More precisely, we will describe a general method of
dening inductively a sequence P = ''P

`: ` of forcings such
that P
+1
can be identied with P

, where

is a good P

-name for
a partially ordered set. This, with help of Theorem 9.5.3, will take care of
the condition P1. The technical diculty in this construction is that we
have to know P

in order to talk about a good P

-name

. So, along with

the sequence P, we have to construct a sequence = ''

`: < `
of appropriate good P

-names. This obscures a simple idea that stands

behind the construction. Thus we will rst describe this construction with
the sequence replaced by a sequence ''Q

,
Q

`: < ` of forcings. In
this case for every we dene
P

S
Q

: S []
<

f[
S
: f

<
Q

& S []
<

(9.13)
and order it by
p

q dom(q) dom(p) &

dom(q)

p()
Q
q()

.
In particular,
P
0
= , P

<
P

for every limit ordinal 0 < , (9.14)

204 9 Forcing
and
P

and

(P

) for every < . (9.15)

Notice also that is a maximal element of every P

with and that

for every <
if r P

extends p P

then r[

also extends p (9.16)

and
if q P

and r P

extends q[

then r q[
(\)
P

extends q. (9.17)
Moreover, if P

+1
= P
+1
` P

= p P
+1
: dom(p) for < then
'P

+1
,
+1
` is isomorphic to 'P

, _` (9.18)
via the isomorphism p 'p[

, p()`. The easy proofs of (9.14)(9.18) are

left as an exercise.
For the general case, when P

-names

do not necessary represent the

forcings from the ground model, the sequence ''P

`: ` will be
dened by simultaneous induction with a sequence ''

`: < `. In
particular, we say that a sequence ''P

`: ` of partially ordered
sets is an -stage forcing iteration (with nite support) if there exists a
sequence ''

`: < ` such that every '

` is a good P

-name
for a partially ordered set, and if for every
P

S[]
<

S
dom(

): S (p[

| p()

(9.19)
is ordered by
p

q dom(q) dom(p) &

dom(q)

p[

| p()

q()

.
It is not dicult to see that the denition (9.19) gives the same notion of
forcing as (9.13) if every

is the standard P

-name

Q

of Q

. Moreover, it
is also easy to notice that the sequence dened by (9.19) has the properties
(9.14)(9.17) as well (the inductive proof of (9.17) is left as an exercise).
Moreover, for P

+1
= p P
+1
: dom(p) a counterpart of (9.18) says
that for every <
'P

+1
,
+1
` is isomorphic to 'P

, _`, (9.20)

9.5 Model for MA+CH 205
where an isomorphism is given by the map p 'p[

, p()`. To see this, it

is enough to notice that for every p, q P

+1
p
+1
q dom(q) dom(p) &

dom(q)

p[

| p()

q()

& p[

| p()

q()
p[

q[

& p[

| p()

q()
'p[

, p()` _ 'q[

, q()`.
The next theorem is an -stage-iteration version of Theorem 9.5.3. It
implies the condition P1, where M

is dened as M[G

] from the theorem.

Theorem 9.5.6 Let M be a countable transitive model of ZFC, let M
be an ordinal number, and let ''P

`: ` M be an -stage forcing
iteration based on the sequence ''

`: < ` of appropriate P

-names.
If G is an M-generic lter in P

then G

= G P

is an M-generic lter
in P

for every < . Moreover, for every < , if

G

= val
G

(p()): p G
+1
& dom(p)
then G

is an M[G

]-generic lter in a forcing P

= 'val
G

), val
G

)`
and M[G
+1
] = M[G

][G

].
Proof To see that G

is a lter in P

take p, q G

G. Then there
exists an r G extending p and q. But r[

belongs to G, since r r[

.
So r[

and, by (9.16), r[

extends p and q. Also, if q G

, p P

,
and q p then p P

, and so p G

. Thus G

is a lter in P

.
To prove that G

is M-generic in P

take a D M that is a dense

subset of P

and notice that

D

= p P

: p[

D M
is dense in P

. Indeed, if q P

then q[

and there exists an r D

with r

q[

. But, by (9.17), p = r q[
(\)
P

extends q and it is easy

to see that p D

. So D

is dense and there exists an r G D

. Hence
r[

D G

.
The proof of the additional part will be done in several steps. First
notice that
P

+1
is dense in P
+1
, (9.21)
since every p P
+1
` P

+1
= P

forces '

` to be a partially ordered
set, so there exists a P

-name such that p |

and p', ` P

+1
extends p.

206 9 Forcing
Next notice that
G

+1
= G P

+1
= G
+1
` P

is also an M-generic lter in P

+1
, since, by (9.21), every dense D M
subset of P

+1
is also dense in P
+1
and D G

+1
= D G
+1
= . Also,
G
+1
= G

+1
p[

: p G

+1
M[G

+1
]. So M[G
+1
] = M[G

+1
].
Now, if i is an isomorphism from (9.20) then
i[G

+1
] = 'p[

, p()`: p G

+1

is an M-generic lter in P

and M[G
+1
] = M[G

+1
] = M[i[G

+1
]].
So, by Theorem 9.5.3, i[G

+1
] = G

, G

is an M[G

]-generic lter in
P

, and M[G
+1
] = M[i[G

+1
]] = M[G

][G

].
The next theorem says that an -stage iteration (with nite support)
of ccc forcings is ccc. This will guarantee that the forcing P

from the
beginning of the section will be ccc. The theorem is a generalization of
Lemma 9.5.5.
Theorem 9.5.7 Let M be a countable transitive model of ZFC, let M
be an ordinal number, and let ''P

`: ` M be an -stage forcing
iteration based on the sequence ''

`: < ` of appropriate P

-names.
If
P

is ccc (9.22)
for every < , then P

is ccc in M.
Proof We will prove, by induction on , that P

is ccc in M. The
proof will be done in M.
Clearly P
0
= is ccc. So let 0 < be such that P

is ccc in M
for every < and let A = p

: <
1
P

. We will nd < <

1
such that p

and p

are compatible.
If is a successor ordinal, say = +1, take C
0
= <
1
: p

.
If [C
0
[ =
1
then A
0
= p

: C
0
P

, and P

is ccc by the inductive

assumption. Hence A
0
A contains two compatible elements with dierent
indices. So assume that [C
0
[ <
1
. Then C
1
=
1
` C
0
is uncountable and
A
1
= p

: C
1
P

+1
. But P

+1
is isomorphic to P

, which, by
Lemma 9.5.5, is ccc. Thus, once more, A
1
A contains two compatible
elements with dierent indices.
If is a limit ordinal, apply to the family dom(p

): <
1
the -
system lemma to nd an uncountable set C
1
and a nite set D
such that dom(p

) dom(p

) = D for all distinct , C. Let < be

such that D and consider p

: C P

. Since, by the inductive

9.5 Model for MA+CH 207
assumption, P

is ccc, we can nd p P

and distinct , C such that

p

and p

. To nish the proof it is enough to notice that

q = p p

[
(\)
p

[
(\)
(9.23)
belongs to P

and extends p

and p

; this follows from (9.17). The details

are left as an exercise.
We say that an -stage forcing iteration ''P

`: ` is a ccc
forcing iteration if it satises condition (9.22).
The next lemma is the last general fact concerning the -stage iteration
and it will imply property P2. It is an analog of Lemma 9.4.3.
Lemma 9.5.8 Let M be a countable transitive model of ZFC, let M
be an innite regular cardinal number, and let ''P

`: ` M be a
-stage ccc forcing iteration. If Y M, G is an M-generic lter in P

,
and S Y is such that M[G] [= [S[ < then there exists a < such
that S M[G

], where G

= G P

.
Proof By Theorem 9.5.7 the forcing P

is ccc. So the cardinals in M and

M[G] are the same, and we will not distinguish between them.
Let X = , where is the cardinality of S in M[G]. Then X M and
[X[ = < . Let f M[G] be a map from X onto S. We will nd a <
such that f M[G

].
By Lemma 9.2.8 we can nd a P

-name M such that f = val

G
(),
'

'x, y`, p`: x X & y Y & p P

, and
() A
x
=

p P

: y Y

'

'x, y`, p`

is an antichain for every

x X.
So every A
x
is countable in M, since P

is ccc in M. Let
A =

xX

pAx
dom(p).
Then A M, A , and [A[ [X[ + < in M, since every set

pAx
dom(p) is countable in M. Thus, by the regularity of , there exists
a < such that A . In particular,

xX
A
x
P

. So

'

'x, y`

, p`: '

'x, y`, p`

M
is a P

-name, where

'x, y`

is a standard P

-name for 'x, y` M, and

f = val
G
() =

'x, y`: p G

'

'x, y`, p`

'x, y`: (p G P

'

'x, y`

, p`

= val
G

() M[G

].

208 9 Forcing
So S = f[X] M[G

].
Now we are ready for the main theorem of this section.
Theorem 9.5.9 It is consistent with ZFC that c =
2
and MA holds.
Proof By the foregoing discussion, it is enough to construct a -stage
ccc nite-support forcing iteration ''P

`: `, with =
M
2
, that
satises P3 and P4. The construction will be done in a countable transitive
model M of ZFC in which 2
1
=
2
(such an M exists by Corollary 9.3.9).
In what follows we will use the notation introduced in the rst part of this
section. We will also use repeatedly the fact, following from (9.15), that
for every any P

-name is a P

-name too.
Let ''

`: < ` M be a one-to-one enumeration of

such that

for every < . By induction on < we will con-

struct a nite-support iteration ''P

`: `, together with a sequence

''

`: < ` of appropriate P

-names, and a sequence 'S

: < `
such that the following inductive conditions are satised for every < .
(I

) [P

[
1
.
(II

) S

= ''

`: < ` contains every pair ', ` with

1
and
'

', `, p`: , & p P

.
( will be used as a P

-name for a partial order on .)

(III

=

is a standard P

-name for =

and

= '

', `, p`

: q A (q

p)
'

', `, p`: < & p B,

where
A = q P

: q | '

` is a ccc forcing
and B = p P

: p is incompatible with every q A. In partic-

ular,
P

| '

` is a ccc forcing. (9.24)

To see that this inductive construction preserves (I

) consider two cases.

If 0 < < is a limit ordinal, then P

=

<
P

has cardinality
1
by the inductive assumption. If = + 1 is a successor ordinal, then
P

= P

, and P

is isomorphic to P

. But by (I

) and (III

) we
have [

[
1
. So [P

[ = [P

[ ([P

[ [

[)
1
as well.

9.5 Model for MA+CH 209
To see that the sequence S

in (II

) can be found note that there are only

2
pairs ', ` to list since, by (I

), the set '

', `, p`: , & p P

has cardinality
1
and so it has at most
2
subsets, as 2
1
=
2
in M.
To nish the construction it is enough to argue for (9.24). For this,
notice rst that
A B is dense in P

(9.25)
and
p | (

) for every p A. (9.26)

To see (9.26) take p
0
A and an M-generic lter H in P

containing
p
0
. Then
val
H
(

) = ', `: '

', `, p`

& p H
= ', `: '

', `, p`

& q A (q

p) & p H
= ', `: '

', `, p`

& q A (q

p) & p H
= ', `: '

', `, p`

& p H
= val
H
(

),
proving (9.26).
To argue for (9.24) let H be an M-generic lter in P

. We have to show
that 'val
H
(

), val
H
(

)` = ', val
H
(

)` is a ccc partially ordered set

in M[H]. But, by (9.25), either AH = , in which case this is true since
an element of A H forces '

` = '

` to be a ccc forcing, or
B H = , in which case val
H
(

) is just a standard well ordering of ,

so it is obviously ccc. This nishes the inductive construction.
Now, to see that P3 is satised note that P

<
P

has cardinality

2
. Thus [(P

[ (
2
)
1
= 2
1
=
2
.
To see that P4 holds let G be an M-generic lter in P

, let N = M[G],
and let ', <` N be such that
N [=
1
and ', <` is ccc.
Choose an
0
< . By Theorem 9.5.6 it is enough to nd
0
< < such
that P

= 'val
G

), val
G

)` = ', <`. But, by Lemma 9.5.8, there is

0
< < such that ', <` M[G

]. Take a P

-name

M such that
val
G

) is equal to < and let

= '

', `, p`: , & p P

& p |

', `

M.
Note that
val
G

() = val
G

)

210 9 Forcing
since for every ', `
', ` val
G

() p G

(p |

', `

) ', ` val
G

).
Next, choose < such that '

` = ', ` and pick < with

'

` = ', `. Then
0
< =

and

= , so val
G

) =
. Also,

= and
val
G

() = val
G

() = val
G

) = (<) M[G

] M[G

]
since G

= G

.
Thus it suces to argue that
val
G

) = val
G

(). (9.27)
But ', <` is a partially ordered set in M[G

], since it is a partially ordered

set in N and this is an absolute property. Also, ', <` is ccc in M[G

],
because every antichain A M[G

] is also an antichain in N and

= [A[
N
= [A[
M
= [A[
M[G

]
,
since ', <` is ccc in N and both extensions of M preserve the cardinals.
Thus there is a p G

that forces this to be true for '

, ` = '

`. In
particular, G

A = so (9.26) and

= imply (9.27). This nishes

the proof.
EXERCISES
1 Complete the details of the proof of (9.8).
2 Complete the details of the proof of Lemma 9.5.1.
3 Prove that the relation _ dened on an iteration P is transitive.
4 Prove (9.10).
5 Complete the proof of Theorem 9.5.3 by showing that the set D from
(9.11) is indeed dense below 'q, `.
6 Prove Theorem 9.5.4.
7 Prove properties (9.14)(9.18).
8 Prove (9.17) for the forcing dened by (9.19).
9 Complete the proof of Theorem 9.5.7 by showing that the condition q
dened by (9.23) has the desired properties.

Appendix A
Axioms of set theory
Axiom 0 (Set Existence) There exists a set:
x(x = x).
Axiom 1 (Extensionality) If x and y have the same elements, then x is
equal to y:
xy [z(z xz y)x = y].
Axiom 2 (Comprehension scheme or schema of separation) For every for-
mula (s, t) with free variables s and t, for every x, and for every
parameter p there exists a set y = u x: (u, p) that contains all
those u x that have property :
xpy [u(u y(u x & (u, p)))].
Axiom 3 (Pairing) For any a and b there exists a set x that contains a
and b:
abx(a x & b x).
Axiom 4 (Union) For every family T there exists a set U containing the
union

T of all elements of T:
TUY x[(x Y & Y T)x U].
Axiom 5 (Power set) For every set X there exists a set P containing the
set {(X) (the power set) of all subsets of X:
XPz [z Xz P].
211

212 Appendix A
To make the statement of the next axiom more readable we introduce
the following abbreviation. We say that y is a successor of x and write
y = S(x) if S(x) = x x, that is,
z[z y (z x z = x)].
Axiom 6 (Innity) (Zermelo 1908) There exists an innite set (of some
special form):
x[z(z = z x) & y xz(z = S(y)z x)].
Axiom 7 (Replacement scheme) (Fraenkel 1922; Skolem 1922) For every
formula (s, t, U, w) with free variables s, t, U, and w, every set A,
and every parameter p if (s, t, A, p) denes a function F on A by
F(x) = y (x, y, A, p), then there exists a set Y containing the
range F[A] = F(x): x A of the function F:
Ap [x A!y(x, y, A, p)Y x Ay Y (x, y, A, p)],
where the quantier !x(x) is an abbreviation for there exists pre-
cisely one x satisfying , that is, is equivalent to the formula
x(x) & xy((x) & (y)x = y).
Axiom 8 (Foundation or regularity) (Skolem 1922; von Neumann 1925)
Every nonempty set has an -minimal element:
x[y(y x)y(y x & z(z x & z y))].
Axiom 9 (Choice) (Levi 1902; Zermelo 1904) For every family T of dis-
joint nonempty sets there exists a selector, that is, a set S that
intersects every x T in precisely one point:
T [x T(x = ) & x Ty T(x = y x y = )]
Sx T!z(z S & z x),
where x y = is an abbreviation for
z(z x & z y).
Using the comprehensive schema for Axioms 0, 3, 4, and 5 we may
easily obtain the following strengthening of them, which is often used as
the original axioms.

Axioms of set theory 213
Axiom 0
t
(Empty set) There exists the empty set :
xy(y x).
Axiom 3
t
(Pairing) For any a and b there exists a set x that contains
precisely a and b.
Axiom 4
t
(Union) (Cantor 1899; Zermelo 1908) For every family T there
exists a set U =

T, the union of all subsets of T.

Axiom 5
t
(Power set) (Zermelo 1908) For every set x there exists a set
Y = {(x), the power set of x, that is, the set of all subsets of x.
The system of Axioms 08 is usually called ZermeloFraenkel set theory
and is abbreviated by ZF. The system of Axioms 09 is usually denoted by
ZFC. Thus, ZFC is the same as ZF+AC, where AC stands for the axiom
of choice.
Historical Remark (Levy 1979) The rst similar system of axioms was
introduced by Zermelo. However, he did not have the axioms of founda-
tion and replacement. Informal versions of the axiom of replacement were
suggested by Cantor (1899) and Mirimano (1917). Formal versions were
introduced by Fraenkel (1922) and Skolem (1922). The axiom of foundation
was added by Skolem (1922) and von Neumann (1925).
Notice that Axioms 2 and 7 are in fact the schemas for innitely many
axioms, one for each formula . Thus theory ZFC has, in fact, innitely
many axioms. Axiom 1 of extensionality is the most fundamental one.
Axioms 0 (or 0
t
) of set existence (of empty set) and 6 of innity are ex-
istence axioms that postulate the existence of some sets. It is obvious
that Axiom 0 follows from Axiom 6. Axioms 2 of comprehension and 7
of replacement are schemas for innitely many axioms. Axioms 3 (or 3
t
)
of pairing, 4 (or 4
t
) of union, and 5 (or 5
t
) of power set are conditional
existence axioms. The existing sets postulated by 3
t
, 4
t
, and 5
t
are unique.
It is not dicult to see that Axiom 3 of pairing follows from the others.
The axiom of choice AC and the axiom of foundation also have the
same conditional existence character. However, the sets existing by them
do not have to be unique. Moreover, the axiom of foundation has a very
set-theoretic meaning and is seldom used outside abstract set theory or
logic. It lets us build hierarchical models of set theory. During this
course we very seldom make use of it. The axiom of choice, on the other
hand, is one of the most important tools in this course. It is true that
its nonconstructive character caused, in the past, some mathematicians to

214 Appendix A
reject it (for example, Borel and Lebesgue). However, this discussion has
been for the most part resolved today in favor of accepting this axiom.
It follows from the foregoing discussion that we can remove Axioms 0
and 3 from 08 and still have the same theory ZF.

Appendix B
Comments on the forcing
method
This appendix contains some comments and explanations for Section 9.1.
It also includes some missing proofs of the results stated there.
We start by explaining a few more terms. In Section 9.1 we said that a
formula is a consequence of a theory T, T , if there is a formal proof
of from T. By a formal proof in this statement we mean a formalization
of what we really do in proving theorems. Thus a proof of from T is a
nite sequence
0
, . . . ,
n
of formulas such that
n
= and a formula
k
can appear in the sequence only because of one of following two reasons:

k
is an axiom, that is, it belongs to T or is a logic axiom; or

k
is obtained from some
i
and
j
(i, j < k) by a rule of detachment
(also called a modus ponens rule), that is, if there exist i, j < k such that

j
has the form
i

k
.
Since any formal proof is a nite sequence of formulas, it can contain only
nitely many sentences from T. Thus, if T then there is a nite sub-
theory T
0
T such that T
0
. In particular, if theory T is inconsistent,
then there is a nite subtheory T
0
of T that is also inconsistent. So theory
T is consistent if and only if every one of its nite subtheories is consistent.
Note also that if
0
, . . . ,
n
is a proof of =
n
in theory T then
(
0
), . . . , (
n
) is a proof of in T, where formula (
i

k
)
is identied with
i
(
k
). Thus, if is a consequence of a nite
theory T
0
=
0
, . . . ,
n
then (
0
( (
n
) )), that is,
(
0
& &
n
) is a consequence of the axioms of logic.
To argue for the forcing principle we will need the following theorem.
215

216 Appendix B
Theorem B.1 Let S and T be two theories and assume that for every -
nite subtheory S
0
of S we can prove in theory T that there exists a nonempty
transitive set M that is a model for S
0
. Then Con(T) implies Con(S).
Proof Let S and T be as in the theorem and assume that S is inconsistent.
It is enough to show that this implies the inconsistency of T.
Since S is inconsistent, there is a nite subtheory S
0
=
1
, . . . ,
n

of S such that S
0
is inconsistent, that is, such that S
0
& for some
sentence . In particular, (
1
& &
n
)(&).
But, by our assumption, in theory T we can prove the existence of a
nonempty transitive set M that is a model for S
0
, that is, such that
M
holds for every from S
0
. In particular,
T M (
M
1
& &
M
n
).
Now, it is enough to notice that the rules of formal deduction are set up
in such a way that if
0
, . . . ,
n
is a formal proof of & from S
0
then

M
0
, . . . ,
M
n
is a formal proof of
M
&
M
from S
M
0
=
M
1
, . . . ,
M
n
. In
particular, the implication
(
M
1
& &
M
n
)(
M
&
M
)
is true for every M. So we have proved in T that there is an M such that

M
1
& &
M
n
, while (
M
1
& &
M
n
)(
M
&
M
) is true for every M.
So M (
M
&
M
) is a consequence of T, that is, T is inconsistent.
Now assume that we have proved the following condition (F) from the
forcing principle:
(F) Every CTM M of ZFC can be extended to a CTM N of ZFC+.
If we could prove in ZFC that
M (M is a CTM for ZFC)
then we would conclude in ZFC that there exists a model N of ZFC+,
and this, by Theorem B.1, implies that Con(ZFC)Con(ZFC+). Un-
fortunately, the existence of a countable transitive model of ZFC cannot
be proved from ZFC axioms (this follows from Theorem 1.1.1). Thus we
need the following more rened argument.
A closer look at the forcing method shows that it lets us extend any
countable transitive set M to another countable transitive set N = M[G].
Then Theorem 9.2.2 asserts that if M satises ZFC then so does M[G].
However, an examination of this proof (which is not included in this text)
shows that checking whether M[G] has a given property is of nitistic
character in the sense that in the proof of the implication
if M is a model for ZFC then M[G] is a model for (B.1)

Comments on the forcing method 217
we use the knowledge that M is a model for a given axiom only for
nitely many axioms of ZFC. In particular, for any sentence for which
(B.1) holds there is a nite subtheory T

of ZFC such that the implication

if M is a model for T

then M[G] is a model for (B.2)

has the same proof as (B.1). To show that such a nice proof of (F) implies
Con(ZFC)Con(ZFC+) we need one more theorem, which will be left
without proof.
Theorem B.2 For every nite subtheory T of ZFC it is provable in ZFC
that there exists a countable transitive model M of T.
Now, to argue for the forcing principle assume, to obtain a contradic-
tion, that (F) holds and ZFC+ is inconsistent. Then there exists a
nite subtheory S of ZFC+ that is also inconsistent. For every from
S let T

be a nite subset of ZFC for which (B.2) holds and let T be the
union of all T

for from S. By Theorem B.2 there is a CTM M for T,

and using (B.2) we conclude that there exists a countable transitive model
N = M[G] for S. Since all of this was proved in ZFC, we can use Theo-
rem B.1 to deduce that Con(ZFC) implies Con(S). But S is inconsistent,
so it can only happen if ZFC is inconsistent. This nishes the argument
for the forcing principle.
We will nish this appendix with the list of
0
-formulas that constitute
the
Proof of Lemma 9.2.5
(0) Formula
0
(x, y) representing x y:
w x(w y).
(1) Formula
1
(x, y) representing y =

x:
w yz x(w z) & z xw z(w y).
(2) Formula
2
(x, y) representing y =

x:
w yz x(w z) & z xw z[u x(w u)(w y)].
(3) Formula
3
(x, y, z) representing z = x y:
w z(w x w y) & w x(w z) & w y(w z).

218 Appendix B
(4) Formula
4
(x, y, z) representing z = x y:
w z(w x & w y) & w x[w yw z].
(5) Formula
5
(x, y, z) representing z = x ` y:
w z(w x & w y) & w x(w yw z).
(6) Formula
6
(x, y, z) representing z = x, y:
x z & y z & t z(t = x t = y).
(7) Formula
7
(z) representing z is an unordered pair:
x zy z
6
(x, y, z).
(8) Formula
8
(x, y, z) representing z = 'x, y` = x, x, y:
u zw z[
6
(x, x, u) &
6
(x, y, w) &
6
(u, w, z)].
(9) Formula
9
(z) representing z is an ordered pair:
w zx wy w
8
(x, y, z).
(10) Formula
10
(x, y, z) representing z = x y:
w zs xt y
8
(s, t, w) & s xt yw z
8
(s, t, w).
(11) Formula
11
(r) representing r is a binary relation:
w r
9
(w).
(12) Formula
12
(d, r) representing d is the domain of a binary rela-
tion r:
x dz rw zy w
8
(x, y, z)
& z rw zx wy w[
8
(x, y, z)x d].
(13) Formula
13
(R, r) representing R is the range of a binary relation r:
y Rz rw zx w
8
(x, y, z)
& z rw zx wy w[
8
(x, y, z)y R].

Comments on the forcing method 219
(14) Formula
14
(f) representing f is a function:

11
(f) & p fq fu pw qa ub ud w
[[
8
(a, b, p) &
8
(a, d, q)]b = d].
(15) Formula
15
(f) representing function f is injective:

14
(f) & p fq fu pw qa ub uc wd w
[[(
8
(a, b, p) &
8
(c, d, q) & b = d)]a = c].
(16) Formula
16
(, P) representing is a partial-order relation on P:

11
() &
12
(P, ) &
13
(P, ) & x Pw
8
(x, x, w)
& x Py Pv w [(
8
(x, y, v) &
8
(y, x, w)) x = y]
& x Py Pz Pv w
[(
8
(x, y, v) &
8
(y, z, w)) u
8
(x, z, u)].
(17) Formula
17
(D, , P) representing D is a dense subset of the par-
tially ordered set 'P, `:

16
(, P) &
0
(D, P) & x Pd Dw
8
(d, x, w).
(18) Formula
18
(A, , P) representing A is an antichain in the partially
ordered set 'P, `:

16
(, P) &
0
(A, P) & a Ab Ax Pu w
[(
8
(x, a, u) &
8
(x, b, u)) a = b].
(19) Formula
19
(x) representing set x is transitive:
y xz y(z x).
(20) Formula
20
() representing is an ordinal number:

19
() & ( = ).
(21) Formula
21
() representing is a limit ordinal number:

20
() & ( ).
(22) Formula
22
() representing = :

21
() & [
21
() x (x = x)].
(23) Formula
23
() representing is a nite ordinal number:

20
() &
22
() & (
22
()).
(24) Formula
23
() representing is a successor ordinal number:

20
() &
21
().

Appendix C
Notation
x y x is an element of y, 6.
the negation of formula , 6.
& the conjunction of formulas and , 6.
the disjunction of formulas and , 6.
the implication, 6.
the equivalence of formulas and , 6.
x the existential quantier, 6.
x the universal quantier, 6.
x A a bounded existential quantier, 6.
x A a bounded universal quantier, 6.
x y x is a subset of y, 6.
the empty set, 7.

T the union of a family T of sets, 8.
{(X) the power set of a set X, 8.
x y the union of sets x and y, 8.
x ` y the dierence of sets x and y, 8.

T the intersection of a family T of sets, 8.
x y the intersection of sets x and y, 9.
220

Notation 221
xy the symmetric dierence of sets x and y, 9.
'a, b` the ordered pair a, a, b, 9.
'a
1
, a
2
, . . . , a
n1
, a
n
` the ordered n-tuple, 10.
X Y the Cartesian product of sets X and Y , 10.
S(x) the successor of x: x x, 10.
dom(R) the domain of a relation (or function) R, 12.
range(R) the range of a relation (or function) R, 12.
R
1
the inverse of a relation (or function) R, 13.
S R the composition of the relations (or functions) R and S, 13.
f : X Y a function from a set X into a set Y , 16.
Y
X
the class of all function from a set X into a set Y , 16.
f[A] the image of a set A with respect to a function f, 16.
f
1
(B) the preimage of a set B with respect to a function f, 16.
f[
A
the restriction of a function f to a set A, 18.

tT
F
t
the union of an indexed family F
t

tT
, 19.

tT
F
t
the intersection of an indexed family F
t

tT
, 19.

tT
F
t
the Cartesian product of an indexed family F
t

tT
, 20.
N the set of natural numbers, 26.
the set of natural numbers, 27;
the order type of an innite strictly increasing sequence, 39;
the rst innite ordinal number, 44.
Z the set of integers, 30.
the set of rational numbers, 30.
R the set of real numbers, 31.
B(p, ) the open ball in R
n
with center p and radius , 32.
int(S) the interior of a set S in R
n
, 33.
cl(S) the closure of a set S in R
n
, 33.

the order type of an innite strictly decreasing sequence, 39.

222 Appendix C
O(x
0
) the initial segment x X: x < x
0
, 40.
Otp(W) the order type of a well-ordered set W, 47.
[A[ the cardinality of a set A, 61.
2

the cardinality of the set {(), 65.

A
the characteristic function of a set A, 65.

+
the cardinal successor of a cardinal , 65.

the th innite cardinal number, 66.

aleph-alpha: the th innite cardinal number, 66.

Z

bet-alpha, 66.
c continuum: the cardinality of the set {(), 66.
the cardinal sum of cardinals and , 68.
the cardinal product of cardinals and , 68.

cardinal exponentiation: the cardinality of the set

, 68.
A
<
the set of all nite sequences with values in A:

n<
A
n
, 71.
[X]

the family of all subsets of X of cardinality , 72.

[X]
<
the family of all subsets of X of cardinality < , 72.
[X]

the family of all subsets of X of cardinality , 72.

((R) the family of all continuous functions f : R R, 73.
cf() the conality of an ordinal number , 74.
cl
J
(Z) closure of a set Z under the action of T, 86.
F
t
limit points of F R
n
, 92.
[T] the smallest -algebra containing T, 94.
Bor the -algebra of Borel sets, 94.

0

subsets of Bor, 94.

0

subsets of Bor, 94.

^T the ideal of nowhere-dense sets, 98.
the -ideal of meager (rst-category) sets, 98.
^ the -ideal of Lebesgue measure-zero (null) subsets of R
n
, 99.

Notation 223
LIN
Q
(B) linear subspace of R over generated by B R, 111.
Func(X, Y ) all partial functions from X into Y , 131.
Func

(X, Y ) all partial functions from X into Y with the domain of

cardinality < , 131.
<

a relation on

, 135.

a linear-order relation on

, 139.
=

an equivalence relation on

, 139.
P = P
0
P
1
the product of partially ordered sets P
0
and P
1
, 149, 190.
trcl(x) the transitive closure of x, 165.
R() 165.
rank(x) the rank of x, 166.
T sentence is a consequence of a theory T, 167.
Con(T) theory T is consistent, 167.

M
the relativization of a formula to a set M, 168.
M [= a formula is true in M, 168.
M[G] a generic extension of model M, 170, 175.

M
1

1
in model M, 175.
c
M
c in model M, 175.
val
G
() valuation of a P-name , 175.
standard P-name for a generic lter in P, 176.
p |
P,M
p forces , 176.
| forcing relation, 176.
P | every p P forces , 177.
|

P
(x
1
, . . . , x
n
) formula for p forces (x
1
, . . . , x
n
), 177.
|

177.
P(Z) Cohen forcing on a set Z, 192.
P iteration of P and a good P-name , 200.
G
0
G
1
200.

References
Cantor, G. 1899. Correspondence between Cantor and Dedekind (in Ger-
man). In Cantor (1932: 44351).
Cantor, G. 1932. Gesammelte Abhandlungen: Mathematischen und philoso-
phischen Inhalts. Berlin: Springer.
Ciesielski, K. 19956. Uniformly antisymmetric functions and K
5
. Real
Anal. Exchange 21: 14753.
Ciesielski, K. 1996. Sum and dierence free partitions of vector spaces.
Colloq. Math. 71: 26371.
Ciesielski, K., and Larson, L. 19934. Uniformly antisymmetric functions.
Real Anal. Exchange 19: 22635.
Ciesielski, K., and Miller, A. W. 19945. Cardinal invariants concerning
functions whose sum is almost continuous. Real Anal. Exchange 20:
65772.
Erd os, P. 1969. Problems and results in chromatic graph theory. In Proof
techniques in graph theory, ed. F. Harary, 2735. New York: Aca-
demic Press.
Erd os, P. 19789. Measure theoretic, combinatorial and number theoretic
problems concerning point sets in Euclidean space. Real Anal. Ex-
change 4: 11338.
Erd os, P., and Kakutani, S. 1943. On non-denumerable graphs. Bull Amer.
Math. Soc. 49: 45761.
Fraenkel, A. A. 1922. Zu den Grundlagen der Cantor-Zermeloschen Men-
genlehre. Math. Ann. 86: 2307.
Frege, G. [1893] 1962. Grundgesetze der Arithmetik. Vol 1. Reprint.
Hildesheim: Olms.
225

226 References
Freiling, C. 1986. Axioms of symmetry: throwing darts at the real number
line. J. Symb. Logic 51: 190200.
Freiling, C. 198990. A converse to a theorem of Sierpi nski on almost
symmetric sets. Real Anal. Exchange 15: 7607.
Friedman, H. 1980. A consistent Fubini-Tonelli theorem for non-measurable
functions. Illinois J. Math. 24: 3905.
Galvin, F. 1980. Chain conditions and products. Fund. Math. 108: 3348.
Kirchheim, B., and Natkaniec, T. 19901. On universally bad Darboux
functions. Real Anal. Exchange 16: 4816.
Komj ath, P., and Shelah, S. 19934. On uniformly antisymmetric functions.
Real Anal. Exchange 19: 21825.
Kunen, K. 1980. Set Theory. Amsterdam: North-Holland.
Levi, B. 1902. Intorno alla teoria degli aggregati. Rendic., 2d ser. 35:
8638.
Levy, A. 1979. Basic Set Theory. New York: Springer-Verlag.
Mazurkiewicz, S. 1914. Sur un ensemble plan (in Polish). Comptes Rendus
Sci. et Lettres de Varsovie 7: 3823. French translation reprinted
in Travaux de Topologie et ses Applications, 467. Warsaw: Polish
Scientic Publishers PWN, 1969.
Mirimano, D. 1917. Les antinomies de Russell et de Burali-Forti et le
probl`eme fondamental de la theorie des ensembles. LEnseignement
Math. 19: 3752.
Royden, H. L. 1988. Real Analysis. 3d ed. New York: Macmillan.
Russell, B. 1903. The Principles of Mathematics. Vol. 1. Cambridge:
Cambridge University Press.
Sierpi nski, W. 1919. Sur un theor`em equivalent ` a lhypoth`ese du continu.
Bull Int. Acad. Sci. Cracovie A: 13. Reprinted in Oeuvres Choisies,
vol. 2, 2724. Warsaw: Polish Scientic Publishers PWN, 1974.
Sierpi nski, W. 1920. Sur les rapports entre lexistence des integrales

1
0
f(x, y)dx,

1
0
f(x, y)dy et

1
0
dx

1
0
f(x, y)dy. Fund. Math. 1:
1427. Reprinted in Oeuvres Choisies, vol. 2, 3415. Warsaw: Polish
Scientic Publishers PWN, 1974.

References 227
Sierpi nski, W. 1936. Sur une fonction non measurable partout presque
symetrique. Acta Litt. Scient (Szeged) 8: 16. Reprinted in Oeuvres
Choisies, vol. 3, 27781. Warsaw: Polish Scientic Publishers PWN,
1974.
Skolem, D. 1922. Einige Bemerkungen zur axiomatischen Begr undung der
Mengenlehre. In Proceedings of the 5th Scandinavian Mathemati-
cians Congress in Helsinki, 1922, 21732.
Stepr ans, J. 1993. A very discontinuous Borel function. J. Symbolic Logic
58: 126883.
von Neumann, J. 1925. Zur Einf uhrung der transniten Zahlen. Acta Sci.
Math. (Szeged) 1: 199208.
von Neumann, J. 1929. Eine Widerspruchsfreiheitsfrage in der axiomatis-
chen Mengenlehre. Crelle J. 160: 227.
Zermelo, E. 1904. Beweis, da jede Menge wohlgeordnet werden kann.
Math. Ann. 59: 51416.
Zermelo, E. 1908. Untersuchungen uber die Grundlagen der Mengenlehre,
I. Math. Ann. 65: 26181.

Index
Abelian group, 55
absolute property, 172, 182
AC, 15
additive function, 111, 117
Jones space of, 114
Alexander subbase theorem, 58
algebraic numbers, 73
algebraically independent subset,
60
almost contained

, 153
almost-disjoint sets, 153
antichain, 140, 174, 177
maximal, 140
antilexicographic-order relation, 43
antisymmetric function, 119
antisymmetric relation, 21
arithmetic
of cardinal numbers, 68
of ordinal numbers, 48
Peano axioms of, 26
Aronszajn tree, 163
axiom of choice, 4, 15, 18, 20, 52,
54, 101
axiom of empty set, 7
axiom of extensionality, 6
axiom of foundation, 37
axiom of innity, 10, 26, 182
axiom of pairing, 8
axiom of power set, 8
axiom of regularity, 37
axiom of set existence, 6
axiom of union, 8
axiom schema of comprehension,
3
axiom schema of separation, 7
Baire category theorem, 33, 98
Baire class-one function, 73
Baire property, 98
Baire-measurable functions, 104
base, 58
for R
n
, 89
basis
Hamel, 111, 115
with Baire property, 117
without Baire property, 113
measurable, 117
nonmeasurable, 113
of a linear space, 55
transcendental, 60
Bernstein set, 1035, 112, 117, 126
bijection, 16
binary relation, 12, 173
B-measurable functions, 104
Borel functions, 104
Borel sets, 89, 97
Borel -algebra Bor, 94
bound variables in a formula, 166
bounded quantier, 168
bounded subset of R
n
, 33
branch of a tree, 163
229

230 Index
Cantor set, 71, 93, 111, 117
CantorBendixson theorem, 92
Cantors theorem, 64
cardinal number, 62
collapsing of, 180, 185
preservation of, 180, 188
regular, 75, 144
successor of, 65
cardinal numbers
arithmetic of, 68
exponentiation of, 68
product of, 68
sum of, 68
union of, 65
cardinality
continuum c, 66, 76, 195
of a set, 61
Cartesian product, 10, 20
Cauchy sequence, 14, 33
ccc forcing, 140, 149, 150, 17981,
195
ccc iteration, 202, 206
ccc linearly ordered set, 154, 162
Ceders theorem, 89
CH, 66
chain in a partially ordered set,
52
characteristic function, 65
choice function, 18
CiesielskiLarson theorem, 119, 121
Ciesielskis theorem, 124
class
equivalence, 14
proper, of sets, 7, 44, 168
quotient, 14
of sets, 7
closed set under the action of T,
85, 97
closed subset of R
n
, 32, 89
closed unbounded subsets of
1
,
157
diagonal intersection of, 161
closure
cl(S) of a set in R
n
, 33
cl
J
(Z) of a set under the ac-
tion of T, 85
transitive, trcl(x), 165
conality cf() of an ordinal num-
ber, 74
Cohen model, 182, 189
Cohen real numbers, 182
collapsing of cardinal numbers, 180,
185
compact subset of R
n
, 33
compact topological space, 58
comparable elements in a partially
ordered set, 140
compatible elements in a partially
ordered set, 140
compatible set, 147
complete Boolean algebra, 140
complete linearly ordered set, 135,
154
complete linear-order relation, 135,
154
complete set of axioms, 4
composition of relations, 13
comprehension, axiom schema of,
3
comprehension scheme, 7
condition, 132
ccc, 140
stronger, 132
conditions
comparable, 140
compatible, 140
incompatible, 140
connected relation, 23
connected subset of R
n
, 33, 114
consistent set of axioms, 4
consistent theory, 167
consistent with ZFC, 66, 133, 141,
155, 168
constants in the language of set
theory, 6, 168, 171, 174

Index 231
continuous function, 33, 72, 107,
111
continuum c, 66, 76, 195
continuum hypothesis, 66, 87, 89,
105, 108, 111, 115, 122,
124, 129, 136, 139, 157,
168, 183
generalized, 66, 107
countable chain condition ccc, 140
countable set, 64
countably closed forcing, 183
cover, 58
Darboux function, 106
decreasing function, 41
strictly, 41, 51
Dedekind cut, 31
denitions by recursion, 49
denitions by transnite induction,
49

0
-formula, 172
-system lemma, 150, 151, 153,
154, 181, 206
dense set, weakly, 139
dense set below p, 183
dense subset of a linearly ordered
set, 133, 135, 154
dense subset of a partially ordered
set, 130, 132, 141, 174
dense subset of R
n
, 33, 79, 99,
117
diagonal intersection of closed and
unbounded sets, 162
diagonalization, 81, 129, 198
diamond principle , 158, 182
-sequence, 158
dierence of sets, 8
symmetric, 9
dierence-free partition, 124
discrete space, 94
discrete subset of R
n
, 89
distance in R
n
, 32, 86
domain of a function, 16
domain of a relation, 12, 173
empty set axiom, 7
equivalence class, 14
equivalence relation, 14, 100
Erd osHajnal theorem, 86, 115
Erd osKakutani theorem, 115
Euclidean space R
n
, 32
exponentiation of cardinal num-
bers, 68
extension
of a function, 18
generic, 170
extensionality axiom, 6
eld, 55
of a relation, 12
lter, 60, 130
T-generic, 132
generic, 169
prime, 60
principal, 131
proper, 60
nite set, 64
rst element, 22
rst-category set, 98, 144
Fodors theorem, 162
forcing, 132
cardinal-collapsing, 180
cardinal-preserving, 180, 188
countably closed, 183
iterated, 189
forcing method, 164
forcing principle, 169
forcing relation |, 177
forest in a graph, 60
formal proof, 167
formula, 6, 166
a consequence of a theory, 167

0
-, 172
|

P
representing forcing re-
lation, 177
relativization of, 168

232 Index
free variable, 166
F

set, 94
FubiniTonelli theorem, 105
function, 16
GCH, 66
G

set, 94, 99
generalized continuum hypothesis,
66, 107
generic extension of a model, 170
generic lter, 132, 169
G odels incompleteness theorem
rst, 5, 167
second, 4, 167
good P-name, 199
graph on a set, 60
bipartite, 116
forest in, 60
greatest element, 22, 39
ground model, 170
group, 55
Abelian, 55
HahnBanach theorem, 54, 56
Hamel basis, 111, 115
with Baire property, 117
without Baire property, 113
measurable, 117
nonmeasurable, 113
Hausdor maximal principle, 54,
140
ideal, 14, 98
of nowhere-dense sets ^T, 98
image, 16
incompatible elements in a par-
tially ordered set, 140
inconsistent theory, 167
increasing function, 41
strictly, 41, 47, 74, 138, 139
independence
algebraic, 60
of the axioms, 5
linear, 55, 113, 123
from ZFC, 5, 66, 110, 164,
166, 168, 183
index set, 19
indexed family, 19
induction
on natural numbers, 27
too short, 130
transnite, 40, 198
denitions by, 49
innite set, 64
innity axiom, 10, 26, 182
initial ordinal number, 62
initial segment, 31, 39
proper, 40, 41
injective function, 16
integers, 30
integral, 105
interior int(S) of a set, 32
intermediate value theorem, 106
intersection of sets, 8, 19
diagonal, 162
inverse relation, 13
isolated point, 89
isomorphism of ordered sets, 39,
41, 47, 134, 135, 185
iterated forcing, 189
iterated integral, 105
iteration of ccc forcings, 202, 206
iteration of partially ordered sets,
200, 204
Jones space, 114
KirchheimNatkaniec theorem, 108
Kuratowskis theorem, 126
language of set theory, 4, 6, 166
largest element, 22
last element, 22
least element, 22
least upper bound, 135
Lebesgue measure, 101

Index 233
Lebesgue measure-zero set, 98, 117,
145, 193, 194
Lebesgue-measurable functions, 104
Lebesgue-measurable Hamel ba-
sis, 117
Lebesgue-measurable sets, 98
lexicographic-order relation, 43
limit points F
t
of F R
n
, 92
linear function, 111
linear functional, 56
linear space, 55
basis of, 55
linear subspace, 55, 111
linearly independent subset, 55,
113, 123
linearly ordered set, 23, 39, 42
linear-order relation, 23, 60, 139,
156
lower bound, 52
MA, 141
map, 16
conal, 74
Martins axiom MA, 141, 155, 195,
196
maximal antichain, 140
maximal element, 22, 39, 53, 54
Mazurkiewiczs theorem, 81
meager set, 98, 144
minimal element, 22, 39
model
Cohen, 182, 189
ground, 170
for theory T, 168
for ZFC, 169
for ZFC+c =
2
, 195
for ZFC+, 182
for ZFC+MA+CH, 196
for ZFC+CH, 168
for ZFC+MA, 195
for ZFC+2
1
=
2
, 188
natural numbers, 4, 25
non-Baire functions, 104
non-Borel functions, 104
nonmeasurable functions, 104
nowhere-constant function, 108
nowhere-dense set, 33, 98, 110,
117
nowheresymmetrically continuous
function, 118
null set, 98, 145, 193, 194
one-to-one function, 16
onto function, 16
open ball B(p, ) in R
n
, 32
open cover, 58
open set in R
n
, 32, 89
order isomorphism, 39, 41, 47, 134,
135, 185
order relation
antilexicographic-, 43
complete, 135, 154
lexicographic-, 43
linear-, 23, 60, 139, 156
partial-, 21, 60, 174
order topology, 154, 157
order type
, 135
, 135

, 51
Otp(W) of a well-ordered set,
47
of a partially ordered set, 38
ordered n-tuple, 10
ordered pair, 9, 173
order-isomorphic partially ordered
sets, 38
ordinal number, 44, 90, 174, 196
conality of, 74
initial, 62
limit, 46, 174
successor of, 46, 174

234 Index
ordinal numbers
arithmetic of, 48
product of, 48
sum of, 48
pairing axiom, 8
partial functions Func(X, Y ) from
X into Y , 131
partially ordered set, 22
partially ordered sets
iteration of, 200, 204
order isomorphic, 38
product of, 149, 150
partial-order relation, 21, 60, 174
partition, 14, 86, 115, 119
dierence-free, 124
sum-free, 124
Peano axioms of arithmetic, 26
perfect subset of R
n
, 89, 90, 101,
102
P-name, 175, 177
good, 199
valuation of, 175
poset, 22
power set, 8
power set axiom, 8
preimage, 16
preorder relation, 24, 139
preservation, in induction proofs,
81
preservation of cardinal numbers,
180, 188
preservation of properties in generic
extensions, 172
pressing-down lemma, 162
principle of mathematical induc-
tion, 26
principle of transnite induction,
40
product
of cardinal numbers, 68
Cartesian, 10, 20
of ordinal numbers, 48
of partially ordered sets, 149,
150
Tychono, 58
product lemma, 190
proper initial segment, 40, 41
quantier, 6, 166
bounded, 168
unbounded, 168
quotient class, 14
range of a function, 16
range of a relation, 12, 174
rank of a set, 166
RasiowaSikorski lemma, 132, 139,
169
rational numbers, 30
real function, 104
real numbers, 25, 31
Cohen, 182
recursion denitions, 49
recursion theorem, 49
reexive relation, 13
regressive function, 162
regular cardinal numbers, 75, 144
relation
antisymmetric, 21
binary, 12, 173
connected, 23
domain of, 12, 173
equivalence, 14, 100
forcing |, 177
|

, 177
inverse, 13
order
antilexicographic-, 43
complete, 135, 154
lexicographic-, 43
linear-, 23, 60, 139, 156
partial-, 21, 60, 174
preorder, 24, 139
range of, 12, 174
reexive, 13

Index 235
restriction of, 23
=

, 139
<

, 136

, 139
symmetric, 13
transitive, 13
well-founded, 37, 163
well-ordering, 38, 90
relations
composition of, 13
in the language of set theory,
168
replacement scheme axiom, 17
restriction of a function, 18
restriction of an order relation, 23
Russells paradox, 3, 7
scale, 136, 141
Schr oderBernstein theorem, 64
selector, 15, 117
sentence, 167
a consequence of a theory, 167
consistent with a theory, 167
independent of a theory, 5,
167
sequence, 18
Cauchy, 14, 33
-, 158
transnite, 48
set-theoretic universe, 7
Sierpi nskiZygmund theorem, 111
Sierpi nskis theorem, 88, 105
-algebra, 93
of Borel sets Bor, 94
of Lebesgue-measurable sets
L, 101
of sets Baire with the Baire
property, 101
-ideal, 98
of rst-category sets, 98, 144
of Lebesgue measure-zero sets,
98, 145, 193, 194
of meager sets, 98, 144
of null sets, 98, 145, 193, 194
smallest element, 22, 39
stationary set, 158, 162
strong measure-zero set, 153
strongly Darboux function, 106
subbase of a topological space, 58
subcover, 58
successor, 10, 26
of a cardinal number, 65
of an ordinal number, 46, 174
sum of cardinal numbers, 68
sum of ordinal numbers, 48
sum-free partition, 124
surjective function, 16
Suslin hypothesis SH, 155
Suslin line, 155
Suslin tree, 163
symmetric dierence, 9
symmetric relation, 13
symmetrically continuous function,
118
nowhere, 118
theory, 167
consistent, 167
inconsistent, 167
topological properties of R
n
, 32
topological space, 58
base of, 58
Borel structure of, 93
compact, 58
discrete, 94
subbase of, 58
topology, 58
order, 154, 157
on R
n
, 32
transcendental basis, 60
transnite induction, 40, 198
denitions by, 49
transnite sequence, 48
transformation, 16
transitive closure trcl(x) of a set,
165

236 Index
transitive model of ZFC, 169
transitive relation, 13
transitive set, 164, 174
tree, 163
Aronszajn, 163
branch of, 163
Suslin, 163
tree construction, 92, 102
Tychono product, 58
Tychono theorem, 54, 58
unbounded quantier, 168
unbounded subset in an ordinal
number, 74
uncountable set, 64
uniformly antisymmetric function,
119
union axiom, 8
union of cardinal numbers, 65
union of sets, 8
upper bound, 52
least, 135
valuation, 175
variable, 6, 166
bound, 166
free, 166
vector space, 55
Vitali set, 117, 126
weakly dense set, 139
well-founded relation, 37, 163
well-ordered set, 38, 39, 42, 47, 51
type of, 47
well-ordering relation, 38, 90
well-ordering theorem, 51
ZermeloFraenkel axioms, 4
Zermelos theorem, 51
ZF, 52
ZFC, 4
Zorns lemma, 53, 54