Recursion in Set Theory

Logic I
Gunter Fuchs
July 22, 2009
2
Contents
1 Syntax 5
1.1 Words . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.2 Formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Sets 9
2.1 Fragments of ZFC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Well-founded Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Ordinal Numbers and Innity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Well-orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 Ordinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.3 The Natural Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.4 Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Cardinal Numbers and Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3 Semantics 31
3.1 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Consequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4 Same same... but dierent 37
4.1 Correctness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.4 Ultralters and Ultraproducts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.4.1 Filters and ultralters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.4.2 Ultraproducts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.4.3 Compactness revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5 Incompleteness 51
5.1 Arithmetic with Exponentiation and Representability . . . . . . . . . . . . . . . 51
5.2 A dierent Godelization: Ackermanns coding of V
. . . . . . . . . . . . . . . . 60
5.3 Godels and Tarskis results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3.1 Recursive sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3
4 CONTENTS
Introduction
These notes were produced on the y in Summer-semester 2009, when I taught the lecture
course Logic I. The situation was complicated because of the heterogeneity of the audiences
backgrounds. Some knew the fundamentals of recursion theory, having attended the Introduction
to computability lecture course of the previous semester, some knew the basics of mathematical
logic from the course Logische Grundlagen, some had attended both of these lectures, some
only one of them, and some had no background whatsoever.
So I decided to approach the topic in a somewhat untraditional way: In the rst chapter, I
describe the syntax of rst order logic and give the denition of the Tait calculus, a cut-free proof
calculus which is the one preferred by proof-theorists. Being cut-free makes this calculus more
amenable to methods from computer science, in particular automated theorem proving. Those
students with a background in logic have not seen this type of calculus, since they previously
worked with a Hilbert-style sequent calculus. So I hope nobody was bored by this chapter.
I follow a very pure approach, in that I dont give any proofs in the rst chapter. The reason
is that I could be asked which theory I am proving the theorems in. So in chapter 2 I develop
basic set theory. The theorems I am stating are proved in (fragments of) set theory, and in
theory, they could be deduced using the Tait calculus. The main tool I need in the that chapter
is the recursion theorem. I develop basic set theoretic concepts up to ordinals, trees and the
axiom of choice with some equivalent formulations.
In chapter 3, I use a weak background set theory to dene the semantic notions like models
and logical consequence.
Chapter 4 is dedicated to the proofs of correctness and completeness of the Tait-calculus. I
prove completeness and compactness rst for countable languages only. In a last section, I prove
the general completeness and compactness theorems for arbitrary languages, using ultraproducts,
and using full ZFC as the background theory.
Finally, in chapter 5, I prove Tarskis undenability of truth theorem and Godels rst in-
completeness theorem. Instead of the traditional Godel coding of formulae, I use Ackermanns
isomorphism which allows to translate relations which are
1
in the hereditarily nite sets to re-
cursive/representable relations. Working in the hereditarily nite sets allows us to apply a weak
set theory (without innity) in order to see that the basic syntactical concepts are representable
in arithmetic.
Please email any mistakes and typos to gfuchs@uni-muenster.de
Gunter Fuchs, July 2009
Chapter 1
Syntax
1.1 Words
Denition 1.1.1. Let be a set. A free semi-group generated by is a structure Z,
), where
: Z
2
Z, satisfying the following axioms:
1. (x
y)
z = x
(y
z).
2. there is an e Z such that for all z Z, z
e = e
z = z. e is uniquely determined, and

well write for the empty word.
3. For every s , there is an s
t
Z, such that the map s s
t
is injective. Moreover,
whenever s
t
= x
y, then either x = s
t
and y = , or x = and y = s
t
.
4. every z Z has the form z = s
t
1
s
t
2
. . .
s
t
n
, where s
1
, . . . , s
n
. This holds for z = e
with n = 0.
5. if s
1
, . . . s
m
and t
1
, . . . , t
n
, then:
s
t
1
. . .
s
t
m
= t
t
1
. . .
t
t
n
i m = n and s
i
= t
i
for all i < n.
Remark 1.1.2. A free semi-group generated by a set is uniquely determined by , up to
isomorphism. This means that any two semi-groups generated by are isomorphic. Thats
why Ill refer to the semi-group generated by , and Ill denote it by
may stand for the

structure Z,
) as well as for the set Z. Also, to save space, Ill drop the connective

, and just
write xy in place of x
y. Dropping the brackets is justied by 1. Ill also identify s and s

t
, for
s , so that
.
Denition 1.1.3. A member s occurs in a word w
if there are words w

0
and w
1
such
that w = w
0
w
1
. An occurrence of s in w is a triple w
0
, s, w
1
) such that w = w
0
w
1
.
1.2 Languages
Denition 1.2.1. A rst order language is a quadruple L = C, P, F, #) such that
1. C, P and F are disjoint sets,
2. # : (P F) N.
5
6 CHAPTER 1. SYNTAX
The alphabet of L, (L), is the least set containing C, P, F and the following distinct elements:
1. v
n
, for n N, in the sense that the map n v
n
exists.
2.

,

, ,

,

,

(,

), ,.
3. =.
1
Ill denote the set v
n
[ n < by Var(L).
The Tait-alphabet of L,
T
(L), is the least set containing C, P, F and the following distinct
elements:
1. v
n
, for n N, in the sense that the map n v
n
exists.
2.

,

,

,

,

(,

), ,. So is missing!
3. = and

,=.
4. An element

P, for every P P, in the sense that the map P

P is injective, and all the
elements

P are new. We shall set

P = P.
1.2.1 Terms
Denition 1.2.2. Let L = C, P, F, #) be a language with alphabet (L). The set of terms of
L is the least subset X of (L)
with the following properties:

1. Var(L) C X,
2. if f F, n = #(f) and t
1
, . . . , t
n
X, then
f
t
1
t
2
, . . . ,
t
n
) X.
1.2.2 Formulae
Denition 1.2.3. Let L = C, P, F, #) be a language with alphabet (L). I shall dene what
atomic formulae and compound formulae are, and at the same time which occurrences of variables
in these formulae are free and which are bound. Then Ill do the same thing for atomic Tait-
formulae and compound Tait-formulae. Every occurrence of a variable will be either free or
bound.
The set of atomic formulae of L is the least subset X of (L)

1. If t
1
and t
2
are terms of L, then t
1
=t
2
X.
2. If P P, n = #(P) and t
1
, . . . , t
n
are terms of L, then P
(t
1
,t
2
, . . . ,t
n
) X.
Every occurrence of a variable in an atomic formula is free.
The set of compound formulae is the least subset X of (L)

1. If
1
and
2
are atomic formulae or members of X, then

(
1
) X, for =

or =

.
An occurrence
(w
0
, v
m
, w
1

2
)) of v
m
in

(
1

2
) is free (bound) if w
0
, v
m
, w
1
) is a free
(bound) occurrence of v
m
in
1
. Analogously, an occurrence
(
0
w
0
, v
m
, w
1
)) of v
m
in
(
1

2
) is free (bound) if w
0
, v
m
, w
1
) is a free (bound) occurrence of v
m
in
2
.
1
This symbol is usually part of the alphabet. In some rare occasions, one may want to work without it though.
Still, the default is that it is part of the alphabet.
1.2. LANGUAGES 7
2. If is an atomic formula or a member of X, then
X. An occurrence w
0
, v
m
, w
1
)
in is free (bound) if w
0
, v
m
, w
1
) is free (bound) in .
3. If is an atomic formula or a member of X, then Qv
m
X, for every m N such that
v
m
does not occur as a bound variable in , and for Q =

or Q =

. Every occurrence of
v
m
in Qv
m
is bound. If n ,= m, then every occurrence Qv
m
w
0
, v
n
, w
1
) of v
n
in Qv
m
is
free (bound) if w
0
, v
n
, w
1
) is a free (bound) occurrence of v
n
in .
The set of atomic Tait-formulae is the least subset X of
T
(L)

1. If t
1
and t
2
are terms of L, then t
1
=t
2
X and t
1

,=t
2
X.
2. If P P, n = #(P) and t
1
, . . . , t
n
are terms of L, then P
(t
1
,t
2
, . . . ,t
n
) X, and also
(t
1
,t
2
, . . . ,t
n
) X.
Every occurrence of a variable in an atomic Tait-formula is free.
The set of compound Tait-formulae is the least subset X of (L)
with the following proper-

ties:
1. If
1
and
2
are atomic formulae or members of X, then

(
1
) X, for =

or =

.
Whether an occurrence of a variable is free or bound is dened as above.
2. If is an atomic formula or a member of X, then Qv
m
X, for every m N such that
v
m
does not occur as a bound variable in , and for Q =

or Q =

. Again, whether an
occurrence of a variable is free or bound is dened as above.
If is a formula (or a Tait-formula), then a variable v is a free variable of if there is a free
occurrence of v in . It is a bound variable of if there is a bound occurrence of v in . It is a
(Tait-)sentence if it has no free variable.
Denition 1.2.4. If is a Tait-formula, then is the Tait-formula arising by swapping the
following symbols:
1.

and

,
2.

and

,
3. = and

,=,
4. P and

P (for P P).
If is a set of Tait-formulae, then set
:= [ .
Denition 1.2.5. Let be a formula or a Tait-formula, v
m
a variable and t a term. Then
(
v
m
/
t
) is the formula which arises from by replacing every free occurrence of v
m
in with
t. More precisely, let = w
0
v
m
w
1
v
m
w
2
. . . v
m
w
k
be maximal, meaning the following: Whenever
u
0
, v
m
, u
1
) is a free occurrence of v
m
in , then there is an l < k such that u
0
= w
0
v
m
. . . w
l
and u
1
= w
l+1
v
m
. . . w
k
. Then (
v
m
/
t
) = w
0
tw
1
tw
2
. . . tw
k
.
If is a set of formulae (or of Tait-formulae), then (
v
m
/
t
) = (
v
m
/
t
) [ .
Usually, I will tacitly assume that no free variable of the term t occurs as a bound variable in
when forming (
v
m
/
t
). Otherwise, this would produce undesired results, for example v
0
v
0
>
v
1
(
v
1
/
v
0
) = v
0
v
0
> v
0
. One might take the substitution to be again otherwise.
8 CHAPTER 1. SYNTAX
1.3 Proofs
For more on the proof calculus introduced in this section, I refer the reader to [Poh09].
Note: When writing down a formula in some given language, I will use the binary connective ,
even though I dont consider it part of the language. It is meant as an abbreviation, as follows:
( ) means ( ).
When using the symbol in Tait-formulae, it is to be understood as follows:
( ) means ( ).
Denition 1.3.1. Fix a language L = L(C, P, F, #). The identity axioms of L are the following
formulae:
Equivalence Relation
(a) v
0
v
1
(v
0
= v
1
v
1
= v
0
),
(b) v
0
v
1
v
2
((v
0
= v
1
v
1
= v
2
) v
0
= v
2
),
(c) v
0
(v
0
= v
0
)
Congruence over functions
For every F F with n = #(F), the sentence
v
0
v
1
. . . v
n1
v
n
v
n+1
. . . v
2n1
(((v
0
= v
n
) (v
1
= v
n+1
) . . . (v
n1
= v
2n1
))
F(v
0
, v
1
, . . . , v
n1
) = F(v
n
, v
n+1
, . . . , v
2n1
)).
Congruence over predicates
For every P P

P with n = #(P), the sentence
v
0
v
1
. . . v
n1
v
n
v
n+1
. . . v
2n1
(((v
0
= v
n
) (v
1
= v
n+1
) . . . (v
n1
= v
2n1
))
(P(v
0
, v
1
, . . . , v
n1
) P(v
n
, v
n+1
, . . . , v
2n1
))).
These formulae can be viewed as being regular formulae or Tait-formulae; both are intertrans-
latable.
Denition 1.3.2. Fix a language L. Say that
T
if is a nite set of Tait-formulae which
belongs to the least collection of nite sets of Tait-formulae X with the following properties:
1. (Tertium non datur) X whenever there is an atomic Tait-formula such that
and .
2. (-rule) If X and X, then X.
3. (-rule) If X or X, then X.
4. (-rule) If (
v
m
/
v
n
) X, where v
n
does not occur as a free variable in any formula
of , then v
m
X.
5. (-rule) If (
x
/
t
) X, for some term t, then x X.
If is a set of L
T
-formulae and is a L
T
-formula, we write if there is a nite set
0
of L
T
-formulae consisting of formulae which are in or an identity-axiom of L, such that

0
.
Chapter 2
Sets
The role of set theory is two-fold: On the one hand, (a fragment of) it serves as the meta-theory.
On the other hand, it can be viewed as a formal theory to which we can apply the logical methods
we are going to develop. The theorems I am going to prove could be proved formally, using the
calculus introduced in the rst chapter. But of course, I am not going to give formal proofs. Its
just good to know that in theory, it would be possible to do so.
The aim is to axiomatize the usage of sets in mathematics without running into contradictions.
Formally, the language we are going to use only consists of one binary predicate, denoted by

.
In addition, we have the equality symbol at our disposal, which belongs to every language (by
our convention). In fact, it will turn out that equality becomes denable in set theory.
We shall adopt the view that everything there is in our mathematical world is a set. As a
rst attempt to axiomatizing set theory, lets look at a quote of Georg Cantor:
Unter einer Menge verstehen wir jede Zusammenfassung M von bestimmten wohlun-
terschiedenen Objekten unserer Anschauung oder unseres Denkens (welche die Ele-
mente von M genannt werden) zu einem Ganzen.
Lets try to formalize this approach rigorously. That x has a property would be formalized
by saying that some formula is true of x. So one might demand that
x [ (x)
is a set.
1
Since everything there is will be a set, this would mean:
y y = x [ (x),
or, in other words:
yx(x y (x)).
Here, I use ( ) as an abbreviation for (( )( )), which, in turn, is an abbreviation
for a longer formula.
The problem is that this leads to a contradiction: Lets apply the above principle to the
formula = (x = x). Wed get that V := x [ x = x is a set. Clearly, it would satisfy
V = V.
1
When writing (y), I mean that y are the free variables of .
9
10 CHAPTER 2. SETS
So by denition of V,
V V.
This contradicts the common view of a set: A set should not be a member of itself. It is not a
formal contradiction, though, and one can formalize set theory without excluding this possibility.
But lets investigate the sets which do not have themselves as a member some more:
A = x [ x / x.
In other words, A is the set consisting precisely of those sets x that satisfy x / x. Adopting the
view that a set is uniquely characterized by its members, this determines A. According to our
strong set existence principle, A would have to be a set. So:
Either A A or A / A.
But by denition of A, we have:
A A A / A.
This is known as Russels antinomy. It is a contradictory statement, and it shows that we have
to use somewhat weaker set existence principles. I shall develop these in the following.
2.1 Fragments of ZFC
Literature: Any book on axiomatic set theory will do, but to be specic, I refer to [Jec03]. The
approach chosen there is close to the one I use.
The language of set theory I shall use is L = L
ST
= C, P, F, #), where C = F = , P =
and #(
) = 2. As usual, I will use the inx notation for the predicates

and =, and also, I
shall drop the dots most of the time. Thus, e.g., I shall write
x(x y x = z)
instead of
x(
(x, y) =(x, z)).

Also, I shall use x, y, z, u, v, w, etc. as meta-variables for variables. So x is really some v
m
, but I
dont say which.
The axioms I am going to introduce together will form the axiom system called ZFC. ZF
stands for Zermelo-Fraenkel and is named after the founders of the system. The C stands for
the axiom of choice that I will introduce much later, in section 2.4.
The rst axiom I want to introduce is the Set Existence Axiom,
x x = x.
This axiom is really a logical axiom, since usually one does not consider models with empty
domain. Sometimes it is formulated to express that there is a set which has no elements, i.e.,
xyy x. This form of the axiom will be a consequence of the other axioms, though.
The second basic axiom I need is the Extensionality Axiom, saying:
xy(z(z x z y) x = y).
It expresses that if two sets have the same members, then they are equal. The converse is true
anyway, by the identity axioms.
2.1. FRAGMENTS OF ZFC 11
Before introducing further axioms, let me introduce a simpler way of writing them down.
Namely, if (x, y
1
, . . . , y
n
) is a formula with free variables shown, then I will call a syntactical
object of the form
x [ (x, y
1
, . . . , y
n
)
a class term. Often I shall write y in place of y
1
, . . . , y
n
. I shall use capital letters for class terms;
for example, I set:
V := x [ x = x.
I shall also consider variables as being class terms. So a class term is either a variable or given by
a formula as above. I am going to use class terms in formulae, but these class terms can always
be eliminated, producing a regular L
ST
-formula. This is how it works. If A is a class term, then
x
A :=
_
x v
m
if A is a variable,
(
u
/
x
) if A is the class term u [ (u, y).
If x occurs as a bound variable in above (in the second case), then this necessitates a renaming
of the bound variables of : Pick l least such that v
l
does not occur in . Let
be the
formula arising from by replacing every occurrence of x in by v
l
; this can be dened just
like the substitution of terms for free variables. Then x u [ (u, y) should be the formula
(
u
/
x
). This phenomenon is referred to as collision of variables, and I shall ignore it in the
future. The cleanest approach is to take what I just described as the ocial denition of (
u
/
x
).
Having understood the problem once is sucient, dealing with it all the time just adds a layer
of notational complexity that conceals sometimes simple ideas and reveals nothing (even though
it of course would be more correct). So the renaming of bound variables happens tacitly... Note
that I havent dened the meaning of A B, if A is not a variable; the x above stands for a
variable only. Before explaining this though, let me stipulate:
A =B :=
_
x = y if A is the variable x and B is the variable y,
x(x A x B) otherwise.
Note that the meaning of x A and x B is dened above. Note also that this interpretation
of A = B conforms with the axiom of extensionality. Now Im ready to explain the meaning of
A B:
A
B := x(x B x = A).
Class terms of course never appear right after a quantier, so it is clear now how to eliminate
class terms from formulae. Note that the introduction of these terms is a mere convenience,
making formulae more easily readable. While were at it, I would like to dene some operations
on class terms. Again, these are just abbreviations.
Denition 2.1.1. Let A and B be class terms.
1. A B := x [ (x A x B),
2. A B := x [ (x A x B),
3. A
c
:= x [ x A,
4. A B := A B
c
.
Also, I want to use the following reserved class terms:
V = x [ x = x, = V
c
= x [ x = x.
12 CHAPTER 2. SETS
It is sometimes useful to allow class terms of the form
A
1
, . . . , A
n
,
where each A
j
is a class-term. The interpretation is:
A
1
, . . . , A
n
= x [ x = A
1
x = A
2
. . . x = A
n
.
Finally, I also want to consider class terms of the form
A[
y
(x, y),
where A is a class term with free variables y. The intended meaning is
A [
y
(x, y) = z [ y((x, y) z = A).
Usually, the variables x will occur as free variables in A as well, but not necessarily. And usually,
A = A(x, y, z) will always be a set when (x, y) holds. The purpose of the subscript y is to
clarify that the variables y are bound.
Note that the set existence axiom can be expressed as V ,= . The next axiom is the Pairing
Axiom:
xyx, y V.
Unraveling this formula, eliminating the class terms, yields a formula equivalent to this:
xyzu(u z (u = x u = y)).
I hope the advantage of using class terms is becoming clearer now. Ill use another operation on
class terms:
Denition 2.1.2. If A is a class term, then let
A be the following class term:

_
A := z [ y(y A z y),
where z and y do not occur in A (as usual). This is called the inner union of A. Sometimes it is
written as
zA
z. Analogously
A is short for
A := z [ y(y A z y).
The remarks above apply. Finally, Ill write A B as a short form for the statement
x(x A x B).
In order to get familiar with these concepts, the reader may try to gure out what

is...
The next axiom Ill introduce is the Union axiom:
x
_
x V.
I trust the reader will be able to retranslate this into a L
ST
-formula.
Lemma 2.1.3. It is provable from the axioms introduced thus far that
1. xy x y V.
2. x
1
x
2
. . . x
n
x
1
, x
2
, . . . , x
n
V.
Proof. This is true in the sense that if is the set of axioms introduced thus far,
xyx y V,
and the same applies to the second sentence. But instead of trying to obtain a formal proof, I
shall argue informally. Much later, we shall see that such informal arguments can be translated
into a formal proof, in principle.
For 1., let x, y be given. By the Pairing Axiom, x, y V. By the Union Axiom,
x, y
V. But by denition,
x, y is the same as x y, so were done.

For 2., let x
1
, x
2
, . . . , x
n
be given. By Pairing, x
1
V, x
2
V, . . ., x
n
V. Applying
1. (n-1) times, it follows that x
1
x
2
. . . x
n
V. But thats just x
1
, x
2
, . . . , x
n
.
The next axiom is a way of avoiding Russels antinomy, while still allowing us to collect sets
which satisfy a certain property into one set. It is called the Separation Axiom. For every class
term A, it says:
x A x V.
More formally, one should demand that x does not occur in A, and one should take the univer-
salization of this formula, meaning that one prexes it with y
1
y
2
. . . y
n
, where y lists the free
variables occurring in A.
I will take this as a convention: If free variables occur in an axiom, then I mean the sentence
obtained by prexing the formula with universal quantiers which bind these variables. Remember
that class terms may contain concealed free variables, as is the case in the Separation Axiom.
So this axiom does not say that the collection of all sets having some property is a set, but
that the collection of all sets having some property and being a member of some xed set is a
set. The idea is that since this xed set is a set, we have understood it well enough that we can
understand the collection of its members which satisfy some property, so that we dont run into
a contradiction.
If is a set of L
ST
-formulae, then -Separation is formulated like the Separation-Axiom, but
only for class terms A whose formula is in .
Lemma 2.1.4. The following are consequences of the axioms introduced thus far:
1. V.
2. x (A x A V).
3. (A ,=
A V).
4. V / V.
5. x x
c
/ V.
Proof. 1. By Set Existence, there exists some set x. By Separation, x V. By denition of
, x = .
2. If A x, then A = A x V, by Separation.
3. Pick a A (so a V). Then
A a, so by 2.,
A V.
4. Suppose v := V V. As in the introduction to this chapter, let A = x [ x / x. Clearly,
A v, so by 2., a := A V. But then, a a a / a.
5. If x
c
V, then x x
c
V, by Lemma 2.1.3, 1. But clearly, x x
c
= V.
The following class term will often be very useful:
14 CHAPTER 2. SETS
Denition 2.1.5.
A, B) := A, A, B.
A, B) is the ordered pair of A and B.
Just as in the case of class terms of the form A, B, the ordered pair only is useful for class
terms A and B which are sets.
Lemma 2.1.6. For x, y, v, w V:
1. x, y) V,
2. (x, y) = v, w) (x = v y = w)).
Proof. Exercise.
One can generalize the formation of ordered pairs to produce nite tupels:
Denition 2.1.7. Set:
x) = x
x
1
, . . . , x
n+1
) = x
1
, x
2
, . . . , x
n+1
)).
Clearly then:
Lemma 2.1.8. For all sets x
1
, . . . , x
n
, y
1
, . . . , y
n
:
1. x
1
, . . . , x
n
) V,
2. if x) = y), then x
1
= y
1
x
2
= y
2
. . . x
n
= y
n
.
Denition 2.1.9.
AB := x, y) [ x A y B,
A
1
. . . A
n
:= A
1
(A
2
. . . A
n
),
A
n
:= A. . . A
. .
n times
.
Denition 2.1.10. A class term R is a relation if R V
2
. We shall often write xRy instead of
x, y) R. R is an n-ary relation if R V
n
(for n 2).
When writing R is a relation, I really mean the formula that R V
2
is a short form for.
Remember that R is either a variable or a class term; in the latter case, R is basically given by
a dening formula.
Denition 2.1.11.
dom(R) = xy [ yRx,
ran(R) = yx [ yRx,
d(R) = dom(R) ran(R),
R A = y, x) [ x A yRx,
RA = y [ x x A yRx,
R
1
= x, y) [ yRx
Denition 2.1.12. If F is a class term, then F is a function stands for the following formula:
F is a relation y
0
y
1
x((y
0
Fx y
1
Fx) y
0
= y
1
).
In case F is a function with x dom(F), I shall write F(x) for the uniquely determined y such
that yFx. If dom(F) V
n
, then I shall write F(x
1
, . . . , x
n
) in place of F(x
1
, . . . , x
n
)).
Now Im ready to introduce the Replacement Axiom:
x (F is a function Fx V).
Again, this is a scheme of axioms, one formula for each class term.
Lemma 2.1.13. The Separation Axiom scheme is provable from the Replacement Axiom scheme.
Proof. Let id = x, x) [ x V; this is clearly a function. Now, given any class term A and any
set a, it follows that idA is also a function, and A a = (idA)a V, by Replacement.
The axioms introduced so far form the weakest fragment of the axioms of set theory we shall
work with. I call it ZF
, which is an ad-hoc name. So it consists of the following axioms:

SetExistence,
Pairing,
Union,
Extensionality,
Separation,
Replacement.
Lemma 2.1.14. The following are consequences of the ZF
axioms:
1. xy x y V,
2. x
1
x
2
. . . x
n
x
1
. . . x
n
V,
3. x(F is a function Fx V).
Proof. 1.) Fix x and y. First note:
z x z V.
To see this, x z. We know already that for any set u, u, z) V. So the following denes a
function with domain x:
F
z
:= u, z), u) [ u x,
i.e., F
z
(u) = u, z), for u x. By Replacement, F
z
x V, and by denition,
F
z
x = u, z) [ u x = x z.
Knowing this, we can now dene a function G as follows:
G := x z, z) [ z y,
16 CHAPTER 2. SETS
i.e., G(z) = x z, for z y. Applying Replacement again yields that Gy V, and by
denition of G, a := Gy = x z [ z y. But then

a V, by the Union axiom, and
a = x y, as is easily checked.
2.) Follows by induction from 1.)
3.) Fix F and x. Then Fx = F ((Fx) x). Fx V by Replacement, so that
(Fx) x V by 1., so that F ((Fx) x) V by Separation.
The next axiom Ill introduce is the Power Set Axiom. Given a class term A, let
T(A) = x [ x A
be its power class. The Power Set Axiom says:
x T(x) V,
i.e., the power class of a set is a set. A lot of set theory can be done without the power set axiom,
so I shall sometimes not assume it.
The last axiom I want to introduce at this early stage has to do with the question raised in
the introduction to this chapter: Can/should there be a set x with x x? The Foundation Axiom
excludes this possibility (among others). To formulate it, I need the following:
Denition 2.1.15. x is -minimal in A if x A and for all y x, y / A.
This is again to be understood as an abbreviation: Given a class term A and a variable x,
x is -minimal in A stands for the following formula:
x A y (y x y A).
The Foundation Axiom now is the following scheme of sentences:
A ,= x x is -minimal in A,
one for each class term. I call the system of axioms comprised of ZF
, together with the

Foundation axiom scheme ZF
F
.
Lemma 2.1.16 (ZF
F
). There is no nite sequence of sets x
1
x
2
. . . x
n
x
1
.
Proof. Otherwise x
1
, x
2
, . . . , x
n
has no -minimal member.
2.2 Well-founded Relations
In this section, I shall work with the axioms introduced thus far, save Power Set and Foundation,
unless stated otherwise. This is what I referred to as ZF
in these notes.
Denition 2.2.1. Given class terms A and R, we say that x is R-minimal in A to express that
R is a relation, x A and for all y, if yRx, then y / A. A relation R is set-like if for all x,
Rx V. Saying that R is well-founded stands for the scheme of formulae
(A ,= x x is R-minimal in A),
one formula for each class term A. R is strongly well-founded if R is well-founded and set-like.
A is R-closed if RA A.
Lemma 2.2.2. Let R V
2
be set-like. Then:
2.2. WELL-FOUNDED RELATIONS 17
1. u Ru V,
2. If u is R-closed, for all u A, then

A and

A are R-closed,
3. Suppose R is well-founded. Then for every set u there is a set v which is minimal wrt. in-
clusion such that u v and v is R-closed.
Remark 2.2.3. We shall see later that the last point of this lemma follows from an axiom that I
have not introduced yet, the innity axiom, without assuming R is well-founded or set-like.
Proof. 1.) Ru =

Rx [ x u =

Fu, where F(x) = Rx, by Replacement and
Union.
2.) This is trivial.
3.) Given any set x, I rst claim that there is a set w such that x w and w is R-closed. For
if not, consider the class A of sets for which this fails. Pick x R-minimal in A. Then Rx V,
and every y Rx (i.e., yRx) is not in A. So, momentarily xing such a y, the class B
y
of sets
z which are R-closed and have y z is non-empty. Hence w
y
:=
B
y
is a set, by 2.) its R-closed
(and by denition, its the -minimal set which has y as a member and which is R-closed). Now
dene F(y) = w
y
for y Rx. Then by Replacement and Union, x
FRx V, and
clearly, this set is R-closed and has x as an element. This is a contradiction.
It now follows that given any set x, there is a least set w
x
(wrt. inclusion) having x as a
member and being R-closed (as in the rst part of the proof, this is just the intersection of all
sets which have these properties; since there is such a set, the intersection is again a set).
Now given an arbitrary set u, one can form

w
x
[ x u =

Gu V, where G(x) :=
w
x
.
Denition 2.2.4. Given class terms F, A and B, write F : A B to express that F is a
function, dom(F) = A and ran(F) B. Write F : A > > B to express that F : A B and
for all x, y A, x ,= y F(x) ,= F(y). Write F : A >> B to express that F : A B and
B = FA. Finally, F : A > >> B expresses that F : A > > B and F : A >> B.
Theorem 2.2.5 (Recursion Theorem). Let R be a strongly well-founded relation, R A
2
. Let
G : AV B. Then there is a unique F : A B such that for all x A,
F(x) = G(x, FRx).
This means that, given class terms R, A, G, there is a class term F such that it is provable
from our axioms, together with the scheme expressing that R is a strongly well-founded relation
and the other assumptions of the theorem, that F satises F(x) = G(x, FRx), for all x A,
and that F is uniquely determined by this (as a class, not as a class term).
Proof. Uniqueness: Let F and F
t
be two such functions. Suppose they were dierent. Let x A
be R-minimal such that F(x) ,= F
t
(x). Then since both satisfy the equation,
F(x) = G(x, FRx) = G(x, F
t
Rx) = F
t
(x),
since by R-minimality of x, FRx = F
t
Rx. So F(x) = F
t
(x) after all.
Existence: Let T be the class consisting of all sets of the form f, u) with:
1. u A is R-closed,
2. f : u B,
3. x u f(x) = G(x, f Rx).
18 CHAPTER 2. SETS
(1) T is a function.
Proof of (1). Suppose f, u), f
t
, u) T. The above proof of uniqueness shows that f = f
t
.
2
(1)
Let D = dom(T), and write T(u) for the unique f with fTu.
(2) If u D and v u is R-closed, then v D and T(v) = (T(u))v.
Proof. Letting f := T(u)v, f, v) clearly satises 1.-3. of the denition of T. The claim follows,
since T is a function. 2
(2)
Let X =
D, F :=
T(u) [ u D. So F is a relation with X = dom(F) and ran(F) B.

(3) If u D, then Fu = T(u).
Proof of (3). By denition of F, T(u) F. For the other direction, assume y, x) Fu, i.e.,
x u and y, x) F. By denition of F again, there is some v D such that y, x) T(v).
Since the domain of T(v) is v, it follows that x v. By Lemma 2.2.2, u v is R-closed. So
u v D and T(u v) = T(u)(u v) = T(v)(u v), by (2). Since x u v, this means that
y = T(v)(x) = T(u)(x), and so, y, x) T(u). 2
(3)
(4) F : X B.
Proof of (4). It has to be shown that F is a function. Suppose y, x), z, x) F. Pick u D
such that x u. Then Fu = T(u), by (3). But y, x), z, x) Fu, so y, x), z, x) T(u). The
latter is a function, by denition, so that it follows that y = z. 2
(4)
(5) For x X, F(x) = G(x, FRx).
Proof of (5). Let x X. Pick u D such that x u. Then F(x) = (Fu)(x) = T(u)(x) =
G(x, T(u)Rx). But since Rx u, as u is R-closed, and since Fu = T(u), it follows that
FRx = T(u)Rx. This shows that F(x) = G(x, FRx), as wished. 2
(5)
(6) X = A.
Proof of (6). Clearly, X A. Suppose X A. Pick x R-minimal in A X. So x / X, but
since Rx A, Rx X. Let u be the least R-closed set which contains Rx; this exists
by Lemma 2.2.2.3. Since X is R-closed (being a union of R-closed sets; see Lemma 2.2.2.2), it
follows that u X (note that uX is an R-closed set (by Lemma 2.2.2.2) that contains Rx,
so that u u X). By (5), it follows that u D and that T(u) = Fu. Set:
v = u x,
g = Fu G(x, FRx), x).
Then clearly, v is R-closed, and g, v) T. So x v D, hence x X, which is a contradiction,
since x was chosen to be R-minimal in A X. This concludes the proof.
Remark 2.2.6. Looking back at the proof of the Recursion Theorem, the formula dening F is:
y = F(x) (x A f (f is a function
dom(f) is R-closed x dom(f)
z dom(f) f(z) = G(z, fRz)
y = f(x)))
2.2. WELL-FOUNDED RELATIONS 19
Denition 2.2.7. U is transitive if
xy ((y U x y) x U).
Denition 2.2.8. Let R U U. Then (U, R) is extensional if for every y, z U,
Ry = Rz y = z.
So this means that members of U are uniquely determined by their R-predecessors. (U, R) is
(strongly) well-founded if R is.
So for example, (V, ) is extensional and strongly well-founded.
Denition 2.2.9. If (U, R), (V, S) are such that R R
2
and S V
2
, then F : (U, R)

(V, S)
means:
1. F : U > >> V ,
2. xy (xRy F(x)SF(y).
For sets u, r) and v, s), saying that u, r)

= v, s) means that there is an f such that f :
(u, r)

(v, s).
Theorem 2.2.10 (Mostowskis Isomorphism Theorem). Suppose (U, R) is extensional and strongly
well-founded. Then there are uniquely determined F, V such that
F : (U, R)

(V, V )
and V is transitive. F satises the following equation, for all x U:
() F(x) = FRx = F(y) [ yRx.
Proof. Existence: By the Recursion Theorem, there exists a unique function F satisfying ().
Set V = FU. It suces to show that (F, V ) is as wished.
(1) V is transitive.
Proof of (1). Let x y V = FU. Let y U be such that y = F( y). By (), F( y) =
F( x) [ xR y, so since x F( y), there is some xR y such that x = F( x). In particular, x FU,
showing that U is transitive. 2
(1)
(2) xRy F(x) F(y).
Proof of (2). By (). 2
(2)
(3) F : U > >> V .
Proof of (3). F is surjective, by denition. Suppose it was not injective. Then pick x U
-minimal such that there is a y U with y ,= x, such that F(x) = F(y). Let y witness this,
i.e., x ,= y, F(x) = F(y). Ill show:
Rx = Ry,
which is a contradiction, since this would imply that x = y, by extensionality of (U, R).
For the direction from left to right, let zRx. Then F(z) F(x) = F(y) = FRy. So
there is z
t
Ry such that F(z) = F(z
t
). Supposing zRy, it follows that z ,= z
t
, since z
t
Ry. So we
20 CHAPTER 2. SETS
have zRx, and there is a w with z ,= w and F(z) = F(w), namely w := z
t
. This contradicts the
R-minimality of x with this property.
The converse is entirely analogous: Let zRy. Then F(z) F(y) = F(x) = FRx, so
there is a z
t
Rx such that F(z
t
) = F(z). Suppose zRx. Since z
t
Rx, this implies that z ,= z
t
. So
we have: z
t
Rx, and there is a w (as witnessed by z) such that z
t
,= w and F(z
t
) = F(w). This
again contradicts the R-minimality of x. 2
(3)
(4) F(x) F(y) xRy
Proof of (4). Let F(x) F(y). Since F(y) = FRy, there is x
t
Ry such that F(x) = F(x
t
).
But since F is injective, x
t
= x. So xRy. 2
(4)
This nishes the proof of existence.
Turning to uniqueness: By the Recursion theorem, there is only one F satisfying (). So it
suces to show:
(5) If W is transitive and G : (U, R)

(W, W), then for all u U, G(u) = G( u) [ uRu.
Proof of (5). Fix u U. If vRu, then G(v) G(u), since G is R-preserving, so this shows that
G( u) [ uRu G(u). For the opposite direction, let v G(u). So v G(u) W, which by
transitivity of W implies that v W. So v has a pre-image under G, as G is onto. So let v U
be such that v = G( v). Since G( v) G(u), it follows that vRu. So v = G( v), where vRu, which
shows that G(u) GRu, as wished. 2
(5)
This nishes the proof.
2.3 Ordinal Numbers and Innity
2.3.1 Well-orders
Denition 2.3.1 (Linear orders, well-orders, connectedness). If R U
2
, then (U, R) is a (strict)
linear order if for all x, y, z U:
1. (xRx) (irreexivity),
2. xRyRz xRz (transitivity),
3. xRy x = y yRx (connectedness).
A linear order (U, R) is a well-order if R is well-founded. It is a strong well-order if R is strongly
well-founded.
While were at it, Ill call a class U connected if (U, U
2
) is connected. This means that
for all x, y U, either x = y, x y, or y x.
Clearly, if (U, R) is linear, then it is extensional. For let x, y U have the same R-
predecessors. Suppose x ,= y. Then xRy or yRx. In the rst case, it follows that xRx,
contradicting irreexivity, and in the second it follows that yRy, again contradicting irreex-
ivity. So the following denition makes sense:
Denition 2.3.2. If (U, R) is a strong well-order, then let otp(U, R), the order-type of (U, R),
be the uniquely determined transitive class such that (, )

= (U, R) (by Mostowskis
Isomorphism Theorem).
2.3. ORDINAL NUMBERS AND INFINITY 21
The natural numbers (N, <) form a well-order. What are the natural numbers? This is a
somewhat philosophical question, but if they exist, then since they form a strong well-order,
they have an order-type, otp(N, <). It is this transitive class that we consider to be the class
of natural numbers. The innity axiom says that it is a set. I will shortly introduce this axiom
more formally.
2.3.2 Ordinals
Lets assume the Foundation Axiom from now on. So the system I work in from now on
is ZF
F
.
I am giving some proofs in these notes that I omit in the lecture...
Denition 2.3.3. On = otp(s) [ s is a well-order. On is the class of ordinal numbers.
So the class of ordinal numbers is the collection of order-types of set-sized well-orders.
Lemma 2.3.4 (Counting Lemma). Let S = (U, R) be a strong well-order. Let = otp(S), and
let F : S

(, ) be the collapse. Then (, ) is a strong well-order. Moreover, if x U,
then, letting u
x
= Rx, it follows that
Fu
x
: u
x
, Ru
x
)

F(x), F(x))
is the Mostowski collapse of the well-order u
x
, Ru
x
). In particular, F(x) On.
Proof. It is obvious that , ) is a strong well-order, being isomorphic to one. Fix x U. It
is easily checked that, letting r
x
= Ru
x
, u
x
, Ru
x
) is a (strong) well-order. I claim that for
y u
x
, F(y) = F(z) [ zr
x
y. Since F is the Mostowski-isomorphism of (U, R), we know that
F(y) = F(z) [ zRy. So it suces to see that Ry = r
x
y. The direction from right to
left is trivial, so suppose zRy. Since r
x
= RRx, to see that zr
x
y, we only need to know
that yRx, but we do know that, since y u
x
. By the way, the reader is invited to check that
Ru
x
= R(u
x
)
2
. In any case, Fu
x
satises the denition of the Mostowski collapse of u
x
, r
x
),
and hence it is the collapse. By denition, F(x) = FRx = Fu
x
, which completes the
proof.
In the future, , , . . . will be reserved for ordinals.
Corollary 2.3.5. If S is a strong well-order, then otp(S) is a transitive subclass of On.
Corollary 2.3.6. On is transitive.
Proof. If On, then = otp(s), for some well-order s, and by the previous corollary, otp(s)
On.
Lemma 2.3.7. Let U be transitive and well-founded. Then:
1. U is connected i (U, U) is a strong well-order.
2. If U is connected, then U = otp(U, U).
Proof. Clear.
Corollary 2.3.8.
On = x [ x is transitive and connected.
Corollary 2.3.9. If is transitive and connected, then On.
Proof. (, ) is a strong well-order, and = [(, )[ On.
22 CHAPTER 2. SETS
Lemma 2.3.10. Let be transitive, connected and well-founded. Let U be transitive with U .
Then either U = or U .
Proof. Assume that U ,= , so that U . By well-foundedness of , let be -minimal in
U.
(1) = / U.
Proof of (1). If U, then U. 2
(1)
(2) U .
Proof of (2). Let U. Since U , . Since too, and since is connected, one
of the following must be true: , , or = . By (1), it cannot be the case that ,
since U. = is impossible, since / U, while U. So the only possibility is that .
2
(2)
(3) U.
Proof of (3). is minimal in U. 2
(3)
So U = .
Denition 2.3.11. For ordinals , , set
<
< = .
Corollary 2.3.12. .
Proof. = = , as is transitive, and = trivially implies .
= is transitive, connected and well-founded, so that the previous lemma applies.
Lemma 2.3.13. If and are ordinals, then or .
Proof. If not, then , and , . This means that
and .
But is transitive (being transitive is the same as being -closed, and the intersection of
R-closed sets is R-closed, for any binary relation). So again, the previous lemma applies, showing
that , and also . So . This contradicts the well-foundedness of
and .
Lemma 2.3.14. (On, <) is a strong well-order, and [(On, <)[ = On.
Proof. By the previous lemma, (On, <) is connected, and weve known for a while that On is
transitive. Since <= On, we also know that it is set-like. We have to show (On, <) is well-
founded. So let , = A On. Pick A. If is not <-minimal in A, then consider a = A .
This is a nonempty subset of , and is well-founded, being an ordinal. Pick , )-minimal
in a. Then is <-minimal in A: If < , then / a, because . So / A, or else
A = a. The second part follows since On is transitive.
Corollary 2.3.15. On / V.
Proof. Otherwise On = [(On, <)[ On, contradicting the Foundation Axiom.
Corollary 2.3.16. Let S be a strong well-order. Then [S[ On or [S[ = On. So, writing for
On and expanding the meaning of <, in the obvious way, [S[ .
Proof. We know already that [S[ On. So if [S[ ,= On, then [S[ On, as On is strongly
well-founded, transitive and connected, and since [S[ is transitive see Lemma 2.3.10.
Lemma 2.3.17. Let A . Then (A, < A) is a strong well-order. Letting F : (A, <
A) be its collapse, it follows that for all A, F() , and that . Hence the
name collapse.
Proof. (A, < A) is a strong well-order: It is set-like by Separation, and it is a well-order, because
(On, <) is.
Turning to the main claim: If not, let be the minimal counterexample in (A, < A). By
denition, F() = F() [ A On. By assumption, F(). So let A be
such that = F(). Then < , and is also a counterexample, contradicting the minimality
of .
Finally, it follows that : If not, then = ran(F). Let A be s.t. F() = .
Then = F() < , contradicting s well-foundedness.
Denition 2.3.18. s = .
I shall often write + 1 instead of s.
Lemma 2.3.19.
1. On.
2. s On.
3. < s.
4. < s .
Proof. For 4.:
< s .
Corollary 2.3.20. s = min [ > .
Proof. Clearly, = s, which shows . Vice versa, if < s, then , so that
,> (or else , contradicting the well-foundedness of On).
2.3.3 The Natural Numbers
Denition 2.3.21.
= On
On [ 0 < s .
is the class of natural numbers. We write m, n, . . . for members of .
So is the least class (wrt. inclusion) thats transitive, well-founded, connected, closed under
x sx and that contains as a member, since it is the intersection of all such classes. It follows
that either = On, or On.
Denition 2.3.22. The axiom of innity (Innity) says: V.
24 CHAPTER 2. SETS
Well not assume this axiom for now. So we continue working in ZF
F
.
Lemma 2.3.23.
1. .
2. n = sn .
Denition 2.3.24. 0 := , 1 := s0, 2 := s1, . . .
Lemma 2.3.25. If n , then either n = 0, or there is an m n with sm = n.
Proof. If not, then picking a least counterexample gives an s-closed member of which contains
0, contradicting the minimality of .
Lemma 2.3.26 (Induction Scheme). If A satises:
1. 0 A,
2. n A sn A,
then A = .
Proof. Otherwise, let n = min( A). Since 0 A, n ,= 0. n = sm, for some m. Then m < n,
so that m A, by minimality of n. But then, by 2., n = sm A after all.
So is the smallest class which contains 0 and is s-closed. For if X is such a class, then
X = , by induction.
There are simple versions of the recursion theorem for , which can be used to dene the
usual arithmetical operations on the natural numbers. This will be done in the exercises.
Denition 2.3.27. Say that x and y are equinumerous i there is an bijection f : x > >> y.
Well write x y for this relation.
Lemma 2.3.28. is an equivalence relation.
Denition 2.3.29. x is nite i there is an n with x n.
Lemma 2.3.30. If m < n < , then m , n. Hence, if x is nite, then there is a unique n with
x n.
Proof. If not, let m be least such that there is an n > m with m n. Clearly, m, n ,= 0.
Let m = s m, n = s n. Let f : m > >> n.
Case 1: f( m) = n.
Then f m : m > >> n, contradicting the minimality of m.
Case 2: f( m) < n.
Then let f(k) = n. Dene g : m n by:
g(l) =
_
f(l) if l ,= k,
f( m) if l = k.
Clearly, g : m > >> n, contradicting the minimality of m.
Denition 2.3.31. For a nite set x, let x, the cardinality of x, be the uniquely determined
n < with x n.
Lemma 2.3.32. Let u n < . Then otp(u, < u)) n. In particular, u is nite. If u n,
then otp(u, <)) < n.
Proof. The rst part of the lemma we know already: Its Lemma 2.3.17. For the second part,
assume the contrary. Let n be minimal such that there is a counterexample u. Then n = sm.
It follows that m u, for otherwise u m, and we already know that then otp(u, < u)) m.
But then, u m m. By minimality of n, it follows that p = otp(u m, <)) < m. Let F
be the Mostowski collapse of u m, <). By the Counting Lemma, F = G(u m), where G is
the collapse of u, <). So G(m) = F(j) [ j u m = p, so that otp(u, <)) = sp. But since
p < m, it follows that sp < n, so that u wasnt a counterexample after all.
Lemma 2.3.33. If v is nite and u v, then u is nite, too. If moreover u v, then u < v.
Proof. For the rst part, let f : v > >> n, where n = v . Let u
t
= fu n. Then
s = u
t
, (u
t
)
2
) is a well-order. Let g : s

otp(s) be its collapse. Then otp(s) n . So
g fu : u > >> otp(s) , proving that u is nite.
For the second part: u
t
n. So m = otp(u
t
, < u
t
)) < n. In particular, u
t
= m.
Corollary 2.3.34. A set a is nite if and only if it is bounded in , i.e., if there is an
n with a n. Also, in that case, a = otp(a, a)).
Proof. Clearly, if a n, then otp(a, <)) n, and so a n, showing that a is nite. It also
follows that otp(a, <)) = a. This is because the collapse of a, <) is a bijection between a and
otp(a, <)), which is a natural number (at most n).
Vice versa, assume that a is nite but unbounded in . Pick such an a of minimal cardinality
n . Let f : a > >> n , n = m + 1. Let a
t
= f
1
m. Then a
t
= a f
1(m), and
fa
t
: a
t
> >> m. By minimality of a, it follows that a
t
is bounded, say by p . But then a
is bounded by p s(f
1
(m)) .
Lemma 2.3.35. If u is nite and f : u >> v, then v is nite, and v u.
Proof. Let g : u > >> n = u. Then g induces a well-order <
on u, dened by x <
y
g(x) g(y). Dene h : v > >> u
t
u by letting h(m) = min
u,<
)
(l u [ f(l) = m). Then
g h : v > >> a n. Letting k : a, <)

l, <) be the collapse, it follows that l n, so that
k g h : v > >> l n, showing that v u.
Another way to put the previous Lemma is to say that the nite sets satisfy Replacement.
I leave it to the reader to check that it is possible, using the recursion theorem, to dene the
usual arithmetic operations like addition and multiplication and exponentiation on the natural
numbers.
Lemma 2.3.36. Let u and v be nite sets.
1. If u and v are disjoint sets, then u v = u +v.
2. u v = u v.
3. If u is nite, then T(u) = 2
u
.
Proof. 1.) By induction on n, show: If u is nite, v = n and u v = , then u v = u +v. For
this, use the fact that + is dened correctly, i.e., that p + (sq) = s(p +q).
2.) By induction on n, show: If u is nite and v is nite with v = n, then u v = u v.
Again, use that p (sq) = (p q) +p.
3.) By induction on n, show: If u is nite with u = n, then T(u) = 2
u
.
Corollary 2.3.37. If u is nite, then T(u) V.
26 CHAPTER 2. SETS
So the nite sets satisfy Power Set also. We can dene recursively:
Denition 2.3.38.
V
0
= ,
V
n+1
= T(V
n
),
V
=
_
n<
V
n
.
Lemma 2.3.39.
1. V
n
is a nite, transitive set.
2. m < n = V
m
V
n
.
3. V
n
= n.
4. Every x V
is nite.
5. If u V
is nite, then u V
.
Proof. 1.-3. are shown by simultaneous induction on n, and 4. follows easily. To see 5., let u V
be nite. Dene F : u by letting F(x) be the least n with x V

n
. Let a = ran(F). So
a V, and actually, F V. Since F : u >> a and u is nite, so is a. So a is a nite set of
natural numbers, hence a is bounded. Let n be such that a n. Then u V
n
, which means
that u V
n+1
V
.
Lemma 2.3.40 (ZF
F
). The following are equivalent:
1. Innity,
2. There is a set which is not nite,
3. V ,= V
.
Proof. 1. = 2.: V is not nite, or else it would be bounded in itself.
2. = 3.: Every member of V
is nite.
3. = 1.: Let x be -minimal in V V
. Then x V
. As before, let F : x be dened by

letting F(y) be the least n such that y V
n
. Let a = ran(F) V. Then a must be unbounded
in , or else x V
would follow. So
a = .
2.3.4 Trees
Trees are very useful in many contexts, such as the analysis of sets of real numbers. Also, the
natural representation of a formula and its subformulae, their subformulae, and so on, is a tree.
Trees can also be used to serve as a model for the space in which a derivation is searched for.
Such trees are called search trees, and I shall make use of such trees later, when proving the
Completeness Theorem for the Tait calculus. Somewhat longer trees occur in combinatorial set
theory as well, and they will come up much later, in the third sequel of this lecture course. The
current subsection should really be called short sequent trees.
Denition 2.3.41. If f is a function with domain and x is a set, let f
x := f x, ).
Denition 2.3.42. A (sequent) tree T on a set X is a set of functions (sequences) f : n X
closed under initial segments. So every member of T is a function whose domain is a natural
number, and whose range is contained in X. Moreover, for all f T and for all natural numbers
m, fm T. If T is a tree and p T, then
succ
T
(p) := x [ p
x T.
T is locally nite if for every p T, succ
T
(p) is nite. A (conal/innite) branch of a tree T is
a function b : X such that for every n , bn T.
Denition 2.3.43. Let T be a tree on X, T V. Given s, t T, I write t >
T
s if s t.
T is well-founded if the structure T, >
T
) (the ordering is reversed here!) is well-founded.
2
Otherwise, the tree is ill-founded.
Lemma 2.3.44. Let u, r) be a structure with r u
2
. Let u be well-ordered by some relation
<
u
. Then u, r) is ill-founded if and only if there is a function f : u such that for all
n < , f(n + 1)rf(n).
Proof. Suppose f : u is such a decreasing sequence. Then the set f has no r-minimal
element, so that u, r) is ill-founded.
Vice versa, suppose u, r) is ill-founded. Now let ,= a T have no >
T
-minimal member.
It is now possible to dene a sequence s
n
[ n < ) recursively, by letting s
0
be the -minimal
member of a. Since a has no r-minimal member, let s
1
be the -minimal member of a such
that s
1
rs
0
. In general, having dened s
n
, we can let s
n+1
be -minimal such that s
n+1
a and
s
n+1
r
T
s
n
. Then f(n) := s
n
is as wished.
Theorem 2.3.45 (Konig, ZF
). Every innite, locally nite tree on a well-ordered set X has

an innite branch.
Proof. Let T be an innite, locally nite tree on X. Consider the set I consisting of those f T
such that the set T
f
:= g T [ f g is innite. It follows that I is a tree: It is obviously
closed under initial segments. Moreover, I is ill-founded: Firstly, I ,= because I, as T is
innite, by assumption. Secondly, if f I, then T
f
is innite. But
T
f
= f
_
xsucc
T
(f)
T
f
x
is a partition of the innite (since f I) set T
f
into nitely many pieces (as T is locally nite,
so that succ
T
(f) is nite). Hence, there must be an x succ
T
(f) such that T
f
x
is innite, in
other word, such that f
x I. So f is not >
T
-minimal in I. Since f was an arbitrary member
of I, this shows that I is ill-founded. T inherits a well-ordering from X in a canonical way: Let
<
X
be a well-order on X. For example, if s, t T such that s ,= t, let s t if dom(s) < dom(t),
or if dom(s) = dom(t) and for the least l such that sl ,= tl, s(l) <
X
t(l). The reader is
invited to check that this is a well-ordering on T. The restriction of this well-ordering to I is of
course a well-order on I, so Lemma 2.3.44 is applicable. It shows that I has an innite branch
>
T
-decreasing sequence, and such a sequence will generate a branch of T in the obvious way.
2
The structure T, <
T
is always well-founded. Note also that since T is a set, the scheme expressing that
T, >
T
is well-founded is actually equivalent to the formula saying that every nonempty subset of T has a
>
T
-minimal member. In this case, I take T, >
T
is well-founded to stand for this formula.
28 CHAPTER 2. SETS
2.4 Cardinal Numbers and Choice
Denition 2.4.1. The axiom of choice (AC) says that every set has a cardinality. I.e., for every
set x, there is a function f and an ordinal such that f : x > >> . The cardinality of a set
x is the least such that there is a function f : x > >> .
Note that if V = V
, then every set is nite, so that the axiom of choice automatically holds.
So AC is only interesting if we assume Innity. I also want to assume Power Set and Foundation.
So the current theory is ZF
F
+ (Innity) + (PowerSet), which is usually referred to as ZF.
Theorem 2.4.2 (ZF). The following are equivalent:
1. AC,
2. every set can be well-ordered (i.e., for all x, there is an r x
2
such that x, r) is a well-
order),
3. every set of nonempty sets has a choice function, i.e., for all x, if y x y ,= , then
there is a function f : x
x, so that y x f(y) y.
3
Proof. 1 = 2: Given a set x, let f : x > >> . Then x inherits a well-order <
x
from , < )
via y <
x
z f(y) < f(z).
2 = 3: Let x be a set of non-empty sets. Let
x, r) be a well-order. Dene F : x
x
by letting F(y) be the r-minimal member of y. F is denable, so that F = F ((
x) x) V.
3 = 1: Let x be a set. We have to nd f and such that f : x > >> . Consider
z := T(x) (so here, the Power Set axiom is used). This is a set of nonempty sets, so there
is a choice function g : z
z. Now we can dene by recursion a function F : On x as

follows:
F() =
_
g(x F) if dened
x otherwise
Let =

[ F() ,= x. Then F : > > x. So F
1
is a function, which means that
= F
1
x V, i.e., := < . It follows that F : >> x as well, so f := F (x ) V
is a bijection from onto x.
Denition 2.4.3. x,
x
) is a partial order if
x
x
2
is reexive and transitive. It is a partial
order in the strict sense if in addition, it is anti-symmetric, meaning that if u
x
v and v
x
u,
then u = v. A chain in a partial order x,
x
) is a set c x such that for any u, v c, u
x
v
or v
x
u. Such a partial order is chain-closed if for any chain c x, c has a bound in x,
x
),
i.e., there is a b x such that for all u c, u
x
b. If x,
x
) is a partial order, write <
x
for the
binary relation on x dened by letting u <
x
v i u
x
v and u ,= v. m x is maximal in x,
x
)
if there is no u x with u >
x
m.
Note: If x,
x
) is a partial order in the strict sense, then x, <
x
) is irreexive and transitive.
Irreexivity is clear by denition. Now assume that u <
x
v and v <
x
w. Then u
x
v and
v
x
w, so that u
x
w, as
x
is transitive. To see that u <
x
w, it must be veried that u ,= w.
But if it were the case that u = w, then wed have u
x
v
x
u, so since x,
x
) is a partial
order in the strict sense, it would follow that u = v, contradicting the assumption that u <
x
v.
Note also that there may be many maximal elements in a chain-closed partial order.
Lemma 2.4.4 (Zorns Lemma, ZF). The following are equivalent:
3
This is usually the ocial denition of the axiom of choice.
2.4. CARDINAL NUMBERS AND CHOICE 29
1. Whenever x,
x
) is a chain-closed partial order in the strict sense (with x ,= ), it has a
maximal element.
2. The axiom of choice.
Proof. 2 = 1: By the axiom of choice, x a well-ordering R of x. Dene a function F : On
x x by setting:
F() =
_
_
_
The R-minimal member u of x with
z F z <
x
u if this exists,
x otherwise.
Let = [ F() ,= x. Then f := F is injective, so < . Moreover, if < < ,
then f() <
x
f(), by denition of F. Also, if is a limit ordinal and , then .
This is because F is a chain in x,
x
), so that it has a bound y. But then this bound is a
strict bound, i.e., y >
x
z, for all z F - this is because if z = F(), for some < , then
z <
x
F( + 1)
x
y, so that z <
x
y. So is a successor ordinal, say = + 1. Then F() is a
maximal element of x, <
x
): Since F( + 1) = x, by denition, there is no y with y >
x
F().
1 = 2: Lets verify the axiom of choice in the form stating every set of nonempty sets has
a choice function. So let u be a set of nonempty sets, and consider the set F consisting of the
functions whose domain is a subset of u and which are choice functions for their domain (which
naturally is a set of nonempty sets). Consider the partial order F, (F
2
)). This is a partial
order in the strict sense, and it is chain closed, as is easily veried (the point is that the union of a
chain of functions is again a function). So let f F be maximal. The claim is that dom(f) = u.
For if not, then let y u dom(f). Since y ,= , pick z y, and set f
t
= f z, y). Then
f
t
F, and f f
t
, contradicting the maximality of f.
So dom(f) = u, which means that f is a choice function for u.
The system ZF + AC is called ZFC. ZF
is ZF
F
+ (Innity), and ZFC
is ZF
+ AC.
30 CHAPTER 2. SETS
Chapter 3
Semantics
In this chapter, I am returning to what was done in the rst chapter. There, I introduced
rst order languages, without talking about the meaning of expressions in such languages. I
augmented these languages with an appropriate notion of proof, but still everything was basically
about words, that is, sequences of symbols. In this chapter, I want to introduce interpretations
of languages, so that, e.g., a formula x gets a meaning, namely that there exists an x such
that is true of x.
I will work in the theory ZF
, although one can work with less in many cases. For example,
ZF
will suce for many purposes, though at the cost of having to deal with proper classes
instead of sets.
It is easy to see that in ZF
, given a set X, one can construct a canonical free halfgroup

generated by X (this will be done in the exercises). Fixing a language L = C, P, F, #), the
collection of L-terms, Term
1
and the collection of L-formulae, Fml
1
, then form sets. The same
is true for the set of Tait-formulae, Fml
T
1
.
3.1 Models
Denition 3.1.1. A model for L is a set of the form/ = X, I) with X ,= and I : CPF
V so that
1. For c C, I(c) X. I will write c
,
for I(c).
2. For P P, I(P) X
#(P)
. I will write P
,
in place of I(P).
3. For F F, I(F) : X
#(F)
X. As expected, F
,
will stand for I(F).
It is sometimes convenient to pretend that = is a member of P and set
=
,
:= x, x) [ x [/[.
Ill write [/[ for X. This is called the universe, or the domain of the model /.
Finally, if P P and

P is the corresponding symbol in the Tait-alphabet of L, then set:
P
,
:= [/[
#(P)
P
,
,
and viewing ,= as =, the corresponding holds:
,=
,
:= x, y) [ x [/[, y [/[ and x ,= y.
31
32 CHAPTER 3. SEMANTICS
I will use the Recursion Theorem in order to dene when a model satises a formula (with
some assignment of its free variables). The relation used will be that of being a subformula. The
point is that it is well-founded. The reader will gladly verify the following:
Lemma 3.1.2. Let v, s) be a well-founded system (i.e., s v
2
and s is well-founded). Let u, r)
be such that r u
2
, and let f : u, r) v, s) be a homomorphism, meaning that if xry, then
f(x)sf(y). Then u, r) is well-founded.
I will state the following two lemmas without proofs:
Lemma 3.1.3 (Unique Readability of Terms). If t is a term in a xed language L, then precisely
one of the following possibilities holds true:
1. t = v
m
, for some (unique) variable v
m
,
2. t = c, for some (unique) constant symbol c,
3. there is a (unique) function symbol F and a (unique) list t
0
, . . . , t
n1
, where n = #(F), such
that t = F(t
0
, . . . , t
n1
). In this case, t
0
, . . . , t
n1
will be called the immediate subterms of
t.
Lemma 3.1.4 (Unique Readability of Formulae). If is a formula in a xed language L, then
precisely one of the following possibilities holds:
1. = P(t
0
, . . . , t
n1
), for a (unique) P P and a (unique) sequence t
0
, . . . , t
n1
with
n = #(P), such that each t
i
is a term. For convenience, I allow the possibility P = = here.
2. there is a (unique) formula (the immediate subformula of ) such that = ,
3. there are (unique) formulae
0
,
1
(the immediate subformulae of ) such that = (
0
1
),
4. there are (unique) formulae
0
,
1
(the immediate subformulae of ) such that = (
0
1
),
5. there is a variable v
m
and a formula (the immediate subformula of ) such that =
v
m
,
6. there is a variable v
m
and a formula (the immediate subformula of ) such that =
v
m
.
Of course, the obvious version of this lemma for Tait-formulae holds as well.
The relation over which I am going to dene whether a formula holds in a given model, with
a given assignment of its free variables, will be that of being an immediate subformula. It is
clear that the formulae in a given language form a set. The map sending a formula to its length
maps an immediate subformula of a formula to a strictly smaller natural number. So since the
natural numbers, together with their natural ordering are well-founded, the formulae, with the
immediate subformula relation, are also well-founded, by Lemma 3.1.2, and the corresponding
statement is true of the terms, with the immediate subterm relation.
Lemma 3.1.5. Let / be a model of some language. An assignment in / is a nite partial
function from the set of variables to [/[. Let Assign
,
be the set of assignments in /.
3.1. MODELS 33
Denition 3.1.6 (Evaluation of Terms). Let / be a model of a xed language L. By recursion
on the immediate subterm relation, dene a function I whose domain is the set Term
1
, so that
for every term t, I(t) : Assign
,
/ [/[.
1
Writing t
,
(a) for (I(t))(a), the denition
proceeds as follows.
t
,
(a) =
_
_
a(v
m
) if t = v
m
and v
m
dom(a),
[/[ if t is a variable not in dom(a),
c
,
if c C,
F
,
(t
,
0
(a), . . . , t
,
n1
(a)) if t = F(t
0
, . . . , t
n1
) and t
,
i
(a) M, for all i < n,
[/[ if for an immediate subterm s of t, s
,
(a) = [/[.
Denition 3.1.7. If a Assign
,
, v
m
is a variable and b [/[, then let a(
v
m
/
b
) be the
assignment with domain dom(a) v
m
dened by:
(a(
v
m
/
b
))(x) =
_
a(x) if x dom(a) and x ,= v
m
,
b if x = v
m
.
Now Im ready to dene when a formula is true in a model, given an assignment of its free
variables:
Denition 3.1.8 (Satisfaction). Let / be an L-model. Dene a function F = F
,
: Fml
T(Assign
,
) by recursion on the immediate subformula relation. F() will be the set of assign-
ments under which is satised in /. Write FV() for the set of free variables occurring in
. If is an atomic formula of the form P(t
0
, . . . , t
n1
), where P P (including the possibility
that P = =), #(P) = n and t
0
, . . . , t
n1
are terms, then let
F() = a Assign
,
[ t
,
0
(a), . . . , t
,
n1
(a)) P
,
FV() dom(a).
Boolean combinations of formulae are dealt with using the following stipulations:
F(
0

1
) := F(
0
) F(
1
),
F(
0

1
) := F(
0
) F(
1
),
F() := a Assign
,
[ (FV() dom(a) a / F()).
Turning to quantications, set:
F(v
m
) := a Assign
,
[ b / a(
v
m
/
b
) F(),
F(v
m
) := a Assign
,
[ b / a(
v
m
/
b
) F().
Finally, I write
/ [= [a]
to express that a F
,
().
This denition can be viewed as dening / [= [a] also for Tait-formulae, by treating

P,

,=
as elements of P. The -step of the denition is then obsolete.
Denition 3.1.9. If a, b Assign
,
and u is a nite set of variables, then a
u
b if au = bu.
If t is a term, I will write a
t
b if a
FV(t)
b. Analogously, if is a formula, then I write a
b
if a
FV()
b.
1
(I(t))(a) = |M| will be used to indicate that there is a free variable in t which is not in the domain of a.
Lemma 3.1.10. Let be a formula, and let a, b Assign
,
be such that FV() dom(a)
dom(b) and a
b. Then
/ [= [a] / [= [b].
Proof. This is a typical example of a proof by induction on . So Ill prove that the lemma holds
true for all atomic formulae, and then that if it holds true for formulae
0
and
1
, then it also
holds true for (
0

1
), (
0

1
),
0
, v
n
0
, v
n
0
, where v
n
does not occur as a bound
variable in
0
. It follows that it holds true for all formulae, since otherwise, one could pick a
formula contradicting it, which is minimal with respect to the immediate subformula relation
(the point is that this relation is well-founded). Then cannot be atomic, and it cannot have
an immediate subformula either, which completes the proof. This is the principle of induction.
As a rst step to proving the lemma, one has to prove the following, by induction on terms:
() If t Term
1
and a, b Assign
,
with FV(t) dom(a) dom(b) and a
t
b, then
t
,
(a) = t
,
(b).
I leave the proof of () (which follows the same pattern as the rest of the proof of the lemma)
to the reader. Assuming (), Ill now prove the lemma. First, suppose = P(t
0
, . . . , t
n1
) is an
atomic formula, and a, b Assign
,
is as in the statement of the lemma. Then
/ [= P(t
0
, . . . , t
n1
)[a] t
,
0
(a), . . . , t
,
n1
(a)) P
,
t
,
0
(b), . . . , t
,
n1
(b)) P
,
/ [= P(t
0
, . . . , t
n1
)[b],
noting that since a
b, it follows that FV(t

i
) dom(a) dom(b) and a
t
i
b for every i < n,
as FV(t
i
) FV(). So () is applicable here, showing that t
,
i
(a) = t
,
i
(b), for i < n.
The induction step covering Boolean combinations is trivial, as is mostly the case. Since this
is the rst occurrence of a proof by induction of formulae, Ill carry it out anyway: Suppose
0
and
1
are formulae for which the claim of the lemma holds. Let a and b be assignments such
that FV(
0

1
(
0
1
)
b. Since FV(
0

1
) = FV(
0
) FV(
1
),
it follows that FV(
i
i
b, for i < 2. So since the lemma holds for
0
and
1
, by assumption, we can apply it to
0
, a, b, and also to
1
, a, b. This yields:
/ [=
i
[a] / [=
i
[b],
for i < 2. So we get:
/ [= (
0

1
)[a] (/ [=
0
[a]) and (/ [=
1
[a])
(/ [=
0
[b]) and (/ [=
1
[b])
/ [= (
0

1
)[b].
The -step is analogous, and the -step is equally trivial. Lets check the v
m
-step. So let
= v
m
, where v
m
does not occur as a bound variable in , and the lemma holds for . Let
a, b be as in the statement of the lemma, for v
m
. Note that then the statement of the lemma
holds for , a(
v
m
/
c
), b(
v
m
/
c
), whenever c [/[. The point is that FV() (dom(a) v
m
)
(dom(b) v
m
). So we know (inductively) that
/ [= [a(
v
m
/
c
)] / [= [b(
v
m
/
c
)]
3.2. CONSEQUENCE 35
in this situation. So assume now that
/ [= v
m
[a].
By denition, this means that there is a c [/[ such that
/ [= [a(
v
m
/
c
)].
By the above, this is equivalent to
/ [= [b(
v
m
/
c
)].
In particular, there is a c with this property, so that
/ [= v
m
[b].
The converse follows by swapping a and b (or, in fact, the above steps work both ways).
The remaining v
m
-step is analogous and left to the reader.
So whether or not a model satises a formula, given an assignment, depends only on the
behavior of the assignment on the formulas free variables. This motivates the following two
denitions.
Denition 3.1.11. Let / be a model. A function a : Var [/[ is called a full assignment
in /. The set of full assignments in / is denoted by Assign
+
,
. If is a formula and a is a full
assignment in /, then
/ [= [a] / [= [a FV()].
Denition 3.1.12. If is a formula with free variables v
m
0
, . . . , v
m
n1
(in increasing order),
then I write
/ [= [b
0
, . . . , b
n1
]
to express that / [= [a], where a(v
m
i
) = b
i
, for all i < n. If is a sentence, i.e., if has no
free variables, then / [= i / [= [].
The reader is invited to check the following:
Lemma 3.1.13. If is a Tait-formula and a Assign
+
,
, for some L-model /, then
/ [= [a] / ,[= [a].
3.2 Consequence
Denition 3.2.1 (Logical Consequence). If is a set of L-formulae or of Tait-formulae, then
let FV() =
FV(). If / is an L-model and a Assign

+
,
, then write / [= [a] to express
that / [= [a], for every . If is an L-formula, then
[= for every L-model / and every a Assign
+
,
,
/ [= [a] = / [= [a].
If = , I write [= instead of [= .
Chapter 4
Same same... but dierent
By now, I have introduced two notions of consequence: Given a theory T in a xed language L,
and an L-sentence , I have dened what it means that is provable from T, T
T
, and what
it means that is a logical consequence, T [= . These are two very dierent concepts, the rst
one being based on purely syntactical concepts such as proofs, and the second one on semantical
concepts such as models. In the following two sections, I am going to show that these concepts
nevertheless are equivalent, for countable T! The direction T = T [= is known as the
Correctness Theorem, while the converse, T [= = T is the Completeness Theorem.
4.1 Correctness
Given a nite set of (Tait-)formulae , let
_
denote the disjunction of all formulae occurring
in (in some xed enumeration). The order in which the formulae are listed is irrelevant. The
meaning is that / [= (
_
)[a] i there is a formula such that / [= [a]. So in fact, it
makes sense to take the disjunction of arbitrarily many formulae, but we wont need to do this
here.
Theorem 4.1.1 (Soundness). If is a (nite) set of Tait-formulae such that
T
, then
[=
_
.
Proof. By induction on rules.
Case 1:
T
because of tertium non datur.
In this case, there is a formula such that and . Now let / be an arbitrary
L-model and a an arbitrary full assignment in /. Then either / [= [a] or / ,[= [a]. But the
latter is equivalent to / [= [a]. So / [= [a] or / [= [a], in particular, / [=
_
.
Case 2:
T
( ) because
T
and
T
, i.e., by applying the -rule.
Let / be an arbitrary L-model with an appropriate assignment a. Inductively, / [=
_
(
)[a] and / [=
_
( )[a]. If there is a formula such that / [= [a], then
/ [=
_
[a], and were done. If not, then since there is a formula
t
in such that
/ [=
t
[a], this must be true for
t
= . Analogously, it follows that / [= [a]. But if / [= [a]
and / [= [a], then this means that / [= ()[a]. In particular, / [=
_
(()), as
claimed.
Case 3:
T
( ) because of the -rule. So we have
T
or
T
.
Inductively, we know that given / and a as before, / [=
_
( )[a] or / [=
_
(
)[a]. Lets assume the rst possibility holds true. If there is some such that / [= [a],
then clearly / [=
_
()[a], so were ne. Otherwise, we know that / [= [a]. But by
37
38 CHAPTER 4. SAME SAME... BUT DIFFERENT
denition of the satisfaction relation, this clearly implies that / [= ( )[a], which, in turn,
implies that / [=
_
( ( ))[a]. The other case is entirely symmetric.
Case 4:
T
v
m
because of the -rule. I.e.,
T
(
v
m
/
v
n
), where v
n
/ FV().
Let appropriate / and a be given. So FV( v
m
) dom(a). Inductively, we know
that for any c [/[,
/ [=
( (
v
m
/
v
n
))[a(
v
n
/
c
)],
since this holds for any assignment. If there is a such that / [= [a(
v
n
/
c
)], for some c,
then since a(
v
n
/
c
)
a (as v
n
is not a free variable of ), it follows that / [= [a] (by Lemma
3.1.10), and hence / [=
_
( v
m
). Now assume the contrary. Then for all c [/[,
/ [= (
v
m
/
v
n
)[a(
v
n
/
c
)].
This is of course equivalent to saying that for all c [/[,
/ [= [a(
v
m
/
c
)].
This means precisely that / [= v
m
[a], as wished.
Case 5:
T
v
m
because of the -rule. I.e.,
T
(
v
m
/
t
), for some t Term
1
.
Suppose /, a are arbitrary, as above. Inductively, we know that / [=
_
((
v
m
/
t
))[a].
As usual, if there is a such that / [= [a], then / [= [a] and were done. Otherwise,
we know that / [= (
v
m
/
t
)[a]. This means that / [= [a(
v
m
/
t
M
(a)
)]. In particular, there is
some c [/[ such that / [= [a(
v
m
/
c
)], namely c = t
,
(a). So by denition, this means that
/ [= v
m
[a], and in particular, / [=
_
( v
m
).
This can be used to prove correctness. Since the traditional Correctness Theorem refers to
regular formulae, not Tait-formulae, a translation procedure is needed:
Denition 4.1.2. Dene a function
T
by induction on formulae as follows:
(P(t
0
, . . . , t
n1
))
T
:= P(t
0
, . . . , t
n1
),
(t
0
=t
1
)
T
:= t
0
=t
1
,
()
T
:= (
T
),
(
0

1
)
T
:= (
T
0

T
1
),
(
0

1
)
T
:= (
T
0

T
1
),
(v
m
)
T
:= v
m
(
T
),
(v
m
)
T
:= v
m
(
T
).
Lemma 4.1.3. Given any L-model / and any L-formula ,
T
is a Tait-formula of L such
that for any full assignment a in /,
/ [= [a] / [=
T
[a].
Proof. Exercise.
There is an obvious translation procedure in the other direction which will be developed as
an exercise, too. So the expressive power of formulae and Tait-formulae is the same.
Denition 4.1.4. For a set Fml
1
, let
T
=
T
[ . If moreover, is an L-formula,
write to mean that
T

T
.
Theorem 4.1.5 (Correctness). Let Fml
1
be a set of formulae. Then
= [= .
4.2. COMPLETENESS 39
Proof. Assume , i.e.,
T

T
. Let
0
be a nite set of formulae each of which is an
identity axiom or a member of
T
such that
T
(
0

T
). Let / be a model and a a full
assignment such that / [= [a], or, equivalently, / [=
T
[a]. Since every model satises the
identity axioms, it follows that / [=
0
[a]. By soundness, we know that / [=
_
(
0
T
)[a].
But since / [=
T
[a], there is no
0
such that / [= [a]. So the only possibility is that
/ [=
T
[a], or equivalently, / [= [a].
4.2 Completeness
To prove the Completeness Theorem, let L be a xed language. Let be a set of L-formulae
which is at most countable (meaning countable or nite). Fix an enumeration

of . Let be
a nite set of L-formulae. It is then easy to see that there are at most countably many terms
which can be built up from constants, function symbols occurring in , and from arbitrary
variables. Let

t := t
i
[ i < ) be an enumeration of these terms which I shall call the relevant
terms.
The aim is to show that if [=
_
, then
T

0
.
1
The argument will proceed as
follows: Supposing there is no such
0
, we will build a tree searching for such a
0
and such
a derivation. Since by assumption, there is no such
0
, the search will be unsuccessful, thus
producing an innite tree. The tree will be locally nite, and hence will have an innite branch.
Such a branch will determine a term model M and a full assignment a from hold, and in which
none of the formulae in holds (under the assignment a).
To dene the tree, view as a nite sequence

which lists all (of the nitely many) formulae
in in some order. If

is some nite list of formulae, let the redex of

, R(
), be the formula
with least index in

which is not atomic.

r
will be the sequence of formulae obtained from

by omitting R(
). The tree S := S
t
we are about to construct is a labeled tree. We construct it
together with a labeling function : S
<
Fml
1
. The tree will be at most 2-splitting, meaning
that every node will have at most two immediate successors. It will be constructed by recursion
on its levels, together with . To start o,
(S
) S and () =

.
Now assume s S. If ran((s)) is an axiom according to tertium non datur (i.e., there is a
Tait-formula such that , ran((s))), then s has no successors in S.
Now suppose this is not the case. Then it is determined which of s
0 and s
1 are in S, and
what the values (s
0) and (s
1) are, by the following clauses. For ease of notation, assuming

(sn) has been dened for all n dom(s), set
(s) =
ndom(s)
ran((sn)).
(S
id
) If R((s)) = , then s
0 S. If X :=
(s)) ,= , then let be the member of X

thats mentioned rst in the xed enumeration of , and set (s
0) = (s)
. If, on
the other hand, X = , then set (s
0) = (s).
(S
) If R((s)) is of the form (

0
1
), then both s
0 and s
1 are in S, and (s
i) = (s)
r
i
,
for i < 2.
(S
) If R((s)) is of the form v

m
, then s
0 S, s
1 / S, and (s
0) = (s)
r
(
v
m
/
v
n
),
where n is least such that v
n
does not occur in (s) .
2
1
Here, I used the notation = { | }, for a set of Tait-formulae.
2
It might happen that all variables occur in . In this case, one can replace every occurrence of v
l
in with
v
2l
, simultaneously, in the beginning. Denoting the resulting sets by
and
, this renaming does not change

(S
) If R((s)) is of the form (

0

1
), then s
0 S and s
1 / S. Let be the rst

formula of
0
and
1
which does not occur in Y :=
(s), if such a formula exists, and

set x := (s)
r
in that case. If there is no such formula, then x := (s)
r
.Let be the
rst formula mentioned in the enumeration of such that does not occur in Y , if there
is such a formula. In this case, let x
t
:= x
. If there is no such formula, let x

t
:= x.
Finally, set (s
0) := x
t
R((s)).
(S
) If R((s)) is of the form v

m
, then s
0 S and s
1 / S. Let t be the rst relevant

term mentioned in the enumeration such that (
v
m
/
t
) /
(s), if such a term exists. In

that case, let x = (s)
r
(
v
m
/
t
), otherwise x = (s)
r
. Further, let be the rst formula
mentioned in the enumeration of such that /
(s), if such a formula exists. In that

case, let x
t
= x
, otherwise, x
t
= x. Finally, (s
0) = x
t
R((s)).
In a way, this tree can be viewed as searching simultaneously for a proof of (
0
) , for some
nite
0
, and (in case this search fails) for a model of
_
.
Lemma 4.2.1 (Syntactical Main Lemma, Sch utte). If S is well-founded, then for every s S,
there is a nite set
s
such that
T

s
ran((s)).
Proof. Suppose this were not the case. Since S is well-founded by assumption, there must be a
>
S
-minimal counterexample (that is, a <
S
-maximal one).
Case 1. s has no successors in S.
This is only possible if ran((s)) is an axiom according to tertium non datur. So in this case,
T
ran((s)), which means that
s
:= is as wished. So s is no counterexample and this case is
excluded.
Case 2. Case 1 fails.
Then we have to distinguish subcases according to the nature of R((s)).
Case 2.1 R((s)) = .
Then s
0 is the sole successor of s in S. By minimality, we know that there is a nite subset
0
such that
T

s
0
ran((s
0)).
If X :=
(s)) ,= , then letting be the member of X thats mentioned rst in the

xed enumeration of , (s
0) = (s)
. So letting
s
:=
s
0
, it follows that
s
ran((s)) =
s
0
ran((s
0)), so that
T

s
ran((s)).
If X = , then (s
0) = (s), so setting
s
:=
s
0
will do.
Case 2.2. R((s)) = (
0

1
).
In this case, s
0 and s
1 both are successors of s in S, and (s
0) = (s)
r
0
, (s
1) =
(s)
r
1
. So by minimality, we know that there are
s
0
and
s
1
such that
T

s
0

ran((s)
r
0
) and
T

s
1
ran((s)
r
1
). Let
s
=
s
0

s
1
. Then by monotonicity
(the structural rule, saying that if
T
and
t
, then
T

t
, which was proven in the
exercises), we have that
T
(
s
ran((s)
r
))
0
and
T
(
s
ran((s)
r
))
1
.
So by the -rule, it follows that
T

s
ran((s)
r
) (
0

1
).
anything about our assumptions. So if |=
, also
|=
, and for a nite subset

0
,
T

0
i
, where
0
is the result of renaming the variables occurring in
0
as described. So we may assume
that there always is such a variable. Remember that (s) is always a nite sequence, by denition.
But (
0

1
) = R((s)), which is omitted in (s)
r
. So this just means that
T

s
ran((s)),
as desired.
Case 2.3 R((s)) = v
m
.
Then s
0 is the only successor of s in S. Moreover, letting n be minimal such that v

n
does not occur in ran((s)) , by denition, (s
0) = (s)
r
(
v
m
/
v
n
). So by minimality,
there is a nite set
s
0
such that
T

s
0
ran((s)
r
) (
v
m
/
v
n
). Since v
n
is a
new variable, the -rule yields that
T

s
0
ran((s)
r
) v
m
, or, in other words, that
T

s
0
ran((s)). So setting
s
:=
s
0
does the job.
Case 2.4 R((s)) = (
0

1
).
Then s
0 S is the unique successor of s in S. Let Y :=
(s).
If at least one of
0
and
1
does not occur in Y , then let be the rst one. Let be the
rst formula in the enumeration of such that does not occur in Y , if it exists. Then by
denition, (s
0) = (s)
r
(
0

1
). Inductively, there is a set
s
0
such that
T

s
0
ran((s
0)) (
0

1
), i.e.,
T

s
0
ran((s)
r
) (
0

1
)
. .
=ran((s))
.
So letting
s
:=
s
0
, we have
T

s
ran((s)) (
0

1
) ,
so that it follows from an application of the -rule that
T

s
ran((s)),
as desired. If there is no as above, one can just let
s
=
s
0
.
If both
0
and
1
occur in ran((s)), then there are again two subcases. The rst is that
there is a formula in such that does not occur in Y . In this case, let be the rst
one mentioned in the enumeration of . Then (s
0) = (s)
r
(
0

1
). In this case, we
have inductively that
T

s
0
ran((s)
r
) (
0

1
). Letting
s
:=
s
0

yields then that
T

s
ran(s), as desired, without applying the -rule. Finally, if there
is no such formula , then (s
0) = (s)
r
(
0

1
), and so, inductively we have that
T
0
ran((s)
r
) (
0

1
)
. .
=ran((s))
, so that
s
:=
s
0
will do the job.
Case 2.5 R((s)) = v
m
.
Then s
0 is the unique successor of s in S. Let t be the rst relevant term mentioned

in the enumeration such that (
v
m
/
t
) /
(s), if such a term exists. In that case, let x =

(s)
r
(
v
m
/
t
), otherwise x = (s)
r
. Further, let be the rst formula mentioned in the enu-
meration of such that /
(s), if such a formula exists. In that case, let x

t
= x
,
otherwise, x
t
= x. Finally, (s
0) = x
t
R((s)).
Inductively, we know that there is a nite set
s
0
such that
T

s
0
ran((s
0)).
This assumption is weakest if both t and exist (by the structural rule/monotonicity). So in
this case, we have
T

s
0
ran((s)
r
) R((s))
. .
=(s)
(
v
m
/
t
).
So setting
s
:=
s
0
, this is the same as to say that
T

s
ran((s)) (
v
m
/
t
)
(if doesnt exist, then taking
s
=
s
0
will do). Applying the -rule yields
T

s
ran((s)),
since v
m
= R((s)) ran((s)). Again, applying the -rule is only necessary if t is dened.
The following is an adaptation of the classical lemma, dealing with pure logic, to logic with
identity.
Lemma 4.2.2 (Semantical Main Lemma, Sch utte). If S is ill-founded and contains the identity
axioms, then there is a model / and a full assignment a in / such that / [= ( )[a].
Proof. Let f : 2 be a conal branch of S. Let (f) :=

ran((fn)) [ n < . The
following properties are crucial:
1. If an atomic formula occurs in ran((fm)), for some m < , then it also occurs in
ran((fn)), for every n > m.
2. (f).
3. If a non-atomic formula occurs in some ran((fm)), then there is an n m such that
= R((fn)).
4. If a formula of the form (
0
1
) occurs in (f), then there is an i < 2 such that
i
occurs
in (f).
5. If a formula of the form v
m
occurs in (f), then there is a variable v
n
such that (
v
m
/
v
n
)
occurs in (f).
6. If a formula of the form (
0

1
) occurs in (f), then
0
,
1
(f).
7. If a formula of the form v
m
occurs in (f), then (
v
m
/
t
) occurs in (f), for every relevant
term t.
Lets check these properties: 1 is clear, because atomic formulae are never discarded. Only
the redex is sometimes thrown away, and this is never an atomic formula, by denition.
To see 2, note that it cannot be the case that only for nitely many n, R(fn) is of the
form v
m
, (
0
1
), or undened. This is because otherwise, there is some n such that for all
m > n, R((fm)) either of the form (
0
1
) or v
i
, which means that (f(m+1)) is either
of the form (s)
r
0
, (s)
r
1
, or (s)
r
(
v
i
/
v
j
). By Konigs Lemma, for every formula ,
there is a number b = b
such that for every sequence

0
,
1
, . . . ,
l
) such that
0
= ,
1
is an immediate subformula of , and so on, it follows that l < b (this is because the tree
of such sequences is locally nite and well-founded, hence nite). It is obvious now that, letting
b =
ran((fn))
b
, the redex of (f(n +b)) would have to be undened, a contradiction.

Property 3 is clear: Suppose a non-atomic formula occurs in ran((fm)). If there were no
n m such that R((fn)) = , then for every n m, there would be a non-atomic formula
which occurs before the rst occurrence of in (fn). But clearly, the number of non-atomic
formulae occurring in (fn) before the rst occurrence of would have to be strictly decreasing
(for n m), which is a contradiction.
Property 4 is clear by denition of S.
Property 5 follows from 3: If v
m
(f), then let k be such that v
m
ran((fk)).
By 3, there is some n > k such that R((fn)) = v
m
, and then, by denition, (
v
m
/
v
l
)
ran((f(n + 1))), for some l.
Property 6 is clear: Let m be least such that R((fm)) = (
0
1
). Then
0
ran((f(m+
1))). Also, by denition, (
0

1
) ran((f(m + 1))). So by 3, there is some m
t
> m such
that again, R((fm
t
)) = (
0

1
). This time, it follows that
1
ran((f(m
t
+ 1))).
A similar argument shows that property 7 holds.
() If is an atomic formula in (f), then / (f).
Proof of (). This is because is also an atomic formula. Suppose both and were in
(f). Once an atomic formula is in (fm), it stays in (meaning it is in (fn), for all n > m).
So pick n large enough so that , (fn). Then ran((fn)) is a tertium non datur
axiom, so that fn is a terminal node of S, and hence, f cant be a conal branch. 2
()
Now dene a relation on relevant terms:
t
0
t
1
t
0

,=t
1
(f).
This is an equivalence relation:
Reexivity: Since v
0
(v
0
= v
0
) is an identity axiom, this formula is in , so v
0
(v
0
= v
0
) =
v
0
(v
0
,= v
0
) (f), by 2. But then by 7, it follows that t ,= t (f), in other words,
that t t.
The proofs of symmetry and transitivity follow the same pattern and are left to the reader.
Moreover, is a congruence relation over the predicates:
(C1) If P P

P with #(P) = n, s
0
, . . . , s
n1
and t
0
, . . . , t
n1
are relevant terms with s
i
t
i
,
for all i < n, then P(s) (f) i P(
t) (f).
Proof. Suppose P(s) (f). By assumption, s
i

,=t
i
(f), for all i < n. Moreover, the Tait
negation of the adequate identity axiom (of the category congruence over predicates) is in ,
and hence in (f):
v
0
v
1
. . . v
n1
v
n
v
n+1
. . . v
2n1
(((v
0
= v
n
) (v
1
= v
n+1
) . . . (v
n1
= v
2n1
))
(P(v
0
, v
1
, . . . , v
n1
) P(v
n
, v
n+1
, . . . , v
2n1
))).
By property 7, the following instance of this formula is in (f):
(s
0
= t
0
s
1
= t
n
. . . s
n1
= t
n1
P(s
0
, s
1
, . . . , s
n1
) P(t
0
, t
1
, . . . , t
n1
)).
By property 4, at least one conjunct of this formula must be in (f). By assumption s
i

,=t
i

(f), for all i < n, so that by (), s
i
=t
i
/ (f). Analogously, P(s
0
, s
1
, . . . , s
n1
) (f)
by assumption, so that P(s
0
, s
1
, . . . , s
n1
) / (f). So the only possibility thats left is that
P(t
0
, t
1
, . . . , t
n1
) (f), which is what we want. The converse is proven analogously.
One has to use the corresponding identity axiom of the category congruence over functions
for the following claim:
(C2) If F F with #(F) = n and s
0
, . . . , s
n1
and t
0
, . . . , t
n1
are relevant terms with s
i
t
i
,
for all i < n, then F(s) F(
t).
Now Im ready to dene the desired model /: Let [/[ consist of the -equivalence classes
of relevant terms. So, writing [t] for the -equivalence class of the relevant term t,
[/[ = [t] [ t is a relevant term.
Set:
F
,
([t
0
], . . . , [t
n1
]) = [F(t
0
, . . . , t
n1
)], for F F such that F occurs in ,
[t
0
], . . . , [t
n1
]) P
,
P(t
0
, . . . , t
n1
) (f), for P P,
c
,
= [c], for relevant c C.
The interpretation of function symbols and constant symbols which dont occur in is
irrelevant and can be prescribed in some simple, arbitrary way.
The correctness of these denitions follows from (C1) and (C2). Dene a full assignment a
in / in the obvious way:
a(v
m
) := [v
m
].
It follows by induction on terms t that
t
,
(a) = [t].
Further, it follows by induction on formulae that:
(D) (f) = / [= [a].
Starting with the atomic case: If = P(t
0
, . . . , t
n
), where P P and (f), then by
denition, / [= P([t
0
], . . . , [t
n1
]), which means precisely that / [= P(t
0
, . . . , t
n1
)[a], by
the previous remark. This argument works for the case P = = as well. Now assume =
P(t
0
, . . . , t
n1
), and (f). By (), / (f). So P(t
0
, . . . , t
n1
) / (f), which means
that by denition of P
,
, [t
0
], . . . , [t
n1
]) / P
,
. So by denition, [t
0
], . . . , [t
n1
])

P
,
, or,
in other words, / [=

P(t
0
, . . . , t
n1
)[a]. Again, this argument works for P = = as well.
In the rest of the argument, the properties 4-7 will be used.
So suppose (
0

1
) (f). By the denition of Tait-negation, this means that (
0

1
) (f). By property 4, let i < 2 be such that
i
(f). Inductively, we know that
/ [=
i
[a]. But then / [= (
0

1
)[a] also.
Now assume that v
m
(f), which means that v
m
(f). By 5, pick v
n
such that
(
v
m
/
v
n
) (f). Inductively, / [= (
v
m
/
v
n
)[a]. In particular, / [= v
m
[a].
If (
0

1
) (f), then this means that (
0

1
) (f), and by 6, this, in turn,
means that both
0
and
1
are in (f), so that inductively, / [=
0
[a] and / [=
1
[a].
This means, of course, that / [= (
0

1
)[a].
If v
m
(f), i.e., v
m
(f), then by 7, (
v
m
/
t
) (f), for every relevant term t.
So inductively, / [= (
v
m
/
t
)[a], for every relevant term t. In other words, / [= [a(
v
m
/
t
M
(a)
)],
for every relevant term t (this is a general fact that can easily be shown by induction on terms).
But since t
,
(a) = [t], and [/[ consists precisely of such [t], this previous statement can be
rephrased as saying that / [= [a(
v
m
/
b
)], for every b [/[. This means that / [= v
m
[a],
as wished.
Now it is an immediate consequence that / [= [a], since if , i.e., (f)
(by property 2), it follows from (D) that / [= [a]. To see that / [= , note that =
ran((f0)) (f). So given , () (f), so that by (D), / [= [a].
Theorem 4.2.3 (Completeness). Let be a countable set of L-formulae and be a formula
such that [= . Then .
Proof. We may assume only variables with even index occur in . Having said this, the
reader may forget about this technicality, see the beginning of the denition of the search tree.
Also, replace by
T
=
T
[ and by
T
, so that we are actually dealing with
Tait-formulae. Let
be , together with the identity axioms.

4.3. COMPACTNESS 45
Let

t enumerate the at most countably many terms which can be built up using constants
and function symbols occurring in
, and variables, the relevant terms. Fix also an

enumeration

of
. Let S = S

],
t
, be the search tree determined by these objects, together
with its labeling function.
S must be well-founded: Otherwise, it is ill-founded, so that the Semantical Main Lemma 4.2.2
applies, giving a model / and a full assignment a with / [= (
)[a], which contradicts

the assumption that [= . So S is well-founded, which means that the Syntactical Main Lemma
4.2.1 applies. This gives, for every s S, a nite set
such that
T

s
ran((s)). For
s = , this means:
is nite and
T

0

T
, where
0
is a nite set of formulae
each of which is an identity axiom or a member of
T
. By denition, this means that .
4.3 Compactness
Work in a xed countable language.
Theorem 4.3.1. If [= , then there is a nite
0
such that
0
[= .
Proof. By the completeness theorem, . This means there is a nite set
0
and a
nite set of identity axioms such that
T
(
0
) . By correctness, this means that
0
[= . But since every model satises the identity axioms, as the identity symbol is always
interpreted by the true identity relation, this implies that
0
[= .
Denition 4.3.2. Let be a set of formulae. is consistent, con(), is the statement that
there is a formula such that it is not the case that .
has a model if there is a model / and a full assignment a in / such that / [= [a].
Theorem 4.3.3. A set of formulae is consistent if and only if it has a model.
Proof. Suppose is not consistent. Then for every , . If there were a model of ,
witnessed by / and a, then it would follow that / [= [a], for every , by correctness. This
is of course a contradiction, since it would follow that / [= [a] and / [= [a], which means
that / ,[= [a]. Vice versa, if there is no model of , then [= , for any formula , by at.
So by completeness, .
Theorem 4.3.4. If every nite subset of has a model, then has a model.
Proof. Suppose has no model. Then is inconsistent. Let be a sentence. Then and
. So there is a nite subset
0
of such that
0
and
0
; I tacitly used the
monotonicity of the deductive calculus here, namely that if
0

1
and
0
, then
1
. It
follows that
0
has no model, a contradiction.
4.4 Ultralters and Ultraproducts
In this section, Ill develop a useful tool for constructing models. As a motivation, suppose
/
i
[ i I) is a sequence of models of a (not necessarily countable) xed language L. We want
to construct a model / such that a sentence is true in / if for almost every i I, is
true in /
i
. The rst step is to make the meaning of for almost every i I precise. This is
done using the concept of ultralters.
4.4.1 Filters and ultralters
Denition 4.4.1. Let I ,= be a set. A lter on I is a set F of subsets of I with the following
properties:
1. / F, I F.
2. If X F and X Y I, then Y F.
3. If X, Y F, then X Y F.
A lter F on I is an ultralter on I if in addition, for every X I, X F or I X F.
F is a principal lter on I if there is an i I such that F = X I [ i X. Clearly, a
principal lter is an ultralter on I.
Lemma 4.4.2. Let F be a lter on I. Then the following are equivalent:
1. F is an ultralter on I.
2. There is no lter G on I with F G. (I.e., F is a maximal lter.)
Proof. 1. = 2.: Suppose there were such a G, Pick X G F. Since F is ultra, I X F.
Since F G, I X G. So = X (I X) G, a contradiction.
2. = 1.: Let X I. Suppose neither X F nor (I X) F. Let G = Y I [ Z
F Z X Y . Obviously, F G. Also, X GF. But G is a lter: I G, since I F G.
Suppose G. Then pick Z F such that Z X = . This means that Z (I X). Since
Z F, it would follow that I X F, a contradiction.
Let Y
0
, Y
1
G. We have to see that Y
0
Y
1
G. Pick Z
0
, Z
1
F such that X Z
0
Y
0
,
XZ
1
Y
1
. Then X(Z
0
Z
1
) (Y
0
Y
1
), and since F is a lter, Z
0
Z
1
F. So Y
0
Y
1
G.
Finally, its obvious that if Y
0
G and Y
0
Y
1
I, it follows that Y
1
G. For if Z witnesses
that Y
0
G, then it also witnesses that Y
1
G.
So altogether, we have that F G, and G is a lter on I, a contradiction.
Theorem 4.4.3 (Ultralter Theorem, ZFC). Suppose I ,= and F is a lter on I. Then there
is an ultralter U on I such that F U.
Proof. Let F be given. Consider the partial order P consisting of all lters on I which contain
F, ordered by inclusion. This is obviously a partial order in the strict sense. Moreover, it is
chain-closed: If C [P[ is a chain, then
C is a lter: Clearly, I
C. /
C, since / H,
for any H C. Given X, Y
C, there is one H C with X, Y H (here it is crucial that C

is a chain!). Hence, XY H, and so, XY
C. Finally, if X
C and X Y I, then
there is H C with X H, so Y H, so Y
C. So by Zorns lemma, there is a maximal

member of P, and by the previous lemma, this maximal member is an ultralter on I.
4.4.2 Ultraproducts
Denition 4.4.4. Let /
i
[ i I) be a sequence of models of a xed language L. Then
iI
[/
i
[ is the set of functions f with domain I such that for all i I, f(i) [/
i
[.
Given a formula with free variables v
m
0
, v
m
1
, . . . , v
m
n1
(in increasing order) and functions
f
0
, . . . , f
n1

iI
[/
i
[, let
[[[f
0
, . . . , f
n1
][[ = i I [ /
i
[= [f
0
(i), . . . , f
n1
(i)].
4.4. ULTRAFILTERS AND ULTRAPRODUCTS 47
Let F be a lter on I. Given f, g
iI
[/
i
[, dene
f
F
g i I [ f(i) = g(i) F.
So f
F
g i [[f = g[[ F.
Lemma 4.4.5. In the situation of the previous denition, if P is an n-ary predicate symbol
(including the possibility that P = =), then for any f
0
, . . . , f
n1
, g
0
, . . . , g
n1

iI
[/
i
[ with
f
0

F
g
0
, . . . , f
n1

F
g
n1
, it follows that
[[P[f
0
, . . . , f
n1
][[ F [[P[g
0
, . . . , g
n1
][[ F.
Proof. The situation is symmetric, so it suces to prove the direction from left to right. Let
X = [[P[f
0
, . . . , f
n1
][[ [[f
0
= g
0
[[ . . . [[f
n1
= g
n1
[[. Since F is a lter, X F. For
i X, since i [[P[f
0
, . . . , f
n1
][[, it follows that /
i
[= P[f
0
(i), . . . , f
n1
(i)]. Since i [[f
0
=
g
0
[[ . . . [[f
n1
= g
n1
[[, it follows that f
0
(i) = g
0
(i), . . . , f
n1
(i) = g
n1
(i). So /
i
[=
P[g
0
(i), . . . , g
n1
(i)]. This means that i [[P[g
0
, . . . , g
n1
[[. So X F, X [[P[g
0
, . . . , g
n1
[[,
which shows that [[P[g
0
, . . . , g
n1
][[ F, as claimed.
Analogously:
Lemma 4.4.6. If H is an n-ary function symbol, then for any f
0
, . . . , f
n1
, g
0
, . . . , g
n1

iI
[/
i
[ with f
0

F
g
0
, . . . , f
n1

F
g
n1
, it follows that
[[H(f
0
, . . . , f
n1
) = H(g
0
, . . . , g
n1
)[[ F,
meaning that the set of i with H
,
i
(f
0
(i), . . . , f
n1
(i)) = H
,
i
(g
0
(i), . . . , g
n1
(i)) is in F.
Moreover,
F
is an equivalence relation on
iI
[/
i
[: Given f, g, h
iI
[/
i
[, it follows
that f
F
f, since [[f = f[[ = I F, f
F
g = g
F
f, since [[f = g[[ = [[g = f[[, and
if f
F
g and g
F
h, it follows that [[f = g[[ [[g = h[[ [[f = h[[ F, so that f
F
h.
This, together with the previous lemma, allows us to turn

iI
[/
i
[ into a model, giving the
correctness of the following denition.
Denition 4.4.7. In the notation of the previous denition, dene a model / :=
iI
/
i
/F
as follows. [/[ consists of the
F
-equivalence classes of functions f

iI
[/
i
[. Write [f]
F
for the
F
-equivalence class of f. Then the interpretations of the symbols in / are given by:
c
,
:= [c
,
i
[ i I)]
F
, for c C,
[f
0
]
F
, . . . , [f
n1
]
F
) P
,
[[P[f
0
, . . . , f
n1
][[ F, for P P with #(P) = n,
F
,
([f
0
]
F
, . . . , [f
n1
]
F
) = [F
,
i
(f
0
(i), . . . , f
n1
(i)) [ i I)]
F
.
The following is the Hauptsatz on ultraproducts.
Theorem 4.4.8 ( Los). Let /
i
[ i I) be a sequence of models of a xed language L, I ,= . Let
U be an ultralter on I. Let be an L-formula with free variables v
m
0
, . . . , v
m
n1
(in increasing
order) and let f
0
, . . . , f
n1

iI
[/
i
[. Then
_
iI
/
i
/U [= [[f
0
]
U
, . . . , [f
n1
]
U
]
_
([[[f
0
, . . . , f
n1
][[ U) .
Proof. Set / :=
iI
/
i
/U.
First, one shows by a straightforward induction on terms t that if

f
U
g, it follows that
[t
,
i
(
f(i)))]
U
= [t
,
i
(g(i)))]
U
. Knowing this, the theorem is now proved by induction on .
Its clear for atomic formulae.
Now consider the case = (
0

1
). Then
/ [= [[f
0
]
U
, . . . , [f
n1
]
U
] / [=
0
[[f
0
]
U
, . . . , [f
n1
]
U
] and / [=
1
[[f
0
]
U
, . . . , [f
n1
]
U
]
[[
0
[f
0
, . . . , f
n1
][[ U and [[
0
[f
0
, . . . , f
n1
][[ U
[[
0
[f
0
, . . . , f
n1
][[ [[
0
[f
0
, . . . , f
n1
][[ U
[[(
0

1
)[f
0
, . . . , f
n1
][[ U
[[[f
0
, . . . , f
n1
][[ U
The -step is similar, with one little additional argument: The rst relevant observation is
that [[(
0
1
)[f
0
, . . . , f
n1
][[ = [[
0
[f
0
, . . . , f
n1
][[[[
1
[f
0
, . . . , f
n1
[[. The additional argument
is that if X Y U, then X U or Y U. The reason is that if both possibilities failed, it
would follow that both I X U and I Y U, since U is an Ultralter. So (I X) (I Y ) =
I (X Y ) U. But then = (X Y ) I (X Y ) U, a contradiction.
It becomes more obvious how the ultralter property of U is used in the -step (and actually,
since (
0

1
) is equivalent to (
0

1
), the -step can be reduced to the -step and the
-step. We get:
/ [= [[f
0
]
U
, . . . , [f
n1
]
U
] / ,[= [[f
0
]
U
, . . . , [f
n1
]
U
]
[[[f
0
, . . . , f
n1
][[ / U
I [[[f
0
, . . . , f
n1
][[ U
[[[f
0
, . . . , f
n1
][[ U.
The -step is of interest, since the axiom of choice is needed here. Consider a formula of the
form v
m
, where wlog m is larger than the index of all the free variables of that formula.
/ [= v
m
[f
1
, . . . , f
n1
] there is a [f]
U
[/[ such that / [= [[f
0
]
U
, . . . , [f
n1
]
U
, [f]
U
]
there is an f
iI
[/
i
[ such that [[[f
0
, . . . , f
n1
, f][[
. .
]]v
m
[f
0
,...,f
n1
]]]
U.
= [[v
m
[f
0
, . . . , f
n1
][[ U.
For the converse, assume X := [[v
m
[f
0
, . . . , f
n1
][[ U. For i X, the set
X
i
:= c [/
i
[ [ /
i
[= [f
0
(i), . . . , f
n1
(i), c]
is nonempty. For i / X, let X
i
= [/
i
[. By the axiom of choice, there is a function f :
I

iI
X
i
. I.e., for all i I, f(i) [/
i
[, and if i X, then f(i) X
i
. Then
clearly, X = [[[f
0
, . . . , f
n1
, f][[ U, which, by induction hypothesis, means that / [=
[[f
0
]
U
, . . . , [f
n1
]
U
, [f]
U
]. So in particular, / [= v
m
[[f
0
]
U
, . . . , [f
n1
]
U
].
The remaining case, v
m
, can in principle be reduced to v
m
, and is left to the reader.
4.4. ULTRAFILTERS AND ULTRAPRODUCTS 49
4.4.3 Compactness revisited
Here is a direct proof of the general version of the compactness theorem.
Theorem 4.4.9 (Compactness Theorem, AC). If is a set formulae of a xed language such
that every nite subset of has a model. Then has a model.
Proof. Let I be the collection of nite subsets of . For I, let /
be a model and a
a full
assignment in /
with /
[= [a
]. For I, let
) = I [ .
Clearly, for , I, it follows that
) ) = ).
Let
F = X I [ I ) X.
It follows that F is a lter (its the lter generated by ) [ I). For example, / F, since
if X F, there is a I s.t. ) X. To see that if X, Y F, then also X Y F,
pick , I s.t. ) X, ) Y . Then ) = ) ) X Y F. By the ultralter
theorem, let U F be an ultralter on I. Let
/ :=
I
/
/U.
For a variable v
n
, dene a
v
n

I
[/
[ by:
a
v
n
() = a
(v
n
).
Dene a full assignment a in / by:
a(v
n
) := [a
v
n
]
U
.
It follows that / [= [a]: Let , where has the free variables v
m
0
, . . . , v
m
n1
(in in-
creasing enumeration). Then for all I with , /
[= [a
], which means that

/
[= [a
(v
m
0
), . . . , a
(v
m
n1
)]. So ) [[[a
v
m
0
, . . . , a
v
m
n1
][[. By the construction
of U, ) U, so [[[a
v
m
0
, . . . , a
v
m
n1
][[ U, which means (by the Los Theorem) that / [=
[[a
v
m
0
]
U
, . . . , [a
v
m
n1
]
U
]. Since a(v
m
i
) = [a
v
m
i
]
U
, this means precisely that / [= [a].
This theorem has the general completeness theorem as a consequence:
Theorem 4.4.10 (Completeness, AC). Assume [= , where is a set of formulae in
an arbitrary xed language. Then .
Proof. By the general compactness theorem, there is a nite set
0
such that
0
[= . For
otherwise, every nite subset of would have a model, which by compactness would
mean that has a model, but this would contradict the assumption that [= . Now
we can apply the completeness theorem to
0
, yielding
0
, in particular, .
Chapter 5
Incompleteness
This exposition of incompleteness phenomena follows [End72].
5.1 Arithmetic with Exponentiation and Representability
The language of number theory is given by: C = 0, P = <, F = S, +, , E, #(<) = 2,
#(S) = 1, #(+) = 2, #() = 2, #(E) = 2.
I will use inx notation for the function symbols + and , and for the predicate symbol <. Of
course, the canonical model for this language is N := , 0, <, S, +, , E) where S is the successor
function and E is exponentiation of natural numbers. We will use a fragment of the theory of
this model, i.e., a subset of the set of sentences true in N. Following [End72], this fragment is
called A
E
. It consists of the following axioms:
(S1) v
0
S(v
0
) = 0
(S2) v
0
v
1
(S(v
0
) = S(v
1
) v
0
= v
1
)
(L1) v
0
v
1
(v
0
< S(v
1
) (v
0
< v
1
v
0
= v
1
))
(L2) v
0
v
0
< 0
(L3) v
0
v
1
(v
0
< v
1
v
0
= v
1
v
1
< v
0
)
(A1) v
0
v
0
+ 0 = v
0
(A2) v
0
v
1
(v
0
+S(v
1
) = S(v
0
+v
1
))
(M1) v
0
v
0
0 = 0
(M2) v
0
v
1
(v
0
S(v
1
) = (v
0
v
1
) +v
0
(E1) v
0
E(v
0
, 0) = S(0)
(E2) v
0
v
1
E(v
0
, S(v
1
)) = E(v
0
, v
1
) v
0
For every natural number n, there is a canonical name in the language of A
E
, namely S
n
(0).
This is called the numeral for n.
51
52 CHAPTER 5. INCOMPLETENESS
Denition 5.1.1. A relation R
n
is represented by a a formula (v
0
, . . . , v
n1
) with the free
variables listed, if for every tuple a
0
, . . . , a
n1
)
n
,
a
0
, . . . , a
n1
) R A
E
(
v
0
/
S
a
0(0)
) . . . (
v
n1
/
S
a
n1
(0)
), and
a
0
, . . . , a
n1
) / R A
E
(
v
0
/
S
a
0(0)
) . . . (
v
n1
/
S
a
n1
(0)
)
R is representable if there is a formula which represents R.
A function f :
n
is represented by a formula if that formula represents the relation
a
0
, . . . , a
n1
, f(a
0
, . . . , a
n1
)) [ a
0
, . . . , a
n1
) dom(f)
which I shall refer to as the graph of f. f is representable if there is a formula which represents
the graph of f.
A formula (v
0
, v
1
, . . . , v
n
) functionally represents the function f :
n
if for all natural
numbers m
0
, . . . , m
n1
,
A
E
v
n
((
v
0
/
S
m
0(0)
) (
v
n1
/
S
m
n1
(0)
) v
n
= S
f(m
0
,...,m
n1
)
(0)).
I shall often write (S
a
0
(0), . . . , S
a
n1
(0)) in place of (
v
0
/
S
a
0(0)
) . . . (
v
n1
/
S
a
n1
(0)
), to save
some space. If a relation R as above is representable, then there is a computer program for an
idealized computer which can check, given a, whether a R or not. This works as follows: Let
witness that R is representable. Given a, let
t
= (S
a
0
(0), . . . , S
a
n1
(0)) (this is a dierent
formula, actually a sentence). The program can now produce the search trees searching for a
proof of A
E

t
and A
E

t
simultaneously. One of the search trees must be nite,
since either A
E
or A
E
. When this occurs, the program halts and outputs yes if the
search tree for
t
was completed and no if it was the other tree. It is less obvious that every
relation that is computable in this way is also representable.
In the following, I will produce some facts about the class of representable relations and
functions.
Denition 5.1.2. A formula (v
0
, . . . , v
n1
) is numeralwise determined by A
E
if for every tuple
a
0
, . . . , a
n1
) of natural numbers, either
A
E
(S
a
0
(0), . . . , S
a
n1
(0))
or
A
E
(S
a
0
(0), . . . , S
a
n1
(0)).
Also, if / is a model of some language, and (v
0
, . . . , v
n1
) is a formula of that language,
then the relation dened in / by is
a
0
, . . . , a
n1
) [/[
n
[ / [= [a
0
, . . . , a
n1
].
In this case, this set is lightface-denable over /. A relation R [/[
i
is boldface-denable
over / if there is a formula as above and members a
i
, a
i1
, . . . , a
n1
of [/[, with i < n, such
that
R = a
0
, . . . , a
i1
) [ / [= [a
0
, . . . , a
n1
].
In this case, I say that R is dened by in the parameters a
i
, . . . , a
n1
over /.
The following lemma describes the connection between denability in N and representability
of a relation.
5.1. ARITHMETIC WITH EXPONENTIATION AND REPRESENTABILITY 53
Lemma 5.1.3. Let (v
0
, . . . , v
n1
) be a formula. There is a relation which is represented by
if and only if is numeralwise determined by A
E
. In that case, the relation represented by is
the relation dened by over N.
Proof. Its clear by the denition of what it means that represents a relation that this implies
is numeralwise determined.
Conversely, suppose is numeralwise determined. Let R = a
0
, . . . , a
n1
)
n
[ A
E

(S
a
0
(0), . . . , S
a
n1
(0)). Clearly, represents R. Moreover, R is dened by over N: Let
a := a
0
, . . . , a
n1
)
n
be given. If a R, then A
E
(S
a
0
(0), . . . , S
a
n1
(0)), and since
N [= A
E
, it follows that N [= (S
a
0
(0), . . . , S
a
n1
(0)), which means that N [= [a
0
, . . . , a
n1
],
since (S
m
(0))
N
= m, for every natural number m. If a / R, then by denition, A
E
,
(S
a
0
(0), . . . , S
a
n1
(0)), so since is numeralwise determined, A
E
(S
a
0
(0), . . . , S
a
n1
(0)),
which means that N [= (S
a
0
(0), . . . , S
a
n1
(0)), so N ,[= [a
0
, . . . , a
n1
].
So in order to see that some relation is representable (and in particular, computable), it
suces to check that it is denable in N by a formula which is numeralwise determined by A
E
.
Thats why it is useful to develop some criteria for when a formula is numeralwise determined
by A
E
.
Lemma 5.1.4. 1. For any natural number n,
A
E
v
0
(v
0
< S
n+1
(0)
jn
v
0
= S
j
(0)).
2. For any variable-free term t, there is a unique natural number n such that A
E
t = S
n
(0).
3. For any quantier-free sentence , A
E
or A
E
(A
E
decides ).
Proof. I shall tacitly use the completeness theorem all the time. Basically, I am arguing by
replacing with [=. This would not be necessary, but it makes the arguments less formalistic
and maybe more transparent. Part of the reason for this is that the choice of the particular
denition of is less canonical than the denition of [=.
For 1., argue by induction on n. Work in an arbitrary model / of A
E
. For n = 0, we have
to show that for any a [/[, a <
,
S(0)
,
i a = 0
,
. From left to right, by (L1) we have that
a <
,
0
,
or a = 0
,
. By (L2), it cannot be that a <
,
0
,
, so it must be that a = 0
,
, i.e.,
a = S
0
(0)
,
. Ill drop the superscripts in similarly simple arguments to follow. The direction
from right to left is clear again by (L1), with v
0
= a and v
1
= 0.
Now suppose 1. has been shown for n, we try to prove it for n + 1. We again use (L1):
/ [= a < S
n+1
(0) means that / [= a < S(S
n
(0)), which by (L1) is equivalent to / [=
a = S
n
(0) a < S
n
(0). Inductively, the latter is equivalent to / [=
_
j<n
v
0
= S
j
(0) in case
n > 0, and it is false in case n = 0 by (L2). Lets take the disjunction
_
j<n
v
0
= S
j
(0) to be
nothing in case n = 0, and an empty disjunction to be false in any model. Then putting the two
disjunctions together gives / [=
_
j<n+1
v
0
= S
j
(0), as desired. All these transformations work
in both directions, so we have shown the desired equivalence.
For 2, uniqueness is easily seen. For suppose A
E
t = S
m
(0) and A
E
t = S
n
(0). Since
N [= A
E
, it follows that N [= S
m
(0) = S
n
(0), i.e., m = n. For existence, argue by induction on t.
If t = 0, then /
E
t = S
0
(0). If t = S(u), where A
E
u = S
m
(0). Then A
E
S(u) = S
m+1
(0).
Now suppose u
0
and u
1
are terms such that A
E
u
0
= S
m
0
(0) and A
E
u
1
= S
m
1
(0), for some
natural numbers m
0
and m
1
. By applying (A2) m
1
times and (A1) once, it follows that
S
m
0
(0) +S
m
1
(0) = S
m
0
(0) +S(S
m
1
1
(0))
= S(S
m
0
(0) +S
m
1
1
(0))
= S(S(S
m
0
(0) +S
m
1
2
(0)))
=
= S
m
1
(S
m
0
(0) + 0)
= S
m
1
(S
m
0
(0))
= S
m
0
+m
1
(0).
So this shows the claim for t = u
0
+u
1
. For t = u
0
u
1
, we use (M2) and (M1) instead of (A2)
and (A1) to conclude that
A
E
S
m
0
(0) S
m
1
(0) = S
m
0
(0) +S
m
0
(0) + +S
m
0
(0)
. .
m
1
times
.
But by the above, A
E
S
m
0
(0) +S
m
0
(0) = S
2m
0
(0), and so on, so that by m
1
1 applications
of the above, it follows that
A
E
S
m
0
(0) S
m
1
(0) = S
m
0
m
1
(0),
which proves the claim for t = u
0
u
1
(0). The case t = u
0
Eu
1
relates to the case t = u
0
u
1
as
the case t = u
0
u
1
relates to the case t = u
0
+u
1
.
For 3., we argue by induction on the sentence .
Let be of the form t
0
= t
1
. By 2., x m
0
and m
1
such that A
E
(t
0
= S
m
0
(0) t
1
=
S
m
1
(0)).
If m
0
= m
1
, then clearly, A
E
t
0
= t
1
, since if / is an arbitrary model of A
E
, / [= t
0
=
S
m
0
(0) = S
m
1
(0) = t
1
. If m
0
,= m
1
, then wlog, let m
0
< m
1
. I claim that A
E
t
0
,= t
1
. If not,
then there would be a model / [= A
E
with / [= t
0
= t
1
. Then, arguing inside /, it follows
that S
m
0
1
(0) = S
m
1
1
(0), by (S2), which, in turn, implies that S
m
0
2
(0) = S
m
1
2
(0), again
by (S2). Applying (S2) m
0
times, it follows that 0 = S
m
1
m
0
(0), and m
1
m
0
> 0. So letting
a = S
m
1
m
0
1
(0), in /, 0 = S(a), which contradicts (S1).
Now let be of the form t
0
< t
1
. Again, x m
0
and m
1
such that A
E
t
i
= S
m
i
(0), by 2.
If m
0
< m
1
, then it follows that A
E
t
0
< t
1
. For if / [= A
E
, then by 1., / [=
x(x < S
m
1
(0)
_
j<m
1
x = S
j
(0). Substituting t
0
for x in the inner formula shows that
/ [= (t
0
< t
1

_
j<m
1
t
0
= S
j
(0)). But the disjunction on the right hand side is true in / by
assumption, as m
0
< m
1
. So / [= t
0
< t
1
, as claimed.
Now suppose m
1
m
0
. I claim that then, A
E
t
0
< t
1
. To see this, let / [= A
E
,
and assume, towards a contradiction that / [= t
0
< t
1
, i.e., / [= S
m
0
(0) < S
m
1
(0), where
m
1
m
0
. As above, arguing in /, it follows that S
m
0
1
(0) < S
m
1
1
(0), by (S2). Applying
(S2) m
1
1 more times, it follows that S
m
0
m
1
(0) < 0, which contradicts the axiom (L2) of A
E
.
So such an / cannot exist, which shows that A
E
t
0
< t
1
, as desired.
Continuing the induction is now a routine matter. Suppose the claim holds for
0
and
1
.
Then it holds for (
0

1
): If A
E

0
and A
E

1
, then A
E
(
0

1
), and we are done.
If not, then wlog A
E

i
, for i = 0 or i = 1. It follows that A
E
(
0

1
), since
i
is
false in any model of A
E
. It follows trivially that the claim holds for
0
also: If A
E

0
, then
A
E
(
0
), and if A
E

0
, then were done anyway. The case (
0

1
) reduces to the
(
0

1
).
Theorem 5.1.5. 1. Any quantier-free formula is numeralwise determined by A
E
.
2. If and are numeralwise determined by A
E
, then so are , ( ), ( ) and
( ).
3. If is numeralwise determined by A
E
, then so are the bounded quantications x(x <
y ) and x(x < y ). Ill write x < y for the rst formula and x < y for
the second.
Proof. For 1, let (x
0
, . . . , x
n1
) be a quantier-free formula. Given m
0
, . . . , m
n1
, and
letting := (S
m
0
(0), . . . , S
m
n1
(0)), it follows that is a quantier-free sentence, so by Lemma
5.1.4.3, either A
E
or A
E
.
For 2, let (v
0
, . . . , v
m1
) and (v
0
, . . . , v
n1
) be numeralwise determined by A
E
. Let
k = max(m, n), and let a
0
, . . . , a
k1
. Let
t
= (
v
0
/
S
a
0(0)
) (
v
m1
/
S
a
m1
(0)
) and
t
=
(
v
0
/
S
a
0(0)
) (
v
n1
/
S
a
n1
(0)
). Then A
E
decides
t
and A
E
decides
t
. Now the argument of
the end of the proof of the last part of Lemma 5.1.4 shows that A
E
also decides (
t

t
),
t
and (
t

t
). So since (
t

t
) is just short for (
t

t
), it follows that A
E
also decides
(
t

t
). This holds for arbitrary a
0
, . . . , a
k1
, so this shows that Boolean combinations
of numeralwise determined formulae are numeralwise determined.
For 3, rst note that (x < y ) is logically equivalent to x < y . By 2, is numer-
alwise determined by A
E
i is. So it suces to prove that if is numeralwise determined by
A
E
, then so is x < y . This is because knowing that, one can argue that given a which is
numeralwise determined by A
E
, it follows that is numeralwise determined by A
E
, hence so
is x < y , which is the same as to say that x < y is numeralwise determined by A
E
,
and hence, so is x < y .
So let be numeralwise determined by A
E
. Let y, z
0
, . . . , z
n1
be the free variables of the
formula x < y . Fix natural numbers a, b
0
, . . . , b
n1
.
Case 1. N [= (x < y )(S
a
(0), S
b
0
(0), . . . , S
b
n1
(0)).
Let c witness this formula, meaning that c < a and
N [= (
x
/
S
c
(0)
)(
y
/
S
a
(0)
)(
z
0
/
S
b
0(0)
) . . . (
z
n1
/
S
b
n1
(0)
).
Since is numeralwise determined, it follows that
A
E
(
x
/
S
c
(0)
)(
y
/
S
a
(0)
)(
z
0
/
S
b
0(0)
) . . . (
z
n1
/
S
b
n1
(0)
),
since it cannot be that A
E
proves the negation of this sentence, as it holds in N. But of course,
A
E
S
c
(0) < S
a
(0) by Lemma 5.1.4.3 as this is an atomic formula which is true in N. But then
clearly,
A
E
(x < y )(
y
/
S
a
(0)
)(
z
0
/
b
0
) . . . (
z
n1
/
b
n1
),
as this is witnessed by S
c
(0).
Case 2. Case 1 fails.
In that case, N [= (x < y )(S
a
(0), S
b
0
(0), . . . , S
b
n1
(0)). I claim that A
E
proves this
sentence. Assuming the contrary, there is a model / [= A
E
in which the negation holds:
/ [= (x < y )(S
a
(0), S
b
0
(0), . . . , S
b
n1
(0)).
Let c [/[ witness this. I.e.,
/ [= c < S
a
(0) and / [= (
y
/
S
a
(0)
)(
z
0
/
S
b
0(0)
) . . . (
z
n1
/
S
b
n1
(0)
)[(
x
/
c
)].
By Lemma 5.1.4.1, there is a natural number c
t
< a such that
/ [= c = S
c
(0).
Hence,
/ [= (
x
/
S
c
(0)
)(
y
/
S
a
(0)
)(
z
0
/
S
b
0(0)
) . . . (
z
n1
/
S
b
n1
(0)
).
Since is numeralwise determined, it follows that this sentence is actually provable in A
E
, or
else the negation would be provable, and hence the negation would have to hold in /. But if it
is provable in A
E
, then it has to hold in N as well:
N [= (
x
/
S
c
(0)
)(
y
/
S
a
(0)
)(
z
0
/
S
b
0(0)
) . . . (
z
n1
/
S
b
n1
(0)
).
But then c
t
witnesses that
N [= (x < y )(S
a
(0), S
b
0
(0), . . . , S
b
n1
(0)),
contrary to our assumption.
Now Ill turn to functional representations.
Lemma 5.1.6. If a formula (v
0
, . . . , v
n
) functionally represents a function f :
n
, then
also represents the graph of f.
Proof. Let m := m
0
, m
1
, . . . , m
n
)
n+1
be given. If m
0
, . . . , m
n
) is in the graph of f, i.e.,
m
n
= f(m
0
, . . . , m
n1
), then since functionally represents f, this means that
A
E
v
n
((
v
0
/
S
m
0(0)
) (
v
n1
/
S
m
n1
(0)
) v
n
= S
f(m
0
,...,m
n1
)
(0)).
So in an arbitrary model / [= A
E
, this sentence is true. Since
/ [= v
n
= S
f(m
0
,...,m
n1
(0)[(
v
n
/
S
f(m
0
,...,m
n1
)
(0)
M
)],
this means that
/ [= (
v
0
/
S
m
0(0)
) (
v
n1
/
S
m
n1
(0)
)(
v
n
/
S
f(m
0
,...,m
n1
)
(0)
),
so that A
E
proves that sentence, as / was arbitrary.
On the other hand, if m
0
, . . . , m
n
) is not in the graph of f, then f(m
0
, . . . , m
n1
) ,= m
n
.
It follows that in the above formula, the right hand side of the equivalence over which the
-quantier is ranging is false if S
m
n
(0) is substituted for v
n
. As before, this means that
A
E
(
v
0
/
S
m
0(0)
) (
v
n1
/
S
v
n1
m
n1
)(
v
n
/
S
f(m
0
,...,m
n1
)
(0)
).
This is the other half of what was to be shown in order to see that represents the graph of
f.
Lemma 5.1.7. If the graph of a function f is representable, f is also functionally representable.
Proof. Let (v
0
, . . . , v
n
) represent the graph of f. Let be the formula
( v
n+1
< v
n
(
v
n
/
v
n+1
)).
So (v
0
, . . . , v
n
) expresses that v
n
is least such that (v
0
, . . . , v
n
) holds. Let natural numbers
m
1
, . . . , m
n
be given. In order to see that functionally represents f, it has to be shown that
A
E
v
n
((
v
0
/
S
m
0(0)
) (
v
n1
/
S
m
n1
(0)
) v
n
= S
f(m
0
,...,m
n1
)
(0)).
So x a model / [= A
E
, and x an element a [/[. We have to see that
/ [= [(
v
0
/
S
m
0(0)
M) (
v
n1
/
S
m
n1
(0)
M)(
v
n
/
a
)] / [= a = S
f(m
0
,...,m
n1
)
(0).
For the direction from right to left: Since represents the graph of f,
/ [= [(
v
0
/
S
m
0(0)
M) (
v
n1
/
S
m
n1
(0)
M)(
v
n
/
a
)],
and there can be no c / such that / [= c < a and
/ [= [(
v
0
/
S
m
0(0)
M) (
v
n1
/
S
m
n1
(0)
M)(
v
n
/
c
)],
for since / [= a = S
f(m
0
,...,m
n1
)
(0), it would follow that / [= c = S
k
(0), for some k <
f(m
0
, . . . , m
n1
) (by Lemma 5.1.4.1), and hence that m
0
, . . . , m
n1
, k) f, as represents the
graph of f as a relation. But this would mean that f(m
0
, . . . , m
n1
) = k < f(m
0
, . . . , m
n1
), a
contradiction.
For the converse, the argument is similar, but let me argue a little less formally. The point
is that there is a natural number k, namely k = f(m
1
, . . . , m
n
), such that
/ [= (S
m
0
(0)
,
, . . . , S
m
n1
(0)
,
, S
k
(0)
,
),
because the graph of f is represented by . Below a numeral, there are only numerals, so since a is
least (in /) such that / [= [S
m
0
(0)
,
, . . . , S
m
n1
(0)
,
, a], it follows that a S
k
(0) in /, and
hence, a is equal to a numeral S
l
(0)
,
, l k. So we have / [= (S
m
0
(0), . . . , S
m
n1
(0), S
l
(0)),
which means that l = f(m
0
, . . . , m
0
) = k, and in particular,
/ [= a = S
f(m
0
,...,m
n1
)
(0).
These facts allow us to formulate some closure properties of the class of representable func-
tions.
Lemma 5.1.8. Suppose f
0
, . . . , f
m1
are representable functions with dom(f
0
) = . . . = dom(f
m1
) =
k
, and that g :
m
is representable. Then the function h = g (f
0
, . . . , f
m1
) :
k
,
dened by
h(a
0
, . . . , a
k1
) := g(f
0
(a
0
, . . . , a
k1
), . . . , f
m1
(a
0
, . . . , a
k1
))
is representable.
Proof. We may by Lemma 5.1.7 choose functional representations
0
, . . . ,
m1
, of f
0
, . . . , f
m1
, g.
Note that the set of free variables of
i
is contained in v
0
, v
1
, . . . , v
k
. Choose variables
x
0
, . . . , x
m1
which do not occur in any of these formulas. Let be the following formula:
x
0
. . . x
m1
_
_
_
_
_
_
0<l<m
l
(
v
k
/
x
l
)
. .
x
l
=f
l
(v
0
,...,v
k1
)
_
_
_ (
v
0
/
x
0
) (
v
m1
/
x
m1
)
. .
v
m
=g(x
0
,...,x
k1
)
_
_
_.
We check that represents the graph of h:
Let a
0
, . . . , a
k1
, b)
k+1
. There are two cases.
Case 1: b = h(a
0
, . . . , a
k1
).
Then let b
0
= f
0
(a
0
, . . . , a
k1
), . . . , b
m
= f
m
(a
0
, . . . , a
k1
). Let / be an arbitrary model of
A
E
. Since
i
represents f
i
, for 0 < i < k, it follows that for each such i, we have:
/ [=
i
(
v
0
/
S
a
0(0)
), . . . , (
v
k1
/
S
a
k1
(0)
)(
v
k
/
S
b
i (0)
),
and trivially also that
/ [=
i
(
v
0
/
S
a
0(0)
), . . . , (
v
k1
/
S
a
k1
(0)
)(
v
k
/
x
i
)(
x
i
/
S
b
i (0)
).
Moreover, since represents g, it follows that
/ [= (
v
0
/
S
b
0(0)
) (
v
m1
/
S
b
m1
(0)
)(
v
m
/
S
b
(0)
),
so that
/ [=
__
i<m
i
(
v
0
/
S
a
0(0)
) . . . (
v
k1
/
S
a
k1
(0)
)(
v
k
/
x
i
)(
x
i
/
S
b
i (0)
)
_
(
v
0
/
S
b
0(0)
) (
v
m1
/
S
b
m1
(0)
)(
v
m
/
S
b
(0)
)
_
.
So the terms S
b
0
(0), . . . , S
b
m1
(0) witness that
/ [= (
v
0
/
a
0
) (
v
k1
/
a
k1
)(
v
m
/
S
b
(0)
).
Since / was arbitrary, this means that
A
E
(
v
0
/
a
0
) (
v
k1
/
a
k1
)(
v
k
/
S
b
(0)
),
as wished.
Case 2: b ,= h(a
0
, . . . , a
k1
).
Let / be an arbitrary model of A
E
. We have to show that
/ [= (
v
0
/
S
a
0(0)
) (
v
k1
/
S
a
k1
(0)
)(
v
0
/
S
b
(0)
).
Assuming the contrary means that there are b
0
, . . . , b
m1
[/[ such that for every i < m,
/ [=
i
(
v
0
/
S
a
0(0)
) (
v
k1
/
S
a
k1
(0)
)[(
v
k
/
b
i
)]
and
/ [= [(
v
0
/
b
0
) (
v
m1
/
b
m1
)(
v
m
/
S
b
(0)
M)].
Since
i
functionally represents f
i
, it follows from the former that
/ [= b
i
= S
f
i
(a
1
,...,a
k
)
(0).
And hence, the latter means that
/ [= (
v
0
/
S
f
0
(a
0
,...,a
k1
)
(0)
) (
v
m1
/
S
f
m1
(a
0
,...,a
k1
)
(0)
)(
v
m
/
S
b
(0)
),
which, since represents g, means that
b = g(f
0
(a
0
, . . . , a
k1
), . . . , f
m1
(a
0
, . . . , a
k1
)) = h(a
0
, . . . , a
k1
),
a contradiction.
Denition 5.1.9. Given a relation R
n
, let
R
be its characteristic function,
R
:
n
2,
dened by:
R
(m
0
, . . . , m
n1
) = 1 i m
0
, . . . , m
n1
) R.
Theorem 5.1.10. A relation R is representable i its characteristic function is.
Proof. From left to right, let (v
0
, . . . , v
n1
) represent R. Let
t
(v
0
, . . . , v
n1
, v
n
) be the formula
(( v
n
= S(0)) ( v
n
= 0)).
Clearly,
t
is numeralwise determined by A
E
, being a Boolean combination of such formulae.
And
t
obviously denes
R
over N.
From right to left, let (v
0
, . . . , v
n
) represent
R
. Let
t
(v
0
, . . . , v
n1
) be the formula
(
v
n
/
S(0)
).
It follows immediately that
t
is numeralwise determined by A
E
, since is. Again,
t
denes R
over N, so were done.
Lemma 5.1.11. Suppose R
n
is representable and f
0
, f
1
, . . . , f
n1
are representable func-
tions with domain
k
. Then so is the relation
S = m
0
, . . . , m
k
) [ f
0
( m), . . . , f
n1
( m)) R.
Proof. Since
S
=
R
(f
0
, . . . , f
n1
),
and
R
, f
0
, . . . , f
n1
are representable,
S
is also representable.
Theorem 5.1.12. Let g :
n+1
be representable, and let g have the property that for
every tuple m
0
, . . . , m
n1
), there is an m such that g(m
0
, . . . , m
n1
, m) = 0. Then the function
f :
n
dened by
f(m
0
, . . . , m
n1
) = minm [ g(m
0
, . . . , m
n1
, m) = 0
is representable.
Proof. Let (v
0
, . . . , v
n+1
) be a functional representation of g. Then the following formula denes
the graph of f over N:
_
(
v
n+1
/
0
) v
n+2
< v
n
(
v
n
/
v
n+2
)(
v
n+1
/
0
)
_
.
I leave it to the reader to convince him- or herself that the formula (
v
n+1
/
0
) is numeralwise
determined. It follows then from Lemma 5.1.5 that the entire formula is numeralwise determined.
Hence, by Lemma 5.1.3, it represents the graph of f.
Theorem 5.1.13. The following relations are representable:
1. P := p [ p is a prime number,
2. Q := p, q) [ p < q and [p, q] P = p, q,
3. The monotone enumeration p of P.
Proof. 1.) P is represented by the following formula:
(v
0
) := (0 < v
0
v
1
< v
0
v
2
< v
0
v
1
v
2
= v
0
).
2.) Here is a formula which represents Q:
((v
0
) (v
1
) v
0
< v
1
v
2
< v
1
(v
0
< v
2
(
v
0
/
v
2
))),
where in (
v
0
/
v
2
), an implicit replacement of the bound variable v
2
occurs, by our convention.
In the future, I shall sometimes abuse notation and write instead:
v
0
P v
1
P v
0
< v
1
v
2
< v
1
(v
0
< v
2
v
2
/ P).
3.) Instead of writing down a formula representing the set in question explicitly, I will use
abbreviations as in the proof of 2.). The reader will no doubt be able to translate this into a
genuine formula. The formula should express the following:
v
1
P and there is an a v
v
2
0
1
such that
1. 2 does not divide a,
2. for all u, v < a, if u, v) Q, then for all i < v
0
:
u
i
divides a v
S(i)
divides a.
3. v
1
is the least prime number r such that r
v
0
divides a.
It is obvious that the relation u, v)
2
[ u divides v is representable (by a bounded formula).
So the above formula results from substituting the representable functions x, y) x
y
2
, x, y)
x
y
and x, y) x
S(y)
into a representable relation. The result is a representable relation, by
Lemma 5.1.11.
To see that the relation dened over N by the above formula actually is the graph of the
monotone enumeration of P, two directions have to be veried. First, assume that p(m) = n.
Then let a = p(0)
0
p(1)
1
p(m)
m
. Clearly, 1.-3. above hold in N with v
0
= m and v
1
= p(m),
and a p(m)
m
2
. For the converse, assume that 1.-3. hold in N, with v
0
= m and v
1
= q, where q
is a prime number, and with a q
m
2
. By induction on i m, it follows that p(i)
i
divides a but
p(i)
i+1
doesnt, using 1. for the start of the induction and 2. for the induction step. It follows
that q = p(m), since p(m)
m
divides a, and it is the least prime number with this property, as is
q, so they must be equal.
5.2 A dierent G odelization: Ackermanns coding of V
Denition 5.2.1 (Ackermann). For a natural number n, let u

n
be the uniquely determined
nite set of natural numbers such that
n =
mu
n
2
m
.
Dene a relation

2
by:
m
n m u
n
.
So m
n i the m-th digit in the binary representation of n is 1. Note that

m
n = m < n.
Theorem 5.2.2.

is representable.
Proof. v
0
v
1
i, in N, there is b p(v
0
)
v
0
v
1
such that:
1. 2
v
1
divides b but 2
S(v
1
)
does not.
5.2. A DIFFERENT G
ODELIZATION: ACKERMANNS CODING OF V
61
2. For all i < v
0
:
p(i)
i
divides b p(i + 1)
[i:2]
divides b.
3. The largest i < S(v
1
) such that p(v
0
)
i
divides b is odd.
The argument is similar to the one establishing that the monotone enumeration of the set of prime
numbers is representable. The methods of the previous section show that the above formula is
numeralwise determined, which means that

is representable.
The following is proved as an exercise:
Theorem 5.2.3 (Ackermann). The structure ,

) is well-founded and extensional. So it is
isomorphic via the Mostowski collapse to a transitive structure. In fact, letting be the collapse,
it follows that
: ,

)

V
, V
).
Denition 5.2.4. A bounded quantier in the language of set theory is a quantier of the form
x y or x y, where x and y are variables. Occurrences of these bounded quantiers can be
eliminated and so these quantiers can be viewed as being abbreviations:
x y means x(x y ),
and
x y means x(x y ).
A
0
-formula is a formula in which all quantiers are bounded. Such formulae are sometimes
also referred to as
0
-formulae.
Theorem 5.2.5. If R V
n
is denable over V
, ) by a
0
-formula (using parameters), then
R :=
1
(a
1
), . . . ,
1
(a
n
)) [ a
1
, . . . , a
n
) R
is representable.
Proof. By induction on formulae in the language of set theory (with bounded quantiers), I
show:
() If is
0
, then there is a numeralwise determined formula in the language of A
E
with
the same free variables as , such that for all a
1
, . . . , a
n
V
,
V
, ) [= [a
1
, . . . , a
n
] N [= [
1
(a
1
), . . . ,
1
(a
n
)].
It will be clear in each case that the displayed claim holds, but Ill check that the formulae
produced are numeralwise determined. To start o, lets deal with the atomic cases. If = x =y,
then we can let = . If = x
y, then let
= x
y,
where x
y stands for the formula representing

according to the previous lemma, with x and
y substituted for v
0
and v
1
, respectively.
Boolean combinations are trivial. Thus,

(
0

1
) = (
0

1
), et cetera. This produces nu-
meralwise determined formulae, since Boolean combinations of numeralwise determined formulae
are numeralwise determined.
Finally, lets turn to bounded quantication. Suppose = x y , where is already
dened. By renaming the bounded variables of if necessary, wma x and y dont occur as
bound variables in . Then set
= x < y(x
y ).
The crucial point here is that m
n = m < n. Since bounded quantication over numeralwise

determined formulae produces numeralwise determined formulae, this produces a numeralwise
determined formula. The case of bounded universal quantication is analogous, so we are done.
Denition 5.2.6. A formula in the language of set theory is
1
if it is of the form x, where
is
0
. Given a model of set theory /, a relation R [/[
n
is
1
(/) if it is denable over
/ by a
1
-formula. It is
1
(/) if both R and R := [/[
n
R are
1
(/).
Ill also say that a formula (x
0
, . . . , x
n1
) is
T
1
, where T is some theory, if there is a
1
-formula (x
0
, . . . , x
n1
) such that T . It is
T
1
if and both are
T
1
.
The importance of
1
(V
)-relations is evidenced by the following theorem:

Theorem 5.2.7. If R V
n
is
1
(V
, )), then

R is representable.
Before beginning the proof, let me remark that we dont need to worry about parameters
that might occur in a
1
(V
) denition of a relation over V
, because the inverse image of such

a parameter is a natural number, represented by a numeral, so that by substituting the numeral
for the preimage of the parameter, it disappears.
Proof. Let R be dened over V
by = x and R by = x, where and are

0
-
formulae. By claim () of the proof of the previous lemma, it follows that

R is dened by x
and R is dened by x
over N. Note that

and

are numeralwise determined formulae, by
the previous lemma. I claim that the following formula
represents

R:
x(
x < x(
(
x
/
x
)
(
x
/
x
))).
To see this, we have to check two things:
First, suppose a

R. Then N [= x
[a]. Let n be the least witness. Then

N [=
[S
a
(0), S
n
(0)] x < n(
(
x
/
x
)(S
a
(0))
(
x
/
x
)(S
a
(0))).
But this is a numeral instance of a numeralwise determined formula, so A
E
proves it. So A
E

(S
a
(0)).
Secondly, suppose a /

R. Then N [= x
(S
a
(0)). Pick the least witness n. So
N [=
(S
a
(0), S
n
(0)) x < S
n
(0)(
(
x
/
x
)[a]
(
x
/
x
)[a]))
. .
.
So A
E
. I claim that A
E

(S
a
(0)). If not, then let / [= A
E
be such that / [=
(S
a
(0)). Pick a witness b [/[, so that
/ [=

(S
a
(0))[b] x < b(
(
x
/
x
)(S
a
(0))
(
x
/
x
)(S
a
(0)))).
Moreover, / [= . It is impossible that / [= S
n
(0) < b, since / [= , and hence, / [=
(S
a
(0), S
n
(0)). So / [= b S
n
(0) (by (L3)). But weve known for a long time that there are
5.2. A DIFFERENT G
ODELIZATION: ACKERMANNS CODING OF V
63
only numerals below numerals (Lemma 5.1.4.1), so that this implies that / [= b = S
l
(0), for
some l n. So then, / [=

(S
a
(0), S
l
(0)). But this is a numeral instance of a numeralwise
determined formula, and hence decided by A
E
. So A
E

(S
a
(0), S
l
(0)), and in particular,
N [=
(S
a
(0), S
l
(0)). So N [= x
(S
a
(0)), i.e., a

R, a contradiction.
Theorem 5.2.8. Let T = ZF
F
+ (Powerset).
1. If (x
0
, . . . , x
n1
) is
T
1
, then so is x
i
, for i < n.
2. If
0
and
1
are
T
1
, then so are (
0

1
) and (
0

1
).
3. If (x
0
, . . . , x
n1
) is
T
1
, and y, z are variables which dont occur as bound variables in ,
then the formulae y z and y z are also
T
1
.
4. If / [= T and F : [/[ [/[ is a function whose graph is
1
(/), then it is
1
(/).
Proof. 1.) Suppose is of the form z(x
0
, . . . , x
n1
). Then, provably in T,
x
i
z(x
0
, . . . , x
n1
) ux
i
uz u(x
0
, . . . , x
n1
),
where u is some new variable that doesnt occur in . This uses the pairing axiom.
2.) Wlog, we may assume that
0
and
1
are
1
-formulae, and also that they dont have a
bound variable in common. Let
0
= x
t
0
and
1
= y
t
1
. Then, provably in T,
(
0

1
) ux uy u(
0

1
),
and analogously for .
3.) Wlog, let be a
1
-formula, say = x
t
(x
0
, . . . , x
n1
), where
t
is
0
. So y z is
equivalent to the formula
y(y z ),
which is
T
1
, by 1. and 2. Now consider the formula y z. Ill need the following basic
tool: Since Im assuming the Power Set axiom, I can dene, by recursion on the ordinals, a
function V
, by letting V
0
= , V
s()
= T(V
) and V
<
V
, for limit ordinals .

I have introduced the rst + 1 steps of this recursion already. Just as then, it follows from
the Foundation axiom that V =

On
V
. Now, working in some model / of T, and xing

a
0
, . . . , a
n1
[/[, assume that
/ [= y zx
t
[a
0
, . . . , a
n1
].
Working in /, let F be the function dened by: F(y) =the least such that if there is a b with
t
[a
0
, . . . , a
n1
(
x
/
b
)], then there is such a b V
. By replacement, the range of F is a set of

ordinals which has a supremum, say . So
/ [= wy zx w
t
[a
0
, . . . , a
n1
],
as witnessed by w = V
,
. Vice versa, if this latter formula holds in / then / models that

y zx
t
[a
0
, . . . , a
n1
] is true. Since / was arbitrary, this shows that the original formula is,
provably in T, equivalent to the
1
-formula wy zx w
t
.
4.) Let dene the graph of F over /. Then a, b
0
, . . . , b
n1
) F i a ,= F(b
0
, . . . , b
n1
)
i
/ [= z(z ,= a (z, b
0
, . . . , b
n1
)).
Theorem 5.2.9. Given
1
-formulae
R
(x
0
, x
1
, . . . , x
n1
, w),
F
i
(y
0
, . . . , y
m1
, y
m
) (for i < n),
there is a
1
-formula (x
0
, . . . , x
n1
, w) such that in the theory ZF
, the following is provable:
i<n
F
i
denes a function
_
_
_z
0
. . . z
m1
((z
0
, . . . , z
m1
, w)
u
0
. . . u
n1
(
i<n
F
i
(
y
0
/
z
0
) (
y
m1
/
z
m1
)(
y
m
/
u
i
)
. .
u
i
=F
i
(z
0
,...,z
m1
)
R
(
x
0
/
u
0
) . . . (
x
n1
/
u
n1
)))
_
_
_.
Proof. We can take the formula
u
0
. . . u
n1
(
i<n
F
i
(
y
0
/
z
0
) (
y
m1
/
z
m1
)(
y
m
/
u
i
)
R
(u
0
, . . . , u
n1
))
as the denition of . This is equivalent to a
1
-formula.
In eect, the previous theorem allows us to substitute
1
-functions in
1
-formulae, and thats
the way in which Ill use it. Here is a collection of T-provably
1
functions and relations. Ill
use class terms as introduced in the chapter on set theory. The previous theorem is very useful
in proving the following lemma.
Lemma 5.2.10. The following formulae are
T
1
:
1. v
1
= v
0
2. v
n
= v
0
, . . . , v
n1
).
3. v
0
is a function.
4. v
1
= ran(v
0
).
5. v
1
= dom(v
0
).
6. v
2
= v
0
v
1
.
7. v
2
= v
0
(v
1
).
Proof. Exercise.
Trusting the reader by now knows what I mean, Ill state the following theorem a little less
formally than Theorem 5.2.9.
Theorem 5.2.11 (T). Suppose G : A V V is dened by a
1
-formula (in parameters),
R AA is strongly well-founded, and the function x A Rx is
1
. Then the uniquely
determined function F : A V (given be the Recursion Theorem) such that
x A F(x) = G(x, FRx)
is a
1
-function.
5.3. G
ODELS AND TARSKIS RESULTS 65

Proof. By the recursion theorem, the following formula denes F:
f((f) f(v
0
) = v
1
),
where (f) expresses the following:
f is a function,
dom(f) is closed under R,
for all x dom(f), f(x) = G(x, fRx).
The second point says: y dom(f)z Ryz dom(f). The third one is obviously
expressible by a
1
-formula.
Theorem 5.2.12. The following functions and relations/sets are
1
(V
):
1. , and the functions +, dened on .
2. The set of symbols of the language of A
E
, in some convenient coding.
3. The set of variables of that language, also the set of constants, the set of function symbols
and the set of relation symbols (the latter three are nite).
4. The free semi-group generated by these symbols, in the following sense: There is a class
Z V
and a function

: Z Z Z such that Z,
) is a free semi-group generated

by , and such that Z and

are
1
(V
).
5. The set of terms and the set of formulae.
6. The function .
5.3 G odels and Tarskis results
Lemma 5.3.1 (Fixed-point lemma). Let (x) be a formula with one free variable x. Then there
is a sentence such that
A
E

_
(
x
/
S
1
()
(0)
)
_
.
Proof. Consider the relation
R = (y), (r), (
y
/
S
r
(0)
)) [ has the one free variable y, and r .
It is easy to see that this relation is
1
(V
). It is actually a function, with the value in the

last component. The expanded function f that maps u, v) to 0 if there is no w such that
u, v, w) R is also
1
(V
) and total, and hence the pullback of that function is representable.

So let (v
0
, v
1
, v
2
) be a functional representation of the pullback

f of f, i.e., the function dened
by

f(
1
(u),
1
(v)) =
1
(f(u, v)). So we have, for all formulae (y) with one free variable y
and all natural numbers r:
(1) A
E
v
2
_
(
v
0
/
S
1
()
(0)
)(
v
1
/
S
r
(0)
) v
2
= S
1
((
y
/
S
r
(0)
))
(0)
_
.
Now set:
(v
1
) := v
2
((
v
0
/
v
1
) (
x
/
v
2
)) .
Let p =
1
(), and set:
:= (
v
1
/
S
p
(0)
).
So by (1), it follows that
(2) / [= v
2
_
(
v
0
/
S
1
()
(0)
)(
v
1
/
S
p
(0)
) v
2
= S
1
()
(0)
_
,
which of course is the same as to say that
(2) / [= v
2
_
(
v
0
/
S
p
(0)
)(
v
1
/
S
p
(0)
) v
2
= S
1
()
(0)
_
,
since p =
1
().
I claim that is as desired. To see this, let / be an arbitrary model of A
E
. It must be
shown that
(/ [= )
_
/ [= (
x
/
S
1
()
(0)
)
_
.
For the direction from left to right, assume that / [= . This means that / [= (
v
1
/
S
p
(0)
), by
denition of . Unraveling further, by denition of , this means that
(3) / [= v
2
_
(
v
0
/
S
p
(0)
)(
v
1
/
S
p
(0)
) (
x
/
v
2
)
_
.
Since / [= A
E
, the sentence that according to (2
t
) is a consequence of A
E
holds in /, so that
/ [= v
2
_
(
v
0
/
S
p
(0)
)(
v
1
/
S
p
(0)
) v
2
= S
1
()
(0)
_
.
In particular,
/ [= (
v
0
/
S
1
()
(0)
)(
v
1
/
S
p
(0)
)(
v
2
/
S
1
()
(0)
),
which by (3) implies that / [= (
x
/
S
1
()
(0)
), as desired.
For the converse, assume that / [= (
x
/
S
1
()
(0)
). We have to show that / [= , which
means, unraveling as above, that (3) holds. In order to do this, we have to show that whenever
a / is such that
/ [= (
v
0
/
S
p
(0)
)(
v
1
/
S
p
(0)
)[(
v
2
/
a
)],
then
/ [= [(
x
/
a
)].
But by (2), we know that if / [= (
v
0
/
S
p
(0)
)(
v
1
/
S
p
(0)
)[(
v
2
/
a
)], it follows that / [= a =
S
1
()
(0), and since our assumption is that / [= (
x
/
S
1
()
(0)
), this means that / [= [(
x
/
a
)],
as desired.
Denition 5.3.2. If / is a model of a language L, then let Th(/) be the set of sentences
which are true in /.
Theorem 5.3.3 (Tarskis undenability of truth). The set

Th(N) is not denable in N.
Proof. Assume the contrary. Note that if Th(N) was bold-face denable, then also lightface,
since every parameter can be replaced by its numeral. So there was a formula (v
0
) such that
for every sentence , we would have:
(N [= )
_
N [= (
v
0
/
S
1
()
(0)
)
_
.
Now, by Lemma 5.3.1, applied to the formula (v
0
), let be a sentence such that A
E
proves
that (
v
0
/
S
1
()
(0)
). Since N [= A
E
, this sentence holds in N, and we get:
N [= N [= (S
1
()
(0)) N ,[= ,
a contradiction.
5.3. G
ODELS AND TARSKIS RESULTS 67

5.3.1 Recursive sets
Denition 5.3.4. A Turing machine is a tuple (of nite sets) of the formT = I, A, S, s
0
, , t
+
, t
),
with A:
1. I is a nite set, called the input alphabet,
2. I A, A is nite; A is called the alphabet, and sometimes A I is called the machine
alphabet,
3. S is a nite set. The members of S are called the states of T.
4. s
0
S is the initial state of T, t
+
S is the accepting state and t
is the declining state.

5. : AS AL, R S is a function, and L, R are distinct xed sets; for deniteness,
let L = 0, R = 1.
A snapshot, or constellation of T is a tuple of the form c = t, p, s) such that t : n A, for
some n , p < n and s S. n is the length of c.
If c = t, p, s) is a snapshot of T of length n, then the T-next snapshot after c is the snapshot
c
t
= t
t
, p
t
, s
t
) of length n
t
dened as follows: Let (t(p), s)) = s, d, x), where d L, R and
x A.
1. s
t
= s,
2. if d = R, then p
t
= p + 1,
3. if d = L and p > 0, then p
t
= p 1,
4. if d = R and p = n 1, then n
t
= n + 1, otherwise n
t
= n,
5. s
t
(l) =
_
_
_
x if l = p,
s(l) if p ,= l < n,
if l = n < n
t
.
An input is a function i : n I. The run of T on input i is the following function r with
domain N + 1 whose values are snapshots of T:
1. r(0) = i, 0, s
0
).
2. If r(n) = t, p, s) and s s
+
, s
, then N = n + 1.
3. If n + 1 < N, then r(n + 1) is the T-next snapshot after r(n).
If r is nite, then let dom(r) = N > 0. The last state of r is the unique member of s
+
, s
which is the last component of r(N 1).

T accepts an input i if the run of T on input i is nite with last state s
+
. It declines an input
i if the run is nite with last state s
. If the run of T on input i is nite, T is said to terminate

on input i.
A set R is recursive i there is a Turing machine T with input alphabet 1 such that
for all n < , if n R, then T accepts the input 1
n
, and if n / R, then T declines the input
1
n
, where 1
n
stands for the unique function i : n 1.
Lemma 5.3.5. The relation R consisting of the tuples T, i, n, c) V
such that T is a Turing

machine, i is an input for T, n is less than the length of the run r of T on input i and c = r(n)
is
1
(V
).
Theorem 5.3.6. If R is recursive, then R is
1
(V
).
Denition 5.3.7. For a set of formulae, let
|
be the set of formulae which are provable
from . is complete if for every sentence , or is in .
Theorem 5.3.8 (Godels rst Incompleteness Theorem). If Th(N) is recursive (meaning
that

is recursive), then
|
is incomplete.
Proof. So let

be recursive. Then by Theorem 5.3.6, is
1
. Suppose
|
was complete. It
would follow that
|
is
1
:
Let

be the monotone enumeration of , according to the ordering <
1
() <
1
(). Since and are
1
(V
), it follows easily that

is also
1
. Analogously, it is easy
to see that there is a
1
(V
) enumeration of the terms of A

E
.
To see that
|
is
1
, let x V
be arbitrary. Then x
|
i x is an A
E
-sentence (which
is
1
(V
)) and the search tree S
x],
t
is nite (this follows from Lemmas 4.2.1 and 4.2.2). That
S
x],
t
is nite can be expressed over V
by a
1
-formula: There is a set b such that b = S
x],
t
(the latter is easily seen to be
1
).
To see that
|
is
1
, again let x V
be arbitrary. Then x
|
i x is not an A
E
-
sentence, or if x
|
. The latter is true because by assumption,
|
is complete. But Ive
already argued that
|
is
1
(V
).
Moreover,
|
= Th(N):
For the inclusion from left to right, let
|
. Since Th(N), N [= , so Th(N).
For the converse, let Th(N). If /
|
, then by completeness of
|
,
|
, so that
by the inclusion from left to right, Th(N), which is a contradiction.
So what I showed at the beginning of the proof is that actually, Th(N) is
1
(V
). Since
: ,

)

V
, ), it follows that

Th(N) is denable in ,

). Let (x) be a denition
(in the pure language, without bounded quantiers). Replacing every occurrence of

by the
denition of

results in a formula
of the language of A
E
that denes

Th(N) over N, which
contradicts Theorem 5.3.3.
Bibliography
[End72] Herbert B. Enderton. A Mathematical Introduction to Logic. Academic Press, New
York, 1972.
[Jec03] Thomas Jech. Set Theory. The Third Millenium Edition. Springer Monographs in
Mathematics. Springer, Berlin, 2003.
[Poh09] Wolfram Pohlers. Proof Theory. The rst Step into Impredicativity. Springer, 2009.
69

Recursion in Set Theory

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Recursion in Set Theory

Uploaded by

Copyright:

Available Formats

Logic I

z = z. e is uniquely determined, and

may stand for the

y. Dropping the brackets is justied by 1. Ill also identify s and s

if there are words w

with the following properties:

with the following properties:

with the following properties:

with the following properties:

with the following proper-

) = 2. As usual, I will use the inx notation for the predicates

(x, y) =(x, z)).

A be the following class term:

x, y is the same as x y, so were done.

, which is an ad-hoc name. So it consists of the following axioms:

, together with the

T(u) [ u D. So F is a relation with X = dom(F) and ran(F) B.

be nite. Dene F : u by letting F(x) be the least n with x V

. As before, let F : x be dened by

). Every innite, locally nite tree on a well-ordered set X has

z. Now we can dene by recursion a function F : On x as

, given a set X, one can construct a canonical free halfgroup

b, it follows that FV(t

FV(). If / is an L-model and a Assign

1) are, by the following clauses. For ease of notation, assuming

(s)) ,= , then let be the member of X

) If R((s)) is of the form (

) If R((s)) is of the form v

, this renaming does not change

) If R((s)) is of the form (

1 / S. Let be the rst

(s), if such a formula exists, and

. If there is no such formula, let x

) If R((s)) is of the form v

1 / S. Let t be the rst relevant

(s), if such a term exists. In

(s), if such a formula exists. In that

0 is the sole successor of s in S. By minimality, we know that there is a nite subset

(s)) ,= , then letting be the member of X thats mentioned rst in the

1 both are successors of s in S, and (s

, and for a nite subset

0 is the only successor of s in S. Moreover, letting n be minimal such that v

0 S is the unique successor of s in S. Let Y :=

0 is the unique successor of s in S. Let t be the rst relevant term mentioned

(s), if such a term exists. In that case, let x =

(s), if such a formula exists. In that case, let x

such that for every sequence

, the redex of (f(n +b)) would have to be undened, a contradiction.

be , together with the identity axioms.

, and variables, the relevant terms. Fix also an

)[a], which contradicts

C, there is one H C with X, Y H (here it is crucial that C

C. So by Zorns lemma, there is a maximal

], which means that

Denition 5.2.1 (Ackermann). For a natural number n, let u

n i the m-th digit in the binary representation of n is 1. Note that

ODELIZATION: ACKERMANNS CODING OF V

y stands for the formula representing

n = m < n. Since bounded quantication over numeralwise

)-relations is evidenced by the following theorem:

) denition of a relation over V

, because the inverse image of such

by = x and R by = x, where and are