Buehler - Classical Metalogic PDF

Classical Metalogic
An Introduction to Classical Model Theory & Computability
Notes by R.J. Buehler
January 5, 2015
2
Preface
What follows are my personal notes created in preparation for the UC-Berkeley Group in Logic preliminary exam. I
am not a computability theorist, nor a model theorist; I am a graduate student with some knowledge who isalasquite
fallible. Accordingly, this text is made available as a convenient reference, set of notes, and summary, but without even
the slight hint of a guarantee that everything contained within is factual and correct (indeed, some areas are entirely
unchanged from the moment I copied them off the blackboard). This said, if you find a serious error, I would greatly
appreciate it if you would let me know so that it can be corrected. The material for these notes derives from a wide
variety of sources:
Lectures by Wes Holliday
Lectures by Antonio Montalban
Lectures by John Steel
Kevin Kellys computability theory notes

David Markers Model Theory: An Introduction
Wilfrid Hodges A Shorter Model Theory
Robert Soares Recursively Enumerable Sets and Degrees"

Richard Kayes Models of Peano Arithmetic
Chang and Keislers Model Theory"
While I certainly hope my notes are beneficial, if youre attempting to learn the contained material for the first time, I
would highly suggest picking up (at least) a copy of Marker and Soares texts in addition.
To those Group in Logic students who may be using these notes to help themselves prepare for their preliminary
exam, chapters 1-5, 7, and 9-18 contain relevant material, as well as chapter 8, section 3.
i
ii PREFACE
Contents
Preface i
1 First-Order Logic 1
1.1 Signatures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2.1 Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2.2 Atomic Formulas and Sentences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.3 Formal Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.4 Standard Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Normal Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
I Classical Model Theory 5

2 The Basics of Models 7
2.1 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Models and Formal Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Model-Theoretic Logical Consequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Theories and Diagrams of Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.2 Canonical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 Relations Between Models 13

3.1 From Homomorphism to Isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Submodels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Preserving Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1 and Preservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.2 Elementary Preservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4 Chains of Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4.1 Elementary Chains of Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5 Reduction and Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4 Compactness and Strong Completeness 21

4.1 Hintikka Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.2 Compactness for First-Order Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Strong Soundness and Completeness for First-Order Logic . . . . . . . . . . . . . . . . . . . . . . . 23
5 Definability 25
5.1 Definability of Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.2 Definable Classes of Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.3 The Skolem Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.3.1 Skolemization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.3.2 The Downward Lwenheim-Skolem Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 27
iii
iv CONTENTS
5.3.3 The Upward Lwenheim-Skolem Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
6 Amalgamation 31
6.1 Elementary Amalgamation Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.2 Strong Elementary Amalgamation Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.3 The Strong Amalgam Property and Algebraicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.4 Interpolation and Definability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
7 Types, Saturation, and Categoricity 37

7.1 Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.1.1 Tuples with the Same Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.1.2 Principal and Supported Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.1.3 Towards Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.2 Omitting Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
7.3 Atomic and Prime Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.4 Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
7.4.1 Homogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.4.2 Universality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
7.5 Categoricity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
8 Other Model-Theoretic Constructions 49

8.1 Frasss Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
8.1.1 Constructing the Frass Limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
8.1.2 Countable Categoricity and Frass Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
8.2 Ultraproducts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
8.2.1 Ultrafilters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
8.3 Ehrenfeucht-Mostowski Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
8.3.1 Indiscernible Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
8.3.2 E-M Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
II Computability Theory 61
9 Of Algorithms 63
9.1 Turing Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
9.1.1 The Church-Turing Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
9.2 Functions and Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
10 The Primitive and -Recursive Functions 65

10.1 The Primitive Recursive Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
10.1.1 Primitive Recursive Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
10.2 Primitive Recursive Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
10.2.1 Piecewise Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
10.2.2 Bounded Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
10.2.3 Coding Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
10.2.4 Simultaneous and Course of Values Recursion . . . . . . . . . . . . . . . . . . . . . . . . . 68
10.3 The Computable Functions? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
10.4 Unbounded Search & -Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
10.5 The Computable Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
10.5.1 Computable Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
CONTENTS v
11 The Fundamental Theorems of Computability Theory 73

11.1 The Padding Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
11.2 The Normal Form Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
11.3 The Enumeration Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
11.4 The sm
n Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
11.5 The Halting Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
11.6 Kleenes Fixed Point Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
12 Non-Computable Relations 81
12.1 A Three-Fold Distinction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
12.1.1 Closure Properties and Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
12.2 Computably Enumerable Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
12.3 Reduction and Computable Separability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
12.4 Many-One Reducibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
12.4.1 The m-Degrees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
12.4.2 Characterizing the 01 /01 -Complete Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
13 Relative Computation and Turing Reducibility 95

13.1 Higher Complexity Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
13.1.1 Closure Properties & Basic Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
13.1.2 Canonical Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
13.1.3 Trees and Tree Arguments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
13.2 Relative Computability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
13.2.1 The Relativized Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
13.2.2 The Relative Arithmetical Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
13.3 Turing Reducibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
13.3.1 The Turing Jump Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
14 Posts Problem & Oracle Constructions 107

14.0.2 Simple Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
14.1 Incomparable Turing Degrees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
14.2 Friedburg Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
14.2.1 Minimal Degrees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
14.2.2 Constructing c.e. sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
III Computability and First-Order Logic 115

15 Languages, Theories, and Computability 117
15.1 Languages, Theories, and Axioms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
15.2 Decidability and Undecidability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
15.3 Proving Decidability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
15.3.1 Quantifier Elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
15.3.2 Model Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
15.3.3 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
15.3.4 The Decidability of Validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
15.4 Proving Undecidability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
16 Arithmetic & Incompleteness 125

16.1 True Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
16.2 First-Order Peano Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
16.3 Robinsons Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
16.4 Gdels Incompleteness Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
16.4.1 Representability in a Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
16.4.2 Representability in Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
vi CONTENTS
16.4.3 The Diagonalization Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

16.4.4 The Incompleteness Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
16.4.5 The Consistency-Strength Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
16.4.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
16.5 Nonstandard Models of Peano Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
16.5.1 PA and End-Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
16.5.2 Overspill and Underspill . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
16.5.3 Cofinal Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
IV Appendices and References 141

17 Orderings 143
17.1 Linear Orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
17.1.1 Discrete Linear Orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
17.1.2 Dense Linear Orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
18 A Brief Synopsis of Set Theory 145

18.1 Ordinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
18.1.1 Transfinite Induction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
18.1.2 Ordinals as Order Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
18.1.3 Ordinal Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
18.2 Cardinal Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
18.2.1 Cardinal Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
18.2.2 Basic Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
18.2.3 Regular and Cofinal Cardinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Chapter 1
First-Order Logic
Although our aim is, quite clearly, to speak eventually of full first-order logic, we begin much more simplywith the
basic building blocks thereof. The first of these basic building blocks is the notion of a signature or vocabulary.
1.1 Signatures
A signature is simply a listing of the symbols with which we will be workingour vocabulary, so to speak. This
vocabulary of symbols is divided into three sets based on what kind of thing a given symbol is intended to represent.
Signature
A signature is a (possibly empty) collection of the following sets:
i. The set of constant symbols; symbols which represent a particular entity
ii. For every n N, a set of n-ary relation symbols; symbols which represent a relation which either holds or
does not hold between tuples of n entities.
iii. For every n N, a set of n-ary function symbols; symbols which represent functions mapping n tuples of
entities to entities.
A signature L with no constants or function symbols is, for obvious reasons, called a relational signature; a
signature with no relation symbols is, for less obvious reasons, sometimes called an algebraic signature.
1.2 Languages
In addition to the symbols provided by the signature, every language also has a stock of variables. Any symbol can be
used as a variable provided that it is not already being used. The actual choice of variables is never important, and so
its common to restrict them to x0 , x1 , . . . with natural number or cardinal subscripts.
1.2.1 Terms
The terms of a signature will eventually play a critical role in our work; intuitively, a term is a combination of function
symbols, constants, and variablessomething which would denote an individual if all the variables were assigned.
1
2 CHAPTER 1. FIRST-ORDER LOGIC
Term
The terms of a signature L are the strings of symbols defined as follows:
Every variable is a term of L

Every constant is a term of L
For n > 0, if f is an n-ary function symbol of L, and t1 , . . . , tn are terms of L, then f (t1 , . . . , tn ) is a term of L
as well.
Nothing else is a term of L.

In other words, the terms of a signature (eventually, language) are simply the constants and variables closed under the
application of function symbols.
Closed Term
A term with no variables
If a term t is introduced as t( x) this means that x is a sequence (x0 , x1 , . . . , xn ) of variables, and every variable which
occurs in t is among the variables in x.
1.2.2 Atomic Formulas and Sentences
Atomic Formula
The atomic formulas of a signature L are the strings of symbols generated by:
If s and t are terms of L, then s = t is an atomic formula of L.
For n > 0, if R is an n-ary relation symbol of L, and t1 , . . . , tn are terms of L, then the expression
R(t1 , . . . , tn ) is an atomic formula of L.
Atomic Sentence
An atomic formula in which there are no variables orequivalentlyall of whose terms are closed.
Just as with terms, if we introduce an atomic formula as ( x), then ( s) means the atomic formula got from by
putting terms from the sequence s in place of all occurrences of the corresponding variables in x.
At times, it will be useful to refer to not only that atomic formulas of a signature, but also negated atomic formulas:
Literal
An atomic formula or negated atomic formula; paralleling earlier use, closed literals are those literals without
variables.
1.2.3 Formal Languages

Let L be an arbitrary signature. To achieve a (canonical) formal language L from the signature L, the symbols of L
will be supplemented by the normal logical connectives; that is,
not
and
or
= equals
for all
there exists
1.2. LANGUAGES 3
The terms, closed terms, atomic formulas, atomic sentences, etc. of the formal language L are defined to be the same
as those of the signature L. Finally, we define L itself as the class of all its well-formed formulas. The well-formed
formulas of L include, for any atomic formula A of L and any well-formed formula , both
A and
which is quite standard. In addition to these, however, we also allow possibly infinitary constructions with a given
cardinal upperbound on conjunction/disjunction and on quantification:
If L and || < , then

L and L
if L and x is a tuple of variables such that | x| < , then
( x) [] L and ( x) [] L
In addition to the well-formed formulas of the language, we also distinguish a particular class of these formulas:
sentences.
Sentence
A formula of a language L with no free variables.
Some thought quickly shows that every language has a unique signature; since our primary topic will be languages,
we will often conflate the two. In addition, although only occasionally observed elsewhere, given a language L
generated from a signature L with conjunction/disjunction and quantification upperbounds / respectively, we will
use L as the name of this language.
Example 1.
L Any number of disj./conj. and any number of quantifiers

L Less than formulas disj./conj. and any number quantifiers
L0 0 Only finitely many formulas disj./conj. and only finitely many quantifiers
L0 0 Only finitely many formulas disj./conj. and no quantifiers (called quantifier-free)
Although the term first-order technically refers to logics quantifying only over objectsa criterion which all of
our languages meet, regardless of the particular , chosenwe will specifically use the term for the language L0 0
which is our primary concern. If an infinitary language is being considered, it will be made explicit. In the L0 0 case,
the more complex definition of language given above simplifies to the one typically given in undergraduate courses:
(
R(t) for t terms
Wffs =
| ( ) | ( ) | x | x for , Wffs
We add the following definitions to our repertoire in order to mirror our natural intuitions on the similarity of
various formulas:
Variant
For a language L and corresponding signature L, we say that two n-ary formulas of L, ( x) and (y), are variants
of one another if ( x) may be obtained from (y) via a uniform replacement of variables (that is, ( x) = ( x)) and
vice-versa.
Note that the relation of variance over the formulas of a language is an equivalence relation; the cardinality of a first-
order language L is actually defined as the number of variant equivalence classes of its formulas. In simpler terms, the
actual choice of variables is immaterial; for our purposes R(x1 , x2 ) may as well be R(x3 , x9 ).
4 CHAPTER 1. FIRST-ORDER LOGIC
1.2.4 Standard Abbreviations

The following is a quick survey of standard abbreviations used here and elsewhere with the formulas of languages.
x,y for x = y
(0 n ) for {0 , . . . , n }
(0 n ) for {0 , . . . , n }
iI i for {i |i I}
iI i for {i |i I}
( ) for ()
( ) for ( ) ( )
x1 x2 . . . xn or x for x1 x2 . . . xn
x1 x2 . . . xn or x for x1 x2 . . . xn
for
Normal conventions for parenthesizing are also in effect; binds tighter than all the other logical symbols while
and bind tighter than and in turn.
1.3 Normal Forms

Propositional and first-order logic possess a number of normal forms; the most important are briefly canvassed below:
Conjunctive Normal Form (CNF)
Any truth functional formula of a first-order language L (i.e. no quantifiers) is equivalent to a formula which is
a conjunction of disjunctions of literals, i.e.
[0 ] [1 ] ... [n ]
where the i are finite sets of literals.
Disjunctive Normal Form (DNF)

Any truth functional formula of a first-order language L (i.e. no quantifiers) is equivalent to a formula which is
a disjunction of conjunctions of literals, i.e.
[0 ] [1 ] ... [n ]
where the i are finite sets of literals.
Prenex Normal Form

Any formula of a first-order language L is equivalent to a formula which has all its quantifiers preceding it, i.e.
x0 x1 . . . xn ( x)
where is quantifier-free.
Part I
Classical Model Theory
5
Chapter 2
The Basics of Models
The most basic example of a model theory is that of propositional logic: Emile Posts much vaunted truth table.
Example 2.
A simple truth table.
p q q p q
T T F F
T F T T
F T F F
F F T F
This simple model theory may be taken as our paradigm; it elegantly represents all facts necessary to determining the
truth or falsity of a given sentence in a formal and concise manner. If more generality were desired (as it may very
well be), imagine assigning truth values to all the atomic propositions available for use. Each of these assignments
gives a different model, a different situation.
2.1 Models
So-called models, interpretations, or structures are the primary topic of model theory; a model will play the role of a
particular instantiation of our signature, one way the constant symbols could have been assigned, one definition each
of the relation symbols could have been given, and one function each of our function symbols could have named. All
of these could, of course, have been different; a model simply gives one way the world could have been.
7
8 CHAPTER 2. THE BASICS OF MODELS
Model
For a given signature L, a model A is an object made up of the following four elements:
i. A set called the domain of A, written dom(A) or dom A; occasionally also referred to as the universe or
carrier of A. The elements of dom(A) are called the elements or entities of the model A. The cardinality of A,
symbolized |A|, is defined to be the cardinality of dom(A) or |dom(A)| .
ii. A referent for each of the constant symbols of the signature L. If c is a constant symbol of the signature, we
write cA for the element named by the constant symbol c in the model A.
iii. For each positive integer n and each n-ary relation symbol of the signature, an extension for this relation
symbol; that is, an n-ary relation on dom(A) (subsets of (dom(A))n ) which the n-ary relation symbol from the
signature names. If R is a relation symbol, we write RA for the relation named by the symbol R in the model
A.
iv. For each positive integer n and n-ary function symbol of the signature, an n-ary function on dom(A) (i.e. a
map from (dom(A))n to dom(A)) which this symbol names. If f is a function symbol, we write f A for the
function named by f in the model A.
As the definition implies, a model is always associated with a signature; to make this connection explicit, given
a signature L, a model A of that signature is called an L-model. Crucially, the domain of a first-order model is not
allowed to be empty, but the extensions of relations are. Finally, we adopt the following notation:
Models are symbolized by capital letters: A, B, C, . . .
Sequences of elements from a model are represented by placing a bar: a = a1 , a2 , . . . , an where a1 , a2 , . . . , an

dom(A). A tuple in/from A is a finite sequence of elements of A; it is an n-tuple if it has exactly length n. In a
minor breach of this notation, we will occasionally make use of an infinite sequence of elements and also denote
it with a, b, c, etc; to avoid confusion, whenever we do so it will be made explicit that the tuples being used are
possibly infinite.
For better or worse, there is no standard way of representing a model; it often falls to the reader to make educated
guesses as to what a writer intended.
2.2 Models and Formal Languages
From a given signature L, we may construct a number of formal languages, most notably L0 0 . Similarly, a signature
L determines a class of L-models. As foreshadowed above, L-models and a formal language L generated from L
share a very special relationship, namely that the class of L-models exhausts all possible interpretations of the formal
language L, it represents every different way things could be with respect to the language L.
To briefly recap, an L-model A provides an interpretation of the proper type for every symbol in the signature L. For
any constant symbols c, this means cA is some element of dom(A); for any n-ary relation symbol R or n-ary function
symbol f , this means an n-ary relation over dom(A), denoted RA , and an n-ary function f A : dom(A)n dom(A),
respectively. In doing so, however, a model does far moreassuming standard first-order logic, it sets a truth value for
every well-formed formula of any language L generated from the signature L. Mathematicians tend to symbolize this
relationship with |=, writing A |= whenever the formula is made true by the model A and saying A models .
We retain their phrasing, but reserve the symbol |= for the model-theoretic consequence relation. In lieu, we adopt ,
following the usage of modal logicians.
For a model A, |A| is the cardinality of A. For a signature L, |L| denotes the cardinality of L if and only if L is
infinite; otherwise, if L is finite, well say |L| = 0 .
2.2. MODELS AND FORMAL LANGUAGES 9
Truth in a Model
Let L be a signature, L a language generated from L, and A an L-model. Furthermore, let t( x) be a term of L and
a a sequence of elements from A at least as long as x. Then, by recursion on the complexity of the term t,
If t( x) is c for some constant c of the signature L, then t A [a] is cA

If t( x) is xi for some variable xi in x, then t A [a] is ai ; that is, the corresponding entry in the tuple a.
If t( x) is f ( s) for some function symbol f and tuple of terms s, then t A [a] is f A (s1A [a], . . . , snA [a])
Note that these three clauses fix the denotation of any term in L when it is paired with a tuple of elements from
dom(A) of appropriate length. This done, we may now define truth in a model or the relation. Again, the
definition is recursive, but now on the complexity of the formula. First, suppose that ( x) is atomic. Then, given a
tuple of elements a at least as long as x,
A R(t1 , . . . , tn )[a] is true for some relation symbol R in L if and only if ht1A [a], . . . , tnA [a]i RA
A t s [a] if and only if t A [a] = sA [a]

Finally, the recursive step:
A [a] if and only if A 6 [a]
A [a] if and only if for every formula ( x) , A [a]
A [a] if and only if for some formula ( x) , A [a]

A y [y, a] if and only if for every element b of A, A [b, a]
A y [y, a] if and only if for some element b of A, A [b, a]
Formula Equivalence in a Model

For a language L, corresponding signature L, and L-model A, we say that two n-ary formulas of L, ( x) and (y),
are equivalent in A if and only if A x [( x) ( x)].
2.2.1 Model-Theoretic Logical Consequence
Model-Theoretic Logical Consequence

Let L be a language, L the corresponding signature, T a set of sentences of L, and a sentence of L. T entails
or is a model-theoretic consequence of T in symbols, T |= if and only if every model of T is also a model of
.
Entailment in hand, a host of familiar logical concepts follow closely behind:
Validity
A sentence of a language L is valid if it is entailed by the empty set; that is, |= . is also called a logical theorem.
Satisfiability
A sentence of a language L is called satisfiable or consistent if there exists an L-model A such that A ;
similarly, a set of sentences of a language L is called satisfiable or consistent if and only if there exists an
L-model A such that A for every .
2.3 Theories
While models are, unsurprisingly, the primary topic of model theory, theories represent a close second place. Theories
along with the model theoretic consequence relationwill bridge the gap between the formal languages and models.
Theory
A set of sentences of a particular formal language L
Extension of a Theory
An extension of a theory T (or supertheory of T ) is a theory T 0 such that T T 0 .
Modeling a Theory
Let L be a language, L the corresponding signature, A an L-structure, and T a theory in L. A T if and only if
A for every sentence T .
Note that it makes sense to speak of a model of a theory (with no element assignment) since every formula under
consideration has no free variables.
2.3.1 Theories and Diagrams of Models

In moving from models to theories, we distinguish three theories of particular importance:
Theory of a Model
For an L-model A, the theory of A in the language L or ThL (A) is the set of all L-sentences such that A .
That is, all L-sentences true in A.
Atomic Diagram
For an L-model A, the atomic diagram of A or Diagat (A) is the theory containing all closed literals of the
language L+ such that A+ , where L+ is the language generated by expanding the signature L to include a new
constant for every element of the domain of A and A+ is A with each of these new constants representing the
intended element of the domain.
Couched in less formal terms, the expansion from L to L+ gives the ability to name every element of the model A.
To make this connection clear, the added constants are often subscripted with the names of the elements themselves,
e.g. we add ca for every a A. Of course, simply expanding the signatureand by extension the languagewith new
constants is not enough; the movement from A to A+ guarantees that each of these constants refers to its intended
element of A. A itself is, after all, not an L+ -model.
Elementary Diagram
For an L-model A, the elementary diagram of A or Diagel (A) is the theory containing all sentences of the
language L+0 0 such that A+ , where L+0 0 is the first-order language generated by expanding the signature L
to include a new constant for every element of the domain of A and A+ is A with each of these new constants
representing the intended element of the domain.
Just as above, the expansion from L to L+ gives the ability to name every element of the model A, while the move-
ment to A+ ensures these constants refer as intended. The new stipulation of L+0 0 above is the reason for the term
elementarya common synonym in model theory for first-order.
Finally, while its clear that, for any model A, Diagat (A) Diagel (A), ThL (A) is constructed using a different
(smaller) language. Thus, it is not true that Diagat (A) ThL (A)although, from the perspective of L+ , ThL (A)
Diagel (A). All three of the theories above will prove useful later, but even now they suggest a distinction in the
2.3. THEORIES 11
sameness of models. In particular, note that for two L-models A and B, B ThL (A) is a strictly weaker condition
than B+ Diagat (A) (for some extension B+ of B) which is in turn strictly weaker than B+ Diagel (A) (for some
extension B+ of B).
L -Equivalent
Two L -models A and B are L -equivalent if and only if T hL (A) = T hL (B), written A L B.
In practice, however, we will constrain ourselves to L0 0 , and so we will almost exclusively use:
Elementary Equivalent
Two L-models A and B are elementary equivalent if and only if ThL0 0 (A) = ThL0 0 (B), written simply as A B.
2.3.2 Canonical Models

The reverse directionfrom theory to modelis, as it turns out, a significantly harder task. Here we consider only the
most simplistic casealthough we return to the topic in our discussion and proof of compactness. Given a theory T of
atomic sentences, the following gives a canonical model A such that A T .
Equality Closed
A theory T of atomic sentences is equality closed if and only if
For every closed term t, t = t is in T

If (x) is an atomic formula, s and t are terms, (s) T , and s = t T , then (t) T .
Theorem 2.1 (Henkin Models for Atomic Theories).

Let T be a consistent theory containing only atomic sentences. Then, there is a model A such that
A T
T is exactly the set of atomic sentences in A.

Every element of A is of the form t A for some closed term t.
Proof Sketch.
First, generate the equality closure T 0 of T . Let the domain of A be the equivalence classes of = over the
closed terms of T 0 . For any n-ary function f and tuple t of n terms, set f (t)A = [ f (t)]; that is, set f (t)A to
be the equivalence class which includes the closed term f (t). Similarly, for any n-ary relation R, set RA to
be the set of tuples t such that R(t) is in T 0 .
The method of proof sketched above is a Henkin construction; later, well see that the same process can be expanded
to construct models for arbitrary consistent theories.
Chapter 3
Relations Between Models
3.1 From Homomorphism to Isomorphism

By earlier, a is ha0 , . . . , an1 i; for a function , we adopt the notation a for the tuple h(a0 ), . . . , (an1 )i.
Homomorphism
Let L be a signature and let A, B be L-models. A homomorphism is a function : dom(A) dom(B) such that:
For each constant symbol c in L, (cA ) = cB

For each n > 0, each n-ary relation symbol R in L, and each n-tuple a from A, if a RA , then a RB .
For each n > 0, each n-ary function symbol f in L, and n-tuple a from A, ( f A (a)) = f B (a)
In other words, the existence of a homomorphism from a model A to a model B implies that some portion of B looks
like Aalbeit in a very weak sense. In particular, over that portion of B the functions of L behave as they do in A
and the relations of B contain at least everything they do in A, although a single element of B may represent multiple
elements in A. This definition can clearly be strengthened, andin factwe do precisely this:
Embedding
Let L be a signature and let A, B be L-models. A function : dom(A) dom(B) is an embedding if and only if it
is injective, a homomorphism, and satisfies a strengthened second condition:
For each n > 0, each n-ary relation symbol R of L and each n-tuple a from A, a RA a RB .
Less formally, the existence of an embedding from a model A to a model B means that B contains an exact copy of A
somewhere inside of it.
Isomorphism
Let L be a signature and let A, B be L-models. A function : dom(A) dom(B) is an isomorphism if and only if
it is a surjective embedding. We say A is isomorphic to B if an isomorphism : A B exists, symbolized A B.
Isomorphism is, as usual, taken to mean that the models A and B are essentially the same; that they dont differ in
any interesting respect. In the future, for the sake of brevity, we denote functions : dom(A) dom(B) simply by
: A B. In addition, the following terminology will occasionally be used: homomorphisms : A A are called
endomorphisms of A; isomorphisms : A A are called automorphisms of A. The identity map on dom(A), IA , is an
automorphism for any model A.
The following follows easily from the definitions given.
13
14 CHAPTER 3. RELATIONS BETWEEN MODELS
Theorem 3.1 (Basic Properties).

Let L be a signature.
(i) If A, B, and C are L-models and : A B, : B C are homomorphisms / embeddings / isomorphisms, then
the composed map is a homomorphism / embedding / isomorphism from A to C
(ii) Let A and B be L-models. If : A B is an isomorphism, then the inverse map 1 : dom(B) dom(A) exists
and is an isomorphism from B to A
(iii) The relation is an equivalence relation on the class of L-models.

(iv) If A, B are L-models, : A B is a homomorphism, and there are homomorphisms : B A and : B A
such that = IA and = IB , then is an isomorphism and = = 1 .
Proof Sketch.
Simply leverage the definitions above
Theorem 3.2 (Basics of Homomorphism).

If : A B is a homomorphism,
(i) For every term t( x) and tuple a A of sufficient length, (t A [a]) = t B (a)
(ii) For every atomic formula ( x) and tuple a A of sufficient length, if A [a], then B [a]
Proof Sketch.
Simply leverage the definitions above
3.2 Submodels
Submodel
A is a submodel of B, often notated A B, if and only if
A and B are both L-models
dom(A) dom(B)
For all constants in L, cA = cB
For each n > 0, each n-ary relation R in L, and each n-tuple a in A, a RA a RB .
For each n > 0, each n-ary function symbols f in L, and each n-tuple a in A, f A (a) = f B (a).
or, equivalently, A and B are L-models with dom(A) dom(B) and the inclusion map i : dom(A) dom(B) is an
embedding.
3.3. PRESERVING FORMULAS 15
Lemma 3.1.
Let B be an L-model and X dom(B). X is the domain of a submodel of B if and only if
For all constants c, cB X.

For every n > 0, every n-ary function symbol f of L, and every n-tuple a from X, f B (a) X.
Proof Sketch.
Simply leverage definitions for both directions.
Generated Submodel
Let A be a submodel of B and Y a set of elements of B. A is the submodel of B generated by Y, notated A = hYiB ,
if and only if it is the smallest submodel of B which contains Y. hYiB is also called the hull of Y in B; we call Y
itself a set of generators for A.
Less formally, a generated submodel is the result of taking all the elements in the provided set, all the constants,
and closing under functions (apply the functions a step at a time to generate a hierarchy, the union is the generated
submodel).
Finitely Generated
A submodel B is finitely generated if and only if it is of the form hYiA for some finite set Y of elements and
L-model A .
Theorem 3.3 (A Bound on Generated Submodels).

Let B be an L-model and Y dom(B). Then, |hYiB | |Y| + |L|.
Proof Sketch.
Its easy to show that Y union an additional individual for every constant symbol in L has cardinality at
worst |Y|+|L|; similarly, the set of functions from L applied to elements of Y can be seen to have cardinality
at most |Y| + |L| + 0 . Altogether, then, the size of the domain of hYiB isusing the lemma abovebounded
by |Y| + |L| + |Y| + |L| + 0 . But this is equivalent to, max(|Y|, |L|, |Y|, |L|, 0 ) which is, of course, simply
max(|Y|, |L|, 0 ). Recall, however, that |L| 0 , and so this is equivalent to max(|Y|, |L|) = |Y| + |L|
3.3 Preserving Formulas

Thus far, our relations between models have focused on connections between the various parts of a modelnot the fact
the models make any sentence of the corresponding language either true or false. Taking this approach offers a new
way of looking at several old relations, as well as a trio of entirely new ones. On the very broadest level, we could
classify maps as to whether or not they preserve particular sets of formulas; that is, given a language L, two L-models
A and B, a set of formulas ( x), and a function : A B, is called a -map if and only if it preserves all the
formulas in ( x). Of course, not all sets of formulas are equally interesting; we begin by formalizing the notion of
preservation and presenting the most ubiquitous classes of formulae.
Preservation
Let L be a language and A, B L-structures. A function : A B preserves a formula ( x) of L if and only if for
every a A, A [a] B [a]
Using our previous work, It follows immediately that homomorphisms preserve atomic formulas while embeddings
preserve both atomic formulas and their negations (literals). Isomorphisms, naturally, preserve all formulae.
3.3.1 and Preservation
Existential
A formula is existential, an 1 -formula (pronounced E-1), or simply an -formula if and only if it is built out of
quantifier free formulas using only , , and . In other words, it can be written in prenex normal form as
x [( x)].
Universal
A formula is universal, an 1 -formula (pronounced A-1), or simply an -formula if and only if it is built out of
quantifier free formulas using only , , and . In other words, it can be written in prenex normal form as
x [( x)].
As the subscripts hint, 1 and 1 formulas are, in fact, members of a natural hierarchy on formulas; while we will use
these two most often, the full context may make our work easier to understand:
.. ..
. .
Quantifier Hierarchy
4 or 4 or
0 and 0 are defined as the quantifier-free formulas.
A formula is in n+1 if and only if it is in the smallest class of 3 or 3 or
formulas which contains the n formulas and is closed under ,
, and adding to the front. 2 or 2 or
A formula is in n+1 if and only if it is in the smallest class of
formulas which contains the n formulas and is closed under , 1 or 1 or
, and adding to the front.
Quantifier-Free
Theorem 3.4 (Embeddings Preserve -Formulas Upwards).

Let L be a first-order language and L the signature associated with it. If A, B are L-models and : A B is an
embedding, then preserves all -formulas.
Proof Sketch.
By induction on the complexity of formulas; note that any existential formula can be written in the form
x( x) where is in CNF or DNF. The base case of the induction is atomic and negated atomic formulas;
from here, its easy to see that and are also preserved. The case is slightly harder and relies on the
definition of an embedding.
Corollary (Submodels Preserve -Formulas Downwards).

Let L be a first-order language, L the signature associated with it, and A, B L-models. Then,
(i) If A is a submodel of B and is an -formula, then for every a A, B [a] A [a].
(ii) If T is an -theory, B T , and A is a submodel of B, then A T as well.
Notation. For a theory T in a language L, the theory containing all of the -sentences of L
which are consequences of T is denoted T .
3.3. PRESERVING FORMULAS 17
Lemma 3.2 (T -Models are Submodels of T -models).

If T is a theory in a first-order language L, then the T -models are precisely the submodels of the T -models.
Proof Sketch.
()
Simply leverage the previous preservation results.
()
Suppose A T for some theory T . We show the existence of a B such that B T and A B. Let U be
the set of all -sentences such that A . Assume T U Diagat (A) doesnt have a model. Then, by
compactness, there is some finite set F U such that T Diagat (A) |= F. Using compactness again,
there is some finite set E Diagat (A) such that T |= E F. But, by the definition of U, F
is logically equivalent to an -formula , and so T |= E but then E as a whole is a -
sentence (recall, E must be quantifier-free). Thus, E T and, by assumption, A+ E .
Therefore, A+ and A Fcontradicting the definition of U. By reductio, T U Diagat (A) must
have a model, and the reduct of this model to our original language fulfills the stipulations for our B.
Theorem 3.5 (os-Tarski Preservation Theorem).

Let L be a first-order language, T a theory in L, and ( x) a set of L-formulas ( x need not be finite); then, the following
are equivalent:
(i) For any two L- models A, B such that A T , B T , and A B, ( x) is preserved downward from B to A; that
is, for every a A, if B [a], then A [a] as well
(ii) For some set of -formulas ( x) ( x need not be finite), T |= x[( x) ( x)]; that is, is equivalent
modulo T to a set of -formulas
Proof Sketch.
()
Simply leverage the earlier -preservation results.
()
The case where ( x) is a theory follows easily by the previous proposition. Consider, then, the case where
( x) is not a theory. Let c be a tuple of new constants (expanding the language) and note that (c) is now
a theory. Note further that, like the more general ( x), (c) is preserved downwards between all models
A, B such that A T , B T , and A B. Let (c) denote (c) . Its trivial that T (c) |= (c). Consider
T (c), by the definition of and the previous lemma, any model of T (c) is submodel of a (c)
modeland thus itself a (c) model. It follows immediately that T (c) |= (c), and so
T |= (c) (c)
Noting that c occurs nowhere in T ,

T |= x[( x) ( x)]
Note that, in the theorem above, ( x), ( x), and x[( x) ( x)] need not actually be well-formed
formulas of L. Nonetheless, their meanings are still clear. More colloquially, the os-Tarski preservation theorem
states that the formulas preserved downwards by the submodel relation are always exactly the 1 formulas (any other
formula preserved downwards by the relation is just equivalent to an 1 formula). This, of course, quickly gives rise
to the following corollary:
Corollary (Embeddings Preserve Exactly 1 Upwards).

If L is a first-order language, T is an L-theory, and ( x) is an L-formula, then the following are equivalent:
(i) ( x) is preserved upwards by embeddings between models of T

(ii) ( x) is equivalent modulo T to an -formula
3.3.2 Elementary Preservation

Arguably the most important class for preservation, however, isnt found in the quantifier hierarchy; rather, it is the
class of all first-order formulas of a language.
Elementary Embedding
For a first-order language L, a function : A B is an elementary embedding if and only if for every formula
( x) of L and every a A, A [a] B [a]. In less formal terms, an elementary embedding preserves the
truth of all formulas applied to elements of A.
Elementary Submodel
Let A, B be L-structures. A is an elementary submodel of B or B is an elementary extension of Anotated A Bif
and only if A B and the inclusion map is an elementary embedding (that is, an elementary inclusion)
Theorem 3.6 (Vaught-Tarski Criterion).

Let L be a first-order language, A, B L-structures, and A B, then the following are equivalent:
(i) A B
(ii) For every formula ( x, y) of L and every tuple a from A, if B y (a, y), then B (a, d) for some element d
from A
Proof Sketch.
()
Simply leverage the definitions
()
Induction on formulas; the un-quantified cases all follow easily using vacuous existentials. For the case,
note that (ii) is easily extended to an if and only if and route through existentials.
That is, being an elementary submodela submodel which preserves all formulasis encompassed simply by the in-
clusion map preserving, for every n, the n formulas downwards.
3.4 Chains of Models
Chain
Let L be a signature, a cardinal, and (Ai : i < ) a sequence of L-structures. The sequence (Ai : i < ) is a chain
if and only if for every i, j such that i < j < , Ai A j .
3.5. REDUCTION AND EXPANSION 19
Union of a Chain [
Given a chain of models (Ai : i < ), the union of (Ai : i < ) is the L-model A = Ai defined as follows:
i<
[
dom(A ) = dom(Ai )
i<
For any constant symbol c, cA = cAi , for any i <

For any n-ary relation symbol R, a RA if and only if there is an Ai such that a RAi
For any n-ary function symbol f , define f A (a) = f Ai (a) where Ai is any model in the chain containing a
Preservation Under Chains

Given a chain (Ai : i < ) of L-models, a formula ( x) is preserved under the chain (Ai : i < ) if and only if
whenever there is a A0 such that Ai [a] for every i < , then A [a] as well.
Theorem 3.7 (Chains Preserve ).

Let L be a first-order language and L the associated signature; if ( x) is an formula, then ( x) is preserved under
chains.
3.4.1 Elementary Chains of Models
Elementary Chain
A chain of L-structures (Ai : i < ) such that the inclusion maps are all elementary inclusions.
Theorem 3.8 (Unions of Elementary Chains are Elementary [ Extensions).

If (Ai : i < ) is an elementary chain of L-structures, then Ai is an elementary extension of every Ai .
i<
Proof Sketch.
Use the os-Tarski theorem and induction on formulas; by the construction of A , one can argue that
the witness appears at some point in the chain and by the transitivity of elementary models/induction
hypothesis, that it a witness appears in the one under consideration. Thus, by the inductive hypothesis, this
last witness appears in A .
3.5 Reduction and Expansion
Reduct, Expansion
Let L, L+ be signatures such that L L+ . Let A be an L-model, B an L+ -model such that the interpretations of the
symbols in L coincide in A and B. A is called an L-reduct of Bin other words, A is B Lwhile B is a
L+ -expansion of A.
Lemma 3.3 (Elementary Diagram Lemma).

Suppose L is a first-order language, A, B are L-structures, c is a tuple of distinct constants not in L, (A, a) and (B, b)
are L(c) structures, and a generates A. Then the following are equivalent:
For every formula ( x) of L, if (A, a) (c), then (B, b) (c)
There is an elementary embedding : A B such that a = b
Note that, by the forward direction from the first to second statement, for any model D of the elementary diagram of
A in L, there is an embedding of A in the reduct D|L.
1 Preservation
Let L be a formal language, L the associated signature, and A, B L-models. A 1 B if and only if every 1 formula
( x) is preserved from A to B.
Lemma 3.4.
Let L be a formal language, L the associated signature. For any theory U and L-model G, if G U, then there exists
an L-model D such that D T and G 1 D.
Lemma 3.5.
Let L be a formal language, L the associated signature, and G, D L-models; if G 1 D, then there exists an L-model E
such that D E and G E.
Proof Sketch.
Let T 0 be the union of the quantifier free formulas of B and elementary diagram of A. Suppose that T 0
is not consistent; noting that constants from one do not appear in the other, derive a contradiction. By
compactness, T 0 has a model.
Theorem 3.9.
Let T be a theory. The following are equivalent:
For every chain of structures (Bi : i < ) if Bi T for all i, then B T as well.
There is an 2 -theory U such that |= T U.
Proof Sketch.
() Let U = { is a 2 sentence such that T |= }. Clearly, T |= U. From here, well build, starting with
A |= U and using the lemmas above, an elementary chain A A1 A2 . . . where each Ai 1 Bi Ai+1 .
Since T is preserved under chains, A |= T and A A , A |= T by the theorem above.
Chapter 4
Compactness and Strong Completeness
As stated above, every structure has a first order theorynamely, Th(A); but, quite obviously, not every theory has a
first order structure (simply consider a theory with two contradictory sentences). Earlier, canonical models provided a
means for constructing a model of any set of atomic sentences; in this section, we first consider what constraints are
needed to guarantee the existence of a model of an arbitrary theory, then use this result to prove two major theorems
for first-order languages.
4.1 Hintikka Sets
Hintikka Set
A theory T in a language L is a Hintikka set if and only if
(H1) For every atomic sentence , if T , then < T .

(H2) For every closed term t of L, t t T
(H3) For all closed terms s, t, and atomic formula ( x) if [s] T and s t T , then [t] T as well.
(H4) If T , then T
(H5) If T , then T . If T , then T for some .

(H6) If T , then there is a such that T . If T , then T for all .
(H7) If x (x) T , then (t) T for every closed term t of L. If x (x) T , then (t) T for some closed
term t of L
(H8) If x (x) T , then (t) T for some closed term t of L. If x (x) T , then (t) T for every closed
term t of L
The condition (H7) above is known as the Henkin propertyafter a famous logicianand the closed term t in the
property which makes x(x) true is commonly called a witness to x(x); its often said in this context that t witnesses
the truth of x(x).
Theorem 4.1 (Hintikka Sets have Models).

If T is a Hintikka set, then the canonical model A of the set of atomic sentences in T is a model of T .
21
22 CHAPTER 4. COMPACTNESS AND STRONG COMPLETENESS
Proof Sketch.
Let U be the set of atomic sentences in T . Observe that, by definition, U is equality closed. So, by a
previous theorem, the canonical model A is such that A U. By construction U is the set of atomic
sentences in A and every a A is of the form t A for some closed term. To show that A T , argue by an
induction on formulas using the Hintikka properties.
Theorem 4.2 (Finite Satisfiability, Completeness, and Witnesses).

Let T be a theory in a first-order language L. If
(i) Every finite subset of T has a model

(ii) For every sentence of L, either or is in T
(iii) For every sentence x (x) T , there exists a closed term t of L such that [t] T (that is, t is a witness for
x (x))
Then, T is a Hintikka set.
Proof Sketch.
Simply prove that each clause of the Hintikka set definition holds.
The conditions (i) (iii) above are commonly referred to as finite satisfiability, completeness, and having witnesses
respectively.
4.2 Compactness for First-Order Logic

Among model theorists, few theorems are as revered or as useful as compactness; compactness itself establishes a
fundamental limit to the expressive capabilities of first-order logic: only finite things are distinguishable. The infinite
all blurs together.
Theorem 4.3 (Compactness for First-Order Logic).

Let T be a theory of a first-order language L. If every finite subset of T has a model, then T has a model.
Proof Sketch.
Show the following:
Extension Lemma: If a theory is finitely satisfiable, then has an extension + in an expanded
first-order language L+ which is finitely satisfiable, complete, and has witnesses.
Then, argue as follows. Let be finitely satisfiable. Then, by the lemma, there is a + in an expanded
language which extends it and is Hintikka (an earlier theorem). By previous work, + has a model A
because it is a Hinikka set; thus, , a subtheory of + , also has a modelnamely A L.
Lemma 4.1 (Extending Finitely Satisfiable Theories).

If a theory in a first-order, countable language is finitely satisfiable, then has an extension + which is finitely
satisfiable, complete, and has witnesses.
4.3. STRONG SOUNDNESS AND COMPLETENESS FOR FIRST-ORDER LOGIC 23
Proof Sketch.
Add 0 many new constants to L, the signature in use; enumerate the sentences of L(0 ), noting that it is
possible to be so since |L| |L(0 )| = 0 . Induct starting with T , adding either a sentence or its negation
at each step according to whichever is finitely satisfiable with the current theory (note that you must prove
it is always possible to add at least one these). In addition, if the added sentence is an existential x[(x)],
also add [ci ], where ci is the least constant of the 0 that were originally added which has not yet appeared
in any sentence of our theory.
4.3 Strong Soundness and Completeness for First-Order Logic

Although our focus thus far has been exclusively on the model-theoretic logical consequence relation, |=, it will be
useful to occasionally pair it with a proof theoretic counterpart, `. This proof theoretic counterpart can be derived from
any of the standard systems; the actual choice doesnt matter for our purposes.
`
For a theory and sentence , ` read proves if and only if there is a proof from the sentences of to .
Theorem 4.4 (Soundness).

If ` , then |= .
Proof Sketch.
Soundness, as usual, is easily established simply by showing that each of the proof rules of a given system
preserves the truth in a model relation, .
Consistent
A set of sentences is consistent if and only if 0 where can be taken as either a primitive that is always
false or an abbreviation for x(x = x) x(x = x).
Theorem 4.5 (Completeness).

If |= , then ` .
Proof Sketch.
With clever usage of compactness, it is enough to show that whenever a theory is consistent, it is satisfiable.
Assume |= ; if is inconsistent, then ` trivially. To see this, assume, then, that is consistent. If
whenever a theory is consistent, it is satisfiable, we now have that is satisfiable. Consider, then, {}.
By assumption, this is unsatisfiable, and soby compactnessthere is a finite subtheory of such that
{} is unsatisfiable. Then, by above, {} is inconsistent. Utilizing the reductio rule of the
proof system (or whatever combination of rules is equivalent to it), we now have a proof establishing ` .
To prove that whenever a theory is consistent, it is satisfiable, first prove that if is consistent, then has
an extension + in an expanded language which is consistent, complete, and has witnesses. From here,
we have that + is Hintikka, and thus has a modelmaking it satisfiable. It follows immediately that is
satisfiable (as earlier, simply take the reduct).
24 CHAPTER 4. COMPACTNESS AND STRONG COMPLETENESS
Chapter 5
Definability
In model theory, definability comes in two distinct forms: the definability of a relation and the definability of a class
of models. Each notion is important in its own right, and so we consider each in turn. The primary distinction between
the two is simply that the former encompasses relations relative to a particular model while the latter describes classes
of models, and is thus relative to only a particular language.
5.1 Definability of Relations

To say an n-ary relation is definable in a some L-model M simply means that there is a formula of L such that all and
only those tuples of elements in the relation make the formula true; more colloquially, there is a formula that describes
the relation. Notationally, (An ) is intended as shorthand for the following set: {a | A [a]}. Of course, this notation
naturally extends to include parameters as well: if (x0 , . . . , xn1 , y) is an L-formula and b is a tuple from A, then
(An , b) = {a | A [a, b]}. Using the above, a formal definition for definable relations is given below:
First-Order Definable Relation without Parameters
Let L be a first-order language and A an L-model. An n-ary relation R is first-order definable without parameters
or -definable (pronounced zero-definable) if and only if there exists an L-formula ( x) such that R = (An )
Expanding on this basic definition, its also possible to define a weaker notion:
First-Order Definable with Parameters
Let L be a first-order language, A an L-structure, and X dom(A) a set of parameters. An n-ary relation R is
first-order definable with parameters or X-definable if and only if there exists an L-formula ( x, y) such that
R = (An , b) for b X.
It may be easiest to think of first-order definability with paramters as simply definability in an expanded signature with
constants for all the elements in the parameter set S . In general, proving that a relation is definable is accomplished
by exhibiting the formula which defines the relation; proving that a relation is not definable by invoking a connection
between automorphisms and definable relations:
Proposition 5.1 (Fixed by Automorphisms Definable Relation).

Let M be an L-model and R an n-ary relation over M. R is first-order definable if and only if R is fixed set-wise by
every automorphism of M
Thus, to show a relation is nondefinable, it suffices to exhibit an automorphism which either moves an element of
relation out of the relation or moves an element not in the relation into it. As one last note, an element a of a model
is said to be definable if and only if the relation {a} is. All of the above, thus, carries over directly into talk about
particular elements.
25
26 CHAPTER 5. DEFINABILITY
Example 3.
For simple examples of definable relations, consider:
(i) and the domain in any model
(ii) Zero in hZ, +i

(iii) > in hN, +, 0i
For simple examples of nondefinable relations, consider:
(i) Zero in hZ, <i
(ii) x1 + 3 = x2 in hN, +, S , 0i
5.2 Definable Classes of Models

One of the topics of preeminent concern in logic is what can be said in a particular logical systemthe topic of this
section.
Basic Elementary Definable

A class of models K is basic elementary definable (abbreviated, EC) if and only if there is a sentence such that
K is the class of all models which make true; that is, K = Mod().
Elementary Definable/Axiomatizable
A class of models K is elementary definable or elementary axiomatizable (abbreviated, EC ) if and only if there is
a theory T such that K is the class of all models which make T true; that is, K = Mod(T ).
Closely connected to these is the theory of a class of models:
Theory of a Class of Models

Let K be a class of L-models. The theory of K in the language Lnotated ThL (K)is the set of all sentences of L
that are true in every model in K.
Note that, for any class K, Th(K) is the largest possible theory which could define K; by simply leveraging the
definitions above, its easy to show that if any theory defines K, then Th(K) does.
While the primary method for showing definability (either elementary or basic elementary) remains simply ex-
hibiting the defining sentence or theory, showing non-definability proves much more difficult. Given a class K, the
standard strategy is to show that Th(K) union a theory whichas a wholeguarantees a model is not in K is finitely
satisfiable by models in K. A simple application of compactness then gives the desired result. The most famous and
important result along this line is owed to Thoralf Skolem.
5.3 The Skolem Theorems

Thoralf Skolem was a Norwegian logician who wrote prolifically during the early parts of the 20th century. Skolem
was, among other things, one of the detractors from the use of so-called uncountable infinites and worked to dissuade
others from using them. One of Skolems major results in this vein was showing that there existed countable models
of the uncountable real numbers; how Skolem did so is our topic for this section.
5.3. THE SKOLEM THEOREMS 27
5.3.1 Skolemization
Skolemization
Let T be a theory of a first-order language L. The Skolemization of T is a theory T + in a first-order language L+
where T T + , L L+ , and
(i) Every L-model A such that A T can be extended to an L+ -model A+ such that A+ T + .
(ii) For every formula ( x, y) of L+ with x non-empty, there is a term t( x) of L+ such that
T + |= x[y ( x, y) ( x, t( x))]
Skolem Terms
The terms t( x) of L+ in condition (ii) above. These terms are usually generated by simply adding, for every
formula ( x, y), a new function f to the signature which takes x and returns the witness y if it exists; in this case,
the f are called Skolem functions
Skolem Theory
A theory T is a Skolem theory orequivalentlyhas Skolem functions if and only if T is a Skolemization of itself.
Note that being a Skolem theory is relative to the language in use; if T is a Skolem theory under L, it will usually not
be a Skolem theory under L+ where L L+ .
Theorem 5.1 (Skolemization Theorem).

If L is a first-order language, then there exists L a language such that L L and a theory of L such that
Every L-model A can be expanded to a model A of
is a Skolem theory in L
|L | = |L|
Proof Sketch.
For every formula ( x, y), create a new function symbol f with arity | x| and add these new functions to
the signature. Make = { x[y( x, y) ( x, f ( x))] : ( x, y) is an L-formula}. Union and repeat.
Corollary.
Let T be a theory in a first-order language L. Then, T has a Skolemization T + in a first-order language L+ with
|L+ | = |L|.
5.3.2 The Downward Lwenheim-Skolem Theorem
Skolem Hull
Let T be a Skolem theory in a first-order language L, A a T -model, and X a subset of the elements of A. The
Skolem Hull of X is hXiA , the submodel of A generated by X.
Theorem 5.2 (Skolem Hulls are Elementary Submodels).

Let T be a Skolem theory in a first-order language L. If A is an T -model and X is a set of elements from A such that
the Skolem hull of X, hXiA , is non-empty, then hXiA A.
Proof Sketch.
Use the Tarski-Vaught criterion, noting that Skolemization is specifically designed to force the presence of
witnesses in the submodel.
Theorem 5.3 (Downward Lwenheim-Skolem Theorem).

Let L be a first-order language, B be an infinite L-structure, and X B. Then, there is an L-structure A such that:
X dom(A)
AB
|A| = max(|X|, |L|, 0 )
Proof Sketch.
Simply leverage the work above; expand A to A use the Skolemization theorem, then take the Skolem
hull around X.
5.3.3 The Upward Lwenheim-Skolem Theorem

Rather than simply proceed to the theorem immediately, we first consider two results of gradually increasing strength,
but similar technique to our eventual result.
Proposition 5.2 (Bigger Models Always Exist).

Let T be a theory which has at least one infinite model. Then, for every , there is a model of T of size .
Proof Sketch.
Just add constants to the signature and sentences saying that no two constants name the same element
(pairwise distinct) to the theory T . Prove this expanded theory is finitely satisfiable and invoke compact-
ness; finally, take the reduct to the original signature.
Fact. If B Diagat (A), then A can be embedded into B; simply consider f (a) = caB .
Proposition 5.3 (Bigger Extensions Always Exist).

Let T be a theory and let A be an infinite model such that A T . Then, for every cardinal , there is a T -model
extending B (A B) such that |B| .
Proof Sketch.
As earlier, introduce new constants to the signature, then add sentences asserting these constants are
all pairwise distinct from one another to T . In addition, however, also add Diagat (A) to T . Prove finite
satisfiability, invoke compactness, take the reduct.
5.3. THE SKOLEM THEOREMS 29
Theorem 5.4 (Upward Lwenheim-Skolem).

For any infinite L-structure A and any cardinal such that |A| and |L| , there is a model B such that A B and
|B| = .
Proof Sketch.
As earlier, add new constants to L, generating a new language L+ . Take Diagel (A) union sentences
asserting that the new constants are pairwise disjoint and show that this is finitely satisfiable. Next, prove
that A B L. While its very possible that B L is too big, simply leverage the downward Lwenheim-
Skolem theorem.
Chapter 6
Amalgamation
Amalgamation is a process centered on the following diagram:
D T
B T C T
Each of A, B, C, and D represents a model while the arrows signify embeddings. The model of primary interest in
the diagram is A; in order to investigate a model, its natural to view the ways in which it can be extendedhence B
and C. But even more information can be gained by looking not only at the extensions themselves, but also how those
extensions of A relate to one another in a model containing both of them: hence, D.
Of course, B and C are seldom allowed to be any extension whatsoever; we will assume there is a theory T such
that B T and C T which we desire to preserve in the amalgam of B and CD above. A final constraint imposed
on the process of amalgamation is, again, quite natural; since the entire process is used to investigate the model A,
the embeddings from B and C into D had better agree on A. In this context, D is said to be an amalgam of B and C
over A. As we will show, first-order logic has a number of amalgamation properties, general processes whichunder
specific assumptionsguarantee the existence of an amalgam with certain properties. As it turns out, these properties
can be leveraged to a variety of uses beyond the investigation of A, the interpolation and definability theorems in the
last section among them.
6.1 Elementary Amalgamation Property
The most basic amalgamation process takes, for a a set of generators of A, the theory T above to be the set of all
sentences in the language L(a)immediately giving that B, C, and D are all L(a)-equivalent to one another. In fact,
we will guarantee an even stronger condition: the embeddings from B and C into D will be elementary. Because we
can, we may as well take A to be a submodel of B and B to be a submodel of D, making the relations along the left
side of the diagram and respectively. Note, however, that doing so only makes things slightly simpler and doesnt
constitute an important extension of our original diagram.
31
32 CHAPTER 6. AMALGAMATION
Theorem 6.1 (Elementary Amalgamation Property).

Let L be a first-order language, let B, C be L-structures, and let a, c be (possibly infinite) tuples of elements of B, C
such that (B, a) (C, c). Then, there exists an amalgam D of B and C over haiB such that B D and the embedding
: C D is elementary. In a picture,
D ThL(a) (B)

B ThL(a) (B) C ThL(a) (B)

haiB
where : hai C is the unique embedding which takes a 7 c
Proof Sketch.
Do it without the constants first. Let T be the union of the elementary diagram of B and the elementary
diagram of C. Suppose T is not consistent; noting that constants from one do not appear in the other, derive
a contradiction.
6.2 Strong Elementary Amalgamation Property

An amalgamation is said to be strong if the overlap between the two extensions B and C of A have minimal overlap in
the amalgam D; that is, the embedded images of B and C in D have only the domain of A in common with one another.
Theorem 6.2 (Strong Elementary Amalgamation Property).

Let A, B, C be L-structures such that A B and A C. Furthermore, let a (not necessarily finite) name all the elements
of A. Then, there is an L-structure D such that B D and there exists an elementary embedding : C D such that
dom(B) (dom(C)) = dom(A).
D D

B C B C 0 = (C)

A A
Proof Sketch.
Let T = Diagel (B) Diagel (C) {b , c: for all b B A and all c C A}. Show that T is finitely sat-
isfiable; in particular, suppose not and note that we need only consider one element from each elementary
diagram (elementary diagrams are closed under conjunction). From here, simply use the universalization
trick: Diagel (B) |= (a, c) i, j ci = b j , and so A |= (a, d), but then B |= (a, d) (use the earlier).
6.3 The Strong Amalgam Property and Algebraicity

In the previous section we showed that, in first-order logic, two elementary extensions of a model always have a strong
amalgam; its natural to ask whether this result can be weakened. It turns out that it canbut not completely. A strong
6.3. THE STRONG AMALGAM PROPERTY AND ALGEBRAICITY 33
amalgam does not, in general, exist for any two extensions of a model. In order to guarantee existence, its necessary
to borrow a notion from mathematics: algebraicity.
Algebraic
Let B be an L-structure, X B, and b an element of B. b is algebraic over X in B if and only if there is a formula
(x, y), tuple a X, and n N, such that B (b, a) n y (y, a)
Algebraic Closure
The algebraic closure of X in B, algB (X), is the set of all b from B which are algebraic over X. A set X is
algebraically closed if and only if X = algB (X)
Proving each of the following is a good exercise in fleshing out an intuitive understanding of algebraic closures:
Fact. For any model B,
X algB (X); for a X, simply consider the formula x y
If B C, then algB (X) = algC (X); use the definition of elementary extension
If Y algB (X), then algB (Y) algB (X);
algB (algB (X)) = algB (X)
dom(halgB (X)iB ) = algB (X)
The major result in this area is, unsurprisingly, a strong amalgamation theorem dealing with algebraically closed sets:
Theorem 6.3 (Strong Elementary Amalgamation over Algebraically Closed Sets).

Let B and C be L-structures and a a sequence of elements in both B and C such that (B, a) (C, a). Then, there exist
an elementary extension D of B and an elementary embedding : C D such that a = a and dom(B) g(dom(C)) =
algB (a).
D ThL(a) (B)

B ThL(a) (B) C ThL(a) (B)

halgB (a)iB
Presented slightly differently,
Lemma 6.1.
Given a model B, a tuple a dom(B), and b dom(B) algB (a),
There exists an L-structure A such that B A, c dom(A) dom(B), and (A, a, c) (B, a, b)
There exists L-structures D and C such that a B D, a C D, and b < dom(C).
Proof Sketch.
Its easier if we absorb X into the language and add a constant cc for c. Let T = Diagel (B) {(cc ) :
(x) tpB (b/)} {c , d: d B}. If this is inconsistent, then Diagel (B) ` (cc ) ki=1 cc = di , thus
B |= x((x) ki=1 x = di ). Prove that b is algebraica contradiction.
Corollary (Strong Amalgamation for Elementary Equivalent Models).

Note that, as a special case of the theorem above, Let B, C be L-models. If B C, then there exists a L-models D and C 0 such that B
B C 0 = algB () = algC ().
Proof Sketch.
We want D to satisfy T = Diagel (C) Diagel (B) {b , c: b B algB (), c C algC ()}. Suppose its
inconsistent. Then there are finite sets Y B algB (), Z C algC () such that Diagel (B) Diagel (C) `
yY,zZ y = z. Thus we have that Y g(Z) , where g is the elementary embedding from C to D.
6.4 Interpolation and Definability
Theorem 6.4 (Elementary Extensions across Language Expansions).

Let B be an L1 -model and C an L2 -model for languages L1 and L2 such that L = L1 L2 . If B L C L, then there
is an L1 L2 -model D such that B D L1 and there is an L2 -elementary embedding from C into D L2 .
Proof Sketch.
Observe that if L2 = L, then L2 Diagel (B) L Diagel (C) is consistent because if (b) ` (z) for
Diagel (B), Diagel (C), then (b) ` x( x), B |= x( x), and C |= x( x)contradicting
C |= (c).
Let B0 be an L1 structure such that B L1 B0 and there is an embedding C L L B0 . Since C L L B0 L ,
there is C0 an L2 structure such that B0 L L C0 L and C L L2 C0 L . Let D = n=1 Bn , an L1 structure;
S
S
we have B L1 D. Observe that D = n=1 Cn , an L2 structure, and thus C L2 D. Observe that gn+1
Theorem 6.5 (Craigs Interpolation Theorem).

Assume L1 , L2 are languages with L = L1 L2 . Let be an L1 -sentence and an L2 -sentence such that ` .
Then, there is an L-sentence such that ` and ` .
Proof Sketch.
Let = { is an L-sentence such that ` } = the set of L-implications of . We show first that |= .
Let C be an L2 structure such that C |= . Consider ThL (C L ). Assume it is not consistent; then
` where ThL (C L ), but then and C |= , so Th(C)a contradiction. Let B be an
L2 structure such that B |= ThL (C L ). We have C L B, so there is an L1 L2 structure D such that
B L1 D L1 and C L2 D L2 . By assumption, B |= , so D |= , so D |= , so C |= .
6.4. INTERPOLATION AND DEFINABILITY 35
Theorem 6.6 (Beths Definability Theorem).

Let L and L+ be signatures such that L L+ . If ( x) is an L+ -formula, then the following are equivalent:
For A, B L+ -structures, if A L = B L, then for a a tuple of elements from A and B, A [a] if and only if
B [a]
There is an L-formula (x) such that ThL (B) ` x[(x) (x)].
Proof Sketch.
Chapter 7
Types, Saturation, and Categoricity
7.1 Types
A type isat basea description of a possible element or sequence of elements in a model, often with the use of
parameters from some distinguished set. More formally, let M be an L-structure and A dom(M). L(A) refers to
the language L augmented by constants for each element of A while ThLA (M) is, of course, the theory of M (or, more
properly, that extension of M wherein the added constants refer as intended) in LA .
Type
A set of L(A) formulas p with x1 , . . . , xn free variables is an n-type in M over A if and only if
Th(M)LA x p( x) is satisfiable
An n-type p is complete if and only if for any formula (x1 , . . . , xn ), either p or p. The set of all complete n
type in M over A is denoted S nM (A)
Proposition 7.1 (Type Finitely Satisfiable).

A set of L(A) formulas p with x1 , . . . , xn free variables is an n-type in M over A if and only if Th(M)LA x p( x) is
finitely satisfiable
The simplest way to generate a complete type is, of course, to describe an element that actually exists in the model.
If a dom(M), then tp M (a/A) is the set of all n-ary L(A)-formulas ( x) such that M [a]; tp M (a/A) is, as our
notation hints, obviously itself a complete type. Not all complete types correspond, however, to a tuple from the
model itself. Those that do are said to be realized in the model; those that do not are omitted.
Proposition 7.2 (Types are Realized in Elementary Extensions).

Let p( x) be an n-type in M over A. Then, there is a model A such that M A and A realizes p
Nota bene. If M N, then ThLA (M) = ThLA (N), and thus S nM (A) = S nN (A). That is, the set
of complete types is constant along the elementary substructure relation.
7.1.1 Tuples with the Same Type

If a complete type is a complete description of a tuple of possible elements in a given model, then two tuples having
the same complete type is, quite naturally, understood as the tuples being possibly identical for the model. Indeed, this
intuition can be made precise:
37
38 CHAPTER 7. TYPES, SATURATION, AND CATEGORICITY
Proposition 7.3 (Same Complete Types Automorphism in Elementary Extension).

If tp M (a/A) = tp M (b/A), then there is a model N such that M N and M has an automorphism such that a = b
Proving the result above requires a new bit of terminology and a clever lemma:
Partial Elementary Map
A partial elementary map from a model N to a model M is a function : B dom(N) dom(M) such that, for
any b B and L,
N [b] M [b]
Lemma 7.1 (Partial Elementary Maps Always Extend in an Elementary Extension).

Let M and N be L-models with B dom(N) and : B dom(M) partial elementary. For any b dom(M), there is
an elmentary extension M 0 of M and 0 : B {b} M 0 extending
Corollary (Partial Elementary Maps Extend Completely in Some Elementary Extension).

Let M and N be L-models with B dom(N) and : B dom(M) partial elementary. There is a model M 0 such that
M M 0 and extends to an elementary embedding : N M
7.1.2 Principal and Supported Types

Let L be a countable language and M an L-model.
Support
Let p( x) be a type of M over A dom(M). p( x) has support ( x) if and only if ( x) is an L(A)-formula(?) such
that ThL(A) (M) { x ( x)} is satisfiable and for every ( x) p( x), T x(( x) ( x)); ( x) itself is called a
support of the type p( x).
Note that the clause ThL(A) (M) { x phi( x)} is satisfiable simply guarantees that ( x) doesnt generate a contradic-
tion, but rather legitimately implies the formulas of p( x) against the backdrop of ThL(A) (M). A special case of support
arises when the supporting formula ( x) is itself a member of the type it supports; in such cases, ( x) is called a
generator for p and the type p( x) is said to be principal.
Principal
Let p( x) be a type of M over A dom(M). p( x) is a principal type if and only if it has support ( x) such that
( x) p( x).
In a striking continuity of definition, a type is said to be unsupported or non-principal if it is not supported or not
principal, respectively. From only the definitions given, several basic facts about supported and principal types are
immediate:
Fact. If a type p( x) over a set A is complete (with respect to all formulas with parameters from A), then it has
support if and only if it is principal over T .
7.1.3 Towards Saturation

All this talk of types and possible elements of a model leads naturally to the construction of models which realize as
many types as possible; our first step in this direction and towards the notion saturation is the following:
Theorem 7.1 (Elementary Extensions which Realize all Types of the Model).
For any L-model A, there is an L-model B such that A B and every type of A is realized in B.
7.1. TYPES 39
Proof Sketch.
List all of the complete types over A: { : }. Let B0 = A. By definition, there is a B1 such that
B0 B1 and B1 realizes 0 . Note further that 1 is finitely realized in A = B0 and B0 B1 thus, 1 is
finitely realized in B1 as well. 1 is thus a complete type of B1 , and there is thus a B2 such that B1 B2
and B2 realizes 1 (and 0 ). Generalize to an inductive argument.
7.2 Omitting Types

Before moving to models which realize as many types as possible, the opposite process bears mentioning. In particular,
by restricting to the countable case it becomes possible to exert near perfect control over which types of a theory T are
and are not omitted in elementary extensions of some model Awith a single exception.
Theorem 7.2 (Countable Omitting Types Theorem).

Let L be a countable, first-order language. If T is an L-theory and 1 , 2 , 3 , . . . are unsupported types over T , then
there is a countable model of T which omits all the types i for i N.
Proof Sketch.
Let L+ = L {c1 , c2 , c3 , . . . }. Generate T + , an L+ theory, that is finitely satisfiable, complete, has witnesses,
T T + , and for every m and every ci there is (x) m such that T + ` (ci ). The term model of T +
is a model of T with the required property.
To construct T + , build a sequence T 0 T 1 . . . of finite sets of L+ sentences such that iT T i is

consistent and at the end we let T + = T i=0 T i . At stage s, suppose we have T s and need T s+1 . If
S
s = 3e, work toward completeness (add the next sentence or its negation). If s = 3e + 1, work towards
adding witnesses. If s = 3e + 2, take the next pair hi, mi. Let = (ci , d) = T s . T is consistent with
(ci , d) and so xy (x, y). Since m is not supported by y(x, y), there is (x) m such that T 0
(y(x, y)) ( x). Thus, T 0 (c, d) (c). So, T T s (ci ) is consistent. Let T s+1 = T s {(ci )}.
We argue along the lines of the omitting types theorem. First, extend L to L+ by adding a constant for
every element (countably many) and countably many new constants. Let T 1 be the empty theory and note
that, by assumption, T T 1 has a model which is an L+ model.
Given T i1 , T i is generated as follows. Noting that T i1 must be finite and that the formulas of L+ are
countable, consider the next formula in the sequence of L+ formulas which hasnt yet been considered, say
i . If T T i1 {i } has a model, define T i = {i } T i1 ; if not, then every model of T T i1 is a
model, and so define T i = {i } T i1 . In either case, T T i has a model.
Next, we search through T i for any formulas of the form x(x). Every time such a formula is found,
take one of our new constants (of which only finitely have been used), say c , to name an element which
satisfies (this must be possible since T T i has a model. Redefine T i = T i {(c )}.
Finally,
Example 4.
Let T = Th(N). Note that ( x) = {x , 0, x , s0, x , ss0,. . . } is finitely realizable and so a type. It
cannot be supported, however, because N omits .
7.3. ATOMIC AND PRIME MODELS 41
7.3 Atomic and Prime Models

Throughout this section, its assumed that L is countable.
Prime Model
A model M is prime if and only if whenever M N, M N
Remark. Its easy to show with Lowenheim-Skolem thatsupposing a countable languageany prime model
must be countable.
Example 5.
The standard model of arithmetic N is prime
Every finite model M is trivially prime
Atomic Model
A model M is atomic if and only if tp M (a) is principal for every a dom(M)
Proposition 7.4 (Prime Atomic).

If a model M is a prime model of Th(M), then M is atomic.
Nota bene. The converse does not hold generally, just in the countable case; consider, for
example, uncountable dense linear orders without endpoints
Proposition 7.5 (Element-wise Definable Atomic).

If every element of a model M is definable, then M is atomic
Theorem 7.3.
Let L be a countable, first-order language and T a complete L-theory with infinite models.
If for every n < 0 , S n (T ) is countable, then T has a countable atomic model
If an L-model A is countable, atomic, and models T , then A is a prime model of T
Proof Sketch.
(1) Let 1 , 2 , . . . be a lits of all non-principle, complete types over T . This list is countable, so use the
omitting types theorem; (2) Argue generally, noting the assumption of completeness, and construct the
embedding piece by piece. In particular, prove and use that if a from A and b from B and tpA (a) = tpB (b),
then for every c A, there is d B such that tpA (ac) = tpB (bd) (note that the types in A are principle).
To do this, not that tpA (ac) is generated by some ( xy). Thus, T |= xy (( xy) ( xy)tpA (ac) ( xy).
A |= y(a, y), so B |= y(b, y). Let d be that y. If we iterate this principle, we obtain an embedding;
moreover, it must be elementary since the elements were chosen to have the same types and thus satisfy
the same formulas.
Lemma 7.2.
Let L be a countable, first-order language. If A = algA (), then A is atomic.
Proof Sketch.
Take a from A such that there is a formula ( x) and n such that A |= (a) and n x ( x). Pick so that
n is minimized. Show that A |= x(( x) ( x)) whenever ( x) tpA (a). Let ( x) tpA (a). Let a1 , . . . , an
be that tuples where is true. Assume A 6|= (ai ) for some i. Then, ( x) ( x) is a formula that a satisfies,
but which has fewer than n tuples which satisfy it totalcontradicting our choice of .
Theorem 7.4.
Let L be a countable, first-order language. Suppose A and B are atomic, countable L-models. Then, if A B, A B.
Proof Sketch.
We adapt the argument showing If A is countable and atomic, then A is prime to go back and forth
between the models and generate an isomorphism. Pick a1 A, b1 B such that tpA (a1 ) = tpB (b1 ). Pick
a2 A, b2 B such that tpA (a1 , a2 ) = tpB (b1 , b2 ). The end result of this procedure is a bijection g(ai ) = bi
which preserves types and is thus an isomorphism.
7.4. SATURATION 43
7.4 Saturation
The end of the previous section saw us realizing all the types of some starting model in one of its elementary extensions;
saturation takes this tactic one step further: realizing the types of a model in that very model.
Saturated
Let be a cardinal. A model M is -saturated if and only if for every A dom(M) with |A| < , every complete
1-type p(x) S nM (A) is realized in M. A model M is saturated simpliciter if and only if M is |M|-saturated
Example 6.
(Q, <) is saturated

(R, <) is not saturated
The countable random graph is saturated
Lemma 7.3 (Alternative Definition of Saturation).

The following are equivalent:
(i) M is -saturated
(ii) For every L-model B and (possibly infinite) tuples a, b of length < such that (M, a) (B, b), if d B, then
there is a c M such that (M, a, c) (B, b, d).
Theorem 7.5 (Saturated Models are Unique).

If M and N are saturated models of a theory T of the same size, then M N.
7.4.1 Homogeneity
Homogeneity is, as the term suggests, a self-similarity property; for a model M to be -homogeneous for some cardinal
, it must be that, when two parts of M of size < are elementary identical (using parameters from these parts), any
addition to one part can be mirrored by an addition on the other. That is, no difference occurs below the level. More
formally,
Homogeneous
Let M be a model and a cardinal. M is -homogeneous if and only if for any possibly infinite tuples a, b from M
with length < , if (M, a) (M, b), then for every d M, there is a c M such that (M, a, c) (M, b, d). M is
homogeneous simpliciter if and only if M is |M|-homogeneous.
Remark. Later, it will be popular to view this process as dealing in partial elementary embeddings; under this
scheme, the above states that any partial elementary embedding : A dom(M) where A dom(M) and |A| < can
be extended by an additional element c so long as |A {c}| < . The utility of this view clearly lies in the ability to
eventually construct a full elementary embedding of M to itself.
The definition of homogeneity should, of course, be quite familiar; it is, after all, a slight weakening of the alter-
native definition proved above for saturation:
Proposition 7.6 (-Saturated -Homogeneous).

If a model M is -saturated, then M is -homogeneous.
The notion of homogenity also has close ties to the notion of type; simply by definition:
Proposition 7.7.
Suppose M is a 0 -homogeneous model and a, b from dom(M). If tp M (a) = tp M (b), then for any c dom(M), there is
a d dom(M) such that tp M (ac) = tp M (bd)
Pushing this idea further and leveraging partial elementary embeddings,
Theorem 7.6.
If M and N are homogeneous, M N, |M| = |N|, and both realize the same n-types over for every n, then M N.
In addition, 0 -homogeneous models are rather ubiquitous entities:
Proposition 7.8 (Cardinality-Preserving, 0 -Homogeneous Elementary Extensions).

For any model M, there is a 0 -homogeneous model N such that M N and |M| = |N|.
Putting the facts above together, a characterization of not only 0 -saturation generally, but also countable saturated
models is possible:
Proposition 7.9 (0 -saturated 0 -homogeneous and S nM () Realized ).

A model M is 0 -saturated if and only if M is 0 -homogeneous and M realizes all types in S nM () for every n
Theorem 7.7 (Countable, Saturated Model S n (T ) is Finite).

A complete theory T has a countable, saturated model if and only if S n (T ) is finite for every n
7.4. SATURATION 45
7.4.2 Universality
Its clear that in moving from -saturation to -homogeneity something has been lost; luckily, that loss can be made
explicit:
-Universal
A model M is -universal for a cardinal if and only if for every B M with |B| < , B M. A model M is
universal simpliciter if and only if M is |M|+ -universal.
Universality is, as it were, the obvious dual to the previously-defined prime; as many models as possible elementary
embed into a universal model.
Theorem 7.8 (-saturated + -universal and -homogeneous).

Let be a cardinal. Then, the following are equivalent:
(i) M is -saturated
(ii) M is + -universal and -homogeneous
Additionally, if > 0 , then
(iii) M is -universal and -homogeneous
Corollary (Saturated Homogeneous and Universal).

A model M is saturated if and only if it is homogeneous and universal
Theorem 7.9 (-Saturated Models Always Exist).

For every model A and every cardinal , there is a model B such that A B and B is -saturated.
Theorem 7.10 (0 -Homogeneous Extensions).

Let A be a model such that |A| 0 and A B for some 0 -homogeneous B. Then, there is a model C which is
0 -homogeneous, A C B, and |C| = |A|.
7.5 Categoricity
Categorical
A theory T is -categorical for a cardinal if and only if T is consistent and for all T -models A, B such that
|A|, |B| = , A B. As usual, a model A is -categorical just in case Th(A) is.
A theory is said to be categorical simpliciter if and only if it is categorical for every cardinal the next result, however,
shows this to be a rather uninteresting result:
Theorem 7.11 (All Categorical Theories are Th(A) for a Finite Model A).
If T is a categorical theory, then T = Th(A) for some finite model A
Restricting to -categorical for some cardinal , however, proves much more nuanced. We already have two ready
examples of 0 -categorical models/theories:
Example 7.
(Q, <) / dense linear orders without endpoints is 0 -categorical

The random graph is 0 -categorical
For a, b dom(A), a and b are automorphic (notated a b) if and only if there is an automophism : A A such
that a = b. Putting together many of the results of this chapter,
Theorem 7.12 (Characterizations of 0 -Categoricity).

Let L be a countable, first-order language and T a complete L-theory with infinite models; then, the following are
equivalent:
(a) For every A T and every n < 0 , (An / ) is finite

(b) For some A T and every n < 0 , (An / ) is finite
(c) For some A T and every n < 0 , A realizes finitely many n-types.
(d) For every n < 0 , S n (T ) is finite
(e) Given variables x, there are only finitely many pairwise non-equivalent formulas ( x) modulo T
(f) For every n < 0 , every type p S n (T ) is principal over T
(g) T is 0 -categorical
7.5. CATEGORICITY 47
Proof Sketch.
(a) (b) is trivial.
(b) (c) since automorphic tuples realize the same type.
(c) (d), fix n. A realizes k n-types. Suppose that |S n (T )| > K. Let p0 , . . . , pk be types in S n (T ). There
is a formula 0j ( x) such that 0j ( x) p j and 0j ( x) < pi , i , j, setting j = 0j i, j 0i p j . Not that
j,i p j pi ki=1 j,i p j pi . For every i, since pi S n (T ) and i pi , T ` x i ( x), and thus
A |= x i ( x) with some ai . Show that each ai realizes a different type for each ia contradiction.
(d) (e), we know there are k types in S n (T ), p1 ( x), . . . , pk ( x). Show that if ( x) and ( x) are such
that i = 1, . . . , k pi pi , then T ` x(( x) ( x)), possibly by contradiction. Argue
combinatorially.
(e) ( f ), let 1 ( x), . . . , k ( x) be a maximal set of inequivalent formulas. Given p( x) S n (T ), take
( x) := ik,i p i ( x) p( x). If ( x) p( x), then for some i, T ` i and i p, so ( x) ( x).
( f ) (g), every model of T is atomic.
(g) ( f ), every type that is not principle can be omitted (be careful with countability), so ( f ) (g)
since we can have one model realizing the non-principle type and another which does notmeaning they
arent isomorphic.
( f ) (d), Suppose S n (T ) is infinite and every type is principle; let p1 ( x), . . . be a listing of the types in
S n (t). Each pi ( x) has a generator i ( x). Note that the set of the negations of these generatorssay ( x)is
finitely satisfiable (each generator must be false in every type it doesnt generate; every type is realized
in some model). Then it can be extended to a complete typebut our list was supposed to contain all the
types!
(d) (a), Note thatby (f)every model of T is atomic. Moreover, if A is atomic, if a, b from A and
tpA (a) = tpA (b), then cdtpA (ac) = tpA (bd). Using thisas earlierwe can construct an automorphism
f : A A with f (a) = b. Noting that there are finitely many types, there must then by finitely many
equivalent classes.
An interesting corollary of the theorem above is that 0 -categorical models are locally finite:
Locally Finite
A model A is locally finite if and only if every finitely-generated submodel of A is itself finite.
Corollary (o -categorical Locally Finite).

If a model A is 0 -categorical, then A is locally finite. In fact, there is a unique function f : N N depending only
on Th(A) with the property that for each n < 0 , f (n) is the least number m such that every n-generator submodel of
A has at most m elements.
Proof Sketch.
Suppose that a1 , . . . , an M generate an infinite model; if b1 , b2 ha1 , . . . , an i and b1 , b2 ,
tp(a1 , . . . , an , b1 ) , tp(a1 , . . . , an , b3 ) because b1 = t1 (a1 , . . . , an ) ....
A deeper result dealing in categoricityand beyond the scope of these notesis Morleys categoricity theorem;
while no proof is even attempted here, the statement is given and some of the basic notions of the theorem are developed
in the next subsections.
Theorem 7.13 (Morleys Categoricity Theorem).

If T is a theory in a countable L which is categorical for some uncountable cardinality, then T is categorical in all
uncountable cardinalities
Chapter 8
Other Model-Theoretic Constructions
Fair warning, the section on Frass limits below has never been edited while the section on ultrafilters is woefully
incomplete.
8.1 Frasss Construction

The countable case is a unique one; it borders on the finitewhere things are, as a general rule, simplerbut also allows
for infinite processes, useful techniques like chains and the models representing their unions. The combination of
these proves powerful and allows for several results not generally applicable. Moreover, since the majority of logics
fall into the countable case, its well worth our time to explore the class of countable models.
This section details a construction technique which generates particularly nice countable models; indeed, much of
our work in the following sections will make use of the properties these countable models possess. Perhaps the most
basic notion well make use of is the age of a model.
Age
Let D be an L-model. The age of D is the class, up to isomorphism, of all finitely generated submodels of D.
Thus, a class of models is an age if and only if it is the age of some model D.
Its easy to show that an age has the following two properties: joint embedding and heredity.
Joint Embedding Property

A class of models K has the joint embedding property (JEP) if and only if for every A, B K, there is C K such
that A and B embed in C.
A
Hereditary Property
A class of models K has the hereditary property (HP) if and only if for every A K and every finitely-generated
submodel B of A, an isomorphic copy of B is in K; in other words, K is closed under the operation of taking
finitely-generated submodels.
Of course, Frass proved more than just that ages have the joint embedding and hereditary properties; ages are
completely characterized by them:
49
50 CHAPTER 8. OTHER MODEL-THEORETIC CONSTRUCTIONS
Theorem 8.1 (Frasss Characterization of Age).

A countable class K of finitely-generated models is an age if and only if it has both the joint embedding and hereditary
properties.
Proof Sketch.
The forward direction is easy, so the reverse is sketched here. Our strategy is simply to take an arbitrary
class of finitely generated models with JEP and HP, then construct the model D for which it is an age.
Assume = {A0 , A1 , . . . } is an arbitrary class of finitely-generated models with HP and JEP. We build a
chain of models as follows:
Let B0 = A0 . In general, let Bn+1 be the model in such that both Bn and An+1 embed in it (JEP). Let
D = i< Bi . If A , then A embeds in D. If C is a finitely generated submodel of D, then all of the
S
generators must be present in some Bn , and thus C Bn . It follows that C and C D.
Alongside the joint embedding and hereditary properties, we single out one last property which will ultimately
prove necessary for our main result:
Amalgamation Property
A class of models K has the amalgamation property if and only if for every A, B, C K such that : A C,
: A B are embeddings, there is a D K and embeddings : B D, : C D such that =
D

B C

A
Although it is tempting to think so, the amalgamation property (AP) doesnt subsume the joint embedding property
as a special case. In addition to these three properties of classes of models, another property of particular models will
also prove useful:
Ultrahomogeneous
A model D is ultrahomogeneous if and only if whenever A and B are finitely-generated submodels of D, every
isomorphism A B extends to an automorphism of D.
The primary result of this section is the construction of a particular modelthe Frass limitfrom a class of count-
able models K which is itself countable; this particular construction is important because the Frass limit is not only
unique up to isomorphism (i.e. categorical), but also possess two other important properties:
Theorem 8.2 (Frasss Theorem).

Let K be a countable class of finitely-generated, countable models with the hereditary, joint embedding, and amalga-
mation properties. Then K is the age of a model D, called the Frass limit of K, which is countable, ultrahomogeneous,
and 0 -categorical.
8.1. FRASSS CONSTRUCTION 51
8.1.1 Constructing the Frass Limit
Weakly Homogeneous
A model D is weakly homogeneous if and only if whenever A and B are finitely-generated submodels of D such
that A B and : A D is an embedding, then there is an embedding : B D extending .
Proposition 8.1 (Ultrahomogeneous Weakly Homogeneous).

If a model A is ultrahomogeneous, then A is also weakly homogeneous
Proof Sketch.
Build a sequence of models B0 B1 . . . all in such that for every A , there is a j where A
embeds in B j and for every C A , every i , and embedding g : C Bi , there is a j and an
embedding g0 : A B j . If we can establish this, then the conclusion follows easily. Note that N4 N,
let (i, j, k0 , k1 ) 7 hi, j, k0 , k1 i. At stage s we will build Bs+1 . If s = 2e, let Bs+1 be obtained using JEP with
Bs and Ae . If s is odd, its of the form 2hi, j, k0 , k1 i + 1. Suppose i, j, k0 , k1 < s. Let C be the jth finitely
generated submodel of Bi . Let f 0 be the k0 th finitely generated submodel of Ak1 . If C C 0 , let Bs+1 be
obtained by using AP on C, Ak+1 , and Bs . If C C 0 , do nothing.
Lemma 8.1.
Let C, D be L-models. If the age(C) age(D) and D is weakly homogeneous, then C embeds in D.
Proof Sketch.
Consider increasingly large elements of age(C) and note that, while each element may be disjoint inside of
D, we can construct an extended embedding around the smaller element which is isomorphic to the larger.
Building successively, we obtain an embedding of C.
Lemma 8.2 (Back and Forth Argument).

Let C, D be L-models. If the age(C) = age(D) and C,D are weakly homogeneous, then for any finitely -generated
submodels A, B of C, D where A B by , extends to an isomorphism + : C D.
Proof Sketch.
Let A, B be arbitrary submodels such that A B. Let A1 be the model generated by A0 and c0 an arbitrary
element of C. Since the ages of C and D are identical, there is a A01 in D such that A1 A01 . Note, however,
that A0 embeds in A01 , and soby weakly homogeneousthere is a B1 such that B1 A01 and B B1 . Let B2
be the model generated by B2 and d0 an arbitrary element of D. Repeat the argument above, going from D
to C.
Corollary (Weakly Homogeneous Ultrahomogeneous).

A model is weakly homogeneous if and only if it is ultrahomogeneous
Simply use the lemma above with the same model for C and D.
Corollary (Frass Limits are Unique).

The Frass limit of a class of models K is 0 -categorical.
8.1.2 Countable Categoricity and Frass Limits

0 -categorical models are handy; luckily, we have a powerful means of constructing them: Frasss method. The only
difficulty to simply importing the Frass limit construction covered earlier wholesale is ensuring that the sizes of the
models in the class are kept under control by the number of generators which give rise to them, a concern which
gives rise to the following property:
Uniformly Locally Finite
A model A is uniformly locally finite if and only if there is a function f : N N such that for every n, if a is an
n-tuple from A, |hai| f (n).
Corollary (0 -Categorical Uniformly Locally Finite).

If a model A is -categorical, then A is uniformly locally finite.
Proof Sketch.
Note that if tp(a1 , . . . , an ) = tp(b1 , . . . , bn ), then ha1 , . . . , an i hb1 , . . . , bn i (think about what happens if
theyre not). Since S n (T ) is finite (say size m), there are at most m distinct submodels generated from n
elementsso set f (n) = m for all n.
Lemma 8.3.
Let a, b be tuples of A. Then the following are equivalent:
(a) hai hbi
(b) qf-tpA (a) = qf-tpA (b)
Proof Sketch.
(a) (b); Take an atomic formula, say R(t1 (a), t2 (a), . . . , tn (a)). Show that M |= R(t1 (a), t2 (a), . . . , tn (a))
if and only if M |= R(t1 (b), t2 (b), . . . , tn (b)), and so all quantifier-free formulas have the same truth values.
(b) (a); We want to produce f . Take t(a), define f (t(a)) = t(b). Consider atomic formulas and
show that f is well-defined, surjective, and injective.
Lemma 8.4.
If hai is finite, then there is a q.f. formula ( x) qf-tp(a) such that M |= x(( x) qf-tp(a)) (i.e. if b M and
M |= (b), then hai hbi).
8.1. FRASSS CONSTRUCTION 53
Proof Sketch.
Let hai = {t1 (a), . . . , tk (a)}. Let ( x) be the conjunctions of all formulas f (ti1 (a), . . . , til (a)) = t s (a) for all
of the l-ary function symbols f and all tuples (i1 , . . . , il ) kl . Repeat this for both R and R as well.
Homogeneous
A model M is homogeneous if and only if a, b if tp M (a) = tp M (b) a b.
Theorem 8.3.
If M is atomic, then M is homogeneous
Proof Sketch.
Assume tp(a) = tp(b). By atomic, there is ( x, y) generating tp(ac). M |= y(a, y), and so M |= y(b, y).
Let d be such that M |= (bd). Note that (bd) now generates the whole type.
Corollary.
If M is 0 -categorical, then M is homogeneous
Theorem 8.4.
Let L be finite. Let be a class of finitely generated models which is uniformly locally finite and has the HP, the JEP,
and the AP. Let M be the Frass limit of , T = Th(M). Then,
(a) T is 0 -categorical
(b) T has quantifier elimination
Proof Sketch.
M is ultrahomogeneous if and only if a, b M if q f tp(a) = q f tp(b) a b if and only if

a, b if q f tp(a) = q f tp(b) cd such that q f tp(a, c) = q f tp(b, d).
q f tp(a) = q f tp(b) (hai, a) (hbi, b)
Let L be finite. If hai is finite, there is a q.f. formula ( x) such that M |= x(( x) q f tp(a))
(a) Note that the Frass limit, M is countable, ultrahomogeneous, and age(M) = . We want to define a
theory T such that the countable models of T are exactly the Frass limit of (because then M |= T and
T Th(M) and if A |= Th(M) A |= T A M).
To say that age(M) , for each A in generated by a A (look at the q f -formula generated above for
model with finite generators). For each hai in , there is a q f formula hai ( x) such that B |= (b) hbi
hai. Let T 0 = { xhai ( x) : hai }.
We want T 1 such that if A |= T 1 , then age(A) . The number of models in with n generators is
|ha1 , . . . , an i| f (n). The number of models in with n generators is less than the total number of L-
models with f (n) elements which is finite. Let B1 , . . . , Bl be all the models in with n generators. For
each Bi , there is a formula Bi ( x) such that M |= Bi ( x) h xi Bi . Let T 1 = { xli=1 Bi ( x) : for n }.
To get ultrahomogeneous, use M is ultrahomogeneous if and only if a, b if q f tp(a) = q f tp(b) cd
such that q f tp(a, c) = q f tp(b, d) and note that all the q f tps are principal. That is q f tp(a) = q f tp(b)
if and only if li=1 Bi (a) Bi (b)
Example 8.
Let be the class of finite graphs with no function symbols. The Frass limit is a graph that has the
property if X, Y R are finite, there is v R such that w X, (v, w) E and w Y, (v, w) < E.
8.2. ULTRAPRODUCTS 55
8.2 Ultraproducts
Ultraproductsalthough considerably more complicated than the procedures given thus farare simply yet another
means of generating new models from existing ones while preserving certain first-order characteristics. The method
originates with Skolem in the 1930s and has been used extensively since os work in 1955. The fundamental
theorem of ultraproducts simply establishes the connection between ultraproducts and first-order characteristics; before
proceeding to it, however, its necessary to establish the notion of an ultrafilter.
8.2.1 Ultrafilters
Recall that P(I) gives the powerset of Ithat is, the set of all subsets of I.
Filter
A filter D over a non-empty set I is a set D P(I) such that
(i) I D
(ii) X, Y D X Y D
(iii) X D and X Y I Y D
That is, a filter D over I contains I, is closed under intersection, and is closed upwards under the subset relation.
Common examples of filters over a set I are:
The trivial filter D = {I}
The improper filter D = P(I) (not actually a filter)
For each Y I, the filter D = {X I : Y X}; called the principal filter generated by Y
The Frchet filter D = {X P(I) : X is cofinite}
Proper Filter
A filter D over a set I is a proper filter if and only if it is not the improper filter
Generated Filter
Given a non-empty set I and any subset X of P(I), the filter generated by X is the filter D given by
{F : X F and F is a filter over I}

A filter D is said to have the finite intersection property if and only if the intersection of any finite number of
elements of D is non-empty.
Proposition 8.2.
Let X be any subset of P(I) and let DX be the filter generated by X. Then,
(i) DX is a filter over I
(ii) D is the set of all Y P(I) such that either Y = I or for some Z1 , . . . , Zn X, Z1 Zn Y
(iii) D is a proper filter if and only if X has the finite intersection property
Ultrafilter
A filter D over a non-empty set I is an ultrafilter over I if and only if for all X P(I),
X DIX <D
Its common to simply say that D is an ultrafilter and assume that I = D.
Proposition 8.3 (Ultrafilters are Maximal Proper Filters).

(i) D is an ultrafilter over I
(ii) D is a maximal proper filter over I
Proof Sketch.
Simply leverage the definition of ultrafilter
Theorem 8.5 (Ultrafilter Theorem).

Let I be a non-empty set. If X P(I) and X has the finite intersection property, then there exists an ultrafilter D over
I such that X D
Corollary (Proper Filters Extend to Ultrafilters).

Let I be a non-empty set. Any proper filter over I can be extended to an ultrafilter over I
8.3. EHRENFEUCHT-MOSTOWSKI MODELS 57
8.3 Ehrenfeucht-Mostowski Models

At base, Ehrenfeucht-Mostowski models boil down to the thought that, given a particular theory T , by building a
model of T around a particular group of elements called an indiscernible sequence, the properties of the resulting
model as a whole can be controlled. In this area, a model A contains a linear ordering X just in case X dom(A), no
further connection with <X is required.
8.3.1 Indiscernible Sequences

Taking a linearly order set (X, <) and some k N, we may generate a new set [(X, <)]k , the set of all <-increasing
k-tuples of elements from X. Let f : [(X, <)]k B be a partition of this set (elements in the same partition are mapped
to the same element of B). It may help to think of f as a coloring of [(X, <)]k using |B| distinct colors. Define
f -Indiscernible
A subset Y of X is f -indiscernible if and only if for any two <-increasing k-tuples a, b from Y, f a = f b. In other
words, f is constant across [(Y, <)]k .
Put slightly differently, an f -indiscernible set is simply a set of elements from the original X all of whose <-increasing
k-tuples reside in the same partition, all of whose <-increasing k-tuples have received the same color. Although it is
far from obvious, the situation described above and combinatorial results guaranteeing the existence of f -indiscernible
subsets of certain sizes paves the way to many model theoretic results.
To introduce a bit more notation, suppose that it can be shown in the situation above that there exists an f -
indiscernible set Y of size . Since all of the above were arbitrary, this holds irrespective of the actual sets/colors
involved and merely in virtue of their relative sizes. Accordingly, let |X| = , |Y| = , and |B| = . Then,
()k
That is, for any linearly order set X of size and set of colors, a coloring of [(X, <)]k is guaranteed to have a set Y
of size at least such that all of [(Y, <)]k has the same color. This notation was chosen by Erdos and Rado so that if
a result of the form above holds, then it remains true for any increase of on the left or decrease of , , or k on the
right. Both the earliest and most important result for our purposes is:
Theorem 8.6 (Ramseys Theorem - Infinite Form).

For all k, n N,
0 (0 )kn
The key move in adapting these results to model theory is to think of our partitions, our colors, as determined by
whether or not the given <-increasing k-tuple satisfies a particular combination of formulas.
-Indiscernible Sequence
Let A be an L-structure and a set of L-formulas all with k free variables. Suppose Y dom(A) and < is a linear
ordering on Y. Y is a -indiscernible sequence in A if and only if for every formula ( x) and every pair of
<-increasing k tuples a, b,
A [a] [b]
Again, the above has little intuitive force in isolation, and so we tip our hand a bit. Given an ordering < on dom(A),
note that for any tuple a [(dom(A), <)]k , either A [a] or A [a] for each . Form a partition, then, on
the tuples in [(dom(A), <)]k so that tuples in the same part all behave the same with respect to every . Each part
in this partition is -indiscernible. In truth, -indiscernability isnt of considerable interest. It is, however, a stepping
stone to something which is:
Indiscernible Sequence
Let A be an L-structure, Y dom(A), and < a linear ordering on Y. (Y, <) is an indiscernible sequence in A if and
only if (Y, <) is simultaneously a {}-indiscernible sequence for every in L.
Theorem 8.7 (Constructing Models with Indiscernible Sequences).

Let T be a theory with infinite models. For any infinite linear order (I, <), there is a model M such that M T and M
contains an indiscernible sequence (xi : i I)
Proof.
Unsurprisingly, by compactness and Ramseys theorem. Expand L to include constants ci for every i I;
take
T 0 := T {ci , c j : i, j I i , j} {[ci1 , . . . , cim ] [c j1 , . . . , c jm ] : i, j [(I, <)]m , any L-formula}
Any model of T 0 is as requested, and so it remains only to show finite satisfiability. Consider any finite
subset of T 0 ; we may assume with no loss that it contains T and sentences dictating that all the mentioned
constants are not equal. All that remains are a finite number of formulas from
{[ci1 , . . . , cim ] [c j1 , . . . , c jm ] : i, j [(I, <)]m , any L-formula}
It follows that there is a maximum arity that occurs among the finite number of , say k. Take a countable
model A of T and consider [(dom(A), <)]k where < is an arbitrary linear ordering on dom(A). Let
denote the union of the and note that we may think of each of the as taking k free variables. Thus,By
Ramseys theorem, there exists a countably infinite -indiscernible sequence Y dom(A). Assign the
constants mentioned so that the ordering <I is preserved by the chosen ordering on Athe result is a model
of the finite subset of T 0 .
8.3.2 E-M Models

Given any consistent theory T , recall that it can be Skolemized; that is, it can be extended to a Skolem theory T in an
extended language L such that, for any model M of T , there is an L -extension of M such that M T . Furthermore,
any generated submodel of M must be an elementary submodel. Last section showed that, given a theory, producing
models with particular indiscernible sequences is a surprisingly easy taskso long as the theory has infinite models.
This one not only defines, at long last, E-M models, but also details the results which give them such importance.
Ehrenfeucht-Mostowski Models
An Ehrenfeucht-Mostowski or E-M model is the Skolem hull of an indiscernible sequence
Lemma 8.5 (Order Indiscernible Automorphisms Induce a Submodel Automorphisms).

Let T be a Skolem theory in L , and suppose that M T . If Y = (yi : i I) is an infinite indiscernible sequence
for M and : I I is an automorphism for the linear order I, then there is an automorphism : hYi M hYi M
extending .
Type of an Indiscernible Sequence

The type of an indiscernible sequence X in a model M is the set of all such that M [a] for a from X
8.3. EHRENFEUCHT-MOSTOWSKI MODELS 59
Lemma 8.6 (Type Preserving Order Changes).

Let T be a Skolem theory in L . Suppose that Y = (yi : i I) is an infinite sequence of order indiscernibles in a
model M of T . If J is any infinite linear order, there exists a T -model N containing a sequence of order indiscernibles
X = (x j : j J) and tp(X) = tp(Y).
Lemma 8.7 (Order Embeddings Induce Submodel Embeddings).

Let T be a Skolem theory in L . Suppose that Y = (yi : i I) is a sequence of order indiscernibles in a model M of T
and X = (X j : j J) is a sequence of order indiscernibles in a model N of T with tp(X) = tp(Y), then any embedding
: I J extends to an elementary embedding : hYi M hXiN .
Part II
Computability Theory
61
Chapter 9
Of Algorithms
Computability theory is, at base, simply an attempt to characterize the intuitive notion of an algorithm, a finite set of
simple, step-by-step directions. Although both computability theory and the notion of algorithm existed long before
computers, both areto the non-logicianmost manifest in these little boxes of metal and plastic. As any computer
scientist will happily tell you, algorithms are wonderfully powerful thingsespecially when they can be executed at
incredibly high speeds. Everything from watching a tv show, to playing a video game, to opening a word document is
governed by step-by-step algorithms encoded into your computer; incredibly complex algorithms to be sure, but still
the step-by-step sets of instructions that constitute our subject matter.
A natural first step towards formally characterizing the intuitive notion of algorithm is to generate a set of intuitive
conditions on how algorithms behave; the list below will serve as our basis:
Algorithm
1. Uniquely determined by a finite set of instructions

2. Takes inputs from some set I and may produce outputs from some set O
3. Given input i I, an algorithm A determines a sequence of execution steps called the computation of A on
input I
4. Given input i I, an algorithm A may reach a halting step with an output o O
In this last and only this last case, an algorithm A is said to output o O on input i I and we write fA (i) = o
to convey this fact. Generating a partial function fA in this manner gives the function computed by A, abbreviated of
course to simply fA . The function fA is partial because nothing requires the algorithm A to have an output for every
i I; thus, dom( fA ) I. As a last bit of symbolism, stating that g is a partial function taking a tuple of k inputs
from a set X and possibly returning an output from a set Y will often be abbreviated, following standard mathematical
practice, as g : X k 7 Y. Future references to g will often make the arity of the function explicit as a superscript, e.g.
gk .
To push forward and fully formalize the notion of algorithm, its necessary to specify a programming language,
a means with which to specify algorithms. There are a number of natural choices, andhistoricallyearly research in
computability theory gave rise to a myriad of suggestions. Luckily, they are all equivalent.
9.1 Turing Machines

Perhaps the best known of these suggestions is the Turing machine. A Turing machine M is:
A finite alphabet A with distinguished symbol A
A finite set of S of states with both a start and a stop state in S
A 2-way infinite tape divided into cells and a reading head
63
64 CHAPTER 9. OF ALGORITHMS
Each cell has some a A written in it (the from earlier can be interpreted as a blank)
A finite set of quintuples of the form: hq, a, b, R, si or hq, a, b, L, si where q, s S , a, b A, and L/R specify
movement right or left of the read/write head. Interpreted, in state q, reading a replace it with b and move one
square right/left, going into state s.
Additionally, if hq, a, b, M, si, hq, a, c, N, ti M, then b = c, M = N, and s = t.

For those unfamiliar, the quintuples specify the behavior of the Turing machine, andalthough tedious in practice
all known algorithms can, in principle, be carried out by an appropriately defined Turing machine.
Turing Computable
A partial function f : Ok 7 I is Turing computable if and only if there exists a Turing machine M such that
f = fM
9.1.1 The Church-Turing Thesis

The much affirmed Church-Turing thesis is the claim that an algorithm is computable in the intuitive sense if and
only if it is Turing computable; that is, for any algorithm A there is a corresponding Turing machine M such that
fAk = f Mk . In support of this idea are the closure properties of the class of Turing computable functions, the lack of a
counterexample, and the equivalence of Turing computability to other means of characterizing computability (register
machines, -recursive derivation, etc.). We will, as usual, assume the thesis throughout our work.
9.2 Functions and Recursion

Studying computability theory via Turing machines is possible, but ultimately proves tedious. Instead, we adopt a
different and more mathematically tractable strategy. The beginnings of this formulation appeared when the output o
of an algorithm A on input i was written earlier as
fA (i) = o
The standard function notation of mathematics, it turns out, carries over rather effortlessly into talk of algorithms, and
so we leverage this conceptualization as the basis of our analysis. Notice, moreover, that while the function fA doesnt
uniquely specify the algorithm Amaybe a different set of instructions produces all the same input-output pairs!it
does capture the fact that an algorithm exists which produces such and such output on such and such input. Some
thought shows, however, that this is all we truly care about; whether there exists any particular set of instructions with
such and such behavior is the intriguing questionand easily captured within our function notation. The fact that we
give up the ability to distinguish between particular sets of instructions doing the same thing is a very small loss in
return for a great deal of simplicity.
Thinking of algorithms as functions naturally leads to the question of what sets the functions map things between;
for complete generality, we might wish to always have the ability to specify an input set I and an output set O such
that the algorithm f under consideration is a partial map of the form I 7 O. This is certainly doable, but a better
alternative is readily available. Rather than allow wildly shifting input and output sets, assume that every function
maps from Nk to N for some k; at first glance, this might seem to unnaturally limit the algorithms we may consider. As
well soon see, however, the limitations imposed by this assumption are actually remarkably weak; simple encoding
schemes will allow us to effortlessly represent any algorithm with both a countable input and output set.
Having solidified the interpretation of algorithms as functions from Nk to N, the task is to lay out those function
operations which allow us to construct all and only the computable functions. This task will carry us through the next
three chapters, and centers on the notion of recursion. Recursion is, at base, simply defining something in terms of its
previous values (often called an inductive definition in mathematics)a process anyone with programming experience
knows well. As an intuitively computable operation, recursionalongside function composition and several simple
starting functionswill give us a first attempt at defining the computable functions, yielding a class known as the
primitive recursive functions.
Chapter 10
The Primitive and -Recursive Functions
To sum the previous discussion, the primitive recursive functions are an attempt to generate formally a class of func-
tions which corresponds exactly to the intuitively computable functions; in order to do so, we select three types of
functions as obviously computable: the constant functions, the successor function, and the projection functions. Next,
we select two means of combining these functions into more complicated functions which are also intuitively com-
putable: function composition and primitive recursion.
10.1 The Primitive Recursive Functions

Formally,
Primitive Recursive Functions
The class of primitive recursive functions is the smallest class C of partial functions f : Nk 7 N for any finite k
containing
f (x1 , . . . , xk ) = n for every k > 0 and n 0; the constant functions
f (n) = n + 1; the successor function

f (x1 , . . . , xk ) = xi for 1 i k; the projection functions
and closed under
i. Function Composition: Let f be a k-ary function and g1 , . . . , gk l-ary functions all in C, then the l-ary
function h(y) = f (g1 (y), . . . , gk (y)) is in C as well.
ii. Primitive Recursion: If h is a k-ary function, g a k + 2-ary function, and both are in C, then
f (0, x) = h( x)
f (y + 1, x) = g(y, f (y, x), x)
Primitive Recursive
A function f : Nk N is primitive recursive if and only if there exists a primitive recursive derivation g1 , . . . , gk
with f = gk where each gi is either one of the basic primitive recursive functions or obtained from a previous g j
( j < i) by composition or primitive recursion.
10.1.1 Primitive Recursive Relations

A k-ary predicate or relation in N is a subset of Nk ; the characteristic function of a k-ary predicate/relation R is, contrary
to the typical definition in other areas of mathematics, a function R : Nk {0, 1} such that
65
66 CHAPTER 10. THE PRIMITIVE AND -RECURSIVE FUNCTIONS
(
0 if x R
R ( x) = 1 if x < R
Primitive Recursive Relation

A relation R is primitive recursive if and only if the characteristic function of the relation, R ( x), is primitive
recursive.
10.2 Primitive Recursive Operations
Lemma 10.1. Y X
If h(z, x) is a primitive recursive function, then g1 (z, y) = h(z, x) and g2 (z, y) = h(z, x) are primitive recursive
x<y x<y
functions as well.
Proof Sketch.
Induction
Lemma 10.2 (Primitive Recursive Operations).

The class of primitive recursive relations is closed under , , , substitution of primitive recursive functions, permu-
tation of variables, adding dummy variables, and bounded quantifiers.
Proof Sketch.
Composition; let P be a k-ary primitive recursive relation and f1 , . . . , fk primitive recursive functions.
Define,
Q(y) if and only if P( f1 (y), . . . , fk (y))
Note that Q (y) = P ( f1 (y), . . . , fk (y)), and thus is primitive recursive.
Bounded Quantifiers; Let R be a primitive recursive relation and define
Q(z, y) if and only if x < y, R(x, z)
Note that Q (z, y) = R (x, z), and thus is primitive recursive.

Y
x<y
10.2. PRIMITIVE RECURSIVE OPERATIONS 67
10.2.1 Piecewise Definition
Lemma 10.3 (Piecewise Definitions with Primitive Recursive Components).

If

h1 (x) if R1 ( x)
2 ( x) if R2 ( x)
h

f ( x) =

..
.

h ( x) if R ( x)
k k
where each hi is a primitive recursive function, each Ri is a primitive recursive relation, and the Ri are both mutually
exclusive and exhaustive, then f is itself primitive recursive.
Proof.
Define f by
R1 ( x)) + h2 ( x)sg(
f ( x) = h1 ( x)sg( R2 ( x)) + + hk ( x)sg(
Rk ( x))
and note that this is primitive recursive.
10.2.2 Bounded Search

Continuing the trend, we now show that the primitive recursive functions are closed under bounded search. The or
search operator takes on the least value satisfying some relation; thus, z(R(z)) is the least z such that R(z); in this
section, the operator will be bounded, represented by <y for a given bound y.
Theorem 10.1 (Bounded Search is Primitive Recursive).

Let R be a primitive recursive relation; if
(
(<y z)[R(z, x)] if such a z exists
f ( x, y) =
y otherwise
then f is a primitive recursive function.
Proof.
f ( x, 0) = 0
(
f ( x, n) if R( f ( x, n), x)
f ( x, n + 1) =
n+1 otherwise

10.2.3 Coding Sequences

Given a finite sequence ha0 , . . . , ak i, we may encode it in a single number as
pha0 , . . . , ak iq = pa00 +1 pa11 +1 . . . pak k +1
where pi denotes the ith prime number. In addition, we set
pq = 0
Note that, for every k,

fk (a0 , . . . , ak ) = pha0 , . . . , ak iq
is a primitive recursive function. More generally, the following primitive recursive predicates and functions are con-
structed to make working with encoded sequences easy.
S eq(n) if and only if n is the encoding of a sequence
k if n = pha0 , . . . , ak1 iq for some sequence ha0 , . . . , ak1 i

(
lh(n) =
0 otherwise
if S eq(n) and i < lh(n)

(
ai
(n)i =
0 otherwise s
Less formally, S eq indicates whether a number encodes a sequence, lh gives the length of the sequence a number
encodes if it encodes one and 0 otherwise, while (n)i returns the ith entry (starting at 0) from the encoded sequence if
i is less than the length of the sequence, 0 otherwise.
Proposition 10.1.
i. S eq is a primitive recursive predicate

ii. lh is a primitive recursive function
iii. The function d(n, i) = (n)i is primitive recursive
Proof.
S eq(n) if and only if n = 0 p n[(Prime(p) Div(n, p)) (q < p[Prime(q) Div(q, n)])
prime(x)Div(n,x) (x)) if S eq(n)
X

sg(
lh(n) =

xn
0

otherwise
Finally, taking f (n) to give the nth prime:
<n x[Div(n, f (i) x ) Div(n, f (i) x+1 )] 1 if i < lh(n)

(
(n)i =
0 otherwise

10.2.4 Simultaneous and Course of Values Recursion
Theorem 10.2 (Simultaneous Recursion is Primitive Recursive).

Suppose that
f (0) = m0
g(0) = n0
f (n + 1) = h1 ( f (n), g(n))
g(n + 1) = h2 ( f (n), g(n))
If h1 , h2 are primitive recursive, then f, g are as well.
10.3. THE COMPUTABLE FUNCTIONS? 69
Proof.
Let f, g, h1 , and h2 be as above. Let f2 be the length 2 encoding function and define as follows:
(0) = f2 (m0 , n0 )
(n + 1) = f2 (h1 (((n))0 , ((n))1 ), h2 (((n))0 , ((n))1 ))

Note that not only is primitive recursive, but (n) also encodes the values f (n), g(n) for every n. Thus,
f (n) and g(n) can be retrieved as follows:
f (n) = ((n))0
g(n) = ((n))1
Both of which are, of course, primitive recursive constructions. By the arbitrary nature of f, g, h1 and h2
the theorem is established.
Theorem 10.3 (Course of Values Recursion is Primitive Recursive).

If f (n+1) is always definable via a single primitive recursion procedure Q given singleton access to all of its preceding
values f (0), . . . , f (n), then f itself is primitive recursive.
Proof.
Let p(n) be the primitive recursive function which produces the nth prime. Define
(0) = (p(0)) f (0)

0
(n + 1) = (n)[(p(n + 1))Q ((n)) ]
where Q0 is the primitive recursive procedure Q with singleton access to all of f s preceding values given
by the decoding function (n)i on (n), whichby constructionencodes each of them. Note that, by as-
sumption, f (0) is definable by a primitive recursive procedure (it has no previous values), and so the brief
use above is not illicit. It follows immediately that is primitive recursive. Using , its now possible to
give a primitive recursive construction of f itself:
f (n) = ((n))n
It follows by the arbitrary nature of f and Q that course of values recursion is primitive recursive, and thus
the theorem is proved.
10.3 The Computable Functions?

The primitive recursive functions, then, are a surprisingly large class; including the vast majority of everyday mathe-
matical functions and operations. Its natural, then, to wonder whether weve already succeeded; perhaps all algorithm
means is primitive recursive function! Life, of course, is seldom so compliant.
Theorem 10.4 (Primitive Recursive Computable).

There exists a computable function which is not primitive recursive.
Proof Sketch.
Utilizing the encoding scheme already defined, we may encode derivations of primitive recursive func-
tions; more importantly, however, we may encode them in such a way that it can be checkedprimitive
recursivelywhether or not a given number e encodes a valid derivation of a primitive recursive function.
Note that, by simply proceeding through the natural numbers, we now have a primitive recursive enumer-
ation within which every last primitive recursive functions encoding must appear. We may go further,
however. Once a derivation is fixed, its certainly algorithmically possible to run the derived function on
any appropriate input; that is, we should be able to construct a computable function such that
e (x) if e encodes a primitive recursive function e

(
(e, x) =
0 otherwise
But, with such a computable in hand, we may channel the spirit of Cantor and define : N 7 N by
(x) = (x, x) + 1
A moments thought shows that not only should be computable, but it also couldnt have been listed
in our enumeration! It follows immediately that there must exist computable functions which are not
primitive recursive, as asserted.
10.4 Unbounded Search & -Recursion

-recursion grants the ability for unbounded search, to simply continue looking until a value is foundor even to
never stop looking at all. This operation represents a fundamental increase in the strength of the operations we allow
ourselves, but at a serious cost; for the very first time, the functions under consideration need not be total, with -
recursion in play its possible that a function is simply undefined on some inputs. In particular, there are two sources
of incompleteness to be wary of; the first arises quite naturally from the unbounded operator alone:
Unbounded -Operator
For R(x1 , . . . , xk , y) a k + 1-ary relation, the expression
y[R( x, y)]
gives the least number y such that, for a given x, R( x, y) holds; if no such y exists, the expression is undetermined.
That is, an expression containing the -operator may be undefined because a satisfying y cannot be found. The
definition of -recursion, however, hints at a more subtle difficulty:
-Recursion
Given (x1 , . . . , xk , y) a k + 1-ary function, the result of -recursion on with respect to y is the k-ary function,
( x) = y[( x, y)= 0 (z y)[( x, z)]]
Note the shift from talk of relations in the definition of the -operator to discussion of a partial function in the definition
of -recursion. -recursion is not restricted to primitive recursive or even -recursive relations; rather, we allow -
recursion over any -recursive functioneven though many are partial. This allowance is necessary to achieve the
class of functions we desire, but requires users of -recursion to keep the (z y)[( x, z) ] clause in mind; while
there might be a least y satisfying ( x, y), -recursion on will be undefined if ( x, z) isnt defined for all lesser natural
numbers.
Despite this, its common to omit the (z y)[( x, z) ] clause for -recursion on primitive recursive or -
recursive relations; since these are defined everywhere, the result always behaves as expected. Although it is tempting
to do so, the same liberties must not be taken with relations that are not -recursive; indeed, doing so represents an
unwarranted expansion of the class of functions under consideration.
10.5. THE COMPUTABLE FUNCTIONS 71
-Recursive
A function f is called -recursive if and only if f can be built using primitive recursive functions by means of
composition, primitive recursion, and -recursion.
With this last addition, its possible to prove thatfor any Turing machine Mthe k-ary function f Mk is -recursive;
in other words, the -recursive functions are the computable functions, assuming the Church-Turing thesis.
Theorem 10.5 (Turing-Computable = -Recursive).

.
Given a partial function f .. Nk N, then f is Turing computable if and only if f is -recursive.
Proof Sketch.
()
.
Fix a Turing machine t and the partial function it defines ft .. N N. Define a primitive recursive predicate
T.Compt (p(x, y, L)q) where C = p(x, y, L)q with L coding a terminating Turing computation with machine
t, input x, and output y:
T.Compt (p(x, y, L)q) if and only if (C[T.Compt (C) (C)0 = x])1
Informally, a terminating Turing comp with (t, x, y) is a sequence L of machine status descriptions. A status
description is a sequence l = p(T, h, i)q where T specifies the tape (a finite sequence of 0s and 1s) , h the
head (< lh(T )), and i (in {0, 1, . . . , s} if machine has states 0 through s) the Turing machines state. Thus,
L = p(p(T, h0 , i0 )q, . . . , p(T, hm , im )q)q and initial case i0 = 1, T 0 = (0, 1, 1, . . . , 1)(x many 1s), and h0 = 0
and end case ik = 0 (termination state), T k = (. . . , 1, 1, 1, . . . , 1, 0 . . . ) (y many 1s) ik case i < lh(L) and
how (T i+1 , hi+1 , ii+1 ) relates to (T 0 , hi , ii )
10.5 The Computable Functions
Computable Functions
The class of functions which can be computed by algorithm; according to the Church-Turing thesis, the term is
synonymous with the -recursive/Turing-computable functions.
Its important to note that a member of the computable functions may be partial; to distinguish the set of total
functions within the set of computable functions, we define
Total Computable Functions
The subset of the computable functions containing all and only the total functions; that is, functions defined for
every n N; also known as the total recursive or recursive functions
10.5.1 Computable Relations

As earlier, a k-ary predicate or relation in N is a subset of Nk ; the characteristic function of a k-ary predicate/relation
R is a function R : Nk {0, 1} such that
(
0 if x R
R ( x) = 1 if x < R
Computable Relation
A relation R is computable if and only if the characteristic function of the relation, R ( x), is a total computable
function; also called recursive.
As a special case, note that for k = 1 we have also given a definition for what it means to say that a set A N is
computable.
Chapter 11
The Fundamental Theorems of

Computability Theory
As the section on primitive recursion argued, the computable functions should, intuitively, be capable of enumerating
themselvesproving that they actually can will occupy most of our time in this section. The constructions of this
section not only give a host of basic results about the computable functions, but also will showcase a number of
recurring tools in computability theory.
11.1 The Padding Lemma

As earlier with the primitive recursive functions, its easy to define a primitive recursive predicate which indicates
whether a given natural number encodes (according to our scheme) a valid derivation of a -recursive function; these
natural numbers are called indices or Gdel numbers.
Lemma 11.1 (The Padding Lemma).

Each computable function e has 0 many indices, and for each e, we may construct an infinite set Ae such that for
every y Ae , y = e .
Proof Sketch.
Simply add extraneous steps to the derivation encoded by e
11.2 The Normal Form Theorem

The normal form theorem is the first major step in constructing an enumeration of the computable functions and does
precisely what its name indicates; it provides a standard format capable of representing every possible computable
function. The key to this form is Kleenes T -predicatea primitive recursive relation T (e, x, C) which verifies C
encodes a computation of the function encoded by e on input x.
Theorem 11.1 (The Normal Form Theorem).

For every k 1, there exists a primitive recursive predicate T k (e, x, C)the Kleene T -predicatewith x = (x0 , . . . , xk1 )
and a 1-ary primitive recursive function U such that
e ( x) = U(C[T k (e, x, C)])
73
74 CHAPTER 11. THE FUNDAMENTAL THEOREMS OF COMPUTABILITY THEORY
As well see, it is Kleenes T predicate and the -recursion which do the work above; the primitive recursive U does
little more that output the appropriate value.
Proof Sketch.
The key to defining Kleenes T predicate is the notion of a computation; given a derivation f0 , . . . , fn , a
computation that fn ( x) = y is a list of statements of the form:
fi0 ( x0 ) = y0
fi1 ( x1 ) = y1
..
.
fn ( x) = y
where each line fia ( xa ) = ya is justified by the derivation of fn and earlier equations on the computation
list. Less formally, a computation should be thought of as a proof that fn ( x) = y for some x, y. Just as with
derivations, constructing a primitive recursive predicate to test for valid computations is little more than
picking a representation scheme and working through the details. Once this is accomplished, Kleenes T
predicate is defined as
0 C encodes a valid computation of e ( x)

(
T k (e, x, C) =
1 otherwise
while U is defined to simply pick out the value of ( x) from C.
11.3 The Enumeration Theorem

A function : Nk+1 7 N is k-ary universal if and only if for any k-ary computable function f , there exists e N such
that (e, x) = f ( x).
Theorem 11.2 (The Enumeration Theorem).

For every k N, there exists a computable k-ary universal function, k .
Proof Sketch.
For any k, simply define
k (e, x) = U(C[T (e, x, C)])
using the k-ary Kleene T -predicate.
Note that in k (e, x) = U(C[T (e, x, C)]) above U and T are primitive recursive relations and appears just
once. It follows immediately that only a single application of the -operator in a derivation suffices to produce any
-recursive function.
Corollary.
Let f : N N be computable, total, and a bijection. Then, f 1 is also computable
11.4. THE S NM THEOREM 75
Proof.
Define -recursive g : N N by, informally, fix e N such that f = 1e (by enum. thm., e exists).
Given y N, to compute g(y), search for some computation C and x N such that C says f (x) = y.
Then, output x. (that defines g(y) = x).
Formally,
C[x, L C[C = p(e , p(x)q, y, L)q Comp(C)]] 1
0
or, since we have total, x[ f (x) = y] (the first construction will skip non-total portions of functions, the
latter may stall).
11.4 The sm
n Theorem
Also known as the parameter or parameterization theorem, the sm n theorem establishes, for any n and m, the existence
of a primitive recursive function smn whichgiven the encoding of an m+n-ary function f and m parameterspicks out
an encoding which represents f , but with its first m inputs fixed to the given parameters. At first blush, this seems like
a rather trivial task for such a vaunted theorem; implicit in this operation, however, is that the sm
n functions effectively
allow traversal through an infinite hierarchy of enumeration functions 1 , 2 , 3 , . . . which are based in the same
encoding scheme. The uses for such traversal are, as we will see, quite extensive.
Theorem 11.3 (sm n Theorem).

For every m, n 1, there is a primitive recursive, m + 1-ary function sm
n such that for every e, a1 , . . . , am , and x1 , . . . , xn
m+n (e, a, x) = n (sm

n (e, a), x)
Proof Sketch.
An especially useful variant of this theorem is the index function theorem:
Theorem 11.4 (Index Function Theorem).

If f (a1 , . . . , am , x1 , . . . , xn ) is a m + n-ary primitive recursive function, then there is a particular primitive recursive
function s(a1 , . . . , am )called the index function of f such that for every a and x,
nS (a) ( x) = f (a, x)
Thus, given f (a, x), we may construct a list of parameterized functions fa ( x) = f (a, x) by varying the parameters a.
Using the theorem above, it follows that there is a primitive recursive function s that will allows us to convert each set
a of parameters into the index of the corresponding parameterized function.
Proof Sketch.
11.5 The Halting Problem

The halting problem for a computable k-ary function e and input x asks whether or not e halts on input x, which of
e ( x) and e ( x) obtains. Of course, this question by itself isnt incredibly interesting; whether there is, for a given
arity k, a k + 1-ary, computable function which is capable of solving the halting problem for any encoding e of a k-ary
computable function and any input x, however, is. That is, we desire a computable H such that
0 k (e, x)
(
H (e, x) =
1 otherwise
As it turns out, this is provably impossible; that is, there exist functions which are demonstrably non-computable.
Theorem 11.5 (The Halting Problem is not Computable).

For any k > 0, there is no computable function which solves the k-ary halting problem.
Proof Sketch.
Suppose H (e, x) solves the k-ary halting problem. Define a new k-ary function f as
0 H (x0 , x)= 1
(
f ( x) =
otherwise
and note that f is clearly computable; it follows that f = e f for some e f . Consider, then, f (e f , . . . , e f );
if H (e f , . . . , e f ) = 1, then f (e f , . . . , e f ) contra H (e f , . . . , e f ). If not H (e f , . . . , e f ) = 1, then
f (e f , . . . , e f ) contra H (e f , . . . , e f ) yet again. Thus, in either case, H (e f , . . . , e f ) answers incorrectly.
Corollary.
Let f be any computable function. Then, there is an inverse g for f such that g is computable, where an inverse is
taken to mean
g(y)= x if and only if y range( f ) y range( f )[ f (g(y)) = y]
Proof Sketch.
See the earlier proof for f total, bijective
11.6 Kleenes Fixed Point Theorem

The following result, owed like much of our work to Kleene, is important enough to have earned the monikers the 2nd
recursion theorem and even just the recursion theorem. v u
Theorem 11.6 (The Kleene Fixed Point Theorem).

Suppose f : N N is a total computable function; then for every k, there exists an e such that
ke ' kf (e)
In other words, e is a fixed point of f vis--vis the enumeration function k .

11.6. KLEENES FIXED POINT THEOREM 77
Proof Sketch.
Suppose, for the moment, that we can computably construct a table of k-ary computable functions (by
specifying their indices in our K-ary enumeration) that has the following property:
(?) For every computable function f , there is a corresponding row r f such that for every n N,
(i) if f (n), the function f (n) is the nth entry in r f
(ii) if f (n) , the nth entry in r f is instead the function which halts on no input
Given such a table and the closure properties of the computable functions, its easy to see that the diagonal
of the table must itself appear as a row of the table. We thus have the following picture:
e1,1 e1,2 e1,3 ...

e2,1 e2,2 e2,3 ...
e3,1 e3,2 e3,3 ...
.. .. .. ..
. . . .
Diagonal, row n: e1,1 e2,2 e3,3 ... en,n ...
.. .. .. ..
. . . .
Moreover, since our construction of the table is computable, we have a means of computably generating
the indices of just the diagonal of our table. Taking an arbitrary total computable f , then, we may construct
a new row f (Diagonal) using our procedure for generating the diagonal row (recall, by (?), there must be
a corresponding row). Note, however, that this new row must intersect the diagonal, giving a fixed point
of precisely the type required:
e1,1 e1,2 e1,3 ...

e2,1 e2,2 e2,3 ...
e3,1 e3,2 e3,3 ...
.. .. .. ..
. . . .
Diagonal, row n: e1,1 e2,2 e3,3 ... en,n ...
.. .. .. ..
. . . .
f(Diagonal), row m: f (e1,1 ) f (e2,2 ) f (e3,3 ) ... f (en,n ) ... em,m ' f (em,m) ...
As it turns out, its easy to generate a table meeting the requirements above; the k = 1 case is given
explicitly, but the strategy easily generalizes. Consider
1 (x,y) (z) 1 (x, y)

( 1
t(x, y, z) =
otherwise
Taking x as our row variable and y as our column, the t above generates a table of the required type;
moreover, the process is computable. Define:
t(x, y, z) = 1 (1 (x, y), z)

2
and so the sm
n theorem guarantees a total computable s1 such that, for any r and c,
t(r, c, z) ' 3et (r, c, z) ' s21 (et ,r,c) (z)
This s21 (et , x, y) allows, then, the effective generation of our table, while s21 (et , x, x) generates the diagonal
indices. We may carry out the remainder of the argument as above; since s21 (et , x, x) generates the diagonal
indices, m isby the construction of tthe index of f (s21 (et , x, x)), while the actual fixed point is given by
s21 (et , m, m).
1s2 (et ,m,m) (z) ' 3 (et , m, m, z)

1
By the definition of et ,
' 1 (1 (m, m), z)
and, finally, by the definition of m,
' 1 ( f (s21 (et , m, m)), z)
' 1f (s2 (et ,m,m)) (z)
1
An equivalent way of viewing the construction above is to notice that the fixed point e above, viewed as a function,
forces itself to be a fixed point relative to f by computing the function f (e ) and then mimicking this functions
behavior.
Corollary.
Let f be total computable; for every k there exists e such that
Wek = W kf(e)
Example 9.
As an alternative to our original proof of Rices Theorem, suppose I is a computable index set that is
non-trivial (, N, ). Get e using recursion theorem, define
b (y) if I(e)
(
f (e) (y) =
a (y) if I(e)

where a < I, b I. If I is computable, then there is a primitive recursive f such that (e) = f (e) .
11.6. KLEENES FIXED POINT THEOREM 79
Example 10.
Is there a partial recursive such that for every e, if We is finite, then (e) and |We | (e)? Define
if (e)
(

We =
{0, . . . , (e) + 1} if (e)
By the S nm theorem, we may obtain such an e. Define f so that
if (e)
(

W f (e) =
{0, . . . , (e) + 1} if (e)
(e) since otherwise We = is finite, and (e). But then, |We | = (e) + 1.
Theorem 11.7 (Uniform Fixed Point Theorem).

Let f (x, y) be total computable. Then, for any k, there is a total computable, unary function 1e (y) such that
k1 (x) ' kf (x,1 (x))

e e
Corollary (Double Fixed Point Theorem).

Suppose f (x, y) and g(x, y) are total computable; then, there is a pair of indices a, b such that
a = f (a,b)
b = g(a,b)
In other words, there are a, b such that Wa = {b} and Wb = {a}.
Proof.
Get a total recursive a = a(b) such that a(b) = f (a(b),b) for every b (by URT). Now, fix b such that
b = g(a(b),b)
by the ordinary recursion theorem. The pair (b, a(b)) then works.
Example 11.
Suppose is computable and A N, and for all e,
We A (e) and (e) A We
Claim: A is productive. The productive function is f where f (y) =? There is a total computable e(y) such
that for every y,
(ey )
(

Wey =
Wy (ey )
(z)
(

W sy (z) =
Wy (z)
where s(y, z) is total computable by S nm . Our productive function is f (y) = (ey ).
Chapter 12
Non-Computable Relations
Two chapters ago, the definition of a computable relation was given; it has been reproduced below for ease of reference:
Computable Relation
A relation R is computable if and only if the characteristic function of the relation, R ( x), is a total computable
function; also called recursive or decidable.
Computable relations, then, have a computable decision procedure for whether any particular input is or is not in the
set; rather appropriately, computable relations are also known as decidable, a computable function can always decide
whether a given input is or is not in the set. This is, undoubtedly, a nice property for a relation to have, but its far from
necessary. Indeed, its natural to immediately distinguish to other classes of relations:
Computably Verifiable
A relation R is computably verifiable if and only if it has a computable verifying function; also known as
semi-recursive or semi-decidable
Computably Refutable
A relation R is computably refutable if and only if it has a computable refuting function; also known as
co-semi-recursive
A verifying function (also known as a positive characteristic function) for a relation R is

(
R ( x) = 0 ifotherwise
+ R( x)
Similarly, a refuting function (also known as a negative characteristic function) for a relation R is
(
otherwise
R ( x) = 1 if R( x)
Before moving forward, we introduce two further notational abbreviations:
We = domain(1e )
Ee = range(1e )
and offer a warning. Up until this point, all relations have been well-behaved, i.e. computable, and so its easy to give
little thought to familiar operations on them. From here forward, however, the relations under consideration need not
be well-behaved, and thusfor example-recursion is no longer always admissible on whichever relation currently
holds our interest. Failure to internalize this expansion in our topic matter is liable to lead to large errors, and it
may serve the reader well to refamiliarize themselves with the limitations of the operations covered thus far vis-a-vis
relations.
81
82 CHAPTER 12. NON-COMPUTABLE RELATIONS
Theorem 12.1 (Computably Verifiable Relations are Domains).

A relation R is computably verifiable if and only if R is domain( f ) for some computable function f ; that is, every R is
We for some e
Proof Sketch.
By definition, a computable relation R has a computable verifying function e but then R simply is We
Corollary (Computably Refutable Relations are the Complements of Domains).

A relation R is computably refutable if and only if R is N domain( f ) for some computable function f ; that is, every
R is N We for some e
12.1 A Three-Fold Distinction

Throughout the next few sections well explore various different ways of viewing and manipulating relations; thus
far, a three-fold distinction has been introduced: only computably verifiable relations, only computably refutable
relations, and both computably verifiable and refutable relations (that is, computable relations). This is one way of
viewing relations, but it is by no means the only reasonable one. An alternative tactand one which ultimately proves
quite valuableis to group relations via the formulas which define them.
Of course, definability is a relative concept; in this case, the language of choice is the language of first-order
arithmetic; our signature contains a constant symbol 0, relation symbol <, and function symbols for multiplication,
addition, and successor. In addition, we assume that all of our work is being carried out in the standard model of the
natural numbers, denoted N. We, of course, make significant use of common abbreviations like , >, (x < y), 102,
5, etc. In short, we work withinhopefully!the readers intuitive understanding of the natural numbers.
Just as in model theory, a relation R( x) on the natural numbers is defined by a formula ( x) in the language of
first-order arithmetic if and only if the relation contains all and only those tuples of natural numbers which satisfy the
formula; that is,
R(n) N (n)
Using this identification of formulas with relations, its natural to generate both the arithmetical hierarchy of formulas
and the arithmetical hierarchy of relations, although only the first level will concern us for the moment. First, we
construct the hierarchy for formulas; call a first-order, LA formula 00 or 00 if and only if it is first-order equivalent to
some ( x) using only the functions/relations of the signature, bounded quantification, and the standard truth-functional
connectives. Base case established, let , be formulas of LA and define:
( x) is 01 ( x) is first-order equivalent to y[( x, y)] where ( x, y) is 00 and x is finite
( x) is 01 ( x) is first-order equivalent to y[( x, y)] where ( x, y) is 00 and x is finite

This establishes the first level of the arithmetical hierarchy of formulas; now, for its relation counterpart. If a relation
on the natural numbers is defined by a 01 /01 formula, we will call the relation itself 01 /01 . Next, define:
A relation R is 01 R is defined by both a 01 and a 01 formula
Alternatively, a relation R is 01 if and only if both R and R are 01 /01 . This gives the basis of the arithmetical
hierarchy of relations; in picture form, we have:
01 01 01 01
00 /00 01
Formula Hierarchy Relation Hierarchy
12.1. A THREE-FOLD DISTINCTION 83
In both cases, the arrows denote set inclusion upwards. Our emphasis will, of course, be on the relation hierarchy; its
important, however, to keep the formula hierarchy in mind since it represents the basis for the former.
Some basic ways of classifying relationscomputably verifiable, computably refutable, 01 , 01 , 01 in hand, its
natural to ask how these classifications relate to one another and our previous work; uncovering and clarifying these
relationships is our next task.
Lemma 12.1 (00 /00 -Definable is Primitive Recursive).

Let R be a relation over N; then, the following are equivalent:
(i) R is primitive recursive
(ii) R is defined by a 00 /00 -formula in N
Proof Sketch.
()
By induction; show that all of the starting primitive recursive functions are definable, then that composition
and primitive recursion are as well.
()
By induction on the complexity of ; note, in particular, that all of the operations we need to close under
have been shown to be primitive recursive.
Theorem 12.2 (01 is Computably Verifiable).

Let R denote a relation on N; then, the following are equivalent:
(i) R is 01
(ii) R is computably verifiable
Proof Sketch.
(i) (ii)
Suppose R is 01 ; that is, R is defined by y( x, y) for ( x, y) a 00 /00 formula. By the lemma above, ( x.y)
defines a primitive recursive relation Q. Thus, we may explicitly define a verifying function for R by:
+R ( x) = y[Q ( x, (y)0 , . . . , (y)k )= 0 (z y)[Q ( x, (z)0 , . . . , (z)k )]]
(ii) (i)
Suppose R = dom(ke ) for some e, k. Then, by our earlier work, R( x) if and only if there is an encoded
computation C such that T k (e, x, C). Recall, however, that T k (e, x, C) is primitive recursive, and thus,
defined by a 00 /00 formula of LA , (y, x, w). Since e is fixed, it follows immediately from the preceding
that R is defined in N by w (e, x, w), a 01 -formula.
Example 12.
Examples of 01 Relations:
(i) Every We under our enumeration scheme
(ii) Kleenes K-predicate; defined by: e K if and only if 1 (e, e)

(iii) The halting predicate; defined by hi, ji H if and only if 1 (i, j)
Corollary (01 is Computably Refutable).

Let R denote a relation over N; then, the following are equivalent:
(i) R is 01
(ii) R is computably refutable
Example 13.
Examples of 01 Relations
(i) Every N We under our enumeration scheme
(ii) Kleenes K-predicate; defined by: e K if and only if 1 (e, e)
(iii) The un-halting predicate; defined by hi, ji H if and only if 1 (i, j)
Corollary (01 is Computable).

A relation R over N is 01 if and only if R is computable.
2
The three-fold distinction present at the beginning of this section thus survives the initial characterization of rela-
tions via formulas; eventually, this simple picture will collapse in the face of an infinite hierarchy of relations, butfor
the momentwe may bask in the ability to sort relations into one of three intuitive classes.
12.1.1 Closure Properties and Connections

With the identification of the 01 relations and computably verifiable relations, 01 relations and computably refutable
relations, and finally 01 relations and computable relations, it remains to detail both which operations each class is
closed under and whichif anyoperations carry us from one class to another.
Theorem 12.3 (01 Closed Operations).

From previous work, 0 or computable relations are closed under all boolean operations, bounded quantifiers, and substitution o
1
computable functions.
Theorem 12.4 (01 , 01 Closed Operations).

In a similar vein, its easy to show Both 0 and 0 are closed under , , bounded quantification, and substitution of partial computabl
1 1
addition, 01 is closed under existential quantification, while 01 is closed under universal quantification
12.2. COMPUTABLY ENUMERABLE RELATIONS 85
Proof Sketch.
Leverage either the definition in terms of computable verifiability or definability by a certain class of
formula.
Moreover, as might be readily surmised from either the opposing natures of verifiability and refutability or existential
and universal quantification, Theorem 12.5 (01 and 01 Relations are Complements).
For any 1 relation R, the relation R is 01 .
0
Proof Sketch.
Simply leverage definitions
12.2 Computably Enumerable Relations

Before moving forward, there is a last class of relations which naturally distinguish themselves: the computably
enumerable relations. The insight behind this class is, quite simply, thatin studying computabilityrelations whose
members can be computably listed deserve special attention.
Computably Enumerable
A set/relation R N is called computably enumerable (often abbreviated to c.e.) if and only if R = or
R = range( f ) for some total computable function f ; in other words, there exists a total computable function
f : N R which enumerates the members of R. Also known as recursively enumerable.
A bit of works shows that, while computable enumerability is a new consideration, the class of computably enu-
merable relations is one with which were already quite familiar.
Graph of a Function
For a k-ary partial function f : Ak 7 B, the graph of f is {h x, f ( x)i : x A and f ( x)}; that is, the k + 1-ary
relation relating elements of dom( f ) to their corresponding range element.
Theorem 12.6 (Computable Function Computably Enumerable Graph).

A partial function f : Nk 7 N is computable if and only if graph( f ) is computably verifiable/01 .
Proof Sketch.
()
Suppose f = ke for some k, e; then, the relation f ( x) = y has positive characteristic function
w[T k (e, x, w) y = U(T k (e, x, w))] 0
which is -recursive, and thus computable.
()
Assume graph( f ) is defined in N by a 01 -formula of LA , say w( x, y, w). Then, ( x, y, w) is a 00 /00 -
formula, and so defines a primitive recursive relation Q( x, y, w). Thus,
f ( x) = w[Q ( x, y, (w)0 , . . . , (w)k )= 0 (z w)[Q ( x, y, (z)0 , . . . , (z)k )]] 0 + y
By inspection, f is computable.
Theorem 12.7 (Selection Theorem).

If R( x, y) is a k + 1-ary, computably verifiable / 01 relation, then there is a computable f such that
(i) dom( f ) = { x | yR( x, y)}

(ii) For a given x, if yR( x, y), then R( x, f ( x))
In other words, the computable f always picks out a witness to yR( x, y) for any input x and is only defined when such
a witness exists.
Proof Sketch.
Simply leverage the previous results and build f
Theorem 12.8 (Computable Enumerability is Computable Verifiability).

R N is computably enumerable if and only if R is computably verifiable / 01 .
Proof Sketch.
()
Construct an existential formula defining R using the Tarski T predicate
()
Define the computable f such that R = ran(F) by using recursion and the operator
We have, then, a plethora of names, but only three distinct classes of relations; since writing computably verifi-
able / 01 / computably enumerable invariably becomes unwieldy, we adopt the convention of referring to our classes
of relations by the names 01 , 01 , and 01 . The alternative characterizations in terms of verifiability, refutability, com-
putability, and enumerability should not, of course, be forgottenbut, one hopes, constant reminders are not particularly
necessary.
Proposition 12.1 (Infinite 01 Relations are the Ranges of Computable, Bijective Functions).
If A N is infinite and 01 , then there exists f : N A such that f is both bijective and total computable.
Theorem 12.9 (Infinite, C.E., and <-Preserving 01 ).

If A N is infinite, then the following are equivalent:
i. A = ran( f ) where f : N N is total computable and <-preserving, n < m f (n) < f (m)
ii. A is a 01 relation
Remark. Note that any finite set is computable; thus, if f : N N is total computable and <-preserving, then
ran( f ) is computable.
12.3. REDUCTION AND COMPUTABLE SEPARABILITY 87
12.3 Reduction and Computable Separability

Reduction for a class of sets is the property that, for any two sets of the class, each can be reduced to a subset of itself
such that the subsets are disjoint and the union of the subsets is unchanged from the union of the original sets; as it
turns out, the set of 01 relations over N is one such class:
Theorem 12.10 (Reduction for 01 ).

If A, B N are 01 , then there exists A0 A and B0 B both 01 such that
(i) A0 B0 = A B
(ii) A0 B0 =
Proof.
Let A = range( f ), B = range(g). Let n A0 k(n = f (k) j < k(n , g( j)), n B0 k(n =
g(k) j k(n , f ( j)); that is, we simply look at which function gets to an element first.
In the context of the arithmetical hierarchy, the natural and immediate companion to the reduction property is the
separation property; since 01 has reduction, we may immediately show that the class of its complements01 has
separation:
Corollary (Separation for 01 ).

If A, B are disjoint 01 sets, then there is a 01 set C such that A C and C B = ; that is, C separates A and B (note
that which 01 set is taken as A and which B is of no consequence).
Proof Sketch.
Apply the reduction theorem to A and B
Strangely enough, the primary utility of this section comes not from the positive results already given, but rather
the following negative result:
Theorem 12.11 (01 does not have Separation).

01 does not have the separation property.
Proof Sketch.
We define two disjoint 01 sets, A and B, as follows:
e A 1 (e, e)= 0
e B 1 (e, e)= 1
Suppose that there is a computable C such that A C and C B = . Then, the following defines a
computable function, and so has some index e0 in our enumeration:
(
1 if x C
e0 (x) =
0 if x < C
Note that e0 A B. If e0 A, then 1 (e0 , e0 ) = 0, and so e0 < Ameaning A * C. Suppose, then, that
e0 B; then, 1 (e0 , e0 ) = 1, and so e0 Cmeaning e0 < B. It must be, then, that no such C exists; that
is, A and B are inseparable.
Corollary (Not All Computable Functions can be Extended).

There is a computable function with no total computable extension
Proof Sketch.
Taking A, B from the proof above, define
(
0 if x A
g(x) =
1 if x B
If g f and f is total computable, then we may use f to define a separating 01 C:
x C f (x)= 0
Fact. If a total computable function sends every member of K into a set A and every member of K into a set B,
12.4 Many-One Reducibility

All our distinctions vis-a-vis relations have thus far collapsed into the basic three-way distinction between verifiable
relations (01 ; computably enumerable), refutable relations (01 ), and relations which are both verifiable and refutable
(01 ). Nonetheless, these classes cannot be entirely homogeneous since not all 01 , nor all 01 relations are 01 . Is this
the only dimension along which relations within each class differ? In other words, are all 01 , but not 01 relations
essentially the same? Are all 01 , but not 01 relations essentially the same? Answering these questions, of course,
requires a notion of what it means for relations to be the same. After presenting an intuitive notion of reductionfor
one relation being at least as hard as anotherwe begin the long process of showing that, surprisingly enough, not all
01 01 relations were created equal and, of course, the same for 01 01 relations.
The firstand most importantof our steps is formalizing precisely what it means for one relation to be at least
as hard as another; its intuitively obvious that such a relation should be both reflexive and transitive. A natural
additional criterion is that the process of reduction should itself be computable; indeed, its a short leap from these
basic considerations to our most basic notion of reducibility:
Many-One Reducible
For two relations A, B N, A is many-one reducible to Bin symbols, A m Bif and only if there is a total
computable function f such that for every e,
e A f (e) B
i.e., f 1 (B) = A. Also known as m-reducible.
Although implicit in much of work until now, its worth noting that only considering 1-ary relations doesnt result in
a loss of generalityencoding sequences is, after all, a primitive recursive operation. The notion of m-reducibility in
hand, its easy to see that it behaves as intended vis-a-vis computable relations:
Proposition 12.2 (Reducible to Computable Computable).

If A m B and B is computable, then A is computable
Proof Sketch.
Simply note that A (x) = B ( f (x)) where f is the function which gives the reduction A m B
12.4. MANY-ONE REDUCIBILITY 89
Before moving forward, we introduce two additional pieces of terminology; although defined in terms of 01 ,
identical definitions hold for the terms 01 -hard and 01 -complete.
01 -Hard
A relation A N is 01 -hard if and only if for every 01 relation B N, B m A; i.e., A is at least as hard to
compute as any 01 relation.
01 -Complete
A relation A N is 01 -complete if and only if A is both 01 -hard and 01 .
Our original question, then, asks whether every relation in 01 01 / 01 01 is, in fact, 01 /01 -complete.
12.4.1 The m-Degrees

To generalize even further, recall that for two relations R and S , if both R m S and S m R, R and S areintuitivelyof
the same computational difficulty; moreover, this relationship is precisely what interests us. We may readily symbolize
such a relationship by dictating the following definition:
R m S if and only if both R m S and S m R for unary relations R, S

It follows immediately from our previous work that m is an equivalence relation, and so divides relations over N into
equivalence classes of computational difficultythe m-degrees.
m-Degree
The m-degree of a relation A is its equivalence class under m ; that is, degm (A) = {B : A m B}
Our goal of completely characterizing the basic relation classes 01 , 01 , and 01 may then be readily rephrased into un-
derstanding the m-degrees of the relations in these classes; in particular, whether or not there are precisely three
m-degrees below the 01 /01 level.
12.4.2 Characterizing the 01 /01 -Complete Sets

Although weve already encountered several 01 -complete and 01 -complete relations, the most useful and intuitive of
these will prove to be Kleenes K predicate and its complement K:
Proposition 12.3 (K is 01 -Complete).
Corollary (K is 01 -Complete).
Index Sets and Rices Theorem
Index Set
I N is an index set if and only if for every a, b,
a = b (a I b I)
In other words, I is an index set if and only if, for some class F contained in the set of computable functions,
I = {a | a F}. There are two index sets that, one hopes, leap immediately to mind: and N. For obvious reasons,
we refer to these two as the trivial index sets.
Theorem 12.12 (Non-Trivial Index Sets are at least 01 /01 -Hard).

For any non-trivial index set I N, either K m I or K m I
Proof Sketch.
Suppose I is a non-trivial index set. Note that there exists a computable function f such that f (x) for all
noting the index sets are closed
inputs x; let e f denote the index of this function. Then, e f I or e f I;
under complementation, assume without loss of generality that e f I. By assumption, I , , so there
must exist some e0 Iand, by I an index set, e0 , e f . Consider the following function:
e f (y)
(
xK
g(x, y) =
otherwise
Note, in particular, that g is computable; it follows by the sm n theorem that there exists a computable
function s11 (x) such that
s11 (x) (y) = g(x, y)
Take the reduction function to be this s11
Corollary (Rices Theorem).

The only computable index sets are the trivial index sets: and N.
Corollary.
then I is 0 -hard.
If I is an index set, I , 0, I , N, and is indexed in I, 1
Theorem 12.13 (Rice-Shapiro).

For any n, if I N is an n-ary, 01 index set, then
e I k I such that nk ne and nk is a function with finite domain
Proof Sketch.
That is, any 01 , k-ary index set contains the index of a function just in case it also contains an index for a finite
function which extends into it.
Productive and Creative Relations
Productive
A relation B N is productive if and only if there is a total computable function f called the productive function
for Bsuch that for every e,
We B f (e) B We
Note that if B is productive, then B is not computably enumerable since, for any c.e. set We , there is always another
element f (e) in B, but not in We ; in other words, productive simply means effectively not computably enumerable
Completely Productive
B is completely productive if and only if there is a total computable f such that
f (e) B f (e) < W(e)
Theorem 12.14.
B is productive if and only if B is completely productive.
Proof.
()
() tricky proof with recursion thm. or (a) K is completely productive f (e) = e K iff e < We (f witnesses
comp. prod.) (b) if K m B, then B comp. prod. (as for prod.)
Proposition 12.4 (K is productive).

K is a productive relation.
Proof Sketch.
Set f (e) = e. Then need, if We K, e K and e < We . But, if e We , e Kbut then e < We . Thus,
e < We , and so e K.
Proposition 12.5 (Productive Relations contain Infinite, 01 Relations).

If B N is a productive relation, then B has an infinite, 01 subset.
Proof Sketch.
Note that, since B is productive, it has a productive function f ; note further that is a computably enumer-
able subset of B. Taking e to be its index, f (e ) is another element of Bmeaning { f (e )} is a c.e. subset
of B! In general, the pattern above can be computably cycled, giving an infinite, c.e. subset of B.
Theorem 12.15 ( K m Productive).

(i) A relation R N is productive
(ii) K is m-reducible to R
[Myhill 1955]
Proof.
Let fm be the many-one reduction of B to C and fB the productive function for B. Define fC (x) =
fm ( fB (s11 (x))) where s11 is defined as follows. Set
g(e, x) = 1 (e, fm (x))

1
Noting that g is computable, by the sm
n theorem there must exist a function s1 such that
s11 (e) (x) = g(e, x)
Noticing that W s11 (e) = fm1 (We ) makes it easy to show that fC is, in fact, a productive function for C.
Corollary (01 -Hard Productive).

Every 01 -complete relation is a productive relation
Creative
A relation A N is creative if and only if A is 01 and A is productive.
Fact. By definition, no creative relation can be computable; indeed, the name creative derives from this built
in non-decidability
Fact. Not every productive set has a creative complement
Computably Isomorphic
Two relations A and B are computably isomorphic if and only if there exists a total computable, bijective function
from A to B
Being computably isomorphic means that two sets are isomorphic from the limited perspective of the computable
functions; that is, there is no difference between the computational difficulty of the two sets according to any of the
measures considered thus far.
Theorem 12.16 ( K m Creative).

(i) A relation R N is creative

(ii) K is m-reducible to R
(iii) R is computably isomorphic to K
[Myhill 1955]
Proof Sketch.
Corollary (01 -Complete Creative).

Every 01 -complete relation is a creative relation; moreover, all of these are computably isomorphic to one another.
They are all the same according to our current measures.
Chapter 13
Relative Computation and Turing

Reducibility
Having explored the bottom of the arithmetical hierarchy, we now set our sights on the hierarchy as a whole; its
not clear, however, that the rest of the arithmetic hierarchy is worth studying. For the lowest level, we had a clear
motivation in the notion of computability; the higher levels, though, dont at first seem to admit anything similar. In
this section, not only is the remainder of the hierarchy generated, but natural problems are shown to reside at levels
beyond 01 and 01 ; in the next section, we take up the task of giving an intuitive framework for thinking about these
higher levels.
13.1 Higher Complexity Relations
As before, call a first-order formula of LA , the language of first-order arithmetic, 00 or 00 if and only if it is logically
equivalent to some using only bounded quantification and the standard truth-functional connectives. Inductively,
is 0n+1 is logically equivalent to x where is 0n and x is finite
is 0n+1 is logically equivalent to x where is 0n and x is finite
This gives rise to an infinite hierarchy of LA formulas; as earlier, we again associate the relations on the natural
numbers defined by each of these formulas in N with the formulas themselves, generating a hierarchy of relations
over N. We also define,
A relation R is 0n R is defined by both a 0n and a 0n formula
Alternatively, a relation R is 0n if and only if both R and R are 0n -relations.
95
96 CHAPTER 13. RELATIVE COMPUTATION AND TURING REDUCIBILITY
.. .. .. ..
. . . .
03 03 03 03
03
02 02 02 02
02
01 01 01 01
00 /00 01
00 /00 = P.R.
13.1.1 Closure Properties & Basic Results

0n , 0n , and 0n relations are closed under , , bounded quantification, substitution of total computable functions, and
substitution of partial computable functions after the n = 1 level. In addition, 0n relations are closed under negation,
0n relations are closed under existential generalization, and 0n relations are closed under universal generalization. All
of these results follow easily by leveraging the formula-based definition of the hierarchy.
Universal Relation
A k + 1-ary relation U( x) is universal for some class of relations D if and only if
i. U is itself in D
ii. Whenever a relation R is in D and k-ary, then there exists an e N such that
R = U(e, y) = {y : U(e, y)}
Theorem 13.1 (0n -Universal Relations Exist).

For any n 1 and any k, there is a 0n -universal k + 1-ary U.
Proof Sketch.
Proceed by induction on n, noting that the n = 1 case has already been established. For the inductive step,
suppose the result holds for n and consider n + 1. Given k, let V be 0n -universal for k + 1-ary (note that
0n -universal immediately guarantees a 0n -universal, namely its negation); define:
U(e, x) yV(e, x, y)
Corollary (0n -Universal Relations Exist).

For any n 1 and any k, there is a 0n -universal k + 1-ary U.
13.1. HIGHER COMPLEXITY RELATIONS 97
Theorem 13.2 (Hierarchy Theorem).

If a k-ary relation U is 0n -universal, then U is not 0n
Proof Sketch.
Suppose that a k + 1-ary 0n -universal U is 0n . Consider the k-ary relation R defined by:
R(e, n) U(e, e, n)
Since U is 0n , R is both 0n and k-ary.
Note that the hierarchy theorem guarantees something weve tactily assumed thus far; the arithemtic hierarchy
defined above never collapses at some level. Going higher always introduces new, previously unseen relations.
Hard
A relation A N is D-hard for some class of relations D if and only if for every relation B N in D, B m A
Corollary (Hard Relation Hard Complement).

A relation R |N is 0n -hard if and only if R is 0n -hard
13.1.2 Canonical Problems

Just as K and K are used as the standard bearers for the 01 and 01 relations generally, there are a host of canonical
problems which are widely known and utilized as the de facto problem in their complexity class:
K(e) e (e)
Tot(e) We = N
Und(e) We =
Inf(e) We is infinite
Fin(e) We is finite
Cof(e) We is cofinite
Onto(e) Ee = N
Rec(e) We is computable
Comp(e) We is Turing complete
Using our work with index sets, its easy to see that all of thesewith the exception of Kare non-computable. Classi-
fying these relations further, however, can become quite involved. Several of these results are presented below:
Proposition 13.1 (Tot is 02 -Complete).

Tot is a 02 -complete relation
Proof Sketch.
For hardness, let B(n) be an arbitrary 02 -relation. B is, by definition, defined by mnR(x, m, n) for some
primitive recursive R. To reduce B to Tot, consider the total computable f which gives
f (e) (m) = nR(e, m, n)

Proposition 13.2 (Fin is 02 -Complete).

Fin is a 02 -complete relation
Proof Sketch.
For hardness, let B(n) be an arbitrary 02 -relation. B is, by definition, defined by nmR(x, n, m) for some
primitive recursive R. To reduce B to Fin, consider the total computable f which gives
(
0 (m y)(n)R(e, m, n)
f (e) (y) =
otherwise
Proposition 13.3 (Cof is 03 -Complete).

Cof is a 03 -complete relation
Proof Sketch.
In general, then, we have the following picture of the arithmetic hierarchy when all of the canonical problems have
been assigned to the appropriate level of the hierarchy:
.. ..
. .
(m Comp) 04 04 (m Comp)
04
(m Co f , Rec) 03 03 (m Co f )
03
(m Fin) 02 02 (m T ot, In f )
02
(m K) 01 01 (m K)
01
Primitive Recursive
Limit Computation
13.1.3 Trees and Tree Arguments
Tree
A tree on a set X is a set T nN n X where nN n X is the set of sequences of values ha0 , . . . , an1 i where each
ai X such that if t T k(t k T )
13.2. RELATIVE COMPUTABILITY 99
We let [T ] = { f : N X|k f k T }
Lemma 13.1 (Konigs Infinity Lemma).

If T is an infinite, finitely branching tree, then [T ] ,
Finitely Branching
T is finitely branching iff s T , {a|sh ai T } is finite.
Proof? ? ? of the defin? above

T : define si T, dom(si ) = i such that {t T |si t} is infinite. (Some Stuff)
If T nN n N is computable (i.e. {psq|s T } is computable) and infinite finitely branching, then
S
i. f [T ] such that graph( f ) is 03
ii. There need not be a 02 branch.
n
{0, 1} Suppose T is computable, finitely branching. T has a 02 branch.).
S
Theorem 13.3 (Consider T nN
A
ll we really need is that we have a recursive function bounding the size of levels?
Proof.
Let f be the leftmost branch, i.e. f o = , f (n) =least i {0, 1} such that f n ` hii has infinitely
many extensions in T . Note that what follows the such that in the previous sentence is equivalent to
k > dom(t)s k {0, 1}[s T t S ]01 . Gives that f is 02 . f (k) = i iff s(seq(s) lh(s) = k + 1 j
lh(s) (somethingaboutthepreviousconditiononS j, (s)) (s)k = i).
Theorem 13.4.
There is a recursive tree T on {0, 1} such that T is infinite and T has no infinite recursive branch.
Proof.
Let A, B be c.e., A B = such that there is no recursive C with A C and C B = . T will be the
tree of all attempts to build C , for some C such that A C and C B = . Formally, let h : N A,
g : N B both surjective, recursive, total. Put s T iff s is a finite sequence of 0s and 1s and
n lh(s)[h(n) dom(s) s(h(n)) = 0 g(n) dom(s) s(g(n)) = 1]. T recursive and check that
[T ] = {C |A C C B = }.
13.2 Relative Computability

Post believed that many-to-one reducibility fails to accurately capture our intuitive notion of reducibility; he proposed
several different notions instead, the most important of which is based in Turings own notion of relative computability
(Soare 47; Cutland 167). At its core, Posts suggestion was to explain reducibility in terms of relative difficulty; the
general thrust of this analysis is that a problem X is reducible to a problem Y if and only if having a solution for Y also
gives a solution for X. This initial idea is, of course, in desperate need of further expansion, but the intuition behind it
should be a natural one.
To phrase this new vision slightly differentlyand draw on terminology we eventually adoptwe may imagine hav-
ing an oracle for a particular problem, call it Y. This oracle, like a rather peculiar character out of Greek mythology,
tells us instantly whether or not any item we present to it is a solution to Y; beyond this, however, our oracle is a
proverbial black box. We do not question how it operates, why it is so magnificently fast, etcetera; we simply know
that it is always correct and always gives its answer instantly. Oracle in hand, the problems that are reducible to Y are
simply classified as all of the problems we may solve with the use of our oracle for Y.
Returning to the formalism weve developed thus far, the analog to a problem is a k-ary relation on Nk ; an oracle
O for a k-ary relation R (the problem above) simply says yes or no for any k-tuple of N according to whether the
tuple is or is not in the relation. This characterization immediately gives rise to a number of possibilities for what, in
particular, we should take an oracle to be.
i. The relation R itself
ii. The characteristic function of R, R
iii. An oracle tape added to a Turing machine with R printed on it
iv. A total function O : N N
Its not hard to see that all of these are reducible to one another; in a by now stereotypical move, its not necessary
to consider k-ary relations at allsimply encode it into a 1-ary relation in the standard way; similarly, taking R is a
very natural choice, but restricting ourselves to a characteristic function still gives the same power as a total function,
so why bother? The oracle tape suggestion is out if only because of our function-based analysis up until now; the
choice, then, is between representing oracles as 1-ary relationsknown, more colloquially, as setsor total functions
from N to N. We take the latter tact, but build in aspects of the others:
Oracle
An oracle O is a any total function O : N N
In starting our study of computability, we selected a set of functions which were intuitively computable, and then
proceeded; now, we take those same functions and simply add one more: the oracle function, O. Just as we constructed
a notion of computable, this immediately gives rise to, for any oracle O, a notion of O-computability.
O-Computable
A partial k-ary function f : Nk 7 N is computable in an oracle O or O-computable if and only if there is a
derivation of f using the primitive recursive functions, O, and the operations of primitive recursion, composition,
and -recursion
Just as earlier, we must invoke the following with only suggestive evidence:
Thesis 13.1 (Relativized Church-Turing Thesis).

A function f is algorithmically computable relative to an oracle O if and only if f is O-computable
13.2.1 The Relativized Theorems

The major theme of this section is that the most important results established for computability generally carry over
wholesale to the notion of O-computability; in most cases the proofs of these theorems are almost verbatim those of
the previous resultnote, in particular, that all are independent of the particular oracle in use.
13.2. RELATIVE COMPUTABILITY 101
Theorem 13.5 (The Relativized Enumeration Theorem).

There exists an k + 1-ary, O-computable function O,k with the property that for any oracle O and any O-computable,
k-ary f , there is an e N for which
O,k (e, x) ' f ( x)
Theorem 13.6 (The Relativized sm n Theorem).

For every m, n 1 there exists an injective, computable, m + 1-ary function sm n such that for all oracles O, and any
e, a1 , . . . , am N,
O,m+n (e, a, x) ' O,n
sm
n (e,a)
( x)
Theorem 13.7 (Kleenes Relativized Fixed-Point Theorem).

For any oracle O, if f (x, y) is an O-computable function, then there is a 1-ary, computable function n such that for any
y and any k,
O,k
n(y) ' f (n(y),y)
O,k
Moreover, the function n does not depend on the oracle O used.
Notation. e (x) < s, sometimes written la Soare as {e}s (x) , indicates that e (x)
without recourse to any number greater than or equal to s; thus, s > 0, e, , x < s, and no
number used in the computation is greater than or equal to s
Theorem 13.8 (Use Principle).

For any oracle O and any e N,
(i) If Oe (x) = y, then (s)( O)[e (x) = y e (x)< s]

(ii) For any s N and sequence , if e (x) = y and e (x)< s, then (t s)( )[e (x) = y e (x)< t]
(iii) For any sequence , if e (x) = y, then (O )[Oe (x) = y]
Proof Sketch.
Although quite strong in the abstract, each of the above follows easily from how weve defined oracle
computation
13.2.2 The Relative Arithmetical Hierarchy

Mimicking our development of computability, we may, for any oracle O, identify a relativized arithmetical hierarchy.
Paralleling our earlier work, we work in the language of first-order arithmetic expanded with our oracle relation O,
notated LA (O). The hierarchy of formulas is again inductively defined by:
A formula of LA (O) is 00 (O)/00 (O) is logically equivalent to a formula using only the boolean
connectives and bounded quantification
A formula of LA (O) is 0n+1 (O) is logically equivalent to x where is 0n (O) and x is finite
A formula of LA (O) is 0n+1 (O) is logically equivalent to x where is 0n (O) and x is finite
As before, we associate the relation defined by a given formula in the standard model N with that formulas position
in the hierarchy, defining
A relation R is 0n (O) R is defined by both a 0n (O) and a 0n (O) formula
This is, of course, equivalent to: a relation R is 0n (O) if and only if both R and R are 0n (O)-relations. We thus have:
.. .. .. ..
. . . .
03 (O) 03 (O) 03 (O) 03 (O)
03 (O)
02 (O) 02 (O) 02 (O) 02 (O)
02 (O)
01 (O) 01 (O) 01 (O) 01 (O)
00 (O)/00 (O) 01 (O)

The closure rules associated with the arithmetical hierarchy carry over verbatim to the relativized arithmetical
hierarchy:
i. The class of 01 (O) predicates is closed under , , substitution of total O-computable functions, substitution of
O-computable functions past the n = 1 level, and existential generalization
ii. The class of 01 (O) predicates is closed under , , substitution of total O-computable functions, substitution of
O-computable functions past the n = 1 level, and universal generalization
iii. The class of 01 (O) predicates are closed under , , , substitution of total O-computable functions, and substi-
tution of O-computable functions past the n = 1 level.
13.3 Turing Reducibility

Having refined the notion of relative computability, its time to return to the impetus for its development.
Turing Reducible
Given two unary relations A, B N, A is Turing reducible to B, notated A T B, if and only if A is
B -computable.
That is, A T B just in case A is decidable given an oracle for B; the following are easy to establish:
13.3. TURING REDUCIBILITY 103
Proposition 13.4 (Characteristics of T ).

The following all hold for Turing reducibility:
(i) T is a preorder
(ii) m T
(iii) For any relation A, A T A
(iv) For any computable relation B, B T A for any relation A
(v) O-Computability is preserved downward under T

(vi) The / distinction is not preserved downward under T
Turing Degree
The equivalence classes generated by the relation of mutual Turing reducibility; that is, for any relation A N,
[A]T = {B N : B T A and A T B}
Two relations with the same Turing degree are, for obvious reasons, said to be Turing equivalent, often notated by
T .
Conventionally, Turing degrees are written in boldface, e.g. a, b, and c, while the class of all Turing degrees is
denoted D. The following should be reminiscent of the terminology used with many-one reducibility: for a class of
relations ,
a is a -degree there is a relation of Turing degree a in

a is -hard every relation in has a Turing degree T a
a is -complete a is both a -degree and -hard
Corollary (Characteristics of Turing Degrees).

The following all hold for Turing degrees:
(i) There is a unique computable degree, denoted 0
(ii) For a relation A N, [A]m [A]T

(iii) [K]T is 01 -complete
(iv) The Turing degrees D form a partially ordered set
(a<b just in case a relation in a T a relation in b)
13.3.1 The Turing Jump Operator

Denoted by the addition of an apostrophe, the Turing jump or simply jump A0 (read A-jump or A-prime) of a relation
A is Kleenes K relation relative to an oracle for A, often abbreviated K A . In other words,
A0 = K A = {x : A,1 (x, x)} = {x : x W xA }

As above, there is a unique Turing degree for the computable relations, 0, and so the unrelativized version of K is 0
This notation can clearly be extended and often is:
A(n+1) = (A(n) )0
Proposition 13.5 (K A T K0A T K1A ).

For any A N, K A is Turing equivalent to each of the following:
(i) K0A = {hx, yi : A,1 (x, y)}
(ii) K1A = {x : W xA,1 , }
Proof Sketch.
Use the relativized sm
n theorem and Myhills isomorphism theorem
Theorem 13.9 (Characteristics of the Turing Jump).

For any relations A, B, and C on N,
(i) A0 is 01 (A)
(ii) A0 T A
(iii) B is 01 (A) if and only if B m A0
(iv) If A is 01 (B) and B T C, then A is 01 (C)
(v) B T A if and only if B0 m A0
(vi) If B T A, then B0 m A0
(vii) A is 01 (B) if and only if A is 01 ( B)
Proof Sketch.
pg. 53 Soare
Corollary (Post Hierarchy Theorem).

For every n 1,
i. (n) is a 0n -complete set.
ii. A relation A N is 0n+1 if and only if A T (n)
Proof Sketch.
13.3. TURING REDUCIBILITY 105
Using the Turing jump operator, we define a parallel operation on Turing degrees:
a0 = degT (A0 ) for any A a
Using our results above, the jump is well-defined on degrees, andby definition0 is the minimal degree. We thus have
an infinite hierarchy of degrees:
0 < 00 < 000 < 0000 < < 0(n) < . . .
In fact, this hierarchy is closely related to another, as the first few levels make clear:
0 = deg() = {A : A is computable}
00 = deg(0 ) = degT (K) = degT (K)
000 = deg(00 ) = degT (Fin) = degT (Tot)
0000 = deg(000 ) = degT (Cof) = degT (Rec)
00000 = deg(0000 ) = degT (Comp) = degT (Comp)
0(n)
..
.
000
All T -Degrees 00 Degrees T 00
0 -Jump
A(0 ) = {hn, ki|k A(n) }
This is the computable join of all A(n) for n N.
Fact. (0 ) m ThN
Chapter 14
Posts Problem & Oracle Constructions
Much of our work thus far has been foundational, establishing the basic results and basic notions of computability
theory. In this chapter, we leverage the results and notions of the last few chapters in order to prove deep results about
the structure of both the Turing and m-degrees.
14.0.2 Simple Relations
Immune
B N is immune if and only if both B is infinite and B has no infinite 01 -subset.
Fact. There are no 01 immune sets.
Proposition 14.1 (01 , Immune Sets Exist).

There is a 01 set that is immune
Proof Sketch.
Priority?
Simple
A set A N is simple if and only if A is 01 and A is immune.
Theorem 14.1 (Simple Sets T ).

For any simple set A N, A T
Theorem 14.2 (Simple Sets T K).

For any simple set A N, A T K
By the theorem above, we have a simple set A.
107
108 CHAPTER 14. POSTS PROBLEM & ORACLE CONSTRUCTIONS
Theorem 14.3 (Post).

There is a simple set.
Proof.
Let f be total recursive, 1-1, and ran( f ) is not recursive. Put D(n) k > n( f (k) < f (n)) (deficiency
stages in the enumeration of ran( f )). Claim: D is simple Proof: D is 01 ; D is infinite: if D has one element,
it has infinitely many since there must always be a next smallest and, by definition, this is also in D. Note
further that there must always be at least one element in D since there must be a smallest element (which,
by definition, is in D).
D is Immune: if not, let B D be infinite r.e. Then, m < ran( f ) iff n(n B m < f (n) k < n( f (k) ,
m)). Note that the right side is 01 .
Recall that fk is the finite function coded by k as a table. Fk = ran( fk ).

Hyper-Immune
A set A is hyper-immune iff
i. A is infinite
ii. Whenever g is total recursive N N, and Fg(m) Fg(n) = for all m , n, then n Fg(n) A = .
Hyper-Simple
A set A is hyper-simple if and only if A is 01 and A is hyper-immune.
14.1 Incomparable Turing Degrees
Incomparable
Two degrees a, b are incomparable, notated a|b, if and only if a b and b a
Proposition 14.2.
If a function f is total, extends every partial function in some A-computable sequence, and is completely determined
by the sequence, then f is A-computable.
Proof Sketch.
To determine any value, simply search the sequence until a partial function which defines it is found (such
a partial function must exist).
Theorem 14.4 (Incomparable Turing Degrees below 00 ).

There are Turing degrees a,b T 00 such that a and b are incomparable [Kleene-Post].
14.1. INCOMPARABLE TURING DEGREES 109
Proof Sketch.
Our proof method is the first in a set of standard proof techniques in computability theory: a finite exten-
sion, oracle construction. Put simply, we build step-by-step the characteristic functions for two relations
A and B, intending to let a,b be the Turing degrees of A, B respectively. Crucially, however, we build A
and B in both a 00 -computable manner (guaranteeing a,b T 00 ) and so that
(i) No B-computable function computes A
(ii) No A-computable function computes B

which, of course, together guarantee incomparability. To guarantee (i) and (ii), we identify a pair of
constraints for every e N:
(S e ) A ; eB
(Re ) B ; eA
We need simply show that A, B can, in fact, be constructed in a 00 manner to meet both Re and S e for every
e N. To start, set A = B = . At any odd stage 2e + 1, requirement S e is satisfied. Taking 2e+1
A , 2e+1
B
to be our initial approximations of A , B (encoded as sequences), we query our -oracle for the truth or
0
falsity of the following 01 , LA sentence:
t[BinS eq() 2e+1

B e (lh(2e+1
A ))< t]
where BinS eq is true if and only if the input is a number encoding a sequence of 0s and 1s. Put less
formally, we ask whether there is an acceptable extension of 2e+1
B such that
Case 1: If the sentence is true, there must be a least pair h0 , t0 i witnessing this; set 2e+1+1
B = 0 and
define 2e+1+1
A as 2e+1
A , but with the addition of
t0 (lh(As )) > 0
( 0
A (lh(A )) = 01 ifotherwise
2e+1+1 2e+1
Case 2: If the sentence is false, define 2e+1+1

A as 2e+1
A and 2e+1+1
B as 2e+1
B , but with the addition of
2e+1+1
A (lh(2e+1
A )) = 0
2e+1+1
B (lh(2e+1
B )) = 0
Reversing the treatment of A and B in the even case gives a complete procedure of the type required.
Theorem 14.5 (Incomparable Turing Degrees).

For any degree c, there are incomparable Turing degrees a,b such that c T a, b and a,b T c0 .
Proof Sketch.
Generalize the proof above
Theorem 14.6 (An Incomparable Degree Always Exists).

For any degree b >T 0, there exists an incomparable degree a <T b0 for b.
Proof Sketch.
Another finite extension, oracle construction; the two requirements necessary correspond to avoiding the
lower cone of degrees with reduce to b and the upper cone of degrees which reduce b.
Remark. If A T B, then A0 m B0 . So we can define [A]0T = [A0 ]T
c T d c0 T d0
So, 00 T d0 for all d.
14.2 Friedburg Completeness

Since 0 is the unique smallest Turing degree, 00 is the unique smallest jump-degree. The next theorem establishes that
not only do jumps take us to progressively higher Turing degrees, but they also exhaust all possible Turing degrees
above 00 :
Theorem 14.7 (Friedburg Completeness Theorem).

For every Turing degree b such that 00 T b, there is a Turing degree a such that a0 =b
Proof Sketch.
Finite extension, oracle construction; we need to construct an A meeting two general constraints: A0 T B
and B T A0 . In a now standard move, well ensure the first on odd steps and the second on even. As well
see, the latter of these is quite easy, while the former requires more thought. Turning to this problem first,
we wish to provide a means for B to decide A0 ; since B is arbitrary, this information must somehow be
built into Awhich, of course, is only useful if B can compute A. Our strategy, then, is to make our entire
construction of A B-computable, then to somehow encode whether or not e A0 into the construction.
Moving to the details of this procedure, suppose we are at an odd step 2e + 1 with an initial approximation
2e+1
A of A . We need a B-decidable question which will somehow indicate whether or not e A0 which
since all we know of B is 0 T Bthis means all we may ask is a strictly existential question. As usual,
we invoke the use properties of relative computation, querying:
( 2e+1
A )(t)[e (e)< t]
If true, we set 2e+2

A = ; if not, we set 2e+2
A = 2e+1
A . Thus, for B to decide whether or not e A0 , it need
simply run the construction of A until the 2e + 1-th step, then check the formula above for truth or falsity
(note that, if false, e < A0 ).
After all the excitement above, building A (B-computably) so that B T A0 is as easy as, for any even step
2e + 2, always extending 2e+2
A by B (e); that is, set
A if x < lh(2e+2
( 2e+2
)
2e+3
A (x) =
B(e)
A
if x = lh(2e+2 )
A
Theorem 14.8 (Relative Friedburg Completeness).

For every Turing degree c and any Turing degree b T c0 , there is a Turing degree a such that c T a and a0 = b
Proof Sketch.
Mimic the proof of the original, replacing 0 with c
14.2. FRIEDBURG COMPLETENESS 111
Corollary (Generalized Friedberg Completeness).

For every n 1 and every degree c, any Turing degree b T c(n) is such that there is a Turing degree a where a(n) = b.
Proof Sketch.
Induction on m
14.2.1 Minimal Degrees
Minimal
A Turing degree a is minimal if and only if 0 <T a and there is no Turing degree b such that
0 <T b <T a
Theorem 14.9 (Spector 1957?).

There is a noncomputable A such that for every B, B <T A(B T or B T A)
Proof.
Construct A by approximating A with computable perfect ? or {0, 1} (i.e. T is a binary tree).
s T ks k T
T is perfect : s T u, v T [s u s v u v]
where r s iff n dom(r) dom(s)[r(n) , s(n)].
T is computable
Notation:
For T a tree, s T ,
T s = {u T |u 6 s}

Requirements on A: Re : Either e A is not total or e A is computable or e A T A.
Set T 0 = nN n {0, 1}. Given T e : want to ensure Re .
S
Case 1: s T e such that u, v s with u, v T k such that ue (k) and ve (k) , then ue (k) = ve (k). in

that case, set T e+1 = (T e ) s for some such s. Re is now guaranteed: suppose A [T e+1 ] and e A
is total.

Then, e (k) = i iff u
A
T e+1 ue (k)= i (note that this is 01 )
Case 2: Otherwise. Well define a T as follows: For r 2< , define (r) = T e
computable, perfect T e+1
by induction of the length of r. Then, take {u|r[u (r)]}. Set () = . Given (r), we need to define
(r ` h0i)
(r ` h1i)
Search for u, v, k showing Case 1 fails for s = (r). For the first ones you find, set u = (r ` h0i),
v = (r ` h1i). Thus completes the definition of . Note that is computable. Thus, set
u T e+1 r(u (r))
r(r 2dom(u)+1 u (r))

Claim: If A [T e+1 ] and were in case 2, then A T e A . To recover A from e A . Suppose we know
that (r) A . Q is (r ` h0i) A or (r ` h1i) A . These two differ on e , and so search for k such
that for u = (r ` h0i), v = (r ` h1i), ue (k), ve (k), and ue (k) , ve (k). Then, (r ` h0i) A if and only

if (r
`
h0i)
e (k) = e A (k) so e A can tell us the answer to our question.
To finish, there is an A s such that A [T e ] for all e (if T i is a subtree of (2< , ) such that for every finite
F such that
iF [T i ] ,
then,
[T i ] ,

14.2.2 Constructing c.e. sets

Constructing c.e. with a prescribed T .
Theorem 14.10.
There is a simple set A such that A0 T 00
00 = A0 A 0
14.2. FRIEDBURG COMPLETENESS 113
Proof.
Well define finite sets A s , s N, A s A s+1 , S `0 A0 computable, A = s A s .
S

Requirements: Pe : ensure We A , (if We infinite) Ne : if you see e As (e), then preserve As k (i.e.
k
restrain all i < k such that if i < A s , then i not in A).

To resolve conflicts: if Pe and Nl want to act in conflict at some stage s, give priority to lower index,
breaking equality in favor of P. If an N is violated, throw it out since it doesnt do anything anymore. Ns
can be issued repeatedly for a given e, each Pe acts at most once (achieving its goal). Each Ne acts finitely
often ( l + 1).
Now, We inf We A , . Take every i We such that no Nl , l < e ever restrains i. i A unless some

other i0 We A T . To see A0 T 0 , to decide whether l A0 , e A (l) using 0 as oracle: go to
a stage s such that if e l is such that Pe ever acts, it has acted already (this needs to ask did Pe ever
A
act?a 01 question). Then, l (l) iff Nl has a restraint in force at some stage greater than s (again, a 01
question)
Get A, B r.e. such that A T B and B T A

The above is called a finite injury priority argument.
Part III
Computability and First-Order Logic
115
Chapter 15
Languages, Theories, and Computability
Historically, logical interest has centered around a handful of first-order theories; in this chapter we present the basic
terminology for working with both first-order logic and theories from a computability perspective, then canvas those
theories which have proven most interesting, proving some basic results about each.
15.1 Languages, Theories, and Axioms

First, we bridge the gap between first-order languagesfinite strings of symbols over varying signatureson one hand
and our theory of computationfunctions from N to Non the other; to do so, we simply follow Gdel and generate an
encoding scheme for the symbols which occur in any language of interest. Given all of our previous work, doing so
is as simple as assigning each distinct symbol its own number via a computable function #; then, strings like xR(x)
can easily be encoded as a sequence of these numbers:
h#(), #(x), #(R), #((), #(x), #())i
This sequence is, of course, itself representable as a single number in the familiar way. To save on space, we wrap
this entire process into the placement of corner quotes around a well-formed formula of the language in question, as
below:
pxR(x)q
Thus, xR(x) is a well-formed formula from some language of interest and pxR(x)q is the natural number which en-
codes it according to some background encoding scheme. The specific encoding scheme used will be left unspecified;
we merely suppose that it is a reasonable one which has the basic properties of traditionally popular schemes. The most
important of these properties is that the language under consideration is computable (also, recursively-presented); that
is, there exists a computable function which, given a natural number, decides whether or not that number encodes a
well-formed formula of the language. As far as naturally occurring languages and theories, this is remarkably weak,
but shouldnonethelessbe kept in mind.
Deductively-Closed Theory
A deductively-closed theory in a language L is a set T of L-sentences such that for all sentences of L, if T |= ,
then T .
Axioms
A set of L-sentences is a set of axioms for a deductively-closed L-theory T if and only if T is the deductive
closure of under some specific consequence relation
Computably Axiomatizable
A deductively-closed theory T in a computable language L is axiomatizable if and only if there exists a
computable set of axioms for T ; that is, there is a computable function which decides whether or not any natural
number encodes an axiom
117
118 CHAPTER 15. LANGUAGES, THEORIES, AND COMPUTABILITY
Historically, one of the largest areas of interaction between computability theory and logic is in proving or dis-
porving the decidability of some particular, usually mathematical, theory.
15.2 Decidability and Undecidability

A theory is decidable if and only if, using our encoding of formulas into natural numbers, the relation defined by
the theory is decidable; that is, determining whether or not any natural number encodes a formula which is in the
theory is a computable process. A word of warning is necessary here; many authors equivocate between theories and
deductively-closed theories or between the decidability of a theory and the decidability of its deductive closure; the
terminology herein is unambiguousother authors may not be so kind.
Theorem 15.1 (Craigs Trick).

For a computable language L, if a theory T is computably enumerable and deductively-closed, then T is computably
axiomatizable.
Proof Sketch.
By assumption, we have a theory T = {n : n N} such that, for some computable f , f (n) = pn q for
every N. Using this f , define, for each n/n , a new formula n as ni=1 n (n conjoined with itself n
times). Define:
= {n |n N}
is decidable since we may simply count the number of conjuncts in any well-formed formula, checking
they are all the same, then use the f above to ensure that the conjunct is the correct one. Note finally that
T = { : |= }
Theorem 15.2 (Complete and Computably Axiomatizable Decidable).

If a consistent theory T is both complete and computably axiomatizable, then T is decidable.
Proof Sketch.
Assume T is consistent since, otherwise, it is trivially decidable. To decide whether T , search for
both a proof of from the computable axioms for and a proof of from the computable axioms for
simultaneously, return whichever terminates first. Note that determining whether or not some encoding
is a valid proof is computable so long as the process for determining whether or not a given premise is
admissible is itself computable.
Of course, not all theories of interest are so well-behaved as to be both complete and computably axiomatizable;
indeed, much work has gone into the question of determining the decidability or undecidability of a theory under more
difficult circumstances; in this vein, we distinguish two notions stronger than simple undecidability:
Essentially Undecidable
A deductively-closed theory T in a computable L is essentially undecidable if and only if T has no consistent,
decidable extension in L.
Strongly Undecidable
A theory T in a computable L is strongly undecidable if and only if, for every deductively-closed L-theory S
such that S T is consistent, S is undecidable. A model A is called strongly undecidable if and only if Th(A) is
strongly undecidable
15.2. DECIDABILITY AND UNDECIDABILITY 119
Essential undecidability captures the idea that there is something about the sentences of the theory which forces
undecidability; there is no way to repair the undecidability of the theory. Strongly undecidable theories, in contrast,
guarantee that no decidable theory is even so much as consistent with thema distinctly stronger claim. Our picture,
then, is as follows:
Undecidable < Essentially Undecidable < Strongly Undecidable
From just our previous work, we thus have:
Corollary (Equivalences for Complete Theories).

For a complete theory T , the following are equivalent:
(i) T is undecidable
(ii) T is essentially undecidable

(iii) T is not computably axiomatizable
This means, for example, that if a theory is known to be both computably axiomatizable and undecidable, then it must
incomplete. The next proposition helps to demonstrate exactly how powerful a property strong undecidability really
is:
Proposition 15.1 (Strongly Undecidable Theory Strongly Undecidable Model).

If T is any deductively-closed theory and A is some model such that A T , then the following are equivalent:
(i) A is strongly undecidable
(ii) T is strongly undecidable
Proof Sketch.
Simply leverage definitions
Theorem 15.3 (Essentially Undecidable and Finitely Axiomatizable Strongly Undecidable).

If a theory T in a computable L is essentially undecidable and finitely axiomatizable, then T is strongly undecidable
as well.
Proof Sketch.
Suppose S is a non-empty, deductively-closed, decidable theory consistent with an essentially undecidable
T and is the finite axiomatization of T . A decidable, consistent extension of T can be constructed as
follows: Proceed through all formulas , if is in S , is in our extension; if not, it is not.
Theorem 15.4 (Incomplete, Decidable Theories have Consistent, Decidable, and Complete Extensions).
Every incomplete, deductively-closed, and decidable theory T has a decidable, consistent, and complete extension
[Tarski, 1949].
Proof Sketch.
By deductively-closed and incomplete, we have that the theory is consistent. We construct the consistent,
complete, and decidable extension as follows: start with our theory and begin moving through the well-
formed sentences of the language, starting with those having least Gdel number; for any given sentence,
add it to our theory if its negation is not a consequence of the theoryotherwise, simply move to the next
sentence. Leverage this process and the deductive-closure of the original theory to prove decidability.
Corollary (Equivalences for Computably Axiomatizable, Deductively-Closed Theories).

For a computably axiomatizable, deductively-closed theory T , the following are equivalent:
(i) T is essentially undecidable
(ii) T is consistent and every consistent, computably axiomatizable, deductively-closed extension of T is incomplete
(iii) T is consistent and no consistent, complete, deductively-closed extension of T is computably axiomatizable
15.3 Proving Decidability

A variety of equivalences and implications in hand, there are three main methods for proving that a particular,
deductively-closed theory is, in fact, decidable:
(i) Quantifier Elimination
(ii) Model-Theoretic
(iii) Interpretations
The first two, quantifier elimination and model-theoretic, seek to leverage both the results above and in the model-
theory portion of this text. In particular, weve shown that a deductively-closed theory which is complete, consistent,
and computably axiomatizable must also be decidable. Since theories of interest are generally both consistent and
computably axiomatizable, both of these methods are often used to demonstrate the last property: completeness.
15.3.1 Quantifier Elimination

Quantifier elimination is, as its name suggests, a method which attempts to eliminate the use of quantifiers in a
language, to make them redundant and unnecessary. In general, this is, of course, impossible; quantifiers add ex-
pressiveness andin the most general settingits impossible to eliminate them without the loss of this expressivity.
Luckily, we often work in a setting far more limited than the most general; in particular, relative to some background
theory T , it may very well be possible to eliminate the use of quantifiers. This is the most natural picture of quantifier
elimination, and theories which behave in this manner are said to have quantifier elimination.
Formally, the method of quantifier elimination is slightly broader than detailed above; rather than the complete
elimination of quantifiers, the method of quantifier elimination allows for a set of formulas, , within which quanti-
fiers are allowed, then reduces all formulas of the language to some boolean combination of these. That is, given a
deductively-closed theory T in a language L, wein a moment of divine inspirationidentify some set of formulas, ,
from L with a peculiar property:
(?) For any formula ( x) of L, there is some boolean combination of formulas from , ( x), such that T |=
x[( x) ( x)]
In other words, the set called an elimination set for T can supply a boolean combination of its formulas which is
T -equivalent to any formula of L. In the particular case of sentences, then, all quantifiers except those occurring in
the basic formulas of are eliminatedwe simply consider the T -equivalent formula from instead!
Of course, the divine inspiration needed to pick can be hard to come by; in general, there are four criterion
good choices tend to satisfy:
15.3. PROVING DECIDABILITY 121
(1) is reasonably small and non-redundant

(2) Formulas in have straightforward meanings
(3) Constructing a boolean combination from T -equivalent to a given sentence is a computable process
(4) Determining whether or not a boolean combination from is a consequence of T is computable.
Given (3) and (4), its easy to see that quantifier elimination shows the decidability of T ; (1) and (2), in contrast,
are empirical observations about sets for which logicians have established (3) and (4). Crucially, even without (3)
and(4), all hope isnt lost for proving decidability; in particular, if it can be shown that T is complete with respect to
all the sentences in (for any sentence, either it or its negation are a consequence of T ), then T must be completea
property that, under many circumstances, readily gives decidability.
To achieve the full flavor of this method, we walk through a proof demonstrating the decidability of Presburger
arithmetic:
Fact. To establish (?), its only necessary to consider an arbitrary formula of the form x[( x)] where is
a set containing only literals
Other natural theories whose decidability can be demonstrated using only the method of quantifier elimination are:
The theory of the standard model of N under the signature containing only a successor function S and constant
0 [Herbrand, 1928]
The theory of the standard model of N under the signature containing only a successor function S , multiplication
function , and constant 0 [Skolem, 1930]
Elementary Theory of Identity; the first-order validities of a language with empty signature [Lowenheim, 1915].
Theory of Finitely Many Sets; the first-order validites of a language with only m unary relations [Lowenheim,
1915]
Theory of Discrete Orders; axiomatized by,
xy[x y y x x = y]
xy[x y y x]
xyz[x y y z x z]
xy[y x x = y]
xy[x < y z(x < z y z)]
xy[y < x zw(z < x (z < w z w))]
[Langford, 1927]
Theory of linear order in Q; i.e. Th(hQ, i)
Theory of Boolean Algebras [Tarski, 1949]
Theory of R; i.e. Th(hR, 0, 1, +, , , i) [Tarski, 1948]
15.3.2 Model Theory

Model theory, as weve seen, has a host of results about theories; thus, one way to show the decidability of a theory
is to leverage these results along with those just established. Most commonly, this is done by showing a theory to be
complete (usually via categoricity and the os-Tarski theorem) or by systematically studying all possible extensions
of a theory (and leveraging our earlier results in this direction).
Its worth noting that, while presented as separate methods, model theory and quantifier elimination often work
hand-in-hand; the process of constructing elimination sets, in particular, benefits greatly from model-theoretic insights
and results.
15.3.3 Interpretation
The method of interpretation is, at heart, the result of a simple, familiar idea: reduction. Suppose that, on one hand,
we have a theory T ? in a computable language L? for which we wish to prove decidability; meanwhile, on the other
hand, we have a provably decidable theory T D formalized in a computable language LD . If there exists a computable
function f : L? LD such that, for every L? ,
T ? f () T D
then T ? must be decidable. Our decision procedure simply translates into LD using f , then leverages the decision
procedure for T D . Of course, in line with our earlier work with reductions, reducibility is a two-way street. In
particular, given two theories T 0 , T 1 in two language L0 , L1 and a computable f : L0 L1 such that
T 0 f () T 1
If T 0 is undecidable, then T 1 is undecidable as wellthe contrapositive of our previous result.

From this general strategy a much more particular approach has developed; rather than work with theories, the
focus is turned, instead, towards models. This gives the computable f needed for the reduction a great deal of structure
and dovetails nicely with already existent model-theoretic practices (indeed, this is where the term interpretation
derives from). On the downside, however, this limits the applicability of interpretations to complete theories (which
determine a model) or properties that interact nicely with models, e.g. strong undecidability and certain combinations
of decidability/essential undecidability.
Formally, the method of interpretation focuses on defining one model within another; that is, given some model A,
we construct an isomorphic, first-order definable copy of another model B within A. This construction is accomplished
in a number of steps, starting with the interpretation function:
Interpretation Function
Given two computable languages L0 and L1 , an interpretation function is a function I : L0 L1 that is, from the
signature of L0 to the signature of L1 such that,
i. I(c) for any constant symbol c is a well-formed formula c (v0 ) of one free variable
ii. I(R) for some n-ary relation symbol R is a well-formed formula R (v0 , . . . , vn1 )
iii. I( f ) for any n-ary function symbol f is a well-formed formula f (v0 , . . . , vn )
In addition, with every interpretation function I we associate a well-formed formula D (v0 ) of L1 with a single free-
variable; in our scheme D (v0 ) will determine the domain of the model we wish to interpret within A while I provides
a simple translation of formulas from L0 to L1 . Thus, given an L1 -structure A, an interpretation function I : L0 L1 ,
and an L1 formula D (v0 ), we may construct an model-like entity AI as follows:
{
Let |AI | be AD ; that is, all elements a of A such that A D [a]
I
For every constant symbol c of L0 , let cA be the unique b AD such that A c [b]
AI I
For every relation symbol R of L0 , let RA be RA (AD )n
I
For every function symbol f of L0 , let f A be defined by
I
f A (a) = b A f (a, b)
In order for AI to be a model, it must be the case that A I where I is defined as follows:
{
v0 [D (v0 )]
I !v0 [c (v0 ) D (v0 )] for every constant symbol c of L0
v0 . . . vn1 [n1
i=0 D (vi ) !vn [ f (v0 , . . . , vn ) D (vn )] for every function symbol f of L0
15.4. PROVING UNDECIDABILITY 123
The first condition on I guarantees that the domain of AI is nonempty, while the second and third guarantee that the
constant and function definitions of AI are respectively acceptable.
Interpretable
Let B be an L0 structure and A a L1 structure. B is interpretable in A if and only if B AI for some interpretation
function I such that A I ; B is computably interpretable in A if and only if B AI for some computable
interpretation function I such that A I
At long last, we turn to how interpretability can be leveraged. If Th(A) or Th(B) is of interest, any interpretation
function I can be used to generate a full, computable map f : L0 L1 reducing Th(B) to Th(A), just as in the
introduction to this section.
Proposition 15.2 (Computably Interpretable, Strongly Undecidable Model Strongly Undecidable).

If A is computably interpretable in B, and A is strongly decidable, then B is strongly undecidable as well
Proposition 15.3 (Computably Interpretable, Essentially Undecidable Essentially Undecidable).

If A is computably interpretable in B, and Th(A) is essentially decidable, then Th(B) is essentially undecidable as well
Proposition 15.4 (Computably Interpretable in a Decidable Model Decidable).

If A is computably interpretable in a decidable model B, then Th(A) decidable
15.3.4 The Decidability of Validity

(i) Let L be recursively presented. Let E be the set of all L-sentences of the form: v where is quantified or
free. Then, S AT over E (that is, { E|S at()}) is computable since if E if and only if has a finite model
(where this latter is a 01 property).
(ii) Let = {| E} or rather are of the form v where is q.f., then S AT is 01 -complete for L
having infinitely many function symbols of each arity and a binary relation symbol. S AT LS T m S AT given
a sentence of LS T . Find such that
|= v1 v2 v3 . . .
where is q.f. This is called prenex form and is computable. We now successivbley replcae quatnifiers with
skolem functions.
1 = v1 v3 v4 . . . (v1 , f2 (v1 ), v3 , . . . )
= v1 v3 v5 . . . (v1 , f2 (v1 ), v3 , f4 (v1 , v3 ), . . . )
2
Then, Bi S AT i+1 S AT
(iii) Suppose that L has only unary predicate symbols and no function symbols. Then S AT L is computable (this is
again a result of the following property has a model iff has a finite model)
(iv) If L has 2 unary function symbols S AT L is 01 complete; if theres just one and nothing else, S atL is computable.
Fact. (can use in exercises) Can use (iv) to get an essentially undecidable, not strongly undecidable, axiomati-
zable T again using inseparable r.e. sets.
15.4 Proving Undecidability

Chapter 16
Arithmetic & Incompleteness
16.1 True Arithmetic

True arithmetic is that understanding of arithmetic and the natural numbers which is taught to us from elementary
through high school; the operations of multiplication, addition, etc. behave as usual while the only elements are those
in N, under the order-type . This is our inuitive picture of the natural numbers. We formalize this picture by selecting
a first-order language in which we will work:
The Language of First-Order Arithmetic
The language of first-order arithmetic, abbreviated LA , is the first-order language generated from the following
signature:
(i) Binary Function Symbols: +,
(ii) Constants: 0, 1
(iii) Binary Relation Symbol: <
This signature is often referred to, for historical reasons, as thesignature of Peano arithmetic.
Fact. Adding exponenetiation to the signature of LA doesnt add expressivity; in other words, exponentiation
is definable in LA
and then designating our intended model of LA as N, the standard model. To complete the picture and return to the
topic of this section, we identify the term true arithmetic with Th(N) in the language LA . Recalling that much of our
work with relations in computability theory is using true arithmetic (and, by extension, LA ), it should now be quite
obvious where the term arithmetic derives from in the arithmetic hierarchy:
Theorem 16.1 (Definable over N Arithmetic).

i. A relation R is definable over N in LA
ii. A relation R is arithmetic; that is R appears in the arithmetic hierarchy at some level n
Indeed, as we saw earlier, the correspondence is even more exact; relations definable in N by a bounded formula
are primitive recursive (although the converse is not true), the 01 -relations are exactly those definable in N by an
existential formula, so on and so forth.
125
126 CHAPTER 16. ARITHMETIC & INCOMPLETENESS
Corollary.
Let
T hN
0
= {pq| is an existential sentence of LENT and N }
1
Then, K m T hN
01
Theorem 16.2.
T hN
0
is 01 -complete
1
Proof.
First, show its 01 by showing:
S at(pq, s) N [(s)0 , . . . , (s)lh(s)1 ] is bounded
is primitive recursive (primitive recursion on pq). Then,
pq T hN
0
= v for some bounded sS at(pq, s)
1
Next, show K m T hN
0
. Let
1
K(n) N [n]
( existential), then
K(n) p(n)q T hN
0 1
Then,
f (n) = p(n)q
is primitive recursive.
Let N + = (N, +, , exp, S , 0, , T k for every k N) where T k is intended to represent the Kleene T -predicate. It
+
follows that T hN
0
is, trivially, 01 -complete.
1
Corollary.
N is strongly undecidable
Validities of a Language
Let L be a first-order language; the validities of L, denoted ValL , is the theory
ValL = { : is a sentence of L and ` }
Corollary (ValLA is 01 -Complete).

The validities of LA , ValLA , form a 01 -complete theory.
Proof.
PA
ThmQ m ValL : T hmQ i f f 7 Q which requires that Q is finitely axiomatized!
16.2. FIRST-ORDER PEANO ARITHMETIC 127
Weve already shown that any such theory is 01 -complete and thus undecidable.
Theorem 16.3.
Suppose N is computably definable in some model D. Let T be axiomatizable and true in D. Then, T is 01 -complete
and D is strongly undecidable
Proof.
Let I: symbols of LPA into wffs in the language L of D be an interpretation function. Let I be the
sentences of L saying this about I (i.e. D |= I iff DI is an LPA structure). Let S iff T I ` I . I is
defined by induction on in such a way that for any D |= I , D |= I iff DI |= . E.g.,
(v)I = v(uI (v) I
( f (v1 , . . . , vn ) = vn+1 )I = If (v1 , . . . vn+1 )

S is a theory: Suppose S ` . Let D |= T I . DI |= S (D |= I for every S .) So, DI |= , so D |= I ,
and thus T I ` I . Moreover, N |= S (DI |= S ). So S is 01 -complete (S is axiomatizable since T
was). S m T since S iff T ` I I . The function f mapping to I I is recursive.
16.2 First-Order Peano Arithmetic

First-order Peano arithmetic, often shortened to simply Peano arithmetic, is, in conception, a dodge; an attempt to
circumvent the difficult philosophical and set-theoretic questions inherent in the original, second-order Dedekind-
Peano induction axiom
P[0 P x[(x P x + 1 P) y[y P]]]
by replacing it with an infinite number of weaker, first-order induction axioms. With the second-order axiom, the
Dedekind-Peano axioms succeed in characterizing N up to isomorphism; as we saw above, the first-order attempt was
doomed to failure from the start. Nonetheless, Peano arithmetic has become a first-order theory of significant import,
characterizing a great deal of the most important aspects of N. Intuitively, Peano arithmetic is true arithmeticif you
can only see the natural numbers; that is, on the natural numbers N, Peano arithmetic is perfectit simply fails to rule
out the existence of elements not in N.
Peano Arithmetic
Peano arithmetic, denoted PA, is the deductively-closed theory in LA axiomatized by the following:
(L1) xyz[x < y y < z x < z]

(Id1) x[(x + 0 = x) (x 0 = 0)]
(L2) x[x x]
(Id2) x[x 1 = x]
(L3) xy[x < y x = y y < x]
(A1) xyz[(x + y) + z = x + (y + z)]
(O1) 0 < 1 x[x > 0 x 1]
(A2) xy[x + y = y + x] along
(O2) x[x 0]
(M1) xyz[(x y) z = x (y z)]
(O3) xyz[x < y x + z < y + z]
(M2) xy[x y = y x]
(O4) xyz[0 < z x < y x z < y z]
(D1) xyz[x (y + z) = x y + x z]
(O5) xy[x < y (z)[x + z = y]]
with an instance of
y[(0, y) x[(x, y) (x + 1, y)] x(x, y)]
for every (x, y) in LA .
Comparing the original Dedekind-Peano second-order induction axiom with the replacement first-order induction
schema, we trade induction over arbitrary subsets of Na philosophically and set-theoretically questionable actfor
induction over the first-order, LA definable subsets of Na safer, but weaker ability. Its precisely this loss whichas
we will seeallows the formulation of so-called non-standard models of PA.
Note that, while not finitely axiomatizable, PA is computably axiomatizable, and thus, by earlier work, computably
enumerable.
16.3 Robinsons Arithmetic

Robinsons arithmetic is the result of stripping all axioms from Peano arithmetic which are not strictly necessary for
the establishment of PAs undecidability and incompleteness, thus giving a minimalist system in which Gdels famous
theorems still hold.
Robinsons Arithmetic
Robinsons arithmetic, denoted Q, is the deductively-closed theory in LA axiomatized by the following:
(L4) xy[x < y + 1 x y]

(Id11 ) x[x + 0 = x]
(L5) x[x = 0 y[x = y + 1]]
(Id12 ) x[x 0 = 0]
(L6) xy[x + 1 = y + 1 x = y]
(A1 ) xy[x + (y + 1) = (x + y) + 1]
(O6) x[x + 1 , 0]
(D1 ) xy[x (y + 1) = (x y) + x]
(O7) xy[x < y + 1 z[(x + z) + 1 = y]]
There are a number of facts about Q which follow just from considering the axioms given; first, Q is finitely axiom-
atizable, and thus computably enumerable. Next, a bit of work shows that all the axioms of Q are either axioms or
consequences of axioms of PA. That is, Q PA and N Q. Finally, Q lacks any induction axiom, meaning its
possible to prove, for every natural number every instance of a property (x), but yet not obtain the universalization
x(x).
16.4 Gdels Incompleteness Theorems
16.4.1 Representability in a Theory
Numeral
The set of numerals of Q/PA is defined inductively as follows:
The constant 0 is a numeral
If is a numeral, then the term S () is a numeral

The numerals of Q/PA, then, are simply the closed terms we instinctively associate with the natural numbers; e.g.
the numeral for 2 is SS0. Since actually writing numerals out is rather tiresome, its common to either overline
or underline a number in order to denote its numeral:
2 = 2 = SS0
We adopt the latter convention.

16.4. GDELS INCOMPLETENESS THEOREMS 129
Representability
Given a theory T in a language L with a suitable notion of numeral,
(i) A k-ary relation R is weakly represented in T by a well-formed formula of L with k free variables if and
only if for any k-tuple of numbers n,
R(n1 , . . . , nk ) T ` [n1 , . . . , nk ]
(ii) A k-ary relation R is strongly represented in T by a well-formed formula of L with k free variables if and
only if for any k-tuple of numbers n,
R(n1 , . . . , nk ) T ` [n1 , . . . , nk ]
R(n1 , . . . , nk ) T ` [n1 , . . . , nk ]
(iii) A k-ary function f is functionally represented in T by a well-formed formula with k + 1 free variables if
and only if for any natural numbers n1 , . . . , nk , m,
f (n) = m T ` [n1 , . . . , nk , m]
f (n) = m T ` (!z)[n1 , . . . , nk , z]
Proposition 16.1 (Strong Representation in PA).

If PA is consistent and R is strongly represented in PA, then:
R(n1 , . . . , nk ) PA ` [n1 , . . . , nk ]
R(n1 , . . . , nk ) PA ` [n1 , . . . , nk ]
Thus, a relation is strongly represented in PA if and only if it is weakly represented by some formula and its
negation is weakly represented by the negation of the same formula.
Proposition 16.2 (Representation in PA).

The following hold for the theory PA defined above:
(i) The identity relation is strongly represented in PA by the formula x = y

(ii) Addition of natural numbers is represented in PA by x + y = z
(iii) Multiplication of natural numbers is represented in PA by x y = z
(iv) The less than relation on the natural numbers is strongly represented in PA by x < y
(v) A relation R is strongly representable in PA if and only if its characteristic function R is representable in PA
(vi) For any LA -formula and any n N,
PA ` (0) (1) (n) x < n + 1 (x)
(vii) For any LA -formula and any n N, if for every m < n, PA ` (m) and PA ` (n) as well, then
PA ` x[(x) y[y < x (y)] x = n]
Proof Sketch.
The above results, though sometimes tedious to prove, all straightforwardly rely on the axioms of PA,
requiring only clever usage there of.
Theorem 16.4 (Computable Functions are PA-Representable).

Every computable function is representable in PA; in fact, every computable function is representable by a 01 -formula
in PA
Proof Sketch.
Leverage the definition of the computable functions; show that the starting functions are representable in
PA, and that representability in PA is closed under the closure operations of the computable functions.
Corollary (Computable Relations are Strongly Represented in PA).

Every computable relation is strongly representable in PA; in fact, every computable relation is strongly representable
by a 01 -formula in PA
Sound
A theory T is D-sound for a class of sentences D if and only if whenever is a D sentence and T ` , then is
true (what this means, of course, depends on context; usually truth in some standard model).
Of course, PA was built with the intent that PA Th(N); unfortunately, this hasnt been proven. Indeed, although
there is a great deal of anecdotal evidence, it isfor all intents and purposesan open question whether or not PA is
even consistent. Soundness results with regard to PA, then, are not nearly as trivial as they may, at first, seem.
Theorem 16.5 (PA is 01 -Sound).

If PA is consistent, then PA is 01 -sound; that isassuming consistencyfor any 01 , LA -formula such that PA ` ,
N
Proof Sketch.
Leverage the representability of computable relations, note that if PA ` [n], then PA ` x(x)
Corollary (PA is 01 -Sound).

If PA is consistent, then PA is 01 -sound
Theorem 16.6 (01 and True in N Provable in PA).

If is a 01 sentence of LA such that N , then PA `
Proof Sketch.
Note that PA strongly represents the formula [a] by itself (our results above translate addition in N
to addition in PA, etc.) where a is the witness to x(x). By our earlier results, PA ` [a], and thus
PA ` x(x).
Theorem 16.7 (Computably Axiomatizable, Extends PA, and 01 -Sound 01 -Complete).

If a deductively-closed theory T is computably axiomatizable, PA T , and T is 01 -sound, then T is 01 -complete.
Proof Sketch.
Computable axiomatizability guarantees that T is 01 , while extending PA and being 01 -sound guarantee
that T is consistent and still correctly proves 01 facts about Nmaking it 01 -hard.
Theorem 16.8 (PA is Essentially Undecidable).

The theory PA is essentially undecidable.
Proof Sketch.
Let A, B N be disjoint, 01 , and computably inseparable. Since A, B are 01 , PA strongly represents them;
say by A (x) and B (x). Any decidable extension of Q can thus computably separate A and B using A ,
B a contradiction.
16.4.2 Representability in Q
Although we present it in less detail, many of the representability results above carry over verbatim into Robinsons
arithmetic; of course, this is because Robinson selected precisely the axioms needed to prove them.
Theorem 16.9 (Computable Functions are Q-Representable).

Every computable function is representable in Q
Corollary (Computable Relations are Strongly Represented in Q).

Every computable relation is strongly representable in Q
Corollary (01 Relations are Weakly Represented in Q).

Every 01 relation is weakly represented in Q by a 01 formula
Theorem 16.10 (Q is 01 -Sound).

If Q is consistent, then Q is 01 -sound; that isassuming consistencyfor any 01 , LA -formula such that Q ` , N
Theorem 16.11 (01 and True in N Provable in Q).

If is a 01 sentence of LA such that N , then Q `
Corollary (Q is 01 -Complete).
The theory Q is 01 -complete
Theorem 16.12 (Computably Axiomatizable, Extends Q, and 01 -Sound 01 -Complete).

If a deductively-closed theory T is computably axiomatizable, Q T , and T is 01 -sound, then T is 01 -complete.
Proof Sketch.
Computable axiomatizability guarantees that T is 01 , while extending Q and being 01 -sound guarantee
that T is consistent and still correctly proves 01 facts about Nmaking it 01 -hard.
Theorem 16.13 (Q is Essentially Undecidable).

The theory Q is essentially undecidable.
Proof Sketch.
Let A, B N be disjoint, 01 , and computably inseparable. Since A, B are 01 , Q strongly represents them;
say by A (x) and B (x). Any decidable extension of Q can thus computably separate A and B using A ,
B a contradiction.
16.4.3 The Diagonalization Lemma

Recall that, while it is often hidden, all of the results thus far make recourse to some fixed, well-chosen, and computable
encoding of the language LA wherein pq is the number encoding the LA -formula . The next result is one of the
most important to the establishment of the incompleteness theorems; it shows the validity of the self-reference which
Gdels proof relies so heavily on.
Lemma 16.1 (Diagonalization Lemma).

Let T be an LA theory that 01 -represents all computable functions. Given a well-formed formula (x) of one free
variable from LA , there is a sentence of LA such that
Q ` (pq)
Moreover, if (x) is 01 , then may be taken to be equivalent to a 01 sentence also.
Proof.
Define
if m encodes a LA -formula with one free variable
(
p[n]q
sub(m, n) =
0 otherwise
Note that sub is computable, and thus functionally represented in Q; suppose that the formula which does
so is sub (x, y, z). This established, let a well-formed formula of LA with one free variable, (x), be given.
Define (v) as
u(sub (v, v, u) (u))
Colloquially, (v) claims that there is a u such that u = pv [pvq]q (v = pv (x)q) and [u]. Let be (pq).
Thus, is
u(sub (pq, pq, u) (u))
and, since Q functionally represents sub,
T ` u[sub (pq, pq, u) u = p(pq)q]
Simply using the definition of ,
T ` u(sub (pq, pq, u) (u))
And, by the formula derived via Q functionally representing sub,
T ` u(sub (pq, pq, u) (u)) u = pq
Thus,
T ` (pq)
To establish the converse, by the definition of ,
T ` (pq) sub (pq, pq, pq) (pq)
And thus,
T ` () u(sub (pq, pq, u) (u))
T ` ()
which gives,
T ` ()

There are a few interesting characteristics of the proof given above; first, the proof in no way depended on what
(x) actually is. The fixed point, while depending on , is not in any way based on the content of . In every case,
there is a fixed point of the same form as above. The fixed point itself, above, is often understood to claim I
have property ; while this is certainly a fair description of how behaves, the direct self-reference is misleading.
actually says "there is something which is both and has properties X" where the properties X are enough for Q
to uniquely determine that something as the encoding of that very sentence. The self-reference utilized above is thus
indirect. Finally, while T was used throughout the proof above, note that both Q and PA meet the criteria for T , and
thus the fixed point lemma holds in both.
16.4.4 The Incompleteness Theorems
Theorem 16.14 (Gdels First Incompleteness Theorem).

Let T be a computably axiomatizable, consistent, deductively-closed LA -theory such that Q T . Then, there is a 01
LA sentence such that N and T 6` .
Proof Sketch.
Given that T is computably axiomatizable, its computable to check, given a reasonable encoding scheme,
whether any given natural number encodes a proof of some particular statement. Thus, the following is a
computable relation:
ProofT (x, y) x is a proof of y from the axioms of T
Since Q strongly represents all computable relations, there must be some LA formula ProofT (x, y) with
the requisite properties. Noting that x[ProofT (x, y)] is an LA -formula with only one free variable, the
diagonalization lemma gives that there is a such that
Q ` x[ProofT (x, pq)]
Then, T 6` since, if it did, the proof must have an encoding e, so that ProofT (e, pq). Then,
Q ` ProofT (e, pq)

Q `
Q`
Thus, T would be inconsistent, and so T 6` . But, N . Note finally that is, by construction, equivalent
to a 01 , namely x[ProofT (x, pq)].
Remark. This is, of course, not quite the oft quoted statement of Gdels first incompleteness theorem. In
particular, it leaves open the possibility that T is, in fact, complete, and that T ` . There are a number of ways to get
around this and establish the incompleteness of T . Strengthening the assumptions to include 01 -soundness suffices, as
well as Gdels tact: -consistency instead of simply consistency.
-Consistency
A theory T which interprets the natural numbers is -consistent if and only if, for every (x), if T ` (0),
T ` (1), T ` (2), . . . , then T 6` x[(x)]
In truth, however, neither are necessary. A clever adjustment to Gdels original proof, known as Rossers trick,
suffices to establish the stronger conclusion without strengthening the assumptions given above:
Theorem 16.15 (Gdel-Rosser Incompleteness Theorem).

Let T be a computably axiomatizable, consistent, deductively-closed LA -theory such that Q T . Then, T is incom-
plete. In particular, there is a 01 LA sentence such that N , T 6` , and T 6` .
Proof Sketch.
Given that T is computably axiomatizable, its computable to check, given a reasonable encoding scheme,
whether any given natural number encodes a proof of some particular statement. Thus, the following is a
computable relation:
ProofT (x, y) x is a proof of y from the axioms of T
Since Q strongly represents all computable relations, there must be some LA formula ProofT (x, y) with
the requisite properties; similarly,
pq x = pq for some LA formula

(
neg(x) =
0 otherwise
is also computable, and thus represented by some LA -formula, neg (x, y). Note that the following, denoted
(y) for short, is an LA -formula with one free variable:
x[ProofT (x, y) (u < x)(z)[ProofT (u, z) neg (y, z)]]
Less formally, the formula (y) above states "for every proof of the formula encoded by y, there is a
shorter proof of the negation of the formula encoded by y". Its this shift from a formula claiming its
own unprovability in Gdels proof to one claiming the earlier provability of its negation which comprises
Rossers contribution. From here, the rest of the proof follows easily.
Notation. The computability of proof for computably axiomatizable theories and the strong
representability of this relation in theories extending Q allowed us, in the proofs above, to talk
about the PA-provability of various sentences in PA itself. We continue this trend by utilizing
the formula which strongly represents provability to define a LA sentence which claims that
T Q is consistent. Let ConT denote
x[ProofT (x, pq)]
Theorem 16.16 (Gdels Second Incompleteness Theorem).

Let T be axiomatizable, consistent, deductively-closed, and PA T . Then, T 6` ConT
Proof Sketch.
Using our work in Gdels first and the fixed point lemma, let be such that
Well show
PA ` ConT
which, by our previous work, implies PA 6` ConT . Throughout the remainder of the proofs, we use T
as an abbreviation of x[ProofT (x, pq)] (the notation is, of course, remnant of modal logic; in particular,
the provability logic GL; see Boolos text on the subject). To start, note that
PA ` T
and, by a reductio on the negation of the consequent,
PA ` T ConT
Thus,
PA ` ConT
For the reverse, note that, by the construction of ,
PA ` T T T
Thus,
PA ` T T
In other words,
PA ` T ConT
Contraposing,
PA ` ConT T
By the definition of ,
PA ` ConT
Putting the two together,
PA ` ConT
Theorem 16.17 (Lbs Theorem).

Suppose
PA ` PA
Then, PA ` .
Proof.
(Kripke)
Show PA {} is inconsistent. By assumption and modus tollens,
PA {} ` PA (pq)
But then,
PA {} ` PA{}
By Gdels second, however, it now follows that PA {} is inconsistent.
Remark. Note that PA only proves the soundness of its provability predicate with respect to a sentence when
it knows that holds, i.e. when the value of PA doesnt matter. Given our previous work, this should make a great
deal of sense; for all PA can prove, it is inconsistent, and so provability cant be depended on to entail truth generally.
16.4.5 The Consistency-Strength Hierarchy

The impetus for the so-called Consistency Strength Hierarchy is simply considering the result of adding successive
consistency claims to a theory extending PA; that is, starting with a consistent, computably axiomatizable, deductively-
closed T extending PA,
T0 = T
T i+1 = T i {ConTi }
Some reflection shows, however, that if wed like hierarchy which doesnt immediately collapse, we need another
condition on T ; after all, sometimes adding a consistency claim as above actually yields inconsistency, e.g. consider
starting with PA {ConPA }. We thus stipulate that N T or, less constrictively, T is 01 -sound.
Of course, this hierarchy over T can be extended transfinitely in the typical way:
[
T = Tn
n<
Note that if the original T is computably axiomatizable, so too is T n for every n, as well as T ; in a similar vein, note
that if T is 01 -sound, then so to is T n for every n N.
The Consistency Strength Hierarchy itself is the result of generalizing across the hierarchies generated by different
extensions of PA:
Consistency Strength Hierarchy
For consistent, 01 -sound, deductively-closed theories T, S extending PA, S Con T if and only if for every finite
F S , there is a finite G T such that PA ` ConG ConF . Note that in the finitely axiomatizable case this
reduces to PA ` ConT ConS . The consistency strength hierarchy is the hierarchy over consistent,
01 -sound,deductively-closed theories extending PA generated by the ordering Con .
Notation. Define
(01 )T := { : is a 01 sentence of LA and T ` }
Theorem 16.18 (Conssitency Strength is 01 -sentence Inclusion).

For any consistent, 01 -sound, deductively-closed theories extending PA, S Con T if and only if (01 )S (01 )T
Proof.
()
Let S ` where is universal. Then have a finite F S such that F ` . Let G finite T such
that F Con G. Then, T ` ConG . (Use proof of PA proving consistency of all finite subsets, basically
the same). So, T ` ConF . Then, T ` because T ` xProvF (x, ). T ` xProvF (x, ), and PA `
xProvF (x, ) ConF )
()
Finish this direction
Example 14.
Let PA T . RosserT = "for every proof of me, there is a shorter (smaller Godel number) proof of my
negation". RosserT is universal. T + RosserT Con T
Remark. if T + ConT is 01 -sound, T + ConT Con T (01 )T +RosserT * (01 )T since RosserT is a coun-
terexample
Proof.
Reason in PA. Assume ConT . We need ConT +RosserT . Suppose not. Then, T ` RosserT , so T `
ProvT (x, RosserT ) for some x. For every y < x, T ` ProvT (y, RosserT ). So, T ` RosserT .
So T is inconsistent. Contradiction. Thus, ConT +RosserT .
16.4.6
If T is 01 -sound, then T is 01 -sound because its consistent. The converse does not, however, hold (PA+ ConPA ).
01 -soundness is actually equivalent to 01 -soundness (Look at this).
Theorem 16.19.
Peano Arithmetic is finitely axiomatizable
Proof.
Assume PA consistent. Let F finite such that F Axioms of PA. Then, PA ` ConF ; bound the complexity
of these sentences (there must be some n such that for all F, is 0n -sentence). Note that satisfaction
for 0n is definable over N . We now have truth for F.
Say by the wff ,
PA ` ()
for all sentences of F. Then,
PA ` ()
Then, we use the formalized in PA soundness of some proof system. Thus, if F ` PA, then F ` ConF , so
F is inconsistenta contradiction.
Remark. Let T be axiomatizable in LPA , PA T .
i. Suppose T ` where is universal. Then, PA + ConT ` (T Con

PA + ConT )
ii. (ConT is sterile) Let T + ConT ` , universal. Then, T ` .
Proof.
Have PA + ConT +ConT ` by (i). But, PA ` PA ` ConT ConT +ConT (formalized Godel 2). So,
PA + ConT ` . So T ` .
16.5 Nonstandard Models of Peano Arithmetic

16.5.1 PA and End-Extensions
Let PA denote the theory axiomatized by all the axioms of PA sans the induction schema.
Initial Segment / End-Extension
For two LA -models M and N with M N, N is an end-extension of M (also, M is an initial segment of
N)notated M e Nif and only if for all n N and m M, N n < m n M
16.5. NONSTANDARD MODELS OF PEANO ARITHMETIC 139
Theorem 16.20 (All Models of PA are End-Extensions of N).

Let M PA . Then, the map N dom(M) given by n 7 n M is an embedding from the standard model onto an initial
segment of M.
Proposition 16.3 (PA End-Extensions Preserve 01 Formulas).

If M e N with M PA and N PA , then M 01 N
16.5.2 Overspill and Underspill
Cut
For a model M of PA, a non-empty subset I of M is a cut if and only if for all x I and all y dom(M),
M y < x y I and I is closed under the successor function.
Lemma 16.2 (Proper Cuts are Not Definable).

If I is a proper cut of a PA model, then I is not definable (even with parameters).
Proof.
Taking such a formula, note that the model makes the antecedent of the induction schema true, and thus
the consequent as wella contradiction.
Corollary (Overspill).
Let M PA be non-standard and let I be a proper cut of M. Suppose a dom(M) and (x, a) is an LA formula such
that M (b, a) for all b I. Then, there is c > I in M such that M (x < c)(x, a)
Corollary (Underspill).
Let a M with M PA, I dom(M) a proper cut, and (x, y) an LA formula. Then,
(i) if M (c, a) for c > I in M, then M (x b)(x, a) for some b I

(ii) if for all c > I in M there exists x > I with M (x, a) x < c, then for all b I there exists y I with
M y b (y, a).
16.5.3 Cofinal Extensions
Cofinal Extension
Suppose that M, N are PA models such that M N. N is a cofinal extension of Mnotated M c f Nif and only
if for every n dom(N), there is an m dom(M) such that N n < m
Theorem 16.21 (Cofinal Extension of PA Model and 00 Preserving Elementary Extension).

If M c f N such that M 00 N, then if M PA, M N
Corollary (PA-Model Inclusions are 00 -Elementary).

Any inclusion between models of PA is 00 -Elementary
Part IV
Appendices and References
141
Chapter 17
Orderings
17.1 Linear Orders

x[x < x]
xy[x < y x = y y < x]
xyz[x < y y < z x < z]
17.1.1 Discrete Linear Orders

There are four kinds of infinite, discrete linear order, all of which generate elementary classes:
(i) Infinite discrete linear orders with a starting point, but no endpoint; e.g., (N, <).
(ii) Infinite discrete linear orders with an endpoint, but no starting point; e.g., (N, <) with the reverse ordering.
(iii) Infinite discrete linear orders with both a starting and ending point; e.g. an order of type (i) followed by an order
of type (ii).
(iv) Infinite discrete linear orders without endpoints
xy[x < y z[x < z z < y]]
xy[y < x z[z < x y < z]]
In all cases, quantifier elimination can be used to show that the theory is complete.
17.1.2 Dense Linear Orders

There are four kinds of dense linear orders, all of which generate elementary classes:
(i) Dense linear orders with a starting point, but no endpoint; e.g., ([0, 1), <).
(ii) Dense linear orders with an endpoint, but no starting point; e.g., ((0, 1], <)
(iii) Dense linear orders with both a starting and ending point; e.g. ([0, 1], <)
(iv) Dense linear orders without endpoints; e.g. (Q, <), (R, <)
xy[x < y z[x < z z < y]]

xy[y < x]
xy[x < y]
143
144 CHAPTER 17. ORDERINGS
The theory of dense linear orders without endpoints is complete (use os-Tarski), 0 -categorical (QE), but not
categorical in any uncountable cardinal. Quantifier elimination is often a popular and productive proof tactic
when working with DLOs; popular examples include
In cases (i)-(iii), quantifier elimination can easily be used to show that the theory is complete.
Corollary.
The theory of partial orders is 01 -complete.
Fact. The theory of linear orders is decidable
Example 15.
(Q, ) is 0 -saturated; (Q, ) is not 1 saturated (look at the type realized by i)
Example 16.
(R, ) is not 1 -saturated (look at the type containing n < x for all n N.
Chapter 18
A Brief Synopsis of Set Theory
In the beginning, there was notation:
The empty set; {}

{x : (x)} The set of elements x such that (x) holds
XY The set of all elements of X not in Y
XY The set X is a subset (not necessarily proper) of Y
x = hx1 , . . . , xn i By induction,
hxi = x
hx1 , x2 i = {x1 , {x1 , x2 }}
hx1 , . . . , xk i = hhx1 , . . . , xk1 i, xk i
XY The Cartesian product of X and Y; {hx, yi : x X and y Y}
X1 Xn ((X1 X2 ) Xn )
Xn By induction,
X1 = X
X k = X k1 X
Similarly, an n-ary relation R over a set X is a subset of X n , while a function f is a binary, functional relation.
f A The function f restricted to the set A

X
Y The set of all functions f : X Y
iI Xi The Cartesian product of a collection of sets {Xi : i I}; defined to be the set of all functions f
with domain I such that f (i) Xi for all i I.
Equivalence Relation
For some set X, a relation which is:
i. Reflexive - xRx for every x X
ii. Symmetric - xRy yRx
iii. Transitive - xRy and yRz xRz
145
146 CHAPTER 18. A BRIEF SYNOPSIS OF SET THEORY
Orderings
For some set X,
i. Partial Ordering - A binary relation R over X which is reflexive, transitive, and anti-symmetric (xRy and yRx
x = y)
ii. Simple Ordering - A partial ordering R of X which is connected (for all x, y X, either xRy or yRx)
iii. Well Ordering - A simple ordering R of X with the property that every nonempty subset of X has an R-least
element (for every other y in the subset, xRy)
iv. Strict Well Ordering - A relation R over X which is irreflexive (xRx for every x X) such that adding
{hx, xi : x X} gives a well ordering
Chains
For some set X,
i. X is a chain if and only if X is simply ordered by
ii. X is a well ordered chain if and only if X is well ordered by
18.1 Ordinals
Ordinal
An ordinal is a set such that and is strictly well ordered by the relation
S
Example 17.
In canonical set theory, the first three ordinals are:
0=
1 = {}
2 = {, {}}
Proposition 18.1 (Basic Properties of Ordinals).
i. Every element of an ordinal is itself an ordinal

ii. If , are ordinals, then if and only if or =
iii. Every ordinal is a well ordered chain
iv. Every set of ordinals is strictly well ordered by
v. Every set of ordinals is a well ordered chain

18.1. ORDINALS 147
Successor Ordinal
For a given ordinal , the successor of is the ordinal
+ 1 = {}
Any ordinal which can be generated as the successor of another is called a successor ordinal.
Limit Ordinal
An ordinal which is not a successor ordinal
Example 18.
The least limit ordinal is , the order type of N under <
In addition, we define the infimum of a nonempty set X of ordinals is the least element of X, while the supremum
is the least ordinal which is greater than or equal to every element of X; in fact,
Proposition 18.2 (Infimum and Supremum).

Let X be a nonempty set of ordinals. Then, X and X are ordinals; in particular, the infimum and supremum
respectively of X.
18.1.1 Transfinite Induction

Let be any property of ordinals. Assume that for all ordinals , if holds for all < , then () holds. Then,
holds of all ordinals.
18.1.2 Ordinals as Order Types
Order Type
A well ordering R over a set X has order type for some ordinal if and only if there is a bijection f : X
such that xRy f (x) f (y); f is called an isomorphism
Proposition 18.3.
Every well ordering has exactly one order type
18.1.3 Ordinal Arithmetic

Addition
For two ordinals , , + is defined as the ordinal such that and is isomorphic to the chain . Transfinite
induction shows that not only does + always exist, it is unique. Intuitively, adding the ordinals and can be
pictured as the ordering generated by pasting and togetherin that order.
.|. . {z ... . . . ...

} | {z }

Example 19.
As the above shows, ordinal addition is not commutative:
+1,
1+=
Multiplication
For two ordinals and , is defined as the ordinal given by the lexicographic ordering over the Cartesian product
. Graphically, this is equivalent to listing distinct copies of the ordinal:
.|. . {z ... . . . ... . . . . . . . . . . . . . . . ...

} | {z } | {z }

| {z }
copies
Some thought shows that while ordinal multiplication is associative, itlike additionisnt commutative.
Example 20.
2,
2=
Exponentiation
For better or worse, ordinal exponentiation lacks an easy geometric presentation; for finite exponents, its easiest to
think of exponentiation in terms of multiplication, e.g. 2 = . The general case is given inductively by
0 = 1
+1 = ( )
if is limit, then is the limit of for all <

18.2. CARDINAL NUMBERS 149
18.2 Cardinal Numbers

Cardinal numbers or cardinals are a generalization of the natural numbers used to measure the sizes of sets; this is
accomplished via the notion of bijection.
Cardinal
Two sets are said to have the same cardinality if and only if there exists a bijection between them; the cardinal
numbers is the hierarchy of these sizes. There are two standard ways of defining precisely what the cardinal
numbers are; the von Neumann cardinal assignment sets each cardinal as the least ordinal with that cardinality.
Alternatively, we may identify cardinals as the equivalence class of sets with the given cardinality (note that this is
not a set and so cannot be used in ZFC)
The th infinite cardinal is denoted by ; the first infinite cardinal is |N| = 0 .
18.2.1 Cardinal Arithmetic

Addition
Given the axiom of choice, every cardinal has a unique successor; set wise, given two disjoint sets X and Y with
cardinalities and respectively, + is simply the cardinality of X Y. Note that this correspond to standard
addition if both and are finite, otherwise max(, ). Unlike ordinal addition, cardinal addition is both associative
and commutative.
Multiplication
Cardinal multiplication is similarly well-behaved; given two disjoint sets X and Y with cardinalities and respec-
tively, is simply the cardinality of the Cartesian product X Y. Finite numbers behave as usual and if both and
are non-zero with at least one infinite, = max(, ). Like cardinal addition, cardinal multiplication maintains both
associativity and commutativity.
Exponentiation
Cardinal exponentiation is defined as follows: given two disjoint sets X and Y with cardinalities and respectively,
is |X Y | where X Y is the set of all functions from Y into X, also commonly written as Y X. Finite numbers again
behave as usual and cardinal exponentiation maintains all of the basic exponent manipulations characteristic of the
finite setting.
18.2.2 Basic Results

The successor of a cardinal , commonly denoted by + , is the least cardinal greater than ; that is, + = +1 . Note,
however, that + = +1 , + 1. Paralleling the ordinal definition, a cardinal is a limit cardinal if and only if it is
not the successor of a cardinal.
Strong Limit Cardinals

An important sequence of cardinals and one intimately connected with the notion of a strong limit cardinal is the
beths; the beths are defined recursively by:
i0 = 0
i+1 = 2i
and when is a limit ordinal, [
i = i
<
Meanwhile,
Strong Limit Cardinal

A cardinal is a strong limit cardinal if and only if for all cardinals < , 2 <
Proposition 18.4.
(i) is an infinite limit cardinal if and only if there is a limit ordinal such that =
(ii) is an infinite strong limit cardinal if and only if there is a limit ordinal such that = i
(iii) For all ordinals , i
(iv) There are arbitrarily large cardinals such that || = = i
As a point of interest, the continuum hypothesis (CH) states that i1 = 1 , and the generalized continuum hypoth-
esis (GCH) states that, for all , 2 = +1 . Thus, GCH implies that for all , i = and that every limit cardinal
is, in fact, a strong limit cardinal.
18.2.3 Regular and Cofinal Cardinals
Cofinality
Let be a limit ordinal; a set X is cofinal in if and only if X and = X. The cofinality of , notated
S
cf(), is the least cardinal such that a set of power is cofinal in .
Regular and Irregular cardinals Cofinality Universes

Buehler - Classical Metalogic PDF

Uploaded by

Copyright:

Available Formats

Buehler - Classical Metalogic PDF

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

Buehler - Classical Metalogic PDF

Uploaded by

Copyright:

Available Formats

Classical Metalogic

An Introduction to Classical Model Theory & Computability

Notes by R.J. Buehler

Kevin Kellys computability theory notes

Robert Soares Recursively Enumerable Sets and Degrees"

I Classical Model Theory 5

3 Relations Between Models 13

4 Compactness and Strong Completeness 21

5.3.3 The Upward Lwenheim-Skolem Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

7 Types, Saturation, and Categoricity 37

8 Other Model-Theoretic Constructions 49

10 The Primitive and -Recursive Functions 65

11 The Fundamental Theorems of Computability Theory 73

13 Relative Computation and Turing Reducibility 95

14 Posts Problem & Oracle Constructions 107

III Computability and First-Order Logic 115

16 Arithmetic & Incompleteness 125

16.4.3 The Diagonalization Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

IV Appendices and References 141

18 A Brief Synopsis of Set Theory 145

Every variable is a term of L

Nothing else is a term of L.

1.2.2 Atomic Formulas and Sentences

1.2.3 Formal Languages

If L and || < , then

L Any number of disj./conj. and any number of quantifiers

1.2.4 Standard Abbreviations

1.3 Normal Forms

where the i are finite sets of literals.

Disjunctive Normal Form (DNF)

where the i are finite sets of literals.

Prenex Normal Form

Classical Model Theory

The Basics of Models

Models are symbolized by capital letters: A, B, C, . . .

Sequences of elements from a model are represented by placing a bar: a = a1 , a2 , . . . , an where a1 , a2 , . . . , an

2.2 Models and Formal Languages

If t( x) is c for some constant c of the signature L, then t A [a] is cA

A t s [a] if and only if t A [a] = sA [a]

A [a] if and only if for some formula ( x) , A [a]

Formula Equivalence in a Model

2.2.1 Model-Theoretic Logical Consequence

Model-Theoretic Logical Consequence

2.3.1 Theories and Diagrams of Models

2.3.2 Canonical Models

For every closed term t, t = t is in T

 Theorem 2.1 (Henkin Models for Atomic Theories).

T is exactly the set of atomic sentences in A.

Relations Between Models

3.1 From Homomorphism to Isomorphism

For each constant symbol c in L, (cA ) = cB

 Theorem 3.1 (Basic Properties).

(iii) The relation  is an equivalence relation on the class of L-models.

 Theorem 3.2 (Basics of Homomorphism).

For all constants c, cB X.

 Theorem 3.3 (A Bound on Generated Submodels).

3.3 Preserving Formulas

3.3.1 and Preservation

 Theorem 3.4 (Embeddings Preserve -Formulas Upwards).

Theorem 2.1 (Henkin Models for Atomic Theories).

Theorem 3.1 (Basic Properties).

(iii) The relation is an equivalence relation on the class of L-models.

Theorem 3.2 (Basics of Homomorphism).

Theorem 3.3 (A Bound on Generated Submodels).

Theorem 3.4 (Embeddings Preserve -Formulas Upwards).

Corollary (Submodels Preserve -Formulas Downwards).

Lemma 3.2 (T -Models are Submodels of T -models).

Theorem 3.5 (os-Tarski Preservation Theorem).

Corollary (Embeddings Preserve Exactly 1 Upwards).

Theorem 3.6 (Vaught-Tarski Criterion).

Theorem 3.7 (Chains Preserve ).

Theorem 3.8 (Unions of Elementary Chains are Elementary [ Extensions).

Lemma 3.3 (Elementary Diagram Lemma).

Theorem 4.1 (Hintikka Sets have Models).

Theorem 4.2 (Finite Satisfiability, Completeness, and Witnesses).

Theorem 4.3 (Compactness for First-Order Logic).

Lemma 4.1 (Extending Finitely Satisfiable Theories).

Theorem 4.4 (Soundness).

Theorem 4.5 (Completeness).

Proposition 5.1 (Fixed by Automorphisms Definable Relation).

Theorem 5.1 (Skolemization Theorem).

Theorem 5.2 (Skolem Hulls are Elementary Submodels).

Theorem 5.3 (Downward Lwenheim-Skolem Theorem).

Proposition 5.2 (Bigger Models Always Exist).

Proposition 5.3 (Bigger Extensions Always Exist).

Theorem 5.4 (Upward Lwenheim-Skolem).

Theorem 6.1 (Elementary Amalgamation Property).

Theorem 6.2 (Strong Elementary Amalgamation Property).

Theorem 6.3 (Strong Elementary Amalgamation over Algebraically Closed Sets).

Corollary (Strong Amalgamation for Elementary Equivalent Models).

Theorem 6.4 (Elementary Extensions across Language Expansions).

Theorem 6.5 (Craigs Interpolation Theorem).

Theorem 6.6 (Beths Definability Theorem).

Proposition 7.1 (Type Finitely Satisfiable).