Tobias Nipkow Lawrence C.

Paulson
Markus Wenzel

HO
e lle L

a b ∀
Is =
α
λ
β →

A Proof Assistant for

Higher-Order Logic

September 11, 2023

Springer-Verlag
Berlin Heidelberg NewYork
London Paris Tokyo
Hong Kong Barcelona
Budapest
Preface

This volume is a self-contained introduction to interactive proof in higher-

order logic (HOL), using the proof assistant Isabelle. It is written for potential
users rather than for our colleagues in the research world.
The book has three parts.
– The first part, Elementary Techniques, shows how to model functional
programs in higher-order logic. Early examples involve lists and the natural
numbers. Most proofs are two steps long, consisting of induction on a
chosen variable followed by the auto tactic. But even this elementary part
covers such advanced topics as nested and mutual recursion.
– The second part, Logic and Sets, presents a collection of lower-level
tactics that you can use to apply rules selectively. It also describes Isa-
belle/HOL’s treatment of sets, functions and relations and explains how to
define sets inductively. One of the examples concerns the theory of model
checking, and another is drawn from a classic textbook on formal languages.
– The third part, Advanced Material, describes a variety of other topics.
Among these are the real numbers, records and overloading. Advanced
techniques for induction and recursion are described. A whole chapter is
devoted to an extended example: the verification of a security protocol.
The typesetting relies on Wenzel’s theory presentation tools. An anno-
tated source file is run, typesetting the theory in the form of a LATEX source
file. This book is derived almost entirely from output generated in this way.
The final chapter of Part I explains how users may produce their own formal
documents in a similar fashion.
Isabelle’s web site1 contains links to the download area and to documenta-
tion and other information. The classic Isabelle user interface is Proof Gen-
eral / Emacs by David Aspinall’s. This book says very little about Proof
General, which has its own documentation.
This tutorial owes a lot to the constant discussions with and the valuable
feedback from the Isabelle group at Munich: Stefan Berghofer, Olaf Müller,
Wolfgang Naraschewski, David von Oheimb, Leonor Prensa Nieto, Cornelia
Pusch, Norbert Schirmer and Martin Strecker. Stephan Merz was also kind
1
https://isabelle.in.tum.de/
iv Preface

enough to read and comment on a draft version. We received comments from

Stefano Bistarelli, Gergely Buday, John Matthews and Tanja Vos.
The research has been funded by many sources, including the dfg
grants NI 491/2, NI 491/3, NI 491/4, NI 491/6, bmbf project Verisoft,
the epsrc grants GR/K57381, GR/K77051, GR/M75440, GR/R01156/01
GR/S57198/01 and by the esprit working groups 21900 and IST-1999-29001
(the Types project).
Table of Contents

Part I. Elementary Techniques

1. The Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Types, Terms and Formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Interaction and Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2. Functional Programming in HOL . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.1 An Introductory Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 An Introductory Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Some Helpful Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Datatypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5.1 Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5.2 The General Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5.3 Primitive Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.4 Case Expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.5 Structural Induction and Case Distinction . . . . . . . . . . . 19
2.5.6 Case Study: Boolean Expressions . . . . . . . . . . . . . . . . . . . 19
2.6 Some Basic Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6.1 Natural Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6.2 Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6.3 Datatype option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7.1 Type Synonyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7.2 Constant Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.8 The Definitional Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3. More Functional Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.1 Simplification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1.1 What is Simplification? . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1.2 Simplification Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
vi Table of Contents

3.1.3 The simp Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.1.4 Adding and Deleting Simplification Rules . . . . . . . . . . . . 29
3.1.5 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.6 Rewriting with Definitions . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.7 Simplifying let-Expressions . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.8 Conditional Simplification Rules . . . . . . . . . . . . . . . . . . . . 31
3.1.9 Automatic Case Splits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.10 Tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.11 Finding Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2 Induction Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Case Study: Compiling Expressions . . . . . . . . . . . . . . . . . . . . . . . 37
3.4 Advanced Datatypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.4.1 Mutual Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.4.2 Nested Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4.3 The Limits of Nested Recursion . . . . . . . . . . . . . . . . . . . . 43
3.4.4 Case Study: Tries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.5 Total Recursive Functions: fun . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.5.2 Termination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5.3 Simplification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5.4 Induction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

4. Presenting Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1 Concrete Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1.1 Infix Annotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1.2 Mathematical Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.1.3 Prefix Annotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.1.4 Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2 Document Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2.1 Isabelle Sessions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2.2 Structure Markup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2.3 Formal Comments and Antiquotations . . . . . . . . . . . . . . 60
4.2.4 Interpretation of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2.5 Suppressing Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

Part II. Logic and Sets

5. The Rules of the Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

5.1 Natural Deduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Introduction Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.3 Elimination Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.4 Destruction Rules: Some Examples . . . . . . . . . . . . . . . . . . . . . . . 71
5.5 Implication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.6 Negation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
 Table of Contents vii

5.7 Interlude: the Basic Methods for Rules . . . . . . . . . . . . . . . . . . . . 75

5.8 Unification and Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.8.1 Substitution and the subst Method . . . . . . . . . . . . . . . . 77
5.8.2 Unification and Its Pitfalls . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.9 Quantifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.9.1 The Universal Introduction Rule . . . . . . . . . . . . . . . . . . . 81
5.9.2 The Universal Elimination Rule . . . . . . . . . . . . . . . . . . . . 81
5.9.3 The Existential Quantifier . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.9.4 Renaming a Bound Variable: rename_tac . . . . . . . . . . . 83
5.9.5 Reusing an Assumption: frule . . . . . . . . . . . . . . . . . . . . . 84
5.9.6 Instantiating a Quantifier Explicitly . . . . . . . . . . . . . . . . 84
5.10 Description Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.10.1 Definite Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.10.2 Indefinite Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.11 Some Proofs That Fail . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.12 Proving Theorems Using the blast Method . . . . . . . . . . . . . . . . 90
5.13 Other Classical Reasoning Methods . . . . . . . . . . . . . . . . . . . . . . . 91
5.14 Finding More Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.15 Forward Proof: Transforming Theorems . . . . . . . . . . . . . . . . . . . 94
5.15.1 Modifying a Theorem using of, where and THEN . . . . . . 94
5.15.2 Modifying a Theorem using OF . . . . . . . . . . . . . . . . . . . . . 96
5.16 Forward Reasoning in a Backward Proof . . . . . . . . . . . . . . . . . . 97
5.16.1 The Method insert . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.16.2 The Method subgoal_tac . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.17 Managing Large Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.17.1 Tacticals, or Control Structures . . . . . . . . . . . . . . . . . . . . 100
5.17.2 Subgoal Numbering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.18 Proving the Correctness of Euclid’s Algorithm . . . . . . . . . . . . . 103

6. Sets, Functions and Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

6.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.1.1 Finite Set Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.1.2 Set Comprehension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.1.3 Binding Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.1.4 Finiteness and Cardinality . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.2 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.2.1 Function Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.2.2 Injections, Surjections, Bijections . . . . . . . . . . . . . . . . . . . 112
6.2.3 Function Image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.3 Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.3.1 Relation Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.3.2 The Reflexive and Transitive Closure . . . . . . . . . . . . . . . 114
6.3.3 A Sample Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.4 Well-Founded Relations and Induction . . . . . . . . . . . . . . . . . . . . 116
6.5 Fixed Point Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
viii Table of Contents

6.5.1 Propositional Dynamic Logic — PDL . . . . . . . . . . . . . . . 118

6.5.2 Computation Tree Logic — CTL . . . . . . . . . . . . . . . . . . . 121

7. Inductively Defined Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

7.1 The Set of Even Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.1.1 Making an Inductive Definition . . . . . . . . . . . . . . . . . . . . 127
7.1.2 Using Introduction Rules . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.1.3 Rule Induction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.1.4 Generalization and Rule Induction . . . . . . . . . . . . . . . . . 129
7.1.5 Rule Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.1.6 Mutually Inductive Definitions . . . . . . . . . . . . . . . . . . . . . 132
7.1.7 Inductively Defined Predicates . . . . . . . . . . . . . . . . . . . . . 132
7.2 The Reflexive Transitive Closure . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.3 Advanced Inductive Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7.3.1 Universal Quantifiers in Introduction Rules . . . . . . . . . 136
7.3.2 Alternative Definition Using a Monotone Function . . . . 137
7.3.3 A Proof of Equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.3.4 Another Example of Rule Inversion . . . . . . . . . . . . . . . . . 140
7.4 Case Study: A Context Free Grammar . . . . . . . . . . . . . . . . . . . . 141

Part III. Advanced Material

8. More about Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

8.1 Pairs and Tuples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.1.1 Pattern Matching with Tuples . . . . . . . . . . . . . . . . . . . . . 149
8.1.2 Theorem Proving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
8.2 Records . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.2.1 Record Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.2.2 Extensible Records and Generic Operations . . . . . . . . . . 153
8.2.3 Record Equality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.2.4 Extending and Truncating Records . . . . . . . . . . . . . . . . . 156
8.3 Type Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.3.1 Overloading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.3.2 Axioms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
8.4 Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
8.4.1 Numeric Literals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
8.4.2 The Type of Natural Numbers, nat . . . . . . . . . . . . . . . . . 165
8.4.3 The Type of Integers, int . . . . . . . . . . . . . . . . . . . . . . . . . 166
8.4.4 The Types of Rational, Real and Complex Numbers . . 167
8.4.5 The Numeric Type Classes . . . . . . . . . . . . . . . . . . . . . . . . 168
8.5 Introducing New Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
8.5.1 Declaring New Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.5.2 Defining New Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
 Table of Contents ix

9. Advanced Simplification and Induction . . . . . . . . . . . . . . . . . . . 175

9.1 Simplification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
9.1.1 Advanced Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
9.1.2 How the Simplifier Works . . . . . . . . . . . . . . . . . . . . . . . . . 177
9.2 Advanced Induction Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 178
9.2.1 Massaging the Proposition . . . . . . . . . . . . . . . . . . . . . . . . . 178
9.2.2 Beyond Structural and Recursion Induction . . . . . . . . . . 180
9.2.3 Derivation of New Induction Schemas . . . . . . . . . . . . . . . 182
9.2.4 CTL Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

10. Case Study: Verifying a Security Protocol . . . . . . . . . . . . . . . . 187

10.1 The Needham-Schroeder Public-Key Protocol . . . . . . . . . . . . . . 187
10.2 Agents and Messages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
10.3 Modelling the Adversary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
10.4 Event Traces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
10.5 Modelling the Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
10.6 Proving Elementary Properties . . . . . . . . . . . . . . . . . . . . . . . . . . 193
10.7 Proving Secrecy Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

A. Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
x Table of Contents
 Part I

Elementary Techniques
1. The Basics

1.1 Introduction

This book is a tutorial on how to use the theorem prover Isabelle/HOL as a

specification and verification system. Isabelle is a generic system for imple-
menting logical formalisms, and Isabelle/HOL is the specialization of Isabelle
for HOL, which abbreviates Higher-Order Logic. We introduce HOL step by
step following the equation

HOL = Functional Programming + Logic.

We do not assume that you are familiar with mathematical logic. However, we
do assume that you are used to logical and set theoretic notation, as covered
in a good discrete mathematics course [34], and that you are familiar with
the basic concepts of functional programming [5, 14, 29, 35]. Although this
tutorial initially concentrates on functional programming, do not be misled:
HOL can express most mathematical concepts, and functional programming
is just one particularly simple and ubiquitous instance.
Isabelle [28] is implemented in ML [19]. This has influenced some of Isa-
belle/HOL’s concrete syntax but is otherwise irrelevant for us: this tutorial
is based on Isabelle/Isar [37], an extension of Isabelle which hides the im-
plementation language almost completely. Thus the full name of the system
should be Isabelle/Isar/HOL, but that is a bit of a mouthful.
There are other implementations of HOL, in particular the one by Mike
Gordon et al., which is usually referred to as “the HOL system” [10]. For
us, HOL refers to the logical system, and sometimes its incarnation Isa-
belle/HOL.
A tutorial is by definition incomplete. Currently the tutorial only intro-
duces the rudiments of Isar’s proof language. To fully exploit the power of
Isar, in particular the ability to write readable and structured proofs, you
should start with Nipkow’s overview [24] and consult the Isabelle/Isar Refer-
ence Manual [37] and Wenzel’s PhD thesis [38] (which discusses many proof
patterns) for further details. If you want to use Isabelle’s ML level directly
(for example for writing your own proof procedures) see the Isabelle Reference
Manual [26]; for details relating to HOL see the Isabelle/HOL manual [25].
All manuals have a comprehensive index.
4 1. The Basics

1.2 Theories

Working with Isabelle means creating theories. Roughly speaking, a theory

is a named collection of types, functions, and theorems, much like a module
in a programming language or a specification in a specification language. In
fact, theories in HOL can be either. The general format of a theory T is
theory T
imports B1 . . . Bn
begin
declarations, definitions, and proofs
end
where B1 … Bn are the names of existing theories that T is based on and
declarations, definitions, and proofs represents the newly introduced concepts
(types, functions etc.) and proofs about them. The Bi are the direct parent
theories of T. Everything defined in the parent theories (and their parents,
recursively) is automatically visible. To avoid name clashes, identifiers can
be qualified by theory names as in T.f and B.f. Each theory T must reside
in a theory file named T.thy.
This tutorial is concerned with introducing you to the different linguistic
constructs that can fill the declarations, definitions, and proofs above. A com-
plete grammar of the basic constructs is found in the Isabelle/Isar Reference
Manual [37].

!! HOL contains a theory Main, the union of all the basic predefined theories like
arithmetic, lists, sets, etc. Unless you know what you are doing, always include
Main as a direct or indirect parent of all your theories.

HOL’s theory collection is available online at

https://isabelle.in.tum.de/library/HOL/

and is recommended browsing. In subdirectory Library you find a growing

library of useful theories that are not part of Main but can be included among
the parents of a theory and will then be loaded automatically.
For the more adventurous, there is the Archive of Formal Proofs, a journal-
like collection of more advanced Isabelle theories:
https://isa-afp.org/

We hope that you will contribute to it yourself one day.

1.3 Types, Terms and Formulae

Embedded in a theory are the types, terms and formulae of HOL. HOL is
a typed logic whose type system resembles that of functional programming
languages like ML or Haskell. Thus there are
 1.3 Types, Terms and Formulae 5

base types, in particular bool, the type of truth values, and nat, the type of
natural numbers.
type constructors, in particular list, the type of lists, and set, the type of
sets. Type constructors are written postfix, e.g. (nat)list is the type
of lists whose elements are natural numbers. Parentheses around single
arguments can be dropped (as in nat list), multiple arguments are sep-
arated by commas (as in (bool,nat)ty ).
function types, denoted by ⇒. In HOL ⇒ represents total functions only. As
is customary, τ1 ⇒ τ2 ⇒ τ3 means τ1 ⇒ (τ2 ⇒ τ3 ). Isabelle also sup-
ports the notation [τ1 , . . . , τn ] ⇒ τ which abbreviates τ1 ⇒ · · · ⇒ τn
⇒ τ.
type variables, denoted by 'a, 'b etc., just like in ML. They give rise to
polymorphic types like 'a ⇒ 'a, the type of the identity function.

!! Types are extremely important because they prevent us from writing nonsense.
Isabelle insists that all terms and formulae must be well-typed and will print an
error message if a type mismatch is encountered. To reduce the amount of explicit
type information that needs to be provided by the user, Isabelle infers the type of
all variables automatically (this is called type inference) and keeps quiet about
it. Occasionally this may lead to misunderstandings between you and the system. If
anything strange happens, we recommend that you ask Isabelle to display all type
information via the Proof General menu item Isabelle > Settings > Show Types (see
Sect. 1.5 for details).
Terms are formed as in functional programming by applying functions
to arguments. If f is a function of type τ1 ⇒ τ2 and t is a term of type
τ1 then f t is a term of type τ2 . HOL also supports infix functions like +
and some basic constructs from functional programming, such as conditional
expressions:
if b then t1 else t2 Here b is of type bool and t1 and t2 are of the same
type.
let x = t in u is equivalent to u where all free occurrences of x have been re-
placed by t. For example, let x = 0 in x+x is equivalent to 0+0. Multiple
bindings are separated by semicolons: let x1 = t1 ;…; xn = tn in u .
case e of c1 ⇒ e1 | … | cn ⇒ en evaluates to ei if e is of the form ci .
Terms may also contain λ-abstractions. For example, λx. x+1 is the func-
tion that takes an argument x and returns x+1. Instead of λx.λy.λz. t we
can write λx y z. t .
Formulae are terms of type bool. There are the basic constants True
and False and the usual logical connectives (in decreasing order of priority):
¬, ∧, ∨, and −→, all of which (except the unary ¬) associate to the right.
In particular A −→ B −→ C means A −→ (B −→ C) and is thus logically
equivalent to A ∧ B −→ C (which is (A ∧ B) −→ C ).
Equality is available in the form of the infix function = of type 'a ⇒ 'a
⇒ bool. Thus t1 = t2 is a formula provided t1 and t2 are terms of the same
6 1. The Basics

type. If t1 and t2 are of type bool then = acts as if-and-only-if. The formula
t1 6= t2 is merely an abbreviation for ¬(t1 = t2 ).
Quantifiers are written as ∀ x. P and ∃ x. P . There is even ∃! x. P , which
means that there exists exactly one x that satisfies P . Nested quantifications
can be abbreviated: ∀ x y z. P means ∀ x.∀ y.∀ z. P .
Despite type inference, it is sometimes necessary to attach explicit type
constraints to a term. The syntax is t::τ as in x < (y::nat). Note that
:: binds weakly and should therefore be enclosed in parentheses. For in-
stance, x < y::nat is ill-typed because it is interpreted as (x < y)::nat. Type
constraints may be needed to disambiguate expressions involving overloaded
functions such as +, * and <. Section 8.3.1 discusses overloading, while Ta-
ble A.2 presents the most important overloaded function symbols.
In general, HOL’s concrete syntax tries to follow the conventions of func-
tional programming and mathematics. Here are the main rules that you
should be familiar with to avoid certain syntactic traps:
– Remember that f t u means (f t) u and not f(t u)!
– Isabelle allows infix functions like +. The prefix form of function application
binds more strongly than anything else and hence f x + y means (f x) + y
and not f(x+y).
– Remember that in HOL if-and-only-if is expressed using equality. But
equality has a high priority, as befitting a relation, while if-and-only-if
typically has the lowest priority. Thus, ¬ ¬ P = P means ¬¬(P = P) and
not (¬¬P) = P. When using = to mean logical equivalence, enclose both
operands in parentheses, as in (A ∧ B) = (B ∧ A).
– Constructs with an opening but without a closing delimiter bind very
weakly and should therefore be enclosed in parentheses if they appear in
subterms, as in (λx. x) = f. This includes if, let, case, λ, and quantifiers.
– Never write λx.x or ∀ x.x=x because x.x is always taken as a single qualified
identifier. Write λx. x and ∀ x. x=x instead.
– Identifiers may contain the characters _ and ', except at the beginning.
For the sake of readability, we use the usual mathematical symbols
throughout the tutorial. Their ascii-equivalents are shown in table A.1 in
the appendix.

!! A particular problem for novices can be the priority of operators. If you are
unsure, use additional parentheses. In those cases where Isabelle echoes your
input, you can see which parentheses are dropped — they were superfluous. If you
are unsure how to interpret Isabelle’s output because you don’t know where the
(dropped) parentheses go, set the Proof General flag Isabelle > Settings > Show
Brackets (see Sect. 1.5).
 1.4 Variables 7

1.4 Variables
Isabelle distinguishes free and bound variables, as is customary. Bound vari-
ables are automatically renamed to avoid clashes with free variables. In ad-
dition, Isabelle has a third kind of variable, called a schematic variable or
unknown, which must have a ? as its first character. Logically, an unknown
is a free variable. But it may be instantiated by another term during the proof
process. For example, the mathematical theorem x = x is represented in Isa-
belle as ?x = ?x, which means that Isabelle can instantiate it arbitrarily. This
is in contrast to ordinary variables, which remain fixed. The programming
language Prolog calls unknowns logical variables.
Most of the time you can and should ignore unknowns and work with
ordinary variables. Just don’t be surprised that after you have finished the
proof of a theorem, Isabelle will turn your free variables into unknowns. It
indicates that Isabelle will automatically instantiate those unknowns suitably
when the theorem is used in some other proof. Note that for readability we
often drop the ? s when displaying a theorem.

!! For historical reasons, Isabelle accepts ? as an ASCII representation of the ∃

symbol. However, the ? character must then be followed by a space, as in ? x.
f(x) = 0. Otherwise, ?x is interpreted as a schematic variable. The preferred ASCII
representation of the ∃ symbol is EX.

1.5 Interaction and Interfaces

The recommended interface for Isabelle/Isar is the (X)Emacs-based Proof
General [1, 2]. Interaction with Isabelle at the shell level, although possible,
should be avoided. Most of the tutorial is independent of the interface and is
phrased in a neutral language. For example, the phrase “to abandon a proof”
corresponds to the obvious action of clicking on the Undo symbol in Proof
General. Proof General specific information is often displayed in paragraphs
identified by a miniature Proof General icon. Here are two examples:
Proof General supports a special font with mathematical symbols known as “x-
symbols”. All symbols have ascii-equivalents: for example, you can enter either
& or \<and> to obtain ∧. For a list of the most frequent symbols see table A.1 in
the appendix.
Note that by default x-symbols are not enabled. You have to switch them on
via the menu item Proof-General > Options > X-Symbols (and save the option via
the top-level Options menu).

Proof General offers the Isabelle menu for displaying information and setting
flags. A particularly useful flag is Isabelle > Settings > Show Types which causes
Isabelle to output the type information that is usually suppressed. This is indis-
pensible in case of errors of all kinds because often the types reveal the source of
the problem. Once you have diagnosed the problem you may no longer want to see
the types because they clutter all output. Simply reset the flag.
8 1. The Basics

1.6 Getting Started

Assuming you have installed Isabelle and Proof General, you start it by typing
Isabelle in a shell window. This launches a Proof General window. By
default, you are in HOL1 .

You can choose a different logic via the Isabelle > Logics menu.

1
This is controlled by the ISABELLE_LOGIC setting, see The Isabelle System Manual
for more details.
2. Functional Programming in HOL

This chapter describes how to write functional programs in HOL and how
to verify them. However, most of the constructs and proof procedures in-
troduced are general and recur in any specification or verification task. We
really should speak of functional modelling rather than functional program-
ming: our primary aim is not to write programs but to design abstract models
of systems. HOL is a specification language that goes well beyond what can
be expressed as a program. However, for the time being we concentrate on
the computable.
If you are a purist functional programmer, please note that all functions in
HOL must be total: they must terminate for all inputs. Lazy data structures
are not directly available.

2.1 An Introductory Theory

Functional programming needs datatypes and functions. Both of them can be

defined in a theory with a syntax reminiscent of languages like ML or Haskell.
As an example consider the theory in figure 2.1. We will now examine it line
by line.
theory ToyList
imports Main
begin

HOL already has a predefined theory of lists called List — ToyList is merely a
small fragment of it chosen as an example. To avoid some ambiguities caused
by defining lists twice, we manipulate the concrete syntax and name space of
theory Main as follows.
no notation Nil ("[]") and Cons (infixr "#" 65) and append (infixr "@" 65)
hide type list
hide const rev

datatype 'a list = Nil ("[]")

| Cons 'a "'a list" (infixr "#" 65)

The datatype list introduces two constructors Nil and Cons, the empty list
and the operator that adds an element to the front of a list. For example,
10 2. Functional Programming in HOL
theory ToyList
imports Main
begin

no_notation Nil ("[]") and Cons (infixr "#" 65) and append (infixr "@" 65)
hide_type list
hide_const rev

datatype 'a list = Nil ("[]")

| Cons 'a "'a list" (infixr "#" 65)

(* This is the append function: *)

primrec app :: "'a list => 'a list => 'a list" (infixr "@" 65)
where
"[] @ ys = ys" |
"(x # xs) @ ys = x # (xs @ ys)"

primrec rev :: "'a list => 'a list" where

"rev [] = []" |
"rev (x # xs) = (rev xs) @ (x # [])"

Fig. 2.1. A Theory of Lists

the term Cons True (Cons False Nil) is a value of type bool list, namely
the list with the elements True and False. Because this notation quickly
becomes unwieldy, the datatype declaration is annotated with an alternative
syntax: instead of Nil and Cons x xs we can write [] and x # xs. In fact,
this alternative syntax is the familiar one. Thus the list Cons True (Cons
False Nil) becomes True # False # []. The annotation infixr means that #
associates to the right: the term x # y # z is read as x # (y # z) and not as
(x # y) # z. The 65 is the priority of the infix #.

!! Syntax annotations can be powerful, but they are difficult to master and are
never necessary. You could drop them from theory ToyList and go back to the
identifiers Nil and Cons. Novices should avoid using syntax annotations in their
own theories.

Next, two functions app and rev are defined recursively, in this order,
because Isabelle insists on definition before use:
primrec app :: "'a list ⇒ 'a list ⇒ 'a list" (infixr "@" 65) where
"[] @ ys = ys" |
"(x # xs) @ ys = x # (xs @ ys)"

primrec rev :: "'a list ⇒ 'a list" where

"rev [] = []" |
"rev (x # xs) = (rev xs) @ (x # [])"

Each function definition is of the form

primrec name :: type (optional syntax) where equations
 2.2 Evaluation 11

The equations must be separated by |. Function app is annotated with con-

crete syntax. Instead of the prefix syntax app xs ys the infix xs @ ys becomes
the preferred form.
The equations for app and rev hardly need comments: app appends two
lists and rev reverses a list. The keyword primrec indicates that the recursion
is of a particularly primitive kind where each recursive call peels off a datatype
constructor from one of the arguments. Thus the recursion always terminates,
i.e. the function is total.
The termination requirement is absolutely essential in HOL, a logic of
total functions. If we were to drop it, inconsistencies would quickly arise: the
“definition” f (n) = f (n) + 1 immediately leads to 0 = 1 by subtracting f (n)
on both sides.

!! As we have indicated, the requirement for total functions is an essential char-

acteristic of HOL. It is only because of totality that reasoning in HOL is com-
paratively easy. More generally, the philosophy in HOL is to refrain from asserting
arbitrary axioms (such as function definitions whose totality has not been proved)
because they quickly lead to inconsistencies. Instead, fixed constructs for introduc-
ing types and functions are offered (such as datatype and primrec) which are
guaranteed to preserve consistency.

A remark about syntax. The textual definition of a theory follows a fixed

syntax with keywords like datatype and end. Embedded in this syntax are
the types and formulae of HOL, whose syntax is extensible (see Sect. 4.1), e.g.
by new user-defined infix operators. To distinguish the two levels, everything
HOL-specific (terms and types) should be enclosed in "…". To lessen this
burden, quotation marks around a single identifier can be dropped, unless
the identifier happens to be a keyword, for example "end". When Isabelle
prints a syntax error message, it refers to the HOL syntax as the inner
syntax and the enclosing theory language as the outer syntax.
Comments must be in enclosed in (*and*).

2.2 Evaluation

Assuming you have processed the declarations and definitions of ToyList

presented so far, you may want to test your functions by running them. For
example, what is the value of rev (True # False # [])? Command
value "rev (True # False # [])"

yields the correct result False # True # []. But we can go beyond mere func-
tional programming and evaluate terms with variables in them, executing
functions symbolically:
value "rev (a # b # c # [])"

yields c # b # a # [].
12 2. Functional Programming in HOL

2.3 An Introductory Proof

Having convinced ourselves (as well as one can by testing) that our definitions
capture our intentions, we are ready to prove a few simple theorems. This
will illustrate not just the basic proof commands but also the typical proof
process.
Main Goal. Our goal is to show that reversing a list twice produces the
original list.
theorem rev_rev [simp]: "rev(rev xs) = xs"

This theorem command does several things:

– It establishes a new theorem to be proved, namely rev (rev xs) = xs.
– It gives that theorem the name rev_rev, for later reference.
– It tells Isabelle (via the bracketed attribute simp ) to take the eventual
theorem as a simplification rule: future proofs involving simplification will
replace occurrences of rev (rev xs) by xs.
The name and the simplification attribute are optional. Isabelle’s response
is to print the initial proof state consisting of some header information (like
how many subgoals there are) followed by
1. rev (rev xs) = xs

For compactness reasons we omit the header in this tutorial. Until we have
finished a proof, the proof state proper always looks like this:
1. G1
.
.
.
n. Gn
The numbered lines contain the subgoals G1 , …, Gn that we need to prove to
establish the main goal. Initially there is only one subgoal, which is identical
with the main goal. (If you always want to see the main goal as well, set the
flag Proof.show_main_goal — this flag used to be set by default.)
Let us now get back to rev (rev xs) = xs. Properties of recursively de-
fined functions are best established by induction. In this case there is nothing
obvious except induction on xs:
apply(induct_tac xs)
This tells Isabelle to perform induction on variable xs. The suffix tac stands
for tactic, a synonym for “theorem proving function”. By default, induction
acts on the first subgoal. The new proof state contains two subgoals, namely
the base case (Nil) and the induction step (Cons):
1. V
rev (rev []) = []
2. x1 x2. rev (rev x2) = x2 =⇒ rev (rev (x1 # x2)) = x1 # x2

The induction step is an example of the general format of a subgoal:

 2.3 An Introductory Proof 13

i. x1 …xn . assumptions =⇒ conclusion

V

The prefix of bound variables x1 … xn can be ignored most of the time, or

V
simply treated as a list of variables local to this subgoal. Their deeper signif-
icance is explained in Chapter 5. The assumptions are the local assumptions
for this subgoal and conclusion is the actual proposition to be proved. Typical
proof steps that add new assumptions are induction and case distinction. In
our example the only assumption is the induction hypothesis rev (rev list)
= list, where list is a variable name chosen by Isabelle. If there are multiple
assumptions, they are enclosed in the bracket pair [[ and ]] and separated by
semicolons.
Let us try to solve both goals automatically:
apply(auto)

This command tells Isabelle to apply a proof strategy called auto to all sub-
goals. Essentially, auto tries to simplify the subgoals. In our case, subgoal 1 is
solved completely (thanks to the equation rev [] = []) and disappears; the
simplified version of subgoal 2 becomes the new subgoal 1:
1. x1 x2. rev (rev x2) = x2 =⇒ rev (rev x2 @ x1 # []) = x1 # x2
V

In order to simplify this subgoal further, a lemma suggests itself.

First Lemma. After abandoning the above proof attempt (at the shell level
type oops) we start a new proof:
lemma rev_app [simp]: "rev(xs @ ys) = (rev ys) @ (rev xs)"
The keywords theorem and lemma are interchangeable and merely indi-
cate the importance we attach to a proposition. Therefore we use the words
theorem and lemma pretty much interchangeably, too.
There are two variables that we could induct on: xs and ys. Because @ is
defined by recursion on the first argument, xs is the correct one:
apply(induct_tac xs)

This time not even the base case is solved automatically:

apply(auto)

1. rev ys = rev ys @ []
A total of 2 subgoals...

Again, we need to abandon this proof attempt and prove another simple
lemma first. In the future the step of abandoning an incomplete proof before
embarking on the proof of a lemma usually remains implicit.
14 2. Functional Programming in HOL

Second Lemma. We again try the canonical proof procedure:

lemma app_Nil2 [simp]: "xs @ [] = xs"
apply(induct_tac xs)
apply(auto)

It works, yielding the desired message No subgoals!:

xs @ [] = xs
No subgoals!

We still need to confirm that the proof is now finished:

done

As a result of that final done, Isabelle associates the lemma just proved with
its name. In this tutorial, we sometimes omit to show that final done if it is
obvious from the context that the proof is finished.
Notice that in lemma app_Nil2, as printed out after the final done, the
free variable xs has been replaced by the unknown ?xs, just as explained in
Sect. 1.4.
Going back to the proof of the first lemma
lemma rev_app [simp]: "rev(xs @ ys) = (rev ys) @ (rev xs)"
apply(induct_tac xs)
apply(auto)

we find that this time auto solves the base case, but the induction step merely
simplifies to
1. x1 x2.
V
rev (x2 @ ys) = rev ys @ rev x2 =⇒
(rev ys @ rev x2) @ x1 # [] = rev ys @ rev x2 @ x1 # []

Now we need to remember that @ associates to the right, and that # and @
have the same priority (namely the 65 in their infixr annotation). Thus the
conclusion really is
(rev ys @ rev list) @ (a # []) = rev ys @ (rev list @ (a # []))

and the missing lemma is associativity of @.

Third Lemma. Abandoning the previous attempt, the canonical proof pro-
cedure succeeds without further ado.
lemma app_assoc [simp]: "(xs @ ys) @ zs = xs @ (ys @ zs)"
apply(induct_tac xs)
apply(auto)
done

Now we can prove the first lemma:

lemma rev_app [simp]: "rev(xs @ ys) = (rev ys) @ (rev xs)"
apply(induct_tac xs)
apply(auto)
done
 2.3 An Introductory Proof 15

Finally, we prove our main theorem:

theorem rev_rev [simp]: "rev(rev xs) = xs"
apply(induct_tac xs)
apply(auto)
done

The final end tells Isabelle to close the current theory because we are finished
with its development:
end
The complete proof script is shown in Fig. 2.2. The concatenation of
Figs. 2.1 and 2.2 constitutes the complete theory ToyList and should reside
in file ToyList.thy.

lemma app_Nil2 [simp]: "xs @ [] = xs"

apply(induct_tac xs)
apply(auto)
done

lemma app_assoc [simp]: "(xs @ ys) @ zs = xs @ (ys @ zs)"

apply(induct_tac xs)
apply(auto)
done

lemma rev_app [simp]: "rev(xs @ ys) = (rev ys) @ (rev xs)"

apply(induct_tac xs)
apply(auto)
done

theorem rev_rev [simp]: "rev(rev xs) = xs"

apply(induct_tac xs)
apply(auto)
done

end

Fig. 2.2. Proofs about Lists

Review This is the end of our toy proof. It should have familiarized you
with
– the standard theorem proving procedure: state a goal (lemma or theorem);
proceed with proof until a separate lemma is required; prove that lemma;
come back to the original goal.
– a specific procedure that works well for functional programs: induction
followed by all-out simplification via auto.
– a basic repertoire of proof commands.
16 2. Functional Programming in HOL

!! It is tempting to think that all lemmas should have the simp attribute just
because this was the case in the example above. However, in that example all
lemmas were equations, and the right-hand side was simpler than the left-hand
side — an ideal situation for simplification purposes. Unless this is clearly the case,
novices should refrain from awarding a lemma the simp attribute, which has a
global effect. Instead, lemmas can be applied locally where they are needed, which
is discussed in the following chapter.

2.4 Some Helpful Commands

This section discusses a few basic commands for manipulating the proof state
and can be skipped by casual readers.
There are two kinds of commands used during a proof: the actual proof
commands and auxiliary commands for examining the proof state and con-
trolling the display. Simple proof commands are of the form apply(method),
where method is typically induct_tac or auto. All such theorem proving oper-
ations are referred to as methods, and further ones are introduced through-
out the tutorial. Unless stated otherwise, you may assume that a method
attacks merely the first subgoal. An exception is auto, which tries to solve all
subgoals.
The most useful auxiliary commands are as follows:
Modifying the order of subgoals: defer moves the first subgoal to the end
and prefer n moves subgoal n to the front.
Printing theorems: thm name1 … namen prints the named theorems.
Reading terms and types: term string reads, type-checks and prints the
given string as a term in the current context; the inferred type is output
as well. typ string reads and prints the given string as a type in the
current context.
Further commands are found in the Isabelle/Isar Reference Manual [37].

Clicking on the State button redisplays the current proof state. This is helpful
in case commands like thm have overwritten it.

We now examine Isabelle’s functional programming constructs systemat-

ically, starting with inductive datatypes.

2.5 Datatypes

Inductive datatypes are part of almost every non-trivial application of HOL.

First we take another look at an important example, the datatype of lists,
before we turn to datatypes in general. The section closes with a case study.
 2.5 Datatypes 17

2.5.1 Lists

Lists are one of the essential datatypes in computing. We expect that you are
already familiar with their basic operations. Theory ToyList is only a small
fragment of HOL’s predefined theory List1 . The latter contains many further
operations. For example, the functions hd (“head”) and tl (“tail”) return
the first element and the remainder of a list. (However, pattern matching is
usually preferable to hd and tl.) Also available are higher-order functions like
map and filter. Theory List also contains more syntactic sugar: [ x1 ,…,xn ]
abbreviates x1 # …# xn #[]. In the rest of the tutorial we always use HOL’s
predefined lists by building on theory Main.

!! Looking ahead to sets and quanifiers in Part II: The best way to express that
some element x is in a list xs is x ∈ set xs, where set is a function that turns
a list into the set of its elements. By the same device you can also write bounded
quantifiers like ∀x ∈ set xs or embed lists in other set expressions.

2.5.2 The General Format

The general HOL datatype definition is of the form

datatype (α1 , . . . , αn ) t = C1 τ11 . . . τ1k1 | . . . | Cm τm1 . . . τmkm

where αi are distinct type variables (the parameters), Ci are distinct con-
structor names and τij are types; it is customary to capitalize the first letter
in constructor names. There are a number of restrictions (such as that the
type should not be empty) detailed elsewhere [25]. Isabelle notifies you if you
violate them.
Laws about datatypes, such as [] 6= x#xs and (x#xs = y#ys) = (x=y
∧ xs=ys), are used automatically during proofs by simplification. The same
is true for the equations in primitive recursive function definitions.
Every2 datatype t comes equipped with a size function from t into the
natural numbers (see Sect. 2.6.1 below). For lists, size is just the length, i.e.
size [] = 0 and size(x # xs) = size xs + 1. In general, size returns

– zero for all constructors that do not have an argument of type t,

– one plus the sum of the sizes of all arguments of type t, for all other
constructors.
Note that because size is defined on every datatype, it is overloaded; on lists
size is also called length, which is not overloaded. Isabelle will always show
size on lists as length.
1
https://isabelle.in.tum.de/library/HOL/List.html
2
Except for advanced datatypes where the recursion involves “⇒” as in Sect. 3.4.3.
18 2. Functional Programming in HOL

2.5.3 Primitive Recursion

Functions on datatypes are usually defined by recursion. In fact, most of

the time they are defined by what is called primitive recursion over some
datatype t. This means that the recursion equations must be of the form

f x1 . . . (C y1 . . . yk ) . . . xn = r

such that C is a constructor of t and all recursive calls of f in r are of the

form f . . . yi . . . for some i. Thus Isabelle immediately sees that f termi-
nates because one (fixed!) argument becomes smaller with every recursive
call. There must be at most one equation for each constructor. Their order
is immaterial. A more general method for defining total recursive functions
is introduced in Sect. 3.5.
Exercise 2.5.1 Define the datatype of binary trees:
datatype 'a tree = Tip | Node "'a tree" 'a "'a tree"

Define a function mirror that mirrors a binary tree by swapping subtrees

recursively. Prove
lemma mirror_mirror: "mirror(mirror t) = t"

Define a function flatten that flattens a tree into a list by traversing it in

infix order. Prove
lemma "flatten(mirror t) = rev(flatten t)"

2.5.4 Case Expressions

HOL also features case-expressions for analyzing elements of a datatype. For

example,
case xs of [] ⇒ [] | y # ys ⇒ y

evaluates to [] if xs is [] and to y if xs is y # ys. (Since the result in both

branches must be of the same type, it follows that y is of type 'a list and
hence that xs is of type 'a list list.)
In general, case expressions are of the form

case e of pattern1 ⇒ e1 | . . . | patternm ⇒ em

Like in functional programming, patterns are expressions consisting of data-

type constructors (e.g. [] and #) and variables, including the wildcard “_”.
Not all cases need to be covered and the order of cases matters. However,
one is well-advised not to wallow in complex patterns because complex case
distinctions tend to induce complex proofs.
 2.5 Datatypes 19

!! Internally Isabelle only knows about exhaustive case expressions with non-
nested patterns: patterni must be of the form Ci xi1 . . . xiki and C1 , . . . , Cm
must be exactly the constructors of the type of e. More complex case expressions are
automatically translated into the simpler form upon parsing but are not translated
back for printing. This may lead to surprising output.

!! Like if, case-expressions may need to be enclosed in parentheses to indicate

their scope.

2.5.5 Structural Induction and Case Distinction

Induction is invoked by induct_tac, as we have seen above; it works for any

datatype. In some cases, induction is overkill and a case distinction over all
constructors of the datatype suffices. This is performed by case_tac. Here is
a trivial example:
lemma "(case xs of [] ⇒ [] | y#ys ⇒ xs) = xs"
apply(case_tac xs)

results in the proof state

1. V
xs = [] =⇒ (case xs of [] ⇒ [] | y # ys ⇒ xs) = xs
2. a list.
xs = a # list =⇒ (case xs of [] ⇒ [] | y # ys ⇒ xs) = xs

which is solved automatically:

apply(auto)

Note that we do not need to give a lemma a name if we do not intend

to refer to it explicitly in the future. Other basic laws about a datatype
are applied automatically during simplification, so no special methods are
provided for them.

!! Induction is only allowed on free (or -bound) variables that should not occur
V
among the assumptions of the subgoal; see Sect. 9.2.1 for details. Case distinc-
tion (case_tac) works for arbitraryV terms, which need to be quoted if they are
non-atomic. However, apart from -bound variables, the terms must not contain
variables that are bound outside. For example, given the goal ∀ xs. xs = [] ∨
(∃ y ys. xs = y # ys), case_tac xs will not work as expected because Isabelle
interprets the xs as a new free variable distinct from the bound xs in the goal.

2.5.6 Case Study: Boolean Expressions

The aim of this case study is twofold: it shows how to model boolean expres-
sions and some algorithms for manipulating them, and it demonstrates the
constructs introduced above.
20 2. Functional Programming in HOL

Modelling Boolean Expressions. We want to represent boolean expres-

sions built up from variables and constants by negation and conjunction. The
following datatype serves exactly that purpose:
datatype boolex = Const bool | Var nat | Neg boolex
| And boolex boolex
The two constants are represented by Const True and Const False. Variables
are represented by terms of the form Var n, where n is a natural number (type
nat). For example, the formula P0 ∧ ¬P1 is represented by the term And (Var
0) (Neg (Var 1)).
The Value of a Boolean Expression. The value of a boolean expression
depends on the value of its variables. Hence the function value takes an addi-
tional parameter, an environment of type nat ⇒ bool, which maps variables
to their values:
primrec "value" :: "boolex ⇒ (nat ⇒ bool) ⇒ bool" where
"value (Const b) env = b" |
"value (Var x) env = env x" |
"value (Neg b) env = (¬ value b env)" |
"value (And b c) env = (value b env ∧ value c env)"

If-Expressions. An alternative and often more efficient (because in a cer-

tain sense canonical) representation are so-called If-expressions built up from
constants (CIF ), variables (VIF ) and conditionals (IF ):
datatype ifex = CIF bool | VIF nat | IF ifex ifex ifex
The evaluation of If-expressions proceeds as for boolex :
primrec valif :: "ifex ⇒ (nat ⇒ bool) ⇒ bool" where
"valif (CIF b) env = b" |
"valif (VIF x) env = env x" |
"valif (IF b t e) env = (if valif b env then valif t env
else valif e env)"

Converting Boolean and If-Expressions. The type boolex is close to

the customary representation of logical formulae, whereas ifex is designed
for efficiency. It is easy to translate from boolex into ifex :
primrec bool2if :: "boolex ⇒ ifex" where
"bool2if (Const b) = CIF b" |
"bool2if (Var x) = VIF x" |
"bool2if (Neg b) = IF (bool2if b) (CIF False) (CIF True)" |
"bool2if (And b c) = IF (bool2if b) (bool2if c) (CIF False)"
At last, we have something we can verify: that bool2if preserves the value
of its argument:
lemma "valif (bool2if b) env = value b env"
The proof is canonical:
apply(induct_tac b)
 2.5 Datatypes 21

apply(auto)
done
In fact, all proofs in this case study look exactly like this. Hence we do not
show them below.
More interesting is the transformation of If-expressions into a normal
form where the first argument of IF cannot be another IF but must be a
constant or variable. Such a normal form can be computed by repeatedly
replacing a subterm of the form IF (IF b x y) z u by IF b (IF x z u) (IF
y z u), which has the same value. The following primitive recursive functions
perform this task:
primrec normif :: "ifex ⇒ ifex ⇒ ifex ⇒ ifex" where
"normif (CIF b) t e = IF (CIF b) t e" |
"normif (VIF x) t e = IF (VIF x) t e" |
"normif (IF b t e) u f = normif b (normif t u f) (normif e u f)"

primrec norm :: "ifex ⇒ ifex" where

"norm (CIF b) = CIF b" |
"norm (VIF x) = VIF x" |
"norm (IF b t e) = normif b (norm t) (norm e)"
Their interplay is tricky; we leave it to you to develop an intuitive under-
standing. Fortunately, Isabelle can help us to verify that the transformation
preserves the value of the expression:
theorem "valif (norm b) env = valif b env"
The proof is canonical, provided we first show the following simplification
lemma, which also helps to understand what normif does:
lemma [simp]:
"∀ t e. valif (normif b t e) env = valif (IF b t e) env"
Note that the lemma does not have a name, but is implicitly used in the
proof of the theorem shown above because of the [simp] attribute.
But how can we be sure that norm really produces a normal form in the
above sense? We define a function that tests If-expressions for normality:
primrec normal :: "ifex ⇒ bool" where
"normal(CIF b) = True" |
"normal(VIF x) = True" |
"normal(IF b t e) = (normal t ∧ normal e ∧
(case b of CIF b ⇒ True | VIF x ⇒ True | IF x y z ⇒ False))"
Now we prove normal (norm b). Of course, this requires a lemma about nor-
mality of normif :
lemma [simp]: "∀ t e. normal(normif b t e) = (normal t ∧ normal e)"

How do we come up with the required lemmas? Try to prove the main
theorems without them and study carefully what auto leaves unproved. This
can provide the clue. The necessity of universal quantification (∀ t e) in the
two lemmas is explained in Sect. 3.2
22 2. Functional Programming in HOL

Exercise 2.5.2 We strengthen the definition of a normal If-expression as

follows: the first argument of all IF s must be a variable. Adapt the above
development to this changed requirement. (Hint: you may need to formulate
some of the goals as implications (−→) rather than equalities (=).)

2.6 Some Basic Types

This section introduces the types of natural numbers and ordered pairs. Also
described is type option, which is useful for modelling exceptional cases.

2.6.1 Natural Numbers

The type nat of natural numbers is predefined to have the constructors 0

and Suc. It behaves approximately as if it were declared like this:
datatype nat = zero ("0") | Suc nat
In particular, there are case-expressions, for example
case n of 0 ⇒ 0 | Suc m ⇒ m
primitive recursion, for example
primrec sum :: "nat ⇒ nat" where
"sum 0 = 0" |
"sum (Suc n) = Suc n + sum n"
and induction, for example
lemma "sum n + sum n = n*(Suc n)"
apply(induct_tac n)
apply(auto)
done
The arithmetic operations + , -, *, div , mod, min and max are predefined,
as are the relations ≤ and < . As usual, m - n = 0 if m < n. There is even a
least number operation LEAST . For example, (LEAST n. 0 < n) = Suc 0.

!! The constants 0 and 1 and the operations +, -, *, min, max, ≤ and < are
overloaded: they are available not just for natural numbers but for other types
as well. For example, given the goal x + 0 = x, there is nothing to indicate that you
are talking about natural numbers. Hence Isabelle can only infer that x is of some
arbitrary type where 0 and + are declared. As a consequence, you will be unable
to prove the goal. To alert you to such pitfalls, Isabelle flags numerals without a
fixed type in its output: x + (0::'a) = x. (In the absence of a numeral, it may
take you some time to realize what has happened if Show Types is not set). In this
particular example, you need to include an explicit type constraint, for example x+0
= (x::nat). If there is enough contextual information this may not be necessary:
Suc x = x automatically implies x::nat because Suc is not overloaded.
For details on overloading see Sect. 8.3.1. Table A.2 in the appendix shows the
most important overloaded operations.
 2.6 Some Basic Types 23

!! The symbols > and ≥ are merely syntax: x > y stands for y < x and similary
for ≥ and ≤.

!! Constant 1::nat is defined to equal Suc 0. This definition (see Sect. 2.7.2) is
unfolded automatically by some tactics (like auto, simp and arith) but not by
others (especially the single step tactics in Chapter 5). If you need the full set of
numerals, see Sect. 8.4.1. Novices are advised to stick to 0 and Suc.

Both auto and simp (a method introduced below, Sect. 3.1) prove simple
arithmetic goals automatically:
lemma "[[ ¬ m < n; m < n + (1::nat) ]] =⇒ m = n"

For efficiency’s sake, this built-in prover ignores quantified formulae, many
logical connectives, and all arithmetic operations apart from addition. In
consequence, auto and simp cannot prove this slightly more complex goal:
lemma "m 6= (n::nat) =⇒ m < n ∨ n < m"

The method arith is more general. It attempts to prove the first subgoal
provided it is a linear arithmetic formula. Such formulas may involve the
usual logical connectives (¬, ∧, ∨, −→, =, ∀ , ∃ ), the relations =, ≤ and <, and
the operations +, -, min and max. For example,
lemma "min i (max j (k*k)) = max (min (k*k) i) (min i (j::nat))"
apply(arith)
succeeds because k * k can be treated as atomic. In contrast,
lemma "n*n = n+1 =⇒ n=0"

is not proved by arith because the proof relies on properties of multiplication.

Only multiplication by numerals (which is the same as iterated addition) is
taken into account.

!! The running time of arith is exponential in the number of occurrences of -,

min and max because they are first eliminated by case distinctions.
If k is a numeral, div k, mod k and k dvd are also supported, where the former
two are eliminated by case distinctions, again blowing up the running time.
If the formula involves quantifiers, arith may take super-exponential time and
space.

2.6.2 Pairs

HOL also has ordered pairs: (a1 ,a2 ) is of type τ1 × τ2 provided each ai
is of type τi . The functions fst and snd extract the components of a pair:
fst(x,y) = x and snd(x,y) = y . Tuples are simulated by pairs nested to the
right: (a1 ,a2 ,a3 ) stands for (a1 ,(a2 ,a3 )) and τ1 × τ2 × τ3 for τ1 × (τ2 × τ3 ).
Therefore we have fst(snd(a1 ,a2 ,a3 )) = a2 .
Remarks:
24 2. Functional Programming in HOL

– There is also the type unit, which contains exactly one element denoted
by (). This type can be viewed as a degenerate product with 0 components.
– Products, like type nat, are datatypes, which means in particular that
induct_tac and case_tac are applicable to terms of product type. Both split
the term into a number of variables corresponding to the tuple structure
(up to 7 components).
– Tuples with more than two or three components become unwieldy; records
are preferable.
For more information on pairs and records see Chapter 8.

2.6.3 Datatype option

Our final datatype is very simple but still eminently useful:

datatype 'a option = None | Some 'a

Frequently one needs to add a distinguished element to some existing type.

For example, type t option can model the result of a computation that may
either terminate with an error (represented by None) or return some value
v (represented by Some v ). Similarly, nat extended with ∞ can be modeled
by type nat option. In both cases one could define a new datatype with
customized constructors like Error and Infinity, but it is often simpler to
use option. For an application see Sect. 3.4.4.

2.7 Definitions

A definition is simply an abbreviation, i.e. a new name for an existing con-

struction. In particular, definitions cannot be recursive. Isabelle offers defini-
tions on the level of types and terms. Those on the type level are called type
synonyms; those on the term level are simply called definitions.

2.7.1 Type Synonyms

Type synonyms are similar to those found in ML. They are created by a
type synonym command:
type synonym number = nat
type synonym gate = "bool ⇒ bool ⇒ bool"
type synonym ('a, 'b) alist = "('a × 'b) list"

Internally all synonyms are fully expanded. As a consequence Isabelle’s out-

put never contains synonyms. Their main purpose is to improve the readabil-
ity of theories. Synonyms can be used just like any other type.
 2.8 The Definitional Approach 25

2.7.2 Constant Definitions

Nonrecursive definitions can be made with the definition command, for

example nand and xor gates (based on type gate above):
definition nand :: gate where "nand A B ≡ ¬(A ∧ B)"
definition xor :: gate where "xor A B ≡ A ∧ ¬B ∨ ¬A ∧ B"

The symbol ≡ is a special form of equality that must be used in constant

definitions. Pattern-matching is not allowed: each definition must be of the
form f x1 . . . xn ≡ t. Section 3.1.6 explains how definitions are used in proofs.
The default name of each definition is f _def, where f is the name of the
defined constant.

!! A common mistake when writing definitions is to introduce extra free variables

on the right-hand side. Consider the following, flawed definition (where dvd
means “divides”):
"prime p ≡ 1 < p ∧ (m dvd p −→ m = 1 ∨ m = p)"
Isabelle rejects this “definition” because of the extra m on the right-hand side, which
would introduce an inconsistency (why?). The correct version is
"prime p ≡ 1 < p ∧ (∀ m. m dvd p −→ m = 1 ∨ m = p)"

2.8 The Definitional Approach

As we pointed out at the beginning of the chapter, asserting arbitrary ax-

ioms such as f (n) = f (n) + 1 can easily lead to contradictions. In order to
avoid this danger, we advocate the definitional rather than the axiomatic ap-
proach: introduce new concepts by definitions. However, Isabelle/HOL seems
to support many richer definitional constructs, such as primrec. The point
is that Isabelle reduces such constructs to first principles. For example, each
primrec function definition is turned into a proper (nonrecursive!) definition
from which the user-supplied recursion equations are automatically proved.
This process is hidden from the user, who does not have to understand the
details. Other commands described later, like fun and inductive, work sim-
ilarly. This strict adherence to the definitional approach reduces the risk of
soundness errors.
3. More Functional Programming

The purpose of this chapter is to deepen your understanding of the con-

cepts encountered so far and to introduce advanced forms of datatypes and
recursive functions. The first two sections give a structured presentation of
theorem proving by simplification (Sect. 3.1) and discuss important heuris-
tics for induction (Sect. 3.2). You can skip them if you are not planning to
perform proofs yourself. We then present a case study: a compiler for ex-
pressions (Sect. 3.3). Advanced datatypes, including those involving function
spaces, are covered in Sect. 3.4; it closes with another case study, search trees
(“tries”). Finally we introduce fun, a general form of recursive function def-
inition that goes well beyond primrec (Sect. 3.5).

3.1 Simplification
So far we have proved our theorems by auto, which simplifies all subgoals.
In fact, auto can do much more than that. To go beyond toy examples, you
need to understand the ingredients of auto. This section covers the method
that auto always applies first, simplification.
Simplification is one of the central theorem proving tools in Isabelle and
many other systems. The tool itself is called the simplifier. This section
introduces the many features of the simplifier and is required reading if you
intend to perform proofs. Later on, Sect. 9.1 explains some more advanced
features and a little bit of how the simplifier works. The serious student should
read that section as well, in particular to understand why the simplifier did
something unexpected.

3.1.1 What is Simplification?

In its most basic form, simplification means repeated application of equations
from left to right. For example, taking the rules for @ and applying them to
the term [0,1] @ [] results in a sequence of simplification steps:
(0#1#[]) @ [] ; 0#((1#[]) @ []) ; 0#(1#([] @ [])) ; 0#1#[]
This is also known as term rewriting and the equations are referred to as
rewrite rules. “Rewriting” is more honest than “simplification” because the
terms do not necessarily become simpler in the process.
28 3. More Functional Programming

The simplifier proves arithmetic goals as described in Sect. 2.6.1 above.

Arithmetic expressions are simplified using built-in procedures that go be-
yond mere rewrite rules. New simplification procedures can be coded and
installed, but they are definitely not a matter for this tutorial.

3.1.2 Simplification Rules

To facilitate simplification, the attribute [simp] declares theorems to be sim-

plification rules, which the simplifier will use automatically. In addition,
datatype and primrec declarations (and a few others) implicitly declare
some simplification rules. Explicit definitions are not declared as simplifica-
tion rules automatically!
Nearly any theorem can become a simplification rule. The simplifier will
try to transform it into an equation. For example, the theorem ¬ P is turned
into P = False. The details are explained in Sect. 9.1.2.
The simplification attribute of theorems can be turned on and off:
declare theorem-name[simp]
declare theorem-name[simp del]
Only equations that really simplify, like rev (rev xs) = xs and xs @ [] = xs,
should be declared as default simplification rules. More specific ones should
only be used selectively and should not be made default. Distributivity laws,
for example, alter the structure of terms and can produce an exponential
blow-up instead of simplification. A default simplification rule may need to
be disabled in certain proofs. Frequent changes in the simplification status of
a theorem may indicate an unwise use of defaults.

!! Simplification can run forever, for example if both f (x) = g(x) and g(x) = f (x)
are simplification rules. It is the user’s responsibility not to include simplification
rules that can lead to nontermination, either on their own or in combination with
other simplification rules.

!! It is inadvisable to toggle the simplification attribute of a theorem from a parent

theory A in a child theory B for good. The reason is that if some theory C
is based both on B and (via a different path) on A, it is not defined what the
simplification attribute of that theorem will be in C : it could be either.

3.1.3 The simp Method

The general format of the simplification method is

simp list of modifiers

where the list of modifiers fine tunes the behaviour and may be empty. Specific
modifiers are discussed below. Most if not all of the proofs seen so far could
 3.1 Simplification 29

have been performed with simp instead of auto, except that simp attacks only
the first subgoal and may thus need to be repeated — use simp_all to simplify
all subgoals. If nothing changes, simp fails.

3.1.4 Adding and Deleting Simplification Rules

If a certain theorem is merely needed in a few proofs by simplification, we do

not need to make it a global simplification rule. Instead we can modify the
set of simplification rules used in a simplification step by adding rules to it
and/or deleting rules from it. The two modifiers for this are
add: list of theorem names
del: list of theorem names
Or you can use a specific list of theorems and omit all others:
only: list of theorem names
In this example, we invoke the simplifier, adding two distributive laws:
apply(simp add: mod_mult_distrib add_mult_distrib)

3.1.5 Assumptions

By default, assumptions are part of the simplification process: they are used
as simplification rules and are simplified themselves. For example:
lemma "[[ xs @ zs = ys @ xs; [] @ xs = [] @ [] ]] =⇒ ys = zs"
apply simp
done
The second assumption simplifies to xs = [], which in turn simplifies the first
assumption to zs = ys, thus reducing the conclusion to ys = ys and hence to
True.
In some cases, using the assumptions can lead to nontermination:
lemma "∀ x. f x = g (f (g x)) =⇒ f [] = f [] @ []"
An unmodified application of simp loops. The culprit is the simplification rule
f x = g (f (g x)), which is extracted from the assumption. (Isabelle notices
certain simple forms of nontermination but not this one.) The problem can
be circumvented by telling the simplifier to ignore the assumptions:
apply(simp (no_asm))
done
Three modifiers influence the treatment of assumptions:
(no_asm) means that assumptions are completely ignored.
(no_asm_simp) means that the assumptions are not simplified but are used
in the simplification of the conclusion.
30 3. More Functional Programming

(no_asm_use) means that the assumptions are simplified but are not used in
the simplification of each other or the conclusion.
Only one of the modifiers is allowed, and it must precede all other modifiers.

3.1.6 Rewriting with Definitions

Constant definitions (Sect. 2.7.2) can be used as simplification rules, but by

default they are not: the simplifier does not expand them automatically.
Definitions are intended for introducing abstract concepts and not merely as
abbreviations. Of course, we need to expand the definition initially, but once
we have proved enough abstract properties of the new constant, we can forget
its original definition. This style makes proofs more robust: if the definition
has to be changed, only the proofs of the abstract properties will be affected.
For example, given
definition xor :: "bool ⇒ bool ⇒ bool" where
"xor A B ≡ (A ∧ ¬B) ∨ (¬A ∧ B)"
we may want to prove
lemma "xor A (¬A)"
Typically, we begin by unfolding some definitions:
apply(simp only: xor_def)
In this particular case, the resulting goal
1. A ∧ ¬ ¬ A ∨ ¬ A ∧ ¬ A
can be proved by simplification. Thus we could have proved the lemma out-
right by
apply(simp add: xor_def)
Of course we can also unfold definitions in the middle of a proof.

!! If you have defined f x y ≡ t then you can only unfold occurrences of f with at
least two arguments. This may be helpful for unfolding f selectively, but it may
also get in the way. Defining f ≡ λx y. t allows to unfold all occurrences of f .
There is also the special method unfold which merely unfolds one or
several definitions, as in apply(unfold xor_def). This is can be useful in
situations where simp does too much. Warning: unfold acts on all subgoals!

3.1.7 Simplifying let-Expressions

Proving a goal containing let-expressions almost invariably requires the let-

constructs to be expanded at some point. Since let…= …in… is just syntac-
tic sugar for the predefined constant Let, expanding let-constructs means
rewriting with Let_def :
 3.1 Simplification 31

lemma "(let xs = [] in xs@ys@xs) = ys"

apply(simp add: Let_def)
done
If, in a particular context, there is no danger of a combinatorial explosion
of nested lets, you could even simplify with Let_def by default:
declare Let_def [simp]

3.1.8 Conditional Simplification Rules

So far all examples of rewrite rules were equations. The simplifier also accepts
conditional equations, for example
lemma hd_Cons_tl[simp]: "xs 6= [] =⇒ hd xs # tl xs = xs"
apply(case_tac xs, simp, simp)
done
Note the use of “,” to string together a sequence of methods. Assuming
that the simplification rule (rev xs = []) = (xs = []) is present as well, the
lemma below is proved by plain simplification:
lemma "xs 6= [] =⇒ hd(rev xs) # tl(rev xs) = rev xs"
The conditional equation hd_Cons_tl above can simplify hd (rev xs) # tl
(rev xs) to rev xs because the corresponding precondition rev xs 6= [] sim-
plifies to xs 6= [], which is exactly the local assumption of the subgoal.

3.1.9 Automatic Case Splits

Goals containing if -expressions are usually proved by case distinction on the

boolean condition. Here is an example:
lemma "∀ xs. if xs = [] then rev xs = [] else rev xs 6= []"
The goal can be split by a special method, split:
apply(split if_split)

1. ∀ xs. (xs = [] −→ rev xs = []) ∧ (xs 6= [] −→ rev xs 6= [])

where if_split is a theorem that expresses splitting of if s. Because splitting
the if s is usually the right proof strategy, the simplifier does it automatically.
Try apply(simp) on the initial goal above.
This splitting idea generalizes from if to case. Let us simplify a case
analysis over lists:
lemma "(case xs of [] ⇒ zs | y#ys ⇒ y#(ys@zs)) = xs@zs"
apply(split list.split)

1. (xs = [] −→ zs = xs @ zs) ∧
(∀ x21 x22. xs = x21 # x22 −→ x21 # x22 @ zs = xs @ zs)
32 3. More Functional Programming

The simplifier does not split case-expressions, as it does if -expressions, be-

cause with recursive datatypes it could lead to nontermination. Instead, the
simplifier has a modifier split for adding splitting rules explicitly. The lemma
above can be proved in one step by
apply(simp split: list.split)
whereas apply(simp) alone will not succeed.
Every datatype t comes with a theorem t .split which can be declared
to be a split rule either locally as above, or by giving it the split attribute
globally:
declare list.split [split]
The split attribute can be removed with the del modifier, either locally
apply(simp split del: if_split)
or globally:
declare list.split [split del]
Polished proofs typically perform splitting within simp rather than in-
voking the split method. However, if a goal contains several if and case
expressions, the split method can be helpful in selectively exploring the ef-
fects of splitting.
The split rules shown above are intended to affect only the subgoal’s
conclusion. If you want to split an if or case-expression in the assumptions,
you have to apply if_split_asm or t .split_asm:
lemma "if xs = [] then ys 6= [] else ys = [] =⇒ xs @ ys 6= []"
apply(split if_split_asm)
Unlike splitting the conclusion, this step creates two separate subgoals, which
here can be solved by simp_all:
1. [[xs = []; ys 6= []]] =⇒ xs @ ys 6= []
2. [[xs =
6 []; ys = []]] =⇒ xs @ ys = 6 []
If you need to split both in the assumptions and the conclusion, use t .splits
which subsumes t .split and t .split_asm. Analogously, there is if_splits.

!! The simplifier merely simplifies the condition of an if but not the then or
else parts. The latter are simplified only after the condition reduces to True
or False, or after splitting. The same is true for case-expressions: only the selector
is simplified at first, until either the expression reduces to one of the cases or it is
split.

3.1.10 Tracing

Using the simplifier effectively may take a bit of experimentation. Set the
Proof General flag Isabelle > Settings > Trace Simplifier to get a better idea
of what is going on:
 3.1 Simplification 33

lemma "rev [a] = []"

apply(simp)
produces the following trace in Proof General’s Trace buffer:
[1]Applying instance of rewrite rule "List.rev.simps_2":
rev (?x1 # ?xs1) ≡ rev ?xs1 @ [?x1]

[1]Rewriting:
rev [a] ≡ rev [] @ [a]

[1]Applying instance of rewrite rule "List.rev.simps_1":

rev [] ≡ []

[1]Rewriting:
rev [] ≡ []

[1]Applying instance of rewrite rule "List.append.append_Nil":

[] @ ?y ≡ ?y

[1]Rewriting:
[] @ [a] ≡ [a]

[1]Applying instance of rewrite rule

?x2 # ?t1 = ?t1 ≡ False

[1]Rewriting:
[a] = [] ≡ False

The trace lists each rule being applied, both in its general form and the
instance being used. The [i] in front (where above i is always 1) indicates
that we are inside the ith invocation of the simplifier. Each attempt to apply
a conditional rule shows the rule followed by the trace of the (recursive!)
simplification of the conditions, the latter prefixed by [i + 1] instead of
[i]. Another source of recursive invocations of the simplifier are proofs of
arithmetic formulae. By default, recursive invocations are not shown, you
must increase the trace depth via Isabelle > Settings > Trace Simplifier Depth.
Many other hints about the simplifier’s actions may appear.
In more complicated cases, the trace can be very lengthy. Thus it is ad-
visable to reset the Trace Simplifier flag after having obtained the desired
trace. Since this is easily forgotten (and may have the unpleasant effect of
swamping the interface with trace information), here is how you can switch
the trace on locally in a proof:
using [[simp_trace=true]]
apply simp
Within the current proof, all simplifications in subsequent proof steps will be
traced, but the text reminds you to remove the using clause after it has done
its job.
34 3. More Functional Programming

3.1.11 Finding Theorems

Isabelle’s large database of proved theorems offers a powerful search engine.

Its chief limitation is its restriction to the theories currently loaded.

The search engine is started by clicking on Proof General’s Find icon. You specify
your search textually in the input buffer at the bottom of the window.
The simplest form of search finds theorems containing specified patterns.
A pattern can be any term (even a single identifier). It may contain “ ”, a
wildcard standing for any term. Here are some examples:
length
"_ # _ = _ # _"
"_ + _"
"_ * (_ - (_::nat))"
Specifying types, as shown in the last example, constrains searches involving
overloaded operators.

!! Always use “ ” rather than variable names: searching for "x + y" will usually
not find any matching theorems because they would need to contain x and y
literally. When searching for infix operators, do not just type in the symbol, such
as +, but a proper term such as "_ + _". This remark applies to more complicated
syntaxes, too.
If you are looking for rewrite rules (possibly conditional) that could sim-
plify some term, prefix the pattern with simp:.
simp: "_ * (_ + _)"
This finds all equations—not just those with a simp attribute—whose con-
clusion has the form
_ * (_ + _) = . . .
It only finds equations that can simplify the given pattern at the root, not
somewhere inside: for example, equations of the form _ + _ = . . . do not
match.
You may also search for theorems by name—you merely need to specify
a substring. For example, you could search for all commutativity theorems
like this:
name: comm
This retrieves all theorems whose name contains comm.
Search criteria can also be negated by prefixing them with “-”. For exam-
ple,
-name: List
finds theorems whose name does not contain List. You can use this to exclude
particular theories from the search: the long name of a theorem contains the
name of the theory it comes from.
 3.2 Induction Heuristics 35

Finallly, different search criteria can be combined arbitrarily. The effect

is conjuctive: Find returns the theorems that satisfy all of the criteria. For
example,
"_ + _" -"_ - _" -simp: "_ * (_ + _)" name: assoc
looks for theorems containing plus but not minus, and which do not simplify
_ * (_ + _) at the root, and whose name contains assoc.
Further search criteria are explained in Sect. 5.14.

Proof General keeps a history of all your search expressions. If you click on Find,
you can use the arrow keys to scroll through previous searches and just modify
them. This saves you having to type in lengthy expressions again and again.

3.2 Induction Heuristics

The purpose of this section is to illustrate some simple heuristics for inductive
proofs. The first one we have already mentioned in our initial example:
Theorems about recursive functions are proved by induction.
In case the function has more than one argument
Do induction on argument number i if the function is defined by
recursion in argument number i.
When we look at the proof of (xs@ys) @ zs = xs @ (ys@zs) in Sect. 2.3 we
find
– @ is recursive in the first argument
– xs occurs only as the first argument of @
– both ys and zs occur at least once as the second argument of @
Hence it is natural to perform induction on xs.
The key heuristic, and the main point of this section, is to generalize the
goal before induction. The reason is simple: if the goal is too specific, the
induction hypothesis is too weak to allow the induction step to go through.
Let us illustrate the idea with an example.
Function rev has quadratic worst-case running time because it calls func-
tion @ for each element of the list and @ is linear in its first argument. A linear
time version of rev reqires an extra argument where the result is accumulated
gradually, using only #:
primrec itrev :: "'a list ⇒ 'a list ⇒ 'a list" where
"itrev [] ys = ys" |
"itrev (x#xs) ys = itrev xs (x#ys)"
The behaviour of itrev is simple: it reverses its first argument by stacking
its elements onto the second argument, and returning that second argument
36 3. More Functional Programming

when the first one becomes empty. Note that itrev is tail-recursive: it can
be compiled into a loop.
Naturally, we would like to show that itrev does indeed reverse its first
argument provided the second one is empty:
lemma "itrev xs [] = rev xs"
There is no choice as to the induction variable, and we immediately simplify:
apply(induct_tac xs, simp_all)
Unfortunately, this attempt does not prove the induction step:
1. a list.
V
itrev list [] = rev list =⇒ itrev list [a] = rev list @ [a]

The induction hypothesis is too weak. The fixed argument, [], prevents it
from rewriting the conclusion. This example suggests a heuristic:
Generalize goals for induction by replacing constants by variables.
Of course one cannot do this naïvely: itrev xs ys = rev xs is just not true.
The correct generalization is
lemma "itrev xs ys = rev xs @ ys"

If ys is replaced by [], the right-hand side simplifies to rev xs, as required.

In this instance it was easy to guess the right generalization. Other situ-
ations can require a good deal of creativity.
Although we now have two variables, only xs is suitable for induction,
and we repeat our proof attempt. Unfortunately, we are still not there:
1. a list.
V
itrev list ys = rev list @ ys =⇒
itrev list (a # ys) = rev list @ a # ys

The induction hypothesis is still too weak, but this time it takes no intuition
to generalize: the problem is that ys is fixed throughout the subgoal, but the
induction hypothesis needs to be applied with a # ys instead of ys. Hence we
prove the theorem for all ys instead of a fixed one:
lemma "∀ ys. itrev xs ys = rev xs @ ys"
This time induction on xs followed by simplification succeeds. This leads to
another heuristic for generalization:
Generalize goals for induction by universally quantifying all free vari-
ables (except the induction variable itself!).
This prevents trivial failures like the one above and does not affect the validity
of the goal. However, this heuristic should not be applied blindly. It is not
always required, and the additional quantifiers can complicate matters in
some cases. The variables that should be quantified are typically those that
change in recursive calls.
 3.3 Case Study: Compiling Expressions 37

A final point worth mentioning is the orientation of the equation we just

proved: the more complex notion (itrev ) is on the left-hand side, the sim-
pler one (rev ) on the right-hand side. This constitutes another, albeit weak
heuristic that is not restricted to induction:
The right-hand side of an equation should (in some sense) be simpler
than the left-hand side.
This heuristic is tricky to apply because it is not obvious that rev xs @ ys is
simpler than itrev xs ys. But see what happens if you try to prove rev xs
@ ys = itrev xs ys!
If you have tried these heuristics and still find your induction does not go
through, and no obvious lemma suggests itself, you may need to generalize
your proposition even further. This requires insight into the problem at hand
and is beyond simple rules of thumb. Additionally, you can read Sect. 9.2 to
learn about some advanced techniques for inductive proofs.

Exercise 3.2.1 Define the following addition function

primrec add :: "nat ⇒ nat ⇒ nat" where
"add m 0 = m" |
"add m (Suc n) = add (Suc m) n"
and prove
lemma "add m n = m+n"

Exercise 3.2.2 In Exercise 2.5.1 we defined a function flatten from trees

to lists. The straightforward version of flatten is based on @ and is thus,
like rev, quadratic. A linear time version of flatten again reqires an extra
argument, the accumulator. Define
flatten2 :: "'a tree ⇒ 'a list ⇒ 'a list"
and prove
lemma "flatten2 t [] = flatten t"

3.3 Case Study: Compiling Expressions

The task is to develop a compiler from a generic type of expressions (built

from variables, constants and binary operations) to a stack machine. This
generic type of expressions is a generalization of the boolean expressions
in Sect. 2.5.6. This time we do not commit ourselves to a particular type of
variables or values but make them type parameters. Neither is there a fixed set
of binary operations: instead the expression contains the appropriate function
itself.
type synonym 'v binop = "'v ⇒ 'v ⇒ 'v"
datatype (dead 'a, 'v) expr = Cex 'v
38 3. More Functional Programming

| Vex 'a
| Bex "'v binop" "('a,'v)expr" "('a,'v)expr"
The three constructors represent constants, variables and the application of
a binary operation to two subexpressions.
The value of an expression with respect to an environment that maps
variables to values is easily defined:
primrec "value" :: "('a,'v)expr ⇒ ('a ⇒ 'v) ⇒ 'v" where
"value (Cex v) env = v" |
"value (Vex a) env = env a" |
"value (Bex f e1 e2) env = f (value e1 env) (value e2 env)"
The stack machine has three instructions: load a constant value onto the
stack, load the contents of an address onto the stack, and apply a binary
operation to the two topmost elements of the stack, replacing them by the
result. As for expr, addresses and values are type parameters:
datatype (dead 'a, 'v) instr = Const 'v
| Load 'a
| Apply "'v binop"
The execution of the stack machine is modelled by a function exec that
takes a list of instructions, a store (modelled as a function from addresses
to values, just like the environment for evaluating expressions), and a stack
(modelled as a list) of values, and returns the stack at the end of the execution
— the store remains unchanged:
primrec exec :: "('a,'v)instr list ⇒ ('a⇒'v) ⇒ 'v list ⇒ 'v list"
where
"exec [] s vs = vs" |
"exec (i#is) s vs = (case i of
Const v ⇒ exec is s (v#vs)
| Load a ⇒ exec is s ((s a)#vs)
| Apply f ⇒ exec is s ((f (hd vs) (hd(tl vs)))#(tl(tl vs))))"
Recall that hd and tl return the first element and the remainder of a list.
Because all functions are total, hd is defined even for the empty list, although
we do not know what the result is. Thus our model of the machine always
terminates properly, although the definition above does not tell us much
about the result in situations where Apply was executed with fewer than two
elements on the stack.
The compiler is a function from expressions to a list of instructions. Its
definition is obvious:
primrec compile :: "('a,'v)expr ⇒ ('a,'v)instr list" where
"compile (Cex v) = [Const v]" |
"compile (Vex a) = [Load a]" |
"compile (Bex f e1 e2) = (compile e2) @ (compile e1) @ [Apply f]"
Now we have to prove the correctness of the compiler, i.e. that the exe-
cution of a compiled expression results in the value of the expression:
theorem "exec (compile e) s [] = [value e s]"
 3.4 Advanced Datatypes 39

This theorem needs to be generalized:

theorem "∀ vs. exec (compile e) s vs = (value e s) # vs"
It will be proved by induction on e followed by simplification. First, we
must prove a lemma about executing the concatenation of two instruction
sequences:
lemma exec_app[simp]:
"∀ vs. exec (xs@ys) s vs = exec ys s (exec xs s vs)"
This requires induction on xs and ordinary simplification for the base cases.
In the induction step, simplification leaves us with a formula that contains
two case-expressions over instructions. Thus we add automatic case splitting,
which finishes the proof:
apply(induct_tac xs, simp, simp split: instr.split)
Note that because both simp_all and auto perform simplification, they can
be modified in the same way as simp. Thus the proof can be rewritten as
apply(induct_tac xs, simp_all split: instr.split)
Although this is more compact, it is less clear for the reader of the proof.
We could now go back and prove exec (compile e) s [] = [value e s]
merely by simplification with the generalized version we just proved. How-
ever, this is unnecessary because the generalized version fully subsumes its
instance.

3.4 Advanced Datatypes

This section presents advanced forms of datatypes: mutual and nested re-
cursion. A series of examples will culminate in a treatment of the trie data
structure.

3.4.1 Mutual Recursion

Sometimes it is necessary to define two datatypes that depend on each other.

This is called mutual recursion. As an example consider a language of
arithmetic and boolean expressions where
– arithmetic expressions contain boolean expressions because there are con-
ditional expressions like “if m < n then n − m else m − n”, and
– boolean expressions contain arithmetic expressions because of comparisons
like “m < n”.
In Isabelle this becomes
datatype 'a aexp = IF "'a bexp" "'a aexp" "'a aexp"
| Sum "'a aexp" "'a aexp"
40 3. More Functional Programming

| Diff "'a aexp" "'a aexp"

| Var 'a
| Num nat
and 'a bexp = Less "'a aexp" "'a aexp"
| And "'a bexp" "'a bexp"
| Neg "'a bexp"
Type aexp is similar to expr in Sect. 3.3, except that we have added an IF
constructor, fixed the values to be of type nat and declared the two binary
operations Sum and Diff. Boolean expressions can be arithmetic comparisons,
conjunctions and negations. The semantics is given by two evaluation func-
tions:
primrec evala :: "'a aexp ⇒ ('a ⇒ nat) ⇒ nat" and
evalb :: "'a bexp ⇒ ('a ⇒ nat) ⇒ bool" where
"evala (IF b a1 a2) env =
(if evalb b env then evala a1 env else evala a2 env)" |
"evala (Sum a1 a2) env = evala a1 env + evala a2 env" |
"evala (Diff a1 a2) env = evala a1 env - evala a2 env" |
"evala (Var v) env = env v" |
"evala (Num n) env = n" |

"evalb (Less a1 a2) env = (evala a1 env < evala a2 env)" |

"evalb (And b1 b2) env = (evalb b1 env ∧ evalb b2 env)" |
"evalb (Neg b) env = (¬ evalb b env)"
Both take an expression and an environment (a mapping from variables
'a to values nat) and return its arithmetic/boolean value. Since the datatypes
are mutually recursive, so are functions that operate on them. Hence they
need to be defined in a single primrec section. Notice the and separating
the declarations of evala and evalb. Their defining equations need not be
split into two groups; the empty line is purely for readability.
In the same fashion we also define two functions that perform substitution:
primrec substa :: "('a ⇒ 'b aexp) ⇒ 'a aexp ⇒ 'b aexp" and
substb :: "('a ⇒ 'b aexp) ⇒ 'a bexp ⇒ 'b bexp" where
"substa s (IF b a1 a2) =
IF (substb s b) (substa s a1) (substa s a2)" |
"substa s (Sum a1 a2) = Sum (substa s a1) (substa s a2)" |
"substa s (Diff a1 a2) = Diff (substa s a1) (substa s a2)" |
"substa s (Var v) = s v" |
"substa s (Num n) = Num n" |

"substb s (Less a1 a2) = Less (substa s a1) (substa s a2)" |

"substb s (And b1 b2) = And (substb s b1) (substb s b2)" |
"substb s (Neg b) = Neg (substb s b)"
Their first argument is a function mapping variables to expressions, the sub-
stitution. It is applied to all variables in the second argument. As a result,
the type of variables in the expression may change from 'a to 'b. Note that
there are only arithmetic and no boolean variables.
Now we can prove a fundamental theorem about the interaction between
evaluation and substitution: applying a substitution s to an expression a
 3.4 Advanced Datatypes 41

and evaluating the result in an environment env yields the same result as
evaluation a in the environment that maps every variable x to the value of
s(x) under env. If you try to prove this separately for arithmetic or boolean
expressions (by induction), you find that you always need the other theorem
in the induction step. Therefore you need to state and prove both theorems
simultaneously:
lemma "evala (substa s a) env = evala a (λx. evala (s x) env) ∧
evalb (substb s b) env = evalb b (λx. evala (s x) env)"
apply(induct_tac a and b)
The resulting 8 goals (one for each constructor) are proved in one fell swoop:
apply simp_all
In general, given n mutually recursive datatypes τ1 , …, τn , an inductive
proof expects a goal of the form

P1 (x1 ) ∧ · · · ∧ Pn (xn )

where each variable xi is of type τi . Induction is started by

apply(induct_tac x1 and … and xn )

Exercise 3.4.1 Define a function norma of type 'a aexp ⇒ 'a aexp that re-
places IF s with complex boolean conditions by nested IF s; it should eliminate
the constructors And and Neg, leaving only Less. Prove that norma preserves
the value of an expression and that the result of norma is really normal, i.e.
no more Ands and Neg s occur in it. (Hint: proceed as in Sect. 2.5.6 and read
the discussion of type annotations following lemma subst_id below).

3.4.2 Nested Recursion

So far, all datatypes had the property that on the right-hand side of their
definition they occurred only at the top-level: directly below a constructor.
Now we consider nested recursion, where the recursive datatype occurs nested
in some other datatype (but not inside itself!). Consider the following model
of terms where function symbols can be applied to a list of arguments:
datatype ('v,'f)"term" = Var 'v | App 'f "('v,'f)term list"
Note that we need to quote term on the left to avoid confusion with the
Isabelle command term. Parameter 'v is the type of variables and 'f the
type of function symbols. A mathematical term like f (x, g(y)) becomes App
f [Var x, App g [Var y]], where f, g, x, y are suitable values, e.g. numbers
or strings.
What complicates the definition of term is the nested occurrence of term
inside list on the right-hand side. In principle, nested recursion can be elim-
inated in favour of mutual recursion by unfolding the offending datatypes,
here list. The result for term would be something like
42 3. More Functional Programming

datatype ('v,'f)"term" = Var 'v | App 'f "('v,'f)term_list"

and ('v,'f)term_list = Nil | Cons "('v,'f)term" "('v,'f)term_list"
Although we do not recommend this unfolding to the user, it shows how to
simulate nested recursion by mutual recursion. Now we return to the initial
definition of term using nested recursion.
Let us define a substitution function on terms. Because terms involve term
lists, we need to define two substitution functions simultaneously:
primrec
subst :: "('v⇒('v,'f)term) ⇒ ('v,'f)term ⇒ ('v,'f)term" and
substs:: "('v⇒('v,'f)term) ⇒ ('v,'f)term list ⇒ ('v,'f)term list"
where
"subst s (Var x) = s x" |
subst_App:
"subst s (App f ts) = App f (substs s ts)" |

"substs s [] = []" |
"substs s (t # ts) = subst s t # substs s ts"

Individual equations in a primrec definition may be named as shown for

subst_App. The significance of this device will become apparent below.
Similarly, when proving a statement about terms inductively, we need
to prove a related statement about term lists simultaneously. For example,
the fact that the identity substitution does not change a term needs to be
strengthened and proved as follows:
lemma subst_id: "subst Var t = (t ::('v,'f)term) ∧
substs Var ts = (ts::('v,'f)term list)"
apply(induct_tac t and ts rule: subst.induct substs.induct, simp_all)
done

Note that Var is the identity substitution because by definition it leaves vari-
ables unchanged: subst Var (Var x) = Var x. Note also that the type anno-
tations are necessary because otherwise there is nothing in the goal to enforce
that both halves of the goal talk about the same type parameters ('v,'f).
As a result, induction would fail because the two halves of the goal would be
unrelated.

Exercise 3.4.2 The fact that substitution distributes over composition can
be expressed roughly as follows:
subst (f ◦ g) t = subst f (subst g t)

Correct this statement (you will find that it does not type-check), strengthen
it, and prove it. (Note: ◦ is function composition; its definition is found in
theorem o_def ).
Exercise 3.4.3 Define a function trev of type ('v, 'f) Nested.term ⇒
('v, 'f) Nested.term that recursively reverses the order of arguments of all
function symbols in a term. Prove that trev (trev t) = t.
 3.4 Advanced Datatypes 43

The experienced functional programmer may feel that our definition of

subst is too complicated in that substs is unnecessary. The App -case can be
defined directly as
subst s (App f ts) = App f (map (subst s) ts)
where map is the standard list function such that map f [x1,...,xn] = [f
x1,...,f xn]. This is true, but Isabelle insists on the conjunctive format.
Fortunately, we can easily prove that the suggested equation holds:

lemma [simp]: "subst s (App f ts) = App f (map (subst s) ts)"

apply(induct_tac ts, simp_all)
done

What is more, we can now disable the old defining equation as a simplification
rule:
declare subst_App [simp del]

The advantage is that now we have replaced substs by map, we can profit from
the large number of pre-proved lemmas about map. Unfortunately, inductive
proofs about type term are still awkward because they expect a conjunc-
tion. One could derive a new induction principle as well (see Sect. 9.2.3), but
simpler is to stop using primrec and to define functions with fun instead.
Simple uses of fun are described in Sect. 3.5 below. Advanced applications,
including functions over nested datatypes like term, are discussed in a sepa-
rate tutorial [17].
Of course, you may also combine mutual and nested recursion of data-
types. For example, constructor Sum in Sect. 3.4.1 could take a list of expres-
sions as its argument: Sum "'a aexp list".

3.4.3 The Limits of Nested Recursion

How far can we push nested recursion? By the unfolding argument above,
we can reduce nested to mutual recursion provided the nested recursion only
involves previously defined datatypes. This does not include functions:
datatype t = C "t ⇒ bool"

This declaration is a real can of worms. In HOL it must be ruled out because
it requires a type t such that t and its power set t ⇒ bool have the same
cardinality — an impossibility. For the same reason it is not possible to allow
recursion involving the type t set, which is isomorphic to t ⇒ bool.
Fortunately, a limited form of recursion involving function spaces is per-
mitted: the recursive type may occur on the right of a function arrow, but
never on the left. Hence the above can of worms is ruled out but the following
example of a potentially infinitely branching tree is accepted:
datatype (dead 'a,'i) bigtree = Tip | Br 'a "'i ⇒ ('a,'i)bigtree"
44 3. More Functional Programming

Parameter 'a is the type of values stored in the Br anches of the tree, whereas
'i is the index type over which the tree branches. If 'i is instantiated to bool,
the result is a binary tree; if it is instantiated to nat, we have an infinitely
branching tree because each node has as many subtrees as there are natural
numbers. How can we possibly write down such a tree? Using functional
notation! For example, the term
Br 0 (λi. Br i (λn. Tip))
of type (nat, nat) bigtree is the tree whose root is labeled with 0 and whose
ith subtree is labeled with i and has merely Tip s as further subtrees.
Function map_bt applies a function to all labels in a bigtree:
primrec map_bt :: "('a ⇒ 'b) ⇒ ('a,'i)bigtree ⇒ ('b,'i)bigtree"
where
"map_bt f Tip = Tip" |
"map_bt f (Br a F) = Br (f a) (λi. map_bt f (F i))"
This is a valid primrec definition because the recursive calls of map_bt involve
only subtrees of F, which is itself a subterm of the left-hand side. Thus termi-
nation is assured. The seasoned functional programmer might try expressing
λi. map_bt f (F i) as map_bt f ◦ F, which Isabelle however will reject. Ap-
plying map_bt to only one of its arguments makes the termination proof less
obvious.
The following lemma has a simple proof by induction:
lemma "map_bt (g o f) T = map_bt g (map_bt f T)"
apply(induct_tac T, simp_all)
done
Because of the function type, the proof state after induction looks unusual.
Notice the quantified induction hypothesis:
1. V
map_bt (g ◦ f) Tip = map_bt g (map_bt f Tip)
2. x1 V F.
( x2a. x2a ∈ range F =⇒
map_bt (g ◦ f) x2a = map_bt g (map_bt f x2a)) =⇒
map_bt (g ◦ f) (Br x1 F) = map_bt g (map_bt f (Br x1 F))

If you need nested recursion on the left of a function arrow, there are
alternatives to pure HOL. In the Logic for Computable Functions (LCF),
types like
datatype lam = C "lam → lam"
do indeed make sense [27]. Note the different arrow, → instead of ⇒, express-
ing the type of continuous functions. There is even a version of LCF on top
of HOL, called HOLCF [20].

3.4.4 Case Study: Tries

Tries are a classic search tree data structure [16] for fast indexing with strings.
Figure 3.1 gives a graphical example of a trie containing the words “all”, “an”,
 3.4 Advanced Datatypes 45

“ape”, “can”, “car” and “cat”. When searching a string in a trie, the letters of
the string are examined sequentially. Each letter determines which subtrie to
search next. In this case study we model tries as a datatype, define a lookup
and an update function, and prove that they behave as expected.

Q
 Q
a c
 QQ
 Q
 QQ
l n p a
 Q
 QQ
l e n r t

Fig. 3.1. A Sample Trie

Proper tries associate some value with each string. Since the information
is stored only in the final node associated with the string, many nodes do not
carry any value. This distinction is modeled with the help of the predefined
datatype option (see Sect. 2.6.3).
To minimize running time, each node of a trie should contain an array that
maps letters to subtries. We have chosen a representation where the subtries
are held in an association list, i.e. a list of (letter,trie) pairs. Abstracting over
the alphabet 'a and the values 'v we define a trie as follows:
datatype ('a,'v)trie = Trie "'v option" "('a * ('a,'v)trie)list"

The first component is the optional value, the second component the associa-
tion list of subtries. This is an example of nested recursion involving products,
which is fine because products are datatypes as well. We define two selector
functions:
primrec "value" :: "('a,'v)trie ⇒ 'v option" where
"value(Trie ov al) = ov"
primrec alist :: "('a,'v)trie ⇒ ('a * ('a,'v)trie)list" where
"alist(Trie ov al) = al"
Association lists come with a generic lookup function. Its result involves type
option because a lookup can fail:
primrec assoc :: "('key * 'val)list ⇒ 'key ⇒ 'val option" where
"assoc [] x = None" |
"assoc (p#ps) x =
(let (a,b) = p in if a=x then Some b else assoc ps x)"

Now we can define the lookup function for tries. It descends into the trie
examining the letters of the search string one by one. As recursion on lists is
simpler than on tries, let us express this as primitive recursion on the search
string argument:
46 3. More Functional Programming

primrec lookup :: "('a,'v)trie ⇒ 'a list ⇒ 'v option" where

"lookup t [] = value t" |
"lookup t (a#as) = (case assoc (alist t) a of
None ⇒ None
| Some at ⇒ lookup at as)"
As a first simple property we prove that looking up a string in the empty
trie Trie None [] always returns None. The proof merely distinguishes the
two cases whether the search string is empty or not:
lemma [simp]: "lookup (Trie None []) as = None"
apply(case_tac as, simp_all)
done

Things begin to get interesting with the definition of an update func-

tion that adds a new (string, value) pair to a trie, overwriting the old value
associated with that string:
primrec update:: "('a,'v)trie ⇒ 'a list ⇒ 'v ⇒ ('a,'v)trie" where
"update t [] v = Trie (Some v) (alist t)" |
"update t (a#as) v =
(let tt = (case assoc (alist t) a of
None ⇒ Trie None [] | Some at ⇒ at)
in Trie (value t) ((a,update tt as v) # alist t))"
The base case is obvious. In the recursive case the subtrie tt associated
with the first letter a is extracted, recursively updated, and then placed in
front of the association list. The old subtrie associated with a is still in the
association list but no longer accessible via assoc. Clearly, there is room here
for optimizations!
Before we start on any proofs about update we tell the simplifier to expand
all lets and to split all case-constructs over options:
declare Let_def[simp] option.split[split]
The reason becomes clear when looking (probably after a failed proof at-
tempt) at the body of update: it contains both let and a case distinction
over type option.
Our main goal is to prove the correct interaction of update and lookup :
theorem "∀ t v bs. lookup (update t as v) bs =
(if as=bs then Some v else lookup t bs)"

Our plan is to induct on as; hence the remaining variables are quantified.
From the definitions it is clear that induction on either as or bs is required.
The choice of as is guided by the intuition that simplification of lookup might
be easier if update has already been simplified, which can only happen if as
is instantiated. The start of the proof is conventional:
apply(induct_tac as, auto)

Unfortunately, this time we are left with three intimidating looking subgoals:
 3.5 Total Recursive Functions: fun 47

1. … =⇒ lookup … bs = lookup t bs
2. … =⇒ lookup … bs = lookup t bs
3. … =⇒ lookup … bs = lookup t bs

Clearly, if we want to make headway we have to instantiate bs as well now.

It turns out that instead of induction, case distinction suffices:
apply(case_tac[!] bs, auto)
done

All methods ending in tac take an optional first argument that specifies the
range of subgoals they are applied to, where [!] means all subgoals, i.e. [1-3]
in our case. Individual subgoal numbers, e.g. [2] are also allowed.
This proof may look surprisingly straightforward. However, note that this
comes at a cost: the proof script is unreadable because the intermediate
proof states are invisible, and we rely on the (possibly brittle) magic of auto
(simp_all will not do — try it) to split the subgoals of the induction up in
such a way that case distinction on bs makes sense and solves the proof.
Exercise 3.4.4 Modify update (and its type) such that it allows both in-
sertion and deletion of entries with a single function. Prove the correspond-
ing version of the main theorem above. Optimize your function such that it
shrinks tries after deletion if possible.
Exercise 3.4.5 Write an improved version of update that does not suffer
from the space leak (pointed out above) caused by not deleting overwritten
entries from the association list. Prove the main theorem for your improved
update.

Exercise 3.4.6 Conceptually, each node contains a mapping from letters to

optional subtries. Above we have implemented this by means of an association
list. Replay the development replacing ('a × ('a, 'v) trie) list with 'a
⇒ ('a, 'v) trie option.

3.5 Total Recursive Functions: fun

Although many total functions have a natural primitive recursive definition,

this is not always the case. Arbitrary total recursive functions can be defined
by means of fun: you can use full pattern matching, recursion need not involve
datatypes, and termination is proved by showing that the arguments of all
recursive calls are smaller in a suitable sense. In this section we restrict our-
selves to functions where Isabelle can prove termination automatically. More
advanced function definitions, including user supplied termination proofs,
nested recursion and partiality, are discussed in a separate tutorial [17].
48 3. More Functional Programming

3.5.1 Definition

Here is a simple example, the Fibonacci function:

fun fib :: "nat ⇒ nat" where
"fib 0 = 0" |
"fib (Suc 0) = 1" |
"fib (Suc(Suc x)) = fib x + fib (Suc x)"

This resembles ordinary functional programming languages. Note the oblig-

atory where and |. Command fun declares and defines the function in one
go. Isabelle establishes termination automatically because fib ’s argument
decreases in every recursive call.
Slightly more interesting is the insertion of a fixed element between any
two elements of a list:
fun sep :: "'a ⇒ 'a list ⇒ 'a list" where
"sep a [] = []" |
"sep a [x] = [x]" |
"sep a (x#y#zs) = x # a # sep a (y#zs)"
This time the length of the list decreases with the recursive call; the first
argument is irrelevant for termination.
Pattern matching need not be exhaustive and may employ wildcards:
fun last :: "'a list ⇒ 'a" where
"last [x] = x" |
"last (_#y#zs) = last (y#zs)"

Overlapping patterns are disambiguated by taking the order of equations

into account, just as in functional programming:
fun sep1 :: "'a ⇒ 'a list ⇒ 'a list" where
"sep1 a (x#y#zs) = x # a # sep1 a (y#zs)" |
"sep1 _ xs = xs"
To guarantee that the second equation can only be applied if the first one
does not match, Isabelle internally replaces the second equation by the two
possibilities that are left: sep1 a [] = [] and sep1 a [x] = [x]. Thus the
functions sep and sep1 are identical.
Because of its pattern matching syntax, fun is also useful for the definition
of non-recursive functions:
fun swap12 :: "'a list ⇒ 'a list" where
"swap12 (x#y#zs) = y#x#zs" |
"swap12 zs = zs"

After a function f has been defined via fun, its defining equations (or vari-
ants derived from them) are available under the name f .simps as theorems.
For example, look (via thm) at sep.simps and sep1.simps to see that they
define the same function. What is more, those equations are automatically
declared as simplification rules.
 3.5 Total Recursive Functions: fun 49

3.5.2 Termination

Isabelle’s automatic termination prover for fun has a fixed notion of the size
(of type nat) of an argument. The size of a natural number is the number
itself. The size of a list is its length. For the general case see Sect. 2.5.2.
A recursive function is accepted if fun can show that the size of one fixed
argument becomes smaller with each recursive call.
More generally, fun allows any lexicographic combination of size measures
in case there are multiple arguments. For example, the following version of
Ackermann’s function is accepted:
fun ack2 :: "nat ⇒ nat ⇒ nat" where
"ack2 n 0 = Suc n" |
"ack2 0 (Suc m) = ack2 (Suc 0) m" |
"ack2 (Suc n) (Suc m) = ack2 (ack2 n (Suc m)) m"

The order of arguments has no influence on whether fun can prove ter-
mination of a function. For more details see elsewhere [6].

3.5.3 Simplification

Upon a successful termination proof, the recursion equations become simpli-

fication rules, just as with primrec. In most cases this works fine, but there
is a subtle problem that must be mentioned: simplification may not terminate
because of automatic splitting of if. Let us look at an example:
fun gcd :: "nat ⇒ nat ⇒ nat" where
"gcd m n = (if n=0 then m else gcd n (m mod n))"

The second argument decreases with each recursive call. The termination
condition
n 6= 0 =⇒ m mod n < n

is proved automatically because it is already present as a lemma in HOL. Thus

the recursion equation becomes a simplification rule. Of course the equation
is nonterminating if we are allowed to unfold the recursive call inside the
else branch, which is why programming languages and our simplifier don’t
do that. Unfortunately the simplifier does something else that leads to the
same problem: it splits each if -expression unless its condition simplifies to
True or False. For example, simplification reduces
fun0.gcd m n = k

in one step to
(if n = 0 then m else fun0.gcd n (m mod n)) = k

where the condition cannot be reduced further, and splitting leads to

(n = 0 −→ m = k) ∧ (n 6= 0 −→ fun0.gcd n (m mod n) = k)
50 3. More Functional Programming

Since the recursive call fun0.gcd n (m mod n) is no longer protected by an if,

it is unfolded again, which leads to an infinite chain of simplification steps.
Fortunately, this problem can be avoided in many different ways.
The most radical solution is to disable the offending theorem if_split,
as shown in Sect. 3.1.9. However, we do not recommend this approach: you
will often have to invoke the rule explicitly when if is involved.
If possible, the definition should be given by pattern matching on the left
rather than if on the right. In the case of fun0.gcd the following alternative
definition suggests itself:
fun gcd1 :: "nat ⇒ nat ⇒ nat" where
"gcd1 m 0 = m" |
"gcd1 m n = gcd1 n (m mod n)"
The order of equations is important: it hides the side condition n 6= 0. Un-
fortunately, not all conditionals can be expressed by pattern matching.
A simple alternative is to replace if by case, which is also available for
bool and is not split automatically:
fun gcd2 :: "nat ⇒ nat ⇒ nat" where
"gcd2 m n = (case n=0 of True ⇒ m | False ⇒ gcd2 n (m mod n))"
This is probably the neatest solution next to pattern matching, and it is
always available.
A final alternative is to replace the offending simplification rules by de-
rived conditional ones. For fun0.gcd it means we have to prove these lemmas:
lemma [simp]: "gcd m 0 = m"
apply(simp)
done

lemma [simp]: "n 6= 0 =⇒ gcd m n = gcd n (m mod n)"

apply(simp)
done
Simplification terminates for these proofs because the condition of the if
simplifies to True or False. Now we can disable the original simplification
rule:
declare gcd.simps [simp del]

3.5.4 Induction

Having defined a function we might like to prove something about it. Since the
function is recursive, the natural proof principle is again induction. But this
time the structural form of induction that comes with datatypes is unlikely
to work well — otherwise we could have defined the function by primrec.
Therefore fun automatically proves a suitable induction rule f .induct that
follows the recursion pattern of the particular function f . We call this re-
cursion induction. Roughly speaking, it requires you to prove for each fun
 3.5 Total Recursive Functions: fun 51

equation that the property you are trying to establish holds for the left-hand
side provided it holds for all recursive calls on the right-hand side. Here is a
simple example involving the predefined map functional on lists:
lemma "map f (sep x xs) = sep (f x) (map f xs)"
Note that map f xs is the result of applying f to all elements of xs. We prove
this lemma by recursion induction over sep :
apply(induct_tac x xs rule: sep.induct)
The resulting proof state has three subgoals corresponding to the three
clauses for sep :
1. Va. map f (sep a []) = sep (f a) (map f [])
V
2. Va x. map f (sep a [x]) = sep (f a) (map f [x])
3. a x y zs.
map f (sep a (y # zs)) = sep (f a) (map f (y # zs)) =⇒
map f (sep a (x # y # zs)) = sep (f a) (map f (x # y # zs))

The rest is pure simplification:

apply simp_all
done

The proof goes smoothly because the induction rule follows the recursion of
sep. Try proving the above lemma by structural induction, and you find that
you need an additional case distinction.
In general, the format of invoking recursion induction is
apply(induct_tac x1 . . . xn rule: f .induct)
where x1 . . . xn is a list of free variables in the subgoal and f the name of a
function that takes n arguments. Usually the subgoal will contain the term
fx1 . . . xn but this need not be the case. The induction rules do not mention
f at all. Here is sep.induct:
[[ Va. P a [];
V

Va x. P a [x];
a x y zs. P a (y # zs) =⇒ P a (x # y # zs)]]
=⇒ P u v

It merely says that in order to prove a property P of u and v you need to prove
it for the three cases where v is the empty list, the singleton list, and the list
with at least two elements. The final case has an induction hypothesis: you
may assume that P holds for the tail of that list.
4. Presenting Theories

By now the reader should have become sufficiently acquainted with elemen-
tary theory development in Isabelle/HOL. The following interlude describes
how to present theories in a typographically pleasing manner. Isabelle pro-
vides a rich infrastructure for concrete syntax of the underlying λ-calculus
language (see Sect. 4.1), as well as document preparation of theory texts based
on existing PDF-LATEX technology (see Sect. 4.2).
As pointed out by Leibniz more than 300 years ago, notions are in princi-
ple more important than notations, but suggestive textual representation of
ideas is vital to reduce the mental effort to comprehend and apply them.

4.1 Concrete Syntax

The core concept of Isabelle’s framework for concrete syntax is that of mixfix
annotations. Associated with any kind of constant declaration, mixfixes
affect both the grammar productions for the parser and output templates for
the pretty printer.
In full generality, parser and pretty printer configuration is a subtle af-
fair [37]. Your syntax specifications need to interact properly with the existing
setup of Isabelle/Pure and Isabelle/HOL. To avoid creating ambiguities with
existing elements, it is particularly important to give new syntactic constructs
the right precedence.
Below we introduce a few simple syntax declaration forms that already
cover many common situations fairly well.

4.1.1 Infix Annotations

Syntax annotations may be included wherever constants are declared, such as

definition and primrec — and also datatype, which declares constructor
operations. Type-constructors may be annotated as well, although this is less
frequently encountered in practice (the infix type × comes to mind).
Infix declarations provide a useful special case of mixfixes. The following
example of the exclusive-or operation on boolean values illustrates typical
infix declarations.
54 4. Presenting Theories

definition xor :: "bool ⇒ bool ⇒ bool" (infixl "[+]" 60)

where "A [+] B ≡ (A ∧ ¬ B) ∨ (¬ A ∧ B)"
Now xor A B and A [+] B refer to the same expression internally. Any curried
function with at least two arguments may be given infix syntax. For partial
applications with fewer than two operands, the operator is enclosed in paren-
theses. For instance, xor without arguments is represented as ([+]); together
with ordinary function application, this turns xor A into ([+]) A.
The keyword infixl seen above specifies an infix operator that is nested
to the left: in iterated applications the more complex expression appears on
the left-hand side, and A [+] B [+] C stands for (A [+] B) [+] C. Similarly,
infixr means nesting to the right, reading A [+] B [+] C as A [+] (B [+] C).
A non-oriented declaration via infix would render A [+] B [+] C illegal, but
demand explicit parentheses to indicate the intended grouping.
The string [+] in our annotation refers to the concrete syntax to repre-
sent the operator (a literal token), while the number 60 determines the prece-
dence of the construct: the syntactic priorities of the arguments and result.
Isabelle/HOL already uses up many popular combinations of ASCII symbols
for its own use, including both + and ++. Longer character combinations are
more likely to be still available for user extensions, such as our [+].
Operator precedences have a range of 0–1000. Very low or high priorities
are reserved for the meta-logic. HOL syntax mainly uses the range of 10–100:
the equality infix = is centered at 50; logical connectives (like ∨ and ∧) are
below 50; algebraic ones (like + and *) are above 50. User syntax should strive
to coexist with common HOL forms, or use the mostly unused range 100–900.

4.1.2 Mathematical Symbols

Concrete syntax based on ASCII characters has inherent limitations. Math-

ematical notation demands a larger repertoire of glyphs. Several standards
of extended character sets have been proposed over decades, but none has
become universally available so far. Isabelle has its own notion of symbols
as the smallest entities of source text, without referring to internal encodings.
There are three kinds of such “generalized characters”:
1. 7-bit ASCII characters
2. named symbols: \<ident>
3. named control symbols: \<^ident>
Here ident is any sequence of letters. This results in an infinite store of
symbols, whose interpretation is left to further front-end tools. For example,
the Isabelle document processor (see Sect. 4.2) display the \<forall> symbol
as ∀ .
A list of standard Isabelle symbols is given in [37]. You may introduce
your own interpretation of further symbols by configuring the appropriate
front-end tool accordingly, e.g. by defining certain LATEX macros (see also
 4.1 Concrete Syntax 55

Sect. 4.2.4). There are also a few predefined control symbols, such as \<^sub>
and \<^sup> for sub- and superscript of the subsequent printable symbol,
respectively. For example, A\<^sup>\<star>, is output as A? .
A number of symbols are considered letters by the Isabelle lexer and can
be used as part of identifiers. These are the greek letters α (\<alpha>), β
(\<beta>), etc. (excluding λ), special letters like A (\<A>) and A (\<AA>).
Moreover the control symbol \<^sub> may be used to subscript a single letter
or digit in the trailing part of an identifier. This means that the input
\<forall>\<alpha>\<^sub>1. \<alpha>\<^sub>1 = \<Pi>\<^sub>\<A>

is recognized as the term ∀ α1 . α1 = Π A by Isabelle.

Replacing our previous definition of xor by the following specifies an Isa-
belle symbol for the new operator:
definition xor :: "bool ⇒ bool ⇒ bool" (infixl "⊕" 60)
where "A ⊕ B ≡ (A ∧ ¬ B) ∨ (¬ A ∧ B)"

It is possible to provide alternative syntax forms through the print mode

concept [37]. By convention, the mode of “xsymbols” is enabled whenever
Proof General’s X-Symbol mode or LATEX output is active. Now consider the
following hybrid declaration of xor :
definition xor :: "bool ⇒ bool ⇒ bool" (infixl "[+]" 60)
where "A [+] B ≡ (A ∧ ¬ B) ∨ (¬ A ∧ B)"

notation (xsymbols) xor (infixl "⊕" 60)

The notation command associates a mixfix annotation with a known con-

stant. The print mode specification, here (xsymbols), is optional.
We may now write A [+] B or A ⊕ B in input, while output uses the nicer
syntax of xsymbols whenever that print mode is active. Such an arrange-
ment is particularly useful for interactive development, where users may type
ASCII text and see mathematical symbols displayed during proofs.

4.1.3 Prefix Annotations

Prefix syntax annotations are another form of mixfixes [37], without any
template arguments or priorities — just some literal syntax. The following
example associates common symbols with the constructors of a datatype.
datatype currency =
Euro nat ("e")
| Pounds nat ("£")
| Yen nat ("U")
| Dollar nat ("$")

Here the mixfix annotations on the rightmost column happen to consist of a

single Isabelle symbol each: \<euro>, \<pounds>, \<yen>, and $. Recall that
a constructor like Euro actually is a function nat ⇒ currency. The expression
56 4. Presenting Theories

Euro 10 will be printed as e 10 ; only the head of the application is subject to

our concrete syntax. This rather simple form already achieves conformance
with notational standards of the European Commission.
Prefix syntax works the same way for other commands that introduce new
constants, e.g. primrec.

4.1.4 Abbreviations

Mixfix syntax annotations merely decorate particular constant application

forms with concrete syntax, for instance replacing xor A B by A ⊕ B. Occa-
sionally, the relationship between some piece of notation and its internal form
is more complicated. Here we need abbreviations.
Command abbreviation introduces an uninterpreted notational constant
as an abbreviation for a complex term. Abbreviations are unfolded upon
parsing and re-introduced upon printing. This provides a simple mechanism
for syntactic macros.
A typical use of abbreviations is to introduce relational notation for mem-
bership in a set of pairs, replacing (x, y) ∈ sim by x ≈ y. We assume that
a constant sim of type ('a × 'a) set has been introduced at this point.
abbreviation sim2 :: "'a ⇒ 'a ⇒ bool" (infix "≈" 50)
where "x ≈ y ≡ (x, y) ∈ sim"

The given meta-equality is used as a rewrite rule after parsing (replacing

x ≈ y by (x,y) ∈ sim) and before printing (turning (x,y) ∈ sim back into
x ≈ y ). The name of the dummy constant sim2 does not matter, as long as
it is unique.
Another common application of abbreviations is to provide variant ver-
sions of fundamental relational expressions, such as 6= for negated equalities.
The following declaration stems from Isabelle/HOL itself:
abbreviation not_equal :: "'a ⇒ 'a ⇒ bool" (infixl "~=" 50)
where "x ~= y ≡ ¬ (x = y)"

notation (xsymbols) not_equal (infix "6=" 50)

The notation 6= is introduced separately to restrict it to the xsymbols mode.

Abbreviations are appropriate when the defined concept is a simple vari-
ation on an existing one. But because of the automatic folding and unfolding
of abbreviations, they do not scale up well to large hierarchies of concepts.
Abbreviations do not replace definitions.
Abbreviations are a simplified form of the general concept of syntax trans-
lations; even heavier transformations may be written in ML [37].
 4.2 Document Preparation 57

4.2 Document Preparation

Isabelle/Isar is centered around the concept of formal proof documents.

The outcome of a formal development effort is meant to be a human-readable
record, presented as browsable PDF file or printed on paper. The overall
document structure follows traditional mathematical articles, with sections,
intermediate explanations, definitions, theorems and proofs.
The Isabelle document preparation system essentially acts as a front-
end to LATEX. After checking specifications and proofs formally, the theory
sources are turned into typesetting instructions in a schematic manner. This
lets you write authentic reports on theory developments with little effort:
many technical consistency checks are handled by the system.
Here is an example to illustrate the idea of Isabelle document preparation.

The following datatype definition of 'a bintree models binary

trees with nodes being decorated by elements of type 'a.
datatype 'a bintree =
Leaf | Branch 'a "'a bintree" "'a bintree"

The datatype induction rule generated here is of the form

[[P Leaf; x1 x2 x3. [[P x2; P x3]] =⇒ P (Branch x1 x2 x3)]] =⇒ P
V
bintree

The above document output has been produced as follows:

text {*
The following datatype definition of @{text "'a bintree"}
models binary trees with nodes being decorated by elements
of type @{typ 'a}.
*}

datatype 'a bintree =

Leaf | Branch 'a "'a bintree" "'a bintree"

text {*
\noindent The datatype induction rule generated here is
of the form @{thm [display] bintree.induct [no_vars]}
*}

Here we have augmented the theory by formal comments (using text blocks),
the informal parts may again refer to formal entities by means of “antiquota-
tions” (such as @{text "'a bintree"} or @{typ 'a}), see also Sect. 4.2.3.
58 4. Presenting Theories

4.2.1 Isabelle Sessions

In contrast to the highly interactive mode of Isabelle/Isar theory develop-

ment, the document preparation stage essentially works in batch-mode. An
Isabelle session consists of a collection of source files that may contribute to
an output document. Each session is derived from a single parent, usually an
object-logic image like HOL. This results in an overall tree structure, which
is reflected by the output location in the file system (the root directory is
determined by the Isabelle settings variable ISABELLE_BROWSER_INFO).
The easiest way to manage Isabelle sessions is via isabelle mkroot (to
generate an initial session source setup) and isabelle build (to run sessions
as specified in the corresponding ROOT file). These Isabelle tools are described
in further detail in the Isabelle System Manual [36].
For example, a new session MySession (with document preparation) may
be produced as follows:

isabelle mkroot MySession

isabelle build -D MySession
The isabelle build job also informs about the file-system location of
the ultimate results. The above dry run should be able to produce some
document.pdf (with dummy title, empty table of contents etc.). Any failure
at this stage usually indicates technical problems of the LATEX installation.
The detailed arrangement of the session sources is as follows.

– Directory MySession holds the required theory files T1 .thy, …, Tn .thy.

– File MySession/ROOT specifies the session options and content, with decla-
rations for all wanted theories; it is sufficient to specify the terminal nodes
of the theory dependency graph.
– Directory MySession/document contains everything required for the LATEX
stage; only root.tex needs to be provided initially.
The latter file holds appropriate LATEX code to commence a document
(\documentclass etc.), and to include the generated files Ti .tex for each
theory. Isabelle will generate a file session.tex holding LATEX commands
to include all generated theory output files in topologically sorted order, so
\input{session} in the body of root.tex does the job in most situations.

One may now start to populate the directory MySession and its ROOT file
accordingly. The file MySession/document/root.tex should also be adapted
at some point; the default version is mostly self-explanatory. Note that
\isabellestyle enables fine-tuning of the general appearance of characters
and mathematical symbols (see also Sect. 4.2.4).
Especially observe the included LATEX packages isabelle (mandatory),
isabellesym (required for mathematical symbols), and the final pdfsetup
(provides sane defaults for hyperref, including URL markup). All three are
 4.2 Document Preparation 59

distributed with Isabelle. Further packages may be required in particular

applications, say for unusual mathematical symbols.
Any additional files for the LATEX stage go into the MySession/document
directory as well. In particular, adding a file named root.bib causes an au-
tomatic run of bibtex to process a bibliographic database; see also isabelle
document [36].
Any failure of the document preparation phase in an Isabelle batch ses-
sion leaves the generated sources in their target location, identified by the
accompanying error message. This lets you trace LATEX problems with the
generated files at hand.

4.2.2 Structure Markup

The large-scale structure of Isabelle documents follows existing LATEX conven-

tions, with chapters, sections, subsubsections etc. The Isar language includes
separate markup commands, which do not affect the formal meaning of a
theory (or proof), but result in corresponding LATEX elements.
From the Isabelle perspective, each markup command takes a single text
argument (delimited by " . . . " or {* . . . *}). After stripping any surrounding
white space, the argument is passed to a LATEX macro \isamarkupXYZ for
command XYZ. These macros are defined in isabelle.sty according to
the meaning given in the rightmost column above.
The following source fragment illustrates structure markup of a theory.
Note that LATEX labels may be included inside of section headings as well.
60 4. Presenting Theories

section {* Some properties of Foo Bar elements *}

theory Foo_Bar
imports Main
begin

subsection {* Basic definitions *}

definition foo :: …

definition bar :: …

subsection {* Derived rules *}

lemma fooI: …
lemma fooE: …

subsection {* Main theorem \label{sec:main-theorem} *}

theorem main: …

end

4.2.3 Formal Comments and Antiquotations

Isabelle source comments, which are of the form (* . . . *), essentially act
like white space and do not really contribute to the content. They mainly
serve technical purposes to mark certain oddities in the raw input text. In
contrast, formal comments are portions of text that are associated with
formal Isabelle/Isar commands (marginal comments), or as standalone
paragraphs within a theory or proof context (text blocks).
Marginal comments are part of each command’s concrete syntax [37]; the
common form is “-- text” where text is delimited by ". . . " or {* . . . *} as
before. Multiple marginal comments may be given at the same time. Here is
a simple example:
lemma "A --> A"
— a triviality of propositional logic
— (should not really bother)
by (rule impI) — implicit assumption step involved here

The above output has been produced as follows:

lemma "A --> A"
-- "a triviality of propositional logic"
-- "(should not really bother)"
by (rule impI) -- "implicit assumption step involved here"
 4.2 Document Preparation 61

From the LATEX viewpoint, “--” acts like a markup command, associated
with the macro \isamarkupcmt (taking a single argument).
Text blocks are introduced by the commands text and txt. Each takes
again a single text argument, which is interpreted as a free-form paragraph
in LATEX (surrounded by some additional vertical space). The typesetting
may be changed by redefining the LATEX environments of isamarkuptext or
isamarkuptxt, respectively (via \renewenvironment).
The text part of Isabelle markup commands essentially inserts quoted ma-
terial into a formal text, mainly for instruction of the reader. An antiquo-
tation is again a formal object embedded into such an informal portion. The
interpretation of antiquotations is limited to some well-formedness checks,
with the result being pretty printed to the resulting document. Quoted text
blocks together with antiquotations provide an attractive means of referring
to formal entities, with good confidence in getting the technical details right
(especially syntax and types).
The general syntax of antiquotations is as follows: @{name arguments},
or @{name [options] arguments} for a comma-separated list of options con-
sisting of a name or name=value each. The syntax of arguments depends on
the kind of antiquotation, it generally follows the same conventions for types,
terms, or theorems as in the formal part of a theory.
This sentence demonstrates quotations and antiquotations: λx y. x is a
well-typed term.
The output above was produced as follows:
text {*
This sentence demonstrates quotations and antiquotations:
@{term "%x y. x"} is a well-typed term.
*}

The notational change from the ASCII character % to the symbol λ reveals
that Isabelle printed this term, after parsing and type-checking. Document
preparation enables symbolic output by default.
The next example includes an option to show the type of all variables.
The antiquotation @{term [show_types] "%x y. x"} produces the output
λ(x::'a) y::'b. x. Type inference has figured out the most general typings
in the present theory context. Terms may acquire different typings due to con-
straints imposed by their environment; within a proof, for example, variables
are given the same types as they have in the main goal statement.
Several further kinds of antiquotations and options are available [37]. Here
are a few commonly used combinations:
62 4. Presenting Theories

@{typ τ } print type τ

@{const c} check existence of c and print it
@{term t} print term t
@{prop φ} print proposition φ
@{prop [display] φ} print large proposition φ (with linebreaks)
@{prop [source] φ} check proposition φ, print its input
@{thm a} print fact a
@{thm a [no_vars]} print fact a, fixing schematic variables
@{thm [source] a} check availability of fact a, print its name
@{text s} print uninterpreted text s
Note that no_vars given above is not an antiquotation option, but an
attribute of the theorem argument given here. This might be useful with a
diagnostic command like thm, too.
The @{text s} antiquotation is particularly interesting. Embedding un-
interpreted text within an informal body might appear useless at first sight.
Here the key virtue is that the string s is processed as Isabelle output, inter-
preting Isabelle symbols appropriately.
For example, @{text "\<forall>\<exists>"} produces ∀ ∃ , according
to the standard interpretation of these symbol (cf. Sect. 4.2.4). Thus we
achieve consistent mathematical notation in both the formal and informal
parts of the document very easily, independently of the term language of Isa-
belle. Manual LATEX code would leave more control over the typesetting, but
is also slightly more tedious.

4.2.4 Interpretation of Symbols

As has been pointed out before (Sect. 4.1.2), Isabelle symbols are the smallest
syntactic entities — a straightforward generalization of ASCII characters.
While Isabelle does not impose any interpretation of the infinite collection of
named symbols, LATEX documents use canonical glyphs for certain standard
symbols [37].
The LATEX code produced from Isabelle text follows a simple scheme. You
can tune the final appearance by redefining certain macros, say in root.tex
of the document.
1. 7-bit ASCII characters: letters A…Z and a…z are output directly, digits
are passed as an argument to the \isadigit macro, other characters are
replaced by specifically named macros of the form \isacharXYZ.
2. Named symbols: \<XYZ> is turned into {\isasymXYZ}; note the additional
braces.
3. Named control symbols: \<^XYZ> is turned into \isactrlXYZ; subsequent
symbols may act as arguments if the control macro is defined accordingly.
You may occasionally wish to give new LATEX interpretations of named
symbols. This merely requires an appropriate definition of \isasymXYZ,
 4.2 Document Preparation 63

for \<XYZ> (see isabelle.sty for working examples). Control symbols are
slightly more difficult to get right, though.
The \isabellestyle macro provides a high-level interface to tune the
general appearance of individual symbols. For example, \isabellestyle{it}
uses the italics text style to mimic the general appearance of the LATEX math
mode; double quotes are not printed at all. The resulting quality of type-
setting is quite good, so this should be the default style for work that gets
distributed to a broader audience.

4.2.5 Suppressing Output

By default, Isabelle’s document system generates a LATEX file for each theory
that gets loaded while running the session. The generated session.tex will
include all of these in order of appearance, which in turn gets included by
the standard root.tex. Certainly one may change the order or suppress
unwanted theories by ignoring session.tex and load individual files directly
in root.tex. On the other hand, such an arrangement requires additional
maintenance whenever the collection of theories changes.
Alternatively, one may tune the theory loading process in ROOT itself:
some sequential order of theories sections may enforce a certain traversal
of the dependency graph, although this could degrade parallel processing.
The nodes of each sub-graph that is specified here are presented in some
topological order of their formal dependencies.
Moreover, the system build option document=false allows to disable doc-
ument generation for some theories. Its usage in the session ROOT is like this:
theories [document = false] T

Theory output may be suppressed more selectively, either via tagged

command regions or ignored material.
Tagged command regions works by annotating commands with named
tags, which correspond to certain LATEX markup that tells how to treat par-
ticular parts of a document when doing the actual type-setting. By default,
certain Isabelle/Isar commands are implicitly marked up using the predefined
tags “theory” (for theory begin and end), “proof ” (for proof commands), and
“ML” (for commands involving ML code). Users may add their own tags us-
ing the %tag notation right after a command name. In the subsequent example
we hide a particularly irrelevant proof:
lemma "x = x"

The original source has been “lemma "x = x" by %invisible (simp)”.
Tags observe the structure of proofs; adjacent commands with the same tag
are joined into a single region. The Isabelle document preparation system
allows the user to specify how to interpret a tagged region, in order to keep,
64 4. Presenting Theories

drop, or fold the corresponding parts of the document. See the Isabelle System
Manual [36] for further details, especially on isabelle build and isabelle
document.
Ignored material is specified by delimiting the original formal source with
special source comments (*<*) and (*>*). These parts are stripped before
the type-setting phase, without affecting the formal checking of the theory, of
course. For example, we may hide parts of a proof that seem unfit for general
public inspection. The following “fully automatic” proof is actually a fake:
lemma "x 6= (0::int) =⇒ 0 < x * x"
by (auto)

The real source of the proof has been as follows:

by (auto(*<*)simp add: zero_less_mult_iff(*>*))

Suppressing portions of printed text demands care. You should not mis-
represent the underlying theory development. It is easy to invalidate the
visible text by hiding references to questionable axioms, for example.
 Part II

Logic and Sets

5. The Rules of the Game

This chapter outlines the concepts and techniques that underlie reasoning
in Isabelle. Until now, we have proved everything using only induction and
simplification, but any serious verification project requires more elaborate
forms of inference. The chapter also introduces the fundamentals of predicate
logic. The first examples in this chapter will consist of detailed, low-level proof
steps. Later, we shall see how to automate such reasoning using the methods
blast, auto and others. Backward or goal-directed proof is our usual style,
but the chapter also introduces forward reasoning, where one theorem is
transformed to yield another.

5.1 Natural Deduction

In Isabelle, proofs are constructed using inference rules. The most familiar
inference rule is probably modus ponens:

P→Q P
Q

This rule says that from P → Q and P we may infer Q.

Natural deduction is an attempt to formalize logic in a way that mirrors
human reasoning patterns. For each logical symbol (say, ∧), there are two
kinds of rules: introduction and elimination rules. The introduction rules
allow us to infer this symbol (say, to infer conjunctions). The elimination
rules allow us to deduce consequences from this symbol. Ideally each rule
should mention one symbol only. For predicate logic this can be done, but
when users define their own concepts they typically have to refer to other
symbols as well. It is best not to be dogmatic. Our system is not based on
pure natural deduction, but includes elements from the sequent calculus and
free-variable tableaux.
Natural deduction generally deserves its name. It is easy to use. Each proof
step consists of identifying the outermost symbol of a formula and applying
the corresponding rule. It creates new subgoals in an obvious way from parts
of the chosen formula. Expanding the definitions of constants can blow up
the goal enormously. Deriving natural deduction rules for such constants
68 5. The Rules of the Game

lets us reason in terms of their key properties, which might otherwise be

obscured by the technicalities of its definition. Natural deduction rules also
lend themselves to automation. Isabelle’s classical reasoner accepts any
suitable collection of natural deduction rules and uses them to search for
proofs automatically. Isabelle is designed around natural deduction and many
of its tools use the terminology of introduction and elimination rules.

5.2 Introduction Rules

An introduction rule tells us when we can infer a formula containing a specific

logical symbol. For example, the conjunction introduction rule says that if
we have P and if we have Q then we have P ∧ Q. In a mathematics text, it
is typically shown like this:
P Q
P ∧Q
The rule introduces the conjunction symbol (∧) in its conclusion. In Isabelle
proofs we mainly reason backwards. When we apply this rule, the subgoal
already has the form of a conjunction; the proof step makes this conjunction
symbol disappear.
In Isabelle notation, the rule looks like this:
[[?P; ?Q]] =⇒ ?P ∧ ?Q (conjI)

Carefully examine the syntax. The premises appear to the left of the arrow
and the conclusion to the right. The premises (if more than one) are grouped
using the fat brackets. The question marks indicate schematic variables
(also called unknowns): they may be replaced by arbitrary formulas. If we
use the rule backwards, Isabelle tries to unify the current subgoal with the
conclusion of the rule, which has the form ?P ∧ ?Q. (Unification is discussed
below, Sect. 5.8.) If successful, it yields new subgoals given by the formulas
assigned to ?P and ?Q.
The following trivial proof illustrates how rules work. It also introduces
a style of indentation. If a command adds a new subgoal, then the next
command’s indentation is increased by one space; if it proves a subgoal, then
the indentation is reduced. This provides the reader with hints about the
subgoal structure.
lemma conj_rule: "[[P; Q]] =⇒ P ∧ (Q ∧ P)"
apply (rule conjI)
apply assumption
apply (rule conjI)
apply assumption
apply assumption

At the start, Isabelle presents us with the assumptions (P and Q ) and with
the goal to be proved, P ∧ (Q ∧ P). We are working backwards, so when we
 5.3 Elimination Rules 69

apply conjunction introduction, the rule removes the outermost occurrence

of the ∧ symbol. To apply a rule to a subgoal, we apply the proof method
rule — here with conjI, the conjunction introduction rule.
1. [[P; Q]] =⇒ P
2. [[P; Q]] =⇒ Q ∧ P

Isabelle leaves two new subgoals: the two halves of the original conjunction.
The first is simply P, which is trivial, since P is among the assumptions.
We can apply the assumption method, which proves a subgoal by finding a
matching assumption.
1. [[P; Q]] =⇒ Q ∧ P
We are left with the subgoal of proving Q ∧ P from the assumptions P and Q.
We apply rule conjI again.
1. [[P; Q]] =⇒ Q
2. [[P; Q]] =⇒ P

We are left with two new subgoals, Q and P, each of which can be proved
using the assumption method.

5.3 Elimination Rules

Elimination rules work in the opposite direction from introduction rules. In

the case of conjunction, there are two such rules. From P ∧ Q we infer P.
also, from P ∧ Q we infer Q:

P ∧Q P ∧Q
P Q

Now consider disjunction. There are two introduction rules, which resem-
ble inverted forms of the conjunction elimination rules:

P Q
P ∨Q P ∨Q

What is the disjunction elimination rule? The situation is rather different

from conjunction. From P ∨ Q we cannot conclude that P is true and we
cannot conclude that Q is true; there are no direct elimination rules of the sort
that we have seen for conjunction. Instead, there is an elimination rule that
works indirectly. If we are trying to prove something else, say R, and we know
that P ∨ Q holds, then we have to consider two cases. We can assume that
P is true and prove R and then assume that Q is true and prove R a second
time. Here we see a fundamental concept used in natural deduction: that of
the assumptions. We have to prove R twice, under different assumptions.
The assumptions are local to these subproofs and are visible nowhere else.
70 5. The Rules of the Game

In a logic text, the disjunction elimination rule might be shown like this:

[P] [Q]
.. ..
.. ..
P ∨Q R R
R

The assumptions [P] and [Q] are bracketed to emphasize that they are local
to their subproofs. In Isabelle notation, the already-familiar =⇒ syntax serves
the same purpose:
[[?P ∨ ?Q; ?P =⇒ ?R; ?Q =⇒ ?R]] =⇒ ?R (disjE)
When we use this sort of elimination rule backwards, it produces a case
split. (We have seen this before, in proofs by induction.) The following proof
illustrates the use of disjunction elimination.
lemma disj_swap: "P ∨ Q =⇒ Q ∨ P"
apply (erule disjE)
apply (rule disjI2)
apply assumption
apply (rule disjI1)
apply assumption

We assume P ∨ Q and must prove Q ∨ P . Our first step uses the disjunction
elimination rule, disjE . We invoke it using erule, a method designed to work
with elimination rules. It looks for an assumption that matches the rule’s
first premise. It deletes the matching assumption, regards the first premise as
proved and returns subgoals corresponding to the remaining premises. When
we apply erule to disjE, only two subgoals result. This is better than applying
it using rule to get three subgoals, then proving the first by assumption:
the other subgoals would have the redundant assumption P ∨ Q . Most of
the time, erule is the best way to use elimination rules, since it replaces
an assumption by its subformulas; only rarely does the original assumption
remain useful.
1. P =⇒ Q ∨ P
2. Q =⇒ Q ∨ P

These are the two subgoals returned by erule. The first assumes P and the
second assumes Q. Tackling the first subgoal, we need to show Q ∨ P . The
second introduction rule (disjI2 ) can reduce this to P, which matches the
assumption. So, we apply the rule method with disjI2 …
1. P =⇒ P
2. Q =⇒ Q ∨ P
…and finish off with the assumption method. We are left with the other sub-
goal, which assumes Q.
1. Q =⇒ Q ∨ P
 5.4 Destruction Rules: Some Examples 71

Its proof is similar, using the introduction rule disjI1.

The result of this proof is a new inference rule disj_swap, which is neither
an introduction nor an elimination rule, but which might be useful. We can
use it to replace any goal of the form Q ∨ P by one of the form P ∨ Q.

5.4 Destruction Rules: Some Examples

Now let us examine the analogous proof for conjunction.
lemma conj_swap: "P ∧ Q =⇒ Q ∧ P"
apply (rule conjI)
apply (drule conjunct2)
apply assumption
apply (drule conjunct1)
apply assumption
Recall that the conjunction elimination rules — whose Isabelle names are
conjunct1 and conjunct2 — simply return the first or second half of a con-
junction. Rules of this sort (where the conclusion is a subformula of a premise)
are called destruction rules because they take apart and destroy a premise.1
The first proof step applies conjunction introduction, leaving two sub-
goals:
1. P ∧ Q =⇒ Q
2. P ∧ Q =⇒ P
To invoke the elimination rule, we apply a new method, drule. Think of
the d as standing for destruction (or direct, if you prefer). Applying the
second conjunction rule using drule replaces the assumption P ∧ Q by Q.
1. Q =⇒ Q
2. P ∧ Q =⇒ P
The resulting subgoal can be proved by applying assumption. The other sub-
goal is similarly proved, using the conjunct1 rule and the assumption method.
Choosing among the methods rule, erule and drule is up to you. Isabelle
does not attempt to work out whether a rule is an introduction rule or an
elimination rule. The method determines how the rule will be interpreted.
Many rules can be used in more than one way. For example, disj_swap can
be applied to assumptions as well as to goals; it replaces any assumption of
the form P ∨ Q by a one of the form Q ∨ P.
Destruction rules are simpler in form than indirect rules such as disjE, but
they can be inconvenient. Each of the conjunction rules discards half of the
1
This Isabelle terminology is not used in standard logic texts, although the dis-
tinction between the two forms of elimination rule is well known. Girard [9, page
74], for example, writes “The elimination rules [for ∨ and ∃] are very bad. What
is catastrophic about them is the parasitic presence of a formula [R] which has
no structural link with the formula which is eliminated.” These Isabelle rules are
inspired by the sequent calculus.
72 5. The Rules of the Game

formula, when usually we want to take both parts of the conjunction as new
assumptions. The easiest way to do so is by using an alternative conjunction
elimination rule that resembles disjE . It is seldom, if ever, seen in logic books.
In Isabelle syntax it looks like this:
[[?P ∧ ?Q; [[?P; ?Q]] =⇒ ?R]] =⇒ ?R (conjE)

Exercise 5.4.1 Use the rule conjE to shorten the proof above.

5.5 Implication

At the start of this chapter, we saw the rule modus ponens. It is, in fact, a
destruction rule. The matching introduction rule looks like this in Isabelle:
(?P =⇒ ?Q) =⇒ ?P −→ ?Q (impI)
And this is modus ponens:
[[?P −→ ?Q; ?P]] =⇒ ?Q (mp)
Here is a proof using the implication rules. This lemma performs a sort
of uncurrying, replacing the two antecedents of a nested implication by a
conjunction. The proof illustrates how assumptions work. At each proof step,
the subgoals inherit the previous assumptions, perhaps with additions or
deletions. Rules such as impI and disjE add assumptions, while applying
erule or drule deletes the matching assumption.
lemma imp_uncurry: "P −→ (Q −→ R) =⇒ P ∧ Q −→ R"
apply (rule impI)
apply (erule conjE)
apply (drule mp)
apply assumption
apply (drule mp)
apply assumption
apply assumption

First, we state the lemma and apply implication introduction (rule impI ),
which moves the conjunction to the assumptions.
1. [[P −→ Q −→ R; P ∧ Q]] =⇒ R

Next, we apply conjunction elimination (erule conjE ), which splits this con-
junction into two parts.
1. [[P −→ Q −→ R; P; Q]] =⇒ R

Now, we work on the assumption P −→ (Q −→ R), where the parentheses

have been inserted for clarity. The nested implication requires two applica-
tions of modus ponens: drule mp. The first use yields the implication Q −→
R, but first we must prove the extra subgoal P, which we do by assumption.
 5.6 Negation 73

1. [[P; Q]] =⇒ P
2. [[P; Q; Q −→ R]] =⇒ R
Repeating these steps for Q −→ R yields the conclusion we seek, namely R.
1. [[P; Q; Q −→ R]] =⇒ R
The symbols =⇒ and −→ both stand for implication, but they differ in
many respects. Isabelle uses =⇒ to express inference rules; the symbol is
built-in and Isabelle’s inference mechanisms treat it specially. On the other
hand, −→ is just one of the many connectives available in higher-order logic.
We reason about it using inference rules such as impI and mp, just as we
reason about the other connectives. You will have to use −→ in any context
that requires a formula of higher-order logic. Use =⇒ to separate a theorem’s
preconditions from its conclusion.
The by command is useful for proofs like these that use assumption heav-
ily. It executes an apply command, then tries to prove all remaining subgoals
using assumption. Since (if successful) it ends the proof, it also replaces the
done symbol. For example, the proof above can be shortened:
lemma imp_uncurry: "P −→ (Q −→ R) =⇒ P ∧ Q −→ R"
apply (rule impI)
apply (erule conjE)
apply (drule mp)
apply assumption
by (drule mp)

We could use by to replace the final apply and done in any proof, but
typically we use it to eliminate calls to assumption. It is also a nice way of
expressing a one-line proof.

5.6 Negation

Negation causes surprising complexity in proofs. Its natural deduction rules

are straightforward, but additional rules seem necessary in order to han-
dle negated assumptions gracefully. This section also illustrates the intro
method: a convenient way of applying introduction rules.
Negation introduction deduces ¬P if assuming P leads to a contradiction.
Negation elimination deduces any formula in the presence of ¬P together
with P:
(?P =⇒ False) =⇒ ¬ ?P (notI)
[[¬ ?P; ?P]] =⇒ ?R (notE)

Classical logic allows us to assume ¬P when attempting to prove P:

(¬ ?P =⇒ ?P) =⇒ ?P (classical)
74 5. The Rules of the Game

The implications P → Q and ¬Q → ¬P are logically equivalent, and

each is called the contrapositive of the other. Four further rules support
reasoning about contrapositives. They differ in the placement of the negation
symbols:
[[?Q; ¬ ?P =⇒ ¬ ?Q]] =⇒ ?P (contrapos_pp)
[[?Q; ?P =⇒ ¬ ?Q]] =⇒ ¬ ?P (contrapos_pn)
[[¬ ?Q; ¬ ?P =⇒ ?Q]] =⇒ ?P (contrapos_np)
[[¬ ?Q; ?P =⇒ ?Q]] =⇒ ¬ ?P (contrapos_nn)
These rules are typically applied using the erule method, where their effect
is to form a contrapositive from an assumption and the goal’s conclusion.
The most important of these is contrapos_np. It is useful for applying
introduction rules to negated assumptions. For instance, the assumption
¬(P → Q) is equivalent to the conclusion P → Q and we might want to
use conjunction introduction on it. Before we can do so, we must move that
assumption so that it becomes the conclusion. The following proof demon-
strates this technique:
lemma "[[¬(P−→Q); ¬(R−→Q)]] =⇒ R"
apply (erule_tac Q = "R−→Q" in contrapos_np)
apply (intro impI)
by (erule notE)
There are two negated assumptions and we need to exchange the conclusion
with the second one. The method erule contrapos_np would select the first
assumption, which we do not want. So we specify the desired assumption
explicitly using a new method, erule_tac. This is the resulting subgoal:
1. [[¬ (P −→ Q); ¬ R]] =⇒ R −→ Q

The former conclusion, namely R, now appears negated among the assump-
tions, while the negated formula R −→ Q becomes the new conclusion.
We can now apply introduction rules. We use the intro method, which
repeatedly applies the given introduction rules. Here its effect is equivalent
to rule impI.
1. [[¬ (P −→ Q); ¬ R; R]] =⇒ Q

We can see a contradiction in the form of assumptions ¬ R and R, which

suggests using negation elimination. If applied on its own, notE will select
the first negated assumption, which is useless. Instead, we invoke the rule
using the by command. Now when Isabelle selects the first assumption, it
tries to prove P −→ Q and fails; it then backtracks, finds the assumption ¬ R
and finally proves R by assumption. That concludes the proof.
The following example may be skipped on a first reading. It involves a
peculiar but important rule, a form of disjunction introduction:
(¬ ?Q =⇒ ?P) =⇒ ?P ∨ ?Q (disjCI)
 5.7 Interlude: the Basic Methods for Rules 75

This rule combines the effects of disjI1 and disjI2. Its great advantage is that
we can remove the disjunction symbol without deciding which disjunction
to prove. This treatment of disjunction is standard in sequent and tableau
calculi.
lemma "(P ∨ Q) ∧ R =⇒ P ∨ (Q ∧ R)"
apply (rule disjCI)
apply (elim conjE disjE)
apply assumption
by (erule contrapos_np, rule conjI)

The first proof step to applies the introduction rules disjCI. The resulting
subgoal has the negative assumption ¬(Q ∧ R).
1. [[(P ∨ Q) ∧ R; ¬ (Q ∧ R)]] =⇒ P

Next we apply the elim method, which repeatedly applies elimination rules;
here, the elimination rules given in the command. One of the subgoals is
trivial (apply assumption), leaving us with one other:
1. [[¬ (Q ∧ R); R; Q]] =⇒ P
Now we must move the formula Q ∧ R to be the conclusion. The combination
(erule contrapos_np, rule conjI)
is robust: the conjI forces the erule to select a conjunction. The two subgoals
are the ones we would expect from applying conjunction introduction to
Q ∧ R:
1. [[R; Q; ¬ P]] =⇒ Q
2. [[R; Q; ¬ P]] =⇒ R
They are proved by assumption, which is implicit in the by command.

5.7 Interlude: the Basic Methods for Rules

We have seen examples of many tactics that operate on individual rules. It
may be helpful to review how they work given an arbitrary rule such as this:
P1 ... Pn
Q

Below, we refer to P1 as the major premise. This concept applies only to

elimination and destruction rules. These rules act upon an instance of their
major premise, typically to replace it by subformulas of itself.
Suppose that the rule above is called R . Here are the basic rule methods,
most of which we have already seen:
– Method rule R unifies Q with the current subgoal, replacing it by n new
subgoals: instances of P1 , …, Pn . This is backward reasoning and is appro-
priate for introduction rules.
76 5. The Rules of the Game

– Method erule R unifies Q with the current subgoal and simultaneously

unifies P1 with some assumption. The subgoal is replaced by the n − 1 new
subgoals of proving instances of P2 , …, Pn , with the matching assump-
tion deleted. It is appropriate for elimination rules. The method (rule R,
assumption) is similar, but it does not delete an assumption.
– Method drule R unifies P1 with some assumption, which it then deletes.
The subgoal is replaced by the n − 1 new subgoals of proving P2 , …, Pn ;
an nth subgoal is like the original one but has an additional assumption:
an instance of Q. It is appropriate for destruction rules.
– Method frule R is like drule R except that the matching assumption is not
deleted. (See Sect. 5.9.5 below.)
Other methods apply a rule while constraining some of its variables. The
typical form is
rule_tac v1 = t1 and …and vk = tk in R

This method behaves like rule R, while instantiating the variables v1 , …, vk

as specified. We similarly have erule_tac, drule_tac and frule_tac. These
methods also let us specify which subgoal to operate on. By default it is
the first subgoal, as with nearly all methods, but we can specify that rule R
should be applied to subgoal number i:
rule_tac [i] R

5.8 Unification and Substitution

As we have seen, Isabelle rules involve schematic variables, which begin with a
question mark and act as placeholders for terms. Unification — well known
to Prolog programmers — is the act of making two terms identical, possibly
replacing their schematic variables by terms. The simplest case is when the
two terms are already the same. Next simplest is pattern-matching, which
replaces variables in only one of the terms. The rule method typically matches
the rule’s conclusion against the current subgoal. The assumption method
matches the current subgoal’s conclusion against each of its assumptions.
Unification can instantiate variables in both terms; the rule method can do
this if the goal itself contains schematic variables. Other occurrences of the
variables in the rule or proof state are updated at the same time.
Schematic variables in goals represent unknown terms. Given a goal such
as ∃x. P, they let us proceed with a proof. They can be filled in later, some-
times in stages and often automatically.

If unification fails when you think it should succeed, try setting the Proof Gen-
eral flag Isabelle > Settings > Trace Unification, which makes Isabelle show the
cause of unification failures (in Proof General’s Trace buffer).
 5.8 Unification and Substitution 77

For example, suppose we are trying to prove this subgoal by assumption:

1. P (a, f (b, g (e, a), b), a) =⇒ P (a, f (b, g (c, a), b), a)

The assumption method having failed, we try again with the flag set:
apply assumption

In this trivial case, the output clearly shows that e clashes with c:
Clash: e =/= c

Isabelle uses higher-order unification, which works in the typed λ-

calculus. The procedure requires search and is potentially undecidable. For
our purposes, however, the differences from ordinary unification are straight-
forward. It handles bound variables correctly, avoiding capture. The two
terms λx. f(x,z) and λy. f(y,z) are trivially unifiable because they differ
only by a bound variable renaming. The two terms λx. ?P and λx. t x are
not unifiable; replacing ?P by t x is forbidden because the free occurrence
of x would become bound. Unfortunately, even if trace_unify_fail is set,
Isabelle displays no information about this type of failure.

!! Higher-order unification sometimes must invent λ-terms to replace function

variables, which can lead to a combinatorial explosion. However, Isabelle proofs
tend to involve easy cases where there are few possibilities for the λ-term being
constructed. In the easiest case, the function variable is applied only to bound
variables, as when we try to unify λx y. f(?h x y) and λx y. f(x+y+a). The
only solution is to replace ?h by λx y. x+y+a. Such cases admit at most one unifier,
like ordinary unification. A harder case is unifying ?h a with a+b; it admits two
solutions for ?h, namely λx. a+b and λx. x+b. Unifying ?h a with a+a+b admits
four solutions; their number is exponential in the number of occurrences of a in the
second term.

5.8.1 Substitution and the subst Method

Isabelle also uses function variables to express substitution. A typical sub-

stitution rule allows us to replace one term by another if we know that two
terms are equal.
s = t P[s/x]
P[t/x]
The rule uses a notation for substitution: P[t/x] is the result of replacing x
by t in P. The rule only substitutes in the positions designated by x. For
example, it can derive symmetry of equality from reflexivity. Using x = s
for P replaces just the first s in s = s by t:

s=t s=s
t=s

The Isabelle version of the substitution rule looks like this:

78 5. The Rules of the Game

[[?t = ?s; ?P ?s]] =⇒ ?P ?t (ssubst)

Crucially, ?P is a function variable. It can be replaced by a λ-term with

one bound variable, whose occurrences identify the places in which s will be
replaced by t. The proof above requires ?P to be replaced by λx. x=s; the
second premise will then be s=s and the conclusion will be t=s.
The simp method also replaces equals by equals, but the substitution rule
gives us more control. Consider this proof:
lemma "[[x = f x; odd(f x)]] =⇒ odd x"
by (erule ssubst)

The assumption x = f x, if used for rewriting, would loop, replacing x by f x

and then by f(f x) and so forth. (Here simp would see the danger and would
re-orient the equality, but in more complicated cases it can be fooled.) When
we apply the substitution rule, Isabelle replaces every x in the subgoal by f
x just once. It cannot loop. The resulting subgoal is trivial by assumption,
so the by command proves it implicitly.
We are using the erule method in a novel way. Hitherto, the conclusion of
the rule was just a variable such as ?R, but it may be any term. The conclusion
is unified with the subgoal just as it would be with the rule method. At
the same time erule looks for an assumption that matches the rule’s first
premise, as usual. With ssubst the effect is to find, use and delete an equality
assumption.
The subst method performs individual substitutions. In simple cases, it
closely resembles a use of the substitution rule. Suppose a proof has reached
this point:
1. [[P x y z; Suc x < y]] =⇒ f z = x * y

Now we wish to apply a commutative law:

?m * ?n = ?n * ?m (mult.commute)
Isabelle rejects our first attempt:
apply (simp add: mult.commute)
The simplifier notices the danger of looping and refuses to apply the rule.2
The subst method applies mult.commute exactly once.
apply (subst mult.commute)
1. [[P x y z; Suc x < y]] =⇒ f z = y * x
As we wanted, x * y has become y * x.
This use of the subst method has the same effect as the command
apply (rule mult.commute [THEN ssubst])
2
More precisely, it only applies such a rule if the new term is smaller under a
specified ordering; here, x * y is already smaller than y * x.
 5.8 Unification and Substitution 79

The attribute THEN, which combines two rules, is described in Sect. 5.15.1 be-
low. The subst method is more powerful than applying the substitution rule.
It can perform substitutions in a subgoal’s assumptions. Moreover, if the sub-
goal contains more than one occurrence of the left-hand side of the equality,
the subst method lets us specify which occurrence should be replaced.

5.8.2 Unification and Its Pitfalls

Higher-order unification can be tricky. Here is an example, which you may

want to skip on your first reading:
lemma "[[x = f x; triple (f x) (f x) x]] =⇒ triple x x x"
apply (erule ssubst)
back
back
back
back
apply assumption
done

By default, Isabelle tries to substitute for all the occurrences. Applying erule
ssubst yields this subgoal:
1. triple (f x) (f x) x =⇒ triple (f x) (f x) (f x)

The substitution should have been done in the first two occurrences of x
only. Isabelle has gone too far. The back command allows us to reject this
possibility and demand a new one:
1. triple (f x) (f x) x =⇒ triple x (f x) (f x)

Now Isabelle has left the first occurrence of x alone. That is promising but it
is not the desired combination. So we use back again:
1. triple (f x) (f x) x =⇒ triple (f x) x (f x)
This also is wrong, so we use back again:
1. triple (f x) (f x) x =⇒ triple x x (f x)
And this one is wrong too. Looking carefully at the series of alternatives, we
see a binary countdown with reversed bits: 111, 011, 101, 001. Invoke back
again:
1. triple (f x) (f x) x =⇒ triple (f x) (f x) x

At last, we have the right combination! This goal follows by assumption.

This example shows that unification can do strange things with function
variables. We were forced to select the right unifier using the back command.
That is all right during exploration, but back should never appear in the final
version of a proof. You can eliminate the need for back by giving Isabelle
less freedom when you apply a rule.
80 5. The Rules of the Game

One way to constrain the inference is by joining two methods in a apply

command. Isabelle applies the first method and then the second. If the second
method fails then Isabelle automatically backtracks. This process continues
until the first method produces an output that the second method can use.
We get a one-line proof of our example:
lemma "[[x = f x; triple (f x) (f x) x]] =⇒ triple x x x"
apply (erule ssubst, assumption)
done

The by command works too, since it backtracks when proving subgoals by

assumption:
lemma "[[x = f x; triple (f x) (f x) x]] =⇒ triple x x x"
by (erule ssubst)

The most general way to constrain unification is by instantiating variables

in the rule. The method rule_tac is similar to rule, but it makes some of
the rule’s variables denote specified terms. Also available are drule_tac and
erule_tac. Here we need erule_tac since above we used erule.
lemma "[[x = f x; triple (f x) (f x) x]] =⇒ triple x x x"
by (erule_tac P = "λu. triple u u x" in ssubst)

To specify a desired substitution requires instantiating the variable ?P with

a λ-expression. The bound variable occurrences in λu. P u u x indicate that
the first two arguments have to be substituted, leaving the third unchanged.
With this instantiation, backtracking is neither necessary nor possible.
An alternative to rule_tac is to use rule with a theorem modified us-
ing of, described in Sect. 5.15
V below. But rule_tac, unlike of, can express
instantiations that refer to -bound variables in the current subgoal.

5.9 Quantifiers

Quantifiers require formalizing syntactic substitution and the notion of arbi-

trary value. Consider the universal quantifier. In a logic book, its introduction
rule looks like this:
P
∀x. P
Typically, a proviso written in English says that x must not occur in the as-
sumptions. This proviso guarantees that x can be regarded as arbitrary, since
it has not been assumed to satisfy any special conditions. Isabelle’s under-
lying formalism, called the meta-logic,V eliminates the need for English. It
provides its own universal quantifier ( ) to express the notion of an arbitrary
value. We have already seen another operator of the meta-logic, namely =⇒,
which expresses inference rules and the treatment of assumptions. The only
other operator in the meta-logic is ≡, which can be used to define constants.
 5.9 Quantifiers 81

5.9.1 The Universal Introduction Rule

Returning to the universal quantifier, we find that having a similar quantifier

as part of the meta-logic makes the introduction rule trivial to express:
( x. ?P x) =⇒ ∀ x. ?P x (allI)
V

The following trivial proof demonstrates how the universal introduction

rule works.
lemma "∀ x. P x −→ P x"
apply (rule allI)
by (rule impI)
The first step invokes the rule by applying the method rule allI.
1. x. P x −→ P x
V

Note that the resulting proof state has a bound variable, namely x. The rule
has replaced the universal quantifier of higher-order logic by Isabelle’s meta-
level quantifier. Our goal is to prove P x −→ P x for arbitrary x ; it is an
implication, so we apply the corresponding introduction rule (impI ).
1. x. P x =⇒ P x
V

This last subgoal is implicitly proved by assumption.

5.9.2 The Universal Elimination Rule

Now consider universal elimination. In a logic text, the rule looks like this:
∀x. P
P[t/x]

The conclusion is P with t substituted for the variable x. Isabelle expresses

substitution using a function variable:
∀ x. ?P x =⇒ ?P ?x (spec)
This destruction rule takes a universally quantified formula and removes the
quantifier, replacing the bound variable x by the schematic variable ?x. Re-
call that a schematic variable starts with a question mark and acts as a
placeholder: it can be replaced by any term.
The universal elimination rule is also available in the standard elimination
format. Like conjE, it never appears in logic books:
[[∀ x. ?P x; ?P ?x =⇒ ?R]] =⇒ ?R (allE)
The methods drule spec and erule allE do precisely the same inference.
To see how ∀-elimination works, let us derive a rule about reducing the
scope of a universal quantifier. In mathematical notation we write
∀x. P → Q
P → ∀x. Q
82 5. The Rules of the Game

with the proviso “x not free in P.” Isabelle’s treatment of substitution makes
the proviso unnecessary. The conclusion is expressed as P −→ (∀ x. Q x).
No substitution for the variable P can introduce a dependence upon x : that
would be a bound variable capture. Let us walk through the proof.
lemma "(∀ x. P −→ Q x) =⇒ P −→ (∀ x. Q x)"
First we apply implies introduction (impI ), which moves the P from the con-
clusion to the assumptions. Then we apply universal introduction (allI ).
apply
V (rule impI, rule allI)
1. x. [[∀ x. P −→ Q x; P]] =⇒ Q x
As before, it replaces the HOL quantifier by a meta-level quantifier, producing
a subgoal that binds the variable x. The leading bound variables (here x )
and the assumptions (here ∀ x. P −→ Q x and P ) form the context for the
conclusion, here Q x. Subgoals inherit the context, although assumptions can
be added or deleted (as we saw earlier), while rules such as allI add bound
variables.
Now, to reason from the universally quantified assumption, we apply the
elimination rule using the drule method. This rule is called spec because it
specializes a universal formula to a particular term.
apply
V (drule spec)
1. x. [[P; P −→ Q (?x2 x)]] =⇒ Q x
Observe how the context has changed. The quantified formula is gone, re-
placed by a new assumption derived from its body. We have removed the
quantifier and replaced the bound variable by the curious term ?x2 x. This
term is a placeholder: it may become any term that can be built from x.
(Formally, ?x2 is an unknown of function type, applied to the argument x.)
This new assumption is an implication, so we can use modus ponens on it,
which concludes the proof.
by (drule mp)
Let us take a closer look at this last step. Modus ponens yields two subgoals:
one where we prove the antecedent (in this case P ) and one where we may
assume the consequent. Both of these subgoals are proved by the assumption
method, which is implicit in the by command. Replacing the by command
by apply (drule mp, assumption) would have left one last subgoal:
1. x. [[P; Q (?x2 x)]] =⇒ Q x
V

The consequent is Q applied to that placeholder. It may be replaced by any

term built from x, and here it should simply be x. The assumption need not
be identical to the conclusion, provided the two formulas are unifiable.

5.9.3 The Existential Quantifier

The concepts just presented also apply to the existential quantifier, whose
introduction rule looks like this in Isabelle:
 5.9 Quantifiers 83

?P ?x =⇒ ∃ x. ?P x (exI)
If we can exhibit some x such that P(x) is true, then ∃x.P(x) is also true.
It is a dual of the universal elimination rule, and logic texts present it using
the same notation for substitution.
The existential elimination rule looks like this in a logic text:
[P]
..
..
∃x. P Q
Q
It looks like this in Isabelle:
[[∃ x. ?P x; x. ?P x =⇒ ?Q]] =⇒ ?Q (exE)
V

Given an existentially quantified theorem and some formula Q to prove, it

creates a new assumption by removing the quantifier. As with the universal
introduction rule, the textbook version imposes a proviso on the quantified
variable, which Isabelle expresses using its meta-logic. It is enough to have a
universal quantifier in the meta-logic; we do not need an existential quantifier
to be built in as well.
Exercise 5.9.1 Prove the lemma
∃x. P ∧ Q(x) =⇒ P ∧ (∃x. Q(x)).
Hint: the proof is similar to the one just above for the universal quantifier.

5.9.4 Renaming a Bound Variable: rename_tac

When you apply a rule such as allI, the quantified variable becomes a new
bound variable of the new subgoal. Isabelle tries to avoid changing its name,
but sometimes it has to choose a new name in order to avoid a clash. The
result may not be ideal:
lemma "x < y =⇒ ∀ x y. P x (f y)"
apply
V (intro allI)
1. xa ya. x < y =⇒ P xa (f ya)
The names x and y were already in use, so the new bound variables are called
xa and ya. You can rename them by invoking rename_tac:
apply
V (rename_tac v w)
1. v w. x < y =⇒ P v (f w)
Recall that rule_tac instantiates a theorem with specified terms. These terms
may involve the goal’s bound variables, but beware of referring to variables
like xa. A future change to your theories could change the set of names pro-
duced at top level, so that xa changes to xb or reverts to x. It is safer to
rename automatically-generated variables before mentioning them.
If the subgoal has more bound variables than there are names given to
rename_tac, the rightmost ones are renamed.
84 5. The Rules of the Game

5.9.5 Reusing an Assumption: frule

Note that drule spec removes the universal quantifier and — as usual with
elimination rules — discards the original formula. Sometimes, a universal
formula has to be kept so that it can be used again. Then we use a new
method: frule. It acts like drule but copies rather than replaces the selected
assumption. The f is for forward.
In this example, going from P a to P(h(h a)) requires two uses of the
quantified assumption, one for each h in h(h a).
lemma "[[∀ x. P x −→ P (h x); P a]] =⇒ P(h (h a))"
Examine the subgoal left by frule:
apply (frule spec)
1. [[∀ x. P x −→ P (h x); P a; P ?x −→ P (h ?x)]] =⇒ P (h (h a))
It is what drule would have left except that the quantified assumption is still
present. Next we apply mp to the implication and the assumption P a:
apply (drule mp, assumption)
1. [[∀ x. P x −→ P (h x); P a; P (h a)]] =⇒ P (h (h a))
We have created the assumption P(h a), which is progress. To continue the
proof, we apply spec again. We shall not need it again, so we can use drule.
apply (drule spec)
1. [[P a; P (h a); P ?x2 −→ P (h ?x2)]] =⇒ P (h (h a))
The new assumption bridges the gap between P(h a) and P(h(h a)).
by (drule mp)

A final remark. Replacing this by command with

apply (drule mp, assumption)
would not work: it would add a second copy of P(h a) instead of the desired
assumption, P(h(h a)). The by command forces Isabelle to backtrack until it
finds the correct one. Alternatively, we could have used the apply command
and bundled the drule mp with two calls of assumption. Or, of course, we
could have given the entire proof to auto.

5.9.6 Instantiating a Quantifier Explicitly

We can prove a theorem of the form ∃x. P x by exhibiting a suitable term t

such that P t is true. Dually, we can use an assumption of the form ∀x. P x to
generate a new assumption P t for a suitable term t. In many cases, Isabelle
makes the correct choice automatically, constructing the term by unification.
In other cases, the required term is not obvious and we must specify it our-
selves. Suitable methods are rule_tac, drule_tac and erule_tac.
We have seen (just above, Sect. 5.9.5) a proof of this lemma:
 5.9 Quantifiers 85

lemma "[[∀ x. P x −→ P (h x); P a]] =⇒ P(h (h a))"

We had reached this subgoal:

1. [[∀ x. P x −→ P (h x); P a; P (h a)]] =⇒ P (h (h a))

The proof requires instantiating the quantified assumption with the term h a.
apply (drule_tac x = "h a" in spec)
1. [[P a; P (h a); P (h a) −→ P (h (h a))]] =⇒ P (h (h a))
We have forced the desired instantiation.
Existential formulas can be instantiated too. The next example uses the
divides relation of number theory:
?m dvd ?n ≡ ∃ k. ?n = ?m * k (dvd_def)

Let us prove that multiplication of natural numbers is monotone with

respect to the divides relation:
lemma mult_dvd_mono: "[[i dvd m; j dvd n]] =⇒ i*j dvd (m*n :: nat)"
apply (simp add: dvd_def)

Unfolding the definition of divides has left this subgoal:

1. [[∃ k. m = i * k; ∃ k. n = j * k]] =⇒ ∃ k. m * n = i * j * k

Next, we eliminate the two existential quantifiers in the assumptions:

apply
V (erule exE)
1. k. [[∃ k. n = j * k; m = i * k]] =⇒ ∃ k. m * n = i * j * k
apply
V (erule exE)
1. k ka. [[m = i * k; n = j * ka]] =⇒ ∃ k. m * n = i * j * k

The term needed to instantiate the remaining quantifier is k*ka. But ka is an

automatically-generated name. As noted above, references to such variable
names makes a proof less resilient to future changes. So, first we rename the
most recent variable to l:
apply
V (rename_tac l)
1. k l. [[m = i * k; n = j * l]] =⇒ ∃ k. m * n = i * j * k

We instantiate the quantifier with k*l:

apply
V (rule_tac x="k*l" in exI)
1. k ka. [[m = i * k; n = j * ka]] =⇒ m * n = i * j * (k * ka)

The rest is automatic, by arithmetic.

apply simp
done
86 5. The Rules of the Game

5.10 Description Operators

HOL provides two description operators. A definite description formalizes
the word “the,” as in “the greatest divisior of n.” It returns an arbitrary
value unless the formula has a unique solution. An indefinite description
formalizes the word “some,” as in “some member of S.” It differs from a
definite description in not requiring the solution to be unique: it uses the
axiom of choice to pick any solution.

!! Description operators can be hard to reason about. Novices should try to avoid
them. Fortunately, descriptions are seldom required.

5.10.1 Definite Descriptions

A definite description is traditionally written ιx.P(x). It denotes the x such

that P(x) is true, provided there exists a unique such x; otherwise, it returns
an arbitrary value of the expected type. Isabelle uses THE for the Greek letter ι.
We reason using this rule, where a is the unique solution:
[[P a; x. P x =⇒ x = a]] =⇒ (THE x. P x) = a (the_equality)
V

For instance, we can define the cardinality of a finite set A to be that n

such that A is in one-to-one correspondence with {1, . . . , n}. We can then
prove that the cardinality of the empty set is zero (since n = 0 satisfies the
description) and proceed to prove other facts.
A more challenging example illustrates how Isabelle/HOL defines the least
number operator, which denotes the least x satisfying P :
(LEAST x. P x) = (THE x. P x ∧ (∀ y. P y −→ x ≤ y))
Let us prove the analogue of the_equality for LEAST .
theorem Least_equality:
"[[P (k::nat); ∀ x. P x −→ k ≤ x]] =⇒ (LEAST x. P x) = k"
apply (simp add: Least_def)

1. [[P k; ∀ x. P x −→ k ≤ x]]
=⇒ (THE x. P x ∧ (∀ y. P y −→ x ≤ y)) = k
The first step has merely unfolded the definition.
apply (rule the_equality)

1. [[P
V k; ∀ x. P x −→ k ≤ x]] =⇒ P k ∧ (∀ y. P y −→ k ≤ y)
2. x. [[P k; ∀ x. P x −→ k ≤ x; P x ∧ (∀ y. P y −→ x ≤ y)]]
=⇒ x = k
As always with the_equality, we must show existence and uniqueness of the
claimed solution, k. Existence, the first subgoal, is trivial. Uniqueness, the
second subgoal, follows by antisymmetry:
[[x ≤ y; y ≤ x]] =⇒ x = y (order_antisym)
 5.10 Description Operators 87

The assumptions imply both k ≤ x and x ≤ k. One call to auto does it all:
by (auto intro: order_antisym)

5.10.2 Indefinite Descriptions

An indefinite description is traditionally written εx.P(x) and is known as

Hilbert’s ε-operator. It denotes some x such that P(x) is true, provided one
exists. Isabelle uses SOME for the Greek letter ε.
Here is the definition of inv ,3 which expresses inverses of functions:
inv f ≡ λy. SOME x. f x = y (inv_def)

Using SOME rather than THE makes inv f behave well even if f is not injective.
As it happens, most useful theorems about inv do assume the function to be
injective.
The inverse of f, when applied to y, returns some x such that f x = y.
For example, we can prove inv Suc really is the inverse of the Suc function
lemma "inv Suc (Suc n) = n"
by (simp add: inv_def)
The proof is a one-liner: the subgoal simplifies to a degenerate application of
SOME, which is then erased. In detail, the left-hand side simplifies to SOME x.
Suc x = Suc n, then to SOME x. x = n and finally to n.
We know nothing about what inv Suc returns when applied to zero. The
proof above still treats SOME as a definite description, since it only reasons
about situations in which the value is described uniquely. Indeed, SOME satis-
fies this rule:
[[P a; x. P x =⇒ x = a]] =⇒ (SOME x. P x) = a (some_equality)
V

To go further is tricky and requires rules such as these:

P x =⇒V P (SOME x. P x) (someI)
[[P a; x. P x =⇒ Q x]] =⇒ Q (SOME x. P x) (someI2)

Rule someI is basic: if anything satisfies P then so does SOME x. P x . The

repetition of P in the conclusion makes it difficult to apply in a backward
proof, so the derived rule someI2 is also provided.
For example, let us prove the axiom of choice:
theorem axiom_of_choice: "(∀ x. ∃ y. P x y) =⇒ ∃ f. ∀ x. P x (f x)"
apply (rule exI, rule allI)

1. x. ∀ x. ∃ y. P x y =⇒ P x (?f x)
V

We have applied the introduction rules; now it is time to apply the elimination
rules.
3
In fact, inv is defined via a second constant inv_into, which we ignore here.
88 5. The Rules of the Game

apply (drule spec, erule exE)

1. x y. P (?x2 x) y =⇒ P x (?f x)
V

The rule someI automatically instantiates f to λx. SOME y. P x y , which is

the choice function. It also instantiates ?x2 x to x.
by (rule someI)

Historical Note. The original purpose of Hilbert’s ε-operator was to ex-

press an existential destruction rule:
∃x. P
P[(εx.P)/ x]

This rule is seldom used for that purpose — it can cause exponential blow-up
— but it is occasionally used as an introduction rule for the ε-operator. Its
name in HOL is someI_ex .

5.11 Some Proofs That Fail

Most of the examples in this tutorial involve proving theorems. But not every
conjecture is true, and it can be instructive to see how proofs fail. Here we
attempt to prove a distributive law involving the existential quantifier and
conjunction.
lemma "(∃ x. P x) ∧ (∃ x. Q x) =⇒ ∃ x. P x ∧ Q x"

The first steps are routine. We apply conjunction elimination to break the as-
sumption into two existentially quantified assumptions. Applying existential
elimination removes one of the quantifiers.
apply (erule conjE)
apply
V (erule exE)
1. x. [[∃ x. Q x; P x]] =⇒ ∃ x. P x ∧ Q x

When we remove the other quantifier, we get a different bound variable in

the subgoal. (The name xa is generated automatically.)
apply
V (erule exE)
1. x xa. [[P x; Q xa]] =⇒ ∃ x. P x ∧ Q x

The proviso of the existential elimination rule has forced the variables to
differ: we can hardly expect two arbitrary values to be equal! There is no
way to prove this subgoal. Removing the conclusion’s existential quantifier
yields two identical placeholders, which can become any term involving the
variables x and xa. We need one to become x and the other to become xa,
but Isabelle requires all instances of a placeholder to be identical.
 5.11 Some Proofs That Fail 89

apply (rule exI)

apply
V (rule conjI)
1. Vx xa. [[P x; Q xa]] =⇒ P (?x3 x xa)
2. x xa. [[P x; Q xa]] =⇒ Q (?x3 x xa)

We can prove either subgoal using the assumption method. If we prove the
first one, the placeholder changes into x.
apply
V assumption
1. x xa. [[P x; Q xa]] =⇒ Q x
We are left with a subgoal that cannot be proved. Applying the assumption
method results in an error message:
*** empty result sequence -- proof command failed

When interacting with Isabelle via the shell interface, you can abandon a
proof using the oops command.
Here is another abortive proof, illustrating the interaction between bound
variables and unknowns. If R is a reflexive relation, is there an x such that
R x y holds for all y? Let us see what happens when we attempt to prove it.
lemma "∀ y. R y y =⇒ ∃ x. ∀ y. R x y"

First, we remove the existential quantifier. The new proof state has an un-
known, namely ?x.
apply (rule exI)
1. ∀ y. R y y =⇒ ∀ y. R ?x y

It looks like we can just apply assumption, but it fails. Isabelle refuses to
substitute y, a bound variable, for ?x ; that would be a bound variable capture.
We can still try to finish the proof in some other way. We remove the universal
quantifier from the conclusion, moving the bound variable y into the subgoal.
But note that it is still bound!
apply
V (rule allI)
1. y. ∀ y. R y y =⇒ R ?x y

Finally, we try to apply our reflexivity assumption. We obtain a new assump-

tion whose identical placeholders may be replaced by any term involving y.
apply
V (drule spec)
1. y. R (?z2 y) (?z2 y) =⇒ R ?x y

This subgoal can only be proved by putting y for all the placeholders, making
the assumption and conclusion become R y y. Isabelle can replace ?z2 y by
y ; this involves instantiating ?z2 to the identity function. But, just as two
steps earlier, Isabelle refuses to substitute y for ?x. This example is typical
of how Isabelle enforces sound quantifier reasoning.
90 5. The Rules of the Game

5.12 Proving Theorems Using the blast Method

It is hard to prove many theorems using the methods described above. A

proof may be hundreds of steps long. You may need to search among different
ways of proving certain subgoals. Often a choice that proves one subgoal
renders another impossible to prove. There are further complications that we
have not discussed, concerning negation and disjunction. Isabelle’s classical
reasoner is a family of tools that perform such proofs automatically. The
most important of these is the blast method.
In this section, we shall first see how to use the classical reasoner in its
default mode and then how to insert additional rules, enabling it to work in
new problem domains.
We begin with examples from pure predicate logic. The following exam-
ple is known as Andrew’s challenge. Peter Andrews designed it to be hard
to prove by automatic means. It is particularly hard for a resolution prover,
where converting the nested biconditionals to clause form produces a combi-
natorial explosion [33]. However, the blast method proves it in a fraction of
a second.
lemma "((∃ x. ∀ y. p(x)=p(y)) = ((∃ x. q(x))=(∀ y. p(y)))) =
((∃ x. ∀ y. q(x)=q(y)) = ((∃ x. p(x))=(∀ y. q(y))))"
by blast

The next example is a logic problem composed by Lewis Carroll. The blast
method finds it trivial. Moreover, it turns out that not all of the assumptions
are necessary. We can experiment with variations of this formula and see
which ones can be proved.
lemma "(∀ x. honest(x) ∧ industrious(x) −→ healthy(x)) ∧
¬ (∃ x. grocer(x) ∧ healthy(x)) ∧
(∀ x. industrious(x) ∧ grocer(x) −→ honest(x)) ∧
(∀ x. cyclist(x) −→ industrious(x)) ∧
(∀ x. ¬healthy(x) ∧ cyclist(x) −→ ¬honest(x))
−→ (∀ x. grocer(x) −→ ¬cyclist(x))"
by blast

The blast method is also effective for set theory, which is described in the
next chapter. The formula below may look horrible, but the blast method
proves it in milliseconds.
lemma "(S i∈I. S
A(i)) ∩ ( j∈J. B(j)) =
S S
( i∈I. j∈J. A(i) ∩ B(j))"
by blast

Few subgoals are couched purely in predicate logic and set theory. We can
extend the scope of the classical reasoner by giving it new rules. Extending
it effectively requires understanding the notions of introduction, elimination
and destruction rules. Moreover, there is a distinction between safe and un-
safe rules. A safe rule is one that can be applied backwards without losing
 5.13 Other Classical Reasoning Methods 91

information; an unsafe rule loses information, perhaps transforming the sub-

goal into one that cannot be proved. The safe/unsafe distinction affects the
proof search: if a proof attempt fails, the classical reasoner backtracks to the
most recent unsafe rule application and makes another choice.
An important special case avoids all these complications. A logical equiv-
alence, which in higher-order logic is an equality between formulas, can be
given to the classical reasoner and simplifier by using the attribute iff . You
should do so if the right hand side of the equivalence is simpler than the
left-hand side.
For example, here is a simple fact about list concatenation. The result of
appending two lists is empty if and only if both of the lists are themselves
empty. Obviously, applying this equivalence will result in a simpler goal.
When stating this lemma, we include the iff attribute. Once we have proved
the lemma, Isabelle will make it known to the classical reasoner (and to the
simplifier).
lemma [iff]: "(xs@ys = []) = (xs=[] ∧ ys=[])"
apply (induct_tac xs)
apply (simp_all)
done
This fact about multiplication is also appropriate for the iff attribute:
(?m * ?n = 0) = (?m = 0 ∨ ?n = 0)
A product is zero if and only if one of the factors is zero. The reasoning
involves a disjunction. Proving new rules for disjunctive reasoning is hard,
but translating to an actual disjunction works: the classical reasoner handles
disjunction properly.
In more detail, this is how the iff attribute works. It converts the equiv-
alence P = Q to a pair of rules: the introduction rule Q =⇒ P and the
destruction rule P =⇒ Q. It gives both to the classical reasoner as safe rules,
ensuring that all occurrences of P in a subgoal are replaced by Q. The simpli-
fier performs the same replacement, since iff gives P = Q to the simplifier.
Classical reasoning is different from simplification. Simplification is deter-
ministic. It applies rewrite rules repeatedly, as long as possible, transforming
a goal into another goal. Classical reasoning uses search and backtracking in
order to prove a goal outright.

5.13 Other Classical Reasoning Methods

The blast method is our main workhorse for proving theorems automatically.
Other components of the classical reasoner interact with the simplifier. Still
others perform classical reasoning to a limited extent, giving the user fine
control over the proof.
Of the latter methods, the most useful is clarify . It performs all obvious
reasoning steps without splitting the goal into multiple parts. It does not
92 5. The Rules of the Game

apply unsafe rules that could render the goal unprovable. By performing the
obvious steps, clarify lays bare the difficult parts of the problem, where
human intervention is necessary.
For example, the following conjecture is false:
lemma "(∀ x. P x) ∧ (∃ x. Q x) −→ (∀ x. P x ∧ Q x)"
apply clarify
The blast method would simply fail, but clarify presents a subgoal that
helps us see why we cannot continue the proof.
1. x xa. [[∀ x. P x; Q xa]] =⇒ P x ∧ Q x
V

The proof must fail because the assumption Q xa and conclusion Q x refer to
distinct bound variables. To reach this state, clarify applied the introduction
rules for −→ and ∀ and the elimination rule for ∧. It did not apply the
introduction rule for ∧ because of its policy never to split goals.
Also available is clarsimp , a method that interleaves clarify and simp.
Also there is safe, which like clarify performs obvious steps but even applies
those that split goals.
The force method applies the classical reasoner and simplifier to one goal.
Unless it can prove the goal, it fails. Contrast that with the auto method,
which also combines classical reasoning with simplification. The latter’s pur-
pose is to prove all the easy subgoals and parts of subgoals. Unfortunately,
it can produce large numbers of new subgoals; also, since it proves some
subgoals and splits others, it obscures the structure of the proof tree. The
force method does not have these drawbacks. Another difference: force tries
harder than auto to prove its goal, so it can take much longer to terminate.
Older components of the classical reasoner have largely been superseded
by blast, but they still have niche applications. Most important among these
are fast and best. While blast searches for proofs using a built-in first-
order reasoner, these earlier methods search for proofs using standard Isabelle
inference. That makes them slower but enables them to work in the presence
of the more unusual features of Isabelle rules, such as type classes and function
unknowns. For example, recall the introduction rule for Hilbert’s ε-operator:
?P ?x =⇒ ?P (SOME x. ?P x) (someI)
The repeated occurrence of the variable ?P makes this rule tricky to apply.
Consider this contrived example:
lemma "[[Q a; P a]]
=⇒ P (SOME x. P x ∧ Q x) ∧ Q (SOME x. P x ∧ Q x)"
apply (rule someI)
We can apply rule someI explicitly. It yields the following subgoal:
1. [[Q a; P a]] =⇒ P ?x ∧ Q ?x
The proof from this point is trivial. Could we have proved the theorem with
a single command? Not using blast: it cannot perform the higher-order uni-
fication needed here. The fast method succeeds:
 5.14 Finding More Theorems 93

apply (fast intro!: someI)

The best method is similar to fast but it uses a best-first search instead
of depth-first search. Accordingly, it is slower but is less susceptible to di-
vergence. Transitivity rules usually cause fast to loop where best can often
manage.
Here is a summary of the classical reasoning methods:
– blast works automatically and is the fastest
– clarify and clarsimp perform obvious steps without splitting the goal;
safe even splits goals
– force uses classical reasoning and simplification to prove a goal; auto is
similar but leaves what it cannot prove
– fast and best are legacy methods that work well with rules involving un-
usual features
A table illustrates the relationships among four of these methods.
no split split
no simp clarify safe
simp clarsimp auto

5.14 Finding More Theorems

In Sect. 3.1.11, we introduced Proof General’s Find button for finding theo-
rems in the database via pattern matching. If we are inside a proof, we can
be more specific; we can search for introduction, elimination and destruction
rules with respect to the current goal. For this purpose, Find provides three
aditional search criteria: intro, elim and dest.
For example, given the goal
1. A ∧ B
you can click on Find and type in the search expression intro. You will
be shown a few rules ending in =⇒ ?P ∧ ?Q, among them conjI . You may
even discover that the very theorem you are trying to prove is already in the
database. Given the goal
1. A −→ A

the search for intro finds not just impI but also imp_refl: ?P −→ ?P.
As before, search criteria can be combined freely: for example,
"_ @ _" intro
searches for all introduction rules that match the current goal and mention
the @ function.
Searching for elimination and destruction rules via elim and dest is anal-
ogous to intro but takes the assumptions into account, too.
94 5. The Rules of the Game

5.15 Forward Proof: Transforming Theorems

Forward proof means deriving new facts from old ones. It is the most funda-
mental type of proof. Backward proof, by working from goals to subgoals, can
help us find a difficult proof. But it is not always the best way of presenting
the proof thus found. Forward proof is particularly good for reasoning from
the general to the specific. For example, consider this distributive law for the
greatest common divisor:

k × gcd(m, n) = gcd(k × m, k × n)

Putting m = 1 we get (since gcd(1, n) = 1 and k × 1 = k)

k = gcd(k, k × n)

We have derived a new fact; if re-oriented, it might be useful for simplification.

After re-orienting it and putting n = 1, we derive another useful law:

gcd(k, k) = k

Substituting values for variables — instantiation — is a forward step. Re-

orientation works by applying the symmetry of equality to an equation, so it
too is a forward step.

5.15.1 Modifying a Theorem using of, where and THEN

Let us reproduce our examples in Isabelle. Recall that in Sect. 3.5.3 we de-
clared the recursive function gcd:
fun gcd :: "nat ⇒ nat ⇒ nat" where
"gcd m n = (if n=0 then m else gcd n (m mod n))"

From this definition, it is possible to prove the distributive law. That takes
us to the starting point for our example.
?k * gcd ?m ?n = gcd (?k * ?m) (?k * ?n) (gcd_mult_distrib2)

The first step in our derivation is to replace ?m by 1. We instantiate the

theorem using of , which identifies variables in order of their appearance from
left to right. In this case, the variables are ?k, ?m and ?n. So, the expression
[of k 1] replaces ?k by k and ?m by 1.
lemmas gcd_mult_0 = gcd_mult_distrib2 [of k 1]

The keyword lemmas declares a new theorem, which can be derived from an
existing one using attributes such as [of k 1]. The command thm gcd_mult_0
displays the result:
k * gcd 1 ?n = gcd (k * 1) (k * ?n)
 5.15 Forward Proof: Transforming Theorems 95

Something is odd: k is an ordinary variable, while ?n is schematic. We did

not specify an instantiation for ?n. In its present form, the theorem does not
allow substitution for k. One solution is to avoid giving an instantiation for
?k : instead of a term we can put an underscore (_). For example,
gcd_mult_distrib2 [of _ 1]
replaces ?m by 1 but leaves ?k unchanged.
An equivalent solution is to use the attribute where.
gcd mult distrib2 [where m=1]
While of refers to variables by their position, where refers to variables by
name. Multiple instantiations are separated by and, as in this example:
gcd mult distrib2 [where m=1 and k=1]
We now continue the present example with the version of gcd_mult_0
shown above, which has k instead of ?k. Once we have replaced ?m by 1, we
must next simplify the theorem gcd_mult_0, performing the steps gcd(1, n) =
1 and k × 1 = k. The simplified attribute takes a theorem and returns the
result of simplifying it, with respect to the default simplification rules:
lemmas gcd_mult_1 = gcd_mult_0 [simplified]
Again, we display the resulting theorem:
k = gcd k (k * ?n)
To re-orient the equation requires the symmetry rule:
?s = ?t =⇒ ?t = ?s (sym)
The following declaration gives our equation to sym:
lemmas gcd_mult = gcd_mult_1 [THEN sym]
Here is the result:
gcd k (k * ?n) = k
THEN sym gives the current theorem to the rule sym and returns the result-
ing conclusion. The effect is to exchange the two operands of the equality.
Typically THEN is used with destruction rules. Also useful is THEN spec, which
removes the quantifier from a theorem of the form ∀x. P, and THEN mp, which
P
converts the implication P → Q into the rule . Similar to mp are the follow-
Q
ing two rules, which extract the two directions of reasoning about a boolean
equivalence:
[[?Q = ?P; ?Q]] =⇒ ?P (iffD1)
[[?P = ?Q; ?Q]] =⇒ ?P (iffD2)
Normally we would never name the intermediate theorems such as gcd_mult_0
and gcd_mult_1 but would combine the three forward steps:
lemmas gcd_mult = gcd_mult_distrib2 [of k 1, simplified, THEN sym]
96 5. The Rules of the Game

The directives, or attributes, are processed from left to right. This declaration
of gcd_mult is equivalent to the previous one.
Such declarations can make the proof script hard to read. Better is to
state the new lemma explicitly and to prove it using a single rule method
whose operand is expressed using forward reasoning:
lemma gcd mult [simp]: "gcd k (k*n) = k"
by (rule gcd_mult_distrib2 [of k 1, simplified, THEN sym])
Compared with the previous proof of gcd_mult, this version shows the reader
what has been proved. Also, the result will be processed in the normal way.
In particular, Isabelle generalizes over all variables: the resulting theorem will
have ?k instead of k .
At the start of this section, we also saw a proof of gcd(k, k) = k. Here is
the Isabelle version:
lemma gcd self [simp]: "gcd k k = k"
by (rule gcd_mult [of k 1, simplified])

!! To give of a nonatomic term, enclose it in quotation marks, as in [of "k*m"].

The term must not contain unknowns: an attribute such as [of "?k*m"] will
be rejected.

5.15.2 Modifying a Theorem using OF

Recall that of generates an instance of a rule by specifying values for its

variables. Analogous is OF, which generates an instance of a rule by specifying
facts for its premises.
We again need the divides relation of number theory, which as we recall
is defined by
?m dvd ?n ≡ ∃ k. ?n = ?m * k (dvd_def)
Suppose, for example, that we have proved the following rule. It states that
if k and n are relatively prime and if k divides m × n then k divides m.
[[gcd ?k ?n = 1; ?k dvd ?m * ?n]] =⇒ ?k dvd ?m (relprime_dvd_mult)
We can use OF to create an instance of this rule. First, we prove an instance
of its first premise:
lemma relprime 20 81: "gcd 20 81 = 1"
by (simp add: gcd.simps)
We have evaluated an application of the gcd function by simplification. Ex-
pression evaluation involving recursive functions is not guaranteed to termi-
nate, and it can be slow; Isabelle performs arithmetic by rewriting symbolic
bit strings. Here, however, the simplification takes less than one second. We
can give this new lemma to OF. The expression
relprime_dvd_mult [OF relprime_20_81]
 5.16 Forward Reasoning in a Backward Proof 97

yields the theorem

20 dvd (?m * 81) =⇒ 20 dvd ?m
OF takes any number of operands. Consider the following facts about the
divides relation:
[[?k dvd ?m; ?k dvd ?n]] =⇒ ?k dvd ?m + ?n (dvd_add)
?m dvd ?m (dvd_refl)
Let us supply dvd_refl for each of the premises of dvd_add:
dvd_add [OF dvd_refl dvd_refl]
Here is the theorem that we have expressed:
?k dvd (?k + ?k)
As with of, we can use the _ symbol to leave some positions unspecified:
dvd_add [OF _ dvd_refl]
The result is
?k dvd ?m =⇒ ?k dvd ?m + ?k
You may have noticed that THEN and OF are based on the same idea,
namely to combine two rules. They differ in the order of the combination
and thus in their effect. We use THEN typically with a destruction rule to
extract a subformula of the current theorem. We use OF with a list of facts
to generate an instance of the current theorem.
Here is a summary of some primitives for forward reasoning:
– of instantiates the variables of a rule to a list of terms
– OF applies a rule to a list of theorems
– THEN gives a theorem to a named rule and returns the conclusion
– simplified applies the simplifier to a theorem
– lemmas assigns a name to the theorem produced by the attributes above

5.16 Forward Reasoning in a Backward Proof

We have seen that the forward proof directives work well within a backward
proof. There are many ways to achieve a forward style using our existing
proof methods. We shall also meet some new methods that perform forward
reasoning.
The methods drule, frule, drule_tac, etc., reason forward from a subgoal.
We have seen them already, using rules such as mp and spec to operate on
formulae. They can also operate on terms, using rules such as these:
x = y =⇒ f x = f y (arg_cong)
i ≤ j =⇒ i * k ≤ j * k (mult_le_mono1)
For example, let us prove a fact about divisibility in the natural numbers:
98 5. The Rules of the Game

lemma "2 ≤ u =⇒ u*m 6= Suc(u*n)"

apply (intro notI)
1. [[2 ≤ u; u * m = Suc (u * n)]] =⇒ False

The key step is to apply the function …mod u to both sides of the equation
u*m = Suc(u*n):
apply (drule_tac f="λx. x mod u" in arg_cong)
1. [[2 ≤ u; u * m mod u = Suc (u * n) mod u]] =⇒ False

Simplification reduces the left side to 0 and the right side to 1, yielding the
required contradiction.
apply (simp add: mod_Suc)
done

Our proof has used a fact about remainder:

Suc m mod n =
(if Suc (m mod n) = n then 0 else Suc (m mod n)) (mod_Suc)

5.16.1 The Method insert

The insert method inserts a given theorem as a new assumption of all sub-
goals. This already is a forward step; moreover, we may (as always when
using a theorem) apply of, THEN and other directives. The new assumption
can then be used to help prove the subgoals.
For example, consider this theorem about the divides relation. The first
proof step inserts the distributive law for gcd. We specify its variables as
shown.
lemma relprime dvd mult:
"[[ gcd k n = 1; k dvd m*n ]] =⇒ k dvd m"
apply (insert gcd_mult_distrib2 [of m k n])

In the resulting subgoal, note how the equation has been inserted:
1. [[gcd k n = 1; k dvd m * n; m * gcd k n = gcd (m * k) (m * n)]]
=⇒ k dvd m

The next proof step utilizes the assumption gcd k n = 1 (note that Suc 0 is
another expression for 1):
apply(simp)
1. [[gcd k n = Suc 0; k dvd m * n; m = gcd (m * k) (m * n)]]
=⇒ k dvd m

Simplification has yielded an equation for m. The rest of the proof is omitted.
Here is another demonstration of insert. Division and remainder obey a
well-known law:
(?m div ?n) * ?n + ?m mod ?n = ?m (div_mult_mod_eq)

We refer to this law explicitly in the following proof:

 5.16 Forward Reasoning in a Backward Proof 99

lemma div_mult_self_is_m:
"0<n =⇒ (m*n) div n = (m::nat)"
apply (insert div_mult_mod_eq [of "m*n" n])
apply (simp)
done
The first step inserts the law, specifying m*n and n for its variables. Notice
that non-trivial expressions must be enclosed in quotation marks. Here is the
resulting subgoal, with its new assumption:
1. [[0 < n; (m * n) div n * n + (m * n) mod n = m * n]]
=⇒ (m * n) div n = m
Simplification reduces (m * n) mod n to zero. Then it cancels the factor n on
both sides of the equation (m * n) div n * n = m * n, proving the theorem.

!! Any unknowns in the theorem given to insert will be universally quantified in

the new assumption.

5.16.2 The Method subgoal_tac

A related method is subgoal_tac, but instead of inserting a theorem as an
assumption, it inserts an arbitrary formula. This formula must be proved
later as a separate subgoal. The idea is to claim that the formula holds on
the basis of the current assumptions, to use this claim to complete the proof,
and finally to justify the claim. It gives the proof some structure. If you find
yourself generating a complex assumption by a long series of forward steps,
consider using subgoal_tac instead: you can state the formula you are aiming
for, and perhaps prove it automatically.
Look at the following example.
lemma "[[(z::int) < 37; 66 < 2*z; z*z 6= 1225; Q(34); Q(36)]]
=⇒ Q(z)"
apply (subgoal_tac "z = 34 ∨ z = 36")
apply blast
apply (subgoal_tac "z 6= 35")
apply arith
apply force
done
The first assumption tells us that z is no greater than 36. The second tells
us that z is at least 34. The third assumption tells us that z cannot be 35,
since 35 × 35 = 1225. So z is either 34 or 36, and since Q holds for both of
those values, we have the conclusion.
The Isabelle proof closely follows this reasoning. The first step is to claim
that z is either 34 or 36. The resulting proof state gives us two subgoals:
1. [[z < 37; 66 < 2 * z; z * z 6= 1225; Q 34; Q 36;
z = 34 ∨ z = 36]]
=⇒ Q z
2. [[z < 37; 66 < 2 * z; z * z =6 1225; Q 34; Q 36]]
=⇒ z = 34 ∨ z = 36
100 5. The Rules of the Game

The first subgoal is trivial (blast), but for the second Isabelle needs help to
eliminate the case z =35. The second invocation of subgoal_tac leaves two
subgoals:
1. [[z < 37; 66 < 2 * z; z * z 6= 1225; Q 34; Q 36;
z 6= 35]]
=⇒ z = 34 ∨ z = 36
2. [[z < 37; 66 < 2 * z; z * z =6 1225; Q 34; Q 36]]
=⇒ z 6= 35

Assuming that z is not 35, the first subgoal follows by linear arithmetic
(arith). For the second subgoal we apply the method force, which proceeds
by assuming that z =35 and arriving at a contradiction.
Summary of these methods:
– insert adds a theorem as a new assumption
– subgoal_tac adds a formula as a new assumption and leaves the subgoal of
proving that formula

5.17 Managing Large Proofs

Naturally you should try to divide proofs into manageable parts. Look for
lemmas that can be proved separately. Sometimes you will observe that they
are instances of much simpler facts. On other occasions, no lemmas suggest
themselves and you are forced to cope with a long proof involving many
subgoals.

5.17.1 Tacticals, or Control Structures

If the proof is long, perhaps it at least has some regularity. Then you can
express it more concisely using tacticals, which provide control structures.
Here is a proof (it would be a one-liner using blast, but forget that) that
contains a series of repeated commands:
lemma "[[P−→Q; Q−→R; R−→S; P]] =⇒ S"
apply (drule mp, assumption)
apply (drule mp, assumption)
apply (drule mp, assumption)
apply (assumption)
done

Each of the three identical commands finds an implication and proves its
antecedent by assumption. The first one finds P−→Q and P, concluding Q ; the
second one concludes R and the third one concludes S. The final step matches
the assumption S with the goal to be proved.
Suffixing a method with a plus sign (+ ) expresses one or more repetitions:
 5.17 Managing Large Proofs 101

lemma "[[P−→Q; Q−→R; R−→S; P]] =⇒ S"

by (drule mp, assumption)+
Using by takes care of the final use of assumption. The new proof is more
concise. It is also more general: the repetitive method works for a chain of
implications having any length, not just three.
Choice is another control structure. Separating two methods by a vertical
bar (|) gives the effect of applying the first method, and if that fails, trying
the second. It can be combined with repetition, when the choice must be
made over and over again. Here is a chain of implications in which most of
the antecedents are proved by assumption, but one is proved by arithmetic:
lemma "[[Q−→R; P−→Q; x<5−→P; Suc x < 5]] =⇒ R"
by (drule mp, (assumption|arith))+
The arith method can prove x < 5 from x +1 < 5, but it cannot duplicate the
effect of assumption. Therefore, we combine these methods using the choice
operator.
A postfixed question mark (? ) expresses zero or one repetitions of a
method. It can also be viewed as the choice between executing a method
and doing nothing. It is useless at top level but can be valuable within other
control structures; for example, (m+)? performs zero or more repetitions of
method m.

5.17.2 Subgoal Numbering

Another problem in large proofs is contending with huge subgoals or many

subgoals. Induction can produce a proof state that looks like this:
1. bigsubgoal1
2. bigsubgoal2
3. bigsubgoal3
4. bigsubgoal4
5. bigsubgoal5
6. bigsubgoal6
If each bigsubgoal is 15 lines or so, the proof state will be too big to scroll
through. By default, Isabelle displays at most 10 subgoals. The pr command
lets you change this limit:
pr 2
1. bigsubgoal1
2. bigsubgoal2
A total of 6 subgoals...

All methods apply to the first subgoal. Sometimes, not only in a large
proof, you may want to focus on some other subgoal. Then you should try
the commands defer or prefer.
In the following example, the first subgoal looks hard, while the others
look as if blast alone could prove them:
102 5. The Rules of the Game

1. hard
2. ¬ ¬ P =⇒ P
3. Q =⇒ Q

The defer command moves the first subgoal into the last position.
defer 1
1. ¬ ¬ P =⇒ P
2. Q =⇒ Q
3. hard

Now we apply blast repeatedly to the easy subgoals:

apply blast+
1. hard

Using defer, we have cleared away the trivial parts of the proof so that we
can devote attention to the difficult part.
The prefer command moves the specified subgoal into the first position.
For example, if you suspect that one of your subgoals is invalid (not a theo-
rem), then you should investigate that subgoal first. If it cannot be proved,
then there is no point in proving the other subgoals.
1. ok1
2. ok2
3. doubtful

We decide to work on the third subgoal.

prefer 3
1. doubtful
2. ok1
3. ok2

If we manage to prove doubtful, then we can work on the other subgoals,

confident that we are not wasting our time. Finally we revise the proof script
to remove the prefer command, since we needed it only to focus our ex-
ploration. The previous example is different: its use of defer stops trivial
subgoals from cluttering the rest of the proof. Even there, we should consider
proving hard as a preliminary lemma. Always seek ways to streamline your
proofs.
Summary:
– the control structures +, ? and | help express complicated proofs
– the pr command can limit the number of subgoals to display
– the defer and prefer commands move a subgoal to the last or first position
Exercise 5.17.1 Explain the use of ? and + in this proof.
lemma "[[P∧Q−→R; P−→Q; P]] =⇒ R"
by (drule mp, (intro conjI)?, assumption+)+
 5.18 Proving the Correctness of Euclid’s Algorithm 103

5.18 Proving the Correctness of Euclid’s Algorithm

A brief development will demonstrate the techniques of this chapter, includ-

ing blast applied with additional rules. We shall also see case_tac used to
perform a Boolean case split.
Let us prove that gcd computes the greatest common divisor of its two ar-
guments. We use induction: gcd.induct is the induction rule returned by fun.
We simplify using rules proved in Sect. 3.5.3, since rewriting by the definition
of gcd can loop.
lemma gcd dvd both: "(gcd m n dvd m) ∧ (gcd m n dvd n)"
The induction formula must be a conjunction. In the inductive step, each
conjunct establishes the other.
1. m n. (n 6= 0 =⇒
V
gcd n (m mod n) dvd n ∧
gcd n (m mod n) dvd m mod n) =⇒
gcd m n dvd m ∧ gcd m n dvd n
The conditional induction hypothesis suggests doing a case analysis on
n=0. We apply case_tac with type bool — and not with a datatype, as we have
done until now. Since nat is a datatype, we could have written case_tac n
instead of case_tac "n=0". However, the definition of gcd makes a Boolean
decision:
"gcd m n = (if n=0 then m else gcd n (m mod n))"

Proofs about a function frequently follow the function’s definition, so we

perform case analysis over the same formula.
apply
V (case_tac "n=0")
1. m n. [[n 6= 0 =⇒
gcd n (m mod n) dvd n ∧ gcd n (m mod n) dvd m mod n;
n = 0]]
=⇒ gcd m n dvd m ∧ gcd m n dvd n
2. m n. [[n 6= 0 =⇒
V
gcd n (m mod n) dvd n ∧ gcd n (m mod n) dvd m mod n;
n 6= 0]]
=⇒ gcd m n dvd m ∧ gcd m n dvd n
Simplification leaves one subgoal:
apply
V (simp_all)
1. m n. [[gcd n (m mod n) dvd n ∧ gcd n (m mod n) dvd m mod n;
0 < n]]
=⇒ gcd n (m mod n) dvd m
Here, we can use blast. One of the assumptions, the induction hypothesis, is
a conjunction. The two divides relationships it asserts are enough to prove
the conclusion, for we have the following theorem at our disposal:
[[?k dvd (?m mod ?n); ?k dvd ?n]] =⇒ ?k dvd ?m (dvd_mod_imp_dvd)
104 5. The Rules of the Game

This theorem can be applied in various ways. As an introduction rule, it

would cause backward chaining from the conclusion (namely ?k dvd ?m) to
the two premises, which also involve the divides relation. This process does
not look promising and could easily loop. More sensible is to apply the rule
in the forward direction; each step would eliminate an occurrence of the mod
symbol, so the process must terminate.
apply (blast dest: dvd_mod_imp_dvd)
done
Attaching the dest attribute to dvd_mod_imp_dvd tells blast to use it as de-
struction rule; that is, in the forward direction.
We have proved a conjunction. Now, let us give names to each of the two
halves:
lemmas gcd_dvd1 [iff] = gcd_dvd_both [THEN conjunct1]
lemmas gcd_dvd2 [iff] = gcd_dvd_both [THEN conjunct2]
Here we see lemmas used with the iff attribute, which supplies the new
theorems to the classical reasoner and the simplifier. Recall that THEN is
frequently used with destruction rules; THEN conjunct1 extracts the first half
of a conjunctive theorem. Given gcd_dvd_both it yields
gcd ?m1 ?n1 dvd ?m1
The variable names ?m1 and ?n1 arise because Isabelle renames schematic
variables to prevent clashes. The second lemmas declaration yields
gcd ?m1 ?n1 dvd ?n1

To complete the verification of the gcd function, we must prove that it

returns the greatest of all the common divisors of its arguments. The proof
is by induction, case analysis and simplification.
lemma gcd greatest [rule format]:
"k dvd m −→ k dvd n −→ k dvd gcd m n"
The goal is expressed using HOL implication, −→, because the induction af-
fects the two preconditions. The directive rule_format tells Isabelle to replace
each −→ by =⇒ before storing the eventual theorem. This directive can also
remove outer universal quantifiers, converting the theorem into the usual for-
mat for inference rules. It can replace any series of applications of THEN to
the rules mp and spec. We did not have to write this:
lemma gcd_greatest [THEN mp, THEN mp]:
"k dvd m −→ k dvd n −→ k dvd gcd m n"
Because we are again reasoning about gcd, we perform the same induction
and case analysis as in the previous proof:
1. m n. [[n =
6 0 =⇒
V
k dvd n −→ k dvd m mod n −→ k dvd gcd n (m mod n);
n = 0]]
 5.18 Proving the Correctness of Euclid’s Algorithm 105

=⇒ k dvd m −→ k dvd n −→ k dvd gcd m n

2. m n. [[n 6= 0 =⇒
V
k dvd n −→ k dvd m mod n −→ k dvd gcd n (m mod n);
n 6= 0]]
=⇒ k dvd m −→ k dvd n −→ k dvd gcd m n
Simplification proves both subgoals.
apply (simp_all add: dvd_mod)
done

In the first, where n=0, the implication becomes trivial: k dvd gcd m n goes
to k dvd m. The second subgoal is proved by an unfolding of gcd, using this
rule about divides:
[[?f dvd ?m; ?f dvd ?n]] =⇒ ?f dvd ?m mod ?n (dvd_mod)

The facts proved above can be summarized as a single logical equivalence.

This step gives us a chance to see another application of blast.
theorem gcd greatest iff [iff]:
"(k dvd gcd m n) = (k dvd m ∧ k dvd n)"
by (blast intro!: gcd_greatest intro: dvd_trans)

This theorem concisely expresses the correctness of the gcd function. We state
it with the iff attribute so that Isabelle can use it to remove some occur-
rences of gcd. The theorem has a one-line proof using blast supplied with
two additional introduction rules. The exclamation mark (intro!) signifies
safe rules, which are applied aggressively. Rules given without the exclama-
tion mark are applied reluctantly and their uses can be undone if the search
backtracks. Here the unsafe rule expresses transitivity of the divides relation:
[[?m dvd ?n; ?n dvd ?p]] =⇒ ?m dvd ?p (dvd_trans)

Applying dvd_trans as an introduction rule entails a risk of looping, for it

multiplies occurrences of the divides symbol. However, this proof relies on
transitivity reasoning. The rule gcd greatest is safe to apply aggressively
because it yields simpler subgoals. The proof implicitly uses gcd_dvd1 and
gcd_dvd2 as safe rules, because they were declared using iff.
6. Sets, Functions and Relations

This chapter describes the formalization of typed set theory, which is the
basis of much else in HOL. For example, an inductive definition yields a set,
and the abstract theories of relations regard a relation as a set of pairs. The
chapter introduces the well-known constants such as union and intersection,
as well as the main operations on relations, such as converse, composition
and transitive closure. Functions are also covered. They are not sets in HOL,
but many of their properties concern sets: the range of a function is a set,
and the inverse image of a function maps sets to sets.
This chapter will be useful to anybody who plans to develop a substantial
proof. Sets are convenient for formalizing computer science concepts such
as grammars, logical calculi and state transition systems. Isabelle can prove
many statements involving sets automatically.
This chapter ends with a case study concerning model checking for the
temporal logic CTL. Most of the other examples are simple. The chapter
presents a small selection of built-in theorems in order to point out some key
properties of the various constants and to introduce you to the notation.
Natural deduction rules are provided for the set theory constants, but
they are seldom used directly, so only a few are presented here.

6.1 Sets
HOL’s set theory should not be confused with traditional, untyped set theory,
in which everything is a set. Our sets are typed. In a given set, all elements
have the same type, say τ , and the set itself has type τ set.
We begin with intersection, union and complement. In addition to
the membership relation, there is a symbol for its negation. These points
can be seen below.
Here are the natural deduction rules for intersection. Note the resemblance
to those for conjunction.
[[c ∈ A; c ∈ B]] =⇒ c ∈ A ∩ B (IntI)
c ∈ A ∩ B =⇒ c ∈ A (IntD1)
c ∈ A ∩ B =⇒ c ∈ B (IntD2)
Here are two of the many installed theorems concerning set complement.
Note that it is denoted by a minus sign.
108 6. Sets, Functions and Relations

(c ∈ - A) = (c ∈
/ A) (Compl_iff)
- (A ∪ B) = - A ∩ - B (Compl_Un)
Set difference is the intersection of a set with the complement of another
set. Here we also see the syntax for the empty set and for the universal set.
A ∩ (B - A) = {} (Diff_disjoint)
A ∪ - A = UNIV (Compl_partition)

The subset relation holds between two sets just if every element of one
is also an element of the other. This relation is reflexive. These are its natural
deduction rules:
( x. x ∈ A =⇒ x ∈ B) =⇒ A ⊆ B (subsetI)
V

[[A ⊆ B; c ∈ A]] =⇒ c ∈ B (subsetD)

In harder proofs, you may need to apply subsetD giving a specific term for c.
However, blast can instantly prove facts such as this one:
(A ∪ B ⊆ C) = (A ⊆ C ∧ B ⊆ C) (Un_subset_iff)

Here is another example, also proved automatically:

lemma "(A ⊆ -B) = (B ⊆ -A)"
by blast

This is the same example using ascii syntax, illustrating a pitfall:

lemma "(A <= -B) = (B <= -A)"

The proof fails. It is not a statement about sets, due to overloading; the
relation symbol <= can be any relation, not just subset. In this general form,
the statement is not valid. Putting in a type constraint forces the variables
to denote sets, allowing the proof to succeed:
lemma "((A:: 'a set) <= -B) = (B <= -A)"

Section 8.3 below describes overloading. Incidentally, A ⊆ -B asserts that the

sets A and B are disjoint.
Two sets are equal if they contain the same elements. This is the principle
of extensionality for sets.
( x. (x ∈ A) = (x ∈ B)) =⇒ A = B (set_ext)
V

Extensionality can be expressed as A = B ⇐⇒ (A ⊆ B) ∧ (B ⊆ A). The

following rules express both directions of this equivalence. Proving a set equa-
tion using equalityI allows the two inclusions to be proved independently.
[[A ⊆ B; B ⊆ A]] =⇒ A = B (equalityI)
[[A = B; [[A ⊆ B; B ⊆ A]] =⇒ P]] =⇒ P (equalityE)
 6.1 Sets 109

6.1.1 Finite Set Notation

Finite sets are expressed using the constant insert, which is a form of union:
insert a A = {a} ∪ A (insert_is_Un)

The finite set expression {a,b} abbreviates insert a (insert b {}). Many
facts about finite sets can be proved automatically:
lemma "{a,b} ∪ {c,d} = {a,b,c,d}"
by blast

Not everything that we would like to prove is valid. Consider this attempt:
lemma "{a,b} ∩ {b,c} = {b}"
apply auto

The proof fails, leaving the subgoal b=c. To see why it fails, consider a correct
version:
lemma "{a,b} ∩ {b,c} = (if a=c then {a,b} else {b})"
apply simp
by blast

Our mistake was to suppose that the various items were distinct. Another
remark: this proof uses two methods, namely simp and blast. Calling simp
eliminates the if -then-else expression, which blast cannot break down. The
combined methods (namely force and auto) can prove this fact in one step.

6.1.2 Set Comprehension

The set comprehension {x. P} expresses the set of all elements that satisfy the
predicate P. Two laws describe the relationship between set comprehension
and the membership relation:
(a ∈ {x. P x}) = P a (mem_Collect_eq)
{x. x ∈ A} = A (Collect_mem_eq)

Facts such as these have trivial proofs:

lemma "{x. P x ∨ x ∈ A} = {x. P x} ∪ A"
lemma "{x. P x −→ Q x} = -{x. P x} ∪ {x. Q x}"

Isabelle has a general syntax for comprehension, which is best described

through an example:
lemma "{p*q | p q. p∈prime ∧ q∈prime} =
{z. ∃ p q. z = p*q ∧ p∈prime ∧ q∈prime}"
The left and right hand sides of this equation are identical. The syntax used
in the left-hand side abbreviates the right-hand side: in this case, all numbers
that are the product of two primes. The syntax provides a neat way of ex-
pressing any set given by an expression built up from variables under specific
110 6. Sets, Functions and Relations

constraints. The drawback is that it hides the true form of the expression,
with its existential quantifiers.
Remark. We do not need sets at all. They are essentially equivalent to
predicate variables, which are allowed in higher-order logic. The main benefit
of sets is their notation; we can write x∈A and {z. P} where predicates would
require writing A(x) and λz. P.

6.1.3 Binding Operators

Universal and existential quantifications may range over sets, with the obvi-
ous meaning. Here are the natural deduction rules for the bounded universal
quantifier. Occasionally you will need to apply bspec with an explicit instan-
tiation of the variable x :
( x. x ∈ A =⇒ P x) =⇒ ∀ x∈A. P x (ballI)
V
[[∀ x∈A. P x; x ∈ A]] =⇒ P x (bspec)
Dually, here are the natural deduction rules for the bounded existential quan-
tifier. You may need to apply bexI with an explicit instantiation:
[[P x; x ∈ A]] V
=⇒ ∃ x∈A. P x (bexI)
[[∃ x∈A. P x; x. [[x ∈ A; P x]] =⇒ Q]] =⇒ Q (bexE)
Unions can be formed over the values of a given set. The syntax is
x∈A. B or UN x:A. B in ascii. Indexed union satisfies this basic law:
S

(b ∈ ( x∈A. B x)) = (∃ x∈A. b ∈ B x) (UN_iff)

S

It has two natural deduction rules similar to those for the existential quanti-
fier. Sometimes UN_I must be applied explicitly:
[[a ∈ A;S b ∈ B a]] =⇒V b ∈ ( x∈A. B x) (UN_I)
S
[[b ∈ ( x∈A. B x); x. [[x ∈ A; b ∈ B x]] =⇒ R]] =⇒ R (UN_E)
The following built-in abbreviation (see Sect. 4.1.4) lets us express the union
over a type:
( x. B x) ≡ ( x∈UNIV. B x)
S S

We may also express the union of a set of sets, written Union C in ascii:
(A ∈ C) = (∃ X∈C. A ∈ X) (Union_iff)
S

Intersections are treated dually, although they seem to be used less often
than unions. The syntax below would be INT x: A. B and Inter C in ascii.
Among others, these theorems are available:
(b ∈ T
( x∈A. B x)) = (∀ x∈A. b ∈ B x) (INT_iff)
T
(A ∈ C) = (∀ X∈C. A ∈ X) (Inter_iff)
Isabelle uses logical equivalences such as those above in automatic proof.
Unions, intersections and so forth are not simply replaced by their definitions.
Instead, membership tests are simplified. For example, x ∈ A ∪ B is replaced
by x ∈ A ∨ x ∈ B.
 6.2 Functions 111

The internal form of a comprehension involves the constant Collect,

which occasionally appears when a goal or theorem is displayed. For example,
Collect P is the same term as {x. P x}. The same thing can happen with
quantifiers: All P is ∀ x. P x and Ex P is ∃ x. P x ; also Ball A P is ∀ x∈A. P
x and Bex A P is ∃ x∈A. P x. For indexed unions and intersections, you may
see the constants UNION and INTER . The internal constant for εx.P(x) is Eps.
We have only scratched the surface of Isabelle/HOL’s set theory, which
provides hundreds of theorems for your use.

6.1.4 Finiteness and Cardinality

The predicate finite holds of all finite sets. Isabelle/HOL includes many
familiar theorems about finiteness and cardinality (card). For example, we
have theorems concerning the cardinalities of unions, intersections and the
powerset:
[[finite A; finite B]]
=⇒ card A + card B = card (A ∪ B) + card (A ∩ B) (card_Un_Int)

finite A =⇒ card (Pow A) = 2 ^ card A (card_Pow)

finite A =⇒
card {B. B ⊆ A ∧ card B = k} = card A choose k (n_subsets)

Writing |A| as n, the

last of these theorems says that the number of k-element
subsets of A is nk .

6.2 Functions

This section describes a few concepts that involve functions. Some of the more
important theorems are given along with the names. A few sample proofs
appear. Unlike with set theory, however, we cannot simply state lemmas and
expect them to be proved using blast.

6.2.1 Function Basics

Two functions are equal if they yield equal results given equal arguments.
This is the principle of extensionality for functions:
( x. f x = g x) =⇒ f = g (ext)
V

Function update is useful for modelling machine states. It has the obvi-
ous definition and many useful facts are proved about it. In particular, the
following equation is installed as a simplification rule:
(f(x:=y)) z = (if z = x then y else f z) (fun_upd_apply)
112 6. Sets, Functions and Relations

Two syntactic points must be noted. In (f(x:=y)) z we are applying an up-

dated function to an argument; the outer parentheses are essential. A series
of two or more updates can be abbreviated as shown on the left-hand side of
this theorem:
f(x:=y, x:=z) = f(x:=z) (fun_upd_upd)
Note also that we can write f(x:=z) with only one pair of parentheses when
it is not being applied to an argument.
The identity function and function composition are defined:
id ≡ λx. x (id_def)
f ◦ g ≡ λx. f (g x) (o_def)
Many familiar theorems concerning the identity and composition are proved.
For example, we have the associativity of composition:
f ◦ (g ◦ h) = f ◦ g ◦ h (o_assoc)

6.2.2 Injections, Surjections, Bijections

A function may be injective, surjective or bijective:

inj_on f A ≡ ∀ x∈A. ∀ y∈A. f x = f y −→ x = y (inj_on_def)
surj f ≡ ∀ y. ∃ x. y = f x (surj_def)
bij f ≡ inj f ∧ surj f (bij_def)
The second argument of inj_on lets us express that a function is injective over
a given set. This refinement is useful in higher-order logic, where functions are
total; in some cases, a function’s natural domain is a subset of its domain type.
Writing inj f abbreviates inj_on f UNIV, for when f is injective everywhere.
The operator inv expresses the inverse of a function. In general the
inverse may not be well behaved. We have the usual laws, such as these:
inj f =⇒ inv f (f x) = x (inv_f_f)
surj f =⇒ f (inv f y) = y (surj_f_inv_f)
bij f =⇒ inv (inv f) = f (inv_inv_eq)
Theorems involving these concepts can be hard to prove. The following
example is easy, but it cannot be proved automatically. To begin with, we
need a law that relates the equality of functions to equality over all arguments:
(f = g) = (∀ x. f x = g x) (fun_eq_iff)
This is just a restatement of extensionality. Our lemma states that an injec-
tion can be cancelled from the left side of function composition:
lemma "inj f =⇒ (f o g = f o h) = (g = h)"
apply (simp add: fun_eq_iff inj_on_def)
apply auto
done
The first step of the proof invokes extensionality and the definitions of
injectiveness and composition. It leaves one subgoal:
 6.3 Relations 113

1. ∀ x y. f x = f y −→ x = y =⇒
(∀ x. f (g x) = f (h x)) = (∀ x. g x = h x)
This can be proved using the auto method.

6.2.3 Function Image

The image of a set under a function is a most useful notion. It has the
obvious definition:
f ` A ≡ {y. ∃ x∈A. y = f x} (image_def)

Here are some of the many facts proved about image:

(f ◦ g) ` r = f ` g ` r (image_compose)
f`(A ∪ B) = f`A ∪ f`B (image_Un)
inj f =⇒ f`(A ∩ B) = f`A ∩ f`B (image_Int)

Laws involving image can often be proved automatically. Here are two
examples, illustrating connections with indexed union and with the general
syntax for comprehension:
lemma "f`A ∪ g`A = ( x∈A. {f x, g x})"
S

lemma "f ` {(x,y). P x y} = {f(x,y) | x y. P x y}"

A function’s range is the set of values that the function can take on. It
is, in fact, the image of the universal set under that function. There is no
constant range. Instead, range abbreviates an application of image to UNIV :
range f f`UNIV

Few theorems are proved specifically for range; in most cases, you should look
for a more general theorem concerning images.
Inverse image is also useful. It is defined as follows:
f -` B ≡ {x. f x ∈ B} (vimage_def)

This is one of the facts proved about it:

f -` (- A) = - f -` A (vimage_Compl)

6.3 Relations

A relation is a set of pairs. As such, the set operations apply to them. For
instance, we may form the union of two relations. Other primitives are defined
specifically for relations.
114 6. Sets, Functions and Relations

6.3.1 Relation Basics

The identity relation, also known as equality, has the obvious definition:
Id ≡ {p. ∃ x. p = (x,x)} (Id_def)
Composition of relations (the infix O ) is also available:
r O s = {(x,z). ∃ y. (x,y) ∈ s ∧ (y,z) ∈ r} (relcomp_unfold)
This is one of the many lemmas proved about these concepts:
R O Id = R (R_O_Id)
Composition is monotonic, as are most of the primitives appearing in this
chapter. We have many theorems similar to the following one:
[[r' ⊆ r; s' ⊆ s]] =⇒ r' O s' ⊆ r O s (relcomp_mono)
The converse or inverse of a relation exchanges the roles of the two
operands. We use the postfix notation r −1 or r^-1 in ASCII.
((a,b) ∈ r −1 ) = ((b,a) ∈ r) (converse_iff)
Here is a typical law proved about converse and composition:
(r O s)−1 = s−1 O r −1 (converse_relcomp)
The image of a set under a relation is defined analogously to image under
a function:
(b ∈ r `` A) = (∃ x∈A. (x,b) ∈ r) (Image_iff)
It satisfies many similar laws.
The domain and range of a relation are defined in the standard way:
(a ∈ Domain r) = (∃ y. (a,y) ∈ r) (Domain_iff)
(a ∈ Range r) = (∃ y. (y,a) ∈ r) (Range_iff)
Iterated composition of a relation is available. The notation overloads that
of exponentiation. Two simplification rules are installed:
R ^ 0 = Id
R ^ Suc n = R O R^n

6.3.2 The Reflexive and Transitive Closure

The reflexive and transitive closure of the relation r is written with a

postfix syntax. In ASCII we write r^* and in symbol notation r ∗ . It is the
least solution of the equation
r ∗ = Id ∪ (r O r ∗ ) (rtrancl_unfold)
Among its basic properties are three that serve as introduction rules:
(a, a) ∈ r ∗ (rtrancl_refl)
p ∈ r =⇒ p ∈ r ∗ (r_into_rtrancl)
[[(a,b) ∈ r ∗ ; (b,c) ∈ r ∗ ]] =⇒ (a,c) ∈ r ∗ (rtrancl_trans)
 6.3 Relations 115

Induction over the reflexive transitive closure is available:

[[(a, b) ∈ r ∗ ; P a; y z. [[(a, y) ∈ r ∗ ; (y, z) ∈ r; P y]] =⇒ P z]]
V
=⇒ P b (rtrancl_induct)

Idempotence is one of the laws proved about the reflexive transitive closure:
(r ∗ )∗ = r ∗ (rtrancl_idemp)

The transitive closure is similar. The ASCII syntax is r^+. It has two
introduction rules:
p ∈ r =⇒ p ∈ r + (r_into_trancl)
[[(a, b) ∈ r + ; (b, c) ∈ r + ]] =⇒ (a, c) ∈ r + (trancl_trans)

The induction rule resembles the one shown above. A typical lemma states
that transitive closure commutes with the converse operator:
(r −1 )+ = (r + )−1 (trancl_converse)

6.3.3 A Sample Proof

The reflexive transitive closure also commutes with the converse operator. Let
us examine the proof. Each direction of the equivalence is proved separately.
The two proofs are almost identical. Here is the first one:
lemma rtrancl_converseD: "(x,y) ∈ (r −1 )∗ =⇒ (y,x) ∈ r ∗ "
apply (erule rtrancl_induct)
apply (rule rtrancl_refl)
apply (blast intro: rtrancl_trans)
done

The first step of the proof applies induction, leaving these subgoals:
1. V
(x, x) ∈ r ∗
2. y z. [[(x,y) ∈ (r −1 )∗ ; (y,z) ∈ r −1 ; (y,x) ∈ r ∗ ]]
=⇒ (z,x) ∈ r ∗

The first subgoal is trivial by reflexivity. The second follows by first elimi-
nating the converse operator, yielding the assumption (z,y) ∈ r, and then
applying the introduction rules shown above. The same proof script handles
the other direction:
lemma rtrancl_converseI: "(y,x) ∈ r ∗ =⇒ (x,y) ∈ (r −1 )∗ "
apply (erule rtrancl_induct)
apply (rule rtrancl_refl)
apply (blast intro: rtrancl_trans)
done

Finally, we combine the two lemmas to prove the desired equation:

lemma rtrancl_converse: "(r −1 )∗ = (r ∗ )−1 "
by (auto intro: rtrancl_converseI dest: rtrancl_converseD)
116 6. Sets, Functions and Relations

!! This trivial proof requires auto rather than blast because of a subtle issue
involving ordered pairs. Here is a subgoal that arises internally after the rules
equalityI and subsetI have been applied:
1. x. x ∈ (r −1 )∗ =⇒ x ∈ (r ∗ )−1
V

We cannot apply rtrancl_converseD. It refers to ordered pairs, while x is a variable

of product type. The simp and blast methods can do nothing, so let us try clarify:
1. a b. (a,b) ∈ (r −1 )∗ =⇒ (b,a) ∈ r ∗
V

Now that x has been replaced by the pair (a,b), we can proceed. Other methods
that split variables in this way are force, auto, fast and best. Section 8.1 will
discuss proof techniques for ordered pairs in more detail.

6.4 Well-Founded Relations and Induction

A well-founded relation captures the notion of a terminating process. Com-

plex recursive functions definitions must specify a well-founded relation that
justifies their termination [17]. Most of the forms of induction found in math-
ematics are merely special cases of induction over a well-founded relation.
Intuitively, the relation ≺ is well-founded if it admits no infinite de-
scending chains
· · · ≺ a2 ≺ a1 ≺ a0 .
Well-foundedness can be hard to show. The various formulations are all com-
plicated. However, often a relation is well-founded by construction. HOL pro-
vides theorems concerning ways of constructing a well-founded relation. The
most familiar way is to specify a measure function f into the natural num-
bers, when x ≺ y ⇐⇒ f x < f y; we write this particular relation as
measure f.

!! You may want to skip the rest of this section until you need to perform a
complex recursive function definition or induction. The induction rule returned
by fun is good enough for most purposes. We use an explicit well-founded induction
only in Sect. 9.2.4.

Isabelle/HOL declares less_than as a relation object, that is, a set of

pairs of natural numbers. Two theorems tell us that this relation behaves as
expected and that it is well-founded:
((x,y) ∈ less_than) = (x < y) (less_than_iff)
wf less_than (wf_less_than)
The notion of measure generalizes to the inverse image of a relation.
Given a relation r and a function f, we express a new relation using f as a
measure. An infinite descending chain on this new relation would give rise
to an infinite descending chain on r. Isabelle/HOL defines this concept and
proves a theorem stating that it preserves well-foundedness:
 6.5 Fixed Point Operators 117

inv_image r f ≡ {(x,y). (f x, f y) ∈ r} (inv_image_def)

wf r =⇒ wf (inv_image r f) (wf_inv_image)
A measure function involves the natural numbers. The relation measure
size justifies primitive recursion and structural induction over a datatype.
Isabelle/HOL defines measure as shown:
measure ≡ inv_image less_than (measure_def)
wf (measure f) (wf_measure)
Of the other constructions, the most important is the lexicographic
product of two relations. It expresses the standard dictionary ordering over
pairs. We write ra <*lex*> rb, where ra and rb are the two operands. The
lexicographic product satisfies the usual definition and it preserves well-
foundedness:
ra <*lex*> rb ≡
{((a,b),(a',b')). (a,a') ∈ ra ∨
a=a' ∧ (b,b') ∈ rb} (lex_prod_def)
[[wf ra; wf rb]] =⇒ wf (ra <*lex*> rb) (wf_lex_prod)
The multiset ordering, useful for hard termination proofs, is available
in the Library [4]. Baader and Nipkow [3, Sect. 2.5] discuss it.
Induction comes in many forms, including traditional mathematical in-
duction, structural induction on lists and induction on size. All are instances
of the following rule, for a suitable well-founded relation ≺:
[∀y. y ≺ x → P(y)]
..
..
P(x)
P(a)
To show P(a) for a particular term a, it suffices to show P(x) for arbi-
trary x under the assumption that P(y) holds for y ≺ x. Intuitively, the
well-foundedness of ≺ ensures that the chains of reasoning are finite.
In Isabelle, the induction rule is expressed like this:
[[wf r; x. ∀ y. (y,x) ∈ r −→ P y =⇒ P x]] =⇒ P a (wf_induct)
V

Here wf r expresses that the relation r is well-founded.

Many familiar induction principles are instances of this rule. For exam-
ple, the predecessor relation on the natural numbers is well-founded; induc-
tion over it is mathematical induction. The “tail of” relation on lists is well-
founded; induction over it is structural induction.

6.5 Fixed Point Operators

Fixed point operators define sets recursively. They are invoked implicitly
when making an inductive definition, as discussed in Chap. 7 below. However,
118 6. Sets, Functions and Relations

they can be used directly, too. The least or strongest fixed point yields an
inductive definition; the greatest or weakest fixed point yields a coinductive
definition. Mathematicians may wish to note that the existence of these fixed
points is guaranteed by the Knaster-Tarski theorem.

!! Casual readers should skip the rest of this section. We use fixed point operators
only in Sect. ??.

The theory applies only to monotonic functions. Isabelle’s definition of

monotone is overloaded over all orderings:
mono f ≡ ∀ A B. A ≤ B −→ f A ≤ f B (mono_def)

For fixed point operators, the ordering will be the subset relation: if A ⊆ B
then we expect f (A) ⊆ f (B). In addition to its definition, monotonicity has
the obvious introduction and destruction rules:
( A B. A ≤ B =⇒ f A ≤ f B) =⇒ mono f (monoI)
V

[[mono f; A ≤ B]] =⇒ f A ≤ f B (monoD)

The most important properties of the least fixed point are that it is a
fixed point and that it enjoys an induction rule:
mono f =⇒ lfp f = f (lfp f) (lfp_unfold)
[[aV ∈ lfp f; mono f;
x. x ∈ f (lfp f ∩ {x. P x}) =⇒ P x]] =⇒ P a (lfp_induct)

The induction rule shown above is more convenient than the basic one derived
from the minimality of lfp . Observe that both theorems demand mono f as
a premise.
The greatest fixed point is similar, but it has a coinduction rule:
mono f =⇒ gfp f = f (gfp f) (gfp_unfold)
[[mono f; a ∈ X; X ⊆ f (X ∪ gfp f)]] =⇒ a ∈ gfp f (coinduct)

A bisimulation is perhaps the best-known concept defined as a greatest

fixed point. Exhibiting a bisimulation to prove the equality of two agents in
a process algebra is an example of coinduction. The coinduction rule can be
strengthened in various ways.

6.5.1 Propositional Dynamic Logic — PDL

The formulae of PDL are built up from atomic propositions via negation and
conjunction and the two temporal connectives AX and EF . Since formulae are
essentially syntax trees, they are naturally modelled as a datatype:1
datatype formula = Atom "atom"
| Neg formula
| And formula formula
1
The customary definition of PDL [12] looks quite different from ours, but the
two are easily shown to be equivalent.
 6.5 Fixed Point Operators 119

| AX formula
| EF formula
This resembles the boolean expression case study in Sect. 2.5.6. A validity
relation between states and formulae specifies the semantics. The syntax an-
notation allows us to write s |= f instead of valid s f . The definition is by
recursion over the syntax:
primrec valid :: "state ⇒ formula ⇒ bool" ("(_ |= _)" [80,80] 80)
where
"s |= Atom a = (a ∈ L s)" |
"s |= Neg f = (¬(s |= f))" |
"s |= And f g = (s |= f ∧ s |= g)" |
"s |= AX f = (∀ t. (s,t) ∈ M −→ t |= f)" |
"s |= EF f = (∃ t. (s,t) ∈ M ∗ ∧ t |= f)"

The first three equations should be self-explanatory. The temporal formula

AX f means that f is true in All neX t states whereas EF f means that there
Exists some F uture state in which f is true. The future is expressed via ∗ ,
the reflexive transitive closure. Because of reflexivity, the future includes the
present.
Now we come to the model checker itself. It maps a formula into the set
of states where the formula is true. It too is defined by recursion over the
syntax:
primrec mc :: "formula ⇒ state set" where
"mc(Atom a) = {s. a ∈ L s}" |
"mc(Neg f) = -mc f" |
"mc(And f g) = mc f ∩ mc g" |
"mc(AX f) = {s. ∀ t. (s,t) ∈ M −→ t ∈ mc f}" |
"mc(EF f) = lfp(λT. mc f ∪ (M −1 `` T))"

Only the equation for EF deserves some comments. Remember that the postfix
−1
and the infix `` are predefined and denote the converse of a relation and
the image of a set under a relation. Thus M −1 `` T is the set of all predecessors
of T and the least fixed point (lfp ) of λT. mc f ∪ M −1 `` T is the least set
T containing mc f and all predecessors of T. If you find it hard to see that
mc (EF f) contains exactly those states from which there is a path to a state
where f is true, do not worry — this will be proved in a moment.
First we prove monotonicity of the function inside lfp in order to make
sure it really has a least fixed point.
lemma mono_ef: "mono(λT. A ∪ (M −1 `` T))"
apply(rule monoI)
apply blast
done
Now we can relate model checking and semantics. For the EF case we need a
separate lemma:
lemma EF_lemma:
"lfp(λT. A ∪ (M −1 `` T)) = {s. ∃ t. (s,t) ∈ M ∗ ∧ t ∈ A}"
120 6. Sets, Functions and Relations

The equality is proved in the canonical fashion by proving that each set
includes the other; the inclusion is shown pointwise:
apply(rule equalityI)
apply(rule subsetI)
apply(simp)
Simplification leaves us with the following first subgoal
1. s. s ∈ lfp (λT. A ∪ M −1 `` T) =⇒ ∃ t. (s, t) ∈ M ∗ ∧ t ∈ A
V
A total of 2 subgoals...

which is proved by lfp -induction:

apply(erule lfp_induct_set)
apply(rule mono_ef)
apply(simp)
Having disposed of the monotonicity subgoal, simplification leaves us with
the following goal:
1. x. x ∈ A ∨
V
x ∈ M −1 `` (lfp (…) ∩ {x. ∃ t. (x, t) ∈ M ∗ ∧ t ∈ A})
=⇒ ∃ t. (x, t) ∈ M ∗ ∧ t ∈ A

It is proved by blast, using the transitivity of M ∗ .

apply(blast intro: rtrancl_trans)

We now return to the second set inclusion subgoal, which is again proved
pointwise:
apply(rule subsetI)
apply(simp, clarify)
After simplification and clarification we are left with
1. x t. [[(x, t) ∈ M ∗ ; t ∈ A]] =⇒ x ∈ lfp (λT. A ∪ M −1 `` T)
V

This goal is proved by induction on (s, t) ∈ M ∗ . But since the model

checker works backwards (from t to s), we cannot use the induction theorem
rtrancl_induct: it works in the forward direction. Fortunately the converse
induction theorem converse_rtrancl_induct already exists:
V b) ∈ r ; P b;
∗
[[(a,
y z. [[(y, z) ∈ r; (z, b) ∈ r ∗ ; P z]] =⇒ P y]]
=⇒ P a

It says that if (a, b) ∈ r ∗ and we know P b then we can infer P a provided

each step backwards from a predecessor z of b preserves P.
apply(erule converse_rtrancl_induct)
The base case
1. x t. t ∈ A =⇒ t ∈ lfp (λT. A ∪ M −1 `` T)
V
A total of 2 subgoals...
 6.5 Fixed Point Operators 121

is solved by unrolling lfp once

apply(subst lfp_unfold[OF mono_ef])

1. x t. t ∈ A =⇒ t ∈ A ∪ M −1 `` lfp (λT. A ∪ M −1 `` T)
V
A total of 2 subgoals...

and disposing of the resulting trivial subgoal automatically:

apply(blast)
The proof of the induction step is identical to the one for the base case:
apply(subst lfp_unfold[OF mono_ef])
apply(blast)
done
The main theorem is proved in the familiar manner: induction followed
by auto augmented with the lemma as a simplification rule.
theorem "mc f = {s. s |= f}"
apply(induct_tac f)
apply(auto simp add: EF_lemma)
done

Exercise 6.5.1 AX has a dual operator EN (“there exists a next state such
that”)2 with the intended semantics
s |= EN f = (∃ t. (s, t) ∈ M ∧ t |= f)
Fortunately, EN f can already be expressed as a PDL formula. How?
Show that the semantics for EF satisfies the following recursion equation:
s |= EF f = (s |= f ∨ s |= EN (EF f))

6.5.2 Computation Tree Logic — CTL

The semantics of PDL only needs reflexive transitive closure. Let us be ad-
venturous and introduce a more expressive temporal operator. We extend the
datatype formula by a new constructor
| AF formula
which stands for “Always in the Future”: on all infinite paths, at some point
the formula holds. Formalizing the notion of an infinite path is easy in HOL:
it is simply a function from nat to state.
definition Paths :: "state ⇒ (nat ⇒ state)set" where
"Paths s ≡ {p. s = p 0 ∧ (∀ i. (p i, p(i+1)) ∈ M)}"
3
This definition allows a succinct statement of the semantics of AF :
2
We cannot use the customary EX: it is reserved as the ascii-equivalent of ∃ .
3
Do not be misled: neither datatypes nor recursive functions can be extended
by new constructors or equations. This is just a trick of the presentation (see
Sect. 4.2.5). In reality one has to define a new datatype and a new function.
122 6. Sets, Functions and Relations

"s |= AF f = (∀ p ∈ Paths s. ∃ i. p i |= f)"

Model checking AF involves a function which is just complicated enough to
warrant a separate definition:
definition af :: "state set ⇒ state set ⇒ state set" where
"af A T ≡ A ∪ {s. ∀ t. (s, t) ∈ M −→ t ∈ T}"

Now we define mc (AF f) as the least set T that includes mc f and all states
all of whose direct successors are in T :
"mc(AF f) = lfp(af(mc f))"
Because af is monotone in its second argument (and also its first, but that
is irrelevant), af A has a least fixed point:
lemma mono_af: "mono(af A)"
apply(simp add: mono_def af_def)
apply blast
done

All we need to prove now is mc (AF f) = {s. s |= AF f}, which states

that mc and |= agree for AF . This time we prove the two inclusions separately,
starting with the easy one:
theorem AF_lemma1: "lfp(af A) ⊆ {s. ∀ p ∈ Paths s. ∃ i. p i ∈ A}"

In contrast to the analogous proof for EF, and just for a change, we do not
use fixed point induction. Park-induction, named after David Park, is weaker
but sufficient for this proof:
f S ≤ S =⇒ lfp f ≤ S (lfp_lowerbound)
The instance of the premise f S ⊆ S is proved pointwise, a decision that auto
takes for us:
apply(rule lfp_lowerbound)
apply(auto simp add: af_def Paths_def)

1. p. [[∀ t. (p 0, t) ∈ M −→
V
(∀ p. t = p 0 ∧ (∀ i. (p i, p (Suc i)) ∈ M) −→
(∃ i. p i ∈ A));
∀ i. (p i, p (Suc i)) ∈ M ]]
=⇒ ∃ i. p i ∈ A

In this remaining case, we set t to p 1. The rest is automatic, which is sur-

prising because it involves finding the instantiation λi. p (i + 1) for ∀ p.
apply(erule_tac x = "p 1" in allE)
apply(auto)
done
The opposite inclusion is proved by contradiction: if some state s is not in
lfp (af A), then we can construct an infinite A-avoiding path starting from s.
The reason is that by unfolding lfp we find that if s is not in lfp (af A),
then s is not in A and there is a direct successor of s that is again not in
 6.5 Fixed Point Operators 123

lfp (af A). Iterating this argument yields the promised infinite A-avoiding
path. Let us formalize this sketch.
The one-step argument in the sketch above is proved by a variant of
contraposition:
lemma not_in_lfp_afD:
"s ∈
/ lfp(af A) =⇒ s ∈/ A ∧ (∃ t. (s,t) ∈ M ∧ t ∈
/ lfp(af A))"
apply(erule contrapos_np)
apply(subst lfp_unfold[OF mono_af])
apply(simp add: af_def)
done

We assume the negation of the conclusion and prove s ∈ lfp (af A). Unfold-
ing lfp once and simplifying with the definition of af finishes the proof.
Now we iterate this process. The following construction of the desired
path is parameterized by a predicate Q that should hold along the path:
primrec path :: "state ⇒ (state ⇒ bool) ⇒ (nat ⇒ state)" where
"path s Q 0 = s" |
"path s Q (Suc n) = (SOME t. (path s Q n,t) ∈ M ∧ Q t)"
Element n + 1 on this path is some arbitrary successor t of element n such
that Q t holds. Remember that SOME t. R t is some arbitrary but fixed t
such that R t holds (see Sect. 5.10). Of course, such a t need not exist, but
that is of no concern to us since we will only use path when a suitable t does
exist.
Let us show that if each state s that satisfies Q has a successor that again
satisfies Q, then there exists an infinite Q -path:
lemma infinity_lemma:
"[[ Q s; ∀ s. Q s −→ (∃ t. (s,t) ∈ M ∧ Q t) ]] =⇒
∃ p∈Paths s. ∀ i. Q(p i)"

First we rephrase the conclusion slightly because we need to prove simulta-

neously both the path property and the fact that Q holds:
apply(subgoal_tac
"∃ p. s = p 0 ∧ (∀ i::nat. (p i, p(i+1)) ∈ M ∧ Q(p i))")
From this proposition the original goal follows easily:
apply(simp add: Paths_def, blast)
The new subgoal is proved by providing the witness path s Q for p :
apply(rule_tac x = "path s Q" in exI)
apply(clarsimp)

After simplification and clarification, the subgoal has the following form:
1. i. [[Q s; ∀ s. Q s −→ (∃ t. (s, t) ∈ M ∧ Q t)]]
V
=⇒ (path s Q i, SOME t. (path s Q i, t) ∈ M ∧ Q t) ∈ M ∧
Q (path s Q i)
It invites a proof by induction on i:
124 6. Sets, Functions and Relations

apply(induct_tac i)
apply(simp)
After simplification, the base case boils down to
1. [[Q s; ∀ s. Q s −→ (∃ t. (s, t) ∈ M ∧ Q t)]]
=⇒ (s, SOME t. (s, t) ∈ M ∧ Q t) ∈ M
A total of 2 subgoals...
The conclusion looks exceedingly trivial: after all, t is chosen such that (s,
t) ∈ M holds. However, we first have to show that such a t actually exists!
This reasoning is embodied in the theorem someI2_ex :
[[∃ a. ?P a; x. ?P x =⇒ ?Q x]] =⇒ ?Q (SOME x. ?P x)
V

When we apply this theorem as an introduction rule, ?P x becomes (s, x) ∈

M ∧ Q x and ?Q x becomes (s, x) ∈ M and we have to prove two subgoals:
∃ a. (s, a) ∈ M ∧ Q a, which follows from the assumptions, and (s, x) ∈ M
∧ Q x =⇒ (s, x) ∈ M, which is trivial. Thus it is not surprising that fast
can prove the base case quickly:
apply(fast intro: someI2_ex)
What is worth noting here is that we have used fast rather than blast. The
reason is that blast would fail because it cannot cope with someI2_ex : unify-
ing its conclusion with the current subgoal is non-trivial because of the nested
schematic variables. For efficiency reasons blast does not even attempt such
unifications. Although fast can in principle cope with complicated unifica-
tion problems, in practice the number of unifiers arising is often prohibitive
and the offending rule may need to be applied explicitly rather than auto-
matically. This is what happens in the step case.
The induction step is similar, but more involved, because now we face
nested occurrences of SOME. As a result, fast is no longer able to solve the
subgoal and we apply someI2_ex by hand. We merely show the proof com-
mands but do not describe the details:
apply(simp)
apply(rule someI2_ex)
apply(blast)
apply(rule someI2_ex)
apply(blast)
apply(blast)
done
Function path has fulfilled its purpose now and can be forgotten. It was
merely defined to provide the witness in the proof of the infinity_lemma.
Aficionados of minimal proofs might like to know that we could have given
the witness without having to define a new function: the term
rec_nat s (λn t. SOME u. (t, u) ∈ M ∧ Q u)
is extensionally equal to path s Q, where rec_nat is the predefined primitive
recursor on nat.
 6.5 Fixed Point Operators 125

At last we can prove the opposite direction of AF_lemma1:

theorem AF_lemma2: "{s. ∀ p ∈ Paths s. ∃ i. p i ∈ A} ⊆ lfp(af A)"

The proof is again pointwise and then by contraposition:

apply(rule subsetI)
apply(erule contrapos_pp)
apply simp

1. x. x ∈
/ lfp (af A) =⇒ ∃ p∈Paths x. ∀ i. p i ∈
/ A
V

Applying the infinity_lemma as a destruction rule leaves two subgoals, the

second premise of infinity_lemma and the original subgoal:
apply(drule infinity_lemma)

1. Vx. ∀ s. s ∈
/ lfp (af A) −→ (∃ t. (s, t) ∈ M ∧ t ∈
/ lfp (af A))
V
2. x. ∃ p∈Paths x. ∀ i. p i ∈/ lfp (af A) =⇒
∃ p∈Paths x. ∀ i. p i ∈
/ A

Both are solved automatically:

apply(auto dest: not_in_lfp_afD)
done

If you find these proofs too complicated, we recommend that you read
Sect. 9.2.4, where we show how inductive definitions lead to simpler argu-
ments.
The main theorem is proved as for PDL, except that we also derive the
necessary equality lfp(af A) = ... by combining AF_lemma1 and AF_lemma2
on the spot:
theorem "mc f = {s. s |= f}"
apply(induct_tac f)
apply(auto simp add: EF_lemma equalityI[OF AF_lemma1 AF_lemma2])
done

The language defined above is not quite CTL. The latter also includes
an until-operator EU f g with semantics “there Exists a path where f is true
U ntil g becomes true”. We need an auxiliary function:
primrec
until:: "state set ⇒ state set ⇒ state ⇒ state list ⇒ bool" where
"until A B s [] = (s ∈ B)" |
"until A B s (t#p) = (s ∈ A ∧ (s,t) ∈ M ∧ until A B t p)"

Expressing the semantics of EU is now straightforward:

s |= EU f g = (∃ p. until {t. t |= f} {t. t |= g} s p)

Note that EU is not definable in terms of the other operators!

Model checking EU is again a least fixed point construction:
mc(EU f g) = lfp(λT. mc g ∪ mc f ∩ (M −1 `` T))
126 6. Sets, Functions and Relations

Exercise 6.5.2 Extend the datatype of formulae by the above until operator
and prove the equivalence between semantics and model checking, i.e. that
mc (EU f g) = {s. s |= EU f g}

For more CTL exercises see, for example, Huth and Ryan [15].
Let us close this section with a few words about the executability of our
model checkers. It is clear that if all sets are finite, they can be represented as
lists and the usual set operations are easily implemented. Only lfp requires
a little thought. Fortunately, theory While_Combinator in the Library [4] pro-
vides a theorem stating that in the case of finite sets and a monotone func-
tion F, the value of lfp F can be computed by iterated application of F to {}
until a fixed point is reached. It is actually possible to generate executable
functional programs from HOL definitions, but that is beyond the scope of
the tutorial.
7. Inductively Defined Sets

This chapter is dedicated to the most important definition principle after

recursive functions and datatypes: inductively defined sets.
We start with a simple example: the set of even numbers. A slightly
more complicated example, the reflexive transitive closure, is the subject
of Sect. 7.2. In particular, some standard induction heuristics are discussed.
Advanced forms of inductive definitions are discussed in Sect. 7.3. To demon-
strate the versatility of inductive definitions, the chapter closes with a case
study from the realm of context-free grammars. The first two sections are
required reading for anybody interested in mathematical modelling.

!! Predicates can also be defined inductively. See Sect. 7.1.7.

7.1 The Set of Even Numbers

The set of even numbers can be inductively defined as the least set containing
0 and closed under the operation +2. Obviously, even can also be expressed
using the divides relation (dvd). We shall prove below that the two formula-
tions coincide. On the way we shall examine the primary means of reasoning
about inductively defined sets: rule induction.

7.1.1 Making an Inductive Definition

Using inductive set, we declare the constant even to be a set of natural

numbers with the desired properties.
inductive set even :: "nat set" where
zero[intro!]: "0 ∈ even" |
step[intro!]: "n ∈ even =⇒ (Suc (Suc n)) ∈ even"
An inductive definition consists of introduction rules. The first one above
states that 0 is even; the second states that if n is even, then so is n + 2.
Given this declaration, Isabelle generates a fixed point definition for Even.even
and proves theorems about it, thus following the definitional approach (see
Sect. 2.8). These theorems include the introduction rules specified in the dec-
laration, an elimination rule for case analysis and an induction rule. We can
128 7. Inductively Defined Sets

refer to these theorems by automatically-generated names. Here are two ex-

amples:
0 ∈ Even.even (even.zero)
n ∈ Even.even =⇒ Suc (Suc n) ∈ Even.even (even.step)
The introduction rules can be given attributes. Here both rules are spec-
ified as intro!, directing the classical reasoner to apply them aggressively.
Obviously, regarding 0 as even is safe. The step rule is also safe because n + 2
is even if and only if n is even. We prove this equivalence later.

7.1.2 Using Introduction Rules

Our first lemma states that numbers of the form 2 × k are even. Introduction
rules are used to show that specific values belong to the inductive set. Such
proofs typically involve induction, perhaps over some other inductive set.
lemma two_times_even[intro!]: "2*k ∈ even"
apply (induct_tac k)
apply auto
done
The first step is induction on the natural number k, which leaves two subgoals:
1. V
2 * 0 ∈ Even.even
2. n. 2 * n ∈ Even.even =⇒ 2 * Suc n ∈ Even.even
Here auto simplifies both subgoals so that they match the introduction rules,
which are then applied automatically.
Our ultimate goal is to prove the equivalence between the traditional
definition of even (using the divides relation) and our inductive definition.
One direction of this equivalence is immediate by the lemma just proved,
whose intro! attribute ensures it is applied automatically.
lemma dvd_imp_even: "2 dvd n =⇒ n ∈ even"
by (auto simp add: dvd_def)

7.1.3 Rule Induction

From the definition of the set Even.even, Isabelle has generated an induction
rule:
[[xV ∈ Even.even; P 0;
n. [[n ∈ Even.even; P n]]
=⇒ P (Suc (Suc n))]]
=⇒ P x (even.induct)
A property P holds for every even number provided it holds for 0 and is closed
under the operation Suc(Suc ·). Then P is closed under the introduction rules
for Even.even, which is the least set closed under those rules. This type of
inductive argument is called rule induction.
 7.1 The Set of Even Numbers 129

Apart from the double application of Suc, the induction rule above resem-
bles the familiar mathematical induction, which indeed is an instance of rule
induction; the natural numbers can be defined inductively to be the least set
containing 0 and closed under Suc.
Induction is the usual way of proving a property of the elements of an
inductively defined set. Let us prove that all members of the set Even.even
are multiples of two.
lemma even_imp_dvd: "n ∈ even =⇒ 2 dvd n"
We begin by applying induction. Note that even.induct has the form of
an elimination rule, so we use the method erule. We get two subgoals:
apply (erule even.induct)

1. V
semiring_parity_class.even 0
2. n. [[n ∈ Even.even; semiring_parity_class.even n]]
=⇒ semiring_parity_class.even (Suc (Suc n))

We unfold the definition of dvd in both subgoals, proving the first one and
simplifying the second:
apply (simp_all add: dvd_def)

1. n. [[n ∈ Even.even; ∃ k. n = 2 * k]] =⇒ ∃ k. Suc (Suc n) = 2 * k

V

The next command eliminates the existential quantifier from the assumption
and replaces n by 2 * k.
apply clarify

1. n k. 2 * k ∈ Even.even =⇒ ∃ ka. Suc (Suc (2 * k)) = 2 * ka

V

To conclude, we tell Isabelle that the desired value is Suc k. With this hint,
the subgoal falls to simp.
apply (rule_tac x = "Suc k" in exI, simp)
Combining the previous two results yields our objective, the equivalence
relating Even.even and dvd.
theorem even_iff_dvd: "(n ∈ even) = (2 dvd n)"
by (blast intro: dvd_imp_even even_imp_dvd)

7.1.4 Generalization and Rule Induction

Before applying induction, we typically must generalize the induction for-

mula. With rule induction, the required generalization can be hard to find
and sometimes requires a complete reformulation of the problem. In this ex-
ample, our first attempt uses the obvious statement of the result. It fails:
lemma "Suc (Suc n) ∈ even =⇒ n ∈ even"
130 7. Inductively Defined Sets

apply (erule even.induct)

oops
Rule induction finds no occurrences of Suc (Suc n) in the conclusion,
which it therefore leaves unchanged. (Look at even.induct to see why this
happens.) We have these subgoals:
1. V
n ∈ Even.even
2. na. [[na ∈ Even.even; n ∈ Even.even]] =⇒ n ∈ Even.even
The first one is hopeless. Rule induction on a non-variable term discards
information, and usually fails. How to deal with such situations in general is
described in Sect. 9.2.1 below. In the current case the solution is easy because
we have the necessary inverse, subtraction:
lemma even_imp_even_minus_2: "n ∈ even =⇒ n - 2 ∈ even"
apply (erule even.induct)
apply auto
done
This lemma is trivially inductive. Here are the subgoals:
1. V
0 - 2 ∈ Even.even
2. n. [[n ∈ Even.even; n - 2 ∈ Even.even]] =⇒ Suc (Suc n) - 2 ∈ Even.even
The first is trivial because 0 - 2 simplifies to 0, which is even. The second is
trivial too: Suc (Suc n) - 2 simplifies to n, matching the assumption.
Using our lemma, we can easily prove the result we originally wanted:
lemma Suc_Suc_even_imp_even: "Suc (Suc n) ∈ even =⇒ n ∈ even"
by (drule even_imp_even_minus_2, simp)
We have just proved the converse of the introduction rule even.step. This
suggests proving the following equivalence. We give it the iff attribute be-
cause of its obvious value for simplification.
lemma [iff]: "((Suc (Suc n)) ∈ even) = (n ∈ even)"
by (blast dest: Suc_Suc_even_imp_even)

7.1.5 Rule Inversion

Case analysis on an inductive definition is called rule inversion. It is fre-
quently used in proofs about operational semantics. It can be highly effective
when it is applied automatically. Let us look at how rule inversion is done in
Isabelle/HOL.
Recall that Even.even is the minimal set closed under these two rules:
0 ∈ Even.even
n ∈ Even.even =⇒ Suc (Suc n) ∈ Even.even
Minimality means that Even.even contains only the elements that these rules
force it to contain. If we are told that a belongs to Even.even then there are
only two possibilities. Either a is 0 or else a has the form Suc (Suc n), for
some suitable n that belongs to Even.even. That is the gist of the cases rule,
which Isabelle proves for us when it accepts an inductive definition:
 7.1 The Set of Even Numbers 131

[[aV ∈ Even.even; a = 0 =⇒ P;
n. [[a = Suc (Suc n); n ∈ Even.even]]
=⇒ P]]
=⇒ P (even.cases)

This general rule is less useful than instances of it for specific patterns. For
example, if a has the form Suc (Suc n) then the first case becomes irrele-
vant, while the second case tells us that n belongs to Even.even. Isabelle will
generate this instance for us:
inductive cases Suc_Suc_cases [elim!]: "Suc(Suc n) ∈ even"

The inductive cases command generates an instance of the cases rule

for the supplied pattern and gives it the supplied name:
[[Suc (Suc n) ∈ Even.even; n ∈ Even.even =⇒ P]] =⇒ P (Suc_Suc_cases)
Applying this as an elimination rule yields one case where even.cases would
yield two. Rule inversion works well when the conclusions of the introduc-
tion rules involve datatype constructors like Suc and # (list “cons”); freeness
reasoning discards all but one or two cases.
In the inductive cases command we supplied an attribute, elim!, in-
dicating that this elimination rule can be applied aggressively. The original
cases rule would loop if used in that manner because the pattern a matches
everything.
The rule Suc_Suc_cases is equivalent to the following implication:
Suc (Suc n) ∈ Even.even =⇒ n ∈ Even.even

Just above we devoted some effort to reaching precisely this result. Yet we
could have obtained it by a one-line declaration, dispensing with the lemma
even_imp_even_minus_2. This example also justifies the terminology rule in-
version: the new rule inverts the introduction rule even.step. In general, a
rule can be inverted when the set of elements it introduces is disjoint from
those of the other introduction rules.
For one-off applications of rule inversion, use the ind_cases method. Here
is an example:
apply (ind_cases "Suc(Suc n) ∈ even")

The specified instance of the cases rule is generated, then applied as an

elimination rule.
To summarize, every inductive definition produces a cases rule. The in-
ductive cases command stores an instance of the cases rule for a given
pattern. Within a proof, the ind_cases method applies an instance of the
cases rule.
The even numbers example has shown how inductive definitions can be
used. Later examples will show that they are actually worth using.
132 7. Inductively Defined Sets

7.1.6 Mutually Inductive Definitions

Just as there are datatypes defined by mutual recursion, there are sets defined
by mutual induction. As a trivial example we consider the even and odd
natural numbers:
inductive set
Even :: "nat set" and
Odd :: "nat set"
where
zero: "0 ∈ Even"
| EvenI: "n ∈ Odd =⇒ Suc n ∈ Even"
| OddI: "n ∈ Even =⇒ Suc n ∈ Odd"

The mutually inductive definition of multiple sets is no different from that

of a single set, except for induction: just as for mutually recursive datatypes,
induction needs to involve all the simultaneously defined sets. In the above
case, the induction rule is called Even_Odd.induct (simply concatenate the
names of the sets involved) and has the conclusion
(?x ∈ Even −→ ?P ?x) ∧ (?y ∈ Odd −→ ?Q ?y)

If we want to prove that all even numbers are divisible by two, we have
to generalize the statement as follows:
lemma "(m ∈ Even −→ 2 dvd m) ∧ (n ∈ Odd −→ 2 dvd (Suc n))"

The proof is by rule induction. Because of the form of the induction theorem,
it is applied by rule rather than erule as for ordinary inductive definitions:
apply(rule Even_Odd.induct)

1. V
even 0
2. Vn. [[n ∈ Odd; even (Suc n)]] =⇒ even (Suc n)
3. n. [[n ∈ Even; even n]] =⇒ even (Suc (Suc n))

The first two subgoals are proved by simplification and the final one can
be proved in the same manner as in Sect. 7.1.3 where the same subgoal was
encountered before. We do not show the proof script.

7.1.7 Inductively Defined Predicates

Instead of a set of even numbers one can also define a predicate on nat:
inductive evn :: "nat ⇒ bool" where
zero: "evn 0" |
step: "evn n =⇒ evn(Suc(Suc n))"

Everything works as before, except that you write inductive instead of in-
ductive set and evn n instead of n ∈ Even. When defining an n-ary rela-
tion as a predicate, it is recommended to curry the predicate: its type should
 7.2 The Reflexive Transitive Closure 133

be τ 1 ⇒ . . . ⇒ τ n ⇒ bool rather than τ 1 × . . . × τ n ⇒ bool. The curried

version facilitates inductions.
When should you choose sets and when predicates? If you intend to com-
bine your notion with set theoretic notation, define it as an inductive set. If
not, define it as an inductive predicate, thus avoiding the ∈ notation. But
note that predicates of more than one argument cannot be combined with
the usual set theoretic operators: P ∪ Q is not well-typed if P, Q :: τ 1 ⇒ τ 2
⇒ bool, you have to write λx y. P x y ∧ Q x y instead.

7.2 The Reflexive Transitive Closure

An inductive definition may accept parameters, so it can express functions
that yield sets. Relations too can be defined inductively, since they are just
sets of pairs. A perfect example is the function that maps a relation to its
reflexive transitive closure. This concept was already introduced in Sect. 6.3,
where the operator ∗ was defined as a least fixed point because inductive
definitions were not yet available. But now they are:
inductive set
rtc :: "('a × 'a)set ⇒ ('a × 'a)set" ("_*" [1000] 999)
for r :: "('a × 'a)set"
where
rtc_refl[iff]: "(x,x) ∈ r*"
| rtc_step: "[[ (x,y) ∈ r; (y,z) ∈ r* ]] =⇒ (x,z) ∈ r*"
The function rtc is annotated with concrete syntax: instead of rtc r we can
write r*. The actual definition consists of two rules. Reflexivity is obvious and
is immediately given the iff attribute to increase automation. The second
rule, rtc_step, says that we can always add one more r -step to the left.
Although we could make rtc_step an introduction rule, this is dangerous: the
recursion in the second premise slows down and may even kill the automatic
tactics.
The above definition of the concept of reflexive transitive closure may be
sufficiently intuitive but it is certainly not the only possible one: for a start,
it does not even mention transitivity. The rest of this section is devoted to
proving that it is equivalent to the standard definition. We start with a simple
lemma:
lemma [intro]: "(x,y) ∈ r =⇒ (x,y) ∈ r*"
by(blast intro: rtc_step)
Although the lemma itself is an unremarkable consequence of the basic rules,
it has the advantage that it can be declared an introduction rule without
the danger of killing the automatic tactics because r* occurs only in the
conclusion and not in the premise. Thus some proofs that would otherwise
need rtc_step can now be found automatically. The proof also shows that
blast is able to handle rtc_step. But some of the other automatic tactics are
134 7. Inductively Defined Sets

more sensitive, and even blast can be lead astray in the presence of large
numbers of rules.
To prove transitivity, we need rule induction, i.e. theorem rtc.induct:
?x2.0) ∈ ?r*; x. ?P x x;
V
[[(?x1.0,
x y z. [[(x, y) ∈ ?r; (y, z) ∈ ?r*; ?P y z]] =⇒ ?P x z]]
V
=⇒ ?P ?x1.0 ?x2.0

It says that ?P holds for an arbitrary pair (?x1.0, ?x2.0) ∈ ?r* if ?P is

preserved by all rules of the inductive definition, i.e. if ?P holds for the con-
clusion provided it holds for the premises. In general, rule induction for an
n-ary inductive relation R expects a premise of the form (x1 , . . . , xn ) ∈ R.
Now we turn to the inductive proof of transitivity:
lemma rtc_trans: "[[ (x,y) ∈ r*; (y,z) ∈ r* ]] =⇒ (x,z) ∈ r*"
apply(erule rtc.induct)

Unfortunately, even the base case is a problem:

1. x. (y, z) ∈ r* =⇒ (x, z) ∈ r*
V
A total of 2 subgoals...
We have to abandon this proof attempt. To understand what is going on,
let us look again at rtc.induct. In the above application of erule, the first
premise of rtc.induct is unified with the first suitable assumption, which is
(x, y) ∈ r* rather than (y, z) ∈ r*. Although that is what we want, it is
merely due to the order in which the assumptions occur in the subgoal, which
it is not good practice to rely on. As a result, ?xb becomes x, ?xa becomes
y and ?P becomes λu v. (u, z) ∈ r*, thus yielding the above subgoal. So
what went wrong?
When looking at the instantiation of ?P we see that it does not depend
on its second parameter at all. The reason is that in our original goal, of
the pair (x, y) only x appears also in the conclusion, but not y. Thus our
induction statement is too general. Fortunately, it can easily be specialized:
transfer the additional premise (y, z) ∈ r* into the conclusion:
lemma rtc_trans[rule_format]:
"(x,y) ∈ r* =⇒ (y,z) ∈ r* −→ (x,z) ∈ r*"

This is not an obscure trick but a generally applicable heuristic:

When proving a statement by rule induction on (x1 , . . . , xn ) ∈ R, pull
all other premises containing any of the xi into the conclusion using
−→.
A similar heuristic for other kinds of inductions is formulated in Sect. 9.2.1.
The rule_format directive turns −→ back into =⇒: in the end we obtain the
original statement of our lemma.
apply(erule rtc.induct)
Now induction produces two subgoals which are both proved automatically:
 7.2 The Reflexive Transitive Closure 135

1. Vx. (x, z) ∈ r* −→ (x, z) ∈ r*

V
2. x y za.
[[(x, y) ∈ r; (y, za) ∈ r*; (za, z) ∈ r* −→ (y, z) ∈ r*]]
=⇒ (za, z) ∈ r* −→ (x, z) ∈ r*

apply(blast)
apply(blast intro: rtc_step)
done

Let us now prove that r* is really the reflexive transitive closure of r, i.e.
the least reflexive and transitive relation containing r. The latter is easily
formalized
inductive set
rtc2 :: "('a × 'a)set ⇒ ('a × 'a)set"
for r :: "('a × 'a)set"
where
"(x,y) ∈ r =⇒ (x,y) ∈ rtc2 r"
| "(x,x) ∈ rtc2 r"
| "[[ (x,y) ∈ rtc2 r; (y,z) ∈ rtc2 r ]] =⇒ (x,z) ∈ rtc2 r"

and the equivalence of the two definitions is easily shown by the obvious rule
inductions:
lemma "(x,y) ∈ rtc2 r =⇒ (x,y) ∈ r*"
apply(erule rtc2.induct)
apply(blast)
apply(blast)
apply(blast intro: rtc_trans)
done

lemma "(x,y) ∈ r* =⇒ (x,y) ∈ rtc2 r"

apply(erule rtc.induct)
apply(blast intro: rtc2.intros)
apply(blast intro: rtc2.intros)
done

So why did we start with the first definition? Because it is simpler. It

contains only two rules, and the single step rule is simpler than transitivity.
As a consequence, rtc.induct is simpler than rtc2.induct. Since inductive
proofs are hard enough anyway, we should always pick the simplest induction
schema available. Hence rtc is the definition of choice.
Exercise 7.2.1 Show that the converse of rtc_step also holds:
[[(x, y) ∈ r*; (y, z) ∈ r]] =⇒ (x, z) ∈ r*

Exercise 7.2.2 Repeat the development of this section, but starting with
a definition of rtc where rtc_step is replaced by its converse as shown in
exercise 7.2.1.
136 7. Inductively Defined Sets

7.3 Advanced Inductive Definitions

The premises of introduction rules may contain universal quantifiers and

monotone functions. A universal quantifier lets the rule refer to any number
of instances of the inductively defined set. A monotone function lets the rule
refer to existing constructions (such as “list of”) over the inductively defined
set. The examples below show how to use the additional expressiveness and
how to reason from the resulting definitions.

7.3.1 Universal Quantifiers in Introduction Rules

As a running example, this section develops the theory of ground terms:

terms constructed from constant and function symbols but not variables. To
simplify matters further, we regard a constant as a function applied to the
null argument list. Let us declare a datatype gterm for the type of ground
terms. It is a type constructor whose argument is a type of function symbols.
datatype 'f gterm = Apply 'f "'f gterm list"

To try it out, we declare a datatype of some integer operations: integer

constants, the unary minus operator and the addition operator.
datatype integer_op = Number int | UnaryMinus | Plus

Now the type integer_op gterm denotes the ground terms built over those
symbols.
The type constructor gterm can be generalized to a function over sets. It
returns the set of ground terms that can be formed over a set F of function
symbols. For example, we could consider the set of ground terms formed from
the finite set {Number 2, UnaryMinus, Plus}.
This concept is inductive. If we have a list args of ground terms over F
and a function symbol f in F, then we can apply f to args to obtain another
ground term. The only difficulty is that the argument list may be of any
length. Hitherto, each rule in an inductive definition referred to the induc-
tively defined set a fixed number of times, typically once or twice. A universal
quantifier in the premise of the introduction rule expresses that every element
of args belongs to our inductively defined set: is a ground term over F. The
function set denotes the set of elements in a given list.
inductive set
gterms :: "'f set ⇒ 'f gterm set"
for F :: "'f set"
where
step[intro!]: "[[∀ t ∈ set args. t ∈ gterms F; f ∈ F]]
=⇒ (Apply f args) ∈ gterms F"
To demonstrate a proof from this definition, let us show that the function
gterms is monotone. We shall need this concept shortly.
 7.3 Advanced Inductive Definitions 137

lemma gterms_mono: "F⊆G =⇒ gterms F ⊆ gterms G"

apply clarify
apply (erule gterms.induct)
apply blast
done
Intuitively, this theorem says that enlarging the set of function symbols
enlarges the set of ground terms. The proof is a trivial rule induction. First
we use the clarify method to assume the existence of an element of gterms
F. (We could have used intro subsetI.) We then apply rule induction. Here
is the resulting subgoal:
1. x args f.
V
[[F ⊆ G; ∀ t∈set args. t ∈ gterms F ∧ t ∈ gterms G; f ∈ F]]
=⇒ Apply f args ∈ gterms G
The assumptions state that f belongs to F, which is included in G, and that
every element of the list args is a ground term over G. The blast method
finds this chain of reasoning easily.

!! Why do we call this function gterms instead of gterm? A constant may have
the same name as a type. However, name clashes could arise in the theorems
that Isabelle generates. Our choice of names keeps gterms.induct separate from
gterm.induct.
Call a term well-formed if each symbol occurring in it is applied to the
correct number of arguments. (This number is called the symbol’s arity.)
We can express well-formedness by generalizing the inductive definition of
gterms. Suppose we are given a function called arity, specifying the arities
of all symbols. In the inductive step, we have a list args of such terms and a
function symbol f. If the length of the list matches the function’s arity then
applying f to args yields a well-formed term.
inductive set
well_formed_gterm :: "('f ⇒ nat) ⇒ 'f gterm set"
for arity :: "'f ⇒ nat"
where
step[intro!]: "[[∀ t ∈ set args. t ∈ well_formed_gterm arity;
length args = arity f]]
=⇒ (Apply f args) ∈ well_formed_gterm arity"
The inductive definition neatly captures the reasoning above. The univer-
sal quantification over the set of arguments expresses that all of them are
well-formed.

7.3.2 Alternative Definition Using a Monotone Function

An inductive definition may refer to the inductively defined set through an

arbitrary monotone function. To demonstrate this powerful feature, let us
change the inductive definition above, replacing the quantifier by a use of the
138 7. Inductively Defined Sets

function lists. This function, from the Isabelle theory of lists, is analogous
to the function gterms declared above: if A is a set then lists A is the set of
lists whose elements belong to A.
In the inductive definition of well-formed terms, examine the one intro-
duction rule. The first premise states that args belongs to the lists of well-
formed terms. This formulation is more direct, if more obscure, than using a
universal quantifier.
inductive set
well_formed_gterm' :: "('f ⇒ nat) ⇒ 'f gterm set"
for arity :: "'f ⇒ nat"
where
step[intro!]: "[[args ∈ lists (well_formed_gterm' arity);
length args = arity f]]
=⇒ (Apply f args) ∈ well_formed_gterm' arity"
monos lists_mono
We cite the theorem lists_mono to justify using the function lists.1
A ⊆ B =⇒ lists A ⊆ lists B (lists_mono)
Why must the function be monotone? An inductive definition describes an
iterative construction: each element of the set is constructed by a finite num-
ber of introduction rule applications. For example, the elements of even are
constructed by finitely many applications of the rules
0 ∈ Even.even
n ∈ Even.even =⇒ Suc (Suc n) ∈ Even.even

All references to a set in its inductive definition must be positive. Applications

of an introduction rule cannot invalidate previous applications, allowing the
construction process to converge. The following pair of rules do not constitute
an inductive definition:
0 ∈ Even.even
n ∈
/ Even.even =⇒ Suc n ∈ Even.even

Showing that 4 is even using these rules requires showing that 3 is not even.
It is far from trivial to show that this set of rules characterizes the even
numbers.
Even with its use of the function lists, the premise of our introduction
rule is positive:
args ∈ lists (well_formed_gterm' arity)
To apply the rule we construct a list args of previously constructed well-
formed terms. We obtain a new term, Apply f args. Because lists is mono-
tone, applications of the rule remain valid as new terms are constructed.
Further lists of well-formed terms become available and none are taken away.

1
This particular theorem is installed by default already, but we include the monos
declaration in order to illustrate its syntax.
 7.3 Advanced Inductive Definitions 139

7.3.3 A Proof of Equivalence

We naturally hope that these two inductive definitions of “well-formed” co-

incide. The equality can be proved by separate inclusions in each direction.
Each is a trivial rule induction.
lemma "well_formed_gterm arity ⊆ well_formed_gterm' arity"
apply clarify
apply (erule well_formed_gterm.induct)
apply auto
done

The clarify method gives us an element of well_formed_gterm arity on

which to perform induction. The resulting subgoal can be proved automati-
cally:
1. x args f.
V
[[∀ t∈set args.
t ∈ well_formed_gterm arity ∧ t ∈ well_formed_gterm' arity;
length args = arity f]]
=⇒ Apply f args ∈ well_formed_gterm' arity

This proof resembles the one given in Sect. 7.3.1 above, especially in the form
of the induction hypothesis. Next, we consider the opposite inclusion:
lemma "well_formed_gterm' arity ⊆ well_formed_gterm arity"
apply clarify
apply (erule well_formed_gterm'.induct)
apply auto
done

The proof script is virtually identical, but the subgoal after applying in-
duction may be surprising:
1. x args f.
V
[[args
∈ lists
(well_formed_gterm' arity ∩
{a. a ∈ well_formed_gterm arity});
length args = arity f]]
=⇒ Apply f args ∈ well_formed_gterm arity

The induction hypothesis contains an application of lists. Using a monotone

function in the inductive definition always has this effect. The subgoal may
look uninviting, but fortunately lists distributes over intersection:
lists (A ∩ B) = lists A ∩ lists B (lists_Int_eq)

Thanks to this default simplification rule, the induction hypothesis is quickly

replaced by its two parts:
args ∈ lists (well_formed_gterm' arity)
args ∈ lists (well_formed_gterm arity)
140 7. Inductively Defined Sets

Invoking the rule well_formed_gterm.step completes the proof. The call to

auto does all this work.
This example is typical of how monotone functions can be used. In par-
ticular, many of them distribute over intersection. Monotonicity implies one
direction of this set equality; we have this theorem:
mono f =⇒ f (A ∩ B) ⊆ f A ∩ f B (mono_Int)

7.3.4 Another Example of Rule Inversion

Does gterms distribute over intersection? We have proved that this function
is monotone, so mono_Int gives one of the inclusions. The opposite inclusion
asserts that if t is a ground term over both of the sets F and G then it is also
a ground term over their intersection, F ∩ G.
lemma gterms_IntI:
"t ∈ gterms F =⇒ t ∈ gterms G −→ t ∈ gterms (F∩G)"

Attempting this proof, we get the assumption Apply f args ∈ gterms G,

which cannot be broken down. It looks like a job for rule inversion:
inductive cases gterm_Apply_elim [elim!]: "Apply f args ∈ gterms F"

Here is the result.

[[Apply f args ∈ gterms F;
[[∀ t∈set args. t ∈ gterms F; f ∈ F]] =⇒ P]]
=⇒ P (gterm_Apply_elim)

This rule replaces an assumption about Apply f args by assumptions about

f and args. No cases are discarded (there was only one to begin with) but
the rule applies specifically to the pattern Apply f args. It can be applied
repeatedly as an elimination rule without looping, so we have given the elim!
attribute.
Now we can prove the other half of that distributive law.
lemma gterms_IntI [rule_format, intro!]:
"t ∈ gterms F =⇒ t ∈ gterms G −→ t ∈ gterms (F∩G)"
apply (erule gterms.induct)
apply blast
done
The proof begins with rule induction over the definition of gterms, which
leaves a single subgoal:
1. args f.
V
[[∀ t∈set args.
t ∈ gterms F ∧ (t ∈ gterms G −→ t ∈ gterms (F ∩ G));
f ∈ F]]
=⇒ Apply f args ∈ gterms G −→
Apply f args ∈ gterms (F ∩ G)
 7.4 Case Study: A Context Free Grammar 141

To prove this, we assume Apply f args ∈ gterms G. Rule inversion, in the

form of gterm_Apply_elim, infers that every element of args belongs to gterms
G ; hence (by the induction hypothesis) it belongs to gterms (F ∩ G). Rule
inversion also yields f ∈ G and hence f ∈ F ∩ G. All of this reasoning is
done by blast.
Our distributive law is a trivial consequence of previously-proved results:
lemma gterms_Int_eq [simp]:
"gterms (F ∩ G) = gterms F ∩ gterms G"
by (blast intro!: mono_Int monoI gterms_mono)

Exercise 7.3.1 A function mapping function symbols to their types is called

a signature. Given a type ranging over type symbols, we can represent a
function’s type by a list of argument types paired with the result type. Com-
plete this inductive definition:
inductive set
well_typed_gterm :: "('f ⇒ 't list * 't) ⇒ ('f gterm * 't)set"
for sig :: "'f ⇒ 't list * 't"

7.4 Case Study: A Context Free Grammar

Grammars are nothing but shorthands for inductive definitions of nontermi-

nals which represent sets of strings. For example, the production A → Bc is
short for
w ∈ B =⇒ wc ∈ A
This section demonstrates this idea with an example due to Hopcroft and
Ullman, a grammar for generating all words with an equal number of a’s
and b’s:

S →  | bA | aB
A → aS | bAA
B → bS | aBB

At the end we say a few words about the relationship between the original
proof [13, p. 81] and our formal version.
We start by fixing the alphabet, which consists only of a’s and b ’s:
datatype alfa = a | b

For convenience we include the following easy lemmas as simplification rules:

lemma [simp]: "(x 6= a) = (x = b) ∧ (x 6= b) = (x = a)"
by (case_tac x, auto)
142 7. Inductively Defined Sets

Words over this alphabet are of type alfa list, and the three nonterminals
are declared as sets of such words. The productions above are recast as a
mutual inductive definition of S, A and B :
inductive set
S :: "alfa list set" and
A :: "alfa list set" and
B :: "alfa list set"
where
"[] ∈ S"
| "w ∈ A =⇒ b#w ∈ S"
| "w ∈ B =⇒ a#w ∈ S"

| "w ∈ S =⇒ a#w ∈ A"

| "[[ v∈A; w∈A ]] =⇒ b#v@w ∈ A"

| "w ∈ S =⇒ b#w ∈ B"

| "[[ v ∈ B; w ∈ B ]] =⇒ a#v@w ∈ B"

First we show that all words in S contain the same number of a’s and b ’s.
Since the definition of S is by mutual induction, so is the proof: we show at
the same time that all words in A contain one more a than b and all words
in B contain one more b than a.
lemma correctness:
"(w ∈ S −→ size[x←w. x=a] = size[x←w. x=b]) ∧
(w ∈ A −→ size[x←w. x=a] = size[x←w. x=b] + 1) ∧
(w ∈ B −→ size[x←w. x=b] = size[x←w. x=a] + 1)"

These propositions are expressed with the help of the predefined filter func-
tion on lists, which has the convenient syntax [x←xs. P x], the list of all el-
ements x in xs such that P x holds. Remember that on lists size and length
are synonymous.
The proof itself is by rule induction and afterwards automatic:
by (rule S_A_B.induct, auto)

This may seem surprising at first, and is indeed an indication of the power
of inductive definitions. But it is also quite straightforward. For example,
consider the production A → bAA: if v, w ∈ A and the elements of A contain
one more a than b’s, then bvw must again contain one more a than b’s.
As usual, the correctness of syntactic descriptions is easy, but complete-
ness is hard: does S contain all words with an equal number of a’s and b ’s? It
turns out that this proof requires the following lemma: every string with two
more a’s than b ’s can be cut somewhere such that each half has one more a
than b. This is best seen by imagining counting the difference between the
number of a’s and b ’s starting at the left end of the word. We start with 0
and end (at the right end) with 2. Since each move to the right increases
or decreases the difference by 1, we must have passed through 1 on our way
from 0 to 2. Formally, we appeal to the following discrete intermediate value
theorem nat0_intermed_int_val
 7.4 Case Study: A Context Free Grammar 143

[[∀ i<n. |f (i + 1) - f i| ≤ 1; f 0 ≤ k; k ≤ f n]]

=⇒ ∃ i≤n. f i = k
where f is of type nat ⇒ int, int are the integers, |.| is the absolute value
function2 , and 1 is the integer 1 (see Sect. 8.4).
First we show that our specific function, the difference between the num-
bers of a’s and b ’s, does indeed only change by 1 in every move to the right.
At this point we also start generalizing from a’s and b ’s to an arbitrary prop-
erty P. Otherwise we would have to prove the desired lemma twice, once as
stated above and once with the roles of a’s and b ’s interchanged.
lemma step1: "∀ i < size w.
|(int(size[x←take (i+1) w. P x])-int(size[x←take (i+1) w. ¬P x]))
- (int(size[x←take i w. P x])-int(size[x←take i w. ¬P x]))| ≤ 1"
The lemma is a bit hard to read because of the coercion function int :: nat
⇒ int. It is required because size returns a natural number, but subtraction
on type nat will do the wrong thing. Function take is predefined and take i
xs is the prefix of length i of xs; below we also need drop i xs, which is what
remains after that prefix has been dropped from xs.
The proof is by induction on w, with a trivial base case, and a not so trivial
induction step. Since it is essentially just arithmetic, we do not discuss it.
apply(induct_tac w)
apply(auto simp add: abs_if take_Cons split: nat.split)
done
Finally we come to the above-mentioned lemma about cutting in half a
word with two more elements of one sort than of the other sort:
lemma part1:
"size[x←w. P x] = size[x←w. ¬P x]+2 =⇒
∃ i≤size w. size[x←take i w. P x] = size[x←take i w. ¬P x]+1"
This is proved by force with the help of the intermediate value theorem,
instantiated appropriately and with its first premise disposed of by lemma
step1:
apply(insert nat0_intermed_int_val[OF step1, of "P" "w" "1"])
by force
Lemma part1 tells us only about the prefix take i w. An easy lemma
deals with the suffix drop i w :
lemma part2:
"[[size[x←take i w @ drop i w. P x] =
size[x←take i w @ drop i w. ¬P x]+2;
size[x←take i w. P x] = size[x←take i w. ¬P x]+1]]
=⇒ size[x←drop i w. P x] = size[x←drop i w. ¬P x]+1"
by(simp del: append_take_drop_id)
In the proof we have disabled the normally useful lemma
2
See Table A.1 in the Appendix for the correct ascii syntax.
144 7. Inductively Defined Sets

take n xs @ drop n xs = xs (append_take_drop_id)

to allow the simplifier to apply the following lemma instead:

[x∈xs@ys. P x] = [x∈xs. P x] @ [x∈ys. P x]

To dispose of trivial cases automatically, the rules of the inductive defini-

tion are declared simplification rules:
declare S_A_B.intros[simp]

This could have been done earlier but was not necessary so far.
The completeness theorem tells us that if a word has the same number of
a’s and b ’s, then it is in S, and similarly for A and B :
theorem completeness:
"(size[x←w. x=a] = size[x←w. x=b] −→ w ∈ S) ∧
(size[x←w. x=a] = size[x←w. x=b] + 1 −→ w ∈ A) ∧
(size[x←w. x=b] = size[x←w. x=a] + 1 −→ w ∈ B)"

The proof is by induction on w. Structural induction would fail here because,

as we can see from the grammar, we need to make bigger steps than merely
appending a single letter at the front. Hence we induct on the length of w,
using the induction rule length_induct:
apply(induct_tac w rule: length_induct)
apply(rename_tac w)

The rule parameter tells induct_tac explicitly which induction rule to use.
For details see Sect. 9.2.2 below. In this case the result is that we may assume
the lemma already holds for all words shorter than w. Because the induction
step renames the induction variable we rename it back to w.
The proof continues with a case distinction on w, on whether w is empty
or not.
apply(case_tac w)
apply(simp_all)

Simplification disposes of the base case and leaves only a conjunction of two
step cases to be proved: if w = a # v and
length (if x = a then [x ∈ v] else []) =
length (if x = b then [x ∈ v] else []) + 2
then b # v ∈ A, and similarly for w = b # v. We only consider the first case
in detail.
After breaking the conjunction up into two cases, we can apply part1 to
the assumption that w contains two more a’s than b ’s.
apply(rule conjI)
apply(clarify)
apply(frule part1[of "λx. x=a", simplified])
apply(clarify)

This yields an index i ≤ length v such that

 7.4 Case Study: A Context Free Grammar 145

length (filter (λx. x = a) (take i v)) =

length (filter (λx. x = b) (take i v)) + 1
With the help of part2 it follows that
length (filter (λx. x = a) (drop i v)) =
length (filter (λx. x = b) (drop i v)) + 1

apply(drule part2[of "λx. x=a", simplified])

apply(assumption)
Now it is time to decompose v in the conclusion b # v ∈ A into take i v @
drop i v,
apply(rule_tac n1=i and t=v in subst[OF append_take_drop_id])
(the variables n1 and t are the result of composing the theorems subst and
append_take_drop_id) after which the appropriate rule of the grammar re-
duces the goal to the two subgoals take i v ∈ A and drop i v ∈ A:
apply(rule S_A_B.intros)
Both subgoals follow from the induction hypothesis because both take i
v and drop i v are shorter than w :
apply(force simp add: min_less_iff_disj)
apply(force split: nat_diff_split)
The case w = b # v is proved analogously:
apply(clarify)
apply(frule part1[of "λx. x=b", simplified])
apply(clarify)
apply(drule part2[of "λx. x=b", simplified])
apply(assumption)
apply(rule_tac n1=i and t=v in subst[OF append_take_drop_id])
apply(rule S_A_B.intros)
apply(force simp add: min_less_iff_disj)
by(force simp add: min_less_iff_disj split: nat_diff_split)
We conclude this section with a comparison of our proof with Hopcroft
and Ullman’s [13, p. 81]. For a start, the textbook grammar, for no good
reason, excludes the empty word, thus complicating matters just a little bit:
they have 8 instead of our 7 productions.
More importantly, the proof itself is different: rather than separating the
two directions, they perform one induction on the length of a word. This
deprives them of the beauty of rule induction, and in the easy direction
(correctness) their reasoning is more detailed than our auto. For the hard
part (completeness), they consider just one of the cases that our simp_all
disposes of automatically. Then they conclude the proof by saying about the
remaining cases: “We do this in a manner similar to our method of proof
for part (1); this part is left to the reader”. But this is precisely the part
that requires the intermediate value theorem and thus is not at all similar to
the other cases (which are automatic in Isabelle). The authors are at least
146 7. Inductively Defined Sets

cavalier about this point and may even have overlooked the slight difficulty
lurking in the omitted cases. Such errors are found in many pen-and-paper
proofs when they are scrutinized formally.
 Part III

Advanced Material
8. More about Types

So far we have learned about a few basic types (for example bool and nat),
type abbreviations (types) and recursive datatypes (datatype). This chap-
ter will introduce more advanced material:
– Pairs (Sect. 8.1) and records (Sect. 8.2), and how to reason about them.
– Type classes: how to specify and reason about axiomatic collections of
types (Sect. 8.3). This section leads on to a discussion of Isabelle’s numeric
types (Sect. 8.4).
– Introducing your own types: how to define types that cannot be constructed
with any of the basic methods (Sect. 8.5).
The material in this section goes beyond the needs of most novices. Serious
users should at least skim the sections as far as type classes. That material
is fairly advanced; read the beginning to understand what it is about, but
consult the rest only when necessary.

8.1 Pairs and Tuples

Ordered pairs were already introduced in Sect. 2.6.2, but only with a minimal
repertoire of operations: pairing and the two projections fst and snd. In
any non-trivial application of pairs you will find that this quickly leads to
unreadable nests of projections. This section introduces syntactic sugar to
overcome this problem: pattern matching with tuples.

8.1.1 Pattern Matching with Tuples

Tuples may be used as patterns in λ-abstractions, for example λ(x,y,z).x+y+z

and λ((x,y),z).x+y+z. In fact, tuple patterns can be used in most variable
binding constructs, and they can be nested. Here are some typical examples:
let (x, y) = f z in (y, x)
case xs of [] ⇒ 0 | (x, y) # zs ⇒ x + y
∀ (x,y)∈A. x=y
{(x,y,z). x=z}
(x, y)∈A {x + y}
S
150 8. More about Types

The intuitive meanings of these expressions should be obvious. Unfortunately,

we need to know in more detail what the notation really stands for once
we have to reason about it. Abstraction over pairs and tuples is merely a
convenient shorthand for a more complex internal representation. Thus the
internal and external form of a term may differ, which can affect proofs. If
you want to avoid this complication, stick to fst and snd and write λp. fst
p + snd p instead of λ(x,y). x+y. These terms are distinct even though they
denote the same function.
Internally, λ(x, y). t becomes case_prod (λx y. t), where split is the
uncurrying function of type ('a ⇒ 'b ⇒ 'c) ⇒ 'a × 'b ⇒ 'c defined as
case_prod = (λc p. c (fst p) (snd p)) (split_def )
Pattern matching in other variable binding constructs is translated similarly.
Thus we need to understand how to reason about such constructs.

8.1.2 Theorem Proving

The most obvious approach is the brute force expansion of split:

lemma "(λ(x,y).x) p = fst p"
by(simp add: split_def)

This works well if rewriting with split_def finishes the proof, as it does
above. But if it does not, you end up with exactly what we are trying to
avoid: nests of fst and snd. Thus this approach is neither elegant nor very
practical in large examples, although it can be effective in small ones.
If we consider why this lemma presents a problem, we realize that we need
to replace variable p by some pair (a, b). Then both sides of the equation
would simplify to a by the simplification rules (case (a, b) of (c, d) ⇒ f
c d) = f a b and fst (x1, x2) = x1. To reason about tuple patterns requires
some way of converting a variable of product type into a pair. In case of a
subterm of the form case p of (x, xa) ⇒ f x xa this is easy: the split rule
prod.split replaces p by a pair:
lemma "(λ(x,y).y) p = snd p"
apply(split prod.split)

1. ∀ x1 x2. p = (x1, x2) −→ x2 = snd p

This subgoal is easily proved by simplification. Thus we could have combined

simplification and splitting in one command that proves the goal outright:
by(simp split: prod.split)

Let us look at a second example:

lemma "let (x,y) = p in fst p = x"
apply(simp only: Let_def)
 8.1 Pairs and Tuples 151

1. case p of (x, y) ⇒ fst p = x

A paired let reduces to a paired λ-abstraction, which can be split as above.

The same is true for paired set comprehension:
lemma "p ∈ {(x,y). x=y} −→ fst p = snd p"
apply simp

1. (case p of (x, xa) ⇒ x = xa) −→ fst p = snd p

Again, simplification produces a term suitable for prod.split as above. If you

are worried about the strange form of the premise: case_prod (=) is short for
λ(x, y). x = y. The same proof procedure works for
lemma "p ∈ {(x,y). x=y} =⇒ fst p = snd p"
except that we now have to use prod.split_asm, because split occurs in the
assumptions.
However, splitting split is not always a solution, as no split may be
present in the goal. Consider the following function:
primrec swap :: "'a × 'b ⇒ 'b × 'a" where "swap (x,y) = (y,x)"
Note that the above primrec definition is admissible because × is a datatype.
When we now try to prove
lemma "swap(swap p) = p"
simplification will do nothing, because the defining equation for swap expects
a pair. Again, we need to turn p into a pair first, but this time there is no
split in sight. The only thing we can do is to split the term by hand:
apply(case_tac p)

1. a b. p = (a, b) =⇒ swap (swap p) = p

V

Again, case_tac is applicable because × is a datatype. The subgoal is easily

proved by simp.
Splitting by case_tac also solves the previous examples and may thus
appear preferable to the more arcane methods introduced first. However, see
the warning about case_tac in Sect. 2.5.5.
Alternatively, you can split all -quantified variables in a goal with the
V

rewrite rule split_paired_all:

lemma " p q. swap(swap p) = q −→ p = q"
V
apply(simp only: split_paired_all)

1. a b aa ba. swap (swap (a, b)) = (aa, ba) −→ (a, b) = (aa, ba)
V

apply simp
done
Note that we have intentionally included only split_paired_all in the first
simplification step, and then we simplify again. This time the reason was not
152 8. More about Types

merely pedagogical: split_paired_all may interfere with other functions of

the simplifier. The following command could fail (here it does not) where two
separate simp applications succeed.
apply(simp add: split_paired_all)
Finally, the simplifier automatically splits all ∀ and ∃ -quantified variables:
lemma "∀ p. ∃ q. swap p = swap q"
by simp
To turn off this automatic splitting, disable the responsible simplification
rules:
(∀ x. P x) = (∀ a b. P (a, b)) (split_paired_All)
(∃ x. P x) = (∃ a b. P (a, b)) (split_paired_Ex )

8.2 Records

Records are familiar from programming languages. A record of n fields is

essentially an n-tuple, but the record’s components have names, which can
make expressions easier to read and reduces the risk of confusing one field
for another.
A record of Isabelle/HOL covers a collection of fields, with select and
update operations. Each field has a specified type, which may be polymorphic.
The field names are part of the record type, and the order of the fields is
significant — as it is in Pascal but not in Standard ML. If two different
record types have field names in common, then the ambiguity is resolved in
the usual way, by qualified names.
Record types can also be defined by extending other record types. Exten-
sible records make use of the reserved pseudo-field more, which is present in
every record type. Generic record operations work on all possible extensions
of a given type scheme; polymorphism takes care of structural sub-typing
behind the scenes. There are also explicit coercion functions between fixed
record types.

8.2.1 Record Basics

Record types are not primitive in Isabelle and have a delicate internal rep-
resentation [21], based on nested copies of the primitive product type. A
record declaration introduces a new record type scheme by specifying its
fields, which are packaged internally to hold up the perception of the record
as a distinguished entity. Here is a simple example:
record point =
Xcoord :: int
 8.2 Records 153

Ycoord :: int

Records of type point have two fields named Xcoord and Ycoord, both of
type int. We now define a constant of type point:
definition pt1 :: point where
"pt1 ≡ (| Xcoord = 999, Ycoord = 23 |)"

We see above the ASCII notation for record brackets. You can also use the
symbolic brackets (| and |). Record type expressions can be also written di-
rectly with individual fields. The type name above is merely an abbreviation.
definition pt2 :: "(|Xcoord :: int, Ycoord :: int|)" where
"pt2 ≡ (|Xcoord = -45, Ycoord = 97|)"
For each field, there is a selector function of the same name. For exam-
ple, if p has type point then Xcoord p denotes the value of the Xcoord field
of p. Expressions involving field selection of explicit records are simplified
automatically:
lemma "Xcoord (|Xcoord = a, Ycoord = b|) = a"
by simp

The update operation is functional. For example, p(|Xcoord := 0|) is a

record whose Xcoord value is zero and whose Ycoord value is copied from p.
Updates of explicit records are also simplified automatically:
lemma "(|Xcoord = a, Ycoord = b|)(|Xcoord := 0|) =
(|Xcoord = 0, Ycoord = b|)"
by simp

!! Field names are declared as constants and can no longer be used as variables.
It would be unwise, for example, to call the fields of type point simply x and y.

8.2.2 Extensible Records and Generic Operations

Now, let us define coloured points (type cpoint) to be points extended with
a field col of type colour :
datatype colour = Red | Green | Blue

record cpoint = point +

col :: colour

The fields of this new type are Xcoord, Ycoord and col, in that order.
definition cpt1 :: cpoint where
"cpt1 ≡ (|Xcoord = 999, Ycoord = 23, col = Green|)"

We can define generic operations that work on arbitrary instances of a

record scheme, e.g. covering point, cpoint, and any further extensions. Every
record structure has an implicit pseudo-field, more, that keeps the extension
154 8. More about Types

as an explicit value. Its type is declared as completely polymorphic: 'a. When

a fixed record value is expressed using just its standard fields, the value of
more is implicitly set to (), the empty tuple, which has type unit. Within the
record brackets, you can refer to the more field by writing “. . . ” (three dots):
lemma "Xcoord (|Xcoord = a, Ycoord = b, . . . = p|) = a"
by simp
This lemma applies to any record whose first two fields are Xcoord
and Ycoord. Note that (|Xcoord = a, Ycoord = b, . . . = ()|) is exactly the
same as (|Xcoord = a, Ycoord = b|). Selectors and updates are always poly-
morphic wrt. the more part of a record scheme, its value is just ignored (for
select) or copied (for update).
The more pseudo-field may be manipulated directly as well, but the iden-
tifier needs to be qualified:
lemma "point.more cpt1 = (|col = Green|)"
by (simp add: cpt1_def)
We see that the colour part attached to this point is a rudimentary record
in its own right, namely (|col = Green|). In order to select or update col, this
fragment needs to be put back into the context of the parent type scheme,
say as more part of another point.
To define generic operations, we need to know a bit more about records.
Our definition of point above has generated two type abbreviations:
point = (|Xcoord :: int, Ycoord :: int|)
'a point_scheme = (|Xcoord :: int, Ycoord :: int, . . . :: 'a|)
Type point is for fixed records having exactly the two fields Xcoord and Ycoord,
while the polymorphic type 'a point_scheme comprises all possible extensions
to those two fields. Note that unit point_scheme coincides with point, and
(|col :: colour|) point_scheme with cpoint.
In the following example we define two operations — methods, if we regard
records as objects — to get and set any point’s Xcoord field.
definition getX :: "'a point_scheme ⇒ int" where
"getX r ≡ Xcoord r"
definition setX :: "'a point_scheme ⇒ int ⇒ 'a point_scheme" where
"setX r a ≡ r(|Xcoord := a|)"
Here is a generic method that modifies a point, incrementing its Xcoord
field. The Ycoord and more fields are copied across. It works for any record
type scheme derived from point (including cpoint etc.):
definition incX :: "'a point_scheme ⇒ 'a point_scheme" where
"incX r ≡
(|Xcoord = Xcoord r + 1, Ycoord = Ycoord r, . . . = point.more r|)"
Generic theorems can be proved about generic methods. This trivial
lemma relates incX to getX and setX :
lemma "incX r = setX r (getX r + 1)"
 8.2 Records 155

by (simp add: getX_def setX_def incX_def)

!! If you use the symbolic record brackets (| and |), then you must also use the
symbolic ellipsis, “. . . ”, rather than three consecutive periods, “...”. Mixing
the ASCII and symbolic versions causes a syntax error. (The two versions are more
distinct on screen than they are on paper.)

8.2.3 Record Equality

Two records are equal if all pairs of corresponding fields are equal. Concrete
record equalities are simplified automatically:
lemma "((|Xcoord = a, Ycoord = b|) = (|Xcoord = a', Ycoord = b'|)) =
(a = a' ∧ b = b')"
by simp
The following equality is similar, but generic, in that r can be any instance
of 'a point_scheme:
lemma "r(|Xcoord := a, Ycoord := b|) = r(|Ycoord := b, Xcoord := a|)"
by simp

We see above the syntax for iterated updates. We could equivalently have
written the left-hand side as r(|Xcoord := a|)(|Ycoord := b|).
Record equality is extensional: a record is determined entirely by the
values of its fields.
lemma "r = (|Xcoord = Xcoord r, Ycoord = Ycoord r|)"
by simp
The generic version of this equality includes the pseudo-field more:
lemma "r = (|Xcoord = Xcoord r, Ycoord = Ycoord r, . . . = point.more r|)"
by simp

The simplifier can prove many record equalities automatically, but general
equality reasoning can be tricky. Consider proving this obvious fact:
lemma "r(|Xcoord := a|) = r(|Xcoord := a'|) =⇒ a = a'"
apply simp?
oops
Here the simplifier can do nothing, since general record equality is not elimi-
nated automatically. One way to proceed is by an explicit forward step that
applies the selector Xcoord to both sides of the assumed record equality:
lemma "r(|Xcoord := a|) = r(|Xcoord := a'|) =⇒ a = a'"
apply (drule_tac f = Xcoord in arg_cong)

1. Xcoord (r(|Xcoord := a|)) = Xcoord (r(|Xcoord := a'|)) =⇒ a = a'

156 8. More about Types

Now, simp will reduce the assumption to the desired conclusion.

apply simp
done
The cases method is preferable to such a forward proof. We state the
desired lemma again:
lemma "r(|Xcoord := a|) = r(|Xcoord := a'|) =⇒ a = a'"
The cases method adds an equality to replace the named record term by
an explicit record expression, listing all fields. It even includes the pseudo-field
more, since the record equality stated here is generic for all extensions.
apply (cases r)

1. Xcoord Ycoord more.

V
[[r(|Xcoord := a|) = r(|Xcoord := a'|);
r = (|Xcoord = Xcoord, Ycoord = Ycoord, . . . = more|)]]
=⇒ a = a'

Again, simp finishes the proof. Because r is now represented as an explicit

record construction, the updates can be applied and the record equality can
be replaced by equality of the corresponding fields (due to injectivity).
apply simp
done
The generic cases method does not admit references to locally bound
parameters of a goal. In longer proof scripts one might have to fall back on
the primitive rule_tac used together with the internal field representation
rules of records. The above use of (cases r) would become (rule_tac r = r
in point.cases_scheme).

8.2.4 Extending and Truncating Records

Each record declaration introduces a number of derived operations to refer

collectively to a record’s fields and to convert between fixed record types.
They can, for instance, convert between types point and cpoint. We can add
a colour to a point or convert a cpoint to a point by forgetting its colour.
– Function make takes as arguments all of the record’s fields (including those
inherited from ancestors). It returns the corresponding record.
– Function fields takes the record’s very own fields and returns a record
fragment consisting of just those fields. This may be filled into the more
part of the parent record scheme.
– Function extend takes two arguments: a record to be extended and a record
containing the new fields.
– Function truncate takes a record (possibly an extension of the original
record type) and returns a fixed record, removing any additional fields.
 8.2 Records 157

These functions provide useful abbreviations for standard record expressions

involving constructors and selectors. The definitions, which are not unfolded
by default, are made available by the collective name of defs (point.defs,
cpoint.defs, etc.). For example, here are the versions of those functions gen-
erated for record point. We omit point.fields, which happens to be the same
as point.make.
point.make Xcoord Ycoord ≡ (|Xcoord = Xcoord, Ycoord = Ycoord|)
point.extend r more ≡
(|Xcoord = Xcoord r, Ycoord = Ycoord r, . . . = more|)
point.truncate r ≡ (|Xcoord = Xcoord r, Ycoord = Ycoord r|)

Contrast those with the corresponding functions for record cpoint. Observe
cpoint.fields in particular.
cpoint.make Xcoord Ycoord col ≡
(|Xcoord = Xcoord, Ycoord = Ycoord, col = col|)
cpoint.fields col ≡ (|col = col|)
cpoint.extend r more ≡
(|Xcoord = Xcoord r, Ycoord = Ycoord r, col = col r, . . . = more|)
cpoint.truncate r ≡
(|Xcoord = Xcoord r, Ycoord = Ycoord r, col = col r|)
To demonstrate these functions, we declare a new coloured point by
extending an ordinary point. Function point.extend augments pt1 with a
colour value, which is converted into an appropriate record fragment by
cpoint.fields.
definition cpt2 :: cpoint where
"cpt2 ≡ point.extend pt1 (cpoint.fields Green)"
The coloured points cpt1 and cpt2 are equal. The proof is trivial, by
unfolding all the definitions. We deliberately omit the definition of pt1 in
order to reveal the underlying comparison on type point.
lemma "cpt1 = cpt2"
apply (simp add: cpt1_def cpt2_def point.defs cpoint.defs)

1. Xcoord pt1 = 999 ∧ Ycoord pt1 = 23

apply (simp add: pt1_def)

done
In the example below, a coloured point is truncated to leave a point. We
use the truncate function of the target record.
lemma "point.truncate cpt2 = pt1"
by (simp add: pt1_def cpt2_def point.defs)

Exercise 8.2.1 Extend record cpoint to have a further field, intensity, of

type nat. Experiment with generic operations (using polymorphic selectors
and updates) and explicit coercions (using extend, truncate etc.) among the
three record types.
158 8. More about Types

Exercise 8.2.2 (For Java programmers.) Model a small class hierarchy using
records.

8.3 Type Classes

The programming language Haskell has popularized the notion of type

classes: a type class is a set of types with a common interface: all types
in that class must provide the functions in the interface. Isabelle offers a sim-
ilar type class concept: in addition, properties (class axioms) can be specified
which any instance of this type class must obey. Thus we can talk about a
type τ being in a class C , which is written τ :: C . This is the case if τ satisfies
the axioms of C . Furthermore, type classes can be organized in a hierarchy.
Thus there is the notion of a class D being a subclass of a class C , written
D < C . This is the case if all axioms of C are also provable in D.
In this section we introduce the most important concepts behind type
classes by means of a running example from algebra. This should give you an
intuition how to use type classes and to understand specifications involving
type classes. Type classes are covered more deeply in a separate tutorial [11].

8.3.1 Overloading

Type classes allow overloading; thus a constant may have multiple definitions
at non-overlapping types.

Overloading. We can introduce a binary infix addition operator ⊕ for ar-

bitrary types by means of a type class:
class plus =
fixes plus :: "'a ⇒ 'a ⇒ 'a" (infixl "⊕" 70)

This introduces a new class plus, along with a constant plus with nice infix
syntax. plus is also named class operation. The type of plus carries a class
constraint "'a :: plus" on its type variable, meaning that only types of class
plus can be instantiated for "'a". To breathe life into plus we need to declare
a type to be an instance of plus:
instantiation nat :: plus
begin

Command instantiation opens a local theory context. Here we can now

instantiate plus on nat:
primrec plus_nat :: "nat ⇒ nat ⇒ nat" where
"(0::nat) ⊕ n = n"
| "Suc m ⊕ n = Suc (m ⊕ n)"
 8.3 Type Classes 159

Note that the name plus carries a suffix _nat; by default, the local name of a
class operation f to be instantiated on type constructor κ is mangled as f_κ.
In case of uncertainty, these names may be inspected using the print context
command.
Although class plus has no axioms, the instantiation must be formally
concluded by a (trivial) instantiation proof “..”:
instance ..

More interesting instance proofs will arise below.

The instantiation is finished by an explicit
end

From now on, terms like Suc (m ⊕ 2) are legal.

instantiation prod :: (plus, plus) plus
begin

Here we instantiate the product type prod to class plus, given that its type
arguments are of class plus:
fun plus_prod :: "'a × 'b ⇒ 'a × 'b ⇒ 'a × 'b" where
"(x, y) ⊕ (w, z) = (x ⊕ w, y ⊕ z)"
Obviously, overloaded specifications may include recursion over the syntactic
structure of types.
instance ..

end

This way we have encoded the canonical lifting of binary operations to prod-
ucts by means of type classes.

8.3.2 Axioms

Attaching axioms to our classes lets us reason on the level of classes. The
results will be applicable to all types in a class, just as in axiomatic mathe-
matics.

!! Proofs in this section use structured Isar proofs, which are not covered in this
tutorial; but see [24].

Semigroups. We specify semigroups as subclass of plus:

class semigroup = plus +
assumes assoc: "(x ⊕ y) ⊕ z = x ⊕ (y ⊕ z)"

This class specification requires that all instances of semigroup obey assoc: " x
V

y z :: 'a::semigroup. (x ⊕ y) ⊕ z = x ⊕ (y ⊕ z)".
160 8. More about Types

We can use this class axiom to derive further abstract theorems relative
to class semigroup :
lemma assoc_left:
fixes x y z :: "'a::semigroup"
shows "x ⊕ (y ⊕ z) = (x ⊕ y) ⊕ z"
using assoc by (rule sym)
The semigroup constraint on type 'a restricts instantiations of 'a to types of
class semigroup and during the proof enables us to use the fact assoc whose
type parameter is itself constrained to class semigroup. The main advantage
of classes is that theorems can be proved in the abstract and freely reused
for each instance.
On instantiation, we have to give a proof that the given operations obey
the class axioms:
instantiation nat :: semigroup
begin

instance proof
The proof opens with a default proof step, which for instance judgements
invokes method intro_classes.
fix m n q :: nat
show "(m ⊕ n) ⊕ q = m ⊕ (n ⊕ q)"
by (induct m) simp_all
qed

end
Again, the interesting things enter the stage with parametric types:
instantiation prod :: (semigroup, semigroup) semigroup
begin

instance proof
fix p 1 p 2 p 3 :: "'a::semigroup × 'b::semigroup"
show "p 1 ⊕ p 2 ⊕ p 3 = p 1 ⊕ (p 2 ⊕ p 3 )"
by (cases p 1 , cases p 2 , cases p 3 ) (simp add: assoc)
Associativity of product semigroups is established using the hypothetical as-
sociativity assoc of the type components, which holds due to the semigroup
constraints imposed on the type components by the instance proposition.
Indeed, this pattern often occurs with parametric types and type classes.
qed

end
Monoids. We define a subclass monoidl (a semigroup with a left-hand neu-
tral) by extending semigroup with one additional parameter neutral together
with its property:
class monoidl = semigroup +
 8.3 Type Classes 161

fixes neutral :: "'a" ("0")

assumes neutl: "0 ⊕ x = x"
Again, we prove some instances, by providing suitable parameter definitions
and proofs for the additional specifications.
instantiation nat :: monoidl
begin

definition
neutral_nat_def: "0 = (0::nat)"

instance proof
fix n :: nat
show "0 ⊕ n = n"
unfolding neutral_nat_def by simp
qed

end
In contrast to the examples above, we here have both specification of class
operations and a non-trivial instance proof.
This covers products as well:
instantiation prod :: (monoidl, monoidl) monoidl
begin

definition
neutral_prod_def: "0 = (0, 0)"

instance proof
fix p :: "'a::monoidl × 'b::monoidl"
show "0 ⊕ p = p"
by (cases p) (simp add: neutral_prod_def neutl)
qed

end
Fully-fledged monoids are modelled by another subclass which does not add
new parameters but tightens the specification:
class monoid = monoidl +
assumes neutr: "x ⊕ 0 = x"
Corresponding instances for nat and products are left as an exercise to the
reader.

Groups. To finish our small algebra example, we add a group class:

class group = monoidl +
fixes inv :: "'a ⇒ 'a" ("÷ _" [81] 80)
assumes invl: "÷ x ⊕ x = 0"
We continue with a further example for abstract proofs relative to type
classes:
162 8. More about Types

lemma left_cancel:
fixes x y z :: "'a::group"
shows "x ⊕ y = x ⊕ z ←→ y = z"
proof
assume "x ⊕ y = x ⊕ z"
then have "÷ x ⊕ (x ⊕ y) = ÷ x ⊕ (x ⊕ z)" by simp
then have "(÷ x ⊕ x) ⊕ y = (÷ x ⊕ x) ⊕ z" by (simp add: assoc)
then show "y = z" by (simp add: invl neutl)
next
assume "y = z"
then show "x ⊕ y = x ⊕ z" by simp
qed

Any group is also a monoid; this can be made explicit by claiming an additional
subclass relation, together with a proof of the logical difference:
instance group ⊆ monoid
proof
fix x
from invl have "÷ x ⊕ x = 0" .
then have "÷ x ⊕ (x ⊕ 0) = ÷ x ⊕ x"
by (simp add: neutl invl assoc [symmetric])
then show "x ⊕ 0 = x" by (simp add: left_cancel)
qed
The proof result is propagated to the type system, making group an instance
of monoid by adding an additional edge to the graph of subclass relation; see
also Figure 8.1.

semigroup semigroup

? ?
monoidl monoidl
B
B
monoid B monoid
B PP
PP
BN P
q
P
group group
Fig. 8.1. Subclass relationship of monoids and groups: before and after establishing
the relationship group ⊆ monoid; transitive edges are left out.

Inconsistencies. The reader may be wondering what happens if we attach

an inconsistent set of axioms to a class. So far we have always avoided to add
new axioms to HOL for fear of inconsistencies and suddenly it seems that we
are throwing all caution to the wind. So why is there no problem?
The point is that by construction, all type variables in the axioms of a
class are automatically constrained with the class being defined (as shown for
 8.4 Numbers 163

axiom refl above). These constraints are always carried around and Isabelle
takes care that they are never lost, unless the type variable is instantiated
with a type that has been shown to belong to that class. Thus you may be
able to prove False from your axioms, but Isabelle will remind you that this
theorem has the hidden hypothesis that the class is non-empty.
Even if each individual class is consistent, intersections of (unrelated)
classes readily become inconsistent in practice. Now we know this need not
worry us.

Syntactic Classes and Predefined Overloading. In our algebra exam-

ple, we have started with a syntactic class plus which only specifies operations
but no axioms; it would have been also possible to start immediately with
class semigroup, specifying the ⊕ operation and associativity at the same
time.
Which approach is more appropriate depends. Usually it is more conve-
nient to introduce operations and axioms in the same class: then the type
checker will automatically insert the corresponding class constraints when-
ever the operations occur, reducing the need of manual annotations. However,
when operations are decorated with popular syntax, syntactic classes can be
an option to re-use the syntax in different contexts; this is indeed the way
most overloaded constants in HOL are introduced, of which the most impor-
tant are listed in Table A.2 in the appendix. Section 8.4.5 covers a range of
corresponding classes with axioms.
Further note that classes may contain axioms but no operations. An ex-
ample is class finite from theory HOL.Finite_Set which specifies a type to
be finite: "finite (UNIV :: 'a::finite set)".

8.4 Numbers

Until now, our numerical examples have used the type of natural num-
bers, nat. This is a recursive datatype generated by the constructors zero
and successor, so it works well with inductive proofs and primitive recursive
function definitions. HOL also provides the type int of integers, which lack
induction but support true subtraction. With subtraction, arithmetic reason-
ing is easier, which makes the integers preferable to the natural numbers for
complicated arithmetic expressions, even if they are non-negative. There are
also the types rat, real and complex : the rational, real and complex num-
bers. Isabelle has no subtyping, so the numeric types are distinct and there
are functions to convert between them. Most numeric operations are over-
loaded: the same symbol can be used at all numeric types. Table A.2 in the
appendix shows the most important operations, together with the priorities
of the infix symbols. Algebraic properties are organized using type classes
around algebraic concepts such as rings and fields; a property such as the
164 8. More about Types

commutativity of addition is a single theorem (add.commute) that applies to

all numeric types.
Many theorems involving numeric types can be proved automatically by
Isabelle’s arithmetic decision procedure, the method arith. Linear arithmetic
comprises addition, subtraction and multiplication by constant factors; sub-
terms involving other operators are regarded as variables. The procedure can
be slow, especially if the subgoal to be proved involves subtraction over type
nat, which causes case splits. On types nat and int, arith can deal with
quantifiers—this is known as Presburger arithmetic—whereas on type real
it cannot.
The simplifier reduces arithmetic expressions in other ways, such as di-
viding through by common factors. For problems that lie outside the scope
of automation, HOL provides hundreds of theorems about multiplication, di-
vision, etc., that can be brought to bear. You can locate them using Proof
General’s Find button. A few lemmas are given below to show what is avail-
able.

8.4.1 Numeric Literals

The constants 0 and 1 are overloaded. They denote zero and one, respectively,
for all numeric types. Other values are expressed by numeric literals, which
consist of one or more decimal digits optionally preceeded by a minus sign
(-). Examples are 2, -3 and 441223334678. Literals are available for the types
of natural numbers, integers, rationals, reals, etc.; they denote integer values
of arbitrary size.
Literals look like constants, but they abbreviate terms representing the
number in a two’s complement binary notation. Isabelle performs arithmetic
on literals by rewriting rather than using the hardware arithmetic. In most
cases arithmetic is fast enough, even for numbers in the millions. The arith-
metic operations provided for literals include addition, subtraction, multipli-
cation, integer division and remainder. Fractions of literals (expressed using
division) are reduced to lowest terms.

!! The arithmetic operators are overloaded, so you must be careful to ensure that
each numeric expression refers to a specific type, if necessary by inserting type
constraints. Here is an example of what can go wrong:

lemma "2 * m = m + m"

Carefully observe how Isabelle displays the subgoal:

1. (2::'a) * m = m + m
The type 'a given for the literal 2 warns us that no numeric type has been specified.
The problem is underspecified. Given a type constraint such as nat, int or real,
it becomes trivial.
 8.4 Numbers 165

!! Numeric literals are not constructors and therefore must not be used in pat-
terns. For example, this declaration is rejected:
function h where
"h 3 = 2"
|"h i = i"
You should use a conditional expression instead:
"h i = (if i = 3 then 2 else i)"

8.4.2 The Type of Natural Numbers, nat

This type requires no introduction: we have been using it from the beginning.
Hundreds of theorems about the natural numbers are proved in the theories
Nat and Divides. Basic properties of addition and multiplication are available
through the axiomatic type class for semirings (Sect. 8.4.5).
Literals. The notational options for the natural numbers are confusing. Re-
call that an overloaded constant can be defined independently for each type;
the definition of 1 for type nat is
1 ≡ Suc 0 (One_nat_def)
This is installed as a simplification rule, so the simplifier will replace every
occurrence of 1::nat by Suc 0. Literals are obviously better than nested Sucs
at expressing large values. But many theorems, including the rewrite rules
for primitive recursive functions, can only be applied to terms of the form
Suc n .
The following default simplification rules replace small literals by zero and
successor:
2 + n = Suc (Suc n) (add_2_eq_Suc)
n + 2 = Suc (Suc n) (add_2_eq_Suc')
It is less easy to transform 100 into Suc 99 (for example), and the simpli-
fier will normally reverse this transformation. Novices should express natural
numbers using 0 and Suc only.
Division. The infix operators div and mod are overloaded. Isabelle/HOL pro-
vides the basic facts about quotient and remainder on the natural numbers:
m mod n = (if m < n then m else (m - n) mod n) (mod_if)
m div n * n + m mod n = m (div_mult_mod_eq)
Many less obvious facts about quotient and remainder are also provided.
Here is a selection:
a * b div c = a * (b div c) + a * (b mod c) div c (div_mult1_eq)
a * b mod c = a * (b mod c) mod c (mod_mult_right_eq)
a div (b*c) = a div b div c (div_mult2_eq)
a mod (b*c) = b * (a div b mod c) + a mod b (mod_mult2_eq)
0 < c =⇒ (c * a) div (c * b) = a div b (div_mult_mult1)
(m mod n) * k = (m * k) mod (n * k) (mod_mult_distrib)
m ≤ n =⇒ m div k ≤ n div k (div_le_mono)
166 8. More about Types

Surprisingly few of these results depend upon the divisors’ being nonzero.
That is because division by zero yields zero:
a div 0 = 0 (DIVISION_BY_ZERO_DIV)
a mod 0 = a (DIVISION_BY_ZERO_MOD)
In div_mult_mult1 above, one of the two divisors (namely c) must still be
nonzero.
The divides relation has the standard definition, which is overloaded over
all numeric types:
m dvd n ≡ ∃ k. n = m * k (dvd_def)
Section 5.18 discusses proofs involving this relation. Here are some of the
facts proved about it:
[[m dvd n; n dvd m]] =⇒ m = n (dvd_antisym)
[[k dvd m; k dvd n]] =⇒ k dvd (m + n) (dvd_add)
Subtraction. There are no negative natural numbers, so m - n equals zero
unless m exceeds n. The following is one of the few facts about m - n that is
not subject to the condition n ≤ m.
(m - n) * k = m * k - n * k (diff_mult_distrib)
Natural number subtraction has few nice properties; often you should remove
it by simplifying with this split rule.
P(a-b) = ((a<b −→ P 0) ∧ (∀ d. a = b+d −→ P d)) (nat_diff_split)
For example, splitting helps to prove the following fact.
lemma "(n - 2) * (n + 2) = n * n - (4::nat)"
apply
V (simp split: nat_diff_split, clarify)
1. d. [[n < 2; n * n = 4 + d]] =⇒ d = 0
The result lies outside the scope of linear arithmetic, but it is easily found if
we explicitly split n<2 as n=0 or n=1:
apply (subgoal_tac "n=0 | n=1", force, arith)
done

8.4.3 The Type of Integers, int

Reasoning methods for the integers resemble those for the natural numbers,
but induction and the constant Suc are not available. HOL provides many
lemmas for proving inequalities involving integer multiplication and division,
similar to those shown above for type nat. The laws of addition, subtraction
and multiplication are available through the axiomatic type class for rings
(Sect. 8.4.5).
The absolute value function abs is overloaded, and is defined for all types
that involve negative numbers, including the integers. The arith method can
prove facts about abs automatically, though as it does so by case analysis,
the cost can be exponential.
 8.4 Numbers 167

lemma "abs (x+y) ≤ abs x + abs (y :: int)"

by arith
For division and remainder, the treatment of negative divisors follows
mathematical practice: the sign of the remainder follows that of the divisor:
0 < b =⇒ 0 ≤ a mod b (pos_mod_sign)
0 < b =⇒ a mod b < b (pos_mod_bound)
b < 0 =⇒ a mod b ≤ 0 (neg_mod_sign)
b < 0 =⇒ b < a mod b (neg_mod_bound)
ML treats negative divisors in the same way, but most computer hardware
treats signed operands using the same rules as for multiplication. Many facts
about quotients and remainders are provided:
(a + b) div c =
a div c + b div c + (a mod c + b mod c) div c (div_add1_eq)
(a + b) mod c = (a mod c + b mod c) mod c (mod_add_eq)

(a * b) div c = a * (b div c) + a * (b mod c) div c (zdiv_zmult1_eq)

(a * b) mod c = a * (b mod c) mod c (zmod_zmult1_eq)

0 < c =⇒ a div (b*c) = a div b div c (zdiv_zmult2_eq)

0 < c =⇒ a mod (b*c) = b*(a div b mod c) + a mod b (zmod_zmult2_eq)

The last two differ from their natural number analogues by requiring c to
be positive. Since division by zero yields zero, we could allow c to be zero.
However, c cannot be negative: a counterexample is a = 7, b = 2 and c = −3,
when the left-hand side of zdiv_zmult2_eq is −2 while the right-hand side
is −1. The prefix z in many theorem names recalls the use of Z to denote the
set of integers.
Induction is less important for integers than it is for the natural numbers,
but it can be valuable if the range of integers has a lower or upper bound.
There are four rules for integer induction, corresponding to the possible re-
lations of the bound (≥, >, ≤ and <):
≤ i; P k; i. i; P i]] =⇒ P(i+1)]] =⇒ P i (int_ge_induct)
V
[[k V [[k ≤
[[k < i; P(k+1);
V i. [[k < i; P i]] =⇒ P(i+1)]] =⇒ P i (int_gr_induct)
[[i ≤ k; P k; i. V [[i ≤ k; P i]] =⇒ P(i-1)]] =⇒ P i (int_le_induct)
[[i < k; P(k-1); i. [[i < k; P i]] =⇒ P(i-1)]] =⇒ P i (int_less_induct)

8.4.4 The Types of Rational, Real and Complex Numbers

These types provide true division, the overloaded operator /, which differs
from the operator div of the natural numbers and integers. The rationals
and reals are dense: between every two distinct numbers lies another. This
property follows from the division laws, since if x 6= y then (x + y)/2 lies
between them:
a < b =⇒ ∃ r. a < r ∧ r < b (dense)
168 8. More about Types

The real numbers are, moreover, complete: every set of reals that is
bounded above has a least upper bound. Completeness distinguishes the reals
from the rationals,
√ for which the set {x | x < 2} has no least upper bound.
2

(It could only be 2, which is irrational.) The formalization of completeness,

which is complicated, can be found in theory RComplete.
Numeric literals for type real have the same syntax as those for type
int and only express integral values. Fractions expressed using the division
operator are automatically simplified to lowest terms:
1. P ((3 / 4) * (8 / 15))
apply simp
1. P (2 / 5)
Exponentiation can express floating-point values such as 2 * 10^6, which will
be simplified to integers.

!! Types rat, real and complex are provided by theory HOL-Complex, which is
Main extended with a definitional development of the rational, real and complex
numbers. Base your theory upon theory Complex_Main, not the usual Main.
Available in the logic HOL-NSA is the theory Hyperreal, which define the
type hypreal of non-standard reals. These hyperreals include infinitesimals,
which represent infinitely small and infinitely large quantities; they facili-
tate proofs about limits, differentiation and integration [8]. The development
defines an infinitely large number, omega and an infinitely small positive num-
ber, epsilon. The relation x ≈ y means “x is infinitely close to y.” Theory
Hyperreal also defines transcendental functions such as sine, cosine, expo-
nential and logarithm — even the versions for type real, because they are
defined using nonstandard limits.

8.4.5 The Numeric Type Classes

Isabelle/HOL organises its numeric theories using axiomatic type classes.

Hundreds of basic properties are proved in the theory Ring_and_Field. These
lemmas are available (as simprules if they were declared as such) for all
numeric types satisfying the necessary axioms. The theory defines dozens of
type classes, such as the following:
– semiring and ordered_semiring : a semiring provides the associative oper-
ators + and *, with 0 and 1 as their respective identities. The operators
enjoy the usual distributive law, and addition (+ ) is also commutative. An
ordered semiring is also linearly ordered, with addition and multiplication
respecting the ordering. Type nat is an ordered semiring.
– ring and ordered_ring : a ring extends a semiring with unary minus (the
additive inverse) and subtraction (both denoted -). An ordered ring in-
cludes the absolute value function, abs. Type int is an ordered ring.
 8.4 Numbers 169

– field and ordered_field: a field extends a ring with the multiplicative

inverse (called simply inverse and division (/ )). An ordered field is based
on an ordered ring. Type complex is a field, while type real is an ordered
field.
– division_by_zero includes all types where inverse 0 = 0 and a / 0 = 0.
These include all of Isabelle’s standard numeric types. However, the basic
properties of fields are derived without assuming division by zero.
Hundreds of basic lemmas are proved, each of which holds for all types in
the corresponding type class. In most cases, it is obvious whether a property
is valid for a particular type. No abstract properties involving subtraction
hold for type nat; instead, theorems such as diff_mult_distrib are proved
specifically about subtraction on type nat. All abstract properties involving
division require a field. Obviously, all properties involving orderings required
an ordered structure.
The class ring_no_zero_divisors of rings without zero divisors satisfies
a number of natural cancellation laws, the first of which characterizes this
class:
(a * b = (0::'a)) = (a = (0::'a) ∨ b = (0::'a)) (mult_eq_0_iff)
(a * c = b * c) = (c = (0::'a) ∨ a = b) (mult_cancel_right)

Setting the flag Isabelle > Settings > Show Sorts will display the type classes of
all type variables.
Here is how the theorem mult_cancel_left appears with the flag set.
((c::'a::ring_no_zero_divisors) * (a::'a::ring_no_zero_divisors) =
c * (b::'a::ring_no_zero_divisors)) =
(c = (0::'a::ring_no_zero_divisors) ∨ a = b)
Simplifying with the AC-Laws. Suppose that two expressions are equal,
differing only in associativity and commutativity of addition. Simplifying
with the following equations sorts the terms and groups them to the right,
making the two expressions identical.
a + b + c = a + (b + c) (add.assoc)
a + b = b + a (add.commute)
a + (b + c) = b + (a + c) (add.left_commute)
The name ac_simps refers to the list of all three theorems; similarly there is
ac_simps. They are all proved for semirings and therefore hold for all numeric
types.
Here is an example of the sorting effect. Start with this goal, which in-
volves type nat.
1. Suc (i + j * l * k + m * n) = f (n * m + i + k * j * l)
Simplify using ac_simps and ac_simps.
apply (simp add: ac_simps ac_simps)
Here is the resulting subgoal.
1. Suc (i + (m * n + j * (k * l))) = f (i + (m * n + j * (k * l)))
170 8. More about Types

Division Laws for Fields. Here is a selection of rules about the division
operator. The following are installed as default simplification rules in order
to express combinations of products and quotients as rational expressions:
a * (b / c) = a * b / c (times_divide_eq_right)
b / c * a = b * a / c (times_divide_eq_left)
a / (b / c) = a * c / b (divide_divide_eq_right)
a / b / c = a / (b * c) (divide_divide_eq_left)
Signs are extracted from quotients in the hope that complementary terms
can then be cancelled:
- (a / b) = - a / b (minus_divide_left)
- (a / b) = a / - b (minus_divide_right)
The following distributive law is available, but it is not installed as a
simplification rule.
(a + b) / c = a / c + b / c (add_divide_distrib)
Absolute Value. The absolute value function abs is available for all ordered
rings, including types int, rat and real. It satisfies many properties, such as
the following:
|x * y| = |x| * |y| (abs_mult)
(|a| ≤ b) = (a ≤ b ∧ - a ≤ b) (abs_le_iff)
|a + b| ≤ |a| + |b| (abs_triangle_ineq)

!! The absolute value bars shown above cannot be typed on a keyboard. They can
be entered using the X-symbol package. In ascii, type abs x to get |x|.

Raising to a Power. Another type class, ordered idom, specifies rings that
also have exponentation to a natural number power, defined using the obvi-
ous primitive recursion. Theory Power proves various theorems, such as the
following.
a ^ (m + n) = a ^ m * a ^ n (power_add)
a ^ (m * n) = (a ^ m) ^ n (power_mult)
|a ^ n| = |a| ^ n (power_abs)

8.5 Introducing New Types

For most applications, a combination of predefined types like bool and ⇒
with recursive datatypes and records is quite sufficient. Very occasionally
you may feel the need for a more advanced type. If you are certain that your
type is not definable by any of the standard means, then read on.

!! Types in HOL must be non-empty; otherwise the quantifier rules would be

unsound, because ∃x. x = x is a theorem.
 8.5 Introducing New Types 171

8.5.1 Declaring New Types

The most trivial way of introducing a new type is by a type declaration:

typedecl my_new_type

This does not define my_new_type at all but merely introduces its name. Thus
we know nothing about this type, except that it is non-empty. Such declara-
tions without definitions are useful if that type can be viewed as a parameter
of the theory. A typical example is given in Sect. ??, where we define a tran-
sition relation over an arbitrary type of states.
In principle we can always get rid of such type declarations by making
those types parameters of every other type, thus keeping the theory generic.
In practice, however, the resulting clutter can make types hard to read.
If you are looking for a quick and dirty way of introducing a new type to-
gether with its properties: declare the type and state its properties as axioms.
Example:
axiomatization where
just_one: "∃ x::my_new_type. ∀ y. x = y"

However, we strongly discourage this approach, except at explorative stages

of your development. It is extremely easy to write down contradictory sets
of axioms, in which case you will be able to prove everything but it will
mean nothing. In the example above, the axiomatic approach is unnecessary:
a one-element type called unit is already defined in HOL.

8.5.2 Defining New Types

Now we come to the most general means of safely introducing a new type,
the type definition. All other means, for example datatype, are based on
it. The principle is extremely simple: any non-empty subset of an existing
type can be turned into a new type. More precisely, the new type is specified
to be isomorphic to some non-empty subset of an existing type.
Let us work a simple example, the definition of a three-element type. It
is easily represented by the first three natural numbers:
typedef three = "{0::nat, 1, 2}"

In order to enforce that the representing set on the right-hand side is non-
empty, this definition actually starts a proof to that effect:
1. ∃ x. x ∈ {0, 1, 2}

Fortunately, this is easy enough to show, even auto could do it. In general,
one has to provide a witness, in our case 0:
apply(rule_tac x = 0 in exI)
by simp
172 8. More about Types

This type definition introduces the new type three and asserts that it is a
copy of the set {0, 1, 2}. This assertion is expressed via a bijection between
the type three and the set {0, 1, 2}. To this end, the command declares the
following constants behind the scenes:
Rep_three :: three ⇒ nat
Abs_three :: nat ⇒ three
The situation is best summarized with the help of the following diagram,
where squares denote types and the irregular region denotes a set:

nat

three {0,1,2}

Finally, typedef asserts that Rep_three is surjective on the subset and

Abs_three and Rep_three are inverses of each other:

Rep_three x ∈ {0, 1, 2} (Rep_three)

Abs_three (Rep_three x) = x (Rep_three_inverse)
y ∈ {0, 1, 2} =⇒ Rep_three (Abs_three y) = y (Abs_three_inverse)
From this example it should be clear what typedef does in general given a
name (here three) and a set (here {0, 1, 2}).
Our next step is to define the basic functions expected on the new type.
Although this depends on the type at hand, the following strategy works well:
– define a small kernel of basic functions that can express all other functions
you anticipate.
– define the kernel in terms of corresponding functions on the representing
type using Abs and Rep to convert between the two levels.
In our example it suffices to give the three elements of type three names:
definition A :: three where "A ≡ Abs_three 0"
definition B :: three where "B ≡ Abs_three 1"
definition C :: three where "C ≡ Abs_three 2"

So far, everything was easy. But it is clear that reasoning about three
will be hell if we have to go back to nat every time. Thus our aim must be to
raise our level of abstraction by deriving enough theorems about type three
to characterize it completely. And those theorems should be phrased in terms
of A, B and C, not Abs_three and Rep_three. Because of the simplicity of the
 8.5 Introducing New Types 173

example, we merely need to prove that A, B and C are distinct and that they
exhaust the type.
In processing our typedef declaration, Isabelle proves several helpful lem-
mas. The first two express injectivity of Rep_three and Abs_three:
(Rep_three x = Rep_three y) = (x = y) (Rep_three_inject)
[[x ∈ {0, 1, 2}; y ∈ {0, 1, 2} ]]
(Abs_three_inject)
=⇒ (Abs_three x = Abs_three y) = (x = y)
The following ones allow to replace some x::three by Abs_three(y::nat), and
conversely y by Rep_three x :

[[y ∈ {0, 1, 2}; x. y = Rep_three x =⇒ P]] =⇒ P (Rep_three_cases)

V

( y. [[x = Abs_three y; y ∈ {0, 1, 2}]] =⇒ P) =⇒ P (Abs_three_cases)

V

[[y ∈ {0, 1, 2}; x. P (Rep_three x)]] =⇒ P y (Rep_three_induct)

V

( y. y ∈ {0, 1, 2} =⇒ P (Abs_three y)) =⇒ P x (Abs_three_induct)

V

These theorems are proved for any type definition, with three replaced by
the name of the type in question.
Distinctness of A, B and C follows immediately if we expand their defini-
tions and rewrite with the injectivity of Abs_three:
lemma "A 6= B ∧ B 6= A ∧ A 6= C ∧ C 6= A ∧ B 6= C ∧ C 6= B"
by(simp add: Abs_three_inject A_def B_def C_def)
Of course we rely on the simplifier to solve goals like 0 6= 1.
The fact that A, B and C exhaust type three is best phrased as a case
distinction theorem: if you want to prove P x (where x is of type three) it
suffices to prove P A, P B and P C :
lemma three_cases: "[[ P A; P B; P C ]] =⇒ P x"
Again this follows easily using the induction principle stemming from the
type definition:
apply(induct_tac x)

1. y. [[P A; P B; P C; y ∈ {0, 1, 2}]] =⇒ P (Abs_three y)

V

Simplification leads to the disjunction y = 0 ∨ y = 1 ∨ y = 2 which auto

separates into three subgoals, each of which is easily solved by simplification:
apply(auto simp add: A_def B_def C_def)
done
This concludes the derivation of the characteristic theorems for type three.
The attentive reader has realized long ago that the above lengthy defini-
tion can be collapsed into one line:
datatype better_three = A | B | C
In fact, the datatype command performs internally more or less the same
derivations as we did, which gives you some idea what life would be like
without datatype.
174 8. More about Types

Although three could be defined in one line, we have chosen this exam-
ple to demonstrate typedef because its simplicity makes the key concepts
particularly easy to grasp. If you would like to see a non-trivial example
that cannot be defined more directly, we recommend the definition of finite
multisets in the Library [4].
Let us conclude by summarizing the above procedure for defining a new
type. Given some abstract axiomatic description P of a type ty in terms of a
set of functions F , this involves three steps:
1. Find an appropriate type τ and subset A which has the desired properties
P, and make a type definition based on this representation.
2. Define the required functions F on ty by lifting analogous functions on
the representation via Abs ty and Rep ty.
3. Prove that P holds for ty by lifting P from the representation.
You can now forget about the representation and work solely in terms of the
abstract functions F and properties P.
9. Advanced Simplification and Induction

Although we have already learned a lot about simplification and induction,

there are some advanced proof techniques that we have not covered yet and
which are worth learning. The sections of this chapter are independent of
each other and can be read in any order.

9.1 Simplification

This section describes features not covered until now. It also outlines the
simplification process itself, which can be helpful when the simplifier does
not do what you expect of it.

9.1.1 Advanced Features

Congruence Rules. While simplifying the conclusion Q of P =⇒ Q, it is

legal to use the assumption P. For =⇒ this policy is hardwired, but contextual
information can also be made available for other operators. For example, xs
= [] −→ xs @ xs = xs simplifies to True because we may use xs = [] when
simplifying xs @ xs = xs. The generation of contextual information during
simplification is controlled by so-called congruence rules. This is the one
for −→:
[[P = P'; P' =⇒ Q = Q']] =⇒ (P −→ Q) = (P' −→ Q')
It should be read as follows: In order to simplify P −→ Q to P' −→ Q', sim-
plify P to P' and assume P' when simplifying Q to Q'.
Here are some more examples. The congruence rules for bounded quanti-
fiers supply contextual information about the bound variable:
[[A = B; x. x ∈ B =⇒ P x = Q x]]
V
=⇒ (∀ x∈A. P x) = (∀ x∈B. Q x)
One congruence rule for conditional expressions supplies contextual informa-
tion for simplifying the then and else cases:
[[b = c; c =⇒ x = u; ¬ c =⇒ y = v]]
=⇒ (if b then x else y) = (if c then u else v)
176 9. Advanced Simplification and Induction

An alternative congruence rule for conditional expressions actually prevents

simplification of some arguments:
b = c =⇒ (if b then x else y) = (if c then x else y)

Only the first argument is simplified; the others remain unchanged. This
makes simplification much faster and is faithful to the evaluation strategy in
programming languages, which is why this is the default congruence rule for
if. Analogous rules control the evaluation of case expressions.
You can declare your own congruence rules with the attribute cong , either
globally, in the usual manner,
declare theorem-name [cong]
or locally in a simp call by adding the modifier
cong: list of theorem names

The effect is reversed by cong del instead of cong.

!! The congruence rule conj_cong

[[P = P'; P' =⇒ Q = Q']] =⇒ (P ∧ Q) = (P' ∧ Q')
is occasionally useful but is not a default rule; you have to declare it explicitly.

Permutative Rewrite Rules. An equation is a permutative rewrite

rule if the left-hand side and right-hand side are the same up to renaming of
variables. The most common permutative rule is commutativity: x + y = y +
x. Other examples include x - y - z = x - z - y in arithmetic and insert
x (insert y A) = insert y (insert x A) for sets. Such rules are problematic
because once they apply, they can be used forever. The simplifier is aware of
this danger and treats permutative rules by means of a special strategy, called
ordered rewriting: a permutative rewrite rule is only applied if the term
becomes smaller with respect to a fixed lexicographic ordering on terms. For
example, commutativity rewrites b + a to a + b, but then stops because a +
b is strictly smaller than b + a. Permutative rewrite rules can be turned into
simplification rules in the usual manner via the simp attribute; the simplifier
recognizes their special status automatically.
Permutative rewrite rules are most effective in the case of associative-
commutative functions. (Associativity by itself is not permutative.) When
dealing with an AC-function f , keep the following points in mind:
– The associative law must always be oriented from left to right, namely
f (f (x, y), z) = f (x, f (y, z)). The opposite orientation, if used with commu-
tativity, can lead to nontermination.
– To complete your set of rewrite rules, you must add not just associativ-
ity (A) and commutativity (C) but also a derived rule, left-commutativ-
ity (LC): f (x, f (y, z)) = f (y, f (x, z)).
 9.1 Simplification 177

Ordered rewriting with the combination of A, C, and LC sorts a term lexi-

cographically:
A C LC
f (f (b, c), a) ; f (b, f (c, a)) ; f (b, f (a, c)) ; f (a, f (b, c))

Note that ordered rewriting for + and * on numbers is rarely necessary

because the built-in arithmetic prover often succeeds without such tricks.

9.1.2 How the Simplifier Works

Roughly speaking, the simplifier proceeds bottom-up: subterms are simplified

first. A conditional equation is only applied if its condition can be proved,
again by simplification. Below we explain some special features of the rewrit-
ing process.

Higher-Order Patterns. So far we have pretended the simplifier can deal

with arbitrary rewrite rules. This is not quite true. For reasons of feasibility,
the simplifier expects the left-hand side of each rule to be a so-called higher-
order pattern [23]. This restricts where unknowns may occur. Higher-order
patterns are terms in β-normal form. (This means there are no subterms of the
form (λx.M )(N ).) Each occurrence of an unknown is of the form ?f x1 . . . xn ,
where the xi are distinct bound variables. Thus all ordinary rewrite rules,
where all unknowns are of base type, for example ?a + ?b + ?c = ?a + (?b
+ ?c), are acceptable: if an unknown is of base type, it cannot have any
arguments. Additionally, the rule (∀ x. ?P x ∧ ?Q x) = ((∀ x. ?P x) ∧ (∀ x.
?Q x)) is also acceptable, in both directions: all arguments of the unknowns
?P and ?Q are distinct bound variables.
If the left-hand side is not a higher-order pattern, all is not lost. The
simplifier will still try to apply the rule provided it matches directly: with-
out much λ-calculus hocus pocus. For example, (?f ?x ∈ range ?f) = True
rewrites g a ∈ range g to True, but will fail to match g(h b) ∈ range(λx.
g(h x)). However, you can eliminate the offending subterms — those that are
not patterns — by adding new variables and conditions. In our example, we
eliminate ?f ?x and obtain ?y = ?f ?x =⇒ (?y ∈ range ?f) = True, which
is fine as a conditional rewrite rule since conditions can be arbitrary terms.
However, this trick is not a panacea because the newly introduced conditions
may be hard to solve.
There is no restriction on the form of the right-hand sides. They may not
contain extraneous term or type variables, though.

The Preprocessor. When a theorem is declared a simplification rule, it

need not be a conditional equation already. The simplifier will turn it into a
set of conditional equations automatically. For example, f x = g x ∧ h x =
k x becomes the two separate simplification rules f x = g x and h x = k x.
In general, the input theorem is converted as follows:
178 9. Advanced Simplification and Induction

¬P 7→ P = False
P −→ Q 7→ P =⇒ Q
P ∧Q 7→ P, Q
∀x. P x 7→ P ?x
∀x ∈ A. P x 7→ ?x ∈ A =⇒ P ?x
if P then Q else R 7→ P =⇒ Q, ¬P =⇒ R

Once this conversion process is finished, all remaining non-equations P are

turned into trivial equations P = True. For example, the formula
(p −→ t = u ∧ ¬ r) ∧ s
is converted into the three rules
p =⇒ t = u, p =⇒ r = False, s = True.

9.2 Advanced Induction Techniques

Now that we have learned about rules and logic, we take another look at
the finer points of induction. We consider two questions: what to do if the
proposition to be proved is not directly amenable to induction (Sect. 9.2.1),
and how to utilize (Sect. 9.2.2) and even derive (Sect. 9.2.3) new induction
schemas. We conclude with an extended example of induction (Sect. 9.2.4).

9.2.1 Massaging the Proposition

Often we have assumed that the theorem to be proved is already in a form

that is amenable to induction, but sometimes it isn’t. Here is an example.
Since hd and last return the first and last element of a non-empty list, this
lemma looks easy to prove:
lemma "xs 6= [] =⇒ hd(rev xs) = last xs"
apply(induct_tac xs)

But induction produces the warning

Induction variable occurs also among premises!
and leads to the base case
1. xs 6= [] =⇒ hd (rev []) = last []
A total of 2 subgoals...

Simplification reduces the base case to this:

1. xs 6= [] =⇒ hd [] = last []
 9.2 Advanced Induction Techniques 179

We cannot prove this equality because we do not know what hd and last
return when applied to [].
We should not have ignored the warning. Because the induction formula
is only the conclusion, induction does not affect the occurrence of xs in the
premises. Thus the case that should have been trivial becomes unprovable.
Fortunately, the solution is easy:1
Pull all occurrences of the induction variable into the conclusion using
−→.
Thus we should state the lemma as an ordinary implication (−→), letting
rule_format (Sect. 5.15) convert the result to the usual =⇒ form:
lemma hd_rev [rule_format]: "xs 6= [] −→ hd(rev xs) = last xs"
This time, induction leaves us with a trivial base case:
1. [] 6= [] −→ hd (rev []) = last []
A total of 2 subgoals...
And auto completes the proof.
If there are multiple premises A1 , …, An containing the induction variable,
you should turn the conclusion C into

A1 −→ · · · An −→ C .

Additionally, you may also have to universally quantify some other variables,
which can yield a fairly complex conclusion. However, rule_format can remove
any number of occurrences of ∀ and −→.
A second reason why your proposition may not be amenable to induction
is that you want to induct on a complex term, rather than a variable. In
general, induction on a term t requires rephrasing the conclusion C as

∀y1 . . . yn . x = t −→ C . (9.1)

where y1 . . . yn are the free variables in t and x is a new variable. Now you
can perform induction on x. An example appears in Sect. 9.2.2 below.
The very same problem may occur in connection with rule induction.
Remember that it requires a premise of the form (x1 , . . . , xk ) ∈ R, where R
is some inductively defined set and the xi are variables. If instead we have a
premise t ∈ R, where t is not just an n-tuple of variables, we replace it with
(x1 , . . . , xk ) ∈ R, and rephrase the conclusion C as

∀y1 . . . yn . (x1 , . . . , xk ) = t −→ C .

For an example see Sect. 9.2.4 below.

Of course, all premises that share free variables with t need to be pulled
into the conclusion as well, under the ∀ , again using −→ as shown above.
1
A similar heuristic applies to rule inductions; see Sect. 7.2.
180 9. Advanced Simplification and Induction

Readers who are puzzled by the form of statement (9.1) above should
remember that the transformation is only performed to permit induction.
Once induction has been applied, the statement can be transformed back
into something quite intuitive. For example, applying wellfounded induction
on x (w.r.t. ≺) to (9.1) and transforming the result a little leads to the goal

y. ∀z. t z ≺ t y −→ C z =⇒ C y
^

where y stands for y1 . . . yn and the dependence of t and C on the free vari-
ables of t has been made explicit. Unfortunately, this induction schema cannot
be expressed as a single theorem because it depends on the number of free
variables in t — the notation y is merely an informal device.

9.2.2 Beyond Structural and Recursion Induction

So far, inductive proofs were by structural induction for primitive recursive
functions and recursion induction for total recursive functions. But some-
times structural induction is awkward and there is no recursive function that
could furnish a more appropriate induction schema. In such cases a general-
purpose induction schema can be helpful. We show how to apply such induc-
tion schemas by an example.
Structural induction on nat is usually known as mathematical induc-
tion. There is also complete induction, where you prove P(n) under the
assumption that P(m) holds for all m < n. In Isabelle, this is the theorem
nat_less_induct:
( n. ∀ m<n. P m =⇒ P n) =⇒ P n
V

As an application, we prove a property of the following function:

axiomatization f :: "nat ⇒ nat"
where f_ax: "f(f(n)) < f(Suc(n))" for n :: nat

!! We discourage the use of axioms because of the danger of inconsistencies. Ax-

iom f_ax does not introduce an inconsistency because, for example, the identity
function satisfies it. Axioms can be useful in exploratory developments, say when
you assume some well-known theorems so that you can quickly demonstrate some
point about methodology. If your example turns into a substantial proof develop-
ment, you should replace axioms by theorems.
The axiom for f implies n ≤ f n, which can be proved by induction on f n.
Following the recipe outlined above, we have to phrase the proposition as
follows to allow induction:
lemma f_incr_lem: "∀ i. k = f i −→ i ≤ f i"
To perform induction on k using nat_less_induct, we use the same general
induction method as for recursion induction (see Sect. 3.5.4):
apply(induct_tac k rule: nat_less_induct)
We get the following proof state:
 9.2 Advanced Induction Techniques 181

1. n. ∀ m<n. ∀ i. m = f i −→ i ≤ f i =⇒ ∀ i. n = f i −→ i ≤ f i
V

After stripping the ∀ i, the proof continues with a case distinction on i. The
case i = 0 is trivial and we focus on the other case:
apply(rule allI)
apply(case_tac i)
apply(simp)

1. n i nat.
V
[[∀ m<n. ∀ i. m = f i −→ i ≤ f i; i = Suc nat]] =⇒ n = f i −→ i ≤
f i

by(blast intro!: f_ax Suc_leI intro: le_less_trans)

If you find the last step puzzling, here are the two lemmas it employs:
m < n =⇒ Suc m ≤ n (Suc_leI)
[[x ≤ y; y < z]] =⇒ x < z (le_less_trans)
The proof goes like this (writing j instead of nat). Since i = Suc j it suf-
fices to show j < f (Suc j), by Suc_leI . This is proved as follows. From f_ax
we have f (f j) < f (Suc j) (1) which implies f j ≤ f (f j) by the induc-
tion hypothesis. Using (1) once more we obtain f j < f (Suc j) (2) by the
transitivity rule le_less_trans. Using the induction hypothesis once more
we obtain j ≤ f j which, together with (2) yields j < f (Suc j) (again by
le_less_trans).
This last step shows both the power and the danger of automatic proofs.
They will usually not tell you how the proof goes, because it can be hard to
translate the internal proof into a human-readable format. Automatic proofs
are easy to write but hard to read and understand.
The desired result, i ≤ f i, follows from f_incr_lem:
lemmas f_incr = f_incr_lem[rule_format, OF refl]
The final refl gets rid of the premise ?k = f ?i. We could have included this
derivation in the original statement of the lemma:
lemma f_incr[rule_format, OF refl]: "∀ i. k = f i −→ i ≤ f i"

Exercise 9.2.1 From the axiom and lemma for f, show that f is the identity
function.
Method induct_tac can be applied with any rule r whose conclusion is of
the form ?P ?x1 . . .?xn , in which case the format is
apply(induct_tac y1 . . . yn rule: r )
where y1 , . . . , yn are variables in the conclusion of the first subgoal.
A further useful induction rule is length_induct, induction on the length
of a list
( xs. ∀ ys. length ys < length xs −→ P ys =⇒ P xs) =⇒ P xs
V
182 9. Advanced Simplification and Induction

which is a special case of measure_induct

( x. ∀ y. f y < f x −→ P y =⇒ P x) =⇒ P a
V

where f may be any function into type nat.

9.2.3 Derivation of New Induction Schemas

Induction schemas are ordinary theorems and you can derive new ones
whenever you wish. This section shows you how, using the example of
nat_less_induct. Assume we only have structural induction available for nat
and want to derive complete induction. We must generalize the statement as
shown:
lemma induct_lem: "( n::nat. ∀ m<n. P m =⇒ P n) =⇒ ∀ m<n. P m"
V
apply(induct_tac n)
The base case is vacuously true. For the induction step (m < Suc n) we dis-
tinguish two cases: case m < n is true by induction hypothesis and case m = n
follows from the assumption, again using the induction hypothesis:
apply(blast)
by(blast elim: less_SucE)

The elimination rule less_SucE expresses the case distinction:

[[m < Suc n; m < n =⇒ P; m = n =⇒ P]] =⇒ P

Now it is straightforward to derive the original version of nat_less_induct

by manipulating the conclusion of the above lemma: instantiate n by Suc n
and m by n and remove the trivial condition n < Suc n. Fortunately, this
happens automatically when we add the lemma as a new premise to the
desired goal:
theorem nat_less_induct: "( n::nat. ∀ m<n. P m =⇒ P n) =⇒ P n"
V
by(insert induct_lem, blast)

HOL already provides the mother of all inductions, well-founded in-

duction (see Sect. 6.4). For example theorem nat_less_induct is a special
case of wf_induct where r is < on nat. The details can be found in theory
Wellfounded_Recursion.

9.2.4 CTL Revisited

The purpose of this section is twofold: to demonstrate some of the induction

principles and heuristics discussed above and to show how inductive defi-
nitions can simplify proofs. In Sect. 6.5.2 we gave a fairly involved proof of
the correctness of a model checker for CTL. In particular the proof of the
infinity_lemma on the way to AF_lemma2 is not as simple as one might expect,
 9.2 Advanced Induction Techniques 183

due to the SOME operator involved. Below we give a simpler proof of AF_lemma2
based on an auxiliary inductive definition.
Let us call a (finite or infinite) path A-avoiding if it does not touch any
node in the set A. Then AF_lemma2 says that if no infinite path from some state
s is A-avoiding, then s ∈ lfp (af A). We prove this by inductively defining
the set Avoid s A of states reachable from s by a finite A-avoiding path:
inductive set
Avoid :: "state ⇒ state set ⇒ state set"
for s :: state and A :: "state set"
where
"s ∈ Avoid s A"
| "[[ t ∈ Avoid s A; t ∈
/ A; (t,u) ∈ M ]] =⇒ u ∈ Avoid s A"
It is easy to see that for any infinite A-avoiding path f with f 0 ∈ Avoid
s A there is an infinite A-avoiding path starting with s because (by definition
of Avoid) there is a finite A-avoiding path from s to f 0. The proof is by
induction on f 0 ∈ Avoid s A. However, this requires the following reformu-
lation, as explained in Sect. 9.2.1 above; the rule_format directive undoes the
reformulation after the proof.
lemma ex_infinite_path[rule_format]:
"t ∈ Avoid s A =⇒
∀ f∈Paths t. (∀ i. f i ∈
/ A) −→ (∃ p∈Paths s. ∀ i. p i ∈
/ A)"
apply(erule Avoid.induct)
apply(blast)
apply(clarify)
apply(drule_tac x = "λi. case i of 0 ⇒ t | Suc i ⇒ f i" in bspec)
apply(simp_all add: Paths_def split: nat.split)
done
The base case (t = s) is trivial and proved by blast. In the induction step,
we have an infinite A-avoiding path f starting from u, a successor of t. Now
we simply instantiate the ∀ f∈Paths t in the induction hypothesis by the
path starting with t and continuing with f. That is what the above λ-term
expresses. Simplification shows that this is a path starting with t and that
the instantiated induction hypothesis implies the conclusion.
Now we come to the key lemma. Assuming that no infinite A-avoiding
path starts from s, we want to show s ∈ lfp (af A). For the inductive proof
this must be generalized to the statement that every point t “between” s and
A, in other words all of Avoid s A, is contained in lfp (af A):
lemma Avoid_in_lfp[rule_format(no_asm)]:
"∀ p∈Paths s. ∃ i. p i ∈ A =⇒ t ∈ Avoid s A −→ t ∈ lfp(af A)"
The proof is by induction on the “distance” between t and A. Remember that
lfp (af A) = A ∪ M −1 `` lfp (af A). If t is already in A, then t ∈ lfp (af
A) is trivial. If t is not in A but all successors are in lfp (af A) (induction
hypothesis), then t ∈ lfp (af A) is again trivial.
The formal counterpart of this proof sketch is a well-founded induction
on M restricted to Avoid s A - A, roughly speaking:
184 9. Advanced Simplification and Induction

{(y, x). (x, y) ∈ M ∧ x ∈ Avoid s A ∧ x ∈

/ A}

As we shall see presently, the absence of infinite A-avoiding paths starting

from s implies well-foundedness of this relation. For the moment we assume
this and proceed with the induction:
apply(subgoal_tac "wf{(y,x). (x,y) ∈ M ∧ x ∈ Avoid s A ∧ x ∈
/ A}")
apply(erule_tac a = t in wf_induct)
apply(clarsimp)

1. t. [[∀ p∈Paths s. ∃ i. p i ∈ A;
V
∀ y. (t, y) ∈ M ∧ t ∈/ A −→
y ∈ Avoid s A −→ y ∈ lfp (af A);
t ∈ Avoid s A]]
=⇒ t ∈ lfp (af A)
2. ∀ p∈Paths s. ∃ i. p i ∈ A =⇒
wf {(y, x). (x, y) ∈ M ∧ x ∈ Avoid s A ∧ x ∈ / A}

Now the induction hypothesis states that if t ∈ / A then all successors of t

that are in Avoid s A are in lfp (af A). Unfolding lfp in the conclusion of
the first subgoal once, we have to prove that t is in A or all successors of
t are in lfp (af A). But if t is not in A, the second Avoid-rule implies that
all successors of t are in Avoid s A, because we also assume t ∈ Avoid s A.
Hence, by the induction hypothesis, all successors of t are indeed in lfp (af
A). Mechanically:
apply(subst lfp_unfold[OF mono_af])
apply(simp (no_asm) add: af_def)
apply(blast intro: Avoid.intros)

Having proved the main goal, we return to the proof obligation that the
relation used above is indeed well-founded. This is proved by contradiction: if
the relation is not well-founded then there exists an infinite A-avoiding path
all in Avoid s A, by theorem wf_iff_no_infinite_down_chain:
wf r = (@ f. ∀ i. (f (Suc i), f i) ∈ r)

From lemma ex_infinite_path the existence of an infinite A-avoiding path

starting in s follows, contradiction.
apply(erule contrapos_pp)
apply(simp add: wf_iff_no_infinite_down_chain)
apply(erule exE)
apply(rule ex_infinite_path)
apply(auto simp add: Paths_def)
done

The (no_asm) modifier of the rule_format directive in the statement of

the lemma means that the assumption is left unchanged; otherwise the ∀ p
would be turned into a p, which would complicate matters below. As it is,
V

Avoid_in_lfp is now
[[∀ p∈Paths s. ∃ i. p i ∈ A; t ∈ Avoid s A]] =⇒ t ∈ lfp (af A)
 9.2 Advanced Induction Techniques 185

The main theorem is simply the corollary where t = s, when the assumption
t ∈ Avoid s A is trivially true by the first Avoid-rule. Isabelle confirms this:
theorem AF_lemma2: "{s. ∀ p ∈ Paths s. ∃ i. p i ∈ A} ⊆ lfp(af A)"
by(auto elim: Avoid_in_lfp intro: Avoid.intros)
10. Case Study: Verifying a Security Protocol

Communications security is an ancient art. Julius Caesar is said to have

encrypted his messages, shifting each letter three places along the alphabet.
Mary Queen of Scots was convicted of treason after a cipher used in her
letters was broken. Today’s postal system incorporates security features. The
envelope provides a degree of secrecy. The signature provides authenticity
(proof of origin), as do departmental stamps and letterheads.
Networks are vulnerable: messages pass through many computers, any of
which might be controlled by an adversary, who thus can capture or redirect
messages. People who wish to communicate securely over such a network can
use cryptography, but if they are to understand each other, they need to
follow a protocol: a pre-arranged sequence of message formats.
Protocols can be attacked in many ways, even if encryption is unbreak-
able. A splicing attack involves an adversary’s sending a message composed
of parts of several old messages. This fake message may have the correct for-
mat, fooling an honest party. The adversary might be able to masquerade as
somebody else, or he might obtain a secret key.
Nonces help prevent splicing attacks. A typical nonce is a 20-byte random
number. Each message that requires a reply incorporates a nonce. The reply
must include a copy of that nonce, to prove that it is not a replay of a past
message. The nonce in the reply must be cryptographically protected, since
otherwise an adversary could easily replace it by a different one. You should
be starting to see that protocol design is tricky!
Researchers are developing methods for proving the correctness of security
protocols. The Needham-Schroeder public-key protocol [22] has become a
standard test case. Proposed in 1978, it was found to be defective nearly two
decades later [18]. This toy protocol will be useful in demonstrating how to
verify protocols using Isabelle.

10.1 The Needham-Schroeder Public-Key Protocol

This protocol uses public-key cryptography. Each person has a private key,
known only to himself, and a public key, known to everybody. If Alice wants
to send Bob a secret message, she encrypts it using Bob’s public key (which
188 10. Case Study: Verifying a Security Protocol

everybody knows), and sends it to Bob. Only Bob has the matching private
key, which is needed in order to decrypt Alice’s message.
The core of the Needham-Schroeder protocol consists of three messages:

1. A → B : {|Na, A|}Kb
2. B → A : {|Na, Nb|}Ka
3. A → B : {|Nb|}Kb

First, let’s understand the notation. In the first message, Alice sends Bob a
message consisting of a nonce generated by Alice (Na) paired with Alice’s
name (A) and encrypted using Bob’s public key (Kb). In the second message,
Bob sends Alice a message consisting of Na paired with a nonce generated
by Bob (Nb), encrypted using Alice’s public key (Ka). In the last message,
Alice returns Nb to Bob, encrypted using his public key.
When Alice receives Message 2, she knows that Bob has acted on her mes-
sage, since only he could have decrypted {|Na, A|}Kb and extracted Na. That
is precisely what nonces are for. Similarly, message 3 assures Bob that Alice
is active. But the protocol was widely believed [7] to satisfy a further prop-
erty: that Na and Nb were secrets shared by Alice and Bob. (Many protocols
generate such shared secrets, which can be used to lessen the reliance on
slow public-key operations.) Lowe found this claim to be false: if Alice runs
the protocol with someone untrustworthy (Charlie say), then he can start
a new run with another agent (Bob say). Charlie uses Alice as an oracle,
masquerading as Alice to Bob [18].

1. A → C : {|Na, A|}Kc 10 . C → B : {|Na, A|}Kb

2. B → A : {|Na, Nb|}Ka
3. A → C : {|Nb|}Kc 30 . C → B : {|Nb|}Kb

In messages 1 and 3, Charlie removes the encryption using his private key
and re-encrypts Alice’s messages using Bob’s public key. Bob is left thinking
he has run the protocol with Alice, which was not Alice’s intention, and Bob
is unaware that the “secret” nonces are known to Charlie. This is a typical
man-in-the-middle attack launched by an insider.
Whether this counts as an attack has been disputed. In protocols of
this type, we normally assume that the other party is honest. To be hon-
est means to obey the protocol rules, so Alice’s running the protocol with
Charlie does not make her dishonest, just careless. After Lowe’s attack, Alice
has no grounds for complaint: this protocol does not have to guarantee any-
thing if you run it with a bad person. Bob does have grounds for complaint,
however: the protocol tells him that he is communicating with Alice (who is
honest) but it does not guarantee secrecy of the nonces.
Lowe also suggested a correction, namely to include Bob’s name in mes-
sage 2:
 10.2 Agents and Messages 189

1. A → B : {|Na, A|}Kb
2. B → A : {|Na, Nb, B|}Ka
3. A → B : {|Nb|}Kb

If Charlie tries the same attack, Alice will receive the message {|Na, Nb, B|}Ka
when she was expecting to receive {|Na, Nb, C }| Ka . She will abandon the run,
and eventually so will Bob. Below, we shall look at parts of this protocol’s
correctness proof.
In ground-breaking work, Lowe [18] showed how such attacks could be
found automatically using a model checker. An alternative, which we shall
examine below, is to prove protocols correct. Proofs can be done under more
realistic assumptions because our model does not have to be finite. The strat-
egy is to formalize the operational semantics of the system and to prove
security properties using rule induction.

10.2 Agents and Messages

All protocol specifications refer to a syntactic theory of messages. Datatype
agent introduces the constant Server (a trusted central machine, needed for
some protocols), an infinite population of friendly agents, and the Spy :
datatype agent = Server | Friend nat | Spy
Keys are just natural numbers. Function invKey maps a public key to the
matching private key, and vice versa:
type synonym key = nat
consts invKey :: "key ⇒ key"
Datatype msg introduces the message forms, which include agent names,
nonces, keys, compound messages, and encryptions.
datatype
msg = Agent agent
| Nonce nat
| Key key
| MPair msg msg
| Crypt key msg
The notation {|X1 , . . . Xn−1 , Xn }| abbreviates MPair X1 . . . (MPair Xn−1 Xn ).
Since datatype constructors are injective, we have the theorem
Crypt K X = Crypt K' X' =⇒ K = K' ∧ X = X'
A ciphertext can be decrypted using only one key and can yield only one
plaintext. In the real world, decryption with the wrong key succeeds but
yields garbage. Our model of encryption is realistic if encryption adds some
redundancy to the plaintext, such as a checksum, so that garbage can be
detected.
190 10. Case Study: Verifying a Security Protocol

10.3 Modelling the Adversary

The spy is part of the system and must be built into the model. He is a mali-
cious user who does not have to follow the protocol. He watches the network
and uses any keys he knows to decrypt messages. Thus he accumulates ad-
ditional keys and nonces. These he can use to compose new messages, which
he may send to anybody.
Two functions enable us to formalize this behaviour: analz and synth.
Each function maps a sets of messages to another set of messages. The set
analz H formalizes what the adversary can learn from the set of messages H .
The closure properties of this set are defined inductively.
inductive set
analz :: "msg set ⇒ msg set"
for H :: "msg set"
where
Inj [intro,simp] : "X ∈ H =⇒ X ∈ analz H"
| Fst: "{|X,Y |} ∈ analz H =⇒ X ∈ analz H"
| Snd: "{|X,Y |} ∈ analz H =⇒ Y ∈ analz H"
| Decrypt [dest]:
"[[Crypt K X ∈ analz H; Key(invKey K) ∈ analz H]]
=⇒ X ∈ analz H"

Note the Decrypt rule: the spy can decrypt a message encrypted with
key K if he has the matching key, K −1 . Properties proved by rule induction
include the following:
G ⊆ H =⇒ analz G ⊆ analz H (analz_mono)
analz (analz H) = analz H (analz_idem)
The set of fake messages that an intruder could invent starting from H is
synth(analz H), where synth H formalizes what the adversary can build from
the set of messages H .
inductive set
synth :: "msg set ⇒ msg set"
for H :: "msg set"
where
Inj [intro]: "X ∈ H =⇒ X ∈ synth H"
| Agent [intro]: "Agent agt ∈ synth H"
| MPair [intro]:
"[[X ∈ synth H; Y ∈ synth H]] =⇒ {|X,Y |} ∈ synth H"
| Crypt [intro]:
"[[X ∈ synth H; Key K ∈ H]] =⇒ Crypt K X ∈ synth H"
The set includes all agent names. Nonces and keys are assumed to be
unguessable, so none are included beyond those already in H . Two elements
of synth H can be combined, and an element can be encrypted using a key
present in H .
Like analz, this set operator is monotone and idempotent. It also satisfies
an interesting equation involving analz :
 10.4 Event Traces 191

analz (synth H) = analz H ∪ synth H (analz_synth)

Rule inversion plays a major role in reasoning about synth, through declara-
tions such as this one:
inductive cases Nonce_synth [elim!]: "Nonce n ∈ synth H"
The resulting elimination rule replaces every assumption of the form Nonce n
∈ synth H by Nonce n ∈ H, expressing that a nonce cannot be guessed.
A third operator, parts, is useful for stating correctness properties. The set
parts H consists of the components of elements of H . This set includes H and
is closed under the projections from a compound message to its immediate
parts. Its definition resembles that of analz except in the rule corresponding
to the constructor Crypt:
Crypt K X ∈ parts H =⇒ X ∈ parts H
The body of an encrypted message is always regarded as part of it. We can
use parts to express general well-formedness properties of a protocol, for
example, that an uncompromised agent’s private key will never be included
as a component of any message.

10.4 Event Traces

The system’s behaviour is formalized as a set of traces of events. The most

important event, Says A B X, expresses A → B : X , which is the attempt
by A to send B the message X . A trace is simply a list, constructed in
reverse using #. Other event types include reception of messages (when we
want to make it explicit) and an agent’s storing a fact.
Sometimes the protocol requires an agent to generate a new nonce. The
probability that a 20-byte random number has appeared before is effectively
zero. To formalize this important property, the set used evs denotes the set of
all items mentioned in the trace evs. The function used has a straightforward
recursive definition. Here is the case for Says event:
used (Says A B X # evs) = parts {X} ∪ used evs
The function knows formalizes an agent’s knowledge. Mostly we only care
about the spy’s knowledge, and knows Spy evs is the set of items available
to the spy in the trace evs. Already in the empty trace, the spy starts with
some secrets at his disposal, such as the private keys of compromised users.
After each Says event, the spy learns the message that was sent:
knows Spy (Says A B X # evs) = insert X (knows Spy evs)
Combinations of functions express other important sets of messages derived
from evs:
– analz (knows Spy evs) is everything that the spy could learn by decryption
192 10. Case Study: Verifying a Security Protocol

– synth (analz (knows Spy evs)) is everything that the spy could generate

The function pubK maps agents to their public keys. The function priK
maps agents to their private keys. It is merely an abbreviation (cf. Sect. 4.1.4)
defined in terms of invKey and pubK.
consts pubK :: "agent ⇒ key"
abbreviation priK :: "agent ⇒ key"
where "priK x ≡ invKey(pubK x)"

The set bad consists of those agents whose private keys are known to the spy.
Two axioms are asserted about the public-key cryptosystem. No two
agents have the same public key, and no private key equals any public key.
axiomatization where
inj_pubK: "inj pubK" and
priK_neq_pubK: "priK A 6= pubK B"

10.5 Modelling the Protocol

inductive set ns_public :: "event list set"

where

Nil: "[] ∈ ns_public"

| Fake: "[[evsf ∈ ns_public; X ∈ synth (analz (knows Spy evsf))]]

=⇒ Says Spy B X # evsf ∈ ns_public"

| NS1: "[[evs1 ∈ ns_public; Nonce NA ∈

/ used evs1]]
=⇒ Says A B (Crypt (pubK B) {|Nonce NA, Agent A|})
# evs1 ∈ ns_public"

| NS2: "[[evs2 ∈ ns_public; Nonce NB ∈ / used evs2;

Says A' B (Crypt (pubK B) {|Nonce NA, Agent A|}) ∈ set evs2]]
=⇒ Says B A (Crypt (pubK A) {|Nonce NA, Nonce NB, Agent B|})
# evs2 ∈ ns_public"

| NS3: "[[evs3 ∈ ns_public;

Says A B (Crypt (pubK B) {|Nonce NA, Agent A|}) ∈ set evs3;
Says B' A (Crypt (pubK A) {|Nonce NA, Nonce NB, Agent B|})
∈ set evs3]]
=⇒ Says A B (Crypt (pubK B) (Nonce NB)) # evs3 ∈ ns_public"

Fig. 10.1. An Inductive Protocol Definition

 10.6 Proving Elementary Properties 193

Let us formalize the Needham-Schroeder public-key protocol, as corrected

by Lowe:

1. A → B : {|Na, A|}Kb
2. B → A : {|Na, Nb, B|}Ka
3. A → B : {|Nb|}Kb

Each protocol step is specified by a rule of an inductive definition. An

event trace has type event list, so we declare the constant ns_public to be
a set of such traces.
Figure 10.1 presents the inductive definition. The Nil rule introduces the
empty trace. The Fake rule models the adversary’s sending a message built
from components taken from past traffic, expressed using the functions synth
and analz. The next three rules model how honest agents would perform the
three protocol steps.
Here is a detailed explanation of rule NS2. A trace containing an event of
the form
Says A' B (Crypt (pubK B) {|Nonce NA, Agent A|})

may be extended by an event of the form

Says B A (Crypt (pubK A) {|Nonce NA, Nonce NB, Agent B|})

where NB is a fresh nonce: Nonce NB ∈ / used evs2. Writing the sender as A'
indicates that B does not know who sent the message. Calling the trace vari-
able evs2 rather than simply evs helps us know where we are in a proof after
many case-splits: every subgoal mentioning evs2 involves message 2 of the
protocol.
Benefits of this approach are simplicity and clarity. The semantic model
is set theory, proofs are by induction and the translation from the informal
notation to the inductive rules is straightforward.

10.6 Proving Elementary Properties

Secrecy properties can be hard to prove. The conclusion of a typical secrecy

theorem is X ∈ / analz (knows Spy evs). The difficulty arises from having to
reason about analz, or less formally, showing that the spy can never learn X.
Much easier is to prove that X can never occur at all. Such regularity prop-
erties are typically expressed using parts rather than analz.
The following lemma states that A’s private key is potentially known to the
spy if and only if A belongs to the set bad of compromised agents. The state-
ment uses parts: the very presence of A’s private key in a message, whether
protected by encryption or not, is enough to confirm that A is compromised.
The proof, like nearly all protocol proofs, is by induction over traces.
194 10. Case Study: Verifying a Security Protocol

lemma Spy_see_priK [simp]:

"evs ∈ ns_public
=⇒ (Key (priK A) ∈ parts (knows Spy evs)) = (A ∈ bad)"
apply (erule ns_public.induct, simp_all)
The induction yields five subgoals, one for each rule in the definition of
ns_public. The idea is to prove that the protocol property holds initially (rule
Nil), is preserved by each of the legitimate protocol steps (rules NS1–3 ), and
even is preserved in the face of anything the spy can do (rule Fake).
The proof is trivial. No legitimate protocol rule sends any keys at all,
so only Fake is relevant. Indeed, simplification leaves only the Fake case, as
indicated by the variable name evsf :
1. evsf X.
V
[[evsf ∈ ns_public;
(Key (priK A) ∈ parts (knows Spy evsf)) = (A ∈ bad);
X ∈ synth (analz (knows Spy evsf))]]
=⇒ (Key (priK A) ∈ parts (insert X (knows Spy evsf))) =
(A ∈ bad)

by blast
The Fake case is proved automatically. If priK A is in the extended trace
then either (1) it was already in the original trace or (2) it was generated by
the spy, who must have known this key already. Either way, the induction
hypothesis applies.
Unicity lemmas are regularity lemmas stating that specified items can
occur only once in a trace. The following lemma states that a nonce cannot
be used both as Na and as Nb unless it is known to the spy. Intuitively,
it holds because honest agents always choose fresh values as nonces; only
the spy might reuse a value, and he doesn’t know this particular value. The
proof script is short: induction, simplification, blast. The first line uses the
rule rev_mp to prepare the induction by moving two assumptions into the
induction formula.
lemma no_nonce_NS1_NS2:
"[[Crypt (pubK C) {|NA', Nonce NA, Agent D|} ∈ parts (knows Spy evs);
Crypt (pubK B) {|Nonce NA, Agent A|} ∈ parts (knows Spy evs);
evs ∈ ns_public]]
=⇒ Nonce NA ∈ analz (knows Spy evs)"
apply (erule rev_mp, erule rev_mp)
apply (erule ns_public.induct, simp_all)
apply (blast intro: analz_insertI)+
done
The following unicity lemma states that, if NA is secret, then its appearance
in any instance of message 1 determines the other components. The proof is
similar to the previous one.
lemma unique_NA:
"[[Crypt(pubK B) {|Nonce NA, Agent A |} ∈ parts(knows Spy evs);
Crypt(pubK B') {|Nonce NA, Agent A'|} ∈ parts(knows Spy evs);
 10.7 Proving Secrecy Theorems 195

Nonce NA ∈
/ analz (knows Spy evs); evs ∈ ns_public]]
=⇒ A=A' ∧ B=B'"

10.7 Proving Secrecy Theorems

The secrecy theorems for Bob (the second participant) are especially impor-
tant because they fail for the original protocol. The following theorem states
that if Bob sends message 2 to Alice, and both agents are uncompromised,
then Bob’s nonce will never reach the spy.
theorem Spy_not_see_NB [dest]:
"[[Says B A (Crypt (pubK A) {|Nonce NA, Nonce NB, Agent B|}) ∈ set evs;
A ∈
/ bad; B ∈ / bad; evs ∈ ns_public]]
=⇒ Nonce NB ∈ / analz (knows Spy evs)"

To prove it, we must formulate the induction properly (one of the assump-
tions mentions evs), apply induction, and simplify:
apply (erule rev_mp, erule ns_public.induct, simp_all)

The proof states are too complicated to present in full. Let’s examine the
simplest subgoal, that for message 1. The following event has just occurred:

1. A0 → B 0 : {|Na 0 , A0}| Kb0

The variables above have been primed because this step belongs to a different
run from that referred to in the theorem statement — the theorem refers to a
past instance of message 2, while this subgoal concerns message 1 being sent
just now. In the Isabelle subgoal, instead of primed variables like B 0 and Na 0
we have Ba and NAa:
1. evs1 NAa Ba.
V
/ bad; B ∈
[[A ∈ / bad; evs1 ∈ ns_public;
Says B A (Crypt (pubK A) {|Nonce NA, Nonce NB, Agent B|})
∈ set evs1 −→
Nonce NB ∈ / analz (knows Spy evs1);
Nonce NAa ∈ / used evs1]]
=⇒ Ba ∈ bad −→
Says B A (Crypt (pubK A) {|Nonce NA, Nonce NB, Agent B|})
∈ set evs1 −→
NB 6= NAa

The simplifier has used a default simplification rule that does a case analysis
for each encrypted message on whether or not the decryption key is compro-
mised.
analz (insert (Crypt K X) H) =
(if Key (invKey K) ∈ analz H
then insert (Crypt K X) (analz (insert X H))
else insert (Crypt K X) (analz H)) (analz_Crypt_if)
196 10. Case Study: Verifying a Security Protocol

The simplifier has also used Spy_see_priK, proved in Sect. 10.6 above, to yield
Ba ∈ bad.
Recall that this subgoal concerns the case where the last message to be
sent was
1. A0 → B 0 : {|Na 0 , A0}| Kb0 .
This message can compromise Nb only if Nb = Na 0 and B 0 is compromised,
allowing the spy to decrypt the message. The Isabelle subgoal says precisely
this, if we allow for its choice of variable names. Proving NB 6= NAa is easy: NB
was sent earlier, while NAa is fresh; formally, we have the assumption Nonce
NAa ∈/ used evs1.
Note that our reasoning concerned B ’s participation in another run.
Agents may engage in several runs concurrently, and some attacks work by
interleaving the messages of two runs. With model checking, this possibility
can cause a state-space explosion, and for us it certainly complicates proofs.
The biggest subgoal concerns message 2. It splits into several cases, such as
whether or not the message just sent is the very message mentioned in the
theorem statement. Some of the cases are proved by unicity, others by the
induction hypothesis. For all those complications, the proofs are automatic
by blast with the theorem no_nonce_NS1_NS2.
The remaining theorems about the protocol are not hard to prove. The
following one asserts a form of authenticity: if B has sent an instance of
message 2 to A and has received the expected reply, then that reply really
originated with A. The proof is a simple induction.
theorem B_trusts_NS3:
"[[Says B A (Crypt (pubK A) {|Nonce NA, Nonce NB, Agent B|}) ∈ set evs;
Says A' B (Crypt (pubK B) (Nonce NB)) ∈ set evs;
A ∈
/ bad; B ∈ / bad; evs ∈ ns_public]]
=⇒ Says A B (Crypt (pubK B) (Nonce NB)) ∈ set evs"
From similar assumptions, we can prove that A started the protocol run
by sending an instance of message 1 involving the nonce NA. For this theorem,
the conclusion is
Says A B (Crypt (pubK B) {|Nonce NA, Agent A|}) ∈ set evs
Analogous theorems can be proved for A, stating that nonce NA remains se-
cret and that message 2 really originates with B. Even the flawed protocol
establishes these properties for A; the flaw only harms the second participant.
Detailed information on this protocol verification technique can be found
elsewhere [30], including proofs of an Internet protocol [31]. We must stress
that the protocol discussed in this chapter is trivial. There are only three
messages; no keys are exchanged; we merely have to prove that encrypted
data remains secret. Real world protocols are much longer and distribute
many secrets to their participants. To be realistic, the model has to include
the possibility of keys being lost dynamically due to carelessness. If those
keys have been used to encrypt other sensitive information, there may be
 197

cascading losses. We may still be able to establish a bound on the losses and
to prove that other protocol runs function correctly [32]. Proofs of real-world
protocols follow the strategy illustrated above, but the subgoals can be much
bigger and there are more of them.
198
 199

You know my methods. Apply them!

Sherlock Holmes
A. Appendix

[[ [| \<lbrakk>
]] |] \<rbrakk>
=⇒ ==> \<Longrightarrow>
!! \<And>
V
≡ == \<equiv>
== \<rightleftharpoons>
* => \<rightharpoonup>
) <= \<leftharpoondown>
λ % \<lambda>
⇒ => \<Rightarrow>
∧ & \<and>
∨ | \<or>
−→ --> \<longrightarrow>
¬ ~ \<not>
6 = ~= \<noteq>
∀ ALL, ! \<forall>
∃ EX, ? \<exists>
∃! EX!, ?! \<exists>!
ε SOME, @ \<epsilon>
◦ o \<circ>
|| abs \<bar> \<bar>
≤ <= \<le>
× * \<times>
∈ : \<in>
∈
/ ~: \<notin>
⊆ <= \<subseteq>
⊂ < \<subset>
∪ Un \<union>
∩ Int \<inter>
UN, Union \<Union>
S
INT, Inter \<Inter>
T
∗
^* \<^sup>*
−1
^-1 \<inverse>

Table A.1. Mathematical Symbols, Their ascii-Equivalents and Internal Names

 Constant Type Syntax
0 'a::zero
1 'a::one
plus 'a::plus ⇒ 'a::plus ⇒ 'a::plus (infixl + 65)
minus 'a::minus ⇒ 'a::minus ⇒ 'a::minus (infixl − 65)
uminus 'a::uminus ⇒ 'a::uminus −x
times 'a::times ⇒ 'a::times ⇒ 'a::times (infixl ∗ 70)
inverse_divide 'a::inverse ⇒ 'a::inverse ⇒ 'a::inverse (infixl / 70)
divide 'a::divide ⇒ 'a::divide ⇒ 'a::divide (infixl div 70)
modulo 'a::modulo ⇒ 'a::modulo ⇒ 'a::modulo (infixl mod 70)
abs 'a::abs ⇒ 'a::abs |x|
sgn 'a::sgn ⇒ 'a::sgn
less_eq 'a::ord ⇒ 'a::ord ⇒ bool (infixl ≤ 50)
less 'a::ord ⇒ 'a::ord ⇒ bool (infixl < 50)
top 'a::top
bot 'a::bot
Table A.2. Important Overloaded Constants in Main

ALL BIT CHR EX GREATEST INT Int LEAST O

OFCLASS PI PROP SIGMA SOME THE TYPE UN Un
WRT case choose div dvd else funcset if in
let mem mod o of op then

Table A.3. Reserved Words in HOL Terms

Bibliography

[1] David Aspinall. Proof General. http://proofgeneral.inf.ed.ac.uk/.

[2] David Aspinall. Proof General: A generic tool for proof development.
In Tools and Algorithms for the Construction and Analysis of Systems
(TACAS), volume 1785 of Lecture Notes in Computer Science, pages
38–42. Springer-Verlag, 2000.
[3] Franz Baader and Tobias Nipkow. Term Rewriting and All That. Cam-
bridge University Press, 1998.
[4] Gertrud Bauer, Tobias Nipkow, David von Oheimb, Lawrence C Paulson,
Thomas M Rasmussen, Christophe Tabacznyj, and Markus Wenzel. The
supplemental Isabelle/HOL library. Part of the Isabelle distribution,
https://isabelle.in.tum.de/library/HOL/Library/document.pdf, 2002.
[5] Richard Bird. Introduction to Functional Programming using Haskell.
Prentice-Hall, 1998.
[6] Lukas Bulwahn, Alexander Krauss, and Tobias Nipkow. Finding lexi-
cographic orders for termination proofs in Isabelle/HOL. In K. Schnei-
der and J. Brandt, editors, Theorem Proving in Higher Order Logics:
TPHOLs 2007, volume 4732 of Lecture Notes in Computer Science, pages
38–53. Springer-Verlag, 2007.
[7] M. Burrows, M. Abadi, and R. M. Needham. A logic of authentication.
Proceedings of the Royal Society of London, 426:233–271, 1989.
[8] Jacques Fleuriot and Lawrence C. Paulson. Mechanizing nonstandard
real analysis. LMS Journal of Computation and Mathematics, 3:140–190,
2000. http://www.lms.ac.uk/jcm/3/lms1999-027/.
[9] Jean-Yves Girard. Proofs and Types. Cambridge University Press, 1989.
Translated by Yves Lafont and Paul Taylor.
[10] M. J. C. Gordon and T. F. Melham, editors. Introduction to HOL:
A Theorem Proving Environment for Higher Order Logic. Cambridge
University Press, 1993.
[11] Florian Haftmann. Haskell-style type classes with Isabelle/Isar. https:
//isabelle.in.tum.de/doc/classes.pdf.
[12] David Harel, Dexter Kozen, and Jerzy Tiuryn. Dynamic Logic. MIT
Press, 2000.
[13] John E. Hopcroft and Jeffrey D. Ullman. Introduction to Automata
Theory, Languages, and Computation. Addison-Wesley, 1979.
204 BIBLIOGRAPHY

[14] Paul Hudak. The Haskell School of Expression. Cambridge University

Press, 2000.
[15] Michael Huth and Mark Ryan. Logic in Computer Science. Modelling
and reasoning about systems. Cambridge University Press, 2000.
[16] Donald E. Knuth. The Art of Computer Programming, Volume 3: Sorting
and Searching. Addison-Wesley, 1975.
[17] Alexander Krauss. Defining Recursive Functions in Isabelle/HOL. https:
//isabelle.in.tum.de/doc/functions.pdf.
[18] Gavin Lowe. Breaking and fixing the Needham-Schroeder public-key
protocol using CSP and FDR. In T. Margaria and B. Steffen, editors,
Tools and Algorithms for the Construction and Analysis of Systems:
second international workshop, TACAS ’96, LNCS 1055, pages 147–166.
Springer, 1996.
[19] Robin Milner, Mads Tofte, and Robert Harper. The Definition of Stan-
dard ML. MIT Press, 1990.
[20] Olaf Müller, Tobias Nipkow, David von Oheimb, and Oscar Slotosch.
HOLCF = HOL + LCF. Journal of Functional Programming, 9:191–223,
1999.
[21] Wolfgang Naraschewski and Markus Wenzel. Object-oriented verifica-
tion based on record subtyping in higher-order logic. In Jim Grundy
and Malcom Newey, editors, Theorem Proving in Higher Order Log-
ics: TPHOLs ’98, volume 1479 of Lecture Notes in Computer Science.
Springer-Verlag, 1998.
[22] Roger M. Needham and Michael D. Schroeder. Using encryption for
authentication in large networks of computers. Communications of the
ACM, 21(12):993–999, December 1978.
[23] Tobias Nipkow. Functional unification of higher-order patterns. In
M. Vardi, editor, Eighth Annual Symposium on Logic in Computer Sci-
ence, pages 64–74. IEEE Computer Society Press, 1993.
[24] Tobias Nipkow. Structured Proofs in Isar/HOL. In H. Geuvers and
F. Wiedijk, editors, Types for Proofs and Programs (TYPES 2002),
volume 2646 of Lecture Notes in Computer Science, pages 259–278.
Springer-Verlag, 2003.
[25] Tobias Nipkow, Lawrence C. Paulson, and Markus Wenzel. Isabelle’s
Logics: HOL. https://isabelle.in.tum.de/doc/logics-HOL.pdf.
[26] Lawrence C. Paulson. The Old Isabelle Reference Manual. https://
isabelle.in.tum.de/doc/ref.pdf.
[27] Lawrence C. Paulson. Logic and Computation: Interactive proof with
Cambridge LCF. Cambridge University Press, 1987.
[28] Lawrence C. Paulson. Isabelle: A Generic Theorem Prover. Springer,
1994. LNCS 828.
[29] Lawrence C. Paulson. ML for the Working Programmer. Cambridge
University Press, 2nd edition, 1996. https://www.cl.cam.ac.uk/~lp15/
MLbook.
 BIBLIOGRAPHY 205

[30] Lawrence C. Paulson. The inductive approach to verifying cryptographic

protocols. Journal of Computer Security, 6:85–128, 1998.
[31] Lawrence C. Paulson. Inductive analysis of the Internet protocol TLS.
ACM Transactions on Information and System Security, 2(3):332–351,
August 1999.
[32] Lawrence C. Paulson. Relations between secrets: Two formal analyses
of the Yahalom protocol. Journal of Computer Security, 9(3):197–216,
2001.
[33] F. J. Pelletier. Seventy-five problems for testing automatic theorem
provers. Journal of Automated Reasoning, 2:191–216, 1986. Errata,
JAR 4 (1988), 235–236 and JAR 18 (1997), 135.
[34] Kenneth H. Rosen. Discrete Mathematics and Its Applications. McGraw-
Hill, 1998.
[35] Simon Thompson. Haskell: The Craft of Functional Programming.
Addison-Wesley, 1999.
[36] Makarius Wenzel. The Isabelle System Manual. https://isabelle.in.tum.
de/doc/system.pdf.
[37] Makarius Wenzel. The Isabelle/Isar Reference Manual. https://isabelle.
in.tum.de/doc/isar-ref.pdf.
[38] Markus Wenzel. Isabelle/Isar — a versatile environment for human-
readable formal proof documents. PhD thesis, Institut für Informatik,
Technische Universität München, 2002. https://mediatum.ub.tum.de/
doc/601724/601724.pdf.
Index

!, 201 abandoning a proof, 13

?, 201 abbreviation (command), 56
∃! , 201 abs (constant), 166, 168, 170
?!, 201 abs, 201
&, 201 absolute value, 166, 170
~, 201 ac_simps (theorems), 169
~=, 201 Ackermann’s function, 49
|, 201 add (modifier), 29
*, 22 add.assoc (theorem), 169
+, 22 add.commute (theorem), 169
-, 22 ALL, 201
<=, 201 All (constant), 111
<, 22 allE (theorem), 81
>, 23 allI (theorem), 81
≥, 23 antiquotation, 61
≤, 22 append function, 11–15
[], 10 apply (command), 16
#, 10 arg_cong (theorem), 97
@, 11, 201 arith (method), 23, 164
/ 201
∈, arithmetic operations
~:, 201 – for nat, 22
S , 201 ascii symbols, 201
T
, 201 Aspinall, David, iii
−1
, 201 associative-commutative function, 176
^-1, 201 assumption (method), 69
∗
, 201 assumptions
^*, 201 – of subgoal, 13
, 13, 201 – renaming, 83
V
!!, 201 – reusing, 84
⇒, 5 auto (method), 39, 93
[|, 201 axclass, 158
|], 201 axiom of choice, 87
%, 201 axiomatic type classes, 158
() (constant), 24
+ (tactical), 100 back (command), 79
<*lex*>, see lexicographic product Ball (constant), 111
? (tactical), 101 ballI (theorem), 110
| (tactical), 101 best (method), 93
Bex (constant), 111
0 (constant), 22, 164 bexE (theorem), 110
1 (constant), 22, 164, 165 bexI (theorem), 110
 Index 207

bij_def (theorem), 112 datatype (command), 9, 39–44

bijections, 112 datatypes, 16–22
binary trees, 18 – and nested recursion, 41, 45
binomial coefficients, 111 – mutually recursive, 39
bisimulations, 118 defer (command), 16, 102
blast (method), 90–91, 93 definition (command), 25
bool (type), 5 Definitional Approach, 25
boolean expressions example, 19–22 definitions, 25
bspec (theorem), 110 – unfolding, 30
by (command), 73 del (modifier), 29
description operators, 86–88
card (constant), 111 descriptions
card_Pow (theorem), 111 – definite, 86
card_Un_Int (theorem), 111 – indefinite, 87
cardinality, 111 dest (attribute), 104
case (symbol), 31, 32 destruction rules, 71–72
case expressions, 5, 6, 18 diff_mult_distrib (theorem), 166
case distinctions, 19 difference
case splits, 31 – of sets, 108
case_tac (method), 19, 103, 151 disjCI (theorem), 74
cases (method), 156 disjE (theorem), 70
clarify (method), 91, 93 div (symbol), 22, 23
clarsimp (method), 92, 93 div_mult_mod_eq (theorem), 165
class, 163 divides relation, 85, 96, 103–105, 166
classical (theorem), 73 division
coinduction, 118 – by negative numbers, 167
Collect (constant), 111 – by zero, 166
comment, 11 – for type nat, 165
compiling expressions example, 37–39 division_by_zero (type class), 169
Compl_iff (theorem), 108 documents, 57
complement domain
– of a set, 107 – of a relation, 114
complex (type), 167–168 Domain_iff (theorem), 114
complex numbers, 167–168 done (command), 14
Complex_Main (theory), 168 drule_tac (method), 76, 98
composition dvd (symbol), 23
– of functions, 112 dvd_add (theorem), 166
– of relations, 114 dvd_antisym (theorem), 166
conclusion dvd_def (theorem), 166
– of subgoal, 13
conditional expressions, see if elim! (attribute), 131
expressions elimination rules, 69–71
conditional simplification rules, 31 end (command), 15
cong (attribute), 176 Eps (constant), 111
congruence rules, 175 equality, 5
conjE (theorem), 72 – of functions, 111
conjI (theorem), 68 – of records, 155
Cons (constant), 9 – of sets, 108
contrapositives, 74 equalityE (theorem), 108
converse equalityI (theorem), 108
– of a relation, 114 erule (method), 70
converse_iff (theorem), 114 erule_tac (method), 76
CTL, 121–126, 182–185 Euclid’s algorithm, 103–105
208 Index

evaluation, 11 id_def (theorem), 112

even numbers identifiers, 6
– defining inductively, 127–131 – qualified, 4
EX, 201 identity function, 112
Ex (constant), 111 identity relation, 114
exE (theorem), 83 if expressions, 5, 6
exI (theorem), 83 – simplification of, 32
ext (theorem), 111 – splitting of, 31, 49
extend (constant), 156 if-and-only-if, 6
extensionality if_split (theorem), 31
– for functions, 111, 112 if_split_asm (theorem), 32
– for records, 155 iff (attribute), 91, 104, 130
– for sets, 108 iffD1 (theorem), 95
EX!, 201 iffD2 (theorem), 95
ignored material, 63
False (constant), 5 image
fast (method), 92, 93, 124 – under a function, 113
Fibonacci function, 48 – under a relation, 114
field (type class), 169 image_def (theorem), 113
fields (constant), 156 Image_iff (theorem), 114
finding theorems, 34, 93 impI (theorem), 72
finite (symbol), 111 implication, 72–73
fixed points, 117–118 imports (command), 4
force (method), 92, 93 ind_cases (method), 131
formal comments, 60 induct_tac (method), 12, 19, 51, 181
formal proof documents, 57 induction, 178–185
formulae, 5–6 – complete, 180
forward proof, 94–100 – deriving new schemas, 182
frule (method), 84 – on a term, 179
frule_tac (method), 76 – recursion, 50–51
fst (constant), 23 – structural, 19
fun (command), 47–51 – well-founded, 117
function (command) induction heuristics, 35–37
– and numeric literals, 165 inductive (command), 132
function types, 5 inductive definition
functions, 111–113 – simultaneous, 142
– total, 11, 47–51 inductive definitions, 127–146
inductive predicates, 132–133
gcd (constant), 94–96, 103–105 inductive cases (command), 131, 140
generalizing for induction, 129 inductive set (command), 127
generalizing induction formulae, 36 infinitely branching trees, 43
Girard, Jean-Yves, 71n infix annotations, 53
Gordon, Mike, 3 infixr (annotation), 10
grammars inj_on_def (theorem), 112
– defining inductively, 141–146 injections, 112
ground terms example, 136–141 insert (constant), 109
insert (method), 98–100
hd (constant), 17, 38 instance, 158
Hilbert’s ε-operator, 87 INT, 201
HOLCF, 44 Int, 201
Hopcroft, J. E., 145 int (type), 166–167
hypreal (type), 168 INT_iff (theorem), 110
IntD1 (theorem), 107
Id_def (theorem), 114
 Index 209

IntD2 (theorem), 107 meta-logic, 80

integers, 166–167 methods, 16
INTER (constant), 111 min (constant), 22, 23
Inter, 201 mixfix annotations, 53
Inter_iff (theorem), 110 mod (symbol), 22, 23
intersection, 107 mod_mult_distrib (theorem), 165
– indexed, 110 model checking example, 118–126
IntI (theorem), 107 modus ponens, 67, 72
intro (method), 74 mono_def (theorem), 118
intro! (attribute), 128 monotone functions, 118, 140
introduction rules, 68–69 – and inductive definitions, 137–138
inv (constant), 87 more (constant), 152, 153
inv_image_def (theorem), 117 mp (theorem), 72
inverse multiset ordering, 117
– of a function, 112
– of a relation, 114 nat (type), 5, 22, 165–166
inverse (constant), 169 nat_less_induct (theorem), 180
inverse image natural deduction, 67–68
– of a function, 113 natural numbers, 22, 165–166
– of a relation, 116 Needham-Schroeder protocol, 187–189
itrev (constant), 35 negation, 73–75
Nil (constant), 9
λ expressions, 5 no_asm (modifier), 29
LCF, 44 no_asm_simp (modifier), 29
LEAST (symbol), 22, 86 no_asm_use (modifier), 30
least number operator, see LEAST no_vars (attribute), 62
Leibniz, Gottfried Wilhelm, 53 non-standard reals, 168
lemma (command), 13 None (constant), 24
lemmas (command), 94, 104 notation (command), 55
length (symbol), 17 notE (theorem), 73
length_induct, 181 notI (theorem), 73
less_than (constant), 116 numbers, 163–170
less_than_iff (theorem), 116 numeric literals, 164–165
let expressions, 5, 6, 30 – for type nat, 165
Let_def (theorem), 30 – for type real, 168
lex_prod_def (theorem), 117
lexicographic product, 117 O (symbol), 114
lfp o, 201
– applications of, see CTL o_def (theorem), 112
linear arithmetic, 22–23, 164 OF (attribute), 96–97
List (theory), 17 of (attribute), 94, 97
list (type), 5, 9, 17 only (modifier), 29
list.split (theorem), 31 oops (command), 13
Lowe, Gavin, 188–189 option (type), 24
ordered rewriting, 176
Main (theory), 4 ordered_field (type class), 169
major premise, 75 ordered_ring (type class), 168
make (constant), 156 ordered_semiring (type class), 168
marginal comments, 60 overloading, 22, 158–159
markup commands, 59 – and arithmetic, 164
max (constant), 22, 23
measure functions, 116 pairs and tuples, 23, 149–152
measure_def (theorem), 117 parent theories, 4
pattern matching
210 Index

– and fun, 48 rewrite rules, 27

patterns – permutative, 176
– higher-order, 177 rewriting, 27
PDL, 118–121 ring (type class), 168
Power (theory), 170 ring_no_zero_divisors (type class),
pr (command), 101 169
prefer (command), 16, 102 rtrancl_refl (theorem), 114
prefix annotation, 55 rtrancl_trans (theorem), 114
primitive recursion, see recursion, rule induction, 128–130
primitive rule inversion, 130–131, 140–141
primrec (command), 11, 39–44 rule_format (attribute), 179
print mode, 55 rule_tac (method), 76
product type, see pairs and tuples – and renaming, 83
Proof General, 7
proof state, 12 safe (method), 92, 93
proofs safe rules, 90
– abandoning, 13 searching theorems, 34, 93
– examples of failing, 88–89 selector
protocols – record, 153
– security, 187–197 semiring (type class), 168
session, 58
quantifiers, 6 set (type), 5, 107
– and inductive definitions, 136–137 set comprehensions, 109–110
– existential, 82–83 set_ext (theorem), 108
– for sets, 110 sets, 107–111
– instantiating, 84 – finite, 111
– universal, 80–82 – notation for finite, 109
Show Types (Proof General), 7
r_into_rtrancl (theorem), 114 show_main_goal (flag), 12
r_into_trancl (theorem), 115 simp (attribute), 12, 28
range simp (method), 28
– of a function, 113 simp del (attribute), 28
– of a relation, 114 simp_all (method), 29, 39
range (symbol), 113 simplification, 27–35, 175–178
Range_iff (theorem), 114 – of let-expressions, 30
rat (type), 167–168 – with definitions, 30
rational numbers, 167–168 – with/of assumptions, 29
real (type), 167–168 simplification rule, 177–178
real numbers, 167–168 simplification rules, 28
record (command), 152 – adding and deleting, 29
records, 152–158 simplified (attribute), 95, 97
– extensible, 153–155 size (constant), 17
recursion snd (constant), 23
– primitive, 18 SOME (symbol), 87
recursion induction, 50–51 SOME, 201
reflexive and transitive closure, 114–116 Some (constant), 24
reflexive transitive closure some_equality (theorem), 87
– defining inductively, 133–135 someI (theorem), 87
relations, 113–116 someI2 (theorem), 87
– well-founded, 116–117 someI_ex (theorem), 88
relcomp_unfold (theorem), 114 source comments, 60
rename_tac (method), 83 spec (theorem), 81
rev (constant), 10–15, 35 split (attribute), 32
 Index 211

split (constant), 150 type inference, 5

split (method), 31, 150 type synonyms, 24
split (modifier), 32 type variables, 5
split rule, 32 type synonym (command), 24
ssubst (theorem), 78 typedecl (command), 171
structural induction, see induction, typedef (command), 171–174
structural types, 4–5
subclasses, 158 – declaring, 171
subgoal numbering, 47 – defining, 171–174
subgoal_tac (method), 99, 100
subgoals, 12 Ullman, J. D., 145
subset relation, 108 UN, 201
subsetD (theorem), 108 Un, 201
subsetI (theorem), 108 UN_E (theorem), 110
subst (method), 78, 79 UN_I (theorem), 110
substitution, 77–80 UN_iff (theorem), 110
Suc (constant), 22 Un_subset_iff (theorem), 108
surj_def (theorem), 112 unfold (method), 30
surjections, 112 unification, 76–79
sym (theorem), 95 UNION (constant), 111
symbols, 54 Union, 201
syntax, 6, 11 union
– indexed, 110
tacticals, 100–101 Union_iff (theorem), 110
tactics, 12 unit (type), 24
tagged command regions, 63 unknowns, 7, 68
term (command), 16 unsafe rules, 91
term rewriting, 27 update
termination, see functions, total – record, 153
terms, 5 updating a function, 111
text, 61
text blocks, 60 variables, 7
THE (symbol), 86 – schematic, 7
the_equality (theorem), 86 – type, 5
THEN (attribute), 95, 97, 104 vimage_def (theorem), 113
theorem (command), 12, 13
theories, 4 wf_induct (theorem), 117
theory (command), 4 wf_inv_image (theorem), 117
theory files, 4 wf_less_than (theorem), 116
thm (command), 16 wf_lex_prod (theorem), 117
tl (constant), 17 wf_measure (theorem), 117
ToyList example, 9–15
tracing the simplifier, 32
trancl_trans (theorem), 115
tries, 44–47
True (constant), 5
truncate (constant), 156
tuples, see pairs and tuples
txt, 61
typ (command), 16
type classes, 163
type constraints, 6
type constructors, 5