Logic PDF

Formal systems, logic and semantics
Daniel Richardson,
Department of Computer Science, University of Bath.
email : [email protected]
September 27, 2006

Contents
1 Formal Systems 5
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Problem Solving . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Design principles of formal systems . . . . . . . . . . . . . . 8
1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Related ideas:Confluence and termination . . . . . . . . . . 9
2 String Rewriting Systems 11

2.1 Rules and derivations . . . . . . . . . . . . . . . . . . . . . 12
2.2 Language generation: Grammars . . . . . . . . . . . . . . . 13
2.3 Problems (which may be discussed in tutorials) . . . . . . 14
2.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Other uses of string rewriting systems . . . . . . . . . . . . 16
2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3 Term languages, substitution and unification 17

3.1 Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Term languages . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3 Interpretation of a term language . . . . . . . . . . . . . . . 19
3.4 Parse Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5 Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6 Valuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.7 Properties of term languages . . . . . . . . . . . . . . . . . 25
3.8 Unification . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.9 The Unification Algorithm . . . . . . . . . . . . . . . . . . 27
3.10 Problems, which may be discussed in problems class . . . . 30
3.11 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.12 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1
2 CONTENTS
4 Predicate Logic 33
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Truth values and predicates . . . . . . . . . . . . . . . . . . 33
4.3 Variables and types . . . . . . . . . . . . . . . . . . . . . . . 34
4.4 Translation from informal to formal language . . . . . . . . 35
4.5 First order languages . . . . . . . . . . . . . . . . . . . . . . 35
4.6 Syntax, substitution . . . . . . . . . . . . . . . . . . . . . . 37
4.7 Free and bound variables . . . . . . . . . . . . . . . . . . . 37
4.8 Semantics for first order languages . . . . . . . . . . . . . . 38
4.9 Examples : LN and LR . . . . . . . . . . . . . . . . . . . . 41
4.10 Uses of these ideas in computing . . . . . . . . . . . . . . . 42
4.11 More Substitution . . . . . . . . . . . . . . . . . . . . . . . 43
4.12 Other kinds of language and other logics . . . . . . . . . . . 44
4.13 Logic Games . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.14 Normal Forms . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.14.1 CNF and DNF . . . . . . . . . . . . . . . . . . . . . 46
4.14.2 Prenex Normal form . . . . . . . . . . . . . . . . . . 48
4.14.3 Skolem form . . . . . . . . . . . . . . . . . . . . . . 49
4.14.4 Clausal form . . . . . . . . . . . . . . . . . . . . . . 51
4.15 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.16 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.17 Revision Problems. Clausal form . . . . . . . . . . . . . . . 55
4.18 Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.19 Examples and Solutions . . . . . . . . . . . . . . . . . . . . 56
5 Semantic Tableaux 59
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2 Semantic tableau rules . . . . . . . . . . . . . . . . . . . . . 63
5.3 Some advice about the use of the rules . . . . . . . . . . . . 66
5.4 How much use is this system? . . . . . . . . . . . . . . . . . 67
5.4.1 Gödel completeness theorem . . . . . . . . . . . . . 67
5.4.2 Unsolvability and intractability . . . . . . . . . . . . 70
5.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.6 Satisfaction Games . . . . . . . . . . . . . . . . . . . . . . . 72
5.7 Revision Problems . . . . . . . . . . . . . . . . . . . . . . . 72
5.8 Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6 Other formal deductive systems for first order logic 75

6.1 Deduction by Resolution and Unification . . . . . . . . . . . 75
6.2 The Sequent Calculus . . . . . . . . . . . . . . . . . . . . . 76
6.3 Other deductive systems . . . . . . . . . . . . . . . . . . . . 78
CONTENTS 3
7 A Formalisation of Mathematics: ZF set theory 79
8 Gödel Incompleteness theorems 85
9 Logic programming ( prolog) 87

9.1 Headed Horn clauses . . . . . . . . . . . . . . . . . . . . . . 87
9.2 Predicates, facts, constants, variables in prolog . . . . . . . 90
9.3 Programs in prolog . . . . . . . . . . . . . . . . . . . . . . . 93
9.3.1 Comments in Prolog . . . . . . . . . . . . . . . . . . 96
9.4 How prolog works . . . . . . . . . . . . . . . . . . . . . . . . 96
9.5 Programs with just facts . . . . . . . . . . . . . . . . . . . . 97
9.6 Backtracking . . . . . . . . . . . . . . . . . . . . . . . . . . 99
9.7 Warning: problems with prolog . . . . . . . . . . . . . . . . 100
9.8 How to write simple prolog . . . . . . . . . . . . . . . . . . 101
9.9 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
9.10 Circular definitions and Recursion . . . . . . . . . . . . . . 105
9.11 The Cut, and negation . . . . . . . . . . . . . . . . . . . . . 106
9.11.1 The Cut and fail combination . . . . . . . . . . . . . 109
9.12 Family Tree . . . . . . . . . . . . . . . . . . . . . . . . . . 111
9.13 Lists, Member(X,Y), Append(X,Y,Z) . . . . . . . . . . . . . 111
9.14 Sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
9.15 Equivalence relations; and how to find your way out of a
labyrinth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
9.15.1 First Attempt . . . . . . . . . . . . . . . . . . . . . . 115
9.15.2 Second attempt . . . . . . . . . . . . . . . . . . . . . 116
9.15.3 Third attempt . . . . . . . . . . . . . . . . . . . . . 117
9.15.4 Fourth attempt . . . . . . . . . . . . . . . . . . . . . 117
9.15.5 Fifth attempt . . . . . . . . . . . . . . . . . . . . . . 118
9.15.6 Sixth Attempt . . . . . . . . . . . . . . . . . . . . . 119
9.15.7 Testing and Proof . . . . . . . . . . . . . . . . . . . 119
9.16 Labyrinth Program . . . . . . . . . . . . . . . . . . . . . . . 119
9.17 Language Recognisers in Prolog . . . . . . . . . . . . . . . . 120
9.18 kitchen table propositional theorem prover . . . . . . . . . 121
9.19 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
9.20 Input and Output in Prolog . . . . . . . . . . . . . . . . . . 124
9.20.1 write(X) . . . . . . . . . . . . . . . . . . . . . . . . . 125
9.20.2 read(X) . . . . . . . . . . . . . . . . . . . . . . . . . 125
9.20.3 Debugging and tracing . . . . . . . . . . . . . . . . . 125
9.20.4 Assertions and Retractions . . . . . . . . . . . . . . 126
9.21 Arithmetic in Prolog . . . . . . . . . . . . . . . . . . . . . . 126
9.22 Typed Prolog . . . . . . . . . . . . . . . . . . . . . . . . . . 128
9.23 Summary of this chapter . . . . . . . . . . . . . . . . . . . . 129
4 CONTENTS
10 Multisorted, higher order languages, and non classical log-

ics 131
10.1 Notation for types . . . . . . . . . . . . . . . . . . . . . . . 132
10.2 Problems with types . . . . . . . . . . . . . . . . . . . . . . 132
10.2.1 Solutions to Types . . . . . . . . . . . . . . . . . . . 133
10.3 Non classical logics . . . . . . . . . . . . . . . . . . . . . . . 133
10.4 Proof Assistants . . . . . . . . . . . . . . . . . . . . . . . . 134
11 Semantics and Specification for Programs 135

11.1 Programming Language Semantics . . . . . . . . . . . . . . 135
11.2 Denotational Semantics . . . . . . . . . . . . . . . . . . . . 136
11.3 Correctness and Completeness of software . . . . . . . . . . 138
11.4 Formal Methods . . . . . . . . . . . . . . . . . . . . . . . . 138
12 References 139
Chapter 1
Formal Systems
1.1 Introduction
There is a long tradition, going back at least to the axiomatic geometry
of the ancient Greeks, of trying to model human thought processes by for-
mal systems. Computers, allowing global communications and requiring
development of standard languages, encourage standard formal representa-
tions of human thinking, reckoning, scheming, calculating and cogitation.
A great deal of work is currently being done in this area, which includes and
exceeds the bounds of the subject usually called artificial intelligence. This
book attempts to describe one of the most successful of the formalisms:
formal deductive systems.
1.2 Problem Solving

When we are absorbed in the business of trying to solve a problem, we are
usually too occupied to observe patterns in what we are doing, or to make
generalisations about how the process works. However, since we wish to
cooperate with others in problem solving and also to use computers to help
us, and to represent our insights computationally, some generalisation is
called for.
Consider the following list of problem solving situations.
1. A person trying to solve the Rubic cube puzzle.
2. Two people playing chess.
3. A mouse, apparently trying to find its way out of a maze.
5
6 CHAPTER 1. FORMAL SYSTEMS
4. A child learning to speak.
5. A mathematician trying to prove a theorem, which he or she is intu-

itively sure is correct.
6. A computer programmer developing a program.
Perhaps you will agree that there are some common features in this
list. It seems that in some ways we can take the mouse in the maze as
typical of problem solving agents. Problem solving seems to be a process of
exploration within some definite space of possibilities. Within this space of
possibilities, some moves are allowed and some are prohibited. What kind
of general representation can we make for this situation?
We will call a position in the space of possibilities a configuration. We
hope to be able to write down the configuration at each moment. The
configuration at a moment in a game of chess, for example, might be the
arrangement of pieces on the board, together with an indication of who
should move next. We think of each configuration as a point, or a posi-
tion in the space of possibilities, and the moves are transitions from one
configuration to another.
Assume that we represent each configuration by a data structure in some
set D of data structures, appropriate to the situation we are considering.
(Here, a data structure could be a number or a string of characters, or a
vector or a matrix, or a tree labelled with data structures, or a directed
graph labelled with data structures, for example.) To say that D is a set
of data structures does not commit us to very much. But we do want to
insist that the items in D are unambiguous and finite and can somehow
be written down. The items in D are syntactic; they are pieces of syntax,
arrangements of symbols.
We must also say how the moves within D are constrained. What
transitions are allowed, and what forbidden? We will suppose that we have
some collection of rules, which we will denote R, which determine which
transitions in D are allowed. If X and Y are in D, we will write
X ⇒R Y
to mean that the rules R allow a transition from X to Y in one step.
Definition 1.1 A formal system is (D, R) where D is a set of data struc-

tures, and R is a set of rules which determine which transitions between
objects in D are allowed.
Definition 1.2 If (D, R) is a formal system, let X ⇒R Y mean that X

and Y are in D and that the transition from X to Y is allowed by the rules
R.
1.2. PROBLEM SOLVING 7
Definition 1.3 If (D, R) is a formal system, let

X ⇒∗R Y
mean that there exists a finite sequence X1 , ..., Xn of items in D so that
X = X1 ⇒R X2 ⇒R X3 ⇒R ... ⇒R Xn = Y .
X ⇒∗R Y means that it is possible to get from X to Y by a finite

sequence of steps which are allowed by the rules R. A sequence of this
kind,
X1 ⇒R X2 ⇒ ... ⇒R Xn
is called a derivation, of length n − 1, in the formal system (D, R).
So X ⇒∗R Y means that there exists a derivation from X to Y . Notice
that derivations of length 0 are allowed. So X ⇒∗R X is always true.
If the set of rules can be understood from the context, we will leave off
the subscript R.
We require that X ⇒R Y be decidable. That is, given X and Y in
D, we should be able to determine whether or not X ⇒R Y . (Otherwise,
what good are the rules?) Moreover, we expect that it should be easy to
decide X ⇒R Y . The decision as to whether or not X ⇒R Y must only
depend on X and Y as syntactic objects. That is, R is not allowed to refer
to any meaning associated with the data structures in D. For example, we
would not allow a rule which said that X ⇒ Y if and only if “ a move from
X to Y is a good idea”. The reason for this is that R is required to be
unambiguous. In this sense, R is formal.
Although we insist that X ⇒R Y should be easy, we expect that X ⇒∗R
Y may be relatively hard. Our claim is that the core of problem solving
has to do with this relation X ⇒∗R Y .
People have probably always invented and puzzled over formal systems,
as described above. Part of their charm is the contrast between the typical
simplicity of X ⇒ Y and the typical complexity of X ⇒∗ Y . Even when D
is finite, X ⇒∗ Y is often extremely interesting, as in chess, for example.
There are two reasons for the development of formal systems (D, R).
One reason is their inherent fascination and beauty; and the other is that
a formal system may have quite simple rules, and yet may capture the
combinatorial features of an infinite and surprising reality. We are trying
to do something which is so ambitious that it is almost preposterous: the
hope is to capture some aspects of our own thought processes, by modelling
them with formal systems. We are trying to make language mechanisms
which behave like thought. The design and study of such formal systems is
the primary motivation of the field of study known as artificial intelligence.
We are trying to catch some aspects of the action of intelligence within
formal systems.
Remarks. If you like geometric pictures, you may visualise a formal

system as a collection of points, some of which are connected by arrows.
The points are called configurations and each one is identified by a finite
piece of syntax. In this sense, a formal system may be considered to be a
certain kind of labelled, possibly infinite, graph. If you know what a Turing
machine is, you will see that a Turing machine is a special kind of formal
system. If you know what a (mathematical) category is, you will see that
if (D, R) is a formal system, then D together with ⇒∗ , is a certain kind of
category.
Examples: Chess is a formal system, with configuration being chess po-
sitions. In general two person games of strategy can usually be modelled
as formal systems. Interaction between a computer system and an environ-
ment can also be modelled in this way. We may consider logical argument
or disputation as a formal system, a sort of game in which only certain
moves are allowed. A non deterministic finite automaton is a formal sys-
tem; so is a pushdown automaton, or a deterministic or non deterministic
Turing machine; the configurations specify the complete situation of the
system at a moment of time. The lambda calculus is a formal system; the
data structures are the lambda terms, the transitions are alpha and beta
reductions. In linear algebra, we may take matrices as configurations, and
row operations as allowed transitions; then A ⇒∗ B means that matrix A
can be changed into matrix B by a sequence of row operations.
We may also model populations of agents by formal systems. A config-
uration gives the internal state of each agent, and any other appropriate
information such as, for example, position. We might, for example if we
were trying to model group behaviour of animals or humans, allow each
agent to observe its immediate surroundings, and to perform certain actions
in relation to other agents, for example offering to shake hands, bowing,
following, avoiding, bluffing, mating, killing, eating, etc. X ⇒ Y means
that the rules allow configuration X to change to configuration Y in one
step. Of course, we may also consider independently acting processes in
computer systems to be agents, whose behaviour is constrained by rules.
1.3 Design principles of formal systems

A specification for a formal system is a statement, hopefully a precise state-
ment, of what we want the formal system to do.
There are accepted criteria for judging how well a formal system (D, R)
meets its specification.
1. Completeness. (D, R) does everything which the specification re-
quires.
1.4. SUMMARY 9
2. Correctness. (D, R) does nothing which the specification prohibits.
3. Simplicity. This is also called the Occam’s razor condition. One

statement of this very old condition of good intellectual design is:
Thou shalt not multiply entities unnecessarily . The formal system
should fulfil its specification, or come close to fulfilling its specification
with minimal complexity.
4. Naturalness. The formal system should be in accord with intuition.
All of these conditions are important. In many cases, in most interesting

cases, they may also, in part, contradict one another. Design of a good
formal system usually involves a trade off between these conditions.
It will not have escaped the thoughtful reader that the relationship
between a formal system and its specification is like the relationship between
a scientific theory and the phenomena which it is designed to explain. One
difference is that science has traditionally believed that its theories are
physically true. Although, when a formal system is extremely good, as
with the formal grammars explained in the next chapter, it is hard not to
suspect that there could be some physical basis for it, most people think
that formal systems are inventions rather than discoveries.
1.4 Summary
A formal system is a set of data structures together with a set of transfor-
mations which act on them.
Important examples are formal grammars for generating formal lan-
guages, formal proof systems in predicate logic, the lambda calculus which
formalises symbolic computation, and populations of autonomous agents.
Also all games.
1.5 Related ideas:Confluence and termination

We will say that a formal system (D, R) is terminating if there is no infinite
derivation in it. This means that no matter what choices are made, every
derivation must eventually stop.
A formal system (D, R) is confluent if whenever A ⇒∗ B and A ⇒∗ C
there exists D so that B ⇒∗ D and C ⇒∗ D. This means that if we started
at configuration A and computed to find C, and computed in another way
to find B, there is some way to bring the two computations back together
to get D.
A formal system (D, R) is locally confluent if whenever A ⇒ B and

A ⇒ C there exists D so that B ⇒∗ D and C ⇒∗ D.
For formal systems used for computation (such as the lambda calculus)
confluence is a very useful property.
Theorem 1.1 A formal system is confluent if it is locally confluent and

terminating.
Chapter 2
String Rewriting Systems
Suppose we start with a language, such as English or Urdu or C or LZF .

Each utterance in the language is a finite string of symbols. Some strings
are in the language and some are not. We may think of the language as a set
of strings of symbols. It seems that if people speak or recognise a language
this means that they somehow know the definition of this possibly infinite
set of strings of symbols. How can these complicated sets be defined? The
best contemporary answer to this puzzle is that although languages may
be infinite they can be generated by finite sets of rules.
String rewriting systems are formal systems whose data structures are
strings of symbols, and whose transformations are rewriting rules, as defined
below.
Let Σ be a finite alphabet of symbols. Let Σ∗ be the set of all finite
length strings which can be formed using symbols from Σ. The elements of
Σ∗ will be called words on Σ.
A Σ rewriting rule is an expression of the form
α;β
where α and β are in Σ . Note that Σ∗ contains the empty word.
∗
A rule α ; β means that it is allowed to replace α by β in any context.

Note that here and in the following we are using Greek letters as linguistic
variables for strings of words in Σ∗ .
Definition 2.1 A Σ rewriting system is a finite list of Σ rewriting rules:

α1 ; β1
.
.
.
11
12 CHAPTER 2. STRING REWRITING SYSTEMS
αn ; βn
A Σ rewriting system is meant to be a set of rules which defines a
computational process on strings of symbols from Σ. It turns out that any
computational process can be seen in this way.
2.1 Rules and derivations

Definition 2.2 Let P =
α1 ; β1
.
.
.
αn ; βn
be a Σ rewriting system, and let A and B be words in Σ∗ .
We will say that B can be obtained from A in P in one step, and write
this as
A ⇒P B
if A is of the form W1 αi W2 , and B is of the form W1 βi W2 , where
αi ; βi is in P . (Note that W1 and/ or W2 can be empty.)
When the rewriting system P is understood, we will leave off the sub-
script P .
As with all formal systems, once we have got a clear notion of one step
derivation, we can iterate this to obtain derivations of any finite length.
Definition 2.3 A derivation in rewriting system P is a list of words A1 , A2 , ..., Ak
so that A1 ⇒ A2 ⇒ ... ⇒ Ak .
Definition 2.4 We will say that there is a derivation which goes from A
to B, and write this as A ⇒∗ B if there exists A1 , ..., An so that
A = A1 ⇒ A2 ⇒ ... ⇒ An ⇒ B
We allow derivations of length zero. So we will always have A ⇒∗ A.
Example 2.1 Suppose Σ = {a, b}, and P is
aba ; b
aaa ;
Let W be the word aaaababbbbabababaaaababa.
In order to apply P to W , we attempt to match left hand sides of the
rules of P with parts of W . There are 6 matches with aba and four matches
with aaa. So there are ten words B with W ⇒ B.
2.2. LANGUAGE GENERATION: GRAMMARS 13
Problem 2.1 Start with W , as given in the example above. Apply the
given rules in any order until you terminate. Why did you eventually ter-
minate? Do you get the same result at termination, no matter in which
order you apply the rules?
2.2 Language generation: Grammars

Definition 2.5 We will say that a word W is terminal in rewriting system
P if there is no word Z so that W ⇒ Z.
Definition 2.6 A grammar is a rewriting system P together with an initial

word I. Such a grammar (P, I) generates the language
{W : I ⇒∗ W, W terminal}.
Example 2.2 Consider the language, L = {ab, aabb, aaabbb, ...}. The lan-
guage consists of the set of words which are formed by a string of a’s followed
by an equally long string of b’s. There are many different grammars which
could be used to generate this language. One would be:
X ; ab
X ; aXb.
The starting word is X. A typical derivation would be:
X ⇒ aXb ⇒ aaXbb ⇒ aaaXbbb ⇒ aaaabbbb
Note that the string on the end of the derivation is terminal and is in
the language L.
Try to prove that the language generated by this grammar is exactly L.
Note that there would be two parts to such a proof. We would want to show
that every terminal string generated by the grammar is in L; and we would
also want to show that every string in L could be generated by this grammar.
So even at this simple level there is a completeness and correctness problem.
2.
2.3 Problems (which may be discussed in tu-

torials)
Problem 2.2 Define a variable to be any upper case letter followed by any
string of letters and digits. Write a grammar which generates this language.
Problem 2.3 Let Σ = {X,0 , 0}.

Describe the Σ language L1 generated by grammar (P, X) where P is:
X ; X0 | 0
(It is cheating in this case to describe the language as the language
generated by the grammar. )
Problem 2.4 Describe the language L2 generated by the grammar (P, X)

where P is
X ; X 0 | (X + X) | 0
Show that ((000 + 0000 ) + 00 )000 is in L2 , by giving a derivation of it from
the initial word X. You may, to shorten the amount you have to write,
use X ≡> Y to mean that Y can be obtained from X by doing several non
overlapping single steps in parallel.
Problem 2.5 Consider languages L such that 0 is in L and whenever a

string α is in L the string α0 is also in L. Give two examples of such
languages. How is L1 special among such languages.
Problem 2.6 Give a grammar for the language which consists of strings
of b’s separated by single a’s.
Problem 2.7 Here are some samples from a language.
1. (((p ∨ q) ∧ ¬r) ∨ (p ∧ q))

2. ¬¬(p ∧ q)
Give a grammar to generate this language. Also give a description of the

language (i.e. a specification), which does not use the idea of a grammar.
Problem 2.8 Here are some samples from a language.
1. (xy)
2. (x(yz))
3. ((xy)z)
4. (λx.(xz))
2.3. PROBLEMS (WHICH MAY BE DISCUSSED IN TUTORIALS) 15
5. ((λx.x)(λy.y))
6. ((x(λy.(yy)))z)
Give a grammar, as simple as possible, which generates all of these. Al-

though the answer to this problem is simple once you see it, it is a harder
problem than the earlier ones. (Of course, if I have made any typing mis-
takes in the sample, this problem is more or less impossible.)
Problem 2.9 Let P1 be a programming language, consisting of statements.

A sequence of statements is also a statement. Statements are closed under
the usual IF and WHILE constructions. Assignments are statements. Vari-
ables are all of type integer. Expressions include the variables and the usual
notation for the integers, and are closed under +, -, *. Give a grammar
(as simple as possible consistent with this description) for the expressions
and also for the statements of P1.
Example 2.3 LZF , the language of Zermelo-Fraenkel set theory, is gen-

erated by the grammar (P, S), where P is as follows.
S ; (S ∨ S) | (S ∧ S) | (¬S) | (S → S) | (S ↔ S) | (∀V ariable)S |
(∃V ariable)S | V ariable = V ariable | V ariable ∈ V ariable
where V ariable is as defined above. Almost all of contemporary mathe-
matics can be written in this famous formal language LZF . This is a very
successful example of formalisation. Note that the fact that mathematics
can be expressed in LZF does not imply that we should use LZF as a work-
ing language of mathematical thought. On the other hand, the existence of
a formalisation is extremely useful, as a last resort, in case of disagreement
about meaning.
As an example of translation from mathematical English into LZF , con-
sider the statement: For any sets X and Y there exists a set Z so that the
elements of Z are either elements of X or elements of Y or both.
In LZF , this would be:
(∀X)(∀Y )(∃Z)(∀W )(W ∈ Z ↔ (W ∈ X ∨ W ∈ Y ))
A formal language is one which can be defined by a grammar, as de-

scribed above.
2.4 Applications
Programming languages are formal languages. For example, the syntacticly
correct programs of C constitute a formal language. It is quite important to
have a compact and unambiguous description of a programming language,
and formal grammars give us exactly that.
Formal languages are also used as a language for specification of pro-
grams. This raises the possibility that part of the process of development,
from specification to working program, could be formalised and automated.
Formal languages are also used in the foundations of mathematics. Part
of the process of formalisation of a field of mathematics is the development
of an appropriate formal language.
Some people think also that natural human languages have some com-
mon formal core. The idea is that there is something like a formal language
which is common to humanity, and that natural languages are obtained
from this by a process of transformation.
2.5 Other uses of string rewriting systems

In this chapter we have concentrated on grammars and languages. However,
string rewriting can be used also as a basis for computation on strings of
symbols. This will be explored later.
2.6 Summary
In string rewriting systems, the data structures are strings of symbols, and
the transformations are string rewriting rules.
Given a string rewriting system P , we will say A ⇒P B if string B can
be obtained from string A by one application of the rules in P . We will say
A ⇒∗ B if strong B can be obtained from string A by a sequence (possibly
of length zero) of applications of the rules.
From an initial word I, a language is generated within a string rewriting
system P . The language is
{W : I ⇒P W, W/terminal}.
The string rewriting system together with the initial word is called a
grammar for the language.
Chapter 3
Term languages,
substitution and
unification
3.1 Function Spaces

A type is a set together with a family of operations defined on the set. An
example is the integers, with the usual operations of addition, subtraction
and multiplication.
If A and B are types, A → B will mean the type of all functions from A
to B. So f : N → N means that f is a function from the natural numbers
to the natural numbers.
If D is any set, f : Dk → D means that f is a function of k arguments
from D which takes values in D.
Let D be a set. Let FD be ∪∞ n
n=0 (D → D). FD will be called the
function space of D. It contains all the functions which apply to arguments
in D and give values in D. Since D may be infinite, we may not be able to
give names to all the objects in FD . What we want therefore is a convenient
notation for some of these functions.
3.2 Term languages

Term languages occur very often in mathematics and computing. The terms
in a term language are generally used to denote functions in some function
space.
17
18CHAPTER 3. TERM LANGUAGES, SUBSTITUTION AND UNIFICATION
Programming languages typically have term languages as expressions.

A term language is obtained by taking a collection of function symbols
and constants and a set of variables and forming all possible correctly for-
matted expressions from the variables and the function symbols, allowing
arbitrary nesting of expressions.
The collection of function symbols and constants is called the signature
of the term language.
Function symbols are also called function names, or function identifiers.
The function symbols are not functions, they are just names, pieces of
syntax.
It is assumed here that we know, for each function symbol in a signature,
the number of arguments which it should have. This number is called the
arity of the function symbol. It is also assumed that we know, for each
function symbol, the correct format for writing the application of it to its
arguments.
For example + usually has arity 2, and sin(X) has arity 1. A function
name f (X, Y ) would have arity 2.
We will consider constants as function symbols of arity 0. So a signature
for a term language is just a set of function symbols.
We can begin with any signature. We do not determine in advance what
the function symbols mean.
Example 3.1 The function names sin(X) and cos(X), together with the
variables generates a term language. One of the terms in this language is
sin(sin(cos(sin(cos(sin(Y ag)))))). This term is meant to stand for some
function of the variable Y ag. If we give sin and cos their usual interpreta-
tions, this term is determined as a real valued function of a real variable.
But sin and cos could be interpreted in other ways, and the term language is
just the set of terms, as syntactic things in themselves, without any special
assumptions about the interpretation of the function symbols.
Term languages are languages with a fairly simple semantics. So this is

a reasonable place to try out some ideas.
Example 3.2 Suppose f (X, Y ), and s(X) are function symbols. The
term language generated by these function symbols consists of all terms
which can be written using any variables, and these two function symbols
and composition. So, for example f (s(s(W )), f (U, f (B, s(W )))) is a term
in this term language. Note that the terms are syntactic things, strings of
symbols. We have not yet decided what these terms mean, if anything. Note
also that we do have a strict notion of syntactic correctness for these terms,
even without being sure what they stand for. For example f (Z, W ( must be
3.3. INTERPRETATION OF A TERM LANGUAGE 19
a typing mistake; it is not a term. Similarly, s(A, B, C) might be a term in

some other term language, but not in this one.
One way to define a term language is by a grammar. For the above

example, the grammar would be:
< term >; f (< term >, < term >) | g(< term >) |< V ariable >
where, < V ariable > means an upper case letter followed by a string of
letters and digits.
Example 3.3 Suppose we take signature [+, ∗,0 , 0] with arities 2,2,1,0.
An example of a term is (X +(Y ∗X)00 ). Another term is 000000000000000000000000000000000 .
3.3 Interpretation of a term language

Definition 3.1 Let A be a term language, with signature Σ = (f1 , ..., fk ).
An interpretation of A is (D, f1 , ..., fk ), where D is a set and f1 , ..., fk are
in the function space of D and, for i = 1, ..., k, if the arity of fi is ni then
fi : Dni → D.
In other words, an interpretation of a term language gives a specific

meaning in some function space to each function symbol in the signature.
Given an interpretation, each term is interpreted as a function in the
function space FD . We can regard the term language as a notation for
some of the functions in the function space. It would also be correct to say
that the functions give a meaning to the term language. Notice that an
interpretation is specified by giving a meaning to the function symbols in
the signature; the meaning of all the terms is then automatically generated.
D is called the domain of the interpretation. This can be any set. The
functions can be any totally defined functions, which take arguments which
are tuples of the right arity in D to match the signature, and take values
in D.
For each term language, you should imagine a spectrum of all possible
interpretations. D could be a finite set, or the natural numbers, or the
reals.
D could even be A itself. That is, the domain could be the terms!
Within this domain, we can interpret a function symbol fi , of arity ni as a
function fi : Ani → A by letting fi map terms T1 , ..., Tni into the term which
is simply obtained by writing fi applied to T1 , ..., Tni . For example f (X, Y )
applied to Z and g(W ) is literally f (Z, g(W )). In this interpretation a term
is regarded both as a syntactic object and also as a definition of a function
from terms to terms. We will call this the literal interpretation of a term
language.
Example 3.4 Consider signature [f, g, s, a] with arities 2,1,1, 0. One of

the terms would be
f (s(X), f (Y, f (Y, Z))).
Even though we do not yet know what f, g, s are, we know how to read
this notation. This is a compound function which has been put together
from basic parts f, g, s.
Example 3.5 Consider signature [(cons X Y ), (car X), (cdr X), nil] with
arities 2, 1, 1, 0. One of the terms in this language is
(cons (cdr X) (cons Y (cons Y Z))).
Note that no predefined meaning is attached to these function symbols.
In fact this term language is essentially the same as the term language
in the previous example, even though the function names and the format
conventions are dissimilar.
In all the following examples of term languages, variables will begin

with upper case letters. Any name which starts with a lower case letter is
assumed to be a function name.
So jane is a constant, Jane is a variable, and lub(X,Y) is a term, in
which the function lub of arity 2 is applied.
3.4 Parse Trees

Consider the term language with grammar
< T >; f (< T >, < T >) | g(< T >) |< V ariable >,
with < V ariable > defined as usual. A typical term in this language is
f (g(X), f (Y, g(X))).
Problem 3.1 Write down three different derivations of the term above in
the grammar of its term language.
It should be clear that although there are different ways of deriving this
term, all the derivations are essentially the same. We could try to make
this clear by combining together several parallel steps:
< T >≡> f (< T >, < T >) ≡> f (g(< T >, f (< T >, < T >))) ≡>
f (g(X), f (Y, g(< T >))) ≡> f (g(X), f (Y, g(X))).
We may also display the derivation as a tree:
f
/\
/ \
3.4. PARSE TREES 21
g f
/ /\
/ / \
X Y g
\
X
These trees are called parse or expression trees. We can think of them
as growing downwards from the initial string.
<T> => f => f => etc

/\ /\
/ \ / \
<T> <T> g f
/ /\
/ / \
<T> <T> <T>
The parse tree of a term shows how the term is derived.
Definition 3.2 The outermost operator of a term is the function name

which appears at the top of the parse tree of the term.
Continuing with this example, suppose we give a semantics to the term

language by specifying that f (X, Y ) is addition and g(X) is multiplication
by 3 over the integers.
It should be clear that each term represents a function, which takes
values in the integers. That is, if we assign integer values to the variables,
we should get an integer value returned.
We can see how this works on the parse tree:
Suppose we have X = 2, Y = 3.
+
/\
/ \
*3 +
/ /\
/ / \
2 3 *3
\
2
As we evaluate the tree, the numbers percolate upwards, until we get

the final value at the top.
In conclusion, grammars generate parse trees from the root downwards;
semantics evaluates and interprets a parse tree, starting from the frontier
and finishing with the return value at the root.
The outermost operator of a term is the one chosen first in the deriva-
tion, and evaluated last.
3.5 Substitution
Substitution of terms for variables inside other terms is a very natural
operation. If N and M are terms in a term language and x is a variable, we
will use the notation N [x := M ] to mean the result of substituting M for all
occurrences of x in N . We can also do several substitutions simultaneously.
Definition 3.3 If N is a term, and x1 ,..., xn are variables, and M1 ,...,

Mn are terms, then
N [x1 := M1 , ..., xn := Mn ]
means the result of simultaneously replacing all occurrences of x1 ,...,
xn by terms M1 ,..., Mn in term N .
Notice that the insistence on simultaneity is important.
Example 3.6 (X + Y )[X := (X ∗ Y ), Y := X] = ((X ∗ Y ) + X)
Problem 3.2 Suppose A, B, C are terms. Which of the following are cor-
rect?
1. (A[x := B])[y := C] = (A[y := C])[x := B]
2. A[x := B, y := C] = A[y := C, x := B]
3. (A[x := B])[y := C] = A[x := B, y := C]
Give counterexamples where possible.

3.6. VALUATIONS 23
A new formal system has popped up here. The data structures are
terms. The transformations are the substitutions. It seems that everybody
who studies mathematics or computing already knows about substitutions.
Are there any hard mathematical problems in this area?
Problem 3.3 Suppose we start with a term and apply a substitution to

it, and then apply another substitution to the result. Can the final result
be obtained from the initial term by one substitution? A simple yes or no
answer will not suffice here. If the answer is obvious, it should also be easy
to give a proof. Try this.
Definition 3.4 If σ1 and σ2 are substitutions, let σ1 ◦ σ2 be the result of

first applying σ1 and then applying σ2 . This will be called the composition
of σ1 and σ2 .
3.6 Valuations
Suppose A is a term language with signature (f1 , ..., fk ). Let D, f1 , ..., fk )
be an interpretation of A. A valuation from D is a function which gives
values in D to a subset of the variables of A. If X1 , ..., Xn are variables in
A, and a1 , ..., an are objects in D, we will use the notation
[X1 := a1 , ..., Xn := an ]
for the valuation which gives value a1 to variable X1 , ... and value an
to variable Xn .
The support of a valuation is the set of variables which are given values
by it.
Let τ be a term of A and set v be a valuation whose support includes
all the variables which occur in A.
We will use the notation
τ (v)
for the value of τ in D which results from setting the values according to
v, and using the given interpretation of the function symbols. If the support
of v does not include all the variables of τ , then τ (v) will be undefined. In
general, given an interpretation, terms denote partially defined functions of
valuations.
Example 3.7 Let E be the term language with function symbols +, −, ∗,

with arity 2, and sin, cos, exp, abs with arity 1, and 0, 1, π with arity zero.
The natural interpretation of E has the the real numbers as its domain and
the usual function and constant meanings. So exp(X)([X := π]) is the
value of exp(X) when X = π, i.e. eπ . This value is a real number, not an
approximation. Note that in this example there is no possibility of writing
down names for all the objects in the domain. Of course, we can write
down names for some of the objects, for example
exp(1) − exp(3 + exp(2)), or exp(exp(π)).
Unfortunately, in the present state of mathematics, we do not know how
to decide whether or not any two given names of this kind denote the same
real number.
We expect that substitution and evaluation will interact in a nice way.
Problem 3.4 Let τ be a term and α a substitution, and v a valuation
whose support includes all the variables which occur in τ or in α. We can
first substitute, and then evaluate. This gives us
τ α(v)
Show that there exists a valuation w which, applied directly to τ , gives
the same result.
Definition 3.5 Let A be a term language and I = (D, f1 , ..., fk ) an inter-
pretation of A. If τ1 and τ2 are two terms of A, we will write
|=I τ1 = τ2
if
τ1 (v) = τ2 (v).
for all valuations v from I whose support includes all the variables of τ1
and τ2 .
|=I τ1 = τ2 will be read: Under interpretation I, terms τ1 and τ2 are
semantically equal. For example, under the usual interpretation of multi-
plication over the reals, X ∗ Y and Y ∗ X are semantically equal. Of course,
X ∗ Y and Y ∗ X are not syntactically equal. We can find an interpretation
in which X ∗ Y and Y ∗ X have different meanings.
A given interpretation induces a notion of semantic equality (in that
interpretation) on a term language.
Problem 3.5 Show that for any term language, A, there exists an inter-
pretation, I in which the notion of semantic equality in I is the same as
literal syntactic identity in A.
Of course we can decide syntactic equality in a term language. Two
terms are syntactically the same only when they are identical. But some-
times semantic equality is difficult to recognise. For example, in the term
language E mentioned earlier, interpreted in the reals in the natural way,
semantic equality is undecidable. That is, there can be no algorithm to
decide whether or not two terms of E, interpreted over the reals, have the
same meaning.
There are even difficulties, as mentioned earlier, in deciding semantic
equality of constants standing for real numbers.
3.7. PROPERTIES OF TERM LANGUAGES 25
Problem 3.6 Assume sqrt(x) means the square root of x, defined for non
negative reals.
What should be the value of x after the following?
If (sqrt(9 + 4*sqrt(2)) = 1+ 2*sqrt(2))

then x:=1
else x:=0;
Problem 3.7 How could a computer correctly solve the previous problem?
Semantics is supposed to give a standpoint for criticism. The above

problem is intended to convince you that an honest semantics should con-
sider interpretations in the real numbers, and not just in an approximation.
On the other hand, there are severe difficulties in this area.
Don’t worry if you can not see how to solve the problem above. No one
else is sure how to do this either. After all, the test could have been much
more difficult, e.g.
(sqrt((112+70*sqrt(2)) +(46+34*sqrt(2))sqrt(5))=
(5+4*sqrt(2))+(3+sqrt(2))*sqrt(5) ?
How to make a computational representation of the real numbers which
is an improvement on bounded precision floating point numbers is a subject
of current research in computing.
3.7 Properties of term languages

A term says how to compute something. Compound terms are built up
from simple ones by function application. If we compare a term language
with the way in which functions are built up from one another in a typi-
cal programming language, such as C, we see that term languages have a
beautiful simplicity and clarity. In evaluation of a parse tree, all subtrees
are independent, and can be evaluated in parallel. In any expression to be
evaluated, we can replace a term by its value without changing the final
result. There is a horrible piece of jargon for this. It is called referential
transparency. Term languages have referential transparency. As far as I
can understand, the original idea here is that we can, so to speak, look
right through the term to see its value in the domain without any loss of
information; we can suppose that the term and its value are the same. On
the other hand, in a C program, a function may depend on the state of the
computer memory, and this may in turn be altered by the act of evaluating
the function. C functions have side effects. So we do not have referential

transparency for C functions in general.
Because of referential transparency, computation in a term language is
naturally parallelisable. It is also relatively feasible to prove properties of
such computation. The main idea of functional programming is to try to
get some of these beneficial properties for programming languages.
3.8 Unification
Definition 3.6 Suppose N and M are terms and σ is a substitution. We
will say that σ unifies M and N if N σ = M σ.
Example 3.8 (X ∧ (Y ∨ Z)) and ((¬W ) ∧ V ) are unified by substitution

[X := ¬horsef eathers, W := horsef eathers, V := (Y ∨ Z)] .
Note that there are other ways of unifying these two terms. In fact
the substitution given has an annoying excess specificness. Where did the
horsef eathers come from? We can say that although this substitution is a
unifier, it is not the most general unifier, since a constant horsef eathers
has been unnecessarily used.
We are already starting to use the concept of most general unifier, but
we do not yet have a definition of it. We will have to return to this question
later. We do need to clarify it. You may be able to do this yourself at this
point if you think very hard about it. A unifier is a substitution which
unifies two terms. What is a most general unifier?
There is another problem here. Given two terms, how do we decide
whether or not they have a unifier, or a most general unifier? If there is
such, how can we find it? We will also return to this later.
For the moment, here is a useful image. Suppose we have two terms,
and we are looking for a unifier. A good way to begin is to superimpose one
parse tree over the other. If they contradict each other, there is no unifier.
If they do not contradict each other, we should be able to see the minimal
substitutions which will make the two trees the same.
Example 3.9 Consider the following list of terms:
1. f (g(X), f (Y, Z))
2. W
3. f (W, f (g(U ), V ))
4. g(T )
3.9. THE UNIFICATION ALGORITHM 27
5. f (Y, Z)
The first two are unified with the substitution [W := f (g(X), f (Y, Z)).
The first and the third can also be unified. To see this, superimpose the parse
trees. The f matches at the top. So now we need to unify W and g(X)
to get the left hand subtree, and having done that we will need to unify the
right hand subtrees. You can see now why we need a most general unifier.
We wish to unify and also leave ourselves as much room to manoeuvre as
possible. (We still do not have a definition!) But you can see what to do,
namely [W := g(X)]. Fortunately there are no W occurrences in the right
hand subtree. We need now to unify f (Y, Z) and f (g(U ), V ). Once again,
the f at the top matches. Possibly you can finish this part of the example
yourself ?
Note that the attempt at unification fails if you get non matching func-
tion names at the top of the parse tree, as in items 4 and 5 above. There
is another way in which unification can fail. Consider terms 1 and 5. We
get [Y := g(X)]. We are then confronted with the alarming problem of try-
ing to unify Z and f (g(X), Z). Since terms are finite, we observe that no
term can ever be a proper subterm of itself; so there is no way that Z and
f (g(X), Z) can be the same.
Definition 3.7 A most general unifier of two terms T and S is a substi-

tution α which is a unifier, and which also has the property that any other
unifier β can be obtained from α by a further substitution. That is, if α is
a most general unifier and T β = Sβ, then there must exist a substitution γ
so that
(T α)γ = T β
and
(Sα)γ = Sβ
Example 3.10 Suppose we wish to unify f (X, g(U, h(U, s), U )) and
f (g(U, V, s), g(W, Y, t))
We see that X must be the same as g(U, V, s), U ), and g(U, h(U, s), U )
must be the same as g(W, Y, t). Thus W and U must be the same, and Y
must be h(U, s), and U must be t. So we end up with
α = (X := g(t, V, s), Y := h(U, s), W := t, U := t)
3.9 The Unification Algorithm

This will apply to two lists of terms (S1 , ..., Sn ) and (T1 , ..., Tm ). The algo-
rithm will either produce a most general unifier of these two lists, or it will
return FAIL. It should fail if and only if no unifier is possible. If we wish
to unify two individual terms, we will just put these two terms in two lists
of length one and apply the algorithm to this pair of lists.
The algorithm will be described recursively.
Basis. This is a collection of simple cases.
1. (Simple Case 1)
unif y((S1 , ..., Sn ), (T1 , ..., Tm )) = F AIL
if n 6= m,( since, of course, a substitution can not alter the number
of terms in a list).
2. (Simple Case 2)
Another easy case occurs when (S1 , ..., Sn ) and (T1 , ...Tm ) are already
equal. In this case no unifier is necessary. We let
unif y((S1 , ..., Sn ), (S1 , ..., Sn )) = [], the empty substitution.
3. (Simple Case 3)
Another simple case occurs when n = m = 1, and one of S1 or T1 is
a variable, say X.
We can assume unif y((S1 ), (T1 )) = unif y((T1 ), (S1 )), so swapping
of arguments is allowed.
Swap arguments if necessary to that S1 = X. We have already dealt
with the earlier simple cases, so we can assume that T1 is not X. We
check to see if X occurs inside T1 . If it does than no unification is
possible, since, for any substitution α, (S1 )α is strictly smaller than
(T1 )α . In this case,
unif y((S1 ), (T1 )) = F AIL.
The other possibility is that X does not occur in T1 . In this case
unif y((S1 ), (T1 )) = [X := T1 ].
Recursive Step .
There are several cases in the recursive step.
1. (Recursive Case 1)
First suppose that n = m and n and m are greater than one, and
none of the simple cases apply.
We are trying to unify
(S1 , ..., Sn )
and
3.9. THE UNIFICATION ALGORITHM 29
(T1 , ..., Tn )
We first find unif y((S1 ), (T1 )). If this fails, then unif y((S1 , ..., Sn ), (T1 , ..., Tn ))
also fails. Suppose, however that unification of S1 and T1 succeeds,
with most general unifier α. Thus (S1 )α = (T1 )α.
We apply α to the rest of the list and continue recursively. That is,
we find
unif y((S2 , ..., Sn )α, (T2 , ..., Tn )α)
If this fails, we return F AIL. If this succeeds with most general

unifier β, then
unif y((S1 , ..., Sn ), (T1 , ..., Tn )) = α ◦ β, the composition of β and α.
2. (Recursive Case 2)
Finally, suppose n = m = 1, and none of the earlier cases apply. In
this case, we want to unify S1 and T1 . We can suppose that neither
is a variable, since this case was dealt with earlier, and we can also
suppose that they are not already equal. If one of them is a constant,
then return F AIL. Suppose that neither is a constant.
It must happen that both S1 and T1 are compound terms obtained
by application of a function symbol to argument lists.
S1 = f (A1 , ..., Aj )
T1 = g(B1 , ..., Bk )
for some function symbols f and g.
If f and g are not the same, then unification fails, and we return
F AIL. If f and g are the same, then we call the unification algorithm
recursively on the argument lists. That is,
unif y((S1 ), (T1 )) = unif y((A1 , ..., Aj ), (B1 , ..., Bk ))
2
It is claimed that:
Theorem
1) The unification algorithm, as given above, always eventually termi-
nates.
2) If the algorithm terminates with a unifier this unifier is a most general
unifier.
3) If the algorithm terminates and returns FAIL, then there is no unifier.
It is not obvious that any of these claims are true. The reader is invited
to try the algorithm on a few cases to see if it works. This may or may not
inspire confidence.
Exercise. Try to prove the first claim above, that the algorithm always
terminates. We can start to think about this by a dialogue between a
defender and a critic of the algorithm. Defender: the algorithm always
terminates. Critic: No it does not. Defender: OK, then give me a simplest
problem in which it does not terminate. Critic: Define simplest. Defender:
Minimum number of variables, and subject to that the minimum length.
Now, whatever the critic responds, the defender has a strategy to win the
argument. Consider each case in turn, using the assumed minimality of the
critic’s response....
The proofs of the other two parts can be done in the same way. For
example, for part 3) the critic would have to claim that there was a case
in which the algorithm did terminate and returned fail, but that there was
a unifier. The defender makes the critic give a minimal explicit case, and
then the defender demolishes the critic. Advice to the reader: you ought
to think about this proof. You ought to be able to prove the correctness of
the algorithm in specific problems.
3.10 Problems, which may be discussed in

problems class
Problem 3.8 bish bash buf, bish bash bish bash buf buf, bish bash bish
bash bish bash buf buf buf : three utterances in some language. Specify
the language with some exact description. Then give a grammar which you
believe generates this language.
Problem 3.9 Give a grammar for the smallest term language which con-
tains all the usual variables, and also 0 and 1, and also +, -, *, written in
the usual way, with arity 2. In which cases can you leave off some brackets
without ambiguity? Give the parse tree for ((X + W ) ∗ (Y − (Z ∗ (X + 1)))).
Suppose X, Y, Z, W have values 2,3,4,5 respectively; use this to attach a
value to every node of the parse tree.
Problem 3.10 Find the composition of the substitution [X := X +Y, Y :=

Z], followed by [Z := Z 2 , Y := Z]
3.11 Applications
Unification is an essential part of prolog, which will be discussed later.
3.12. SUMMARY 31
3.12 Summary
Term languages are the sets of expressions which can be built up from an
initial signature of function names, and a set of variables. Within program-
ming languages, term languages occur as sets of expressions. An interpreta-
tion of a term language is given by giving a domain, D, and by interpreting
each function name as a function defined over that domain. For the sake of
simplicity, we are, at this stage, assuming that the values returned by our
functions are also in the domain D. The domain may be a set, such as the
real numbers, which is important in applications, but for which we do not
have an adequate notation.
Each expression in a term language has a unique parse tree. Given an
interpretation, terms in a term language can be evaluated by working up
the associated parse trees.
If I is an interpretation and τ1 and τ2 are terms |=I τ1 = τ2 means that τ1
and τ2 have the same meaning (or are semantically equal) in interpretation
I.
We have defined substitution as an operation on terms.
We have defined a unifier of two terms to be a substitution which makes
them the same. A unifier α of two terms A and B is a most general unifier
if any unifier of A and B can be obtained as α followed by some other
substitution.
The unification algorithm, given two terms A and B, either produces a
most general unifier for A and B, or, in case no unifier exists, fails.
Chapter 4
Predicate Logic
4.1 Introduction
In English and other human natural languages complex sentences and prop-
erties are built up by combining simpler ones using certain logical operators.
There are surprisingly few of these operators. On the other hand, they are
quite ambiguous in natural speech.
In order to begin formalisation of reasoning, we consider, in this chapter,
formal versions of the logical operators.
4.2 Truth values and predicates

In the following we will make the classical assumption that there are only
two truth values True and False, which we will write as T and F . (Note
that in other versions of logic, this assumption is not made.)
Definition 4.1 A predicate is a function whose codomain is the truth val-

ues {T, F }.
So predicates are usually represented as statements with variables in

them. When the variables are given values the statement should evaluate
as either true or false.
So, for example, uncle(X, Y ), which says that X is an uncle of Y , defines
a predicate. For any particular X and Y , this statement will either be true
or false.
The number of variables which occur in a predicate and which can be
instantiated in this way is called the arity of the predicate. So, for example,
33
34 CHAPTER 4. PREDICATE LOGIC
red(X), meaning that X is red, has arity 1; and between(X, Y, Z), meaning
that X is between Y and Z, has arity 3.
There are a number of formats for writing predicates. We will often put
the predicate name first and follow it by the arguments enclosed in brackets,
as in uncle(X, Y ). This is called prefix format. On the other hand, many
commonly occurring predicates use other formats. For example
X = Y , meaning that X and Y are the same
X < Y , meaning that X is less than Y .
Predicates of arity 1, such as even(X), are sometimes called properties.
Predicates can be combined together using logical operators. We will
consider the following list of operators:
1. and, written symbolically as ∧

2. or , written symbolically as ∨
3. not, written symbolically as ¬
4. implies, written symbolically as →
5. if and only if, written symbolically as ↔
6. for all, written symbolically as ∀
7. there exists, written symbolically as ∃.
In the following, we will assume that you know the truth tables for
∧, ∨, ¬, →, ↔, and that you have some working familiarity with the quan-
tifiers ∀, and ∃. If you are not happy with this, you should revise.
Note especially that the truth table for p → q is the same as the truth
table for ¬p ∨ q. This is somewhat counterintuitive. (In this case Occam’s
razor has been given priority over naturalness.) We do assume that the
truth value of a compound expression should be determined from the truth
values of its constituents, without knowing anything else about possible en-
tanglement of meaning. In other words we assume referential transparency
for the logical operators. Since some false propositions imply true ones,
we have to agree that (F → T ) evaluates as T ; and similarly (F → F )
evaluates at T . Continuing with this, we get p → q evaluates as true when
p is false, or when q is true.
4.3 Variables and types

We can of course give types to variables, and write quantifiers in the form
(∀X : N )A(X) for example. Everything we do below can also be done in
4.4. TRANSLATION FROM INFORMAL TO FORMAL LANGUAGE35
these types languages. However, in order to try to keep the notation fairly
simple, we will in this chapter assume that all variables have the same type,
which will not be declared. Each variable X ranges over some fixed domain
D. The domains are the same for all the variables.
4.4 Translation from informal to formal lan-

guage
We can now take a piece of text written in an informal language, try to
identify what the basic predicates are, and represent the text as a logical
combination of the basic predicates.
Example 4.1 Anyone who loves Jane is brave. There seem to be two pred-
icates at work here loves(X, Y ), which says that X loves Y , and brave(X),
which says that X is brave. We want to say that if X loves Jane, no matter
who or what X is, then X must be brave.
(∀X)(loves(X, jane) → brave(X))
Note that the variable X is not given a type. It just ranges over some
universe. All we know about jane is that jane is also in the same universe.
Example 4.2 Z is the union of X and Y . We could of course take this

as a basic predicate, or it could be expressed using equality and a function
symbol for union. We could also express this in terms of set membership.
In this case we want to say that any element of Z is either an element of
X or an element of Y , and vica versa.
(∀W )(W ∈ Z ↔ (W ∈ X ∨ W ∈ Y ))
One of the values of formalisation is that it uncovers ambiguities in the

original statement. It turns out that people are able to maintain mutual
misunderstandings over remarkably long periods of time without realizing
it. There is no cure for this, but formalisation is one proven technique for
diminishing this type of ambiguity.
It is quite hard to learn to do this type of translation correctly. Try
some of the examples below.
4.5 First order languages

We saw earlier how a collection of function symbols generates, in a natural
way, a term language. Once we give an interpretation to the basic function
symbols, all the terms in the term language take on meaning as functions.
Suppose now that we are given a list of function symbols and a list
of predicate symbols. As before the function symbols will generate a term
language. We can write down predicates by applying the predicate symbols
to the terms in the term language. These are called atomic formulae, since
there is no way to break them down logically. The atomic formulae can
then be combined together, using the logical operators. The resulting set of
expressions are called formulae. A first order language is the set of formulae
generated from a given list of function symbols and list of predicate symbols.
In first order languages, all the variables have the same type. In the
interpretations of these languages there is only one universal domain, over
which all variables range. Later on we will see more complicated languages:
multisorted languages in which there are several different types of variable,
and higher order languages in which we have variables for sets or functions
as well as for objects.
The lists of function symbols and predicate symbols which specify a first
order language is called its signature.
The term language always includes the variables. We can therefore allow
the list of function symbols to be empty, since we still get infinitely many
terms. But we can not use an empty predicate symbol list, since in that
case we would not get any atomic formulae or formulae. Since equality
is used so often in mathematics and computing, we will assume that the
predicate symbol list always has = in it.
An example of a first order language is LZF , given earlier. In this case,
the term language is just the set of variables. The list of predicate symbols
is [=, ∈], both with arity 2.
In any first order language, the formulae are generated from the atomic
formulae in the same way, shown in the following grammar.
< S >; (< S > ∧ < S >) | (< S > ∨ < S >) | (¬ < S >) |
(< S >→< S >) | (< S >↔< S >) | (∀ < V ariable >) < S >| (∃ <
V ariable >< S >|< Atomicf ormula >
Example 4.3 LN , the first order language of arithmetic, has the signature
with function symbols [+, ∗,0 , 0], and predicate symbol [=]. Function symbols
+ and ∗ are written in the usual way, so that for example, ((X +Y )∗(X ∗Z))
is a term. The intended meaning of X 0 is X + 1, the successor of X. The
arity zero predicate symbol, 0, is intended to mean zero.
An example of an atomic formula would be
(000 ∗ X) = Y .
An example of a formula would be
(∃Y )(000 ∗ Y = X).
Under the usual interpretation, this means that X is even.
4.6. SYNTAX, SUBSTITUTION 37
4.6 Syntax, substitution

A first order language consists of a set of terms, a set of atomic formulae
and a set of formulae. When such a language is interpreted, the terms
return values in some domain D, but the formulae (including the atomic
ones) return truth values. When an interpretation is given, the variables
are all assumed to range over the common domain of the interpretation.
It would make sense to substitute one term inside another one. It would
also make sense to substitute a term for a variable inside a formula. For
example, in LN , we could have
(∃W )(X + W 0 = Y )
and substitute Z ∗ Y for Y to obtain
(∃W )(X + W 0 = Z ∗ Y )
On the other hand, it would not make sense to try to apply a function
symbol to a formula. In general formulae may not be substituted for vari-
ables in terms, since the value returned by a formula is a truth value, and
the function symbols are defined on domain values, not truth values.
4.7 Free and bound variables

Definition 4.2 In an expression of the form (∀X)wff or (∃X)wff , the
subexpression, wff is called the scope of the quantifier.
Definition 4.3 An occurrence of a variable, X in a formula wff is bound

if it occurs in (∀X) or (∃X), or in the scope of a quantifier (∀X) or (∃X)
in wff .
Definition 4.4 An occurrence of a variable is free if it is not bound.
Example 4.4 (∀Z)((∃W )((W + Z) = X) → (∃W 1)((W 1 + Z) = Y )). In

this formula, all occurrences of X and Y are free , and all other occurrences
of other variables are bound. So this formula represents a predicate of arity
two.
Definition 4.5 A sentence is a formula with no free variables.
If A(X1 , ..., Xn ) is a formula with free variables X1 , ..., Xn , the sentence

(∀X1 )(∀X2 )...(∀Xn )A(X1 , ..., Xn )
is called the universal closure of A(X1 , ..., Xn ).
On the other hand, the sentence
(∃X1 )...(∃Xn )A(X1 , ..., Xn )
is called the existential closure of A(X1 , ..., Xn ).
Problem 4.1 LN , the language of arithmetic, with signature (0,0 , +, ∗)

with arities (0, 1, 2, 2) respectively, is a typical first order language. What
are the variables, constants, terms, atomic formulae and formulae of this
language? Give a grammar for each, and an example of each.
4.8 Semantics for first order languages

We do not assume any predefined meaning for the function symbols and
predicate symbols which occur in a first order language.
The value of a term in a first order language, or the truth value of a
formula is likely to depend upon the interpretation given to the function
symbols and the predicate symbols of the language.
Example 4.5 LG , the first order language of group theory, has term lan-
guage generated by the usual variables, 1 (meant to be the identity), X ◦ Y ,
meant to be the group operation, and X −1 meant to be the group inverse.
The only predicate symbol needed in LG is =, meant to be equality.
However, in an interpretation of LG , ◦ can be any operation : D2 →
D, and X −1 can be any unary operation. Even though we want to use
LG to talk about groups, we do not impose any of our ideas on the set of
allowable interpretations. The groups are the interpretations which satisfy
the axioms.
We do not even insist that = be interpreted as the usual equality. This
means that in this context when we give the group axioms we need to include
some statements which characterise equality, as follows.
1. X = X
2. X = Y → Y = X
3. (X = Y ∧ Y = Z) → X = Z
4. X = Y → X −1 = Y −1
5. (X = Y ∧ Z = W ) → X ◦ Z = Y ◦ W
6. X ◦ (Y ◦ Z) = (X ◦ Y ) ◦ Z
7. X ◦ 1 = X ∧ 1 ◦ X = X
8. X ◦ X −1 = 1 ∧ X −1 ◦ X = 1
4.8. SEMANTICS FOR FIRST ORDER LANGUAGES 39
To give an interpretation for a first order language we have to specify

a set D which is called the domain of the interpretation, and we have to
interpret each function symbol and predicate symbol in the signature of the
language as a function or predicate of the appropriate arity defined over
D. If a function symbol f has arity n, we have to interpret f as a function
f : Dn → D. We always assume that such functions are totally defined and
that they return values in D. If p is a predicate symbol of arity m, we have
to interpret p as a predicate p : Dm → {T, F }.
An interpretation of a first order language, L, is given by a set D called
the domain of the interpretation, together with a list of functions and pred-
icates defined over D, corresponding to the list of function symbols and
predicate symbols in the signature of L.
Let I be an interpretation of a first order language L.
Note that the interpretations are supposed to be mathematical objects,
whereas the signatures and languages are supposed to be syntactic objects.
An interpretation I is also called a structure.
If I is an interpretation of a first order language, a valuation from I
is a function which maps a subset of the variables of A into values in the
domain of I. The support of a valuation is the set of variables to which
it gives values. We will write valuations as if they were substitutions, e.g.
[X1 := d1 , . . . , Xk := dk ].
Definition 4.6 Suppose I is an interpretation of a first order language L,
v is a valuation from I, τ is a term of L and A is a formula of L. Suppose
that the variables of τ and the free variables of A are included in the support
of v. Then
τ (v) is the value in the domain of I of the term τ under valuation v.
A(v) is the truth value of A in interpretation I under valuation v.
Of course a formula may be true in one interpretation and false in oth-
ers. We are especially interested in formulae which are true in all possible
interpretations.
If S is a formula of some first order language, L, and I is an interpreta-
tion of L, we will write
|=I S
to mean that S is true in I. If S has free variables,
|=I S means that S is true for all possible values of the free variables in
the domain of I. In other words,
|=I S means S(v) = T for all valuations from I whose support contains
all the free variables of S.
S is true in I if and only if the universal closure of S is true in I
Of course the truth or falsity of S may depend upon how the predicate
symbols, function symbols and constants are interpreted in I.
Definition 4.7 We will say |= S if |=I S for all possible interpretations I.

In this case we will say that S is logically valid.
|= S means that S is logically valid. Intuitively, this is meant to express
the idea that S is true by virtue of its logical form, independently of the
interpretation of its function symbols and predicate symbols.
Definition 4.8 We will say that a formula S is satisfiable if there exists
an interpretation I and a valuation from the domain of I which makes S
true.
Intuitively, to say that S is satisfiable means that S is true in some
possible world, with some evaluation of the free variables. Note that |= S
if and only if ¬S is not satisfiable.
An important open problem in computer science is either to give a
practical algorithm to decide whether or not if a quantifier free formula is
satisfiable, or to show that no such algorithm exists. (A practical algorithm,
in this context, would be one which terminated with the right result in a
number of steps which was bounded by some polynomial in the length of
the input.) This is a version of the famous P = N P question.
Example 4.6 A formula can look correct and yet not be logically valid.
For example
(X < Y ∧ Y < Z) → X < Z.
This is not logically valid, since we are free to interpret < as any binary
relation.
Definition 4.9 Suppose Γ is a list of formulae in a first order language
and M is an interpretation of the language. We will say that M is a model
of Γ if every formula of Γ is true in M .
M is a model of Γ if and only if M is a model of the universal closure
of Γ.
Example 4.7 Let M = (D, ◦, X −1 , =, 1) be an interpretation of LG in
which D is the non zero real numbers, ◦ means multiplication and X −1 is
interpreted as 1/X, and = is the usual equality. This is a model for the
axioms of group theory.
Definition 4.10 Suppose Γ is a list of formulae, and S is a formula in
some first order language. We will say
Γ |= S
if S is true in every model of Γ. In this case we will also say that S is
a logical consequence of Γ.
4.9. EXAMPLES : LN AND LR 41
The usual situation in mathematics is that we have a list Γ of axioms,

and we are trying to decide whether or not some formula S is a logical
consequence of Γ. We all know that such problems can be very hard.
In mathematics and in computing, when we give a list of specifying
axioms, we hope to be completely explicit. The reason for allowing non
intuitive interpretations of a first order language is to expose any lack of
explicitness in sets of axioms. We do not want to build in behind the scenes
assumptions which are not explicitly written.
Note: In Γ |= S, we can replace all the formulae by their universal
closures without altering the situation.
Definition 4.11 Two formulae A and B are logically equivalent if

|= (A ↔ B).
4.9 Examples : LN and LR

LN , defined earlier is the first order language of arithmetic. In the standard
interpretation of this language, the domain is the natural numbers, includ-
ing zero, + and ∗ have their usual interpretation, and X 0 means X + 1, and
0 and 1 have the usual meaning. This structure will be called N.
Suppose S is a formula of LN . The notation |=N S means that S is true
in the standard interpretation.
There is a famous set of axioms, called the Peano postulates, which is
intended to capture the truths of arithmetic. This includes some simple
statements which are essentially recursive definitions of + and ∗,
1. X 0 = Y 0 → X = Y
2. ¬(0 = X 0 )
3. ¬(X = 0) → (∃Y )(X = Y 0 )
4. X + 0 = X
5. (X + Y 0 ) = (X + Y )0
6. X ∗ 0 = 0
7. X ∗ Y 0 = X ∗ Y + X
To this we add some axioms for equality
1. X = X
2. X = Y → Y = X
3. X = Y ∧ Y = Z → X = Z
4. X = Y → X 0 = Y 0
5. X = Y ∧ Z = W → (X + Z = Y + W ∧ X ∗ Z = Y ∗ W )
and also an infinite list of induction axioms:

((A(0) ∧ (∀X)(A(X) → A(X 0 ))) → (∀X)A(X))
where A(X) can be any formula of LN . Almost any known theorem
which can be stated in LN can be proved from the Peano postulates. We
will see later, however, that this axiom set does not succeed in capturing
all the truths of arithmetic.
As usual with axioms, we assume that all the statements are universally
true. That is, X 0 = Y 0 → X = Y means that (∀X)(∀Y )(X 0 = Y 0 → X =
Y ). A structure in which all of these axioms are true (i.e. universally true)
is called a model of the Peano postulates. We can see that there is at least
one such model, N.
Another famous first order language is LR , the first order language
of real algebra and geometry. This has functions symbols +, −, ∗, and
predicate symbols =, <, all with arity 2, and has constants 0 and 1. The
standard interpretation of this is (R, +, −, ∗, < . =, 0, 1), that is, the real
numbers with the usual meanings of the symbols. In this language, with
the standard interpretation, we could say that (X, Y ) was inside the unit
circle by the atomic formula
X ∗ X + Y ∗ Y < 1.
Unlike the situation with the natural numbers, the truths in this stan-
dard interpretation do have a complete axiomatisation.
4.10 Uses of these ideas in computing

There are a number of ways in which these ideas are used in computing
practice. Just as the designer of a bridge may try to prove that it is un-
likely to collapse, designers of computer systems try to prove that they
are unlikely to fail, or that the code which they produce has some other
desirable properties. We may consider a piece of code as a transformation
which changes the values of some variables, i.e. as a transformation on
valuations. In order to reason about the properties of code, the usual ap-
proach is to consider triples, (A, P, B) consisting of a precondition A, the
code P , and a postcondition B. The precondition and postcondition are
usually expressed in some formal language. The code is correct relative to
this triple if whenever we start the code with a valuation satisfying the pre-
condition, then, if the code terminates, the valuation on termination must
4.11. MORE SUBSTITUTION 43
satisfy the postcondition. Such a triple may also be considered as a formal

specification for the code.
Usually in this sort of application we are interested in truth in some
particular interpretation, rather than in all possible interpretations. For
example, if we were dealing with the programming language P1 defined
earlier, we would consider the standard interpretation over the integers,
with plus, minus, and times having the usual meanings.
4.11 More Substitution

Suppose A us a formula, and T1 , ..., Tn are terms, and X1 , ..., Xn are vari-
ables. Let A[X1 := T1 , ..., Xn := Tn ] be the result of substitution every free
occurrence of X1 , ..., Xn by T1 , ..., Tn respectively. As before in the term
languages, all the substitutions are supposed to be done simultaneously.
But the situation is much more complicated now than it was previously,
because of the quantifiers.
Example 4.8 Let A be (∀X)(∃Y )(B(X, Z) → C(Z, Y )).

Then A[Z := W ] would be (∀X)(∃Y )(B(X, W ) → C(W, Y )).
We can also form A[Z := g(Y )]. But there is something potentially
wrong about this substitution. This is because the variable Y in g(Y ) will
accidentally fall into the scope of the (∃Y ) quantifier.
Let (∀X)A be a formula which is true in interpretation I. We might

expect this truth to be preserved under substitution. So we might suppose
A[X := T e] to be true in I for any term T e. But the example above shows
that we may need to be careful about these substitutions. Here is another
example:
Example 4.9 Suppose we consider the language of arithmetic LN , and

take the standard interpretation, in which the domain is the natural num-
bers, = is equality, and +, ∗, 0, 1 have their usual meaning, and X 0 means
X + 1. Call this interpretation N.
(∃Y )(Y = X) is true in N. Now substitute Y 0 for X.
(∃Y )(Y = X)[X := Y 0 ] is (∃Y )(Y = Y 0 ), which is a false sentence in
N.
Why do we (incorrectly ) expect that if (∀X)A is true in I then A[X :=

τ ] should be true in I for any term τ ? We think:
If A[X := τ ] is not true in I, then there must exist a valuation v so that
A[X := τ ](v) = F . We expect that the result of first substituting and
then evaluating should be an evaluation. So this would imply
A(w) = F for some valuation w. But this can not happen if (∀X)A is
true in I.
Obviously our expectations are wrong here. It seems that some substi-
tutions followed by evaluation are not evaluations. Such substitutions are
a common source of error in mathematics and computing.
The problem is always caused by the fact that a variable in the substi-
tuted term is quantified inside the formula, and accidentally falls into the
scope of the quantifier after substitution.
Definition 4.12 A term t is free for a variable X in a formula A(X) if

there is no free occurrence of X in A(X) inside the scope of a quantifier,
where the quantified variable also occurs in the term t.
This very unpleasant definition has the following good consequence. If

|=M A(X) and if term t is free for X in A(X), then |=M A(t).
4.12 Other kinds of language and other logics

Suppose we have some domain D. We will call the elements of D objects
of order zero. Predicates over D and functions which take arguments in
D and return values in D will be called objects of order 1. These are the
predicates and functions which are represented in first order languages. In
general an order n + 1 predicate is a predicate which speaks about objects
of order n; and an order n + 1 function is a function which takes arguments
of order n and returns values of order n. Higher order languages can have
quantifiers over higher order objects. For example, a nice way to express
induction in the natural numbers would be:
For all predicates P over the natural numbers, if P (0) is true and
(∀n)(P (n) → P (n + 1) then (∀n)P (n) is true.
This is a second order statement. It is not possible to express this in
LN , the first order language of arithmetic.
Other kinds of logical connectives, and other kinds of logic, are also
important in applications. For example, in temporal logic we might want
to say that some formula A will become true at some time in the future,
or might become true. In the modal logic of interacting agents, we might
want to say that agent X believes that formula A is true; or that agent
X believes that formula A might possibly be true. Many or the artificial
systems which we create are nondeterministic, since they have autonomous
independently acting subsystems, and what might happen is as important
as what will definitely happen. For all of these ideas there are notations
and corresponding notions of logical consequence.
4.13. LOGIC GAMES 45
In this course we will work on the simplest case of first order languages
and classical logic.
4.13 Logic Games

Let S be a formula in a first order language L, and let M be an interpreta-
tion of L. Imagine a game played between two people, P, the proponent ,
and O, the opponent. P wants S to be true , and O desires falsity of S. The
verification game is played using the formula S and the interpretation M .
If S is an atomic sentence, P wins immediately if M |= S, and otherwise
O wins immediately. If there are free variables, the opponent O chooses
values for them in M , and the value is substituted into S. If S is (∀X)A
then the opponent chooses a value for X. If S is (∃X)A, the proponent P is
allowed to choose a value for X. If S is A ∨ B the proponent chooses either
A or B, and play continues with this choice. If S is A ∧ B, the opponent
chooses either A or B and play continues with this choice. If S is ¬A, the
proponent and the opponent switch roles, and play continues. A → B is
dealt with as ¬A ∨ B.
From this we get:
M |= S if and only if the proponent has a winning strategy in this game.
If S is a sentence, then M |= ¬S if and only if the opponent has a
winning strategy in this game.
Try the following in the problems class, with one team playing proponent
and one playing opponent. Which has a winning strategy? What is it?
Example 4.10 Let S be (∀X)(∀Y )(r(X, Y ) ∨ (∃Z)(r(X, Z) ∧ f (Z, Y )))

Let M have domain {1, 2, 3, 4}. Assume r(3, 1), r(1, 2), r(2, 3), r(4, 3), r(4, 2).
Next let the tutor change M and S and repeat the game.
4.14 Normal Forms

Within a first order language there are many formulae which are logically
equivalent to any given formula. Given a formula, we would, of course,
like to be able to transform it, preserving logical equivalence, into some
simplified standard form.
These standard forms are also called normal forms, and there are quite
a few of them: conjunctive normal form (CNF), disjunctive normal form
(DNF), Prenex normal form, Skolem form, and clausal form.
4.14.1 CNF and DNF

CNF and DNF are obtained by rearranging propositional operators ∧, ∨, →
, ↔, ¬ between formulae.
We will say that a formula S is a conjunction of formulae A1 , ..., An if
S has the form
A1 ∧ A2 ∧ ... ∧ An .
We
Vn will write this as
i=1 Ai
We will say that a formula of the form A1 ∨ A2 ∨ ... ∨ An is a disjunction
of formulae
Wn A1 , ..., An , and write this as
i=1 Ai
Define a literal to be either an atomic formula or the negation of an
atomic formula.
Definition 4.13 A formula is in conjunctive normal form (CNF) if it is a

conjunction of disjunctions of literals. A formula is in disjunctive normal
form (DNF) if it is a disjunction of conjunctions of literals.
So
Vn a formula
Wmi in CNF has the form
i=1 j=1 A ij , where each Aij is a literal.
That is, a formula in CNF all the ands on the outside, and all the
negations applying directly to atomic formulae.
On
Wn theVmother hand a formula in DNF has the form
i
i=1 j=1 A ij , where each AIj is a literal.
Theorem 4.1 Let S be any quantifier free formula. We can find a formula
in CNF which is logically equivalent to S. We can also find a formula in
DNF which is logically equivalent to S.
proof.
We will give an algorithm for this.
Algorithm 4.1 DNF

Input: quantifier free formula S.
Output: DN F (S) logically equivalent to S.
Method: Find the atomic formulae A1 , ..., An of S. Construct a truth
table, showing how the truth value of S depends on the truth values of these
atomic formulae. This truth table will have n + 1 columns, for the truth
values of A1 , ..., An , S respectively, and 2n rows. Let the rows which make
S true be R1 , ..., Rk . We have
S is true if and only if
R1 happens or R2 happens or ... Rk happens.
4.14. NORMAL FORMS 47
So return DN F (S) as
W
happens(Ri )
Vn write happens(Ri )
and
j=1 Bij
where Bij = Ai if Ri gives Aj value True, and Bij = (¬Ai ) if Ri gives
Aj value False.
2
The CNF algorithm does this the other way around.
Algorithm 4.2 CNF

Input: quantifier free formula S.
Output: CN F (S) logically equivalent to S.
Method: Find the atomic formulae A1 , ..., An of S. Construct a truth
table, showing how the truth value of S depends on the truth values of these
atomic formulae. This truth table will have n + 1 columns, for the truth
values of A1 , ..., An , S respectively, and 2n rows. Let the rows which make
S false be R1 , ..., Rk . We have
S is true if and only if
R1 does not happen and R2 does not happen and ... Rk does not happen.
V return DN F (S) as
So
nothappens(Ri )
and
Wn write nothappens(Ri )
j=1 Bij
where Bij = Ai if Ri gives Aj value False, and Bij = (¬Ai ) if Ri gives
Aj value True.
Example 4.11 Consider C ↔ D. Construct the truth table. DNF for

this is (C ∧ D) ∨ ¬C ∧ ¬D). CNF for this is (C ∨ ¬D) ∧ (¬C ∨ D).
Note that the method given above can not be called practical, since the
amount of work to be done increases with the number of rows of the truth
table and this increases exponentially with the number of atomic formulae
in the problem.
There are many other ways of finding CNF and DNF, which are not
unique. None of the other known methods are practical, i.e. have complex-
ity bounded by some polynomial in the size of the input.
Exercise: Write specifications for the DNF and the CNF algorithms.
What does it mean to say that the algorithms are correct with respect to the
specifications? What does it mean to say that the algorithms are complete
with respect to the specifications? Do the algorithms always terminate?
4.14.2 Prenex Normal form

A formula is in prenex normal form if all the quantifiers are in the front.
This means that the formula has the form
Q1 Q2 ...Qk M
where each Qi has the form (∃Xi ) or (∀Xi ), and Xi is a variable, and
M is quantifier free.
Note: you may need to use all the correct bracketing to be sure that a
formula is in prenex normal form.
For example,
(∀X)(A(X) → B(X)) is in prenex normal form. However
((∀X)A(X) → B(X)) is not in prenex normal form, since B(X) is not
in the scope of the quantifier in this case.
To say that a formula is in prenex normal form means that all the out-
ermost operators are quantifiers. In ((∀X)A(X) → B(X)), the outermost
operator is →.
Given any formula S, we can find a logically equivalent formula in prenex
normal form. This is done by progressively moving the quantifiers outside
the propositional operators, as explained below.
In order to do this correctly, it is sometimes necessary to rename bound
variables. (This is also called α conversion.) The idea is to replace (∀X)A(X)
by (∀Y )A(Y ), where Y is a new variable, which was not used previously.
In this case (∀X)A(X) has the same meaning (i.e. is logically equivalent
to) (∀Y )A(Y ). Similarly (∃X)A(X) is logically equivalent to (∃Y )A(Y ),
provided that Y is a new variable.
For example (∃X)(X = W 0 ) has the same meaning as (∃Y )(Y = W 0 ).
We can rename the bound variable X as Y , since Y does not occur in
the original formula. On the other hand, great confusion can result by
renaming bound variables with variables which are already present. For
example (∃X)(X = W 0 ) does not have anything like the same meaning as
(∃W )(W = W 0 ).
Algorithm 4.3 Input: Formula S

Output: a formula in prenex normal form logically equivalent to S.
Method:
1. Replace all subformulae using ↔ using

(A ↔ B) ; (A → B) ∧ (B → A)
2. Replace all subformulae using → using

(A → B) ; (¬A ∨ B)
3. Rename all bound variables so that no variable occurs both bound and
free in any subformula, and so that no two occurrences of the same
variable are in the scopes of different quantifiers.
4. Move negations inside quantifiers with

¬(∀X)A(X) ; (∃X)(¬A(X))
¬(∃X)A(X) ; (∀X)(¬A(X))
5. Move all ∧ and ∨ inside all quantifiers with

((QX)A opB) ; (QX)(A opB) and
(A op(QX)B) ; (QX)(AopB)
4.14.3 Skolem form

It is fairly difficult to understand formulae with alternations of quantifiers.
A famous example of this is the definition of continuity of a function f (X)
at a point X = a.:
(∀)(∃δ)(∀X)((( > 0 → δ > 0) ∧ (| X − a |< δ) → (| f (X) − f (a) |< )).
Suppose we are asked what this means. We might say that if we were
given an > 0 we could find a number δ > 0 so that, for all X , | f (X) −
f (a) |< if | X − a |< δ. This seems equally hard to understand. How can
we simplify this?
An important point is that the number δ depends on the number , but
not on the number X. So we might proceed by introducing a function δ().
We would then have
(∀)(∀X)
( > 0 → δ() > 0∧
(| X − a |< δ() →| f (X) − f (a) |< .
Of course it is still complicated. But using this function seems to im-
prove things a bit. The function δ() is an example of a Skolem function.
A formula is in Skolem form if it is in prenex normal form and only uses
universal quantifiers. In order to put a formula into Skolem form we get
rid of existential quantifiers by using new function symbols, such as δ() in
the example above..
Suppose a sentence is true in some interpretation M , and is in Skolem
form. In terms of the verification game between Proponent and Opponent
mentioned earlier, the Skolem functions give a winning strategy to the
Proponent.
Example 4.12 Consider the formula

(∀X)(∀Y )(∃Z)(X ≤ Z ∧ Y ≤ Z)
This is in prenex normal form. The Z alleged to exist depends on X
and Y . If this statement is true, there must be some function Z = f (X, Y )
which makes it true. The Skolem form of the statement is:
(∀X)(∀Y )(X ≤ f (X, Y ) ∧ Y ≤ f (X, Y ))
The existential quantifier has been eliminated by introducing a new func-
tion f (X, Y ).
It is fairly easy to see how to find a Skolem form for a formula in prenex
normal form. Suppose our formula is:
(∀X1 )(∀X2 )...(∀Xn )(∃Y )M (W1 , ..., Wk , X1 , ..., Xn , Y )
where M (W1 , ..., Wk , X1 , ..., Xn , Y ) is another formula in prenex normal
form, having free variables W1 , ..., Wk , X1 , ..., Xn , Y . Suppose we wish to
introduce a Skolem function to get rid of the (∃Y ).
The value Y which is asserted to exist by the original formula may
depend on W1 , ..., Wk , X1 , ..., Xn . We pick a new function symbol for our
Skolem function. Not trying to be very original here, suppose we choose
f . It will have arity k + n, where k is the number of free variables in the
original formula and n is the number of universal quantifiers to the left of
(∃Y ). Replacing this existential quantifier, we get:
(∀X1 )(∀X2 )...(∀Xn )M (W1 , ..., Wk , X1 , ..., Xn , f (W1 , ..., Wk , X1 , ..., Xn ))
Inside M there may remain more existential quantifiers. So this process
of inventing new function symbols and replacing existential quantifiers is
repeated until we get Skolem form.
Example 4.13 A Skolem form of

(∀X)(∃Y )(∀Z)(∀W )(∃V )A(X, Y, Z, W, V, U )
would be
(∀X)(∀Z)(∀W )A(X, dog(X, U ), Z, W, cat(X, Z, W, U ), U )
using cat and dog as new function symbols.
Skolem form is not unique, since we can use any new function symbol
we choose. It is essential to use a function symbol which has not been used
previously.
The Skolem form of a formula logically implies the original formula.
However the converse is false. The original formula does not imply the
Skolem formula. The Skolem form is slightly stronger. So we have lost
logical equivalence. On the other hand, the difference is small.
Suppose we have an interpretation, I, in which the original formula is
true. The Skolem form can not even be interpreted in I as it stands, since
we have used new function symbols. But since the original formula is true,
we can extend I, defining the Skolem functions so that the Skolem form is
true.
Theorem 4.2 Suppose sk(A) is a Skolem form of formula A. sk(A) |= A.

Furthermore, if I is an interpretation in which A is true, then I can be
extended, by appropriate definition of the Skolem function symbols, to an
interpretation I∗ so that sk(A) is true in I∗.
If we regard a formula A as an axiom which makes some demands on

reality, it seems reasonable to say that a Skolem form of A makes the same
demands, except for notation. If it is possible to satisfy A, it is also possible
to satisfy a Skolem form of A, and vica versa.
Note that a formula is satisfiable if and only if its Skolem form is satis-
fiable.
Remark: we can keep logical equivalence if we are willing to go up
to a higher order language. For example (∀X)(∃Y )a(X, Y ) has Skolem
form (∀X)a(X, f (X)), where f (X) is a Skolem function. If we allowed
quantification over function variables we would have
(∃f )(∀X)a(X, f (X))
and this is logically equivalent to the original statement.
4.14.4 Clausal form

Definition 4.14 A clause is a formula of the form
(A1 ∧ A2 ∧ ... ∧ Ar ) → (B1 ∨ B2 ∨ ... ∨ Bs ), where A1 , ..., Ar , B1 , ..., Bs
are atomic formulae.
In a clause, a conjunction of atomic formulae implies a disjunction of

atomic formulae. It allowed for for either the conjunction or the disjunction
to be empty. The clause
(A1 ∧ A2 ∧ ... ∧ Ar ) → (B1 ∨ B2 ∨ ... ∨ Bs )
is logically equivalent to the disjunction of literals
(¬A1 ∨ ¬A2 ∨ ... ∨ ¬Ar ∨ B1 ∨ B2 ∨ ... ∨ Bs ).
There is no mechanical method for putting a sentence into the simplest
form. The nearest approach we have to this is clausal form. Given a
sentence A, a clausal form is obtained as follows.
1. Find prenex normal form of A
2. Find Skolem form. We now have something of the form

(∀X1 )...(∀Xn )M (X1 , ..., Xn ),
where M is quantifier free.
3. Delete the universal quantifiers. We now have a quantifier free formula

M (X1 , ..., Xn )
4. Put M into conjunctive normal form. We get C1 ∧ C2 ∧ ... ∧ Cj , where
each Ci is a disjunction of literals.
5. Write the conjunction as a list Γ = [C1 , ..., Cj ]
6. Write each disjunction ¬A1 ∨ ¬A2 ∨ ... ∨ ¬Ar ∨ B1 ∨ ... ∨ Bs in the
form of a clause
(A1 ∧ A2 ∧ ... ∧ Ar ) → (B1 ∨ ... ∨ Bs ).
The result is a clausal form of A.
This method only makes sense for sentences. That is, we have to start
with A which has no free variables.
We end up with a list of quantifier free clauses. Except for the fact
that we may have introduced new function symbols, the clausal form has
the same meaning as A. On the other hand, the clausal form is often a lot
easier to understand than the original.
The clausal form process may be regarded as a simplification procedure.
Example 4.14 Consider again the continuity definition:

(∀)(∃δ)(∀X)((( > 0 → δ > 0) ∧ (| X − a |< δ) → (| f (X) − f (a) |< )).
A clausal form for this is:
> 0 → δ() > 0
(| X − a |< δ()) → (| f (X) − f (a) |< )).
4.15. PROBLEMS 53
4.15 Problems
Problem 4.2 (6) Put the following statements into conjunctive normal
form, and also disjunctive normal form. Check that each statement has the
same truth table as its conjunctive normal form.
a) (¬a ∧ b) → ¬b)
b) (a ∧ ((¬p) ↔ ¬q))
Problem 4.3 (4)

Find clausal form for ((a ∨ b) → c)
Problem 4.4 (45) Translate the following statements into some first order
language. Note that in e) and h) you may need to use equality.
a) Not all toothless animals which have feathers can fly.
b) If some human being is in prison than all human beings are in prison.
c) It is not true that every short person who is not crazy likes some
other person who is either not short or crazy.
d) Gorillas are hairier than billiard balls.
e) Everyone is either loved or hated by some other person.
f ) There is a town in Spain in which there is a barber who shaves ev-
eryone in the town who does not shave themselves.
g) Everyone is happy except for Jane.
h) There is only one thing with wheels in the shed and that is a bicycle
with a flat tire.
i) For every X and Y there is a Z so that X ≤ Z and Y ≤ Z, and if
W is such that X ≤ W and Y ≤ W then Z ≤ W .
note: such a Z is called the least upper bound of X and Y . Can you see
how to establish that least upper bounds are unique?
Problem 4.5 (5) a) Repeat the first two parts of previous problem (i.e.
the statements about feathered animals and prisoners) without using any
existential quantifier.
b) (5) Repeat the first two parts of the previous problem without using
any universal quantifier.
Problem 4.6 (20)

Put the statements a,b,c,d,e above into clausal form.
4.16 Summary
In this chapter we have defined first order languages and given them a
semantics. We have defined the notions of logical consequence and logical
equivalence.
We have also developed a number of normal forms for formulae in first
order languages: conjunctive normal form, disjunctive normal form, prenex
normal form, Skolem form, and clausal form.
The most valuable and interesting of the normal forms is clausal form.
The clausal form of a sentence is a list of clauses, each clause being in the
form:
V W
Ai → Bi
where each Ai and Bi is an atomic formula. Thus clausal form uses no
quantifiers, and seems not to use negation. The crucial step in the reduction
of a sentence to clausal form is replacement of existential quantifiers by use
of Skolem functions.
The clausal form of a sentence is weakly equivalent to the original sen-
tence in this sense: Any interpretation which makes the clausal form true
also makes the original sentence true. Any interpretation which makes the
original sentence true can be extended, by definition of the Skolem func-
tions, to an interpretation which makes the clausal form true. In particular,
the original sentence can be satisfied if and only if the clausal form can be
satisfied.
4.17. REVISION PROBLEMS. CLAUSAL FORM 55
4.17 Revision Problems. Clausal form

Put the following into clausal form.
Problem 4.7 No man is an island
Problem 4.8 Every dog, except for Hugo, is frightened of some cat. But
Hugo is not frightened of any cat.
Problem 4.9 If you can get from A to B by rail and you can get from B
to C by rail, then you can get from A to C by rail, except on Wednesday.
Problem 4.10 Jane only likes men who are crazy.
Problem 4.11 Every bar stool has a person sitting on it, and every such
person has a mother, alive or dead.
Problem 4.12 Substitution α is a most general unifier of terms T and S.
Problem 4.13 If A(0) is true and if A(x) implies A(x + 1) for all x, then
A(x) is true for all x.
Problem 4.14 (∀)( > 0 → (∃δ)(δ > 0 ∧ (∀X)(| X − X0 |< δ →| f (X) −

f (X0 ) |< )))
Problem 4.15 ((∃X)A(X) → (∃X)B(X))
Problem 4.16 ((∀X)A(X) → (∀X)B(X))
Problem 4.17 ((∃X)A(X) → ((∃X)B(X) → (∃X)C(X)))
4.18 Solutions
1. m(X) ∧ is(X) →
2. d(X) → X = h ∨ c(γ(X))
d(X) → X = h ∨ f right(X, γ(X))
3. r(A, B) ∧ r(B, C) → wednesday ∨ r(A, C)
4. male(X) ∧ likes(j, X) → crazy(X)
5. barstool(B) → human(α(B))
barstool(B) → mother(µ(B), α(B))
barstool(B) → d(µ(B)) ∨ al(µ(B))
6. → Tα = Sα
Tβ = Sβ → Tα◦δ(α,β) = Tβ
7. A(0) → A(τ ) ∨ A(Y )

A(0) ∧ A(τ + 1) → A(Y )
8. see previous notes
9. B(X) → C(τ (X))
10. A(τ ) → B(Y )
11. A(X) ∧ B(Y ) → C(τ (X, Y ))
4.19 Examples and Solutions

1. Translate the following into some first order language. In each case,
give the signature of the language. Where there seems to be ambiguity
in the English, point this out and explain which possibility you have
chosen. (30)
(a) If some of Al’s chickens have mange, then all of Al’s chickens
have mange.
(∃X)(a(X)∧ch(X)∧mg(X)) → (∀X)((a(X)∧ch(X)) → mg(X))
(b) Some people like other people who do not like anyone.
(∃X)(p(X)∧(∃Y )(p(Y )∧likes(X, Y )∧(∀Z)(p(Z) → ¬likes(Y, Z))))
(c) George must be eliminated unless Alice only saw one kangaroo.
I translate “A unless B” as “A or B”.
elim(g) ∨ ((∃X)(saw(a, X) ∧ kang(X)) ∧ (∀Y )(∀W )(kang(Y ) ∧
kang(W ) ∧ saw(a, Y ) ∧ saw(a, W )) → Y = W )
2. Put the statements of the previous section into clausal form. (30)
Prenex normal form is
(∀X)(∀Y )((a(X) ∧ ch(X) ∧ mg(X) ∧ a(Y ) ∧ ch(Y )) → mg(Y ))
a)
Clausal form is
(a(X) ∧ ch(X) ∧ mg(X) ∧ a(Y ) ∧ ch(Y )) → mg(Y )
b) Clausal form is
→ p(τ )
4.19. EXAMPLES AND SOLUTIONS 57
→ p(ρ)
→ likes(τ, ρ)
p(Z) ∧ likes(ρ, Z) →
c) Clausal form is
→ elim(g) ∨ saw(a, τ )
→ elim(g) ∨ kang(τ )
saw(a, X) ∧ saw(a, W ) ∧ kang(X) ∧ kang(W ) → elim(g) ∨ X = W
3. Put the following statement into conjunctive normal form: (((p∧q) →

r) → ¬q) (10)
(p ∨ ¬q) ∧ (¬r ∨ ¬q)
4. Put the following statement into Skolem form:
((∀X)a(X) → (∀X)(∃Z)b(X, Z)) (10)
(∃X)(∀Y )(∃Z)(a(X) → b(Y, Z)) is prenex normal form.
Skolem form is
(∀Y )(a(τ ) → b(Y, h(Y ))
Chapter 5
Semantic Tableaux
Assume that we have a set of axioms Γ, written in some first order language,
for some field of knowledge. Imagine that a lot of work and experience has
gone into Γ. We believe that Γ summarises every known truth in some area
which concerns us.
The nice feature of this situation is that Γ may be quite small. It is a
compact representation for all of its logical consequences. We believe that
the information we want is in Γ. But how do we get this information?
It becomes clear that we need to be able to answer the following ques-
tion.
Given Γ, and formula A, decide whether or not Γ |= A.
This is called the logical consequence problem, for the predicate calculus.
Of course this may be hard. As if it were not hard enough, another
problem immediately suggests itself.
Given formula A(X1 , ..., Xn ) with free variables X1 , ..., Xn , find,
if possible, values v1 , ..., vn so that Γ |= A(v1 , ..., vn ).
If we are to make some progress with these problems computationally,
we need some formalisation of the concept of proof.
There are a number of formal systems for deduction in predicate logic.
One technique is called forward chaining. A forward chaining system is
given by a certain number of rules of inference. Sometimes, for example,
such systems include the following rule, which is called modus ponens:
———————————-
A, (A → B)
————————-
B
———————————-
59
60 CHAPTER 5. SEMANTIC TABLEAUX
This says that if we have proved A and also (A → B), we may conclude
B.
In a forward chaining system, we start with the axioms Γ and apply the
rules of inference. We regard all the formulae in Γ as true, and every time
we apply a rule of inference we enlarge the set of known truths. We will
say Γ ` A if we can eventually derive A from Γ using the rules of inference.
If we think that A really is a logical consequence of Γ, we must try to get
from Γ to A using the rules of inference.
There are several severe difficulties here. Even if we know that Γ |= A,
it may be extremely difficult to find the right way to apply the rules of
inference to get from Γ to A. If we are not confident that A is a logical
consequence of Γ, our problem is even worse. A forward chaining system
will never establish that A is not a logical consequence of Γ.
Backward chaining systems are based on the method of proof by con-
tradiction. We begin by assuming that A is not a logical consequence of Γ
and proceed to attempt to construct a counterexample. (A counterexample
would be an interpretation in which Γ is true but A is false.) If our attempt
to construct a counterexample is eventually blocked by contradictions, we
conclude that there is no counterexample, and thus our original assumption
was incorrect, and Γ |= A. On the other hand, if some branch of the con-
struction never gets blocked, the construction, in the limit, should produce
a counterexample, and in this case A is not a logical consequence of Γ.
As mentioned above there are a large number of formalisations of proof
in predicate logic. The semantic tableaux method is one of these. This is
a backward chaining system.
5.1 Introduction
We should sort out the notation first. Γ |= A means that A is a logical
consequence of Γ. As explained previously, this means that if I is any
interpretation which makes Γ true, then I must make A true. So Γ |= A is
about semantics. It means that in reality A is true whenever Γ is true.
On the other hand, we will have a syntactic definition of proof.
Definition 5.1 If Γ is a list of sentences and S is a sentence, we will write
Γ`S
if we can prove S from Γ in the semantic tableaux system.
Note that at this stage we are only dealing with sentences. So we will
need to replace our axioms by their universal closure.
5.1. INTRODUCTION 61
Γ ` A means that there is a formal proof of A from Γ. We now have to

explain what such a formal proof would be.
Since we are trying to construct counterexamples, we will have to explore
a number of alternatives. To reflect this the data structures we will use will
be trees labelled with formulae.
We will first give some examples, and then describe the technique more
precisely.
Example 5.1 Suppose Γ is the empty set, and A is (p → (q → p)). We

want to know whether or not (p → (q → p)) is logically valid. We are
using proof by contradiction. So assume that it is false. This gives us the
beginning of our proof tree, i.e. one node.
(1) not (p->(q->p))
The form of the supposedly false expression is A → B. How can such a

formula be false? (Review the truth table definition of implication.) In fact
there is only one way that (A → B) can be false. It must be that A is true
and B is false. So we can see how to continue our construction.
(1) not (p->(q->p))

|
(2, from 1) p
|
(3, from 1) not (q->p)
We now continue the construction at node (3).
(1) not (p->(q->p))

|
(2, from 1) p
|
(3, from 1) not (q->p)
|
(4, from 3) q
|
(5, from 3) not p , contradicts (2)
The construction ends in a contradiction. That is, node (2) tells us

that p is true, and node (5) tells us that p is false. Therefore, there is no
counterexample. Thus |= (p → (q → p)).
Furthermore, the tree shown above is a proof of this in the semantic
tableau system.
(Note that in these examples “not p” is used instead of ¬p, purely for
the convenience of the typist. In later semantic tableaux, “(For all X)” and
“(Exists X)” will be used instead of (∀X) and (∃X) for the same reason.)
Here is another example, in which there is some branching.
Example 5.2 Decide whether or not |= ((p → q) → q).
(1) not ((p->q)->q)

|
(2, from 1) (p ->q)
|
(3, from 1) not q
At this point we have (p → q) asserted to be true. How can an im-

plication be true? There are two ways in which this can happen. (Please
review truth table definition of implication.) Either the premise is false or
the conclusion is true. So we continue the construction by branching to
reflect this.
(1) not((p->q)->q)
|
(2, from 1) (p ->q)
|
(3, from 1) not q
/ \
/ \
(4, from 2) not p (4, from 2) q contradicts 3)

5.2. SEMANTIC TABLEAU RULES 63
Please look at the tree below node 3). It is supposed to say that either p
is false or q is true.
We now have a contradiction on one branch. There is no contradiction
on the other branch. There is also nothing more we can do. So the con-
struction on the left hand branch has ended in a counterexample. Reading
down the left hand branch, we can see that p and q are both false. (Check
that this is a counterexample.)
In a tree, the node at the top is called the root. The node or nodes
immediately below a node are called its children. A node with no children
is called a leaf. The collection of leaves is called the frontier of the tree. A
path which starts at the root and goes all the way to the frontier will be
called a branch. If a node is labelled with formula A, we will say that A is
asserted at the node. If a node is labelled with ¬A, we will say that A is
denied at the node.
In a semantic tableau, we will say that a branch is closed if it contains
a contradiction. This means that some formula is both asserted and denied
on the branch. The tableau is closed if all branches are closed.
Suppose we wish to try to decide whether or not Γ |= A. Assume that
A is a sentence, and all the formulae in Γ are also sentences. The semantic
tableau method is the following. Form an initial semantic tableau with all
the formulae of Γ asserted, and A denied. Then apply the semantic tableau
rules, which will be described below. If a closed tableau is eventually ob-
tained, it follows that no counterexample is possible, and so A is a logical
consequence of Γ. The closed tableau is a proof of A from axioms Γ.
5.2 Semantic tableau rules

1. If (A → B) is denied, add nodes to assert A and deny B.
2. If (A → B) is asserted, form two branches and add a node denying A
on one and a node asserting B on the other.
3. If ¬A is denied, add a node asserting A
4. If (A ∧ B) is asserted, add two nodes, one asserting A and the other
asserting B.
5. If (A ∧ B) is denied, form two branches, and deny A on one and deny

B on the other.
6. If (A ∨ B) is denied, add two nodes, one denying A and the other

denying B .
7. If (A ∨ B) is asserted, form two branches, and assert A on one, and

assert B on the other.
8. If (A ↔ B) is asserted, form two branches, and deny both A and B

on one and assert both A and B on the other.
9. If (A ↔ B) is denied, form two branches and assert A and deny B on

one and assert B and deny A on the other.
10. If (∀X)A(X) is asserted, and if t is any variable free term, assert A(t).
This may be done for any number of variable free terms t.
11. If (∀X)A(X) is denied invent a new Skolem constant c, and deny A(c)
.
12. If (∃X)A(X) is asserted, invent a new Skolem constant c and assert

A(c).
13. If (∃X)A(X) is denied, and if t is any variable free term, deny A(t).
This may be done fro any number of variable free terms t.
Rules 10) and 13) are called substitution rules. They may be applied any
number of times. Some proofs need many substitutions. Note that only
variable free terms may be substituted. So, for example, in arithmetic,
(0000000 + 000 ) may be substituted, but (X + 00 ) may not be substituted. This
restriction ensures that all the formulae in a semantic tableau are sentences.
Variable free terms may involve arbitrarily many function symbols and
constants, and may be nested to any depth. Rule 10) says that if we assert
(∀X)A(X) and if t is any variable free term, then we may also assert A(t).
Rule 13) says that if we deny (∃X)A(X) and if t is any variable free term,
then we may also deny A(t).
In rules 11) and 12), a new Skolem constant is introduced. It is impor-
tant to realize that the constant which is introduced must not have been
previously used anywhere in the tree. The constant names an object which
either makes (∃X)A(X) true, or makes (∀X)A(X) false. If we correctly
assert (∃X)A(X), then there must exist an object which makes A(X) true;
we can call such an object by any name we choose, provided only that
this name has not been used previously. Similarly, if we correctly deny
5.2. SEMANTIC TABLEAU RULES 65
(∀X)A(X), it must follow that there is at least one object which makes
A(X) false; and we can invent a new name for such an object.
Note that a denied disjunction behaves like a conjunction of denials.
That is ¬(A∨B) is the same as ¬A∧¬B. Furthermore, a denied conjunction
behaves like a disjunction of denials.
The universal quantifier ∀ is, in a sense, just a big conjunction, and ∃ is,
in a sense, just a big disjunction. So a universal quantifier asserted behaves
like an existential quantifier denied.
Example 5.3 Suppose we want to test whether or not (∀X)(A(X) →
B(X)) logically implies ((∀X)A(X) → (∀X)B(X)). We start by asserting
(∀X)(A(X) → B(X)) and denying ((∀X)A(X) → (∀X)B(X))
(1) (Forall X)(A(X) -> B(X))

|
(2) not((Forall X)A(X) -> (Forall X)B(X))
|
(3, from 2) (Forall X)A(X)
|
(4, from 2) not(Forall X)B(X))
|
(5, from 4, add new constant c) not B(c)
|
(6, from 3, substitution) A(c)
|
(7, from 1, substitution) A(c) -> B(c)
/ \
(8, from 7) B(c) (8, from 7) not A(c), contradicts (6)

contradicts (5)
Note that the Skolem constant c does not occur above node 5), where
it was introduced, using rule 13). All branches are closed. So there is no
counterexample. Therefore
(∀X)(A(X) → B(X)) |= (∀X)A(X) → (∀X)B(X). The above tree is a
proof of this in the semantic tableau system.
Definition 5.2 If T1 and T2 are semantic tableaux, we will say T1 ⇒ T2
if T2 can be obtained from T1 by one application of the rules.
Definition 5.3 Let Γ be a set of sentences, and let A be a sentence in some

first order language. We will say Γ ` A in the semantic tableau system if
T1 ⇒∗ Tk where Tk is closed, and T1 is the initial tableau with Γ asserted
and A denied.
5.3 Some advice about the use of the rules

The semantic tableau rules may be applied in any order. But it seems
advisable to apply certain rules before others. With this in mind, we can
divide the rules into categories.
1. Propositional rules (involving ∧, ∨, ¬, →) which do not increase the

number of branches.
2. Quantifier rules involving introduction of a new Skolem constant.
3. Propositional rules which may increase the number of branches.
4. Substitution rules (which involve choose a variable free term and sub-
stituting it). That is, when (∀X)A(X) is asserted, or (∃X)A(X) is
denied.
On each branch, the rules in the first three categories only need to be
applied once. The substitution rules can be applied any number of times
with different terms on any branch. Since we do not want to have a lot
of branches, it seems a good idea to make all applications of rules in the
first two categories, before doing anything else. In general, it also seems
a good idea to do all possible category 3) operations, before doing any
substitutions. Of course there are cases in which this advice is not useful.
Whatever strategy we follow, we do not want indefinitely to defer taking
any allowed operation. Everything which can be done on any open branch
should eventually be done. Unless of course the branch gets closed for
some other reason. Since if there are function symbols in our language,
there may be infinitely many variable free terms, and thus infinitely many
possible substitutions, a perverse user of the semantic tableau system could
continue forever without getting a proof by insisting on doing one useless
substitution after another, even though closure might be obtained in some
other way.
Definition 5.4 We will say that the semantic tableau construction is done
systematically if any operation which is possible on any open branch (in-
cluding all substitutions) is eventually done, unless all branches extending
it get closed first.
5.4. HOW MUCH USE IS THIS SYSTEM? 67
We will return later to the important question of how to choose useful

substitutions, from among the many possibilities. You may already guess
that unification will be involved in this.
Before reading the next section, you should probably try the following.
Note that in the one case in which validity can be proved, a stubborn user
of the system can avoid ever getting closure, and thus may never find the
proof, by insisting on doing an infinite sequence of useless substitutions.
Problem 5.1 Either give a proof of validity or a counterexample.
1. ((∃X)a(X) → (∃X)b(X)) → (∃X)(a(X) → b(X))

2. ((∃X)(a(X) → b(X)) → (∃X)a(X) → (∃X)b(X))
5.4 How much use is this system?

5.4.1 Gödel completeness theorem
To begin with, we have an extraordinary theorem, due to Kurt Gödel.
Assume that we are using a first order language with at least one con-
stant, so that we have something to substitute to begin with. We can always
add such a constant to any first order language.
Theorem 5.1 Gödel’s completeness and correctness theorem !!
Γ |= S
if and only if
Γ`S
for any list of sentences Γ and any S in any first order language.
We will eventually sketch a proof of this. Before the proof, we need

some definitions and a lemma.
Definition 5.5 We will say that an interpretation I satisfies a branch α

of a semantic tableau, T , if every sentence asserted in α is true in I and
every sentence denied in α is false in I.
Definition 5.6 We will say that a semantic tableau T is satisfiable if there

exists an interpretation I and a branch α of T so that I satisfies α.
Lemma 5.1 One hop Lemma

If T1 ⇒ T2 and T1 is satisfiable, then T2 is satisfiable.
We suppose that T1 is satisfiable. So there is some branch α1 in T1 so

that α1 is satisfiable.
T2 is obtained from T1 in one step, by one application of the rules. This
means that there is some path β1 in T1 so that T2 is obtained from T1 by
an extension of β1 . If α1 6= β1 , then the branch α1 is still in T2 , and so T2
remains satisfiable. Now consider the case in which α1 = β1 .
We need to do is show that the rules preserve satisfiability. This can
be done by considering each rule in turn. Suppose T1 ⇒ T2 using rule 1).
Then branch α1 of T1 contains a node labelled ¬(A → B). T2 is obtained
by adding two nodes to α1 , and denying A on one and asserting B on the
other. The branch α1 in T1 is extended to a branch α2 in T2 . We suppose
α1 is satisfiable. The satisfying interpretation makes A → B false. Thus it
must make A true and B false. Thus it must satisfy the extension of α1 in
T2 . Thus T2 is satisfiable. The other cases are similar.
Corollary 5.1 If T1 ⇒∗ Tk and T1 is satisfiable, then Tk is satisfiable.
Proof. By induction on the number of steps in the derivation, using the

one hop lemma.
2.
Correctness
We can now prove the correctness part of the theorem above.
Suppose Γ ` A. Let T1 be the initial tableau with Γ asserted and
A denied. T1 ⇒∗ Tk , where Tk is closed. Since Tk is closed, it has a
contradiction on every branch, so it cannot be satisfied. By the corollary
above, T1 is not satisfiable. Thus, no interpretation which makes Γ true
can make A false. Therefore, any interpretation which makes Γ true also
makes A true. But this says exactly that A is a logical consequence of Γ.
Thus Γ |= A.
So far we have Γ ` A implies Γ |= A. In other words, we have correct-
ness.
Completeness
We need to show that Γ ` A whenever A is a logical consequence of Γ.
This is the completeness part. The proof also uses contradiction ! Assume
that Γ ` A is false. This means that no closed tree can be obtained if we
start with an initial tableau with Γ asserted and A denied. That means
that no closed tree can be obtained from this initial tableau no matter how
we apply the rules. Suppose we apply the rules systematically, as explained
in the previous section; this means that if any rule can be applied at some
node it is eventually applied, unless all the branches through the node are
closed. We do not get a closed tree. If the construction terminates, we
get a counterexample, and in this case A is not a logical consequence of Γ.
5.4. HOW MUCH USE IS THIS SYSTEM? 69
Suppose however that the construction never terminates, no matter what

we do. The construction builds a tree with at least one infinite branch.
Such an infinite branch actually describes a counterexample, I claim.
Suppose T1 ⇒ T2 ⇒ T3 ⇒ ....Tk ⇒ Tk+1 ⇒ ... is the sequence of trees
obtained by our non terminating systematic derivation. Let Tω be the union
of all these trees. Tω has at least one infinite open branch. Let α be such
a branch. The claim is that if the derivation has been done systematically,
that is if any operation which could be done on α is eventually done, then
α describes an interpretation, I, which satisfies Tω , and thus satisfies T1 ,
since T1 is actually the initial part of Tω . For the domain of I we take
the constant terms appearing in α, i.e. all the terms which do not have
variables in them. Let D be this domain. D is syntactic, a set of strings
of characters, but it is still quite acceptable as a mathematical domain.
Note that D can have compound terms as well as constants. In fact, D is
closed under application of the function symbols, since we were obliged to
try all possible substitutions in the rules. If f is a function symbol of arity
n, and t1 , ..., tn are in D, define the function f (t1 , ..., tn ) to be f (t1 , ..., tn ).
(Nothing could be simpler!) So far we have got a domain and defined the
functions. We only need to define the predicates. If p is a predicate symbol
of arity n define p(t1 , ..., tn ) to be true if and only if p(t1 , ..., tn ) appears
asserted of α. We now have an interpretation I. It can be shown that I
satisfies α.
Lemma 5.2 If B is asserted on α then |=I B.

If B is denied on α then |=I ¬B.
Proof. The proof is by induction on the number of logical operators in

B. For the basis of the induction, consider the case in which there are no
logical operators in B, i.e. B is atomic. So B is just a predicate symbol
applied to a list of terms. Note that I has been defined in order to verity
the statement of the lemma. Now for the induction step. We will need to
use complete induction. We therefore suppose that the lemma is true for
sentences B with less than k logical operators, and we suppose that B has
k logical operators. There are now seven cases, depending on the outermost
operator of B. They are all quite easy.
Case B has the form C → D. Assume B is asserted on α. Then, since
the construction is assumed to be systematic, either C is denied on α or D
is asserted. If C is denied, then, by induction hypothesis, |=I ¬C. If D is
asserted, then, by induction hypothesis, |=I D. In either case |=I (C → D).
Thus |=I B, which was to be proved.
The other cases you can do yourself.
2
So A is, again, not a logical consequence of Γ.

2
We mentioned earlier that there are a number of formal systems for proof
in predicate logic. The Gödel theorem is true of all of them. (Of course the
proofs are different.) So all these systems are, in a sense, equivalent.
Can we actually use a system such as semantic tableaux for deciding
whether or not a sentence is a logical consequence of a set of axioms?
5.4.2 Unsolvability and intractability

The bad news is:
Theorem 5.2 The problem of deciding whether or not an arbitrary sen-
tence A of predicate logic is logically valid is recursively unsolvable.
This means that there is no algorithm to decide whether or not |= A.
So obviously there can never be a computer program which, given a set
Γ of axioms and a sentence A, will decide whether or not Γ |= A.
Even though we have a complete and sound formal system for deduction
in predicate logic, we have to give up any hope of completely solving the
logical consequence problem. On the other hand, there may be special
versions of this problem which are of practical interest and which we can
solve in practice.
5.5 Problems
Problem 5.2 Give a definition for:
term t is free for variable X in formula A(X).
This should be done in such a way that it makes sense to substitute t
for X in A(X).
Problem 5.3 Find a counterexample to the following, or give a proof. In
each case, complete the semantic tableau construction, and state how many
branches there are in the completed tree. How many are closed? Verify that
each branch which is not closed defines a counterexample.
a) (((p → ¬q) → p) → q)
b) ((q → ¬p) → ((p → (¬q))))
c) ((p → q) → ((¬p) → (¬q)))
Problem 5.4 If we know that either A is true or B is true and we know
that either B is true or C is false, and we know that A and C can’t both be
true, then it must happen that A is true. Either give a formal proof of this,
or a counterexample.
5.5. PROBLEMS 71
Problem 5.5 Show whether or not the following set of statements is con-
sistent.
A and B and C implies that D is false unless E is true. E and not B
implies C. A and B are equivalent. C is true.
Problem 5.6 Prove this or construct a counterexample:

A and (B or C or D) is logically equivalent to (A and B) or (A and C)
or (A and D).
———————-
Either show that the following statements are logically valid, or provide
a counterexample.
Problem 5.7 Esmerelda is a duck and all ducks like ponds so Esmerelda
likes ponds.
Problem 5.8 If all burglars are barbers and all barbers are bakers then all
burglars are bakers.
Problem 5.9 If all burglars are barbers and no bakers are not barbers then
some bakers are not burglars.
Problem 5.10 If all burglars are barbers and some barbers are bakers and
some baker is not a burglar then some barber is not a burglar.
5.6 Satisfaction Games

We earlier saw that truth of a statement in an interpretation can been
understood in terms of a game between two people, a proponent and an
opponent.
In a similar way we can understand the semantic tableaux method in
terms of a game between two people a builder and a critic. Suppose we are
given an initial set of sentences Γ. We don’t have an interpretation. The
builder wants to build an interpretation in which all the sentences in Γ are
true. The critic wants this to fail. He hopes to confound the builder. When
there is a branch the builder is allowed to choose which branch to follow.
The builder also invents new names for the Skolem constants which have to
be introduced. Otherwise the critic (with malicious intent) chooses which
rule to apply next, and which substitutions to make. The original set of
sentences are not satisfiable if and only if the critic has a winning strategy.
Otherwise the builder has a winning strategy. But the builder wins if the
game goes on forever. Even teh best human player imaginable will make
mistakes in this very difficult game.
Example 5.4 Γ = {((∀X)(∀Y )(r(X, Y ) → r(Y, X)), (∃X)¬r(X, X), (∀X)(∀Y )(∀Z)((r(X, Y )∧
r(Y, Z)) → r(X, Z))}.
5.7 Revision Problems

Problem 5.11 The two statements (∀X)(∃Y )p(X, Y ), and (∃Y )(∀X)p(X, Y )
are not logically equivalent. Show that one of them implies the other using
semantic tableaux. Give a counterexample to show that the implication in
the other direction is not valid.
Consider the following three statements.
1. ((∀X)a(X) → (∀X)b(X))
2. (∃X)(∀Y )(a(X) → b(Y ))
3. (∀Y )(∃X)(a(X) → b(Y ))
Problem 5.12 Construct parse trees for each of the three statements above.
Show the scope of each quantifier.
Problem 5.13 Use semantic tableaux to show that the first statement im-
plies the second.
5.8. SOLUTIONS 73
Problem 5.14 Use semantic tableaux to show that the first statement im-
plies the third.
Problem 5.15 Show that the second statement implies the first.
Problem 5.16 Are the three statements logically equivalent?
Problem 5.17 Find prenex normal form of

((∃X)a(X) → (∃X)b(X))
Problem 5.18 In our prenex normal form algorithm, we proceeded by first

getting rid of implication, replacing A → B by (¬A ∨ B), and then moving
negations inside and then moving quantifiers outside of conjunctions and
disjunctions. Show how to modify the algorithm, leaving implications in
place, and moving quantifiers outside of implications.
5.8 Solutions
1. To show
((∀X)a(X) → (∀X)b(X)) |= (∃X)(∀Y )(a(X) → b(Y ))
1) (Forall X)a(X) -->(Forall X) b(X)

2) not(Exists X)(Forall Y)(a(X) --> b(Y))
/ \
3) not (Forall X) a(X) from 1 3) (Forall X) b(X) from 1

| |
4) not a(tau) from 3, Skolem tau 4) not (Forall Y)
(a(rho) -->b(Y)) sub 2
|
|
5) not (Forall Y)(a(tau) --> b(Y)) subs in 2 5) not(a(rho) --> b(alph))
from 4, Skolem alph
| |
6) not(a(tau) --> b(sig)) from 5, Skolem sig 6) a(rho) from 5
| |
7) a(tau), from 6, contradiction 7) not b(alph) from 5
| |
8) not b(sig) from 6 8) b(alph) from 3, subs
Contradiction Contradiction
The tree is closed. So there is no counterexample. Thus

((∀X)a(X) → (∀X)b(X)) |= (∃X)(∀Y )(a(X) → b(Y ))
5.9 Summary
We have developed a complete and sound proof system for predicate logic.
This is based on proof by contradiction. Given a set of axioms Γ and a
sentence A, we attempt systematically to construct a counterexample, i.e.
an interpretation in which A is false although Γ is true.
The data structures in this semantic tableau system are trees labelled
with assertions or denials of formulae. The transformations are rules which
extend the trees.
Chapter 6
Other formal deductive

systems for first order
logic
6.1 Deduction by Resolution and Unification

Suppose Γ is a set of sentences and A is a sentence. If A is not a logical
consequence of Γ then there is an interpretation in which Γ is true and A
is false. Thus Γ |= A if and only if Γ, ¬A is not satisfiable. So the logical
consequence problem reduces to the problem of deciding whether or not a
set of formulae is satisfiable.
Let Γ be any set of sentences, and let c(Γ) be the clausal form of Γ. Γ
is satisfiable if and only if c(Γ) is satisfiable. (To say that c(Γ) is satisfi-
able means that there is some interpretation in which all the clauses are
universally true.) So the logical consequence problem reduces to deciding
whether or not a clausal form is satisfiable.
Resolution is a deduction rule for clauses. Suppose we have two clauses:
Γ1 → ∆ 1 , C
C, Γ2 → ∆2
with some formula C appearing on the left of one and on the right of
the other. Assume that Γ1 and Γ2 are true. Then either the disjunction of
∆1 is true or C is true. If C is true, then the disjunction of ∆2 must be
true. So we get the new clause
Γ1 , Γ2 → ∆1 , ∆2
and we say that the new clause is obtained by resolution of the original
75
76CHAPTER 6. OTHER FORMAL DEDUCTIVE SYSTEMS FOR FIRST ORDER LOGIC
two clauses. Note that although the new clause may be longer, it has one
less atomic formula. We start with a pool of clauses in the original clausal
form, and we add to the pool by using resolution. If we eventually get to
the contradictory clause
→
which has empty left and right hand sides, this implies that the original
clausal form was contradictory, and thus not satisfiable.
We may also use substitution on clauses to obtain new clauses. In
order to apply resolution we need pairs of clauses in which some formula
C appears on the left of one and on the right of the other. The unification
algorithm can be used to find substitutions which create these matches.
Resolution and unification can be used together to make a complete and
correct deductive system for clausal forms. This is especially easy to apply
in the special case in which all the clauses have exactly one atomic formula
on the right hand side. This special case is the basis of the programming
language prolog, which will be discussed in detail later.
6.2 The Sequent Calculus

In this section we will briefly give yet another complete and correct formal
deductive system for predicate logic. This other system is called the sequent
calculus. In the sequent calculus we start with a set of axioms which are
obviously valid. The axioms are also considered to be theorems. We obtain
new theorems by applying certain rules of inference.
The basic statement in this system is the sequent. A sequent has the
form:
A1 , ..., An ` B1 , ..., Bm
where A1 , ..., An , B1 , ..., Bm are formulae in some first order language.
We will interpret such a statement as meaning:
A1 ∧ ... ∧ An |= B1 ∨ ... ∨ Bm
That is, the disjunction of the right hand side is a logical consequence
of the conjunction of the left hand side. In general, we will consider the left
and right hand sides of a sequent as sets; that is, the order of the formulae
will not matter. We will consider an empty set on the left hand side as
Truth, and the empty set on the right hand side as Falsity. Thus
` A, B, C
means that A ∨ B ∨ C is logically valid. On the other hand
A, B, C `
means that A ∧ B ∧ C is contradictory; that is, it is not possible to
satisfy A, B, C simultaneously.
6.2. THE SEQUENT CALCULUS 77
A sequent is like a clause except that quantifiers and negation are al-
lowed.
A sequent is obviously valid if the same formula appears on both left
and right hand sides.
We will use Γ and ∆ as variables for sets of formulae.
In order to make this system similar to the semantic tableaux system,
we will assume that all the formulae in sequents are sentences. That is, no
free variables are allowed.
Definition 6.1 A sequent of the form

Γ, A ` ∆, A
is an axiom of the sequent calculus.
The rules of inference will allow us to prove some sequents from others.
We have two rules for negation:
One of these rules is:
Γ ` ∆, A
————————–negation left
¬A, Γ ` ∆
The other rule for negation is:
A, Γ ` ∆
————————negation right
Γ ` ∆, ¬A
Please check that negation left and negation right are correct. Do you
get a strange feeling of familiarity when you look at these? What rules
should we have for implication?
A, Γ ` ∆, B
———————–implication right
Γ ` ∆, A → B
There is also an implication left rule, but it needs two sequents as
premises.
(Γ ` ∆, A), (B, Γ ` ∆)
——————————————————-implication left
A → B, Γ ` ∆
Please check again that these rules are correct.
Perhaps you can see the connection with semantic tableaux. Consider
a branch of a semantic tableau. Let ∆n be a subset of the formulae which
occur on the branch and which begin with negation. Let Γ be the other
formulae which occur on the branch. Obtain ∆ from ∆n by removing the
initial negation. Now write
Γ ` ∆.
78CHAPTER 6. OTHER FORMAL DEDUCTIVE SYSTEMS FOR FIRST ORDER LOGIC
This translates branches into sequents. The set Γ are the statements
asserted on the branch and the ∆ are the statements which are denied. A
branch is impossible to satisfy if and only if the corresponding sequent is
valid.
Notice that a branch is closed if and only if it contains a contradiction
if and only if it translates into an axiom in the sequent calculus.
You can now see: the rules for the sequent calculus are obtained just by
turning the semantic tableau rules upside down. A semantic tableau proof
is just a sequent proof turned upside down. As an exercise, you should
make sure that you can write down all the rules of inference for the sequent
calculus. Look at the quantifiers!
Please note that if you actually want to find a proof of a given sequent in
the sequent calculus, the best way to do this is usually to tackle the problem
backwards using semantic tableaux. Once you get a closed tableau, you turn
this upside down to get a sequent proof.
As an example, we could prove A ` (B → A) in the semantic tableau
system, and then obtain the following two step proof in the sequent calculus:
———–
A, B, ` (B → A), A Axiom
A ` (B → A) implication: right
————————–
We could also have used the simpler axiom:
A, B ` A
6.3 Other deductive systems

There are many other correct and complete deductive systems for first order
logic. An important example is natural deduction, which is described in is
the book by Huth and Ryan mentioned in the references.
Chapter 7
A Formalisation of
Mathematics: ZF set
theory
There have been several attempts to formalise mathematics. The most

successful of these is Zermelo Fraenkel set theory. Anyone who has studied
pure or discrete mathematics will be aware of the existence of this system
as a mathematical background and framework. It has had a very strong
influence on the development of mathematics. A rough description of ZF
set theory is given below in a few paragraphs. This famous formal system
consists of a formal language, LZF , together with a set of axioms and some
formal rules of deduction.
The language LZF consists of all statements, called formulae, which
can be built up from equality statements, such as X = Y , and set member-
ship statements, such as X ∈ Y , using the logical operators not, or, and,
implies, iff, for all and there exists, which are written ¬, ∨, ∧, →, ↔, ∀, ∃,
respectively. For example, to say that X = Y if and only if every element
of X is an element of Y and vica versa, we could write, in LZF :
(X = Y ↔ (∀W )(W ∈ X ↔ W ∈ Y )).
The axioms for ZF set theory say, essentially, that certain basic sets
exist, and that sets are closed under certain operations. Here is a list of the
axioms, with some redundancies.
1. Equality axiom. For sets A and B, A = B if and only if (∀X)(X ∈
A ↔ X ∈ B).
2. Singleton axiom. For any A , there is a set {A}, called singleton A,
79
80CHAPTER 7. A FORMALISATION OF MATHEMATICS: ZF SET THEORY
which is the set whose only element is A.
3. Union axiom. If A and B are sets, so also is A ∪ B, the union of A

and B. It is also assumed that if F is a set whose elements are also
sets, then there is a set ∪F = {Y : (∃X)(X ∈ F ∧ Y ∈ X}. In other
words, ∪F is the union of all the elements of F .
4. Intersection axiom. If A and B are sets, so also is A ∩ B, the intersec-

tion of A and B. It is also assumed that if F is a set whose elements
are also sets, then there is a set ∩F = {Y : (∀X)(X ∈ F → Y ∈ X).
5. Set difference axiom. If A and B are sets, so also is A/B = {X : X ∈

A ∧ ¬X ∈ B}. This is the set difference of A and B.
6. Ordered pair axiom. For any A and B, there is an ordered pair (A, B),
with first element A and second element B. (A, B) = (C, D) if and
only if A = C and B = D.
7. Cartesian product axiom. If A and B are sets, so is A × B , the

Cartesian product of A and B, i.e. the set of ordered pairs (a, b),
with a in A and b in B.
8. Power set axiom. If A is a set, so is P (A), the set of all subsets of a

set A. This is called the power set of A.
9. Function space axiom. If A and B are sets, so is A → B, the set of

functions from A to B.
10. Comprehension axiom: If A is a set and C(X) is a condition on X

which we can write in LZF , then we can define {X : X ∈ A ∧ C(X)}.
The comprehension axiom picks out that subset of A for which C(X)
is true.
11. Replacement axiom: It is also assumed that if we have succeeded,

within LZF , in defining a function f (x) on a set A, then we can
construct a new set as {f (x) : x ∈ A}, i.e. the image of f applied to
A.
12. Axiom of choice: If F is a set whose elements are disjoint non empty
sets, there exists a set choice(F ) which has exactly one element in
common with each element of F .
13. There is an empty set, denoted ∅, which has no elements. Also, there
exists the set of natural numbers N = {0, 1, 2, ...}.
81
The attentive reader may notice that some of the axioms, as given above
can be proved from others. For example, the existence of A ∩ B can be
proved from the axiom of comprehension, since A ∩ B = {X : X ∈ A ∧ X ∈
B}. And in fact the comprehension axiom itself can be proved from the
axiom of replacement. I have stated the axioms with these redundancies
since I thought they might help the reader to see what was happening. The
main sense of the axioms is that we can build up complex sets from simpler
ones using common operations.
The axioms of ZF set theory are written in our special language, LZF .
For example, the axiom which says that for any set A there exists a set,
singleton A, whose only element is A, would be written as follows:
(∀A)(∃B)(∀W )(W ∈ B ↔ W = A)
In order to write the axioms entirely in LZF , rather than in the mixture
of LZF and informal mathematical English used above, we need to define
functions, ordered pairs, and the natural numbers in terms of LZF . That
is to say, we need to express these ideas in terms of set membership and
equality.
The definition of function is especially important.
Definition 7.1 A set f is a function if f is a set of ordered pairs such that

whenever (X, Y ) ∈ f and (X, Z) ∈ f then Y = Z.
In other words a function is defined to be a set of ordered pairs with

the property that the second element is determined by the first element.
We will write f : A → B and say that f is a function with domain A
and codomain B if A is the set of first elements of f and if all the second
elements of f are in B.
Roughly speaking, ZF set theory considers functions not as processes
but as look-up tables. For example, the squaring function on the natural
numbers is the set
{(n, n2 ) : n ∈ {N}}
In order to write the above axioms in LZF , we also need to represent
ordered pairs in general and the natural numbers in particular as sets.
There are a number of ways of doing this. For example, we could define
the ordered pair (X, Y ) to be the same as the set {{X}, {X, Y }}.
To represent the natural numbers we could identify 0 with ∅, and n + 1
with n ∪ {n}. Then N is the smallest set so that ∅ ∈ N and whenever
X ∈ N so is X ∪ {X}.
In the usual version of ZF set theory all the above representations are
used.
From the natural numbers, the axioms allow us to construct the set Z
of integers, the set Q of rational numbers, and the set R of real numbers.
The rules of deduction are, of course, formal rules, which can be applied
mechanically. The theorems of ZF set theory are the set of formulae which
can be obtained by starting with the axioms and applying the rules of
inference.
The original specification for this formal system was quite amazing. It
was that the theorems of this system should contain no contradictions,
and should contain the translations into LZF of all the currently accepted
theorems of mathematics, together with all the translations into LZF of all
theorems of mathematics which may become accepted in the future. So ZF
set theory is an attempt to characterise the apparently ultimate infinite
object of mathematical reality.
The Occam’s razor principle was applied very seriously by the develop-
ers of ZF set theory. Since every statement in LZF is built up from set
membership and equality, and the logical operations ∧, ∨, ¬, →, ↔, ∀, ∃, all
mathematical ideas must also be constructed from these elements.
In order to get simplicity, some aspects of naturalness were sacrificed.
For example, in order to keep the basic elements minimal, the natural
numbers are built up from the empty set in a way which may seem arbitrary:
0 is ∅, 1 is {O}, 2 is {O, {O}}, and, in general, n + 1 is n ∪ {n}.
So in ZF set theory, The natural numbers is
{O, {O}, {O, {O}}, ...}
which does not seem friendly or necessary. We do lose something if we
adopt a nasty notation, such as this. A good notation encourages correct
and fluent thinking.
Some of these problems of notation can be overcome without much
difficulty, by extending LZF , adding names of important objects and oper-
ations. So we can define 0 to mean ∅, 1 to mean {O}, and so on. Similarly,
without serious difficulty, we can extend the language LZF to include fa-
miliar notations such as {X}, X ∪ Y, X ∩ Y, X ⊆ Y , etc. The point is that
these notations can be unambiguously defined in the original LZF . For
example, we can define: X ⊆ Y to mean (∀W )(W ∈ X → W ∈ Y ).
There is a more serious problem about the definition of functions in ZF
set theory.
The designers of this system restricted themselves to what could be
said with set membership, equality, and the logical operators. Once this
restriction had been accepted, they had to explain the notion of function
in these terms.
The problem with defining a function as a kind of set is that, intuitively,
functions are active and dynamic, and sets are static. However Occam’s
razor was given priority over intuition in this case. Consequently, a function
in ZF set theory is defined to be a set of ordered pairs, corresponding to
the graph of the intuitive function, as described above.
83
The ZF notion of function is useful but not entirely satisfactory for

computing. From the point of view of computing, it might be worth having
a more complicated idea which was closer to intuition. The lambda calculus,
discussed in another course, is an attempt to supply such an idea.
Unfortunately, the completeness and correctness status of ZF set theory
is also less than ideal. We do not know whether or not ZF set theory is
consistent. So 0 = 1 might be a theorem, for all we know.
We do know, due to the wonderful work of K. Gödel, that ZF set theory
can not prove itself consistent, unless it is actually inconsistent. So either
there is a mathematical truth which is not a theorem, or correctness fails.
This is discussed later.
Nevertheless, ZF set theory is a great accomplishment. It does seem
successfully to formalise almost all of contemporary mathematics. It gives
us a universally recognised language for mathematics and a universally
recognised standard for rigorous proof. If two mathematicians disagree
about the meaning of a term or about the validity of a proof, they can,
in principle, resolve their differences by a process, called formalisation, of
translating their ideas into ZF set theory. The existence of the ZF formal-
ism is one reason for the great progress which has been made in mathematics
in the last few generations.
A dialect of LZF , called the Z specification language, can also be used
to write unambiguous specifications for computer programs. Since it is
formal, the Z specification language can also be read, transformed, and
processed by computer.
Chapter 8
Gödel Incompleteness
theorems
We previously defined LN , the language of first order arithmetic, and de-

scribed an infinite axiom set, Γ, which is called the Peano postulates. A
sentence, S, of LN is provable from the Peano postulates if
Γ`S
We are really interested in the truths in the standard interpretation N.
We have Γ ` S implies Γ |= S and this implies |=N S.
Theorem 8.1 (Gödel Incompleteness theorem for Arithmetic.) There is

a sentence S which is true in the standard interpretation but cannot be
proved from the Peano postulates.
Here is a sketch of the proof.

The main idea is to find a way to look at formulae of LN so that they
can be understood as making comments about their own provability.
Gödel begins by showing that each formula, A of LN can be coded as a
natural number n = g(A). The number n is now called the Gödel number
of the formula A. Similarly, we can find Gödel numbers for proofs from the
Peano postulates. This means that we can find a formula P r(X, Y ) with
two variables, which is true for natural numbers X and Y if and only if X is
the Gödel number of a proof from the Peano postulates of a formula which
has the Gödel number Y . Please notice that P r(X, Y ) is an actual formula
of LN , and it says something about the natural numbers X and Y in terms
of addition, multiplication, equality, and the usual logical operators. But
we know that it has another interpretation also, because of the way it was
85
86 CHAPTER 8. GÖDEL INCOMPLETENESS THEOREMS
constructed. This one piece of syntax with two different meanings is typical
Gödelian lateral thinking.
We can say that Y is the number of a provable formula by (∃X)P r(X, Y ).
We can now find a formula diag(X) which says that X is the Gödel
number of a formula A(Z) with one free variable and if n = g(A(Z) then
A(n) is not provable from the Peano postulates. Now let n = g(diag(X)).
We see that diag(n) is a sentence of LN . One of its meanings is that its
own self is not provable. diag(n) is either true or false in the standard
interpretation. If it is false, it is also provable from the Peano postulates.
But the Peano postulates are all true in the standard interpretation, and
deduction preserves truth. Therefore |=N diag(n). Also diag(n) is not
provable from the Peano postulates.
Remarks:
These results also apply to Zermelo Fraenkel set theory. In fact there is
no recursive axiomatisation which is complete for any extension of Peano
arithmetic.
It may be called common sense that human knowledge is limited. The
incompleteness theorems of Gödel prove this rigorously.
Chapter 9
Logic programming (
prolog)
What sort of computational use can we make of the semantic tableau idea?
9.1 Headed Horn clauses

Suppose we have some finite set of axioms Γ, and we wish to somehow
compute the logical consequences of these. Note that we are not working
in the sequent calculus at this point. Γ is just a set of formulae; in fact we
will assume that the axioms in Γ are all sentences. It seems reasonable to
begin by putting Γ in clausal form.
Suppose we have done this. In what cases is it feasible to use the se-
mantic tableau method to find the consequences of Γ? To put this question
in another way, when the semantic tableau method goes wrong, how does
it do this?
The problem with the semantic tableau method is that it creates a lot
of branches. In some cases these branches proliferate until they become
unmanageable. In order to implement a theorem prover for predicate logic,
it seems necessary to look for special cases in which we hope the compu-
tation will not get too complicated. (Or at least we can regard this as a
reasonable first step.)
Definition 9.1 A headed Horn clause is one of the form:
(A1 ∧ A2 ∧ ... ∧ Ak ) → B
Headed Horn clauses are especially simple because the conclusion is a
single statement, rather than a list of alternatives.
87
88 CHAPTER 9. LOGIC PROGRAMMING ( PROLOG)
Example 9.1 We will write a headed Horn clause (A1∧A2∧...∧An) → B

in the form B if A1, A2, ... An.
Suppose our axioms are (p ∧ r) → q, r → p, and r. Can we prove q?.
We can proceed as follows:
(1) q if p,r
|
(2) p if r
|
(3) r
|
(4) not q
/ \
(5) q (5)not( p,r)

contr / \
(6) not p (6) not r, contr

/ \
(7, from 2) p (7) not r contr

contr
We take the not q of the original tableau as a goal. We match this with
one of the conclusions of the clauses which are asserted, in this case with
the conclusion of clause (1). The premises of clause (1) then become sub
goals. So we have subgoals p and r. Proceeding recursively, we match p
with one of the conclusions of the clauses. In this case the match is with
the conclusion of clause (2). We generate another subgoal, namely r. This
matches with one of our original premises, so one branch gets closed. The
next branch gets closed in the same way.
It seems that in the case when our axioms are Headed Horn clauses, we
have a reasonable search strategy. That is, we seem to know what to do
next. The main idea of prolog is to try to exploit this.
From now on, we will suppose Γ is a list of headed Horn clauses which
are universally quantified, and A is an existentially quantified conjunction
of atomic formulae.
We are trying to decide Γ |= A.
9.1. HEADED HORN CLAUSES 89
From now on we will also omit universal quantifiers of statements as-

serted in our trees, and we will omit existential quantifiers on statements
which are denied. (So free variables in asserted formulae are tacitly as-
sumed to be universally quantified, and free variables ion formulae which
are denied are tacitly assumed to be existentially quantified. )
Example 9.2 Suppose Γ is

{(∀X)(∀Y )(duck(X) ∧ pond(Y )) → likes(X, Y )
duck(esmerelda), pond(estero), pond(emeryville)},
and A is (∃Z)likes(Z, estero).
We would write the initial tree as:
(1) likes (X,Y if duck(X), pond(Y)
|
(2) duck (esmerelda)
|
(3)pond(estero)
|
(4) pond(emeryville)
(5) not likes(Z, estero)
Our convention that constants begin with lower case and variables begin
with upper case is very useful here, since it allows us to distinguish constants
and variables.
We wish to decide whether or not Γ |= A. We will need to look for good

substitutions which may lead to a closed tree.
The prolog strategy is the following.
Suppose Γ is C1, C2,..., Ck, where each Ci is a headed Horn clause.
1. Put Γ, the list of headed Horn clauses asserted and ¬A, a conjunction
of atomic formulae denied in the initial tree.
2. If A is the conjunction of n atomic formulae, split ¬A into n branches.
The i th branch has Γ asserted and has the i th conjunct of A denied.
3. Pick the leftmost open branch. Let’s say this has A1 denied.
4. Scan down the list Γ. For each Ci in Γ, try to find a substitution

which unifies A1 and the conclusion of Ci.
5. If we get to the end of Γ without success, return FAIL.
6. Suppose we find a substitution α which unifies A1 and the conclusion
of Ci. Let Ci be:
B if R1, R2, ..., Ra
so Bα= A1 α. Split Ci in the subteee below A1 so that we get:
() A1
/ \
not B not(R1 , R2 , ... Ra)
Make the substitution α so that the left hand branch of this closes.
On the right hand branch we have a new denied conjunction of atomic
formulae
not (R1, R2, ... , Ra) α
7. Continue recursively with the new tree, after application of substitu-
tion α.
8. If we do not get closure of the new subtree below A1, backtrack, delete
this subtree, undo the substitution α and continue scanning Γ below
Ci.
9. If we do get closure, report the final substitution of terms for variables.
9.2 Predicates, facts, constants, variables in

prolog
In the following, a “character” is either a letter or a digit.
A constant in prolog is either a number, such as “776”, or a string of
characters beginning with a lower case letter. So, for example, “jane” is a
constant, but “Jane” is not a constant.
A variable in prolog is a string of characters beginning with an upper
case letter.
9.2. PREDICATES, FACTS, CONSTANTS, VARIABLES IN PROLOG91
Prolog also has function and predicate symbols, just like first order
languages. We can use any string of characters starting with a lower case
letter as the name for a function or a predicate. As in any term language,
we can have complex terms built up out of simpler ones. So for example
f (g(X, hat), Y, Z, h(p, q))
is a possible prolog term. Prolog also has some built in function symbols.
Lists are especially important. In prolog, lists are written surrounded with
square brackets and separated by commas. So, for example,
[cat,dog,horse]
is a list of three items. Lists can also have other lists as components.
So
[hat,[big,yellow],got]
is a list whose first component is a constant, and whose second compo-
nent is a list. Lists are terms and lists can have any terms as components.
There is am empty list, written [ ].
Predicate names in prolog are strings of characters beginning with lower
case letters.
You should try to keep in mind the distinction between predicates, which
are semantic things, functions whose codomain is {T, F }, and predicate
names, which are defined by syntax.
Predicate expressions are usually written as predicate names followed
by a list of terms enclosed in brackets. For example
red(cup)
says that some particular object, called “cup” is red.
There is no systematic way to distinguish function names from predicate
names in prolog.
Prolog has a few built in predicates, whose names are already deter-
mined. An important built in predicate is equality, written
X = Y.
Inequality, written
X \= Y.
is also built in.

As long as we avoid names of built in predicates, we can invent predicate
names to suit ourselves. Of course it is necessary to use whatever names
we invent in a consistent manner. Suppose, for example, we wish to write
some axioms for motherhood. We might begin by asking ourselves: what
is the arity of motherhood? We seem to need a predicate of arity 2, which
says that one person is the mother of another. We could decide to write
this as
qqqqqzsz(A,B).
However, this will be hard to remember, and it seems better to use a

more natural name. We could call our predicate
mother(A,B).
This is clearly meant to say either that A is the mother of B, or vica

versa. Obviously, we have to determine which we mean. So we have to
invent some convention to determine, in our own minds, the order of the
arguments. We might decide, for example, that mother(A,B) should mean
that A is the mother of B. Having set up this convention, we have to hold
to it.
A fact in prolog is a predicate expression in which all the terms are
constant. For example
mother(louise, mabel)
is a fact. To say that this is a fact does not, in this context, imply that
we think it is true. It may be true or it may be false. A fact, in this context,
is an expression which has the form
predicate-name(constant-term,..., constant-term).
To give another example,
likes(joe, alice)
which is intended to say that joe likes alice, is a fact.

A predicate of arity zero is written as
p
9.3. PROGRAMS IN PROLOG 93
9.3 Programs in prolog

A prolog program may be regarded as a set of axioms for some field of
knowledge. When the program is consulted we ask whether some statement
is a logical consequence of the axioms. If the statement has variables in it
we ask for assignments of values to the variables so that the statement is a
logical consequence of the axioms.
A prolog program consists of a list of statements, with full stops at the
end of each. Example:
——————-
mortal(X) :- human(X).
featherless(socrates).
bipedal(socrates).
animal(socrates).
human(X) :- featherless(X), bipedal(X), animal(X).
——————–
As mentioned above, terms starting with upper case letters, such as X,

are variables, and other names such as socrates are constants. human(X)
is a predicate expression with variable X, meaning X is human. Note that
predicate names, such as human, mortal, featherless, should start with
lower case letters.
In prolog, the sign :- is interpreted as if. So the first line in the example
program above means X is mortal if X is human, for all X. The next three
lines say that socrates is a featherless bipedal animal. The comma , is
interpreted as and so the last line says that X is human if X is a featherless
bipedal animal, for all X.
Among the five statements in the above prolog program, the statements
on lines 2,3,4 are facts. The other statements are called rules. A rule in
prolog, in general, is a statement of the form
p :- q1,q2,..., qn.
where
p, q1, q2,..., qn
are predicate expressions. The meaning of this rule is that p is true when-
ever all of q1,..., qn are true.
Statements in prolog are only of two types, either facts or rules. Prolog
statements are also called clauses.
Both rules and facts are headed Horn clauses.
In a rule, such as
mortal(X) :- human(X).
the left hand side is called the conclusion. A rule is supposed to state a
truth and also show a possible way to prove the conclusion. So, for example,
human(X):- featherless(X), bipedal(X), animal(X).
expresses the truth that any featherless bipedal animal is human, and
also tells us that if we have an X and want to show
human(X)
we can do this by checking
featherless(X), bipedal(X), animal(X).
Note that different occurrences of a variable in a single rule must refer

to the same object. If we have the above rule and we want to prove
9.3. PROGRAMS IN PROLOG 95
human(fred).
we need to check
featherless(fred), bipedal(fred), animal(fred).
On the other hand, occurrences of the same variable in different rules are
not linked at all. The program given above would have the same meaning
if the first line were changed to
mortal(Y):- human(Y).
Exercise: You can create a prolog program with any editor. Try this.
Once a program has been written, we may wish to ask questions about
its logical consequences. This process is called consultation and is done
with a prolog interpreter and/or compiler.
The reader should at this point discover how to get access to some
version of prolog.
At Bath University on the BUCS machines, prolog can be run by typing
pl.
Suppose you have written a prolog program. Give whatever local com-
mand is necessary to run prolog.
Once prolog is running, type
[file].
if file is the name of the program you have written. You can then ask
questions, and you should get logical consequences of the facts and rules
you have given. For example if you ask mortal(socrates). after reading in
the above example, it should, (after thinking for a while), say yes. If you
ask
mortal(X).
it should eventually satisfy this with
X=socrates.
To get out of prolog type
halt.
Note that a full stop is necessary after each statement.

Now, try to create a program and consult it.
9.3.1 Comments in Prolog

For benefit of a human reader, comments may be added to prolog code as
follows:
/* This is a comment which might continue over

several lines */
% This is also a comment, but on one line.
9.4 How prolog works

Define the head of a headed Horn clause to be the conclusion, and define
the tail to be the list of premises.
So, for example, in
perfect :- beautiful, surprising, useful, at-right-time.
the head is
perfect
and the tail is
beautiful, surprising, useful, at-right-time.
In section 9.1, we explained how prolog works in terms of semantic
tableaux. We can also understand prolog’s working as a special case of
resolution and unification. We can also understand how prolog works in
terms of a technique for satisfying goals, given a program Γ. Each clause
in Γ is a headed Horn clause. Such a clause,
A : −B1 , . . . , Bn
means that if our goal is to staisfy the head A, we can do this by
satisfying the tail B1 , . . . , Bn . (There may, of course, be other ways to
9.5. PROGRAMS WITH JUST FACTS 97
satisfy A.) So from goal A we can get subgoals B1 , . . . , Bn , and we continue

recursively.
Suppose G is our current list of goals. We try to find a substitution α
so that Γ |= Gα . We deal with each goal in G in turn. Suppose the first
goal is G1 . We try to match G1 with the head of a clause in Γ, starting
at the beginning of Γ. Matching is done by unification. Suppose the first
match is found a line k of Γ. Suppose this is
A : −B1 , . . . , Bn
The unification algorithm (assumed to succeed in this case) gives us
a substitution α which is a most general unifier between G1 and A. So
we take (B1 )α , . . . , (Bn )α as subgoals. We remember (in case we have to
backtrack later) that substitution α was applied at line k. We continue
recursively, attempting to satisfy the subgoals.
Suppose this succeeds, with final substitution γ, (which includes α). So
Γ |= (G1 )γ . We now apply γ to the tail of G and continue recursively. If
the tail is empty, we report the final substitution γ so that Γ |= Gγ .
Suppose however that we fail to satisfy subgoals (B1 , . . . , Bn )α . In this
case we backtrack. This means that we undo substitution α and return to
line k of Γ. We look below line k for another match with goal G1 . If none
is found, we report “no”. This means that Γ does not logically imply the
existence of a way to satisfy G.
9.5 Programs with just facts

According to what was said above, a prolog program is a list of statements,
which are either facts or rules. In particular, a list of facts is a prolog
program. Suppose, for example, we had a list of facts of the form
criminal(X), meaning that X is a criminal
locate(X,City, Date), meaning that X is known to have been in City on
Date, and
associate(X,Y), meaning that persons X and Y are associates.
We could just list all our facts in a program. For example,
—————–
criminal(boxcarjoe).
criminal(bigred).
criminal(eaglehat).
locate(bigred,ny,july3).
locate(bigred,sf,june4).
associate(bigred,boxcarjoe).
locate(boxcarjoe,milan,june5).
etc.
——————–
We store this in some file. We then call prolog. We read in the file with
[’filename’].
We can then ask questions. If we ask, for example,
criminal(bigred).
we get the answer
yes.
We can also ask questions with variables, such as
criminal(X).
In this case we would get the first criminal
X = boxcarjoe.
In general, given a query with variables in it, prolog attempts to find

values of the variables which make the query true. If we want another
example, we type
and we get the next crook, bigred. If we type ; again we get the next
one. When the list is exhausted, we get answer
9.6. BACKTRACKING 99
no.
We can also ask compound queries. Suppose, for example we are con-
cerned to discover a criminal associate of bigred who was in milan on some
date.
We ask
associate(bigred,X), criminal(X), locate(X,milan,Date).
and we will get an answer such as
X=boxcarjoe, Date=june5.
9.6 Backtracking
When we were working with semantic tableaux, we made substitutions and
attempted to close branches of our trees. If a substitution did not result
in closure, we tried another substitution; we never deleted anything. But
prolog does delete the results of unsuccessful substitutions.
Suppose the current goal of prolog is query q(X1 , ..., Xn ). prolog scans
down its program trying to unify the query with one of its facts or rules.
Suppose the first match is with the head of a rule, with unifier σ, at line
k of the program. The unifier σ is applied to the tail of the rule, and
the conditions in this tail become new goals. Prolog continues recursively,
starting with the first subgoal in the tail. If this attempt ultimately fails,
prolog backtracks. This means that it discards the substitution σ and
resumes its attempt to match the goal q(X1 , ..., Xn ) just below the last
match, i.e. at line k + 1 in the program.
Consider, for example, the query
criminal(X), locate(X,ny,july3).
with the program given above. This will first try the substitution [X :=
boxcarjoe]; but this will not succeed, and backtracking will ensue.
9.7 Warning: problems with prolog

It would be nice if prolog really implemented predicate logic. However, this
is not the case. Prolog sometimes works correctly and sometimes does not.
Since we have a definition of logical consequence, we know exactly what
it means to say that prolog works correctly. We would hope that prolog
would show that A is a consequence of Γ if and only if Γ |= A, as long as Γ
is a list of headed Horn clauses, and A is atomic. But no such luck.
Prolog uses a certain strategy for trying to get closure of its proof trees.
This enables it to finish quickly in some cases, but causes problems in other
cases. For example, suppose we are trying to decide whether or not
(p → p) |= p.
We could do this by semantic tableau. We assert (p → p) and deny p ;
we get one split, and stop. One branch is closed, the other is open, and by
setting p to FALSE we get a counterexample.
What does prolog do?
It begins with
(1) p:- p.
|
(2) not p.
The goal is p. This is matched with the head of p:-p, and we obtain
a subgoal of p. Unfortunately, prolog now continues recursively, and so it
never terminates. It comes to no conclusion. Prolog is unable to recognise
that it is considering a goal which it has considered previously.
Obviously there are many ways in which prolog can tie itself into such
infinite loops, and it frequently does so. In predicate logic, we often consider
symmetric relations. For example,
married(X,Y):- married(Y,X).
To write such a rule in prolog would be to invite non termination.
To give another example, we could define fatherhood in terms of male-
ness and parenthood.
father(X,Y):- male(X), parent(X,Y).
Of course it is also true that fatherhood implies parenthood. We would
include this if we were writing axioms of fatherhood in predicate logic. But
in a prolog program, we would be unwise to have both
father(X,Y):- male(X), parent(X,Y).
9.8. HOW TO WRITE SIMPLE PROLOG 101
and
parent(X,Y):- father(X,Y).
since this would be likely to cause non termination.
Such difficulties are very annoying. There are ways of alleviating such
problems, but the alleviations are almost as irritating as the original prob-
lems. It seems to me that it is fair to say that prolog is a reasonable
first step toward implementing predicate logic, but that, at present, it is
extremely limited.
9.8 How to write simple prolog

Suppose we have a situation which we wish to describe in prolog. We first
decide what are the important predicates in the situation. We then think
about the predicates to try to see which ones can be defined in terms of
the others. We should arrange the predicates in a hierarchy of complexity,
with complex predicates defined in terms of simpler ones. Usually there
are many different ways to do this! If possible, all the facts should be
entered using the simplest predicates. If possible, the simplest predicates
should be logically independent, so that no single piece of information is
held in two different ways. These are desirable features, which can’t all be
realized in every case. So writing a prolog program is a problem in design,
and different people will have different solutions, and some solutions will
be simpler and more elegant than others. There may not exist any perfect
solution.
In general, it is best to put the facts before the rules in a prolog program.
It is best to put definitions of relatively simple predicates before definitions
of relatively complex predicates. In other words, the order of the program
should reflect, if possible, the hierarchy of complexity which you are using.
This usually makes the program easier to read for a human being. And in
some cases, as we will see, it may also improve the computational behaviour
of the program.
9.9 Examples
Imagine a small community where everyone is related to everyone else in
several different ways. We wish to express the relationships in a prolog
program.
Some of the predicates we might be concerned with here are: par-
ent(X,Y), father(X,Y), mother(X,Y), male(X), female(X), sister(X,Y), grand-
father(X,Y), aunt(X,Y), cousin(X,Y). Evidently there are many logical re-
lationships, for example:
———————–
parent(X,Y) :- mother(X,Y).
parent(X,Y) :- father(X,Y).
grandfather(X,Y) :- father(X,Z),
parent(Z,Y).
————————–
The problem is to decide which are the basic predicates, and then to
build up definitions of the more complex predicates in terms of the simpler
ones.
To begin with, suppose we took the above definition of parent in terms
of father and mother, and of grandfather in terms of father and parent.
Suppose we added some facts to the program given above.
———————-
father(jud,rubel).
father(rubel,jubel).
parent(jubel,marx).
parent(rubel,jed).
father(jubel,lu).
parent(marx,mabel).
parent(X,Y) :- mother(X,Y).
parent(X,Y) :- father(X,Y).
grandfather(X,Y) :- father(X,Z),
parent(Z,Y).
——————-
9.9. EXAMPLES 103
If the above sample were consulted with
grandfather(rubel, lu).
the answer should be
yes.
We can also have questions with variables in them, e.g.
grandfather(X,Y).
In this case, prolog will try to find values of X and Y which make this
true. It might reply
X=jud, Y=jubel.
If you then type
you get another possibility such as
X=rubel, Y=marx.
By giving more semicolons, you should get all the possibilities, one after
the other.
You can also have compound queries, such as
grandfather(X,Y), parent(Y,lu).
And again all the possibilities should be generated,e.g. X=jud,Y=jubel.

Here is another example.
Example 9.3 —————————
cat(bags).
cat(felix).
bird(harold).
hunt(X,Y) : - cat(X), bird(Y).
————————–
Suppose we ask the question:
hunt(felix, harold).
Prolog matches this goal with the left hand side of the first rule, using
unifying substitution α = (X, f elix), (Y, harold). It then has two subgoals
cat(felix), bird(harold).
It satisfies these from left to right, and eventually says “yes”.

In this case there were no variables in the query. If we had asked
hunt(felix,W).
prolog would have eventually said
yes, W=harold.
You can see what prolog is doing, step by step, by turning on the trace.
This is done as follows:
9.10. CIRCULAR DEFINITIONS AND RECURSION 105
trace, hunt(felix,harold).
It may happen that the same variable is used in two different statements
in a program. In order to prevent clashes in the substitutions, most versions
of prolog will begin by renaming the variables. So you will find that new,
perhaps peculiar variable names are used in trace.
The trace can be turned off by calling
notrace.
9.10 Circular definitions and Recursion

You are lost in the country and you ask someone how to get to the old
Stanley place. “You go down Stanley road, “ you are told. So you ask:
How do you get to Stanley Road? “It goes right past the old Stanley
Place”, you are told.
This type of muddled definition or description will be called circular.
The person who has given it actually has the information which is wanted,
and has in fact said something possibly useful, but it is not sufficient to find
what we want. Compare that with the following, possibly from an obscure
nineteenth century anthropological work.
“If the relation Kram holds between a man and woman they are not
allowed to marry. Kram holds between a man and woman if they share the
same mother. Kram also holds between a man and woman, if Kram holds
between the mother of the man and the father of the woman. Kram never
happens for any other reasons.”
This definition might at first sight seem to be circular. Kram is defined
in terms of another instance of Kram. If you look at it carefully, however,
you will see that the explanation is arranged quite subtly to give us just
the right amount of information to let us know what Kram is. All the
information has been conveyed in a very compact way in this definition.
It is quite common for concepts to be self referential. A predicate at
one set of arguments is defined in terms of itself, but with other instances
of the arguments. This is a powerful technique but is obviously easy to
misuse. The problem is that a person attempting to apply the definition
may be sent off into considering an infinite loop of possibilities. A definition
or description which is self referential but eventually terminating is called
recursive. On the other hand, a muddled self referential definition will be

called circular.
Consider a definition of n!. We could say that 0! = 1. Also, (X 0 )! =
0
(X ∗ (X!)). This recursive definition gives us an algorithm to calculate n!
for any natural number n. We get
7! = 7 ∗ 6! = 7 ∗ 6 ∗ 5! = 7 ∗ 6 ∗ 5 ∗ 4! = 7 ∗ 6 ∗ 5 ∗ 4 ∗ 3! = 7 ∗ 6 ∗ 5 ∗ 4 ∗ 3 ∗ 2! =
7 ∗ 6 ∗ 5 ∗ 4 ∗ 3 ∗ 2 ∗ 1! = 7 ∗ 6 ∗ 5 ∗ 4 ∗ 3 ∗ 2 ∗ 1 ∗ 0! = 7 ∗ 6 ∗ 5 ∗ 4 ∗ 3 ∗ 2 ∗ 1.
In terms of prolog, we will say that a rule is recursive if the same predi-
cate appears on both the left and right hand sides. Of course the arguments
on the two sides may be different. In fact if the arguments are not different,
the rule is obviously circular, i.e. muddled.
Suppose, for example that we have a prolog program in which par-
ent(X,Y) is defined. We want to define “Z is a descendent of X”; This
should mean that either X is a parent of Z, or X is a parent of a parent of
Z, or X is a parent of a parent of Z or ....
We can express this recursively as follows.
———————-
descendent(Z,X) :- parent(X,Z).
descendent(Z,X) :- parent(X,Y), descendent(Z,Y).
———————
9.11 The Cut, and negation

Before reading this section, you should be sure that you understand the
idea of backtracking, which was explained above.
Example 9.4 ——————-
q(a).
s(a).
p(X) :- q(X), r(X).
p(Y) :- s(Y).
9.11. THE CUT, AND NEGATION 107
——————
Suppose we ask
p(a).
Prolog will find a match with the first rule, using α = {(X, a)}. It then
has subgoals
q(a), r(a).
It satisfies q(a), but then fails to satisfy r(a). So it backtracks, and dis-
cards the substitution α; it gives up on the first rule and tries the second. It
matches, with substitution {(Y, a)}, and gets subgoal s(a), which it satisfies.
So it says “yes”.
Prolog’s search for ways to satisfy its goal, is like the exploration of a
maze. The process of backtracking is like the retracing of steps in a maze,
after a dead end has been found.
Backtracking may be inhibited by writing
!
The instruction ! is called a cut. An example of how this is used would
be:
p :- q1, !, q2
This statement would tell prolog that if it is trying to satisfy p and has
got as far as q1, it can’t backtrack in order to satisfy p. The only way to
satisfy p is then to satisfy q2.
Continuing with the maze analogy, the cut is like a one way door in the
maze of possibilities. Notice that this has absolutely no axiomatic analogy.
There is no such thing as a cut, or anything like a cut, in a set of axioms.
With the appearance of the cut, prolog blatantly diverges from its original
idea. Nevertheless, the cut is interesting and useful.
Example 9.5 —————-
a :- b,c.
c:- d,!,e.
c:- f.
a:- d,f.
b.
d.
f.
—————–
The goal
c
will fail, but the goal
a
will succeed.
We can understand the cut more fully by returning to semantic tableaux.

As a goal, the cut always succeeds immediately. A branch with
!
at the leaf denied is just declared to be closed. As a predicate then
!
is like
True.
So a cut has no effect until a situation arises in which prolog tries to
backtrack through the cut. The cut then comes into effect. Normally prolog
would backtrack to the last goal, deleting the tree and the substitutions as
it goes; this usually means reconsidering the branch immediately to the left
of the current one. In the presence of the cut however, prolog backtracks up
the tree to the parent goal, deletes the parent goal and the whole subtree of
subgoals below it, and backtracks from there, as if no match for the parent
goal could be found in the whole program.
Example 9.6 Suppose, for example, we defined

————–
brother(X,Y):- father(Z,X), father(Z,Y), male(X), X \=Y.
brother(X,Y) :- mother(Z,X), mother(Z,Y), male(X), X \=Y.
————–
Suppose you are trying to decide whether or not
brother(a,b)
and you already know that a and b have a common father. You need now
to check that a is male and not the same as b. We know however that there
is no use looking for another common father, or in looking for a common
mother, since that will still leave us with the same subgoals. So we can add
cuts which speed up the computation without changing its results.
——————————–
9.11. THE CUT, AND NEGATION 109
brother(X,Y):- father(Z,X), father(Z,Y), !, male(X), X \=Y.
brother(X,Y) :- mother(Z,X), mother(Z,Y), !, male(X), X \=Y.
————————–
This is what we might call an innocuous cut, since we believe it never

can change the results of a computation. There are, however, other uses of
the cut which are not innocuous in this way.
9.11.1 The Cut and fail combination

There is a predicate
fail
which is always false.
It is clear that with cut and fail we can, in effect, build mazes with
traps in them. This can be used to express a sort of logical negation. For
example:
————————
p :- q, !, fail.
p.
————————
This means that if q can be satisfied, p must fail. On the other hand,
if q can not be satisfied, p is true. In this situation, p means that prolog
can’t prove q. We could think of p as
not `prolog q.
For example, prolog has a built in equality predicate, written
X == Y.
We can use this, together with the cut and fail combination to define
inequality, as follows.
————————
notequal(X,Y) :- X == Y, !, fail.
notequal(X,Y).
———————–
Another way of thinking of the above is that we take inequality to be
true by default; it will be false only when we can prove equality.
To give another example, we could define femaleness as the negation of
maleness, or vica verse
————————
female(X) :- male(X), !, fail.
female(X).
———————–
Note that the whole effect is ruined if we change the order of the state-
ments in a prolog program of this kind. So the meaning of such a prolog
program depends on the order of the statements in it.
Another annoying feature of the cut is that it interferes with the capacity
of a prolog program to generate a list of substitutions which satisfy a given
condition. This is because the
;
depends on backtracking, which may be inhibited. For example, the
above program will not generate a list of females. It can only be used to
test whether or not a known individual is female.
To make the example even more annoying, consider
—————–
female(X) :- male(X), !, fail.

female(X).
female(louise).
male(john).
——————–
If this program is given query
9.12. FAMILY TREE 111
female(X).
it will incorrectly say no.
9.12 Family Tree

Problem 9.1 Finish the family tree example, which is started above, con-
taining information about real or imaginary people over a number of gener-
ations. Define at least ten predicates, and include at least thirty facts. As
well as a program, draw a family tree for your community.
9.13 Lists, Member(X,Y), Append(X,Y,Z)

A list is a type of data structure. It is supposed to represent a set whose
elements are arranged in some order.
We assume that there is such a thing as the empty list. This is denoted
by [ ]. [ ] is a peculiar object but it is evidently necessary.
In Prolog all lists are finite.
In Prolog a list is written as a finite sequence of terms
[t1 , t2 , ..., tn ]
separated by commas and proceeded by a [ and followed by a ].
The terms t1 , t2 , ..., tn in list [t1 , t2 , ..., tn ] are called the components of
the list. The components of a list may also be lists.
Examples of Lists
[a,b,c] is a list with three components.

[[cat,X],Y] is a list with two components. The first component is also a
list.
[[the, quick, brown, fox], [jumped], [over, the, lazy, dog]] is a list with
three components.
The head of a list [t1 , t2 , ..., tn ] is the first component, t1 . The tail is
the rest of the list, [t2 , t3 , ..., tn ].
We use
[X | Y]
to denote the list with head X and tail Y .
[X | Y]
is usually read as “X cons Y”.

More examples
[a,b,c,d] has head a and tail [b,c,d].

[[a,a],[b,c],d] had head [a,a] and tail [[b,c],d].
[a | [b,c]] = [a,b,c]
[cat,cow,pig] = [cat | [cow,pig]] = [cat | [cow | [pig]]] =

[cat | [cow | [pig | [ ] ]]]
Prolog represents lists internally as built up from the cons operation.

[p, imp, [q, imp, p] ] is a list with three components. It is not the same
list as [[q,imp,p], imp, p], i.e. order is important in lists. A list may also
have repeated components. So [a,a,b] is not the same as [a,b].
Predicates involving lists are usually defined by recursion. For example,

—————-
samelength([ ], [ ]).
samelength([X | List1], [Y | List2])

:-samelength(List1,List2).
———————-
First the predicate is defined in the simplest case, when both lists are
empty. Then if we are given two lists, neither of which is empty, we chop
off their heads, and compare the lengths of the tails.
Most versions of prolog have built in predicates
member(X,Y)
and
append(X,Y,Z).
The predicate member(X,Y) is true if and only if Y is a list and X is a
component of Y. If our prolog does not have member(X,Y) already defined,
we can define it as follows:
————————
member(X, [X | Z]).
member(X, [W | Z]) :- member(X,Z).

9.14. SORTING 113
———————-
The idea here is: X is a member of Y if X is the first component of Y.
Otherwise check to see if X is a member of the tail of Y.
The predicate append(X,Y,Z) is true if X, Y, and Z are all lists, and
the list Z is obtained by appending list Y to list X.
We could give a recursive definition. The idea here is first to think of the
simplest possible case: append(X,Y,Z) is true if X is empty and Y and Z
are the same. We then express a more complicated case in terms of simpler
cases. Suppose X is not empty, but the head of X is the same as the head
of Z. We chop off these two heads and continue recursively.
—————————-
append( [ ], Y, Y).
append([X | Y], Z, [X | W]) :- append (Y,Z,W).
—————————
Exercise. Try to see what prolog actually does when given a question
such as:
append([cat],[dog,horse],[cat,dog,horse]).
9.14 Sorting
Prolog has built in predicates X = < Y and X < Y which have the usual
meaning as applied to numbers. We will say a list of numbers is monotone
non decreasing if every number in the list is less than or equal to the next
number, if any, in the list. We can define this as follows:
—————
mnd([X ]).
mnd([X | [Y | Z]]) :- X = < Y, mnd([Y | Z]).
————————-
There is an algorithm called bubble sort, which takes a list, X, of num-
bers and rearranges it to get a monotone non decreasing list, Y . If the first
list, X, is already monotone non decreasing, then bubble sort does nothing
and Y is the same as X. On the other hand, if X is not already monotone
non decreasing, bubble sort finds a pair of numbers in X which is out of
order, swaps them to get list Z and then applies bubble sort recursively to
Z.
Problem 9.2 Let bubble(X, Y ) mean that X is a list of numbers and Y is

a monotone non decreasing list which is obtained from X by bubble sort.
Define bubble(X, Y ) in prolog without using the cut or negation. (Hint:
If a list X is not in order, there is an adjacent pair of numbers in the list
which is out of order. )
If you give it a specific list X, and leave Y as a variable, it should return
a sorted list. For example
bubble([3, 5, 4], Y )
should get the response Y = [3, 4, 5].
This example shows how an algorithm, such as bubble sort, can be

expressed in a declarative style. However it should be clear that a solution
to the above problem actually relies on knowledge of how prolog behaves.
So the distinction between procedural and imperative styles tends to break
down when examined.
A solution to the above problem may look like a list of axioms for the
predicate bubble(X, Y ), but it is also a detailed specification of prolog’s
behaviour.
9.15 Equivalence relations; and how to find

your way out of a labyrinth
A labyrinth is defined by a data structure with finitely many nodes and
finitely many arrows. We will say that if an arrow goes from node X to
node Y then it is possible to go either from X to Y or from Y to X in one
step.
The labyrinth problem is to decide, given two nodes Z and W , whether
or not there is a path from Z to W , and if there is such a path to find it.
A slightly harder problem we will call the labyrinth problem with mon-
sters. In this case we age given a labyrinth, and a subset A of nodes which
we wish to avoid. The problem is, given two nodes, to decide whether or
not there is a path from one to the other which does not go through any
node in the subset A.
Both these problems are extremely ancient. The main difficulty is that
although the labyrinth is finite, it seems possible that in the search for a
path we might go around in circles forever.
In the following, we will make a series of attempts to solve this problem.
None of the proposed solutions are correct, except possibly the last one.
9.15. EQUIVALENCE RELATIONS; AND HOW TO FIND YOUR WAY OUT OF A LABYRINTH115
This section traces the evolution of a process of development of a solution

to this problem. To fix ideas, please stop at this point and create an example
labyrinth.
9.15.1 First Attempt

To begin with, we will represent the immediate connections by a list of
facts. This is:
—————————-
c(i,n).
c(t,e).
c(f,g).
c(w,e).
c(w,l).
etc.
—————————–
A binary relation on a domain D is a set of ordered pairs of elements
from D. We may represent a binary relation by a predicate of arity 2.
An equivalence relation on a domain D is a binary relation r(X, Y ),
defined over D so that the statements
1) r is reflexive, i.e.:
r(X, X)
2) r is symmetric, i.e.:
r(X, Y ) → r(Y, X)
3) r is transitive, i.e.:
(r(X, Y ) ∧ r(Y, Z)) → r(X, Z)
are true over D. This means that these statements are true for all pos-
sible values of the variables X and Y in D.
There is a path from a node X to another node Y if and only if we can
prove r(X, Y ) from the axioms for an equivalence relation, and
(∀X)(∀Y )(c(X, Y ) → r(X, Y ))
and the above list of facts. Therefore we might try to solve the labyrinth
problem with the following program.
.
—————
facts as above
r(X,X).
r(X,Y) :- r(Y,X).
r(X,Z) :- r(X,Y), r(Y,Z).
r(X,Y) :- c(X,Y).
————–
However this will certainly not work in prolog.
Prolog, given these axioms and a few facts, will just go into an infinite
loop. For example, trying to decide r(a,b), prolog might choose the second
of the two rules, and take as subgoal r(b,a); in trying to satisfy this, it
might again use the second rule, and find r(a,b) as a sub-sub goal, etc.
The fact that prolog falls on its face when given one of the simplest
axiom systems in mathematics shows that prolog does not fulfil its ideal
of directly representing logic. On the other hand, the fact that prolog gets
mixed up here can also be seen as a criticism of the usual way of doing
mathematics. The standard definition does not really give us an effective
definition of an equivalence relation.
Although the first attempt does not work, it seems that we have learned
something.
9.15.2 Second attempt

We have a problem with saying directly that r(X,Y) is symmetric. It would
not have been necessary to say that r(X,Y) was symmetric if all the arrows
originally went in both directions. If that were true, then r(X,Y) would
turn out to be symmetric, since it is defined symmetrically from c(X,Y).
So we could avoid having to say that r(X,Y) is symmetric just by putting
all the arrows in twice. A better solution is just to say that r(X,Y) is to be
true if either c(X,Y) is true or c(Y,X) is true.
So we get the following:
—————
facts as above
r(X,X).
r(X,Z) :- r(X,Y), r(Y,Z).
r(X,Y) :- c(X,Y).
r(X,Y) :- c(Y,X).
————–
Problem 9.3 Why does this not work?

9.15. EQUIVALENCE RELATIONS; AND HOW TO FIND YOUR WAY OUT OF A LABYRINTH117
9.15.3 Third attempt

It seems we should put the recursive step last. So we get
—————
facts as above
r(X,Y) :- c(X,Y).
r(X,Y) :- c(Y,X).
r(X,X).
r(X,Z) :- r(X,Y), r(Y,Z).
————–
Remarks. This is an improvement. It sometimes works and sometimes
does not work.
Problem 9.4 Give an example in which the above program does not work.
9.15.4 Fourth attempt

It seems that we need to keep track of the path and avoid going in loops.
We can use a variable
Route
to be the path of intermediate steps between the start and the finish.
The predicate
path(X,Y,Route)
will mean that Route is a path of intermediate steps which goes from X
to Y.
Take out the recursive step in the program above, and replace it with.
——————–
path(X,Y,[ ]):- r(X,Y).

path(X,Y,[Z | Route ]):- r(X,Z), nonmember(Z,Route),path(Z,Y,Route).
nonmember(X,R) :- member(X,R), !, fail.
nonmember(X,R).
————————-
Problem 9.5 What does the above program do?

Problem 9.6 What happens if we change the definition of nonmember(X,Y)

as follows.
—————–
nonmember(X,[ ]).
nonmember(X, [Y | Z]) :- X \=Y, nonmember(X,Z).
—————
9.15.5 Fifth attempt

Define
path(X, Z, Route, Avoid)
to mean that Route is a path of intermediate steps which goes from X
to Z but avoids going through any element in the list Avoid.
The definition is:
——————-
path(X, Z, [ ], Avoid) :- r(X,Z).

path(X,Z, [ A | B], Avoid) :- r(X,A),
nonmember(A, Avoid), path(A, Z, B, [A | Avoid]).
nonmember(X,Y) :- member(X,Y) , !, fail.
nonmember(X,Y).
————————–
Then to find a route, for example, from a to z, we ask:
path(a, z, Route, [ ]).
Prolog will give us the route if there is one, and will inform us if there
is none.
Problem 9.7 Do you believe the above claim? What would it mean to say
that the path finding program was correct? Try to state this carefully. Note
that if we ask
path(a,z,Route,A)
the program will fall on its face. How does it fall on its face? The area
to be avoided must be set initially to a constant. Once you have decided
how to state correctness, try either to prove or disprove it.
Would you be worried if your life depended on the correctness of a four
line program written by an expert?
9.16. LABYRINTH PROGRAM 119
Problem 9.8 Try moving the innocuous statement

r(X,X).
around in the program. Try listing all possible paths from one point to
another. On the basis of this, criticise the above program.
9.15.6 Sixth Attempt

Please write this yourself.
9.15.7 Testing and Proof

The series of examples above is supposed to convince you that there can
be short programs which look right to begin with but do not work. Also,
there is no way to prove that a program is correct by testing it. All a test
can do is reveal that a program is wrong.
Also it is quite possible for an experienced and conscientious person to
believe that an incorrect program is correct.
When you write a program and believe that it is correct, that belief
should contain at least a vague idea for a proof of correctness. It seems
that the only way to make any progress with this situation is to work
toward making these vague ideas about correctness more explicit. It is
usually quite difficult even to state clearly what we mean by correctness.
Problem 9.9 If you think the program above, which you wrote, is correct,
try to say why it is correct. If you are not able to say why it is correct,
but you think it is correct, can you offer any justification to support your
assertion? Can you at least say what correctness would be in this case?
It has been known from the beginning of the history of science that
anecdotal evidence, in situations in which the variables are not controlled,
and in which the witnesses are not neutral, is not reliable at all.
9.16 Labyrinth Program

Problem 9.10 Read the discussion of how to find your way out of a labyrinth
in the section 9.15. On the basis of this, write a short program in Prolog
which solves the problem of finding a path from one place to another in a
labyrinth. Explain just what you think your program does. Give some evi-
dence, in the form of tests and results, or, where possible, proofs, that your
claims are true.
9.17 Language Recognisers in Prolog

A language recogniser is a program or an algorithm which recognises whether
or not a given expression is a grammatical member of some formal language.
Consider, for example, the statement forms involving p,q, and r, as
defined by the grammar:
sf ; (sf ∨ sf ) | (sf ∧ sf ) | (sf → sf ) | (¬sf ) | (sf ↔ sf ) | p | q | r.
If we represent strings of symbols in a language by lists, it is fairly easy
to write language recognisers in prolog for context free languages, such as
those defining statement forms.
We could represent a statement form such as (p → (q ∨ r)) by a list
[p, imp, [q, or, r]]
using “imp” for →, and “or” for ∨. A language recogniser for the
propositional statement forms involving p, q, and r can be obtained directly
from the grammar.
————————-
sf([X, imp, Y]) :- sf(X), sf(Y).

sf([X, or, Y]) :- sf(X), sf(Y).
sf([X, and, Y]) :- sf(X), sf(Y).
sf([X, iff, Y]) :- sf(X), sf(Y).
sf([neg, X]) :- sf(X).
sf(X) :- member(X, [p,q,r]).
—————————-
The number of proposition names can be increased just by adding to
the list in the last line. If we really need infinitely many proposition names,
we could include
sf(X) :- atom(X).
using the built in prolog predicate atom(X), which will return true for
any sequence of characters beginning with a lower case letter.
The above prolog program will not only recognise grammatical state-
ment forms in variables p, q, r, but it will also generate them. For example,
if we ask a question
sf(X).
prolog will find an X which is a statement form. Presumably it will say
yes. X=p.
9.18. KITCHEN TABLE PROPOSITIONAL THEOREM PROVER 121
If we say
;
we will get another grammatical statement form, presumably q, and if
we continue to request more, we will get r, [p, imp, p], etc.
So this program appears both to generate and also to recognise state-
ment forms.
Most context free languages can be dealt with in this way. (Some extra
difficulty may occur when the grammar allows erasure of some grammatical
symbols, i.e.
V := λ
or when the rewrite rules contain a loop. )
9.18 kitchen table propositional theorem prover

We can also implement the semantic tableau method, as so far described,
in prolog.
In the following, A is supposed to be the list of statements asserted in
a semantic tableau, and D is supposed to be the list of statements denied.
The predicate contra(A,D) means that if the construction is started with
initial tableau (A,D), all branches will end in contradiction.
————————–
contra(A,D) :- member(X,A), member (X,D).

contra(A,D) :- remove(A,[neg,X ],A1),contra(A1,[X | D]).
contra(A,D) :- remove(D, [neg,X ],D1), contra([X | A],D1).
contra(A,D) :- remove(D, [X,imp,Y ],D1),
contra([X | A ], [Y | D1 ]).
contra(A,D) :- remove(A, [ X,imp,Y ],A1),
contra(A1,[X | D ],
contra( [Y | A1 ],D).
remove(List,Item,Shorterlist) :- append(A, [Item | B ],List),
append(A,B,Shorterlist).
————————–
The six lines of this program are supposed to correspond to the semantic
tableau rules for negation and implication. You will need to add some more
rules for other logical operators. To try to prove (p → p) you would query:
contra([ ],[[p, imp ,p]]), and hope to get the answer yes. Since the semantic
tableau system was sound and complete, this program is also a sound and
complete formal system for the propositional calculus.
As it stands this kitchen table theorem prover does a lot of unnecessary
backtracking. You may wish to add some cuts to speed things up. Another
idea is arranging the statements so that the branching is done as late as
possible. You may also want to use some write statements, to see what is
happening.
Note that although prolog is defined only using implication and conjunc-
tion in a very special form, we have obtained a complete theorem prover
for the propositional calculus inside prolog. In a sense this is a triumph.
We have encapsulated almost everything we know about the propositional
calculus in a few lines of code.
9.19 Problems
Problem 9.11 What is the relation between the Assertion and Denial lists
which occur in the theorem prover above, and the branches which occurred
in the semantic tableau algorithm?
Problem 9.12 Explain what prolog will do with
contra([p],[[q,imp,p]]).
Problem 9.13 Suppose the propositional theorem prover is given
contra([],[[[p,imp,q],imp,[q,imp,r]]])
Following what prolog is doing, find the first subgoal
contra(A,D)
which fails. This means that there are only variables left in the live
part of the branch, and that the same variable does not occur on both sides.
Check that this gives a counterexample. What does prolog do next? Modify
the code, using cuts, so that as soon as prolog finds a counterexample, it
stops the computation and prints the counterexample.
To get something printed on the screen, you can use the built in
write(X)
which will write whatever term X is bound to. For example
write(A, ’are true’)
will print the current list of Assertions followed by the text “are true”.
9.19. PROBLEMS 123
Problem 9.14 Write a prolog program which will check whether or not an
expression represents a grammatically correct formula of LN , the language
of first order single sorted arithmetic. (Such a program is called a language
recogniser.) Then translate the following statements into LN , give them to
your language recogniser, and see if you were correct, at least grammatically.
If not, rewrite your translations, correcting the grammar. You can save
yourself some time by telling your language recogniser to report errors as it
finds them.
a) X is prime.
b) X can be written as the sum of three squares in just one way.
c) some linear combination of X and Y can not be written as the sum
of two squares unless Z is even.
d)X is congruent to Y modulo Z
e) There are infinitely many prime numbers.
f ) There are infinitely many prime numbers P such that P+2 is also
prime.
g)There are only finitely many prime numbers P so that P+2 is also
prime.
Problem 9.15 Implement the propositional theorem prover which was de-
scribed above, and apply it to formulae of LN . As a result you should be
able to prove all tautologies in LN , a tautology being a formula obtained by
substitution from a statement form tautology.
Problem 9.16 * Invent some axioms, which seem correct to you, and add
them to your theorem prover so that you can prove some statements which
are true in the normal sense of the word about the natural numbers, but
which are not tautologies. Try to make your theorem prover as good as you
can, in a couple of hours of work.
Try your theorem prover on f ) and g) in the above list.
What does “true in the normal sense” mean in the above? Is everything
which your theorem prover proves true in the normal sense? Do you think
there are any statements which are true in the normal sense but which are
not consequences of your theorem prover? Please think quite hard about this
and explain your opinion, using common sense arguments. We will return
to this later.
Problem 9.17 Give a recursive definition of finite tree and translate this
into prolog.
Problem 9.18 Write a prolog program which finds conjunctive normal

form for a statement form, represented as a list. Use the rewriting method
rather than the truth table method. cnf (X, Y ) should be true if X is a list
corresponding to a statement form and Y is obtained by applying a sequence

of rewrite rules, according to some priority, starting with X, and Y is in
conjunctive normal form. This problem is solved, in ML, in the book by
Paulson. Your program will probably not successfully test whether or not Y
is a possible conjunctive normal form for X; however given X the program
should compute Y which is one of the possible conjunctive normal forms for
X.
This problem convinces me that pure declarativeness should not be

taken as an ideal feature of programming. The predicate
Y is a possible conjunctive normal form for X
is much harder to describe than the conjunctive normal form algorithm.
In general it quite often happens that the set of solutions to some prob-
lem is quite complicated, but a particular solution can be found relatively
easily by performing a certain sequence of operations.
There is a saying that there are two types of mathematicians, Greek and
Babylonian, Greek ones being concerned with finding essential relationships
in reality, and Babylonian ones being concerned with calculations. In my
opinion this is a false distinction; and attempting to make sharp distinction
between procedural and declarative programming is equally flawed.
9.20 Input and Output in Prolog

When prolog is running it will have at each instant a current input stream,
which is initially from the keyboard, and a current output stream, which is
initially the display.
We can switch streams as follows. Suppose X is instantiated to a file-
name. Then
tell(X)
switches the current output stream to X. The first time this goal occurs
a new file with name X is created.
Similarly
see(X)
switches the current input stream to X.
Prolog has built in predicates
write(X)
and
read(X)
which write to and read from the current input and output streams.
9.20. INPUT AND OUTPUT IN PROLOG 125
9.20.1 write(X)
If a variable X is instantiated to a term, then write(X) will succeed and
cause the term to be printed on the current output stream. So, for example,
the items in a list, L, would be printed one after another by
expose(L)
with the following definition
--------------
expose([]).
expose([X | Y]) :- write(X), expose(Y).
---------------
In connection with this, there is another useful built in predicate

nl
which means “new line” and has the effect that all succeeding output is
printed on the next line.
Another built in predicate is
tab(N)
which, when N is instantiated to a natural number, causes the cursor
to move to the right by N spaces.
9.20.2 read(X)
The predicate
read(X)
will attempt to unify X with the next term that is input on the current
input stream. The term must be followed by a full stop and a non printing
character, such as a space or a return.
9.20.3 Debugging and tracing

You can get a full running report of what prolog is doing with the instruction
trace.
When this occurs as a goal the tracing facility is turned on. The result
of this is that prolog will write every goal which it attempts to satisfy,
including subgoals, failures and backtracking. The trace can be turned off
with
notrace.
The problem with trace is that such a large amount of information is

reported. You may only wish to see what prolog is doing in regard to certain
goals. Suppose we are interested in a predicate
p(X,Y,Z)
with arity 3. We could ask for a report of attempts to satisfy goals
involving this predicate with the instruction
spy(p/3)
and this can be turned off with
nospy(p/3).
In general the correct syntax for the spy argument is
predicate-name/arity.
So, for example if we wanted to follow what prolog was doing with
append(X,Y,Z), we could use
spy(append/3).
9.20.4 Assertions and Retractions

Suppose that in the middle of running prolog we decide that we would like
to add a new fact, say
p(a,b).
to our program. We could of course stop, add the new fact and rerun
prolog. However it is also possible to add the fact immediately, as follows
asserta(p(a,b)).
In general if X is a clause, we can add X to the beginning of the program
with
asserta(X).
A clause X can be added to the end of the program with
assertz(X).
A clause X can be removed from the program with
retract(X).
You can see what clauses are currently in the program with the query
clause(Head,Tail).
9.21 Arithmetic in Prolog

Prolog has built in all the usual binary predicates for inequality between
integers: X < Y, X =<Y, X> Y, X>=Y.
Prolog also has built in function symbols for some arithmetic operations:
X + Y, X - Y, X * Y, X / Y, X mod Y.
9.21. ARITHMETIC IN PROLOG 127
There is also an assignment command in prolog which sets a variable X

to the value of an expression, E, which can be evaluated to give an integer
or a real number. This is written
X is E.
Thus
X is (1+2)*3.
succeeds with X=9. Similarly
(2+3) < 4*5.
succeeds, first evaluating 2+3 and then evaluating 4*5 and then com-
paring.
Notice however that we need another version of equality, since, for ex-
ample
4 = 2+2.
will fail, since the response of prolog to this will be to try to unify 4
and 2+2, which is not possible. What we need is a binary predicate which
evaluates two arithmetic expressions and then checks to see if the results
are the same. This is written
X =:=Y.
Example 9.7 Suppose we want to define a predicate sum(X,Y) meaning

that X is a natural number and Y is the sum of the natural numbers from
0 to X inclusive. This could be done as follows.
———————-
sum(X,0):- X = < 0.
sum(X,Y) :- Z is X-1, sum(Z,W),Y is W +X.
—————–
If we give this the query
sum(999,S).
we will get the answer S=999500. The program, given X, will find Y
so that sum(X,Y) is true. However it will not work in reverse. So the
program, given S=999500 will not find X so that sum(X,S). This is because
when the program gets to line 2, it will try to evaluate X-1 and fail since X
is uninstantiated.
Example 9.8 The following program finds the greatest common divisor K
of two integers I and J.
——————–
gcd(I,0,I).
gcd(I,J,K) :- R is I mod J, gcd(J,R,K).
———————
As above this only works when I and J are instantiated.
Prolog has a built in predicate

integer(X)
which succeeds if X is a whole number. If X is an expression, it is not
evaluated inside this predicate. So integer(9) will succeed but integer (7+2)
will fail. Also, if X is uninstantiated, integer(X) will fail. In particular the
goal
integer(X).
will not result in prolog’s choosing an integer.
Suppose we wanted a predicate which would list the natural numbers.
We could write this as follows.
——————
natural(0).
natural(X):- X is Y+1, natural(Y).
—————–
9.22 Typed Prolog

There are several logic programming languages which have attempted to
improve on prolog. A particularly interesting one is the Gödel system. See
[Hill and Lloyd,1994].
Prolog is a language with only one type. A variable in prolog stands for
some unspecified object, and the objects are understood to belong to one
big domain.
The Gödel language on the other hand, has a multiplicity of types.
There are constructors, such as, for example, lists, which produce new
types from old ones. So if we had a type, integers, we might also have lists
of integers, lists of lists of integers, lists of lists of lists of integers, lists
of integers and lists of integers, and so on. In fact, as long as there is one
constructor and one type to begin with, an infinite number of distinct types
are generated.
9.23. SUMMARY OF THIS CHAPTER 129
Another difference between the Gödel language and prolog is that the
former has abandoned the cut, but uses a related construction called com-
mit. A virtue of the commit operation is that programs written with it
have a meaning which is not entirely destroyed by changing the order of
the statements in the program.
Gödel allows definitions with quantifiers. This implies that a full, com-
plete implementation of the language can not be specified.
Gödel also continues to follow the extreme form of the declarative ideal,
which is that the procedural part of a program should be automatically and
unobtrusively derived from the declarative part. My opinion is that this is
a mistake.
9.23 Summary of this chapter

A prolog program is a compact way to represent information about a situ-
ation.
A predicate is a function whose codomain is the truth values. Prolog
is concerned with representing logical relationships between predicates by
rules and facts.
A fact is a single predicate expression.
A rule is an implication, written in the form
p :- q1, q2, ..., qn.
where p, q1, q2,..., qn are all predicate expressions. The rule
p :- q1,q2,..., qn.
means that p is true whenever all of q1,q2,..., qn are true.
From a procedural point of view, such a rule says that in order to find
values which make p true, we should search for values which make q1, q2,...,
qn true.
In order to create a prolog program about a situation, do the following.
• Decide what are the important predicates.
• Arrange the predicates in a hierarchy of complexity.
• Define more complex predicates in terms of simpler ones, using rules.
• Enter facts about predicates which are as simple as possible. Put the
facts at the top of the program.
The declarative meaning of a prolog program is as a list of axioms.

The procedural meaning is the computation prolog will do given a query.
This procedure may be understood either in terms of semantic tableaux
(explained in section 9.1), or in terms of an algorithm for satisfaction of

goals (explained in section 9.4).
Lists in prolog may be written as a sequence of terms, separated by
commas and enclosed in square brackets.
[X | Y],
read as X cons Y is the list with head X and tail Y.

Predicates on lists are usually defined by recursion. An important ex-
ample is append(X,Y,Z) which is defined by two rules
append( [], X, X).

append([X | Y ], Z, [ X | W ]) :- append(Y,Z,W).
Terms in prolog include constants and variables and the set of terms is
closed under formation of lists and application of function symbols.
Two terms A and B are unified by a substitution α if Aα = Bα .
A substitution α is a most general unifier of two terms A and B if
• α unifies A and B.
• If β is any other unifier of A and B, there is a substitution δ so that
α◦δ =β
We have an algorithm to find a most general unifier of any two terms.

When prolog matches a goal with a fact or the head of a rule, it does so by
finding a most general unifier.
Using lists and recursion, we can write short programs in prolog with
with very powerful computational behaviour. As an example of this, the
propositional part of the semantic tableau algorithm is implemented in
prolog. We can also easily write recognisers for context free languages in
prolog.
On the other hand, there are many short programs in prolog which look
right at first but are badly flawed. There are also programs which may be
correct but for which correctness is hard to prove.
It should be pointed out that it is unscientific to claim that a program
is somehow validated by reported correctness on some examples invented
by the writer of the program.
Prolog programs using recursion and cuts should be viewed with espe-
cially sharp scepticism.
Chapter 10
Multisorted, higher order

languages, and non
classical logics
Up to this point we have been dealing with first order languages. In this
situation there is only one type of variable, and in the semantics there is
only one universal domain of objects. This is also the situation in prolog.
We have seen that it is possible to express almost anything in this frame-
work. However in computing languages, as in real life, we often wish to use
languages in which variables are given different types to show that they are
intended to range over different domains. For example, we might have a
type for the integers and another type for the reals, and various functions
which go from one domain to the other. In general, we will write X : A to
mean that variable X has type A. So for example X : integer means that
X is an integer. A multisorted language is a first order language in which
the variables and function and predicate symbols have all been typed. We
can also have a multisorted term language.
We may also wish to have variables for sets or functions, and to quantify
over these variables. For example we can obtain higher order arithmetic by
extending LN by adding variables for sets of natural numbers, and a new
predicate for set membership.
Bound variables can also be given types. So, for example,
(∀X : A)p(X)
means that p(X) is true for all X of type A. Similarly,
(∃X : A)p(X)
means that there exists and X of type A so that p(X). Many of the
131
132CHAPTER 10. MULTISORTED, HIGHER ORDER LANGUAGES, AND NON CLASSICA
ideas and methods we have considered for first order languages naturally
extend to these typed languages. For example, CNF, DNF, prenex normal
form, Skolem form, clausal form. The semantic tableau method also extends
in a natural way. However, since the types usually have a fixed intended
interpretation, the semantic tableau method is no longer complete.
10.1 Notation for types

We will assume certain basic types, such as integer, real, bool (for the truth
values), nat(for the natural numbers). More complex types will be formed
by type constructors, some of which are in the following list.
Suppose A and B are types.
1. A → B is the type of all functions from A to B.
2. A × B is the type of all ordered pairs (x, y) where x : A and y : B.
3. A∪B is the disjoint union of A and B. This is obtained by first making

two copies of A and B in such a way that they do not intersect, and
then taking the union. So, for example, integer ∪ real is the disjoint
union of the integers and the reals. The integer 5 in this context is
not the same as the real number 5.0.
4. P (A) is the type of all subsets of A.
5. A list is the type of lists of elements from A
So, for example, the factorial function has type nat → nat. This would
be written f act : nat → nat. The logical operator of conjunction has type
bool × bool → bool. Addition over the real numbers has type real × real →
real. Reversal of a list of natural numbers has type list nat → list nat.
10.2 Problems with types

Problem 10.1 We will consider two types to be essentially the same if we
can find a natural bijection from one to the other. Suppose A, B, C are
types. Are the following pairs of types essentially the same? Why or why
not? (If you say there is a bijection, you must define it!) If you say that
they are not the same, you should give an example to show this.
• (A × B) × C and A × (B × C)?
• What about A → (B → C) and (A × B) → C?

10.3. NON CLASSICAL LOGICS 133
• Is A × (B ∪ C) essentially the same as A × B ∪ A × C?
• (A → (B → C)) and ((A → B) → C) ?
10.2.1 Solutions to Types

In the following, maps are written on the left.
1. ((A × B) × C) is in 1-1 correspondence with (A × (B × C)), via the

bijection f defined as follows.
Pick x in ((A×B)×C. x has the form ((a, b), c), where a : A, b : B, c :
C. Define f (x) = (a, (b, c)). Note f (x) : (A×(B ×C)). We can define
function g in the other direction as follows. g((a, (b, c))) = ((a, b), c).
We have f ◦ g is the identity on (A × (B × C)), and g ◦ f is the identity
on ((A × B) × C). (Check this by calculation, from the definitions.)
Thus g is a two sided inverse for f and f is a bijection.
2. There is a bijection in this case. Define

f : (A → (B → C)) → ((A × B) → C) as follows. Pick x : A → (B →
C). Define f (x) = y : (A × B) → C by
y(a, b) = x(a)(b)
for a : A, b : B. Note that x(a) : B → C, so this makes sense.
As above, we define an inverse, which we will call g. Pick arbitrary
y : (A × B) → C. Define g(y) = x : A → (B → C) by
x(a)(b) = y(a, b)
for a : A, b : B. f ◦ g and g ◦ f are the identity on their respective
domains. So f is a bijection. (You may feel that this is not a natural
bijection. Note, however, that it is quite universal, i.e. independent
of A, B, C.)
3. Same sets. The identity is a bijection.
4. Not the same cardinality in most cases. So there is no bijection in

most cases.
10.3 Non classical logics

We may also wish to question our original assumptions about the truth val-
ues. Are there really only two truth values? Do we believe in the excluded
middle principle which we have used? If we find that ¬A is impossible,
134CHAPTER 10. MULTISORTED, HIGHER ORDER LANGUAGES, AND NON CLASSICA
does this imply that A is true? To see some of the developments and appli-
cations of this, the reader may refer to Logic in Computer Science by Huth
and Ryan.
10.4 Proof Assistants

Even in the case of first order languages, we are not so far able successfully
to automate the whole deductive process in a practical way. We have seen
that prolog is quite limited since it always uses the same search strategy,
and that the semantic tableaux method somewhat computationally diffi-
cult, since it leaves so many choices open. The multisorted higher order
situation, which is what we usually want for applications, is certainly not
going to be easier. A great deal of work has been done to develop systems in
which proofs are constructed by interaction between a human being and a
computer. The computer has at its disposal something like unification and
the semantic tableau system. The human gives the initial proof goal. If the
computer gets stuck, the human can suggest lemmas which are, hopefully,
intermediate steps toward proving the goal. Example systems of this kind
are HOL, Isabelle, and COQ. HOL means higher order logic. Isabelle is a
proof assistant for higher order logic based on HOL. Isabelle also does first
order logic and Zermelo Fraenkel set theory. It is written in the language
ML. See
www.cl.cam.ac.uk/Research/HVG/Isabelle
for documentation and a tutorial.

Chapter 11
Semantics and
Specification for Programs
Consider an imperative programming language, such as C. In such a lan-

guage, a program is a list of instructions, which is given structure by branch-
ing and looping constructions. A typical line of such a program is an as-
signment, in which the value of a variable is reset. The program itself is
just a piece of text. We would like to give a semantics to such programs.
That is, we would like to be able to give such a program a precise interpre-
tation, as some kind of mathematical object. We all have some such object
in mind when we discuss what a program is meant to do. If we are able to
clarify our ideas in this area, we may also in some cases be able to construct
mathematical proofs of the correctness of computer programs or parts of
computer programs. Even if we are not able to prove all the results we
would like to prove, the existence of a formalisation has a wholesome effect
on the whole subject of computing, since two practitioners who disagree
about some fundamental issue can dispute in a common framework. This
is also true for disputes between client and programmer. There is obviously
a practical use for precise specification.
11.1 Programming Language Semantics

A programming language is, in a sense, similar to a first order language in
that it has a signature, which consists of a list of function symbols and a
list of predicate symbols. There are two important differences however.
1. First order languages only have variables of one type. In a first order
135
136CHAPTER 11. SEMANTICS AND SPECIFICATION FOR PROGRAMS
language, a variable stands for some unspecified object in the domain

of the interpretation. We do not determine in advance what the
domain of the interpretation is. It is true that some programming
languages, such as prolog for example, or some versions of Lisp, are
also untyped. However, most programming languages give types to
variables, as discussed previously. For this reason we need to use
multi sorted languages.
2. Formulae in first order languages are interpreted as statements. They

return truth values. A set of formulae is interpreted as a list of condi-
tions. On the other hand, a piece of software is naturally interpreted
as an action of some sort. The software does something. The action
of the software is the interpretation.
There are a number of differing approaches to the semantics for com-

puter software. Two important examples are operational semantics and
denotational semantics.
Operational semantics interpretes software as operations of some ma-
chine M . A compiler or interpreter from a programming language L for
a machine M translates code in L into sequences of machine instructions
for M , which are then realised as machine operations. The operational
semantics for a language L is different for each target machine M .
Denotational semantics, on the other hand, gives each part of the signa-
ture of a programming language L a mathematical interpretation. Types
are interpreted as mathematical sets. Function symbols are interpreted as
functions. Predicate symbols are interpreted as predicates.
11.2 Denotational Semantics

We previously gave a semantics to formulae in a first order language. Given
a formula of the language, we can find the function symbols, predicate
symbols and constants which appear in the formula; we can call this the
signature of the formula. We can analyse a piece of code or pseudocode in
the same way. Consider, for example,
gcd(x:integer, y:integer):integer;
r:=remainder(x,y);
while (r not = 0)
x:=y;
y:=r;
11.2. DENOTATIONAL SEMANTICS 137
r:=remainder(x,y);
return y;
We find a constant, 0, the equality predicate and one function symbol,

remainder(x,y). We can give types to all of these. So far so good. However,
the program does not seem to be a formula. It is an action rather than a
statement. Nevertheless, we can begin by considering an interpretation
for the language with this signature. Thus we need a domain, D, and an
interpretation for = and for remainder(x, y).
The code we are looking at has type declarations, for example x :
integer. This says that we are not free to choose the domain for x. In
general types in computing are sets which are associated with standard op-
erations and predicates. So when you know the type of an object, you also
know how to apply the common functions to it. In particular, we would re-
gard equality and remainder(x, y), i.e. the remainder when x is divided by
y as fixed by convention in the integer type. Also, 0 is fixed by convention.
So the whole interpretation is fixed.
We previously defined a valuation to be a function which gave values in
the domain to some subset of of the variables. As before, we have valuations.
In the semantics of programming languages, these are usually called states.
An example of a state would be {x := 52, y := 13}.
If we start with any state, and apply a program, it will either not ter-
minate or it will terminate with some other state. This is the basic idea of
denotational semantics: computer code is interpreted as a partially defined
transformation between states. We would say that the code denotes the
transformation.
Once we are clear about the interpretation, the transformation is spec-
ified. We then have a clear semantics for programming languages.
138CHAPTER 11. SEMANTICS AND SPECIFICATION FOR PROGRAMS
11.3 Correctness and Completeness of soft-

ware
Let α be a piece of software with inputs X1 , ..., Xn and outputs X10 , ..., Xn0 .
Let ∆ be a specification for α.
−−−−−−−−−−−−−−−−−−∆−−−−−−−−−−−−−−−−−−
Declarations
X1 : A1 , . . . , Xn : An
——————————————-
Predicates
P recondition(X1 , . . . , Xn )
P ostcondition(X1 , . . . Xn , X10 , . . . , Xn )
———————————————-
We will say that α is correct with respect to ∆ if whenever α is given
(X1 , . . . , Xn ) of the right types and (X1 , . . . , Xn ) satisfies the precondi-
tion, and α terminates with (X10 , . . . , Xn0 ), then (X1 , . . . , Xn , X10 , . . . , Xn0 )
satisfies the postcondition.
We will say that an input (X1 , . . . , Xn ) is feasible for ∆ if there exist
output values (X10 , . . . , Xn0 ) so that (X1 , . . . , Xn , X10 , . . . , Xn0 ) satisfies ∆.
We will say that α is complete with respect to ∆ if α terminates for
every input (X1 , . . . , Xn ) which is feasible for ∆.
11.4 Formal Methods

In some situations, where software reliability is very important, we may
try to prove, with some computer assistance, that a piece of code has cer-
tain properties. You now have all the ideas which you need to begin to
understand how such a system could work. Look up Isabelle or COQ for
examples. If software engineering is ever to be comparable in reliability to
ordinary engineering, these techniques will need to be developed.
Chapter 12
References
Boolos, G. S., Jeffrey, R., Computability and Logic, Cambridge University

Press, 1999
Bratko, I., Prolog programming for artificial intelligence, Addison Wes-
ley
Huth, M. R. A., Ryan, M. D., Logic in Computer Science, Cambridge
University Press, 2000
See www.cs.bham.ac.uk/research/lcs/
Follow wwwtutor link for review questions about Huth and Ryan book,
also useful for this course.
See
www.afm.sbu.ac.uk
for many links concerning formal methods.
See
turing.wins.uva.nl/~johan/Phil.298.html
139
140 CHAPTER 12. REFERENCES
for Johan van Bentham’s excellent notes on logic in games.
See
mas.colognet.org/objectives.html
for information about agents and logic.
Chomsky, N, Aspects of the Theory of Syntax. This argues for the

connection between formal grammars and natural languages.
Slonneger, K., Kurz, B.L., Formal Syntax and Semantics of Program-
ming Languages, Addison Wesley, 1995. This has a nice discussion of the
lambda calculus, and of Scott domains. A beta reduction program is writ-
ten in prolog.
Stoy, J.E., Denotational Semantics: The Scott-Strachey Approach to
Programming Language Theory, MIT Press, 1977. This is an advanced
text.
Winskel, Glyn, The formal semantics of programming languages, The
MIT press, Foundations of computing series, Cambridge Mass, 1993.
Wooldridge, Michael, An Introduction to MultiAgent Systems, John
Wiley, 2002. Chapters 3 and 12 describe how some deductive ability can
be built in to agents. The main idea is that selection of an action is a form
of deduction.

Logic PDF

Uploaded by

Copyright:

Available Formats

Logic PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Logic PDF

Uploaded by

Copyright:

Available Formats

Formal systems, logic and semantics

September 27, 2006

2 String Rewriting Systems 11

3 Term languages, substitution and unification 17

6 Other formal deductive systems for first order logic 75

7 A Formalisation of Mathematics: ZF set theory 79

8 Gödel Incompleteness theorems 85

9 Logic programming ( prolog) 87

10 Multisorted, higher order languages, and non classical log-

11 Semantics and Specification for Programs 135

1.2 Problem Solving

1. A person trying to solve the Rubic cube puzzle.

2. Two people playing chess.

3. A mouse, apparently trying to find its way out of a maze.

4. A child learning to speak.

5. A mathematician trying to prove a theorem, which he or she is intu-

6. A computer programmer developing a program.

Definition 1.1 A formal system is (D, R) where D is a set of data struc-

Definition 1.2 If (D, R) is a formal system, let X ⇒R Y mean that X

Definition 1.3 If (D, R) is a formal system, let

X ⇒∗R Y means that it is possible to get from X to Y by a finite

Remarks. If you like geometric pictures, you may visualise a formal

1.3 Design principles of formal systems

2. Correctness. (D, R) does nothing which the specification prohibits.

3. Simplicity. This is also called the Occam’s razor condition. One

4. Naturalness. The formal system should be in accord with intuition.

All of these conditions are important. In many cases, in most interesting

1.5 Related ideas:Confluence and termination

A formal system (D, R) is locally confluent if whenever A ⇒ B and

Theorem 1.1 A formal system is confluent if it is locally confluent and

String Rewriting Systems

Suppose we start with a language, such as English or Urdu or C or LZF .

A rule α ; β means that it is allowed to replace α by β in any context.

Definition 2.1 A Σ rewriting system is a finite list of Σ rewriting rules:

2.1 Rules and derivations

2.2 Language generation: Grammars

Definition 2.6 A grammar is a rewriting system P together with an initial

2.3 Problems (which may be discussed in tu-

Problem 2.3 Let Σ = {X,0 , 0}.

Problem 2.4 Describe the language L2 generated by the grammar (P, X)

Problem 2.5 Consider languages L such that 0 is in L and whenever a

Problem 2.7 Here are some samples from a language.

1. (((p ∨ q) ∧ ¬r) ∨ (p ∧ q))

Give a grammar to generate this language. Also give a description of the

Problem 2.8 Here are some samples from a language.

Give a grammar, as simple as possible, which generates all of these. Al-

Problem 2.9 Let P1 be a programming language, consisting of statements.

Example 2.3 LZF , the language of Zermelo-Fraenkel set theory, is gen-

A formal language is one which can be defined by a grammar, as de-

2.5 Other uses of string rewriting systems

3.1 Function Spaces

3.2 Term languages

Programming languages typically have term languages as expressions.

Term languages are languages with a fairly simple semantics. So this is

a typing mistake; it is not a term. Similarly, s(A, B, C) might be a term in

One way to define a term language is by a grammar. For the above

3.3 Interpretation of a term language

In other words, an interpretation of a term language gives a specific

Example 3.4 Consider signature [f, g, s, a] with arities 2,1,1, 0. One of

If (sqrt(9 + 4sqrt(2)) = 1+ 2sqrt(2))