BASIC SIMPLE TYPE THEORY

Cambridge Tracts in Theoretical Computer Science

Editorial Board

S. Abramsky, Department of Computing Science, Edinburgh University

P. H. Aczel, Department of Computer Science, University of Manchester
J. W. de Bakker, Centrum voor Wiskunde en Informatica, Amsterdam
J. A. Goguen, Programming Research Group, University of Oxford
Y. Gurevich, Department of Electrical Engineering and Computer Science, University
of Michigan
J. V. Tucker, Department of Mathematics and Computer Science, University College
of Swansea

Titles in the series

1. G. Chaitin Algorithmic Information Theory
2. L. C. Paulson Logic and Computation
3. M. Spivey Understanding Z
4. G. Revesz Lambda Calculus, Combinators and Functional Programming
5. A. Ramsay Formal Methods in Artificial Intelligence
6. S. Vickers Topology via Logic
7. J: Y. Girard, Y. Lafont & P. Taylor Proofs and Types
8. J. Clifford Formal Semantics & Pragmatics for Natural Language Processing
9. M. Winslett Updating Logical Databases
10. K. McEvoy & J. V. Tucker (eds) Theoretical Foundations of VLSI Design
11. T. H. Tse A Unifying Framework for Stuctured Analysis and Design Models
12. G. Brewka Nonmonotonic Reasoning
14. S. G. Hoggar Mathematics for Computer Graphics
15. S. Dasgupta Design Theory and Computer Science
17. J. C. M. Baeten (ed) Applications of Process Algebra
18. J. C. M. Baeten & W. P. Weijland Process Algebra
19. M. Manzano Extensions of First Order Logic
21. D. A. Wolfram The Clausal Theory of Types
22. V. Stoltenberg-Hansen, I. Lindstrom & E. Griffor Mathematical Theory of Domains
23. E.-R. Olderog Nets, Terms and Formulas
26. P. D. Mosses Action Semantics
27. W. H. Hesselink Programs, Recursion and Unbounded Choice
28. P. Padawitz Deductive and Declarative Programming
29. P. Gardenfors (ed) Belief Revision
30. M. Anthony & N. Biggs Computational Learning Theory
31. T. F. Melham Higher Order Logic and Hardware Verification
32. R. L. Carpenter The Logic of Typed Feature Structures
33. E. G. Manes Predicate Transformer Semantics
34. F. Nielson & H. R. Nielson Two Level Functional Languages
35. L. Feijs & H. Jonkers Formal Specification and Design
36. S. Mauw & G. J. Veltink (eds) Algebraic Specification of Communication Protocols
37. V. Stavridou Formal Methods in Circuit Design
38. N. Shankar Metamathematics, Machines and Godel's Proof
39. J. B. Paris The Uncertain Reasoner's Companion
40. J. Desel & J. Esparza Free Choice Petri Nets
41. J.-J. Ch. Meyer & W. van der Hoek Epistemic Logic for AI and Computer Science
42. J. R. Hindley Basic Simple Type Theory
BASIC SIMPLE TYPE THEORY

J. Roger Hindley
University of Wales, Swansea

CAMBRIDGE
UNIVERSITY PRESS
PUBLISHED BY THE PRESS SYNDICATE OF THE UNIVERSITY OF CAMBRIDGE
The Pitt Building, Trumpington Street, Cambridge CB2 1RP, United Kingdom

CAMBRIDGE UNIVERSITY PRESS

The Edinburgh Building, Cambridge CB2 2RU, United Kingdom
40 West 20th Street, New York, NY 10011-4211, USA
10 Stamford Road, Oakleigh, Melbourne 3166, Australia

© Cambridge University Press 1997

This book is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.

First published 1997

Typeset in 10/13 point Times

A catalogue record for this book is available from the British Library

Library of Congress Cataloguing in Publication data

Hindley, J. Roger.
Basic simple type theory / J. Roger Hindley.
p. cm. - (Cambridge tracts in theoretical computer science ; 42)
Includes bibliographical references and index.
ISBN 0 521 46518 4
1. Programming languages (Electronic computers) 2. Type theory.
1. Title. II. Series.
QA76.7.H55 1996
005.13-dc2O 95-9058 CIP

ISBN 0 521 46518 4 hardback

Transferred to digital printing 2002

To Carol
 Contents

Introduction page ix
1 The type-free A-calculus 1

1A A-terms and their structure

1B /3-reduction and #-normal forms 4
1C ri- and fn-reductions 7
1D Restricted 2-terms 10

2 Assigning types to terms 12

2A The system TAA 12
2B The subject-construction theorem 20
2C Subject reduction and expansion 24
2D The typable terms 27

3 The principal-type algorithm 30

3A Principal types and their history 31
3B Type-substitutions 34
3C Motivating the PT algorithm 38
3D Unification 40
3E The PT algorithm 44
4 Type assignment with equality 52
4A The equality rule 52
4B Semantics and completeness 57

5 A version using typed terms 63

5A Typed terms 63
5B Reducing typed terms 67
5C Normalization theorems 71

6 The correspondence with implication 74

6A Intuitionist implicational logic 74
6B The Curry-Howard isomorphism 79
6C Some weaker logics 85
6D Axiom-based versions 88

7 The converse principal-type algorithm 93

7A The converse PT theorems 93
7B Identifications 95

vii
Vlll Contents
7C The converse PT proof 96
7D Condensed detachment 102

8 Counting a type's inhabitants 108

8A Inhabitants 108
8B Examples of the search strategy 114
8C The search algorithm 118
8D The Counting algorithm 124
8E The structure of a nf-scheme 127
8F Stretching, shrinking and completeness 132

9 Technical details 140

9A The structure of a term 140
9B Residuals 144
9C The structure of a TA2-deduction 148
9D The structure of a type 151
9E The condensed structure of a type 153
9F Imitating combinatory logic in A-calculus 157

Answers to starred exercises 161

Bibliography 169
Table of principal types 177
Index 179
 Introduction

This book is not about type theories in general but about one very neat and special
system called "TA" for "type-assignment". Its types contain type-variables and
arrows but nothing else, and its terms are built by A-abstraction and application
from term-variables and nothing else. Its expressive power is close to that of the
system called simple type theory that originated with Alonzo Church.
TA is polymorphic in the sense that a term can have more than one type, indeed
an infinite number of types. On the other hand the system has no V-types and
hence it is weaker than the strong polymorphic theories in current use in logic and
programming. However, it lies at the core of nearly every one of them and its
properties are so distinctive and even enjoyable that I believe the system is worth
isolating and studying on its own. That is the aim of this book. In it I hope to try
to pass on to the reader the pleasure the system's properties have given me.
TA is also an excellent training ground for learning the techniques of type-theory
as a whole. Its methods and algorithms are not trivial but the main lines of most
of them become clear once the basic concepts have been understood. Many ideas
that are complicated and tedious to formulate for stronger type-theories, and many
complex techniques for analysing structures in these theories, appear in TA in a very
clean and neat stripped-down form. This book will take advantage of this neatness
to introduce some of the most important type-theoretic techniques with particular
emphasis on explaining why things happen the way they do.
Thus the reader who learns the basic techniques of type-theory from TA will
acquire a very good foundation for the study of other type-systems.
Type theories in general date back to the philosopher Bertrand Russell and
beyond.' They were used in the early 1900's for the very specific purpose of getting
round the paradoxes that had shaken the foundations of mathematics at that time,
but their use was later widened until they came to be part of the logicians' standard
bag of technical tools, especially in proof-theory. (Their use in combinatory logic
dates back to Curry 1934 and in A-calculus to Church 1940.) However, they remained
a relatively specialist tool until around the 1970's.
About that time the need for stronger programming languages brought type-
theories to the attention of computer scientists, and several of the new languages
developed in the 1970's and 80's were built on a type-theory base. These languages

1 See for example Russell 1903 Appendix B and the comments in Gandy 1977 and Church 1976.

ix
x Introduction

have proved themselves in many applications and have now become well established
in the research community (and are even becoming known outside it!). The chief
example is ML, developed at Edinburgh University by the group led by Robin
Milner, but others include HOL (Cambridge University), Miranda (Regd trademark,
Research Software Ltd.) and Nuprl (Cornell University).
The system TA is, with slight modifications, a common part of all of these. Indeed,
in its early days it was studied mainly as a prelude to studies of stronger systems
and this is the way it was treated in Curry and Feys 1958.1 But from the 1960's
onward it gradually became clear that TA was not as trivial as it had at first seemed
and was worth isolating and studying in its own right. Natural questions about
TA turned out to be much harder to answer than expected. Their answers are not
completed even today, but from them have come some very interesting techniques
that have had applications elsewhere, such as type-checking algorithms and filter
2-models.
In fact more is now known about TA than can fit into a book of reasonable
length. The present book will therefore be very selective. Although covering all the
main basic properties of TA it will focus on the following three algorithms. The first
is well known but the other two are scarcely known at all and their consequences are
still not nearly fully understood, though they were first discovered well over fifteen
years ago.
(1) Type-checking or principal-type algorithm. This algorithm is the core of the
type-checking algorithm used in ML. It takes a 2-term and decides whether a type
can be assigned to it and, if so, outputs the most general such type (the principal
type of the term). The method behind the algorithm was sketched in Curry and
Feys 1958 and versions of the algorithm itself have appeared in Morris 1968, Curry
1969, Hindley 1969 and Milner 1978.
(2) Converse principal-type algorithm. This algorithm takes a closed A-term M
and any type i which can be assigned to M, and outputs a closed 2-term M` such
that T is the principal type of M. Via the formulae-as-types correspondence with
propositional logic it leads to a completeness proof for a variant of the Resolution
rule called the rule of condensed detachment in a system of implicational logic. In
fact several converse principal-type algorithms are known, each producing an M'
with slightly different properties and giving completeness for a slightly different
logic. (Hindley 1969, Meyer and Bunder 1988, Mints and Tammet 1991.)
(3) Inhabitant-counting algorithm. A normal inhabitant of a type i is a closed
A-term M in #-normal form to which 'r can be assigned. The counting-algorithm
takes a type r and outputs the number of its normal inhabitants (0, 1, 2,... or
infinity, modulo changes of bound variables), and then lists these one by one; in
particular it decides in a finite time whether this list will be infinite or not. It is
like the known algorithm for deciding whether a regular language is infinite but
with extra procedures to deal with bound variables. It is also like known algorithms
for deciding provability in Intuitionist propositional logic but with extra procedures

Barendregt 1992 is a good survey of type-theories which shows the relations between TA2 and others;
see especially §3.1 where TA,t is called 2-*-Curry. Other modern introductions to type theories in
general are Andrews 1986, Girard et al. 1989, Krivine 1990, Mitchell 1990 and 1996, Constable 1991,
Gallier 1993, Nerode and Odifreddi 199-, Scedrov 1990.
 Introduction xi

added to do the counting. It originated in Ben-Yelles 1979 though part of the proof
that it works is due to Hirokawa 1993c.
Each of the above algorithms will be presented in full with a proof of correctness
included.

Acknowledgements

I am very grateful to all who have helped in producing this book, especially the
following.
For invitations to give the lecture-courses from which this book has (very slowly)
developed: Institut d' Informatique, Universite de Tizi-Ouzou, Algeria (1988), and
the U.K. Science and Engineering Research Council's "Logic for I.T." Initiative
(1990).
For a period of reduced teaching duties without which the book would not have
been possible: my colleagues in the Mathematics Department, University of Wales
Swansea.
For finance and accommodation which facilitated some of the meetings and
discussions involved: the European Community's "Esprit" Basic Research Actions
3230 and 7232, the Australian Research Grants Scheme, and the Mathematics
Department of the University of Wollongong.
For useful criticisms, discussions, and helpful suggestions: Yohji Akama, Choukri-
Bey Ben-Yelles, Martin Bunder, Naim cakman, Felice Cardone, Mariangiola
Dezani-Ciancaglini, Roy Dyckhoff, Fritz Henglein, Sachio Hirokawa, Hans Leiss,
Mohamed Mezghiche, Seref Mirasyedioglu, Gordon Plotkin, Adrian Rezus, Jon
Seldin, Masako Takahashi, Anne Troelstra, Werner Wolff, Marek Zaionc and
anonymous referees. (Any errors remaining in the work are my own responsi-
bility, however.)
For perseverance: the staff of Cambridge University Press, especially David
Tranah.
And last but not least, for a healthy combination of support and cynicism, my
wife Carol.

Notes on the text

Chapter 1 contains a very short summary of all the basic facts about A-calculus
needed in this book, though the reader is assumed to have met A-calculus before.
Further information can be found in standard text-books; for example (in English)
Barendregt 1984, Curry and Feys 1958, Hindley and Seldin 1986, Revesz 1988,
Hankin 1994, or (in French or English) Krivine 1990, or (in Japanese) Takahashi
1991.
Chapter 9 at the end of the book collects together some technical details that are
needed in the correctness-proofs of the main algorithms: readers who prefer to omit
these proofs can omit this chapter too.
Some exercises are provided in the text. Answers to starred ones are at the end of
the book.
Throughout this book "HS 86" refers to Hindley and Seldin 1986.
 1

The type-free A-calculus

The R-calculus is a family of prototype programming languages invented by a

logician, Alonzo Church, in the 1930's. Their main feature is that they are higher-
order ; that is, they give a systematic notation for operators whose input and output
values may be other operators. Also they are functional, that is they are based on
the notion of function or operator and include notation for function-application and
abstraction.
This book will be about the simplest of these languages, the pure A-calculus, in
which R-terms are formed by application and abstraction from variables only. No
atomic constants will be allowed.

1A d-terms and their structure

IAI Definition (2-terms) An infinite sequence of term-variables is assumed to be
given. Then linguistic expressions called A-terms are defined thus:
(i) each term-variable is a 2-term, called an atom or atomic term;
(ii) if M and N are A-terms then (MN) is a A-term called an application;
(iii) if x is a term-variable and M is a A-term then (Ax M) is a A-term called an
abstract or a A-abstract.
A composite R-term is a R-term that is not an atom.

1A1.1 Notation Term-variables are denoted by "u", "v", "w", "x", "y", "z", with or
without number-subscripts. Distinct letters denote distinct variables unless otherwise
stated.
Arbitrary R-terms are denoted by "L", "M", "N", "P", "Q", "R", "S", "T", with
or without number-subscripts. For "2-term" we shall usually say just "term".
Syntactic identity: "M = N" will mean that M is the same expression as N (if M
and N are terms or other expressions). But for identity of numbers, sets, etc. we
shall say "=" as usual.
Parentheses and repeated A's will often be omitted in such a way that, for example,
Axyz-M = (2x-(2y-(1z-M))), MNPQ - (((MN)P)Q).
(The rule for restoring parentheses omitted from MNPQ is called association to the
left.)

1
2 1 The type free A-calculus
1A2 Definition The length, IMI, of a A-term M is the number of occurrences of
variables in M; in detail, define
Ixl = 1, IMNI = IMI + INI, 1 + IMI.

1A2.1 Example 5.

1A3 Definition (Subterms) The subterms of a term M are defined by induction on

IMI as follows:
(i) an atom is a subterm of itself;
(ii) if M Ax-P, its subterms are M and all subterms of P ;
(iii) if M P1P2, its subterms are all the subterms of P1, all those of P2, and M
itself.

1A3.1 Example If M = its subterms are x, y, yx, x(yx),

Az-x(yx) and M itself. (But not z.)

1A4 Notation (Occurrences, components) A subterm of a term M may have more

than one occurrence in M; for example the term
(Ax yx)(Az . x(yx))

contains two occurrences of yx and three of x. The precise definition of "occurrence"

is written out in 9A2, but the reader who already has a good intuitive idea of the
occurrence-concept will go a long way without needing to look at this definition.
In this book occurrences will be underlined to distinguish them from subterms;
for example we may say
"Let P be any occurrence of P in M".
An occurrence of Ax will be called an abstractor, and the occurrence of x in it
will be called a binding occurrence of x.
All the occurrences of terms in M, other than binding occurrences of variables,
will be called components of M.

1A5 Definition (Body, scope, covering abstractors) Let be a component of a

term M. The displayed component P is called the body of or the scope of the
abstractor Ax.
The covering abstractors of a component R of M are the abstractors in M whose
scopes contain R.

1A6 Definition (Free, bound) A non-binding variable-occurrence x in a term M is

said to be bound in M if it is in the scope of an occurrence of Ax in M, otherwise
it is free in M.
A variable x is said to be bound in M if M contains an occurrence of Ax; and x
is said to be free in M if M contains a free occurrence of x. The set of all variables
free in M is called
FV(M).
 1A A-terms and their structure 3

1A6.1 Warning Two distinct concepts have been defined here, free/bound occur-
rences and free/bound variables. A variable x may be both free and bound in M,
for example if M - x in M cannot be both
free and bound.
Also note that x is said to be bound in even though its only occurrence there
is a binding one.

1A7 Definition (Substitution) Define [N/x]M to be the result of substituting N for

each free occurrence of x in M and making any changes of bound variables needed
to prevent variables free in N from becoming bound in [N/x]M. More precisely,
define for all N, x, P, Q and all y # x
(i) [N/x]x N,
(ii) [N/x]Y = Y,
(iii) [N/x](P,Q) = (([Nlx]P)([Nlx]Q)),
(iv) Ax-P,
(v) [Nlx](AY'P) = fx FV(P),
(vi) if x E FV(P) and y V FV(N),
(vii) if x E FV(P) and y E FV(N).
(In (vii) z is the first variable in the sequence given in 1A1 which does not occur
free in NP.)

1A7.1 Notation (Simultaneous substitution) For any N1,...,Nn and any distinct
xl,...,xn, the result of simultaneously substituting N1 for xj,N2 for x2,... in M, and
changing bound variables to avoid clashes, is defined similarly to [N/x]M. (For a
neat definition see Stoughton 1988 §2.) It is called

[Ni l x 1, ... , Nn/xn]M.

1A8 Definition (Changing bound variables, a-conversion) Let y FV(M); then we

say

(a) Ax'M =« AY' [Ylx]M,

and the act of replacing an occurrence of in a term by is called a
change of bound variables. If P changes to Q by a finite (perhaps empty) series of
changes of bound variables we say P a-converts to Q or

P=a Q.

1A8.1 Note Some basic lemmas about a-conversion and substitution are given in
HS 86 §1B. Two simple properties that will be needed here are
(i) P Q IPI = IQI,
(ii) P Q FV(P) = FV(Q)

1A9 Definition A term M has a bound-variable clash if M contains an abstractor

Ax and a (free, bound or binding) occurrence of x that is not in its scope.
4 1 The type free A-calculus
Examples of terms with bound-variable clashes are
Ax-Ay-Ax-N,

We shall be mainly interested in terms without such clashes.

1A9.1 Lemma Every term can be a-converted to a term without bound-variable clashes.

Proof By the lemmas in HS 86 §1B.

1A10 Definition (Closed terms) A closed term or combinator is a term in which no

variable occurs free.

1A10.1 Example The following closed terms will be used in examples and results
throughout this book.
B Axyz-x(yz), B' Axyz-y(xz),
C I Ax-x,
K S
W AxY'xyy, Y = Ax-(AY'x(YY))(AY'x(YY)),
0 - 1 -
n (nx's applied toy)
(Y is Curry's fixed-point combinator, see HS 86 Ch.3 §3B for background; the terms
n are the Church numerals for n = 0, 1, 2,..., see HS 86 Def. 4.2.)

1B fl-reduction and #-normal forms

This section outlines the definition and main properties of the term-rewriting pro-
cedure called Q-reduction. Further details can be found in many other books, for
example HS 86 Chs. 1-6 and Barendregt 1984 Chs. 3 and 11-14.

1B1 Definition (#-contraction) A l3-redex is any term its contractum is

[Nlx]M and its re-write rule is
Dip [N/x]M.
Iff P contains a /3-redex-occurrence R and Q is the result of replacing
this by [N/x]M, we say P #-contracts to Q (P r>lp Q) and we call the triple (P, R Q)
a #-contraction of P.

1B1.1 Lemma P r.1Q Q FV(P) ? FV(Q).

1B2 Definition (#-reduction) A #-reduction of a term P is a finite or infinite sequence

of #-contractions with form
(i) (P1, R1, Q1), (P2, R2, Q2), .. .
where Pl -a P and Q; -a Pi+1 for i = 1,2,.... (The empty sequence is allowed.) We
say a finite reduction is from P to Q if either it has n >_ 1 contractions and Q" -a Q
 I B #-reduction and #-normal forms 5

or it is empty and P -« Q. A reduction from P to Q is said to terminate or end at

Q. Iff there is a reduction from P to Q we say P #-reduces to Q, or
P 1>S Q.

Note that a-conversions are allowed in a #-reduction.

1B3 Definition The length of a fl-reduction is the number of its #-contractions (finite
or oo). A reduction with maximal length is one that continues as long as there are
redexes to contract (i.e. one that either is infinite or ends at a term containing no
redexes).

1B4 Definition (#-conversion) Iff we can change P to Q by a finite sequence of

#-reductions and reversed #-reductions, we say P #-converts to Q, or P is #-equal
to Q, or
P=RQ.
A reversed #-reduction is called a #-expansion.

1B4.1 Exercise For every term F let XF - YF where Y is the fixed-point combinator
defined in 1A10.1; show that
FXF =, XF.

1B5 Church-Rosser Theorem for f (i) If M ># P and M >p Q (see Fig. 1 B5a) then
there exists T such that
P 1'p T, Q ># T.

(ii) If P =# Q (see Fig. IB5b) then there exists T such that

P 'fl T, Q 'p T.

Proof of 1B5 (i) See HS 86 Appendix 1 or Barendregt 1984 §3.2. (ii) This is deduced
from (i) as suggested in Fig. IB5b.
6 I The type free .1-calculus

P \ /\ /,O\\ %
Q

1B6 Definition (#-normal forms) A fl-normal form is a term that contains no /3-
redexes. The class of all fl-nf's is called /3-nf. We say a term M has /3-nf N
if
Mr>jNandNE/3-nf.

1B6.1 Note Roughly speaking, a reduction can be thought of as a computation and

a fl-nf as its result. One main aim when designing a type-theory is to give it the
property that every computation can be pursued to a result if the operator wishes,
i.e. that every term with a type has a fl-nf. This gives normal forms even more
significance in a type-theory than they already have in a type-free theory.
(Terms in general do not necessarily have /3-nf's of course. The simplest term
without one is

1B7 NF-Uniqueness Lemma Modulo a-conversion, a term M has at most one f3-nf.

Proof An easy application of the Church-Rosser theorem.

1B7.1 Notation If M has a fl-nf it will be called M*p.

1B8 Definition (Leftmost reductions) The leftmost /3-redex-occurrence in a term P is

the fl-redex-occurrence whose leftmost parenthesis is to the left of all the parentheses
in all the other fl-redex-occurrences in P.
The leftmost /3-reduction of a term P is a f3-reduction of P with maximal length,
say
(P1, R1, Q1), (P2, Rz, Q2),

such that R. is the leftmost fl-redex-occurrence in Pi for all i >_ 1 (and P, a-converts
to P and Pi+1 or-converts to Q; for all i >_ 1).

1B9 Leftmost-reduction Theorem A term M has a f3-nf M* fl iff the leftmost I

reduction of M is finite and ends at M*#.
 IC rl- and firs-reductions 7

Proof See Curry and Feys 1958 §4E Cor. I.I. (In fact this result is an immediate
corollary of a slightly deeper result called the standardization theorem ; for the latter
see Curry and Feys 1958 §4E Thm. 1 or Barendregt 1984 Thm. 11.4.7, or the
particularly clear proof in Mitschke 1979 Thm. 7.)

1B9.1 Example The leftmost reduction of the fixed-point combinator Y in 1A10.1 is

easily seen to be infinite, so Y has no f3-nf.

1B9.2 Note (Seeking f3-normal forms) The leftmost reduction of a term M is com-
pletely determined by M, so by 1B9 it gives an algorithm for seeking M*#: if M*Q
exists the leftmost reduction of M will end at M*#, and if not, this reduction will
be infinite. Of course this algorithm does not decide in finite time whether M has a
fl-nf; and in fact this cannot be done, as the set of terms with normal forms is not
recursive. (See e.g. HS 86 Cor 5.6.2 or Barendregt 1984 Thm. 6.6.5.)

1B10 Lemma (Structure of a /3-normal form) Every f3-nf N can be expressed

uniquely in the form
(i) N- (m > 0, n > 0),
where N1....,Nn are 13-nf's. And if N is closed then y E {xl,...,x,n}.

Proof Easy induction on INI. Note the uniqueness.

11310.1 Note The following special cases of 11310 are worth mention:
m = n = 0: N y (an atom);
m = 0, n > 1: N = yNl ... Nn (an application);
m 1: N AxI ... (an abstract);
m > 1, n = 0: N Axe ... xn, - y (called an abstracted atom).

11310.2 Exercise Prove that f3-nf is the smallest class of terms satisfying (i) and (ii)
below:
(i) all variables are in fl-nf;
(ii) for all m,n>_ 0 with m+n > 1, and all
N1,...,Nn E f3-nf E f3-nf.

1C I- and #I-reductions
This section sketches the most basic properties of rl- and frl-reductions. For more
details see HS 86 Ch. 7 and Barendregt 1984 §15.1.

1C1 Definition (q-reduction, I -conversion) An u-redex is any term Ax Mx with

x FV(M); its re-write rule is
Ax-Mx >11 M.
Its contractum is M. The definitions of q-contracts, q-reduces (>,), u-converts (=h),
etc. are like those of the corresponding /3-concepts in 1B.
8 1 The type-free A-calculus
1C2 Lemma All q-reductions are finite; in fact an n-reduction P r7 Q must have
length < JP1/2.

Proof Each q-contraction reduces IPA to IPA -2.

1C3 Definition The qfamily {P}7 of a term P is the set of all terms Q such that
P -7 Q.

1C3.1 Note By 1C2, {P},, is finite.

1C4 Church-Rosser Theorem for I If P =,, Q then there exists T such that
P>7T, Qr>7T.

Proof Straightforward. (Barendregt 1984 Lemma 3.3.7.)

1C5 Definition (#I-reduction, fq-conversion) A jJq-redex is any /3- or q-redex. The

definitions of #9-contracts, jtq-reduces (r'p7), /q-converts (=p7) are like those of the
corresponding #-concepts in 113.

1C5.1 Lemma P >p7 Q FV(P) 2 FV(Q).

1C5.2 Note A Jrl-reduction may have a-steps as well as /3 and q. The following
theorem says that all its rl-steps can be postponed to the end of the reduction.

1C6 q-Postponement Theorem If M Dp7 N then there exists a term P such that
M Np P r7 N.

Proof Nederpelt 1973 Thm. 7.28 or Barendregt 1984 Cor. 15.1.6.

1C7 Commuting Lemma If M Np P and M r'7 Q (see Fig. 1 C7a) then there exists
a term T such that
P D7 T, Q Np T.

Proof Barendregt 1984 Lemma 3.3.8.

1C7.1 Corollary If M Dp7 P and M Np Q then there exists a term T such that
P Pp T, Q Np7 T.

Proof By 1B5, 1C4 and 1C7.

1C8 Church-Rosser Theorem for /q (i) If M Np7 P and M Dp7 Q then there exists
T such that
P Dd7 T, Q t>#7 T.
 IC q- and iq-reductions 9

3T
Fig. lC7a.

(ii) If P =p, Q then there exists T such that

P DBn T, Q >p, T.

Proof (i) From 1B5, 1C4, 1C6, 1C7. (ii) From (i) as in Fig. 1B5b.

1C9 Definition (Jiq- and q-normal forms) A ft-normal form ($q-nf) is a term
without frl-redexes. The class of all #q-nf's is called /Jq-nf. We say M has fq-nf N'
if
MDRaN, NE f3j-nf.
Similarly we define 1-normal form, q-nf, and M has q-nf N.

1C9.1 Notation The f q-nf and q-nf of a term M are unique modulo =-,, by the
Church-Rosser theorems for fn and q ; they will be called
M*Pn, M*?1*

1C9.2 Lemma (i) An q-reduction of a fl-nf cannot create new f3-redexes; more precisely

M E fl-nf and M >n N N E/3-nf.

(ii) For every M,M*#,, is the q-nf of M*p; i.e. M*p, - (M*#)*,,.

Proof (i) It is easy to check all possible cases. (ii) By 1C2, M*fl has an q-nf (M*#)*,,
and this is a f3q-nf by (i).

1C9.3 Corollary If N is a fl-nf then all the members of its ?I-family are fl-nf's and
exactly one of them is a fq-nf, namely N*,,.

1C9.4 Lemma A term has a $q-nf iff it has a f3-nf.

Proof For "only if", see Curry et al. 1972 §11E Lemma 13.1 or Barendregt 1984
Cor. 15.1.5. For "if", see 1C9.2. (By the way, do not confuse the present lemma with
a claim that a term is in fl-nf if it is in Pq-nf, which is of course false!)
10 1 The type free A-calculus
1C9.5 Note (Seeking $q-normal-forms) To seek for M*f,, reduce M by its leftmost
/3-reduction. If this is finite, it must end at M*p and then the leftmost 11-reduction
will reach an rl-nf in < IM*f 1/2 steps, by 1C2. If the leftmost #-reduction of M is
infinite, M*s does not exist and hence by 1C9.4 neither does M*#,,. Of course this
procedure does not decide in finite time whether M*R, exists; see the comment in
1 B9.2.

1D Restricted A-terms

The following restricted classes of A-terms will play a role later in the correspondence
between type-assignment and propositional logic.

ID1 Definition (Al-terms) A 2-term P is called a 2l--term if, for each subterm with
form in P, x occurs free in M at least once.

1D1.1 Note The 1I-terms are the terms that were originally studied by Church.
They have the property that if a Al-term has a normal form, so have all its
subterms (Church 1941 §7, Thm. 7 XXXII). Church restricted his system to 2I-terms
because he regarded terms without normal forms as meaningless and preferred that
meaningful terms did not have meaningless subterms. The 21-terms are discussed in
detail in Barendregt 1984 Ch. 9.
The standard example of a non-Al-term is K =

1D1.2 Notation Sometimes unrestricted A-terms are called 2K-terms, and the unre-
stricted A-calculus the 2K-calculus, to contrast with 2I-terms and to emphasise the
absence of restriction.

1D2 Definition (BCKA-terms) A BCKA-term is a A-term P such that

(i) for each subterm of P, x occurs free in M at most once,
(ii) each free variable of P has just one occurrence free in P.

1D2.1 Examples Of the terms in the list in 1A10.1 the following are BCKA-terms:
B= B' C
I K n= (n = 0 or 1).
And the following are not:
S Axy xyy,
n = Axy x"y (n >_ 2).

1D2.2 Lemma The class of all BCKA-terms is closed under abstraction, i.e. if M is a
BCKI-term then so is Ax M for every variable x.
.

Proof By 1D2(ii), x occurs free at most once in M.

1D2.3 Notes (i) In contrast to the above lemma the class of all AI-terms is only
closed under abstractions such that x occurs free in M.
 1D Restricted A-terms 11

(ii) The BCKA-terms are so called because the closed terms in this class correspond
to combinations of three combinators called "B", "C" and "K" in combinatory logic
(see 9F for details). They have also sometimes been called linear A-terms but this
name is nowadays usually applied to the following class.

1D3 Definition (BCIA-terms) A BCIA-term or linear A-term is a A-term P such that

(i) for each subterm of P, x occurs free in M exactly once,
(ii) each free variable of P has just one occurrence free in P.
Clearly every BCIA-term is a BCKA-term, but the BCKA-term K is not a BCI1.-
term; in fact a term is a BCIA-term if it is both a Al-term and a BCKA-term. The
closed BCIA-terms correspond to combinations of the combinatory called B, C and
I in combinatory logic; details are in 9F.

1D4 Lemma Each of the three classes (Al-terms, BCKA-terms and BCIA-terms) is
closed under f q-reduction, i.e. every term obtained by flu-reducing a member of the
class is also in the class.

Proof Straightforward.

1D5 Definition A f3-contraction [N/x]M is said to cancel N if x does

not occur free in M; it is said to duplicate N if x has at least two free occurrences
in M.
A #-reduction is non-duplicating if none of its contractions duplicates; it is
non-cancelling if none cancels.

1D6 Lemma Every f3-reduction of a AI-term is non-cancelling; every one of a BCKA-

term is non-duplicating, and every one of a BCIA-term is both.
 2

Assigning types to terms

The topic of this book is one of the simplest current type-theories. It was called TA
in the Introduction but in fact it comes in two forms, TAC for combinatory logic
and TA2 for A-calculus. Since most readers probably know A-calculus better than
combinatory logic, only TAx will be described here. (The reader who wishes to see
an outline of TAc can find one in HS 86 Ch.14; most of its properties are parallel
to those of TA2.)
The present chapter consists of a definition and description of TA2. It is close to
the treatment in HS 86 Ch. 15 but differs in some technical details.

2A The system TAt

2A1 Definition (Types) An infinite sequence of type-variables is assumed to be given,
distinct from the term-variables. Types are linguistic expressions defined thus:
(i) each type-variable is a type (called an atom);
(ii) if o and -r are types then (o->r) is a type (called a composite type).

2A1.1 Notation Type-variables are denoted by "a", "b", "c", "d", "e", "f", "g", with
or without number-subscripts, and distinct letters denote distinct variables unless
otherwise stated.
Arbitrary types are denoted by lower-case Greek letters except "A".
Parentheses will often (but not always) be omitted from types, and the reader
should restore omitted ones in such a way that, for example,

P->a-t = (P-(Q->r)).
This restoration rule is called association to the right.'

2A1.2 Informal interpretation To interpret types we think of each type-variable as

a set and a--+T as a set of functions from o into T. The precise nature of this set
of functions (all functions, all functions definable in some given system, etc.) will
depend on the particular interpretation we may have in mind.
It is the opposite of the rule for terms! The reason originates in the fact that terms follow the common
notation convention in which an operator's input is written on the right, but in a type o-t the type
of the input is on the left of that of the output.

12
 2A The system TAA 13

2A2 Definition The total number of occurrences of type-variables in a type T will

be called ITI or the length of T; more precisely, define

Ial 1, IP-*al IPI + lal.

The number of distinct type-variables occurring in T will be called
IITII

and the set of all these variables will be called

Vars(T).

2A2.1 Example If T = (a-+b-*c)--+(a-*b)--+a-*c, then

ITI = 7, IITII = 3, Vars(T) = {a,b,c}.

The structure of an arbitrary type is analysed in detail in 9D-E. The lemmas there
will be used in some later chapters but not in this one.

2A3 Discussion (The Church and Curry approaches) In current use there are two
main ways of introducing types into A-calculus, one attributable to Alonzo Church
and the other to Haskell Curry.
The former goes back to a type-system introduced in Church 1940. In it, the
definition of "A-term" is restricted by giving each term a unique type as part of
its structure and saying that an application PQ is only defined when P has a
function-type a--+T and Q the appropriate argument-type a.1
The effect of Church's restriction can be seen on which is a well-formed
term in type-free A-calculus but represents the abstract concept of self-application, a
concept whose meaningfulness may well be questioned. Self-application was involved
in most of the paradoxes that were discovered in mathematics in the early 1900's,
and Bertrand Russell devised the first of all type-theories specifically as a language
in which these paradoxes could not be expressed. In Church's typed A-calculus each
variable has a unique type, so if x has a function-type a-->T it cannot also have type
a, and so the application xx cannot be defined as a typed term. Hence also Ax xx
cannot be a typed term.
Curry took a different approach. He pointed out that if we wish to ask questions
about the meaningfulness of then we need a language in which these questions
can be expressed. And Church's type-theory by itself is not adequate for this, because
we have just seen that is excluded from it. Curry proposed a language which
would include all the type-free A-terms, and a type-theory which would contain rules
assigning types to some of these terms but not to others. The term would not
be given a type by these rules, but would still remain in the system and hence be
discussable. (Curry and Feys 1958 §0B, p.5.)
Along with this change Curry proposed another, which is best understood by
looking at the identity-combinator as an example.
In Church's type-theory there is no term Ax-x. Instead, for each type o there is
For a definition of typed A-term and a few examples see HS 86 §13A; for another version, with
motivation and more details, see Barendregt 1992 §3.2.
14 2 Assigning types to terms
a variable x° with type o and a term with type Informally, this term
denotes the identity function on whatever set S may be denoted by a. Call this
function Is; the only objects it accepts as inputs are members of S, and Is(x) = x for
all x E S. Thus Church's theory has an infinite number of identity functions, one for
each set S. This agrees with the view of functions taken by most mathematicians:
each function is seen as a set of ordered pairs with a domain and range built into
its definition, and the identity functions Is and IT on two distinct sets S and T are
viewed as different functions.
But this view is not entirely satisfying; an alternative and perhaps more natural
view is to see all the separate identity-functions Is, IT, etc. as special cases of one
intuitive concept, the operation of doing nothing. If we admit that such a concept
exists, even though only in an imprecise sense, then a type-theory that tries to make
it precise by splitting it into an infinite number of different special cases at the
beginning will seem at the very best inefficient.
Curry's aim was a type-theory in which the identity-concept would be expressed
by just one term Ax-x, to which an infinite number of types would be assigned
by suitable formal rules. Types would contain variables, and if a term M received
a type T it would also receive all substitution-instances of T. This kind of theory
will be called here a type-assignment theory or a Curry-style type-theory. (It is the
ancestor of polymorphic type-theories.) In contrast, a theory in which each term has
a unique built-in type will be called a typed-term theory or a Church-style theory.'
TA2 will be a type-assignment theory.

2A4 Definition A type-assignment is any expression

M:T

where M is a A-term and T is a type; we call M its subject and T its predicate.

("M: T" should be read informally as "assign to M the type T" or "M has type T"
or "M denotes a member of whatever set T denotes".)

2A5 Definition A type-context r is any finite, perhaps empty, set of type-assignments

F = lxl:pi, ..., xm:pm}

whose subjects are term-variables and which is monovalent or consistent in the sense
that no variable is the subject of more than one assignment. For any such IF define

Subjects(F) = {xi,...,xm}.
i
Curry's and Church's lines of thought were not really as distinct as the above seems to imply. In
particular Church did not ignore the possibility that a single identity-concept might be formalizable
instead of a multitude of particular identity-functions. Indeed his first systems of A-calculus in the
1930's were part of an attempt to formalize exactly this single-identity view of functions in a type-free
theory, and one of the best available expositions of this view is in the introduction to his book Church
1941. Only after his attempt to do this in an extremely general setting proved inconsistent did Church
turn to type-theory and a more restricted approach to functions. Also Curry's type-theories began
their development in some of his earliest work and were not simply a response to Church's; see Curry
1934.
 2A The system TA, 15

2A5.1 Notation The result of removing from r the assignment whose subject is x (if
r has one) is called

(If x 0 Subjects(F) we define F - x = F.) The result of removing from F all

assignments xi:pi with xi 0 FV(M) (where M is a given term) is called

FPM

or "r restricted to M". And F is called an "M-context" (for a given M) if

Subjects(F) = FV(M).

2A5.2 Note A type-context F is a set, not a sequence. Hence it does not change
when its members are permuted or repeated. To implement TA2 as a practical
system we would have to represent F by an expression in some language and
include rewrite-rules to permute F's members and make and remove repetitions.
Such rules would obscure the main themes of this book so they have been avoided
here by simply assuming that contexts are sets)

2A6 Definition We say Ft is consistent with F2 if Fl UI72 is consistent; and Ft,..., Fn

are mutually consistent if their union is consistent.

2A7 Definition (TA2-formulae) For any F, M and T the triple (F, M, T) is called a
TAx formula and is written as
F --* M :'r
(or just --+ M:T when F is empty). We shall call M the subject of this formula and T
its predicate (despite the fact that in general it contains other subjects and predicates
too, namely those of the assignments in F).

2A7.1 Notation The following abbreviations will often be used:

x1:Ul,...,xn:Qn F- M:T for {x1:61,...,xn:6n} I- M:T,

F, Y1:ai,...,Yn:on I- M:T for r U {y1:vt,...,Yn:an} I- M:T.

2A8 Definition (The system TAB,) TA2 has an infinite set of axioms and two
deduction-rules (called (-'.E) or --*-elimination and (-*I) or -'.-introduction), as
follows.
Axioms of TA2: for every term-variable x and every type r, TA,1 has an axiom

x:T '--> x:T.

Type-contexts are also called environments in the literature. They play a different role from the sets
called bases in HS 86 Chs.14-15: there a basis was a set of axioms for a theory, whereas here a context
will be used as a set of assumptions for a particular deduction in a theory.
16 2 Assigning types to terms
Deduction-rules of TAx:

171 H P :(o--->T) r2 Q
(-E) r1 U r2 H (PQ) : T,
[if 171 U r2 is consistent]

F H P :-r
[if r is consistent with x:oj
F-x F-+ (Ax.P):(r->r).

A TAx-deduction A is a tree of TA2-formulae, those at the tops of branches being

axioms and those below being deduced from those immediately above them by a
rule. (A detailed definition of such deductions is given in 9C1.) The bottom formula
in A is called its conclusion; if it is

r --> M :T

we call A a deduction of r H M:T or a deduction of M:T from r, and say that

r f-a M:T is TAx-deductible. In the special case r = 0, A may be called a proof of
the assignment M:T.

2A8.1 Note (Rule (-+I)) The condition in (-*1) that r be consistent with x:a means
that either r contains x:a or r contains no assignment at all whose subject is x. In
the first case the rule is said to discharge or cancel x from F. In the second case it is
said to discharge x vacuously.
In these two cases the rule takes two slightly different forms which may be
displayed as follows (using "r1" below to correspond to "F - x" above).

r1,x:a H P:T
(*l)main [ifx Subjects(r1)]
rl (Ax.P):(Q->T),

[ifx 0 Subjects(rl)]
F, '-+ (Ax.P):(a-*T).

2A8.2 Example Let B - as in the list in 1A10.1; the following is a

deduction of
-+ B:
(In it, "r" will denote the set {x:a-*b, y: c--+a, z:c}.)
y:c-+a F-+ y:c-+a z:c ---> z:c
(-+E)
x:a-+b H x:a->b y:c-+a, z:c i-+ yz:a
(->E)
F (x(yz)):b

r-z -+ (Az-x(yz)):c-->b

r - z - y H (Ayz -x(yz)):(c-+a)
ti (Axyz-x(yz)):(a->b)->(c-*a)-,c-->b
 2A The system TA2 17

2A8.3 Example Let I - Ax-x; the following is a deduction of

H I : a-+a.
x:a F--+ x:a
i--> a-+a
(-'I) ma i n

2A8.4 Example Let K = the following is a deduction of

--+ K:a- b-+a.
x:a --+ x:a
(+I)vac
x:a --+
(-+I)main
-4

2A8.5 Example The following is a deduction of

F--+ II: a-+a.
x:a--+a x:a-+a x:a '--+ x:a (-*1)

(-E)
H
2A8.6 Remark (Self-application) The above example gave a type to a term involving
self-application, namely II. This was done by giving a different type to each of the
two occurrences of I, and to do this we had to give two different types to the one
variable x; but there was no inconsistency problem when (-+E) was applied because
the two applications of (--+I) above (-+E) removed x from the contexts on the left
of " -4 ". Similarly it is possible to give types to several other self-applications in
TA2, for example KK and BB.
This may seem surprising, in view of the claim in 2A3 that the original aim
of a type-theory was to avoid self-application. But in fact the "dangerous" self-
application to be avoided is not any one simple particular case like II, but the overall
general concept of self-application as represented by the term Ax.xx. And 2x.xx
does not receive a type in TAz.
To see this, suppose there were a TA2-deduction of
H
for some r. Then its last step would have to be an application of (-+I) to a deduction
of

x:p I--+ xx:a

for some p and a such that i - p-+6; and the last step in this deduction would
have to be an application of (-+E) to two deductions of
x : al-+6 i--+ x : 61x : 61 H x : 61
for some 61. But 61-+a # at so the consistency condition in (-+E) would be violated
by these deductions.
18 2 Assigning types to terms
Thus the consistency condition in (-->E) prevents from having a type. This
is in fact its main purpose.'

2A8.7 Exercise* Deduce the following in TA2, where B' - C=

S- and W - as in 1A10.1.
(i) --* B': (a-+b)-+(b-->c)-->a-*c,
(ii) --> C: (a-->b-->c)-+b-+a-->c,
(iii) -> S: (a->b-*c)- *(a-*b)--+a-->c,
(iv) H W: (a->a-*b)-a-b.
2A8.8 Exercise* Deduce the following in TA2, where
P - (Avxyz-v(Y(vxz)))I, Q ,.xyz- 1(Y(Ixz));

(i) --> P : (a--+b)-*(b-+a-.b)--.a-.a-->b;

(ii) H Q: (a-*b)-+(b-+c)->a--*c.

2A8.9 Note (Comparison with HS 86) The format of TAx is what is known as the
"Natural Deduction" style and was originated by Gerhard Gentzen in his thesis
Gentzen 1935. The system called "TA2" in HS 86 §15B is another variant of the
same style; its main differences from the above system TA1 are as follows.
(i) In HS 86 the discharging of assumptions by rule (--+I) was shown by enclosing
the assumption in brackets at the top of the deduction-tree. But here the set of
undischarged assumptions at each stage of the deduction is displayed on the left
of the " -4 " symbol and when rule (-*I) is used this set is simply reduced. This
notation is perhaps more explicit than that in HS 86 and is in common use in recent
literature. In both notations deductions have the same tree-structure.
(ii) The version in HS 86 included an a-rule that is not in the present version. This
was to ensure that the set of provable formulae would be closed under a-conversion
even when the basis of axioms was not. But there are no axioms here in the sense of
HS 86 so a-closure will turn out to be provable without adding an a-rule; see 2B6.

2A9 Definition Let r be a type-context. Iff there is a TA2-deduction of a formula

r' '-- M:T for some F' c F we shall say
I' F-,1 M : T.

In the special case r = 0 we shall say M has type T in TA2, ors is a type of M
in TAx, or
F-2 M: T.

The phrase "in TA2" may be omitted when no confusion is likely.

2A9.1 Lemma (Weakening) I' PA M:T, r+ ? I' . 17' PA M:T.

There is at least one interesting type-theory in which this consistency condition is relaxed, the theory
of intersection-types that originated in Coppo and Dezani 1978 and Salle 1978. In this theory xx
receives a type and types play a significantly more complex role than in TA2, see for example the
comment in Hindley 1992 §1.1.
 2A The system TA2 19

Proof Trivial from 2A9. 0

2A9.2 Warning Do not confuse " H " with "F-". The former is part of the language
of TA2 and serves merely to separate two parts of a formula. But "I-A" is part of
the meta-language in which TA,2 is described, and asserts the existence of a TA,2-
deduction; it is the traditional deducibility symbol.' In particular do not confuse
the two statements
(a) the. formula F H M: r is deducible,
(b) F F-2 M:r.
Statement (b) has the weakening property, as we have just seen in 2A9.1. But
(a) does not; the rules of TA2 have been formulated so that if (a) holds then the
subjects of r coincide exactly with the free variables of M, and we cannot modify
a deduction of F F-+ M:T to make a deduction of r+ '--> M:T if r+ ? F.2
The next two lemmas express the above comments more formally.

2A10 Lemma If F H M:2 is deducible in TA2 then Subjects (F) = FV(M).

Proof By an easy induction on lengths of deductions.

2A11 Lemma (i) F I-2 M:r iff Subjects(F) 2 FV(M) and there exists a TA,2-
deduction of the formula F r M --* M:2.
(ii) (3F)(F I-A M:'t) (3F){F is an M-context and F I-2 M: i}.
(iii) For closed terms M,
(317)(F F-,2 M: T) : F-Z M: T.

Proof By 2A9 and 2A10. (The definition of r (M is in 2A5.1.)

2A12 Historical Comment Although this book focuses on A-calculus, type-assignment

in fact began in its sister-theory, combinatory logic. The first systems appeared in
Curry 1934 and were developed further in Curry and Feys 1958 Chs. 8-10, Seldin
1968 and Curry et al. 1972 Ch. 14, though in his earlier work Curry was aiming at
building type-theories with the greatest possible generality and strength and simple
type-assignment formed only a small part of each of these. However, when he found
his strongest system inconsistent in 1954 he turned to the study of weaker ones
and gradually realized that their -*-fragment formed a very neat core system that
was worth studying on its own and stood a good chance of having some practical
value.
This system was essentially a combinatory-logic analogue of TA2: Curry called
it "modified basic functionality" though later in HS86 it was called "TAc". Its basic
properties first appeared in Curry and Feys 1958 §§8C and 9A-F, though even then
those authors' main interest was in a slightly stronger system obtained by adding to
1 Many works use I-" where this book uses " H ", and introduce no special notation for deducibility.
2 Many other versions of Natural Deduction in the literature do not have this restriction but I believe
its use slightly simplifies the proofs of some properties that depend on analysing the structure of a
deduction.
20 2 Assigning types to terms
TAc an equality-invariance rule (see Chapter 4 below). The first papers to feature
TAc exclusively were Curry 1969 and Hindley 1969.
As for TAx itself, its first appearance was in Curry and Feys 1958 §9F1 under the
name "F1(2)T", though none of its properties were stated at that date other than
a theorem relating it to TAc. Its properties as an independent system were not
described until ten years later in the theses Morris 1968 and Seldin 1968. Seldin's
results on TA2 appeared in Curry et al. 1972 §14D, but Morris' thesis was never pub-
lished. The next study devoted to TA1 was another unpublished thesis, Ben-Yelles
1979; some of its material will be the subject of Chapter 8 below.
For subsequent work on TA,1 see the references in Chapters 3-9 below.

2B The subject-construction theorem

Deductions in TAx have one very important property that is not shared by de-
ductions in many more complex type-theories; the tree-structure of a deduction of
F H M:T follows the tree-structure of M exactly. To make this correspondence
precise we really need the detailed definition of construction-tree of a term given in
9A4 and that of a deduction given in 9C1; but the following example gives a very
good idea of what it means.

2B1 Example Let B A deduction of a type-assignment for B was

shown in 2A8.2. If all but the subject is erased from each formula in this deduction
the result is the tree shown in Fig. 2Bla. This is the same as the construction-tree of
B (with certain details called position-labels omitted, for these see the full definition
of construction-tree in 9A4).

y z
x
yz
x(yz)

),z.x(yz)
?,yz.x(yz)
T
kxyz.x(yz)
Fig. 2B l a.

The following theorem describes the deductions-to-terms correspondence formally.

2B2 Subject-construction Theorem (Seldin 1968 §3D Thm. 1, Curry et al. 1972 §14D
Thm. 1.) Let A be a TA2-deduction of a formula 1 --> M:T.
(i) If we remove from each formula in A everything except its subject, A changes to
a tree of terms which is exactly the construction-tree for M.
 2B The subject-construction theorem 21

(ii) If M is an atom, say M x, then F = {x: T} and A contains only one formula,
namely the axiom
x:T --+ X: T.

(iii) If M - PQ the last step in A must be an application of (-+E) to two formulae

with form
r rP --+ P:a-+T, r [ Q F-. Q: a,
for some a.
(iv) If M = Ax-P then T must have form p--+a; further, if x E FV(P) the last step
in A must be an application of (->I)main to
F,x:p F-+ P:a,
and if x FV(P) the last step in A must be an application of (+I)vac to

r F--+ P: a.

Proof Induction on IMI. Parts (i)-(iii) follow immediately from the full definition
of deduction in 9C1. For (iv): if M = then by 2A8.1 the last step in A must
have one of the forms
F,x:p --+ P: or r --+ P : a
r i--* Ax.P:p-+a F i--i
and by 2A10 (->I) main is used when x E FV(P) and (-+I)vac is used otherwise. Hence
result.

2B2.1 Warning (Deductions are not unique) Given M, let A be a deduction of

F M:T. By the subject-construction theorem the structure of M determines both
the tree-structure of A and the terms at all the nodes in A. But this does not mean
that the whole of A is completely determined by its conclusion, because there is
some freedom of choice of the types assigned to terms at non-bottom nodes in A.
For example, let F = 0 and
M (.1xy-y)(Az-z), T =- a-+a

and consider the deduction in Fig. 2B2.1a; the type a in that figure can be arbitrary.

y:a h-+ y:a

(axy-y):a-+a z:a '-+ z:a

(-+1)

Fig. 2B2.1 a.
22 2 Assigning types to terms
However, if M is a normal form or a AI-term this freedom will disappear and
A will be completely determined by M, as we shall see in the next lemma and the
exercise below it.

2B3 Lemma (Uniqueness of deductions for of s) (Ben-Yelles 1979 Cor. 3.2.) Let
M be a /3-nf and A a TAB,-deduction of I' F--> M:T. Then
(i) every type in A has an occurrence in T or in a type in F,
(ii) 0 is unique, i.e. if A' is also a deduction of r H M:T then A' - A.

Proof' Use induction on IMI. The cases M = y and M - are easy. Since M is
a /3-nf, by 1B10 the only other possible case is
M yPl ... Pn
In this case any deduction A of F H M:T must contain an axiom
Y : (PI -P.-T) Y : (P1-'...-'Pn-T),
as well as n deductions 01, ... , On giving
r1 - P1:P1, ... >Fn - Pn:Pn
followed by n applications of (-FE) to deduce
{Y : (P1- ... -->Pn-T)} U F1 U ... U Tn --> (YP1 ... Pn) T.

And F must be
{Y : (P1-... ->P.-'T)} U 171 U ... U Fn-
To prove (i): by part (i) of the induction hypothesis every type in a Ai occurs in
p; or F, and hence in F; also the type of y occurs in F. Hence (i) holds.
To prove (ii): the argument above shows that A' must use the same rules at the
same positions as in A. And the type assigned to y in A' is determined by F and
the assumption that type-contexts are consistent; then the types of P1,...,Pn are
determined by the type of y.

2B3.1 Note (Subformula property) Part (i) of 2B3 corresponds to what is usually
called in logic the subformula property; this says that in a Natural Deduction system
every formula in an irreducible deduction occurs in either the conclusion or an
undischarged assumption. (The correspondence between types and propositional
logic will be fully described in Chapter 6.)
In contrast the TAI-deduction in Fig. 2B2.1a contains a type a that does not
occur in an undischarged assumption or the conclusion.

2B3.2 Exercise* (Uniqueness of deductions for AI-terms) Show that if A is a TA2-

deduction of F F--. M:T and M is a 21-term, then A is unique; i.e. if A' is also a
TA1-deduction of F F-* M:T then A' _ A. (Hint (Thierry Coquand): use 2B3 and
some facts from 2C and 2D below, and the leftmost-reduction theorem (1B9), plus
some thought on the form that a leftmost reduction must have; see the Answers for
details.)
 2B The subject-construction theorem 23

The following three lemmas will be needed in the next section. The first is a
special case of the third but is stated separately because it is needed in the proof of
the third.

2B4 First Substitution Lemma for Deductions Let F - M:T and let [y/x]F be
the result of substituting y for a term-variable x in F. if either of the following holds:
(i) y Subjects(F),
(ii) y and x receive the same type in F,
then
[Y/x]r H2 ([Ylx]M) : T.

Proof First, in both cases (i) and (ii) [y/x]F satisfies the consistency condition for
contexts. Next, by 2A 11 there is a deduction of
F- F+ M : T
for some F- s F with Subjects (F-) = FV(M). Then [y/x]F- is consistent. An
induction on IMI then shows that
[Y/x]F l-z ([Ylx]M) : T.
Then the weakening lemma (2A9.1) gives the result.

2B5 a-Invariance Lemma If F -2 P:T and P -« Q then F I-i Q:T.

Proof [Depends on Section 9C] It is enough to prove the result for one change of
bound variable, say the replacement of a component of P by with
y ¢ FV(M). If P the result follows using 2B4. If is a proper part of
P, use 9C5 (a replacement lemma).

2B6 Second Substitution Lemma for Deductions Let F1 be consistent with F2 and let
F1,x:o i-,, M:T, F2 I-,t N:a
Then
F1 U F2 I-,, [N/x]M : T.

Proof Assume x E FV(M). (If not, the result holds trivially.) By 2B5 we can assume
no variable bound in M is free in xN. In this case [N/x]M is simply the result of
replacing each free x in M by N with no accompanying changes of bound variables.
And by 2A11 we can assume that
Subjects(Fi) U {x} = FV(M),
Subjects(F2) = FV(N).

The result is then proved by a straightforward induction on IMI.

By the way, y might be bound in M. To deal with the case that M = Ay P, it is necessary to use the
induction hypothesis twice and use the fact that I [z/y]PI = IPI.
24 2 Assigning types to terms
2C Subject reduction and expansion
Besides avoiding logical paradoxes another main purpose of type-theories is to avoid
errors of mis-matching in programming. If a term P has a type T we can think
of P as being in some sense "safe". If P represents a stage in some computation
which continues by fn-reducing P, we would like to know that all later stages in
the computation are just as safe as P. The following theorem guarantees this.

2C1 Subject-reduction Theorem (Morris 1968 §4D Thm. 1, Seldin 1968 §3D Thm. 2.)
If F -1 P :T and P >f, Q then
F [-A Q: T.

Proof [Depends on lemmas in 9C] First, by 2A11 there exists a deduction A of

F- t--* P:T for some f- s F with Subjects (f-) = FV(P). By 1C5.1, FV(P)
FV(Q). Hence by 9C5 it is enough to prove the theorem when P is a /3- or ?I-redex
and Q is its contractum.

Case 1: P - (AxM)N, Q = [N/x]M. If x r= FV(M) then by the subject-construction

theorem (2B2) the lower steps of A must have form
f1,x:Q H M:T
(->I)main
f1 H (2x'M):(o->T) f2 -- N:a
(-.E)
171U1-2 '-p

where F1 U F2 = f-. Then 2B6 applied to the deductions for M and N gives
f1Uf2F[N/x]M:T
If x V FV(M) the proof is similar.

Case 2: P = Ax-Mx, Q = M, x FV(M). Then r = p-+Q for some p and o, and

by 2B2 the last steps in A must have form
f- i-* M:p- *a x:p F-. x:p
(-E)
F-, x:p --* Mx:Q
(->I)main
f- -->
Thus A contains a deduction off - H W r as required.

The subject-reduction theorem has a partial analogue for expansion as follows.

2C2 Subject-expansion Theorem If f 1-2 Q:T and P >p Q by non-duplicating and

non-cancelling contractions, then
f 1--2 P :T:

Proof Exercise. This theorem is a special case of Curry et al. 1972 p.315, §14D
Thm. 3 (= Seldin 1968 §3D Thm. 3).
 2C Subject reduction and expansion 25

2C2.1 Corollary If P is a closed BCIA-term and P .p Q then

FI-2P:i Ft-2Q:i

Proof For use 1D6 and 2C2; for "=" use 2C1.
The subject-expansion theorem can be extended to some cancelling contractions
under suitable restrictions. (For example see Curry et at. 1972 §14D Thm. 3 or
Hindley 1989 Thm. 3.3.) But it cannot be extended to arbitrary contractions, as the
following examples show.

2C2.2 Example P >1# Q by a cancelling contraction and Q has a type but P has no
type:
P = ()uv-v)(ilx-xx), Q - 2v-v.
We have I-x Q: a-+a by 2A8.3. But no TA2-deduction has a conclusion with
form H-+ P : z. Because such a deduction would have to contain a deduction
of --+ a for some a and this is impossible by 2A8.6.

2C2.3 Example P tip Q by a duplicating contraction and Q has a type but P has
none:
P = (Ax-xx)I, Q = 11.
We have I-A Q: a--+a by 2A8.5. But P has no type because has none (by 2A8.6).

2C2.4 Example P > p Q by a cancellation, P and Q both have types, but Q has more
types than P :
P - 2xyz-(Au-y)(xz), Q = 1xyz'y.
It is easy to prove that
12 P: (c-*d)->b-*c-+b, I-A Q: a->b-*c-*b;

and an application of the principal-type algorithm (3E1) will show that the types
possessed by P are exactly the substitution-instances of the one shown above, and
similarly for Q. Hence P cannot have the type displayed for Q. (Roughly speaking,
the underlying reason is that x has a function position in P and must therefore be
assumed to have a function-type c--+d; since x does not occur at all in Q the type
of Q has no such limitation.)

2C2.5 Example P >1# Q by a duplication, P and Q both have types, but Q has more
types than P :
P =_ Q

By 2A8.8 we have
I-A P: (a--+b) --- (b--+a--+b)--+a-+a-+b,
I-A Q: (a-+b)-*(b--.c)-+a-->c;
26 2 Assigning types to terms
and an application of the principal-type algorithm (3E1) will show that P cannot
have the type displayed for Q. (The underlying reason is that the two v's in P must
receive the same type whereas the two I's in Q are not so limited.)

2C2.6 Example P rl-contracts to Q, P and Q both have types, but Q has more types
than P:
P - 2xy-xy, Q = AY.Y
It is easy to see that
F-2 P : (a-*b)-*a-*b, F-2 Q : a--+a,

and that a TA2-deduction of H P : a-*a is impossible (because x is in a function

position in P).

2C3 Definition (Types(M)) If M is closed, define Types(M) to be the set of all T

such that F-A M: T.

We shall see in Chapter 3 that if Types(M) is not empty its members are exactly
the substitution-instances of one type, the principal type of M; hence Types(M) is
either empty or infinite.

2C3.1 Lemma Let P be closed. Then

(i) PE>#Q Types(P) c Types(Q),
(ii) if P r'Q Q by a non-cancelling and non-duplicating reduction, then

Types (P) = Types(Q)-

Proof By 2C 1 and 2C2.

2C3.2 Note (Conversion-invariance) Examples 2C2.2-2C2.6 show that we do not

always have
M =# N Types(M) = Types(N).

Even worse, we shall see an example in 7A2.1 where M =p N but

Types(M) fl Types(N) = 0.

Thus Types(M) is very definitely not invariant under conversion.

From a theoretical point of view this seems unsatisfactory. In fact, continuing
the Church-versus-Curry discussion from 2A3, it must be admitted that conversion-
sensitivity of Types(M) is the main disadvantage of a Curry-style type system. In a
Church system the type of M does not change with conversion (because all terms are
typed, including /3-redexes, and this fact restricts the /3-reduction rule and prevents
type-changes). But when we move to a Curry-style system to get the extra expressive
power provided by its polymorphism, we do not get it for free, and the price we pay
is that Types(M) can change with conversion.
 2D The typable terms 27

However, Chapter 4 will describe the effect of adding a new rule to TA,t to
overcome this defect, and it will give theoretical evidence to suggest that perhaps
the price is not so high after all.
In practice too the conversion-sensitivity of Types(M) has turned out to be a
very small problem. Indeed, if one views an assignment M: a-fr as saying that the
application of M to every term with type v is "safe" in some sense, then the most
important practical property of a type-system is the subject-reduction theorem, which
says that if M has type a--+T it will not lose this safety-feature during a reduction.
If Types(M) happens to increase as M is reduced this is not a drawback but simply
means that M is becoming safer. In particular, practical programming languages
like ML and its relatives operate very successfully without conversion-invariance.

2D The typable terms

The system TAx divides the A-terms in a natural way into two complementary
classes: those which can receive types, such as Axyz x(yz), and those which cannot,
such as The former may be regarded as "safe" in the sense that if a term has
a type we know there is a way of assigning types to all its components that avoids
mis-matches of types. The following is a precise definition of this class.

2D1 Definition A term M is called (TAI-) typable or stratified if there exist r and
r such that
r1-2M:T.

2D2 Lemma The class of all TA2-typable terms is closed under the following oper-
ations:

(i) taking subterms (i.e. all subterms of a typable term are typable);
(ii) /3rl-reduction;
(iii) non-cancelling and non-duplicating /3-expansion;
(iv) A-abstraction (i.e. if M is typable so is Ax-M).

Proof (i) by 2B2. (ii) by 2C1. (iii) by 2C2. (iv) by rule (-+I).

2D3 Theorem The class of all TA2-typable terms is decidable; that is, there is an
algorithm which decides whether a given term is typable in TA1.

Proof The principal-type algorithm (3E1) will be a suitable decision-procedure.

2D4 Remark (Normalization) A property that nearly every type-theory in the

literature possesses is the weak normalization (WN) property, which says that every
typable term can be reduced to a normal form. Many type-theories also have the
strong normalization (SN) property, which says that all reductions of a typable term
28 2 Assigning types to terms
are finite. Both WN and SN can be regarded as safety-features of the type-theory
in question: if reductions are viewed as imitating the process of computing values,
WN says that a computation can always be continued to a result if we wish and
SN says that all computations terminate.
The next theorems state precisely the position for TA2.

2D5 Weak Normalization (WN) Theorem (Turing 1942, Curry and Feys 1958, etc.)
Every TA2-typable term has both a f3-nf and a /3n-nf.

Proof See 5C1 and 5C1.1 for a proof (from Turing 1942), and 5C1.2 for historical
notes.

2D5.1 Example By 2D5 the fixed-point combinator Y in lA10.1 is not typable in

TA2; because by 1B9.1, Y has no f-nf.

2D6 Strong Normalization (SN) Theorem (Sanchis 1967, Diller 1968, etc.) if M is
a TA2-typable term, every f3rl-reduction that starts at M is finite.

Proof There are many proofs in the literature besides those of Sanchis and Diller;
for example HS 86 Appendix 2 contains an accessible one for /3 in Thm. A2.3 and
one for /3ri in Thm. A2.4. For references to some others see 5C2.2.

2D6.1 Note Since SN implies WN there is no real need for a separate treatment of
WN. But the Turing proof of WN in 5C1 is both simpler and older than any proof
of SN. Further, most applications of normalization turn out to be of WN rather
than SN. The following are a couple of such applications.

2D7 Theorem There is a decision-procedure for /3-equality of TA2-typable terms; i.e.

an algorithm which, given any typable terms P and Q, will decide whether P Q.
Similarly for & -equality.

Proof Reduce P and Q to their fl-nf's (which exist by WN, and can be found using
leftmost reductions, by 1B9) and see whether they differ.

2D7.1 Note The complexity of the above decision-procedure can be measured in

terms of the Grzegorczyk hierarchy 6°, 6,1, 62, ... of sets of primitive recursive
functions (Grzegorczyk 1953 p.29): in fact Statman 1979b pp. 73-75 points out that
the procedure can be programmed to operate on a Turing machine in 64 time but
no decision-procedure for /3-equality of typable terms can be made to operate in 63
time. (The members of 63 are known as elementary functions.)

2D8 Theorem Every BCK2-term (as defined in I D2) is typable.

Proof Hindley 1989 Thm. 4.1, depending on WN.

 2D The typable terms 29

2D8.1 Note The BCKA-terms are terms without multiple occurrences of variables
(except possibly for binding occurrences), so the above theorem connects untypabil-
ity with multiple occurrences of variables. On the other hand not every term with
multiple occurrences is untypable; consider S Axyzxz(yz) in 2A8.7(iii) for example.
 3

The principal-type algorithm

A typable term has in general an infinite set of types in TAR. For example if I =_
it is possible to assign to I every type with form 6-*6, by the following deduction:

x:a --> x:a

-,I)
(-I)
--* o-- a
(like 2A8.3). But all the types in this infinite set are substitution-instances of the one
type
a-*a,

and it is easy to see that I has no other types than these. (In fact every deduction
for I must have the simple form shown above, by the subject-construction theorem.)
The type a---+a is called a principal type for I.
The aim of the present chapter is to show that the existence of a principal type
is a property of all typable terms, not just I. In effect the principal type of a term
is the most general type it can receive in TAR, and the principal type theorem will
say that every typable term has one. Further, and most important in practice, an
algorithm will be described for finding it.
This algorithm will decide whether a given term M is typable and, if the answer
is "yes", will output a principal type for M. Such algorithms are usually called
type-checking or principal-type or PT algorithms. The existence of a PT algorithm
is what gives TAR and its extensions such as ML their practical value, since if
the typability of a program is regarded as a safety criterion the programmer will
want to be able to decide effectively whether a newly created program satisfies this
criterion.
The PT algorithm below will be easy to describe and even easier to apply in
practice. But to prove the PT theorem a bald statement of the algorithm will not
be enough; we shall need also a proof that the algorithm is correct, i.e. that it does
what it claims to do. In the account below a correctness proof will be included with
the algorithm in the form of comments to each of its steps, explaining the purpose
and effect of each step as the reader meets it.
A little knowledge of substitution, unification and most general unifiers will
be needed so introductions to these will be included before the statement of the
algorithm.

30
 3A Principal types and their history 31

3A Principal types and their history

3A1 Definition (Type-substitutions) A (type-) substitution s is any expression

[at lal,... , Qnl an],

where a1,..., an are distinct type-variables and 61, ... , an are any types. For any T
define s(T) to be the type obtained by simultaneously substituting Q1 for al, ... , an
for an throughout T. In more detail, define
(i) s(a7) = a+,
(ii) s(b) b if b is an atom {al,... an
(iii) §(P-*a) =
We call s(T) an instance of T.

3A1.1 Notation Letters "r", "s", "t", "u", "d" will denote type-substitutions. If
s - [at/al,... , Qn/an] a frequent alternative notation to s(T) will be

[al /al, ... , Qnl an]T.

Recall that the set of all variables occurring in a type T is called Vars(T). The sets of
all type-variables occurring in a finite sequence (T1,...,Tn) of types, or in a deduction
A, are called respectively
Vars(TI...... ,), Vars(A).

3A2 Definition The action of a substitution s is extended to finite sequences of

types, to contexts and to TA,1-formulae thus:

§((21,...92,,)) = (§(TI),.... §(Tn)),

s(F) = {Xi:s(TI),...,Xm:§(2m)} tfF = {X1:T1,...,Xm:Tm},

s(F --> M:T) = s(F) --> M:s(T).

We also extend s to act on deductions A by defining s(A) to be the result of

applying s to every TA2-formula in A. We call s(A) an instance of A. (Similarly for
instances of type-sequences, etc.)

3A2.1 Notes (i) The consistency of F implies that of s(F).

(ii) If A is a TA1-deduction then so is s(A), because the side-conditions in rules
(-*E) and (-+I) in 2A8 remain true after s has been applied. Hence
F E- M:T S(F) E- M:s(T),
and the set of all types assigned to a term M is closed under substitution.

3A2.2 Warning Two distinct concepts of substitution into deductions have now
been mentioned, for term-variables in 2B4 and for type-variables above. Note that
Vars(A) is a set of type-variables not term-variables, and when s is applied to A the
terms in A are completely unchanged.
32 3 The principal-type algorithm
3A3 Definition (Principal types) In TA1, a principal type or PT of a term M is a
type T such that
(i) I I-2 M:T for some F,
(ii) if I'' I-A M:a for some I'' and a then a is an instance of T.

3A3.1 Note By 3A2.1(ii) a type T is a PT of M if, for all types a,

(3F)(F' 1-1 M:a) a is an instance of T.

Thus a PT of M can be thought of as completely characterizing the set of all types

assignable to M.

3A3.2 Notation It will be shown in 3B8.2 that a term's principal type is unique
(modulo substitutions of distinct variables for distinct variables), so we shall often
say "the principal type of M" or
PT(M).

3A4 Definition (Principal pairs) A principal pair for a term M is a pair (F, T) such
that the formula F H M:T is TA1-deducible and every other TA1-deducible formula
r" H M:a is an instance of F M:T.

3A5 Definition (Principal deductions) A principal deduction for a term M is a

deduction A of a formula IF F-. M:T such that every other deduction whose
conclusion's subject is M is an instance of A. To abbreviate "there exists a principal
deduction of r H M:T" we shall say
F I-p M:T.

3A5.1 Notes (i) If A is a principal deduction of r --p M:T then clearly 'r is a principal
type of M and (F, T) is a principal pair for M. In fact the PT theorem will prove that
every typable term has not only a principal type but a principal deduction, and the
PT algorithm will be seen to construct principal deductions as well as types. This
observation will slightly simplify the algorithm's correctness-proof. Thus a typable
term M will be shown to have not only a most general type but a deduction whose
every step is most general.
(ii) Although principal types have been studied since 1969, principal deductions
were almost entirely neglected until about 1990 when the structure of a principal
deduction was characterized by Sachio Hirokawa (Hirokawa 1991a Thms. 1 and
2). This work is beyond the scope of the present book but it is one of the more
interesting recent developments in the study of TAz and has led to new results on
principal types (for example those in Hirokawa 1991b-c, 1993a) and to simplified
proofs of some old results.1
To be precise, Hirokawa 1991a uses a weaker definition of principal deduction than that in 3A5 above:
Hirokawa calls a deduction of r H M:t principal if the formula F M:t is a principal pair in
the sense of 3A4. But his characterization theorems can easily be modified to fit the definition in 3A5
by extending the conditions in them to apply to all type-variables in a deduction, not just those in its
conclusion and undischarged assumptions.
 3A Principal types and their history 33

3A6 Principal Type (PT) Theorem Every typable term has a principal deduction
and a principal type in TA2. Further, there is an algorithm that will decide whether a
given A-term M is typable in TAx, and if the answer is "yes" will output a principal
deduction and principal type for M.

Proof See the PT algorithm and correctness-proof in 3E.

3A6.1 Exercise* Let B - The type assigned to B in Example 2A8.2 was

(a-*b)-> (c-+a)->c-+b.

Using the subject-construction theorem (2B2), show that this type is principal for B.
That is, show that every type assigned to B in TA2 must have form

(P- x)-(t-P)->Q-*t,
where p, U, T are arbitrary types.

3A7 Historical Comment The PT problem, that of deciding whether a term is

typable and finding its principal type if it is, is one that is crucial to many type-
systems, and several different PT algorithms have appeared in the literature over
the years. One of the best known is probably Algorithm W of Damas and Mil-
ner 1982. This was devised originally by Robin Milner for the language ML
which includes the present system and a version first appeared in print in Milner
1978.
But the history of PT algorithms goes back a long way before 1978. There seem to
have been two mainly independent strands of development, one in A-calculus with
combinatory logic and the other in propositional logic. We shall leave the latter
until Chapter 7, but it is almost certainly the older of the two; indeed what amounts
to the core of a PT algorithm was formulated in propositional logic as long ago as
1957 and was actually run on a computer at that date, and the informal idea behind
such algorithms may go back as far as the 1920's (see 7D3).
The first informal use of a PT algorithm in combinatory logic or A-calculus
was probably due to Haskell Curry in the 1950's. Particular examples of type-
computations occur in Curry and Feys 1958 (§9B2, pp.284-293) and show very
clearly the equation-solving method on which several later formal algorithms were
based. But, as mentioned earlier, Curry was mainly interested in stronger systems
at that time and only afterwards began to look at simple type-assignment on its
own. When he did, he wrote up his earlier methods as a formal equation-solving PT
algorithm for TAc and added a proof of its correctness; these survive in his notes
Curry 1966 and were published in Curry 1969.
A PT algorithm and a correctness-proof were also included in an account of TAc
in Hindley 1969, but these differed from Curry's in that the algorithm made explicit
use of the unification algorithm of Robinson 1965 to save itself some work.'
The preparation of Curry 1969 and Hindley 1969 took place in about 1967 with the authors discussing
matters frequently but deliberately keeping their approaches different.
34 3 The principal-type algorithm
On the A-calculus side the first PT algorithm was due to James H. Mor-
ris and was written out with a correctness proof in his thesis Morris 1968.
This was an equation-solving algorithm and like all such, including Curry's, it
automatically did all the unification it needed without calling on an external
algorithm.'
For about ten years after 1968 type-assignment remained a very specialist topic,
then the development of ML by Robin Milner and his group led to a surge of interest
in assignment systems and PT algorithms. Milner's own algorithm was invented as
a preliminary to this development, and was applied to-types and pair-types in a
A-calculus with the extra operator let that is characteristic of ML. It was written out
formally in Milner 1978 and rewritten and extended in Damas and Milner 1982.2
Milner's algorithm, like Hindley's, depended on Robinson's unification algorithm to
save work.3
Since Milner 1978 many modified or extended PT algorithms have been published,
some depending on external unification algorithms and others being self-sufficient.
See Tiuryn 1990 for a survey.

3A7.1 Note The PT algorithm in this chapter will use the method of Hindley 1969
and Milner 1978 rather than equation-solving.
As mentioned above this method depends on a unification algorithm given in
advance, but unification algorithms are widely available as packages in practice, so
this feature makes a PT algorithm easy to fit into an already given system in a
practical implementation.
This method will also turn out to be well suited to deal with the case where
the term whose PT is being computed is a combination of other terms Pt,...,P
whose PT's are already known. This situation is common in practice, where a
library of terms and their PT's can be built up and used in determining the PT's of
new terms, and one of the original motivations for the method was a belief that it
would probably use such accumulated information more efficiently than a straight
equation-solving algorithm.
However, the account in this chapter will not be concerned with maximizing
efficiency, but only with the (usually incompatible) aim of making the PT algorithm's
structure and motivation as clear as possible.

3B Type-substitutions

To make the statement of the PT algorithm reasonably concise it will help to

introduce some preliminary definitions and lemmas on type-substitutions.
Morris 1968 was independent of Curry's 1966-67 work and must have been prepared at about the
same time. Although Morris 1968 was never published there are later accounts of equation-solving
A-calculus algorithms available in the literature; for example there is one with a correctness-proof in
Wand 1987.
2 Milner's algorithm was discovered independently of earlier work. Milner 1978 did not contain a
correctness-proof but there was one in the thesis Damas 1984 to which Damas and Milner 1982
referred.
3 The history of these early algorithms seems to be a tale of repeated re-discovery of very similar ideas.
This is not really very surprising, as the ideas are very simple once they are approached in the right
way; it is their applications that make them important, not any intrinsic subtlety.
 3B Type-substitutions 35

3B1 Notation As defined in 3A1 a substitution is just a finite sequence of instructions

§ _ [at /al, ... , 6n/an] saying "simultaneously substitute of for at, ... , a, for an". The
following extra notation will be useful.'
The case n = 0 will be allowed and called the empty substitution, e. So
e(T) _ T.

If n = 1, s will be called a single substitution.

Each part-expression a1/al of s will be called a component of s, and called trivial
if o, = ai.
If all trivial components are deleted from s the resulting substitution will be called
the nontrivial kernel of s.
The set {a,,...,an} will be called Dom(s) or the variable-domain of s.
And Vars(ot,...,an) will be called Range(s) or the variable-range of s.

3B2 Definition Substitutions s and t are extensionally equivalent (s =ext t) if s(T)

I(T) for all T.

3B2.1 Lemma (i) b Dom(s) s(b) _ b.

(ii) § =ext tiff s and t have the same non-trivial kernel.

3B3 Definition (Restriction, s r V) If s - [at/at,...,Un/an] and V is a given set

of variables, the restriction s r V of s to V is the substitution consisting of the
components of/ai of s such that ai E V.

3B3.1 Lemma (s [ Vars(T))(T) = s(T).

3B4 Definition (Union) If s _ [at/at,...,vn/an] and t _ [T,/bt,...,Tp/bp] and either

a1.... , an, bt, ... , by are all distinct or ai = b1 (Ti = r1, define

sUt = [al/al,...,Qnlan,Tllbl,...,Tplbp]
(with repetitions omitted).

3B4.1 Lemma (i) Dom(s U t) = Dom(s) U Dom(t).

(ii) If § =ext s' and I =ext t' and s U t is defined, then so is s' U C and
§ U t =ext § U t'.

The next definition will be the composition of two substitutions s and t, a

simultaneous substitution that will have the same effect as applying t and s in
succession. To motivate it, consider the case
§ - [(c-+d)/a, (b->a)/b], I - [(b->a)/b]
and let T = a-+b. The result of applying first t then s is easily seen to be
s(t(T)) _ (c->d)-(b--*a)-+(c->d);
There seems to be no standard substitution notation in the literature.
36 3 The principal-type algorithm
the problem is to find a simultaneous substitution having the same effect. A naive
attempt to use s U t would fail, because
(s U t)(T) = s(T) _ (c--+d)-+b-+a.
But there does exist a suitable substitution, namely
[(c-'d)l a, (s(b-*a))l b]
This special case is generalized as follows.

3B5 Definition (Composition) If s and t are any substitutions, say

s= t = [Tt/b1,...,Tp/bp],
define
s o t = [ai, /ai...... air, l ai,,, s(T1)/b1,... , s(T4)/bp]
where {ai,,...,ai,,} = Dom(s) -Dom(t) and 0< h< n.

3B5.1 Lemma (i) Dom(s o t) = Dom(s) U Dom(t).

(ii) (s o t)(T) = S(t(T)).
(iii) r o (s o t) =ext(r o s) o t.
(iv) s =ext s', t =ext t' § o t =ext s' o
(v) By (ii), an instance of an instance of T is an instance of T.

3B5.2 Exercise* (i) Write out s o t in the special case that s - [alb] and I = [b/a],
and verify that (s o t)(T) = s(t(T)) in the case T - a-*b.
(ii) Show that the action of any s on a given type T can be expressed as a
composition sl o ... o sk of single substitutions, in the sense that
(sl o ... o §k)(T) = §(T)
The next lemma will play an important role in the correctness proof of the PT
algorithm: it says that if a composition sot is "extended" to r U (sot), the extended
substitution can also be expressed as a composition with t (under certain conditions
on r to prevent clashes).

3B6 Composition-extension Lemma Let r, s, t be substitutions such that

(i) Dom(r) n (Dom(s) U Dom(t)) = t0,
(ii) Dom(r) n Range(t) = 0.
Then r U (s o t) and (r U s) o t are both defined and
rU(sot) - (rUs)ot
Proof Suppose r, s, t are, respectively,
[p1/al,.. , Pr/ar], [al /bl, , as/bs], [T1/C1,...,Tt/ct]
with r, s, t >_ 0, and suppose
(1)V Dom(s)-Dom(1)={bi...... bi,,} (h >_ 0).
 3B Type-substitutions 37

Then r U (s o 1) is

[Pl/al,...,Pr/ar, vil/bi1...... in/bin,

(This is defined, since Dom(s o 1) = Dom(s) U Dom(1) by 3B5.1(i) and the latter is
disjoint from Dom(r) by assumption (i).) On the other hand (r U s) o I is

[Pi/al,..., Pr/ar, vi/bl,...,vs/bs] o [Tl/ci,...,ir/ce]

(This is defined because r U s is defined, by (i).) Now by (i) and (1),
Dom (r U s) - Dom (1) = { a, , ... , ar, bi,, ... , bi,, }

so by the definition of composition, (r U s) o t is

[Pl/ai,...,Pr/ar, ai,/bi,,...ain/bi,,, (rUs)(T1)/ci,...,(rUs)(Tt)/ce]

And (r U s)(Tj) = s(T1) for j = 1,..., t by assumption (ii). Hence result.
To end the section a proof will be given that the PT of a term is unique in a
certain sense. The definition of PT carries with it an intuitive feeling that this should
be so, and the following definitions will help to make this intuition precise.

3B7 Definition A variables-for-variables substitution is a substitution s

[bi /al, ... , b are variables (not necessarily distinct).
If bi,...,b,, are distinct, s is called one-to-one, and if also {al,...,an} = Vars(T)
for a given type T, s is called a renaming (of the variables) in T. (Renamings in
deductions and finite type-sequences are defined similarly.)
If s is one-to-one its inverse, is defined by

S_' = [al/bi,...,an/bn]
3137.1 Lemma If s is a renaming in T then s-1 is a renaming in s(T) and
s '(s(T)) = T.
3B7.2 Warning A one-to-one substitution may be a renaming in one type but not
in another. For example [b/a] is a renaming in a--+a but not in a-- b.

3B8 Definition We say a is an alphabetic variant of r, or a and r are identical modulo

renaming, if a - s(T) for some renaming s in T. (Alphabetic variants of deductions
and of finite type-sequences are defined similarly.)

3138.1 Lemma (i) a is an alphabetic variant of T iff a and T are instances of each
other.
(ii) Part (i) also holds for deductions and finite type-sequences.

Proof For "only if" in (i) use 3B7.1. For "if", let Vars(T) = where
al,...,a are distinct, and suppose v - s(T) and T - 1(a), for substitutions s and It
with
Dom(s) = Vars(T), Dom(t) = Vars(a).
38 3 The principal-type algorithm
Then t(s(ai)) = ai for i = 1,...,n. Hence §(a;) cannot be composite, otherwise
9(§(a;)) would be composite too. Thus s has form

§ ° [bl/al,...,bn/a],
and Vars(a) = {bl,...,bn}. Also bl,...,bn are distinct, because if bi = bj then
a; = ff(s(ai)) = ff(bi) = ¢(bi) = tr(s(ai)) = ap

Thus a is an alphabetic variant of T. 0

3B8.2 Corollary (Uniqueness) Let T be a principal type of a term M. Then T is unique
modulo renaming, in the sense that another type a is a principal type of M iff a is an
alphabetic variant of T. Principal deductions are unique in a similar sense.

3B8.3 Lemma For each finite set of type-variables at....,am and each type T there is
an alphabetic variant of T that contains none of al, ... , am.

3C Motivating the PT algorithm

The two main steps in the PT algorithm will be the computation of
PT from P T (P) and P T (Q). The first will be
no problem but the second will need a little work which in fact will be the core of
the PT algorithm; the aim of the present section is to motivate the method used in
this core.

3C1 Discussion Suppose we are trying to decide whether an application PQ is

typable in TA2, and we already know that
PT(P) = p-*a, PT(Q) = T.
(And for simplicity suppose PQ has no free variables.) Then by 3A3.1 the types
assignable to P are generated from p-->a by substitution and those assignable to Q
are generated from T. Hence, if we can find substitutions sI and §2 such that

sl(p) ° §2(T),
we will be able to deduce a type for PQ by rule (-*E), thus:
- P:940-'si(a) '-' Q:s2(T) (-E)
F-> PQ:sI(a).
Conversely, by the subject-construction theorem (2B2(iii)), every type deduced for
PQ must have been obtained from instances of p--+u and T by (->E) in this way.
Thus the problem of deciding whether PQ is typable reduces to that of finding §I
and §2 such that §I(p) §2(T). This suggests the next two definitions.

3C2 Definition (Common instances) (i) Iff v =_ §I(p) §2(T) we call v a common
instance (c.i.) of the pair (p, T), and we call (sI,§2) a pair of converging substitutions
for (p, T).
 3C Motivating the PT algorithm 39

(ii) Iff (VI,...,vn) we call (vl,...,vn) a common

instance of (p1,..., p,,) and (t1...... n). Common instances of pairs of deductions are
defined similarly.

3C2.1 Example A common instance of the pair (a->(b-*c), (a-+b)->a) is the type
(where /3, y, b are any given types), and the corresponding
converging substitutions are
sl = fl/b, y/c), §2 = [(l3-'Y)la, Slb].

3C2.2 Note Not every pair of types has a common instance. For example the pair
(a-*a, (b-*b)-*b) has none, because if §i (a-+a) §2((b-+b)-*b) we would have

si(a) = s2(b-*b), si(a) = §2(b),

which would imply the impossible identity s2(b) - §2(b)-+s2(b).

3C3 Definition (M.g.c.i.) (i) A most general common instance (m.g.c.i.) of (p, T) is
a common instance vo such that every other common instance is an instance of vo.
If vo is an m.g.c.i. of (p, T) we shall call any pair (51,52) such that S1 (P) _ §2(T) vo
an m.g.c.i.-generator for (p, T).
(ii) M.g.c.i.'s of pairs of type-sequences and pairs of deductions are defined
similarly.

3C3.1 Exercise* Show that the pair (a-*(b-*c), (a--+b)-->a) in 3C2.1 has the following
as an m.g.c.i.:
vo =
3C3.2 Lemma (i) M.g.c.i.'s are unique modulo renaming. That is, if v is an m.g.c.i. of
(p, T) the other m.g.c.i.'s of (p,T) are alphabetic variants of'v.
(ii) If p' and T' are alphabetic variants of p and T respectively, then (p, T') has the
same common instances and the same m.g.c.i.'s as (p,T).
(iii) Similarly for m.g.c.i.'s of deductions and finite type-sequences.

3C4 Discussion It will be shown later that every pair (p, T) with a common instance
has an m.g.c.i. Given this fact, the discussion in 3C1 suggests that if we know
PT(P) _ p- +a and PT(Q) - T and we know somehow that (p,T) has a common
instance, we can compute PT(PQ) by just constructing the m.g.c.i. of (p, T), say

V = sl(P) = §2(T),
and then letting PT(PQ) _ si(a).
And this is indeed true, provided we avoid one small snag. Suppose u contains
some variables bl,...,bk that do not occur in p, and by bad luck the m.g.c.i. v we
have constructed also contains some of these variables. Then si(a) might contain
two occurrences of one variable b;, one originally in o and the other introduced into
it by sl. In this case sl(a) would not be the most general type assignable to PQ,
because we could change v to an alphabetic variant v' - s,(p) with no variables in
40 3 The principal-type algorithm
common with a, and then the corresponding sp(a) would be a type of P Q that was
not an instance of Si (a).
As a concrete example, let P - Q =_ Ax-x; it will be shown in 3E that
PT(P) - a-+(b-4a), PT(Q) - b-*b.
Thus in this case p ° a, a = b-- a, T = b--+b. Clearly an m.g.c.i. of (p, T) is v = b-+b,
obtained from p by the substitution §1 - [(b-*b)/a]. And
§I(a) = si(b-+a) - b-.(b-+b).
It is easy to see that b-*(b-*b) is assignable to PQ. But it is not the principal type
of PQ. Because if we change s1 to a new substitution s' by replacing b by a new
variable c that does not already occur in a, we get
IA P : (c-->c)->(b->(c--c)), la Q: c->c,
and hence PQ has a type b-*(c-*c) that is not an instance of b-+(b->b).
To avoid this snag the PT algorithm will be careful to choose an m.g.c.i. v of
(p, T) such that
(1) Vars(v) n (Vars(a) - Vars(p)) = 0.
Given this precaution, we have now reduced the problem of finding PT(PQ) to
that of finding an m.g.c.i. Of (p, T). More precisely, we need an algorithm to decide
whether (p,T) has a common instance and, if the answer is "yes", to output a most
general one. A suitable algorithm will be given in the next section. It will not be
direct, but will apply the unification algorithm, which has the advantage that its main
properties are so widely known that they will only need to be outlined below.

3D Unification
Most readers have probably met unification before. This section merely summarizes
the relevant definitions and basic properties for the reader who has not. The account
is based on the classical one in Robinson 1965. (An alternative account is in Aho et
al. 1986 §6.7, a thorough survey of major results is in Baader and Siekmann 1994,
and a survey of the various applications of unification is in Knight 1989.)

3D1 Definition (Unifiers) (i) Iff there is a substitution s such that s(p) __ §(T) we
say (p, T) is unifiable; we call any such s a unifier of (p, T) and call s(p) a unification
of (p, T).
(ii) A unifier of a pair of sequences ((pi,..., pn), (TI,..., T,,)), both with the same
length, is an s such that
§((pl,...,pn)) = s((TI.....rn))

3D1.1 Note A unification of (p, T) is just a common instance obtained by making

the same substitution in p as in T. But not every common instance is a unification;
indeed a pair (p,T) may have a common instance but not be unifiable at all. For
example, consider
p = a--+(b-+c), T - (a-+b)-+a;
 3D Unification 41

this pair was shown in 3C2.1 to have a common instance, but no s can exist such
that §(p) - §(T), because the latter would imply the impossible identity
s(a) - §(a)-*s(b).

3D1.2 Note The problem of finding unifiers for pairs of type-sequences can be
reduced to that for pairs of types as follows. Given two sequences (P1,...,pn) and
(ti..... Tn), choose a variable b not occurring in any of these types and define
p* = P1--+...->Pn-+b, T - T1-->...->Tn-fib;
then the given pair of sequences is unified by a substitution s if (p*,T*) is unified
by s r Vars(pl,...,Pn,T1,...,Tn).

3D2 Definition (M.g.u.) (i) A most general unifier (m.g.u.) of (p, T) is a unifier u
such that for every other unifier s of (p, T) we have

§(P) = § (u(P))
for some s'. If v - u(p) for some m.g.u. U of (p,T) we shall call v a most general
unification (m.g.u.) of (p, T).
(ii) M.g.u.'s of pairs of type-sequences or deductions are defined similarly.

3D2.1 Exercise* Prove that the pair (a-*(b->b), (c--+c)-+a) is unifiable with a most
general unifier u = [(b->b)/a, b/c], and with the corresponding most general unifi-
cation being
(b-*b)-*(b-*b).

3D2.2 Lemma A type p is a most general unification of (p, T) iff µ is a unification of

(p, T) and all other unifications of (p, T) are instances of p.

3D2.3 Lemma If u is an m.g.u. of (p, T) and v is a renaming of variables in u(p), then

v o u is an m.g.u. of (p, T).

3D2.4 Notation From now on we shall often speak of "the" m.g.u. of (p, T) as if
m.g.u.'s were unique. (By 3D2.2, 3B8.1 and 3D2.3 they are unique modulo renaming.)

3D2.5 Lemma (Avoiding variables) (i) Let V be any finite set of type-variables. If
(p, T) has an m.g.u. u, then it has an m.g.u. u' such that
Dom(u) = Vars(p) U Vars(T), Range(u') n v = 0.
(ii) Similarly for m.g.u.'s of pairs of type-sequences or deductions.

Proof If u contains a component v/a with a Vars(p) U Vars(T) we can remove

this component from u without affecting u(p) or U(T). We can also add trivial
components to enlarge Dom(u) if necessary, and the result can easily be seen to be
still an m.g.u. We can then change the range of u by a renaming. 0
42 3 The principal-type algorithm
3D2.6 Note (M.g.u.'s and m.g.c.i.'s) The m.g.u. of a pair (p,T) may differ from its
m.g.c.i. For instance the m.g.u. of (a->(b-*b), (c-*c)-*a) was shown in 3D2.1 to be
(b-*b)-*(b-*b); but in contrast its m.g.c.i. is easily seen to be (c-*c)-*(b->b). And
as noted in 3D1.1, it is also possible for (p, T) to have an m.g.c.i. but no m.g.u. at all.
However, in the special case that p and T have no common variables every
common instance is also a unification, because if
v = §i(P) §2(0
and Dom(§t) c Vars(p) and Dom(§2) c Vars(T) then §a U §2 is defined and

v = (§i U §2)(P) = (§a U §2)(T)

This fact gives us the following lemma.

3D3 M.G.U.-M.G.C.I. Lemma (i) If p and r have no common variables, (p, T) has
an m.g.u. ii f it has an m.g.c.i., and the two are identical.
(ii) For all p and T: if we change r to an alphabetic variant T* with no variables
in common with p, the unifications of (p,T"`) will be exactly the common instances of
(p, T) and the m.g.u. Of (p,T*) will be the m.g.c.i. Of (p, T).
(iii) Similarly for pairs of type-sequences or deductions.

Thus the problem of finding m.g.c.i.'s has now been reduced to that of finding
m.g.u.'s of pairs (p,T) with no variables in common. But searching for m.g.u.'s of
pairs, with or without variables in common, can be done by an algorithm as follows.

3D4 Unification Theorem (J. A. Robinson) (i) There is an algorithm which decides
whether a pair of types (p, T) has a unifier, and, if the answer is "yes", constructs its
m.g.u.
(ii) If a pair (p,T) has a unifier it has an m.g.u.
(iii) Parts (i)-(ii) hold also for pairs of deductions and for pairs of finite type-
sequences.

Proof (i) For Robinson's algorithm see 3D5 below; for a proof of its correctness
see Robinson 1965 §5 pp. 32-33.
(ii)-(iii) Like (i).

3D5 Unification Algorithm (Robinson 1965 §5.) Input: any pair (p, T) of types.
Intended output: either a correct statement that (p, T) is not unifiable or an m.g.u. au
of (p, T). [The algorithm will build au in stages auo, wit,..., each auk being a composition
of auk-1 with a new substitution: at the k-th stage it will test whether auk(p) - 0uk(T),
and if the answer is "yes" it will choose u = uk and stop; but if not, it will "extend"
uk and go to the next stage.]
Step 0. Choose k = 0 and uo = e (the empty substitution).
Step k+1. Given k and auk, construct pk = auk(p) and Tk = 0uk(T), and apply'the
comparison procedure below to (pk,Tk). That procedure will output either a correct
statement that pk Tk or a disagreement pair (a, a) (see below) such that a * a.
If Pk - Tk, choose u = uk.
 3D Unification 43

If pk * Tk and the output of the comparison procedure is (a, a), decide whether
a E Vars(a).1
If a E Vars(a), state that (p, r) is not unifiable and stop.
If a ¢ Vars(a), then replace k by k + 1, choose uk+1 = [a/a] o uuk, and go to
Step k + 2.
Comparison Procedure. Given a pair (p, v) of types, write p and v as symbol-strings,
say
p - 51 ... 5m, V = tl ... to (m, n ! 1)
where each o f sl,... , sm, t1, ... , t is an occurrence of a parenthesis, arrow or
variable.
If p v, state that p - v and stop.
If y * v, choose the least p < Min{m, n} such that s, # tp; it is not hard
to show that one of sa, t, must be a variable and the other must be a left
parenthesis or a different variable. Further, sp can be shown to be the leftmost
symbol of a unique subtype p" of p. (If sv is a variable, p" - se.) Similarly
to is the leftmost symbol of a unique subtype v" of v. Choose one of p",
v" that is a variable and call it "a". (If both are variables, choose the one
that is first in the sequence given in Definition 2A1.) Then call the remaining
member of (p", v") "a"; the pair (a, a) is called the disagreement pair for
(p, v)

3D5.1 Note To prove that p exists in the case that p # v in the comparison
procedure we must show that it is not possible to have
tl...tn = SI ... Smtm+l ... tn
with it > m. This is left as a (rather dull) exercise for the reader.

3D5.2 Note (History) As mentioned earlier, the above unification algorithm is due
to J. A. Robinson, see Robinson 1965 §5; it was the first to be published complete
with a correctness-proof. Robinson dates its initial implementation to 1962, but he
credits a less smooth algorithm implicit in Prawitz 1960 with influence on its origin
and mentions that the history of unification algorithms goes back at least as far
as Herbrand 1930.2 Another early unification algorithm is implicit in Maslov 1964.
For more on the early history see the start of §3 in Baader and Siekmann 1994.
Although Robinson's algorithm is easy to describe it is not particularly efficient
to run, and many better algorithms have been published and implemented since the
1960's. Useful surveys are in Knight 1989 and Baader and Siekmann 1994. A good
discussion of efficiency is in §3.3 of the latter; here it is enough just to mention that
there are algorithms in Paterson and Wegman 1978 §4 and Martelli and Montanari
1982 that run in linear time if their inputs (p, T) and outputs au are coded in a suitably
compact way, but that without some coding even the mere printing-out of au(p) may
take exponential time in the worst cases. The problem of deciding whether a given
pair (p,T) can be unified is known to be PTIME-complete (Dwork et al. 1984).
This decision is called the "occurs check".
Z Robinson 1966, Robinson 1979 p. 292.
44 3 The principal-type algorithm
3D5.3 Corollary (i) There exists an algorithm which decides whether a pair (p, T) has
a common instance, and if the answer is "yes", outputs an m.g.c.i. and a pair (§1, §2)
of m.g.c.i.-generators for (p, T).
(ii) Every pair of types with a common instance has an m.g.c.i.

Proof By 3D3(ii) and the unification algorithm.

3E The PT algorithm

The PT theorem (3A6) stated that there is an algorithm that decides whether a given
untyped term M is typable in TAR, and outputs a principal deduction and principal
type for M if the answer is "yes". A suitable algorithm will now be described. The
proof that it satisfies 3A6 will be given as a series of notes between the algorithm's
steps; these notes will also help to motivate the steps.

3E1 Principal Type (PT) Algorithm (Hindley 1969 §3.) Input: any A-term M, closed
or not.
Intended output: either a principal deduction AM for M or a correct statement
that M is not typable.

Case I. If M is a variable, say M x, choose AM to be the one formula deduction

x:a --> x:a, where a is any type-variable. [Clearly AM is principal for M.]

Case II. If M - and x e FV(P), say FV(P) = {x,xi,...,x,}, apply the

algorithm to P. If P is not typable, neither is M. If P has a principal deduction
A. its conclusion must have form

x:a, x1:a1,...,xt:at

.for some types a, cu, ... , at, fl. Apply rule to make a deduction of

x1:a1,...,xt:at '-+ (Ax.P):a-fl.

Call this deduction ARp

Justification of Case II. We must show that the above ARC is principal for
Let A be any other deduction of a type for Ax-P. By the subject-construction
theorem (2B2) the last step in A must be by rule (-*I)main, with form

x:a', x17[1,...,xt:7rt i--a P:/3'

x1:7[1,...,xt:71t ti
for some types a', 7C1, ... , It,, /3'. Now the steps of A above the last one form a
deduction A' that assigns a type to P ; but P has a principal deduction AP, so
 3E The PT algorithm 45

A' _ §(Ap) for some s; hence in particular

cc' = §(a), TIl = §(al), ... , nt = s(at), aa
1'' = §(Y )
RR

Consequently A =
Case III. If M - 2x P and x FV(P), say FV(P) = {xl,...,xt}, apply the
algorithm to P. If P is not typable then M is not. If P has a principal deduction
Ap its conclusion must have form
xl:al,...,xt:at H P:/3
for some types al, ... , at, /3. Choose a new type-variable d not in Ap and apply
vacuously discharging x:d, to get a deduction of
xl:al,...,xt:at '--* (.lx

Call this deduction Ap.

Justification of Case III. We must show that the above Daxr is principal. Let 0 be
any other deduction of a type for By 2B2 its last step must be an application
of rule (j)vac, with form
xl:nl,...,xt:nt '.-+ P: f3'

xl:ni,...,xt:nt '-- (2x.P):a'-./3'

for some a', nl..... ins, /3'. The steps above the last one form a deduction S' of a type
for P, /and A is principal, so A' _ s(Ap) for some §. Choose s' - § U [a'/d] ; then
A = §'(Alx.P )
Case IV. If M - PQ, apply the algorithm to P and Q. If P or Q is untypable then
so is M. If P and Q are both typable, suppose they have principal deductions Ap
and AQ. First rename type-variables, if necessary, to ensure that A and AQ have
no common type-variables. Next list the free term-variables in P and those in Q
(noting that these lists may overlap); say
FV(P) = lul,...,up>N't>...,wr
r
(p,r > 0),
FV(Q) = {vii...,vq,w1,...,wr} (q? 0),
where u1.... , up, vl..... vq, W1 ...., wr are distinct.
Subcase IVa: M . PQ and PT(P) is composite, say PT(P) - p-+Q. Then the
conclusions of A and AQ have form, respectively,

(1) ul:O1,...,up:Bp, wl:ipl,...,wr:1pr '--

(2) vt:4'1i...,vq:4q,wl:X1,...,wr:Xr
tt
h-> Q:t.

Apply the unification algorithm (3D5) to the pair of sequences

(3) (ipl,..., l1r,P), (XI,...,Xr,r)

Subcase IVal: the pair (3) has no unifier.

Then PQ is not typable. [See the
justification below.]
46 3 The principal-type algorithm
Subcase IVa2: the pair (3) has a unifier. Then the unification algorithm gives a most
general unifier uu; apply a renaming if necessary (3D2.5) to ensure that

(4) Dom(u) = Vars(ipl, Xr, T),

(5) Range(u) fl V = 0,
where

(6) V = (Vars(AP) U Vars(AQ)) - Dom(u).

Then apply u to Ap and AQ; this changes their conclusions to, respectively,
u1:6* ,...,uP:BP, Wl :tpi ,...,Wr:tpr r-* P:p*-.v*,
vl:(al*,...,vq:(a9*, WI:XI*,...,Wr:Xr* Q:2*,
F-i

where Bi =- u(Bi), etc. And by the definition of uu we have

* * * * *
Wi = Xl , . tr - Xr , P =7
Hence (-->E) can be applied to the conclusions of u(Ap) and u(AQ). Call the
resulting combined deduction APQ; its conclusion is

(7) u1:0*.... I UP OP

Justification of Subcase IVa. For IVal we must prove that if PQ is typable then
(wl, . , Wr, p) and (Xi, ... , Xr, i) have a unifier, and for IVa2 we must prove that the
above APQ is principal for PQ.
Justification of IVal. If PQ is typable, there is a deduction A whose conclusion
has form

(8) U1:2r1,...,UP:7rP, vl:µl,...,vy:uq, wl:vl,...,Wr:Vr F-s PQ:/3

for some types 7t1, ... , 7tp, etc. By the subject-construction theorem (2B2), A must have
been built by applying rule (--* E) to two deductions Al and A2 whose conclusions
are

(9) UI:7tl,...,up:7tp, W1:V1,...,Wr:Vr f-a P:

(10) v1:1,11,...,vy:zq, W1:Vl,...,Wr:Vr I--.. Q:a

for some type a. But Ap and AQ are principal deductions for P and Q, so
Al = ri(Ap), A2 = r2(AQ)
for some substitutions r1 and r2 such that
Dom(rl) = Vars(Ap), Dom(r2) = Vars(AQ).
F
Roughly speaking, (5) and (6) say that when u is applied to Ap and AQ any new variables it introduces
will differ from all those already in Ap and AQ, and hence no unnecessary identifications of variables
will be made. (Cf. the motivation in 3C4.) By the way, if the aim of the algorithm had been to
construct merely a principal type and not a principal deduction, (5) could have been weakened by
re-defining V to consist of just the variables (if any) in a that do not occur in the types in (3). (Cf. (1)
in 3C4.)
 3E The PT algorithm 47

Let r - rl U r2. (This is a well-defined substitution since Ap and OQ were chosen

such that Vars(Ap) n Vars(OQ) = 0.) Then
(11) Al = r(Ap), 02 = r(OQ),
(12) Dom(r) = Vars(Op) U Vars(AQ).
Now look at V1,...,Vr and a in (9). By comparing (9) and (1), since O1 - r(Op) we
have
(13) (V1,...,Vr,a) = (r(W1),...,r(>Vr),r(P))
And by comparing (10) and (12), since
A2 =_ r(DQ),

(14) (V1,...,Vr,a) = (r(XI),...,r(Xr),r(T))-

Thus 41, ... , ivr, p) and (XI, ... , Xr, T) have a unifier, namely r.
Justification of IVa2. We must prove that ApQ is principal for PQ, i.e. that an
arbitrary deduction A of (8) is an instance of ApQ. For any such A, define Al, A2
and r as above. By (6) and (12) we can express r as r =_ r' U r", where
(15) r' - r V r" - r rpom(u).
By (4), (13) and (14), r" is a unifier of
(VP1,...,lPr,P), (XI,...,Xr,T).
But u is an m.g.u. of this pair of sequences, so by the definition of m.g.u. (3D2) there
exists s such that r"(a) _ s(u(a)) for all a E Dom(u). We can clearly also assume
(16) Dom(s) c Range(u).
It follows that
(17) r" =ext § o U.

Since r - r' U r", we have by 3B4.1(ii),

(18) r=extr'U(§ou).
Now r', s, u satisfy conditions (i) and (ii) in the composition-extension lemma (3B6).
In fact 3B6(i) holds because by (15) and (16) we have
Dom(r') n (Dom(s) U Dom(u)) c V n (Range(u) U Dom(u))
= 0 by (5), (6),
and 3B6(ii) holds by (5) and (15). Hence by (18) and 3B6,
r =ext(G' U s) o U.

Therefore by (11),
Al - (r' U s)(u(Ap)), A2 = (p U s)(u(AQ))
But A is a combination of Al and A2 by rule and ApQ is a similar combination
of u(Ap) and u(AQ). Hence
A = (r' U s)(ApQ).
48 3 The principal-type algorithm
Subcase IVb: M = PQ and PT(P) is atomic, say PT(P) = b. Then the conclusions
of A and AQ have form, respectively,

(19) u1:B1,...,uP:O" W1:W1,...,wr:Wr i-- P:b,

(20) v1:4)1,...,vq:4)q, W1:X1,...,Wr:Xr H Q:T.

Choose any variable c Vars(SP) U Vars(OQ) and apply the unification algorithm
to the pair of sequences

(21) (Wl,...,Wr,b), (Xl,...,Xr,t +c).

Subsubcase IVbl: the pair (21) has no unifier. Then PQ is not typable. [See the
justification below]
Subsubcase IVb2: the pair (21) has a unifier. Then the unification algorithm gives
an m.g.u. u; apply 3D2.5 to ensure that

(22) Dom(u) = Vars(W1,

(23) Range(u) r1V = 0,

where V is the same as in (6). Then apply a to Op and L. By the definition

of u,

u(b) = u(T-'c) = u(T)--'u(c),

and thus the conclusions of u(Ap) and u(AQ) are

u1 :e1 , ... , uP :BP , w1 :W1 , ... , Wr :W,* Fa P : T* -+C*,

vl:4l,...,vq:4) , W1:X1,...,Wr:Xr Q:T ,

where * denotes application of the substitution u, and

WI - Xl > ,Wr = Xr
Choose APQ to be the deduction obtained by applying rule to u(Ap) and
u(OQ); its conclusion is

ul:01 ,...,uP:o , v1 :01 ,...,vq:4q, W1:XI,...,Wr:Xr i-a PQ:C*.

Justification of Subcase M. For IVb1 we must prove that if PQ is typable then

and have a unifier, and for IVb2 we must prove that
OPQ is principal for PQ.
Justification of IVbl. If PQ is typable there is a 0 whose conclusion has the form
(8). By 2B2, 0 must have been built by applying (--*E) to two deductions Al and A2
with conclusions (9) and (10). But P and Q have principal deductions Ap and AQ,
so

(24) Al = a'1(AP), A2 = r2(AQ)

 3E The PT algorithm 49
for some substitutions rl and r2 such that Dom(rl) = Vars(Ap) and Dom(r2) _
Vars(AQ). Note in particular that

a--*# = ri(b), a = r2(T)-

Define
r - rl U r2 U [/3/c].
This is well-defined because c Dom(rl) U Dom(r2). And r(c) and by (24),
r(Ap) A1, r(AQ) = A2,
r(b) a-+/3 = r(T-ac).
Also
Dom(r) = Vars(Ap) U Vars(AQ) U f c}.
Now look at vl,...,vr and a- f3 in (9). By the above, we have
(25) (v1,...,Vr,a--->/3) = (r(Wt),...>r(Wr),r(b)),
(26) (VI,..., Vr, a-->/3) = (r(Xl),..., r(Xr), r(T-'c)).
Thus ( W 1 ,- .. , Y )r, b) and (X1, ... , Xr, T--+c) have a unifier, namely r.
Justification of IVb2. To prove ApQ principal we must prove that an arbitrary
deduction A of (8) is an instance of ApQ. For any such A, define A1, A2 and r as
above. We can express r as VU r", where
r' __ r rV, r" = r rpom(au),
where V and Dom(u) were defined in (6) and (22). By (25) and (26), r" unifies

(Wl,...,Wr,b), (X1,...,Xr,T-C).
Hence, just as in the justification of IVa2, r" =ext s o u for some s such that
(27) Dom(s) 9 Range(u).
Thus r =ext r' U (s o u). Using (23) and (27) it is easy to see that r', s, u satisfy the
conditions in the composition-extension lemma (3B6), so by that lemma,
r=ext(0 U §) o U.

Hence, similarly to the justification of IVa2, A __ (r' U s)(ApQ).

3E1.1 Notes (i) The PT algorithm could have been shortened by combining Sub-
cases IVa and IVb as was done in Hindley 1969: in IVa the unification algorithm
could have been applied to
(W1,...,Wr,P-'Q), (X1,...,Xr,T-C)
where c is a new variable, and this modification would have made IVa exactly
parallel to IVb and the latter could have been omitted. But this chapter's aim is
to explain clearly what is involved in computing a PT and not just to give the
slickest presentation, and IVa seems easier to understand when not combined with
IVb.
50 3 The principal-type algorithm
(ii) At the start of Case IV the algorithm could have been made more direct
by omitting the renaming of the type-variables in A p and AQ and using common
instances instead of unifiers in Subcases IVa and IVb. For example in IVa, instead
of seeking an m.g.u. u of the sequences
(p1, ... , wr, p), (XI, ... xr, T),
we could have sought an m.g.c.i. The justification of the algorithm would then have
rested on 3D5.3, the analogue of the unification theorem for common instances.

3E2 Exercise Show that the closed terms in Table 3E2a have the principal types
shown there. (The answers to (1)-(7) are displayed in 2A8.2-4 and in the answer to
Exercise 2A8.7; these deductions can easily be shown to be principal. The others
can be checked by applying the PT algorithm.)
By the way, the terms (10) and (11) in Table 3E2a are not in #-normal form; we
shall see in 8B4 and 8B7 that there are no closed fl-nfs with the same PT's as these
terms.
An extended table of terms and their principal types is given at the end of the
book for ease of reference.

Table 3E2a.
Term Principal type
(1) B = 2xyz-x(yz)
(2) B' = 2xyz'Y(xz)
(3) C (a-b-.c)-b-a-c
(4) a- a
(5) K a-b-*a
(6) S = (a-.b- c)-
(7) W )xY'xYY (a-+a-*b)-*a-*b
(8) 0 a-.b-b
(9) 1 -
(10) (.1xyz-K(xy)(xz))I a- a- a
(11) Rxy (2z x)(Yx) a-(a-b)- a
(12) )xyz-xy(xz) [untypable]
(13) [untypable]

3E3 Theorem The relation I-x M:T is decidable, i.e. there is an algorithm which
accepts any M and T and decides whether or not I-x M:T.

Proof Apply the PT algorithm to M. If M is not closed or not typable we cannot

have t2 M:T. If M is closed and typable, decide whether T is an instance of PT(M).
0
 3E The PT algorithm 51

3E4 Further Reading For an alternative introduction to PT's and type-checking see
Aho et al. 1986 Ch. 6. The early literature on PT algorithms was mentioned in 3A7
above. The more recent literature on their uses and properties is fairly extensive and
perhaps the best place to start would be the survey Tiuryn 1990 and the papers in
its bibliography, as well as Giannini et al. 1993.
Accounts of PT algorithms that include correctness-proofs are relatively rare,
however; besides the proofs in Curry 1969 and Hindley 1969 §3 (and the unpublished
ones in Morris 1968 and Damas 1984) there is one in Wand 1987.

3E4.1 Note (Efficiency) Results on efficiency and complexity of PT algorithms,

mainly focusing on ML, are contained in Kanellakis and Mitchell 1989, Mairson
1990, Kfoury et al. 1990 and Kanellakis et al. 1991. In summary, there is a PT
algorithm which can decide in linear time whether a A-term containing no constants
is typable, and if it is, will print out its PT in linear time coded in a suitably
compressed form (as a directed acyclic graph or DAG). Also the typability problem
for pure 2-terms is PTIME-complete just like the unification problem (3D5.2).
(Proofs of this are in Tyszkiewicz 1988 and Mitchell 1996 Ch. 11.)
But when ML's let operator is added to the term-language the corresponding
typability problem becomes PSPACE-hard (Kanellakis and Mitchell 1989) and
DEXPTIME-complete (Mairson 1990, Kfoury et al. 1990).
 4

Type assignment with equality

Let M be a closed term and Types(M) be the set of all types assignable to M in
TA2. As we saw in Chapter 2 this set might not stay invariant during a /t-conversion
of M. In fact Examples 2C2.2-6 and Note 2C3.2 showed that this set may increase
during a reduction of M and sometimes M can even be converted to a term with
no types at all in common with M.
As noted in 2C3.2 this lack of equality-invariance has not hindered TA2 from a
practical point of view, nor its descendant ML, as the main need in both systems
has been simply for the subject-reduction theorem. However, on the theoretical side,
equality-invariance of Types(M) seems a desirable property from the viewpoint of
any of the standard 2-calculus semantics (except possibly an operational semantics):
if we believe terms represent functions of some sort and equal terms represent the
same function, then equal terms should have the same types.
So this chapter will describe a system obtained by adding an equality-invariance
rule to TA2. More precisely, two systems will be described in parallel, one for =p
and the other for =#,; both will have similar properties.

4A The equality rule

4A1 Definition (The systems TAx+p, TA2+p,,) We define TAt+p by adding to rules
(->E) and (-+I) in Definition 2A8 the new rule
(Eqp) r --* M:T M =p N
r '--> N:T.
Iff a formula t' --> M:T is provable in TA1+p for some T'' c F we say
t F-fl M:T.
And if there exist F and T such that F I-p M:T we call M TAz+p-typable.
We define TA2+p,,, I-p. and TAz+p,,-typability similarly but with "=pq" instead of
"=p", calling the corresponding new rule "(Eqp,)".t

4A1.1 Notation When confusion is unlikely (Eqp) and (Eqp,) may both be called
just "(Eq)".
Strictly speaking, to make the (Eq)-rules meaningful we must also add the axioms and rules for =p
and =p to the definitions of TAA+f and TAx+pq in some way; the details of this are left to the reader.

52
 4A The equality rule 53

The name "TAx+N[n]" will be used to denote TA2+p and TAx+p, simultaneously
when stating results that hold for both systems. Deducibility notation in such
statements will be

To emphasize the contrast with F# and F we shall call F2, in the present chapter
only,

Fnoeq

4A1.2 Example Some terms become typable in TAx+9[n] that were not typable in
TA2. For example, consider the term

P-
which was shown in 2C2.2 to have no type although it reduces to I which has type
a-*a. In TA,2+p[n] we can assign a type to P by the following deduction.

v:a H v:a
F-. Av - v =p(.1uv v)(2x xx)
(Eqp)
F-->

4A 1.3 Remark By adding (Eq) to TAx we have trivially solved the problem of
making the set of a term's types invariant with respect to conversion. But for this
easy gain we pay at least two prices.
(i) Rule (Eq) is undecidable. That is, there is no algorithm that will apply to
arbitrary F, T, M, N and decide whether the formula F i-+ N:i is deducible from
F --. M:2 by (Eq). (Because the relations =p and =p,, are undecidable.) This defect
can be remedied by replacing (Eq) by a series of rules, each one corresponding
to an axiom in a suitable definition of =p or =#,,, but this solution is far from
neat.
(ii) The subject-construction theorem (2B2) breaks down: when there are (Eq)-
steps present the structure of a deduction is no longer dictated by that of its subject.
Also when the proof of a formula F H M:T contains (Eq) we cannot infer that
Subjects(F) = FV(M) as we did in Lemma 2A10.
At first sight (i) and (ii) are serious drawbacks. However, the following theorems
will show that (ii), at least, is not as bad as it might seem; we shall see that in fact
rule (Eq) plays an unexpectedly small role in deductions and that TA,A+B[n] is tied to
TA2 very closely indeed.

4A2 Equality-postponement Theorem In TAx+9[nl every deduction of F F-> M:'r can

be changed into one in which the only use of (Eq) is at the end. That is, if F Fp[ ] M:i
then there exists a term M' =#I,l M such that

F [-no eq M' :T.

54 4 Type assignment with equality
Proof Let A be a TA2+P[gl-deduction of r H M:t. If (Eq) occurs in A above an
occurrence of (--*E), the relevant part of A will look like this:
F, P:p-+Q P =6[n] P' 172 ' Q:P Q =6[g] Q'
(Eq) (Eq)
r, P':p F2 Q':p

r, u r2 ti (P'Q'):Q.

Replace this part by the following, noting that the assumptions and the conclusion
stay the same.
Ti F-+ r2 i-- Q:p

r, u r2 H (PQ):a PQ =9[n1 P'Q'

(Eq)
r, U r2 -+ (P'Q'):Q.

If (Eq) occurs in A above an occurrence of (->I)main, the relevant part of A will

look like this:
r,x:p H P:a P =#[g] P' (Eq)
r,x:p H P':a
(--")main
I" H

Replace this part by the following:

r, x : p i-+ P :6
HI)main
F -+ AX-P =#[,,) AX-P'
(Eq)
r-x H
If (Eq) occurs above (-.I)vac make a similar replacement. (By the way, (-.I)vac is
distinguished from HI)main in Definition 2A8 by whether x occurs in the context
on its top line, not by whether x E FV(P), so converting P' to P will not change
an application of (->I)vac to one of (-+I)main or vice versa.')
Finally, if (Eq) occurs in A above another (Eq), combine them into one in the
obvious way. The above replacement-procedure will terminate in a deduction with
(Eq) at the end (if (Eq) occurs in it at all).

4A2.1 Note The above proof depends on little more than the very simple fact that
if P =6[n] P' and Q =#[n] Q' then

PQ=#[q] P'Q', =P[ql 2x'P'.

But as a consequence it allows us to extend several important results from TA2 to
TA2+p[ql, as follows.

In TA2, x occurred in the context on the top line of (-.I) iff x E FV(P), but as remarked in 4A1.3(ii),
this does not necessarily hold when (Eq) is present.
 4A The equality rule 55

4A3 Weak Normalization Theorem for TAA+p[n] Every TA,A+p[n]-typable term M has
a f[q]-nf M*, and furthermore
F 1-p[n] M:'r t=om r F-no eq M* :'r.

Proof Let F I-p[n] M:T. Then by 4A2 there exists M' =p[n] M such that
(1) r 1-no eq M':T.
By the weak normalization theorem for TA2 (2D5), M' has a of M* ; and by the
Church-Rosser theorem M' reduces to M* and M* is also the of of M. They by (1)
and the subject-reduction theorem (2C1),
(2) r F-no eq M*:T.

Conversely, if (2) holds then r 1-p[n] M:'r by (Eq).

4A3.1 Warning The weak normalization theorem cannot be strengthened to strong

normalization. For example let
M =_
Then M =p I so M is typable in both TA1+p and TA2+pn. But M has an infinite
reduction, so strong normalization fails for both systems.

4A3.2 Corollary The relation F I-p[nl M:-r is equivalent to each of the following:
(i) F F-no eq M* :t,
(ii) F r M* F-no eq M* :T,
(iii) F F-p[n] M* :T,
(iv) F r M* I-p[n] M* :'r,
(v) r r M* I-p[nl M :T'
(vi) F rM F-p[n1 M:T.

Proof Each of (i)-(vi) implies F F-p[n1 M:T by (Eq) and the weakening property of
" F-" (which holds for F-p[n] just as for I-2, see 2A9.1).
For the converse, the relation F F-p[n] M:T implies (i) by 4A3. And (i) implies (ii)
by 2A11. Next, (ii) implies (iv) trivially, and (iv) implies (iii), (v) and (vi) by (Eq)
and weakening.

4A3.3 Note When F F-#[n] M:T we cannot infer that Subjects(F) 2 FV(M), but by
the above corollary and 2A11 we can infer that
Subjects(F) 2 FV(M*).
The next theorem will express the content of Corollary 4A3.2 very neatly for the
case that M is closed.

4A4 Definition If M is closed, the set of all T such that I-g[,] M:T holds will be
called
Types p[n] (M).
56 4 Type assignment with equality
To emphasize the contrast, the set Types(M) defined in 2C3 will be called here
Types no eq(M).

4A5 Theorem If M is closed and has /3- and f rl-nf's M*p and M*p, respectively,
then
(i) TypeSp(M) = Typesp(M*p) = TYpesno eq(M*p),
(ii) Typespq(M) = Types#,(M*R,) = TYpesno eq(M*pn).

Proof By 4A3.2.

4A6 Definition (Principal types in TA2+p[n]) In TA,2+p[n] a type T is a principal type

of a term M if
(i) r --* M:T is deducible for some F,
(ii) if F' -+ M:T' is deducible then T' is an instance of T.

4A7 PT Theorem for TAx+p[a] In TAx+p[a] every typable term M has a principal
type, and it is the same as the principal type in TAx of the /3[q]-nf of M.

Proof By 4A3.2 and the PT theorem for TA2 (3A6).

4A8 Remark Results 4A2-7 show that the effect of adding rule (Eq) has been much
less than we might have feared: (Eq) simply transfers types from normal forms to
the terms that reduce to them, and consequently TA2+p[n] is little more than the
theory of assigning types to f3[ry]-normal forms in TAx.
However, this is not quite the whole story. Rule (Eq) was motivated at the start
of this section by semantic considerations and the semantics of TAx+p[n] will be
investigated in the next section.
But first two more syntactic results will be stated. The first is an analogue for
TA,1+p of the subject-reduction theorem for TA2: if it merely asserted closure under
'p it would be trivial but it says slightly more.

4A9 Theorem (il-closure of TA1+p) If F [-p P:T and P r>p, Q, then

r F-Q Q :T.

Proof If F [-p P :T then by 4A3, P has a fl-nf P*p such that

(1) I [-no eq P*p:T.
Now P >p, Q and P >p P*#, so by 1C7.1 there exists T such that
Q >S T, P*# DRn T.

(In fact P*p >, T by 1C9.2.) By (1) and subject-reduction (2C1),

F [-no eq T :T.
Hence by (Eqp),
r I-RQ:T.
 4B Semantics and completeness 57

4A10 Theorem (Undecidability of TA2+p[n]) There is no algorithm that accepts ar-

bitrary F, M, r as inputs and decides whether I, f[n] M:r.

Proof Let F = 0 and T=- a--+a. For any M, if Hp[n] M: a-*a then M is closed and
by 4A3, M has a of M* which is closed and has the property
Hoo eq M* : a-->a

But by the subject-construction theorem (2B2) it is easy to see that a closed /3[r]]-nf
with type a--+a in TA2 must have form for some x. Hence
[n] M:a-*a > M=d[n] I
Thus a test for Hpp[n] would give a test for convertibility to I, contrary to standard
undecidability results (e.g. Barendregt 1984 Thm. 6.6.2 or HS 86 Cor. 5.6.1).

4B Semantics and completeness

In this section TAZ+p will be interpreted in an arbitrary A-model and its rules proved
sound and complete, and the same will be done for TA,1+6n in extensional A-models.
The concept of A-model has been motivated, defined and thoroughly discussed in
Barendregt 1984 Ch. 5 (especially Def. 5.2.7), and in HS 86 Ch. 11. There are
several definitions of A-model in these sources, all equivalent, and the one presented
below has been chosen mainly because it emphasizes the concepts needed in the
completeness proof. It is from HS 86 Def. 11.3.

4B1 Definition (Semantic environments) Given a non-empty set D, a semantic en-

vironment in D is any function E that assigns to each term-variable v a member
E(v) of D. If E is a semantic environment, x a variable and d e D, the semantic
environment
[d/x]E
is defined to be the same as E except that it assigns d to x. (By the way, if E(x) = d
then [d/x]E = E.)

4B2 Definition (A-models) A A-model is a triple.9 = (D, , I I) where D is a set with

at least two members, is a 2-place function such that
d1,d2 E D dl-d2 E D,
and I I is a function that assigns to each E and each term M a member of D called
IMIE, such that
(i) WE = E(x) ,for all term-variables x,
(n) IPQIE =
(iii) [PI[d/x]E for all d E D,
(iv) IMIE, = IMIE2 if E1(v) = E2(v) for all v E FV(M),
(v) IMIE = I NI E if M =a N,
(vi) IAX-PIE = IAr-QIE if IPI [d/x]E = 1Q°[d/x]E for all d e D.
58 4 Type assignment with equality
4132.1 Notation The function - is called application and I I is called the interpretation
mapping.
Note that in the above definition as elsewhere in this book, "=" without a
subscript "fl" or " /rl" means "is identical to".

4B3 Definition A A-model -9 = (D, , I I) is called extensional if, for all di and
d2 E D,
(Ve E D)(di-e = d2-e) di = d2.

4B4 Lemma (i) Every A-model -9 satisfies the theory of of fl-equality in the sense
that
M =s N (VE) IMIE = INIE
(ii) A A-model -9 is extensional iff it satisfies f rl in the sense that

M=pqN (VE) IMIE = INIE

Proof HS 86 Thms. 11.12 and 11.15.

4B4.1 Note A A-model is a model of untyped A-calculus, i.e. every term receives
an interpretation in D whether it is typable or not. This agrees with the Curry
approach to type-theory described in 2A3. In contrast, in a model of Church's
type-theory only typed terms are interpreted, and instead of one domain D the
model has a distinct domain for each type. The definition of model for Church's
system can be found in Henkin 1950.
Types are interpreted in an arbitrary A-model as follows.

4B5 Definition (Interpreting types) Let ' = (D, , [I) be any A-model. An inter-
pretation of the type-variables is any function V that assigns a subset of D to each
type-variable. Each such V generates an interpretation of all the types, I Iv, defined
as follows:
(i) Ialy = V (a) for all type-variables a,
(ii) Ia->Tlv = {d E D : (Vd' E Iolv)(d-d' E ITIv)}.

4B6 Definition (Satisfaction) Let -9 = (D, , 11) be a A-model. A type-assignment

M:T is said to be satisfied in -9 by E and V if

IMIE E ITIV.
A set r of type-assignments is said to be satisfied if all its members are satisfied. A
formula r H M:T is said to be valid in .9 if every pair E, V satisfying F in -9 also
satisfies M:T in -9.
If I' --> M:T is valid in all A-models (respectively, all extensional A-models), we
say
F Ip M:T, IT =Bn M:T.
 4B Semantics and completeness 59

4136.1 Note The above definition of satisfaction is called the simple semantics for
type-assignment. There are several other semantics-definitions in the literature, for
example see Hindley 1983 §§4-5 and Mitchell 1988 §3.

Many different particular A-models are described in the literature (for some
examples see Barendregt 1984 Ch. 18 and HS 86 Ch. 12), but we shall not need to
know about them here. In fact the only model used in the completeness-proof below
will be the simplest of all kinds of model, a term-model.

4B7 Definition (The term-models TM#, TMpn) The domain D of TM# is the set
of all fl-equality-classes of A-terms: in more detail, for each M we define

[M]P = {P : P =s M},

and then define

D = {[M]0:M is a A-term}.

Application in TM# is defined thus:

[M]'-[N]l = [MN]I.
Terms are interpreted in TM# thus: if FV(M) = {xl,...,xn} and E(x1) = [Qi]fl for
i = 1, ... , n (for some terms Q1, ... , Q,,), define

IMIE = [[Q1/xl,...>Qnlxn]M]'.

The definition of TM& is similar, starting with [M]1n = {P : P =Qn M}.

4B7.1 Definition A semantic environment in TMp[n] called Eo is defined by setting

[x]Il[nl
E0(x) =

4B7.2 Lemma (i) TM# and T M#,, are A-models and T M#, is extensional.
(ii) In TMp[nl we have ME,= [M]PIll for all M.

Proof Routine. (For (i) see, e.g., Barendregt 1984 Prop. 5.2.12.)

4B8 Soundness Theorem Let I' be any type-context, M any A-term and r any type.
Then

I' Hp[nl M:-r F =p[nl M:i.

Proof Straightforward. (See, e.g., Ben-Yelles 1979 Thm. 4.17.)

4B9 Completeness Theorem Let r be any type-context, M any 2-term and i any
type. Then

r 9[n] M:i IF Ha[nl M:t.

60 4 Type assignment with equality
Proof [Hindley 1983a §3.] We shall prove the #-case; the proof of the /n-case
is almost identical. By the definition of "type-context", F is a finite set of type-
assignments
F = {x1 :Pl, ... , xm :Pm },

where xl,... , xm are distinct. Extend F to an infinite set F+ in which every type in
the language of TA2 is assigned to an infinite number of term-variables (and no
variable receives more than one type).
This can be done as follows. First list all the types as an infinite sequence, say
'G1,t2,.... Then note that in 1Al we assumed the A-language to contain an infinite
number of term-variables, so there are an infinite number of variables distinct from
xl,... , xm and from the free variables of M; choose from these an infinite number
of disjoint infinite sequences, say

{v1,1, v1,2.... }, {v2,1, v2,2, ...}, {v3,1, 03,2, ...}.....

Then define
F+=FU{v,,;:i;: i> 1,.j>_ 1}.

In what follows, the notation "F+ F-p P:a" (for any given P and a) will mean
that there is a finite subset r* of F+ such that r* I-p P:a.
In TM # define an interpretation vo of the type-variables by setting, for each
type-variable a,
vo(a) = {[P]: t+ F-# P :a}.

Extend vo to an interpretation I L. of all types as in 4B5. We shall now prove that,

for all a and all P :
(1) [P]f E IcrL,o F+ F-p P:v;
i.e. that the interpretation of a is exactly the set of all terms which receive type a
when the assignments in F+ are assumed.

Proof of (1) Use induction on a. If a is an atom we have (1) by the definition of

vo.
For the induction step let a By the definition of TM and of the
interpretation of l; we have

(VQ) ([Q], e I No [PQI13 E InIV,)

and by the induction hypothesis the right-hand side is equivalent to

(VQ) (r+ 1-0 Q: F+ F-fl PQ:n)

So to complete the induction step it is enough to prove
(2)
t b
((dQ) (F+ F-P Q:
F+ F-a
F+ F-p PQ:n)
P:(-*j).
 4B Semantics and completeness 61

Proof of (2) For ". ", use rule (-*E). To prove "=>", suppose the upper clause
in (2) holds and, as a special case, take Q to be a variable z not occurring in P
and such that r+ contains the assignment z exists.)
Then by the upper clause in (2),
I"+ Hp Pz:tl.
Hence by rule (-+1),
F+-z I-p
This means that there is a finite subset F* of F+ - z such that
F* I-p
hence by 4B9
F* I-p P: ->rl,

which gives the "=>" part of (2). Hence (2) holds and the proof of (1) is complete.

Deduction of the theorem from (1) If I' (=p M:T then in particular, for TMp and E0
and V O we have
[M]p E ITIVo
Hence by (1),
I" I-p M:z.
That is F* I-p M:i for some finite F* c F+. Hence by 4A3.2(vi),
(3) F* tM I-p M:r.
But F* P M = F P M because F* s r+ and the extra term-variables in F+ were
chosen to be distinct from those in M. Hence by the weakening-property of I-,
F 1-fl M:i.

4B10 Remark (Too many completeness-proofs?) In the above completeness-proof

the full strength of the assumption that F --* M:i was valid in every model was
not used; all we needed was that it was valid in one model, the term-model. There
is another completeness proof in Barendregt et al. 1983, which uses a more complex
kind of term model called a filter model. And there is a third proof in Coppo
1984 using the graph model Pw. It looks almost as if the choice of model is quite
irrelevant to the success of the completeness-proof.
Furthermore, TAA+p and TAA+p, can be proved sound and complete for at least
three other variant definitions of the semantics, see Hindley 1983a §5, Hindley 1983b,
and Mitchell 1988 §§3-4.
This richness of completeness-proofs says in effect that the differences which
undoubtedly exist between these various kinds of models and definitions of semantics
cannot be expressed in the rather limited language of TA2+p and TA2+p,.
62 4 Type assignment with equality
4B11 Remark By the way, although earlier remarks in this chapter may have given
the impression that without equality-invariance a type-theory has no reasonable
semantics, this is not strictly true. Plotkin 1994 has described a semantics for TA2
(and some extensions) based on modelling the concept of reduction rather than
equality. Plotkin's approach seems closer to the "spirit" of TA2; in fact TA2 has
at its heart the notion of reduction rather than conversion, and has always been
seen primarily as a tool for eliminating type-conflicts in practical computations or
reductions, rather than as a formal theory of equality.
 5

A version using typed terms

In Chapter 2 some care was taken to distinguish the Curry and Church approaches
to type-theory from each other. Curry's approach involved assigning types to pre-
existing untyped terms with each term receiving either an infinite set of types or
none at all, whereas in Church's the terms were defined with built-in types with
each term having a single type (see 2A3). In Curry's approach the types contained
variables, in Church's they contained only constants.
This book focuses on the Curry approach. However, even in this approach it turns
out to be very useful to introduce a typed-term language as an alternative notation
for TA2-deductions. Although the tree-notation introduced in Chapter 2 shows very
clearly what assumptions are needed in deducing what conclusions, it takes up a lot
of space and is hard to visualise when the deduction is in any way complicated. And
when manipulations and reductions of deductions are under discussion it is almost
unmanageable. A much more compact alternative notation is needed, and this is
what the typed terms in the present chapter will give.
We shall also define reduction of typed terms; typed terms will be shown in the
next chapter to encode deductions in propositional logic as well as in TA2, and
their reduction will be essentially the same as the reduction of deductions that is a
standard tool in proof theory.
By the way, a cynical reader might think we are simply abandoning Curry's
approach here and replacing it by Church's, but this is not so; the main positive
feature of Curry's approach is the expressive power gained from its use of type-
variables and the presence of an underlying language of untyped terms, and we shall
not abandon these. All we shall do is replace space-hungry deduction-tree diagrams
by neat and compact typed terms. And these terms will differ in several important
ways from those in a true Church-style system (see 5A1.5 for example); they will
simply be codes for TA2-deductions, nothing more.

5A Typed terms

In this section we shall define a system of typed terms isomorphic to TA2-deductions:

for each F a set TT(I') of typed terms will be defined that will encode deductions
of formulae of form
I'- -- M:T (F- s F).

63
64 5 A version using typed terms
It might be useful to read the detailed definition of a TA2-deduction in 9C in
parallel with this section.

5A1 Definition (Typed terms, TT(F)) Given a type-context F the set TT(F) of typed
terms relative to F is a set of expressions defined thus:
(i) if F contains x:a then the expression x° is in TT(F) and is called a typed
variable;
(ii) if F1 U F2 is consistent and M°-'T E TT(F1) and N° E TT(r2), then

(M°-'TN°)T E TT(F1 U r2);

(iii) if r is consistent with {x:a} and MT E TT(r), then

(Ax° MT )°-,T E TT(r - x).

If MT is a typed term (relative to some r), i is called the type of MT.

5A1.1 Notation Typed terms will be abbreviated using the same conventions as
for untyped terms. Also some type-superscripts may be omitted. For example,
depending on which of its types are to be emphasised in a particular discussion, a
typed term (xP-OyP)° may be called any of

(xy)°, x' yP, xP"°y.

5A1.2 Exercise (Compare 2A8.2-4.) Show that TT(O) contains the typed terms
(2xa-b yc-.azc C. (xa-b (yc-aZC)a)b)(a-.b)-(c-+a)-c-.b
(Axa . xa)a-.a

(Axayb. xa)a~b~a

5A1.3 Note (i) In the following term x appears decorated with two different types:
(x(a-a)_b(Axa.xa)a-.a)b
Qb =
Despite this, Qb is a genuine typed term, and in fact it is easy to verify that
Qb E TT(r), where F = {x: (a-*a)-*b}. Further, Qb translates into a genuine TA,2
-deduction in an obvious way.
(ii) In contrast, the following expression is not a typed term (and does not translate
into a TA2-deduction):
(xa-'bxI.

5A1.4 Warning The subterms of a term in TT(r) need not be in TT(I'). For
example we have
(2xa,xa)a-.a E TT(O)

but the subterm xa is not in TT(O). (It is in TT({x:a}).)

 5A Typed terms 65

5A1.5 Warning If M` and N° are typed terms we cannot always say that
(M°-'tN°)T is one. (Because if M°-'T E TT(r1) and N° E TT(F2) we cannot
apply 5A1(ii) unless F1 U F2 is consistent.)
This is a crucial difference between the present typed terms and those of Church,
for example in Church 1940; the latter are typed in an "absolute" sense and have
the property that if M°-'T and N° are typed then so is (M°-'tN°)t, but the present
terms are only typed relatively to a given F.

5A2 Definition (Type-erasing) The type-erasure Mf of Mt is the untyped term

obtained by erasing all types from Mt.
We shall see in 5A5.1 that Mf is the subject of the conclusion of the TA2-deduction
that Mt encodes.

5A3 Lemma If Mt E TT(I'), then FV(Mf) c Subjects(F) and

(i) Mt E TT(r'+) for all contexts F+ ? F,
(ii) Mt E TT(F rMf),
(iii) If Mt E TT(F') then r' r Mf = F r Mt .

Proof Straightforward induction on JMf I.

5A4 Definition (Minimum context, Con(Mt )) If Mt E TT(r), its minimum context

is
Con(Mt) = F rMt.

5A4.1 Note Con(Mt) is determined uniquely by Mt. In particular it is independent

of F, in the sense that if Mr E TT(F) n TT(r') then IF rM1 = r' t Mf (by 5A3(iii)).
We shall see in 5A7 that if Mt encodes a TA2-deduction of F --* Mf :T then
r = Con(MT). Also it is easy to see that
(i) for all typed terms M°-'t and N°
(a) (M°-'tN°)t is a typed term if Con(M°-'t) is consistent with Con(N°),
(b) if Con(Ma-t) is consistent with Con(N°) then

Con((M "tN°)t) = Con(M°-'t) U Con(N°);

(ii) for all typed terms Mt
(a) (2x°.Mt)Q-'t is a typed term if Con(Mt) is consistent with x:Q (i.e. if
Con(MI) either contains x:a or does not contain x at all),
(b) if Con(Mt) is consistent with x:6 then
Con((1x°.MT)°-'t) = Con(Mt) - X.

The translation mappings between typed terms and TA2-deductions will now be
defined; they are very straightforward but will be given here in full for the sake of
precision.
66 5 A version using typed terms
5A5 Definition (Translating deductions to typed terms) Let A be a TAR-deduction
with conclusion F --> M:T; define a typed term To E TT(F) by following the clauses
of the definition of A (see 9C1) thus:
(i) if A is an axiom x:T H x:T, define To - xt;
(ii) if A is the result of applying (-+E) to deductions A1, A2 of
I'1 I--) P : a--+T, r2 I-)
Q:a,

and TZT T E TT(F1) and Toe E TT(172), define

To = (TA"' 1 )t;
(iii) if A is the result of applying rule (-+I) to a deduction Al of
F I-- P :U'

and T - p--+a and F U {x:p} is consistent and Tot E TT(F), define

To = (AxP.Te1)P-Q

5A5.1 Lemma If A is a TAA-deduction oft I-+ M:T then

TAT E TT(F), TA = M.

5A6 Definition (Translating typed terms to deductions) To each typed term MT we

associate a TAR-deduction A(Mt) as follows:
(i) if Mt = xt define A(Mt) to be the atomic deduction
x:T I-+ x:T;

(ii) if MT = (P°-'tQ°)t, P°-'t E TT(I'1), Q° E TT(I 2), i 1 U F2 is consistent, and

0(P°-'t) and 0(Q°) are deductions of
F' '-. p4"t : a-+r, I'2 F-+ Qa:a

for some r' - 171 and F' s F2, define NA MT) to be the deduction obtained by
applying (-+E) to A(P°-'t) and A(Qa);
(iii) if MT = (2xP Pa)P- T p- o, P° E TT(F), r is consistent with {x:p}, and
A(Pa) is a deduction of
F' i--> Pd:a
for some t' s F, define A(Mt) to be the deduction obtained by applying (-AI)
to A(Pa), discharging x.

5A7 Theorem The translation from A to To is a one-to-one correspondence between

all TAR-deductions and all typed terms, and its inverse is the translation from Mt to
A(MT). In particular,
(i) if A is a deduction of FI-+ M:T then
TAT E TT(F), TX M, A(TA) A,
 5B Reducing typed terms 67

(ii) if MT E TT(F) then Con(MT) c F and L(MT) is a deduction of

Con(MT) F-+ M{:'r.

Proof (i) By 5A5.1 and 5A6. (ii) Easy induction on IMfI.

5B Reducing typed terms

Most of the term-structure concepts introduced in Section 1A, and those in 9A, can
be extended from untyped to typed terms in an obvious way. For example position,
subterm, occurs and component are all defined for typed terms just as for untyped
terms in 9A.
The only difference is that in a typed term every subterm is decorated with a
type-superscript; in particular an atom has form xT not just x. For example if
MT = (Xa-'by°)b we shall say that x°-b and ya occur in M1 but not that x and y
occur in MT.
Some of the concepts in IA and 9A do not translate to typed terms quite so
unambiguously however, so their typed definitions will be given below and their
main properties will be stated.

5B1 Definition The length, IMTI, of a typed term MT is the same as JM11, see 1A2.

5B2 Definition (Replacement) Let P° _ (Pa, p, MT) be a component of a typed term

MT (see 9A2-3), and let T° be any typed term with the same type as P°. Define
{T°/Pa}PMT
to be the result of replacing P° by T° at position p in MT, just as for untyped terms
(see 9A6), but only when (i) or (ii) below holds:'
(i) Con(Tc) s Con(P°);
(ii) the set is consistent, where L1,',...,.lx"
are the abstractors in MT whose scopes contain P°.

5B2.1 Lemma (i) If 5B2(i) holds then {Ta/P°}PMT is a well-defined typed term and
Con({Ta/P°}PMT) 9 Con (MT).
(ii) If 5B2(ii) holds then {T°/P°}PMT is well-defined and
Con({Ta/P°}PMT) c Con(MT) U Con(T°).

Proof Like the proofs of 9C5 and 9C6.

5B2.2 Note If neither (i) nor (ii) holds in 5B2 we do not define {Ta/P°}PMT. In
this case the replacement of P° by T° in MT will still produce an expression of
some kind but it might not be a typed term. For example let
(xa-bya)b, Ta =
MT = P° = yQ,
xa.

Compare the conditions in the replacement lemmas for deductions, 9C5-6.

68 5 A version using typed terms
Then replacing ya by x° changes Mt to (xa-,bxa)b, which fails to be a typed term
because x:a-*b is inconsistent with x:a. (See 5A1.5.)
Also if p # o we do not define {TP/Pa}PMt.

5B3 Definition Free and bound variable-occurrences, and the set FV(Mt), are defined
just as in 1A6. (But all occurrences are now typed.)

5B3.1 Notes (i) x° E FV(Mt) x:v E Con(Mt).

(ii) x E FV(M1) (3a)(xa E FV(Mt)).
(iii) By (i) and the consistency of contexts,
xa,xP E FV(Mt) 6 = p.
(iv) But a typed term can contain inconsistently typed variables if one variable is
free and the other is bound (or if both are bound). An example is the term Qb in
5A1.3:
Qb =

5B4 Definition Mt has no bound-variable clashes if Mf has no such clashes (as

defined in 1A9).

5B4.1 Example Consider the term Qb above and its type-erasure:

(x(a-a)-b(Axa.xa)a_a)b,
Qb = Q' = x(2x.x).
This erasure has a clash so the above definition says Qb has a clash. (We might
feel like saying that Qb is without clashes because xa-a-b and xa are distinct typed
variables, but this would make the theory of typed terms diverge too far from that
of TA2-deductions. Definition 5B4 says in effect that a typed term has no clashes if
it represents a deduction whose conclusion has no clashes.)

5B5 Definition (Typed substitution) We define [Na/xa]Mi by the following clauses.

(Compare untyped substitution, 1A7.) These clauses will not cover all cases but all
those in which (2xa.Mt)Na has no bound-variable clashes, which are all the cases
needed in this book.
(i) [Na/xa]xa = Na;
(ii) [Nalxa]Yt = Yt ifx #Y;
([Nalxa]PP-,t[Na/xa]QP)t;
(iii) [N'/xa](PP~tQP)t =
(iv)
(v) [Nalxa](.Y (,ly if x # y, x FV(P');
(vi) [Nalxa](Ay''Pn)1-1
=
ifx # y, x E FV(P'i), y FV(Nd);
(vii)
if x # y, x E FV(P'i), y E FV(Nd).
(In (vii) z is the first untyped variable not occurring in NtPI.)
 5B Reducing typed terms 69

5B5.1 Notes The restrictions in the above clauses may seem over-strong in that they
involve untyped variables and terms; but if they were weakened the last part of the
useful lemma 5B5.2 below would fail.
Clause (iv) includes the case 1 # o. But in this case x° cannot occur free in P",
because if it did then would not be a typed term (see Definition 5A1(iii)).
Thus substituting for x° should change nothing. And this is exactly what (iv) says.
No attempt has been made to define [NP/x'] when p # o or [N9/xa]xt when
'r*0.
5B5.2 Lemma If ((2xa.Mt)Na)t is a typed term with no bound-variable clashes, then
(i) [Na/xa]Mt is defined and is a typed term with type r,
(ii) [Na/x°]Mt - Mt ifxa 0 FV(Mt),
(iii) Con([N'/xa]Mt) c (Con(Mt) - x) U Con(N'),
(iv) ([Na/xa]Mt)f =a [N'1/x]Mf.

Proof Parts (i)-(iv) are proved together by a straightforward but boring induction on
IMt 1. (The assumption about ((2xa.Mt)Na)t implies in particular that Con(MI) - x
is consistent with Con(Na) and that x does not occur in Nd.)

5B6 Definition (Typed a-conversion) The relation -a is defined just as for untyped
terms (see 1A8), using replacements with form
(a)
Axa'Mt AY°'[Ya/xa]Mt FV(Mf))
(y

5B6.1 Lemma If Ax°.Mt is a typed term and y FV(Mf ), then [ya/xa]MT is defined
and both sides of (a) are typed terms with the same type and minimum context. Hence
the class of all typed terms is closed under a-conversion, and

Pt =a Qt' i - i and Con(Pt) = Con (Q").

5B6.2 Warning The condition "y FV(M1 )" in (a) cannot be weakened to "ya 0
FV(Mt)". Because if it were, we could a-convert a typed term to an expression that
was not one, thus:
Axa-+b , xa-b ya Aya-b, ya.b ya
=a

5B7 Definition (Typed redexes and reduction) Typed >p and v#,, are defined just like
the untyped relations in 1B and 1C. In particular typed f- and q-redexes have form

((2x°'Mt)a~tNo)t,
with x 0 FV(P'N'f), and their contracta are, respectively,
[Nalxo]Mt, Pa-.t

And, just as in Chapter 1, a contraction is a replacement of a redex by its contractum.

A typed fl-nf (q-nf, /q-nf) is a typed term containing no fl-(rl-, $i -) redexes.
70 5 A version using typed terms
5B7.1 Lemma (i) If a /1- or q-contraction in a typed term Pt changes Pt to an ex-
pression X, then X is also a well-defined typed term with type T.
(ii) If Pt Dan Qt, then
(a) Con(PT) 2 Con(Qt),
(b) Pf r131 Qf.

Proof Prove (i) and (ii) together: for $ use 5B5.2 and 5B2.1(i), and for n the proof
is straightforward.

5B7.2 Note (i) When typed terms are interpreted as TA2-deductions a /3-reduction
of a typed term corresponds to a reduction of a deduction. (The idea of reducing
deductions originated in Prawitz 1965 and there is a modern account in Troelstra
and van Dalen 1988 Vol. 2 Ch. 10 §2; cf. also HS86 Def. 15.30.) In this interpretation
the above lemma says that if we reduce a deduction of F H P:r, the result will
still be a genuine deduction and will have conclusion I'` i--). Q:T for some F* c--- r
and some Q obtained by #-reducing P.
(ii) Part (ii) of the above lemma is rather like the subject-reduction theorem
for TA2-deductions (2C1), but is not an exact analogue. A closer analogue is the
following lemma which will also be used later.

5B7.3 Lemma (i) Let Pt be a typed term and Pf D19 Q by contracting a f-redex R
at a position r in Pf. Then Pt contains at position r a typed $-redex R° such that
Ra = R, and contracting this R° changes Pt to a typed term Qt such that
Qf = Q.
(ii) The same holds true for rl-redexes.

Proof Same as the proof of 2C1 but with typed terms instead of TA2-deductions.
0
5B7.4 Lemma (i) A typed term Pt has a /3-reduction with length n iff Pf has a 1-
reduction with length n.
(ii) A typed term Pt is a f-nf iff Pf is a f-nf.
(iii) Both (i) and (iii) hold for n.

Proof Note first that if a typed term Pt contains a fl-redex then so does P1. Then
(i)-(iii) follow from this and 5B7.3.

5B8 Typed Church-Rosser Theorem (i) If Mt >p Pt and Mt >p Qt then there exists
a typed term Tt such that Pt >p Tt and Qt us P.
(ii) The same holds for t1- and fltl -reductions.

Proof Like the untyped Church-Rosser theorems in 113 and 1C.

5138.1 Warning (No typed convertibility relation) No attempt is made here to define
a convertibility relation =p or =p, on typed terms by expansions and contractions
 5C Normalization theorems 71

of redexes. First, we shall not need one. Second, since the typed terms in this
chapter correspond exactly to TAT-deductions, any attempt to define such a typed
equality would meet the same type-variation problems as were discussed for TAT in
2C. In fact, f3-expanding a typed term may lead to an expression which is no longer
a typed term (cf. 2C2.2-6).

5C Normalization theorems
Along with most other type-theories, TAB, has what is called the weak normalization
property:
(WN) every typable term has a normal form.
And many of them, including TA2, also have the strong normalization property:
(SN) all reductions of a typable term are finite.
These properties were stated for TA2 in Theorems 2D5 and 2D6 without proof. In
the present section a proof will be given of the analogue of WN for typed terms, and
WN for typable terms will be deduced from it. No proof of SN will be given here
as there is already a detailed proof in our main reference (see HS86 Appendix 2).

5C1 Weak Normalization (WN) Theorem (Turing 1942, Curry and Feys 1958
Cor. 9F9.2, etc.) Every typed term P" has both a fl-nf and a f3n-nf.

Proof The existence of a frl-nf follows from that of a fl-nf by the typed analogue of
1C9.4. It remains to prove that P" has a fl-nf, i.e. to give a strategy for reducing P"
and to prove this reduction terminates. This will be done by the method of Turing
1942, the earliest proof known and also the simplest. Some knowledge of 9B will be
needed.
Let P' E TT(r). Then each fl-redex-occurrence RT in P" has form
RT = ((,.xo'MT)a-.TNa)T
We shall call the function part of RT and its type a--+T the dominant type
of RT, and the number of atom-occurrences in Q-*r the degree of RT :
Deg(RT) = lo-,rl
Now let the contraction of RT change P" to Q', and consider the residuals in
Q" of fl-redex-occurrences in P'. (Residuals are defined in 9B2.) By 9B2.2(iii) the
degree of each residual is the same as that of the residual's parent in PR.
Further, by the conclusion of 9B5.1, for each newly created fl-redex-occurrence in
Q" (i.e. one that is not a residual of one in P') its function-part is an occurrence
of either Na or [N/x]MT ; thus its type is either a or z, and so the degree of each
newly created redex-occurrence is strictly less than the degree of RT.
Now let d(PI) be the maximum of the degrees of all fl-redex-occurrences in P"
and let S be the set of all fl-redex-occurrences in P" with this degree. Reduce P"
by a minimal complete development of S as defined in 9B4. By 9B4 and 9B4.1 this
reduction is finite and its result is a term Pi containing no residuals of members of
72 5 A version using typed terms
S. All the redexes in P, are either newly created or residuals of redexes not in 5;
hence
d(P1) < d(P").
Continue this procedure by making further minimal complete developments of sets
of redexes with maximum degree; since each development strictly decreases d(P")
we must eventually obtain a fl-normal form of P".

5C1.1 Corollary The WN theorem for typable terms (2D5) holds.

Proof By 5B7.4 and the WN theorem for typed terms.

5C1.2 Historical notes The first known proof of WN for a type-theory equivalent
to TAx was written around 1941 or '42 in unpublished notes by Alan Turing. (See
Gandy 1980a for a transcript of Turing's notes, and Gandy 1977 pp. 178-180 for
some comments on Turing's work in type-theory.)
But the first proof to be actually published was carried out by Curry in the late
1950's by a completely different technique (Curry and Feys 1958 §9F, especially
Cor. 9F9.2). Its key step was a cut-elimination theorem for a particular formulation
of TACL.'
Also there is a WN theorem in Prawitz 1965 for reducing proofs in logic which is
essentially equivalent to WN for TA2.
From the 1960's onwards the Turing method of proof was re-discovered, probably
independently, several times (Morris 1968 §4F Thm. 2, Andrews 1971 Prop. 2.7.3,
for example), and WN theorems were proved by various methods for many stronger
type-theories than TA2.
In particular, the mid-60's saw a spate of proofs of WN for typed .1-calculi
enhanced by primitive recursion operators (for example those in Tait 1965, Hanatani
1966, Hinata 1967, Sanchis 1967, Tait 1967, Diller 1968, Dragalin 1968 and Howard
1970); see Troelstra 1973 §§2.2.1-2.3.13 or HS86 Ch. 18 for descriptions of the
background setting. At least two of these also included proofs of SN (see below).

5C2 Strong Normalization (SN) Theorem (Sanchis 1967 Thm. 8, Diller 1968 §6,
etc.) Let P" E TT(F) for some F; then, for /3- and for /hj-reductions,
(i) all reductions of P" are finite,
(ii) there is an algorithm which accepts P" as input and outputs a number k(P") such
that all reductions of P" have length 5 k(P").

Proof For (i) there is an accessible proof in HS86 Appendix 2: see Thm. A2.3 for
/3, and Thm. A2.4 for fl q. Alternatively, see Barendregt 1992 Thm. 4.3.6.
For (ii), simply construct a finitely branching tree of reductions starting at P" by
doing all possible contractions at each step. By (i) this tree's branches are all finite
Cut-elimination originated in the study of predicate logic in Gentzen 1935, and the relation between
cut-elimination and normalization is explored in Zucker 1974. It depends on the correspondence
between formulae and types to be described in the next chapter. Cut-elimination theorems for versions
of TAx can be found in Seldin 1977 and 1978.
 SC Normalization theorems 73

in length, and the famous Konig's Lemma on trees states that in this case the whole
tree must be finite. Hence its branches can all be measured and the maximum of
their lengths determined. Call this number k(P').
But this algorithm is of course inefficient and the proof that k(P") is well-defined
depends on Konig's Lemma which is usually regarded as non-constructive. To
remedy these defects several workers have devised constructive proofs of (ii); see
Mints 1979, Gandy 1980b, de Vrijer 1987 and Schwichtenberg 1991. These proofs
give more efficient ways of computing suitable bounds k(PE) (though the bounds
they give are not always the least possible, see comments in de Vrijer 1987).

5C2.1 Corollary The SN theorem.for typable terms (2D6) holds.

Proof By 5B7.4(i) and 5C2.

5C2.2 Note The history of SN began over twenty years later than that of WN. The
first known explicit SN proofs were in Sanchis 1967 Thm. 8 and Diller 1968 §6, the
former for weak reduction in combinatory logic enhanced by primitive recursion
operators and the latter for A$-reduction similarly enhanced.'
Most published proofs have depended on defining what is usually known as a
computability predicate by a suitable induction, and proving first the SN property
for all computable terms and then the computability of all terms P" by induction
on JP"i. (The proof in HS86 Appendix 2 is typical of this approach, cf. also the WN
proof in Tait 1967.) This method has been successfully adapted to so many different
type-theories over the years that it is now in effect the standard one. In the early
1970's it was applied by Jean-Yves Girard in a generalised and strengthened form
to prove SN for his strong second-order type-theory known as System F (Girard et
al. 1989 Ch. 14), and it has since been applied to other second-order type-theories.
An analysis with references is in Gallier 1990. Another account is in Barendregt
1992; Barendregt's System A2 is equivalent to Girard's System F.

1 In each case manuscript versions were made available a couple of years before publication.
 6

The correspondence with implication

One of the most interesting facts about TA,2 is that there is a very close correspon-
dence between this system and propositional logic, in which the types assignable
to closed terms in TA1 turn out to be exactly the formulae provable in a certain
formal logic of implication. This correspondence is often called the Curry-Howard
isomorphism or the formulae-as-types isomorphism, and will be studied in this chapter.
The logic involved in this correspondence is not the classical logic of truth-tables
however, but that of the intuitionist philosophers; it will be defined in the first
section below.
The Curry-Howard isomorphism was first hinted at in print in Curry 1934 p.588,
and was made explicit in Curry 1942 p.60 footnote 28 and Curry and Feys 1958 §9E.
But it was viewed there as no more than a curiosity. The first people to see that it
could be extended to other connectives and quantifiers and used as a technical tool to
derive results were N. G. de Bruijn, William Howard and H. Lauchli in the 1960's.
See Howard 1969, de Bruijn 1980 (an introduction to de Bruijn's AUTOMATH
project which began in the 1960's), and Lauchli 1965.
This chapter will also define three rather interesting subsystems of intuitionist logic
and show that they correspond to the three restricted classes of 2-terms defined in
Section 1D. This correspondence was first noted by Carew Meredith in unpublished
work around 1951 and was explored in detail in the thesis Rezus 1981.

6A Intuitionist implicational logic

In this section we shall define the implicational fragment of intuitionist logic.

The original formulations of this logic in the 1920's were axiom-based systems and
corresponded more closely to combinatory logic than A-calculus, but it will be defined
here in the more modern "Natural Deduction" style due to Gerhard Gentzen.
There is an elementary introduction to Natural Deduction in van Dalen 1980 §1.5,
and more detailed accounts are in Gentzen 1935, Prawitz 1965, and Troelstra and
van Dalen 1988 Chs. 2 and 10. But the style is so like the definition of TA2 in 2A8
that the short introduction below will probably be plenty enough for the reader.[

This similarity of style is no coincidence; when the earliest formulations of TA,2 were written Gentzen's
methods were just beginning to be known more widely; indeed Curry was one of the first to understand
their importance.

74
 6A Intuitionist implicational logic 75

As mentioned earlier, the formulae provable in the logic defined below will turn
out to be exactly the types of closed A-terms. However, we can also go one step
deeper than this and get a correspondence between the proofs of these formulae and
the A-terms themselves. But to make this deeper correspondence work we must be
very careful about the definition of "proof" and "deduction" in a Natural Deduction
system, and we must clarify some features of such deductions that are not often
emphasised in the standard literature on the topic.
Therefore some of the definitions below will involve position-labels and other
tedious details that have so far been confined to Chapter 9. These will be needed
when determining exactly which deductions correspond to which A-terms.

6A1 Definition Implicational formulae (denoted by lower-case Greek letters other

than "A") are built from propositional variables (denoted by "a", "b", "c", ...) using
the implication connective "->" thus: if a and r are formulae, then so is (a--*r). (An
infinite sequence of propositional variables is assumed to be given.)

6A1.1 Note (i) Of course implicational formulae are exactly the same as types.
Parentheses will be omitted using the same association-to-the-right convention as
for types, see 2A1.1.

(ii) There are no propositional constants and no other connectives than

(Because this book is only concerned with the Curry-Howard correspondence in its
simplest and smoothest form.)

6A1.2 Background The intuitionist philosophy of mathematics was developed in the

early 1900's by L. E. J. Brouwer and his group in Holland, and a very readable
introduction is in Heyting 1955. One feature of this philosophy is that some propo-
sitional formulae that are usually accepted as universally valid, and are tautologies
in the usual truth-table sense, are regarded by the intuitionists as having only limited
validity. One such formula is

(PL) ((a-+b)-->a)--*a,

which is known as Peirce's law. The logic developed by the intuitionists has attracted
continuing interest over the years from many logicians and computer scientists quite
independently of their philosophical views. For example many of the polymorphic
type-theories in the current literature have intuitionist logic as their basis, not
classical logic.

The formulae provable by the rules in the next definition can be shown to
coincide exactly with the implicational formulae that the intuitionists accept as
universally valid. (See Troelstra and van Dalen 1988, Ch.2 (System IPC) and Ch.10
§5 (separation theorem).)
76 6 The correspondence with implication
6A2 Definition Intuitionist logic (or more precisely, its implicational fragment) has
the following two rules:
Q-+T a
(-+E) : (-+I) : [o]

o.-+T.

Each application of (-+I) is said to discharge (or cancel) some, all, or none of the
occurrences of a above i, and must be accompanied by a discharge label that lists
all the occurrences of a it discharges (see below for details). If none are discharged
the application of (-+I) is called vacuous. Discharged occurrences of a at the tops
of branches must be marked by enclosing them in brackets.1
Before formally defining deductions and proofs in this system it will be helpful
to look at some examples of how the rules are used. In these examples a position-
notation like that in 9A1-4 will be included in the deductions to allow the discharge
labels to say precisely which occurrences have been discharged. The position of each
formula will be written in parentheses beside it, and discharge labels will be written
between braces "{ }".

6A2.1 Example A proof of (a--+a-->c)-+a-+c.

[a-+a--+c] (0011) [a] (0012)
(-+E)
a--+c (001) [a] (002)
(-+E)
c (00)
(-+I)
a-+c (0) {discharging a at 0012, 002}
(-+I)
(a--+a-+c)--+a.-a-+c (0) {discharging a-+a-+c at 00111

6A2.2 Example A proof of (a-+a-+c)-+a--+a-+c.

[a-+a-+c] (00011) [a] (00012

a--+c (0001) [a] (0002)

(-+E)
c (000)
(-+I)
a--+c (00) {discharging a at position 00012}
(-+I)
(a--+a--+c) (0) {discharging a at 0002}
(-+I)
(a-+a-+c)--+a-+a-*c (0) {discharging a-*a--*c at 000111

6A3 Discussion (Partial discharging) The proof in 6A2.2 contains an example of

what will be called partial discharging. Suppose rule (-+I) in 6A2 is applied to a
deduction of -r in which or has n >_ 1 occurrences as assumptions. Definition
The system defined above is often called minimal logic. But we shall later be interested in subsystems
of this system so "minimal" seems inappropriate here.
 6A Intuitionist implicational logic 77

6A2 allows (->I) to leave some of undischarged. For example in 6A2.2

the uppermost application of (-*1) has two occurrences of a as assumptions above
it but only discharges one.
It is standard practice to allow this in a Natural Deduction system, although
few introductory accounts of Natural Deduction emphasise that this is so. Here we
shall say that a system which allows it obeys the partial discharge convention. This
convention is discussed in Prawitz 1965 Ch. I §4, Leivant 1979, and Troelstra and
van Dalen 1988 Vol. 2, pp.559-560 and p.568 Ex. 10.8.1. (In the extreme, it allows us
even to make a vacuous discharge of a when n > 1; this would deduce o- ,r
and leave all of undischarged! An example of this extreme is in
6A3.3.)
In any system obeying this convention each occurrence of (->I) in a deduc-
tion must carry a label listing all the occurrences of a that it discharges. These
discharge-labels play so little role in standard accounts of natural deduction
that they are almost always omitted, but when we come to make a correspon-
dence between proofs and 2-terms they will be essential because proofs with
different discharge-labels will correspond to different terms. We shall see this in
6B.
By the way, partial discharging is forbidden in TA2. In fact (-*1) in 2A8 has
an effect equivalent to discharging every occurrence of x:a in F that has not
been discharged earlier; we shall say TA2 obeys the complete discharge conven-
tion.

6A3.1 Example The following is a proof of (a--*.a->c)-->a-*a->c that is identical to

that in 6A2.2 except for having different discharge-labels.
[a-->a--*c] (00011) [a] (00012) (--+E)

arc (0001) [a] (0002)

(-+E)
c (000)
(-*1)
a->c (00) {discharging a at 00012, 0002}
(-*1)
a-+a-+c (0) {discharging a vacuously}
(-*1)
(a-- a-*c)->a-*a-+c (0) (discharging a--+a-+c at 00011}

6A3.2 Note Comparing 6A2.2 with 6A3.1 might suggest that a succession of
partial discharges of the same formula can always be replaced by one complete
discharge followed by some vacuous ones. And this suggestion is correct. (Though
it would fail in restricted logics that forbid vacuous discharging.) Thus if partial
discharging were forbidden the class of all provable formulae would not be reduced.
However, the class of proofs would then become more restricted than the class of
A-terms.

6A3.3 Example The following proof illustrates an earlier remark that the partial
discharge convention allows vacuous discharging of a even when the set of available
78 6 The correspondence with implication
occurrences of a is not empty.

[a] (00) (-1)

a-*a (0) {discharging a vacuously}
(-,I)
a->a--*a (0) {discharging a at position 00}

Hopefully this and the earlier examples have given the reader some feeling for
the concepts involved in a natural deduction. The formal definition of "natural
deduction" will now be given for precision's sake, though the reader may omit
it without much loss. In it, a natural deduction will be a tree in which each
node will carry either two or three labels: (1) a formula, (2) a position, and
(3) either a discharge-label (if the node represents the conclusion of (-+I)) or a
pair of brackets (if the node represents a discharged assumption), or no third
label.

6A4 Definition Natural deductions in the Intuitionist logic of implication (or logic
deductions for short) are trees, defined thus.

(i) An atomic deduction is a single node with two labels r, 0, where t is a formula
and 0 is the empty position.

(ii) If Al and A2 are natural deductions whose bottom nodes are labelled respec-
tively by o-*t, 0 and a, 0 (and possibly a third label each), define a new deduction
by first putting "1" on the left end of each position-label in Al and "2" on the
left end of each position-label in A2, next doing the same to the positions in any
discharge-labels in Al and A2, and finally placing an extra node beneath the two
modified deductions, as shown below.

Modified . . Modified .

° °2
6T `2
(iii) If A, is a natural deduction whose bottom node is labelled by r, 0 (and
possibly a third label), and c is any formula, define a new deduction thus. First
choose a (possibly empty) set of unbracketed occurrences of a at the tops of branches
in A, and label them with brackets, then put "0" on the left end of every position
in Al, then place an extra node beneath the modified Ai as shown in the diagram
below, and finally give this node three labels: (1) a-->i, (2) 0, and (3) a discharge-label
listing the positions of the chosen occurrences of o. (If the set of chosen occurrences
is empty we make the discharge-label say "vacuous".)
 6B The Curry-Howard isomorphism 79

T 0

I
0 {discharging assumptions at p ...,pk }

6A4.1 Notation In practice natural deductions are never displayed in the space-
wasting form shown in these diagrams, but in the horizontal-line form used in the
examples earlier.

6A4.2 Definition Two occurrences of an assumption at branch-tops in a natural

deduction A will be called discharge-equivalent if they are discharged by the same
occurrence of (-.I), i.e. if their names appear on the same non-vacuous discharge-
label in A.

6A5 Definition (Deducibility and provability) Iff there is a deduction whose con-
clusion is T and whose undischarged assumptions are occurrences of members of
we say
Fint T
(This notation may be used when a,_., on are not distinct.) Iff n = 0 the deduction
is called a proof and T a provable formula or theorem, and we say
dint T.

6B The Curry-Howard isomorphism

The similarity between rules (-+E) and (->I) in the previous section and the TA2-
rules (-+E) and (-+I) in Chapter 2 suggests that every TA2-deduction might be
changed into a logic deduction by stripping off all the contexts and terms in it.
This is in fact true, modulo a few minor points like differences in the discharge
conventions. The details are as follows.

6B1 Definition (Curry-Howard mapping from lambda to logic) If A is a TA,2-deduc-

tion of r --> M:T (see 2A8), the corresponding logic deduction AL is defined thus.
(i) If M x and A is x:T H x:T, choose AL to be just T.
(ii) If M = PQ and r = r, U r2 and the last step in A has form

(Al) (02)
r, H P:a- T r2 H Q:a
(-.E)
F, U F2 H (PQ):T,
construct AL by applying rule (-*E) of 6A2 to A,L and A2L.
80 6 The correspondence with implication
(iii) If M - Ax-P, r = p--+a, r = F'-x and the last step in A is
(A')

r' F--+ P
F'-x --> (Ax'P): (p-+v),

construct AL by applying rule (-+I) of 6A2 to AL and discharging all occurrences

of p in AL whose positions are the same as the positions of the free occurrences
of x in P.

6B1.1 Example If A, is the following TA2-deduction, (AI)L is the deduction in 6A2.2.

x:a-+a-+c --> x:a-+a-- c z:a --> z:a
(-+E)
x:a-).a-+c, z:a i-* y:a H y:a
(-+ E )
x:a->a->c, y:a, z:a ti (xzy):c
I { di sch arg i ng z}
x:a-+a-*c, y:a i--*
I { di sch arg i ng y}
x:a-+a-+c --+
I {di sc h arging x}
F-+

6B1.2 Example If A2 is the TA1-deduction below, (A2)L is the deduction in 6A3.1.

x:a-+a-+c '--+ x:a--+a-+c z:a --+ z:a
(-E)
x:a--+a-+c, z:a F-+ (xz):a-+c z:a F--+ z:a
(-+E)
x:a-+a-+c, z:a F--+ (xzz):c
(-+I) {discharging z}
x:a-+a-+c F-+
(--+I) {vac. dis. y:a}
x:a-*a-+c F-. (Ayz xzz) :a-*a-+c
(-+I) {discharging x}
F-,

6B1.3 Note The A-terms to which AI and A2 assign types in the two above examples
are distinct, although they have the same tree-structure. Correspondingly (AI)L and
(A2)L are distinct, though they only differ in their discharge-labels.

6B2 Lambda-to-Logic Lemma Let x1,...,xn be distinct and A a TA2-deduction of

xl:Pl,...,xn:pn F-+ M :T.
Then AL is a natural deduction in Intuitionist logic and gives
PI,...,Pn H T.

Proof Induction on IMO. El

 6B The Curry-Howard isomorphism 81

6132.1 Exercise By applying the Curry-Howard mapping to the TAA-proofs in 2A8.2-

4 and the answers to 2A8.7, show that the following formulae are provable in Intu-
itionist logic. (Their names below date from the 1950's and are now standard in the
logic literature; of course they are not accidental.)

(B) (a-+b)-+(c-+a)-*c->b,
(B') (a-->b)-+(b-*c)-->a-*c,
(I) a--+a,
(K) a--*b-+a,
(C)
(W) (a->a->b)->a--*b.

6B3 Warning The Curry-Howard mapping is not one-to-one. For example consider
the following TA2-deductions A3 and A4 which assign types to xyz and xyy.
A3: x:a-+a-+c --+ x:a-+a-+c y:a --+ y:a
(-+E)
x:a-+a--+c, y:a H xy:a-+c z:a '-+ z:a
(-+E)
x:a-*a-+c, y:a, z:a F+ xyz:c

A4: x:a-+a-c F--+ x:a-+a-+c y:a -+ y:a

(-+E)
x:a-4a-+c, y:a - xy:a-+c y:a '--* y:a
(-*E)
x:a->a-+c, y:a --> xyy:c

It is easy to check that the Curry-Howard mapping assigns the same logic
deduction to both, namely
a-+a-+c a
(-+E)
a-+c a
(---+E)
c

(More generally, if A is a TAA-deduction of 1 i--+ M:T and x and y receive the

same type in F and we change A to a deduction of
[v / x, v / y ] I- - [v / x, v / y ] M :T,

then AL will not change.)

Since the Curry-Howard mapping is not one-to-one it cannot have an inverse, i.e.
there is no logic-to-lambda mapping ( )x such that
(1) (AA)L A (for all logic deductions A),
(2) (AL)A A (for all TAA-deductions A).
However there is a "one-sided inverse", a mapping ( )A satisfying (1) alone. In fact
there are several such inverses, and the next definition will describe one of them; it
will be obtained by choosing term-variables to be distinct whenever possible.
82 6 The correspondence with implication
6B4 Definition (A logic-to-lambda mapping) To each natural deduction A in Intu-
itionist logic we assign a TA2-deduction AA thus.
(i) If A is a single formula i, choose any term-variable x and let AA be

x:T H x:T.
(ii) If the last step in A is (-*E) applied to the conclusions of deductions A' and
A", and A' and A" have been defined and are deductions of
I'' --* M : o -vr, r" H N :a,
then replace all term-variables in A" by distinct new ones (to make it have no
term-variables in common with A', neither free nor bound), and apply the TA2-rule
(-+E); call the resulting deduction Ax.
(iii) If the last step in A is an occurrence of (--+I) with form

a deduction A'
{discharging k >_ 0 occurrences p1,...,pk of p},
p-+a

and A' has been defined and its conclusion is

where v1....,vk are distinct from each other and from the variables in F, and vi,...,vk
occur free in P at the same positions as pi, ... , pk have in A' (each v; occurring only
once in P), then proceed as follows.
If k >_ 1, replace all of v1,...,vk in Az by one new variable x that does not occur
elsewhere in A'; this changes A' to a deduction of
f,x:p > P*:v (P* = [x/vi,...,x/vk]P)
where x occurs in P * at exactly the same positions as p1,. .. , pk have in A. Then
apply (-*I)ma;,, to this modified A' and call the resulting deduction A2. Its conclusion
will be

If k = 0, the conclusion of A' must be simply IF -+ P :a; choose a new variable

x not in A' and apply (-+I),,ac to deduce
t i-+ (.1x P) : p-+r

6B4.1 Note In the above definition clause (ii) makes as many variables as possible
distinct, and then (iii) identifies some of these again. Lemma 6B5 will show that A2
is well-defined for every A.

6B4.2 Examples (i) If A is the following logic deduction, then AA is the TA2-
deduction A3 in 6B3 (modulo replacements of distinct term-variables by distinct
 6B The Curry-Howard isomorphism 83

term-variables).
a-*a--*c a
(-E)
a-+c a
(-+E)
c

(ii) If A is the logic deduction in 6A2.2, AA is the TA2-deduction in 6B1.1 (modulo

=a)
(iii) If A is the logic deduction in 6A3.1, AA is the TA2-deduction in 6B1.2 (modulo
=a)
6B4.3 Exercise* Construct and compare (A5)2 and (A6)2, where A5 is the deduction
in 6A3.3 and A6 is the following deduction.

[a] (00)
(-'I)
a-+a (0) a
(-+I)
a vacouously}

6B5 Logic-to-Lambda Lemma Let al_., an H i by a logic deduction A, and for

each i < n let

be those occurrences of 61 that are undischarged assumptions at branch-tops in A. Then

AA is defined and is a TA2-deduction whose conclusion has form
x1 t:oli...,x1mi:v1, ... xn,l:o ...... :an '- M:2
where each x1j occurs exactly once in M (at the same position as q.J has in A) and
M has no bound-variable clashes. Also

(AA)L = A.

Proof Straightforward induction on the number of steps in A. The change of

variables in 6B4(ii) ensures that M has no clashes of bound variables and has only
one occurrence of each of its free variables.

6135.1 Note The Curry-Howard mapping from A to AL is usually called the Curry-
Howard isomorphism: how far can it justify this description? There are three levels
to consider:
(i) provable formulae H types of closed terms,
(ii) logic proofs -> TAx-proofs,
(iii) logic deductions +- TA1-deductions.
By 6B5, the AL-mapping has a one-sided inverse AA such that (A2)L A, but for it
to be an isomorphism in any real sense we should also have (AL)z A. We have
seen in 6B3 that this breaks down on level (iii). But on levels (i) and (ii) it holds
true, as the next theorem will show.
84 6 The correspondence with implication
But before that theorem a lemma will describe exactly how far the identity
(AL)z = A breaks down.

6B6 Lemma Let A be a TA2-deduction oft --> P :,r. Then (AL)2 differs from A only
in that some term-variables distinct in (OL)2 may be identical in A. In more detail:
(i) (AL)2 is a TA1-deduction of a formula with form

r' H M:T
where, just as in 6B5, M has no bound-variable clashes, FV(M) = {xl,...,xn}, and
each x; occurs just once in M; also, for some vl,...,v, not necessarily distinct,

P =a [vl/xl] ... [vn/xn]M,

A =a [VI/XI] ... [vn/xn](AL)l;

(ii) if A is a proof of H P:t then (AL)2 is a proof of i--+ P':i for some P' =_a P.

Proof Part (i) is proved by induction on IMO with three cases following the clauses
of 6B1. (The notation "[VIN] ... [vn/xn](AL)A" means the deduction obtained by
doing the substitutions [v1/x1],..., [vn/xn] in all subjects in (AL)2.)
Part (ii) is a special case of (i).

6B7 Curry-Howard Theorem (i) The provable formulae of Intuitionist implicational

logic are exactly the types of closed R-terms.
(ii) The relation 61, ... , Qn I- T holds in Intuitionist logic iff there exist M and
x1,...,xn (distinct) such that

x1:Q1,...,xn:Qn I-A M:T.

(iii) The A-to-AL mapping is a one-to-one correspondence between TAa-proofs and

proofs in the Natural Deduction version of Intuitionist Implicational logic, and the
A-to-A2 mapping is its inverse (modulo That is, for all TAx-proofs A,

(AL)x = A
(modulo -a in subjects in (OL)x), and for all logic proofs A,

(Da)L = 0.

Proof For (i) and (iii) use 6B2 and 6B6. For (ii) "if", use 6B2.
For (ii) "only if", let 61, ... , an I- t be obtained by a logic deduction A, and apply
rule (->I) n times to change A into a proof, A% of

?1--> ... -Qn--+T.

Then construct (A*)a. By 6B5 its conclusion will be

 6C Some weaker logics 85

for some closed A-term N with no bound-variable clashes. Hence by the subject-
construction theorem, N must have form Axl ... for some M and some distinct
xl,... , x,, and (1 * )2, must contain the formula
x1:61i...,xn:Un I-2 M:i.
(By the way, M has no bound-variable clashes because N has none.)

6137.1 Note The one-to-one correspondence in 5A7 between TA2-deductions and

the typed terms of Chapter 5 can be combined with the above mappings to give a
one-to-one correspondence between closed typed terms and logic proofs.

6B7.2 Corollary To decide whether an implicational formula T is provable in Intuition-

ist logic it is enough to decide whether there exists a closed term M such that -2 M:i,
or equivalently, whether there exists a closed typed term with type T.

6B7.3 Note An algorithm to decide whether there exists a closed typed term Mt will
be described in 8D5.2. Via the above corollary this will give a decision-algorithm
for Intuitionist implicational logic. Of course decision-procedures for this logic have
been known for a long time,' but the one in 8D5.2 will also count the number of
closed terms Mt in #-normal form and will generate them one by one. This can be
regarded as counting and generating all the irreducible Natural Deduction proofs
of T if we think of each Mt as representing a proof via the Curry-Howard mapping.

6B7.4 Corollary Peirce's Law, the formula introduced in 6A1.2, is not

provable in Intuitionist logic.

Proof By 8B6, Peirce's Law is not the type of a closed normal form; hence by the
SN theorem it cannot be the type of any closed term.

6B8 Note The Curry-Howard theorem has shown that the Intuitionist logic of pure
implication corresponds in a very neat and natural way to the pure A-calculus. Over
the years this correspondence has been extended to more expressive logics and even
to quite strong mathematical systems, and this has led to some useful techniques in
proof theory, for example methods for extracting programs from proofs of existence.
(See Crossley and Shepherdson 1993 for a modern account.)
But although each of these extensions is very important from a practical point of
view, few of them have the directness and neatness of the original "core" system.

6C Some weaker logics

At the end of Chapter 1 three restricted classes of A-terms were defined: the Al-terms
(roughly speaking, terms without vacuous binding), the BCKA-terms (terms in which
For example an old one is in Gentzen 1935 pp.405-407 (pp.83-105 of the English translation). A more
efficient one is in Dyckhoff 1992 p.800 (the note after the proof of Thm. 1); it has been implemented
in the University of St Andrews proof-discovery program MacLogic. Another one is in Hudelmaier
1993 along with a proof that it operates in O(n log n)-space. And one involving an explicit translation
into A-calculus is in Zaionc 1988.
86 6 The correspondence with implication
each subterm has at most one non-binding occurrence of each variable), and the
BCIA-terms (those that are both BCK2- and 2I-terms).
Suppose we restrict TA2 by looking only at terms in one of these three classes;
does the corresponding logic obtained via the Curry-Howard mapping have any
interesting properties? The answer turns out to be "yes" for each class, as we shall
see in this section.'

6C1 Definition (The relevance logic R.) The definition of R.. is exactly like that
of Intuitionist implicational logic in 6A2 except that vacuous discharging is not
allowed. That is, when a--*r is the conclusion of rule (-+I) its discharge-label must
show at least one occurrence of or.

6C1.1 Note (Motivation for R.) In one important view of implication, a formula
a--+T should not be provable unless a is in some way relevant to T. In this view the
formula

is not universally valid, because it says in essence that if a statement a is true then
every other statement b implies it, even when b has no connection with the meaning
of a. There have been many different attempts to capture the notion of relevant
implication by formal systems and R_ is one of the earliest and simplest. In it,
roughly speaking, we can only prove a--+-r when a has actually been used in the
deduction of r.
More about R_ can be found in Anderson and Belnap 1975, especially §§3, 7, 8.3,
8.4, 8.18 and 9-14, and in Anderson et al. 1992, especially §§47.1, 63.2 and 71.2. The
system has been invented independently at several different times; see Moh 1950,
Church 1951 and the discussion in Dosen 1992b. It is known to be decidable (i.e.
there is an algorithm to decide whether a given formula is provable in R.); see
Anderson and Belnap 1975 §13 and Anderson et al. 1992 §63.2

6C1.2 Examples Of the proofs in the examples in 6A, those in 6A2.1 and 6A2.2 are
correct R--proofs. In particular, by 6A2.2, the formula

is provable in R. In contrast the proof of the same formula in 6A3.1 contains a

vacuous discharge and hence is not an R.-proof. Nor is the proof of a--+a--+a in
6A3.3.
Of the proofs obtained in Exercise 6B2.1, those of (B), (B'), (I), (C), and (W) are
easily seen to be R--proofs, but that of (K) is not.

6C2 Definition (BCK-logic) The system that will be called here BCK--logic (or more
precisely, the implicational fragment of BCK-logic) is defined exactly like Intuitionist
The logics defined in the present section and some others like them have come to be known as
substructural logics.
2 Though in the literature R.. is often presented as the implicational fragment of a system R which has
other connectives besides "-+" and has been proved undecidable (Urquhart 1984 or Anderson et al.
1992 §65).
 6C Some weaker logics 87

logic in 6A2, except that multiple discharging is not allowed. That is, when (-I) is
used its discharge-label must either be vacuous or contain only one occurrence of v.

6C2.1 Note Roughly speaking, BCK-logic is a system in which an assumption

cannot be used more than once; it is a logic of non-re-usable information. Interest
in such logics dates back at least as far as Fitch 1936, and results on BCK
logic can be found in a number of sources, for example Meredith and Prior
1963 §§7-8, Jaskowski 1963, Is6ki and Tanaka 1978, Bunder 1986, 1991, Komori
1987, 1989, Blok and Pigozzi 1989, Hindley 1989, 1993, Dosen 1992a, Hirokawa
1991c, 1992b, 1993a-b, Komori and Hirokawa 1993, and those mentioned in their
bibliographies. The reason for the name "BCK" will be made clear in the next
section.

6C2.2 Examples Of the proofs constructed in Exercise 6B2.1, it is easy to see that
those of (B), (B'), (I), (K) and (C) are correct BCK-proofs.
But the proof of (W) in 6A2.1 is not, because it contains a multiple discharge.
Of the two proofs of in 6A2.2 and 6A3.1, the former is a
BCK-proof but the latter is not.
The proof of a->a-+a in 6A3.3 is a BCK-proof.

6C3 Definition (BCI-logic) The definition of BCI-logic (to be precise, its implica-
tional fragment) is like 6A2 except that both vacuous and multiple discharging are
forbidden; i.e., when rule (-+I) is used its discharge-label must contain exactly one
occurrence of a.

60.1 Note BCI-logic is a relevance logic of non-re-usable information. Most work

on this system has been done in parallel with work on BCK-logic and the majority
of the above references for BCK-logic treat both systems. Other references are, for
example, Hirokawa 1991b, Avron 1992.
Incidentally, the BCI system above coincides with the implicational fragment of
the linear logic of Girard 1987. Although the latter is a much deeper theory than
any logic confined to a single connective and has been studied in many publications,
a few of these publications give its implicational fragment some separate attention;
an example is Avron 1988.

6C3.2 Examples Of the proofs constructed in Exercise 6B2.1, those of (B), (B'), (I),
and (C) are correct BCI-proofs.
But the proofs of (K) and (W) are not.
Of the two proofs of (a-+a-.c)--+a-*a-+c in 6A2.2 and 6A3.1, the former is a
BCI-proof but the latter is not.
The proof of a--+a-+a in 6A3.3 is not a BCI-proof.

6C4 Lemma The sets of provable formulae of the four logics defined in 6B-C are
related as follows:
(i) BCI c BCK c Intuitionist,
(ii) BCI c R, c Intuitionist.
88 6 The correspondence with implication
6C5 Refined Curry-Howard Theorem (i) The provable formulae of the three logics
defined in 6C1-3 are exactly the types of the following A-terms:
R, : types of the closed 2I-terms;
BCK-logic : types of the closed BCKA-terms;
BCI-logic : types of the closed BCI2-terms.
(ii) The relation a 1, ... , Qn T holds in R_ BCK- or BCI-logic iff there exist M
and x1,..., x (distinct) such that
xI:a1,...,xn:an l- M:T

and M is, respectively, a 21-term, BCK2-term or BCIA-term.

Proof By Lemmas 6C5.1 and 6C5.2 below.

6C5.1 Lemma Let A be a TAx-deduction oft --> M:T and let AL be defined as in
6B1. Then
(i) f M is a 2l-term, AL is an Rte-deduction,
(ii) f M is a BCKA-term, AL is a BCK-deduction,
(iii) f M is a BCI2-term, AL is a BCI-deduction.

Proof Straightforward induction on IMI.

6C5.2 Lemma Let A be an R.-, BCK- or BCI-deduction with Ax defined as in 6B4,

and let the conclusion of AA be F --> M:T. Then
(i) if A is an R.-deduction, M is a RI-term,
(ii) if A is a BCK-deduction, M is a BCKA-term,
(iii) if A is a BCI-deduction, M is a BCI2-term.

Proof Straightforward induction on the number of steps in A.

6D Axiom-based versions

As a matter of historical fact the four logics discussed in this chapter all made their
first appearance not in the Natural Deduction form given above but in an axiom-
based version. These axiom-based versions will play a role in the next chapter, so
as a preparation they will be defined and discussed in this section.
But before defining the four logics separately some of the properties of axiom-
based systems in general will be described, relative to an arbitrary set A of axioms.
Axiomatic systems are often called Hilbert-style systems to contrast with Natural
Deduction.

6D1 Definition (A-logics) Let A be any set of implicational formulae that are
tautologies in the classical truth-table sense. Then A generates the following Hilbert-
style system, which will be called the corresponding A-logic.
Axioms: the members of A.
 6D Axiom-based versions 89

Deduction-rules :
(-.E)
T

(Sub) T
[if s is a substitution and no variable in
S(T)
Dom(s) occurs in a non-axiom assumption
in the deduction above the line]
Deductions in an A-logic are trees, with axioms and assumptions at the tops of
branches and the conclusion at the bottom of the tree. The notation
a1,...,a8 F-A T

means that there is a deduction whose non-axiom assumptions are some or all of
61, ... , 6n and whose conclusion is T. need not all be distinct.)
When n = 0 the deduction is called a proof of T and we call T a provable formula
or theorem of the A-logic in question, and we say

The set of all theorems in an A-logic may be called At

6D1.1 Note (i) Rule (-*E) is often called modus ponens or the detachment rule. In
it, U--+T is called its major premise and T its minor premise.
(ii) Rule (Sub) is the substitution rule. Its side-condition is standard in propositional
logic and says, roughly speaking, that substitutions may be made for variables
occurring only in axioms. It excludes such deductions as

a
[b/a].
b

6D1.2 Example Let A contain the formulae (C) and (K) shown in 6B2.1. Then the
following deduction gives I-A a-'a. (Both its substitution-steps use the substitution
[(a-->b-+a)/b, a/c].)
(C) (K)
(a->b-*c)->b-*a-*c
(Sub) (Sub)
(a-'(a->b-+a)-.a)-'(a->b->a)-*a-*a a-'(a->b->a)-'a
(-+E) (K)
(a-'b-'a)-'a-+a a-+b-+a
(->E)
a-*a

6D1.3 Warning It is natural to expect the theory of A-logics to be fairly simple,

because every such logic is a subsystem of classical propositional logic and the latter
is very well known. But this is not so. For example consider the provability problem
for a set A (the problem of deciding, for each type T, whether T is provable in A-logic
or not). For at least one A-logic it is not yet known whether its provability problem
is decidable although the logic has been studied for over 30 years (see 7D11(ii),
90 6 The correspondence with implication
system T.). Further, for every recursively enumerable degree of unsolvability,
however high, there is a finite set A whose provability problem has that degree
(Gladstone 1965 §6 Thm 1).

6D2 Definition In any A-logic a substitutions first deduction is a deduction in which

rule (Sub) is only applied to axioms.

6132.1 Lemma In any A-logic, every deduction A can be replaced by a substitutions-

first deduction A* with the same assumptions, axioms and conclusion.

Proof Suppose rule (Sub) is applied below an application of rule (-+E), as follows:

a-*r or
(->E)
T
(Sub)
§(i)
Then (Sub) can be moved up above (-->E), thus:
6--+i a
(Sub) - (Sub)
(-+E)
§(i)

Also two successive (Sub)'s can be combined into one. The moving-up procedure
ends when all (Sub)'s are at the tops of branches in the deduction-tree. By the
restriction on (Sub) in 6D1, the top formula of each of these branches cannot be a
non-axiom assumption, so it must be an axiom. D

6D3 Definition Hilbert-style Intuitionist logic (of implication only) is the A-logic
whose A has just the following four members:
(B)
(C)
(K) a--,.b--+a,
(W) (a->a->b)->a-+b.

6D4 Definition Hilbert-style R. is the A-logic for A = {(B),(C),(I),(W)}, where (I)

is

(I) a--+a.

6D5 Definition Hilbert-style BCK-logic (of implication only) is the A-logic for the
set A = {(B), (C), (K)}.

6D6 Definition Hilbert-style BCI-logic (of implication only) is the A-logic for the
set A = {(B), (C), (I)}.
 6D Axiom-based versions 91

6D6.1 Note (i) By Example 6D1.2, (I) is provable in Hilbert-style BCK-logic and
Intuitionist logic.
(ii) Of course (B), (C), (I), (K) and (W) are the principal types of the 2-terms B,
C, I, K and W, see Table 3E2a. But each one also expresses a property of implication
that has its own interest quite independently of type-theory. Roughly speaking, (I) is
the reflexivity property of implication, (C) states that hypotheses can be commuted,
(K) that redundant hypotheses can be added, (W) that duplicates can be removed.
Finally (B) can be viewed either as a transitivity property of implication or as a
"right-handed" replacement property which says that if a--+b holds, then a may be
replaced by b in the formula c--+a.
(iii) The next theorem will describe the precise connection between A-logics and
A-calculi; it will be a Hilbert-style analogue of the Curry-Howard theorem at the
end of the previous section. As mentioned earlier, the type-system that corresponds
most naturally to a Hilbert-style logic is combinatory logic not A-calculus, but we
can build a correspondence with A-calculus if we are careful; the key will be the
concept of applicative combination defined in 9F1.

6D7 Theorem (Curry-Howard for Hilbert systems) Let {Cl, C2,. ..} be a finite or
infinite set of typable closed A-terms and let A = {y1,y2.... } where y, PT(C,). Then
(i) the theorems of A-logic are exactly the types of the typable applicative combi-
nations of C1, C2, ...,
(ii) the relation el,...,an I-A T holds ii f there exist an applicative combination M of
C1, C2.... and some distinct term-variables xl, ... , xn, such that
xl:Ql,...,xn:Q I-A M:T.

Proof Part (i) is the case n = 0 of (ii). To prove (ii) we first prove "if". Let M be
an applicative combination of xl,... , xn, Cl, C2, ..., and let A be a TA2-deduction of
(1) xl:oi,...,xn:Qn i-- M:T.
Corresponding to each occurrence of a C; in M there will be an occurrence of
i-, C,:s(y,) in A for some substitution s. Remove from A all steps above these
occurrences of C1, C2, ..., and replace each formula H C; :s(yi) by the type s(yi).
Then replace every other formula in A, say IF H N:p, by the type p. (Cf. 6B1(i)-(ii).)
The result is a Hilbert-style deduction giving
(2) Q1,...,an I-A T.

To prove "only if" in (ii), let 01.... , an I- T in A-logic. Then by 6D2.1 there
is a deduction A of T in which (Sub) is only applied to axioms. Change A to a
TA2-deduction thus. First choose some distinct term-variables xl,...,xn and replace
each undischarged branch-top occurrence of each Q, in A by
xi:a, h-) x,:Q,.

Next, note that each application of (Sub) in A will be applied to an axiom to give,
say, yj I- s(yj); replace it by a TAA-proof of H Cj:s(yj). Then replace the logic rule
by the TA2-rule (-+E) throughout. The result is a TA2-deduction of (1) for
some term M as required.
92 6 The correspondence with implication
6137.1 Corollary The relation al, ... , Qn F- i holds in the Hilbert version of Intuitionist
logic, R., BCK- or BCI-logic if
x1:al,...,x,,:a F-A M:T

holds for some distinct X1,...,xn and some M which is an applicative combination of
x1i ... , xn and, respectively, {B, C, K, W}, {B, C, I, W}, {B, C, K} or {B, C,1}.

6D8 Theorem (Hilbert-Gentzen link) For the Intuitionist, R.-, BCK- and BCI-
logics the relation
at,...,an F- 2
holds in the Natural Deduction version tiff it holds in the Hilbert version.

Proof By the Curry-Howard theorems for Natural Deduction (6B7(ii), 6C5(ii)), the
combinatory completeness theorems (9F3, 9F5), and Corollary 6D7.1.

6D8.1 Note The Hilbert-Gentzen link is usually proved directly without going
through A-calculus. The key step in a direct proof is a result usually called the
deduction theorem which is a close analogue of the combinatory completeness the-
orems in 9F. (The deduction theorem is treated in many standard introductions
to classical logic, see for example Hamilton 1988 Proposition 2.8; a study of the
theorem in some different logics is in Bunder 1982.)
 The converse principal-type algorithm

This chapter combines the theme of propositional logic from Chapter 6 with that of
principal types from Chapter 3. We saw in the Curry-Howard theorem (6B7) that
the types of the closed terms are exactly the theorems of the intuitionist logic of
implication: hence the principal types of these terms must form a subset of these
theorems, and the very natural question arises of just how large this subset is. Do its
members form an aristocracy distinguished in some structural way from the general
rabble of theorems, or can every theorem be a principal type?
The main result of the chapter will show that there is in fact no aristocracy: if a
type T is assignable to a closed term M but is not the principal type of M, then it is
the principal type of another closed term M*.
The proof will include an algorithm to construct M* when T and M are given.
To build M* an occurrence of M will be combined with some extra terms chosen
from a certain carefully defined stock of "building blocks", closed terms with known
principal types; and the main aim of this chapter will be to build M* from as
restricted a set of building blocks as possible.
The algorithm in the earliest known proof needed full A-calculus (Hindley 1969),
but two later ones used only Al-terms as building blocks (Mints and Tammet 1991,
Hirokawa 1992a §3) and another used an even more restricted class (Meyer and
Bunder 1988 §9). The algorithm below will be based on the latter very economical
one.
When it was first proved the existence of M* seemed nothing more than a mere
technical curiosity, but several years later it was discovered by David Meredith to
be equivalent to the completeness of a neat form of resolution rule for Intuitionist
implicational logic. This rule and its completeness problem will be described at the
end of the chapter.

7A The converse PT theorems

In fact the situation is not quite as simple as suggested above, as the converse PT
theorems for full A-calculus and AI-calculus are very slightly different. They will be
stated separately here.
First, recall that the class of unrestricted A-terms is sometimes called the AK-calculus
(because it contains K - Axxy x) and its terms AK-terms to distinguish them from
restricted classes of terms.

93
94 7 The converse principal-type algorithm
7A1 Definition A AK-PT [respectively, AI-PT, BCKA-PT, BCIA-PT] is the principal
type of a closed AK-term [Al-term, BCKA-term, BCIA-term].
A BCKW-PT[BCIW-P T, etc] is the principal type of a BCKW-combination
[BCIW-combination, etc] as defined in 9F1.

7A1.1 Lemma A type is a AK-PT, AI-PT, BCKA-PT or BCIA-PT iff it is, respec-
tively, a BCKW-, BCIW-, BCK- or BCI-PT.

Proof By 9F3 and 9175.

7A2 Converse PT Theorem for AK (i) Every type of a closed AK-term is also a
AK-PT.
(ii) Further, there is an algorithm which accepts any typable closed AK-term M and
any type T of M, and constructs a closed AK-term M* whose PT is T.
(iii) Furthermore, M" >p M.

Proof There is proof in Hindley 1969 Thm. 3 of the corresponding result for
combinatory logic. This translates straightforwardly into A-calculus. It contains an
algorithm that constructs, for every type a, a closed AK-term IQ such that
PT(IQ) = Q-+o, IQ c'f I.

The last step is to choose M* - ITM. Then an application of the PT algorithm

(3E1) shows that PT(M*) - T.
(An alternative proof of (i)-(ii) would be to apply Theorem 7A3 below to construct
IQ, but this would not give (iii) but only M* t>11n M.)

Before going on to the Al-theorem we pause to look at a particularly neat and

interesting application of the AK-theorem. The following corollary shows even more
clearly than Examples 2C2.2-6 how widely a term's set of types can vary when the
term is /3- or rl-converted.

7A2.1 Corollary There exist two typable closed AK-terms P and P' such that P =0 P'
but P and P' have no types in common at all.

Proof The following two types have no substitution-instances in common (cf.

3C2.2):
a - a-+a, T-
Hence also the following have no instances in common:
a--+Q, T--+T.

But by 7A2 there exist IQ and It such that IQ =p I =p IT and

PT(IQ) = or--+a, PT(IT) = T--+T.
Define P -1Q, P' -1t.
 7B Identifications 95

7A3 Converse PT Theorem For Al (i) Every type of a closed AI-term is also a 21-PT.
(ii) Further, there is an algorithm which accepts any typable closed Al-term M and
any type r of M, and builds a closed AI-term M* whose PT is T.
(iii) Furthermore, M* Pan M.

7A3.1 Proof-note An algorithm and correctness proof will be given in 7C. Note
that (iii) is weaker than the corresponding result for AK in 7A2(iii): in AK we have
t>p but in Al the proof will only give DRn
The first known algorithm for Al was devised in 1986 by Robert Meyer (Meyer
and Bunder 1988 §9), and another was invented independently by Grigori Mints
and Tanel Tammet (Mints and Tammet 1991 §2). Both of these originated in the
context of Hilbert-style axiom-based logics and although they can easily be trans-
lated into A-calculus they lose some of their directness when this is done. In 1990
a direct and natural A-algorithm was described by Sachio Hirokawa (Hirokawa
1992a §3).
The most economical of these algorithms is the Meyer one, in the sense that it
does not need all the Al-terms as building blocks but only a certain well-defined
proper subset of them. So this algorithm is the one that will be described in
7C.
All three algorithms produce terms M* satisfying (iii) as well as (ii). Though (iii)
was first proved for the Hirokawa algorithm (Hirokawa 1992a §3), a proof will be
given below for the Meyer algorithm and a careful reading of Mints and Tammet
1991 shows that (iii) holds for that algorithm too.

7A3.1 Question Can 7A3(iii) be strengthened to say "M* >0 M"?

7A3.2 Exercise* If the answer to the above question is "yes" then as a consequence
there exist two /3-equal typable Al-terms with no types in common (cf. Corol-
lary 7A2.1 for AK). Show that this consequence holds even if the answer to the
above question is "no", by constructing two suitable Al-terms directly. (Hint (Martin
Bunder): consider the term P in 2A8.8.)

7B Identifications

The construction of M* from M and T in the proof of 7A3 will depend on viewing T
as the result of applying certain substitutions to a particularly simple form of type.
The present section introduces the notation needed for this.
First recall the substitution notation introduced in 3B. In particular (from 3B7) a
variables-for-variables substitution is one with form

s = [blbal,...,bn/an]
where bl,... , b are any variables; and a renaming in T (where T is a given type) is
a variables-for-variables substitution such that Vars(T) and bl,...,b
are distinct.
96 7 The converse principal-type algorithm
7B1 Definition For any type T, an identification in T (sometimes called a contraction
in T) is any substitution [b/a] such that a, b both occur in T. A type a is obtained
from ,r by identifications if

v = [b./a/,](... ([bt/ai](T))...)
where [bl/al] is an identification in T, [b2/a2] is an identification in [bl/at]T, etc.

7B1.1 Example If T - (a-+b-+c)-+(a-+b)-+a-+c then [b/a] is an identification in T,

and
[b/a]T - (b-4b--+c)--+(b-+b)--+b--+c.

7B1.2 Lemma Every variables-for-variables substitution can be performed by a renam-

ing followed by a series of identifications. More precisely, if p - §(T) and s substitutes
variables. for variables, then

p = [bn/an](...([bl/al](v(T)))...)
where n > 0, v is a renaming in T, and each [b;/a;] is an identification in

[b1_i /a,_i](... (d(T)) ...).

7B2 Definition A type T is skeletal if each variable in T occurs exactly once. (The
property of being skeletal is sometimes called the 1 -property.)

7B3 Lemma Every type r can be obtained by identifications from a skeletal type T°
(which will be called a skeleton of T).

Proof If T is skeletal choose T° - T. If not, then for each variable b with two or more
occurrences in r, replace all but one of these occurrences by distinct new variablesE

7C The converse PT proof

This section contains a proof of Theorem 7A3 obtained by translating the proof
in Meyer and Bunder 1988 §9 into A-calculus. The strategy will be to show first
that every type of a closed 2I-term can be obtained from a AI-PT by a series of
identifications, and then that identifications preserve the property of being a 2I-PT.
But a preliminary step will be to prove the special case of the theorem in which T
has form 9-.0 where 0 is skeletal.

7C1 Lemma Let 0 be any skeletal type. Then is the PT of a closed RI-term IB
such that

Proof We shall construct Io by induction on 101. If 0 is an atom, say 0 - a, choose

(1) la - I.
 7C The converse PT proof 97

Next, suppose 0 _ p--+o- and IP and Ia have already been built such that

(2) PT(IP) = P-'P, PT(I,,)

(3) IP D#n I, IQ D6n I.

Choose
(4) IP_Q = Axy-I"(x(IPy))
From (2) and the PT algorithm (3E1) it is easy to deduce that
(5) PT(IP. ) = (P-'v)-'(p-'6).
(Note that the proof of (5) depends on the fact that Vars(p) n Vars(a) = 0, which
holds because p-*a is skeletal.) Also, by (3),

IP-a DNn '1xy'I(x(ly))

i.p Axy-xy
C>n I. 0
7C2 Lemma Every type of a closed AI-term can be obtained from a AI-PT by identi-
fications. In detail: if M is a closed AI-term and I-2 M:T, there exists a closed typable
.1I-term M+ such that
(i) T sk(sk_t(...si(PT(M+))...)) for some identifications §1,...,§k,
(ii) M+ Dan M,
(iii) if T is skeletal then PT(M+) _ T.

Proof First construct a skeleton T° of r as in 7B3. (If T itself is skeletal, T° =_ 'r.) By

7C1 there is a closed 2I-term ITo with
(1) PT(I,.) _ ITo Din I.

Let F-x M:T. Then M has a PT, call it a. And by renaming variables in a we can
ensure that
(2) Vars(a) n Vars(T°) = 0.

Define
(3) M+ _ I,. M.
Then (ii) holds by (1). To prove (i), apply the PT algorithm to M+. The first step
is to find an m.g.u. of (T°, a). To see that one exists, note that T is an instance of
T° by the definition of T°, and T is an instance of a because a is principal, so (T°, a)
has a common instance, namely r; and by (2) this instance is a unification of (T°, a) ;
hence (T°, a) has an m.g.u. T+ and r is an instance of T+, say T _ s(T+). Then
(4) PT(M+) = T+.
To complete the proof of (i) the next step is to show that s is a variables-for-
variables substitution. For this it is enough to show that
(5) ITI < IT+I,
98 7 The converse principal-type algorithm
because if s substituted a composite type for a variable in T+ it would increase IT+I.
But ITI = IT°I because T is obtained from t° by identifications; and T+ is an instance
of T° since T+ unifies (T°, a), so IT+I > IT°I; hence (5) holds.
Thus r is obtained from a 2I-PT T+ by a variables-for-variables substitution s.
By 7B1.2, s can be split into a renaming (which changes T+ to another PT of M+)
followed by some identifications. Hence (i) holds.
To prove (iii) note that if T is skeletal then T° - T, so M+ = ITM and the PT of
the latter is clearly r.

7C2.1 Note By the above lemma, the proof of Theorem 7A3 will be complete if we
can prove that identifications preserve the property of being a 2I-PT. This will be
done in the next three lemmas.
In them the following notation will be useful: if a1,...,a are distinct and each
has exactly one occurrence in a given type T, we may say
T(al,...,an), T(o ,...,a,)

for r and respectively.

7C3 Single-Replacement Lemma If T is skeletal and a E Vars(T) and b 1 Vars(T),

there exists a closed 2I-term Q whose PT is
(i) (a-'b)*--'T(a)-'T(b),
where (a->b)* - a-+b if the sole occurrence of a in r is positive in the sense of
Definition 9D3, but (a->b)* __ b--+a if it is negative. Further,
(ii) QI >p,11.

Proof Q will be constructed by induction on In.

Basis: T - a. Then (a->b)* - a-*b. Choose Q - then QI r>p I, and by the
PT algorithm (3E1),
PT(Q) - (a->b)->(a->b).
Induction step: T p-+a. Then p and a are skeletal, so by 7C1 there exist 11-terms
IP and I, with
(1) PT(IP) = P-'P, PT(Ia) = o-'Q,
(2) IP>Q71, IQ>p1I.

Case 1: the sole occurrence of a in T is positive and is in p. Then (a-+b)* __ a-*b.

But the positive occurrence of a in T corresponds to a negative occurrence in p, so
by the induction hypothesis there is a 2I-term QP- such that
(3) PT(QP-) = (b-a)-P(a)-&),
and Qp-I >R, I. By renaming the variables in (3) we get
PT(QP-) =
Choose
(4) Q = 2xyz'Io(Y(QP-xz))
 7C The converse PT proof 99

Since (a-+b)* __ a--+b, we must prove that

(5) PT(Q) = (a-*b)-+(p(a)-a)->(p(b)-*a).
But this comes easily from applying the PT algorithm to Q. Finally it is simple to
check that QI D6n I, using (2).

Case 2: the occurrence of a in T is positive and is in a. Then a occurs positively in

a, so by the induction hypothesis there exists Q,,+ such that
(6) PT(Q,+) =
and QQ+I t>#,, I. Choose

(7) Q = 2xYz'Q,+x(Y(Ipz))
Then QI D6n I easily. Now (a-+b)* - a--+b since a occurs positively in T; and by the
PT algorithm and perhaps some renaming we get
(8) PT(Q) _
Case 3: the occurrence of a in T is negative and is in p. Then a occurs positively
in p, so by the induction hypothesis there exists Qp+ such that
(9) PT(Qp+) ° (a-'b)-'p(a)-'p(b),
and Qp+I Dpn I. Choose
(10) Q = 2xYz.IQ(Y(Qp+xz)).
Then QI upn I. Now (a-+b)* - b-+a since a occurs negatively in T; and by the PT
algorithm and perhaps some renaming, we get
(11) PT(Q) =
Case 4: the occurrence of a in T is negative and is in a. Then a occurs negatively
in a, so the induction hypothesis gives Q,_ such that
(12) PT(QQ_) - (b->a)- a(a)- a(b)
and Qa_I D#n I. Choose
(13) Q

Then QI DBa I. Now b--+a since a occurs negatively in T; and by the PT

algorithm and perhaps some renaming we get
(14) PT(Q) = (b-- a)-'(p-'a(a))-+(p-'a(b)).

7C4 Double-Replacement Lemma If T is skeletal and at, a2 are distinct variables in

T, and bl, b2 are distinct variables not in T, then there is a closed 21-term R whose PT
is

(i) (at-+bt)*-+(a2-+b2)*-+T(at,a2)-+T(bt,b2),
where, for each i < 2, (a;-+b;)* - a1-+b, if the occurrence of ai in T is positive, but
(ai-+bi)* - b,-*a, if it is negative. Also
(ii) RII >pn 1.
100 7 The converse principal-type algorithm
Proof By 7C3 there exist AI-terms P and Q such that
PT(P) = (ai->bi)*-+T(ai,a2)-'T(bi,a2),
PT(Q) = (a2-b2)*-*T(bi,a2)_ r(bi,b2),
and PI D#n I and QI D6q I. Choose
(1) R - lxyz-Qy(Pxz).
7C5 Identification Lemma If T is a AI-PT then so are all types obtained from T
by identifications. Further, there is an algorithm which, when given a Al-term M and
T - P T (M) and a - [b/a]T with a, b E Vars(T), will construct a Al-term N such that
PT(N) - a, N>#n M.

Proof Let T = PT(M) for some closed AI-term M and let a =_ [b/a] T where a,
b E Vars(T). We must prove that a is a AI-PT. Let
Vars(T) = {a,b,cl,...,ck}
where a, b, cl,... , ck are distinct, and let p, q, n1,. .. , nk respectively be the number
of occurrences of each of these variables in T.
Define T° to be the result of replacing each occurrence of each variable in T by a
distinct new variable; say

where al,...,ap replace the occurrences of a, etc. Then replace al and bl in T° by

two new variables f and g distinct from all those above, and call the result T':
T ti (f,a2i...,ap,g,62,...,bq,C1,1,...,,C1,n ... , Ck,1,...,Ck,nt)
Apply the double-replacement lemma (7C4) to al, bl, f, g and T this gives a closed
AI-term R such that
(1) PT(R) = (al-*f)*-'(b1-*g)*->(T°-'T')
where (a1__ f)*, (bl-+g)* are defined as in 7C4 depending on whether al, b1 occur
positively or negatively in T°. Define
(2) N-
We shall see that PT(N) - [b/a]T. First, the PT algorithm gives
(al-+f)*-*[al/bi,f/g](T -+r ).
Hence
[al/f][al/bl,f/g](r°-),T)
(3) [ai/bl]T°-- [ai/f,ai/g]T'
To finish computing PT(N) we need an m.g.c.i. of ([al/bi]T°,T). Now [al/bl]T° is
T al,a2,...,a.,a,,v2,...,bq,c1,1,...,Chn ... , Ck,1,...,Ck,nA)

and T is
T a,a,...,a,b,b,...,b,C],...,C], . .. , Ck,...,Ck .
 7C The converse PT proof 101

Consider any common instance p of ([al/bl]T°,T): since [al/bl]T° has al in the

position that was occupied by bl in T°, the types in p at the al- and bi-positions
must be the same. Also, since p is an instance of T, the types in p at the positions
of al, ... , ap must be the same, and those at the positions of b1,.. . , bq must be the
same. Hence an m.g.c.i. of ([al/bl]T°,T) is
T (b,b,...,b,b,b,...,b,cl,...,c1, ... , Ck,...,Ck).
The substitution that generates this from [al/bl]T° is
[b/ai, b/a2, ..., b/ap, b/b2,... , b/bq, Cl /Cl,l, .. J.
Hence by the PT algorithm and (3), P T(N) is
[b/ai, b/a2, ... , b/a,, b/b2, ..., b/b9, cl /cl,i,...]([ai If, al /g]i ),
which is
T°(
b,b,...,b,b,b,...,b,cl,...,cl, ... , Ck,...,Ck >

and this is [b/a]T as required.

Finally, N Dpn M by (2) and 7C4(ii).

The following lemma and algorithm complete the proof of Theorem 7A3.

7C6 Lemma Algorithm 7C7 below accepts any T and any M such that F-A M:-r, and
outputs a closed term M* such that PT(M*) - T and M* t>s,i M. Further, if M is a
t1-term so is M*.

Proof See 7C7 and 7C1-5.

7C7 Converse PT Algorithm (Meyer and Bunder 1988 §9.) Input: any pair (T, M)
such that I-A M:T. (The relation I-x M:T is decidable by 3E3.)
Step 1. Follow the proof of 7C2 to construct M+ and T+ - PT(M+) and to find a
series o f identifications S1, ... , Sk (k >_ 0) such that

T - Sk(... (Sl (T+)) ...).

Step 2. If k = 0 in Step 1, choose M* - M+. If k 1, apply 7C5 first to T+, then
to sl(T+), then to S2(sl(T+)), etc., and choose M* to be the last term N produced.

7C7.1 Note By the proofs of 7C1-5 we can see that the M* produced by the above
algorithm has form
M* = Tk(Tk-l(... (T1(I,,-M))...))
where T1..... Tk and ITo are typable closed AI-terms that flry-reduce to I.

7C7.2 Exercise* Apply the above algorithm to construct a term M* whose PT is T

in the following cases:
(i) T M I;

(ii) T b-+b--*b, M K.
102 7 The converse principal-type algorithm
(By the way, in (ii) M is not a 2I-term; but the algorithm works despite this, though
in such a case it produces an M* that is also not a 2I-term.)

7C7.3 Note As mentioned earlier the above converse PT proof is not quite the
one in Meyer and Bunder 1988 §9 but a 2-adaptation of it. The unadapted Meyer-
Bunder proof is slightly stronger, and shows that the converse PT theorem holds
not only for the class of all AI-terms but for a more restricted class too, namely the
BB'IW-combinations as defined in 9171.1 For completeness' sake this result will now
be stated as a formal theorem.

7C8 Theorem (Meyer and Bunder 1988 §9.) (i) Every type of a BB'IW-combination
is also a BB'IW-P T.
(ii) In detail: there is an algorithm which accepts any typable closed term M and any
type T of M, and outputs an applicative combination M* of M and the closed terms
B, B', I, W, such that
PT(M*) __ T, M* t'p,1 M.

Proof Using the PT algorithm and some patience, it is easy to show that the proofs
of Lemmas 7C1-5 stay valid if the definitions of Ip-Q, Q, R and N in them are
replaced by the following:
in 7C1 (4): I,, B(B'B)(BBB')IpI,,,
in 7C3 (4): Q = BB(B(B'B')(BBB))IQQP_,
in 7C3 (7): Q BB(B(B'B)(BBB'))IpQQ+,
in 7C3 (10): Q BB(B(B'B')(BBB))IQQp+,
in 7C3 (13): Q BB(B(B'B)(BBB'))IpQQ_,
in 7C4 (1): R B(B'B')(BB')PQ,
in 7C5 (2): N = WRIM.

7D Condensed detachment

We now come to the connection with the resolution-style rule for propositional logic
mentioned in the introduction to this Chapter. This rule was originally devised in
the early 1950's under the name of the condensed detachment rule. It will be de-
scribed here in the context of the A-logic generated by an arbitrary set A of axioms,
see 6D1.

7D1 Definition Rule (D), the condensed detachment rule, makes the deduction
p-+o T
(D)
D(p-.o)T
where D(p--*r)T is an implicational formula defined thus: if p and T have a common
This class is a proper subclass of the closed Al-terms because it does not contain C - (When
C xyz is #-reduced the occurrence of z moves away from the rightmost position, and it is easy to see
that no BOW-combination has this property.)
 7D Condensed detachment 103

instance, construct an m.g.c.i. V - §1(p) §2(T) of (p, T) such that

(i) Vars(v) fl (Vars(a) - Vars(p)) = 0,
(ii) Dom(sl) = Vars(p), Dom(s2) = Vars(T),
and define
(iii) D(p-*a)T - si(a).
(If p and r have no common instance D(p-+a)T is undefined and rule (D) is not
applicable.)

7D1.1 Lemma If D(p-*Q)T is defined, it is the most general conclusion that can be
obtained by (-+E) from instances of p--+o and r using an instance of p-+a as major
premise. More precisely, if v §1(p) s2(T) is an m.g.c.i. of (p,T) satisfying 7D1(i)-
(ii) and r1 and r2 are any substitutions such that ri(p) - r2(T), then r1(o) is an instance
of si(Q).

Proof The definition of D(p-+a)T is almost identical to the procedure described

in 3C4 for computing PT(PQ) when PT(P) - p--+a and PT(Q) - T; hence the
argument in 3C4 proves the present lemma. (By the way, condition 7D1(i) corre-
sponds to (1) in 3C4.)

7D1.2 Notes (i) D(p-+a)T is unique modulo renaming of variables (by 3C3.2), so in
future we may think of it as uniquely defined.
(ii) D(p-+a)T is also independent of renaming in p--+a and T; that is, if D(p-+6)T
is defined and p'-*c' and T' are alphabetic variants of p--+u and r respectively then
D(p'->a')T' is defined and is an alphabetic variant of D(p-+v)T. (By 3C3.2).
(iii) In any system whose rules include (-+E) and (Sub) we can derive rule (D) as
follows.
(-auE) P" T
(Sub) §2(T)
§1(Q)-+§i(6)
(§2(P) = §2(T))
si(a)
The following lemma expresses formally the fact that the definition of D(p-+o)T
is the same as the procedure in 3C4 for computing PT(PQ) when PT(P) - p--+Q
and PT(Q) - T.

7D2 Lemma (D. Meredith.) If P and Q are typable closed terms with P T(P) = p-+o
and P T (Q) =-T, then
(i) PQ is typable iff D(p-+a)T is defined,
(ii) PT(PQ) - D(p-+Q)T modulo permutation of variables.

7D3 Historical Note Although rule (D) is so close to the PT algorithm each of the
two concepts was developed in total ignorance of work on the other until around
1978, when David Meredith was the first to realise the parallel. (Hindley and
Meredith 1990.)
104 7 The converse principal-type algorithm
Rule (D) was invented by Carew Meredith, and David Meredith (his cousin)
recalled learning it from him in 1954. Its first appearance in print was in Lemmon et
al. 1957 §9. A definition of the rule with historical comments appears in D. Meredith
1977 and further historical comments are in Kalman 1983.
Precursors of the rule date back perhaps as far as the 1920's, however. In Poland, a
particular D-computation was published in Lukasiewicz 1939 (p. 276 of the English
translation); and there is indirect evidence' that Alfred Tarski, then a colleague of
Lukasiewicz, might have formulated the D-concept (and therefore perhaps also an
algorithm for computing D(p- a)T) even before 1930.
Carew Meredith's work on propositional logics involved computing D(p- a)T in
many special cases, for example those in Lemmon et al. 1957, but he does not seem
to have written out his method as a formal algorithm, at least not in print. However,
in 1957 an algorithm for constructing D(p-*a)T was formalized by David Meredith
and implemented by him as a program for the computer UNIVAC 1.2
Now by 7D2 the D-construction is the same as the core of the PT algorithm;
hence David Meredith's 1957 D-program was probably in essence the first formal
PT algorithm. It was also the first one to be run on a machine. (Cf. Comment 3A7
on the history of PT algorithms.)

7D4 Definition (Condensed logics) For any set A of formulae, condensed A-logic is
obtained by replacing the two rules (-*E) and (Sub) in the definition of "A-logic"
in 6D1 by the single rule (D). We shall call its deductions D-deductions; they are
trees constructed from axioms and assumptions by rule (D). For deducibility, etc.
in a condensed logic we shall say D-deducible, D -provable, D -proof, D-theorem, and
shall use the notation
U1,...,u, E- T, [-AD T.

The set of all theorems of condensed A-logic may be called ADF.

7D4.1 Note By 7D1.2(iii), ADI- c Al-.

7D5 Definition Rule (D) is said to be complete for a given set A of axioms (or Al is
said to be D-complete) if
ADF-
= Ak.

7135.1 Note The concepts of condensed logic and D-completeness have not been
defined for Natural Deduction systems but only for Hilbert-style (axiom-based)
logics.

7D5.2 Question It is natural to ask which sets A are D-complete: in particular,

which of the axiom-sets for Intuitionist, R.-, BCK- and BCI-logics in 6D3-6 are
D-complete? This is the question that the converse PT theorem will help to answer,
and the key to the answer is the following preliminary theorem.
1 Mentioned in Kalman 1983 p. 447.
2 Noted in Kalman 1983 p. 449 and in correspondence from David Meredith to the present author.
 7D Condensed detachment 105

7D6 Meredith's Curry-Howard Theorem (D. Meredith.) Let {C1, C2,...} be a finite
or infinite set of typable closed A-terms, let PT(C;) - yi, and let Al = {y1,y2.... }.
Then the theorems of condensed A-logic are exactly the PTs of the typable applicative
combinations of C1, C2. ....

Proof By 7D2.

7D7 Theorem The axiom-set {(B), (C), (I), (W)} for R_, given in 6D4 is D-complete.

Proof Let T be provable in R.. Then by 6D7.1, r is the type of an applicative

combination M of B, C, I and W. Hence by 7C8, r - PT(M*) for some applicative
combination M* of M, B, B', I and W. Let M** be the result of replacing B' in M*
by the following combination of C, B and I:
B(B(CIB)B)(CI).

It is straightforward to check that this combination has the same PT as B' (in fact
it also reduces to B'); so PT(M**) -PT(M*) - T. Hence r is a BCIW-PT. But by
7D6 every BCIW-PT is a D-theorem of the logic whose axioms are the PT's of B, C,
I, W, and this logic is R..

The next step in answering Question 7D5.2 is to prove the D-completeness of the
axiom-set {(B), (C), (K), (W)} for Intuitionist logic. This logic is stronger than Rte,
and if we view the D-completeness of a set of axioms as saying that deductions
obtained by rule (Sub) can be imitated using rule (D) in combination with some
of the axioms, it is natural to conjecture that if we strengthen a D-complete set it
will remain D-complete. The following definition and lemma make this conjecture
precise.

7D8 Definition For any sets Al and B of formulae: B-logic is called an extension of
A-logic if B' =2 A'; it is called a D-extension of Al-logic if
AD-

7D8.1 Note If B 2 A then B-logic is both an extension and a D-extension of

A-logic.

7D9 D-Extension Lemma If A is D-complete and a-*a is an A-theorem, then every

set B whose logic is a D-extension of A-logic is D-complete.

Proof Let BDF 2 ADI-;

to prove B is D-complete we must show that if A is a
deduction from B by rules (-*E) and (Sub), each application of (-*E) or (Sub) in A
can be replaced by one of (D).
is already a special case of (D).
Case (Sub). It is enough to show that if s is any substitution then
DF _ §(2) E BDI-.
106 7 The converse principal-type algorithm
First note that a-+a E A' by assumption, so by (Sub),
§(z) - s(a) E V-.
But we have assumed AF = A°- and ADS- BD-, so

s(2)->s(r) E BD-.

Now it is easy to see from Definition 7D1 that

D(s(i)->s(r), r) = S(T).
Hence by rule (D) applied to the D-theorems S(T)--+S(T) and T we get

s(2) E I°F.

7D10 Theorem The axiom-set {(B), (C), (K), (W)} for Intuitionist logic given in 6D3
is D-complete.

Proof By 7D7 and 7D9, since this logic is a D-extension of R..

7D11 Note (Other D-complete logics) (i) Classical logic. The implicational fragment
of classical logic is a Hilbert-style system defined by the axioms (B), (C), (K), (W)
and
(PL)
((PL) is called Peirce's law, see .6A1.2 and 6B7.4.) It can be shown that a formula
is provable in this system if it is a tautology in the usual truth-table sense. (Prior
1955 Part I, Ch. III.) By 7D9-10 this system is D-complete.
(ii) Ticket entailment. A Hilbert-style logic defined by the axioms (B), (B'), (I) and
(W) listed in 6B2.1 was introduced in Anderson and Belnap 1975 (see especially
Ch. 1 §§6, 8.3.2), where it was called T_, or the logic of ticket entailment. The details
of T_, and its motivation are not the concern of this book, but in a sense (B) and
(B') are "right-handed" and "left-handed" replacement properties: if a-+b holds, (B)
says that a can be replaced by b in the formula c-.a and (B') says that b can be
replaced by a in a-*c. (Meanings for (I) and (W) were suggested in 6D6.1(ii).) This
logic can be shown to be weaker than R_, but by 7D6 and 7C8 it is nevertheless
D-complete. 1
(iii) A set of axioms whose logic is strictly weaker than T. was proved D-complete
by N. Megill in unpublished notes in 1993, and an infinite series of ever weakening
D-complete axiom-sets has since been constructed (Megill and Bunder 1996).

But not all axiom-sets are D-complete, as the next theorem will show.

7D12 Theorem The axiom-sets {(B), (C), (K)} and {(B), (C), (I)} given in 6D5-6 are
D-incomplete.

1 By the way, it is not yet known whether there is a decision-procedure for provability in T...,.
 7D Condensed detachment 107

Proof By 7D9 it is enough to prove the result for {(B), (C), (K)}. And by 7D6 it
is enough to find a type T that can be assigned to an applicative combination of
B, C, K but is not the PT of such a combination. One such type (pointed out by
A. Wronski, cf. Bunder 1986) is
T = ((a-*a)--*a)-,a.
It is easy to deduce from the types of C and K in Table 3E2a that
I-1 (C(CKK)(CKK)):T.

On the other hand, T cannot be the PT of an applicative combination M of B, C, K.

Because if it were, M would be a BCKA-term and hence MI would be one too by
the definition of BCKA-terms in 1D2; then by 2D8 MI would be typable. The last
step in a deduction of a type for MI must have form
' -. M: (p-'p)-'r t- I: P-*p
(-FE)
--> MI:Q.
Hence (p-*p)-+i would have to be an instance of P T(M) which is T. But then p-pp
would be an instance of (a-*a)-->a, which is impossible.

7D12.1 Warning D-completeness or -incompleteness is a property of sets of axioms,

not of the logics they generate. In fact it is possible to have two sets A and B of
axioms with AF = BF but such that A is D-complete and B is not. One example is
B = {(B), (C), (K)} and A = BF. There are also examples with A and B both finite
(Martin Bunder, unpublished notes 1994).
This brings our excursion into propositional logic to an end. (Except that if one
identifies A-terms with deductions the whole of this book is a look at the proof-theory
of propositional logic from a particular point of view.)
 8

Counting a type's inhabitants

Given a type r, how many closed terms can receive type r in TA2? As stated this
question is trivial, since if the answer is not zero it is always infinite; for example
the type a--+a can be assigned to all members of the sequence I, II, III, etc. But if we
change the question to ask only for terms in normal form, the answer is often finite
and interesting patterns show up which are still not completely understood.
The aim of this chapter is to describe an algorithm from Ben-Yelles 1979 that
answers the "how many" question for normal forms. For each r it will decide in a
finite number of steps whether the number of closed $-normal forms that receive
type r is finite or infinite, will compute this number in the finite case, and will list
all the relevant terms in both cases.
Ben-Yelles' algorithm can be used in particular to test whether the number of
terms with type r is zero or not, and as mentioned in 6B7.3 this gives a test for
provability in Intuitionist implicational logic.
The first section below will describe the sets to be counted. The next will show
some examples of the algorithm's strategy in action. Then in 8C-D the algorithm
will be stated formally, and the rest of the chapter will be occupied by a proof that
the algorithm does what it claims to do.t

8A Inhabitants
This section gives precise definitions and notations for the sets to be counted.

8A1 Definition (Inhabitants) An untyped inhabitant of r is a closed term M such

that I-A M:-r. A typed inhabitant of r is a closed typed term Mt. The sets of all typed
and untyped inhabitants of r will be called Habst(r) and Habsu(r) respectively, or,
when no confusion is likely, just
Habs(r).
A (#-)normal inhabitant of a type is an inhabitant in fl-nf. The sets of all typed and
untyped normal inhabitants of r will be called Nhabst(r) and Nhabsu(r) respectively,
or, when no confusion is likely, just
Nhabs(r ).
This fact was first proved mostly in Ben-Yelles 1979 and partly in Hirokawa 1993c.

108
 8A Inhabitants 109

A Pq-normal inhabitant of a type is an inhabitant in /3rl-nf. The sets of all typed and
untyped fl q-normal inhabitants of T will both be called
Nhabs,(r ).

A type with at least one inhabitant is said to be inhabited.

As remarked earlier the aim of this chapter is to count /3-normal inhabitants.
The following lemma will show that it does not matter whether we count typed or
untyped inhabitants.

8A2 Lemma If MT E Nhabst(T) then Mf E Nhabs (T); further, the type-erasing

mapping is a one-to-one correspondence between the typed and the untyped #-normal
inhabitants Of T (modulo The same holds for /3,1-normal inhabitants.

Proof Let M E $-nf; then M inhabits T if there exists a proof A of H M:T; and
by 2B3 this A is uniquely determined by M. And such proofs correspond one-to-one
with typed closed terms by 5A7.

8A2.1 Notation All terms in this chapter will be typed unless explicitly stated
otherwise. But for ease of reading they will often be written with some or all of
their types omitted.

8A3 Definition The number (0, 1, 2.... or oo) of members of a set S, counted modulo
ma if § is a set of A-terms, is called the cardinality of S or

#(S).
For #(Nhabs(T)) and #(Nhabs,(i)) we shall usually say just

#(T), #n(T).

8A4 Definition (Counting, enumerating) A distinction will be made in this book

between counting and enumerating a set S of #-normal inhabitants of a type T:
(i) to count S will mean to compute #(S) after a finite number of steps (even
when #(S) = oo);
(ii) to enumerate or list S will mean to enumerate S in the usual recursion-
theoretic sense, i.e. to output a sequence consisting of all the members of S (and no
non-members!), continuing for ever if S is infinite.

8A4.1 Comment The aim of Ben-Yelles' algorithm is to count inhabitants as well

as enumerate them. Mere enumeration would be easy: we could simply list all
closed typed fl-nf's in some standard order and for each one decide whether it is an
inhabitant of a given T by looking at its type. But counting is not so easy: we must
find a way of enumerating Nhabs(T) which will tell us after only a finite number of
steps whether the enumeration will continue for ever or not.
The strategy will be to do the counting in order of increasing depth in a sense to
be defined below, and to first count certain inhabitants called long/3-nf's from which
110 8 Counting a type's inhabitants
all others can be generated by ti-reduction. Before defining "depth" and "long" we
shall look briefly at the structure of arbitrary typed fl-nf's (cf. untyped fl-nf's, 1B10).

8A5 Lemma (Structure of a typed fl-nf) Let r be a type-context. Every f3-nf Mt E

TT(r) can be expressed uniquely in the form

2x21 ..
xTm.(V(P1......
MPn)T*)(21--*
(1) 1 m \\ 1 n
where m >_ 0, n >_ 0, and
(ii) T Tl-+... +Tm->r
for some T*, possibly composite, and each is a f3-nf that is typed relative to
(iii) f U {xl :Tl,...,xm:Tm}
(and the set displayed in (iii) is consistent). Further, if Mt is closed then m >_ 1 and
there is an i < m such that
(iv) V ° xi, Ti = Pl--'...-+Pn-+T

Proof Straightforward induction on JMl1. Note that m, n etc. are determined

uniquely, just as in 1B10 for untyped $-nf's.

8A5.1 Notation In the term in 8A5(i) the displayed occurrences of

1 ,..., Axt'^
Axt' m+ e Min,Mn
are called respectively the term's initial abstractors, head and arguments.

8A6 Definition (Depth(Mt )) Define the depth of a typed or untyped fi-nf thus:
(i) Depth(y) = Depth(Axl ... x,n y) = 0;
(ii) Depth(Axl ... xm yM, ... Mn) = 1 + Max Depth(Mj) if n > 0.
15j5n

8A6.1 Examples
Depth 2.

8A6.2 Lemma Depth(Mt) = Depth(Mk) < IM'1.

8A7 Definition (Long Jl-nf s) A typed /i-nf Mt is called long or maximal if every
variable-occurrence z in Mt is followed by the longest sequence of arguments allowed
by its type, i.e. if each component with form (zPl ... Pn)(n >_ 0) that is not in a
function position has atomic type.' An untyped fl-nf M is called long relative to a
type T if it is the type-erasure of a typed long /3-nf Mt. (By 8A2 Mt is unique.)
The sets of all long normal inhabitants of T (typed'or untyped) will both be called
Long(T ).

Sometimes "long ?I-normal form" is used in the literature for long fl-nf's but this is misleading, as these
terms are /3-nf's but not necessarily ry-nrs.
 8A Inhabitants 111

8A7.1 Example Let s = ((a-*b)->c)-*(a->b)-*c. Then the following normal inhabi-

tant of s is not long:
Mt - Ax(a-.b)-.cya-.b.x(a-.b)-cya-b

(because ya-.b has a type which "demands" an argument but none is provided). On
the other hand the following one is long:
Nt -
8A7.2 Definition The sets of all long normal inhabitants of s (typed or untyped)
with depth < d will both be called
Longs, d).

The next lemma shows that if we could enumerate long normal inhabitants the
others would be obtainable from these by il-reduction.

8A8 Lemma (Completeness of Long(s)) (Ben-Yelles 1979 Lemma 3.9.) Every nor-
mal inhabitant Of T can be rl-expanded to a long normal inhabitant of s. And this long
inhabitant is unique (modulo -a); i.e.
{Mt,Nt E Long(s) and Mt =,1 Nt} Mr -a N.
Proof Let Pt E Nhabs(s). First we must rl-expand Pt to a term Pt+ E Long(s).
Then we must show that Pt+ is unique, i.e. that
(1) Mt E Long(s), Mt C>,1 Pt Mt Pt+
Suppose Pt contains a short component, i.e. a component with form

(yQ1 ... Qn)° (n >_ 0)

that is not in function position and whose type a is composite, say

a (k >_ 1).

Choose distinct new variables zl, ... , zk not occurring in Pf and replace this com-
ponent by
(AZ1" ... Zkop '((yQI ... Qn) aZ1a, ... Z)
6k') a e

This replacement is an rl-expansion; it is easy to check that its result is still a

fl-normal form. (cf. Example 8A7.1.)
Make similar replacements until there are no short components left in Pt. (Each
replacement may introduce new short components with types Ql,... , ak, but these
types are shorter than a so the replacement process will terminate.)
Call the result Pt+. The proof that this term satisfies (1) is a routine induction on
IPtI.

8A8.1 Corollary Nhabs(r) = 0 Long (T) = 0.

112 8 Counting a type's inhabitants

Long normal inhabs: M

1
M2

(3r1-nfs: M1*11 M2*1

Fig. 8A 10a.

8A9 Definition (q-family) The set of all terms obtained by j-reducing Mt will be
called the q-family of MT (just as for untyped terms in 1C3), or
{Mt }n.

8A9.1 Note Let Mt E TT(I') for some F. Then {M'},, is finite (by the typed
analogue of 1C3.1) and all its members are in TT(I') (by 5B7.1). If Mt is a f3-nf
then so are all the members of {Mt}n (by the typed analogue of 1C9.3). Hence
(i) MT E Nhabs(T) {Mt}n S Nhabs(T).
Also, if MT is a /3-nf its q-family contains exactly one f3 -nf (by the typed analogue
of 1C9.3). Finally, each normal inhabitant of r is in the ?I-family of exactly one long
normal inhabitant, by 8A8.
The following lemma summarises this situation.

8A10 Lemma (i) The q families of the long typed normal inhabitants of T partition
Nhabs(T) into non-overlapping finite subsets, each q -family containing just one long
member and just one /3ri-nf. (See Fig. 8A10a.)
(ii) #(T) is finite or infinite or zero according as #, (T) is finite or infinite or zero.
(iii) #n(T) = #(Long(T)) .

8A11 Definition (Principal inhabitants) An untyped inhabitant M of T is called

principal if r is a principal type of M. A typed inhabitant Mt of T is called principal
if the deduction of H M':t that it encodes (see 5A7) is principal.
The sets of all principal inhabitants of r (typed or untyped) will both be called
P rinc(t ).

The sets of all principal /3-normal inhabitants of r (typed or untyped) will both
be called
Nprinc(t ).

8A11.1 Lemma Mt is a typed principal l3-normal inhabitant Of T iff M1 is an untyped

principal fl-normal inhabitant of T.
 8A Inhabitants 113

Fig. 8A12a.

8A11.2 Lemma Let MT+ be the unique member of Long(T) to which MTri-expands
(see 8A8). Then
MT E Nprinc(T) MT+ E Nprinc(T).

Proof The it-expansion in the proof of 8A8 preserves principality because of the
way the types given to zl, ... , zk are determined by the type v of the component that
is replaced.

8A12 Comment (Sets of of s) Three sets of fl-nf's have been defined so far in this
section, namely
Nhabs(T), Long(T), Nprinc(T),

and the sets of all f t -nf's in these sets will be called respectively
Nhabs,(T), Long,(T), Nprinc,,(T).

To clarify the relations between these six sets consider the type
T - (a-+a-->a)->a->a-*a.

(See Fig. 8A12a.) For this T the six sets are all distinct and in fact there is a term in
every space in Fig. 8A12a except one. In detail:
(i) Axa_a-+a.xa _ a-.a E Nhabs,, - (Nprinc U Long);
(ii) A,xa-'a-'aya.(xy)a-'a E Nhabs - Nhabs,, - (Nprinc U Long);
(iii) 1xa-'a-+a yaza . (xyZ)a E Long - Nhabs, - Nprinc ;
(iv) 1xa-.atiaya.(x(xyy))a-a E Nprincn - Long,,;

(v) 2xa-.a-ayaZa.(x(xyy)z)a E Nprinc fl Long - Nhabs,;

(vi) .xa-.a-.ayaZa.(xZ(xyy))a E Nprinc,, fl Long, ;
(vii) E Long, - Nprinc,,.
114 8 Counting a type's inhabitants
Proof of 8A12 (i) The term shown in (i) is clearly a Prl-nf and is not long. It fails
to encode a principal deduction for 2x x because the P T (Ax x) is not T but a--+a.
(ii) This term is obtained by il-expanding the term in (i).
(iii) This term is obtained by n-expanding (i) until it becomes long.
(iv) This term is easily shown to be principal by the PT algorithm (3E1). It fails
to be long because its second x from the right has only one argument.
(v) This term is obtained from (iv) by il-expansion; both occurrences of x now
have two arguments.
(vi) This term is like (v) but z and xyy have been reversed to make it an 11-nf.
(vii) This term is clearly long. However, its PT is not T but

8A12.1 Exercise* In Fig. 8A12a one of the eight spaces is empty: show that if T is
changed to the following type then every space in that figure will contain a term:
T-
8A13 Remark (Sets containing non-nf's) This chapter is concerned mainly with
counting normal forms so the sets Habs(T) and Princ(T) will play almost no role.
However, it is worth noting their relation to Nhabs(T) and Nprinc(T) before going on.
(i) If T has an inhabitant M then M has a fl-nf M*p by WN (5C1) and M*p is
also an inhabitant of T by subject-reduction (2C1); hence
Habs(T) 0 Nhabs(T) # 0.
(ii) Next, by the converse PT theorem (7A2) every type with an inhabitant has a
principal one, so
Habs(T) # 0 . Princ(T) # 0.
(iii) In contrast it is not true that Habs(T) # 0 Nprinc(T) 0. It is not even true
that Princ(T) # 0 Nprinc(T) # 0. Because T may have an inhabitant M, even a
principal one, such that PT(M) changes when M is reduced to M*p. An example is
T = a->a--*a;
by 8B4 this type's only normal inhabitants are Axy x and Axy y and it is easy to
check that neither of these is principal; on the other hand Table 3E2a(10) showed
that T has a non-normal principal inhabitant
(Axyz' K(xy)(xz))I.

8B Examples of the search strategy

The core of the counting algorithm will be a search algorithm which will seek long
normal inhabitants of T with increasing depths d = 0, 1, 2,.... The strategy of this al-
gorithm will depend on nothing more than some surprisingly simple comments about
the structure of long typed terms. These comments will be collected in 8B2 below
and then applied in some examples which will show the search strategy in action.
 8B Examples of the search strategy 115

8B1 Lemma Every type r can be expressed uniquely in the following form, where
m >_ 0 and e is an atom:
2=
Proof Easy induction on ITI

8131.1 Notation The occurrences of and e displayed above will be called

the premises and conclusion (or tail) of T respectively, and m will be called the arity
of T. (See Definition 9E5 for more details.)
Two type-occurrences will be called isomorphic if they are occurrences of the
same type. (cf. Definition (iv).) Iff the tail-components of o and T are isomorphic
we may say
Tail(a) = Tail(T).

8B2 Comments (Long typed P-nf s) Let T be any type; say T has form
T- (m >_ 0, e an atom)

and let MI be any fl-nf with type T. By 8A5, MT has form

(2xll 11 fnn1T*)(TI-+ ...Tk->T*)

T*
where 0 < k:5 m and __ If MT is long (see 8A7), then
(i) k = m,
(ii) T* __ e,
(iii) the types of xl,... , xm coincide with the premises of T,
(iv) the tail of the type of v is isomorphic to that of T,
(v) if MT is closed then m >_ 1 and v is an xi (1 < i< m) and
Ti - PV->...-+Pn-+e.
The following examples show how the above comments are applied.

8B3 Example (A type T with #(T) = 1) (Ben-Yelles 1979 p. 42.) The following type
has exactly one normal inhabitant:
T-
And its normal inhabitant (which is also both long and principal) is
ST Axa-.b-.cya-»bza.xZ(yz).

Proof We shall start by proving that Long(T) = {ST}.

Step 1. First look at the structure of T: in the notation of 8B2 we have m = 3 and
e - c, Tl =_ a--+b--+c, T2 - a--+b, T3 - a.
Hence any MT E Long(T) must have just three initial abstracted variables, say
zx33.(v(v1-...P"-c)M°' ... Mn")c)(TI-T2-T3~C).
MT
(AX 1X
116 8 Counting a type's inhabitants
Now by 8B2(iv), v must be one of xl, X2, X3 whose type's tail is isomorphic to that
of T. The tail of T is c, and the only Ti whose tail is an occurrence of c is Tl, so v must
be xl. Since Tl has two premises, xl must be followed by exactly two arguments.
Hence M must have form
M= Axi-b-cx2-bxa.xl-b-c uaVb
(1)
Step 2. Search for suitable Ua and Vb. First Ua: its type is an atom so Ua cannot
be an abstract, by 8B2(i). Thus Ua has form

Ua = (WPl ... Pr)a (r > 0)

where w is an xi whose type's tail is isomorphic to the tail of the type of U. This
tail is an occurrence of a, so the only possibility is

W x3.

Since the type of x3 has no premises we have r = 0 and hence

(2) Ua = x3.
Next search for Vb. Since b is an atom, Vb cannot be an abstract and its head must
be an xi whose type's tail is an occurrence of b. The only possibility is x2i and the
type of x2 allows only one argument, so
(3) Vb = x2-,bWa
for some W a.
Step 3. Search for Wa. Just as for Ua, the only possibility is
(4) W a = x3.
Conclusion. Modulo =_a there is at most one term in Long(T), namely
St
=
It is easy to check that St is actually in Long(T). Hence, by 8A10 and the fact that
St is n-irreducible,
Nhabs(T) = {St}.

By the PT algorithm it is easy to check that r is a principal type of St; hence

Nprinc(r) = {St}. 0
8B4 Example (A type r with #(T) = m) For each m >_ 2 the following r has exactly
m normal inhabitants and all are long and non-principal:
T m a-.... (m + 1 a's).

In fact its normal inhabitants are the following m terms (called projectors or selec-
tors):

II"n = Axj...xa.x° (1 <i<m).

 8B Examples of the search strategy 117

Proof By 8B2, any MT E Long(T) must have form

MT=_xl ...xm
a a
VVI ... V
where v =_ xi for some i < m. But the types of x1,. .. , x, have no premises, so n = 0.
Hence
MT =
It is easy to see that every such selector is in fact in Long (t). Also all these selectors
are rl-irreducible, so by 8A10,
Nhabs (T) = Long (T).

Now m > 2, so by the PT algorithm no selector is in Nprinc(T); hence

Nprinc(T) = 0.

8B5 Example (Types i with #(r) = 0) (i) No atomic type has inhabitants.
(ii) No type that is skeletal (i.e. in which each atom occurs at most once, cf. 7B2)
has inhabitants.

Proof (i) Every type with an inhabitant has a normal one by 8A13(i). But every
type with a normal inhabitant has arity m >_ 1 by 8B2(iii).
(ii) Let T be skeletal and let T - T i -+... (m >_ 1). If r had inhabitants it
would have at least one long normal one by 8A8, and by 8B2 this inhabitant would
have form
Axl ... xm xiMi ... M.
with xi having type Ti and the tail of Ti being an occurrence of e. But T is skeletal,
so e cannot occur in any Ti. Hence T has no inhabitants.

8B5.1 Corollary Intuitionist implicational logic is consistent in the sense that not all
.formulae are provable.

Proof If an atomic formula e were provable it would be the type of a closed term
MI by 6B7.1, contradicting 8B5.

8B6 Example (Peirce's law, cf. 6A1.2) The type r = ((a-+b)->a)-+a has no inhabi-
tants.

Proof Suppose MT E Long(T). Then by 8B2, MT must have form

ilx(a-.b)-a.ll
MT Ul ... Un (n > 0).
And v = x, since MT is closed. Hence n = 1, since the type of x has only one
premise. Thus
MT = Ax(a-b)-.a. x(a-b)-a Ua-.b

for some Ua-b. Since a--+b has just one premise, U'-.b must have form
Ua-b = Vr)b (r>_0).
118 8 Counting a type's inhabitants
Since Mi is closed, w must be x or y. But w must have a type whose tail is an
occurrence of b and neither x nor y has such a type, so no suitable U°-b exists.
Thus Long(r) _ 0 and hence Nhabs(r) _ 0 by 8A8.1. Hence Habs(r) _ 0 by 8A13(i).
0
8B7 Exercise* Verify the information in Table 8B7a by proving that the types
shown there have no other inhabitants than those displayed. The notation used in
the table's last row is:
PO AX-XI,
Pn = (n z 1),
Qn,i = (Ayn x(Rz Yi)) ...)) (n > 1,1 < i < n).

Table 8B7a. Some types and their normal inhabitants

Type r Nhabs(r) Long(i) Nprinc(z) Notes

1 a None None None See Ex. 8B5
2 a-+a I I I I

3 a-b-*a K K K K
4 (a-*b)-(c-a)-+c-*b B B B B
5 (a-*b-c)->b-a-+c C C C 2xyz xzy
6 (a-*b-*c)- (a-*b)-a-c S S S S
see Ex. 8B3
7 (a-+a--b)-+a-+b W W W W Axy.xyy
8 a-* ... -+a-+a II.,....n II; None i - xl ... Xm'xi
m

{m+la's,m>-2} see Ex. 8B4

9 (a-+a)-a-+a n=
10 (a- 1, 1 1 1 1-
11 ((a->b)-*a)-*a None None None x is Peirce's law
see Ex. 8B6
12 ((a-*a)-+a)---*a Po, P., Po, P., n>_ 1,1 <I<n;
see Ex. 8B7 for
Po, P. Q,,,i

8C The search algorithm

The search strategy introduced informally in the last section will be crystallised into
a formal algorithm in this one. The treatment will be in full detail as a preparation
for the correctness-proof at the end of the chapter, but the core strategy is very
simple as the examples in the last section have shown, and the reader who prefers
to stay on an informal level should omit the present section and move on to the
counting algorithm in 8D.
The method of the search algorithm will be to first look for members of
 8C The search algorithm 119

Long(s) with depth d = 0, then d = 1, etc. The search will output a sequence
d('r, 0), .Sd7(t, 2),... of finite sets of expressions that will serve as "approxima-
tions" to these members. Those in .W(T,d) will look like typed fl-nf's with depth < d
but may have some "holes" to be filled in by the algorithm at a later stage. These
holes will be represented by new symbols called meta-variables and the approxima-
tions will be called nf-schemes. (In Example 8B3 the expression on the right-hand
side of (1) was in effect a nf-scheme and "U" and "V" were its meta-variables.)
To construct JV('r, d+1) the approximations in d(x, d) will be extended by replacing
their meta-variables by certain chosen nf-schemes with depth 1.

8C1 Definition (Nf-schemes) We assume given an infinite sequence of expressions

called meta-variables, distinct from each other and from term-variables. Meta-
variables will be denoted by

Nf-schemes are defined like terms in 1A1 except that they may contain meta-
variables or term-variables or both, and must satisfy the following restrictions:
(i) each nf-scheme is a $-nf without bound-variable clashes;
(ii) meta-variables do not bind, i.e. AV is forbidden;
(iii) in a composite nf-scheme meta-variables only occur in argument positions (as
defined in 9A1);
(iv) each meta-variable in a nf-scheme occurs only once.
Proper nf-schemes are those that contain at least one meta-variable.

8C1.1 Examples The expression V2) is a nf-scheme. So also is the atomic

expression V. But the following are not nf-schemes:
2.V V, Vxy, Ax. V.

8C1.2 Notes (i) The reader familiar with "context" notation will see that a nf-
scheme with k meta-variables is essentially a context containing k different kinds of
hole.
(ii) The restrictions in 8C1 are imposed simply because no wider class of expres-
sions will be needed below.
(iii) Non-proper nf-schemes are just terms (in /3-nf without bound-variable clashes).
(iv) In what follows, a new variable or meta-variable will be one that has not been
used earlier in the chapter.
(v) Most of the term-notation introduced in Chapter 1 extends to nf-schemes in
an obvious way, and the same holds for the definition of depth in 8A6. But a few
concepts need defining afresh and the next three definitions will cover these.

8C2 Definition A closed nf-scheme is one containing no free variables. (But possibly
containing meta-variables, for example Ax-xV and V are closed.)

8C3 Definition (Typed nf-schemes, TNS(F)) In this chapter a type-context F is

any set of assignments whose subjects are meta-variables or variables, such that
120 8 Counting a type's inhabitants
no subject receives more than one type in i' (cf. 2A5). The set TNS(i') of all
typed nf-schemes relative to F is defined just like TT(F) in 5A1 but satisfying the
restrictions in the definition of "nf-scheme" (8C1).
The typed-term notation in Chapter 5 will be used here for typed nf-schemes also.

8C4 Definition (Long typed nf-schemes) A typed nf-scheme XT is long if each

component of XT with form
(n>0)
that is not in a function position has atomic type. (cf. Definition 8A7.)

8C4.1 Example The nf-scheme (i) below is long, though in contrast the term (ii) is
not:
x(a-.b)-.c Va-.b
(i)
(ii) x(a-.b)-.cZa-.b

The description of the search algorithm will now be given: it will begin with a
precise statement of what the algorithm does (in the following theorem), followed
by the algorithm itself.

8C5 Search Theorem for Long(t) (Ben-Yelles 1979 §§3.14-3.17.) The search al-
gorithm in 8C6 below accepts as input any composite type t and outputs a finite or
infinite sequence o f sets d(T, d) (d = 0, 1, 2, ...) such that ,for all d >- 0,
(i) each member of d(t, d) is a closed long typed nf-scheme with type t, and is either
(a) a proper nf-scheme with depth d, or
(b) a term with depth d - 1;

(ii) 9/(T, d) is finite;

(iii) Long ('t,d) C &/(t,0) U... U.d(t,d+ 1);
(iv) if we call the set of all terms in d(T, d) "dterms(t, d) ", then
Long(r) = USVterms(T,d)
dzO

8C5.1 Proof-note Parts (i)-(ii) will be proved by induction on d in the notes

accompanying the steps of the search algorithm.
Part (iii) states that the algorithm is "complete", i.e. finds everything it is looking
for; its proof is postponed to 8F1.
Part (iv) follows easily from (i)-(iii).

8C6 Search Algorithm (Ben-Yelles 1979 §§3.14-17.) Input: any type T. If t is an

atom it has no inhabitants (by 8B5). If t is composite, proceed as follows.
Step 0. Choose any meta-variable V and define
,4(T,0) = {VT}.
 8C The search algorithm 121

Note. This .W('t, 0) trivially satisfies 80(i)-(ii). (The algorithm may be seen as
building approximations to an unknown term Mt; Vt is the weakest approximation
and represents the fact that at this stage we know nothing at all about Mt other
than its type.)
Step d + 1. Assume sl(r,d) has been defined and satisfies 8C5(i)-(ii).
Substep I. If C(T,d) = 0 or no member of .c/(r,d) contains meta-variables then
stop. (In this case .W(r,d+ 1) is undefined and the algorithm's output is just the
finite sequence d(i, 0), ... , d ('C, d).)
Substep II. Otherwise, begin the construction of d(,r,d+ 1) by listing the proper
nf-schemes in d (?, d) and applying IIa-IIb below to each one.
Subsubstep IIa. Given any proper Xt E d(r,d), list the meta-variables in Xt;
say they are

Vi'..., Vyq (q ?1),

and apply Ilal-M2 to each one.

Part IIal. Given any meta-variable VP in an Xt E Qf(T,d), say

(1) p= (m>0);
first list all i < m for which Tail(aj) = a = Tail(p). (If there are none or
m = 0, go direct to IIa2.) For each such i, ai has form

(ni > 0).

Define
il1...V aini'')a
(3) yip Ax1'...xm'(xi
where the x's and V's are distinct new variables and meta-variables. (Yif
is called a suitable replacement for VP.)
Part IIa2. List the abstractors that cover the (unique) occurrence of VP in
Xt, in the order they occur in XT from left to right; say they are

(4) Azi', ... , ?[` (t > 0).

List all j < t (if any) such that Tail (Cj) = a. For each such j,1 j has form

(5)
Define
j (hj >_0).

(6) Zj Ax" Vii'...Vjhj i)a

where the x's and V's are distinct and new. (Ze is called a suitable
replacement. for VP.)
122 8 Counting a type's inhabitants
Notes. (i) It is easy to see that each Y;P defined in IIal and each Zf defined in
IIa2 is a long nf-scheme with depth < 1 and the same type as VP, so VP can be
replaced by Y;P or Z' in XT without violating type-restrictions or the restrictions
in the definition of long nf-scheme (8C1, 8C3-4). And the result of making such a
replacement will clearly have depth < d + 1.
(ii) YAP and Z° need not contain meta-variables. In fact YAP is without meta-
variables if o; is an atom, and in this case

Similarly Zf is without meta-variables if ( is an atom, and in this case

C
Zf Ax?'1 m
xOm .

Hence if Y;P or Zf is without meta-variables its depth is 0 and if VP is replaced by

it then Depth(XT) will not increase.
(iii) The total number of suitable replacements (Y's and Z's) for VP is at most
m+t.
Subsubstep IIb. When IIal-IIa2 have been applied to all the meta-variables
in XT the result is a list of suitable replacements for each V; in X.
If one or more of V1,...,V9 has no suitable replacements, abandon XT,
calling it a reject, and start applying Substep II to the next member of
d). (A reject will generate no members of W('t, d + 1).)
If all of 17j,..., V9 in XT have suitable replacements, XT is called extend-
able; in this case list all possible sequences

(7) (W1',..., WyP")

where W; is a suitable replacement for V; for i = 1,. .. , q. For each sequence

(7) construct a new nf-scheme X*T from XT by simultaneously replacing V;
by W; in XT for i = 1,..., q. (Call each sequence (7) a suitable multi-
replacement and call X *T an extension of XI; if this extension is a term
call it a success.)
Notes. (i) The number of extensions of XT is finite since each is generated by a
suitable multi-replacement and the number of these is clearly finite (for a given XT).
(ii) To construct an extension each meta-variable in XT is replaced by either a
nf-scheme with depth 1 or a term with depth 0. If the latter holds for all the
meta-variables in XT the extension has depth < d and is a success. If the former
holds for at least one meta-variable, the extension has depth < d + 1 and contains
meta-variables.
Substep III. Finally, if the set d(r,d) contains at least one proper nf-scheme,
define d(r,d+1) to be the set containing all the extensions of all the extendable
proper nf-schemes in W(r,d).
Notes. (i) By the notes after IIa and IIb above it is easy to check that sd(r, d + 1)
satisfies 8C5(i) and (ii).
(ii) By the way, d('t, d + 1) does not contain d(T, d) as a subset.
 8C The search algorithm 123

8C6.1 Example (cf. Example 8B3.) Applying the search algorithm to the type
T - (a-*b--*c)-+(a--+b)->a--+c
produces the following sets:
SI(T, 0) = {VTj,
{1x1-b-cxz-bx3.xa-b-cVl Vz
d(T,2) = l 1 2 3 1 3( 2

(T, 3) = a-.b-.c a-.b a, a-b-c a a-b a

1
x2 x3 x1 x3(x2 x3)}

8C6.2 Example (cf. Example 8B4.) Applying the search algorithm to the type
T =_ a-+ ... --+a-+a

(with m+1 a's) produces the following sets:

sd(T,0) = {VT},
(T,1) = {(2xi ... xm'xi), ... , (Axj ... xmxm)}.

8C6.3 Example (cf. Example 8B6.) Applying the search algorithm to the type
T ((a-+b)-+a)->a
produces the following sets:
,%/(T,0) = {VT},
,V(T 1) = {Ax(a-.b)-.a,x(a-+b)-ava_.b}
1 1 1

d(T, 2) = 0.

8C6.4 Exercise* List members of the sets sad(T,0),szl(T,1).... for all the types in
Table 8B7a.

8C6.5 Comments (i) Besides Ben-Yelles' 1979 search algorithm there are others
in Zaionc 1985 and 1988 and in Takahashi, Akama and Hirokawa 1994. Also
procedures essentially equivalent to the main parts of 8C6 are used in some decision-
algorithms for provability in intuitionist logic, cf. 6B7.3, and in algorithms that seek
to unify pairs of typed A-terms, for example the original one in Huet 1975 §3.
(ii) The algorithm in 8C6 can be viewed as a context-free grammar for generating
a language whose expressions are exactly the inhabitants of T. In the standard
notation for context-free grammars (see for example Hopcroft and Ullman 1979
Ch. 4) the meta-variables in 8C6 would be called non-terminal variables and each
suitable replacement defined in Subsubstep Ila of 8C6 would determine a production
in the grammar. The set of all such productions need not be finite, however, so the
original definition of "grammar" must be relaxed if this view is to be made precise.
This view was taken in Huet 1975 §3, Zaionc 1985 §3 and 1988 §5, and was used
and analysed in detail in Takahashi, Akama and Hirokawa 1994.
(iii) A very smooth description of a search algorithm in terms of solving sets of
polynomial equations over the domain {0, 1, 2, ... , co} is in Zaionc 199-.
124 8 Counting a type's inhabitants
8D The Counting algorithm
In this section the search algorithm will be extended to count and enumerate normal
inhabitants. The extension will be very simple indeed to state; but the proof that it
works will need some care and will be postponed to Sections 8E-F.
As noted in 8A2, it does not matter whether typed or untyped inhabitants are
counted since there is a one-to-one correspondence between the two: in this section
"Long(T)" and "Nhabs(T)" will denote sets of typed terms.

8D1 Definition (i) For the sets .QI(T, d) and dterms(T, d) introduced in 8C5, define
d(T, < d) = S /(T, 0) U ... U a (T, d),
dterms(T, < d) = dterms(T, 0) U ... U Wterms(T, d).

(ii) For any T, recall from 2A2 that ITI is the number of atom-occurrences in T
and IITII is the number of distinct atoms in T; define

®(T) = ITI X IITII.

8D2 Stretching Lemma If Long(T) has a member MT with depth d >_ lit II then it has
members with depths greater than any given integer, and hence is infinite.

Proof-note The proof will be given in 8F2 (cf. Ben-Yelles 1979 §§3.20-3.21). It will
show that if MT is in Long(T) and has depth d >_ lit Ii then MT must contain two
distinct components B and B' with the same type and with one inside the other. Then
replacing the smaller one by the larger will change MT to a new term M`T deeper than
MT, and it will be easy to check that M'T is a genuine typed term and is in Long(T).
Then a similar replacement in M'T will change it to a still deeper term, and so on.1

8D3 Shrinking Lemma If Long(T) has a member MT with depth >_ 11(T) then it has
a member NT with
®(T) - IITII < Depth(NT) < ®(T).

Proof-note The proof will be given in 8F3 (cf. Ben-Yelles 1979 Thm. 3.28). Just
as for the previous lemma, MT will be seen to contain two components B and B'
with the same type and one inside the other; but now M*T will be constructed by
replacing the larger by the smaller. A problem is that if B and B' are not chosen
carefully this replacement risks deleting some abstractors and giving a non-closed
M*T: to prevent this, and to ensure that the depth of M*r lies in the range required,
B and B' will be chosen only after a close analysis of the structure of MT.2

I
This lemma is very like the pumping lemma in the theory of context-free grammars and regular
languages, see for example Hopcroft and Ullman 1979 §3.1 and Comment 8C6.5(ii) above. And the
next lemma will also be very like one in that theory, see Hopcroft and Ullman 1979 §3.3 Theorem 3.7,
though its proof here will be more complicated due to the presence of bound variables. For a
precise statement of the analogy between typed 2-calculus and the theory of context-free grammars
see Takahashi et al. 1994, where a generalised concept of context-free grammar is introduced and
generalizations of the classical results are proved which imply the corresponding 2-results.
2
Ben-Yelles 1979 Thm. 3.28, whose lines the proof in 8173 will follow, was actually slightly weaker than
the above lemma. The first proof of a full-strength shrinking lemma was outlined in Hirokawa 1993c;
it was different from the one to be given in 8F3.
 8D The Counting algorithm 125

8D3.1 Corollary If Long(T) has a member Mt with depth > D(T) then it has a
member Nt with
IITII <_ Depth(Nt) < D(T).

Proof If Long(r) has a member then T is composite by 8B5. Hence ITI > 2, so

D(T) - IITII >_ 211211- 11TII = IITII.

The result then follows by the shrinking lemma.

8D4 Theorem (Counting long normal inhabitants) (Ben-Yelles 1979 Cor. 3.33.)
When given a type r the algorithm in 8D5 below outputs #(Long(T)) and an enumer-
ation of Long(T).

Proof See the notes in brackets in the following algorithm.

8D5 Counting Algorithm for Long(r) If T is an atom, Long(T) is empty by 8B5. If

T is composite, apply the search algorithm (8C6) to r; this outputs a finite or infinite
sequence of sets
.(T, d) (d = 0,1, 2, ...).

Stop the search algorithm at d = D(T) and enumerate dterms(T, < D(T)).
[By 8C5 this set is finite and contains all members of Long(T) with
depth < D(T).]
Case I: (Vterms(T, < D(T)) = 0. Then Long(T) = 0.
[By 8D3, if Long(T) had a member with depth >_ D(T) it would have
one with depth < D(T).]
Case II: c/terms(T, < D(T)) has a member with depth >_ IITII. Then by 8D2 Long(T) is
infinite. To enumerate Long(T), apply the search algorithm to enumerate dterms(T, d)
ford=0,1,2,....
[By 8C5(iv) the union of these sets is Long(T).]
Case III: dterms(T,< D(T)) has members but they all have depth < IITII. Then
Long(T) = dterms(T, < D(T)), which is finite.
[The only way for Long(T) to differ from this set would be for Long(T)
to have members with depth d >_ D(T), but by 8D3.1 it would then
have a member with lit II < d < D(T) contrary to the assumption of
the present case.]

8D5.1 Corollary (Counting fn-normal inhabitants) The algorithm in 8D5 can be

used to count and enumerate Nhabsq(T) for every T.

Proof By 8A10 the members of Nhabs,t(T) are the q-nf's of those of Long(T). And
by 8A10(iii), #,t(T) = #(Nhabsq(T)) = #(Long(T)).

8D5.2 Corollary (Emptiness test) The algorithm in 8D5 can be used to decide whether
a type T has no inhabitants.
126 8 Counting a type's inhabitants
Proof By 8A13(i), Habs(T) = 0 e:> #(Nhabs(T)) = #(T) = 0. And by 8A10(ii) and
(iii), #(T) = 0 a #(Long(T)) = 0.

8D6 Theorem (Counting all normal inhabitants) (Ben-Yelles 1979 Ch. 3.) When
given a type T the algorithm in 8D7 below outputs #(T) and an enumeration of the set
Nhabs(T) of all the /3-normal inhabitants of T.

8D7 Counting Algorithm for Nhabs(r) If T is an atom then Nhabs(T) is empty by

8B5. If r is composite, apply Algorithm 8D5 to count Long(T).
Case 1: If Long(T) = 0 then Nhabs(T) = 0. [By 8A8.1]
Case II: If Long(T) # 0 then Nhabs(T) is counted and enumerated by counting and
enumerating Long(T) and enumerating the n family of each member of Long(T).
[By 8A10 these rl-families are finite and their union is Nhabs(T).]
Before ending this section we shall digress from the main theme to look at an
interesting application of the counting algorithm to types that contain only one
atom. Such types have been studied and used as tools by several workers, for
example see Statman 1979a, 1979b and 1980, Dekkers 1988, and Zaionc 1987b and
1990.

8D8 Definition A type T will be called monatomic if IITII = 1, i.e. if only one atom
occurs in T.

8D9 Theorem (Ben-Yelles 1979 §3.25.) Let r be a monatomic type with the form
T -TI ->...Tm->a (m > 0). Then
(i) if at least one Ti is composite, #(T) is either oo or 0;
(ii) if TI - ... - Tm = a then #(T) = M.

Proof Part (ii) is just 8B4. To prove (i), assume Nhabs(T) is finite and non-empty
and assume r has form
T=
with m >_ 1 and at least one premise composite. By 8A8 Long(T) is also finite and
has a member MT. By 8D2, Depth(MT) < fITli = 1, so MT must have form
MT = AXT I 1
XT'"
m ' xTI
P (1<p<m).
And TP = a since MT is long. Choose a composite premise of T; say it is Ti and has
form
(m;>_ 1).

For j = 1, ... , m, each T;j has form

T+,i = Tf,i,i-*...-->T;,i,m+,;-'a (m;j > 0).

For each j < m, choose distinct new variables yl,... , ym J and define
P.+,i
- T+i,l
Y1 .. Ym; T 1,i,m;' j . Xap
 8E The structure of a nf-scheme 127

where xP is the rightmost variable in MT (and is therefore distinct from the y's).
Then define

NT - AxT1.. xtmm. x ``P Ti j ... P miTi,m.ja

1 i 1

Clearly NT E Long(T). But Depth(NT) > 1 since NT contains mi >_ 1 arguments.

Hence Long(T) is infinite by 8D2. 0
8D10 Problems The above algorithms have solved the problem of counting normal
inhabitants. But the following related problems are still open.
(i) Counting principal normal inhabitants (Nprinc(T), Nprinc,,(r)). I know of no way
to count these sets that covers the case when Nhabs(T) is infinite. Hirokawa 1991a
contains a structural analysis of principal deductions that may help. It is applied to
give interesting results in Hirokawa 1991a-c, 1992b and 1993a-b, and these results
include stating some sufficient conditions for Nprinc(T) to have only one member.
(ii) Counting principal inhabitants in general (Princ(T)). Each member of Princ(T)
must have a fl-nf and that of must have type T, but since /3-reduction does not
preserve principality there is no direct connection between Princ(T) and Nprinc(T).
By the way, if Nprinc(T) = 0 but Princ(T) # 0 then T has the curious property of
being the principal type of a non-normal form but not of a normal form. (An
example of such a type is a-*a-*a, see 8A13(iii).) Is there any neat characterization
of the types that have this property?
(iii) Counting inhabitants in restricted classes of terms, for example the Al-terms,
BCKA-terms or BCIA-terms introduced in 1D. The search algorithm would give no
trouble in such a class but the crucial shrinking lemma (8D3) would need a new
proof for each restricted class, and indeed might fail for some classes. An example
of a result for a restricted class is the proof in Komori and Hirokawa 1993 that for
BCKA-terms #(T) is always finite. An example of a currently unsolved problem for a
restricted class is that of finding a decision-procedure for the Hilbert-style logic T.
whose axioms are the principal types of B, B', I, W (see 7D11(ii)). This is equivalent
to finding an algorithm to decide whether a given type T has any inhabitants that
are applicative combinations of these four terms.

8D11 Comment (Efficiency) At first glance the algorithms in this section may seem
remarkably efficient, as they terminate in only ITI X IITII steps. But this impression
is totally misleading, as each one of these steps involves listing a set d(T,d) which,
although finite, may be very large indeed. In fact any counting algorithm will decide
as a special case whether Nhabs(T) is non-empty, which is equivalent to deciding
whether T is provable in Intuitionist logic (by 6B7.2), and this decision problem is
polynomial-space complete (Statman 1979a Prop. 4).

8E The structure of a nf-scheme

The earlier sections left the stretching and shrinking lemmas unproved (8D2, 8D3),
as well as the "completeness" part of the search theorem for Long(T) (8C5 (iii)).
The aim of the next section will be to fill in these gaps, and the present one
128 8 Counting a type's inhabitants

.Tree for

m n Y
v 0 1 j
1
n-1 Tree for
0 1 2
mn-1
O1

vYl

vYl ..Yn

X xl...xIn(vYI ...Yn) 0
Fig. 8131a.

will lay the foundation by analysing the structure of an arbitrary long typed nf-
scheme.
A key role will be played by a slightly strengthened form of the subformula
property (2B3.1). That property says in effect that the types of all the components
of a closed fl-nf Mi are subtypes of r, and this implies that all the successes
produced by the search algorithm, growing deeper and deeper, have the types of
their components drawn from the same fixed finite set. This limitation is the source
of the bounds in the stretching and shrinking lemmas.

8E1 Notation Recall that a nf-scheme is essentially a $-nf that may contain meta-
variables under certain restrictions, see Definitions 8C1 (untyped) and 8C3 (typed).
The early parts of the present section will apply to both typed and untyped nf-
schemes, so types will be omitted when nf-schemes are written. But later parts will
apply only to typed nf-schemes and types will then be displayed.
We shall need the notation for positions, components and construction-trees
introduced in 9A1-4.
In writing positions a sequence of n 0's may be written as on (with 00 and
similarly for l's and 2's.
Recall from 8A5 that every non-atomic nf-scheme X can be expressed uniquely
in the form
(1) X- (m+n>_ 1),
where v is one of xl,...,xn if X is closed. The construction-tree of such an X is
shown in Fig. 8Ela. The head and arguments of X are v_ and Y1,...,Yn. (If X is an
atom its head is X and it has no arguments.) Note that the position of L is
(2) 0"`1n-`2 (1 < i < n).
 8E The structure of a nf-scheme 129

x(Xu.uV1 )V2

?.y.x(Xu.uV1)
2
X xy.x(k u.uVI )V2

Fig. 8E2.1a. Construction-tree for

8E2 Definition (Subarguments) A subargument of a typed or untyped nf-scheme X

is a component that is an argument of X or an argument of a proper component
of X.

8E2.1 Lemma A component Y of a typed or untyped nf-scheme X is a subargument

f its position is not 0 and the last symbol in its position is 2.

Proof Induction on IXI. As an example, consider the subarguments in the tree in

Fig. 8E2.1a.

8E2.2 Note (i) All occurrences of meta-variables in a composite nf-scheme are

subarguments (by restriction 8C1(iii) in the definition of nf-scheme).
(ii) A subargument of a subargument of X is a subargument of X.

8E3 Definition (Relative depth) The 2-length of a position-string p is the number of

2's in p. The depth in X of a subargument Z of X is the 2-length of its position (i.e.
the number of right-hand choices made when travelling up the tree of X from the
bottom node to Z. cf. Fig. 8E2.1a).

8E3.1 Lemma Let X be a typed or untyped nf-scheme with Depth(X) > 1. (Depth is
defined in 8A6.) Then
(i) Depth(X) is the maximum of the depths in X of all subarguments of X,
(ii) X has at least one subargument whose depth in X is the same as Depth(X), and
each such subargument is an atom or abstracted atom.

Proof By induction on IXI, using 8A6.

130 8 Counting a type's inhabitants
8E4 Definition (Argument-branch) If Z is a subargument of a typed or untyped
nf-scheme X, the argument-branch from X to Z is the sequence

ZZo,Z1,...,Z) (k > 1)
such that Zo - X and Z1 is an argument of Z.-1 for i = 1,...,k, and Zk = Z. It is
called unextendable if Z is an atom or abstracted atom. Its length is k (not k + 1).

8E4.1 Lemma For any typed or untyped nf-scheme X:

(i) the depth in X of a subargument Z is the same as the length of the argument-
branch from X to Z ;
(ii) Depth(X) is the maximum of the lengths of all the argument-branches in X.

Proof For (i) use induction on IXI; for (ii) use 8E3.1.

8E5 Definition (IA, CA) Let Z be a subargument of a typed or untyped nf-scheme

X; say
Z = Axl ... xm yZi ... Zn (m > 0, n > 0).
The Initial Abstractors sequence IA(Z) is the (possibly empty) sequence
IA(Z) = (xl,...,xm).
The Covering Abstractors sequence CA(Z,X) is defined to be
CA(Z X) = (zl,...,zq),
where Azl,...,AZq are the abstractors in X whose scopes contain Z. written in the
order they occur in X from left to right. Also define:
Length(IA(Z)) = m, Length(CA(Z X)) = q.

8E5.1 Note (i) If X has no bound-variable clashes the member of IA(Z) are distinct
and so are those of CA(Z X).
(ii) IA(Z) and CA(Z X) are sequences of variables not components.
(iii) For typed nf-schemes each variable in IA(Z) or CA(Z X) is typed.
(iv) If the argument-branch from X to Z is (Zo,...,Z) (k > 1), then
CA(Z X) = IA(Zo) * ... * IA(Zk_l)
where "*" denotes concatenation of sequences. (Because the abstractors whose
scopes contain Z are exactly the initial abstractors of
The next definition and lemma will have meaning for typed nf-schemes only. The
lemma will be the strengthened form of the subformula property mentioned at the
start of the section, and will lead to a computation of upper bounds for several key
sequences and sets. Part of it will use notation from 9E.

8E6 Definition (IAT) Let Z° be a subargument of a typed nf-scheme XT; say

Z' _ 2xj'... xm yZ1... Zn (m > 0, n > 0).
 8E The structure of a nf-scheme 131

The Initial Abstractors' Types sequence IAT(Z°) is defined to be

IAT(Z°) =

Length (IAT(Z°)) = m.

8E7 Enhanced Subformula Lemma (cf. Ben-Yelles 1979 Lemma 3.31.) If Z° is a

subargument of a closed long typed nf-scheme XT, then
(i) a occurs as a positive subpremise in 2 (as defined in 9E6-8),
(ii) if a is an atom, IAT(Z°) = 0,
(iii) if a is composite, IAT(Z') E NSS(T) (defined in 9E9),
(iv) NSS (a) s NSS (T).

Proof Since Z° is long, IAT(Z°) coincides with the sequence of all premises of a,
so (ii) holds. Also if or is composite we have
(1) IAT(Z°) E NSS(a)
by the definition of NSS (a) in 9E9. Now (i) implies (iv) by 9E9.2(iii), and (iv) and
(1) imply (iii). Hence only (i) remains to be proved.
The proof of (i) is an induction on IXTI. To make this work we shall prove
If XT is a long member of TNS(r) (defined in 8C3) and
r = {UI:BI,...,up:O ,V1:01,...,Vq:¢q}
(2) and Z' is a subargument of XT, then a occurs as a positive
subpremise of
elm...-ABP-->T.
Basis. If XT is an atom the conclusion of (2) holds vacuously.
Induction step. Let XT have form
Xnn)e)(TI-...-+T,n--*e)
(3) (A.xl' ... x;;
where m, n > 0 and T - TI-+... -T,,, -e. Then either y - xi for some i < m or y ui
for some i< p. If y - xi then
(4) Ti PI -+ ... ->Pn-+e
and if y - ui then
(5) ei = Pl-+...->Pn->e.
In both cases each of p1i..., Pn occurs as a positive subpremise of
(6) e1--+...->O
Now Z° must be in an for some j < n. If Z° then a pi and the
conclusion of (2) follows by the above. Next, suppose Z° is a subargument of XP'.
Note that
E TNS({xl :Ti,...,x,n:T,n} U F).
Hence, by the induction hypothesis, a occurs as a positive subpremise of
T1-> -+eP-+P).
Thus a occurs as a positive subpremise of (6), giving (2). 0
132 8 Counting a type's inhabitants
8E7.1 Corollary If Xt is a closed long typed nf-scheme, the type of each meta-variable
in XT either occurs as a positive subpremise of T or is T itself.

Proof By 8E2.2(i) and 8E7(i).

The main effect of 8E7 is to connect IAT(Z°), which in general depends on the
structure of ZI and hence implicitly on that of XT, with NSS(T) which depends on r
and nothing else. The next corollary will use this to deduce reasonably neat bounds
for IA(Z°) and CA(Z,XT).

8E7.2 Corollary If XT is a closed long typed nf-scheme and Z° is a subargument of

XT or Z° XT, then
(i) Length(IA(Z°)) = Length(IAT(Z')) < ITI - 1,
(ii) Length (CA(Z,XT)) < (1r - 1) x Depth(XT).
Further, if 2v°',...,Av°' are all the abstractors in XT (not just its initial ones), then
(iii) {pi...., pr} has < ITI - 1 distinct members.

Proof For (i): Length (IAT(Z°)) < ITI - 1 by 8E7(iii) and 9E9.3(iv).
For (ii): If Z - X the left side of (ii) is 0. If Z # X let (Zo,... , Z) (k >- 1) be the
argument-branch from X to Z; then by 8E5.1(iv)
Length(CA(Z,X)) = Length(IA(Zo)) + + Length(IA(Zk_l))
< k(JTJ - 1)
by (i). But Depth(X) > k by 8E4.1(ii), so (ii) holds.
For (iii): Each pi is in IAT(XT) or in IAT(YB) for some subargument YB of XT;
and in both cases pi E U NSS (T) (in the former case trivially, and in the latter case
by 8E7(iii)). Then use 9E9.3(iii).

8E7.3 Exercise* Show that if T is composite and d >_ 1 and .nI(,r,d) is defined, then
(i) each XT E d) contains < T1d meta-variables,
(ii) #(_q/(T, d)) < (d x
ITI)(IT1"

') x #(S/(T, d - 1)),

(iii) #(-%/(T, d)) < 1 x 2ITI x 3(1T12) x ... x d(ITI`'-') x TJ(1+JTJ+IT12+ +1T1d-').

8F Stretching, shrinking and completeness

This section fills in the three gaps that were left in the verification of the counting
algorithm in 8C-D: the stretching and shrinking lemmas and the "completeness"
part of the search theorem.

8F1 Search-Completeness Lemma Part (iii) of the search theorem 8C5 holds; i.e. if
T is composite and d > 0, then
Long (T,d) 9 d(T, < d + 1).
 8F Stretching, shrinking and completeness 133

Proof The lemma will be proved by induction on d but to make the induction work
we must prove something slightly stronger. Recall that Long(T,d) is the set of all
long inhabitants of T with depth < d (8A7.2). Let L*(T,d) be the set of all long typed
closed nf-schemes XT such that Depth(XT) = d and
(1) ( J(a) XT is proper and all its meta-variables have depth d in XT,
t (b) all subarguments with depth d in XT are meta-variables.
We shall prove both the inclusions
(2) L*(T, d) 9 sl(T, < d),
(3) Long (T, d) c d + 1),
where (2) is understood modulo renaming of meta-variables (i.e. (2) says that if
XT E L*(T, d) then d(T, < d) contains either XT or a nf-scheme that differs from XT
only by replacing its meta-variables by distinct others.)
Basis: d = 0. For (2): the only nf-schemes with depth 0 are meta-variables;
also .sa l(T,0) _ {VT} by Step 0 of the search algorithm (8C6), so (2) holds modulo
renaming.
For (3): let T - (m >_ 1), and let MT E Long(T,0). Then MT is a
term and has form
(4) AY1 ... Ym'. Y ` (1 < i < m, Ti = e).
Now d(,r,0) = {VT}. To construct C/(T,1) apply 8C6 Step 1: Part Ilal therein will
output (4) as a suitable replacement for VT because the tail of Ti is isomorphic to
that of T. Hence .9/(T, 1) is defined and contains MT.
Induction step: d to d+1. For (2): let X E L*(T,d+1). (Types will not be displayed
from now on.) Then Depth(X) = d + 1, so by 8E3.1(ii) X has a subargument whose
depth in X is d + 1, and so by 8E4.1 X has one whose depth in X is d. List all
such subarguments (without repetitions); say they are W1,..., W (r > 1). Clearly
W1,..., Wr are disjoint components since they all have the same depth d in X; and
since Depth(X) = d + 1 we have Depth(Wi) < 1 for each i. Also since X satisfies
(la) and (lb) relative to d + 1, no Wi can be a meta-variable; hence each Wi must
have form
(5) Wi = Axi,1 ... xi,m^ - Yi Vi,l ... Vj,nr (mi, ni > 0).
Let X' be the result of replacing each W i in X by a distinct new meta-variable V,
with the same type as W-.
Then X' is a nf-scheme. (It satisfies restriction (iii) in Definition 8C1 because each
Wi is a subargument.) And it is clearly long and closed and has depth d. Also, by its
construction X' satisfies condition (lb) for membership of L*(T,d). It also satisfies
(la), because if it contained a meta-variable-occurrence V at a depth < d such a V
could not be a Vi and hence would occur also in X at a depth < d, contrary to the
assumption that X satisfies (la) relative to d + 1.
Hence X' E L*(T,d). Therefore by the induction hypothesis there is an X" E
_%/(T,< d) that is identical to X' except perhaps for alphabetic variation of meta-
variables. Apply Step d + 1 of Algorithm 8C6 to each Vi in X". Since Wi has form
(5) and is part of a closed nf-scheme, namely X, it is easy to see that Wi is a suitable
replacement for Vi (modulo renamings in Wi). Hence the algorithm will give X as
an extension of X", so X E W(T, < d + 1), giving (2) for d + 1.
134 8 Counting a type's inhabitants
For (3): let M E Long(T,d + 1). Then by 8E3.1 M has a subargument whose
depth in M is d + 1. List all such subarguments without repetitions; say they are
Ul,..., U (r >_ 1). Clearly El__ U are disjoint; and since Depth(M) = d+ 1 each
Ui must have depth 0 and hence must have form
(6) Ui = 2xi,l ... xi,m,.yi (mi > 0).
Define M' to be the result of replacing each U. in M by a distinct new meta-variable
V. with the same type as U. Then M' is a genuine nf-scheme. And it is clearly long
and closed and has depth d + 1. Also M' satisfies condition (lb) for membership of
L* (T, d + 1), and satisfies (1a) because all its meta-variables have been introduced by
the above replacements. Hence M' E L*(T, d + 1).
Therefore by (2) for d + 1 there is an M", differing from M' only by renaming
meta-variables, such that
M" E d + 1).

Apply Step d + 1 of Algorithm 8C6 to M": it is easy to check that U. is a suitable

replacement for Vi in M", so the algorithm will give M as an extension of M".
Hence M E 4(T,< d + 2), proving (3) for d + 1.

8F2 Detailed Stretching Lemma (cf. 8D2.) If Long (T) has a member Mt with depth
>- IITII then
(i) there exists M*t E Long(T) with Depth (M*t) > Depth(MT) + 1,
(ii) Long (T) is infinite.

Proof [Ben-Yelles 1979 §§3.20-3.21.] We shall prove (i) and then (ii) will follow by
repetition. Types of typed terms will be omitted.
To prove (i), let M be a typed closed long fl-nf with type T and without bound-
variable clashes. Let d = Depth(M) >- IITII > 1. By 8E4.1, M has at least one
argument-branch with length d. Choose any such branch, call it
(1) (No,...,Nd),
where No M and Ni+l is an argument of Ni for i = 0, ... , d -1. Each of No, ... , Nd
must have form
(2) Ni = 2xi,l ... xi,m,.yiPi,l ... Pi,n, (mi, ni >- 0),
and for i < d - 1 we have Ni+1 = Pit, for some ki < ni (and hence ni > 1). Since
d = Depth(M) the last member of (1) has no arguments, so nd = 0.
For i = 0, ... , d let Bi be the body of Ni ; that is
(3) Bi = r
-Pill ...Pin..
Since Ni is long, the type of Bi is an atom. And this atom occurs in T by 2B3(i). But
IITII < d and there are d + 1 components B3,...,Bd, so at least two of these must
have the same type.
Choose any pair (p, p + r) such that BP and BP+, have the same type (and r > 1).
Note that BP properly contains Bp+r and
Depth(Bp) >- r + Depth(Bp+r).
 8F Stretching, shrinking and completeness 135

Define M* to be the result of replacing Bp+r in M by a copy of B,, (after changing

bound variables in this copy to avoid clashes). Then M* has an argument-branch
with length d + r. (Its members are
No,...,Np+rlNpp+1,...,N°,
where for i = 0,..., p + r each N` has the same position in M* as Ni had in M, and
for j = p + 1, ... , d we have N = Np) Hence by 8E4.1,
Depth(M*) >d + r >d + 1.
To complete the proof of (i) it only remains to show that M* is a genuine typed
term. This will be done by applying Lemma 5B2.1(ii) on replacement in typed terms.
First, for i = 0, ... , d let F be the context that assigns to the initial abstractors of
Ni the types they have in M. Since M has no bound-variable clashes the variables
in I'o,...,rd are all distinct, so
(4) 1'o U ... U I'd is consistent.
Also every variable free in B; is bound in one of No_., Ni because M is closed and
Bi is in Ni. Hence, by the definitions of typed term (5A1) and Con (5A4),
(5) B1 E TT(FO U ... U ri), Con(B;) fo U ... U Fi.
To apply 5B2.1(ii) it is enough to show that the set
(6) Con(Bp) U Con(M) U FO U ... U Fp+r
is consistent. (The abstractors in M whose scopes contain Bp+r are exactly the initial
abstractors of N0....,Np+r.) But M is closed, so Con(M) = 0. And by (5),
Con(Bp) s roU...Ufp s foU...UFp+r.
Thus (6) is consistent by (4).

8F3 Detailed Shrinking Lemma (cf. 8D3.) If Long(T) has a member Mt with depth
>- D(T) then
M*T
(i) it has a member with

Depth(Mt) - IITII <_ Depth(M*t) < Depth(MT),

(ii) it has a member NT with

D(T) - IITII <- Depth(Nt) < D(T).

Proof [cf. Ben-Yelles 1979 Thm. 3.28.] Part (ii) is proved by repeating (i) and taking
the first output with depth < D(-r).
Part (i) is proved as follows. (Types of typed terms will be omitted.) Let M be
a member of Long(T) without bound-variable clashes. Let d = Depth(M) > D(T).
Incidentally D(T) >- 2 because T is composite since atomic types have no inhabitants.
By 8E4.1, M has at least one argument-branch with length d; to reduce the depth
of M we must shrink all these branches. Consider any such branch; just as in the
proof of the stretching lemma 8F2 it has form
(1) (No,...,Nd),
136 8 Counting a type's inhabitants
where N M and Ni+1 is an argument of Ni for i = 0, ... , d - 1. And
(2) Ni = )xi,1 ... xi,m, *yiPi,I ... Pi,,,, (mi, ni > 0)
for 0 < i < d - 1. Let the type of Ni be
(3) pi = Pi,l --+ --+Pi,mr--+ai

Then since Ni is long, the types of Xi,1,xi,2,... are exactly pi,1,pi,2.... ; that is, using
the "IAT" notation introduced in 8E6,
IAT(Ni) =
Just as in the proof of 8F2 let B be the body of N. for i = 0,...,d. Then the type
of Bi is the tail of the type of NO namely ai.
Define a sequence of integers do,d1,... thus: do = 0 and d3+l is the least i > dj
such that IAT(Ni) differs from all of
(4) IAT(Ndo), IAT(Nd,),..., IAT(Nd1).
Let n be the greatest integer such that d is defined. The branch (1) has only d
members after No, so n< d and d < d. Then
(5) 0=do <d1 <... <dn <d,
and for 0< i< d, IAT(Ni) is identical to one of
(6) IAT(Nd0), IAT(Nd1), ...,
Also the n+1 IAT's in (6) are distinct, and by 8E7 they are either empty or members
of NSS (T). Hence by 9E9.3(ii),
(7) n + 1 <- ITI.
Now do,. .. , do partition the set {0, 1.....d} into the following n+ 1 non-empty sets
which will be called IAT-intervals:
Oj ={dj,dj +1,...,dj+1-1} (0<j<n-1),
On = {dn,dn+ 1,...,d}.

If lj contains two numbers p, p + r such that r > 1 and B,, and Bp+r have the same
type (i.e. ap = ap+r) we shall call (p, p + r) a tail-repetition. It will be called minimal
iff there is no other tail-repetition (p',q') with
p<p'<q'<p+r.
Now an Bj that contains no tail-repetitions must have < ITI1 members. Because
for such an lj the atoms
ad,...,adi+,-1

must all be distinct, and by (3) each ai occurs in pi, which occurs in T by 8E7, and
there are only IITII distinct atoms in T. This argument also shows that for a minimal
tail-repetition (p, p + r) we have
(8) r < IITII.
 8F Stretching, shrinking and completeness 137

Now there are n + 1 IAT-intervals in the given branch and n + 1 < ITI by (7), so if
no interval contained a tail-repetition the branch would have < ITI X IITII members.
But the branch has d + 1 members and

d + 1 = Depth(M) + 1 >_ ®(T) + 1 > ITI X IITII.

Hence at least one IAT-interval contains a tail-repetition.

We start to build M* as follows. In the given branch take the last Oj containing
a tail-repetition, choose a minimal tail-repetition (p, p + r) in it, and change M to a
new term M' by replacing BP by Bp+r.
To see that M' is a genuine typed term we apply 5B2.1(ii) similarly to the proof
of 8F2. Using the notation of (4) in that proof, it is enough to show that the set
(9) Con(Bp+r) U Con(M) U ro u ... U rp
is consistent. But, as in the proof of 8F2, Con(M) = 0 and Con(Bp+r) is a subset of
ro U ... U rp+r, so (9) is consistent by (4) in the proof of 8F2.
It is straightforward to prove that M' is a long /3-nf with the same type as M.
Also that IM'I < IMI.
But M' might not be closed, because the change from M to M' has removed the
initial abstractors of Np+1,...,Np+r from M, and so some free variable-occurrences
in Bp+r that were bound in M might now be free in M'. To close M', apply the
following procedure to every such variable-occurrence.
Let v be free in the occurrence of Bp+r in M' that has replaced BP in M, and let v
be also free in M'. Then v does not occur in a covering abstractor of this occurrence
of Bp+r in M. But these covering abstractors are exactly the initial abstractors of
No_., NP in M, so
v IA(N) U ... U IA(NP).

But M is closed, so the free v in B p+r in M must be in the scope of a .iv in one of
IA(No),... , IA(N p+r). Hence v occurs in IA(Nh) for some h with p + 1 < h < p + r;
in the notation of (2) we have
V ° Xh,k
for some k < mh. And the type of v is Ph,k E IAT(Nh). Now by the definition of
do,..., d,,, and the fact that the tail-repetition (p, p + r) we are eliminating is in an
interval Oj,IAT(Nh) coincides with one of

IAT(Ndo),..., IAT(4);
say IAT(Nh) = IAT(Ndq) for some q < j. Hence there is a variable

Xdq,k E IA(Ndq)

with the same type as v. Replace v by this variable throughout M'. The result will
be a long fl-nf with the same type and depth as M' and containing one less free
variable.
138 8 Counting a type's inhabitants
Similarly replace every variable of BP+, that is free in M' by a new one which has
the same type but is bound in M'. The result will be a long $-nf M" with the same
type and depth as M' and which is closed.'
Now M" has been obtained by removing a type-repetition from an argument-
branch in M which originally contained d subarguments. And by (8) the number of
subarguments removed is < IITII. Hence
(10) d - IITII <- Depth(M") < d.
If Depth(M") < d; define M* - M". If not, select a branch in M" with length
d and apply the above removal procedure to it, then continue shortening branches
with length d until there are none left. (This process must terminate because each
removal strictly reduces IMI.) Define M* to be the first term produced by this
procedure whose depth is less than d. Then

d - IITII <-Depth(M*) <d

as required.

8F3.1 Example Let T = (a-*a)-*a-+a, and let Mr be a typed version of the Church
numeral for the number four, namely
M, =
Then
IITI I = 1, ITI = 4, D(T) = ITI X IITII = 4.

And Depth(M) = 4, so the above shrinking procedure can be applied to M. There

is only one argument-branch in M containing four subarguments, and its members
are
u3v, u2v, uv, V.

(Call them No_., NN4 respectively.) And

IAT (No) = (a-*a, a)
IAT(Ni) = IAT(N2) = IAT(N3) = IAT(N4) = 0.
Thus the only change in IAT(N;) comes at i = 1, so in the notation of the proof of
8F3 we have

n = 1, do = 0, di = 1, to = {0}, l I _ { 1, 2, 3, 4}.

There are three minimal tail-repetitions in l and the last one is (3, 4). The proof of
8F3 removes this by replacing uv by v_; this changes M to

M*

And Depth(M*) = 3 < D(T).

No claim is made that M" is related to M' by a-conversion or any other semantically significant
relation!
 8F Stretching, shrinking and completeness 139

8F3.2 Warning As mentioned in 8D10(iii) the proof of the shrinking lemma does
not necessarily apply to restricted systems of A-terms, for example the Al-calculus.
In fact there is no guarantee that if we shrink a Al-term the result will still be a
AI-term, since shrinking may cut out some variables.
 9

Technical details

To avoid interrupting the main lines of thought in the earlier chapters some concepts
were defined there only in outline and their main properties were stated without
proof. This chapter gives the full definitions and proofs. It should be read only
as required to follow the arguments in the other chapters. Its sections are largely
independent of each other.

9A The structure of a term

From the viewpoint of logical order this section is best read between 1A3 and 1A4.
As remarked in 1A4, a subterm of a term may have more than one occurrence.
The present section introduces a precise notation to distinguish such different occur-
rences; it is rather clumsy and the reader should avoid using it whenever possible,
but in some proofs its precision will be vital. The first step is to define a set of ex-
pressions called positions that can be assigned to different occurrences of a subterm
to show where they occur.

9A1 Definition (Positions) A position p = ii ... i is any finite (perhaps empty) string
of symbols such that i1,..., irri_1 are integers and in, is either an integer or an asterisk,
*. Its length is m, and if m = 0 we say p = 0.
If m 1 and n, = 1 we call p a function position;
if m 1 and i,,, = 2 we call p an argument position;
if m > 1 and ,n = 0 we call p a body position; and
if m 1 and n, = * we call p an abstractor position.
(Positions containing integers >_ 3 will be used later but not in the present section.)
The concatenation pq of positions p = i1 ... i, and q = j1... j is defined thus:
p0 = p, Oq = q, and if m, n > 1 and in * define
pq = i1 ... imjl ... jn.
(pq is undefined if m,n >_ 1 and in, _ *.)
A refinement of p is any position with form pq; it is proper if q * 0.
Two positions p = i1 ... in and q = j1 ... j, are said to diverge if neither is a
refinement of the other, i.e. if there exists h such that
ih # jh, 1 < h < Min {m, n}.

140
 9A The structure of a term 141

9A2 Definition (Occurs, occurrence) The phrase "P occurs in M at position p" (or
"M contains P at position p") is defined by induction on the length of p, thus:

(i) M occurs in M at position 0;

(ii) if Pi P2 occurs in M at p, then P, occurs in M at pi for i = 1,2;
(iii) if occurs in M at p, then Q occurs in M at p0 and x occurs in M at p*.
An occurrence of P in M is a triple (P, p, M) such that P occurs in M at p. If
P - x and the last symbol of p is * we call (x, p, M) a binding occurrence of x or an
occurrence of ).x.
An abstractor is an occurrence of Ax for some x.
A subterm of M is any term P that occurs in M.

9A2.1 Note It is easy to prove that at most one term can occur in M at one position.
(We assume of course that a term cannot be simultaneously an application and an
abstraction, or a composite term and an atom; also that

P1P2 Q1Q2 P1 Qi & P2 Q2,

x = y & P = Q.)

9A2.2 Notation An occurrence (P, p, M) may be called simply "P" when no confu-
sion is likely. A binding occurence of x may be called either of

9A3 Definition (Components) A component of a term M is any occurrence in M

other than a binding occurrence of a variable, i.e. any occurrence (P, p, M) such
that the last symbol in p is not *.
A component is called proper if it is not (M, 0, M).

9A3.1 Note The reason that binding occurrences of variables are denied the name
"components" in the above definition is that they play a very different role from
other occurrences. In a term such as we shall think of x and y as being the
material from which the term is built, but think of z as being part of one of the
operations which do the building.

9A4 Definition (The construction-tree of a term) We can display the structure of a

term as a tree with each node carrying two labels, a position and a subterm that
occurs at that position. This tree is defined for an arbitrary term M as follows.
(i) If M = x, its tree is a single node labelled with x and the empty position, thus:

x.0.
(ii) If M = PQ, its tree is obtained by first concatenating "1" onto the left end of
each position-label in the tree for P, then concatenating "2" onto the left of each
142 9 Technical details
position-label in the tree for Q, and then placing an extra node beneath the two
modified trees, as shown below.

Modified tree....* Modified tree .

for P for Q

PQ 0
(iii) If M =_ Ax-P, its tree is obtained by first concatenating "0" onto the left end
of each position-label in the tree for P, and then placing an extra node beneath the
modified tree, as shown below.

Modified tree .
for P

X x.P 0
9A4.1 Example The construction-tree of is shown in Fig. 9A4.1a.

(Xx.yx)(X z.x(yx)) 0
Fig. 9A4.1a. Construction-tree of

9A4.2 Note Each component has a node on the construction-tree; but binding
occurrences of variables are not components and do not have their own nodes.

9A5 Definition (Relations between components) Let P = (P, p, M) and R =_ (R, r, M)

be components of a term M:
(i) if r is a refinement of p we may say any of

P contains R R is a part of P R is in P;
 9A The structure of a term 143

(ii) if r is a proper refinement of p we may say either of

P properly contains R R is a proper part of P ;
(iii) if p and r diverge we may say either of
P is disjoint from R P does not overlap R;

(iv) if R - P (i.e. R, P are occurrences of the same term) we may say

R is isomorphic to P R = P.

9A5.1 Notes (i) If P is not disjoint from R it is easy to see that either P is in R
or R is in P. Hence if P, .... P, are components of a term M and no two are
disjoint they must be linearly ordered by the relation "is in"; i.e. there must exist a
permutation (i1....,in) of (1,...,n) such that
P 1
is in P1-2 in P1.3 in ... in P. .

(ii) If P is a proper part of R then P # R.

9A6 Definition (Replacement in a term) Let P - (P, p, M) be a component of a

term M. The result of replacing P by a term T will be called
{T/P}pM.
Its detailed definition proceeds by induction on the length of p, thus:
(i) if p - 0: then P - M; define
{T/M}pM - T;
(ii) if p ql: then M contains a term PQ at position q; define
{T/P}q1M - {(TQ)/(PQ)}qM;
(iii) if p q2: then M contains a term QP at q; define
{T/P}q2M - {(QT)/(QP)}qM;
(iv) if p q0: then M contains a term at q; define
{T/P}qoM -
(The case p - q* is excluded above because then P would be a binding occurrence
of a variable and not a component; {T/P}pM is undefined in this case.)

9A6.1 Example Let M - P - (yx,10,M), T - xxyz. Then

{T/P},oM -
9A6.2 Note In substitution we change bound variables to avoid clashes but in
replacement we do not.
144 9 Technical details
9A6.3 Note Of course replacing one component of a term can disturb or destroy
other components. To see what happens in two important cases, let P - (P, p, M)
and R - (R, r, M) and let
M* _ {T/P}PM.
(i) If P is disjoint from R it is not hard to show that M* contains an occurrence
of R at position r (i.e., roughly speaking, replacing P does not change its disjoint
neighbours.)
(ii) If P is in R, then p - rq for some position q such that P occurs in R at
position q, and it is not hard to show that M* contains, at position r, the term
{T/P}9R.

9B Residuals
This section summarises some properties of #-contractions needed in the proof of
the weak normalisation theorem in 5C. The full theory of #-reduction is in fact
quite deep (see Barendregt 1984 Chapters 3 and 11-14), but none of it is used in
this book except the following few very basic ideas.
Everything in this section is valid for both the untyped terms of Chapter 1 and
the typed terms of Chapter 5.

9131 Notation Recall the definitions of ji-redex and #-contraction in 1B1. (In this
section "#" will usually be omitted.) A redex-occurrence
R_
is a particular occurrence of a redex in a term. The notations function part and
argument part of R will be used here for and N respectively.
Recall from 1132 that a reduction is a finite or infinite sequence of contractions
(P1, R1, Q1), (P2, R2, Q2)....
where Pl _a P and Ql =a Pi+1 for i = 1,2,.... This definition allows a reduction
to make a-conversions before or after each of its contractions, but the reader may
safely ignore these and concentrate on the contractions; the next lemma will say
contractions are unaffected by a-conversions in a certain precise sense.
Recall from 1B3 that the length of a reduction is the number of its contractions.
(And a-conversions are not counted.)
In this section "c" will denote an arbitrary contraction and "r" an arbitrary
reduction.
If rl is a finite reduction from a term P to a term Q and r2 is a reduction of 0,
the reduction consisting of rl followed by r2 will be called
rl + r2.

9B1.1 Lemma (a-invariance) If P =a P' and P contains a /3-redex-occurrence R _

(R, p, P) whose contraction changes P to Q, then P' contains a /3-redex-occurrence
(R', p, P') whose contraction changes P' to a term Q' _a Q.
 9B Residuals 145

Proof-note Two cases must be considered: (i) P' comes from P by replacing
by Ay [y/x]M, (ii) P comes from P by replacing Ay [y/x]M by Ax M (where
y FV(M)). The proof is boring but can be slightly simplified by restricting rule
(a) as follows (cf. HS 86 Remark 1.21):

(ao) replace Ax M by Ay [y/x] M if y does not occur at all in M.

(It is straightforward to prove that every re-writing by (a) can be done by a series of
re-writings by (ao).) The reverse of an (oco)-replacement is also an (oco)-replacement,
so case (ii) above becomes redundant.

9B1.2 Warning The above lemma does not claim that R' -a R. In fact the following
is an example where this fails:
P - Ax-R, R =_ P' _
The next definition will describe what happens to a redex-occurrence S in a term
P when another redex-occurrence R in P is contracted.

9B2 Definition (Residuals) Let R - (R, p, P) and S - (S, q, P) be redex-occurrences

in a term P, say
R- S-
Let P change to P' by contracting R. The residuals of S with respect to R are certain
redex-occurrences in P defined as follows (from Curry and Feys 1958 §4B1).
Case 1: R does not overlap S. Then contracting R leaves S unchanged; more
precisely, by 9A6.3(i), P' contains an occurrence of S at position q; we call this
occurrence the residual of S.
Case 2: R S. Then contracting R is the same as contracting S; we say S has no
residual.
Case 3: R S and R is in S. Then R must be in H or in L and contracting R
will change (Ay-H)L to a component (Ay-H)L or of P. We call this the
residual of S.
Case 4: R =# S and S is in R. Then S must be in M or in N.
Subcase 4a: S is in M. The contraction of R changes to [N/x]M and
changes S to a redex-occurrence S' with one of the forms
[N/x]S [Nix][zj/yl] ... [zn/yn]S, S.

(The second form will apply if some changes of bound variables are needed during
the substitution [N/x]M; the third if S is in the scope of an occurrence of Ax in
M.) We call this S' the residual of S.
Subcase 4b: S is in N. Let M contain k >_ 0 free occurrences of x. The contrac-
tion of R changes to [N/x]M which contains k substituted occurrences of
N, each containing an occurrence of S. We call these occurrences of S the residuals
of S.
146 9 Technical details
9132.1 Notation The set of all the residuals of S with respect to R will be called
"S/R" (or "S/1;" if is the contraction (P,R,P')).
A redex-occurrence S will be called the parent of its residuals.

9B2.2 Note (i) If S is not a part of R then SIR has exactly one member (because
then Case 2 and Subcase 4b cannot happen).
(ii) Each member of SIR is an occurrence of a term where z y except
possibly in Case 4a if changes of bound variables are needed, and H' and L' are
either H and L or terms obtained from them by substitutions.
(iii) The definition of residuals is meaningful for the typed-term system in Chap-
ter 5 as well as for untyped terms. If P is a typed term then by (ii) each residual
of S in P' has the same type as S. Further, its function-part has the
same type as the function-part of S. This fact is needed in the proof of the
weak normalisation theorem in 5C.

9B3 Definition Let r be a reduction from P to Q and let S be a redex-occurrence

in P. The set S/r of the residuals of S with respect to r is a set of redex-occurrences
in Q defined thus:
(i) If r has no contractions and no a-conversions (so P = Q), define
S/r = {S}.
(ii) If r = r° + l; where u°° reduces P to a term Q° and l; is a contraction reducing
Q° to Q, and S/r° = {51,...,5}, define
s/(r° + ) = S, /l: u S2/ u ... u S
(iii) If r ends with an a-conversion Q' Q and the residuals of S in Q' are S1,...,S ,
define the residuals of S in Q to be the redex-occurrences in Q with the same
positions as 51,...,Sn in Q'. (These exist by 9B1.1.)
Finally, if _ {51,...,5,} is a set of redex-occurrences in P, define
S/r = SI/rU...US,/v.
9B4 Definition (M.C.D.'s) Let S = {S,, .. , be a set of redex-occurrences in a
term P. A development of S is a finite or infinite reduction b1 + 2 + s +... (with
a-conversions allowed), such that l;1 E S and for n >_ 1,
n+1 E
A complete development of S is a finite development 1;1 + + n such that
SA( l + ... + Sn) = 0-
A minimal member of S is one that does not contain another member. A minimal
complete development (m.c.d.) of S is a complete development 1:1 +... + l;,, such
that 1 is minimal in S and 1j+1 is minimal in S/(gl +... + ;) for i = 1, ... , n - 1
(and the only a-conversions are at the end of i;0-
If P changes to Q by an m.c.d. of some set of redex-occurrences in P we shall say
P >mcd Q.
 9B Residuals 147

9134.1 Lemma Every set S of redex-occurrences in a term has an m.c.d..

Proof Let S = Since S is finite it has at least one minimal member;

choose one, say Si, and contract it; by 9B2.2(i) each S, (i * j) will have just one
residual, so S/Sj will have only n - 1 members. By continuing this process we must
eventually reach a term containing no residuals of S.

9B4.2 Notes (i) If S is empty an m.c.d. of S has no contractions.

(ii) A single contraction is an m.c.d. of a one-member set.
(iii) The main properties of developments are described in Barendregt 1984 §11.3.
In particular it is shown there that infinite developments are impossible even in the
untyped A-calculus. Also the concept of m.c.d. is used in HS86 Appendix 1 as the
basis of a proof of the Church-Rosser theorem.

9B5 Definition (Newly created redex-occurrences) If P changes to P' by contracting

a redex-occurrence R and P' contains a redex-occurrence S that is not a residual of
a redex-occurrence in P, we say S has been newly created by the contraction of R.

9135.1 Discussion The proof of the weak normalisation theorem in 5C depends on

knowing how a newly created redex-occurrence can be produced, so the possible
ways of doing this will be analysed here. Let R be a redex-occurrence at a position
p in a term P, say
R-
Let P change to P' by contracting R; then at position p in P' there is an occurrence
of

[N/x] M.
Let S be any newly created redex-occurrence in P'; say
S =_

Now S must overlap [N/x]M, otherwise it would be the residual of a redex-

occurrence in P that did not overlap R. Hence either S properly contains [N/x]M
or S is in [N/x]M.

Case 1: S properly contains [N/x]M. If [N/x]M were in H or L then S would

not be new; hence [N/x]M = Thus S = ([N/x]M)L and P must contain
RL. Now the only way to obtain [N/x] M - Az H is either to have M - x and
N - Az H or to have M =_ Ay H' for some y and H'.
Subcase 1a: M - x, N = Az H. Then R =- (Ax x)N and the contraction of R
proceeds thus:
RL - frp NL - (Az-H)L - S.
Subcase lb: M - AY-H. Then R - (Ax (Ay H'))N and the contraction of R
proceeds thus:
RL - Np S.
148 9 Technical details
Case 2: S is in [N/x]M. Let x1,...,Xk be the free occurrences of x in M in P,
and let Nl,...,Nk be the corresponding occurrences of N in [N/x]M in P. If
S were in an N; or were disjoint from all of N1,...,N , it is easy to see that S
would not be new. Hence at least one N. must be in S. If each such N- was in H
or L S would be a residual of a redex-occurrence in M so one N, must coincide
with Az H. This can only happen if M contains a component with form xL' for
some L', and
N- S - [N/x](xL'), L - [N/x]L'.
Conclusion: in each of the above cases the function-part of the newly created
redex S is an occurrence of N or of [N/x]M.

9C The structure of a TAA-deduction

This section should be read in parallel with the outline definition of "TA2-deduction"
in 2A8. That outline is filled in here and some lemmas are proved for use in
Chapters 5 and 8.
A TAx-deduction as defined below will be a slightly more elaborate object than the
trees shown in the examples in Chapter 2, and each deduction-tree in that chapter
should be regarded as an abbreviation for a deduction as defined below.

9C1 Definition (TAx-deductions) TAx-deductions are labelled trees defined as fol-

lows, each node carrying two labels, a TAA-formula and a position.
(i) An atomic deduction is a single node labelled with an axiom and the empty
position, thus:
x:T r-* x:r 0.

(ii) If Ai and A2 are deductions whose bottom nodes are labelled by the position
0 and the formulae, respectively,
rl ra P:(v-.r), F2 F-., Q:v,
and F1 U F2 is consistent, then a new deduction called (A1A2) is constructed by first
putting "I" on the left end of each position-label in At, then putting "2" on the left
end of each position-label in A2, and then placing an extra node beneath the two
modified deductions, as shown below.

Modified Modified
2

17I H P: (T-)i 2HQ 2

i U r2 H PQ: t 0
(iii) If Al is a deduction whose bottom node is labelled by the position 0 and the
formula F H P :T, and r is consistent with x:a, then a new deduction called (.1x01)
 9C The structure of a TAR-deduction 149

is constructed by first putting "0" on the left end of each position-label in At and
then placing an extra node beneath the modified deduction, as shown below.

9C1.1 Example For examples of deductions see 2A8.2-5 and the answers to Exer-
cises 2A8.7-8. (The deductions in those examples omit all position-labels and are
displayed in the standard space-saving way using horizontal lines instead of the
node-and-line way shown in the diagrams above, but this is merely a matter of
representation and is not intended to conflict with the definition.)

9C1.2 Note If we remove contexts and types from all the TAR-formulae in a
deduction-tree whose conclusion is IF --> M:r, it will be transformed into the
construction-tree of M as defined in 9A4.

9C2 Definition The length, 1Al, of a TAR-deduction A is defined thus: JAI = 1 for
atomic A, and for composite A

IAAA21 = lAi l + IA2l, I2x-Aj l =1 + JAI 1.

Clearly JAI = IMI if the conclusion of A is F F--* M:i. (For IMI see 1A2.)

9C3 Definition The set of all type-variables occurring in a deduction A will be called

Vars(A).'

9C4 Discussion (Replacement in a deduction) Let M contain a component P =

(P, p, M) and let A be a TAR-deduction of

r H M:r.
Then from 9C1 it is easy to see that A contains a deduction Ap of a formula
(1) Fp H P :Up
for some Fp and op.
It is natural to expect that replacing P by a new term T with the same type
will leave the type of M unchanged. But the types of P, T and M all depend on
contexts: F for M, ['p for P, and an unspecified one for T. So our expectation is
imprecise. It will be made precise and proved in the two replacement lemmas below.

Warning: deductions 0 contain term-variables as well as type-variables but only the latter are included
in Vars(L).
150 9 Technical details
The first step towards precision is to clarify the relationship between rp and F.
If P is not in the body of a A-abstract in M then clearly
(2) rpgr.
But now suppose P is in the bodies of some A-abstracts in M; say there are n
distinct such abstracts:
1x,Nt, ... , -1x :Nn,

with P in N; for i = 1, ..., n. By 9A5.1 these components must be linearly ordered

by the relation "is in"; by renaming them if necessary, we can assume that
P is in Ni in 2x1Nl in N2 in 2x2N2 ... in Nn in Ax --N...

By the subject-construction theorem (2B2) i must contain n applications of (-1)

below (1), and each of these has one of the following forms, depending on whether
xi E FV(Ni) or not:
Ni:gi Fi H Ni:q,
(3)
Fi - ri H
(Here 1 < i< n.) Hence
(4) Fp F U {xl
If M has no bound-variable clashes then xl,... , xn are distinct and the right-hand
side of (4) is consistent. (But if M has bound-variable clashes an might be
inconsistent with an (j * i) or even with a member of F.)
The replacement lemmas will now be stated. The first one is easy and is used in
proving the subject-reduction theorem (2C1). The second is more complex to state
but gives a stronger result; it is applied in 5B2.1(ii) which is used in 8F2.

9C5 First Replacement Lemma for Deductions Let 0 be a TAA-deduction of the

formula r F-> M:T and let M contain a component P with a position p. Then A
contains a deduction Op of a formula with form
rp ' -> P :ap

for some rp and vp; let {T/P}PM be the result of replacing P at p by a term T such
that
FT FA T :Qp
for some FT c rp. Then
F 1-1 {T/P}pM:i.

Proof Let r;. FT P T. Then Subjects(FT) = FV(T) and by 2A11(i) there is a

deduction AT of
FT H T :up.
Replace Ap by AT in A and modify the contexts and subjects in all the formulae
 9D The structure of a type 151

below Ap accordingly. Since FT s Fp none of these modifications enlarges a

context, so no inconsistencies are introduced. Therefore
F H {T/P}pM:T.

9C6 Second Replacement Lemma for Deductions Let A, M, P be as in the first

replacement lemma, and let {T /P }PM be the result of replacing P at p by a term T
such that
(i) rT 1-A T :ap
for some FT such that
(ii) rT U IT U KK
is consistent,
where xl,... , xn (n 0) are the binding variables of the 2-abstracts in M whose bodies
contain P, and l;l,...,l;n are their types in A (see 9C4 for details). Then
(iii) F U FT [-a {T/P}PM:T.

Proof Just as in the previous lemma, by 2A11(i) we can assume Subjects(FT) =

FV(T). Replace Ap by AT in A and modify the contexts and subjects in all the
formulae below Op accordingly. No inconsistencies will be introduced, because by
9C4(4) the only variables not in F that could cause problems are x l, ... , x, .

9C6.1 Corollary The conclusion of Lemma 9C6 also holds if instead of 9C6(ii) we
assume that FU FT is consistent and none of X1.... , xn occurs free in T.

Proof If Subjects(FT) = FV(T) the above assumption implies that IT satisfies 9C6
(ii).

9D The structure of a type

This section and the next give two alternative approaches to the structure of an
arbitrary type. The approach in the next section is used in Chapter 8; that in this
one is more basic and is included mainly as a contrast and introduction to the other
approach.
From the viewpoint of logical order the present section fits between 2A2 and 2A3.

9D1 Definition (Occurs, occurrence) Here positions are the same as in 9A1; however,
positions containing * are not needed in the present section. The phrase "a occurs
in z at position p" is defined thus (cf. 9A2):
(i) T occurs in T at position 0;
(ii) if al --.v2 occurs in T at p, then a; occurs in T at pi (i = 1, 2).
A triple (a, p, T) such that a occurs in T at p is called an occurrence of a in T, or a
component of T. A type that occurs in T is called a subtype of T.

9D1.1 Notation An occurrence (a, p,T) may be called simply "a" when no confusion
is likely.
Recall from 2A2 that the number of occurrences of variables in T is called DTI and
152 9 Technical details
the set of these variables is called Vars(T). The set of all variables occurring in a
finite sequence (ti,. .. , will be called

Vars (T 1, ... , T,,).

9D2 Definition (Construction-tree of a type) The construction-tree of a type T is a

labelled tree showing how T is built up (cf. the tree of a term in 9A4). Each node
has two labels, a type and a position, and the tree is defined thus:
(i) If T is an atom, say T = e, its tree has just one node:
e.0.
(ii) If T =_ p-.a, its tree is built from the trees for p and a by first putting "1" on
the left end of each position-label in the tree for p, next putting "2" on the left end
of each position-label in the tree for a, and then placing an extra node beneath the
two modified trees, thus:

Modified tree....* . Modified tree....-'

for p for a

p-*a 0
9D2.1 Example Let T - (a-(b--+c))-+((a-+b)-*(a-*c)); its tree is as shown in
Fig. 9D2.1a.

0
(a-. (b-+ c))- ((a-. b)-* (ate c))

Fig. 9D2.1 a.

9D3 Definition (Positive and negative occurrences) Let a - (a, p, T) be an occurrence

of a. We call a positive (and say that a occurs positively in r) if p = 0 or the number
of 1's in p is even. We call a negative and say that a occurs negatively in T if the
number of l's in p is odd.
 9E The condensed structure of a type 153

9D3.1 Example The positive occurrences in the type r shown in Example 9D2.1 are

(T, 0, T), (a, l 1, T), (b,121, T), ((a-*b)--(a-*c), 2, r),

(a, 211, T), (a-*c, 22, T), (c, 222, T).

All other occurrences in this type T are negative.

The reader familiar with the usual concept of positive and negative occurrences
in propositional logic can easily check that the above definition agrees with it.

9E The condensed structure of a type

The present section analyses the structure of an arbitrary type in a more compact
way than the last, and computes a bound on the number of its components of a
certain kind. This bound is used in the proof of 8F3.
The starting point is a remark in 8B1 that every type T can be expressed uniquely
in the form

(1)

where m >_ 0 and e is an atom. In this section we shall view a type as having
been built up from its atoms by the operation of constructing (1) from ti...... ,,,
and e instead of the more usual operation of constructing from p and u.
Corresponding to this new view a new construction-tree of a type can be defined
that is more condensed than the usual one and in which there is no bound on the
number of branches that may start from a node. The following definitions formalize
this view.

9E1 Notation Just as in 9A1, a position is any finite (perhaps empty) string p =
it ... iof integers and *'s such that ij,..., The present section will use the
notation of 9A1, with the following exceptions:

if m 1 and in, p will be called a premise position;

if m I and p will be called a conclusion or tail position.

9E2 Definition (Condensed tree of a type) The condensed construction-tree of a type

T is defined by induction on 1Tj, thus. (Each of its nodes is labelled with a type and
a position.)
(i) If T is an atom e, its condensed tree is a single node:

(ii) If T =_ (n > 1), construct its condensed tree from the condensed
trees of T1,...,T,,, by first replacing each position-label p in the tree of T; by
154 9 Technical details
(m + 1 - i)p (for i = 1, ... , m), and then combining the modified trees as follows:

Modified Modified
tree for tree for

9E2.1 Note The use of (m + 1- i)p in (ii) above has the effect of assigning positions
to Tl,...,Tm backwards, giving position 1 to Tm, 2 to and m to TI. This
makes it easier to relate positions in the condensed tree of a--*r to those of a and
T, though we shall not need this facility in this book.
As an example, Fig. 9E2.1a shows the condensed tree for the type

T (a->(b->c))->((a-b)-(a->c))
(a-.b-->c)-+(a-*b)--a--*c.

9E3 Definition (S-subtypes) The significant subtypes or s-subtypes of a type r are

the types that label the nodes in the condensed tree of T. A proper s-subtype of T is
an s-subtype # T.

9E3.1 Example The s-subtypes of the type T in Fig. 9E2.1a are

a, b, c, a-*b-.c, a-*b, T.

Note that although b-+c is a subtype of T in the usually accepted sense (9D1) it is
not an s-subtype because it does not correspond to a node on the condensed tree.

(a_b-4c)_(a-+b)-> a-* c

Fig. 9E2.1a.
 9E The condensed structure of a type 155

9E4 Definition (S-components) Iff a node on the condensed tree of T is labelled with
a type a and a position p we call the triple (a, p, T) an s-component of -r. (Thus an
s-component is a particular occurrence of an s-subtype.)

9E4.1 Notation S-components are distinguished from s-subtypes by underlining

their names. The phrases "p contains g", "p properly contains a" etc. are defined
here as in 9A5.

9E5 Definition (Premises, tail) If p is a composite s-component of a type T and

p =_ (m >_ 1), the s-components pl,..., p,, are called the premises of
p and a is called the conclusion or tail-component of p, or just
Tail(p).

9E5.1 Lemma Two distinct s-components of a type T cannot have the same tail-
component.

Proof Induction on itt, using the fact that if T = Tl--*... the only s-
components containing a are r and a and a is not the tail of itself because atoms
do not have tails (by 9E5).

9E5.2 Warning The above lemma does not say that the tails of two distinct s-
components p and g cannot be occurrences of the same atom. That is, using
the =-notation of (iv), the lemma forbids Tail(p) = Tail(g) but does not forbid
Tail(p) = Tail(g).

9E6 Definition (Subpremises, subtails) An s-component of T is called a subpremise

or subtail of T according as it is a premise or tail of another s-component of T. The
sets of all subpremises and all subtails of T will be called, respectively,
Subpremises(T), Subtails(r).

9E6.1 Example The type T - (a--+b--+c)--+(a-+b)-+a-+c in Fig. 9E2.1a has six sub-
premises, namely all three a's and
3, T), (a-*b, 2, T), (b, 31, T).

It has three subtails, namely

(b, 2*, T), (c, 3*, T), (c, *, T)

9E6.2 Notes (i) A proper s-component g = (a, p, T) (p * 0) is a subpremise iff p is a

premise position and a subtail iff p is a tail position.
(ii) Each s-component of T is either a subtail, an atomic subpremise, a composite
subpremise, or T itself, and cannot be more than one of these.
(iii) If T is composite, its leftmost atom-occurrence is a subpremise and its
rightmost is a subtail.
(iv) An atom has no subpremises or subtails.
156 9 Technical details
9E6.3 Lemma If T is composite, then
(i) #(Subtails(T)) no. of composite s-components of r,
(ii) 1 + no. of composite sub premises ofT,
(iii) ITI - no. of atomic sub premises of T,
(iv) ITI -
(v) #(Subpremises(T)) ITI -
(vi) no. of s-components of T < 21TI-1.

Proof For (i): use 9E5.1. For (ii): each composite s-component is either a subpremise
or r itself. For (iii): use 9E6.2(ii). For (iv): use 9E6.2(iii). For (v): subtract (iii) from
(ii). For (vi): use 9E6.2(ii), adding (iv) to (v) and adding 1 for T itself.

9E7 Definition Order(T), the order of r, is 1 + the length of the longest position on
the condensed tree of T. In detail: Order(e) = 1 for atoms e, and for composite types
Order 1+Max {Order(Ti),...,Order(T)}.

9E7.1 Example Order((a->b-+c)-*(a-*b)--*a-*c) = 3.

9E8 Definition (Positive and negative s-components) An s-component a of r is called

positive or negative according as the number of non-asterisk symbols in its position
is even or odd. If v is positive we say a occurs positively in T, otherwise a occurs
negatively in T.

9E8.1 Example If r - (a-+b-.c)->(a->b)-*a-*c, see Fig. 9E2.1a, its positive s-

components are
(r, 0, r), (c, *, r), (a, 32, r), (b, 31, r), (a, 21, r).

9E8.2 Notes (i) It is straightforward to show that an s-component a is positive or

negative according as the corresponding component in the more usual sense (9D3)
is positive or negative.
(ii) A subpremise of T is positive if its position has even length. (Because the
position of a subpremise contains no *'s.)
The following set plays a role in Chapter 8.

9E9 Definition (NSS(T)) (cf. Ben-Yelles 1979 Def. 3.36.) If r is composite, NSS(T)
is the set of all finite sequences (n >: 1) such that T contains a positive
composite s-component with form

for some atom a. Each member of NSS (T) is called a negative subpremise-sequence
(because it is a sequence of terms that have occurrences as negative subpremises in
T).
The set of all the members of the sequences in NSS(T) will be called
U NSS (r).
 9F Imitating combinatory logic in A-calculus 157

9E9.1 Example If r - (a-+(b->d-+c)-+d)-+(a->b-*c)-->d-+d, we have

NSS (T) = { (a->(b->d-*c)--*d, d), (b, d) },
U NSS (T) = la-+(b-+d--+c)-+d, a-*b-*c, d, b}.

9E9.2 Notes (i) If T is composite, say T - Tl-+...--+T,,,-+e (m >_ 1), and each T; has
form
Ti - T1,1-+...--+Ti,m.-+ei (m1>0),
then

NSS(T) = {(T1,...,Tm)} U U{NSS(Tij):1 < j < mi, l < i < m}.

(ii) If T is an atom, NSS (T) = 0.
(iii) If p occurs positively in r then NSS (p) s NSS (r).

9E9.3 Lemma If r is composite, then

(i) #(NSS (T)) < no. of positive composite s-components of r,
(ii) < DTI - 1,
(iii) #(UNSS(T)) < DTI - 1,
(iv) (61, ... , 6,) E NSS (T) = n< ITI - 1.

P r o o f For (i): each (61, ... , E NSS (T) is obtained from a distinct positive
composite s-component of T. For (ii): use 9E6.3 (i) and (iv). For (iii): note that
U NSS (T) c Sub premises (T)
and use 9E6.3 (v). For (iv): use the definition of NSS(T). 0

9F Imitating combinatory logic in A-calculus

This section is used in 6D and 7D. The converse PT theorems in 7A can be seen as
completeness theorems for the rule (D) described in 7D 1; but that rule is formulated
for Hilbert-style axiom-based versions of logic not Natural Deduction versions,
and in the correspondence between logic and types the Hilbert style corresponds
more closely to combinatory logic than to A-calculus. It is possible to describe this
correspondence in a A-context but to do so we must first use A-terms to imitate a
little of standard combinatory logic technique; this is done in the present section.
Recall the following closed A-terms:
B = Axyz-x(yz), B' Axyz-y(xz), C I

K= S= W=
9F1 Definition (S-combinations) If S is a set of A-terms, an S-combination, or
applicative combination of members of 5, is a A-term built from some or all of the
members of S by application only. An S-and-variables combination is an applicative
combination of members of S and variables.
For subsets of {B, B', C, I, K, S, W} the S-combinations will be called BCK-combina-
tions, BCIW-combinations, etc.
158 9 Technical details

9F1.1 Example If S = {B, C, K} then CKK and B are S-combinations and CKx, xy
and CKK are S-and-variables combinations. But lx-BC is not an S-combination or
an S-and-variables combination.
The theory called combinatory logic was originally developed from the idea
that 2-abstraction can be imitated by building combinations of a very limited set
of operators called basic combinators. (See HS 86 Ch. 2.) In many accounts of
the theory just two basic combinators called S and K are assumed, with similar
reduction-properties to the 2-terms S and K akove. But the version below uses B, C,
K and W instead; this makes it easier to discuss particular subsystems.

9F2 Definition For 2-terms P and Q, we shall say that P fl-reduces to Q with strong
type-invariance if
(i) P>pQ,
(ii) FV(P) = FV(Q),
(iii) for all I' and i, f F-2 P :T t F-1 Q:i.
We shall say P #-converts to Q with strong type-invariance if P =p Q and (ii) and
(iii) hold.

9F3 Combinatory Completeness Theorem There is an algorithm which accepts any

A-term M and constructs a BCKW-and-variables combination M* which /f-reduces to
M with strong type-invariance.

Proof The algorithm will construct M* by induction on IMI. If M is a variable

choose M* = M.
If M - NP and N* and P* have been constructed, choose M* N*P*. (To
see that N*P* satisfies 9F2 (iii), use the induction hypothesis and the subject-
construction theorem, 2B2(iii).)
Now assume M =- 1xN and N* has been constructed. By the induction hypothesis,
(1) FV(M), >p M.
Also, by the subject-construction theorem (2B2), F 1--2 M:T if 2 = p-pa for some p
and a such that
I-,x:p F-2 N:a
and by the induction hypothesis the latter is equivalent to
(2) F,x:p F-2 N* :Q.
We cannot simply choose M* _- lxN* because this is not an applicative combination.
Instead, an applicative combination called "[x] N*" will be constructed, such that
(3) FV([x] N*) = FV(N*) - {x},
(4)
(5) I' F-2, i = p--+a and F,x:p F-2 N*:o.
Then M* will be chosen to be and 9F2 (i)-(iii) will follow from (1)-(5).
 9F Imitating combinatory logic in 2-calculus 159

Before constructing define

(6) 1* = CKK, S* B(B(BW)C)(BB).
It is routine to check that I * and S * #-reduce to I and S with strong type-invariance.
Then is constructed by the following algorithm (from Curry and Feys 1958
§6A3, omitting (c)).
(a) If x V FV(N*): KN*;
(b) if N* x:
[x].N*

I*;
(d) if N* PQ,x FV(P),x E FV(Q):
(e) if N* PQ,x E FV(P),x FV(Q):
(f) if N* PQ,x E FV(P),x E FV(Q):
It is easy to see that is defined for every applicative combination N* of B,
C, K, W and variables, and a routine induction on IN* j shows that (3)-(5) hold.

9F4 Definition A set S of typable closed 2-terms is called a typable basis for a set 1_
of 2-terms if there is an algorithm that accepts any member of L and constructs an
S-and-variables combination M* which fl-reduces to M with strong type-invariance.

9F5 Partial Completeness Theorem (i) {B, C, I, W} is a typable basis.for the set of
all 2I-terms (defined in 1D1).
(ii) {B,C,K} is a typable basis for the set of all BCK2-terms (1D2).
(iii) {B, C, I} is a typable basis for the set of all BC12-terms (1D3).

Proof (i) Modify the algorithm given in 9F3 by omitting (a) and replacing I * by I
in (b).
(ii) Modify the algorithm in 9F3 by omitting (f).
(iii) In 9F3, omit (a) and (f) and replace I* by I in (b).
 Answers to starred exercises

2A8.7 Recall that (i) B' - Axyz.y(xz); (ii) C - (iii) S - (iv)

W- the deductions are:
x:a-+b u-+ x:a-.b z:a H z:a
(-.E)
y:b-.c H y:b---+c x:a-+b, z:a H xz:b
(-.E)
x:a-+b, y:b-+c, z:a F-+ y(xz):c
(-+I) 3 times
u--+ (Axyz'y(xz)):(a->b)->(b--+c)-+a--+c

x:a-+b--+c x:a-+b-.c z:a H z:a

(--+E)
x:a-+b-+c, z:a i--+ xz:b-+c y:b -+ y:b
(-+E)
x:a-+b-+c, y:b, z:a H xzy:c
(-+1) 3 times
i--+

x:a-*b-+c x:a-+b--+c z:a z:a (-+E) y :a-+b H y :a-+b z :a '-+ z :a (--+E)

x:a-+b--+c, z:a H xz:b--+c y:a-*b, z:a i-+ yz:b
(-+E)
x:a--.b-+c, y:a-*b, z:a H xz(yz):c
(-+I) 3 times
H
x:a-*a-+b H x:a-+a--*b y:a H y:a
(--+E)
x:a-'a-+b, y:a F-+ xy:b y:a H y:a
(- *E)
x:a-+a-+b, y:a H xyy:b
(-+I) twice
+ (Axy-xyy):(a-+a-+b)--+a-+b

2A8.8 The deductions are too wide to show in full here, so the context and arrow
will be omitted from each formula. (This cannot be done for arbitrary deductions

161
162 Answers to starred exercises
without ambiguity but those below are simple enough.) First note that, by deductions
like 2A8.3,
I- I :(a-+b)---).a-+b, I- I :c--+c

(i) Let M - 2vxyz v(y(vxz)) and P - MI.

v:(a-+b)->a-+b x:a-*b
(-+E)
vx:a--+b z:a
(-E)
y :b-+a-+b vxz :b
(-E)
v:(a-.b)-*a-*b y(vxz):a-+b
(-+E)
v(y(vxz)):a--+b
(--+I) 4 times
I:(a-*b)-*a->b

(ii) Let Q -2xyz-I(y(Ixz)).

x:a-*b (-+E)
Ix :a-+b z :a
(-E)
y :b-+c Ixz :b
(-+E)
I:c-+c y(Ixz):b
(-*E)
I(ylxz)):c
(-+I) 3 times
(2xyz I(y(Ixz))):(a-+b)-+(b-+c)-*a->c

2B3.2 (T. Coquand, informal correspondence 1993.) Let F F-+ M:T be the conclusion
of A and A. Then M has a /3-nf M* by WN (2D5), and by 1B9 there is a leftmost
reduction p from M to M*. Such a reduction consists of the following successive
parts (each of which may be empty):
(1) A reduction (called a head reduction of order 0) in which each step has form
(2y P )QR1... Rk >1 ([Q/y]P)R1... Rk (k > 0).

(2) A series of reductions pl,..., pm (m >_ 0), where p, has form

where N > N' by a head reduction of order 0. (We call pi a head reduction of
order i.)
(3) A reduction (called an internal reduction) with form
D 2xl...x,n-yN1*...Nn* = M*,
where Ni P N1* by a leftmost reduction, for i = 1, ... , n.
 Answers to starred exercises 163

By the proof of the Subject-reduction theorem 2C1, A and A' reduce to deductions
of
r* H M*:T, r'* H M*:t,
for some F* s r and r'* r. (Reducing a deduction is like reducing a typed term,
5B8.) But r* = r'* = r, because FV(M*) = FV(M) since reductions of Al-terms
do not cancel, and so
Subjects(r*) = Subjects(r'*) = FV (M*) = FV(M) = Subjects(r).
Hence by 2B4, A and A' both reduce to the same deduction.
Now prove A = A' by induction on IM *I with an induction on the length of the
reduction p in both the basis and induction steps. (The case p = 0 is 2B4.)
Basis : M* - y. Then (2) and (3) are empty. For each step (1) in p, if the deduction
for the right side of (1) is unique, then the types and deductions for Q, P, R1, ... , Rk
are uniquely determined. (Note that Q occurs in [Q/y]P because y occurs in P.)
Hence the deduction for the left side is uniquely determined.
Induction step: M* - Axl ... xn,.yNi * . . . Nn*, m + n > 1. For steps (1) use the
above argument and for (2) and (3) use the hypothesis of the induction on JM*I.
3A6.1 Let B - Axyz.x(yz) and let A be any TAx-deduction of a type l; for B. By
2B2(iv) the last three steps in A must be applications of rule (-*1), with form

x:a,Y:/3,z:a --* x(yz):T

x:a, y:fi H Az x(yz) : a->t

x:a H
i--+

for some a, /3, a, r. Then by 2B2(iii) the step above these must be (-*E), and its
premises must have form
H x:P-*T, Y:$,z:a H yz:p
for some p. Hence a - p-nT. Also by 2B2(iii) the step above yz:p must be (-*E),
with premises of form
y:a--.p H y:a-'p, z:a --> z:Q.

Hence /3 - c-*p, and the type assigned to B must have form

3B5.2 (i) [a/b] o [b/a] - [a/b, a/a]. For example if t = a-*b, we have
([a/b] o [b/a])(T) a--*a, [a/b]([b/a](T)) = [alb](b-*b) = a--+a.
(Warning : [alb] o [b/a] * [a/a] ! To see this, apply [a/a] to T.)
(ii) Let s - [al /al,... , a /a ]. If Range(s) n Dom(s) = 0, define
St = [QI/al] o ... o [Un/an]
164 Answers to starred exercises
If Range(s) fl Dom(s) * 0, choose distinct b1,..., bn 0 Range(s) U Dom(s) U Vars(T),
and define c = [bi/a,]o, for each i < n; then define
ST =ext [al/bl] o o [an/bn] o [al /al] ° ° [Qn /an]-

(Warning: The dependence of si on r is real; the definition of s' includes the clause
"bl,...,bn 0 Vars(r)" and this clause is more important than it looks: no claim is
made that there is a composition s'" of single substitutions such that s'(r) - s(r)
for all T.)
3C3.1 Let p - a-+(b-+c), r (a->b)--+a, vo - By 3C2.1, vo is
a c.i. of p and T. To show vo is most general, let v be any other c.i., say

v = si(a)->(§1(b)--+si(c)) = (§2(a)-*§2(b))-+§2(a)
Then
s1(a) = §2(a)-'§2(b), §2(a) = si(b)->sl(c),
so

sl(a) _ (st(b)-->sl(c))--+s2(b),
and hence v is an instance of vo.
3D2.1 Let p - a->(b-+b), r - (c-+c)--+a, au =_ [(b->b)/a, b/c]. First, u clearly unifies
{ p, r}. To prove that u is most general, let s be any other unifier, say

s = [a/a, Q/b, y/c],

where
(Y-'Y)-a.
Then, by comparing the two sides of this identity, we have
a = Y-'Y, a = -'/
Hence $ - y and
s = [($- $)/a, /3/b, /3/c] = [/3l b] o u.

6B4.3 Respectively, (AS)A and (A6)A are

x:a F-+ x:a x:a H x:a

(-+I)vac (-'I)main
x:a .--.

7A3.2 (M. W. Bunder) Choose P - P' - (2vxyz.vy(x(vz)))I. Then

P =# P', and by an easy computation
PT(P) (a->b)-*(b->a-.b)->a->a-+b,
PT(P') ((c->d)-c)-+(c-+d)--.(c-+d)-+d.
 Answers to starred exercises 165

These two types have no common substitution-instance, because if such an instance

were obtained by substitutions [a/a, f /b] and [y/c, 5/d] we would get
a = y-*b, a--+f = 6,
which would imply the impossible identity a - y--+a--+/t.
7C7.2 (i) Let M - I and T = (b-.b)-*b-+b. Define i° - (a-+b)-+c-+d; then r° is
skeletal and changes to T when we make the identifications [b/a], [b/c], [b/d]. The
algorithm begins by applying the proof of 7C2, which begins by applying the proof
of 7C1 to build a term whose PT is T°-+T°; and the procedure in that proof gives
I, -
Then, following the proof of 7C2, the algorithm defines
M+ - IToM =_ IT-'-

It then applies the PT algorithm to compute T+ - PT(M+); in fact in this case

T+ = T° and the identifications s1,..., sk in 7C7 are [b/a], [b/c], [b/d] (and k = 3).
The next step is to apply the proof of 7C5 to build three terms N1, N2, N3 in turn,
such that
PT(NI) [b/d]T+ (a-*b)->c-*b,
PT(N2) [b/c][b/d]T+ (a-+b)->b-+b,
PT(N3) = [b/a] [b/c] [b/d]T+ (b-+b)-+b-+b.
To do this, the algorithm first applies 7C4 to obtain three terms R1, R2, R3 such that

PT(RI) = (d->.f)-->(g->b)-*((a->b)->c-*d)-+((a-*g)->c->f),
PT(R2) = (f--+c)-+(g-+bl)-+((a->bl)-+c-->b2)-'((a-*g)-*f-*b2),
PT(R3) = (a-+f)-'(g-'bl)-'((a-Bbl)-+b2-b3)-'((f-'g)-,b2-.b3).

Then the algorithm defines

NI - N2 - N3 -
Finally it defines M* - N3. (By the way, the algorithm is not claimed to be efficient!
There exists a much shorter M* than the one above, namely M*
(ii) Let M - K, ,r - b--+b--+b. Define T° - a--+b--+c; then T° has the 1-property and
changes to r under the identifications [b/a], [b/c]. The algorithm begins by applying
the proof of 7C1 to build a term whose PT is T°-+T°; this term is
IT' -
Then, following the proof of 7C2, the algorithm defines
M+-IToM-IT0K
and computes T+ - PT(M+); in this case T+ - a-+b-+a and for the identifications
sl, ... , sk in 7C7 we have k = 1 and sl - [b/a]. The next step is to apply 7C5's
proof to build N such that
PT(N) __ [b/a]T+ b-+b-+b.
166 Answers to starred exercises
To do this, the algorithm first applies 7C4 to obtain a term R such that

PT(R) = (f-*a1)-*(g-*b)->(al->b-->a2)-+(f-*g-*a2),
and then defines N - (Ax'Rxx)I(I,.K).
8A12.1 The eight regions contain the following terms in order from left to right.
Top row: Axy'xy, Axyz'x(xyyy)z, 2xyzu'x(xyyy)zu, Axyzu'xyzu;
bottom row: Ax-x, Axyz'xz(xyyy), Axyzu'xu(xyyy)z, Axyzu'uxyz.

8B7 For items 1, 6, 8, 11 in Table 8B7a see 8B3-8B6. For the rest, see the answer to
8C6.4. (In item 12, PO Nprinc(T) since the PT of Po is ((a--+a)--+b)-+b by the PT
algorithm, 3E1.)
8C6.4 For rows 6, 8 and 11 of Table 8B7a see Examples 8C6.1-3. The other rows
are dealt with below. (For ease of reading, types are omitted and x'M is used for
x(x(... (xM) ...)) with d x's.)
1. d(T,0) = {V}, ,q/(T,1) = 0.
2. &/('r,0) _ {V}, .sJI(T,1) _ {Ax1'x1}.
3. .Q/' (T,0) _ {V}, Ql(T,1) _ {Ax1x2'xl}.
4. d(T,0) _ {V}, d(T,1) _ {Ax1x2x3'x1 VI 1, .9/(T,2) = {Ax1x2x3'x1(x2V2)},
,SV(T, 3)= {Ax1x2x3'x1(x2x3)}
5. .Ql(T,0) = {V}, Q/(T,1) = {Ax1x2x3'x1 V1 V2}, JV(T,2) = {Axlx2x3'xlx3x2}.
7. crl(T,0) = {V}, d(T,1) = {Axix2'x1 V1 V2}, 4(T,2)= {Ax1x2'x1x2x2}.
9. 5z/(T,O) = {V}, .21(T, 1) = {Axix2'xl V1, Axtx2'x2}
srl(T,d) = {2x1x2'xiVd, Axlx2'xi-1x2} for all d > 2.
10. Al(T,O) = {V}, .sal(T, 1) = V1}, sl(T,2) = {Ax1x2'xlx2}.
12. d(T,0) = {V}, s.V/(T,1) = {Ax'xVl}
d(T,2) = Ax'x(Ayryi)}
d(T,3) = 2x'x(2y1'x(2y2'xV3))
2x'x(AY1'x(2Y2'Y1)), Ax'x(2Y1'x(2Y2'Y2))}
,(V(T,d) _ {2x'x(Ayl'x(... (2yd-1'xVd)...)),
Ax'x(Ay1'x(... (AYd-2'x(AYd-1'Y1))...)), ...
2x'x(Ay1'x(... (AYd-2'x(AYd-1'Yd-1))...))} for all d 4.

8E7.3 For (iii), use (ii) and the fact that #(szl(T,0)) = 1 (since d(T,0) = {VT} by
Step 0 of Algorithm 8C6).
For (i), use induction on d. The basis is trivial since d(T, 0) = {V' J.
For the induction step (d to d+ 1), let XT E d(T, d) contain q metavariables where
1 < q < ITId, and let VP be one of these.
Consider Part Hal of Step d + 1 of Algorithm 8C6: using the notation of IIal,
note that each suitable replacement YfP generated by IIal for VP contains < n1
metavariables, where n1 is the arity of a1. But o1 occurs in p which occurs in T by
8E7.1, so
ni < jail - 1 < ITI - 1 < ITI.
 Answers to starred exercises 167

Next consider Part IIa2 in 8C6. Using the notation of IIa2, note that each suitable
replacement Zj generated by IIa2 for VP contains < h metavariables, where hj is
the arity of Cj. But Cj occurs in r by 2B3(i), so
hj < K,I-1 < ITI-1 < ITI.
Thus each metavariable in XT is replaced by < ITI new ones, so the total number
of metavariables in the resulting extension of XT is < gITI. Hence (i) holds.
To prove (ii), look at the above induction step in more detail. When IIal in 8C6
is applied to VP the number of suitable replacements it generates is < m, where m
is the arity of p. We have
m< IPI-1 < ITI-1.
When IIa2 in 8C6 is applied to VP the number of suitable replacements it generates
is < t, and by 8E7.2(ii),
t < (ITI - 1)Depth(XT).
Thus the total number of suitable replacements for VP is less than or equal to
IT) - 1 +(ITI - 1)d, < ITI(d+ 1).
But there are q metavariables in XT, so when we apply IIa2 the total number of
resulting extensions of XT is less than (ITI(d + 1))9. Hence
#(d(T, d + 1)) < (ITI(d + 1))9 x #(d(T, d))
Then (ii) follows by (i).
 Bibliography

References to unpublished manuscripts have been avoided as far as possible, as

they are likely to be unavailable to most readers. A few have been included for
historical and other reasons, however.

AHO, A. V., SETHI, R., ULLMAN, J. D. [1986] Compilers, Addison-Wesley Co., USA 1986.
ANDERSON, A. R., BELNAP, N. D. [1975] Entailment, Vol. I, Princeton University Press,
USA 1975.
ANDERSON, A. R., BELNAP, N. D., DUNN, J. M. [1992] Entailment, Vol. II, Princeton
University Press, USA 1992.
ANDREWS, P. B. [1965] A transfinite type theory with type variables, North-Holland Co.,
Netherlands 1965.
ANDREWS, P. B. [1971] Resolution in type theory, J. Symbolic Logic 36 (1971), 414-432.
ANDREWS, P. B. [1986] An Introduction to Mathematical Logic and Type Theory; to Truth
through Proof, Academic Press, USA and UK 1986.
AVRON, A. [1988] The semantics and proof theory of linear logic, Theoretical Computer
Science 57 (1988), 161-184.
AVRON, A. [1992] Axiomatic systems, deduction and implication, J. Logic and
Computation 2 (1992), 51-98.
BAADER, F., SIEKMANN, J. H. [1994] Unification theory, in Handbook of Logic in
Artificial Intelligence, Vol. 2: Deduction Methodologies, ed. D. M. Gabbay, C. J. Hogger,
J. A. Robinson, Oxford University Press, UK 1994, pp. 41-126.
BARENDREGT, H. P. [1984] The Lambda Calculus, North-Holland Co., Netherlands, 2nd.
edition 1984.
BARENDREGT, H. P. [1992] Lambda calculi with types, in Handbook of Logic in Computer
Science, Vol. 2, ed. S. Abramsky et at., Clarendon Press, UK 1992, pp. 117-309.
BARENDREGT, H. P., COPPO, M., DEZANI, M. [1983] A filter lambda model and the
completeness of type assignment, J. Symbolic Logic 48 (1983), 931-940.
BEN-YELLES, C: B. [1979] Type-assignment in the lambda-calculus; syntax and semantics,
thesis 1979, Mathematics Dept., University of Wales Swansea, Swansea SA2 8PP, UK.
BLOK, W. J., PIGOZZI, D. [1989] Algebraizable Logics, Memoirs of the American
Mathematical Society No. 396 (1989), Amer. Math. Soc., Providence, R.I. 02901, USA.
de BRUIJN, N. G. [1980] A survey of the project AUTOMATH, in To H. B. Curry, ed.
J. P. Seldin, J. R. Hindley, Academic Press, UK 1980, pp. 579-606.
BUNDER, M. W. [1982] Deduction theorems for weak implicational logics, Studia Logica 41
(1982), 95-108.
BUNDER, M. W. [1986] Review no. 03004, in Zentralblatt fdr Mathematik 574 (1986),
10-11, of Bunder, M. W. and Meyer, R. K. A result for combinatory, BCK logics and
BCK algebras, Logique et Analyse 28 (1985), 33-40.
BUNDER, M. W. [1991] Corrections to some results for BCK logics and algebras, Logique et
Analyse 31 (1991), 115-122.

169
170 Bibliography

BUNDER, M. W. [1996] Standardization of proofs in propositional logic, in Logic and

Reality (Proceedings of the 1989 conference in Christchurch, N. Z. in memory of A. N.
Prior), ed. J. Copeland, Oxford University Press, UK, 1996.
BUNDER, M. W. [1993] Theorems in classical logic are substitution instances of theorems in
condensed BCI logic, in Substructural logics, ed. P. Schroeder-Heister, K. Dosen,
Clarendon Press, UK 1993, pp. 43-62.
CHURCH, A. [1940] A .formulation of the simple theory of types, J. Symbolic Logic 5 (1940),
56-68.
CHURCH, A. [1941] The Calculi of Lambda Conversion, Princeton University Press, 1941.
Reprinted by University Microfilms, Inc., Ann Arbor, Michigan, USA 1963.
CHURCH, A. [1951] The weak theory of implication, in Kontrolliertes Denken,
Untersuchungen zum Logikkalkul and zur Logik der Einzelwissenschaften (Festgabe zum
60. Geburtstag von Prof. W. Britzelmayr), ed. A. Menne, A. Wilhelmy, H. Angstl;
publisher Kommissionsverlag Karl Alber, Miinchen 1951 (Sonderdruck); pp. 22-37.
CHURCH, A. [1976] SchrSder's anticipation of the simple theory of types, Erkenntnis 10
(1976), 407-411. (Presented at the 5th Internat. Congress for the Unity of Science, 1939,
and intended for publication in J. Unified Science (Erkenntnis) Vol. 9 which did not
appear.)
CONSTABLE, R. L. [1991] Type theory as a foundation for computer science, in Theoretical
Aspects of Computer Software, ed. by T. Ito, A. R. Meyer, Lecture Notes in Computer
Science, Springer-Verlag, Germany, No. 526 (1991), 226-243.
COPPO, M. [1984] Completeness of type-assignment in continuous lambda models, Theoretical
Computer Science 29 (1984), 309-324.
COPPO, M., DEZANI, M. [1978] A new type-assignment for lambda terms, Archiv fur
Mathematische Logik 19 (1978), 139-156.
CROSSLEY, J. N., SHEPHERDSON, J. C. [1993] Extracting programs from proofs by an
extension of the Curry-Howard process, in Logical Methods, ed. J. N. Crossley,
J. B. Remuel, R. A. Shore, M. E. Sweedler, Birkhauser, Boston, Mass., USA 1993,
pp. 222-288. '
CURRY, H. B. [1934] Functionality in combinatory logic, Proc. National Academy of
Sciences of the USA 20 (1934), 584-590.
CURRY, H. B. [1942] The combinatory foundations of mathematical logic, J. Symbolic Logic
7 (1942), 49-64.
CURRY, H. B. [1966] Technique.for evaluating principal functional character, Notes dated
March 17th. 1966, in Curry archive, Pennsylvania State University, Univ. Park, PA
16802, USA
CURRY, H. B. [1969] Modified basic functionality in combinatory logic, Dialectica 23 (1969),
83-92.
CURRY, H. B., FEYS, R. [1958] Combinatory Logic, Vol. I, North-Holland Co.,
Netherlands 1958.
CURRY, H. B., HINDLEY, J. R., SELDIN, J. P. [1972] Combinatory Logic, Vol. II,
North-Holland Co., Netherlands 1972.
van DALEN, D. [1980] Logic and Structure, Springer-Verlag, Germany, 1980.
DAMAS, L. [1984] Type Assignment in Programming Languages, Ph.D. thesis, Computer
Science Dept., University of Edinburgh, King's Buildings, Mayfield Rd., Edinburgh, UK
1984.
DAMAS, L., MILNER, R. [1982] Principal type-schemes for functional programming
languages, in Proc. Ninth Annual A.C.M. Symposium on the Principles of Programming
Languages (P.O.P.L.), 1982, Assoc. for Computing Machinery, New York, USA 1982,
pp. 207-212.
DEKKERS, W. [1988] Reducibility of types in typed lambda calculus, Information and
Computation 77 (1988), 131-137.
DILLER, J. [1968] Zur Berechenbarkeit primitiv-rekursiver Funktionale endlicher Typen, in
Contributions to Mathematical Logic, ed. H.A. Schmidt et al., North-Holland Co.,
Netherlands 1968, pp. 109-120.
 Bibliography 171

DOSEN, K. [1992a] Modal translations in substructural logics, J. Philosophical Logic 21

(1992), 283-336.
DOSEN, K. [1992b] The first axiomatization of relevant logic, J. Philosophical Logic 21
(1992), 339-356.
DRAGALIN, A. G. [1968] The computation of primitive recursive terms of finite type, and
primitive recursive realization, Zapiski Naucnyh Seminarov Leningradskogo Otdelenija
Matematiceskogo Instituta im V.I. Steklova Akademii Nauk S.S.S.R. (L.O.M.I.) 8
(1968), 32-45.
DWORK, C., KANELLAKIS, P. C., MITCHELL, J. C. [1984] On the sequential nature of
unification, J. Logic Programming 1 (1984), 35-50.
DYCKHOFF, R. [1992] Contraction free sequent calculi for intuitionistic logic, J. Symbolic
Logic 57 (1992), 795-807.
FITCH, F. B. [1936] A system of formal logic without an analogue to the Curry W operator,
J. Symbolic Logic 1 (1936), 92-100.
FORTUNE, S., LEIVANT, D., O'DONNELL, M. [1983] The expressiveness of simple and
second-order type structures, J. Assoc. for Computing Machinery 30 (1980) 151-185.
GALLIER, J. H. [1990] On Girard's "Candidats de deductibilite", in Odifreddi 1990,
pp. 123-203.
GALLIER, J. H. [1993] Constructive logics. Part 1: a tutorial on proof systems and typed
A-calculi, Theoretical Computer Science 110 (1993), 249-339.
GANDY, R. O. [1977] The simple theory of types, in Logic Colloquium 76, ed. R. O. Gandy,
M. Hyland, North-Holland Co., Netherlands 1977, pp. 173-181.
GANDY, R. O. [1980a] An early proof of normalization by A. M. Turing, in To H. B. Curry,
ed. J. R. Hindley, J. P. Seldin, Academic Press, UK 1980, pp. 453-455.
GANDY, R. O. [1980b] Proofs of strong normalization, in To H. B. Curry, ed. J. R. Hindley,
J. P. Seldin, Academic Press, UK. 1980, pp. 457-477.
GENTZEN, G. [1935] Untersuchungen fiber das logische Schliessen, Math. Zeitschrift 39
(1935), 176-210, 405-431. English transl. as Investigations into logical deduction it, The
Collected Papers of Gerhard Gentzen, ed. M. E. Szabo, North-Holland Co., Netherlands
1969.
GIANNINI, P., HONSELL, F., RONCHI DELLA ROCCA, S. [1993] Type inference: some
results, some problems, Fundamenta Informaticae 19 (1993), 87-125.
GIRARD, J. -Y. [1987] Linear logic, Theoretical Computer Science 50 (1987), 1-101.
GIRARD, J. -Y., LAFONT, Y., TAYLOR, P. [1989] Proofs and Types, Cambridge
University Press, UK 1989.
GLADSTONE, M. D. [1965] Some ways of constructing a propositional calculus of any
required degree of unsolvability, Trans. American Math. Soc. 118 (1965), 192-210.
GRZEGORCZYK, A. [1953] Some Classes of Recursive Functions, Rozprawy
Matematyczne, Panstwowe Wydawnictwo Naukowe, Warszawa, Poland 1953.
HAMILTON, A. [1988] Logic for Mathematicians, Cambridge University Press, UK 1978,
revised edition 1988.
HANATANI, Y. [1966] Calculability des fonctionnels recursives primitives de type fini sur les
nombres naturels, Ann. Japan Assoc. for the Philosophy of Science 3 (1966), 19-30.
Revised version: Calculability of the primitive recursive functionals of finite type over the
natural numbers, in Proof Theory Symposium, Kiel 1974, ed. by J. Diller, G. H. Muller,
Lecture Notes in Mathematics 500 (1975), Springer-Verlag, Germany, pp. 152-163.
HANKIN, C. [1994] Lambda Calculi, Clarendon Press, Oxford, UK 1994.
HENKIN, L. [1950] Completeness in the theory of types, J. Symbolic Logic 15 (1950), 81-91.
HERBRAND, J. [1930] Recherches sur la theorie de la demonstration, Trav. Soc. Sci. Lett.
Varsovie, Classe III Sci. Math. Phys., No. 33 (1930), English transl. in Jacques Herbrand
Logical Writings, ed. W.D. Goldfarb, Harvard University Press, USA 1971, pp. 44-202.
HEYTING, A. [1955] Intuitionism, an Introduction, North-Holland Co., Netherlands 1955
(3rd edition 1971).
172 Bibliography

HINATA, S. [1967] Calculability of primitive recursive functionals of finite type, Sci. Report
Tokyo Kyoiku Daigaku, Section A, 9 (226) (1967), 42-59.
HINDLEY, J. R. [1969] The principal type-scheme of an object in combinatory logic, Trans.
American Math. Soc. 146 (1969), 29-60.
HINDLEY, J. R. [1983a] The completeness theorem for typing 2-terms, Theoretical Computer
Science 22 (1983), 1-17.
HINDLEY, J. R. [1983b] Curry's type-rules are complete with respect to the F-semantics too,
Theoretical Computer Science 22 (1983), 127-133.
HINDLEY, J. R. [1989] BCK-combinators and linear 2-terms have types, Theoretical
Computer Science 64 (1989), 97-106.
HINDLEY, J. R. [1992] Types with intersection, an introduction, Formal Aspects of
Computing 4 (1992), 470-486.
HINDLEY, J. R. [1993] BCK and BCI logics, condensed detachment and the 2-property,
Notre Dame J. Formal Logic 34 (1993), 231-250.
HINDLEY, J. R., MEREDITH, D. [1990] Principal type-schemes and condensed detachment,
J. Symbolic Logic 55 (1990), 90-105.
HINDLEY, J. R., SELDIN, J. P. [1986] [HS 86] Introduction to Combinators and A-calculus,
Cambridge University Press, UK 1986.
HIROKAWA, S. [1991a] Principal type assignment to lambda terms, International Journal of
the Foundations of Computer Science 2 (1991), 149-162.
HIROKAWA, S. [1991b] Principal type-schemes of BCI-lambda-terms, in Theoretical Aspects
of Computer Software, ed. by T. Ito, A. R. Meyer, Lecture Notes in Computer Science,
Springer-Verlag, Germany, No. 526 (1991), 633-650.
HIROKAWA, S. [1991c] BCK-formulas having unique proofs, in Category Theory and
Computer Science, ed. by D. H. Pitt et al., Lecture Notes in Computer Science,
Springer-Verlag, Germany, No. 530 (1991), 106-120.
HIROKAWA, S. [1992a] The converse principal type-scheme theorem in lambda calculus,
Studia Logica 51 (1992), 83-95.
HIROKAWA, S. [1992b] Balanced formulas, BCK-minimal formulas and their proofs, in
Logical Foundations of Computer Science - Tver '92, ed. by A. Nerode, M. Taitslin,
Lecture Notes in Computer Science, Springer-Verlag, Germany, No. 620 (1992), 198-208.
HIROKAWA, S. [1993a] Principal types of BCK-lambda-terms, Theoretical Computer
Science 107 (1993), 253-276.
HIROKAWA, S. [1993b] The relevance graph of a BCK-formula, J. Logic and Computation
3 (1993), 269-285.
HIROKAWA, S. [1993c] The number of proofs for an implicational formula, J. Symbolic
Logic 58 (1993), 1117 (abstract only; fuller MS informally circulated 1991).
HOPCROFT, J. E., ULLMAN, J. D., [1979] Introduction to Automata Theory, Languages,
and Computation, Addison-Wesley, USA 1979.
HOWARD, W. [1969] The formulae-as-types notion of construction, MS 1969, publ. in To
H. B. Curry, ed. J. R. Hindley, J. P. Seldin, Academic Press, UK 1980, pp. 479-490.
HOWARD, W. [1970] Assignment of ordinals to terms for primitive recursive functionals of
finite type, in Intuitionism and Proof Theory (Proc. Buffalo Conference 1968), ed.
A. Kino, J. Myhill, R. Vesley, North-Holland Co., Netherlands, 1970, pp. 443-458.
HUDELMAIER, J. [1993] An O(nlogn)-space decision procedure for intuitionistic
propositional logic, J. Logic and Computation 3 (1993), 63-75.
HUET, G. [1975] A unification algorithm for typed 2-calculus, Theoretical Computer Science
1 (1975), 27-57.
ISEKI, K., TANAKA, S. [1978] An introduction to the theory of BCK-algebras,
Mathematica Japonica 23 (1978), 1-26.
JASKOWSKI, S. [1963] Uber Tautologien, in welchen keine Variable mehr als zweimal
vorkommt, Zeitschrift fur Mathematische Logik 9 (1963), 219-228.
KALMAN, I. A. [1982] The two-property and condensed detachment, Studia Logica 41
(1982), 173-179.
 Bibliography 173

KALMAN, J. A. [1983] Condensed detachment as a rule of inference, Studia Logica 42

(1983), 443-451.
KANELLAKIS, P., MAIRSON, H., MITCHELL, J. C. [1991] Unification and ML-type
reconstruction, in Computational Logic, Essays in Honour of Alan Robinson, ed.
J. -L. Lassez, G. Plotkin, M.I.T. Press, Cambridge, Mass., USA 1991, pp. 444-478.
KANELLAKIS, P., MITCHELL, J. C. [1989] Polymorphic unification and ML typing, in
Proceedings of the 16th Annual A.C.M. Symposium on the Principles of Programming
Languages (P.O.P.L), 1989, Assoc. for Computing Machinery, New York, USA, 1989
pp. 105-113.
KFOURY, A. J., TIURYN, J., URZYCZYN, P. [1990] ML typability is
DEXPTIME-complete, Lecture Notes in Computer Science 431 (1990), Springer-Verlag,
Germany, pp. 206-220.
KNIGHT, K. [1989] Unification: a multidisciplinary survey, A.C.M. Computing Surveys 21
(1989), 93-124.
KOMORI, Y. [1987] BCK algebras and lambda calculus, in Proc. 10th Symposium on
Semigroups, Sakado 1986, publisher Josai University, Sakado, Saitama, Japan, 1987,
pp. 5-11.
KOMORI, Y. [1989] Illative combinatory logic based on BCK-logic, Mathematica Japonica
34 (1989), 585-596.
KOMORI, Y., HIROKAWA, S. [1993] The number of proofs for a BCK-formula, J.
Symbolic Logic 58 (1993), 626-628.
KRIVINE, J. -L. [1990] Lambda-calcul, types et modeles, Masson, France, 1990. English
transl. Lambda-Calculus, Types and Models, Ellis-Horwood, USA 1993.
LAMBEK, J., SCOTT, P. J. [1986] Introduction to Higher-order Categorical Logic,
Cambridge University Press, UK, 1986.
LAUCHLI, H. [1965] Intuitionistic propositional calculus and definably non-empty terms
(abstract), J. Symbolic Logic 30 (1965), 263.
LEIVANT, D. [1979] Assumption classes in natural deduction, Zeitschrift fur Mathematische
Logik 25 (1979), 1-4.
LEMMON, E. J., MEREDITH, C. A., MEREDITH, D., PRIOR, A. N., THOMAS, I.
[1957] Calculi of pure strict implication, mimeographed 1957, publ. in Philosophical
logic, ed. Davis, Hockney, Wilson, D. Reidel Co., Netherlands 1969, pp. 215-250.
LUKASIEWICZ, J. [1939] Der .4quivalenzenkalkul, Collectanea Logica 1 (1939), 145-169.
English transl. as The equivalential calculus, in Jan Lukasiewicz Selected Works, ed.
L. Borkowski, North-Holland Co., Netherlands 1970, pp. 250-277.
MAIRSON, H. [1990] Deciding ML typability is complete for deterministic exponential time,
in Proceedings of the 17th Annual A.C.M. Symposium on the Principles of
Programming Languages (P.O.P.L.), 1990, Assoc. for Computing Machinery, New
York, USA 1990, pp. 382-401.
MARTELLI, A., MONTANARI, U. [1982] An efficient unification algorithm, A.C.M.
Transactions on Programming Languages and Systems 4 (1982), 258-282.
MARTIN-LOF, P. [1975] An intuitionistic theory of types: predicative part, in Logic
Colloquium 173, ed. H. E. Rose et al., North-Holland Co., Netherlands 1975, pp. 73-118.
MASLOV, S. [1964] An inverse method of establishing deducibilities in the classical predicate
calculus, Soviet Mathematics 5 (1964), 1420-1424. (English transl. of Doklady Akad.
Nauk S.S.S.R. 159 (1964), 17-20.)
MEGILL, N. D., BUNDER, M. W. [1996] Weaker D-complete logics, Journal of the Interest
Group on Propositional Logics 4 (1996), 215-225.
MEREDITH, C. A., PRIOR, A. N. [1963] Notes on the axiomatics of the propositional
calculus, Notre Dame J. Formal Logic 4 (1963), 171-187.
MEREDITH, D. [1977] In memoriam Carew Arthur Meredith, Notre Dame J. Formal Logic
18 (1977), 513-516.
MEYER, R. K., BUNDER, M. W. [1988] Condensed detachment and combinators, Tech.
Report TR-ARP-8/88 (1988), Research School of the Social Sciences, Australian
National University, Canberra, Australia.
174 Bibliography

MILNER, R. [1978] A theory of type polymorphism in programming, J. Computer and

System Sciences 17 (1978), 348-375.
MINTS, G. [1979] A primitive recursive bound of strong normalization for predicate calculus,
Zapiski Naucnyh Seminarov Leningradskogo Otdelenija Matematiceskogo Instituta im
V. I. Steklova Akademii Nauk S.S.S.R. (L.O.M.I.) 88 (1979), 131-135 (in Russian with
English summary).
MINTS, G., TAMMET, T. [1991] Condensed detachment is complete for relevance logic: a
computer-aided proof, J. Automated Reasoning 7 (1991), 587-596.
MITCHELL, J. C. [1988] Polymorphic type inference and containment, Information and
Computation 76 (1988), 211-249.
MITCHELL, J. C. [1990] Type systems for programming languages, in Handbook of
Theoretical Computer Science Vol. B, Formal Models and Semantics, ed. J. van
Leeuwen, Elsevier, Netherlands 1990, pp. 365-458.
MITCHELL, J. C. [1996] Foundations for Programming Languages, M.I.T. Press, USA 1996.
MITSCHKE, G. [1979] The standardization theorem for the R-calculus, Zeitschrift fur
Mathematische Logik 25 (1979), 29-31.
MOH, S: K. [1950] The deduction theorems and two new logical systems, Methodos 2 (1950),
56-75.
MORRIS, J. H. [1968] Lambda-calculus Models of Programming Languages, Ph.D. thesis,
Massachusetts Institute of Technology, Cambridge, Mass., USA 1968.
NEDERPELT, R. P. [1973] Strong Normalization in a Typed Lambda-calculus with
Lambda-structured Types, Ph.D. thesis, Tech. Hogeschool, Eindhoven, Netherlands 1973.
NERODE, A., ODIFREDDI, P., [199-] Lambda Calculi and Constructive Logics, Cornell
University Press, USA, to appear 1997.
ODIFREDDI, P. [1990] (editor) Logic and Computer Science, Academic Press, USA 1990
(Series APIC Studies in Data Processing No. 31).
PATERSON, M., WEGMAN, M. [1978] Linear Unification, J. Computer and System
Sciences 16 (1978), 158-167.
PLOTKIN, G. D. [1994] A semantics for static type-inference, Information and Computation
109 (1994), 256-299.
PRAWITZ, D. [1960] An improved proof procedure, Theoria 26 (1960), 102-139.
PRAWITZ, D. [1965] Natural Deduction, Almqvist and Wiksell, Sweden 1965.
PRIOR, A. N. [1955] Formal Logic, Clarendon Press, UK 1955 (2nd. edition. 1962).
REVESZ, G. [1988] Lambda Calculus, Combinators and Functional Programming,
Cambridge University Press, UK 1988.
REYNOLDS, J. [1985] Three approaches to type structure, in Mathematical Foundations of
Software Development, Vol. 1 (CAAP '85), ed. by H. Ehrig et al., Lecture Notes in
Computer Science, Springer-Verlag, Germany, No. 185 (1985), 97-138.
REZUS, A. [1981] Lambda-conversion and logic, Ph.D. thesis, Universiteit Utrecht, Math.
Inst., Budapestlaan 6, de Uithof, Utrecht, Netherlands.
ROBINSON, J. A. [1965] A machine-oriented logic based on the resolution principle, J. Assoc.
for Computing Machinery 12 (1965), 23-41.
ROBINSON, J. A. [1966] Review of Prawitz 1960, J. Symbolic Logic 31 (1966), 126-127.
ROBINSON, J. A. [1979] Logic: Form and Function. The Mechanization of Deductive
Reasoning, Edinburgh University Press, UK 1979.
RUSSELL, B. [1903] Principles of Mathematics, Allen and Unwin, UK 1903. (1992 edition:
Routledge and Kegan Paul, UK.)
SALLE, P. [1978] Une extension de la theorie des types, in Automata, Languages and
Programming, ed. by G. Ausiello, C. Bohm, Lecture Notes in Computer Science,
Springer-Verlag, Germany, No. 62 (1978), 398-410.
SANCHIS, L. E. [1967] Functionals defined by recursion, Notre Dame J. Formal Logic 8
(1967), 161-174.
 Bibliography 175

SCEDROV, A. [1990] A guide to polymorphic types, in Odifreddi 1990, pp. 387-420.

SCHWICHTENBERG, H. [1991] An upper bound for reduction sequences in typed
lambda-calculus, Archiv fur Math. Logik 30 (1991), 405-408.
SELDIN, J. P. [1968] Studies in Illative Combinatory Logic, Ph.D. thesis, University of
Amsterdam, Netherlands 1968.
SELDIN, J. P. [1977] A sequent calculus for type assignment, J. Symbolic Logic 42 (1977),
11-28.
SELDIN, J. P. [1978] A sequent calculus formulation of type assignment with equality rules
for the ,1/3-calculus, J. Symbolic Logic 43 (1978), 643-649.
STATMAN, R. [1979a] Intuitionistic propositional logic is polynomial-space complete,
Theoretical Computer Science 9 (1979), 67-72.
STATMAN, R. [1979b] The typed A-calculus is not elementary recursive, Theoretical
Computer Science 9 (1979), 73-81.
STATMAN, R. [1980] On the existence of closed terms in the typed A-calculus; Part 1 in To
H.B. Curry, ed. J.R. Hindley and J.P. Seldin, Academic Press, UK 1980, pp. 511-534;
Part 2 in Theoretical Computer Science 15 (1981), 329-338.
STOUGHTON, A. [1988] Substitution revisited, Theoretical Computer Science 59 (1988),
317-325.
TAIT, W. W. [1965] Infinitely long terms of transfinite type, in Formal Systems and Recursive
Functions, ed. J. N. Crossley, M. A. Dummett, North-Holland Co., Netherlands 1965,
pp. 176-185.
TAIT, W. W. [1967] Intensional interpretations of functionals of finite type, J. Symbolic Logic
32 (1967), 198-212.
TAKAHASHI, M. H. [1991] Theory of Computation, Computability and Lambda-calculus,
Kindai Kagaku Sha, Tokyo, Japan 1991 (in Japanese).
TAKAHASHI, M. H., AKAMA, Y., HIROKAWA, S. [1994] Normal proofs and their
grammar, in Theoretical Aspects of Computer Software (TACS '94), ed. by M. Hagiya,
J. C. Mitchell, Lecture Notes in Computer Science, Springer-Verlag, Germany, No. 789
(1994), 465-493.
TIURYN, J. [1990] Type inference problems: a survey, in Mathematical Foundations of
Computer Science 1990, ed. by B. Rovan, Lecture Notes in Computer Science,
Springer-Verlag, Germany, No. 452 (1990), 105-120.
TROELSTRA, A. S. [1973] (editor) Metamathematical Investigation of Intuitionistic
Arithmetic and Analysis, Lecture Notes in Mathematics, Springer-Verlag, Germany,
No. 344 (1973).
TROELSTRA, A. S., van DALEN, D. [1988] Constructivism in Mathematics, an
Introduction, Vols. 1 & 2, North-Holland Co., Netherlands 1988.
TURING, A. M. [1942] Notes published in Gandy 1980a.
TYSZKIEWICZ, J. [1988] Complexity of Type Inference in Finitely Typed Lambda
Calculus, Master's thesis, University of Warsaw, Poland 1988.
URQUHART, A. [1984] The undecidability of entailment and relevant implication, J.
Symbolic Logic 49 (1984), 1059-1073.
URQUHART, A. [1995] The complexity of propositional proofs, Bull. Symbolic Logic 1
(1995), 425-467.
DE VRIJER, R. [1987] Exactly estimating functionals and strong normalization, Proc.
Koninklijke Nederlandse Akademie van Wetenschappen, Series A, 90 (1987), 479-493.
WAND, M. [1987] A simple algorithm and proof for type inference, Fundamenta Informaticae
10 (1987),115-122.
ZAIONC, M. [1985] The set of unifiers in typed A-calculus as regular expression, in Rewriting
Techniques and Applications, ed. by J.-P. Jouannaud, Lecture Notes in Computer
Science, Springer-Verlag, Germany, No. 202 (1985), pp. 430-440.
ZAIONC, M. [1987a] The regular expression descriptions of unifier set in the typed
A-calculus, Fundamenta Informaticae 10 (1987), 309-322.
176 Bibliography

ZAIONC, M. [1987b] Word operations definable in the typed ,1-calculus, Theoretical

Computer Science 52 (1987), 1-14.
ZAIONC, M. [1988] Mechanical procedure for proof construction via closed terms in typed
R-calculus, J. Automated Reasoning 4 (1988), 173-190.
ZAIONC, M. [1990] A characterization of lambda definable tree operations, Information and
Computation 89 (1990), 35-46.
ZAIONC, M. [199-] Fix point technique for counting terms in typed 2-calculus, Fundamenta
Informaticae (to appear).
ZUCKER, J. [1974] The correspondence between cut-elimination and normalization, Annals of
Mathematical Logic 7 (1974): Part I, pp. 1-112; Part II, pp. 113-155.
 Table of principal types

This table gives a (not necessarily unique) principal inhabitant of each of some
types. It gives a fl-nf where possible, but if there is no normal principal inhabitant
it gives a non-normal one. It includes most of the PT inform Atio. i in Tables 3E2a
and 8B7a.
Type A principal inhabitant
a-*a
a-*b-*a K
a-*a-*a (Axyz K(xy)(xz))I [Nprinc = 0]
(a-.b)-*a-*b
(a-.b)-.a->a Axy Ky(xy) [Nprinc = 0]
(a-.a)-.a-*a [and
a-*(a-*b)-*a 2xy Kx(yx) [Nprinc = 0]
a-*a-4b-*b [Nprinc = 0]
((a-a)-+a)-a , others see 8B7a(12)]
(a-.a-*b)-*a-.b )xY'xYY = W
(a-*a-*a)-*a-+a
(a-.b)-*a-+a-*b 2xyz K(xy)(xz) [Nprinc = 0]
C
(a1-+...
b- aj -* ...
(a-.b)-.(c-*a)-*c-b 2xyz x(yz) = B
(a-.b)-.(b-.c)-*a-+c 2xyzy(xz) = B'
S
(a- b- 1xyz xyzz

b)- (e- b--+d)- c- a-+d

(a-*b)-*(c I -* ... ,luvwi ... w(ux)
cj- ... c, a-*d
((a-*b)-*c-*a)-*(a-*b)-*c-*b
(a-*a')-*(b--*b')-*(a'- b)- a--+b'
(a-*b)-*(c-*d)-*c-*(d--*a)--*b )luvwx u(x(vw))
(a-+b-+c)-+a-*(d-*b)-*d-*c

177
178 Table of principal types

Table of principal types - continued

Type A principal inhabitant

Auvwx u(vx)w

c)- (d-+a)-+(d-+b)-*d-.c Auvwx u(vx)(wx)

Auvwxy u(x(vwy))

a3 b
(a-*b1-+... b,-+a--+d Auvwl ...
(a-+b)-+((c1 -+ ...
(c1-+...
(a-+b-+c)-+((d, -+... f )-+ ...Y,,z'xY1 ... yn(uzw))
b-+(d1-+...
 Index

A-logics (see axiom-based logics) BCK-logic (also see BCK-) 90

A°I-

104 D-incompleteness of 106-7

A'- 89 condensed logics 104
o/(T,d) 120 Intuitionist logic (also see Intuit.) 90
d) 124 D-completeness of 106
dterms(T,d) 120 link with Natural Deduction 92
terms(T, <_ d) 124 link with types of A-terms 91
abstract (A-abstract) 1 R. (also see R.) 90
abstracted atom 7 D-completeness of 105
abstractor 2, 141 T. (logic of ticket entailment) 106
covering 2, 130 axioms
initial 110, 130 ofTAA 15
abstractor position 140 (B) (B'), (C), (I), (K), (W) 81, 90
algorithms
for counting inhabitants 108-139, esp. 125-6 B (closed A-term) 4
Search algo. in 120-2 principal type of 16, 33, 50
main searching strategy 114-7 deduction for PT of 16
Converse PT algo. 96-101, esp. 101 (B) (axiom) 81, 90
Principal-type (PT) algo. 30-51, esp. 44-9 B' (closed A-term) 4
historical comment on 33-4 principal type of 18, 50
main step in 38 deduction for PT of 161
Type-checking algo. (= PT algo.) (B') (axiom) 81
Unification algo. 42-3 basic combinators, sets S of 157-8
alphabetic variant 37 basis, typable 159
application (term) 1 BCI-logic
application in a A-model 58 Natural Deduction version 87
applicative combination 157 axiom-based version 90
argument in a normal form 110, 128 D-incompleteness of 106-7
subargument 129 link with types of A-terms 88
argument-branch 130 BCIW-PT 94
argument-part of a redex 144 BCIA-terms 11
argument position 140 BCIA-PT 94
arity of a type 115 BCK-logic
assignment of types (see TAA) Natural Deduction version 86
association rule for parentheses axiom-based version 90
in terms 1 D-incompleteness of 106-7
in types 12 link with types of A-terms 88
atom, in a term 1 BCKW-PT 94
abstracted 7 BCKA-terms 10
typed 67 typability of all 28
atom, in a type 12 BCKA-PT 94
atomic deduction 78, 148 binding occurrence 2, 141
atomic term (see atom in a term) body of an abstract 2
axiom-based logics 88-92 body position 140
A-logics, definition 88-9 bound variable-occurrence 2
BCI-logic (also see BCI-) 90 changing (see a-conversion)
D-incompleteness of 106-7 bound-variable clash 3
in typed terms 68

179
180 Index

C (closed A-term) 4 construction-tree

principal type of 18, 50 of a term 141-2
deduction for 161 of a type 152
(C) (axiom) 81, 90 condensed version 153-4
CA() 130 contains (said of components) 142
cancelled assumptions (see discharging) context, type- 14
cancelling reduction 11 for nf-schemes 119-20
cardinality of a set of terms 109 M-context 15
Changing bound variables (see a-conversion) minimum context, Con( ) 65
Church numerals 4, 118 context-free grammars 123, 124
Church-Rosser theorem contraction (also see S-, flry-, rl-
for fl 5 #-contracts 4
for flry and ry 8 for typed terms 69
for typed terms 70 fill-contracts 8
Church-style type-theories 14 ry-contracts 7
c.i. (= common instance) 38-9 for typed terms 69
clash of bound variables 3 contraction (kind of substitution) 96
in typed terms 68 contractum
Classical logic, D-completeness of 106 of a fl-redex 4, 69
closed nf-scheme 119 of an q-redex 7, 69
closed term 4 Converse principal-type algorithm 96-101, esp.
combination of given terms 157 101
combinator (= closed l-term) 4 Converse principal-type theorem
B, B', C, I, K, S, W, Y 4, 50 for BB' IW- combinations 102
fixed-point (see Y) for Al 95
combinatory logic (CL) proof of 96-101
imitated in 2-calculus 157-9 for AK 94
combinatory completeness 158, 159 conversion
typable basis for 159 a- (see a-conversion)
type-system TAB for CL 19 /3- (also see #-) 5
common instance 38-9 /3rl- (also see flry-) 8
Commuting lemma for fi and rl 8 rl- (also see n-) 7
complete development 146 conversion-sensitivity of Types( ) 26
complete discharge convention 77 corresponding logic deduction AL 79
completeness counting (contrast to enumerating) 109
combinatory 158-9 Counting theorem and algorithm
comp. lemma for Search algo. 132 for Long(2) 125
of rule (D) (see D-completeness) for Nhabs(s) 126
Completeness theorem (semantic) 59 for Nhabsq(t) 125
component of a substitution 35 Search algo. 120-2
component of a term 2, 141 main search strategy 114-7
contains another component 142 covering abstractor 2
does not overlap another 143 CA() 130
is a part of another 142 Curry-Howard mapping 79-85, esp. 79, cf. also 74
is a proper part of another 143 is an isomorphism of proofs 83
is disjoint from another 143 is not 1-1 for deductions 81
is in another 142 one-sided inverse of 82, 84
is isomorphic to another 143 Curry-Howard theorem 84
proper 141 for axiom-based logics 91
properly contains another 143 for condensed logics 105
component of a type 151 for weaker logics 88
s-component 155 Curry-style type-theory (also see TAR) 14
composite term 1 cut-elimination, to prove normalisn. 72
composite type 12
composition of substitutions (o) 36 (D) (condensed detachment rule) 102-7, esp. 102
Composition-extension lemma 36 historical note on 103-4
Con( ) (minimum context) 65 D(p--.a) s 102-3
concatenation of positions 140 D( ) 124
conclusion of a type (see tail) D-completeness 104
conclusion (= tail) position 153 of intuitionist logic 106
Condensed Detachment (see (D), rule) of R.. 105
condensed logics 104 of some other logics 106
link with principal types 105 preserved by D-extension 105
condensed tree of a type 153-4 fails for BCI, BCK logics 106-7
consistent (for contexts) 14, 15 D-deducible, D-deduction 104
D-extension of a logic 105
 Index 181

D-incompleteness of BCI, BCK logics 106 identification (kind of substitution) 96

D-proof 104 Identification lemma 100
D-theorem 104 implicational formulae 75
Deduction theorem mentioned 92 implicational logics (see logics)
deduction-rules (see rules) in (said of components) 142
deductions inhabitant 108-39, esp. 108-9
in a Natural Deduction logic 78 normal 108-9
in axiom-based logics 89 long 110
in TAx (see TAx-deduction) principal 112
substitutions-first 90 Inhabitant-counting algorithm
depth of a /3-nf 110 for Long(r) 125
depth of a subargument in a of 129 for Nhabs(r) 126
Depth( ) (see depth of a fi-nf) for Nhabsq(r) 125
detachment rules (see (-*E), (D)) Search algo. in 120-2
development 146 main search strategy in 114-7
disagreement pair 43 inhabitants of particular types (tables of) 118,
discharge convention, complete 77 177-8
discharge convention, partial 77 inhabited type 109
discharge label 76, 78 initial abstractors 110, 130
discharge-equivalent 79 instance (= substitution-inst.) 31
discharging 16, 76-8 common 38-9
vacuous 16,76-8 interpretations in a A-model
disjoint components 143 of terms 57-8
diverging positions 140 of types 58
Dom( ) 35 intersection-types 18
duplicating reduction 11 Intuitionist logic (implicational)
axiom-based version 90
Enhanced Subformula lemma 131 D-completeness of 106
enumerating (contrast to counting) 109 Natural Deduction version 76
environment, semantic 57 consistency of 117
(Eqp), (Egp,t), rules 52 decidability of 85
(Eq) denotes both 52 link with types 84
equal (see /3-, P7-, ii-conversion) inverse of a substitution 37
Equality-postponement theorem 53 isomorphic occurrences 115, 43
erasing types 65
expansion, /3- 5 K (closed A-term) 4
Subject-expansion theorem 24 principal type of 17, 50
extendable nf-scheme 122 deduction for 17
extensions of a logic 105 (K) (axiom) 81, 90
extensions of a nf-scheme 122
extensional A-model 58 lambda-to-logic (see Curry-Howard)
extensionally equivalent substitutions 35 leftmost p-reduction 6
length
fixed-point combinator (see Y) of a deduction 149
formula, implicational 75 of a position 140
formula, TAx- 15 of a reduction 5
formulae-as-types (also see Curry-Howard) 74 of a term 2
free variable-occurrence 2, 68 a typed term 67
function position 140 of a type 13
function-part of a redex 144 of an argument-branch 130
FV( ) 2,68 of CA( ), IA( ) 130
of IAT( ) 131
Habs( ), Habs,( ), 108 2-length of a position 129
head of a normal form 110, 128 linear logic, implicational fragment 87
Hilbert-Gentzen link theorem 92 linear l-terms 11
Hilbert-style (see axiom-based logics) listing (= enumerating) 109
HS86 (refers to Hindley & Seldin 1986 in logic deductions (see deductions)
Bibliography) logic-to-lambda mapping 82
logics, implicational 74-92, 102-7
I (closed A-term) 4 (also see BCI-, BCK-, Classical, Intuitionist,
principal type of 17, 50 Rte, T..)
(I) (axiom) 81, 90 axiom-based (= A-logics) 88-9
IA( ) 130 Natural Deduction based 76-9, 86-7
IAT() 130-1 condensed logics 104
IAT-intervals 136 link with principal types 105
identical modulo renaming 37 extensions, D-extensions 105
182 Index

long /3-normal form 110 occurrences in a term 2, 141

Long() 110 abstractors 2, 141
counting theorem and algo. for 125 binding occurrences 2, 141
search algorithm for 120-2 occurrences of ix 141
in restricted classes of terms 127 occurrences in a type 151
Long(, d) 111 negative, positive 152, 156
Longq( ) 113 occurs (in a term) 2, 141
occurs (in a type) 151
m.g.c.i. (= most general common instance) 39 negatively, positively 152, 156
m.g.c.i.-generator 39 Order of a type, Order( ) 156
m.g.u. (= most general unification) 41 overlaps, does not overlap 143
m.g.u. (= most general unifier) 41
major premise 89 parent of a residual 146
maximal-length reduction 5 parentheses, omission from terms 1
maximal (= long) normal form 110 parentheses, omission from types 12
MCD 146 part of (said of components) 142
metavariables 119 Partial completeness theorem 159
minimal complete development (see MCD) partial discharge convention, partial discharging
minimal logic (see Intuitionist logic) 76-7
minimal member 146 Peirce's Law (PL) 75
minimal tail-repetition 136 intuitionistic unprovability 85, 117
minimum context 65 (PL) (see Peirce's Law)
minor premise 89 positions 140, 153
ML (programming language) x, 30, 52 abstractor pos., argument pos., body pos.,
model, A- 57 function pos. 140
extensional 58 concatenation of two pos. 140
modified basic functionality 19 conclusion or tail pos., premise pos. 153
refinement of a pos. 140
Modus Ponens (also see (-.E)) 89
monovalent (= consistent) 14 diverging pair of pos. 140
positive occurrence 152
most general common instance (see m.g.c.i.)
positive s-component 156
most general unification 41 Postponement theorem for ry 8
most general unifier 41
predicate 14, 15
premise position 153
Natural Deduction premises of a type 115, 155
in implicational logics 76-9, 86-7 subpremises 155
in TA2 15-16, 18 Princ( ) 112, 127
negative occurrence 152 principal deduction 32
negative s-component 156 principal inhabitants of a type 112
negative subpremise-sequence 156 principal pair 32
new metavariable or variable 119 principal type (see PT)
newly-created redex-occurrence 147 proofs (see also deductions)
of (= normal form, see fl-nf, $ry-nf, q-nf) in a Natural Deduction logic 79
nf-scheme 119 in an axiom-based logic 89
closed 119 in TAx 16
long typed 120 proper component of a term 141
proper 119 proper nf-scheme 119
typed 119 proper part of a component 143
Nhabs( ), Nhabst( ), 108 proper refinement of a position 140
counting theorem and algo. for 126 proper s-subtype 154
searching for members 114-22 properly contains 143
Nhabsq( ) 113 propositional variables 75
counting 125 provability problem 89
non-cancelling reduction 11 provable formula (= theorem) 79, 89
non-duplicating reduction 11 PT (= principal type)
non-overlapping components 143 definition in TAx 32
normal form (see /3-, /3q- ry-nf) of B, I, K (with deductions) 16-17
normal inhabitant (see Nhabs( )) of B',C,S,W 18,161
normalisation, strong (see SN) tables of PT's 50, 177-8
normalisation, weak (see WN) in TA,t+9 and TAx+#q 56
Nprinc( ) 112, 127 BCIA-PT, BCKW-PT, etc. 94
Nprincq( ) 113, 127 AI-PT, AK-PT as is 94
NSS( ) 156 P T( ) (= principal type of) 32
numerals of Church 4 PT algorithm 30-51, esp. 44-9
principal types of 50, 118 main step in 38
R (see numerals of Church) historical note 33-4
 Index 183

PT algorithm, converse (see Converse Shrinking lemma, statement 124

principal-type algorithm) proof 135-8
PT theorem significant subtype (see s-subtype)
for TAA 33 simple semantics 59
for TAx+p and TA,t+g,, 56 simple type theory (see also TAI) ix
Pumping lemma 124 simultaneous substitution
pure A-calculus 1 in terms 3
in types 31, 35
R-, (relevance logic) single substitution in types 35
axiom-based version 90 skeletal type 96
D-completeness of 105 skeleton of a type 96
Natural Deduction version 86 SN (strong normalization) 27-8
link with types of Al-terms 88 SN theorem
Range() 35 for typable terms in TAA 28
redex for typed terms 72
/3- (also see /3-redex) 4 historical note 73
Qn 8 Standardization theorem mentioned 7
n 7 stratified term (= typable term, q.v.)
typed fl- , n- 69 Stretching lemma, statement 124
redex-occurrences, notation for 144 proof 134-5
reduces, reductions strong normalization (see SN)
j3- (also see fl-reduction) 4 strong type-invariance 158
A- 8 structure of fl-normal form 7
n- (also see n-reduction) 7 typed 110
length of a reduction 5 long typed 115
reduction of a deduction 70 (Sub), substitution rule 89
reduction of subjects 24 subarguments 129
refinement of a position 140 Subformula lemma, enhanced 131
proper refinement 140 subformula property 22
regular languages 124 subject 14, 15
reject, in Search algorithm 122 Subject-construction theorem 20-1
relevance logic (see R-) Subject-expansion theorem 24
replacement in a term 143 Subject-reduction theorem
typed 67 for #-reduction 24
in a deduction 149 for n-reduction in TAx+B 56
replacement lemmas 150-1 subpremises, Subpremises( ) 155
replacements, suitable 121 substitution-instance (see instance)
residuals of a redex-occurrence 145 substitution into terms
of a set 146 def for untyped terms 3
rules of deduction simultaneous 3
(D), condensed detachment 102-3 def. for typed terms 68
historical note 103-4 subs. lemmas for deductions 23
(Eq), (Eq6), (Egdn) 52 substitution rule (Sub) 89
(Sub) in propositional logics 89 substitutions-first deduction 90
(-.E) in propositional logics 76, 89 substitutions into types 31, 35
(.-.E) in TAA 16 components of a sub. 35
(-.I) in TAA 16 composition of two subs. (o) 36
(-'1)main, (-'I)vac 16 contractions (= identifications, q.v.)
extensional equivalence (=ext) 35
S (closed A-term) 4 identifications 96
principal type of 18 inverse of a one-to-one sub. 37
deduction for 161 one-to-one sub. 37
S-and-variables combination 157 renamings 37
S-combination 157 single sub. 35
s-component 155 trivial compons. of a sub. 35
negative, positive 156 union of two subs. (U) 35
s-subtype 154 variable-domain (Dom( )) 35
satisfaction in a A-model 58 variable-range (Range( )) 35
scope 2 subtails, Subtails( ) 155
Search algorithm 120-2 subterm 2, 141
main strategy of 114-7 subtype 151
Search-completeness lemma 132 s-subtype 154
self-application 17 success made in Search algo. 122
semantic environment 57 suitable multi-replacements 122
semantics, simple 59 suitable replacements 121
184 Index

T., ticket entailment logic 106 f3-, q-redex, reduction 69

TAc 19-20 /3-, q-, $n-normal form 69
TA2 15-16 structure of /3-nf 110
axioms and rules for 15-16 types 12
basic facts about 12-29 atomic 12
comparison with HS86 version 18 composite 12
TA2- construction-tree of 152
-deducible formula 16 condensed 153-4
-deduction 16 inhabitants, tables of 50, 177-8
detailed definition 148-9 principal (also see PT) 32
length of 149 with intersection 18
translated to a typed term 66 Types( ) 26
-formula 15 Typesno eq( ) 56
-typable term 27 Typesp( ), Typespn( ) 55
TA2+p, TA2+pn 52
TA2+p[o] (for both TA2+p, TA2+pn) 52 unextendable argument-branch 130
TA2+p-typable, TA2}pn-typable 52 unifiable 40
tail of a type, Tail( ) 115, 155 unification 40
subtail 155 most general (= m.g.u.) 41
tail position 153 Unification algorithm 42-3
term (means A-term) 1 historical note on 43
abstract 1 use in PT algorithm 34, 45-6
application I Unification theorem 42
closed 4 unifier 40
composite 1 most general (= m.g.u.) 41
construction-tree of 141-2 union of substitutions (U) 35
subterm 2 uniqueness of TA2-deductions
typable (see typable term) holds for nf's and Al-terms 22
typed (see typed term)
term models 59 vacuous discharging 16, 76, 78
term-variable 1 valid in a A-model 58
terminates, said of a reduction 5 variable
theorem (= provable formula) bound 2,68
in a Natural Deduction logic 79 changing (see a-conversion)
in an axiom-based logic 89 free 2, 68
ticket entailment logic (T.) 106 new 119
TMM#, TMMpn (term models) 59 propositional 75
TMpln] (for both TMp, TMpn) 59 term- 1
TNS(I-) 119-20 type- 12
trivial component of a substitution 35 interpreting in a A-model 58
TT(r) 64 typed 64
typable basis 159 variable-domain of a substitution 35
typable term variable-range of a substitution 35
in TA2 27 Vars( ) 13, 31, 149
in TA2+p, TA2+p,r 52
type-assignment (expression) 14 W (combinator) 4
type-assignment theory (see TA2) principal type of 18
Type-checking algorithm (see PT algo.) deduction for 161
type-context 14 (W) (axiom) 81, 90
for nf-schemes 119 weak normalization (see WN)
minimum, Con() 65 Weakening lemma 18
type-erasure 65 WN (= weak normalization) 27-8
type-invariance, strong 158 WN theorem
type-substitutions (see subs. into types) for TA2-typable terms 28
type-theories in general x for TA2+p and TA2+pn 55
Church-style and Curry-style 13-14 for typed terms 71
type-variables 12 historical note 72
interpretation in a A-model 58
typed Y (fixed-point combinator) 4
nf-scheme 120 untypability of 28
substitution 68
term 63-73, esp. 64
encoding a TA2-deduction 66 a-conversion 3
variable 64 of typed terms 69
a-conversion 69 a-invariance lemma for TA2 23
 Index 185

#-contraction 4 ii-Postponement theorem 8

of typed terms 69 n-redex 7
Q-conversion 5 typed 69
Church-Rosser theorem for 5 n-reduction 7
decidable in TAx 28 Church-Rosser theorem for 8
of typed terms 69 Commuting lemma with fl 8
rule (Eqf) in TA,t+p 52
with strong type-invariance 158 A-abstract 1
fl-equality (see fl-conversion) A-calculus, definition 1

#-expansion 5 basic facts 1-11

sometimes preserves types 24 A-model 57
fl-nf (see #-normal forms) extensional 58
fl-nf (set of all fl-normal forms) 6 A-term (see term)
structural characterization 7 Al-calculus, Al-terms 10
#-normal forms 6 Al-PT 94
structure of 7 AK-calculus, AK-terms 10
typed 69 AK-PT 94
structure of 110
long typed 110
structure of long typed 115 0 (Church numeral zero) 4
fl-redex 4 principal type of 50
typed 69 1-property 96
fl-redex-occurrence 144
1 (Church numeral 1) 4
argument-part 144
function-part 144
principal type of 50
newly-created 147
2-length of a position 129
residuals of 145
fl-reduction 4 oo (possible length of a reduction) 5
Church-Rosser theorem for 5 ao (possible no. of inhabitants) 109
leftmost red. 6
length of a red. 5 #( ) (cardinality of a set of terms) 109
non-cancelling red. 11
non-duplicating red. 11 (-+E), -* -elimination rule
preserves types (Subject-reduction thm.) 24 in TAx 16
red. with strong type-invariance 158 in propositional logics 76, 89
typed red. 69 (-I), - -introduction rule
fln-contraction 8 in TAx 16
fin-conversion 8
Church-Rosser theorem for 8 in propositional logics 76
rule (Egp,t) in TAx+Rn 52 discharging assumptions 76-9
Pry-equality (see Bn-conversion) (-'I)main, (-'I)vac 16
fln-nf (see Bn-normal forms)
fin-nf (set of all fln-normal forms) 9 * (for concatenation) 130
fln-normal forms 9 (symbol in a position) 140, 153
seeking 10 . (meaning #-normal form of) 6
typed 69
(meaning fln-normal form of) 9
fln-redex 8 .n (meaning n-normal form of) 9
#n-reduction 8
Church-Rosser theorem for 8
Commuting lemma for 8 (identity of terms etc.) 1

n-Postponement theorem for 8 - (a-convertibility) 3

r (for arbitrary type-contexts) 14 (isomorphism of components) 115, 143

r-x 1s
r )M 15 = (identity of nos., sets, etc.) 1

=ext 35
AL (see Curry-Howard mapping) (see #-conversion)
A2 (logic-to-lambda mapping) 82 =fin (see fln-conversion)
=9[n] (used for both =p, =pq) 53
n-contraction, conversion 7
=n (see n-conversion)
typed 69
n-family 8
rt9 (fl-contracts to, see fl-contraction)
typed 112
n-nf (see n-normal forms) r-p (/f-reduces to, see #-reduction)
n-normal forms 9 t>f,, (flu-reduces to, see flu-reduction)
typed 69 E>n (n-reduces to, see n-reduction)
186 Index

i 15 { / } (replacement notation) 67, 143

contrasted with H 19 { }, (q-family) 8, 112
I
(length of a deduction) 149
F-A 89 (length of a term) 2
HAD 104 of a typed term 67
Hint 79 I I (length of a type) 13
Hno eq 53 II II 13
HP 32
Hp 52 M1 (type-erasing) 65
Hs,t 52
HA (also see TAx, A2-deduction) 18 U (union of substitutions) 35

kp, kpn 58 o (composition of substitutions) 36

[ / ] (substitution notation) 3 (restriction of a substitution) 35

in typed terms 68 [ (restriction of a context) 15