ontological proof

© All Rights Reserved

31 vues

ontological proof

© All Rights Reserved

- A Proof Of God
- wnewsletter april 1-5 2019 2 autosaved autosaved
- Words and Worlds
- ch01
- Modal Logic
- Operation Research07.04.14
- Tut+4+-+Sol
- Angelelli the Hierarchy of Classes of Johann Caspar Sulzer 1755
- Gibbard (1974).pdf
- Symbolic Logic - VENN.pdf
- 1010.1282
- cis342-Ch02b
- What is Model Theory (Jhonatan Kirby)
- Nominal Restriction
- At least Some
- Lexical Semantics Klp 8
- First
- Bitsat 2014 Syllabus for English Proficiency and Logical Reasoning
- Givon Logic Pragmatics Epistemology
- A Study of Tarka and Its Role in Indian Logic

Vous êtes sur la page 1sur 198

Melvin Fitting

TYPES, TABLEAUS, AND GO DEL'S GOD

TRENDS IN LOGIC

Studia Logica Library

VOLUME 13

Managing Editor

Ryszard Wojcicki, Institute of Philosophy and Sociology,

Polish Academy of Sciences, Warsaw, Poland

Editors

Daniele Mundici, Department of Computer Sciences, University of Milan, Italy

Ewa Orlowska, National Institute of Telecommunications,

Warsaw, Poland

Graham Priest, Department of Philosophy, University of Queensland,

Brisbane, Australia

Krister Segerberg, Department of Philosophy, Uppsala University,

Sweden

Alasdair Urquhart, Department of Philosophy, University of Toronto, Canada

Heinrich Wansing, Institute of Philosophy, Dresden University of Technology,

Germany

SCOPE OF THE SERIES

Trends in Logic is a bookseries covering essentially the same area as the journal

Studia Logica - that is, contemporary formal logic and its applications and

relations to other disciplines. These include artificial intelligence, informatics,

cognitive science, philosophy of science, and the philosophy of language.

However, this list is not exhaustive, moreover, the range of applications, com-

parisons and sources of inspiration is open and evolves over time.

Volume Editor

Heinrich Wansing

The titles published in this series are listed at the end of this volume.

MELVIN FITTING

Lehman College and the Graduate Center,

City University of New York, U.S.A.

TYPES, TABLEAUS,

AND GO DEL'S GOD

~ ~ ~

''

KLUWER ACADEMIC PUBLISHERS

DORDRECHT/BOSTON/LONDON

A C.I.P. Catalogue record for this book is available from the Library of Congress.

ISBN 1-4020-0604-7

Published by Kluwer Academic Publishers,

P.O. Box 17, 3300 AA Dordrecht, The Netherlands.

Sold and distributed in North, Central and South America

by Kluwer Academic Publishers,

101 Philip Drive, Norwell, MA 02061, U.S.A.

In all other countries, sold and distributed

by Kluwer Academic Publishers,

P.O. Box 322, 3300 AH Dordrecht, The Netherlands.

Printed on acid-free paper

All Rights Reserved

2002 Kluwer Academic Publishers

No part of this work may be reproduced, stored in a retrieval system, or transmitted

in any form or by any means, electronic, mechanical, photocopying, microfilming, recording

or otherwise, without written permission from the Publisher, with the exception

of any material supplied specifically for the purpose of being entered

and executed on a computer system, for exclusive use by the purchaser of the work.

Printed in the Netherlands.

Contents

PREFACE Xl

Part I CLASSICAL LOGIC

1. CLASSICAL LOGIC-SYNTAX 3

1 Terms and Formulas 3

2 Substitutions 8

2. CLASSICAL LOGIC-SEMANTICS 11

1 Classical Models 11

2 Truth in a Model 12

3 Problems 15

3.1 Compactness 15

3.2 Strong Completeness 16

3.3 Weak Completeness 16

3.4 And Worse 17

4 Henkin Models 19

5 Generalized Henkin Models 24

6 A Few Technical Results 29

6.1 Terms and Formulas 29

6.2 Extensional Models 29

6.3 Language Extensions 30

3. CLASSICAL LOGIC-BASIC TABLEAUS 33

1 A Different Language 33

2 Basic Tableaus 35

3 Tableau Examples 37

v

vi TYPES, TABLEAUS, AND GODEL'S GOD

4. SOUNDNESS AND COMPLETENESS 43

1 Soundness 43

2 Completeness 46

2.1 Hintikka Sets 47

2.2 Pseudo-Models 48

2.3 Substitution and Pseudo-Models 52

2.4 Hintikka Sets and Pseudo-Models 59

2.5 Pseudo-Models are Models 62

2.6 Completeness At Last 63

3 Miscellaneous Model Theory 66

5. EQUALITY 69

1 Adding Equality 69

2 Derived Rules and Tableau Examples 69

3 Soundness and Completeness 73

6. EXTENSIONALITY 77

1 Adding Extensionality 77

2 A Derived Rule and an Example 77

3 Soundness and Completeness 79

Part II MODAL LOGIC

7. MODAL LOGIC, SYNTAX AND SEMANTICS 83

1 Introduction 83

2 Types and Syntax 86

3 Constant Domains and Varying Domains 89

4 Standard Modal Models 90

5 Truth in a Model 92

6 Validity and Consequence 94

7 Examples 95

8 Related Systems 101

9 Henkin/Kripke Models 102

8. MODAL TABLEAUS 105

1 The Rules 105

1.1 Prefixes 105

1.2 Propositional Rules 107

Contents vii

1.3 Modal Rules 107

1.4 Quantifier Rules 108

1.5 Abstraction Rules 109

1.6 Atomic Rules 109

1.7 Proofs and Derivations 110

2 Tableau Examples 111

3 A Few Derived Rules 113

9. MISCELLANEOUS MATTERS 115

1 Equality 115

1.1 Equality Axioms 115

1.2 Extensionality 117

2 De Re and De Dicta 118

3 Rigidity 121

4 Stability Conditions 124

5 Definite Descriptions 125

6 Choice Functions 128

Part III ONTOLOGICAL ARGUMENTS

10. GODEL'S ARGUMENT, BACKGROUND 133

1 Introduction 133

2 Anselm 134

3 Descartes 134

4 Leibniz 137

5 Godel 138

6 Godel's Argument, Informally 139

11. GODEL'S ARGUMENT, FORMALLY 145

1 General Plan 145

2 Positiveness 145

3 Possibly God Exists 150

4 Objections 152

5 Essence 156

6 Necessarily God Exists 160

7 Going Further 162

7.1 Monotheism 162

viii TYPES, TABLEAUS, AND GODEL'S GOD

8

9

10

11

7.2 Positive Properties are Necessarily Instantiated

More Objections

A Solution

Anderson's Alternative

Conclusion

REFERENCES

INDEX

162

163

164

169

171

173

179

Truth did not come into the world naked,

but it came in types and images.

One will not receive truth in any other way.

The Gospel of Philip

[Rob77]

Preface

What's Here

This is a book about intensional logic. It also provides a thorough

look at higher-type classical logic, including tableaus and a complete-

ness proof for them. It also provides a formal examination of the Godel

ontological argument. These are not disparate topics. Higher-type clas-

sical logic is intensional logic with the intensional features removed, so

this is a good place to start. Ontological arguments, Godel's in partic-

ular, are natural examples of intensional logic at work, so this is a good

place to finish.

The term formal logic covers a broad range of inventions. At one

end are small, special-purpose systems; at the other are rich, expressive

ones. Higher-type modal logic-intensional logic-is one of the rich ones.

Originating with Carnap and Montague, it has been applied to provide

a semantics for natural language, to model intensional notions, and to

treat long-standing philosophical problems. Recently it has also supplied

a semantic foundation for some complex database systems. But besides

being rich and expressive, it is also tremendously complex, and requires

patience and sympathy on the part of its students.

There are two quite different reasons to be interested in a logic. There

is its formal machinery for its own sake, and there is using the formal

machinery to address problems from the outside world. The mechanism

of higher-type modal logic is complex and requires serious mathematics

to develop properly. Models are not simple to define, and tableau sys-

tems are quite elaborate. A completeness argument, to connect the two,

is difficult. But, the machinery is of considerable interest, if this is the

sort of thing you have a considerable interest in. If you are such a reader,

applications concerning the existence of God can simply be skipped. On

the other hand, if philosophical applications are what you are after, the

xi

Xll TYPES, TABLEAUS, AND GO DEL'S GOD

Godel ontological argument is a prime example. If this is the kind of

reader you are, much of the mathematical background can be taken on

faith, so to speak. It is a rare reader who will be interested equally in

both the formal and the applied aspects of intensional logic. In a sense,

then, this book has no audience--there are separate audiences for dif-

ferent parts of it. (But I encourage these audiences to do some 'crossing

over.')

If you are interested in ontological arguments for their own sakes, start

with Part III, and pick up material from earlier chapters as it is needed.

If you are interested in the mathematical details of the formal system,

its semantics and its proof theory, Parts I and II will be of interest-you

can skimp on reading Part III. Part I is entirely devoted to classical

logic, and Part II to modal. Here is a more detailed summary.

Part I presents higher-type classical logic. It begins with a discussion

of syntax matters, Chapter 1. I present types in Schutte's style, rather

than following Church. Types can be somewhat daunting and I've tried

to make things go as smoothly as I can.

Chapter 2 examines semantics in considerable detail. What are some-

times called "true" higher-order models are presented first. After this,

Henkin's generalization is given, and finally a non-extensional version of

Henkin models is defined. Henkin himself mentioned such models, but

knowledge of them does not seem to be widespread. They are natural,

and should become more familiar to the logic community-the philo-

sophical logic community in particular.

Classical higher-order tableaus are formulated in Chapter 3. These

are not original here--versions can be found in several places. A number

of worked out examples of tableau proofs are given, and more are in

exercises. The system is best understood if used. I do not attempt a

consideration of automation-the system is designed entirely for human

application. There is even some discussion of why.

Soundness and completeness are proved in Chapter 4. Tableaus are

complete with respect to non-extensional Henkin models. The com-

pleteness argument is not original; it is, however, intricate, and detailed

proofs are scarce in the literature.

After the hard work has been done, equality and extensionality are

easy to add using axioms, and this is done in Chapters 5 and 6. And this

concludes Part I. Except for the explicit formulation of non-extensional

models, the material in Part I is not original-see [Tak67, Pra68, Tol75,

And86, Sha91, Lei94, Koh95, Man96], for example.

Part II is devoted to the complications that modality brings. Chap-

ter 7 adds the usual box and diamond to the syntax, and possible worlds

PREFACE xiii

to the semantics. It is now that choices must be made, since quantified

modal logic is not a thing, but a multitude.

First, at ground level quantifiers could be actualist or possibilist-

they can range over what actually exists at a world, or over what might

exist. This corresponds to the varying domain, constant domain split

familiar to many from first-order modal discussions. However, either an

actualist or a possibilist approach can simulate the other. I opt for a

possibilist approach, with an explicit existence predicate, because it is

technically simpler.

Next, we must go up the ladder of higher types. Doing so exten-

sionally, as in classical logic, means we take subsets of the ground-level

domain, subsets of these, and so on. Going up intensionally, as Montague

did, means we introduce functions from possible worlds to sets of ground-

level objects, functions from possible worlds to sets of such things, and

so on. What is presented here mixes the two notions-both extensional

and intensional objects are present. I refer you to [FitOOb, FitOOa] for ap-

plications of these ideas to database theory-intensional and extensional

objects make natural sense even in such a context.

Classical tableau rules are adapted in Chapter 8, using prefixes, to pro-

duce modal systems. While the modal tableau rules are rather straight-

forward, they are new to the literature, and should be of interest. Since

things are already quite complex, no completeness proof is given. If it

were given, it would be a direct extension of the classical proof of Part I.

Using modal semantics and tableaus, in Chapter 9 I discuss the re-

lationships between rigidity, de re and de dicto usages, and what I call

Godel's stability conditions, which arise in his proof of the existence of

God. I also relate all this to definite descriptions. While this is not deep

material, much of it does not seem to have been noted before, and many

should find it of some significance.

Finally, Part III is devoted to ontological proofs. Chapter 10 gives

a brief history and analysis of arguments of Anselm, Descartes, and

Leibniz. This is followed by a longer, still informal, presentation of

the Godel argument itself. Formal methods are applied in Chapter 11,

where Godel's proof is examined in great detail. While Godel's argument

is formally correct, some fundamental flaws are pointed out. One, noted

by Sobel, is that it is too strong-the modal system collapses. This

could be seen as showing that free will is incompatible with Godel's

assumptions. Some ways out of this are explored. Another flaw is equally

serious: Godel assumes as an axiom something directly equivalent to a

key conclusion of his argument. The problematic axiom is related to a

principle Leibniz proposed as a way of dealing with a hole he found in

an ontological proof of Descartes. Descartes, Leibniz, and Godel (and

XlV TYPES, TABLEAUS, AND GODEL'S GOD

also Anselm) all have proofs that stick at the same point: showing that

the existence of God is possible.

If the Godel argument is what you are interested in, start with Part III,

and pick up earlier material as needed. Many of the uses of the formal-

ism are relatively intuitive. Indeed, in Godel's notes on his ontological

argument, formal machinery is never discussed, yet it is possible to get

a sense of what it is about anyway.

How Did This Get Written?

Having just completed work on a book about first-order modal logic,

[FM98], a look at higher-order modal logic suggested itself. I thought I

would use Godel's ontological argument as a paradigm, because it is one

of the few examples I have run across that makes essential use of higher-

order modal constructs. Godel's argument for the existence of God is not

particularly well-known, but there is a growing body of literature on it.

This literature sometimes gives formalizations of Godel's rather sketchy

ideas-generally along natural deduction or axiomatic lines. My idea

was, I would design a tableau system within which the argument could

be formalized, and this might lead to a nice paper illustrating the use of

tableau methods. First, give tableau rules, then give Godel's proof.

One cannot really develop semantic tableaus without a semantics be-

hind it. The semantics of higher-order modal logic turned out to be

of considerable intricacy, far beyond what could even be sketched in a

paper. Clearly, an extended discussion of the semantics for higher-order

modal logic was needed before the tableau rules could be motivated.

I soon realized that in presenting higher-order modal logic, I was try-

ing to explicate ideas corning from two quite different sources. On the

one hand, there are essentially modal problems, some of which already

arise at the first-order level and have little to do with higher-order con-

structs. On the other hand, a number of higher-order modal complexities

also manifest themselves in a classical setting, and can be discussed more

clearly without modalities complicating things. So I decided that before

modal operators were introduced, I would give a thorough presentation

of a semantics and tableau system for higher-order classical logic. There

are already treatments of tableau, or Gentzen, systems for higher-order

classical logic in the literature, but I felt it would be useful to give things

in full hr"e. Detailed completeness proofs are hard to find, for instance.

Higher-order classical logic already has its hidden pitfalls. It is com-

mon knowledge, so to speak, that "true" higher-order classical models

cannot correspond to any proof procedure. Henkin models are what is

needed. But a "natural" formulation of tableaus is not complete with

respect to Henkin models either. This is something known to experts-

PREFACE XV

it was not known to me when I started this book. A broader notion

of Henkin model (also due to Henkin) is needed, a non-extensional ver-

sion. Such models should be better known since they are actually quite

plausible things, and address problems that, while not common in math-

ematics, do arise in linguistic applications of logic.

In the 1960's, cut-elimination theorems were proved for higher-order

classical logic, using semantic methods that relied on non-extensional

models. In effect, these cut-elimination proofs concealed a completeness

argument within them, but the general notion of non-extensional model

was not formulated abstractly--only the specific structure constructed

by the completeness argument was considered. In short, a completeness

theorem was never stated, only a consequence, albeit a very important

one. So I found myself required to formulate a general notion of classical

non-extensional Henkin model, then prove completeness for a suitable

classical tableau system. After this I could move on to discuss modality.

What sort of modal features did I want? Formalizations of the Godel

argument by others had generally used some version of an intensional

logic, with origins in work of Carnap, [Car 56], developed and applied by

Montague, [Mon60, Mon68, Mon70], and formally elaborated in [Gal75].

After several preliminary attempts I decided this logic was not quite

what I wanted. In it, semantically speaking, all objects are intensional.

I decided I needed a logic containing both intensional and extensional ob-

jects. Of course, one could bring extensional objects into the Montague

setting by identifying them with objects that are rigid, in an appropriate

sense, but it seemed much more natural to have extensional objects from

the start. Thus the modal logic given in the second part of this book is

somewhat different from what has been previously considered.

Once I had formulated the modal logic I wanted, tableau rules were

easy, and I could finally formalize the Godel argument. What began as

a short paper had turned into a book. My after-the-fact justification is

that there are few treatments of higher-order logic at all, and fewer still

of higher-order modal logic. It is a rare flower in a remote field. But it

is a pretty flower.

Acknowledgments

An earlier draft of this work was on my web page for some time, and I

was given several helpful suggestions as a result. In particular I want to

thank Peter Hajek, Oliver Kutz, Paul Gilmore, and especially Howard

Sobel.

I

CLASSICAL LOGIC

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 1

CLASSICAL LOGIC-SYNTAX

1. Terms and Formulas

The formulation of a higher-order logic allows some freedom-there

are certain places where choices can be made. Several of these choices

produce equivalent results. Before getting to the formal machinery, I

informally set out my decisions on these matters. Other treatments may

make different choices, but ultimately it is largely a matter of conve-

nience that is involved.

Often classical first-order logic is formulated with a rich variety of

terms, built up from constant symbols and variables using function sym-

bols. Since higher-order constructs are already complicated, I have de-

cided to have constant symbols but not function symbols. If necessary

for some purpose, it is not a major issue to add them-doing so yields

a conservative extension.

Higher-order logic can be formulated with or without explicit abstrac-

tion machinery. Speaking informally, one wants to make sure that every

formula specifies a class, but there are two ways of making this happen.

One is to assume comprehension axioms, formulas of the general form:

where cp(x1, ... , xn) is a formula with free variables as indicated. Such

axioms ensure that to each formula corresponds an 'object.' The other

approach is to elaborate the term-forming machinery, so that there is

an explicit name for the object specified by a formula cp. This involves

predicate abstraction, or >-.-abstraction:

3

4 TYPES, TABLEAUS, AND GODEL'S GOD

The two approaches are equivalent in a direct way. I have chosen to use

explicit abstracts for several reasons. First, axioms are not as natural

when tableau systems are the proof machinery of choice. And second,

predicate abstraction has already played a major role in earlier inves-

tigations of modal logic [FM98], and makes discussion of major issues

considerably easier here.

Finally, one can characterize higher-order formulas more-or-less the

way it is done in the first-order setting, taking quantifiers and connec-

tives as "logical constants." This is the approach of [Sch60]. Alterna-

tively, following [Chu40], one can think of quantifiers and connectives as

constants of the language, which itself is formulated in lambda-calculus

style. In this book I take the first approach, though one can make

arguments for the second on grounds of elegance and economy. My jus-

tification is that doing things the way that has become standard for

first-order logic will be less confusing to the reader.

Recently one further alternative has become available. In [Gil99,

GilOl], Paul Gilmore has shown that by a relatively simple change, a

system of classical higher-order logic can be developed allowing a con-

trolled degree of impredicativity-typing rules can be relaxed to permit

the formation of certain useful sentences that are not "legal" in the ap-

proach presented here. This, in turn, allows a more natural development

of arithmetic in the higher-order setting. I do not follow Gilmore's ap-

proach here, but I recommend it for study. Much of what I develop

carries over quite directly.

So these are my choices: no function symbols, explicit predicate ab-

straction, quantifiers and connectives as in the first-order setting, and

no impredicativity. With this out of the way I can begin presenting the

formal syntactical machinery.

In first-order logic, relation symbols have an arity--some are one-

place, some are two-place, and so on. In higher-order logic this simple

idea gets replaced by a typing mechanism, which is considerably more

complex. Terms, and certain other items, are assigned types, and rules

of formation make use of these types to ensure that things fit together

properly. I begin by saying what the types are.

DEFINITION 1.1 (TYPE) 0 is a type. If t1, ... , tn are types, (tl, ... , tn)

is a type. I generally use t, t1, t2, t', etc. to represent types.

An object of type 0 is intended to be a ground-level object-it corre-

sponds to the designation of a constant symbol or variable in standard

first-order logic. An object of type (t

1

, ... , tn) is a predicate that takes

n arguments, of types t

1

, ... , tn respectively. Thus a constant symbol

of type (0, 0, 0), say, would be called a three-place relation symbol in

CLASSICAL LOGIC-SYNTAX 5

standard first-order logic-it applies to three ground-level arguments.

But now we can have relation symbols of types such as ( (0), (0, 0), 0), to

which nothing in first-order logic corresponds.

DEFINITION 1.2 (L(C)) Let C be a set of constant symbols with a type

associated to each, containing at least an equality symbol =(t,t) for each

type t. I denote the classical higher-order language built up from C by

L( C). The rest of this section amounts to the formal characterization of

L(C).

For each type t I assume there are infinitely many variable symbols

of that type. I generally use letters from the beginning of the Greek

alphabet to represent variables, with the type written as a superscript:

at, j3t, '"'/, . . . . Likewise I generally use letters from the uppercase

Latin alphabet as constant symbols, again with the type written as a

superscript: At, Bt, ct, Dt, .... As noted, equality is primitive, so for

each type t there is a constant symbol =(t,t) of type (t, t). Often types

can be inferred from context, and so superscripts will be omitted where

possible, in the interests of uncluttered notation.

Sometimes it is helpful to refer to the order of a term or formula-first-

order, second-order, and so on. It is simplest to define this terminology

first for types themselves.

DEFINITION 1.3 (ORDER) Type 0 is of order 0. Type (t1, ... , tn) has

as its order the maximum of the orders of t1, ... , tn, plus one.

Thus (0, 0) is of order 1, or first-order. Likewise (0, (0, 0)) is of order

2, or second-order. Types will play the fundamental role, but order

provides a convenient way of referring to the maximum complexity of

some construct. When I talk about the order of a constant or variable,

I mean the order of its type. Likewise once formulas are defined, I may

refer to the order of the formula, by which I mean the highest order of

a typed part of it.

Next I define the class of formulas, and their free variables. This

definition is more complex than the corresponding first-order version

because the notion of term cannot be defined first; both term and formula

must be defined together. And to define both, I need the auxiliary notion

of predicate abstract which is, itself, part of a mutual recursion involving

Definitions 1.4, 1.5, and 1.6.

DEFINITION 1.4 (PREDICATE ABSTRACT OF L(C)) Let <f? be a for-

mula of L(C) and a1, ... , an be a sequence of distinct variables of

types h, ... , tn respectively. (.Aa1, ... , an.<P) is a predicate abstract of

6 TYPES, TABLEAUS, AND GODEL'S GOD

L( C). Its type is (t1, ... , tn), and its free variable occurrences are the

free variable occurrences in the except for occurrences of the

variables a1, . . . , an.

DEFINITION 1.5 (TERM OF L(C)) Terms of each type are character-

ized as follows.

1 A constant symbol of L(C) or variable is a term of L(C). If it is a

constant symbol, it has no free variable occurrences. If it is a variable,

it has one free variable occurrence, itself.

2 A predicate abstract of L( C) is a term of L( C). Its free variable

occurrences were defined above.

T is used, with and without subscripts, to stand for terms.

DEFINITION 1. 6 (FORMULA OF L( C)) The notion of formula is given

as follows.

1 If T is a term of type (t1, ... , tn), and TI, ... , Tn is a sequence of

terms of types t1, ... , tn respectively, then r( TI, ... , Tn) is a formula

(atomic) of L( C). The free variable occurrences in it are the free

variable occurrences of r, n, ... , Tn

2 is a formula of L( C) so is The free variable occurrences of

are those of<>.

3 and w are formulas of L(C) so is 1\ w). The free variable

occurrences of ( 1\ w) are those together with those of w.

4 is a formula of L( C) and a is a variable then is a formula

of L(C). The free variable occurrences are those except

for occurrences of a.

EXAMPLE 1. 7 Suppose a(o,o} is a variable of type (0, 0) (and so first-

order), (3 is a variable of type 0, and 'Y((O,O},O} is a variable of type

((0,0),0) (second-order). Both (3 and 'Y((O,O},o} are terms. Then the

expression 'Y( (o,o} ,o} (a (O,O}, (3) is an atomic formula. Generally I will

write the simpler looking 'Y( a, (3), and give the information contained

in the superscripts in a separate description. Since this atomic formula

contains a variable 'Y of order 2, it is referred to as a second-order atomic

formula.

DEFINITION 1.8 (SENTENCE) A formula with no free variables is a sen-

tence.

CLASSICAL LOGIC-SYNTAX 7

One can think of V, ::=>,:=,and 3 as defined symbols, with their usual

definitions. But sometimes it is convenient to take them as primitive-!

do whatever is most useful at the time. Also square and curly paren-

theses are used, as well as the official round ones, to aid readability.

And finally, I write the equality symbol in. infix position, following stan-

dard convention. Thus, for example, I write (a/ =(t,t) f3t) in place of

=(t,t) (

0

t, f3t).

Several examples involving just first and second-order notions will be

considered, so a few special alphabets are introduced informally, to make

reading the examples a little easier.

Order Constants Variables

0 a, b, c, ... x, y, z, ...

1 A,B,C, ... X, Y,Z, ...

2 A,B,C, ... X,Y,Z, ...

ExAMPLE 1.9 For this example I give explicit type information (in su-

perscripts), until the end of the example. After this I omit the super-

scripts, and say in English what is needed to restore them.

Suppose x

0

, X(O), and X((O)) are variables (the first is of order 0, the

second is of order 1, and the third is of order 2). Also suppose p((O)) and

g

0

are constant symbols of L(C) (the first is of order 2 and the second

is of order 0).

1 Both X((O))(X(

0

)) and X(

0

)(x

0

) are atomic formulas. All variables

present have free occurrences.

2 (>..X((o)).X((O))(X(

0

))) is a predicate abstract, of type (((0))). Only

the occurrence of X(O) is free.

3 SinceP((O)) isoftype ((0)), (>..X((O)).X((O))(X(

0

)))(P((O))) is a formula.

Only X(O) is free.

4 [(>..X((O)) .X((O))(X(

0

)))(P((O)))::) X(

0

)(x

0

)] is a formula. The only free

variable occurrences are those of X(

0

) and x

0

.

5 (VX(

0

))[(>..X((O)).X((O))(X(

0

)))(P((O))) ::=> X(

0

)(x

0

)] is a formula. The

only free variable occurrence is that of x

0

.

6 (>..x

0

.(\fX(

0

))[(>..X((O)).X((O))(X(

0

)))(P((O))) ::=> X(

0

)(x

0

)]) is a predi-

cate abstract. It has no free variable occurrences, and is of type (0).

The type machinery is needed to guarantee that what is written is well-

formed. Now that the exercise above has been gone through, I will

display the predicate abstract without superscripts, as

(>..x.('v'X)[(>..X.X(X))(P) ::=> X(x)]),

8 TYPES, TABLEAUS, AND GODEL'S GOD

leaving types to be inferred, or explained in words, as necessary.

In first-order logic, facts about formulas and terms are often proved

by induction based on complexity. And complexity for a formula is often

measured by the number of logical connectives and quantifiers, or what

amounts to the same thing, by how "far away" the formula is from being

atomic. In a higher-order setting these notions diverge.

DEFINITION 1.10 (DEGREE) By the degree of a formula or term is

meant the number of propositional connectives, quantifiers, and lambda-

symbols it contains.

Note that since an atomic formula can involve terms containing pred-

icate abstracts which, in turn, involve other formulas, the degree of an

atomic formula need not be 0, as in the first-order case.

2. Substitutions

Formulas can contain free variables, and terms that are very complex

can be substituted for them. The notion of substitution is a fundamental

one, and this section is devoted to it. In a general way, I follow the

treatment in [Fit96].

DEFINITION 1.11 (SUBSTITUTION) A substitution is a mapping from

the set of variables to the set of terms of L( C) such that variables of

type t map to terms of type t.

I generally denote substitutions by u, with and without subscripts.

Also I generally write xu rather than u(x). Most concern is with substi-

tutions having finite support, that is, they are the identity on all but a

finite number of variables. A special notation is used for the finite sup-

port substitution that maps each O:i to Ti and is the identity otherwise:

{ o:l/r1, ... , O:n/Tn}

The action of substitutions on variables is readily extended to terms

and formulas generally.

DEFINITION 1.12 For a substitution u, by Ua

1

, ... ,an is meant the substi-

tution that is like u except that it is the identity on 0:1, . . . , O:n.

DEFINITION 1.13 Let u be a substitution. The action of u is extended

recursively as follows.

1 Au = A for a constant symbol A.

2 (.Ao:1, .. , O:n.<I>)u = (.Ao:1, .. , O:n.<I>Ua1, ... ,an)

CLASSICAL LOGIC-SYNTAX

3 [7(71, ... , 7n)]u = 70"(710", ... , 7n0").

4 [<P]u = [<Pu].

5 (<P 1\ w)u = (<Pu 1\ \llu).

6 [('v'a)<P]u = ('v'a)[<Pua]

9

EXAMPLE 1.14 Let <P be the formula (3a((O)))[o/(

0

))((>.,6.')'(

0

)(,6)))]

and let u be any substitution such that 'Y(o) u = 7(o). I compute <Pu.

For convenience I omit type-indicating superscripts, but even so, the

notion is a bit much. Sorry.

(3a) [a( (>.,6. 1(,6))) ]u = (3a) [a( (>.,6. 1(,6))) ]u a

= (3a)[aua( (>.,6.')'(,6))ua)]

= (3a)[a( (>.,6.')'(,6))ua)]

= (3a)[a( (>.,6.['Y(,6)]ua,{J) )]

= (3a)[a( (>.,6.(!'ua,{J)(,6ua,{J)) )]

= (3a)[a( (>.,6.7(,6)) )]

The connection between substitutions and free variable occurrences is

simple: it is only the free occurrences that can be changed by substitu-

tions. I leave the proof to you as an exercise.

PROPOSITION 1.15 Let 0"1 and 0"2 be substitutions.

1 If u1 and u2 agree on the free variables of the term 7 then 70"1 = 70"2.

2 If 0"1 and u2 agree on the free variables of the formula <P then <Pu1 =

<Pu2.

Not all substitutions are appropriate in all settings; some do not prop-

erly respect the role of bound variables, in the sense that they may re-

place a free occurrence of a variable in a formula with another variable

that is "captured" by a quantifier or predicate abstract of the formula.

The substitutions that are acceptable are called free substitutions. These

play a significant role throughout what follows.

DEFINITION 1.16 (FREE SUBSTITUTION) The following character-

izes when a substitution u is free for a formula or term.

1 u is free for a variable or constant.

2 O" is free for (>.a1, ... , an.<P) if O"a1, ... ,an is free for <P, and if ,6 is

any free variable of (>.a1, ... , an.<P) then ,6u does not contain any of

a1, ... , an free.

10 TYPES, TABLEAUS, AND GODEL'S GOD

3 a is free for ,q, if a is free for <P.

4 a is free for ( <P 1\ 'l1) if a is free for <P and a is free for 'l1.

5 a is free for (Ya)<P if CJ

01

is free for <P, and if f3 is a free variable of

(Ya)<P then f3a does not contain a free.

With the action of substitutions extended to all terms, composition

of substitutions is easily defined.

DEFINITION 1.17 (SUBSTITUTION COMPOSITION) Let a

1

and a

2

be

substitutions. Their composition is the mapping defined by: a(aw2) =

(aa1)a2, for variables a.

It is not generally the case that T(a1a2) = (Ta1)a2, for terms T, and

similarly for formulas. But it is when appropriate freeness conditions

are imposed.

THEOREM 1.18 Substitution is under the following cir-

cumstances.

1 If a1 is free for the formula <P, and a2 is free for the formula <Pa1,

then (<Pa1)a2 = <P(aw2).

2 If a1 is free for the term T, and a2 is free for the term TCJI, then

(Ta1)a2 = T(a1a2).

The proof of this is essentially the same as in the first-order setting.

Rather than giving it here, I refer you to the proof of Theorem 5.2.13 in

[Fit96].

Exercises

EXERCISE 2.1 Prove Proposition 1.15 by induction on degree. Conclude

that if <P is a sentence then <Pa = <P for every substitution a.

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 2

CLASSICAL LOGIC-SEMANTICS

1. Classical Models

Defining the semantics of any higher-order logic is relatively compli-

cated. Since modalities add special complexities, it is fortunate I can

discuss underlying classical issues before bringing them into the picture.

In this Chapter the "real" notion of higher-order model is defined first,

and truth in them is characterized. Then Henkin's modification of these

models is considered-sometimes these are called general models-as

well as a non-extensional version of them.

I don't want just syntactic objects, terms, to have types. I want

sets and relations to have them too. After all, we think of terms as

designating sets and relations, and we want type information to move

back and forth between syntactic object and its designation.

DEFINITION 2.1 (RELATION TYPES) Let 8 be a non-empty set. For

each type t the collection [t, S] is defined as follows.

1 [0,8] = s.

2 [ (t1, ... , tn), S] is the collection of all subsets of [t1, S] X x [tn, S].

0 is an object of type t over S if 0 E [t, S]. 0 is systematically used,

with or without subscripts, to stand for objects in this sense.

For example, a member of [ (0, 0), S] is a subset of S x S, and in

standard first-order logic it would simply be called a two-place relation

on S. But now relations of relations are allowed, and even more complex

things as well, so terminology gets more complicated.

A classical model consists of an underlying domain, thought of as the

"ground level objects," and an interpretation, assigning some denota-

11

12 TYPES, TABLEAUS, AND GO DEL'S GOD

tion in the model to each constant symbol of the language. But that

denotation must be consistent with type information.

DEFINITION 2.2 (CLASSICAL MODEL) A higher-order classical model

for L(C) is a structure M = (D,I), where D is a non-empty set called

the domain of the model, and I is a mapping, the interpretation, meeting

the following conditions.

1 If A is a constant symbol of L(C) of type t, I(A) E [t, D].

2 If = is the equality constant symbol of type (t, t) then I(=) is the

equality relation on [t, D].

2. Truth in a Model

Assume M = (D,I) is a classical model for a language L(C). It is

time to say which sentences of the language, or more generally, which

formulas with free variables, are true in M. This is symbolized by M lf-v

<P. Informally it can be read: the formula <Pis true in the model M, with

respect to the valuation v which assigns meanings to free variables. But

as will be seen, to properly define this one must also assign denotations

to all terms. The denotation of a term of type t will be an object of

type t over D. And this can not be done first, independently. The

assignment of denotations to terms, and the determination of formula

truth constitutes a mutually recursive pair of definitions, just as was the

case for the syntactic notions of term and formula in Section 1. Still, it

is all rather straightforward.

DEFINITION 2.3 (VALUATION) The mapping v is a valuation in the

classical model M = (D,I) if v assigns to each variable o/ of type t

some object of type t, that is, v((i) E [t, D].

DEFINITION 2.4 (VARIANT) A valuation w is an o:-variant of a valua-

tion v if v and w agree on all variables except possibly o:. More generally,

w is an 0:1, ... , O:n-variant if v and w agree on all variables except pos-

sibly 0:1, ... , O:n

Now, the following two definitions constitute a single recursive char-

acterization.

DEFINITION 2.5 (DENOTATION OF A TERM) Let M = (D,I) be a

classical model, and let v be a valuation in it. A mapping is defined,

( v *I), assigning to each term of L( C) a denotation for that term.

1 If A is a constant symbol of L(C) then (v * I)(A) =I( A).

CLASSICAL LOGIC-SEMANTICS

2 Ifa is a variable then (v*I)(a) =v(a).

3 If (.Aa1, ... ,an.q,) is a predicate abstract of L(C) of type t, then

(v *I)( (.Aa1, ... , an.q,)) is the following member of [t, V]:

{ (w(a1), ... , w(an)) Jw is an a1, ... , an variant of v

and M if-w q,}

13

DEFINITION 2.6 (TRUTH OF A FORMULA) Again let M = (V,I) be a

classical model, and let v be a valuation in it. The notion of formula q,

of L(C) being true in model M with respect to v, denoted M if-v q,, is

characterized as follows.

1 For terms T, T1, ... ,Tn, M if-v T(T1, ... ,Tn) provided

( ( v * I) ( T1) , .. . , ( v * I) ( T n)) E ( v * I) ( T) .

2 M if-v --,q, if it is not the case that M if-v q,,

3M if-v q, 1\ 'lt if M if-v q, and M if-v 'lt.

4 M if-v (Va)q, if M if-v' q, for every a-variant v' of v.

There is an alternative notation that makes evaluating the truth of

formulas in models somewhat easier.

DEFINITION 2. 7 (SPECIAL NOTATION) Suppose v is a valuation, and w

is the a1, ... , an variant ofv such that w(al) = 01, ... , w(an) =On.

Then, if M if-w q, this may be symbolized by

Now part 3 of Definition 2.5 can be restated as follows.

3 (v*I)((.Aa1, ... ,an,q,)) =

{ (01, ... , On) J M if-v q,[al/01, ... , an/On]}

Likewise part 4 of Definition 2.6 becomes

4 M if-v (Va)q, if M if-v q,[a/0] for every object 0 of the same type as

a.

Defined symbols l i k ~ and 3 have their expected behavior, which are

explicitly stated below. Alternately, this can be considered an extension

of the definition above.

5 M if-v q, V 'lt if M if-v q, or M if-v 'lt.

6 M if-v q, ~ 'lt if M if-v q, implies M if-v 'lt.

14 TYPES, TABLEAUS, AND GODEL'S GOD

7 M lf--v <I>= W if M lf--v <I> iff M lf--v W.

8 M lf--v (3a)<P if M lf--v' <I> for some a-variant v' of v; equivalently if

M lf--v <P[a/0] for some object 0 of the same type as a.

As in first-order logic, if <I> has no free variables, M lf--v <I> holds for

some v if and only if it holds for every v. Thus for sentences (closed

formulas), truth in a model does not depend on a choice of valuation.

DEFINITION 2.8 (VALIDITY, SATISFIABILITY, CONSEQUENCE) Let <l>

be a formula and S be a set of formulas.

1 <I> is valid if M lf--v <I> for every classical model M and valuation v.

2 S is satisfiable if there is some model M and some valuation v such

that M If-v <p for every <p E S.

3 <I> is a consequence of S provided, for every model M and every

valuation v, if M lf--v <p for all <pES, then M lf--v <I>.

The definitions above are of some complexity. Here is an example to

help clarify their workings.

EXAMPLE 2.9 This example shows a formula that is valid and involves

equality. In it, cis a constant symbol of type 0.

The expression (AX.(:lx)X(x)) is a predicate abstract of type ((0)),

where X is of type (0) and x is of type 0. Intuitively it is the "being

instantiated" predicate. Likewise the expression (>.x.x = c) is a predicate

abstract of type (0), where x and care of type 0. Intuitively this is the

"being c" predicate. Since this predicate is, in fact, instantiated (by

whatever c designates), the fir:;;t predicate abstract correctly applies to

it. That is, one should have the validity of the following.

(>.X.(3x)X(x))((>.x.x =c)) (2.1)

I now verify this validity. Suppose there is a model M = (V, I). I show

the formula is true in M with respect to an arbitrary valuation v. To do

this, I investigate the behavior, in M, of parts of the formula, building

up to the whole thing.

First, recalling that the interpretation of an equality symbol is by the

equality relation of the appropriate type, we have the following.

(v *I)((>.x.x =c))= {o I M lf--v (x = c)[x/o]}

= {o I o =I(c)}

= {I(c)}

CLASSICAL LOGIC-SEMANTICS

We also have the following.

(v *I)((.AX.(3x)X(x))) = {0 I M 11-v (3x)X(x)[Xj0]}

= {0 I M 11-v X(x)[X/O,xjo] for some o}

= { 0 I o E 0 for some o}

= {o I o # 0}

Now we have (2.1) because

M 11-v (.AX.(3x)X(x))((.Ax.x =c)):}

15

(v *I)((.Ax.x =c)) E (v *I)((.AX.(3x)X(x))) :} {I(c)} E {0 I 0 # 0}.

You might try verifying, in a similar way, the validity of the following .

.

(.AX.(3x)X(x))((.Ax.(x = x)))

3. Problems

First-order classical logic has many nice features that do not carry

over to higher-order versions. This is well-known, and partly accounts

for the general emphasis on first-order. I sketch a few of the higher-order

problems here.

3.1 Compactness

The compactness theorem for first-order logic says a set of formu-

las is satisfiable if every finite subset is. This is a fundamental tool

for the construction of models of various kinds-non-standard models

of analysis, for instance. The higher-order analog does not hold, and

counter-examples are easy to come by. Here is one.

The Dedekind characterization of infinity is: a set is infinite if it can be

put into a 1-1 correspondence with a proper subset. Consequently, a set

is finite if any 1-1 mapping from it to itself can not be to a proper subset,

i.e. must be onto. This can be said easily, as a second-order formula.

Since function symbols are not available, I make do with relation symbols

in the usual way-the following formula is true in a model if and only if

the domain of the model is finite.

(\fX)[(function(X) 1\ one-one(X)) :J onto(X)] (2.2)

In (2.2) the following abbreviations are used.

function(X) for (\fx)(3y)(\fz)[X(x, z) = (z = y)]

one-one(X) for (\fx)(\fy)(\fz){[X(x, z) 1\ X(y, z)] :J (x = y)}

onto(X) for (\fy)(3x)X(x, y)

16 TYPES, TABLEAUS, AND GODEL'S GOD

Also, define the following infinite list of formulas, where x ol y abbrevi-

ates (x = y).

A2 = (:3xl)(:Jx2)[xl ol x2]

A3 = (:Jx1)(:Jx2)(:lx3)[(x1 ol x2) 1\ (x1 ol x3) 1\ (x2 ol x3)]

So An is true in a model if and only if the domain of the model contains

at least n members.

Now, the set consisting of (2.2) and all of A2, A3, ... , is certainly

not satisfiable, but every finite subset is, so compactness fails. (In first-

order classical logic this example turns around, and shows finiteness has

no first-order characterization.)

3.2 Strong Completeness

A proof procedure is said to be (sound and) strongly complete if <.!.>

has a derivation from a set S exactly when <.!.> is a logical consequence

of S. Classical first-order logic has many proof procedures that are

strongly complete for it, but there is no such proof procedure for higher-

order logic. To see this, one doesn't need an exact definition of proof

procedure-it is enough that proofs be finite objects.

Let S be the set of formulas defined in Section 3.1, a set that is not

satisfiable though every finite subset is. And let j_ be <.!.> 1\ <.!.>, for some

formula <.!.>. The formula j_ is a logical consequence of S, since it is true in

every model in which the members of S are true, namely none. If there

were a strongly complete proof procedure, j_ would have a derivation

from S. That derivation, being a finite object, could only use a finite

subset of S, say So. Then j_ would be a logical consequence of So, and so

So could not be satisfiable (otherwise there would be a model in which

j_ was true). But every finite subset of S is satisfiable. Conclusion: no

strongly complete proof procedure can exist for higher-order classical

logic.

3.3 Weak Completeness

A proof procedure is (sound and) weakly complete if it proves exactly

the valid formulas. A strongly complete proof procedure is automatically

weakly complete (just use the empty set of premises). Higher-order

classical logic does not even possess a weakly complete proof procedure.

To show this the Incompleteness Theorem can be used.

The idea is to write a single formula that characterizes the natural

numbers-a second-order formula will do. One needs a constant symbol

CLASSICAL LOGIC-SEMANTICS 17

of type 0 to represent the number 0 and, to thoroughly overload notation,

I use 0 for this. Also a successor function is needed, but since we do not

have function symbols in this language, it is simulated with a relation

symbol S, technically a constant symbol of type (0, 0). In addition to

the abbreviations of Section 3.1, the following is needed.

0-exclude(S)

inductive-set(P, S)

induction(S)

for (Vx)-.S(x, 0)

for P(O) 1\ (Vx)[P(x) :J (3y)(S(x, y) 1\ P(y))]

for (VP)[inductive-set(P, S) :J (Vx)P(x)]

Now, let integer(S) be the formula

function(S) 1\ one-one(S) 1\ 0-exclude(S) 1\ induction(S)

It is not hard to show that integer(S) is true in a model (V,I) if and

only if the domain Vis (isomorphic to) the natural numbers, with I(S)

as successor. Consequently for any sentence <P of arithmetic, <P is true

of the natural numbers if and only if integer( B) :J <P is valid.

It is a standard requirement that the set of ( Godel numbers of) the-

orems of a proof procedure must be recursively enumerable, so if there

were a weakly complete proof procedure for higher-order classical logic,

the set of valid formulas would be recursively enumerable. The recursive

enumerability of the following set would then be an easy consequence:

the set of sentences <P such that integer(S) :J <P is valid. But, as noted

above, this is just the set of true sentences of arithmetic, and this is

not a recursively enumerable set. Conclusion: no weakly complete proof

procedure can exist for higher-order classical logic.

3.4 And Worse

I have been discussing higher-order classical logic, particularly its

models, using conventional informal mathematics of the sort that ev-

ery mathematician applies in papers and books. But certain areas of

mathematics-certainly formal logic is among them-are close to foun-

dational issues, and one needs to be careful. It is generally understood

that informal mathematics can be formalized in set theory, and this is

commonly taken to be Zermelo-Fraenkel set theory, or a variant of it.

Let us suppose, for the time being, that the development so far has been

within such a framework.

One of the famous problems associated with set theory is Cantor's

continuum hypothesis. It is the statement that there are no sets inter-

mediate in size between a countable set and its powerset. A little more

formally, it says:

18 TYPES, TABLEAUS, AND GO DEL'S GOD

Let X be a set, and let P(X) be its powerset. If X is countable, then any

infinite subset Y of P(X) either is in a 1-1 correspondence with X, or is in a

1-1 correspondence with P(X).

(The generalized continuum hypothesis is the natural extension of this

to uncountable infinite sets as well, but the simple continuum hypothesis

will do for present purposes.) Now, a difficulty for set theory is this: the

continuum hypothesis has been proved to be undecidable on the basis of

the generally accepted axioms for Zermelo-Fraenkel set theory. That is

(assuming the axioms for set theory are consistent) there is a model of

the Zermelo-Fraenkel axioms in which the continuum hypothesis is true,

and there is another in which it is false.

The problem for us is that the continuum hypothesis can be stated

as a sentence of higher-order classical logic. I briefly sketch how. First,

one can say the domain of a model is countable by saying there is a

relation that orders it isomorphically to the natural numbers. Using a

formula from Section 3.3, the following will do: (3a(O,O))integer(a(O,O)).

Next, one can identify a subset of the domain with an object of type

(0). Then the collection of all subsets of the domain is an object of type

( (0) ), so the following says there is a powerset for the domain of a model:

(

3

,6( (o))) (v

1

(o) ),B( (o)) (

1

(o)).

Having shown how to start, I leave the rest of the details to you. Write

a sentence saying: if the domain is countable then there is a powerset

for the domain and, for every infinite subset of that powerset, either

there is a 1-1 correspondence between it and the domain, or there is

a 1-1 correspondence between it and the powerset. You can say a set

is infinite using the negation of a formula from Section 3.1. And the

existence of a 1-1 correspondence amounts to the existence of a binary

relation meeting certain appropriate conditions. Let us call the sentence

that is the higher-order formalization of the continuum hypothesis CH.

Now, the real problem is: is the sentence CH valid or not? There are

the following not very palatable options.

1 Assume the foundation for informal mathematics is Zermelo-Fraenkel

set theory, formulated axiomatically. In this case neither CH nor its

negation can be shown to be valid, since the continuum hypothesis is

consistent with, but independent of, the Zermelo-Fraenkel axioms.

2 Assume that informal mathematics is being done in some particular

model for the Zermelo-Fraenkel axioms. In this case, CH is definitely

valid, or its negation is, but it depends on which Zermelo-Fraenkel

model is being considered.

CLASSICAL LOGIC-SEMANTICS 19

3 Assume that higher-order classical logic itself supplies the theoretical

foundations for mathematics. In this case CH either is valid or its

negation is, but which is it?

I have reached perhaps the most basic difficulty of all with classical

higher-order logic. Not only is there no proof procedure that will allow

us to prove every valid formula, the very status of validity for some

important formulas is unclear.

4. Henkin Models

As we saw in the previous section, higher-order classical logic is diffi-

cult to work with. Indeed, difficulties already appear at the second-order

level. Not only does it lack a complete proof procedure, but the very

notion of validity touches on profound foundational issues. Nonetheless,

there are several sound proof procedures for the logic-any formula that

has a proof must be valid, though not every valid formula will have a

proof. So, there are certainly fragments of higher-order logic that we

can hope to make use of.

In a sense, too many formulas of higher-order classical logic are valid,

so no proof procedure can be adequate to prove them all. Henkin broad-

ened the notion of higher-order model [Hen50] in a natural way, which

will be described shortly. With this broader notion there are more

models, hence fewer valid formulas, since there are more candidates for

counter-models. Henkin called his extension of the semantics general

models-! will call them Henkin models.

Henkin's idea seems straightforward, after years of getting used to

it. Given a domain V, a universal quantifier whose variable is of type

0, (Vx), ranges over the members of V. If we have a universal quanti-

fier, (VX), whose variable is of type (0), it ranges over the collection of

properties of V, or equivalently, over the subsets of V. The problem of

just what subsets an infinite set has is actually a deep one. The inde-

pendence of Cantor's continuum hypothesis is one manifestation of this

problem. Methods for establishing consistency and independence results

in set theory can be used to produce models with considerable variation

in the powerset of an infinite set. Henkin essentially said that, instead

of trying to work with all subsets of V, we should work with enough of

them, that is, we should take (V X) as ranging over some collection of

subsets of V, not necessarily all of them, but containing enough to satisfy

natural closure properties. Think of the collection as being intermediate

between all subsets and all definable subsets.

In a higher-order model as defined earlier, there is a domain, V, and

this determines the range of quantification for each type. Specifically,

20 TYPES, TABLEAUS, AND GODEL'S GOD

we thought of a quantifier (Vo:t) as ranging over the members of [t, V].

This time around a function is introduced, which I call a Henkin do-

main function and denote by 'It, explicitly giving us the range for each

quantifier type. Then Henkin frames are defined. This basic machinery

is needed before it can be specified what it means to have enough sets

available at each type.

DEFINITION 2.10 (HENKIN DOMAIN FUNCTION) 1t is a Henkin do-

main function if 1t is a function whose domain is the collection of types

and, for each type (t1, ... , tn), 'It( (t1, ... ,tn)) is some non-empty col-

lection of subsets of'Jt(ti) x x 1t(tn)

Sets of the form 1t(t) are called Henkin domains. The key point

is allowing some of the subsets of 1t(t1) x x 7t(tn)-the definition

of [t, V] had the word all at the corresponding point. Obviously the

function 1t(t) = [t, V] is a Henkin domain function. In fact, if 1t is

any Henkin domain function, and 7t(O) = V, then for every type t,

1t(t) [t, V], with equality holding at t = 0.

DEFINITION 2.11 (HENKIN FRAME) The structure M = (7t,I) is a

Henkin frame for a language L( C) if it meets the following conditions.

1 1t is a Henkin domain function.

2 If A is a constant symbol of L(C) of type t, I(A) E 7t(t).

3 I(=(t,t}) is the equality relation on 1t(t) for each type t.

The notion of valuation must be suitably restricted, of course.

DEFINITION 2.12 (VALUATION) V is a valuation in a Henkin frame

M = (7t,I) if v maps each variable of type t to some member of 1t(t).

Now, what will make a Henkin frame into a Henkin model? Let's try

a first attempt at a characterization. (This is not the "official" one, how-

ever. That will come later.) Definition 2.5, for the meaning of a term,

carries over almost word for word to a Henkin frame M. Also Defini-

tion 2.6, for truth in a model, carries over to M, with one restrictive

change. Item 4, the universal quantifier condition, gets replaced with

the following.

4'. Let M = ('It, I) be a Henkin frame and let o:t be a variable of type t.

M lf-v (Vo:t)<I> if M lf-v <I>[o:t ;ot] for every ot E 1t(t), or equivalently,

if M lf-v' <P for every at-variant v' of v such that v'(o:t) E 7t(t).

The revised version of item 4 above says that quantifiers of type t range

over just 1t(t) and not over all objects of type t.

CLASSICAL LOGIC-SEMANTICS 21

But there is a fundamental problem. Let M = (1t, I) be a Henkin

frame, and suppose (.Xa1, ... , is a predicate abstract-to keep

things simple for now, assume itself contains no abstracts. Then

for any valuation v, the characterization above determines whether or

not M 11-v Now, according to Definition 2.5, the meaning (v *

I)((.Xa1, ... to be assigned to the abstract, is {(01, ... ,On) I

M 11-v ... , an/On]}. The trouble is, we have no guarantee that

this set will be a member of the appropriate Henkin domain. If it is not

a member, quantifiers cannot include it in their ranges. If this happens,

we lose the validity of formulas like (Va) \]! (a) :::> \]! ( ( .Xa1, .. . ,

The whole business becomes somewhat problematic since formulas like

this clearly ought to be valid.

What must be done is impose enough closure conditions on the Henkin

domains of a Henkin frame to ensure that predicate abstracts always des-

ignate objects that are present in the Henkin domains. There are several

ways this can be done. Algebraic closure conditions can be formulated

directly, though this takes some effort. I follow a different route that is

somewhat easier. Essentially, I first allow predicate abstracts to desig-

nate members of Henkin domains in some arbitrary way, then. I add the

requirement that they be the "right" members.

DEFINITION 2.13 (ABSTRACTION DESIGNATION FUNCTION) A is an

abstraction designation function in the Henkin frame M = (1t, I) with

respect to the language L (C) if, for each valuation v in M, and each

type t predicate abstract (.Xa1, ... , an. <I?) of L(C), A(v, (.Xa1, ... , an.<P))

is some member of1t(t).

Think of an abstraction designation function as providing a "meaning"

for each predicate abstract. For the time being, such meanings can

be quite arbitrary, except that they must be members of appropriate

Henkin domains. Now earlier definitions get modified in straightforward

ways (and these are the "official" versions). Definition 2.5 becomes the

following.

DEFINITION 2.14 (DENOTATION OF A TERM IN A HENKIN FRAME)

Let M = (1t,I) be a Henkin frame, let v be a valuation, and let A be

an abstraction designation function. A mapping, (v *I* A), is defined

assigning to each term of L( C) a denotation for that term.

1 If A is a constant symbol of L(C) then (v *I* A)(A) = I(A).

2 If a is a variable then (v *I* A)(a) = v(a).

3 If (.Xa1, ... , is a predicate abstract of L(C), then

(v*I*A)((.Xa1, ... =A(v,(.Xa1, ...

22 TYPES, TABLEAUS, AND GODEL'S GOD

And Definition 2.6 becomes the following.

DEFINITION 2.15 (TRUTH OF A FORMULA IN A HENKIN FRAME)

Let M = (H,I) be a Henkin frame, let v be a valuation, and A be an

abstraction designation function. A formula <1> of L( C) is true in model

M with respect to v and A, denoted M If-v,A <1>, if the following holds.

1 For terms T, T!, ... ,Tn, M lf-v,A T(TI, ... ,Tn) provided

((v *I* A)(TI), ... , (v *I* A)(Tn)) E (v *I* A)(T).

2 M lf-v,A --,q> if it is not the case that M lf-v,A <1>.

3M lf-v,A <1> 1\ \]i if M lf-v,A <1> and M lf-v,A W.

4 M lf-v,A ('v'at)<P if M lf-v,A <P[at /0] for every 0 E 1t(t).

Now we can impose a requirement that designations of predicate ab-

stracts be "correct."

DEFINITION 2.16 (PROPER ABSTRACTION DESIGNATION FUNCTION)

Let M = (1t, I) be a Henkin frame and let A be an abstraction desig-

nation function in it, with respect to L( C). A is proper provided the

following is the case. For each predicate abstract (Aal, ... , an.<P) (with

ai of type ti) and for each valuation v we have

(v *I* A)( (Aal, ... , an.<P)) =

{ (01, ... , On) E 1t(t1) X X 1t(tn) J M if-v,A <P[ai/01, ... , an/On]}.

DEFINITION 2.17 (HENKIN MODEL) Let M be a Henkin frame, and let

A be an abstraction designation function in M. If A is proper, (M, A)

is a Henkin model.

For a given Henkin frame M it may be the case that no proper ab-

straction designation function exists. But, if one does exist it must be

unique.

PROPOSITION 2.18 Let M = (1t,I) be a Henkin frame and let both A

and A' be proper abstraction designation functions, with respect to L( C).

Then A= A'.

Proof The following two items are shown simultaneously, by induction

on degree (Definition 1.10). From this the Proposition follows immedi-

ately.

M lf-v,A <1> } M lf-v,A' <1>

(v*I*A)(T) = (v*I*A')(T)

(2.3)

(2.4)

CLASSICAL LOGIC-SEMANTICS 23

Suppose (2.3) and (2.4) are known for formulas and terms whose de-

gree is < k. It will be shown they hold for degree k too, beginning with

(2.4).

Suppose T is a term of degree k. Since k could be 0, T could be a

constant symbol or a variable. If it is a constant symbol, ( v *I* A) ( T) =

I( T) = ( v *I * A') ( T). Similarly if T is a variable. Finally, T could be a

predicate abstract, (>.a1, ... , an.<I>), in which case <I> must be a formula

of degree < k, so using the induction hypothesis with (2.3) we have

(v *I* A)( > . a ~ , ... , an.<I>)) =

{ (01, ... , On) I M 11-v,A <I>[ai/01, ... , an/On]}=

{ 0 ~ , ... , On) I M 11-v,A' <I>[ai/01, ... , an/On]}=

(v*I*A')((>.a1, ,an.<I>))

Thus (2.4) holds for terms of degree ~ k.

Now assume <I> is a formula of degree k. There are several cases,

depending on the form of <I>.

If <I> is atomic, it is T( T1, ... , Tn) where T, T1, ... , Tn are all of degree ~

k. Since (2.4) holds for terms of degree< k by assumption, and for terms

of degree = k by the proof above,

M 11-v,A T(T1, ... , Tn) {:}

((v *I* A)(T1), ... , (v *I* A)(Tn)) E (v *I* A)(T) {:}

((v*I*A')(T1), ... ,(v*I*A')(Tn)) E (v*I*A')(T) {:}

M 11-v,A' T(T1, ... , Tn)

If <I> is a negation, conjunction, or universally quantified formula, the

result follows easily using the fact that (2.3) holds for its subformulas

(which are of lower degree), by the induction hypothesis.

We thus have (2.3) for formulas of degree k, and this concludes the

induction.

Note on Induction Proofs The pattern of the induction proof above

will recur many times, with little variation of structure. We go from

terms and formulas of degrees < k to terms of degrees ~ k, and then to

formulas of degrees ~ k.

The Proposition above allows us to give the following extension of

Definition 2.17.

DEFINITION 2.19 (HENKIN MODEL) If (M,A) is a Henkin model, the

proper abstraction designation function A is uniquely determined, so we

24 TYPES, TABLEAUS, AND GODEL'S GOD

will say the Henkin frame M itself is a Henkin model, and write M lf-v <P

forM lf-v,A <P.

Suppose (V, I) is some classical model, as defined in Section 1. Set

H(t) = [t, V] for all types t. This gives us a Henkin domain function.

And it is easy to see that (H, I) will be a Henkin model. In fact, a

sentence <P is true in (H, I), as defined in this section, exactly when it

is true in the classical model (V, I), as defined in Section 2. This says

that "true" higher-order models are among the Henkin models. The

real question is, are there any other Henkin models? The answer is, yes.

The proof of the completeness theorem for tableaus will yield this as a

byproduct.

DEFINITION 2.20 (STANDARD MODEL) A Henkin model M

is a standard model if1i(t) = [t, V] for all types t.

(H,I)

Since standard models are among the Henkin models, any formula

that is true in all Henkin models must be true in all standard models

as well. But there is the possibility (a fact, as it happens) that there

are formulas true in all standard models that are not true in all Henkin

models. That is, the set of Henkin-valid formulas (Definition 2.29) is a

subset of the set of valid formulas (Definition 2.8), and in fact turns out

to be a proper subset. By decreasing the set of validities, it opens up the

possibility (again a fact, as it happens) that there may be a complete

proof procedure with respect to this more restricted version of validity.

5. Generalized Henkin Models

Unlike standard higher-order models, Henkin models are allowed to

have some, but not necessarily all, of the relations permissible in prin-

ciple at each type. This means there are more possibilities for Henkin

models than for standard models. Even so, the objects in the domains of

Henkin models are sets, and this imposes a restriction that we may want

to avoid in certain circumstances. Sets are extensional objects-that is,

a set is completely determined by its membership. Using the language

of properties rather than sets, two extensional properties that apply to

exactly the same things must be identical, and hence must have the same

properties applying to them. Working with sets is sufficient for math-

ematics, but it is not always the right choice in every situation. Even

if the terms "human being" and "featherless biped" happen to have the

same extension, we might not wish to identify them. As another exam-

ple, the properties of being the morning star and being the evening star

have the same extension, but were thought of as distinct properties by

the ancient Babylonians.

CLASSICAL LOGIC-SEMANTICS 25

Henkin himself [Hen50] noted the possibility of a more general notion

than what I am calling a Henkin model, "The axioms of extensionality

can be dropped if we are willing to admit models whose domains contain

functions which are regarded as d!stinct even though they have the same

value for every argument." Even so, extensionality has commonly been

built into the treatment of Henkin models in the literature-[And72]

is one of the rare instances where a model without extensionality is

constructed. As it happens, we will have need for a non-extensional

version in carrying out the completeness proof for tableaus. Since such

models are also of intrinsic interest, they are presented in some detail in

this section.

For Henkin frames, simply specifying the members of the Henkin do-

mains tells us much. Since they are sets, there is a notion of membership,

and it can be used in the definition of truth for atomic formulas. That is,

sets come with their extensions fully determined. If we move away from

sets this machinery becomes unavailable, and we must fill the gap with

something else-I make use of an explicit extension function, denoted E.

That is, for an arbitrary object 0, E(O) gives us the extension of 0. I

also allow the possibility that equality may not behave as expected-!

allow for non-normal frames and models.

DEFINITION 2.21 (GENERALIZED HENKIN FRAME) M = (1-l,I,E) is

called a generalized Henkin frame for a language L( C) if it meets the

following conditions.

1 1t is a function whose domain is the collection of types.

2 For each type t, 1-l(t) is some non-empty collection of objects (not

necessarily sets).

3 If A is a constant symbol of L(C) of type t, I(A) E 1-l(t).

4 For each type t = (h, ... , tn), E maps 1-l(t) to subsets of 1-l(tl) x

X1t(tn)

In addition, M is normal ifE(I(=(t,t))) is the equality relation on 1-l(t)

for each type t.

Much of this definition is similar to that of Henkin frame. The mem-

bers of 1-l(t) are the objects of type t (which now need not be sets). The

new item is the mapping E. Think of E(O) as the extension of the object

0.

Generalized Henkin models are built out of generalized Henkin frames.

Much of the machinery is almost identical with that for Henkin models,

26 TYPES, TABLEAUS, AND GODEL'S GOD

but there are curious twists, so things are presented in detail, rather

than just referring to earlier definitions. The definition of valuation is

the same as before.

DEFINITION 2.22 (VALUATION) The function vis a valuation in a gen-

eralized Henkin frame M = ('H, I,) if v maps each variable of type t

to some member of1i(t).

Next, just as with Henkin models, a function is needed that provides

designations for predicate abstracts, then later we can require that it

give us the "right" values. The wording is the same as before.

DEFINITION 2.23 (ABSTRACTION DESIGNATION FUNCTION) A is an

abstraction designation function in the generalized Henkin frame M =

('H, I,), with respect to the language L( C) provided, for each valuation

v in M, and for each predicate abstract (Aal, ... , an.<l?) of L(C) of type

t, A(v, (Aal, ... ,an.<I>)) is some member of1i(t).

Term denotation is like before-terms designate objects in the Henkin

domains.

DEFINITION 2.24 (DENOTATION OF A TERM) Let M = ('H,I,) be a

generalized Henkin frame, let v be a valuation, and let A be an abstrac-

tion designation function. A mapping, (v *I* A), is defined assigning

to each term of L( C) a denotation for that term.

1 If A is a constant symbol of L(C) then (v *I* A)(A) =I( A).

2 If a is a variable then (v *I* A)(a) = v(a).

3 If (Aal, ... , an. <I>) is a predicate abstract of L( C), then

(v*I*A)((Aal, ,an.<I>)) =A(v,(Aal, ,an-<I>)).

The following has a few changes from the earlier definition-to take

the extension function into account the atomic case has been modified.

DEFINITION 2.25 (TRUTH OF A FORMULA) Again let M = ('H,I,)

be a generalized Henkin frame, let v be a valuation, and A be an ab-

straction designation function. A formula <I> of L( C) is true in model

M with respect to v and A, denoted M lf-v,A <I>, provided the following.

1 For an atomic formula, M lf-v,A T(TI, ... , Tn) provided

((v *I* A)(TI), ... , (v *I* A)(Tn)) E E((v *I* A)(T)).

2 M lf-v,A ...,<f> if it is not the case that M lf-v,A <I>.

3 M If-v,A <I> 1\ \ll if M If-v,A <I> and M If-v,A \ll.

CLASSICAL LOGIC-SEMANTICS 27

4 M lf-v,A (Vat)<P if M lf-v,A <P[at /Ot] for every ot E 7-l(t).

In item 1 above, T(Tl, ... , Tn) is true if the designation of (TI, ... , Tn)

is in the extension of the designation of T. For Henkin frames, we were

dealing with sets, and extensions were for free. Now we are dealing with

arbitrary objects, and we must explicitly invoke the extension function

.

I am about to impose a "correctness" requirement, analogous to Def-

inition 2.16, but now there are three parts. The first part is similar to

that for Henkin models, except that the extension function is invoked.

The other parts need some comment. Suppose we have two predicate

abstracts ( >.a1, . . . , an. <P) and ( >.a1, . . . , an.\]!). In a Henkin model, if <P

and \]! are equivalent formulas, they will be true of the same objects and

so the two predicate abstracts will designate the same thing, since they

have the same extensions. But now we are explicitly allowing predicate

abstracts having the same extension to denote different objects. Still,

we don't want the designation of objects by predicate abstracts to be

entirely arbitrary-! will require equi-designation under circumstances

of "structural similarity."

DEFINITION 2.26 Let M be a generalized Henkin frame (or a Henkin

frame), and let A be an abstraction designation function in it. For each

valuation v and substitution a-, define a new valuation vu by:

Thus vu assigns to a variable a the "meaning" of the term aa-.

DEFINITION 2.27 (PROPER ABSTRACTION DESIGNATION FUNCTION)

Let M = (7-l, I,) be a generalized Henkin frame and let A be an ab-

straction designation function in it, with respect to L( C). A is proper

provided, for each predicate abstract (>.a1, ... , an.<P) we have

1 ((v*I*A)((>.ab ,an.<P))) =

{ (01, ... , On) I M lf-v,A <P[ai/01, ... , an/On]}

2 If v and w agree on the free variables of (>.a1, ... , an.<P) then

A(v,(>.al, ,an.<P)) =A(w,(>.al, ,an.<P))

3 If a- is a substitution that is free for the term (>.a1, ... , an. <P), then

A( v, (>.a1, ... , an.<P)a-) = A(vu, (>.a1, ... , an.<P))

The technical significance of items 2 and 3 above will be seen in the

next section. When using Henkin frames, if a proper abstraction des-

ignation function exists, it is unique (Proposition 2.18). But with a

28 TYPES, TABLEAUS, AND GODEL'S GOD

generalized Henkin frame, it is entirely possible for there to be more

than one proper abstraction designation function. Since there is this

possibility, we must specify which one to use--the frame alone does not

determine it.

DEFINITION 2.28 (GENERALIZED HENKIN MODEL) Let M be a gen-

eralized Henkin frame, and let A be an abstraction designation function

in M. If A is proper, (M, A) is a generalized Henkin model.

Finally Definition 2.8 is broadened to the entire class of generalized

Henkin models.

DEFINITION 2.29 (VALIDITY, SATISFIABILITY, CONSEQUENCE) Let <I>

be a formula and S be a set of formulas of L( C).

1 <I> is valid in generalized Henkin models if M lf-v,A <I> for every gen-

eralized Henkin model (M,A) for L(C) and valuation v.

2 S is satisfiable in a generalized Henkin model (M, A) for L( C) if

there is some valuation v such that M If-v,A <p for every <p E S.

3 <I> is a generalized Henkin consequence of S provided, for every gen-

eralized Henkin model (M,A) for L(C) and every valuation v, if

M lf-v,A <p for all <p E S, then M lf-v,A <I>.

Similar terminology is used when confining things to generalized Henkin

models that are normal, or to Henkin models themselves.

We saw in Section 4 that the notion of Henkin model extended that

of "true" higher-order model, since "true" models can be identified with

standard Henkin models. In a similar way the notion of generalized

Henkin model extends that of Henkin model, since Henkin models cor-

respond to what will be called extensional generalized Henkin models

(the definition is in the next section). Verifying this is postponed since it

requires us to show that parts 2 and 3 of Definition 2.27 hold for Henkin

models, and this involves some technical work. Assuming the result for

the moment, it follows that there are generalized Henkin models because

there are Henkin models; and we know there are Henkin models because

there are standard models. The question is: have the various generaliza-

tions really generalized anything? In fact, they have. It is a consequence

of the completeness proofs, which are given later, that there are Henkin

models that are not standard, and there are generalized Henkin models

that are not extensional Henkin models.

CLASSICAL LOGIC-SEMANTICS 29

6. A Few Technical Results

There are several results of a rather technical nature that, nonetheless,

are of fundamental importance. In fact, one of the propositions below

allows us to show that Henkin models are (isomorphically) among the

generalized Henkin models. Since we do not yet know this, we must

treat Henkin models and generalized Henkin models separately for the

time being.

6.1 Terms and Formulas

I leave the proof of the following Proposition as an exercise-see the

proof of Proposition 4.15 as a guide. The proof for generalized Henkin

models is similar to that for Henkin models except for the induction

step involving terms that are predicate abstracts, where a reduction to a

simpler case is not possible. But for generalized Henkin models, we are

given what we need for this step as part of the definition. (See part 2 of

Definition 2.27).

PROPOSITION 2.30 Let (M, A} be either a Henkin model or a gener-

alized Henkin model, and let v and w be valuations.

1 If v and w agree on the free variables of the term T

(v*I*A)(T) = (w*I*A)(T).

2 If v and w agree on the free variables of the formula <P

M lf--v,A <P {=::} M lf--w,A <P.

Next I state a result that will be used in the next Chapter to establish

the soundness of the tableau system.

PROPOSITION 2.31 Let (M, A} be either a Henkin model or a gener-

alized Henkin model. For any substitution a and valuation v:

1 If a is free for the term T then

(v *I*A)(Ta) = (vu *I*A)(T).

2 If a is free for the formula <P then

M lf--v,A <Pa {=::} M lf--v,.,A <P.

Once again I omit the proof, and refer you to Proposition 4.16 for

a similar argument. (For generalized Henkin models, part 3 of Defini-

tion 2.27 is needed.)

6.2 Extensional Models

Among Henkin models the standard ones correspond to "true" higher-

order models. A similar phenomenon occurs here-among the general-

ized Henkin models certain ones correspond to Henkin models.

30 TYPES, TABLEAUS, AND GODEL'S GOD

DEFINITION 2.32 (EXTENSIONAL) The generalized Henkin frame ('H,

I,) is extensional provided that (0) = (0') implies 0 = O' for all

objects 0 and O'. A generalized Henkin model is extensional if its frame

is.

Suppose M = ('H, I) is a Henkin frame (Definition 2.11). M can be

converted into a generalized Henkin frame M' = ('H, I,) by setting

( 0) = 0 for each object 0 of non-zero type. That is, we specify an ex-

tension function that gives us the usual set-theoretic notion of extension.

It is easy to check that if (M, A) is a Henkin model then (M', A) is a

generalized Henkin model-part 1 of Proposition 2.31 directly gives us

part 3 of Definition 2.27, and likewise Proposition 2.30 gives us part 2.

Obviously (M', A) is extensional. And equally obviously, evaluation of

truth in the original Henkin model and in the generalized Henkin model

just constructed is essentially the same.

Conversely, suppose M = ('H, I,) is a generalized Henkin frame that

is extensional. Inductively define a mapping () as follows. For objects

0 of type 0, B(O) = 0. And for an object 0 of type (tt, ... ,tn), set

B(O) = { (B(Ot), ... , B(On)) I (Ot, ... , On) E (0)}. Define a new

domain function 'H' by setting 'H'(t) = {B(O) I 0 E 'H(t)}. Using the

fact that M is extensional, it is not hard to show that () is 1-1 and

onto between 'H(t) and 'H'(t), for each type t. Finally, for each term T,

set I'(T) = B(I(T)). This gives us a Henkin frame ('H',I'). Thus, in

effect, each generalized Henkin frame that is extensional is isomorphic

to a Henkin frame as defined earlier.

From now on I will treat Henkin models as being generalized Henkin

models that are extensional, when it is convenient to do so.

6.3 Language Extensions

Part of the definition of (generalized) Henkin model is that each pred-

icate abstract must have an interpretation that is an object with the

"right" extension. But what predicate abstracts there are depends on

what the language is. Given a language L( C), one would expect mod-

els to depend on the collection of constants-members of C-which the

interpretation function, I, deals with. One would not expect the choice

of free variables of L(C) to matter, but this is not entirely clear, since

predicatr: abstracts can involve free variables. It is important to know

that the choice of free variables, in fact, does not matter, since the ma-

chinery of tableau proofs will require the addition of new free variables

to the language.

In what follows, L( C) is the basic language, and L + (C) is like L( C),

with new variables added, but with the understanding that these new

CLASSICAL LOGIC-SEMANTICS 31

variables are never quantified or >.-bound. (This all takes on a significant

role in the next chapter.) I note the fundamental problem: even with

the restrictions imposed on the additional variables, the collection of

predicate abstracts of L + (C) properly extends that of L( C).

PROPOSITION 2.33 Each generalized Henkin model with respect to L(C)

can be converted into a generalized Henkin model with respect to L+(c)

so that truth values for formulas of L( C) are preserved.

There are two immediate consequences of this Proposition that I want

to state, before I sketch its proof. First, any set S of sentences of L( C)

that is satisfiable in some generalized Henkin model with respect to

L( C) is also satisfiable in some generalized Henkin model with respect to

L+(C). And second, any sentence <I> of L(C) that is valid in all general-

ized Henkin models with respect to L + (C) is also valid in all generalized

Henkin models with respect to L( C) (because an L( C) countermodel

can be converted into a L+(c) countermodel).

Proof The proof basically amounts to replacing the new variables of

L + (C) by some from L( C), to determine behavior of predicate abstracts.

I only sketch the general outlines. Let M = (H, I,) be a generalized

Henkin frame, and let (M, A) be a generalized Henkin model with re-

spect to L(C).

Recall the notational convention: {/h / a1, . . . , f3n /an} is the substi-

tution that replaces each f3i by the corresponding ai. Also, if v is a

valuation, by v{fil/al, ... , fin/an} I mean the valuation v' such that

v'(ai) = v(fii), and on other free variables, v' and v agree.

Now we extend A to an abstraction designation function, A', suitable

for L+(C). For each predicate abstract (A"Yl, ... ,')'k.<I>) of L+(C), and

for each valuation v with respect to L+(C), do the following. Let (31,

. . . , fin be all the free variables of <I> that are in the language L + (C)

but not in L( C), and let a1, ... , an be a list of variables of L( C) of the

same corresponding types, that do not occur in <I>, free or bound. Now,

set

A' ( v, (>.1'1, ... , '/'k<I>)) =

A(v{(Jl/al, ... ,fin/an}, (A')'l, ... ,')'k.<I>{fil/al, ... ,fin/an}))

It can be shown that this is a proper definition, in the sense that it does

not depend on the particular choice of free variables to replace the f3i

Now it is possible to show that (M, A') is a generalized Henkin model

with respect to L + (C), and truth values of sentences of L( C) evaluate

the same with respect to A and A'. One must show a more general

result, involving formulas with free variables. The details are messy,

and I omit them.

32 TYPES, TABLEAUS, AND GODEL'S GOD

Finally, Proposition 2.33 has a kind of converse. Together they say

the difference between L(C) and L+(C) doesn't matter semantically. I

omit its proof altogether.

PROPOSITION 2.34 A generalized Henkin model with respect to L+(c)

can be converted into a generalized Henkin model with respect to L( C)

so that truth values for formulas of L( C) are preserved.

Exercises

EXERCISE 6.1 Give a proof of Proposition 2.30.

EXERCISE 6.2 Give a proof of Proposition 2.31.

EXERCISE 6.3 Supply details for a proof that each generalized Henkin

frame that is extensional is isomorphic to a Henkin frame.

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 3

CLASSICAL LOGIC-BASIC TABLEAUS

Several varieties of proof procedures have been developed for first-

order classical logic. Among them the semantic tableau procedure has

a considerable attraction, [Smu68, Fit96]. It is intuitive, close to the

intended semantics, and is automatable. For higher-order classical logic,

semantic tableaus are not as often seen-most treatments in the liter-

ature are axiomatic. Among the notable exceptions are [Tol75, Smi93,

Koh95, GilOl]. In fact, semantic tableaus retain much of their first-order

ability to charm, and they are what I present here. Automatability be-

comes more problematic, however, for reasons that will become clear

as we proceed. Consequently the presentation should be thought of as

meant for human use, and intelligence in the construction of proofs is

expected.

This chapter examines what I call a basic tableau system; rules are

lifted from those of first-order classical logic, and two straightforward

rules for predicate abstracts are added. It is a higher-order version of the

second-order system given in [Tol75]. I will show it corresponds to the

generalized Henkin models from Section 5 of Chapter 2. In Chapters 5

and 6 I make additions to the system to expand its class of theorems

and narrow its semantics to Henkin models.

1. A Different Language

In creating tableau proofs I use a modified version of the language

defined in Chapter 2. That is, I give tableau proofs of sentences from the

original language L( C), but the proofs themselves can involve formulas

from a broader language that is called L + (C). Before presenting the

tableau rules, I describe the way in which the language is extended for

proof purposes.

33

34 TYPES, TABLEAUS, AND GODEL'S GOD

Existential quantifiers are treated at higher orders exactly as they are

in the first-order case. If we know an existentially quantified formula is

true, a new symbol is introduced into the language for which we say, in

effect, let that be something whose value makes the formula true. As

usual, newness is critical. For this purpose it is convenient to enhance the

collection of free variables by adding a second kind, called parameters.

DEFINITION 3.1 (PARAMETERS) In L(C), for each type t there is an

infinite collection of free variables of that type. The language L + (C)

differs from L( C) in that, for each t there is also a second infinite list

of free variables of type t, called parameters, a list disjoint from that

of the free variables of L( C) itself. Parameters may appear in formulas

in the same way as the original list of free variables but they are never

quantified or A bound. p, q, P, Q, ... are used to represent parameters.

Parameters appear in tableau proofs. They do not appear in the

sentences being proved. Since they come from an alphabet distinct from

the original free variables, an alphabet that is never quantified or A

bound, we never need to worry about whether the introduction of a

parameter will lead to its inadvertent capture by a quantifier or a A-

introducing them will always involve a free substitution. Thus rules that

involve them can be relatively simple.

Special Terminology Technically, parameters are a special kind of

free variable. But to keep terminology simple, I will continue to use

the phrase free variable for the free variables of L( C) only, and when I

want to include parameters in the discussion I will explicitly say so.

The notion of truth in generalized Henkin models must also be ad-

justed to take formulas of +(c) into account. As I have just noted,

parameters are special free variables, so when dealing semantically with

L + (C), valuations must be defined for parameters as well as for the

free variables of L( C). Essentially, the difference between a generalized

Henkin frame and a generalized Henkin model lies in the requirement

that the extension of a formula appearing in a predicate abstract must

correspond to the designation of that abstract, which is a member of the

appropriate Henkin domain. In L + (C) there are parameters, so there

are more formulas and predicate abstracts than in L( C). Then requiring

that something be a generalized Henkin model with respect to L + (C) is

apparently a stronger condition than requiring it be one with respect to

L( C), though Section 6 establishes that this is not actually so.

DEFINITION 3.2 (GROUNDED) A term or a formula of +(C) is

grounded if it contains no free variables of L( C), though it may con-

tain parameters.

CLASSICAL LOGIC-BASIC TABLEAUS 35

The notion of grounded extends the notion of closed. Specifically, a

grounded formula of L + (C) that happens to be a formula of L( C) is a

closed formula of L( C), and similarly for terms.

2. Basic Tableaus

I now present the basic tableau system. It does not contain machinery

for dealing with equality-that comes in Chapter 5. The rules come

from [Tol75], where they were given for second-order logic. These rules,

in turn, trace back to the sequent-style higher-order rules of [Pra68] and

[Tak67].

All tableau proofs are proofs of sentences-closed formulas-of L( C).

A tableau proof of q> is a tree that has --,q> at its root, grounded formulas

of L + (C) at all nodes, is constructed following certain branch extension

rules to be given below, and is closed, which means it embodies a con-

tradiction. Such a tree intuitively says --,q> cannot happen, and so q> is

valid.

The branch extension rules for propositional connectives are quite

straightforward and well-known. Here they are, including rules for vari-

ous defined connectives.

DEFINITION 3.3 (CONJUNCTIVE RULES)

XI\Y

X

y

(X V Y)

-.x

--,y

(X =:J Y)

X

--,y

For the conjunctive rules, if the formula above the line appears on a

branch of a tableau, the items below the line may be added to the end of

the branch. The rule for double negation is of the same nature, except

that only a single added item is involved.

DEFINITION 3.4 (DOUBLE NEGATION RULE)

x

X

Next come the disjunctive rules. For these, if the formula above the

line appears on a tableau branch, the end node can have two children

added, labeled respectively with the two items shown below the line in

the rule. In this case one says there is tableau branching.

36 TYPES, TABLEAUS, AND GODEL'S GOD

DEFINITION 3.5 (DISJUNCTIVE RULES)

XVY

XIY

(X A Y)

XiY

(X = Y)

(X Y) I (Y X)

This completes the propositional connective rules. The motivation

should be intuitively obvious. For instance, if X A Y is true in a model,

both X and Y are true there, and so a branch containing X A Y can

be extended with X and Y. If X V Y is true in a model, one of them

is true there. The corresponding tableau rule says if X V Y occurs on a

branch, the branch splits using X and Y as the two cases. One or the

other represents the "correct" situation.

Though the universal quantifier has been taken as basic, it is con-

venient, and just as easy, to have tableau rules for both universal and

existential quantifiers directly. To state the rules simply, I use the fol-

lowing convention. Suppose <I>( li) is a formula in which the variable at,

of type t, may have free occurrences. And suppose Tt is a term of type

t. Then <I>( Tt) is the result of carrying out the substitution {at /Tt} in

<I>( at), replacing all free occurrences of at with occurrences of Tt. Now,

here are the existential quantifier rules.

DEFINITION 3.6 (EXISTENTIAL RULES) In the following, pt is a param-

eter of type t that is new to the tableau branch.

(:Jat)<P( at)

<P(pt)

(Vat)<P( at)

<I>(pt)

The rules above embody the familiar notion of existential instantia-

tion. Since the convention is that parameters are never quantified or >.-

bound, we don't have to worry about accidental variable capture. More

precisely, in the rules above, the substitution {at jpt} is free for the for-

mula <P(at).

The universal rules are somewhat more straightforward. Once again, .

note that in them the substitution {at /Tt} is free for the formula <I>( at).

DEFINITION 3.7 (UNIVERSAL RULES)

grounded term of type t of +(C).

In the following, Tt is any

(Vat) <I>( at)

<l>(Tt)

(3at)<I>( at)

<l>( Tt)

CLASSICAL LOGIC-BASIC TABLEAUS 37

Finally we have the rules for predicate abstracts. Earlier notation

is extended a bit, so that if q,(a1, ... , an) is a formula, a1, ... , an

are distinct free variables, and T1, ... , Tn are grounded terms of the

same respective types as a1, ... , an, then q,(rl, ,Tn) is the result of

simultaneously substituting each Ti for all free occurrences of ai in q,,

DEFINITION 3.8 (ABSTRACT RULES)

(Aal, ... , an.q,(a1, ... , ctn))(TI, ... , Tn)

q,(Tl, , Tn)

(Aa1, ... ,an.q,(a1, ... ,an))(r1, ... ,Tn)

--,q,(Tl, ... , Tn)

Now what, exactly, constitutes a proof.

DEFINITION 3.9 (CLOSURE) A tableau branch is closed if it contains

q, and --,q,, where q, is a grounded formula. A tableau is closed if each

branch is closed.

DEFINITION 3.10 (TABLEAU PROOF) For a sentence q, of L(C), a

closed tableau beginning with ,q, is a proof of q,.

DEFINITION 3.11 (TABLEAU DERIVATION) A tableau derivation of a

sentence q, from a set of sentences S, all of L( C), is a closed tableau

beginning with --,q,, allowing the additional rule: at any point any member

of S can be added to the end of any open branch.

This concludes the presentation of the tableau rules. In the next

section I give several examples of tableaus. Classical first-order tableau

rules, as in [Smu68, Fit96] are only involve subformulas

of the formula being proved. (It is not the case with the cut rule, but

this is an eliminable rule.) Higher-order rules, for the most part, have

an analytic nature as well. The important exception is the rule for the

universal quantifier. It allows us to pass from (Vat)q,(at) to q,(rt) where

Tt is an arbitrary grounded term. Since terms can involve predicate

abstracts, applications of this rule can introduce formulas that are not

subformulas of the one being they may be much more

complicated. There is no way around this. In a sense, the introduction

of predicate abstracts embodies the "creative element" of mathematics.

3. Tableau Examples

Tableaus for first-order classical logic are well-known, but the ab-

straction rules of the previous section are not as widely familiar. I give

38 TYPES, TABLEAUS, AND GO DEL'S GOD

a number of examples illustrating their uses. The first embodies the

principle behind many diagonal arguments in mathematics.

EXAMPLE 3.12 Suppose there is a way of matching subsets of some

set V with members of V. Let us call a member of V associated with

a particular subset a code for that subset. It is required that every

member of V must be a code, and nothing can be a code for more than

one subset, though it is allowed that some subsets can have more than

one code. Then, some subset of V must lack a code. (One consequence

of this is Cantor's Theorem: a set and its power set cannot be in a 1-1

correspondence.)

To formulate this, let R(x, y) represent the relation: y is in the subset

that has x as its code; so (>.y.R(x,y)) represents the set coded by x.

Then the following second-order sentence does the job.

('v'R)(3X)('v'x)[(>.y.R(x, y)) =X] (3.1)

This formulation contains equality. I have not given rules for equality

yet, so I give an alternative formulation that does not involve it.

('v'R)(3X)('v'x)(3y){[R(x, y) 1\ X(y)] V [R(x, y) 1\ X(y)]} (3.2)

I give a proof of (3.2). It is contained in Figure 3.1. In it, 2 is from 1

by an existential rule ( P is a new parameter); 3 is from 2 by a universal

rule ((>.x.P(x,x)) is a grounded term); 4 is from 3 by an existential

rule (p is another new parameter); 5 is from 4 by a universal rule (p is a

grounded term); 6 and 7 are from 5 by a conjunction rule; 8 and 9 are

from 6 by a disjunction rule; 10 is from 9 by double negation; 11 and 12

are from 7 by a disjunction rule, as are 13 and 14; 15 is from 12 by an

abstract rule, as is 16 from 10. Closure is by 8 and 11, 8 and 15, 13 and

16, and 10 and 14.

A key feature in the tableau proof of (3.2) is the use of (>.x. P(x, x))

in an application of a universal rule. This, in fact, is the heart of diagonal

arguments and amounts to looking at the collection of things that do not

belong to the set they code. The choice of such abstracts at key points

of proofs is the distilled essence of mathematical thinking-everything

else is mechanical. It is the need for such choices that stands in the way

of fully automating higher-order proof search.

Next is an example that comes out of propositional modal logic. Some

knowledge of Kripke semantics will be needed in order to understand the

background explanation, though not the tableau proof. See [HC96, pp

188-190] for a fuller treatment.

O'l

M

os

&S

C,)

0

0

t3

-.(\IR)(:JX)(\Ix)(:Jy){[R(x, y) 1\ -.X(y)] V [-.R(x, y) 1\ X(y)]} 1.

-.(:JX)(\Ix)(:Jy){[P(x, y) 1\ -.X(y)] V [-,P(x, y) 1\ X(y)]} 2.

-.(\lx)(:Jy){[P(x, y) 1\ -.(>.x.-.P(x, x)}(y)] V [-.P(x, y) 1\ (>.x.-.P(x, x))(y)]} 3.

-.(:Jy){[P(p, y) 1\ -.(>.x.-.P(x, x))(y)] V [-.P(p, y) 1\ (>.x.-.P(x, x))(y)]} 4.

-.{[P(p,p) 1\ -.(>.x.-.P(x, x))(p)] V [-.P(p,p) 1\ (>.x.-.P(x, x))(p)]} 5.

-.[P(p,p) 1\ -.(>.x.-.P(x,x))(p)] 6.

?,p) A

-.P(p,p) 8.

/

-.-.(>.x.-.P(x, x))(p) 9.

(>.x.-.P(x, x))(p) 10.

/

-.-.P(p,p) 11. -.(>.x.-.P(x, x))(p) 12. -.-.P(p,p) 13. -.(>.x.-.P(x, x))(p) 14.

-.-.P(p,p) 15. -.P(p,p) 16.

Figure 3.1. Tableau Proof of (\fR)(3X)(\fx)(3y){[R(x, y) 1\ X(y)] V [R(x, y) 1\ X(y)]}

40 TYPES, TABLEAUS, AND GODEL'S GOD

EXAMPLE 3.13 It is a well-known result of modal model theory that a

relational frame is reflexive if and only if every instance of DP => P is

valid in it. I want to give a formal version of this using the machinery of

higher-order classical logic. Suppose we think of the type 0 domain of a

higher-order classical model as being the set of possible worlds of a rela-

tional frame. Let us think of the atomic formula P(x) as telling us that

P is true at world x, and R( x, y) as saying y is a world accessible from

x. Then making use of the usual Kripke semantics, ('v'y)[R(x, y) => P(y)]

corresponds to P being true at every world accessible from x, and hence

to DP being true at world x, where R plays the role of the accessibil-

ity relation. Then further, saying DP => P is true at x corresponds to

('v'y)[R(x, y) => P(y)] => P(x). We want to say that if this happens at

every world, and for all P, the relation R must be reflexive, and con-

versely. Specifically, I give a tableau proof of the following. In it, take

R to be a constant symbol.

('v'x)R(x, x) = ('v'P)('v'x){('v'y)[R(x, y) => P(y)] => P(x)} (3.3)

Actually, the implication from left to right is straightforward-! sup-

ply a tableau proof from right to left.

--,{('v'P)('v'x){('v'y)[R(x, y) => P(y)] => P(x)} => ('v'x)R(x, x)} 1.

('v'P)('v'x){(Vy)[R(x,y) => P(y)] => P(x)} 2.

--,(\fx)R(x, x) 3.

--,R(p, p) 4.

('v'x){('v'y)[R(x, y) => (Az.R(p, z))(y)] => (Az.R(p, z))(x)} 5.

(\ly)[7 (Az.R(p,z))(y)] ::> ~ ) p ) 6.

--,(\fy)[R(p, y) => (Az.R(p, z))(y)]

--,[R(p, q) => (Az.R(p, z))(q) 9.

R(p, q) 10.

--,(Az.R(p, z))(q) 11.

--,R(p, q) 12.

7. (Az.R(p, z))(p) 8.

R(p,p) 13.

In this, 2 and 3 are from 1 by a conjunctive rule; 4 is from 3 by an

existential rule (p is a new parameter); 5 is from 2 by a universal rule

( (Az.R(p, z)) is a grounded term); 6 is from 5 by a universal rule (p is

a grounded term); 7 and 8 are from 6 by a disjunctive rule; 9 is from 7

by an existential rule ( q is a new parameter); 10 and 11 are from 9 by a

conjunction rule; 12 is from 11 and 13 is from 8 by abstract rules.

CLASSICAL LOGIC-BASIC TABLEAUS 41

The last example is a version of the famous Knaster-Tarski theorem

[Tar55].

EXAMPLE 3.14 Let 1) be a set and let F be a function from its powerset

to itself. F is called monotone provided, for each P, Q 1), if P

Q then F(P) F(Q). Theorem: any monotone function F on the

powerset of 1) has a fixed point, that is, there is a set C such that

F(C) =C. (Actually the Knaster-Tarski theorem says much more, but

this will do for present purposes.)

I now give a formalization of this theorem. Since function symbols

are not available, I restate it using relation symbols, and it is not even

necessary to require functionality for them. Now, (Vx)(P(x) :J Q(x))

will serve to formalize P Q. If F(P, x) is used to formalize that x is

in the set F(P), then (Vx)(P(x) :J Q(x)) :J (Vx)(F(P,x) :J F(Q,x))

says we have monotonicity. Then, the following embodies a version of

the Knaster-Tarski theorem (F is a constant symbol).

(VP)(VQ)[(Vx)(P(x) :J Q(x)) :J

(Vx)(F(P, x) :J F(Q, x))] :J (3S)('v'x)(F(S, x) = S(x))

(3.4)

I leave the construction of a tableau proof of this to you as an exer-

cise, but I give the following hint. Let <P(P, x) abbreviate the formula

(Vy)(F(P,y) :J P(y)) :J P(x). An appropriate term to consider during

a universal rule application is: (>..x.(VP)<P(P, x)).

A comment on the hint above. Rewriting ('v'y) ( F( P, y) :J P(y)) using

conventional function notation: it says F(P) P. Then <P(P, x) says

that x belongs to a set P if P meets the condition F(P) P. Then

further, (VP)<P(P, x) says that x is in n{P I F(P) P}. So finally,

(>..x.(VP)<P(P, x)) represents the set n{P I F(P) P} itself. In the

most common proof of the Knaster-Tarski theorem, one proceeds by

showing this set, in fact, is a fixed point of F.

Example 3.14 once again illustrates a fundamental point about higher-

order tableaus. They mechanize routine steps, but do not substitute

for mathematical insight. The choice of which predicate abstract to use

during an application of a universal rule really contains, in distilled form,

the essence of a standard mathematical argument.

The problem of what choice to make when instantiating a universal

quantifier also arises in first-order logic, but there is a way around it--one

uses free variables when instantiating, then one determines later which

values to choose for them [Fit96]. This last step, picking values, involves

unification, the solving of equations involving first-order terms. There

are several unification algorithms to do this, all of which accomplish

42 TYPES, TABLEAUS, AND GO DEL'S GOD

the following: given two terms, if there is a choice of values for their

free variables that makes the terms identical, the algorithm finds the

most general such choice; and if the terms cannot be made identical, the

algorithm reports this fact. Unification is at the heart of every first-order

theorem prover.

If we attempt a similar strategy in automating higher-order logic, we

immediately run into an obstacle at this point. The problem of unifi-

cation for higher-order terms is undecidable! This was shown for third-

order terms in [Hue73], and improved to show unification for second-

order terms is already undecidable, in [Gol81]. This does not mean

the situation is completely hopeless. While first-order unification is de-

cidable, and second-order is not, still there is a kind of semi-decision

procedure, [Hue75]. Two free-variable tableau systems for higher-order

classical logic, using unification, are presented in [Koh95]. The use of

higher-order unification in this way traces back to resolution work of

[And71] and [Hue72]. But finally, technical issues aside, we always come

back to the observation made above: the choice of predicate abstract to

use in instantiating a universally quantified formula often embodies the

mathematical "essence" of a proof. Too much should not be expected

from the purely mechanical.

Exercises

EXERCISE 3.1 Extending the ideas of Example 3.13, give tableau proofs

of the following.

1 (symmetry)

('v'x)('v'y)[R(x,y) :J R(y,x)] =

('v'P)('v'x){(3y)[R(x, y) 1\ ('v'z)(R(y, z) :J P(z))] :J P(x)}

2 (transitivity)

('v'x)('v'y)('v'z)[(R(x, y) 1\ R(y, z)) :J R(x, z)] =

('v'P)('v'x){('v'y)[R(x,y) :J P(y)] :J

('v'y)('v'z)[(R(x, y) 1\ R(y, z)) :J P(z)]}

EXERCISE 3.2 Give the tableau proof to complete Example 3.14.

EXERCISE 3.3 ContinuingwithExample3.14, the set n{P I F(P) P}

is not only a fixed point of monotonic F, it is the smallest one. Dually,

U{P I P F(P)} is also a fixed point, the largest one. Give a tableau

proof of (3.4) based on this idea.

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 4

SOUNDNESS AND COMPLETENESS

This chapter contains a proof that the basic tableau rules are sound

and complete with respect to generalized Henkin models. Soundness is

by the "usual" argument, is straightforward, and is what I begin with.

Completeness is something else altogether. For that I use the ideas

developed simultaneously in [Tak67, Pra68], where they were applied to

give a non-constructive proof of a cut elimination theorem.

1. Soundness

Soundness means that any sentence having a tableau proof must be

valid. Tableau soundness arguments follow the same pattern for all log-

ics: some notion of satisfiability is defined for tableaus; then satisfiability

is shown to be preserved by each tableau rule application. Note that in

the following, L + (C) is used rather than L( C), because formulas of the

larger language L + (C) can occur in tableaus.

DEFINITION 4.1 (TABLEAU 8ATISFIABILITY) A tableau branch is sat-

isfiable if the set of formulas on it is satisfiable in a generalized Henkin

model for L+(C) (see Definition 2.29). A tableau is satisfiable if some

branch is satisfiable.

Now, two key facts about these notions easily give us soundness. For

the first, a closed tableau branch contains some formula and its negation,

hence cannot be satisfiable. Since a closed tableau has every branch

closed, we immediately have the following.

LEMMA 4.2 A closed tableau cannot be satisfiable.

The second key fact takes more work to prove, but the work is spread

over several cases, each of which is rather simple.

43

44 TYPES, TABLEAUS, AND GO DEL'S GOD

LEMMA 4.3 If a branch extension rule is applied to a satisfiable tableau,

the result is another satisfiable tableau.

Proof Suppose T is a satisfiable tableau. Then it has some satisfiable

branch, say 13. Also suppose some branch extension rule is applied toT

to produce a new tableau, T'. It must be shown that T' is satisfiable.

The rule that was applied to turn T into T' may have been applied

on a branch other than 13. In this case 13 is still a branch of T', and

of course is still satisfiable, so T' is satisfiable. Now, for the rest of the

proof assume a branch extension rule has been applied to the satisfiable

branch B itself. And to be specific, say all the grounded formulas on

13 are true in the generalized Henkin model (M, A) with respect to the

valuation v, where M = (1t,I, E).

There are several cases, depending on which branch extension rule

was applied. I consider only a few of these cases and leave the rest to

you.

Disjunction Suppose the grounded formula X V Y occurred on 13 and

a rule was applied to it. Then in T' the branch 13 has been replaced

with two branches: 13 lengthened with X, and 13 lengthened with Y.

All formulas on 13 are true in (M, A) with respect to valuation v,

hence M lf--v,A XV Y. Then either M lf--v,A X or M lf--v,A Y. In the

first case, all members of 13 lengthened with X, and in the second

case, all members of B lengthened with Y, are true in (M, A) with

respect to v. Either way, some branch ofT' is satisfiable.

Existential Quantifier Suppose the grounded formula (3a)<P(a) oc-

curred on Band a rule was applied to it, so that in T' branch B has

been lengthened with <P(p) where p is a parameter new to B, of the

same type as a.

Since all formulas on Bare true in (M, A) with respect to v, M lf--v,A

(3a)<P(a). Then, by definition of truth in a model, there must be

some a-variant w of v such that M lf--w,A <P(a). Let a = {p/a}-

the substitution that replaces p by a-and consider the valuation

wa (Definition 2.26). I claim all formulas on 13 extended with <P(p)

are true in (M, A) with respect to wa, so the extended branch is

satisfiable.

First of all, v and w agree on all variables except a. It is easy to

see that w and wa agree on all variables except p, so the only vari-

ables on which v and wa can differ are a and p. But a does not

occur free in any formula on B, since these formulas are all grounded.

And p does not occur either, since p was new to the branch. Conse-

SOUNDNESS AND COMPLETENESS 45

quently all formulas on B are true in (M, A) with respect to wa, by

Proposition 2.30.

Finally, note that since p did not occur in (:lo:)<P(o:), then <P(o:) =

<P(p)a. We have M if-w,A <P(o:), and by Proposition 2.31

M if-w,A <P(o:) {:} M if-w,A <P(p)a

{:} M if-wu,A <P(p).

This completes the argument for the existential case.

Abstraction Suppose the grounded formula

occurred on B, and a rule was applied to it, so that in T' branch B

has been lengthened with <P(Tt, ... , Tn) We are assuming that the

formulas on B are all true in (M, A) with respect to valuation v. I

will show that this extends to include <P(Tt, ... ,Tn) as well.

Let a = { o:l/Tt, ... , o:n/ Tn}. This is free for <P( 0:1, ... , o:n) because

Tt, . . . , T n must be grounded, and parameters are never quantified or

>.-bound. Now consider the valuation va. Note the following useful

items.

1 va(o:i) = (v*I*A)(o:ia) = (v*I*A)(Ti)

2 If {3 is different from 0:1, ... , O:n, va ({3) = ( v * I * A) ({3a) =

(v*I*A)(f3) =v({3).

For this to be the case

((v *I* A)(Tt), ... , (v *I* A)(Tn)) E

E((v*I*A)((>.o:l, ,o:n.<P(o:l, ,o:n)))).

Since we have a generalized Henkin model, A is proper, so

E((v*I*A)((>.o:l, ,o:n.<P(o:l, ,o:n)))) =

{ (w(o:1), ... , w(o:n)) I w is an 0:1, ... , O:n-variant of v

and M if-w,A <P(o:1, , O:n)}

and consequently M if-w,A <P(o:1, ... , O:n) where w is the 0:1, ... ,

an-variant of v such that w(o:t) = (v *I* A)(Tt), ... , w(o:n) =

46 TYPES, TABLEAUS, AND GODEL'S GOD

(v *I* A)(Tn) But, by items 1 and 2 above, vu itself is this a1, ... ,

an-variant of v. We thus have

Now, by Proposition 2.31,

that is,

There are other cases-! leave them to you.

THEOREM 4.4 (SOUNDNESS) If a sentence <P of L(C) has a tableau

proof, <P must be true in all generalized Henkin models with respect to

L(C).

Proof Suppose <P has a tableau proof, but is not true in every gen-

eralized Henkin model with respect to L( C)-I derive a contradiction.

Since <P is not true in every generalized Henkin model with respect to

L( C), { -.<P} is satisfiable, and by Proposition 2.33, is so in a generalized

Henkin model with respect to L + (C). A tableau proof of <P begins with

a tableau consisting of a single branch, containing the single formula

--.<P, so this must be a satisfiable tableau. As we apply branch extension

rules, we continue to get satisfiable tableaus, by Lemma 4.3. Since <P

is provable, we can get a closed tableau. Hence there must be a closed,

satisfiable tableau, which is impossible according to Lemma 4.2.

Essentially the same argument also establishes the following.

THEOREM 4.5 Let S be a set of sentences and <P be a single sentence

of L(C). If <P has a tableau derivation from S, then <P is a generalized

Henkin consequence of S.

2. Completeness

The proof of completeness, for basic tableaus, with respect to gen-

eralized Henkin models, is of considerable intricacy. It is spread over

several subsections, each devoted to a single aspect of it. All the ba-

sic ideas go back to [Tak67, Pra68], where they were used to establish

non-constructively a cut-elimination theorem for higher-order Gentzen

systems. I also use aspects of the (second-order) presentation of [Tol75],

in particular the central goal, for us, is to prove that something called

SOUNDNESS AND COMPLETENESS 47

a Hintikka set is satisfiable. This contains the essence of the proofs

of [Tak67, Pra68]. [And71] abstracted the Takahashi, Prawitz ideas to

prove a higher-order Model Existence Theorem which could simply have

been cited, but the ideas of the completeness proof are pretty and deserve

to be better known, hence the full presentation.

In outline, the completeness proof is as follows. In Section 2.1 the

notion of a Hintikka set is defined: it is a set of grounded formulas of

L + (C) meeting certain closure conditions bearing an obvious relation-

ship to the tableau rules. In Section 2.2 pseudo-models are introduced.

These are the closest we come, in higher-order logic, to the Herbrand

models familiar in the first-order setting. Unfortunately, they will not

look like proper models in the higher-order sense, because objects as-

signed as meanings for predicate abstracts might lie outside the range

allowed for quantifiers. In Section 2.3 some rather technical (but impor-

tant) results about the behavior of substitution in pseudo-models are

shown. In Section 2.4 it is established that each Hintikka set is satisfi-

able in some pseudo-model. Section 2.5 shows that pseudo-models, in

fact, are proper generalized Henkin models after all, and so each Hin-

tikka set is satisfiable in such a model. Finally in Section 2.6 it is shown

how to extract a Hintikka set from a failed tableau proof attempt, and

this puts the last step in place for the completeness proof.

2.1 Hintikka Sets

Hintikka sets are fairly familiar from propositional and first-order

logics-see [Fit96] and [Smu68] for instance. They play a similar role

in the higher-order case, though arguments about them are much more

complex. You should note that the basic tableau rules all correspond

directly to Hintikka set conditions (I omit the connective = as a small

convenience).

DEFINITION 4.6 (HINTIKKA SET) A non-empty setH of grounded for-

mulas of L + (C) is a Hintikka set if it meets the following conditions.

1 Atomic Case. If <I> is atomic, not both <I> E H and <I> E H.

2 Conjunctive Cases.

(a) If (<I> 1\ w) E H then <I> E H and WE H.

{b) If (<I> V w) E H then <I> E H and W E H.

{c) If (<I> :J w) E H then <I> E H and W E H.

3 Disjunctive Cases.

{a) If (<P V w) E H then either <P E H or WE H.

48 TYPES, TABLEAUS, AND GODEL'S GOD

(b) If --, ( <P 1\ \ll) E H then either --,<fl E H or --, \ll E H.

(c) If (<P :J \ll) E H then either --,q, E H or \ll E H.

4 Double Negation Case. If --,--,q, E H then <P E H.

5 Universal Cases.

(a) If (Vat)<P(ci) E H then <P(rt) E H for every grounded term Tt.

(b) If (3at)<P(at) E H then <P(rt) E H for every grounded term

Tt.

6 Existential Cases.

(a) If (3at)<P( at) E H then <P(pt) E H for at least one parameter pt.

(b) If (Vat)<P(at) E H then <P(pt) E H for at least one parameter

pt.

7 Abstraction Cases.

(a) If (>.a1, ... , ... , an))(rl, ... , Tn) E H, then

<P(r1, ... , Tn) E H.

(b) If(>.al, ,an.<P(a1, ... ,an))(TI, ,Tn) E H, then

--,<fl(TI, ... ,Tn) E H.

This completes the definition of Hintikka sets. The task of relating

them to models begins in the next subsection.

2.2 Pseudo-Models

The eventual goal is to construct a generalized Henkin model, starting

with a Hintikka set. To do this a pseudo-model is first created, something

that is much like a generalized Henkin model but with one significant

difference: predicate abstracts are allowed to take on values that may

lie outside the range of the quantifiers! This will pose no problems for

the definition of truth in a pseudo-model since, for example, r1(r2) can

still be taken to be true if the value assigned to T2 is in the extension

of the value assigned to TI, whether or not these values are in quantifier

ranges. Eventually it will be shown that we can dispose of the "pseudo"

qualification on a pseudo-model.

I begin by defining entities of each type. These are the things that

can serve as values of predicate abstracts. In some ways the collection

of entities is an analog of a Herbrand universe, familiar from treatments

of first-order logic.

SOUNDNESS AND COMPLETENESS 49

DEFINITION 4. 7 (ENTITIES OF TYPE t) The notion of entity of type t

is defined inductively, on the complexity oft.

1 Suppose t = 0. If T is a grounded term of type t (thus a constant or

parameter of type 0}, T is an entity of type t.

2 Suppose t = (t1, ... , tn) and the collection of entities of type ti has

been defined for each i = 1, ... , n. Then (T, S) is an entity of type t

provided T is a grounded term of type t, and S is a set whose members

are of the form (E1, ... , En), where each Ei is an entity of type ti.

I also define two mappings on entities.

DEFINITION 4.8 (, T) If the entity E is of a type other than 0, it is

of the form (T, S); then T(E) = T and E(E) = S. I refer to E(E) as

the extension of E. The definition of T (but not of E) is extended to

entities of type 0 as well. If E is of type 0 it is, itself, a grounded term

of L+(C); in this case T(E) =E.

The idea is, if (T, S) is an entity of type t, it is something that could

serve as a semantic value for the term T, with the extension explicitly

coded in.

One problem with entities is that Hintikka sets play no role-the

collection of entities is the same no matter what Hintikka set we may

have. Presumably, if we are trying to construct a model from a given

Hintikka set, that should place some restrictions on what entities we

want to consider. The next definition separates out those entities that

will be in the range of quantifiers-it makes direct use of a Hintikka set.

It is these entities that will make up the Henkin domains of a model.

DEFINITION 4.9 (POSSIBLE VALUE) Let H be a Hintikka set. For

each grounded term T, define a collection of possible values ofT relative

to H. This is done inductively, on type complexity.

1 If T is a grounded term of type 0, the only possible value ofT relative

to H is T itself.

2 Suppose T is a grounded term of type (t1, ... , tn), and possible values

relative to H have been specified for all grounded terms of types t1,

... , tn. Then, an entity (T, S) is a possible value ofT relative to

H provided, for all grounded terms TI, ... , Tn of types t1, ... , tn

respectively, and for all possible values E1, ... , En of TI, ... , Tn:

(a) If T(TI, ... , Tn) E H then (E1, ... , En) E S.

(b) If T(TI, ... , Tn) E H then (E1, ... , En) S.

50 TYPES, TABLEAUS, AND GO DEL'S GOD

E is a possible value if it is a possible value for some grounded term.

Roughly the idea is, any possible value for T should have in its extension

all those things the Hintikka set H requires, and should omit all the

things H forbids. Any entity that meets these conditions will serve as a

possible value. Clearly each possible value of a grounded term of type

t, relative to a Hintikka set H, is an entity of type t. Item 1 of the

definition of Hintikka set ensures that part 2 above is meaningful.

Now that we have the notion of possible value, Henkin domains for

our pseudo-models can be defined.

DEFINITION 4.10 (RELATIVE HENKIN DOMAINS) Let H be a Hintikka

set. A mapping, 'HH is defined, from types to entities, as follows. For

each type t, 'HH(t) is the set of all entities of type t that are possible

values relative to H.

The languages L( C) and L + (C) are allowed to contain constant sym-

bols. How to interpret these is rather arbitrary, within broad limits.

DEFINITION 4.11 (ALLOWED INTERPRETATION) Let H be a Hintikka

set. A mapping I is an allowed interpretation relative to H provided I

assigns to each constant symbol A of type t some possible value for A,

relative to H.

We now have all the machinery needed to characterize an important

class of generalized Henkin frames arising from Hintikka sets.

DEFINITION 4.12 (RELATIVE GENERALIZED HENKIN FRAME) Let H

be a Hintikka set. M = ('HH, I,) is a generalized Henkin frame relative

to H provided:

1 'HH is the relative Henkin domain function of Definition 4.10;

2 I is an allowed interpretation relative to H, Definition 4.11;

3 is the extension function of Definition 4.8.

To produce our pseudo-models we need some notion of an abstraction

designation function. To define this we first need a little more machinery.

DEFINITION 4.13 (-v) Let v be a valuation in some generalized Henkin

frame relative to a Hintikka set H. Define a substitution -v as follows:

a'V = T(v(a)).

Thus, ifv(o:) = (T,S) then o:-v = T. Ifv(o:) = C, of type 0, then o:-v =

C. Note that -v substitutes grounded terms of L+(C) for variables, and

SOUNDNESS AND COMPLETENESS 51

so if T is an arbitrary term, T+v" must be a grounded term. Similarly

for formulas. Then, for any formula <I> and any valuation v, <I>+v" is

something that could potentially be a member of a Hintikka set.

Now everything is in place to define the notion of pseudo-model. Here

is a simultaneous recursive definition of truth, and of an abstraction

designation function. I denote the abstraction designation function by

AH, reflecting its dependence on the Hintikka set H, and ( v *I* AH) is

defined from it, a valuation v, and an interpretation I in the customary

way. I note that valuations have their standard meaning: they map

variables to members of Henkin domains. They do not map to arbitrary

entities.

DEFINITION 4.14 (PSEUDO-MODEL) Let H be a Hintikka set and let

M = (1-lH,I,) be a generalized Henkin frame relative to H. Truth of

formulas, and an abstraction designation function, are characterized as

follows.

1 For atomic formulas of L+(C), M lf--v,AH T(TI, ... , Tn)) if

((v*I*AH)(TI), ... ,(v*I*AH)(Tn)) E((v*I*AH)(T)),

2 M lf--v,AH X if M IYv,AH X.

3M lf--v,AH X 1\ Y if M lf--v,AH X and M lf--v,AH Y.

4 M lf--v,AH (Vo/)<I> if M lf--v,AH <I>[o:t ;at] for every ot E 1-lH(t).

5 Let T = (>.o:1, ... , O:n.<I>) be a predicate abstract of type (t1, ... , tn).

SetAH(v,(>.o:l, ,o:n.<I>)) = (T+v",S) where

S = { (01, ... , On) E 1iH(ti) X X 1iH(tn) I

M lf--v,AH <I>[o:I/01, , O:n/On]}.

The structure (M, AH) is a pseudo-model, relative to the Hintikka set

H.

The definition above has the usual complex recursive structure, with

truth at the atomic level needing ( v * I * AH) and hence AH, and the

characterization of AH itself needing the notion of truth for formulas.

Of course what makes it work is the fact that, in every case, behavior

of some construct on a formula or term requires constructs involving

simpler formulas and terms.

The key point is, why is this called a pseudo-model, and not simply a

model? The answer is, we have the characterization of the abstraction

designation function backwards here. In Chapter 2 we assumed we had a

function A that mapped valuations and abstracts to members of Henkin

52 TYPES, TABLEAUS, AND GODEL'S GOD

domains. Then we imposed a condition that A map to the "right"

members, so that our intuitions concerning abstracts would be respected.

Here we made that intuitive condition the defining property, reversing

the usual order of things. But now we have no guarantee that AH must

assign values that are in the Henkin domains. Looking at part 5 of the

definition above, it is clear that, for an abstract T of type (t1, ... , tn),

E(AH(v, T)) will be a subset of 'HH(tl) x x 'HH(tn), but we do not

know that AH(v,T) will be a possible value, and hence a member of

'HH( (t1, ... , tn) ). In short, while quantifiers range over Henkin domains

(condition 4 above), for all we know some terms-abstracts-can have

values that fall outside them. As a matter of fact, it will be proved

that this does not happen, but it is not obvious, and it is not easy to

establish.

Exercises

EXERCISE 2.1 Show that if entity E is a possible value, then E must

be a possible value of T(E).

2.3 Substitution and Pseudo-Models

In this subsection valuations and substitutions are shown to be well-

behaved with respect to pseudo-models. It should be noted again that

valuations always mean valuations in a pseudo-model-they map vari-

ables to members of Henkin domains, to possible values. They do not

map to arbitrary entities. The proofs below are rather technical, so I

begin with the statements of the two Propositions to be established, af-

ter which their proofs are given, broken into a number of Lemmas. On a

first reading you might want to just read the Propositions and skip over

the proofs.

The first item should be compared with Proposition 2.30.

PROPOSITION 4.15 Let H be a Hintikka set, let M = ('HH,I,E) be

a generalized frame relative to H, and let (M, AH) be a pseudo-model

relative to H. Also let v and w be valuations.

1 If v and w agree on the free variables of the term T

(v*I*AH)(T) = (w*I*AH)(T).

2 If v and w agree on the free variables of the formula <I>

M II-v,AH <I> ::::::} M II-w,AH <I>

The second item is an analog to Proposition 2.31. Definition 2.26 is

carried over to the present setting: given (M, AH), for each valuation v

and substitution CJ, the valuation vu is defined by o:va = ( v *I *AH) ( o:o-).

SOUNDNESS AND COMPLETENESS 53

PROPOSITION 4.16 Again let H be a Hintikka set, let M = (1lH,I,)

be a generalized frame relative to H, and let (M, AH) be a pseudo-model

relative to H. For any substitution CJ and valuation v:

1 If CJ is free for the term T then

(v * T * AH)(TCJ) = (vu * T * AH)(T).

2 If CJ is free for the formula <P then

M lf--v,AH <PCJ {::::::::} M lf--v",AH <P.

Now I turn to the proofs, which are given in considerable detail since

these results are critical to the completeness argument, and I want the

reasoning on record. On a first reading, skip the proofs and move on to

the next section.

Proof of 4.15 Suppose the result is known for terms and formulas

whose degree is < k. I show the result is also true for those of degree k

itself, beginning with terms.

Assume Tis a term of degree k, and v and w agree on the free variables

and parameters of T. If k happens to be 0, T is a constant symbol,

variable, or parameter. In these cases the result is immediate.

Now suppose k =/= 0, and soT= (.Xa1, ... , an.<P), where <Pis of degree

< k. Say (v *T*AH)((.Xal, ,an.<P)) = AH(v,(Aai, ,an.<P)) =

(a,S) and (w*T*AH)((Aal, ,an.<P)) =AH(w,(.Xal, ,an.<P)) =

(a', S'). We must show a= a' and S = S'.

Suppose a is a variable or parameter that occurs free in T. Then

air = T(v(a)) = T(w(a)) = a ~ using the assumption that v and

w must agree on a. By definition, a = T'v and a' = T ~ and the

substitutions 'v and ~ agree on the free variables of T, so a = a' by

Proposition 1.15.

Next we suppose (E1, ... , En) E S, and so M lf--v,AH <P[al/ E1, ... ,

an/ En] Since v and w agree on the free variables and parameters of

(.Xa1, ... , an.<P), they agree on the free variables and parameters of <P,

other than a1, ... , an. Then by the induction hypothesis, M lf--w,AH

<P[al/ E1, ... , an/ En], and it follows that (E1, ... , En) E S'. Thus S ~

S'. A similar argument shows S' ~ S.

This completes the induction step for terms, and I turn next to formu-

las. Suppose <P is of degree k and v and w agree on the free variables and

parameters of <P. By the induction hypothesis, we have the Proposition

for terms and formulas of degree< k, and by what was just shown, we

also have it for terms of degree k itself. Now we have several cases.

Suppose <Pis atomic, To(TI, ... , Tn), where each Ti must be of degree

~ k. Then, using the induction hypothesis and what was just proved,

54 TYPES, TABLEAUS, AND GODEL'S GOD

M lf--v,AH To(T1, ... , Tn) ::>

((v *I* AH)(T1), ... , (v *I* AH)(Tn)) E ((v *I* AH)(To)) ::>

((w *I* AH)(T1), ... , (w *I* AH)(Tn)) E ((w *I* AH)(To)) ::>

M lf--w,AH To(T1, ... , Tn)

The various non-atomic cases are left to you.

Next we have several preliminary results, leading up to the proof of

Proposition 4.16. Recall Definition 1.12: (Ja

1

, ... ,an is the substitution

that is like(} except that it is the identity on 0:1, ... , O:n.

LEMMA 4.17 Let H be a Hintikka set, let M = (HH,I, ) be a general-

ized frame relative to H, and let (M, AH) be a pseudo-model relative to

H. Let (..Xo:1, ... , O:n.<P) be a term of L+(C), and suppose the substitu-

tion (} is free for it, and is the identity map on parameters and variables

that do not occur free in (..Xo:1, ... ,o:n.<P). Then, for any valuation v:

1 If w is an 0:1, ... , O:n variant of v then wu"'l ,a.n is an 0:1, ... , O:n

variant of vu.

2 Conversely, if u is an 0:1, ... , O:n variant of vu then u = wu"'l , ... "'n

for some 0:1, ... , O:n variant w of v.

3 vu<>1, ... ,<>n (o:i) = v(o:i), fori= 1, ... , n.

Proof

Part 1. Suppose w is an 0:1, ... , O:n variant of v. Let {3 be a variable or

parameter other than 0:1, ... , O:n. It must be shown that wu<>1>-- ,<>n ({3) =

va({3). Here are the steps; the reasons follow.

w(T<>l, ... ,<>n ({3) = ( w *I* AH )(f3(Jal, ,an)

= ( w * I * AH) ({3(})

= (v*I*AH)(f3(})

= vu ({3)

(4.1)

(4.2)

(4.3)

(4.4)

Above, (4.1) is the definition of wu<>l, ... ,<>n, and (4.2) is because {3 is

different from 0:1, ... , O:n. Also (4.4) follows from (4.3) by the definition

of vu. The key item is the equality of (4.2) and (4.3), and for this it is

enough to show v and w agree on the free variables and parameters of

{3(}, and then appeal to Proposition 4.15. The argument for this follows.

If {3 does not occur free in (..Xo:1, ... , O:n.<P), {3(} = {3 by assumption.

Also we are assuming {3 is different from o:

1

, ... , O:n, and v and w agree

on all variables except o:

1

, . . . , O:n, so v and w must agree on {3, and

SOUNDNESS AND COMPLETENESS 55

hence trivially they agree on the free variables and parameters of {JCJ in

this case.

Now suppose {3 does occur free in (Ao:1, ... , Since CJ is free for

(Ao:1, ... , {JCJ cannot contain any of 0:1, ... , O:n free. Once again

v and w must agree on the free variables and parameters of {JCJ, since v

and w can only differ on 0:1, . . . , O:n.

Part 2. Suppose u is an 0:1, ... , O:n variant of va. Define a valuation

w as follows.

w(o:i) = u(o:i)

w(f3) = v(f3)

i = 1, ... ,n

{3 ::/= 0:1, . . . , O:n

By definition, w is an 0:1, ... , O:n variant of v. I will show wa"'I>-- "'n = u.

The argument is in two parts.

wa"'l"'n(o:i) = (w*L*AH)(O:Wat, ... ,aJ (4.5)

=(w*L*AH)(o:i) (4.6)

= w(o:i) (4.7)

= u(o:i) (4.8)

In this, (4.5) is by definition of wa"'l .. "'n. Then (4.6) is because CJa

1

, ... ,an

is the identity on O:i Next, (4.7) is because o:i is a variable, and finally

( 4.8) is by definition of w.

Now suppose f3 ::/= 0:1, ... , O:n.

wa"'l>"'n(f3) = (w*I*AH)(f3CJa

1

, . ,an) (4.9)

=(w*I*AH)(f3CJ) (4.10)

=(v*I*AH)(j3CJ) (4.11)

= va(f3) (4.12)

= u(f3) (4.13)

Here (4.9) is by definition of wa"'l .. "'n. Then (4.10) is by definition of

CJat, ... ,an Next, (4.11) follows exactly as (4.3) did above. Finally (4.12)

is by definition of va, and (4.13) is because u and va are 0:1, ... , O:n

variants.

Part 3. Va"'l>"'n(o:i) = (V*L*AH)(o:iCJa

1

, ... ,an) = (v*L*AH)(o:i) =

v(o:i)

LEMMA 4.18 Let H be a Hintikka set, let M = (1tH,I,) be a gener-

alized frame relative to H, and let (M, AH) be a pseudo-model relative

to H. For any term T of L+(C), = T((v *I* AH)(T)).

56 TYPES, TABLEAUS, AND GODEL'S GOD

Proof Suppose first that T is a predicate abstract. Then by Defini-

tion 4.14, (v *I* AH)(T) = AH(v,T) = (T*v,S) for a certain setS,

and so T((v *I* AH)(T)) = T*v. If T is a variable or parameter,

T*v = T(v(T)) by definition of tv", and v(T) = (v *I* AH)(T) by def-

inition of ( v * I * AH) again, for variables. If T is a constant symbol,

T*v = T, and also T((v *I* AH)(T)) = T(I(T)) = T because I is an

allowed interpretation.

The proof of Proposition 4.16 is by an induction on degree. Since the

steps are somewhat complex, I have separated the significant parts out,

in the following two Lemmas.

LEMMA 4.19 Let H be a Hintikka set, let M = ('HH, I,) be a general-

ized frame relative to H, and let (M, AH) be a pseudo-model relative to

H. Assume that for each formula <I> of degree < k, whenever substitution

C5 is free for <I> then

M lf-v,A <l>C5 {::} M lf-v",A <I>.

Then for each term T of degree k, and for each substitution C5 that is free

forT,

( v *I * AH) ( TC5) = ( vu *I * AH) ( T).

Proof Assume the hypothesis concerning formulas, and suppose T is a

term of degree k. If k is 0, T must be a constant symbol, a variable, or a

parameter. If Tis a variable or parameter, say a, then (vu *I*AH)(a) =

vu (a) = ( v *I* AH) ( ae5). The case of a constant symbol is trivial.

Now suppose k > 0, and so T must be of the form (>.a1, ... , an.<I>),

where <I> is a formula whose degree is < k. And suppose C5 is free for

(>.a1, ... , an. <I>). Using the definition of substitution and Definition 4.14:

where

(v*I*AH)((>.a1, ... ,an.<I>)C5)

= (v*I*AH)((>.a1, ... ,an.<I>C5a

1

, ... ,an))

=AH(v,(>.a1, ,an.<I>C5a

1

, ... ,aJ)

=(a, S)

a= (>.a1, ... ,an.<I>C5a

1

, ... ,an)+v

S = { (01, ... , On) E 'HH(t1) X .. X 'HH(tn) I

M lf-v,AH <l>C5a1 , ... ,an[al/01, , an/On]}.

SOUNDNESS AND COMPLETENESS

Similarly:

where

I ( )+a

a = >.a1. ... , an.<I> v

S' = { (01, ... , On) E 'HH(tl) X X 'HH(tn) J

M 11-v",AH <I>[al/01, ... , an/On]}.

So, we must show a= a' and S = S'.

Part 1, a= a'.

First of all,

a= (>.al, ... ,an.<I>aal,,an)1J

= ((>.al, ,an.<I>))a)1J

= (>.a1, ... , an.<I>)(a11)

57

(4.14)

(4.15)

In this, (4.14) is by definition of substitution. Recall we are assuming

that a is free for (>.a1. ... , an. <I>). Also, since 11 replaces variables by

grounded terms, and parameters are never bound, substitution 11 is free

for (>.a1, ... , an.<I>)a. Then (4.15) follows by Theorem 1.18.

So, to show a= a' it is enough to show the substitutions a11 and if

are the same. Let (3 be a variable or parameter. f3(a11) = ([3a)11 by

definition of composition for substitutions. And, using Definition 4.13,

(3if = T(v

17

({3)) = T((v *I* AH)(f3a)). Finally, (f3a)1J and T((v *I*

AH)(f3a)) are the same, by Lemma 4.18.

We thus have shown that a= a'.

Part 2, S = S'.

Using Proposition 1.15 it can be assumed that a is the identity on

variables and parameters that do not occur free in (>.a1, ... , an. <I>).

58 TYPES, TABLEAUS, AND GO DEL'S GOD

8= {(01, ... 0n) I M lf-v,AH <Paa

1

, ... ,an[o:I/01, ,o:n/On]} (4.16)

= { (w(o:1), ... , w(o:n)) I M lf-w,AH <Paa

1

, ... ,an

where w is an 0:1, ... , O:n variant of v} ( 4.17)

= {(w(o:1), ... ,w(o:n)) I M lf-wu"l .. "n,AH <P

where w is an 0:1, ... , O:n variant of v} ( 4.18)

= { (wa"l'"' ,<>n (o:I), ... 'wa"l'"' ,<>n (o:n)) I M If-wO'"l"' ,<>n ,AH <I>

where w is an 0:1, ... , O:n variant of v} ( 4.19)

= { (u(o:1), ... , u(o:n)) I M lf-u,AH <P

where u is an 0:1, ... , O:n variant of va} ( 4.20)

= { (01, .. , On) I M lf-vu,AH <P[o:I/01, ... , O:n/On]} (4.21)

Above, (4.17) is just ( 4.16) rewritten. Since a is free for A o : ~ , ... , O:n.<P),

by Definition 1.16, aa

1

, ... ,an must be free for <P, and since <P must be of

degree < k, ( 4.17) and ( 4.18) are equal by the hypothesis of the Lemma.

Then (4.18) and (4.19) are equal by part 3 of Lemma 4.17. The equality

of (4.19) and (4.20) follows by parts 1 and 2 of Lemma 4.17. Finally,

(4.21) is (4.20) rewritten.

LEMMA 4.20 Let H be a Hintikka set, let M = (HH,I,t:) be a gener-

alized frame relative to H, and let (M, AH) be a pseudo-model relative

to H. Assume that

( 4.22)

for each substitution a that is free for <I>, provided <I> is of degree < k.

Also assume that

( v *I* AH) (TO") = ( va *I * AH) ( T)

(4.23)

for each substitution that is free forT, provided T is of degree ::; k. Then

(4.22} also holds for each formula <P of degree k itself.

Proof Assume the hypothesis. Suppose <P is a formula of degree k.

There are several cases depending on the form of <P.

SOUNDNESS AND COMPLETENESS 59

If <I> is of the form r(r1, ... , Tn), each of the terms r, r1, ... , Tn must

be of degree :=:; k. Then, using hypothesis ( 4.23) about terms,

M lf-v,AH <I>a {:::> M lf-v,AH (r(TI, ... , Tn)))a

{:::> M lf-v,AH ra(r1a, ... ,Tna)

{:::> ((v*I*AH)(rla), ... ,(v*I*AH)(rna))

E ((v *I* AH)(ra))

{:::> ( ( Va * I * A H) ( TI) , . . . , ( Va * I * A H) ( T n))

E ((va *I* AH)(r))

{:::> M lf-v",AH r(TI, ... , Tn)

{:::> M If-v" ,AH <I>

Next, if <I> is a propositional combination of simpler formulas, the

argument is straightforward using the hypothesis about formulas, and is

left to you. Finally, if <I> is of the form (Va)w the argument, in outline,

is as follows.

M lf-v,AH [(Va)w]a {:::> M lf-v,AH (Va)[Waa]

{:::> M lf-w,AH Waa for every valuation

w that is an a-variant of v

{:::> M If-w"" ,AH W for every valuation

w that is an a-variant of v

{:::> M lf-u,AH W for every valuation

u that is an a-variant of va

{:::> M lf-v",AH (Va)w

(4.24)

( 4.25)

(4.26)

In this, the equivalence of ( 4.24) and ( 4.25) is by the hypothesis about

formulas, and the equivalence of ( 4.25) and ( 4.26) is by a result similar

to that stated in Lemma 4.17, but for quantified formulas rather than

for predicate abstracts.

Finally, the central item we have been aiming at.

Proof of 4.16 The proof, of course, is by induction on degree. Sup-

pose the result is known for formulas and terms of degree < k. Then

by Lemma 4.19 the result holds for terms of degree k, and then by

Lemma 4.20 it holds for formulas of degree k as well.

2.4 Hintikka Sets and Pseudo-Models

Given a Hintikka set, we now know how to create a pseudo-model from

it. It would be nice if the various formulas in the Hintikka set turned

60 TYPES, TABLEAUS, AND GODEL'S GOD

out to be true in that pseudo-model. This happens, and will be shown

below. But we still have the troublesome feature of pseudo-models that

quantifiers range over members of the Henkin domains, but predicate

abstracts can have as values entities that might lie outside them. It

would be nice if the values assigned to predicate abstracts turned out

to be possible values after all. This too happens to be the case, and

will also be shown below. In fact, both of the things we desire will be

shown simultaneously, in one big result. Then we can conclude that each

Hintikka set is satisfiable in a pseudo-model that actually is a generalized

Henkin model.

THEOREM 4.21 Let H be a Hintikka set, let M = (HH,I,) be a gen-

eralized frame relative to H, and let (M, AH) be a pseudo-model relative

to H. Also let v be a valuation in the pseudo-model.

1 For each term T of L + (C), ( v *I* AH) ( T) is a possible value for TlJ.

2 For each formula <J> of L+(C), if <J>"V E H then M lf-v,AH <J>.

Proof Both parts of the theorem are shown together by a simultaneous

induction on degree. Assume they hold for formulas and terms of degree

< k. It will first be shown that item 1 holds for terms of degree k; then

it will be shown that item 2 holds for formulas of degree k.

Part 1. Let T be a term of degree k. If k happens to be 0, T is a

constant symbol, variable, or parameter. If T is a constant symbol A,

A"V =A, and (v *I* AH)(A) = I(A), which is a possible value of A

because I is an allowed interpretation. If Tis a variable or parameter, a,

( v *I* AH) (a) = v( a) is some possible value E because v is a valuation

in the pseudo-model. But then a"V = T(v(a)) = T(E), and E is a

possible value of T(E) by Exercise 2.1.

Now suppose T = (.Aa1, ... , an.<I>). Then (v*I*AH)(r) = AH(v, r) =

(r"V, S) where

S = { (Ot, ... , On) E HH(ti) X .. X HH(tn) I

M lf-v,AH <J>[ai/01, .. , an/On]}

= { (w(ai), ... , w(an)) I M lf-w,AH <I>

where w is an a1, ... , an variant of v}.

I show (r"V, S) is a possible value of r"V relative to H. To do this it

must be shown that if E1 is a possible value for T1, ... , En is a possible

value for Tn, then

SOUNDNESS AND COMPLETENESS 61

2 (71f)(71, ... , 7n) E H implies (E1, ... , En) fl. S.

I show the first of these; the second is similar.

So assume E1 is a possible value for 71, . . . , En is a possible value for

7n, and (71f)(71,. ,7n) E H. That is,

[(>.a1, ... ,an.<I>)1f](71, ... ,7n) E H.

By definition of substitution we have

(.Xa1, , an.<I>+v a1 , ... ,an)(71, , 7n) E H.

Since H is a Hintikka set, it follows that (Definition 4.6, part 7)

[<I>+v a1, ... ,an]{al/71, ... , an/7n} E H.

Since 71, ... , 7n are grounded terms, they do not contain any of a1, ... ,

an free. Now, let w be the a1. ... , an-variant of v such that w(a

1

) = E

1

,

... , w(an) =En. Since Ei is a possible value for the grounded term 7i

it follows that ai w = 7i. And if j3 f= a1, ... , an then j3w = j31J. Then

[<I> tv a

1

, ... ,an]{ al/71, ... , an/Tn} = <I>w so

<I>w E H.

<I> must be of lower degree than (>.a1, ... ,an.<I>), that is, k, so the in-

duction hypothesis applies and

M 11-w,AH <I>.

Then (w(a1), ... , w(an)) E S, so (E1, ... , En) E S, which is what we

wanted.

This concludes the induction step for terms.

Part 2. Let <I> be a formula of degree k. By the induction hypothesis

the result holds for formulas and terms of degree < k, and by part 1 of

the proof it also holds for terms of degree k. Now we have several cases,

depending on the form of <I>. I only present a few of them.

Suppose <I> is 7o(71, ... , 7n) and [To(71, ... , 7n)]tv E H. That is,

Each 7i is of degree ::; k so by the induction hypothesis, each ( v * I *

AH)(7i) is a possible value for 7i1f. It follows immediately from the

definition of possible value (Definition 4.9) and the definition of (Def-

inition 4.8) that

62 TYPES, TABLEAUS, AND GODEL'S GOD

and so

Suppose <P is X 1\ Y, and (X 1\ Y) 'v E H. By definition of substitution,

(X'v 1\ Y'v) E H. Since H is a Hintikka set, X'v E H and Y'v E H.

But each of X and Y is of lower degree than <P, so by the induction

hypothesis, M 11-v,AH X and M 11-v,AH Y. It follows that M 11-v,AH

X 1\ Y.

Suppose <P is (\io:)w(o:) and [(\io:)w(o:)]'v E H. By definition of

substitution, (\io:)[w'v a](o:) E H. Let w be an arbitrary a-variant of v

that assigns to o: the possible value E. Since E is a possible value, it is

the possible value of some grounded term, say T. Now by definition of

Hintikka set, [w'v a](T) E H. We have o:w = T, and if {3 -=f. o:, {Jw =

{J'v, so [w(o:)]w = [w'v a](T), and hence [w(o:)]w E H. But W(o:) is

oflower degree than <P, so by the induction hypothesis, M 11-w,AH w(o:).

Since w was arbitrary, M 11-v,AH (\io:)w(o:).

The other cases are similar and are omitted.

COROLLARY 4.22 Let H be a Hintikka set, let M = (1-lH,I,) be a

generalized frame relative to H, and let (M, AH) be a pseudo-model

relative to H. Then H is satisfied in the pseudo-model (M, AH). More

specifically, let v be any valuation in this pseudo-model that assigns to

each parameter p some possible value for p; then if <P E H, M lf-v,AH <P.

Proof If v assigns to each parameter p some possible value for p, then

p'v = T(v(p)) = p. Consequently for each grounded formula <P we have

<P'v = <P. The result then follows from part 2 of Theorem 4.21.

2.5 Pseudo-Models are Models

So far, a satisfiability result has been shown using pseudo-models.

But along the way everything needed to show that pseudo-models are

actually models has been established. Since this is an important fact, I

give it a section of its own though, as I said, the work has already been

done.

THEOREM 4. 23 Let H be a Hintikka set, let M = (1-lH, I,) be a gen-

eralized frame relative to H, and let (M, AH) be a pseudo-model relative

to H. Then (M, AH) is a generalized Henkin model.

Proof We need that AH is an abstraction designation function, Def-

inition 2.23. Specifically, we need that it maps predicate abstracts to

SOUNDNESS AND COMPLETENESS 63

members of Henkin domains. But part 1 of Theorem 4.21 takes care of

this.

And we need that AH is a proper abstraction designation function,

Definition 2.27. There are three conditions that must be met. The

first, that abstracts map to the 'right' values, is taken care of by the

way we defined AH in pseudo-models. The other two conditions have

to do with the behavior of substitution, and these are taken care of by

Propositions 4.15 and 4.16.

COROLLARY 4.24

Henkin model.

Every Hintikka set is satisfiable in a generalized

Proof By Corollary 4.22 and the Theorem above. (Recall, it was shown

in Section 6 that a choice between L( C) and L + (C) was not significant

when considering models for formulas from the language L(C).)

2.6 Completeness At Last

Most of the work of showing completeness is over. All that is left is

to connect Hintikka sets with tableaus. This can be done in either of

two ways. One could give a systematic tableau construction procedure,

designed to ensure everything that can be done is eventually done in fact.

Then one would show that the set of formulas on an unclosed branch

of such a tableau is a Hintikka set. This approach involves considerable

attention to detail, and is not what I have chosen to do here. The other

technique involves maximal consistent sets, much like in the standard

axiomatic approach. Things must be adapted to tableaus, of course, but

this is the direction I picked because it is considerably simpler.

DEFINITION 4.25 (CONSISTENCY) Call a setS of grounded formulas of

L + (C) consistent if no basic tableau beginning with any finite subset of

S closes. If S is not consistent, call it inconsistent.

Thus a set S is inconsistent if there is a closed tableau beginning with

some finite subset.

DEFINITION 4.26 (MAXIMAL CONSISTENCY) A set S is maximally

consistent if it is consistent but no proper extension of it is consistent.

For propositional logic, working with maximal consistent sets is suf-

ficient to prove completeness, but with quantifiers involved, more is

needed.

DEFINITION 4.27 (E-COMPLETE) A set S of grounded formulas of

L + (C) is E-complete if:

64 TYPES, TABLEAUS, AND GODEL'S GOD

1 --{1/o:)cll(o:) E S implies --,cp(p) E S for some parameter p.

2 (3o:)cll(o:) E S implies cll(p) E S for some parameter p.

It will be shown that lots of maximal consistent, E-complete sets

exist, and they are Hintikka sets. From this, completeness follows eas-

ily. The primary difference between a tableau completeness proof and

an axiomatic one is that with tableaus, maximal consistency and E-

completeness give us the implications that make up the definition of a

Hintikka set, while in an axiomatic version, these implications become

equivalences. The stronger version, in fact, is more than is needed. But

now, to work.

PROPOSITION 4.28 If S is a consistent set of closed formulas of L(C),

S can be extended to a maximal consistent, E-complete set of grounded

formulas of L+(c).

Proof The set of grounded formulas of L + (C) is countable; let 'lT 1, 'lT 2,

W3, ... be an enumeration of all of them. Also, let PI, P2, P3, ... be an

enumeration of all parameters of L+(C) of all types. Now we construct

a sequence of sets of formulas. Each set in the sequence will meet two

conditions: it is consistent, and infinitely many parameters of each type

do not appear in it. Here is the construction.

Let 8

0

= S. This is consistent by hypothesis, and contains no param-

eters at all, so both of the conditions are met.

Suppose Sn has been defined, and the conditions are met.

1 If Sn U {'l'n+I} is not consistent, let Sn+l = Sn.

2 If Sn U {'l'n+I} is consistent, and Wn+l is not an existentially quan-

tified formula or the negation of a universally quantified formula, let

Sn+l = Sn U {'l'n+I}

3 Finally, if Sn U {'l'n+I} is consistent, and Wn+l is (3o:)cll(o:), choose

the first parameter p in the enumeration of parameters, of the same

type as o:, that does not appear in Sn or in (3o:)cll(o:), and set Sn+l =

Sn U {(3o:)cll(o:), cll(p)}. And similarly if Wn+l is (Vo:)cll(o:).

Note that Sn+l meets the conditions again. In case 3, consistency needs

a small argument, which I leave to you.

Finally, let 8

00

be SoUS1 US2U .... I leave to you the easy verification

that 8

00

will be consistent, E-complete, and maximal.

PROPOSITION 4.29 If S is a set of grounded formulas of L+(C) that is

maximal consistent and E-complete, S is a Hintikka set.

SOUNDNESS AND COMPLETENESS 65

Proof Let S satisfy the hypothesis of the Proposition. It is a simple

matter to verify that S meets each of the Hintikka set conditions. One

is presented as an example.

Suppose we have (>.a1, ... , an.<I>(a1, ... , an)}(TI, ... , Tn) E S, but

<I>(TI, ... , Tn) rf. S; we derive a contradiction.

If S U { <I>(TI, ... , Tn)} were consistent, <I>(TI, ... , Tn) would be in S,

since S is maximally consistent. Consequently S U {<I>( 71, ... , Tn)} is

not consistent, so there is a closed tableau for some finite subset, which

must include <I>(TI, ... ,Tn), since S itself is consistent. Thus there are

formulas X1, ... , Xk E S such that there is a closed tableau, call it

T, beginning with xl, ... ' xk, <I>(TI, ... 'Tn) Now, since we have

(>.a1, ... , an.<I>(a1, ... , an)}(TI, ... , Tn) E S we can construct a tableau

as follows. Begin with

xk k.

(>.a1, ... , an.<I>(a1, ... , an)}(TI, ... , Tn) k + 1.

<l>(TI, ... , Tn) k + 2.

In this, the first k + 1 lines are members of S. Line k + 2 is from k + 1 by

an abstract rule. Now continue this tableau to closure by copying over

the steps of tableau T.

This shows there is a closed tableau for a finite subset of S itself, so

S must be inconsistent, which is a contradiction.

Now, finally, we get" the completeness results.

THEOREM 4.30 Let <I> be a closed formula and letS be a set of closed

formulas, all of L( C).

1 If <I> is valid in generalized Henkin models, <I> has a basic tableau proof.

2 If <I> is a generalized Henkin consequence of S, <I> has a basic tableau

derivation from S.

Proof Suppose there is no basic tableau derivation of <I> from S. Then

there is no closed tableau for ...,<f>, allowing members of S to be added

to the ends of open branches. It follows that S U { ...,<f>} is consistent. It

can be extended to a maximal consistent, E-complete set H, by Propo-

sition 4.28. The set His a Hintikka set, by Proposition 4.29. Then by

Corollary 4.24, SU{ -,<f>} is satisfiable in some generalized Henkin model,

and consequently <I> is not a generalized Henkin consequence of S. This

establishes part 2; part 1 has a simpler proof.

66 TYPES, TABLEAUS, AND GODEL'S GOD

3. Miscellaneous Model Theory

Two of the main results about first-order logic are the Compactness

and the Lowenheim-Skolem theorem. I already noted, in Section 3,

that compactness does not hold for "true" higher-order logic. It is also

easy to verify that the Lowenheim-Skolem theorem does not hold, since

one can write a formula asserting an uncountable object exists. But

things are very different if generalized Henkin models are used, instead

of standard models. Then both theorems hold, just as in the first-order

case. Compactness is easy to verify, now that completeness has been

shown. Lowenheim-Skolem takes more work.

THEOREM 4.31 (COMPACTNESS) LetS be a set of closed formulas of

L( C). If every finite subset of S is satisfiable in some generalized Henkin

model, so is S itself.

Proof Suppose S is not satisfiable in any generalized Henkin model-!

show some finite subset of S is also not satisfiable.

Let ..L abbreviate X 1\ X, where X is some arbitrary closed formula

of L( C). Since S is not satisfiable in any generalized Henkin model, ..L is

true in every model in which the members of S are true (since there are

none), so ..Lis a generalized Henkin consequence of S. By Completeness,

..L has a basic tableau derivation from S. A closed tableau, being a

finite object, can use only a finite subset So of S. Now ..L has a basic

tableau derivation from So, so by Soundness, ..Lis a generalized Henkin

consequence of So. If So were satisfiable in some generalized Henkin

model, ..L would be true in it, which is not possible. Consequently S

0

is

unsatisfiable.

The Lowenheim-Skolem theorem for first-order classical logic follows

easily from the observation that models constructed in completeness

proofs are countable. This does not apply directly to the generalized

Henkin models constructed using tableaus. The reason is very simple.

I showed how to construct a generalized Henkin frame M = (1-lH, I,)

starting with a Hintikka set H. In this frame, the Henkin domains

consisted of possible values for grounded terms, Definition 4.9. It is easy

to see that 1-lH(O) must be countable. But say T is a grounded term of

type (0) such that no formulas of the form r(ro) or r(ro) occur in H.

(This can certainly happen-take the Hintikka set H to be the empty

set!) Then (r, S) is a possible value for T for every subset S of 1-lH(O),

so 1-lH( (0)) is uncountable.

We need some way around this difficulty. The main tool is contained

in the following.

SOUNDNESS AND COMPLETENESS 67

THEOREM 4.32 (CUT-ELIMINATION) LetS be a finite set of grounded

formulas of L + (C). If there is a closed tableau beginning with S U { <P},

and a closed tableau beginning with S U { --,cp}, then there is a closed

tableau beginning with S.

This Theorem is a version of Gentzen's famous Haputsatz, or cut elim-

ination theorem, for higher-order logic. It is an important result about

classical first-order logic that closed tableaus for SU{ <P} and for SU{ --,cp}

can be constructively converted into one for S. There is no constructive

proof for the higher-order case, but the result can be obtained provided

we are willing to drop constructivity. Such a proof was given in [Pra68]

and in [Tak67], and their argument has appeared here, in disguise, as

a completeness proof. To finish things off I sketch the remaining ideas

involved in a proof of the Theorem.

Proof Suppose there are closed tableaus for S U { <P} and for S U { --,cp}.

Then neither set is satisfiable. It follows that S itself is not satisfiable,

for if there were a generalized Henkin model in which its members were

true, one of <P or --,cp would be true there. It remains to show that the

unsatisfiability of S implies there must be a closed tableau beginning

with S.

Suppose the contrary: there is no closed tableau beginning with S,

so that S is a consistent set. Proposition 4.28 says a consistent set of

L( C) sentences can be extended to a maximal consistent, E-complete

set-the same proof can easily be made to work even if the set contains

parameters, provided it omits infinitely many of them. Since S is finite,

it certainly omits infinitely many parameters, so we can extend it to

a maximal consistent, E-complete set, which must be a Hintikka set.

Corollary 4.24 says Hintikka sets are satisfiable. Since Sis a subset of a

satisfiable set, it too must be satisfiable, but it is not. This contradiction

concludes the proof.

This immediately gives us the following important result.

COROLLARY 4.33 (CUT RULE) The addition of the following Cut Rule

to the basic tableau system does not change the class of provable formulas:

at any point split a branch, and add --,cp to one fork, and <P to the other,

where <P can be any grounded formula.

The way this result is most often used is embodied in the following.

COROLLARY 4.34 If <I> has a tableau proof, <I> can be added as a line to

any tableau, without expanding the class of provable formulas.

68 TYPES, TABLEAUS, AND GO DEL'S GOD

Proof Suppose has a tableau proof, and so there is a closed tableau

for And now suppose we are constructing another tableau, and we

wish to use in that construction, We can proceed as follows.

That is, we have used an application of a cut. Now, on the left branch,

introduce the steps appropriate to close it, which exist because we are

assuming there is a closed tableau This leaves the right branch.

The net effect has been to add to the tableau.

Now, go back through the proof of completeness given earlier. Propo-

sition 4.28 said we could extend a consistent set to a maximal consistent,

E-complete one. Using the work above, it follows that a maximal con-

sistent set must contain either or for every grounded formula

Since this is the case, each grounded term can, in fact, have only one pos-

sible value associated with it. Thus the particular model constructed in

the completeness argument must have countable Henkin domains, since

the family of grounded terms for each type is countable. We thus have

the following.

THEOREM 4.35 (LOWENHEIM-SKOLEM) Let 8 be a set of closed for-

mulas of L(C). If S is satisfiable in some generalized Henkin model, S

is satisfiable in a generalized Henkin model whose domain function 1t

meets the condition that 1t(t) is countable for every type t.

The results above have both good and bad points. It is obviously good

to be able to prove such powerful model-theoretic facts about a logic-it

provides tools for the construction of useful models. The bad side is

that Lindstrom's Theorem says, since the version of higher-order logic

based on generalized Henkin models satisfies the theorems above, it is

simply an equivalent of first-order logic. This does not mean nothing has

been gained. The higher-order formalism is natural for the expression

of things whose translation into first-order versions would be unnatural.

And finally, if a sentence is not provable, it must have a generalized

Henkin counter-model, but if it is provable, it must be true in all gen-

eralized Henkin models, and among these are the standard higher-order

models! Thus we have a means of getting at higher-order validities-we

just can't get at all of them this way.

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 5

EQUALITY

The basic tableau rules of Chapter 3 do not give any special role to

equality. It is time to bring it into the picture. This is done by adding

axioms to the tableau system, which has the effect of narrowing things

to normal generalized Henkin models. In addition, some useful derived

tableau rules will be presented.

1. Adding Equality

Leibniz's principle is that objects are equal just in case they have

the same properties. This principle is most easily embodied in axioms,

rather than in tableau-style rules.

DEFINITION 5.1 (EQUALITY AXIOMS) Each sentence of the following

form is an equality axiom:

(Va)(V,B)[(a =,B)= (V1)(1(a) :) 1(,8))]

In this, = is of type (t, t), for some t, then a and ,B are of type t and 1

is of type (t). EQ denotes the set of equality axioms.

I will show that a closed formula <I> of L( C) is valid in normal gen-

eralized Henkin models if and only if <I> has a tableau derivation from

EQ. But before that is done I give some handy derived tableau rules,

and examples of their use.

2. Derived Rules and Tableau Examples.

There are two derived rules involving equality that are more "tableau-

like" in flavor, and are what I primarily use in constructing tableau

proofs and derivations. I do not know if they can serve as full replace-

ments for the official Equality Axioms, since I have been unable to prove

69

70 TYPES, TABLEAUS, AND GODEL'S GOD

a completeness theorem using them. Nonetheless, the derived rules be-

low are the ones I generally use in practice.

DEFINITION 5.2 (DERIVED REFLEXIVITY RULE) For a grounded term

T of L + (C), at any point in a proof ( T = T) may be added to the end of

a tableau branch. Schematically,

Justification of Derived Reflexivity Rule Let T be a grounded

term of type t. (T = T) can be added to the end of a branch via the

following sequence of steps.

(\fa)(\f/3)[(a = /3) = (\f'Y)(r(a) :J 'Y(/3))] 1.

(\ff3)[(T = /3) := (\f'Y)('y(T) :J 'Y(/3))] 2.

[(T = T) := (\f'Y)('y(T) :J /(T))] 3.

[(T = T) :J (\fr)(r(T) :J 'Y(T))] 4.

[(\f'Y)('y(T) :J 'Y(T)) ::J (T = T)] 5.

/

(\f'Y)(/(T) :J 'Y(T)) 6. (T = T) 7.

In this, 1 is an equality axiom; 2 is from 1 and 3 is from 2 by universal

rules; 4 and 5 are from 3 by a conjunction rule; 6 and 7 are from 5 by

a disjunction rule. Clearly the left branch continues to closure. The

remaining open branch, the right one, indeed, has (T = T) on it.

The next rule embodies the familiar notion of substitutivity of equals

for equals.

DEFINITION 5.3 (DERIVED SUBSTITUTIVITY RULE) Suppose <J?(a)

is a formula of L + (C) in which the variable a may have free occur-

rences, but no other variables occur free. Also suppose Tl and T2 are

grounded terms of the same type as a. As usual, let <I>( Tl) denote the

result of replacing free occurrences of a in <I>( a) with occurrences of T1;

and similarly for <I>(T2). Then, if both <P(TI) and (Tl = T2) occur on a

tableau branch, <I>( T2) can be added to the branch end. Schematically,

<J?( T1)

(T1=T2)

<J?( T2)

Justification of Derived Substitutivity Rule Assume T

1

and T

2

are grounded terms of type t, and <I>(TI) and (T1 = T2) occur on a tableau

EQUALITY

branch. I show <P( 72) can be added to the end of the branch.

(Va)(V,B)[(a = ,6) = (Vr)('y(a) ::::> 1(,6))] 1.

(\f ,6) [ ( 71 = ,6) := (\fr) ('y( 71) ::::> 1'(,6))] 2.

[(71 = 72) := (\7'1')(1'(71) ::::> 1'(72))] 3.

[(71 = 72) ::::> (\f/)('y(7I) ::::> 1'(72))] 4.

[('v'r)('y(7I) ::::> 1'(72)) ::::> (71 = 72)] 5.

~

(71 = 72) 6. (\f/)('y(7I) ::::> /(72)) 7.

(Aa.<P(a))(7I) ::::> (Aa.<l>(a))(72) 8.

/ ~

(Aa. <P( a)) ( 7I) 9. (Aa.<P( a)) ( 72) 10.

<1>(71) 11. <1>(72) 12.

71

In this, 1 is an equality axiom; 2 is from 1 and 3 is from 2 by universal

rules; 4 and 5 are from 3 by a conjunction rule; 6 and 7 are from 4

by a disjunction rule; 8 is from 7 by a universal rule, using the term

(Aa.<P(a)); 9 and 10 are from 8 by a disjunction rule; 11 is from 9 and

12 is from 10 by a predicate abstract rule. The two left branches are

closed, leaving the right one which contains <1>(72)

Now I give several examples of tableau derivations using the derived

rules. The first example is (intentionally) a simple one. It appeared

earlier as Example 2.9, where an informal reading was given, and validity

was shown directly.

EXAMPLE 5.4 Here is a proof of (AX.(3x)X(x))((Ax.x =c)).

(AX.(3x)X(x))( (Ax.x =c)) 1.

(3x)(Ax.x = c)(x) 2.

(Ax.x = c)(c) 3.

(c=c) 4.

(c=c) 5.

In this, 2 is from 1 by an abstract rule; 3 is from 2 by a universal rule;

4 is from 3 by an abstract rule, and 5 is by the derived reflexivity rule.

The next example shows how, by using the derived rules, we can

reverse things and prove a version of the equality axioms.

72 TYPES, TABLEAUS, AND GODEL'S GOD

EXAMPLE 5.5 I give a tableau proof (using derived rules, not axioms)

of

(Va)(V,B)[(Vr)('y(a) ~ 1(,8)) ~ a = ,B)].

--,(\fa)(V,B)[(V')')('Y(a) ~ 1(,8)) ~ (a= ,B)] 1.

--,(\f,B)[(Vr)(r(P) ~ 1(,8)) ~ (P =,B)] 2.

--,[(Vr)('y(P) :::> r(Q)) ~ (P = Q)] 3.

(Vr)('y(P) ~

1

(Q)) 4.

--,(P = Q) 5.

(AX.--,(X = Q))(P) ~ (AX.--,(X = Q))(Q) 6.

/ ~

--,(AX.--,(X = Q))(P) 7.

--,--,(P = Q) 9.

(AX.--,(X = Q))(Q) 8.

--,( Q = Q) 10.

(Q = Q) 11.

Here 2 is from 1, and 3 is from 2 by an existential rule (P and Q

are new parameters of the appropriate type); 4 and 5 are from 3 by a

conjunctive rule; 6 is from 4 by a universal rule, using the grounded term

(AX.--,(X = Q)); 7 and 8 are from 6 by a disjunction rule; 9 and 10 are

from 7 and 8 by abstract rules; 11 is the derived reflexivity rule.

Though the derived tableau rules for equality allow us to prove the

axioms, it does not follow they are their equivalent. To establish that,

we would need to have a cut elimination theorem for the tableau system

with the equality rules. And the way to prove cut elimination is to first

have a completeness proof. I conjecture that such a completeness result

is provable, but I don't know how to do it.

Exercises

EXERCISE 2.1 Prove the following characterization of equality-it says

it is the smallest reflexive relation.

(Vx)(Vy){(x = y) = (VR)[(Vz)R(z,z) ~ R(x,y)]}

EXERCISE 2.2 Give a tableau derivation of the following from EQ.

(Va)(V,B)[(a =,B) ~ (Vr)(a('y) = ,8(1))]

More generally, one can do the same with the following.

EQUALITY 73

3. Soundness and Completeness

The results of this section combine to prove the following.

THEOREM 5.6 Let <P be a closed formula and let S be a set of closed

formulas of L( C).

1 <P is valid in all normal generalized Henkin models if and only if <P

has a tableau derivation from EQ.

2 <P is a consequence of S with respect to normal generalized Henkin

models if and only if <P has a tableau derivation from S U EQ.

The theorem above combines soundness and completeness. One di-

rection, soundness, is almost immediate. Every equality axiom is true in

every normal generalized Henkin model, so the implications from right

to left in Theorem 5.6 follow immediately from Theorems 4.4 and 4.5.

As usual, the completeness direction is more work. The key item is to

prove the following Proposition. Once we have it, completeness follows

immediately using part 2 of Theorem 4.30.

PROPOSITION 5. 7 Given a generalized Henkin model in which all mem-

bers ofEQ are true, there is a normal generalized Henkin model in which

exactly the same closed formulas of L( C) are true.

The rest of this section is given over to a proof of Proposition 5. 7 -it is

broken up into constructions and Lemmas. The ideas are the same as in

Godel's original completeness proof for first-order logic with equality-

bring equivalence classes into the picture.

For the rest of this section, assume (M, A) is a generalized Henkin

model, M = (H,I,), and all members of EQ are true in this model.

For 01,02 E H(t), let us write 01 =r 02 as a more readable alter-

native notation for (01,02) E I(=(t,t)). Thus =r is the interpretation

of the equality constant symbol (of a particular type, which will be in-

dicated only if needed). Since all equality axioms are true in (M, A), it

is an easy consequence that =r is an equivalence relation.

For each 0 E H ( t), let 0 be the equivalence class determined by 0,

that is, 0 = { O' I 0 =r O'}. Define a new Henkin domain mapping

by setting H(t) = {0 I 0 E H(t)}. Also, define a new interpretation

by setting I(A) to be the equivalence class containing I(A), that is,

I(A) = I(A).

LEMMA 5.8 If 01 = 02 then (01) = (02)

74 TYPES, TABLEAUS, AND GODEL'S GOD

Proof Suppose 01 = 02, that is, 01 =I 02, and say 01 and 02 are of

type (t). In Exercise 2.2 you were asked to give a tableau derivation of

(Va)(Vj])[(a = (3) ::J (V'Y)(a('Y) = {3(/))] from EQ. Then by soundness,

this sentence is valid in (M, A). It follows that (V'Y)(a('Y) = (3(!)) is

also true with respect to any valuation assigning 01 to a and 02 to (3.

From this we immediately get that the sets (01) and (02) must be the

same. A similar argument applies if 01 and 02 are of type (tb ... , tn)

The Lemma above justifies the following. For 0 E 1{ ( (t1, ... , tn)), set

(0) = { (01, ... , On) I (01, ... , On) E (0)}.

We have now created a new generalized Henkin frame M = (?t,I,).

LEMMA 5.9 The generalized Henkin frame M = (1t,I, ) is normal.

Proof The following calculation establishes normality.

(01, 02) E E(I(=)) {:} (01, 02) E E(I(=))

{:} (01,02) E E(I(=))

{:} 01 =I 02

{:} 01 = 02

For each valuation v in M, let v be the corresponding valuation in

M given by v(a) = v(a). It is easy to see that each valuation in M is

v for some valuation v in M.

LEMMA 5.10 Let T be a predicate abstract. If v1 = v2 then A( v1, T) =I

A(v2, r).

Proof For convenience say T just has one free variable, 1; the more gen-

eral case is treated similarly. From now on I'll write T as r('Y). Assume

v1 = v2, hence in particular, v1 (!) =I v2 (!); I'll show A( v1, r('Y)) =I

A(v2, r('Y)).

Let a and (3 be variables of the same type as 1, that do not occur

in r('Y) (free or bound). Since all members of EQ are true in (M, A),

(Va)(Vf3)[(a = (3) ::J (r(a) = r((3))] is true in it, and hence r(a) = r((3)

is true in (M, A) with respect to any valuation v such that v(a) =I v(f3).

Set w to be a particular valuation such that w(a) = v1(!), w(f3) =

v2(!), and otherwise w is arbitrary. Since v1('Y) =I v2(!), we have

w(a) =I w((3), so by the paragraph above, r(a) = r((3) is true in (M, A)

with respect tow, in other words, A(w,r(a)) =I A(w,r(f3)). Also

EQUALITY 75

= v1(a{ah}) = v1(r) = w(a). Likewise = w(,6).

Thus and w agree on the free variables of T(a), and and

w agree on the free variables of T(,6).

Now since A is proper, we can make use of the conditions of Defini-

tion 2.27, and we have the following.

A(v1,T(r)) = A(vi,T(a){ah})

=

= A(w, T(a))

=I A( w, T(,6))

=A( T(,6))

= A( V2' T(,6){,6 h})

= A(v2, T(r))

This Lemma justifies the following. Define an abstraction designation

function by A(v, (-Xa1, ... ,an.<P)) = A(v, (-Xa1, ... ,an.<P)).

Now the final step.

LEMMA 5.11 For each valuation v in M:

1 (v *I* A)(T) = (v *I* A)(T) for each term T of L(C);

2 M lf-v,A <P {:::::::::> M lf-v,A <P for each formula <P of L(C);

3 A is proper, and hence (M, A) is a generalized Henkin model.

Proof Part 1 follows for variables, constant symbols, and predicate ab-

stracts by definition of v, I, and A respectively.

Part 2 is by an induction on the degree of <P, which I leave to you.

Finally, for part 3 it is necessary to verify the three parts of Defini-

tion 2.27. I check part 3 and leave the other parts to you. Let 0' be a

substitution that is free for (>.a1, ... , an.<P), a term which I abbreviate

as T. It must be shown that A(v, TO') = A(W, T). We have the following.

A(v, TO')= A(v, TO')

= A(vu,T)

= A(vu,T)

76 TYPES, TABLEAUS, AND GO DEL'S GOD

But also we have the following, for each variable o:.

Exercises

va ( o:) = va ( o:)

= v(o:o-)

= v(o:o-)

= VU(o:)

EXERCISE 3.1 Give the details of the proof that =x is an equivalence

relation.

EXERCISE 3.2 Supply the proof of part 2 of Lemma 5.11.

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 6

EXTENSIONALITY

Extensionality says that properties applying to the same objects are

identical. Just as was done with equality in Chapter 5, extensionality

is added via axioms. Throughout this chapter it is, of course, assumed

that equality is available.

1. Adding Extensionality

The extensionality axioms simply assert the equality of co-extensional

properties.

DEFINITION 6.1 (EXTENSIONALITY AXIOMS) Each sentence of the fol-

lowing form is an extensionality axiom, where a and (3 are of type

(tl, ,tn), '/'1 is oftypetl, ... , 'T'n is oftypetn.

EXT denotes the set of extensionality axioms.

I will show that a closed formula <I> of L( C) is valid in normal Henkin

models (note that I've dropped the qualifier "generalized") if and only

if <I> has a tableau derivation from EQ U EXT. But first some examples.

2. A Derived Rule and an Example

Extensionality was embodied in a set of axioms. There is a derived

tableau rule that expresses the same idea in a rather more useful form.

DEFINITION 6.2 (DERIVED EXTENSIONALITY RULE) Suppose T1 and

T2 are two grounded terms, both of type (t1, ... , tn) At any point in a

tableau construction the end of a branch can be split, with one fork labeled

77

78 TYPES, TABLEAUS, AND GODEL'S GOD

(71 = 72), and the other fork labeled (7I(PI, ... ,Pn) = 72(PI, ... ,Pn))

where Pl, ... , Pn are parameters (of appropriate types) new to the

branch. Schematically, for new parameters:

The justification of this rule is quite straightforward, and I leave it as

an exercise. Here is an example that illustrates the use of this Derived

Extensionality Rule.

EXAMPLE 6.3 I give a proof of the following formula.

(\fx) [(>.X.X(x))(P) = (>.X.X(x))(Q)] :::>

(>.X, X, Y.X(X) :::> X(Y))(P,P,Q)

{(Vx) [(>.X.X(x))(P) = (>.X.X(x))(Q)]

:::> (>.X, X, Y.X(X) :::> X(Y))(P, P, Q)} 1.

(\fx) [(>.X.X(x))(P) = (>.X.X(x))(Q)] 2.

(>.X,X,Y.X(X) :::> X(Y))(P,P,Q) 3.

--, [P(P) :::> P(Q)] 4.

P(P) 5.

P(Q) 6.

--, [P(p) = Q(p)] 7.

(.XX.X(p))(P) =

(.XY.Y(p))(Q) 10.

P=Q 8.

P(Q) 9.

In this, 2 and 3 are from 1 by a conjunctive rule; 4 is from 3 by an

abstract rule; 5 and 6 are from 4 by a conjunctive rule. Now I apply the

extensionality rule. Take 7

1

to be P and 72 to be Q, both of which are

grounded, and take p to be a new parameter. We get a split to 7 and 8.

Item 9 is from 5 and 8 by substitutivity, and the right branch is closed.

Item 10 is from 2 by a universal rule. The left branch can be continued

to closure. I leave this to you.

Exercises

EXERCISE 2.1 Give a proof of formula (3.1) from Example 3.12.

EXERCISE 2.2 Show that the rule contained in Definition 6.2 is, in fact,

a derived rule, using EXT.

EXTENSIONALITY 79

3. Soundness and Completeness

I sketch a proof that the sentences having tableau proofs using EQ U

EXT as axioms are exactly the sentences valid in normal Henkin models

(and similarly for derivability as well).

Soundness takes very little work. It just amounts to the observation

that all members of EQ U EXT are valid in normal Henkin models.

Completeness also takes very little work. Using results of Chapter 4,

if a sentence If? does not have a tableau proof using EQ U EXT as axioms,

there is a generalized Henkin model in which If? is false, but in which all

of EQ U EXT are true. I'll show it follows that If? is false in a normal

Henkin model.

Say (M, A), where M = (1t, I,), is a generalized Henkin model

in which the members of EQ U EXT are true but If? is false. Since the

members of EQ are true, by results of Chapter 5 we can take (M, A) to

be normal. I claim it is also extensional in the sense of Definition 2.32,

that is, if (0) = (0') then 0 = 0', where 0 and O' are objects in the

model domain. I now show this.

Suppose (0) = (0'), where 0 and O' are of type (t) for simplicity

(the general case is similar). The following is a member of EXT (in it, a

and {3 are of type (t), and 1 is of type t)

(Va)(V,B){('v'!)[a(/) = ,8(/)] :) [a= ,8]}

and so this sentence is true in (M, A). Let v be a valuation such that

v(a) = 0 and v(f3) = 0'. Then

But since (0) = (0') it is easy to see we also have

and so M lf-v,A a= ,6. Since (M,A) is normal, it follows that v(a) =

v({3), that is, 0 = 0'.

Since (M, A) is extensional, it is isomorphic to a Henkin model, as

was shown in Section 6. And trivially, isomorphism preserves sentence

truth.

II

MODAL LOGIC

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 7

MODAL LOGIC

SYNTAX AND SEMANTICS

1. Introduction

The second part of this book investigates a logic of intensions and

extensions, using a possible world semantics. For purposes of background

discussion in this section, I will assume you have some general familiarity

with possible worlds at least informally. Technical details are postponed

till after that.

First, a point about terminology. The intensional/extensional distinc-

tion is an old one. Unfortunately, the word "extensionality" has already

been given a technical meaning in Part I, where Henkin models that did

or did not satisfy the extensionality axioms were considered. The use of

"extension" in this part, while related, is not the same. I briefly tried

using the word "denotation" here, but finally it seemed unnatural, and

I resigned myself to using the word "extension" after all. As a matter of

fact, the Axioms of Extensionality will be assumed throughout Part II

for those terms that will be called extensional, so any confusion of mean-

ings between the classical and the modal settings should be minimal.

The machinery in Part I had no place for intensions-meanings. In a

normal Henkin model, if terms intended to denote the morning star and

the evening star have the same extension, as they do in the real world,

they are equal, and so share all properties. They cannot be distin-

guished. Montague and his students, notably Gallin, developed a purely

intensional logic. In this, extensions could only be handled indirectly-

in some sense an extension could be an intension that did not vary with

circumstances. While this can be made to work, it treats extensions

as second class objects, and leads to a rather complicated development.

83

84 TYPES, TABLEAUS, AND GODEL'S GOD

What is presented here is a modification of the Montague/Gallin ap-

proach, in which both extensions and intensions are first class objects.

What are the underlying intuitions? An extensional object will be

much as it was in Part I: a set or relation in the usual sense. The added

construct is that of intensional object, or concept, and this is treated

in the Carnap tradition. A phrase like, say, "the royal family of Eng-

land," has a meaning, an intension. At any particular moment, that

meaning can be used to determine a particular set of people, constitut-

ing its extension. But that extension will vary with time. For other

phrases, there may be different mechanisms for determining extensions

as circumstances vary. The one thing common to all such intensional

phrases is that they, somehow, induce mappings from circumstances to

extensions. Abstracting to the minimum useful structure, in a possible

world model an intensional object will be a function from possible worlds

to extensional objects.

Here is an example using the terminology just introduced. Suppose we

take possible worlds as people, with an 85 accessibility

person is accessible to every other person. And suppose the ground-level

domain is a bunch of real-world objects. Any one person will classify

some of those objects as being red. Because of differences in vision,

and perhaps culture, this classification may vary from person to person.

Nonetheless, there is a common concept of red, or else communication

would not be possible. We can identify it with the function that maps

each person to the set of objects that person classifies as red. And

similarly for other colors. In addition, each person has a notion of color,

though this too may vary from person to person. One person may think

of ultra-violet as a color, another not. We can think of the color concept

as a mapping from persons to the set of colors for that person. If we

assert that red is a color for a particular person, we mean the red concept

is in the extension of the color concept for that person. The extension

of the red concept for that person plays no role for this purpose.

Sometimes extensions are needed too. Certainly if we ask someone

whether or not some object is red, the extension of the red concept, for

that person, is needed to answer the question. Here is another example

in this direction. Assume the word "tall" has a definite, non-fuzzy,

meaning. Say everybody gets together and votes on which people are

tall, or say there is a tallness czar who decides to whom the adjective

applies. The key point is that the meaning of "tall," even though precise,

drifts with time. Average height of the general population has increased

over the last several generations, so someone who once was considered

tall might not be considered so today.

MODAL LOGIC, SYNTAX AND SEMANTICS 85

Now suppose I say, "Someday everybody will be tall." There is more

than one ambiguity here. On the one hand I might mean that at some

point in the future, everybody then alive will be a tall person. On the

other hand I might mean that everybody now alive will grow, and so

at some point everybody now alive will be a tall person. Let us now

read modal operators temporally, so that OX informally means that X

is true and will remain true, and OX means that X either is true or will

be true at some point in the future. Also, let us use T(x) as a tallness

predicate. (The examples that follow assume an actualist reading of the

quantifiers, and eventually I will adopt a version of a possibilist reading.

For present purposes, this is a point of no fundamental importance. For

now, think in terms of varying domain models, with quantifiers ranging

over different domains at different worlds.) The two readings of the

sentence are easily expressed as follows.

('v'x)OT(x)

0(\lx)T(x)

(7.1)

(7.2)

Formula (7.1) refers to those alive now, and says at some point they

will all be tall. Formula (7.2) refers to those alive at some point in the

future and asserts, of them, that they will be tall. All this is standard;

the problem is with the adjective "tall." Do we mean that at some point

in the future everybody (read either way) will be tall as they use the

word in the future, or as we use the word now? If we interpret things

intensionally, T(x) at a possible world would be understood according

to that world's meaning of tall. There is no way, using the present

machinery, to formalize the assertion that, at some point in the future,

everybody will be tall as we understand the term. But this is what is

most likely meant if someone says, "Someday everybody will be tall."

Here is another example, one that goes the other way. Suppose a

member of the Republican Party, call him R, says, "necessarily the pro-

posed tax cut is a good thing." Suppose we take as the possible worlds

of a model the collection of all Republicans, and assume a sentence is

true at a world if that Republican thinks the sentence true. (We as-

sume Republicans are entirely rational, so we don't have to worry about

contradictory beliefs.) Let us now take OX to mean that every Repub-

lican thinks X is the case, which means X is necessary for Republicans.

(Technically, this gives us an 85 modality.) How do we formalize the

sentence above? Let c be a constant symbol whose intended meaning

is, "the proposed tax cut," and let G be a "goodness" predicate. Then

OG(c) seems reasonable as a formalization. What should it mean to say

it is true for R?

86 TYPES, TABLEAUS, AND GODEL'S GOD

One possibility is that R means every Republican thinks the tax cut

is good, as R understands the word good. This may not be what was

meant. After all, R might consider something good only if it personally

benefitted him. Another Republican might think something good if it

eventually benefitted the poor. Such a Republican probably would not

think a tax cut good simply because it benefitted R but he might believe

it would eventually benefit the poor, and so would be good in his own

sense. Probably R is saying that every Republican thinks a tax cut is

good, for his own personal reasons. The notion of what is good can

vary from Republican to Republican, provided they all agree that the

proposed tax cut is a good thing. But the mere fact that we can consider

more than one reading tells us that a simple formalization like DG(c) is

not sufficient.

Here will be presented a logic of both intension and extension, of

both sense and reference. In one of the examples above, color is an in-

tensional object. It is a function from persons to sets of concepts like

red, blue, and so on. As such, it is the same function for each person.

The extension of color for a particular person is the color function eval-

uated at that person, and thus it is a particular set of concepts, such

as red but not infra-red, and so on, quite possibly different from person

to person. We need a logic in which both intensions and extensions are

first-class objects. The machinery for doing this makes for complicated

looking formulas. But I point out, in everyday discourse all the machin-

ery is present but hidden-we infer it from our knowledge of what we

think must have been meant. Formalization naturally requires complex

machinery-it is making explicit what our minds do automatically.

2. Types and Syntax

Now begins the formal treatment, starting with the notion of type. I

want it to include the types of classical logic, as defined in Section 1.

I also want it to include the purely intensional types of the Montague

tradition, as given in [Gal75].

DEFINITION 7.1 (TYPE) The notion of a type, extensional and inten-

sional, is given by the following.

1 0 is an extensional type.

2 If t1, ... , tn are types, extensional or intensional, (t1, ... , tn) is an

extensional type.

3 If t is an extensional type, jt is an intensional type.

A type is an intensional or an extensional type.

MODAL LOGIC, SYNTAX AND SEMANTICS 87

The ideas behind the definition above are these. As usual, 0 is to

be the type of ground-level objects, unanalyzed "things." The type

(t1, ... , tn) is intended to be analogous to types in part I. The type jt is

the new piece of machinery-an object of such a type will be a function

on possible worlds. Recall the example involving colors from Section 1;

it can be used to give a sense of how these types are intended to be

applied. In that example, real-world objects are those of type 0. A set

of real-world objects is of type (0) so, for instance, the set of objects some

particular person considers red is of this type; this is the extension of

red for that person. The intensional object red, mapping each person to

that person's set ofred objects, is of type j(O). A set of such intensional

objects is of type (j(O)), so for a particular person, that person's set of

colors is of this type-the extension of color for that person. Finally,

the intensional object color, mapping each person to that person's set of

colors, is an object of type i(j(O)).

For another example, assume possible worlds are possible situations,

and the ground-level objects include people. In each particular situa-

tion, there is a tallest person in the world. The tallest person, in each

situation, is an object of type 0. The tallest person concept is an object

of type jO-it associates with each possible world the tallest person in

that possible world.

As a final example example, suppose t is an extensional type, so that

jt is intensional. The two-place relation: the intensional object X of type

jt has the extensional object y of type t as its extension, is a relation of

type (jt, t).

The language of Part I must be expanded to allow for modality. Just

as classically, C is a set of constant symbols of various types, containing

at least an equality symbol =(t,t) for each type t, though the set of types is

now larger. Note that the equality symbols themselves are of extensional

type. Using them we can form the intensional terms (>.x, y.x = y) and

(>.x, y.D(x = y)), as needed. We also have variables of each type. There

is one new piece of machinery, an operator l, which plays a role in term

formation. As usual, terms and formulas must be defined together in a

mutual recursion.

DEFINITION 7.2 (TERM OF L(C)) Terms are characterized as follows.

1 A constant symbol or variable of L( C) of type t is a term of L( C) of

type t. If it is a constant symbol, it has no free variable occurrences.

If it is a variable, it has one free variable occurrence, itself.

2 If <I> is a formula of L( C) and 0:1, ... , O:n is a sequence of distinct

variables of types t1, ... , tn respectively, then (>..o:1, ... , O:n.<I>) is a

88 TYPES, TABLEAUS, AND GODEL'S GOD

term of L( C) of the intensional type j (t1, ... , tn). Its free variable

occurrences are the free variable occurrences of <I>, except for occur-

rences of the variables a1, ... , an.

3 If 7 is a term of L(C) of type jt then 17 is a term of type t. It has

the same free variable occurrences that 7 has.

The predicate abstract (Aa1, ... , an. <I>) is of type j(t1, ... , tn) above,

and not of type (t1, ... , tn), essentially because <I> can vary its meaning

from world to world, and so (Aa1, ... , an.<I>) itself is world dependent.

Case 3 above makes use of what may be called an extension-of op-

erator, converting a term of an intensional type to a term of the corre-

sponding extensional one. Continuing with the color example, suppose

r is the intensional notion of red, of type j(O), mapping each person to

that person's set of red objects. Then for a particular person, 1r is that

person's set of red objects-the extension of r for that person, and an

extensional object of type (0).

Of course the symbols j and 1 were chosen to suggest their roles-in

a sense 1 'cancels' j. Nonetheless, 1 is a symbol of the language, while j

occurs in the metalanguage, as part of the typing mechanism.

DEFINITION 7.3 (MODAL FORMULA OF L(C)) The definition of for-

mula of L( C) is as follows:

1 If 7 is a term of either type (t1, ... , tn) or type i (t1, ... , tn), and

71, ... , 7n is a sequence of terms of types t1, ... , tn respectively,

then 7(71, ... , 7n) is a formula {atomic) of L(C). The free variable

occurrences in it are the free variable occurrences of 7, 71, ... , 7n

2 If <I> is a formula of L( C) so is -,q>. The free variable occurrences of

-,q> are those of <I>.

3 If <I> and \lT are formulas of L( C) so is (<I> 1\ \lf). The free variable

occurrences of (<I> 1\ \lf) are those of <I> together with those of \lf.

4 If <I> is a formula of L(C) and a is a variable then (\fa)<I> is a formula

of L( C). The free variable occurrences of (\fa )<I> are those of <I>, except

for occurrences of a.

5 If <I> is a formula of L( C) so is D<I>. The free variable occurrences of

D<I> are those of <I>.

Item 1 above needs some comment, and again the example concerning

colors should help make things clear. Suppose r is the intensional notion

of red, of type j(O). And suppose cis an extensional notion of color,

MODAL LOGIC, SYNTAX AND SEMANTICS 89

the set of colors for a particular person-call the person George. Also

let C be the intensional version of color, mapping each person to that

person's extension of color. cis of type (i(O) ), and Cis of type j(j(O)) I

take both C(r) and c(r) to be atomic formulas. If we ask whether they

are true for George, no matter which formula we use, we are asking if r

is a color for George. But if we ask whether they are true for Natasha,

we are asking different questions. C(r) is true for Natasha if r is a color

for Natasha, while c(r) is true for Natasha if r is a color for George.

No matter which, both c(r) and C(r) make sense, and are considered

well-formed.

I use 0 to abbreviate ...,o..., in the usual way, or I tacitly treat it as

primitive, as is convenient at the time. And of course other propositional

connectives and the existential quantifier will be introduced as needed.

Likewise outer parentheses will often be dropped.

3. Constant Domains and Varying Domains

Should quantifiers range over what does exist, or over what might

exist? That is, should they be actualist or possibilist? This is really a

first-order question. A flying horse may or may not exist. In the world

of mythology, such a being does exist. In the present world, it does

not. But the property of being a flying horse does not exist in some

worlds and lack existence in others. In the present world nothing has

the flying-horse property, but that does not mean the property itself is

non-existent. Thus actual/possible existence issues really concern type 0

objects, so the discussion that follows assumes a first-order setting.

As presented in [HC96] and [FM98], the distinction between actualist

and possibilist quantification can be seen to be that between varying

domain modal models and constant domain ones. In a varying domain

modal model, one can think of the domain associated with a world as

what actually exists at that world, and it is this domain that a quanti-

fier ranges over when interpreted at that world. In a constant domain

model one can think of the common domain as representing what does

or could exist, and this is the same from world to world. Of course a

choice between constant and varying domain models makes a substan-

tial difference: both the Barcan formula and its converse are valid in a

constant domain setting, but neither is in a varying domain one.

As it happens, while a choice between constant and varying domain

models makes a difference technically, at a deeper level such a choice is

essentially an arbitrary one. If we choose varying domains as basic, we

can restrict attention to constant domain models by requiring the Barcan

formula and its converse to hold. (Technically this requirement involves

an infinite set of formulas, but if equality is available a single formula will

90 TYPES, TABLEAUS, AND GODEL'S GOD

do.) Thus when using actualist quantification, we can still determine

constant domain validity. The other direction is even easier. If we

have possibilist-constant domain-:-quantification we can also determine

varying domain validity. And on this topic I present a somewhat more

detailed discussion.

Suppose quantification is taken in a possibilist sense--domains are

constant. Nonetheless, at each world we can intuitively divide the com-

mon domain into what 'actually' exists at that world and what does

not. Introduce a predicate symbol E of type j(O) for this purpose. At a

particular world, E(x) is true if x has as its value an object one thinks

of as existing at that world, and is false otherwise. Then the effect of

varying domain quantification can be had by relativising all quantifiers

to E. That is, replace (Vx)r.p by (Vx)(E(x) :) r.p) and replace (3x)r.p by

(3x)(E(x) 1\ r.p). What we get, at least intuitively, simulates an actualist

version of quantification.

All this can be turned into a formal result. Suppose we denote the

relativization of a first-order formula r.p, as described above, by r.pE. It

can be shown that r.p is valid in all varying domain models if and only if

r.pE is valid in all constant domain models. Possibilist quantification can

simulate actualist quantification. I note in passing that [Coc69] actually

has two kinds of quantifiers, corresponding to actualist and possibilist,

though it is observed that a quantifier relativization of the sort described

above could be used instead.

The discussion above was in a first-order setting. As observed earlier,

when higher types are present the actualist/possibilist distinction is only

an issue for type 0 objects. I have made the choice to use possibilist

type 0 quantifiers. The justification is that, first, such quantifiers are

easier to work with, and second, they can simulate actualist quantifiers,

so nothing is lost. When I say they are easier to work with, I mean

that both the semantics and the tableau rules are simpler. So there is

considerable gain, and no loss.

Officially, from now on the formal language will be assumed to contain

a special constant symbol, E, of type j (0), which will be understood

informally as an existence predicate.

4. Standard Modal Models

I begin the formal presentation of semantics for higher-order modal

logic with the modal analog of standard models. The new piece of se-

mantical machinery added to that for classical logic is the possible world

structure.

MODAL LOGIC, SYNTAX AND SEMANTICS 91

DEFINITION 7.4 (KRIPKE FRAME) A Kripke frame is a structure

(Q, R). In it, g is a non-empty set (of possible worlds), and R is a

binary relation on g (called accessibility). An augmented frame is a

structure (Q, R, V) where (Q, R) is a frame, and V is a non-empty set,

the (ground-level) domain.

The notion of a Kripke frame should be familiar from propositional

modal logic treatments, and I do not elaborate on it. As usual, different

restrictions on R give rise to different modal logics. The only two I will

be interested in are K, for which there are no restrictions on R, and

85, for which R is an equivalence relation. Note that the ground-level

domain, V, is not world dependent, since the choice was to take type-0

quantification as possibilist and not actualist.

Next I say what the objects of each type are, relative to a choice of

ground-level domain. This is analogous to what was done in Part I,

in Definition 2.1. To make things easier to state, I use some standard

notation from set theory. The first item is something that was used

before, but I include it here for completeness sake.

1 For sets A1, ... , An, A1 x x An is the collection of all n-tuples of

the form (a1, ... , an), where a1 E A1, ... , an E An. The 1-tuple (a)

is generally identified with a.

2 For a set A, P(A) is the power set of A, the collection of all subsets

of A.

3 For sets A and B, A B is the function space, the set of all functions

from B to A.

DEFINITION 7.5 (OBJECTS, EXTENSIONAL AND INTENSIONAL) Let

g be a non-empty set (of possible worlds) and let V be a non-empty

set (the ground-level domain). For each type t, I define the collection

[t, V, Q], of objects of type t with respect to V and Q, as follows.

1 [O,V,Q] =V.

2 [(t1, ... , tn), V, Q] = 'P([t1, V, Q] X" X [tn, V, Q]).

3 [jt, V, Q] = [t, V, Q]g.

0 is an object of type t if 0 E [t, V, Q]. 0 is an intensional or exten-

sional object according to whether its type is intensional or extensional.

As before, 0 is used, with or without subscripts, to stand for objects.

Now the final notion of the section.

92 TYPES, TABLEAUS, AND GO DEL'S GOD

DEFINITION 7.6 (MODAL MODEL) A (higher-order) modal model for

L(C) is a structure M = (Q, R, V,I), where (Q, R, V) is an augmented

frame and I is an interpretation .. The interpretation I must meet the

following conditions.

1 If At is a constant symbol of type t, I(At) is an object of type t, that

is, I(At) E [t, V, Q].

2 If =(t,t) is an equality constant symbol, I( =(t,t)) is the equality relation

on [t, V,Q].

5. Truth in a Model

In this section I say how truth is to be assigned to formulas, at worlds,

in models, and how values should be assigned to terms. I lead up to a

proper definition after a few preliminary notions.

DEFINITION 7.7 ((MODAL) VALUATION) The mapping v is a modal

valuation in the modal model M = (9, R, V,I) if v assigns to each

variable at of type t some object of type t, that is, v( at) E [t, V, Q]. The

notion of a variant valuation is defined exactly as classically.

A term like lr is intended to designate the extension of the intensional

object designated by T. To determine this a context is needed-the

designation of T where, under what circumstances? The notation I'll

use for a designation function is (v *I* f)(T), where vis a valuation,

I is an interpretation, and r is a context, a possible world. (In fact the

context only matters for terms of the form lT.)

In specifying the designation of a term, the predicate abstract case

requires information about formula truth. This is more complex than

classically, again because a context must be specified-truth under what

circumstances, in which possible world. The notation for this is a mod-

ification of what was used earlier. I'll write M, r lf-v <P to mean formula

<P is true, with respect to valuation v, in model M, at possible world r.

Just as classically, term designation and formula truth must be defined

together, in one big recursion. Definitions 7.8 and 7.9 really are two parts

of the same definition-they have been separated for pedagogical reasons

only.

Recall the special notation that was introduced in Definition 2. 7. That

is extended to the present setting in the obvious way:

means M, r lf-w <P where w is the a1, ... , an variant of v such that

w(ai) = 01, ... , w(an) =On.

MODAL LOGIC, SYNTAX AND SEMANTICS 93

DEFINITION 7.8 (DESIGNATION OF A TERM) Let M = (Q, 'R, V,I)

be a modal model, let v be a valuation in it, and let r E g be a possible

world. Define a mapping ( v * I * r), assigning to each term an object

that is the designation of that term.

1 If A is a constant symbol of L(C) then (v *I* r)(A) = I(A).

2 If a is a variable then (v *I* r)(a) = v(a).

31fT is a term oftype jt then (v*I*r)(lT) = (v*I*r)(T)(r)

4 If(>-.a1, ... ,an.4>) isapredicateabstractofL(C) oftypej(tl, ... ,tn),

then ( v *I* r) ( (>-.a1, ... , an. 4>)) is the function f on possible worlds

given by the following.

Item 3 is a little awkward to read. ( v *I * r) ( T) (r) means: evaluate

T using ( v * I * r), getting a function, an intension, then evaluate that

function at r. Generally the simpler notation ( v * I * r) ( T, r) will be

used for this. Similarly for v(a, r) and I( A, r), when a and A are of

intensional type.

Item 4 tells us this is part of a mutual recursion-Definition 7.9 be-

low is the other part. Without using the special notation, part 4 of

Definition 7.8 reads as follows.

4 If (>-.a1, ... , an.4>) is a predicate abstract of L(C) of type j(t1, ... , tn),

then ( v * I * r) ( (>-.a1, ... , an. <I>)) is the function that assigns to an

arbitrary the following member of [ (t1, ... , tn), V, Q]:

{(w(a1), ... ,w(an)) I w is an a1, ... ,an variant ofv

and 11--w <I>}

The next item should be compared with Definition 2.6: worlds (con-

texts) must now be taken into account.

DEFINITION 7.9 (TRUTH OF A FORMULA) Let M = (Q, 'R, V,I) be a

modal model, and let v be a valuation in it. The notion of formula

4> being true at world r of g in model M with respect to v, denoted

M, r 11-v 4>, is characterized as follows.

1 For an atomic formula T( Tl, . .. , Tn),

{a) lfT is of an extensional type, M,r 11-v T(TI, ... ,Tn) provided

((v*I*r)(TI), ... ,(v*I*r)(Tn)) E (v*I*r)(T).

94 TYPES, TABLEAUS, AND GODEL'S GOD

(b) If T is of an intensional type, M, r lf-v T(T1, ... 'Tn) provided

M' r If-v (1 T) ( 71' . . . ' T n). This reduces things to the previous

case.

2 M, r If-v --,<fl if it is not the case that M, r If-v <P.

3 M,r lf-v <P 1\ w if M,r lf-v <P and M,r lf-v w.

4 M, r lf-v (Va)<P if M, r lf-v <P[a/0] for all objects 0 of the same type

as a.

5 M, r lf-v D<P if M, D. lf-v <P for all D. E g such that rnD..

As usual, other connectives, quantifiers, and modal operators can be

introduced via definitions, with the expected behavior. For instance:

M, r lf-v O<P if M, D. lf-v <P for some D. E g such that rnD..

Note that part 4, the quantifier case, can also be given as follows.

4 M, r lf-v (Va)<P if M, r lf-v' <P for every a-variant v' of v.

It follows from the definitions, that M, r lf-v ... , an/On]

if and only if M, D. lf-v <P[ai/01, ... , an/On] for all D. such that rnD..

It also follows from the definitions that (.Aa.<P(a))(r) and <P(r) are equiv-

alent under certain circumstances. For instance, this is the case if T is

a constant symbol of either intensional or extensional type. It is not

the case if T involves the extension-of operator, l Rather than give

exact conditions, I give the following, which is more useful for present

purposes-it is an immediate consequence of the definitions above.

PROPOSITION 7.10 Suppose that (v *I* r)(rl) = 01, ... , (v *I*

r)(rn) =On in model M. Then

M, r lf-v (.Aa1, ... 'an.<P)(TI, ... 'Tn) {:;} M, r lf-v <P[ai/01, ... 'an/On]

6. Validity and Consequence

Validity in a modal setting is now straightforward to define, but con-

sequence has a few surprises, so I've devoted a separate section to the

matter. Let us begin with what is simple.

DEFINITION 7.11 (VALIDITY) A formula <P of L(C) is valid in a model

M = (g, n, V, I) provided M, r If-v <P for all valuations v and all worlds

rEg.

<I> is K-valid if it is valid in all models, and is S5-valid if it is valid in

all models for which the accessibility relation is an equivalence relation.

MODAL LOGIC, SYNTAX AND SEMANTICS 95

I'll only be interested in sentences, for which the choice of valuation

is essentially irrelevant. In addition to the notion of validity in models,

a notion of frame validity can also be introduced. This is an important

concept, but it is not needed for the purposes of this book and I omit

discussion of it.

Logical consequence is inherently more complex than validity. When

asking about the consequences of some set of sentences, one must dis-

tinguish between sentences like "it is raining" and "rain is wet." The

first sentence, "it is raining" may be true in present circumstances, but

certainly it is not the case always. The second sentence, "rain is wet,"

presumably is true no matter what-it is independent of circumstance.

This gives rise to an important formal distinction. I'll say something is

a global assumption if we take it to be the case at every possible world,

and something is a local assumption if we take it to be the case in the

current possible world. The two kinds of assumptions behave quite dif-

ferently. Clearly global assumptions are taken to be not only true, but

necessary. That is not the case with local assumptions.

In the following, I'll use the expression 'true at a world' with the

obvious meaning.

DEFINITION 7.12 (CONSEQUENCE) LetL be one ofK orS5, letS and

U be sets of sentences of L( C), and <I> be a single sentence. <I> is a

consequence of global assumptions S and local assumptions U provided

that for every L model M = (g, R, 'D,I), if members of S are true at

every world in g, then <I> is true at each world at which members of U

are true.

Factual items would, most naturally, be local assumptions, while log-

ical principles would be global. Tableau rules differ for the two kinds of

assumptions. [Fit83] has a detailed discussion of the notions, including

appropriate versions of the deduction theorem. [Fit93] has a somewhat

more abbreviated treatment.

7. Examples

Here is a simple informal example to start with. Suppose we take

possible worlds to be points in time (within a reasonable range from near

past to near future). Also take the accessibility relation to always hold,

so that D<I> means <I> holds at all times. Does there exist, now, somebody

whose parents are necessarily not alive? Certainly-the oldest person

in the world. After all, the oldest person can never have living parents.

But on the other hand, there was a time when the oldest person had

living parents. There seems to be a discrepancy here.

96 TYPES, TABLEAUS, AND GODEL'S GOD

Say cp(x) is read as "x has no living parents." We are asking about

the truth of (3x)Dcp(x). The key question is, what type of variable

is x? If we think of the quantifier as ranging over objects-so x is of

type 0-then when we say the oldest person in the world instantiates

the existential quantifier we are saying a particular person does so. If

we designate the oldest person now as the value of x, instantiating the

existential quantifier, while cp(x) is certainly true now for this value of

x, there are earlier worlds in which the person who is the oldest now

had living parents. Thus we do not have Ocp(x), where x has as value

the oldest person in the present world. The proposed instantiation for

the existential quantifier does not work. More generally, it is easy to see

that (3x)Dcp(x) can never be true, now or at any other point of time,

provided we think of quantifiers as ranging over objects or individuals.

On the other hand if quantifiers range over individual concepts-so

that x is of type jO-we would certainly have the truth of (3x)Dcp(x)

since taking the value of x to be the oldest-person concept would serve

as a correct instantiation of Ocp(x).

The type theory of [Bre72] makes intensional objects basic. The

second-order logic of [Coc69] quantifies over extensional objects at the

first-order level, and over intensional objects at the second-order level.

The higher-order modal logic of [Fit98], which is a forerunner to this

book, had quantification only over extensional objects. The first-order

treatment of [FM98] involves a kind of mixed system, and more will be

said about it shortly. The system of [FitOOb] has quantifiers over types

0 and jO. Clearly many variations are possible.

Now for some further examples, which will be treated more formally.

ExAMPLE 7.13 Suppose x is a variable of type 0 and P is a constant

symbol of type j(O). The following formula is valid, where X is of type

j(O).

(.XX.0(3x)X(x))(P) :::> 0(-AX.(:Jx)X(x))(P) (7.3)

I leave it to you to verify the validity of this-one way is to show both

the antecedent and the consequent are equivalent to O(:Jx)P(x). On the

other hand, the following formula is not valid, where X is of type (0).

(.XX.O(:Jx)X(x))(lP) :::> 0(-AX.(:Jx)X(x))(lP) (7.4)

Here is an informal illustration to help you understand intuitively why

this formula is invalid. Suppose that on an island there are two societies,

MODAL LOGIC, SYNTAX AND SEMANTICS 97

optimists and pessimists, separated by a volcano. Let us say the opti-

mists, while generally positive in their outlook, are quite insecure and

don't accept something as possible (even if true) unless the pessimists

believe it. Further, the optimists think the volcano is beautiful, while

the pessimists think nothing is beautiful.

Take for P the concept of beauty-it maps each society to the set of

things that society accepts as beautiful. For the optimists the formula

(.XX.O(:Jx)X(x))(lP) is true for the following reasons. In the optimist

society the extension of P is the set consisting of the volcano, so the

formula asserts O(:Jx)X(x) is the case, when X is understood to be

that set. For optimists, (3x)X(x) is possible if the pessimists believe

it, and even the pessimists would agree that something is in the set

consisting of the volcano. On the other hand, O(.XX.(:lx)X(x))(lP) is

not true for the optimists, because (.XX.(:Jx)X(x))(lP) is not the case

for the pessimists, and this happens because the pessimists do not think

anything is beautiful.

This informal example can be turned into a formal argument. Here

is a model, M = (g, 'R, 'D,I), in which (7.4) is not valid. The collection

of worlds, Q, contains two members, r and with Think of

r as the optimists and as the pessimists. The domain, 'D is the set

{7} (think of the number 7 as the volcano). I show 7 available at both

worlds as a reminder that domains are constant. The constant symbol

P is interpreted to be a type j(O) object: the function that is {7} at r

and 0 at Thus I(P) is true of 7 at r, and of nothing at This

gives us the model presented schematically below.

r [2] I(P, r) = {7}

1

[2] = 0

The first claim is that, for an arbitrary valuation v, we have

M, r 11-v (.XX.O(:lx)X(x))(lP). (7.5)

Since (v *I* r)(lP) = (v *I* r)(P, r) = I(P, r) = {7}, by Proposi-

tion 7.10 we will have (7.5) provided we have

M,r 11-v O(:lx)X(x)[X/{7}] (7.6)

98 TYPES, TABLEAUS, AND GODEL'S GOD

which will be the case provided we have

M, 11-v (3x)X(x)[X/{7}].

But, since 7 E {7}, we have

M, 11-v X(x)[X/{7}, x/7]

and hence we have (7.7).

We thus have established (7.5). Next it is shown that

M, r IYv O(.XX.(3x)X(x))(!P)

which, together with (7.5), gives us the invalidity of (7.4).

Well, suppose otherwise, that is, suppose we had

M, r 11-v O(.XX.(3x)X(x))(lP).

Then we must have

M, 11-v (.XX.(3x)X(x))(!P),

and so, since (v *I* = 0,

M, 11-v (3x)X(x)[X/0].

(7.7)

(7.8)

(7.9)

(7.10)

(7.11)

(7.12)

It is easy to see we can not have this, and thus we have (7.9).

EXAMPLE 7.14 This example is one that is unexpected on superficial

consideration, although deeper thought says it should not be. The fol-

lowing formula is valid, with types of variables and constants as in ( 7.4).

(.XX.0(3x)X(x))(!P) :J (.XX.(3x)X(x))(!P) (7.13)

To show validity, suppose M = (Q, R, V,I) is an arbitrary model,

r E g is an arbitrary world, and v is an arbitrary valuation. Suppose

M, r 11-v (.XX.0(3x)X(x))(tp). (7.14)

Then

M, r 11-v 0(3x)X(x)[X/O] (7.15)

where 0 = ( v *I* r) (lP) = I(P, r). Then, for some E g such that

M, 11-v (3x)X(x)[Xj0] (7.16)

MODAL LOGIC, SYNTAX AND SEMANTICS 99

and so, for some object o we have

M, Llll-v X(x)[X/0, xjo]. (7.17)

For (7.17) to be the case, we must have o E 0. Now,

M, r 11-v X(x)[X/0, xjo] (7.18)

since o E 0. Consequently

M, r 11-v (3x)X(x)[X/O] (7.19)

and finally,

M, r 11-v (>.X.(3x)X(X))(!P) (7.20)

since 0 = (v *I* f)(JP).

Since we went from (7.14) to (7.20), the validity of (7.13) has been

established.

Some comments on the example above. The point is, the term 1 P

is given broad scope in both the antecedent and the consequent of the

implication. This essentially says its meaning in alternative worlds will

be the same as in the present world. Under these circumstances, exis-

tence of something falling under 1P in an alternate world is equivalent

to existence of something falling under !P in the present world. This is

just a formal variation on the old observation that in Kripke models, if

relation symbols could not vary their interpretation from world to world,

modal operators would have no visible effect.

The distinction between intensional and extensional types is complex.

The following two examples should help make clear the role of the 1

operator.

EXAMPLE 7.15 Let x and c be of type jO, and P be of type j(jO). The

following formula is valid.

DP(c) :J (3x)DP(x) (7.21)

To show (7.21) is valid let M = (g, R, D,I) be an arbitrary model,

r be an arbitrary world in g, and v be an arbitrary valuation. Suppose

we had the following.

M, r 11-v DP(c) (7.22)

Let Ll be an arbitrary world such that fRLl. We must have

100 TYPES, TABLEAUS, AND GO DEL'S GOD

M, 11-v P(c) (7.23)

from which it follows that

M, 11-v P(x)[xji(c)]. (7.24)

Since was arbitrary, we have

M, r 11-v DP(x)[x/I(c)] (7.25)

and hence

M, r 11-v (3x)OP(x) (7.26)

Since we went from (7.22) to (7.26), the validity of (7.21) has been

established.

EXAMPLE 7.16 This continues the previous example. Let c again be of

type jO, but now let x be of type 0 and P be of type j(O). The following

formula is not valid.

DP(lc) :=) (3x)DP(x) (7.27)

To show the non-validity of (7.27) a specific model, M = (Q, n, 1J,I),

is constructed. In this model, g consists of three possible worlds: r,

n. We have rnn, and R holds in no other cases. The domain 1J

is {1, 2}. I interprets c by a function that is 1 at 2 at n, and either

1 or 2 at r (it won't matter). Likewise I interprets P by the function

that is {1} at {2} at n, and some arbitrary value at r. Here is the

model schematically.

= 1

= {1}

n@J

I(c,n) = 2

I(P,O) = {2}

MODAL LOGIC, SYNTAX AND SEMANTICS 101

I leave it to you to check that (7.27) is not valid in this model.

EXAMPLE 7.17 The last example is in three parts. Consider the follow-

ing three formulas, in which x, y, and z are of type 0, and X, Y, and Z

are of type jO.

(V'Z)(..\x.D(..\y.x = y)(lZ))(lZ)

(V'z)(..\x.D(..\y.x = y)(z))(z)

(V'Z)(..\X.D(..\Y.X = Y)(Z))(Z)

(7.28)

(7.29)

(7.30)

Of the formulas above, (7.28) is not valid, but (7.29) and (7.30) are both

valid. I leave the work to you. I note that in [FM98] it was shown

that, in a first-order setting, the constructions used above relate directly

to rigidity. Both extensional and intensional objects, as such, are the

same from world to world, but the extensional object designated by an

intensional object can vary. This is what the example illustrates.

Exercises

EXERCISE 7.1 Show that formula (7.3) is valid.

EXERCISE 7.2 This is a variation on formula (7.13); the formula looks

the same, but the types are different. Show the validity of

(..\X.0(3x)X(x))(jP) :::> (..\X.(3x)X(x))(jP)

where xis of type jO and Pis of type j(jO). The fact that ground level

quantification is possibilist---constant domain-will be needed.

EXERCISE 7.3 Show the validity of the following, which looks a little

like a version of the Barcan formula: 0(3x)P(x) :::> (3X)OP(lX). In

this x is of type 0, X is of type jO and P is of type j(O).

EXERCISE 7.4 Show the non-validity of the following, where x is of type

0, X is of type (0), and Pis of type j(O).

O(:lx)(..\X.X(x))(jP) :::> (:lx)(..\X.OX(x))(lP)

EXERCISE 7.5 Verify the claims made in Example 7.17.

8. Related Systems

There have been many other versions of quantified modal logics in

the literature. Here I briefly say how a few of them relate to the one

presented here.

102 TYPES, TABLEAUS, AND GO DEL'S GOD

First-order modal logic, as given in [Fit83] or [HC96] (which also

contains a discussion of intensional objects) has variables and constant

symbols of type 0, and predicate symbols of types j(O, 0, ... , 0). Thus

quantification is over ground-level objects; constant symbols designate

such objects and hence are rigid. Predicates, of course, vary in meaning

from world to world-they are intensional. Treating them extensionally

would force modal logic to collapse to classical.

In [FM98], conventional first-order modal logic is extended by allowing

non-rigid terms, and an abstraction mechanism. Relating things to the

present system, variables are still of type 0, but constant symbols are

of type jO: they are individual concepts. Allowing intensional constant

symbols greatly enhances the expressibility of the language. Predicate

symbols are still of types j(O, 0, ... , 0). The fit between intension and

extension is achieved by treating (Ax.<I>) (c), where cis a constant symbol,

as if it were (Ax.<I>)(lc) in the present system. In effect, this means

the logic of [FM98] can be embedded in the higher-type version given

here. (Actually, this is not quite correct, since the logic of [FM98] allows

function symbols, and partial designation, neither of which is the case

here. But with these exceptions noted, the embedding claim is correct.)

Montague proposed a higher-order modal logic specifically as a logic

of intensions, in [Mon60, Mon68, Mon70]. It is presented most fully in

[Gal75]. Essentially it is the present system with only intensional types

(except at the lowest level). More specifically, define a Gallin/Montague

type, as follows.

1 0 is a Gallin/Montague type.

2 If t1, ... , tn are Gallin/Montague types, so is j(t1, ... , tn)

Then the logic of [Gal75] can be identified with the sublogic of the system

given here, in which all constant symbols and variables are restricted

to be of some Gallin/Montague type. Indeed, the present system was

created by adding extensional types to the logic of Gallin and Montague.

Bressan is a pioneer in the study of higher-order modal logics [Bre72].

I must confess that I do not fully understand his presentation. It is an 85

system rather like that of Gallin, though Gallin's is for a broader variety

of logics. In it extensional objects are not explicitly present, but rather

are identified with constant intensional objects. Also abstractions are

not taken as primitive, but are defined in terms of definite descriptions.

9. Henkin/Kripke Models

In the classical case there were good reasons for introducing non-

standard higher-order models, and those same reasons apply in the

MODAL LOGIC, SYNTAX AND SEMANTICS 103

modal case as well. Since modal versions of Henkin and generalized

Henkin models are relatively straightforward extensions of the classical

versions, I confine things to a brief sketch, and refer to Part I and your

intelligence for the details.

Definition 7.4 specified Kripke frames and augmented Kripke frames.

What takes the place of augmented Kripke frames is the following.

DEFINITION 7.18 (HENKIN/KRIPKE FRAME) Let (Q, R) be a Kripke

frame. 1t is a Henkin domain function in this frame if it is a func-

tion on the collection of types and:

1 'H(O) is some non-empty set.

2 'H((tl, ,tn)) ~ P('H(h) X X 'H(tn)); 'H((tl, ,tn)) =/= 0.

3 'H(jt) ~ ['H(t)]9; 'H(jt) =I= 0.

T is an interpretation if it maps each constant symbol of L( C) of type

t to a member of'H(t). Finally, M = (Q, R, 'H,T) is a Henkin/Kripke

frame for L(C).

If items 2 and 3 above hold with=, and not just ~ t h e Henkin/Kripke

model is standard. Standard models correspond exactly to the models

defined in Section 4.

DEFINITION 7.19 (ABSTRACTION DESIGNATION FUNCTION) Function

A is an abstraction designation function in the Henkin/Kripke frame

M = (Q, R, 1t, T), with respect to the language L(C), provided that for

each valuation v in M and for each predicate abstract (>.a1, ... , an.cl>)

of L(C) of type t, A(v, (>.a1, ... , an.cl>)) is some object of type t in M.

Term designation gets the obvious modification.

DEFINITION 7.20 (DESIGNATION OF A TERM)

Let M = (Q, R, 1t,T) be a Henkin/Kripke frame with A an abstraction

designation function in it. For each valuation v, define a mapping ( v *

T * r *A) assigning to each term a designation for that term, in the

context (possible world} r.

1 If A is a constant symbol of L(C) then (v * T * r * A)(A) =I( A).

2 If a is a variable then (v * T * r * A)(a) = v(a).

3 Ifr is a term of type jt then (v*T*r*A)(lr) = (v*T*r*A)(r)(r).

4 If(>.al. ,an.cl>) isapredicateabstractofL(C) oftypej(tl, ,tn),

then (v*T*r*A)((>.al, ,an.cl>)) =A(v,(>.al, ,an.cl>)).

104 TYPES, TABLEAUS, AND GODEL'S GOD

As usual, (v *I* r * A)(T, r) is written for (v *I* r * A)(T)(r). Now

truth, at a world, also has the expected characterization.

DEFINITION 7.21 (TRUTH OF A FoRMULA) Let M = (Q, R, 1l,I) be

a Henkin/ Kripke frame, let A be an abstraction designation function,

and let v be a valuation.

1 For an atomic formula T(TI, ... , Tn),

(a) lfT is of an intensional type, M,r lf-v T(TI, ... ,Tn) provided

((v*I*r*A)(TI), ... ,(v*I*r*A)(Tn)) E (v*I*r*A)(T,r).

{b) lfT is of an extensional type, M,r lf-v T(TI, ... ,Tn) provided

( ( v * I * r * A) ( TI)' .. . ' ( v * I * r * A) ( T n)) E ( v * I * r * A) ( T).

2 M, r lf-v,A <I> if it is not the case that M, r lf-v,A <I>.

3 M, r lf-v,A <I> 1\ w if M, r lf-v,A <I> and M, r lf-v,A w.

4 For a of type t, M, r lf-v,A (Va)<I> if M, r lf-v,A <I>[ a/ OJ for every

0 E 1l(t).

5 M, r lf-v,A D<I> if M, lf-v,A <I> for a l l ~ E g such that r n ~

Finally, the following should be no surprise.

DEFINITION 7.22 (HENKIN/KRIPKE MODEL)

(M, A) is a Henkin/Kripke model provided that, for each predicate ab-

stract (.Xa1, ... , an.<I>) of L(C) of type jt, A(v, (.Xa1, ... , an.<I>)) is the

function f given by the following:

The various theorems concerning uniqueness of an abstraction des-

ignation function, if one exists, and the good behavior of substitution

(Section 6) all carry over to the modal setting. I leave this to you.

The semantics just presented is extensional, in the sense of Part I. A

modal analog of generalized Henkin models can also be developed, along

the lines of Section 5. Objects in the Henkin domains are no longer sets,

and an explicit extension function must be added. The generalization is

straightforward but complex, and I also leave this to you.

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 8

MODAL TABLEAUS

1. The Rules

There are several varieties of tableaus for modal logic. This book uses

a version of prefixed tableaus. These incorporate a kind of naming device

for possible worlds into the tableau mechanism, and do so in such a way

that syntactic features of prefixes reflect semantic features of worlds.

Prefixed tableau systems exist for most standard modal logics. Here I

only give versions for K and 85 since these are the extreme cases. I

refer you to the literature for modifications appropriate for other modal

logics-see [FM98] for instance.

1.1 Prefixes

There are two versions of what are called prefixes. The version for K

is more complex, and variations on it also serve for many other modal

logics. The version for 85 is simplicity itself.

DEFINITION 8.1 (PREFIX) A K prefix is a finite sequence of positive

integers, written with periods as separators {1.2.1.1 is an example). An

85 prefix is a single positive integer.

Think of prefixes as naming worlds in some (unspecified) model. Prefix

structure is intended to embody information about accessibility between

worlds. For K, think of the prefixes 1.2.1.1, 1.2.1.2, 1.2.1.3, etc. as

naming worlds accessible from the world that 1.2.1 names. For 85 one

can take each world as being accessible from each world, so prefixes are

simpler. Prefixes have two uses in tableau proofs, qualifying formulas

and qualifying terms. I begin with terms.

105

106 TYPES, TABLEAUS, AND GODEL'S GOD

As was done classically, a larger language allowing parameters is used

for tableau proofs, with parameters for each type. But in addition,

an intensional term T is allowed to have a prefix. If we think of CJ as

designating a possible world, we should think of CJ T as representing the

extensional object that r designates at CT. Formally, if Tis of type jt, then

CJ T is of type t. But writing prefixes in front of terms makes formulas

even more unreadable than they already are. Instead, in an abuse of

language, I have chosen to write prefixes on terms as subscripts, Tr,-,

though of course the idea is the same, and I still often refer to them as

prefixes. So, if one thinks of CJ as designating possible world r, and r as

having the function f as its meaning, then Ta should be thought of as

designating the object f(r).

By L + (C) is meant L( C) enlarged with parameters, and allowing

prefixes (written as subscripts) on terms of intensional type (this includes

parameters, but prefixes will not be needed on free variables that are not

parameters). This extends the classical version of L + (C), since prefixes

are permitted now. But just as classically, in proving a closed formula

of L( C) it is formulas of L + (C) that will appear in proofs.

I said prefixes had two roles. Qualifying formulas is the main one.

DEFINITION 8.2 (PREFIXED FORMULA) A prefixed formula is an ex-

pression of the form CJ <I>, where CJ is a prefix and <I> is a formula of

L+(c).

Think of CJ <I> as saying that formula <I> is true at the world that CJ

names. Note that this use of prefixes does not compound, that is, CJ <I>

is a prefixed formula if <I> is a formula, and not something built up from

prefixed formulas.

DEFINITION 8.3 (GROUNDED) l call a term or a formula of L+(C)

grounded if it contains no free variables, though it may contain param-

eters.

As usual, tableau proofs are proofs of sentences--closed formulas-of

L(C). In the tableau, prefixed grounded formulas of L+(C) may appear.

To construct a tableau proof of <I>, begin with a tree that has 1 <I> at

its root, and nothing else. Think of 1 as an arbitrary world. This initial

tableau intuitively asserts that <I> is false at some world of some model,

the world designated by 1. Next the tree is expanded according to branch

extension rules to be given below. If we produce a tree that is closed,

which means it embodies a contradiction, we have a proof of <I>.

MODAL TABLEAUS 107

1.2 Propositional Rules

Since the modal tableau rules are rather complex, I've divided their

presentation into categories, beginning here with the propositional ones.

These are much as in the classical case, except that prefixes must be

"carried along." In these, and throughout, I use a, a', and the like to

stand for prefixes.

DEFINITION 8.4 (CONJUNCTIVE RULES) For any prefix a,

aXI\Y

aX

aY

a (X V Y)

aX

aY

a (X ::) Y)

aX

aY

aX =:Y

aX:=>Y

aY:=>X

DEFINITION 8.5 (DOUBLE NEGATION RULE) For any prefix a,

aX

aX

DEFINITION 8.6 (DISJUNCTIVE RULES) For any prefix a,

aXVY

aX laY

aX::)Y

a .x I a Y

a (X 1\ Y)

a X I a ,y

a (X = Y)

a (X ::) Y) I a (Y ::) X)

This completes the classical connective rules. The motivation should

be intuitively obvious. For instance, if X 1\ Y is true at a world named

by a, both X and Y are true there, and so a branch containing a X 1\ Y

can be extended with a X and a Y.

1.3 Modal Rules

Naturally the rules for modalities differ between the two logics we are

considering. It is here that the structure of prefixes plays a role. The

idea is, if OX is true at a world, X is true at some accessible world,

and we can introduce a name-prefix-for this world. The name should

be a new one, and the prefix structure should reflect the fact that it is

accessible from the world at which OX is true.

DEFINITION 8. 7 (POSSIBILITY RULES FOR K) If the prefix a.n is new

to the branch,

a OX

a.nX

aDX

a.n X

108 TYPES, TABLEAUS, AND GODEL'S GOD

DEFINITION 8.8 (POSSIBILITY RULES FOR 85) If the positive integer

n is new to the branch,

aOX aDX

nX nX

Notice that for both logics there is a newness condition. This implic-

itly treats 0 as a kind of existential quantifier. Correspondingly, the

following rules treat 0 as a version of the universal quantifier.

DEFINITION 8.9 (NECESSITY RULES FORK) If the prefix a.n already

occurs on the branch,

a OX

a.nX

aOX

a.nX

DEFINITION 8.10 (NECESSITY RULES FOR 85) For any positive inte-

ger n that already occurs on the branch,

a OX

nX

aOX

nX

Many examples of the application of these propositional and modal

rules can be found in [FM98J. I do not give any here. Rather, tableau

examples will be given after the full higher-type system has been intro-

duced.

1.4 Quantifier Rules

For the existential quantifier rules parameters must be introduced,

just as in the classical case. Thus proofs of sentences of L( C) are forced

to be in the larger language L + (C).

DEFINITION 8.11 (EXISTENTIAL RULES) In the following, pt is a pa-

rameter of type t that is new to the tableau branch.

a (:l<i)cJ>( oJ)

a cf>(pt)

a (Vo:t)cf>( o:t)

a -,cf>(pt)

Terms of the form lT may vary their denotation from world to world

of a model, because the extension of the intensional term T can change

from world to world. Such terms should not be used when instantiating

a universally quantified formula.

DEFINITION 8.12 (RELATIVIZED TERM) If T is a grounded intensional

term, lT is a relativized term.

MODAL TABLEAUS 109

DEFINITION 8.13 (UNIVERSAL RULES) In the following, 7t is any

grounded term of type t that is not relativized.

a ('v'at)<P(at)

a <I>(7t)

1.5 Abstraction Rules

a (3at)<I>( at)

a <l>( 7t)

The rules for predicate abstracts essentially correspond to Proposi-

tion 7.10. Note the presence of a subscript (prefix) on the predicate

abstract. We must know at what world the abstract is to be evaluated

before doing so. The next subsection provides machinery for the intro-

duction of these subscripts. Note that the subscript on the abstract, and

the prefix for the entire formula need not be the same.

DEFINITION 8.14 (ABSTRACT RULES) In the following, 71, ... ,7n are

non-relativized terms.

a' (>.a1, ... , an.<P( a1, ... , an)) u( 71, ... , 7n)

a<fl(71, ... ,7n)

a' (.Aa1, ... , an.<P( a1, ... , an)) u( 71, ... , 7n)

a --,<fl(71, ... , 7n)

1.6 Atomic Rules

Unlike classically, much can be done with atomic formulas in a modal

tableau besides just using them to close branches. The first atomic rule

says that, at a world, an intensional predicate applies to terms if those

terms are in the extension of the predicate at that world. It corresponds

to part 1a of Definition 7.9.

DEFINITION 8.15 (INTENSIONAL PREDICATION RULES) Let 7 be a

grounded intensional term, and 71, ... , 7n be arbitrary grounded terms.

a7(71, ... , 7n) a7(71, ,7n)

a (17)(71, ... , 7n) a(l7)(71, ,7n)

Relativized terms denote different objects in different worlds. In

tableaus, their behavior depends on the prefix of the formula in which

they appear. This leads us to the evaluation of relativized terms at pre-

fixes. Think of 7@a as 7 evaluated at a. On non-relativized terms, such

evaluation has no effect-their meaning is world independent.

DEFINITION 8.16 (EVALUATION AT A PREFIX) Let a be a prefix.

110 TYPES, TABLEAUS, AND GODEL'S GOD

1 For a relativized term lT, set (lT)@O" = Tu.

2 For a non-relativized term T, set T@O" = T.

The next rule covers the case of an extensional predicate applying to

terms. This corresponds to part lb of Definition 7.9.

DEFINITION 8.17 (EXTENSIONAL PREDICATION RULES) Let T be a

grounded extensional term, and Tl, ... , Tn be arbitrary grounded terms.

Here is a simple example of how these rules work. Suppose A is of

intensional type j(O) and b is of type 0. If O" A(b) occurs on a branch,

we may add O" (lA)(b) by an Intensional Predication Rule. Now the

Extensional Predication Rule applies; (lA)@O" = A,. and b@O" = b, so

we may add O" Au(b). Think of this as saying, since A(b) is true at the

world that O" designates, then b is in the extension of A at that world,

an extension represented by Au.

Finally, there are atomic formulas that must evaluate the same way

no matter what world is involved.

DEFINITION 8.18 (WORLD INDEPENDENT) We call an atomic formula

T(T1, ... ,Tn) world independent if none ofT, Tl, ... , Tn is relativized,

and T is of extensional type.

DEFINITION 8.19 (WORLD SHIFT RULES) Let T(Tl, ... , Tn) be world

independent. If 0"

1

already occurs on the branch,

O"T(Tl,. ,Tn)

0"

1

T( Tt, ... , Tn)

0"-,T(Tl, ,Tn)

0"

1

-,T( Tl, ... , Tn)

1. 7 Proofs and Derivations

I'll begin with the easy part.

DEFINITION 8.20 (CLOSURE) A tableau branch is closed if it contains

O" w and O" -.w, for some formula W of L+(C). A tableau is closed if each

branch is closed.

DEFINITION 8.21 (TABLEAU PROOF) For a sentence <P of L(C), a

closed tableau beginning with 1 -.<P is a proof of <P.

A brief discussion of the complexities of modal consequence is in Sec-

tion 6. More discussion can be found in [Fit83, Fit93, FM98]. Corre-

sponding to the local/global semantic distinction of Definition 7.12, we

have the following tableau version.

MODAL TABLEAUS 111

DEFINITION 8.22 (LOCAL AND GLOBAL ASSUMPTIONS) Let 8 and U

be sets of sentences of L( C). A tableau uses S as global assumptions and

U as local assumptions if the following two tableau rules are admitted.

Local Assumption Rule If Y is any member of U then 1 Y can be

added to the end of any open branch.

Global Assumption Rule If Y is any member of S then a Y can be

added to the end of any open branch on which a appears as a prefix.

DEFINITION 8.23 (TABLEAU DERIVATION) A sentence <I> has a deriva-

tion from global assumptions S and local assumptions U if there is a

closed tableau beginning with 1 --,<]>, allowing the use of U and S as local

and global assumptions respectively.

This concludes the presentation of the basic tableau rules. It is a

rather complex system. In Section 2 I give a few examples of proofs us-

ing the rules. I omit soundness and completeness proofs. The arguments

are elaborations of those given earlier for classical logic. Complexity of

presentation goes up, but no fundamentally new ideas arise. Conse-

quently they are left as a huge exercise.

There is one important consequence of the completeness proofs that

we will need, however, and that is the fact that the system has the cut-

elimination property-see Theorem 4.32. Just as in the classical case

(Corollary 4.34), it is a consequence of this that any previously proved

result can simply be introduced into a tableau.

2. Tableau Examples

Tableaus for classical logic are well-known, and even for propositional

modal logics they are rather familiar. The abstraction and predication

rules of the previous section are new, and I give two examples illustrat-

ing their uses. The examples use the K rules; I do not give examples

specifically for S5 here.

EXAMPLE 8.24 This provides a proof for (7.3) which was verified valid

in Example 7.13. The formula is

(>.X.()(:Jx)X(x))(P) :J ()(>.X.(:lx)X(x))(P)

in which x is a variable of type 0 and X is a variable and P a constant

symbol, both of type j(O).

112 TYPES, TABLEAUS, AND GO DEL'S GOD

1 ..,[(.XX.O(::Jx)X(x))(P) :J 0(-XX.(::Jx)X(x))(P)] 1.

1 (.XX.O(::Jx)X(x))(P) 2.

1 --,0(-XX.(::Ix)X(x))(P). 3.

1 l(.XX.O(::Jx)X(x))(P) 4.

1 (.XX.O(::Jx)X(x))I(P) 5.

1 O(::Jx)P(x) 6.

1.1 (::lx)P(x) 7.

1.1--,(.XX.(::Jx)X(x))(P) 8.

1.1--, l(.XX.(::Ix)X(x))(P) 9.

1.1--,(.XX.(::Jx)X(x))u(P) 10.

1.1--,(::lx)P(x) 11.

In this, 2 and 3 are from 1 by a conjunctive rule; 4 is from 2 by inten-

sional predication; 5 is from 4 by extensional predication; 6 is from 5 by

predicate abstraction; 7 is from 6 by a possibility rule; 8 is from 3 by a

necessitation rule; 9 is from 8 by intensional predication; 10 is from 9 by

extensional predication; and 11 is from 10 by predicate abstraction.

It should be obvious that useful derived rules could be introduced.

For instance, the passage from 2 to 4 to 5 to 6 could be collapsed. Such

rules are given in the next section.

EXAMPLE 8.25 Here is a proof of (7.13), which was shown to be valid

earlier.

{.XX.O(::Jx)X(x))(jP) :J (.XX.(::Ix)X(x))(jP)

See Example 7.14 for a discussion of the significance of this formula.

1 --,[(.XX.O(::Ix)X(x))(jP) :J (.XX.(::Jx)X(x))(jP)] 1.

1 (.XX.O(::Jx)X(x))(jP) 2.

1 --,(.XX.(:3x)X(x))(1P) 3.

1 1(-XX.O(::Ix)X(x))(jP) 4.

1 --, 1(-XX.(::Jx)X(x))(lP) 5.

1 (.XX.O(::Jx)X(x))I(Pl) 6.

1 --,(.XX.(::Ix)X(x))I(PI) 7.

1 O(::Jx)P1(x) 8.

1 --,(::Jx)PI(x) 9.

1.1 (:3x)P

1

(x) 10.

1.1 P1(p) 11.

1 P1(p) 12.

1 --,pl (p) 13.

In this, 2 and 3 are from 1 by a conjunction rule; 4 is from 2 and 5 is from

3 by intensional predication; 6 is from 4 and 7 is from 5 by extensional

MODAL TABLEAUS 113

predication; 8 is from 6 and 9 is from 7 by predicate abstraction; 10 is

from 8 by a possibility rule; 11 is from 10 by an existential rule; 12 is

from 11 by a world shift rule; and 13 is from 9 by a universal rule.

Exercises

EXERCISE 2.1 Give a tableau proof of the following

(>.X.O(:lx)X(x))(tP) :J (>.X.(:lx)X(x))(lP)

where xis of type jO, X is of type (jO) and Pis of type j(jO).

EXERCISE 2.2 Give a tableau proof of the following

O(:lx)P(x) :J (:lX)OP(lX)

where x is of type 0, X is of type jO and Pis of type j(O).

3. A Few Derived Rules

The tableau examples in the previous section are short, but already

quite complicated to read. In the interests of keeping things relatively

simple, a few derived rules are introduced which serve to abbreviate

routine steps.

DEFINITION 8.26 (DERIVED CLOSURE RULE) Suppose X is a world

independent atomic formula. A branch closes if it contains a X and

a' -,x.

The justification for this is easy. Using the World Shift Rule, if a X

is on a branch, we can add a' X, and then the branch closes according

to the official closure rule.

The official rule concerning intensional predication has a slightly more

efficient version, in which we first apply intensional, then extensional

predication rules.

DEFINITION 8.27 (DERIVED INTENSIONAL PREDICATION RULE) Let

T be a grounded intensional term, and T

1

, ... , Tn be arbitrary grounded

terms.

Also here are two derived rules for predicate abstracts, one in which

the abstract has a prefix (subscript), one in which it does not.

114 TYPES, TABLEAUS, AND GODEL'S GOD

DEFINITION 8.28 (DERIVED SUBSCRIPTED ABSTRACT RULE)

In the following, 71, ... , 7n are arbitrary grounded terms.

a' (.Xa1, ... , an.<P(a1, ... , an))u(71, ... , 7n)

a<P(71@a', ... ,7n@a')

a'(Aa1, ,an.<P(a1, ,an))u(71, ,7n)

a -.<P(71 @a', ... , 7n@a')

This abbreviates the application of the extensional predication rule, fol-

lowed by predicate abstraction.

DEFINITION 8.29 (DERIVED UNSUBSCRIPTED ABSTRACT RULE)

In the following, 71, ... , 7n are grounded terms.

a(.Aat, ... ,an.<P(a1, ,an))(7t, ... ,7n)

a <P(71 @a, ... , 7n@a)

a (.Xa1, ... , an.<P(a1, ... , an))( 71, ... , 7n)

IT iP( 71 @a, ... , 7n@a)

This rule abbreviates successive applications of intensional predication,

extensional predication, and predicate abstraction.

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 9

MISCELLANEOUS MATTERS

This chapter is something of a grab-bag. Some familiar topics, like

equality, and some less familiar, like choice functions, are discussed.

1. Equality

The tableau rules of the previous chapter do not mention equality or

extensionality. These are treated exactly as in the classical setting, via

axioms, though as we will see, extensionality requires some care.

1.1 Equality Axioms

If we want to take equality into account, we use the Equality Axioms,

Definition 5.1, as global assumptions. From here on these will be assumed

in this book.

In Chapter 5 I presented some tableau rules that were derivable classi-

cally provided equality axioms were allowed. In the modal setting these

rules (with prefixes added, of course) are also derived rules. They are

stated again for reference.

Reflexivity Rule For a grounded, non-relativized term 7, and a prefix

a that is already present on the branch,

Substitutivity Rule For grounded, non-relativized terms 71 and 72,

a <P(Tl)

a (71 = 72)

a IP(72)

115

116 TYPES, TABLEAUS, AND GODEL'S GOD

Here is an example that uses equality. To help understand what the

example says, and see why it ought to be valid, I give an informal inter-

pretation for it.

Suppose we read modal operators temporally, so that OX means X

will be the case no matter what the future brings, and OX means the

future could turn out to be one in which X is true. Let p be a type

jO constant symbol intended to be read, "the President of the United

States." Thus pis an individual concept, and designates different people

in different possible futures.

Now, call a person Presidential material if the person could be Pres-

ident (say the person meets all the legal requirements, such as being at

least 35, not having already served twice, and so on). Being Presidential

material is a property of persons. If we assume we have a model whose

domain is the population of the United States, being Presidential ma-

terial is a type j (0) object and is expressed by the following abstract,

where xis of type 0.

(>.x.O(lp = x))

Informally, this predicate applies to a person at a particular time if

there is some possible future in which that person is the President of the

United States.

Next, call a property of persons statesmanlike if it will always ap-

ply to the President. Thus we are using statesmanlike as a property

of properties of persons-being diplomatic is hopefully a statesmanlike

property, for instance. As such, being statesmanlike is of type j((O)).

It is expressed by the following abstract, where X is of type (0), and

applies to those properties that will always belong to the President, no

matter who that will be.

(>.X.DX(lp))

Now, the extension of the property of being Presidential material is a

statesmanlike property since, no matter who turns out to be President,

that person must have been of Presidential material. The following gives

a tableau verification for this.

EXAMPLE 9.1 Here is a proof of the formula:

(>.X.DX(lp))(t(>.x.O(lp = x)))

MISCELLANEOUS MATTERS 117

1 (.AX.DX(lp))(l(.Ax.O(lp = x))) 1.

1 D(.Ax.O(lp = x))I(lp) 2.

1.1(.Ax.O(lP = :t))I(lP) 3.

1.1(Ax.O(lp = x))I(pu) 4.

1 O(lP=pu) 5.

1.1(1P = Pu) 6.

1.1(Pu = Pu) 7.

1.1 (pu = Pu) 8.

In this 2 is from 1 by the derived unsubscripted abstraction rule; 3 is

from 2 by a possibility rule; 4 is from 3 by extensional predication; 5 is

from 4 by a predicate abstract rule; 6 is from 5 by a necessity rule; 7 is

from 6 by extensional predication; 8 is by the derived reflexivity rule.

1.2 Extensionality

Extensionality can, of course, be imposed by assuming the Extension-

ality Axioms of Chapter 6, Definition 6.1, as global assumptions. The

trouble is, doing so for intensional terms yields undesirable results, as

the following shows.

PROPOSITION 9.2 Assume the Extensionality Axioms apply to inten-

sional terms. If a and f3 are of intensional type j(t), then the following

is valid.

(Va)(V,B)[(la =l/3) ~ (a= ,B)]

The proof of this is left to you. It is almost immediate, using the

Intensional Predication Rules. The problem with this result is, it tells

us that if two intensional objects happen to coincide in extension at

some world, then they are identical and hence coincide at every world.

Clearly this is undesirable, so extensionality for intensional terms is not

assumed.

If two intensional objects agree in extension at every possible world of

a model they are, in fact, the same. Saying this requires a quantification

over possible worlds, which we cannot do. The following is as close as

we can come.

DEFINITION 9.3 (EXTENSIONALITY FOR INTENSIONAL TERMS) For a

and f3 of the same intensional type,

(Va)(V,B)[D(la =l/3) ~ a = ,B)]

I will assume this at some points, but I will be explicit when. For

extensional terms, the extensionality axioms pose no difficulty and will

always be assumed. Let me make this official.

118 TYPES, TABLEAUS, AND GODEL'S GOD

Extensionality Assumptions From now on, the extensionality ax-

ioms will be assumed for extensional terms as global assumptions. For

intensional terms extensionality, Definition 9.3, will only be assumed if

explicitly stated.

I restate the extensionality axioms here for convenience.

DEFINITION 9.4 (EXTENSIONALITY FOR EXTENSIONAL TERMS)

Each sentence of the following form is an extensionality axiom, where

a and {3 are of type (t1, ... , tn), 11 is of type t1, ... , In is of type tn.

('v'a)('v'{3){('v'11) ('v'ln)[a(/'1, . , In)= {3(/'1, ... , In)] :J [a= {3]}

In Chapter 6 a derived tableau rule for extensionality was given, as-

suming the extensionality axioms. Once again, it is still a derived rule

for modal tableaus. Here is a statement of it.

Extensionality Rule For grounded, non-relativized extensional terms

T1 and T2, and for parameters P1, ... , Pn that are new to the branch,

(}--, [T1(P1, ,pn) = T2(P1, ,Pn)JI (} (T1 = T2)

2. De Re and De Dicto

Loosely speaking, asserting the necessary truth of a sentence is a de

dicta usage of necessity; for example, "it is necessary that the President

of the United States is a citizen of the United States." This asserts the

necessary truth of the sentence, "the President of the United States is

a citizen of the United States." For this to be the case, it must be so

under all circumstances, no matter who is President, and since being a

citizen of the United States is a requirement for the Presidency, this is

the case. Ascribing to an object a necessary property is a de re usage; for

example, "it is a necessary truth, of the President of the United States,

that he is at least 50 years old." This asserts, of the President, that he

is and always will be at least 50 years old. Since the President, at the

time of writing, is Bill Clinton, and he is at the moment 53 years old

and will never be younger than this, this assertion is correct. But since

the Constitution of the United States only requires that a President be

at least 35, the assertion may not be true in the future, for a different

President. If an object is identified using an intensional term, it makes a

serious difference whether that term is used in a de dicta or a de re con-

text, as the examples involving the Presidency illustrate. In this section

the formal relationship between the two notions is explored. As will be

seen over the next several sections, this also relates to other interesting

concepts that have been part of historic philosophical discourse.

MISCELLANEOUS MATTERS 119

In the next few paragraphs, f3 is of some extensional type t, and Tis

of the corresponding intensional type jt.

Consider the expression (.Xf3.D<l>(f3))(1 T), where <P(/3) is some for-

mula with only f3 free (for simplicity). Say the expression is true at

a world of a modal model. (I use a generalized Henkin/Kripke model,

but what is said applies to any version--extensional, standard-just as

well.) Thus suppose M, f lf-v,A (.Xf3.D<l>(f3))(lT). Let Or be the object

that T designates at r, that is, (v*I*r*A)(T,f) =Or. Then we have

M, r lf-v,A D<l>(/3)[/3/0r]. So at every alternative world, .6., we have

M, .6. lf-v,A <P(/3)[!3/0r], that is <P(/3) is true, at .6., of the object that T

designates at r. The effect is that (.Xf3.D<l>(f3))(1T) asserts, of the object

designated by T at r that it has a necessary property. This is a de re

use of necessity-ascribing a necessary property to a thing.

Next consider the expression D(.Xf3.<l>(f3))(1 T). This asserts the ne-

cessity of a sentence. It is a de dicta use of necessity-applying it

to a sentence, a dictum. And in general the behavior is quite dif-

ferent from the de re version. If M,r lf-v,A D(.Xf3.<l>(f3))(1 T), then

at each alternative world .6. we have M, .6. lf-v,A (.Xf3.<l>(f3))(1 T), and

so M, .6. lf-v,A <P(f3)[w/O.!l], where oil is the designation of T at .6.,

something that depends on .6.. We can thus think of the assertion

D(.X/3.<1>(/3))(1 T) as being concerned with the sense ofT and not just

with the object it happens to denote in "our" world-we use the local

desigmrtion of T, which can vary from world to world.

One remarkable thing about de re and de dicta is that, if either hap-

pens to imply the other, for a particular term, then the two turn out to

be equivalent for that term. The following makes this precise. In the

next section the phenomena is linked to the notion of rigidity.

DEFINITION 9.5 (De Re/ De Dicta) LetT be a term of intensional type

jt, f3 be a variable of type t, and o: be a variable of type j(t). In a model:

1 de re is equivalent to de dicto for T if the following is valid.

(Vo:)[(.Xf3.Do:(f3))(1T) = D(.Xf3.o:(f3))(1T)]

2 de re implies de dicto for T if the following is valid.

(Vo:)[(.Xf3.Do:(f3))(1T) ::) D(.Xf3.o:(f3))(1T)]

3 de dicto implies de re for T if the following is valid.

(Vo:)[D(.A,B.o:(,B))(lT) :J (.A,B.Do:(,B))(tr)]

The formulas above are allowed to be open--free variables may be

present. Equivalently, one can work with universal closures. In [FM98]

120 TYPES, TABLEAUS, AND GODEL'S GOD

we used schemas instead of the formulas given above, because that was

a first-order treatment and we did not have the higher-type quantifier

(Vo:) available. The interesting fact about the three notions above is:

they all say the same thing.

PROPOSITION 9.6 For any intensional term T, the following are equiv-

alent (in K).

1 de dicto is equivalent to de re for T

2 de dicto implies de re for T

3 de re implies de dicto for T

Proof Obviously item 1 implies items 2 and 3. I give a tableau proof,

in K, showing that item 2 implies item 3. A similar argument, which

I leave to you, shows that item 3 implies item 2, and this is enough to

complete the proof of the Proposition. To keep things simple, assume T

has no free variables. Here is a closed tableau for (Vo:)[(.A,B.Do:(,6))(l

T) => D(.A,B.o:(,B))(lT)], (negation of) de re implies de dicto. In it, at a

certain point, use is made of an instance of the de dicto implies de re

schema. The tableau begins as follows.

1 (Vo:) [(.A,B.Do:(,B))(lT) => O(.A,B.o:(,B))(lT)] 1.

1 ..., [(.A,B.D<P(,B))(lT) => D(.A,B.<I>(,B))(lT)j 2.

1 (_A,6.0<l>(,6))(1T) 3.

1 D(.A,B.<l>(,B))(lT) 4.

1 D<I>( Tl) 5.

1.1(A,6.<J?(,6))(1T) 6.

1.1<l>(Tu) 7.

1.1 <f>(Tl) 8.

1 (Vo:)[O(.A,B.o:(,B))(lT) =:> (.A,B.Do:(,B))(lT)] 9.

1 D(.A,B.(AJ'.<l>(J') => <P(lT))(,B))(lT)

(.A,B.D(.A')'.<P(/') =:> <P(lT))(,6))(lT) 10.

Item 2 is from 1 by an existential rule, using <P as a new parameter of

type j(t); items 3 and 4 are from 2 by a conjunctive rule; 5 is from 3

by an unsubscripted abstract rule; 6 is from 4 by a possibility rule; 7 is

from 6 by an unsubscripted abstract rule; 8 is from 5 by a necessity rule.

Item 9 is a de dicto implies de re formula; and item 10 is from 9 by a

universal rule, using (.A')'.<P(/') =:> <P(lT)) to instantiate the quantifier.

Using item 10, the tableau splits into two branches. I first present the

left one, and afterwards the right.

MISCELLANEOUS MATTERS

1 D(.A,6.(A'Y.cf>(r) :) cf>(lT))(,6))(lT) 11.

1.2 (A,6.(A/.cf>(r) :) cf>(lT))(,6))(1T) 12.

1.2 (A/.cf>(r) :) cf>(lT))(Tl.2) 13.

1.2 [cf>(TL2) :) cf>(lT)) 14.

1.2 cf>( TL2) 15.

1.2 --,cf>(lT) 16.

1.2 cf>1.2(T1.2) 17.

1.2 --,cf>L2( 71.2) 18.

121

Item 11 is from 10 by a disjunctive rule (recall, this is the left branch);

12 is from 11 by a possibility rule; 13 is from 12 and 14 is from 13 by

an unsubscripted abstract rule; 15 and 16 are from 14 by a conjunctive

rule; 17 is from 15 and 18 is from 16 by a derived intensional predication

rule. The branch is closed because of 17 and 18.

Now I show the right branch, below item 10.

1 (.A,6.0(A/.cf>(r) :) cf>(lT))(,6))(1T) 19.

1 D(.A,.cf>('y) :) cf>(lT))(Tl) 20.

1.1 (A/.cf>(r) :) cf>(lT))(Tl) 21.

1.1 cf>(Tl) :) cf>(lT) 22.

/

1.1-.cf>(Tl) 23. 1.1 cf>(lT) 24.

1.1 cf>u ( Tu) 25.

1.1 cf>u ( Tu) 26.

In this part, 19 is from 10 by a disjunctive rule; 20 is from 19 by an

unsubscripted abstract rule; 21 is from 20 by a necessity rule; 22 is

from 21 by an unsubscripted abstract rule; 23 and 24 are from 22 by a

disjunctive rule; 25 is from 24 and 26 is from 7 by a derived intensional

predication rule. Closure is by 8 and 23, and by 25 and 26.

Exercises

EXERCISE 2.1 Give the tableau proof needed to complete the argument

for Proposition 9.6.

3. Rigidity

In [Kri80] the philosophical ramifications of the notion of rigidity are

discussed at some length, with a key claim being that names are rigid.

The setting is first-order modal logic, treated informally. A term is taken

to be rigid if it designates the same thing in all possible worlds. In [FM98]

122 TYPES, TABLEAUS, AND GO DEL'S GOD

we modified this notion somewhat so that a formal investigation could

more readily be carried out-we called a term rigid if it designated the

same thing in any two possible worlds that were related by accessibility.

The idea is that the behavior of a term in an unrelated world should

have no "visible" effect. It is this modified notion of rigidity that is used

here, and it will be seen that it can be expressed directly if equality is

available. (Whether models are standard, Henkin, or generalized Henkin

does not matter for what we are about to do, only that they are normal.)

For the rest of this section, normality is assumed.

DEFINITION 9. 7 The intensional term T is rigid in a normal model if

the following is valid in it.

It is easy to see that the formula asserting rigidity of T is true at a

world r of a normal model if and only if, at each world accessible from

r, T designates the same object that it designates at r itself. Thus

asserting validity for the rigidity formula indeed captures the notion of

rigidity for terms that we have in mind.

If an intensional term is rigid, it does not matter in which possible

world we determine its designation. But then, if both necessitation and

designation by a rigid intensional term are involved in the same formula,

it should not matter whether we determine what the term designates

before or after we move to alternative worlds when taking necessitation

into account. In other words, for rigid intensional terms the de re/ de

dicto distinction should vanish. In fact it does, and as it happens, the

converse is also the case. The following is a higher order version of a

first order argument from [FM98].

PROPOSITION 9.8 In K, the intensional term T is rigid if and only if

the de re/de dicto distinction vanishes, that is, if and only if any (and

hence all) parts of Proposition 9. 6 hold.

Proof This is shown by proving two implications, using tableau rules

for K including rules for equality.

Let A be the formula (A,6.0(,8 =lT))(lT) and let B be the formula

('v'a)[D(A,6.a(,B))(T) :J (A,6.0a(,6))(T)]. A says T is rigid, while B says

de dicto implies de re for T. I first give a tableau proof of A :J B.

MISCELLANEOUS MATTERS

1 (A :J B) 1.

1 (,\,6.0(,6 =17))(17) 2.

1 (Va)[0(,\,6.a(,6))(17) :J (>.,6.0a(,6)}(17)] 3.

1 [0(>.,6.<1>(,6))(17) :J (>.,6.0<1>({3))(17)] 4.

1 0(,\,6.<1>(,6))(17) 5.

1 (A,6.0<I>(J3)}(17) 6.

1 D<I>( 71) 7.

1.1<1>(71) 8.

1.1 (,\,6.<1>({3))(17) 9.

1.1 <l>( 71.1) 10.

1 0(71 =17) 11.

1.1 (71 =17) 12.

1.1 71 = 71.1 13.

1.1 <l> ( 71.1) 14.

123

In this tableau, 2 and 3 are from 1 by a conjunctive rule; 4 is from 3

by an existential rule, with <I> as a new (intensional) parameter; 5 and 6

are from 4 by a conjunctive rule; 7 is from 6 by a derived unsubscripted

abstract rule; 8 is from 7 by a possibility rule; 9 is from 5 by a necessity

rule; 10 is from 9 and 11 is from 2 by a derived unsubscripted abstract

rule; 12 is from 11 by a necessity rule; 13 is from 12 by a derived unsub-

scripted abstract rule; and 14 is from 8 and 13 by a derived substitutivity

rule for equality.

Finally I give a tableau proof of B :=:> A.

1 (B :J A) 1.

1 (Va)[D(>.,6.a(,6)}(17) :J (>.,6.0a(,6))(l7)] 2.

1 (A,6.0(,6 =17)}(17) 3.

1 0(,\,6.(Af'. 17 = 1)(,6)}(17) :J (,\,6.0(Af'. 17 = /)(,6)}(17) 4.

1 0(71 =17) 5.

1.1(71 =17) 6.

1.1(71 = 71.1) 7.

/

1 0(>.,6.(Af'. 17 = /)(,6))(17)

8.

1.2 (A,6.(Af'. 17 = 1)(,6))(17)

1.2(Af'.l7 = 1)(71.2) 10.

1.2 (17 = 71.2) 11.

1.2 ( 71.2 = 71.2) 12.

1.2 71.2 = 71.2 13.

1 (,\,6.0(Af'. 17 = 1)(,6)}(17)

14.

9. 1 D(>.,. 17 = ,)( 71) 15.

1.1 (Af'.l7=f'}(71) 16.

1.1 (17=71) 17.

1.1 71.1 = 71 18.

1.1(71 = 71) 19.

1.1 71 = 71 20.

124 TYPES, TABLEAUS, AND GODEL'S GOD

In this, 2 and 3 are from 1 by a conjunctive rule; 4 is from 2 by a

universal rule, instantiating with the term (>."f. l T = 'Y); 5 is from 3

by an unsubscripted abstract rule; 6 is from 5 by a possibility rule; 7

is from 6 by an unsubscripted abstract rule; 8 and 14 are from 4 by

a disjunctive rule; 9 is from 8 by a possibility rule; 10 is from 9, and

11 is from 10 by an unsubscripted abstract rule; 12 is from 11 by an

extensional predication rule; 13 is by reflexivity; 15 is from 14 by an

unsubscripted abstract rule; 16 is from 15 by a necessity rule; 17 is from

16 by an unsubscripted abstract rule; 18 is from 17 by an extensional

predication rule; 19 is from 7 and 18 by substitutivity; and 20 is by

reflexivity.

4. Stability Conditions

In his ontological argument Godel makes essential use of what he

called "positiveness," which is a property of properties of things. He-

does not define the notion, instead he makes various axiomatic assump-

tions concerning it. Among these are: if a property is positive, it is

necessarily so; and if a property is not positive, it is necessarily not pos-

itive. (His justification for these was the cryptic remark, "because it

follows from the nature of the property.") Suppose we use the second-

order constant symbol P to represent positiveness, and take it to be

of type j (i (0)). Godel stated his conditions more or less as follows,

with quantifiers implied: P(X) ::J DP(X) and P(X) ::J DP(X). The

second of these is equivalent to ()P(X) ::J P(X), and this form will be

used in what follows. Positiveness is a second-order notion, but Godel's

conditions can be extended to other orders as well. I call the resulting

notion stability, which is not terminology that Godel used.

DEFINITION 9.9 (STABILITY) LetT be a term of type j(t). T satisfies

the stability conditions in a model provided the following are valid in

that model.

(Va)[T(a) :::> DT(a)]

(Va)[()T(a) ::J T(a)]

The stability conditions come in pairs. In S5, however, these pairs

collapse.

PROPOSITION 9.10 In S5, (\la)[T(a) ::J DT(a)] and (\la)[()T(a) :::> T(a)]

are equivalent.

Proof Suppose (Va)[T(a) ::J DT(a)]. Contraposition gives (Va)[DT(a)

::J T(a)]. From necessitation and converse Barcan, (Va)D[DT(a) :::>

MISCELLANEOUS MATTERS 125

T(a)], and so ('v'a)[DDT(a) :J DT(a)], equivalently, ('v'a)[DOT(a)

:J DT(a)]. But in S5, X :J DOX is valid, hence we have ('v'a)[T(a) :J

DT( a)]. By contra position again, ('v' a)[ DT( a) :J T( a)], and hence

('v'a)[OT(a) :J T(a)]. The converse direction is similar.

In the stability conditions, T is being predicated of other things. On

the other hand, to say T is rigid, or that the de re /de dicta distinction

vanishes for T, involves other things being predicated of T. Here is the

fundamental connection between stability and earlier items.

THEOREM 9.11 An intensional term T is rigid if and only if it satisfies

the stability conditions.

Proof This is most easily established using tableaus. And it is a good

workout. I leave it to you to supply the details.

Exercises

EXERCISE 4.1 Complete the proof of Theorem 9.11 by giving appropri-

ate closed tableaus. Recall that extensionality is assumed for extensional

terms, and we have the derived extensionality rule given in Definition 6.2.

5. Definite Descriptions

As is well-known, Russell treated definite descriptions by translat-

ing them away, [Rus05]. His familiar example, "The King of France is

bald," is handled by eliminating the definite description, "the King of

France," in context, to produce the sentence "exactly one thing Kings

France, and that thing is bald." It is also possible to treat definite

descriptions as first-class terms, making them a primitive part of the

language. In [FM98] we showed how both of these approaches extend to

first-order modal logic. Further extending this dual treatment to higher-

order modal logic adds greatly to the complexity, so I confine things to

a Russell-style version here.

Suppose we have a formula <P, and we form the expression m.<P, which

is read as the a such that <P, and is called a definite description. Syntac-

tically it is treated like a term. Its free variables are those of <P, except

for a, and its type is the type of a. In a more formal presentation, all

this would have been built into the definition of term and formula given

earlier, but doing so adds much complexity at the start of the subject,

so I am taking the easier route of explaining now what could have been

done.

DEFINITION 9.12 (DESCRIPTION DESIGNATION) The definite descrip-

tion 1a.<P designates, or is defined at the possible world r of M =

126 TYPES, TABLEAUS, AND GODEL'S GOD

(Q, R, 1-l,I) if

M, r 11-v (3,6)('v'8)[(>.a.<P)(8) = (,6 = 8)]

where ,6 and 8 are not free in <P.

Next, the behavior of definite descriptions in context is treated in

Russell's style. As he so famously noted, scope issues are fundamen-

tal. There is a difference between "the King of France is non-bald,"

which is false since there is no King of France, and "it is not the

case that the King of France is bald," which is true. Formally, it is

the difference between (>.x.-.B(x))( 1y.K(y)) and (>.x.B(x)) ( 1y.K(y) ).

There is a similar distinction to be made between (>.x.DB(x))(1y.K(y))

and D(>.x.B(x))(1y.K(y)) since definite descriptions generally act non-

rigidly, and so the de rej de dicta issue arises.

Note that in all the examples above, scope of a definite description was

indicated by the use of a predicate abstract. Now (>.x.DB(x))(1y.K(y))

is atomic, as are (>.x.B(x))(1y.K(y)) and (>.x.-.B(x))(1y.K(y)). It is

enough for us to specify how definite descriptions behave in atomic con-

texts, and everything else follows automatically. But even at the atomic

level, a definite description can occur in a variety of ways. For instance,

in To(TI) either, or both, of To and TI could be descriptions. There are

several ways of dealing with this, all of which lead to equivalent results.

I'll use a Russell-style translation directly in the simplest case, and re-

duce other situations to that.

DEFINITION 9.13 (DESCRIPTIONS IN ATOMIC CONTEXT) Let m.<P be

a definite description, and let ,6 and 8 be variables of the same type as

a, that do not occur free in <P or in any of the terms Ti below.

1 To ( m. <P) is an abbreviation for

(3,6){('v'8)[(>.a.<P)(8) = (,6 = 8)] 1\ To(,6)}.

2 To(TI. ... , m.<P, ... , Tn) is an abbreviation for

(>.,6.To(TI, ... , ,6, ... , Tn))(7a.<fl).

3 (7a.<P)(TI, ,Tn) is an abbreviation/or

(>.,6.,6( TI, ... , Tn)) ( ?a.<fl).

4 To(TI, ... , l(m.<P), ... , Tn) is an abbreviation for

MISCELLANEOUS MATTERS 127

5 l( m. <I>) ( T1, . . . , T n) is an abbreviation for

(.\,8.(1,6) ( Tl, ... , Tn) )( m.<I>).

The definition above provides a routine for the elimination of defi-

nite descriptions. The problem is, there may be more than one way

of following the routine. For instance, consider the atomic formula

(1x.A(x))(1y.B(y)), which contains two definite descriptions. If we elim-

inate (1y.B(y)) first, beginning with an application of part 1 of the def-

inition, and then eliminate (1x.A(x)), we wind up with the following.

(:Jz1){('v'z2)[(.\y.B(y))(z2) = (z1 = z2)]/\

(3z4){('v'zs)[(.\x.A(x))(zs) = (z4 = zs)]/\ (.\z3.Z3(zl))(z4)}}

(9.1)

On the other hand, we might choose to eliminate 1x.A(x) first, beginning

with part 3 of the definition. If so, after a few steps we wind up with

the following.

(:Jz2){('v'z3)[(.\x.A(x))(z3) = (z2 = z3)]/\

(.\zl-(3z4){('v'zs)[(.\y.B(y))(zs) = (z4 = zs)]/\ z1(z4)})(z2)}

(9.2)

Fortunately, (9.1) and (9.2) are equivalent. In general, the elimina-

tion procedure is confluent-different reduction sequences for the same

atomic formula always lead to equivalent results.

In a sense there are two kinds of definite descriptions, intensional and

extensional, depending on the type of the variable a in 10:. <I>. Extensional

definite descriptions are rather well-behaved, and I say little about them,

but for intensional ones, some interesting issues can be raised. In Def-

inition 9. 7 I characterized a formal notion of rigidity. That definition

can be extended to definite descriptions: call m.<I> rigid at a world if the

following is true at that world.

(.\,B.D(,B =1( 10:. <I>))) (l( m.<I>)).

Informally speaking, to say this is true at a world r amounts to saying:

m.<I> designates at world r, m.<I> designates at all worlds accessible from

r, and at rand every world accessible from it, 1o:.<I> designates the same

thing. The following Proposition is an alternative characterization.

PROPOSITION 9.14 The formula (.\,8.0(,8 =1(m.<I>)))(1(1o:.<I>)) is equiv-

alent in K to the conjunction of the following three formulas.

1 (:3,8) ('v'8) [(.\a. <I>) ( 8) = (,B = 8)]

128 TYPES, TABLEAUS, AND GODEL'S GOD

2 (\1,6)[(-\a.<P)(,B) :J 0(-\a.<P)(,B)]

3 (\1,6)[0(-\a.<P)(,B) :J (-\a.<P)(,B)] ..

In other words, this Proposition says ( 1a. <P) is rigid if and only if

( m.<P) designates and (Aa.<P) satisfies the stability conditions.

Exercises

EXERCISE 5.1 Show the equivalence of (9.1) and (9.2). (For this classi-

cal tableaus can be used, since modal operators do not explicitly appear.)

EXERCISE 5.2 Use K tableaus to prove Proposition 9.14. (This is a

long exercise.)

6. Choice Functions

In a Henkin/Kripke model, not all the objects of a standard model

need be present. We would like some mechanism to ensure that many

are, so non-standard models may have a sufficiently rich universe. Ab-

straction provides one way of doing this. If <P is a formula, there must

be an intensional object in a Henkin/Kripke model to serve as the des-

ignation for (-\a.<P), and so in turn there must be extensional objects

to supply the designations for t (-\a.<P) at each particular world. But

for some purposes this is still not enough. In effect, the example just

given starts with an intensional object, and moves to extensional objects

derivatively. We need some machinery for moving in the other direction

as well.

Suppose, in a Henkin/Kripke model, we have somehow picked out

an extensional object of the same type at each world-say we call the

object we choose at world r, Or. It seems plausible that there should

be an intensional object: the chosen object. That is, there should be

an intensional object f whose value, at each world r, is the object Or.

More generally, suppose at each world we have selected a non-empty set

of extensional objects, all of the same type. Say at world r we select

the set Sr. Again it seems plausible that there should be an intensional

object-a selected object-a mapping f whose value at each world r is

some member of Sr.

Given the formal machinery up to this point, the existence of the

intensional objects posited above cannot be guaranteed. (At least, I be-

lieve this to be the case. I do not have a proof.) To postulate existence

of such intensional objects using some sort of axiom requires quantifica-

tion over possible worlds, which we cannot do, but we can approximate

to it by use of the 0 operator. What we wind up with is the following

MISCELLANEOUS MATTERS 129

postulate, which I call a choice axiom because, in effect, it posits the

existence of choice functions in the standard set-theoretic sense.

DEFINITION 9.15 (CHOICE AXIOM) Let t be an extensional type, and

let o: be of type j(t), f3 be of type t, and 'Y be of type jt. The following is

the choice axiom of type t.

(Vo:)[D(:3{3)o:(f3) ::::> (:3'Y)Do:(h)]

Informally, the axiom says that if, at each world the set of things such

that o: is non-empty-0(:3f3)o:(f3)-then there is a choice function 'Y that

picks out something such that o: at each world-(::l"f)Do:(h). I give one

example of a Choice Axiom application. Suppose o: is an extensional

variable, and m.<I> designates in every possible world. That is, in each

possible world, the <I> is meaningful. Then, plausibly, there should be an

intensional object that, in each world, designates the thing that is the <I>

of that world-that is, the term ?(.D(Ao:.<I>)(t() should also designate.

More loosely, the <I> concept should also designate. Recall, Definition 9.12

says what it means for a definite description to designate, and since

(A(.0(Ao:.<I>)(K))("7) = D(Ao:.<I>)(t'f]), things can be simplified a little.

PROPOSITION 9.16 Assume the Choice Axiom (Definition 9.15} and

Extensionality for Intensional Terms (Definition 9.3). Assume a, {3,

and 8 are of extensional type t, and 'Y and "7 are of type jt. The follow-

ing is valid in all K models.

D(:3{3)(V8)[(Ao:.<I>)(8) = ({3 = 8)] ::::> (:3'Y)(V"7)[D(Ao:.<I>)(t"7) = ('Y = 17)]

Proof Assume D(:3f3)(V8)[(Ao:.<I>)(8) = ({3 = 8)] is true at a possible

world. I show that (:3'Y)(V7])[D(Ao:.<I>)(1"7) = ('Y = 17)] must also be true

there. Start with

D(:3f3)(V8)[(Ao:.<I>)(8) = ({3 = 8)] (9.3)

which is equivalent to

0(:3{3){ (Ao:.<I>)(f3) 1\ (V8)[(Ao:.<I>)(8) ::::> ({3 = 8)]}. (9.4)

Instantiating the universal quantifier in the choice axiom with

(A"7.(Ao:.<I>)("7) 1\ (V8)[(Ao:.<I>)(8) ::::> ("7 = 8)])

(9.4) implies

130 TYPES, TABLEAUS, AND GODEL'S GOD

(:l!)D{(>,a.<P)(h) 1\ (V8)[(>,a.<P)(8):) (h = 8)]} (9.5)

which is equivalent to

(:J!){D(>,a.<P)(h) 1\ D(V8)[(>,a.<P)(8) :) (h = 8)]}. (9.6)

Since the Barcan and converse Barcan formulas are valid in the seman-

tics, this is equivalent to

(:3/){D(>,a.<P)(h) 1\ (V8)D[(>,a.<P)(8) :) (h = 8)]}. (9.7)

This, in turn, implies the following formula. I leave the justification to

you.

(:l!){D(>,a.<P)(h) 1\ (V1J)D[(>,a.<P)(11J) :) (h =11J)]}. (9.8)

Using distributivity of necessity over implication, this implies

(:3/){D(>,a.<P)(h) 1\ (V1J)[O(>,a.<P)(11J) :) D(h =11J)]} (9.9)

and using Extensionality for Intensional Terms, this implies

(:l{){D(>,a.<P)(h) 1\ (V1J)[D(>.a.<P)(l1J) :) (t = 17)]} (9.10)

which is equivalent to

(:l!)(V1J)[O(>,a.<P)(11J) = (t = 17)] (9.11)

and we are done.

Exercises

EXERCISE 6.1 Give tableau proofs of the Barcan formula, and of the

converse Barcan formula.

EXERCISE 6.2 Give a tableau proof to show (9. 7) implies (9.8).

III

ONTOLOGICAL ARGUMENTS

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 10

..

GODEL'S ARGUMENT, BACKGROUND

1. Introduction

There are many directions from which people have tried to prove the

existence of God. There have been arguments based on design: a com-

plex universe must have had a designer. There have been attempts to

show that the existence of an ethical sense implies the existence of God.

There have been arguments based on causality: trace the chain of effect

and cause backward and one must reach a first cause. Ontological ar-

guments seek to establish the existence of God based on pure logic: the

principles of reasoning require that God be part of ones ontology.

For religion, as contrasted with philosophy or logic, it does not matter

if proofs for God's existence have holes. Religious belief, like much that

is fundamentally human, is not really the product of reason. We are

emotional animals, and one of the uses of proof, in the various senses

above, is to sway emotion. Proof is often just a rhetorical device, one

among many.

But this takes us too far afield. Here we are interested in ontological

arguments only. Independently of whether one believes their conclu-

sion to be true, the logical machinery used in such arguments is often

ingenious, and merits serious study. It is generally accepted that such

arguments contain flaws, but saying exactly where the flaw lies is not

easy, and is subject to controversy. It happens that different analyses

of the same argument will locate an error at different points. Often

this happens because the notions involved in a particular ontological

argument are vague and subject to interpretation. Godel's ontological

argument is rather unique in that it is entirely precise-the premises are

clearly set forth, and the reasoning can be formalized. But we will see

133

134 TYPES, TABLEAUS, AND GODEL'S GOD

that here too there is room for interpretation, and things are not as clear

as they first seem.

2. Anselm

Historically, the first ontological argument is that of St. Anselm (1033

- 1109), given in his book Proslogion. Here is the argument itself, in a

somewhat technical translation from [Cha79].

Even the Fool, then, is forced to agree that something-than-which-nothing-

greater-can-be-thought exists in the mind, since he understands this when

he hears it, and whatever is understood is in the mind. And surely that-

than-which-a-greater-cannot-be-thought cannot exist in the mind alone. For

if it exists solely in the mind even, it can be thought to exist in reality also,

which is greater. If then that-than-which-a-greater-cannot-be-thought exists

in the mind alone, this same that-than-which-a-greater-cannot-be-thought is

that-than-which-a-greater-can-be-thought. But this is obviously impossible.

Therefore there is absolutely no doubt that something-than-which-a-greater-

cannot-be-thought exists both in mind and in reality.

Put into more modern terms, Anselm speaks of a maximally con-

ceivable being. This term-maximally conceivable being-must denote

something, since ''whatever is understood is in the mind." But a maxi-

mally conceivable being must have the property of existence, because if

it did not, we could conceive of a greater being, namely one that also

had the existence property.

My understanding of this is that, read with some charity, it shows the

phrase "maximally conceivable being," if it designates anything, must

designate something that exists. The flaw lies in the failure to properly

verify that the phrase designates at all-to show it is not in the same

category as "the round square." Indeed, Anselm's way of justifying this,

by claiming that it exists in the mind, is exactly what was attacked by

his contemporary Gaunilo, in his counter-argument, A Reply on Behalf

of the Fool. A modern translation of this can also be found in [Cha79].

Anselm's argument was the ancestor of various later versions, all of

which involve some notion of maximality. An easily accessible discussion

of the family of ontological arguments in general is in the on-line Stanford

Encyclopedia of Philosophy [Opp96b], and [Opp95, Pla65] are recom-

mended as more detailed studies. A full examination of the Anselm ar-

gument can be found in [Har65]. In addition, a detailed book in progress

is available on the internet, [SobOl]-the Anselm argument is discussed

in Chapter 2, "Classical Ontological Arguments."

3. Descartes

Descartes (1598 - 1650) gave different versions of an ontological ar-

gument. Here is one, in which he defines God to be a being whose

GODEL'S ARGUMENT, BACKGROUND 135

necessary existence is part of the definition. It is from the Appendix to

his replies to the Second Objections to his Meditations, [Des51]. I omit

the Definitions and Axioms to which the quote refers.

Proposition I

The existence of God is known from the consideration of his nature alone.

Demonstration

To say that an attribute is contained in the nature or in the concept of a

thing is the same as to say that this attribute is true of this thing, and that

it may be affirmed to be in it (Definition IX).

But necessary existence is contained in the nature or in the concept of God

(by Axiom X).

Hence it may with truth be said that necessary existence is in God, or that

God exists.

Here is a somewhat different argument, using existence rather than nec-

essary existence. This version is from The Meditations, book V, [Des51].

. . . because I cannot conceive God unless as existing, it follows that ex-

istence is inseparable from him, and therefore that he really exists; not that

this is brought about by my thought, or that it imposes any necessity on

things, but, on the contrary, the necessity which lies in the thing itself, that

is, the necessity of the existence of God, determines me to think in this way,

for it is not in my power to conceive a God without existence, that is a being

supremely perfect, and yet devoid of an absolute perfection, as I am free to

imagine a horse with or without wings.

Taking some liberties with the first of the Descartes proofs above: God

is the most perfect being, the being having all perfections, and among

these is necessary existence. Put a little differently, necessary existence

is part of the essence of God. And here we have reached an ontolog-

ical argument that can be easily formalized. Recall the discussion in

Chapter 7, Section 3. The type-0 objects are possibilist-they represent

what might exist, not what does. If we want to relativize things to what

actually exists, we need a type-(0) "existence" predicate, E, about which

nothing special need be postulated at this point. Now, suppose we define

God to be the necessarily existent being, that is, the being g such that

DE(g). If such a being exists, it must satisfy its defining property, and

hence we have

E(g) ::::> DE(g). (10.1)

Given (10.1), using the rule of necessitation, we have the following.

D[E(g) ::::> DE(g)] (10.2)

136 TYPES, TABLEAUS, AND GODEL'S GOD

From (10.2), using the K principle D(P :::) Q) :::) ( ()P :::) OQ) we have

the next implication.

OE(g) :::) ODE(g) (10.3)

Finally we use something peculiar to 85 (and some slightly weaker

logics, a point of no importance here). The principle needed is ODP :::)

DP, and so from (10.3) we have the following.

OE(g) :::) DE(g) (10.4)

We thus have a proof that God's existence is necessary, if possible.

And, again following Descartes loosely, God's existence is possible be-

cause possibility is identified with conceivability, and we may take it for

granted that God is conceivable.

Russell's treatment of definite descriptions applies quite well in a

modal setting-Chapter 9, Section 5. The use of g above was an informal

way of avoiding a formal definite description-note that I gave no real

prooffor (10.1). Let us recast the argument using definite descriptions-

the necessarily existent being is m.DE(a) and I assume g is an abbrevi-

ation for this type-0 term. Now (10.1) unabbreviates to the following.

:::) DE(1a.DE(a)). (10.5)

This is not a valid formula of K, but that logic is too weak anyway, given

the step from (10.3) to (10.4) above. But (10.5) is valid in 85, a fact

I leave to you as an exercise. In fact, using 85, the argument above is

entirely correct!

The real problem with the Descartes argument lies in the assump-

tion that God's existence is possible. In 85 both OE(g) :::) E(g) and

E(g) :::) OE(g) are trivially valid. Since OE(g) :::) DE(g) has been shown

to be valid, we have the equivalence of E(g), OE(g), and DE(g)! Thus,

assuming God's existence is possible is simply equivalent to assuming

God exists. This is an interesting conclusion for its own sake, but as an

argument for the existence of God, it is unconvincing.

Exercises

EXERCISE 3.1 Give an 85 tableau proof of the following, where P and

Q are type-(0) constant symbols.

P(m.DQ(a)) :::) DQ(m.DQ(a))

From this it follows that (10.5) is valid in 85.

GODEL'S ARGUMENT, BACKGROUND 137

EXERCISE 3.2 Construct a model to show

E(m.DE(a)) DE(1a.OE(a)).

is not valid inK.

EXERCISE 3.3 Formula 10.5 can also be written as

(.X,8.E(,8))(1a.OE(a)) D(.X,B.E(,B))(m.OE(a))

which, by the previous exercise, is not K valid. Show the following

variant is valid (a K tableau proof is probably easiest).

(.X,8.E(,8))(1a.DE(a)) (.X,B.OE(,B))(m.OE(a))

EXERCISE 3.4 Show why the valid K formula of Exercise 3.3 can not

be used in a Descartes-style argument.

4. Leibniz

Leibniz (1646 - 1716) partly accepted the Descartes argument from

The Meditations, mentioned in the previous section. But he also clearly

identified the critical issue: one must establish the possibility of God's

existence. The following is from Two Notations for Discussion with

Spinoza, [Lei56].

Descartes' reasoning about the existence of a most perfect being assumed

that such a being can be conceived or is possible. If it is granted that there

is such a concept, it follows at once that this being exists, because we set up

this very concept in such a way that it at once contains existence. But it

is asked whether it is in our power to set up such a being, or whether such

a concept has reality and can be conceived clearly and distinctly, without

contradiction. For opponents will say that such a concept of a most perfect

being, or a being which exists through its essence, is a chimera. Nor does

it suffice for Descartes to appeal to experience and allege that he experiences

this very concept in himself, clearly and distinctly. This is not to complete the

demonstration but to break it off, unless he shows a way in which others can

also arrive at an experience of this kind. For whenever we inject experience

into our demonstrations, we ought to show how others can produce the same

experience, unless we are trying to convince them solely through our own

authority.

Leibniz's remedy amounted to an attempt to prove that God's ex-

istence is possible, where God is defined to be the being having all

perfections-again a maximality notion. Intuitively, a perfection is an

atomic property that is, in some sense, good to have, positive. Leib-

niz based his proof on the compatibility of all perfections, from which

he took it to follow that all perfections could reside in a being-God's

138 TYPES, TABLEAUS, AND GODEL'S GOD

existence is possible. Here is another quote from Two Notations for

Discussion with Spinoza, [Lei56].

By a perfection I mean every simple quality which is positive and absolute

or which expresses whatever it expresses without any limits. But because a

quality of this kind is simple, it is unanalyzable or indefinable .... From this

it is not difficult to show that all perfections are compatible with each other

or can be in the same subject.

Leibniz goes on to provide a detailed proof of the compatibility of all

perfections, though it is not a proof in any modern sense. Indeed, it

is not clear how a proper proof could be given at all, using the vague

notion of perfection presented above. I omit his proof here. The point

for us is that, as we will see, precisely this point is central to Godel's

argument as well.

5. Godel

Godel (1906- 1978) was heir to the profound developments in math-

ematics of the late nineteenth and early twentieth centuries, which often

involved moves to greater degrees of abstraction. In particular, he was

influenced by David Hilbert and his school. In the tradition of Hilbert's

book, Foundations of Geometry, Godel avoided Leibniz's problems com-

pletely, by going around them. It is as if he said, "I don't know what

a perfection is, but based on my understanding of it intuitively, it must

have certain properties," and he proceeded to write out a list of axioms.

This neatly divides his ontological argument into two parts. First, based

on your understanding, do you accept the axioms. This is an issue of

personal intuitions and is not, itself, subject to proof. Second, does the

desired conclusion follow from the axioms. This is an issue of rigor and

the use of formal methods, and is what will primarily concern us here.

Godel's particular version of the argument is a direct descendent of

that of Leibniz, which in turn derives from one of Descartes. These argu-

ments all have a two-part structure: prove God's existence is necessary,

if possible; and prove God's existence is possible.

Godel worked on his ontological argument over many years. According

to [Ada95], there is a partial version in his papers dated about 1941.

In 1970, believing he would die soon, Godel showed his proof to Dana

Scott. In fact Godel did not die until1978, but he never published on the

matter. Information about the proof spread via a seminar conducted by

Dana Scott, and his slightly different version became public knowledge.

Godel's proof appeared in print in [Sob87], based on a few pages of

Godel's handwritten notes. Scott also wrote some brief notes, based

on his conversation with Godel, and [Sob87] provides these as well. In

fact, [Sob87] has served as something of a Bible (pun intended) for the

GODEL'S ARGUMENT, BACKGROUND 139

Godel ontological argument. Finally the publication of Godel's collected

works has brought a definitive version before the public, [G70]. Still,

the notion of a definitive version is rather elusive in this case. Godel's

manuscript provides almost no explanation or motivation. It amounts

to an invitation to others to elaborate.

Godel's argument is modal and at least second-order, since in his

definition of God there is an explicit quantification over properties. Work

on the Kripke semantics of modal logic was relatively new at the time

Godel wrote his notes, and the complexity of quantification in modal

contexts was perhaps not well appreciated. Consequently, the exact

logic Godel had in mind is unclear.

Subsequently several people took up the challenge of putting the

Godel argument on a firm foundation and exposing any hidden as-

sumptions. People have generally used the second-order modal logic

of [Coc69], sometimes rather informally. [Sob87], playing Gaunilo to

Godel's Anselm, showed the argument could be applied to prove more

than one would want. Sobel's discussion has been greatly extended in

[SobOl], Chapter 4; Chapter 3 is also relevant here. [AG96] showed

that one could view a part of the argument not as second-order, but as

third-order. Many others contributed, among which I mention [And90,

Haj96b]. Postings on the internet are, by nature, somewhat ephemeral,

but interesting discussions of the Godel argument, intended for a general

audience, can be found at [SmaOl] as well as at [OppOl]. In addition,

there are [Opp96b] and [SobOl]. The present chapter and the next can

be thought of as part of the continuing tradition of explicating Godel.

6. Godel's Argument, Informally

Before we get to precise details in the next Chapter, it would be good

to run through Godel's argument informally to establish the general

outline, since it is considerably more complex than the versions we have

seen to this point.

To begin with, Godel takes over the notion of perfection, but with

some changes. For Leibniz, perfections were atomic properties, and any

combination of them was compatible and thus could apply to some ob-

ject. They could be freely combined, a little like the atomic facts about

the world that one finds in Wittgenstein's Tractatus. Since this is the

case, why not form a new collection, consisting of all the various com-

binations of perfections, each combination of which Leibniz considers

possible. Godel found it convenient to do this, and used the term pos-

itiveness for the resulting notion. Thus we should think of a positive

property, in Godel's sense, as some conjunction of perfections in Leib-

140 TYPES, TABLEAUS, AND GODEL'S GOD

niz's sense. At least, I am assuming this to be the case-Godel says

nothing explicit about the matter.

The most notable difference between Godel and Leibniz is that, where

Leibniz tried to use what are essentially informal notions in a rigorous

way, Godel introduces formal axioms concerning them. Here are Godel's

axioms (or their equivalents), and his argument, set forth in everyday

English. A formalized version will be found in the next Chapter. The

Godel argument has the familiar two-part structure: God's existence is

possible; and God's existence is necessary, if possible. I'll take these in

order.

I'll begin with the axioms for positiveness. The first is rather strong.

(I have made no attempt to follow Godel's numbering of axioms and

propositions, and in some cases I have adopted equivalents or elabora-

tions of what Godel used.)

INFORMAL AXIOM 1 Exactly one of a property or its complement is pos-

itive.

It follows that there must be positive properties. If we call a property

that is not positive negative, it also follows that there are negative prop-

erties. By Informal Axiom 1, a negative property can also be described

as one whose complement is positive.

Suppose we say property P entails property Q if, necessarily, every-

thing having P also has Q.

INFORMAL AXIOM 2 Any property entailed by a positive property is pos-

itive.

This brings us to our first interesting result.

INFORMAL PROPOSITION 1 Any positive property is possibly instanti-

ated. That is, if P is positive, it is possible that something has property

P.

Proof Suppose P is positive. Let N be some negative property (the

complement of P will do). It cannot be that P entails N, or else N

would be positive. So it is not necessary that everything having P has

N, that is, it is possible that something has P without having N. So it

is possible that something hasP.

Leibniz attempted a proof that "all perfections are compatible with

each other or can be in the same subject," that is, having all perfec-

tions is a possibly instantiated property. Godel instead simply takes

GODEL'S ARGUMENT, BACKGROUND 141

the following as an axiom-it is an immediate consequence, using In-

formal Proposition 1, that having all positive properties is a possibly

instantiated property.

INFORMAL AXIOM 3 The conjunction of any collection of positive prop-

erties is positive.

This is a problematic axiom, in part because there are infinitely many

positive properties, and we cannot form an infinite conjunction (unless

we are willing to allow an infinitary language). There are ways around

this, but there is a deeper problem as well-we will see that this axiom

is equivalent to Godel's desired conclusion (given Godel's other assump-

tions). But further discussion of this point must wait till later on. For

now we adopt the axiom and work with it in an informal sense.

Now Godel defines God, or rather, defines the property of being God-

like, essentially the same way Leibniz did.

INFORMAL DEFINITION 2 A God is any being that has every positive

property.

This gives us part one of the argument rather easily.

INFORMAL PROPOSITION 3 It is possible that a God exists.

Proof By Informal Axiom 3, the conjunction of all positive properties

is a positive property. But by Definition 2, this property-maximal

positiveness-is what makes one a God. Since the property is positive,

it is possibly instantiated, by Informal Proposition 1.

There are also a few technical assumptions concerning positiveness,

whose role is not apparent in the informal presentation given here. Their

significance will be seen when we come to the formalization in the next

Chapter. Here is one.

INFORMAL AXIOM 4 Any positive property is necessarily so, and any

negative property is necessarily so.

Now we move on to the second part of the argument, showing God's

existence is necessary, if possible. Here Godel's proof is quite different

from that of Descartes, and rather ingenious. To carry out the argument,

Godel introduces a pair of notions that are of interest in their own right.

INFORMAL DEFINITION 4 A property G is the essence of an object g if:

1 g has property G;

142 TYPES, TABLEAUS, AND GODEL'S GOD

2 G entails every property of g.

Strictly speaking, in the definition above I should have said an essence

rather than the essence, but it is an easy argument that essences are

unique, if they exist at all. Very simply, if an object g had two essences,

P and Q, each would be a property of g by part 1, and then each would

entail the other by part 2. Godel does not, in general, assume that

objects have essences, but for an object that happens to be a God, there

is a clear candidate for the essence.

INFORMAL PROPOSITION 5 If g is a God, the essence of g is being a

God.

Proof Let's state what we must show a little more precisely. Suppose

G is the conjunction of all positive properties, so having property G is

what it means to be a God. It must be shown that if an object g has

property G, then G is the essence of g.

Suppose g has property G. Then automatically we have part 1 of

Informal Definition 4.

Suppose also that P is some property of g. By Informal Axiom 1, if

P were not positive its complement would be. Since g has all positive

properties, g then would have the property complementary toP. Since

we are assuming g has P itself, we would have a contradiction. It follows

that P must be positive. Since G is the conjunction of all positive

properties, clearly G entails P. Since P was arbitrary, G entails every

property of g, and we have part 2 of Informal Definition 4.

Here is the second of Godel's two new notions.

INFORMAL DEFINITION 6 An object g has the property of necessarily

existing if the essence of g is necessarily instantiated.

And here is the last of G6del's axioms.

INFORMAL AXIOM 5 Necessary existence, itself, is a positive property.

INFORMAL PROPOSITION 7 If a God exists, a God exists necessarily.

Proof Suppose a God exists, say object g is a God. Then g has all

positive properties, and these include necessary existence by Informal

Axiom 5. Then the essence of g is necessarily instantiated, by Informal

Definition 6. But the essence of g is being a God, by Informal Propo-

sition 5. Thus the property of being a God is necessarily instantiated .

GODEL'S ARGUMENT, BACKGROUND 143

Now we present the second part of the ontological proof.

INFORMAL PROPOSITION 8 If it is possible that a God exists, it is nec-

essary that a God exists (assuming the logic is 85).

Proof In any modal logic at least as strong as K, if P :::::> Q is valid, so is

OP :::::> OQ. Then by Informal Proposition 7, if it is possible that a God

exists, it is possibly necessary that a God exists. In 85, ODP :::::> DP is

valid, and the conclusion follows.

Finally, by Informal Propositions 3 and 8, we have our conclusion.

INFORMAL THEOREM 9 Assuming all the axioms, and assuming the un-

derlying logic is 85, a God necessarily exists.

One final remark before moving on. I've been referring to a God,

rather than to the God. As a matter of fact uniqueness is easy to estab-

lish, provided we make use of Leibniz's condition that having the same

properties ensures identity. Let G be the property of being Godlike-

the maximal positive property-and suppose both g1 and g2 possess this

property. By Informal Proposition 5, G must be the essence of both g

1

and g2. Now, if P is any property of g1, G must entail P, by part 2

of Informal Definition 4. Since G is a property of g2, by part 1 of the

same Informal Definition, P must also be a property of g2. Similarly,

any property of g2 must be a property of g1. Since g1 and g2 have the

same properties, they are identical.

This concludes the informal presentation of Godel's ontological argu-

ment. It is clear it is of a more complex nature than those that histor-

ically preceded it. But an informal presentation is simply not enough.

God is in the details, so to speak, and details demand a formal ap-

proach. In the next Chapter I'll go through the argument again, more

slowly, working things through in the intensional logic developed earlier

in Part II.

Exercises

EXERCISE 6.1 Show that only God can have a positive essence. (This

exercise is due to Ioachim Teodora Adelaida of Bucharest.)

M. Fitting, Types, Tableaus, and Gdel's God

Kluwer Academic Publishers 2002

Chapter 11

..

GODEL'S ARGUMENT, FORMALLY

1. General Plan

The last Chapter ended with an informal presentation of Godel's argu-

ment. This one is devoted to a formalized version. I'll also consider some

objections and modifications. There are two kinds of objections. One

amounts to saying that Godel committed the same fallacy Descartes did:

assuming something equivalent to God's existence. Nonetheless, again

as in the Descartes case, much of the argument is of interest even if it

falls short of establishing the desired conclusion. The second kind of

objection is that Godel's axioms are too strong, and lead to a collapse

of the modal system involved. Various extensions and modifications of

Godel's axioms have been proposed, to avoid this modal collapse. I'll

discuss these, and propose a modification of my own. Now down to de-

tails, with the proof of God's possible existence coming first. I will not

try to match the numbering of the informal axioms in the last chapter,

but I will refer to them when appropriate.

2. Positiveness

God, if one exists, will be taken to be an object of type 0. We are

interested in the intensional properties of this object, properties of type

j(O). Among these properties are the ones Godel calls positive, and which

we can think of as conjunctive combinations of Leibniz's perfections. At

least that is how I understand positiveness. Godel's ideas on the subject

are given almost no explanation in his manuscript-here is what is said,

using the translation of [G70].

145

146 TYPES, TABLEAUS, AND GODEL'S GOD

Positive means positive in the moral aesthetic sense (independently of the

accidental structure of the world). Only then [are] the axioms true. It may also

mean pure 'attribution' as opposed to 'privation' (or containing privation).

This is not something I profess to understand. But what is significant is

that, rather than attempting to define positiveness, Godel characterized

it axiomatically. In this section I present his basic axioms concerning

the notion, and I explore some of their consequences.

DEFINITION 11.1 (POSITIVE) A constant symbol P of type j (j (0)) zs

designated to represent positiveness. It is an intensional property of

intensional properties. Informally, P is positive if we have P(P).

It is convenient to introduce the following abbreviation.

DEFINITION 11.2 (NEGATIVE) If T is a term of type j(O), take T as

short for (Ax.T(x)). Call T negative if T is positive.

Loosely, at a world in a model, T denotes the complement of whatever T

denotes. It is easy to check formally that T = ( T), given extensionality

for intensional terms, Definition 9.3.

Godel assumes that, for each P, exactly one of it or its negation must

be positive. Godel's axiom (which he actually stated using exclusive-or)

can be broken into two implications. Here they have been formulated as

two separate axioms, since they play different roles.

AXIOM 11.3 (FORMALIZING INFORMAL AXIOM 1)

A (VX)[P(X) :J P(X)]

B (VX)[P(X) :J P(X)]

Of these, Axiom 11.3A is certainly plausible: contradictory items should

not both be positive. But Axiom 11.3B is more problematic: it says

one of a property or its complement must be positive. We might think

of the notion of a maximal consistent set of formulas-familiar from

the Lindenbaum/Henkin approach to proving classical completeness-as

suggestive of what Godel had in mind. There are some cryptic remarks

of Godel relating disjunctive normal forms and positiveness, but these

have not served as aids to my understanding. At any rate, these are the

basic assumptions.

The next assumption concerning positiveness is a monotonicity condi-

tion: a property that is entailed by a positive property is, itself, positive.

Here it is, more or less as Godel gave it.

[P(X) 1\ D(Vx)(X(x) :J Y(x))] :J P(Y)

GODEL'S ARGUMENT, FORMALLY 147

In this formula, x is a free variable of type 0. For us, type-0 quantification

is possibilist, while for Godel it must have been actualist. I am assuming

this because his conclusion, that God exists, is stated using an existential

quantifier, and a possibilist quantifier would have been too weak for the

purpose. For us, existence must be made explicit using the existence

predicate E, relativizing the ('v'x) quantifier to E. Since this relativization

comes up frequently, it is best to make an official definition.

DEFINITION 11.4 (EXISTENTIAL RELATIVIZATION) ('v'Ex)<I> abbrevi-

ates (Vx)[E(x) :J <I>], and (3Ex)<I> abbreviates (3x)[E(x) 1\ <I>].

AXIOM 11.5 (FORMALIZING INFORMAL AXIOM 2)

In the following, x is of type 0, X and Y are of type i(O).

('v'X)(VY){[P(X) 1\ D('v'Ex)(X(x) :J Y(x))] :J P(Y)}

At one point in his proof, Godel asserts that (>.x.x = x) must be

positive if anything is, and (>.x.x = x) must be negative. This is easy

to see: P( (>.x.x = x)) is valid if anything is positive because anything

strictly implies a validity, and we have Axiom 11.5. The assertion that

(>.x.x = x) is negative is equivalent to the assertion that (>.x.x = x) is

positive. We thus have the following consequences of Axiom 11.5.

PROPOSITION 11.6 Assuming Axiom 11.5:

1 (3X)P(X) :J P((>.x.x = x));

2 (3X)P(X) :J P((Ax.x = x)).

PROPOSITION 11.7 Assuming Axioms 11.3A and 11.5:

(3X)P(X) :J P( (>.x.x = x) ).

Now we have a result from which the possible existence of God will

follow immediately, given one more key assumption about positiveness.

PROPOSITION 11.8 (FORMALIZING INFORMAL PROPOSITION 1)

Assuming Axioms 11.3A and 11.5, ('v'X){P(X) :J 0(3Ex)X(x)}.

Proof The idea has already been explained, in the proof of Informal

Proposition 1 in Section 6. This time I give a formal tableau, which is

displayed in Figure 11.1. In it use is made of one of the Propositions

above. Item 1 negates the proposition in unabbreviated form. Item 2 is

from 1 by an existential rule (with P as a new parameter); 3 and 4 are

from 2 by a conjunctive rule; 5 is Axiom 1; 6 is from 5 and 7 is from 6

by universal rules; 8 and 9 are from 7 by a disjunctive rule; 10 and 11

148 TYPES, TABLEAUS, AND GODEL'S GOD

are from 8 by a disjunctive rule; 12 is from 11 by a possibility rule; 13

is from 12 by an existential rule (with pas a new parameter, and some

tinkering with E); 14 and 15 are from 13 by a conjunctive rule; 16 is

from 4 by a necessity rule; 17 is from 16 by a universal rule (and some

tinkering with E again); 18 is Proposition 11.7; 19 and 20 are from 18

by a disjunctive rule; 21 is from 19 by a universal rule.

Leibniz attempted to prove that perfections are mutually compatible,

basing his proof on the idea that perfections can only be purely positive

qualities and so none can negate the others. For Godel, rather than

proving any two perfections could apply to the same object, Godel as-

sumes the positive properties are closed under conjunction. This turns

out to be a critical assumption. In stating the assumption, read X 1\ Y

as abbreviating (.Xx.X(x) 1\ Y(x)).

AXIOM 11.9 (FORMALIZING INFORMAL AXIOM 3)

(VX)(VY){[P(X) 1\ P(Y)] ::::> P(X 1\ Y)}

Godel immediately adds that this axiom should hold for any number

of summands. Of course one can deal with a finite number of them

by repeated use of Axiom 11.9 as stated-the serious issue is that of

an infinite number, which Godel needs. [AG96] gives a version of the

axiom which directly postulates that the conjunction of any collection

of positive properties is positive. Note that it is a third-order axiom.

For reading ease I use the following two abbreviations.

1 Z applies only to positive properties (Z, like P, is of type j(j(O))):

pos(Z) {::} (\iX)[Z(X) ::::> P(X)]

2 X applies to those objects which possess exactly the properties falling

under Z-roughly, X is the (necessary) intersection of Z. (In this,

Z is of type i(j(O)), X is of type j(O), and x is of type 0.)

(X intersection of Z) {::} D(\ix){X(x) = (W)[Z(Y) ::::> Y(x)]}

AXIOM 11.10 (ALSO FORMALIZING INFORMAL AXIOM 3)

(\iZ){pos(Z) ::::> (\iX)[(Xintersection of Z) ::::> P(X)]}.

Axiom 11.10 implies Axiom 11.9. I leave the verification to you. I'll

finish this section with two technical assumptions that Godel makes

"because it follows from the nature of the property." I don't understand

this terse explanation, but here are the assumptions.

(\iX)[P(X) ::::> DP(X)]

(\iX)[P(X) ::::> 0-.P(X)]

GODEL'S ARGUMENT, FORMALLY 149

u

150 TYPES, TABLEAUS, AND GODEL'S GOD

If the underlying logic is just K, equivalence of these two assumptions

follows from Axioms 11.3A and 11.3B. And if the underlying logic is 85,

as it must be for part of Godel's argument, equivalence also follows by

Proposition 9.10. Consequently the version used here can be simplified.

AXIOM 11.11 (FORMALIZING INFORMAL AXIOM 4)

('v'X)[P(X) :J DP(X)].

P has been taken to be an intensional object, of type j(j(O)). Ax-

iom 11.11 and Theorem 9.11 tells us that Pis rigid. In effect the inten-

sionality of P is illusory-since it is rigid it could just as well have been

an extensional object of type (j(O)).

Exercises

EXERCISE 2.1 Give a tableau proof that (.Ax.(x = x)) = (.Ax.x = x).

More generally, show that for a type (0) term T, (T) = T.

EXERCISE 2.2 Show that ('v'X)[P(X) :J DP(X)] follows from Ax-

iom 11.11 together with Axioms 11.3A and 11.3B.

EXERCISE 2.3 Show Axiom 11.10 implies Axiom 11.9. Hint: use equal-

ity.

3. Possibly God Exists

Godel defines something to be Godlike if it possesses all positive prop-

erties.

DEFINITION 11.12 (FORMALIZING INFORMAL DEFINITION 2)

G is the following type j(O) term, where Y is type j(O).

(.Ax.('v'Y)[P(Y) :J Y(x)]).

Given certain earlier assumptions, anything having all positive prop-

erties can only have positive properties. Perhaps the easiest way to state

this formally is to introduce a second notion of Godlikeness, and prove

equivalence.

DEFINITION 11.13 (ALSO FORMALIZING INFORMAL DEFINITION 2)

G* is the type j(O) term

(>.x.(W)[P(Y) = Y(x)]).

The following result is easily proved; I leave it to you as an exercise.

GODEL'S ARGUMENT, FORMALLY 151

PROPOSITION 11.14 Assume Axiom 11.3B, ('v'X)[P(X) :J P(X)].

InK, with this assumption, ('v'x)[G(x) = G*(x)].

Axiom 11.3B is a little problematic, but it is essential to the Proposi-

tion above. If, eventually, we show something having property G exists,

and G and G* are equivalent, we will know that something having prop-

erty G* exists. But the converse is also the case: if something having

property G* exists, Axiom 11.3B is the case, even if the existence in

question is possibilist. Here is a formal statement of this. Once again I

leave the proof to you.

PROPOSITION 11.15 InK, (3x)G*(x) :J (VX)[P(X) :J 'P(X)].

Now we can show that God's existence is possible. Godel assumes the

conjunction of any family of positive properties is positive. Since G* is,

in effect, the conjunction of all positive properties, it must be positive,

and hence so must G be.

PROPOSITION 11.16 InK Axiom 11.10 implies P(G).

Once again I leave the formal verification to you. What must be

shown is the following.

('v'Z)('v'X){[pos(Z) 1\ (X intersection of Z)] :J P(X)} :J P(G)

Essentially, this is the case because, as is easy to verify, we have each of

pos(P) and ( G intersection of P).

Now the possibility of God's existence is easy. In fact, it can be proved

with an actualist quantifier, though only the weaker possibilist version

is really needed for the rest of the argument.

THEOREM 11.17 Assume Axioms 11.3A, 11.5, and 11.10. InK both

of the following are consequences. 0(3Ex)G(x) and 0(3x)G(x).

Proof By Proposition 11.8,

('v'X){P(X) :J 0(3Ex)X(x)},

hence trivially,

('v'X){P(X) :J 0(3x)X(x)}.

By the Proposition above, P(G). The result is immediate.

Note that the full strength of Proposition 11.8 was not really needed

for the possibilist conclusion. In fact, if we modify Axiom 11.5 so that

quantification is possibilist,

(VX)(W){[P(X) 1\ D('v'x)(X(x) :J Y(x))J :J P(Y)}

152 TYPES, TABLEAUS, AND GODEL'S GOD

we would still be able to prove Proposition 11.8 in the weaker form

(\fX){P(X) :J 0(:3x)X(x)}

and the Godel proof would still go through.

Exercises

EXERCISE 3.1 Give a tableau proof that G entails any positive property:

(\fX){P(X) :J D(\fy)[G(y) :J X(y)]}. You will need Axiom 11.11.

EXERCISE 3.2 Give a tableau proof for Proposition 11.14.

EXERCISE 3.3 Give a tableau proof for Proposition 11.15.

EXERCISE 3.4 Give a tableau proof for Proposition 11.16.

EXERCISE 3.5 Give a tableau proof of

(\fZ)(\fX){[pos(Z) 1\ (X intersection of Z)] :J P(X)} :J P(G).

4. Objections

Godel replaced Leibniz's attempted proof of the compatibility of per-

fections by an outright assumption, given here as Axiom 11.10. Dana

Scott, apparently noting that the only use Godel makes of this Axiom

is to show being Godlike is positive, proposed taking P( G) itself as an

axiom. Indeed, Scott maintains that the Godel argument really amounts

to an elaborate begging of the question-God's existence is simply being

assumed in an indirect way. In fact, it is precisely at the present point

in the argument that Scott's claim can be localized. Godel's assumption

that the family of positive properties is closed under conjunction turns

out to be equivalent to the possibility of God's existence, a point also

made in [SobOl].

We will see, later on, Godel's proof that God's existence is necessary,

if possible, is correct. It is substantially different from that of Descartes,

and has many points of intrinsic interest. What is curious is that the

proof as a whole breaks down at precisely the same point as that of

Descartes: God's possible existence is simply assumed, though in a dis-

guised form.

The rest of this section provides a formal proof of the claims just

made. Enough tableau proofs have been given in full, by now, so that

abbreviations can be introduced as an aid to presentation. Before giving

the main result of this section, I introduce some simple conventions for

shortening displayed tableau derivations.

GODEL'S ARGUMENT, FORMALLY 153

If a X and a X :J Y occur on a branch, a Y can be added. Schemat-

ically,

aX

aX :J Y

aY

The justification for this is as follows.

aX 1.

a X :J Y 2.

a X 3. aY 4.

The left branch is closed, and the branch below 4 continues as if we had

used the derived rule.

Here are a few more derived rules, whose justification I leave to you.

aX

a (X 1\ Y) :J Z

aY :J Z

aX

aX=:Y

aY

a ('v'o:1) ('v'o:n)<P( 0:1, , O:n)

a<P(T1, ... ,Tn)

for any grounded

terms T1, ... , Tn

aX

a X=:Y

a--,Y

a (:lo:1) (:lo:n)<P( 0:1, , O:n)

a<P(P1, ... ,Pn)

for any new, distinct

parameters P1, ... , Pn

Now, here is the promised proof of equivalence.

THEOREM 11.18 Assume all the Axioms to this point, except for Ax-

iom 11.10 and Axiom 11.9. The following are equivalent, using 85:

1 Axiom 11.10;

2 P(G);

3 O(:JEx)G(x);

154 TYPES, TABLEAUS, AND GODEL'S GOD

4 0(3x)G(x).

Proof We already know 1 implies 2, this is Proposition 11.16. Likewise

3 follows from 2, by Theorem 11.17. And the implication of 4 from 3 is

trivial.

Showing that 4 implies 2 is straightforward, using the fact that G

and G* are equivalent, and the fact that positiveness is rigid. Here is a

tableau derivation.

1 0(3x)G(x) 1.

1 P(G) 2.

1.1 (3x)G(x) 3.

1.1 G(g) 4.

1.1 ('v'x)[G(x) = G*(x)] 5.

1.1 [G(g) = G*(g)] 6.

1.1 G*(g) 7.

1.1 (>.x.('v'Y)[P(Y) = Y(x)])(g) 8.

1.1 (W)[P(Y) = Y(g)] 9.

1.1 [P(G) = G(g)] 10.

1.1 P(G) 11.

1 (VX)[P(X) :J 0-.P(X) 12.

1 P( G) :J DP( G) 13.

1 DP( G) 14.

1.1P(G) 15.

Item 3 is from 1 by a possibility rule; 4 is from 3 by an existential

rule, with g as a new parameter; 5 is Proposition 11.14, and note that

the modal version of Corollary 4.34 is being used here; 6 is from 5 by a

universal rule; 7 is from 4 and 6 by a derived rule; 8 is 7 unabbreviated;

9 is from 8 by an abstraction rule; 10 is from 9 by a universal rule; 11 is

from 10 and 4 by a derived rule; 12 is an equivalent of Axiom 11.11; 13

is from 12 by a universal rule; 14 is from 2 and 13 by a derived rule; 15

is from 14 by a necessity rule.

Showing 2 implies 1 informally is also not hard. If C is any collection

of positive properties, G entails every member of C by Exercise 3.1.

It follows that G also entails the conjunction of C. Since 2 says G is

positive, the conjunction of C is positive by Axiom 11.5. The informal

argument just sketched can be turned into a proper tableau proof. In

Figure 11.2 I give a proof that 2 implies Axiom 11.9, and I'll leave the

argument for Axiom 11.10 as an exercise.

In Figure 11.2, item 3 is from 2 by a (derived) existential rule; 4 and 5

are from 3, and 6 and 7 are from 4 by conjunctive rules; 8 is Axiom 11.5;

9 is from 8 by a derived universal rule; 10 is from 1 and 9 by a derived

GODEL'S ARGUMENT, FORMALLY

1 P(G) 1.

1 -{v'X)(W){[P(X) 1\ P(Y)] :::> P(X 1\ Y)} 2.

1 {[P(A) 1\ P(B)] :::> P(A 1\ B)} 3.

1 P(A) 1\ P(B) 4.

1 P(A 1\ B) 5.

1 P(A) 6.

1 P(B) 7.

1 ('v'X)('v'Y){[P(X) 1\ D('v'Ex)(X(x) :::> Y(x))] :::> P(Y)} 8.

1 [P(G) 1\ D('v'Ex)(G(x) :::>(A 1\ B)(x))] :::> P(A 1\ B) 9.

1 D('v'Ex)(G(x) :::>(A 1\ B)(x)) :::> P(A 1\ B) 10.

1 D('v'Ex)(G(x) :::>(A 1\ B)(x)) 11.

1.1(\:IEx)(G(x) :::>(A 1\ B))(x) 12.

1.1(\:lx)[E(x) :::> (G(x) :::> (A 1\ B)(x))] 13.

1.1-,[E(c) :::> (G(c) :::>(A 1\ B)(c))] 14.

1.1 E(c) 15.

1.1(G(c) :::> (A 1\ B)( c)) 16.

1.1 G(c) 17.

1.1-,(AAB)(c) 18.

1 ('v'X)[P(X) :::> DP(X)] 19.

1 P(A) :::> DP(A) 20.

1 P(B) :::> DP(B) 21.

1 DP(A) 22.

1 DP(B) 23.

1.1 P(A) 24.

1.1 P(B) 25.

1.1 (>.x.('v'Y)[P(Y) :::> Y(x)])(c) 26.

1.1 ('v'Y)[P(Y) :::> Y(c)] 27.

1.1 P(A) :::> A(c) 28.

1.1 P(B) :::> B(c) 29.

1.1 A(c) 30.

1.1 B(c) 31.

1.1(.\x.A(x) 1\ B(x))(c) 32.

1.1[A(c) 1\ B(c)] 33.

~

1.1A(c) 34. 1.1B(c) 35.

Figure 11.2. Proof that item 2 implies Axiom 11.9

155

156 TYPES, TABLEAUS, AND GODEL'S GOD

rule; 11 is from 5 and 10 by a derived rule; 12 is from 11 by a possibility

rule; 13 is 12 unabbreviated; 14 is from 13 by an existential rule; 15

and 16 are from 14, and 17 and 18 are from 16 by conjunctive rules; 19

is Axiom 11.11; 20 and 21 are from 19 by universal rules; 22 is from 6

and 20, and 23 is from 7 and 21, by derived rules; 24 is from 22 and

25 is from 23 by necessity rules; 26 is 17 unabbreviated; 27 is from 26

by an abstraction rule; 28 and 29 are from 27 by universal rules; 30 is

from 24 and 28, and 31 is from 25 and 29 by derived rules; 32 is 18

unabbreviated; 33 is from 32 by an abstraction rule; 34 and 35 are from

33 by a disjunctive rule.

Exercises

EXERCISE 4.1 Give a tableau proof showing that 0(3x)G(x) implies

Axiom 11.10.

5. Essence

Even though we ran into the old Descartes problem with half of the

Godel argument, we should not abandon the enterprise. The other half

contains interesting concepts and arguments. This is the half in which

it is shown that God's existence is necessary, if possible. For starters,

Godel defines a notion of essence that plays a central role, and is of

interest in its own right. [Haz98] makes a case for calling Godel's notion

character, reserving the term essence for something else. I follow Godel's

terminology. The essence of something, x, is a property that entails every

property that x possesses. Godel says it as follows.

cp Ess x = (\17/!){7/J(x) 0(\fy)[cp(y) 7/J(y)]}

As just given, it does not follow that the essence of x must be a property

that x possesses. Dana Scott assumed this was simply a slip on the part

of Godel, and inserted a conjunct cp(x) into the definition. I will follow

him in this.

cp Ess x = cp(x) 1\ (V7j!){7j!(x) D('v'y)[cp(y) 7/J(y)]}

Godel states cp Ess x as a formula rather than a term-in the version

in this book an explicit predicate abstract is used. Also, I assume the

type-0 quantifier that appears is actualist, and so in my version the

existence predicate, E, must appear. (P, q) is intended to assert that

P is the essence of q.

DEFINITION 11.19 (ESSENCE, FORMALIZING INFORMAL DEF. 4)

abbreviates the following type i (i (0), 0) term, in which Z is of type

GODEL'S ARGUMENT, FORMALLY

i(O) and w is of type 0:

(.XY,x.Y(x) A ('v'Z){Z(x):) O('v'Ew)[Y(w):) Z(w)]})

157

The property of being Godlike was defined earlier, Definition 11.12.

A central fact about Godlikeness, from Godel's notes, is that it is the

essence of any being that is Godlike.

THEOREM 11.20 (FORMALIZING INFORMAL PROPOSITION 5)

Assume Axioms 11.3B and 11.11. InK the following is provable. (Note

that x is of type 0.)

('v'x)[G(x) :) ( G, x)].

Rather than giving a direct proof, if we use Proposition 11.14 it follows

from a similar result concerning G*, provided Axiom 11.3B is assumed.

Since such a result has a somewhat simpler proof, this is what is actually

shown.

THEOREM 11.21 InK the following is provable, assuming Axiom 11.11.

(\fx)[G*(x):) (G*,x)].

Proof Here is a closed K tableau to establish the theorem.

1 ('v'x )[G* (x) :) ( G*, x )] 1.

1 -.., [ G* (g) :) ( G*, g)] 2.

1 G*(g) 3.

1-.E(G*,g) 4.

1-.{G*(g) A ('v'Z){Z(g):) D('v'Ew)[G*(w):) Z(w)]}} 5.

~

1-.G*(g) 6. 1(\IZ){Z(g):) O('v'Ew)[G*(w):) Z(w)]} 7.

Item 2 is from 1 by an existential rule, with g a new parameter; 3 and

4 are from 2 by a conjunction rule; 5 is from 4 by a derived unsubscripted

abstract rule; 6 and 7 are from 5 by a disjunction rule. The left branch

is closed. I continue with the right branch, below item 7.

158 TYPES, TABLEAUS, AND GO DEL'S GOD

1 {Q(g) :J D(VEw)[G*(w) :J Q(w)]} 8.

1 Q(g) 9.

1 D(VEw)[G*(w) j Q(w)] 10.

1.1(V'Ew)[G*(w) :J Q(w)] 11.

1.1{E(a) :J [G*(a) :J Q(a)]} 12.

1.1 E(a) 13.

1.1--,[G*(a) :J Q(a)] 14.

1.1 G*(a) 15.

1.1 Q( a) 16.

1 (VY)[P(Y) = Y(g)] 17.

1 P(Q) = Q(g) 18.

1 P(Q) 19.

1.1 (VY)[P(Y) = Y(a)] 20.

1.1 P(Q) = Q(a) 21.

1 (VY)[P(Y) :J DP(Y)] 22.

1 P( Q) :J DP( Q) 23.

1 DP(Q) 24.

1.1 P(Q) 25.

1.1 Q(a) 26.

Item 8 is from 7 by an existential rule, with Q a new parameter; 9 and

10 are from 8 by a conjunction rule; 11 is from 10 by a possibility rule; 12

is from 11 by an existential rule; 13 and 14 are from 12 by a conjunctive

rule, as are 15 and 16 from 14; 17 is from 3 by a derived unsubscripted

abstract rule; 18 is from 17 by a universal rule; 19 is from 9 and 18 by an

earlier derived rule; 20 is from 15 by a derived unsubscripted abstract

rule; 21 is from 20 by a universal rule; 22 is Axiom 11.11; 23 is from 22

by a universal rule; 24 is from 19 and 23 by a derived rule; 25 is from 24

by a necessity rule; 26 is from 21 and 25 by a derived rule. The branch

is closed by 16 and 26.

In the notes Dana Scott made when Godel showed him his proof, there

are two observations concerning essences. One is that something can

have only one essence. The other is that an essence must be a complete

characterization. Here are versions of these results. I begin by showing

that any two essences of the same thing are necessarily equivalent.

THEOREM 11.22 Assume the modal logic is K. The following is prov-

able.

(VX)(VY)(Vz){[t'(X, z) 1\ t'(Y, z)] :J D(VEw)[X(w) = Y(w)]}

Proof The idea behind the proof is straightforward. If P and Q are

essences of the same object, each must entail the other. I give a tableau

GODEL'S ARGUMENT, FORMALLY 159

proof mainly to provide another example of such. It starts by negating

the formula, applying existential rules three times, introducing new pa-

rameters P, Q, and a, then applying various propositional rules. Omit-

ting all this, we get to items 1 - 3 below.

1 E(P, a) 1.

1 E(Q, a) 2.

1D('v'Ew)[P(w) = Q(w)] 3.

1 P(a) 4.

1 ('v'Z)[Z(a) :J D('v'Ew)[P(w) :J Z(w)]] 5.

1 Q(a) 6.

1 ('v'Z)[Z(a) :J D('v'Ew)[Q(w) :J Z(w)Jl 7.

1 Q(a) :J D('v'Ew)[P(w) :J Q(w)] 8.

1 P(a) :J D('v'Ew)[Q(w) :J P(w)] 9.

~

1Q(a) 10. 1 D('v'Ew)[P(w) :J Q(w)] 11.

~

1P(a) 12. 1 D('v'Ew)[Q(w) :J P(w)] 13.

Items 4 and 5 are from 1 by an abstraction rule (and a propositional

rule), 6 and 7 are from 2 the same way; 8 is from 5 and 9 is from 7

by universal rules; 10 and 11 are from 8, and 12 and 13 are from 9 by

disjunction rules. The left branch is closed, by 6 and 10. The middle

branch is closed by 4 and 12. I continue with the rightmost branch,

below item 13.

1.1('v'Ew)[P(w) = Q(w)] 14.

1.1{E(b) :J [P(b) = Q(b)]} 15.

1.1 E(b) 16.

1.1[P(b) = Q(b)] 17.

~

1.1 P(b) 18. 1.1P(b) 20.

1.1Q(b) 19. 1.1 Q(b) 21.

Item 14 is from 3 by a possibility rule; 15 is from 14 by an existential

rule; 16 and 17 are from 15 by a conjunction rule; 18, 19, 20, 21 are from

17 by successive propositional rules. I show how the left branch can

be continued to closure; the right branch has a symmetric construction

which I omit.

160 TYPES, TABLEAUS, AND GODEL'S GOD

1.1 (VEw)[P(w) :J Q(w)] 22.

1.1 E(b) :J [P(b) :J Q(b)] 23.

~

1.1--,E(b) 24. 1.1 P(b) :J Q(b) 25.

~

1.1--,P(b) 26. 1.1 Q(b) 27.

Item 22 is from 11 by a necessitation rule; 23 is from 22 by a universal

rule; 24 and 25 are from 23 by a disjunction rule, as are 26 and 27 from

25. The left branch is closed by 16 and 24, the middle branch is closed

by 18 and 26, and the right branch is closed by 19 and 27.

Now, here is the second of Scott's observations: if X is the essence of

y, only y can have X as a property.

THEOREM 11.23 Assume the modal logic is K, including equality. The

following is valid.

(VX)(Vy){ (X, y) :J D(VEz)[X(z) :J (y = z)]}

This can be proved using tableaus-! leave it to you as an exercise.

Exercises

EXERCISE 5.1 Give a tableau proof for Theorem 11.23. Hint: for a

parameter c, one can consider the property of being, or not being, c,

that is, (.Xx.x =c) and (.Xx.x #c). Either property can be used in the

proof.

EXERCISE 5.2 Give a tableau proof to establish Theorem 11.20 directly,

without using G*.

6. Necessarily God Exists

In this section I present a version of Godel's argument that God's

possible existence implies His necessary existence. It begins with the

introduction of an auxiliary notion that Godel calls necessary existence.

DEFINITION 11.24 (NECESSARY EXISTENCE)

(Formalizing Informal Definition 6) N abbreviates the following

type i(O) term:

(.Xx.(VY)[(Y, x) :J 0(3Ez)Y(z)]).

GODEL'S ARGUMENT, FORMALLY 161

The idea is, something has the property N of necessary existence pro-

vided any essence of it is necessarily instantiated. Godel now makes a

crucial assumption: necessary existence is positive.

AXIOM 11.25 (FORMALIZING INFORMAL AXIOM 5)

'P(N).

Given this final axiom, Godel shows that if (some) God exists, that

existence cannot be contingent. An informal sketch of the proof was

given in Section 6 of Chapter 10, and it can be turned into a formal

proof-see Informal Propositions 7 and 8. I will leave the details as

exercises, since you have seen lots of worked out tableaus now. Here is

a proper statement of Godel's result, with all the assumptions explicitly

stated. Nate that the necessary actualist existence of God follows from

His possibilist existence.

THEOREM 11.26 (FORMALIZING INFORMAL PROPOSITION 7)

Assume Axioms 11.3B, 11.11, and 11.25. In the logic K,

(3x)G(x) ::) D(3Ex)G(x).

I leave it to you to prove this, using the informal sketch as a guide.

Now Godel's argument can be completed.

THEOREM 11.27 (FORMALIZING INFORMAL PROPOSITION 8)

Assume Axioms 11.3B, 11.11, and 11.25. In the logic 85,

0(3x)G(x) ::) D(3Ex)G(x).

Proof From Theorem 11.26,

(3x)G(x) ::) D(3Ex)G(x).

By necessitation,

D[(3x)G(x) ::) D(3Ex)G(x)].

By the K validity D(A::) B)::) (OA::) OB),

0(3x)G(x) ::) OD(3Ex)G(x).

Finally, in 85, ODA::) DA, so we conclude

0(3x)G(x) ::) D(3Ex)G(x) .

Now we are at the end of the argument.

COROLLARY 11.28 Assume all the Axioms. In the logic 85,

D(3Ex)G(x).

Proof By Theorems 11.27 and 11.17.

162 TYPES, TABLEAUS, AND GODEL'S GOD

Exercises

EXERCISE 6.1 Give a tableau proof to show Theorem 11.26. Use various

earlier results as assumptions in the tableau.

7. Going Further

Godel's axioms admit more consequences than just those of the onto-

logical argument. In this section a few of them are presented.

7.1 Monotheism

Does there exist exactly one God? The following says "yes." You are

asked to prove it, as Exercise 7.1.

PROPOSITION 11.29 (3x)('v'y)[G(y) =: (y = x)j.

This Proposition has a curious Corollary. Since type-0 quantification

is possibilist, it makes sense to ask if there are gods that happen to be

non-existent. But Corollary 11.28 tells us there is an existent God, and

the Proposition above tells us it is the only one God, existent or not.

Consequently we have the following.

COROLLARY 11.30 ('v'x)[G(x) :J E(x)j.

Proposition 11.29 tells us we can apply the machinery of definite de-

scriptions. By Definition 9.12, 1x.(W)[P(Y) :J Y(x)] always designates,

and consequently so does 1x.G(x). Proposition 9.14 tells us this will be

a rigid designator provided G(x) is stable. It follows from Sobel's argu-

ment in Section 8 that it, and everything else, is. But alternative versions

of Godel's axioms have been proposed-! will discuss some below-and

using them the stability of G(x) does not seem to be the case. That

is, it seems to be compatible with the axioms of Godel (as modified by

others) that, while the existence of God is necessary, a particular being

that is God need not be God necessarily. If this is not the case, and a

proof has been missed, I invite the reader to correct the situation.

7.2 Positive Properties are Necessarily

Instantiated

Proposition 11.8 says that positive properties are possibly instanti-

ated. In [Sob87], it is observed that a consequence of Corollary 11.28 is

that every positive property is necessarily instantiated.

PROPOSITION 11.31 ('v'X){'P(X) :J 0(3Ex)X(x)}.

I leave the easy proof of this to you.

GODEL'S ARGUMENT, FORMALLY 163

Exercises

EXERCISE 7.1 Give a tableau proof for Proposition 11.29. Hint: you

will need Corollary 11.28, Theorem 11.20, and Theorem 11.23.

EXERCISE 7.2 Provide a tableau proof for Proposition 11.31. Hint: by

Corollary 11.28, a Godlike being necessarily exists. Such a being has

all positive properties, so every positive property is instantiated. Now,

build this into a tableau.

8. More Objections

In Section 4 we saw that one of Godel's Axioms was equivalent to

the possible existence of God. Other objections have been raised that

are equally as serious. Chapter 4 of [SobOl] discusses problems with

Axiom 11.25, that necessary existence is positive. I do not take this

point up here. But also in [SobOl], and earlier in [Sob87], it was argued

that Godel's axiom system is so strong it implies that whatever is the

case is so of necessity, Q ::J DQ. In other words, the modal system

collapses. In still other, more controversial, words, there is no free will.

Roughly speaking, the idea of Sobel's proof is this. God, having all

positive properties, must possess the property of having any given truth

be the case. Since God's existence is necessary, anything that is a truth

must necessarily be a truth.

Here is a version of the argument given by Sobel. For simplicity,

assume Q is a formula that contains no free variables. By Theorem 11.20,

(\fx)[G(x) ::J ( G, x)]. (11.1)

Using the definition of, we have as a consequence

(\fx){G(x) ::J ('v'Z){Z(x) ::J D('v'Ew)[G(w) ::J Z(w)]}}. (11.2)

There is a minor nuisance to deal with. In the formula (11.2) I would

like to instantiate the quantifier ('v'Z) with Q, but this is not a 'legal'

term, so instead I use the term (>..y.Q) to instantiate. In it, y is of type

0, and so (>..y.Q) is of type i(O). We get the following consequence.

(\fx){G(x) ::J {(>..y.Q)(x) ::J D('v'Ew)[G(w) ::J (>..y.Q)(w)]}}. (11.3)

Now to undo the technicality just introduced, note that since y does not

occur free in Q, (>..y.Q)(x) = (>..y.Q)(w) = Q, and so we have

(\fx){G(x) ::J {Q ::J D('v'Ew)[G(w) ::J Q]}}. (11.4)

164 TYPES, TABLEAUS, AND GODEL'S GOD

Since x does not occur free in the consequent, (11.4) is equivalent to the

following:

(3x)G(x) :::> {Q :::> D('v'Ew)(G(w) :::> Q)}. (11.5)

We have Corollary 11.28, from which

(3x)G(x) (11.6)

follows. Then from (11.5) and (11.6) we have

Q :::> D('v'Ew) ( G( w) :::> Q). (11.7)

Since Q has no free variables, (11.7) is equivalent to the following:

Q :::> D[(3Ew)G(w) :::> Q]. (11.8)

Using the distributivity of D over implication, (11.8) gives us

Q :::> [D(3Ew)G(w) :::> DQ]. (11.9)

Finally (11.9), and Corollary 11.28 again, give the intended result,

Q :::> DQ. (11.10)

Most people have taken this as a counter to Godel's argument-if the

axioms are strong enough to admit such a consequence, something is

wrong. In the next two sections I explore some ways out of the difficulty.

9. A Solution

Sobel's demonstration that the Godel axioms imply no free will rather

takes the fun out of things. In this section I propose one solution to the

problem. I don't profess to understand its implications fully. I am

presenting it to the reader, hoping for comments and insights in return.

Throughout, it has been assumed that Godel had in mind inten-

sional properties when talking about positiveness and essence. But,

suppose not-suppose extensional properties were intended. I reformu-

late Godel's argument under this alternative interpretation. It is one

way of solving the problem Sobel raised.

GODEL'S ARGUMENT, FORMALLY 165

In this section only I will take P to be a constant symbol of type

i((O)). Axiom 11.5 gets replaced with the following.

Revised Axiom 11.5 In the formula below, xis of type 0, and X and

Y are of type (0).

(VX)(VY){[P(X) A D(V'Ex)(X(x) :J Y(x))] :J P(Y)}

Note that this has the same form as Axiom 11.5, but the types of

variables X and Y are now extensional rather than intensional. This

will be the general pattern for changes. The definition of negative, for

instance, is modified as follows. For a term T of type (0), take T as

short for l(.Xx.T(x)). Then Axioms 11.3A and 11.3B, 11.10, and 11.11,

all have their original form, but with variables changed from intensional

to extensional type.

The analog of Proposition 11.8 still holds, but with extensional vari-

ables involved.

(VX){P(X) :J 0(3Ex)X(x)}

Analogs of G and G* are defined in the expected way. G is the fol-

lowing type j(O) term, where Y is type (0) and, as noted before, Pis of

type j((O)).

(.Xx.(VY)[P(Y) :J Y(x)])

Likewise G* is the type i(O) term

(.Xx.(W)[P(Y) = Y(x)]).

One can still prove (Vx)[G(x) = G*(x)].

Essence must be redefined, but again it is only variable types that are

changed. now abbreviates the following type j( (0), 0) term, in which

Z is of type (0) and w is of type 0:

(.XY, x.Y(x) A (VZ){Z(x) :J D(V'Ew)[Y(w) :J Z(w)]})

Theorem 11.21 plays an essential role in the Godel proof, and it too

continues to hold, in a slightly modified form:

(Vx)[G*(x) :J (1G*, x)] .

.

I leave the proof of this to you-it is similar to the earlier one.

Of course we must modify the definition of Necessary Existence, to use

the revised version of essence, and Axiom 11.25 as well, to use the mod-

ified definition of Necessary Existence. For this section, N abbreviates

the following type j(O) term, in which Y is of type (0):

(>.x.('VY)[E(Y, x) :J D(3Ez)Y(z)).

166 TYPES, TABLEAUS, AND GODEL'S GOD

Revised Axiom 11.25 is P(lN), where N is as just modified.

With this established, the rest of Godel's argument carries over di-

rectly, giving us the following.

0(:3E z) {tG*) (z)

The final step is the easy proof that this implies the desired 0(:3Ez)G*(z),

and hence O(::JEz)G(z), and I leave this to you.

So, we have the conclusion of Godel's argument. Finally, here is a

model, adapted from [And90], that shows Sobel's continuation no longer

applies.

EXAMPLE 11.32 Construct a standard S5 model as follows. There are

two possible worlds, call them r and D.. The accessibility relation always

holds. The type-0 domain is the set {a, b}. Since this is a standard

model, the remaining types are fully determined.

The existence predicate, E, is interpreted to have extension {a, b} at

r and {a} at D.. Informally, all type-0 objects exist at r, but only a

exists at D..

Call a type-(0) object positive if it applies to a. Interpret P so that

at each world its extension is the collection of positive type-(0) objects;

that is, at each world P designates {{a}, {a, b}}.

This finishes the definition of the model. I leave the following facts

about it for you to verify.

1 The designation of G in this model is rigid, with {a} as its extension

at both worlds.

2 The designation of is also rigid, with extension { ( {a}, a), ( { b}, b)}

at each world. Loosely, the essence of a is {a} and the essence of b is

{b}.

3 The designation of N is also rigid, with extension {a} at each world.

4 All the Axioms are valid, as modified in this section.

Now take Q to be the closed formula (::JEx)(::JEy)-,(x = y). Since it

asserts two objects actually exist, it is true at r, but not at D., and hence

Q :J OQ is not true at r.

We now know that Sobel's argument must break down in the present

system, but it is instructive to try to reproduce the earlier proof, and see

just where things go wrong. The attempted argument takes on a rather

formidable appearance--you might want to skip to the last paragraph

and read the conclusion, before going through the details.

GODEL'S ARGUMENT, FORMALLY 167

We try to prove Q ::) DQ, starting more or less as we did before.

(Vx)[G(i) ::) (lG, x)] (11.11)

which, unabbreviated, is

(Vx)[G(x) ::)

(.XY,x.Y(x) 1\ (VZ){Z(x)::) D(VEw)[Y(w)::) Z(w)]})(1G,x)] (

1

1.

12

)

where Y and Z are of type (0), unlike in (11.2) where they were of type

j(O).

The variable xis of type 0, and it is easy to show the following simpler

formula is a consequence of (11.12).

(\ix)[G(x) ::)

(.XY.Y(x) 1\ (VZ){Z(x) ::) D(VEw)[Y(w) ::) Z(w)]})(1G)]

(11.13)

From this we trivially get the following.

(Vx)[G(x) ::)

(.XY.(VZ){Z(x)::) D(VEw)[Y(w)::) Z(w)]})(1G)]

(11.14)

Next, in the argument of Section 8, we instantiated the quantifier (V Z)

with the term (.Xy.Q). Of course we cannot do that now, since (.Xy.Q)

is an intensional term, while the present quantifier (V Z) is extensional.

Apply the extension-of operator, getting 1(-Xy.Q), and use this instead.

But universal instantiation involving relativized terms is a little tricky.

If 1 T is a relativized term of the same type as Z, (VZ)cp(Z) ::) cp(l T)

is not generally valid. What is valid is ('v'Z)cp(Z) ::) (.XZ.cp(Z))(1T). So

what we get from formula (11.14) when we instantiate the quantifier is

the following consequence.

(Vx)[G(x) ::)

(.XY, Z.Z(x) ::) D(VEw)[Y(w) ::) Z(w)])(1G, 1(-Xy.Q) )]

(11.15)

Distributing the abstraction, this is equivalent to the following.

(Vx ){ G(x) ::)

[(.XY, Z.Z(x))(lG, 1(-Xy.Q))::) (11.16)

(.XY, Z.D('v'Ew)(Y(w)::) Z(w)))(lG, 1(-Xy.Q))]}

168 TYPES, TABLEAUS, AND GODEL'S GOD

The variable x does not occur free in (>.y.Q) and Y does not occur

in Z(x), so (>.Y,Z.Z(x))(lG,l(>.y.Q)) is simply equivalent to Q, and

(11.16) reduces to the following.

('v'x){G(x) :J

[Q :J (>.Y, Z.D('v'Ew)(Y(w) :J Z(w)))(lG, l(>.y.Q) )]}

(11.17)

From this we get

(3x)G(x) :J

[Q :J (>.Y, Z.D('v'Ew)(Y(w) :J Z(w)))(lG, l(>.y.Q) )]

(11.18)

and since we have (3x)G(x), we also have

Q :J (>.Y, Z.D('v'Ew)(Y(w) :J Z(w)))(lG, l(>.y.Q) ). (11.19)

Since Q has no free variables, (11.19) can be shown to be equivalent to

the following (where a constant symbol a has been introduced to keep

formula formation correct).

Q :J (>.Y, Z.D((3Ew)Y(w) :J Z(a)))(lG, l(>.y.Q)).

Using the distributivity of 0 over implication, (11.20) gives us

Q :J (>.Y, Z.D(3Ew)Y(w) :J DZ(a))(lG, l(>.y.Q)).

From (11.21) we get

Q :J[(>.Y, Z.D(3Ew)Y(w))(lG, l(>.y.Q)) :J

(>.Y, Z.DZ(a))(lG, l(>.y.Q) )].

(11.20)

(11.21)

(11.22)

Because Z has no free occurrences in D(3Ew)Y(w) and Y has no free

occurrences in Z(a), (11.22) can be simplified to

Q :J[(>.YD(3Ew)Y(w))(lG) :J

(>.Z.DZ(a))(l(>.y.Q) )].

(11.23)

I don't know the status of (>.Y.D(3Ew)Y(w))(lG), that is, whether or

not it follows from the axioms used in this section. It does hold provided

GODEL'S ARGUMENT, FORMALLY 169

G is rigid, so in particular, it holds in the model of Example 11.32.

Consequently, in settings like that model (11.23) reduces to the following.

Q:) (>.Z.DZ(a))(l(>.y.Q)). (11.24)

But (>.Z.DZ(a))(l (.>.y.Q)) is not equivalent to DQ, and that's an

end of it. Expressing the essential idea of (.>.Z.DZ(a))(l(.>.y.Q)) with

somewhat informal notation, we might write it as (>.Z.DZ)(lQ), and so

what has been established, assuming rigidity of G, is

Q :) (.>.Z.DZ)(lQ) (11.25)

and this is quite different from Q :) DQ. In the abstract, the variable

Z is assigned the current version of Q-its truth value in the present

world. Perhaps an example will make clear what is happening. Suppose

it is the case, in the real world, that it is raining-take this as Q. If

we had validity of Q :) DQ, it would necessarily be raining-DQ-and

so in every alternative world, it would be raining. But what we have is

Q :) (.>.Z.DZ)(lQ), and since Q is assumed to hold in the real world, we

conclude (>.Z.DZ)(lQ). This conclusion asserts something more like: if

it is raining in the real world, then in every alternative world it is true

that it is raining in the real world. As it happens, this is trivially correct,

and says nothing about whether or not it is raining in any alternative

world.

10. Anderson's Alternative

One solution to the objection Sobel raised has been presented. In

[And90] some different, quite reasonable, modifications to the Godel

axioms are proposed that also manage to avoid Sobel's conclusion. For

this section I return to the use of intensional variables.

Axiom 11.3B is something of a problem. Essentially it says, everything

must be either positive or negative. As Anderson observes, why can't

something be indifferent? Anderson drops Axiom 11.3B.

The most fundamental change, however, is elsewhere. Definition 11.12

and its alternative, Definition 11.13, are discarded. Instead there is a

requirement that a Godlike being have positive properties necessarily.

DEFINITION 11.33 (GODLIKE, ANDERSON VERSION) GA is the type j

(0) term

(.Xx.(W)[P(Y) := DY(x)]).

There is a corresponding change in the definitions of essence and nec-

essary existence. Definition 11.19 gets replaced by the following

170 TYPES, TABLEAUS, AND GO DEL'S GOD

DEFINITION 11.34 (ESSENCE, ANDERSON VERSION) [A abbreviates

the following type i(i(O), 0) term:

(.XY, x.(VZ){DZ(x) =: D(VEw)[Y(w) :J Z(w)]})

Notice several key things about this definition. The Scott addition,

that the essence of an object actually apply to the object, is dropped.

A necessity operator has been introduced that was not present in the

definition of . And finally, an implication in the definition of has

been replaced by an equivalence.

The definition of necessary existence, Definition 11.24, is replaced by

a version of the same form, except that Anderson's definition of essence

is used in place of that of Godel.

DEFINITION 11.35 (NECESSARY EXISTENCE, ANDERSON VERSION)

NA abbreviates the following type j(O) term:

(.Xx.(VY)[A(Y, x) :J D(3Ez)Y(z)).

Now, what happens to earlier reasoning? Of course Proposition 11.8

still holds, since Axioms 11.3A and 11.5 remain unaffected. Theo-

rem 11.20 turns into the following.

THEOREM 11.36 In S5 the following is provable.

(Vx)[GA(x) :J eA(GA,x)].

I leave it to you to verify the theorem, using tableaus say.

Next, Anderson replaces Axiom 11.25 with a corresponding version

asserting that his modification of necessary existence is positive.

AXIOM 11.37 (ANDERSON'S VERSION OF 11.25) P(NA).

Now Theorem 11.26 turns into the following.

THEOREM 11.38 Assume Axioms 11.11 and 11.37. In the logic S5,

(3x)GA(x) :J 0(3Ex)GA(x).

Once again, I leave the proof to you. These are the main items. The

rest of the ontological argument goes through as before. At the end, we

have the following.

THEOREM 11.39 Assume all the Axioms 11.3A, 11.5, 11.10, 11.11, and

11.37. In the logic S5,

GODEL'S ARGUMENT, FORMALLY 171

Thus the desired necessary existence follows, and with one fewer axiom

(though with more complex definitions). And a model, closely related to

the one given in the previous section, can be constructed to show that

these axioms do not yield Sobel's undesirable conclusion-see [And90]

for details.

Exercises

EXERCISE 10.1 Supply a tableau argument for Theorem 11.36. Do the

same for Theorem 11.38.

11. Conclusion

Godel's proof, and criticisms of it, have inspired interesting work.

Some was mentioned above. More remains to be done. Here I briefly

summarize some directions that might profitably be explored.

[Haj95] studies the role of the comprehension axioms-work that is

summarized in [Haj96b]. Completely general comprehension axioms are

implicit in my presentation, they are present as the assumption that

every abstract has a meaning. Hajek confines things to a second-order

intensional logic, augmented with one third-order constant to handle

positiveness. In this setting Hajek introduces what he calls a cautious

comprehension schema:

('v'x)[G(x) :J (D<I>(x) V 0-.<I>(x))] :J (:3Y)D('v'x)[Y(x) = <I>(x)].

Hajek shows that Godel's axioms do not lead to a proof of Q :J DQ,

provided cautious comprehension replaces full comprehension, but the

necessary existence of God still can be concluded.

Hajek refutes a claim by Magari, [Mag88], that a subset of Godel's

axiom system is sufficient for the ontological argument. But he also

shows Magari's claim does apply to Anderson's system. And he shows

that Godel's axioms, with cautious comprehension, can be interpreted

in Anderson's system, with full comprehension.

The results of Hajek assume an underlying model with constant do-

mains but no existence predicate, and only intensional properties. It is

not clear what happens if these assumptions are modified.

In Section 7, some further consequences of Godel's axioms were dis-

cussed. I don't know what happens to these when the axioms are modi-

fied in the various ways suggested here and in the previous two sections.

Nor do I know the relationships, if any, between the extensional-property

approach I suggested, and Anderson's version.

Finally, and most entertainingly, I refer you to an examination of

ontological arguments and counter-arguments in the form of a series of

172 TYPES, TABLEAUS, AND GO DEL'S GOD

puzzles, in [Smu83], Chapter 10. You should find this fun, and a bit of

a relief after the rather heavy going of the book you just finished.

References

(Ada95]

(AG96]

(And71]

[And72]

[And86]

[And90]

(Bre72]

(Car 56]

(Cha79]

[Chu40]

Note to reader: At the end of each bibliogra-

phy item is a list of the pages on which there

is a reference to the item.

Robert Merrihew Adams. Introductory note to *1970. In Feferman

et al. (FJWDG+95], pages 388-402. pages 138

C. Anthony Anderson and Michael Gettings. Gooel's ontological

proof revisited. pages 167-172, 1996. In [Haj96a]. pages 139, 148

Peter B. Andrews. Resolution in type theory. Journal of Symbolic

Logic, 36(3):414-432, 1971. pages 42, 47

Peter B. Andrews. General models and extensionality. Journal of

Symbolic Logic, 37(2):395-397, 1972. pages 25

Peter B. Andrews. An Introduction to Mathematical Logic and Type

Theory: To Truth Through Proof Academic Press, Orlando, Florida,

1986. pages xii

C. Anthony Anderson. Some emendations of Gooel's ontological

proof. Faith and philosophy, 7:291-303, 1990. pages 139, 166, 169,

171

Aldo Bressan. A General Interpreted Modal Calculus. Yale Univer-

sity Press, 1972. pages 96, 102

Rudolph Carnap. Meaning and Necessity. University of Chicago

Press, Chicago, 2 edition, 1956. pages xv

M. J. Charlesworth. St. Anselm's Proslogion. University of Notre

Dame Press, Notre Dame, IN, 1979. First edition, Oxford University

Press, 1965. pages 134, 134

Alonzo Church. A formulation of the simple theory of types. Journal

of Symbolic Logic, 5:56-68, 1940. pages 4

173

174

[Coc69]

[Des51]

[Fit83]

[Fit93]

[Fit96]

[Fit98]

[FitOOa]

[FitOOb]

[FJWDG+95]

[FM98]

[G70]

[Gal75]

[GHR94]

TYPES, TABLEAUS, AND GODEL'S GOD

Nino B. Cocchiarella. A completeness theorem in second order

modal logic. Theoria, 35:81-103, 1969. pages 90, 96, 139

Rene Descartes. A Discourse on Method and Selected Writings.

Everyman's Library. E. P. Dutton, New York, 1951. Translation by

John Veitch. pages 135, 135

Melvin C. Fitting. Proof Methods for Modal and Intuitionistic Log-

ics. D. Reidel Publishing Co., Dordrecht, 1983. pages 95, 102,

110

Melvin C. Fitting. Basic modal logic. pages 368-448. 1993. In

[GHR94] Volume 1. pages 95, 110

Melvin C. Fitting. First-Order Logic and Automated Theorem Prov-

ing. Springer-Verlag, 1996. First edition, 1990. pages 8, 10, 33, 37,

41, 47

Melvin C. Fitting. Higher-order modal logic-a sketch. In Ricardo

Caferra and Gernot Salzer, editors, Automated Deduction in Classi-

cal and Non-Classical Logics, pages 23-38. Springer Lecture Notes

in Artificial Intelligence, 1761, 1998. pages 96

Melvin C. Fitting. Databases and higher types. In John Lloyd,

Veronica Dahl, Ulrich Furbach, Manfred Kerber, Kung-Kiu Lau,

Catuscia Palamidessi, Luis Moniz Pereira, Yehoshua Sagiv, and Pe-

ter Stuckey, editors, Computational Logic-CL 2000, pages 41-52.

Springer Lecture Notes in Artificial Intelligence, 1861, 2000. pages

xiii

Melvin C. Fitting. Modality and databases. In Roy Dyckhoff, editor,

Automated Reasoning with Analytic Tableaux and Related Methods,

pages 19-39. Springer Lecture Notes in Artificial Intelligence, 1847,

2000. pages xiii, 96

Solomon Feferman, Jr. John W. Dawson, Warren Goldfarb, Charles

Parsons, and Robert N. Solovay, editors. Kurt Godel Collected

Works, Volume III, Unpublished Essays and Lectures. Oxford Uni-

versity Press, New York, 1995. pages 173, 174

Melvin C. Fitting and Richard Mendelsohn. First-Order Modal

Logic. Kluwer, 1998. Paperback, 1999. pages xiv, 4, 89, 96, 101,

102, 102, 102, 105, 108, 110, 119, 122, 122, 125

Kurt Gooel. Ontological proof. In Feferman et al. [FJWDG+95],

pages 403-404. pages 139, 145

D. Gallin. Intensional and Higher-Order Modal Logic. North-

Holland, 1975. pages xv, 86, 102, 102

Dov M. Gabbay, C. J. Hogger, and J. A. Robinson, editors. Hand-

book of Logic in Artificial Intelligence and Logic Programming. Ox-

ford University Press, 1993-94. pages 17 4, 176

REFERENCES 175

[Gil99] Paul C. Gilmore. Partial functions in an impredicative simple the-

ory of types. In Neil V. Murray, editor,. Automated Reasoning

with Analytic Tableaux and Related Methods, number 1617 in Lec-

ture Notes in Artificial Intelligence, pages 186--201, Berlin, 1999.

Springer. pages 4

[Gil01] Paul C. Gilmore. An intensional type theory: motivation and cut-

elimination. To appear, Journal of Symbolic Logic, 2001. pages 4,

33

[Gol81] Warren D. Goldfarb. The undecidability of the second-order uni-

fication problem. Theoretical Computer Science, 13:225-230, 1981.

pages 42

[Haj95] Petr Hajek. Der Mathematiker und die Frage der Existenz Gottes.

To appear in Wahrheit und Beweisbarkeit- Leben und Werk von K.

Godel, Buld, Kohler, and Schimanowicz., 1995. pages 171

[Haj96a] Petr Hajek, editor. Godel '96, Berlin, 1996. Springer. pages 173

[Haj96b] Petr Hajek. Magari and others on Gooel's ontological proof. In

Ursini and Agliano, editors, Logic and Algebra, pages 125-136. Mar-

cel Dekker, Inc., 1996. pages 139, 171

[Har65] Charles Hartshorne. Anselm's Discovery: A Re-Examination of the

Ontological Proof for God's Existence. Open Court, LaSalle, IL,

1965. pages 134

[Haz98] A. P. Hazen. On Gooel's ontological proof. Australasian Journal of

Philosophy, 76(3):361-377, 1998. pages 156

[HC96] G. E. Hughes and M. J. Cresswell. A New Introduction to Modal

Logic. Routledge, London and New York, 1996. pages 38, 89, 102

[Hen50] Leon Henkin. Completeness in the theory of types. Journal of

Symbolic Logic, 15:81-91, 1950. pages 19, 25

[Hue72] Gerard P. Huet. Constrained Resolution: A Complete Method for

Higher Order Logic. PhD thesis, Case Western Reserve University,

1972. pages 42

[Hue73] Gerard P. Huet. The undecidability of unification in third-order

logic. Information and Control, 22:257-267, 1973. pages 42

[Hue75] Gerard P. Huet. A unification algorithm for typed >.-calculus. The-

oretical Computer Science, 1:27-57, 1975. pages 42

[Koh95] Michael Kohlhase. Higher-order tableaux. In Peter Baumgartner,

Reiner Hiihnle, and Joachim Posegga, editors, Theorem Proving

with Analytic Tableaux and Related Methods, volume 918 of Lec-

ture Notes in Artificial Intelligence, pages 294-309, Berlin, 1995.

Springer-Verlag. pages xii, 33, 42

176

[Kri80]

(Lei 56]

[Lei94]

(Mag88]

[Man96]

(Mon60]

(Mon68]

(Mon70]

(Opp95]

[Opp96a]

[Opp96b]

(Opp01]

(Pla65]

(Pra68]

[Rob77]

TYPES, TABLEAUS, AND GODEL'S GOD

Saul Kripke. Naming and Necessity. Harvard University Press, 1980.

pages 121

Gottfried Wilhelm Leibniz. Philosophical Papers and Letters. D.

Reidel, Dordrecht-Holland, 2 edition, 1956. Translated and Edited

by Leroy E. Loemker. pages 137, 138

Daniel Leivant. Higher order logic. pages 229-321. 1994. In [GHR94]

Volume 2. pages xii

R. Magari. Logica e teofilia. Notizie di logica, VII(4), 1988. pages

171

Maria Manzano. Extensions of First Order Logic. Cambridge Uni-

versity Press, Cambridge, UK, 1996. pages xii

Richard Montague. On the nature of certain philosophical enti-

ties. The Monist, 53:159-194, 1960. Reprinted in (Tho74], 148-187.

pages xv, 102

Richard Montague. Pragmatics. pages 102-122. 1968. In Contempo-

rary Philosophy: A Survey, R. Klibansky editor, Florence, La Nuova

Italia Editrice, 1968. Reprinted in [Tho74], 95-118. pages xv, 102

Richard Montague. Pragmatics and intensional logic. Synthese,

22:68-94, 1970. Reprinted in [Tho74], 119-147. pages xv, 102

Graham Oppy. Ontological Arguments and Belief in God. Cam-

bridge University Press, Cambridge, UK, 1995. pages 134

Graham Oppy. Gooelian ontological arguments. Analysis,

56(4):226-230, October 1996. pages 176

Graham Oppy. Ontological arguments. The Stanford Encyclopedia

of Philosophy, Edward N. Zalta (ed.), March 13 1996. Available on

the internet. URL = http:/ /plato.stanford.edu/entries/ontological-

arguments/. pages 134, 139

Graham Oppy. Gooelian ontological arguments. Available on the

Internet. URL =

http://www .infidels.org/library /modern/ graham_oppy / godel.ht,

2001. Originally published as (Opp96a]. pages 139

Alvin Plantinga. The ontological argument, from St. Anselm to con-

temporary philosophers. Anchor-Doubleday, Garden City, NY, 1965.

pages 134

Dag Prawitz. Hauptsatz for higher order logic. Journal of Symbolic

Logic, 33:452-457, 1968. pages xii, 35, 43, 46, 47, 67

The Gospel of Philip. In James M. Robinson, editor, The Nag Ham-

madi Library in English. Harper and Row, New York, 1977. pages

ix

REFERENCES 177

[Rus05] Bertrand Russell. On denoting. Mind, 14:479-493, 1905. Reprinted

in Robert C. Marsh, ed., Logic and Knowledge: Essays 1901-1950,

by Bertmnd Russell, Allen & Unwin, London, 1956. pages 125

[Sch60] Kurt Schutte. Syntactical and semantical properties of simple type

theory. Journal of Symbolic Logic, 25(4):305-326, December 1960.

pages 4

[Sha91] Stewart Shapiro. Foundations Without Foundationalism: A Case

for Second-Order Logic. Number 17 in Oxford Logic Guides. Oxford

University Press, Oxford, UK, 1991. pages xii

[SmaOl]

[Smi93]

[Smu68]

[Smu83]

(Sob87]

(SobOl]

[Tak67]

[Tar 55]

[Tho74]

(Tho87]

(Tol75]

Christopher Small. Kurt Gooel's ontological argument. Available

on the internet. URL =

http:/ /www.stats.uwaterloo.ca/ "-'Cgsmall/ontology.html, 2001.

pages 139

Perry Smith. Higher-order logic, model theory, recursion theory,

and proof theory. Unpublished manuscript, 1993. pages 33

Raymond M. Smullyan. First-Order Logic. Springer-Verlag, Berlin,

1968. Revised Edition, Dover Press, New York, 1994. pages 33, 37,

47

Raymond M. Smullyan. 5000 B. C. and Other Philosophical Fan-

tasies. St. Martin's Press, New York, 1983. pages 172

Jordan Howard Sobel. Gooel's ontological proof. pages 241-261.

1987. In [Tho87]. pages 138, 138, 138, 139, 162, 163

Jordan Howard Sobel. Logic and Theism. Available on the internet.

URL = www.scar.utoronto.ca/"-'sobel/, 2001. pages 134, 139, 139,

152, 163, 163

Moto-o Takahashi. A proof of cut-elimination theorem in simple

type theory. J. Math. Soc. Japan, 19:399-410, 1967. pages xii, 35,

43, 46, 47, 67

Alfred Tarski. A lattice-theoretical theorem and its applications.

Pacific Journal of Mathematics, 5:285-309, 1955. pages 41

Richmond H. Thomason, editor. Formal Philosophy, Selected Pa-

pers of Richard Montague. Yale University Press, New Haven and

London, 1974. pages 176, 176, 176

Judith Jarvis Thomson. On Being and Saying: Essays for Richard

Cartwright. MIT Press, Cambridge, MA, 1987. pages 177

Sue Toledo. Tableau Systems for First Order Number Theory and

Certain Higher Order Theories. Number 447 in Lecture Notes in

Mathematics. Springer-Verlag, Berlin, Heidelberg, NY, 1975. pages

xii, 33, 33, 35, 46

Index

abstraction designation function, 21, 26,

103

proper, 22, 27

accessibility, 91

Anderson, C. A., 169-171

Anselm, 134

cautious comprehension, 171

character, 156

choice axiom, 129

choice function, 128-130

closed, 37, 110

compact, 15, 66

complete, 46, 73

strong, 16

weak, 16

composition, 10

comprehension axiom, 3

concept, 84

consequence, 14, 28, 95

consistent, 63

maximal, 63

constant domain, 89

constant symbol, 5, 87

continuum hypothesis, 17

cut rule, 67

cut-elimination, 66

de dicto, 118-121

de re, 118-121

Dedekind, R., 15

defined at, 125

definite description, 125-128

degree, 8

Descartes, R., 134-136, 152, 156

description designation, 125

designates, 125

domain, 91

domain function

Henkin, 20, 103

179

E-complete, 63

entity, 48, 49

equality, 115

equality axioms, 69, 115

essence, 141, 142, 156, 170

evaluation at a prefix, 109

existential relativization, 147

extensional object, 84, 91

extensionality, 117

assumptions, 118

axioms, 77

for extensional terms, 118

for intensional terms, 117

finite support, 8

formula, 6

modal, 88

prefixed, 106

truth, 13, 22, 26, 93, 104

frame

augmented, 91

extensional, 30

Henkin, 20

generalized, 25

relative generalized, 50

Henkin/Kripke, 103

Kripke, 91

free, 9

free variable, 6

Gaunilo, 134

global assumption, 95, 111

Gi:idel, K., 138-143, 145, 147, 148, 150,

152, 156, 158, 162-164, 166,

171

grounded, 34, 106

Hajek, P., 171

Henkin domain

180

relative, 50

Hintikka set, 47

impredicativity, 4

inconsistent, 63

intensional object, 84, 91

interpretation, 11, 30, 51, 73, 92, 103

allowed, 50

K, 105

L(C), 5

.X abstraction, 3

Leibniz, G., 137-140, 145, 148, 152

Lindstrom, P., 68

local assumption, 95, 111

Liiwenheim-Skolem, 66, 68

Magari, R., 171

model

classical, 12

extensional, 30

general, 19

generalized Henkin, 28, 104

Henkin, 19, 22, 23

Henkin/Kripke, 104

modal, 91

standard, 24

monotheism, 162

necessary existence, 142, 160, 170

negative, 141, 146

non-rigid, 102

normal, 25

order, 5

parameter, 34, 108

perfection, 137-139

positive, 138-142, 145, 146, 162

possible value, 49

possible world, 91

predicate abstract, 5

predicate abstraction, 3

prefix, 105

pseudo-model, 47, 48, 51

quantification

actualist, 89, 91

possibilist, 89, 91

rigid, 121-124

rule

abstract, 37, 109

branch extension, 35

conjunctive, 35, 107

derived

TYPES, TABLEAUS, AND GO DEL'S GOD

closure, 113

extensionality, 77

intensional predication, 113

reflexivity, 70

subscripted abstract, 114

substitutivity, 70

unsubscripted abstract, 114

disjunctive, 36, 107

double negation, 35, 107

existential, 36, 108

extensional, 118

extensional predication, 110

intensional predication, 109

necessity, 108

possibility, 107, 108

reflexivity, 115

substitutivity, 115

universal, 36, 109

world shift, 110

Russell, B., 125, 126, 136

85, 105

satisfiability, 14, 28

Scott, D., 138, 152, 156, 158

sentence, 6

Sobel, J. H., 163, 164, 166, 171

sound, 43, 46, 73

stability, 124-125

substitution, 8

free, 9

tableau, 33

basic, 35

derivation, 37, 111

prefixed, 105

proof, 37, 110

satisfiable, 43

term, 6, 87

denotation, 12, 21, 26

designation, 93, 103

relativized, 108

type, 4, 86

extensional, 86

Gallin/Montague, 102

intensional, 86

relation, 11

validity, 14, 28, 94

valuation, 12, 20, 26, 92

variable, 5

variant, 12

varying domain, 89

Wittgenstein, L., 139

world independent, 110

Zermelo-Fraenkel set theory, 17

TRENDS IN LOGIC

1. G. Schurz: The Is-Ought Problem. An Investigation in Philosophical Logic. 1997

ISBN 0-7923-4410-3

2. E. Ejerhed and S. Lindstrom (eds.): Logic, Action and Cognition. Essays in Philo-

sophical Logic. 1997 ISBN 0-7923-4560-6

3. H. Wansing: Displaying Modal Logic. 1998 ISBN 0-7923-5205-X

4. P. Hajek: Metamathematics of Fuzzy Logic. 1998 ISBN 0-7923-5238-6

5. H.J. Ohlbach and U. Reyle (eds.): Logic, Language and Reasoning. Essays in Honour

ofDov Gabbay. 1999 ISBN 0-7923-5687-X

6. K. Dosen: Cut Elimination in Categories. 2000 ISBN 0-7923-5720-5

7. R.L.O. Cignoli, I.M.L. D' Ottaviano and D. Mundici: Algebraic Foundations of many-

valued Reasoning. 2000 ISBN 0-7923-6009-5

8. E.P. Klement, R. Mesiar and E. Pap: Triangular Norms. 2000

ISBN 0-7923-6416-3

9. V.F. Hendricks: The Convergence of Scientific Knowledge. A View From the Limit.

2001 ISBN 0-7923-6929-7

10. J. Czelakowski: Protoalgebraic Logics. 2001 ISBN 0-7923-6940-8

11. G. Gerla: Fuzzy Logic. Mathematical Tools for Approximate Reasoning. 2001

ISBN 0-7923-6941-6

12. M. Fitting: Types, Tableaus, and Godel's God. 2002 ISBN 1-4020-0604-7

KLUWER ACADEMIC PUBLISHERS - DORDRECHT I BOSTON I LONDON

- A Proof Of GodTransféré parMarsha Pine
- wnewsletter april 1-5 2019 2 autosaved autosavedTransféré parapi-329583662
- Words and WorldsTransféré parMedway08
- ch01Transféré parCheng Kin Meng
- Modal LogicTransféré parNathan Steinman
- Operation Research07.04.14Transféré parsandeep
- Tut+4+-+SolTransféré parhoangduc007
- Angelelli the Hierarchy of Classes of Johann Caspar Sulzer 1755Transféré parecrcau
- Gibbard (1974).pdfTransféré parJames Newgas
- Symbolic Logic - VENN.pdfTransféré parjr1234
- 1010.1282Transféré parTSaros Spiridon
- cis342-Ch02bTransféré parlionheart100
- What is Model Theory (Jhonatan Kirby)Transféré parSol Amortegui
- Nominal RestrictionTransféré parhector3nu3ez
- At least SomeTransféré parFrancesco-Alessio Ursini
- Lexical Semantics Klp 8Transféré parPutri Suryani
- FirstTransféré parabidrauf
- Bitsat 2014 Syllabus for English Proficiency and Logical ReasoningTransféré parNiranjan Savarirajalu
- Givon Logic Pragmatics EpistemologyTransféré paridschun
- A Study of Tarka and Its Role in Indian LogicTransféré parRohit
- Edward N. Zalta-Basic Concepts in Modal Logic (1995)Transféré parLuiz Helvécio Marques Segundo
- Monotheism Polytheism and Modality 1Transféré parJaiadeva Gosvami Seus
- Trends in Linguistics Studies and Monographs 202--Modality and Ellipsis (Diachronic and Synchronic Evidence) by Remus GergelTransféré parFrank Tatel
- Logic Logic is m and Word FormaTransféré parShiva Nagri
- BolzanoTransféré parPranay Kumar
- log algebraica abstracta introtuction.pdfTransféré parMichael Rivadeneira
- final-spring98.pdfTransféré parchemchemha
- l (1)Transféré parnidhi
- MathTransféré parDave Francis Felix
- cv (1).pdfTransféré parBenjamin Ibarra Quintana

- La Kabylie et les Coutumes kabyles 1/3, par Hanoteau et Letourneux, 1893Transféré parTamkaṛḍit - la Bibliothèque amazighe (berbère) internationale
- StieltjesReconstructed (1)Transféré pararrupe2
- LerayTransféré pararrupe2
- Meyer God ExistsTransféré pararrupe2
- Strategies for Reducing Binge Drinking and a Hook Up Culture on CampusTransféré pararrupe2
- Schelling OntologicalTransféré pararrupe2
- 45316780 the Unity of Mathematics in Honor of the Ninetieth Birthday of I M GelfandTransféré pararrupe2
- On Firmer Fundaments of AstrologyTransféré pararrupe2
- Documenting Riemann'SCOMPLEXTransféré pararrupe2
- 206998234 Riemann B Collected PapersTransféré pararrupe2
- esquisseengTransféré pararrupe2
- Quantum PigeonTransféré pararrupe2
- Pesic SampleTransféré pararrupe2
- Water is BestTransféré pararrupe2
- Spiritual DiaryTransféré pararrupe2
- DirectoriesTransféré pararrupe2
- 12.2.HollywoodTransféré pararrupe2
- Gordon Moore 1965 "Moore's Law" articleTransféré parGina Smith
- jmms-vol4-no2-pp082-096-collinsTransféré pararrupe2
- EulerTransféré pararrupe2
- Euler Force SpeedTransféré pararrupe2
- Gauss Potential (Eng)Transféré pararrupe2
- Wraith ArtifactsTransféré pararrupe2

- Logic GateTransféré parAkihiko Jiraiya Nelvin
- MTH202 Short Questions AnswersTransféré parcs619finalproject.com
- Lesson 4 Python Programming Automate the Boring Stuff With PythonTransféré parAmir Dinata
- 2016082212544800012269_MATH6077 _ PPT3 _ W3 _ R0Transféré parEvian Putra Setiawan
- MathematicianTransféré parmary
- ITC16 (Discrete Structures)Transféré parMELJUN CORTES, MBA,MPA
- Quiz 1 With AnswersTransféré parjavanilla
- Cmos 3Transféré parnandanvr
- Random set theoryTransféré parTytus Metrycki
- Dynamic ItabTransféré parSanket Kohli
- project on logic gatesTransféré parSonu Singhal
- Tapani Hyttinen, Olivier Lessmann and Saharon Shelah- Interpreting Groups and Fields in Some Nonelementary ClassesTransféré parUltmas
- Advanced CalculusTransféré parElliot Lu
- Categorical PropositionTransféré parPepz Montuya Lupeba
- Auto Skrip tTransféré parMuhammad Waqas
- 12 TypesTransféré parDaWheng Vargas
- FuzzificationTransféré parAvnesh Suryavanshi
- ArraysTransféré parAsafAhmad
- Discrete Mathematics_ Chapter 7 Posets Lattices & Boolean AlgeTransféré parWon Raima
- graph coloringTransféré parapi-19981779
- Xy IrrationalTransféré parZed Elspeth
- Nand and NOR Experiment.docx RevisedTransféré parjoy escandor
- Somw Weaker Forms Fuzzy Alpha Continuous MappingTransféré parDr.Hakeem Ahmed Othman
- Some New Exercises on Computability Theory Associated with Turing Machine.Transféré parEditor IJRITCC
- Homotopy Type TheoryTransféré paralfredtarski276
- Binary Search TreesTransféré parAsafAhmad
- Chapter 4 - Fundamentals of LogicTransféré parbob
- Business Mathematics LogicTransféré parSufi Khan
- Digital Design Fundamentals (Ch.3 - Gate Level Minimization)Transféré parAnonymous U9IGEgIh7V
- MatTransféré parBastian g

## Bien plus que des documents.

Découvrez tout ce que Scribd a à offrir, dont les livres et les livres audio des principaux éditeurs.

Annulez à tout moment.