This book, written by one of philosophy's pre-eminent logicians,

argues that many of the basic assumptions commonly made in logic,

the foundations of mathematics and metaphysics are in need of
change. It is therefore a book of critical importance to logical theory
and the philosophy of mathematics.
Jaakko Hintikka proposes a new basic first-order logic and uses it
to explore the foundations of mathematics. This new logic enables
logicians to express on the first-order level such concepts as equi-
cardinality, infinity and truth in the same language. The famous
impossibility results by Gödel and Tarski that have dominated the
field for the past sixty years turn out to be much less significant than
has been thought. All of ordinary mathematics can in principle be
done on this first-order level, thus dispensing with all problems
concerning the existence of sets and other higher-order entities.
THE PRINCIPLES OF MATHEMATICS REVISITED
THE PRINCIPLES OF
MATHEMATICS
REVISITED

JAAKKO HINTJKKA
Boston University

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

CAMBRIDGE UNIVERSITY PRESS

The Edinburgh Building, Cambridge CB2 2RU, UK http://www.cup.cam.ac.uk
40 West 20th Street, New York, NY 10011-4211, USA http://www.cup.org
10 Stamford Road, Oaldeigh, Melbourne 3166, Australia
Ruiz deAlarcón 13, 28014 Madrid, Spain

© Cambridge University Press 1996

This book is in copyright. Subject to statutory exception

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

First published 1996

First paperback edition 1998
Reprinted 1999

Typeset in 'Fimes

Libra,y of Congress Cataloging in Publication data is available.

A catalog record for this book is available from the British Library.

ISBN 0521 49692 6 hardback

ISBN 0521 62498 3 paperback

Transferred to digital printing 2003

 Contents

Introduction page vii

1 The functions of logic and the problem of truth

definition 1

2 The game of logic 22

3 Frege's fallacy foiled: Independence-friendly logic 46
4 The joys of independence: Some uses of IF logic 72
5 The complexities of completeness 88
6 Who's afraid of Alfred Tarski? Truth definitions for
IF first-order languages 105
7 The liar belied: Negation in IF logic 131
8 Axiomatic set theory: Fraenkelstein's monster? 163
9 IF logic as a framework for mathematical theorizing 183
10 Constructivism reconstructed 211
11 The epistemology of mathematical objects 235
Appendix (by Gabriel Sandu) 254

References 271
Index of names 281
Index of subjects and titles 285

V
 Introduction

The title of this work is modeled on Bertrand Russell's 1903 book

The Principles of Mathematics. What is the connection? As I see it,
Russell's book was an important step in his struggle to liberate
himself from traditional approaches to logic and the foundations of
mathematics and to replace them by an approach using as its main
tool, and deriving its inspiration from, the new logic created by Frege
and Peano. In the Principles, Russell is not yet actually constructing
the new foundation for mathematics which he later built with A. N.
Whitehead. The Principles is not the Principia. What Russell is doing
in the 1903 book is to examine the conceptual problems that arise in
the foundations of logic and mathematics, expose the difficulties in
the earlier views and by so doing try to find guidelines for the right
approach.
In this book, I am hoping in the same spirit to prepare the ground
for the next revolution (in Jefferson's sense rather than Lenin's) in the
foundations of mathematics. As in Russell, this involves both a criti-
cal and a constructive task, even though they cannot be separated
from each other. Indeed, if there had not been the danger of
confusing bibliographers, I would have given this book the double
entendre title The Principles of Mathematics Revis(it)ed.
The critical part of my agenda is the easier one to describe. Indeed,
if! were Thomas Kuhn, I could give a most concise description of the
state of the art in the foundations of logic and mathematics. Almost
everyone in the field assumes that he or she is doing normal science,
even though in reality a serious crisis is about to break out. Hence
what is needed here is crisis science in Kuhn's sense, and not normal
vii
Viii INTRODUCTION

science with its safe, or at least generally accepted, conceptual

framework and its agreed-upon criteria of success. Most of the
received views concerning the overall situation in the foundations of
logic and mathematics are not only open to doubt but are arguably
(and often demonstrably) false. Indeed, arguments and demonstra-
tions to this effect will be presented in this work.
The dogmas that are ripe for rejection include the following
commonplaces: The basic part of logic, the true elementary logic, is
ordinary first-order logic, at least if you are a classicist. If you are an
intuitionist, you use instead Heyting's intuitionist first-order logic.
In either case, our basic elementary logic admits of a complete
axiomatization.
On the level of first-order logic, you can only formulate formal
rules of inference, that is, treat logic syntactically. For (and with this
point I agree) in order to do semantics (model theory), you need
a truth definition for the language in question. According to the
current views, such a truth definition cannot be formulated in the
same language, only in a stronger metalanguage. For first-order
languages, such a truth definition must naturally be formulated in set
theory or in a second-order language. In both cases, we are dealing
with a mathematical rather than a purely logical language. Model
theory is thus inevitably a mathematical discipline, and not a purely
logical one. Moreover, what a formal truth definition can do is only
an abstract correlation between sentences and the facts that make
them true. They cannot provide an explanation of what it is that
makes a sentence true. Formal truth definitions cannot show, either,
how sentences are actually verified.
Nontrivial first-order mathematical theories, including elementary
arithmetic, are inevitably incomplete in all the interesting senses of
the term. There are no absolute principles to guide the search for
stronger deductive axioms for mathematical theories.
First-order logic is incapable of dealing with the most characteris-
tic concepts and modes of inference in mathematics, such as math-
ematical induction, infinity, equicardinality, well-ordering, power
set formation, and so forth. Hence mathematical thinking involves
essentially higher-order entities of some sort or the other, be they
sets, classes, relations, predicates, and so forth, in the strong sense of
involving quantification over them. Formally speaking, mathemat-
ics can be done on the first-order level only if the values of individual
variables include higher-order entities, such as sets. It is therefore
 INTRODUCTION ix

theoretically illuminating to formulate mathematical theories in

set-theoretical terms. Axiomatic set theory is accordingly a natural
framework for mathematical theorizing.
Negation is a simple concept which involves merely a reversal of
truth-values, true and false. It is a characteristic feature of construc-
tivistic approaches to logic to deny the law of excluded middle.
Another trademark of constructivism is the rejection of the axiom of
choice. In general, the way to implement a constructivistic interpre-
tation of logic is to change the rules of inference or perhaps to change
some of the clauses in a truth definition.
Furthermore, it is generally (though not universally) assumed by
linguists, logicians and philosophers that one should try to abide by
the so-called principle of compositionality in one's semantics. It says
that the semantical value of a given complex expression is always
a function of the semantical values of its component expressions. The
function in question is determined by the syntactical form of the
given expression.
In this work, I will argue that not only one or two of these
widespread views are mistaken, but that all of them are wrong.
Moreover, I will show what the correct views are that should replace
the erroneous dogmas listed above. I am not a deconstructivist or
a skeptic. On the contrary, I am trying to wake my fellow philos-
ophers of mathematics from their skeptical slumbers and to point
out to them a wealth of new constructive possibilities in the founda-
tions of mathematics. This will be done by developing a new and
better basic logic to replace ordinary first-order logic. This received
basic logic owes its existence to an unduly restrictive formulation of
the formation rules for first-order logic probably motivated by an
uncritical acceptance of the dogma of compositionality. The new
logic which is free of these restrictions provides the insights that in
my judgment are needed in the foundations of logic and mathemat-
ics. Ironically, the received first-order logic which I am demoting is
the core part of the then new logic to whose introduction Russell was
paving the way in the Principles of Mathematics. The admirers of
Russell can nevertheless find some consolation in the closeness of my
new logic to the kind of logic Russell was propagating. The two
nevertheless differ significantly in their philosophical and other
general theoretical implications and suggestions.
In my judgment, even the critical aspects of this book will have
a liberating effect. The reason is that many of the dogmas I have just
X INTRODUCTION

listed are essentially restrictive, including theorems as to what

cannot be done in logic or in mathematics. Refuting them is therefore
an act of liberation. For a significant example, it turns out that in one
perfectly natural sense of the word mathematics can in principle be
done on the first-order level. In general my liberated first-order logic
can do much more in mathematics than philosophers have recently
thought of as being possible.
The strategy of this logician's liberation movement does not rely
on complicated technical results, but on a careful examination of
several of the key concepts we employ in logic and mathematics,
including the following concepts: quantifier, scope, logical priority
and logical dependence, completeness, truth, negation, constructiv-
ity, and knowledge of objects as distinguished from knowledge of
facts. A patient analysis of these ideas can uncover in them greater
conceptual riches than philosophers and mathematicians have sus-
pected. It has recently been fashionable to badmouth Carnap's idea
of philosophy of science as a logical analysis of the language of
science, and by analogy of the philosophy of mathematics as the
logical analysis of the "language" of mathematics that is, of the
basic concepts of mathematics and metamathematics. Undoubtedly
Carnap and his colleagues were neither sensitive enough to the
deeper conceptual issues nor in possession of sufficiently powerful
logical tools. Nonetheless, objectively speaking a study of the roles of
language, truth and logic in mathematics seems to me by far the most
fruitful approach to the philosophical nature of mathematical the-
ories. For a brief while! was playing with the idea of calling this book
Language, Truth and Logic in Mathematics.
This strategy implies that my book will be more like Russell's
Principles than his Principia in another respect, too. Most of the time
I am not developing my ideas here in the form of a mathematical or
logical treatise, with full formal details and explicit proof. This work
is a philosophical essay, and not a research paper or treatise in logic
or mathematics. Even though I will try to explain all the main formal
details that I need in this book, I will accordingly do so only in so far
as they seem to be necessary for the purpose of understanding my
overall line of thought.
In thinking about the subject matter — or, rather, subject mat-
ters — of this book over the years! have profited from exchanges with
more people than I can recall, let alone list here. Three of them
occupy a special position, however, because they have made a contri-
 INTRODUCTION Xl

bution of one kind or the other to the genesis of this book. Some of
the points I am making here I have argued with or, rather,
against — one of my oldest American friends, Burton Dreben, ever
since I came to know him in 1954. After forty years, I believe I have
finally found the clinching arguments for my line of thought.
Most of the new ideas expounded in this book I have developed in
close cooperation with Gabriel Sandu. In some cases, I cannot any
longer say who came upon which idea first. I can only hope that he
will soon publish his version of some of the main results of this book.
Sandu has also contributed to this book substantially through
specific comments, criticisms and suggestions. He has written an
important comparison between some of our results and Kripke's
approach to truth which is printed below as an appendix to this
book. In brief, this book owes more to Gabriel Sandu than I am
likely to be aware of myself. Furthermore, Janne Hiipakka has been
of invaluable help, not only in preparing the manuscript but also in
correcting mistakes and making suggestions.
In working on this book, I have come to appreciate more and
more the importance of the issues raised by the classics of twentieth-
century philosophy of mathematics, especially Hilbert, Tarski,
Carnap and Gödel. While I frequently disagree with each of them,
the questions they were asking are much more important than most
of the issues that have been debated recently, and eminently worth
going back to for problems and ideas. I only regret that I have
developed my own ideas far too late to discuss them with the likes of
Tarski, Carnap or Gödel.
I have not received any direct grant support or release time while
writing this book. Indirect support includes a grant from the Acad-
emy of Finland which has made possible the work of Sandu and
Hiipakka. Travel grants from the Academy of Finland have also
facilitated several of my working trips to Finland. Boston University
has provided a congenial working environment. In fact, the origins
of this book can be traced back to the insight that self-applied truth
definitions are possible for independence-friendly languages; an
insight which first occurred to me in the midst of a seminar I was
conducting at Boston University in 1991. I have also had an
opportunity to help organize several symposia under the auspices of
the Boston Colloquium for Philosophy and History of Science
which have provided stimuli for the ideas that have reached fruition
in this book.
Xii INTRODUCTION

Relatively little of the material published here has been available

earlier. Much of Chapter 10 has been published under the title
"Constructivism Aufgehoben" in the proceedings of the "Logica 94"
conference held in the Czech Republic in June 1994. A version of
Chapter 2 has appeared in Dialectica vol.49 (1995), pp. 229—249
under the title, "The Games of Logic and the Games of Inquiry."
Some examples and some of the other material in this book, especial-
ly in Chapter 7, are lifted from the preprint entitled Defining Truth,
the Whole Truth and Nothing But the Truth, published in the
preprint series of the Department of Philosophy, University of
Helsinki, in 1991. The rest of my literary "thefts" from my own earlier
works are unintentional.
Finally my thanks are due to Cambridge University Press for
accepting my book for publication in their program. It is a special
compliment for a work on the foundations of mathematics to share
a publisher with Russell and Whitehead.
 1

The Functions of Logic and the Problem

of Truth Definition

The most pervasive misconception about the role of logic in math-

ematical theorizing may turn out to be the most important one.
Admittedly, this mistake is easily dressed up to look like a biased
emphasis or an exaggerated division of labor. But it is nonetheless
a serious matter. It can be exposed by asking the naive-sounding
question: What can logic do for a mathematician? What is the role of
logic in mathematics?
As a case study, I propose to consider one of the most influential
works in the foundations of mathematics and one of the last major
works that does not use the resources of modern logic. This work is
Hubert's Foundations of Geometry (1899). What Hilbert does there is
to present an axiomatization (axiom system) of Euclidean geometry.
Such a system is a nonlogical axiom system. It is a systematization of
the truths (ordinary scientific or mathematical truths, not logical
truths) of some discipline, usually some branch of mathematics or
science. The systematization is obtained by, as it were, compressing
all the truths about the subject matter into a finite (or recursively
enumerable) set of axioms. In a certain sense, they are supposed to
tell you everything there is to be told about this subject matter. Such
an axiomatization, if complete, will give you an overview of the
entire field in question. If you have reached in your investigation into
this field a complete axiom system, then the rest of your work will
consist in merely teasing out the logical consequences of the axioms.
You do not any longer need any new observations, experiments or
other inputs from reality. It suffices to study the axioms; you no
longer need to study the reality they represent. Obviously, this
1
2 THE PRINCIPLES OF MATHEMATICS REVISITED

intellectual mastery of an entire discipline is the most important

attraction of the entire axiomatic method.
Philosophers sometimes think of the axiomatic method as a way of
justifying the truths that an axiom system captures as its theorems. If
so, the axioms have to be more obvious than the theorems, and the
derivation of theorems from axioms has to preserve truth. The latter
requirement is discussed below. As to the former, the obviousness
requirement plays no role in the most important scientific theories. No
one has ever claimed that Maxwell's or Schrodinger's equations are
intuitively obvious. The interest of such fundamental physical equa-
tions is not even diminished essentially by the knowledge that they are
only approximately true. The explanation is that such equations still
offer an overview of a wide range of phenomena. They are means of
the intellectual mastery of the part of reality they deal with. Generally
speaking, this task of intellectual mastery is a much more important
motivation of the axiomatic method than a quest of certainty.
Hilbert's axiomatization of geometry is an extreme example of
this fact. Hilbert does not even raise the question whether the axioms
of Euclidean geometry are true in actual physical space. All he is
interested in is the structure delineated by the axioms. This structure
is spelled out by the theorems of the axiom system. Whether this
structure is instantiated by what we call points, lines and planes or by
entities of some other kind is immaterial to his purpose.
But what is thejob of logic in this enterprise? In Hilbert's treatise it
remains tacit or at least unformalized. Not a single symbol of formal
logic disfigures Hubert's pages. Hilbert's monograph could have
been written even if Boole, Frege and Cantor had never been born.
Yet logic is an all-important presence in Hubert's treatise in more
than one way.
One way is completely obvious to any contemporary philosopher,
logician or mathematician. In fact, this role of logic is so obvious that
subsequent thinkers tend to take it for granted, thereby missing
much of the historical significance of Hilbert's achievement. Hubert
is envisaging what is nowadays called a purely logical axiom system.
That is to say, all substantive assumptions are codified in the axioms,
whereas all the theorems are derived from the axioms by purely
logical means. This idea may be taken to be the gist of the entire
axiomatic method. Familiarity with this idea is so complete in our
days that we easily forget what a bold novelty such a treatment was
in Hilbert's historical situation. One index of this novelty is the label
 THE FUNCTIONS OF LOGIC 3

that Hubert's contemporaries put on his approach, calling it "for-

malistic". Of course this term is completely undeserved. When
Hubert said that instead of points, lines, and circles he could have
spoken of chairs, tables, and beermugs, he was merely highlighting
the purely logical nature of the derivation of theorems from axioms.1
Because of the purely logical nature of Hilbert's axiom system, all
proofs from axioms could be represented in an explicit formal (but of
course interpreted) logical notation, without any relevant difference.
This idea of a purely logical consequence was familiar to Aristotle,
though not to such intervening thinkers as Kant. This purely logical
character of the passage from the axioms to theorems does not mean
that the axioms and the theorems must be uninterpreted. It does not
mean that the axioms cannot be true (materially true). In fact,
Hilbert elsewhere (1918, p. 149) indicated quite explicitly that in any
actual application of geometry, the truth of the axioms is a material
(empirical) matter. Even the actual truth of the axioms of continuity
has to be empirically ascertained.2
A further distinction is nevertheless in order. It is important to
realize that a nonlogical axiom system can be either interpreted, as in
applied geometry or thermodynamics, or uninterpreted, as in set
theory or lattice theory. The only difference between the two is that
in the latter the fundamental nonlogical concepts are left uninter-
preted. This does not make any difference to the derivation of con-
sequences (e.g., theorems) from premises (e.g., axioms). Such deriva-
tions proceed exactly in the same way in the two cases, as long as they
are purely logical. Hence in purely logically formulated nonlogical
axiom systems the contrast between interpreted and uninterpreted
systems does not matter. Hilbert's provocative slogan about the
geometry of chairs, tables, and beermugs can be understood as
flaunting this parity of interpreted and uninterpreted axiom systems
when it comes to the derivation of theorems, and hence indirectly
illustrating the purely logical character of geometrical proofs.
The same point can be put in a slightly different way. The question
whether logical inferences can be captured by completely formal
(computable, recursive) rules is independent of the question whether
the language in which those inferences are drawn is "formal", that is,
uninterpreted, or "informal", that is, interpreted. The question of
interpretation pertains to the nonlogical constants of the language.
Logical constants are assumed to have the same meaning in either
case.
4 THE PRINCIPLES OF MATHEMATICS REVISITED

Uninterpreted nonlogical axiom systems may be thought of as

pertaining to certain structures as such, while the corresponding
interpreted systems deal with the actual instantiations of these
structures.
Here, then, is an example of the role of logic which understandably
is on top of the minds of most of my fellow philosophers, logicians
and mathematicians. Logic is the study of the relations of logical
consequence, that is, of relations of implication or entailment. Its
concrete manifestation is an ability to perform logical inferences,
that is, to draw deductive conclusions. I will call this the deductive
function of logic.
The tools needed for this job are normally thought of as being
collected into sundry axiomatizations of logic. The term "axiomatiza-
tion" as used here is nevertheless a courtesy title. For a so-called
axiomatization of some part of logic is merely a method of enumerat-
ing recursively all the logical truths expressible in some explicit
("formal") language or the other. This enumeration is usually made
to appear uniform with nonlogical axiom systems, like Hilbert's
system of geometry. That is to say, the enumeration is obtained by
putting forward completely formal axioms to start the enumeration
from, and equally formal rules of inference whose repeated applica-
tions are the means of accomplishing the enumeration. In spite of
this similarity in expositional style, one must be aware of the
fundamental differences between nonlogical and logical axiom sys-
tems. Nonlogical axiomatization trades in ordinary nonlogical (ma-
terial or mathematical) truth, logical axiomatization in so-called
logical truth. Nonlogical axiom systems normally have an intended
interpretation, even when the derivations from axioms are purely
logical, whereas an axiomatization of logic has to be purely formal to
generate the mechanical enumerability. A nonlogical axiom system
can be formulated without saying anything about any explicitly
formulated logic. In fact, this is what Hilbert does in the Grundlagen
(1899). Even though all his derivations of theorems from axioms are
carried through in conformity with impeccable logic, not a single
explicit rule of logical inference is ever appealed to by Hilbert in his
famous book.
From an axiomatization of logic, patterns of valid logical infer-
ence are obtained as a special case in the form of conditionals

(1.1)
 THE FUNCTIONS OF LOGIC 5

If and only if such a conditional (with S1 as the antecedent and S2 as

the consequent) is logically true (valid, as a sometime alternative
term goes) is an inference from S1 to S2 valid.
Nevertheless in the course of this book I will point out difficulties
with this characterization of valid inferences in terms of logical truth.
I will return to this matter at the end of Chapter 7.
It is important to realize that this task of spelling out the valid
inference-patterns cannot be entrusted to the ill-named "rules of
inference" of an axiomatization of logic. For the raison d'être of such
"rules of inference" is to serve as a vehicle for enumerating logical
truths, and not material or mathematical truths. Hence the only
thing required of them is that they preserve logical truth. For this
purpose, it is not even necessary that they preserve material truth,
that is, truth simpliciter. There are in fact plenty of examples of
so-called inference rules used in the axiomatization of some part of
logic that do not preserve truth as such. The best known cases in
point are probably the rules of necessitation in modal logic. In them,
one may for instance "infer" the logical truth of
N(S1 S2) (1.2)

from the logical truth of

(1.3)

where N is the necessity operator. But of course from the plain

(natural) truth of (1.3) one cannot infer the truth of (1.2). Similar
examples can also be found in ordinary nonmodal logic.
Thus a philosopher should be extra careful with such terms as
"rule of inference" or even "rule of logical inference". They can in the
wrong hands be viciously ambiguous. At the very least, we have to
distinguish carefully between a logical proof (from premises) of
material truth and logical (pseudo) proof of a logical truth. The latter
is best looked upon merely as a segment of the recursive enumer-
ation of logical truths.
For a while, mathematicians tended to take the basic rules of valid
logical inference for granted. Sometimes they were thought of as
being too obvious to need any particularly explicit treatment. Some-
times it was assumed — and is still being assumed that Frege and
Russell accomplished this task once and for all. Or if they did not,
then Hilbert's henchman Wilhelm Ackermann did so. For what
6 THE PRINCIPLES OF MATHEMATICS REVISITED

Ackermann did in 1928 under Hubert's supervision was to formulate

for the first time that basic ground-floor part of logic (as it seems to
be) that everybody and his sister have taken for granted ever since.
This is what is usually known as first-order logic, quantification
theory or (lower) predicate calculus. It used to be presented as if it
resulted from a minor regimentation of our ordinary-language use of
everys and somes. This idea seems to be vindicated by Chomsky's use
of quantificational logic as the main medium of his born-again
logical forms (LF's) of natural-language sentences (see, e.g.,
Chomsky 1977, p. 59). Small wonder, therefore, that first-order logic
is generally taken to be the safe, unproblematic core area of logic,
and that the languages relying on it are considered as a natural
vehicle of our normal reasoning and thinking. A number of concrete
developments in the foundations of mathematics in fact prompted
logicians and mathematicians to present explicit axiomatizations of
this part of logic separate from those higher echelons of logic which
involve quantification over such abstract entities as predicates,
classes, sets, and/or relations.
It is in fact easy to see why first-order logic at first seems to be
a logician's dream come true. First and foremost this logic admits of
a complete axiomatization. This seems at first sight to effectively
fulfill the hopes of Hilbert and others that there is an unproblematic
and completely axiomatizable basic logic. The existence of such
a logic was in fact one of the presuppositions of what is known as
Hubert's program — a program, that is, of proving the consistency of
certain mathematical theories, primarily arithmetic and analysis, by
showing that one cannot formally derive a contradiction from their
axioms. If the logic that is being used is not complete, then the entire
project loses its point for there may then be unprovable contradic-
tions among the logical consequences of the axioms. Luckily, as it
seems, first-order logic was shown to be completely axiomatizable
by Gödel in 1930. Furthermore, first-order logic can be shown to
admit all sorts of nice metalogical results, such as compactness (an
infinite set of sentences is consistent if and only if all its finite subsets
are), upwards Löwenheim—Skolem theorem (a consistent finite set of
sentences has a countable model), separation theorem (if i and t are
consistent sets of formulas but c u 'r is inconsistent, then for some
"separation formula" S in the shared vocabulary of a and t we have
aI—S, ti— S), interpolation theorem (if I—(S1 nontrivially,
then for some formula I in the shared vocabulary of S1 and S2,
 THE FUNCTIONS OF LOGIC 7

I), —(I S2)), Beth's theorem (implicit definability implies

explicit definability), and so forth. In brief, first-order logic does not
only seem basic, it is very nearly looks like a logician's paradise.
Admittedly, constructivistic logicians and mathematicians have
sought to change some of the rules of first-order logic. But this does
not seem to have shaken most philosophers' beliefs that in some size,
shape or form first-order logic is the true unproblematic core area of
logic.
Philosophically, first-order logic owes its special status largely to
the fact that it is in a sense a nominalistic enterprise. It involves
quantification only over individuals, that is, particular objects or
entities that are dealt with as if they were particulars. This is in fact
why this logic is called first-order logic.
However, storm clouds begin to gather as soon as logicians
venture beyond the enchanted land of first-order logic. And they
have to do so, for unfortunately first-order logic soon turns out to be
too weak for most mathematical purposes. Its resources do not
suffice to characterize fully such crucial concepts as mathematical
induction, well-ordering, finiteness, cardinality, power set, and so
forth. First-order logic is thus insufficient for most purposes of actual
mathematical theorizing. Moreover, the terrain beyond it has pro-
ved uncharted and treacherous. It is notorious that those logical
methods which go beyond first-order logic give rise to serious
problems, no matter whether we are dealing with set theory or
higher-order logic (type theory). There does not at first seem to be
much to choose between the two. Second-order logic has been
branded by Quine as set theory in sheep's clothing. And there is in
fact a multitude of overlapping problems concerning the two. In
highlighting these problems, philosophers of yore used to point to
paradoxes sometimes amounting to outright contradictions. The
sad truth is that even if we rid ourselves of all threatening inconsist-
encies, maybe by some sort of type stratification (as in second-order
logic and more generally in higher-order logic), we still have the
problem of finding out which formulas are logically true. Eliminat-
ing what is logically false does not give us a means of ascertaining
what is logically true. Often these problems are quite specific. Is the
axiom of choice valid? What axioms do we need in axiomatic set
theory? How are higher-order quantifiers to be understood? The list
of serious problems can be continued ad nauseam, if not ad in-
finitum.
8 THE PRINCIPLES OF MATHEMATICS REVISITED

If! ended my survey of the role of logic in mathematical theories

here, perhaps with some further elaboration of the difficulties
connected with the uses of logic beyond first-order theories, few
philosophers would be surprised. However, such a procedure would
leave half of the real story untold, and indeed the more basic half of
the story of the uses of logic in mathematics.
In order to enable you to see what I mean, I can ask: What would
you have to do in order to turn Hilbert's axiomatization of geometry
into a state of the art logic-based axiom system? What I have
discussed so far is what has to be done to Hilbert's argumentation,
that is, to the lines of reasoning which lead him from the axioms to
the theorems. Most of such reasoning turns out to utilize first-order
reasoning, with a smattering of number-theoretical and set-
theoretical reasoning thrown in. But even before we can hope to
express Hilbert's inferences in the language of logic, we have to
express in such a language the ultimate premises of his inferences
that is to say, the axioms of geometry. And here is the other, unjustly
underemphasized, second function of logical concepts in mathemat-
ics. They are relied on essentially in the very formulation of math-
ematical theories. We can make the axioms of a typical mathematical
theory say what they say only by using suitable logical concepts,
such as quantifiers and logical connectives.
This fact is vividly illustrated once again by Hilbert's treatment of
geometry. His axioms deal with certain specified relations between-
ness, equidistance, and so forth, between certain specified kinds of
objects points, lines, and so forth. What the axioms say about these
entities is formulated by means of the basic logical notions, mainly
quantifiers and propositional connectives. Of course, Hilbert does
not use any explicit logical symbols, but formulates his axioms by
means of ordinary mathematical language. But anyone who has
a modicum of elementary logic can write down a first-order formula-
tion of all of Hilbert's axioms in fifteen minutes, with two exceptions.
They are the Archimedean axiom, which relies on the notion of
natural number, and the so-called axiom of completeness, which
turns on the notion of the maximality of a model.
The Archimedean axiom says that from a given line segment one
can reach any given point on the same line by extending the segment
by its own length a finite number of times. The axiom of complete-
ness was a second thought on Hilbert's part. It appears first in
Hubert (1900), and makes its entry into Hilbert's geometrical mono-
 THE FUNCTIONS OF LOGIC 9

graph in its French translation (1900) and then in the second

German edition (1903). Its import can be expressed by saying that
the intended models of the other axioms must be maximal in the
sense that no new geometrical objects can be added to them without
making one of these axioms false.
More generally, much of the foundational work that has been
done since Cauchy by working mathematicians consisted in express-
ing in first-order logical terms the precise contents of different
mathematical concepts. A familiar but instructive case in point is
offered by the so-called s-5 definitions of concepts like continuity
and differentiation. (I will return to them in Chapter 2.) Another
example is found in the different and increasingly general concepts of
integral that were developed in the late nineteenth and early twenti-
eth centuries. In developing such ideas, mathematicians were not
engaged in the discovery of new mathematical truths. They were
engaged in analyzing different mathematical concepts in logical
terms.
I will call this function of logic (logical concepts) in expressing the
content of mathematical propositions its descriptive function. Many
interesting phenomena in the foundations of mathematics become
easier to understand in the light of the tension which there often is in
evidence between this descriptive function and the deductive func-
tion of logic in mathematics. Arguably, the descriptive function is the
more basic of the two functions! have distinguished from each other.
If mathematical propositions were not expressed in terms of logical
concepts, their inferential relationships would not be possible to
handle by means of logic.
What I have called the descriptive function of logic can be put into
service as a tool of conceptual analysis. This possibility is illustrated
by examples from the history of mathematics like the ones just
mentioned, but it is not restricted to mathematics. It is what under-
lies Hilbert's less than completely happy terminology when he calls
his axioms implicit definitions of the geometrical concepts involved
in them.
It is important to realize that this descriptive function of logic in
formulating mathematical axioms is the same — and hence equally
indispensable no matter whether the axioms belong to an inter-
preted axiom system, for instance, to the axioms of thermodynamics
or interpreted geometry, or to an uninterpreted axiom system, such
as the axioms of group theory, theory of fields, or lattice theory. Even
10 THE PRINCIPLES OF MATHEMATICS REVISITED

abstract mathematical theories can be thought of as explications of

certain intuitive concepts, topology as an explication of the concept
of continuity, group theory of the idea of symmetry, lattice theory
of the notions related to the idea of ordering, and so forth. In all these
cases, for the explicatory purpose logical notions of some sort are clearly
a must. Thus we can see that those philosophers who maintain that
logical notions like quantifiers do not have the same sense in mathe-
matical theories as they have in everyday life are not only off the mark
but off the map.3 On the contrary, the representation of mathemat-
ical propositions in formal or informal language is predicated on the
assumption that logical constants are being used with their normal
meaning.
A realization of the difference between the two functions of logic in
mathematics prompts further questions. The main one concerns the
way different parts and aspects of logic serve the two (or three) func-
tions. Which parts serve which function? What is required of logic
that it can serve one or the other of the two functions? Or, must
there for some reason be just one indivisible logic serving both the
descriptive and the deductive purpose? These are fundamental
questions in the philosophy of mathematics, and yet they have been
scarcely raised in earlier literature. I will return to them at the end of
Chapters 9 and 10.
Acknowledging the descriptive function of logic in mathematical
theories in fact occasions an addition to what was said earlier of the
purely logical character of an axiom system like Hilbert's
Grundlagen. Not only must it be the case that all the proofs of
theorems from axioms are conducted purely logically; it must also be
required that the representational task of the axioms is accom-
plished by purely logical means. In practice this normally means that
the only nonlogical notions in the axiom system are certain unde-
fined properties and relations among the objects of the theory in
question. Of them, only such things are assumed that are explicitly
stated in the axioms.
This purely logical character of the formulation of axioms is only
partly necessitated by the requirement that proofs of theorems from
axioms must be purely logical. It is part and parcel of Hilbert's
conception of the axiomatic method. A partial reason is that the only
entities that the axioms may mention must be the objects of the
theory. For instance, in the axioms of geometry we must not postu-
late a correspondence between points on a line and real numbers, for
 THE FUNCTIONS OF LOGIC 11

the latter are not objects of geometry. In actual historical fact,

Hubert's choice of his axioms seems to have been deeply influenced
by the requirement that they must be purely logical in the sense
indicated here.
The neglect of the descriptive function of logical is especially
striking in recent discussions of the philosophical problems of
cognitive science. A central role is played there by the notion of
representation, especially mental but to some extent also linguistic.
Now, what would a typical situation of linguistic representation
look like? On the basis of what was said earlier, representation by
means of first-order logic should certainly be the test case. There may
not be any completely presuppositionless Adamic discourse to serve
as a testing-ground of theories of representation, but in the first-
order heirs to Frege's we seem to have — Gott sei
Lob a simple mode of representation which everybody admits is
a basic one. Or is it the case that we have to exempt a great many
cognitive scientists and their pet philosophers from the scope of that
"everybody"? In reading the relevant literature, it is hard to avoid the
impression that many contemporary philosophers of cognitive
science are desperately trying to restrict their attention to such forms
of representation as do not involve logic. If so, we are witnessing
a comedy in which one fraction of philosophers scrupulously tries to
avoid discussing the very cases of representation which others treat
as the paradigmatic ones. Or, is it perhaps that philosophers of
cognition have not acknowledged the descriptive function of logic,
which amounts in effect to its representative function? If so, they are
disregarding the medium of representation for the overwhelming
majority of our advanced theoretical knowledge.
And even if a philosopher believes that the all-important mental
information processing uses some means other than linguistic or
logical representation, he or she should still offer a realistic alterna-
tive account of the cognitive processes that can be carried out by
means of first-order languages.
The systematic study of the deductive function of logic is known
as proof theory. The systematic study of the descriptive function of
logic is known as model theory or logical semantics. Several highly
influential logicians and philosophers used to maintain and in
some cases still maintain—the impossibility of model theory as
a large-scale philosophically relevant systematic enterprise. Some
used to deny the very possibility of model theory outright, some its
12 THE PRINCIPLES OF MATHEMATICS REVISITED

possibility in the all-important case of our actual working (or

perhaps better, thinking) language, while still others merely deny
its philosophical relevance. In different variants, such doubts have
been expressed by Frege, Russell, Wittgenstein, the Carnap of the
early thirties, Quine and Church. The way in which this tradition
was gradually and partially overcome is studied briefly in Hintikka
(1988b). This underestimation of model theory among philosophi-
cally oriented logicians can typically be traced back to a failure to
appreciate the descriptive (representational) function of logic.
Notwithstanding such doubts, model theory has apparently
proved possible, because actual, as the scholastics used to say. But
questions can still be raised about its philosophical relevance.
What precisely are the conceptual presuppositions of model theory?
The crucial concept in any such theory is, unsurprisingly, the
concept of model. The idea is to discuss what a sentence S says by
associating with it a class of structures, also known as models,
scenarios, systems, possible worlds, worlds, or what not. Let us
call this class M(S). Strictly speaking, there are two related senses
of model here, depending on whether in the passage from S to
M(S) the nonlogical constants of S are allowed to be reinterpreted.
Models of the second kind are of course the same as the models of
the first kind modulo isomorphism. In the foundations of mathemat-
ics, the distinction between the two kinds of models makes little
difference.
But how is M(S) specified? In order to specify it, we must obvious-
ly do two things. First, we have to be given some class (set, space) of
models, that is, structures of the appropriate sort. Second, by
reference to S we have to give the criteria as to when a given member
M of qualifies as one of the models of S.
The former question (the question of the choice of did not
attract much attention on the part of logicians until the development
of abstract (alias model-theoretical) I ogics (see here Barwise and
Feferman 1985). Independently of them, certain interesting candi-
dates for a nonstandard choice of have meanwhile been proposed.
They can involve on the one hand novel structures that were not
earlier considered as models, such as the urn models of Veikko
Rantala (1975) that will be explained in Chapter 5. On the other
hand they can involve deliberate restrictions on the space of models.
One especially interesting restriction is obtained by requiring that
the surviving models should possess certain suitable extremality
 THE FUNCTIONS OF LOGIC 13

(maximality and minimality) properties (see Hintikka 1993b). More

work is nevertheless in order in this direction. Perhaps the "special
models"just mentioned could in the future serve usefully as test cases
in the very, very abstract theory of abstract logics.
I will return briefly to the ideas of maximality and minimality in
Chapter 9.
For the purposes of my line of thought here, the central question is
the second one, the question as to when a model M is a model of
a sentence S. Here a partial answer is obvious. In the basic sense of
model, M is a model of S if and only ifS is true in M. And the condi-
tions of a sentence being true in a model is what truth definitions
codify. Hence the specification of the all-important relations of being
a model of is essentially a matter of truth definitions.
The question of the possibility of truth definitions, and the
question of the presuppositions of such truth definitions, are thus of
major interest to the entire foundation of logic and mathematics.
The philosophical viability of model theory stands or falls with the
philosophical viability of truth definitions, and the relative depend-
ence or independence of model theory on other approaches to
foundations is a matter of whether those other approaches are
presupposed in the relevant truth definitions. Even though Tarski
himself never highlighted the fact (as far as I know), it is no accident
that the same thinker developed both the first explicit truth defini-
tions and later (with his students and associates) the contemporary
model theory in the narrower technical sense.
The most commonly used type of truth definition was introduced
by Alfred Tarski in 1935 — or perhaps 1933, if we heed the publica-
tion date of the Polish original. As will be explained more fully in
Chapter 5, the guiding principle of Tarski's enterprise was what
logicians at the time (and later) usually called the idea of recursive
definition. Later linguists are wont to call it compositionality. This
principle says that the semantic attributes of a complex expression
are functions of the semantic attributes of its constituent expressions.
Applied to truth definitions, it says that the truth-value of a sentence
is determined by the semantical attributes of its constituent expres-
sions. One perhaps hopes to be able to say here: it depends only, on
the truth-values of its constituent expressions. Unfortunately the
constituent expressions of quantified sentences typically contain free
variables. They are open formulas, not sentences, and therefore
cannot have a truth-value.
14 THE PRINCIPLES OF MATHEMATICS REViSITED

This explains the first main feature of Tarski's truth definition.

The truth-value of a sentence is defined by him with the help of
another notion which applies also to open formulas, namely, the
notion of satisfaction. The definition is applied from inside out,
beginning with the simplest (atomic) formulas in the sentence in
question. A crucial role is played by the notion of a valuation, which is
indeed an assignment of individuals to each individual constant and
each individual variable of the language in question as its values. This
reliance of Tarski-type definitions of truth on the auxiliary concept of
satisfaction has, strangely enough, led some scholars to deny that
Tarski's truth definition is compositional.
Very briefly, a Tarski-type truth definition can be explained by
keeping in mind that truth is relative to a model M and a valuation v.
A valuation assigns to each nonlogical primitive symbol, including the
individual variables x1, x2,.. . ,x1,..., an entity of the appropriate type
from the model M. A sentence (closed formula) is true if and only if
there is a valuation that satisfies it. Satisfaction is defined recursively
in the obvious way. For instance, (8x35[x1] is satisfied by a valuation
v if and only if there is a valuation which differs from v only for the
argument x1 and which satisfies S[x1]. Likewise, v satisfies if
and only if every valuation that differs from v only on x1 satisfies S[x1].
For propositional connectives, the usual truth-table conditions are
used to characterize satisfaction.
The satisfaction of an atomic formula is defined in the usual way.
For instance, R(x1, is satisfied by v if and only if > e v(R).
Putting all these stipulations together, we can indeed arrive at a recur-
sive truth definition.
The further details of Tarski's procedure do not have to concern us
here, but only the main features of his truth definition. If a truth
definition is formulated explicitly in a metalanguage, it is natural to
assume that that metalanguage contains elementary arithmetic. Then
one can use the normal technique of Gödel numbering to discuss the
syntax of the first-order language in question. If this language contains
a finite number of predicate and function symbols, then the logical
type of the valuation function v is essentially a mapping from natural
numbers (Gödel numbers of symbols and formulas) into the individ-
uals of the domain do(M) of the model in question. The truth predicate
itself which emerges from a Tarski-type treatment therefore has what
logicians call a form. In other words, it has the form of a second-
order existential quantifier (or a finite string of such quantifiers)
 THE FUNCTIONS OF LOGIC 15

followed by a first-order formula. All the quantifiers in this formula

range over either natural numbers or else over the (other) individuals
in the domain do(M) of the model in question. In other words, they are
first-order quantifiers.
Indeed, we can see the most important feature of Tarski-type truth
definitions for first-order languages: the definition itself is formulated
in a second-order language, that is to say, in a language in which one
can quantify over valuations. This feature of Tarski's truth definition
was not an accidental one, as Tarski himself proved. he proved
that, given certain assumptions, a truth definition can be given for
a language only in a stronger metalanguage. This is Tarski's famous
impossibility result. It is closely related to Gödel's (1931) incomplete-
ness results. Indeed, it has been established that Gödel first arrived at
his incompleteness results by discovering the undefinability of arith-
metical truth in a first-order arithmetical language.
The details of Tarski-type truth definitions and of Tarski's impossi-
bility result are familiar from the literature, and therefore will not be
elaborated here (cf. Ebbinghaus, Flum, and Thomas 1984, Ch. 3;
Mostowski 1965, Ch. 3).
Truth definitions are important for several different reasons,
though those reasons are not all equally directly relevant to the
foundations of mathematics. Without the notion of truth there is little
hope of capturing such basic concepts of logic as validity (truth in
every model) and logical consequence. Furthermore, a truth definition
is calculated to specify the truth-conditions of different sentences S.
These truth-conditions are closely related to the notion of meaning, in
the sense of sentence meaning. For what our assertively uttered
sentence S in effect says can be paraphrased by saying that what
S asserts is that its truth-conditions are satisfied. Thus to know the
truth-conditions of S is to know what S means.
The only major qualification needed here has already been made.
Truth-conditions deal with sentence meaning, not symbol meaning.
The latter has to be treated separately, and it must be taken for granted
in the formulation of truth-conditions and truth definitions. This point
is useful to keep in mind. Some philosophers have, for instance, tried to
criticize Tarski-type truth definitions for illicit reliance on the concept
of meaning. The reliance is there, but it is not illicit because it is not
circular. Tarski's project is nothing more and nothing less than to define
truth-conditions in terms of symbol meaning, that is, sentence meaning
in terms of symbol meaning. There is nothing wrong in such an attempt.
16 THE PRINCIPLES OF MATHEMATICS REVISITED

Tarski's and Gödel's negative results have been taken to have

major philosophical and other general theoretical implications. The
undefinability of such basic metalogical concepts as truth, validity
and logical consequence on the first-order level shows that ordinary
first-order logic is in an important sense not self-sufficient. Different
philosophers have sought to draw different conclusions from this
failure. Some rule out metalogic as a purely logical discipline,
maintaining that it belongs to set theory and hence to mathematics.
Others, like Putnam (1971), argue that logic must be taken to
comprehend also higher-order logic and perhaps, by implication,
parts of set theory. Generally speaking, Tarski's result seems to
confirm one's worst fears about the dependence of model theory on
higher-order logic and thereby on questions of sets and set existence.
It has even been alleged that this makes model theory little more
than a part of set theory. Indeed, the apparent dependence of
Tarski-type truth definitions on set theory is in my view one of the
most disconcerting features of the current scene in logic and in the
foundations of mathematics. I am sorely tempted to call it "Tarski's
curse". It inflicts model theory with all the problems and uncertain-
ties of the foundations of set theory. More generally, Tarski's curse
might be understood as the undefinability of truth for a given
language in that language (given Tarski's assumptions). The import-
ance of this negative result cannot be exaggerated. One of its first
victims was Carnap's grandiose vision of a single universal language
in which Hubert's and Gödel's formalization techniques would also
enable us to discuss its own semantics. In a more general perspective,
Tarski's undefinability result inevitably gives every model theorist
and most semanticists a bad intellectual conscience. The explicit
formal languages Tarski's result pertains to were not constructed to
be merely logicians' playthings. They were to be better tools, better
object languages for the scientific and mathematical enterprise. But
if a model theorist decides to study one of them, he or she will then
have either to use a stronger metalanguage for the purpose or else to
leave the metalanguage informal. In the former case, we have the
blind leading the blind or, more specifically, the semantics of a lan-
guage being studied by means of a more mysterious language, while
in the latter case the semanticist has simply given up his or her
professional responsibilities.
Second, one can try to apply Tarski's result to our actual working
language, called "colloquial language" by Tarski himself. The appli-
 THE FUNCTIONS OF LOGIC 17

cation is not unproblematic, for Tarski's theorem is formulated so as

to deal only with formal (but interpreted) languages satisfying
certain explicit conditions. But assuming that the conditions of
Tarski's result are satisfied by our ordinary working language, then
we cannot define truth for this language. The main characteristic of
our own actual language, duly emphasized by Tarski, is its universal-
ity. There is therrefore no stronger metalanguage beyond (or over) it
in which the notion of truth for this universal language could be
defined.
Hence presumably the assumptions on which Tarski's theorem
rests do not apply to natural languages. But if so, we have reached
a major negative metatheorem concerning what can be done by way
of an explicit semantical theory of ordinary language. And even if we
consider, instead of our everyday ("colloquial") discourse, the con-
ceptual system codified in the language of science, we still obtain
remarkable negative results. As Tarski himself expressed it on one
occasion (Tarski 1992, in a letter to Neurath, 7 September 1936):
There still remains the problem of universal language. It appears to me, that
this problem has been completely cleared up through the discussions of the
Varsovians (Lesniewski and myself) and also the Viennese (of Gödel and
Carnap): one cannot manage with a universal language. Otherwise one
would have to forego [sic] introducing and making precise the most
important syntactical and semantical concepts ("true", "analytic", "syn-
thetic", "consequence", etc.). One might now think that this circumstance is
of no special importance for the actual sciences, that therefore for the
purposes of actual sciences a single universal language can suffice entirely.
Even this opinion appears to me as incorrect, and for the following reason
in particular: to pursue actual sciences, something like physics, one must
have available an extended mathematical apparatus; now we know how-
ever, that for every language (therefore also for the presumed "universal
language") one can give entirely elementary number-theoretic concepts,
respectively sentences, which in this language cannot be made precise,
respectively cannot be proved.
More generally, the undefinability of truth can be considered
a paradigmatic instance of the important but frequently unacknow-
ledged assumption of the ineffability of semantics. This is not an
occasion to examine the full role of this assumption in the general
philosophy of the last hundred-odd years. Some aspects of its career
are examined in Hintikka (1988b) and in Hintikka and Hintikka
(1986, Ch. 1). However, it is relevant to note that this assumption
18 THE PRINCIPLES OF MATHEMATICS REVISITED

dominated the early development of contemporary logical theory. It

was embraced, by, among others, Frege, Wittgenstein, the Vienna
Circle during its "formal mode of speech" years, Quine and Church.
We have also seen that Tarski upheld a version of the same ineffabil-
ity thesis as applied to his "colloquial language". I have argued
elsewhere that the idea of the ineffability of semantics, in a somewhat
generalized form, is the mainstay of the methodology of hermeneuti-
cal and deconstructivist approaches to philosophy. A great deal thus
rests on the question of the definability of truth, especially on the
question whether we can define truth for a realistic working lan-
guage in that language itself. It is generally thought that all that there
is unproblematical to logic (at least to the kind of logic a mathema-
tician has occasion to use) is first-order logic. Everything else
depends on a set theory, formulated as a first-order axiomatic
theory. But set theory is not a part of logic, but a part of mathemat-
ics. Hence it cannot provide any absolute basis for the rest of
mathematics, and is itself beset with all the problems of set existence.
In this book, I will show that this defeatist picture is wrong. Not
surprisingly, a central part of this program concerns the ideas of
truth and truth definition. It turns out, however, that in order to do
so I will first have to revise some of our common ideas about truth
and other semantical ideas, as well as about the foundations of logic.
Tarski's result seems to suggest — and even to establish a nega-
tive answer to the problem of a realistic, philosophically interesting
definability of truth. As such, Tarski's result is but a member of a
family of apparently negative results which also include Gödel's
incompleteness results and Lindström's (1969) theorem, according
to which there cannot be a stronger logic than first-order logic
(satisfying certain conditions) which has the same nice properties as
first-order logic, principally compactness and the upward Löwen-
heim—Skolem property. Such negative results have dominated the
thinking of philosophers of mathematics in recent decades. In my
considered judgment, their importance has been vastly exaggerated.
A third function of logic in mathematics has cropped up in the
course of my discussion above. It is the use of first-order logic as
a medium of axiomatic set theory. This set theory is in turn supposed
to serve as the universal framework of all mathematics.
This conception of set theory as the lingua universalis (or at least as
a lingua franca) of mathematics is not universally accepted, and it is
in any case riddled with difficulties. These problems will be discussed
 THE FUNCTIONS OF LOGIC 19

more fully later in this book, especially in Chapter 8. As a preview of

what is to follow, the following problem areas are perhaps worth
mentioning here:

(I) The current conception of set theory as axiomatic and deductive

theory is quite different from the conception of set theory
prevalent earlier. Mathematicians like Hubert thought of set
theory, not as a theory among many, but as a super-theory,
a theory of all theories, which were thought of model-theoreti-
cally rather than deductively.
(ii) For reasons voiced so forcefully by Tarski, the conceptions of
a universal theory and universal language for all mathematics of
any sort is wrought with serious problems. For one thing, such
a universal theory must inevitably be deductively incomplete.
(iii) Axiomatic set theory likewise is inevitably incomplete deduct-
ively. Because of this incompleteness, the necessary assumptions
of set existence are extremely tricky.

However, this is not the end of the woes of axiomatic set theory. The
full horror story is not revealed until we take a closer look at the
different notions of completeness and incompleteness in Chapter 5
and then apply them to axiomatic set theory in Chapter 8.
Thus it is generally (but not universally) thought that the basic
difficulties in the foundations of mathematics are problems of set
existence, and more generally that the true core area of mathematics
is set theory.
At this point it may be instructive to return to the relation of the
first two functions of logic to each other. Obviously, the descriptive
function is the basic one. Needless to say, one can try to study
possible inferences by reference to their perceived appeal to us and
systematize those that seem to be acceptable. The current euphem-
ism for such an appeal to people's more or less educated prejudices is
"intuition". Such a study of our "logical intuitions" nevertheless
soon reaches a point where we need a firmer foundation for our
system of logical inferences. The people's, and even the philosophical
logician's, so-called intuitions have turned out to be fallible, and
even mathematicians have not reached unanimity as to which
principles of inference they should rely on. If an example is needed,
then the checkered history of the axiom of choice serves as a case in
point.
20 THE PRINCIPLES OF MATHEMATICS REVISITED

The right prescription for these inferential woes is to pay attention

to the descriptive function of logic. This is what shows which
putative logical inferences really preserve truth — and why. For
instance, why is it that we can infer S1 and from (S1 & S2)? Because
of the meaning of "&" as the connective which combines sentences in
such a way that both of them have to be true in order for the
combination to be true.
More generally, the basis of all model theory is the relation of
a sentence S to the set M(S) of its models. A putative inference from
to is then valid if and only if M(S1) M(S2).
Many variations are possible here, but they do not affect the main
point. Logical inferences are based on the meaning of the symbols
they involve, and purely logical inferences rely only on them.
I am painfully aware that distinctions of the kind I am making
here have been challenged by certain philosophers. The bases of their
theses are abstract views of meaning and the evidential status of
meaning ascriptions. Here I am talking about the concrete issues in
the foundations of logic and mathematics. And there the distinctions
I am talking about have an unmistakable import.
This can be illustrated by pushing my line of thought further.
From what! have said, it follows that a study of the inferences people
are inclined to draw qualifies as a genuine logic only in so far as those
inferences are based on the descriptive function of part of the logic in
question. What this means in practice can be seen for instance from
the so-called nonmonotonic logics. They may be magnificently
interesting and important, but they are not logics in the sense just
explained.
In the premises and conclusions of nonmonotonic inferences the
logical constants clearly have their normal meaning. Hence the
unusual principles of inference studied in these "logics" are not based
on the same model-theoretical meaning of logical notions as in our
ordinary logic. They are based on something else. And it is not very
hard to see what this "something else" is. For instance, in what is
known as circumscriptive inferences, the "something else" is the
assumption that the premises in some sense supply all the informa-
tion that is relevant to the subject matter. In some cases, this means
that all the relevant individuals are mentioned in the premises or that
their existence is implied by the premises (cf. Hintikka 1988c). Such
assumptions are interesting, and they can be important both in
theory and in practice, but they belong to the study of human
 THE FUNCTIONS OF LOGIC 21

communication, and not to logic. Aristotle would have dealt with

them in the Topics, not in the Analytics.
This is not merely a terminological matter. It makes a difference to
the way the "nonmonotonic inferences" are studied. For instance,
inferences by circumscription ought to be studied by reference to
their model-theoretical basis and not merely by postulating certain
new patterns of inference. What we have in "nonmonotonic logics" is
a clear instance of overemphasizing the deductive function of logic at
the expense of its descriptive function.
In reality, the core area of logic is the one where all the valid
inferences are based on the model-theoretical meaning of logical
constants. Whether one wants to call what goes beyond this core
area "logic" or not is a matter of intellectual taste. Someone else
might all it heuristics or formalized psychology of reasoning. The
important thing is to be clear about what one is doing.

Notes
1
For this famous quip, see Blumenthal (1935, p. 403) and Toepell (1986, p. 42).
A more pedestrian but at the same time perhaps more persuasive evidence for
Hubert's concern with the purely logical relations between the axioms and the
theorems is forthcoming from the development of Hilbert's thinking (Toepell
1986). Of importance, a major role in Hubert's development was played by
questions as to which axioms are those that results like Pascal's Theorem and
Desargue's Theorem depend or do not depend on.
2
This question of the status of the axioms of continuity was at the time a matter of
considerable interest. In the philosophy of physics, thinkers like Mach had
flaunted the idea of a presuppositionless, purely descriptive phenomenological
science typically describing the relevant phenomena by means of differential
equations. Against them, Boltzmann pointed out vigorously that even the
applicability of differential equations rested on nontrivial empirical assumptions
concerning the phenomena, namely, on assumptions of continuity and differen-
tiability (see, e.g., Boltzmann 1905). It would be interesting to know who
influenced whom in this emphasis on the empirical content of continuity assump-
tions.
Benacerraf (1973) in Benacerraf and Putnam (1983) attributes such a view to
David Hilbert. In my view, this is a radical misinterpretation of Hilbert's ideas; see
Hintikka (1996) and Chapter 9.
 2

The Game of Logic

The pivotal role of truth definitions in the foundations of logic and

mathematics prompts the question whether they can be freed from
the severe limitations which Tarski's impossibility result apparently
imposes on them — and whether they can be freed from other alleged
or real defects that critics claim to have found in them.
One defect for which Tarski-type truth definitions are blamed is
that of excessive abstractness. It has been alleged by, among others,
soi-disant intuitionists and contructivists, that such definitions mere-
ly characterize a certain abstract relationship between sentences and
facts. But such definitions leave unexplained, so this line of thought
goes, as to what it is that makes this relation a truth relation. In
particular, such abstract relations are unrelated to the activities by
means of which we actually verify and falsify sentences of this or that
language, whether a natural language or a formal (but interpreted)
one. As Wittgenstein might have put it, each expression belongs to
some language game which gives that expression its meaning. A
specification of truth-conditions does not provide us with such a
game, as Michael Dummett has doggedly argued time and again
(see, e.g., Dummett 1978, 1991).
Criticisms like these have a good deal to say for themselves. There
is much to be said for the fundamental Wittgensteinian idea that all
meaning is mediated by certain complexes of rule-governed human
activities which Wittgenstein called language games. Much of
Wittgenstein's late philosophy is devoted to defending this funda-
mental vision against actual or potential usually potential criti-
cisms, as is argued in Hintikka (1993b).
22
 THE GAME OF LOGIC 23

The kinds of criticisms I am talking about are often expressed in

terms of a need to replace a truth-conditional semantics by a verifi-
cationist one. The philosophers stressing such a need nevertheless
uniformly overlook the fact (pointed out in Hintikka 1987) that the
constant between truth-conditional and verificationist semantics is
not exclusive. For perfectly good truth-conditions can in principle be
defined in terms of the very activities of verification and falsification.
Indeed, it can be argued that such a synthesis is implicit in Wittgen-
stein's philosophy of language (ci Hintikka and Hintikka 1986,
Ch. 8). The deep idea in Wittgenstein is not that language can be
used in a variety of ways, most of them nondescriptive; rather,
Wittgenstein's point is that descriptive meaning itself has to be
mediated by rule-governed human activities, that is, by language
games. Moreover, the first language games Wittgenstein considered
were games of verification and falsification. In such games, meaning
can be both truth-conditional and verificationist, in that the truth-
conditions themselves are, as it were, created and maintained by
language games of verification and falsification.
Wittgenstein himself does not pay much systematic attention to
such Janus-faced language games, and his self-designated followers
have almost totally failed to appreciate them. In this chapter, I will
nevertheless show that the idea of language games can be made
a cornerstone of an extremely interesting logico-semantical theory.
In doing so, I will also uncover an important additional ambiguity,
this time an ambiguity affecting the notion of verification.
Furthermore, it will be shown more by example than by argu-
ment — that the involvement of humanly playable language games
does not make a concept of truth any the less objective or realistic.
These programmatic remarks, however, need to be put into
practice. There is an obvious way of dealing with the difficulties
about truth definitions. It is to confront the problem directly and to
ask: What are the relevant language games, anyway, that constitute
the notion of truth? How do we in fact verify and falsify sentences?
Let us take a simple example. How can (and must) you verify an
existential sentence of the following form?
(Bx)S[x] (2.1)
where S [x] is quantifier-free? The answer is obvious. In order to
verify (2.1), one must find an individual, say b, such that
S[b] (2.2)
24 THE PRINCIPLES OF MATHEMATICS REVISITED

is true. Here etymology serves to illustrate epistemology. In several

languages, existence is expressed by a locution whose literal transla-
tion would be "one can find". For the quality of the pudding its proof
may be in eating it, but when it comes to existence the proof of the
pudding is in finding it.
Propositional connectives can be treated in the same manner
almost a fortiori. For instance, if you have to verify a disjunction
(S1 V what you have to do is no more and no less than to choose
one of the disjuncts and s2 and verify it.
But what about more complex cases? I will use a formal (but
interpreted) first-order language L as an example. In order to speak
of truth and falsity in connection with such a language, some model
M of L (which you can think of either as "the actual world" or as
a given fixed "possible world") must be specified in which truth or
falsity of the sentences of L is being considered. The domain of
individuals of M is called do(M). That L has been interpreted on
M means that each atomic sentence (or identity) in the vocabulary of
L plus a finite number of individual constants (names of members of
do(M)) has a definite truth-value, true or false.
Consider now as an example a sentence of L of the form
(Vx)(By)S[x, y] (2.3)

What is needed for me to be in a position to verify (2.3)? The answer is

obvious. Clearly I must be able, given any value of x, say a, to find
a value of y, say b, such that S[a, b] is true. The only difference as
compared with (2.1) is that now the individual to be looked for
depends on the individual given to the verifier as the value a of the
variable x.
What is needed for the purpose of making the finding of a suitable
ha veritable test case of the truth of(2.3)? Clearly we have a test case
on our hands if the value a of x is chosen in the most unfavorable way
as far as the interests of the verifier are concerned. Descartes might
have conceptualized this idea by letting the choice of a be made by
a maim genie. It is nevertheless more useful to pick up a clue from
John von Neumann rather than René Descartes and to think of that
critical choice made by an imaginary opponent in a strategic game.
The natural way of generalizing and systematizing observations
of the kind just made is therefore to define certain two-person games
of verification and falsification. The two players may be called the
initial verifier and the initial fals(fier. I have called such games
 THE GAME OF LOGIC 25

semantical games and an approach to the semantics of both formal

and natural languages game-theoretical semantics, in short GTS.
The semantical game G(50) associated with a sentence S0 begins
with At each stage of the game, the players are considering some
sentence or other S1. The entire game is played on some given model
M of the underlying language.
On the basis of what has been said, the rules for semantical games
are thoroughly unsurprising:

(R. v) G(S1 v begins with the choice by the verifier of (i = 1 or

2). The rest of the game is as in G(51).
(R.&) G(51 & begins with the choice by the falsifier of S1(i = 1
or 2). The rest of the game is as in G(SJ.
(G.E) G((lx)S[x]) begins with the choice by the verifier of a mem-
ber of do(M). If the name of this individual is b, the rest of the
game is as in G(S[b]).
(G.A) G((Vx)S[x]) likewise, except that falsifier makes the choice.
(R. p.s) G( 5) is like G(S), except that the roles of the two players
(as defined by these rules) are interchanged.
(R.At) If A is a true atomic sentence (or identity), the verifier wins
G(A) and the falsifier loses it. If A is a false atomic sentence
(or identity), vice versa.

Since each application of one of the rules (R. V )—(R. —') eliminates
one logical constant, any game G(S) reaches in a finite number of
moves a situation in which (R.At) applies — that is, a situation which
shows which player wins.
It is to be noted that the name b mentioned in (R.E) and (R.A) need
not belong to L. However, because of the finite length of any play of
a semantical game, only a finite number of new names is needed in
the language necessary to cope with any given play of a semantical
game.
The rule (R.At) requires a special comment. It introduces an
apparent circularity into my treatment in that it contains a reference
to the truth or falsity of atomic sentences. However, as was pointed
out above, the concept of truth can be applied to the relevant atomic
sentences as soon as all the nonlogical constants of the given
sentence have been interpreted on the given model M with respect to
which the truth or falsity of S is being evaluated and on which G(S) is
26 THE PRINCIPLES OF MATHEMATICS REVISITED

being played. This interpretation is part and parcel of the definition

of M. It is determined by the meanings of the nonlogical constants
of S.
What (R.At) hence codifies is a kind of division of labor. The
game-theoretical analysis of truth takes the meanings of primitive
nonlogical constants of an interpreted first-order language for
granted. This fixes the truth-values of the relevant atomic sentences,
that is, of all sentences that can serve as endpoints of a semantical
game. What my characterization does is to extend the notion of truth
to all other sentences of the language in question.
The fact that I am thus restricting my task does not mean that
I consider a further model-theoretical analysis of meanings unnec-
essary. My only reason for the restriction is that otherwise the scope
of my enterprise would become unrealistically and unmanageably
large.
It is important to realize, however, that what is taken for granted
here is merely symbol meaning. The notion of sentence meaning is
inextricably tied to the notion of truth. As one might say, a sentence
means what it means by showing us what the world is like when the
sentence is true. Thus the notion of truth is the be-all and end-all of
sentence meaning in general.
To return to the rules of semantical games, it is important to
realize that the notion of truth is in no way involved in the explicit
formulation of the rules that govern the way moves are made in
semantical games. Only on a heuristic level can it be said that their
guiding idea is that S is true if and only if the applications of the game
rules can always be chosen by the initial verifier so as to be truth-
preserving. This heuristic idea leads also to the following game-
theoretical truth definition for applied first-order languages:
(R.T) S is true in M if and only if there exists a winning strategy for
the initial verifier in the game G(S) when played on M.
Correspondingly, the falsity of a sentence can be defined
(R.F) S is false in M if and only if there exists a winning strategy in
G(S) for the initial falsifier.
There is an apparently small but in reality most consequential
difference between the ideas represented here and those of construe-
tivists like Dummett. They are not averse to using notions from
strategic games in explaining their ideas but they give the game
 THE GAME OF LOGIC 27

analogy a wrong, or perhaps a far too simplistic, turn. For instance

Dummett writes (1978, p. 19):
The comparison between the notion of truth and that of winning a game
still seems to me a good one.

But this specific analogy is a bad one. The interesting analogy is

between the notion of truth and the existence of a winning strategy.
In this respect, semantical games differ essentially from the "games"
of formal proof. There the analogue to logical truth is winning in
a single play of the game of proof-searching. All of this illustrates the
subtleties — and the importance — of the apparently obvious truth
definition (R.T).
These semantical games and the truth definition based on them
can be extended in different directions. A similar approach can also
be used in the semantics of natural languages. The treatment of first-
order languages by means of semantical games is a paradigmatic
example of what has been game-theoretical semantics (GTS). An
extensive survey of GTS is presented in Hintikka and Sandu (1996).
This is not the occasion to advertise the merits of GTS. Its
applications for natural languages speak for themselves. They are
partially expounded in Hintikka and Kulas (1983, 1985) and in
Hintikka and Sandu (1991). The present book can be considered
another application and further development of the ideas of game-
theoretical semantics. This further development is not prompted by
a feeling of satisfaction with the existing theory, but by the questions
and puzzles it gives rise to. At the same time, an examination of these
open questions helps to put GTS itself into sharper focus.
The puzzles I am talking about are typically not difficulties for the
development of the theory, but curious phenomena that suggest that
further explanations are needed. In particular, my game-theoretical
truth definition and the semantical games on which it is based
require — and deserve a number of further comments.
(i) The truth definition uses the notion of a winning strategy. Here
the notion of strategy is used in its normal game-theoretical
sense, which can be understood on the basis of the everyday
sense of the word "strategy" but is stricter than that. In my sense,
a strategy for a player is a rule that determines which move that
player should make in any possible situation that can come up in
the course of a play of that game.
28 THE PRINCIPLES OF MATHEMATICS REVISITED

This notion of strategy is the central concept in the math-

ematical theory of games. By using it, any game can be assumed
to be represented in a normal form in which it consists simply of
the choice of a strategy by each player. Together, these choices
completely determine the course of a play of the game, including
an answer to the question of who wins and who loses. A winning
strategy for a player is one which results in that player winning
no matter which strategy is chosen by the other player or
players.
These notions are rather abstract, even though the starting-
points of the abstraction are familiar and clear. This abstract-
ness is a problem for a philosophical analyst, but it is also an
opportunity for further development. I will avail myself of this
opportunity in Chapter 10.
(ii) This kind of truth definition is not restricted to formal (but
interpreted) first-order languages but can be extended to various
other logical languages. It can also be extended to natural
languages. Even though quantifiers (quantifier phrases) behave
in certain respects differently in natural languages from the way
they behave in the usual formal first-order languages, a treatment
can be presented for them, too, in the same game-theoretical
spirit. What is especially important here is that the very same
truth definition applies there, too, without any changes. In other
words, even though the game rules for particular moves are
different, precisely the same characterization of truth and falsity
can also be used in the semantics of natural languages.
(iii) Thus we have reached a semantical treatment of first-order
languages and a characterization of truth which is in many ways
a most satisfactory one. Its naturalness can be illustrated by tel-
ling evidence. The naturalness of the game-theoretical treatment
of quantifiers is illustrated by the fact that it was put forward
completely explicitly by C. S. Peirce (Hilpinen 1983) and that it
has been spontaneously resorted to by logicians and mathema-
ticians practically always when the usual Tarski-type truth
definitions do not apply; and indeed sometimes when they do
apply, as for instance in the Diophantine games of number of
theorists like Jones (1974).

Some of the reasons why Tarski-type truth definitions fail will be

discussed in the later chapters of this book. Some others can be noted
 THE GAME OF LOGIC 29

here. One of them is due to the fact that Tarski-type truth definitions
start from the truth-conditions of the simplest (atomic) sentences
and work their way recursively to the complex ones. This presup-
poses that there always are fixed starting-points for such a pro-
cedure; in other words, that the formulas of one's language are
well-founded as set theorists would say. It is nevertheless possible to
introduce, use, and to study languages which do not satisfy this
requirement. Cases in point are game quantifier languages and more
generally the infinitely deep languages first introduced in Hintikka
and Rantala (1976). For such languages, it is impossible to give
Tarski-type truth definitions. In contrast, game-theoretical charac-
terizations of truth are perfectly possible. The only novelty is that
some plays of a game can now be infinitely long. But, for a game
theorist this is no obstacle to a definition of winning and losing. And
once these notions are defined, the rest of one's GTS operates as of
old.
More generally speaking, GTS is little more than a systematiz-
ation of the mathematicians' time-honored ways of using quantifiers
and of thinking of them. Careful mathematicians habitually use
locutions like "given any value of x, one can find a value of y such
that .. ." A typical context of such talk is the so-called epsilon-delta
definition of a notion like limit or derivative. An independent
testimony might be more persuasive here than my own words.
Speaking of the concept of limit, Ian Stewart writes in a recent book:
Finally• Karl Weierstrass sorted out the muddle in 1850 or thereabouts by
taking the phrase 'as near as we please' seriously. How near do we please?
He treated a variable, not as a quantity actively changing, but simply as
a static symbol for any member of a set of possible values. (Stewart 1992, p.
105)

In other words, Weierstrass used quantifiers to analyze the concept

of limit. But how did Weierstrass treat quantifiers? Stewart con-
tinues:
A function f(x) approaches a limit Las x approaches a value a if, given any
positive number g, the difference f(x) — lAs less than e whenever x — a is less
than some number ö depending on t. It's like a game: 'You tell me how close
you want f(x) to be to L; then I'll tell you how close x has to be to a.' Player
Epsilon says how near he pleases; then Delta is free to seek his own pleasure.
If Delta always has a winning strategy, then f(x) tends to the limit L.
(Stewart 1992, pp. 105—106)
30 THE PRINCIPLES OF MATHEMATICS REVISITED

The only word in Stewart's account that I do not like is 'like' (in "like
a game"), for what he describes is precisely the truth-condition for
the e —(5 quantifier definition of limit in game-theoretical semantics.
However, the game-theoretical treatment of truth in interpreted
first-order languages does not yet satisfy everything that philos-
ophers might legitimately ask of it. First, what my treatment yields
are truth-conditions for different first-order sentences. They are not
united into a genuine truth definition or truth predicate. Such a defini-
tion must be formulated in a metalanguage in which we can speak of
the syntax of the given first-order object language. Now we can
discuss the syntax of a given first-order language in another first-
order metalanguage, provided that the latter contains a modicum of
elementary arithmetic, for instance by using the well-known tech-
nique of Gödel numbering. A truth definition will then have to
consist in the definition of a number-theoretical predicate T(x) which
applies to the Gödel number of a sentence if and only if this sentence
is true in the model under consideration. Such definitions for inter-
preted first-order object languages cannot be formulated in a first-
order metalanguage. A fortiori, a truth definition for a given inter-
preted first-order language cannot be formulated in that language
itself.
In these respects, my game-theoretical truth-conditions do not
alone help us with the crucial problems as indicated in Chapter 1.
(iv) What the game-theoretical approach does tell us is what the
truth-conditions of first-order sentences are like. These truth-
conditions are formulated in terms of strategies of the two players.
Now the notion of strategy is itself amenable to a logical analysis and
to a formulation in logical terms. Suppose that a first-order sentence
S is in a negation normal form (i.e., all negation signs prefixed to
atomic formulas or identities). Since every first-order formula can be
brought to this form by an effective procedure, this is not a restrictive
assumption. Then a strategy for the initial verifier is defined by
a finite set of functions (known as choice functions or Skolem func-
tions) whose values tell which individual the verifier is to choose at
each of his or her moves. These moves are in S connected with its
existential quantifiers and disjunctions. The arguments of these
functions are the individuals chosen by the falsifier up to that point
in a play of the game G(S). Choice functions are second-order
entities, and their existence or nonexistence can be expressed by
a second-order sentence. In this way, the game-theoretical truth-
 THE GAME OF LOGIC 31

condition of S can be expressed by a second-order sentence S*, which

can be considered as a translation of S.
Speaking more explicitly, S* can be obtained through the follow-
ing steps:
(a) Let (ax) be an existential quantifier occurring in S within the
scope of the universal quantifiers (Vy1), (Vy,),.. . Replace
each occurrence of x bound to (3x) by f(y1, Y2, .
where f is
a new function symbol, different for different existential quanti-
fiers. Omit the quantifier (3x).
These functions are usually called in logic the Skolem functions
of S.
(b) Let (S1 v S2) be a disjunction occurring within the scope of the
universal quantifiers (Vy1), (Vy2),. (Vy,J. Replace the disjunc-
. .

tion by
v
(2.4)

where g is a new function symbol, different for different disjunc-

tions and different from the functions f mentioned in (a).
In this work, I will extend the usual terminology somewhat
and also call the g functions Skolem functions.
(c) Prefix the resulting formula by
. (2.5)

where f1,f2,... are all the functions introduced in (a) and

g1, g2,... all the functions introduced in (b).
The result S* will be called the second-order translation of S. It
expresses the game-theoretical truth-condition of S. It states how
the truth of S is connected with the semantical games of verifica-
tion and falsification described earlier.
(v) It may nevertheless be questioned whether the concept of truth
in general is really illuminated by the game-theoretical conditions.
The job that they do is to specify what quantificational sentences
mean by specifying their truth-conditions. The notion of truth is here
a merely auxiliary one, it seems. In other words, the first-order
semantical games seem to be language games for quantifiers, and not
for the concept of truth. This is apparently in keeping with the nature
of these games as games of seeking and finding. The conceptual
connection between quantifiers and the activities of seeking and
32 THE PRINCIPLES OF MATHEMATICS REVISITED

finding is easy to appreciate, but there does not seem to be any

equally natural link between semantical games and the notion of
truth in general. This can be thought of as being illustrated also by
the impossibility of defining truth of quantificational sentences in
those first-order languages which receive their meaning from my
semantical games. One can suspect here, as Wit tgenstein would have
done, that the concept of truth can only receive a use — and ergo
a meaning — in the context of certain other language games. Hence
a great deal of further work needs to be done here, unless we are
ready to acquiesce to all the different kinds of incompletenesses
that prevail here. I shall take up these questions later in this
work.
Another set of puzzles concerns the relation of GTS to construc-
tivistic ideas. In a sense, I seem to have realized the constructivists'
dream. I have shown the extremely close connection which obtains
between the game-theoretically defined concept of truth and the
activities (semantical games) by means of which the truth and falsity
of our sentence is established. The lack of such a connection in earlier
truth definitions has been the favorite target of constructivists. Now
this objection is completely eliminated.
But, paradoxically, none of the consequences which constructivists
have been arguing for follow from the game-theoretical truth
definition. In fact, as far as first-order logic is concerned, the game-
theoretical truth definition presented earlier in this chapter is equiv-
alent to the usual Tarski-type truth definition, assuming the axiom of
choice. Indeed, the second-order truth-condition of a given sentence
S defined earlier in this chapter is equivalent to 5, assuming the
normal (standard) interpretation of second-order quantifiers. Are
the constructivists' aims vacuous? Something strange is clearly going
on here. I will return to this puzzle in Chapter 10.
One particular facet of constructivistic ideas turns out to be
a teaser, too. Given the definitions of truth and falsity (R.T) and (R.F)
presented earlier, there is in general no reason to believe that the law
of excluded middle should hold. For according to (R.T) and (R.F) it
holds for a given sentence S only if either the initial verifier or initial
falsifier has a winning strategy in G(S). But we know from game
theory that there are many two-person zero-sum games in which
neither has a winning strategy. A simple example is a game in which
each of the two players chooses a natural number, independently of
the other's choice. The player with the larger number wins.
 THE GAME OF LOGIC 33

Games in which one of the two players has a winning strategy are
said to be determined. The assumption that the one or the other
player has a winning strategy is known as a determinacy (or deter-
minateness) assumption. Some such assumptions can be extremely
strong, as is known for instance from the different versions of the
axiom of determinacy in set theory (cf. here Fenstad 1971).
Hence the law of excluded middle cannot in general be expected to
hold if truth is defined game-theoretically. This should warm the
heart of every true constructivist, for the law of excluded middle has
long been their favorite target of criticism. Yet in that supposed core
area of contemporary logic, ordinary first-order logic, tertium non
datur does hold. An optimist might speak here of a fortunate coin-
cidence, while a pessimist might be led to wonder whether first-order
logic is really representative at all of the different intriguing things
that GTS shows can happen in logic. I will return to this matter, too,
later in Chapter 7. There, and elsewhere in the rest of this book, it will
turn out that ordinary first-order logic is a fool's paradise in that it
offers a poor and indeed misleading sample of the variety of things
that can happen in logic in general.
Yet another Pandora's box of puzzling questions concerns the
overall character of semantical games, especially as to what they are
not. Earlier, I referred to them as activities of attempted verification
and falsification. Yet this identification is by no means unproble-
matic. The very terms "verification" and "falsification" have to be
handled with great care. For what kinds of activities do we usually
think of as being involved in the verification and falsification of
propositions? Any ordinary answer is likely to include at least the
following two kinds of processes:
(a) logical (deductive) inferences
(b) different kinds of scientific inference, for instance, inductive
inferences
Yet it is important to realize that semantical games are different
from both these kinds of activities. How? Why? If semantical games
do not codify the ways in which we in our actual epistemological
practice verify and falsify sentences, then what light can they shed on
a realistic notion of truth?
This rhetorical question can be answered in two parts. First, the
activity of logically proving something is a language game of its own,
with its own rules different from the rules of semantical games.
34 THE PRINCIPLES OF MATHEMATICS REVISITED

Second, and most important, this language game is parasitic on

semantical games. Indeed, an attempt to prove that S1 logically
implies S2 can be thought of as a frustrated attempt to construct
a model ("possible world") in which S1 is true but S2 is not. If truth is
understood game-theoretically, then this means constructing
a model (world) M in which the initial verifier has a winning strategy
in the game G(S1) played on M but not in the game G(S2) played
likewise on M. For the purposes of ordinary first-order logic, the
rules for such attempted model construction can be read from the
rules for semantical games formulated above. The result is a com-
plete set of tobleau rules for first-order logic. By turning them upside
down we then obtain a somewhat more familiar looking set of
sequent calculus rules for first-order logic.
Hence ordinary deductive logic does not present an alternative to
semantical games for the process of verification and falsification. The
rules of deductive logic are themselves parasitic on the rules for
semantical games.
Somewhat similar points can be made of the processes that are
usually said to be the ways of verifying and falsifying propositions
for instance scientific hypotheses. They, too, are, conceptually
speaking, parastic on semantical games.
We have here a situation reminiscent of Meno's paradox, but
without a paradoxical conclusion. In order to find out whether
a proposition is true, one has to know what it means for it to be true.
What this means is that the language games of actual scientific or
other real-life verification depend on the semantical language games
that are constitutive of a sentence's truth in the first place.
For instance, assume that I have to verify in actual scientific or
everyday practice a functional dependence statement of the
form
(Vx)(Ry)S[x, y] (2.3)

(with x and y taking real-number values). In order to verify (2.3),

I naturally do not undertake to play a game against an imaginary or
real opponent, waiting for him or her or it to choose a real-number
value of x for me to respond to. What I typically do is to try to find
out what the function is that relates x to y in (2.1), for instance, by
means of a controlled experiment with x as the controlled variable
and y as the observed variable. Suppose that that experiment yields
the function g(x) = y as its result. This means that g(x) is a Skolem
 THE GAME OF LOGIC 35

function of (2.3), that is, that the following is true

(Vx)S[x, g(x)] (2.6)

But this means that g(x) is the strategy function (or a part of one)
which enables me to win the semantical game connected with (2.3),
that is, whose existence is the truth-condition of (2.3) according to
the game-theoretical truth definition.
This can be generalized. The task of actually verifying a sentence
S in the sense of getting to know its truth does not mean that the
initial verifier wins a play of the game G(S). Such a win may be due
merely to good luck or to a failure of one's opponent to pursue an
optimal strategy. Coming to know that S is true means finding
a winning strategy for this game. The latter enterprise makes sense in
principle only to someone who masters the game G(S) whose
strategies I am talking about. Hence the games of actual verification
and falsification are based on semantical games, which are concep-
tually the more fundamental kind of game. The games of actual
verification and falsification are secondary language games parasitic
on the semantical games I have explained. Confusing the two kinds
of games with each other is like confusing actual warfare in the
trenches with the task of a general staff planning a campaign.
In sum, even though a real-life process of verification does not in-
volve playing a semantical game, it aims at producing the very
information a winning strategy function whose existence guaran-
tees truth according to GTS.
These observations are instructive in several ways. They show,
first of all, that the very concepts of verification and falsification are
ambiguous. The terminology (and ideology) of language games
serves us well in explaining this point. There are on the one hand
those games which serve to define truth. I have shown that semanti-
cal games can play this role. But because of this very function, such
games cannot be thought of as games of establishing (proving,
verifying) a sentence whose truth-conditions are, so to speak, already
known, without falling prey to Meno's paradox. Are such games to
be called games of verification? Either answer can be defended, but
the interesting point is that both answers are possible.
On the other hand there are games where the truth-conditions of
the relevant sentences are taken for granted, and the game serves
merely to enable the inquirer to come to know that these conditions
are in fact satisfied. It is again a good question whether such games
36 THE PRINCIPLES OF MATHEMATICS REVISITED

should be called games of verification and falsification. An edge is

given to this question by the fact that such games are not merely
games of verification and falsification. They serve a purpose that
goes beyond that of characterizing the notions of truth and falsity.
They serve the purpose of coming to know the truth of the sentences
in question.
Perhaps both of these two kinds of games deserve to be related to
the notion of truth. In the former, what is at issue is the very notion of
truth, veritas, whereas in the latter we are dealing with some particu-
lar truth, verum.
What has been found strongly suggests that much of the recent
talk of verification, falsification, assertibility conditions and so forth,
is based on a confusion between the two related but importantly
different kinds of language games which I have just distinguished
from each other. The "real-life" or "real science" verification pro-
cesses can in fact be subsumed under the approach to inquiry and
reasoning in general which I have called the interrogative model of
inquiry. In it, an inquirer reaches new information through a series of
questions put to a suitable source of information ("oracle"), inter-
spersed with logical inferences from answers previously received
(plus initial premises). This model is really much more than a mere
"model"; it is a new overall framework for epistemology. I have
studied it in a number of publications,jointly with my associates (see,
e.g., Hintikka 1985, 1988a). The interrogative model of inquiry can
also be formulated as a game (language game). Prima facie, it does
not look altogether different from the semantical games being
examined here. Hence a confusion between the two is natural, albeit
inexcusable. In reality, the two kinds of games are conceptually
different from each other in extremely important ways. For one
thing, interrogative games depend, as was indicated, on rules of
logical inference. If they are changed, what can be verified by means
of interrogative inquiry will also change. In fact, because of the
epistemic element of interrogative inquiry, it may very well be argued
that the usual deductive rules of first-order logic have to be changed
in order to serve this purpose. If interrogative games were the logical
home of the notion of truth, then changes in the rules of logical
inference would affect the concept of truth. This would be in keeping
with the thinking of constructivists like Dummett, who propose to
discuss the "enigma" of truth by reference to rules of logical infer-
ence, and in fact advocate modifying the classical rules of first-order
 THE GAME OF LOGIC 37

deductive inference. But once we realize that the language games

which are constitutive of the notion of truth are semantical games,
and not the games of interrogative inquiry, then the basis of this
constructivistic line of thought vanishes. For semantical games are
more basic than the "games" of formal logical proof. You may change
the rules of deductive inferences by changing the rules of semantical
games. Later, in Chapter 10, I will give concrete examples. But there is
no way in evidence by which we could naturally do the converse.
Moreover, the move-by-move rules of semantical games of the kind
formulated above are so obvious that they offer scarcely any foot-
hold for a critic to try to change them. Even the changes that will be
considered experimentally in Chapter 10 will not affect these move-
by-move rules.
One can say more here, however. If you seriously examine interro-
gative games, it will become overwhelmingly clear that the interro-
gative inquiry they model can only be understood and studied
adequately as aiming at knowledge, and not merely at truth. Other-
wise, there is no hope of understanding such crucial notions as the
question—answer relationship, the criteria of conclusive answer-
hood, the presuppositions of different questions, and so forth. Most
important, without acknowledging the epistemic element in interro-
gative inquiry we cannot cope with the use of such inquiry for the
all-important purpose of answering questions, and not merely deriv-
ing given conclusions from given premises. Hence interrogative
inquiry must be viewed as an essentially epistemic language game,
aiming at knowledge and not merely at truth. And by the parity of
cases, we must view the sundry processes which pass as verification
and/or confirmation in ordinary usage and of which the interroga-
tive model is a generalized model.
Thus we have to distinguish from each other three different kinds
of language games. Their most important characteristics can be
summed up in the accompanying table on p. 38.
In this table, the question mark indicates that winning a play of
the game G(S) is not a hallmark of the truth of S. If it were, Meno's
paradox would be applicable and semantical games could not be
used to define truth because that notion would be needed to define
and to understand semantical games. As it is, truth is defined
indirectly as the existence of a winning strategy.
We might also speak of epistemic games as truth-seeking games
and of semantical games as truth-constituting games.
38 THE PRINCIPLES OF MATHEMATICS REVISITED

Interrogative
GAME TYPE Semantical game Proof game game

WHAT ? logical truth knowledge of

WINNING truths
SHOWS

CRITERION truth of the closure of the closure of the

OF output sentence game tableau game tableau
WINNING

OPERATIONAL- seeking and attempted questioning

IZATION finding counter-model plus logical
construction inference

There are closer relationships between the three kinds of games.

A proof "game" connected with Scan be thought of as an attempted
construction of a model in which S is not true, that is, in which there
is no winning strategy for the initial verifier. An interrogative game is
like a proof game except that answers to questions can be added as
new premises. In spite of these connections, it is philosophically very
important to distinguish them clearly from each other. Semantical
games are outdoor games. They are played among the objects of the
language one speaks, and they consist largely of the choices of the
two players between different objects. In contrast, proof games are
indoor games. They are played with pencil and paper, with chalk and
chalkboard, or these days most likely with a computer.
Note also that none of the three types of games is a dialogical
game in the literal sense of the word. The last two are "games against
nature" in a game theorist's sense. For instance, all that the inquirer's
opponent does in an interrogative game is occasionally to answer
questions.
Once you develop a theory of interrogative games of inquiry, you
will also see that constructivists like Dummett are making further
unwarranted and distortive assumptions concerning the character of
these "games against nature". Even though I have not found a
"smoking gun" statement in his writings, a number of things Dummett
says (e.g., 1978, p. 227) makes sense to me only if the empirical input
 THE GAME OF LOGIC 39

into the process of verification and falsification consists entirely of

observations whose expressions in language are assumed to be
atomic propositions. This fits very well into Dummett's paradigm
case of arithmetic, where the directly verifiable propositions are
numerical ones. In general, this assumption amounts to claiming
that the only answers one can hope to obtain in interrogative games
are particular truths. I have called this assumption the atomistic
postulate. A closer examination shows, however, that not only is this
assumption unwarranted, but that it has distorted philosophers'
entire view of scientific inquiry. In this respect, too, Dummett's ideas
of what the processes ("language games") are like which we actually
use to verify and falsify our sentences are inaccurate.
Of course Dummett and his ilk have not enjoyed the privilege of
considering my interrogative model in so many words. What they
have had in mind is some, more or less, unarticulated idea of how we
in practice verify and falsify scientific and everyday propositions. But
the interrogative model is calculated to capture all such activities.
Indeed, the only main assumption I have to make for the applicabil-
ity of the interrogative approach is that all new information im-
ported into the argument can be thought of as being obtained as an
answer to an implicit (or explicit) question addressed to a suitable
source of information. For this to be possible, little more is required
than that the inquirer is aware where he or she gets his or her
information from. Hence the interrogative model applies extremely
widely to all sorts of different processes of knowledge-seeking and
verification, presumably including what constructivists have in
mind. Hence they are not necessarily wrong about our actual
processes of inquiry. Their mistake is to assume that these processes
are constitutive of the notion of truth. They have missed the point of
Meno's paradox. To seek truth, one has to know what truth is.
Hence the usual scientific and/or everyday procedures of verifica-
tion cannot serve to define truth. This suggests that semantical
games might ultimately be used to define an actual truth predicate.
Whether my arguments show that semantical games can do so
depends on what will be found about such notions as Skolem
function. I will return to such questions in Chapter 6.
The distinction between semantical games on the one hand and
"games" of formal proof as well as interrogative games on the other
hand is thus, in any case, a crucial one. At the same time, the
distinction puts the nature of semantical games into sharper profile.
40 THE PRINCIPLES OF MATHEMATICS REVISITED

What consequences or suggestions does the game-theoretical

approach (and its success) to logic yield? The rest of this book is but
one long argument calculated to answer this question. One specific
suggestion might nevertheless be registered here. In order to see
what it is, we might consider what GTS says of truth of sentences of
the form
(Vx)(Jy)S[x, y] (2.3)

where S [x, y] is (for simplicity only) assumed not to contain quanti-

fiers. According to GTS, (2.3) is true if and only if there exists
a winning strategy for the initial verifier in the correlated game. Such
a strategy will tell among other things how to choose the value of
y depending on the value of x. This part of the verifier's strategy can
therefore be codified in a function f(x) = y. For such a function to
be part of a winning strategy means that the following sentence is
true:
(]f)(Vx)S[x, f(x)] (2.7)

Conversely, (2.7) is obviouly true if (2.3) is.

Hence (2.3) and (2.7) must be true simultaneously. But the equival-
ence of (2.3) and (2.7) is but a variant of the axiom of choice. Hence
GTS resoundingly vindicates this controversial axiom. Moreover,
the reasons for the validity of the axiom of choice are purely logical.
The axiom is but a corollary of the game-theoretical definition of
truth. Hence GTS realizes Hilbert's (1923, p. 157) bold conjecture,
not only vindicating the axiom of choice but showing that it is as
unexceptionable a truth as 2 + 2 =4.
In one sense, it is an understatement to say that the game-
theoretical viewpoint vindicates the axiom of choice because, by the
same token, it vindicates any equivalence between any old first-order
sentence and its second-order "translation" (truth-condition). Some
such equivalences are known to be stronger than the plain-vanilla
equivalences between (2.3) and (2.7), where S[x,y] does not contain
quantifiers or disjunctions (in its negation normal form). For in-
stance, such simple equivalences cannot enforce the existence of
nonrecursive functions while some of the more complex equivalen-
ces can (seô Kreisel 1953; Mostowski 1955). Hence we apparently
have here a hierarchy of stronger and stronger assumptions avail-
able to mathematicians. However, a separate investigation is still
needed to see whether the more complex equivalences are actually
 THE GAME OF LOGIC 41

stronger assumptions than the ordinary axiom of choice in a set-

theoretical context. What has been seen in any case shows that the
axiom of choice cannot be discussed in isolation. It is a member of
a series of assumptions that can all be justified in the same way.
Moreover, its justification or rejection is tied to more general
questions of the nature of truth, especially to questions concerning
the game-theoretical analysis of truth. In so far as I can vindicate the
general ideas of GTS, I can to the same extent vindicate the axiom of
choice. In Chapter 10 it will turn out that even certain interesting
constructivistic changes in my game-theoretical approach fail to
provide any reasons for rejecting the axiom of choice.
Gödel has advocated a cultivation of our mathematical intuition
in ways that would eventually enable us to see the truth or falsity of
different set-theoretical assumptions. In the spirit of Gödel's idea, it
may be said that game-theoretical semantics provides us with the
intuitions that are needed to vindicate the axiom of choice plus
a number of related assumptions. I believe that this is a fair way of
describing the situation. Moreover, the intuitions in question have
been seen to be solidly grounded on the ways in which we actually
deal with the concept of truth.
One can add a slightly different twist to the axiom of choice by
recalling that (2.7) is but a restatement of the truth of (2.3) from
a game-theoretical viewpoint. Hence the equivalence of (2.7) and
(2.3) is for a game-theoretical semanticist an instance of Tarski's
so-called T-schema. This schema is usually said to be of the form
3 is true i-+S

where 6 is a quote or a description of S. More generally, we can speak

of an instance of the T-schema in our extended sense whenever an
equivalence is set up between a sentence and its truth-condition. The
equivalence between (2.7) and (2.3) is a case in point. Hence the
axiom of choice is from the game-theoretical viewpoint closely
related to the T-schema, a relationship which illustrates the intu-
itiveness of the axiom of choice in the framework of GTS.
The same intuitiveness is likewise associated with the stronger
forms of the axiom of choice obtained by setting up equivalences
between more complex first-order sentences and their respective
game-theoretical truth-conditions.
A skeptical reader might very well remain unconvinced. But even
the most dogmatic skeptic — or should I say, the most skeptical
42 THE PRINCIPLES OF MATHEMATICS REVISITED

skeptic? — now knows what he or she must focus their critical

scrutiny on. All evidence for GTS is ipso facto evidence for the axiom
of choice. Any interesting criticism of the axiom of choice must be
based on an examination of the basic ideas of GTS, and in particular
of the game-theoretical definition of truth as the existence of a win-
ning strategy for the initial verifier in a semantical game. Indeed,
I will myself subject this definition to a closer examination in
Chapter 10. Hence the reader will have to remain in suspense, till the
last episode of the story of the perils and triumphs of the axiom of
choice.
Furthermore, it is not only the concepts of verification and
falsification (as well as the contrast between truth-conditional and
verificationist semantics) that are shown by GTS to need further
attention and further distinctions. It is likewise seen that the received
trichotomy of syntax — semantics — pragmatics has to be recon-
sidered, for where does game-theoretical semantics belong in this
classification? It deals with some of the basic semantical relations
between language and the world, and hence should belong to
semantics. At the same time, it deals with the uses of language in
certain language games. Hence Charles Morris (1938), who is pri-
manly responsible for the unfortunate syntax — semantics pragma-
tics trichotomy, would presumably relegate it to what Yehoshua
Bar-Hillel used to call the pragmatic wastepaper basket.
The right conclusion from my case study is that the traditional
distinction between semantics and pragmatics involves two separate
fallacies. First, crucial semantical relations like that of truth can exist
(and perhaps must exist) only in the form of certain rule-governed
human activities a la Wittgensteinian language games. Being activities
of language use, they would have to be pigeonholed as belonging to
pragmatics. In spite of this, their applicability to the question of the
truth of a given sentence in some particular world does not depend on
the human agents who implement these games. Their applicability
depends only on the structure of the "world" (model) in question.
Hence the contrast between semantics and pragmatics must be
defined in some other way, lest we end up saying that semantics is
a part of pragmatics. Personally I would not mind such a conclusion
very much, as long as the other fallacy muddling these two notions is
avoided. This fallacy consists in thinking that all theories of language
must inevitably involve the language users, and are therefore part and
parcel of the psychology and sociology of language. This view is as
 THE GAME OF LOGIC 43

fallacious as the claim would be that the study of syntax is a part of

graphology (or in the present day and age, of the technology of
computer displays and computer printouts). In both cases, we can
(and do) abstract from the idiosyncracies of the persons (and com-
puters) in question and concentrate on the general rules that govern
writing or language use, respectively. If a dramatization of this fact is
needed, then the reader is invited to think of how automata could be
programmed to play semantical games as they perfectly well could be
programmed to do. The hardware of such robots would be irrelevant
to the semantical and logical situation — only the software would
matter.
A further important contribution which the game-theoretical ap-
proach can make to the philosophical clarification of logic is an
account of how the bound variables of quantification really work.
What does such a variable "mean"? Such questions appear easy to
answer only as long as you do not raise them in so many words. It is far
from clear how a sentence in (say) a first-order formal notation is in
fact processed psycholinguistically, in spite of the frequent practice of
linguists to use in effect first-order logic as the medium of semantical
representation, or as the medium of representing logical form. One's
first idea is to think of bound variables as some sort of referring terms.
But even though are similarities between the variables of quantifica-
tion and ordinary singular terms, there are also differences.
Many of the same puzzles which arise in connection with the formal
variables of quantification also arise in connection with the quantifier
phrases of natural languages. This is not surprising, for one of the
main uses of formal quantification theory is supposed to be to serve as
the framework of semantical representation into which the quantifier
sentences of natural languages can be translated (cf. Russell 1905).
For instance, two occurrences of the same singular term refer to the
same individual. But two occurrences of the same quantifier phrase (or
the same variable) need not do so. This can be illustrated by the
contrast between the following sentences:
John admires John. (2.8)

Everybody admires everybody. (2.9)

If you want to find a sentence which is semantically parallel to (2.8), it

is not the syntactic analogue (2.9) to (2.8) but rather
Everybody admires himself. (2.10)
44 THE PRINCIPLES OF MATHEMATICS REVISITED

In game-theoretical Semantics, bound variables are placeholders

for the names of the individuals which the players choose during
a play of a semantical game. They are thus relative to such a play. They
can be said to stand for particular individuals within the context of
one particular play, but not absolutely. This helps to explain both the
similarities and the dissimilarities between quantificational variables
and ordinary singular terms. Further details are found in Hintikka
and Kulas (1985).
For instance, each occurrence of a quantifier phrase has to be
replaced by the name of an individual independently of others, even if
the two phrases are formally identical. This explains the semantical
difference between such sentences as (2.8) and (2.9).
In general, the game-theoretical approach facilitates a theoretically
satisfying account of the behavior of the variables of quantification
in formal languages and of quantifier phrases in natural languages,
including several aspects of anaphora in natural languages (see here,
e.g., Hintikka 1976a and Hintikka and Kulas 1985). This clarifica-
tion of the basic notations of the logic of quantification by means of
GTS is a natural continuation of the analysis of basic mathematical
concepts by means of quantifiers carried out by the likes of Weier-
strass. In this respect, the quote from Ian Stewart earlier in this
chapter is a most instructive one.
The possibility of a game-theoretical concept of truth which ac-
cords with our natural concept of truth, together with the distinction
between semantical (truth-conditioning) and interrogative (truth-
seeking) games also has profound philosophical repercussions. For
one thing, it shows what is true and what is false in pragmatist
conceptions of truth. What is true is that to speak of truth is not to
speak of any independently existing correspondence relations be-
tween language and the world. There are no such relations. Or, as
Wittgenstein once put it, the correspondence between language and
the world is established only by the use of our language — that is, by
semantical games. Truth is literally constituted by certain human
rule-governed activities.
What is false in pragmatist ideas about truth is the claim that the
relevant activities are the activities by means of which we typically
find out what is true — that is to say, verify, falsify, confirm, discon-
firm and so forth, our propositions. This claim is based on overlook-
ing the all-important distinction between truth-establishing games
(that is, semantical games) and truth-seeking games (that is, interro-
 THE GAME OF LOGIC 45

gative or perhaps other epistemic games). For all their formal

similarities and partial mutual determinations, these two kinds of
language games are fundamentally different from each other philo-
sophically and should be kept strictly apart. And when they are
distinguished from each other and their true nature recognized, then
the pragmatist claim is seen to be false. Our actual truth-seeking prac-
tices, whether or not they are relative to a historical era, epistemic or
scientific community, social class or gender, are not constitutive of
our normal concept of truth that is, of the concept of truth.
 3

Frege's Fallacy Foiled:

Independence-Friendly Logic

Several of the problems uncovered in the first two chapters above

can be traced to a common source. This source is a somewhat
surprising one. If there is an almost universally accepted dogma in
contemporary logical theory and analytic philosophy, it is the status
of the usual first-order logic as the unproblematic core area of logic.
"If I don't understand it, I do not understand anything," fellow
philosopher once said to me in discussing first-order logic. And when
on another occasion I expressed reservations concerning the claims
of the received first-order logic to be the true Sprachiogik, the logic of
our natural language, a well-known philosopher's response was:
"Nothing is sacred in philosophy any longer." He might have added,
"Or in linguistics, either," with a side glance at the role of the usual
quantificational (first-order) notation as the preferred medium of
representing logical form (LF) in Chomsky's linguistic theorizing.
Yet the received first-order logic, as formulated by the likes of
Frege, Russell and Whitehead, or Hilbert and Ackermann, involves
an important restrictive assumption which is largely unmotivated —
or perhaps rather motivated by wrong reasons. This assumption is
the hidden source of the problems registered above. For simplicity,
I will refer to it as Frege's fallacy or mistake. Though I will not
discuss the historical appropriateness or inappropriateness of that
label here, it is of some interest to note that Peirce would have found
it much easier to avoid this mistake — if he committed it in the first
place.
In order to diagnose Frege's mistake, we have to go back to the
basic ideas of first-order logic. This part of logic is also known as
46
 INDEPENDENCE-FRIENDLY LOGIC 47

quantification theory. It is in fact clear that the notions of existential

and universal quantification play a crucial role in serving both the
deductive and the descriptive function of logic, so much so that the
birth of contemporary logic is usually identified with the roughly
simultaneous discovery of the notion of quantifier by Frege and
Peirce.
But to say that first-order logic is the logic of quantifiers is to utter
a half-truth. If you consider the two quantifiers in isolation from
each other you are not going to obtain an interesting logic that is
to say, a logic which could serve the purposes of a nontrivial
language. All we would get is some mild generalization of syllogistic
logic, for instance, monadic predicate logic.
No, the real source of the expressive power of first-order logic lies
not in the notion of quantifier per se, but in the idea of a dependent
quant4ier. If you want to give a simple but representative example of
a quantificational sentence, it will not be a syllogistic premise with
just one universal or existential quantifier, but a sentence where an
existential quantifier depends on a universal one, such as
y] (3.1)

Without such dependent quantifiers, first-order languages would be

pitifully weak in expressive power. For instance, without them we
cannot express any functional dependencies in a purely quantifica-
tional language. The fact that existential quantification can be
hidden under a functional notation does not change the situation.
Such hidden quantifiers are forced out of the closet when function
symbols are eliminated in favor of predicate symbols.
Hence it is no exaggeration to say that to understand first-order
logic is to understand the notion of a dependent quantifier. If my
learned friend just quoted is right, if you do not understand the
notion of a dependent quantifier you do not understand anything.
One important service which game-theoretical semantics per-
forms is to explain the nature of the dependence in question. In
game-theoretical terms, this dependence is informational dependence.
It means that when a player of a semantical game makes a move
prompted by a dependent quantifier, then he or she knows what the
earlier move is like that was prompted by the quantifier it depends
on. In the technical jargon of game theory, the earlier move is in the
information set of the other (later) move. For instance, in playing
a game with (3.1), when the verifier chooses a value for y, he or she
48 THE PRINCIPLES OF MATHEMATICS REVISITED

knows which individual was chosen by the falsifier as a value of x and

can use this information in making his or her choice of the value of y.
This game-theoretical explanation of the nature of dependent
quantifiers is no mean feat. At least it is superior to such meaner
explanations of the nature of quantifiers as Frege's suggestion to
construe first-order quantifiers as second-order predicates (predi-
cates of one-place predicates), which serve to tell whether the given
predicate is empty or nonempty, exceptionless or exception-
admitting. Such explanations overlook the fact that quantifiers can
syntactically speaking be applied to complex or simple predicates
with more than one argument-place. This oversight in effect means
disregarding the possibility of dependent quantifiers, and hence
disregarding the true source of the expressive strength of first-order
languages. It is not only the case that there are certain arcane
phenomena in the semantics of natural language quantifiers that are
not done full justice to by considering quantifiers as higher-order
predicates. By so doing one just cannot bring out fully the most
important features of the behavior of quantifiers, as I will argue later
in this paper.
In the notation of the usual first-order logic, quantifier depend-
ence is indicated by parentheses. Each quantifier is assigned a pair of
parentheses indicating its scope. A quantifier occurring within the
scope of another is dependent on it. This is not the only way of
indicating dependence, however. For instance, in the gametheoretical
second-order translations of first-order sentences the dependence of
an existential quantifier on universal ones is shown by the selection
of arguments of the Skolem function that replaces it. For instance in

y, z, u] (3.2)

the dependencies are indicated by the tacit parentheses delineating

the scopes of the different quantifiers. In the second-order transla-
tion of(3.2), to wit,

(3f)(Jg)(Vx)(Vz)S{x, f(x), z, g(x, z)] (3.3)

the dependencies are indicated by argument selection rather than

parentheses. In general, the variable x bound to a given existential
quantifier (3x) is replaced in the second-order translation by a func-
tional term f(z1, z2,...), precisely when the universal quantifiers
within whose scope (ax) occurs are (Vz1), (Vz2),.. (I am once again
.
 INDEPENDENCE-FRIENDLY LOGIC 49

assuming that the sentence in question is in the negation normal

form.)
In contrast,
y, z, u] (3.4)

is equivalent to
f(x, z), z, g(x, z)] (3.5)

where the new argument z inf(x, z) shows that (3y) depends in (3.4)
also on (Vz) and not only on (Vx), as in (3.2).
But the two methods of indicating dependence and independence
are not equivalent. The Frege—Russell use of parentheses is more
restrictive than the use of Skolem functions. For instance, consider
the following second-order sentence which is on a par with (3.3) and
(3.5):

(3f)(3g)(Vx)(Vz)S[x, fix), z, g(z)] (3.6)

From the game-theoretical viewpoint, (3.6) is in perfect order. It

expresses the existence of a winning strategy in a semantical game
that is easily defined and implemented. The move-by-move rules of
this game are the same as in the GTS of ordinary first-order logic. In
this game, the initial falsifier chooses two individuals b, d to serve as
the respective values of x and z. Then the initial verifier chooses
a value of y, say c, knowing only the value of x; and chooses a value of
u, say e, knowing only the value of z, in both cases with an eye to
winning the game with S[b, c, d, e]. Admittedly, in order to imple-
ment such a game in a form which involves human players, we must
think of the initial verifier as a team of at least two human beings.
This is no hindrance to a game-theoretical logician, however. All
that it means is that my semantical games can occasionally look
more like bridge, where each "player" in the technical jargon of game
theory is a pair of human beings possessing somewhat different
information, than chess, where the game-theoretical usage of
"player" coincides with the colloquial one. (More generally speak-
ing, a game theorist's "players" can humanly speaking be teams of
several men or women.)
But although there is a well-defined semantical game that corre-
sponds to (3.6) and lends it a well-defined truth-condition, there is no
formula of the traditional first-order logic which is equivalent to
(3.6). This can be seen by trying to order linearly the quantifiers (Vx),
50 THE PRINCIPLES OF MATHEMATICS REVISITED

If we indicate the order by > we can argue as follows:

(Vz), (3y), (eu).
Since y =f(x) depends on x but not on z, we must have
('lx) > (Vy) > (Vz) (3.7)

Since u = g(z) depends on z but not on x, we must have

(Vz)>(3u)>(Vx) (3.8)

But (3.7) and (3.8) are incompatible, thus proving my point.

If we devise a notation which allows partial rather than linear
ordering, then we can easily formulate equivalents to (3.6). The
following two-dimensional formula fits the bill:

u] (3.9)

It will be shown further on, however, that partially ordered quanti-

fier structures are not the most natural way of capturing the
phenomenon of informational independence.
Thus, we have arrived at a diagnosis of Frege's mistake. This
diagnosis is virtually forced on us by the game-theoretical viewpoint.
The first question any game theorist worth his or her utility matrix
will ask about semantical games is: Are they games with perfect
information or not? Looked upon from our anachronistic game-
theoretical perspective, Frege in effect answered: Logic is a game
with perfect information. In spite of its prima facie plausibility, this
answer is nevertheless sub specie logicae not only arbitrary and
restrictive but wrong, in that in cuts off from the purview of logical
methods a large class of actual uses of our logical concepts. The fact
that those uses are descriptive rather than deductive ought not to
have deterred Frege, either. The crucial fact here is that the ideas
Frege was overlooking are implicitly an integral part and parcel of
a logic which everybody has practised and which almost everybody
thinks of as the unproblematic core area of logic, namely, the logic of
quantifiers.
Hence what is to be done here is clear. We have to extend our
familiar traditional first-order logic into a stronger logic which
allows for information independence where the received Frege—
Russell notation forbids it. The result is a logic which will be called
independence-friendly (IF) first-order logic. It is fully as basic as
ordinary first-order logic. It is our true elementary logic. Everything
you need to understand it, you already need to understand tradi-
 INDEPENDENCE-FRIENDLY LOGIC 51

tional Frege—Russell first-order logic. To quote my own earlier

exposition, IF first-order logic is a true Mafia logic: It is a logic you
cannot refuse to understand.
IF first-order logic can be implemented in several different ways,
some of which will be discussed further on. The formulation used
here involves one new item of notation, for which I will use the slash
'7". It will be used to exempt a quantifier or a connective (or, for that
matter, any logically active ingredient of a sentence) from the scope
of another one. Thus, for instance
(Vx)(Ry/Vx)S[x, y] (3.10)

islike (3.1) except that the choice by the verifier of a value of y must
be made in ignorance of the value of x. In this case, the former choice
might as well be made before the latter. In other words, (3.10) is
logically equivalent to
y] (3.11)
In this particular case, the slash notation does not yield anything that
we could not express without it. But in other cases, the independence-
friendly notation enables us to express what cannot be expressed in
the old notation. For instance, (3.6) can now be expressed on the
first-order level as
(Vx)(Vz)(By/Vz)(3u/Vx)S[x, y, z, u] (3.12)
It is seen here how the slash notation serves to omit potential
arguments of the Skolem (strategy) functions associated with a first-
order sentence.
The quantifier structure exemplified by (3.12) (or, equivalently, by
(3.9)) is known as the Henkin quantifier.
The slash notation, unlike the branching notation, can (and must)
also be applied naturally to propositional connectives. For instance,
in
(Vx)(S1[x]( v /Vx)S2[x]) (3.13)

the choice between S1 [x] and S2[x] must be made independently of

the value of x. Hence (3.13) is equivalent to
(Vx)S1[x] v (Vx)S2[x] (3.14)
But in other cases the slash notation is uneliminable (on the first-
order level) even when applied to propositional connectives. An
52 THE PRINCIPLES OF MATHEMATICS REVISITED

example is
(S1 [x, y, z] (v/Vx) S2[x, y, z]) (3.15)

In a second-order translation, (3.15) becomes

[x,f(x), z] & (g(z) = 0))
v (S2 [x,f(x), z] & (g(z) 0))) (3.16)
which shows the parallelism between (3.15) and (3.12).
It turns out that for the purposes of IF first-order we do not even
have to deal with all the different possible types of independence. It
also suffices to explain the slash notation for formulas in negation
normal form that is, in a form in which all negation-signs are
prefixed to atomic formulas or identities. Rules for transforming
formulas into negation normal form and out of it are the same
in IF first-order logic as in its traditional variant.
Thus we can explain the formalism of IF first-order logic as
follows:
Let S0 be a formula of ordinary first-order logic in negation
normal form. A formula of IF first-order logic is obtained by any
finite number of the following steps:
(a) If(3y)51 [y] occurs in S0 within the scope of a number of universal
quantifiers which include (Yx1), (Vx2)..., then it may be replaced
by
(3y/Vx1,Vx2,. . (3.17)

(b) If(S1 v 52) occurs in So within the scope of a number of universal

quantifiers which include (Vx1),(Vx2)..., then it may be replaced
by
(S1(v/Vx1,Vx2,...)S2) (3.18)

Languages using IF first-order logic are called IF first-order lan-

guages.
IF first-order languages will be my main conceptual tool in the rest
of this work. They deserve (and need) a number of further explana-
tions. Earlier discussions of independence-friendly first-order logic
are found in Sandu (1993) and Hintikka (1995b).
(i) A few remarks are in order concerning the formation rules for
IF first-order logic. They will turn out to be more interesting than
what might first appear.
 INDEPENDENCE-FRIENDLY LOGIC 53

In the form envisaged in (a) and (b), those formation rules for IF
first-order languages that introduce slashes are of an unusual form.
They take an already formed sentence (closed formula) and, instead
of using it as a building-block for more complex expressions, modify
certain subformulas of the expression in question. Moreover, these
changes inside the given expression are context-sensitive. What
changes are admissible depends on the context of the subformula in
question. For instance, whether (dy) can be replaced (salva well-
formedness) by depends on whether (Vx) occurs somewhere
in the larger formula with (3y) within its scope.
I could eliminate this defect (if it is a defect) of the formation rules
for IF logic by using a slightly different notation. In it, the indepen-
dence of (Jy) from (Yx) (where the former occurs in the scope of the
latter) is not expressed by replacing (dy) by but by replacing
(Vx) by It is obvious how this notation can be used in all the
different cases.
For instance, (3.10) can be written in the alternative notation as
y] (3.19)

while (3.12) would become

y, z, u] (3.20)

and (3.14) would become

(Vx//v)(S1[x] v S2[x]) (3.21)

The double-slash notation is in fact somewhat more natural than

the single-slash one. Unfortunately, I did not realize this fact when
I chose my notation, and at least for the purposes of this book it is
too late to change now. This does not matter very much, however,
for the interesting thing here is a comparison between the two
notations.
The naturalness of the double-slash notation is shown by the fact
that the formation rules for IF first-order languages would be
context-independent. Over and above this advantage, the double-
slash notation has other advantages. For instance, it would enable us
to formulate various simple logical laws. For instance, the equival-
ence of (3.13) and (3.14) will then be an instance of a distribution law
for (Vx// v):
(Vx//v)(S1[x] v S2[x]) (3.21)
54 THE PRINCIPLES OF MATHEMATICS REVISITED

will be equivalent to
(Vx)S1{x] v (Vx)S2[x] (3.22)

Even though the difference between / and //is in some sense only
a matter of notation, it will nevertheless turn out in Chapter 6 that it
illustrates certain important methodological problems.
(ii) Frege's fallacy is connected with an impotant unclarity con-
cerning the notion of scope. This notion is used widely both in logic
and in linguistics, and it is normally taken for granted. When it
comes to first-order languages (and to their counterparts in natural
languages), the very notion of scope nevertheless involves certain
serious problems. They are once again brought out into the open by
the game-theoretical approach. It shows at once that the traditional
notion of scope has two entirely different functions when it is applied
to quantifiers.
First, it is used to indicate the order in which the game rules are
applied to a given formula. In other words, the nesting of scopes
serves to indicate logical priority. This notion of scope will be called
priority scope. In the received notation of first-order logic, priority
scope is indicated by the inclusion of the scope of a quantifier (or
a connective) in the scope of another. The latter will then have the
right of way when it comes to game rule application. In order to serve
this purpose, the scopes of quantifiers and of connectives must be
partially ordered — that is, they must be nested. Furthermore, only
the relative order of the scopes matters and not their actual extent in
a sentence or discourse.
But the scopes of quantifiers are also supposed to indicate some-
thing else. When the scope of a quantifier (Qx) is indicated by a
pair of parentheses, then the segment of the sentence within those
parentheses is supposed to be where the variable x is bound to the
quantifier in question. These variables are often compared to
anaphoric pronouns. The scope of a quantifier is then the part of
a sentence where such a pronoun can have the quantifier in question
as its head. This kind of scope will be called the binding scope. The
distinction between priority scope and binding scope is related to the
difference between Chomsky's (198!) notions of government and
binding, and could be used to elucidate what Chomsky is trying to
get at in his theory.
Once the distinction is made, then it is seen that there is no reason
why priority scope and binding scope should always go together.
 INDEPENDENCE-FRIENDLY LOGIC 55

For instance, there is no reason why the scopes of two quantifiers (in
the purely formal sense of scope in logic) could be only partially
overlapping, even though the Frege—Russell notation requires that
they be either nonoverlapping or else nested (one of them contained
in the other). It can even be argued that the notion of binding scope is
not an unanalyzable one in natural languages. Furthermore, the
binding scope of a quantifier need not be a continuous segment of
a sentence or a discourse.
Once the situation is thought of in these terms, then it is seen that
the usual notion of scope as it is being actually employed is a royal
mess. This mess assumes a somewhat different complexion for
logicians and linguists. For logicians running together the two senses
of scope has been an unnecessary obstacle to their theorizing, but for
linguists it has actually harmed their theories, in that it has led them
to overlook certain telling phenomena of our Sprachiogik.
Frege's mistake can thus be seen as being a result of conflating the
two uses of parentheses. The most natural way of using them for the
purpose of indicating logical priority is to require that the scopes of
quantifiers are nested. But to do so precludes a merely partial
overlapping of quantifier scopes, even though such a violation of the
nesting requirement is a perfectly natural thing from the vantage
point of binding scope.
If Frege's needless restrictions on the use of scope indicators are
removed, we then obtain a first-order language that is as strong as IF
first-order language. If quantifiers are indexed together with the
corresponding parentheses, we could then for instance express (3.12)
as follows:
(Vx)1 y, z, u])1)2 (3.23)

Here the scope of the initial universal quantifier (Vx) (with an index
(Vx)1) consists of two segments of the formula in question. The
former is

and the latter is

1(S[x, y, u, z])1

where in both segments the subscripts merely indicate that the

parentheses in question belong to (Vx) = (Vx)1. In contrast, the scope
56 THE PRINCIPLES OF MATHEMATICS REVISITED

of (Vz) (Vz)2 is a continuous one, consisting of

2((3u)S[x, y, z, u] )2
But by thus lifting the quantifier (au) from the scope of (Vx) we
accomplish precisely the same as is done in my notation by writing it
as
This shows how fundamental Frege's fallacy was. It could be
corrected even without introducing any new symbols into our logic,
by merely changing the use of the innocent-looking device of par-
entheses. (The subscripts are not any more meaningful in themselves
than connecting lines.) At the same time (3.23) shows that such a
procedure would be so clumsy as to be impracticable.
No matter which notation is used, the step from ordinary first-
order logic to IF first-order logic is little more than the recognition of
the fact that the two kinds of scope do not have to go together. For
instance, in our slash notation the formal scope (indicated by
parentheses) can be taken to be the binding scope, while the slash
notation involves a recognition of the fact that priority scope does
not automatically follow the binding scope.
(iii) A special case of the phenomenon of informational indepen-
dence has been studied by logicians ever since Henkin (1961) under
the heading "branching quantifiers" (cf. e.g., Barwise 1979, Enderton
1970, Walkoe 1970). The connection between the two can be seen
from an example. As was noted previously, our familiar Henkin
quantifier sentence
y, z, u] (3.12)

can be written, instead of the slash notation, in a nonlinear notation

as follows:

z, (3.9)
} S[x, y, uIJ
Here the mutual dependencies and independencies of different quan-
tifiers are indicated by a partial ordering of those quantifiers. This
idea can easily be generalized. However, it does not yield a really
satisfactory analysis of the situation. For one thing, the branching
notation has not been used in connection with propositional con-
nectives, which can also exhibit informational independence, until
very recently and for the first time apparently in Sandu and
Väänänen (1992).
 INDEPENDENCE-FRIENDLY LOGIC 57

Also, the partial ordering of quantifiers corresponds in a game-

theoretical approach to a partial ordering of their information sets
with respect to class-inclusion. But in general there is no reason
whatever to think that the information sets of different moves in
a game are even partially ordered. Hence branching or partially
ordered quantifiers are not representative of the general situation.
Admittedly, it turns out that in the special case of quantifiers all their
possible patterns of dependence and independence do reduce to a
suitable partial ordering. However, this is not true in general.
Furthermore, one has to handle negation very cagily in the context
of informationally independent quantifiers, as will be shown in Chap-
ter 7. It is not even clear that all the earlier treatments of branching
quantifiers handle negation in a consistent manner.
Most importantly, branching is a matter of the interplay of quanti-
fiers, and not of the interpretation of quantifiers taken one by one. For
this reason, generalizations of the notion of quantifier taken in
isolation from other quantifiers do not help essentially in understand-
ing branching and other informationally indpendent quantifiers. For
instance, in spite of the other kinds of success that the theory of
generalized quantifiers has enjoyed, it has not helped essentially to
understand branching quantifiers. Admittedly, there exist treatments
of inlormationally independent quantifiers in the tradition of general-
ized quantifiers (see, for example, Westerstahi 1989 and Keenan and
Westerstahl 1996). They do not bring out the generality of the
phenomenon of independence, however, or even reach a fully satis-
factory treatment of the semantics of informationally independent
quantifiers.
However, some of the general issues that are prompted by the
phenomenon of informational independence have been discussed by
philosophers and logicians in the earlier literature in the special case of
branching quantifier structures. Among other things, several of the
most important perspectives of IF first-order languages are generaliz-
ations of the properties of branching quantifiers. I will return to them
later in this chapter.
(iv) The best measure of the naturalness of IF first-order logic is the
simplicity of the (game-theoretical) semantics that governs it. How are
the semantical game rules for IF first-order languages related to those
of ordinary first-order languages? The relationship is an interesting
one and a simple relationship at that. It is one of identity. The rules
for making moves in a semantical game in IF first-order logic are
58 THE PRINCIPLES OF MATHEMATICS REVISITED

precisely the same as those used in ordinary first-order logic, except

that imperfect information is allowed. The extensive form of a game
connected with an IF sentence is the same as that of the corresponding
ordinary first-order formula, except for the fact that the information
sets of some of the moves do not contain all the earlier moves. (The
information set of a move includes all and only those earlier moves
which are known to the player making the move.)
Moreover, it is important to realize that this failure of perfect
information has nothing subjective about it. It is not a matter of
a limitation of human knowledge or human ingenuity. It is a feature of
the combinatorial situations that can come up in a semantical game.
The truth or falsity of a sentence S of an IF language is an objective
fact about the world which S can be used to speak of. This objectivity
of truth in IF first-order logic will be further illustrated in Chapter 7.
(v) It is to be noted that the irreducibly independence-friendly
formulas are typically stronger than the corresponding formulas of
traditional first-order logic. The reason is clear from a game-
theoretical perspective. From this viewpoint, a sentence asserts the
existence of a winning strategy for the initial verifier in the correspond-
ing game. The presence of informational independence means the use
of fewer arguments in some of the strategy functions; in other words,
the use of a strategy based on less information. Hence the existence of
such a strategy is a stronger claim than the existence of the analogous
strategy based on more information. It thus turns out that IF first-
order languages are not only stronger than ordinary ones, but ex-
tremely strong.
The greater expressive strength of IF first-order logic as compared
to its older brother has a concrete meaning. It does not mean only, or
in the first place, greater deductive resources. What it means is that
there are classes (kinds) of structures (even finite ones) that simply
cannot be described by means of the resources of ordinary first-order
logic. Particular examples will be discussed in Chapter 7. This kind of
advantage of IF first-order logic is naturally of great potential interest
both theoretically and in practice — that is, in view of the role of logic
in mathematical theorizing.
An example may help the reader to appreciate this point. Consider
the following formula

(Vx)(Vz)(3y/Vz)(Bu/Vx)(((x =
& H(x, y) & H(z, u) & — H(x, u) & H(z, y)) (3.24)
 INDEPENDENCE-FRIENDLY LOGIC 59

It is easily seen to be equivalent to

z) (f(x) f(z))) & H(x, f(x))

& H(z,f(z)) & H(x,f(z)) & H(z,f(x))) (3.25)

In other words, what (3.24) says is that one can correlate with each
x an individual f(x) such that the correlates of two different individ-
uals are different and differ from each other in the way indicated by
(3.25). Thus, if H(x, y) is read "y is a hobby of x's", then (3.24) says
something like "everybody has a unique hobby."
It can be shown that (3.24)—(3.25) do not reduce to any expressions
of an ordinary first-order language. What (3.24) does is therefore to
capture as its models a class of structures with a relatively simple
characteristic common feature. Yet this class of structures cannot be
captured by ordinary first-order logic. Thus the basis of the greater
strength of IF first-order logic is its greater representational capacity,
and not its ability to offer to a logician new deductive gimmicks.
(vi) In spite of their strength, IF first-order languages have a num-
ber of welcome properties. Among other results, IF first-order logic
has the following properties:
(A) It is compact. This fact is easily proved. Let a be an infinite set of
formulas of IF first-order logic. Let a* be the set of second-order
translations of all the members of a, with different initial existential
function quantifiers using different function variables. Let a** be the
result of omitting all these initial function quantifiers from all the
members of This a** is a set of ordinary first-order formulas.
Clearly, if one of a, a* and a** is satisfiable, then all these are. Because
ordinary first-order logic is compact, a* * is consistent if and only if all
its finite subsets are consistent. But clearly such a subset is
consistent if and only ii its parent subset x of a is consistent. And x is
consistent by hypothesis.
(B) IF first-order logic has the (downwards) Löwenheim—Skolem
property. In order to prove this let a be a finite set of formulas of an IF
first-order language, and let a* and a** be formed as before. Clearly a,
a* and a** are satisfiable in the same domains. But in virtue of the
downwards Löwenheim—Skolem theorem for ordinary first-order
logic, a** is satisfiable in a countable domain if satisfiable at all.
(C) The separation theorem holds in a strengthened form. To see
this, assume that a and are sets of formulas of an IF first-order
language satisfying the following conditions:
60 THE PRINCIPLES OF MATHEMATICS REVISITED

(i) a and r are both consistent.

(ii) a U r is inconsistent.
Then there is a formula of the corresponding ordinary first-order
language (i.e., a slash-free formula) S such that
(iii)tI—S
(iv) All the nonlogical constants of S occur in the members of both
a and 'r.
The separation theorem can be proved by transforming a and t into
a* and a** as well as into t4' and as before. It is easily seen that the
separation theorem of ordinary first-order logic applies to a** and
r**, yielding an ordinary first-order formula S0 as the separation
formula. This S0 is easily seen to serve also as the separation formula
claimed to exist in the theorem.
(D) IF first-order languages admit of the Skolem normal form in
a stronger sense than ordinary first-order languages. In them, the
Skolem normal form S2 of a given formula S1, which is a formula of
the form
(3.26)

is satisfiable if and only if S1 is satisfiable. But S1 and s2 need not be

logically equivalent nor have the same vocabulary. But in IF first-order
logic, each formula can easily be seen to have an equivalent formula
which is like (3.26) except that every (3Yk) is replaced by

(3y/Vxkl,Vxk2,...) (3.27)

where

Xk2,. . .} c {x1, x2,. .}

.
(3.28)

The given formula S1 can be transformed into S2 by first transforming

it into its prenex form, that is, into a form where all quantifiers are at
the beginning of the formula. Then one can extend the scope of
a universal quantifier (Vx) over a number of existential quantifiers
(9y2),. . . if one replaces them by The same
can be done to several universal quantifiers at the same time. This
yields the desired form.
In results like (d) we can also include within their scope the initially
tacit "Skolem functions" associated with disjunctions. For instance,
the Skolem normal form of
{x, y, z] ( v /Vx)S2[x, y]) (3.29)
 INDEPENDENCE-FRIENDLY LOGIC 61

can be

Vz)((S1 [x, y, z]&(u = 0))

v(S2[x,y] (3.30)

(E) Beth's theorem on definability holds in IF first-order logic in

a sharper form. This theorem says that an "implicit" definability of
a constant, say a one-place predicate P. on the basis of a consistent
theory T[P], entails its explicit definability on the basis of the same
theory. In other words, assume that

(T[P] &T[P'] ) (Vx)(P(x)4—*P'(x))

F— (3.31)

where P' is a new one-place predicate constant. Then for a suitable

definiens (complex formula) D[x] with x as its only individual van-
able but without P (or P') we have
T[P] (3.32)

This result can be proved for IF first-order logic on the basis of the
separation theorem in precisely the same way as the corresponding result
for ordinary first-order logic. From the proof (and the separation the-
orem) it can be seen that the definiens D[x] can always be chosen to be
a formula of ordinary first-order logic.
(F) Just as every ordinary first-order sentence was seen in Chapter 2
to have a second-order translation, in the same way each sentence of
an IF first-order language can be seen to have a similar second-order
translation. The translation is obtained in the same way as in ordinary
first-order languages (see Chapter 2). The only difference is that the
arguments x1, x2,. . . are dropped from the function correlated with
Vx2,...) or with (v Vx2,...).
(G) As in ordinary first-order logic, the second-order translations of
IF first-order sentences are of the form. Unlike ordinary first-order
logic, however, IF first-order logic also allows for the converse
translation. In other words, each sentence has a translation (logical
equivalent) in IF first-order language. Even though this result is
known from the literature, it is still useful to see how it can be
established. 1
A sentence has the form of a sequence of second-order existential
quantifiers followed by a first-order formula. First, we replace all
predicate variables by function variables. This can be done one by one.
62 THE PRINCIPLES OF MATHEMATICS REVISITED

For instance, suppose we have a formula of the form

(3.33)

where X is a one-place predicate variable. Then we can replace (3.33)

by the formula
(3.34)

where S'[f] is obtained from SEX] by replacing every atomic formula

of the form X(y) or X(b) by
f(y)=o (3.35)
or
f(b)=O (3.36)

respectively.
Alter eliminating all predicate variables in this way, all the formulas
we need to deal with are of the form
. . (3fk)(Vxl)(Vx2).. . (Vx1)S[f1, fk' x1, x2,. . . , x1]
(3.37)

where SE] is a quantifier-free ordinary first-order formula.

Here (3.37) can be expressed in an equivalent form in the IF
first-order notation if the following two conditions are satisfied:
(a) There is no nesting of functionsf,.
(b) Each occurs in S[] with only one sequence of arguments,
say, with the variables x11, missing from its set of arguments.
For then we can replace (3.37) simply by
(Vx1)(Yx2). . . (Vx,). . . . .)S* (3.38)

where S* results from replacing each f, followed by its arguments by y1.

What has to be shown is therefore how an arbitrary (3.37) can be
brought to a form where (a)—(b) are satisfied.
I will illustrate this transition by means of two examples. First,
consider a formula of the form
f(x2)] (3.39)

Here (a) is satisfied. In order to bring it to a form where (b) is also

satisfied, we can rewrite it as
= x2) (f(x1) = g(x2))) & S[f(x1),g(x2)])
(3.40)
 INDEPENDENCE-FRIENDLY LOGIC 63

This is easily seen to be equivalent to

= x2) D (Yi = Y2)) & S[y1, .v2])
(3.41)

For the second example, consider a formula of the form

(3.42)

Here condition (a) is not satisfied. In order to meet this condition, we

can consider the formula
=f1(x1)) & (X3 =f2(x2))) D S[x3])
(3.43)

which satisfies both (a) and (b). Indeed, (3.43) is logically equivalent
to

1Vx2Vx3)
(((x1 =x3)i-4y1 =y3))&((x2=x4)4—4y2=y4)&
(((Y2=X2) & (Y2 = x3))DS[y2])) (3.44)

In neither example are the other singular terms occurring in S

affected. We can therefore eliminate one by one all the violations of
the requirements (a) and (b) mentioned earlier in this section. This
process terminates after a finite number of steps, yielding the desired
translation.
(vi) At this point, it is in order to look back at the precise way the
information sets of different moves are determined in semantical
games with IF sentences. The small extra specification that is needed
is that moves connected with existential quantifiers are always
independent of earlier moves with existential quantifiers. This as-
sumption was tacitly made in the second-order translation of IF
first-order sentences explained earlier in this chapter. The reason for
this provision is that otherwise "forbidden" dependencies of existen-
tial quantifiers on universal quantifiers could be created through the
mediation of intervening existential quantifiers.
This point can be illustrated as follows: If the provision is heeded,
the second-order translation of
y, z, u] (3.45)
64 THE PRINCIPLES OF MATHEMATICS REVISITED

will be
f(x), z, g(z)J (3.46)
If it is not heeded, the translation will be
f(x), z, h(z, f(x))] (3.47)
The two second-order translations (3.46) and (3.47) are not equiv-
alent. Obviously (3.46) is at least as strong as (3.47). Without the
provision, there is no way of capturing the stronger statement (3.46)
in IF first-order logic. Hence the extra provision has to be adopted.
I will not try to prove this impossibility here — that is to say, the
impossibility of expressing (3.46) in an IF notation without the
convention of omitting the initial verifier's moves from information
sets. What can be easily shown is that we get a different interpreta-
tion with and without the convention. For the purpose, it suffices to
provide an example where the second-order translations (3.46) and
(3.47) are not equivalent. Such an example is
(Vx)(Vz)(3y/Vz)(3u/Vx)((y x)&(u z) & ((x = = u)) (3.48)
With the convention, the nonempty models of (3.48) are all infinite
ones. Without the convention, (3.48) is satisfied in the following
two-element model M:
do(M) = (3.49)
f(a)—=b f(b)=a
g(a,b)=b g(b,a)=a
g(a,a)=b g(b,b)=a
It is easy to see that f and g satisfy the formula
(Vx)(Vz)((x = = g(z, f(x)))) (3.50)

which is a subformula of the translation of (3.48) to a second-order

form without the convention. This translation is of course
= z)*-÷(f(x) = g(z,f(x)))) (3.51)
All this illustrates the indispensability of the convention for a reason-
able semantics of IF first-order logic.
(viii) It is important to understand precisely what the relation-
ship is between IF first-order logic and its traditional order brother.
The following are among the most relevant aspects of this relation-
ship:
 INDEPENDENCE-FRIENDLY LOGIC 65

(a) Ordinary first-order logic is a special case of IF first-order logic.

(b) No ideas are involved in IF first-order logic that were not needed
for the understanding of ordinary first-order logic.
(c) Technically speaking IF first-order logic is a conservative exten-
sion of ordinary first-order logic.
IF first-order languages nevertheless differ strikingly from the
ordinary first-order logic in several ways. Three of them are so
important as to serve naturally as lead-ins into the next three
chapters. These three forms of unfamiliar logical behavior are the
following:
(a) IF first-order logic does not admit of a complete axiomatization.
(b) IF first-order logic does not admit of a Tarski-type truth defini-
tion.
(c) The law of excluded middle does not hold in IF first-order logic.
Combined with the status of IF first-order logic as the natural
basic core area of logic, these observations will be seen to put the
entire foundations of logic in a new light, as I will show in the
subsequent chapters of this book. Among the minor new bridges one
has to cross here is the precise definition of notions like logical
consequence and logical equivalence. The novelty is due to the
failure of the law of excluded middle, which forces us to associate
with each sentence S not only the class M(S) of all models in which it
is true, but also the class M(S) of all models in which it is false, for the
latter is no longer a mere complement of the former. This prompts
the question: Are two sentences logically equivalent if and only if
they are true in the same models (on the same interpretation), or are
they to be so called if and only if they are true and false in the same
models? In this work, I will consistently opt for the former alterna-
tive, and for its analogies with other concepts of the metatheory of
logic.
Indeed, an Argus-eyed reader may already have noticed that
I tacitly relied on such an understanding of the basic metalogical
notions like logical equivalence in the preceding arguments that
established some of the basic logical properties of IF first-order
languages. Moreover, and most importantly, an IF first-order sen-
tence and its second-order translation are true in the same models,
but they are not false in the models. This follows from the fact
that tertium non datur holds ordinary second-order logic, but not
in IF first-order logic.
66 THE PRINCIPLES OF MATHEMATICS REVISITED

By means of IF first-order logic we can overcome most of the

limitations of ordinary first-order logic which were pointed out in
earlier chapters. IF first-order logic is thus the key tool in my
approach to the foundations of logic and mathematics. Earlier,
I argued for the status of IF first-order logic as the core area of all
logic. This special status of the notion of informational independence
can be illustrated further by pointing out the ubiquity of the phe-
nomenon of informational independence in the semantics of natural
languages. This ubiquity is not obvious prima facie, although the
success of GTS makes it predictable. The reasons for its inconspicu-
ousness are diagnosed in Hintikka (1990).
(ix) Some philosophers have seen in the absence of a complete
proof procedure the basis of a serious objection to IF first-order
logic. Such a way of thinking is based on a number of serious
mistakes. Apparently the skeptics think that they cannot understand
a language unless they understand its logic; that they do not
understand its logic unless they understand what its logical truths
are; and that they do not understand the logical truths of a part of
logic unless they have a complete axiomatization of such truths
(valid formulas). The last two steps of this line of thought are both
fallacious. First, what is needed to understand a language is to
understand the notion of truth (truth simpliciter, not logical truth) as
applied to it. The reason is that what a sentence says is in effect that
the world is such that it is true in the world. This point is completely
obvious, and it has been emphasized by a variety of prominent
philosophers all the way from Frege to Davidson. I do not see much
point in arguing for it elaborately here.
Since IF first-order languages admit truth definitions expressible
in the language itself, as will be explained in a later chapter, they
cannot possibly be faulted on this point.
The subtler mistake seems to be that in order to understand logic
one has to understand its logical truths. This is at best a half-truth.
One possible mistake here is to think that logical truths are merely
a subclass of truths. Hence, according to this line of thought, a
satisfactory account of truth for a language ought to yield as
a special case an account of logical truth for this language. But
logical truth simply is not a species of (plain) truth, notwithstanding
the views of Frege and Russell. Logical truths are not truths about
this world of ours. They are truths about all possible worlds, truths
on any interpretation of nonlogical constants. This notion is an
 INDEPENDENCE-FRIENDLY LOGIC 67

animal of an entirely different color from ordinary truth in some one

world. In this sense, to call valid sentences "logically true" is a mis-
nomer. Several philosophers who swear by the ordinary notion of
truth are nonetheless more than a little skeptical of the ultimate
viability of the very notion of logical truth.
Moreover, it simply is not true that the only way of coming to
master the notion of logical truth is to provide a recursive enumer-
ation of logically true sentences. Such an enumeration is neither
necessary nor sufficient for the purpose. On the one hand, even if we
can give such an enumeration, it is not automatically clear what it is
that is being enumerated. On the other hand, there are other ways
than recursive enumeration of coming to understand the idea of
logical truth. Indeed, if you understand the ordinary notion of truth
in a model and also understand what the totality of all models is, you
can characterize valid (logically true) sentences that are true (in the
ordinary sense) in all models. This way of coming to understand the
notion of logical truth does not presuppose the axiomatizability of
all logical truths nor indeed any logical proof method at all.
It is fairly clear that many of the philosophers who object to IF
first-order logic on the grounds of its incompleteness are assuming
some version of the syndrome which I have called belief in the
universality of language. Among its manifestations is the assumption
that all semantics is ineffable and the thesis that properly speaking
we can speak of only this world of ours. If the latter thesis is true,
logical truths must be a subclass of ordinary truths, as Frege and
Russell thought, for there is nothing else for them to be true about.
Whatever one thinks of such theses, it is inappropriate to base one's
arguments on them here. The reason is that the question of the
definability or undefinability of truth is one of the most important
test cases as to whether the universality assumption is correct or not.
Hence to assume universality in the present instance is in effect to
argue in a vicious (or at best uninformative) circle.
Some philosophers, for instance Quine, have expressed the fear
that a logic which is not axiomatizable is too unwieldy to be viable.
IF first-order logic offers a counterexample to such paranoia. How
manageable IF first-order logic is conceptually, is attested to by the
fact that most of the "nice" metatheorems that hold in ordinary
first-order logic also hold in IF first-order logic, as we have seen.
(x) The apprehensions prompted by the nonaxiomatizability of
the logical truths of IF first-order logic can perhaps be partly allayed
68 THE PRINCIPLES OF MATHEMATICS REVISITED

by considering the question of the axiomatizability of the inconsist-

ent (unsatisfiable) formulas of IF first-order logic. A moment's
thought shows that the class of inconsistent formulas of our new
logic is indeed axiomatizable. This can be shown and an axiomat-
ization can be found — along the usual way. Given a finite set a of
formulas, we can formulate a set of rules for adding new formulas to
it along the same lines as, say, in the step-by-step rules for construct-
ing a model set alias "Hintikka set" in the sense of Smullyan (1968).
Intuitively, we can think of following these rules as an attempt to
describe more and more fully a model in which all the members of
a are true. The rules for doing so are, formally speaking, inverses of
certain disproof rules. It is not difficult to show (though I will not do
so here) that these rules can be chosen so as to be complete. That is to
say, they can be chosen in such a way that they bring out the
inconsistency of any inconsistent a that can be uncovered by a finite
number of applications of the construction rules: all the different
ways of trying to construct a model for the members of a will run into
a dead end, just as in ordinary first-order logic.
The only reason why this complete disproof procedure does not
yield a proof procedure in IF first-order logic is the behavior of
negation. For a complete proof procedure, say, a procedure for
proving a conclusion C from a set a of premises, we need rules for
trying to construct a model in which the members of a are true but
C not true. In general, alas, there is no contradictory negation of C in
the relevant IF first-order language (i.e., no formula which would be
truejust in case C is not) for which we could try to construct a model
jointly with a. It is for this reason that a complete disproof procedure
does not translate into a complete proof procedure in the teeth of
a failure of a tertium non datur.
Most of the details of a disproof procedure of the kind just
envisaged are not interesting. There nevertheless is one point worth
noting in it. In the usual proof and disproof procedures, the lion's
share of the work is done by instantiation rules or by rules equivalent
to them, especially by the rule of existential instantiation or its dual,
universal generalization. In the usual simple form of the rule of
existential instantiation, a new individual constant is brought in to
replace the variable bound to an existential quantifier.
In IF logic, such a simple rule of existential instantiation is no
longer adequate. Such a rule cannot discriminate different kinds of
dependencies of an existential quantifier on universal quantifiers
 INDEPENDENCE-FRIENDLY LOGIC 69

with a wider scope. Instead of the former rule which simply takes us
from an existentially quantified sentence
(±c)S[x] (3.52)

to its substitution-instance
S[b] (3.53)

where b is a new individual constant, we must have a rule which takes

us from a formula
S1 =S1[(ly/Vx1,Vx2,...)S2[y]] (3.54)

(in the negation normal form) which contains an existentially quanti-

fled subformula to a new sentence of the form
(3.55)

where (Vz1), (Vz2),... are all the universal quantifiers other than
(Vx1), (Vx2),.. within the scope of which (]y/Vx1, Vx2,...) occurs in
.

s2 and where f is a new function constant. This rule might be called

the rule of functional instantiation. We also need a stipulation to the
effect that a rule of universal instantiation must not be applied to
a formula (in negation normal form) to which the rule of functional
instantiation can be applied. Otherwise we cannot reach a sound and
complete system of rules of disproof for IF first-order logic along
anything like the usual lines.
This result might seem to be rather quaint, but on a closer scrutiny
it turns out to throw light on general issues concerning the nature of
quantifiers. If the idea of quantifiers as higher-order predicates is
right, then a first-order existential quantifier prefixed to an open
formula says merely that the (usually complex) predicate defined by
that open formula is not empty. The resulting logic so understood
must be exhausted by the usual rule of existential instantiation which
introduces one new individual constant, for the nonemptyness of
a predicate can be expressed by using such a constant. However, we
just saw that such a simple instantiation rule is not sufficient or
always admissible. Why not? The reason obviously is that Hilbert
was right and Frege wrong that is, that quantifiers are in reality
tacit choice functions that depend on some outside universal quanti-
fiers but not on others. The insufficiency of the usual rule of
existential instantiation in IF first-order logic thus constitutes telling
70 THE PRINCIPLES OF MATHEMATICS REVISITED

evidence against the idea that quantifiers are merely higher-order

predicates.
The naturalness of the rule of functional instantiation can be
illustrated in a variety of ways. For instance, a perceptive reader may
have noticed that it is very closely related to the way most meta-
theoretical results were proved, as in Section (a). Furthermore, the
old-fashioned rule of existential instantiation can be thought of as
a limiting case of functional instantiation and the instantiating
individual constant as a function constant with zero arguments.
Thus the sufficiency of the rule of existential instantiation in ordi-
nary first-order logic can be thought of as yet another piece of
undeserved luck which has befallen this particular logic, but which is
not representative of the real situation in logic in general.
Conversely, the new rule of functional instantiation is a most
natural generalization of the old rule of existential instantiation.
Other changes in the familiar rules of logical inference may likewise
be needed in IF first-order logic. For instance, we may choose to
stipulate that the inclusion of the syntactical scope of (Q2x) within
the scope of (Q1x) automatically means that the former depends
on the latter. If so, the usual rules for extending and con-
tracting the scope of the quantifier over another quantifier may have
to be reformulated.
Another dimension of the need for a rule of functional instanti-
ation is that it illustrates the failure of compositionality in IF
first-order logic. In the most natural rules of disproof for this logic,
what is substituted for an existentially bound variable depends
for its context on the entire initial formula to be tested for possible
inconsistencies.
Yet another implication of the need for a functional instantiation
rule pertains to the relation of the semantical games of verification
and falsification to the games of formal proof and disproof. At first
sight, it looks as if one could read from the rules of verification games
rules for constructing a world in which there exists a winning
strategy for the initial verifier. In a certain sense this holds for
disproof procedures, but not in as direct a sense as might first seem to
be the case. For the existence of a winning strategy is not a matter of
the possibility of making moves in a certain way independently
of other moves. The existence of a strategy involves dependencies of
moves on or of each other. Such (in)dependencies can be captured
only by means of functions.
 INDEPENDENCE-FRIENDLY LOGIC 71

Thus in a very real sense IF logic does not only force us to rethink
the semantics of our basic logic; it forces us also to reconsider our
rules of logical inference. In terms of my anachronistic indictment of
Frege, his mistake is not limited to formulating his formation rules
too restrictively. If he had removed the restrictions, he would also
have had to strengthen his rules of inference.

Note
1
This seems to have been established first by Walkoe (1970); but see also Enderton
(1970) and Barwise (1979).
 4

The Joys of Independence:

Some Uses of IF Logic

So far, I have been talking about what IF logic is like. The next ques-
tion is what such a logic can do. This question admits of a variety of
partial answers already at the present stage of the exploration that is
being carried out here.
Perhaps the most general key to the usefulness of a recognition of
informational independence is the ubiquity of such independence.
Many philosophers still seem to think of branching quantifiers — and
informational independence in general as a marginal phenom-
enon, a logician's curiosity without deeper theoretical interest. In
view of such views, it is important to realize that informational
independence is in reality a widespread and important feature of the
semantics of natural languages. It can be shown to play a crucial role
in epistemic logic, in the theory of questions and answers, in the de
dicto versus de re distinction, and so forth.
The prevalence of informational independence does not come as
a surprise to anyone who thinks game-theoretically about the
semantics of natural languages. The leading idea of such a game-
theoretical approach is to associate a game rule to each different
structural feature and to each lexical item of a natural language. But
as soon as that has been done, you have — or, rather, language users
have — the option of considering some moves as being made on the
basis of nonmaximal information sets, no matter what the structural
feature or lexical item in question is. Hence the prevalence of
independence is virtually a corollary of the basic idea of GTS,
assuming of course its success in elucidating the different aspects of
the semantics of natural languages.
72
 JOYS OF INDEPENDENCE: SOME USES OF IF LOGIC 73

The ubiquity of the phenomenon of informational independence

in natural language nevertheless has been hidden by the fact that
informational independence is not indicated in natural languages by
any uniform syntactical device. The reasons for this secretiveness of
natural language grammar turn out in fact to be highly interest-
ing from the viewpoint of linguistic theorizing (see Hintikka
1990).
A terminological comment may be in order here. At least one
observer has in discussion called IF first-order logic "deviant". The
epithet is undeserved. IF first-order logic is a conservative extension
of ordinary first-order logic. It contains classical first-order logic as
a part, and hence cannot be called nonclassical except on the pain of
deviant terminology. More than that: even though the extension
beyond ordinary first-order logic which IF first-order logic repre-
sents exhibits nonclassical behavior, this extension is forced on us by
perfectly classical principles. What happened was that I assumed
some perfectly classical laws for the nonordinary part of IF first-
order logic, such as DeMorgan's laws, the law of double negation,
and so forth. But then it turned out that those assumptions entailed
a violation of another allegedly classical law, namely, the law of
excluded middle. Hence attributes like "deviant" and "nonclassical"
when applied to IF first-order logic are apt to direct a hearer's
attention to a diametrically wrong direction.
One can also point out that, in principle, interpreted IF first-order
sentences are not connected with experience in a more distant man-
ner than ordinary first-order sentences. I have pointed out on earlier
occasions, for instance in Hintikka (1988a), that even sentences with
a complex quantifier prefix can be thought of as having been learned
directly. In particular, what a controlled experiment shows in
a simple two-variable case is how the observed variable depends on
the control variable. The outcome of such an experiment is therefore
a sentence of the form
y] (4.1)

normally with some restriction x1 <x < x2 on x. More complex

experimental setups can yield more complex propositions. A mo-
ment's thought shows that they can likewise yield an IF first-order
sentence, for example,
y, z, u] (4.2)
74 THE PRINCIPLES OF MATHEMATICS REVISITED

Here x and z are the control variables, y and u the observed ones.
What the experiment will have to show is, among other things, that
y does not depend on z or u on x. Establishing such an independence
experimentally may be difficult in practice, but it is not impossible in
principle.
One can say more here, however. One surprise is not enough. It
was seen in the preceding chapter that things are not in logic as we
used to think. The true basic logic is not ordinary first-order logic,
but independence-friendly first-order logic. But things are not as we
used to think in mathematics, either. The usual picture of mathemat-
ical practice is of a mathematician carrying out proofs in terms of
first-order logic. Whatever assumptions one has to make beyond
first-order logic are set-theoretical, and the entire enterprise can in
principle be carried out within an axiomatic set theory.
This is a misleading picture of mathematics. One reason why it is
misleading is that mathematics is full of conceptualizations which in
the last analysis must be expressed by means of IF first-order logic
rather than of ordinary first-order logics, even though this ingredient
is seldom acknowledged in so many words. Perhaps the most
familiar example is offered by the notion of uniform differentiability.
The function f(x) is differentiable at each point of an interval
x1 <x <x2 if and only if
<x <x2) & (Izi <IoI))
D(I((f(x+z)—f(x))/z)---yI <IrD) (4.3)

The function is uniformly differentiable in the same interval if and

only if the same condition is satisfied with replaced by In
this case of uniform differentiability, a confusion between dependent
and independent quantifiers has actually manifested itself in the
history of mathematics (see, for instance, Grabiner 1981, p. 133).
In general, the mathematical notion of uniformity is closely
related to the idea of informational independence.
Such examples are easily multiplied. In many cases, the use of IF
quantifiers is hidden by the use of function symbols. Consider
a second-order formula where two functions are asserted to exist.
Suppose the argument sets of those functions are not linearly
ordered by class-inclusion. Then the resulting expression will have
the logical form
(4.4)
 JOYS OF INDEPENDENCE: SOME USES OF IF LOGIC 75

where the sets {x11, x12,. . .} are not linearly ordered by class inclu-
sion. Then the translation of (4.4) to the first-order level will nor-
mally involve irreducibly IF quantifiers.
Even a sentence asserting the existence of a single function, whose
argument sets at its different occurrences are not linearly ordered by
class inclusion, can be a case in point. An example might be of the
form
(]f)(Vx)(Vz)S[x, z,.f(x), f(z)] (4.5)

Here (4.5) is easily seen to be equivalent to

z) (y = u))) (4.6)

Such examples perhaps illustrate why informational independence

is not an unknown phenomenon in working mathematics. One
cannot perhaps hope to find examples of branching first-order
quantifier structures explicitly mentioned in mathematical treatises,
but one should not be surprised to find mathematical results which
assert the existence of two functions of one variable whose values for
respectively different arguments can be related to each other in a
specifiable way, as in a sentence like
f(x), z, g(z)] (4.7)

And, as has been seen, to assert (4.7) is equivalent to asserting the

independence-friendly first-order sentence
y, z, u] (4.2)

The connection between the notion of informational indepen-

dence and the relation between the argument sets of different func-
tions can be put to use for clarificatory purposes that transcend the
scope of logic and mathematics. Consider, as an example, the
question as to when two physical systems and E2 are independent
of each other. Let us assume, for the sake of analysis, that these
systems and their behavior are specified in first-order terms. Then
one may be tempted to say that and E2 are independent if and
only if the quantifiers characterizing one range over a set of individ-
uals different from, and exclusive of, the range of values of quantifiers
characterizing the other. We can now see that this requirement is not
always sufficient. For the relative order of the members of the two
sets of characterizing quantifiers may still create dependencies be-
tween the two sets, and hence between the two systems. Hence, for
76 THE PRINCIPLES OF MATHEMATICS REVISITED

a satisfactory characterization of the mutual independence of the two

systems, it must also be required that the quantifiers characterizing one
are informationally independent of those characterizing the other.
An especially interesting and tacit use of informationally indepen-
dent quantifiers is found in the mathematics of quantum theory. Let
us assume that we are there considering a systems whose state
depends among others on two variables x, z. Let y and u be variables
for the observed values of x and z, respectively. Let S[x, y, z, u] be
a specification of how x, y, z, u are related to each other. In practice,
S[x, y, z, u] includes a specification of the state of the system in
question, including the values of x and z. It will also include a
specification of the projection operators which lead from the state
vector to the possible observed values y and u.
The crucial question is whether x and z are simultaneously
measurable. This question means asking whether y can depend only
on x and u only on z. Otherwise, the observed values cannot help to
specify the actual ones. But this question amounts to asking whether
the following sentence is true:

(]f)(3g)(Vx)(Vz)S[x, f(x), z, g(z)] (4.7)

If x and z are conjugate variables, the well-known answer is that (4.7)

is not true. Of course, this does not preclude that one can measure
either x or z. However, if one then tries to measure the other one, too,
the result depends also on the first variable. In other words, the
following sentences can be true even though (4.7) is not:
(Bf)(Rg)(Vx)(Vz)S[x, f(x), z, g(z, x)] (4.8)

(Jf)(dg)(Vx)QVz)S[x, f(x, z), z, g(z)] (4.9)

But this possibility is well known from the theory of IF logic. In fact,
(4.7) is equivalent to the IF sentence
(Vx)(Vz)(ay/Vz)(Ju/Vx)S[x, y, z, u] (4.2)

while (8)—(9) are equivalent to the following ordinary first-order

sentences:
(Vx)(Jy)(Vz) (Ju)S[x, y, z, u] (4.10)

y, z, u] (4.11)

As usual, the IF formula (4.2) is stronger than the corresponding

 JOYS OF INDEPENDENCE: SOME USES OF IF LOGIC 77

ordinary first-order formulas (4.10)—(4.11). Thus (4.2) = (4.7) can fail

to be true even though (4.10) (4.8) and (4.11) (4.9) are true.
Moreover, it need not even be the case that (4.7) is false; it may simply
fail to be true. (The reader will find more about this possibility in
Chapter 7.)
Thus certain initially puzzling aspects of the conceptual situation
in quantum theory, especially the impossibility of measuring conju-
gate variables at the same time, can be understood as examples of the
characteristic behavior of quantifiers in IF logic. This fact, together
with my earlier explanation of how naturally the distinctive features
of IF logic emerge out of game-theoretical semantics, helps to
demystify the indeterminacy phenomena of quantum theory.
There is another connection here with what is found in quantum
theory. The formulas (4.8) = (4.10) and (4.9) = (4.11) can be taken to
pertain to the two different orders of the two projective operations
that characterize the measurement of x and z, respectively. The result
is different in the two cases in that (4.8) = (4.10) and (4.9) = (4.11) are
not equivalent. This amounts to the phenomenon of noncom-
mutativity, which again turns out to be a manifestation of a simple
feature of IF logic. How close we are here to the actual mathematics
of quantum theory is shown by the well-known results (see, e.g.,
Hughes 1989, pp. 102—104) to the effect that compatible observables
commute. For one can define compatible observables as two observ-
ables both of which depend functionally on one and the same third
observable.
It is worth pointing out that in analyzing and explicating such
notions as indeterminacy and failure of commutativity I have not
appealed in the least to the physical situation in quantum theory, for
instance to the alleged influence of the measurement apparatus on
the object of measurement. I do not need such hypotheses here.
In mathematics, unlike physics, the informational independence
of quantifiers ranging over different classes of individuals is some-
times taken for granted. As a result, an entire and extensive math-
ematical theory can be essentially a theory of certain kinds of
informationally independent quantifiers. An important example is
offered by what is known as the Ramsey Theory in combinatorics. It
was launched by Frank Ramsey (1930) as a by-product of his work
on the decision problem of first-order logic, and rediscovered in
1933 by a group of young mathematicians that included Paul Erdös
and George Szekeres. The first problem that caught the attention of
78 THE PRINCIPLES OF MATHEMATICS REVISITED

one of them (Esther Klein) can serve to illustrate the role of informa-
tionally independent quantifiers in Ramsey Theory. The problem
was to prove the following theorem: If five points all lie in a plane but
no three of them in a straight line, then four of the five always form
a convex quadrilateral (cf. Graham and Spencer 1990, p. 114). The
five original points can be thought of as being introduced by five
universal quantifiers, say (Vx1), (Vx2), (Vx3), (Vx4) and (Vx5). The
choice of the four can be formalized by an existential quantifier
specifying the excluded point, say
(4.12)

The quantifiers by means of which it is stipulated that the other four

form a convex quadrilateral clearly can — and must — be taken to be
independent of the fifth and of (dy).
Erdös and Klein quickly generalized the problem of 1 + 2" 2
given points and a convex k-sided polygon. In recalling the event,
Szekeres subsequently wrote
We soon realized that simple-minded argument would not do and there
was a feeling of excitement that a new type of geometrical problem emerged
from our circle. (Graham and Spencer 1990, p. 114).
In the light of hindsight, we can say that the "new type of• . . problem"
was a problem involving informationally independent quantifiers.
This is the explanation of the nontriviality of proofs of relatively
simple results in Ramsey Theory which was noted by Szekeres. In
fact, the entire Ramsey Theory is shot through with informationally
independent quantifiers.
This can be illustrated and documented by considering as an
example one of the main results of Ramsey Theory, known as the
Hales—Jewett Theorem. It can be formulated in the conventional
terminology as follows:
For all r, t there exists N' = HJ(r, t) so that for N N', the
following holds: If the vertices of are r-colored there exists
a monochromatic line.
I have followed here the formulation of Graham, Rotschild, and
Spencer (1990, pp. 32—35). One does not even have to know the
precise definitions of the concepts used here to realize what is going
on. (As an aid to visualization, is roughly speaking an N-
dimensional cube with edges of t units.) It is easily seen that the
Hales—Jewett Theorem can be reformulated as follows:
 JOYS OF INDEPENDENCE: SOME USES OF IF LOGIC 79

For all t, the following holds: For each R there exists N', and for
each together with a r-colo ring of its vertices, there exists a line
I in such that, if N N' and R r, then 1 is monochromatic.
Here "existing for" means being functionally dependent on.
The point is that N' is a function only oft and R while I is a function
of only t, N, and the r-coloring. For instance, for 1 we can take the line
in that has the largest number of elements of the same color. This
brings into the open that there is a Henkin quantifier structure as
part of the logical form of the Hales—Jewett Theorem.
Similar comments can be made of several other theorems of
Ramsey Theory. As was mentioned, it seems to me that the overall
character of Ramsey Theory is closely related to the pervasive use of
informationally independent quantifiers in it. In fact, some of the
peculiarities of informationally independent quantifiers have been
noted precisely in the context of Ramsey Theory by logicians and
linguists studying generalized quantifiers (see Keenan and
Westerstahl, 1996). They have not pointed out the generality of the
phenomenon in question, however.
The combinatorial complexity of IF first-order logic is related to
certain types of computational complexity. This matter is studied in
Hintikka and Sandu (1995). They consider a first-order arithmetical
language L. With each formula F or L, they correlate a computer
architecture capable of computing the Skolem functions of F (in the
extended sense of Skolem function explained in Chapter 2). For an
open formula F[x] with, say, one free variable x, this computer
architecture will then be able to compute for any given value x,
whether this value satisfies F[x] or not. For a closed formula
(sentence) this computer can calculate the truth-value of the sentence
in question.
This correlation between logical formulas and computer architec-
tures is a generalization of the antediluvian or, rather, pre-von
Neumann correlation between truth-functions and switching cir-
cuits. As was just presented, it is not complete when one goes beyond
propositional logic in the sense that many consistent first-order
sentences do not have recursive Skolem functions. No computer
architecture is then correlated with them, notwithstanding their
consistency. I will return to this matter in Chapter 10. It does not
invalidate the use to which I am putting the (admittedly partial)
correlation between first-order formulas and computing architec-
tures.
80 THE PRINCIPLES OF MATHEMATICS REVISITED

How does this correlation behave? Suppose I am given two

formulas. To each I can correlate a certain computer, but I can also
combine the two formulas into a more complex one. This combined
formula will be correlated (with a usually more complex) computer
architecture.
This correlation can be extended to simple and more complex
functional expressions (functional terms). Then the correlate of such
a term will be simply a computer architecture that will compute the
function in question.
Any such functional expression f0[x] can also be correlated with
the formula
(Vx)(By)(y =f'0[x]) (4.13)

Clearly the Skolem function of (4.13) can be chosen to be f0[x].

However, here we are considering formula—architecture correla-
tions, and using the extension to functional expressions only as an
expository device. Strictly speaking you should always think of all
functions as having been eliminated from formulas I am discussing in
terms of predicates in the usual way, rewriting, for example, f(x) = y
as F(x, y) where it is assumed that
(Vx)(Jy)F(x, y) (4.14)

(Vx)(Vy)(Vz)((F(x, y) & F(x, z)) D y = z) (4.15)

The failure of an analogy between first-order formulas and first-

order functional expressions illustrates how easily one loses sight of
informational independence by switching one's attention to func-
tions. It is as if one had already opened the door a little bit to the
axiom of choice merely by allowing function symbols into one's
first-order language.
But suppose I start from two computer architectures, each corre-
lated with a certain ordinary first-order formula. Furthermore,
suppose that these computer architectures are then combined into
a more complex one. Is there still an ordinary first-order formula
correlated with the combined architecture? The right answer is: It
depends. It is not hard to see that if the two given computers are
combined sequentially, the result is a correlate of a more complex
ordinary first-order formula. For instance, if the two formulas are
functional terms [x] and f2 [x], then the sequential combination
corresponds to f2 [f1 [x]].
 JOYS OF INDEPENDENCE: SOME USES OF IF LOGIC 81

It is easily seen, however, that when two computer architectures

are combined in parallel (e.g., as components of a more comprehen-
sive architecture) the result is no longer always a correlate of an
ordinary first-order expression. For instance, consider the following
combined architecture, where C1, C2, and C0 are correlates of
certain functional expressions f1[x],f2[x], and f0[x]:
Cl C0

f1[x]

C2 f0[f1[x],f2[x]J
f2['I

Thus there is a functional expression correlated with the com-

bined architecture; but there is no longer an ordinary first-order
formula correlated with it. For suppose C1 and C2 are correlated
with
and (4.16)

(Vz)(au)S1[z, u] (4.17)

Then in the formula correlated with the combined architecture (Ry)

must depend on (Yx) but not on (Vz). Likewise, (Bu) must depend on
(Vz) but not on (Vx). Such a pattern of dependencies cannot be
represented by any ordinary quantifier prefix.
However, it can obviously be expressed by means of the IF prefix
(Vx)(Vz)(Jy/Vz)(du/Vx) (4.18)

Hence a parallel combination of architectures corresponds, not

always to an ordinary first-order formula, but (always) to an IF
first-order formula. In this interesting sense, IF first -order logic is the
logic of parallel processing. For all practical purposes, it is such a
logic in the same sense in which truth-function theory is the logic of
switching circuits, albeit with the qualification to be registered in
Chapter 10.
Independence-friendly logic can also be used to discuss matters of
principle in the foundations of mathematics. Among other things, it
throws additional interesting light on the axiom of choice. It was
noted in the preceding chapter that in IF first-order logic we can
82 THE PRINCIPLES OF MATHEMATICS REVISITED

have a complete disproof procedure, even though there is no com-

plete proof procedure. Such a complete disproof procedure never-
theless cannot be obtained simply by extending the techniques of
first-order reasoning to IF first-order logic. The usual rule of instan-
tiation, especially the rule for existential instantiation, cannot do the
whole job. The reason is intrinsic to the very ideas of quantifier
dependence and independence. If we try to deal with quantifiers in IF
first-order logic by means of successive instantiations, we lose track
of which instantiations depend on which. We need a rule of func-
tional instantiation for the purpose, as was indicated in the preced-
ing chapter. The fact that it is possible to use only individual
instantiation in ordinary first-order logic is a fortunate but atypical
and accidental feature of this logic. In this respect, as in several
others (cf. Chapter 2), ordinary first-order logic is simply too simple
to serve as a paradigm case of what real logic is like in general.
In fact, we could use the rule of functional instantiation (cf.
Chapter 3) also in ordinary first-order logic, and obtain an elegant
proof method. Perhaps we ought to do so, not only because of the
technical merits of the resulting proof method but primarily because
it would bring out more perspicuously than the received methods the
all-important relations of dependence and independence between
different quantifiers. In any case, it is amply clear that the rule of
functional instantiation formulated in Chapter 3 is an archetypally
first-order (quantificational) rule. It is an immediate and natural
expression of the normal interpretation of quantifiers as codifying
choice functions of a certain kind (cf. here Hintikka and Sandu 1994).
If so, then the axiom of choice is vindicated, for, when applied to
a sentence of the form
(Vx)(3y)S[x, y] (4.1)

where S[x,y] is in a negation normal form with no existential

quantifiers and no disjunctions, it yields as a consequence
(Vx)S[x, f(x)] (4.19)

where f is a new function constant. But such inferences are precisely

what the axiom of choice justifies.
All this could — and should — be said with reference to ordinary
first-order logic. What IF first-order logic does is, in the first place, to
illustrate the organic place of the rule of functional instantiation in
dealing with quantifiers of any kind by pointing out that in IF
 JOYS OF INDEPENDENCE: SOME USES OF IF LOGIC 83

first-order logic some such rule is indispensable. The specific contri-

bution that IF first-order logic makes here is to show that more
complex and potentially more powerful variations of the axiom of
choice can bejustified in the same way. The inferences they authorize
come about by applying the rule of functional instantiation to
sentences more complex than (4.1).
Thus we obtain striking evidence not only of the acceptability of
the axiom of choice but also of its character as a purely logical
principle. Indeed, the rule of functional instantiation can be con-
sidered a generalization of the axiom of choice at the same time as it
can be considered a generalization of the usual rule of existential
instantiation in terms of individuals.
The most telling context nevertheless, in which the notion of
informational independence plays a crucial role is, epistemic dis-
course, including the use of questions and answers. The notion of
independence does not only offer a new and sharper tool for discus-
sing epistemic logic and the logic and semantics of questions and
answers. This notion is indispensable for the purpose of capturing
the central concepts that figure in our use of such ideas as knowledge
as well as question and answer. Here I will discuss only the logic of
knowledge usually referred to as epistemic logic.
The role of informational independence in epistemic logic is not
only important in its own right — it will later turn out to offer insights
into the epistemology of mathematics (see Chapter 11). It is there-
fore in order to sketch briefly the role of informational independence
in epistemic logic (for the concepts used here, see Hintikka 1992).
As usual, I take the notion of knows that as the sole primitive. The
sentence a knows that S will be represented as KaS. In addition to the
notion of knowing that, we used the notion of epistemic possibility:
PbS means it is possible,for all that b knows, that S. Since the only case
considered here is one in which only one person's knowledge is being
discussed, the subscript indicating the agent (knower) can sometimes
be omitted.
It can be said that a statement like KaS deals with the knowledge of
facts (or of propositions). The logic of such knowledge is relatively
easy to master. I am taking here the usual semantics of such notions
for granted. What is not equally clear is how our knowledge of entities
(or of objects) of different kinds (other than facts, in case you think of
facts as entities) can be understood and how it is related (to the extent
it is related) to our knowledge of facts.
84 THE PRINCIPLES OF MATHEMATICS REVISITED

The most important kind of knowledge involving individuals,

functions, and other entities (different from facts) is the kind of know-
ledge which in English is expressed by the different wh-constructions.
How are they to be expressed in an explicit logical notation?
Earlier, attempts were made to deal with these constructions in
terms of K plus ordinary first-order quantifiers. Several construc-
tions can apparently be dealt with in such a manner, including the
following:
a knows whether or not S = KaS v Ka S (4.20)

a knows whether S1 or S2 = KaS1 v KaS2 (4.21)

a knows who satisfies the condition SEx] = (4.22)

But more complex kinds of knowledge cannot be so dealt with.

Schematic examples illustrating this include the following:
a knows to whom (y) each individual (x) bears the relation
S(x,y) (4.23)

a knows whether each individual (x) satisfies the condition

S1[x] or S2[x] (4.24)

It can be shown that the meaning of these sentences cannot be

accounted for in terms of an interplay of K and first-order quantifiers
(and propositional connectives). In terms of conventional logical
notation, they can be expressed only by resorting to second-order
quantification. Indeed, (4.23)—(4.24) are equivalent to the follow-
ing:
f(x)) (4.25)
& (f(x) = 0)) v (S2[x] & (f(x) 0))) (4.26)

But this poses additional problems. For one thing, there is no trace of
higher-order quantification in the English sentences (4.23)—(4.24)
any more than there is such a trace in (4.20)—(4.21). Why, then, must
we use second-order quantifiers in (4.23) and (4.24)? And why should
(4.23)—(4.24) have to be dealt with in an essentially different way from
(4.21)—(4.22), when there is no clue to such a difference in these
ordinary-language knowledge statements?
A method is found in this apparent madness when we realize that
the English wh-elements, logically speaking, merely indicate inde-
pendence of a sentence-initial K. This is true of such "propositional"
 JOYS OF INDEPENDENCE: SOME USES OF IF LOGIC 85

questions as (4.20)—(4.21) which are usually not included among

wh-questions as well as of such wh-questions proper as (4.22)—(4.23).
(It is not even clear how (4.24) should be labeled in the received
classification.) Thus (4.20)—(4.24) can be represented as follows:
K(S( v /K) S) (4.27)

K(S1( v /K)S2) (4.28)

(4.29)

y] (4.30)

K(Vx)(S1[x]( v /K)S2[x]) (4.31)

Here (4.25)—(4.26) are simply the second-order translations of (4.30)—

(4.31). We can now see why (4.23)—(4.24) do not allow for a formula-
tion in terms of a linear sequence of quantifiers plus K. If we try to
express (4.30) on the first-order level without independent quanti-
fiers, we run into an unsolvable dilemma. Since depends on (Vx),
it should come later than (Vx). But since it is independent of K, it
should precede K and hence also (Vx). Similar remarks apply to
(4.24) and (4.31).
The similarity between IF epistemic languages and English be-
comes even more pronounced when we use the double-slash nota-
tion. This is illustrated by the following translations:
a knows whether or not S = (Ka// v )(S v 5) (4.32)

a knows whether S1 or = (Ka// v )(S1 v S2) (4.33)

a knows who (x) is such that SEx] = (4.34)

a knows to whom (y) each individual (x) bears the

relation S[x, y] = y] (4.35)

a knows whether each individual (x) satisfies the condition

S1[x] or S2[x] = (Ka// v )(Vx)(S1[x] v S2[x])
(4.36)

This is not the right occasion to develop further the theory of

epistemic logic resulting from these ideas. There nevertheless is one
massive fact that will later be relevant to my investigations in this
book. In order to see what it is, we may consider the two epistemic
86 THE PRINCIPLES OF MATHEMATICS REVISITED

statements
K(Jx)S[x] (4.37)

K(Bx/K)S[x] (4.29)

What does their truth mean model-theoretically? What (4.37) says is

that in each model (world, scenario) compatible with what is known
there is an individual x satisfying S[x]. In other words, it is known that
there is an x such that S[x]. What (4.29) says is that I can choose an
individual x, independently of whatever world my opponent chooses,
such that S[x] is true in that world. This clearly means that it is known
who such an x is, The important semantical difference is that of these
two the former (4.37) depends only on what is true in each of the worlds
considered alone. No comparisons between different worlds are made.
The latter, that is (4.29), depends also on what counts as the same
individual in different worlds. In order to specify the semantics of
knows that statements, the former is enough. In order to specify the
semantics of knows + wh-constructions we also need criteria of cross-
world identity. They are not fixed by what is true in each world taken
alone.
In brief, the criteria of knowing that do not completely determine
what counts as knowledge of individuals and other kinds of entities,
for instance what counts as knowing who, what, when, where, and so
forth. This result is eminently in agreement with what is found in
ordinary discourse. In order to understand, for instance, a knows who
statement, one must know what the criteria of identification are that
the speaker or writer is presupposing. In so far as we are dealing with
objective meaning, these criteria are objective. Moreover, the same
criteria must be used throughout any coherent argument or dis-
course. But they do not reduce to the criteria of knowing that
(knowing facts). Accordingly, different criteria can be adopted, and
are adopted, on different occasions. Knowing who someone is means
somewhat different things to passport authorities, the FBI, or to the
registrar of your university; and it means radically different things to
the occult writer who once asked in a Boston Globe ad: "Who were
you in 1493?" and to the anonymous British male chauvinist pig who
coined the saying, "Be nice to young girls. You never know who they
will be." A choice between all these different criteria of identification
is determined not by some eternal and immutable logic,just because
the underlying logic will remain the same structurally no matter
 JOYS OF INDEPENDENCE: SOME USES OF IF LOGIC 87

which choice is made; the choice is rather guided by general epi-

stemological considerations plus the specific purpose of the dis-
course in question.
This observation can be extended to knowledge of higher-order
entities, for instance, knowledge of functions. This is an important
point, for functions are inevitably involved as soon as an existential
quantifier depends on a universal one. For instance, in order to be in
a position to say that a scientist s knows that a variable y depends on
another, say x, in a certain way specified by S{x, y] it does not suffice
that there exists a function f such that he knows that
K,(Vx)S[x, f(x)] (4.38)

This function f must furthermore be known to s. Only then does

(4.38) imply
y] (4.39)
And the fact thatf is known to s is expressed by
= y) (4.40)

Now the truth-condition of statements like (4.39) are not determined

by the truth-conditions of unslashed statements of the form In
order to apply epistemic concepts to functions, we must therefore
have independent criteria as to when they are known.
 5

The Complexities of Completeness

In view of what was said in Chapters 3—4, there is no doubt about the
character of IF first-order logic as the basic area of our logic. But
what is this logic like? What is new about it? In Chapters 3—4, a
number of partial answers were given to these questions. Perhaps the
most general one concerns the expressive strength of our new basic
logic. In view of the close similarity between ordinary and IF
first-order logic, it may come as a surprise that IF first-order logic is
much stronger than its more restricted traditional version. How
much stronger will become clearer in the course of my examination
of its properties and applications.
One consequence of the strength of IF first-order logic is that it
does not admit of a complete axiomatization. The set of valid formulas
of IF first-order logic is not recursively enumerable. Hence there is
no finite (or recursive) set of axioms from which all valid sentences of
this logic can be derived as theorems by means of completely formal
(recursive) rules of inference. Thus the first remarkable property of
IF first-order logic is that, unlike its special case of ordinary first-
order logic, it does not admit of a complete axiomatization. The
reasons for this incompleteness will be explained in Chapter 7.
How are we to react to this incompleteness? Different perspectives
are possible here. Purely technically, in view of the power of IF
first-order logic, the failure of this logic to be axiomatizable is
perhaps not entirely surprising. Yet this failure (and this power) goes
against a philosophical tradition of long standing. According to this
traditional way of thinking, incompleteness is a problem for math-
ematicians, not for logicians. It was in many ways shocking of Gödel
88
 COMPLEXITIES OF COMPLETENESS 89

to prove that a mathematical theory as simple as elementary arith-

metic is incomplete. Yet the nearly universal assumption has been
that this incompleteness is somehow attributable to the character of
elementary arithmetic as a mathematical theory. Logicians and
philosophers could find some consolation in the thought that pure
logic, which typically was taken to be first-order logic, remained
untainted by the dreaded virus of incompleteness.
Admittedly, certain areas of what is commonly called logic, for
instance second-order logic, have been known to be incomplete. But
the typical reaction to their incompleteness has been to reclassify
them as mathematical rather than purely logical theories, perhaps
amounting to set theory (or a part thereof) considered as a math-
ematical theory.
This kind of reaction to Gödel's incompleteness result is not
a peculiarity of twentieth-century philosophers. It is in reality part
and parcel of a long tradition which goes back at least to Descartes.
Descartes rejected traditional syllogistic logic as being trivial and
fruitless, while extolling the power of mathematical reasoning, es-
pecially of the analytic method of Greek geometers.1 In a similar vein,
Kant declared logic as being based on the law of contradiction and
yielding only analytic truths, in contradistinction to the synthetic
a priori truths of mathematics. Against such a background, it was
not surprising that Gödelian incompleteness was thought of as being
due to the status of elementary arithmetic as a mathematical theory.
Yet a closer analysis of the history of this tradition soon produces
warning signals. Whatever the right interpretation of the ancient
analytic method is, and whatever one thinks of its use in the hands of
Descartes and his contemporaries, it is a fact that it was used by
Greek geometers in the context of reasoning which can be captured
by the resources of our ordinary first-order logic. Furthermore, for
Kant the gist of the mathematical method consisted in the use of
instantiation rules, which with our twentieth-century hindsight are
again part and parcel of logic in our sense, that is, of first-order logic.
Hence it is clear that at the earlier stages of this tradition the
boundary between logic and mathematics was drawn in an essential-
ly different way from what twentieth-century philosophers are used
to. Yet this historical discrepancy has not really been acknowledged
by most philosophers. They go on thinking in terms of an essential
difference between unproblematic logic and treacherous mathemat-
ics. This may, among other things, explain why Hilbert's ideas have
90 THE PRINCIPLES OF MATHEMATICS REVISITED

not been appreciated more widely. One thing that he tried to do was
to break down the traditional dichotomy between logical and
mathematical reasoning.
Against this background, it is a shock indeed that the most basic
part of our logic, IF first-order logic, is incomplete. This shock is not
just a matter of being unused to the idea that a genuine part of logic
could be incomplete. For one thing, it forces us to look at the entire
history of the foundations of mathematics in a new light. For
instance, suppose, as a light-hearted thought-experiment, that Frege
had corrected his fallacy, admitted informationally independent
quantifiers and realized that the resulting logic is not axiomatizable.
This would have made nonsense of his entire project. For what
would then have been the point of reducing mathematics to logic, if
the logic needed for that purpose cannot be axiomatized? How could
Frege have possibly been able to characterize the logic which the
reduction aims at, since he avoided semantical methods and used
only formal (syntactical) ones? Again, Hilbert could not have
conceived of his famous project of proving semantical (model-theor-
etical) consistency of mathematical theories by establishing their
proof-theoretical consistency, if he had realized that the true basic
logic he needed in those theories would not have a complete
proof-theoretical axiomatization.
It is also instructive here to recall what was found in Chapter 2. It
was pointed out there that, in a perspective provided by GTS,
ordinary first-order logic is in many respects unrepresentative of the
conceptual situation in general. One of the reasons why IF first-
order logic is interesting is that it is more representative of the whole
range of things that can happen in logic in genera! than ordinary
first-order logic. It represents a concrete example of features that are
characteristic of logic in general but are not exhibited by its more
traditional cousins. Incompleteness is one such feature. Some others
will be examined in the next couple of chapters.
In order to understand the conceptual situation created by the
incompleteness of IF first-order logic, we have to also take into
account Gödel's celebrated impossibility results (1931). I do not
think it is any exaggeration to say that the community of philos-
ophers and mathematicians has still not managed to cope with the
true import of Gödel's discoveries and to draw the right conse-
quences from them for the future development of the foundations of
mathematics.
 COMPLEXITIES OF COMPLETENESS 91

First, we have to realize that there are several entirely different

notions of completeness. They do not even apply to the same sort of
thing. For instance, what Gödel was dealing with was the incom-
pleteness of a nonlogical (mathematical) theory, namely, elementary
arithmetic, whereas the incompleteness of IF first-order logic per-
tains to a part of logic, namely, to the nonaxiomatizability of a
logical theory. It is a serious mistake to assimilate these two kinds of
theories to each other. Such an assimilation in effect means over-
looking the distinction between the two functions of logic in mathe-
matics which was emphasized in Chapter 1.
We have to distinguish from each other, at the very least, the
following senses of completeness:

(a) Descriptive completeness. It is a property of a nonlogical axiom

system T. It means that the models of T comprise only the
intended models. If there is only one intended model (modulo
isomorphism), descriptive completeness means categoricity, with
the intended model as the sole model of T
Notice that no reference to any method of logical proof is
made in this characterization. All that is needed is the relation of
a sentence to its models plus the idea of an intended model.
If, following the terminology of logicians, the set of intended
models is said to constitute a theory, descriptive completeness
means that this theory is axiomatizable. This is precisely the sense
in which Euclid and Hubert were trying to axiomatize geometry.
It is also the sense in which we speak of the axiomatization of
some physical theory. In that case, the intended models are
simply all the physically possible systems of the relevant
kind.
(b) Semantical completeness. It is a property of a so-called axiomatiz-
ation of (some part of) logic. It means that all the valid sentences
of the underlying language can be obtained as theorems from the
so-called axioms of that system of logic by means of its inference
rules. If we assume the usual permissive standards of axiomatiz-
ation, then there exists a complete axiomatization for some part
of logic if and only if the set of logically true (valid) sentences (of
that part of logic) is recursively enumerable.
(c) Deductive completeness is a property of a nonlogical axiom
system T together with an axiomatization of logic (method of
formal logical proof). It means that from T one can prove by
92 THE PRINCIPLES OF MATHEMATICS REVISITED

means of the underlying logic C or C for each sentence C of the

language in question.
(d) There is a further use of the term "complete" in connection with
nonlogical axiom systems. It can be called Hilbertian complete-
ness, for the locus classicus of this sense of the term is the so-called
axiom of completeness first used in the second edition of Hubert's
Grundlagen der Geometrie (1899, second ed. 1902). Strictly speak-
ing, though, the axiom occurs first in Hilbert (1900). The difficulty
of understanding the notion of completeness is illustrated by the
confusions surrounding Hilbert's "unglückseliges Axiom"
(Freudenthal 1957, p. 117). The sense of completeness it involves
was first assimilated to the other sense of completeness. Only in
Baldus (1928) do we find a clear statement of the difference be-
tween Hilbertian completeness and other senses of completeness.
Hilbert's axiom is in effect a maximality assumption. It says
that the intended models of the axiom systems are such that no
new geometrical objects can be added to them without violating
the (other) axioms of the system. Maximality assumptions of this
kind play a potentially important role in the foundations of
mathematics but the maximality of a model with respect to the
individuals it contains is a fundamentally different idea from the
maximality of the system, which is what the descriptive and
deductive senses of completeness deal with.

The difference between these different senses of completeness will

become clearer as we proceed. Philosophers may have been confused
here by the fact that there are interrelations between these different
kinds of completeness and incompleteness. For instance, let us
assume that there is only one intended model for a nonlogical theory
T which is based on a semantically complete logic L. Then T is
descriptively complete only if it is deductively complete. However, if
L is not semantically complete, then T can be descriptively complete
even when it is deductively incomplete.
The differences between (a)—(c) are not merely technical. They are
manifestations of the different purposes which logic can serve in
mathematical thinking. These purposes were discussed in Chapter 1.
Completeness in the sense (a) pertains to the descriptive function of
logic in mathematics which was described there. If a mathematician
has reached a descriptively complete axiom system, then he or she
has reached an intellectual mastery of the range of structures that are
 COMPLEXITIES OF COMPLETENESS 93

exemplified by the intended models of T The mathematician has

separated these structures from all the others and expressed what is
characteristic of them.
In contrast to such descriptive completeness (a), deductive com-
pleteness (c) deals with something else. It deals with the deductive
function of logic, which in this case amounts to a mathematician's
ability to deal computationally with the given theory T. Complete-
ness in this sense means that there is an algorithm for separating
what is true and what is false in what can be said of the models of T
Such an algorithm takes the form of a set of rules for deriving
theorems from axioms. An emphasis on this kind of completeness
(deductive completeness) often reflects a conception according to
which the essence of mathematical practice is theorem-proving. It
also means emphasizing the deductive function of logic at the
expense of its descriptive function.
The general distinction of which the contrast between descriptive
completeness and deductive completeness is a special case is not new.
As an example, logicians almost routinely make a distinction be-
tween descriptive complexity and computational complexity. The
former deals with the power of a logic to capture different kinds of
structures "out there", while the latter deals with the complexity of
the different derivations of theorems from axioms. This obviously
parallels the distinction between descriptive completeness on the one
hand and deductive (as well as semantical) completeness on the other
(see, e.g., Dawar and Hella forthcoming).
What are the main consequences of Gödel's incompleteness the-
orem? What did he really prove? First and foremost, it is important
to realize that the only kind of incompleteness that he established
directly is the deductive incompleteness of elementary arithmetic.
That is, there is no recursive enumeration of the true sentences of
elementary arithmetic implemented by an axiomatization of the
underlying logic. In other words, no consistent axiomatization T of
elementary logic, taken together with an explicitly axiomatized logic
L, enables us to prove (i.e. to prove from T by means of the logic L)
either S or —' S for each sentence S expressible in the language of
elementary arithmetic.
As was just noted, from this deductive incompleteness one can
infer the inevitability of the descriptive incompleteness of elementary
arithmetic only if L is semantically complete. This was the case with
the logic Gödel actually used. He used ordinary first-order logic,
94 THE PRINCIPLES OF MATHEMATICS REVISITED

which he had proved semantically complete himself in Gödel (1930).

But the observations made in this work put the matter in a new light.
It has been seen that the completeness of ordinary first-order logic is
merely a by-product of the unnecessary and unnatural restrictions
that Frege and Russell imposed on the formation rules of this logic in
ruling out independence-friendliness. Hence it is likely that we have
to use a semantically incomplete logic anyway. And if so, Gödel's
results do not imply anything about the prospects of formulating
descriptively complete axiom systems for elementary arithmetic, not
even on the usual first-order level.
Here we can see one kind of impact that IF first-order logic has in
virtue of its very existence. One implicit reason why the possibility of
formulating descriptively complete nonlogical theories based on
a semantically complete logic has not been taken seriously has been
noted previously. It is the belief that no logic worth its name can be
incomplete. IF first-order logic provides a concrete counterexample
to such beliefs, even though it does not itself provide the possibility of
a descriptively complete axiomatization of elementary arithmetic.
In fact, it will be shown later in this work how such semantically
complete axiomatizations can in fact be formulated by methods
suggested by the game-theoretical approach to logic and mathemat-
ics. In this way, the developments reported in this book put Gödel's
incompleteness results in a new and interesting perspective.
From this perspective, it must be said that the philosophical
community has not acknowledged how limited the significance of
Gödel's incompleteness result really is. This point is related to what
was said earlier of the difference in the significance of the different
kinds of completeness, especially the difference between descriptive
and deductive completeness. This misperception has deeper roots.
Philosophers have been impressed by Gödel's result because they
have overestimated the importance of deductive and computational
techniques in mathematics. They have been seduced by the oversim-
plified picture of mathematical activity as mere theorem-proving. In
reality, it is clear that deductive incompleteness is not the most
important kind of incompleteness. What attracted mathematicians,
scientists and philosophers in the first place to the axiomatic method
was the possibility of an intellectual mastery of a whole area of
important truths, say all geometrical truths. A complete axiom
system was a means of reaching that mastery. But this kind of
mastery is in the last analysis a mastery of the models of the axioms,
 COMPLEXITIES OF COMPLETENESS 95

and not of the things we (or our computers) can deductively prove
about them. Such mastery can manifest itself in ways other than
a mechanical derivation of theorems from axioms. Contrary to the
oversimplified picture that most philosophers have of mathematical
practice, much of what a mathematician actually does is not to
derive theorems from axioms. This point is especially clear in those
branches of mathematics which do in fact start from a given set of
axioms, for instance, the theory of groups. What a group theorist or
a mathematician working in a similar field does is only to a limited
extent theorem-proving in the strict sense of the term. Most of it can
from a logical point of view be characterized only as a metatheory of
group theory. It is a quest for an overview of all the models of a given
theory, and not of its deductive consequences. A mathematician for
instance classifies groups and proves representation theorems for
different kinds of groups. Such theorizing is largely independent of the
deductive resources which the mathematician in question commands.
The kind of completeness relevant to this kind of enterprise is
obviously descriptive completeness. This notion is definable com-
pletely independently of any axiomatization of the underlying logic,
and hence independently of all questions of deductive completeness.
It admittedly presupposes that we understand the sentence—model
relationship. But this we need to master in order to understand the
meanings of our sentences in the first place. It does not presuppose,
as we just saw, that the axiom system in question is deductively
complete.
Hence Gödel's incompleteness result does not touch directly on
the most important sense of completeness and incompleteness,
namely, descriptive completeness and incompleteness. This is an
important observation, for Gödel's result has generally been taken
to point out limits to what can be done by means of logic. The
explanation of this discrepancy in perception is that the distinction
between the different kinds of completeness has not been made
clearly by the majority of philosophers and mathematicians. This
failure has even led to the panic reaction of doubting even the
viability of the concept of arithmetical truth. In reality, Gödel's
incompleteness result casts absolutely no shadow on the notion of
truth. All that it says is that the whole set of arithmetical truths
cannot be listed, one by one, by a Turing machine.
Thus we have reached an interesting insight into Gödel's incom-
pleteness theorem. This result pertains to the deductive and corn-
96 THE PRINCIPLES OF MATHEMATICS REVISITED

putational ways of dealing with nonlogical axiom systems. Noting

this true address of Gödel's result helps us to understand the urgency
of the theoretical problems which Gödel's incompleteness theorem
has forced us to face. In many ways these problems have become
increasingly more pressing by the developments which have oc-
curred since Gödel reached his theorem. In the meantime, computa-
tional and deductive ways of dealing with problems of virtually any
kind have become increasingly more common in almost all walks of
life because of the explosive development of computer technology.
Such computational techniques are prevalent in areas like cognitive
science and artificial intelligence. It follows from Gödel's and
Tarski's results (or is at least strongly suggested by them) that
computational techniques can never exhaust the theory of the
phenomena in question. Hence the status of most of the computa-
tional approaches to different kinds of inquiry are bound to stay
forever in a limbo of incompleteness. This seems to deprive computa-
tional methods of all deeper theoretical interest. The entire develop-
ment of computer-based technology for dealing with language and
thinking is thus not only under suspicion but already indicted for
being ad hoc, a mere collection of tricks and rules of thumb put
together for the purposes of application. Such overall suspicions are
of course unjustified. However, philosophically speaking a justifica-
tion presupposes a shift of emphasis away from attempts to capture
arithmetical truths, or whatever phenomena one may be studying, in
one fell swoop computationally to the never-ending pursuit of better
and better approximations of the target phenomena.
Thus in a sense Gödel's real message was merely the deductive
incompleteness of elementary arithmetic. As such, his famous result
affects only the prospects of our deductive mastery of elementary
arithmetic and not our ability to deal with this branch of mathemat-
ics axiomatically or descriptively, unless such axiomatic treatments
are constrained to use only ordinary first-order logic.
Admittedly, Gödel's incompleteness results are closely related to
the undefinability of arithmetical truth in elementary arithmetic
itself. However, this aspect of Gödel's results can be judged in the
same way as Tarski's result (1935, 1956a) concerning the undefina-
bility of truth for a language in that language itself, and will be so
judged in Chapter 6.
Meanwhile, a few comments of another kind may be in order.
A confusion between deductive and semantical completeness (and
 COMPLEXITIES OF COMPLETENESS 97

incompleteness) is also what underlies the fashionable comparisons

between non-Euclidean geometries and nonclassical logics. The
possibility of non-Euclidean geometries is a consequence of the
descriptive incompleteness of an axiomatic geometry without the
parallel postulate. But different kinds of logics are not models of any
one incomplete theory that could be completed by adding new
axioms. In order to obtain different logics, it is not enough to choose
different axioms. For this purpose, you have to define different
notions of model and/or truth. You cannot as much as speak of the
semantical completeness of a logical axiomatization without having
defined the notions of model and truth and, by their means, the
notion of validity.
The incompleteness that attaches to IF first-order logic is nat-
urally semantical incompleteness. As such, it has few implications for
the completeness or incompleteness of nonlogical theories. On the
contrary, the enhanced strength of IF first-order logic makes it
ceteris paribus a better tool for formulating descriptively complete
nonlogical theories, notwithstanding its semantical incompleteness.
Indeed, as soon as we have a viable notion of truth-in-a-model at
our disposal, and also some characterization of the totality of all
models of the given language, then we can set up an axiom set for
elementary arithmetic hoping realistically that it might then turn out
to be complete. The intended kind of completeness I have called it
descriptive completeness — will then be different from deductive
completeness. For one thing, nothing is said in the characterization
of this kind of completeness of any actual deductive method; the
question of deductive completeness does not even have to be raised
in order to be able to speak of descriptive completeness. Descriptive
completeness is in a sense a purely model-theoretical notion.
Unfortunately this implies that descriptive completeness is sub-
ject to "Tarski's curse" discussed in Chapter 1. The notion of valid-
ity, that is, truth in every model, cannot be characterized without
characterizing the notion of truth; which in turn cannot be defined (if
Tarski's result is the last word on the subject) without resorting to
a stronger metalanguage, for instance, to higher-order logic or set
theory. And that appeal to set theory seems to entail all the prob-
lems, puzzles and paradoxes that set theory is ripe with.
One thing that I will do in this book is to study different ways out
of this problem. For instance, as far as Gödel's immediate subject-
matter, elementary arithmetic, is concerned, descriptive complete-
98 THE PRINCIPLES OF MATHEMATICS REVISITED

ness would be easy to reach if we were allowed to ascend to the

second-order level. Essentially, all we have to do is to replace in
a Peano-type axiom system the schema for induction by the second-
order formulation of the principle of induction:
(VX)((X(0) & (Vy)(X(y) X(y+ 1))) (Vy)X(y)) (5.1)

Of course, the standard interpretation in Henkin's (1950) sense of the

second-order language being used must be presupposed. Unfortu-
nately, as was pointed out in Chapter 1, second-order logic apparently
involves all the vexing problems of set existence, and is therefore
deemed by many philosophers to be beyond the pale of pure logic.
And even on the first-order level, there are ways of obtaining
descriptively complete axiomatizations of elementary arithmetic. It
follows from Gödel's result that such axiomatizations must not use
ordinary first-order logic or any other deductively complete logic.
But it is not difficult to find other kinds of first-order logic to fit the
bill. Indeed, in Chapter 7 it will be shown how elementary arithmetic
can be completely axiomatized (in the descriptive sense of complete-
ness) on the first-order level. Another, and somewhat more exotic,
way of doing so will be briefly discussed in Chapter 10.
There are other ways of illustrating the fact that Gödel's results do
not cover all the bases even in the foundations of arithmetic. For
instance, even if elementary arithmetic were deductively complete, it
would not yet be descriptively complete. In more literal terms, even
the deductively complete theory which is true in the structure of
natural numbers is not descriptively complete.2 (Needless to say, this
theory is not axiomatizable.) It admits of several different non-
isomorphic models; it does not have the structure formed by natural
numbers (with addition and multiplication) as its only model. In-
deed, the descriptive incompleteness of Peano arithmetic is an
almost direct consequence of the compactness of first-order logic. It
does not need Gödel's elaborate argument for its proof. All this
shows how many important issues are left untouched by Gödel's
results.
In a different direction, it is known that there are deductively
complete and hence decidable theories of reals and of sufficiently
elementary geometry (see, e.g., Tarski 1951, 1959). These theories are
nevertheless — or, for that very reason descriptively incomplete.
They admit of different nonisomorphic models, and hence will not
satisfy a mathematician who like Hubert is committed first and
 COMPLEXITIES OF COMPLETENESS 99

foremost to the study of certain kinds of structures, such as the

structure of the "real" reals. Such a mathematician will not be
impressed by the deductive completeness of a Tarski-type theory of
reals.
One upshot of this incomplete examination of the problems of
incompleteness is a virtual reversal of the traditional picture sketched
above. Incompleteness is a logical rather than a mathematical
phenomenon. There are in principle no obstacles for mathema-
ticians to reach the kind of completeness that really matters to them,
namely descriptive completeness. In subsequent chapters, I will try
to say a little bit more of the concrete ways of realizing such
completeness. But in order to reach such completeness, the underly-
ing logic must be strengthened so as to become semantically incom-
plete.
In this respect, IF first-order logic is tailor-made to satisfy the
existing need of such a strong and natural, but at the same time,
semantically incomplete logic.
These brief remarks enable us to describe the general problem
situation that Gödel's incompleteness result forces on us. The crucial
feature of the situation is not that we cannot somehow set up
descriptively complete systems of axioms for the mathematical
theories we are interested in I will soon return to the question as to
how we can actually set them up. The possibility of descriptively
complete axiomatizations nevertheless does not diminish the im-
portance of the problem of finding actual deductive ways of dealing
with the structures that the axiom system deals with. Hence the main
post-Gödelian, not to say postmodern, foundational problem is to
look for new deductive methods and to analyze them. Gödel's results
show that this cannot be done mechanically, much less all at once,
but of course this is precisely the reason why this new "fundamental
problem of the philosophy of mathematics" is an intriguing one. In
practice, such stronger aids of deduction can often be codified in the
form of new axioms for the mathematical theory in question. Hence
the task I am talking about here is not entirely unlike the task of
finding stronger and stronger axioms of set theory.
Here the impact of Gödel's incompleteness theorem is reinforced
by that of the incompleteness of IF first-order logic. In the preceding
chapter, it was seen that mathematics is shot through with concepts
that rely on IF first-order logic. What this means is that the full logic
that is needed in most of the advanced mathematical theories is not
100 THE PRINCIPLES OF MATHEMATICS REVISITED

amenable to a semantically complete axiomatization. Mathematical

reasoning, for instance mathematical theorem-proving, cannot be
reduced to the deduction of new results by means of a logic that can
be completely captured by any set of recursive rules. Moreover, the
new principles of reasoning whose need is always lurking just under
the surface are in principle but logical rather than set-theoretical.
For this reason, ordinary first-order axiomatic set theory is an
inappropriate framework for mathematical theorizing when it comes
to finding new and stronger principles of mathematical reasoning,
even if the need for finding ever stronger set-theoretical assumptions
is acknowledged.
More generally speaking, as soon as we realize that the fundamen-
tal logic of mathematics is IF first-order logic rather than its old-
fashioned brother, we lose any and every reason to formulate
deductively complete theories in mathematics, except for atypical
special cases. For an axiomatic theory can be deductively complete
only if the underlying logic is semantically complete when restricted
to deductions from the axioms of that theory. And usually this is
possible only if that basic logic is semantically complete in general.
For instance, one can now state concisely one reason why set
theory is inappropriate as a basis of mathematics. The problem is not
that set theory has led to paradoxes or threatens to lead to para-
doxes. The problem is not the deductive incompleteness of set
theory, either. It is its descriptive incompleteness. This is the reason
why set theory cannot provide aims, much less guidelines, for a
search of increasingly stronger new assumptions which can be
thought of as providing better deductive methods for dealing with it.
This descriptive incompleteness pertains in the first place to first-
order axiomatizations of set theory. But in its usual forms axiomatic
set theory does not have really natural higher-order models, either.
I will return to the prospects of axiomatic set theory in Chapter 8.
So what should a mathematician do? When a logician or a math-
ematician sets up an axiomatic system these days his or her main aim
invariably is to make it deductively complete. I can envisage a future
in which the typical aim of a mathematician or logician is to
formulate descriptively complete axiom systems which are deduct-
ively incomplete. This is the kind of axiom system that can best serve
the crucial purpose of guiding the mathematicians' and logicians'
never-ending quest for increasingly stronger deductive principles for
that particular theory. Descriptive completeness provides then the
 COMPLEXITIES OF COMPLETENESS 101

aim which can help to direct the logicians' search for stronger logical
principles that would provide more deductive power.
Such descriptively complete axiom systems must nonetheless be
on the first-order level. For otherwise they will be subject to Tarski's
curse: their model theory would arguably involve set-theoretical
concepts.
This observation provides an answer to a question that may
already have occurred to the reader. Is not the task of finding new
axioms addressed every day by working logicians, at least tacitly?
For the obvious suggestion here is that guidelines for the search for
new deductive principles are provided by the model theory for that
part of mathematics in question. For instance, in elementary arith-
metic we have a nice truth definition which can guide a logician's or
mathematician's search for stronger axioms.
Somewhat similar things can be said of a logician's quest for
semantically complete axiomatizations of some part of logic. This
kind of completeness is, as I have shown, a mirage aready on the level
of IF first-order logic. What can be done, however, is to give a non-
axiomatic characterization of logical truth (validity) and to use it as
a guideline for looking for stronger deductive principles. In practice,
it is difficult (and pointless) to try to distinguish the quest for stronger
deductive principles sharply from the quest for stronger descriptive
axioms.
But this cannot be the whole story. For one thing, not all model
theory helps us in our search for the right deductive principles. For
the purpose, we must have some idea of what the intended models of
the relevant theory are so that we can characterize truths in those
models. Furthermore, we must gain some new insights into the
structure of the intended models in order to see what follows
logically from the (descriptively complete) axioms characterizing
those structures. Thus it is not the whole story that deductive logic is
a means of understanding better the structures a mathematical
axiom system specifies. Sometimes, the traffic flows in the other
direction: better grasp of those structures is needed to see what the
deductive consequences of the axioms are.
Admittedly, this "reverse logic" situation does not seem to arise
very often in mathematical practice. But perhaps this rarity is some-
thing of an illusion. For instance, when number-theoretical results
are proved by analytical means what happens is that the structure of
natural numbers is considered as part of a more complicated struc-
102 THE PRINCIPLES OF MATHEMATICS REVISITED

ture, insights into which provide insights also into the structure of
natural numbers. Another example is provided by the way in which
the continuity assumptions concerning real numbers were originally
inspired by the analogy between the structure of reals and the
structure of a geometrical line.
A search for stronger deductive assumptions is also closely related
to that fetish of recent philosophy of mathematics, namely, math-
ematical practice. As was pointed out above, a mathematician is
typically searching for an insight into the structure of all models of
a given theory. Only a minuscule part of a practising mathema-
tician's actual work consists in deductively deriving theorems from
axioms, philosophers' preconceptions notwithstanding. Admittedly,
such representation theorems do not usually give rise to new axioms
for the original given theory or to new logical principles. Neverthe-
less, when the given theory is deductively incomplete, such meta-
theorems may be one way of finding new axioms. I am even tempted
to suggest that the role Gödel assigned to intuition in the founda-
tions of mathematics should really be understood as pointing out the
importance of this metatheoretical and model-theoretical way of
deriving better deductive principles for the study of certain kinds of
structures from insights concerning these structures directly.
I will return to the problem of finding stronger assumptions in
Chapters 8—9. Further reasons will be given there as to why axiom-
atic set theory serves singularly badly the new purposes of founda-
tional research that Gödel's incompleteness results force logicians to
face.
These viewpoints can be illustrated in terms of the very distinction
made earlier in this chapter between different senses of completeness.
Doubts are perhaps prompted by my characterization of descriptive
completeness, the reason being the reference it contains to the class
of intended models. These intended models are sometimes referred
to also as standard models. However, in this usage the term "stan-
dard" is merely a euphemism for "intended", for it cannot in general
be identified with any of the senses of standardness (such as Henkin's
in his 1950 paper) that can be strictly defined. It is not even clear what
the relationship between these different senses of standardness is, or
may be.
Hence any reference to intended models in the definition of
descriptive completeness might seem extrinsic to the true founda-
tions of mathematics, and perhaps even dangerously unclear. How
 COMPLEXITIES OF COMPLETENESS 103

can we characterize such models with sight unseen, so to speak? One

way of allaying such doubts is to point out that the idea of the
intended model need not be any more dubious than the idea of
the model in general. As was in effect pointed out in Chapter 1, the
general concept of model also involves prelogical ideas concerning
what a structure must be like in order to be acceptable as a model
simpliciter. One way of seeing this is to examine how the notion of
model can naturally be varied. Such variation helps us to understand
what the unvaried notion of model involves.
One can for instance, widen the concepts of model and obtain
a weaker logic as a consequence that is, a logic in which fewer
formulas are valid because there are more models for them to fail to
be true in. A way of doing so is easily uncovered by means of GTS. It
shows that our usual logic is based on an important assumption. The
different choices made by the two players can be thought of in the
same way as probability theorists think of their sampling pro-
cedures, to wit, as "draws" (selections) of "balls" (individuals) from
an "urn" (domain of a model). Normally, it is assumed that the
population of balls in an urn stays the same between successive
draws in the logical terminology; that the model does not change
between the successive moves in a semantical game. This assumption
is not unavoidable. Probability theorists have in fact devised "urn
models" which are changed in accordance with some fixed law
between the draws. In the same way, we can allow the model on
which a semantical game is based to change in certain specified ways
between the successive moves of the game. The result is a well-
defined logic which can be used in the study of the foundations of
mathematics (see Hintikka 1975, Rantala 1975).
In the other direction, one can change one's logic by imposing
extra conditions on it. This will result in there being more logical
truths than before, for there are now fewer acceptable models in
which a formula can fail to be true. Once again, the resulting new
logic can have interesting uses in the study of the foundations of
mathematics. An especially interesting possibility in this direction is
to implement maximality and minimality assumptions concerning
models, not by means of explicit axioms or other kinds of sentences,
but as a model of the underlying language in the first place. This
possibility is tentatively explored in Hintikka (1993a).
Implicit in the traditional notion of a model is a rejection of
changes of both these two kinds. Hence this traditional notion of
104 THE PRINCIPLES OF MATHEMATICS REVISITED

model involves highly nontrivial prelogical assumptions. It is not

any more of a dangerous assumption to think that somehow a nar-
rower class of intended models of some mathematical theory is given
to us. Certainly we have as sharp an idea of the structure of natural
numbers as we have of traditional models as distinguished from their
rivals.
It seems to me that the usefulness of our prelogical ideas about
certain classes of structures could be exploited more widely than
they have in fact been utilized. For instance, the second number class
(the structure of countable ordinals) is in our day usually studied as
an aspect of axiomatic set theory. I would not be surprised if it turned
out to be helpful to draft into the service of some set theorists in
a more direct way the intuitions we have about the second number
class, by considering its model theory directly in its own right.
We have thus seen that the rumors of the demise of completeness
in mathematics have been greatly exaggerated. The most important
kind of completeness in mathematics is alive and well in the land of
model theory. However, the independence of that land still appears
to remain in jeopardy, as was emphasized in Chapter 1. For its basic
notion, the notion of truth in a model, still seems to us to take refuge
in higher-order considerations which appear to invite an invasion of
higher-order logical (or else set-theoretical) foreign armies. My next
task hence is to see how this danger can be avoided.

Notes
For the history of the analytic method of ancient Greek geometers and for its
subsequent history, cf. Hintikka and Renies (1974). For Kant, see Hintikka (1973).
2
In this sense, a "theory" is simply a deductively closed set of formulas.
 7

The Liar Belied: Negation in IF Logic

(A) Tertium datur

So far so good. A truth predicate can after all be defined in an IF
first-order language for the Gödel numbers of the sentences of that
very same language. But what about the arguments which Tarski
and others have presented against the definability of truth for
a formal language in that language itself? The best known of these
arguments are closely related to the time-honored or perhaps rather
time-abused liar paradox (for this paradox, see Martin 1978 and
1984). They are illustrated by an outline of an argument which
establishes the indefinability of truth for a first-order arithmetical
language in that language. Such an argument can use the technique
of Gödel numbering. In rough outline, the argument runs as follows:
Let n be the numeral representing the number n. According to
the diagonal lemma, for any expression S[x] in the language of
elementary arithmetic containing x as its only variable, there is
a number n represented by the numeral n satisfying the following
condition:
g(S[n])=n (7.1)

where g(S [n]) is the Gödel number of the expression obtained from
S[x] through replacing x by n. Assume now that there is a truth
predicate T[x] in the arithmetical language in question, that is,
a numerical predicate T[x] such that T[n] is a true arithmetical
statement if and only if n is the Gödel number of a true arithmetical
statement. Then we could apply the diagonal lemma to and
obtain a number d such that the Gödel number of
(7.2)

131
132 THE PRINCIPLES OF MATHEMATICS REVISITED

is d. Hence what (7.2) says is that the sentence with the Gödel number
d is false. But that sentence is (7.2), which yields a contradiction.
This line of thought is similar to the usual proofs of Gödel's
incompleteness theorem except that there the actual provability
predicate and not the hypothetical truth predicate is being consid-
ered.
What happens to this highly important line of thought when we
move from ordinary first-order arithmetic to IF first-order arithme-
tic? At first sight this argument might seem impossible to default.
After all, IF first-order logic is stronger than the ordinary one, and
hence the paradox might seem to be impossible to avoid also in IF
first-order languages.
In order to see why the liar paradox does not arise in IF first-order
languages, we have to go back to the basic ideas of game-theoretical
semantics. In it, the truth of a sentence S is defined as the existence of
a winning strategy for the initial verifier in the corresponding
semantical game G(S). The falsity of S is defined as the existence of
a winning strategy of the initial falsifier in G(S). Negation is handled
by stipulating that G( S) is like G(S) except that the roles of the two
players have been exchanged.
These stipulations are simple and natural, but they have a striking
consequence. This consequence concerns the law of excluded middle
or the law of bivalence, as some people prefer to put it. This principle
of tertium non datur becomes a determinacy assumption in the sense
of game theory. In other words, it asserts that there always exists
a winning strategy for one or the other of the two players of
a semantical game. As we know from game theory (and from set
theory), determinacy assumptions are normally far from obvious.
They often fail. They represent strong assumptions which allow
a quantifier inversion, in that a determinacy assumption enables one
to infer from the nonexistence of a winning strategy for one of the
players the existence of a winning strategy for the other one. Sight
unseen, there is therefore no reason to think that the law of excluded
middle should hold in my semantical games. It is only by a lucky
coincidence, so to speak, that it holds in ordinary first-order logic. In
IF first-order logic, it is easily seen to fail, and to fail in a radical
manner.
In order to see this, we can recall that many of the "nice"
metatheorems that hold in ordinary first-order logic continue to
hold in IF, sometimes in a strengthened form. One of them is the
 NEGATION IN IF LOGIC 133

separation theorem, which says that two jointly inconsistent but

separately consistent sets of formulas of an IF first-order language,
say o and x, can always be "separated" by a single ordinary first-order
formula F in the sense that r I— F and x I— F.
Assume now that the contradictory negation of an IF first-order
sentence is somehow expressible in the same language, say by the
sentence S1. Then by applying the separation theorem to {S1} and
{S2} it is seen that both of them must be sentences of ordinary
first-order logic. For then there is a separation formula which
belongs to an ordinary first-order language such that
sit—so

F-

But since I— it follows that S1 is equivalent to and to

This shows that the only IF first-order formulas whose
contradictory negation is expressible in the IF language are precisely
the ordinary first-order formulas. In this sense, the law of excluded
middle holds only in the fragment of an IF language consisting of
ordinary first-order language.
The upshot is that in IF first-order languages, the law of excluded
middle inevitably fails. This result requires a number of comments
and explanations.
(i) This result is not due to our arbitrary choice to define falsity in
a certain way for IF first-order languages, namely, in the way
explained above. It is independent of our terminology. The real issue
is how negation behaves in IF languages. Indeed, whatever interest
my observations may have is due to the fact that we did not assume
any breakdown of the law of excluded middle or any truth-value
gaps or a third truth-value. What we have is an inevitable conse-
quence of the most natural semantics for IF first-order languages,
namely, the game-theoretical one. Indeed, since neither Tarski-type
truth definitions nor substitutional ones are available in IF first-
order languages, we scarcely have any choice here but to use some
version of game-theoretical semantics.
(ii) This observation can be sharpened. Let us merely assume that
the usual interconnections hold between negation and the other
logical constants. These interconnections are codified in the law of
double negation, in De Morgan's laws and in the interdefinability of
the two quantifiers. These laws certainly hold for classical negation.
134 THE PRINCIPLES OF MATHEMATICS REVISITED

Assume also that the usual game-theoretical rules for the other
notions (&, v, V) apply. They, too, merely incorporate perfectly
classical assumptions. We then need a rule for negation only for the
trivial case of a negated atomic sentence. Yet the result is the same
treatment for negation as was outlined above. Inter alia, the law of
excluded middle fails. Somehow an apparently nonclassical logic
results from entirely classical assumptions.
Thus we can see that the prima facie surprising behavior of
negation in IF first-order logic is not due to our game rule for
negation, which reverses the roles of the two players. This rule can be
replaced by the perfectly classical rules which merely push negation
signs deeper into a formula. Given the other game rules, the two
kinds of rules for negation are simply equivalent.
In a sense illustrated by this line of thought, we cannot help
treating negation in the way indicated above, which leads to the
inevitable result that S is not the contradicatory negation of S but
a stronger (dual) negation.
(iii) The kind of failure of the principle of excluded middle that we
are dealing with here is an unavoidable combinatorial consequence
of the way quantifiers and other concepts interact with each other. It
has nothing to do with the limitations of human knowledge. Epi-
stemic failures of tertium non datur should be studied in epistemic
logic, not here. The failure I have described has nothing to do with
some particular subject matter, for example, with the future as
distinguished from the past. It has nothing to do with infinity, either.
Below (see (v)) a finite miniexample is given of a sentence pertaining
to a universe of six individuals where one can literally see (at least
after a few moments thought) that a certain sentence is neither true
nor false. You can come to see virtually directly that any attempt to
verify a certain simple sentence can be defeated and that any attempt
to falsify it can likewise be defeated. What else can any man (or, as
Samuel Johnson might add, any woman or any child) reasonably say
here except that such a sentence is neither true nor false? Thus a little
reflection can easily bring home to you the naturalness, nay, the
inevitability of our treatment of negation.
(iv) Unlike other approaches, such as intuitionism, our treatment
of a logic in which the law of excluded middle fails does not involve
any tampering with the definitory laws of ordinary first-order logic.
The game rules for semantical games connected with IF first-order
sentences are precisely the same as those for ordinary first-order
 NEGATION IN IF LOGIC 135

languages, except for allowing that certain moves are made by the
initial verifier in ignorance of certain earlier ones. Moreover, the
logical truths and valid rules of inference of ordinary first-order logic
remain unproblematically valid.
(v) It might be instructive to illustrate these remarks by a simple
example. For the purpose, consider a miniuniverse consisting of six
entities, three gentlemen (Alan, Brian, and Cecil) who between them
have three hobbies, namely, riding, sailing, and tennis. Indicating
hobbies by an arrow, we can sum up the relationships holding here
by means of the following diagram:

Alan
Brian = b . s = sailing (D)
Cecil = C. • t = tennis
For simplicity, I will tacitly restrict the variables x and z to gentle-
men and the variables y and u to hobbies. Consider, then, the
following statements:
y) & H(z, u) & (y u)) (7.3)

y) & H(z, u) & (y u)) (7.4)

y) & H(z, u)) (y = u)) (7.5)

Here (7.3) says that any two different gentlemen have respective
hobbies that are different, that is, that no two different gentlemen
have all their hobbies in common. Moreover, by putting x = z we see
that (7.3) implies that each gentleman has at least two hobbies. It is
immediately seen that (7.3) is true in the model (D). It is also seen that
(7.5) is the contradictory negation of (7.3) and hence false in (D).
Clearly, (7.4) is true if and only if there are function f and g such
that
(Vx)(Vz)(H(x, f(x)) & H(z, g(z)) & (f(x) g(z))) (7.6)

Can such functions be defined in the model D? Let us see. By

symmetry it suffices to consider the case
f(a) = r (7.7)

Then a substitution of a for both x and z in (7.6) yields

H(a, f(a)) & H(a, g(a)) & (f(a) g(a)) (7.8)
136 THE PRINCIPLES OF MATHEMATICS REVISITED

This is possible only if

g(a)=s (7.9)

By putting x = a, z = b we obtain
H(a, f(a)) & H(b, g(b)) & (f(a) g(b)) (7.10)
This is possible only if
g(b)=t (7.11)

In the same way as in (7.9), it can be shown that

f(b)=r (7.12)

In the same way as in (7.11) we can show that

f(c)=s (7.13)

But then we can substitute c and a for x and z, respectively, and

obtain
H(c, f(c)) & H(a,g(a)) & (f(c) # g(a)) (7.14)

which is in view of (7.7) possible only if

(7.15)

But this contradicts (7.13), showing that (7.4) is not true in D.

By reviewing this line of thought a couple of times you can reach
a point at which you can directly see the failure of (7.4) to be true
simply by inspecting the model. In other words, you can train
yourself literally to see from the model (D) that any strategy of
attempted verification of(7.4) can be defeated by a suitable counter-
strategy.
Now, it can easily be seen that (7.5) is not only the (dual) negation
of (7.3) but also of (7.4). The reason is that a winning strategy must
work no matter what moves my opponent makes, independently of
their dependence on (or independence of) others. Hence neither (7.4)
nor its negation (7.5) is true — that is, the law of excluded middle fails
for (7.4).
This simple example illustrates several important facts, some of
them explained above. It shows that the failure of the law of excluded
middle is by no means assumed in our treatment. It is a consequence
of a certain combinatorial feature of the behavior of negation in IF
logic. In intuitive terms, this combinatorial fact can be interpreted as
 NEGATION IN IF LOGIC 137

saying that every serious attempt to verify (7.2) in (D) can be defeated
and that every serious attempt to falsify it in (D) can be likewise
defeated. By serious attempt, I mean here an attempt governed by an
explicit strategy. In the case on hand, the relevant combinatorial fact
is a finitary one. This shows that infinity cannot in any way be
blamed for the failure of tertium non datur, even if in other cases the
combinatorics involved is infinitary.
To put the same point in a different way, the applicability or
nonapplicability of tertium non datur to a given sentence in a certain
model ("world") is an objective fact about that model or world. It
does not depend on any limitations of one's knowledge about the
world, nor does it imply that there are such limitations.
More generally, the example illustrated the fact that questions
concerning the logical truth and satisfiability of IF first-order for-
mulas are in a reasonable sense combinatorial. They are questions
concerning the possibility of certain kinds of structures of individ-
uals. These "kinds" are characterized in a way similar to, say, the
structures involved in Ramsey-type theorems. These structures can
be finite, but they may of course also be infinite. In neither case are
any totalities of sets or functions or any other kinds of higher-order
entities involved.
What distinguishes our example from others is its simplicity. By
looking at (D) and by reviewing the argument presented above in
your mind, you can reach a point where you can almost literally see
the failure of the principle of excluded middle. Even though the same
cannot be said of more complex cases, the difference lies merely in
a contingent limitation of human capacities, and not in the subject
matter.
It is even possible in principle to formulate examples of sufficiently
simple infinite structures where the failure of tertium non datur could
be seen equally directly.
All this illustrates the fact that our truth definition deals with the
strategies of semantic games and not with their definitory rules.
Later, in Chapter 10, an important possibility will be found of
changing our truth definition without changing the move-by-move
definitory rules of semantical games.
Here the reader is perhaps already catching a glimpse of why it is
hopeless to try to implement constructivistic ideas by somehow
manipulating the definitory rules of the games of formal proof. Even
when we move from such formal games to their real basis in the form
138 THE PRINCIPLES OF MATHEMATICS REVISITED

of semantical games it still does not help to try to vary the definitory
rules of such games. Conversely, our results illustrate at the same
time the advantages of a strategic viewpoint in logical and linguistic
theory.
(vi) In order to avoid misunderstanding, it is in order to point out
that the failure of the law of excluded middle for a certain sentence
S is no guarantee that its contradictory negation is not expressible in
a IF first-order language. For instance, it is easily seen that in many
models a sentence of the following form is neither true nor false:
y] (7.16)
where A[x,y] is atomic. Yet (7.16) is easily seen to be logically
equivalent to
y] (7.17)
and hence to have as its contradictory negation
A[x, y] (7.18)
What happens in the more complicated example just presented is
not only that tertium non datur fails, but that the contradictory
negation of a sentence is not expressible in the IF first-order lan-
guage in question.
The behavior of negation in IF first-order languages has a number
of consequences, including the following:
(a) Unless conditionals are treated separately, material implication
(truth-functional conditional) is not expressible in general in IF
first-order languages. For instance, (F G) and F v G) ex-
press something stronger than the customary material implica-
tion of truth-function theory.
It would be interesting to examine whether this fact might
have tacitly been responsible for the unease many philosophers
have felt about truth-functional conditionals.
(b) The same applies to equivalences. If (S1 *-+S2) is construed as

(S1 & S2) v & S2) (7.19)

then its logical truth entails that S1 and S2 both satisfy the law of
excluded middle and hence belong to ordinary first-order logic.
(c) As an application of these observations, consider Tarski's T-
sentences. They are substitution-instances of the schema
(T) fl is true
 NEGATION IN IF LOGIC 139

where "H" is a placeholder for a quote or a structural description

of the sentence which is to replace "p." (T) is equivalent to
(T)* (H is true & p) v (fl is not true & p)

(T)* clearly entails

(7.20)
In other words, any sentence whose truth can be characterized by
means of an explicitly formulated instance of the T-schema satis-
fies the law of excluded middle; it has a contradictory negation in
the same language. But if so, Tarski's T-schema is useless in its
usual form in IF logic. What it tries to express is true and impor-
tant, namely, that a sentence S and its truth-condition c(S) must
be true simultaneously. But the usual formal expressions of the
T-schema require more. It requires that S and c(S) must also be
false simultaneously. This requirement cannot be sustained in IF
first-order logic. Hence the entire matter of the T-schema has to
be handled with caution. It is unproblematic only in ordinary
first-order languages.
As I have pointed out before (Hintikka 1976b), the T-schema fails
for another reason, too. Like other features of Tarski's approach, it
presupposes compositionality. When compositionality fails, T-
schema may also fail. Cases in point are found when the T-schema is
formulated in natural languages, for example, as follows:
(T)** H is true if and only if p
with "H" and "p" working as before, but this time "p" working as
a placeholder for English sentences (free of indexicals). By substitu-
ting "anyone can become a millionaire" for "p" and its quote for "H"
and by taking only one half of the equivalence we can see that (T)**
entails
"anyone can become a millionaire" is true if anyone can
become a millionaire (7.21)

But (7.21) is not true: one person making a cool million does not
imply that everyone can do it.
The explanation is that the game rule for any has priority over the
game rule for conditionals. From that it follows that the force of any
depends on where it occurs in a sentence, for instance, in a condi-
tional sentence. In other words, it violates compositionality.
140 THE PRINCIPLES OF MATHEMATICS REVISITED

Because of this scope peculiarity of any, it has a wider scope in

(7.21) than thereby making (7.21) false.
This counterexample to Tarski's T-schema might seem to turn on
a peculiarity of the English any that has no general theoretical
interest. That would be a misperception, however. What (7.21) really
illustrates is the failure of compositionality in natural languages,
ergo an extremely interesting general phenomenon.
In sum, the T-schema is virtually useless for a discussion of truth
and truth definitions in general. The presuppositions it is predicated
on cannot be expected to hold in general outside ordinary first-order
languages.
The failure of the T-schema to serve as a satisfactory criterion of
the characterization of truth also means that disquotational ac-
counts of truth are seriously incomplete. They presuppose composi-
tionality and tertium non datur, neither of which can be expected to
hold in languages of more than minimal expressible power and
surely not in ordinary language. Consequently, Tarski's T-schema
cannot serve as a condition on our conception of truth in general.
The failure of truth-functional conditions to be expressible in IF
first-order languages is a more significant fact than one might first
suspect. This significance can be seen by asking: What do we need
truth-functional conditionals for, anyway? An important part of the
answer is: We need them for the purposes of logical proof. For
instance, without a contradictory negation at our disposal, we can-
not use modus ponens, except in a weakened form.
The same point can be put in another way. The formulas that can
serve to justify inferential steps from one sentence to another, as the
truth-functional conditional "if S1, then S2" would serve to facilitate
an inference from S1 to S2, turn out to be extremely strong sentences
semantically. They cannot be expressed in ordinary first-order lan-
guage, and they cannot be expressed in IF first-order language, either.
Later in this chapter we shall see an example of a still stronger
language in which they can be so expressed.
This strength and subtlety of inferential relationships, reflected by
their failure to be expressed except in very strong languages, makes it
highly unlikely that we humans could have direct unaided insights
into their status as "inference tickets". And if so, another common
view of logic bites the dust. It is the idea of logic as a systematization of
our intuitions about relationships of logical consequence between
propositions. There is precious little reason to think that we could
 NEGATION IN IF LOGIC 141

have viable "intuitions" about matters so complex that they cannot

even be expressed in an IF first-order language.
This observation further enhances the status of model-theoretical
concepts and theories as compared with the role of proof-theoretical
(inferential) approaches to logic.
An additional observation is relevant here. In Chapter 3 it was
seen that IF first-order logic enables us to capture, as the models
of a formula, classes of structures that cannot be likewise captured
by any ordinary first-order formula. Let L be such a class of struc-
tures, captured by the formula F(X). Then it follows from what
has been seen that the complement of Z cannot be captured by any
IF first-order formula, For this complement would have to be cap-
tured by the contradictory negation of F(X), which is not available
unless F(X) is a formula of ordinary first-order logic. But then it does
not enhance the representational capacities of ordinary first-order
logic.
The behavior of negation in IF first-order logic also shows how
this logic can escape the clutches of the third main result (besides
Gödel's and Tarski's) that apparently severely restricts what can be
done by means of logic. This result is known as Lindström's Theorem
(see Lindström 1969; cf. Ebbinghaus, Flum and Thomas 1984, Ch. 12).
It is a result in the theory of abstract (model-theoretical) logics
which says, roughly speaking, that the ordinary first-order logic is in
a natural sense the strongest possible logic satisfying certain condi-
tions. This sense is simply an explication of the idea that a stronger
logic enables one to make finer distinctions among models than
a weaker one.
Of the conditions of Lindström's Theorem, most attention has been
directed at the assumptions that the logic in question is compact and
that it validates the Löwenheim—Skolem theorem. Since IF first-order
logic satisfies these conditions and yet is stronger than it in the
relevant sense of the term, these two assumptions cannot be all the
important ones. Indeed, the joker in the pack turns out to be
Lindström's assumption that negation behaves like contradictory
negation. This assumption is not satisfied by IF first-order logic, and
for this reason it can be stronger than ordinary first-order logic,
Lindström's Theorem notwithstanding. Thus negation once again
plays a crucial role in the foundations of logic. And, once again,
apparent restrictions on the power of logic turn out not to be real
ones.
142 THE PRINCIPLES OF MATHEMATICS REVISITED

(B) The liar is merely a nontruth-teller

In the light of our results, we can now see why it is that the liar paradox
does not arise in IF first-order languages. The status of the diagonal
lemma strictly speaking needs a review. But since it basically expresses
a combinatorial fact about elementary number theory, there is no
reason to expect that it does not hold in IF first-order logic. But what
escape hatches does its validity leave us with?
Let us see. Let us apply the diagonal lemma to our truth predicate
T. Then we obtain the result that there exists a number h such that h is
the Gödel number of
(7.22)

Then it is easily seen in the usual way that if

T[h] (7.23)

is false, it is true, and vice versa. But (7.23) need not be either true or
false. Hence no contradiction ensues. The liar paradox does not vitiate
the truth definition we have given for suitable IF first-order lan-
guages. The liar sentence merely turns out to be neither true nor
false.
This resolution of the liar paradox is so simple as to prompt
surprise and even disbelief. A few comments are therefore in order.
(a) There is nothing mysterious about the fact that (7.23) is neither
true nor false. That it is neither is merely a combinatorial feature of the
structures that (7.23) speaks of. In principle, this fact is as straightfor-
ward as our example about (D) above. It means that each strategy of
verification for (7.23) can be defeated and that each strategy of
falsification can likewise be frustrated. That we cannot see this fact
directly, as we could see the corresponding fact in the case of (7.4) in
relation to (D), is merely due to the greater complexity of (7.23) as
compared to (7.4). Otherwise the two are on a par. One might even
suggest that the dramatization of the properties (7.23) via a compari-
son with the liar paradox is a partial way of highlighting this intuitive
feature of (7.23). My point here is closely related to the fact that
Gddel's impossibility theorem, which he himself compared with the
liar paradox, can be proved constructivistically.
(b) Furthermore, the combinatorial features of (7.23) are indepen-
dent of our terminology. They cannot be changed, for example, by
defining falsity in a way different from ours. The crucial point is that
 NEGATION IN IF LOGIC 143

there is basically only one way of characterizing negation in IF

first-order logic.
It has sometimes been suggested (cf. e.g., Simmons 1990, P. 290) as
a solution for paradoxes like that of the liar that there is in the last
semantical analysis no paradox-generating concept to be expressed.
What has been seen shows that this is not the case in the approach
developed here. For the existence of a winning strategy for the initial
verifier is a definite combinatorial property of models, no matter
whether the game in question is determinate or not. Hence, in a very
concrete sense, there is such a concept of truth to be defined for IF
first-order languages.
(c) It was assumed above in the argument concerning (7.22)—(7.23)
that the diagonal lemma applies when elementary arithmetic is
developed within the framework of IF first-order logic instead of an
ordinary one. Even though the correctness of this assumption is fairly
obvious, it is still worth noting here. The reason why the diagonal
lemma applies is that the syntax of IF first-order arithmetical lan-
guage can be handled (via a suitable Gödel numbering) in its own
ordinary first-order-logic fragment. Hence one can represent the
diagonal functions (cf. Mendelson 1987, p. 155) in the language as
before, for it deals simply with the syntax of the arithmetical language
in question. Furthermore, the representability of the diagonal func-
tions is the only substantial assumption needed for the diagonal
lemma, which is therefore applicable in IF first-order arithmetical
languages.
The same remarks apply mutatis mutandis to the fragments of
second-order arithmetical languages, unsurprisingly, because they are
equivalent to the corresponding IF first-order languages.
(d) If you take a wider view of the situation, you will find that not
only should you have expected the liar paradox not to touch our truth
definition but that you should have expected it to be immune to all
paradoxes. In order to see this, you may have a look at the second-
order truth definition as presented in Chapter 6. What it says is that
there exists a predicate (of the Gödel numbers of sentences) satisfying
certain conditions such that x has this predicate. What are the relevant
conditions? They spell out precisely the features which our intuitive
concept of truth has, applied to the given IF first-order language sans
any special truth predicate. There is no self-reference here to generate
paradoxes or contradictions. Unless our intuitive conception of truth
is seriously flawed, this second-order definition cannot go astray, for it
144 THE PRINCIPLES OF MATHEMATICS REVISITED

merely captures what we mean by the truth of an IF first-order

sentence. Hence no paradoxes can be generated by such a definition.
And the possibility of translating the second-order truth definition
back into the IF first-order language in question is simply a matter of
a well-known logical result.
(e) The ease at which the liar paradox was avoided in IF first-order
languages suggests that it does not pose a serious threat against truth
definitions or against the possibility of a consistent concept of truth
also in natural languages. Instead, it suggests an important general
point. In some sense, the liar paradox has more to do with our concept
of negation than with our concept of truth. This is suggested also by
the way in which liar-like arguments are used by Gödel and Tarski. At
bottom, these uses are based on the diagonal lemma which is a combi-
natorial fact about first-order arithmetical languages. There is no
leeway with respect to such a metatheorem. A fortiori, any truth
predicate is vulnerable to the diagonal lemma. There is no way of
avoiding this application. Hence the only possible way out is to
challenge the true-false dichotomy.
In a constructive vein, the failure of the liar paradox in IF first-order
languages, notwithstanding the possibility of being able to define their
own truth, throws some further light on the prospects of giving truth
definitions for natural languages. Indeed, liar-like paradoxes and the
diagonal arguments which they rely on are often adduced as a reason
for the impossibility of defining truth for natural languages.
What has been presented is more than just an example of a rich
language for which truth can be defined in the language itseW
Independently of all questions concerning the relations of IF first-
order languages and all questions as to whether any truth definitions
like the ones expounded here are possible in natural languages, the
definability of truth for IF first-order languages has certain massive
consequences. It explodes once and for all the myth that semantical
concepts like truth cannot be expressed in the (sufficiently rich)
language to which they are applied, as well as the myth that a Tarski-
type hierarchy of metalanguages is inevitable for the purpose of
escaping semantical paradoxes, as has been alleged (cf. Simmons 1990,
especially pp. 296—299).
Our results incidentally show that the main competing approaches
to truth fail to do the same job as the account offered here. This holds
both of redundancy "theories" of truth including disquotational
ones, and of inductive truth definitions. If they are formulated in the
 NEGATION IN IF LOGIC 145

same language as the one to which the proposed truth predicate

pertains, then they are subject to the liar-type counterexamples
discussed here.
(f) A wary reader might nevertheless still feel insecure about the
dangers of possible paradoxes. One source of worry is that, even
though IF first-order logic is (semantically) incomplete there still
exists a complete disproof procedure. Does this completeness perhaps
make IF first-order logic too strong to be safe?
The answer is that the complete disproof method does not give rise
to any paradoxes, but instead gives rise to interesting results. For the
purpose, consider a nonlogical finitely axiomatizable theory T which
contains elementary arithmetic and which is formulated in an IF
first-order language. The general disproof procedure for IF first-order
logic yields a disproof procedure for the theory T, in the sense of
a complete proof procedure for establishing inconsistency with the
axioms of T Using this technique of Gödel numbering, we can discuss
the metatheory of Tin T itself. The complete disproof procedure just
mentioned gives rise to a predicate Disp(x), which is true of the Gödel
number g(S) of a sentence S if S is disprovable on the basis of T,
assuming that T is consistent. Because of the completeness of the
procedure, Disp(x) holds of the Gödel number g(S) of S if and only if
S is inconsistent with T If we apply the diagonal lemma to the formula
Disp(x), we obtain a formula
Disp(n) (7.24)

where the Gödel number of (7.24) is n and where n is the numerals

representing n. What can we say to (7.24)? If it is true, then it is
disprovable. But since the disproof procedure in question clearly
produces only false sentences in the models of T, then (7.24) must be
false in such models. Hence it cannot be true in them.
But fairly obviously (7.24) is, like the comparable predicate in
ordinary first-order languages, expressible in an ordinary first-order
form. Hence its strong and contradictory negations coincide. There-
fore there exists a sentence,
Disp(n) (7.25)

which is true but whose negation is not disprovable. Hence the theory
T is consistent but incomplete in the sense that the negation (7.24) of
(7.25) cannot be shown by our disproof procedure to be inconsistent
with T
146 THE PRINCIPLES OF MATHEMATICS REVISITED

Thus the existence of a complete disproof procedure does not result

in paradoxes but in a remarkable positive result. It shows that if
a statement (e.g., a finitely axiomatizable theory) T expressible in IF
first-order language is consistent (in the weak sense of not being
logically false), its consistency in this sense can be proved by establish-
ing the truth of a certain particular sentence in the same language. In
this sense, problems of consistency, like problems of truth, can always
be formulated in the same IF language as the original theory T At the
same time, we can see that T is inevitably incomplete deductively:
there will be sentences that are false (in the domain of natural
numbers) but not disprovable on the basis of T.
The behavior of negation in IF first-order languages is also one of
the sources of the limitations of IF first-order languages mentioned in
Chapter 4 above. One aspect of these limitations confronts us in the
theory of definitions. Indeed, IF first-order logic is likely to put the
entire role of definitions in logic and mathematics in a new light. The
current view of explicit definitions is that they are noncreative.
Whatever can be proved (or defined) by using a definition must be
provable (or definable) without using it. Neither of these requirements
is satisfied in IF first-order logic. In order to see this, consider an
irreducibly IF first-order formula S(x) with x as its only free variable.
Suppose we want to abbreviate S[x] by introducing a new one-place
predicate constant P(x) as a shorthand. There is of course no harm in
doing so, as long as we realize that the constant so introduced is not on
a par logically with the primitive nonlogical constants of the underly-
ing language. For such an introduced constant the law of excluded
middle does not hold, unlike the original nonlogical constants of the
language. From that it follows that the explicit definition of P(x)
cannot be expressed by a sentence like
(7.26)

for (7.26) implies that the law of excluded middle applies to S[x].
Furthermore, we cannot use the definiens P(x) in other formulas
that are abbreviated definitionally. Such dependent abbreviative defi-
nitions can introduce false assumptions into one's argument.
These remarks suggest a way of extending IF first-order logic in the
form it has been defined here. This way consists in allowing the
nonlogical constants (predicates and functions) to be only partially
defined, so that they need not satisfy the law of excluded middle. This
would open interesting possibilities of further developing IF first-
 NEGATION IN IF LOGIC 147

order logic as has been explained in this book. Exploiting these

possibilities would bring the theory of IF first-order logic in cooper-
ation with the theories of partiality and partial logics. (They are
surveyed in a masterly fashion by Fenstad, 1996.) In spite of the great
potential interest of such developments, I will not indulge in them
here. Instead, I will explain a different kind of extension of the original
IF first-order logic of the Chapter 3 vintage.

(C) Contradictio ex machina

At this point of my mystery story of the missing negation, I fully expect
Lieutenant Colombo to scratch his head and say to me: "Yes, but there
is still one question that bothers me. Why cannot we simply introduce
a contradictory negation by a fiat? All we need to give a semantical
meaning to such a negation, say —i S, is to stipulate that —i S is true
just in case S is not true. What's wrong with such a procedure?".
As always, Colombo is right. There is nothing wrong introducing a
contradictory (weak) negation as it were an afterthought. The result
will be called an extended IF first-order Ian guage. Its logical vocabu-
lary is the same as that of our basic IF first-order language, that is,
&, v, V and =, plus the contradictory negation—i. These languages
have several interesting properties, some of which will be discussed
later. Their semantics is determined by the earlier rules of semantical
games for IF first-order logic plus a semantical rule for contradictory
negation. However, there are conditions which such a semantical rule
must satisfy.
(a) The only semantical rule which one can give to the contradictory
negation is the negative one:
i S is true if and only ifS is not true, otherwise false. (R.—i)

In particular, no game rules can be used to define contradictory negation.

One way of seeing this is to point out that the contradictory negation
should certainly obey Dc Morgan's laws, the law of double negation and
the usual laws for negated quantifiers (i.e., the equivalence of with
(Vx) and of (Vx) with (ax) —i). But it can easily be seen that if you
assume these laws as a part of your system of game rules, then you end up
with the strong (dual) negation and not the contradictory one. Hence, in
view of these observations, if contradictory negation could be defined by
semantical game rules, it would have to coincide with our dual (strong)
negation, which in fact it does not.
148 THE PRINCIPLES OF MATHEMATICS REVISITED

(b) In the sense that can be seen from (a), contradictory negation is
in a game-theoretical treatment of the semantics of IF first-order
languages inevitably parasitic on the dual negation.
(c) Another consequence of (a) is that contradictory negation can
syntactically speaking occur only in front of an entire sentence (closed
formula). It cannot be prefixed to an open formula, for then you would
need a game rule to handle a substitution-instance of that open
formula when you reach it in a semantical game.
There are certain exceptions to this restriction, but they do not
change the main point. For one thing, —i can occur in front of an
atomic formula for no further rule applications are needed in that case.
Furthermore, one can obviously allow contradictory negation to
occur within the scope of &, v and —i. Even in the quantificational
contexts, the difference between and i is sometimes immaterial.
Consider, for instance, a game of the form The first move
here is made by the initial verifier. He (she, it) has to choose a member
bedo(M) (of the domain of the model in question) such that he (she, it)
has a winning strategy in G(S[b]). But this means that the initial
verifier fails to have a winning strategy available in [x]) if and
only if there is none in any game of the form G(S[d])for any dedo(M).
But what this means is that
i (7.27)

is true if and only if

(Vx)i S[x] (7.28)

is likewise true, with the initial quantifier (Vx) being governed by its
usual game rule. In other words, the symbol combination
always makes perfect sense when it occurs sentence-initially, unlike (as
you can easily see) the symbol combination (Vx)i.
There is one more exception to the ban on contradictory negation
inside formulas. When a negation-sign is prefixed to an atomic
formula, it clearly does not matter whether it is viewed as expressing
a dual or a contradictory negation. This observation does not matter
a great deal for the theory of IF formal languages, but it helps us to
understand some aspects of negation in natural languages.
The result of adding the contradictory negation i to IF first-order
logic creates a type of logic which may at first seem somewhat strange,
both syntactically and semantically. Syntactically, because the way
i can occur is restricted in an unusual way. Semantically, because the
 NEGATION IN IF LOGIC 149

semantical rules for n is of a kind different from the semantical rules

for the other logical constants. Methodologically, it is especially
interesting to see the interplay between syntactical and semantical
rules here. For instance, if we had tried to introduce i in such a way
that it could occur in front of any formula or subformula also insider
a sentence, we could not have given semantical rules for the kind of
negation so introduced. One corollary therefore is that we cannot
assume that each logical constant, not even each propositional con-
nective, can be introduced by recursive formation rules of a usual sort,
in spite of the glib assumption of some philosophers that we must
always be able to do so. It could be argued that this syntactical dogma
has been instrumental in smuggling the hopeless semantical as-
sumption of compositionality into the philosophers' and linguists'
theories.
(d) Strictly speaking, we thus have two different extensions of IF
first-order logic. If we merely add all contradictory negations of IF
first-order sentences (closed formulas), we obtain what we will call
extended IF logic. If we also add truth-functional combinations of
(formulas obtainable by means of &, v and i from) IF first-order
formulas, we obtain what will be called truth-functionally (TF) ext-
ended IF first-order logic.
(e) By moving to extended IF first-order logic, one of the unsatisfac-
tory features of its unextended variant can be avoided. In unextended
IF logic, no deduction theorem holds. Given two sentences S1 and S2,
there is in general no sentence S0 which is valid if and only ifS2 is true
in every model in which S1 is true. For instance, (S1 S2) is valid if and
only if S2 is true in each model in which is not false, which is in
general a different matter altogether.
In contrast, in extended IF first-order logic, i v fills the bill,
being true if and only if S1 is not true or S2 is true.
This shows the rationale of the extended IF first-order logic as
compared to its unextended version. Looking back at the two main
functions of logic that were distinguished from each other in Chap-
ter 1, we can see that the notion of contradictory negation and
therefore the extension of IF first-order logic is vital for the deductive
function of logic. Without the extension, the deduction theorem does
not hold. And its not holding means that an inference from the truth of
S1 to the truth of S2 is not guaranteed by the truth of any formula of
the language in question. Hence one of the purposes of the extension is
to facilitate the deductive function of logic.
150 THE PRINCIPLES OF MATHEMATICS REVISITED

(f) The introduction of weak (contradictory) negation nevertheless

necessitates a review of the basic concepts of IF first-order logic. The
reason is that I have so far dealt with two sentences that the true in the
same models as being logically equivalent, even though they can be
false in different sets of models. Two sentences that are logically
equivalent have equivalent contradictory negations, but not always
equivalent strong (dual) negations. Moreover, second-order transla-
tions of first-order sentences preserve only truth, contradictory nega-
tion, and logical equivalence as it has so far been defined, but not
necessarily falsity. Hence two first-order sentences may have the same
second-order translations, but different strong negations. According-
ly, two first-order sentences may be logically equivalent but one of
them can be impossible to represent as an ordinary first-order sen-
tence in the sense that its strong negation differs from contradictory
negation.
All this may at first seem confusing, but there is nothing paradoxi-
cal about it. What follows is that we must be careful when using the
notion of equivalence as applied in IF first-order logic. It might in
fact be appropriate to distinguish the strong equivalence of S1 and
which preserves truth and falsity, from their weak equivalence, which
is the sense of logical equivalence so far employed and which means
only that S1 and are true in the same models.
One consequence that follows from here is that the logical equiv-
alence of the second-order translations of S1 and guarantees only
their weak equivalence, but not their strong equivalence. Thus in
a sense IF first-order logic seems richer than second-order logic in
that it facilitates finer distinctions than the latter. This is only
apparent richness, however, for the same distinctions could be made
on the second-order level, albeit by using nonstandard notions of
negation.
After having explained the idea of extended IF first-order logic, it
is time for me to return to the crucial question of truth definitions.
We have to ask: Why doesn't the contradictory negation recreate the
liar paradox? Indeed, some of the versions of the paradox are
formulated precisely in terms of Gödel-type sentences which say, so
to speak, "I am not true" rather than "I am false".
But how could we construct such a sentence? Presumably by
means of the diagonal lemma. But in order to use the diagonal
lemma here, we would have to apply it to an open formula. i
where T[x] is the truth predicate. But such a formula is ruled out by
 NEGATION IN IF LOGIC 151

the restrictions as to where —i can occur in well-formed formulas. It

cannot occur prefixed to open formulas.
(g) What this means is that it makes sense to apply the notions of
truth and falsity also to an extended IF first-order language. It must
be noted, however, that extended IF first-order languages do not
allow for a truth definition that can be formulated in the same
language. For, if we were to do so, we would have to say in that
language that, for the Gödel number g(—i S) the contradictory
negation —i S of any formula S,
T(g(i S)) if and only if not T(g(S)) (7.29)
where q(S) is the Gödel number of S. But we cannot replace "not" in
(7.29) by because it would then be prefixed to an open formula,
and in particular to an open formula in the scope of a sentence-initial
universal quantifier.
I do not see that this constitutes an objection to extended IF
languages. Just because the notion of contradictory negation is so
natural, I do not see that it has to be captured by a truth definition.
For another thing, any finite number of sentences of the form
T(g(i T(g(S)) (7.30)
can be added to an extended IF first-order language without any
problems. Furthermore, in extended IF first-order languages we can
have a partial truth definition which applies to all sentences (closed
formulas) S that do not contain contradictory negation Since
each i occurs prefixed to such an S, this truth definition indirectly
determines the truth and falsity of all sentences of extended IF
first-order languages, and hence does the job of a truth definition as
fully as can be expected. In this qualified sense, truth definition is
possible also in extended IF first-order languages.
In the sense that appears from these remarks, a truth definition for
an unextended IF first-order language indirectly defines the notion
of truth also for the corresponding extended IF first-order language.
This is in fact the best one can possibly do here, in the sense that
the limitations that rule out a translation of a universal generaliz-
ation of (7.29) in one's object language are unavoidable. For if these
restrictions were removed, we could formulate a truth predicate for
an extended IF first-order language. All we have to do is to add to the
conjunction of clauses in the predicate something like the following:
(Vx)(Vy)((Neg(x,y) (, X(x)4-+X(y))) (7.31)
152 THE PRINCIPLES OF MATHEMATICS REVISITED

where Neg(x, y) is the relation that holds between x and y if and only
if y is the Gödel number of the contradictory negation of the formula
with the Gödel number x. This relation is obviously representable as
soon as our language contains a modicum of elementary number
theory.
But the extended truth predicate allows for the liar paradox. The
paradox is obtained by applying the diagonal lemma to the formula
—i where Tr*(x) is the extended truth predicate. The diagonal
lemma yields a number n such that
g(_iTr*(n))=n (7.32)

But (7.32) is paradoxical. Hence the restrictions mentioned are

unavoidable.
(h) Consider now the fragment of extended IF first-order logic
which contains only sentences of the form S, where S is a sentence
of the unextended IF first-order logic. This fragment has several
remarkable properties. For instance, we can test whether i S is
logically true by trying to construct a model for S. As was pointed
out in Chapter 1, there is a complete set of rules for doing so. Hence,
—iS is logically true (i.e., true in every model of the underlying
language) if and only if this construction process comes to a dead
end. But this means that the logic of sentences of the form n S (where
S is an IF first-order sentence) is semantically complete.
(i) The results of the last two sections will be put in an interesting
perspective by further observations. It was said earlier that IF first-
order logic is much stronger than ordinary first-order logic. Ex-
tended IF first-order logic is even stronger. How strong? That
depends on how you measure strength. In Chapter 9 it will be
pointed out that there is a sense in which one can do all of ordinary
mathematics by means of the extended IF first-order logic, not just
notationally but model-theoretically. Among other things, for each
normal mathematical theory there is one in an extended IF first-
order language having mutatis mutandis the same models.
What such results imply is that extended IF first-order logic is an
infinitely better candidate for the role of a mathematical mat hesis
universalis than, for example, set-theoretical languages. The reason is
that by giving a characterization of truth for extended IF first-order
languages (albeit only via a truth definition for the corresponding
unextended IF first-order language) we can in effect fulfill Carnap's
dream and give a characterization of mathematical truth (math-
 NEGATION IN IF LOGIC 153

ematical validity, if you prefer), and to give it in the same language by

means of which you can in principle do your mathematics. For any
sufficiently rich mathematical theory one can then characterize the
concept of truth for it in that very same language. This is something that
you just cannot do fully in set theory, as will be shown in Chapter 8.
How important this fact is can be seen by recalling the deductive
incompleteness of most mathematical theories. What this means is
that an important part of a mathematician's work is the search for
new deductive principles suitable for the theory that he or she is
pursuing. My results show that if that theory is formulated in an
extended IF language, that search can be conducted by means of the
very same language. A language like that of set theory which does
not allow the formulation of descriptively complete theories does not
enjoy a comparable privilege. Thus my results have important
consequences for the formulations of mathematics in general.
As an example of these consequences, it may be pointed out that in
extended IF logic one can formulate a descriptively complete axiom
system for elementary number theory. In order to do so, one
basically needs, over and above the usual first-order Peano axioms,
an axiom of induction as strong as its second-order version
(VX)((X(O) & (Vy)(X(y) X(y + 1)) D (Vy)X(y) (7.33)

But (7.33) is a sentence, hence the contradictory negation of

a sentence. Since this sentence has an equivalent IF first-order
translation, (7.33) has an equivalent in the corresponding extended
IF first-order language. And this implies that elementary number
theory admits of a descriptively complete axiomatization by means
of extended IF first-order logic.

(D) Negation is a Siamese Twin concept

Our results put the entire concept of negation within a surprisingly
new light. Logicians and philosophers sometimes think of negation
as a simple and unproblematic idea. All that is involved, so it might
seem, is an inversion of truth-values. For instance, in the Tract atus
Wittgenstein held that the negation of a pictorially interpreted
sentence is not only also a picture, but the same picture, only with its
polarity reversed. Perhaps logicians and philosophers should have
taken a cue from linguists, who have found ordinary-language
negation a complex and puzzling phenomenon.
154 THE PRINCIPLES OF MATHEMATICS REVISITED

Our results lead to a striking conclusion. What they show is that in

any sufficiently rich language, there will be two notions of
negation present. Or if you prefer a different formulation, our ordi-
nary concept of negation is intrinsically ambiguous. The reason is
that one of the central things we certainly want to express in our
language is the contradictory negation. But it was found earlier in
this chapter (cf. (a)—(b) above) that a contradictory negation is not
self-sufficient. In order to have actual rules for dealing with negation,
one must also have the dual negation present, however implicitly.
What is truly remarkable here is not just the possibility of having
two different notions of negation present in one's language, but the
virtual necessity of doing so in sufficiently strong languages. The
crucial fact here is not that we can introduce a contradictory
negation in the way we have done. Rather, it is the fact that it must be
so introduced. The contradictory negation cannot be introduced by
means of any game rules. Hence it has to be different from the logical-
ly primary (dual) negation that can be defined by the game rules.
Another important fact is that the strong (dual) negation can be
introduced in only one way. Its properties are completely fixed in an
IF first-order language.
What this implies for the semantics of natural languages is clear.
We might as a thought-experiment think of the language community
as if it were facing the task of creating a sufficiently strong artificial
language for roughly the same purpose as a natural language. Such
an imaginary legislative linguist faces an interesting problem. On the
one hand, the most important thing that has to be expressed in such
a language is clearly the contradictory negation. But it cannot be the
only notion of negation present. Something else has to be involved in
the language, however tacitly. For we just cannot formulate ad-
equate semantical or syntactical rules for the language in question in
terms of the contradictory negation alone. Without the dual nega-
tion, one's language is just not adequate for the purposes of logical
and semantical processing. (For instance, the rules of semantical
games have to be formulated in terms of the dual negation.) If our
imaginary linguistic inventor uses only one symbol for negation,
then he or she has to put that one symbol to two essentially different
uses. This inevitably leads to complications of some sort or the
other.
Now the language community is in an essentially similar position
as the imaginary language designer, for in natural languages we seem
 NEGATION IN IF LOGIC 155

to be dealing with one negation only, namely, the contradictory one.

Accordingly, the behavior of negation in natural languages is likely
to exhibit curiosities caused by the need of also using, however
tacitly, the other (dual) negation. This suggests an interesting per-
spective from which to view the phenomenon of negation in natural
languages.
I cannot attempt here a full-scale discussion of negation in natural
languages. I am in any case fully convinced that the results reached in
this chapter will also be an essential part of any half-way satisfactory
theory of negation when it comes to natural languages. For one
thing, in spite of the confusing complexities of natural languages,
several features of the behavior of negation in English become more
understandable when they are compared with what we find in
extended IF languages.
Among other things, we obtain an interesting perspective on the
bifurcation of natural-language negation into verbal and sentential
negation. It would be a serious oversimplification simply to assimi-
late this dichotomy to the distinction between strong and weak
negation in extended IF languages; but the two contrasts are not
unrelated, either. When a negation-sign occurs in front of an atomic
formula in a formal language, the difference between strong (fal-
sifying) negation and weak (contradictory) negation disappears.
Hence verbal negation, which very roughly speaking negates the
unquantified part of a sentence, can be construed as the contradic-
tory negation which the natural language negative ingredient obvi-
ously expresses in the first place.
At the same time, the vagaries of contradictory negation in
natural languages like English become understandable. The basic
fact is that explicit ordinary-language negation is the contradictory
one. Complications arise partly because such a negation is not
self-sufficient, and partly because it cannot always occur meaning-
fully in all contexts in which it should, purely syntactically speaking,
be admissible. Hence it may be expected that there are semantical
regularities which semantically speaking move either to a sentence-
initial position or to a position which corresponds to fronting an
atomic formula. For instance, the fact that we can easily form the
contradictory of an English sentence by prefixing it by "It is not the
case that" is like introducing a sentence-initial contradictory nega-
tion into an IF first-order logic without being able to introduce it
into any position except the sentence-initial one. The difficulty or
156 THE PRINCIPLES OF MATHEMATICS REVISITED

perhaps even impossibility of giving effectively applicable rules for

forming contradictory negations in English is in keeping with the
inevitable absence of contradictory negation from an entire IF
first-order language, without extending it by a new type of formation
rules and by a new type of truth-condition.
There even seems to be linguistic regularities in English which are
closely related to the awkwardness of locating negation between
a sentence-initial position and a minimal-scope position in an
English sentence, apart from truth-functional compounding. One
regularity is seen by raising the following question: When can not be
prefixed to a quantifier phrase in English? The following examples
illustrate some of the acceptable constructions:
Not every Scotsman is stingy. (7.34)
Not a single student failed.
Not many runners managed to finish.
In contrast, the following constructions are unacceptable:
not any (7.35)
not each
not some
not few
not several
What makes the difference? It turns out that the quantifiers
involved in (7.35) all have a wider scope than negation while those in
(7.34) do not. Hence the constructions (7.35), if they were admissible,
could push negation into the sentence-internal no-man's-land. In
contrast, the quantifiers occurring in (7.34) do not have the right of
way in relation to negation. Therefore they let negation stay
sentence-initial, and hence presumably a contradictory negation.
This explanation is reinforced by observing that the constructions
instantiated in (7.34) are ceteris paribus admissible only in the subject
position, but not in the object position. For instance, one can say
Not every hunter shot a lion. (7.36)

but not
A hunter shot not every lion. (7.37)

The reason is that if not occurs in a phrase which occupies an object

position, it can be within the scope of other quantifiers, typically
 NEGATION IN IF LOGIC 157

quantifiers in the subject-position phrase. This would once again

land negation in a position where contradictory negation cannot
occur, if we assume that the English is like extended IF language.
This regularity has nothing to do with the "intended" meaning of
expressions like (7.37), for it is acceptable to say
Not every lion was shot by a hunter. (7.38)
There are other linguistic regularities which become understand-
able in the light of the behavior of negation in IF languages. One of
them is the fact that overt negation can be a barrier to anaphora.
This fact has been noted in the literature, and Irene Heim (1982, pp.
115—117) has even used its alleged inexplicability in game-theoretical
semantics as a basis for an attempted criticism of GTS. Now it can be
seen that such attempted objections can be turned into evidence for
game-theoretical semantics, and not against it.
This phenomenon is exemplified by the fact that anaphora is
possible in the former of the following two sentences, but not in the
latter:
Some student passed the examination. She must have studied
very hard. (7.39)
Not every student failed to pass the examination. She must
have studied very hard. (7.40)
Since (7.39) and (7.40) are logically equivalent, it might at first seem
puzzling that anaphora is acceptable in one of them but not in the
other. This fact is nevertheless easily accounted for. In Hintikka and
Kulas (1985), a theory of anaphora is developed within GTS. Apart
from the details the basic idea is clear, and it is the only thing I need
to evoke here. The idea is that, in any play of a semantical game,
anaphoric pronouns refer to certain individuals introduced earlier in
the course of the same play. For instance in (7.39), the first game rule
to be applied is the rule for some. It involves a choice of an individual,
say Susan, from the relevant domain, whereupon the game is con-
tinued with respect to
Susan passed the examination. Susan is a student. She must
have studied very hard. (7.41)
When some time later in the game the pronoun She is dealt with,
there will be an eligible value for it available, namely, Susan. Hence
anaphora is possible in (7.39).
158 THE PRINCIPLES OF MATHEMATICS REVISITED

In contrast, (7.40) is not interpreted directly through any game

played with it. Since its first sentence obviously involves a (sentence-
initial) contradictory negation, it is interpreted only indirectly as the
contradictory negation of
Every student failed to pass the examination. (7.42)

Even if a game with (7.42) is involved in the interpretation of (7.40), it

does not automatically restore the possibility of anaphora in (7.40).
There are two concurrent reasons for this. On the one hand, what the
first sentence of (7.40) says is that there is no winning strategy for
a game with (7.42). Hence it does not give any clues as to what kind of
play with (7.42) might be relevant here, and consequently what
individuals might be chosen by the players in such a game. Therefore
there are no well-defined individuals in the offing to serve as the
value of every in (7.40) and hence as a value of the anaphoric pronoun
She.
On the other hand, individuals introduced by applications of rules
for universal quantifiers (like the every in (7.42)) are not automati-
cally available as values of anaphoric pronouns. (I will not discuss
here the reasons and limits for this regularity.)
In any case, this regularity has nothing to do with the quantifier
every as such. For we have a similar situation with pairs of sentences
like the following:
Few students passed the examination. They must have
studied very hard to be able to do so. (7.43)

Not many students passed the examination. They must

have studied very hard to be able to do so. (7.44)

There are prima facie exceptions to the rule that negation is

a barrier to anaphora. They are nevertheless precisely the kinds of
exceptions that literally prove the rule. In fact, they are examples in
which the semantical game with the unnegated sentence is brought
in and considered explicitly or implicitly, usually to provide a reason
why the unnegated sentence is false. The following is a case in point:
Nobody stole your diamonds, for he would have had
to scale a ten-foot wall to do it. (7.45)

An even simpler regularity which becomes understandable from

our vantage point is the semantical ordering principle that gives
 NEGATION IN IF LOGIC 159

negation priority over every also when the negation ingredient is

within the syntactical scope of every. For instance,
Everything that glitters is not gold. (7.46)

does not have the logical form

(Vx)(x glitters —i (x is gold)) (7.47)
but rather
—i (Vx)(x glitters x is gold) (7.48)
Why? Because the former reading would place a contradictory
negation in the scope of a universal quantifier where it is not
interpretable in all similar sentences. What is interesting about this
phenomenon is that it occurs in a wide variety of languages, some of
them completely unrelated to each other. It therefore cries out for
a theoretical explanation. The results reached in this chapter provide
such an explanation.
This list of explanations of natural language phenomena could be
continued. Similar problems have in fact been noted by linguists, for
instance in connection with the phenomena of neg-raising (cf. e.g.,
Horn 1989, pp. 308—330). Many of them admit of explanations along
the same lines as the phenomena noted here. What has been said
suffices to show that IF first-order logic will play a crucial role in
such a study.
It is to be noted also that all we need to solve the informal liar-type
paradoxes is to acknowledge that negation behaves in natural
languages in essentially the same way as in extended IF first-
order languages. An analysis of the self-referential mechanism
which leads to such prima facie paradoxes is needed only for the
purpose of arguing for the naturalness of this solution, and not
for the solution itself. For instance, consider the archetypal liar
sentence
Sentence (7.49) is false. (7.49)
What kind of procedure is involved in trying to verify (7.49)? Such
a language game would have to involve some rule like the follow-
ing:
When the game has reached a sentence of the form, "Sentence
(n) is false", the verifier has to look up sentence #n. The
game is then continued with respect to its negation. (7.50)
160 THE PRINCIPLES OF MATHEMATICS REVISITED

A mirror-image rule would be

When the game has reached a sentence of the form, "Sentence
(n) is true", the falsifier has to look up sentence #n. The
game is then continued with respect to it. (7.51)

Applied to (7.49), a game with the rule (7.50) leads to a loop, and
hence to an infinite play of the game. This does not by itself prevent
us from speaking of winning and losing here. We know from game
theory that there can be perfectly reasonable ways of defining who
wins and who loses in infinite and not only finite games. However, in
this particular case, the infinite play is symmetrical with respect to
the two players. Hence there is no reasonable rule for winning or
losing that would declare either player the winner. Hence (7.49)
cannot be verified, and by symmetry it cannot be falsified, either.
Therefore (7.49) cannot naturally be considered either true or false.
A similar line of thought can be applied to
The sentence (7.52) is true. (7.52)

In this case, however, the infinite play of the game is not symmetrical.
One of the players, namely, nature (the initial falsifier) can be
"blamed" for the play's infinitude. This can naturally be used as
a basis for stipulating a sentence like (7.52) to be true. The idea is that
if one player bears the whole responsibility for the play going on to
infinity, then the blameless one wins (cf. here Hintikka and Rantala,
1976).
Hence an attempt to vindicate the strengthened liar paradox by
considering a sentence like
The sentence (7.53) is not true. (7.53)

where "not true" expresses contradictory negation also fails, for

(7.53) is simply false.
You do not have to like the details of this treatment of informal
liar-type paradoxes. For even though this solution might very well
look like an ad hoc procedure, it in any case serves to illustrate the
ease with which liar-type paradoxes can be dealt with in natural
languages as soon as it is admitted that negation behaves in natural
languages basically in the same way as in extended IF first-order
languages.
The kinds of definitions of winning and losing and a fortiori of
truth and falsity are interesting in principle (among other things)
 NEGATION IN IF LOGIC 161

because they show that there are reasonable ways of defining truth
and falsity also for what Kripke calls ungrounded sentences contain-
ing the truth predicate (see Kripke in Martin 1984, p. 57).
(j) Among the other features of the concept of negation that have
to be deconstructed is the so-called law of excluded middle. Since
there must be two different concepts of negation in a sufficiently
strong language, the question of the validity of the law of excluded
middle is intrinsically and inevitably ambiguous between the two
negations.
Moreover, for each of the two negations, the corresponding
question concerning the tertium non datur admits of a simple answer.
For the contradictory negation, the law of excluded middle holds
virtually per definitionem. For the dual (strong) negation, it fails
almost equally trivially. What is particularly remarkable is that both
answers represent straightforward logico-combinatorial facts which
are not tied to any particular epistemological or other philosophical
assumptions.
One remarkable thing about the failure of tertium non datur in IF
first-order logic is that this logic is in some sense perfectly classical.
The failure is a consequence of changes in the ordinary first-order
logic which were motivated completely independently of all "non-
classical" ideas, such as constructivism or intuitionism. Hence my
results put the entire contrast between classical and nonclassical
logic in a new light. Or, rather, they confound it totally by showing
that the received logic (received ultimately from Frege and Russell)
does not deserve the epithet "classical" except in the sense of being
the logic that is taught in classrooms these days.
This point is best explored in connection with the reconstructions
of constructivistic ideas discussed in Chapters 10 and 11.
(k) In a truth-functionally extended IF first-order language, we
can finally express normal truth-functional conditionals, for
example,
(i S1 v S,) (7.54)

which can be abbreviated

(S1 DTS2) (7.55)
Hence a truth-functionally extended IF first-order language is a hap-
pier medium of actually proving propositions logically than the
unextended IF first-order logic. This gain in syntactical manipulabil-
162 THE PRINCIPLES OF MATHEMATICS REVISITED

ity is nevertheless inversely related to the perspicuity of the semanti-

cal situation in the two types of languages. In the original model-T
variant of IF first-order languages we did not have truth-functional
conditionals, which makes a proof theorist's life difficult, but we have
all sorts of nice model-theoretical results. In contrast, in truth-
functionally extended IF first-order languages we do have truth-
functional connectives, ordinary modus ponens and so forth, making
chains of actual logical inferences much easier. But that means being
banished from the model-theoretical Eden of the original plain IF
first-order logic. There is a price to be paid here, no matter which
way we go. A main restriction that must be observed here is that the
truth-functional conditional D1 can only be used outside quanti-
fiers, like the truth-functional negation. Fortunately, this presents no
new obstacles to the use of conditionals in actual logical inferences.
(1) My results admit of a striking formulation in Wittgensteinian
terms. Wittgenstein spoke of the limits of language. Usually, the
limits of what a formal language can express are thought of in terms
of limitations to the expressive power of the language or perhaps in
terms of limitations of the totality of entities it can deal with. What
we have found shows that one of the true limits of language lies in the
inexpressibility of contradictory negation or, if such a negation is
introduced by afiat, in the consequences of that forced expressibility.
For instance, the crucial but inevitable restriction on IF first-order
logic is the inexpressibility of contradictory negation, outside the
ordinary first-order fragment. And the introduction of contradictory
negation destroys most of the nice metalogical properties of IF
first-order logic, including the definability of truth in the language
itself.
 6

Who's Afraid of Alfred Tarski?

Truth Definitions for IF First-Order Languages

In earlier chapters, it was seen how crucial the concept of truth is

in the foundations of logic and mathematics. An especially central
role was seen to be played by the question of the definability of truth.
The new languages explained in Chapter 3, IF first-order langua-
ges, open up a way of actually freeing the expressibility of the con-
cept of truth from the serious problems that have been seen to beset
it.
As was pointed out in Chapter 1, discussions of truth and its
definability have in the last sixty years been profoundly influenced
by the ideas of Alfred Tarski (1935, 1956a).' This influence has been
strong and widespread. The following aspects of Tarski's work are
especially relevant here:
(i) In his classic monograph, Tarski showed how the concept of
truth can be defined explicitly and precisely for certain types of
formal languages.
The nature of Tarski-type truth definitions was explained
briefly in Chapter 1. Among other things, the following features
of these definitions were pointed out:
(ii) A Tarski-type definition of truth is an indirect one, defining the
notion of truth in tandem with the notion of satisfaction.
(iii) Tarski proved, given certain assumptions, that a definition of
truth for a formal language can only be formulated in a metalan-
guage which is stronger in certain specifiable respects than the
given object language. In our day, Tarski-type truth predicates
for first-order languages are usually expressed by second-
order formulas — that is to say, by formulas whose only non—first-
105
106 THE PRINCIPLES OF MATHEMATICS REVISITED

order ingredient is a formula-initial string of second-order exis-

tential quantifiers.
(iv) Tarski argued that the concept of truth cannot be defined for
what he called the colloquial language.
(v) A less deep but eminently popular suggestion of Tarski's was to
use as a touchstone for truth definitions that they have as their
consequence all the relevant instances of the so-called T-schema:

(T) H is true
where "H" is a placeholder for a quote or a structural description
of the sentence which is to replace "p."

It will be argued in this chapter that we have to revise our ideas

about all the five features (i)—(v) of Tarski-type truth definitions.
It will also be seen that truth definitions can be freed from the
problems or at least limitations — which they were seen in
Chapter 1 to be subject to.
Tarski's procedure is generally taken to be normative. However, it
is seldom, if ever, explained what the reason for this normative
character of Tarski's definition is, or even what Tarski's own ration-
ale was in proceeding in the way he did. There is nevertheless a
simple answer available, which is the same for both questions. This
answer is closely related to the presuppositions of the approach
used here.
Historically, an important part of Tarski's background was an
early form of the idea of categorial grammar. Now any approach like
categorial grammar presupposes certain things of the semantics of
the language whose grammar it is supposed to be. It presupposes
that the language in question satisfies the requirement which lin-
guists usually call compositionality. (For this notion, see Partee 1984
and Pelletier 1994.) Philosophers sometimes refer to it as the Frege
principle, and Frege did indeed pay lip service to it. As usually
formulated, it says that the meaning of a complex expression is
a function of the meanings of its constituent expressions. It is
nevertheless appropriate to extend the idea of compositionality to
say that all the relevant semantical attributes of an expression (and
not only just its meaning or truth) are functions of the semantical
attributes of its constituent expressions (not just of its meaning). The
major impact of the principle of compositionality on logical and
linguistic theorizing is to allow what is usually referred to as
 TRUTH-DEFINITIONS FOR IF LANGUAGES 107

recursive definitions (or other kinds of recursive characterization) of

the relevant semantical properties. That is to say, we can specify the
conditions for the applicability of the semantical attribute in ques-
tion to the simplest possible expressions, and then specify, step by
step, for each operation of forming complex expressions out of the
simpler ones, how applicability of that semantical attribute to the
complex expression is derived on the semantical attributes of its
constituent expressions.
This idea of compositionality has loomed large in recent linguistic
and philosophical theorizing. Conformity with the principle of
compositionality is generally (though not universally) thought of as
a major desideratum which a satisfactory linguistic theory should
satisfy. Donald Davidson apparently at one time even thought of
compositionality as a precondition for the learnability of language
(see Davidson 1965).
The real impact of the principle of compositionality nevertheless
has not been emphasized in recent discussion. It is illustrated by the
role of the principle in facilitating recursive definitions of semantical
attributes. Such recursive definitions proceed from simpler expres-
sions to more complicated ones. They are not possible unless the
attribute to be defined is semantically speaking context-independent.
The main function of the principle of compositionality is to secure
such semantical context-independence.
It seems to me that the principle of compositionality was one of
the most important presuppositions of Tarski's work on the concept
of truth. For one thing, it was fairly obvious that it was the principle
of compositionality that motivated Tarski's reliance on recursive
definitions in his approach to truth. Furthermore, it is the same
commitment to inside-out procedure that forced Tarski to define
truth for his formal languages, not alone, but in tandem with the
notion of satisfaction. In a compositional procedure, one defines the
semantical attributes of an expression in terms of the semantical
attributes of its component expressions. But truth can only be
attributed to sentences (closed formulas), whose component expres-
sions are often not sentences but open formulas. For this reason,
truth can be defined compositionally only in conjunction with some
other semantical concept like satisfaction. This observation helps us
to understand Tarski's procedure in his classic paper. It also offers
a glimpse of the difficulties in which adherence to the principle of
compositionality easily leads a linguistically minded logician or
108 THE PRINCIPLES OF MATHEMATICS REVISITED

logically minded linguist. For clearly it would be preferable to define

the notion of truth directly without using the concept of satisfaction
as a middleman. For one thing, we are likely to have sharper pre-
theoretical ideas about the notion of truth than about the notion of
satisfaction.
Tarski points out himself the roots of his ideas in the notion of
semantical category which goes back of Lesniewski and indirectly to
Husserl (cf. Tarski 1956a, p. 215). As is spelled out in various explicit
developments of categorial grammar, the leading idea here is to
mirror the semantical structure of a sentence by the way in which it is
constructed syntactically from its basic components. Whether or not
such parallelism between semantics and syntax inextricably presup-
poses compositionality, there is little doubt that in praising the
naturalness of the notion of semantical category Tarski is in effect
extolling the virtues of compositionality.
In short, adherence to the principle of compositionality is the
answer, or at least a large part of the answer, to the question as to
why Tarski formulated his truth definition in the way he did.
Furthermore, I suspect that the same principle was operative in
Tarski's claim that truth is not definable for our ordinary "collo-
quial" language. Tarski blamed this alleged impossibility on the
irregularities of natural languages. I suspect that a failure to conform
to the principle of compositionality was the first and foremost
"irregularity" that Tarski discerned in natural languages.
It is not that prima facie counterexamples to compositionality
have not been noted in the literature. The nonsynonymy of the
following pairs of sentences is a case in point:
Mary will be surprised if anyone comes. (6.1)

Mary will be surprised if everyone comes. (6.2)

Likewise, the context affects the meaning of a sentence like

Jim can beat anyone. (6.3)

when it is imbedded in a belief sentence:

John doesn't believe that Jim can beat anyone. (6.4)

For what John is said to disbelieve in (6.4) is not (6.3) but

Jim can beat someone. (6.5)
 TRUTH-DEFINITIONS FOR IF LANGUAGES 109

It is even possible to give a systematic account of the reason why

compositionality fails in sentences like (6.1) or (6.4). An important
explanation for one type of counterexample lies in what I have called
the special ordering principles of GTS. They do not respect the order
in which an expression is assembled out of its component expres-
sions (see here Hintikka and Kulas 1983, pp. 233—234; 1985, pp.
180—181). In spite of all this, linguists have not taken prima facie
counterexamples to compositionality as seriously as they ought to
have done. Maybe they have never read those chapters of the
treatises in the philosophy of science that deal with the significance
of anomalies in heralding theory changes.
One of the most important kinds of impacts that IF first-order
logic has on our ideas about logic and language, is that it shows once
and for all the utter futility of trying to abide by the principle of
compositionality in our linguistic and logical theorizing. The reason
stares you in the face from the explanations that were given in
Chapter 3 of the nature of IF logic. By their very nature, all such
instances of quantifier independence (of any other kind of informa-
tional independence in logic) as cannot be dealt with in ordinary
first-order logic violate the principle of compositionality.
This could be dramatized by changing the notation that has been
used here along the lines mentioned in Chapter 3. Instead of
appending to each quantifier an indication of which earlier quanti-
fiers it is independent of, we could add to each quantifier a list of later
quantifiers that are exempted from its scope. For instance, instead of
writing my favorite sample formula in the notation so far used as
y, z, u] (6.6)

I could write it as
y, z, u] (6.7)

where the double slash//is the converse of the independence relation /.

I have argued in print that compositionality is a lost cause in the
study of the semantics of natural languages (see Hintikka and Kulas
1983, Ch. 10). This is not an impossibility claim. If you are a die-hard
believer in compositionality, there is a way in which you can always
enforce it. The prima facie violations of compositionality are due to
the semantical interaction of an expression with its context, as we
have seen. Now in principle you can try to build the laws covering
that interaction into the semantical entity you consider as the vali
110 THE PRINCIPLES OF MATHEMATICS REVISITED

of the expression in question. In practice, however, this would result

in intolerably complicated and completely unnatural semantics.
This unnaturalness is illustrated by the fact that a similar act of
semantical desperation is impossible in suitable formal languages.
There the resurrection of compositionality is made impossible by the
tacit conventions that govern the interpretation of logical formulas.
No perverse ingenuity can make the semantics of sentences like
(6.6)—(6.7) conform to compositionality.
The commitment of Tarski-type truth definitions to composi-
tionality is so deep that the glory and misery of the entire principle of
compositionality is to a large extent measured by the success and
shortcomings of Tarski-type truth definitions. Since I am in this
chapter showing the very important limitations of Tarski-type defini-
tions, I am ipso facto exposing certain serious theoretical shortcom-
ings of the principle of compositionality.
Conversely, to look at the bright side of things, the very existence
of IF first-order logic is an eloquent proof that a rejection of
compositionality is no obstacle to the formulation of a simple and
powerful logic. Indeed, the best argument against compositionality
as a general linguistic principle is the success of independence-
friendly logic in its different variants in the logical analysis of various
important concepts and of their manifestations in natural languages.
This is not the occasion to tell that success story, which is in fact still
continuing. A cumulative evidence of different applications is in any
case sufficiently impressive for me to rest my case against composi-
tionality for the purposes of this book (see here, e.g., Hintikka and
Kulas 1983 and 1985, Hintikka and Sandu 1991 and 1996).
The dependence of Tarski-type truth definitions on the assump-
tion of compositionality makes them suspect even when they do
work.
Because IF first-order languages are not compositional, Tarski-
type truth definitions do not apply to them. You just cannot give
satisfaction-conditions for expressions like
y, z, u] (6.8)

when you do not know what kinds of dependence and independence

relations obtain there between (Ry), (au) and the quantifiers further
out to which x and z might be bound (cf. 6.7).
This failure of Tarski-type truth definitions in IF first-order logic
is nevertheless perhaps less obvious than it might at first seem. It is at
 TRUTH-DEFINITIONS FOR IF LANGUAGES 111

its clearest when we use the double-slash notation. For then there is
no difference between two component expressions whose variables
are in one of them independent of certain quantifiers further out,
while in the other they are bound to such quantifiers. For instance,
the universal-quantifier-free constituent expressions of the following
two expressions are identical:
z, u] (6.9)

y, z, u] (6.7)

If compositionality were in force, the same semantical entity should

be assigned to
y, z, u] (6.8)

in the two cases. But this would make it impossible to distinguish

(6.9) and (6.7) semantically if we abide compositionally and build our
truth-conditions and satisfaction-conditions for (6.9) and (6.7) from
inside out. For the "insides" of(6.9) and (6.7) are identical, and by the
time we come to the slashed quantifiers it is too late to try to
distinguish the two from each other semantically.
When the single-slash notation is used, we can make a distinction
etween the semantical entities assigned, for example, to
y, z, u] (6.10)

and
y, z, U] (6.11)

But this possibility will not help compositionalists decisively. As

a minor point, compositionality will not be true in the literal sense of
the word, for (6.11) is not a component expression of anything as it is
not a well-formed formula. Whether or not it can occur as a part of
a well-formed sentence depends on that part of the sentence which
lies outside (6.11).
This failure of compositionality might seem to be mere quibbling.
It is nevertheless indicative of the situation. In any case, there is
another reason for the impossibility of Tarski-type truth definitions
for IF first-order sentences. The difference between (6.10) and (6.11)
lies in the question whether the value of u depends on the value of x.
No matter how this dependence or independence is expressed in
assigning a semantical value to (6.11), by the same token this
112 THE PRINCIPLES OF MATHEMATICS REVISITED

assignment should indicate whether the value of u depends on those

of y and z. But this question does not depend on the expression (6.11)
alone, as compositionality requires. Because of the convention that
was shown in Chapter 3 to be needed in IF first-order languages, an
answer to the question depends also on whether y and z are bound to
universal or existential quantifiers. And this cannot be seen from the
component expression in question alone. Likewise, it cannot be seen
from (6.11) as to which of the variables x, y, z depend on which ones.
Yet this is needed if we want to assign a class of valuations to (6.11) as
those that satisfy it.
As a consequence, there is no realistic hope of formulating compo-
sitional truth-conditions for IF first-order sentences, even though
I have not given a strict impossibility proof to that effect.
This failure of Tarski-type truth definitions is an extremely serious
black mark against them. Most of the earlier failures of such
definitions took place at the outer reaches of ongoing logical re-
search and involved concepts that are at least prima facie unfamil-
iar — unless you are Leibniz and relish the kind of infinite analysis
that infinitely deep languages codify. In contrast, IF logic was seen
to be our true elementary logic involving nothing more than quanti-
fiers and propositional connectives. Hence the failure of Tarski-
type truth definitions in IF first-order logic shows that they do not
have a serious claim to being the basic normative kind of truth
definition.
In contrast to Tarski-type truth definitions, game-theoretical
characterizations of truth are easily available for IF first-order
languages. This is not much of a surprise, for a semantical game
starts from an entire sentence and not from its simplest constituents,
and gradually anatomizes it into simpler and simpler sentences. In
such an outside-in process, context-dependencies can easily be taken
into account.
Indeed, for each first-order sentence S (whether IF or not) its
second-order translation can serve as its truth-condition. For, as we
saw in Chapter 3, such a "translation" expresses precisely the
existence of a winning strategy for the initial verifier in the game G(S)
correlated with S. This is how truth is characterized in the first place
in GTS. The remaining part of the task of formulating an actual truth
definition is in effect to integrate all the truth-conditions into a single
definition of truth. It will turn out, Alfred Tarski notwithstanding,
that such a truth definition can be given for representative IF
first-order language in that language itself.
 TRUTH-DEFINITIONS FOR IF LANGUAGES 113

Let us approach such truth definitions step by step. I will first

discuss the definability of arithmetical truth. There the model in
which the truth or falsity of the sentences of the underlying language
is evaluated is the structure of natural numbers. First, I will likewise
assume that the language that is being used is an ordinary first-order
arithmetical language. Later, I will expand my horizon to take in
independence-friendly quantifiers and connectives also, plus the
fragment of the corresponding second-order language.
I assume that my readers are familiar with the technique of Gödel
numbering. The basic idea is completely straightforward. What is
involved is a coding of the syntax of the given formal object language
in question in the language of elementary arithmetic. For the purpose
of this discussion, it will be assumed that the object language itself
contains some suitable formalization of elementary (first-order) arith-
metic. The reason is to show that for certain object languages the truth
predicate can be defined in the object language itself.
The coding of the syntax of the object language by means of the
arithmetical part of itself can be done in different ways. All that
counts here is that certain formal relations between formulas can be
expressed in the language itself. In my metalogical discussion, I will
refer to the Gödel number of a formula S by means of corner quotes.
In other words, is the Gödel number of S. The numeral which
represents a natural number n will be referred to as n.
Then the properties of, and relationships between, different ex-
pressions whose representability plays a role in my line of thought
are relations between expressions of the following forms:
(i) n and n (6.12)
(ii) (S1 & S2) and S2
, S1

(iii) (S1 v S2) , S1 and S2

(iv) (Vx)S[x] n and S[n]
,

(v) , n and S[n]

If R(x, y) is a primitive two-place relation,

(vi) R(n, m) , n, m

Likewise for other primitive predicates and primitive functions.

(vii) Likewise for negated primitive predicates and functions.
(viii) , when S2 is the negation normal form of S1

This relation will be called R(x, y).

114 THE PRINCIPLES OF MATHEMATICS REVISITED

Later, the Gödel numbering technique will be also applied to the

part of the corresponding second-order language with indepen-
dence-friendly quantifiers allowed. (It will be assumed for simplicity
that all the initial second-order existential quantifiers are function
quantifiers.) Then we will consider relations between the following:
(ix) 1(Vy)(Vu)(3z/Vu)(3t/Vy)S[y, z, u, t)]1, k, 1, m, n and rS[k, I, m, n]l

How do I know that all these relations are representable in

elementary number theory and hence in the given IF first-order
language? A blow-by-blow account of an argument to this effect
would be too cumbersome to be given here. Fortunately, there is an
argument which is not only persuasive but resorted to by logicians
every day. It is an appeal to Church's thesis. It is known that all
recursive relations are representable in elementary number theory
(cf. e.g., Mendelson 1987, P. 143, Proposition 3.23). Church's thesis
says that every effectively (mechanically) decidable relation is recur-
sive. Since all the relations discussed here are fairly obviously
mechanically decidable, they are representable in the given IF first-
order language, for it was assumed to contain elementary arithmetic.
In some cases, the representability of these relations in a suitable
language of arithmetic is known from the literature.
For the purposes of a perspicuous exposition, I will not introduce
any special notation for the relations (i)—(ix) but use the corner quote
notation instead.
Once we have all the relations (i)—(ix) expressible in the language,
it is easy to see how a truth predicate can be formulated for different
languages. I will first show how a truth predicate for an ordinary
first-order language L can be formulated in the corresponding IF
first-order language. I will first formulate this predicate in the
fragment of the corresponding second-order languages. Once this is
done, one can simply translate that predicate back to the corre-
sponding IF first-order language. By a corresponding language, I
mean of course a language with the same nonlogical primitives.
The second-order truth predicate has an obvious intuitive mean-
ing. It says, when applied to the Gödel number y of a sentence, that
there exists a one-place predicate X which behaves in the way a truth
predicate should, and that y has this predicate. In other words, the
truth predicate has the form
& X(y)) (6.13)
 TRUTH-DEFINITIONS FOR IF LANGUAGES 115

Here TR[X] serves to guarantee that there is a winning strategy for

the initial verifier in a semantical game. In fact, it is a conjunction of
the following formulas:

(a) (Vx)(Vy)(Vz)((x = r(S1 & S2)1&y = rs11 &z = (X(x) D

(X(y) & X(z))))
(b) likewise for disjunction
(c) (Vy)(Vz)(Vu)((y = r(Vx)S[x]1 & u = 1S[z]1 & X(y)) X(u))
(d) (Vy)(Vz)(Vu)((y = r(R4s[x]l & u = rS[z]1 & X(y)) D X(u))
(e) If R is a two-place primitive relation, (Vx)(Vy)(X( R(x, y)
R(x, y)); likewise for other primitive predicates and negated
primitive predicates.
(f) (Vx)(Vy)((x = 1S11 & y = & N(x,y)) (X(x)t-÷X(y)))
where N(x, y) is the relation of the Gödel number of a sentence
to the Gödel number of its negation normal form

Once you understand this definition of the truth predicate (6.13),

you will realize that it does its job. For the property of being true (as
attributed to the Gödel numbers of sentences of a given ordinary
first-order language with finite number of primitive predicates and
functions) satisfies TR[X]. Hence if the sentence with the Gödel
number y is true, it satisfies (6.13).
Conversely, if it satisfies (6.13), from TR[X] you can see that there
is a winning strategy for the initial verifier. That player can always
make his or her choices in such a way that the Gödel number of the
resulting sentence has the property X. Indeed, the truth predicate
(6.13) is little more than a way of spelling out more fully the game-
theoretical truth definition explained in Chapter 2.
This observation can be formulated in a somewhat different way.
Consider, for the purpose, the statement

& (6.14)

in which the truth predicate is applied to the Gödel number of the

sentence S. It is easily seen that (6.14) is logically equivalent to the
game-theoretical (second-order) truth-condition of 5, as explained in
Chapter 2. This helps to explain the sense in which the truth predicate
(6.13) merely serves to integrate the game-theoretical truth-conditions
for different sentences.
116 THE PRINCIPLES OF MATHEMATICS REVISITED

The most remarkable thing about the truth predicate (6.13) is that
it is of the form. (The only higher-order quantifier in (6.13) is the
initial second-order existential quantifier (3X).) From this it follows
that it can be translated into the corresponding IF first-order
language2. Thus a truth predicate for a given ordinary first-order
language can be defined in the corresponding IF first-order lan-
guage; in other words without any higher-order quantifiers. Or, in
other words, without having to raise any questions about the
existence of higher-order entities.
How can we generalize the truth predicate (6.13) to other first-
order languages besides purely arithmetical ones? It turns out that
the natural way of doing so involves going back to the familiar
Tarski-type procedure of first defining satisfaction for the language
in question. The definition can be formulated in the corresponding
second-order language. The satisfaction relation Sat(x, f) holds
between the Gödel number x = IS1 of a formula S and a valuation
function f ifand only if f satisfies S in the sense explained in Chapter
1. As was also pointed out in Chapter 1 the valuation function f is
a mapping from natural numbers to the individuals of the model in
question. I assume here and in the following that the given first-order
language contains elementary arithmetic. Then the valuation func-
tion f is just an ordinary second-order object, and the existentially
quantified formula
1S1,f) (6.15)

is simply a formula of the corresponding second-order language.

When S is a closed formula (i.e., a sentence), (6.15) asserts that S is
true.
So why did I offer a different truth predicate (6.13) for arithmetical
languages? My reasons were pedagogical. First, the direct truth
predicate shows vividly how extremely close the notion of truth is to
both our ordinary ideas about truth and to the basic ideas of the
game-theoretical approach represented in this book.
Second, I wanted to bring out the fact that the difficulty of
extending the truth predicate (6.13) to other first-order languages
besides arithmetical ones is in a sense merely technical, and has
nothing to do with the deep problems of compositionality or of the
allegedly unavoidable second-order character of truth definitions for
first-order languages. The feature of (6.13) that makes a generaliz-
ation difficult is the quantification over numerals (names of num-
 TRUTH-DEFINITIONS FOR IF LANGUAGES 117

bers) in clauses (c) and (d) of the truth predicate. In order to extend
such a truth predicate to nonarithmetical languages, one needs to
have a name in the language for each individual in the models of the
theory in question, and one needs even to be able to quantify over the
class of all such names, just as in elementary arithmetic we have
a numeral representing each number and can quantify over the set of
all such numerals.
No matter whether we use the truth predicate (6.13) or (6.15), it is
of form, and hence can be translated to the corresponding IF
first-order language. In other words, either truth predicate is ex-
pressed without any quantification over higher-order entities. They
are purely nominalistic (combinatorial), and free from any set-
theoretical commitments. This is the first step in my exorcism of
Tarski's curse. Other advantages ensue from my definition not
presupposing compositionality. They will be discussed later in this
chapter.
The truth predicate (6.13) does not work in IF first-order lan-
guages. However, it is easily modified so as to do so. Consider for the
purpose an arithmetical IF first-order language. In such a language,
we can define a pair predicate in the usual way. Now it has been
shown by Krynicki (1993) that in an IF first-order language with
a pair predicate all independent quantifiers can be expressed in terms
of the familiar quantifier combination
(6.16)

known as the Henkin quantifier. Moreover, the elimination of other

independent quantifiers in favor of Henkin quantifiers can be taken
to be effective. Hence we can strengthen the normal form predicate
N(x,y) as follows: It will now say that y is the Gödel number
of a normal form of the sentence with the Gödel number x
where S2 satisfies the following requirements:
(i) S2 is logically equivalent with S1.
(ii) is obtained from S1 by a recursive procedure.
(iii) S2 is in the negation normal form.
(iv) In S2. each slashed quantifier occurs in a Henkin quantifier
(6.16).
(v) All slashed disjunctions are eliminated from
After this redefinition of N(x, y), clause (f) in the definition of
TR(X) can read as it did before. The main novelty that is needed is an
118 THE PRINCIPLES OF MATHEMATICS REVISITED

additional clause for Henkin quantifier formulas. It can be for-

mulated as follows:
= r(Yy)(Vu)(3z/Vu)(Jt/\/y)S[y, z, u, t]1 & X(x))
= 1S[y, f(y), u, g(u)]1 & X(x)) (*)

A moment's thought shows that (*) does indeed yield the right
truth-condition for Henkin quantifier sentences. With the changes
indicated, (6.13) serves as a truth predicate for the arithmetical IF
first-order language in question.
The main difference is that (6.13) does not any longer contain only
one initial second-order quantifier Second-order quantifiers
can also be introduced by (g), as shown by the quantifiers and
(ag) there. But when these quantifiers are moved to an initial (prenex)
position in (6.13), they still remain existential. Hence (6.13) as a whole
is of a form. Hence (6.13) can be translated back into the original
IF first-order language. All told, it follows that the truth predicate for
the given model of an IF first-order language is expressible in the very
same IF first -order language.
This truth predicate ((6.13) modified) does the job it was supposed
to do. It provides a truth definition for a suitable IF first-order
language in that language itself. Of course, the term "definition" has
to be dealt with gingerly here. What has been shown is that there is
a complex predicate (of the Gödel numbers of the sentences of the
given IF first-order language) which applies to a number if and only
if it is the Gödel number of a true sentence. Nothing is said of falsity
here, and no explicit formal definitions are set up.
Like the earlier truth definition for ordinary first-order languages,
the new one can be generalized from arithmetical languages to a
large class of other IF first-order theories. This class includes all
first-order theories which contain elementary arithmetic and also
contain a pair predicate for all their individuals. In view of Krynicki's
(1993) result, the latter assumption again implies that all slashed
expressions occur in Henkin quantifiers.
Hence the only novelty needed here (as compared to a corre-
sponding definition for ordinary first-order theories) is a clause in
the characterization of satisfaction which takes care of Henkin
quantifier formulas, in rough analogy to the clause (*) given above.
This is a straightforward matter. All we need to do is to say that
a valuation v satisfies (6.6) if and only if there are functions f(x) and
 TRUTH-DEFINITIONS FOR IF LANGUAGES 119

g(z) such that for every x and z, S[x,y,z,u] is satisfied by the

valuation which is like v except that the value of y is f(x) and the
value of u is g(z). This turns my satisfaction predicate Sat(x, v) itself
into form. But this fact does not mean that the truth predicate
(6.15) is not also in form.
The reliance of my truth predicate (6.13) on existential function
quantifiers brings the truth definition presented here even closer to
the explanations given in Chapter 2 as to what truth means in GTS.
In fact, the most remarkable thing about the truth predicate (6.13) is
its close relationship to semantical games, especially to the initial
verifier's strategies in these games. Indeed the entire definition is little
more than an explicit formulation of the idea that the truth of
a sentence S means the existence of a winning strategy for the initial
verifier in the corresponding semantical game. The clauses for
propositional connectives say in effect that the verifier must always
be able to make a move connected with them in accordance with
a winning strategy, in order for the propositional compound to be
true. And the clause for quantifiers likewise says that the verifier of
a true sentence must be able to choose in quantifier moves in
accordance with a winning strategy if the quantified sentence is to be
true. Because of the failure of compositionality in IF languages, this
requirement must sometimes be made for several quantifiers at the
same time (as in the Henkin quantifier), and not for each quantifier
one by one.
It can in fact be shown that the sentence in which the truth
predicate (6.13) is applied to the Gödel number of some particular
sentence S is equivalent to the second-order sentence which states
the game-theoretical truth condition of S.
Similar things can be said of the more general truth predicate
(6.15), even though the detour via the notion of satisfaction tends to
obscure them somewhat. It is for the purpose of avoiding this un-
necessary partial obfuscation that I once again formulate my truth
predicate first for IF first-order arithmetical languages, where the
nature of my truth predicate can be appreciated most easily.
Since semantical games are essentially activities of verification
and falsification, what all this means is that my truth predicate ties
the concept of truth essentially to the activities by means of which the
sentences of the relevant language are verified and falsified. The fact
that these activities are not what in everyday life (and in everyday
science) is most frequently meant by the verification, falsification,
120 THE PRINCIPLES OF MATHEMATICS REVISITED

confirmation, or disconfirmation of propositions does not in-

validate my point. The reasons for the basic position of semantical
games in the constitution (to use the Husserlian term) of the concept
of truth were explained in Chapter 2. With the qualifications ex-
pressed there, my truth definition can literally be taken to say that
a sentence is true if and only if it can in principle be verified, that is, if
and only if there exists a winning strategy for the initial verifier.
Admittedly, the remarks just made relate directly to the second-
order (sf) formulation of my truth predicate. But the IF first-order
truth definition is simply a translation of this predicate into the
corresponding IF first-order language.
It follows that my truth predicate refutes an important (and
fashionable) objection to explicit formal truth definitions. This
objection is the main reason why Tarski-type truth definitions are
these days routinely claimed to be philosophically irrelevant. It is
sometimes put by saying that Tarski-type truth definitions merely
characterize a certain abstract relationship between sentences and
reality, but that they do not provide any reasons why this relation-
ship should be taken to be the notion of truth. In another formula-
tion, it has been alleged that what is wrong with Tarski-type truth
definitions is that they are completely unrelated to the activites by
means of which we actually establish the truth (or falsity) of different
sentences. A Wittgensteinian might continue this line of thought and
aver that without such a link between truth definitions and the
activities of verification and falsification, it does not make sense to
speak of teaching, learning, understanding or mastering the concept
of truth so defined. As one recent writer has put it, Tarski-type truth
definitions do not show "what it is in virtue of which a sentence has
any truth-condition at all or in virtue of which it has the particular
truth-condition it happens to have" (Cummins 1989, p. 11).
Whether or not these criticisms are in fact applicable to Tarski-
type truth definitions need not be discussed here. The crucial thing is
that the truth definition and the truth-conditions explained in this
work answer precisely those questions which Tarski-type definitions
have been claiming to be incapable of answering. For instance, why
does a quantificational sentence S have the truth-condition it in fact
has? Now what does the truth-condition say? It says that there exists
a winning strategy for the initial verifier in the corresponding game
G(S). But how is the game G(S) determined, besides of course the
model (world) M with respect to which it is played? Its several moves
 TRUTH-DEFINITIONS FOR IF LANGUAGES 121

and their order are determined by the syntactical structure of S. In

this way, S has the truth-condition it has in virtue of its syntactical
structure, but only indirectly, because of the way this structure
determines the structure of the verification "game" G(S).
In Chapter 10 I will show how the game-theoretical truth defini-
tion and game-theoretical truth-conditions open the door for further
pertinent questions concerning the strategies that are their pivotal
element; to wit, whether those strategies may perhaps be subject to
some specifiable limitations.
More generally speaking, my truth definition does precisely what
Tarski-type truth definitions were accused of not doing. They relate
the notion of truth directly to the activities (semantical games) by
means of which we in a sense verify and falsify sentences. Even the
qualifications which are implicit in the phrase "in a sense" and which
were explained in Chapter 2, do not invalidate my point.
Indeed, we have a near-paradox on our hands. Emphasis on an
inseparable connection between truth and the activities of demon-
strating truth is a characteristic of the constructivistic approach to
the foundations of logic and mathematics. Hence we seem to have
taken sides with the constructivists. Yet the outcome is an extension
of classical logic capable of serving classical mathematics. This
near-paradox occasions a fresh look at constructivistic ideas which
will be undertaken in Chapter 10.
The independence of my truth definition, and indeed of the entire
GTS, of the assumption of compositionality is not merely a philo-
sophical or architectonic virtue. It extends the range of applicability
of truth definitions. It was explained earlier why their commitment
to compositionality prevents Tarski-type truth definitions from
being applicable to IF first-order languages. This fact has been
tacitly acknowledged by logicians in that special case of IF first-
order languages, that is, languages with partially ordered quantifier
prefixes, which they have studied since 1960 or so. In fact, all truth-
conditions that have been offered for them are in explicitly game-
theoretical terms, albeit formulated independently of my GTS (see
Henkin 1961 and cf. e.g., Barwise 1979, Enderton 1970 and Walkoe
1970).
But this is not the only advantage offered by GTS and by my truth
predicate. Since Tarski-type truth definitions toe the line of composi-
tionality, they must operate from inside out in a given sentence.
Hence they need starting-points for the definition in the form of
122 THE PRINCIPLES OF MATHEMATICS REVISITED

unanalyzable atomic sentences. In contrast, game-theoretical truth

definitions of the kind considered in this work operate from outside
in. The game starts from the given sentence, and in each move it is
replaced by a simpler sentence as an input for the next move. Now, as
every game theorist knows, there is in principle no obstacle to
defining winning and losing strategies, and so forth, also for infinitely
deep languages. Hence GTS and my truth definition can in principle
be applied to such Leibnitian languages as allow for infinitely deep
formal expressions. This possibility has been put into practice by the
group of logicians in Helsinki led by Jouko Väänänen and Juha
Oikkonen who have systematically studied such infinitely deep
languages. Needless to say, all their truth-conditions have been game-
theoretical, as are the truth definitions used in the special case of
infinitely deep languages known as game quantifier languages,
whose study antedates the more general study of infinitely deep
languages.
Thus working logicians have repeatedly found Tarski-type ap-
proaches inapplicable in different circumstances, and invariably
they have spontaneously resorted instead to game-theoretical con-
ceptualizations.
The game-theoretical approach to truth and truth predicates has
also distinct advantages over the approaches which operate with
sequences of partial truth predicates and which were initiated by
Kripke (1975). In an appendix to this book, Gabriel Sandu presents
a detailed analysis of some of these advantages.
Furthermore, a question that may have bothered my readers
receives a simple answer through a recognition of the basic role of
the strategic viewpoint. The truth definition presented here is as clear
as you can ever hope to get when presented in its second-order form.
Yet its translation into IF first-order logic results in a complex and
clumsy sentence which I have not even dared to try to write down
explicitly. Similar things can be said of the truth-conditions of
particular first-order sentences. Why this intuitive preference for the
second-order formulation?
An answer is implicit in the original preformal game-theoretical
characterization of truth. This characterization is in terms of the
existence of winning strategies for the initial verifier. These strategies
are expressed by (finite sets of) functions. Hence the most direct
("intuitive") formulation of game-theoretical truth-conditions is in
second-order (Ei) terms, and one's reasoning about first-order truth
 TRUTH-DEFINITIONS FOR IF LANGUAGES 123

is clear and iibersichtlich only as long as it is likewise conducted on

the second-order level.
However, this does not make any substantial difference to the
systematic situation. A language can in principle be thought of as
a notational shorthand alternative for the corresponding IF first-
order language. At worst I can defend myself by appealing to
something like Aristotle's distinction between what is primary for us
humans and what is primary in the order of nature or in any case,
in the order of logic.
At the same time, we can see that there is a solid systematic reason
for developing truth-theory in practice on the second-order level.
For it is on that level that we have most immediate access to
strategies for the players in semantical games.
To return to the starting point of this chapter, Tarski's approach
to truth definitions, we have seen that Tarski-type truth definitions
can, and must, be replaced by much more flexible definitions as do
not presuppose compositionality (cf. item (i) in the initial list).
As to (ii), the usual detour via the notion of satisfaction in
formulating the truth predicate is needed only because in most
nonarithmetical languages there is no name for each member of the
domain of individuals of the model in which truth is being defined.
As to the crucial claim (iii), it has been seen that the assumptions
of Tarski's impossibility theorem are not satisfied by IF first-order
languages and that it is in fact possible to define a truth predicate for
a suitable IF first-order language in that language itself.
In Chapter 7 it will be seen that the T-schema is not appropriate
as an adequacy criterion for truth definitions in general.
This leaves open only question (iv), that is, the definability of truth
in ordinary ("colloquial") language. At first sight, I seem to be
vulnerable to the same objection as has been repeatedly leveled at
Tarski. From the fact that truth is definable for certain formalized
(but interpreted) languages in the language itself it does not follow
that truth is likewise definable for our actual working language.
However, it is interesting to see that the evidential situation concern-
ing the definability of truth for natural languages is radically differ-
ent vis-â-vis my truth definitions from what it was for Tarski-type
ones. As was pointed out, Tarski's doubts concerning the definability
of truth for our "colloquial language" were probably prompted by
the failure of compositionality in natural languages. In my treat-
ment, such a failure becomes a reason for, instead of a reason against,
124 THE PRINCIPLES OF MATHEMATICS REVISITED

expecting that truth is definable for natural languages. For dispensing

with compositionality was precisely what enabled us to extend the
scope of our truth definitions beyond the purview of Tarski-type ones.
Again, I argued ages ago (Hintikka 1974) that informationally
independent quantifiers occur in natural languages. Some of my
examples raised eyebrows initially, but the general upshot of my
arguments is by now generally accepted. What has not been equally
generally acknowledged yet is that different variations of informa-
tional independence play a ubiquitous role in the semantics of
natural languages. They turn out to be at the bottom of such diverse
semantical phenomena as the Sprachiogik of our epistemic vocabu-
lary, of the de dicto versus de re distinction, and of the so-called
nonstandard quantifiers. In the next chapter, it will be argued that
negation behaves in natural languages very much in the same way as
it does in IF first-order logic.
Thus the prima facie evidence for or against the applicability of
my type of truth definition to natural languages is the reverse of what
it is for Tarski-type truth definitions. It suggests strongly a positive
answer for my definition and a negative one for Tarski's definition.
One can say more than this, however. In a sense, my truth
definition for IF first-order languages provides a paradigmatic
example of a truth definition for a language with a relatively poor
syntax, which through its very form shows how it can be extended to
syntactically richer languages or language fragments. In order to see
this, let us have a look at the truth predicate itself in its second-order
form. Its initial clauses tell essentially that for atomic sentences we
are adopting a redundancy (disquotational) treatment of truth. The
other clauses specify the conditions of the existence of a winning
strategy for the initial verifier for sentences containing quantifiers
and propositional connectives. Hence the upshot of the truth predi-
cate seems to be merely to spell out how the truth of an IF first-order
sentence depends on the quantifiers and propositional connectives it
contains. Hence it might seem that my truth predicate can be defined
only for quantificational languages, IF or not, and it apparently has
little relevance to other kinds of languages. The treatment of quanti-
fiers can be extended to natural languages without any problems in
principle, even though the actual logical behavior of ordinary lan-
guage quantifiers is quite different from that of the formalized
quantifiers of first-order languages. But at first sight it is hard to see
how my treatment could be extended to richer languages.
 TRUTH-DEFINITIONS FOR IF LANGUAGES 125

This objection is inconclusive, however. Assume, for the sake of

the argument, that we enrich a first-order language further by
introducing fresh logical constants. Assume, likewise for the sake of
the argument, that these new constants admit of a game-theoretical
treatment along the same lines as quantifiers and propositional
connectives; that is to say, by means of an affective game rule which
syntactically speaking simplifies the sentences that the players are
considering. Assuming that an application of a game rule by the
verifier can be thought of as an attempted step toward verification,
then the definition simply spells out the contribution of different
logical and nonlogical constants to the truth-conditions of the
language in question. Hence it should not be any surprise, and much
less an objection, that for languages which have been developed for
the explicit purpose of studying the semantics of quantifiers and
connectives, the first-order languages, only the contribution of
quantifiers and propositional connectives is spelled out in so many
words. What my truth definition provides is therefore a paradigm
case of truth definitions, an example whose purview is automatically
extended when other logical and nonlogical constants are brought
within the scope of semantical analysis. The crucial idea of truth as
the existence of a winning strategy remains completely unmodified.
What such extensions require is the possibility of extending the
game-theoretical treatment to logical and nonlogical constants
other than quantifiers and propositional connectives. There are
enough examples offered by game-theoretical analyses of different
natural-language expressions to suggest strongly that such a treat-
ment is possible over a wide front of semantical phenomena. Such
analyses as can already be found in literature include anaphora,
genitives, prepositional phrases, questions and answers, epistemic
verbs, and so forth (cf. here Hintikka and Kulas 1983 and 1985).
There is another, independent line of argument for the general
significance of the truth definitions explained in this chapter. A par-
tial answer can be formulated in terms of the power and central role
of IF first-order languages. In twentieth-century philosophy and
logic, ordinary first-order logic has often been taken to be the basic
logic. Sometimes it has even been claimed that it is the logic of
ordinary language. Frege thought that his is a univer-
sal "conceptual notation" or mathesis universalis in Leibniz's sense.
Frege, and most of his followers, have presented first-order lan-
guages as resulting from a minor regimentation of ordinary lan-
126 THE PRINCIPLES OF MATHEMATICS REVISITED

guage, involving mostly the elimination of the unfortunate ambi-

guities and other imperfections which beset the unpurified language
of our tribe. Admittedly, Frege's logic is not a pure first-order logic,
but it contains higher-order ingredients. But since he can be shown
to have assumed a nonstandard interpretation of the higher-order
part of his logic, he might as well have construed his mathesis
universalis as a many-sorted first-order language (see Hintikka and
Sandu 1991). Moreover, it will in any case be seen in Chapter 9 that
one can in effect do everything in the IF first-order logic that can be
done in a higher-order logic.
Among other things, it is in a sense assumed widely these days that
all of mathematics can be done by means of ordinary first-order logic
for it is widely believed that all mathematics can be done in terms of
an axiomatic set theory. Moreover, this axiomatization is typically
taken to be a first-order one. This means that the only logic it
involves is the usual first-order logic. The other elements, including
the axioms, are according to this view specifically mathematical
rather than logical, thus leaving first-order logic as the only logic of
mathematics.
Moreover, Frege's (and his followers') claim that ordinary first-
order logic is our true Sprachiogik seems to be supported by
Chomskian linguistics. Chomsky's recent version of the old idea of
logical form, teasingly called just LF, is essentially like a formula of
first-order logic (see here, e.g., Chomsky 1986 and cf. Hintikka 1989).
The few steps beyond ordinary first-order logic that Chomsky or his
followers have taken go unsurprisingly in the direction of IF first-
order logic.
Even if these arguments have an aura of ad hominem around them,
others are readily forthcoming. In Chapter 9 it will be seen that in
a perfectly good sense most of ordinary mathematics can be done in
IF first-order languages. I have shown earlier that informationally
independent quantifiers occur in natural languages. In the next
chapter, it will be seen that negation behaves in natural languages
very much like the negation (or negations) of IF first-order language.
Even though such arguments are suggestive rather than binding,
they do show at the very least the very general interest of my truth
predicate.
Philosophically the most important upshot of the results of this
chapter is undoubtedly that they disprove, once and for all, the myth
that the notion of truth in a sufficiently strong language is inexpres-
 TRUTH-DEFINITIONS FOR IF LANGUAGES 127

sible in that language itself. Applied to our actual working language,

this myth entails the companion myth of the ineffability of truth in
the ordinary sense of the word. These myths are ready to be
discarded. Not only has it been shown in this work that such
a self-applicable notion of truth is possible in some recondite formal
languages the construction presented above can serve as a para-
digm case for the definition of a truth predicate for any language
whose logic contains IF first-order logic. And this logic is, as I have
argued, nothing but our most basic ground-floor elementary logic.
Hence the strong suggestion of what has been seen here is virtually
the contrary of the ineffability myth. As soon as you understand your
own language and its logic, you have all that it takes to under-
stand and even to define the concept of truth, or so the suggestion
goes.
This general point can be illustrated with reference to a more
specific philosophical problem. It may be called the two hats prob-
lem. These two hats are apparently worn at the same time by my
semantical games for first-order logic. The language games of seek-
ing and finding are in the first place games for quantifiers and
propositional connectives. The player's moves in them are governed
by the rules that characterize the meaning of quantifiers and con-
nectives. Semantical games of the kind I have described are the
logical home of quantifiers and connectives.
But if so, how can these very same games also serve to give an
altogether different kind of concept its meaning, namely, the concept
of truth — at least the notion of truth as applied to first-order
languages. How can one and the same language game serve to lend
a meaning to two different kinds of concepts, one of which (the
concept of truth) seems to be a metalogical one? This two hats pro-
blem might also be called Wittgenstein's problem (cf. here Hintikka
and Hintikka 1986, Ch. 1). For Wittgenstein insisted that you cannot
speak meaningfully and nontrivially of the truth of the sentences of
a language in that language itself. Or, since for Wittgenstein there is
ultimately only one language ("the only language that I under-
stand"), we cannot speak of truth nontrivially, period. What looks
like a metalogical discourse pertaining to the truth and falsity of
a fragment of language is for Wittgenstein merely a different "calcu-
lus," a different language based on a different language game. How,
then, can the meaning of first-order logical constant and the notion
of truth as applied to first-order languages be constituted by the
128 THE PRINCIPLES OF MATHEMATICS REVISITED

same language games? Doesn't speaking of truth take us ipso facto to

a metatheoretical level?
A clue to an answer to this question is the close connection there is
between the notions of truth and meaning. What a sentence says is
that its truth-conditions are satisfied. Hence it is impossible to try to
detach the notion of truth as applied to a quantificational sentence to
the language games that lend quantifiers their meaning.
This answer can be elaborated further. For the purpose, it is
handy to recall the distinction between definitory rules of a game
from the strategic rules or principles governing it. The former specify
which moves are permitted and which ones are not. The latter, the
strategic rules, specify which strategies (and a fortiori which rules)
are better or worse than others. The several definitory rules of
semantical games characterize the meaning of the different quanti-
fiers and propositional connectives. In contrast, we have seen that
the notion of truth for quantificational sentences is to be character-
ized in terms of the sets of strategies that are open to the initial
verifier to use. Thus the step from the meanings of quantifiers to the
truth of quantificational sentences is not a step to a metalogical
level. It is, rather, a step from considerations pertaining to the
particular moves in a semantical game to concepts pertaining to
the strategies of the initial verifier in those games.
The reason why it is impossible to detach the meaning of quanti-
fiers from the notion of truth is that some degree of the mastery of the
strategies of a game is unavoidable for the understanding of a game.
If you only know how chessmen are moved on the board you cannot
as much as say that you know how to play chess, unless you have
some idea of the better and the worse moves and sequences of moves.
No one would deign to play chess with you. This illustrates the fact
that some degree of understanding of the strategies of a game is an
integral part of the conceptual mastery of that game.
But if so, we then have here an example of the general insight
mentioned above. Understanding a quantificational language
means mastering the language games that give quantifiers their
meaning. Now, such mastery was just seen to involve a grasp of the
strategies available to the players of a semantical game. But those
strategies are just what is needed to understand the concept of truth
as applied to quantificational languages.
One of the main features of the truth predicate presented here is
that it is in first-order terms. As was pointed out previously, Tarski-
 TRUTH-DEFINITIONS FOR IF LANGUAGES 129

type truth definitions are of a second-order form. The reason why

this makes a crucial difference to the foundations of logic and mathe-
matics was explained in Chapter 1. It was explained there in what
sense all model theory depends on truth definitions. As long as these
definitions can only be given on second-order level, then model theory
depends on second-order logic. And even if second-order logic is not
merely "set theory in sheep's clothing", it still admittedly involves
many of the same problems of set existence that makes set theory not
only a difficult but positively frustrating exercise.
Against this backdrop, one can see the significance of my truth
predicate. It is nothing more and nothing less than the declaration of
independence of model theory. It shows that one can develop a model
theoryfor the powerful IF first-order languages on the first -order level,
ergo independently of all questions of sets and set existence. All the
quantifiers in the IF first-order version of my truth predicate range
over individuals. If Quine is right (as I think he is, if rightly
understood) and to be is to be a value of a quantified variable, then
our truth definition (and indeed the entire IF first-order logic) does
not take up the question of the existence of higher-order entities like
sets at all. Consequently, all apprehensions concerning the purely
logical status of model theory are groundless. Model theory of
first-order logic is part of logic, and not a proper part of mathema-
tics. The problems which are caused by the apparent dependence of
the model theory of first-order logic on set theory (or on higher-
order logic) can thus be solved, and Tarski's curse be exorcised. This
makes a substantial difference to the foundations of mathematics. If
you return to Chapter 5 and review what was found there, you can
see in how many different ways freedom from Tarski's curse enhan-
ces the self-sufficiency of first-order logic as contrasted to set theory.
Of course, my saying this does not mean that I advocate doing
model theory in everyday practice in IF first-order languages. That
would be impracticable. What is at issue is the conceptual nature of
model theory and not the heuristic framework best suited for the
actual nitty-gritty work in model theory.
The conceptual independence of model theory from set theory is
only one aspect of a more comprehensive reevaluation that is needed
in our ideas about the foundations of logic and mathematics. This
reevaluation will be continue :1 in Chapter 8.
The philosophical and suggestions of the definabil-
ity of a truth predicate in suitable IF first-order languages are too
130 THE PRINCIPLES OF MATHEMATICS REVISITED

sweeping to be exhausted here. Let me mention only a couple of new

tacit perspectives. Suppose that it could be shown that our ordinary
language (or some significant fragment thereof) is inevitably like IF
first-order languages as far as the notion of truth is concerned. This is
suggested by what has been found in this chapter, and it will be
reinforced by what is found in Chapter 7. Then the definability of
truth would be equally inevitable, and hence it could be argued that
merely by using our own language we are inextricably committed to
a realistic concept of truth, namely, the one definable along the lines
I have followed. This would be an important conclusion indeed.
Even though it is beside the main purpose of this book to examine
such issues here at any greater length, I can point out that this line of
thought promises to support very strongly what was said at the end
of Chapter 2 of the independence of our normal notion of truth from
the epistemic institutions of verifying and falsifying propositions.
Truth is in a deep sense a logical rather than an epistemological
notion.
To put the same point in other terms, anyone who is using
a language with a minimally rich expressive power (logically speak-
ing) is committed to a concept of truth which in the current crude
classification would be called a correspondence view of truth. This
would be the case even if that language were only used for the
purpose of Rortian dialogues for dialogue's sake. Any sufficiently
strong language would be a mirror of the world, Rorty notwith-
standing.

Notes
The only book-length discussion of Tarski's truth-definition seems to be Moreno
(1992). It supplies further references to the literature.
2
See Walkoe (1970).
For infinitely deep languages, see the bibliography in the introduction to Tuuri
(1990). For game quantifier languages, see Ph. G. Kolaitis, "Game Quantifica-
tion", Chapter 10 in Barwise and Feferman (1985).
 8

Axiomatic Set Theory:

Fraenkelstein's Monster?

The backbone of the approach to the foundations of mathematics

developed in this book is game-theoretical semantics and the new
logic it has inspired, the independence-friendly logic. In the preced-
ing chapters, GTS has opened new perspectives in several directions.
IF logic was seen to put the whole array of questions concerning
completeness and incompleteness in logic as well as in mathematics
in a new light. It frees the model theory of the fetters of set theory by
facilitating the formulation of a truth predicate for a suitable IF
first-order language in the language itself. Furthermore, in IF first-
order languages the important idea of negation turns out to behave
in an unexpected way, thereby suggesting that the same might be
true of natural languages. But even after all these results many
important questions still remain unanswered, including compara-
tive ones. IF logic may have some advantages as a framework of
mathematical theorizing, but what has been seen scarcely seems to
affect the virtual monopoly of set theory in that department. Admit-
tedly, there is no longer any need to resort to set theory or even
to higher-order logic for the purpose of doing model theory of
first-order logic. But this relative self-sufficiency of IF first-order
languages does not prove their superiority over set-theoretical lan-
guages. On the contrary, set theory is still the medium of choice for
many mathematicians, philosophers and logicians for model-
theoretical theorizing. In this chapter, it will be argued that this
privileged role of axiomatic first-order set theory becomes extremely
dubious in the light of the insights that the game-theoretical
approach has yielded, and will yield.
163
164 THE PRINCIPLES OF MATHEMATICS REVISITED

It seems to me that history is very much on my side. This role of set

theory as universal mathematics goes back to the founding fathers of
set theory. Cantor thought that any structure could in principle be
dealt with in set theory (see Dauben 1979, pp. 229—230). This was not
his private theory, either. Set theory was thought of as providing
models for all possible mathematical theories. One context in which
there is a need for different kinds of models is created by various
consistency proofs. Beltrami and Felix Klein had impressed the
mathematical world by presenting a consistency proof for an axiom-
atic non-Euclidean plane geometry relative to the Euclidean one.
They had constructed a model for a nonEuclidean geometry within
a model of the Euclidean one (cf. Torretti 1978, pp. 115—142). But
other varieties of non-Eucidean geometry did not admit such models.
Where, then, can we find the models that serve to show the consistency
of such geometry? In set theory, it was thought. This is the reason why
Cantorian set theory seemed a veritable paradise to David Hubert
(1925, p. 170) who was looking for consistency proofs for all and
sundry mathematical theories. For Aristotle, the hand was the tool of
all tools and the soul the form of all forms. In a similar manner, for the
creators of set theory this theory was calculated to give us the model of
all models. Indeed, in Hubert's thinking, besides the universe of dis-
course of some one mathematical theory which we might be consider-
ing, we also have to keep an eye on a more extensive domain, namely,
the universe of all possible structures. When Hubert sometimes seemed
to infer existence from consistency, incidentally upsetting poor Frege no
end, he meant existence in the super-universe of all structures.1
In the course of the subsequent development of axiomatic set
theory, this idea of the universality of the universe of set theory was
nevertheless lost. The genesis of axiomatic set theory was in effect an
attempt to make sure, not that the universe of set theory contains all
possible structures but only that it contains all structures represent-
able in the language of set theory. In the simplest case, this means
trying to capture in the universe of one's set theory all sets (that is, all
subsets of the domain of individuals) definable in the set theory in
question rather than all sets, period. This is illustrated by the role of
the comprehension axiom or comprehension schema in the axiom-
atic version of set theory. This axiom schema can be expressed in its
unrestricted form as

(8.1)
 AXIOMATIC SET THEORY 165

where S[x] is any well-formed formula in the language of set theory

with x as its only free variable. The paradoxes forced set theorists to
give up the unrestricted form of(8.1). It can nevertheless be said that
set theorists have tried to capture as many sets as possible by means
of.(suitably restricted forms of) the comprehension schema (8.1).
They have tried to impose as few restrictions on it as is deemed safe.
The scheme (8.1) has thus been their main tool in investigating the
difficult problems of set existence. In typical axiom systems of set
theory, all axioms of set existence are special cases of(8.1), with one
exception. This exception is the axiom of choice, which has been
argued (and will be argued) in this book to be unproblematic. Yet,
even if we could look away from the restrictions that are needed in
(8.1), we cannot capture by its means all the sets we are considering —
that is, all subsets of the domain of individuals of the intended model
of axiomatic set theory. Thus the paradoxes of set theory which have
necessitated restrictions in the comprehension axiom are only
a part and the easier part — of the problem of constructing an
adequate set theory. The more difficult part is to capture the sets
not representable by explicit formulas of one's set-theoretical lan-
guage.
There is a sense in which the use of the comprehension axiom (8.1)
is based on a fallacy, or at least on dated assumptions. The richness
that the universe of set theory was originally thought of as having
was model-theoretical in nature, as was seen earlier. In contrast, the
comprehension axiom (8.1) is an axiomatic and syntactic device to
enforce as much richness in the universe of one's set theory as
possible. It is perhaps understandable that Zermelo should have
thought in 1908 that he could enforce model-theoretical plenitude by
axiomatic means. But no one who has internalized Gödel's incom-
pleteness results can reasonably hope to do so any longer.
This matter can be looked at from another point of view. The
contrast we have found between the original Cantorian "naive" set
theory and axiomatic set theory is clearly related to the distinction
between the standard and nonstandard interpretation of higher-
order logic. This distinction was first given these labels by Henkin
(1950) even though some earlier thinkers, especially Frank Ramsey
(1925), had been aware of it and though it had played an important
role in earlier foundational discussions (see Hintikka 1995a). It
affects all higher-order quantifiers, but it can be explained by
reference to a single second-order one-place predicate variable (set
166 THE PRINCIPLES OF MATHEMATICS REVISITED

variable), say X. What are the (extensions of the) values of such

a variable? The standard interpretation says that they are all the
subsets of the domain of individuals do(M) of the model M in
question — that is, that the range of X is the power set of do(M). If,
instead of this interpretation, we restrict the values of X to some
designated subset of the power set of do(M), we obtain a nonstan-
dard interpretation of the variable X. The most common restriction
is to the sets definable to the theory in question.
A similar distinction can be applied to any higher-order variable.
It is usually required that the set of designated sets (of any logical
type) is closed with respect to Boolean operations and projective
(quantificational) ones. When I speak of the descriptive complete-
ness of axiomatic set theory, it is of course by reference to some kind
of standard interpretation. However, the analogy between the differ-
ent interpretations of set theory and the different interpretations (i.e.,
the standard vs. nonstandard contrast) of higher-order logic is not
perfect. The standard versus nonstandard distinction can be only
locally applied to models of axiomatic set theory. We can speak of
standardness locally, for instance, in connection with a given set, by
reference to the question as to whether all its (extensionally possible)
subsets are represented in some given model of one's axiom system
for set theory, or to the question whether the set of natural numbers
as reconstructed in axiomatic set theory contains only "standard"
natural numbers. In contrast, the entire set-theoretical universe is
inevitably nonstandard. Either the set of all sets or its power set fails
to exist, though both are extensionally possible and hence ought to
exist in a standard model.
What was pointed out is that an ordinary first-order axiomatic set
theory cannot guarantee full standardness. In this sense, axiomatic
set theory fails in its attempt to recover the Cantorian paradise.
Different versions of the axiom of comprehension inevitably fail to
do their intended job of capturing all the sets that we would like to
include in the universe of set theory, not because they are not the
right formulations of the axiom, but because they are first-order
axioms. As a consequence, axiomatic set theory is inevitably incom-
plete descriptively. And since it uses first-order logic, which is
semantically complete, set theory is also deductively incomplete. (If
you have not thought of this incompleteness before, recall that you
can do elementary number theory by means of axiomatic set theory,
and then apply Gödel's incompleteness theorem.) The crucial ques-
 AXIOMATIC SET THEORY 167

tion is: How serious is this failure? A related question that arises here
is: Why bother with an unavoidably incomplete theory? The most
respectable prima facie reason that I can see to cultivate incomplete
axiomatic set theory is to think of it as an approximation to the real
thing, an approximation which can be made closer and closer by the
addition of new axioms. This way of thinking nevertheless involves
a serious mistake. To speak of approximations presupposes that
there is something to approximate. And the lack of a standard model
of axiomatic set theory means that there is apparently nothing to
approximate. We have no unproblematically clear idea of what the
intended models of set theory are. Hence it makes no sense to speak
of additional axioms that capture the structure of the intended
model of set theory more and more closely.
This point is worth spelling out more fully. A comparison between
set theory and normal mathematical theories may be instructive.
When it comes to many other theories we often have a good idea of
what the relevant models are — for instance, what the structures are
that are exhibited by natural numbers, by real numbers, by the
second number class, and so forth. We know what the different
structures are that we want to study in group theory, lattice theory or
topology. In contrast, we do not have a sharp idea of what the model
of set theory intended by Cantor and his contemporaries is like. This
uncertainty is partly a consequence of the paradoxes that arose in
early set theories. It is usually formulated by speaking of the
difficulty of deciding which axioms we should adopt in an axiomatic
set theory. This diagnosis nevertheless deals with the symptoms of
the problem rather than with their cause. There cannot exist a com-
plete first-order axiom system for set theory, anyway. Hence the
focus of attention should be on the principles of looking for new
axioms rather than thejustification of any particular candidates. For
instance, the idea that set-theoretical axioms can be compared with
each other in terms of the "naturalness" of their consequences is
theoretically very shallow. How do the people who advocate this
idea know that naturalness is not in the eye of the set theorist, and
insensitive to the model-theoretical realities of the situation? Instead
of relying on such intuitions, we must realize that to have some idea
of what axioms are supposed to hold in set theory is in effect to know
something about what the intended models of set theory are like.
And such knowledge is precisely what philosophers do not have, at
least not with sufficient clarity.
168 THE PRINCIPLES OF MATHEMATICS REVISITED

One reply that I will inevitably receive is that the cumulative

hierarchy of sets provides the intended model I claim to be missing
from the foundations of set theory (for able expositions of this
conception of set, see Boolos 1971 and Parsons 1977). According to
this idea, the universe of set theory is built up by a step-by-step (but
infinite) process. At each stage, the new sets that are introduced are
all possible collections of sets already reached.
The rejoinder to this objection is that the cumulative hierarchy is
not by any stretch of the imagination what the architects of set
theory had in mind. Their idea of the set-theoretical universe is that it
is the richest possible one. Whatever can exist does exist in the rich
and plentiful world of set theory. Like the Lovejoyful god who is so
generous that he grants the gift of existence to everything that can
possibly exist, a Cantorian set theorist is expected to assume that all
possible sets are present in his universe of all sets (cf. Dauben 1979,
pp. 225—236; Lovejoy 1936).
Yet the cumulative hierarchy is ontologically stingy. It may be
compared to Gödel's constructive model of set theory (Gödel 1940)
even though it is not identical with Gödel's. And the main virtue of
Gödel's model is ontological economy. All sets existing in that model
are obtained by a narrowly defined construction process, which can
be described in set theory itself. The famous axiom of constructibility
says that this process exhausts the entire universe of sets. The
procedure for building up the constructive hierarchy is a more
generous one, but it, too, rules out a tremendous variety of potential
sets (and their structures). As Hallett (1984, Ch. 6, esp. p. 223) points
out very clearly, even though the defenders of the cumulative
hierarchy use expressions like "all possible sets", their force is not the
same as in ontologically more generous and more realistic ap-
proaches to set existence.

Boolos uses the same term "all possible collections" [as Wang and
Shoenfield] in his account of the intuitive hierarchy. If we are really forming
sets in stages, then it may be that at stage s we do form all possible sets of
objects formed at previous stages. But here "all possible" will not be very
many, certainly not many compared with the "all possible" of the later
"formal" (impredicative) stage of the theory (Hallett 1984).
In general, it is hard to reconcile an emphasis on predicativity or
even on any stage-by-stage generation of sets with the Cantorian
idea of set theory as the study of all possible structures. Among those
 AXIOMATIC SET THEORY 169

possible structures there are very large sets. The iterative (cumulat-
ive) conception of sets does not capture them automatically. They can
only be captured by special axioms of large cardinals (cf. Kanamori
and Magidor 1978). As Gödel emphasizes (1947, reprint p. 476),
beyond a certain point the further iteration of the "set of" operation
depends on assumptions of large cardinals. The need of such axioms
shows that the iterative conception of set is not self-explanatory and
certainly does not alone capture the notion of set that Cantor was
trying to implement. Even if you do not actually need impredicative
sets for certain specific purposes in set theory or elsewhere in
mathematics, it is still hard to deny that there are possible structures
that can only be characterized impredicatively.
Even without much detailed argument, it is thus clear that the
iterative model is not what the founding fathers of set theory had in
mind or that such a model of set theory is especially well-suited for
the purpose of serving as a framework of mathematics, including the
search for new axioms. Such a search is best given a firm direction by
some sufficiently explicit characterization of the intended model,
which is just what has not been provided by set theorists.
Other candidates for the role of the intended model of set theory
will be examined later and found wanting.
In discussing the foundations of set theory, philosophers and
mathematicians often talk about the search for new axioms. But such
talk is misguided, for nobody knows what one is supposed to look
for. In order to illustrate the axiomatic and hypothetical approach to
mathematical theorizing, Bertrand Russell once called a mathema-
tician a chap who never knows what he is talking about. In the same
spirit of friendly teasing, a set theorist would have to be called a guy
(or a gal) who never knows what he or she is looking for.
In brief, axiomatic set theory does not have a model theory which
would enable us to characterize its intended models. The deeper
reason for this need of a solid model-theoretical foundation of
whatever logic or theory that we are envisaging as the ultimate
framework of mathematical theories, is seen from the conclusions
reached in Chapter 5. There it was seen that the first and foremost
goal of mathematical theorizing is a descriptively complete theory,
and not a deductively complete one, as many philosophers would
like to have it. Now, the notion of descriptive completeness is a
model-theoretical one, presupposing both a truth definition and a
notion of intended model.
170 THE PRINCIPLES OF MATHEMATICS REVISITED

But the difficulties with axiomatic set theory cut deeper than our
ignorance (or indecision) of what its intended models are like. We
know that whatever the details of the standard models of set theory
are or may be, these models cannot be fully captured by an axiomatic
set theory formulated on the first-order level. First-order axiomatic
set theories are not only deductively incomplete as they stand; they
are inevitably incomplete descriptively, no matter what new axioms
are added to them. One way of seeing this is to go back to the basic
question: Why should we cultivate set theory? The reason why we
were first led to consider set theory in the line of thought presented in
this book was that set theory (or its equivalent) was apparently
needed for the model theory of our basic logic, that is, ordinary
first-order logic. This motivation for taking up set theory has been
widespread among philosophically oriented logicians, and probably
still is. Often it takes a more general form. One of the main jobs that
set theory is drafted to do is to enable us to deal with model-
theoretical questions in different parts of logic. And this conception
of the job that set theory can do is not restricted to logic. The so-
called structuralist philosophy of science is little more than an
application of this idea (or presumption) to the epistemology of
scientific knowledge.
This alleged rationale of set theory is not a genuine one. There is in
the literature a widespread tendency to give set theory the credit for
theories which have been developed by using set theory as their
metatheory. For instance, in Moore (1994, p. 635) we read:
Set theory influenced logic, both through its semantics, by expanding the
possible models of various theories and by the formal definition of a model;
and through its syntax, by allowing for logical languages in which formulas
can be infinite in length or in which the number of symbols is uncountable.

The tacit syllogism on which such claims rest is something like this:
logical languages which cannot be dealt with syntactically must be
dealt with model-theoretically. Now set theory is the natural me-
dium of model-theoretical conceptualizations. Hence it is by means
of set theory that theories like infinitary logics have been developed
and are being cultivated. The fallacy here is that nothing is said or
done to exclude other possible metalanguages and metatheories.
And what has been seen in this book makes this question a most
pertinent one. By showing in Chapter 6 how a truth predicate can be
defined for a suitable IF first-order language I removed the main
 AXIOMATIC SET THEORY 171

obstacle from the path of a model theory of IF first-order logic done

in IF first-order logic itself.
The table can in fact be turned on set theory here. Much more can
be said than merely to remove one of the many possible reasons for
cultivating set theory axiomatically, or in some other rigorous
fashion needed in a genuine model theory. First-order logic was
originally suspect because on apparently could do its model theory
in first-order logic itself. If this was a legitimate prima facie com-
plaint, then it is also fair game to ask whether the model theory of
axiomatic set theory can be done in axiomatic set theory itself. This
question leads us to consider the possibility of defining a truth
predicate for axiomatic set theory in that very same theory. For
without a truth predicate, there is little hope of developing anything
like a viable model theory. Furthermore, this question is made
especially pertinent by the reputed role of set theory as the medium
(or at least as a surrogate) of model theory.
It is in fact possible to bring the results reached in Chapter 6 to
bear on this task of doing model theory of axiomatic set theory in the
same theory. Indeed, axiomatic set theory is an ordinary first-order
theory with a single nonlogical predicate containing elementary
arithmetic. Hence a truth predicate can be formulated for it in the
fragment of the corresponding second-order language, that is, in the
form
(3f)Sat[x, f] (8.2)

But what (8.2) expresses in a sense should be expressible also in the

first-order set theory itself. Indeed (8.2) is equivalent to a statement of
the form
(3X)S[X] (8.3)

where X is a one-place second-order predicate variable. From (8.3) we

can obtain a related but not always logically equivalent (first-order) set-
theoretical statement by replacing every subformula of (8.3) of the form
X(y) (or X(a)) by (or aex) and by replacing (JX) simply by (3x). In
this sense, we can in fact formulate a truth predicate for first-order
axiomatic set theory in the same set theory itself.
But this observation needs to be put in perspective. At first sight,
it seems to violate Tarski's impossibility result according to which
the concept of truth in a first-order theory cannot be defined in that
theory itself.
172 THE PRINCIPLES OF MATHEMATICS REVISITED

A beginning of a solution to this puzzle lies in the fact that no truth

definition has been given for axiomatic set theory in that theory
itself— only a truth predicate. And from the mere formulation of a
complex predicate no contradiction can follow.
But in what sense, if any, does Tarski's result then rule out the
possibility of self-applicable truth predicates? In order to see what
this sense is, the reader is invited to go back to the argument at the
beginning of Chapter 7 which served to show that the truth pred-
icate (8.1) considered there leads to a contradiction. We implicitly
used there the assumption that the sentence T[n] constructed
by means of the diagonal lemma is logically equivalent to (or at
least has the same truth-value in the intended model as) the arith-
metical sentence that says that the sentence with its very own Gödel
number g( T[n]) = n is true. But T[n] says that the same
sentence is false; hence a contradiction which can be taken to have
the form
T[nJ+-'T[g(-- TEn])] (8.4)

where the right side is of course equivalent to T[n].

The equation of the truth-values of the two sentences T[n] and
T[n])] is a consequence of the adequacy conditions on any
candidate for the role of a truth predicate. The operative adequacy
condition is, to all practical purposes, Tarski's familiar T-schema
(see Chapter 6) applied to the game-theoretical truth predicate.
Both are incomplete, but neither leads to a contradiction.
This point can be generalized. A truth predicate has consequences
only in conjunction with an adequacy condition which is essentially
tantamount to Tarski's T-schema. This condition requires that any
sentence S is true if and only if the sentence T(g(S)) is true. The latter
sentence says of course that the truth predicate applies to the Gödel
number g(S) of S.
The truth predicate defined in Chapter 6 for suitable IF first-order
languages thus accomplished something that Tarski's impossibility
result at first sight discourages us from expecting. This something is
a truth predicate T[x] which applies to g(S) if and only if S is true
according to the game-theoretical conception of truth — which
I have argued is our natural conception of truth. The reason why we
cannot express this as a formal equivalence
Si—*T{g(S)] (8.5)
 AXIOMATIC SET THEORY 173

is that (8.5) says too much. It asserts also that S and T[g(S)] are false
simultaneously. This unwanted implication cannot be avoided in
(unextended) IF first-order languages. In such languages the bite of
a truth predicate thus lies in its satisfying the adequacy requirement
that motivates Tarski's T-schema. However, this requirement can-
not be formulated in unextended IF first-order languages as a formal
equivalence schema. In this way, new light is thrown both on
Tarski's intentions and on their implementation.
What does follow from all this when applied to axiomatic set
theory? The first and foremost point here is that, contrary views
notwithstanding (cf. McGee 1991, pp. 75—76), there is nothing wrong
as such with the truth predicate for axiomatic set theory which was
explained earlier in this chapter, and which is formulated in axio-
matic set theory itself. In fact, by means of this truth predicate one
can do a fair amount of model theory of set theory in set theory itself.
We can for instance define what it means for an element of the set-
theoretical universe to satisfy a given formula with one free variable.
We can then consider the totality of elements satisfying the for-
mula, and so on. This is the objective basis for the surprisingly
popular casting of set theory in the role of a do-it-yourself model
theory. The implicit idea undoubtedly is in a sense one of economy.
Who needs fancy model theory when there already exists a set theory
serving as the universal framework of mathematical, and even
scientific, theorizing and allowing us to do its model theory in the set
theory itself? In fact, much of the actual work in basic model theory
of axiomatic set theory can, in this spirit, be thought of as being done
by means of set theory itself. Even though such model theory of set
theory is ordinarily done informally, it is not hard to believe that
much of it can also be done in terms of a suitable axiomatic set
theory.
What such a self-applied theory for axiomatic set theory cannot
assume are all the instances of Tarski's T-schema, applied to the
game-theoretical truth predicate. This follows from Tarski's very
own impossibility result. Alternatively, and more directly, we can use
the same kind of constructive argument as is usually employed in
proving Gödel's first incompleteness theorem by means of the
diagonal lemma. What we need to do is to apply this lemma to the
negation of the quasi-predicate of truth that I showed above how to
formulate for axiomatic set theory in set theory itself. Let this
predicate be T[x] and let g(S) be the Gödel number of S. Then by ti
174 THE PRINCIPLES OF MATHEMATICS REVISITED

diagonal lemma there is a number n such that

=T[n] (8.6)

where n is the numeral representing n. Along the lines indicated

above one can see that T[n] is true but that the sentence
T[n])) = T[nJ (8.7)

which attributes the truth predicate to its Gödel number is false.

Thus the game-theoretical truth predicate I am considering can-
not do its job for all set-theoretical sentences. It will fail to yield the
right truth-condition for the liar-type sentence whose Gödel
number is the number n yielded by the diagonal lemma. In this sense,
the model theory in question is inevitably incomplete. This destroys
the main rationale of set theory, namely, its role as the allegedly
canonical framework for model theory. The failure of any self-
applied set-theoretical truth predicate means that set theory cannot
adequately serve as its own model theory, any more than ordinary
first-order logic can.
But incompleteness is the least of the woes of axiomatic set theory.
The failure of Tarski's T-schema for the sentence S (constructed by
means of the diagonal lemma) means that there inevitably exists in
any model of set theory a set-theoretical sentence which is true in the
ordinary sense of the word but whose truth-condition, when ex-
pressed set-theoretically, is false. Now the truth-condition of only
first-order sentence asserts the existence of the Skolem functions that
codify a winning strategy for the initial verifier in the semantical
game connected with it. Considered as a set-theoretical statement,
S can fail to be true only if the truth-making strategy functions
(Skolem functions) do not exist in a model of the set theory in
question, even though they are extensionally possible "functions in
extension", as Russell would have expressed the point. And such
a failure means that in axiomatic set theory, a sentence is not always
equivalent to the set-theoretical counterpart of its second-order
translation.
This observation has several different consequences illustrating
the inadequacy of axiomatic set theory. What just has been seen
means that intuitively speaking my neo-Gödelian sentence Sn
should be true, in that the Skolem functions which it asserts to
exist in set theory actually do exist in a purely extensional sense. And
axiomatic set theory was originally calculated to capture precisely
 AXIOMATIC SET THEORY 175

all extensionally possible sets — or at least as many of them

as we can.
This might not yet seem to set the Skolem functions asserted to
exist in S, apart from the nonexisting sets and functions which are
ruled out by the well-known paradoxes. But the present failure is far
more poignant than the traditional ones. What is so special about
the failure of the equivalence of Sn and its truth-condition T[n]? One
thing that is special about them is that equivalences between a first-
order sentence and its second-order truth-condition were seen ear-
lier in Chapter 2 to be nothing but instances of the axiom of choice or
of its generalizations. These generalizations can be justified in terms
of my game-theoretical approach in the same way as the axiom of
choice itself. In other words, in equivalences of sentences like my
with their truth-conditions we are merely dealing with instances of
the translations of generalizations of the axiom of choice into a set-
theoretical language. And yet some of them were seen to fail inevi-
tably. Hence the failure of these generalizations in axiomatiè set
theory shows that the axiom of choice should be really rejected by
axiomatic set theorists, for its motivation is, in effect, rejected by
them. To put the same point in other terms, the fact that the axiom of
choice is compatible with the usual axioms of set theory (see Gödel
1940) is merely a lucky accident having nothing to do with the
theoretical motivation of the axiom of choice. One can in fact say
that the spirit of axiomatic set theory is deeply antithetical to the
axiom of choice. Since this axiom was seen to have an impeccable
theoretical motivation, we have here a strong reason to be wary of
axiomatic set theory in general.
As a by-product of this line of thought, we have arrived at an
argument to show that my generalization of the axiom of choice is
actually stronger than the axiom itself. This verifies the surmise
mentioned in Chapter 2. For the axiom of choice, even in its global
form, holds at least some of the models of axiomatic set theory,
whereas some instance or other of its generalization has been seen to
fail in each model.
This point can be put in a different way. Gödel advocated the
cultivation of our "mathematical intuition" in such a manner that it
will hopefully show us the validity of new axioms of set theory. One
of the most firmly entrenched sets of intuitions we seem to have in set
theory are the ones that lead to the axiom of choice. I have shown
above in a Gödelian spirit that these same intuitions justify stronger
176 THE PRINCIPLES OF MATHEMATICS REVISITED

set-theoretical assumptions as well, namely, the set-theoretical counter-

parts to equivalences between a first-order set-theoretical sentence
and its second-order truth-condition. However, these stronger set-
theoretical assumptions turn out to contradict the received axioms
of set theory. This means that a most natural implementation of
Gödel's idea does not lead to an improved set theory, but to the
rejection of first-order axiomatic set theory.
Here the reader can perhaps begin to see the general reasons for
my suspicions of axiomatic set theory. It was originally constructed
to embody precisely the way of thinking represented among other
things by the axiom of choice. As it flexes its deductive muscle, it
clearly begins to turn menacingly against the intentions of its origi-
nators. It is turning into a veritable Frankenstein's (Fraenkelstein's)
monster which threatens to destroy what it was constructed to serve
and to defend.
Thus we have arrived at a serious indictment against axiomatic set
theory. For each model of axiomatic set theory, one can find a
sentence S in the language of a set theory which is true interpreta-
tionally but false in that model. In brief, axiomatic set theory cannot
be true on its own intended interpretation. When we try to construct
the liar paradox for axiomatic set theory, the liar turns out to be
axiomatic set theory itself.
A possible rejoinder to this indictment is to say: What else is
new? The old paradoxes already show that there are set-theoretical
propositions which we are intuitively tempted to accept but which
cannot be assumed to be true in the models of set theory. This
rejoinder does not cut much ice, however. In the old paradoxes, we
are typically dealing with the failure of an instance of the unrestricted
axiom of comprehension to be true. In other words, we are dealing
with an apparently well-formed concept whose extension is not
well-defined. For instance, there is no well-defined class which is an
extension of (xEx). In contrast the extensions of the functions
whose existence is asserted by the sentence S conducted above are
unproblematic. The only problem is that these classes cannot be
reified into sets which could exist in the models of the set theory in
question.
More generally speaking, the earlier paradoxes of set theory are
all caused by there being, so to speak, too many sets required to exist
by some set-theoretical assumption or the other. The charge I am
making against the usual formulations of set theory is that it does not
 AXIOMATIC SET THEORY 177

allow the existence of functions (sets) which ought to exist, and in this
sense assumes too few sets to exist.
Indeed, what leads one to say that the Skolem functions that are
asserted to exist in S actually do so are not a Gödelian set theorist's
mysterious intuitions. They exist by the same criteria as are applied
every day by mathematicians whenever they speak of the existence of
sets and functions. (At least, all classical mathematicians use such
criteria; and I will take care of the constructivists and the intuition-
ists in the last two chapters of this book.) And the criteria they
explicitly or (more likely) implicitly use do not in the present case
depend on any assumptions concerning the existence of sets or of any
other higher-order entities. This can be seen by observing that the
second-order sentence which asserts the existence of the relevant
Skolem functions for T[n] is naturally of the E form. Hence it is
logically equivalent to an IF first-order sentence F. Since F is
first-order, its truth is a purely combinatorial matter, independent of
what one thinks of the existence or nonexistence of sets and func-
tions.
Thus the new difficulties I have pointed out in axiomatic set
theory are not merely an extension of the well-known traditional
ones. The only reason why they do not make axiomatic set theory
totally useless for all mathematical purposes in one fell swoop is that
they concern the entire set-theoretical universe. This is because
a self-sufficient truth definition must involve an entire model. But
what if you try to formulate a truth predicate for those parts of the
set-theoretical universe that are supposed to be used directly in
mathematics, such as the theories of natural numbers, real numbers,
real functions and so forth? The answer is that such a truth predicate
must refer to other parts of the given model of set theory. For
instance, for elementary number theory, the Skolem functions over
which you are quantifying in your truth predicate are second-order
entities, not natural numbers. Hence axiomatic set theory escapes
the ultimate ignominy of implying false propositions about ordinary
mathematical entities only by a tacit stratification of one's math-
ematical universe into what, in effect, corresponds to a logician's
types. But such a stratification itself militates against the allegedly
type-free spirit of set theory, which was supposed to be type-free.
The reaction of a confirmed set theorist to such criticisms of his or
her métier might perhaps be an instance of the old rule: If you cannot
beat them, join them. In the next chapter it will be shown that IF
178 THE PRINCIPLES OF MATHEMATICS REVISITED

first-order logic greatly extends the sphere of what logic can do in the
foundations of mathematics. Why not join forces and simply harness
IF logic to the service of axiomatic set theory? Alas, this strategy fails
unless the foundations of axiomatic set theory itself are radically
revised. The rest of this chapter is devoted to exposing the difficulties
of any straightforward merger of IF logic and set theory.
There is in fact an important reason why axiomatic set theory in
its present form is incapable of using the help offered by IF logic.
This handicap is due to certain basic features of axiomatic set theory.
As was pointed out earlier in this chapter, the main strategy of set
theorists is to capture sets as fully as possible as extensions of
formulas of the language of set theory. The use of the axiom of
comprehension noted earlier is but one manifestation of this strat-
egy. The fundamental fact nevertheless is that this technique be-
comes largely inapplicable when IF first-order logic is being used. If
brief, the usual techniques of axiomatic set theory are in a sense
incompatible with IF first-order logic.
What I mean by this is the following: When a set is captured by
means of an ordinary first-order formula F[x], this can be expressed
by the general equivalence
(Vx)(xec*-+F{xJ) (8.8)

This equivalence can be used as a premise in set-theoretical reasoning.

But if F[x] is an irreducible independence-friendly formula, then (8.8)
fails to be true. The reason is that it implies that F[x] obeys the law of
excluded middle for all values of x, which an irreducibly independence-
friendly formula does not do, as was explained in Chapter 7.
The underlying reason for this situation is that in the usual
axiomatic set theories, the membership relation e is construed as
a primitive predicate. Every assertion of set membership is therefore
an atomic formula. And the law of excluded middle applies to atomic
formulas. Hence we can never truly say in a set-theoretical language
that a set captures the extension of an irreducibly independence-
friendly formula. Hence the entire strategy of trying to capture sets as
extensions of formulas fails to do its job when the logic it is based on
includes independent quantifiers.
This fact is but a particular manifestation of a more general
phenomenon. One of the most crucial questions in the foundations
of axiomatic set theory is how sets are to be captured. As was
indicated in the beginning of this chapter, the most common type of
 AXIOMATIC SET THEORY 179

answer we find in the literature is: As extensions of predicates

well-defined in the language of set theory, or at least as extensions of
some suitably restricted well-defined predicates. Perhaps the clearest
expression is Skolem's and Weyl's explication of Zermelo's (1908)
notion of a definit predicate. Zermelo had introduced this notion to
capture the sets whose existence is asserted by his Aussonderun-
gsaxiom. Skolem (1922) and Weyl (1917) in effect identified definit
predicates with those representable in the language of their first-
order set theory.
More generally, there is an influential tradition which equates the
existence of higher-order entities with their representability in a suit-
able language. This tradition has influenced in various ways the
development of axiomatic set theory. It is nevertheless hard to find
theoretically motivated reasons for the theses of this tradition.
Now we are beginning to see why. When we move from ordinary
first-order logic to IF first-order logic, every reason disappears for
thinking of well-formedness of complex predicates as the benchmark
of their capability to capture a set. Assume that SEx] is a complex
predicate of an IF first-order language, with x as its only free
individual variable. Then the corresponding set s presumably would
be characterized by the sentence
(8.9)

But this sentence can be true only if S[x] satisfies the law of excluded
middle, which is often not the case in IF first-order languages. Hence
the extension of S[x] is a set only if S[x] obeys tertium non datur.
Even if S[x] is perfectly well-formed, it does not capture a set unless
it is a formula of an ordinary first-order language. Thus in an IF
first-order language of set theory, the well-formedness of a formula
does not guarantee the existence of the corresponding set.
This failure of the purported link between well-formedness and set
existence does not have anything to do with the reification involved
in the idea of sets as individuals over which one can quantify. Sup-
pose that we modify (8.9)so that, instead of characterizing the set s, it
now defines a new one-place predicate P. for instance as follows:
(8.10)

Then we are not doing any better than with (8.9). For, like (8.9), the
explicit definition (8.10) presupposes that S[x] conforms to the law
of excluded middle.
180 THE PRINCIPLES OF MATHEMATICS REVISITED

The trouble here cannot be blamed on the absence of a purely

truth-functional equivalence, either. In order to see this, we may
introduce a new connective defined semantically as follows:
is true if and only if (8.11)

(a) both S1 and S2 are true

or
(b) both S1 and S2 are not true (i.e., either false or neither true-
nor-false)
What will happen? if we now use instead of in the explicit
definition of a new predicate? Consider the result of this change in
(8.10), that is
(8.12)

Intuitively speaking, for each value b P(b) is true if and only if S[b] is
true; otherwise it is false. Let us now apply the definition schema (8.12)
to the truth predicate Tr[x]. Then there is in the extended language
a predicate constant P(x) such that
(8.13)

By the diagonal lemma, there is a number n such that n is the Gödel

number of —' p(n), where n is the numeral representing n. If P(n) were
true, it would follow from (8.13) that it is false. But if i—' P(n) were not
true, it would likewise follow from (8.13) that it would be true. Hence
definitions of the form (8.12) are not always admissible.
The existence (admissibility) of definitions like (8.12) is what
Russell's Axiom of Reducibility (Russell and Whitehead 1910—13,
Vol. 1, pp. 55—59) is supposed to guarantee. What has been found is
that in IF first-order logic no principle remotely like Russell's
reducibility axiom holds.
In different terminology, we can say that the failure of well-
formedness to guarantee set or predicate existence is no abject that
even the safest vehicle of well-definedness, an explicit definition, does
not guarantee set existence or even the existence of a quantifier-free
predicate.
These observations are most suggestive. They show that some of
the favorite methods used by logicians and mathematicians cannot
be extended beyond ordinary first-order theories. Explicit defini-
tions cannot be used without strict precautions outside ordinary
 AXIOMATIC SET THEORY 181

first-order logic. And the tactic of set theorists in trying to access sets
via the predicates whose extensions they are supposed to be does not
always work after the possibility of informational independence is
recognized. This in turn shows that Frege's entire approach to the
foundations of logic and mathematics is subject to severe limitations,
for the only role that sets play in his approach is as course-of-values
of predicates ("functions").
Another way of putting essentially the same point is to say that the
usual theory of definition is one of the many features of ordinary
first-order logic that cannot be extended beyond its idiosyncratic
boundaries.
There is a more upbeat way of looking at these results. What has
been found is that the power of IF first-order logic in dealing with
complex predicates extends way beyond predicates that are well-
defined for the purposes of set theorists. Hence once again IF first-
order logic has been found to be a better conceptual tool in the
foundations of mathematics than received set theory.
This suggests a major new way of improving the existing tech-
niques of set theoretical conceptualization and argumentation. This
way consists in departing radically from Cantor's requirement that
set membership must be well-defined, and in allowing for sets which
are only partially defined in the sense that the law of excluded middle
does not apply to membership in such a set: besides definite members
and nonmembers there are also potential elements which are neither.
Such a set theory would be in the spirit of the recently kindled
interest among logicians, mathematicians and formal linguists in
partiality, evinced and surveyed by Fenstad (1996). Indeed, such an
IF set theory would provide a convincing example of the ways in
which allowing partiality can actually strengthen significantly one's
conceptual resources.
A few further comments may be in order. Set theory is sometimes
accused of reifying sets into objects. It is not clear precisely what such
accusations can mean. The treatment of membership relation as
a primitive relation on a par with others offers a possible precisation
of such accusations. For what it amounts to is a treatment of the
relation of a set to each one of its members in the same way as any
relation between two individuals (objects) is treated. And if this is
what the reification of sets means, it is indeed a mistaken or at least
ill-advised procedure, for it cuts axiomatic set theory off from some
of its most important potential resources.
182 THE PRINCIPLES OF MATHEMATICS REVISITED

Axiomatic set theory is thus subject to various theoretical prob-

lems as a framework for doing mathematics, and may even turn out
to be totally unacceptable in its present form. Furthermore, I have
not as much as touched on the difficulties of actually choosing the
axioms of set theory. What assumptions about set existence should
we make? How can we decide? Enough has been said in any case to
justify a search for a better framework for mathematical theorizing.

Note
This conception of mathematical existence as pertaining to some kind of super-
universe of all possible structures played an important role in the foundations of
mathematics around the turn of the century. It was seldom, if ever, discussed
explicitly, and later philosophers and historians have neglected it virtually
completely. Yet it rears its head in the writings of the foremost mathematicians
and logicians discussing the foundations of mathematics of that period. The
following is a brief list of some of the more conspicuous appearances of this idea:
Dauben (1979, pp. 145—146, on Cantor); Frege and Hilbert (Frege 1980, pp. 43—
44); Poincaré (1905—6, p. 819; 1952, pp. 151—152); Tarski (1935, p. 318; 1956,
p. 199).
 9

IF Logic as a Framework for

Mathematical Theorizing

In earlier chapters, especially in Chapters 3, 4, and 7, IF first-order

logic was introduced and its mathematical properties were briefly
studied. The crucially important question here is, nevertheless, not
what mathematics can do for your logic, but what your logic can do
for mathematics. What can logic any logic in principle do in and
for mathematics?
Here we are in the proximity of some of the main traditional, not
to say nostalgic, issues in the foundations of mathematics. For one
thing, the logicist philosophers used to have a simple answer to the
questionjust posed. Their answer was: Everything. Or, as it also used
to be put: Mathematics can be reduced to logic. Unfortunately, the
nostalgia of this answer is not what it was in the good old days. It was
then thought that "the logic" could be formulated as an axiom
system on the same level as different mathematical axiom systems, so
that the purported reduction could be discussed as a reduction of
one axiomatic system to another.
What has been found in this book shows the hopelessness of such
thinking. Not only is a complete axiomatization impossible for the
basic part of logic, IF first-order logic but even when there exists an
axiomatization of some part of logic, it is an axiomatization in
a radically different sense from axiom systems for nonlogical the-
ories. This disparity makes it nonsensical to speak of a reduction of
a nonlogical axiom system to a logical one.
Yet the issues raised by the logicists are relevant and important.
However, instead of trying to formulate them as questions of the
reducibility of one axiom system to another, we should formulate
183
184 THE PRINCIPLES OF MATHEMATICS REVISITED

them as questions about the status of mathematical concepts and

mathematical modes of reasoning vis-â-vis logic. Thus we may — and
must — ask two different questions:

(a) Can one define the crucial mathematical concepts in logical

terms?
(b) Can one express the modes of inference used in mathematics in
logical terms?
It is important to realize that what has been found in this book
changes the received force of the questions (a)—(b). In particular:
(i) In reality, (a) involves two different questions. They are: (al) Can
one define in purely logical terms such concepts as natural
number and real number? (a2) Can one formulate a descriptively
complete axiom system for different mathematical theories?
(ii) Since logic itself is incomplete, the force of (b) is not to ask
whether mathematical modes of inference can be captured by
deductive rules of logic, but rather whether they can be captured
as semantically valid logical inferences.
Here we encounter a problem situation that has prevailed in
twentieth-century philosophical thinking since the demise of the
older logicist program. A philosopher's typical idea of logic is the
ordinary first-order logic. There is in fact much to be said for it as
a framework for mathematical theorizing. Because this logic is
semantically complete, a descriptively complete axiomatization of
any given mathematical theory by means of ordinary first-order
logic automatically yields a deductively complete axiomatization.
Most philosophers would in fact agree that it would be splendid if we
could do all of mathematics within the framework of first-order
logic — wenn das Wenn im Wege nicht ware. But the sad truth is that
all of mathematics just cannot be done in terms of ordinary first-
order logic, or apparently on the first-order level in general. There
are crucially important concepts and modes of inference in mathe-
matics that cannot be captured by means of ordinary first-order
logic. They include the principle of mathematical induction and the
notions of finiteness, infinity, well-ordering, cardinality, power set,
and so forth. And the way these notions are not captured by ordinary
first-order logic is much deeper than merely not being able to set up
complete axiom systems incorporating these notions. The impossi-
bility is model-theoretical. By means of ordinary first-order sen-
 MATHEMATICAL THEORIZING 185

tences one just cannot capture the right class of structures as their
models. For one uncharacteristically simple example, there is no
formula of ordinary first-order logic with identity = as its only
predicate which is true if and only if the universe of discourse is
infinite.
Jon Barwise has expressed this shortcoming of ordinary first-
order logic as follows:

Paging through any modern mathematical book, one comes across concept
after concept that cannot be expressed in [ordinary] first-order logic,
concepts from set-theory (like infinite set, countable set), from analysis (like
set of measure 0 or having the Baire property), and from probability theory
(like random variable and having probability greater than some real number r)
are central notions in mathematics which, on the mathematician-in-the-
street view, have their own logic. (Barwise in Barwise and Feferman 1985,
pp. 5—6)

I will return to Barwise's statement later.

The need to cope with all these different kinds of mathematical
concepts and mathematical modes of reasoning imposes require-
ments on logic that ordinary first-order logic cannot satisfy. But it is
not clear what kind of logic, if any, or what kind of use of logic can do
so. This is what I meant by the current tension between mathematics
and logic in the foundations of mathematics.
Because of this failure of ordinary first-order logic to serve the
needs of mathematics, it is usually thought that mathematical
reasoning is irreducibly set-theoretical or that it involves otherwise
higher-order notions like concepts, functions or predicates. This
lends a tremendous prima facie importance to the third use of logic in
mathematics mentioned at the end of Chapter 1, namely, to the use of
logic (usually ordinary first-order logic) as a framework for doing
axiomatic set theory. Indeed, we have here an explanation for the
allegedly central role of set theory in the foundations of mathemat-
ics. Even though there is little agreement among the philosophical
theories, in practice there is a great deal of uniformity as to how
mathematics should in principle be done. It is precisely axiomatic set
theory which is generally (but not universally) thought of as the
framework in which all serious mathematical theorizing should — or
at least could take place. Unfortunately, it was found in the
preceding chapter that axiomatic set theory in its present form is not
capable of doing what it must be capable of doing to qualify as a
186 THE PRINCIPLES OF MATHEMATICS REVISITED

framework for mathematical theorizing. Hence we have to look for

some other sources of greater logical horsepower than what is
provided for by ordinary first-order logic.
How soon the need for a standard higher-order logic arises can be
seen by considering a Frege-type definition of number as the class of
all equicardinal sets. Equicardinality cannot be defined in ordinary
first-order terms. We not only use higher-order logic in the custom-
ary definition of equicardinality which says that two classes are
equicardinal if and only if there exists a one-to-one relation mapping
them onto each other. The notion of higher-order existence must be
taken in the standard sense. For otherwise there might very well exist
two classes which are equicardinal in the intended pretheoretical
sense but which do not have any nonstandardly existing relation to
map them on each other one-to-one. (In the most common variant of
nonstandard higher-order logic, the higher-order variables range
over those higher-order entities that can be defined in the theory in
question.) If brief, equicardinality is apparently an essentially (that is,
standardly) higher-order or perhaps set-theoretical notion. And if so,
Frege's and Russell's famous reduction could not reduce mathemat-
ics all the way to logic (even if it were otherwise successful), only to
the branch of mathematics called set theory — or misleadingly
labeled "higher order logic". Thus no reduction of mathematics to
anything else is accomplished — or so it seems.
Here IF first-order logic puts things in a radically new light. To
take but one example, even though it may be a short step for one
logic, it is a potentially long step for mathkind to discover that
equicardinality can be defined in IF first-order logic for the exten-
sions of two ordinary first-order predicates, simple or complex,
say F1[x] and F2[x]. The following formula will serve the pur-
pose:

D F2[y]) &
(F2[z] F1[uJ) & ((y = z)*—*(u = x))) (9.1)

A moment's thought will convince you that (9.1) does the job it was
supposed to do. This thought might be facilitated by the second-
order translation of (9.1):

(Jf)(Rg)(Vx)(z)(F1[x] z F2[f(x]) &

(F2[z] F1[g(z)J) & ((f(x) = z)4-+(g(z) = x))) (9.2)
 MATHEMATICAL THEORIZING 187

As a special case we obtain an expression for the equicardinality of

two sets cc and /3.
D ye/i) &
(zefi D uGac) & ((y = z)4-.(u = x))) (9.3)

If we try to remove the restriction of F1 [x] and F2[x] to ordinary

first-order predicates, then we will find out that our definition works
only if F1 [x] and F2 [x] are ordinary first-order formulas.
One interesting fact here is that the equicardinality of two sets
cc and /3 cannot be expressed by an ordinary first-order formula even
for the cases where cc and /3 are finite. Hence we have here an example
of the ways in which IF first-order logic gives us additional power
even for the study of finite structures.
Similar things can be said of the notion of infinity. In mathematics,
we are dealing with an infinite universe. But even before we can face
the question as to whether or not the infinity of the universe of
mathematics has to be assumed or the question whether this is
a logical or a mathematical assumption, we have to see whether or
not the assumption itself can be expressed in purely logical terms, or
whether the mere formulation of the concept of infinity requires the
help of set theory. In ordinary first-order logic, the infinity of the
universe cannot be expressed without the help of nonlogical con-
stants. However, in IF first-order logic the infinity of the universe of
discourse can be expressed by using = as the only predicate. The
following is a case in point:
y) & (z y) & ((x = z)4-*(y = u)))
(9.4)

This is easily seen to be equivalent to

& ((f(x) =f(z))+-*(x = z))) (9.5)

If (9.4) or (9.5) is true, then the universe must be either empty or

infinite.
By relativizing the quantifiers in (9.4) to the extension of a given
ordinary first-order predicate F[x] (simple or complex) we can
likewise define the infinity of this extension in first-order terms. If the
restriction to ordinary first-order formulas is removed, we obtain
a characterization of the infinity of the class of individuals which
does not make F[x] false.
188 THE PRINCIPLES OF MATHEMATICS REVISITED

It is of some philosophical interest to take a closer look at

sentences like (9.4). This sentence contains no predicates other than
the identity relation. Yet it has a highly nontrivial meaning. Where
does this nontrivial meaning come from? The only candidates are the
notion of identity and quantifiers (plus propositional connectives).
Their meaning thus has to be understood independently of any
constant predicates. This suggests an interesting conclusion con-
cerning the notion of identity which is codified in first-order logic.
This notion is not based on, or otherwise derivative from, any
particular constant predicates, such as so-called sortal predicates.
The notion of identity in a given domain can be understood indepen-
dently of one's understanding of any particular predicates of the
members of that domain — indeed, independently of one's under-
standing of any nonlogical predicates at all.
Moreover, a wealth of other mathematical concepts can be de-
fined in IF first-order logic or in its extended version. They include
the following:
(i) The notion of well-ordering. Consider a discrete ordering de-
fined by a (possibly complex) relation R[x,y] expressible in
ordinary first-order language. We can then express in an IF
first-order language the fact that R[x, y] is not a well-ordering.
For the purpose, what we have to express is that there exists an
infinite descending sequence of individuals. This can be ex-
pressed as follows:
y&z u & R[y, x] &
R{u,zJ (9.6)

A moment's reflection easily shows that (9.6) can be used to

express the failure of R to be a well-ordering, in that it says that
there is an infinite descending sequence of individuals in the
order R. Hence the concept of well-ordering can be expressed in
extended IF first-order logic.
(ii) Likewise, the principle of mathematical induction can be for-
mulated in extended IF first-order logic, while its contradictory
negation is expressible in (unextended) IF first-order language.
(iii) The notion of power set can be characterized in extended IF
first-order logic in a sense (see here Krynicki and Väänänen,
1989). The sense in question is that for two simple predicates
A(x), B(x) we can express by means of extended IF first-order
 MATHEMATICAL THEORIZING 189

logic that the cardinality of 8(x) is where is the cardinality of

A(x).
This result shows incidentally that the Löwenheim—Skolem
theorem does not hold in extended IF logic.
(iv) The Bolzano—Weierstrass Theorem is expressible in extended
IF first-order logic. This theorem says that every bounded
infinite set of real numbers has an accumulation point. It has
accordingly the form
(VX)((Inf(X)&Bound(X))
D (ay)(y is an accumulation point for X)) (9.7)
Here the notion of boundedness (Bound) and accumulation
point can be defined in ordinary first-order logic. In (i) above, it
was seen that infinity can be defined in (unextended) IF first-order
logic, ergo by a formula. Hence the Bolzano — Weierstrass
Theorem can be formulated as a HI formula. This suffices to
show that the theorem can be formulated in extended IF first-
order logic.
(v) In topology, the notion of open set can be characterized by
means of IF first-order logic. For instance, we can say that X is
open if there is a function that maps each element of X on
a neighborhood which it is a member of and which is a subset
of X. This can be expressed by a second-order formula,
ergo by an IF first-order formula.
(vi) In the same direction, the topological notion of continuity can
be characterized in extended IF first-order logic. In order to see
how this can be done, it suffices to recall that a function f is
continuous in the topological sense if and only if the inverse of
each open set is open. By what was found in (v), this is expres-
sible in truth-functionally extended IF first-order logic.
(vii) A number of apparently mathematical (set-theoretical) truths
also turn out to be truths of IF first-order logic. A case in point is
the principle of transfinite induction. It can be taken to be an
inference from
= z4-÷y = u) & (A(x) A(y)) &

(A(z)nA(u))&ycx&u<z] (9.8)
to
= zi-*y = u) & y .c x & u <z]
(9.9)
190 THE PRINCIPLES OF MATHEMATICS REVISITED

In (9.8)—(9.9) the ordering relation is taken to be <.An explana-

tion might run as follows: The contradictory negation of (9.9)
says that there is no infinite descending chain. The contradictory
negation of(9.8) says that there is no infinite descending chain of
As — in other words, that there is a smallest A.

These examples show that much more mathematics can be for-

mulated by means of extended (and in some cases even unextended)
IF logic than can be formulated in ordinary first-order logic. It is
amusing to compare the above sample of examples (i)—(vii) with
Jon Barwise's list (quoted earlier in this chapter) of notions that
transcend the powers of logic in its accustomed sense. Several of his
prize specimens of characteristically mathematical concepts have
been shown to be expressible in extended or unextended IF first-
order logic. In Chapters 3 and 4, it was shown that these logics are
so-called with an as good or better right than ordinary first-order
logic.
In conventional philosophical terms, the tremendous increase
in conceptualizing power which IF first-order logic gives could
perhaps be said to aid and abet a logicist philosophy of mathe-
matics. Such statements have to be taken with a pinch of salt,
in that the historical movement known as logicism was using higher-
order logic (and calling it "logic," instead of restricting the term
to first-order conceptualizations) and was also committed deeply
to a nonstandard interpretation of their higher-order logic. How-
ever, in the reconstructed sense of logicism explained earlier in
this chapter, the increased power of IF first-order logic as com-
pared to ordinary first-order logic shows that the new logic indeed
helps greatly the logicist cause. Both conceptualizations and
modes of inference which were earlier impossible to capture on
the first-order level can now be captured by means of IF first-order
logic.
What has been found in any case forces us to radically revise our
ideas of the borderline between logic and mathematics. Several of the
most crucial concepts and modes of inference of mathematics that
have been believed to be beyond the powers of logic to capture have
turned out to be expressible in IF first-order logic.
For instance, the main obstacle to a purely logical definition of
number has been the impossibility of characterizing equicardinality
in terms of ordinary first-order logic. It is generally thought that such
 MATHEMATICAL THEORIZING 191

a characterization inevitably involves quantification over higher-

order entities. IF first-order logic provides a way of defining equicar-
dinality in entirely logical terms and hence removes this particular
obstacle to a purely logical definition of number.
Similarly, the observations made above concerning the principle
of mathematical induction have sometimes been claimed to be
a specifically mathematical principle of reasoning. Now it can be
seen that this principle can be formulated on a first-order and
therefore uncontroversially logical level.
Likewise the notion of power set is one of the most crucial in set
theory. The possibility of characterizing it and the other notions
dealt with above, may thus be seen as a partial and qualified vindic-
ation of the belief of early set theorists that they were dealing with
purely logical conceptualizations.
Hence examples like (ii)—(vi) are not just isolated instances of what
can be done by means of IF first-order logic. They go a long way
toward showing that the inevitable shortcomings of axiomatic
first-order set theory can be overcome by means of IF first-order
logic. For instance, the existence of uncountable sets can be guaran-
teed by means of suitable axioms that can be formulated in terms
of extended IF first-order logic. This follows from what was said
earlier about the power set (see (vi)). My point is not spoiled by the
fact that the existence of nondenumerable sets cannot be expressed
in an unextended IF first-order logic, since such a logic admits of the
Löwenheim—Skolem theorem. Notwithstanding this restriction, the
extended IF first-order logic was seen to provide the requisite means
of capturing such notions as well-ordering.
Thus IF first-order logic apparently can help set theory precisely
where it needs help most badly, namely, in providing standard
formulations of central mathematical concepts. This suggests that
the next step forward in the foundations of mathematics might very
well be to simply combine axiomatic set theory and IF logic or,
rather, to use IF first-order logic as the basic logic of our familiar
axiomatic set theory. This step looks tempting but it is not feasible,
as was shown in the preceding chapter.
In other respects, too, it might very well look as if the euphoria
that was generated by the first victories of IF first-order logic is soon
dissipated. Not only is it impossible to combine ordinary axiomatic
set theory with this new logic. What is more, IF first-order logic is
equivalent only to a small fragment of second-order logic, namely,
192 THE PRINCIPLES OF MATHEMATICS REVISITED

the fragment. Even by means of extended IF first-order logic, our

logic can only capture the u fragment of second-order logic.
Indeed, second-order logic (and, more generally speaking, higher-
order logic) seems to be emerging as the superior rival to IF
first-order logic. In the preceding chapter, axiomatic set theory in its
present form was found to be seriously wanting as a framework
for working mathematics. But the criticisms offered there do not
apply against second-order or higher-order logic. On the contrary,
what has been found in this book might first seem to enhance the
status of second-order logic in the foundations of mathematics. For
instance, in Chapter 5 I emphasized the importance of putting
forward descriptively complete (albeit deductively incomplete)
theories to guide us in extending the scope of actual deductively
developed mathematics. In the preceding chapter, it was shown that
axiomatic set theory cannot perform this service in its present form.
A much more promising candidate for this job is offered by second-
order logic, of course with what Henkin (1950) called the standard
interpretation. (This interpretation was explained briefly in the
preceding chapter.) Stewart Shapiro has recently (1991) argued
ably for second-order logic as the appropriate "foundations with-
out foundationalism" for mathematics. Indeed, a great deal of
mathematics can be formulated in second-order terms in a much
stronger sense than set theory allows us to do. Not only does
second-order logic provide you with languages in which practically
any mathematical theory can be expressed — what is much more, this
second-order formulation, unlike the set-theoretical one, is model-
theoretically faithful to the original. In other words, when a math-
ematical statement S can be formulated in second-order logic it
normally has precisely the models it is normally thought of as
having. Hence, for a mathematical statement S to be a theorem of
a theory characterizable by a conjunction of axioms T all of which
are expressible in second-order language, it is necessary and suffi-
cient that the conditional (TD S) be logically true. All questions
concerning such a mathematical theory can thus in principle be
reduced to questions concerning the logical truths of second-order
logic.
As was mentioned earlier, this remark seems to apply to practi-
cally all mathematical theories and hence to mathematical proposi-
tions that can be formulated in them. Cases in point include, for
instance, many famous unsolved mathematical propositions from
 MATHEMATICAL THEORIZING 193

Goldbach's Conjecture to the Special Continuum Hypothesis. In

principle, though not necessarily in practice, the second-order logi-
cal formulation provides a clear-cut goal for a search for a proof of
such conjectures. Needless to say, the guidelines we can hope to
obtain are indirect ones. They are criteria for the acceptability
(validity) of new rules of inference that can facilitate actual deductive
proofs, rather than tactical principles for carrying out proofs by
means of already known and accepted rules.
Just how much mathematics can thus be handled in second-order
logic requires careful examination. It is not difficult to argue, as is for
instance shown in Shapiro (1991, Ch. 5; cf. also Shapiro 1985) that all
the most common set-theoretical conceptualizations, including
minimal closure, cardinality, continuum hypothesis, well-ordering,
axiom of choice, and well-foundedness, can all be captured by means
of second-order logic. By their means, obviously a great deal of
actual mathematics can be formulated. It looks pretty much as if the
entire classical analysis can be formulated in second-order terms. It
is also easy to see that general topology can be done in this frame-
work. I will return to the limitations of such formulations at the end
of this chapter.
But many hard-nosed logicians will not be happy with the propo-
sal of using a second-order language as a medium for their math-
ematical theorizing, and for a good reason. In order for such a
language to serve its purpose, its second-order variables must be
taken in their standard sense. They must be taken to range over all
extensionally possible entities of the appropriate type (sets, func-
tions, etc.). For instance, one's function variables have to be taken to
range over all arbitrary functions, however noncomputable (cf. here
Hintikka 1995a). But if so, we face all the problems connected with
the ideas of arbitrary set and arbitrary function (strictly speaking all
arbitrary subsets of a given infinite set and all arbitrary functions
from a given infinite set to another given set). I can indicate this kind
of commitment to arbitrary higher-order entities by saying that it
involves the idea of "all sets". Another way of expressing myself
might be to speak of the standard interpretation in Henkin's sense.
But whatever the name that this idea passes under, its smell is equally
foul to many logicians. And there is a great deal to be said for their
perceptions. The idea of the totality of all (sub)sets is indeed a hard
one to master. Therefore, even though a second-order formulation of
mathematical propositions and conjectures may be superior to
194 THE PRINCIPLES OF MATHEMATICS REVISITED

treating them in an axiomatic set theory, it involves further serious

problems.
Higher-order logic, like axiomatic set theory, can of course be
given a nonstandard interpretation. Indeed, the best known in-
carnation of higher-order logic, the theory of types of Russell and
Whitehead (1910—13), was intended to have a nonstandard interpre-
tation (cf. Ramsey 1925).1 But as Frank Ramsey was not slow to
point out, this interpretation is just not in agreement with the
mathematicians' prevalent way of thinking. For this reason, the
nonstandard interpretation of higher-order logic loses all the advan-
tages it lots had on the standard interpretation in guiding mathema-
ticians in their search for stronger axioms of set existence, and for
other mathematical and logical truths that could provide stronger
deductive principles.
Since nonstandard interpretations do not help us to deal with the
problems of set existence, what can? Here IF logic seems to offer its
services to us. As long as we can stay on the level of first-order logic,
independence-friendly or not, then problems of set existence do not
arise. We do not have to open the Gordian knot of set existence since
it was not tied in the first place.
What can be done by means of IF first-order logic? Relatively
little, it might seem. Second-order logic cannot be reduced in any
direct sense to IF first-order logic. As was pointed out, only the
fragment of second-order logic can be translated into IF first-
order logic. The fact that contradictory negations of IF first-order
logic are translatable to fragments of second-order logic, and
vice versa, does not change the picture essentially.2
However, there is an indirect sense in which mathematical
theories and mathematical problems can be reduced by means of
IF first-order logic. What we can do is to try to reconstruct a given
higher-order theory T by means of ordinary first-order logic as
a many-sorted first-order theory. In this reconstruction, each of the
types needed in T will become a sort of its own. The different types
can be thought of as forming a special syncategorematic sort (set) t.
The structure of this set can be expressed by means of explicit
first-order axioms. Furthermore, we can introduce a function
t which maps each entity (other than a type) on its type. Each sort
corresponds to the set of all entities mapped by t on the same type.
Among other things, there will then be a defined relation c(x, y)
which holds if and only if the type of y is the type of sets of entities of
 MATHEMATICAL THEORIZING 195

the same type as x. All relations between entities belonging to

different sorts can be handled by a suitable new predicate E(x, y)
(also written xey) representing membership. Obviously, we can then
rewrite T as a theory T* about our new many-sorted first-order
models. The result will then be somewhat like a stratified set theory,
with the different types of sets separated from each other by having
different sortal predicates, that is, being mapped by t. on different
types.
This way of dealing with higher-order logic as if it were a many-
sorted first-order logic is in many ways a most natural one. The ease
with which we can do it may very well have been one of the reasons
why the distinction between first-order logic and higher-order logic
crystallized so slowly in the consciousness of most logicians and
philosophers, as emphasized by Moore (1988). It is not offered here
as any great novelty. It may even be argued that this is not unlike
how early contemporary logicians like Frege and Russell viewed
their higher-order logics.
In any case, the crucial question is: Why does not the many-sorted
first-order theory T* capture everything there is to be captured by
means of the second-order theory T? The answer is that the many-
sorted first-order formulation does not implement the standard
interpretation which is needed to make T a genuinely second-order
theory. For there is in T* no way of guaranteeing that for each
extensionally possible class of first-order entities (individuals or
tuples of individuals) there is a second-order entity having them (and
only them) as its elements, and likewise for all higher-order entities.
But now comes the pleasant surprise. To implement the standard
interpretation, it suffices to introduce a finite number of additional
second-order axioms, all of which have the form
(VX)(Vw)(Rz)(3y)(Vx)((X(z) (t(z) = w)) D (a(w,y) &
(xey+-*X(x)))) (9.10)

where c(x, y) is the predicate explained above. If desired, they can

even be combined in a single axiom with the same lone initial
second-order quantifier (YX). But any finite conjunction J of sen-
tences of the form (9.10) is of the 111 form, and hence a very special
kind of second-order sentence.
Here the axioms of the form (9.10) are clearly closely related to the
comprehension axiom of axiomatic set theory. However, the com-
prehension axiom of any first-order axiomatization of set theory (or
196 THE PRINCIPLES OF MATHEMATICS REVISITED

any substitution-instance of it, if the comprehension principle is

formulated as an axiom schema) is an ordinary first-order sentence.
Hence it cannot do the same job as (9.10), for the whole point of (9.10)
is that the quantifier (VX) is a second-order one and it has to be taken
in the standard sense.
By conjoining T* and J we obtain a theory (J & T*) which does
the same job as the original theory T. It is not only expressible in
a second-order language; it is expressible in the Hf fragment of that
language.
In the way — and in the sense —just sketched, the entire higher-
order logic can be reduced to the Hf fragment of second-order logic.
The possibility of this "transcendental reduction" (as Husserl might
have said here) has not attracted much attention. It nevertheless
promises some very real advantages.
The appearance of IF logics on the scene throws new light on the
situation. The Hf character of the reducts means that the new theory
is translatable into the language of extended IF first-order logic. In
the sense which appears from these observations, every mathemat-
ical theory which can be formulated in a higher-order language can
also be formulated in extended IF first-order language. This result
can be extended to all theories expressible in terms of finite types. As
was indicated above, this covers most of the actual mathematical
theories. In this qualified sense, virtually all of classical mathematics
can in principle be done in extended IF first-order logic.
Furthermore, the question whether a given proposition C follows
logically from T is equivalent to the question whether the following
sentence is logically true (valid):
(J&T*)DC (9.11)

where J is the conjunction of all sentences of the form (9.10) for all the
different types (sorts) needed in T*. Since we are dealing here with
normal second-order logic, the conditional can be taken to be the usual
truth-functional one, equivalent to
—i(J&T*)vC (9.12)

where iis the contradictory negation. Moreover, T* and C are

(many-sorted) first-order sentences. Therefore (9.12) is logically
equivalent to a sentence. But this means that it has a translation
into an (unextended) independence-friendlyfirst-order language.
 MATHEMATICAL THEORIZING 197

Combined with our earlier remarks on the way mathematical

theories and problems can be represented in second-order lan-
guages, we thus obtain a remarkable result. In an interesting sense,
a great many mathematical theories can be formulated in an ex-
tended IF first-order language. Such a theory can be represented in
the form T, where T is a sentence of an (unextended) IF first-order
language and -i the contradictory negation. This includes all the
mathematics that can be done in terms of a theory of finite types.3
Likewise, a great many mathematical problems can be taken to relate
to the logical status of a sentence of an unextended IF first-order
language. For instance, if the problem is to prove that a conjecture
C follows from second-order axioms T, then there is an IF first-order
sentence, S such that S is logically true (valid) if and only if C is
a logical consequence of T.
The same argument can easily be applied to the entire theory of
finite types instead of to second-order logic.
In the special case in which the original mathematical theory T,
expressed in the theory of finite types, is categorical, the proposition
C is true in the sole model of T (modulo isomorphism) if TI— C. But
this was seen to hold if and only if a certain IF first-order sentence is
logically true. Thus for categorical theories mathematical truth can
in a sense be equated with logical truth in IF first-order logic. In
particular, the truth of any proposition of elementary arithmetic is
equivalent to the logical truth of an IF first-order sentence. In
general, it is an interesting historical question whether this subtle
model-theoretical connection between mathematical truth and
material truth has tacitly been instrumental in confusing people
about the relationship of ordinary truth to logical truth, especially
as many of the best known mathematical theories are, when
formulated in higher-order terms, categorical. Cases in point are
elementary arithmetic, the theory of reals, elementary geometry,
and so forth.
For instance, Peano arithmetic becomes categorical when com-
plemented by a second-order induction axiom. This axiom system
can be seen to be equivalent to a sentence, say N0. Hence the
truth of any unproved arithmetical conjecture F, say Goldbach's
Conjecture, is equivalent to the validity of (N0 F). But this is
equivalent to a sentence and a fortiori equivalent to a sentence F0
of IF first-order language. The truth of Goldbach's Conjecture is
thus equivalent to the validity of an IF first-order formula.
198 THE PRINCIPLES OF MATHEMATICS REVISITED

The upshot of this line of thought is thus a kind of reduction of the

entire finite theory of types, with standard interpretation, to IF first-
order logic. Since most of mathematics can in principle be expressed
in a standardly interpreted finite theory of types, this reduction
throws some interesting light on mathematics in general. For what
can we say of the output sentences of this reduction? They are IF
first-order sentences. All their bound variables range over individ-
uals. This should warm the heart of every philosophical nominalist.
More importantly, their interpretation is completely free of the
logical problems that beset the notion of all subsets of a given infinite
set. An IF first-order sentence is valid if and only if a certain
relational structure cannot help being instantiated in every model.
The problem whether a given IF first-order sentence is valid or not is
therefore a combinatorial problem in a sufficiently wide sense of the
term. This sense is in fact not as loose as it might at first be suspected
to be. How natural it is to think of what is going on in ordinary
first-order logic as being fundamentally combinatorial is shown
vividly by Hao Wang's (1990, essays 10—11) reduction of the decision
problem for ordinary first-order logic to domino problems. And in
this respect IF first-order logic does not differ in principle from the
ordinary one. Indeed, it can be shown that the decision problem for
IF first-order logic can be reduced to domino problems that incor-
porate the characteristic failure of perfect information: some speci-
fied domino tiles must be fitted in without knowing what has
happened elsewhere in the construction (tiling).
Thus, it is not hard to convince oneself that the additional expres-
sive power of IF first-order logic, over and above that of the ordinary
one, is combinatorial in nature, rather than set-theoretical. The overt
novelty of IF logics is a greater freedom of the various dependence
relations between quantifiers and connectives. This is a combina-
torial matter, and not a question of the existence of infinite sets of
different kinds. IF first-order logic is in fact one of the few extensions
of first-order logic which adds to its power in dealing with finite
structures (models). This additional power is exploited by, among
others, Blass and Gurevich (1986). A feeling for this extra force
of IF logic can perhaps be garnered also from examples like (D) in
Chapter 2.
Conversely, the question whether a sentence of IF first-order logic
reduces to an ordinary sentence is a purely combinatorial one. In
order to see this, assume that an IF first-order sentence is given in its
 MATHEMATICAL THEORIZING 199

Skolem normal form

(Vx1)(Vx2). ,
(9.13)

Here each set X1 = .

a subset of {x1,x2
.} is A
sufficient condition for (9.13) to be equivalent to an ordinary first-
order sentence is that the finite sets can be linearly ordered by
class-inclusion. It can be shown that this comes close to being the
most general sufficient condition. In fact, it is the widest sufficient
condition that can be formulated by the sole means of the sequence
of prenex quantifiers.
Our reduction is thus in effect a reduction of practically all
familiar mathematics to combinatorial theory, in a sufficiently wide
sense of the term. This result throws interesting light on the nature of
mathematics in general and on its relation to logic. Among other
things, it shows that all mathematical reasoning can in principle be
considered as being logical in nature. It also shows that the concep-
tual problems connected with the idea of all subsets of a given infinite
set can be completely dispensed with in most of the foundations of
mathematics. Admittedly, mathematical problems are not automati-
cally solved by reducing them to problems concerning IF first-order
logic. However, such a reduction opens various possibilities of
conceptualization. For instance, the structure of the reduct (i.e., of
the IF first-order sentence which is the output of the reduction
process) will reflect the (combinatorial) difficulty of the problem
codified in the original sentence.
My approach has affinities with Hilbert's approach to the founda-
tion of mathematics, which I have discussed elsewhere (see Hintikka
1996). Rightly understood, Hilbert's emphasis was not on formalism
but on the combinatorial (in contrast to set-theoretical) thinking as
the true basis of mathematics.4 This agrees with my thesis that
mathematics is at bottom combinatorial rather than set-theoretical
in character. For instance, our truth definition was formulated by
reference to the combinatorial properties of a truth predicate as
applied to the Gödel numbers of sentences. This is reminiscent of
Hilbert's idea of using the combinatorial properties of a completely
formal language as the basis of his foundational approach. This idea
is not, philosophically speaking, formalistic at all. On the contrary,
we have seen that it supports a realistic conception of truth. The
main difference between us and Hubert is that Hilbert thought that
200 THE PRINCIPLES OF MATHEMATICS REVISITED

finite combinatorics is all we need. It is not, and the most interesting

aspects of IF first-order logic embody problems of infinitary combi-
natorics. In so far as we can look away from this difference, our ap-
proach can nevertheless perhaps even be thought of as a vindication
of Hilbert's approach.
Hilbert's intentions are more clearly seen from his criticisms of his
predecessors than from his constructive suggestions. In Hilbert
(1922, p. 162) he criticizes Dedekind and Frege because they have
operated with general concepts.
Frege has tried to find a foundation for the theory of numbers in pure logic,
Dedekind in set theory as a branch of pure logic, neither has reached his
goal.

The reason is that they have operated in abstract terms with

extensions and intensions of concepts, which according to Hilbert is
insufficient and dangerous. In the place of such general concepts as
the principal objects of a mathematician's attention Hilbert wants to
put
certain discrete extralogical objects which are presented intuitively as
immediate experiences before all thought.
These are obviously individuals in contradistinction to general
concepts. Hubert is therefore advocating as a basis of mathematics
and logic a study of concrete objects that we can grasp directly and
operate with. It does not seem to me far-fetched to call such a vision
of the foundations of mathematics combinatorial.
Admittedly the sense of "combinatorial" in which I am using the
term here is somewhat vague. I believe nevertheless that there are
enough connections between what I have in mind and combinatorics
in the generally acknowledged sense, not only to make my usage
defensible but to make it interesting. Combinatorial theory, especial-
ly Ramsey theory, has turned out to have uses in many different
branches of mathematics, perhaps especially striking in logical
number theory (see here Graham, Rotschild, and Spencer 1990, Ch.
6). In a different direction, infinitary combinatorics, especially parti-
tion theory, is playing an increasingly important role in set theory, as
witnessed by titles like Combinatorial Set Theory (Erdös et al. 1984;
cf. Kleinberg 1977; cf. Williams 1977). It is perhaps symbolic that
Ramsey theory started with his 1930 paper which was devoted to the
Entscheidungsproblem of logic.
 MATHEMATICAL THEORIZING 201

The sense of "combinatorial" presupposed here comes close in

meaning to "first-order", in that first-order reasoning involves only
structures of individuals (particulars), independently of the existence
or nonexistence of any properties or relations or sets or classes, or
any other higher-type entities. Unfortunately, the term "first-order"
is too closely associated with the received attempts to systematize
this basic part of logic. As has been shown in this book, these
attempts do not exhaust all genuinely first-order reasoning. Indeed,
IF first-order logic is neglected in the usual treatments of "first-
order" logic.
The same point can be illustrated by means of historical example.
I have suggested that one of the leading ideas of David Hilbert's
thinking was to base all mathematics on combinatorial realities. It is
therefore in keeping with the link between what I here call combina-
torial reasoning and first-order logic that one of the crucial steps in
the crystallization of first-order logic was taken in Hilbert and
Ackermann (1928; cf. Moore 1988). But Hilbert was not satisfied
with that systematization. He tried to make more explicit the nature
of quantifiers as embodiments of certain choice functions by means
of his epsilon-calculus. As noted above, Hubert hoped in this way to
justify the apparently set-theoretical axiom of choice. This attempt
can be thought of as another way of trying to implement the same
ideas as are systematized in game-theoretical semantics. Moreover,
Hilbert was not completely happy, either, with the treatment of
negation that is codified in the usual systems of first-order logic. For
such reasons, it might be misleading to refer to his logical preferences
as first-order ones. I can only hope that the term "combinatorial"
does not prompt other misunderstandings.
It must be emphasized that my reduction of mathematics to IF
first-order logic is not a translation. One way of describing what is
happening is to consider a given higher-order axiom system T It
specifies a class of (higher-order) structures as its models. The reduc-
tion I have described means that these structures can be captured by
means of extended IF first-order formulas. However, they do not
appear as classes of models of given formulas, but as structures
embedded in such models.
In brief, for each mathematical axiom system A1 expressible in
terms of standardly interpreted higher-order logic, I have shown
how to specify an axiom system A2 formulated in terms of extended
IF first-order logic in which A1 admits of a relative interpretation. In
202 THE PRINCIPLES OF MATHEMATICS REVISITED

this sense, the reduction I have sketched is a reduction by relative

interpretation. Such reductions are sometimes called conceptual
reductions (see, e.g., Feferman 1993, P. 148). As was indicated in the
preceding chapter, such reductions are not very popular these days.
The reason is that such reductions all too often are not what
Feferman calls foundational reductions. Speaking of relative inter-
pretations, he writes (Feferman 1993, p. 148):
a familiar example is that [interpretation] of Peano Arithmetic PA . in
. .

Zermelo—Fraenkel set theory ZF..., where the natural numbers are inter-
preted as finite ordinals. This is a conceptual reduction of number theory to
set theory, but not a foundational reduction, because the latter system is
justified only by an uncountable infinitary framework whereas the former is
justified simply by a countable infinitary framework.

I am here concerned with justification in a sense that is somewhat

different from that of Feferman's. My focus is not the contrast
between countable and uncountable infinity, but the problem of the
existence of higher-order entities. If this is one's focus, then the
reduction I have shown how to carry out is indeed a foundational
reduction with a vengeance. The reduct is more fundamental in an
important respect than the theory to be reduced. Indeed, in the
reduction all questions concerning the existence of higher-order
entities are replaced by questions concerning the truth of extended
IF first-order sentences (in case we want to know if the axioms are
true) and by questions concerning the validity of unextended IF
first-order sentences (if we are interested in matters of theoremhood).
If this is not a foundational reduction, it is hard for me to think what
can be.
These observations can be illustrated and put into perspective by
relating them to the so-called Skolem paradox on the one hand and
the distinction between the two functions of logic in mathematics (cf.
Chapter 1) on the other.
The Skolem paradox is less a paradox than an insight into the
limitations of ordinary first-order logic. It is prompted by the
Löwenheim-Skolem theorem which says that if a first-order sentence
is satisfiable in an infinite model it is satisfied (true) in a countable
model. What follows from this metatheorem is that first-order
languages are not a satisfactory vehicle for discussing noncountable
sets, for no first-order formula or axiom system can distinguish
a noncountable set from countable ones. In the preceding chapter, it
 MATHEMATICAL THEORIZING 203

was indicated how these limitations apparently are manifested in

(first-order) axiomatic set theory. Such applications have probably
contributed to the air of paradox about the Löwenheim—Skolem
theorem. This reputation is in reality thoroughly undeserved.
As was pointed out in Chapter 3, the Ldwenheim—Skolem the-
orem holds in IF first-order logic. How, then, can it serve as the
universal medium of mathematical problems, as was suggested?
What I have shown is that any mathematical theory finitely axio-
matizable by means of higher-order logic can be formulated (in the
sense explained above) as a contradictory negation of an IF first-
order sentence, in such a way that any putative theorem can be taken
to be an ordinary first-order sentence (or an unextended IF first-
order sentence). Then the claim that it really is a theorem is tanta-
mount to claiming that an unextended IF first-order sentence is valid
(logically true).
A mathematical theory formulated in this way by means of
extended IF first-order logic can very well have only uncountable
models, for the Löwenheim—Skolem theorem does not apply to
extended IF logic. At the same time, as was seen, questions of
theoremhood can be reformulated so as to pertain to the validity of
unextended IF formulas only.
This analysis of the situation can be thought of as a resolution of
the Skolem "paradox". It shows that (and in what sense) mathemat-
ical theories can deal with uncountable structures even though their
logic at least the logic of theoremhood can be handled in a logic
that admits of the Löwenheim—Skolem theorem. In other words, we
can now see how mathematical theories can deal not only with sets
but even with uncountable sets, and yet mathematical reasoning is
essentially combinatorial.
There is also something of a correlation here with the first two
functions of logic in mathematics as outlined in Chapter 1. The
descriptive function requires extended IF first-order logic whereas
the consequence relations (questions of theoremhood) turn on the
validity of unextended IF first-order formulas. These observations
will nevertheless be placed in a somewhat different light in the last
section of the next chapter.
One important facet of the possibility of using IF first-order logic
as the framework of mathematics is that then the metatheory of
mathematical theories can be developed by means of the same
logical tools as were used to develop these theories in the first place.
204 THE PRINCIPLES OF MATHEMATICS REVISITED

In brief, the logic of mathematical theories can be self-applied. The

mathematics of metamathematics (cf. Rasiowa and Sikorski 1963)
can in principle be taken to be the same mathematics as it serves us to
metatheorize about.
This can be illustrated by considering the most important specific
results of self-applied logic and mathematics. For instance, my
qualified reduction of a large part of mathematics to IF first-order
logic can be cast into sharper relief by our other results, especially by
our truth definition for IF first-order languages. The possibility of
such a truth definition means that important aspects of an IF
first-order language can be studied by means of that very language.
Even though this possibility does not cover everything in the meta-
theory of the mathematical theories in question, it is in principle
highly interesting because it has often been thought that such self-
study is impossible. More specifically, this impossibility is thought of
as being implied by Gödel's results. For instance, van der Waerdcn
(1985, p. 157) writes:
From this Holder concludes that it is impossible to comprehend the whole
of mathematics by means of a logical formalism, because logical consider-
ations concerning the scope and limits of the formalism necessarily tran-
scend the formalism and yet belong to mathematics. This conclusion is fully
confirmed by later investigations of Kurt GOdel.

Such claims are at best seriously misleading. GOdel did not prove
the descriptive incompleteness of any mathematical theory an sich,
without reference to some underlying logic or the other. On the
contrary, we can see that some of the most important facts about
a suitable logical language, namely, the truth-conditions of its
sentences, can after all be formulated in the language itself.
But suppose I am asked to respond to Tarski's letter to Neurath of
September 7, 1936, quoted in Chapter 1, and to say whether the
problem of a universal language had been definitively disposed of.
What do my results imply concerning the possibility of a universal
language? Is such a language possible? I am tempted to answer, yes
and no. However, in this case yes is closer to the truth than no. What
we cannot have is a single universal language that need not be
extended further. The first truth definition that I explained in
Chapter 6 presupposes that any actual finite language system must
be capable of being extended by adding new individual constants.
Moreover, and even more importantly, there is no hope of formula-
 MATHEMATICAL THEORIZING 205

ting a single language in which all mathematical theories could be

discussed, a language in which the kind of "model of all models"
could be described that set theory was originally thought of as being.
But even though we cannot have a universal language of mathe-
matics, we can have a universal logic. As such a logic, extended IF
first-order logic, has been seen to fill the bill. As has been shown, any
normal mathematical theory can in principle be formulated in
a language whose logic is extended IF first-order logic. If we want to
formulate, study and apply another mathematical theory, we can
formulate it in a similar but different language. Moreover, we can
pool any finite number of such languages together by the simple
device of relativization. (It is worth pointing out that the painlessness
of such a unification is due to the fact that the theories in question
are first-order theories.) After the merger we can study the relation-
ship of the different theories in question in the resulting richer
language.
In so far as such as open and potentially forever growing sequence
of bigger and better languages counts as one language, it qualifies as
the universal language of mathematics. Admittedly, it is not one
language by the letter of most current definitions, and admittedly
different mathematical theories occupy themselves with different
parts of it. Since this language is open one cannot discuss it as a
completed whole, even though at any stage of its evolution the
syntax and semantics of the language fragment so far reached can be
discussed in the fragment itself.
In other words, I am not advocating a universal language of
mathematics in the strict sense of the word. Different branches of
mathematics have different primitive notions. Even if some of them,
for instance the notion of cardinality, could be defined in purely
logical terms, I still do not see any theoretical or practical payoff in
doing so. There does not seem to be any point in trying to combine
different mathematical theories somehow so as to become parts of
some huge super-theory.
What I have argued is that most (and perhaps all) of mathematics
can in principle be done by means of one and the same logic, IF
first-order logic. Moreover, that logic is a genuine article and not
a disguised version of set theory because it is first-order logic and
therefore free of the philosophical problems that have beset set
theory or type theory. It has the same — and a better — claim to the
title of logic as ordinary first-order logic.
206 THE PRINCIPLES OF MATHEMATICS REVISITED

Needless to say, I am advocating this purely logical nature of

mathematical theorizing as a rational reconstruction calculated to
solve philosophical and other theoretical problems. As to practical-
ities, presumably the most intuitive way of mathematical theorizing
would be to carry it out on a second-order level, or maybe with
a simple type theory as its logic; in other words, not unlike what is
done in general topology. Only when theoretical problems that have
to do with the status of higher-order entities that have to do arise will
there be a reason to resort to the kind of reduction to IF first-order
logic described in this chapter.
The overall conception of mathematical theorizing that evolves
from these observations is not entirely unlike Hubert's idea of the
axiomatic method as used for mathematical purposes. This corn pan-
son will have to be a mutatis mutandis one, however, for many of the
features of Hubert's views have been dropped or modified here. They
include for-reaching changes in the nature of the basic logic, and the
abandonment of set theory as a framework for mathematical theo-
rizing, and so forth.
My reduction of all mathematical problems to questions concern-
ing the validity of sentences of an IF first-order language has both
mathematical and philosophical significance. One kind of math-
ematical significance is that it shows that practically all mathemat-
ical problems are at bottom combinatorial rather than set-theoretical.
This implies that the notion of truth applies in mathematical theories.
If you look at set theory, especially in its familiar axiomatic dress,
you have a theory whose intended models are not clearly under-
stood, so that the choice of stronger assumptions seems not to be
guided by questions of truth and falsity but by some vague "intu-
itions", or else by considerations of mathematical taste and expedi-
ency. In contrast, combinatorial problems are clear-cut. Either there
exists a structure of a certain kind or else there does not exist one.
Either your jigsaw puzzle or tiling task can be completed or else it
cannot be. We have in such cases a razor-sharp characterization of
the structures whose existence we are speaking of. This provides no
reason whatsoever for dispensing with the notion of truth. The search
for stronger deductive premises will be guided by one's combinatorial
experience.
This can be illustrated by means of examples. For instance, the
stronghold of the defenders of the notion of truth in mathematical
contexts has always been the structure of natural numbers. We have
 MATHEMATICAL THEORIZING 207

such a marvelous familiarity with it that elementary arithmetic is

spontaneously thought of as dealing with truth and falsity in this
given structure (intended model). Calculatory practice does not
define this structure — it merely helps to provide insights into the
combinatorial structure of this model.
Now we can see that practically all of mathematics can be taken to
deal with similar problems. The structures involved are likely to be
less familiar than that of natural numbers, but the character of the
problems is in principle similar. In a typical mathematical problem,
we are dealing with an axiom system and a putative theorem. The
combinatorial insights needed to decide the status of that alleged
theorem concern the models of the axiom system. They are parallel
to our insights into the structure of natural numbers. They are
supplemented ad hoc by insights into the combinatorial claims made
by the particular putative theorem in question.
In a more philosophical direction, the reduction I have examined
means that abstract entities are not indispensable in mathematics in
the last analysis (pardon the pun). Frege was wrong; mathematics is
not a study of general concepts, but of structures consisting of
particulars (individuals).
The results so far reached suggest a view of mathematics which is
more pluralistic than the traditional idea of mathematics as being
done, not only by means of, but within the scope of logic, set theory
or type theory. This traditional picture is inappropriate in several
different ways. For one thing, there is no reason to think that all
mathematical concepts can be defined within one and the same
language, be it logical or set-theoretical. Different mathematical
theories study structures of different kinds. There is no reasonable
hope to be able somehow to integrate the models of all the different
mathematical theories into a single set-theoretical super-universe, as
the early set-theorists thought they could do.
Furthermore, the results reached in the preceding chapter and
more generally in this book do not only show that axiomatic set
theory is an ill-conceived and dispensable basis for mathematics.
They also show that one of the main rivals of set theory, to wit,
higher-order logics, is likewise dispensable. Both set theory and type
theory are theoretically unsatisfactory in that they construe math-
ematical reasoning and mathematical concept formation as involving
essentially higher-order entities. In reality, as has been seen, practi-
cally all mathematical problems can be construed as problems in IF
208 THE PRINCIPLES OF MATHEMATICS REVISITED

first-order logic, and all mathematical axiomatizations can in principle

be carried out by means of extended IF first-order logic. This makes
type theory as fully unemployed philosophically as set theory is.
Do all mathematical theories and conceptualizations fall within
the scope of our results? Can they all be captured, in the sense
indicated, in extended IF first-order logic? It is not clear what the
answer is, partly because it is not clear what is supposed to count as
mathematics. The success of extended IF first-order logic in captur-
ing set-theoretical and second-order conceptualizations suggests
a positive answer. There nevertheless seem to be ideas even in
perfectly familiar mathematical theories that transcend the purview
of the results discussed here. Hence a more appropriate question
might be: How far do the results obtained in this chapter extend? The
realm of mathematics is a free country. There is no way of anticipa-
ting what kinds of conceptualizations mathematicians might decide
to use. In the next chapter, I will in fact explore a different and novel
perspective on the foundations of logic and mathematics. Hence it is
impossible to say anything absolute here. By and large, it neverthe-
less seems that the observations made in this chapter can be ex-
tended further in several important directions. In the next chapter,
it turns out that there is an eminently natural constructivistic
approach to logic that facilitates a line of thought similar to the one
carried out here.
Furthermore, what is probably the most striking prima facie
exception to what I have said in this chapter can also be shown to be
amenable to treatment on my terms. Even though an explicit
treatment would take us too far afield, this case is important enough
to be mentioned here.
Probably the best known and most important types of mathe-
matical assumption that do not seem to admit of a formulation
in terms of higher-order logic are maximality and minimality
assumptions. And probably the best known example of such as-
sumptions is in turn the so-called axiom of completeness which
was used by Hubert (1899) (see Chapter 5). (Strictly speaking, the
reference should be to the second through the sixth editions of
Hilbert's classic, for the axiom of completeness made its appear-
ance only in the second (1902) edition, and made its exit in favor
of an apparently more frugal assumption in the seventh edition in
1930.) This assumption says in effect that the intended models of
Hilbert's axiom system are maximal in the sense that mathematical
 MATHEMATICAL THEORIZING 209

objects cannot be added to them without violating the other assump-

tions.
Needless to say, maximality assumptions have elsewhere played
an important role in mathematics, too, often in combination with
minimality assumptions. They open in fact an interesting perspective
onto the foundations of mathematics. This perspective is discussed
in Hintikka (1993a).
The reason why such maximality assumptions cannot be directly
formulated in a higher-order logic is obvious. Saying that a model
M is maximal apparently involves quantification over individuals
and/or sets outside the domain do(M) of M, while an axiomatic
specification of M can only be made by means of quantifiers ranging
over entities in M.
It can nevertheless be shown that such maximality assumptions
can also be brought within the scope of the treatment outlined in this
chapter, at least in the cases where they typically seem to have been
used. If maximality and minimality assumptions are added to ordi-
nary first-order theories, they can easily be handled along the lines
indicated in this chapter, thus extending significantly the scope of the
treatment proposed here.
With these qualifications, I venture to say that most of the usual
mathematics can be done within the framework of IF logic. More
specifically, typical mathematical problems can all be reduced to
questions concerning the logical truth of different formulas of (un-
extended) IF first-order logic. In this sense, mathematics can in
principle be done on the first-order level, and in this sense, mathe-
matics can be thought of as a combinatorial rather that a set-
theoretical exercise.
My reduction of all mathematical theorizing to IF first-order
logic needs to be viewed in a wider perspective. This perspective is
provided by the distinction that I made in Chapter 1 between the
descriptive and the deductive functions of logic and mathematics. If
you review the reduction carried out in this chapter, you will see
immediately that it concerns only the descriptive function of logic in
mathematics. It concerns the question of what kind of logic is needed
to capture and to master intellectually the structures (or classes of
structures) mathematicians might be interested in.
In contrast, my reduction does not mean that we can restrict to the
first-order level the tools needed for the purpose of dealing with
mathematical theories deductively. Indeed, I have apparently
210 THE PRINCIPLES OF MATHEMATICS REVISITED

strayed unrealistically far from all questions concerning actual

logical inferences and hence far from the deductive function of logic.
For instance, in ordinary first-order logic, the truth of the condi-
tional (S1 S2) is precisely what is needed to move from the truth of
S1 to the truth of S2. In contrast, the truth of (S1 D S2) (i.e., of
v S2) in IF first-order logic is much more than what is needed
to make this move in IF first-order logic.
Hence something more has to be said about the deductive func-
tion of logic from the viewpoint of the game-theoretical approach to
logic and mathematics. This task will not be attempted in this book,
although some light on the matter will be shed by my last two chapters.
There, a new aspect of the foundations of logic and mathematics will
be examined, namely, the interpretation and implementation of the
claims of the constructivists.

Notes
1
Warren Goldfarb (1989) may be right in defending Russell against Ramsey if we assume
Russell's nonstandard interpretation in the first place. But in a deeper sense, it was precisely
the nonstandard interpretation that Ramsey was criticizing.
2
For further information about the scope of and logics, see Moschovakis (1974),
especially Chapter 7.
By the theory of type 1 mean theory of simple types in contradistinction to Russel's ramified
theory of types.
The term "combinatorial" has to be handled with care. What I want to emphasize by the use
of the term is the combinatorics of the objects of mathematical theories, and not the
combinatorics of the formulas dealing with those objects. This emphasis is therefore
radically different from the motivation of the use of the same term by Benacerraf (1973).
 10

Constructivism Reconstructed

The approach represented in this book has a strong spiritual kinship

with constructivistic ideas. This kinship can be illustrated in a va-
riety of ways. One of the basic ideas of constructivists like Michael
Dummett (1978, 1993) is that meaning has to be mediated by teach-
able, learnable, and practicable human activities. This is precisely the
job which semantical games do in game-theoretical semantics. These
games can be thought of as being a variety of Wittgensteinian
language games. Now these very same Wittgensteinian ideas have
been one of the main sources of inspiration to contemporary con-
structivists. In view of this close relationship of my ideas to those of
the constructivists, it is in order to ask what relevance the concepts
and results reached here might have to the prospects of a construc-
tivis tic theory of the foundations of mathematics.
The answer to this question is not immediately obvious. It might
seem that the results reached in the earlier chapters of this book
entail a virtual Aufliebung of all constructivistic approaches to the
foundations of mathematics. This loaded Hegelian term is appropri-
ate because it almost looks as if I had perhaps vindicated the
constructivistic approach to logic and mathematics by refuting it. As
was mentioned, the basic ideas of my approach are very much in the
spirit of constructivistic ways of thinking. Yet I have ended up
apparently rejecting many of the characteristic tenets of construc-
tivism. Admittedly, I have not aimed at refuting a constructivistic
stance tout court, and I do not claim to have done so. However,
I have apparently realized some of the main ends of the construc-
tivists without committing myself to the consequences which
211
212 THE PRINCIPLES OF MATHEMATICS REVISITED

constructivists themselves embrace. On the one hand, a perceptive

reader has undoubtedly noted that I have satisfied some of the most
characteristic desiderata of constructivists like Dummett. In par-
ticular, in my approach to meaning, truth is grounded on certain
humanly playable language games, just as Dr. Dummett ordered.
On the other hand, I have ended up rejecting virtually all the
consequences which constructivists hold dear. I have abided by
classical mathematics. Admittedly, I suspended, so to speak, tem-
porarily the law of excluded middle. But it was shown that this
rejection of tertium non datur for the strong negation is an inevitable
feature of every natural basic logic and that it therefore has nothing
whatsoever to do with constructivism. Indeed, in the preceding
chapter it was in effect shown that you can do all of classical
mathematics in a language in which tertium non datur fails. And to
add logical injury to philosophical insult, I showed how to rein-
troduce unreconstructed contradictory negation with its undiluted
law of excluded middle.
Moreover, I have repeatedly criticized the cornerstone of the
usual strategy of the constructivists. This strategy amounts to
characterizing a constructivistic interpretation of logic and/or
mathematics by presenting a nonclassical set of rules of logical proof.
This methodology has been used by Heyting (1956) in presenting
a formal system of intuitionistic logic as well as by such later
constructivists as Dummett and Prawitz. It was shown earlier that
such changes in the rules of logical proof cannot be what the
constructivists are really trying to do.
Moreover, the rules of semantical games should likewise be
acceptable to a constructivist. In order to verify an existential
sentence (dx)S[x] I have to find an individual b such that I can verify
(win in the game played with) S[b]. What could be a more construc-
tivistic requirement than that? Likewise, in the verification game
v connected with a disjunction (S1 v the verifier must
choose or S2 such that the game connected with it (i.e., G(S1) or
G(S2)) can be won by the verifier. Again, there does not seem to be
anything here to alienate a constructivist.
Also, the rules of my semantical games can be taught and learned.
Two human beings can play these games against each other. Indeed,
this is how Peirce thought of his semantical games (connected with
quantifiers) as being played (see Hilpinen 1983). Semantical games
are also finite in that each one comes to an end after a finite, pre-
 CONSTRUCT! VISM RECONSTRUCTED 213

dictable number of moves. There is no opening here for a con struc-

tivistic criticism of game-theoretical semantics.
At first sight, the approach represented here might seem to be
committed to nonconstructive assumptions in other ways. For
instance, the main vehicle of my frequent trips between first-order
and second-order logic has been the principle of choice, which is
often considered a paradigm example of a nonconstructive assump-
tion. For I have used as a second-order translation of any first-order
sentence S of IF (or ordinary) first-order language the sentence that
asserts the existence of the Skolem functions for S. (Plus, of course,
the functions which are needed to take care of dependent disjunc-
tions.) Now even in the case of a simple sentence of the form
y] (10.1)

its second-order counterpart is

f(x)] (10.2)

But the equivalence of sentences like (10.1) and (10.2) is a form of the
principle of the choice and hence seems to involve a heavy dose of set
theory. In Chapter 2 it was nevertheless shown that this apparent use
of the axiom of choice is harmless, being merely an implementation
of the general idea of game-theoretical semantics. What is more,
I will show later in this chapter that such uses of the principle of
choice can be vindicated even from a constructivistic viewpoint.
All this amounts to a serious indictment of the usual ways of
implementing constructivistic ideas. However, this does not imply
that constructivism as such is wrong. Indeed, my game-theoretical
approach helps us to see what the true nature and the true prospects
of constructivism are. This includes changes in the logic and seman-
tics which I have expounded and used in earlier chapters. What is
more, we obtain a new kind of putative argument for a constructivis-
tic approach to logic and the foundations of mathematics. These new
opportunities for a constructivist are based squarely on the basic
ideas of game-theoretical semantics. In the light of these ideas, it can
be seen that a nonstandard interpretation of logic and mathematics
cannot be implemented merely by modifying the definitory rules of
the game of logical proof, or much less by modifying the rules of my
semantical games of verification and falsification. So what else can
a constructivist do? Here the notion of strategy, which is at the heart
of all nontrivial uses of game-theoretical ideas, once again becomes
214 THE PRINCIPLES OF MATHEMATICS REVISITED

crucial. Even if we cannot change the move-by-move rules, we can

alter the strategy sets which are available to the players. Now in GTS
the truth of a sentence S is defined as the existence of a winning
strategy for the initial verifier in the corresponding semantical game
G(S). This winning strategy can be represented by a function (or
finite number of functions). In the simplest — yet fully representa-
tive — case they are functions from positive integers to positive
integers. But what kinds of functions? Here my game-theoretical
viewpoint facilitates a "transcendental deduction" of the way in
which constructivistic ideas in the foundations of logic should be
implemented. Or, more explicitly speaking, it shows two different
ways which constructivists can take, a narrower and a broader
one.
The key to this "transcendental deduction" is unsurprisingly the
game-theoretical definition of truth as the existence of a winning
strategy for the initial verifier. Here the word "existence" has to be
taken literally. It might be tempting to formulate the truth-condition
by saying that a sentence S is true if and only if the initial verifier
"has" a winning strategy in the game G(S). This formulation would
be at least seriously misleading. For, in order to reach the classical
conception of truth we must stick to the letter of the definition. There
may exist a winning strategy for the initial verifier in the abstract
sense of the existence of the relevant strategy function without any
actual player being cognizant of it, or perhaps even without any
player being able to know the winning strategy. This provides an
opening to a constructivist. He or she can require that the verifying
strategy functions be knowable functions in the sense that an actual
human player can play the semantical game in question in accord-
ance with them.
This idea can be interpreted in at least two different ways. Either
of them leads to an interesting implementation of constructivistic
ideas. In this chapter I will consider first one particular interpreta-
tion of the requirement of playability and leave the others for the
next chapter. The demand of playability might seem to imply that
the set of the initial verifier's strategies must be restricted. For it does
not seem to make any sense to think of any actual player as following
a nonconstructive (nonrecursive) strategy. How can I possibly follow
in practice such a strategy when there is no effective way for me to
find out (or perhaps even know) in general what my next move will
be? Hence the basic ideas of the entire game-theoretical approach
 CONSTRUCTIVISM RECONSTRUCTED 215

apparently motivate an important change in the semantics of our

first-order languages (independence-friendly or not) and in their
logic. The resulting semantics is just like my earlier game-theoretical
semantics, except that the initial verifier's ("myself's") strategies are
restricted to recursive ones. This is a perfectly well-defined change. It
leaves the notation used here completely unchanged (independently
of whether the slash symbol is present or not). It also leaves all the
game rules (rules for making moves in a semantical game) un-
touched. Hence it represents an unusual and subtle kind of change in
our semantics and our logic. Its result will be called a constructivistic
game-theoretical semantics.
The change involved in the transition to the new version of GTS is
motivated by precisely the kind of argument which appeals to
constructivists, and which according to them ought to appeal to
everybody. For the basis of my argument was the requirement that
the semantical games that are the foundations of our semantics
and logic must be playable by actual human beings, at least in
principle. This playability of our "language games" is one of the most
characteristic features of the thought of both Wittgenstein and
Dummett.
I will postpone to a later part of this chapter a discussion whether
the constructivistic GTS can really be adequately motivated by the
line of thought just outlined. Meanwhile, I will take the constructiv-
istic GTS as an implementation of the con structivistic approach to
logic and foundations of mathematics and study its properties.
Naturally lam especially interested in the question as to which of the
results concerning the foundations of mathematics reached in the
preceding chapter remain valid. The following are among the most
salient points that can be made here:

(i) The motivation for a constructivistic GTS sketched above

might at first seem to be far removed from the intuitions of the
intuitionists and other constructivists. It is nevertheless easy to
see that our motivation is in reality very close to that of the
constructivists. This can be seen by examining what the restric-
tion to recursive strategies amounts to in particular cases. For
instance, in (10.1) it means that (10.1) is true only if the truth-
making value of y can be calculated from any given value of
x by some Turing machine. But this is very much like what
the constructivistics require in stipulating that all existential
216 THE PRINCIPLES OF MATHEMATICS REVISITED

statements must be effective ones. Likewise, a dependent dis-

junction like
(Vx)(S1[x] v S2[x]) (10.3)

will now be equivalent to

(af)(vx)((S1[x]&f(x) =0) v (S2[x]&f(x) (10.4)

where f is a recursive function. But this means that the choice

of the truth-making disjunct can be made effectively, by calcu-
lating the value of f(x) for the given argument x. And what this
means is that the choice between disjuncts in (10.3) admits of
a constructive decision principle.
If there is an apparent difference between constructivistic
GTS and the ideas of soi-disant constructivists, it is due to a
mistake of theirs concerning the way in which cons tructivistic
ideas should be implemented. They apparently think that such
an implementation must take the form of changing the classical
definitory rules of the "game" of logical and mathematical
proof. What such proofs establish is logical truth or logical
consequence, and not its material truth which is the real issue
here. Admittedly, the rules of logical proof have to be changed
when we move from classical first-order logic to constructivis-
tic logic of the kind explained here. However, the requisite
changes are governed by principles quite different from the
ones constructivists suppose. Later, it will be seen that the
changes intuitionists make in the rules of logical proof are not
the right ones anyway. Indeed, in proofs serving to establish
logical truth, it scarcely even makes sense to require that an
existential statement like be constructive in the sense
of always giving me a recipe for buttonholing a case in point.
In such a proof, we are not arguing about the truth, for
example, of an existential statement but about its
possible truth. For such a purpose, what can possibly be wrong
with my saying that I choose to argue about an arbitrary
specimen of the xs satisfying S[x]? In constrast, when it
comes to the ordinary material truth of existentially quantified
formulas, especially dependent ones, the constructivistic
requirements suddenly begin to make sense. Hence the con-
structivists' concerns about the rule of existential instantiation
in proofs of logical truth are misplaced.
 CONSTRUCTIVISM RECONSTRUCTED 217

(ii) The step from the ordinary GTS to the constructivistic one
makes a difference already to ordinary first-order logic. It was
shown a long time ago that there are arithmetical sentences
that are satisfiable in the domain of natural numbers but which
do not have any recursive Skolem functions to satisfy them (see
Kreisel 1953, Mostowski 1955). This means that when we move
to constructivistic GTS there will be fewer arithmetical sen-
tences true in the domain of natural numbers than before.
The same restriction also lends a new interpretation to IF
first-order logic, not just to ordinary first-order logic. In both
cases, the consequences of the new interpretation have to be
studied separately.
(iii) It is natural to define the falsity of a sentence in constructivistic
GTS as the existence of a recursive winning strategy for the
initial falsifier, that is, a recursive strategy which wins against
any strategy (recursive or not) of the initial verifier. If so, there
will be sentences of ordinary first-order logic which are neither
true nor false in certain models. Thus, constructivists are not
entirely wrong in focusing on the failure of tertium non datur as
a possible symptom of a constructivistic approach. Unfortu-
nately for them, the law of excluded middle fails in perfectly
classical nonconstructivistic logic as soon as informational
independence is allowed. Hence the failure of the law of
excluded middle is not a sufficient condition of a constructivis-
tic logic, in our sense of constructivism. The failure of tertium
non datur is a fetish, and not a touchstone, of constructivism.
(iv) Now we can see another aspect in which traditional construc-
tivists are barking up the wrong rules. It is not the definitory
rules of semantical games that need to be changed, even if we
adopt a constructivist stance. It was seen earlier that the rules
of logical proof for a constructivistic logic will have to be
different from what they are in classical logic. With respect to
the definitory (move-by-move) rules of semantical games, the
situation is different. When we step from the classical first-
order logic to constructivistic logic, no changes whatsoever are
needed in the rules that govern the way the two players make
their moves. Thus the step from classical to constructivistic
logic is as if we could change chess into a different game
without modifying its usual rules for making moves, for mating
and so forth, merely by putting restrictions on the strategies
218 THE PRINCIPLES OF MATHEMATICS REVISITED

chessplayers may use. This invariance of the concrete rules for

playing a semantical game shows that classicial and construe-
tivistic logics are on a par when it comes to teachability,
learnability and the rest of actual playability, apart from
questions of strategy. If it is possible for human beings to learn
the rules of constructivistic logic, then it is equally possible for
them to learn the rules of classical semantical games. It is
perfectly possible for actual human beings to learn and master
the definitory rules (step-by-step rules) of semantical games
which give rise to classical logic. The same holds for my
constructivistic logic. It is therefore beside the point for the
constructivists to try to alter any move-by-move rules of
semantical games. What is at issue here, and what is much more
important conceptually, is the question as to what strategies we
actual human beings can use in semantical games. If construe-
tivism is to be justified, it is here that its justification must lie.
This point can be generalized. It is in principle possible to
think of some automaton or other as following any consistent
set of definitory rules of a game, or at least structurally similar
rules as long as the applicability of those rules is mechanically
decidable, as they are in semantical games and in logical proofs.
Hence it is very hard to think of any reason why any game with
decidable definitory rules could not be playable in principle.
Yet some constructivists are in effect trying to argue that the
classical games of theorem-proving and/or verification, even
though they have perfectly well-defined decidable defmnitory
rules, are not playable by human beings. They are accordingly
trying to prove the impossible. Instead, they ought to have
concentrated on the strategic rules of the relevant games right
from the beginning.
(v) In any case, constructivism in my sense is not aiding or abetting
antirealism in any size, shape or form. Many of the desiderata
that antirealists have aimed at were built right into game-
theoretical semantics, but they did not in the end have the
antirealistic conclusions that philosophers like Dummett have
expected them to have. The constructivistic interpretation
outlined above does not change the situation very much. For
instance, constructivistic truth-conditions are completely on
a par with classical (game-theoretical) ones. One can perhaps
try to argue that any game-theoretical truth-condition is an
 CONSTRUCTIVISM RECONSTRUCTED 219

antirealist one in that it involves games playable by human

beings. But what matters are the rules of semantical games, and
not the psyche, the epistemic state or the cognitive capacity of
the players. The truth of a first-order sentence in a given model
is a combinatorial fact about this model. Whether this fact
obtains or not is independent of whether any human being (or
any robot, for that matter) ever plays the relevant semantical
games.
(vi) The ideas which I have codified in my constructivistic logic and
semantics are so natural that it is no big surprise that they are
not completely news to mathematicians. What is new is the
general formulation of the ideas involved, One particular
illustration is worth mentioning here. Semantical games have
been used in the form of what are known as Diophantine games
in number theory. They were introduced by James P. Jones
(1974). A handy exposition of their main features is found in
Matiyasevich (1993). They are straightforward instances of the
semantical games relied on and discussed in this book, played
with number-theoretical formulas on the domain of natural
numbers. The reason why they are relevant to the present
chapter is that the Diophantine games defined by relatively
simple number-theoretical equations exemplify the difference
between constructivistic and classical games discussed here, in
that in relatively simple games there exist winning strategies for
one of the players, but no recursive ones. Indeed, this is a reason
why Diophantine games are used in studying questions per-
taining to the effective solvability of Diophantine and other
equations. Other games where similar nonrecursivity phenom-
ena are encountered are studied by, among other authors,
Rabin (1957).
Speaking more generally, most of the entire study of the
effective solvability of mathematical problems, such as the
study of Hilbert's tenth problem, can be cast in the form of
a study of the constructivistic truth of mathematical state-
ments in the sense involved here.
Incidentally, such illustrations also serve to drive home the
objectivity of the basic notions of constructivistic GTS. By and
large, these studies show that the law of excluded middle will
fail for surprisingly simple arithmetical statements in the sense
that these statements fail to be true or false on the constructivis-
220 THE PRINCIPLES OF MATHEMATICS REVISITED

tic interpretation. While this fact does not by itself prove any-
thing, it certainly suggests that this constructivistic interpreta-
tion is out of step with our natural ideas about truth as applied
to number-theoretical statements. Indeed, the solvability of an
equation means the constructivistic existence of a solution. The
very question of effective solvability would become empty.
(vii) On the philosophical level, it might nevertheless seem as if this
irrelevance of GTS and of constructivism to the problem of
realism could be mooted. For our game-theoretical truth-
conditions are formulated in terms of strategies of verification.
Strategies are essentially functions of a suitable kind. If an
antirealist is also a nominalist, then he or she cannot accept
such entities as functions and cannot quantify over them.
Hence for such a thinker there apparently is no way of defining
realism in the way we have indicated. This loophole is closed,
however, by the truth definition which was explained earlier for
IF first-order languages. It is on the first-order level, and hence
it does not presuppose any higher-order entities like functions.
Hence all reliance on the existence of higher-order entities can
be avoided, and a fortiori the assumption of nominalism does
not strengthen the hand of an antirealist.
(viii) Admittedly, I have used second-order logic liberally to moti-
vate the approach I have proposed and the definitions and
other analyses it gives rise to. This ascent to second-order logic
cannot be faulted for its apparent use of higher-order entities,
for it was seen that these higher-order entities cancel out from
our most important end-products, such as truth definitions.
But it might still seem dubious for procedural reasons. The
second-order translations of first-order sentences are accept-
able, so it may appear, only if we presuppose the axiom of
choice, which is often taken to be an archetypically non-
constructive assumption.
We are in this quandary, it seems, even in dealing with the
very simplest sentences with dependent quantifiers. Consider,
for instance, a sentence of the form
y] (10.1)
where S [x, y] is quantifier-free. Its translation (looking away
from dependent disjunctions) is
f(x)] (10.2)
 CONSTRUCTIVISM RECONSTRUCTED 221

But the validity of the translation presupposes that (10.1) and

(10.2) are logically equivalent. And this equivalence not only
presupposes the axiom of choice; it is (a form of) the axiom of
choice. Hence my entire procedure, involving as it does con-
stant commuting between first-order and second-order levels
of the kind illustrated by (10.1)—(10.2), seems to rely on a bridge
which cannot bear any constructivistic traffic.
At this very point, the game-theoretical approach puts the
axiom of choice in an interesting light. For what is the relation
of (10.1) to (10.2) in my constructivistic GTS? In (10.2), the
function variable f must of course be now restricted to recur-
sive values (recursive Skolem functions). Furthermore, (10.1) is
now true if and only if there is a recursive winning strategy for
its initial verifier. Such a winning strategy is partially codified
in a Skolem function f satisfying
(Vx)S[x, f(x)] (10.5)

But to say that there is a recursive function of this kind is pre

cisely to assert (10.2), constructivistically interpreted. Hence
(10.1) and (10.2), as they stand, are equivalent also according to
constructivistic GTS.
In brief, if the principle of choice is formulated as an explicit
axiom, it remains valid on my constructivistic interpretation.
The idea that the principle fails when constructivistically inter-
preted is due to an optical illusion or perhaps rather, to mix
metaphors, to a vicious double standard. When (10.2) is rejec-
ted as not being equivalent to (10.1), the function quantifier
is in effect interpreted constructivistically but the first-
order quantifier (dy) classically. Such an incoherent procedure
can only create confusion, however. When all quantifiers are
interpreted constructivistically, no matter what their level, the
principle of choice changes from a daring hypothesis to a mere
definition of a constructivistic interpretation of first-order
quantifiers. This is the vindication of the axiom of choice
anticipated above. It shows that I did not make any noncon-
structivistic commitments when I used equivalences like that
between (10.1) and (10.2) in my argumentation.
A review of my vindication of the axiom of choice shows that
its nerve is simply the game-theoretical definition of truth as
the existence of a winning strategy for the initial verifier. This
222 THE PRINCIPLES OF MATHEMATICS REVISITED

illustrates what was said in Chapter 3 of how the axiom of

choice is little more than one particular facet of the game-
theoretical approach to logic in general. Ironically, this ap-
proach was originally inspired by constructivistic ideas in the
guise of Wittgenstein's notion of language-games.
These observations confute those philosophers and math-
ematicians who have expressed doubts about explicitly for-
mulated versions of the principle (axiom) of choice. They
should encourage those latter-day intuitionists who have
quietly sought to reinstate the principle of choice. An example
of such recantations of the intuitionists' initial criticism of
the axiom of choice is found in Dummett (1977, pp. 52—54).
Dummett motivates his acceptance of the axiom of choice by
borrowing in effect a page of my book, saying that the axiom of
choice "is only dubious under a half-hearted platonistic inter-
pretation of the quantifiers." The thrust of my remarks would
also have delighted Hilbert, who once expressed (as we saw) the
hope that the principle of choice could be shown to be as
obvious as 2 + 2 = 4. In constructivistic GTS it virtually is just
that, quite as much as it is in classical GTS. Thus we are ready
to put to rest for good all the doubts concerning the axiom of
choice in the foundations of mathematics, except for an epitaph
to be presented in Chapter 11.
(ix) Almost by the same token, it can now be seen that my
constructivistic logic does not offer any aid and comfort to
axiomatic set theory. The hopes that were tentatively raised in
Chapter 8 turn out to be false. The suggestion was that by
interpreting quantifiers constructivistically we could vindicate
the intuitively accurate truth predicate for axiomatic set
theory which turned out to be classically unacceptable in
Chapter 8.
(x) My constructivistic interpretation of logic might at first sight
appear not only new-fangled but somewhat strange. It is
therefore in order to point out that it has a respectable, albeit
only partial, precedent. It is Gödel's interpretation (1958) of
first-order logic and arithmetic, known variously as his func-
tional interpretation or the Dialectica interpretation, in honor
of the forum of its initial appearance.
Gödel's interpretation begins just like mine. Each first-
order sentence S is replaced (interpreted) by a second-order
 CONSTRUCTIVISM RECONSTRUCTED 223

sentence S* which asserts the existence of the Skolem function

of S (plus, of course, the similar "Skolem functions" of the
disjunctions of S). Then the range of the function quantifier
which asserts this existence is restricted to recursive functions.
This is in effect the same as we have done.
In Gödel's Dialectica interpretation, however, something
else is also done. In effect, negations and conditionals are
translated into the second-order forms by means of nonstan-
dard rules. These rules are chosen in such a way that the trans-
lation process can lead us beyond second-order formulas
to even higher-order ones. Basically, the rule for negation
interprets it as the contradictory negation, and the rule for
conditional is chosen so that F and (F = 0) are equiva-
1

lent.
These two rules can nevertheless be chosen in a different way
which is arguably much more natural. If Gödel's functional
interpretation is modified in this way, it comes closer to ours
(see Hintikka (1993c).
(xi) Gödel's functional interpretation illustrates how our construc-
tivistic interpretation can be extended to higher-order logics.
What we need to do is to translate each statement of order n to
a statement of order n + 1 all of whose existential quantifiers
are initial quantifiers of order n + 1 (at most). This translation
can be carried out in just the same way as in transforming
first-order sentences into equivalent second-order sen-
tences. Then all functions and functionals are restricted to
recursive ones, just as in Gödel's interpretation. Since the
axiom of choice is available, much of traditional mathematics
remains in force on this interpretation.
(xii) The constructivistic logic envisaged here has certain major
advantages over ordinary first-order logic. One of them is
a corollary to Tennenbaum's old result (1959) to the effect that
the only recursive model of Peano arithmetic is the standard
(intended) one. By a recursive model one means here a model in
which the basic relations of elementary arithmetic (sum of,
product of) are recursive. But they are recursive if and only if
the corresponding Skolem functions can be recursive. And to
require the existence of recursive Skolem functions is precisely
what characterizes the constructivistic interpretation of logic
examined here.
224 THE PRINCIPLES OF MATHEMATICS REVISITED

In brief, by means of my constructivistic logic, one can

formulate a descriptively complete (categorical) axiom system
for elementary arithmetic.
From this result it follows of course that one cannot find
a semantically complete axiomatization of the constructivistic
logic tentatively proposed here. For if it were possible to find
such an axiomatization, elementary arithmetic would admit of
a deductively complete axiomatization and hence of a decision
procedure. This is in fact an example of the kind of tradeoff
between descriptive and semantical completeness that was
discussed in Chapter 5. It fulfills, at least by way of an example,
the promissory note issued in the same chapter to the effect that
there are descriptively complete theories of elementary arith-
metic (based of course on a semantically incomplete logic)
already at the first-order level.
(xiii) One of the advantages of my constructivistic interpretation of
IF first-order logic is that it tightens up the correlation between
(ordinary and IF) first-order formulas and computer architec-
tures explained in Chapter 4. It was pointed out there that this
correlation leaves some first-order formulas high and dry in
that their Skolem functions are not computable, in which case
no computer architecture can be associated with them. What
the constructivistic first-order logic envisaged here accom-
plishes is to require computability of the Skolem functions of
a sentence as a necessary condition of its satisfiability. This
removes the unassociated formulas from the correlation and
thus a correspondence between all satisfiable first-order for-
mulas and certain computer architectures. This goes for both
ordinary and IF first-order formulas. The details are explained
in somewhat greater detail in Hintikka and Sandu (1995).
More generally, the constructivistic interpretation of logic
considered here is very much in the spirit of the times, in the
sense that it puts a premium on the computability of the
mathematical entities we are dealing with. This is compatible
with, if not justified by, the practical importance of computers
and computation in contemporary life.

After all these specific comments on the reconstruction of con-

structivistic logic and mathematics considered here, it is time to
return to the general moral of my story. One large-scale question
 CONSTRUCTIVISM RECONSTRUCTED 225

that arises here is: What is changed and what remains unchanged
when a constructivistic interpretation is adopted? This question has
to be answered case by case.
Perhaps the most intriguing fact about IF first-order logic is that
the entire fragment of second-order logic can be translated into it.
By reviewing the translation process it can be seen that this trans-
latability remains valid also when a constructivistic interpretation
(in the sense of cons tructivistic GTS) is presupposed.
This observation implies that what was said in the preceding
chapter about the reducibility of most of classical mathematics to IF
first-order logic (and what was said about the sense in which this
reducibility holds) remains valid in my constructivistic reinterpreta-
tion of logic and mathematics.
For instance, the same line of thought as before shows that for
each higher-order statement S0 we can correlate a statement. The
question whether a certain other higher-order statement S1 logically
follows from S0 is then equivalent to a question concerning the
logical truth of a statement. And as was seen in the preceding
section, this question is tantamount to the question whether a cer-
tain IF first-order statement is logically true. All of this remains valid
even if all the higher-order variables are restricted to recursive
values.
Thus on the constructivistic interpretation, too, most of ordinary
mathematics can be seen to be combinatorial in the sense explained
in the preceding chapter.
What is not preserved in the transition to the constructivistic
interpretation is the truth predicate discussed in Chapter 6. When
one tries to reconstruct this predicate, much of the construction goes
through without any problems. For instance, we can give a charac-
terization of what is required of a one-place predicate X (of the
Gödel numbers of formulas) to be a truth predicate in the same way
as before. But in order for this truth definition to work, the truth
predicate must be recursive, for otherwise it would not be a value of
the existential quantifier (3X). Even though the matter needs further
study, there does not seem to be any reason to expect that a construc-
tivistic truth predicate is recursive any more than a classical truth
predicate.
If so, interesting suggestions offer themselves here. Maybe the idea
of the ineffability of truth has another source besides a belief in the
universality of language, namely, a belief in constructivism. This
226 THE PRINCIPLES OF MATHEMATICS REVISITED

would for instance provide the later Wittgenstein a specific reason to

believe in the ineffability of semantics.
These remarks barely scratch the surface of the complex of
questions concerning the nature of the newly reconstructed con-
structivistic mathematics — that is, a mathematics using my con-
structivistic logic as its basis. Further work is needed in this
direction. Meanwhile, we have to face squarely the basic question: Is
constructivism correct, at least in its reconstructed form? What
makes this question poignant here is that at the beginning of this
chapter! gave an argument in favor of constructivism. Moreover, it
is obviously a telling argument. It can be summed up in the rhetori-
cal question: Can it possibly make sense to play a game with
a nonrecursive strategy? The only reasonable answer that any man
(and as Dr. Johnson might have added here, any woman or any
child) can give seems to be that it does not make any sense whatso-
ever. How can I possibly say that I am playing a game with a fixed
strategy when I do not have (even in principle) any way of actually
calculating what my next move will be?
In spite of the apparent force of these persuasive questions, they
do not close the issue. We are here obviously dealing with problems
concerning the basic concepts of game theory. And of those basic
concepts the most basic one is the notion of strategy. It would not be
much of an exaggeration to say that the mathematical theory of
games was born the moment John von Neumann (1928) — or was it
Borel? — conceived of the current abstract notion of strategy and
began to use it to reduce games to their normal form.
But the seeds of dissension lie in this very notion of strategy. On
a closer look, it is seen to involve a sweeping and possibly unrealistic
abstraction. This abstraction is codified in the notion of strategy
employed routinely by game theorists. This concept is used by them
to reduce an entire play of a game to a choice of strategy by each of
the players. And that certainly looks like an overabstraction if ever
there was one. In chess, a grandmaster may perhaps choose his or her
strategy (selection of moves) for the first six, eight, ten or twelve
moves, as witnessed by the chess clock. In a World Championship
match, the first several moves of a game are sometimes made
virtually instantaneously, thus revealing an antecedent partial strat-
egy choice. But after that, even grandmasters have to start creating
their strategies "across the board", as the saying goes. In doing so,
the player in question uses his or her partial knowledge of the
 CONSTRUCTIVISM RECONSTRUCTED 227

opponent's strategy, as it is revealed in the play so far. This can (and

should) be generalized to all games of strategy. In all but the very
simplest ones, players do not choose their strategies once and for all,
but create their strategies in the course of a play of the game in
question. This strategy is properly so-called in a very real sense, for it
typically involves considerations concerning what my opponent
would do if! made a certain move, what I could do in response, and
so on. (Of course, the strategy in question is according to the strict
game-theoretical definition at best only a partial one.) But if a
player's strategy is created in this way, there is no good reason to
assume that the resulting eventual strategy is recursive. There even
exist probabilistic criteria which tell you, given an initial segment of
the values of an integral-valued function, how likely it is to be
a nonrecursive one. Applied to the strategy function of a player of
a semantical game, such criteria might even justify an empirical
finding (albeit only a probabilistic one) to the effect that a player is
employing a nonrecursive strategy in a play of a semantical game.
Ironically, the line of thought proposed in this chapter on behalf of
constructivists can be turned into an argument against them. It was
just seen that there is nothing incoherent about a human being
playing a game without an antecedently decided strategy, recursive
or not. This is just like a chessplayer creating his or her strategy
across the board. What one cannot do is to program an automaton
to play a game without a recursive strategy. Even though many
constructivists emphasize the role of human thinking and human
constructions in the "game" of mathematics, the main argument on
their behalf virtually amounts to arguing that since robots have to be
constructivists, we humans, too, have to be constructivists. For the
only strategies that a digital automaton can be programmed to play
in accordance with are the recursive ones. This is precisely what the
proconstructivist argument sketched above has alleged to be the
inevitable human predicament. The constructivists are in effect
imputing to human beings those very limitations that characterize
digital computers. It is thus the classical mathematicians who have
more faith in human creativity than constructivists, contrary to the
occasional claims of the latter.
It is thus difficult to decide whether or not the constructivistic
limitation of verificatory strategies to recursive ones is sound or not.
What has been seen nevertheless shows that the constructivistic
answer cannot be accepted simply on the basis of its greater realism,
228 THE PRINCIPLES OF MATHEMATICS REVISITED

as far as the actual playing of a semantical game is concerned. On the

contrary, the restriction to recursive strategies arguably overlooks
the entire realistic possibility of an across-the-board construction of
strategies in semantical games. However, further thought is perhaps
needed here.
My reconstruction of constructivism brings to the open another
subtle problem. The requirement that constructivistic language
games must in principle be actually implementable can be turned
against the constructivists. For this purpose I only need to look at
semantical games from the vantage point of a referee or perhaps
a Quinean jungle linguist, rather than that of a player. How can a
referee decide on the only basis that he or she has, to wit, on the basis
of a player's behavior, whether the player in question is breaking the
rules of the game by using a nonrecursive strategy? In a concrete
implementation of a semantical game, such a determination must be
made on the basis of a finite number of moves. But any finite
sequence of moves is compatible with a recursive strategy. Hence
a restriction of the use of recursive strategies seems unenforceable.
This problem might at first seem like mere quibbling. In practice,
as was noted earlier, it is by no means impossible to draw at least-
probabilistic inferences from a player's moves to the strategy that he
or she is using. In Las Vegas casinos, pit bosses can spot a "counter",
that is, a player who uses a system at blackjack tables. A theoretical
basis for such inferences is provided by the studies of Martin—Löf
(1966, 1970) and others of the statistical properties of recursive
sequences of numbers. Such statistical properties naturally manifest
themselves in finite sequences of events (see here Fine 1973, Ch. 5 and
Schnorr 1971).
But the comparable implementation of the recursivity restriction
in semantical games can only be probabilistic, and not strict. This
introduces a new and unfamiliar complication in that the outcome of
a play of the game will depend on how the rules of the game are
enforced. Such games may or may not be a territory on which game
theorists fear to tread, but so far they do not seem to have done so in
a way that would help us here with our present problems. Hence the
conceptual problems involved in the implementation of a recursivity
restriction do cast a shadow on constructivistic philosophies of logic
and mathematics.
There is another kind of reason against restricting Skolem func-
tions to recursive ones. Such a restriction would make it extremely
 CONSTRUCTIVISM RECONSTRUCTED 229

hard to develop a mathematics that could serve the purposes of

physical theory. It has been shown that certain perfectly determinis-
tic systems in classical mechanics exhibit nonrecursive behavior (see
Ekeland 1988, pp. 59—6 1). Such a behavior would be impossible or at
the very least extremely awkward to handle in mathematics where
functions are restricted to computable ones.
It is in any case important to realize what precisely is involved in
the transition from a classical first-order logic (IF or not) to the
corresponding constructivistic logic. In a sense, both logics are
arguably constructivistic. Both kinds of logic are grounded on
certain "language games". Moreover, these language games are the
same on the classical and on the constructivistic side of the fence, in
the sense that precisely the same moves are admissible in the two
cases and that winning and losing are defined in the same way.
Hence, pace many constructivists, if constructivistic logic is teach-
able, learnable, recognizable from peoples' behavior, and actually
practicable by human beings, then so is the corresponding classical
logic. What makes the difference is the use we make of the language
games in question for the purpose of defining a notion like truth.
Even though the constructivistic logic developed here is an emi-
nently natural and interesting way of implementing constructivistic
ideas, it does not match everything that intuitionists and other
constructivists say about their logic. However, it is not obvious what
such discrepancies prove. In some cases at least, they serve to
illustrate the confusions of the constructivists. For instance, it is
clearly in accordance with the intentions of the constructivists that
more is required of the constructivistic truth of a mathematical
theorem than of its classical truth. This idea is implemented by the
constructivists discussed here. Yet the notion of constructive truth so
implemented is not in agreement with the fine print of all of the
so-called constructivistic theories of logic and mathematics.
For instance, when the constructivists' aims are realized along the
lines proposed here, we cannot eliminate the undecidability of
mathematical problems in the way that some constructivists wanted
to do. However, to my mind this result only goes to show that the
constructivists' aims were incompletely analyzed. Decidability is
closely related to deductive completeness, and such completeness
is shown by Gödel's results (and by the role of IF first-order logic
as the basic logic of mathematics) to be a pie in the sky. What
remains of the constructivists' motivation is therefore an emphasis
230 THE PRINCIPLES OF MATHEMATICS REVISITED

on the epistemic element in mathematical theories. But this idea can,

and should, be implemented in an entirely different way. For one
thing, the epistemic element should be dealt with above the board
and not merely tacitly. In the next chapter, I will show how this can
be done in a simple and elegant way.
In sum, the claims of the constructivists to be able to give us
a satisfactory general perspective on logic and mathematics seem to
me seriously flawed. However, if we lower our sights somewhat, the
situation can be seen to change. If we restrict our attention to the
deductive function of logic in mathematics, then constructivistic
ideas will suddenly be seen to fill a genuine need.
It was noted at the end of Chapter 9 that the game-theoretical
approach apparently has little to contribute to, or even to say about,
the deductive function of logic. This admission might have struck the
reader as being premature. Indeed, the game-theoretical conception
of truth might prima facie seem to show precisely what there is
to be done for the purpose of facilitating the deductive task of logic.
What is needed is an analysis of what it takes to infer the truth of
a sentence, say of
u] (10.6)

from the truth of another one, say of

(klx)(By)S1 [x, y] (10.7)

For simplicity, I will in this example disregard the role of all

quantifiers and connectives in S1 and
The truth of (10.7) means that there is a winning strategy for the
initial verifier in the correlated game. If the simplifying assumptions
just indicated are made, this means the existence of a function f such
that
(Vx)S1 [x, f(x)] (10.8)
Similarly, the truth of (10.6) means that there exists a function g such
that
(Vz)S2[z, g(z)] (10.9)
We can therefore get from the truth of (10.6) to the truth of (10.7) by
means of a function which takes us from a winning strategy in the
game correlated with (10.7), that is, from the function f, to a winning
strategy in the game correlated with (10.6), that is, to the function g.
 CONSTRUCT! VISM RECONSTRUCTED 231

This means obviously the existence of a functional that yields g as

a function off. This can be expressed by the third-order sentence
[(Vx)S1 [x, f(x)] zJ (10.10)

More generally, we can in our metalogical notation associate with

each formula two metalogical variables — one for the strategy func-
tion of each of the two players. If the strategy functions of the initial
verifier and the initial falsifier associated with the premise (say A) are
and and those associated with the conclusion (say B) are 4) and
we can then say in general that the inference-ticket conditional that
mediates the passage from the premise to the conclusion is
i/i)] (10.11)

In ordinary higher-order logic this would be equivalent to

'i) vi)] (10.12)

and hence to
i/i)) vi)] (10.13)

This, then, might seem to be the kind of conditional that could serve
as an inference-ticket for the purposes of logical deduction. Indeed,
(10.13) is formally identical with Gödel's interpretation of condi-
tionals.
Unfortunately, in ordinary higher-order logic, (10.13) can be seen
to reduce back to ordinary truth-functional conditional, and hence
not to give us anything new.
The reasons for this failure of the analysis (10.13) to yield anything
useful for the deductive function of logic go back to the very
definition of truth as the existence of a winning strategy for the initial
verifier in a semantical game. As long as existence is taken in a
standard sense, it does not follow that the verifier "has" a winning
strategy in the sense of knowing what such a strategy is — or even
being in a position to know what it is. In deduction, we are moving
from the known truth of the premise or premises to the known truth
of the conclusion. In order to handle such matters, it is mandatory at
least to restrict the initial verifier's winning strategies to those that he
or she can know. And this apparently entails that these strategies can
be represented by recursive (computable) functions.
This means using the constructivistic logic outlined earlier in this
chapter. In general, while it was not possible to find convincing
232 THE PRINCIPLES OF MATHEMATICS REVISITED

reasons for adopting a constructivistic standpoint in general in the

foundations of logic and mathematics, there seems to be a great deal
to be said for adopting it for the purposes of the deductive function of
logic in mathematics.
In the particular matterjust discussed, we can use our constructiv-
istic logic to get a nontrivial analysis of the conditionals which can
operate as a bridge from the known truth of a sentence (in the sense
that a winning strategy in the corresponding game is known) to the
known truth of another sentence. In the case of a specific sentence
like (10.13) expressing the conditional which can bridge such inferen-
ces, this constructivistic logic amounts to restricting the functions
and functionals in it to recursive ones.
I will not pursue the technical possibilities that are opened up
here. However, it is of great interest to examine the general theoreti-
cal perspectives that are revealed by my observations.
First, in this way constructivistic ideas can be seen to play a major
legitimate role in a mathematician's work in any case. They are
needed to understand and to master the deductive task of logic in
mathematics. At the same time, this role is different from what
constructivistic philosophers like to think. For one thing, the role of
constructivistic notions has nothing to do within the meaning of
mathematical statements. Meaning is a matter of the descriptive
function of logic in mathematics. It is a matter of sentence—model
relationships, ultimately a matter of truth definitions. It is not a
matter of deductive relationships between propositions.
Second, the analysis of conditionals (10.10), combined with the
constructivistic logic outlined above, can be thought of as yielding
Gödel's so-called functional interpretation of first-order logic. The
line of thought just indicated provides a new theoretical justification
for Gödel's interpretation and at the same time shows its roots in the
game-theoretical approach to logic.
It may be noted that the Gödelian conditional (10.10) is equiva-
lent to the metatheoretical sentence
B(4, i/i)] (10.14)

which prima fade involves no step to a higher type in the type

hierarchy. But if (10.10) is itself to occur in a conditional, it has to be
brought to a form where existential quantifiers precede universal
ones. This form is (10.13), which shows that this construal of iterated
conditionals leads to higher and higher types. Such iterated condi-
 CONSTRUCT! VISM RECONSTRUCTED 233

tionals are arguably indispensable in any reasonable treatment of

the deductive function of logic, and the ascent in type hierarchy
appears to be unavoidable.
This result confirms the suspicions voiced at the end of Chapter 9.
The reduction of mathematical theories to the first-order level which
was outlined in Chapter 9 applies only to the descriptive function of
logic in mathematics. In contrast, an attempt to serve increasingly
better the deductive task of logic will lead us in the opposite direc-
tion, that is, to higher and higher types, just as in Gödel's functional
interpretation.
We can thus see that the distinction made in Chapter 1 between
the two main functions of logic in mathematics does in fact matter.
The descriptive and the deductive functions are most naturally
served by different conceptualizations. It might indeed be salutary
for philosophers of mathematics to keep this distinction more firmly
in mind than what is customary these days.
The schizophrenic relation of logic to the type hierarchy
which has been found is highly interesting philosophically and
historically. It is perhaps not too far-fetched to see in Gödel's proof
of the deductive incompleteness of elementary arithmetic an indica-
tion of the need of having to climb higher and higher in the type
hierarchy for deductive purposes. This point can be generalized
in an interesting way. Mathematics is often thought of as a science
of abstract objects and abstract structures. This abstractness
is presumably greater the higher we climb in the type (order)
hierarchy. Godel for one entertained a concept of abstraction
to which this relationship to type hierarchy can be attributed.
In contrast, logic tends to be thought of, at least ideally, as a ma-
nipulation of concrete symbols. Such an ideal of logic was represent-
ed by the Vienna positivists' idea of the logical syntax, and by
Hubert.
What we have found turns this relationship between mathematics
and logic upside down. For the primary descriptive function of logic
in mathematics, all that we need conceptually are combinatorial
structures consisting of particulars belonging to the rock bottom of
the type (order) hierarchy. In contrast, the deductive treatment of
such mathematical structures forces us to consider increasingly more
and more abstract (higher-type) entities. It may thus be said that the
ontology of mathematics can be restricted to the world of combina-
torics and of particular nominalistic entities, whereas the deductive
234 THE PRINCIPLES OF MATHEMATICS REVISITED

technology of mathematics forces us to higher and higher levels of

abstraction.
Even though these results have to be handled with considerable
caution, they suffice to show how thoroughly misleading, not to say
mistaken, our received ideas about logic, mathematics and their
interrelations are.
Some of these comments can be given a more personal address.
For instance, Gödel not only maintained a Platonistic ontology of
mathematics but thought that an ascent to higher and higher ab-
straction was a way of solving foundational problems. If! am right,
the need of abstractive ascent is due only to the needs of actual
deductive inferences; Platonistic theorizing, which Gödel advocated
in some of his pronouncements, does not require such an ascent.
Also, it is ironic in a historical perspective to see that Hilbert's first
and foremost interest was in the descriptive function of logic. This is
in evidence not only in his work in the foundations of geometry but
perhaps even more clearly in his general remarks on the axiomatic
method. The function of a mathematical theory is to capture a struc-
ture or class of structures by completely logical means. The develop-
ment of metamathematics and proof theory which was prompted by
Hubert's foundational project was a mere by-product of his overall
enterprise. In a sense that appears from what has been said, proof
theory is an enterprise in a direction quite different from filbert's
central interests.
 11

The Epistemology of Mathematical Objects

At this point, it behooves me to reflect on what I am doing and to

generalize the questions that I have raised. It was indicated above
that the constructivistic interpretation of logic and mathematics
outlined in the preceding chapter is not without precedents. The
philosophical motivation of the constructivistic interpretation dis-
cussed there is new, or at least much more fully articulated than the
expressed earlier motivations of similar views. But the technical
implementation of my interpretation does not stray very far from
Gödel's Dialectica interpretation (1958) of first-order elementary
arithmetic or from Kleene's realizability interpretation (see Kleene
1952, sec. 82).
However, it is possible to generalize the entire constructivistic
interpretation considered here in a way which is more radically new
and which opens the door to a deeper motivation of constructivism.
This motivation can be seen by criticizing the way I have tried
experimentally to present a raison d'être for the one particular
version of constructivism in the preceding chapter. This rationale
turned on the alleged impossibility of playing a semantical game in
accordance with a nonrecursive strategy. It was found unconvincing.
A working mathematician might very well resent the tacit limita-
tions that my tentative argument imposes on a mathematician's
ability to play a semantical game. Surely a competent mathema-
tician is not restricted in his or her mastery of strategies to such simple
ones as to be codifiable by recursive functions — that is, to functions
that even a stupid machine can handle. Even if a mathematician
cannot program a computer to calculate a function for all argument
235
236 THE PRINCIPLES OF MATHEMATICS REVISITED

values, he or she can do all sorts of other things with it. The value of
this strategy function can perhaps be calculated for most (or perhaps
almost all) of the argument values. The qualitative behavior of the
strategy function can be mastered even if it is not computable by
a digital automaton. It can even in principle be calculable by an
analogical computing device in contradistinction to a digital one. In
brief, a nonrecursive function can be mastered intellectually in a way
that would authorize us to say that a semantical game is played in
accordance with a strategy codified by this function.
This line of thought is clearly related to what was said in Chap-
ter 10 about the possibility of creating a strategy "across the board".
It can be used to evade the criticisms that were rehearsed there
against such "classical" GTS as dispenses with the recursivity re-
quirement.
But although this line of argument tells against a restriction of
strategy functions to recursive ones, it is prima facie compatible with
some restrictions on those functions. In fact, it is easily turned into
a criticism of the classical interpretation of logic and mathematics
along the same lines as were followed above, but imposing a weaker
restriction on the strategy (Skolem) functions relied on in a game-
theoretical approach to the foundations of mathematics. The argu-
ment would then be that the verifier can meaningfully be said to play
a semantical game in accordance with a strategy only if this strategy
is codifiable in a function which can in principle be mastered
mathematically. Surely the very idea of strategy as being used by
a human being implies a conceptual control of the successive moves
made in accordance with that strategy.
But this argument, even though eminently plausible, now seems to
be so vague as to be useless. It is not that one cannot think of what
might be involved in mastering a function mathematically. The
trouble is that one can think of a large number of different ways of
understanding this idea of mastery. What I will do here is to focus on
what seems to be the greatest common denominator of these differ-
ent ways and to interpret one's mastery of a function to mean that
this function is a known one more accurately speaking, to mean
that one knows which function it is. At first sight, this might only seem
to introduce afaçon de parler; for where is the real link between this
location and our actual notion of knowledge? Even though philos-
ophers have analyzed the notion of knowledge intensively (or at least
extensively), it still does not seem to be sharp enough to be useful here.
 EPISTEMOLOGY OF MATHEMATICAL OBJECTS 237

A look at the historical situation may provide some clues here. So

far, I have not sought to make any distinction between construe-
tivism and intuitionism. This might seem comme ilfaut in view of the
current usage in which all and sundry constructivistic tendencies are
called "intuitionistic". This usage nevertheless fails to do justice to
the specific views of the real intuitionists who are following or, at
least, are inspired by Brouwer. His views are not necessarily done
justice to by calling them constructivistic. Typical constructivistics
pay attention in the first place to logical and mathematical behav-
ior — that is, to what can be done in logic and mathematics. The
question of the possibility of actually playing semantical games with
arbitrary strategies and the question of how different strategies are
betrayed by a player's behavior are typical constructivistic ques-
tions. In contrast, Brouwer is inquiring into a mathematician's
thought rather than into his or her behavior, and he is raising
questions about what we know and can know rather than questions
as to what we can do. This is connected with the fact that the true
intuitionists have not been happy with the development of the
notion of recursivity (Turing machine computability or equivalent)
as an explanation of the kind of knowability they have in mind. Their
reasoning is revealing. It is not enough that a set of equations in fact
enables us to carry out the calculation needed to determine the value
of a function for each argument value. We must be able to know that
it does so.
More generally, no reader of Brouwer's writings (1975) can fail to
be impressed by the role of epistemic ideas in his argumentation. An
especially telling case in point is Brouwer's characteristic and cru-
cially important technique of "weak counterexamples" (see Brouwer
1908c in Brouwer 1975; Troelstra and van Dalen 1988, pp. 8—16; van
Stigt 1990, pp. 252—255). Such a counterexample to a logical or
mathematical principle does not show that the principle is false in
any usual sense. It shows the unacceptability of the principle by
showing that its acceptance implies that "we ought to have certain
knowledge. . which in fact we do not possess" (Troelstra and van
.

Dalen 1988, p. 11). This, then, seems to be the dividing line between
constructivists and true intuitionists: intuitionists emphasize the role
of knowledge in mathematics while constructivists put a premium
on the effectiveness of constructions and other operations. Construe-
tivists are do-it-yourself mathematicians, intuitionists are know-it-
yourself ones. In view of this contrast, it is not surprising that it has
238 THE PRINCIPLES OF MATHEMATICS REVISITED

been possible to interpret intuitionistic logic in epistemic logic. This

result has to be handled with great care, however, for it will be shown
later in this chapter that neither one of these two logics is adequate in
its present (axiomatic) shape.
In my considered view, intuitionists have done foundational
discussion a monumental disservice in not bringing out into the
open the epistemic element, which according to their own views
there is in mathematical theorems. I also believe that they have done
their own ideas a disservice by not so doing. They apparently
thought that there inevitably is such an epistemic element, in effect
an implied claim to knowledge, in all mathematical statements. If so,
their tacit epistemic claims would be commensurable with the
propositions of classical mathematicians only if they, too, have been
making tacit epistemic claims. However, I fail to find any strong
arguments for such commensurability in the writings of the genuine
intuitionists. It is perhaps possible to try to view some of Brouwer's
pronouncements as steps in this direction. If so, they are scarcely
explicit enough to be discussed and evaluated here.
It is not too far-fetched to see in the intuitionists' failure to
acknowledge the epistemic dimension of their thinking a confusion
between the different language games distinguished from each other
in Chapter 2. What gives mathematical statements their meaning are
the semantical games on which the truth definition used here is
based. However, it is obvious that intuitionists like Brouwer have
thought of mathematical activity as a process of actually coming to
know mathematical truths and mathematical objects. In Chapter 2
it was seen that these two kinds of activities ("language games")
have to be sharply distinguished from each other in the interest of
conceptual clarity.
This criticism puts an onus on me to spell out precisely what this
epistemic dimension of mathematics alleged by the intuitionists is,
and how it can be implemented. At first sight, this might seem
a pointless undertaking. As an example, we might ask how the
epistemic element might be expressed by means of received epistemic
logic. There you find only one epistemic ingredient. It is the "knows
that" or "it is known that" operator K. If the knower is to be
indicated, a subscript can serve the purpose, so that KaS is to be read
approximately

a knows that S
 EPISTEMOLOGY OF MATHEMATICAL OBJECTS 239

But such an epistemic ingredient does not seem to help us at all. For
in the usual treatments of epistemic logic KaS is logically true (valid)
if and only if the nonepistemic sentence S is logically true. Hence the
introduction of the epistemic element does not seem to serve any
purpose.
This mildly paradoxical-looking result is not merely a matter of
the deductive or inferential relationships between different state-
ments. It is firmly rooted in the model theory of epistemic logic. The
one safe conceptual anchor one has there is the idea that knowing
something, say, knowing that S, means being able to rule out
legitimately all the states of affairs, courses of events and other
scenarios in which S fails to be true. No matter what questions can be
raised about the legitimacy of such exclusion of scenarios (the
philosophers' ill-named "possible worlds"), the resulting epistemic
logic remains unaffected.
At this point the aficionados of epistemic logic will undoubtedly
expect me to evoke the notorious paradox of logical omniscience.
This so-called paradox is a simple corollary to the model-theoretical-
conception of knowledge just outlined. As a moment's reflection
shows, this conception implies, so it seems, that whoever knows
something knows all the logical consequences of what he or she
knows.
This diagnosis of the so-called paradox immediately reveals a
cure. This cure lies in the study of the processes through which one in
fact comes to know the logical consequences of one's premises. This
prescription is not hard to use. It can even be handled model-
theoretically, as has been shown by the work represented in Rantala
(1975) and Hintikka (1975).
This is nevertheless not an occasion to follow up such a line of
thought. The main reason is that it is not what intuitionists have
had in mind. In the paradox of logical omniscience, in the simplest
cases at least, we are dealing with the fact that the application of
logical rules of inference does not happen in one fell swoop, but is
a matter of step-by-step applications. However, what the intuition-
ists are doing is not a study of such stepwise procedures. Rather, they
propose to change the very rules of logical inference. This is an
approach that is entirely different from the ideas of Rantala and
Hintikka who focus on the fine structure of inference rule applica-
tions. It has to be conceptualized and otherwise dealt with in some
other manner.
240 THE PRINCIPLES OF MATHEMATICS REVISITED

But how? If the intuitionists were not dealing with knowing that,
then what was the epistemic element they were implicitly importing
into mathematics?
Here the developments which are made possible by the notion of
informational independence and which were outlined at the end of
Chapter 4 play a crucial role. What was seen there is how to imple-
ment an idea which the received epistemic logic does not capture.
This is the idea of knowledge of entities (objects, things) as distin-
guished from knowledge of facts (propositions, truth of sentences).
The latter kind of knowledge is relative to a space of scenarios
("possible world") on which an alternativeness ("accessibility") rela-
tion is defined. The former kind of knowledge is relative to a set of
"world lines" of cross-identification which define which denizens of
the different possible worlds count as identical (manifestations of the
same entity). As was emphasized in Chapter 4, the way these "world
lines" are drawn is largely independent of the truth-conditions of
knowing that statements.
This enables me to put forward an interpretation or rather a
rational reconstruction — of the main thrust of the intuitionists'
thinking. They were indeed tacitly introducing an epistemic element
into mathematics. But this epistemic element is not a knowledge of
truths, but a knowledge of entities. It should not be left implicit,
either, as intuitionists were doing. It can and should be imple-
mented along the lines of the epistemic logic sketched in Chapter 4.
This new logic is made possible by the main new conceptual tool
introduced in this book, namely, the notion of informational inde-
pendence.
Before developing this approach systematically, it is in order to
recall how it hangs together with the line of thought initiated in the
preceding chapter. There the idea was entertained of restricting
strategy functions to recursive functions. It was argued that this is far
too restrictive a limitation. Instead, it was suggested that the strategy
functions should be restricted to known functions. This is the restric-
tion which has meanwhile been found to be in agreement with the in-
tuitions of intuitionists. It is the idea that is being implemented here.
Using the notation and the observations explained in Chapter 4,
we can express what it means for a function f to be known. The
following are the equivalent formulations of this idea:

K(Vx)(Jy/K)(f(x) = y) (11.1)
 EPISTEMOLOGY OF MATHEMATICAL OBJECTS 241

K(Rg/K)(Vx)(f(x) = g(x)) (11.2)

= g(x)) (11.3)

(11.5)
Here K is interpreted in a game-theoretical context as a universal
quantifier ranging over epistemic alternatives to the world which the
players are considering at the time: the falsifier chooses such an
alternative, by reference to which the game is continued.
This suggestion does not yield a unique interpretation of what it
means to master a function intellectually, that is, to "know which
function it is". On the contrary, the criteria for knowing a function
are left almost completely open. Indeed, it is one of the fundamental
insights prompted by the semantics of epistemic logic that the
truth-criteria of knowing + wh-constructions are largely independent
of those of knowing that statements. In the independence-friendly
notation, this means that the truth-conditions of K-statements
containing slashed existential quantifiers (ax/K) are left
underdetermined by the truth-conditions of unslashed K- statements.
The underlying semantical reason is that they depend on criteria of
identification across models (worlds). In contrast, slashed disjunc-
tions (v /K) depend on identification conditions of functions only
indirectly by being under the scope of universal quantifiers.
This inevitable underdetermination opens the door to several
interesting lines of thought. It shows that the basic ideas of intuition-
ists are open to different specific interpretations, depending on which
criteria are applied to our knowledge of mathematical objects. It
looks hopeless to try to find a consensus as to which interpretation is
the right one. However, this does not spoil the usefulness of the
epistemic interpretation schema for the purpose of conceptual
clarification. It has already been seen that one can use this epistemic
interpretation schema to discuss a number of different intuitionistic
and constructivistic ideas. What is more, one can even establish
general logical results which are independent of what counts as
a function being known.
In order to reach such results, I have to be a little more explicit
than I have been so far. I will generalize the intuitionists' way of
looking at mathematical theories and theorems into something that
might be called an epistemic (or intuitionistic) interpretation of
mathematics and logic. Starting from the simplest case, what is the
242 THE PRINCIPLES OF MATHEMATICS REVISITED

epistemic interpretation of a first-order mathematical statement S?

For simplicity, I will assume that S is in a negation normal form.
Then we can reinterpret S as "really" meaning
KS* (11.6)
where S* is formed from S by replacing every existential quantifier
(ax) by (2x/K) (and likewise for higher-order quantifiers if they are
used) and every disjunction v by ( v /K). This amounts to requiring
that all the functions which operationalize existential statements
(Skolem functions) are known functions.
This new interpretation can naturally be extended to IF first-
order logic and to mathematical theories and statements that can be
expressed by its means. I will call it the epistemic interpretation of
logic and of mathematical theories. It is important to realize what is
new in it. In the earlier literature on the subject, when epistemic
considerations were brought to bear on the foundations, the ques-
tion explicitly or implicitly raised was: What does it mean to know
a mathematical truth? Moreover, this question was discussed as if it
merely involved knowing that a certain mathematical proposition
holds. What has come up here is an altogether different dimension of
the epistemology of mathematics. We are now dealing with the
question of our knowledge of mathematical objects such as functions.
One remarkable but largely overlooked fact here is that these two
epistemic dimensions are largely independent of each other. Yet,
there is plenty of evidence that in the earlier literature the two kinds
of epistemic issues were confused with each other. A case in point is
offered by the usual criticisms of the axiom of choice discussed
below.
The epistemic interpretation of mathematics offered here is cal-
culated to incorporate both questions of our knowledge of math-
ematical truths and questions of our knowledge of mathematical
objects. However, the precise criteria of the latter kind of knowledge
are left open by the epistemic interpretation, which therefore is an
interpretational schema rather than a unique interpretation.
Thus a rough-and-ready distinction can be made between con-
structivistic and intuitionistic approaches to the foundations of
mathematics. The former emphasize the role of what a mathema-
tician can in fact do, while the latter emphasize what a mathema-
tician can know. The relationship between the two is a delicate
matter, so much so that it is not even clear as to which is the more
 EPISTEMOLOGY OF MATHEMATICAL OBJECTS 243

general view. On the one hand, the constructivistic game-theoretical

semantics for mathematics characterized in the preceding chapter
can be thought of as a special kind of the epistemic interpretation. It
is obtained by stipulating that all and only recursive functions are
"really known". A part of the interest of the constructivistic interpre-
tation is thus due to the fact that it exemplifies the epistemic
interpretation of mathematics.
In this perspective, the epistemic interpretation of logic and
mathematics is much broader than the constructivistic one defined
in the preceding chapter. For instance, it is quite clear that the hard-
core intuitionists rejected the idea that recursivity could serve as the
explication of constructivity. Their views therefore fall within the
purview of the epistemic interpretation, but not of the constructivis-
tic one. More generally speaking, one important difference between
classical and intuitionistic mathematicians is that the former are
satisfied with knowing mathematical truths while the latter also want
to know mathematical objects.
On the other hand, the emphasis of Brouwerian intuitionists on
the epistemic element in mathematics is seen to lead inevitably to an
emphasis on our knowledge of mathematical objects. This emphasis
is quite different from an emphasis on our actual ways of coming to
know mathematical truths, and in a sense is much more specific.
The epistemic interpretation of mathematics offers a new and
powerful tool for future studies of the epistemology of mathematics.
A few examples may illustrate its relevance. By means of epistemic
logic, the epistemic interpretation can be illuminated by the same
means as ordinary logic, namely, through second-order translations.
For instance, on the epistemic interpretation
y] (11.7)

becomes (if we look away from dependent disjunctions) a shorthand

for
y] (11.8)

which is equivalent to
f(x)] (11.9)

and to
f(x)] (11.10)
244 THE PRINCIPLES OF MATHEMATICS REVISITED

But (11.10) is the "translation" (interpretation) of

(3f)(Vx)S[x, f(x)] (11.11)
Simple though these observations are, they have interesting
consequences. What the equivalence of (11.8) and (11.10) shows is
that the epistemic interpretation validates the axiom of choice, no
matter how a function being known is understood, as long as the
epistemic interpretation is applied consistently to all quantifiers.
What the critics of the axiom of choice have in effect done, looked at
from the vantage point of the epistemic interpretation, is to confuse
with each other
(11.12)
and
K(Vx)(3y/K)S[x, y] (11.8)
which is equivalent to
f(x)] (11.10)

In other words, the critics of the axiom of choice have tacitly taken
the problem of whether one knows the mathematical statement
y] to involve also the question as to whether the choice
function f (i.e., a certain mathematical object) is known or not. This
exemplifies the general confusion mentioned earlier.
These observations throw a great deal of light on the philosophi-
cal stance of different mathematicians and philosophers. We can for
instance see that it is not constructivism but intuitionism, with its
emphasis on our knowledge of mathematical objects, that can lead
to a rejection of the axiom of choice. Furthermore, from my epi-
stemic vantage point it is highly instructive to see how a prominent
defender of the axiom of choice rests his case on the idea that the
existence of a mathematical object "is a fact like any other"
(Hadamard, quoted in Moore 1982, p. 317; emphasis added). The
whole claim of the intuitionists is that we are, in assumptions like the
axiom of choice, not dealing with knowledge of facts but with
knowledge of mathematical objects.
For another instance, Dummett's philosophical position can, in
the light of what has been found, seem to be constructivistic rather
than intuitionistic according to the distinction adumbrated in this
chapter. This is shown not only by Dummett's acceptance of the
axiom of choice (see the preceding chapter) but also by his insistence
 EPISTEMOLOGY OF MATHEMATICAL OBJECTS 245

that there is nothing special about the semantics of dependent

quantifiers like (]y) in
(Vx)(Jy)S[x, y] (11.7)
beyond the general requirement of constructivistic justifiability or in
Dummett's term, assertability (Dummett 1977, p. 451; cf. Troelstra
1977, pp. 154—156).
Thus the epistemic interpretation of mathematics enables us to
tell precisely what is valid and what is invalid in the axiom of choice.
Likewise, we can see by means of the epistemic interpretation what is
acceptable and what is not in another bugbear of the intuitionists,
namely, the method of indirect proofs. At first sight, the critics of
indirect proofs have once again got hold of the wrong end of a sticky
concept. For suppose that a logician has taken the negation S of
a statement and showed that it leads to a contradiction, in other
words, shown the following:
(11.13)
But(1 1.13) is equivalent to 5, so that the indirect prover needs only to
add "q.e.d." And it does not matter whether or not we prefix K to
(11.13).
However, if in KS there are slashed expressions (Jx/K) or ( v /K),
the negation of S that figures in an indirect proof cannot simply be
taken to be S or 5, for neither of these makes any sense standing
alone. The only negation that can be relevant here is (SI) where (SI)
is obtained from S by omitting all independence indications (i.e., all
slashes with their right-hand accompaniments). Then an indirect
proof involves an inference from (SI) to S. Such inferences are
not always valid.
For instance, take the epistemic statement
y] (11.8)
The only negation that can be drafted into the service of an indirect
proof is here
(Bx)(Yy)'—'S[x,y] (11.14)
If this is reduced to impossibility, then what has been reached is
K S[x,y] (11.15)
which is equivalent to
K(Vx)(ly)S[x, y] (11.12)
246 THE PRINCIPLES OF MATHEMATICS REVISITED

But (11.12) is weaker than (11.8). What is known according to the

former is that everything, say x, is related to something, say y, such
that S[x,y]. What is known according to the latter is which indivi-
dual y it is that each given x bears this relation to.
One can perhaps say that indirect proofs can in this way of
looking at them yield knowledge that certain mathematical proposi-
tions are true, but without knowledge as to why they are true, in the
sense that the proof does not show what the Skolem functions are
that provide the true-making substitution-values of the quantified
propositions in question. The naturalness of the epistemic interpre-
tation of mathematics is illustrated by the fact that formulations of
the kind just used have time-honored precedents in the foundational
discussion. For instance, one classical formulation of what proofs by
contradiction do runs as follows: "Even if they do not give us the
cause of why a certain affection is to be predicated of a subject, they
nonetheless give us a reason by which we know that a certain state of
affairs holds" (Mancosu 1991, p. 34).
Once again, intuitionists have seen something interesting about
mathematical reasoning. Once again, their insight turns out to be
epistemic in nature. And once again the way intuitionists have tried
to spell out their insight is woefully inadequate.
Another application of the epistemic interpretation of logic en-
ables us to interpret certain nonstandard readings of higher-order
quantifiers. This interpretation results from taking a second-order
sentence S, which we can assume to be in the negation normal form,
prefixing it by K, and replacing all second-order qunatifiers like (IX)
and (VY) by (RX/K) and (VY/K), respectively. This idea is readily
extended to all higher-order logic. I believe that the resulting interpre-
tation is what some apparent adherents of a nonstandard interpre-
tation of logical and mathematical propositions have had in mind.
Other applications throw light on the idea of intuitionistic logic.
In fact, from the vantage point of the epistemic a;nroach to intu-
itionism we can now see certain serious defects in the attempts that
have been made to formulate explicitly an intuitionistic logic, such as
Heyting's (1956). By and large, what has happened is the same
mistake as was diagnosed above. At best, some unarticulated non-
standard sense of knowing that is captured by intuitionistic logic, but
not a constructivistic sense of knowing what, for example, knowing
which function captures a certain regularity. Yet it is the latter issue
that is really interesting in the epistemology of mathematics.
 EPISTEMOLOGY OF MATHEMATICAL OBJECTS 247

For instance, if a formula like

(11.7)
is supposed to capture the force of
(11.8)
then universal instantiation will not be a valid rule of logical
inference. For (11.8) does not logically imply
(11.16)
for a given b, as a semantical analysis of the situation will show.
Indeed, if we put S[x,y] = (x = y), then (11.8) will say (roughly
speaking) that the identity of all individuals is known under some
guise or other. From that it does not follow that the identity of the
reference of any old nonempty name is known, which is what (11.7)
says.
But if (11.8) does not imply (11.16), then (11.7) should not imply
(11.17)
either, on the intuitionistic interpretation. Yet in intuitionistic logic
it does, for universal instantiation is valid in Heyting's logic. Hence
intuitionistic logic is seriously inadequate as a true epistemic logic of
mathematics. It does not capture the kind of reasoning which is for
instance presupposed in intuitionistic and constructivistic criticism
of the principle of choice.
This important result calls for a few comments. First, it is indepen-
dent of the use of explicit epistemic operators. The same change in
the rules of quantifier instantiation is needed also in IF first-order
logic, as was pointed out in Chapter 3. Hence, it is occasioned by
the use of informational independence rather than by the use of
epistemic notions. Of course, we do not have a complete set of rules
for establishing validity in IF first-order logic, only of rules for
establishing inconsistency.
The failure of Heyting's intuitionistic logic admits of an explana-
tion, if not an excuse. The crucial inference from (11.8) to (11.16) can
be restored by an additional premise
(11.18)
which says that it is known what (which individual) b is. And of
course any old intuitionist would admit that even if he had an
effective way of finding a number y, given a number x, such that
248 THE PRINCIPLES OF MATHEMATICS REVISITED

S[x, y] is the case, it does not follow that he can do that for b unless
he knows which number b is. Hence the failure of Heyting's logic
makes perfect sense to an intuitionist.
This intuitionistic meaning of my criticism enables an intuitionist
to offer a prima facie reply to my criticism. He can reply that the way
his logic is supposed to be applied is different from what I have been
assuming. He is willing to substitute for individual constants like the
b in (11.17) only such terms whose reference is known to him. This
rejoinder nevertheless does not save intuitionistic logic. For one
thing, it makes intuitionistic logic incommensurable with ordinary
first-order logic. The two are then applied in different ways, for
classical logic admits also singular terms whose reference is not
known to the logician. Indeed, the use of "arbitrary individuals" is
a time-honored ploy in logic. Hence intuitionistic logic is not, on the
construal which the intuitionistic response relies on, any longer
a rival to classical logic. It is a specialized logic relying on unspoken
epistemic assumptions for its applicability.
For another thing, similar failures can be pointed out in intuition-
istic logic which are independent of the use of individual constants as
substitution-values. Hence, in brief, Heyting's intuitionistic logic fails
its self-imposed task by criteria that intuitionists should accept them-
selves.
As was indicated in Chapter 3, this defect in the usual instanti-
ation rules is easily corrected by modifying the instantiation rules
which the classicists and the intuitionists have both used. What is
needed is a logic in which existential instantiation is effected by
means of Skolem functions and in which universal instantiation
cannot precede existential instantiation. The requisite changes in
instantiation rules are indicated in Chapters 3 and 4. They will do
the job here. For instance, (11.7) will now imply
(Vx)S[x,f(x)] (11.19)
which is a proxy for
K(Vx)S{x, f(x)] (11.20)
Universal instantiation now yields
S[b,f(b)] (11.21)

which represents
KS[b, f(b)] (11.22)
 EPISTEMOLOGY OF MATHEMATICAL OBJECTS 249

As long as the function f is assumed to be known, this yields the right

inferential relationships.
The real defect of Heyting's logic does not lie in its choice of in-
appropriate rules of inference as much as in its failure to incorporate
the correct instantiation rules. This failure cannot be corrected by
changing its mode of application.
Such modifications in the instantiation rules for quantifiers suffice
to restore our disproof rules to a state of semantical completeness.
However, it is not obvious that there can exist a semantically complete
axiomatization for epistemic logic when it comes to the axiomatiz-
ation of logical truths. This is a problem even if existential quantifiers
and disjunctions are only allowed to be independent of sentence-
initial epistemic K-operators but not independent of other quantifiers.
An informal explanation of this problem can be given by pointing out
that a sentence-initial K is to all effects and model-theoretical pur-
poses a universal quantifier. Hence a sentence of the form
[x, y, z] (11.23)
seems to have the same structure as the Henkin prefix,

(11.24)

And it is known that the logic of Henkin quantifier sentences is not

axiomatizable.
Thus Heyting's idea of implementing intuitionistic logic in the
customary way, that is, by axiomatizing the logical truths of such
a logic, needs a much better justification than it has been given by
intuitionists themselves. Intuitionistic logic and mathematics are
better viewed model-theoretically than axiomatically.
These observations provide examples of interesting results that hold
independently of how functions are thought of as being identified.
One of the virtues of the epistemic interpretation is its generality.
The epistemic interpretation allows us to restrict the functions
(Skolem functions) which figure in second-order translations of
first-order sentences to any class of function that can be considered
as "being known". This class could be a class of function wider than
the class of recursive ones, Instead of requiring computability, we
could merely require of "known" functions that they can be mastered
theoretically in some suitable sense. All such variants of the epistemic
interpretation can be dealt with by means of a suitable epistemic logic.
250 THE PRINCIPLES OF MATHEMATICS REVISITED

However, some possible restrictions could on the contrary be more

severe than a restriction to recursive functions. For instance, some-
one might want to require some kind of real-time computability.
The observations made in Chapter 4 throw further interesting
light on the theoretical situation concerning these different epistemic
interpretations. It was seen in Chapter 4 that the criteria of knowing
that (knowledge of propositions) do not determine the criteria of
knowing what or which (knowledge of objects). This implies that
epistemic interpretations of mathematics are in a sense inevitably
underdetermined. One mathematician may impose more or less
strict requirements on what counts as knowing what given math-
ematical functions are, than a colleague of his or hers. And the
considerations they can use in order to resolve their disagreement
will have to be quite different from the ones used to decide whether
a given mathematical proposition is true or not.
From a more general vantage point, the epistemic interpretation
of mathematics also offers an interesting general perspective on the
epistemology of mathematics. One thing that we might want to
do — or at least try out — is to use the epistemic interpretation of
mathematics, not as a characterization of what it means for math-
ematical propositions to be true, but what it means for them to be
known. This possibility is a most natural one, independently of one's
stance, in the foundations of mathematics. Even the most Platonistic
mathematician might, if only in his or her off-duty hours, raise the
question as to in what sense it can be claimed that we know the
mathematical propositions we are familiar with. Is it really enough
to know that in some Platonic heaven there exists verifying strategies
for such propositions? Should we also know what those strategies
are? Very plausibly, it can be suggested that for the purpose of "really
knowing", for instance, the statement
y] (11.7)

it is not enough that there exists a choice function f which makes

true the formula
(Vx)S[x, f(x)] (11.19)
We also have to know an actual instance of a function f which
does this.
Thus my epistemic interpretation of mathematics has a great deal
of interest for any epistemologist of mathematics, Platonist, con-
 EPISTEMOLOGY OF MATHEMATICAL OBJECTS 251

structivist or whatnot. And if so, epistemic logic is already in

a position to provide a major insight into the epistemology of
mathematics. Without trying to explain the details here, it shows
that all questions about whether a certain mathematical proposition
S is known, reduce to two simpler ones. They are (i) the question of
the truth of S (and/or of the truth of certain related propositions) and
(ii) the question as to whether the Skolem functions involved in S are
known ones or not.
This provides an extremely interesting perspective on the entire
epistemology of mathematics. It shows that the question as to what
functions are "known" (in the sense explained above) is not only
central in some arcane reinterpretation of logic and mathematics — it
is the central question in all epistemology of mathematics.
This observation in turn throws some light on the main lines of the
history of mathematics. It shows that a crucial role in the develop-
ment of mathematical knowledge has been played by the gradual
expansion of the notion of function (see here Youschkevitch 1976,
Hintikka 1995a). And by expansion I do not mean in the first place
the widening of the explicit or implicit definition of an arbitrary
function. Rather, what has happened is a broadening of the range of
functions that can be said to be known by mathematicians. In other
words, what is crucial is the growth of the class of functions that can
be mastered intellectually by means of the available mathematical
concepts and theories. By choosing the class of known functions in
different ways we can thus model the different stages of the historical
development of mathematics with the help of my epistemic interpre-
tation of mathematics.
From the vantage point of the epistemic interpretation, a logical
axiomatization of a mathematical theory (or even of mathematics at
large) in its epistemic dress becomes less of a codification of eternal
truths in a permanent fixed language than a report of the current
state of mathematical knowledge. For the epistemic impact of the
axiom system does not depend only on the truth of the axioms, but
also on the class of functions that are known to the users of the axiom
system. And this class is not specified by the axiom system itself but
only presupposed there. Moreover, it has been seen that such
questions of our knowledge of mathematical objects can arise also
when there are no explicit function quantifiers in the axiom system in
question. All that is needed for the purpose is the presence of
dependent existential quantifiers.
252 THE PRINCIPLES OF MATHEMATICS REVISITED

As far as the ontology of mathematics is concerned, an epistemic

interpretation of mathematics is in any case, in one important respect,
on a different footing from the constructivistic one. In the constructiv-
istic one, the range of function variables was restricted to recursive
ones. If"to be is to be a value of a bound variable", the constructivis-
tic interpretation in effect denies the existence of nonrecursive
functions, at least in so far as strategy functions are concerned.
In contrast, the crucial question in the epistemic interpretation is
which functions are known and which ones are not known. Unless
one assumes an unrealistically strict positivistic position, this does
not presuppose a denial of the existence of functions of other kinds.
On the contrary, the way known functions were delineated was by
means of function quantifiers ranging over a wider class of arbitrary
functions. Thus one can even think of the task of mathematical
research as involving an attempt to bring as many of these strange
"arbitrary functions" to the fold of intellectually mastered and, in
that sense, "known" mathematical objects. This task presupposes,
ontologically speaking, existence rather than the inexistence of all
and sundry arbitrary functions. In general, it makes sense to speak of
mathematical objects as being known or not known and of our
coming to know new mathematical objects only if one assumes that
they actually exist even when they are not known. A defender of the
category of arbitrary functions might thus accuse his or her con-
structivistic critics of a confusion between epistemological and
ontological ideas. In the last analysis, logical rather than Freudian,
the epistemic interpretation on mathematics thus aids and abets the
classical view of mathematical existence, as long as we realize that
this view has to be complemented by epistemic considerations.
The same point can be put in different terms. Once the epistemic
element in mathematics is made explicit, the intuitionistic criticisms
of the classical sense of knowing that and hence of classical mathe-
matics lose their sting. This applies also to the questions whether
certain kinds of mathematical objects in fact exist or not. Indeed, an
approach to the epistemology of mathematical objects (such as we
found the genuine Brouwerian intuitionist undertaking to be)
scarcely makes sense unless the unqualified existence of mathemat-
ical objects is taken in the classical sense. Thus the ultimate thrust of
the intuitionists' philosophy, rightly understood, supports rather
than undermines the kind of classical view of the ontology of
mathematics that goes together with the standard interpretation of
 EPISTEMOLOGY OF MATHEMATICAL OBJECTS 253

the existence of higher-order entities. Intuitionists are in the last

analysis closet realists.
This point is obscured in the literature on the foundations of
mathematics by a failure to distinguish from each other the question
as to which mathematical objects exist and the question as to which
ones are known. This confusion is not a recent phenomenon. On the
contrary, realizing that such a confusion has actually affected math-
ematicians helps us to understand what otherwise might appear to
be contradictory statements by one and the same mathematician.
A representative example is offered by no lesser a figure than
Leonhard Euler. On the one hand, he has been hailed not merely as
a precursor but even as a full-fledged believer in arbitrary functions.
And we do in fact find statements like the following:
If some quantities so depend on other quantities that as the latter are
changed the former undergo change, then the former quantities are called
functions of the latter. This denomination is of the broadest nature and
comprises every method by means of which one quantity could be deter-
mined by others. If, therefore, x denotes a variable quantity, then all
quantifiers which depend upon x in any way or are determined by it are
called functions of it.
(Institutiones calculi quoted after Bottazzini 1986, p. 33.)
This sounds conclusive (that is, general) enough. But in Euler (1988)
we read:
A function of a variable quantity is an analytic expression composed in any
way whatsoever of the variable quantity and numbers or constant quantities.
The obvious explanation is that Euler thought only of functions
represented by analytic expressions as being known to him. So why
did he bother to preserve the possibility of a wider conception of
a function? Once again the actual facts of the history of mathematics
yield an interesting clue. His work on the problem of the vibrating
string had led him to the conclusion that there may in fact be modes
of dependence in physical nature that are discontinuous and hence
cannot be captured by analytic expressions (see Bottazzini 1986, pp.
2 1—33). In other words, problems of applied mathematics made him
aware that there may very well exist functions that are not yet
known, even though Euler did not express himself in this way.
This example illustrates the fact that the epistemic approach to
the foundations of mathematics is very closely related to the realities
of the actual historical development of mathematics.
 APPENDIX

IF First-Order Logic, Kripke,

and 3-Valued Logic

GABRIEL SANDU

In this appendix we show how the result of the definability of truth in

IF first-order logic in the sense discussed in Chapter 6 relates to
Kripke's result of the definability of truth in partial models. For this
purpose, we will give a short description of partial models together
with Kripke's result. Before doing that, we shall have to recast the
description of IF first-order logic in a form which makes comparison
with partial models possible. (The notation used here should be
self-explanatory even though it differs slightly from the one used in
the bulk of this book.)

1. IF first-order logic
Let us fix an IF first-order language L in an arbitrary signature. As
we have seen, this language is nothing else but an ordinary first-order
language enriched with a slash "/". In this new language we shall
have formulas like
Vx2yVz(3w/Vx)Ø (A.l.1)
or
Vx3yVz(4 (v/Vx) i/i) (A.1.2)

We shall denote formulas like (A.!.!) by and formulas

like (A.!.2) by vi). More generally, we shall adopt the
254
 APPENDIX 255

following conventions:
. Xj1.. YmZ1. . . ;W1. .. denotes
Yx1. . . . . . . zk(]wlI Vx1. . . . .(]w1/Vx1. . .

. . . . YmZ1. . . ifr) denotes

.
. . . w1çb denotes

• X,,V)Y1. .
. . Zk(VW1/]X1. . . xe). . . (Vw,/Vx1. . .

. . X,j/1.. YmZi. . . denotes

]Xl...XflVY1...YmJZI...Zk(4(A/]Xl...Xfl)I1I)
We let the formulas of L be the smallest class of formulas
containing the usual atomic formulas and closed under V, A,],
V, Gr7m, and m, k, 1 1). Notice that =
= Vxçb, ,/,) = (4' A = A i/i).
With every formula = x1,. . , of L and model M of L, .

a game M, g) is associated, where g is an assignment

restricted to the free variables of 4 with values in dom(M). If
g(x1) = a1,.. , = we prefer to use the notation
. (M,
a1,... instead of M, g). The game is played by two
players: Myself (the initial verifier) and Nature (the initial falsifier).
The definition of M, a1,. . . , is by induction on the
complexity of 4i:
Definition 1.
(i) M, g), for atomic, contains no moves. If (M, g)
then the player who is the verifier wins M, g);
otherwise, the player who is the falsifier wins.
(ii) G(i 4, M, g) is identical with M, g), except that the players
exchange roles.
(iii) v ), M,g) starts by the verifier choosing ie{O, 1}. The
game goes on as G(41, M, g).
(iv) G((40 A M,g) starts by the falsifier choosing ie{O, 1). The
game goes on as M, q).
(v) G(3x4,M,g) starts with the verifier choosing an element
a e dom(M). The game goes on as in M, gu{(x,a)}).
(vi) M, g) starts with the falsifier choosing an element
a dom(M). The game goes on as in G(4, M, g u { (x, a) }).
256 APPENDIX

(vii) . . Xpj)1. w1çb, M,g) is played in the

. . ZkW1...
following way: the falsifier chooses a sequence a of length n,
then the verifier chooses a sequence b of length m, after which
the verifier chooses a sequence ë of length k, and then finally the
verifier chooses d= (d1 ,. d1). The game goes on as in G(4, M,
. . ,

(viii) . . . YmZp . . . M, g) is played exactly

as . . M,g), except that the . .. w14,
moves made by the verifier in the former are made by the
falsifier in the latter.
(ix) . . . M, g)is played
ly to . . . . . w,4, M, g), except that in
. .

the last move the verifier chooses ie {O, 1 }, and the game goes on
asin
(x) The description of the game . . . . YmZI.. .

M, g) should be obvious by now.

The notion of strategy is crucial here. A strategy for a player p (p is

either Myself or Nature) in the game G(4, M, g) is a set
M,g)) = of functions f1,.. ., each corresponding to
a move M1 of the player p in the game. Usually, each f1 is defined on the
set of all the possible moves of the opponent of p before the move
with the following exceptions: in the game . . x,y1...

. . ZkW1... w,4, M, g), the associated with the move of the

verifier prompted by w. is not defined on but on
(dom(M))k. In other words this game is one of imperfect information.
Similarly for the game . x,,y1... YmZI.. M, g). Of
. .

course the same applies to the moves of the falsifier in

. YmZp . ;w1.. . w140, M, g),
. . and in . x,,y1... .

YmZ1. . . ), M, g).
A strategy for a player in the game G(4, M, g) is a winning one if
the player wins G(4, M, g) no matter what the moves of the opponent
are.
The following proposition is straightforward from the definitions:

Proposition 1. For any formula 4, model M of L, and assignment gin

M (restricted to the free variables of it is not possible that both
Myself and Nature have a winning strategy in G(4, M, g).

In Chapter 2 the truth of a sentence 4 in the model M was defined

as the existence of a winning strategy for Myself in the game G(4, M).
 APPENDIX 257

It is convenient, for later comparisons, to have a notation for both

truth and falsity in M. For this purpose, whenever 4' is a formula, we
use M to mean is true in M', and M to mean '4i is
false in M'.

Definition 2. Let 4 be a formula of L, M a model of L, and g an

assignment in M restricted to the free variables of 0. Then

(i) (M, g) + if Myself has a winning strategy in G(4, M, g).

(ii) (M, g) - if Nature has a winning strategy in G(0, M, g).
We shall now define by double induction two mappings * and #,
the truth-preserving and the falsity-preserving mapping, respective-
ly, which map formulas of L into formulas of L:

Definition 3. The mappings * and # are defined as follows:

(i) = 0; q5# = 4, for atomic. i

(ii) (i 0)* = (4#; (—i 4')# =

(iii) (4 v = (4* v v = (4# A
(iv) (4 A = A A ,1i)# (4# v
(v) = = VxO#.
(vi) (VxO)* =
(Vx4)# =
(vii) . . x,y1.. . . ZkW1.
= . . ZkW1.. .

ZkW1... w1)#
= .
Xj1.. YmZ1 . . W,çb#.
(viii) . . ZkW1. .
—
G*n,m
k,1 XnYi.
YmZj
X,y1. . YmZ1. . .

(ix) . . y,,,z1. . .

—
— Vn,m
k,2X1. . X,j'1.
.

— v*n.m .x,,y1. .YmZ1. A#

— k,2
k k2 YmZ1 .
— v*N,m X1. X,j'1. .
— k,2
.XnYl...YmZl...Zk(OO,Ol))#
= U
258 APPENDIX

A straightforward induction on the complexity of 4) shows

that
Proposition 4. For any formula 4), model M of L, and assignment g:
(I) Myself hs a winning strategy in G(4), M, g) if Myself has a win-
fling strategy in G(4)*, M, g).
(ii) Naure has a winning strategy in G(4), M, g) if Myself has a win-
ning strategy in G(4)#, M, g).
Corollary. For any formula 4):
(i) (M, g) + if (M, g) +
(ii) (M, g) GTS4> — if (M, g)
Proof: from Proposition 4 and Definition 1.
Proposition 5. Let 4> be an arbitrary IF first-order sentence. Then the
following holds for any model M:
(i) if MIGTS4)
(ii) MI'GTS(i4))
(iii) MIGTS(4) v iffMI=GTS4)+ or MIrGTSI/,+
(iv) v and
(v) A and
(vi) A 'I') jffMtrGTS4) or
Proof: (i) From the Corollary and Definition 3 we have
MI'GTS(l if 4))*)+ if MtrGTS(4)#)+ if
All the other cases are similar.
We know that the law of excluded middle fails in IF first-order
logic. This raises the question — what is the valuation schema that is
satisfied by the Boolean connectives in this logic.
In order to give an answer to this question, let 4) M denote the
II 1

truth-value of the L-sentence 4) in the model M. Putting

14>11M
1, if (Hence, by Proposition 1: not MIZGTS4))
14>11M0, if MIZGTS4> (Hence, by Proposition 1: not MIrGTS4>+)

andnotMIGTS4)
it is easy to check, using Proposition 5, that the Boolean connectives
satisfy the Strong Kleene valuation schema (conjunction, and
implication are defined via negation and disjunction in the usual
 APPENDIX 259

way):
i/i i4 —i4i (4)A41) (4)—34,)
110
4)

100 0
1
1

1
1

0
1

0 0
1

1? 0 ? 1 ? ?

01 0 0 0
001
1 1 1

1 0 0 1 1

0? 1 ? ? 0 1 ?

?1 ? 0 1 ? 1 ?

?0 ? 1 ? 0 ? ?

?? ? ? ? ? ? ? R

It will be seen below that the partial logic used by Kripke in his result
about the definability of truth satisfies the Strong Kleene valuation
schema too.

2. IF first-order logic and the definability of truth and falsity

In this section we shall resume the argument of Chapter 6 showing
that an IF first-order language (in the signature of Peano Arithmetic)
admits of a truth-predicate in this language itself.
So let the signature S of the language L be that of Peano
Arithmetic (i.e., S = {0, 1, and <}).
A couple of remarks are in order here.
Any formula of the form ..YmZ1. ..Zk(4)o,4)1) is de-
.

finable from . ZkW1. . w,4) and the standard

. . . . .

quantifiers:
. . . . .

. XnYi.. . YmZ1.. . ZkWj((W1 = 0 A v (w1 = I A

(A.2. 1)

So in the sequel we shall disregard formulas . .

Zk(4)O,
Also, we shall assume that the formulas of L are closed only
with respect to instead of being closed with
respect to every . x,,y1.. . YmZ1. . . ZkWI.. w14). (cf. our remarks
. .

below).
Let us fix a Gödel numbering of L. We let te1 be the Gödel number
of the expression e. We shall use the following recursive functions
260 APPENDIX

and relations (in addition to *, and #):

ct(n): n is a closed term of L
v(n): the value of n, if n is a closed term of L
fml(n): ii is an IF first-order formula of L
st(n): n is an IF first-order sentence of L
Sub(m, n,p): the substitution function whose value is 14(p)1, if m is the
Godel number of the formula 4(x), and n is the Gödel number of the
free variable x; and m otherwise (p is the numeral corresponding top).
.
A El-formula over L is a formula having the form
where i4' is a first-order formula (in the vocabulary of L plus
.1
tJ 1'" ''in
Let L(X) (X is a second-order variable) be the conjunction of the
sentences
Vx(st(x) —. (X(x).-* X(x*))) (A.2.2)
Vx, y((ct(x) A ct(y)) —* (X(L(x = y)1)—* v(x)

= v(y)) A (X(1(x cv(y)) A (X(H (x <y)1)t-rl v(x)

<v(y)) A (X('i (x = y)1)÷-ci v(x) = v(y))) (A.2.3)

Vx,y((st(x) A st(y))-.(X(t(x A y)1)—*X(x) A X(y))) (A.2.4)

Vx, y((st(x) A st(y)) —÷ (XQ(x v y)1) X(x) v X(y))) (A.2.5)

Vy((f 1(y) A —*(X(1(Vx1y1) -+ VxX(Sub(y, x)))) (A.2.6)

Vy((f 1(y) A —* (X(1(1x1y1) -+ 3xX(Sub(y, 1x11, x)))) (A.2.7)

'cly((fl(y) A
JfgVx1z1X(Sub(Sub(Sub(Sub(y, Ey1], f(x1)), g(z1)),
1x11, t), 1z11, z1)))) (A.2.8)
—÷
Vy((fI(y) A —*

Jfx1Vy1w1X(Sub(Sub(Sub(Sub(y, 1z11,

1y11, y1), 1w11, wi)))

.
(A.2.9)

Remark. The expressions 'i', 'A', 'v', 'V', ']', 'GI'I', and 'Gt'1"
occurring within the corners u do not have their usual meaning but
are function symbols which name primitive recursive functions. Thus
 APPENDIX 261

'—i' denotes a function f defined by: f(n) = H 4)', if n is the Gddel

number of the formula 4); and n otherwise. 'V' denotes a binary
function, g defined by: g(n, m) = tVx4)', if n is the Gödel number of the
formula 4) and m is the Gödel number of the variable x; and
n otherwise. Similarly, 'Gd' denotes a function h defined by:
if n is the Gödel number of the
formula 4), m is the Gddel number of the variable x1, p is the Gödel
number of the variable y1, q is the Gödel number of the variable z1,
and m is the Gödel number of the variable w1; and n otherwise. The
other cases should be obvious by now.
Proposition 1. Let 0(x) be the A X(x)), and
N the standard model of L. Then for every IF first-order formula 4)
(A.2.1O)

(In other words, 0(x) defines explicitly 'truth in N' for the language L.)
Proof of (A.2.10): Assume and, in order to prove
A let A = 4) is a sentence of L and
An induction on the complexity of 4) shows that
(N, A) L(X) A X(14)'). In the induction, it is enough to consider
only the cases for 4) being in negation normal form, since 4) is in
X only if 4)* is in X, and 4)* is in negation normal form. Let us
prove two cases.
Assume 4) is Vxi/i and N)rGTS(VX4,)+. The inductive hypothesis is
IfNIrGTSX+ then (N,A)IrL(X) A (+)
for every sentence x of complexity less than that of Vx4i.
From NIrGTS(VX4,)+, it follows that Also, since Myself
has a winning strategy in N), Myself has to win against every
move of Nature, that is, Myself has a winning strategy in every game
N, n), for every neN. Hence (N, n) for every n, or
equivalently, for every n. Hence, by ( +), for
every nEN, that is, (N,
Assume now 4) is and
(x1y1z1w1))t Then for every m picked up by Nature, Myself finds
p = f(m), and for every n picked up by Nature, Myself finds s
(f, and g are part of the winning strategy of Myself in the game) such
that Myself wins the game G(4,(x1, y1,z1, w1), N, m, p, n, s), that is, (N,
rn, p. n, z1w1))+. Equivalently, N p, n,
262 APPENDIX

By the inductive hypothesis, A('*(m, p. n, s)'), for every quad-

ruple m, p, n, SEN, such that p = f(m), and s = g(n). Hence
(N, A) 1y', f(x i)), 1w11, g(z1)),
x1), 1z11, z1)))).
The other cases are similar.
In the other direction, assume N fr (]X)(L(X) A We prove
by induction on the complexity of 4) that N IZGTS4) The only
complicated cases are for 4) being and
Suppose 4) is Vxtfr. By our assumption, there is a set A such that
(N,A)frL(X) A X('Vx*1) (A.2.1l)

which implies
(N,A)bL(X) A (A.2.12)

for all neN. Hence

A (A.2.13)

for all neN. By the inductive hypothesis

(A.2.14)

for all neN, that is, for all neN. Thus Myself has
a winning strategy in every game GØ/4x), N, n), for arbitrary n. Then
it is trivial to see that Myself has a winning strategy in G(Vxifr, N). For
let Nature choose an individual neN and the game go on with
N, n). Then Myself does not need to do anything else than
follow his winning strategy in GØfr(x), N, n).
Suppose now that 4) is and that
N fr(RX)(L(X) A (A.2.15)

In a similar way to the previous case, we get that

(N, A) fr L(X) A Ey11, f(x1)),

1w11, g(z1)), x1), z1)))) (A.2.16)

Hence
A X(1ifr(m, p, it, s') (A.2.17)

for every quadruple (m, p, it, s) which is such that p = f(m), and
s = g(n), for some functions f and g. As above, by the inductive
 APPENDIX 263

hypothesis we have
p, n, (equivalently (N, rn, p, n, s)

(A.2.18)

for every quadruple (in, p, n, s) such that p = f(m), and s = g(n). Again,
it is easy to see that Myself has a winning strategy in
z 1z1 N). For let Nature choose arbitrary m; then Myself
chooses p such that p = f(m). Then let Nature choose again, n;
Myself chooses s such that s = g(n), and the game goes on with
m, p, n, s). In this game, let Myself follow his
winning strategy which exists by (A.2.17).
The formula 0(x) which defines 'truth in N' for the IF first-order
language L (strictly speaking, for the subfragment of L closed with
respect to the standard quantifiers, the Boolean connectives and
is a Ei-formula. However, it was pointed out in
Chapter 6 that each such formula is equivalent to an IF first-order
formula. To show that 0(x) is equivalent to a formula of the relevant
IF first-order subfragment (i.e., the subfragment closed with respect
to the standard quantifiers, the Boolean connectives and
GNx1y1z1w1) one needs the following further observations.
In Walkoe (1970) an effective procedure is given which translates
every Ei-formula 4 of an arbitrary language L into a formula
which is logically equivalent to such that belongs to a language
L* which extends ordinary first-order logic with all Henkin quanti-
fiers (more exactly, there is an additional clause saying that, if is
a first-order formula, then Q4 is a formula of L*, where Q is defined
below).
A Henkin quantifier Q is usually defined as a triple Q =
(AQ, EQ, DQ), where AQ and EQ are disjoint sets of variables (universal
and existential variables, respectively) and DQ (the dependency
relation) is a partial ordering on AQ x EQ. Intuitively, if (x, y)EDQ
then the existentially quantified variable y is in the scope of the
universally quantified variable x. For instance, the Henkin quanti-
fier Q = ({x, z}, {y, w}, { (x, y), (z, w) }) says that the quantifier is in
the scope of Vx, and that is in the scope of \fz. The interpretation
of a Henkin quantifier is given through a second-order translation.
For instance, the interpretation of Q is given by

Qxyzwçb(x, y, z, f(x), z, g(z)) (A.2. 19)

264 APPENDIX

In Krynicki (1993) it is shown that in models with definable pairing

functions every formula of L* is materially equivalent to a formula of
the subfragment of L* in which the only Henkin quantifier is Q
(from A.2.19). On the other side, Q is definable in IF first-order
language, that is
Qxyzw4.(x, y, z, w) (A.2.20)

Thus, since N has definable pairing functions, we have shown that

the formula 0(x) is equivalent in N with a formula of the IF first-
order language L. U

3. Partial models
The easiest way to introduce partial models is as models of a prop-
ositional language L consisting of a set P of primitive propositional
symbols The sentences of L are built up from P using
negation and disjunction in the usual way. A partial model M of L is
atripleM =(A, V),whereAisasubsetofPand and
disjoint subsets of M. The idea is that represents the true
sentences of L, and V the false ones. For S a sentence of L we use
to mean 'S is true in A', and AIZPMS mean 'S is false in
to
A'. Truth and falsity of a sentence S in A are defined by induction in
the following way:
primitive

if SeV,S primitive
MtZPM(1R)+

if

if MIZPMR and
Conjunction, implication and double implication are defined from
negation and disjunction in the usual way. It is easy to build up
partial models A or Lin which the law of excluded middle fails, that
is, we have neither nor MIPMS even at the level of atomic
propositional symbols.
In case L is a first-order language (containing only relational
symbols), a partial model for it is defined analogously. That is, each
 APPENDIX 265

relational symbol R is assigned both an extension RM+ and

a counterextension RM -. The inductive definition of satisfaction (with
respect to an assignment) is then defined analogously with the
propositional case. The new clauses are
if there is an aedom(M) such that

(M, (M,

a first-order language in an arbitrary signature, and

N a partial model for it. A game-theoretical interpretation for L is
analogous to the game-theoretical interpretation on classical models
for L. The only difference is with atomic formulas. That is, clause (i)
of Definition 1 in section 1 is now replaced by
(1) contains no moves. If then
the player who is the verifier wins G(R(t1, . ., ta), M, g); II' .

.
. then the player who is the falsifier wins
G(R(t1,. .. , ta), J%4, g).

The clause for identity is the same as in the case of classical models.
The proof of the next proposition is straightforward.
Proposition 1. For any first-order formula of L, partial model M of
L and assignment g in M:
iff(M, (*)

iff(M, (**)

Partiality in the models can be combined with partiality of informa-

tion, that is, we can give a game-theoretical interpretation for an IF
first-order language L and a partial model M of L. All the relevant
definitions remain the same.

4. Kripke's result
Let L be a first-order language in the signature of Peano Arithmetic.
We enrich L with a truth-predicate Tr and denote the resulting
language by Lt We expand the classsical model N of L to a classical
model (N, E) of (E is the interpretation of Tr). We get a partial
model (N, (E, D)) of by adding a set D disjoint from E to be the
counterextension of Tr. Intuitively, E is intended to consist of the
266 APPENDIX

names (Gödel numbers) of the sentences of which are true in (N,

(E, D)), and D of those sentences (and possibly other individuals in N)
to which Tr does not apply. Notice that in this partial model, the
counterextensions of all the predicate symbols, except Tr, are the
complements of the respective extensions.
Tarski's classical result is that we cannot find a classical model
(N, E) of such that E is the set of the Gödel numbers of the
sentences of true in (N, E).
Kripke's well-known result is that we can find a partial model
which has the analogous property, that is, we can find (E, D) such
that
(N, (E, D)) + if (N, (E, D)) + (A.4. 1)

(N, (E, D)) — if (N, (E, D)) Ir — (A.4.2)

Such an (E, D) is called a fixed point for The existence of a

minimal fixed point can be proved by a straightforward transfinite
construction: We let
= (N,

= {nGN: i st(n)} u (N, — }

We then set = and = By cardinality con-

siderations, there are EA and DA such that for all K> = and
= DK. The pair (E2, D2) is the minimal fixed point.
The fixed point (E, D) for Kripke's first-order language has
been axiomatically characterized by Feferman (1984):
Let KF(Tr) (Kripke—Feferman axioms) be the conjunction of the
following sentences in the language
Vx(Tr(x) (A.4.3)

(4 is an atomic sentence ofLor the negation of an atomic sentence

of L -+ 4)) (A.4.4)

Vx(ct(x) —* 4—*Tr(x))) (A.4.5)

Vx(ct(x) —+ Tr(x)1)4-+ Tr(H x1))) (A.4.6)

Vx'c/y((st(x) A st(y))—*(Tr(t(x v v Tr(y)))) (A.4.7)

Vx'c/y((st(x) A v A

(A.4.8)
 APPENDIX 267

Vx(st(x) —+ (Tr(H H (A.4.9)

Vy(fl(y) A ix', x))) (A.4.1O)

Vy(fi(y) A st(H (Tr(H y1, 1x1, x)))

(A.4.11)

—i A Tr(H x1)) (A.4. 12)

Feferman (1984) proved the following:

Theorem 1. Let E be a subset of N. Then the classical model (N, E) is

a model of KF(Tr) iff(N, E, {nonsentences of u {'4': H 4'eE}) is
a fixed point for Lt

5. GTS and partial models

It is time to relate our result on the definability of truth to Kripke's
result. But first, it might be interesting to point out the simi-
larities and the dissimilarities between the two kinds of par-
tiality.
In the case of partial predicates (Kripke), we start with a new
primitive predicate Tr. Then we can find a partial interpretation
(E, D) for it such that "truth-in-(N, (E, D))" (in the sense of partial
models) is defined by Tr. In the case of partiality of information
(GTS), we are able to find a formula 0(x) which defines "truth-in-N"
(in the sense of GTS).
The truth-defining formula 0(x) cannot be replaced by a predicate
symbol by the standard technique, that is, by means of an explicit
definition
(A.5.1)

because, given the fact that 0(x) does not satisfy the law of excluded
middle, the predicate S would have to be partially interpreted. If we
want to reduce 0(x) to a predicate, we then have to combine the two
kinds of partiality, that is, to have an IF language which is game-
theoretically interpreted, and in addition to allow for the predicate
symbols of L to be partially interpreted. We indicated at the end of
section 3 how this can be done.
So let L be an IF language (in the signature of Peano Arithmetic)
and let us add to it a predicate symbol Tr. We denote the resulting
268 APPENDIX

language by Then, by a transfinite construction similar to that

for partial models in the preceding section, it can be shown that there
is a fixed point (E, D) for
(N, (E, D)) + if (N, (E, D)) + (A.5.2)

(N, (E, D)) if (N, (E, D)) — (A.5.3)

A syntactical characterization of a fix point, in the style of

Feferman can be obtained by adding to the Kripke—Feferman
axioms from the preceding section the sentences
Vy((fl(y) A

Tr(Sub(Sub(Sub(Sub(y, 1y11, f(x 1, g(z

x1), z1)))) (A.5.4)

Vy((fl(y) A st('i —+(Tr(1—i

1Tr(Sub(Sub(Sub(Sub(H y1, tx xj)' tz f(y1)),

Yi)' tti1, t1)))).
.
(A.5.5)

We denote the result by KF*(Tr). We then have

Theorem 1. Let L be an IF language (in the signature of Peano
Arithmetic), N the standard model of L, and Tr a new predicate
symbol. Then for any A dom(N) we have
(*) (N, A)I KF*(Tr) if (N, A, {nonsentences} u is
a fixed point for Lu {Tr}.
Proof:
Let K = {nonsentences of Lu {Tr} } u I—iWe assume
(N, A) Ir KF*(Tr), and then show by induction on the complexity of
that (A, K) is a fixed point, that is
(N,(A,K))IrGTS4+ if (A.5.6)

(A.5.7)

For all the first-order clauses, the proof is identical with that given
for Feferman for partial models in the first-order case. So it is enough
to consider the two additional clauses related to IF languages.
However, in order to make the connection with the proof given by
 APPENDIX 269

Feferman more transparent, we shall also review the clause for

negation.
Suppose 4) is—i i/i, and (N, (A, K)) Then, by the defini-
tion of a fixed point, H 4i1eA. Conversely, assume H Then by
the definition of K, 'iJi1eK, and by the inductive hypothesis,
Hence, by Definition 2 (i) of section 1,
(N,(A,
In order to prove (7), assume Hence, by
Definition 2 (ii), (N, (A, K)) By the inductive hypothesis,
hence (N, A) Tr(1i/i1). By (8) of KF(Tr), (N, A) (Tr(H H
Hence H H ifr"eA, and thus H The converse is similar.
Suppose that 4) is G1:1x1y1z1t14, and assume
Then there are functions f and g such that for
all natural numbers n and m: (N, (A, K)) f(n), m, g(m))
By the inductive hypothesis, f(n), m, g(m))'€A. We thus showed
that there are functions f and g such that for all n and
!n 1y11, f(n)), g(m)), Ex1], n), m)eA. But
(N, A) KF*(Tr), hence by (A.5.4) we get (N, A) x1y1z 1t1 4)'),
that is,
Conversely, assume EGNxiyiziti4JIeA. Then (since A is the
interpretation of Tr) (N, A) ix 1y1z1 t1 ifr'). But (N, A) satisfies
KF*(Tr), therefore there are functions f and g such that
(N, A) Ir p, n, s)1), for all quadruples (m, p, n ,s) such that p =
f(m), and s = g(n). Hence p, n, s)'eA. By the inductive hypoth-
esis (N,(A, p, n, s) such
that p = f(m), and s = g(n). But then it is easy to see that Myself has
a winning strategy in G(G1: (N, (A, K)). Therefore
(N, (A, This proves (A.5.6).
In order to prove (A.5.7), assume (N,(A, x1y1z1t14)
Then, by the definition of the fixed point,
Conversely, assume Then, by the definition of
K, Hence (N, By
(A.5.5), there is an m and a function f such that for all p and s we have
(N, A) Tr(H ,ji(in, p. f(p), s)'). Hence L-_i i/i(m, p. f(p), s)1eA. By the
inductive hypothesis, (*) (N, (A, n, s). But then it is
easy to see that Nature has a winning strategy in G(GHx1y1zit14i,
(N, (A, K))). In his first move, let Nature choose an m as above. Then
for any arbitrary choice p made by Myself, let Nature choose
We then know by (*) above, that for whatever s chosen by Myself,
Nature has a winning strategy in G(4), (N, (A, K), tn, p, n, s)). N
270 APPENDIX

It is straightforward to show that, if(A, K) is a fixed point, then the

axioms KF*(Tr) are true in (N, A).
The result we reached here is stronger than Kripke's given the fact
that the partially interpreted IF languages are much stronger than
partially interpreted first-order languages.

*I am grateful to Jouko Väänänen, Taneli Huuskonen, and Kerkko Luosto

for useful suggestions concerning the material of this appendix.
 References

Baldus, Richard, 1928. "Zur Axiomatik der Geometrie 1: Uber Hilberts

Vollständigkeitsaxiom," Mat hematische Annalen 100, 32 1—333.
Barwise, Jon, 1979. "On Branching Quantifiers in English," Journal of Philosophical
Logic 8, 47—80.
Barwise, Jon, 1985. "Model-Theoretic Logics: Background and Aims," in Barwise
and Feferman 1985, 3—23.
Barwise, Jon, and Solomon Feferman (eds.), 1985. Model-Theoretic Logics,
Springer, New York.
Benacerraf, Paul, 1973. "Mathematical Truth," Journal of Philosophy 70, 66 1—680;
reprinted in Benacerraf and Putnam (1983), 403—420.
Benacerraf, Paul, and Hilary Putnam (eds.), 1964. Philosophy of Mathematics,
Prentice-Hall, Englewood Cliffs, NJ. Second ed. 1983, Cambridge University
Press, Cambridge.
van Benthem, Johan, and Alice ter Meulen (eds.), 1996. Handbook of Logic and
Language, Elsevier, Amsterdam.
Blass, Andreas, and Yuri Gurevich, 1986. "Henkin Quantifiers and Complete
Problems," Annals of Pure and Applied Logic 32, 1—16.
Blumenthal, Otto, 1935. "Lebensgeschichte [Hilberts]," in Hilbert (1935) 388—429.
Boltzmann, Ludwig, 1905. Populdre Schriften, Verlag von Johann Ambrosius
Baerth, Leipzig.
Boolos, George, 1971. "The Iterative Conception of Set," Journal of Philosophy 68,
215—232. Reprinted in Benacerraf and Putnam (1983), 486—502.
Bottazzini, Umberto, 1986. The "Higher Calculus": A History of Real and Complex
Analysis from Euler to Weierstrass, Springer, Berlin.
Brouwer, L. E. J., 1975. Collected Works I, ed. by A. Heyting, North-Holland,
Amsterdam.
Carnap, Rudolf, 1934. Logische Syntax der Sprache, Springer, Wien.
Chomsky, Noam, 1977. Essays on Form and Interpretation, North-Holland,
Amsterdam.
Chomsky, Noam, 1981. Lectures on Government and Binding, Foris, Dordrecht.

271
272 REFERENCES

Chomsky, Noam, 1986. Knowledge of Language, Praeger, New York.

Cohen, Paul, 1966. Set Theory and the Continuum Hypothesis, Benjamin,
New York.
Cummins, Robert, 1989. Meaning and Mental Representation, MIT Press,
Cambridge, MA.
Dauben, J. W., 1979. Geory Cantor: His Mathematics and Philosophy of the Infinite,
Harvard University Press, Cambridge, MA.
Davidson, Donald, 1965. "Theories of Meaning and Learnable Languages,"
in Yehoshua Bar-Hillel (ed.), Logic, Methodology and Philosophy of Science,
Proceedings of the 1964 International Congress, North-Holland, Amsterdam,
383—394.
Dawar, Anuj, and Lauri Hella, forthcoming. "The Expressive Power of Finitely
Many Generalized Quantifiers."
Devlin, Keith, 1984. Constructibility, Springer, New York.
Dummett, Michael, 1977. Elements of Intuitionism, Clarendon Press, Oxford.
Dummett, Michael, 1978. Truth and Other Enigmas, Duckworth, London.
Dummett, Michael, 1991. The Logical Basis of Metaphysics, Harvard University
Press, Cambridge, MA.
Dummett, Michael, 1993. The Seas of Language, Clarendon Press, Oxford.
Easton, W. B., 1964. Powers of Regular Cardinals, Ph. D. Dissertation, Princeton
University.
Ebbinghaus, H. D., J. Flum and W. Thomas, 1984. Mathematical Logic, Springer,
New York.
Ekeland, Ivar, 1988. Mathematics and the Unexpected, University of Chicago Press,
Chicago.
Enderton, H. B., 1970. "Finite Partially Ordered Quantifiers,"
Mathematische Logik und Grundlagen der Mat hematik 16, 393—397.
Engdahl, Elisabeth, 1986. Constituent Questions, D. Reidei, Dordrecht.
Erdös, P., A. Hajnal, A. Mate, and R. Rado, 1984. Combinatorial Set Theory: Partial
Relations for Cardinals, North-Holland, Amsterdam.
Euler, Leonhard, 1988. Introduction to Analysis of the Infinite I, translated by John
D. Blanton, Springer, Berlin. (Original 1748.)
Feferman, Solomon, 1984. "Towards useful Type-Free Theories, I," Journal of
Symbolic Logic 49, 75—111.
Feferman, Solomon, 1993. "What Rests on What? The Proof-Theoretic Analysis of
Mathematics," in J. Czermak (ed.), Philosophy of Mathematics: Proceedings of the
15th International Wittgenstein Symposium. Hölder—Pichler---Tempsky, Vienna,
pp. 147—171.
Feigner, Ulrich, 1971. "Comparison of the Axioms of Logical and Universal
Choice," Fundamenta Mathematicae 71, 43—62.
Fenstad, Jens—Erik, 1971. "The Axiom of Determinateness," in J.-E. Fenstad (ed.),
Proceedings of the Second Scandinavian Logic Symposium, North-Holland,
Amsterdam.
Fenstad, Jens-Erik, 1996. "Partiality," in van Benthem and ter Meulen, 1996.
Fernández Moreno, Luis, 1992. Wahr he it und Korrespondenz bei Tarski,
Konigshausen & Neumann, Würzburg.
Fine, Terrence L., 1973. Theories of Probability, Academic Press, New York.
 REFERENCES 273

Frege, Gottlob, 1879. eine der aritmetischen nachgebildete Formel-

sprache des reinen Denkens, Nebert, Halle.
Frege, Gottlob, 1980. Philosophical and Mathematical Correspondence, Basil
Blackwell, Oxford.
Freudenthal, Hans, 1957. "Zur Geschichte der Grundlagen der Geometrie," Nieuw
Archiefvoor Wiskunde 5, 105—142.
Gaifman, Haim, 1975. "Global and Local Choice Functions," Israel Journal of
Mathematics 22, 257—265.
Gödel, Kurt, 1930. "Die Vollständigkeit der Axiome des logischen Funktionen-
kalküls," Monatshefte für Mathematik und Physik 37, 340—360; reprinted in
English translation in Gödel 1986, 102—123 (see also 44—101).
Gödel, Kurt, 1931. "Uber formal unentscheidbare Sätze der Principia Mat hematica
und verwandter Systeme 1," Monatshefte für Mathematik und Physik 38, 173—
198; reprinted in English translation in Gödel 1986, 144—195.
Gödel, Kurt, 1940. The Consistency of the Axiom of Choice and of the Generalized
Continuum Hypothesis (Annals of Mathematical Studies 3), Princeton University
Press, Princeton.
Gödel, Kurt, 1947. "What Is Cantor's Continuum Problem?" American Mat hemat-
ical Monthly 54, 515—525. Reprinted in an expanded form in Benacerraf and
Putnam (1983), 470—485.
Gödel, Kurt, 1958. "Uber eine bisher noch nicht benützte Erweiterung des finiten
Standpunktes," Dialectica 12,280—287; reprinted in English translation in Godel
1990, 241—251.
Godel, Kurt, 1986. Collected Works, vol. 1, Oxford University Press, New York.
Gödel, Kurt, 1990. Collected Works, vol. 2, Oxford University Press, New York.
Goldfarb, Warren, 1979. "Logic in the Twenties: The Nature of the Quantifier," The
Journal of Symbolic Logic 44, 35 1—368.
Goldfarb, Warren, 1989. "Russell's Reasons for Ramification," in C. Wade Savage
and C. Anthony Anderson (eds.), Rereading Russell (Minnesota Studies in
Philosophy of Science 12), University of Minnesota Press, Minneapolis, 24—40.
Grabiner, Judith V., 1981. The Origins of Cauchy's Rigorous Calculus, MIT Press,
Cambridge.
Graham, Ronald L., Bruce L. Rotschild, and Joel H. Spencer, 1990. Ramsey Theory,
2nd ed., John Wiley & Sons, New York.
Graham, Ronald L., and Joel H. Spencer, 1990. "Ramsey Theory," Scientific
American 262, no. 1 (July 1990), 112—117.
Hallett, Michael, 1984. Cantorian Set Theory and Limitation of Size, Clarendon
Press, Oxford.
Heim, Irene R., 1982. The Semantics of Definite and Indefinite Noun Phrases,
Dissertation at the University of Massachusetts, Amherst, MA.
Henkin, Leon, 1950. "Completeness in the Theory of Types," Journal of Symbolic
Logic 15, 81—91.
Henkin, Leon, 1961. "Some Remarks on Infinitely Long Formulas;' in (no editor
given) Injinitistic Methods, Pergamon Press, Oxford, 167—183.
Heyting, Arend, 1931. "Die intuitionistische Grundlegung der Mathematik," Erken-
ntnis 2, 106—115; English translation in Benacerraf and Putnam, 1983, 52—61.
Heyting, Arend, 1956. Intuitionism: An Introduction, North-Holland, Amsterdam.
274 REFERENCES

Hubert, David, 1899. "Grundlagen der Geometrie," in B. G. Teubner (ed.), Fes-

zur Freier der Enthüllung des Gauss-Weber Denkmals in Gottingen,
Leipzig, pp. 3—92.
Hubert, David, 1900. "Uber den Zahibegriff," Jahresbericht des Deutschen
Mathematiker Vereinigung 8, 180—184.
Hubert, David, 1918. "Axiomatisches Denken," Mathematische Annalen 78, 405—
415.
Hubert, David, 1922. "Neubegrundung der Mathematik. Erste Mitteilung," Abhan-
dlungen aus dem Mat hematischen Seminar der Hamburg Universität 1, 157—177.
Hubert, David, 1925. "Uber das Unendliche," Mat hematische Annalen 95, 161—190.
English translation in Benacerraf and Putnam 1983, 183—201.
Hilbert, David, 1935. Gesammelte Abhandlungen 3, Springer, Berlin.
Hilbert, David, and W. Ackermann, 1928. Grundzüge der theoretischen Logik,
Springer, Berlin.
Hilpinen, Risto, 1983. "On C. S. Peirce's Theory of the Proposition: Peirce as
a Precursor of Game-Theoretical Semantics," in Eugene Freeman, (ed.), The
Relevance of Charles Peirce, The Hegeler Institute, La Salle, Illinois, 264—270.
Hintikka, Jaakko, 1973. Logic, Language-Games and Information: Kantian Themes
in the Philosophy of Logic, Clarendon Press, Oxford.
Hintikka, Jaakko, 1974. "Quantifiers vs. Quantification Theory," Linguistic Inquiry
5, 153—177,
Hintikka, Jaakko, 1975. "Impossible Possible Worlds Vindicated," Journal of
Philosophical Logic 4, 475—484.
Hintikka, Jaakko, 1976a. "Quantifiers in Logic and Quantifiers in Natural
Languages," in Stephan Körner (ed.), Philosophy of Logic, Basil Blackwell,
Oxford.
Hintikka, Jaakko, 1976b. "The Prospects of Convention T," Dialectica 30, 61—63.
Hintikka, Jaakko, 1979. "Quantifiers in Natural Languages." In Saarinen (1979),
8 1—117.
Hintikka, Jaakko, 1985, "A Spectrum of Logics for Questioning," Philosophica 35,
135—150.
Hintikka, Jaakko, 1987. "Game-Theoretical Semantics as a Synthesis of Truth-
Conditional and Verificationist Theories of Meaning," in Ernest LePore (ed.),
New Directions in Semantics, Academic Press, London.
Hintikka, Jaakko, 1988a. "What Is the Logic of Experimental Inquiry?" Synthese
74, 173—190.
Hintikka, Jaakko, 1988b. "On the Development of the Model-Theoretical View-
point in Logical Theory," Synthese 77, 1—36.
Hintikka, Jaakko, 1988c. "Model Minimization — An Alternative to Circumscrip-
tion," Journal of Automated Reasoning 4, 1—13.
Hintikka, Jaakko, 1989. "Logical Form and Linguistic Theory," in Alex George
(ed.), Reflections on Chomsky, Basil Blackwell, Oxford, 41—57.
Hintikka, Jaakko, 1990. "Paradigms for Language Theory," Acta Philosophica
Fennica 49, 181—209.
Hintikka, Jaakko, 1992. "Different Constructions in Terms of 'Knows' ", in
Jonathan Dancy and Ernest Sosa, eds., A Companion to Epistemology, Basil
Blackwell, Oxford, 99—104.
 REFERENCES 275

Hintikka, Jaakko, 1993a. "New Foundations for Mathematical Theories," in

J.Oikkonen and J. Väänänen, eds., Logic Colloquium '90 (Lecture Notes in
Logic 2), Springer, Berlin, 122—144.
Hintikka, Jaakko, 1993b. "The Original Sinn of Wittgenstein's Philosophy of
Mathematics," in Klaus Puhi (ed.), Wittgenstein's Philosophy of Mathematics,
Holder—Pichler--Tempsky, Vienna, 24—51.
Hintikka, Jaakko, 1993c. "Gödel's Functional Interpretation in a Wider Perspec-
tive," in H.-D. Schwabl (ed.), Yearbook 1991 of the Kurt Gödel Society, Vienna,
5—39.
Hintikka, Jaakko, 1995a. "The Standard vs. Nonstandard Distinction: A Water-
shed in the Foundations of Mathematics." in Jaakko Hintikka (ed.), From
Dedekind to Gödel, Kluwer Academic, Dordrecht, 2 1—44.
Hintikka, Jaakko, 1995b. "What Is Elementary Logic? Independence-Friendly
Logic as the True Core Arc of Logic," in Kostas Gavrogtu et at. (eds.), Physics,
Philosophy, and the Community, Kluwer Academic, Dordrecht,
301—326.
Hintikka, Jaakko, 1996. "Hilbert Vindicated?" Synthese (forthcoming).
Hintikka, Jaakko, and Jack Kulas, 1983. The Game of Language, D. Reidel,
Dordrecht.
Hintikka, Jaakko, and Jack Kulas, 1985. Anaphora and Definite Descriptions: Two
Applications of Game- Theoretical Semantics, D. Reidel, Dordrecht.
Hintikka, Jaakko, and Veikko Rántala, 1976. "A New Approach to Infinitary
Languages," Annals of Mathematical Logic, 10,95—115.
Hintikka, Jaakko, and Unto Remes, 1974. The Method of Analysis, D. Reidel,
Dordrecht.
Hintikka, Jaakko, and Gabriel Sandu, 1991. On the Methodology of Linguistics:
A Case Study, Basil Blackwell, Oxford.
Hintikka, Jaakko, and Gabriel Sandu, 1994. "What Is a Quantifier?," Synthese 98,
113— 129.
Hintikka, Jaakko, and Gabriel Sandu, 1995. "What Is the Logic of Parallel
Processing?" International Journal of Foundations of Computer Science, 6,
27—49.
Hintikka, Jaakko, and Gabriel Sandu, 1996. "Game-theoretical Semantics," in van
Benthem and ter Meulen, 1996.
Hintikka, Merrill B., and Jaakko Hintikka, 1986, Investigating H4ttgenstein, Basil
Blackwell, Oxford.
Horn, Laurence R., 1989. A Natural History of Negation, University of Chicago
Press.
Hughes, R. I. G., 1989. The Structure and Interpretation of Quantum Mechanics,
Harvard University Press, Cambridge, MA.
Jackendoff, Ray S., 1972. Semantic Interpretation in Generative Grammar, MIT
Press, Cambridge, MA.
Jeroslow, R. G., 1973. "Redundancies in the Hilbert-Bernays Derivability Condi-
tions for Gödel's Second Incompleteness Theorem," The Journal of Symbolic
Logic 38, 359—367.
Jones, James P., 1974. "Recursive Undecidability — An Exposition," The American
Mathematical Monthly 81, 724—738.
276 REFERENCES

Kanamori, Akihiro, 1994. The Higher Infinite, Springer-Verlag, Berlin.

Kanamori, Akihiro, and M. Magidor, 1978. "The Evolution of Large Cardinal
Axioms in Set Theory," in G. H. Muller and Dana Scott (eds.), Higher Set Theory
(Lecture Notes in Mathematics 699), Springer, New York, 99—275.
Keenan, Edward, and Dag Westerstahl, 1996. "Generalized Quantifiers," in van
Benthem and ter Meulen, 1996.
Kleene, Stephen C., 1952. Introduction to Metamathematics, Van Nostrand, New
York.
Kleinberg, Eugene M., 1977. Infinitary Combinatorics and the Axiom of Determinate-
ness (Lecture Notes in Mathematics 612), Springer, Berlin.
Kreisel, Georg, 1953. "A Note on Arithmetic Models for Consistent Formulae of the
Predicate Calculus II," in the Proceedings of the XI International Congress of
Philosophy 14, Amsterdam & Louvain, 37—47.
Kripke, Saul, 1975. "Outline of a Theory of Truth," Journal of Philosophy 72,
690—716; reprinted in Martin 1984, 53—81.
Krynicki, Michael, 1993. "Hierarchies of Partially Ordered Connectives and Quan-
tifiers," Mathematical Logic Quarterly 39, 287—294.
Krynicki, Michael, and Jouko Väänänen, 1989. "Henkin and Function Quanti-
fiers," Annals of Pure and Applied Logic 43, 273—292.
Kunen, Kenneth, 1980. Set Theory: An Introduction to Independence Proofs,
North-Holland, Amsterdam.
Kusch, Martin, 1989. Language as the Universal Medium vs. Logic as Calculus:
A Study of Husserl, Heidegger and Gadamer, Kiuwer, Dordrecht.
Lindström, Per, 1969. "On Extensions of Elementary Logic," Theoria 35, 1—il.
Lovejoy, A. 0., 1936. The Great Chain of Being, Harvard University Press,
Cambridge, MA.
Mancosu, Paolo, 1991. "On the Status of Proofs by Contradiction in the Seven-
teenth Century," Synthese 88, 15—41.
Martin, Robert L. (ed.), 1978. The Paradox of the Liar, Ridgeview, Atascadero, CA.
Martin, Robert L. (ed.), 1984. Recent Essays on Truth and the Liar Paradox,
Clarendon Press, Oxford.
Martin—Löf, Per, 1966. "The Definition of Random Sequences," Information and
Control 6, 602—619.
Martin—Löf, Per, 1970. "On the Notion of Randomnesss," in A. Kino, J. Myhill, and
R. E. Vesley (eds.), Intuitionism and Proof Theory, North-Holland, Amsterdam,
73—78,
Martin—Löf, Per, 1984. Intuitionistic Type Theory, Bibliopolis, Napoli.
Matiyasevich, Yuri V., 1993. Hubert's Tenth Problem, MIT Press, Cambridge,
MA.
McGee, Vann, 1991. Truth, Vagueness and Paradox, Hackett, Indianapolis.
Mendelson, Elliott, 1987. Introduction to Mathematical Logic, 3rd ed., Wadsworth
& Brooks/Cole, Monterey, CA.
Moore, Gregory H., 1982. Zermelo's Axiom of Choice, Springer, Berlin.
Moore, Gregory H., 1988. "The Emergence of First-Order Logic," in William
Aspray and Philip Kitcher (eds.), History and Philosophy of Modern Mathematics
(Minnesota Studies in the Philosophy of Science XI), University of Minnesota
Press, Minneapolis, 95—135.
 REFERENCES 277

Moore, Gregory H., 1994. "Logic and Set Theory," in I. Grattan-Guinness (ed.),
Companion Encyclopedia of the History and Philosophy of the Mathematical
Sciences, vol. 1, Routledge, London, 635—643.
Moreno, see Fernández Moreno.
Morris, Charles, 1938. Foundations of the Theory of Signs (International Encyclo-
pedia of Unified Science 1, no. 2), University of Chicago Press, Chicago.
Moschovakis, Y. N., 1974. Elementary Induction on Abstract Structures, North-
Holland, Amsterdam.
Mostowski, Andrzej, 1955. "A Formula with no Recursively Enumerable Model,"
Fundamenta Mat hematicae 42, 125—140.
Mostowski, Andrzej, 1965. Thirty Years of Foundational Studies (Acta Philosophica
Fennica 17), Helsinki.
von Neumann, John, 1928. "Zur Theorie der Gesellschaftsspiele," Mat hematische
Annalen 100, 295—320.
von Neumann, John, and Oskar Morgenstern, 1944. Theory of Games and Economic
Behavior, Princeton University Press, Princeton.
Parsons, Charles, 1977. "What Is the Iterative Conception of Set?" in Robert Butts
and Jaakko Hintikka (eds.), Logic, Foundations of Mathematics and Computabil-
ity Theory (Proceedings of the 5th International Congress of Logic, Methodology
and Philosophy of Science 1975 1), Dordrecht Reidel, Dordrecht, 335—367.
Reprinted in Benacerraf and Putnam (1983), 503—529.
Partee, Barbara, 1984. "Compositionality," in F. Landman and F. Veltman (eds.),
Varieties of Formal Semantics, Foris, Dordrecht, 281—312.
Pelletier, Francis Jeifry, 1994. "The Principle of Semantic Compositionality," Topoi
13, 11—24.
Poincaré, Henri, 1905—06. "Les mathematiques et Ia logique," Revue de métaphys-
ique et de morale 13, 8 15—835 and 14, 17—34. English translation in Science and
Method, Dover, New York, 1952, 143—159.
Prawitz, Dag, 1965, Natural Deduction. A Proof-Theoretical Study, Acta Univer-
sitatis Stockholmiensis, Stockholm.
Putnam, Hilary, 1971. Philosophy of Logic, Harper & Row, New York.
Rabin, Michael, 1957. "Effective Computability of Winning Strategies," in M.
Dresher, A. W. Tucker, and P. Wolff (eds.), Contributions to the Theory of Games
III (Annals of Mathematics Studies 39), Princeton Univrsity Press, 147—157.
Ramsey, Frank. P., 1925. "The Foundations of Mathematics," Proceedings of the
London Mathematical Society, Ser. II, vol. 25, 338—384.
Ramsey, Frank P., 1930. "On a Problem of Formal Logic," Proceedings of the
London Mathematical Society, Ser. 2, vol. 30, Part 4, 338—384.
Rantala, Veikko, 1975. "Urn Models: A New Kind of Non-Standard Model for
First-Order Logic," Journal of Philosophical Logic 4, 455—474.
Rasiowa, Helena, and Roman Sikorski, 1963. The Mathematics of Metamathematics,
Polska Akademia Nauk, Warsaw.
Russell, Bertrand, 1903. The Principles of Mathematics, Allen & Unwin, London.
Russell, Bertrand, 1905. "On Denoting," Mind 14, 479—493.
Russell, Bertrand, and A. N. Whitehead, 1910—13. Principia Mathematica I—Ill,
Cambridge University Press, Cambridge.
Saarinen, Esa (ed.), 1979. Game-Theoretical Semantics, D. Reidel, Dordrecht.
278 REFERENCES

Sandu, Gabriel, 1993. "On the Logic of Informational Independence and Its
Applications," Journal of Philosophical Logic 22, 29—60.
Sandu, Gabriel, 1996. "IF First-Order Logic and Truth-Definitions," Journal of
Philosophical Logic.
Sandu, Gabriel, and Jouko Väänänen, 1992. "Partially Ordered Connectives,"
für Mathematische Logik und Grundlagen der Mathematik 38, 361—
372.
Schnorr, Claus Peter, 1971. Zufdlligkeit und Wahrscheinlichkeit, (Lecture Notes in
Mathematics 218) Springer, Berlin.
Shapiro, Stewart, 1985. "Second-Order Languages and Mathematical Practice,"
The Journal of Symbolic Logic 50, 714—742.
Shapiro, Stewart, 1991, Foundations without Foundationalism, Clarendon Press,
Oxford.
Simmons, Keith, 1990. "The Diagonal Argument and the Liar," Journal of Philo-
sophical Logic 19, 277—303.
Simons, Peter, 1992. Philosophy and Logic in Central Europefrom Bolzano to Tarski,
Kluwer Academic, Dordrecht.
Skolem, Thoralf, 1922. "Einige Bemerkungen zur axiomatischen Begrundung der
Mengenlehre," Proceedings of the Scandinavian Mathematics Congress,
Helsinki, 1922, 217—232.
Smullyan, Raymond, 1968. First-Order Logic, Springer, Berlin.
Stewart, Ian, 1992. The Problems of Mathematics, new edition, Oxford University
Press, New York.
van Stigt, Walter P., 1990. Brouwer's Intuitionism, North-Holland, Amsterdam.
Szabo, M. E. (ed.), 1969. The Collected Papers of Gerhard Gentzen, North-Holland,
Amsterdam.
Tarski, Alfred, 1933. "Projcie prawdy w jezykach nauk deducyjnych," Prace
Towarzystaw Naukowego Warszawsiego, wydizial III, no. 34; English translation
in Tarski 1956a; German translation as Tarski 1935.
Tarski, Alfred, 1935. "Der Wahrheitsbegriff in den formalisierten Sprachen," Studia
philosophica 1, 26 1—405.
Tarski, Alfred, 1951. A Decision Met hod for Elementary Algebra and Geometry, 2nd
ed. University of California Press, Berkeley and Los Angeles.
Tarski, Alfred, 1956a. "The Concept of Truth in Formalized Languages," in Tarski
1956b, 152—278.
Tarski, Alfred, 1956b. Logic, Semantics, Metamathematics: Papers from 1923 to
1938, Clarendon Press, Oxford.
Tarski, Alfred, 1959. "What Is Elementary Geometry?" in L. Henkin, P. Suppes,
and A. Tarski (eds.), The Axiomatic Method, North-Holland, Amsterdam, 16—29.
Tarski, Alfred, 1992. "Drei Briefe an Otto Neurath," Grazer Philosophische Studien
43, 1—32.
Tennenbaum, Stanley, 1959. "Non-Archimedean Models for Arithmetic," Notices
of the American Mathematical Society 6, 270.
Toepell, M.-M., Uber die Entstehung von David Hilberts "Grundlagen der Geomet-
ne," Vandenhoek & Ruprecht, Göttingen, 1986.
Torretti, Roberto, 1978. Philosophy of Geometry from Rietnann to Poincaré, D.
Reidel, Dordrecht,
 REFERENCES 279

Troelstra, A. S., 1977. Choice Sequences: A Chapter of Intuitionistic Mathematics,

Clarendon Press, Oxford.
Troelstra, A. S., and Dirk van Dalen, 1988. Constructivism in Mathematics: An
Introduction, North-Holland, Amsterdam.
Tuuri, Heikki, 1990. Infinitary Languages and Ehrenfeucht-Fraisse Games, Disserta-
tion, University of Helsinki.
van der Waerden, B. L., 1985. A History of Algebra, Springer, Berlin.
Walkoe, W. Jr., 1970. "Finite Partially Ordered Quantification," Journal of Sym-
bolic Logic 35, 535—555.
Wang, Hao, 1990. Computation, Logic, Philosophy: A Collection of Essays, Kiuwer
Academic, Dordrecht.
Westerstahi, Dag, 1989, "Quantifiers in Formal and Natural Languages," in
D. Gabbay and F. Guenther (eds.), Handbook of Philosophical Logic IV, Reidel,
D., Dordrecht, 1—131.
Weyl, Hermann, 1917. Das Kontinuum, Veit, Leipzig.
Williams, Neil H., 1977. Combinatorial Set Theory, North-Holland, Amsterdam.
Wittgenstein, Ludwig, 1922. Tractatus Logico-Philosophicus. The German text of
Ludwig Wittgenstein's Logisch-philosophische Abhandlung with English transla-
tion by C. K. Ogden, Routledge & Kegan Paul, London.
Youschkevitch, A. P., 1976. "The Concept of Function up to the middle of the 19th
century," Archive for the History of Exact Sciences 16, 37—85.
Zermelo, Ernst, 1908. "Untersuchungen über die Grundlagen der Mengenlehre I,"
Mat hematische Annalen 65, 261—281.
 Index of Names

Ackermann, Wilhelm, 5, 6, 46, 201, 274 Church, Alonzo, 12, 18, 114
Anderson, C. Anthony, 273 Cohen, Paul, 272
Archimedes, 8 Cummins, Robert, 120, 272
Aristotle, 3, 21, 123, 164
Aspray, William, 276 van Dalen, Dirk, 237, 279
Dancy, Jonathan, 274
Baire, René, 185 Dauben, Joseph W., 164, 168, 182, 272
Baldus, Richard, 92, 271 Davidson, Donald, 66, 106, 272
Bar-Hillel, Yehoshua, 42 Dawar, Anuj, 93, 272
Barwise, Jon, 12, 56, 71, 121, 185, 190, Devlin, Keith, 272
271 Dedekind, Richard, 200
Beltrami, Eugenio, 164 DeMorgan, Augustus, 73, 133, 147
Benacerraf, Paul, 21, 210, 271, 273, Desargues, Girard, 21
274, 277 Descartes, René, 24, 89
van Benthem, Johan, 271, 275 Dreben, Burton, xi
Beth, Evert W. 6, 61 Dresher, M., 277
Blass, Andreas, 198, 271 Dummett, Michael, 22,26,27, 36, 38,39,
Blumenthal, Otto, 21, 271 211, 212, 215, 218, 222, 244, 245,
Boltzmann, Ludwig, 21, 271 272
Bolzano, Bernhard, 189
Boole, George, 2, 166, 258, 263 Easton, William B., 272
Boolos, George, 168, 271 Ebbinghaus, Heinz-Dieter, 15, 141,
Borel, Emile, 226 272
Bottazzini, Umberto, 253, 271 Ekeland, ivar, 229, 272
Brouwer, Luitzen E. J., 237, 238, 252 Enderton, H. B., 56, 71, 121, 272
Butts, Robert, 277 Erdös, Paul, 77, 78, 200, 272
Euclid, 1, 2, 91, 164
Cantor,Georg,2, 164—169,181,182,273 Euler, Leonhard, 253, 271, 272
Carnap, Rudolf, x, xi, 12, 16, 17, 152
Cauchy, Augustin-Louis, 9, 273 Feferman, Solomon, 12, 185, 201, 202,
Chomsky, Noam, 6, 46, 54, 126 266—268, 271, 272

Index prepared by Risto Vilkko

281
282 INDEX OF NAMES

Feigner Ulrich, 272 83, 103, 104, 109, 110, 124—127,

Fenstad, Jens-Erik, 33, 147, 181 139, 157, 160, 165, 193, 199, 208,
Fernández Moreno, Luis, 130, 272 223, 224, 239, 251, 274, 275, 277
Fine, Terrence, 228, 272 Hintikka, Merrill B., 17, 23, 127, 275
Flum, Jörg, 15, 141 Holder, (Ludwig) Otto, 204
Fraenkel, Adolf, 163, 176, 202 Horn, Laurence R., 275
Freeman, Eugene, 274 Husserl, Edmund, 108, 120, 196
Frege, Gottiob, vii, 2, 5, 11, 12, 18, 46— Hughes, R. I. G., 77, 275
51, 54—56, 66, 67, 69, 71, 90, 94, Huuskonen, Taneli, 270
106, 125, 126, 161, 164, 181, 182,
186, 195, 200, 207, Jackendorif, Ray S, 275
Freud, Sigmund, 252 Jefferson, Thomas, vii
Freudenthal, Hans, 92, 273 Jeroslow, R. G., 275
Jewett, R. I., 78, 79
Gabbay, Dov M., 279 Jones, James, P., 28, 219, 275
Gaifman, Haim, 273
Gavroglu, Kostis, 275 Kanamori, Akihiro, 169, 276
Georg, Alex, 274 Kant, Immanuel, 3, 89, 104
Gödel, Kurt, xi, 6, 14—18, 30, 31, 41, Keenan, Edward, 57, 79, 276
88—91,93—99,102,113—119,131, Kino, A., 276
132, 141—145, 150—152, 165, 166, Kitcher, Philip, 276
168, 169, 172—177, 199, 204, 222, Kleene, Stephen, C., 235, 258, 259, 276
223, 225,229,231—235,259—261, Klein, Felix, 164
265, 266, 273 Klein, Esther, 78
Goldbach, C., 192, 197 Kleinberg, Eugene M., 200, 276
Goldfarb, Warren, 210, 273 Kolaitis, Phokion G., 130
Grabiner, Judith V., 74, 273 KOrner, Stephan, 274
Graham, Ronald, L., 78, 200, 273 Kreisel, George, 40, 217, 276
Grattan-Guinness, Ivor, 277 Kripke, Saul, xi, 122, 161, 254, 259,
Guenther, Franz, 279 266—268, 270, 276
Gurevich, Yuri, 198, 271 Krynicki, Michael, 117, 118, 188, 276
Kuhn, Thomas, vii
Hadamard, Jacques, 244 Kunen, Kenneth, 276
Hales, A. W., 78, 79 Kulas, Jack, 27, 44, 109, 110, 125, 157,
Hallett, Michael, 168, 273 275
Hegel, Georg W. Friedrich, 211 Kusch, Martin, 276
Heim, Irene, 157, 273
Hella, Lauri, 93 Landman, Fred, 277
Henkin, Leon, 51, 56, 79, 98, 102, 117— Leibniz, Gottfried W., 112, 122, 125
119, 121, 165, 192, 193,249,263, Lenin, Vladimir I., vii
271, 273, 278 LePore, Ernest, 274
Heyting, Arend, viii, 212, 246—249, 271, Lesniewski, Stanislaw, 17, 108
273 LindstrOm Per, 18, 141, 276
Hiipakka, Janne, xi Lovejoy, Arthur 0., 168, 276
Hubert, David, xi, 1—6, 8—11, 16, 19, 21, LOwenheim, Leopold, 6, 18,59,141,189,
40, 46, 69, 89—92, 98, 164, 182, 191, 202, 203
199—201, 206, 208, 219, 222, 233, Luosto, Kerkko, 270
234, 271, 274
Hilpinen, Risto, 28, 212, 274 Mach, Ernst, 21
Hintikka, Jaakko, 12, 13, 17, 20—23, 27, Magidor, Menachem, 169, 276
29, 36, 44, 52, 66, 68, 73, 79, 82, Mancosu, Paolo, 246, 276
 INDEX OF NAMES 283

Martin, Robert L., 131, 161, 276 Savage, C. Wade, 273

Martin-Löf, Per, 228, 276 Schnorr, Claus Peter, 228, 278
Matiyasevich, Yuri V., 219, 276 Schrödinger, Erwin, 2
Maxwell, James Clerk, 2 Schwabl, H.-D, 275
McGee, Vann, 173, 276 Scott, Dana, 276
Mendelson, Elliott, 114, 143, 276 Shapiro, Stewart, 192, 193, 278
ter Meulen, Alice, 271, 275, 276 Shoenfield, Joseph, 168
Moore, Gregory, 170, 195, 201, 244, Sikorski, Roman, 203, 277
276, 277 Simmons, Keith, 143, 278
Morgenstern, Oskar, 277 Simons, Peter, 278
Morris, Charles, 42, 277 Skolem, Thoralf, 6, 18, 30, 34, 39,48,49,
Moschovakis, Yannis N., 210, 277 51, 59, 60, 79, 80, 141, 174, 175,
Mostowski, Andrzej, 15, 40, 217, 277 177, 179, 189, 191, 198, 202, 203,
Muller, Gert H., 276 213, 217, 221, 223, 224, 228, 236,
Myhill, John, 276 242, 246, 248, 249, 251, 278
Smullyan, Raymond, 68, 278
von Neumann, John, 24, 79, 226, 277 Sosa, Ernst, 274
Neurath, Otto, 17, 204 Spencer, Joel H., 78, 200, 273
Ogden, C. K., 279 Stewart, Ian, 29, 30, 44, 278
Oikkonen, Juha, 122, 275 van Stigt, Walter P., 278
Szabo, M. E., 278
Parsons, Charels, 168, 277 Suppes, Patrick, 278
Partee, Barbara, 106, 277 Szekeres, George, 77, 78
Pascal, Blaise, 21
Peano, Giuseppe, vii, 98, 153, 197, 202, Tarski, Alfred, xi, 13—19, 22, 28, 29, 32,
223, 259, 265, 267, 268 41,96—99,101,105—108,110—
Peirce, Charles S., 28, 46, 212, 274 112,116, 117, 120—424, 128—131,
Pelletier, Francis J., 106 133, 138—141, 144,171—174, 182,
Poincaré, Henri, 182, 277 204, 266, 278
Prawitz, Dag, 212, 277 Tennenbaum, Stanley, 223, 278
Puhi, Klaus, 275 Teubner, B. G., 274
Putnam, Hilary, 16, 21, 271, 274, 277 Thomas, Wolfgang, 15, 141
Toepell, Michael-Markus, 21, 278
Quine, W.V., 7, 12, 18, 67, 129, 228 Torretti, Roberto, 164, 279
Rabin, Michael, 219, 277 Troelstra, Anne S., 237, 245, 279
Ramsey, Frank, 77—79, 137, 165, 194, Turing, Alan, 95, 215, 237
200, 210, 273, 277 Tucker, A. W., 277
Rantala, Veikko, 12, 29, 103, 160, 239, Tuuri, Heikki, 130, 279
275, 277
Rasiowa, Helena, 203, 277 Väänänen, Jouko, 56, 122, 188,270,275,
Remes, Unto, 104, 275 276, 278
Rorty, Richard, 130 Veltman, Frank, 277
Rotschild, Bruce L., 78, 200, 273 Versley, Richard E., 276
Russell, Bertrand, vii, ix, x, xii, 5, 12, 43,
46, 47, 49—51, 55, 66, 67, 94, 161, van der Waerden, B. L., 204, 279
169, 174, 180, 186, 194, 195, 210, Walkoe, Wilbur John Jr., 56, 71, 121,
273, 277 279
Wang, Hao, 168, 198, 279
Saarinen, Esa, 274, 278 Weierstrass, Karl, 29, 44, 189, 271
Sandu, Gabriel, xi, 27, 52, 56,79, 82, 110, Westerstahl, Dag, 57, 79, 276, 279
122, 126, 224, 254, 275, 278 Weyl, Hermann, 179, 279
284 INDEX OF NAMES

Williams, Neil H., 200, 279 Wolff, P., 277

Wittgenstein, Ludwig, 12, 18, 22, 23, 32,
42, 44, 127, 162, 211, 215, 222, Youschkevitch, A. P., 251, 279
226, 279
Whitehead, Alfred N., vii, xii, 46, 180, Zermelo, Ernst, 251, 279
194, 277
 Index of Subjects and Titles

analysis, conceptual, 9 of reducibility (Russell's), 180

arithmetic, 79 of set existence, 194
elementary, viii, 14, 30, 89, 90, axiomatization, 1, 9, 65—68, 88—104,
93—98, 101, 113, 116, 118, 143, 126, 145, 146, 183, 251, 266
145, 171, 197, 206, 223, 224, of arithmetic, 94
233, 235 complete, viii, 6, 65—67, 153, 184, 224,
first-order (IF), 132 249
Peano, 98, 197,202,223,259,265,267, of geometry (Euclidean), 2
269 of geometry (Hubert's), 2, 3, 8, 10,
undefinability of 15 11
axiom of logic, 4, 5
Archimedean, 8 nonlogical, 1, 3, 4,
Aussonderung (Zermelo's), 179
of choice, ix, 7, 19, 32, 40—42, 80, 82, Begriffsschr(ft, 11, 125
83, 165, 175, 176, 193, 201, 213,
220—223, 244 Church's Thesis, 114
deductive, viii completeness, x, 1, 19, 67, 88—104,
of determinacy, 33 145, 163, 166, 169, 170, 192,
of geometry, 8 233
of induction, 153 deductive, 91, 92
Kripke-Feferman, 268 descriptive, 91
of large cardinals, 169 Hilbertian, 92
of choice, ix, 7, 19, 32, 40—42, 80, 82, semantical, 91
83, 165, 175, 176, 193, 201, 213, compositionality, 13, 106—112,116,118,
220—223, 244 121, 139, 140, 149
of completeness, 8, 92, 208 principle of, ix
of comprehension, 165, 176, 178, conjectures, 193
195 Goldbach's, 192, 197
of constructibility, 168 constructivism, ix, x, 7, 38, 39, 121,
of continuity, 3 137, 142, 161, 177, 210—237,
Peano, 153 241, 243—245, 250

Index prepared by Risto Vilkko

285
286 INDEX OF SUBJECTS AND TITLES

continuity, 9, 10 standard vs. nonstandard, 165, 166,

continuum, 193 192—196, 201, 210, 213, 246
course-of-values (Frege's Werthverldufe), intuitionism, 134, 161, 177, 229,
181 237—247, 252
classes, viii, 12
knowledge of objects vs. knowledge of
Definition e — ö, 9, 29 facts, x, 86, 252
dependence vs. independence, x, 72—87, language games, 32—39, 159, 212, 229,
of quantifiers, 47—53, 56—58, 70 238
differentiation, 9 Wittgensteinian, 22, 23, 42, 211, 222
distinction law
de dicto vs. de re, 72, 124 de Morgan's, 73, 133, 147
definitory vs. strategic rules, 128 of double negation, 73, 133, 147
logical vs. mathematical reasoning, of the excluded middle, see tertium
90, non datur
semantics vs. pragmatics, 42 lemma, diagonal, 131, 142—145, 150,
152, 172—174, 180
equations, 2, 185 logic
equicardinality, viii, 186, 187, 190, 191 epistemic, 83—87, 238—253
functions of, 1—21
finiteness, 7 deductive/descriptive, 4, 9, 10, 19, 20,
functions, 40, 193, 253 46, 92, 149, 209, 210, 230—234
in extension (Russell), 174 higher-order, 6, 7, 16, 97, 104, 126,
Skolem, 30, 31, 34, 35, 39, 48, 49, 51, 163, 165, 186, 190, 192, 194, 195,
60,79,80, 174,175, 177,213,217, 201, 203, 207—209, 223, 231
221, 223, 224, 228, 236, 242, 246, independence-friendly (IF), 46—163,
248, 249, 251 178—210,217,229,241,247,254—
270
geometry, 9, 10, 78, 89, 97, 234 intuitionistic, viii, 134, 212, 238, 246—
applied, 3 249
elementary, 197 modal, 5
Euclidean, 1, 2, 164 model-theoretical, 12, 13
nonEuclidean, 97, 164 ordinary first-order, viii, ix, 6, 7, 9,
Grundlagen der Geometrie, 1, 4, 10, 92 11, 16, 18, 33, 34, 36, 43, 46—51,
54—6 1, 65—70, 73, 74, 77, 82, 89,
Impossibility result (Tarski's), 15, 22, 90,93, 94, 96, 109, 125—129, 132,
123, 171, 172 135, 138, 141, 143, 152, 163, 170,
induction 171, 174, 184—190, 194, 195, 198,
mathematical, viii, 7, 184, 188, 191 201—205, 209, 213, 216, 217,223,
transfinite, 189 229
inference, 4, 5, 20, 21, 33, 37, 71, 88 second-order, 7, 65, 89, 98, 129, 150,
Hilbert's, 8 191—197, 213, 220, 225
modes of, viii, 184, 190 syllogistic, 89
infinity, viii, 184, 187 logicism, 183, 190
interpretation, 3, 4
constructivistic, ix, 224, 225, 235, method
251 analytic, 89
epistemic, 242—246, 250—253 axiomatic, 2, 10, 94, 206, 234
Gödel's functional (Dialectica), 222, deductive, 100
223, 232, 233, 235 mathematical, 89
intuitionistic, 247 semantical vs. syntactical, 90
 INDEX OF SUBJECTS AND TITLES 287

model set
concept of, 12 power, viii, 7, 184, 188, 191
interrogative, 36, 39 Hintika, 68
urn, 12, 103 set theory, viii, 3, 8, 16, 33, 89, 97, 99,
model theory, viii, 11—13, 16, 19—21, 85, 129, 132, 152, 185, 186, 189, 191,
90, 102, 104, 129, 141, 152, 162, 198—207, 213
163, 165, 170—174, 184, 192, 197, axiomatic, ix, 7, 18, 19, 74, 100, 102,
239, 249 104, 126, 163—182, 185, 191—195,
202, 207, 222
nominalism, 7, 117, 198, 220, 234, 250 slash-notation (/), 51—53, 254
numbers Sprachlogik, 46, 55, 124, 126
natural, 8, 102, 104, 116, 166, 177, 184,
206, 207, 217, 219 tertium non datur, ix, 32, 33, 65, 68,
real, 10, 102, 177, 184, 185, 189 73, 131-146, 161, 178, 179,
Gödel, 14, 30,113—119,131,132,142, 181, 212, 217, 219, 258,
143, 145, 151, 152, 172—174, 180, 267
225, 259—261, 265, 266 theorem
Beth's, 7, 61
paradox, 7, 97, 101, 121, 145, 165, 167, Bolzano—Weierstrass, 189
175, 239 deduction, 149
liar, 131, 132, 142—144, 150, 152, 159, Desargues', 21
160, 176 Godel's incompleteness, 18, 93, 95,
of logical omniscience, 239 96, 132, 165, 166, 173
Meno's, 34, 35, 38, 39 Gödel's impossibility, 142
Skolem, 202, 203 Hales—Jewett, 78, 79
possible worlds, 12, 34, 66, 239, 240 interpolation, 6
predicates, viii, 6, 185—188 Lindström, 18, 141
Principia Mathematica, vii, x Löwenheim—Skolem, 6, 18, 59, 141,
Principles of Mathematics, vii, ix, x 191, 202, 203
proof, 5, 10, 38 Pascal's, 21
Ramsey-type, 137
relation, viii, 4,6, 8, 113—115, 135, 188, separation, 6, 61, 133
194 Tarski's, 16
epsilon, 115, 135, 188, 194, 178 theory
elementary number, 153
semantical games, 25—27, 3 1—45, 57, 58, game, see semantics, GTS
63, 70, 115, 119—123, 128, 132, group, 9, 10, 167
134, 137, 138, 147, 148, 154, 157, lattice, 3, 9, 10, 167
158, 174, 211—219, 227, 228, 231, model, see model theory
235—238 of anaphora (GTS), 157, 158
semantics of fields, 9
game-theoretical (GTS), 25—50, 54, of finite types, 197
57, 58, 72, 77, 90, 103, 109, 112, of reals, 197
115, 119—122,125,132—134,148, of types (Russell's), 194
157, 160, 163, 201, 210, 211, proof, 11, 141, 162, 234
2 13—222, 225—227, 230,232, 236, quantum, 76, 77
241, 243, 254—270 Ramsey, 77—79, 200
ineffability of, 17, 18 set, see set theory
logical, 11 type, 205, 207
truth-conditional, 23 Tract atus Logico-Philosophicus,
verificationist, 23 153
288 INDEX OF SUBJECTS AND TITLES

truth definition, viii, ix, 1—45, 105—130, T-schema (Tarski's), 41, 106, 138—140,
144 172—174
Tarski's, 13—18, 22, 28, 29, 32, 65, 105—
112, 120—124, 128, 129, 133, 266 universal language, 17—19, 204, 205
truth definition, (cont.)
recursive, 14 Vienna Circle, 17, 18, 233
game-theoretical, 26, 27,30,32,34,35,
41, 119—122, 172—174, 220, 221 Warsaw School, 17
pragmatist, 44 well-ordering, viii, 7, 184, 188, 193