(Jaakko Hintikka Selected Papers 3) Jaakko Hintikka (Auth.) - Language, Truth and Logic in Mathematics-Springer Netherlands (1998)

LANGUAGE, TRU1H AND LOGIC IN MATHEMATICS
JAAKKO HINTIKKA SELECTED PAPERS
VOLUME 3
The titles published in this series are listed at the end of this volume
JAAKKO HINTIKKA
Boston University
LANGUAGE, TRUTH
AND LOGIC
IN MATHEMATICS
SPRINGER-SCIENCE+BUSINESS MEDIA, B.V.

A C.I.P. Catalogue record for this book is available from the Library of Congress.
ISBN 978-90-481-4923-0 ISBN 978-94-017-2045-8 (eBook)

DOI 10.1007/978-94-017-2045-8
Printed on acid-free paper
All Rights Reserved

1998 Springer Science+Business Media Dordrecht
Originally published by Kluwer Academic Publishers in 1998
No part of the material protected by this copyright notice may be reproduced or
utilized in any form or by any means, electronic or mechanical,
including photocopying, recording or by any information storage and
retrieval system, without written permission from the copyright owner.
TABLE OF CONTENTS
ORIGIN OF THE ESSAYS vii
INTRODUCTION ix
1. "What Is Elementary Logic? Independence-Friendly Logic as the True

Core Area of Logic"
2. (with Gabriel Sandu) "A Revolution in Logic?" 27
3. "A Revolution in the Foundations of Mathematics?" 45
4. "Is There Completeness in Mathematics after G6del?" 62
5. "Hilbert Vindicated?" 84
6. "Standard vs. Nonstandard Distinction: A Watershed in the Foundations of

Mathematics" 106
7. "Standard vs. Nonstandard Logic: Higher-Order, Modal, and First-Order

Logics" 130
8. (with Gabriel Sandu) "The Skeleton in Frege's Cupboard: The Standard

versus Nonstandard Distinction" 144
9. (with Arto Mutanen) "An Alternative Concept of Computability" 174
10. (with Gabriel Sandu) "What is the Logic of Parallel Processing?" 189
11. "Model Minimization - An Alternative to Circumscription" 212
12. "New Foundations for Mathematical Theories" 225

ORIGIN OF THE ESSAYS
The following list indicates the first publication forums of the different essays included
in the present volume (the forthcoming publication forum, if an essay ,lppears here for the
first time):
1. "What Is Elementary Logic? Independence-Friendly Logic as the True Core Area

of Logic", in K. Gavroglu et aI., editors, Physics, Philosophy and the Scientific
Community, Kluwer Academic, Dordrecht, 1995, pp. 301-326.
2. (with Gabriel Sandu) "A Revolution in Logic?", Nordic Journal for Philosophical
Logic (new).
3. "A Revolution in the Foundations of Mathematics?", Synthese vol. 111 (1997), pp.
155-170.
4. "Is There Completeness in Mathematics after Godel?", Philosphical Topics vol. 17,
no. 2 (1989), pp. 69-90.
5. "Hilbert Vindicated?", Synthese vol. 110 (1997), pp. 15-36.
6. "Standard vs. Nonstandard Distinction: A Watershed in the Foundations of
Mathematics", in Jaakko Hintikka, editor, From Dedekind to Giidel: Essays on the
Development of the Foundations of Mathematics, Kluwer Academic, Dordrecht,
1995, pp. 21-44.
7. "Standard vs. Nonstandard Logic: Higher-Order, Modal, and First-Order Logics",
in E. Agazzi, editor, Modern Logic, D. Reidel, Dordrecht, 1981, pp. 283-296.
8. (with Gabriel Sandu) "The Skeleton in Frege's Cupboard: The Standard versus
Nonstandard Distinction", Journal of Philosophy vol. 89 (1992), pp. 290-315. (A
Postscript has been added.)
9. (with Arto Mutanen) "An Alternative Concept of Computability", not previously
published.
10. (with Gabriel Sandu) "What is the Logic of Parallel Processing?", International
Journal of the Foundations of Computer Science vol. 6 (1995), pp. 27-49.
11. "Model Minimization - An Alternative to Circumscription", Journal of Automated
Reasoning vol. 4 (1988), pp. 1-13.
12. "New Foundations for Mathematical Theories", in 1. Vamanen and 1. Oikkonen,
editors, Logic Colloquium 90, ASL summer meeting in Helsinki, Lecture Notes in
Logic vol. 2, Springer, Berlin, 1993, pp. 122-144.
All the previously published essays appear here with the permission of the respective
copyright owners, if any. These permissions are most gratefully acknowledged.
INTRODUCTION
One can distinguish, roughly speaking, two different approaches to the

philosophy of mathematics. On the one hand, some philosophers (and some
mathematicians) take the nature and the results of mathematicians' activities as
given, and go on to ask what philosophical morals one might perhaps find in
their story. On the other hand, some philosophers, logicians and mathematicians
have tried or are trying to subject the very concepts which mathematicians are
using in their work to critical scrutiny. In practice this usually means
scrutinizing the logical and linguistic tools mathematicians wield. Such scrutiny
can scarcely help relying on philosophical ideas and principles. In other words it
can scarcely help being literally a study of language, truth and logic in
mathematics, albeit not necessarily in the spirit of AJ. Ayer.
As its title indicates, the essays included in the present volume represent the
latter approach. In most of them one of the fundamental concepts in the
foundations of mathematics and logic is subjected to a scrutiny from a largely
novel point of view. Typically, it turns out that the concept in question is in need
of a revision or reconsideration or at least can be given a new twist. The results
of such a re-examination are not primarily critical, however, but typically open
up new constructive possibilities. The consequences of such deconstructions and
reconstructions are often quite sweeping, and are explored in the same paper or
in others.
For instance, in essays 1-3 the core subject area of contemporary
mathematical logic, quantification theory, is shown to have been only partially
and inadequately captured by the received first-order logic. The changes
necessitated by this prima facie modest insight in the foundations of
mathematics tum out to be both deep and extensive.
In essay 4, the notion of completeness is subjected to an analysis and its
different variants are distinguished from each other. The results of this quite
elementary analysis show that much of the prolonged discussions of the
implications of Godel-type incompleteness results are not only superficial but
positively misleading. There is much more completeness in mathematics after
COdel than one would ever have gathered from popular expositions.
In essay 5, Hilbert's philosophy of mathematics is viewed from the vantage
point of the results obtained in the other papers. Once again, the popular cliche
of Hilbert as a "formalist" is shown to hide rather than to reveal the true
dynamics of his thought.
Essays 6-8 explore the role of the contrast between standar(i and nonstandard
interpretations of higher-order logics in the history of the foundations of
ix
x LANGUAGE, TRUTH AND LOGIC
mathematics. Its nature and its implications have remained unacknowledged to

a surprising extent, in spite of the importance of the role which it has tacitly
played.
In essay 9, an alternative concept of computability is formulated. Even though
the idea on this new concept is quite simple (it has in fact been utilized ad hoc in
earlier literature), it has important philosophical implications. For instance, it
implies an important qualification to Church's thesis.
In essay 10, the new logic explored in the first few essays is shown to
constitute in a natural sense a logic of parallel processing. In essay 11, computer
scientists' notion of circumscription is interpreted model-theoretically as
minimality .
Finally, essay 12 explores the possibility of varying our logic by imposing
suitable a priori constraints on what is considered a model of a logical formula.
Once again, a familiar concept turns out to allow interesting, theoretically
motivated variation.
I am calling the papers printed or reprinted here essays not because of their
literary format, but because they are essays in the etymological sense: new
exploratory ventures, not unlike Descartes' Essais de eette methode.
As with the earlier volumes in this series, I want to acknowledge my debts to
my co-authors, my secretary and my publisher. The actual editing of this volume
was facilitated by sabbatical support by Boston University and by travel grants
by the Academy of Finland.
1
WHAT IS ELEMENTARY LOGIC?

INDEPENDENCE-FRIENDLY LOGIC AS THE
TRUE CORE AREA OF LOGIC
1. IS ORDINARY FIRST-ORDER LOGIC

THE TRUE ELEMENTARY LOGIC?
My ultimate aim in this paper is to show that conventional wisdom is

seriously wrong when it comes to logic. I want to challenge the tradi-
tional ideas as to what the most basic part of logic is like and how it
ought to be studied. What is more, I will actually prove that these
widely accepted ideas are mistaken.
Now what is the most basic part of contemporary logic, the true
elementary logic? Most philosophers, and most logicians, would undoubt-
edly answer: first-order logic, also known as quantification theory, lower
predicate calculus or - nomen non est omen - elementary logic. This part
of logic was first developed explicitly by Frege. Indeed, the often-
repeated claim that Frege is the true founder of modem logic is largely
predicated on the assumption that first-order logic, that versatile creation
of Frege's, is the true ground-floor part of logic.'
The belief in the status of quantification theory (first-order logic) as
the true elementary (basic) logic is not unanimous. For instance, some
philosophers and mathematicians have argued that second-order logic
is the appropriate universal medium of mathematical theorizing.2 I shall
postpone my comment on the claims of second-order logic to a later stage
of my line of thought. For my present purposes, it suffices to assume that
first-order logic is a part of the genuine basic logic.
This assumption looks so trivial that it is seldom explicitly stated
by philosophers. It is codified in virtually all introductory textbooks of
formal logic which are nothing but expositions of first-order logic
preceded by an ancillary presentation of propositional logic. Occasionally,
however, the claims of first-order logic are stated in so many words.
"If I don't understand first-order logic, I don't understand anything", a
well-known philosopher-logician once said to me in discussion, undoubt-
edly echoing the views of a large number of other philosophers, not
only of some of his Harvard colleagues.
J. Hintikka, Language, Truth and Logic in Mathematics
Springer Science+Business Media Dordrecht 1998
2 LANGUAGE, TRUTH AND LOGIC
The claims of first-order logic to a special status are often reinforced

by, and sometimes probably based on, the idea that it is part and parcel
of the logic of our natural language. The most influential (though
cautious) proponent of this view is probably Noam Chomsky. His version
of the time-honored idea of logical form (Chomsky's LF) is essentially
the idea of a logical form of a first-order formula. 3 Some of Chomsky's
followers have in fact argued for an even more intimate relation between
first-order logic, as illustrated by a recent book title Logic as Grammar. 4
The "logic" intended here is essentially first-order logic.
Elsewhere I have shown that the claims of first-order logic to be the
true Sprachlogik are to a considerable extent unfounded. 5 For instance,
anaphoric pronouns do not behave in natural languages like variables
of.quantification theory, conventional wisdom notwithstanding, but more
like quaint definite descriptions. 6 I shall not review my earlier argu-
ments here, however. It suffices for my present purposes to note that
the claims of first-order logic to a specially central role in logic cannot
be based on its especially close relation to natural language and its
logic.
Another line of argument for the sufficiency of first-order logic as
the true basic logic relies on its alleged capacity of accommodating all
mathematical reasoning. This type of argument proceeds step by step. On
the one hand, it is argued that all mathematics can be carried out in terms
of set theory.' On the other hand, it is argued (or assumed) that set theory
itself can, and should, be thought of as a first-order axiomatic theory.s
Once again, this is too big an issue to be argued adequately here. It
suffices to point out that it is by no means obvious that all mathematics
should be approached via set theory. Not only are there serious alter-
natives to the status of first-order axiomatic set theory as the true logic
of mathematics, such as categories and topoi,9 as well as a suitable
Bourbaki-type approach, 10 I shall in fact try to show on another occasion
that, independently of the claims of such rivals, set theory is not the
natural foundation of mathematical theories. Hence, the self-sufficiency
of first-order logic as the basic logic cannot be argued by reference to
set theory, either.
However, the claims of first-order logic as the true elementary logic
might seem to be safe enough without such supporting arguments from
linguistics or foundations of mathematics. It is an incontrovertible fact
that even those philosophers who plead for the claims of suitable exten-
sions of first-order logic to the title of general logic do usually consider
WHAT IS ELEMENTARY LOGIC? 3
first-order logic as the core area of general logic. It is this belief that I
want to shake.
The first main point I shall argue for is that the usual formulation
of first-order logic incorporates completely arbitrary restrictions. As soon
as you understand the usual form of first-order logic., you ipso facto
understand logical ideas that take you beyond it. They should therefore
be incorporated in our basic general logic on a par with the ideas of
traditional first-order logic. The only reason why they have not been
codified in our ground-floor general logic is a number of arbitrary nota-
tional conventions which have no foundation in the true order of things
or perhaps rather in the true order of logic. Hence, even though the
conventional first-order logic is part of the true elementary logic, it is not
all of it.
2. QUANTIFIER DEPENDENCE AS THE ESSENCE

OF FIRST-ORDER LOGIC
In order to argue for this claim, we have to ask: What is needed to under-
stand first-order logic? What is its conceptual gist? The obvious answer
is: quantifiers. First-order logic does not have the alias "quantification
theory" for nothing.
But this answer is seriously incomplete. There is more to first-order
logic than the license to speak of all members of a domain and of at least
one of its denizens. By means of these two ideas taken separately, we
can obtain little more than the old syllogistic (monadic) logic. Something
else is needed to provide first-order logic the escape velocity needed
to propel it beyond the trivial realm of syllogistic reasoning.
Where does this extra force come from? The answer is not trivial,
although there is not likely to be serious disagreemfmt about it. The
"secret" of quantification theory, the idea that enables us for instance
to formulate highly interesting mathematical theories by its means, is
the notion of dependent quantifier. Because of the availability of this
idea, we can say in an applied quantificational language, not only that
everything is so-and-so or that something is thus-and-so, but that for each
individual there is at least one individual related to it in such-and-such
a way. Thus it is the notion of a dependent quantifier that enables us
to deal with functional relationships by means of first-order logic. It
is not merely that we can admit relations into first-order languages; the
logic of relations would be seriously handicapped without the idea of
quantifier dependence. This is the reason why the order of dissimilar

quantifiers matters, so that (3x)(V'y)Rxy and (V'y)(3x)Rxy have to be
distinguished sharply from each other. It is also the reason why on the
semantical plane we need the notion of satisfaction over and above the
notion of truth.
The notational implementation of quantificational dependence is
carried out by means of the notion of scope. In first-order logic, each
quantifier (Q.) comes fully equipped with a pair of parentheses following
it which define the scope of (Q.). Thus in (Q.)(-), (-) is the scope of
(Q.). Any other quantifier which occurs within the scope of (Q.) depends
on (Q.). Any quantifier which occurs outside the scope of (Q.) is inde-
pendent of (Q.).
The idea of independence is of course familiar to my readers. For
instance, what is said in
(1) (3y)Rya
is that there exists an individual which is related to the individual a in
a certain way. What. is said in
(2) (V'x) (3y)Ryx
is that for each individual, call it x, there exists an individual which is
so related to it. The choice of the value of "y" which makes (2) true
depends on the choice of the value "x".
The insight that the true gist of first-logic lies in quantifier depen-
dencies is not new. Among others, it is discussed in a most illuminating
way by Warren Goldfarb, who argues persuasively that this insight was
one of the motive forces of Hilbert's work in logic, culminating in his
epsilon-calculus. II
The functions that serve to codify truth-making choices of the values
of dependent quantifiers are known as Skolem functions.
3. QUANTIFIER INDEPENDENCE AND ITS CODIFICATION
SO far, everything that has been said has been relatively unsurprising.
How can it be. then, that the innocent-sounding remarks I have made
point beyond the traditional first-order logic? The answer is exceed-
ingly simple. To understand customary first-order logic means essentially
to understand quantifier dependence. To understand quantifier depen-
dence is the same thing as to understand quantifier independence: they
are the two sides of the same conceptual coin. Hence to understand
first-order logic presupposes understanding the idea of quantifier inde-
pendence.
Now how are quantifier dependence and independence expressed in
logic? Bracketing is one partial way of signalling quantifier depen-
dence, but neither the only possible way nor an exhaustive one. In the
light of what has been said, we are not taking a single step beyond
what is needed to master the ordinary first-order logic if we amplify
its notational apparatus by indicating explicitly quantifier independence.
Because of the presence of the old device of parentheses, independence
needs to be expressed only for ordered pairs of quantifiers where the
latter of the two quantifiers would otherwise depend on the former, i.e.,
where the latter occurs within the scope of the former.
For the purpose I propose to introduce a simple slash notation. It
can be explained as follows:
Assume that F is a well-formed formula of the usual first-order logic
containing a quantifier (Qo) which occurs within the scope of quanti-
fiers QI)' (QJ, ... (QJ, plus possibly some others. Then we obtain a
well-formed formula F* by replacing, in F, the quantifier (QO> by
(3) (QrlQlt Q2' ... , QJ.
Naturally, more than one quantifier in F can be independent of others
in this way.
The intended interpretation of F* is determined by the idea that in it
(Qo) or, rather, (3), is to be taken to be independent of (QI)' (QJ, ... ,
(~). Otherwise, F* is to be understood in the same way as F.
How this intuitive idea can be implemented by an explicit seman-
tical treatment will be discussed later. (See Section 8 below.) The result
of carrying out the possibility of independent quantifiers (and of
extending this idea back to propositional logic), starting from the usual
first-order logic, will be called independence-friendly (IF) logic. More
explicitly, such a logic will be called an IF first-order logic. Languages,
including interpreted languages, whose logic is an IF logic, will be called
IF languages.
Somewhat fuller explanations of the syntax and the semantics of IF
first-order languages will be provided later. For my purposes in this paper,
including the definition of IF first-order languages, it suffices to consider
only such formulas whose negation normally do not contain indepen-
dencies other than that of existential quantifiers.
The relation of IF first-order languages to ordinary first-order lan-

guages is thus like that of a game with imperfect information to the
corresponding game with perfect information. As an example, you can
think of the relation of the game of bridge to a simplified form of bridge
in which all the players know the distribution of cards. What is impor-
tant to realize is that if you understand one game, you automatically
understand the other one, too. (Just think of how actual bridge games
are analyzed by reference to diagrams showing all the hands.)
What has been said can be illustrated by means of particular examples.
For instance, consider an ordinary first-order sentence like
(4) ctf'x) (3y)Rxy.
Its meaning is clear to everyone who understands first-order logic.
But what would happen if the existential quantifier (3y) were made
independence of ("Ix)? In other words, how are we to understand the
following?
(5) (Vx)(3yNx)Rxy
Consider, for a moment, what (5) says. Unlike the situation in (4), in
(5) the truth-making choice of the value of "y" does not depend on the
value of "x". Hence it might as well be made prior to the latter choice.
In both cases, the formula which is to be made true is the same, viz,
Rxy. Thus it can be seen that the force of (5) is the same as that of
(6) (3y)(Vx)Rxy.
Thus, anyone who understands ordinary first-order logic will readily
see, not only what (5) means, but that it is logically equivalent with
the ordinary independence-free first-order formula (6). In other cases, the
meaning of IF sentences is equally easy to understand, but they no longer
reduce to ordinary first-order logic. The simplest example of such a
formula is the following:
(7) ('Vx)(Vz)(3yNz)(3ufv'x)S[x, y, z, u].
In (7), the truth-making choice of the value of "y" depends on the value
of "x" but not on that of "z", and vice versa for "u".
It can be seen in a simple fashion that (7) cannot be expressed in a
linear form without the independence indicator. For in such a represen-
tation, the quantifier ('Ix) must precede (3y), for. the latter depends on
the former. Likewise, (Vz) must precede (3u). Again, (3y) must precede
(V'z). for it is independent of the latter. By the same token (3u) must
precede (V'x). These requirements may be symbolized as follows
(8) (V'x) ::> (3y). (V'z) ::> (3u)
(3y) ::> (V'z). (3u) ::> (V'x)
However. they are easily seen to be incompatible with the requirement

of linearity.12
A knowledgeable reader has long since noted that IF logic I am out-
lining comprises what is known as the logic of partially ordered quantifier
prefixes. popularly but inaccurately known as the logic of branching
quantifiers. 13 For instance. (7) is tantamount to what is known in the trade
as the Henkin quantifier. It can be written in a self-explanatory notation
as
(9) (V'x)(3y) '-....

/S[x, y, z, u]
(V'z)(3u)
Such a notation enhances further the case with which IF formulas can
be understood. by displaying graphically the relative dependencies and
independencies of the several quantifiers.
The idea of branching enables me to illustrate in yet another way
the extent to which the principles needed to understand IF logic are
already needed to understand the normal independence-free first-order
logic. In a perfectly good sense, branching quantifier structures have
always been part and parcel of the first-order notation. The branching
of quantifiers to the right has always been admissible and understood
without any special explanations. Quantifiers in different branches
(branching to the right) are simply subordinate quantifiers whose scopes
do not overlap. If one understands rightwards-branching quantifiers
without any special instructions, by the same token one can appreciate
the meaning of leftwards-branching quantifiers without being told how
they are to be understood.' The fact that quantifier structures branching
to the right reduce to the usual linear independence-free quantifier
prefixes (e.g., as in a prenex normal form) does not invalidate my point,
for we do not understand them by means of a translation to a linear
form. 14
However, the IF logic envisaged here is in several respects richer

than the logic of partially order quantifier prefixes. For one thing, the
ideas of dependence and independence can be extended to the proposi-
tional connectives &, V.I~ They, too, can be dependent on each other
and on quantifiers. For them, independence can be indicated by the
same slash notation as in the case of quantifiers with one exception (or,
rather,addition). Since different occurrences of the ampersand or of the
wedge are not always distinguished from each other typographically or
by their location in a sentence, it may be necessary to use subscripts to
keep them apart from each other.
For instance, we may have expressions like the following:
(10) ('v'x)(A[x] (vl'tx) B[xD
(11) (A1[x] (v/&) A[x]) & (B1[x] (v/&) B 2[x])
(12) ('v'x)('v'y)Alty](vl'tx)A2[x,y])(vl'ty)(B 1[x.y ](vl'tx) B2[x,y]).

It easily turns out that the first two of these reduce to ordinary first-
order notation but the third one does not. 16
In a similar way, quantifiers mayor may not be independent of con-
nectives within the scope of which they occur. Such independence can
likewise be indicated by the slash notation.
One can extend the idea of independence even to negation. Nota-
tionally, we can simply extend the slash notation to negation. What the
interpretation of this notation is has to be discussed separately. I shall not
do so in this paper.
Thus it can be seen that understanding quantifier independence is an
integral part of understanding ordinary quantification theory. If the con-
ventional first-order logic is the ground-floor logic, or part of it, then
so is the IF first-order logic.
4. THE SCOPE OF SCOPE
But this result poses a problem. How come, in view of the fact that the
idea of independence is implicit in ordinary first-order logic, that is not
indicated in the usual formulations of this logic? The right answer, it
seems to me, is the notorious Montague reply: historical accident. 17 Or,
perhaps I should say instead: arbitrary choice of notation which unneces-
sarily and artificially restricts what can be expressed in first-order logic.
Above, I hinted at this fact by saying that in the usual first-order logic
we have a way of signalling dependence but not a general way of

signalling independence. This point deserves to be spelled out more
fully.
The joker in the pack of the basic concepts of conventional first-
order logic is the notion of scope. It is one of the received ideas which
are taken virtually" for granted but which soon turns out to involve a
mare's nest of problems. I have spelled out some of those problems
elsewhere. 18 A sense of the problem situation can perhaps be conveyed
to the reader by a quick comparison between customary formal languages
and natural languages. In the former, scope is usually indicated by a
pair of parentheses (brackets) attached to a quantifier or to a proposi-
tional connective. In a natural languages, there are no parentheses
available that could be used in the same way. (Customary parentheses
are not scope indicators, but a discourse device, indicating as it were
that certain contributions are not a part of the speaker's or writer's main
line of thought.) How can natura1language get along without the usual
scope indicators?
A full answer cannot be given here. I have provided one on an earlier
occasion. Instead, a few more specific comments can be made. One
function that parentheses have in the usual formal languages is to indicate
the syntactical limits of binding. This is in fact how the customary
quantificational notation is usually explained. Let (Qx) be a quantifier.
The explicit or implicit parentheses attaching to it, as in
(13) (Qx)(S[xD
demarcate (so we are told) that segment of the formula in which the
variable "x" is bound to that (occurrence of the) quantifier (Qx).
Sometimes this "scope" is compared with the segment of a sentence or
discourse in a natural language in which anaphoric pronouns are "bound"
to a "head. 19
Be this comparison between the variables of quantification and
anaphoric pronouns as it may, we can see at once that the usual quan-
tificational scope notation is arbitrarily restrictive. Accepting for the
moment the idea of binding, we can see that there are a number of unwar-
ranted assumptions built into the customary use of parentheses. Given
a quantifier (Qx), the pair of parentheses indicating its scope are supposed
to demarcate a segment of the formula in which variables are bound to
it. However, the usual notation presupposes that a quantifier scope (i)"
begins immediately after the occurrences of the quantifier in question;
and (ii) is continuous (without lacunae). Moreover, (iii) quantifier scopes

are assumed to be nested, that is to say, assumed never to overlap only
partially. If the scopes of two quantifiers overlap, then the scope of one
of them must lie completely within the scope of this other.
None of these three assumptions is lent any a priori support by the
very explanations that are given of the meaning of parentheses and thus
are supposed to enable us to understand them. For instance (cf. (i
there is no law, logical. legal or divine. that says that the part of a sentence
in which a variable is bound to a given quantifier (Qx) must being imme-
diately after (Qx). This requirement is made awkward already by the
fact that the conventional parentheses notation has systemically speaking
other functions than to indicate the limits of binding. In fact, its first
and foremost function is to show the relative logical priority of the
different quantifiers. Such collateral functions may make it awkward
to assume that the "scope" of a quantifier is always and everywhere
adjacent to it.
Likewise (ii) represents a substantial assumption to require that the
binding range of a quantifier is continuous. It is even fairly obvious
that the corresponding assumption is false in natural languages.20 It is
fairly easy to find examples of a discourse where an anaphoric pronoun
momentarily changes heads, only for the speaker to resume later to
refer to the original head.
Furthennore, (iii). and most importantly, there is no a priori reason
to assume that the ranges of binding for different quantifiers must be
nested. In tenns of dependence. this would mean assuming that quanti-
fier dependencies are always transitive. Why should they be? From game
theory, it is known that informational dependencies of the kind we are
dealing with here need not be transitive. 21 But if they are not, quanti-
fier "scopes" will not be partially ordered, contrary to what they are
assumed to be in the customary notation for quantification theory.
If a quick example is needed for the difficulties that beset the notion
of scope as applied to natural languages. a variant of the so-called Bach-
Peters sentences will serve the purpose. 22 Consider the following perfectly
understandable English sentence:
(14) The boy who was fooling her kissed the girl who loved him.
How are the quantifier scopes supposed to be ordered in (14)? There is
no reasonable way of making them linear in this usual independence-free
first-order logic, no matter how you analyze the definite descriptions.
However, when it comes to logic it might seem that [ am here quib-

bling about a minor point of notation. In order to see whether the point
is in reality so small, let us see what happens if we give up the third
assumption (iii). If we do so, we need some notational convention to
indicate which pair of parentheses or brackets goes together with which
quantifier. This can be done by using different kinds of parentheses or
more generally, by co-indexing parentheses and quantifiers.
Then we can construct well-fonned fonnulas like the following:
(15) ('Vx)3y)[Axy) & ('Vz)3u)[Bzu) & R[y, um.
Of course, (15) could be written more pedantically (and less perspicu-
ously) as
(16) ('v'X)l(l(3Y)z(ZAxy)1 & ('v'z)3(3(3u).(.Bzu)3 & R[y, u]h).
But the $64,000 question still remains: Does the liberalization of the
use of parentheses represent more than a: minor change in the notational
conventions, perhaps something of the same order of (un)importance
as the use of the Polish notation?
The answer is that the notational change has indeed significant con-
sequences. It has to all practical purpose the' same consequences as the
step to an IF first-order language.
This can perhaps be seen from an example. Let us assume that we
have a language that includes the pairing function z =p(x.y) (expressing
the fact that z is the ordered pair (x, y}) and the left and the right
unpacking functions l(z) and r(z). Thus, e.g.
(17) ('v'x)('v'y)('v'z)x = l(z & (y = r(z) H (z =p(x, y
Then we can write out sentences like the following:
(18) ('v'x)3y)[(x =ley~) & ('v'z)3u)[(z =l(z & S[l(y), r(y), l(u),
r(u)]]].
This is seen to be of the same form as (15). But a moment's reflection
also shows that it has the same force as the Henkin quantifier sentence
(7) or (8).
Thus the unavoidable liberalization of the use of parentheses once
again leads us to IF first-order logic. This logic is therefore installed more
and more finnly as the true basic logic.
S. THE FAILURE OF AXIOMATIC METHODS IN IF LOGIC
What has been established so far? It may have disturbed some of my

readers that we have to dethrone conventional first-order logic from its
place of honor as our basic general logic and to install in its stead the
IF first-order logic in the way I have indicated. But this is not the end
of my story. The most important shock (or, for the happy few, the reas-
surance) is still to come. We encounter it when we actually begin to study
the IF first-order logic sketched above, to gain a general view of it, to
develop a metatheory for it, etc. How can we to do so? Here comes
the real shocker: We cannot study IF first-order logic satisfactorily by
means of the most common tools of customary logical theory. For what
are those tools? On the formal (syntactical) level, the most important con-
ceptual tool is the deductive-axiomatic method. What that means is that
we try to enumerate recursively all valid logical truths as theorems deriv-
able from formally characterized axioms by means of purely formal rules,
the so-called "rules of inference". As GOdel showed in 1930, all logical
truths of the ordinary first-order logic can be so derived, i.e., there are
complete axiomatizations of first-order logic.
It quickly turns out however, that the IF first-order logic is not axiom-
atizable. Perhaps the easiest way of seeing this intuitively is to see what
happens if you try to treat higher-order logics simply as many-sorted first-
order 10gics.23 You can capture everythingsalva satisfiability in first-order
terms except the idea that there must be a higher-order entity (of a suitable
type) corresponding to each extensionally possible class of (n-tuples
of) lower-order entities. For instance, there must be (on the standard
interpretation of second-order logic) a second-order entity corresponding
to each class of individuals.
This requirement can be expressed by means of one single second-
order universal quantifier. Hence the decision problem (for satisfiability)
for a higher-order logic can be reduced to the decision problem for
second-order sentences of the form
(19) ('<tX)S[X]
where the initial universal quantifier is the only higher-order quan-
tifier.
Now it can be shown (this is implicit in Enderton's results)24 that
each sentence (19) is equivalent with the contradictory negation of an
IF first-order sentence. Hence the decision problem of validity for higher-
order logic reduces to that of IF logic. From this it follows that IF logic
cannot be axiomatizable.
6. QUALMS ALLAYED
It is not that the failure of the axiomatic-deductive techniques in (a special

case of an) IF logic has gone unnoticed. In fact, Quine has relied on
this very failure in his attempt to discredit in effect IF logic as being
an integral part of logic in contradistinction to mathematics. 25
Quine's argument is seriously inconclusive. He tries to conclude from
the possibility of representing the branching quantifier formula
(20) ('<i x)(3y)
('<iz)(3u)
'"
/
S[x, y, z, u]
in a second-order form
(21) (3j)(3h)('<ix)('<iz)S[[x,j(x), z. h(z)]
that (20) is in reality a higher-order (and hence, according to Quine's
lights, "mathematical") statement. But precisely in the same way we
can represent any first-order statement in a second-order form. Does
that show that all of first-order logic is really "mathematics"? If QUine's
argument is valid, it would show just that, providing us with a neat
reductio refutation of Quine's position.
What is wrong with QUine's argument is that it confuses the question
as to what is needed to understand a branching-quantifier sentence with
the question as to what is needed to master the set of logical truths of
the theory of branching quantifiers. It is virtually a confusion between
the notions of truth and logical truth. The axiomatizability (or non-
axiomatizability) of a branch of logic pertains to the latter, that is, to
the treatment of logical truths, not to the former, that is, to the treat-
ment of truth simpliciter. Yet it is the latter question alone that is relevant
to the question of understanding a language. To understand a sentence
is to know what things are like when it is true, not what they are like
when it is logically true. Thus it is the question of understanding of actual
truth of sentences that is crucial here. One could even turn the tables here
on Quine and recall his reservations about any sharp distinction between
analytic and synthetic truths. What is so especially important about the
axiomatization of logical truths if there is no important distinction
between them and the synthetic (empirical) ones? Complete axiomati-
zation fails quickly outside pure logic anyway. Why should there be
any problem about that in logic?
I suspect that the likes of Quine are suffering from methodological
insecurity. If we cannot study logic by means of the axiomatic-deduc-
tive method, they seem to think, there is nothing equally clear-cut and
familiar that we can resort to - or so it seems. The antidote to this inse-
curity is obviously to show that there are plenty of ways of handling
and even mastering IF logic, even though its logical truths are not
recursively enumerable. Indeed, that very nonaxiomatizability makes it
possible to use IF logic for what I consider the most important task of
logical theory: the discovery of successively stronger and stronger logical
principles. 26
This argument might seem to be too sophisticated, too ad hominem,
to be convincing. It is nevertheless based on a solid insight.
The right method of freeing philosophical logicians from this method-
ological insecurity would be to show how IF logic can be successfully
studied notwithstanding the absence of a completely axiomatization.
On another occasion, I shall try to do so and even to indicate how IF
logic can be employed to facilitate the discovery of new logical princi-
ples.
7. THE FAILURE OF TARSKI-TYPE TRUTH-DEFINITIONS

IN IF LOGIC
We have discovered that axiomatic-deductive methods cannot serve to

build an exhaustive theory of IF first-order logic. But perhaps semantical
(model-theoretical) methods can do the trick? Well, what are those
methods? Ninety-nine percent of my readers will answer: They are
methods based on Tarski-type truth-definitions. 27 And now comes the
next shock: Tarski-type truth-definitions are not applicable to IF first-
order logic, either.
Why not? The reason is that Tarski-type truth-definitions rely on
another widely (but not universally) accepted principle of language
theory, a principle whose role has been thought to be to keep, if not
the body and the soul of language theory together, than at least to keep
syntax and semantics together. It is most commonly called the prin-

ciple of compositionality.28 Among its aliases is "Frege Principle". In one
of its' formulations it says that the meaning of a complex expression is
a function of the meanings of its constituent parts. In practice, what it
amounts to is an inside-out principle that implies a kind of semantical
context-independence. If the meaning of a complex expression depends
only on the meanings of its parts, it can never depend on its context in
a still more comprehensive expression. Thus much of the actual force
of the principle of compositionality lies in ruling out semantical context
dependencies. 29
Whether the principle of compositionality can always be honored in
one's semantical theory, come hell or high water, will not be discussed
here. Suffice it to observe that in an obvious sense a Tarski-type truth-
definition presupposes compositionality. In such a truth-definition, the
concepts of truth and satisfaction for a complex expression are defined
recursively in terms of the truth and satisfaction of certain simpler expres-
sions. In brief, a Tarski-type truth-definition works its way from inside
out, and hence cannot accommodate any real semantical context-depen-
dence.
But the very idea of quantifier independence violates the principle
of compositionality. For the force of an independent quantifier (QlQ2)
depends on another quantifier (Q2) which occurs, not within the scope
of the former, but outside this scope, in other words in the context of
(QlQ2). This clearly violates compositionality, and it. is the ultimate
reason why Tarski-type truth-definitions do not work in an IF first-
order 10gic. 30 By the same token, any approach to formal semantics,
such as Montague semantics, which relies on compositionality, is inca-
pable of handling IF quantifiers, at least without some special measures,
such as taking refuge in higher order logic or some comparable counsel
of despair.
Thus the role of IF first-order logic as the true elementary logic has
a general methodological moral. It suggests that we should not put much
trust in the principle of compositionality. This suggestion pertains in
the first place to the semantics of formal languages. However, to the
extent we can argue that an IF logic is a better framework of seman-
tical representation for natural languages than ordinary first-order logic,
to the same extent the injunction against compositionality applies also
to the semantical treatment of natural languages.
Although the point will not be argued here, there is in fact ample
evidence to suggest that independence of the kind we are dealing with

here is a pervasive phenomenon in natural languages. Hence the status
of compositionality in IF logics is relevant also to the semantics of natural
languages.
In general, the role of informational independence in natural languages
needs a closer study than can be devoted to it here. In particular, we need
a fully argued-for answer to the question: Why did the presence of
informational independence in natural languages elude so long the atten-
tion of linguists, logicians, and philosophers? The answer that seems
to be current is: Because it is a relatively rare and marginal phenomenon.
There is reason to think that the true answer is almost diametrically
opposite. 31 It seems to me that informational independence is one of
the most common and most spontaneously understood aspects of the
semantics of natural languages. It escaped attention partly because of
its very familiarity and of the very ease of its interpretation. An impor-
tant contributing factor is that independence is not uniformly marked
syntactically in natural languages. However, this absence of syntactical
expressions of independence admits of an interesting explanation, and
hence does not count as an argument against the ambiguity and natu-
ralness of the phenomenon of independence in natural languages. 32
In any case, the failure of compositionality in formal languages already
has interesting consequences. One thing it shows is that the so-called sub-
stitutional interpretation of quantifiers does not work even in elementary
logic, now that we have seen that this elementary logic has to include
IF first-order logic. 33 The reason is that the substitutional interpretation
is usually taken to conform to the principle of compositionality. If so, the
question whether x, y, z and u satisfy the formula S[x, y, z, u] as it
occurs in (7) or in its independence-free variant
(22) (\fx)(\fz)(3y)(3u)S[x, y, z, u]
depends only on its substitution-instances S[a, b, c, d]. But the relevant
substitution-instances are different in (7) and in (22) and they are not
shown by the formula S[x, y, z, u] alone. You also have to know the
dependencies and independencies between the quantifiers to which
x, y, z and u are bound.
This failure of our most cherished methods in systematic logical theory
undoubtedly is not easy to countenance. Sixty years after Godel's results
we have perhaps become prepared to come across logics that cannot
be axiomatized, but even then it is a surprise to see that our most basic
logic is among them. And several logicians and philosophers have been
so impressed by Tarski-type truth-definitions that they consider them
the alpha and the omega of logical semantics. For them, it is bad enough
a shock to find that Tarski-type truth-definitions cannot be used (at least
not in any direct manner) in such exotic looking logics as the theory
of branching quantifiers or infinitely deep logics. So much the worse
a shock if Tarski-type truth-definitions cannot be used in the most
fundamental region of logic which does not go beyond concepts which
we have learned to understand and to master so completely that we are
virtually taking them for granted.
8. GAME-THEORETICAL SEMANTICS TO THE RESCUE
In spite of the failure of Tarski-type truth-definitions in a first-order IF

logic, this logic can easily be treated semantically. This fact should
give us a pause, for it shows strikingly that Tarski-type truth-defini-
tions do not constitute a universally applicable approach to logical
semantics nor indeed the most basic approach.
One approach that is applicable to IF logics is what I have called game-
theoretical semantics (GTS).34 On an earlier occasion, I have suggested
that, far from being a contrived and artificial approach to semantics, it
is merely a systematization of the ways in which mathematicians (and
other folk, for that matter) spontaneously speak and think of (depen-
dent) quantifiers. 3s (Subsequently, of course, this systematization has been
extended to the semantics of expressions other than quantifiers.)
The need to resort to GTS in IF logic further strengthens the justifi-
cation of this way of looking at GTS. That GTS had not been formulated
explicity any earlier can be attributed to its very naturalness. It was
literally too obvious a way of thinking to have caught the attention of
mathematicians and logicians. It is nevertheless significant that whenever
Tarski-type methods failed, logicians spontaneously resorted to game-
theoretical conceptualizations well before the explicit development of
GTS. Examples are offered by the theory of partially ordered quantifiers,
game quantifiers, and Godel-type functional interpretations.
Epistemologically, GTS can be thought of as a way of relating the
truth-conditions of a sentence S to the processes by means of which
the truth or falsity of S can be ascertained. 36 They are in GTS con-
ceived of as games between an initial verifier or "myself" and an initial
falsifier or "nature". The game associated with S will be called G(S).
It starts with S, and its rules can be gathered from the basic ideas just
adumbrated. They can easily be formulated for the ordinary first-order
logic.
A background explanation is nevertheless needed. The game G(S) is
played on a given model M of the language (Ls) of S. That M is a
model of (Ls) implies that all the nonlogical symbols of (Ls) have been
interpreted on M. This means that all the atomic sentences of (LJ have
been assigned a definite truth-value. The same goes for all the atomic
sentences of any extension of (LI ) obtained by adjoining to it a finite
number of names of the members of the domain do(M) of M. Thus the
function of a game-theoretical truth definition, like a Tarski-type truth-
definition, is in effect to extend the concepts of truth and falsity from
atomic sentences to all others.
The game rules can be formulated as follows:
CR.v) G(S. v S2) begins with a choice of Sj (i = 1 or 2) by myself.
The game is continued as in G(Sj)
(R.&) G(S. & S2) begins with a choice of Sj (i =1 or 2) by nature.
The game is continued as in G(S;)
(R.E.) G3.x)So[x]) begins with a choice of a member of do(M)
by myself. Let "b" be a name of the individual chosen. Then
the game is continued as in G(So[b D.
(R.U) Likewise, except that nature chooses b.
(R.-) G( -So) is like G(So), except that the roles of the two players,
as defined by these rules, are reversed.
(R.A) G(A) is won by myself (and lost by nature) if A is true. If
A is false, nature wins and myself loses.
(R.true) S is true iff there exists a winning strategy for myself in
G(S).
(RJalse) S is false iff there exists a winning strategy for nature in
G(S).37
It is easy to see that in ordinary first-order logic this truth-definition
agrees with the usual (Tarski-type) truth-definition, if it is assumed that
all (mathematically possible) strategies are allowed for myself. If
myself's strategies are restricted to recursive ones, we obtain a dif-
ferent concept of truth and falsity. This is one of the ideas on which
Godel's famous functional interpretation of first-order liogic and first-

order arithmetic is based. 38
In the rules so formulated it was tacitly assumed that the games G(S)
are games with perfect information. What happens in IF logic is simply
that this assumption is given up. The slash notation will then be a way
of specifying the information sets of the different moves (in the usual
game-theoretical sense of the term). For instance, the information set
of
(24) (QJQI' Q2' ... , Qx)
(where Qo, QI' Q2' ... , QJ are quantifiers or connectives) consists in
moves connected with all the quantifiers and connectives within the scope
of which (24) occurs, except for QI' Q2' ... , Qk.
Through this simple stipulation GTS can be extended to IF languages.
Thus we have explained completely how an IF first-order language
is to be interpreted. Each sentence of such a language has been assigned
a definite truth-value.
9. THE GENERALITY OF THE PHENOMENON OF

INFORMA TION INDEPENDENCE
At the same time, game-theoretical semantics shows that the phenom-

enon of independence is not a curious, isolated feature of the behavior
of quantifiers. It is pervasive phenomenon which can come up in virtu-
ally any part of the semantics of formal as well as natural languages.
In order to see this, you only have to realize the general explanatory
strategy involved in a game-theoretical approach to our Sprachlogik.39
This strategy consists in associating a rule for semantical games with
a wide variety of different logical and nonlogical symbols (words, con-
structions, formal symbols, etc.). The idea is to capture the semantics
of the symbol in question by means of this game rule.
This strategy can be shown to work in a wide variety of different cases.
A game rule is naturally associated with the usual truth-functional con-
nectives, as illustrated by (R.v), (R.&) and (R.-) above. Likewise, a game
rule can be associated with many of the logical or nonlogical symbols
whose introduction gives rise to well-known extensions of first-order
logic, including the symbols for necessity, knowledge (knowing that),
belief, obligation, etc. Indeed, the possibility of so doing is implicit in
the usual possible-worlds semantics for these notions. Indeed, opera-
tors like necessity are in that semantics in effect universal quantifiers

ranging over alternative worlds. 40
It is not my purpose in this work to discuss in detail the game-theo-
retical treatment of the different extensions of first-order logic. In part,
such a treatment is typically quite obvious. In part, it is (or will be)
actually worked out elsewhere. Even in cases where the role of game-
theoretical ideas is not obvious, there are explicit examples of how they
can be used in a highly interesting manner. The so-called functional inter-
pretations of first-order logic and arithmetic, launched by Godel in
1957, are cases in point,41 as was first pointed out by Dana Scott in
1968.42
In a different direction, when GTS is applied to natural languages, a
game rule not only can but must be associated with a number of further
words, including pronouns43 and prepositions ..... Furthermore, it can be
argued that a move (and hence an application of a game rule) must be
associated even with our most primitive expressions, proper names and
words expressing primitive predicates. 4s The background of this prima
facie surprising phenomenon is possible-worlds semantics, where the
value (interpretation) of any symbol, including a proper name, must be
specified for each relevant possible world.
The same reason is operative in formal languages, where a move
thus can be associated with the interpretation of a proper name in each
model or "world" we may have to consider.
We do not have to consider the details of the rules governing such
moves. The only fact that is relevant here is the existence of a game
rule governing the move in question. For as soon as the semantics of a
symbol is handled by means of a game rule, the notion of informational
independence is ipso facto applicable. In the theory of natural lan-
guages, this leads to the expectation that informational independence is
a widespread phenomenon, which rears its head in a wide variety of
different categories. That this expectation is in fact fulfilled, has been
argued extensively by laakko Hintikka and Gabriel Sandu. 46
In formal languages, this means that the slash notation can be applied
to expressions of all sorts of different logical types, with an interpreta-
tion which is obtained automatically as a corollary to any satisfactory
game-theoretical treatment of the semantics of the formal language in
question. 47
We have already relied on this line of thought in applying the slash
notation to propositional connectives.48 Of the multitude of extensions
of IF first-order logic, I mention here only IF (first-order) lepistemic logic.

It turns out to be especially important because of its applications to
epistemology 49 and to the logic and semantics of questions and answers.~o
In such extensions, there is a game rule associated with each of the
additional ingredients of the language, e.g., with the epistemic operator
"K." which is to be read "x knows that". Hence we must: consider addi-
tional types of independencies. For instance, an existential quantifier
can be independent of K, as in
(23) Kb(Vx) (3y!Kb )S [x, y]
It turns out that such existential quantifiers independent of a sentence-
initial K playa major role in the theory of questions and answers in
natural languages.~l
Thus IF first-order logic can stake a claim to being the true basic logic
also in the sense that it gives rise to various extended IF logics pre-
cisely in the same way in which ordinary independence-free first-order
logic can be extended in different directions.
But are there restrictions as to how far IF first-order logic can be
extended? It turns out that the main restriction is to be found in a sur-
prising direction: negation. As a preparation for a study of this restriction,
I shall offer here a couple of comments on the game-theoretical truth-
definition. The rule (T.-) defines the concept of negation - or, perhaps
I should say, a concept of negation - for IF first-order languages, by
defining uniquely (jointly with the other rules, of course) what it means
for a negated sentence to be true. In the usual independence-free first-
order logic this concept of negation agrees with the usuall one. However,
it is easy to see that in the IF first-order language the game-theoretical
concept of negation defined by (R.-) is not the contraditctory negation.
For this reason, it has to be handled with care. For instance, even though
the law of contradiction holds for it, the law of excluded middle does
not do so in general.
A closer examination nevertheless shows that the failure of tertium
non datur is practically the only major change caused by the step from
ordinary first-order logic to IF one. Practically all the "nice" metatheo-
retical results that do not depend on the presence of contradictory
negation hold in IF first-order logic, such as compactness, upwards
Lowenheim-Skolem theorem, axiomatizability of logically false sen-
tences, separation theorem, Beth's theorem on definability. Some of these
results hold even in a stronger form. For instance, in the separation
theorem the separating formula can always be chosen to be an ordinary

first-order one.
The exploration of all these opportunities is too large a task to be
undertaken in one paper, however.
NOTES
Even though this paper does not come as close to Bob Cohen's central interests as I
hoped, its history perhaps makes it a suitable tribute to him. It was written at his request
and presented under the auspices of Bob's Boston Colloquium for the Philosophy of
Science as my unofficial inaugural lecture at Boston University, where I had moved to,
to a large extent because of Bob's persuasiveness. lowe him much more than this modest
paper can even symbolize.
I The gradual development of the conception of first-order logic is an intricate subject
of which the final truth has not yet been told. Meanwhile, you can have a glimpse of
the problems from studies like Gregory H. Moore, "The Emergence of First-Order Logic",
in William Aspray and Philip Kitcher, editors, History and Philosop'ty of Modern
Mathematics (Minnesota Studies in the Philosophy of Science, vol. II, University of
Minnesota Press, Minneapolis. 1988, pp.95-135). Moore's paper is to be read with caution,
however, for he is unaware of some of the most important conceptual points concerning
the idea of flfst-order logic. For one thing, he does not even mention Henkin's distinc-
tion between standard and non-standard interpretations of higher-order logic. Yet it is
possible to reconstruct a higher-order language, with a suitable non-standard interpreta-
tion, as a many-sorted "flfst-order" language, as far as logic is concerned. Cf. also below,
especially sec. 16.
For some reasons for the ambivalence of Frege' s status, d. Jaakko Hintikka "The Place
of C. S. Peirce in the History of Logical Theory", forthcoming.
2 Cf., e.g., Stewart Shapiro, "Second-order Languages and Mathematical Practice",
Journal of Symbolic Logic SO (1985), pp. 714-42; Georg Kreisel, "Informal Rigor and
Completeness Proofs", in Imre Lakatos, editor, Problems in the Philosophy of Mathematics,
North-Holland, Amsterdam, 1967, pp. 138-86.
3 See here Jaakko Hintikka, "Logical Form and Linguistic Theory", in Alex George,
editor, Reflections on Chomsky, Basil Blackwell, Oxford, 1989, pp. 41-57.
4 Norbert Hornstein, Logic as Grammar, The MIT Press, Cambridge, 1984; cf. Robert
May, Logical Form, The MIT Press, Cambridge, 1985.
~ See note 3 above and also Jaakko Hintikka and Gabriel Sandu, On the Methodology
of Linguistics: A Case Study, Basil Blackwell, Oxford, 1991.
6 See Jaakko Hintikka and Jack Kulas, Anaphora and Definite Descriptions, D. Reidel,
Dordrecht, 1985.
7 Cf. statements like the following: "Among the many branches of modem mathematics
set theory occupies a unique place: with rare exceptions the entities which are studied
and analyzed in mathematics may be regarded as certain particular sets or classes of
objects." (See Patrick C. Suppes, Axiomatic Set Theory, Dover, New York, 1972, p. 1.)
8 The way in which first-order logic came to be considered the right foundation for set
theory is described in Gregory H. Moore, op. cit., note 1 above.
9 See, e.g., S. MacLane, Categoriu for the Working Mathematician, Springer-Verlag,

Berlin, 1971; Jon Barwise, editor, Handbook of Mathematical Lo~'ic, North-Holland,
Amsterdam, 1977, chapters A8 and 06.
10 See, e.g., Jean Dieudonn~, A Panorama of Pure Mathematics. As seen by N. Bourbaki,
Academic Press, New York, 1982 (with further references to the literature).
11 Warren Goldfarb, "Logic in the Twenties: the Nature of the Quantifier", Journal of
Symbolic Logic 44 (1979), pp. 351-68.
12 This line of thought actually gives you a general method of dec:iding as to when a
first-order formula with informationally independent quantifiers reduces to the linear form.
13 Cf. here the brief bibliography of Jaaklco Hintikka and Jack Kulas, The Game of
lAnguage, D. Reidel, Dordrecht, 1983, pp.300-3.
14 It can be argued that, on the contrary, what the cash value of first-order sentences
is, is brought out more clearly by pushing quantifiers deeper and deeJ)(:r into the formulas,
as in the distributive normal forms. This normally increases left-ta-right branching, not
decreases it.
For distributive normal forms, see Jaakko Hintikka, Distribute Normal Forms (Acta
Philosophica Fennica.. vol. 6), Societas Philosophica Fennica.. Helsinki, 1953. A good expo-
sition is also contained in Veikko Rantala, Aspects of DeJinability (Acta Philosophica
Fennica, vol. 29, nos. 2-3), Societas Philosophica Fennica.. Helsinki., 1977.
U See Gabriel Sandu and Jonko VlUtnilnen, "Partially Ordered Connectives", Zeitschrift
fUr mathematische Logik und Grundlagen der Mathemati1c 38 (1992), pp. 361-372; Gabriel
Sandu, "On the Logic of Informational Independence", Journal of Philosophical Logic
22 (1993), pp. 29-60.
16 See the papers referred to in note 15.
17 According to the story (I was not an eyewitness), Richard Montague was once
criticized on the grounds that his universal grammar did not explain the properties of actual
languages (humanly possible languages) as distinguished from logical languages and from
computer languages. ''Why is it that we speak English and not ALGOL?" - Montague's
sincere reply was, "Historical accident."
One can challenge Montague's reply, but in the case of the usual scope conventions
of formal logic this notorious answer can be argued to be literally true; cf. below.
IS Jaakko Hintikka, "Is Scope a Viable Concept in Semantics?", in ESCOL '86:
Proceedings of the Third Eastern States Conference on Linguistics, ESCOL, Columbus,
OH, 1987, pp. 259-70.
19 This is in effect the way quantifier scopes are explained in textbooks of logic. In
Hintikka and Kulas (note 6 above), it is shown that it embodies a seriously mistaken
view of the way pronouns and quantifiers operate in natural languages.
20 This point is argued in Hintikka and Kulas, note 6 above, especially pp. 137-8.
21 The information set of a move may be any subset of the set of c:arlier moves. There
is hence no reasons in general why the information sets of successive moves should be
ordered even partially.
As it happens, in the special case of ordinary quantifiers, all possible types of infor-
mational independence can be shown to reduce to partial ordering. However, this special
case is not representative of the general conceptual situation.
22 Bach-Peters sentences were first introduced as primafacie counterexamples to certain
types of generative grammars. Independently of this original use of theirs, they can be
employed also as counterexamples of certain types of rules for translating from natural
languages to a logical notation.
13 Cf. here Jaakko Hintikka, "Reductions in the Theory of Types", Acta Philosophica
Fennica 8 (1955) pp. 57-115.
2. H. B. Enderton, "Finite Partially-Ordered Quantifiers", Zeitschrift fUr nwthenwtische
Logik IUId Gnmdlagen der Mathemotik 16 (1970), pp. 393-7.
2S W. V. Quine, Philosophy of Logic, Prentice-Hall, Englewood Cliffs, N.J., 1970,
p.91.
Quine bases his negative attitude to branching quantifiers on the alleged fact that
the loss of completeness makes a big change in the resulting logic. But this is a strictly
circular argument. On the contrary, the preservation of such important results as the
separation theorem in an IF logic can be used to argue that the inevitable step from ordinary
first-order logic t<? an IF first-order logic is not a momentous step from the vantage
point of a deeper and more sophisticated logical theory. Cf. below, secs. 10 and 12.
26 This task of logical theory has not been given its proper due. It is perhaps consid-
ered most often in a context of an "experimentalist" and antirealist view of mathematics.
But, as GOdel's example shows, it can be combined with a realist and even Platonist
attitude to mathematical b'Uth. Indeed, semantical (and in that sense "realistic") consid-
erations are especially well suited to sharpen our ideas about possible stronger principles.
An example of how the methods used in this work (cf. sec. 8 below) help to motivate
stronger logical principles is provided by the second-order translation to be outlined in
sec. 11. The translation of
('Vx)(3y)S[x. y]
is
(3f)('Vx)S[x. /Xl.
If you simply require the two to be equivalent, you receive a form of the principle of
choice.
27 These methods go back to Alfred Tarski. "The Concept of Truth in Formalized
Languages", in A. Tarski, Logic. Senwntics. Metamathematics, Clarendon Press. Oxford.
1956. ch. 8. (The Polish original was drafted in 1930 and published in 1933; the German
version appeared in 1935.)
2& Cf. here Barbara Partee, "Compositionality", in F. Landtrnan and F. Veltman, editors.
Varieties of Formal Senwntics, Foris, Dordrecht, 1984, pp. 281-313; Barbara H. Partee,
Alice ter Meulen and Robert E. Wall, Mathenwtical Methods in Linguistics, Kluwer,
Dordrecht, 1990, section 13 ..1 (pp. 317-38).
29 See here Jaakko Hintikka, ''Theories of Truth and Learnable Languages", ch. 10 in
Hintikka and Kulas (1983), note 13 above, pp. 259-92.
30 Cf. here Jon Barwise, "On Branching Quantifiers in English", Journal of Philosophical
Logic 8 (1979), pp. 47-80.
Of course, linearity can be restored by moving to higher-order translations. But it
can be argued that the price of this move is far too heavy if one's aim is to develop a
psycbolinguistically viable theory of what is going on in ordinary discourse.
31 See here Jaalcko Hintikka and Gabriel Sandu, "Informational Independence as a
Semantical Phenomenon", in J. E. Fenstad et al., editors, Logic" Methodology and

Philosophy of Science VIfI, Elsevier, Amstcrdam, 1989, pp. 571-89; Gabricl Sandu, "On
the Logic of Informational Indcpcndencc", Journal of Philosophical Logic 22 (1993)
pp.29-60.
32 See here Jaakko Hintikka, "Paradigms for Languagc Theory", Acta Philosophica
Fennica vol. 49 (1990), pp. 181-209.
II This "interprctation" has been entertained inter alia, by Stanislaw Lesniewski, W.
V. Quine (with reservations), Ruth Marcus and Saul Kripke. Here we finally have
a conclusive argument to show wby it is not a viable account of the semantics of
quantifiers.
Cf. W. V. Quine, Ontological Relativity, Columbia U.P., New York, 1969, pp. 63-7,
104-8; Ruth Marcus, "Modalities and Intensional Languages", Synthese 13 (1961), pp.
303-22; Saul Kripkc, "Speaker's Reference and Semantic Reference", in P. A. French
et al., editors, Midwest Studies in Philosophy vol. 2, University of Minnesota Press, Morris,
MI, 1977, pp. 255-76; D. Grover, J. Camp and N. Belnap, "A Presentential Theory of
Truth", Philosophical Studies 27 (1975), pp. 73-125; T. Baldwin, "Can There Be a
Substantive Theory of Truth?" Recherche sur la philosophie et Ie langage 10, Universit6
des Sciences Sociales de Grenoble, Grenoble, 1989.
Cf. here also D. Gottlieb, Ontological Economy: Substitutional Quantification and
Mathematics, Oxford University Press, Oxford, 1980.
In my judgement, the time has come, not only to kill the so-called substitutional
interpretation, but to bury it.
34 See herc Hintikka and Kulas (1983), note 13 above; Esa Saarinen, editor, Game-
Theoretical Semantics, D. Reidel, Dordrecbt, 1979.
" Cf. Jaakko Hintikka, "On the Development of the Model-Theoretic Viewpoint in
Logical Theory", Synthese 77 (1988), pp. 1-36.
16 As sucb, tbey are systematizations and further developments of certain types of
"language-games" in the sense of Ludwig Wittgenstein. (For them, see Merrill B. Hintikka
and Jaillo Hintikka, Investigating Witlgenstein, Basil Blackwell, Oxford, 1986.) The
failure of the soi-disant Wiugensteinians to put his important and promising concept to
systematic use is but one example of their failure (or refusal) to recognize the construc-
tive potentialities of Wittgenstein's ideas.
37 Clauses (R. true) and (R. false) can be viewed as the true core in the often misused
idea tbat a proposition is true if it "can be verified" or if "there exists a method of
verifying it". Here the existence of a winning strategy explicates the idea of "can be
verified" or "there exists a method of verification", which was left confused or miscon-
strued in the earlier uses of the same idea.
There might seem to be an alternative to (R. false) namely to define S to be false iff
it is not true. This would make it impossiblc to deal with thc falsity of propositions directly
by mcans of game rules, bowever. Below, it will be seen what other consequences the
choice between the two conceptions of falsity has; see especially sec. 15.
38 See Kurt GMel, "On a Hitherto Unexploited Extension of the Finitary Standpoint",
lournal of Philosophical Logic 9 (1980), pp. 133-42. (Translation of GOdel's original
1956 article in Dialectica 11, pp. 280-7, with a bibliography.) See also Kurt GOdel,
Collected Works vol. 2 (edited by Solomon Feferman et al.), Oxford University Press,
Oxford, 1990, pp. 217-53.
19 This is the leading idea of the applications of game-theoretical semantics to natural

languages. For examples of this strategy, see Hintikka and Kulas (198S), note 6 above.
40 This is the gist of the semantics of model logics originally developed by Kanger,
Montague and Hintikka. However, necessity-type operators do not interact with ordinary
quantifiers quite in the same way they interact among themselves. The reason is that
, each possible alternative world (i.e., each value of a necessity-type "universal quanti-
fier") has a domain of individuals (i.e., range of values of ordinary quantifiers) of its
own, possibly different from others. This fact has important consequences. For instance,
not all types of inter- and independencies can any longer be reduced to partial order-
ings.
41 See note 38 above.
41 Dana Scott, "A Game-Theoretical Interpretation of Logical Formulae", McCarthy
Seminar, Stanford, 1967, publisbed in the Yearbook 1991 of Kurt GiJdel Society, Vienna,
1993, pp. 47-48.
4) See Hintikka and Kulas, note 6 above, especially chapter 2.
44 Op. cit., especially pp. 172-9.
45 Op. cit., especially pp. S7, IS9.
See note 31 above.
7 In other words, the slasb simply temporarily exempts a quantifier, connective or
some other expression from the scope of another quantifier (or similar) within whose scope
it would otherwise be.
.. Cf. note IS above.
49 For instance. different constructions in terms of knows cannot be analyzed without
independence-friendly logic. Cf. here Jaakko Hintikka, "Different Constructions in Terms
of 'Knows'", in Jonathan Dancy and Ernest Sosa", editors, Companion to Epistemology,
Basil Blackwell. Oxford, 1992. pp. 99-104.
The need for independence-friendly logic has further repercussions in epistemology
and philosophy of science.
~ What is true of subordinate questions with knows is true of questions in general (cf.
note 49): independence-friendly logic is needed to cope with their logic and semantics.
51 For instance. (23) is the logical fonn of such wh-questions in natural languages as
have an outside universal quantifier, e.g.
Whom does each person admire most?
There are also natural-language questions whose analysis requires independent proposi-
tional connectives. e.g.
Whom does each person admire more, his or her father or his or her mother?
2
A REVOLUTION IN LOGIC?*
Logic might at first sight seem an unlikely arena for revolutions, even for
revolutions in Jefferson's sense rather than Lenin's. Kant maintained that
Aristotelian logic had not changed in two thousand years and could never
change.! Even though Kant's view of Aristotle's logic has been thoroughly
discredited (even by Aristotle's own standards) both systematically and
historically,2 most contemporary philosophers and linguists are adopting the
same Kantian attitude to Frege's logic, or strictly speaking rather to that
part of Frege's logic that has come to be known variously as first-order logic,
quantification theory, or lower predicate calculus. When a critic once
suggested to a prominent philosopher of language that one of the
cornerstones of ordinary first-order logic, Frege's treatment of verbs for
being like is, does not capture the true Sprachlogik, he looked at the speaker
with an expression of horror - mock horror, we hope - and said, "Nothing is
sacred in philosophy any longer!"
Yet Frege's formulation of first-order logic contains a fundamental error. It
is manifested already in his formation rules. And this same virus has
infected all the subsequent formulations and versions of first-order logic. In
order to see what Frege's mistake is, we have to go back to the basics and ask
what first-order logic is all about. The answer is obvious. First-order logic is
about quantifiers. It is not for nothing a.k.a. quantification theory. But what
is often overlooked is that quantification theory is not a study of quantifiers
in splendid isolation from each other. If you use quantiiiers one by one
unrelated to each other, the only logic you end up with is monadic first-order
logic or some other mild generalization of Aristotelian syllogistic. The real
power of first-order logic lies in the use of dependent quantifiers, as in
(1) (\ix)(3y)S[x,y]
where the truth-making choice of a value of y depends OIL the value of x.

Without dependent quantifiers, one cannot even express the functional
dependence of a variable on another. Quantifier dependence therefore is the
veritable secret of the strength of first-order logic. One can almost say that to
understand first-order logic is to understand the notion of quantifier
dependence. Several of the typical applications of first-order logic in
* Written jointly with Gabriel Sandu

philosophical analysis, such as uncovering quantifier-switch fallacies, exploit

directly the idea of quantifier dependence.
But to understand quantifier dependence is ipso facto to understand the
notion of quantifier independence. The latter is simply the same concept as
the former in a different dress. Hence it does not mean transcending in the
least the conceptual repertoire of ordinary quantification theory if we
introduce into it an explicit independence indicator by means of which the
independence of (QIX) of another quantifier (Qzx) is expressed by writing it
(QlxlQzx).
The same slash notation can be applied to propositional connectives. It is
obvious what is meant by such formulas as
(2) ("i/X) (A(x) (vNx)B(x
or
(3) ("i/x)(3y)("i/Z)(Sl[X,y,z](vNx)S2[X,y,Z])
It turns out that if we limit our attention to formulas in negation normal

form (Le. in a form where all negation-signs precede immediately atomic
formulas or identities), it suffices to consider only two kinds of uses of the
slash, viz. (3x1"i/y) and (vNy).
The result of adding the slash to the conceptual arsenal of the received
first-order logic results in what will be called independence-friendly (IF) first-
order logic.
What is IF first-order logic like? Does it really go beyond from received
first-order logic? In simplest cases, the new notation does not yield anything
new. For instance,
(4) ("i/x) (3yNx)S[x,y]
is obviously logically equivalent with
(5) (3y)("i/x)S[x,y]
Likewise (2) is equivalent with
(6) ("i/x)A(x) v ("i/x)B(x)

A REVOLUTION IN LOGIC 29
But in more complex cases something remarkable happens. Consider, for

instance, the sentence
(7) ('vIx)('vIz)(3yNz)(3uNx)S[x,y,z,u]
which is equivalent with
(8) ('vi x)(3y ) ('viz) (3 uNx)S[x,y ,z,u].
Can (7) be expressed without the slash notation? If so, there would be a
linear ordering of the four initial quantifiers ('vIx) , (3y) , ('viz), (3u) that would
result in the same logical force as (7). But what kind of ordering might that
be? Since (3y) depends in (7) on ('vIx) but not on ('viz), the relative order of the
three would have to be
(9) ('vIx)(3y)('vIz).
But since (3u) depends on ('viz) but not on ('vIx) , their relative order would
have to be ('vIz)(3u)('vIx), which is incompatible with (9).
Likewise, it can be shown that (3) cannot be expressed 'without the slash
notation (or equivalent).
Here we are beginning to see the whole horror of Freg:e's mistake. The
notation he introduced (like the later notation of Russell's lmd Whitehead's)
arbitrarily forbids certain perfectly possible patterns of dependence and
independence between quantifiers or between connectives and quantifiers.
These patterns are the ones represented by irreducible slashed quantifier
combinations (and similar combinations involving quantifiers and
connectives). Anyone who understands received first-order logic understands
sentences which involve such patterns, for instance understands sentences of
the form (7) or (8), in the concrete sense of understanding the precise
conditions that their truth imposes on reality. Hence they ought to be
expressible in the language of our basic logic. And the only way of doing so is
to build our true logic of quantification so as to dispense with the artificial
restrictions Frege imposes on the received first-order logic. The real logic of
quantification, in other words the real ground floor of the 'edifice of logic, is
not the ordinary first-order logic. It is IF first-order logic. Terminologically
speaking, we are doing ourselves injustice by calling the received first-order
logic "ordinary". The limitations we have been discussing make Fregean logic
systematically speaking quite extraordinary, as is also reflected by the fact
that the properties of the received first-order logic do not reflect at all
faithfully what one can expect to happen in logic in generaP
Furthermore, when we said that Frege committed his mistake in
formulating his formation rules, the statement was intended literally. In
order to reach IF first-order logic, it is not necessary to introduce any new
notation like our slash notation. All that is needed is to formulate the rules
for parentheses (scope) more liberally than Frege and Russell. If you reflect
on the matter (and even if you don't), there is absolutely no reason why the
so-called scope of a quantifier, expressed by the associated pair of
parentheses, should comprise a continuous segment of a formula adjacent to
the quantifier and following it.4 Expressions violating this requirement (and
consequentj.y violating the usual formation rules) can have a perfectly
understandable seman tical interpretation. Indeed, (8) might as well be
written as follows:
(10) ('v'x) [ (3y)('v'z)](3u)[S[x,y,z,u]]
where the two outer pairs of square brackets indicate the (discontinuous)
scope of ('v'x). In more complicated cases, alas, the liberated parentheses
notation turns out to be unintuitive to the point of being unreadable. Hence
for practical reasons we prefer the slash notation.
A fragment of IF first-order logic is known to the cognoscendi as the logic
of partially ordered (''branching'') quantifiers. 5 This logic is nevertheless not
entirely representative of the conceptual situation in general. For one thing,
propositional connectives can exhibit the same independence phenomena as
quantifiers. 6 Moreover, even though quantifier dependencies and
independencies can always be dealt with in terms of partial ordering, this is
not true of other concepts. For there is in general no reason why the
dependence oflogical notions on each other should be transitive.
Furthermore, the treatment of negation in the context of independent
quantifiers requires more attention than it has been paid in the theory of
partially ordered quantifiers. 7
Frege's mistake cannot be excused as a minor oversight, either. By
excluding the possibility of genuine independence between quantifiers (and
by implication between other concepts) Frege excluded from the purview of
logic a wealth of important subjects. Many logicians are still apt to consider
the logic of branching quantifiers and also IF first-order logic as a marginal
curiosity. Nothing could be further from truth. One could write an entire
book about the impact of IF first-order logic on the foundations of
mathematics. (Indeed, Jaakko Hintikka has done just that. See note 3.) But
the impact of this new logic extends much more widely. Far from being a
recondite linguistic phenomenon, independence of the sort IF first-order logic
deals with is a frequent and important feature of natural-language
semantics. 8 Without the notion of independence, we cannot fully understand
the logic of such concepts as belief, knowledge, questions and answers, or the
de dicta vs. de re contrast. 9
Frege could not resort to the higher-order component of his logic, either, to
do the same job as IF first-order logic. What gives the impression that he
could do so is that IF first-order sentences have second-order counterparts
asserting the existence of their Skolem functions. For instance, (7)
corresponds to
(ll) (3f) (3g) (Vx)(Vz)S[x,f(x) ,z,g(z)]
But (11) is logically equivalent with (7) only if the secondorder quantifiers
(3f), (3g) are given the standard interpretation in Henkin's sense. 10 And we
have shown in an earlier paper l l that Frege assumed a nonstandard
interpretation of higher-order quantifiers, including second-order ones. More
generally speaking, that means that there is no serious obstacle to treating
Fregean higher-order logic as many-sorted first-order logic, with different
types playing the role of the different sorts.
In first-order logic itself, the phenomenon of independence occasions
changes in the most basic laws oflogical inference. (Such laws were of course
precisely what Frege was trying to capture by means of his Begriffsschrift.)
Existential quantifiers cannot any longer be treated (along the lines of
familiar tableau or Gentzen-type methods)12 by an ordinary rule of
instantiation for formula-initial quantifiers. In order to see this, consider an
existential formula
(12) (3xNzl, VZ2, ... )S2[X]
occurring in a wider context
(We are of course assuming that (12) is in negation normal form, i.e. that all
its negation-signs precede immediately atomic formulas Olr identities.) If we
try to set up instantiation rules that deal only with formulas of the form (11),
we are in trouble. For in the transition from (ll) to (12) in, say, a tableau-
type procedure, individual constants will have to be substituted for Zl, Z2, ...
and more generally for all variables bound to quantifiers whose scope
includes (11). But the result of carrying out all these substitutions in (11) is
ill-formed (since there no longer are any quantifiers (VZl), (VZ2) , ... for (3x) to
be independent of), and has to be replaced by a formula of the form
(3x)S*[x]
where all the free variables of (11) have been replaced by individual
constants. But then the difference between the outer quantifiers in (12)
which (3x) depends or doesn't depend on will be erased.
What has to be done is obviously to formulate a rule of instantiation that
allows us to move from (12) to
where (VYl) , (VY2) , ... are all the universal quantifiers other than (VZl), (VZ2) ,
... within the scope of which (3xNzl, VZ2, ... ) occurs in (12). In (14), {must be a
new function symbol. Notice that in spite of the introduction of a new
function symbol, (14) is first-order. It does not involve any quantification over
functions or other higher-order entities.
The usual rule of existential instantiation can be considered a special case
of the new rule. In this case, (3x) does not depend on any outside quantifiers.
Then the function {(yl, Y2, ... ) of (14) has no arguments and accordingly
becomes a new individual constant.
This change in the rule of existential instantiation can be thought of as
merely spelling out the way quantifiers depend on each other. It is only in
virtue of an undeserved luck that in ordinary first-order logic we can get
away without an instantiation rule that does not introduce new function
constants.
Moreover, Frege is guilty of a premeditated ideocide rather than an
involuntary one. His rejection of independent quantifiers can be motivated by
one or both of two general theoretical ideas, both of which Frege espoused in
so many words. One of them is Frege's interpretation of quantifiers as
higher-order predicates. On this view, what an existential statement like (13)
says is that the (usually complex) predicate S*[x] is not empty. Now for
essentially the same reasons as were seen to prevent an adequate treatment
of existential quantifiers by the sole means of the usual rule of existential
instantiation, the Fregean treatment of existence fails to do justice to
independent quantifiers. At bottom, the interpretation of quantifiers as
higher-order predicates is but an example of philosophers' unfortunate
tendency to try to deal without quantifiers without taking into account their
interaction. Another one is the presumption that we can understand
quantifiers in the sole terms of their "ranging over" a class of values. l3
In this department, the sins of one's intellectual ancestors are visited on
the theories of their scholarly children and grandchildren. The most
intensively cultivated recent treatment of quantifiers in general has been the
so-called theory of generalized quantifiers. l4 It relies crucially on the Fregean
interpretation of quantifiers as higher-order predicates. Small wonder,
therefore, that this theory has not been general enough to deal with
independent quantifiers. The best that the representatives of this tradition
have been able to do is to formulate interpretations of certain particular
combinations of quantifiers involving independent ones (e!.g. combinations
like (7 as unanalyzed units, without showing how this interpretation
depends on the semantics of their constituent expressions. Such a treatment
fails to be a genuine theory of independent quantifiers by a long shot. For
instance, it is helpless to say anything about independence relations between
quantifiers and propositional connectives.
But the possibility of independent quantifiers militates against even more
generally received semantical principles. This can be seen most vividly by
changing our notation and associating the independence indicator (changed
now to a double slash) to the quantifier that another one is independent of.
Thus instead of (4) we could now write
(15) (';;7x/13y) (3y)S[x,y)
and instead of (7)
(16) (';;7 x/13u) (';;7 zlI:3y)(:Jy) (3 u)S[x,y ,z,u].
The double-slash notation makes explicit what is less conspicuous in our

original notation. One of the most striking things conceptually about IF first-
order logic is that it violates what is often known as the principle of
compositionality.l5 This principle can be formulated by saying that according
to it the crucial semantical attributes of an expression are functions of the
semantical attributes of its constituent expressions. Another, somewhat more
illuminating, way of putting it is to say that the principle is an assertion of
seman tical context-independence. According to it, the semantical
interpretation of an expression must not depend on its context. The reason is
clear. If it did, the semantical attributes of the expression in question could
34 LANGUAGE, TRUlli AND LOGIC
not always be functions of (functionally determined by) those of its

constituent expressions.
But in IF first-order logic we do find semantical context-dependence. For
instance, the very same expressions (3y)S[x,y] and (3y)(3u)S[x,y,z,u] occur not
only as subformulas of (15) and (16) but as subformulas of (1) and of
(17) ('v"x) ('v"z) (3y) (3 u)S[x,y,z, u]
respectively. Yet their semantical behaviour depends on the context, for only
from the context can one see which outside quantifiers (3y) and (3u) depend
and do not depend on. Hence IF first-order logic offers a clear-cut counter-
example to the principle ofcompositionality.
Even though there may be some room for doubt - or at least need for
some further explanations Frege accepted the principle of
comp osition ality, so much so that it is sometimes referred to as the Frege
principle. I6 Small wonder, then, that Frege did not countenance independent
quantifiers. They might have prompted him to give up the principle of
composi tionali ty.
We have to be very careful here, however. As Jaakko Hintikka 17 has
pointed out, compositionality can always be reinstated by hook or crook by
building the laws of contextual interdependence into the semantical
attributes of the interacting notions. It appears that Wilfrid Hodges lB has in
fact succeeded in doing so in the case of IF quantifiers. It remains to be
examined precisely what the price is that Hodges has to pay for restoring
compositionality in his formal semantics. In other words, a closer
examination is needed to answer the question as to in what precise sense IF
logic violates compositionality. One concrete sense is in any case the need of
inference rules (like the transition from (12) to (14 that are context-
sensitive.
In many other respects, too, the properties of IF first-order logic need (and
deserve) a much more extensive discussion than can be presented in the
confines of one paper.l9 Independence-friendly first-order logic has many of
the same pleasant metalogical properties as the received first-order logic. It
is compact, and such results as the LOwenheim-Skolem theorem and the
separation theorem hold in it. It also admits of a complete disproof procedure,
assuming of course a reformulation of the rule of the existential instantiation
as a rule of functional instantiation explained above. In other words, the set
of inconsistent formulas is recursively enumerable. But there is one massive
difference between IF first-order logic and its predecessors. IF first-order
logic does not admit of a complete axiomatization. It is inevitably
incomplete. 20 The class of valid IF first-order logic formulas is not recursively

enumerable.
The title of our paper speaks of a revolution. Indeed, IF first-order logic
has many revolutionary features. It will take more than one monograph to
spell them out. However, there does not seem any doubt that the semantical
incompleteness of our basic logic - for that's what IF first-order logic was
just found to be - is the most profoundly revolutionary feature of the new
logic. Here we can only briefly indicate some of the main consequences of this
particular aspect ofIF first-order logic.
The historical reason for the revolutionary consequences of the
incompleteness of IF first-order logic is that our basic logie has traditionally
been assumed to be complete in the familiar sense that we can give a
complete set of purely formal rules of logical inference. Without this
assumption, some of the main foundational projects in the last hundred and
fifty years make little sense. Frege's enterprise in logic consisted precisely in
the presentation of a complete and completely formal axiom system of 10gic. 21
Indeed, he believed that no other approach to logic than a purely formal
(symbolic) and axiomatic one is possible, because the semantics of a language
cannot according to his lights be expressed in that language. Hence his
mission in logic was doomed to remain unaccomplished. Furthermore, in a
sense the completeness of his logic was also a presupposition of Frege's
foundational project. For the form in which he conceived of the reduction of
all mathematical concepts and modes of reasoning to logic was a reduction of
mathematics to a system of logic. If that system is not complete, there will
not be anyone system to reduce mathematics to. There will always be modes
of reasoning whose status is not yet determined. Are they logical or
irreducibly mathematical? A Poincare would have been all too quick to opt for
the second alternative.
Even more blatantly, Hilbert's proof-theoretical project was predicated on
the possibility of a complete axiomatization of the rules of logical inference.
For the way he wanted to prove the consistency of various mathematical
theories was to show that one cannot formally prove a contradictory
conclusion from their axioms. Without the completeness of our rules of formal
proof this project is vacuous. For no matter how one formal rules of proof are
formulated, a contradiction may be hidden in those consequences of the
axiom system that are not deducible by means of the rules of inference so far
formulated.
That major earthquake of twentieth-century 10 IP-C , Godel's first
incompleteness theorem, has unfortunately served only to strengthen the
illusion of the completeness of our basic logic. For it has been generally
36 LANGUAGE. TRU1H AND LOGIC
assumed that what GOdel proved was the inevitable incompleteness of

relatively simple mathematical theories in every interesting sense of the
term.
But if our logic itself is incomplete, we have to reconsider the entire
situation in the foundations of logic and mathematics. For one thing, as
Jaakko Hintikka has pointed out,22 absolutely speaking Go del's
incompleteness theorem establishes only the deductive incompleteness of
elementary arithmetic. In other words, we cannot formally derive by means
of the underlying logic 8 or -,8 for each closed sentence 8. This deductive
incompleteness implies the descriptive incompleteness (nonaxiomatizability)
of elementary arithmetic only if the underlying logic is semantically
complete. Hence the incompleteness of IF first-order logic opens up a realistic
possibility that we might be able to formulate descriptively (model
theoretically) complete axiom systems for various nontrivial mathematical
theories already on the first-order level without violating Go del's
incompleteness theorem. All we need to do is to use IF first-order logic or
some suitable extension of it instead of traditional first-order logic as our
basic logic. This does not necessitate any excuses, for it was argued earlier in
this paper that we have to use such a logic in any case. We can even study
such descriptively complete axiom systems model-theoretically and ascertain
in this way some of their properties without having to worry about the rules
of logical proof at all.
It is well known that descriptively complete but deductively incomplete
axiomatizations of practically all mathematical theories can be formulated in
higher-order logic. But such formulations involve all the dangers,
uncertainties and other problems that pertain to the existence assumptions
concerning higher-order entities, including sets. Since the kind of
descriptively complete axiomatization envisaged here would be first-order, it
could be free of all difficulties concerning the existence of sets or other
higher-type entities.
It turns out that IF first-order logic alone does not in fact allow for a
descriptively complete axiomatization of elementary arithmetic. However,
two different kinds of natural extensions of IF first-order logic facilitate such
an axiomatization. We can either add to this logic sentence-initial
contradictory negation or else restrict the initial verifier's strategies to
Turing machine computable ones. In either way, we can reach a descriptively
complete first-order axiomatization of elementary arithmetic.
Moreover, we have to have a new look at the familiar GOdelian statement
that there are unprovable truths of elementary arithmetic. This adage turns
out to be a half-truth. Indeed, our observations imply that the very notion of
logical proof has to be reconsidered. It is in these days often asserted or

alleged that the notion of proof has lost its central position in mathematics as
a consequence of Godel's results. 23 Nothing is further from truth. Because we
can formulate descriptively complete axiomatic theories, the task of a
mathematical or scientific theorist can still be thought of as a derivation of
logical consequences from the axioms. And such a derivation can still be
considered as a purely logical proof. The requirement that characterizes its
different steps is still the same as of old, viz. truth preservation. All the
former rules concerning such truth preservation remain valid, and new ones
can in principle always be discovered model-theoretically. Each legitimate
step in a proof is still an instance of a purely formal, truth preserving pattern
of inference.
Thus there is a plain, not to say common, sense in which it is false to say
that "we know from Godel's incompleteness theorem that every consistent
proof procedure is bound to leave infinitely many closed sentences of classical
mathematics indemonstrable and irrefutable". This is true only if proof
procedures are restricted to recursively enumerable ones. In this sense, we on
the contrary know that for each truth of elementary arit.hmetic there is a
valid formal proof of it which utilizes only a finite number of deductive
principles. Each of those principles can be discovered and formulated. In this
sense, there are no unprovable truths of elementary arithmetic. This fact is
perfectly compatible with the Godelian fact that the entire set of all valid
rules of deductive inferences is not recursively enumerable. What is new as
compared with provability by means of ordinary first-order logic is merely
that a full set of purely formal rules for such proofs cannot be specified once
and for all. In proving genuinely new results one must always be ready in
principle to countenance steps which are not in accordance with the
previously acknowledged rules of inference. But as long as they are truth-
preserving, such new inference patterns must of course be considered as
legitimate rules of logical proof. This legitimacy means the same as of old,
viz. that all steps of the same form preserve truth. If by the completeness of
formal logic one means the possibility of expressing each and every valid
logical inference as an instance of truth-preserving formal inference pattern,
then all valid logical inferences are formal and formal IO!,ric is in this sense
complete. Perhaps the generally accepted terminology is not as descriptive as
it could be. Perhaps we ought to change it and to say that IF first-order logic
is "formally complete" but not "computable". To put the same point in
different terms, the inevitable incompleteness that is found in logic and in
the foundations of mathematics is not a symptom of any intrinsic limitation
to what can be done by means of logic or mathematics. It is a limitation to

what can be done by means of digital computers.
What the recent and current unease with the traditional conception of
proof in mathematics is a symptom of is not any flaw in the notion of logical
proof, but the need of making the notion of proof more realistic along the
lines just adumbrated, which merely means not restricting the legitimate
steps in a logical argument to some mechanically specifiable list of inference
patterns.
But the most serious and most widespread philosophical mistake in this
direction still remains to be diagnosed. It can be brought out by asking: How
are the new inferential principles (new logical truths) discovered? Here an
extreme form of the belief in the completeness of logic has led to the idea
that, since the introduction of new principles of proof cannot be done by
means of preprogrammable rules, their introduction must be in principle a
nonlogical matter, perhaps a matter of a special intuition, of an arbitrary
decision, or of a conjecture later to be tested and possibly modified. The need
of such constant amplification of the inferential basis of logical and
mathematical proof has even been alleged to cast doubts on the uses of the
notion of truth in logic and in mathematics.
These views are wrong, and invidiously so. A particularly pernicious
variant is the postulation of a special mental capacity, sometimes referred to
as mathematical intuition, as a source of new mathematical truths. 24 Now an
unanalyzed postulation of such a mysterious capacity amounts to an abject
lowering of intellectual standards. If you look at the earlier, respectable uses
of the notion of intuition in the history of philosophy, you will find that its
role as a source of insight or knowledge was always supposed to be backed up
by some philosophical or psychological theory. For Aristotle, the backing was
his theory of thinking as an authentic realization of forms in one's soul, for
Descartes it was the doctrine of innate ideas, and for Kant it was his
transcendental assumption that the forms of space and time have been
imposed on objects by our faculty of sense-perception, which makes these
forms recoverable in intuition. But in recent discussions mathematical
intuition has all too often been resorted to also when no such theoretical
backing is in the offing.
In sober reality, the ways of introducing and justifying new inference
patterns and new logical truths are much more mundane than an appeal to
an extralogical intuition. In fact, we suspect that most appeals to the so-
called mathematical intuition are merely tacit appeals to model-theoretical or
proof-theoretical considerations. Such arguments can sometimes even be
formulated in the very same theory whose axioms and rules of inference one
is studying. If an example is needed, G6del's very proof of the incompleteness

of any given axiomatization AX of elementary arithmetic can serve as a case
in point. In it, an argument expressible in elementary arithmetic itself leads
us to a truth of elementary arithmetic which is not derivable from AX and
hence can serve to strengthen it. Admittedly, G6del's argument is proof-
theoretical rather than model-theoretical. For this reason, philosophers have
not been tempted to think of it as an appeal to intuition. But it is not
impossible to construct a model-theoretical argument which helps to discover
new inferential principles or to vindicate old ones under criticism. In fact, IF
first-order logic opens up new possibilities in this direction. The reason is
that several crucial model-theoretical concepts, including the concepts of
truth and satisfaction, for a suitable first-order logic can be expressed in the
very same language.
An instructive object lesson in the utter futility of trying to appeal to
mathematical or logical intuition is offered by the curious history of the
axiom of choice. 25 Different competent mathematicians have had
contradictory "intuitions" about it. Zermelo for one introduced it as an
obviously valid set-theoretical axiom. Hilbert's intuition seems to have told
him that the axiom of choice is as obviously valid as 2 + 2 = 4, and valid in
the same sense. 26 Wisely, he treated this view as a hunch to be established
through a suitable formulation of the axiom rather than an infallible dictum.
The same desideratum has been flaunted by Dana Scott as a hope that so far
remains unfulfilled. Others, notably Brouwer and his fellow intuitionists
rejected the axiom. Worse still, several of the mathematicians whose
intuitions (or perhaps semi-intuitions) told them to ]ceject the axiom
subsequently turned out to have used the axiom themsellves in their own
mathematical reasoning. Which intuitions of these mathematicians are
genuine ones, the ones that they had tacitly appealed to in their working
reasoning, or the ones prompted by their reflections on the general and
abstract formulation of the axiom of choice?
This whole mare's nest of problems can be clarified, not by an appeal to
intuitions, but by logical, especially model-theoretical, analyses. The basic
ideas of these analyses have been explained by Hintikka. 27 Without trying to
summarize the whole argument here, it can be mentioned that several of our
prima facie intuitions have to be re-educated in the li~iht of the model-
theoretical analyses. For instance, the most popular "intuitions" that have
led to the rejection of the axiom of choice turn out to be based on outright
confusion, viz. based on giving different quantifiers in a common formulation
of the axiom a different interpretation. The formulation we have in mind is
second-order axiom schema
(18) (Vx)(3y)S[x,Y]::J (3f)(Vx)S[x,f(x)].
Critics have turned out to have implicitly given the first-order quantifiers
(Vx), (3y) a classical interpretation but the second-order quantifier (3f) a
nonclassical one. If all quantifiers are given the same interpretation the
axiom of choice holds. This apparently has been noted by Dummett whose
mature intuitions in this matter hence differ from those of other
intuitionists. 28 In other cases, too, an alleged intuitive justification for
rejecting the axiom has proved to be fallacious. Most importantly, the
intuitions of the intuitionists can be shown to pertain to an altogether
different subject matter than the intuitions of the typical nonintuitionist. 29
The latter are concerned with our knowledge of mathematical facts (truths)
while the former are concerned with our knowledge of mathematical objects.
Hence the two parties have been arguing at cross-purposes. For instance, on
the case of the axiom of choice the intuitionists are asking whether we know
the choice function while the classical mathematicians are asking whether
we know that such choice function exists. The confusion between these two
issues thus exemplifies a confusion between two kinds of knowledge, viz.
knowledge of truths and knowledge of objects.
But even after all the alleged intuitive testimony against the axiom of
choice has been discredited, don't we still need intuition to decide whether it
is acceptable or not? No, what we need is a grasp of the meaning of
quantifiers. Earlier in this paper, it was pointed out that an essential
component of the logical behaviour of quantifiers is the way they depend on
each other. For instance, it is part and parcel of the meaning of quantifiers
that in a sentence like
(1) (Vx) (3y)S[x,y]
the truth-making choice of the value of y depends on the choice of the value
of x. Hence (1) is true if and only if there is a function that implements that
choice. In other words, (1) is true if and only if the following second-order
sentence is true:
(19) (3f)(Vx)S[x,f(x)]
But the implication from (1) to (19), that is (18), is a form of the axiom of
choice. This axiom is in other words but a manifestation of the interaction
between different quantifiers that is a part of their meaning. The axiom of
choice is therefore not only valid, but a valid logical principle, just as Hilbert
surmised.
The fact that the validity of the axiom of choice has not been more
universally recognized has a prima facie tempting but fallacious reason. This
usually tacit reason is a belief in the principle of compositionality. According
to it, the meaning of a quantifier must not depend on what there is outside
its scope. But in a perfectly good sense, the truth-making choice of a variable
bound to an existential quantifier, say to (3y) in (1), depends on something
outside its scope, viz. on the value of the universal quantifiers within the
scope of which it occurs, in (1), on the value of x. Thus in the last analysis the
interplay of quantifiers amounts to living (albeit oblique) testimony against
the principle of compositionality. This testimony should be enough to put to
rest all realistic hopes of maintaining compositionality in any strict sense of
the word in one's logical semantics.
When this is realized, the last objections to the axiom of choice in the form
of (18) evaporate and it is seen that the axiom of choice is valid. If it is
construed as a rule of inference, that rule is truth-preserving.
But there is even more to be said here. Model-theoretic considerations
suggest interesting distinctions between different forms of the axiom of
choice and related assumptions. The line of thought that justifies (18) can be
used to vindicate any inference from a first-order sentence (IF or not) to the
second-order sentence that asserts the existence of its Skolem function.
Accordingly, they should all be valid. However, they cannot all be true in any
one model of axiomatic set theory.so Hence our model-theor,etical "intuitions"
yield a serious objection to existing first-order axiomatizations of set theory.
Thus the model theory of IF first-order logic yields a highly interesting
possibility of formulating new, stronger axioms of set theory while at the
same time leading us to the problem of reconciling those intuitively valid
axioms with the usual first-order axiomatizations of set theory.
Thus in this entire matter of the axiom of choice, it llS the appeals to
intuition that have led mathematicians on various wild goose chases, leaving
the resulting mess to be clarified by model-theoretical analysis. All this
illustrates vividly our suggestion that so-called intuitions in mathematics are
little more than tacit appeals to model-theoretical considerations. Such
"intuitions" can accordingly be sharpened and re-edueated by model-
theoretical clarification of the relevant concepts. Generally speaking, the
time is ripe to say to the philosophers who have been discussing the mythical
faculty of mathematical intuition: Put up or shut up. So far they have not
come up with any unequivocal insight into the subjects they have promised
us intuitions about.
42 LANGUAGE, TRUrn AND LOGIC
NOTES
1 Immanuel Kant, Kritik der reinen Vernunft B 7-8.

2 Cf. Jaakko Hintikka, "Aristotle's Incontinent Logician", Ajatus vol. 37 (1978), pp.
48-65, where Hintikka argues that Aristotle's system of logic was an unhappy
compromise between the different guiding ideas he was trying to implement.
3 Cf. Jaakko Hintikka, The Principles of Mathematics Revisited, Cambridge, V.P.,
1996.
4 See J aakko Hintikka, "No Scope for Scope" (forthcoming). Note that the usual
scope notation for propositional connectives violates already some of these
requirements.
5 Branching quantifiers were introduced by Leon Henkin, "Some Remarks on
Infinitely Long Formulas", in In/initistic Methods, Warsaw, 1959, pp. 167-183. For
their theory, see e.g. W. Walkoe, "Finite Partially Ordered Quantification", Journal
of Symbolic Logic vol. 35 (1970), pp. 535-550; Herbert Enderton, "Finite Partially-
Ordered Quantifiers", Zeitschrift filr Mathematische Logik und Crundlagen der
Mathematik vol. 16 (1970), pp. 393-397; and Jon Barwise, "Some Applications of
Henkin Quantifiers", Israel Journal of Mathematics vol. 25 (1976), pp. 47-80.
6 See Gabriel Sandu and Jouko Viiiiniinen, "Partially Ordered Connectives",
Zeitschrift filr Mathematische Logik und Crundlagen der Mathematik vol. 38 (1992),
pp. 361-372.
7 For the behaviour of negation in IF logic, see Hintikka, op. cit. note 3, chapter 7.
It is shown there that the law of excluded middle fails in IF logic and that the only
negation for which usual semanticallaws can be formulated is therefore a strong
(dual) negation, not the contradictory negation. From this it follows in turn that the
decision problem for validity does not reduce to the decision problem for
inconsistency. For a similar reason, Lindstrom's well-known theorem does not apply
to IF logic.
S See Jaakko Hintikka and Gabriel Sandu, "Informational Independence as a
Seman tical Phenomenon", in J.E. Fenstad et al., editors, Logic, Methodology and
Philosophy of Science, Elsevier, Amsterdam, 1989, pp. 571-589.
9 See here Jaakko Hintikka and Gabriel Sandu, "Knowledge Acknowledged",
Philosophy and Phenomenological Research vol. 56 (1996), pp. 251-275.
10 See Leon Henkin, "Completeness in the Theory of Types", Journal of Symbolic
Logic vol. 15 (1950), pp. 81-91 and cf. Jaakko Hintikka, "Standard vs. Nonstandard
Distinction: A Watershed in the Foundations of Mathematics", in Jaakko Hintikka,
editor, From Dedell-ind to Cadel, Kluwer Academic, 1995, pp. 21-44.
11 See Jaakko Hintikka and Gabriel Sandu, "The Skeleton in Frege's Cupboard:
Standard vs. Nonstandard Distinction", Journal of Philosophy vol. 89 (1992), pp.

290-315.
12 The classical formulation of the tableau method in E.W. Beth, "Semantical
Entailment and Formal Derivability", Mededelingen van de Koninkijke Nederlandse

Akademie van Wetenschappen, Md. Letterkunde, N.R. vol. 18, 1995, 309-342.
13 See here Jaakko Hintikka and Gabriel Sandu, "The Fallacies of the New Theory
of Reference", Synthese vol. 104 (1995), pp. 245-283.

14 For a survey of the theory of generalized quantifiers, see M. Krynicki, M.
Mostowski and L.K. Szczerba, editors, Quantifiers: Logics, Models and Computation,
Kluwer Academic Publishers, 1995.
15 For this principle, see Barbara Partee, "Compositionality", in F. Landman and F.
Veltman, editors, Varieties of Formal Semantics, Dordrecht, Foris, 1984, pp. 281-
312; Francis J. Pelletier, "The Principle of Semantic Compositionality", Topoi vol. 13
(1994), pp. 11-24; and Wilfrid Hodges, "Compositional Semantics for a Language of
Imperfect Information", forthcoming.
16 See Michael Dummett, Frege: Philosophy of Language, Duckworth, London, 1973.
17 Jaakko Hintikka, "Theories of Truth and Learnable Lanb'Uages", in Jaakko
Hintikka and Jack Kulas, The Game of Language, D. Reidel, Dordrecht, 1985, ch.
10.
18 Op. cit. note 13 above.
19 For a tentative exploration, see Jaakko Hintikka, "What Is Elementary Logic?",
in Kostas Gavroglu, John Stachel and Marx Wartofsky, editors, Physics, Philosophy
and the Scientific Community, Kluwer Academic Publishers, Dordrecht, 1994, pp.
301-326.
20 The easiest way to see this incompleteness is the following: The socalled Henkin
quantifier H acting on four variables is defined as follows
(Hxyzu)S[x,y,z,u] B (3f)(3g)(Vx)(Vz)S[x,f(x),z,g(z)].
The Henkin quantifier is definable in IF logic in an obvious way:
(Hxyzu)S[x,y,Z,u] B (Vx)(Vz)(3yNz)(3uNx)S[x,y,z,u]
It is well known (cf. M. Krynicki and A. Lachlan, "On the Semantics of Henkin
Quantifier", Journal of Symbolic Logic vol. 44 (1979), pp. 184-200) that the logic
which extends first-order logic with the Henkin quantifier is incomplete. From this it
also follows that IF logic is incomplete.
21 See here Jean van Heijenoort, "Logic as Calculus and Logic as Language",
Synthese vol. 17 (1967), pp. 324-330.

44 LANGUAGE, TRUTII AND LOGIC
22 Jaakko Hintikka, The Principles of Mathematics Revisited (note 3), chapter 6.

23 There has been in the last few years an extensive discussion of the status of rigor
and proofs in the columns of journals addressed by mathematicians to
mathematicians. See e.g. Arthur Jaffe and Frank Quinn, "'Theoretical
Mathematics': Toward a Cultural Synthesis of Mathematics and Theoretical
Physics", Bulletin (New Series) of the American Mathematical Society vol. 29 (1993),
pp. 1-13; William P. Thurston, "On Proof and Progress in Mathematics", ibid. vol. 30
(1994), pp. 161-177, and the other responses to Jaffe and Quinn, loco cit. pp. 178-
211; Bonnie Gold, "What Is the Philosophy of Mathematics and What Should It
BeT, The Mathematical Intelligencer vol. 16, no. 3 (1994), pp. 20-24; and cf. John
Horgan, "The Death of Proof', Scientific American vol. 269, no. 4 (October 1993), pp.
92-103.
24 The notion of intuition and its alleged uses in mathematics are in a dire need of
critical scrutiny. Among the main sources of contemporary philosophers' ideas about
intuitions are Kantian Anschauunge and Godel's emphasis on mathematical
intuition. Neither one is currently being looked at in the right historical and
systematic perspective, however.
25 The history of the axiom of choice is studied by Gregory H. Moore, Zermelo's
Axiom of Choice, Springer-Verlag, Berlin, 1982.

26 David Hilbert, "Neubegriindung der Mathematik", in Gesammelte Abhandlungen
vol. 3, Springer-Verlag, Berlin, 1935, pp. 157-177. (See especially p. 157.)

27 See op. cit., note 3 above, chapters 2, 8 and 11.
28 Michael Dummett, Elements of Intuitionism, Clarendon Press, 1977, pp. 52-54.
29 See op. cit., note 3, chapter 11.
30 See op. cit., note 3, chapter 8.

3
A REVOLUTION IN THE FOUNDATIONS OF MATHEMATICS?
1. FIRST-ORDER LOGIC AND ITS SHORTCOMINGS
The standard contemporary view of the foundations of mathematics and of the

role of logic in it is well known and fumly entrenched. The ground floor of the
edifice of mathematics is on this view our basic logic, that is to say first-order
logic, a.k.a. quantification theory or lower predicate calculus. It has various
desirable features, such as completeness, compactness, Lowenheim-Skolem
property and the validity of the separation theorem. Indeed, it is the strongest
possible logic in the sense of abstract (model-theoretical) logic that has the
pleasant properties of compactness and Uiwenheim-Skolem property, assuming
only the usual behavior of propositional connectives plus a few plausible
structural properties. This is what the famous theorem of Lindstrom's shows.!
This theorem seems to assign a special position to ordinary Just-order logic. At
the same time, it is a kind of impossibility theorem, showing that we cannot hope
to strengthen first-order logic without losing some of its desirable properties.
First-order logic was first formulated explicitly by Frege as a part of his more
comprehensive Begriffsschrift. Frege had to go further, however. His logic is not
first-order, but higher-order. 2 In other words, Frege's logic allows quantification
not only over individuals (particulars), but also over higher-order entities, such
as functions and other concepts applying to individuals. The problems caused by
this transgression beyond first-order concepts will be discussed later in this
paper.
It is sometimes said that the idea of a freestanding first-order logic was not
known to Frege and that it crystallized only later, for the first time apparently in
Hilbert's and Ackermann's 1928 textbook (l\:1oore 1988). Maybe so. But even if
Frege should have entertained the idea of pure first-order logic, he would have
had plenty of good prima facie reasons to go beyond what is in our days known
as first-order logic. The most basic reason is that ordinary :fi.:rst-order logic does
not suffice for mathematics. Another reason is that it is not even self-sufficient.
The first of these two failures is illustrated by the fact that several of the most
basic concepts of all mathematics cannot be expressed by means of ordinary first-
order logic. Perhaps the most fundamental concept in the foundations of
mathematics, the concept of equicardinality, cannot be expressed in first-order

terms. When do the extensions of two concepts A and B have the same
cardinality? If and only if there are functions f and (its inverse) g such that
(1) (\fx)(\fz)A(x) :::J B(f(x & (B(z) :::J A(g(z))) & z =f(x B (x =g(z
Here we cannot dispense with the quantification over the functions f and g.
Other concepts that likewise cannot be expressed in first-order terms include
infinity, continuity in the sense of general topology, mathematical induction, etc.
No wonder that Frege resorted to higher-order logic in his attempted reduction
of mathematics to logic. For one thing, he needed the concept of equicardinality
for his definition of number.
For another thing, first-order languages are not self-sufficient in the sense
that the model theory of a first-order language or first-order axiom system
cannot be formulated in first-order terms. The most central concept of all model
theory (logical semantics), the concept of truth, can be defined for first-order
languages along the lines Tarski staked out (Tarski 1956). Alas, these lines lead
us away from a first-order language to the corresponding second-order language.
A Tarski-type truth predicate for a first-order language is a second-order
predicate asserting the existence of a suitable kind of valuation, that is, of a
function from the expressions of the language to their potential values in the
given model.
2. TRADITIONAL PICTURE OF THE FOUNDATIONS OF MATHEMATICS
Hence we apparently need higher-order logic in our mathematical theorizing.

But higher-order logic is not only inevitably incomplete. It is entangled with all
the problems with the existence of higher-order entities like sets and functions.
Hence it might seem only fair to fess up and admit that Quine is right in calling
higher-order logic set theory in sheep's clothing. The proper framework for
mathematical theorizing therefore appears to be set theory - a view which is
currently shared by many, perhaps most, mathematicians. Because in set theOlY
we need axioms over and above those of first-order logic, it is usually considered
a mathematical rather than logical theory.
The study of set theory is usually conducted in the same way as that of any
other mathematical theory, that is to say, in the form of an axiomatic theory
using first-order logic. Higher-order logic is not used, for the whole idea of
axiomatic set theory is to dispense with it. Instead, the only logic used to
regulate the consequences of set-theoretical axioms is the ordinary first-order
logic, once again reflecting the dogma that this logic is our natural basic logic.
A REVOLUTION IN THE FOUNDATIONS OF MATHEMATICS? 47
First-order axiom atiz ations of nontrivial mathematieal theories are

nevertheless inevitably incomplete, no matter whether set theory is being
resorted to or not. 3 For one thing, first-order set theories cannot do the same job
as higher-order logics in one important respect. No explicit first-order
axiomatization can capture the intended standard interpretation of higher-order
quantifiers. In the (ordinary first-order) axiomatization of set theory, there will
inevitably be models in which e.g. functions variables (like the t, g of (1 do not
range over literally all extensionally possible functions, but only some specified
subset of the set of all such functions. But this means that formulas like (1)
cannot quite do the job they were drafted to do in all models of the underlying
language. For instance, the extensions of A(x) and B(x) might be equinumerous,
and yet there might not exist any functions {, g (in the sense of the first-order
theory) doing the one-to-one correlation expressed by (1).
Now the first-order character of the usual axiomatizations of set theory means
that set theory cannot capture the standard interpretation of higher-order
quantifiers. 4 Model-theoretically speaking, any first-order axiomatization of set
theory admits of nonstandard models which violate the standard interpretation
of set theory thought of as a substitute for higher-order logic.
Moreover, since first-order logic cannot even capture the principle of
mathematical induction, even first-order theories such elementary mathematical
theories as arithmetic are bound to be incomplete, as Godel showed in 1931
(Go del 1986a). This holds of course also when elementary number theOlY is
reconstructed within first-order axiomatic set theory.
Thus the usual line of thought results in a two-tier model of the foundations of
mathematics. The ground floor according to this view is first-order logic, which is
semantically complete but not strong enough to cope with actual mathematical
conceptualizations. The next floor is set theOlY and/or higher-order logic. On this
level, we can express the basic mathematical concepts, but we cannot hope to
achieve completeness. Nevertheless, axiomatic set theory is in this view the best
general framework for mathematical theorizing that we have available to us.
On this view, the main foundational problems can be located. They pertain to
the higher levels of the structure I have described. In other words, practically all
the difficulties in the foundations of mathematics seem to lie in formulating
complete mathematical theories. The task of studying the logical consequences of
axiomatic theories can apparently take place by means of a simple logic that can
be mastered once and for all in the sense of admitting a semantically complete
axiomatization, that is, a recursive enumeration of all valid formulas.
Even though on this view logic is unproblematic, its scope is severely limited.
One might even say that the picture I have painted is overshadowed by several
impossibility results, including Godel's incompleteness theorem, Tarsk]'s result

that truth can be defined for a first-order theory only in a stronger theory,
Lindstrom's theorem, and the impossibility of expressing notions like
equicardinality on the first-order level.
This generally accepted picture has one thing against it. It is wrong. It rests
on a woefully inadequate analysis of the entire problem situation in the
foundations of logic and mathematics. It is based on an unnecessarily restrictive
idea of what our basic logic is like. It relies on an ambiguous and consequently
misleading picture of what completeness means. Moreover, and perhaps most
fundamentally, it relies on an unclear idea what the ends of mathematical
theorizing are in the first place.
r will explain these four points in the rest of this paper.
3. INDEPENDENCEFRIENDLY FIRSTORDER LOGIC
To take the first point first, whether we credit the ordinary first-order logic to
Frege or to Hilbert and Ackermann, it does not do the job fully adequately that it
was calculated to do. 5 For what is first-order logic supposed to do? It is supposed
to be the logic of quantifiers. Now the source of the expressive power of
quantifiers is their interplay. Quantifiers cannot be adequately characterized as
higher-order predicates as Frege thought, or as "ranging over" a class of values,
as most philosophers seem to think in these days. Quantification theory is
essentially a study of the interplay of different quantifiers, their dependencies
and independencies. And the construal of quantifiers as higher-order predicates
cannot do justice to this interplay.
It has turned out that Frege's approach to logic, including the formation rules
for the first-order fragment of his total logic, rules out certain combinatorially
possible and easily interpretable patterns of dependence and independence
between quantifiers. The simplest instances of a pattern not expressible in
ordinary first-order logic are the so-called Henkin quantifier formulas
expressible as the following two-dimensional "branching quantifier" formula:
(2) (Vx) (3y)
(Vz)(3u)
> S[x,y,z,u]
Here the truth making value of y depends only on x and the similar value of u
only on z. Hence (3) is equivalent to the following second-order formula
(3) (?Jf)(?Jg)('Vx)('Vz) S[x,f(x),z,g(z)]
which is easily seen to be irreducible to any ordinary firstorder fonnula. In

practice, it is simplest to introduce a special symbol to indicate the independence
of a quantifier, say (?Jy), of another one, say ('Vz) , of which it otherwise depends
on, by writing it (?JyNz). Then (2) = (3) can be written on the first-order level
linearly as
(4) ('Vx)('Vz)(?JyNz)(?JuNx) S[x,y,z,u]
When this slash notation is extended also to propositional connectives, we obtain

what has been called independence-friendly (IF) first-order logic (Hintikka
1995a; Hintikka 1996, Chaps. 3,7).
This logic has a better claim to be our true basic logic than ordinary first-order
logic, in that it is free from the needless and arbitrary restrictions that beset
ordinary first-order logic. IF first-order logic has several of the same desirable
metalogical properties as its received counterpart. It is compact and has the
Lowenheim-Skolem property. Separation theorem holds in it in an even stronger
fonn than before. The reason it does not violate Lindstrom's Theorem is that the
law of excluded middle fails in IF first-order logic, which it is one of the
apparently minor premises of Lindstrom's Theorem but whieh turns out to be
crucially important for its applicability.
Both ordinary and IF first-order logic admit of a simple translation to the
corresponding second-order language. The translation t(S) of a sentence S simply
asserts that its Skolem functions exist. In other words, if S is of the fonn
S[(?JX)SI [x]] , where (?JX)SI[X] occurs within the scope of the universal quantifiers
('iYl), ('iY2), ... , it is replaced by
(?Jf)S[S 1 [f(yl,Y2, ... )]]
where f is a new function symbol expressing one of the Skolem functions of S. (It
is of course assumed that S is in the negation nonnal fonn, i.e., that all its
negation signs are prefixed to atomic sentences and that the only propositional
connectives in S are &, v.) In order to reach t(S), this elimination of first-order
existential quantifiers must be applied to all of them. Likewise, whenever S is of
thefonn
where (Sl v S2) occurs within the scope of ('l:fYl) , 'l:fY2), ... , it is replaced by
S[SI & g(yl, Y2, ...) = 0) V (S2 & g(yl, Y2, ... ) "" 0)].
In other words we must also associate a "Skolem function" to each disjunction

dependent on universal quantifiers. It is immediately seen that the translation
t(S) belongs to what is known as the ~l fragment of the second-order language in
question, that is, its only second-order ingredient is an initial sequence of
second-order existential quantifiers. The remarkable thing here is that this
fragment can be translated back into the corresponding IF first-order language.
The failure of the law of excluded middle is an inevitable consequence of the
most obvious and most natural seman tical rules for IF logic. (Hintikka 1996,
Chap. 7.) Admittedly, I can extend the IF first-order logic by introducing
contradictory negation by fiat. But then it turns out that this contradictory
negation --, cannot be given any semantical rules except for saying that it is the
contradictory negation. As a consequence, it can occur only sentence-initially (or
fronting an atomic formula). In the resulting extended IF first-order logic, most of
the "nice" metatheorems listed above fail.
One of the most remarkable things about IF first-order logic is that it extends
what can be done in the foundations of mathematics by purely logical means.
Consider for instance, the following statement
(5) ('l:fx)('l:fz)(3yl'Iiz)(:Jul'lix)Ax) :::J B(y &

(B(z) :::J A(u & (y =z) B (u =x)
The second-order translation of (5) is
(6) (3f)(3g)('l:fx)('l:fz)A(x) :::J B(j(x) &

(B(z) :::J A(g(z) & z =f(x B (x =g(z
A comparison with (1) shows that (6) in fact expresses the equicardinality of the
extensions of A(x) and B(x). At the same time, (5) is perfectly first-order. Its
quantifiers "range over" individuals, not over higher-order entities. It expresses
a combinatorial fact about the model ("world") in question. It could even be
expressed by merely revising the received scope conventions of first-order logic.
Indeed, we could reach the entire IF first-order language in this way, merely by
allowing that the scope associated with a quantifier need not be a continuous
segment of the formula in question. For instance, the Henkin quantifier formula
(2) (or (4 could then be written as
(7) ('v'x){(3y)('v'z)} (3 u){S[x,y,z,y]}
where the brackets {} indicate the scope of (Vx).

Consider next the statement
(8) ('v'x)(Vz)(3yNz)(~uNx)y -:t:- x) &

(u -:t:- z) & x =z) ~ (y = u)
This is easily seen to be equivalent to
(9) (3f)(Vx)('v'z)(f(x) -:t:- x) & x = z) ~ (j(x) = f(z
which clearly is true if and only if the universe of discourse is infinite (or empty).
Likewise, the topological notion of continuity can be expressed by means of IF
first-order logic, as can be several other basic mathematical concepts that could
not be expressed in terms of ordinary first-order logic. In extended IF first-order
logic, even more concepts become expressible. They include mathematical
induction, well-ordering, power set, etc.
One might nevertheless doubt the power of IF first-order logic as a tool in
mathematics. It corresponds after all only to a small fragment of second-order
logic, viz. to its :El fragment. Even extended IF first-order logie does not extend
any further than to the:El u nl fragment of second-order logic. Neither one goes
beyond second-order logic in contradistinction from a full higher-order logic. I
will return later to this apparent restriction.
4. DIFFERENT KINDS OF COMPLETENESS AND INCOMPLETENESS
In any case, IF first-order logic extends significantly the range of what logic can
do in the foundations of mathematics. But it seems to have an important
shortcoming: It is incomplete. It is impossible to give a list of axioms from which
all the valid formulas of IF first-order logic can be derived by purely formal
rules. Or, to speak the language of mathematical logicians, the set of all valid
formulas of IF first-order logic is not recursively enumerable.
Far from being a defect, however, this feature of IF first-order logic can be
turned into a forceful reminder of the fact that the notion of completeness
traditionally used in the foundations of mathematics is a mess, albeit perhaps
not a hopeless mess. If the reader has found it hard to grasp what is meant by
different references to completeness and incompleteness earlier in this paper, he
or she has had a good reason to feel at sea. As I have pointed out on an earlier
occaslOn, (Hintikka 1989; 1996, Chap, 5) there are at least three entirely
different notions of completeness and incompleteness. They even apply to
different kinds of theories
(i) Descriptive completeness applies to nonlogical axiom systems. It means that
the models of the system include all and only intended models.
This is a model-theoretical notion in the sense that only the notions of truth
and validity are involved in it, but not any axiomatization oflogic.
(ii) Semantic completeness is a property of a so-called axiomatization of some
part of logic. It says that the axiomatization in question effects a recursive
enumeration of all valid formulas.
(iii) Deductive completeness is a property of a nonlogical axiom system
together with an axiomatization of the underlying logic. It says that from the
axioms of this system one can logically prove either S or ~S for each sentence S
of the language in question.
The drastic differences between these notions can be illustrated by relaxing
our distinctions to Godel's first incompleteness theorem. The theorem establishes
the incompleteness of elementary (first-order) arithmetic, but in what sense? The
answer is clear. Godel showed the deductive incompleteness of elementary
arithmetic. But this is not the only incompleteness in town, and it concerns more
the computational manipulability of proofs in first-order logic than its power to
capture the right structures as its models. This latter power is at issue in
descriptive completeness and incompleteness, not in deductive ones. So the sixty-
four-thousand-dollar question becomes: Does Godel's result entail the descriptive
incompleteness of arithmetic? A closer examination shows that it does so only if
the underlying logic is semantically complete. Of course, the logic Godel was
relying on, ordinary first-order logic, had just been proved complete by Godel
himself. But there is nothing in Godel's result that precludes the possibility that
by using some other kind of logic, which would have to be semantically
incomplete, we could formulate a descriptively complete axiomatization of
elementary arithmetic.
This is one of the many directions in which IF logic opens new possibilities. Its
semantical incompleteness turns out to be a hidden asset, in that it might open
new possibilities of formulating descriptively complete theories. Actually, it
turns out that although unextended IF first-order logic does not allow a complete
axiomatization of elementary number theory, extended IF first-order logic does
so. Indeed, one way of obtaining a descriptively complete axiomatization is as
follows:
Let rp be the sentence

(jZ)[(VX)(VY)(I}'11 (VNY)1}'12)(vNx){J21 (vNy) 1J'22)]
where
1}'11: -,( x =y) 1}'12: -,( x =0)
({J21: -,(y =z) ({J22: -,(x =y) & -, (y =f(x
Let '" be the conjunction of the sentence s
(Vx)-,(f(x) =0), (Vx)(Vy)(f(x) =f(y) ::J X =y), (Vx)(x:t: 0 ::J ~y)(j(y) =x.
It has been shown by Sandu and Vaananen that M~ (I}' & ~'1) iff M is a non-
standard model of arithmetic (Sandu and Vi:Uinanen 1992). Hence M~ -,(1}' & "')
(where -, is contradictory negation) iff M is the standard model. (lowe this
observation to Gabriel Sandu.)
5. AXIOMATIZING MATHEMATICAL THEORIES
This does not yet show anything about the prospects of formulating descriptively
complete axiomatic theories by means of IF logic in general. This matter can be
put into a clear perspective by noting that practically all usual mathematical
theories admit of a descriptively complete axiomatization in higher-order logic.
There are, for instance, scarcely any major unsolved problems in mathematics
that cannot be expressed faithfully in terms of second-order or higher-order logic
(cf. Shapiro 1991). This is possible because higher-order logics are semantically
incomplete.
This descriptive completeness is nevertheless bought at a very high price.
This price is not the loss of semantical completeness but the vexing problems
concerning set existence and more generally the existence of higher-order
entities that are inevitable in higher-order logic. These difficulties are the
justification of branding higher-order logic "set theory in sheeps' clothing". For it
is in set theory that the questions of set existence come to a head. These
problems are often thought of as being mainly about how to avoid paradoxes in
making existence assumptions in set theory. In reality, the really puzzling
difficulties in set theory concern the formulation of sufficiently strong
assumptions of set existence, as will be emphasized below in connection with the
requirement of standard interpretation.
As a consequence of the haunting difficulties, it becomes important to inquire
whether descriptively complete axiomatizations for mathematical theories can be
formulated by means of a suitable first-order logic, which of course has to be

semantically incomplete,6
Here I can turn an old confusion into a successful strategy. In the light of
hindsight it can be said that Frege and Russell dealt - or tried to deal - with
higher-order logics as if they were many-sorted first-order logics, with each type
constituting a separate "sort". The interrelations of different sorts can be
handled in first-order terms. So what did Frege and Russell miss? What they
missed was the intended standard interpretation of higher-order quantifiers.
Standardness is here understood in Henkin's sense, as a requirement that
bindable variables of any given type range over all the extensionally possible
sets of entities of the appropriate lower type (Henkin 1950; Hintikka 1995a). For
instance, set variables range standardly over the entire power set of the domain,
that is, all extensionally possible sets, and not only over sets that e.g. exist in the
sense of some set theory. Likewise, function variables range over what are
sometimes called all arbitrary functions, relation variables over "relations in
extension", as Russell called them, and so on. Standardness cannot be enforced
by first-order means, for instance through first-order axiomatic set theory, as is
illustrated by the so-called Skolem paradox. Because of this failure of first-order
logic to capture the standard interpretation, dealing with higher-order logic as a
many-sorted first-order logic cannot ever succeed completely. As was spelled out
by Ramsey (1931), many of the shortcomings of Russell's and Whitehead's
Principia can be traced to their failure to countenance the standard
interpretation of higher-order quantifiers.
What is remarkable here is that the many-sorted first-order treatment almost
succeeds. For the only thing in higher-order theories which cannot be captured
by means of ordinmy (many-sorted) first-order logic is precisely the requirement
that for each extensionally possible collection of lower-type entities there
corresponds an entity of the appropriate higher type. This point can be
illustrated by reference to what happens in general topology (Kelley 1955).
There all of the most elementary concepts can be formulated in terms of ordinary
first-order quantification over points on the one hand and sets on the other
hand. This includes such concepts as base, closure, openness, etc. But when we
proceed further, at some point we begin to need the assumption that
quantification over, say, the subsets of a given set really means quantification
over all the extensionally possible subsets. One place where this is likely to
happen is in connection with the concept of connectedness.
Obviously, the missing standardness requirements can be formulated quite
simply (cf. Hintikka 1996, Chap. 9). For each type (higher than that of
individuals and n-tuples of individuals), it suffices to require that it contains
A REVOLUTION IN THE FOUNDATIONS OF MATHEMAT1CS? 55
representatives of all the sets of entities of the appropriate lower type. And such
a requirement can be expressed by n1 second-order statements. We have no
reason to resort to any higher-type conceptualizations for the purposes of the
sortal reconstruction of higher-order logic and higher-order theories. We do not
need any more complex second-order formulas, either. With a modicum of stage-
setting, all such n1 statements can even be integrated into one single n1
statement.
These simple observations have striking consequences. They imply among
other things that each mathematical theory that can be formulated by means of
higher-order logic can be construed as a n1 theory. The set of structures that the
given higher-order theory can capture as the set of its models ean be captured by
a n1 theory, albeit imbedded in the additional structure which the many-sortal
reconstruction introduces. In this sense, n1 theories are all that is needed in
ordinary mathematics. Indeed, there is a sense in which the basic form of
mathematical reasoning viz. mathematical induction, formulated generally as
the closure of a certain set under some mathematical operations belongs to n1
logic and hence to extended IF first-order logic. For it was shoVim by Moschovakis
(1974) that any such inductive theory is equivalent to a n1 sentence.
But each such n1 theory can be translated so as to become a theory expressed
in extended IF first-order logic. Hence there is a sense in which all ordinary
mathematics can be reduced to extended IF first-order logic.
Moreover, in the many-sorted first-order reconstruction of a higher-order
theory, all the putative theorems can obviously be expressed as ordinary first-
order statements. The question of the status of such a putative theorem T in an
axiom system X thus becomes a question concerning the validity of
(10) (X:=J 1)
where X is n1 and T first-order. But that means that (10) is itself I1. Hence it
has a translation in IF first-order logic. Hence there is a sense in which every
problem of ordinary mathematics is equivalent with the problem of validity for a
sentence of IF first-order logic.
In the sense that appears from these remarks, IF first-order logic is the only
logic that is in principle needed in ordinary mathematics.
I am not suggesting for a moment that IF first-order logic is a practical
framework for doing mathematics. For such purposes, a judic:ious use of second-
order logic, typically restricted to I1 and n1 conceptualizations, seems to be the
best bet. In practice, the upshot would probably be something rather like general
topology (Kelley 1955). The main advantages of my reduction of mathematical
theories to IF first-order level are philosophical and theoretical. The main payoff
is a complete liberation of mathematical theorizing from all problems of set
existence and of all problems concerning the existence of higher-order entities in
general. The question of the validity of a sentence of IF first-order logic is a
purely combinatorial one: It concerns the possibility or impossibility of different
structures of individuals (particulars). It is nominalistic in Quine's sense. It is to
such combinatorial problems that I am reducing all questions of theoremhood in
mathematical theories. This is a tremendous advantage in principle, for the
problems of set existence are notoriously the most confused and confusing ones
in the foundations of mathematics.
6. IF FIRST-ORDER LOGIC IS SELFAPPLICABLE MODEL-THEORETICALLY
So far, I have not said anything of the second massive reason why ordinary first-
order logic has not been thought of a self-sufficient foundation for mathematical
reasoning. This reason is that the model theory of ordinary first-order logic
cannot be done by means of itself. The crucial concept of all model theory is the
concept of truth (truth in a model), and it was noted earlier that for Tarski-type
truth definitions we need second-order concepts.
In this respect, too, IF first-order logic puts things in a new perspective. The
result has been investigated in some detail elsewhere (cf. Hintikka 1996, Chap.
6). Suffice it here to indicate only the most general features of the situation. Both
Tarski-type truth-predicates for a given finite first-order language and the game-
theoretical ones that are in many ways superior to the Tarskian ones can be
expressed as 2:1 statements in the corresponding second-order language. This
holds also for truth-definitions for an (unextended) IF first-order language. But
such 2:1 statements can be translated back into the correlated IF first-order
language.
In such a language we therefore can formulate a truth-predicate for itself.
Paradoxes are avoided because of the inevitable failure of the law of excluded
middle in IF first-order logic (Hintikka 1996, Chap. 7). This failure creates the
truth-value gaps that are needed to avoid a contradiction.
In extended IF first-order logic we have a contradictory negation present.
However, it can occur only sentence-initially. This makes it impossible to apply
the diagonal lemma so as to create a liar-type sentence that would give rise to a
contradiction.
Truth-predicates are not all that there is to the model theory of a language or
a theory. However, their possibility is an eloquent indication that we do not need
set theory or higher-order logic for the main ingredients of a model theory of a
given first-order theory. Furthermore, it is easily seen that many other familiar
notions of model theory can be formulated in terms of a (possibly extended) IF
first-order language.
7. A NEW PERSPECTIVE ON THE FOUNDATIONS OF MATHEMATICS
The total picture of mathematical thinking that is suggested by these results

differs sharply from the usual one. The core of mathematical thinking is not set
theoretical. We can in principle dispense with set-theoretical assumptions and
concepts altogether. Instead, mathematical thinking can - and perhaps should
- be thought of as combinatorial, dealing with the struct.ures of particular
objects, especially with questions as to which configurations of individuals are
possible and impossible. These are essentially the same kinds of questions or are
dealt with in received first-order logic.
These results might in one respect seem too good, or perhaps too simple, to be
true. If all of ordinary mathematics can be done in extended IF first-order logic
or, equivalently, in the Ll u III fragment of second-order logic, what use and
what interest does the rest of higher-order logic have? Surely it has some content
and use beyond its simplest special case.
I am not denying that higher-order logic has uses beyond the Ll u II1 case.
But the crucial question is: What kind of use? By and large, logic has two kinds
of uses in its applications, especially in mathematics (Hintikka 1996, Chap. 1).
They can be characterized as follows:
(a) Descriptive use. This is the contribution oflogic to the specification of the kind
of structure or structures that a mathematician or a scientist wants to study.
One's aim in this direction can for instance be to formulate a descriptively
complete axiom system. An example might be offered e.g. by the use of
quantifiers and other logical constants in the axioms of a mathematical or
scientific theory. The descriptive function of logic could and perhaps should be
called its theoretical function.
(b) Deductive use. In this employment, logical concepts are used to study valid
inferential relationships between propositions.
This function might also be called the computational or explicative function of

logic in axiomatic theories. What has been seen is that the task (a) can typically
be performed by logic already on the first-order leveL But from this it does not
follow that the deductive task can be so performed. Indeed, it can be seen that
the deductive task is awkward to fulfill as soon as we go beyond ordinary first-

order logic. For what a logician would like to do is to express in his or her
language the proposition that says that if SI is true, then S2 is true. In ordinary
propositional logic, this is expressed by
But in IF first-order logic (11 ) does not semantically speaking say that S2 is true
if SI is. What it says is that either Sl is false or S2 is true. Since in IF first-order
logic there are propositions that are neither true nor false, this is an
unnecessarily strong requirement. What is needed in the study of logical
inference is a.statement whose truth authorizes us to infer the truth of S2 from
the truth of SI. Such a statement can be formulated in higher-order terms.
Expressed in game-theoretical jargon it will say that there is an effective
(recursive) functional J which from a winning veri:i.er's strategy cp in the game
SI[cp, IfI1 associated with SI yields a winning veri:i.er's strategy J(cp) in the game
S2[ ip, IfI1 associated with S2 (Hintikka 1993). Winning strategies are in each case
characterized by the fact that they result in a win even when one's opponent is
aware of them. The possibility of infering the truth of S2 from the truth of SI will
thus be expressible by
However, it is natural to require also that the knowledge of a strategy falsifying

S2 should enable us to find a strategy falsifying Sl. This is captured by changing
(12) into
which is Godel's functional interpretation for conditionals (Go del 1986b). It can
be seen that (12) is of a higher order (type) than Sl or S2. When conditionals are
nested in a sentence, as we might very well want to nest them for purposes of
inference, we are pushed to even higher types.
What this means is that for inferential and deductive purposes, there can be
plenty of reasons to use higher-order logics. The overall picture of the
foundations of mathematics which we thus arrive at is almost diametrically
opposite to the traditional one, explained above. The descriptive, that is, model-
specifying tasks in mathematics can be accomplished by means of relatively
elementary (first-order) logic. These descriptive tasks are paramount on the level
A REVOLUTION IN TIlE FOUNDATIONS OF MATIiEMATICS? 59
of theorizing, be this theorizing mathematical, scientific or philosophical. It is the

inferential, deductive and computational aspects of the task of a mathematician
that force us to consider increasingly higher-order logics and draft them to our
service. What these computational aspects amount to is to elicit consequences of
a theory which has already been formulated concerning a class of structures
(models) which already have been captured by the theory, assuming of course
descriptive completeness. Hence the use of higher-order logics belongs to the
technology of logic rather than its theory.
Obviously neither the descriptive nor the deductive task of logic should be
neglected. In our day and age, which is overwhelmingly the age of computers
and computing, equally obviously it is the descriptive and theoretical function
that is in danger of being neglected. 7
NOTES
1 See Per Lindstrom, (1969), pp. 111. The most accessible formulation of his result is
found in H.D. Ebbinghaus, J. Flum and W. Thomas, 1984, chapter 12.
2 An important qualification needed here is that Frege adopu!d a nonstandard
interpretation (in Henkin's sense, see Note 4 below and Henkin 1950) which means that
his higher-order logic could be dealt with like a many-sorted first-order logic. See Jaakko
Hintikka and Gabriel Sandu.
3 Indeed, if a first-order theory is deductively incomplete, then it is also descriptively
incomplete as pointed out. On the other side, if a theory is descriptively complete, in the
sense of e.g. characterizing up to isomorphism the standard model of arithmetic, then
the theory is semantically incomplete, hence it cannot be a first-order theory. This
follows from facts proved in Jon Barwise and S. Feferman, 1985.
4 For the standard vs. nonstandard distinction, see Jaakko Hintikka,(1995a) pp. 21-44.
5 With the following, cf. Jaakko Hintikka (1996), chapters 3-4 as well as "A revolution
in logic?" (forthcoming).
6 By the results mentioned at the end of Section 4 above, if we have a descriptively
complete axiomatization of a non-trivial mathematical theory, the theory must be
semantically incomplete, hence it cannot be an ordinary first-order theory. (Of course it
can be an IF first-order theory.)
7 In working on this paper, I have greatly profited from the advice of Dr. Gabriel
Sandu.
REFERENCES
Barwise, J. and S. Feferman, 1985, Model-Theoretical Logics, Springer-Verlag, New

York.
Ebbinghaus, H.D., J. Flum, and W. Thomas, 1984, Mathematical Logic, Springer-
Verlag, New York.
COdel, K., 1986a, "Uber formal unentscheindbare Siitze der Principia Mathematica und
verwandter Systeme", in K. COde!, editor, Collected Works Vol. 1, Oxford University
Press, New York, pp. 144-195.
COde!, K., 1986b, Collected Works Vol. 2, Oxford University Press, New York.
Henkin, L., 1950, "Completeness in the Theory of Types", Journal of Symbolic Logic Vol.
14, pp. 81-91.
Hintikka, J., 1989, "Is there Completeness in Mathematics after CO del?", Philosophical
Topics Vol. 17, pp. 69-90.
Hintikka, J., 1993, "COdel's Functional Interpretation in a Wider Perspective", Yearbook
1991 of the Kurt Godel Society, The Kurt COdel Society, Vienna, pp. 5-43.
Hintikka, J., 1995a, "Standard vs. Nonstandard Distinction: A Watershed in the
Foundations of Mathematics", in J. Hintikka, editor, From Dedekind to Godel,
Kluwer Academic, Dordrecht, pp. 21-44.
Hintikka, J., 1995b, "What is Elementary Logic? Independence-Friendly Logic as the
True Core Area of Logic", in K. Gavroglu, J. Stache!, and M. Wartofsky, editors,
Physics, Philosophy and the Scientific Community, Kluwer Academic Publishers,
Dordrecht, pp. 301-326.
Hintikka, J., 1996, The Principles of Mathematics Revisited, Cambridge University
Press, Cambridge.
Hintikka, J. and G. Sandu, 1992, "The Skeleton in Frege's Cupboard: The Standard vs.
Non-standard Distinction", Journal of Philosophy Vol. 89, 290-315.
Kelley, J., 1955, General Topology, D. van Nostrand, Princeton.
Lindstrom, Per, 1969, "On Extensions of Elementary Logic", in W. Aspray and P.
Kitcher, editors, History and Philosophy of Modern Mathematics, University of
Minnesota Press, Minneapolis, pp. 95-135.
Moschovakis, Y., 1974, Elementary Induction on Abstract Structures, North-Holland,
Amsterdam.
Ramsey, F.P., 1931, "The Foundations of Mathematics", in R.B. Braithwaite, editor, The
Foundation;> of Mathematics and Other Logical Essays, Routledge and Keegan Paul,
London.
Sandu, G. and J. Viiiiniinen, 1992, "Partially Ordered Connectives", Zeitschrift fUr
mathematische Logik und Gnmdlagen der Mathematik Vol. 38, pp. 631-72.
Shapiro, S., 1991, Foundations without Foundationalism, Clarendon Press, Oxford.
A REVOLUTION IN 1HE FOUNDATIONS OF MA1HEMATICS? 61
Tarski, A., "The Concept of Truth in Formalized Languages", in Logic, Semantics,

Metamathematics: Papers from 1923 to 1938, Clarendon Press, Oxford, pp. 152178.
4
IS THERE COMPLETENESS IN MATHEMATICS AFTER GODEL?
TWO FUNCTIONS OF LOGIC IN MATHEMATICS
What do mathematicians do? There are many ways of approaching this

que~tion, but one kind of answer is easy to give. Mathematicians study
structures of different kinds: for instance f natural numbers, real numb-
ers (the continuum), geometrical structures, groups, lattices, topological
spaces, etc. For these structures, they develop corresponding theories,
such as number theory, real analysis (theories of measure and integration),
Euclidean geometry, group theory, lattice theory, topology, etc. In the late
nineteenth century, a large and central class of foundational problems
came up concerning the nature, the basic assumptions, and the presuppo-
sitions of theories of this general kind. These problems included questions
concerning the axiomatization of important mathematical theories, the
definitions of their basic concepts (e.g., natural numbers, real numbers,
different geometrical objects), the relation of these mathematical theories
to logic, etc. These questions are clearly among the foundational problems
which working mathematicians are likely to find relevant. Hilbert's efforts
in his axiomatization of elementary geometry constitute a representative
example of this early foundational work.l
My starting point in this work is the question: How can logic help
mathematicians in their studies? This is essentially a question concerning
the role of logic in the foundational inquiries. Clearly, modem logic made
IS THERE COMPLETENESS IN MATHEMATICS AFTER GODEL? 63
many of its first inroads into the foundations of mathematics precisely as

a helpmeet to attempted solutions of the classical foundational problems
indicated above. The role of logic in mathematics is somewhat ambiv-
alent, however. Since World War II, philosophers' and even logicians'
interest in foundational questions has apparently waned considerably.
My work here represents an attempt to return to some of the classi-
cal foundational problems in a new way. I propose to rethink the role
of logic in the foundation of mathematics and thereby to gain a novel
perspective on some of the classical problems. It is a sobering thought
that we still do not have good solutions to all these classical problems.
The main reason for the decreased attention paid by philosophers and
logicians to the good old-fashioned foundational problems is that the tools
of modem logic have by and large proved less successful in dealing with
actual mathematical theories and their foundations than was first hoped.
This relative failure is connected with the first and foremost initial use
of logic in mathematics. This use was to put logical tools to the service
of an explicit axiomatization of different mathematical theories. More or
less fully formalized axiom systems for different mathematical theories
were put forward, including Peano arithmetic and its variants, Hilbert's
axiomatization of elementary geometry, and axiomatic set theory, among
others. The results of these axiomatization efforts have been rather frustrat-
ing in the main. Even as simple and fundamental a theory as elementary
number theory cannot be given a complete axiomatization -- or so Gooel's
famous incompleteness theorem is supposed to show. Small wonder, there-
fore, that more challenging mathematical theories (for instance, set theory)
don't seem to admit anything remotely like a complete axiomatization.
Indeed, the value of logical axiomatization for mathematical purposes
seems to be severely limited, at least on the usual first-order level.
How frustrating the apparent consequences of Gooel' s incompleteness
result are can perhaps be seen by considering any explicit axiom system
for elementary number theory (of the kind normally presented in logical
and mathematical literature). When a mathematician uses a quantifier in
such an axiomatic theory (i.e., speaks of all numbers or some numbers) he
or she has no guarantee that what is involved are just the natural numbers
mathematicians are really interested in. Inevitably, the "all numbers"
involved must in principle be capable of including also "non-natural"
or "non-standard" numbers infinitely greater than the "real" natural
numbers. A mathematician will be, in brief, a victim of Russell's onetime
quip: he is a man who never knows what he is talking about.
Here one can begin to see how this view of the implications of Gooel' s
theorem discourages realistic interpretations of mathematics. If by speaking
of "all natural numbers" we don't really mean just all natural numbers,
how can quantifiers be understood realistically when they occur in math-
ematical theories? We shall return to the question of realism in section VII.
My main thesis is that this entire pessimistic view is mistaken.

There are much better prospects for getting a grip of the classi-
cal problems in the foundations of mathematics by means of the
tools of modern logic than most recent observers have realized. In
order to make use of these possibilities, however, we must take
a fresh view of logic itself and of its role in mathematics.
What can logic do for mathematics? The roles of logical ideas and
concepts in mathematics have been many and many-faceted. There are
nevertheless two uses of logic in mathematics which everybody will
recognize, but which have not usually been separated from each other
sufficiently sharply. Their interrelations have not received their due atten-
tion, either. These two are:
(i) The uses of logical notions, such as quantifiers, propositional con-
nectives, the general concepts of relation and function (codified in the
use of relation and function variables), etc., for the purpose of capturing
certain structures, viz., the different structures studied in various math-
ematical theories. The pursuit of this task typically leads to the formulation
of axiom systems which use the logical concepts just mentioned for the
different mathematical theories. The sundry axiomatizations mentioned
above hence offer examples of this function of logic in mathematics. I
shall call this the descriptive use of logic in mathematics.
(ii) In order to facilitate, systematize, and criticize mathematicians'
reasoning about the structures they are interested in, logicians have isolated
various valid inference patterns, systematized them, and even constructed
various ways of mechanically generating an infinity of such inference
patterns. I shall call this the deductive use of logic in mathematics. The
typical form of a systematization of this deductive function of logic in
mathematics consists of a number of valid logical axioms from which
other valid formulas (or valid inference patterns) can be derived by means
of so-called rules of inference. The derived formulas or patterns are often
called logical theorems. Such theories of systematizing inference patterns
(or alternatively logically valid formulas) are thus know as axiomatizations
of (some part of) logic.
It is important to realize that such an axiomatization of logic is an alto-
gether different enterprise from the axiomatization of various particular
mathematical theories. The latter belongs to the descriptive task of logic
in mathematics. The former is a part of the deductive task of logic. For
instance, logical axioms ought to be separated sharply from mathematical
axioms. The two perform an essentially different function in mathematics.
IS THERE COMPLETENESS IN MATHEMATICS AFrER GODEL? 65
II
MODELS AND COMPLETENESS
The first item on my agenda is to examine the interrelations IQf the descrip-
tive and the deductive roles of logic in mathematics. Each of the two roles
is based on the concept of a model. but in a different way. Given a sentence
or formula S. a certain class of structures M(S) is thereby determined.
They are called the models of s. If S is an interpreted sentence, the
members of M(S) are the models (sometimes dubbed "possible worlds")
in which S is true in the usual sense. If S is a logical formula, ME M(S)
if and only if S is satisfied by every assignment of values to free variables
in M; and so on. (More about truth and satisfiability later.)
A putative inference from a sentence S to another sentjence S' is said
to be valid if and only if
(1) M(S) ~ M(S')
A formula S is contradictory iff
(2) M(S) = 0 (empty set)
A formula S is valid iff M(S) consists of all models (of the language in
which S is formulated).
An axiomatization of some part of logic is inferentially or, as it is
often put, semantically complete iff all valid formulas or all valid inference
patterns (in the underlying language) can be derived in it as theorems.
Models are crucial for both tasks of logic in mathematics, but in different
ways. Given a system of mathematical axioms, say S, it captllI1es certain struc-
tures, viz., those exemplified by its different models M E M(S). A mathematical
axiom system S is complete - I shall call it descriptively complete - if and
only if M(S) coincides with the class of intended structures, i.e., the structures S
was designed to capture. Whenever M(S) consists of one member only, modulo
isomorphism, S is said to be categorical. In other words, a categorical theory
(axiomatic theory) determines its models uniquely up to isomorphism.
There is a third kind of completeness which combines ingredients from
the descriptive and from the deductive dimensions of logic. If a mathemati-
cal theory S is combined with an axiomatization of logic, we can say that
S is deductively complete iff, for each sentence C (in the language of the
theory S), either C or . . ., C can be derived from S by mt:ans of logical
theorems, i.e., iff I- (S ::> C) or I- (S ::> . . ., C) holds for each such C.
The deducti ve completeness of a nonlogical (e. g. , mathematical) theory
presupposes in a sense the semantical completeness of the underlying
logic. For if S is deductively complete, the underlying logic is obviously
semantically complete, at least as far as the consequences. of S are con-
cerned. For if anything more were provable from a deductively complete

theory S, this theory would be inconsistent.
Of course, over and above this semantical completeness of the under-
lying logic, the deductive completeness of S puts a strong requirement on
the models of S. It is important to realize what precisely this requirement
is. The deductive completeness of S implies that the same formulas (of the
language of S) are true (satisfiable) in all the models of S. This does not
entail that S is categorical, however. It implies a weaker similarity between
all the models of S than isomorphism. This weaker similarity is known in
the trade as elementary equivalence.
The need of consciousness-raising among philosophers and mathema-
ticians is vividly demonstrated by the fact that differences between these
entirely different kinds of completeness are seldom spelled out clearly, and
by the fact that their interrelations are not attended to seriously enough in
spite of the fact there is room for radical rethinking concerning them. (Cf.
section IV.)
III
FROM ONE WORLD (MODEL) TO MANY
Presenting these concepts and their definitions explicitly is likely to bore

many contemporary readers. The preceding section I in effect merely
spelled out the basic ideas of a model-theoretical (semantical) way of
looking at the foundations of logic. The main thing that evolves from
the definitions is the absolutely crucial role of the sentence-model rela-
tion. (I shall return to this role soon.) This relation both underlies the
descriptive function of logic (as you saw) and also determines the aims
of the deductive task of logic in mathematics. As was pointed out above,
it is part and parcel of the model-theoretic way of thinking that neither the
idea of specifying certain structures nor the idea of valid inference makes
any respectable sense except in model-theoretic terms.
The impression of triviality that the explanations I have so far offered
probably have generated can perhaps be best dispelled by recalling that
the model-theoretical conception of logic which the definitions sketched
above codify has not always been taken for granted. On the contrary, it has
evolved slowly in the course of the twentieth century. 2 This evolution still
has not been completed in two different senses. On the one hand, the full
force of the developments at the cutting edge of the conceptual innovation
has not yet hit many philosopher-logicians, including some of the most
influential ones. It seems to me that confusion is especially rampant in the
foundations of modal and intensional logics. On the other hand, the evolu-
tion of a consistently model-theoretic idea of logic has not yet reached its
natural. not to say inevitable. conclusion. Indeed. one of my main aims in

this work is to indicate a major new step further in the same direction.
The starting point on this development was a view. represented among
others by Frege. which dispensed with the idea of model altogether. 3
So-called logical truths were not thought of as being essentially different
from plain ordinary truths except for being more general. (Indeed. logical
truths were thought of as the only completely general truths). Hence logical
axiomatization did not differ. according to the representatives of this view.
from nonlogical (e.g . mathematical) axiomatization. Moreover, neither
kind ofaxiomatization needed a model-theoretical foundation. it was
alleged. Typically, all sentence-model or theory-model relations were
thought of as being impossible to speak of in language and hence impos-
sible to theorize about.
This view. or syndrome of views, has been slowly (and partially)
replaced by the model-theoretical approach outlined earlier in this section.
This change is a change in kind and not a change in degref:. At first sight,
it might seem that the model-theoretical viewpoint merely represents one
possible interpretation of the idea of generality as the hallmark of logical
truth, in that in the model-theoretical approach generality is interpreted
as generality with respect to models and not just over individuals in
one model (world), as in Frege. Why cannot these new-fangled model-
theorists' alternative models simply play the same role as the more remote
Platonic regions of Frege's unitary universe? This attempt to assimilate
the two viewpoints to each other is deeply mistaken, however. On the
model-theoretical view, the totality of models never comes into play in
the language. A sentence specifies its models by putting conditions on
each of these models, as it were one by one. For instance, and most
importantly, the (first-order) quantifiers used in a theory S never range
over anything more than the individuals of some one model (however
arbitrarily selected) with respect to which S has been interpreted. Here
the contrast with Frege is at its sharpest; for Frege all quantifiers in a
well-formed language ranged over literally all entities in the word, not
forgetting even such arcane "objects" as "The True" and "The False"
(the truth values).
Philosophers sometimes talk as if the idea of possible world were
merely a gimmick designed to extend model-theoretical conceptualizations
from the basic extensional logic to modal logics. This extension is
a relatively trivial one, however. The much more important fact is
that all model theory, including the model theory of garden-variety
first-order logic (quantification theory, predicate cakulus) involves
a multitude of models, which are misleadingly dubbed "possible
worlds" by philosophers. What happens in the transition to the
semantics of modal and intensional logics is merely that those models
are no longer considered one at a time, but in rdation to each
other. What really counts is nevertheless the multiplicity, not the interrela-
tions. In this sense, all model theory is "possible worlds semantics."
In fact, the development away from the Frege paradigm has been
essentially the growth of the multiple-worlds or multiple-models idea.
For Frege and others in the same tradition, the interpretation of one's
language is cut and dried. It cannot be varied, and since it is designed to
speak of this actual world, all talk of other interpretations (other models)
is either nonsense or at best (as for Quine) merely a useful technique for
formal logicians in the development of their theory, devoid of any general
theoretical significance. The development I have mentioned thus means
that logicians have started taking systematic liberties and varying the inter-
pretation of one's language. The model-theoretical viewpoint described
above is little more than a systematic way of carrying out the program of
systematic variation in the interpretation of one's language.
IV
VARYING SENTENCE-MODEL RELATIONS
As we mentioned, this development has not yet reached its "logical"

conclusion. What has happened in the development of contemporary
logical semantics (model theory) is that the set of models (worlds) on
which a logical language is interpreted is allowed to vary and that in some
conceptualizations (model theory proper) the interpretation of nonlogical
constants is allowed to vary. Some hardy souls have even begun to experi-
ment with nonstandard interpretations of logical concepts.
All these are piecemeal and as it were local variations, however. They
pertain to the models that are at the receiving end of the sentence-model
relations, not to these relations themselves. In the present literature the
sentence-model relation itself has usually been taken for granted. Yet
there is nothing God-given or logic-given about it. There is no theoreti-
cal objection whatsoever to experimentally varying it, too. Indeed, the
main thrust of this work is to propose that this sentence-model relation
be changed, by imposing suitable extra conditions on the models of a
theory S, over and above those that everybody is assuming in these days.
From the perspective which I am representing here, this step appears as a
step further in the same direction as the main conceptual developmental
of logical theory in the twentieth century. I shall also argue that certain
further restrictions on models are indeed tacit in traditional mathematical
conceptualizations, or can be considered rational reconstructions of tradi-
tional mathematical ideas.
What difference can "new" restraints on the models of a theory or
simply of a sentence make? Let us see what the situation is. Given a
sentence F (we can think of it as the conjunction of axioms of some
mathematical theory), it determines as was explained earlier, a class of
IS THERE COMPLETENESS IN MATHEMATICS AFfER GODEL? 69
models M(F) in accordance with the usual sentence model relationship.

Assume that this relation is made more stringent in the simplest possible
way, that is, by omitting certain structures from being potential models
of any theory. For the remaining structures, their status as models of
this or that sentence remains unaffected. This status can, for exam-
ple, be governed by the usual Tarski-type truth conditions. After the
change, F detennines a small class N(F) as its models, i.e.,
Here
(3) N(F) ~ M(F).
What this means is that the descriptive task of logic becomes easier. A
theory restricts the class of its models more narrowly than before. In
particular, more mathematical theories are likely to admit of a categorial
axiomatization than before.
What happens to the deductive task of logic in mathematics? Suppose
that G was a valid consequence of F in the "old" logic, i.e., that
(4) M(F) k M(G)
Since all models in the new restricted sense remain models in the old sense,
(4) implies
(5) N(F) ~ N(G).
This means that old valid consequences remain valid. However, (5) does
not usually imply (4). That means that new valid logical consequences can
come about as a result of the restriction. In other words, after the restriction
there are more valid logical consequence relations and more logical truths
to be captured by a logical axiomatization than before. In still other words,
the deductive task of logic becomes ceteris paribus more difficult. Further
restrictions on models will not only result in a new logic. They will also
result in a logic that is richer and hence more difficult than before. I am
tempted to speak here of truly mathematical logic in contradistinction
to the "old" general or fonnal logic. ("Mathematical logic" means
here the logic of mathematics, not a logic using mathematical tools.)
There will not be a unique mathematical logic in this sense, however,
because different further restraints on models can yield different logics.
This shows why it is important to distinguish the descriptive and the
deductive tasks of logic from each other. Not only are they different; they
can be traded for each other (to some extent). By making the one task
more difficult we can make the other one easier. One of the many ways
of fonnulating my aim in this work is to say that I want to strike a different
balance between the descriptive and the deductive task of logic from what
their relation is usually (and largely unwittingly) taken to be.
What has been seen here shows that this can be done by imposing new
general restraints on the models of a given sentence S. The significance of
the enterprise will then depend on how natural and theoretically motivated
these additional restraints are.
My basic idea of varying sentence-model relations is closely related

to the leading idea of so-called abstract logics (also known as model-
theoretical logics) which have become important in general logical theory
in the last decade or SO.4 I shall not examine this relationship here, how-
ever. It seems to me that the study of abstract logics might profit from
the philosophical and foundational perspective outlined here and perhaps
also from the specific examples of logics with an unusual concept of
model which the "extremalist" approach advocated here can give rise to.
V
GODEL'S INCOMPLETENESS RESULT AND
THE TWO FUNCTIONS OF LOGIC IN MATHEMATICS
One place where the relation of the two tasks of logic in mathematics
becomes crucially important is GOOel's famous incompleteness result. 5 It
is usually taken to imply that descriptive completeness is impossible to
achieve in as simple a theory as elementary arithmetic, let alone in more
complicated (more interesting) mathematical theories. No axiomatization
can ever weed out all unintended models, it is said.
This claim is seriously wrong, however. There is nothing in GOOel's
results that rules out the descriptive completeness of nontrivial mathemati-
cal theories (not even on first-order level). What Godel proved was that
elementary number theory is incomplete in the third of the different senses
mentioned above, i.e., deductively incomplete. Now it was pointed out
above that the deductive completeness of a theory S requires two things:
(i) a partial semantical completeness of a fragment of the underlying logic,
and (ii) a specification of the models of S, i.e., descriptive completeness.
The deductive completeness of S may fail because either one of these two
fails, i.e., it may fail in two different ways. GOOel's result does not tell us
in which way. It is only in conjunction with the semantical completeness of
the logic that GOOel is relying on - which is ordinary first-order logic -
that GOdel's result implies that one cannot have a descriptively complete
axiomatization.
In this sense, the impact of GOOel' s incompleteness theorem for number
theory depends crucially on the completeness of first-order logic, which
GOOel had proved just prior to his incompleteness result. 6 Hence, the
repercussions of GOOel's theorem will be quite different if the semantical
completeness of the underlying logic is given up. If this logic is replaced
by a different one which is not semantically complete, we face a new
situation. And the semantical completeness of first-order logic is easily
sacrificed by changing our concept of model, e.g., along the lines indi-
cated above, that is, by imposing further general restrictions on the models
of one's theories. Such restrictions can even give rise, as I hope to show
IS THERE COMPLETENESS IN MATHEMATICS AFTER GOOEL? 71
in a sequel to this work, to categorical theories of elementary arithmetic

and of various other mathematical theories which are often thought of as
being impossible to axiomatize completely on the first-order level. In this
sense, Godel did not show that the descriptive task of logic in mathematics
is impossible to carry out completely. What his incompleteness result shows
in effect is that if you reach a descriptively complete theory of elementary
arithmetic (or of any mathematical theory containing elementary arithme-
tic), then your underlying logic cannot be semantically comJPlete. You can
have descriptive completeness only at the price of semantical complete-
ness. You can have either kind of completeness. but not both.
This result is miles apart from the casual misinterpretation of Godel's
result as ruling out descriptive completeness. Hence my observation illus-
trates vividly the opposition of the two tasks of logic in malthematics, and
of their interchangability.
One reason for the interest of this interchangability is that if I offer a
working mathematician a choice between, on the one hand, a descrip-
tively complete theory of some mathematically interesting structure. com-
bined with methods of reasoning about it that cannot be completely
axiomatized logically, and, on the other hand, an inevitably incomplete
theory based on a semantically complete logic, he or she will certainly
choose the former (descriptively complete axiomatization). In contrast,
logicians and philosophers naturally aim at a (semantically) complete
theory of logical inferences in general. However, this tempting gener-
ality may be an illusion. In any case, it seems to me that philosophers'
professional emphasis on the theory of valid inferences has made them
less than fully receptive to alternative viewpoints which (I am argu-
ing) are much more useful in trying to understand actual mathematics.
The conception of mathematical activity which results from my sugges-
tions may at first seem strange to some of my readers, especially those
coming from philosophy. They are likely to think that constructing proofs
according to a fixed set of rules is not only an essential part of mathemati-
cal activity but its end-all and be-all. (In other words, the deductive task
of logic is taken to be the fundamental one.) I am of course not denying
the role of proofs in mathematical literature, only the possibility of an
exhaustive set of mechanical rules for such proofs. Moreover, in a large
number of relatively easy cases the principles we actually have formulated
suffice for the desired proofs. This is probably what prompts many phi-
losophers' oversimplified idea of the allegedly all-important role of proofs
in mathematics. However, when it comes to nontrivial problems, the situa-
tion will be different. The semantical incompleteness which we have to
accept in order to reach descriptive completeness shifts the main focus
from the constructions of proofs to whatever it takes to find more and
more proof principles (or other equivalent new assumptions). In the next
section I shall offer a few hints as to what these activities may look like.
It may also be argued that the current emphasis on mathematical proofs
is partly a historical accident whose roots go back ultimately to ancient
Greek geometry. It would be an interesting thought-experiment to try to
think how creative mathematics could be practiced without any explicit
concept of proof. A real-life embodiment of such an imagined experiment
was of course the famous Srinivasa Ramanujan. It is reported that even
after his creative collaboration with Hardy had begun "they had to come to
terms with the difference in their education. Ramanujan was self-taught: he
knew nothing of the model rigor: in a sense he did not know what a proof
was [emphasis added]. "7 Here we have a striking reminder of the fact that
the predominance of the deductive task of logic in mathematics may very
well have been exaggerated in recent philosophical discussion.
VI
"UBER EINE BISHER NOCH NICH BENUTzE ERWEITERUNG
DES KONSTRUKTIVISTISCHEN STANDPUNKTES"
How, then, can the underlying logic be changed so as to shift the balance
of power between the two tasks of logic? If we are to make logic serve
better its first descriptive task, we must clearly have a closer look at how
a logical formula is supposed to determine its models, for the basic idea of
the descriptive task of logic was to capture certain mathematical structures
as models of suitable logical formulas.
Now how does a formula (say, a first-order one) determine its models?
One possible answer, indeed the answer which is probably likeliest to
occur to a logician or philosopher, is to say that the connection between
a sentence S and the set M(S) of its models is effected by a Tarski-type
truth-definition.8 M(S) consists of those and only those models in which
S is true according to the truth-definition. (For other formulas F, M(F) is
the set of models in which F is satisfiable.)
This relation between S and M(S) is highly nonconstructive, however.
A given S does not give us, on the basis of a Tarski-type truth-definition
alone, any way of constructing M(S) or any member of M(S) even asymp-
tomatically, for, on a Tarski-type truth-definition, the truth or falsity of a
quantified sentence is essentially defined in terms of the truth or falsity of
all the substitution-instances of its unquantified part. (It is usually objected
here that even the totality of substitution-instances with respect to actual
names is not enough. This objection is not what I have in mind, but it
would only add more grist to my mill.) This point is connected with the
fact that a Tarski-type truth-definition for a given theory can !be formulated
in a richer one.
A parallel way of making the same (or closely related) point is to
emphasize that Tarski-type truth-definitions, when they are considered in
their role as bridging a theory and its set of models, give precious little
help in understanding how we can use the theory for the purpose of actually
acquiring knowledge about its models. 9
What better bridges are there, then, between a formal mathematical
theory and the class of its models? One set of bridges is created by
the various proof techniques in which a logical proof is construed as
a totally frustrated countermodel construction. Cases in point are model
set constructions \0 and Herbrand expansions. I I In them, we do in fact
seek to construct, step by step, better and better approximations towards
the different kinds of models-at least to their structures--that a given
sentence F (in a logical notation) can have. Indeed, it is easy to see that
each finite or countable model of a given first-order F can be reached in
this way in the sense that an isomorphic replica of each such model can
be reached by pushing the construction to countable infinity. In this sense,
we can think of the construction processes as the way in which a first-order
theory T determines its models.
But when we look at the actual structure of these construction pro-
cesses, they appear messy and aimless. They exhibit a pur,e laissez faire
principle: any way of introducing new elements into the model or other-
wise continuing the construction is as good as any other.
In view of this constructional anarchy, one should not find GOdel-type
incompleteness results entirely surprising. No wonder that the construction
process will not yield the intended structure as its only outcome if so many
liberties are allowed in the course of the construction.
Thus it appears not only possible but desirable to try to change the
relation of a mathematical theory or other mathematical proposition T to
the set M(T) of its models by imposing constraints on the model construc-
tion process. What constraints? They will have to be somehow intrinsic
to the construction idea. Several different constraints can undoubtedly be
motivated. However, we must have some starting point in considering
these restraints. Now I doubt that there are any constraints simpler or
more natural than the requirement of parsimony: Don't multiply entities
in your model without necessity! Don't introduce individuals into the
model you are constructing unless you have to! This is one of the con-
straints I shall study elsewhere in greater detail. I shall call it the paucity
principle.
If there is an equally natural restraint, it is the requirement that
no opportunities for introducing different kinds of individuals are to be
left unused in the eventual outcome. If the paucity principle demands
that the model to be constructed by the poorest possible one, the new
principle, which I shall call the principle of plenitude, requires that it be
the richest possible one qualitatively.
The classic example of an axiom which attempts to codify an assump-
tion of plenitude is Hilbert's Axiom of Completeness in his so-called
axiomatization of geometry (See work cited in note I above, editions 2-6) .
.In different ways, paucity and plenitude assumptions played a major role
in the classical foundational work, as witnessed inter alia by Dedekind's
ideas.
The paucity principle is a kind of minimality requirement. The prin-
ciple of plenitude is likewise a maximality assumption. Minimality and
maximality assumptions used to be called extremality assumptions. At
least one major logician believed that the future of the foundations of
mathematics lies in such extremality assumptions. In a letter to Ulam,
Godel writes, apropos von Neumann's axiomatization of set theory:
The great interest which this axiom [in von Neumann's axiomatization
of set theory1lies in the fact that it is a maximum principle somewhat
similar to Hilbert's axiom of completeness in geometry. For, roughly
speaking, it says that any set which does not, in a certain defined
way, imply an inconsistency exists. Its being a maximum principle
also explains the fact that this axiom implies the axiom of choice. I
believe that the basic problems of abstract set theory, such as Cantor's
continuum problem, will be solved satisfactorily only with the help of
stronger axioms of this kind, which in a sense are opposite or com-
plementary to the constructivistic interpretation of mathematics. 12
Unlike maximality principles, minimality assumptions of course are not
opposite to the constructivistic approach to mathematics.
A closer analysis shows that in some cases, for instance in elementary
number theory, the paucity principle (suitably implemented) does the job.
In other cases, for instance in the theory of reals, we have to impose the
principle of paucity on one part of the intended structure, viz., on the class
of natural numbers, while imposing the principle of plenitude on another
part of the structure, viz., on the reals proper. Either the paucity principle
or the principle of plenitude alone is relatively easy to implement, but the
interaction of the two creates a much subtler situation.
For either principle, the operative question concerns the way in which
it is implemented. This question, which is clearly the heart of my
"extremalist" program in the foundations of mathematics, is too large
to be discussed here adequately.
It is of interest to see how naturally both of these restraints on the
models of a theory can thus be explained by references to the process of
constructing the (countable) models of a first-order theory. This is what
I meant by the words "eine . . . Erweiterung des konstruktivistischen
Standpunktes" in the title of this section. 13 It remains to be seen, however,
by what other means these "special models" (as they are labeled in model
theory) 14 can be approached, and how they can be motivate:d further.
These restraints on Herbrand-type model constructions can be thought
of as the new element in the way in which mathematical propositions
determine their models. It remains to investigate precisely how these
restrictions are to be implemented. Among other things, iit will turn out
that both the principle of paucity and the principle of plenitude allow more
than one essentially different interpretation.
It is to be noted that the constructional restrictions I have spoken
of cannot in interesting cases be just limitations or individual steps of
construction. Rather, they are characterizations of the end product of the
entire process. Because of this, it requires a separate investigation in the
case of each different type of restraint as to how the restraint manifests
itself in the course of the step-by-step construction process itself.
VII
PHILOSOPHICAL IMPLICATIONS
Even though the points I have so far made in this essay are straight-
forward, they have certain philosophical implications. One reason for this
is that the allegedly inevitable descriptive incompleteness of mathematical
theories - and other theories - has been used as a putative argument in
philosophical discussions. For instance, some philosophers, among others
Hilary Putnam, have argued against a realistic construal of scientific theo-
ries on the ground that such a theory (allegedly) will never be able to
specify the real structure of the world anyway, because it will necessarily
have to be incomplete and hence compatible with several different possibil-
ities as to how things really are. 15 As we have seen, however, there is no
such necessity. Mathematical and empirical theories alike can be descrip-
tively complete (categorical) if it is antecedently understood that certain
further restrictions are imposed on their models. Admittedly, the logic we
have to use in dealing with such theories may then be unaxiomatizable.
But this fact is not relevant to the issue of realism any longer. It is not a
restraint on our possibilities of reaching out to reality and specifying its
structure; it is a limitation on the exhaustiveness of mechanical ways of
eliciting consequences of our theories. In other words, Putnam is relying
on a received idea of a model of a theory, which is not defended in any way
and which I find arbitrary and antiquated.
The persuasiveness of my counter-argument to the criticisms of scien-
tific realism just mentioned is clearly contingent on the intelligibility and
naturalness of the further restrictions imposed on the models of theories. I
shall return to this matter later in subsequent papers. In particular, I shall
try to show that the new restraints on models I am considering in this essay
are not arbitrary, but have in fact, been implicit in accepted mathematical
practice.
VIII
RESTRICTIONS AND NONSTANDARD MODELS
A closely related way of looking at the restrictions which I want to find

and impose on the models of first-order theories is not in terms of the con-
struction processes which take us from a theory to its models but in terms
of the structures which mathematicians are interested in. When a theory is
calculated to capture such a structure, the model (or models) exemplifying
this structure is called the standard model (or standard models). The others
are called nonstandard models. The descriptive incompleteness of a theory
means that it does not rule out all nonstandard models.
Hence the ultimate aim of the restrictions on (first-order) models can
be said to eliminate all nonstandard models of different mathematical
theories, formulated by means of first-order logic. At first sight, this
might seem to be very easy. The very fact that we can speak of the
intended or standard model of a certain first-order theory presupposes that
we can pick it out and distinguish from various nonstandard ones. What's
the difficulty here?
One difficulty is in seeing what the different ways have in common
that mathematicians use to rule out nonstandard models of different math-
ematical theories. In other words, the problem we face here is: What do
the different so-called standard models have in common? Is each of them
chosen ad hoc. or are there general principles which govern the exclusion
of nonstandard models? If all such choices are ad hoc. there is no hope
of developing a general method of ruling out nonstandard models. What,
if anything, is common to the standard models of, say, Peano arithmetic,
of the theory of the real line, or of the Hilbertian axioms of geometry?
Why is it so hard to decide which the standard (intended) models of
axiomatic set theory are? These are obviously important questions in any
approach to the philosophy of mathematics. It is a part of the project I am
proposing here to answer them and to argue that there are general prin-
ciples underlying the apparently different ways in which standard models
of diverse mathematical theories are distinguished from the nonstandard
ones. Basically, I propose to argue that extremality principles (the ideas
of paucity and plenitude) suffice for the task in certain mathematically
central cases. However, before they can do this, they will have to be
analyzed and developed essentially further, which is too large a task to
be undertaken here.
IX
WHY NOT SECOND-ORDER LOGIC?
But what's wrong with the most obvious way which logic gives us for
the purpose of separating the nonstandard goat from the standard sheep?
What is wrong with resorting to second-order logic and using its power to
capture the intended model or models? At the outset of modern philosophy
of mathematics, Frege used second-order concepts in defining natural
numbers. It is easy to turn Peano-type axiomatizations of elementary
arithmetic into complete ones by replacing the first-order axiom scheme
for mathematical induction
(6) (A[O] & (Vx)(A[x]::J A[x + 1]) ::J (Vx)A[x]
by the second-order sentence
(7) (VA)[(A[O] & (Vx)(A[x] ::J A[x + I])) ::J (Vx)A[x]].
It is also clear that by using second-order formulations we can obtain
descriptively complete axiomatizations of various important mathematical
structures.
However, this course is not satisfactory in all respects. For one thing,
the ascent to second-order logic helps us to reach descriptive completeness
only if second-order sentences are given what is usually called, following
Henkin, their standard interpretation. 16 (It is important to realize that this
represents a separate and prima facie unrelated use of the t:nns "standard
interpretation" and "standard modeL") What standard interpretation in
Henkin's sense means is essentially that higher-order variables are taken
to range over all extensionally possible entities of the appropriate type,
not just over entities representable in language. For instance, the universal
quantifier "(VA)" in (7) has to be taken to range over all sets (i.e., all
collections of elements in the domain), whether or not they can be captured
by anyone formula of the language in question.
Without this restriction to the standard interpretation of higher-order
variables, second-order logic does not help us to reach descriptive com-
pleteness. Without the standard interpretation, e.g., (7) will not necessarily
take us beyond (6).
Why cannot we be satisfied with the standard interpretation of high-
er-order variables? There are several different reasons, :mphasized by
different philosophers and logicians.
(i) The restriction to the standard model (in Henkin's sense) is an
abstract restriction unrelated to the process of model construction. It is very
hard to see what concrete consequences it can have for these processes.
(ii) The reliance on the totality of all extensionally possible entities of
a certain type seems especially dangerous when second-order procedures
like mathematical induction are used to define certain sets. This set, which
is only now being introduced, must already figure as one of the values of
the set variable by means of which it is being defined. In other words,

unlimited second-order conceptualizations are impredicative, and have
been rejected by several mathematicians, logicians, and philosophers for
this very reason.
(iii) Even if second-logic with the standard interpretation (in Henkin's
sense) is acceptable, it does not seem to help us with the problem of
uncovering a unifonn rationale for the different ways in which the stand-
ard models of different mathematical theories are picked out. The use of
second-order variables in the different cases does not appear to have any
uniform motivation.
Another way of putting the same point is to say that it is not clear
what, if anything, connects Henkin's notion of the standard interpretation
of higher-order logics with the idea that certain first-order mathematical
theories have designated or "standard" models. It is not clear what the two
notions of standardness are supposed to have in common, except the term.
There is a natural way of extending Henkin's standard-nonstandard
distinction for higher-order logics to modal logics and to the first-order
one. However, this extension does not seem to contribute anything to
the understanding of the nature of standard models needed in interesting
mathematical theories. 17
One of the aims of the project I am proposing here is thus to show
that the Henkin sense of standardness has very real connections with
the sense of standardness used in connection with models of particular
mathematical theories. The connection is mediated, not by a higher-order
formulation of mathematical theories, but by the extremality restrictions on
their (first-model) models.
A further objection to attempted second-order fonnulations of the
restrictions on models is that it is not obvious in all cases that they are
possible in the first place. For instance, in order to say that a certain
model is maximal you apparently have to quantify over entities outside
that model. But second-order fonnulations do not enable us to do so any
more than first-order ones.
X
IS MATHEMATICS REDUCIBLE TO LOGIC?
The problems I am discussing in this essay can also be related to the

old question whether mathematics is reducible to logic. Of course, this
question can be interpreted in a number of different ways. One of them
turns it virtually into a fonn of my own basic problem.
The reducibility of mathematics can be taken to mean that interesting
mathematical structures can be captured by means of purely logical consid-
erations. If these "purely logical considerations" are taken to be the usual
IS THERE COMPLETENESS IN MATHEMATICS AFrER GODEL? 79
sentence-model relation, mathematical theorizing is shown to be impos-

sible to reduce to logic by GOdel's theorems. However, this is a
relatively uninteresting sense of reducibility. A more interesting one
is obtained by asking whether there are further restraints on mod-
els (i) which would enable us to reach descriptive completeness
for important mathematical theories and (ii) which themselves tum
solely on general ideas. In this case, demonstrating the reducibility
thesis is essentially tantamount to finding restrictions on models
that are not tied to any particular mathematical subject matter.
This general relation of my project to the traditional reducibility ques-
tion is reflected by several detailed problems. For instance, Frege saw it as
a part of his task "to reduce the argument from ~ to (~ + 1) [Le.,
mathematical induction], which on the fact of it is peculiar to mathematics,
to the general laws of logic. "18 Now it can be shown that one of the main
types of inferences which extremality restraints on models serve to validate
is precisely the inference based on mathematical induction. Hence Frege's
reducibility problem becomes essentially a question concerning the status
of the extremality principles.
What is their status? At first sight it might seem that minimality and
maximality are so abstract general ideas that they must surely be consid-
ered "logical" rather than "mathematical." If this were the last word on
the subject, then a kind of reducibility of the crucial mathematical theories
to logic would be arguably true.
This is not the only way of looking upon the reducibility problem.
As was mentioned, an essential ingredient in the extremality conditions
is the requirement that the class of natural numbers be minimal (in a
sense that remains to be spelled out). It is the interplay of this requirement
with sundry collateral maximality requirements that lends such central
mathematical theories as the theory of the real numbers t.heir distinctive
character. And the minimality of the intended models of elementary
arithmetic can obviously be considered a defining feature of natural numb-
ers. Hence, far from affecting a reduction of mathematics to logic, my
extremalist program assigns the concept of natural number the pride of
place also in mathematical theories which prima facie deal primarily with
entities of a different kind altogether, such as theory of real numbers and
even set theory. In brief, the extremalist program amounts to a resounding
affirmation of the importance of natural numbers in math{:matics, via the
role of the minimality assumptions in different mathematical theories. In
this sense, it might even be said to vindicate the old idea of mathematics
as the science of numbers-primarily natural numbers.
XI
WHAT DOES A MATHEMATICIAN REALLY DO?
According to the picture of mathematical thinking which I am suggesting, a

mathematician does not set up an ordinary first-order axiom system and then
to proceed to derive theorems from the axioms by means of a finite, closed
list of inference rules. This picture, which goes back to the Greeks, is shown
to be insufficient by GOOel's results. I am not suggesting, either, that math-
ematicians in general argue in second-order terms, that is to say, in terms of
all extensionally possible subsets of a given set (and likewise for other types
of higher-order entities). Rather, what they do according to my model is to
set up certain meta-axioms, as it were, general restrictions on the models
of first-order theories. These restrictions are not axioms, because they
are not restrictions on particular models or models of particular theories.
They are characteristic of mathematical thinking for they are vital for the
purpose of capturing uniquely the structure or structures a mathematician
wants to study. These restrictions are reflected partly in such principles
as mathematical induction and the principles needed to characterize the
continuum (continuity assumptions). Such principles are therefore char-
acteristic of mathematical thinking, as distinguished from purely logical
reasoning, just as mathematicians like Poincare have claimed. 19
The fact that the operative restrictions usually make it impossible to
have a complete axiomatization of logic means that the idea of logical
proof (proving theorems) is much less central in mathematical thinking
than most philosophers realize. The idea of theorem-proving as the end-all
and be-all of mathematical thinking has to be given up. At the elementary
stages of the unfolding of a mathematical theory, theorems can admittedly
be proved by the familiar means of "general" logic. Beyond a certain
point, however, new principles (new modes of inference) have to be
introduced. They cannot always be seen to be valid. Hence they have
to be used in practice only experimentally, pending a fuller examination
of their consequences. Mathematics is creative precisely in the sense that
such tentatively accepted principles are in principle always needed. One
kind of mathematical genius consists in the gift of hitting upon such valid
new principles.
One upshot of my approach is thus to encourage what has been called
an experimental view to the foundations of mathematics, and even to
provide for the first time a solid theoretical foundation for it. This inter-
pretation of the foundations of mathematics has recently been discussed
by Penelope Maddy in an interesting way. 20 She finds illustrations of the
experimental attitude in the choice of different set-theoretical as sump-
tions. There are better illustrations, however, in the sense of being more
revealing of the theoretical motivation of the experimental attitude.
Hartley Rogers argued as early as mid-fifties for an experimental
view of mathematical assumptions by reference to the conceptual
situation in arithmetic. Unfortunately, he never published his argu-
ments. 21 I am including some excerpts as an appendix to this paper.
APPENDIX
Excerpt from Lecture Notes by Hartley Rogers, lr. (1956)

As we make successive extensions (of an axiomatized system ENT*
of elementary number theory), of possibly stronger kinds. whatever
"intuitive truth" our added statements possess is (in part) based on
empirical evidence and hence suffers a corresponding decay as we
go to further extensions. Indeed. at any stage beyond (the very
first extension step), a formula asserting the consistency of a new
extension can prove to be a factual falsehood without in any way
impugning the consistency of our metalanguage. The words "in
part" were inserted above for the following reasons. A person
attempting to extend ENT* is in the position of a natural scientist
attempting to widen a theory in order to describe and predict a wider
range of natural phenomena. In such activity, criteria of simplicity
and elegance as well as factual evidence appear to playa role. Thus
adopting the statement wei as true is partly a matter of empirical
evidence and partly a matter of symmetry and economy. (It can be
argued, of course, that all these considerations should be viewed
as one and the same.) Further discussion can thus lead directly
into problems of scientified induction that are a chief wncem of
philosophers of science.
Extensions of ENT* by an ideal statement such as "ENT* is
sound" suggests further analogy to the natural sciences in that it
is similar to the adoption of a scientific hypothesis which employs
deeper and less directly observable concepts. We do not have direct
empirical evidence; but rather we seek (support in) empirical evi-
dence for real consequences of the ideal statement. Criteria of
simplicity (aesthetic criteria, some would call them) play an even
bigger role in the case of such ideal statements. Adoption of new
ideal statements is envisioned by some commentators with words
such as "always possible to discover essentially new modes of
inference to broaden mathematics" (like e.g., the axiom of choice).
We take no issue with such comments except insofar as we feel it
desirable that the tome of "I don't know it now but I'll recognize
it when I see it" should be balanced by emphasis on the role of
empirical consequences.
In Hartley Rogers statement, ENT* is an axiomatization of elementary
number theory, we 1 is the number-theoretical statement that the axiomatic
theory ENT* is w-consistent.
I don't agree with Rogers' emphasis on matters of economy, sym-

metry, and aesthetics in the choice of mathematical assumptions, but
this is merely because I don't believe in their role in the methodology
of science, either. The important points here are that the search for
stronger and stronger mathematical assumptions is like the search for
stronger empirical theories in several respects: both are experimental
and always subject to the test of further hard evidence; both are open-
ended; and in both cases we are dealing with an objective reality "out
there." (Emphasis on this last point is of course more mine than Rogers'.)
NOTES
I. David Hilbert, Foundations of Geometry. translated by Leo Unger. tt:nth edition.

revised and enlarged by Paul Bemays (La Salle, Illinois: Open Court, 1971) Gt:nnan
original (first edition), 1899.
2. See here Jaakko Hintikka. "On tht: Dt:velopment of the Model-Theoretic Viewpoint
in Logical Theory," Synthese 77 (1988): 1-36.
3. Jean van Heijenoort, "Logic as Language and Logic as Calculus." Snlthese 17
(1967): 324-330; Warren Goldfarb. "Logic in the Twenties," Journal of Symbolic
Logic 44 (1979): 351-368.
4. For a survey, see J. Barwise and S. Fefennan (eds.), Model-Theoretic Logics. (New
York: Springer-Verlag, 1985).
5. See Kurt Gooel, "On Fonnally Undecidable Propositions of Principia Mathematica
and Related Systems," ed. Jean van Heijemoort, From Frege to Godel Source Book
in Mathematical Logic 1879-1931 (Cambridge: Harvard University Press, 1967)
592-616.
6. Kurt Gooel "The Completeness of the Axioms of the Functional Calculus of Logic".
van Heijenoort, 582-591 .
7. C. P. Snow, "Foreword" to G. H. Hardy, A Mathematician's Apology (Cambridge:
Cambridge University Press, 1967) 36.
8. These definitions go back to Alfred Tarski, "The Concept of Truth in Fonnalized
Languages," Logic, Semantics, Metamathematics (Oxford: Clarendon Press, 1952)
152-278.
9. How unruly the construction process may have to be is shown by the fact that
in nonstandard models of elementary arithmetic the basic arithmetical relations
are nonrecursive. See Solomon Fefennann, "Arithmetically Dt:finable Models of
Fonnalized Arithmetic," Notices of the American Mathematical Society 5 (\958):
679-680; Dana Scott, "On Constructing Models for Arithmetic," lnjinitistic Methods
(Oxford: Pergamon Press, 1959) 235-255; Solomon Fefennan, Dana Scott, and S.
Tennenbaum, "Models of Arithmetic via Function Rings," Notices of the American
Mathematical Society 6 (1959): 173.
10. See Jaakko Hintikka, "Fonn and Content in Quantification Theory," Acta Philosophi-
cal Fennica 8 (1955): 5-55.
11. See Jacques Herbrand, Logical Writings, ed. Warren D. Goldfarb (Cambridge:
Harvard University Press, 1971).
12. Stanislaw Ulam, "John von Neumann, 1903-1957," Bulletin of the American Math-
ematical Society 64 (1958, May Supplement): 1-49.
13. This title is of course a variant of the title of Gooe!' s famous paper, "Uber eine bisher
noch nicht benutzte Erweiterung des finiten Standpunktes," Dialectia 12 (1958):
280-287; English translation (with a bibliography) in Journal of Philosophical Logic
9 (1980): 133-142.
14. See any good introduction to model theory, e.g., C. C. Chang, and H. J. Keisler.
Model Theory (Amsterdam: North-Holland. 1973).
15. Hilary Putnam, "Models and Reality." Journal of Symbolic Logic 45 (1981): 464-
482. Actually. Putnam's arguments are subtler than what I have indkated. turning on
the implications of Skolem-Uiwenheim theorems rather than on those of incomplete-
ness. This additional sophistication does not change the overall picture, however.
16. Leon Henkin, "Completeness in the Theory of Types," Journal of Symbolic Logic
15 (1950): 81-91 (Cf. Peter Andrews, "General Models and Extensionality", ibid.
37 (1972): 395-397).
17. Jaakko Hintikka. "Standard vs. Nonstandard Logic." ed. Evandro Agazzi. Modern
Logic: A Survey (Dordrecht: D. Reidel. 1981) 283-296.
18. Gottlob Frege. The Foundations of Arithmetic. German text with a translation by J. L.
Austin (Oxford: Basil Blackwell, 1959) iv.
19. H. Poincare. Science and Hypothesis (New York: Dover. 1952) 1-16.
20. See Penelope Maddy. "Believing the Axioms I-II." Journal of Symbolic Logic 53
(1988): 481-511 and 736-764; "New Directions in the Philosophy of Mathematics."
PSA /984. ed. P. Kitcher (East Lansing. MI: Philosophy of Science Association.
1985) 425-447.
21. The excerpt is from Hartley Rogers' mimeographed lectures at MIT in 1956. Most of
these notes were published later as Hartley Rogers, Jr., Theory of Recursive Functions
and Effective Computability (New York: McGraw-Hili, 1967).
5
HILBERT VINDICATED?
Hilbert's philosophy of mathematics is almost universally labeled formal-

ism. I will argue in this paper that such a classification is highly misleading.
It does not do justice to the leading ideas of Hilbert's thinking about the
foundations of mathematics. His so-called fonnalism was the result of
several independent ideas most of which he could have maintained even
if he had given up his fonnalism. This holds, I will argue, of the most
interesting and most characteristic of his foundational ideas. Among such
leading ideas of Hilbert's that survive the elimination of fonnalism there
is his axiomatic ideal, which in fact was for him not a mere ideal, of math-
ematics and science. 1 There is also, it seems to me, another leading idea
which is much more elusive than the axiomatic idea. It will be the focus of
this paper. It might tentatively be called the idea of mathematical reasoning
as combinatorial reasoning.
A starting-point is obtained by asking: If fonnalism was not central to
Hilbert's philosophy of mathematics, how come that he has been classified
as one?
In a historical and critical perspective (in the sense of the good old
Gennan expression "historisch-kritisch") one can distinguish five different
features of his approach to the foundations of mathematics that have at
different times occasioned the application ofthe epithet "fonnalist" to him.
I will suggest that in reality none of them alone justifies the application of
the label "fonnalist" to Hilbert. These excuses for calling him a fonnalist
include the following:
(a) Hilbert's use of a purely logical and axiomatic approach in the foun-
dations of geometry as well as in the other parts of mathematics and
even SCIence.
(b) Hilbert's (less than fully articulated) early idea of a set-theoretical
universe as a "structure of all structures".
(c) Hilbert's strategy of proving the consistency of various mathematical
theories in a purely proof-theoretical manner, that is, by fonnalizing
the logic used in a given mathematical theory and then showing that
one cannot derive a contradiction from the axioms of the given theory
by means of the fonnal rules of the logic in question.
HILBERT VINDICATED? 85
One can consider (b)-{c) as further elaborations ot~ or add-ons to, the
basic axiomatic idea (a).
The fourth idea goes in a direction different from the axiomatic idea.
(d) Hilbert's preference of combinatorial over set-theoretical reasoning in

the foundations of mathematics.
As a fifth assumption one can mention
(e) Hilbert's finitism.
Once again this idea was perhaps not completely articulated by Hilbert.
In this paper, I will concentrate mainly on the fourth alleged reason (d)
for calling Hilbert a formalist. One reason for doing so is that it has not
received adequate attention in the literature. A further reason is that it is
related to certain ongoing developments in the foundations of mathematics.
Another reason is that this idea is the one which in conjunction with others,
notable (e), actually led Hilbert to a kind of formalist position.
Of the other alleged reasons for calling Hilbert a formalist, (a) does
not cut much ice. 2 What is involved in it is that according to Hilbert the
derivation from the axioms of a mathematical system is a purely logical
manner. Hence it is independent of the meanings of the nonlogical constants
of the mathematical system in question. For instance, as flif as the derivation
of theorems from axioms is concerned, we might as well speak of tables,
chairs and beermugs instead of points, lines and circles, as Hilbert put it
in his provocative way.3 Of course, in an actual logical deduction we are
likely to replace nonlogical terms by variables ofthe appropriate type. This
feature of Hilbert's approach is what prompted Russell (1937, vi) to call
him a formalist and to object to his views.
However, such a conception of a purely logical axiom system does not
imply formalism. Aristotle and Euclid used variables in their deductions
of theorems from axioms without being formalists. After Hilbert's time,
the resurrected idea of a purely logical ~xiom system has become a vir-
tual commonplace. The reason why it prompted the charge of formalism
is its short-term novelty. The idea that mathematical arguments are pure-
ly logical had been abandoned in the nineteenth century both by many
mathematicians and by many philosophers.
In reality Hilbert's early conception of an axiom syst1em was a squarely
model-theoretical one, not a proof-theoretical one. From a systematical
vantage point, it does not even begin to justify calling him a formalist.
The second alleged reason (b) for calling Hilbert a formalist pertains
to a view which he apparently abandoned himself later. It, too, is squarely
model-theoretical or at least set-theoretical. It amounts to considering, in

the same way as Cantor apparently did,4 the set-theoretical universe as the
collection of all possible structures (models). 5 Hence, as soon as an axiom
system is consistent, there exists a model for it in that universe. It is in this
way, and for this reason, that consistency implies existence. Once again,
Hilbert's ideas are model-theoretical rather than formalistic in any strict
sense. If there is something unusual about Hilbert's view here, they are his
views of set theory and the set-theoretical universe, not about mathematical
existence or mathematical truth.
The third allegedly formalistic ingredient (c) in Hilbert's thinking was
originally little more than a ploy to prove consistency of mathematical
axiom systems and thereby the existence of models for them. The idea was
to formalize the logical reasoning used in deriving theorems from axioms
in some mathematical axiom system and then prove by considering such
formalized proofs that by their means one cannot derive an inconsistency
from the axioms. If the formalized logic is complete, then from such
a formal consistency one can infer consistency in the model-theoretical
sense of satisfiability, that is, the existence of models of the axiom system
in question. This is in a nutshell the famous Hilbertian project in the
foundations of mathematics. 6
This is again a poor excuse for calling Hilbert a formalist in his phi-
losophy of mathematics. Hilbert's original ambition is squarely a model-
theoretical one, viz., to prove the existence of models for various axiom
systems. The means he used in this enterprise may include a formalistic
element, but this element pertains to Hilbert's philosophy of logic rather
than to his philosophy of mathematics. Hilbert's grand strategy presup-
poses the completeness (complete formalizability) of the logic we use in
mathematics. It may justify calling Hilbert's conception of logic formal-
istic, but not his conception of mathematics (mathematical theorizing).
Understood from the vantage point of Hilbert's original motivation, the
ends of his famous foundational project are not formalistic.
In the light of such observations, Georg Kreisel was amply justified
in saying that "there is no evidence in Hilbert's writings of the kind of
formalist view suggested by Brouwer when he called Hilbert's approach
'formalism' ".7 This point can be spelled out further by pointing out that
the consistency proofs of the kind Hilbert envisaged had nothing to do with
the epistemology of mathematics in the sense of concerning the truth, fal-
sity, probability or reliability of any particular mathematical (or scientific)
axioms.
In other words, for Hilbert the question of the consistency of an axiom
system was completely different from the question of the truth or falsity
(or reliability or certainty) of its different axioms in their application to

real world. The truth or falsity of anyone axiom had to be established
empirically. The truth of Euclid's fifth postulate in the actual universe had
to be tested by such means as are used to test Einstein's general theory
of relativity (which of course was in historical reality as much Hilbert's
theory as it was Einstein's). The same need of empirical verification goes,
Hilbert points out explicitly, even for the axioms of continuity.8 All this
illustrates the robust model-theoretical nature of Hilbert's conception of
an axiomatic theory.
A consistency proof for a mathematical axiom system was not thought
of by Hilbert as pertaining to the question whether its axioms are true or
not, but to the question whether the axiom system has a subject matter in
the first place which could make it true or false.
However, the last two ingredients (d)-{e) of Hilbert's views together
pushed him to a kind of formalist position, even though neither of them
does so alone. Once again, the point is about mathematical thinking and
reasoning rather than about the contents of mathematical theories. The
thesis (d) says that mathematical reasoning can be thought of as pertaining
to the ways particular concrete individuals can be combined together into
structures of different kinds. If these structures are finite, they can be
thought of as being exemplified, not only by tables, chairs and beermugs,
but more relevantly by formal symbols on paper or on a computer screen.
I will discuss this motivation of Hilbert's formalism in the course of this
paper. Clearly, neither (d) or (e) alone suffices to lead anyone to a genuinely
formalist position, even though in combination they may be thought of as
doing so.
A contributing factor to the confusions surrounding Hilbert's alleged
formalism is that his most important statement of his formalism is relatively
little known. It appears not to have been translated into English. This
locus classicus of Hilbert's formalism deserves - and ne:eds- a number of
comments. The crucial passage is found in Hilbert (1922,161-163). Hilbert
is considering there the problem of consistency proofs as a paradigm
problem in the foundations of mathematics. An abridged translation might
run as follows:
The importance ('If our problem of the consistency of axioms is admittedly acknowledged
by philosophers; but I do not find anywhere the philosophical literature, either, any clear
requirement that the problem needs a solution in the mathematical sense. In contrast, our
question is essentially affected by the older attempts to find a basis for number theory and
analysis in set theory and also a basis of all theories in pure logic.
Frege has attempted to build number theory on pure logic and Dedekind on set theory
construed as a chapter of pure logic; but neither has reached his aim. Frege did not handle
carefully enough the usual concept formations of logic in their application to mathematics;
for instance, he considered the extension of a concept to be given directly (ohne weiteres),
in such a way that he believed that these extensions could without restriction be taken
themselves to be objects. He was so to speak a victim of I!xtreme conceptual realism.
Dedekind met with a similar fate. His classical mistake was to take as his starting-point the
system of all objects ....
This is in my view the present state of the question as far as the foundations of mathe-
matics are concerned. Accordingly studies of the foundations can only reach a satisfactory
outcome through a solution of the problem of the consistency of the axioms of analysis. If
we succeed in showing this consistency, then we can conclude that mathematical proposi-
tions are in fact indubitable and final truths - a result which is of the greatest significance
to us also because of its general philosophical character. We tum now to the solution of this
problem.
As we saw, abstract operating with the extensions and contents of general concepts
has turned out to be insufficient and dangerous. As a precondition of the application of
logical inferences and the implementation of logical operations, something must already
be represented [to the mind], viz., certain nonlogical discrete objects which are present in
all thinking intuitively and in immediate experience. If logical inference is to be reliable,
it must be possible to review them and all their parts completely. and their structure, their
individuation and their order are given to us together with the objects themselves as some-
thing that cannot be reduced to anything else. In that I adopt this point of view, the objects
of number theory are - in a precise opposition to Frege and Dedekind - [number] symbols
themselves. Their fonn can be recognized universally and with certainty independently of
location and time, independently of how the symbols are produced and independently of
minor variations in the execution of this production. Here is the philosophical attitude that I
consider indispensable for the foundations of pure mathematics - as well as for all scientific
thought, understanding and communication: In the beginning - so to speak - there was the
symbol.
Certain comments are in order here.
(1) One of the most remarkable features of these statements of Hilbert's is

that there is in it no mention of formalism in the strict sense, that is, of the
requirement that the signs must not mean anything. This absent thesis, the
thesis of formalism, comes in only through the fact that symbols happen
to be handy "nonlogical discrete objects which are present in all thinking
intuitively and in immediate experience".
(2) Hilbert's comments in the quoted passage have to be taken in the con-
text of the rest of his ideas and in the context of his development. They also
illustrate the first four excuses for calling Hilbert a formalist mentioned
above. Hilbert's first and foremost emphasis in the foundations of mathe-
matics (and in the foundations of science in general) is on the axiomatic
method. His second and almost as important emphasis is on the purely
logical character of satisfactory axiom systems. Furthermore, Hilbert rep-
resented (as was seen earlier) the idea that purely logical inferences must
be characterizable and recognizable purely formally, on the basis of the
symbols alone that are being used. Hence Hilbert's tacit "syllogism" runs
somewhat as follows: The basic clarified form of mathematical theoriz-
ing is a purely logical axiom system. All logical reasoning can be carried
out purely formally. Hence in a logically clarified mathematical theory
all reasoning can be carried out purely formally, as if it involved merely
manipulating certain symbols. So understood, Hilbert is not saying that
mathematical symbols do not mean anything. All that he is saying is that
in order to be on the safe side in one's logical inferences in mathematics,
one must be able to follow the rules of logical inferences blindly, as if the
symbols had no meaning. As Poincare saw, this idea is present as early
as in the Grundlagen der Geometrie. As far as Hilbert's axiom system is
concerned, "we might put the axioms into a reasoning apparatus like the
logical machine of Stanley levons, and see all geometry come out of it"
(P(,incare 1902, 150 of the English translation).
Many ofthe statements Hilbert makes in the quoted passage can natural-
ly be understood as merely reiterating these points. Otherwise the explicit
generality of his statements would be incomprehensible. How can "all
scientific thought and communication" be purely formalistic? A purely
formalistic philosophy of mathematics is possible because it is actual, if
not in the writings of Hilbert, then at least in the writings of someone
like Haskell B. Curry (e.g., Curry 1954). But a formalistic philosophy of
physics does not make any sense.
By far the most natural way of taking Hilbert's words, at least their
main thrust, is to take them to take off from the formal character of all
valid logical inference. In other words, Hilbert is in the quoted passage
expressing the idea that all logical reasoning can be carried out purely
formally. This point, whether or not it is all that Hilbert is here saying, has
nothing to do with a formalistic theory of mathematics, as little as it has
to do with a formalistic theory of physics. Thus Hilbert may be committed
to a formalistic philosophy of logic, but scarcely to anything that can
plausibly be called a formalistic philosophy of mathematics. Moreover,
the same idea of a purely formal character of logical laws was shared
with Hilbert by the likes of Frege, Wittgenstein and Carnap. One does not
even have to attribute a major historical inaccuracy to Hilbert, as might
first seem to be necessary because of the contrast he sets up between
his ideas and those of Frege. For even though both maintained a purely
formal character of the actual logical laws (e.g., the rules of inference used
to derive theorems from axioms), they reached this view from opposite
directions, Frege from a belief in the universality of his Begriffsschrift but
Hilbert from an essentially model-theoretical point of view.
(3) This suggestion can be strengthened by considering the immediate

context of Hilbert's utterances. Hilbert is in the quoted passage obviously
thinking primarily in terms of the problem of proving the consistency of
arithmetic and analysis. Hence his statement of his "formalistic" program
in the quoted passage is merely a description of the strategy of his attempted
consistency proofs, not a philosophical manifesto. In brief, the abstraction
from the meanings of symbols and the exclusive concentration on their
formal properties is (in part at least) merely one possible way of carrying
out a consistency proof
The same point can be placed in a wider context. Hilbert's so-called for-
malism was.not calculated to eliminate nonconstructive existence proofs,
but to vindicate them.
What philosophers have failed to see is that Hilbert's so-called formal-
ism was like Russell's "reduction to acquaintance". He did not deny the role
of non formal reasoning in mathematics or its validity. He wanted to show
the acceptability of such reasoning by means of formalistic reasoning.
This is not the whole story, however.
(4) Hilbert's remarks are addressed in so many words to Frege's and

Dedekind's attempts to base arithmetic and set theory on "reine Logik"
and even reduce them to logic. He is hence talking about the basic logic
that is needed in the foundations of mathematics, not about what is going
on in real working mathematics. Insofar as Hilbert really represents a
formalistic position, he is a formalist in the philosophy of logic. not in the
philosophy of mathematics. In the quoted passage Hilbert is not offering
formalism of any sort as an alternative to the usual reduction of the more
advanced parts of mathematics like analysis to arithmetic andlor set theory.
He is talking about the reduction of arithmetic and set theory to logic.
(5) The contrast which Hilbert sets up between his own approach and those
of Frege and Dedekind is clear. According to Hilbert, what was insuffi-
cient and insecure in their procedure was that they operated with abstract
concepts, either with their extensions (Umflinge) or with their intensions
(Inhalte). Hilbert wants to resort instead to operating with directly given
discrete nonlogical objects.
As an example of such concrete, given objects Hilbert chooses cer-
tain symbols. The reason for this choice is their concrete givenness and
manipulability. For these properties, the question as to whether the sym-
bols in question represent something or not is totally immaterial. Symbols
are not drafted to Hilbert's service because they are purely formal, i.e.,
nonrepresentational. Rather, they are preferred by Hilbert because they are
intuitively given to us in direct experience and hence admit of a complete

overvIew.
Thus Hilbert's emphasis on the role of "nonlogical discrete objects" in
logic and the foundations of mathematics is directed squarely against the
likes ofFrege and Dedekind with their use of abstract concepts. Hilbert was
favoring first-order reasoning over set theory or higher-order logic, albeit
not necessarily in the form of what now goes by the name of first-order
logic. Hence the term "formalism" is once again seriously misleading when
applied to Hilbert's views. Even though he does not use the term himself
and even though it is subject to misunderstanding, it seems to me that the
term combinatorial is a much happier one than "formalistic".
The same point can be put differently. It is clear that Hilbert viewed
analysis, set theory, and parts of logic as involving "ideal elements"
which are not self-explanatory and whose role in mathematics ought to
be explained by reference to concrete combinatorial operations. As was
mentioned above, it is for instance very interesting to see that all the mis-
takes Hilbert attributes to Frege and Dedekind concern set-theoretical or
higher-order matters, not their views on ordinary first-order logic.
(6) As an epistemological program, Hilbert's suggestions may be com-

pared with the program of such fellow reductionists as Russell or Husserl.
Hilbert is envisaging a kind of reduction of all mathematical reasoning to
something like Russellian acquaintance. The crucial question in Hilbert
is when the kind of reduction we are dealing with cannot be continued
any further. His view is that we have reached a rock bottom with objects
which are directly given of us and which are under our control. We have
to realize that the Hilbertian reduction is merely a structural one. He is
dealing with logical inferences that are independent of the subject matter
to which it is applied, be it points, lines and circles or tables, chairs and
beermugs. Therefore Hilbert does not need acquaintanct~ with any partic-
ular objects. All that is needed are some objects that are so to speak clear
and distinct. Symbols are Hilbert's candidates for this role merely because
they are directly given to us and completely in our control. Their being
merely formal in the sense of not having a meaning is irrdevant to Hilbert.
In Russell, the inverse of his famous reduction to acquaintance, that is,
the logical construction of objects of description out of objects of acquain-
tance, was calculated to replace inferences to the existence of the objects
of description and hence to vindicate objects of description. In Hilbert,
too, showing how higher-order reasoning can be reduced to combinatorial
reasoning was an essential step in proving that such reasoning cannot lead
to contradictions and thereby in vindicating higher-order reasoning.
(7) What does Hilbert want to say about logical reasoning? A careful
scrutiny of Hilbert's statements fails to reveal any doubt of the actual
validity of our ordinary interpreted principles of logical reasoning, which
according to Hilbert apparently include the axiom of choice. He is not
trying to revise our logic or to replace interpreted logic by something else.
He is trying to justify it by showing its combinatorial character.
In fact, Hilbert spells out almost in so many words that he is not
trying to replace the ordinary methods of logical inference by anything
less but to formulate the rules of inference explicitly ("formally") and to
understand them better. Admittedly, he was fully aware of the view that
"mathematicians [have] paid little attention to the validity of their deductive
methods" and that this neglect is what led to set-theoretical paradoxes. Yet
he firmly believed that there is an unproblematic core of sound deductive
principles.
Does material [or, perhaps more accurately translated, interpreted (inhaltlich)) logical
deduction somehow deceive us or leave us at lurch when we apply it to real things or
events? No! Material logical deduction is indispensable. (Hilbert 1926, 191 of the English
translation)
Moreover, Hilbert believed that an essentially combinatorially interpreted

deductive logic can be adequately captured by the usual formalization of
logic .
. .. we find logical calculus already worked out in advance .... We possess in the logical
calculus a symbolic language which can transform mathematical statements into formulas
and express logical deductions by means of formal procedures .... Material [interpreted]
deduction is thus replaced by a formal procedure governed by rules. The rigorous transition
from a naive to a formal treatment is affected, therefore, both for the axioms. .. and for
the logical calculus (which was originally supposed to be merely a different language).
Hilbert's strategy makes no sense unless it is assumed that the interpreted

logical deductions he proposes to formalize capture all the valid methods of
purely logical deduction. (See Hilbert 1926, 197 of the English translation.)
Another relevant observation here is that Hilbert cared a lot of the actual
correctness of interpreted principles of reasoning. The prime example is
the axiom of choice. But if so, he could not have wanted to restrict our
actual reasoning to finitistic formal reasoning. Such reasoning was for him
merely a ploy for consistency proofs. Once again, a Hilbertian reduction
is calculated to vindicate what is being reduced, not to eliminate it.
(8) This line of thought can be continued. Taken strictly literally, Hilbert's
formulation allows for the possibility that the symbols he is considering
have a meaning and that the concrete manipulations we perform on them
likewise has an intuitive content, e.g., that they can be interpreted as log-
ical inferences. The only condition is that we are dealing with concrete
operations on given discrete objects.
In particular, Hilbert's application of his "formalistic" ideas to number
theory, with number symbols playing the role of the concrete objects Hilbert
wants to focus on, is not the only possible way of putting his combinatorial
approach to work. Furthermore, it is not a precondition of this application
that we forget about the meaning of numerical symbols. Hilbert's emphasis
is merely of actual operations on concrete given objects. There could in
principle be other applications of the same basic ideas.
It is important to realize that this is not merely an abstract possibility
of interpreting Hilbert's words. Interesting and important developments in
the philosophy of logic turn out to exemplify my interpretation of Hilbert's
statement, and large segments of his own foundational work turns out to
be in keeping with the interpretation.
(9) As an example of conceptualizations falling under Hilbert's formu-

lation, let us consider the Hintikka-Beth perspective on first-order proofs
(see Hintikka 1955; Beth 1955). An attempted proof of say, an implica-
tion from F to G, is on this interpretation a step-by-step procedure aiming
to construct a description of a model in which F is true but G false. If
such an attempt is inevitably frustrated in all directions, a countermodel is
impossible, and G is therefore implied logically by F.
Hence this procedure has a clear interpretation. At the: same time, it can
be carried out as a concrete operation whose rules refer only to the signs
(symbols) involved. Not only can the rules which govern the countermodel
(tableau, model set) construction be formulated purely formally. If the
countermodel construction does succeed, one can use a device originating
from Henkin's (1949) completeness proof and use the set of formulas
(configurations of symbols) which results from the constructions as its
own model. Hence the entire procedure can be thought of as an operation
on concrete symbols from beginning to end.
Even if some countermodels take an infinite number of steps to complete
in this way, every step and every initial segment of the process is finite,
and governed by purely formal operating rules. This includes all ordinary
first-order proofs, for they correspond to attempted constructions that reach
a dead end after a finite number of steps in all directions.
Hence Hilbert might as well have considered the entire first-order as a
safe and unproblematic part of logic. It has only individual variables and
quantifiers ranging over individuals, but no variables ranging over concepts
or their extensions. First-order logic is not a logic of abstract general
concepts. It can be viewed as a logic which turns on instantiation rules, that

is, rules governing the choice of concrete objects calculated of instantiate
general concepts (usually complex ones). We may perhaps formulate this
(feasible) view by saying that first-order logic is a combinatorial rather than
conceptual discipline within logic. even though the term "combinatorial"
has to be taken here with a grain of salt.
The only thing that could have prevented Hilbert from adopting this
way oflooking upon first-order logic is his finitism (e). For the models (or
countermodels) constructed by means of the tableau rules or by means of
the rules of the tree method are not always completed in a finite number of
steps. Such models can perhaps be called combinatorial, but they are not
finitistic.
(10) If you want to see a small scale example of Hilbert's thinking, the
usual naive rule of existential instantiation and its philosophical vindication
from the Beth-Hintikka viewpoint will fill the bill. Suppose that you move
from an existential formula (:Jx)S[x] to a substitution instance of S[x],
say S[a], in virtue of the rule of existential instantiation. Suppose further
that your first-order language in question is an interpreted one, so that
individual constants actually refer to certain objects. But the "dummy
name" a introduced in existential instantiation cannot do so. A logical rule
does not provide an actual example of the entities that satisfy S[x]. The
symbol a plays in a sense a purely formal role in the rest of the logical
argument. Here, then, is a familiar example of how the manipulation of a
"purely formal" symbol can serve the purposes of logical reasoning, just
as Hilbert envisaged.
Admittedly, even the "dummy names" of existential instantiation have
an interpretation in a wider sense of the word, not as standing for some
mythical "arbitrary objects" but as ingredients of an attempted experi-
mental countermodel construction a la Beth. But this interpretation only
strengthens Hilbert's hand, for the only model in the literal sense of the
word that is being constructed is a set of formulas. By an interesting switch
of perspective (first employed by Henkin), the formulas of this set are re-
interpreted so as to speak of the purely formal entity that this set itself is.
Hence this interpretation does not diminish the value of existential instan-
tiation as an illustration of Hilbert's point about the uses of symbols qua
symbols in reasoning.
(11) A perceptive reader will have noticed the similarity, indeed virtual
identity, of my way of viewing first-order logic and Kant's conception of
the characteristically mathematical method (see here Hintikka 1973). As
I have shown, Kant considered it to be a characteristic of the method of

mathematicians that they consider their general concepts through particular
representatives of these concepts. Kant referred to this as a method of
constructing those concepts. This is closely similar to Hilbert's idea to base
the logic that we need in mathematics on operation on concrete particulars.
There is an important difference, however, between Kant and Hilbert.
Hilbert's attempt to find precedents to his views in Kant may perhaps have
helped to hide the fact that he was dealing here with the interpretation of
logical thinking rather than mathematical one, as Kant was. But this con-
fusion is partly terminological only. It does not affect the insight we have
reached in the nature of first-order logic as satisfying Hilbert's require-
ments on the kind of reasoning that can be used in a Neubegriindung der
Mathematik. Perhaps we should paraphrase or perhaps parody the Genesis
once again: In the beginning there was first-order logic. Or, in less poetic
terms part of Hilbert's formalism is merely a preference of first-order logic
over higher-order logics and over set theory in the foundations of math-
ematics. In the passage quoted earlier, Hilbert even lists three misleading
ideas. They involve: (i) Begriffsrealismus, that is, reliance on concepts; (ii)
reification of extensions of concepts into objects; (iii) the idea of a System
der allen Dinge, i.e., the idea of the totality of all (possible) objects. It
is instructive to see that all these dubious ideas go beyond the realm of
first-order logic.
(12) The perspective reached by this line of thought seems to accord

well with what Hilbert was trying to do. Much of his concrete efforts in
logic consisted in attempts to extend the unproblematic character of first-
order reasoning to other kinds of reasoning used in mathematics, especially
set-theoretical reasoning.
In these efforts, an important role was played by Hilbert's way oflooking
at first-order reasoning. As Warren Goldfarb (1979) has pointed out, Hilbert
saw the nature of quantifiers as tacitly embodying certain choice functions
(Skolem functions). The function of quantifiers is thus closely related to the
principle of choice. The idea of choice is indeed closely tied to Hilbert's
vision of a combinatorial basis of mathematics. 9 The trouble with the
axiom of choice is that it either involves the introduction of a new concept
(the choice function) or else an infinite number of choices. Hilbert tried
to solve these problems by first in effect postulating a universal choice
function in the form of his epsilon-terms, and then trying to vindicate this
procedure by showing that only a finite number of choices are involved in
anyone logical argument.
A separate examination is required to answer the question whether

Hilbert succeeded in his attempt - whether, for instance, he realized his
ambition of showing that in suitable formulation the axiom of choice is as
undeniably valid as 2 + 2 =4. 10 For the present purposes, it suffices to have
shown that Hilbert himself construed his "formalism" in the same spirit as
we. It would not even be to far fetched to suggest that one of the aims of
Hilbert's famous program was to show the combinatorial character of all
mathematics.
(13) The comparison with Kant mentioned briefly is important enough

historically to be deepened further. In a sense, Hilbert is reversing the
Kantian view of the relation of mathematics to logic. For Kant, logic
deals with general concepts while mathematics uses particular concrete
representatives of general concepts, i.e., instantiation rules. For Hilbert, it
is logic that has to be based on concrete individuals instantiating certain
complex concepts codified by open formulas. In contrast, mathematics
involves certain general concepts to be analyzed in terms of logic and
set theory, such as limit, continuity, differentiability, integral, etc. This
inversion of Kant's conception of mathematical reasoning and its relation
to logic makes doubly ironic the attribution of the label "formalist" to
Hilbert. For Kant, the kind of use particular concrete representatives of
general concepts that Hilbert emphasizes in his combinatorial view of
logic is precisely what he means by the use of intuitions. Hence, in the
historically accurate Kantian terminology, Hilbert was asserting that the
basis of logic and mathematics is the use of intuitions, which Kant even
defines as particular representatives of general concepts.
(14) Suggestions have in fact been made in the literature to interpret

Hilbert's program in roughly the same spirit as I am doing, that is, as
emphasizing the combinatorial character of the logical basis of mathemat-
ics outlined in the preceding section. For instance, Georg Kreisel (1958,
210 of the reprint) notes, speaking of Hilbert's basic assumptions and
noting the failure of Hilbert's original finitistic methods to do their job,
that
. " instead of having a single kind of elementary reasoning whereby we understand the
use of transfinite symbols, there will now be methods of reasoning involving a hierarchy of
conceptions such as, e.g., more and more abstract conceptions of "construction", and we
have a hierarchy of Hilbert programmes of discovering the appropriate complex of such
methods which is needed for understanding the use of transfinite symbols in given systems
(modified Hilbert programme). [Emphasis Kreisel's]
In this paper I am mainly interested in a couple of members of the kind of

hierarchy of methods Kreisel envisages. I am interpreting "the use of trans-
finite symbols" to amount to an essential use of set-theoretical or higher-

order concepts, and the elementary methods Kreisel mentions as first-order
methods, which could also be called combinatorial (in a sufficiently wide
sense). The initial passage quoted from Hilbert and my comments on it will
show you how close this way of looking at Hilbert's words is to their lit-
eral meaning. The main remaining difference between Hilbert and myself
would then be that he restricted his attention to finite combinatorics while
I am also considering infinitary combinatorial methods.
(15) In sum, it is Hilbert's emphasis on the concrete operations on sym-

bols as the true foundation of logic that comes closest to justifying the
application of the term "formalist" to him. And, in conjunction with his
belief (e) in the finitary character of rock-bottom, fully analyzed mathe-
matical reasoning, it does lead him to a form of formalism. But even so,
this label can be misleading. It should be understood by reference to his
historical background in mathematics and in the light of his own work on
the foundations of mathematics. In particular, these sources of information
and insight should be kept in mind in examining and evaluating Hilbert's
program in the foundations of mathematics.
(16) But is Hilbert's vision of the combinatorial character of the true

logical basis of mathematics true, or is it a mirage? Kreisel (1958, 213 of
the reprint) has said that even after Godel's results,
When asked "What is mathematics about?" Hilbert could still have said: about arithmetico-
combinatorial facts of finitist mathematics....
Unfortunately,
Hilbert's answer is simply not true even for the very weak sense of "equivalence of content"
expressed in statements offormal deducibility and nondeducibility (Ioc. cit.).
My answer to the question "What is mathematics all about?" is a vari-

ant of Hilbert's, to wit, "about arithmetico-combinatorial facts, whether
finitistic or not". Since the relevant sense of "arithmetico-combinatorial"
here comes closely to "first-order", the crucial question becomes: Can all
mathematical reasoning be understood as first-order reasoning?
It might at first seem that there is no hope whatsoever of vindicating
the idea that logically speaking one could reconstruct all mathematical
reasoning on the first-order level. Most of the characteristically mathe-
matical concepts and modes of reasoning cannot be expressed by means
of ordinary first-order logic, including mathematical induction, infinity,
equicardinality, power set, well-ordering, etc. The most commonly used
conceptual framework for mathematics is set theory, whose very name cod-
ifies the idea of mathematical reasoning as dealing with such universals as
sets (and possibly concepts) in contradiction to such concrete particulars

as Hilbert put his faith on.
Unlikely or not, recent work by myself and my associates throws some
sharp light on this question and (with certain not unimportant qualifica-
tions) even vindicates Hilbert's vision. The main qualification is that we
cannot restrict ourselves, pace Hilbert, to finite combinatorics but must
indulge in infinitary combinatorics.
But even so, the shortcomings of the ordinary first-order logic are so
conspicuous that any attempt to think of it as a viable framework of all
mathematics is bound to seem quixotic. It is here that we come to the
decisive new insight that eluded Hilbert, as it has eluded other recent
foundationalists. The ordinary codification of first-order logic is not an
adequate logic of first-order (combinatorial) reasoning. To put the point
in different tenns when Frege and Russell fonnulated what is now called
first-order logic (quantification theory), they made a serious mistake. They
fonnulated their logic too narrowly. When these unnecessary limitations
are removed, we obtain a stronger logic. It might be called independence-
friendly (IF) first-order logic. I I Unlike ordinary first-order logic, it can
serve as a philosophically satisfactory framework of all mathematics. All
nonnal mathematics can in principle be thought of as being conducted
within the framework of IF first-order logic. Needless to say, this is neither
heuristically nor expositionally the natural procedure.)
This claim prompts two questions: (i) What is IF first-order logic? (ii)
In what sense can it serve as a framework of all mathematics? Let us take
these two questions in order. Since I will be dealing with this point in some
more detail elsewhere, I can perhaps be rather brief here.
(i) Hilbert's very own emphasis on quantifiers as codifying choice func-

tions provides a clue to the idea of infonnational independence. A choice
is always make on the basis of certain given infonnation, that is, depend-
ing on certain given parameters. What are these parameters in the case of
first-order logic? Consider, as an example, the sentence
(1) (\fx)(:3y)(\fz)(:3u)S[x, y, z, u]
Here the choice of the value of y depends on the value of x, and the choice
of the value of u depends on the values of both x and z. This is reflected
by the second-order Skolem-function translation of (1),12 which is
(2) (:3f)(:3g)(\fx)(\fz)S[x, f(x), z, g(x, z)]

Here the order of the different quantifiers in (1) is reflected in (2) by the
fact that the function f which replaces y has x as its sole argument whereas
the function g replacing u has both x and z as its arguments. But from
the viewpoint of quantifiers as embodiments of choice functions, it would
make perfect sense to require instead that u depend only on z, i.e., that the
function replacing u have z as its only argument. Then (2) changes into
(3) (3f)(3g)(Vx)(Vz)S[x. t(x), z, g(z)]
But what will now be the first-order counterpart to (3) (like (1) to (2? A
moment's thought shows that the obvious intuitive import of (3) cannot be
captured by means of any linear sequence of quantifiers. One thing that we
can do is to use branching quantifiers:
(Vx)(3y) }
(4) (Vz)(3u) S[x, y. z. u]
Instead, and much more systematically, we can introduce a special slash

notation (3u jVx) to exempt (3u) from the scope of ('Ix). Then (3) and (4)
are equivalent to
(5) (Vx)(3y)(Vz)(3ujVx)S[x, y, z, u]
as well as equivalent to
(6) (Vx)(Vz)(3yjVz)(3ujVx)S[x, y, z, u]
When this idea (and this notation) is used systematically, we obtain what
I have called independence-friendly (IF) first-order logic.
An important additional explanation must be added, however. In If
logic, dependent and independent disjunctions have to be dealt with in
the same way, mutatis mutandis, as dependent and independent existential
quantifiers. For instance, we can have sentences like
(7) (Vx)(3y)(Vz)(SJ [x, y, z](V jVX)S2[X, y, z])
which is equivalent with
(8) (3j)(3g)(Vx)(Vz)((SJ [x, j(x), z] & g(z) =0) V

(S2[X, j(x), z] & g(z) :f: 0))
but not with any ordinary first-order sentence.
Such independent disjunctions are an integral part ofIF first-order logic.
(ii) But in what sense can IF first-order logic serve as a framework for
"all" mathematics? An answer to this question can be given in two steps.
(a) First, it can be shown that every EI second-order sentence can be

translated into an IF first-order language. This result has been known
for a while. It goes back at least to Walkoe (1970). (b) Second, there is
a sense in which each question concerning the theoremhood of a normal
mathematical proposition is equivalent to a question concerning the validity
I
of a E second-order sentence. This reduction holds for any mathematical
theory (finite axiom system) T that can be formulated in the theory of finite
types.
What we can do then is to reconstruct the requisite part of the theory
of types as a many-sorted first-order theory, in an obvious way. Then the
only thing that is not captured by doing so is the requirement that for each
class of lower-order entities there must exist a higher-order entity having
them, and only them, as its members. This can be done by means of a finite
number of second-order sentences of the form
(9) (V'X)S[X]
where S[X] is a first-order formula. In this way T is transformed into a III

second-order sentence T* .
Then the question whether a sentence C is a theorem ofT becomes the
question whether a second-order sentence of the form
(10) (T*:J C*)
is logically true (valid), where C* is the translation of C into the language

of the many-sorted first-order theory. But (8) is of the EI form, and hence
translatable into an IF first-order language. Hence the question whether
C is a theorem of T equals the question whether a certain sentence of IF
first-order logic is valid.
Questions of this kind are not finitary, for IF first-order logic is not
axiomatizable, but they are in an obvious sense combinatorial. They con-
cern the possibility or impossibility of certain relational structures of indi-
viduals. They do not involve any quantification over sets or any other
higher-order entities. Hence this reduction shows a sense in which practi-
cally all mathematical problems can be taken to be at bottom combinatorial
questions.
But does that reduction of all classical mathematics (i.e., all mathe-
matics that can be expressed in higher-order logic) vindicate Hilbert? It
remains to be studied in what sense (if any) IF first-order languages can
help us to fulfill Hilbert's dream of showing the consistency of arithmetic
and analysis. What is clear in any case is that it does not save Hilbert's
specific program of finding finitistic consistency proofs for arithmetic.
But that specific program is impossible anyway. Hilbert believed that the
kind of logic needed in mathematical reasoning is axiomatizable. Indeed,
Hilbert (1918, 153) believed that such an axiomatization had already been
accomplished.
The completion of this grand enterprise of Russell's of an axiomafi::ation of logic can be

viewed as the crowning achievement of the project (Werk) ofaxiomatization in general.
Alas, Hilbert's trust in a deductively (semantically) complete logic of

mathematics turned out to be misplaced. Godel's results show Hilbert's
beliefsimply not true. This point is reinforced by IF first-order logic. Even
though it is our basic ground-floor logic, in that it is the true logic of
quantifiers, it is not axiomatizable.
This leads immediately to the nod question. What is a realistic way of
fulfilling as much of Hilbert's vision as possible? What is the best thing that
can be done? More analysis and more discussion is undoubtedly needed
to answer that question. But a couple of facts seem to be clear to me.
First, in his own statements, like the one quoted in the beginning of this
work, Hilbert emphasizes the need of a combinatorial foundation of the
logic needed in mathematics as much as, and more than, the allegedly
purely formal nature of mathematical reasoning, which he interpreted as
reasoning by means of an axiomatizable logic. In this respect, the reduction
to IF first-order logic seems to me to vindicate the essential elements of
Hilbert's vision as fully as it is possible to vindicate Hilbert in any sense.
There exists of course alternative post-Godelian reconstructions of
Hilbert's program. Most of them are nevertheless proof-theoretical. While
acknowledging their great interest, it seems to me that the only way to
clear-cut progress here is a model-theoretical one. And this is the way
which I have followed. The question I have shown how to reduce to IF
first-order logic is precisely: Are all the models of a given axiom system
included in the models of a putative theorem? And leach such question
turns out to be equivalent to a question concerning the validity (truth in
each model) of some one sentence of IF first-order logic.
But a critic might still object to my vindication of combinatorial rea-
soning as the "tool of all tools" of mathematical thinking. For a critic might
question the status of my reasoning in this paper about IF first-order logic
and more generally the status of model-theoretical and other metatheoret-
ical reasoning about my IF first-order logic. The cornerstone of all model
theory is the notion of truth in a model. And when one looks at that,
one immediately sees that Tarski-type truth-definitions inevitably involve
second-order logic. Likewise, so do my truth-conditions for IF first-order
sentences.
This seems to entail a catastrophe for my variant Hilbertian program

of construing all mathematical reasoning as being at bottom combinato-
rial, that is, If first-order reasoning. Not only does a second-order truth-
definition threaten to transcend the scope of first-order conceptualizations.
Most philosophers and logicians think of second-order logic as involving
most of the same problems of set existence as set theory. In Quine's words,
second-order logic is for them "set theory in sheep's clothing". Thus there
seems to be a type of mathematical reasoning, viz., a metamathematical
one, which belies my thesis of the essentially combin~torial character of
mathematics.
It is here that the virtues of IF first-order logic are seen at their best
(cf. here Hintikka 1996, Chap. 6). A moment's look at truth-conditions for
IF first-order sentences, e.g., a look at such second-order sentences as (3)
I
or (8), shows that they are of the E form. As was mentioned above, that
implies that they can be translated into an IF first-order language.
Furthermore, it can be shown that these truth-conditions for IF first-
order sentences can in the most natural sense imaginable be combined
into a truth predicate definable in the very same IF first-order language,
assuming of course that it is rich enough to allow us to express its own
syntax, for instance via the technique of Godel numbering. In brief, the
notion of truth for a suitable IF first-order language can be handled in that
very same language. More generally, there are in principle no obstacles
to developing a model-theoretical metatheory of a suitable IF first-order
language in that same language. IF first-order languages can be self-applied
not only syntactically but semantically.
Thus IF first-order logic also enables us to carry out an essential part
of Hilbert's overall vision. It enables us to construct a metalanguage for
the language of each mathematical theory, a metalanguage which does
not involve any objectionable set-theoretical or higher-order elements.
In fact, if the language of the given theory is rich enough, then it can
serve, after having been reconstructed in IF first-order terms, as its own
metalanguage. Admittedly, such languages are not finitistic and hence do
not fulfil Hilbert's daydream. But in a deeper sense, it seems to me, they
do vindicate his combinatorial vision.
In modifying the biblical opening line for the purpose of his motto,
Hilbert perhaps ought to have opted for the other pole of Quine's contrast:
In the beginning there was the particular object.
NOTES
1 Hilbert thought that the axiomatic method was needed to identify the substantive assump-
tions of a scientific, e.g., physical, theory. A statement to this effect is found in Frege (1976,
68-69) (letter to Frege, 29 December 1899). The same point is also the burden of the
sixth of his famous list of mathematical problems (Hilbert 1900). Hilbert practiced what
he preached for instance in his axiomatization of thermodynamics; see Hilbert (1912).
2 For instance, in the index of Torretti (1978) we find a reference to Hilbert's formalism".
But all that there is in the text is a description of how "Hilbert proves the consistency of his
axiom system by proposing an interpretation which satisfied it", followed by a statement
that in his "later life, Hilbert devoted much effort to prove the consistency of arithmetic
directly, by constructing a sound, complete, syntactically consistent formalization of it".
But the notion of completeness used here by Torretti refers to the model-theoretical notion
of logical consequence and hence is not a formalistic one.
3 For the background of this quip, see Toepell (1986a, 40-43).
4 Cf. Dauben (1979, 142-148).
5 Indirect support for attributing this idea to Hilbert is found in Tarski (1956, 199), who
describes Hilbert's project as capturing the idea of truth for a class of models ("individual
domains") under the cover of the term "general validity".
6 This background of Hilbert's later program in the foundations of mathematics is often
neglected. Against this background, it is seen that Hilbert's program had nothing to do with
the applicability of mathematics to reality, with the reliability of interpreted logical reason-
ing in mathematics, etc. Here my interpretation is in contrast to such works as Detlefsen
(1986).
7 See Kreisel (1958, 346).
8 See Hilbert (1918, 149).
9 With the following, cf. Hilbert and Bemays (Vol. 2; 1034-1039); Leisenring (1969).
10 See Hilbert (1922, 157; 1923, 151-152).
II For it, see Hintikka (1995, 1996, Chaps. 3-4).
12 All such translations are (both in ordinary first-order logic and in IF first-order logic) of
the ~ I form, that is, they have the form of a string of second-order existential quantifiers
followed by a first-order formula.
13 Cf. here Hintikka (1996, Chap. 9).
REFERENCES
Benacerraf, Paul and Hilary Putnam (eds): 1983, Philosophy of Mathematics, 2nd ed.,
Cambridge University Press.
Beth, Evert: 1955, 'Semantic Entailment and Formal Derivability', Mededelingen van de
Koninklijke Nederlandse Akademie van Wetenschappen, Afd. Letterkunde, N.R., Vol.18,
No.13, pp. 309-342.
Curry, Haskell B.: 1954, 'Remarks on the Definition and Nature of Mathematics ',Dialectica
8, 228-33; reprinted in Benacerraf and Putnam, 1983, pp. 202-206.
Dauben, 1. W.: 1979, Georg Cantor, Harvard University Press, Cambridge, Massachusetts.
Detlefsen, Michael: 1986, Hilbert's Program, D. Reidel, Dordrecht.
Frege, Gottlob: 1976, in G. Gabriel et al. (eds.), Wissenschaftliche Briefwechsel, F. Meiner,
Hamburg.
Goldfarb, Warren: 1979, 'Logic in the Twenties: The Nature of the Quantifier', Journal of
Symbolic Logic 44, 351-68.
Henkin, Leon: 1949, 'The Completeness of the First-Order Functional Calculus', Journal
of Symbolic Logic 14, 159-66.
Hilbert, David: 1899, Grundlagen der Geometrie, originally in Festschrift ::u Feier der
Enthiillung des Gauss-Weber-Denkmals, Teubner, Leipzig. Several later revised editions,
including second, 1903, seventh, 1930, tenth, 1968.
Hilbert, David: 1990 'Mathematische Probleme', Nachrichten der Kgl. Gesellschaft der
WissellSchaften =u G6ttingen, Math.-phys. Klasse 3,253--297.
Hilbert. David: 1912. 'Begriindung der kinetischen Gastheorie', Mathematische Annalen
72,562-77.
Hilbert, David: 1918. 'Axiomatisches Denken', Mathematische Annalen 78, 40.>-15.
Hilbert, David: 1922, 'Neubegriindung der Mathematik' ('A New Foundation for Mathe-
matics') Abhandlungen aus dem Mathematischen Seminar der Hamburg Universitiit 1,
157-177.
Hilbert, David: 1923, 'Die Logische Grundlagen der Mathematik', Mathematische Annalen
88,151-165.
Hilbert, David: 1926, 'Ober das unendliche', Mathematiscile Annalen 95, 161-190, in P.
Benacerrafand H. Putnam (trans.leds.), Philosophy of Mathematics: Selected Readings,
2nd ed., Cambridge University Press, 1983, pp.183--20 I, esp. pp.190--191.
Hilbert, David: 1935, Gesammelte Abhandlungen, Vol. 3, Springer-Verlag, Berlin.
Hilbert, David: 1972 (1912), 'Foundations of the Kinetic Theory of Gases', in Stephen
Brush (ed.), Kinetic Theory, Pergamon Press, New York, pp. 89-101. (English translation
of the 1912 original.)
Hilbert, David and Paul Bernays: 1934-1939, Grundlagen der Mathematik, Springer-
Verlag, Berlin.
Hintikka, laakko: 1955, 'Form and Content in Quantification Theory', Acta Philosopica
Fennica 8, II-55.
Hintikka, laakko: 1973, Logie, Language-Games and Information, Clarendon Press,
Oxford.
Hintikka,laakko: 1995, 'What is Elementary Logic? Independence-Friendly Logic as the
True Core Area of Logic', in K. Gavroglu et al. (eds.), Physics, Philosophy and the
SCientific Community, Kluwer, Dordrecht, pp. 301-326.
Hintikka, laakko: 1996, The Principles of Mathematics Revisited, Cambridge University
Press.
Kreisel, Georg: 1958, 'Hilbert's Programme', Dialectica 12, 346-72; reprinted with revi-
sions in Benacerraf and Putnam, 1983, pp. 207-238.
Leisenring, A. c.: 1969, Mathematical Logic and Hilbert's e-Symbol, MacDonald Technical
& Scientific, London.
Peckhaus, Volker: 1990, Hilbertprogramm und Kritische Philosophie, Vandenhoeck &
Ruprecht, Gottingen.
Poincare, Henri: 1902, 'Review of Hilbert's Grundlagen der Geometrie', original French in
Bulletin des Sciences Mathematiques 26, 249-272; English translation in Philip Ehrlich
(ed.), Real Numbers, Generalizations of the Reals, and Theories of Continua, Kluwer,
Dordrecht, 1994, pp. 147-168.
Russell, Bertrand: 1938, The Principles of MathematiCS, 2nd ed. (I st ed. 1903), George
Allen & Unwin, London.
Sina.. eur, Hourya: 1993, 'Du formalisme ala constructivite: Le finitisme', Revue Interna-
(ionale de Philosophie 47, 251-83.
Tarski, Alfred: 1956, Logic, Semantics, Metamathematics, Clarendon Press, Oxford.

Toepe\l, M.-M.: 1986a, Uber die Entstehung von David Hilbert s 'Grundlagen der Geome-
trie', Vandenhoeck & Ruprecht, Gottingen.
Toepel\, M.-M.: I 986b, 'On the Origins of David Hilbert's "Grundlagen der Geometrie" "
Archive for the History of Exact Sciences 35, 329-44.
Torretti, Roberto: 1978, Philosophy of Mathematics from Riemann to Poincare, D. Reidel,
Dordrecht.
Walkoe, W. Jr.: 1970, 'Finite Partially Ordered Quantification', Journal o/Symbolic Logic
35, 53~55.
Department of Philosophy
Boston University
Boston, MA 02215
USA
hintikka@bv.edu
6
STANDARD VS. NONSTANDARD DISTINCTION:

A WATERSHED IN THE FOUNDATIONS OF MATHEMATICS
I. THE HENKIN DISTINCTION
In this paper, I will discuss a conceptual distinction, or a contrast, between

two opposing ideas, that has played an extremely important role in the
foundations of mathematics. The distinction was first formulated
explicitly, though not quite generally, by Leon Henkin in 1950.' He called
it a distinction between the standard and the nonstandard interpretation
of higher-order logic. I will follow his terminology, even though it may
not be the most fortunate one. One reason for saying that this nomen-
clature is not entirely happy is that it is not clear which interpretation
is the "standard" one in the sense of being a more common one histor-
ically. Another reason is that one can easily characterize more than one
nonstandard interpretation of higher-order logic, even though Henkin
considered only one. 2 Furthermore, it turns out that the distinction (rightly
understood) is not restricted to higher-order logics. 3 Last but not least,
it is far from clear how Henkin's notion of standard interpretation or
standard model is related to logicians' idea of the standard model of such
first-order theories as elementary arithmetic, in which usage "standard
model" means simply "intended model".
In spite of these partly terminological qualifications, Henkin's dis-
tinction is a milestone in the clarification of the conceptual foundations
of logic and mathematics.
Henkin's distinction can be explained without excessive formal detai1. 4
It concerns in the first place the interpretation of higher-order quantifiers.
Indeed, anyone who uses such quantifiers faces a momentous choice. Let
us consider as an example of a second-order quantifier which involves
a one-place class or predicate variable, say X. Its values can be taken
to be either classes of individuals or properties of individuals (i.e.,
concepts applying to individuals). In either case, the same dilemma
confronts a higher-order logician, even if its horns look different on
the two interpretations.
If the values of X are thought of as being classes, then the crucial
question is whether the range of a quantifier containing X is the entire

STANDARD VS. NONSTANDARD DISTINCTION 107
power set P(do(M of the relevant domain do(M) of individuals, or

only some designated subset of P(do(M. In other words, the question
is whether the values of X are arbitrary extensionally possible classes
or whether only some such classes are accepted as values of X. This
distinction is extended as a matter of course to variables of other higher-
order types. The former alternative results in what is was called by
Henkin the standard interpretation of higher-order logic, the latter in a
nonstandard interpretation.
To reach a full-fledged standard or nonstandard interpretation, this
same distinction has to be applied, of course, to variables of each logical
type higher than that of individuals. An especially important case in point
is the interpretation of function variables, as witnessed by the fact that
a specific nonstandard interpretation is sometimes imposed on them on
purpose. The best known essay in this direction is probably G5del's
functional interpretation of first-order logic and arithmetic. S There the
values of function variables are restricted to recursive functions of the
appropriate type. However, notwithstanding the variety of different types
of entities and their intricate interrelations, the apparently special case
of one single second-order quantifier with one-place predicate or class
variable is fully representative of the entire theory of finite types. For
there are results which show that if the standard interpretation is granted
to one single second-order one-place quantifier, the entire higher-order
logic (theory of finite types) with the standard interpretation can be
reconstructed by its means. 6 Of course, by the same token the inter-
pretation of a single function variable determines in a sense the inter-
pretation, of the entire higher-order logic, for classes can always be
handled by means of their characteristic functions.
The interpretational dilemma cannot be avoided by switching to
variables which range over properties or relations in contradistinction
to their extensions. Then the question whether the standard interpretation
is assumed becomes the question as to whether for each class (poten-
tial extension) C, there is a property which has C as its extension or,
in Frege's term, as its value range. 7
As a bargain, we are put into the position of being able to distinguish
from each other several different but related questions which are some-
times confused with each other. One can ask, as we just did, whether it
can be assumed that for each class (potential extension) there is a concept
which picks up this class as its extension. But one can also ask the mirror-
image question: Is there an extension for each concept? It turns out to
be extremely important to separate the two questions from each other. 8
We can also formulate a third question, viz.: Are two higher-order

entities (e.g. concepts) with the same extension identical?9 This question
is often combined with the first two questions and formulated as the
query: Is higher-order logic extensional? It turns out, however, that
often the philosopher who asks such a question really has in mind only
one of the three questions. It is also important to realize that an
"extensionalist" answer to the question concerning the identity conditions
of higher-order entities by no means prejudices his answer to either of
the two mirror-image questions. Likewise, the question as to whether
each concept has a corresponding extension is different from the question
whether each class is the extension of some concept (and likewise for
other higher-order entities).
2. DIFFERENT NONSTANDARD INTERPRETATIONS
A nonstandard interpretation can be of many different kinds. The relevant

range of X is sometimes thought of as a part of the specification of the
model M in which a higher-order formula is being interpreted, sometimes
as a restraint on the interpretation. Usually, the range of X and the ranges
of other higher-order variables cannot be selected completely arbitrarily,
however, but are subject to certain closure conditions. For instance, the
totality of these different ranges is often assumed to be closed with respect
to Boolean operations and to projective ones, in other words, the usual
formation rules are assumed to preserve interpretability.
Sometimes, a nonstandard interpretation is guided by the idea that only
such properties, relations and functions can be assumed to exist as can
be defined or otherwise captured by a suitable expression of one's
language. We might call this the definability interpretation. In the case
of infinite models, this leads inevitably to a nonstandard interpretation,
for there can be only a countable number of such definitions or char-
acterizations available for this purpose. Hence they cannot capture all the
subsets of do(M), for there is an uncountable number of such subsets.
Indeed, this variant of nonstandard interpretation was the only one
considered by Henkin, and it is sometimes assumed that it is the only
possible nonstandard interpretation of higher-order logics. This view is
nevertheless wrong, and perniciously wrong. A nonstandard interpreta-
tion is not necessarily a substitutional interpretation. For instance, one
might assume that not even all higher-order entities representable in a
language exist unproblematic ally and restrict that privilege only to entities
that admit of a predicative representation.
What is true is that historically speaking the kind of nonstandard

interpretation Henkin considered has been the most important one, and
has guided the thinking of prominent mathematicians. For instance, we
read in a letter from Lebesgue to Borel: 10
... it is impossible to demonstrate the existence of an object without defining it. (Emphasis
in the original.)
But what does it mean to define an object?
Here the word define always means to name a property characteri.:ing what is defined.
Likewise we read in a letter by Hadamard: II
... debate is the same as the one which arose between Riemann and his predecessors
over the notion of function ... Tannery's arbitrary choices lead to a number ... which
we would be incapable of defining.
It is nevertheless crucial to realize that the definability interpretation is

not the only possible nonstandard interpretation. In brief, the non-
standard vs. standard distinction is not the same as the distinction
between, on the one hand, variables each of whose values always has
a representative in language (which can serve as a substitution value
for the variable) and, on the other linguistic representations. The latter
issue can be called the problem of nonlinguistic existence (of values of
higher-order variables). It may historically speaking have been the source
of the acceptance of standard vs. nonstandard distinction, but system-
atically speaking it is independent of this distinction. 12 In general,
representability in a language on the basis of a theory depends on the
language and on the theory, while the choice between the standard and
a nonstandard interpretation faces us in using any higher-order language.
But what might the other nonstandard interpretations look like? A
taxonomy of nonstandard interpretations should probably be attempted
only by model-theoretical logicians in the sense of Barwise and
Feferman. 13 One particular alternative nonstandard interpretation never-
theless deserves to be mentioned separately. It consists in restricting
the values of higher-order variables to recursive entiti.es, for instance,
recursive functions and functionals. This type of nonstandard interpre-
tation is instantiated by Godel's famous Dialectica interpretation, also
known as Godel's functional interpretation, mentioned above. Such an
interpretation is normally different from, and more restrictive than, the
definability interpretation of higher-order variables. Interpretations of this

kind have in fact several highly interesting properties. '
One reason why the distinction between different kinds of non-
standard interpretations matters is that they behave differently. For one
thing, Henkin showed that higher-order logic is axiomatizable on the
definability interpretation, while suitable variants of a Godel-type
functional interpretation are easily seen to be unaxiomatizable.
3. ST ANDARD INTERPRETATION AND THE IDEAS OF

ARBITRARY FUNCTION AND ARBITRARY SEQUENCE
Since the standard vs. nonstandard distinction was formulated explic-

itly as late as in 1950, it may be asked how it could have played any
role in earlier foundational discussions. What I shall show is that, appear-
ances (and the absence of a uniform terminology) notwithstanding, the
distinction played an extremely crucial role in the development of logical
and foundational studies long before its formulation by Henkin.
One particular manifestation of the distinction in earlier discussion
is especially straightforward. As can be seen from the explanations given
above, the idea of a standard interpretation of higher-order quantifiers
is virtually identical with the idea of a completely arbitrary function.
A more common locution for what was referred to above as an exten-
sionally possible function is in fact precisely an "arbitrary" function.
Hence the gradual development of what in effect was the notion of
standard interpretation can be partially followed by tracing the history
of the notion of a (completely) arbitrary function.
In a wider historical perspective, the growth of the idea of an arbitrary
function is one of the most important conceptual developments in
mathematics. This development has not always been fully appreciated.
For instance, the concept of arbitrary function is often credited to
Dirichlet, even though in reality it was formulated as early as 1755 by
Euler. In his lnstitutiones calculi differentialis, Euler wrote: 14
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 determined by others. If, therefore, x denotes a variable quantity,
then al1 quantities which depend upon x in any way or are determined by it are called
functions of it.
Similar quite general characterizations of a function are found in other

STANDARD VS. NONSTANDARD DISTINCTION III
mathematicians, for instance in Lobachevsky. According to I. H. Anellis,

Lobachevsky thought that "general thought demands that a function If]
of x be called a number which is given for each x and which changes .
. . together with X".IS The case of Lobachevsky is interesting because
the development of non-Euclidean geometries was one of the factors that
prompted mathematicians to consider a wider class of structures than
before, for instance because not all non-Euclidian geometries can be
modelled within the Euclidian one.
In Euler's time, the concept of arbitrary function had relatively little
use in state-of-the art mathematical research, even though he did find
some uses for his widened concept of a function. (See below.) In his own
work, Euler in effect restricted his attention to functions which have a
finite number of discontinuities between which the function is con-
tinuous and otherwise analytically defined. 16 Actual work both in the
foundations of mathematics and at the cutting edge of the subject
nevertheless gradually raised the consciousness of most mathematicians
in the course of the nineteenth century. This development is rife with
tensions and ambiguities which are intimately connected with the standard
vs. nonstandard contrast. On the one hand, the innocent looking Eulerian
definition of an arbitrary function turned not to admit of surprisingly
unruly denizens into the world of mathematics. Many of the results that
had been taken more or less for granted turned out to hold only for
functions with specifiable properties. For instance, it turned out that even
a continuous function need not be differentiable.
On the other hand, over the years more and more of the prima facie
pathological functions turned out to be relevant to the problems of
actual mathematical research. For instance, some nontrivial results turned
out to hold on surprisingly weak conditions. Harold M. Edwards writes: 17
What brought about the change was the discovery of the undreamed-of generality with
which some theorems of analysis are valid. The assumptions needed to prove that the
Fourier series of a function converges to the function were found to be so weak that
mathematicians were led to the outer reaches of what was conceivable as a "function",
and encouraged to discard any notion that a function needed to be given by a fonnula
or by a precisely fonnulated process.
Other examples are offered by functions defined by arbitrary power series

or by different kinds of integrals. Such developments brought more and
more of the "arbitrary" functions to the purview of mathematics.
As a consequence of such developments, a lively interest arose in
the concept of function and in the question as to how wide a class of
functions, possibly including all arbitrary function, mathematicians

should be concerned with. A beginning was made as early as in the
mid-eighteenth century, "when the need arose to consider functions that
presented themselves not as analytic expressions but as solutions to partial
differential equations from physics. These functions might not be given
by explicit formulas at all". Indeed, Euler's motivation in extending
the concept of function may have been a case in point, for "in his
pioneering study of partial differential equations of 1734, Euler admitted
'arbitrary functions' into the integral solutions".t8 The telltale symptom
in these developments is not so much the acceptance or rejection of
the concept of an arbitrary function, as the growing awareness that in
mathematics we must be prepared to deal with increasingly unruly func-
tions and perhaps even that there is in principle no limit to this expansion
of the class of functions mathematicians have to countenance.
The tension generated by this growing awareness is in evidence in
several leading mathematicians. For instance, Weierstrass considered
the idea of arbitrary real function but rejected it as being so vague and
so general as not to enable us to find any interesting things to say about
them as such. The other side of the Weierstrassian idea of function,
however, was his desire to consider as large a class of functions as
possible, to find "the largest class of functions for which one can give
an analytic representation and which can most fully satisfy the needs
of analysis":9 Kronecker, in contrast, rejected clearly the idea of an
arbitrary function. As Harold M. Edwards puts it, "what is missing, and
what is excluded by his [Kronecker's] principles is integrals of 'arbitrary'
functions or sums of 'arbitrary' infinite series".20
There is nevertheless a subtler way in which the idea of an arbitrary
function insinuated itself into the thinking of mathematicians, indepen-
dently of whether they made much use of it or not. This way was opened
simply by the development of a rigorous treatment of the foundations
of the calculus by Cauchy and others. It instructive here to think of the
format of the familiar " - 0 definitions" of, say, continuity or differ-
entiability. In them, we are considering a function for, and we specify
by means of the logical concepts of quantifiers (usually left unformal-
ized, of course) we specify when j(x) is continuous or differentiable
at xo.
Part of the original motivation might have been to limit mathemati-
cians' attention to the well-behaved functions they wanted to consider.
But to carry out the limitation in this way meant to employ a wider
conception of a function out of whose scope the good guys were picked
out by the defining condition. Even though the new foundations of

calculus were not initially put to such uses, they ipso facto provided some
conceptual tools for dealing with arbitrary functions. As Wittgenstein
would say here, if it makes sense to say that a function is continuous,
it must make sense to say that a function is not continuous.
However, there are even more basic ways in which the standard vs
nonstandard contrast entered mathematicians' foundational work and
foundational disputes in the nineteenth century. The contrast does not
matter only to the idea of a mapping (function) from real to reals. It
also affects the idea of a function of integers with integral values. As a
special case, it thus pertains to the idea of a sequence (an arbitrary
sequence) of natural numbers.
Here we are in fact dealing with a real watershed in nineteenth-century
foundational discussions. The issues can be illustrated in terms of an
example. It is known that Kronecker rejected, and criticized Weierstrass'
theory of analytic functions. This might at first sight seem puzzling.
Kronecker is often said to have been a "finitist", but he certainly accepted
the infinite set of natural numbers. And the notion of an arbitrary
(countable) sequence of integers does not involve any more dangerous
sense of infinity than the set of natural numbers. Hence we face the puzzle
of explaining precisely in what sense Weierstrass went beyond what
Kronecker found acceptable. Weierstrass was not dealing with uncount-
able infinities, as Cantor was, nor did he seem to be relying on any
dangerous infinitistic methods of reasoning. So what was the problem
with "the modern theory of functions" according to Kronecker?
Weierstrass' theory was basically a theory of power series with arbi-
trary integral coefficients. 21 This notion seems innocent enough, but it
was Kronecker's stumbling-block. In his criticism of Cantor and
Dedekind he put his cards on the table: 22
Even the general concept of an infinitary series, for example, one which according to
definite powers of variables is in my opinion only permissible with the reservation that
in each particular case, on the basis of arithmetical laws of constructing terms (or
coefficients), just as above, certain assumptions must be shown to hold which are
applicable to the series like finite expressions ....
Thus Kronecker's hete noire was not the infinitude of Weierstrass' power
series, but the assumption that the sequences of coefficients in different
power series included arbitrary sequences of integers. In other words,
what he rejected was a special case of the standard interpretation. For,
even though Weierstrass did not find much use for the idea of an arbi-
trary real-valued function (with real arguments) in his theory of func-

tions, he was relying on the idea of an arbitrary sequence of integers, that
is, of an arbitrary function from natural numbers to integers. That this
idea was what bothered Kronecker is also seen from his requirement
that there must be an arithmetical law governing the construction of
the successive coefficients of a power series. Thus what Kronecker in
effect was objecting to was the standard interpretation of the notion
"any sequence of natural numbers".
Likewise, Dedekind's analysis of real numbers utilized arbitrary
sequences of rationals, and was therefore also rejected by Kronecker.
Furthermore, the standard vs. nonstandard distinction has conse-
quences even for the ontology of mathematics. When set theory was
developed, it become clear that quantification over all reals is tantamount
to standard quantification over sets of integers. A mathematician or
philosopher who rejects all standard interpretation of higher-order
quantifiers must therefore reject also unrestricted quantification over
reals. Among other things, this observation puts in an interesting light
Kronecker's idea that only natural numbers truly exist.
Thus the contrast I am discussing can be seen to enter one of the
most central issues in mathematicians' foundational work in the late
nineteenth century.
4. SET THEORY AND THE TWO INTERPRET A TIONS
In this perspective, the original idea underlying Cantorian set theory

meant a resolute decision to go whole hog, to adopt the unrestricted
notion of an arbitrary function, and to make systematic use of it. Of
course, the idea of (an arbitrary) set is more prominent in set theory
than that of arbitrary function. This makes little difference, however,
for the two are interdefinable. Indeed, in my very explanation above of
the standard interpretation I made use of the set-theoretical notion of
power set. This explanation presupposed that in the definition of a power
set the sets presupposed there were arbitrary ones. Even without exten-
sive documentation it is eminently clear that the conception of set which
Cantor was trying to implement was the "standard" one, that is, the
conception of an arbitrary set. (Some nice documentation is in fact
provided in Hallett.)23 In other words, the early set theorists wanted to
develop an approach such that in the models of set theory every exten-
sionally possible class, Russell's "class in extension", is present.
But even before set theorists were ready to face this task, they had
a series of other problems in their hands. What the paradoxes of set theory
seemed to show is that not even all classes definable in the language
of set theory could be assumed to exist. As we might put it, with the
benefit of hindsight, even Henkin's one and only nonstandard interpre-
tation thus seemed to be too permissive, not to mention the standard
interpretation.
The main strategy which was used - and is still being used - for the
purpose of developing a set theory free of paradoxes was to formulate
it as an axiomatic theory. Moreover, this theory was thought of - and
is still so thought - as a first-order theory. What the original reasons
and the causes of this practice were is not crucially important here.
They may have included Frege's treatment of sets as complex individ-
uals, more accurately, as value-ranges of concepts. It is also patent that
the entire distinction between first-order logic and higher-order logic was
unclear at best to many of the mathematicians and logicians involved.
Indeed, the first axiom system for set theory was proposed by Zermelo,
who never reached a clear understanding of the differenc:e between first-
order and higher-order axiomatizations.
But if set theory is thought of as a first-order theory, the usual way
of distinguishing between standard and nonstandard interpretations is lost.
The only notion of model that makes unproblematic sense is some
suitable nonstandard notion. Thus the all too familiar idea of axiomatic
set theory has as a matter of historical fact aided and abetted mightily
a nonstandard view of set theory. Even if a set theorist formulates, say,
the power set axiom and requires that the elements of the power set of
a given infinite set S are all (and only) subsets of S, there is nothing
in his or her axioms that guarantees that those subsets include all the
relevant "sets in extension".
Even the particular way in which existential assumptions were intro-
duced by the early axiomatists of set theory encouraged them to think
in terms of nonstandard models. Zermelo's idea was that those sets
exist that are picked out by certain properties which he called definit.
What his successors, such as Fraenkel, Skolem and Weyl, did in different
ways was, roughly speaking, to identify definit properties with those
representable (definable) in the language of axiomatic set theory. This
pushed the model theory of axiomatic set theory very close to the non-
standard model theory of higher-order logics. Indeed, if one is willing
to countenance nonstandard models of higher-order logic, one loses much
of one's motivation to climb to higher-order logic in th(~ first place.
Hence the development of axiomatic set theory has encouraged logi-
cians and mathematicians to think in nonstandard terms. This has directed

set theorists' attention away from their original main task of imple-
menting the standard interpretation. Of course, at first sight it makes little
sense to speak of standard interpretation of a first-order axiomatic set
theory. However, model-theoretically speaking one can require that each
extensionally possible class is picked out by some set that exists in the
model in question. In other words, one can still try to require somehow
that all extensionally possible subsets of each given set are present in
the intended models. Latter-day set theorists simply have not had the
courage of Cantor's convictions. They have played with different ways
of postulating sets of larger and larger cardinality. However, this is a
different enterprise altogether from guaranteeing the presence of all the
subsets of an already given infinite model.
If one is in an exceptionally generous mood, one can perhaps consider
some of the post-Cantorian developments in set theory, especially the
introduction of the axiom of choice and more generally the genesis of
the cumulative concept of set, which are usually thought of in their merely
negative role as ways of avoiding paradoxes, as being in reality attempts
to approximate Cantor's ideas more and more closely. But if one does
that, one has to admit that the approximations have remained quite far
from their "standard" target. The result is in fact much more like the
definitory nonstandard interpretation than the standard one.
One may even suspect philosophers of set theory of a confusion
between two different senses of extensionality distinguished in section
1 above. Virtually the entire struggle to eliminate paradoxes from set
theory by axiomatizing it can be viewed as an attempt to secure the
existence of the extension of each concept - or at least as many concepts
as possible. In the midst of that struggle, the reverse question as to
whether there is a concept for each extension was largely forgotten.
5. TYPE DISTINCTIONS AS A WAY OUT
Ironically, a solution to the problems caused by paradoxes was avail-

able at a relatively early stage of the history of this subject. Admittedly,
this solution amounts to cutting the Gordian knot rather than unravel-
ling it. It is in a sense the precise opposite to what happens in set theory.
In set theory, sets of individuals are treated logically speaking on the
same level as individuals; and the same for entities of other logical types.
Even Frege, though he draws a hard-and-fast line between objects and
functions, thought he had to relate functions systematically to objects
through their value-ranges. This plunged him straight into the bottom-
less pit of Russell's paradox.
By the simple expedient of distinguishing between entities of different
logical levels (orders, types), you can in one fell swoop rid yourself of
all fear of paradoxes. What is more, you do not have to worry about
an infinite regress - or, rather, infinite ascent - for second-order logic
turns out to be amply sufficient for the purposes of mathematicians, if
we assume the standard interpretation. For then we can easily formu-
late fully such crucial. mathematical proof principles as complete
induction and the axiom of choice. You can also formulate in a second-
order logic (with the standard interpretation) virtually all major unsolved
mathematical problems. Moreover, this formulation is not merely a
linguistic matter, a possibility of having a rich enough language to
express the relevant concepts in a suitable notation. What is much more
important, and what lends second-order logic a tremendous advantage
over set theory as a foundation of mathematics, is that mathematical
problems become well-defined problems concerning the model-theoret-
ical properties of specific second-order formulas, such as validity or
satisfiability.
What this approach does not give you is a complete axiomatization
of your logic. But this incompleteness is an inescapable fact of life which
logician and philosopher had better learn to live with and perhaps even
to love. It can even be shown that the usual first-order logic is completely
axiomatizable only because its formation rules (which go back to Frege)
arbitrarily restrict the combinatorial resources of the resulting language. 24
Hence the nonaxiomatizability of standard second-order logic can
scarcely count against it.
What is surprising to me is that the virtues of the second-order
approach have been recognized only very slowly. Indeed, the firs~ book
I am aware of that is in its entirely devoted to arguing for second-order
logic as a right medium of mathematics only came out in 1991. 25 This
slowness is partly due to the slowness on the part of logicians, mathe-
maticians, and philosophers to recognize the difference between the
standard and the nonstandard interpretation of second-order logic. For
only on the standard interpretation is there a major difference between
the usual axiomatic set theory, which is a first-order theory and hence
inevitably incomplete deductively, and standardly interpreted second-
order logic, which is semantically incomplete but which allows for a
descriptively complete formulation of all the usual mathematical theories.
What I am talking about here obviously is some variant, and possibly
some fragment, of type theory. A theory of types was formulated as early

as 1908 by Bertrand RusselI. 26 Some of the virtues of a type-theoret-
ical approach were also duly recognized and pointed out by Russell,
including the freedom of any fear of contradictions. So why was this
expedient not adopted then more widely? Why did it not help to solve
the problems of set theorists or at least to get rid of them?
Here we come to a crucial reason for the importance of the standard
vs. nonstandard distinction. An important part of Henkin's achievement
was to show that on a suitable nonstandard interpretation higher-order
logic admits of a complete axiomatization. In contrast, the incomplete-
ness of second-order logic on its standard interpretation is an easy
corollary to the other incompleteness result. This might seem to be a
reason to prefer a nonstandard interpretation, were it not for the fact
that many of the most crucial conceptualizations in mathematics can
be adequately captured only by means of the standard interpretation. This
includes such ideas as the principle of mathematical induction and the
concept of cardinality.
For instance, suppose that one formulates the principle of induction
by means of the second-order axiom
(VX) X(D) & (Vy) (X(y) :J X(y + I) :J (Vy)X(y
Then on a suitable nonstandard interpretation there may in fact exist a

model for a second-order arithmetic and a class of elements of that model
which includes and is closed with respect to x + 1 but which does
not exhaust the domain of the model because there are in the model
nonstandard integers beyond 0, 1, 2, ...
Likewise, a believer in a nonstandard interpretation cannot define
the cardinality of a set in terms of one-to-one mappings between the
elements of two classes, for two classes might then be in the ordinary
sense of the word equinumerous even though no such mapping exists
according to the nonstandard interpretation which is being presupposed.
6. RUSSELL ASSUMES A NONSTANDARD INTERPRETATION
Hence, second-order logic (or perhaps type theory) can serve as a foun-
dation of mathematics only if the standard interpretation is adopted.
And what happened historically speaking is that Bertrand Russell, the
founder of the theory of types, believed in a nonstandard interpreta-
tion, not in the standard one. Hence it is not too much of an exaggeration
to say that the entire Russell-Whitehead project, like Frege's project

before it, was doomed to fail before it started. Even when they formu-
lated and proved in their system central principles of mathematics, for
instance, the principle of mathematical induction, this was a hollow
victory, for, on the tacit nonstandard interpretation they were assuming,
their very formulations of such principles did not mean what they seemed
to say.
The presence of a nonstandard interpretation in Russell's and
Whitehead's system is seen by considering their ramified theory of
types. 27 Because they formulate their discussion in terms of the exis-
tence of propositional functions rather than of functions in the extensional
sense - Russell calls them functions-in-extension - the mechanism of
their nonstandard interpretation might not be quite clear at first. What
happens is this: classes enter into the system of the Principia only as
extensions of propositional functions. But because of the ramified hier-
archy, there is no way in the Principia to speak of all propositional
functions. Any variable for such functions belongs somewhere in the
ramified hierarchy, and hence takes as its values only such proposi-
tional functions as occur lower in the hierarchy. And without speaking
of all propositional functions, one cannot speak of a.ll functions-in-
extension, either.
Russell's and Whitehead's cure, the axiom of reducibility, does not
bring about the standard interpretation, either. What it does is to say
that each extension of a higher-type function is the extension of some
function of the lowest type, called an elementary fum:tion. What the
axiom of reducibility accomplishes is to make sure that the set of classes
that one can consider in the Principia (Le., that "exist'" in it) is closed
with respect to projective operations (quantification). It restricts the
kind of nonstandard interpretation Russell and Whitehead are in effect
assuming, but it does not in any way guarantee standard interpretation.
Indeed, if the standard interpretation were assumed for quantifiers
ranging over elementary functions, there would not be any need to assume
the axiom of reducibility. For whatever class is captured by a function
of a higher ramified type, there would already exist a. class captured
by an elementary function. Hence the need of the axiom of reducibility
shows that Russell and Whitehead are in effect assuming a nonstan-
dard interpretation. .
But even the axiom of reducibility could not do the whole job of
reconstructing received mathematics. Admittedly, Russell claimed that
he can save the principle mathematical induction in the Principia system
even without the axiom of reducibility.28 But this is a Pyrrhic victory.

What can be so saved is the proposition which, if interpreted in the
standard way, codifies the principle of mathematical induction. But
since Russell and Whitehead in effect assume a nonstandard interpreta
tion, Russell's result does not really vindicate mathematical induction.
And for this purpose, the axiom of reducibility is of no help whatso-
ever.
7. RAMSEY'S CONTRIBUTION
Russell's nonstandard preferences were diagnosed with remarkable clarity

by Frank Ramsey. His 1925 essay "The Foundations of Mathematics,,29
contains the clearest statement of the standard vs. nonstandard distinc-
tion to be found in the literature before Henkin, even though Ramsey
did not propose terminological labels for the two types of interpretations.
Ramsey's description of the problem situation is a model of lucidity.30
The theory of Principia Mathematica is that every class or aggregate (I use the words
as synonyms) is defined by a propositional function - that is, consists of the values of x
for which '~' is true, where '~' is a symbol which expresses a proposition if any
symbol of appropriate type is substituted for 'x'. This amounts to saying that every class
has a defining property. Let us take the class consisting of a and b; why, it may be
asked, must there be a function <\lX such that 'cjIa', 'cpb' are true, but all other '<\lX's false?
This is answered by giving as such a function 'x - a.V.x - b'. Let us for the present neglect
the difficulties connected with identity, and accept this answer; it shows us that any
finite class is defined by a propositional function construed by means of identity; but as
regards infinite classes it leaves us exactly where we were before, that is, without any
reason to suppose that they are all defined by propositional functions, for it is impos-
sible to write down an infinite series of identities. To this it will be answered that a
class can only be given to us either by enumeration of its members, in which case it
must be finite, or by giving a propositional function which defines it. So that we cannot
be in any way concerned with infinite classes or aggregates, if such there be, which are
not defined by propositional functions. (For short I shall call such classes 'indefinable
classes'.) But this argument contains a common mistake, for it supposes that, because
we cannot consider a thing individually, we can have no concern with it all. Thus, although
an infinite indefinable class cannot be mentioned by itself, it is nevertheless involved in
any statement beginning' All classes' or 'There is a class such that', and if indefinable
classes are excluded the meaning of all such statements will be fundamentally altered.
Ramsey grasps here with a remarkable clarity the crucial point of the
standard vs. nonstandard distinction, viz. that any attempt to restrict
the classes we are considering to those that can be captured by predicates
will inevitably alter the sense of such expressions as "all classes", that
is, alter the interpretation of higher-order quantifiers.
In view of the clarity and force of Ramsey's poi.nt, it might not

historically speaking be entirely unfair to call the standard vs. non-
standard distinction the Ramsey distinction rather than the Henkin
distinction.
Ramsey also recognizes that Russell's procedure of introducing classes
only via propositional functions is out of step with mathematicians'
thinking. Ramsey speaks of
a fundamental characteristic of modern analysis which we have now to emphasize. This

characteristic may be called extensionality . .. we are using 'extension' in its logical sense,
in which the extension of a predicate is a class, that of a relation a class of ordered couples;
so that in calling mathematics extensional we mean that it deals not with predicates but
with classes, not with relations in the ordinary sense but with possible correlations, or
"relations in extension" as Mr. Russell calls them. 31
Ramsey did not only offer a diagnosis as to what was ailing Russell's
and Whitehead's Principia Mathematica from the viewpoint of "modern
analysis", that is, of the standard interpretation. He set out to specify how
the Principia might be cured of the nonstandard contamination. This is
precisely what Ramsey's elimination of the ramified theory of types
amounted to. It meant adhering strictly to the standard interpretation
and dispensing with all conceptualizations that depended on non-
standard assumptions. The casualties of this purge naturally included
the axiom of reducibility. Indeed, it seems to have been the role of this
axiom that directed Ramsey's attention to the standard vs. nonstandard
distinction.
Ramsey's resolute stand in favor of the standard interpretation led him
into a dispute with Ludwig Wittgenstein concerning the definition of
identity.32 This quaint-looking controversy was more than a storm in a
Cantabridgean teacup - or was it an Austrian wineglass? For it may
very well have been instrumental in Ramsey's conversion to a form of
constructivism in his philosophy of mathematics.
One might even suggest that the entire contrast between first-order
logic and higher-order logic remains a distinction without difference
unless the standard interpretation is adopted. 33 Formally speaking, a
distinction between first-order logic and second-order logic is present
as soon as a distinction is made between first-order and second-order
variables. And this kind of distinction was made by as early a logician
as Frege. But as long as a sufficiently parsimonious nonstandard inter-
pretation is adhered to, one can treat higher-order logic simply as a
many-sorted first-order logic. Henkin's completeness. proof for non-
standardly interpreted type theory is simply one particular instance of the

successes of this strategy. In fact, contrary to the initial disbelief of many
interpreters, this is essentially how Wittgenstein in the Tractatus viewed
his "objects".34 They were in one sense of all the possible logical types,
but they were treated by means of what amounted to one and the same
logic. More specifically, there were in his universe none of the system-
atic dependencies of entities of different types which the standard
interpretation presupposes.
Thus in one sense the idea of first-order logic was born as soon as
modem logic was around. But in another sense it came about - or came
in to its own - only when Henkin formulated explicitly the standard
vs. nonstandard distinction.
8. THE STANDARD VS. NONSTANDARD DISTINCTION AND THE

AXIOM OF CHOICE
So far, I have not taken up what might seem the most conspicuous
manifestation of the standard vs. nonstandard distinction in the history
of mathematics. This manifestation is the role of the distinction in shaping
mathematicians', logicians' and philosophers' attitude to the axiom of
choice. 3s And it is in fact obvious (at least to the cognoscenti) that such
an influence has indeed been operative. It is in evidence in the case of
Bertrand Russell, whose adherence to a nonstandard interpretation was
noted above. In view of this preference, it should not come as a surprise
that Russell conspicuously shunned the axiom of choice, which he mostly
considered in the form of what he called the multiplicative axiom. On
one occasion Russell explains his doubts about the axiom by way of
an example of a millionaire who owned a countable infinity of pairs of
boots and pairs of socks.36
The problem is: How many boots had he, and how many socks? One would naturally
suppose that he had twice as many boots and twice as many socks as he had pairs of
each and that he therefore had Ko of each, since that number is not increased by doubling.
But this is an instance of the difficulty, already noted, of connecting the sum of v classes
each having Jl terms with Jl x v. Sometimes this can be done, sometime it cannot. In
our case it can be done with boots, but not with socks, except by some very artificial
device. The reason for the difference is this: Among boots we can distinguish right and
left and therefore we can make a selection of one from each pair, namely, we choose
all the right boots or all the left boots; but with socks no such principle of selection suggests
itself, and we cannot be sure, unless we assume the multiplicative axiom [i.e., the axiom
of choice], that there is any class consisting of one sock out of each pair. Henc'e the
problem.
Here we can see that for Russell an infinite class of socks cannot be
assumed to exist unless we can somehow pick it out by means of some
distinctive property. This shows clearly the role of the nonstandard
interpretation in his thinking.
Likewise, the quotations from Hadamard given above which show
his trust in a nonstandard interpretation are taken from his letter to
Borel in which he criticizes Zermelo's axiom of choice and Zermelo's
use of it in the proof of the well-ordering theorem.
Gregory Moore sums up the criticisms of the axiom of choice by Baire,
Borel, Lebesgue, Peano and Russell by saying that
the Axiom did not provide a rule by which we can carry out the choices. 37
This is of course correct. Admittedly, though, one needs here a great deal
of care in interpreting different mathematicians' and logicians' verbal
formulations. For instance, if one considers the concepts of rule and
function as being equivalent, we are back at the problem of arbitrary
functions, with the critics of the axiom of choice embracing a non-
standard interpretation (viz. the definability interpretation).
But things are not as simple with the axiom of choice as first meets
the eye. One interesting use of the axiom of choice is to provide a kind
of second-order interpretation of first-order logic. On this interpreta-
tion, a first-order statement S is interpreted by the second-order statement
S* which asserts the existence of the Skolem functions for S. If the axiom
of choice is assumed, Sand S* are equivalent. For a simple example,
(Vx) (3y) F[x, y] (1)
(where F is quantifier-free) is interpreted by
(3f) (Vx) F[x,j(x)] (2)
Now if the function quantifiers in the second-order translation S* are

interpreted in a nonstandard fashion, we obtain a statement which now
has a force different from S. But here we meet an elegant idea which
seems to originate with Godel. 38 It is to use a nonstandard interpreta-
tion of S* as defining a constructivistic interpretation also of S. (In
order to carry out this idea fully, we must actually consider also the
interpretation of propositional connectives; but this complication does
not have to concern us here.) This idea is especially appealing if the
particular nonstandard interpretation opted for restricts the values of

the function variables of S* to recursive functions, as Godel does.
But if we do so, the result is that the axiom of choice is acceptable
after all. For what it now does, in justifying the equivalence of S and
S*, is merely to implement a nonstandard (e.g., constructivistic) inter-
pretation of first-order logic, especially of first-order quantifiers. Hence
a denial of the axiom of choice is arguably based on a confusion. In order
to deny it, first-order quantifiers must be interpreted classically (i.e., in
the standard way) while higher-order quantifiers are given nonstandard
interpretation. The relationship of (1) and (2) serves as a case in point.
As a by-product, the standard vs. nonstandard distinction can be
extended to first-order logic. Contrary to the impression which my
initial explanations may have created, the standard vs. nonstandard
distinction is thus not restricted to higher-order logic. 39
These topical considerations help to bring several different histor-
ical developments into a sharper focus. For instance, the line of thought
just expounded apparently has been instrumental in turning some intu-
itionists away from Brouwer's rejection of the axiom of choice. This
can be seen, e.g., from Dummett's writings. 40
More generally but perhaps somewhat less conspicuously, my analysis
throws some light on the curious history of the axiom of choice. One
of the curiosities here is that a large number of the vocal critics of the
axiom of choice were in a subsequent analysis caught in flagrante delicto
as having used the principle tacitly in their own research. This is
documented carefully in Moore. 41 Such inconsistencies are not what we
should expect of highly sophisticated mathematicians, and they require
therefore some deeper explanation. Even though the historical material
might repay a closer study, it seems that this prima facie inconsistency
of the critics of the axiom of choice becomes less surprising and less
inconsistent than might first seem to be the case when we realize that
their real target was the standard interpretation and not the axiom of
choice.
9. NATURAL VS. FORMAL LANGUAGE
This discussion of the axiom of choice can be generalized. Whenever

we are dealing with an explicit formula which in so many words or
implicitly involves higher-order quantifiers, in many cases a classical
mathematician and constructivist can agree verbally concerning its
validity, even though they interpret the formula differently. They will still
disagree sometimes, but the disagreements are far rarer than might first
seem to be the case. The initially surprising agreement on the axiom
of choice is merely a case in point.
This observation puts certain segments of the classical foundational
discussions into an interesting perspective. Among other things, it shows
why a constructivist needs to consider explicit formulas that can be given
a nonstandard interpretation. This was certainly Krofi(~cker's strategy.
Harold M. Edwards writes on Kronecker: 42
In his works he is always specific. Formulas abound in his papers. Dedekind abhorred
formulas and tried to avoid them. Kronecker was opposite. He once said that he felt
that the essence of mathematical truth lay in formulas .... The wish to get rid of formulas
was, it seems to me; what brought set theory into being, Set theory is what remains
after formulas are banished. How can an arbitrary function be described, other than as
a set of ordered pairs? Since Kronecker wished to place formulas at the heart of his
mathematics, this motive for set theory would not have existed for him.
Of course, the contrast between formulas and ordinary-language expres-

sions should strictly speaking make no difference. In principle,
ordinary-language quantifier sentences can also be interpreted in a non-
standard way. In practice, however, such a reading is felt to be extremely
awkward. If I express (1) by saying,
for each value of x, there exists a value of y such that S[x, y]
no one will take me as really adding, satta voce,
and I can find that value of y constructivistically.
In contrast, a nonstandard interpretation of (2) is eminently natural. In

fact, a formulation like (2) virtually challenges a reader to ask, "There
exists what kind of function?". Thus ordinary language is a much more
natural medium for expressing the standard interpretation than a formal
language, and the use of formulas can accordingly be a symptom of a
10. AN EPISTEMOLOGICAL INTERPRETATION OF THE

DISTINCTION
But which interpretation is the right one? And what is my own position
vis-a-vis the distinction? Very briefly, it may be argued that the distinction
should be viewed, not as an ontological distinction, but as a partly

epistemological one. Then the standard quantifiers range over all the
objectively existing higher-order entities, whereas the nonstandard
quantifiers range over only those higher-order entities which are known
to us. Without arguing extensively for this construal of the distinction,
I can at least illustrate it by pointing out how natural it is in the context
of the history of mathematics. The widening of the class of functions,
sequences, and other higher-order entities can now be viewed as what
it in some sense obviously is, to wit, not as importing new entities to
the purview of mathematicians' attention, but as a gradual coming to
know more and more such entities.
In any case, this construal does justice to the role of mathematics in
the exploration of the natural world. Known entities play an important
role in answering questions, including the questions put to nature in
the form of an experiment.n For only entities known to you can be
referred in a reply if it is to answer fully your question. A reply to any
question is a genuine (conclusive) answer only if the questioner knows
what is being referred to. If a foreigner asks you: "Who is the leading
candidate for the Democratic presidential nomination?" and you answer,
"The governor of Arkansas", the questioner will be satisfied only if
he or she knows who the governor of Arkansas is. 44 Likewise, when
an experimental inquirer asks how the observed variable depends on
the controlled variable, merely plotting the dependence on graph paper
is not a real answer to the question. For that purpose, the inquirer must
know what function it is that the graph represents, in a stronger sense
than just being able to correlate with each other all the associated
argument values and function values. The more functions the inquirer
masters mathematically, the more experimental "answers" there are that
satisfy him or her. Thus the development of mathematics, in particular
the extension of the range of functions that are brought within the scope
of mathematical treatment, can be seen to extend directly the power of
experimental science. This offers in fact an account of the sense in which
pressures by mathematical physicists forced the widening of the set of
bona fide functions through a mathematical study of them.
Then, from the vantage point adumbrated here, the values of non-
standard function variables are precisely the functions known in the
relevant sense of the word.
Thus it is in effect as if all the functions in the standard sense can
be said to exist, but also as if only some of them are at anyone time
known to an inquirer. The latter constitute the range of nonstandard
variables. Maybe the standard interpretation can be said to be correct

ontologically but the nonstandard one epistemologically. It follows that,
since our knowledge can (and does) grow, there is no single non-
standard interpretation we must adopt.
For a philosophical analyst of the foundations of mathematics this
suggestion presents a challenge, viz. the challenge to develop a viable
higher-order epistemic logic in which the standard VS. nonstandard
distinction could be studied.
Boston University
NOTES
1 Leon Henkin: 1950, 'Completeness in the Theory of Types', Journal of Symbolic Logic
IS, 81-91. For a correction to Henkin's paper, see Peter B. Andrews: 1972, 'General
Models and Extensionality', Journal of Symbolic Logic 37,395-397.
2 See section 2.
3 Cf. here Jaakko Hintikka: 1980, 'Standard vs. Nonstandard Logie: Higher Order, Modal
and First-Order Logics', in E. Agazzi (ed.), Modern Logic: A Survey, D. Reidel, Dordrecht,
283-296.
4 The following paragraphs as well as section 3 below follow closdy the exposition in
Jaakko Hintikka and Gabriel Sandu: 1992, 'The Skeleton in Frege's Cupboard: The
Standard versus Nonstandard Distinction', Journal of Philosophy 89, 290-315.
S See 'Uher eine bisher noch nieht beniitzte Erweiterung des finitf:n Standpunktes', in
Solomon Feferman et al. (eds.), Kurt Godel: Collected Works, 1lol. 2: Publications
1938-1974, Oxford University Press, New York, 1990, pp. 240--251. (Cf. also pp.
217-241.)
6 Cf., e.g., Jaakko Hintikka: 1955, 'Reductions in the Theory of Types', Acta Philosophica
Fennica S, 59-115.
7 Cf. here Hintikka and Sandu, op. cit. note 4.
8 Cf. the end of section 4 below.
9 This question is tantamount to the question of the validity of Leibniz's Law,-
10 Quoted in Gregory H. Moore: 1982, Zermelo's Axiom of Choice, Springer-Verlag,
Berlin-Heidelberg-New York, p. 314.
II Quoted in op. cit., p. 318.
12 For instance, it appears that some commentators have misinterpreted Frege because
he does not assume the definability interpretation. (For Frege, functions exist objec- .
tively independently of their representability in language.) From thi~; they have in effect
mistakenly inferred that Frege accepted the standard interpretation. See here Hintikka
and Sandu, op. cit.
13 Jon Barwise and Solomon Feferman (eds.): 1986, Model-theoretical Logics, Springer-
Verlag, Berlin-Heidelberg-New York.
14 Quoted in Umberto Bottazzini, The "Higher CalcuLus": A History of Real and
Complex AnaLysis from Euler to Weierstrass, Springer-Verlag, Berlin-Heidelberg-New
York, 1986, p. 33.
I' I. H. Anellis, A History of Mathematical Logic in Russia and the Soviet Union,
unpublished.
16 See L. Euler: 1990, Introduction to Analysis of the Infinite, Book II, translated by John
D. Blanton, Springer-Verlag, Berlin-Heidelberg-New York, p. 6, section 9.
17 See his paper, 'Mathematical Ideas, Ideas, and Ideology', The Mathematical
Intelligencer 14(2) (Spring 1992), 6-19 (here p. 7b).
18 The first quotation is from Judith V. Grabiner: 1981, The Origins of Cauchy's Rigorous
Calculus, The MIT Press, Cambridge MA, pp. 89-90. The second is from Thomas
Hawkins: 1970, Lebesgue's Theory of Integration, University of Wisconsin Press, Madison,
p.4.
19 P. Dugac: 1973, 'Elements d'analyse de Karl Weierstrass", Archive of the History
of Exact Sciences 10, 41-176. (See p. 71; quoted in Bottazzini, op. cit., p. 199.)
20 In 'Kronecker's View of the Foundations of Mathematics', in David E. Rowe and John
McCleary (eds.), The History of Modern Mathematics, vol. I, Academic Press, San
Diego, pp. 67-77. (See here p. 74.)
21 For Weierstrass's work, see Felix Klein: 1927, Vorlesungen iiber die Entwicklung
der Mathematik im 19. Jahrhundert, vol. 1, Springer-Verlag, Berlin-Heidelberg, pp.
276-295.
22 Leopold Kronecker: 1886, 'Uber einige Anwendungen der Modulsysteme auf
elementare algebraische Fragen', Journal fUr reine und angewandte Mathematik, vol.
99, pp. 329-371, especially p. 336. Quoted in Joseph W. Dauben: 1979, Georg Cantor:
His Mathematics and Philosophy of the Infinite, Harvard U.P., Cambridge MA, p. 68.
23 Michael Hallett: 1984, Cantorian Set Theory and Limitation of Size, Clarendon Press,
Oxford.
24 See here Jaakko Hintikka: 1994, 'What is Elementary Logic? Independence-friendly
Logic as the True Core Area of Logic', in K. Gavroglu et al. (eds.), Physics, Philosophy
and Scientific Community: Essays in Honor of Robert S. Cohen, Kluwer Academic,
2' See Stewart Shapiro: 1991, Foundations without Foundationalism, Clarendon Press,
Oxford.
26 Bertrand Russell: 1908, 'Mathematical Logic as Based on the Theory of Types' ,
American Journal of Mathematics, vol. 30, pp. 222-262, reprinted in Bertrand Russell:
1956, Logic and Knowledge: Essays 1901-1950, ed. by Robert C. Marsh, Allen &
Unwin, London, pp. 59-102.
27 Bertrand Russell and Alfred North Whitehead: 1910-1913, Principia Mathematica
I-Ill, Cambridge University Press, Cambridge; second ed., 1927.
28 Op. cit., note 26, second edition, vol. I, Appendix B.
29 Frank P. Ramsey: 1925, 'The Foundations of Mathematics', Proceedings of the London
Mathematical Society, Ser. 2, vol. 25, part 5, pp. 338-384. Reprinted (among other places)
in F. P. Ramsey: 1978, Foundations, ed. by D. H. Mellor, Routledge and Kegan Paul,
London, pp. 152-212.
)0 Op. cit., p. 173 of the reprint.
31 Op. cit., p. 165 of the reprint.
32 See Maria Carla Galavotti (ed.): 1991, Frank Plumpton Ramsey, Notes on Philosophy,
Probability and Mathematics, Bibliopolis, Napoli, Appendix, and Mathieu Marion's
contribution to the present volume.
33 For a discussion of the history of this contrast, see Gregory H. Moore: 1988, "The
Emergence of First-Order Logic", in William Aspray and Philip Kitcher (eds.), History
44 lAAKKO HINTIKKA
and Philosophy of Modern Mathematics (Minnesota Studies in the Philosophy of Science,

vol. II), University of Minnesota Press, Minneapolis, 1988, pp. 95-135.
34 See here Merrill B. Hintikka and Jaakko Hintikka: 1986, Investligating Wittgenstein,
Basil Blackwell, Oxford, chapters 2 and 4.
3S Cf. Gregory H. Moore, op. cit., note \0 above.
36 Bertrand Russell: 1919, Introduction to Mathematical Philosophy, Allen & Unwin.
London, chapter 12, especially p. 126.
37 Op. cit.. p. 309.
38 See op. cit.. note 5 above, and cf. laakko Hintikka: 1993. "GOdeJ's Functional
Interpretation in Perspective', in M. D. Schwabl (ed.), Yearbook ofthl~ Kurt Gode/ Society,
Vienna. pp. 5-43.
39 See Hintikka, op. cit., note 3 above.
40 Michael Dummett, Elements of Intuitionism, Clarendon Press, Oxford, 1977, pp. 52-53
and 314.
41 Op. cit., note \0 above, especially pp. 64-76.
42 Op. cit., note 20 above, p. 71.
43 See here laakko Hintikka: 1988, 'What Is the Logic of Exp:rimental Inquiry?',
Synthese 74, 173-190.
44 For an early discussion of the nature and role of conclusiveness conditions, see Jaakko
Hintikka, The Semantics of Questions and the Questions of Semantics (Acta Philosophica
Fennica vol. 28, no. 4) Societas Philosophic a Fennica, Helsinki, 1976, especially ch. 3.
The analysis presented there is now being generalized. especially to questions whose
answers are functions.
7
STANDARD VS. NONSTANDARD LOGIC:

HIGHER-ORDER, MODAL, AND FIRST-ORDER LOGICS
Model-theoretical (seman tical) treatments of modal logic have enjoyed spec-

tacular success ever since the pioneering work by Stig Kanger in 1957.1 Quine
and others have admittedly proffered sundry philosophical objections to
modal logic and its semantics but they have not impeded the overwhelming
progress either of the seman tical theory of intensional (modal) logics or ofits
applications.
There nevertheless exists an important prima facie objection to more
of the current versions of this approach which quite surprisingly has not been
presented in its full generality in the literature. (For partial exceptions, see
above, note 1, and see below, especially notes 6 and 7.) It is directed
primarily against the current treatments of logical modalities in seman tical
terms. In order to see what this objection is, let us first recall the key idea on
which the whole possible-worlds semantics is based. This basic idea is to
require for the truth of
(1) Necp
(where 'Nec' expresses necessity) in a world Wo nothing more and nothing
less than the truth of p (Le., p simpliciter) in all worlds alternative to woo
likewise, the truth of
(2) Possp
in Wo (where 'Poss' expressed possibility) amounts to the truth ofp is at least
one such alternative. These characterizations presuppose that, instead of
considering just one world (model) at a time, we are considering a set of
models on which a two-place relation (the relation of being an alternative to)
has been defmed.
So far, so good. But what are these alternative worlds (models)? Here
we come to a veritable skeleton in the cupboard of semanticists of modal
logic: to a problem which is of the utmost importance both historically and
systematically but which has scarcely been discussed in the literature. Kripke
has always assumed that any old subset of some given set of models will do
as the set of alternatives, as long as it contains wo. 2 Or strictly speaking,
he has assumed that any set Wo of models together with an arbitrary reflexive

STANDARD VS. NONSTANDARD LOGIC 131
relation of alternativeness defmed on it is a/rame in which a modal sentence

can be evaluated for truth (with respect to each member Wo of Wo). I shall
refer to a truth-definition of this sort (for modal sentences) as a K-type truth-
defmition. ('K' for Kripke, of course.)
Almost all subsequent logicians have followed Kripke here, who was anti-
cipated among others by Guillaume and Hintikka.3
However, the arbitrariness of the choice of alternatives involved in a K-
type truth defmition appears completely unmotivated, especially in the case
of logical modalities. When we say that something is logicaliy necessary, we
do not mean just that it is true in each member of some arbitrary set of alter-
natives. Rather, we mean that it is true in each logically possible alternative.
In other words, it seems that we oUght to impose much more stringent
requirements on frames than is the case in a K-type semantics for modal
logics. We ought to require at the very least that Wo contains an instance of
each logically possible kind of model with a ftxed set of individuals 10 (or
with a subset of 10 ) as its domain of individuals. By a logically possible type
of model we of course mean a model reachable by choosing the extensions
(interpretations) of nonlogical symbols in some suitable way. Moreover, an
instance of every such model type must occur among the alternatives of each
member of Wo. We shall call such frames loframes. ('L' stands for logically
possible worlds.) The semantics based on them will be called L-semantics.
This was the semantics presupposed by Kanger in 1957 and also by
Montague in his earliest work on modallogic.4 Thus we ha.ve in fact not only
one but two quite unlike traditions in the foundations. of modal logics, the
one relying on K-semantics and the other on L-semantics. Later, however,
Montague switched over from L-semantics to K-semantics, without discus-
sing his reasons. S L-semantics has nevertheless made a couple of brief appe-
arences in the writings of Montague's associates and former students, especi-
ally David Kaplan 6 and Nino Cocchiarella. 7
The two traditions are connected with a number of other issues. K-
semantics was initially made interesting mathematically by being a natural
extension of earlier work by Tarski and his associates on Boolean algebras
with operators. 8 K-semantics is indeed unproblematic in connection with
intensional logics in the narrower sense, e.g., in connection with epistemic
logic, doxastic logic, and the logic of other propositional attitudes. It is
mostly in connection with logical modalities and with whatever analytical
modalities are involved in the analysis of linguistic meanings that K-semantics
loses its plausibility .
Almost all the actual work in the semantics of modal notions has been
based on K-semantics rather than L-semantics. It is therefore in order to ask
what will happen if we impose simUar requirements in ordinary modal logic

of logical modalities? The details of the situation require a more extensive
study than can be undertaken here. I shall make only a few general comments.
First, it may be observed that the additional requirement that is imposed
on frames by L-semantic go way beyond all the requirements that can be im-
posed on frames in K-semantics by putting further conditions on the alter-
nativeness relation. For instance, it is pretty generally agreed that in the case
of logical modalities alternative ness relation must also be transitive and
symmetric, Le., an equivalence relation. These additional requirements do not
change the picture essentially but only complicate it in certain ways which I
shall not try to discuss here.
The next main point is an obvious one. It can be expressed in the fonn of
an analogy:
-frames _ standard semantics
K-frames - nonstandard semantics
In more picturesque tenns we can put the same equation as follows.
Kanger Tarski
Kripke = Henkin
For the tenns 'standard' and 'nonstandard' are here used in the sense of
Henkin.9 These notions were defmed by him to apply in the first place only
to models for higher-order logics. In a standard model M 0 quantification over
(say) one-place predicates of individuals is understood as quantification over
all (and only) extensionally possible subsets of some given set of individuals
10 , In a nonstandard model M 1 , such quantifiers range over some fixed (Le.,
fIXed relative to M 1) subset of the power set of 10 , (This difference extends
naturally to other kinds of higher-type variable.) The analogy with L-frames
vs. K-frames is obvious.
This is an instance of a much more sweeping point. The distinction Henkin
made has so far been applied only to higher-order logics. It is true that the
labels 'standard' and 'nonstandard' have occasionally been used also in the
connection with different first-order models, especially models of
arithmetical and other mathematical axiom systems. These uses are not all
connected with each other or with Henkin's definition in any systematic
manner, however. Often, it would be less misleading to use the tenns
'intended' and 'unintended' instead of 'standard' and 'nonstandard' for these
models. It is not my purpose here to try to capture this variety of different
nonstandard models in tenns of anyone concept. I don't think that it can be
done, and some of the other kinds of nonstandard conceptions of models
defmitely deserve a separate treatment.
In this respect, a radical change has been brought already by the game-
theoretical semantics for first-order logic (and parts of natural languages)
which I have outlined in earlier writings. 10 It would take me too far to study
this new semantics in its entirety here or even to give a explicit definition of
it. Suffice it to say that this semantics yields a translation of first-order logic
into a fragment of second-order logic, possibly (depending on certain fine
points in the semantical interpretation of propOsitional connectives) into a
fragment of higher-order logic (simple type theory). Given a prenex first-
order sentence p, its translation asserts the existence of such associated
Skolem functions as make the quantifier-free part of p true for any choice
of the values of the universally quantified variables. For inst.ance,
(4) ('tx)(3y)(\Iz)(3u) M(x, y, z, u)
will translate into
(5) (3J)(3g)('tx)('tz) M(x, I(x), z, g(x, z.
Returning for a moment to the game-theoretical semantics for first-order
languages, we can think of the value of existentially quantified variables as
being chosen successively by myself in a little game against Nature who
chooses the values of universally bound variables. Then (5) will say that I
have a winning strategy in the game associated with (4). This will be the
general form of the defInition of truth in game-theoretical semantics: a
sentence is true if I have a winning strategy in the correlated game, false if
Nature has one. (Since it is not in general true that either player has a winning
strategy in an infinite two-person game, we see already here how a door is
opened by game-theoretical semantics for treating such nonclassical logics as
are not subject to the law of bivalence.) II
The game just sketched is readily extended to propositional connectives
and to quantifiers in a noninitial (nonprenex) position. For instance, a dis-
junction marks my move. I choose a disjunct with respect to which the game
is then continued. Similarly for conjunctions, except that Nature chooses
the conjunct.
A possible elaboration of these games is obtained by dividing them into
subgames. 12 After such a subgame, one of the players divulges his strategy
in the subgame. (We know from game theory that to playa game is to choose
a strategy for it. Hence to play out a game to the ~itter end is to divulge one's
strategy in it.) Then a player's subsequent moves may depend on that stra-
tegy. Since a strategy in the sub game is the function that tells one which
moves to make depending on what has happened earlier on the game and
since a player's overall strategy will be defined by a functional (function

whose arguments are functions). In this way various functional interpreta-
tions of first-order sentences can be obtained, depending on the details of the
game rules. 1 !
Independently of these details, the explicit truth-condition for each first-
order sentence can always be expressed by a second-order or higher-order sen-
tence. For what this truth-condition amounts to is the existence of a winning
strategy for myself, i.e., a strategy which wins against any strategy of
Nature's. Since strategies are represented by functions or functionals, the
truth-condition can be expressed in higher-order logic.
This translation makes the standard-nonstandard distinction automatically
applicable to first-order logic, too, independently of the details of the game-
theoretical interpretation. Depending on whether the second-order translation
(or the higher order translation) of a first-order sentence is given standard or
nonstandard interpretation, we likewise obtain a similar distinction between
standard and nonstandard interpretations of jirst-order logic. Moreover, to
both these interpretations my game-theoretical semantics assigns not only
a clear semantic meaning but a concrete pragmatic content. 14
Conversely, any distinction along these lines between standard and non-
standard interpretations of ordinary (linear) first-order logic induces as a mat-
ter of course a distinction between the corresponding interpretations of the
logic of finite partially ordered quantifiers. As I have pointed out elsewhere,lS
the logic of such quantifiers (called f.p.o. quantification theory) is 'ahnost'
as strong as second-order logic, in the sense that the decision problem for the
latter reduces to the decision problem for the former. Hence the standard-
nonstandard distinction, when made along the lines I am suggesting for linear
first-order sentences, in effect forces us to a Henkin-type distinction in
higher-order as well.16
The first main point I am making in this paper is thus that the standard-
nonstandard distinction also carries over to modal logics and to first-order
logic in an important way.
But is it a distinction with a real difference? For instance, is there any bite
in the putative criticism of the modal logic mentioned in the beginning of
this paper? Does the difference between K-interpretation and L-interpretation
really matter?
It is obvious that quite a number of changes take place already in proposi-
tionallogic when we switch from K-Iogic to L-Iogic. For whenever a complex
sentence F is satisfiable, the modalized sentence Poss(F) will be logically true
(valid) in L-semantics. In propositional logic, L-Iogic is nevertheless axiomati-
zable: 7 Obviously, modalized predicate logic (quantificaltion theory) with

L-semantics cannot be completely axiomatized. For otherwise we could
decide the logical truth of any plain first~rder sentence F by grinding out
theorems of the complete axiomatization until For Poss(-F) has made an
appearance. Of course such a decision method is known to be impossible.
This does not yet say very much of the expressive power of, say, a first-
order quantification theory cum modality with L-semantics. Does the switch
trom K-semantics to L-semantics matter greatly here?
At first sight, the answer might seem to be negative. In some parts of
customary modal logics, the choice between the K-interpretation and the L-
interpretation apparently does not affect expressive power greatly. However,
in other cases there seem to be important differences. The whole matter is
complex, and requires a fuller investigation than it can be given here. Suf-
fice it to point out some connections between this problem and certain other
important issues concerning the foundations of modallogics. 18
As the reader can gather from any explanation of Henkin's standard-
nonstandard distinction, the crucial question is whethelr an L-semantics
enables us to quantify over all the subsets of a given set (and of course over
them only). At first blush, this seems easy to do. For the sentence
(6) (Vx) (Poss(Ax) :JAx)
says that all individuals that possibly are A's actually are A's. Hence the set of
all A's in anyone alternative to the given (actual) world is a subset of the set
of actual A's. Since these alternatives encompass alllo~cally (extensionally)
possible worlds, each subset will show up in some alternative. Hence to
quantify over these alternatives is in effect to quantify over all the subsets of
the set of actual A's - or so it seems - and over them only. And such quanti-
fication over alternatives is precisely what modal operators do.
This line of thought has a gaping hole in it, however. It is seen by asking:
What are the values of the bound variable 'x' in (6)? Thl~ above argument
remains valid if it is assumed that all the individuals that can be members
of an alternative world are also members of the actual world. Merely allowing
these individuals to fail to exist in some worlds does not affect its validity,
for we can then change (6) to say that any individual that exists and is an A
is some alternative exists also in the actual world and is in it an A.
However, I have argued for independent reasons that the most natural
way of looking at (6) is to take 'x' to range over individuals which not only
exist in but can be cross-identified between the worlds as a member of which
x is being considered in (6). These worlds are obviously the actual one and
the relevant alternative. But to quantify over such cross-identifiable
individuals leaves it completely open as to what we want to say of individuals

which exist in an alternative WI to the actual world Wo but which cannot be
cross-identified between WI and Wo. Such individuals cannot be ruled out by
a fiat, and they may include plenty of A's. Hence in this kind of semantics
(6) does not enable us to consider all the subsets of the set of all the actual
A's.
Thus the introduction of L-semantics brings out a remarkable fact about
the interpretation of quantifiers in modal logics. Far from having only philo-
sophical motivation and philosophical implications, the problems of quanti-
fying in and cross-identification have remarkable technical repercussions as
well. On these problems apparently hangs the question whether standard
modal logics of the conventional sort reduce to nonstandard ones.
One can look at the relation of standard and nonstandard modal logics
from a slightly different viewpoint, viz. in terms of what the syntax of con-
ventional modal logics enables us to express. The basic reason why the differen-
ce between K-frames and L-frames does not seem to matter in much ordinary
modal logic (with a suitable interpretation of quantifiers) is that certain kinds
of dependence of the domain d I of individuals of an alternative world
(model) WI on the domain do of a world Wo to which it is an alternative
cannot be expressed in the language of conventional modal logic. We can say
in such a language that individuals existing in Wo also exist in some or
every WI, and that these individuals have (absolutely or at certain conditions)
certain properties in WI. However, we cannot express any converse
dependence. We cannot take some individual in some particular WI and go on
to say what it is like in the actual world. (The world of conventional modal
logic is like Thomas Wolfe's: In it, you cannot go back home again.) Hence
d I can for instance by any old superset of do. The effects of this one-sided
dependence seem to be quite subtle, and they sometimes eliminate some of
the relevant differences between K-frames and L-frames.
However, the situation is changed radically by the introduction of the so-
called backwards-looking operators. I 9 They enrich our syntax in precisely the
way needed here. What they accomplish is just to enable us to take an
individual in one of the worlds alternative to wo, say in WI , and to bring it
back to the actual world, that is, to say something about it, qua member of
Wo. For instance if 'D' is such an operator.
(7) Poss [(Vx)(Ax ::> DBx)]
says that there is at least one alternative world such that whatever is there A
is actually B. Likewise,
(8) Nec[(Vx)(Ax::> DAx]
says that the set of A's in any alternative is a subset of all the actual A's.
The latter example shows how, in the presence of backwards-looking opera-
tors L-semantics enables us to quantify over all the subdasses of a given
class (and over them only). Thus it is fairly clear that a modal logic with the
usual (linear) first-order quantifiers and backwards-looking operators is in
effect as strong as the whole second-order logic with standard interpretation.
If a detailed argument is needed, my 1955 reduction of higher-order logics to
a fragment of second-order logic 20 (both with the standard interpretation)
produces a reduct which is expressed without too much trouble in quantified
modal logic with backwards-looking, operators. The nerve of the translation
is to replace quantification over all subsets of a given set by a use of the
necessity-operator. And that can readily be accomplised by the means
envisaged. For instance, a standard second-order sentence
(9) (ttX) F (X)
where 'X' is a one-place predicate variable (ranging over predicates of indivi-
duls) and where F does not contain any higher-order quantifiers will translate
as something like the following:
(10) Nec F' (X)" Nec (ttx) [X(x)::> DX(x)] " C
where F' is like F except that all quantifiers have been relativized to X and
where C is a conjunction of sentences of the form
Nec(ttx) [Rxy=DRxy]
which say that the different predicates R in F obtain between exactly the
same individuals in the actual world and all of its alternatives. The second
conjunct of (10) says that the extension of X in the altematives is a subset
of its extension in the actual world. Since we are dealing with an L-Iogic, each
such subset is the interpretation of X in some altemative or other, and hence
the first conjunct of (10) ensures that F(X) is true for all values of X.
The transition from (9) to (10) or vice versa obviously preserves satisfia-
bility. Such a transition does not work when the quantifier is in a noninitial
position. What the reduction just mentioned 20 accomplishes is to bring
(sa/va satisfiability) every higher-order sentence to a form in which the
maneuver exemplified by the transition from (9) to (10) is possible. For its
upshot is in each case a formula which contains only one one-place bound
second-order variable (bound to an initial universial quantifiler) over and above
free predicates and which is satisfiable if the original formula is.
One may still have legitimate compunctions about the: intepretation of
quantifiers in (10). However, a simple change rectifies the situation. One of

the prima facie problems with (10) is that in the second conjunct x is appa-
rently being considered as a member both of a selected alternative world and
of the actual one. Hence 'x' apparently should be restricted to only such
values as can be reidentified in the actual world and the given alternative.
A slight modification of the second conjunct removes this defect, however.
We can write it as follows:
(11) Nec{Vx) [X(x) :>(3z)(z = x 1\ DX(z]
Likewise, each conjunct in C can be rewritten as follows:

Nec [{Vx){Vy)(Rxy :>(3z)(3u)(z =x 1\ u =Y 1\ DRzu]
{Vx){Vy)(Rxy :>(3z (3u)(z = x 1\ u = Y 1\ Nec Rzu)
It follows inter alia that a system of quantified (first-order) modal logic

with backwards-looking operators (call it B) with the L-interpretation cannot
be axiomatized. What its semantics is is obvious at once, however, on the basis
of what has been said.
The following general point is suggested by our observations. It has
sometimes been said or thought that the logic of logical modalities is
somehow more powerful or philosophically deeper than the usual first-order
logic. Historically speaking, such claims have often been formulated in terms
of the alleged superiority of intensional logic over extensional logic. Most
of the earlier treatments of modal logics have nevertheless failed to lead to
the deepest problems of logic. The main exceptions to his negative judge-
ment, for instance intuitionistic logics, have, in my opinion, moved much
more in the direction of epistemic concepts than in the realm of logical or
analytical modalities. Now we can see how these old claims for the special
significance of intensional logic can perhaps be partly vindicated. This
vindication is predicated on two separate ideas: (i) the standard interpreta-
tion of modal logic (i.e., L-semantics); (li) the idea of a backwards-looking
operator. Of these, (i) is essentially based on an extension of Henkin's dis-
tinction between standard and non-standard interpretation to modal logic,
and (ii) has gradually evolved from the work of Hans Kamp, David
Kaplan, Esa Saarinen, and others.
The relation of the distinction standard vs. nonstandard to the old con
trast of intensions with extensions is riddled with ambivalences, however.
There is a sense in which the standard interpretation of higher-order logics
turns on an extensional viewpoint. This sense is seen by observing that on the
standard interpretation predicate variables {higher-order variables) range
over all extensionally possible values. A vindication of the intensionalistic

position which depends essentially on the standard interpretation therefore
smacks of a Pyrrhic victory.
But if B-logics are essentially (Le., as far as their decision problems are con
cerned) equivalent to second-order logic (with the standard inte'rpretation),
what possible uses does their study promise? An obvious answer is that
studying them will be an important part of any attempt to clarify the concept
of logical necessity. Over and above this obvious purpose, there nevertheless
is another direction which appears to be worth a serious investigation. Vir-
tually all the most important set-theoretical problems are tantamount to the
question of the logical truth of some secondorder sentence. One difficulty
in trying to solve these problems is the difficulty of fmdin,g intuitions which
would somehow enable us to see which assumptions are acceptable and which
ones are not. It may be worthwhile to attempt to see whether a 'translation'
of second-order logic into a modal logic with backwards.looking operators
will sharpen our perception in this difficult field. I do not consider this very
likely, but in view of the great potential payoff I think an attempt should be
made. The best chance we have is probably to wed modal logic with back
wards-looking operators to game-theoretical semantics. Game-theoretical
intuitions have elsewhere yielded new ways of approaching strong set-
theoretical hypotheses. Maybe my B-logics will offer a framework for syste-
matizing some of this work.
An even more general point that can now be better appreciated is the po-
tential generality and importance of the standard-nonstandard distinction,
which is seen to permeate not only higher-order logics but first-order logic
and modal logics as well.
An important qualification to - or perhaps rather an ela1boration of - this
claim is nevertheless in order. I have been speaking as if there existed such a
thing as the nonstandard interpretation. Now the only sense in which this is
the case is so trivial as to be uninteresting. The only patently unique non-
standard interpretation is the minimal one, viz. the one in which literally
any subset S ~ P(l) of some power set P(l) is acceptable as the value range
of predicate variables or on which any set of models is acceptable as the set
of alternatives to a given model in a frame. This is in many ways a trivial
interpretation, which comes very close to reducing all highe:r-order logics and
modal logics to many-sorted first-order logic, at least insofar as purely logical
(deductive and model-theoretical in contradiction to pragmatical) aspects of
the situation are concerned. Usually some conditions, for instance closure
with respect to Boolean operations and projective operations, are imposed
on S ~ pel). For instance, in this case we obtain the usual higher-order logics
with nonstandard interpretation. In contrast to traditional higher-order logics,
in first-order logic a different condition on S is eminently natural. There
game-theoretical semantics strongly encourages us to limit S to recursive sets,
and mutatis mutandis for other higher-type entities, especially for functions
representing players' strategies. The motivation should be clear: how can any-
one expect to playa game using nonrecursive strategies?
Other restrictions on S (or on its counterparts of in modal logics) are of
course possible. The question as to what the natural, important, or otherwise
interesting restraints on S are is connected in many interesting ways with the
central problems in modem logic and the foundations of mathematics. I have
in this paper established connections betwe"en the different manifestations
of this question (of the choice of a nonstandard interpretation) in the realm
not only if higher-order logics but also of first-order logic and modal logics.
These connections may be hoped to be useful in throwing light on the choice
of the right interpretations.
One especially interesting further question is the following. In discussing
the distinction between K-frames and L-frames I have spoken apparently as
a matter of course of a fixed domain I of individuals shared by all members
of a frame Wo. (Admittedly, not all members of I have to exist in all the
members of Wo.) This assumption is not unproblematic, however, and has
in effect been challenged for philosophical reasons. What happens if the assump-
tion of a fixed I is given up? I don't know, but it is obvious that an inquiry
into this matter promises results which are relevant to the philosophical
controversies concerning 'prefabricated' individuals, 'possible individuals', and
similar matters. It is very likely that leaving I completely open results in inco-
herence. Ifit does, then philosophical arguments which one can easily give for
the freedom of the choice of I for different members of one and the same
frame tend to m"ake suspect the unlimited and unqualified notion of purely
logical modality.
In spite of the programmatic character of the last few remarks, it seems to
me that we have seen enough to appreciate the great interest of the generali-
zation of the standard-nonstandard distinction which has been outlined in
this paper.
A couple of complementary remarks may serve to throw this whole
complex of issues into a firmer perspective. The standard vs. nonstandard
distinction has recently played an interesting role in propositional modal
logics (and propositional tense logics). The reason is that in modal logics
a propositional variable is usually taken to range over all subsets of the frame.
This corresponds to the standard interpretation. For many reasons, it may be

more natural to let them range over some subset of the power set of the
frame. This step, which has been taken by S. K. Thomason and others,21
leads us to a counterpart of the nonstandard interpretation. The resulting
distinction is different from what is drawn in this paper, however, even
though it, too, serves to illustrate the great importance of the difference
between standard and nonstandard interpretations in Henkin's sense.
But do we need the distinction in the applications of modal logic and of
first-order logic? The following ingenious observation, which is due to Lauri
Carlson,22 illustrates the relevance of the distinction by means of a small
example. In many languages, for instance in English, there is a construction
which amounts to second~>rder quantification (over the subsets of a given
set). It is the following construction:
Some X's are such that - Y 1 - each one of them - Y 2 some one
ofthem-Y3 ,
A case in point could be
(12) Some fleas are such that for each of them we can find some one of
them who is smaller than it.
If the relation of being smaller than is asymmetric and transitive, the only
viable choices for 'some fleas' which make (12) true are infmite classes of
fleas.
A moment's thought convinces one that the construction (12) amounts to
second-order quantification, and my 1955 results suggest that the specific
kind of quantification involved in it is enough to capture the full complexity
of second-order logic, as far as questions of satisfiability are concerned.
From this fact it follows that the interpretation of second-order
quantifiers (standard vs. nonstandard) is relevant to the semantics or ordinary
English. This observation is independent of game-theoretical semantics, which
in its own way puts a premium on the standard-nonstandard distinction.
Another main door through which the standard-nonstandard distinction
enters recent applications of logic to linguistics is Montague semantics.23 Its
logic is higher-order intensional logic. Being higher-order logic, it is highly
sensitive to the difference between standard and nonstandard semantics. Yet
I cannot find any traces of awareness of the difference in the secondary
literature over and above those already mentioned (see notes 6 and 7 above).
Here is another important task for applied logicians. I doubt that the adequ-
acy of Montague semantics as a viable model of natural language can be asses-
sed without close attention to the standard-nonstandard distinction and to

the problem of choosing between different varieties of nonstandard interpre-
tation.
Flon'da State University
NOTES
I Stig Kanger, Provability in Logic. Stockholm Studies in Philosophy, Vol. I, Stock-

holm, 1957. Kanger deserves much more credit for developing a viable semantics for
modal logics than is given to him in the literature. For instance, the main novelty of such
semantics as compared with Carnap's old ideas is the use of the altemativeness relation.
(See below for an explanation of this concept.) Kanger introduced this idea in the
literature and used it in his work before anyone else, e.g., five years before Kripke.
1 See Saul Kripke, 'A Completeness Theorem in Modal Logic', J. Symbolic Logic 24
(1959), 1-4; 'Semantical Considerations on Modal Logic', Acta PhiJosophica Fennica

16 (1963), 83-94; 'Seman tical Analysis of Modal Logic 1', Zeitschrift fUr mathematische
Logik und Grundlagen der Mathematik 9 (1963),67-96; 'Semantical Analysis of Modal
Logic II', in I. W. Addison, L. Henkin, and A. Tarski (eds.), The Theory of Models, North-
Holland, Amsterdam, 1965, pp. 206-220; 'The Undecidability of Monadic Modal Quan-
tification Theory', Zeitschrift fiir mathematische Logik und Grundlagen der Mathematik
8 (1962), 113-116; 'Semantical Analysis of Intutionistic Logic', in 1. N. Crossley and
Michael Dummett (eds), Formal Systems and Recursive Functions, North-Holland,
Amsterdam, 1965, pp. 92-130.
3 Marcel Guillaume, 'Rapports entre calculs propositionels modaux et topologie impliques
par certaines extensions de la methode des tableaux: Systeme de Feys-von Wright',
Comptes rendus des seances de l'Academie des Science (Paris) 246 (1958), 1140-1142;
'Systeme S4 de Lewis', ibid., 2207-2210; 'Systeme S5 de Lewis', ibid.. 247 (1958),
1282-1283; laakko Hintikka, 'Quantifiers in Deontic Logic', Societas Scientariarum
Fennica. Commentationes humanarum litterarum, Vol. 23, 1957, No.4; 'Modality and
Quantification', Theoria 27 (1961), 119-128; 'The Modes of Modality', Acta
Philosophica Fennica 16 (1963),65-82.
4 See Richmond Thomason (ed.), Formal Philosophy: Selected Papers of Richard Mon-
tague, Yale University Press, New Haven, 1974, Chapters 1-2.
5 Op. cit., Chapters 3-8.
6 David Kaplan, UCLA dissertation, 1964.
7 Nino Cocchiarella; 'On the Primary and Secondary Semantics of Logical Necessity',
Journal of Philosophic Logic 4 (1975), 13-27; 'Logical Atomism and Modal Logic',
Philosophia 4 (1974),40-66.
3 Alfred Tarski and Bjami lonsson, 'Boolean Algebras with Operators I-II', American
Journal of Mathematics 73 (1951), and 74 (1952).

9 See Leon Henkin, 'Completeness in the Theory of Types', J. Symbolic Logic 15
(1950), 81-91. (please note that Peter Andrews has discovered a flaw in Henkin's
original argument and has shown how to repair it.)
10 See especially laakko Hintikka, 'Quantifiers in Logic and Quantifiers in Natural

Languages', in S. Komer (ed.), Philosophy of Logic, BlackweU's, Oxford, 1976,
pp.208-232; 'Quantifiers vs. Quantification Theory', LinguistiC Inquiry 5 (1974),
153-177; Logic, Language-Games, and Information, Clarendon Press, Oxford, 1973.
Much of the relevant literature has now been collected in Esa Saarinen (ed.), Game-
Theoretical Semantics, D. Reidel, Dordrecht, 1978.
I I Cf. here laakko Hintikka and Veikko Rantala, 'A New Approach to In finitary
Languages', Annals of Mathematical Logic 10 (1976), 95-115.

12 For an explicit discussion of this idea see laakko Hintikka and Lauri Carlson, 'Con-
ditionals, Generic Quantifiers, and Other Applications of Subgames', in A. Margalit

(ed.), Meaning and Use, D. Reidel, Dordrecht, 1978.
J3 As Dana Scott has pointed out in an unpublished note, one can in this way also
obtain GOdel's functional interpretation of first-order logic and arithmetic. See Kurt
Godel, 'Eine bisber noch nicht beniitzte Erweiterung des finiten Standpunktes', in
Logica: Studiiz Paul Bemays Dedicata, Editions du Griffon, Neuchatel, 1959, pp. 76-83_
14 Cf. my paper 'Language Games', in Essays on Wittgenstein in Honour of G. H. von
Wright (Acta Philosophica Fennica, Vol. 28, Nos. 1-3), North-Holland, Amsterdam,
1976, pp_ 105-125.
15 'Quantifiers vs. Quantification Theory', Linguistic Inquiry 5 (1974), 153 -177.
16 Of course this is not the only nor the most natural way of imposing the distinction
on second-order logic.
17 See David Kaplan, op. cit. (note 6 above).
18 I have discussed these issues in the essays collected in Models for Modalities,
D. Reidel, Dordrecht, 1969, and The Intentions of Intentionality and Other New Models
for Modalities, D. Reidel, Dordrecht, 1975.
19 For backwards-looking operators, see Esa Saarinen, 'Backwards.-Looking Operators
in Tense Logic and Natural Language', in laakko Hintikka et. Gil (eds.), Essays on
Mathematical and Philosophical Logic, D. Reidel, Dordrecht, 1978, pp. 341-367; and
Esa Saarinen, 'Intentional Identity Interpreted', Linguistic and Philosophy 2 (1978),
151-223, with further references to the literature. The initiators of the diesa seem to
have been Hans Kamp and David Kaplan.
20 'Reductions in the Theory of Types', Acta Philosophica Fennica 8 (1955), 56-115.
21 See S. K. Thomason, 'Semantic Analysis of Tense Logics', J. Symbolic Logic 37
(1972), 150-158; 'Noncompactness in Propositional Modal Logic', ibid., 716-720; 'An

Incompleteness Theorem in Modal Logic', Theoria 40 (1974), 30-34.
22 Personal communcation.
23 Barbara Hall Partee (ed.), Montague Grammar, Academic Press, New York, 1976,
and note 4 above.
8
THE SKELETON IN FREGE'S CUPBOARD:

THE STANDARD VERSUS NONSTANDARD DISTINCTION*
A nyone who uses higher-order logic faces a momentous

choice concerning the interpretation of higher-order quan-
tifiers. Let us consider as an example a second-order quan-
tifier that involves a one-place class or predicate variable, say, X. Its
values can be taken to be either classes of individuals or properties
(concepts) of individuals. In either case, the same dilemma con-
fronts a higher-order logician, even if its horns look different on the
two interpretations.
If the values of X are thought of as being classes, then the crucial
question is whether the range of a quantifier containing X is the
entire power set P(do(M of the relevant domain do(M) of individ-
uals, or only some designated subset of P(do(M. In other words,
the question is whether the values of X are arbitrary extensionally
possible classes or whether only some such classes are accepted as
values of X. The former alternative results in what is usually called
the standard interpretation of higher-order logic, the latter in a
A nonstandard interpretation can be of many different kinds. The
relevant range of X is sometimes thought of as a part of the specifica-
tion of the model M in which a higher-order formula is being inter-
preted, sometimes as a restraint on the interpretation. Often, the
range of X and the ranges of other higher-order variables cannot be
selected completely arbitrary, but are subject to certain closure con-
ditions. For instance, the totality of these different ranges is as-
sumed to be closed with respect to Boolean operations and to pro-
jective ones; in other words, the usual formation rules are assumed
to preserve interpretability.1
The same distinction has to be made for variables of each logical
type. An especially important case in point is the interpretation of
function variables, as witnessed by the fact that a specific nonstan-
dard interpretation is sometimes imposed on them on purpose. The
I Leon Henkin, "Completeness in the Theory of Types," The Journal of Sym-
bolic Logic, xv (1950): 81-91. For an important correction to Henkin's paper,

see Peter B. Andrews, "General Models and Extensionality," The Journal of
Symbolic Logic, XXXVII (1972): 395-7.

THE SKELETON IN FREGE'S CUPBOARD 145
best known essay in this direction is probably Kurt Godel's2 func-

tional interpretation of first-order logic and arithmetic. There the
values of function variables are restricted to recursive functions of
the appropriate type. Notwithstanding the variety of different types
of entities and their intricate interrelations, the apparently special
case of one single second-order quantifier with a onc:~-place predi-
cate or class variable is fully representative of the entire theory of
finite types. For there are results that show that, if the standard
interpretation is granted to one single second-order one-place
quantifier, the entire higher-order logic (theory of finite types) with
the standard interpretation can be reconstructed by it.s means. 3 Of
course, by the same token the interpretation of a single-function
variable determines in a sense the interpretation of the entire
higher-order logic, for classes can always be handled by means of
their characteristic functions.
Philosophers have not paid adequate attention to the standard
versus nonstandard distinction. The main. reason is undoubtedly
their slowness to appreciate the importance of the model-theoretical
viewpoint in logical theory." Be this as it may, the contrast was
spelled out explicitly by Leon Henkin in 1950. The fact that he
considered only one nonstandard interpretation does not detract
from his achievement.
I. CONCEPTS VERSUS THEIR EXTENSIONS
The interpretational dilemma cannot be avoided by switching to
variables that range over properties or relations in contradistinction
to their extensions. Then the question whether the standard inter-
pretation is assumed becomes the question as to whether for each
class (potential extension), C, there is a property that has C as its
extension or, in Gottlob Frege's term, as its course-of-values.
This distinction is extended as a matter of course to variables of
other higher-order types. As a bargain, we are put into the position
of being able to distinguish the main problem addressed in this
paper from two others. In this paper, we are asking whether Frege
assumed that for each potential extension there is a concept that has
this extension as its value range. One can also ask the mirror-image
question: Is there an extension for each concept? It turns out to be
2 "On a Hitherto Unexploited Extension of the Finitary Standpoint," Journal

of Philosophical Logic, IX (1980): 133-42.
3 For such results, see Hintikka, "Reductions in the Theory of Types," Acta
Philosophica Fennica, v (1955): 59-115.
4 This is documented in Hintikka, "On the Development of the Model-Theo-
retic Viewpoint in Logical Theory," Synthese, LXXVII (1988): 1-36.
extremely important to separate the two questions from each other,

both in discussing Frege and in discussing other logicians and mathe-
maticians (cf. below).
We can also formulate a third question: Are two concepts with the
same extension identical? This question is often combined with the
first two and formulated as the query: Is logic extensional? It turns
out, however, that often the philosophers who ask this question
really have in mind only one of the three questions. It is also impor-
tant to realize that an "extensionalist" answer to the question con-
cerning the identity conditions of functions (including concepts),
which is Frege's answer, by no means prejudices his answer to either
of the other two questions.
II. THE STANDARD VERSUS NONSTANDARD DISTINCTION
AND THE EXPRESSIBILITY OF CONCEPTS IN LANGUAGE
Sometimes, a nonstandard interpretation is guided by the idea that
only such properties, relations, and functions can be assumed to
exist as can be defined or otherwise captured by a suitable expres-
sion of one's language. In the case of theories with infinite models,
this leads inevitably to a nonstandard interpretation, for there can
be only a countable number of such definitions or characterizations
available for this purpose. Hence they cannot capture all the subsets
of do(M), for there is an uncountable number of them.
This particular nonstandard interpretation was the only one con-
sidered by Henkin. More generally, it seems to be the only one that
is considered in most of the literature where the Henkin distinction
is relied on or discussed. It is nevertheless important to realize that
other nonstandard interpretations are possible. What is crucial is to
realize that the nonstandard versus standard distinction is not the
same as the distinction between, on the one hand, variables each of
whose values always has a representative in language which can serve
as a substitution value for the variable and, on the other, variables
whose values can exist independently of their linguistic representa-
tions. The latter issue can be called the problem of nonlinguistic
existence (of values of higher-order variables). Historically, it may
have been the source of the standard versus nonstandard distinc-
tion, but, systematically, it is independent of this distinction. In gen-
eral, representability in a language on the basis of a theory depends
on the language and on the theory, while the choice between the
standard and a nonstandard interpretation faces us in using any
higher-order language.
A confusion between the different questions we have just distin-
gUished is not merely an abstract possibility. The fallacy of not sepa-
rating them has been committed even by knowledgeable philo-

sophers. For instance, in an article in Erkenntnis, Rudolf Carnap 5
criticizes Frank Ramsey's reliance on the standard interpretation
and writes:
I think we should not let ourselves be seduced by it into accepting
Ramsey's basic premise, viz., that the totalities of properties already
exist before their characterization by definition ...
I think we ought to hold fast to Frege's dictum that in mathematics
only that may be taken to exist whose existence has been !proved (and
he meant proved in finitely many steps) (ibid., p. 102).
Here Carnap is clearly misinterpreting Ramsey's position. In so far

as we are dealing with what Ramsey called predicates and relations
"in the ordinary sense," Ramsey apparently accepts the idea that
there has to be an expression for them. In this sense, Ramsey in fact
discusses how functions are constructed by constructing expressions
for them. Ramsey's point is that our entire logic has to be reinter-
preted so as to deal, "not with relations in ordinary sense, but with
possible correlations, or 'relations in extension' as Mr. Russell calls
them," and likewise for entities of other logical types. In fact, Car-
nap seems to misinterpret Frege's position, too, or at least he gives a
wrong description of it, as we shall see.
A terminological point is in order here. By calling one of the two
contrasting interpretations "standard," we are not passing any judg-
ment as to which view ought to be adopted or which one is histori-
cally the usual one. We are merely following the terminology of
Henkin, who was the first logician to articulate fully the standard
versus nonstandard distinction.
III. THE STANDARD VERSUS NONSTANDARD DISTINCTION
AND SET THEORY
On the basis of what has been said, it might seem that the standard
interpretation has set theory as its main fortress. In the first section
of this paper, we explained what the standard interpretation is by
means of the familiar set-theoretical notion of the power set. It is
undoubtly true that Georg Cantor intuitively assumed some sort of
standard interpretation in his thinking. It can even be argued, as
Michael Hallett 6 has aptly done, that the actual generated principles
for transfinite ordinals go "most of the way toward acceptance of
the existence of arbitrary sub-domains" and therefore toward the
standard interpretation. As Hallett writes:
5 "Die logizistische Grundlegung der Mathematik," Erkenntnis, II (1931): 91-
105.
6 Cantorian Set Theory and Limitation of Size (New York: Oxford, 1990).
Cantor asserts that the two principles of generation [he had formu-
lated] 'give us the ability to break through every barrier in the forma-
tion of real, whole numbers' (the italics are Cantor's). This certainly
suggests a belief that these methods take the ordinal number sequence
arbitrarily far (ibid., p. 58).
But the real question here is whether later set theorists were able to
implement Cantor's intuitive ideas so as to turn what Hallett guard-
edly refers to as "most of the way" into all the way to Cantor's
implicitly "standard" way of thinking about sets and their existence.
The right answer to this question is a resounding no.
One way of looking at contemporary axiomatic set theory is to
view it as an attempt to deal with such higher-order entities as sets
on the first-order level. What inevitably gets lost in such an attempt
is precisely the possibility of a standard interpretation of variables
ranging over higher-order entities. This problem manifests itself,
among other things, in the form of the Skolem paradox. Even when
you seem to be able to prove the existence of all sorts of uncount-
able sets in your axiom system, it turns out that there are, according
to the downward Lowenheim-Skolem theorem, models of the same
axiom system in which those very sets are countable. Indeed, the
very paradox about the Lowenheim-Skolem theorem lies in the fact
that what in the system is called countability or uncountability is that
only on the standard interpretation of set variables understood as
higher-order ones.
It does not automatically help, either, to move to a higher-order
level: higher-order logics can typically be treated as if they were
many-sorted first-order logics as long as the standard interpreta-
tion is not assumed. Thus, in a sense it is the standard versus non-
standard interpretation that constitutes the real watershed between
first-order logic and higher-order logics. At least this distinction is
much more consequential than a mere distinction between types
(order levels). Since Frege's logic was a higher-order one, we are
already beginning to see that the standard versus nonstandard dis-
tinction puts his entire project into a sharp new light.
IV. THE IMPORTANCE OF THE DISTINCTION
We called the standard versus nonstandard distinction momentous.
This was in reality an understatement. Each horn of the dilemma
turns out to be very sharp, in that the prospects of higher-order
logic and of its applications depend essentially on one's choice be-
tween the two kinds of interpretations, in more than one way.
For one important thing, second-order logic with a standard in-
terpretation is easily seen to be semantically incomplete. In other
words, the set of valid formulas is not recursively enumerable on this

interpretation. There cannot exist a finite (or recursively enumer-
able) set of logical axioms and inference rules that would generate
all and only valid formulas as theorems. In contrast, Henkin showed
that, on one especially natural nonstandard interpreta.tion, higher-
order logic does after all admit of a semantically complete axiomati-
zation (op. cit., p. 81). Many other kinds of nonstandard interpreta-
tion likewise yield axiomatizable logics. 7 This does not mean that we
should prefer a nonstandard interpretation. For only the standard
interpretation of second-order logic enables us to use it for the most
important purposes it can serve in the foundations of mathematics.
If we assume the standard interpretation, we can easily formulate
descriptively complete and indeed categorical axiomatizations for
such crucial mathematical theories as number theory and the theory
of real numbers. s Likewise, the principle of mathematical induction
is easily formulated in a second-order logic with standard interpre-
tation, but cannot be formulated in an axiomatizable nonstandard
higher-order logic. 9 Thus, for most serious foundational purposes
we have to assume the standard interpretation of second-order
logic. 10
V. THE ROLE OF THE DISTINCTION IN FOUNDATIONAL DISCUSSIONS
Even though the standard versus nonstandard distinction has not
attracted much attention from philosophers, it has played an impor-
tant, albeit only partially articulated, role in the actual foundational
discussions. For instance, the standard versus the nonstandard con-
trast is the most important determinant of a mathematician's atti-
tude to the axiom of choice. On the standard interpretation, the
7 See also Hintikka, "Standard vs. Nonstandard Logic: Higher Order, Modal
and First-Order Logics," in Modern Logic: A Survey, E. Agazzi, ed. (Boston:
Reidel, 1980), pp. 283-96.
8 Henkin, p. 89. See also Steward Shapiro, Fotmdations without Foundation-
alism (New York: Oxford, 1991).
9 For this point, see Shapiro; and also Thoralf Skolem, "Uber einige Crund-
lagenfragen der Mathematik," Shrifter utgitt av det Norske Videnskaps-Aka-
demi, 1,4 (1929): 1-49.
10 The whole business ofaxiomatizability is nevertheless extremely subtle. Per-
haps the most interesting nonstandard interpretation of first-order logic is ob-
tained by restricting Skolem functions to recursive ones, somewhat as in Codel's
functional interpretation, but without any unusual interpretation of negation and
conditionals. Then the resulting logic is nonaxiomatizable, as was first pointed out
by Robert Vaught in "Sentences True in All Constructible Models," The Journal
of Symbolic Logic, xxv (1960): 39-53. The Peano axiomatization of elementary
arithmetic is then complete, however, as was first pointed out by TenneOOaum.
For this later point, see Richard Kaye, Models of Peano Arithmelic (New York:
Oxford, 1991), esp. pp. 153-7.
axiom of choice is trivially acceptable, whereas for an advocate of a

nonstandard interpretation, there is prima facie little reason to be-
lieve in the axiom of choice. This important background of the ca-
reer of the axiom of choice is almost universally neglected. II
Likewise, the standard versus nonstandard distinction prejudices
one's judgment of Bertrand Russell's axiom of reducibility, which
roughly says that for each impredicatively defined set there exists an
equivalent predicatively definable class. For a believer in the stan-
dard interpretation, a class is there, determined by its members, no
matter how it is defined in some one language. Hence, for the theo-
rist of the standard sort, the "axiom of reducibility" is trivially true,
indeed redundant, as was forcefully pointed out by Ramsey.12 In
general, Ramsey's revision of Russell's ramified theory of types is to
all intents and purposes a self-conscious step from a nonstandard to
a standard higher-order logic.
VI. STANDAR,D INTERPRETATION
AND THE IDEA OF ARBITRARY FUNCTION
We can say that there were two long-range developments of mathe-
maticians' thinking about the foundations of their own field which
led to the idea of standard interpretation:
(i) The development of the idea of an arbitrary function.

(ii) The development of the idea of arbitrary set.
Thinking about (i) and (ii) was conditioned not so much by philo-
sophical questions as by the demands of mathematical theorizing
itself. Let us tum first to the notion of arbitrary function.
The development of the idea of an arbitrary function is one of the
most important conceptual developments in mathematics. This de-
velopment has not always been fully appreciated. For instance, the
concept of arbitrary function is often credited to P. G. L. Dirichlet,
even though in reality it was formulated as early as 1755 by L. Euler.
In his Institutiones calculi differentialis, Euler wrote:
If some quantities so depend on other quantities that is 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
II This is the case even with Gregory Moore's important book, Zermelo's Ax-
iom of Choice: Its Origins, Development and Influence (New York: Springer,
1982).
12 "The Foundations of Mathematics," Proceedings of the London Mathemati-
cal Society, II, 25, (1925): 338-84.
could be determined by others. If, therefore, x denotes a variable

quantity, then all quantities which depend upon x in any way or are
determined by it are called functions of it. U
Similar quite general characterizations of a function are found in

other mathematicians, for instance in N. I. Lobachevsky. According
to I. H. Anellis,14 Lobachevsky thought that "general thought de-
mands that a function [f] of x be called a number which is given for
each x and which changes ... together with x." The case of Loba-
chevsky is interesting because the development of non-Euclidean
geometries was one of the factors that prompted mathematicians to
consider a wider class of structures than before: for instance, be-
cause not all non-Euclidean geometries can be modeled within the
Euclidian one.
In Euler's time, the concept of arbitrary function had little use in
state-of-the-art mathematical research. Actual work both in the
foundations of mathematics and at the cutting edge of the subject
gradually raised the consciousness of most mathemat.icians in the
course of the nineteenth century. The innocent looking Eulerian
definition of an arbitrary function turned out to admit of surpris-
ingly unruly denizens of the world of mathematics. For instance,
many of the results that had been taken more or less for granted
turned out to hold only for functions with specifiable properties.
For instance, it turned out that even a continuous function need not
be differentiable. This development enhanced, on the other hand,
the interest of such results as could be proved for arbitrary func-
tions or at least for a large class of functions. Important examples
are provided by functions that can be represented by means of
Fourier series and by functions that are integrable, on one of the
various extended definitions of integration.
As a consequence of such developments, a lively interest arose in
the concept of function and in the question as to how wide a class of
functions (possibly including all arbitrary functions) mathematicians
should be concerned with. The acceptance or rejection of the con-
U This passage from lnstitutiones calculi differentialis is translated by A. P.

Yushevich: "The Concept of Function up to the Middle of the 191,h," Arch. Hist.
Exact. Sci., XVI (1970), p. 70. For additional information on the development of
the idea of an arbitrary function, see also Umberto Bouazzini, The Higher Cal-
culus: A History of Real and Complex Analysis from Euler to Weierstrass,
(New York: Springer, 1986).
14 A History of Mathematical Logic 'in Russia and the Soviet Union, unpub-
lished.
cept of arbitrary function was, for such reasons, an important water-

shed in the foundational discussions in the nineteenth century.
The tension generated by this question is in evidence in several
leading mathematicians. For instance, K. T. Weierstrass considered
the idea of arbitrary function, but rejected it as being so vague and
so general as not to enable us to find any interesting things to say
about it as such. The other side of the Weierstrassian idea of func-
tion, however, was his desire to consider as large a class of functions
as possible, to find "the largest class of functions for which one can
give an analytic representation and which can most fully satisfy the
needs of analysis."15
L. Kronecker, in contrast, rejected clearly the idea of an arbitrary
function. As Harold M. Edwards 16 puts it, "what is missing, and
what is excluded by his [Kronecker'S] principles is integrals of 'arbi-
trary' functions or sums of 'arbitrary' infinite series" (ibid., p. 74).
All these developments are highly relevant here, for the concep-
tion of the standard interpretation of higher-order quantifiers (and
variables) is, to all purposes, equivalent with the notion of an arbi-
trary function or the notion of an arbitrary set, as was indicated in
section I above. For the same reason it is not anachronistic to relate
foundational discussions of the late nineteenth century to the stan-
dard versus nonstandard distinction, even though it entered into
actual discussion disguised by a different terminology.
VII. PRIMACY OF CONCEPTS OVER THEIR EXTENSIONS IN FREGE
After all these preparations, we can turn to the question whether
Frege assumed the standard interpretation of higher-order quanti-
fiers or a nonstandard one. This question is especially poignant
when applied to Frege, for his logic contains explicitly formulated
higher-order quantifiers. We shall show that Frege lacked both the
idea of arbitrary function and the idea of arbitrary set, and hence in
effect opted for a nonstandard interpretation of higher-order quan-
tifiers. Let us deal first with the latter question.
For Frege, classes as classes come into play only through being
Wertverliiufe of functions (in Frege's wide sense of the term 'func-
tion'). He distinguished very clearly a concept, denoted by a concept
word '~(~)' from its extension, denoted by a course-of-values-name
15 See P. Dugac, "Elements d'analyse de Karl Weierstrass," Archive for History

of Exact Sciences, X (1973): 41-176.
16 "Kronecker's view on the Foundations of Mathematics," in The History of
Modern Mathematics, Vol. I: Ideas and Their Reception, David E. Rowe and
John McCleary, eds. (San Diego: Academic, 1989).
'E~(E)'Y As is well-known, one basic difference in Frege between a

concept and the extension corresponding to it is that the latter is an
object whereas the former is not. A concept is a function whose
value is always a truth value. The extension of the concept, or alter-
natively the course of values of the function which the concept is, is
an object that is determined by the arguments and the correspond-
ing values of the function for these arguments. It is important to
keep in mind that both concepts and courses of values are exten-
sional (ibid., pp. XXXIX, 36).
The relation between a concept and its course of values is, con-
trary to what one might think, logical and not empirical. That is, the
extension of a concept can only be apprehended by our logical facul-
ties starting out from the concept. 18 Accordingly, a course of values
does not exist independently, but only tied to a concept (function)
which is logically more primary than it:
... In fact I do hold that the concept is logically prior to its extension,
and I regard as futile the attempt to base the extension of a concept or
a class not on the concept but on individual things. 19
Consequently, we have no means of recognizing a course of values
except through a concept:
We have only a means of always recognizing a course-of-values if it is
designated by a name like "E4>(E)", by which it is always recognizable as
a course-of-values. But we can neither decide, so far, whether an object
is a course-of-values that is not given us as such, and to what function it
may correspond, nor decide in general whether a given course-of-
values has a given property unless we know that this property is con-
nected with a property of the corresponding function. 2o
It has been argued by Tyler Burge 21 that, in framing his conception
of class, Frege was the member of two traditions. The first was the
intensional tradition originated by Leibniz and continued by J. S.
Mill, G. Boole, and E. Schroder. Roughly, the main conception em-
17 Frege, The Basic Laws of Arithmetic: Exposition of the S,Ystem, Montgo-

mery Furth, trans. (Berkeley: California UP, 1967), p. 36.
18 Frege, Posthumous Writings, H. Hermes, ed., P. Long and R. White, trans.
(Cambridge: Blackwell, 1979), p. 181.
19 Frege, "A Critical Elucidation of Some Points in E. Schroeder's Algebra der
Logik," in Translations from the Philosophical Writings of Gottlob Frege, P.
Geatch and M. Black, eds. (New York: Oxford, 1952), pp. 86-106, esp. p. 106.
20 Frege, The Basic Laws of Arithmetic, p. 46.
21 "Frege on Extensions of Concepts, From 1884 to 1903," Philosophical He-
view, LXIII (1984): 1-34.
bodied in this tradition was that intension, in the form of properties,

characteristic marks, etc., determine extension. The second tradi-
tion was the extensional one represented, among others, by Cantor,
Richard Dedekind, and others. The main feature of this approach
was that sets and numbers are not defined by appeal to properties,
concepts, etc., as in the intensional tradition, but by "abstraction"
(cf. section x below).
In his definition of a class, Frege adopted the former conception
and he continuously criticized the extensionalists' idea that a class is
formed by grouping together elements. It is important to realize
that this antiextensionalism is compatible with Frege's view that the
identity conditions for concepts are extensional. Frege's "intension-
alism" (i.e., his view that concepts are logically prior to their exten-
sions) can be seen in different ways. For instance, in his letter of July
24, 1902, Russell wrote Frege that he had difficulties "to see what a
class really is if it does not really consist of objects but is nevertheless
supposed to be the same for two concepts with the same ex-
tension. "22
Frege's answer is contained in his letter to Russell of July, 28,
1902. It is worth spelling out in some detail. He distinguished
wholes (systems) from classes, using the following examples, among
others:
(1) The Romans conquered Gaul.
(2) The class of prime numbers comprises infinitely many objects.
In the first case, according to Frege, the Romans form a whole, that
is, a physical body held together by relations that are essential to it,
such as institutions, customs, laws. It is characteristic of wholes,
Frege writes, that the parts of any part of a whole are again parts of
the whole. Thus, the parts of the regiments are the battalions, but
also the companies and the soldiers.
This is not so for classes, Frege avers. A class is determined by its
members. The only members of the class of prime numbers are the
prime numbers. The members of a class of regiments include only
the regiments, not any individual soldiers. A class is a logical object,
not a physical object like a whole (system). Asked how we apprehend
classes, Frege gave the following unambiguous answer:
We apprehend them as extensions of concepts, or more generally, as
ranges of values of functions (ibid., p. 141).
22 Frege, Philosophical and Mathematical Correspondence, Gottfried Gabriel

et a!., eds. (Cambridge: Blackwell, 1980), p. 139.
(Here and below, "ranges of values of functions" is synonymous

with "courses of values of functions.")
This is not supposed to be interpreted as the priority of concepts
only in explaining or understanding what classes are, but as an onto-
logical principle as well:
... for me the extension of a concept or a class is only a special case of
a range of values (ibid., p. 142).
Thus, a class which is not the extension of some concept is a contra-
diction in terms within the Fregean system, since every class has to
be introduced via a concept. A class cannot be given through its
elements, only through a rule that determines elementhood. 211
Now, there were at least two different reasons for Frege to reject
the individuation of classes through their members. The first one is
connected with the definition of the empty class:
Of course, one must not then regard a class as constituted by the
objects (individual, entities), that belong to it; for removing the objects
one would then also be removing the class constituted by them. In-
stead, one must regard the class as constituted by the characteristic
marks, i.e., the properties which an object must have if it is to belong to
it. It can then happen that these properties contradict one another, or
that there occurs no object that combines them in itself. The class is
then empty but without being logically objectionable for that reason.24
The second reason has to do with the individuation of infinite
classes. Frege thought that an infinite class cannot be given solely by
its members because of the finiteness of the human intellect. In-
stead, such a class can be justified only as derived from a concept
involved in thought. 25
VIII. FREGE LACKED THE NOTION OF ARBITRARY SET
The intensionalists' and, in particular, Frege's conception of set ex-
plained above was criticized by the so-called extensionalists, that is,
set theorists like Cantor and Dedekind who shared with each other a
different concept of set. Their main criticism against Frege was that
he allowed an extension for every concept, which led to Russell's
paradox. For instance, in his review of Frege (1884), Cantor26 criti-
23 See Frege, Posthumous Writings, p. 34; and The Foundations of Arithme-

tic: A Logico-Mathematical Inquiry into the Concept of a Number, J. L. Austin,
trans. (Cambridge: Blackwell, 1960).
24 See the undated letter to Peano, in Frege, Correspondence, p. 109.
25 Posthumous Writings, p. 34; and Foundations, p. 98. '
26 "Rezension der Schrift von G. Frege Die Grundlagen der Arithmetik," in
Gesammelte Abhandlungen mathematischen und philosophischen Inhalts, E.
Zermelo, ed. (Berlin: Springer, 1932).
cized Frege for defining classes as extensions of concepts. He

pointed out that, in general, "the 'extension of a concept' is some-
thing quantitatively completely undetermined" and for this reason
not all extensions of concepts are sets" (ibid., pp. 440-1).
In the same spirit, E. Zermel027 wrote:
... it follows further as already appears in a much simpler way from

Russell's antinomy, to be sure that it is not permissible to treat the
extension of every arbitrary notion as a set and therefore that the
customary definition of a set is too wide (ibid., p. 107).
Even if Frege's definition of a set was too wide in one direction, in

another direction it was too narrow. The inseparable Fregean link-
age between sets and their defining properties prevented him from
generating new sets by operations acting solely on the members of
the already existing sets. This fact is correctly emphasized by Hao
Wang28 when he writes:
He [Frege] identifies sets with their extensions and treats them as indi-
viduals (on the same level with individuals). This suggests immediately
the idea of a type hierarchy of extensions, since extensions of predi-
cates seem to be more closely related to predicates than to individuals
.... But such a conception has little that is positive to say about how
sets are to be generated (ibid., p. 210).
Frege was thus unable to operate with arbitrary extensions as Can-

tor and Dedekind did. What is more, he tied extensions so closely to
concepts that he could not reach the modern (generative) concept
of set which or more or less completely implemented by the axioma-
tized set theories of Zermelo, J. von Neumann, and others. Al-
though in these theories strong limitations on the formation of sets
still exist (as compared with Cantor's system), still one has some
liberty, starting with a given set, to form new sets. Thus, one basic
ingredient of the standard interpretation, i.e., the notion of arbi-
trary set, is lacking from Frege's system.
One might still claim that, although Frege could not form arbi-
trary sets merely by applying certain operations on the members of
an already existing set, there is nothing in Frege's theory which
prevents the arbitrary sets to exist as value ranges of some suitable
concepts. Hence, for Frege, the question whether the standard in-
27 "Neuer Beweis fUr die Moglichkeit einer Wohlordnung," Mathematische

Annalen, LXV (1908): 107-28.
28 From Mathematics to Philosophy (New York: Routledge, 1974).
terpretation is correct becomes simply and solely the question

whether there is a concept (function) for each potential extension,
that is, whether for each class (class-in-extension) there is a concept
whose course of values this class is. Thus, the answer to the question
"Which of the two interpretations Frege was committed to?" de-
pends crucially on Frege's conception of function.
IX. THE OBJECTIVE EXISTENCE OF FUNCTIONS
DOES NOT IMPLY STANDARDNESS
Michael Dummett 29 claims unambigously, not only that. Frege oper-
ated with higher-order quantification, but that for Frege the range
of a function quantifier included all "arbitrary functions":
For Frege had not the slightest qualm about the legitimacy or intelligi-
bility of higher-level quantification: he used it from the first, in Be-
griffsschrift, freely and without apology, and did not eVlen see first-
order logic as constituting a fragment having any special significance
(ibid., p. 223).
And it is true enough, in a sense, that, once we know what objects there
are, then we also know what functions there are, at least, so long as we
are prepared, as Frege was, to admit all 'arbitrary functions' defined
over all objects (ibid., p. 177).
That Frege can give no consistent justification for interpreting second-

level quantification as varying over the full impredicative totalities of
first-level concepts, relations and functions is, of course, no. ground for
denying him the right to employ second-level quantification at all
(ibid., p. 541).
Dummett does not provide any references to Frege's texts in sup-

port of his claims. It is nevertheless not difficult to see what he has
in mind. On page 177, Dummett is expounding Frege's idea of the
objective existence of functions, his "conception of the function as
belonging to the 'realm of reference'." He points out that "for
Frege the only possible route to grasp what a function is is via an
understanding of the working of functional expressions." But this
does not mean, Dummett in effect avers, that functions according to
Frege are merely counterparts to functional expressions, that they
are not inhabitants of the real universe. Dummett is undoubtly
correct here. But what he is doing is merely a way of asserting the
objective, nonlinguistic existence of functions independently of lan-
guage. They are, in Dummett's words, mappings of ot~ects on ob-
29 Frege's Philosophy of Language (London: Duckworth, 1973).

jects. Hence Dummett's real point is that Frege maintains the non-
linguistic existence of functions. Nothing follows from this point
concerning Frege's alleged recognition of arbitrary function, i.e., of
any possible correlation of objects and arguments. Dummett's move
from the mode of existence of functions as objective correlations of
objects to other objects to the claim that there was such a correla-
tion or mapping for each possible pairing of argument values and
function values is a non sequitur. He is apparently assuming that
there is only one possible nonstandard interpretation, the one Hen-
kin later considered. For only on this assumption could Dummett
infer from Frege's rejection of this particular nonstandard interpre-
tation that Frege accepted the standard one. At best, Dummett
manages to refute Carnap's view (see note 5 above) that, according
to Frege, only provably existing functions really exist.
The other quotes from Frege which Dummett uses are not any
more convincing. On the contrary, on page 541, Dummett even
admits that Frege's "theory of reference cannot of itself, provide
any temptation to violate the vicious-circle principle," that is, to
assume the existence of impredicative totalities. But the very point
of the standard interpretation is to assert the existence of every
possible class independently of its characterization in language.
Hence the standard interpretation would in fact provide a tempta-
tion of the very kind Dummett denied Frege. 30
If Dummett credited Frege with the notion of arbitrary function,
Nino Cocchiarella31 credits him with no more and no less than the
formulation of standard second-order logic:
... both Frege and Russell maintain that the logistic framework within
which classical mathematics is to be represented must consist at least of
a second order predicate logic where quantification is not only with
respect to the role of singular terms, but to that of predicates as well.
Indeed, it was Frege himself who first formulated and developed stan-
dard second order predicate logic, and he did so precisely as a frame-
work within which classical mathematics was to be reduced to logic
(ibid., p. 198).
No reference to Frege's texts is provided by Cocchiarella.
30 In fact, the only axioms of set theory which give us sufficient liberty in form-
ing more comprehensive sets are the axiom of subsets and the power-set axiom,
which. as is well-known. involve impredicativities.
31 "Frege, Russell and Logicism: A Logical Reconstruction," in Frege Synthe-
sized, L. Haaparanta and J. Hintikka, eds. (Boston: Reidel, 1986). pp. 197-252.
It is true that Frege used higher-order logic as a framework for

reducing mathematics to logic, and that this can be done only by
giving higher-order logic the standard interpretation. But there is
no evidence in Cocchiarella or anywhere else that Frege actually
perceived the need of the standard interpretation for this purpose
or actually assumed it.
x. FREGE'S CRITICISMS OF "ABSTRACTION"
What made possible the conception of an arbitrary set was the grad-
ual disentanglement of the notion of set from intensional ingre-
dients such as concepts, properties, etc., and the definition of set-
hood in an alternative way. One such alternative way attempted by
some foundationalists consisted in thinking of classes as mere collec-
tions of objects obtained by "abstraction." For instanoe, abstraction
was typically the procedure through which Cantor claimed he could
reach his notion of cardinal number which he defined as "the gen-
eral concept which arises from the set M when we abstract both
from the nature of the various elements m and from the order in
which they are given" (op. cit., p. 282). But abstraction was not, as
such, limited only to the explanation of what numbers are but could
also be used to explain what classes are. Thus, Edmund Husserls2
argued that a collection (totality) consists of contents that are to be
thought without any relation or connection between them.
Frege criticized consistently and constantly other logicians' and
mathematicians' attempts to introduce arbitrary classes by abstrac-
tion. For instance, he criticized Husserl for mistakenly defending a
notion of totality in which "the contents are merely supposed to be
thought or presented together, without any relation or connection
whatever being presented between them."ss For his part, Frege con-
fesses that he is unable "to disregard the connection without losing
the whole" (ibid.).
Frege's forceful criticisms of abstraction have usually been taken
to be merely a part of his overall campaign against psychologism in
logic and mathematics. s" This is a dangerous oversimplification.
Frege's arguments, no matter how they were formulated, had a di-
rect impact on the understanding of his substantial views of logic.
32 Philosophie der Arithmetik: psychologische und logische Untersuchung,

Vol. I (Halle a.d. Saale: Pfeffer, 1891), p. 79.
33 "Review of Dr. Husserl's Philosophie der Arithmetik, " Mind, LXXX (1972),
p. 329. .
34 For some of Frege's criticisms against psychologism, see ibid., pp. 321-37;
Foundations, p. 63; Collected Papers in Mathematics, Logic and Philosophy,
Brian McGuinness et aI., eds. (Cambridge: Blackwell, 1984), pp. 178-81.
They are, in effect, arguments for the primacy of concepts over

their potential extensions. Hence they have an indirect impact also
on his views on higher-order logic. They show that Frege did not
appreciate the notion of arbitrary set, and they show once more how
unwilling he was to accept what later was to be called the standard
interpretation.
XI. THE lWO FACETS OF EXTENSIONALITY
AND THE PROBLEMS OF SET THEORY
Mathematically and logically, the most important target of Frege's
criticisms of abstraction was undoubtedly Cantor. A delicate touch is
needed, however, in discussing the conceptual history of set theory
and its relations to Frege's ideas. Earlier in this paper, we pointed
out the deep kinship between the standard ("extensional") interpre-
tation of higher-order logics and the intuitions on which set theory
was originally based. This standardness, we pointed out, was lost in
the course of the development of the contemporary (first-order)
axiomatic set theory.
Yet a perspective that at first sight might appear to be diametri-
cally opposite to ours has been presented and ably argued for by
Hallett (op. cit.). In his story, set theory was originally afflicted with
a serious case of intensionalism. This malady was gradually cured in
the course of the development of axiomatic set theory, which slowly
led to the all-too-familiar cumulative conception of sets. According
to Hallett, Frege was on the side of extensionalist angels, and not
the bogeyman that he may look like in our scenario. Indeed, Hallett
(op. cit., p. 46) brackets Frege in so many words with those nine-
teenth-century logicians and mathematicians, exemplified by Boole
and Dedekind, who "operated with arbitrary extensions."
Furthermore, according to Hallett, the liberation of set theory
from the dominance of the intensionalist conception was a painful
and slow process. This can be seen from the fact that ingredients of
the intensionalist conception are plenty in Cantor and still, surpris-
ingly, in Zermelo. For instance, in early Cantor a set was defined as a
"totality of elements which can be united to a whole through a
law."35 Also, in Zermelo's formulation of the subset (separation)
axiom, what yields the subset is a "property."36 It is only with Sko-
lem that set theory was definitively purged of intensional entities
like properties, which were replaced by predicates of the given
language. 37
55 "Uber unendliche, lineare Punktmannigfaltigkeiten 5," in op. cit.

56 See Zermelo, p. 183; and Hallett, p. 267.
57 A more detailed description of Skolem's ideas is contained in Hallett, p. 281.
It is nevertheless easy to reconcile Hallett's perspective with ours.

What has happened is that Hallett tells only a half of the full story,
whereas we are in this paper focusing on the other half. We have
been asking: Does there exist, according to Frege or whoever else, a
concept for each potential extension? Hallett is restricting his atten-
tion to an entirely different and in fact reverse question: Does there
exist an extension for each concept? As was pointed out above,
both questions have sometimes been discussed under the undif-
ferentiated heading of "extensionalism," and even assimilated to
each other.
In the course of the actual development of set theory, it was the
second problem that was gradually forced upon mathematicians and
logicians. This happened largely as a response to the paradoxes of
set theory, which showed the inconsistency of certain kinds of as-
sumptions of the existence of sets (corresponding to apparently irre-
proachable concepts). The result was a preoccupation with the con-
ditions of the existence of extensions, given the corresponding con-
cepts. What Hallett argues is that the development of the cumulative
concept of set was looked upon as a process of liberalization of such
existence conditions. We have no quarrel with his argument.
Yet, as far as the other problem is concerned, the upshot of the
development of axiomatic set theory was to give up the standard
interpretation and hence to give up Cantor's original vision of set
theory. Even set theorists like von Neumann, who were beginning to
see the distance between Zermelo's axiomatic system and its Cantor-
ian ancestor and who tried to construct a richer set-theoretical uni-
verse, could never return to Cantor's conception of arbitrary set.
This fact is pointed out by A. Fraenkel and Y. Bar-Hillel58 when they
discuss von Neumann's system:
... The characteristic property of a function, namely being determined
by the course of its values, is expressed by an appropriate axiom, ... it
states that functions with the same course of values are equal-but not
that to each course of values a function is assigned, which would mean
a return to Cantor's conception of an arbitrary set (ibid., p. 99; italics
added).
Thus, it is important to realize that Hallett's is not the only relevant
perspective here nor, we believe, the most important one. What
happens in type-theoretical higher-order logic is that the Gordian
58 Foundations of Set Theory (Amsterdam: North Holland, 1958), esp. ch. II,
sects. 6-7.
knot of paradoxes is cut open once and for all, not unraveled by
means of sophisticated axioms. The type distinctions guarantee that
the existence of the appropriate extension for each concept of the
same type can never cause any problems. The entire attention
switches to the problem of capturing those potential extensions
which prima facie are not extensions of anyone concept. Probably
the simplest way of accomplishing this task is to adopt a nonstandard
interpretation. Piecemeal adoption of new first-order (or nons tan-
dardly interpreted higherorder) axioms will not make essential dif-
ference. What this means for someone like Frege (for whom sets
exists only as extensions of the corresponding concepts) is that an
important problem here is to guarantee the existence of suitable
concepts.
XII. FREGE AND THE TWO FACETS OF EXTENSIONALITY
In the light of these observations, you can now see that Hallett's
perspective is perfectly compatible with ours, as long as it is under-
stood that he is dealing with only one kind of problem in the foun-
dations of set theory. Indeed, we can now appreciate better the
discussions in which Frege was involved.
Hallett is entirely right in emphasizing that Frege wanted to as-
sume an extension (course of values) for each and every function (in
Frege's sense), including every concept. In fact, that assumption is
built into the very notation of the Grundgesetze. This assumption
was dragged out to the open and criticized by Cantor, who declares
it "an unhappy idea":
... to take "the extension of a concept" as the foundation of the

number concept. He [Frege] overlooks the fact that in general the
"extension of a concept" is something quantitatively undetermined.
Only in certain cases is the "extension of a concept" quantitatively
determined; then it surely has, if it is finite, a definite natural number,
and if infinite, a definite power [Machtigkeit, i.e., cardinality]. For such
quantitative determination of the "extension of a concept", the con-
cepts "number" and "power" must previously be already given from
somewhere else, and it is a reversal of the proper order when one
undertakes to base the latter concepts on the concept "extension of a
concept" (op. cit., p. 440).
Here we see both the fact that Frege was indeed on the side of the
"extensionalists" in the issue on which Hallett is concentrating. At
the same time, the quote supports our thesis of the priority of con-
cepts for Frege, in contradistinction to the priority of classes (exten-
sions) for Cantor.
But if so, is there any trace in Frege of the real problems he has to
face? Does he realize that. if he prioritizes concepts over their ex-
tensions, he has to provide enough concepts (and thereby enough
sets) to satisfy the existential needs of actual mathematical theoriz-
ing, among other things to satisfy the needs of set theory? At the
first sight, Frege does not evince any awareness of such problems. In
a deeper sense, however, there are signs of a such awareness in
Frege. Indeed, the very priority of concepts over their extensions
provides a due. For one does not normally speak of assuming the
existence of a concept but of defining the concept. Frege dearly
thinks that a definition can bring new sets into a logician's reason-
ing. This is seen from his discussion in Grundlagen, section 88, of
the analytic versus synthetic distinction and its relation to defini-
tions. The upshot of this discussion is that what Freg'e calls defini-
tions can extend our knowledge and hence be synthetic in Kant's
sense. Admittedly, Frege says that the consequences of such defini-
tions can be proved by purely logical means and hence are analytic
in his sense. But here we must remember that for Frege logical
truths are simply the most general truths about this world. There-
fore, they presumably can involve existential truths. It seems to us
that definitions are for Frege a way of eliciting the existential impli-
cations of what for him were purely logical truths.
This is strongly suggested in Frege's own illustration:
If we represent the concepts (or their extensions) by figures or areas in

a plane, then the concept defined by a simple list of characteristics
corresponds to the area common to all the areas representing the de-
fining characteristics; it is enclosed by segments of their boundary
lines. With a definition like this, therefore, what we do-in terms of
our illustration-is to use the lines already given in a new way for the
purpose of demarcating an area ... But the most fruitful type of defini-
tion is a matter of drawing boundary lines that were not previously
given at all. What we shall be able to infer from it, can be inspected in
advance; here, we are not simply taking out of the box again what we
have just put into it (Foundations, pp. 100-1).
But if so, we can predict what the focus must have been for Frege's
doubts about the way in which Cantor and others were introducing
the sets they needed in their theories. The prediction is that Frege
must have been worried about the way these logicians and gentle-
men defined their basic concepts. This would-be prediction is in-
deed fulfilled. For Frege's criticism of Cantor is leveled prominently
at his use of definitions. This criticism is found in an especiaHy ~harp
formulation in Frege's originally unpublished review of Cantor. 39 It

involves prominently an attack on Cantor's notion of abstraction,
thus indicating its tie to Frege's rejection of the standard interpre-
tation.
The relation of Frege's notion of definition to existence assump-
tions deserves a fuller examination than we can give it here. But
even without such a further investigation, his discussion of other
logicians' and set theorists' definitions is suggestive of this half-
awareness of the crucial (for him) problem of the existence of con-
cepts. Unfortunately, this problem was swept out of the focus of
Frege's attention by the more immediately urgent problem of the
existence of extensions for (so to speak already existing) concepts.
XIII. "CONSISTENCY DOES NOT IMPLY EXISTENCE" AND
THE STANDARD VERSUS NONSTANDARD DISTINCTION
There are other ways of seeing that Frege was inextricably commit-
ted to a nonstandard interpretation. On several occasions, Frege
bases his arguments on the principle that we must not infer exis-
tence from consistency (freedom of contradiction). For instance, he
writes:
Suppose we have proved that in an equilateral, rectangular triangle the
square on the hypotenuse is twice as large as the square on one of the
sides, which is easier than proving Pythagoras' theorem in general. We
can now conclude that the proposition does not contain a contradic-
tion either within itself or in relation to the geometrical axioms. But
can we conclude further: therefore, Pythagoras' theorem is true? I
cannot accept such a method of inference from lack of contradiction
to truth. 40
The principle that Frege rules out appears to be what would be
needed to guarantee that there is a concept for any class (possible
extension). What this possibility means in this case is that, if a con-
cept should somehow be introduced which actually has this class as
its extension, no contradiction would come about. Hence the for-
bidden principle could guarantee the existence of the "missing"
concept.
Admittedly, Frege does not in so many words apply his principle
"existence does not follow from consistency" to deny that there is a
concept for each class, having this class as this extension. But the
39 "Entwurf zu einer Besprechung von Cantors GesammeIten Abhandlungen

zur Lehre vom Transfiniten," in Gott/ob Frege: Nachgelassene Schriften, H.
Hermes, F. Kambartel, and F. Kaulbach, eds. (Hamburg: Meiner, 1969), pp. 76-
80.
40 Correspondence, p. 48.
variety of different cases to which he does apply his principle

strongly suggests that he was committed to applying it also so as to
rule out the standard interpretation.
In the Grundlagen, Frege admittedly uses his principle only to
criticize H. Hankel's somewhat cavalier procedure in extending the
concept of number by using as his only firm guideline the preserva-
tion of the main laws governing numbers. It might seem that no
sweeping conclusions can be drawn from Frege's facile triumph. But
surely Frege could and should have tried to criticize more significant
thinkers than Hankel in the same way. What Hankel does is not so
entirely different from Dedekind's procedure. Dedekind, too, is in
effect relying on the consistency of the result when he introduces a
real to correspond to any cut, even when it does not correspond to a
real number. And if Frege had so extended his criticisms to Dede-
kind, their force would once again come very close to the denial of
the standard interpretation.
XIV. FREGE DISAVOWS THE CONCEPT OF ARBITRARY FUNCTION
The most convincing evidence for ascribing a nonstandard interpre-
tation to Frege is mediated by the connection pointed out earlier
between the standard interpretation and the idea of arbitrary func-
tion. Frege's conception of function is shown not only by what he
says about it but also how he applies the concept of function. In his
early paper, "Methods of Calculation Based on an Extension of the
Concept of Quality,"41 Frege claims to have proved "some general
propositions ... valid for all functions and functional systems"
(ibid., p. 92). Yet in actually proving his results, Frege is freely
assuming that the functions he is dealing with are differentiable
(ibid., p. 60). In other words, Frege's "all functions" are not all
arbitrary functions in the usual sense of the word. Frege's concep-
tion of function is obviously light years away from the concept of
arbitrary function in the sense that is related to the standard inter-
pretation.
Even more striking than the narrowness of Frege's working con-
cept of a function is his obliviousness to the problems afflicting this
notion. As was noted earlier, discussions concerning the general
concept of function were perhaps the most important ingredient in
the lively foundational discussions among mathematicians in Frege's
lifetime. Hence when Frege takes up in so many words the critical
question "What is a function?"42 one might therefore expect some
41 In Collected Papers, pp. 50-92.

42 Ibid., pp. 285-92.
awareness of the actual cases related to mathematical research in

such areas as Fourier analysis. This expectation is not fulfilled. In-
stead, we find in effect an unmistakeable rejection of the concept of
arbitrary function in the relevant sense of the word. Frege formu-
lates, modifying E. Czuber's definition, the idea of an arbitrary
function as follows:
With every number of an x-range there is correlated a number. I call
the totality of these numbers the y-range (ibid., p. 289).
Frege calls our attention, quite correctly, to the notion of correla-
tion involved in this concept of function:
The heart of the matter really lies in quite a different place, viz. hidden
in the word 'correlated' (ibid.).
He notes that his opponent leaves the nature of this correlation
completely open, quoting Czuber as saying that
The above definition involves no assertion as to the law of correlation,
which is indicated in the most general way by the characteristic f; this
can be set up in the most various ways {ibid.}.
Frege immediately focuses on the idea of "law":
Correlation, then, takes place according to a law, and different laws of
this sort can be thought of. In that case, the expression y' is a function
of x' has no sense, unless it is completed by the law of correlation
(ibid.).
But how can we do that? Frege answers:
Our general way of expressing such a law of correlation is an equation
in which the letter 'y' stands on the left side whereas on the right there
appears a mathematical expression consisting of numerals, and the let-
ter 'x', e.g. 'y = x 2 + 3x' {ibid.}.
Here Frege is not speaking of "correlations" as being exhibited by
objectively existing entities out there in the Platonic realm of num-
bers, called functions, but by actual mathematical equations ex-
pressed in a suitable notation.
This by itself does not yet imply that Frege rejected the idea of
arbitrary function. For it leaves open the possibility that laws defin-
able by equations do not exhaust the range of all functions he is
countenancing. Helpfully, Frege nevertheless clarifies his position
immediately. He continues:
Functions have indeed been defined as being such mathematical ex-
pressions. In recent times this concept has been found too narrow.
However, this difficulty could easily be avoided by introducing new

signs into the symbolic language of arithmetic (ibid" p. 290).
This is a mind-boggling statement. It shows Frege's mind at work. It
shows that there is no niche in his world for the notion of an arbi-
trary function in the sense in which Euler, ezuber, and others were
thinking of this notion. For to think that the range of such arbitrary
functions could be covered by introducing new signs into the nota-
tion of arithmetic is not to understand this notion. Hence we can
safely conclude that, in the logic to which Frege tried to reduce all
mathematics (except geometry), there is no room for the idea of an
arbitrary "function-in-extension" and hence no room for the stan-
dard interpretation.
It is not that Frege accepts uncritically the characterization of the
largest relevant class of functions by means of equations. Indeed, he
immediately goes on to criticize the quoted characterization. But
this criticism has nothing to do with the narrowness of the class of
functions so definable. It deals with the fact that the proposed char-
acterization obscures the "distinction between sign and the thing
signified," and hence obscures the objective existence of functions
independently of any particular way they are formally expressed.
What is interesting about this justified criticism by Frege is that its
upshot obviously has nothing to do with the range of functions that
exist or do not exist. Problems of that kind, Frege has already pro-
claimed, "can easily be avoided by introducing new signs into the
symbolic language of arithmetic." Hence Dummett and others are
simply wrong if they think that the objective existence of functions,
or even their mode of objective existence as correlations of objects
with objects, prejudices Frege in favor of the standard interpre-
tation.
At the same time, the latest displayed quote from Frege boggles
the mind in that it demonstrates Frege's alienation from the actual
working problems in the foundations of mathematics of his time.
Frege is in effect telling Dirichlet, Weierstrass, Dedekind, H.
Poincare, et alia that the problems they have been agonizing about
could easily by solved by introducing new signs into the notation of
arithmetic. It is perhaps not surprising that Frege's foundational
ideas did not attract more attention among the practicing founda-
tionalists than they in fact did.
xv. FREGE'S REDUCTION 1ST PROGRAM IS A NONSTARTER
What have we accomplished so far? And why have we done what we
have done? Why worry about whether Frege accepted the standard
or nonstandard interpretation of higher-order logic? After all, the
distinction was articulated much later, and hence Frege could

scarcely have expected to pay much attention to this distinction,
which might seem rather pedantic. Yet there is a massive reason for
raising the question we have raised. The evaluation of Frege 's en-
tire life work depends crucially on the answer. Frege tried to re-
duce mathematics to logic. In the process, he used crucially higher-
order quantifiers. For instance, he defined "The number which be-
longs to the concept F is 00 1" as "there exists a relation which
correlates one-to-one the objects falling under the concept F with
the finite Numbers" (Foundations, p. 98). Despite the fact that
Frege claimed that his terminology "diverges to some extent" from
Cantor's, in the Grundlagen he welcomes Cantor's achievements
because "they have led to the construction of a purely arithmetical
route to higher transfinite Numbers (powers)" (ibid., p. 96). It has
been emphasized in this paper, not for the first time, that the diver-
gence between Frege and Cantor is much deeper than merely a
terminological one. The divergence spreads from two radically dif-
ferent conceptions of a set, one that ties sets to concepts and proper-
ties, the other that generates sets freely through operations acting
on the members of some already existing sets. Likewise, it makes
more than a terminological difference whether one defines a one-
to-one correspondence between two sets as any arbitrary one-to-one
function in extension mapping them on each other, as Cantor envis-
aged it, or as a similar function picked up by a concept, as Frege in
effect did.
Also, when Dedekind characterized the real numbers in effect by
means of the cut principle, which says that every bounded (from the
above) set of reals has a least upper bound, this characterization is a
categorical one only if the sets involved are arbitrary ones and not
only such sets as are courses of values of concepts expressible in the
language of arithmetic. Thus, there is a deep sense in which Frege's
system is not adequate for interpreting results in contemporary set
theory and mathematical theorizing, for instance in real analysis.
XVI. FREGE'S PLACE IN HISTORY
Even more generally, serious questions must be raised in the light of
our results and of other results about Frege's stature as a philoso-
pher of mathematics and as an analyst of the actual working founda-
tions of mathematics. Frege created our received conception of lo-
gic, not just in the sense of conceiving the general idea but in 'the
sense of actually implementing it. This is a lasting achievement. Yet
it turns out that his creation, the customary idea of first-order logic,
is in certain respects too restrictive and perhaps even arbitrarily
restricted. As Hintikka 43 has recently shown, Frege and others have

arbitrarily excluded quantifier independence from their logic. As we
have seen, Frege's failure to grasp the idea of the standard interpre-
tation of higher-order logic turns his entire foundational project
into a hopeless daydream. Elsewhere, Hintikka44 has pointed out
how totally blind Frege was to what must be considered the most
significant line of development in modern foundational studies, the
gradual systematization of the model-theoretic viewpoint. In particu-
lar, Frege totally failed to understand the early uses of this viewpoint
by David Hilbert, especially in Hilbert's work in the axiomatic foun-
dations of geometry.
As Hilbert's work was easily the most influential kind of founda-
tional work among mathematicians and scientists in the early de-
cades of this century, this means that Frege was totally blind to one
of the most important lines of thought in the foundations of mathe-
matics in this century. What all of this adds up to is that Frege failed
to relate his own work to much, not to say most, of the foundational
work that really mattered to actual mathematical research and to
mathematicians' own efforts to clarify the fundamentals of their
own discipline. It seems to us unfortunate that philosophers habitu-
ally go to Frege for their problems and for the concepts that could
help us to cope with them. Frege was far too myopic to be a fruitful
source of concepts, ideas, and problems. It is likely to turn out that
some of his contemporaries can offer far better ideas for future
work in the foundations of mathematics than Frege did.
4S "What is Elementary Logic?" Proceedings of the Boston Colloquium in Phil-

osophy and History of Science (Dordrecht: Kluwer, forthcoming).
44 "On the Development of the Model-Theoretical Tradition in Logical
Theory," Synthese, LXXVII (1988): 1-36.
POSTSCRIPT
The publication (in 1992) of the article translated above produced a minor
firestorm of protests. No fewer than six philosophers sprung to arms to
defend good old Frege in public. Unfortunately for the critics, none of the
responses to our article turns out to be persuasive in the least. Examples of
the shortcomings of the critics are provided below. What is especially striking
about the responses is nevertheless the degree to which some of them were
predicated on the views on the foundations of mathematics that can only be
considered the current orthodoxy. One strange thing here is that that
orthodoxy would have been deeply repugnant to Frege. Frege sought to
reduce mathematics to logic - to his logic. And this logic was an
axiomatically developed higher-order logic which incorporated our ordinary
first-order logic as a part. Now even though the higher-order component of
Frege's logic can be compared with set theory and perhaps even be said to
overlap with it, it was still part of logic for Frege. Otherwise his entire
foundational program would have made no sense. The idea that set theory
conceived as just another mathematical theory would be the best we can do
in the foundations of mathematics would have been totally foreign to him.
For him, unlike the current orthodoxy according to which, in Quine's words,
higher-order logic is set theory in sheep's clothing; set theory would have
been higher-order logic in wolfs clothing. It is therefore in order to consider
briefly how our results in the original paper affect the location of Frege's
efforts on the map of foundational studies and affect the overall evaluation of
his thought in the light of other recent and ongoing research.
Why did Frege need higher-order logic, anyway? He was trying to reduce
mathematics to logic, and for the purpose he had to define the concept of
cardinal number. For that purpose, he needed to express in his symbolism
when two classes are equinumerous ("gleichzahlig", in Frege's jargon). Given
the rest of Frege's thought, he could do it only by quantifying over relations,
i.e. on the second-order level.
The expression
"the concept F is equal (gleichzahlig) to the concept G'
is to mean according to Frege the same as the expression
"there exists a relation <I> which correlates one to one the objects falling
under the concept F with the objects falling under the concept G'
In order for this definition to serve its purpose, there must actually exist,
intuitively speaking, a relation doing the correlation as soon as the two
concepts are extensionally speaking equinumerous. Since Frege's higher-
order variables range over concepts (e.g. relations) rather than their
extensions, this is what he has to assume for his definition to do its job. But
from Frege's general assumptions it does not follow that there should always
exist a relation implementing the correlation, and given those assumptions it
is hard to see that such existence would even be plausible. It is of course
precisely such existence that the standard reading of higher-order quantifiers
guarantees. But Frege does not assume the standard interpretation, as we
argued in the original paper.
In fact, additional reasons for the same conclusion are easily found.
Perhaps the most obvious one is based on Frege's own procedure in the
Grundgesetze. He puts forward there no explicit existence assumptions for
sets. Sets enter his treatment only as value-ranges of concepts. As a
consequence, only those sets are forced to exist by Frege's axioms as are
value-ranges of some concept expressible in the system. This means that a
Fregean system of higher-order logic cannot rule out nonstandard models, in
the sense that among its models there will inevitably be some nonstandard
ones. This point is independent of the question as to whether according to
Frege only such concepts (and other "functions" in Frege's sense) exist as can
be expressed in his language.
Hence Frege could only hope to rule out nonstandard models
interpretationally. But this was virtually impossible for him to do. The
reason is that his official treatment of his logic was purely axiomatic and
formal. Interpretational points would not have made any difference to him
unless they forced on him an assumption that could be expressed as an
explicit assumption or definition. In any case, such interpretational
pronouncements by Frege as we can find point precisely to the wrong
direction, that is, toward a nonstandard interpretation. In a Fregean
perspective, the crucial question is whether for any (finite or infinite)
collection of objects there exists a concept which picks out all those objects
and only them. Now the kind of operation by means of which Cantor and
others thought they could move from given collections of individuals to the
abstract entities they were dealing with, such as sets and numbers, was
abstraction. Frege recognized abstraction as a viable procedure, but argued
passionately that it cannot do the job assigned to it by the likes of Cantor.
For one thing, abstraction can only yield concepts, not obj,cts like classes or
numbers.
With ordinary abstraction we start out by comparing objects a, b, C, and

find that they agree in many properties but differ in others. We abstract
from the latter and arrive at a concept ~ under which a and b and c all fall.
Now this concept has neither the properties abstracted from nor those
common to a, b and c. The concept "book", for instance, no more consists of
printed sheets than the concept "female mammal" bears young and suckles
them with milk secreted from its glands; for it has no glands. (Posthumous
Writings p. 71.)
Admittedly, concepts can at a second remove yield classes as their value-

ranges. But if abstraction works in the way Frege indicates in the quote,
there plainly is no reason to expect that the totality of those concepts that a,
band c share do not apply to other objects as well. And if so, we do not obtain
a concept that has all and only the initially given objects falling under it,
that is, we do not have a concept for each class having that class as its value-
range. As a consequence, there is no reason to assume that the standard
interpretation applies. In other words, the only way of making sure that this
does not happen would be to assume that for any collection of objects there is
a concept that applies to all of them, and only to them. But this amounts to
assuming the standard interpretation. What is especially relevant here-
amounts to assuming abstraction can yield an object of the kind that Cantor
and Dedekind needed. Hence Frege's criticism of abstraction is in effect a
criticism of the standard interpretation. Contrary to what has sometimes
been claimed, it is not a part of Frege's criticism ofpsychologism in logic.
Thus it might seem that a definitive judgment of history goes against
Frege. Not only can he be seen to be left with only a nonstandard
interpretation of higher-order quantifiers at his disposal. This failure seems
to be fatal to his entire logicist project, in that it spoils his definitions of
equinumerosity and a fortiori of number.
Furthermore, the definitions of equinumerosity and number are not the
only places where Frege would apparently need the standard interpretation
for the purpose of reconstructing mathematics within logic. For instance,
there is obviously no hope of defending the axiom of choice on Fregean
principles.
This is not the end of the story, however. Rather it directs our attention to
certain interesting recent double-edged developments. They are double-edged
in that they tend to vindicate (with certain qualifications) Frege's logicism,
and yet at the same time reveal a different and perhaps even more
fundamental shortcoming of Frege's. Frege needed higher-order quantifiers
and by implication standard interpretation primarily for the definition of
equinumerosity. But it has turned out that one can after all define the
equinumerosity of A(x) and B(x) on the first-order level provided that one
allows for informationally independent quantifiers. Indeeci, using the slash
notation familiar from our earlier publications, the following sentence does
the job:
('v'x)(Vz)(3yNz)(3uNx)A(x) ::J B(y & (B(z) ::J A(u & x = u) ~ (y =z)

Thus Frege could have got around (to some appreciable degree at least) his
lack of a standard interpretation of higher-order quantifiers by dispensing
with them altogether. However, he would have had to use informationally
independent first-order quantifiers to compensate for this lack. And he did
not employ them, even though they are part and parcel of the same group of
ideas as ordinary first-order quantifiers.
Furthermore, one can trace Frege's failure to accommodate independent
quantifiers to his static conception of quantifiers as higher-order predicates.
This kind of conception is incapable of doing full justice to the interplay of
quantifiers with each other. Such as relations of dependence and
independence between them. Frege's treatment of quantifiers rules out
certain configurations of dependence and independence relations among
quantifiers. If they are allowed, either by means of the slash notation or by a
liberated use of parentheses, it becomes possible to characterize not only
equinumerosity but several other crucial mathematical concepts, e.g. infinity,
by means of first-order quantifiers. This pre-empties all questions of
standardness or nonstandardness, at the very least in any sense that might
be relevant to Frege.
One implication of such results is generally speaking to encourage a
logicist approach to the foundations of mathematics. In general as Jaakko
Hintikka has shown in his recent book, in a suitable perspective there is
little in mathematics that cannot in principle be conceptualized in terms of
first-order logic, amplified by allowing informational independence. In this
sense, recent developments tend to support a logicist approach rather than
the orthodoxy relying on set theory as a framework in the theoretical
foundations of mathematics.
The irony of the situation is that Frege's greatest achievement is often
thought of to be his systematization of the logic of quantifiers. From the
vantage point of the observations just made, Frege's comprehension of what
is involved in the real logic of quantification was seriously defective.
9
AN ALTERNATIVE CONCEPT OF COMPUTABILITY*
1. RECURSIVITY AND CHURCH'S THESIS
The fonnation of the current concept of computability (recursivity) is one of

the major achievements of twentieth-century logical theory (see e.g. Davis
1965, Rogers 1967, ch. 1). What is especially impressive about this notion is
its robustness. Originally, different logicians proposed different and prima
facie completely unrelated concepts of computability, among them
computability by a Turing machine, definability ala Kleene through a finite
number of recursion equations and expressibility in Church's lambda-
calculus. As if by miracle, these different characterizations of computability
turned out to coincide. This consilience of apparently unrelated definitions
is a telling indication that the notion so captured is indeed theoretically
important. It lends prima facie support to the thesis of Church (1936) that
mechanically determined (finitely computable) functions (in a pretheoretical
sense) are precisely the recursive ones.
This notion turns out to be robust in another way, too. The definition of
Turing machine computability, which is one of the several equivalent
characterizations of recursivity, is invariant also with respect to the
different possible characterizations of the idealized machine itself. For
instance, the number of infinite tapes used by the machine is irrelevant, as
long as the number is finite (cf. Turing 1936). To some extent the invariance
of recursivity with respect to conceptually different characterizations can be
considered as a reflection of this machine-theoretical invariance of the
notion of Turing machine computability.
The notion of recursivity (Turing machine computability) is generally
assumed to be the natural explication not only of the idea of mechanical
computability in any reasonable sense of "mechanical" but also of the idea of
mechanical detenninacy. For instance, in discussions of whether the human
mind can be a machine it has usually been assumed that Turing machine
computability exhausts the range of machine-like determination.
Furthennore, the recursivity of solutions to differential equations
describing a physical system is sometimes taken to be a necessary and
sufficient condition of the system's being detenninistic. Much of the
presumed general theoretical interest of the notion of recursivity is thus due
* Written jointly with Arto Mutanen

AN ALTERNATIVE CONCEPT OF COMPUTABILITY 175
to its status as an explication of the idea of a function's being mechanically

detennined.
This claim by the notion of recursivity to being the right explication of
our pretheoretical idea of being mechanically detennined is codified in what
is generally known as Church's Thesis.
In this paper it will be shown that recursivity is not the only natural
explication of this kind nor even the most fundamental one theoretically.
For the purpose, we obviously have to fonnulate an alternative concept of
computability. In this enterprise, we are not presupposing any new
hardware in our new definition of computability. AB a consequence, we can
in the usual way enumerate recursively the partial functions that are
computable in our new sense simply by enumerating recursively the Turing
machines that are supposed to compute them. What we do want to point out
is that there is more than one sense in which the same idealized hardware
can be used to compute a function.
2. WHAT IS MECHANICAL COMPUTABILITY?
In what sense does a Turing machine compute a recursive function? One is

perhaps tempted to say something like this: Consider a Turing machine
with two tapes, a working tape and a book-keeping (result-recording) tape.
The machine can write on and erase from either tape. In view of the
robustness of the ordinary notion of Turing machine computability this is
not a real restriction. What is written on the book-keeping tape can be
assumed to be equations of the fonn f(a) = b where a and b are natural
numbers. These equations serve to define the function to be computed. The
Turing machine in question computes the function if and only if all (and
only) such true equations appear on the tape when the maehine has run its
course. This means that eaeh true equation f(a) = b will oceur on the result
tape from some finite stage on for good. Needless to say, only one equation
of the fonn f(a) = b must eventually be on the result tape for any given a
from that point on. If no such equation ever occurs on the result tape for a
given a or if it keeps forever changing, f(x) is not defined for x = a.
If this is what you are tempted to say, you are wrong as far as the
received notion of Turing machine computability is coneerned. You are
defining a concept of computability that is wider than recUlrsivity. In order
to obtain the usual notion of recursivity (computability), we have to impose
additional requirements.
One way of implementing them is to require that the machine never
erases anything from the result tape. The point of this requirement is
obvious. What is required for the usual notion of computability is not only
that each true equation f(a) = b is on the result tape from some finite stage
on, but that we know when the machine has reached this stage. The
requirement of nonerasure is one possible way of implementing this
desideratum. Alternatively, we could of course allow erasure and stipulate
instead that f(x) has the value b if and only if the equation f(a) = b is the
last equation of the form f(a) = x produced by the machine. (This means that
no further equations of this form are produced by the machine from some
point on.) In order to reach the usual notion of recursivity, we then have to
impose the further requirement that one can decide effectively when the
last equation has been reached. This epistemic requirement is obviously
desirable for the practical uses of the results of computation, for we do not
have enough time to wait for the machine to run its course to infinity.
Needless to say, in many practical situations we nevertheless can decide
whether an equation on the result tape will stay put there indefinitely even
without a mechanical method of deciding it for all cases. Hence giving up
the epistemic desideratum does not even rule out all practical uses of our
machine computations.
However, in principle there will always be an experimental element to
such computations. For our wider concept of computability differs from
recursivity only when the length of the computation needed to reach the
definitive value does not admit of recursive upper bound, such a
computation thus grows very fast in length. In other words, computations in
our sense are typically very long precisely in those cases in which they
cannot be reduced to ordinary Turing machine computations.
However, it may even be argued that this epistemic requirement makes
the usual characterization of recursivity epistemologically circular. (In order
to decide mechanically what the value of f(x) is for a, we have to decide
mechanically how far the search for that value need to be carried out.) We
will not press this point here, however.
In any case, however, the main impact of dispensing with the epistemic
requirement affects the theoretical uses of the notion of computability,
especially for uses where the identification of computability with
mechanical determination plays a role. In such contexts, the restriction
involved in the received notion of recursivity is unnatural. For the
restriction just explained is essentially an epistemic requirement, asking for
a guarantee that we know when a trial-and-error construction of a function
has reached the true value f(a) = b of a function f for the argument a. Such
an epistemic requirement is not implied in the idea that the values of fare
mechanically determined.
3. TRIAL-AND-ERROR COMPUTABILITY AND SATISFIABILITY
By simply omitting the epistemic requirement, we obtain a new, wider

sense of computability. We will refer to it as tae-computability, short for
trial-and-error computability. The name is motivated by the fact that
erasure from the result tape is permitted by our definition. Such erasure
can be thought of as an acknowledgment on the part of the machine that its
trial choice of a line of computation has been in error and. that it is using
that recognition of an error to try a different line of computation.
The notion of tae-computability is arguably more fundamental
theoretically than recursivity, in that the latter notion is obtained from it by
adding the epistemic requirement that we can always recognize when the
machine has reached the right equation f(a) = b for a given argument value
a.
In this paper, we will not try to investigate the properties of tae-
computability in any great depth. Instead, we will make certain general
points about this notion and comment on its potential theoretical
significance.
First, how do we know that tae-computability is a wider notion than
recursivity? One way of answering this question is to prove our first result
concerning tae-computability.
Theorem 1. Each satisfiable formula S of first-order logic has at least one

model where its Skolem functions are tae-computable.
This theorem can be seen to hold by considering first one of the standard
cut-free methods of proof for first-order logic without identity, e.g. some
variant of the tree method type attempt to prove ...,S. (We can assume that S
is in a negation normal form.) Such an attempt is tantamount to an
attempted construction of a model for S. In it, we may assume that all
existential quantifiers have been replaced by Skolem functions. Whenever a
new term t is introduced, it is assigned an individual b as a value, and the
equation t = b is printed on the result tape.
Some rule must be chosen to select the formula to which a rule is to be
applied next. For instance, we can stipulate that the oldest available (and
not yet used) rule application always be carried out first.
A critical role is played in such a tree method by applications of the rule
for disjunctions. Each disjunction that appears during the construction
results in dividing the construction into two branches. In the usual
applications of the tree method, all resulting construction branches are
constructed at the same time. Here we are instead assuming that one
particular disjunct is always chosen to be tried first, for instance the

leftmost one. (Actually any recursive rule will do.) If the resulting
construction branch ends up in a contradiction, we return to the last node
where a disjunction rule was applied and where the other disjunct has not
yet been selected. All the intermediate entries on the result tape since this
node are erased, and the other disjunct is used to continue another
construction branch.
This procedure is clearly completely deterministic, and by Church's thesis
it can be implemented by a Turing machine. (Even apart from Church's
thesis, such an implementation can be seen to be possible.) Even more
clearly, the formulas that stay put on the result tape constitute a model set
which shows that S is satisfiable. Among the members of this set there are
equations of the form f(a) = b where f is the function to be computed. They
determine partially the function f.
Because of the upwards Skolem-LOwenheim theorem, we can assume
that the individuals used are natural numbers. Then those equations of the
form f(al, a2, ... , ak) = b (where f(xI, X2, ... , Xk) is one of the Skolem functions
of 8), which stay put on the result tape from some finite stage on define
partly a Skolem function which together with its brothers satisfies S. These
functions are obviously tae-computable.
If we want the procedure to yield total and not only partial functions, we
can stipulate that for the n:th tuple of values in some recursive ordering,
say <aI, a2, ... , ak>, we add at the n:th stage to the output tape the equation
f(al, a3, ... , ak) = n for each Skolem function with k arguments, in case no
similar equation f(cI, C2, ... , Ck) = d occurs on that tape. If an equation of this
form occurs later on the tape, the "artificial" equation f(al, a2, ... , ak) = n is
omitted.
This argument does not work in first-order logic with identity, for then
equations introducing values for new terms involving Skolem functions may
conflict with equations resulting from the formulas to be satisfied. However,
what we can do is to translate the initial formula to be satisfied to a form in
which it has no identity signs and only exclusive quantifiers (cf. Hintikka
1956). Then the argument goes through as before. (In the same way, some of
the known results from the literature can also be extended to first-order
logic with identity; cf. Putnam 1965, p. 51; Mostowski 1965, p. 59.)
Putting all these observations together, we can see that a set of verifying
Skolem functions for S is tae-computable in the sense defined above. This
holds for any satisfiable formula S. Yet it is known that for some satisfiable
formulas there exist no models with recursively enumerable predicates and
hence no sets of recursive Skolem functions (see Kreisel 1953, Mostowski

1955). Hence the concept of tae-computability is wider than recursivity.
A word of warning is needed here. If the set of formulas for which Skolem
functions are being constructed include the axioms of elementary
arithmetic, then there is no guarantee that the resulting model is the
standard one.
Since independence-friendly (IF) first-order logic is comp act, we can do
the same for all satisfiable IF first-order formulas, proving:
Theorem 2. Each satisfiable formula of independence-friendly first order

logic has at least one model where its Skolem functions are
tae-computable.
(For IF first-order logic, see Hintikka 1995 and 1996, chapters 3-4.) The
construction of tae-computable Skolem functions for satisfiable first-order
formulas is worth discussing further. Which set of Skolem functions for S is
obtained from a tree method construction beginning with S is determined by
the choice of which disjunctive branch is tried first in the construction.
Usually different recursively determined totalities of such choices give rise
to different sets of verifying Skolem functions for S, each of them tae-
computable. We can thus correlate with each satisfiable S the class of such
sets of tae-computable Skolem functions and by implication the set of such
computer architectures as can tae-compute all those Skolem functions. Thus
we can associate with each satisfiable first-order formula S the class of
computer architectures capable of producing all its tae-computable Skolem
function sets.
This correlation of first-order logical formulas with classes of computer
architectures will be examined further in another paper (see Hintikka and
Sandu, forthcoming). It can be thought of as a natural extension of the time-
honored correlation of truth-functions and combinations of logic gates. It
also extends the correlation used by Hintikka and Sandu (1995) in their
work in which they show a connection between the basic ideas of IF first-
order logic and those of parallel processing. They show that IF first-order
logic is in a most natural sense the logic of parallel processiLng. In general,
the correlation we have defined brings the theory of first-order logic
(including IF first-order logic) and the theory of computability closer to each
other than before. This is illustrated by the earlier result by Hintikka and
Sandu just mentioned.
180 LANGUAGE, TRUlH AND LOGIC
4. IMPLICATIONS OF TAE-COMPUTABILITY
An interesting foundational application may affect interpretations of

constructivistic ideas in logic. Hintikka has shown (1996, ch. 10) how
constructivistic ideas can be implemented in logic by restricting all models
of a formula S to those in which the Skolem functions of S are recursive. A
less restrictive interpretation is obtained by requiring merely that the
Skolem functions of S be tae-computable in the (admissible) models of S.
This interpretation is intriguing in that Theorem 1 implies that the
resulting logic is in the case of ordinary first-order logic the same as
classical logic. We thus obtain the following interpretational analogy:
tae-computatility = classical logic

rucursivity constructivistic logic
This analogy shows the important position of tae-computability in

systematic logical theory. In this way, the notion of tae-computability brings
to light interesting but largely unexploited connections between logical
theory and the theory of computability. They deserve a separate study (see
Hintikka and Sandu, forthcoming). Here we will concentrate on the
philosophical implications and applications of the notion of tae-
computability.
Among other things, the theoretical consequences of our alternative
concept of computability concern Church's thesis. It asserts that Turing
machine computability (recursivity) is the right explication of our
pretheoretical notion of mechanical calculability (mechanical
determination) of a function. Will Church's thesis remain valid after our
results? Our qualified answer is: No, it is not valid. What our observations
show is that our pretheoretical notion of mechanical calculability is
ambiguous. Does it include the epistemic restriction which leads to Turing
machine computability rather than tae-computability? We do not think that
an unqualified answer is possible. We doubt that our pretheoretical ideas of
mechanical calculability are so sharp as to allow for only one explication.
However, the lines of thought sketched above strongly suggest that the
notion of tae-computability is theoretically more fundamental than the
concept of recursivity. In particular, what can be said in the light of the
observations made above is that if what is emphasized is the mechanical
determination of a function, tae-computability is a more natural candidate
for an explication of our pretheoretical idea of computability than
recursivity. Hence there not only is a possible sense, but even a natural
sense, in which Church's thesis fails.
It should be fairly obvious without much argument that recursivity
cannot be a satisfactory explication of the idea of a purely mechanically
determined function. One way of seeing this is to consider halting problems.
Given a Turing machine (say one whose number is m), programmed to
calculate a function and a number n, will it produce a function value fm(n)?
The answer to this question is completely ("mechanically") determined. It
depends only on the machine, which is determined by m, and on the number
n. Consider the function H(m,n) which has the designated value, say 1,
whenever the machine calculates fm(n) and no value elsewhere. This
function is easily seen to be tae-computable. (See Theorem 3 below.) But it
is not recursive, even though it is obviously mechanically determined in a
perfectly natural sense.
What is possible and indeed important is to see which applications of the
notion of computability are affected by the acknowledgment of the notion of
tae-computability. And part of the story here is clear. Our notion of tae-
computability catches better than the notion of Turing machine
computability the idea of a function's being "deterministic" in the sense of
purely mechanically determined. For obviously the compu tabon of any trial-
and-error-computable function is inexorably fixed, even if it is not recursive.
The process of tae-computation does not leave any room for free choice or
uncertainty. The entire process is completely fixed a priori. Hence it is the
notion of tae-computability, not the notion of Turing machine computability
that is relevant to questions whether certain systems, say physical systems
governed by systems of differential equations, are deterministic or not. The
additional requirement codified in the usual notion of Turing machine
computability, viz. that we have a way of knowing when the machine has
reached the right function value for each given argument value must be
looked upon from the vantage point of the mechanical determination of the
function in question as mere idle curiosity of us mortal human beings.
In principle, one could even try to use tae-computations for the purpose of
actually finding out the values of a function. This "finding out" will
inevitably have as it were an experimental component, in that we cannot be
certain in all cases that the interim values of the function to be computed
are the right (final) ones. This does not prevent them from knowing that one
has reached the definitive value for most or perhaps even for practically all
argument values. Hence it might seem that tae-computability is relevant
also for practical calculations. Above we indicated another reason, however,
why tae-computability does not seem to be of much practical use. The
reason is that characteristically tae-computability, like the tae-computation

of models for ordinary first-order formulas, is extremely slow. It remains to
be studied, however, whether this obstacle can in some cases be removed.l
Hence the main uses of the new notion of computability are likely to be
theoretical. Some of them might be worth mentioning here. Among the
notions that can be defined in terms of tae-computability is a new notion of
axiomatizability (if this term is still appropriate). We might say that a
complete theory (a deductively closed class of sentences) is tae-
axiomatizable if and only if it is tae-computable.
The new notion of computability also shows how careful we have to be in
speaking of computability and related notions. Recursively unsolvable
problems are sometimes described by saying that for them we can never
reach a solution. In the case of some such problems, it would be more
accurate to say that we can indeed reach the solution sooner or later, but
that we cannot always tell whether we have reached it or not.
5. TOWARD A THEORY OF TAE-COMPUTABILITY
More generally speaking we can develop on the basis of the new notion of
tae-computability a partial analogue to the usual theory of recursive
functions. In this task we are helped by the fact that the notion of tae-
computability is not completely new in the literature, even though it has not
been put forward (as far as we know) in general terms as a form of
computability. The first use of the notion seems to be Putnam (1965), who
uses it to solve a number of problems concerning the arithmetical
complexity of predicates in models of first-order formulas.
The task of developing a full-fledged theory of tae-computability is
nevertheless far too large to be carried out adequately within the scope of
one paper. Some indications can nevertheless be given as to what the
resulting theory will look like. We can for instance define a set to be tae-
enumerable if and only if its characteristic function is tae-computable, and
define a set to be tae-decidable if and only if both it and its complement are
tae-enumerable. (By sets we mean in this paper always sets of natural
numbers.) A problem concerning the membership in a set is said to be tae-
decidable if and only if this set is tae-decidable.
In particular, we can formulate analogues to many of the classical
decision problems in the form of problems concerning tae-decidability. For
instance, there is an analogue to Hilbert's Tenth Problem (see Matiyasevich
1993) for tae-decidability instead of recursive decidability. In view of the
theoretical interest of the notion of tae-computability, these new decision
problems are of a considerable importance in the foundations of logic and

mathematics.
These decidability problems will have to be answered by means of
argument unlike the customary ones. One reason is that the most familiar
kinds of diagonal argument are not applicable here. For instance, suppose
we are asking whether the Turing machine with the number n defining a
partial function fn(x) yields a value for the argument m in the sense of tae-
computability. This is our new counterpart to the classical halting problem.
In trying to reproduce the usual argument for the unsolvability of the
halting problem, we cannot any longer use as a part of the instructions for
computing the diagonal function the fact that fn(x) is defined for a given
value of x, for instance n, the obvious reason being that even if fn(x) is
defined, there need not be any way in which we (or the machine) could
recognize this fact and use it as a guideline of further computation. And
without such diagonal arguments, the new decision problems are clearly a
new ballgame.
Some light is thrown on the situation by:
Theorem 3. The set L of Turing machines which define a function f which is

determined for a given value a is tae-enumerable.
Proof: Add to the given Turing machine the following specification:
On a separate tape, the machine prints 1 when it first reaches
an interim value b = f(a) for fat a. It changes this value from n
to n + 1 every time b = f(a) is changed by the machine to
something else, e.g. c = f(a) , c "* b. At each calculation step
which does not change b = f(a) , the machine prints n - 1
instead of n unless n = 1, in which case it does not change n.
Clearly there is 1 permanently on the auxiliary tape from some point on

if and only if the given Turing machine tae-computes a value f(a) for f at the
argument value a.
This theorem shows that in a sense the counterpart of the classical
halting problem for tae-computability is tae-decidable. Of course the
problem of "halting" does not any longer concern stopping but constancy of
the value the machine has reached.
Some basic facts about the theory of tae-computability are easily
established. Among them, there are the following results:
Theorem 4. If the sets A and Bare tae-enumerable, then so are (A n B)

and (AuB).
Theorem 5. The tae-enumerable sets form a lattice.
Theorem 6. If the functions f, g are tae-computable, then so is f(g(x.
Theorem 7. Being tae-computable is an arithmetical predicate. In fact, it is

a L2 predicate.
Proof: From the definition oftae-computatility.
This theorem is closely related to Putnam's (1965) theorem 1 which says

that a predicate Pis tae-decidable if and only if P E L2 " rh.
Theorem 8. The class of arithmetical truths is not tae-enumerable.

Proof: The class of arithmetical truths is not arithmetical.
Theorem 9. If A is recursively enumerable, the complement of A is tae-

enumerable.
Proof: Assume that a Turing machine is given that recursively
enumerates A, for instance in the sense that it computes a
function f(x) such that fen) = 1 occurs on the result tape if and
only if n E A. Change the instructions in such a way that if fen)
= 1 has not occurred on the result tape at stage n, f(n) = 0 will
be written on the result tape. When f(m) = 1 occurs, any
equation f(m) = 0 is erased from the result tape.
From Theorem 9 it follows that every finitely axiomatizable theory is tae-

decidable.
It is even possible to prove suitable fixpoint theorems. For example, let <pn
be the n:th partial tae function in some effective G6del numbering of such
functions. Then one can prove the following:
Theorem 10. l.t f(x) be any total tae-computable function. Then there exists
n such that <pn = q>{(n)
Proof: The same, mutatis mutandis, as for the usual recursion
theorem (see e.g. Rogers 1967, p. 180).
While a great deal of work remains to be done on the notion of tae-

computability, it can already be seen what some of the main features of the
theory of tae-computability are.
Since the notion of tae-computability relies on the same hardware
(Turing machines) as the notion of recursivity, those aspects of the received
AN ALTERNATIVE CONCEPT OF COMPUTABll.ITY 185
theory of computability that depend only on the hardware in question can

be extended virtually intact to the theory of tae-computability. This applies
for instance to theorems asserting the existence of universal machines and
to various fixpoint theorems.
On the other hand, since we have no mechanical way of deciding when a
tae-computation has reached its definitive conclusion, we cannot use such
an outcome as an input into a tae-computation. This implies that the usual
kinds of diagonal arguments are inapplicable. The sample theorems listed
above illustrate both of these restraints. In particular, the failure of the
diagonal constructions means that the unsolvability of the tae-equivalent to
the halting problem cannot any longer be proved and in fact does not hold.
On the other hand, fixpoint theorems turn on the recursive character of the
relevant p arts of the theory of Turing machines which makes them a fortiori
subject to the theory oftae-computability.
6. GENERAL TIIEORETICAL REPERCUSSIONS
What remains is to examine some of the philosophical and other theoretical

consequences of the theory of tae-computability. For instance, what has
been found in this paper shows the need of distinguishing from each other
two kinds of questions. On the one hand, there are questions of the
constructive character of the models of a sentence, for instance the
conjunction of the axioms of a finitely axiomatizable theory. On the other
hand, there is the question of the recursivity of the set of sentences true in
the models of a finite axiom system. We can in some important cases
construct effectively a model for the system while being unable to effectively
enumerate the sentences true in that modeL This shows that there are
other ways of studying structures (models) than to prove theorems about
them.
For instance, we can purely mechanically construct a model (the
standard model) for (first-order) Peano arithmetic, even though we cannot
recursively enumerate all arithmetical truths. Thus to prove theorems in a
formal axiom system is not the only "purely mechanical" way in which we
can come to know a structure (e.g. the structure of natural numbers) or a
class of structures.
For instance, I can arguably come to know the consistency of elementary
arithmetic in a purely mechanical fashion by showing how to construct a
model for it. The construction can be carried out by a Turing machine. It is
the possibility of such a construction and not an intuition that is in reality
what is likely to convince you and me of the consistency of Peano

arithmetic.
This implies that all arguments to the effect that the human mind is not
a mechanism that are based on Godel's theorems must be reconsidered.
They all assume in effect that the only way of coming to know an
arithmetical truth mechanically is to prove it in a formal system of
arithmetic.
Another class of arguments that turn out not to hold water are those
which rely on the identification of mechanical determination with
recursivity. For instance, in Godel's obituary, Kreisel (1980, p. 216) in affect
identifies being mechanical as being possible to simulate by a digital
computer. It is' not clear what the force of the term "mechanical" is supposed
to be here. What is clear is that no inference from nonrecursivity to
indeterminism can be valid, for a nonrecursive function can still be
mechanical in that it is tae-computable. For instance, assume that the
solution of a differential equation describing the behavior of a physical
system is a nonrecursive function. On the basis of this fact alone, one
cannot conclude that the system in question is indeterministic.
There is a third conclusion that can be drawn here. It is a corollary to the
failure of diagonal arguments for tae-computability. This conclusion is that
there is no incompatability between computability determinism of the kind
that has been argued for e.g. by Roger Penrose (1989), (cf. also Calude et.
al., 1995). The argument is in effect that if I could determine
computationally the future course of events, I could use the results of such a
computation to change that course. But if the computability in question is
tae-computability, then I will not always know when the computation has
reached its definitive value. I cannot therefore use the eventual result of the
computation as a basis of action. Since tae-computation is deterministic,
there is no incompatability between computability and determinism.
NOTE
1 It appears that a quantum computation could in principle carry out all the
alternative constructions (calculations) at the same time (see Lloyd 1995). If so,
quantum computation could in principle facilitate (speed up) tae-computations
significan tly.
REFERENCES
Calude, Cristian, et al., 1995, "Strong Determinism vs. Computability", in W.

DePauli-Schimanovich et al., editors, The Foundational Debate, Kluwer
Academic, Dordrecht, pp. 115-13l.
Church, Alonzo, 1936, "An Unsolvable Problem in Elementary Number Theory",
American Journal of Mathematics vol. 58, pp. 345-363.
Davis, Martin, editor, 1965, The Undecidable, Raven Press, Hewlett, N.Y.
Hintikka, Jaakko, 1996, The Principles of Mathematics Revisited,. Cambridge U.P.,
xii + 288 pp.
Hintikka, Jaakko, 1995, "What Is Elementary Logic?' in K. Gavroglu et al.,
editors, Physics, Philosophy, and the Scientific Community, Kluwer Academic,
Hintikka, Jaakko, 1956, "Identity, Variables and Impredicative Definitions",
Journal of Symbolic Logic vol. 21, pp. 225-245.
Hintikka, Jaakko, and Gabriel Sandu, 1995, "What Is the Logic of Parallel
Processing?', International Journal of Foundations of Compu,ter Science vol. 6,
pp.27-49.
Hintikka, Jaakko and Gabriel Sandu, forthcoming, Computability and Logical
Theory.
Kreisel, Georg, 1980, "Kurt Godel 1906-1978", Biographical Memoirs of Fellows of
the Royal Society vol. 26 (December 1980), pp. 148-224.
Kreisel, Georg, 1953, "A Note on the Arithmetical Models for Consistent Formulae
of the Predicate Calculus", Proceedings of the XI International Congress of
Philosophy vol. 14, Amsterdam, pp. 37-47.
Lloyd, Seth, 1995, "Quantum-Mechanical Computers", Scientific American,
October 1995, pp. 140-145.
Matiyasevich, Yuri V., 1993, Hilbert's Tenth Problem, MIT Press, Cambridge.
Mostowski, Andrzef, 1965, Thirty Years of Foundational Studies, (Acta
Philosophica Fennica vol. 17), Societas Philosophic a Fennica, Helsinki.
Mostowski, Andrzej, 1955, "A Formulae with no Recursively Enumerable Model",
Fundamenta Mathematicae vol. 42, pp. 125-140.
Penrose, Roger, 1989, The Emperor's New Mind, Oxford University Press, New
York.
Putnam, Hilary, 1965, "Trial and Error Predicates and the Solution to a Problem
of Mostowski", Journal of Symbolic Logic vol. 30, pp. 49-57.
Rogers, Hartley Jr., 1967, Theory of Recursive Functions and Effective
Computability, McGraw-Hill, New York.
Turing, Alan, 1936, "On Computable Numbers, with an Application to the
Entscheidungsproblem", Proceedings of the London Mathema.tical Society, Ser.
2, vol. 42, pp. 230-265; vol. 43, pp. 544-546
188 LANGUAGE, TRUlH AND LOGIC
Wang, Hao, 1974 From Mathematics to Philosophy, Routledge & Kegan Paul,
London.
10
WHAT IS THE LOGIC OF PARALLEL PROCESSING?*
Received 16 June 1993

Revised 2 December 1994
Communicated by Barry Richards
ABSTRACT
We can associate with each consistent formula F of first-order logic a computing de-
vice as its representation. This computing device is one which will calculate the Skolem
functions of F (for a denumerable domain). When two such devices are operating in
parallel, the resulting architecture does not necessarily represent any ordinary first-order
formula, but it will represent a formula in independence-friendly (IF) logic, which hence
can be considered as a true logic of parallel processing. In order to preserve repre-
sentability by a digital automaton (Turing machine), a nonstandard (constructivistic)
interpretation of the logic in question has to be adopted. It is obtained by restricting
the Skolem functions available to verify a formula F to recursive ones, as in the GOdel's
Di,decticG interpretation.
1. Two Different Kinds of Advantages of Parallel Processing

In the last couple of decades, computer scientists' attention has switched from
sequential to parallel processing. It is not completely clear, however, what precisely
the rationale of this change is supposed to be. By and large, one can distinguish
two different trends. On the one hand, there is the connectionist idea of relying on
models which involve a large number of units each with a rather simple mode of
operation. The advantages of parallel processing are then thought of as being due
to the independent interaction of these simple units. The precise structure of the
complex of units and the precise manner in which they are related to each other do
not play much of a role in the superiority of parallel processing according to this
view.
This connectionist idea has to be distinguished, on the other hand, from the
idea that the superiority of parallel processing is essentially structural, perhaps
manifesting itself already in -computer architectures of certain relatively modest
minimal complexity.
The possible advantages of massive parallel processing have been discussed ex-
tensively in the literature. 1 In contrast, very little has apparently been accomplished

by way of understanding the structural advantages of parallel processing. Of course,

there is no a priori reason why parallel processing could not have both kinds of
advantages over sequential processing. However, it would be useful, for theoretical
as well as practical purposes, to have some idea of the relative importance of both of
these two kinds of advantages. That would also help us in choosing our conceptual
tools. If the advantages of parallel processing lie in the interaction of a large number
of simple units, the appropriate conceptual tools are (roughly speaking), statistical,
whereas structural comparisons are better studied in logical and algebraic terms.
In this paper we look for a new principle which allows us to compare the struc-
tural advantages of different computer architectures with each other. Several such
principles can be envisaged. For instance, one can think about the class of func-
tions that can Be computed by an automaton. But this is not a useful basis of
comparison between linear and parallel processing, for every recursive function can
be computed by a simple universal Turing machine. Or, one can think about the
time, or the space (memory) of computation. But this measures the computational
complexity of an automaton rather than its structural complexity.
The principle we propose here consists of correlating with each consistent formula
of first-order logic a class of computing devices. For simplicity, we can think of a
first-order language which is applied to elementary arithmetic, so that the values
of all the individual variables are natural numbers, even though this feature is
not essential. Roughly speaking, we correlate with the formula the class of such
computing devices that can calculate the Skolem functions which verify the formula
in question. It turns out that, in order to carry out this correlation, we have to
modify the interpretation of first-order logic in a certain way which in fact admits
independent motivation.
Once the correlation is established, we can discuss the properties of a number
of different kinds of computer architecture by examining the logical properties of
the formulas correlated with them, for instance in terms of their logical strength.
Such comparisons clearly have some general interest, especially if we think of the
computational operations as somehow representing such human reasoning as can be
expressed in terms of first-order languages.
In this paper we will restrict ourselves to only one particular correlation between
logical formulas and computer architecture .. This correlation appears to be a new
one. Relying on it we will show that computer architecture which incorporates
parallel processing is more powerful than linear architecture in the sense that the
logical formulas correlated with the former kind of architecture are sometimes logi-
cally much stronger than any other formulas correlated with sequential architecture.
To what extent this answers our question, we leave to our readers to decide.
2. First-Order Sentenc!'ls and Their Second-Order Translation

How can different first-order propositions be represented by computing devices?
A part of the answer is easy and old. Ever since the prehistoric era of computer
science, everybody and her brother has been told how different truth-functions
WHAT IS THE LOGIC OF PARALLEL PROCESSING? 191
(propositional logic formulas) can be represented by switching circuits, i.e. physical

devices which obey the rules of Boolean algebra. In case that electronic rather
than mechanical switches are used, then we are in the more general case of digital
electronic circuits.
As soon as we go beyond propositional logic, however, the situation changes.
There is no equally clear answer available in the literature as to how the formulas
of first-order logic (quantification theory) can likewise have an instructive represen-
tation in terms of computer architecture.
The first main point we are making in this paper is to offer a simple way of repre-
senting first-order formulas in terms of computer architecture. Our answer is based
on an important connection between existential quantifiers and choice functions
(Skolem functions). An example will bring home the basic idea.
For statement like
(V'x}(3y)F[x, y] (2.1)
to be true, there must exist a function I which, given the argument x, will as its
value yield I(x) such that
(V'x}F[x,/(x)] (2.2)
is true. The idea we are implementing here is that (2.1) is to be represented by a
device which can compute a suitable I. The naturalness of this is seen by observing
that, if the axiom of choice is assumed, (2.1) is equivalent with the second-order
sentence
(3f)(Vx)F[x, I(x)] (2.3)
which asserts the existence of the Skolem function I.
A similar second-order translation can be defined for all sentences of any given
first-order language. Let us suppose that we are given a sentence S in such a
language. For simplicity, we shall assume that S contains only the connectives "","
and V, and that it is in a negation normal form, that is, that all its negation-signs
are prefixed to an atomic formula. Then a second-order translation S of S can be
formed as follows:
(i) For each existential quantifier (3Yj) which occurs in the scope of the universal
quantifiers (V'Xl)(V'X2) ... (V'x,) (and of them only), we omit (3Yj) and replace
all the occurrences of the variable Yi bound to it by
(2.4)
where Ii is a new function symbol, different for different quantifiers (3Yi).

(ii) For each disjunction
(2.5)
occurring in S within tlie scope of the universal quantifiers ('VX1)(V'X2) ... (V'x,)
(and of them only), we introduce a new function symbol
(2.6)
different for different occurrences of V and different from the functions (2.4).
The disjunction (2.5) is replaced by
(iii) All the universal quantifiers are moved to the beginning of the formula.
(iv) The resulting formula is prefixed by
(3/d(3h) ... (3gd(3g 2 ) . , (2.8)
where h, 12, ... are all the function symbols introduced in (i), and 911 92, ...
all the symbols introduced in (ii).
S' is clearly equivalent with S. The functions I; are known as the Skolem
functions of S. We shall call the functions Ii and 9k collectively as the choice
functions of S.
For instance, consider the following formula:
(\t'x)(3y)(\t'z)(3u)(Fdx, y, z] V F2 [y, z, u]) , (2.9)
where Fl and F2 do not contain quantifiers or disjunctions. It is equivalent with
(3/d(3/2)(3g)(\t'x)(\t'z)((Fdx, hex), z]
"9(X, z) = 1) V (F2[h(x), z, hex, z)]" 9(X, z) f. 1. (2.10)
The functions h, 12, '" ,91,92, '" playa crucial role in our automaton-
theoretical interpretation of first-order formulas. Their central role in first-order
logic is nevertheless no novelty. Ever since Hilbert,2 perceptive logicians have re-
alized that the secret of first-order reasoning lies in its tacit reliance on choice
functions like the /'s and 9'S in (2.10).
Now one simple way of thinking of the realization of logical formulas in com-
putational terms is to associate with any first-order formula S, a computing device
which can calculate those functions h, 12, ... ,91, 92, . .. as a figure in its second-
order translation. The formula S is then true if and only if there exists a computing
device which can implement the functions h, 12, ... ,91, 92, ... , in the sense that it
can calculate functions whose values satisfy the quantifier-free part of S. Thus what
S will then assert is the existence of a computing device for its choice functions.
Whether or not such functions exist in a given model depends of course on the
model. Hence to assert the existence of functions like It, 12, ... , 91, 92, ... makes
a statement about the model, just like any old sentence of an interpreted first-order
language.
We can think (and speak) of the computing device associated in this way with
a logical formula as representing or modeling the logical relationships expressed by
the formula.
WHAT IS THE LOGIC OF PARALLEL PROCESSING! 193
Some interesting suggestions ensue from what has been said in this section. One
of the sources of philosophical interest in the old switching circuit representation
of propositional logic is that it apparently allows a correlation between certain
semantical distinctions, for instance between disjunctions and conjunctions, and
structural differences in the hardware, for instance the differenc:e between parallel
and sequential gates. By seeing which kind of gate is activated, one can so to
speak see whether the computer has in mind a disjunction or a conjunction. The
same possibility of correlating semantical distinctions and differences in hardware
is extended here in principle to first-order logic.
3. Computational Realizability Presupposes a Non-Classical

Interpretation of First-Order Logic
What kind of interpretation of first-order logic do we obtain by means of the
correlation we have explained? Obviously, we obtain the classical interpretation if
and only if each consistent sentence has as its representation an actually realizable
computer architecture. The most natural interpretation of this realizability is ob-
viously computability by a Turing machine. This means restricting the range of
second-order choice function variables to recursive functions.
What happens when such a restriction is imposed? The rlestriction means a
non-classical reinterpretation of our initial first-order logic. In order to carry out
the correlation between logical formulas and computing devices that we are using
here, we must equate a given first-order formula S and the second-order formula S
which asserts the existence of the Skolem functions of S. If thl~ meaning of S is
changed by restricting the range of the existential function quantifiers to recursive
functions, the meaning of S changes, since we insist on the equivalence of S and
S. At first sight, the price to be paid for such a reinterpretation is a high one.
The price one pays is to have to adopt a nonstandard interpretation of first-order
logic. In reality, however, this reinterpretation merely removes an obstacle from the
way of our correlation between first-order formulas and computing devices. In the
unrestricted case we will have consistent first-order sentences describing processing
units which cannot be realized by any Turing machine. That this is so follows from
the well-known result that there are first-order sentences which cannot be satisfied
by any interpretation on which its choice functions are recursive or even recursively
enumerable. 3 In other words, there are consistent logical formulas that cannot be
correlated with any possible logical device. Then the correlation between first-order
formulas and certain actual computing devices could be only a partial one.
In order to see how this anomaly could be removed, it is useful to look at the way
we correlated computing devices with first-order formulas, or their Skolem equiv-
alents. There we started with a device which computed a func:tion f satisfying a
certain condition F. But if -this device is a Turing machine, then it can compute
only recursive functions. Thus it is natural to require that in the formula corre-
lated with the computing device in question the choice functions are restricted to
recursive ones. We generalize this requirement to every choice function appearing
in the Skolem translation of a first-order formula and then face the consequences.
The good news is that on this (nonstandard) interpretation every satisfiable first-
order formula will have a computer-theoretical meaning in asserting the existence of
an actually realizable computing device (realizable in the same sense as all Turing
machines are realizable). The bad news is that we then are dealing with a non-
standard first-order logic. But there need not be anything wrong with this kind of
nonstandardness. In any case, it has respectable precedents. What we are doing
is to (re}interpret first-order formulas via their second-order translation, such as
(2.8). When the range of the choice function variables needed in the second-order
translation of first-order formulas is restricted to recursive functions, then our inter-
pretation of first-order logic is a simplified version of the interpretation exemplified
by GOdel's famous Dialectica interpretation of first-order logic and arithmetic. 4 On
our interpretation, as on GOdel's, each formula F is taken to be equivalent to its
second-order translation. The second-order translation simply asserts the existence
of Skolem functions for F. It has the form
where i:$ m, j:$ n, {XiI' xh, , .. } ~ {Xl, X2, ,., ,xn }.

The crucial step in our interpretation, as well as in GOdel's, is to restrict the
range of the function quantifiers in (3.1) to recursive functions. But, unlike GOdel,
we do not use a nonstandard interpretation of conditionals, which prejudges his
particular interpretation of negation which is different from ours. 5 FurthermOl:e,
unlike GOdel, we do not go beyond the second-order leve1. 6
We shall call our interpretation simply the functional interpretation of ordinary
first-order logic and ordinary first-order languages. This interpretation has not
been studied extensively. It differs from the normal, "standard" interpretation
(i.e. the interpretation on which there is no restriction whatsoever on the range
of the function variables), but it is not easy to see by how much. It has, in any
case, several interesting features. It is known that the logic resulting from this
interpretation is not axiomatizable, that is, the class of all formulas true in all
models is not recursively enumerable. 7 On the plus side, it is known that by means
of this logic it is possible to present a descriptively complete axiom system for
elementary arithmetic. 8
In any case, the logical relations between first-order sentences will be mirrored by
certain relations of processing power between the corresponding units. For instance,
if
(YX)(3y}Cdx, y] (3.2)
logically implies
(3.3)
(no quantifier in C 1 [x, y] or C2 [x, yD, then the processing power of the unit corre-
lated with (3.2) is in an obvious sense greater than that correlated with (3.3). Thus
the processing power is in an obvious sense measured by the logical strength of the
correlated first-order statement.
4. Logical Formulas Corresponding to Subunits of Processing

The correlation between first-order formulas and certain kinds of computing de-
vices which was set up in the preceding section does not yet do the job we set out
to do. It assigns a computing device to any first-order formula. What we have to
do now is to see how the correlation works the other way around, i.e. whether and
if so a computing device can be assigned to a given first-order formula. Obviously,
this correlation depends on the structure of the computing device which can, for
instance, instantiate either sequential or parallel processing. How are such distinc-
tions reflected in the structure of the logical formula (if any) which is correlated
with the computing architecture having one or the other structure?
In order to answer this question, consider more generally a distributed computing
network. It consists of certain processing elements called units. A logical formula
can be correlated with each of them. However, they can also interact with each
other. Because of this interaction, we can also try to correlate a formula with
the entire network. How can it be built up? We are not interested to describe the
internal operations of the units, but the structural complexity of the whole network.
For this purpose, it is enough to take the unit as the proverbial black box. All that
it does is to take an input x and transform it into an output y. What x and y are
does not matter. The values of these variables are simply all the relevant kinds of
inputs and outputs. We can think of them as numerical variables, but our line of
thought is not predicted on that assumption.
We shall assume, however, that processing units are all homogenous, i.e. that
the output of a unit can always be the input of another unit. We shall also assume
that certain conditions can be imposed on inputs and outputs, i.e. certain things
can be said of them. If the inputs and outputs are numerical, then the conditions
on them will be suitable numerical relations that are usually E'JCpressed by means
of certain basic relations and their Boolean combinations. In other words, we shall
assume that we are given the nonquantified part of a first-order language to describe
the inputs and outputs.
What a unit U does can now be described as follows:
It takes an input x (or several inputs, xl. X2, , x/c) and produces from it (or
them) an output y (or several outputs Yl, Y2, ... ,YI) related to the inputs in the
way expressed by the condition
F[x, y] (4.1)
or, more generally, by the condition
.F[Xl' ... ,X/c, Yl, ... ,yd. (4.2)
In other words, a unit U computes a function f which for any input x, satisfies the
condition
F[x, f(x)] (4.3)
or, more generally, computes the functions 91, ... ,91 which satisfy the condition
(4.4)
Then what the unit U does can be "expressed" by the first-order statement
('v'x)(3y)F[x, yj (4.5)
or, alternatively, by its Skolem from
(3f)('v'x)F[x, lex)]. (4.6)
Similarly, in the general case, the unit U can be associated with the first-order
formula
(4.7)
or, alternatively, with its Skolem form
That we are dealing so far with abstract specifications of what a unit might ~e
able to do and not with actual algorithms for computation is reflected by the fact
that in (4.6)-(4.8) we have (existentially quantified) function variables rather than
specific functions.
This correlation between computational systems and first-order statements is in
the spirit of the explanation by J. L. McClelland et al., that "in some cases, the
units (of a PDP model) can stand for possible hypotheses ... ".9
5. Parallel Processing Mirrored in Logical Formulas

All that was said in the preceding section is essentially that logical formulas
can be correlated not only with entire processing devices, but also with the units
(subunits) of which these computing devices consist. But what are the logical ex-
pressions which correspond to complex computing devices built from simpler units?
At this point we encounter a major difference between propositional logic and
first-order logic. If one combines in any way whatsoever, either sequential or in
parallel, two computing devices representing truth-functions, one obtains a device
which likewise represents a truth-function. That is, the correspondence between a
computing device calculating truth-functions and formulas of the propositional cal-
culus obtains both ways. This correspondence does not hold for the representation
of ordinary first-order logic -explained above. To any formula of first-order logic, a
computing device can be associated. However, the association does not work the
other way around, especially if we consider a network of processing units and not
only simple ones as in Sec. 4 above. As one can easily see, when two such simple
units each correlated with a first-order formula are combined sequentially, the result
is still the correlate of a first-order formula determined easily on the basis of the
given ones. However, when they are combined in the parallel fashion, the resulting
computing architecture can no longer be associated with any ordinary first-order
formula. It will be shown here that it nevertheless can be correlated with a logical
formula in a stronger but natural logic. This logic is the independence-friendly (IF)
logic which has been studied elsewhere by Hintikka and Sandu.
Pursuing this agenda, consider the system which consists of two sequentially
connected units, as follows:
x -IUd- y = hex) Y -lu 1- z = hey)

2
(Vx)(3y)Cdx, y] (Vy)(3z)C2[y, z] (5.1)
(3fd(Vx)Cdx, !l(x)] (3h)(Vy)C2[Y, hey)]
The sequential combination corresponds to the statement
(3fd(3h)(vx)(Cdx, h(x)] " C2 [h(x), h(!l(x)]). (5.2)
This second-order statement has a first-order equivalent, viz.
(Vx)(3y)(3z)(Cdx, y]" C2[y, z]). (5.3)
But consider now a system consisting of two parallel units U1 , U2 Its output will
be an ordered pair of numbers (or other objects of computation) (y, u) obtained
from two separate inputs x, z. The four objects x, y, z, u will satisfy a certain
condition Co[x, y, z, u]. The situation can be represented as follows
(Vx)(3y)C1 [x, y] = (3fd(Vx)C1 [x, h (x)]

x-IUd-Y
Co[x, y, z, u] (5.4)
/
z ..... IU 1..... u
2
Here Co may be taken to be implemented by a third unit, perhaps (if Co is

quantifier-free) a device for calculating the Boolean truth-value of Co!x, y, z, u]
(depending on the values of x, y, z, u).
Then what the combined system Uo does can obviously be described by the
following second-order formula
(3/d(3h)(V'x)(V'z)Co[X, hex), z, h(z)]. (5.5)
But it is known that statements of the form (5.5) are not, in general, first-order
definable, not even if we restrict Co[x, y, Z, 1] to arithmetical relations. What (5.5)
says can be expressed by the following branching-quantifier expression 10
{ V'x 3Y } Co [x, y, z, 1] (5.6)

V'z 31
whose interpretation is given by
{
V'x 3Y } Co[x, y, Z, 1] > (3/tl(3/2)(V'x)(V'z)Co[x, hex), z, h(z)]. (5.7)
V'z 31
Equation (5.5) can also be expressed in the IF first-order logic. l l The basic elements
of this logic are described below.
6. IF First-Order Logic
An idea which goes back to Henkin (1959) is to interpret sentences in prenex
form like
(V'x)(V'z)(3y)(31)C[x, Z, y, 1] (6.1)
by a game G6.1), M) of perfect information played on a model M in which (6.1) is
interpreted. The game is played by two players, Nature who tries to show that (6.1)
is false in M, and Myself who tries to show that (6.1) is true. In the game, Nature
chooses a and b from M corresponding to the two universal quantifiers, and Myself
chooses c and d corresponding to the existential quantifiers. If M F CIa, b, c, dj,
then Myself wins the game. Otherwise Nature wins the game. Equation (6.1) is said
to be true in M if and only if Myself has a winning strategy in the game G6.1), M).
In this case, the winning strategy of Myself reduces to two functions and I and g.
Thus (6.1) is true in M if and only if M F (3f)(3g)(V'x)(V'z)C(x, z, I(x, z), g(x, z)].
Later on the game-interpretation was extended by Hintikka to the truth-
functional connectives. The subtle case is that for negation. The two players
are now assigned roles, so that Myself has, at the beginning of a game, the role
of the verifier, and Nature that of the falsifier. Negation prompts a switch of the
roles. The universal quantifiers and conjunctions prompt moves by the falsifier (in
the case of a conjunction, the falsifier chooses one of the conjuncts)j the existen-
tial quantifiers and disjunctions prompt moves by the verifier (for a disjunction, the
verifier chooses a disjunct). The rule for winning the game is slightly changed, i.e. if
M F CIa, b, c, dj, then the verifier winsj otherwise the falsifier wins. The definition
of truth in a model remains the same.
It is well-known that the game-theoretical interpretation and the Tarski-type

interpretation of first-order logic are equivalent. 12
Now another generalization of the game-interpretation is to allow for games of
imperfect information. 13 Thus we could play the game associated with (6.1) in such
a way that the verifier, when choosing c, does not know the b chosen by the falsifier,
and when choosing d, does not know the a chosen by the falsifier. In this case,
the verifier's winning strategy in the game will consist of two functions I and 9 of
smaller arity, such that M 1= C[x, z, I(x), g(z)].
In order to mark the informationally independent moves in the syntax, we
shall have, in addition to all the well-formed expressions of the usual first-order
logic, expressions which contain occurrences of logical constants involving a slash:
(3x/'1Yl!'" ,'1YIe), (V/'1Yb'" ,'1YIe), ('1X/3YL,'" ,3YIe), (/l.f3YL,'" ,3YIe)'
For instance, the intended meaning of
('1x)('1z)(3y/'1z)(3u/'1x)Co[x, z, y, u] (6.2)
is that the existential quantifier (3y) is not in the scope of the universal quantifier
('1z), and the existential quantifier (3u) is not in the scope of the universal quan-
tifier ('1x). That is, (6.2) will be interpreted by the game of imperfect information
described above.
More generally the formulas containing occurrences of the slash will be inter-
preted by games of imperfect information. The rules of the game will be identical
with the previous ones. The only essential difference is in the definition of a strategy.
As usual, truth of a sentence in a model M is defined as the existence of a
winning strategy for Myself in the semantical game associated with that sentence
played in M. Accordingly, the truth of (6.2) will be expressed by the second-order
sentence
(3f)(3g)('1x)('1z)Co[x, I(x), z, g(z)] , (6.3)
and the truth of e.g.
('1x)('1z)(3y)(Co[x, z, y](v/'1x)Cdx, z, y]) (6.4)
is expressed by
(3f)(3g)('1x)('1z)CJ(%)[x, z, g(x, z)], (6.5)
where I is a function from the Cartesian product of the universe of the model to
the set {a, I}.
An interesting fact about this new logic is the consequence of the combination
of the treatment of negation as role swapping and of informational independencies
between the logical constants.
One such consequence is described by the following theorem
Theorem. For any IF first-order sentence 8 and model M
(i) Myself has a winning strategy in G( ...,8, M) iff Nature has a winning strategy
in G(8, M).
(li) Mysell has a winning strategy in G(S, M) iff Nature has a winning strategy in
G(-.S, M).
From the above Theorem, it follows that the law of the excluded middle becomes
in IF first-order logic the principle of the determinateness of games. This principle
=
fails in this logic. For instance, neither (Vx)(3y/Vx)x Y nor -.(Vx)(3y/Vx)x Y =
is true in any model M which has at least two elements. 14
Thus negation in IF first-order logic is not the classical contradictory negation.
It is, however, a strong, dual negation which satisfies de Morgan's laws. This can
be seen in the following way. Let * be the following mapping which maps each IF
first-order sentence S into its dual S*
S* = -.S, S atomic;
(-.st = S*;
3X/VY1, ... ,VYIe)F)* = (Vx/3Y1I ... ,3YIe)r ;
Vx/3Ylt ... ,3YIe)Ft = (3x/VY1I ... ,vYIe)r ;
(F(v /VYlt ... ,VYIe)Gt = (r(A/3Y1I ... , 3YIe)G*);
(F(/\/3Y1'''' ,3YIe)Gt = (F*(V/VY1I'" ,VYIe)G*).
We call S* the dual of S. Notice that S* is in negation normal form. The next
theorem is straightforward from the definitions
Theorem. For any IF first-order sentence S and model M
Using the last theorem we can encode any IF first-order formula of a language
L into what is known as E~-formula over L (Le. a formula of the form 3h I..<p,
where <P is an ordinary first-order formula in the language L U {h, ... , In}).
The E~ -formula S(2) equivalent with the given IF first-order one S can be
effectively formed as follows
(i) Replace S with its dual S*.
(li) In S drop all the logical constants of the form (VX/3Y1, ... , 3YIe) and (A/3Y1,
... ,3YIe)' Call the result S(l).
(iii) In S(l), perform the following two operations: (a) replace each existential
quantifier of the form (3X/VY1, ... , VYIe) with (3/), and each occurrence of the
variable x with l(z1I .... , zm}. Here Z1I'" ,Zm are the universally quantified
variables within whose scope the quantifier (3X/VY1, ... , VYIe) occurs except for
Y1I ... ,YIe; (b) replace every disjunction of the form (Sl(V/VY1, ... ,VYIe)S2)
by Sl /\ g(z1I ... ,zm) = 1)V(S2 /\ g(z1I ... ,zm) = 2, where Zl, ... ,Zm
are the variables bound to those universal quantifiers within whose scope
(v IVY1, ... ,"lYle) occurs, except for Y1, ... ,YIe. Also 9 must be a new (dif-
ferent) function symbol. Furthermore, S(l) is prefixed by (3g).
The result is the L~-formula S(2) we are looking for.
The fact that S is equivalent with S follows from the first Theorem above. The
equivalence between S and S(l) follows from the fact that the strategy of Nature
does not matter in the truth-definition. The equivalence between S(l) and S(2)
follows from the definition of a winning strategy.
The encoding of the IF first-order logic into L~ -logic has the consequence that
we get for the former all the metalogical properties of the latter: Compactness,
LOwenheim-Skolem Property, and Separation Property.
7. The Advantages of Independent Quantifiers

The important question to be raised is: Can informationally independent quan-
tifiers do any work that ordinary ones cannot do? The answer is definitely yes. The
uses of IF first-order logic in the foundations of mathematics will be discussed by
Jaakko Hintikka in his forthcoming work. There it is shown that there is an in-
teresting subdiscipline of combinatorial theory where informationally independent
quantifiers are used routinely, albeit tacitly. This subdiscipline is known as the
Ramsey theory.1S
By allowing a processing system to consist of parallel subsystems, we can express
stronger requirements that cannot be expressed in terms of sequential architecture.
This means being able to represent through actual or possible processing systems
logical statements whose expressive power goes beyond that of ordinary first-order
logic. In other words, we can thus represent by means of possible parallel processing
systems modes of thinking and reasoning which are beyond the purview of first-order
logic.
It is important to be clear about what is involved here. The advantages which
the theory of parallel processing offers as a modeling device over, say, the theory of
Turing machines do not lie in the fact that parallel processing devices offer possi-
bilities of computating functions which Turing machines cannot compute. What is
involved is that the analysis of computational information processing through Tur-
ing machines does not enable a theorist to analyze the computational processing in
the right way, i.e. in a way which would facilitate the modeling of non-elementary
logical relations by means of computational processing systems.
8. Conditions of Reducibility
The combinatorial character of the reasons for preferring parallel processing as a
modeling device can be seen "especially clearly from the general conditions on which
a second-order (L~ _) statement of the form
where each set of arguments {Xi!, Xi2 ... } is a su bset of the set {Xl, X2 ... }, can be
reduced to a usual first-order form. l6
It can in fact be shown that (8.1) reduces to a first-order form if and only if
the functions It, h, can be ordered in a suitable way. Assuming that It, h, ...
itself is the order, a sufficient condition is simply that
{Xi" Xi 2 ... } ~ {Xi+l,l, Xi+2,2, ... }. (8.2)
In other words, this condition requires that the arguments of the earlier functions
are always included among the arguments of later functions, thus enabling the
computation to proceed linearly.
In applying this sufficient condition, it can be observed that whenever an exis-
tential quantifier (3Yj) occurs outside the scope of a universal quantifier ('Ix;) in the
original given first-order sentence, then the choice of the values of Xi on the part of
Nature does not affect Myself's choice of the value of Yj in the first place. In such
circumstances, Xi can be added to the arguments of the function!; corresponding
to (3Yj). This extends the range of applicability of the sufficient condition (8.2).
The reason why the sufficient condition of reducibility (8.2) does not, as such,
yield a necessary condition is that even when (3Yj) occurs within the scope of (\t'Xi)
in a given first-order sentence, the value of the choice function Ij associated with
(3Yj) need not really depend on Xi. A simple example of such merely apparent
dependence is the logical truth of
(\t'x)(3y)(A(y) -t A(x (8.3)
which has the choice function representation
(3f)(\t'x)(A(f(x -t A(x (8.4)
but also the simpler one

(3y)(\t'x)(A(y) -t A(x. (8.5)
Such hidden independence occurs when the variables bound to the two quanti-
fiers neither occur in the same atomic formula nor are connected by a sequence of
bound variables whose successive members do so occur. Thus we can transform the
sufficient condition (8.2) into a necessary and sufficient one.
Independently of this residual qualification, the combinatorial condition (8.2) is
especially suggestive in that a violation of the ordering requirement in effect means
that the computations codified in the functions h, are done partly in parallel. Hence
the step from ordinary first-order logic which allows for quantifier independence is
connected very closely with the very idea of parallel processing.
9. Independence-Friendly Logic as a Logic of Parallel Processing

These observations prompt a few interesting conclusions. First, ordinary first-
order logic is not adequate as a tool in the theory of parallel processing. We need
a logic that is at least as strong as (has at least the expressive power of) the IF
first-order logic. Hence, since parallel processing is highly important in AI, the
logical tools needed in the study of AI must be stronger than first-order logic.
This observation is interesting in view of the widespread use of first-order logic
in AI, database theory, etc. 17
A second conclusion is prompted by the question: What is the natural logic that
corresponds to parallel processing if first-order logic fails, as we saw?
One possible candidate would be a fragment of second-order logic, for instance
the L~-logic in which the IF first-order is encodable. However, we know that this
logic is not closed with respect to classical (contradictory) negation, and in addition,
does not have any other natural concept of negation.
Another possible candidate would be the logic of partially ordered quantifiers,
or Henkin quantifiers as they are sometimes called. IS A Henkin quantifier G;:;n Xl ..
XnYI ... YmZI . zktl .. t/ denotes the generalized quantifier whose meaning is
defined by
when x and z denote and abbreviate the sequences Xl .. Xn and ZI ... Zk respec-
tively. That is, in the branching notation illustrated above, G;';"~'I ... XnYI .. YmZI
.. Zkti ... t/ can be rewritten as
(9.2)
It is easy to see that the Henkin quantifier G;:;"XI'" XnYI .. YmZI .

zkti ... tz is definable in IF first-order logic
Again, we resist this alternative because it introduces an analyzable notion as

a primitive. In addition, although the logic with Henkin quantifiers has a well-
defined concept of negation, in fact, it is closed with respect to classical negation,
this logic lacks the metalogical properties of the IF first-order logic mentioned above
(i.e. Compactness, L6wenheim-Skolem and Separation Theorem).
We think that there is a more systematic way of building the requisite logic
starting from ordinary first-order logic. This way consists simply of a systematic
use of the independence (slash) notation explained above, while sticking to first-
order logic in all other respects. 19
Even though notationally an IF first-order logic is but a mild extension of the
ordinary first-order logic, it adds a great deal to the expressive power of our logic.
Indeed, it has been shown that the decision problem of the logic with Henkin
quantifiers, which is, as we saw above a fragment of the IF first-order lOgic, is
as difficult as the decision problem for the entire second-order logic with standard
interpretation).2o Hence the increase in logical power in the transition to an IF
logic is truly formidable. What is most important for our purposes is that the ex-
pressive power of the IF first-order logic is much greater than that of first-order
logic even if we restrict attention to finite models. This follows from the fact that
in the IF first-order logic we can express so-called NP-complete predicates, such
as the 3-colorability of a graph, the satisfiability problem for NAND circuits, the
disconnectedness of a graph, etc. 21 (d. next section for an illustrative example).
It may be instructive to relate our observations to the tasks of computational
modeling in general. For instance, in any computational level theory in the sense
of David Marr we have to discuss what the computational tasks are that a machine
or an organ has to perform. 22 In Marr's own words, we have to study "the logic
of the strategy by which it (the task of computation) can be carried out". The
correlation of certain statements in a formal language and certain computational
tasks that we have expounded can serve as a tool for studying in general terms the
logic of such strategies. By means of this correlation, we can express the kinds of
requirements or computational tasks which a computational level theory deals with.
The main thesis propounded here concerns the kind of logic that is adequate for
this task. What we are proposing is that as soon as parallel processing is involved,
any adequate logic must allow for quantifier independence.
In view of the naturalness of an IF first-order logic as a logic of computational
information processing, how come its role has not been recognized earlier? A partial
explanation lies in the fact that in natural language informational independence is
not marked at all syntactically. (The reasons for this fact are likely to be quite inter-
esting from a linguistic viewpoint. 23 ) Hence, when information processing, including
parallel processing, is discussed in normal English prose, the role of informational
independence easily escapes one's notice.
10. An Illustrative Application

In order to illustrate our observations, let us consider an application.
Consider an information processing system of which seven units and the flow of
information between them are sketched in the following picture.
~ al
'\.
a3
/
~ a2 '\.
a7 ~
~ a4 /
'\.
a6
/
~ a5
That is, what each unit does can be described by the following: it receives two
inputs coming from two other distinct units; after making its own computation,
it sends the output to another distinct unit. The output of the whole system is
computed by unit a7. We can think that the processing units al-a4 are functioning
"in parallel", and so do the units as and a6. We can now put some constraints
on the values of the inputs and outputs of the system: Each input (and output)
is either the truth-value True or False; the output computed by each unit has to
satisfy the truth-table for NAND (negated conjunctions).
The informational system described above may be thought of as a model M
for a language L in a signature consisting of a binary relation I and an individual
constant p. The universe of M is the set S of units: S :::: {aI, ... , a7, ... }, its
output unit a7 is the interpretation of p, and the pairs of units sending inputs to
one another are the interpretation of I (the intended meaning of l(x, y) is that the
unit x sends one of two inputs True, False to the unit y).
Suppose now that our informational system computes the truth-value True (at
the unit a7) for some arbitrary inputs. Then we claim that there is an IF first-order
sentence in the signature {I, p} which is "made true" by the system. The sentence
in question is
('Ix ) (\fy)(\fz){ {[Cooo(v j\fx\fy)COOI](V j\fx\fz)[COlO (v j\fx\fy)COll ]}
(10.1)
(v j\fy\fz){[CIOO(v j\fx\fy)C lOI ](V j\fx\fz)[CllO(v j\fx\fy)Cm ]}} ,
where
COOO : -.l(x, y) V -.l(y, z) V x = Y

COOl: y =f. z 1\ z =f. X
COlO: E : x =f. y 1\ Y =f. z
Call: x =f. Y 1\ z =f. x 1\ (-.Z(x, z) V -.l(y, z
C lOO : x =f. y 1\ z =f. x
ClQl: x =f. y 1\ Y =f. z 1\ (-.l(x, z) V -.l(y, z
CllO: y =f. z 1\ z =f. x
C 1l1 : (-.Z(x, z) V -.l(y, z 1\ x =f. p.
Now the fact that the system computes the truth-value True for some inputs
means that there is a function g* defined on the universe of M such that g* (p) =
True and which satisfies the following four sentences (the truth-table for NAND)
l(x, z) 1\ l(y, z) 1\ g*(x) :::: True 1\ g*(y) = True -+ g*(z) = False (10.2)
l(x, z) 1\ l(y, z) 1\ g*(x) :::: True 1\ g*(y) = False -+ =.True

g*(z) (10.3)
l(x, z) 1\ l(y, z) 1\ g*(x) :::: False 1\ g*(y) = True -+ g*(z) = False (10.4)
l(x, z) 1\ l(y, z) 1\ g*.(x) :::: False 1\ g*(y) = False -+ g*(z) :::: False. (10.5)
In order to show that (10.1) is true in the M model, we have to find three functions
J, g, h defined on the universe of M with values in the set {O, I} such that, for any
arbitrary a, b, c in M: M F C'(a)g(b)h(c)(X, y, z, pl.
We define f = 9 = h by
f(a) = 0 ..... g*(a) = True
f(a) = 1 ..... g*(a) = False
for every a in M.
There are eight cases. We consider only the case: f(a) = f(b) = f(c) = O.
We have to show that M F CJ(a)g(b)h(c)(a, b, c), i.e. M F Cooo(a, b, c). Suppose
l(a, b) 1\ l(b, c) 1\ a = b. Then, by the definition of f, g*(a) = g*(b) = g*(c) = True,
and by (10.2), g*(c) = False. Whence f(c) = 1, a contradiction. All the other cases
are similar.
Conversely, suppose (10.1) is true in a model M which can be thought of as
an informational processing system described above. We will show that there is an
assignment g* of truth-values to the units of the system which gives to p the output
True and which requires the NAND-gates to produce the correct outputs for any
inputs.
The truth of (10.1) means the existence of a winning strategy of Myself in the
semantical game associated with it. This winning strategy will consist of three
functions f, g, h : M -+ {O, I} such that for any a, b, c, E M, we have: M F
CJ(a)g(b)h(c)(a, b, c, p).
Claim 1. f = 9 = h
= =
Proof of Claim 1: Assume the contrary, i.e. a b c but f(a) =1= g(b) =1= h(c). In
this case the sequence (a, b, c) should satisfy one of the formulas: COOl, COlO, Call,
C lOO , C lOl , C llO , a contradiction.
Claim 2. f(p) = 0
Proof of Claim 2: If f(p) = 1, then given Claim 1, the sequence (p, p, p) should
satisfy C 111 which is impossible.
We now define the assignment g*
g*(a) = True ..... f(a) = O.
Let us show that g* respects the truth-table for NAND, i.e. (10.2)-(10.5) are satis-
fied. We prove (10.2): Let c be an arbitrary unit and a, b it inputs (thUS a =1= b) and
assume the antecedent of (10.2) is true. If the consequent of (10.2) is false, then
f(a) = 0, f(b) = 1 and f(c) = 1. But then Call shows that either a or b is not an
input to c, a contradiction. Equations (1O.3}-(1O.5) are proved in a similar way.
There is an alternative way to look at our example. We can think of the infor-
mational system considered here as what in computer science is called an electronic
(Boolean) circuit with NAND-gates. In such a case, we just showed that the prop-
erty "Decide whether such a circuit has the value True for some inputs" is expressible
in IF first-order logic. It is known that this property (called an N P-property in
complexity theory) is not expressible in ordinary first-order logic. 24
11. Independence-Friendly Logic Functionally Interpreted

An important further point remains to be made here. We have to combine the
results of Sees. 4 and 7.
In Secs. 6-7 we argued that independence-friendly first-order logic is a logic of
parallel processing. However, this thesis has to be understood in the light of what
was found in Sec. 5 above. In order to be a logic of parallel processing, IF first-
order logic has to be functionally interpreted, in other words, its choice (Skolem)
functions have to be restricted to recursive functions.
Ordinary first-order logic has a rich metatheory, much of which can be extended
to the IF first-order logic. 25 In contrast, relatively little is known directly about
the functionally interpreted first-order logic, and even less about functionally inter-
preted IF first-order logic, which we have argued is a logic of parallel computation.
This is not the place to try to develop a metatheory for it, either. However, we
must make sure of one thing, viz. that it really gives us something new, logically
speaking. For from the fact that IF first-order logic is a strong logic, much stronger
than ordinary first-order logic, it does not automatically follow that so is its func-
tionally interpreted brother. In other words, the question has to be raised whether
the advantages of an IF logic remain there also on the functional interpretation
which changes both ordinary first-order logic and its IF brother.
Even though the question undoubtedly deserves a deeper an.alysis, it is easily
seen that the answer is affirmative. An IF first-order logic is stronger than ordinary
first-order logic also when the functional interpretation is adopted.
One way of seeing this is by way of an example. Consider fi:rst-order formulas
whose only predicate is "=". It is easily seen that, on the standard interpretation
of such a formula, it is satisfiable in a finite model if is satisfiable at all. Hence
its Skolem functions are recursive, which means that it is satisfiable also on the
functional interpretation. Hence there is no difference between the logic of ordinary
pure identity formulas on the standard interpretation and on the functional inter-
pretation. In particular, by means of such formulas, one cannot express the infinity
of the domain.
On the other hand, consider the following IF first-order formula
(3w)(V'x)(V'z)(3yjVz)(3u/V'x)x = z ..... y = u) ..... (x f:. wand z f:. w)). (11.1)
It is satisfiable on the normal interpretation of IF first-order logic only in an infinite

domain (i.e. Eq. (11.1) is true in a model M if and only if there is a bijection from
M onto a proper subset of M). A fortiori, it can be satisfied,. if at all, only in
an infinite domain also on our nonstandard (functional) interpretation. Its second-
order translation is
(3w)(3f)(3g)(V'x)(V'z)x'= z ..... f(x) = g(z +-+ (x f:. wand z f:. w)). (11.2)
It can be seen that (11.2) is satisfied by taking w = 0, f(x) = g(x) = x + 1,

which is a recursive function. Hence (11.2) expresses the infinity of the domain also
in functionally interpreted IF logic, which shows that this logic is stronger than
functionally interpreted first-order logic. Thus what was said earlier of the greater
expressive power of IF first-order logic in comparison with the usual (independence-
free) one still holds after we have imposed the functional interpretation of both.
12. Conclusions
What have we shown? We have shown how to establish a correspondence be-
tween functionally interpreted first-order formulas and certain kinds of computer
architecture. We have shown that, in order to allow a general correspondence of this
kind which is closed with respect to parallel combinations of computational units,
we must extend the first-order language in question so as to allow informational in-
dependence of existential quantifiers and disjunctions of universal quantifiers. The
resulting functionally interpreted independence-friendly first-order logic is thus seen
to be a true logic of parallel processing. We might call this logic the functional IF
first-order logic.
Thus we also obtain one possible way of measuring the power of different com-
putational devices, viz. measuring it by the logical strength of the corresponding
formula. This formula has to belong to the functional IF first-order logic. Measured
in this way, suitable parallel processing devices turn out to be stronger than any
sequential ones.
Further study is needed to show how this measure is related to others, such as the
speed of computation. It looks as if there might very well be connections between
our measure of strength and the length of computation. For instance, whenever a
conditional
(12.1)
is logically true in ordinary first-order logic, there is a connection between the com-
plexity of a cut-free deduction of (12.1), the formal complexity of the interpolation
formula 1 between Fl and F2 , and the logical strength of 1.26
This answers one of our initial questions. There is in fact a structural reason why
parallel processing is superior to sequential processing. This reason is completely
independent of the statistical behavior of large-scale parallel distributive processing
systems. It can manifest itself already when three processing units are combined.
At the same time, it presupposes certain things of the internal structure of these
units. Hence the advantages of parallel processing are not only statistical.
These results can be deepened and developed further by studying the function-
ally interpreted IF first-order logic sketched above. From the (deductive) incom-
pleteness of this logic, it follows that such a study has always something new to
offer in principle.
One particular intriguing direction of further study is suggested by Hintikka's
results concerning the definability of truth for a suitable IF first-order language in
that language itself. This suggests that a suitable IF language can be self applied,
in the sense that it can be used to discuss its own semantics. The theoretical and
perhaps even practical advantages of so doing might be substantial.
Our use of logical tools and in particular the extension of ordinary first-order
logic to which we were forced throws some light on another controversial issue.
Some of the proponents of PDP as a tool of modeling human thought or as a tool in
AI have belittled the applications (and applicability) of logic for these purposes. If
by logic they simply mean customary first-order logic, we can see the point of their
criticisms. However, it seems that at least one underlying reason for the superiority
of parallel processing can only be spelled out clearly by reference to logic.
Notes
1. For a survey of results in this area, see e.g. the article of J. L. McLelland, D. E.
Rumelhart and G. E. Hinton, "The appeal of parallel distributed processing", in David
E. Rumelhart et al., Parallel Distributing Processing, vol. 1, Foundations, MIT Press,
Cambridge, 1986.
2. See, e.g. Warren D. Goldfarb, "Logic in the twenties: The nature of a quantifier", J.
Symbolic Logic 44 (1979) 351-68.
3. This result goes back to A. Mostowski, "On a system of axioms which has no recursively
enumerable model" , Fundamenta Mathematicae 40 (1953) 56-61, and "A formula with
no recursively enumerable model", ibid. 42 (1955) 125-140, and to G. Kreisel, "A note
on arithemetic models for constant formulae of the predicate calculus", in Proc. XIth
Int. Congress of Philosophy, vol. 14, Amsterdam and Louvain, 19113, pp. 39-49.
4. See Kurt Godel, "On a hitherto unexploited extension of the finitary standpoint", J.
Philosophical Logic 9, 133-142.
5. It will nevertheless turn out that negation behaves in certain respects in a nonclassical
way in the languages we will end up embracing.
6. The reasons why we can (and why Godel cannot) avoid the climb towards higher
orders are spelled out in Jaakko Hintikka, "Godel's functional interpretation in a wider
perspective", forthcoming in the Proc. 1991 Meeting of the Int. Godel Society. The
simpler functional interpretation proposed here is also mentioned there and commented
on.
7. See e.g. R. L. Vaught, "Sentences true in all constructible models", in J. Symbolic
Logic 25 (1960), 39-53.
8. This follows from the observation, first made by Stanley Tennenbaum in "Non-
archimedean models for arithmetic", Notices of the American Mathematical Society
6 (1959) 270, that the only model of Peano arithmetic where the interpretations of
sum and product are recursive is the standard one. For the entire range of questions
arising here, see Richard Kaye, Models of Peano Arithmetic, Clarendon Press, Oxford,
1991, especially 11.3.
9. Op. cit., p. 10.
10. For partially ordered quantifier structures, including branching quantifiers, see, e.g.
Leon Henkin, "Some remarks on infinitely long formulas", in lnfinitistic Methods,
Pergamon Press, Oxford, 1959, pp. 167-83 (the first paper on t.he subject) or Jon
Barwise, "On branching quantifiers in English", J. Philosophical Logic 8 (1979) 47-80
(with further reference).
11. Cf. Jaakko Hintikka, "What is elementary logic", forthcoming; Jaakko Hintikka and
Gabriel Sandu, "Informatiqnal independence as a semantical phenomenon", in Logic
Methodology and Philosophy of Science VIII, Jens Erik Fenstad et al., Elsevier Science
Publishers, Amsterdam, 1989, pp. 571-589.
12. The proof can be found in The Game of Language, Jaakko Hintikka and Jack Kulas, D.
Reidel, Dordrecht, 1983, and in Hodges, Wilfrid, 1989, "Elementary predicate logic",
in F. Gabbay and F. Guenther, eds., Handbook of Philosophical Logic, D. Reidel,

Dordrecht, Reidel.
13. These games have been introduced in Jaakko Hintikka, and Gabriel Sandu, 1989,
"Information independence as a semantical phenomenon", in J. E. Fenstad et al.,
eds., Logic Methodology and PhilolJOphy of Science VIII, Elsevier Science Publishers,
pp. 571-589.
14. This example comes from A. Blass and Y. Gurevich, "Henkin quantifiers and complete
problems", in Annals Pure App. Logic 32 (1986) 1-16.
15. What we mean by the use of independent quantifiers in Ramsey Theory can be seen
from an example. As such an example, let us consider one of the central results of Ram-
sey Theory, the Hales-Jewett theorem. (See Ronald L. Graham et al., Ramsey Theory,
John Wiley, New York, 1980.) It can be formulated in the conventional terminology
as follows:
For all r, t there exists N' = H J(r, t) so that for N ~ N', the following holds: If
the vertices of Ct' are r-colored there exists a monochromatic line.
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, Ct' 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:
For all t, the following holds: for each R there exists N', and for each Ct' together
with an r-coloring of its vertices, there exists a line 1 in cf' 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 of only t and R while 1 is a function of only t, N and the r-coloring. For
instance, for 1 we can take the line in cf' that has the largest number of elements of
the same color. This brings it to the open that there is a Henkin quantifier structure
as a part of the logical form of the Hales-Jewett theorem.
16. This follows from the results of W. Walkoe, Jr., "Finite partially ordered quantifica-
tion", J. Symbolic Logic 35 (1970) 535-555.
17. The contrast we have in mind here is that between ordinary first-order logic and higher-
order logic (or such extensions in the direction of higher-order logic as IF logics). The
usual modal logics, epistemic logic, etc., are in the sense intended here in the same
boat as ordinary first-order logic.
With this provisio in mind, one can see that the majority of papers in such AI
oriented volumes as Joseph Y. Halpern, ed., Theoretical Aspects of Reasoning About
Knowledge: Proc. of the 1986 Conj., Morgan Kaufmann, Los Altos, CA, 1986, or its
successor volume, Moshe Y. Vardi, ibid: Proc. Second Conj., Morgan Kaufmann, Los
Altos, CA, 1988, use an essentially first-order logic.
18. For Henkin quantifiers, see Note 10 above, and also M. Krynicki, "Hierarchies of
partially ordered connectives and quantifiers", in Mathematical Logic Quarterley 39
(1993) 287-294.
19. See Jaakko Hintikka, "What is elementary logic? Independence-friendly logic as the
true core area of logic", forthcoming.
20. A proof is sketched in Jaakko Hintikka, "Quantifiers versus quantification theory",
Linguistic Inquiry 5 (1974) 153-77. Another proof may be found in M. Krynicki and
A. Lachlan, "On the sema.litics of the Henkin quantifier", J. Symbolic Logic 44 (1979)
184-200.
21. See the article of Blass and Gurevich in Note 14 above.
22. See David Marr, Vision, W. H. Freeman, San Francisco, 1982, especially p. 25.
23. See Jaakko Hintikka, "Paradigms for language theory" Acta Philosophica Fennica
(1990) 181-209.
24. For the relation between Henkin quantifiers and NP-problems, see Blass and Gurevich,
Note 21 above.
25. For the metatheory of IF first-order logic, see Gabriel Sandu, "On the logic of in-
formational independence and its applications", J. Philosophical Logic, forthcoming;
Jaakko Hintikka, "What is elementary logic? Independence-friendly logic as the true
core area of logic", forthcoming, and Jaakko Hintikka, "Defining truth, the whole
truth and nothing but the truth" , forthcoming (preprint available as Reports from the
Department of Philosophy, University of Helsinki, no. 2, 1991).
26. For this purpose, we have to use interpolation formulas which differ somewhat from
the familiar Craigean ones. See Antti Koura and Jaakko Hintikka, "On the difficulty
of deductions" forthcoming.
11
MODEL MINIMIZATION - AN ALTERNATIVE TO

CIRCUMSCRIPTION
(Received: January 6 1987)
Abstract. The idea underlying John McCarthy's notion of circumscription is interpreted, for formulas with
finite models, as asking whether the conclusion C is true in all the minimal finite models of the premise T.
A way of modifying one of the usual proof procedures for first-order logic (the tableau method) is given
which captures this idea. The result is shown to differ from the consequences of McCarthy's circumscription
schema. The resulting proof procedure is extended to the case in which it is also required that the extensions
of the primitive predicates are minimal. For formulas with only infinite models, the idea on which the
concept of circumscription is based is tantamount to the author's idea of restricting models to minimal
ones.
Key words: circumscription, model, minimal model, non-monotonic reasoning, tableau method, database
theory.
1. SmaU Is Beautiful - In Informal Reasoning

John McCarthy's notion of circumscription' is based on an extremely interesting
insight into spontaneous human reasoning, If he is right, we human reasoners (and,
McCarthy adds, intelligent computer programs) tend to assume "that the objects they
can detennine to have certain properties or relations are the only objects that do." In
other words (my words rather than McCarthy's), human thinkers characteristically
assume that the intended models of our sentences (formulas) are the minimal ones,
containing only such individuals (of different kinds) as we are forced to include in
them by our explicit assumptions or by the logical consequences of such assumptions.
By the same token, only such kinds of individuals are instantiated in the preferred
models as have to be exemplified in these models.
I am convinced that McCarthy has seen a prevalent and important feature of
human reasoning, and hence of the kind of reasoning students of AI want to under-
stand and to capture. McCarthy has himself presented persuasive examples to show
the presence of this feature in a variety of different types of reasoning. Further
evidence is not hard to come by. At least anecdotal evidence is, for instance, provided
by the widespread use (or misuse, as most logicians would call it) of a definitory "if"
when "if and only if" is strictly speaking meant, in contexts like "x belongs to the set
to be defined if it satisfies the condition S[x]."
Independently of McCarthy, I have been led to a similar idea in the foundations of
mathematics. However, at first I applied the idea only to theories with exclusively
infinite models, I shall argue elsewhere that the best way of handling the logic of
MODEL MINIMIZATION - AN ALTERNATIVE TO CIRCUMSCRIPTION 213
certain important mathematical theories (with only infinite models) is to restrict by a

fiat all relevant models to models which are minimal in a certain specifiable sense. 2 As
such minimal models, prime models might first seem to do a minimally satisfactory
job. but on a closer look a somewhat more demanding property of models. viz., their
being superprime (in a sense that can be defined) turns out to be the best explication
of the minimality idea. In this way, mathematical theories (e.g., elementary number
theory) become descriptively complete (Le., they now rule out all nonstandard
models). The price one must inevitably pay is that the underlying logic becomes
incomplete.
This sense ofminimality applies only to infinite models, however. whereas McCarthy
is concerned also, and apparently primarily. with finite models. This prompts the
question as to how the minimality idea can be carried out so as to apply also to finite
models. One can also ask whether there is a more general treatment whose special
cases include both the finite and the infinite case.
2. Circumscription
John McCarthy's answer to this question is his idea of circumscription. Since it is
familiar from the literature, it suffices here to indicate only the main lines of this
answer.
McCarthy interprets the minimality idea on which the notion of circumscription is
based to mean that "the objects that can be shown to have a certain property P by
reasoning from certain facts A are all the objects that satisfy p".3 This leads him to
formulate circumscriptive inference by means of the following schema:
(A[4>] & (Vx)(4>[x] ::l Px ::l (Vx)(Px ::l 4>[x)) (I)
Here A[P] is a complex formula containing the one-place predicate P and A[4>] the
same formula with a complex expression 4> (with "x" as its only free variable)
replacing P.
This schema can be obviously generalized so as to apply to many-placed predicates
and furthermore to several predicates at the same time.
However, the concept of circumscription does not seem to capture the intended idea
optimally. There are at least two kinds of problems about it. First, in some cases it
seems to yield too strong implications. Second, it does not suffice to restrict models
to minimal ones.
As to the former kind of problem, several different difficulties will be discussed later
in this paper. Suffice it here to indicate the general trend of these difficulties by means
of a not altogether serious example. Let the circumspection schema deal with Alice's
"theory" of eating and perceiving, as Lewis Carroll had programmed her. Indeed, let
4>(x) stand for "Alice eats x" and Px "Alice sees x". Then by means of the circum-
scription schema it can be proved that, in the immortal words of the Mad Hatter
rebuking Alice, "you might as well say. , . that 'I see what I eat' is the same thing as
'I eat what I see'. "4 This conclusion is, of course, quite as absurd as Lewis Carroll
intended it to be.
As to the infinite case, McCarthy shows how his circumspection schema leads to the
schema of mathematical induction. It is known, however, that the schema for
mathematical induction does not suffice to restrict the models of an axiomatic theory
of elementary arithmetic to the minimal model, which in this case is the intended
structure N of natural numbers. In this case, the circumscription schema thus does
some of the work that is needed here, but it does not do the whole job.
Further objections to the circumscription idea will be presented later.
3. Minimization Implemented
Can we do better? The idea of model mlOlmlzation easily guides us to a more
satisfactory treatment in the finite case. In order to see what is involved, let us assume
that we are trying to prove a conclusion C from a finite set of premises T in a
first-order language, say, by means of the Beth tableau method. 5 Normally, the tableau
construction begins with T on the top of the left column and C in the right column.
The construction proceeds step by step by means of familiar rules. and the aim is to
close the tableau by closing all its subtableaux. These rules are but the usual Gentzen-
type rules of logical inference read upside down, with formulas in the left column
playing the role of formulas in front of the sequent sign and with formulas in the right
column playing the role of formulas that follow the sequent sign.
The model minimization idea can now be taken into account by initially modifying
the tableau procedure as follows:
(I) The tableau construction rules are modified in such an (obvious) way that no
formulas are transferred from one column to another.
(2) A sub tableau is closed also when Sand NS, for any formula S, both occur in
the same column, not just (as Beth stipulates) when S ocurs in both columns.
(3) Those entries in the left column which are independent of the presence of C in
the right column are said to form the designated part of the column. Clearly, (3)
does not affect the tableau construction as such at all.
(4) The only rule of tableau construction that actually has to be modified is the rule
for existential instantiation in the left column (ElL). What it normally says is
the following:
ElL: If(3x)S[x] occurs in the left column ofa sub tableau while no formula of the form
S[a] occurs in it, then one may add S[b] to the same column, where 'b' is a new
constant.
Even this rule technically need not be modified. What has to be done is to amplify
it by stipulating that our application of ElL creates a number of additional sub tableau
construction lines, alternative to the one initiated by ElL in its usual form (see above).
They will be called the ghosts of the normal sublableau, which we shall call the primary
construction (primary subtableau). They are not related to each other or to the
primary subtableau disjunctively, as the alternatives created by an application of the
disjunction rule (in the left column) are. but in a more complex way soon to be
explained.
Each ghost of a primary suhtahleau shares the initial part with the primary one. [t
differs from the primary one, created by an application ofE[l. by the fact that instead
of the new formula S[b] it initially contains a formula of the form S[a;]. where a, is
one of the constants occurring in the designated part of the subtableau. For each such
a; we have a different ghost initiated by S[a;).
Ghost tableaux are closed in the same way as primary ones. After a ghost subtableau
is closed. it is henceforth disregarded completely, and said to be dead.
Each ghost sub tableau is constructed precisely in the same way as the corresponding
primary sub tableau. Each formula in a ghost has a counterpart in the primary
subtableau and vice versa. This is easily seen to be always possible. except in one case.
this is when in the primary sub tableau E[l is applied. Assuming that the correspond-
ing existentially quantified sentence in the ghost construction is (3y)S[y. a,]. it may
then happen that there already is a formula of the form S[c,. a, I in the left column of
the ghost.
Then the ghost construction is continued simply by considering S[cj , a,] (for each
c,) as the counterpart to the new formula S[d, a;] introduced by Ell in the primary
construction. Thus the analogy between a primary construction and its ghosts can
always be maintained.
We shall restrict the ghost generation to applications of Ell in the designated part
of the overall tableau construction.
Notice that ghost tableaux can form a branching hierarchy: there may be ghosts of
ghosts of ghosts ....
4. Mini-Consequence Defined
It is not very hard to see how the intuitive idea of model minimization can be
implemented by means of the ghost constructions. For the purpose, let us see what
can happen in a tableau construction which starts with the premise T and the
conclusion C. let's assume that we are considering one subtableau s.
let's first assume that the designated part ds of s is completed, i.e., that it reaches
a stage at which no further applications of tableau rules to ds are possible while the
designated part ds of the tableau is still open, i.e., while no contradictions are present.
([n other words, in ds no formula occurs in both columns and for no F do both F and
r::-vF occur in the same column.) Then it is seen that ds describes a finite model M,
of T. The interesting question is whether M, is minimal.
In order to determine whether it is, we can examine all the ghosts of ds. Clearly, no
tableau rules can be applied to them, either. Hence there are two possibilities concern-
ing one of the ghost constructions g: (i) it is closed; (ii) it is completed. '
If (i), g can be disregarded. If (ii), g defines a smaller finite model of T. In order to
determine whether it is minimal, we have to examine the ghosts of g. If one of them
is completed. we have found a still sma1ler model of T. Eventually this process will
come to an end with a minimal completed tableau construction for T which defines
a minimal model Mo of T.
This is the story of what happens if s is completed. If s is closed, then so clearly are
all its ghosts. and s therefore yields no model for T.
Irs goes on to infinity without being closed or completed, then it does not yield a
finite model. But even then, one of the ghosts of s. say g. can sometimes be completed
after a finite number of tableau construction steps. Then we can find a minimal finite
model by treating g in the same way as s was treated above in the case when it was
completed.
If neither s nor any of its ghosts is ever completed, we do not obtain a minimal
model from the modified tableau construction. (Cf Section 10 below.)
Consider now one of the cases where the moified tableau construction has produced
a minimal model, i.e . where we have a construction which has come to an end with
a completed open sub tableau or, rather. completed designated part of the construc-
tion. where no ghost of that construction line is similarly completed. Let the indi-
vidual constants occurring in the completed construction be ai' a2' ... , ak. The
minimality assumption can then be brought to play by continuing the sub tableau as
follows:
(i) Add to the left column
('v'x)(x = a l v x = a2 v ... v x = ak).
Alternatively, for first-order logic without identity, we must replace any fonnula or
subfonnula in the sub tableau of the fonn (3x)51x] by
and every fonnula or subformula of the fonn ('v'x)S[x] by
(ii) Continue the sub tableau normally.
It is of course understood that only the designated part of the left column of the
sub tableau in question comes to play in the continuation of the tableau construction.
I shall call this the minimization rule. It is calculated to capture the same idea as
others have tried to capture by means of the schema of circumscription.
lfin the entire tableau beginning with Tin the left column and Cin the right column
every stage of each subtableau has at least one continuation which can be closed by
means of the minimization rule or else can be closed altogether, then I shall say that
C is a mini-consequence of T. Of course, in each subtableau, the minimization rule is to
be used at most once. The notion of mini-consequence is the alternative I am offering
to trying to capture the model minimization idea by means of the procedure of
circumscription.
5. Examples and Observations

Several different observations are in order here.
First, the notion of mini-consequence yields in many cases the same results as
reasoning by circumscription. For instance, consider McCarthy's sample theory6
consisting (in a slightly different notation) of the conjunction
Isblock(a) & Isblock(b) & Isblock(c). (2)
If this is used as my T, and if the conjunction rule is applied to (2), the tableau
construction will come to an end with (2), Isblock(a), Isblock(b), and Isblock(c) in the
left column. All these formulas will be designated ones. Then the minimization rule
enables us to add to the left column
(Vx)(x = a v x = b v x = c). (3)
But this means that (3) is a mini-consequence of (2), just as it can be obtained from
(2) by circumscriptive inference.
However, in the case of the following formula its mini-consequences differ from its
circumscriptive consequences:
Isblock(a) v Isblock(b). (4)
By starting from (4) in the left column we arrive quickly at a situation in which the
tableau construction has generated two sub tableaux each with the following formula
inserted in virtue of the minimization rule into their left column:
(Vx)(x = a v x = b). (5)
This of course means that (5) is a mini-consequence of (4).

In contrast to this, the circumscription schema yields the conclusion
(Vx)(x = a) v (Vx)(x = b) (6)
which is not equivalent with (5).

Which conclusion is "right", the mini-consequence (5) or the circumscriptive one
(6)? Even though no absolute decision is possible here, it seems that the mini-
consequence (5) is much more natural than the circumscriptive consequence (6). For
suppose that someone actually puts forward (4). Clearly the propounder is envisaging
a situation in which both a and b are present. He or she might very well be reluctant
to introduce any other individuals, but most of us innocent bystanders would surely
think that the speaker would be welshing on his or her existential presuppositions if
he or she tried to retract an existential commitment either to a or to b. Hence (5) rather
than (6) presumably is the right consequence.
It can also be seen, by means of suitable examples, that a logic based on the notion
of mini-consequence is not monotonic. In this respect, too, it is like circumscription.
In order to see this, consider three sentences, T " T 2 , T), which say, intuitively
speaking,
T, = there exist at least n individuals with the property P;

T2 = there exist at least n - I individuals with the property P;
T3 there exist precisely n - I individuals with the property P.
For instance, if n = 2, we have:
T, (3x)(3y)(Px & Py & x =f. y)
T2 (3x)Px
T3 (3x)(Px & (Vy)(Py :::> x = y
Let us now consider what the minimal models of T, - T3 are like, in a sense of
"minimal model" which is intuitively obvious in the present application and which
will be discussed further in Section 6 below. In the minimal model of T, there are
precisely two P's. Hence T2 is true in it, i.e., T2 is a mini-consequence of T, . The
minimal model of T2 has precisely one P, wherefore T3 is true in it. Hence T3 is a
mini-consequence of T2 Yet T3 is not a mini-consequence of T, .
As an additional example, consider the theory T defined by the following axioms:
(Vx)(3y) Rxy (7)
(3y)(Vx) NRxy (8)
(Vx)(Vy)(Vz)(Rxy & Rxz :::> y = z) (9)
(Vx)(Vy)(Rxy :::> NRyx) (10)
For vividness, let us say that y is a successor to x whenever R'Cy. (The converse of
this relation will naturally be called precedence.) Then (7)-(10) say, intuitively, that
everyone has a successor, that someone is not preceded by anyone, that no one has
more than one successor, and that successor relation is antisymmetric. Then a
sentence C is a mini-consequence of T iff it is true in the model of the following
structure:
0---+10 ,0 (II)
~/
o
If a requirement is added to T saying that there must be at least 5 individuals in its
models, then the C is a mini-consequence of T iff it is true in both of the following
structures:
(a) 0 ----+1 ----+1 1 0 (12)
"o' /
Notice that while T allows there to be more than one individual without pre-
decessors, in the model that serves to characterize its mini-consequences in this way,
there is only one such individual.
6. Towards a Model Theory of Mini-Consequence

A model theory can be developed for the notion of mini-consequence. For the
purpose, we have to define a few concepts. Assume. for the sake of argument, that we
are dealing with a first-order language L where the only nonlogical concept is a
two-place relation R. (Our definitions are easily generalized from this case to an
unrestricted first-order language.) Assume also that a model M of L and a mapping
f of the domain of individuals do(M) of M into itself is given. Then it is said thatf
defines an m-automorphism iff the following conditions are satisfied.
(i) Iff a, bE do(M) and iff(a) =F f(b). then Rf(a)f(b) iff there are a', b' E do(M) such
thatf(a) = f(a'),j(b) = f(b'}, Ra'b'.
(ii) If!(a) = !(b) = c, then it must be the case that Raa iff Rbb and that Rcc iff Raa
and Rbb.
We shall say that a finite model M is minimal iff there is no m-automorphism which
maps do(M) into a proper subset of do(M). The following is the basic result that can
then be established in this direction:
Completeness for mini-consequences: C is a mini-consequence of T iff it is true in
each minimal model of T.
For instance, the reason why (12) (a)-(b) are non minimal models of the theory
defined by (7)-( 10) is that they each admit of an m-automorphism which identifies the
individuals represented by solid dots.
The result just stated is essentially a completeness result for my modified tableau
rules, including the mini-consequence rule. In fact, this completeness is an almost
direct consequence of the way the concept of mini-consequence was defined above. It
is easily seen that, in order to prove the desired result, it suffices to prove the following:
LEMMA. I : Assume that M is a minimal model of T. Then the applications of our

modified tableau rules can always be chosen so that in at least one sub tableau, or in at
least one of its ghosts, all the designated formulas are true in Mr.
This can be proved by induction on the length of the subtableau construction
considering the different tableau rules. The only tricky case is ElL (the left-handed
existential instantation rule, modified as indicated). By the inductive hypothesis the
formula (3x)S[x] to be instantiated is true in M. Now assume that none of the
individual constants a, already in the designated part of the tableau can make S[a;]
true in M. Then some other individual b in the domain of M (not yet named by the
constants in the tableau) will do so. In applying ElL to (3x)S[x] so as to yield a new
formula of the form S[d], we can take "d" to be the name of b.
Again, assume that S[x] is satisfied in M by some individual a j whose name already
occurs in the designated part of some subtableau. Then adding a formula S[b] with a
new symbol "b" to the left column will not necessarily preserve the truth of all
the formulas of the designated part in M. However, they are all true in M as far as
the ghost subtableau is concerned which is generated by the addition of S[a;] to the
tableau instead of S[b].
It is easily seen that this tableau construction comes to an end in a finite number
of moves in the sense that the designated part of the left column cannot be extended.
Then it is seen that this distinguished part is a model set with the same individuals as
M and with all its members true in M, in brief, it characterizes completely the
model M.
Since each minimal model M is obtainable in the way specified by Lemma I, C is
a mini-consequence of T iff it is true in all of them, as claimed.
7. Mini-Consequences Differ from Inferences by Circumscription

As a further example, consider a primitive "theory" T of biological family relation-
ships within the human race which guarantees that the relation "ancestor" of (the
ancestral of the relation of being a parent of) is irreflexive, asymmetrical and tran-
sitive. Let us further add an axiom which says that Adam and Eve are the only people
without parents. It is easily seen that then it is a mini-consequence of T that everybody
is a descendant of Adam an"d Eve. For in trying to construct a model for Tyou must
introduce parents for each human being (different from Adam and Eve), their parents,
etc., without being able to return to earlier members of the sequence. Hence the only
way in which the model to be constructed can be finite is for each such sequence of
parents to come to an end with Adam and Eve.
An interesting thing about this conclusion is that the circumscription scheme does
not validate it. For fairly obviously, all that you can prove by means of the circum-
scription schema is that each human being either has a father and a mother or else
is Adam and Eve. From this it does not follow, of course, that there cannot be infinite
chains of earlier and earlier ancestors.
Thus the notion of mini-consequence is stronger than that of inference by circum-
scription, in the sense of validating more inferences.
This example seems to show the superiority of mini-inference over circumscription.
F or clearly the same idea of not multiplying entities without necessity which motivates
the schema of circumscription also motivates the mini-consequence that all human
beings are descendants of Adam and Eve. A model in which this statement is true is
more economical than a model in which it is false; furthermore, there is no necessity
of assuming infinitely long chains of ancestors. Hence by the same heuristic principles
as guide inferences by circumscription, the biblical mini-consequence ought to be
valid. It is, but it is not validated by the schema of circumscription.
In this example, the mini-consequence in question cannot be established in my
approach by a finite argument. However, this does not spoil the counter-example. For
instance, I can add the assumption that every descendant of Adam or Eve is separated
from him or her by fewer than n generations. Then the mini-consequence in question
can actually be proved by a finite argument (tableau construction) without destroying
its role as a counter-example.
Even though I have formulated this example in somewhat frivolous-looking terms,
the underlying point is a serious one. It has, for instance, been argued that Aristotle
in fact assumed a principle of philosophical reasoning which amounts in effect to the
kind of application of the idea of mini-consequence which is at issue in my Adam-and-
Eve example.' More precisely, it has been argued that Aristotle assumed that in a
partly ordered set all chains must be finite, which comes close to minimality require-
ments of the kind we are dealing with here.
8. Mini-consequence and database theory

It seems to me that a concept like mini-consequence or circumscription is destined to
play an important role in database theory. 8 My reasons for this belief can be explained
as follows:
(i) Database theory should be developed in such a way that the data in question do
not include only atomic (particular) propositions (negated or unnegatc.d), but can
also include general propositions, more generally, arbitrary first-order propo-
sitions. This is necessary if database-driven reasoning is to be successfully appli-
cable to scientific or clinical reasoning.
(ii) If this is done, it will be awkward to define a database logically by means of
closure axioms, i.e., axioms which include a sentence of the form
(Vx)(x = a l v x = a2 v ... v x = ak ) (13)
where ai' a2, ... , ak are all the individuals named in the database.
For one thing, this procedure can easily tum a perfectly acceptable database into
an inconsistent one. It also makes it difficult to develop a natural theory of operations
on databases, such as adding new data to it or dropping some of the data.
(iii) For these reasons, it seems to me that the closure of a database should be
accomplished by means of an optional extra assumption which can be used only
when there is some reason to think that the database is supposed to be complete.
(iv) Even then, the closure requirement should not restrict the model that the data-
base defines to individuals mentioned in the database. It should also contain at
least those individuals whose existence is logically implied by the database.
The notion of mini-consequences is intended to be a way of reaching these ends. It

seems to realize very well the idea that each database has to be considered as telling
us everything that needs to be told about its subject matter. At least it realizes this idea
better than closure axioms.
9. Minimizing Predicate Extensions

The ideas so far developed can be extended in different directions. One of them is the
following: It can be argued, along the same lines as are followed above and are also
followed by McCarthy, that in ordinary reasoning human thinkers often assume, not
only that the intended models are as frugal as they can be with respect to the existence
of individuals, but also that the extensions of certain given predicates, or maybe just
the extension of a single given predicate P, can only be as large as they must be.
This idea can easily be captured by the same method as was used to formulate the
notion of mini-consequence. The modified tableau-building procedure outlined above
leads (where it comes to an end with completed constructions) to a specification of a
number of alternative minimal models of T in which C must be true in order to be a
mini-consequence of T. These models are not fully specified by the construction,
however. Consider, for instance, the case of a one-place predicate P occurring in T.
In the description of a model (model set) reached by the procedure, it is specified that
certain individuals C], C2, must have P and that certain others d], d2 , must not
have it, in other words, that the following must be true in the model:
However, the extension of P is otherwise left open.

Now clearly we can implement the idea that the extension of P is assumed to be
minimal by requiring that only C], C2 , have this predicate in the minimal model
we have reached. This means changing the minimization rule in such a way that all
the formulas NPe] , NPe2' ... are also added to the left column, where the e, of all
the individuals in the minimal model different both from all the Cj and from all the dj ,
when the construction of the designated part of a sub tableau comes to a completion.
Whenever C can be derived from T by the latest modified tableau method, C is true
in all the minimal finite models of T in which the extension of P is as small as possible.
This is easily generalized so as to implement the minimization of any finite number
of predicate extensions. Thus the idea of minimizing predicate extension is also easily
captured by our technique.
10. Infinite Case

The idea of model minimization can thus be implemented fully and without difficulties
whenever the relevant minimal models can be finite. It appears that in most actual
applications we are in fact dealing with the finite case. But what happens when we
have a set of first-order premises (assumptions) which have only infinite models? A
case in point would be a Peano-type axiomatization of elementary number theory.
Does the idea of model minimization have interesting applications to such cases?
The answer is an emphatic yes. In fact, I have argued elsewhere that the most
promising approach to theories like Peano arithmetic is precisely to restrict the
models to be considered once and for ali to minimal ones. 9 This prompts of course
the question: minimal in what sense? The cardinality of the domain of a model no
longer serves as an index of its "size" in the sense which is relevant to its minimality.
for clearly. e.g . one countable model can be richer than another countable model.
Hence. some sort of qualitative minimality is what counts here. Now the finest
qualitative distinctions among different individuals in the domain of a model are
apparently those effected by the so-called types (in the sense of model theory).lo
Minimality will then presumably mean being what model theorists call primacy (being
a prime model). In fact, it turns out that the concepts of type and of primacy have to
be generalized. I I This gives rise to a notion stronger than primacy, which I have called
superprimacy.12
It can be shown that an overall restriction on models to superprime ones enables
ut to formulate a complete (in the sense of categorical) theory of elementary number
theory}3 In this sense, the idea of model minimization turns out to be extremely
important in the foundations of mathematics. This seems to me to vindicate in a most
impressive way John McCarthy's original intuition.
The other side of the Godelian coin is that we cannot of course obtain a complete
axiomatization in a sense that involves a complete proof procedure. 14 Hence the
modified proof procedure sketched in this paper for mini-consequences cannot be
generalized to the infinite case.
The operative question becomes, rather, the problem of looking for such stronger
rules of proof as would capture partly (but increasingly extensively) the effects of the
overall restriction of models to superprime ones. Various schemata of induction can
be thought of as (partial) ways of doing so. This intriguing problem lies beyond the
scope of the present paper, however.
Acknowledgement
The work reported here was made possible by NSF Grant # 1ST - 8310936
(Information Science and Technology), Principal Investigators Jaakko Hintikka and
C. J. B. Macmillan.
Notes
I See McCarthy, John. 'Circumscription - a form of non-monotonic reasoning'. Artificial Intelligence 13,
27-39 (1980).
2 See 'Extremality assumptions in the foundations of mathematics'. in PSA /986. vol. 2 (eds. A. Fine and
M. Forbes), Philosophy of Science Association. East Lansing, MI 1987; and 'Is there life in the foundations
of mathematics after G6del?' (forthcoming).
) Op. Cit . p. 27.
See The Annotated Alice (ed. Martin Gardner), Penguin, New York (1965), p. 95.
S Beth. Evert W . 'Semantic entailment and formal derivability', Medede/ingen van de Koninklijke Neder-
lamlse Akademie van Wetenschappen N.R. 18, no. 13, Amsterdam. 309-342 (1955).
6 Op. Cit., p. 32. .
7 See Beth. Evert W . Foundations of Mathematics, Harper and Row, New York (1966), pp. 9-12. (Actu
ally, Beth formulates his principle in a somewhat different way. What is given here seems to be me to be
the gist of the principle he has in mind.)
I C/. e.g . Reiter. Raymond. 'Towards a logical reconstruction of relational database theory'. in On
Conceptual Modelling (eds. M. L. Brodie. J. Mylopoulos and J. W. Schmidt). Springer. New York (1984).
pp. 191-238. (See especially pp. 191. 218 and 227-229 for the "closed world assumption".)
9 Op cit . (note 2 above).
10 See, e.g . Chang. C. c.. and J. J. Keisler. Model Theory. North-Holland. Amsterdam (1973), pp. 77-78.
II Op. cit. (note 2 above).
12 Op. cit. (note 2 above).

13 Op. cit. (note 2 above). It is important here to keep apart the several senses of completeness. all the more
so because they are frequently confused with each other. What is at issue here is the power of a first-order
axiom system (theory) to restrict its models to the intended (standard) ones. This has to be distinguished
from the completeness of a formalization of logic. which means recursive enumerability of all (and only)
valid formulas. and from the deductive completeness of a formalized system. which means the formal
derivability of C or evC for each C in the language of the system .
.. It is only in this sence that Good proved the incompleteness of elementary arithmetic.
12
NEW FOUNDATIONS FOR MATHEMATICAL THEORIES
1. The motivation. In this paper, I shall outline a new approach

to the logical foundations of mathematical theories. One way of looking at its
motivation is as follows (I am following here Hintikka, 1989):
In the foundational work around 1900, e.g. in Hilbert's Foundations of
geometry, a crucial role was played by assumptions of extremality (i.e., mini-
mality and maximality). For instance, Hilbert's so-called axiom of completeness
is a maximality assumption. The Archimedean axiom can be thought of as a
minimality assumption, the principle of induction likewise as a minimality ax-
iom, and Dedekind's assumption of the existence of a real for each cut as a
maximality assumption. Slowly, it has become clear to everybody that such
extremality assumptions cannot normally be expressed as ordinary first-order
axioms. To what extent they can or cannot be expressed in other ways, e.g. as
higher-order axioms or set-theoretical axioms, and to what extent we should try
to express them in such ways, will not be discussed here. In any case, in spite
of the tremendous prima facie interest and power of extremality assumptions,
they have not attracted much interest lately.
The approach proposed and outlined here relies crucially on extremality
assumptions but seeks to implement them in a new way on a first-order level.
Instead of introducing extremality assumptions on the top of a ready-made logic
as explicit axioms, I propose to build them into the very logic we are employ-
ing, thus by-passing the difficulties the earlier uses of extremality assumptions
encountered.
A logic is in effect specified by a space n of models together with a def-
inition of what it means for a statement (closed formula) to be true in a model
MEn (and for a formula to be satisfied with in M). I shall not modify the
latter ingredient. Instead, I propose to modify the usual space of models (of a
given first-order language L) in the simplest possible way, viz. by omitting some
of its members.
Even though this kind of modification looks innocuous, it facilitates a
radical new look at the prospects of mathematical and logical theories. Most
importantly, the possibility of reaching completeness can be profoundly affected.
What are the different kinds of completeness relevant here? Here are four
candidates, which have not always been distinguished from each other sufficiently
clearly:
(1) DeJcriptive completeneJJ. It is an attribute of a non-logical theory. It
means that the theory has as its models only the intended (standard)
ones, i.e., that it has no non-standard ones. If there is only one standard

model, a descriptively complete theory must be categorical. Here the

notion of standardness has to be characterized independently.
(2) Semantical completene33. It is a property of an aziomatization of (some
branch of) logic. It means that the theorems of the axiomatization exhaust
all valid formulas, where validity means truth in all the models of the space
O. Thus semantical completeness amounts essentially to the recursive
enumerabilityof the set of valid formulas.
(3) Deductive completeneu. It is a property of a non-logical axiom 3ydem
together with an aziomatization of logic. It means that, for each statement
C, either Cor -c can be derived from the (non-logical) axioms by means
of the given logic.
(4) Hilbert'3 so-called mom of completene33 is in effect a maximality assump-
tion in a non-logical axiom system. While Hilbert's own intentions are not
clear, this axiom can be taken to say that one cannot add new individuals
to a (standard or intended) model without violating other axioms:
What happens when the space of models 0 is replaced by some 0* C O?
(1) Descriptive completeness becomes ceten., paribU3 easier to reach since
some (or all) of non-standard models in 0 may belong to 0 - 0*.
(2) Since there are fewer models, there are ceten., paribU3 more valid
formulas (i.e., formulas true in all of them). Hence semantical completeness can
become more difficult to achieve.
In other words, by moving from 0 to a suitable n* (0" C 0) we can
trade in the semantical compfeteness of our underlying logic in order to achieve
the descriptive completeness of suitable mathematical theories. I have argued
elsewhere that this would represent a major gain in philosophical and conceptual
clarity. (See Hintikka 1989.)
For example, Godel showed elementary arithmetic to be incomplete in
the sense (3) (deductive incompleteness). From this it does not by itself follow
that elementary arithmetic is descriptively incomplete. This does follow if the
underlying logic is complete, which Godel had proved (for first-order logic) prior
to proving the the incompleteness of elementary arithmetic. However, if we
are willing to change this logic (strengthen it) so as to render it semantically
incomplete, we can very well hope to reach a descriptively complete first-order
theory of arithmetic. This in fact turns out to be possible.
More generally, by means of suitable extremality restrictions on models,
it will turn out to be possible to formulate categorical first-order axiom systems
inter alia for elementary number theory, the theory of reals, Euclidean geometry,
and the second number class (countable ordinals). (See 5 below.)
(3) Deductive completeness, being a kind of combination of descriptive
and semantical completeness, is not necessarily affected by any trade-off between
the other two kinds of completeness (1)-(2).
(4) The restriction of the space of models to a suitable subset serves the
same purpose as the axiom of completeness, and is supposed to replace any such
explicit axiom.
NEW FOUNDATIONS FOR MATHEMATICAL THEORIES 227
2. Model theory via constituents. The crucial question obviously

is how the ideas of minimality and maximality can be implemented in precise
and general terms. Since we are speaking of the minimality and maximality of
models (or parts thereof), the obvious resource here is the theory of models. I
shall review some of the basic ideas of model theory, but for the sake of certain
further developments I shall do so in an unfamiliar way. I shall use in the review
the technique of constituents and distributive normal forms. Even though this
technique may in the last analysis be dispensable in favor of more commonly
employed conceptualizations (e.g. back-and-forth techniques), it offers heuristic
advantages by allowing an almost geometrical (tree-theoretical) visualization of
the logical relationships under scrutiny in this paper.
An approach to model theory via constituents leads us straight to the field
of stability theory. However, it is in fact quicker and much more perspicuous to
develop the necessary theory directly here without going by way of stability
theory. (For stability theory, see e.g. Baldwin, 1988.)
In what follows, it is assumed that we are dealing with a given first-order
language L with a finite list of predicate constants, no function symbols, and
an unspecified supply of individual constants. I shall deal only with languages
without identity. It turns out that the presence or absence of identity does not
matter very much for the central purposes of this work.
A constituent L with the free variables Xl, X2, .. , Xk will be expressed as
follows:
(2.1 )
It is a well-formed formula which has a number k of arguments Xl, X2, ... , Xk
and it is also characterized by its depth d. The subscript i serves to distinguish
different constituents with the same arguments and with the same depth from
each other.
Constituents can be defined recursively as follows:
(2.2) C(O)[X},X2, ... , Xk]
is of the form
(2.3) 1\(j)Aj[Xt.X2, ... ,Xk]
jEJ
where the Aj[XI, X2, ... , Xk], j E J, are all the different atomic formulas that can
be formed from the predicate constants of L and of Xl, X2, . .. , Xk, and where each
(j) is either", or nothing, depending on j.
(2.4)
is of the form
(2.5)
1\ (3y )Cjd)[y, Xl, X2, . .. , Xk] & (Vy) VCjd)[y,XI,X2, ... ,Xk]
jEJ iEJ
& C,(O)[XI,X2, . ,XA,].

Here the last conjunct is simply some one constituent without quantifiers with
as its only free individual symbols. The index set J is a subset of
Xl, X2,' , Xl:
the set of the subscripts of all the different constituents
(2.6)
Intuitively, a constituent like (2.5) of depth d + 1 tells us what kinds of
individuals there exist (in relation to Xl, X2, ,X k) and do not exist. The latter
is accomplished in the universally quantified disjunction of (2.5) by saying that
each individual must be of one of the kinds listed in the first of the three conjuncts
in (2.5). Here the "kinds of individuals" are in turn specified by constituents
(2.6) of a lesser depth d.
Each constituent (2.5) thus has a tree structure where the nodes of this
labeled tree are constituents of increasingly smaller depth each occurring in its
predecessor. Intuitively, each branch of such a tree describes a sequence of d + 1
individuals that you can find in a model of L in which (2.5) is true. The tree
structure show how the initial segments of such sequences limit their possible
continuations.
In a sense, a constituent thus presents an explicit description of certain
salient structural features of a model M in which it is true. The constituent tells
you which (ramified) sequences of individuals (up to the length d + 1) you can
hope to find in a model in which it is true.
In this work, the term "constituent" will also be applied to substitution-
instances of (2.4) with respect to the individual constants of L, i.e., to formulas
like
(2.7)
or
(2.8)
If identity is present in L, the definition of a constituent can be changed
as follows: (2.2) is now of the form
m;cn
(2.9) /\(j)Aj[Xt, X2, ... ,Xk) & /\ (x m -I xn)
jEJ m,n=:;k
and (2.5) is now of the form
(2.10) /\(3y)C)d)[y,XI,X2, ... ,Xk] &
jEJ
m=k
(Vy)( /\ (y -I xm) :> Vc(d)[y, Xl, X2, ... ,Xk)) & cfO)[xt, X2,'" ,Xk).
m=l jEJ
What this means is that in the presence of identity constituents can be

written precisely in the same way as in the identity-free case provided that an
exclusive interpretation of quantifiers and free individual variables is adopted.
In the rest of this paper, I shall assume that identity is not present.
The concept of (quantificational) depth of deS) of a formula S can be

defined for arbitrary formulas as follows:
(d.i) IT there are no quantifiers in S, d( S) = o.
(d.ii) d(Sl & S2) = d(Sl V S2) = max[d(St},d(S2)]
(d.iii) d( ",S) = d( S)
= =
(d.iv) d3x)S[x]) d'9'x)S[x]) d(S[x]) + 1
It is easily seen that this definition agrees with the way the notion of depth was
used in connection with constituents.
In discussing the identity of constituents we shall consider (i) the order
of disjuncts and conjuncts, (ii) the choice of bound variables, and (iii) possible
repetitions of identical (modulo (i)-(ii members as inessential. IT this idea is
used in the numbering (indexing) constituents we can prove the following:
LEMMA 2.1: Hi #= j,
(2.11) cfd)[x1!X2, ... ,Xk] ~ ",Cjd)[X1.X2, ... ,Xk].
This is easily proved by induction on d. It is also obvious on the basis of
the intuitive meaning of constituent.
We can also prove
LEMMA 2.2: H S is a closed formula of L of depth d, then for each i
either
(2.12)
or
(2.13)
This, too, can be proved by induction on d.
The same can be proved for formulas S[Xl,X2, ... ,Xk] and constituents
C(d)[Xl,X2, ... ,Xk] having the same free variables Xl,X2, ... ,Xk. We shall call
this result Lemma 2.3.
In particular,
LEMMA 2.4: For each constituent of the form
(2.14) Crd+l)[Xl,X2, ... ,Xk]
there is precisely one constituent
(2.15) CJd)[X1.X2, ... ,Xk]
such that (2.14) logically implies (2.15). For other values of j, (2.14) logically
implies the negation of (2.15).
In the former case (2.15) can b~ obtained from (2.14) by omitting it from
all constituents of depth 1, together with connectives that thereby become idle,
and all repetitions. It is obvious that the result is implied by (2.14).
What Lemma 2.4 says is in effect that you can omit the last layer of
quantifiers from any constituent and obtain a shallower one which is implied
by the original. In fact you can omit anyone layer of quantifiers in a given
constituent.
Together Lemmas 2.1-2.4 entail

LEMMA 2.5: Each consistent formula S(d)[XI, X2, , Xk] of depth d with
the free variables Xl, X2, . , Xk is logically equivalent with a disjunction of con-
stituents of the form
cjd)[Xl, X2, , Xk].
Not only can we omit layers of quantifiers from a constituent; we can

likewise omit arguments from it.
LEMMA 2.6: Given a consistent constituent
(2.16) c1d)[x1!X2, ,Xk],
consider the constituent

(2.17)
obtained from (2.16) by omitting from it all atomic formulas containing Xk, all
connectives which thereby become vacant, and all repetitions. Then (2.17) is the
only constituent of the form
(2.18)
which is implied by (2.16).
Proof: It is again obvious that (2.16) implies (2.17). IT it implied any
other constituent of form (2.18), it would be inconsistent by Lemma 2.1.
Several of the basic concepts of model theory are easily and naturally
defined by reference to constituents.
A consistent sequence of constituents
(2.19) c~(~)[xt, X2, . , Xk]
=
with a fixed k (k > 0), but with an ever increasing d 1,2,3, ... defines a k-type.
It is easily shown that this definition is equivalent with the usual one, according
to which a k-type is the maximal consistent set of formulas with Xl, X2, , XI<
as their only free variables.
When k = 0, we have a sequence of closed constituents
(2.20) Ci(d)
(~ (d = 1,2, ....)
From Lemma 2.5, it is seen that (2.20) defines a complete theory, and that each
complete theory can be represented in this way.
Notice that each k-type (2.19) implies a unique complete theory (2.20).
For each member of the sequence (2.19) implies a unique constituent without
any individual constants in virtue of Lemma 2.6. Those types (2.19) which so
imply (2.20) are the only ones consistent with (2.20).
The k-type (2.19) is compatible with the complete theory iff constituents
(2.19) all occur in the successive members of (2.20).
The types compatible ~ith (2.20) will be called the types of the complete
theory (2.20). Each type satisfied in a model of (2.20) is a type of (2.20), but
all the types of (2.20) need not be satisfied in a given model of (2.20). The
question as to which of them are satisfied is one of the central ones in model
theory. Different kinds of models are distinguished from each other by the types
that are satisfied in them.
One particularly useful result concerning constituents is the following:
LEMMA 2.7: Given a complete theory (2.20), a model M of (2.20), and
a constituent C (or a substitution instance of a constituent containing names of
members of dom(M)), if C is compatible with the set of sentences true in M,
C is satisfied in M. For constituents without names, it suffices to assume that
they are compatible with Th(M).
We can here perhaps see some of the advantages of the use of constituents.
All the lemmas of this section can be seen to be valid directly on the basis of
the import of a constituent. (Cf. the explanation of the meaning of the tree
structure of a constituent given above.) Lemma 2.7 is a case in point, though
perhaps slightly less obvious at first than the earlier lemmas.
Other results can likewise be read off from the intuitive meaning of a
constituent, albeit not equally directly. As an example of such a result, we can
mention the following result off almost immediately from the intuitive meaning
of a constituent in the following:
LEMMA 2.8: Assume that
(2.21) Cjd)[al, a2, ... , ak)
is compatible with
(2.22) C}~d+k) .
Then (2.22) implies

(2.23) (3xd(3x2)'" (3Xk)Cid)[xl! X2, ... , Xk).
Proof (informal): In exploring a world in which (2.22) is true, we can
come upon Xl = aI, X2 = a2, ... , Xk = ak in this order. If (2.21) is likewise true
in the same world, as it can be if (2.21) and (2.22) are compatible, the rest of
the world is described by cid) [al' a2, ... , ak).
This lemma holds by the same token if there are additional free variables
or constant parameters in (2.21) and (2.22).
Many of the well-known results in model theory are proved easily and in
a perspicuous way by means of constituents. As an example of the use of con-
stituents to systematize old results and to obtain new ones, Rantala's monograph
Aspects of definability (1977) can be mentioned.
More illustrations of the use of constituents are offered in the next few
sections.
3. A wrong implementation of the extremality idea. At this
point, it might seem to be easy to implement extremality conditions on models.'
The natural way to interpret our extremality requirements is to say the following:
A model is minimal iff as few kind., of individual., as possible are instantiated in
it; a model is maximal iff as many kind., of individual., as possible are instantiated
in it. Then the prima facie plausible idea is to take the concept of type defined
above as the explication of the pre-theoretical idea of a kind of individuals (or

a kind of k-tuples of individuals). Then the question raised at the end of the
preceding chapter (Which types are satisfied in a model?) would become highly
relevant to the extremality project.
What can we say by way of a response to this suggestion? In order to
answer the question, we need a few further concepts. It may happen that the
k-type
(3.1) C~(1)[Xl,X2"",Xk] (d=I,2, ... )
compatible with the complete theory Tj
(3.2) CJtJ) (d = 1,2, ... )
stops branching from some point on. Then there is an initial segment of (3.1)
such that only one continuation of it is compatible with (3.2). Such a type is
called atomic in Tj = (3.2).
It may happen that each initial segment of each k-type (for each k) com-
patible with (3.2) is consistent with an atomic k-type. Then the entire complete
theory (3.2) will be called atomic.
Clearly, each atomic k-type compatible with (3.2) must be satisfied in
each model of (3.2). The interesting question is whether any other types need
to be satisfied. This question turns out to be more complicated than one might
first suspect. A model M is called atomic iff each k-tuple of the elements of
the domain dom{M) of M satisfies an atomic k-type. A partial answer to the
question just posed is given by
LEMMA 3.1: A complete theory has a countable atomic model iff it is
atomic.
In the other direction, there are models M such that each k-type, for
each k, compatible with the complete theory Th(M) true in M, is satisfied.
Such models are called weakly $aturated.
A related requirement is the following: Suppose a k-type ttlxl,X2, ... , Xk]
is compatible with a (k + I)-type t2[Xt,X2, ... ,Xk,XHd in Th(M) and that
a},a2, ... ,ak E dom{M) satisfy tt[at,a2, ... ,akj. If there always exists aHl E
dom{M) such that al,a2, ... ,ak,aHl satisfy t2[at,a2, ... ,ak,aHd, then a
weakly saturated model M is called $aturated.
Saturated models are interesting "special models." It is not difficult to
prove that any consistent complete theory has such a model.
Atomic models ar~r seem to be--minimal models in some reasonable
sense, and saturated models seem to be maximal models in an equally clear
sense. The idea is this: It seems that I-types constitute the finest partition
of individuals into different "kinds" that can be affected by first-order means;
and mutati$ mutandi$ for k-types with k > 1. Hence it seems that the poorest
models one can characterize by first-order means are the ones in which only
those "kinds" (i.e., types) are exemplified which must be satisfied in any case,
i.e., atomic models. Likewise, it appears that the richest models that can be dealt
with on the first-order level are the ones in which all the different "kinds" (i.e.,
NEW FOUNDATIONS FOR MAHIBMATICAL THEORIES 233
types) are instantiated, perhaps with the proviso that these types are instantiated
so as to allow all possible steps from k-types to (k + 1)-types. In other words,
the richest models seem to be wea.kly saturated or perhaps saturated models.
In brief, the concepts of atomicity and saturation seem to be the natu-
ral explications of the ideas of minima.lity and maximality that are guiding my
thinking. Yet they do not do this job well at all. Extrema.lity requirements so
interpreted do not allow us to capture the intended (sta.ndard) models in the
interesting ca.ses.
For insta.nce, we ca.nnot in this way capture naturally the intended "stan-
da.rd" model of Pea.no arithmetic. On the contrary, it is known (see, e.g., Chang
and Keisler, Example 3.4.5) that any consistent complete extension of Peano
arithmetic is an atomic theory a.nd hence ha.s atomic models. This holds also
for complete theories not true in the intended structure of natural numbers.
Hence the atomicity requirement does not do the job of capturing the structure
of natural numbers.
Another example is offered by the (first-order) theory of dense linear order.
It has a model which has the structure of the rationals. This model is at the
same time an atomic model a.nd a saturated one. But it is not really a minimal
model in some intuitive sense, for you can omit elements from it and yet preserve
its status a.s a model. It is not really a maximal model, either, in some striking
sense, because it can be embedded in a richer one, viz. the structure of the reals,
which is not isomorphic with it.
The notions of atomicity a.nd saturation of course do not exhaust the
resources of contemporary model theory. For instance, there is the notion of
prime model.
DEFINITION: A model Mo of a theory T is a prime model iff it can be
elementarily embedded in every model M of T.
Prime models might look like plausible candidates for the role of a minimal
model. However, on a closer look even the notion of prime model is not an
adequate explication of the idea of a minimal model. For one thing, the way this
notion is usually introduced is not useful to us a.s such. What we are looking
for are some intuitive structural cha.racterizations of minimality and maxima.lity,
a.nd the notion of primeness does not give us such a cha.racterization.
When I say this I mea.n the following: What made the idea of atomicity
so appea.ling is that there is a clea.n syntactically definable notion of a kind of
individual which enabled us to speak of a model in which a.minimum of such
"kinds" were insta.ntiated. More ~enerally, what we a.re looking for a.re cha.rac-
terizations of a minimal model M in terms of the constituent representation of
the complete theory Th(M) true in M. For the constituent representation is
in some obvious sense a.n explicit descriptjon of the most ea.sily understandable
features of the structure of M. In a sense, therefore, we want to have a cha.rac-
terization of minimality whose applicability can so to speak seem directly from
the theory Th(M). Now it surely cannot be seen directly from model M itself
or from the theory Th(M) whether M can be embedded elementarily in certain
other models.
234 NEW FOUNDATIONS FOR MATlIEMATICAL THEORIES
Of course, another way of defining primeness might do the trick. But the
most prominent alternative characterization of a prime model, viz. to character-
ize it as a countable atomic model, does not fare much better. From a theory it
is very hard to see directly what the cardinality of its several models might be.
Moreover, the notion of prime model is subject to most of the same ob-
jections as were marshaled above against atomic models as implementations of
minimality. For instance, even though the structure of natural numbers N is the
unique prime model of the Peano arithmetic, it is not the prime model of all the
consistent extensions of this arithmetic, viz. of those which are not true in N.
The most flagrant source of dissatisfaction is the fact that a prime model
of a complete theory might be elementary equivalent with a proper submodel
of itself. An example is offered by the theory of dense linear order, where a
prime model, for instance, the structure of the rationals, could obviously be
elementarily equivalent with its proper submodel. Hence prime models are not
always minimal models in any intuitive sense of the word.
4. Super models. The explanation of the failure of special models to
implement the extremality idea is not very hard to see. Types are not the right
explication of the idea of "kinds of individuals" existing in a model M. A type,
say a one-type, characterizes a kind of individual in so far as this individual is
considered alone. In order to catch full the idea of a kind of individual, we have
to consider them also in relation to the other individuals in the model.
This refined idea of a "kind of individual" can be captured by means of
the following definition:
Let M be a model and let aI, a2, . .. be a sequence of members of the
domain dom(M) of M. Let
(4.1) Ci(~,k)[x,aI ,a2, ... ,akJ (d = 1,2, ... , k = 0,1,2, ... )
be a (double) sequence of mutually consistent constituents compatible with
the complete theory Th(M) true in M. Assume also that the constituents
citl,k)[{x},aI,a2, ... ,akJ are all true in M. Then (4.1) is said to define "1.-
pertype in A = {aI, a2, ... } relative to M.
The justification of formulating the definition of a supertype in this way
is that the theory defined by (4.1) clearly does not depend on the order of the
at, a2, ..
The corresponding sequence with individual variables instead of con-
stants, i.e.,
(4.2) (d)
C;(d,k)
[
x, YI, Y2, Yk J
can be called the "tructure of the "upertype (4.1), alias a "upertype "tructure.
Many of the same things can be said mutati" mutand~ of supertypes as
can be said of types. For instance, the supertype (4.1) is said to be atomic iff it
stops branching after a certain point. More explicitly, (4.1) is atomic iff it has a
member
(4.3)
LANGUAGE, TRUTH AND LOGIC 235
such that for any e and any c}, C2 CI E dom( M) there is only one constituent
of the form
(4.4)
compatible with (4.3).
One of the basic properties of supertypes is the following
LEMMA 4.1: Given M and a supertype (4.1) compatible with the com-
plete theory Th( M) true in M, each member of the sequence (4.1 ) (and hence
each initial segment of (4.1 is~atisfied in M by some individual b, i.e., there is
b E dom( M) such that
(4.5) M 1= (d) [ 1
Ci(d.II:) b,al,a2,'" ,all: .
This follows clearly from Lemma 2.7.

Hence in a sense each initial segment of each supertype compatible with
Th(M) is satisfied in M. The only open question is whether the entire supertype
is.
From what has been said it follows that each atomic supertype is satisfied.
Assume now that (4.1) defines an atomic supertype in M and that
(4.6) C~(!./)[x, aI, a2,' , a,]
is the member of (4.1) after which (4.1) no longer branches. We shall say that
(4.6) determine" the atomic supertype. Let us also assume that Th(M), repre-
sented in the form (3.2), is the complete theory true in M.
One the assumptions just stated, we have
LEMMA 4.2: Let
(4.7) C~e+f)[x, aI, a2, , a" b1, b2 , , bml
be any constituent compatible with (4.6) and Th(M). Then we must have j =
i(e + f, I + m), i.e., (4.7) must be a member of (4.1).
Proof: This is what it means for (4.1) to stop branching at (4.6).
LEMMA 4.3: On the same assumptions, each member of (4. 1) later than
(4.6) is equivalent with all of its successors, given Th(M).
Proof: Each member of (4.1) is implied by its successors by Lemmas 2.4
and 2.6. Hence what we have to prove is that it implies them, given Th(M).
For this purpose, it suffices to show that (4.6) is not compatible (together with
Th(M with any other constituent of the form
(4.8)
In order to see this, let (4.8) be compatible with Th(M) and (4.6). Then
by the same reasoning as in Lemma 2.7, there are b1 , ~, , bm E dom( M)
such that (4.7) occurs in some supertype (4.1) of M. But if so, by Lemma 4.2,
j = i( e + f, I + m), in other words, there is only one constituent of form (4.8)
compatible with Th(M) and (4.6).
236 NEW FOUNDATIONS FOR MATHEMATICAL THEORIES
LEMMA 4.4: If Th(M) implies that, in a sequence like (4.1), each mem-
ber is equivalent with its successors from (4.6) on, then (4.6) determines an
atomic supertype.
Proof: If (4.6) is equivalent with each if its successors, say (4. 7) with
j < i(e + f, 1+ m), then by Lemma 2.1 it is incompatible with (4.7) with any
other j. In other words, (4.1) can be continued from (4.6) in only one way, i.e.,
(4.6) determines an atomic supertype.
THEOREM 4.1: On the same assumptions as in Lemma 4.2,
(4.9)
Th{M) f- (4.6) :::> (Vyd{VY2)" . (VYm)C~te"1],'+m)[X, at, a2, ... ,ai, YI, Y2, ... , Ym]
Moreover, if (4.9) holds for all F, M, (4.6) determines an atomic supertype.
Proof: From Lemma 4.3 we have
(4.10) Th{M) f- ((4.6) :::> Cf::J!l+m) [x, ai, a2, .. . ,ai, bl'~'''' ,bmD
From this (4.9) follows by first-order logic.
Conversely, if (4.9), then (4.6) implies (given Th{M all its successors.
By Lemma 4.4, it determines an atomic supertype.
THEOREM 4.2: On the same assumptions
(4.11) Th{M) f- {VZd{VZ2)"'{VZI){C~t!,/)[X,Z}'Z2""'Z,]:::>
{VYI )(VY2)' .. (VYm)C~te"1].l+m)[x, Zl, Z2,' .. ,ZI, Yl, Y2, ... ,Ym])
Conversely, if (4.11) holds, (4.6) determines an atomic supertype, provided that
(4.6) is true in M.
Proof: In the same way as in Theorem 4.1.
THEOREM 4.3: Assume that (4.6) determines an atomic supertype.
Then the same supertype structure is determined by a constituent of the form
(4.12) cje+I)[x].
Proof: Consider
(4.13) c~te~l,,)[x,a},a2, ... ,al]'
By Lemma 4.3, (4.13) is equivalent with (4.6), given Th(M). The formula
(4.14) c~te~l,I)[X, {ad, {a2}, .. . ,{ad]
is of the form (4.12), and in fact can serve as (4.12).
In order to show that this is what we want, it suffices in virtue of Lemma
4.4 that for each f and m there is a constituent of the form
(4.15) c.<e+I+J)[X,Yl,Y2, ... ,Ym]
implied by (4.14), given Th(M).
Now, in virtue of Lemma 2.8, (4.14) implies
(4.16) (3zt}(3z 2 )'" (3ZI)C~(~./)[x, Zb Z2,'" Z,].
In virtue of Lemma 4.3, there is a unique constituent of the form (4.15)

implied by Th(M) and
(4.17) c~t!.I}[x, Z1I Z2, .. , ZIJ.
But there obviously is such a unique constituent implied by Th(M) and (4.14).
Results like Theorem 4.3 are interesting in a wider perspective. At first
sight, it might seem that the switch from special models to superspecial models
destroys the strategic advantages offered by concepts like type and atomicity.
The crucial thing about them is how they help us to read off (as it were) the
structural (model-theoretical) properties of the models of a theory, especially a
complete theory, from the syntactical structure of this theory. The way in which
the atomicity of a model of a complete theory hangs together with the structure
of types in the constituent representation of this theory is a typical example of
this strategy.
It might seem that the way supertypes are defined deprives us of the use of
this syntax-to-models strategy. For supertypes are defined by reference to some
given model. Hence it seems circular to study supertypes and their interrelations
for the purpose of gaining insights into the structure of models.
What the theorems just proved show is that this impression is mistaken.
Even though supertypes are defined by relation to one particular model, some
of their most crucial properties depend only on the structure of the complete
theory Th(M) true in M. In particular, what the atomic super types of M are is
in a certain sense completely determined by Th(M). For instance, each atomic
supertype is determined by a constituent of form (4.12) or C~d)[xl occurring in
the constituent representation of the given theory. Here (4.12) does not depend
on any particular member of dom(M).
One application of these observations is that we can define the notion of
"uperatomicity for a complete theory in analogy with the definition of atomicity
of complete theories. the complete theory Th(M) true in M is superatomic
iff each initial segment of each "upertype structure compatible with Th(M) is
compatible with (the structure of) an atomic supertype.
Then we can also define a superatomic model. A "uperatomic model M is
one in which only atomic supertypes are satisfied, and each different supertype
by precisely one individual. Then we can also prove easily the following:
THEOREM 4.4: Each superatomic complete theory has a superatomic
model.
It is also easily seen that this superatomic model is uniquely determined
(up to isomorphism).
THEOREM 4.5: A model M of a complete theory T = Th(M) is super-
atomic iff it is a prime model but none of its proper elementary submodels is
prJme.
In order to prove Theorem 4.5, we can first prove
LEMMA 4.5: A superatomic model M of a complete theory T = Th(M)
is prime.
In order to show that M is prime, we have to show that it can be ele-

mentarily embedded in an arbitrary model M* of T. Now each member hi of
dom(M) satisfies a supertype determined by a constituent of the form
(4.18) C:(i)[x).
But, by Lemma 4.1 (4.18) must be satisfied by some individual, call it hi in M*.
The mapping of each hi on hi defines an embedding of Minto M*. It is easily
seen that this is an elementary embedding.
To return to the proof of Theorem 4.5 it is clear that if we try to map
M elementarily into itself, each hi E dom(M) must be mapped on itself. For
the image hi must satisfy the same constituent (4.18) as hi, the mapping being
an elementary embedding. But the only member of dom(M) to satisfy this
constituent is hi itself. This proves Theorem 4.5.
In this way we can also see that one of the main shortcomings of prime
models as explications of minimality is overcome by superatomic models.
This is prima facie a major difference between atomic and superatomic
models in that in an atomic model, every k-type, for k = 1,2, ... , satisfied in
it must be atomic, whereas the definition of a supertype involves directly only
formulas with one free individual variable. These apparently correspond to one-
types only.
To reassure the reader, we can prove
LEMMA 4.6: A superatomic model is atomic.
In order to show this, assume first that M is a superatomic model of
Th(M). Then each element of dom(M) satisfies an atomic supertype. If
(al' a2, .. , ak) is a k-tuple of such elements, one can see by the same line of
argument as was given fro Theorem 4.1 that there is a consistent structure
(4.19) C~d)[X1.X2, ,XA:)
satisfied by (all a2, ... , ak) (and hence compatible with Th(M such that each
formula
(4.20)
defines a supertype. Then for each e there is only one constituent of the form
(4.21) c(d+e)[Xl,X2, ,XA:]
compatible with (4.18) and with Th(M). In other words, (4.18) determines an
atomic type.
This proof illustrates how we can get along in our "supertheory" by means
of supertypes with only one free variable, i.e., with what prima lacie should be
called one-supertypes. Just because in supertypes we heed the relation of the
kinds of individuals characterized by them to other individuals in the models,
we do not need k-supertypes with k ~ 2.
In order to extend one horizon to arbitrary theories instead of just com-
plete ones, we must first extend our main concepts.
A supertype is said to be .strongly atomic if it stops branching in all models
compatible with one of its initial segments.
More explicitly, a constituent

(4.22)
determines a strongly atomic supertype iff it determines an atomic supertype in
any model of (4.22).
A theory T is 3trongly 3uperatomic iff each initial segment of each super-
type structure compatible with T is compatible with (the structure of) a strongly
atomic supertype.
A strongly superatomic model M is one in which only strongly atomic
supertypes are satisfied, and each of them by precisely one individual.
It is now easy to prove suitable extensions of our earlier results, for in-
stance
THEOREM 4.6: Each complete theory compatible with a strongly su-
peratomic theory has a strongly superatomic model.
The most natural generalization of the notion of saturation (of a model
M) is not equally directly connected with the structure of the complete theory
Th{M).
DEFINITION: A model M is ab301utely 3uper3aturated iff each supertype

compatible with Th{M) is satisfied.
A model M is supersaturated relative to a set of individuals A = {ai} (i E
I), A dome M) iff each supertype (4.1) with aI, a2, . .. E A compatible with
~
Th(M) is satisfied in M.
These notions still rely fairly heavily on the particular model M. How-
ever, the insights so far reached enable us to define a somewhat less demanding
characteristic of a model which will turn out to be most useful.
DEFINITION: A model M is atomically 3uper3aturated iff

(i) Each atomic supertype structure is satisfied in M by precisely one indi-
vidual Ql j and
(ii) M is supersaturated with respect to the set A = {Qi} of the individuals
satisfying the different atomic supertype structures.
It is easily seen that the following theorem holds:
THEOREM 4.7: Each complete theory has an atomically saturated

model.
Many results familiar from the traditional model theory have related re-
sults that can be proved for supertypes. Here I shall mention only one as an
example.
THEOREM 4.8: From a model Ml of a complete theory T one can omit

a countable number of individuals, each satisfying only nonatomic supertypes,
and obtain a model M2 of T which is an elementary submodel of MI.
5. Super special models as implementations of extremality re-

quirements. The concepts defined in the preceding section serve as excellent
explications of the notions of minimality and maximality which are the focal
ideas of this study. The minimality requirement (principle of paucity) is natu-
rally captured by the idea of strong superatomicity, which literally amounts to
imposing on models the least possible qualitative variety in so far as the relevant
qualitative differences are understood by references to supertypes. This proce-
dure is vindicated by the fact that superatomic models of complete theories turn
out to be the minimal prime models, i.e., models elementarily embeddable into
any model of the given complete theory, but not into any of their own proper
submodels.
Strongly superatomic models turn out to be capable of doing the kind of
job they were cast to do. For one thing, a Peano-type axiomatization of elemen-
tary arithmetic turns out to be categorical and have the structure N of natural
numbers as its sole model, if the space of models is restricted to strongly super-
atomic ones. The only strongly superatomic model of the Peano axiomatization
of elementary number theory can be shown to be the structure N of natural num-
bers, if the space of models is restricted to strongly superatomic models. This
does not conflict with Godel's incompleteness result, because the new "paucity
logic" is not axiomatizable. Hence the new perspective on elementary arith-
metic does not automatically create new avenues of actually establishing new
number-theoretical results. What it nevertheless can in principle do is to facil-
itate the discovery of stronger and stronger proof methods. For the search for
such methods can now be guided by clear-cut seman tical considerations.
But how do I know that the structure of natural numbers N is the only
superatomic model of the Peano axioms? These axioms are compatible with a
number of different complete theories, only one of which is true in N. We can
call it Th(N). It is easily seen to be strongly superatomic. It has different non-
isomorphic models, of which N is one. It is easily seen that N is in fact the only
superatomic model of Th(N).
But what about the other complete theories compatible with Peano arith-
metic? How do we know that they do not have strongly superatomic models,
too?
Perhaps the quickest way of seeing that they do not is to note a trivial-
looking property of strongly superatomic models. Let each member of such a
model M, say b, be correlated one-to-one with one of the constituents C}dl[x]
which determines the strongly atomic supertype that bi satisfies. Given two
such individuals bi and b2 , the second one being correlated with CJe)[x], one can
construct effectively the formula that determines the strongly atomic supertypes
satisfied by their sum, likewise for their product. This means that sum and
product are recursive relations in a strongly superatomic model.
The details of this argument are given in an appendix below.
Now Tennenbaum has shown (see Tennenbaum 1959; Feferman 1958;
Scott 1959; Kaye 1991, p. 153) that sum and product are not recursive in any
non-standard model of Peano arithmetic (in the sense of relative recursivity just
LANGUAGE. TRUTH AND LOGIC 241
explained). From this it follows, together with the observations just made, that
non-standard models of Peano arithmetic cannot be superatomic, just as was
claimed.
Things are somewhat more complicated with respect to the notion of
maximality (principle of plenitude). Hilbert's completeness axiom amounted to
requiring that any attempted adjunction of a new individual to the intended
model must lead to a violation of the other axioms. But such requirements can-
not be satisfied in first-order theories in view of the upwards Skolem-LOwenheim
theorem. In Hilbert's axiomatization of geometry, his completeness axiom has
the intended effect only because he had also assumed the Archimedean axiom
and also tacitly interpreted the notion of natural number involved in the Archi-
medean axiom in the standard sense. Fortunately, the intended maximality
conditions can typically be interpreted so as to require only maximal qualitative
richness, not necessarily the presence of the maximal selection of individuals in
the intended model or models. Hence the natural course for us here is, if we
want to keep our conceptualizations generally applicable, to require only maxi-
mal qualitative richness but not completeness in Hilbert's strong sense. But this
does not really mean giving up Hilbert's original ideas. For even geometry, the
function of the completeness axiom is to enforce continuity, not to restrict the
"size" of the universe of discourse. Indeed, Hilbert's completeness axiom can be
replaced by a pair of assumptions that can be roughly expressed as follows:
(H.I) If two points have the same relations to all other points and lines, they
are identical.
(H.2) If M is a model of the other axioms and if there is a set of relations between
an unspecified individual x and the members of M which is compatible
with the other axioms and with the diagram of M, there exists in M an
individual with these relations.
As you can easily see, (H.I) follows from other axioms. (Axioms of inci-
dence and order suffice for the purpose.) Hence the import of the completeness
axiom is essentially (H.2), which is an assumption of maximal qualitative rich-
ness rather than of maximal size as far as individuals 'are concerned. In fact,
the force (H.2) is easily seen to amount to requiring that maximal number of
supertypes be instantiated, compatible with the other assumptions.
Hence we can safely think of the maximality idea as being captured by
a requirement of maximal qualitative richness. But in the preceding section we
found that we have a genuine choice here. We can require either absolute super-
saturation or atomic supersaturation of our models. The difference between the
two appears nevertheless to be relatively unimportant. For one thing, the former
implies the latter. Furthermore it will turn out that even atomic supersaturation
is quite a strong assumption.
More has to be done here, however, than to explicate the twin notions
of minimality and maximality. In the most interesting mathematical theories
beyond elementary number theory, such as the theory of reals, axiomatic ge-
ometry, and set theory, the crucial thing turns out to be neither minimality
assumptions nor maximality assumptions, but their interaction. Typically, we
can assume that we are dealing with a theory which contains a one-place pred-
icate, say N(x), for natural numbers, and suitable axioms for natural numbers.
(That is, when the axioms of the theory are relativized to N(x), they must yield
as consequences a reasonable axiomatization of natural numbers.) Notice that
N(x) does not necessarily have to be a primitive predicate. Then, we obviously
have to assume that part of a model of the theory which corresponds to natural
numbers is superatomic but that the rest of the model is maximal. But maximal
in what sense? The crucial fact here is that we cannot simply assume that the
individuals satisfying "-'N(x) form a atomically supersaturated model, for that
may be incompatible with the requirement that {x : M F N(x)} is superatomic.
We can only require that the model realizes a maximal number of supertypes
(either absolutely or reelative to the set of individuals satisfying superatomic
types) compatible with the requirement that {x: M F N(x)} be superatomic.
We have thus motivated the following definitions:
Let us assume that we are given a model M of a complete theory Th( M)
which contains a one-place predicate N(x) for natural numbers. Let us assume
further that the theory Th(M) as restricted to {x: M F N(x)} is superatomic.
Then the model M is abJolutely Hilbertian iff the following requirements are also
satisfied:
(i) M restricted to {x : M F N(x)} is superatomic.
(ii) A maximal. subset of supertypes compatible with (i) are instantiated in
M.
M iR an atomically Hilbertian model iff the following conditions are sat-
isfied:
(i) As before.
(ii)* A maximal number of supertypes relative to the set of individuals satis-
fying a superatomic type are instantiated in M.
For instance, consider a set of axioms for real numbers which includes a
predicate N(x) for natural numbers. Then (i) becomes essentially the Archime-
dean axiom. By the usual Dedekind-type line of thought, one can then show
that the structure of the (actual) reals is the only one which is also atomically
Hilbertian.
Essentially, the same also happens in Hilbert's axiomatization of geom-
etry. Hilbert needs the Archimedean axiom (utilizing the standard concept of
natural number) to force as it were the multiples of the unit line to match the
structure of natural numbers, and the axiom of completeness to ensure conti-
nuity. The latter point is especially clear in Hilbert (1900), where the axiom
of completeness first made its appearance (as an ruoom for the theory of reals
rather than as an axiom of geometry). Hilbert's formulation of his axiom there
also makes it clear that he thought of it as a maximality assumptioil.
If we give up requirement (i), we can obtain sundry non-standard models
of reals. If we give up (ii), we need not any longer have all "real reals" in our
model. Depending on the axioinatization, it may be sufficient, e.g., for the model
to contain only all algebraic numbers.
Thus, we can again reach one of our main objectives. If we restrict the
models of (the first-order language of) a theory of reals to atomically Hilbertian
ones, then any reasonable theory of reals is categorical and yields the intended
structure of reals as its only model (up to isomorphism). This descriptive com-
pleteness is not due to the requirement of minimality (superatomicity) alone,
nor to the requirement of maximality (atomic supersaturation), but to the com-
bination of the two in the requirement of the (atomically) Hilbertian character
of the models.
An interesting pitfall here is that a complete theory need not have a
unique richest model of a given cardinality, either in the sense of being absolutely
Hilbertian or atomically Hilbertian. IT we think of the supertypes compatible
with the given theory satisfied one by one, then the ultimate outcome can so to
speak depend on the order in which they are satisfied.
An interesting situation arises when the same ideas are extended to a
sui table axiomatization of set theory. In order to apply the ideas sketched here,
we have to assume that a predicate N(z) for natural numbers is included in
the language of the set theory or can be defined as the basis of the axiomatiza-
tion. Then we can again stipulate that the models be restricted to atomically or
absolutely Hilbertian ones, and see what happens.
I cannot here try to answer this question in general. Certain things are
nevertheless relatively easy to see. Perhaps the most interesting perspective of-
fered by our observation, is that in set theory, too, the greatest subtlety is due to
the interplay of minimality and maximality requirements. On the one hand, one
can construct poor (small) models of, say, ZF set theory which have a clear-cut
structure but which clearly are not what is intended. On the other hand, at-
tempts to enlarge the universe of set theory have not yielded any ultimate clarity
either. It seems to me that the real source of difficulties in set theory is that the
requirements of poverty and plenitude have to be balanced against each other.
For another example, we can construct a theory of finite types as a many-sorted
first-order theory. We might, e.g., assume that there is a primitive predicate
N(z) in the language for natural numbers which are among the individuals. IT
we then require that the models of a suitable axiomatization of such a type theory
are automatically Hilbertian, the resulting theory has all sorts of nice features.
For instance, the Denumerable Axiom of Choice is valid and so is the Principle
of Dependent Choices for subsets of natural numbers. Furthermore, it will be
easy to give a descriptively complete and indeed categorical axiomatization for
a theory of the second number class (countable ordinals).
In this kind of many-sorted first-order reconstruction of type theory we
can even start from an axiomatization of a discrete linear order (with an initial
element) for natural numbers. Its only superatomic model is clearly {O, 1,2, ... }
with successor as the only relation. It is easily seen, however, that functions for
addition, multiplication, etc., all necessarily exist in all the models of the full
axiomatization, as indeed do all recursive functions. Thus the existence of the
usual arithmetical functions does not even have to be assumed; it follows logically
from the axioms. This would speak for a partial reducibility of mathematics to
logic, if it were not for the fact that certain mathematical assumptions were built
right into our concept of model and hence into our concept of logic.
A more sweeping philosophical perspective which opens here takes up
an issue which was mooted by Hilbert and Kronecker. For Kronecker, natural
numbers were the be-all if not the end-all of mathematics. In contrast, one of
Hilbert's acknowledged aims in his Grundlagen der Geometrie was to show
that there can be important mathematical theories which do not involve the
concept of natural numbers at all. (See here Blumenthal 1922, p. 68.)
If the approach advocated here is right, there is more to be said for Kro-
necker and less for Hilbert than has been generally acknowledged. If the subtlety
of advanced mathematical theories lies in the interplay of superatomicity require-
ment for the natural numbers with suitable maximality assumptions, then the
concept of natural numbers is after all essentially involved in these mathematical
theories, via the requirement of superatomicity. Mathematics looks more like a
science of (natural) numbers than it has in a long time.
On a more technical level, there does not seem to be any obstacles in
principle to use the time-honored strategy of using set theory to speak of its own
semantics. In this way, e.g., Godel captured his own metalogical construction
of a constructible model in an explicit axiom. If this strategy works here, the
requirement that only atomically Hilbertian models are considered would' be ex-
pressible by an explicit set-theoretical axiom of the old style (without restrictions
on the usual set of models). This axiom would be eminently acceptable, for (i)
merely spells out the nature of natural numbers while (ii) follows from the idea
that set theory is the theory of all sets, that in the world of set theory what can
exist doe" exist.
Whether or not we can along these lines solve the outstanding problems of
set theory remains to be seen. I do not seem to be the only one who thinks that
they can be so solved. Godel once Wrote to Ulam, apropo" John von Neumann's
axiomatization of set theory:
The great interest which this axiom [in von Neumann's axioma-
tization of set theory] has lies in the fact that it is a maximum
principle somewhat similar to Hilbert's axiom of completeness in
geometry. For, roughly speaking, it says that any set which does
not, in a certain defined way, imply an inconsistency exists. Its
being a maximum principle also explains the fact that this axiom
implies the axiom of choice. I believe that the basic problems of
abstract set theory, such as Cantor's continuum problem, will be
solved satisfactorily only with the help of stronger axioms of thi.s
kind, which in a sense are opposite or complementary to the con-
structivistic interpretation of mathematics ..(See Ulam 1958.)
What Godel misses here is the crucial interplay between maximality and
minimality assumptions, though his remark on von Neumann type axioms being
complementary to constructivistic ideas perhaps suggests some degree of aware-
ness of this fact.
Appendix. Let us assume that a strongly superatomic model M ofPeano
arithmetic has been given and that a one-to-one correlation has been established
between all natural numbers n and all the strongly atomic supertypes of M.
More explicitly, let the correlate <pen) of each n be one of the constituents with
one free individual variable that determine a strongly atomic supertype in M,
different supertypes for different values of n. (Equivalently, the correlate of n

could be the Godel number of this constituent.) This correlation establishes
an isomorphism between M and a certain relational structure. (Cf. here Kaye
1991, especially sec. 11.3.) In this isomorphism, a certain numerical relation will
correspond to the relation of being the sum of in M, i.e., the relation which
holds between three individuals in dome M) say a, h, c. when S( a. h. c) is true in
M. where 5 is the expression of the sum of in Peano arithmetic. The question
is whether this relation is recursive.
In order to show that it is. let us suppose that we are given three con-
stituents each of which is correlated with some natural number by II' and each of
which therefore determines a strongly atomic supertype. Let these constituents
be
(1) C~")[z]
(2) C~e)[y]
(3) C~f)[z].
From (1) and (2) we can form the formulas:
(4) C~")[z] & C~e)[y] & S(z.y.z)
(5) (3z)(3y){C~")[z] & C~")[y] & S(z.y.z

By as~umption. (1) and (2) each determines a strongly atomic supertype.
Clearly (5) is satisfied by the sum of the two individuals in dom(M) which satisfy
(1) and (2). What we are interested in here is whether it is possible to determine
recursively (effectively) whether (3) is also satisfied by this sum.
For the purpose, we shall first show that (5) determines a strongly atomic
supertype. In order to prove this. assume that M* is a model in which (5) is
satisfied. Hence (4) is also satisfied in M*. What has to be shown is that, given
al. a2 ... al: E dome M*) and 9 ~ max( d, e) +2. there is only one constituent of
the form
(6) C!g)[z,al.a2 ... al:]
compatible with (5) and Th(M*).
In order for (6) to be compatible with (5). it must be compatible with (1)
and (2). Since (1) and (2) both determine strongly atomic supertypes, there is
a unique constituent of the form
(7) C (g+I)[ y.z.al,a2.. al: ]
compatible with (5) and Th(M). But since it follows from the axioms of Peano
arithmetic that the sum of two numbers is uniquely determined. there is in (7)
one and only one constituent of the form
(8) CJg)[z. y. z. at. a2 ... al:]
compatible with (4). Hence the constituent
(9)
246 NEW FOUNDAnONS FOR MA THEMATICAL THEORIES
is the only constituent with the parameters aI, a:lt ... , all E dom(M*) and 9
compatible with (5) and Th(M*). In other words, it is the only constituent
which can serve as (6), which is therefore uniquely determined. This is just
what was to be proved.
It is important to realize that this part of the overall proof is not supposed
to be effective.
Consider now the unique constituent
(10) ci.m&X(d,e
compatible with (4). It is true in M, and it can be obtained effectively from (4).
Assume first that f < max(d, e). Because (3) determines a strongly
atomic supertype in M, there is in (9) a unique constituent of the form
(11) cimax(d,e)-l)[x]
compatible with (3). It can be found effectively, given (3) and (14), and it clearly
determines a strongly atomic supertype.
Assume then that f ? max(d, e). Then there is a unique constituent of
the form (11) in (10) compatible with (3). In this case it can be found simply
by omitting layers of quantifiers from (3), hence effectively.
In either case, since (1) and (2) both determine a strongly atomic super-
type in M, there is a unique constituent of the form
(12) C~m&X(d.e[z, y]
compatible with (4). It can be found effectively as follows: First we convert (4)
into its distributive normal form
(13) V Cfmax(d,e[z, y]
This can be done effectively. However, we cannot in general know which

disjuncts in (13) are consistent and which ones are not. This uncertainty can
be eliminated simply by grinding out the logical consequences of (13) jointly
with (4) one after the other until only one survives undisproved. But it follows
from the axioms of Peano arithmetic that the sum of two individuals is uniquely
determined. Hence there is in (12) a unique constituent of the form
(14) c~max(d,e)-l)[X, z, y]
But since (14) is uniquely determined, then so is
(15) c~max(d,e)-l)[x, {y}, {z}]
Now this constituent can be compared with (11) effectively for identity. IT
the two are identical, (3) and (5) determine the same strongly atomic supertype,
if not, they do not.
By reviewing the argument, it is easily seen that this determination can
be made effectively. By Church's thesis, sum will therefore be a recursive relation
in M, which was to be proved.
Acknowledgments. In working on this paper, I have profited from the
comments, suggestions and criticisms by Professor Jouko Viiii.niinen, by several
members of his research group in Helsinki, and by Professors David McCarty

and Philip Ehrlich. They are not responsible for any errors, however. My par-
ticipation in the ASL European Summer Conference in 1990 was facilitated by
a travel grant from the Academy of Finland, and my work by research support
from Boston University.
REFERENCES
JOHN D. BALDWIN, Fundamentals of Stability Theory, Springer-
Verlag, Berlin, 1988.
OTTO BLUMENTHAL, David Hilbert, Die Naturwissenschaften, vol.
10 (1922), pp. 67-72.
C. C. CHANG and H. J. KEISLER, Model Theory, North-Holland, Am-
sterdam, 1973.
SOLOMON FEFERMAN, Arithmetically Definable Modelj of Formalized
Arithmetic, Notices of the American Mathematical Society, vol. 5 (1958),
pp. 679-680.
DAVID HILBERT, Foundations of Geometry, tr. by Leo Unger, tenth
ed., Open Court, La Salle, 1971. German original Grundlagen der Geome-
trie, 1899.
DAVID HILBERT, Uber den ZahlbegritJ, Jahresberichte der Deutschen
Mathematiker- Vereinigung, vol. 8 (1900), pp. 180-184.
JAAKKO HINTIKKA, lJ there Completenejj in Mathematicj after Godel?
Philosophical Topics, vol. 17, no. 2 (1989) pp. 69-90.
RICHARD KAYE, Models of Peano Arithmetic, Clarendon Press, Ox-
ford, 1991.
VEIKKO RANTALA, Aspects of Definability (Acta Philosophica Fen-
nica, vol. 29, nos. 2-3), North-Holland, Amsterdam, 1977.
HARTLEY ROGERS, Jr., Theory of Recursive Functions and Effec-
tive Computability, McGraw-Hill, New York, 1967.
DANA SCOTT, On Corutructing Modelj for Arithmetic, in Infinitistic
Methods, Pergamon Press, Oxford, 1959, pp. 235-255.
S. TENNENBAUM, Non-archimedean Modelj for Arithmetic, Notices of
the American Mathematical Society, vol. 6 (1959), p. 270.
STANISLAW ULAM, John von Neumann, 1903-1957, Bulletin of the
American Mathematical Society, vol. 64 (1958, May Supplement), pp. 1-
49.
Department of Philosophy
Boston University
Boston, MA 02215, USA
JAAKKO HINTIKKA SELECTED PAPERS
1. 1. Hintikka: Ludwig Wittgenstein: Half-Truths and One-and-a-Half-

Truths. 1996 ISBN 0-7923-4091-4
2. 1. Hintikka: Lingua Universalis vs. Calculus Ratiocinator: An Ultimate
Presupposition of Twentieth-Century Philosophy. 1997
ISBN 0-7923-4246-1
3. 1. Hintikka: Language, Truth and Logic in Mathematics. 1998
ISBN 0-7923-4766-8
4. J. Hintikka: Paradigms for Language Theory and Other Essays. 1998
ISBN 0-7923-4780-3
KLUWER ACADEMIC PUBLISHERS - DORDRECHT / BOSTON / LONDON

(Jaakko Hintikka Selected Papers 3) Jaakko Hintikka (Auth.) - Language, Truth and Logic in Mathematics-Springer Netherlands (1998)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Jaakko Hintikka Selected Papers 3) Jaakko Hintikka (Auth.) - Language, Truth and Logic in Mathematics-Springer Netherlands (1998)

Uploaded by

Copyright:

Available Formats

LANGUAGE, TRU1H AND LOGIC IN MATHEMATICS

JAAKKO HINTIKKA SELECTED PAPERS

SPRINGER-SCIENCE+BUSINESS MEDIA, B.V.

ISBN 978-90-481-4923-0 ISBN 978-94-017-2045-8 (eBook)

Printed on acid-free paper

All Rights Reserved

ORIGIN OF THE ESSAYS vii

1. "What Is Elementary Logic? Independence-Friendly Logic as the True

2. (with Gabriel Sandu) "A Revolution in Logic?" 27

3. "A Revolution in the Foundations of Mathematics?" 45

4. "Is There Completeness in Mathematics after G6del?" 62

6. "Standard vs. Nonstandard Distinction: A Watershed in the Foundations of

7. "Standard vs. Nonstandard Logic: Higher-Order, Modal, and First-Order

8. (with Gabriel Sandu) "The Skeleton in Frege's Cupboard: The Standard

9. (with Arto Mutanen) "An Alternative Concept of Computability" 174

11. "Model Minimization - An Alternative to Circumscription" 212

12. "New Foundations for Mathematical Theories" 225

1. "What Is Elementary Logic? Independence-Friendly Logic as the True Core Area

One can distinguish, roughly speaking, two different approaches to the

mathematics. Its nature and its implications have remained unacknowledged to

WHAT IS ELEMENTARY LOGIC?

1. IS ORDINARY FIRST-ORDER LOGIC

My ultimate aim in this paper is to show that conventional wisdom is

The claims of first-order logic to a special status are often reinforced

2. QUANTIFIER DEPENDENCE AS THE ESSENCE

quantifier dependence. This is the reason why the order of dissimilar

3. QUANTIFIER INDEPENDENCE AND ITS CODIFICATION

The relation of IF first-order languages to ordinary first-order lan-

However. they are easily seen to be incompatible with the requirement

(9) (V'x)(3y) '-....

However, the IF logic envisaged here is in several respects richer

(11) (A1[x] (v/&) A[x]) & (B1[x] (v/&) B 2[x])

(12) ('v'x)('v'y)Alty](vl'tx)A2[x,y])(vl'ty)(B 1[x.y ](vl'tx) B2[x,y]).

4. THE SCOPE OF SCOPE

we have a way of signalling dependence but not a general way of

and (ii) is continuous (without lacunae). Moreover, (iii) quantifier scopes

However, when it comes to logic it might seem that [ am here quib-

S. THE FAILURE OF AXIOMATIC METHODS IN IF LOGIC

What has been established so far? It may have disturbed some of my

It is not that the failure of the axiomatic-deductive techniques in (a special

(20) ('<i x)(3y)

7. THE FAILURE OF TARSKI-TYPE TRUTH-DEFINITIONS

We have discovered that axiomatic-deductive methods cannot serve to

syntax and semantics together. It is most commonly called the prin-

evidence to suggest that independence of the kind we are dealing with

8. GAME-THEORETICAL SEMANTICS TO THE RESCUE

In spite of the failure of Tarski-type truth-definitions in a first-order IF

Godel's famous functional interpretation of first-order liogic and first-

9. THE GENERALITY OF THE PHENOMENON OF

At the same time, game-theoretical semantics shows that the phenom-

tors like necessity are in that semantics in effect universal quantifiers

of IF first-order logic, I mention here only IF (first-order) lepistemic logic.

theorem the separating formula can always be chosen to be an ordinary

9 See, e.g., S. MacLane, Categoriu for the Working Mathematician, Springer-Verlag,

Semantical Phenomenon", in J. E. Fenstad et al., editors, Logic" Methodology and

19 This is the leading idea of the applications of game-theoretical semantics to natural

Whom does each person admire most?

where the truth-making choice of a value of y depends OIL the value of x.

* Written jointly with Gabriel Sandu

J. Hintikka, Language, Truth and Logic in Mathematics

philosophical analysis, such as uncovering quantifier-switch fallacies, exploit

(2) ("i/X) (A(x) (vNx)B(x

It turns out that if we limit our attention to formulas in negation normal