Professional Documents
Culture Documents
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
INTRODUCTION ix
5. "Hilbert Vindicated?" 84
10. (with Gabriel Sandu) "What is the Logic of Parallel Processing?" 189
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):
All the previously published essays appear here with the permission of the respective
copyright owners, if any. These permissions are most gratefully acknowledged.
INTRODUCTION
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.
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
4 LANGUAGE, TRUTH AND LOGIC
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
WHAT IS ELEMENTARY LOGIC? 5
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.
6 LANGUAGE, TRUTH AND LOGIC
(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)
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
8 LANGUAGE, TRUTH AND LOGIC
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
WHAT IS ELEMENTARY LOGIC? 9
order logic reduces to that of IF logic. From this it follows that IF logic
cannot be axiomatizable.
6. QUALMS ALLAYED
('<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
14 LANGUAGE, TRUTH AND LOGIC
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.
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.
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
WHAT IS ELEMENTARY LOGIC? 19
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.
WHAT IS ELEMENTARY LOGIC? 23
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
WHAT IS ELEMENTARY LOGIC? 25
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]
or
(3) ("i/x)(3y)("i/Z)(Sl[X,y,z](vNx)S2[X,y,Z])
(5) (3y)("i/x)S[x,y]
(7) ('vIx)('vIz)(3yNz)(3uNx)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
30 LANGUAGE, TRUTH AND LOGIC
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:
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
A REVOLUTION IN LOGIC 31
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
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
(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, ...
32 LANGUAGE, TRUTH AND LOGIC
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
A REVOLUTION IN LOGIC 33
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
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
A REVOLUTION IN LOGIC 35
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
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
A REVOLUTION IN LOGIC 41
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
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.
A REVOLUTION IN LOGIC 43
11 See Jaakko Hintikka and Gabriel Sandu, "The Skeleton in Frege's Cupboard:
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.
Hintikka and Jack Kulas, The Game of Language, D. Reidel, Dordrecht, 1985, ch.
10.
18 Op. cit. note 13 above.
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
(Hxyzu)S[x,y,z,u] B (3f)(3g)(Vx)(Vz)S[x,f(x),z,g(z)].
(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",
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
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.
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:
(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
A REVOLUTION IN THE FOUNDATIONS OF MATHEMATICS? 49
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
50 LANGUAGE, TRUTH AND LOGIC
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)].
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.
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
52 LANGUAGE, TRUTH AND LOGIC
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:
where
1}'11: -,( x =y) 1}'12: -,( x =0)
({J21: -,(y =z) ({J22: -,(x =y) & -, (y =f(x
(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.)
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
54 LANGUAGE, TRUTH AND LOGIC
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
56 LANGUAGE, TRUTH AND LOGIC
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.
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
A REVOLUTION IN THE FOUNDATIONS OF MATHEMATICS? 57
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.
(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.
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
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
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.
60 LANGUAGE, TRUTH AND LOGIC
REFERENCES
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-
66 LANGUAGE, TRUTH AND LOGIC
III
FROM ONE WORLD (MODEL) TO MANY
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
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
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
IS THERE COMPLETENESS IN MATHEMATICS AFTER GODEL? 73
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
74 LANGUAGE, TRUTH AND LOGIC
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,
IS THERE COMPLETENESS IN MATHEMATICS AFfER GODEL? 75
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
76 LANGUAGE, TRUTH AND LOGIC
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
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
78 LANGUAGE, TRUTH AND LOGIC
X
IS MATHEMATICS REDUCIBLE TO LOGIC?
XI
WHAT DOES A MATHEMATICIAN REALLY DO?
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
NOTES
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?
(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.
J. Hintikka, Language, Truth and Logic in Mathematics
Springer Science+Business Media Dordrecht 1998
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.
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
86 LANGUAGE, TRUTH AND LOGIC
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;
88 LANGUAGE, TRUTH AND LOGIC
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.
(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
HILBERT VINDICATED? 89
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.
90 LANGUAGE, TRUTH AND 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
HILBERT VINDICATED? 91
(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)
. .. 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).
(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
HILBERT VINDICATED? 93
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.
(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
HILBERT VINDICATED? 95
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.).
(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
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
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]
(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
(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.
100 LANGUAGE, TRUTH AND LOGIC
(9) (V'X)S[X]
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.
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.
104 LANGUAGE, TRUTH AND LOGIC
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.
HILBERT VINDICATED? 105
Department of Philosophy
Boston University
Boston, MA 02215
USA
hintikka@bv.edu
6
... it is impossible to demonstrate the existence of an object without defining it. (Emphasis
in the original.)
Here the word define always means to name a property characteri.:ing what is defined.
... 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.
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.
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.
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-
114 LANGUAGE, TRUTH AND LOGIC
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-
116 LANGUAGE, TRUTH AND LOGIC
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
118 LANGUAGE, TRUTH AND LOGIC
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.
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
STANDARD VS. NONSTANDARD DISTINCTION 119
7. RAMSEY'S CONTRIBUTION
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.
STANDARD VS. NONSTANDARD DISTINCTION 121
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-
122 LANGUAGE, TRUTH AND LOGIC
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.
STANDARD VS. NONSTANDARD DISTINCTION 123
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,
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.
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
126 LANGUAGE, TRUTH AND LOGIC
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.
128 LANGUAGE, TRUTH AND LOGIC
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,
Dordrecht, pp. 301-326.
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
STANDARD VS. NONSTANDARD DISTINCTION 129
44 lAAKKO HINTIKKA
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
134 LANGUAGE, TRUTH AND LOGIC
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
138 LANGUAGE, TRUTH AND LOGIC
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.
STANDARD VS. NONSTANDARD LOGIC 141
NOTES
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
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
23 Barbara Hall Partee (ed.), Montague Grammar, Academic Press, New York, 1976,
and note 4 above.
8
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
THE SKELETON IN FREGE'S CUPBOARD 149
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.
150 LANGUAGE, TRUTH AND LOGIC
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:
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.
THE SKELETON IN FREGE'S CUPBOARD 151
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).
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).
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).
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.
THE SKELETON IN FREGE'S CUPBOARD 159
58 Foundations of Set Theory (Amsterdam: North Holland, 1958), esp. ch. II,
sects. 6-7.
162 LANGUAGE, TRUTH AND LOGIC
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":
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.
THE SKELETON IN FREGE'S CUPBOARD 163
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:
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
164 LANGUAGE, TRUTH AND LOGIC
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
"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'
THE SKELETON IN FREGE'S CUPBOARD 171
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.
172 LANGUAGE, TRUTH AND LOGIC
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:
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.
AN ALTERNATIVE CONCEPT OF COMPUTABILITY 177
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
178 LANGUAGE, TRUTII AND LOGIC
(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
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
182 LANGUAGE, TRUTH AND LOGIC
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
AN ALTERNATIVE CONCEPT OF COMPUTABILITY 183
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).
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.
AN ALTERNATIVE CONCEPT OF COMPUTABILITY 187
REFERENCES
Wang, Hao, 1974 From Mathematics to Philosophy, Routledge & Kegan Paul,
London.
10
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.
(2.4)
(2.6)
192 LANGUAGE, TRUTH AND LOGIC
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
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:
(3/d(3/2)(3g)(\t'x)(\t'z)((Fdx, hex), z]
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.
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
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
WHAT IS THE LOGIC OF PARALLEL PROCESSING? 195
the processing power is in an obvious sense measured by the logical strength of the
correlated first-order statement.
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)
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
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:
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
Then what the combined system Uo does can obviously be described by the
following second-order formula
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] > (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.
WHAT IS THE LOGIC OF PARALLEL PROCESSING? 199
('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.
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).
200 LANGUAGE, TRUTH AND LOGIC
(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*);
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
WHAT IS THE LOGIC OF PARALLEL PROCESSING? 201
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.
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
202 LANGUAGE, TRUTH AND LOGIC
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
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
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)
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.
~ al
'\.
a3
/
~ a2 '\.
a7 ~
~ a4 /
'\.
a6
/
~ a5
WHAT IS THE LOGIC OF PARALLEL PROCESSING? 205
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
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) :::: 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.
206 LANGUAGE, TRUTH AND LOGIC
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*
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
WHAT IS THE LOGIC OF PARALLEL PROCESSING? 207
(3w)(3f)(3g)(V'x)(V'z)x'= z ..... f(x) = g(z +-+ (x f:. wand z f:. w)). (11.2)
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.
WHAT IS THE LOGIC OF PARALLEL PROCESSING? 209
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",
210 LANGUAGE, TRUTH AND LOGIC
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
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.
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.
214 LANGUAGE, TRUTH AND LOGIC
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
MODEL MINIMIZATION - AN ALTERNATIVE TO CIRCUMSCRIPTION 215
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
216 LANGUAGE, TRUTH AND LOGIC
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:
Alternatively, for first-order logic without identity, we must replace any fonnula or
subfonnula in the sub tableau of the fonn (3x)51x] by
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.
MODEL MINIMIZATION - AN ALTERNATIVE TO CIRCUMSCRIPTION 217
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
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:
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:
T2 (3x)Px
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)
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' /
MODEL MINIMIZATION - AN ALTERNATIVE TO CIRCUMSCRIPTION 219
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.
(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:
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.
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.
(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 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.)
224 LANGUAGE, TRUTH AND LOGIC
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).
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
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
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
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) .
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: .
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,].
LANGUAGE, TRUTH AND LOGIC 237
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.
LANGUAGE, TRUTH AND LOGIC 239
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-
242 NEW FOUNDATIONS FOR MATHEMATICAL THEORIES
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
LANGUAGE, TRUTH AND LOGIC 243
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
244 NEW FOUNDATIONS FOR MATHEMATICAL THEORIES
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,
LANGUAGE, TRUTH AND LOGIC 245
(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)
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]
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