Journal of Philosophy, Inc.

Are Mathematical Truths Synthetic a Priori?

Author(s): Jaakko Hintikka
Source: The Journal of Philosophy, Vol. 65, No. 20, Sixty-Fifth Annual Meeting of the
American Philosophical Association Eastern Division (Oct. 24, 1968), pp. 640-651
Published by: Journal of Philosophy, Inc.
Stable URL: http://www.jstor.org/stable/2024317
Accessed: 23-03-2015 15:59 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact [email protected].

Journal of Philosophy, Inc. is collaborating with JSTOR to digitize, preserve and extend access to The Journal of Philosophy.

http://www.jstor.org

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
640 THE JOURNAL OF PHILOSOPHY

caveats: as, for example, that D contain only projectable predicates,

that some restrictions be placed upon the cases when two events are
to be counted as being of the same type, etc. None of this need con-
cern us here, however, since our goal is only the modest one of dis-
playing the inference from the computational character of machines
to the tacit knowledge of the organisms they simulate as recogniz-
ably similar to orthodox patterns of scientific inference, whatever
the detailed analysis of such patterns of inference may turn out
to be.
If machines and organisms can produce behaviors of the same
type and if descriptions of machine computations in terms of the
rules, instructions, etc., that they employ are true descriptions of the
etiology of their output, then the principle that licenses inferences
from like effects to like causes must license us to infer that the tacit
knowledge of organisms is represented by the programs of the ma-
chines that simulate their behavior. Of course this inference is ana-
lytic only where we know that a given simulation is optimal, and
since claims for optimality are always empirical claims, we must run
an inductive risk when we infer from the operations of a well-evi-
denced simulation to the tacit knowledge of an organism. Since the
inference from like effects to like causes is an inductive inference,
all that a well-evidenced simulation gives us is inductive evidence
for believing that we have discovered something about the tacit
knoweldge of the organism we are studying. What more would it be
rational to want? What more would it be possible to have?
JERRY A. FODOR
Massachusetts Institute of Technology

ARE MATHEMATICAL TRUTHS SYNTHETIC A PRIORI? *

T HE outlines of the partial answer that I can offer to this

question have been argued for in my earlier papers.' In the
present one, I shall first summarize their relevant aspects.
The question was initially posed by Kant, and most existing discus-
* To be presented in an APA symposium on Mathematical Truth, December 29,
1968. Commentators will be Richard Montague and Charles D. Parsons.
1 See the following papers of mine: "Are Logical Truths Analytic?," Philosophi-
cal Review, Lxxiv, 2 (April 1965): 178-203; "Kant's 'New Method of Thought' and
His Theory of Mathematics," Ajatus, xxvii (1965): 37-47; "An Analysis of Analy-
ticity," in Deskription, Analytizitdt und Existenz, 3-4 Forschungsgesprdch des
internationalen Forschungszentrums Salzburg, ed. Paul Weingartner (Salzburg
and Munich: Verlag Anton Pustet, 1966), pp. 193-214; "Are Logical Truths Tau-
tologies?," ibid., pp. 215-233; "Kant Vindicated," ibid., pp. 234-253; "Kant and
the Tradition of Analysis," ibid., pp. 254-272; "Kant on the Mathematical
Method," The Monist, Li, 3 (July 1967): 352-375; "On Kant's Notion of Intuition

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
 ARE MATHEMATICALTRUTHS SYNTHETICA PRIORI? 641

sions of it refer in so many words to Kant. On the pain of gross his-

torical distortion, one cannot therefore help discussing the question
in Kantian terms. Now the examples of mathematical reasoning that
Kant mentions and discusses are typically formalizable in first-order
logic. Hence any historically accurate reading of the question turns
it into a problem concerning the status of logical as much as mathe-
matical truths. Again, by 'synthetic truths' Kant did not mean truths
that do not turn solely on the meanings of the terms they contain,
as a contemporary philosopher is likely to mean. I have argued that
the best explication we can offer of Kant's notion of an analytic
truth (in first-order logic) is what I have called a surface tautology.
Interpreted in this way, Kant's doctrine of the existence of synthetic
a priori truths in what he took to be mathematics turns out to be
correct in an almost trivial fashion, for there are easily any number
of valid (and provable) sentences of first-order logic that are not sur-
face tautologies.
Instead of offering yet another exposition of these points, I shall
in this paper comment on one particular aspect of the situation
which I have not elaborated elsewhere and which seems to possess a
great potential interest from a general philosophical point of view.
This aspect is the nature of the kind of information (surface infor-
mation) that goes together with the notion of surface tautology. An
examination of this concept seems to open wide philosophical per-
spectives which are highly relevant to the traditional discussion of
the possibility of synthetic a priori truths and to its background in
idealistic philosophy.
For the purpose, let me first recapitulate how the concept of sur-
face information can be defined and how it behaves.2 Let us assume
that we are dealing with a given fixed first-order language with a
finite number of predicates but (for simplicity) without individual
constants. Define the depth of a sentence s as the length of the long-
est chain of nested and connected quantifiers in s. (Two quantifiers
which contain the bindable variables x and y and of which the lat-
ter occurs within the scope of the former, are connected if there are

(Anschauung)," forthcoming in a collection of papers on Kant's Critique of Pure

Reason, eds. Terence Penelhum and J. H. MacIntosh, in the series Wadsworth
Studies in Philosophical Criticism (Belmont, Calif.: Wadsworth).
2 A background for the following discussion is given by the following papers
(in addition to those mentioned in the preceding footnote): "Distributive Normal
Forms in First-order Logic," in J. N. Crossley and M. A. E. Dummett, eds., Formal
Systems and Recursive Functions: Proceedings of the Eighth Logic Colloquium,
Oxford, July 1963 (Amsterdam: North-Holland, 1965), pp. 47-90; and "Distribu-
tive Normal Forms and Deductive Interpolation," Zeitschrift fur mathematische
Logik und Grundlagen der Mathematik, x (1964): 185-191.

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
642 THE JOURNAL OF PHILOSOPHY

quantifiers which occur within the scope of the former and which
contain the variables zl, Z2, - - . zk such that x and zl, z4 and zi1
,

(i = 1, 2, . .. k - 1), Zk, and y occur in the same atomic formula in

s.) Each closed sentence of depth d can be effectively transformed
into a disjunction of pairwise exclusive sentences of an especially
simple structure, called (closed) constituents of the same depth. Like-
wise, each open sentence of depth d (with certain individual vari-
ables zl, . . . , zk) can be transformed into a similar distributive nor-
mal form (i.e., into a disjunction of constituents with the same depth
and with the same variables). Some of these are effectively recog-
nized as being, in a specifiable sense, trivially inconsistent. Roughly
speaking, this means that two parts (not necessarily consecutive) of
such a constituent contradict each other propositionally. Since there
is no decision procedure for the whole of first-order logic, there must
exist nontrivially inconsistent constituents, too, which cannot be ef-
fectively recognized.
There exists, however, a systematic procedure for weeding out
more and more inconsistent but not trivially inconsistent constitu-
ents. All we have to do is to keep on adding to the depth of our con-
stituents (while preserving of course the same predicates and the
same individual variables). At each addition to the depth of a con-
stituent, it is usually split into a disjunction of a number of deeper
constituents, which are said to be subordinate to the given one.
These can be tested for trivial inconsistency, and it may happen that
they are all trivially inconsistent although the original, shallower
constituent was not. Moreover, it can be shown that at a sufficiently
great depth the inconsistency of each inconsistent constituent and,
hence, the inconsistency of each inconsistent sentence, is eventually
betrayed in this way. This result, which I have proved in detail else-
where, constitutes a completeness theorem for our method of enu-
merating inconsistencies.
If a probability-like measure (system of weights) is defined on the
set of all consistent sentences, the weight of each consistent constitu-
ent is split up between all its consistent subordinate constituents (of
a fixed depth). Such weights will be called measures of inductive
probability, and by their means one can define in the usual way mea-
sures of information, called depth information.3 One especially sim-
ple way is to put contdepth (s) = 1 - Pind (S) = Pind (~-s), where 'cont'
means "informative content" and 'Pid' "inductive probability."
8 For the different kinds of information,cf. my paper "The Varieties of Infor-
mation and ScientificExplanation,"in B. van Rootselaar, ed., Logic, Methodol-
ogy, and Philosophy of Science, Proceedings of the 1967 International Congress
in Amsterdam(Amsterdam:North-Holland, 1968),pp. 151-171.

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
 ARE MATHEMATICAL TRUTHS SYNTHETIC A PRIORI? 643

It is remarkable that all natural-looking principles of assigning

measures of inductive probability and depth information are non-
recursive. For instance, if the weight w of each constituent of depth
d is always divided evenly between all the consistent subordinate
constituents of depth d + 1, one can tell the weight of one of the
latter constituents (as a function of w) if and only if one knows the
number of these consistent subordinate constituents. However, be-
ing able to tell this number is easily seen to be tantamount to being
able to tell which of the constituents in question are inconsistent
and, hence, tantamount to having a decision procedure for first-
orderlogic.4 And this, of course,we cannot accomplishrecursively.
For this reason, measures of depth information cannot in general be
effectively calculated. Hence they cannot in any realistic sense be
operated with directly. They do not reflect faithfully the realities
one has to deal with in logic and in mathematics.
This becomes especially striking if one recalls the close relation-
ship between the concept of information and the idea of the elim-
ination of uncertainty. Part of the uncertainty we inevitably have
to face is the uncertainty as to which constituents (and other sen-
tences) are inconsistent and which of them are consistent. In order
to take into account the elimination of this kind of uncertainty in
our measures of information, we have to use other systems of weights
than those given to us by measures of inductive probability.
A much better candidate for a genuinely realistic measure of in-
formative content is obtained by assigning a nonzero weight to each
constituent that is not trivially inconsistent. It may be stipulated
that, whenever a constituent of depth d is split up into a disjunc-
tion of a number of constituents (none of which is trivially incon-
sistent) of depth d + 1, its weight is divided between these (each of
them receiving a nonzero weight) by some definite principle. When
the sum of weights is normalized to 1, we again obtain a probability-
like measure Psurf, which will be called surface probability or pre-
logical probability. This may be thought of as the degree of belief
it is rational to associate with a sentence before one has done to it
4 Assuming that we know the number of consistent constituents (and therefore
also the number of inconsistent ones) of a given depth, a simple method of
deciding which constituents of this depth are consistent is to list all valid sen-
tences in order until the right number of constituents are among their negations.
Since we know this number, we know that the remaining constituents are all con-
sistent. In the same way, one can see more generally that the decision problem of
the axiom system whose only nonlogical axiom is a constituent Co(d)(of depth d,
say) is of the same degree of unsolvability as the function (of e) that indicates
how many of the constituents subordinate to Co(d) and of depth d + e are incon-
sistent.

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
644 THE JOURNAL OF PHILOSOPHY

any of the many things that logic enables us to do in order to spell

out its implications more and more fully. (Hence the half-serious
term 'prelogical probability'.) In terms of such probability-like mea-
sures, we can in the usual way define measures of surface informa-
tion. One such measure is contsurf(s) = I - Psurf (s).
What has been said does not yet suffice to explain fully how these
measures are obtained, for so far we have not said anything about
what happens to the weight of a constituent CO4' of depth d which
is not trivially inconsistent but whose subordinate constituents of
depth d + 1 are all trivially inconsistent. It turns out unnatural to
let the weight of CO'('simply get lost. The natural procedure seems
to be to reassign the weight of CO'W' to its several "next of kin." By
this I mean the following: Constituents that are not trivially incon-
sistent form a tree. When we have a situation of the kind just dis-
cussed, a branch comes to an end with the constituent CO"'W in ques-
tion. When this happens, we trace the branch in question back till
we come to the most recent branching point-say, to the constituent
C0(c) (c < d)-from which at least two such branches emerge as
reach down to depth d + 1. The weight of COd) is then divided
among all the constituents Ci(o+1) subordinate to C0o")which have
at least one such branch going through them. This division follows
of course some specified principles. For instance, if we otherwise fol-
low even distribution of weights to subordinate constituents of the
next greater depth, we can follow even distribution here, too.
This suffices to define measures of surface probability and surface
information. By a surface tautology of depth d we can now mean
the disjunction of all the constituents of depth d (with the given
predicates and variables, of course) which are not trivially incon-
sistent, and, more generally, any sentence of depth d that has this
disjunction as its distributive normal form. So defined, surface tau-
tologies are precisely those sentences whose surface information is
zero. This preserves the important systematic connection between
the notion of informativeness and the notions of tautology and ana-
lyticity which was already relied on by Kant when he called ana-
lytic judgments merely "explicative" (Erlduterungsurteile).
An arbitrary surface tautology of depth d will be called in the
sequel t"d).
Among sentences s of depth d or less, the conditional probability
Psurf (slt(d)) (defined in the usual way) satisfies the customary Kol-
mogorov axioms of probability calculus (with the exception of
countable additivity). In view of the familiar betting-theoretical

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
 ARE MATHEMATICAL TRUTHS SYNTHETIC A PRIORI? 645

motivation of these axioms,5 our observations imply that these con-

ditional probabilities can be thought of as possible betting ratios for
a rational agent who thinks of all those events as possible which are
described by such constituents of depth d as are not trivially incon-
sistent. (Rationality may here be taken to mean simply an ability to
avoid Dutch Books made against oneself.) This throws some light
on the nature of our notions of surface probability and surface in-
formation. Notice, incidentally, that Psurf (sIt(d+e)) is simply the sur-
face probability of the distributive normal form of s at depth d + e.
Further light is thrown on our notions by an examination of what
happens to contsurf(SItd+e)) = 1-Psurf (sIt(d+e)) when e grows. The
direction of change depends on the particular sentence in question.
We do not always have contsurf (sI t(d+e)) < contsurf (sj t(d+e+1)). This
inequality nevertheless holds, even as a strict inequality, whenever
one of the constituents in the distributive normal form of s (at the
original depth d) becomes trivially inconsistent for the first time at
depth d + e + 1 while the same happens to no constituent of the
same depth that does not occur in the normal form of s. (This is
very much in keeping with the;idea of information as elimination
of uncertainty. The information s conveys to us-relative to t(,d+e)-
grows whenever we can omit one of the possibilities s seemed to al-
low to begin with and among which we had to be uncertain, while
no competing possibility is likewise omitted.) Moreover, from the
completeness theorem mentioned earlier it follows that the limit
lim8.,0 contsurf (SIt(d+e))is a measure of depth content. What is more,
the distribution principle that gives us this depth information is re-
lated in a simple way to the distribution principles used in assigning
the surface information in question. For instance, if Psurf(s) is based
on the principle of even distribution of weights where we move from
a given depth to the next greater depth, then so is the pind (s) which
we obtain as lim0... Psurf(sIt(d+e)). More generally, one can easily see
that ratios between weights are in a certain natural sense preserved
when one thus moves from Psurf (s) to the corresponding Pind (s). In
the sense that appears from these remarks, we can then say that
depth content of a sentence s is the limit to which surface content
converges when we gradually draw out from s all its more or less
hidden implications. In short, depth information is the limit of sur-
face information.
5 For the original works by Ramsey and De Finetti, see the handy anthology,
H. E. Kyburg and H. E. Smokler, eds., Studies in Subjective Probability (New
York:John Wiley, 1964).For a more recent treatment,see the papers by John G.
Kemeny,Abner Shimony,and R. ShermanLehman in Journal of SymbolicLogic,
xx (1955);3 (September):263-273; 1 (March): 1-28; and 3 (September):251-262,
respectively.

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
646 THE JOURNAL OF PHILOSOPHY

These observations perhaps help to illustrate the naturalness of

our measures of surface information as a measure of the kind of in-
formation we actually (and not just "implicitly") have. In view of
the obvious need of realistic measures of this kind, it is not surpris-
ing that some of the theoreticians of subjective probability have re-
cently expressed interest in (not explicitly specified) probability-like
measures which are supposed to resemble ours in not being in-
variant with respect to logical equivalence.6 Perhaps our measures
of "prelogical" probability are answers to their prayers.
There is nothing subjective about our measures, however. Rather,
they give us perfectly objective-and, I have suggested, fairly real-
istic-measures of information which show that there are senses of
information in which logical and mathematical reasoning yields
new information in a perfectly objective sense of 'information'. This
provides a definitive counterexample to the often-repeated neoposi-
tivistic thesis that the only sense in which logical or mathematical
reasoning gives us new information is purely psychological or sub-
jective, and thus also provides a partial answer to the Kantian ques-
tion to which this symposium is dedicated. Elsewhere, I have ex-
plored the relation of my answer to the traditional philosophy of
mathematics, especially to Kant.
All this leaves open a group of absolutely crucial questions, how-
ever, which have to be raised before we can hope to understand
fully the notion of surface information. Speaking of information
prompts the question: Information about what? How do the opera-
tions that increase our surface information enhance our grasp of
some subject matter or other? In the last analysis, one would like to
see how an increase in surface information enables us to deal more
efficiently with the reality-the "external" nonconceptual world
that we presumably are primarily interested in.
In order to answer these questions, let us assume that I receive an
item of information from a source that I know to be absolutely re-
liable, and let us assume that this news item is expressed in the
form of a first-order sentence s of depth d. What does that sentence
"really" tell me about the aspect of reality of which it ostensibly
speaks? This can in principle be spelled out by seeing which possi-
bilities concerning the world s admits and which possibilities it ex-
cludes, i.e., by transforming s into its distributive normal form of
depth d. (This point can be greatly strengthened. One might say, for
instance, that to specify what a sentence s "really" tells me about
6 See Leonard J. Savage, "Difficulties in the Theory of Personal Probability,"
Philosophy of Science, xxxiv, 4 (December 1967): 305-310, and Ian Hacking,
"Slightly More Realistic Personal Probability," ibid.: 311-325.

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
 ARE MATHEMATICAL TRUTHS SYNTHETIC A PRIORI? 647

reality is to specifywhat kinds of individualsI can expect to find in

the world when s is true and, after I have found one of them, how
the further kinds of individuals are related to it which I may come
across,etc. But what the distributivenormal form of s does is just to
answerall such questions as fully as one can do without going be-
yond the depth of s, i.e., going beyond the resourcesof expression
alreadyused in s.) However, this normal form does not yet give me
everythingI can extractfrom s without the benefit of furtherfactual
discoveriesand without the benefit of further messageswith a fac-
tual content. The reason for this is that some of the possibilities
that s seems to admit at depth d may turn out to be only apparent.
I have tried to illustratethis state of affairsby comparingconstitu-
ents not to pictures of reality, but to recipes for constructingsuch
pictures.7This analogy can in fact be profitablydeveloped further
in severaldirections.The main point relevant here is that some of
these recipes may misfire in the sense that no picture can be con-
structed by their means. From the recipe alone one cannot see
whetherthis is the case or not. Barringfurtherinformation (surface
informationl),I thereforehave to be preparedfor the truth of any
constituentin the normal form of s that is not trivially inconsistent.
The reality and even urgency of this need merits some emphasis.
For instance,in the normalform of s theremay be some constituents
assertingthe existenceof certainkinds of individuals (or n-tuplesof
individuals).Even though all these constituentsare in the last analy-
sis inconsistent,I cannot discount them as long as I have not ac-
tually ascertainedtheir inconsistency.Before I have done that, I
may even find myself making practical preparations for actually
running into the kinds of individuals my (inconsistent)constituents
assertto exist. (A rich stock of detailed illustrationsof the phenom-
enon involved here, drawn from a slightly different departmentof
first-orderlogic, are offeredby the domino problemsof Hao Wang
and his collaborators.8In the same way as one cannot see from a
constituentwhether it yields a consistent "picture"of the world or
not, in the same way one cannot see from one of these domino prob-
lems whether it admits of a positive solution or not. For my pur-
poses, the domino problemsare illustrativein two differentrespects.
First, a completed domino problem- i.e., an euclidean plane filled
with domino chips of specifiablekinds-is closely reminiscentof a
"pictureof the world," especially of a completed jigsaw-puzzlepic-
ture. Second, the domino problems show how very far one some-
7 See my "Are Logical Truths Analytic?" (fi. 1).
8 See Hao Wang's paper, "Remarks on Machines, Sets, and the Decision Prob-
lem," in Crossley and Dummett, op. cit., pp. 304-320.

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
648 THE JOURNAL OF PHILOSOPHY

times has to carry out an attempted "picture construction" before

one can see its impossibility, thus strikingly illustrating the kind of
combinatorial information needed to rule out some merely appar-
ent alternatives.)
The unavoidability of this predicament follows from the un-
decidability of first-order logic, which is thus seen to contain impor-
tant morals for our concept of information.
Our description of the actual situation one faces when one re-
ceives a piece of information expressed in first-order language also
shows what the practical (pragmatic) advantages are that accrue
from an increase in surface information. It was pointed out above
that, as far as s and its successively deeper and deeper normal forms
are concerned, we obtain more surface information essentially in-
sofar as we can eliminate some of the constituents in its normal
form at the initial depth d as being inconsistent. The more such in-
consistent constituents we can eliminate, the more narrowly we can
restrict the range of eventualities we have to be prepared for when
we know that s is the case. It is true that no new empirical observa-
tions and no further messages with a factual import are needed for
the elimination. The fact nevertheless remains that this elimination
is not automatic and that it is not a merely subjective process of get-
ting rid of mental blocks that cloud our vision of what there already
is in front of us. The elimination involves further work whose ex-
tent can be objectively measured.
Thus we can see a good reason for saying that the uncertainty we
get rid of when our surface information grows is uncertainty con-
cerning the reality our sentences speak of. The same is presumably
true of the concept of information involved here. (This could also
be construed as a reason for saying that nontrivial logical and math-
ematical knowledge is "synthetic" in a rather striking sense of the
word.) At the same time, it is clear that the insights we obtain when
we gain new surface information are in some important sense con-
ceptual, almost linguistic. They pertain to the ways in which our
language can or cannot represent reality. They are not comparable
to having a look at a picture so as to see what that part of the world
is like which the picture represents; they are, rather, like realizing
that certain structures that might at first seem to be pictures do not
really stand for anything. The information gained thus seems to be
purely conceptual.
This double nature of surface information might seem very puz-
zling, almost paradoxical. Rightly viewed, it nevertheless points to
a feature of the conceptual situation we are studying which seems

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
 ARE MATHEMATICAL TRUTHS SYNTHETIC A PRIORI? 649

to me highly interesting. The paradoxical double nature of surface

information illustrates the important fact that whatever we try to
say about reality is inextricably interwoven with the contributions
of our own conceptual system. What we can directly and imme-
diately express or grasp is a complex outcome which embodies both
elements due to nonconceptual reality and elements due to the con-
ceptual system we are relying on. The better we know the concep-
tual system, the more fully we know what is contributed to this
complex outcome by our conceptual system. By discounting this we
can therefore ipso facto know to the same extent what is contributed
to it by the reality we in the first place want to describe (or want to
have described to us). For this reason, the deeper understanding of
our own conceptual system-in the case at hand, the wider mas-
tery of first-order logic-which an increase in surface information
amounts to can at the same time mean an elimination of uncer-
tainty concerning the objective state of affairs "out there" in reality.
The metaphor that inevitably suggests itself here is the following:
We do not and cannot "touch" the reality directly, but only by
means of a conceptual system. This system works like a highly com-
plex instrument that connects our knowledge with the reality this
knowledge is about. This instrument is so intricate that we do not
know which of its registrations are due to the influence of the reality
we are interested in and which of them merely reflect the mode of
functioning of the instrument itself. The better we know the instru-
ment, the more of the merely apparent registrations we can disre-
gard. This also means that we can use the instrument more effi-
ciently than before for the purpose of coming to know the reality
its feelers touch.
Thus we can see how surface information may legitimately be
thought of both as conceptual information and at the same time as
information concerning objective reality. Strictly speaking, this is of
course true only of that increase in surface information which takes
place when a sentence is expanded into a disjunction of increasingly
deeper constituents.
The most interesting feature of the situation is the inextricability
of conceptual elements from those contributed by mind-indepen-
dent reality. One reason why this is interesting is that it is closely
related to a well-known thesis of certain traditional philosophers
who are sometimes inaccurately referred to as idealists. These phi-
losophers have claimed that all our knowledge-or at least all our
"better" knowledge, which for Kant meant synthetic knowledge a
priori-presupposes, and depends on, concepts that are of the na-

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
 650 THE JOURNAL OF PHILOSOPHY

ture of the mind's own creations. As a corollary to this dependence

of all our knowledge on our own concepts, it is maintained that
"things in themselves," as they would be independent of all our
concepts and therefore of our own activity, are unknowable and in-
describable. At best, one school of thought avers, the concept of a
Ding an sich is useful as an idealized limit that our knowledge can
approximate but never fully reach. In brief, reality is for these phi-
losophers inseparable from our concepts-and vice versa.
The simple but representative case of first-order languages offers
us a handy testing ground of these "idealistic" theses. In view of
what has been said, these theses can in fact be understood so as to
be in certain respects essentially correct. In terms of the metaphor,
it is the case that (in general) we cannot effectively decide which
registrations of the apparatus to which we are comparing our con-
ceptual system are informative about "objective" reality and which
registrations are, in contrast, merely due to the mode of operation
of our registering apparatus. In a sense, whatever we may try to say
of Dinge an sich in first-order terms is normally shot through with
elements (apparent possibilities) which are entirely due to the spe-
cific way in which first-ordersentences are related to the reality they
strive to mirror. In this sense, reality and our concepts are inextrica-
bly interwoven with each other in all nontrivial use of first-order
discourse. Furthermore, the central role of this discourse strongly
suggests that my point can be generalized.
The inextricability of conceptual from objective factors in first-
order languages is due to the undecidability of first-order logic.
Looked upon from the point of view of the present paper, Church's
undecidability result thus turns out to have decidedly idealistic im-
plications. Our modem version of a time-honored idealistic thesis
undoubtedly appears undramatic as compared with the imagery
that is usually associated with the original version or versions. For
instance, the "things in themselves" that our version may perhaps
be claimed to involve do not constitute a special class of unknow-
able but nevertheless in some strange way causally active entities.
Speaking of one's knowledge of "things in themselves" means for
us merely a fagon de parler-a counterfactual way of speaking-of
one's knowledge such as it would be if all elements contributed to it
by our conceptual apparatus were eliminated. In short, it means for
us speaking of depth information instead of surface information.
There is nothing illegitimate about doing so, as long as we realize
that there is no way of actually (i.e., effectively) dealing with the
nonrecursive concept of depth information.

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions
 FORMALIZATION IN SCIENCE 65I

In spite of this undramatic appearance of our quasi-idealistic con-

clusions, it seems to me that they are in a deep sense connected with
what is true and important in the original thesis. Among other
things, I believe that thinking of Dinge an sich as a separate, un-
knowable class of entities has always been a case of fallacious hy-
postatization. If this connection really exists, it is not surprising
that as a by-product we can also partially vindicate the old Kantian
doctrine of the synthetic a priori character of nontrivial mathemati-
cal (and, for us, logical) truths.
JAAKKO HINTIKKA
University of Helsinki and Stanford University

THE DESIRABILITY OF FORMALIZATION IN SCIENCE *

T would be amusing to put the question whether formalization

in science is desirable to Archimedes in Sicily or some three
hundred years later to Ptolemy in Alexandria. I can imagine
Archimedes, in a characteristic turn of phrase, saying that no man
of eminence in philosophy would ask such a question. It would be
like asking if one wanted an actual demonstration in geometry as
opposed to a suggestive but informal method. I think even Ptolemy,
though much more deeply involved in empirical observations and
the complicated problem of fitting theory to data, would answer in
the same vein. In other words, I want to claim that the only branches
of quantitative science seriously developed in ancient times were re-
garded as extensions of geometry and treated with the same degree
of formality. I am thinking particularly of Archimedes' work in stat-
ics; his treatise on the equilibrium of planes is the first systematic
treatise in mathematical physics. Also I am thinking of Ptolemy's
Almagest, which is unquestionably the greatest scientific work of
antiquity, if we mean by science, as opposed to mathematics, the de-
velopment of theory and the confronting of theory with quantitative
data. For this ancient tradition of scientific analysis, there was no
proper way to think other than in terms of the formal methods of
geometry. This pattern that originates with Archimedes has a long
and continuous history through the Arabic and Latin tradition of
the Middle Ages, leading without a break down to Newton's Prin-
cipia in the seventeenth century. Newton's formal geometrical meth-
ods and his careful consideration of data are very much in the spirit
of Ptolemy's Almagest. For Newton too, there is little question of
* To be presented in an "A symposium on Formalization in Science, December
29, 1968. Commentators will be Carl G. Hempel and Henry E. Kyburg, Jr.

This content downloaded from 128.235.251.160 on Mon, 23 Mar 2015 15:59:51 UTC
All use subject to JSTOR Terms and Conditions