Epistemology, Methodology, and Philosophy of Science

EPISTEMOLOGY,
METHODOLOGY, AND
PHILOSOPHY OF SCIENCE
Essays in Honour of Carl G. Hempel

on the Occasion of His 80th Birthday,
January 8th 1985
Edited by
W. K. ESSLER
University of Frankfurt, Frankfurt am Main
H. PUTNAM
Harvard University, Cambridge, Mass.
and
W. STEGMULLER
University of Munich
Reprinted from
Erkenntnis, Vol. 22, Nos. 1,2 and 3
SPRlNGER-SClENCE+BUSINESS MEDIA, B.V.

ISBN 978-90-481-84\0-1 ISBN 978-94-017-1456-3 (eBook)
DOI 10.1007/978-94-017-1456-3
Ali Rights Reserved

© 1985 by Springer Science+Business Media Dordrecht
Originally published by D. Reîdel Publishing Company, Dordrecht, Holland in 1985
Softcover reprint of the hardcover 1st edition 1985
No part of the material protected by thîs copyright notice may be reproduced or
utilized in any form or by any means, electronic or mechanical,
including photocopying, recording()f b"y any information storage and
retrieval system, without written permission from the copyright owner
ERKENNTNIS / Volume 22 Nos. 1,2, and 3 January 1985
EPISTEMOLOGY, METHODOLOGY, AND

PHILOSOPHY OF SCIENCE
Essays in Honour of Carl G. Hempel on the Occasion

of His 80th Birthday, January 8th 1985
Edited by
w. K. ESSLER, H. PUTNAM, and W. STEGMULLER
Preface 1
HAIM GAIFMAN / On Inductive Support and Some Recent Tricks 5
CLARK GL YMOUR / Inductive Inference in the Limit 23
D. COST ANTINI / Probability and Laws 33
EV AN ORO AGAZZI / Commensurability, Incommensurability and
Cumulativity in Scientific Knowledge 51
EIKE VON SAVIGNY / Social Habits and Enlightened Cooperation:
Do Humans Measure Up to Lewis Conventions? 79
C. U LlSES M OU LIN ES / Theoretical Terms and Bridge Principles:
A Critique of Hempel's (Self-) Criticisms 97
ANDREAS KAMLAH / On Reduction of Theories 119
HI L A R Y PU TN A M / Reflexive Reflections 143
ANNETTE BAIER / Explaining the Actions of the Explainers 155
NORETTA KOERTGE / On Explaining Beliefs 175
PATRICK SUPPES / Explaining the Unpredictable 187
RICHARD N. BOYD / The Logician's Dilemma: Deductive Logic,
Inductive Inference and Logical Empiricism 197
B. KANITSCHEIDER / Explanation in Physical Cosmology: Essay
in Honor of C. G. Hempel's Eightieth Birthday 253
NELSON GOODMAN / Statements and Pictures 265
RAIMO TUOMELA / Truth and Best Explanation 271
RAINER W. TRAPP / Utility Theory and Preference Logic 301
RUDOLF HALLER / Der erste Wiener Kreis 341
NICHOLAS RESCHER / Are Synoptic Questions lllegitimate? 359
W. K. ESSLER / On Determining Dispositions 365
ULRICH BLAU / Die Logik der Unbestimmtheiten und Paradoxien 369
HANS LENK / Bemerkungen zur pragmatisch-epistemischen Wende
in der Wissenschaftstheoretischen Analyse der Ereigniserkliirungen 461
MICHAEL KUTTNER / Zur Verteidigung einiger Hempelscher Thesen
gegen Kritiken Stegmiillers 475
PREFACE
Professor C.G. Hempel (known to a host of admirers and friends as 'Peter'

Hempel) is one of the most esteemed and best loved philosophers in the
world. If an Empiricist Saint were not somewhat of a Meinongian Impos-
sible Object, one might describe Peter Hempel as an Empiricist Saint. In-
deed, he is as admired for his brilliance, intellectual flexibility, and crea-
tivity as he is for his warmth, kindness, and integrity, and does not the
presence of so many wonderful qualities in one human being assume the
dimensions of an impossibility? But Peter Hempel is not only possible but
actual!
One of us (Hilary Putnam) remembers vividly the occasion on which he
first witnessed Hempel 'in action'. It was 1950, and Quine had begun to
attack the analytic/synthetic distinction (a distinction which Carnap and
Reichenbach had made a cornerstone, if not the keystone, of Logical Em-
piricist philosophy). Hempel, who is as quick to accept any idea that seems
to contain real substance and insight as he is to demolish ideas that are
empty or confused, was one of the first leading philosophers outside of
Quine's immediate circle to join Quine in his attack. Hempel had come to
Los Angeles (where Reichenbach taught) on a visit, and a small group
consisting of Reichenbach and a few of his graduate students were gath-
ered together in Reichenbach's home to hear Hempel defend the new posi-
tion. The scene was unforgettable! Putnam still remembers the way in
which many of those present tried to argue against Hempel's brilliant and
forceful attack on the distinction, and the clarity and eloquence with which
Hempel maintained his view. One student asked whether the distinction
could not be drawn in a formalized language, even if it is unclear or in-
applicable in the case of natural language? 'The disease is a hereditary
one', was Hempel's reply. 'Every formalized language is ultimately inter-
preted in natural language. If there is no analytic/synthetic distinction in
some natural language, then there is no such distinction in the case of
languages which we use to formalize natural language assertions either.'
But Hempel's fame does not rest on his exposition of the ideas of others
Erkenntnis 22 (1985) 1.
2 PREFACE
(although Peter Hempel has always remembered and lived up to the great
ideal of the Vienna Circle, that analytic philosophy is cooperative work
and not a competition between a number of famous 'Herr Professors').
The best known of all accounts of scientific explanation is the 'Hem-
pelian account', referred to in countless books and journal articles. For
decades it has been defended, criticized, extended, emended or rejected by
authors in the social sciences, history, physical and biological science as
well as by phiIosophers of science. Moreover;"many of the most important
extensions (e.g. to the logic of statistical explanaticm and to the logic of
functional explanation) have come from Hempel himself. Almost as widely
quoted are Hempel's writings on confi~ation - especially his celebrated
'Paradox of the Raven'. (According to traditional accoUnts, every obser-
vation of an A which has the property B 'confirms' the statement that all
As are Bs, at least as long as nothing has been observed which is an A and
not a B. Now, let A l:>e the property of being Non-Black and let B be the
property of being a N on-Raven. Then, if the traditional doctrine is correct,
every observation of a Non-Black thing which is not a Raven, e.g., the
observation of a green automobile, confirms All Non-Black things are
Non-Ravens. But it is also a principle of inductive logic that if data confirm
a hypothesis S tQen they equally confirm any logically equivalent Qypoth-
esis Sf. Since All Non-Black things are Non-Ravens is logically equivalent
to All Ravens are Black, it follows that the observation of a green auto-
mobile confirms All Ravens are Black, which is violently counterintuitive!)
Like the Russell Paradox inset theory, Hempel's Paradox shows that what
we naively want to maintain cannot be maintained; and it turns out to be
extraordinarily difficult to discover just what principles of inductive logic
we should now maintain.
The fact is that Peter Hempel's achievements cover the entire range of
topics in the methodology of science, including th~ related areas of phi-
losophy of language and theory of knowledge. He has published at every
level from the specialist monograph to the introductory textbook, and he
has more than once produced a'work which represented the 'state of pres-
ent knowledge' in its area (until it was superseded by another work from
Hempel's pen). One thinks, for example, of the way in which Hempel's
monograph on Concept Formation in Empirical Science stood as the great
summary of the Logical Empiricist investigations until Hempel himself
gave up the hope for an 'empiricist criterion of significance' in his magis-
PREFACE 3
terial 'Problems and Changes in the Empiricist Criterion of Significance'.

In recent years Hempel has shown an interest in and sympathy with more
'historical' and less 'logistic' approaches to the philosophy of science; but
this does not mean that he does not detect and criticize weak arguments
or exaggerated claims when they are put forward by proponents of such
approaches. Above all, he has been willing to show us where we are up
against difficulties that none of the present approaches seems likely ro re-
solve. In a true sense he seems to all of us to be our teacher, whether or
not we were ever his students in a formal sense. There is only one Peter
Hempel; long may he flourish.
With love and admiration,
University of Frankfurt, Frankfurt am Main w. K. ESSLER

Harvard University, Cambridge, Massachusetts H. PUTNAM
University of Munich W. STEGMULLER
c. G. HEMPEL
HAIM GAIFMAN
ON INDUCTIVE SUPPORT AND SOME RECENT TRICKS*
This paper is about some aspects of induction. The tricks come only at the
end, where the Popper-Miller proof of the impossibility of inductive prob-
ability is analysed. In the first section inductive inference is placed in the
broader perspective of pattern recognition and reasoning by analogy.
Newton's derivation of his law of gravity furnishes a good illustration.
The second section is concerned with the role of subjective probability. In
particular, we show what can happen in a methodology of science which
denies this role in scientific research. Another short observation concerns
the concept of empirical content; finally, some points are made concerning
the probabilities of universal generalizations which cover a potential in-
finity of cases. The third section focuses on simple inductive generaliz-
ations. The fourth section is concerned with the problem of expressing
inductive support by quantitative probabilities.
1. EXPECTATIONS OF REGULARITY
Inductive inference is the inference from particular cases to a general law,

or to other unobserved cases of a similar nature. Recognized already by
Aristotle, it became a central philosophical issue in the 17th century with
the rise of modern science and the empiricist point of view. Induction has
been regarded as a basic method of acquiring general knowledge in the
empirical domain; knowledge which, in principle, does not attain full cer-
tainty but which can be reconfirmed by repeated evidence to the point of
becoming almost certain, or certain for all practical purposes. Repeated
patterns give rise to and reinforce conscious expectations just as, on a
more primitive level, they create and reinforce unconscious habits. There
has been abundant research into the learning mechanisms of animals, from
worms to humans 1 • The essential role of repeating the same pattern in
order to have it absorbed or "learnt" is clear beyond doubt.
One difficulty in realizing the extent of induction is its very elementary
nature and its being related to non-conscious habit forming and semi-
Erkenntnis 22 (1985) 5-21 0165-{)106/85/0220-0005 $01.70

© 1985 by D. Reidel Publishing Company
6 HAIM GAIFMAN
conscious expectations. It took a Humean analysis to trace back our basic

everyday convictions and habits of thought to pattern repetitions in the
surrounding world. Examples of broken regularities can illustrate this on
a more modest scale: "What happened, why doesn't Mr. Jones carry his
umbrella today?"; as a matter of fact there was no reason that he should
carry an umbrella on such a fine morning, but having observed him to
carry it on all previous occasions, rain or no rain, we have come to regard
Mr. Jones + umbrella as part of the natural order. Exceptional behaviour
is sometimes explained merely by pointing to a repetitive pattern: "Why
does she wear such a strange hat?" "Oh she always wears something like
this" .
Inductive reasoning becomes an explicit tool when one tries to discover,
or confirm, new basic regularities in the experimental data. In Book III of
his Principia Newton reports as follows:
I tried experiments with gold, silver, lead, glass, sand, common salt, wood, water, and wheat.
I provided two wooden boxes, round and equal. I filled the one with wood and suspended
an equal weight of gold (as exactly as I could) in the center of oscillation of the other. The
boxes hanging by equal threads of 11 feet made a couple of pendulums perfectly equal in
weight and figure, and equally receiving the resistance of the air. And, placing the one by the
other, I observed them to play together forwards and backwards, for a10ng time with equal
vibrations.
By straightforward induction we are lead to conclude that the oscillations

of all such pendulums depend neither on the kind of matter in their weights
nor on its quantity. Newton's induction is however much more sophisti-
cated. He interprets the results in terms of the theoretical framework al-
ready laid out in the previous books. Equal oscillations imply equal ac-
celerations towards the earth. And by his second law of motion, / = ma,
the equality of accelerations, a1 = a2, implies that the forces are propor-
tional to the masses, /1//2 = mdm2' Thus he concludes that every body
on earth is subject to a force towards the earth's center proportional to its
mass. This provides a starting point for a still more comprehensive induc-
tive reasoning. First he suggests a thought experiment in which the bodies
are removed to the vicinity of the moon. The validity of the general law
implies that also at that distance free bodies, including the moon, should
fall towards the earth with equal accelerations.
Then he takes the crucial step of replacing the earth and its falling bodies
by Jupiter and its revolving moons. By decomposing a revolving motion
ON INDUCTIVE SUPPORT 7
into its tangential and radial components the body can be regarded as
gravitating towards the center of revolutions. The observational fact that
the periods of Jupiter's moons are related as the 3/2 powers of their dis-
tances implies mathematically that their accelerations towards Jupiter are
inversely proportional to their squared distances, a = c/d 2 • Hence, as far as
these moons are concerned, each is subject to a force towards Jupiter pro-
portional to its mass. The same argument is then applied to the sun and
its planets. Furthermore it is argued that with Jupiter also its moons are
subject to forces towards the sun proportional to their masses; otherwise
they would not revolve in stable orbits around Jupiter. This also applies
to Saturn and its satellites and to the earth and its moon. Having placed
a:11 these side by side, inductive reasoning points to the law that bodies in
general are subject to forces of gravitation towards each planet and to-
wards the sun, forces proportional to their masses. Combining this with
his third law of action and reaction, Newton concludes that the bodies
must likewise attract the planets. Finally, regarding the planets themselves
as composed of parts which are bodies, Newton arrives at his concluding
generalization that all bodies attract each other according to this law.
Going over the Principia's reasoning one realizes the true meaning of
Newton's famous and often misunderstood statement hypotheses nonfingo
(I frame no hypotheses), made in connection with the law of gravity. New-
ton does not claim that his law follows from the evidence by pure deduc-
tion, but that it is the product of sound reasoning based on uncovering
essential patterns of the experimental data. He distinguishes it from "hy-
potheses making" which, in this context, means reckless speculation un-
guided by the evidence. At Newton's time, force was conceived as some-
thing transferred by direct contact between bodies. The abstract notion of
a force acting at a distance (i.e., without intervening matter) went against
this basic intuition. Newton decreed that basic intuition should give way
to the new concept needed for the general law, a law which expresses an
essential regularity of the data, as it is interpreted in terms of the theoret-
ical model. The new concept is, so to speak, distilled from this pattern.
Newton, who was well aware of the problem surrounding gravitation,
tried, without success, to explain it by reduction to familiar notions, in a
way which keeps touch with observed phenomena. ("But hitherto I have
not been able to discover the cause of those properties of gravity from
phenomena"). He decided that the concept can stand on its own feet.
8 HAIM GAIFMAN
Pattern recognition, reasoning by analogy and inductive inference be-

long together. Simple inductions are elementary forms of pattern recog-
nition; on the other hand, reasoning based on sophisticated patterns and
abstract analogies can be viewed as inductive generalizations applied to
complex predicates of higher orders. Such analogies are not likely to com-
mand general agreement. What to one is a natural generalization from the
data may appear to another as a surprising hypothesis. The so-called "in-
ductivist" and "hypothetico-deductivist" methodologies are branches of
the same tree, not the different growths that some make them out to be.
Neither are these methodologies restricted to the empirical domain.
There are abundant examples of mathematical hypotheses that have been
suggested mostly because certain laws were observed to hold in all known
cases; for example Goldbach's conjecture that every even number greater
than 2 is a sum of two primes. (It is only to be expected that some should
be later refuted by counterexamples, such as the hypothesis of Fermat and
Mersenne that all numbers of the form 22 " + 1 are prime.) The four colour
conjecture is an example of a successful hypothesis of this kind which took
more than 130 years to prove; noteworthy is also the fact that the proof
involves the checking of a very large number of particular cases, by com-
puter. Polya points out 2 highly instructive examples of induction and
reasoning by analogy in mathematics. Thus, by treating sin x/x as a po-
lynomial whose roots are ± n, ± 2n, ... Euler arrived at the formula
sin x/x = n00 (

2
1 -x-) .
n2 n 2
":1
What distinguishes mathematics from empirical science is the availabil-

ity of a deductive apparatus which serves as a chief and final method of
establishing truth. But common to all science is a search for regular pat-
terns based on a belief that there is something to be searched, that is to
say - that, viewed correctly, the subject matter exhibits lawlike behaviour.
Some regularity is needed for the very definition of our "raw data" (this
is the valid part of Kant's insight). But even without going so far we can
see that while we can imagine different kinds and degrees of lawlikeness,
it is difficult to picture complete chaos. For example "nature never repeats
itself' is either selfcontradicting through self-reference, or a statement
which by denying order at a given level sets it up on a higher level (It IS
always true that patterns of a given kind do not repeat). Or consider the
randomness of a data sequence obtained by tossing a fair coin. It signifies

total disorder as far as deterministic patterns are considered, but satisfies
extremely binding statistical laws.
2. PLAUSIBILITY
An essential aspect of induction is that it provides the hypothesis in ques-

tion with a certain amount of plausibility and that our confidence grows
with additional confirming tests. It is this aspect which came mostly under
attack by Popper and Lakatos. Yet any adequate philosophy of science
must take into account the crucial role played by plausibility, or subjective
probability, qualitative and vague as it might be. A good illustration of
the point is provided by the very attempts to expel subjective probability
from the philosophy of science. Popper's methodology recommends the
choosing of the boldest most informative hypothesis, or more generally -
theory, for the purpose of testing; the theory's rating is eventually deter-
mined by the severe tests which it has survived. This prescription which
embodies some basic valid insights leaves us with infinitely many good
candidates for testing, as we can easily tum out extremely powerful un-
refuted and testable theories, one after the other. Consider for example
the following:
There is a so-called "mental force" which, by combining with magnetic
fields, produces changes in the field of gravity. The present lack of con-
firming evidence (except, perhaps, for some controversial reports of levi-
tation) is explained by the weakness of the prevailing mental field. In fact,
the current accepted theory is a limiting case obtained by letting the mental
field tend to zero (it is very easy to produce equations which have this
effect). At the same time our theory is highly testable. For it also predicts
that quite noticeable changes of gravity can be achieved by placing twelve
meditating Buddhist monks in a regular dodecagon, within a powerful
magnetic field to be produced by a large electromagnet. The scientific rev-
olution that such a successful experiment will unleash is out of all pro-
portion with the mere 200,000 dollars which the national research agency
is required to put up (the cost of traveling, special accommodations and
health insurance for the twelve monks, use of one of the largest existing
electromagnets and other expenses). Should the experiment refute the pre-
diction we shall abandon this falsified theory, being the honest empiricists
10 HAIM GAIFMAN
that we are, without attempting any of the ad-hoc saving stratagems that
Popper condemns. Tomorrow we shall make up another, totally different,
breathtaking hypothesis.
Once the crucial element of subjective probability is ignored such a pro-
cedure becomes legitimate. It is because scientific research is conducted
along lines deemed to have an appreciable chance of success that experi-
ments of the meditating monks variety do not arise. To be sure, our
science-fiction theory reveals itself as an arbitrary concoction. But then
why should arbitrary concoctions be rejected if not for our estimate that
they lack the minimal required chance of being true, or of having valuable
true consequences? Our attitude would have been different had the theory
and experiment been proposed by some crank, McX, who had an impres-
sive list of unexplained and astounding successes behind him. For in that
case, instead of considering the plausibility of the theory as such, we will
consider the probability of a much more primitive conjecture: "All the
predictions of McX are true".
The wide subject of subjective probability in the philosophy of science
is beyond this paper's scope. Let me only make two short observations.
First, the decision to test some hypothesis, h, is determined by the interplay
of three factors: (i) The prior probability of success, or, more generally,
the probabilities of the various outcomes. (ii) The new information, or
empirical content, that is to be gained. (iii) The cost in terms of time, effort
and resources. Had we had quantitative measures of these factors we could
have treated the problem as one of maximizing expected utility.
The informativeness of an hypothesis is indicated by comparing the ini-
tial distribution, p( ), over some field of relevant statements (or events),
with the conditional probability distribution, given h, p( Ih). As a rule, the
change from p( ) to p( Ih) cannot be too great if h's initial probability,
p(h), is very near 1. But it does not follow that a very small p(h) implies
a high empirical content, as Popper (and at a certain time Carnap) sug-
gested. The conjecture of getting straight heads in the next 15 coin tosses
has extremely small probability; but, as long as it does not affect the prob-
ability distribution over those crucial statements that are relevant in the
given context, its empirical content is rather low. Were this event to hap-
pen, I shall merely conclude that my coin is biased. The empirical content
is therefore determined by h's implications in the presupposed context. If
it is to be measured by a number than this number cannot be a function
of p(h), but some measure reflecting the change from p( ) to p( \h) over
some field of events. 3
The second observation concerns the assignment of probabilities to uni-
versal empirical hypotheses, h, ofthe form "for all x, P (x) " , which involve
a (potentially) infinite number of cases. There is an old argument (sub-
scribed to, at a certain time, even by de-Finetti) purporting to show that,
because of the unbounded number of cases, the probabilities should be
zero. Anyone betting on h, so the argument runs, will lose whenever a
counterexample turns up; but he shall have to wait until all the infinitely
many cases are verified, one by one, before he can collect. Hence the bet
is irrational at any positive odds. Indeed, if these be the conditions of
winning, the bet is irrational, but then the betting odds have little to do
with the probability of h. When betting on a proposition one actually bets
on its being proved against its being refuted; for the payoff takes place
only after a proof of it or of its negation has been obtained. In standard
betting situations the gap between truth and provability is ignored, because
provability poses no problems: we are sure to find out the outcome of a
race, of an election, or of a coin toss. But occasionally the method of
deciding the bet is crucial and it may happen that the ratio of p(h) to
p(-,h) is altogether different from the ratio of p('h' will be proved) to
p('-,h' will be proved); it is this last ratio which detennines the rational
odds.
Suppose, for example, that expert A believes that the crash of a new test
plane was caused by an explosion of overheated inflamable fluid in a par-
ticular part of the plane. Expert B thinks differently and proposes a bet.
The only way of deciding the question is by playing the blackbox tape on
which all the flight data has been automatically recorded. But the blackbox
was located in the vicinity of the very part in which, according to A, the
explosion took place. Hence, if A is wrong, the blackbox, when found
among the debris, will show it. But if A is right then, in all probability,
the blackbox has been destroyed and the bet will remain unsettled. (The
destruction of the blackbox is not sufficient for proving A's claim, since
it could have been caused by other factors). It is therefore irrational for
A to accept odds which reflect his subjective probabilities. This example
has nothing to do with having to verify infinitely many instances, but the
root of the dilemma is the same: the gap between truth and provability.
In the case of a universal hypothesis we are better off than in the last
12 HAIM GAIFMAN
example, for we can still use bets to give direct concrete meanings to the
probabilities; instead of a single bet we have to use a system: Let P*(n) be
the conjunction P(1) 1\ ••• 1\ P(n), let PI be A's probability of P(1) and
Pn+rhis conditional probability of P*(n + 1), given P*(n). Consider the
system in which the first bet is on P(1) with odds PI:l-PI and the (n +
l)'th bet is on P*(n + 1), with odds Pn+I:1-Pn+h conditional on P*(n)
(i.e., the bet is called off if P*(n) in false). A should accept the system and
be indifferent to choices of sides in the separate bets. His probability as-
signment to the universal hypothesis is the limit, say P, of the products
PI ..... Pn as n ~ 00. If P i= 0 then, assuming the stakes in the separate
bets to be bounded, the possible total win or loss in the infinite system is
bounded as well. (If in the n'th bet the stakes are S . (1 - Pn) and S . Pn
then the total win or loss is ::; S· L (1 - Pn) + Sand P i= 0 implies

n=1
L (1 - Pn) < 00).
n=1
3. SIMPLE INDUCTION
The significance of simple inductive generalizations, such as "all ravens

are black" can be fully appreciated only in the wider context of reasoning
by analogy, pattern recognition and higher order inductions. Let me how-
ever consider simple, or basic induction as such.
Usually it is to be found at the more elementary stages of learning. It
becomes rarer as the theoretical framework develops. Suppose, for ex-
ample, that a sample of a new chemical compound has dissolved in water.
From this we conclude that all samples of the same compound will dissolve
under the same conditions. It may look like an inductive generalization,
but actually we are employing a theory which tells us that solubility (like
other properties) depends on the chemical structure and not on the par-
ticular sample. The very fact that we have generalized from a single in-
stance indicates the presence of some background theory. For, as a rule,
basic induction proceeds at a slower rate. (I shall not go here into the
problem of kinds. Even if we grant that, a priori, matter is conceived in
terms of kinds, it does not follow that the chemical structure by itself
determines the kind). Simple inductions played a crucial role in the devel-
opment of our theory; but further developments employ higher order and
more sophisticated reasonings by analogy.
The example from Newton's Principia is most striking, as it illustrates

induction on several levels within a reasoning which combines it with an
already advanced theoretical framework. It starts at the basic level of enu-
merating gold, silver, lead, glass, sand, common salt, wood, water and
wheat as confirming instances - a starting point which is unthinkable in
later physics - and it reaches to the stars.
A classical example of simple induction has been the firm belief in the
permanence of the day-night pattern and the periodic succession of sea-
sons: "While the earth remaineth, seed-time and harvest and cold and heat
and summer and winter, and day and night shall not cease" (Genesis 8.22).
This generalization has nowadays been modified and it is derived from
scientific theory. The solar system model and the laws of motion imply
that, short of a cosInic catastrophe, the earth will continue to revolve
around the sun and around itself; and the more recent theory of astro-
physics yields estimates of the sun's life time. Thus, while basic induction
remains embedded in the very foundations of the theory (where it can be
revealed by a Humean analysis) it is eliminated at the surface.
To be sure, the Bible provides its own theoretical account: the unbreak-
ing pattern is guaranteed by God's proInise to Noah. Nonetheless I think
that the ancient belief, which was common to humans in general, is a clear
case of surface induction; it probably continues to be so today for most
people. Ancient geographic exploration revealed exceptions to the simple
day-night pattern and led to a modified law correlating the pattern with
geographic latitude; this, again, was accepted by virtue of straightforward
induction4 . Babylonian and Egyptian astronomy (from what we know of
them) exemplify this sort of establishing surface regularities which may be
quite intricate.
Trying to find a living example of surface induction, it seems to me that
the best candidate is the belief in "All men are mortal". Or, to be precise
and to avoid ambiguous reference to future men, my belief that none of
the people living today existed at the time of Newton. Taking into account
that there are more than 4.5 billion people today, this appears to be quite
a strong assertion. When asked for a reason I can only point out at the
accumulated data and, by straightforward simple induction, argue that we
have before us a law of nature. Attempts to explain it (e.g., by liInited cell
renewal) are relatively recent and the fact remains that at present we have
no derivation of it other than direct generalization from the data.
14 HAIM GAIFMAN
Current statistical practice is a good source of illustrations of the in-

ductive principle, though not in the simple form just now considered. As-
sume that a manufacturer of light bulbs has to decide whether to replace
machine A by the better reputed machine B. Choosing and analyzing sam-
ples of say 1000 bulbs, the statistician will end by adopting a certain sta-
tisticals hypothesis which, together with the cost of change, will determine
the decision. The decision presupposes that a certain statistical law will
hold; that is to say - a certain statistical pattern, or regularity, is projected
from the 1000 observed cases to the next say 2,000,000 ones. The story is
not so simple, in as much as the projection is not a clear cut deterministic
forecast; it comes with margins of error and probabilities attached to them.
Nonetheless there is an obvious inductive principle at work. In a Bayesian
framework, assuming that we have a prior distribution to start with, the
picture looks different. It appears that instead of projecting an hypothesis
we are only computing conditional probabilities. But the inductive prin-
ciple enters into the choice of the prior probability. This prior distribution
is determined by the kind of regularities that we expect; for an analysis of
this point see my 'Subjective Probability Natural Predicates and Hempel's
Ravens's.
To sum up, if clear-cut cases of simple surface induction are not fre-
quent, it is because this mode combines at various levels with more so-
phisticated modes of reasoning and with all kinds of theoretical knowl-
edge. Like hydrogen on earth it is very common but hard to find in pure
form.
4. INDUCTIVE SUPPORT AS PROBABILITY INCREASE
In the course of induction new confirming evidence makes the hypothesis

in question more credible, or more reliable, or more acceptable, in some
sense of acceptance. This is usually described as the inductive support of
the hypothesis by the evidence. The somewhat loose, unspecified notion
can be made more precise if we express inductive support as an increase
in subjective probability, in particular - when a quantitative measure of
probability is available. With h - the hypothesis, e - the new evidence and
p( ) - the probability function, a probability increase means that p(hle)
> p(h). Here we have amalgamated the existing background knowledge
into 'p'; i.e., p(h) is the probability of h, given the background knowledge,
and p(hle) is the probability of h given this background and e. (With the
background knowledge explicitly displayed by 'b' the inequality becomes
p(hlb /\ e) > p(hlb); but henceforth I shall leave it undisplayed).
My aim is to clarify some points concerning the relation between in-
ductive support and probability increase. In particular, I shall show that
some recently advanced arguments against probabilistic expression of in-
ductive support are totally unfounded.
The significance of a probability change depends highly on the given
context. For example, how does an increase from 0.7 and 0.9 compare
with an increase from 0.1 to 0.4? The first increase is smaller than the
second if "amount of increase" is measured by the difference p(hle) -
p(h); and it is considerably smaller if it is measured by p(hle)/p(h). But in
the context of decision making, where the outcome depends on the truth
of h, it may well happen that an increase from 0.7 to 0.9 makes all the
difference between deciding to do A or to do B, whereas a change from
0.1 to 0.4 is insufficient to make any change of decision. Hence, when
speaking of amount of increase I only assume some unspecified way of
measuring, or comparing, probability changes which is appropriate to the
given context6 •
In general, probabilistic support, i.e., an increase from p(h) to p(hle),
does not in itself indicate that we have a case of induction. For one thing,
if p(h) < 1 and if, given the background knowledge, e implies h, then
p(hle) = 1; but the increase from p(h) to p(hle) does not signify here an
inductive support (One might describe it as a "deductive increase").
If h is a universal generalization and e is one particular instance of it
then h implies e and, by Bayes' law, p(hle) = p(h)/p(e). Hence, with pee)
< 1, we have p(h) < p(hle). Sometimes this corresponds to inductive
support, e.g., the support of "All ravens are black" by an additional con-
firming instance. But the obtaining of this formal scheme does not necess-
arily mean that we have a case of induction. For we can form universal
hypotheses using perverse arbitrary predicates, like Goodman's 'grue'. In
general for any two statements whatsoever, et,e2, if h = el /\ e2, then
p(hlel) ~ p(h); the inequality is strictifp(h) > Oandp(el) < 1. Evidently,
in most cases, we do not want to regard this as a case of inductive support.
But with a specially constructed predicate P we can recast el /\ e2 in the
logically equivalent form: "All physical bodies are P's" and el as "The
planet Jupiter is a P"; hence it falls under the general formal scheme. (To
16 HAIM GAIFMAN
do so let "x is P" be equivalent, by definition, to "x is the planet Jupiter

and el is true, or x is any other physical body and ez is true").
These considerations show that in the case of a universal generalization,
h, and a confirming instance, e, the inductive support is expressed better
by considering the universal generalization restricted to all the remaining
unobserved cases, h'; the change fromp(h') to p(h'le) is a better indicator
than the increase from p(h) to p(hle). For example, the evidence: "the
emerald b is green" increases the probability of "all emeralds, except those
observed already and except b, are green"; but when 'green' is replaced by
'grue' this no longer holds and the probability decreases (unless b is ob-
served after 2000 A.D.). Here I assume of course that p( ) is a reasonable
function which reflects our basic intuitions. It is always possible to define
a perverse function so that the roles of 'green' and 'grue' are reversed.
Probabilities can be used to express and to measure expectations of
regularity'. But probability theory, in itself, does not rule in favour of
regularities of any particular kind. The natural predicates and the patterns
that will serve us as a starting point are determined by epistemic factors.
Hence it is to be expected that factors of this kind will enter into deter-
mining whether a certain probability increase signifies inductive support
or not. To what extent can we read the picture from the probability func-
tion alone and from the formal logical (say, first-order) structure of the
statements over which it is defined? This is an open question. In any case,
we would have to consider a probability which is defined on comprehensive
domains, incomparably richer than the sentential combinations of the hy-
pothesis h and the evidence e.
So far I have interpreted 'the support of h bye' as the change in the
standing of h which is caused bye. There is also another meaning of'sup-
port'. When a lawyer claims that the evidence supports his client's version,
he means first and foremost that, with the evidence given, his client's ver-
sion looks good, i.e., appears very credible. Whether it was as credible
without the evidence is of secondary importance. Let me use 'supportc' for
support as an indicator of change of status and let me use 'support" for
support as an indicator of final status. The evidence, e, supports. the hy-
pothesis, h, to the extent that it has improved h's standing; it supportsf h
to the extent that h has a good standing given the evidence. In ordinary
usage 'support' is ambiguous and quite often it connotes both meanings,
namely: e has improved the standing of h and, consequently, h enjoys now
quite a good standing.
A large part of the Popper-Camap controversy of the fifty's was due to

nothing more than the overlooking of this elementary distinction and to
confusing these two meanings of 'support'. By now the situation has been
clarified, but the same sort of confusion reappears occasionally.
Note that the distinction between supportc and support! applies to all
kinds of support probabilistic or not. In terms of probabilities, supportc
is given by the change fromp(h) to p(hle), whereas support! is given simply
by p(hle). One may claim that a certain reading ('supportc' or'support/)
is more appropriate than another for some particular notion of support.
Whatever the merits of the claim, one should, in order to be consistent,
use the same reading when comparing this notion with other support no-
tions.
Isaac Levi has recently suggested 8 that if, given e, two hypotheses are
equivalent then they have the same inductive support. This is true only if
'support' is read as 'support/. If hI and h2 are equivalent given e, then
indeed their final standings, after obtaining e, are trivially the same. But
their initial standings, before the obtaining of e, might have been quite
different; in which case their supportsc must be different as well. This is
true no matter what notion of support one has in mind. Isaac Levi notes
however cases in which hypotheses which are equivalent given e have dif-
ferent probabilistic support bye. But here he interprets 'support' as a
change from p(h) to p(h Ie), i.e., as supportc. This brings him to the erro-
neous conclusion that inductive support cannot be probabilistic. In fact,
if hI is equivalent to h2 given e then they do have the same probabilistic
support!, p(hIle) = p(h2Ie). Given that the first two tosses yield 'heads',
the conjecture of obtaining three straight 'heads' is equivalent to the con-
jecture of obtaining 'heads' on the third toss. Shall we then say that they
have the same support by the evidence? Yes - if by 'support' we mean
support" no - if we mean supportc. And this is true for support in general,
probabilistic, inductive or what have you.
Recently Popper and Miller have proposed a surprising proof which

shows, so they claim, that inductive support is not probabilistic9 • It runs
as follows:
Let h be some hypothesis and e some evidence. Define hI = D! e v h,
h2 = D! -, e v h. Then h is logically equivalent to hI " h2 and e
implies hI. It follows (trivially) that, when e is given, h is equivalent to h2
18 HAIM GAIFMAN
and hI and h2 are probabilistically independent:

P(hl /\ h2 I e) = p(hIle)·p(h2Ie).
Moreover hI is the strongest statement which is logically implied either by

e or by h. Also, if we keep one of the h;'s fixed, then the other hi is the
weakest statement whose conjunction with the first is logically equivalent
to h. Therefore, the argument goes, h has been split into two factors, one
of these, e v h, is logically implied by e hence deductively supported by
it, and the other, -, e v h, "contains all of h that goes beyond e". Hence,
the argument continues, the question whether the inequality p(hle) >
p(h) can be interpreted as inductive support reduces to the following: "Can
e, in this case, provide any support for the factor -, e v h, which in the
presence of e, is alone needed to obtain h?" An easy calculation shows
that p(-, e v hIe) :S p(-, e v h) and, except for the case where either
p(e) = 1 or p(hle) = 1, the inequality is strict. This is claimed to prove
that "all probabilistic support is purely deductive" a result which is "com-
pletely devastating to the inductive interpretation of probability".
First we note that Popper and Miller take the inequality p(hle) > p(e)
as a criterion of probabilistic support. 'Support' is therefore to be read as
'supportc', it indicates the change of status. Hence the emphasis laid on h
being equivalent to h2 given e is altogether pointless. The equivalence given
e implies nothing concerning the supports, of hand h2 bye. But repeated
references to it in the proof suggest to the reader that it implies somehow
the equality of the inductive supports given to h and to h2 • Popper and
Miller do not make explicitly this wrong claim, but apparently the phrasing
was enough to mislead Isaac Levi.
The probabilistic independence of hI and h2 given e is pointless as well.
Actually it amounts to a triviality, for hI is equivalent to a tautology given
e, and a tautology and any other statement are probabilistically indepen-
dent. It can be shown lo that, without presupposing e, hI and h2 are never
probabilistically independent except for the uninteresting case where either
p(hle) = 1 or p(hl-,e) = 1. Probabilistic independence is, however, irrel-
evant altogether to the question at hand.
Why then should the inductive supports of hand h2 by e be the same?
We are given the hint that h2' i.e., -, e v h, "contains all of h that goes
beyond e". Why? Presumably, because it is the weakest statement whose
conjunction with e v h is equivalent to h. Now induction can indeed be
described as affirming "the part that goes beyond the evidence"; but
"going beyond the evidence" has in the context of induction quite a dif-
ferent meaning. Let h be the hypothesis that all the bills in a given bundle
are forged and let e be the statement that the first bill is forged. Then
obviously the part, say h', of h which goes beyond e and whose standing
is improved inductively by e is: "all the rest of the bills are forged", not
h2 which reads: "the first bill is not forged or all the bills are forged".
(Again, all the three statements h, h' and h2 become equivalent when e is
given, but the changes in their standing which are caused by e are quite
different; even without computing probabilities one can see that h2 looks
better without e than with it).
Thus, the formal criterion by which Popper and Miller define "going
beyond the evidence" has nothing to do with inductive support. Their
factoring of h (i.e., its representation as h1 A h 2) is a purely formal exercise
which has no bearing on the issue. If we want to use factorizations then
the factoring which is relevant for inductive reasoning is, in the example
just given, e A h'; for it is the support of h' which expresses the inductive
element of the inference. Aside from this factorization there are, as Jeffrey
pointed out, many (in fact, infinitely many) non-equivalent factorizations
in which the first factor is implied bye. The second, "non-deductive" factor
will get as a rule varying degrees of support bye, covering quite a wide
range.
Factorization has in general a strong appeal, for it is a fundamental tool
in science and in conceptual analysis. We decompose complex impressions
into their simpler constituents, forces - into components, numbers - into
their prime factors, etc. etc. But it is an amusing and not too difficult
exercise to produce ad-hoc factorizations from which most astounding
conclusions follow. Here is an illustration:
It is an empirical fact that the velocity of a free falling body increases
with the height of the fall. After falling (in vacuum) a distance d, the
velocity, v, is given by the formula:
(i) v = J2gd
Let us define b i = Dpl, b 2 = D/ v/d, then by pure mathematics:
(ii) v = bi . b2
The first factor increases with d, by pure mathematics (or even pure logic).
20 HAIM GAIFMAN
The second which is equal to v/d consists of "all of v that goes beyond d",
or, to paraphrase another expression of the Popper-Miller proof, it is the
"product-wise ampliative component". Our empirical formula (i) implies
that b 2 = .J2g/d. Thus b 2 decreases with d. The increase of v with dis

therefore represented as the final outcome of two contributing effects: a
purely mathematical increase of b i and an empirical decrease of the am-
pliative b 2 • Therefore the increase of v with d has a purely mathematical
origin!
The Popper-Miller proof is as valid as that.
NOTES
• Research supported by a DFG grant.

1 For example, see R. R. Bush and F. Mosteller, Stochastic Models for Learning, Wiley &
Sons, New York, 1955.
z In his Mathematics and Plausible Reasoning, Princeton University Press, Princeton, 1954
and Mathematical Discovery, Wiley & Sons, New York, 1962.
3 To be more precise, h is informative to the extent that p( Ih) is less evenly spread than
p( ). One way of measuring "evenness of spread" is by entropy: there are others. Being less
evenly spread means that the probability distribution implies a stronger factural commitment,
or, in other words, has more empirical content. In my 'Towards a Unified Concept of Prob-
ability', (Technical report, January 1984; to appear also in the Proceedings of the 1983 In-
ternational Congress for Logic Philosophy and Methodology of Science) there is a detailed
discussion of these points.
4 In Conjectures and Refutations, chapter I, IV, Popper uses the exception to the ordinary
day-night pattern in the northern zone, as an argument against inductive reasoning; he claims
that this reasoning led to wrong conclusions. It is difficult to see the point of the argument.
Inductive reasoning is not guaranteed to be fool proof. No empirical method is. The example,
on the contrary, illustrates the flexibility of induction and its power to readjust the pattern.
For the new, more complicated, law which correlates the day-night pattern with geographic
latitude is again accepted by induction. Theoretical models and explanations come later.
S In Erkenntnis 14 (1979), 105-147.
6 In 'Subjective Probability, etc.', ibid., I have used the ratio (I - p(h»/(I - p(hle» as a
measure of the support given by e to h. It seems a reasonable measure and, unlike p(hle)
- p(h) or p(hle)/p(h), fits a wide range of contexts. But, as noticed in that article, it also has
its limitations.
7 For a discussion of this point see 'Subjective Probability etc.', ibid.
8 In 'Popper and Miller Are Right!', mimeographed note July 1983.
9 'A Proof of the Impossibility ofInductive Probability', letter to Nature 302,21 April 1981,
687-88. The proof has been also presented in 'The Calculus of Probability Forbids Ampliative
Probabilistic Induction', a paper contributed by Popper to the 1983 Congress of Logic Meth-
odology and Philosophy of Science. Abstracts, Vol. 1, pp. 250--252.
10 Let Xl = p(e), Xz = p(-,e), Yl = p(e A h), Y2 = p(-,e A h). Then P(hl A h z) =
p(h) = Yl + Y2 and p(hl ) = Xl + Y2, P(h2) = X2 + Yl. Since Xl + X2 = I, we get by
elementary algebra: p(h l ) . P(h2) - P(hl /\ h2) = (Xl - Y2) (X2 - Y2). This value is > 0,
unless Xl = YI or X2 = Y2.
Manuscript received 25 June 1984
Institute of Mathematics and Computer Science

The Hebrew University
Givat Ram 91904
Jerusalem
Israel
CLARK GLYMOUR
INDUCTIVE INFERENCE IN THE LIMIT
There is an enormous and interesting theoretical literature 1 on inductive

inference which remains largely unknown to philosophers of science, even
though a philosopher, Hilary Putnam, may be said to have initiated it2 •
The work in this tradition concerns algorithms for inferring recursive func-
tions from finite samples oftheir graphs. Informally, an inductive machine
is an algorithm which is given larger and larger samples of the graph of
a partial or total recursive function which the machine attempts to identify.
The machine or algorithm produces at various stages a program which
computes a recursive function. The machine is said to identify the target
function if at some point it produces a program which computes that very
function, and thereafter, no matter how much more of the graph of the
target function it sees, continues to produce the same program. A weaker
notion, that of behaviorally correct identification, does not require that in
order to identify the target function the machine converge to a single pro-
gram. Instead it requires only that the machine converge to a (possibly
infinite) set of programs, all of which compute the target function. A class
of functions is said to be identified by a machine if the machine identifies
every function in the class. This framework has been adapted to charac-
terize abstract versions of learning languages in the limit. A natural ques-
tion is whether a similar framework can be developed for learning first
order theories. What follows is an attempt to present such a conception,
and to describe what place Hempel's ideas about confirmation have within
it. There are many definitions, some examples, no theorems, and a great
deal left to be investigated.
A theory can be regarded as a recursively axiomatizable deductively
closed set of first order sentences. It can also be regarded as the collection
of all models of such a set. We might think of the theory as a formalization
of celestial mechanics, for example, and its models as all of the various
n-body solutions to the Newtonian equations. In practice we get evidence
for a theory by determining particular features of systems to which the
theory is supposed to apply. We test theories of celestial mechanics by
Erkenntnis 22 (1985) 23-31. 0165-0106/85/0220-0023 $00.90

24 C.GLYMOUR
determining features of various satellite systems, or of the galaxy, or of

the entire cosmos. By analogy, we should perhaps think of the evidence
to be presented to an induction machine as a sequence of particular facts
from various models of the theory. I will represent particular facts by
Hempel's basic sentences, that is, by atomic sentences or negations of
atomic sentences. So the evidence in principle available to an induction
machine is a sequence of basic sentences, with each sentence identified as
to the model from which it is taken. Given appropriate finite initial seg-
ments of such a sequence, the machine must guess a description of the
theory from whose models the evidence is taken.
The first difficulty in making this idea precise is that the facts implicit
in all models of a theory cannot be enumerated. How much can the set of
all models of a theory be reduced without loss of information? Any gen-
eralized elementary class is closed under model isomorphism. Suppose we
dispense with isomorphic copies and restrict attention to one representa-
tive from each isomorphism class of the theory. The upwards Lowenheim-
Skolem theorem reminds us that unless the theory has only finite models,
there will always be in any reduced elementary class models that caimot
be enumerated. The downwards Lowenheim Skolem theorem reminds us
that every model with an uncountable domain has a countable, elemen-
tarily equivalent submodel. But even if we confine ourselves to represen-
tatives of isomorphism classes of models with countable domains, the facts
still cannot be enumerated unless the theory has only a countable collec-
tion of (not elementarily equivalent) models with countable domains. To
present the evidence for a theory we must choose some countable collec-
tion of countable models from which the evidence is to be taken:
DEFINITION: An evidence frame for a theory is a countable collection

of countable models of the theory, no two of which are elementarily equiv-
alent.
Some theories, of course, have only a countable number of elementarily

inequivalent countable models, and for these theories there is a frame in
which all of the facts can be enumerated.
Given a frame, the models within that frame can be enumerated, al-
though the enumeration cannot always be recursive. How are the facts
within the various models of the theory to be represented? Assume a first
INDUCTIVE INFERENCE 25
order language L, and add to L a denumerably infinite collection of in-

dividual constants. Call the extended language LC. If M is a model in the
frame of a theory, expand M by assigning to each individual in the domain
of M a unique individual constant which names that individual. Call the
resulting expansion of M, M*.
DEFINITION: The diagram, D(M), of a model M is the set of all quan-

tifier free atomic or negated atomic sentences in LC which are true in M*.
We can recursively assign a natural number, a Godel number for ex-

ample, to every well-formed sentence in L or in LC. The diagram of a
model M can then be represented by its characteristic function.
Since given a frame the models in the frame can be enumerated and the
diagram of each model in the frame can be enumerated, we can enumerate
all the facts the frame encompasses. In addition it seems natural to suppose
that when all of the facts of a finite model have occurred in a presentation
of a frame, that feature is somehow noted.
DEFINITION: A presentation of a frame for T is a sequence S or ordered

pairs (i, n) or (i, #) such that
i. [iji is in a member of S] is an enumeration of the members of the
frame;
ii. if (i, n) is in S then n is the number of a sentence in some D(M) with
M in the frame, and i is the number of M;
iii. for every M in the frame and every sentence pin D(M), there occurs
in S a pair (i, n) such that i is the number of M and n is the number of p;
iv. if M is a finite model with index i, and all individual constants oc-
curring in sentences in D(M) occur in the n element initial segment of S
but not in the n - 1 initial segment of S, then the n + 1 element of S is
(i, #).
I will refer to the first member of an element of a pair (i, n) as the index
of the second, and although the second member is a number I will usually
describe it instead as the sentence having that number. In the study of
machines for learning classes of recursive functions, the data can be pre-
sented, well, recursively. A frame cannot in general be presented recur-
sively, and we must think of the evidence as provided by nature's oracle.
An inductive inference machine must guess a description of an axiom
26 C.GLYMOUR
system for the theory from whose models the evidence is taken. We can
think of the description of the axioms either as an effective procedure for
constructing an axiom set, or as a program for computing the character-
istic function of a recursive set of axioms. Informally I think of a machine
the first way, formally the second way:
DEFINITION: A machine, R, for a first order language L is a partial

function on finite initial segments of presentations of frames for theories
in L such that
i. If F is the set of all finite initial segments of presentations, R is partial
recursive in F;
ii. for any sequence Sin F, R(S), if defined, is a program which com-
putes the characteristic function of a set of sentences in L.
Hempel's theory of confirmation provides the basis for an inductive

inference machine. Indeed, for a class of theories it is the best possible
machine. The Hempel machine works as follows:
Let Sn be an initial segment of a presentation of a frame for theory T.
Let Ki be the conjunction of sentences in Sn with index i. Substitute distinct
variables for occurrences of distinct individual constants in Ki and form
the existential closure, which I denote by EKi. Note that if there are m
distinct individual constants occurring in sentences with index i, EKi en-
tails that there are at least m distinct things. If there are m distinct indi-
vidual constants occurring in sentences in Sn with index i, let Ai be the
sentence that says there are no more than m distinct individuals.
For input Sn the Hempel algorithm produces the disjunction
Vi (Ai & EKl)
where i ranges over indices occurring in Sn.

Hempel's confirmation theory takes a sentence to be confirmed if it is
true in every model of the evidence in which each element of the domain
is named by some individual constant occurring non-vacuously in the evi-
dence. The Hempel algorithm simply extends the idea to the predicate
calculus with identity, and supposes that the machine conjectures an ax-
iomatization of exactly those sentences which are confirmed by the evi-
dence. What theories can the Hempel algorithm identify? To answer that
we must first say what we mean by the question. I will distinguish three
senses in which a machine could be said to identify a theory. CR(Sn)
denotes the set of sentences for which the value of the characteristic func-
tion computed by program R(Sn) is unity:
1. Machine R L-identifies theory T if and only if there exists a frame
for T such that for every presentation P of the frame there exists an N
such that for all n > N CR(Pn) and CR(PN) are logically equivalent and
axiomatize T.
2. Machine R W-identifies T if and only if there exists a frame for T
such that for every presentation Pof the frame there exists an axiomati-
zation AX of T and for every sentence A in AX there exists an N such that
for all n > N A is in CR(Pn) and for every sentence B not in AX there
exists an N such that for all n > N B is not in CR(Pn).
3. Machine R V-identifies T if and only if there exists a frame for T
such that for every presentation P of the frame for every sentence A in T
there exists an N such that for all n > N CR(Pn) entails A and for every
sentence B not in T there exists an N such that for all n > N CR(Pn) does
not entail B.
4. A machine R (L, W, V) identifies a class of theories if and only if R
(L, W, V) identifies every theory in the class.
The differences among these definitions can be illustrated by considering

the Hempel algorithm. The Hempel algorithm L-identifies every theory
and thus every class of theories having only finite models. But the Hempel
algorithm cannot L identify any theory having an infinite model. Neither
can it W-identify a theory having an infinite model. Consider, however the
theory with no non-logical terms which says that there are an infinite num-
ber of objects. The theory is recursively but not finitely axiomatizable and
it can be V-identified by the Hempel algorithm.
Other algorithms for inductive inference are possible, of course, and an
interesting question concerns for what classes of theories there exist ma-
chines adequate to identify them. I will not attempt to answer the question
save to remark on one obvious negative result giving a sufficient condition
for a class of theories not to be identifiable by any machine.
Suppose two theories, T and T' have the property that for every pair of
countable models M of T and M' of T' and for every n > 0 every rest-
riction of M to an n member subdomain is isomorphic to every restriction
of M' to an n member subdomain. Then no class of theories containing
both T and T' can be identified. It is clear that at no stage of a presentation
can a machine determine whether the model is of T or of T'. The theories
28 C.GLYMOUR
of linear discrete and dense order with first but no last element have this
property. We can imagine a modification of the capacities of a machine
that would make it possible to discriminate between discrete and dense
order. Suppose machines did not have to receive evidence passively but
could at any stage in the presentation of the evidence, for any quantifier-
free open formula expressible in LC, ask whether a model with a particular
index contains a sequence of individuals satisfying the description. Sup-
pose further the machine receives an answer: either the report that there
is no such sequence, or a sentence which instantiates the description and
is true in the model. With that information one can devise an algorithm
which can distinguish between discrete and dense orders. Call such ma-
chines Q machines.
A machine R would be ideal if no other machine could identify every
class of theories R identifies plus some more. There are an abundance of
results about ideal learning machines in the recursive function theory lit-
erature. In most senses of ideal, there aren't any ideal machines for infer-
ring recursive functions. It seems unlikely that better results will hold for
machines that infer first order theories, but I have no proofs to give. Cer-
tainly the Hempel algorithm is not ideal. Consider an algorithm that defers
conjecturing that every individual in a domain is among those named by
an individual constant occurring in the evidence already seen: it produces
Eki but not Ai. Such an algorithm can still conjecture Ai when it receives
notice that all individuals in a model with index i have been mentioned in
the evid~nce, and it can therefore still identify all theories with only finite
models. However, when the individuals in a model with index i have not
all been named, the algorithm is free to conjecture sentences for i which
have only infinite models.
There are many potentially interesting generalizations of the framework
just described:
1. In reality, not all features of a structure which models a theory are
usually available to observation. One can allow that the possible theories
may contain predicates and function symbols not included in the language
in which the evidence is presented, or equivalently, that the frame consists
not of models for a first order theory, but of reducts of such models. The
reducts may be uniform - that is, the same relations may be deleted from
all models in the frame - or different relations may be deleted from dif-
ferent models of the same theory. More realistically still, the reducts may
be both non-uniform and partial; that is, within a model in the frame, part
of the extension of some relations may be deleted, so that every presen-

tation of the frame fails to specify whether certain predicates apply or do
not apply to some sequences of individuals.
2. Whenever the evidence is fragmentary, as in each of the circum-
stances in 1, underdetermination will be rife. For every theory there will
be a class of alternative theories and no machine will be able to descri-
minate among them, even in the limit. One can still ask whether formal
versions of methodological virtues and presumptions, such as simplicity,
testability, explanatory power, or cosmological principles, permit the iden-
tification of theories having those virtues. Further, even when because of
the fragmentary character of the evidence a class of theories cannot be
identified, it may still happen that there is a machine which for each theory
determines, not an axiomatization of the theory but an axiomatization of
part of the theory.
3. One can fix a particular theory, or even a particular model of a theory
such as the standard models for set theory or for number theory, and ask
whether first order extensions of the theory (or expansions of the model)
are identifiable. For example, one could fix the natural numbers, and ex-
pand the structure by introducing a base function for some recursive func-
tion. A machine is then presented with facts about the expanded model,
and since the machine "knows" it is dealing with an expansion of number
theory, the only facts that matter are those in the graph of the function.
The machine must then determine a program which computes a set of
axioms which when added to number theory has the expansion of the
standard model as its only model which reduces to the standard model.
This is but a variant of the problem addressed in the recursion theoretic
literature on learning in the limit.
4. Rather than attempting to characterize axioms for a theory, a ma-
chine might instead attempt to characterize axioms for sentences of a
special form that are consequences of the theory and hence true in its
models. For example, a machine might attempt to find an axiomatization
of all universal sentences that are consequences of a particular theory.
Using Hempel's satisfaction criterion of confirmation, Kevin Kelly has
developed an efficient algorithm for determining an axiomatization of the
universal sentences true in a model.
The framework I have described for the study of algorithms for inferring
first order theories is in the spirit of Hempel's philosophy of science: there
are no a priori restrictions on the logical form of the axioms of a theory,
30 C.GLYMOUR
or on the theories themselves, save that they be first order. Satisfaction

criteria are obviously appealing in this framework and we have seen that
Hempe1's satisfaction criterion of confirmation yields an inference algo-
rithm. Of course, the framework is not confined to the analysis of algo-
rithms that use satisfaction criteria. There is a previous framework for
analyzing the problem of inferring first order theories in the limit. That
account, due to Ehud Shapir0 3 , is in the spirit of Popper's philosophy of
science in something like the way that the framework given here is in the
spirit of Hempel's. Shapiro takes the learning task to be that of charac-
terizing features of a particular model. Presentations of the evidence con-
sists of a sequence of pairs <s, v), where s is a sentence in a language Lo
(which Shapiro usually takes to be singular atomic sentences or negations
of atomic sentences in a first-order language with identity and function
symbols) and v is either "true" or "false" accordingly as s is true or false
in the target model. The theories are expressed in some language Lh which
is an extension of Lo. A machine for the language produces from appro-
priate finite sequences of a presentation of the facts about a model a finite
collection of first order sentences. Shapiro's machines are permitted to ask
whether particular sentences in Lo are true in the model, and receive ap-
propriate answers. Shapiro says that a machine identifies a model provided
for each presentation of the facts in the model there is an N such that for
all n > N the machine output from n facts is the same, and that output
is a finite set of sentences that (i) is true in the target model and (ii) logically
implies all sentences of Lo true in the model. Shapiro describes algorithms
for this framework which identify certain special classes of models, and a
computer program which implements a Popperian algorithm for inferring
sets of Hom sentences.
Shapiro's work is important and impressive. The framework presented
here is, I believe, nonetheless more general and less artificial than Shapiro's
model inference problem. There are several obvious differences in the
frameworks, and one that is not obvious but which is the most important.
I would permit many models, not just one, and recursively axiomatizable
theories not just finitely axiomatizable theories. I do not require for iden-
tification that a machine fix on a particular axiom system, but merely that
it fix on logically equivalent axiom systems. While I believe in each of these
respects Shapiro's requirements are gratuitously restrictive and less real-
istic, these differences by themselves are perhaps not so important. For
example, since the models are always countable they could all be coded in
a single expansion of the natural numbers. The important difference is that

Shapiro's criterion for identification requires that a machine produce a
finite set of sentences that is true in the model and that entails a Lo sen-
tences true in the model.
Fulfilling this goal is the principal object of the various induction al-
gorithms he describes. The goal is inappropriate, both philosophically and
formally. Philosophically, we certainly do not require that our scientific
theories entail every relevant particular fact, only relations among such
facts, or relationships which the facts instantiate. Formally, the second
clause of Shapiro's identification criterion is vacuous for theories that con-
tain no individual constants, and inadequate for theories which fail to
contain for each individual in the target domain a term naming that in-
dividual. Thus if the model were simply an order, whether dense or dis-
crete, without endpoints, there would be no facts in Shapiro's sense to
present. If we added to the language an infinite set of constants naming
each individual in the domain, no finite set of axioms in the language
could entail all of the facts in a dense ordering. It is no accident therefore
that Shapiro's examples are confined to models of theories for which there
are a finite set of universal sentences generating a term naming each object
in the domain of the model.
NOTES
1 For a recent review see D. Angluin and C. Smith, A Survey of Inductive Inferences: Theory
and Methods, Technical Report 250, Yale University Department of Computer Science, Oc-
tober 1982.
2 H. Putnam, Probability and Confirmation in Mathematics, Matter and Method, Cambridge
University Press, Cambridge, 1975. Another philosopher who was made substantial contri-
butions to the subject is Scott Weinstein. See D. Osherson and S. Weinstein, 'Formal Learn-
ing Theory', in M. Gazzaniga and G. Miller (eds.), Handbook ofCognitive Neurology, Plenum
Press, New York, 1983.
3 E. Shapiro, Inductive Inference of Theories from Facts, Research Report, Yale University
Department of Computer Science, 1982.
4 I thank Robert Daley, Richmond Thomason, Kenneth Manders and Kevin Kelly for help-
ful discussions.
Received 4 July 1984
2510 Cathedral of Learning

Center of Philosophy of Science
University of Pittsburgh
Pittsburgh, PA 15260
U.S.A.
D. COSTANTINI
PROBABILITY AND LA WS*
1. The year 1945 was an important one for inductive logic. Of the many
papers dedicated to this discipline in that year, two were crucial. In a sense,
both pointed out paradoxical aspects of confirmation. One paper was writ-
ten by HempeP and, among many other things, gives notice of an argu-
ment that became famous as Hempel's paradox. The other was written by
Carnap2 and the paradoxical side of this paper came from the fact that,
according to the first explicitly formulated confirmation function, c*, the
degree of confirmation of an hypothesis of universal form, in an infinite
language, is always zero: the paradox of null probability.
The two paradoxes were not new. On the one hand, that of Hempel
which was probably verbally communicated by this author to Hosiasson-
Lindenbaum, had already been solved by her3. On the other hand, Carnap
was not the first to notice that in an infinite population laws have a prob-
ability equal to zero. At least Pearson, Broad and Jeffreys, applying La-
place's rule of succession to an infinite junction, had reached the same
result. Moreover Jeffreys had gone further, suggesting a solution which
was taken up twenty years later by Hintikka.
However, the papers of Hempel and Carnap, the first from a general
point of view and the second from a more specific one, opened a new era
for the studies on inductive logic. This was mainly because they had clearly
stated the difficulties which arise in dealing with laws. In spite of the
numerous attempts made to deal with the problem posed by laws, it has
still not been solved. In the present paper, developing the ideas on which
were founded the solutions proposed by Hempel and Carnap to their res-
pective "paradoxes", we try to give an unified approach to laws. However
in this context a warning is needed. The approach we have in mind does
not cover laws varying with time. That is we do not intend to treat laws
of evolution like the law which governs for example the production of
wheat of a given parcel of land.
2. Neither Hempel nor Carnap specify the ambit the term "law" was in-
Erkenntnis 22 (1985) 33-49. 0165-{)106/85/0220-0033 $01.70
34 D. COSTANTINI
tended to deal with. However they have in mind physical theories as is

suggested by their frequent references to coordinate languages. But, pass-
ing to the formalization, they make use of monadic predicate languages.
This restricts their attention to the probabilistic approach to science. This
limitation, far from being a weakness, is the strength of the inductive logic
which they were constructing. It pointed out, at least implicitly, that in-
ductive logic refers to statistical problems: it represents the foundations of
statistical inferences.
The paradox of Hempel has been of great value in making clear that
inductive logic does not refer to physics, at least to that part of physics
which can be developed without probabilistic notions. To illustrate this
we quore some considerations of Toraldo di Francia4 , one of the few phy-
sicists who have showed interest in inductive logic.
The author, who is a physicist, has several times experienced the feeling that the procedure
of scientific induction, as discussed by many epistemologists does not entirely coincide with
the tool applied by modern physicists [...J A queer situation occurs, especially when a phi-
losopher recognizes the fundational shakiness of a given procedure, and takes for granted
that physicists are still applying it, while physicists are well aware that the procedure can not
be applied at all.
To give a very rough but illuminating example of what physicists are not doing anymore,
suppose that an astrophysicist has observed 100 stars taken at random in the sky and has
found that all of them have some property A. He never concludes that all stars have property
A, and, what is most important, he will not even say that it is probable that all stars have
property A.
After having recalled the paradoxes of Hempel and Goodman, Toraldo

di Francia concludes
A number of ingenious ways have been suggested to solve these paradoxes so as to maintain
the validity of the above [NicodJ criterion of confirmation. Our "solution" is very simple: the
criterion is unsound and cannot be applied with any confidence.
Toraldo di Francia refers to physics, i.e. Nicod's criterion does not work
in physics. On the contrary, it works in statistics and this is our philo-
sophical starting point. This criterion is an essential tool in statistics and,
what is more, in connection with laws. But before going further we must
say something about statistical inference.
3. In statistics one deals with two types of problems: estimations and tests
of significance. Problems of the second type aim at the falsification of
PROBABILITY AND LAWS 35
hypotheses and are dealt with by the use of deductive statistical inferences,
i.e. inferences making use of likelihoods. The method used in accomplish-
ing these inferences is the hypothetico-deductive method. These inferences
are based on hypothetical probabilities.
Problems of the first type aim at determining probabilities of hypotheses.
This can be done either in an inductive way (inductive inferences) or by
combining induction with hypothetical assumptions (hypothetico-induc-
tive inferences). In order to explain this assertion we recall that there are
essentially two types of estimation. In the first type, the Bayesian estima-
tion, one assumes the validity of a "model", that is one makes specific
assumptions about the population, i.e. one assumes the validity of speci-
fications which say something about the probability distribution of the
observations. This assumption represents the hypothetical side of the
Bayesian estimation. The specifications give the distribution of the obser-
vations except for a parameter. The parameter is known to belong to a
fixed region. The inductive side of the Bayesian estimation lies in intro-
ducing a probability distribution, the prior distribution, for the unknown
parameter. By means of the likelihood, determined with the model and the
observations, the prior is modified into a posterior distribution.
The other type of estimation amounts to estimating the composition of
a population or, in statistical terms, the related unknown distribution func-
tion. This is a purely inductive problem in which there is no room for
hypothetical assumptions. This is the problem faced by Camaps and it is
with reference to this problem that we intend to account for laws.
Hence statistical problems can be divided into two classes: problems in
which one assumes the truth of a (family of) distribution(s), and problems
in which such an assumption is not made. In performing a hypothetico-
inductive inference one assumes the truth of a family of distributions char-
acterized by a parameter in order to estimate the value of the parameter.
In performing a hypothetico-deductive inference one assumes the truth of
a distribution in order to falsify it.
Not assuming the truth of any distribution all that remains to be done
is to estimate the composition of the population we are observing, i.e. to
make an inductive inference.
The following objection can be made. It is impossible to perform any
estimation without assuming (the truth of) a distribution, namely the prior
distribution. Hence there is no difference between estimation and signifi-
cance tests.
36 D. COSTANTINI
We think that this is not the case and that the difference between in-
ductive and deductive arguments in statistics can be made clear just by
considering the different roles played by the assumption of the truth of a
distribution. In a test of significance the distribution which is assumed to
be true at the beginning is not changed by observation; in statistical terms,
independence holds. After having observed the sample either the distri-
bution still holds, and in this case it is corroborated, or it does not hold,
and in this case it is falsified. Falsification is a yes-or-no procedure. In
testing there is no room for gradual learning. In performing an estimation
the prior distribution, the truth of which is assumed at the beginning,
changes gradually with observations; in statistical terms, independence
does not hold. The observation of every individual modifies the prior dis-
tribution, and it is the sequence of all distributions, i.e. the process, which
characterizes the estimation, not its prior distribution, the task of which
is merely to allow the process to start.
In our opinion it is very important to distinguish clearly between prob-
lems in which we make hypothetical assumptions on the populations we
are observing and problems in which we do not make such assumptions.
As we have seen we can learn from experience using both deductive and
inductive arguments. In the first case we have no limitations, as the results
are conditional to the hypothetical assumptions we have made. In the
second case we must remain as near as possible to what we have actually
observed.
Any distribution involving an infinite number of attributes or ruling an
infinite number of individuals is derived from theoretical assumptions.
Hence, such a distribution seems to be peculiar to the hypothetical way of
learning from experience, that is to testing or to Bayesian estimation. On
the other hand, in estimating the composition of a population we cannot
make any specification of the population: we only work on what we are
observing. It is well known that physical quantities are always measured
up to a certain precision. That is every measurement ends with an interval
(r - 8, r + 8) where 8 represents the precision of the instrument used in
order to accomplish the measurement. Thus what we actually observe is
a finite number of individuals which can have only a finite number of
attributes.
Therefore in carrying out inductive inferences it seems necessary to con-
sider only families with a finite number of attributes. Not considering such
families and considering an infinite number of attributes gives rise to two

problems which do not seem to be well suited regarding inductive infer-
ences. First, we can no longer use probabilities in expressing prior infor-
mation. In this case the use of a distribution function only hides the need
to introduce densities. Secondly, as a matter of fact we can no longer
modify all the assumptions - the probability or density values we have
assigned to the attributes - which we have made. And to modify the prior
assessment is an essential feature of inductive inferences. Also considering
an infinite population takes us out of the realm of concrete facts. As a
consequence probabilities are no longer of use. We shall return to this
topic later.
To conclude the present section, we stress that our approach to laws is
purely inductive, that is in connection with the estimation of a distribution
function. Tests of significance and Bayesian estimations are points which
we shall not enter into, with the exception of a brief sketch of Jeffreys'
(Bayesian) method of testing.
4. In Hempel's opinion, his paradox is due to misguided intuition rather

than to a logical defect. As Hosiasson-Lindenbaum pointed out, the class
of non-black objects is much larger than that of raven-objects. Therefore
finding a non-black non-raven object actually increases the probabilities
of both generalizations, but by a very small amount. In other words, ob-
serving a black raven and a non-black non-raven object increases both
equivalent generalizations, but the increase due to the first observation is
much larger than that due to the second.
This solution was disregarded by Hempel in view of the logical structure
of· the language intended as a co-ordinate language. Hempel suggested
another solution, i.e. to specify the field of application of a law. However
he also discarded this solution because of the way in which laws are used
in physical theories.
We have clearly stated we have statistics in mind. Every statistical prob-
lem refers to a well specified population, i.e. to a well specified field of
application. In statistics the population is fixed before experimenting. No
statistician, aiming at determining the failure rate of a lot of washing-
machines measures the size of butterflies. Moreover the population is sup-
posed to have at most, and as a limiting case, countably many individuals.
Therefore we accept both solutions. Firstly, as was suggested by Hos-
38 D. COSTANTINI
iasson-Lindenbaum, we suppose that it is useless to observe butterflies in

order to determine the failure rate of washing-machines. Secondly, as was
suggested by Hempel, we suppose that for any law the field of application
is specified.
However, before discussing laws we must say something about the lan-
guage we use to formulate laws. L is a monadic first order language which,
beside the usual connectives, quantifiers and individual variables, contains
countably many individual constants at, ... , a" .... L consists of three
languages. Ll contains a finite number of monadic predicate constants Ql,
... , Qj, ... , Qk whose index set is N k. L2 contains countably many monadic
predicate constants whose index set is N. L3 contains a continuum ofmon-
adic predicate constants whose index set is R. A model is a sequence Z:
N -+ I, I is N k , N or R. Z is the set of all such sequences. (Z: Z(z) = J) is
the atomic proposition Qjai. Ert = (Qfli;j E 1. Eat = .u (Ert) is the set of
.eN
the atomic propositions. The set of molecular propositions EmoJ is the field
generated by Eat on Z. The set of proposition E is the u-field generated by
Eat on Z.
From the assumptions we have made, it follows that in "All Pare Q"
the predicate constant P serves merely to specify the population to which
the law applies. Hence, as the population is fixed before experimenting, a
deterministic law can be formulated as
(1)
This is not the only type of deterministic law we can consider. Propositions
of the following type are also deterministic laws
(2) .n -,Qai) n (. n Qai)

( .eNe .eN-Ne
in which Ne is the set of exceptions. When we cannot or do not want to

continue to state which members of the population are exceptions and we
limit our attention to the relative frequency, we have a statistical law. If
the population is finite and its member are N in number, then a law of this
type can be formulated as
(3) (Z : (Z)r~ = i/N), i E ~N

(Z)r~ is the relative frequency of individual with Q in the first N terms of

z.
It is easy to realize that in the case we are now examining, i.e. finite
population and two predicate constants Q and -, Q, the numbers 0, lin,
2/N, ... , I describe all possible laws. 0 and I describe deterministic laws
whereas the other N-l fractions describe statistical laws.
In an infinite population a statistical law can be formulated as
(4) n u n (Z: I (Z)r~ - p I < I/h), 0 ~ p ~ 1.

h m N>m
It is worth noting that in this case the number 1 does not refer only to the
deterministic law asserting that all individuals are Q. More explicitly, 1
represents a deterministic law as well as all statistical laws in which the
number of exceptions is not large enough, i.e. the relative frequency of
exceptions tends to zero as N grows without limit. Thus when we are
considering an infinite population and we decide to deal with it in fre-
quency terms, we may only consider the order of magnitude of exceptions,
i.e. we can no longer distinguish laws in which the relative frequency of
exceptions tends to zero as N -+ IX).
In the multinomial case, i.e. I = N k , for deterministic laws with or with-
out exceptions we can restate what we have already said. Statistical laws
can be formulated as
(5)
In the case in which there are countably many predicate constants, i.e. I
= N, statistical laws can be formulated as (5) with N instead of N k •
As we have seen, it is possible to pass from deterministic laws to statis-
tical laws with countably many attributes without any gap. Starting from
deterministic laws we have considered exceptions, i.e. the other attribute,
then a finite number of attributes and, finally, countably many attributes.
On the contrary we find a real gap passing from laws with countably many
attributes to laws with a continuum of attributes, i.e. I = R. A law of this
type can be stated as
(6) j(j), j E R
but (6) cannot be formulated in our symbolism. Naturally it is possible to

40 D. COSTANTINI
f
+00
introduce a density directly, i.e. a function f{j) ~ 0 such that f{j) dj

-00
= 1, but in a sense this introduction is arbitrary. Sometimes this arbi-

trariness can be overcome in that some densities can be reached as limiting
cases of laws of type (5). In these cases we again have a gradual way of
passing from deterministic laws to statistical laws with a continuum of
attributes. In any case it is a gradualness that must be proved every time.
Given that "mortal" means "dead before 500 years old", "All men are
mortal" is an example of (1). For Christians "Jesus Christ is the sole im-
mortal man" is an example of (2). "In Italy the relative frequency of death
before 20 years old is 0.02" is an example of (3). "The population of all
infinite washing-machines manufactured by a given firm is such that the
limit of those machines that break down before 1000 hours operation is
0.001" is an example of (4). The life tables are example oflaws of type (5)
referring to finite populations, i.e. the p/s are relative frequencies. "The
infinite population of all army corps are such that the relative frequency
of such corps having a given number of men being killed by horse-kicks
in an year tends to a limit" is an example of a statistical law with countably
many attributes. "The infinite population of all measurements of a quan-
tity is such that the relative frequency of whatever deviation from the mean
value of the quantity measured in terms of the standard deviation tends
to a limit" is an example of (6).
5. In the previous section starting from deterministic laws we reached sta-
tisticallaws with countably many attributes in a gradual way. Moreover
sometimes this gradualness holds also for statistical laws with a continuum
of attributes. At first sight, this suggests that there is something wrong in
giving deterministic laws a privileged status as was done by Hintikka6 •
However there are other and more sound arguments in favour of this
suggestion.
Limiting our attention to the binomial case, for the sake of simplicity,
it is easy to realize that (1) is a special case 0(4). Interpreting Qx as "the
x-th washing machine works (at least for 10,000 hours)", (1) asserts that
all washing-machines work. If we are estimating the failure probability of
a population of washing machines, why, as suggested by Hintikka, must
we privilege (1), i.e. assign it a finite probability? Why not, as suggested
by Spohn?, also privilege (2)? Or (4) with p = 1/2 as was suggested by

Essler8 ? And why not privilege any other value of p?
Our answer is that such probability assessments have no place neither
in inductive inferences nor in hypothetico-inductive inferences. As we al-
ready said, Jeffreys was the first to suggest priors with jumpS 9. However
his suggestion does not refer to the estimation of laws, but to tests of
significance. In section 3 we have noted that testing is a deductive proce-
dure. This does not mean that it is impossible to accomplish a test in a
Bayesian framework. On the contrary, this can be done in a very natural
way, by privileging the hypothesis to be tested, i.e. the null hypothesis, by
assigning it a finite prior probability. This is precisely what Jeffreys did.
As he explicitly affirms 10
The essential feature [of tests of significance] is that we express ignorance of whether the new
parameter is needed by taking half the prior probability for it as concentrated in the value
indicated by the null hypothesis, and distributing the other half over the range possible.
To assign a finite prior probability to a hypothesis is tantamount to priv-

ileging this hypothesis in order to falsify it. The assignment of finite prob-
ability to a hypothesis picked out from a continuum, transforms this hy-
pothesis into a sort of target to be shot at by experimental observations.
The moral of this is that prior with jumps related with problems of
estimation does not belong to the inductive tradition. Nonetheless even in
this tradition one can deal with laws, i.e. it is possible to obtain definite
suggestions about the validity of a law.
This can be done if we give up the queer pretence of dealing with infinite
populations using probabilities and return to the way in which such popu-
lations have always been dealt with, i.e. to densities. This is precisely what
we intend to do. But before we do so we consider another way of inter-
preting laws.
6. The way in which in section 4 we have dealt with laws can be qualified
as the universal interpretation of laws. However there is also another way
of interpreting laws which can be qualified as the individual interpretation
of laws. According to the first, a law refers to all individuals of a popu-
lation; according to the second, a law refers to each individual of a popu-
lation. As an example we consider the law "All men are mortal". Accord-
ing to the universal interpretation this law amounts to affirming that Adam
42 D. COSTANTINI
will die, Eve will die, Cain will die and so on listing all men in the past,
present and future. On the other hand, according to the individual inter-
pretation to affirm "All men are moral" is tantamount to affirming that
the next man we shall observe, irrespective of the fact that the observation
will take place here or there, now or later, will die. That is, for the universal
interpretation stating a law is the same as stating a characteristic of popu-
lation members considered as a whole; for the generic interpretation, stat-
ing a law is the same as stating a characteristic of a generic individual of
the population.
Undoubtedly the universal interpretation is that which one thinks of
first when speaking of laws. But it is not difficult to realize that sometimes
also the individual interpretation is used. But there is more. Statistical laws
are usually expressed as individual laws. The cumbersome way in which
at the end of section 4 we have formulated statistical laws makes clear that
handling the universal interPretation with these laws is not usual. In fact
one says: "The probability of failure before 1000 hours of a washing-ma-
chine is 0.001", "The probability of an army corp having a given number
of men killed by horse-kicks is given by the Poisson distribution" or "The
probability of a deviation from the mean is given by the Gauss distribu-
tion".
The individual interpretation of a deterministic law is
(7) for each i E N, Qfli
(1) and (7) pave the same probability only if we are dealing with a deter-
ministic law that we suppose to be true. In this case, for each i E N,
C(QfliIZ) is equal to CC~l QflilZ ). c(.I.) is a relative probability, since

both probabilities are equal to 1. However if for each i E N, C(QfliIZ)
= r, 0 ~r~ 1, it may happen that CC~l QflilZ) = O. Passing to

statistical laws, we have seen that the universal interpretation of such laws
amounts to stating a characteristic of all members of a population, or,
more precisely, to stating that the limit of the relative frequency of an
attribute Qj in the population is equal to Pj,j E I. On the other hand, the
individual interpretation amounts to stating that for each element of the
population, irrespective of what element we are considering, the proba-
bility of having Qj is Pi> j E I, that is to state that
for each i E N, C(QfliIZ) = Pj.
For the sake of simplicity, we have considered a priori probabilities of

laws, i.e. probabilities given the evidence Z describing the void sample.
The matter does not change if we consider posterior probabilities, i.e.
probabilities whose evidences are not void.
If we interpret laws in the universal way the sole way to assign, con-
sidering an infinite population, a finite posterior probability to a law, is to
assign it a finite prior probability. However if we want to avoid this un-
usual method of estimation we must discard probabilities in favour of
densities. More precisely, probabilities work only if we consider finite
populations; if we want to take into account infinite populations we must
pass to densities, i.e. we must take into account the limit, if it exists, of the
probability we are considering. Of course, the convergence we must con-
sider is the convergence in distribution, i.e. the limit of the sequence of
distribution functions which arises as the number of individuals tends to
infinity.
7. In this section we state formally the two interpretations of laws. We

restrict our attention to statistical laws as deterministic laws are only a
particular case of those laws.
Firstly, we mention some symbols. sb = (0,0, ... , 1, ... ,0), the 1 is in
the j-th place. s\f> = (sb; j E Nk) 0Q is the set of non-negative rational
°
numbers and 0Qk is the set of all k-tuples of numbers of 0Q. For any n
> 0, °Qn = (x; X E 0Q and for some m, ~ m ~ n, x = min). For any
n > 0, 0Q~ = (x ; X E 0Qk where for every Xj E x, j E N1c, Xj E °Qn and
k
L Xj = 1).
j=l
A finite I-instance law is a pair «(Pj;j E Nk),S~l»

A finite n-instance law is a pair «(Pj;j E Nq), 0«) where q =("+~-1).
sb can be interpreted as a degenerating distribution, namely the distri-
bution of a sample of size 1 whose unique individual bears the j-th attribute
of the family. Or, which is the same, sb can be interpreted as the distri-
bution of a random variable being equal to Qj with probability 1. Thus a
finite I-instance law is an assessment of probability to a set of degenerating
distributions. A finite n-instance law is an assessment of probability to a
set of distributions, namely to all distributions that can be originated by
the individuals of a sample of size n. When n -+ 00, 0Q~ becomes °Rk, i.e.
44 D. COSTANTINI
the set of all k-tuples of non-negative real numbers with sum equal to I
and the related probability distribution become, if the sequence of distri-
bution functions converges, a k-dimensional distribution function. Ob-
viously, according to the universal interpretation of laws, a law is an n-
instance law with n -+ 00, whereas according to the individual interpre-
tation, a law is an I-instance law.
Obviously it is possible to define the concepts of denumerable I-instance
and n-instance laws as well as those of continuous I-instance and n-in-
stance laws. We do not deal with these laws as they are used in Bayesian
estimations and tests of significance.
To conclude this section we notice that the considered concepts are de-
fined irrespective of the evidence available. When the evidence is void, then
we are dealing with an a priori law. If the evidence is not void, we are
dealing with a posterior law. Finally, if for allj E Nk , Pj = Sj/s, i.e. Pj is
the relative frequency of the attribute Qj in the sample, then we are dealing
with an empirica! law.
8. In order to avoid unnecessary complications, we consider the binomial

case. Let us suppose that the population we are observing contains a finite
number of individuals that can have only two attributes Q and, Q and
that we have observed a sample of S individuals of which S1. 0 ~ S1 ~ S,
are Q and S2 = S - S1 are 'Q.
The problem we have in mind is to estimate the law ruling the individ-
uals of the population from which the sample has been drawn. The first
step in performing an estimation is to fix the probability-function to be
used. We suppose that C, the probability-function we are using, is regular,
symmetric and invariant, i.e. that C is a A.-function l l .
If we are estimating a law as a I-instance law, then the pair is
• • (1) _ Sj + ')IjA.
(8) ((Pj,] E N 2), So ), where Pj - l
S + I\.
(8) is obtained from the a priori law

((Pj; j EN 2), s1>1\ where Pj = ')Ij
and the empirical law

((Pj; j EN 2), s1>1\ where Pj = Sj/s.
The cardinality of the population has no role in estimating a I-instance

law.
On the other hand, the cardinality becomes important if we are esti-
mating an n-instance law. Let us suppose first that the cardinality of the
population is N + s. In this case we estimate the N-instance law as
«Pj;j E 0NN); 0Q~)
where
The estimation is performed using (8) and the multiplicative axiom. It is

worth-while noting that, irrespective of the value Pj of (8), if N is suffi-
ciently large, than the p/s of (9) are all small. But this does not mean that
all p/s are on a par. If we make a suitable choice of A. there can also be
large differences among the p/s. A few examples will clarify the matter
better than many words. Suppose
Y1 = Y2 = 1/2, s = S1 = 10
and that the n-instance law we are considering is relative to a population
of 1010 individuals. The following table shows the probabilities according
to various values of A. of the population whose individuals bear all Q, let
us denote this population with S(o» and the ratio between these probabil-
ities and that of the population whose individuals bear all Q but one, let
us denote this population with S(1). Notice that all other possible popu-
lations S(i)' 2 ~ i ~ 1010, have probabilities less than that of S(1).
TABLE I
1 probability ratio
2 0.0108 1.01
0.5 0.3122 2
0.1 0.7879 20
0.01 0.9764 200
What Table I shows is plain and has been already clearly noted by Essler 12 •
It is sufficient to choose a low value of A. in order to obtain a large posterior
probability of the deterministic n-instance law not falsified by evidence.
46 D. COSTANTINI
This is already true for a small amount of evidence and a comparatively

large population. Moreover all other probabilities become very small with
regard to that of the not falsified deterministic law.
9. Let us now suppose that the population we are considering contains

countably many individuals. In this case we have a continuum of possible
populations. Each population is represented by a real number r E [0,1],
namely the limit of the relative frequency of individuals bearing Q.
To take account of this case we consider an n-instance law with n -+
00. As is well-known, the probability of such laws, deterministic or not,
tends towards zero irrespective of the sample we have observed. This re-
sult, with special attention to deterministic laws, i.e. to the paradox of null
probability, has been seen as the end of the road for inductive logic, or,
at least, as the mark of the inadequacy of the A.-system. This is not our
opinion. All this result says is that probabilities do not account for infinite
populations. But this is not at all new as it has been known since De
Moivre's time. In dealing with an infinite number of individuals we must
take densities into account.
To this end we consider the probabilities of relative frequencies of Q in
the possible populations. That is, we consider the random variable whose
values are in 0Q; = (i/n; i E ON,,). Let X be this random variable, C,,(X)
its distribution and F,,(x) its distribution function , i.e.
F,,(x) = L C,,(i/n)
i/,,<%
If n -+ 00, then C,,(i/n) -+ 0 for every i E ON". But if we consider F,,(x) we

see that:
(i) by symmetry, F,,(x) converge towards a unique function F(x);
(ii) by invariance, F(x) is a beta distribution, i.e. a distribution function
whose density is
The convergence towards zero of the probability of each law is a very

natural consequence of the fact that we are performing an estimation re-
lated to a continuum of possible laws. To focus on the paradox of null
probability is the best way to disregard the. most interesting side of the
A-continuum. To realize this it is useful to notice what happens if we take

into account the sample observed. In this case (16) becomes
f(x I S1. S2) = (B(11A + S1. 12A + S2»-1
Again the values of densities at various x are not all on a par. In this case
also a few examples will clarify the matter better than many words. We
suppose that we have observed a sample of 1000 individuals all bearing
the attribute Q. We pose
11 = 12 = 1/2 and A = 1/100
We note, firstly, thatf(xIIOOO,O) is an L-curve which has the maximum at
x = 0. That is, the laws which will most likely be true are all laws in which
the limits of the relative frequencies of individuals with Q are 0. As we
°
have already said, it is impossible in an infinite population to characterize
the deterministic laws with a number, or I. Moreover the curve is sharply
peaked. The following table gives the probabilities that the distribution
function corresponding to f(xllOOO,O) assigns to the intervals of the first
column.
TABLE II
interval probability
[0,0.001] 0.99989
[0,0.0001] 0.9990
[0,0.00001] 0.9979
If we suppose that A = 1/1000, the curve is as before but even more

sharpened as is shown by
TABLE III
interval probability
[0,0.0001] 0.999989
[0,0.00001] 0.99979
48 D. COSTANTINI
Once again what the tables show is plain. It is sufficient to choose a low
value of A to obtain precise suggestions about the validity of a law.
10. The A-system works in all cases. It supplies, by the I-instance law, the
probability that any new individual bears an attribute. If the population
is finite, it supplies, by the n-instance law, all probabilities of all possible
populations consistent with evidence. If we suppose that we are dealing
with an infinite population, it supplies the densities of all populations.
In our opinion, the concept of the n-instance law is the most important
concept which we have considered in the present paper. This conviction
is derived from the increasing importance in science of predictive inference.
In the first place, this is the case of statistics, moreover it must not be
forgotten that elementary-particle statistics may also be obtained through
predictive inferences 13 . However, in a precise sense, Camap was right in
asserting that "The most important special case of the predictive inference
is the singular predictive inference"14. In fact this inference amounts to
determining a I-instance law, the first step in determining n-instance laws
and, in case of convergence, density functions.
Regarding limit distributions we note that the existence of densities, i.e.
of laws with a continuum of attributes, does not contradict what we have
already said in section 3. Laws of type (6) may well arise in the course of
the estimation of laws. However their rise is due to the introduction of
hypothetical assumptions about the population. The first assumption re-
gards the growth without limit of individuals. Other assumptions regard
the change in the unit of measure, that originate a continuous scale, and
the (implicit) claim of no limitation in the precision of the measurement.
This is the case we have considered in order to reach the Beta distribution
(I/n to be measured with any degree of precision). This is also the case of
the normal distributions (1/ In 1112 to be measured with any degree of
precision). Finally, another assumption imposes a suitable behaviour to
the population when it grows without limit. For example, the assumption
of the validity at infinity of the symmetry condition that we have also
implicitly made. To this regard it is often forgotten that without the as-
sumption that the phenomenon we are observing leads to an infinitely long
symmetric sequence, de Finetti's theorem does not hold.In our opinion,
these arguments are strong clues in favour of the claim regarding the fin-
itary nature of every inductive inference as has been expressed by Essler1 5:
PROBABILITY AND LAWS. 49
The presumption that the domains [...] are denumerably infinite is wholly unconvincing. We
mortals have always encountered but finite numbers of [...] things and will in the future
continue to be confronted with mere finite numbers of them.
An assertion that we fully share.
NOTES AND REFERENCES
* The author wishes to thank E. Casari, M. L. Dalla Chiara, M. C. Galavotti, T. A. F.

Kuipers and W. Spohn for much helpful discussion and correspondence.
1 Hempel, C. G., 'Studies in the Logic of Confirmation', Mind 54 (1945), 1-26,97-120.
2 Camap, R., 'On Inductive Logic', Philosophy of Science 12 (1945), 72-97.
3 Hosiasson-Lindenbaum, J., 'On Confirmation', The Journal of Symbolic Logic 5 (1940),
133-148.
4 Toraldo di Francia, G., 'Induction in Physics', Rivista del Nuovo Cimento 4 (1974), 144-
165.
5 It is worthwhile to note that also T. S. Fergusson, 'A Bayesian Analysis of Some Non-
parametric Problems', The Annals of Statistics 1 (1973), 209-230, has faced the same problem
essentially reaching the same conclusions.
6 Hintikka, J., 'A Two-dimensional Continuum ofInductive Methods', Aspects of Inductive
Logic (Hintikka and Suppes eds.), North-Holland, Amsterdam, 1966.
7 Spohn, W., 'Analogy and Inductive Logic: A Note on Niiniluoto', Erkenntnis 16 (1981),
35-52.
8 Essler, W. K., 'The Camapian Tolerance in the Foundations of Probability', Probability,
Statistics and Inductive Logic (Agazzi and Costantini, eds.) Special issue of Epistemologia
(forthcoming).
9 Jeffreys, H., Theory of Probability (1939), Oxford Univ. Press, Oxford 1961.
10 Ibid, p. 249.
11 Costantini, D., 'The Relevance Quotient', Erkenntnis 14 (1979), 149-157.
12 Essler, K. W., 'Induktion und Induktive Logik, II', Statistica 44 (1984),2.
13 Costantini, D., Galavotti, M. C., Rosa, R., 'A set of Ground Hypotheses for Elementary
Particle Statistics', II Nuovo Cimento 74 B (1983), 151-158.
14 Camap, R., The Logical Foundations of Probability (1950), The Univ. of Chicago Press,
Chicago, 1963, p. 568.

15 Essler, W. K., 'Hintikka versus Camap', Erkenntnis 9 (1975), 229-233.
Received 4 April 1984.
Universita degli Studi di Bologna

Dipartimento di Scienze Statistiche 'Paolo Fortunati'
Via Belle Arti 41
40126 Bologna
Italy
EVANDRO AGAZZI
COMMENSURABILITY, INCOMMENSURABILITY, AND

CUMULATIVITY IN SCIENTIFIC KNOWLEDGE
ABSTRACT. Until the middle of the present century it was a commonly accepted opinion
that theory change in science was the expression of cumulative progress consisting in the
acquisition of new truths and the elimination of old errors. Logical empiricists developed
this idea through a deductive model, saying that a theory T superseding a theory T must be
able logically to explain whatever Texplained and something more as well. Popper too shared
this model, but stressed that T explains the old known facts in its own new way. The further
pursual of this line quickly led to the thesis of the non-<:omparability or incommensurability
of theories: if T and T are different, then the very concepts which have the same denomi-
nation in both actually have different meanings; in such a way any sentence whatever has
different meanings in T and in T and cannot serve to compare them. Owing to this, the
deductive model was abandoned as a tool for understanding theory change and scientific
progress, and other models were proposed by people such as Lakatos, Kuhn, Feyerabend,
Sneed and Stegmiiller. The common feature of all these new positions may be seen in the
claim that no possibility exists of interpreting theory change in terms of the cumulative
acquisition of truth. It seems to us that the older and the newer positions are one-sided, and,
in order to eliminate their respective shortcomings, we propose to interpret theory change in
a new way.
The starting point consists in recognizing that every scientific discipline singles out its
specific domain of objects by selecting a few specific predicates for its discourse. Some of
these predicates must be operational (that is, directly bound to testing operations) and they
determine the objects of the theory concerned. In the case of a transition from T to T, we
must consider whether or not the operational predicates remain unchanged, in the sense of
being still related to the same operations. If they do not change in their relation to operations,
then T and T are comparable (and may sometimes appear as compatible, sometimes as
incompatible). If the operational predicates are not all identical in Tand T, the two theories
show a rather high degree of incommensurability, and this happens because they do not refer
to the same objects. Theory change means in this case change of objects. But now we can see
that even incommensurability is compatible with progress conceived as the accumulation of
truth. Indeed, T and T remain true about their respective objects (T does not disprove 1),
and the global amount of truth acquired is increased.
In other words, scientific progress does not consist in a purely logical relationship between
theories, and moreover it is not linear. Yet it exists and may even be interpreted as an
accumulation of truth, provided we do not forget that every scientific theory is true only
about its own specific objects.
It may be pointed out that the solution advocated here relies upon a limitation of the
theory-Iadeness of scientific concepts, which involves a reconsideration of their semantic
status and a new approach to the question of 'theoretical concepts'. First of all, the feature
of being theoretical is attributed to a concept not absolutely, but relatively, yet in a sense
different from Sneeds's: indeed every theory is basically characterized by its 'operational'
concepts, and the non-operational are said to be 'theoretical', this distinction clearly de-
pending on every particular theory. For the operational concepts it happens that their mean-
Erkenntnis 22 (1985) 51-77. 0165-{)106j85j0220--0051 $02.70

52 EVANDRO AGAZZI
ing splits into two parts: one I call the 'referential' component and the other the 'contextual'
component. It is only this second which is actually theory-laden, while the first represents a
'stable core' of meaning, which may remain invariant within different theories and in such
a way allow theory comparison. Hence it is possible to admit that every scientific concept is
theory-laden, but at the same time to admit as well that this feature is only partial in the
case of the operational concepts. This amounts to c1arning that the 'sentential view' of scien-
tific theories may be kept, but that the empirical side also preserves its full role and, in
particular, is essential in every question concerning theory change, theory comparison and
evaluation of scientific progress.
FROM A STRUCTURALIST TO A HISTORICIST CONSIDERATION OF

SCIENCE
It is certainly not difficult to recongnize that a very important and dis-

tinctive turn has occurred in the philosophy of science during the second
half of the present century. We might point out its main feature as a tran-
sition from a static to a dynamic way of considering the nature of science,
which finds its most typical expression in the fact that the attention of
philosophers of science has become more and more concentrated on the
problem of theory change. This trend may also be interpreted as a tran-
sition from a structuralist to a historicist approach to science, but this is
rather a second aspect of it, running parallel to the first. As a matter of
fact, one may say that the Popperian philosophy of science, by laying so
much stress on falsification, had already insisted on theory change. How-
ever it nevertheless remained very far from an historicist way of thinking,
as no truly 'historical' element was considered of relevance to understand-
ing the structure of scientific theories (which Popper interpreted in much
the same way as did the logical empiricists) or for understanding, let alone
for explaining, theory change. Even Lakatos, with his proposed ' rational
reconstruction' of the history of science, was inclined towards a structur-
alist interpretation of history, rather than towards a historicization of
scientific structures. The same is also true, at least to a large extent, with
regard to Kuhn, who takes more a sociological, rather than a truly his-
torical outlook in the consideration of the dynamics of science. Of course,
this is completely true in the case of scholars such as Sneed and Stegmiiller,
whose special effort in recent years has been that of showing how theory
change, or theory dynamics, may be totally embedded in a 'structuralist'
view of theories. A historicist outlook, on the other hand, is to be found
explicitly in the works of people such as Hiibner, for whom the problem
SCIENTIFIC KNOWLEDGE 53
of theory change is simply an aspect of the much wider dependence of

science on the whole of its historical context, which essentially affects the
cognitive features of the different scientific theories and systems and not
only their environmental and social surroundings. Perhaps the most con-
vincing evidence for the spreading of this historicist approach is given by
the current interplay between history of science and philosophy of science,
which is occurring throughout the world, and which entails in particular
an increasing estimation of the importance of history of science in doing
reliable philosophy of science.
A result of the development of the historicist trend, in both of its aspects,
has been the increasing attention that has been paid to the problem of
'progress' in science, which has resulted in a wide discussion where dif-
ferent positions emerge. Unfortunately this discussion has .not produced
much clarification yet, owing not only to the complexity of the concept of
progress itself, but especially to the fact that several different planes are
interfering with one another. It would be useful to distinguish these planes
- not in order to keep them separated, but rather with a view to analyzing
them closely and with the appropriate methodological tools. Just to give
an idea of this complexity, let us mention some of he most important
among these planes, by listing a number of problems or questions to which
they correspond:
(a) What is science?
(b) What is a scientific theory?
(c) Why is a theory accepted?
(d) Why is a theory abandoned or dismissed?
(e) Can theory change be interpreted as progress, and how?
(f) What does progress mean in addition to simple change?
(g) To what extent does progress entail comparison?
(h) Can such a comparison be understood as theory comparison?
(i) If so, what are the criteria of comparison and what are their possible
limitations?
G) Is progress necessarily cumulative or not?
(k) Is there a non-linear progress?
(1) What are the relations between the idea of progress and the notion
of truth?
(m) What, if any, are the ontological implications of progress in science?
As we said, this list is by no means complete, especially because there
54 EVANDRO AGAZZI
are many other questions, which may be considered either as subquestions

to the listed ones or even as independent questions of their own. For ex-
ample, question (c) has certainly important links with the following prob-
lems: that of the genesis of theories, that of their dependence on the his-
torical context, that of their philosophical presuppositions, question (1)
involves, among other things, the problems of the denotation and meaning
of concepts and sentences; question (m) gives rise to the debate on scientific
realism; the problem of observational and theoretical terms is involved in
more than one of the listed questions, and so on.
Our proposal here is certainly not that of analyzing all these items: we
shall confine ourselves to a short discussion of a problematic area spanning
more or less from question (e) to question (1) on our list, while other points
might enter our considerations only occasionally and mainly as auxiliary
remarks.
SOME RESTRICTIONS OF OUR ANALYSIS
In this spirit we may begin by saying that we are of the opinion that
question (a) admits of a composite answer, according to which science is,
from a certain point of view, a human activity and, from another point of
view, a system of knowledge. The first aspect implies that science, like all
human activities, is 'performed' in view of many personal and social goals
(and this legitimates debates such as those on the 'neutrality' and re-
sponsibility of science, which also affect the notion of scientific 'progress'
and give it a very special meaning). On the other hand, it must be recog-
nized that the most specific of these goals is that of producing a body of
reliable knowledge in a great variety of fields, and this leads to the second
qualification of science we mentioned (science as a system of knowledge),
which is practically the one traditionally considered by the philosophy of
science (at least when professionally understood). Without underestimat-
ing the legitimacy and the importance of the first qualification, we want
to make clear that science will be considered here only under the second
viewpoint, as the title of this paper explicitly indicates. Yet this delimita-
tion is still too vague, as science does not include all kinds of knowledge,
but only those which meet the requirements of what we could call 'objec-
tivity' and 'rigour'. We shall here feel free from providing an exact speci-
fication of these two requirements of science (we have provided it else-
where) and we shall simply say that their fulfilment obtains at best in what
are usually called 'scientific theories'. For that reason we accept to adopt
a (conscious) limitation, i.e. to take into consideration scientific knowledge
only as far as it is expressed in scientific theories, in this way leaving aside
all those forms of accidental, scattered and unsystematic knowledge which
certainly deserve consideration when a comprehensive discourse on scien-
tific progress is made. A further limitation appears advisable for practical
reasons: we shall concentrate only on 'empirical theories', i.e. on those in
which sense data (of a very special kind) are considered to provide us with
substantial knowledge; and we shall ignore 'formal theories', not because
we are of the opinion that they do not contain knowledge, but because
they require quite a peculiar kind of study: the problem with which we are
concerned is already sufficiently complicated in the domain of empirical
theories, and its solution does not depend on its being treated together
with the question of 'formal knowledge'. It would be too optimistic and
perhaps naive to believe that after these important delimitations our field
of inquiry should be determined unambiguously and with total clarity.
Indeed the question: 'What is a scientific theory' is by no means a simple
and pacific one for, as is well known, the traditional 'statement view of
theories' has been stronly challenged in the past few years, but has not
been superseded by any other view which could be credited with a general
acceptance. This is, by the way, the reason why it would not really prove
profitable to explicitly discuss this question here, although some element
of this discussion will appear in our subsequent presentation.
The problems of why a theory is accepted or abandoned are among the
most frequently debated in the literature of recent years, but they are also
those in which it is possible to see how much confusion has been caused
by the interference and the uncontrolled exchange of different planes of
discussion. These problems may indeed be discussed on a factual plane;
they may been seen under a psychological or a sociological viewpoint; they
involve epistemic attitudes; they include logical aspects, and they have a
very important pragmatic component. Unfortunately, it too often happens
that different scholars lay stress on one single plane as if it were the only
which matters, trying to discredit the other approaches. What they are
actually able to do is to show the limitations of some of the other ap-
proaches (if this were to be taken as an all-explaining tool), but they are
not fair enough as to recognize the aspects of the problem which are rather
56 EVANDRO AGAZZI
well accounted for by the other approaches and, moreover, they seem
unaware of the shortcomings that their own approach shows, if it too is
taken as all-explanatory. An advantage of our further discussion will be
that we shall certainly consider the question of scientific progress in con-
nection (indeed in strict connection) with the question of theory change,
but without having to answer the question 'why' theory change has oc-
curred: it will be enough for our purposes to see 'that' it has occurred, and
to examine the situation which is determined within scientific knowledge
by virtue of this change.
THE NOTION OF PROGRESS
It has been typical of our Western culture to believe that the situation
emerging from this change always be the manifestation of 'progress', and
this concept has been endowed with a rich display of shades of meaning
the analysis of which we must leave to the historian of ideas. Indeed, owing
to the thematic delimitation we have introduced here, we shall simply be
concerned with the idea of progress as applied to scientific knowledge,
which means, as we have already said, to knowledge organized in the form
of scientific theories. Under this restriction the problem of scientific pro-
gress seems to become very simple, as everybody seems ready to accept
that scientific progress may perhaps be questionable if we consider all sides
of science, and especially those which concern its social impact, but its
cognitive side seems to be 'progressive' in a very patent and undeniable
sense.
This first impression has to be treated with caution: however when we
consider, as we have accepted to do, scientific knowledge as it is expressed
in scientific theories, what is really evident is that in every discipline there
is a succession of theories in time, that the new are 'different' from the old,
that a 'change' occurs, but this does not by itself imply 'progress'. In order
for change to be considered as progress, the factual ascertainment of its
having occurred must be accompanied by a value-judgment of some sort,
enabling us to claim that the new situation is 'better' than the previous
one. The difficulty lies precisely in the determination of this 'better'. Let
us note that this difficulty does not vanish when we enter the domain of
knowledge (which might seem to have dispensed with this kind of question,
being, as is usually said, 'value-free'). Indeed, it is precisely when we con-
sider scientific knowledge that we are very naturally inclined to share the
opinion that we know 'more' and 'better' than people belonging to older
generations. Now even the exact determination of this 'more' is by no
means unproblematic, for under its seemingly purely quantitative cate-
gorization it actually conceals much more, i.e. the idea of a 'selective'
process, consisting at least in the elimination of old errors and the continual
addition of new truths. The idea of the 'better' is then more or less vaguely
understood in ter.ms of deeper insight or of better correlation between the
different truths, and its expression is perhaps given by the familiar image
of science as a great and complex building, to the erection of which every
single scientist contributes by adding his own brick, while every generation
of scientists contributes by adding something like a new floor with beau-
tiful rooms. Roughly speaking, there is some kind of common consensus
in the appreciation of the development of scientific knowledge not simply
as change, but rather as linear and cumulative progress.
THE DEDUCTIVE MODEL OF SCIENTIFIC PROGRESS
This commonsense idea became very naturally the tacit guideline for those
people who first elaborated the conception that scientific knowledge ac-
tually consists of scientific theories, and who gave a characterization of
scientific theory which has remained valid and 'classical' for quite a long
time. According to this view, which was mainly developed in the specu-
lations of the logical empiricist philosophy of science beginning with the
Vienna Circle and continuing in particular with the Anglo-Saxon analyti-
cal philosophy of our century, scientific theories are conceived of as being
systems of logically connected sentences having the task of providing an
explanation of observed facts. In order to reach this goal, sentences were
divided into two classes: those which were purely descriptive of observed
facts (the explananda), and those which were introduced as hypotheses
from which formal deductive chains could be started which would end with
the factual sentences. This very familiar picture may be classified as ex-
pressing at the same time the so-called 'sentential view' of theories and the
'deductive model' of scientific explanation. It is also quite understandable
that these two patterns should be adopted when the problem of theory
change came to be investigated within the said school, and that they should
also be used in order to show when and why theory change occurs, and
58 EVANDRO AGAZZI
how this may be interpreted as progress. However, these scholars remained

unaware of some subtle but effective differences which manifest themselves
when relations 'between' different theories are considered, rather than re-
lations 'within' a single theory. Indeed, if a theory is essentially conceived
as being a system of sentences (which, as we shall see, is by no means so
naive and old fashioned as many believe to-day), it seems very appropriate
to consider the relations between the sentences which make up a theory
as being logical in nature, as few other kinds of relations might be seen
between sentences, and they would certainly not be very significant from
the point of view of expressing knowledge. Furthermore, the most typical
and best known logical relations between sentences are surely the deductive
ones.
What is questionable, however, is whether a model which is suitable for
expressing the relationships 'within' a system of sentences can be correctly
extended so as to express the relationships among the whole systems of
sentences that are scientific theories, when globally understood. The logical
empiricists, and also several philosophers of science who cannot be classi-
fied among them, seem to have too hastily accepted the legitimacy of such
an extension. This has ultimately led to a serious impasse, the cause of
which has been seen by many to lie in the 'statement view' of theories, but
which, if better understood, is rather to be identified with the generaliz-
ation of the 'deductive model' to theories as wholes, rather than limiting
it simply to the expression of relations between sentences. Even if we
accept theories to be systems of sentences, we cannot claim either that
theories themselves are sentences, nor that their mutual relations may be
'reduced' to those existing between their sentences, although the consider-
ation of these relationships may be of interest in several respects.
Let us now briefly sketch how the deductive model has been typically
applied in order to account for theory change, and especially in order to
interpret it as 'progress': what we are going to say is so well known, that
we shall confine ourselves to very few indications. 1 According to the said
model, the transition from an existing theory T to a succeeding theory T
could be seen as legitimate and implying progress mainly in two ways. The
first was the possibility of considering the new theory T as including the
old theory T as a subtheory (which means, in short, that the axioms or
the general laws of Tmay be proved as theorems of T). Even this minimal
condition of a syntactic nature, however, reveals itself as much too de-
manding, for it further implies that the primitive notions of T be definable

in terms of the primitive notions of T and the actual fulfilment of these
requirements shows that such a change, by which progress is seen as an
'embedding', actually occurs very seldom. Let us stress, however, that this
first idea was the most consistent application of the deductive model capa-
ble of preserving the global character of theories: in fact, one can say that
the underlying hope was that the old theory, as a whole, could become a
proper part of the new theory and that this would have been a real relation
between systems and not between their elements (as is the set-theoretic
relation of inclusion if compared with that of membership). The failure of
this hope was simply the indication that such an 'inclusion' can hardly be
interpreted as logical dependence or deducibility.
This kind of relation between theories being patently too optimistic, a
second less engaging one was proposed, which tries formally to express
the familiar and commonsense opinion according to which a new theory
T has the right to supersede an old theory T if T is able to explain all the
facts which T was able to explain as well as some facts that T was unable
to explain. This second version of the 'deductive model' is indeed rather
different from the first, as it no longer implies a kind of relation between
theories considered 'as a whole', but tries to compare theories simply by
considering their deductive power with regard to single sentences. If we
indicate by e any empirical sentence belonging to T and bye' any empirical
sentence belonging to T, the two variants of the deductive model can be
presented as follows:
(A) TI- T
(B) "'Ie (T I- e ~ T I- e) /\ 3e' (T I- e' /\ T 'r-f e')
Let us remark that by writing T'r-f e', i.e. by saying that e' is 'not de-
ducible' from T, we actually include two different subcases: the one in
which the negation of e' would be deducible from T, and the one in which
neither e' nor its negation are deducible from T. It is because of this
double possibility that the variant (B) was also accepted by people such
as Popper, who could not accept an intertheoretic relation of the kind of
(A). Indeed, according to his view, a new theory T supersedes an old
theory T because it emerges from a falsification of T, which means that T
is able to explain an empirical fact which goes against T (and which does
not simply lie outside T). In such a way T and T are seen as imcompatible,
60 EVANDRO AGAZZI
and this excludes any possibility of subtheory relation, but the variant (B)
may still be accepted in the sense of its first subcase, as we pointed out
above. However, owing to the incompatibility of T and T, it was necessary
to point out that T must be able to explain in its own new way and ac-
cording to its own new point of view the explananda with respect to which
T was already successful, and in addition new explananda with respect to
which T was not successful and even falsified.
This general idea was further developed by Popper, who introduced
some criteria for measuring the 'truth content' of different theories, in
order to establish the progressive character of T with respect to T, and
also proposed his theory of verisimilitude or approximation to truth. It lies
outside the scope of this paper to discuss these controversial theses of
Popper: we only want to note that Popper shared with the logical empir-
icist approach two essential tenets. The first is the conception of theories
as deductive systems intended to give an explanation of facts in the sense
described above; the second was that comparison between theories, and
thus the problem of justifying and interpreting theory change, was to be
approached according to a deductive model focused on the relation of
logical deducibility between the axioms of the theories and some single
sentences belonging to them.
It is exactly the deductive model which was later recognized as fully
inadequate to account for scientific progress, and its crisis was generally
interpreted as a refutation of the 'sentential view' of theories as such. We
are of the opinion that the question of the sentential view of theories re-
quires a different kind of discussion, but we shall postpone this point now.
THEORY-LADEN NESS AND INCOMPARABILITY
The Popperian conception contained a delicate point which eventually led

to a complete revision of the entire question. Indeed, when Popper said
that T must be able to account for the explananda of T in its own new
way and according to its own new point of view, he was admitting by this
very fact that a change of meaning occurs when we pass from T to T. But
then scheme (B) above cannot be preserved any longer, for it would be a
logical mistake to overlook this change of meaning in a formal
deduction. In other words, it is a mistake to assume that T and T are able
to explain the same explanandum e, because e means two different things
when it is considered as a sentence of T and when it is considered as a

sentence of T. This is due to the fact that no concept in a theory has an
independent meaning, its meaning always being context-dependent (or
theory-laden, as it became customary to say sometime later). The conse-
quence is that two different theories cannot really be compared: they are
imcommensurable, as Kuhn and Feyerabend in particular have so fre-
quently stressed. It is also well known that this incommensurability or
incomparability was almost automatically interpreted as an impossibility
of speaking of progress in science in any correct sense, for in order to
speak of progress, as something different from pure change, we need com-
parison.
It lies outside the scope of this paper to investigate how people such as
Lakatos, Kuhn and Feyerabend proposed alternative views for interpret-
ing theory change. We need only note that the consequence of their crit-
icism was a generalized disInissal of the deductive model presented above
and, more generally, the denial of the idea that inter-theoretic relations
may correctly be understood as logical relations at all. This is so whether
they be relations of logical consequence and compatibility or of logical
opposition and incompatibility, for both presuppose the condition of se-
mantical uniformity which does not hold, owing to the context--<lependen-
ce of the meanings involved. A more general consequence of this crisis has
been the disInission of the conception that scientific theories are systems
of sentences or statements (because, after all, the relationships between
statements cannot help being of a logical character in the last analysis).
Along this line of thought a non-statement view of scientific theories
(which was able nevertheless to preserve all the features of the formal
rigour which are traditional in the professional philosophy of science) has
been developed especially by people such as Sneed and Stegmiiller, who
have also proposed new ways for accounting for theory change. However
we are not going to discuss these conceptions here. 2
One main purpose of this brief historical account was to show the variety
of perspectives involved in this succession of approaches and models,
which undoubtedly leaves us with a feeling of uneasiness: indeed we cannot
help recognizing that the criticisms addressed to some older views are per-
tinent, but we also feel on the other hand that these views contained some
ligitimate claims which do not appear as sufficiently accounted for within
the new models, in spite of the conceptual advancement they entail in other
62 EVANDRO AGAZZI
respects. We may venture to say that the older and the newer conceptions
are all rather one-sided, in the sense that they correctly capture some fea-
ture of scientific theories and of scientific change, but emphasize them in
such a way as to disregard or at least underestimate other no less correct
and essential features.
THE ONE-SIDEDNESS OF THE EXAMINED POSITIONS
What we want to say can be illustrated by means of some simple examples.

Let us consider for instance the large discredit given to the statement view
of theories: this certainly appears as legitimate to the extent that it ex-
presses a reaction against that purely linguistic consideration of science
which was typical of the logical empiricist school, and in this sense it cor-
rectly stresses that scientific theories are not just systems of statements.
But this correct perception cannot prevent us from recognizing that scien-
tific theories are certainly also systems of statements, for scientific knowl-
edge cannot help being formulated knowledge, and this necessarily implies
the use of sentences; also, as a matter of fact, we must be able to recognize
what a theory says in order to discuss it, to put it to test, to obtain from
it predictions, and so on. When we mentioned this question above we said
that the dismission of the statement view of theories had probably resulted
from the criticism addressed against the 'deductive model' of scientific
theories, and this might give the impression that the statement view could
perhaps be preserved, provided we dismiss only the deductive model. Yet
this conviction would be wrong, for we must recognize (unless we advocate
a purely instrumentalist view of science) that one of the most specific aims
of scientific theories is that of providing rational understanding and con-
sistent harmonization of what is factually known within a certain domain
of research. But this is rather satisfactorily expressed by the notion of
'scientific explanation' such as it has been codified, but certainly not in-
vented, by the logical empiricists; and this explanation is convincingly ex-
pressed in the form of a logical deduction. It follows that a complete dis-
mission of the deductive model would imply that the concept of scientific
explanation be dismissed as well, or at least that it be replaced by a really
adequate counterpart. However, after having recognized the great impor-
tance of the deductive aspect of scientific theories, which corresponds to
the great role that logic has to play in them, we must remain able to
perceive that this role is not all--embracing and that experimental evidence
e.g. plays a no less important role in science. Then, however, we are im-
mediately confronted with the difficulties of the complex relations that
exist between these two sides of science, which in particular cannot be
separated by a clear-<:ut distinction. If this fact already suggests that we
not overemphasize deduction within theories, even more caution has to be
adopted when we take into consideration relations between different the-
ories; and we have already said above that the deductive model shows its
weakness especially here. Still it seems to us that the alternative approaches
which have been proposed have their own drawbacks.
In order to see this matter more clearly, some distinct aspects must be
analyzed.
(i) If two theories T and T exist as a matter of pure fact, the problem
of comparing them from a purely logical point of view is quite legitimate
and has even received considerable attention in mathematical logic, where
it has been shown when, how and to what extent such a comparison is
possible, if we understand it as something that should be expressed by
means of formal deductive tools or by means of model-theoretic ap-
proaches. Yet this problem has very little to do with the question of theory
change, and this for two distinct reasons: firstly because it is not said that
T emerged as a kind of modification of dismission of T (they might well
be two rival theories existing at the same time, as has so often been the
case in the history of science). And secondly, because claiming that theory
change is essentially or primarily a question of logic means a transition to
quite a different problem. This problem is that of investigating the reasons
or the motives of the said change; and while it is perfectly correct to study
which deductive or, more generally, which logical relations exist between
two theories, it would be at least very debatable to maintain that the one
has replaced the other because of certain logical imperatives.
(ii) The second problem just mentioned might be characterized as the
question about when and why a theory change occurs, and here we have
many divergent answers. However it must be stressed that this variety of
answers and their discord are, at least to a great extent, the consequence
of two different attitudes which may be adopted, both of which are legit-
imate, provided they do not claim to be the only legitimate attitudes: we
would call them, respectively, the descriptive and the normative ap-
proaches. According to the descriptive approach, our question has to be
64 EVANDRO AGAZZI
answered by a careful historical investigation; and from this it certainly

appears that theory change has occurred in science in connection and be-
cause of a great variety of factors, among which logical rigour and respect
for empirical evidence often playa very marginal role. Kuhn, Feyerabend
and several other scholars who have placed great importance on the his-
torical aspect of science, have laid so much stress on this aspect that they
almost completely disclaimed the thesis that logical and empirical require-
ments have something to do with the evolution of science. On the other
hand, those scholars who adopt the 'normative' attitude are surely ready
to admit that the actual historical development of science shows many
examples of a kind of 'laxist' divergence from what 'ought to be' the cor-
rect 'scientific' behaviour in abandoning a theory and accepting a new one
at its place. But they claim that this fact should not prevent us from re-
cognizing that the authentic and legitimate ground for justifying and eval-
uating scientific change is and especially ought to be that provided by a
convergence of purely logical and empirical requirements. It would be a
waste of time trying to give reasons for the one or the other of these two
attitudes, simply because they are both right to a certain extent, but cannot
account for the phenomenon of scientific change in its entirety. Logical
and empirical requirements alone are surely not sufficient to determine
theory change, and this is not so because scientists are laxist is their actual
praxis, but because these requirements are never fully cogent, as has been
shown ad abundantiam in the discussions of recent years. On the other
hand, all the other factors, be they of socio-cultural, philosophical, meta-
physical or pragmatic nature, though being of great importance in 'prim-
ing' and also in orienting scientific change, are not sufficient to 'accom-
plish' it, unless logical and empirical requirements are satisfied at least in
the final stage of the process. We might even venture to say that the ex-
tralogical conditions are more relevant in the case of 'macrochanges' (in-
volving transitions between large and very general theories) while logical
and empirical considerations play a more important role in the case of
'microchanges' (involving restricted theories which remain within the
framework of the more general ones).
(iii) A third set of questions arises when we consider the situation after
the change has occured, and these are independent of the possible reasons
which eventually led to the change. Now we have two distinct theories T
and T, and our problem has at least three aspects: what are the relations
between T and T; do these relations provide a possibility of comparison;

and, does this comparison open the possibility of speaking of progress?
Unfortunately, most of the discussions of the past years suffer from a
confusion between these distinct questions, as is shown by the fact that
quite frequently the inability of some relations to provide criteria of com-
parison was taken as an impossibility of comparing theories at all, by the
fact that incompatibility of theories was often confused with incompara-
bility, that continuity was confused with cumulativity, and that the lack
of cumulativity was confused with lack of progress.
A NEW APPROACH TO THE DISTINCTION BETWEEN OBSERVATIONAL

AND THEORETICAL CONCEPTS
Before trying to disentangle this skein, we want to point out a feature

which is responsible, in our opinion, for many difficulties occurring in the
evaluation of the role played by logic and experience in the analysis of
scientific change. This feature is particularly visible in the traditional prob-
lem concerning the relation between observational and theoretical con-
cepts in science. The accepted presupposition was at the beginning twofold:
the two classes of concepts are mutually independent; and the empirical
evidence is univocal, while the logical constructs which make up theories
are of very different kinds. The first presupposition was gradually de-
stroyed through the well known discussions which eventually led to the
admission that all concepts in science are theory-laden and that, in such
a way, no absolute separation may be drawn between observational and
theoretical terms. As a consequence it was claimed that the same variability
which characterizes the theoretical contexts also affects the empirical evi-
dence, so that this evidence could no longer be considered as an indepen-
dent basis for comparing theories. To our mind the real weak point was
the second assumption, in the sense that empirical evidence in science is
not at all univocal, but is already subject to a rich variability in itself and
quite independently of the theoretical framework. But this relativity of the
empirical evidence, precisely because it does not depend on the theoretical
context, gives us the possibility of restricting the impact of the theory-
ladenness thesis, and may grant the possibility of theory comparison. This
relativity of the empirical evidence is expressed by the fact that in every
scientific theory it is not the empirical evidence 'in general' which plays a
66 EVANDRO AGAZZI
role, but only a very restricted and specific empirical evidence, which re-
sults from standardized operational procedures of observation and mea-
surement. These procedures are certainly bound to some theoretical con-
text (the context which allowed the design of the instruments and of their
use), but this is not the theoretical context of the theory where these in-
struments are used as tools for providing the empirical evidence. For that
reason we abandon the terminology of 'observational' terms and adopt
instead the terminology of 'operational' terms; and this is not just a matter
of words, as will be seen in the sequel. The consequence is that the dis-
tinction between operational and theoretical terms is not absolute, because
it is relative to the particular theory considered; however this relativity
does not imply that the distinction not be" clear and unambiguous within
any single theory. In this way we reverse the well known thesis of Sneed,
according to which theoretical terms are simply T-theoretical (i.e. theo-
retical relative to a particular theory T, according to some criteria which
we need not mention here), while nothing is said regarding what is empir-
ical about the non-theoretical terms. According to our approach, opera-
tional terms are relative to the particular theory where they occur, and
they provide a foundation for the empirical claim of the theory, while the
theoretical terms are simply those which are non-operational (of course,
relative to the said theory). By stressing this we can avoid the total rela-
tivity in the meaning of all scientific concepts: it may be admitted that one
and the same concept changes its meaning at least to a certain extent
when passing from one theory to another, but this does not imply such a
radical modification as would prevent us from comparing theories. This
means that incommensurability is not a necessary consequence of the cor-
rectly stressed existence ora certain semantic relativity of scientific con-
cepts.
An important feature of this approach is that it enables us to recover
a great part of the 'sentential view of theories', without reducing it to a
purely logical .feature. Indeed we shall consider the different scientific the-
ories as systems of sentences which are intended to speak truly about a
specific domain of objects. The problem is that of correctly understanding
how this domain of objects has to be conceived and determined and it is
here that operations playa decisive and central role. In this regard I allow
myself to refer to some previously published research, in order to avoid
too many details. 3 However, since most of these publications are in
Italian, I feel obliged to give a brief account of these ideas by way of

introduction.
OPERATIONAL AND THEORETICAL PREDICATES
According to this approach, every scientific discipline is characterized by

its investigating reality from a selected specific point of view, which per-
forms a kind of 'cut' in the realm of reality, clipping out only certain
aspects of it and leaving the others out of consideration. Concretely speak-
ing, this happens by selecting a restricted set of specific predicates (or con-
cepts) to be used in the description of reality, in such a way that all the
sentences belonging to the particular discipline involved contain (at least
in principle) only such predicates. For example, classical mechanics is char-
acterized by speaking of reality only by means of predicates such as 'mass',
'length', 'time duration', and other concepts that may be obtained from
these by means of appropriate definitions. It is imperative for any empirical
science that at least some of these predicates be directly connected with
some operational testing procedures: we shall call these concepts basic
predicates, intending by that that they determine the domain of objects of
the given discipline. Additional predicates may then be introduced, but
always through a net of relations (be they of a general logical or of a
specific mathematical nature) which must enable one to reach sentences
entirely constituted of basic predicates by an explicit deduction, starting
from whatever other sentence of the theory. In view of this we may also
call operational predicates the basic predicates and theoretical predicates
the others, owing to the fact that these are introduced by means of logical
or mathematical relations which constitute the actual structure of the
'theory' involved. As is clear, no predicate is operational or theoretical as
such or in itself, but only with respect to the particular discipline under
consideration. On the other hand, the obligation imposed on the sentences
of the discipline to be explicitly linkable with sentences entirely constituted
of operational predicates corresponds to the obvious condition that every
sentence of the discipline must speak about the intended objects of the disci-
pline. According to this approach, the meaning of the predicates occurring
in a given scientific theory may be envisaged as follows: in the case of the
operational predicates there is a part of this meaning which is directly
bound with their definitory operations, and which may be called its 're-
68 EVANDRO AGAZZI
ferential part'; besides this there is another part which depends on the
context of the whole theory, and which comes from the net of logical
relationships that link the basic predicates reciprocally and with the other
theoretical predicates (we shall call this their 'contextual part'. In the case
of the theoretical predicates, their entire meaning depends on the context
ofthe theory, being influenced in particular by the logical relations existing
with the operational as well as with the other theoretical predicates. There-
fore they are endowed with a 'contextual meaning' only.
This remark is very important, as it enables us to see that in the case of
the operational or basic predicates there is a part of their meaning which
is not context-dependent (or 'theory-laden'). It follows that, if we can
consider sentences entirely constituted of operational predicates, we can
restrict our attention to that part of their meaning which only depends on
the operational meaning of their predicates, that is to say, which simply
express their 'referential meaning'. It is certainly easy to recognize that
'meaning' has been understood intensionally, in the above, but this is no
special feature of our discourse, as all talk about context-dependence or
theory-Iadeness of the meaning of concepts necessarily refers to the inten-
sional aspect of this meaning. What we explicitly add to this feature is
emphasis on the fact that referring to reality by means of certain standard
operations belongs to the intension of some concepts, and represents a part
of this intension which is not sensitive to the rest of it. This claim seems
to us very well grounded in the concrete analysis of any operational con-
cept. On the other hand, it should be superfluous to remark that the fact
that measuring instruments and the standardized way of using them de-
pend, generally speaking, 'on theories' does no harm. Indeed, these are
'other' theories, and therefore do not affect the pure operational character
of the referential meaning within the theory involved, as we have already
remarked.
THEOR Y COMPARISON
With this premiss borne in mind, we can proceed to consider the possibility
of comparing theories. Let us assume that two theories contain exactly the
same operational predicates, defined by means of the same operations. In
this case the two theories speak about the same domain of objects, and we
can try to look for a totally operational sentence which e.g. is a logical
consequence of the hypotheses admitted by the first theory, while being

rejected by the second theory. Being totally operational, this sentence may
be tested by means of the operational testing procedures equally admitted
by both theories, and the result of the test will enable a discriminating
judgment in favour of the one or the other. In this case we should say that
the two theories are comparable and prove to be incompatible. In case the
basic predicates are the same, and no operational sentence may be found
that supports one theory and disproves the other, we should say that, as
far as we were able to see, the two theories are comparable and compatible.
Let us now consider the case where two theories contain operational
predicates which are not completely identical: according to our approach
we shall simply say that they do not speak about the same objects, and
because of that that they are to be considered as norrcomparable or in-
commensurable; but this situation, as we have seen, is not the only one
possible, contrary to what is maintained nowadays by so many philos-
ophers of science, and this is true because the meaning of the scientific
concepts if not always totally contexHiependent. Moreover, we may also
say that incommensurability is in a way a matter of degree, for it may
happen that two theories, though being incommensurable in a very strict
sense, share a great many operational predicates, which might allow at
least some comparison of certain of their sentences.
THE IDEA OF SCIENTIFIC TRUTH REVISITED
But let us now try better to understand what incommensurability really

means. Usually it is understood as a kind of evidence that no certitude
may be reached in science, because no theory can disprove another. But,
as we have seen, the deepest sense of this incommensurability is that the
two theories refer to different objects, and this may express a richness rather
than a poverty of science. Transition from one theory to another does not
mean a more or less arbitrary change of opinion or of taste, but rather the
focussing of our attention on new domains of objects. The most important
consequence of this fact is that we are entitled again to speak of truth in
science and even to speak of a preservation of truth under the changing
dynamics of theories.
When philosophers and scientists began to say that science is unable to
reach truth, they did so because they felt discouraged by the fact that even
70 EVANDRO AGAZZI
the best established scientific theories, such as for instance classical me-
chanics, had been found false and rejected by the introduction, say, of
relativity theory or quantum mechanics. This impression was wrong, as
these people forgot that a sentence or a system of sentences is never true
or false 'in itself, but always with regard to some domain of reference. In
particular, the sentences of a given scientific theory aim at being true not
'in general' or 'as such', but simply with regard to their intended objects.
Hence, the sentences of classical mechanics cannot be disproved by quan-
tum mechanics, because the transition from the first to the second means
a change of the domain of objects, as can be seen from the fact that some
predicates which are operational in the one are not operational in the
other, and also by the fact that those which may be operational in both
are actually related to different kinds of operations in the one and in the
other respectively (this is why, e.g., the problem of measurement has so
many complex implications in quantum physics, which it did not have in
classical physics). Therefore quantum mechanics cannot prove that classi-
cal mechanics is false about its (classical) objects. As a matter of fact, we
can claim that a theory which has successfully passed several severe tests,
which has given rise to a rich display of applications, which has been
confirmed by a great variety of independent indirect controls, etc. may be
considered, with 'practical certitude', as being true of its objects and re-
maining true for ever about them. In this sense we must say that classical
mechanics is still true and will remain true of its objects (this is why, by
the way, we still use classical mechanics in so many technological appli-
cations as well as in celestial mechanics, in astrophysics, etc.).
A LEGITIMATE SENSE OF CUMULATIVE SCIENTIFIC PROGRESS
We are now in the position of understanding why it is legitimate to speak

of scientific progress even in some cumulative sense.
This happens because in the history of science there are several theories
which have established a rich set of true sentences about certain specific
domains of objects, and this truth is never destroyed by the fact that other
theories have proposed new systems of true sentences about new domains
of objects: quite the contrary, the new truths remain together with the old
ones and complement them. The result is that the global amount of human
knowledge is increased, in the sense that more and more aspects of reality
become known as a result of this proliferation of viewpoints, leading to a

proliferation of domains of investigation.
It may be clear by now why this progress cannot be understood as a
purely logical fact. Indeed, we have seen that theory change very often
means the opening of a new domain of inquiry, that is the investigation
of a new domain of reference, and this is by no means a matter of pure
logic, as it necessarily involves on the one hand the invention of new per-
spectives, which are very seldom a kind of logical consequence, general-
ization, or particularization of already existing ideas. On the other hand,
these perspectives, in order to become effective, must be endowed with
appropriate operational support, and this again oversteps the bounds of
logic.
Why must we say that this progress is not necessarily linear? This hap-
pens because the different domains of objects are very seldom, so to speak,
'embedded' into one another: this may occur sometimes, when the basic
predicates of a certain discipline prove capable of totally expressing the
concepts of another discipline, which becomes in such a way a subtheory
of the first. But this is a very rare case indeed. The most general situation
is that the different domains of objects are essentially separated. However
this situation too must be seen cum grano salis. In several cases we can
see that the respective domains are not totally disconnected, that they
admit of a certain overlapping, that there are some borderline problems
which may be investigated by the tools of two disciplines. This was prob-
ably the case with classical and quantum mechanics: the problems which
led to the creation of quantum mechanics were at the beginning such that
a treatment of them in terms of classical mechanics was not completely
impossible, though at the price of several ad hoc adjustments. This was
the symptom that cassical mechanics had, so to speak, reached its bor-
derline, and that its capacity of mastering the class of new phenomena
which were being discovered was exhausted. The perception of this diffi-
culty led to the opening of new viewpoints, which ultimately developed
into the creation of a new theory, with new operational procedures and
hence with new objects of its own (micro-objects, as they are usually called,
as distinct from the macro-objects). We could also say that science admits
of both linear and non-linear progress. The first occurs within a given
theory when no change occurs in its domain of objects. The second occurs
when the adoption of a new theory also implies a transition to a new
72 EVANDRO AGAZZI
domain of objects. What is interesting here is that both forms of theory

change are compatible with scientific truth and with the idea that progress
in science also means some kind of cumulation of truth. In the first case
this cumulation concerns truth about the same domain of objects (as we
have said, classical mechanics is still a domain of research where new truths
are being discovered). In the second case the cumulation of truth means
the discovery of new truths inside new domains of objects. Once one has
understood that truth is relative (in the sense of being relative to the objects
investigated) one can understand theory change, be it continuous or dis-
continuous, or be it in terms of commensurability or of incommensura-
bility, as not compelling us to give up the idea of scientific truth. Instead
of discussing in detail all the different possibilities of combination between
comparability and incomparability, compatibility and incompatibility, lin-
ear, non-linear, and cumulative progress, we prefer to present a few dia-
grams in which the situation can be captured in a more articulated and
synoptic way.
DIAGRAMS
1. The meaning of the operational concepts in two different theories T and

T'
Explications
(a) Thb Th 2, Th 3, and OPb Op2 are the theoretical and the operational
concepts in T.
Th' b Th' 2, Th' 3, and Op' bOp' 2 are the theoretical and the operational
concepts in T'.
(b) The lines - and -.-.- indicate the formal (i.e. mathematical or log-
ical) relations existing between the different concepts.
(c) The lines - - - indicate the non-formal (i.e. referential) relations
existing between the operational concepts and their 'defining' concrete
operations.
(d) It is supposed that all concepts are indicated in T and T' by the same
name. Still their meaning is different at least because of their different for-
mal contexts in the two theories.
Comments
(a) The global meanings of Op1 and Op2 are certainly different from
those of Op'1 and Op' 2, owing to their different contexts. In particular,
though Op1 and Op'1 appear to be bound to Th1 and Th2 (Th'1 and
Th' 2) by the same formal relations f1 and f2' it happens that the meanings
ofTh 1 and Th' band ofTh2 and Th' 2 are different owing to their different
contextual definitions. In fact:
(i) Th1 is directly connected with Th2 by means of f5, and directly
connected with Th3 by means of f4 (it is also indirectly connected with
Th3 by means of f5 and f6), while Th'1 is directly connected with Th' 2
by means of another function g1 and it is indirectly connected with Th' 3
according to two different patterns, i.e. via g1 and f6 and via g1 and g2.
(ii) The referential meanings of Op2 and Op' 2 are different, as they are
bound to two different operations W4 and Wn. But this difference in
meaning also affects Th3 and Th' 3 by virtue of the functional relation
f3, so that the meanings of Th3 and Th' 3 are really quite different; and
this fact is certain to affect the whole theoretical context of T and T'.
(b) As to the referential meanings of Op b OP2, Op' band Op' 2, some
distinction has to be made:
(i) Op1 and Op'1 have the same referential meaning, as they are directly
related to the very same operations Wb W2, W3.
74 EVANDRO AGAZZI
(ii) Op2 and Op' 2 have different referential meanings, as they are related
to two different operations W4 and Wn.
(c) If we now remember the strict relationship existing between the
operations and the objects of theories, it is easy to see that this has auto-
matic counterparts on the'referential meanings of their operational con-
cepts. The resulting possibilities are sketched in the following diagram.
2. Relationships between the domains of objects of two theories T and T'.
(a) inclusion (b) identity
(c) partial overlapping

00(d) total disjunction
Comments
(a) This case occurs when all the operational concepts of T also appear
with the same referential meaning in T' while T' contains other operational
concepts of its own. We shall call the two theories locally comparable..
(b) In this case all the operational concepts in T and T' have the same
referential meaning. Therefore T and T' deal with the same domain of
objects and are called fully comparable.
(c) Here T and T' have in common at least some operational concepts
with the same referential meaning, while other operational concepts (even
if they are labelled with the same name) have actually different referential
meanings. This is in particular the case presented in diagram 1. We shall
call the two theories partially comparable.
(d) In this case all the operational concepts have different referential
meanings in T and T', which means that they deal with fully different
domains of objects. We shall call them incomparable (or incommensurable).
As is clear, incommensurability is by no means the only possible case, but
just one among four possibilities.
3. Comparability and incomparability versus compatibility and incompati-

bility.
A. They are also compatible:
(a) Tand T' are (i) case (2a): T' is an extension of T
comparable (ii) case (2b): T and T' are complementary
(iii) case (2c): T and T' are partially compatible
(existence of some correspondence principle)
B. They are incompatible:

(i) T' falsifies T and embeds the objects of T in a
broader, but not radically different, domain of
objects
(ii) T' falsifies T and keeps the same domain of
objects
(b) Tand T' are It (i) they are neither compatible nor incompatible
imcomparable: follows (ii) T' does not falsify T
(case (2d» (iii) T and T' are both simultaneously "true" about
their respective objects.
4. Types of progress involved

(a) Case 3a Ai: Tand T' are comparable, compatible, and Tc. T': con-
tinuous linear and cumulative progress.
(b) Case 3a Aii: T and T' are comparable, compatible and complemen-
tary: continuous, non-linear, cumulative progress.
(c) Case 3a Aiii: T and T' are partially comparable and partially com-
patible: discontinuous, non-linear and cumulative progress.
(d) Case 3a B: T' falsifies T and replaces it: discontinuous non-<:umu-
lative progress.
(e) Case 3b: T and T' are not comparable and neither is falsified: dis-
continuous cumulative progress.
Comments
It may be said that the typical logical empiricist position was limited to
our case 4a; the Popperian doctrine was limited to case 4d; the 'incom-
76 EVANDRO AGAZZI
mensurability' doctrine of Kuhn and Feyerabend was limited to case 4e,

but it failed to recognize here the possibility of progress owing to a lack
of distinction between incompatibility and incomparability.
FINAL REMARK
It is perhaps appropriate to conclude this paper by saying something more

about the problem of the 'sentential view' of scientific theories.
As can easily be seen from the whole of our considerations, we share
the opinion that scientific theories are fundamentally (althouth perhaps not
exclusively) systems of sentences, provided we do not forget that they are
also and necessarily systems of sentences about some domain of intended
objects. To disregard or to deny this fact seems to us a serious flaw, which
in particular makes it impossible to explain why, after all, scientific theories
and in general scientific knowledge is presented in papers and books, which
cannot help but be composed of sentences (and this independently of the
already considered major reason that scientific knowledge must be an ex-
plicitly formulated one, which again necessarily implies the use of sen-
tences).
But, on the other hand, this does not mean that we have to accept the
deductive model for understanding scientific change. The reason is that
such a model is limited to the consideration of the possible syntactical
relationships between theories, which actually playa rather negligible role
in scientific change. But if we concentrate our attention upon the semant-
ical (meaning change) and referential (change of objects) properties of
these systems of sentences, a quite satisfactory possibility for understand-
ing theory change is certainly obtained. As we have seen, this happens
because the possibility of standardizing access to the referents (thanks to
the scientific operational procedures) provides every science with a certain
'stability of meaning', at least as far as some of its concepts are
concerned. In this sense our approach may be seen as a vindication of the
'empiria of science' after many years in which the 'logic of science' has
been so dominant. This vindication of the empirical aspect of science, how-
ever, must not be confused with the old approach, according to which the
empirical burden was totally placed upon observations; this was too weak
a ground, as the long and largely sterile debate on the difference between
observational and theoretical terms has shown. Our choice in favour of
operations (with its methodological specifications which could not be fully

spelled out in this paper) enables us to take fully into consideration, on
the one hand, that experience is always performed from some 'viewpoint',
within a certain 'Gestalt', but on the other hand that it provides US, owing
to its anchorage in certain referents, with a 'stable core' of meaning which,
among other things, allows us to speak again of 'scientific progress' in a
consciously recognized multiplicity of forms and meanings.
NOTES
1 For a good account and discussion of this 'deductive model' see e.g., C. Dilworth, Scientific
Progress, Reidel, Dordrecht, 1981.
2 For a brief but rigorous account of this position, and of its relation to most of the other
positions mentioned here, see in particular W. Stegmiiller, The Structuralist View of Theories,
Springer, Berlin 1979.
3 A rather detailed presentation of these ideas may be found in several publications of mine
which have appeared in Italian, including the last chapter of my book, Temi e problemi di
filosofia dellafisica. Milano, 1969, Abete, Roma 19742 , the essay'L'epistemologia contem-
poranea: il concetto attuale di scienza', in Scienza e filosofia oggi, Massimo, Milano 1980,
pp. 7-20, and 'Proposta di una nuova caratterizzazione dell' oggettivita scientifica', in Iti-
nerari (1979) nn. 1-2, pp. 113-143. In other languages, see my: 'Les criteres semantiques pour
la constitution de l'objet scientifique', in La semantique dans les sciences, Office International
de Librairie, Bruxelles 1978, pp. 13-29; 'The Concept of Empirical Data. Proposal for an
Intensional Semantics of Empirical Theories' in M. Przelecki et al. (eds), Formal Methods in
the Methodology of Empirical Sciences, Reidel, Dordrecht 1976, pp. 153-157; 'Subjectivity,
Objectivity, and Ontological Commitment in the Empirical Sciences', in R.E. Butts and J.
Hintikka (eds.), Historical and Philosophical Dimensions of Logic, Methodology and Philos-
ophy of Science, Reidel, Dordrecht 1977, pp. 159-171; 'Eine Deutung der wissenschaftlichen
Objektivitiit', in Allgemeine Zeitschriftfiir Philosophie' 3 (1978), 20-47; and 'Realism v Nauke
i Istoriceskaja Priroda Naukovo Posnanija', in Voprosi Filosofii 6 (1980), 136-144.
Received 24 April 1984
Philosophisches Seminar
Universitiit Fribourg
CH-1700 Fribourg
Switzerland
EIKE VON SAVIGNY
SOCIAL HABITS AND ENLIGHTENED COOPERATION:

DO HUMANS MEASURE UP TO LEWIS CONVENTIONS?
Since Convention appeared in 1969, David Lewis' conceptI has been dis-
cussed from different points ofview. 2 Partisans seem to have been attract-
ed, partly by the bounty of consequences which flow from the assumption
that a behavioral regularity is a convention in a population, partly by the
beautiful picture of man becoming social by acting rationally. Critiques
have concentrated on well-defined points without, however, shaking the
overall picture. 3 This is what I shall attempt to do: I shall argue that if the
background ideas of Lewis' analysis are taken seriously, being party to a
convention is beyond the average human's reach.
1. LEWIS' ANALYSIS
I shall begin by sketching out the analysis in light of its presumed back-
ground ideas. Consider the following stories: (Ceteris paribus require-
ments, notorious from rational decision theory, are stated once only, and
in brackets.)
(1) One agent,4 one solution: (Above all) Mother wants to enjoy Child's
company once a day. Her problem is how to achieve this. If she knows
she can (best) entice Child by offering a hot meal at 7 p.m. (Child grows
hungry at 7 p.m. and comes to the dinner table when hungry), then it is
rational for her to offer a meal at 7. Therefore, if she is rational, she will
offer a hot meal at 7.
(2) Two agents, one solution: Mother wants to enjoy Child's company,
Child wants a good meal. Mother knows she can enjoy Child's company
by serving a good meal when Child shows up; Child knows he can get
food by showing up when Mother serves dinner. Both know that dinner
takes about thirty minutes, that Mother cannot be finished before 7 and
that Child watches T.V. at 7.30 p.m. Mother and Child know everything
stated so far. If each takes the other for rational and if each knows they
do this, then it is rational for Mother to serve dinner at 7 and for Child
to show up; for it is then rational for them to rely on each other's doing
Erkenntnis 22 (1985) 79--96. 0165--0106/85/0220-0079 S01.80

80 EIKE VON SAVIGNY
what is required, in which case doing their own part is as rational as is

Mother's serving dinner in (I).
(3) One agent, two best solutions: Mother persists with her love-for-
food idea; however, Child can be won over by food all day. Since Mother
does not care whether dinner is at 7 or 8, she has the additional "problem"
of choosing the time. This she can solve by flipping a coin.
(4) Two agents, two best solutions: Same story as in (2); however, 8
o'clock is also available for both. Mother and Child now face an additional
problem. As in (3), and unlike (2), there is more than one solution, i.e.
more than one optimal combination of serving dinner at 7 or 8 and show-
ing up at 7 or 8; but unlike (2), neither Mother nor Child can independently
choose what to do. Lewis calls this kind of situation a "coordination prob-
lem".
This may - and in our case will- be a "recurrent coordination problem";
as long as Mother longs for Child's company and Child seeks food at
home, it will recur every day. In principle, the problem can be solved anew
each day, e.g. by Mother leaving a note when she leaves in the morning.
In many cases, however, a different solution will emerge. By whatever
process, Child comes to expect that Mother will set the table at 7, and
Mother comes to expect that Child will show up at 7. Every evening at 7,
Child shows up because of his expectation and Mother sets the table be-
cause of hers. The recurrent coordination problem is thus solved by "a
regularity in behavior produced by a system of expectations" 5. If we apply
Lewis' definition 6 to our case, we get:
Serving dinner at 7, on Mother's part, and showing up at 7, on Child's
part, is a convention between Mother and Child because:
(Kernel) They behave in this way; they expect each other to behave such;
both prefer simultaneity of dinner serving and showing up to non-simul-
taneity; each would opt for 8 o'clock on condition that the other does.
(Common knowledge of kernel) Both have noticed that they have kept
this regularity in the past; because both noticed that they have, they may
safely assume each other to have noticed that they did; and because they
did, they may safely assume
- that they are disposed to behave in this way,
- that they expect each other to behave such,
- that they prefer simultaneity to chaos,
- and that each would opt for 8, if the other did.
SOCIAL HABITS 81
Why common knowledge of the kernel? It makes operative expeCtations

rationaF. It is not only rational for Child to show up at 7 o'clock because
he expects Mother to serve dinner; it is also rational for him to expect this
because, by virtue of his share of common knowledge, he may safely as-
sume that she has rational expectations and preferences which will lead
her to serve dinner.
The leading ideas in this analysis are:
1. Conventional behavior is a kind of collective problem-solving. If no
convention is followed, everyone is likely to be worse off; conventions are
followed because people want to make sure that they are doing as well as
possible.
2. Conventions are solutions about which it would make sense to "con-
vene" even for rational people who always choose the best solution; for
there is more than one best solution, and which one should be choosen
depends on the choices of others. In this sense, conventions are arbitrary.
3. In a vague sense of "know", people can or do know their conventions
in a way which bolsters the impression of conventions as cleverly applied
tools.
My critique will be directed against the first and third ideas. The more
they get set in social habits, the less rationally people act, and usually they
do not know what they are then doing. (Because of this ignorance, phi-
losophers of language are deluded in the hopes they place in Lewis con-
ventions; this I will show in the concluding section.) I shall not question
the second idea; what remains of it, without problem-solving and knowl-
edge, is that most groups could do with systems of social rules which differ
from their actual ones, without being much worse off. This one would
hardly doubt.
2. NO RIGID PROBLEM-SOLVING
For conventional behavior to be collective problem-solving of the type

illustrated in sect. 1" parties to a convention should be sufficiently rational
to do what, according to their actual expectations, will bring about what
they prefer most. If in our above story, Child is feeble-minded (in the
clinical sense) but regularly shows up at the table when he smells dinner,
we would say that he reacts to the smell rather than that he pays attention
to what is necessary for satisfying his hunger. If Mother is unaware of the
82 EIKE VON SA VIGNY
probability of Child's showing up, it would be inappropriate to say that

she set the table because she expected him to come, even if she could have
known it. If, unknown to both Mother and Child, it is useful for them to
meet once a day, we would not say that, for this reason, meeting satisfies
one of their preferences.
Lewis starts from a picture of man acting rationally, a picture which he
takes for granted. However, people might grant the contrary, and I think
it is an open question which side the burden of proof lies with. Man may
be viewed according to a general character which I hope to call to mind
by recalling only some of its traits. One such trait is that in reflective
behavior there is a wide-spread reluctance to even consider unwelcome
side-effects, if the main object is attractive; another an almost universal
lack of consistency, e.g. where the methodical search for a leak or for a
short-circuit is spoiled by spontaneously following up sparks of genius
which prove useless. A third and perhaps contrary one is sticking to one's
premeditated course regardless of how the situation is developing, frequent
not only with single individuals but also with decision processes involving
several people where prior decisions often count as sacrosanct. A fourth
feature may be illustrated by one of Lewis' examples of conventional be-
haviorS: If people are out in the forest to gather something - wood, mush-
rooms, blueberries - what you observe is a strong tendency for them to
flock together and to work over the same places and bushes; exactly the
same thing happens when the dinner table is cleared - as soon as someone
begins to gather up plates, someone else invariably will follow suit. (Recall
how people cease to care for what they do, or worse, for what is done to
them, if they are many.) This list is quite unsystematic; its sole function is
to get hold of a general character of human behavior (much as we may
cite single features of a person's behavior in order to remind someone of
this person's character), a character which I take to be manifest from
everyday experience, even if it does not figure prominently in rational de-
cision theory.
Perhaps it is also useful to remember from the start that problem solving
differs from problem avoiding. Take, for instance, a colony of ants which,
one sunny summer, has settled beneath a large rock. Remove the rock and
you will find that the ants, in what might anthropomorphically be de-
scribed as a cooperative, disciplined, patriotic endeavor, rescue their eggs
from the dangerous exposure, reassembling in less than an hour to a safer
SOCIAL HABITS 83
place. However, if we assume - as perhaps we should - that their behavior

is purely instinctive in contrast to intelligent, we would not say that they
solved their problem collectively; rather we would say that each individ-
ual's repertoire of instinctive behavior is such that in this situation the
outcome of all is favorable for the colony's and, thereby, for each indi-
vidual's continued existence. Of course, if they judged the situation as we
would judge it in their place, if they loved life as we love life, and if they
were as prudent as we sometimes are (though typically not in such panic
situations), then we could say in a very literal sense that they solved their
problem 9 . However, assuming that their behavior is purely instinctive runs
counter to such assumptions.
Thus it does not seem obvious at all that in dealing with problems,
people generally solve them rationally. But this becomes definitely doubt-
ful when habitual behavior is concerned. I think it can be granted, as a
matter of empirical fact, that the regularities which Lewis has in mind are
habitual for members of the population. (Rarely instantiated habits are
just as reputable as are rarely instantiated regularities.) But then average
humans are unlikely to be sufficiently prone to make use of whatever share
of rationality they possess, to continuously solve their recurrent problems,
if this is to be distinguished from behavior which merely avoids them.
Most recurring difficulties are escaped by sheer habit; that a habit may be
learned does not add to the intelligence which is manifested in displaying
it. Rationality may have been manifested in the course of learning; but
once habitual, the behavior manifests no more rationality than does in-
stinctive behavior. This would be different if problem avoiding habits were
usually controlled in the sense that casual failures were met with quick
corrections. 1o But everyday experience teaches us a different lesson: Too
often, habitual problem avoiding behavior has first to be broken in order
to allow for useful plasticity in behavior. The more rigid behavior is, the
less easily can it be construed as applying solutions to problems.
In the second place, there is no convincing evidence that expectations as
to what other people will do are really operative when we do things which
are useful only if other people do certain other things. Such expectations
are operative if, when they go wrong, the behavior which they are held
responsible for is adjusted. This can frequently be found with new behavior
in new situations; therefore it is frequently plausible that new behavior in
new situations is actually guided by expectations. But the contrary holds
84 EIKE VON SAVIGNY
for habitual behavior in recurrent situations; thus, even if present, the

expectations are not operative.
Lewis more than once addresses himself to the problem of habitual be-
havior. He rightly claims that it can be rationally anchored in the agent's
preferences and expectations:
An action may be rational, and may be explained by the agent's beliefs and desires, even
though that action was done by habit ... A habit may be under the agent's rational control
in this sense: if that habit ever ceased to serve the agent's desires according to his beliefs, it
would at once be overridden and corrected ...
However, he vastly overstates the extent to which this happens:

... whatever may be the habitual processes that actually control our choices, if they started
tending to go against our beliefs and desires they soon would be overridden, corrected, and
retrained ... 11
I think that, in order to get a more adequate view of this matter, we ought
to distinguish some points in a possibly continuous scale of repetitive be-
havior:
1. Behavior may be repetitive because the same solution is attentively
brought about every time. Once the situation were relevantly different,
another solution would be found for this situation (i.e. for the first situa-
tion which is relevantly different).
2. Behavior may be repetitive because the agent relies on his strategy.
If a failure were to discredit the strategy, it would be corrected after the
first failure (i.e. a new solution would be tried for the second situation
which is relevantly different).
3. Behavior may be repetitive because the agent is set in his habit. He
reacts to the results of his behavior in the same way in which he reacts to
the results of other processes which he has got used to although he could
avoid them. Bad results have the effects of mishaps rather than offailures.
Many circumstances are relevant for whether or not measures are taken
against these mishaps and whether or not such measures consist in chang-
ing one's habit; the latter is just one of the possible starting points. Hope-
fully, the habit may be broken after a good many bad results.
4. Behavior may be repetitive because it is socially sanctioned. The
agent is not only set in his habit; the situation he faces is such that his
habit counts as unalterable. For him, it is no more among those elements
of the situation by which he can control its results. 12
SOCIAL HABITS 8S
Lewis does not require conventional behavior to be socially sanctioned;

but he agrees that in fact it will be 13 • - I submit that behavior of type I
rationally solves problems and that behavior of type 2 is problem-solving
in the sense of being controlled by a problem solving habit. However,
behavior of type 3 is no more problem-solving than, e.g., instinctive be-
havior, and behavior oftype 4 is far beyond anything similar. If it is favor-
able, it avoids problems; that is all.
Thirdly, since for all members of a modestly large group, it is terribly
difficult to pinpoint preferences regarding numerous possible individual
actions, construing their behavior as satisfying their preferences assumes
an air of arbitrariness. This arbitrariness is reduced if14 one imputes to all
group members a common desire which they all promote by conforming
to the regularity in question; most of Lewis' examples are of this sort. This
move may be more plausible in some cases than in others. Among car
drivers, there is a group-wide desire to avoid collisions, and this group-
wide desire is somehow operative in each and every driver (disregarding
suicides, and disregarding difficulties with "operative" for habitual behav-
ior). However, we have no reason to suppose that there is any such group-
wide desire, operative in each set of parents, which is achieved by the
socially sanctioned parental behavior of bringing up one's children, con-
servative family ideology notwithstanding. Bringing up one's children has
social functions, of course; it secures the expectation of there being people
to feed the aged and thereby secures group stability; moreover, it secures
survival of enough offspring, thereby securing group survival. However,
none of these goals are operative in many parents, let alone in all. Re-
member that if "A's behavior furthers 0 which is useful for A" is held to
suffice for "A reasonably behaves as he does because he knows that he
furthers his goal 0 in this way", then problem-solving is reduced to prob-
lem avoiding.
There is a further, and pretty intricate, problem about behavior satis-
fying one's preferences, if this behavior is socially sanctioned. Imagine that
in a population there exist preferences such that if R were a prevailing
regularity in behavior, R would satisfy the preferences about as well as
any alternative could do; R, therefore, might possibly become convention-
al. Let us now compare two different ways for the situation to develop: In
the first story, the preferences of several people change (from whatever
causes) in such a way that R is no longer a solution to their coordination
86 EIKE VON SAVIGNY
problem. In the second story, R becomes a convention, and becomes so-

cially sanctioned as such. Here the change in preferences occurring in the
first story is suppressed precisely because (this is a causal "because") acting
in accordance with them is socially enforced. Now this is quite generally
the case - people adjust their preferences to what they are permitted to do.
But then the convention sustains preferences rather than satisfying them.
Welcoming whatever result one arrives at by some action is quite different
from so acting for the sake of it.
3. KNOWLEDGE OF CONVENTIONS
In applying Lewis' definition to the Mother-Child example in sect. 1, I

slurred over a margin of interpretation which opens in his definitions of
"common knowledge" and of "indicating", especially when these are con-
sidered alongside his interpretive remarks and applications to examples.
Taken as it stands, the definition requires only that there is evidence, avail-
able to every member of the group in question, from which each can ra-
tionally conclude, 1. that what we termed the convention's"kernel" holds,
and 2. that every other member has his evidence too. Whether or not, by
using their evidence, they will eventually arrive at knowledge about what
others would like to happen if ... , or what they expect each other to do
if ... , or what they expect each other to like to happen if ... , will depend
on two steps being taken. They must be sufficiently rational and informed
to conclude in a first step, from the evidence at hand, that the "kernel"
holds and that the others have this evidence too; and starting from the
facts that the kernel holds and that the others have their evidence too, in
a second step, they must arrive at the said bits of actual knowledge. If, by
some mishap, they fail to come across such knowledge, then the letter of
the definition need not be violated; its spirit patently would.
It is even a matter of dispute 15 whether, as Lewis intends his definition
to be understood, members of a population could have the required evi-
dence without arriving at actual knowledge; for "What A indicates to x
will depend ... on x's inductive standards and background informa-
tion."16 Apparently A must be such that the first step is actually taken
by every member of the group. However, Lewis does not seem to consider
this as a restriction on the evidence in question, but - at least if the evidence
consists in past conformity to a regularity in behavior - as a more or less
SOCIAL HABITS 87
trivial part of the mental equipment of the group members. What they
need for the second step, namely
... mutual ascription of some common inductive standards and background information,
rationality, mutual ascription of rationality, and so on 17
is not required by the definition (not even in the indirect way just illus-
trated), and there is no general discussion in Convention as to how likely
group members are to satisfy such assumptions. However, the whole of
pp. 24 to 76 of the book leaves no room for doubt that, in Lewis' view,
all but "children and the feeble-minded" 18 are sufficiently rational, have
sufficiently high opinions of each others' intelligence, and have enough
background information to take the second step.19 Therefore, even if the
definition does not require parties to a convention to know what they are
intended to know, it is understood that normal people who satisfy the
definition are such that they have the intended knowledge.
Considering two extreme interpretations, I shall argue two points: (1)
Past conformity does not suffice as a basis of knowledge of conventions;
common knowledge as Lewis defines it does not result in what he intends.
(2) There is no basis for common knowledge; common knowledge as Lewis
intends it does not result from anything like its intended basis. I shall also
argue a third point which attacks this idea independently of Lewis' defi-
nition: (3) Knowledge of conventions in an everyday sense can exist, but
cannot be regarded as a general feature of conventions.
Where "knowledge" is used in what follows, I agree that one may know
something without being able to put it into words 20 and that knowledge
may be latent in contrast to being just thought about. I do not think that,
with non-verbal knowledge, there is a telling difference between knowledge
in sensu composito and in sensu diviso 21 unless knowledge in sensu diviso
is knowing how in Ryle's sense (on which see below); thus, conceding
non-verbal knowledge in sensu diviso amounts to no more than conceding
knowing how. I do not agree, however, that knowledge may be potential
in contrast to actual: if someone only could know something but just does
not, then even if he is in a position to acquire the knowledge, he has none
so far, be this because he does not make use of available evidence, e.g. in
failing to look up in the dictionary which is right before his nose, or be-
cause he is too dull or too tired, or because he fails to have opinions which
informed people ought to have or has opinions which informed people
ought not to have. "Potential knowledge" would be a misleading term for

avoidable ignorance. In Cooper's words:
The expression "merely potential knowledge" ... is scarcely clear, and one's first reaction
might be that potential knowledge is no more knowledge than a potential clientele is a clien-
tele. I think first reactions are right. 22
The argument on point (1) attacks the plausibility of the following pre-
diction: Group members have witnessed passed conformity with R (which
satisfies the "kernel" of convention); therefore, they are likely to know
that R is a convention in their group. Argument: The prediction makes
use of factual premises on common inductive standards, sufficient back-
ground information (especially on each others' preferences), and mutual
ascription of inductive standards, which are empirically untrustworthy or
at least unwarrented. Conclusion: The accessibility of past conformity with
R to members of the group is no basis for assuming that they know R to
be a convention.
Why are the factual premises dubious? On the level of everyday experi-
ence, there is good evidence that different people who have been confront-
ed with past conformity to R will seem to "witness" different regularities
(they interpret the behavior in divergent ways), that their beliefs about
what others prefer and are likely to do under different circumstances differ
wildly, and that some are more prone than others to rely on their fellows'
intelligence, just as some are more likely to be relied upon than others are.
Prejudice is particularly strong when one tries to make sense of the be-
havior of others; common sense abounds with proverbs in this regard.
However, common sense might be considered a bad source (if it is right
on our point, it is likely to be one). Let us therefore tum to surprising
experimental findings in the study of social cognition:
Jones, Davis, and Gergen ... showed that people were more able to attribute a stable per-
sonality disposition to an actor who had behaved in a manner contrary to his best interests
in a social situation ... than when he acted in accord with his best interests. . .. Kelley ...
suggested: To the extent that we can conclude that the behavior in the situation was one of
low distinctiveness, low consensus, and high consistency, then [sic!] we can attribute the be-
havior to the stable properties of the actor. ... 23
The selective italics are all mine; they are to suggest the possibility that it
is difficult to acquire expectations regarding other people's behavior if they
act according to their preferences, if their behavior is salient, and if they
behave in the same ways as the others do.
SOCIAL HABITS 89
Now the very untrustworthiness of the ancillary premises is crucial for

the argument on point (2). The above prediction of knowledge is an in-
direct argument: Given certain information, we are expected without fur-
ther investigation to rely on conventions being known. In contrast, one
may think of a direct argument: The members of the group in question
present a peculiar appearance which can best be understood on the as-
sumption that they have common knowledge of their conventions which
is based on evidence about everyone's past behavior. Whatever this pe-
culiar appearance may be, it must consist of facts about the group mem-
bers' behavior which are best interpreted as manifesting knowledge which
is actually based on that evidence. It will not actually be so based (in
contrast to warranted), unless the above ancillary premises refer to dis-
positions which the group members really have. Since we have no reason
to assume they do, we cannot regard such an interpretation as the best
one. (Inferences to p as the best explanation of q depend on the plausibility
of the further premises which are needed in the explanation.)
The difference which should be stressed between some knowledge being
actually based on bits of evidence and its being warranted by this evidence
might become more translucent if we point out that knowledge need not
even be actually based on the data the initial reception of which was once
responsible for acquiring the relevant beliefs. It seems appropriate to add
this remark here, because Lewis might possibly like to view his basis for
common knowledge as a starting point in a learning process. Compare the
following passage (the verbal/nonverbal question is not relevant for the
present context):
Like it or not, we have plenty of knowledge we cannot put into words. And plenty of our
knowledge, in words or not, is based on evidence we cannot hope to report. Our beliefs are
formed under the influence of impressions left by a body of past experiences, but it is only
occasionally that these impressions allow us to report the experience that created them. You
probably believe that Kamchatka exists. Your belief is justified, for it is based on evidence:
mostly your exposure to various books and to incidents that confirm the reliability of such
books. Try, then, to make a convincing case for the existence of Kamchatka by reporting
parts of your experience. There is no reason why our knowledge of our conventions should
be especially privileged. Like any other knowledge we have, it can be tacit, or based on tacitly
known evidence, or both.24
I fully agree that most of our knowledge is nonverbal. However, if Lewis

speaks of "ta~t1y known evidence" and "evidence we cannot hope to re-
port", and if he equates "based on evidence" with "formed under the
influence of impressions left by a body of past experience" which "created"

the impressions, he is obviously stretching the relation of evidence which
justifies belief so far as to cover experiences which cause belief In short, he
interprets having learned that p as having evidence for p; or, as in the rowing
case 2S , perceiving that p as having evidence for p. However, learning is a
process which causally creates our knowledge; any different process with
the same effect would do just as well. Whether or not our knowledge is
justified by our learning data is thus at best of accidental interest for its
base. Whether or not it is actually so based depends, for instance, on
whether or not new data, which would have cut off the learning process
if it had been among the learning data, would lead us to change our minds
if we were now confronted with it. This is anything but obvious.
The general idea of common knowledge of conventions, or at least that
of knowledge of them, might to a certain extent be saved, even if we drop
the requirement that it be based on particular evidence in a certain way.
Let us therefore turn to point (3) and ask: Do people generally know their
conventions, however based their knowledge may be? In order to avoid
difficulties about which definition of convention is to underlie our ques-
tion, let us resort to a minimum: Do people generally know many of the
sorts of behavior which are sanctioned among them?
In one sense they do: They know how to behave, for this is just a way
of expressing that they behave in the ways which are sanctioned in their
group. However, we must not be misled by the suggestiveness of the fol-
lowing chain of reformulations:
behaving correctly
knowing how to behave
knowing how one has to behave
knowing the correct behavioral patterns
knowing which behavioral patterns are correct behavioral pat-
terns
knowing what behaving correctly looks like
Kemmerling 26 suggests that Lewis hovers between knowing how and
knowing that, and he rightly remarks that knowing how would not do
justice to Lewis' manifest idea that knowledge of convention should, for
rational agents, justify conformative behavior. In order to betray interest-
ing knowledge of socially sanctioned behavior, agents who know their
SOCIAL HABITS 91
behavior to be correct, should differ in behavior from agents who do not;

still, the difference must make sense for nonverbal knowledge.
Let us look at sample differences which, even if consisting of verbal
behavior, make sense for nonverbal knowledge of one's socially sanctioned
behavior. Somebody who not only is a member of a group which follows
R, but also (even if nonverbally) knows that R is followed, is prepared
(1) to fight for the replacement of R with R' if he prefers R' over-R and
(2) to fight for the propagation of R in other groups, if he wishes all
groups to follow his own group's conventions.
However, for too many people and regularities, neither (1) nor (2) are
actually exhibited. People who sincerely prefer equal rights for women fail
to fight for emancipation because they wrongly believe that discrimination
is not conventional. University professors sincerely propagate the principle
of achievement as a mechanism for vocational selection because they
wrongly believe that it is a rule, worthy of emulating, which they them-
selves observe (instead of the principle that you scratch my back and I'll
scratch yours). In short, one wide-spread cause for people's ignorance of
their conventions consists in the fact that actual conventions tend to di-
verge from ideologies.
Now, if it is not knowledge -let alone rationally acquired rational ex-
pectations - that launches and stabilizes even our useful sanctioned be-
havior, then what else could explain this miracle of nature? I have no
competent opinion. 2 7 The theory of origin and conservation of social me-
tastability 28 is beyond the professional limits of a philosopher. In analyz-
ing concepts, philosophers may indeed betray their liking for a picture of
human nature: they reserve the right for man, viewed in a certain way, to
satisfy the requirements which they lay down in their analysis of an essen-
tially human feature. I think we ought neither to exclude nor to be afraid
of the idea that behavior among humans comes to be and remains socially
sanctioned in ways which resemble those of other gregarious animals. Man
is not an animal sociale by virtue of being an animal rationale: it may well
be that being social is more fundamental than rationality.
4. CAN MEANING SNEAK IN VIA COMMON KNOWLEDGE?
What is common knowledge good for in a definition of convention? If we

do not care about the exact definition of "common knowledge" but con-
sider the ideas which are evoked by words like "know", then we would
feel that group members who share common knowledge of the kernel of
their conventions solve their coordination problems in an extra-rational
way: They know they have a coordination problem, they know why it is
one, and their expectations (in which light their behavior is rational any-
way) are warranted by what they know. This is a pleasing view for philos-
ophers who think that man is a rational animal. With Lewis, however,
there are more practical uses of common knowledge: A minor service of
the condition is that some unwelcome examples are exciuded29 ; the major
one is the bridge it builds from convention to meaning. Lewis tries to prove
that a speaker who utters a sign in accordance with a convention of lan-
guage30 thereby means the conventional meaning of that sign; in his proof,
common knowledge plays an essential part.
This, of course, is a delightful result. If what we normally regard as
linguistic utterances can be shown to be actions by which the· speaker
means something thanks to the fact that he sticks to a convention, then
we seem to gain a deep insight into what the meaningfulness of language
owes to its conventionality.
All well and good - if by independent reasons common knowledge can
be shown to be a common feature of conventions which govern linguistic
behavior. I have tried to show in Section 3 that, in the sense of (nonverbal,
latent, but:) actual knowledge, it is a very improbable feature of human
behavior; and it is precisely this sense in which common knowledge is
required for Lewis' proof. To show this, I shall examine one detail.
Lewis tries to prove that the speaker's doing u (signaling) after observing
s, upon observing which the audience conventionally responds by r, is
covered by the three Gricean intentions for his meaning nonnaturally that
the audience should respond by r, and that it is covered by the second
intention in particular, viz. the speaker's intention that his audience recog-
nize the speaker's intention to produce r by doing u. Here is his proof ("I"
is the speaker, "you" his audience):
I expect you to infer s upon observing that I do G. I expect you to recognize my desire to
produce r, conditionally upon s. I expect you to recognize my expectation that I can produce
r, by doing G. SO I expect you to recognize my intention to produce r, when you observe that
I do G.31
The second and third premises are to be provided by common knowledge

(the first one is in the kernel). They require actual expectations 32 ; merely
SOCIAL HABITS 93
being entitled to expectations would not suffice for actually having inten-
tions. Even if Lewis were right in saying
The intention with which I do (1 can be established by examining the practical reasoning that
justifies me in doing it. I need not actually go through that reasoning to have an intention;
actions done without deliberations are often done with definite intentions,33
we ought not to conclude that, because thinking something out is more

than is required for thoughtful behavior, being entitled to believe some-
thing could substitute for believing it. Where belief is required for inten-
tion, availability of evidence will not do in its place. 34
Consider the following sequence:
In every instance of S, x has witnessed that p.
In every instance of S, x has noticed that p.
Of every instance of S, x knows that p.
In every instance of S, x expects that p.
Take any of its steps, and it will make sense to claim that many people
frequently fail to take it. However, since each of the four words leaves
room for interpretation, an unexpecting witness can hardly be described as
having failed to take exactly one of the steps (rather than another). Thus
the feeling arises that the break in the sequence is not real, the connection
still potential. (Every heap is potentially a void place.) Perhaps this elu-
siveness extends, via intending, to meaning. That nonnatural meaning ex-
ists, by and large, only potentially I am the last sceptic to doubt.
I conclude that, if defined as availability of evidence, common knowl-
edge cannot play the part of importing meaning via convention; common
knowledge in a stronger sense might suffice, but is unlikely to exist. This
motive for incorporating it in a definition of convention is therefore void.
NOTES
1 Convention, chs. I and 2. Many of the considerations that follow bear equally on the
partly parallel analysis found in Schiffer's Meaning.
2 For a particularly perspicuous exposition, see Kemmerling, Konvention und sprachliche
Kommunikation. This small monograph contains much substantiiU. criticism, as well as an
improvement which saves what Kemmerling takes to be Lewis' central ideas.
3 Jamieson, "David Lewis on Convention", has been answered by Lewis in "Convention:
Reply to Jamieson"; see further Lewis' replies to several objections in "Languages and Lan-
guage". Questions regarding the status of his analysis vis-a-vis related concepts have largely
been dealt within advance in Convention, ch. 3; this discussion relates partly to terminology,
partly to the more substantial question of the actual extension of Lewis' concept. Gilbert's
point (in "Agreements, Conventions, and Language", Sections 4 and 5) that conventions of
the "first time when x happens" type (which are all due to explicit agreement) may, by their
very formulation, extend to exactly one situation only, does not run counter to the spirit of
Lewis' definition (which excludes them). For the agents' disposition is still one to react to a
type of situation: a situation of the type F such that no other of type F has preceded it. It is
this disposition that counts, and it can obviously fire also in the second situation if the first
has gone unnoticed, or been forgotten. For further critical discussions, see the references
below. I disclaim most of my critical remarks in "Listener-Oriented Versus Speaker-Oriented
Analysis of Conventional Meaning", p. 73 f.
4 Child is an agent too; but his problem is not considered.
5 Convention, p. 118.
6 Convention, p. 78, in combination with the definition of "common knowledge", p. 56, and
of "indicating", p. 52 f. I presuppose that past conformity is the "basis" for common knowl-
edge of conventions referred to in the definition of common knowledge. Because the example
is a two-alternative two-person problem, the application is simpler than the general definition
which runs as follows: A regularity R in the behavior of members of a population P when
they are agents in a recurrent situation S is a convention if and only if
(1) almost everyone conforms to R;
(2) almost everyone expects almost everyone else to conform to R;
(3) almost everyone has approximately the same preferences regarding all possible com-
binations of actions;
(4) almost everyone prefers that anyone more conform to R on condition that almost
everyone conform to R;
(5) almost everyone would prefer that anyone more conform to R', on condition that
almost everyone conform to R' (where conforming to R excludes conforming to R');
(6) and a state of affairs A holds such that
(a) everyone in P has reason to believe that A holds;
(b) if any member of P had reason to believe that A held, it would thereby have reason
to believe that everyone in P had reason to believe that A held;
(c) if any member of P had reason to believe that A held, it would thereby have reason
to believe that (1) to (5).
In "Languages and Language", p. 5, condition (4) is replaced by the following: the belief
that others conform gives everyone a good and decisive reason to conform himself. The new
condition is intended to cover also regularities in belief acquisition; this extension is of no
concern in the present context.
7 In "Languages and Language", p. 6, Lewis states a different motive: Common knowledge
of the kernel is to ensure its stability. I think that such an empirical claim can be convincing
only if rationality is taken to be wide-spread and effective to a high degree; and it is this
assumption which strikes the eye when Lewis hopes to ensure the stability of a convention
by other means than, say, common habit.
S Convention, p. 45. I do not wish to deny that a more economical division of labor would
develop if Princeton staff is camping.
9 This does not mean that intelligence is a matter of inner phenomena of judging, reflecting
and the like! Differences in observable behavior can be big enough.
10 Please note that rational behavior is not thought to be necessarily backed by reasoning
processes.
11 "Languages and Language", pp. 25 f., and Convention, p. 141. (See also "Convention:
Reply to Jamieson", pp. 114 f.) At the close of both passages, I have omitted' the words "by
conscious reasoning". It would be beside the point to refute Lewis by pointing out that
learning by failure almost never involves conscious reasoning.
SOCIAL HABITS 95
12 If such a situation were construed as a change in the agent's beliefs about what he can
do, I should propose to loose interest. There is of course no doubt that, for even the wildest
bit of behavior, there is a consistent function which describes it as rational in terms of some
beliefs and of some desires.
13 Convention,p. 100.
14 - and only if, as Kemmerling has convincingiy argued for coordination problems. See
Konvention und sprachliche Kommunikation, pp. 35-37.
IS In "Lewis on Our Knowledge of Conventions", Cooper claims that "Lewis does not
distinguish between the case where a person is acquainted with what is in fact evidence for
a conclusion, and the case where he is acquainted with it as evidence. Someone may be
surrounded by people who regularly do R, who expect one another to do it, etc., yet be
sublimely unaware of all this." (p. 258, author's italics.)
16 p. 53, immediately following the definition of "indicating". I thank Andreas Kemmerling
for pointing out that Lewis might intend this sentence as an interpretation rather than as a
factual claim.
17 Convention, p. 57 f.
18 p. 62, fn. I, and p. 75.
19 There is but one indication to the contrary: "[Knowledge of our conventions] may be
merely potential knowledge. We must have evidence from which we could reach the conclu-
sion that any of our conventions meets the defining conditions for a convention, but we may
not have done the reasoning to reach the conclusion." (p. 63.) If "potential" means "not
arrived at by reasoning", then the sentence fits in with the remainder of the text. However,
knowledge which is based on evidence may be arrived at without reasoning from this evidence,
and knowledge, which if existing would be based on available evidence, may be missing for
causes other than a failure to reason. If this is what "potential" means, then the cited passage
is strongly inconsistent with the remainder of the text in which Lewis patently takes for
granted that in matters of social behavior, people - aside from children and feeble-minded
- believe what they are entitled to believe.
20 See Convention, pp. 63 f.
21 See Convention, pp. 64-68. Knowledge in sensu diviso is all Lewis requires.
22 "Lewis on Our Knowledge of Conventions", p. 256. The aim of Cooper's argument differs
from mine; he tries to show that if the definition is satisfied, then people know their conven-
tion in a full-blooded sense of "know". He remarks (ibid.) that this result tends to narrow
the extension of the concept of convention, if compared to Lewis' "poor sort of knowledge";
however, he does not venture a guess as to how wide-spread full-blooded knowledge of
conventions, and thus the existence of Lewis conventions, might in fact be.
23 Joel Cooper, "Cognitive Theories in Social Psychology", International Encyclopedia of
Psychiatry, Psychology, Psychoanalysis, and Neurology, III, Aesculapius Pub!., New York
1977,212-215 (p. 213). I thank Prof. Alois Angieitner, Bielefeld, for drawing my attention
to the relevance for background knowledge, in the present context, of social cognition.
24 Convention,p. 64.
2S Convention, p. 63.
26 Konvention und sprachliche Kommunikation, p. 27 f. In "Languages and Language", p. 25,
Lewis states a condition, sufficient for knowledge of conventions, which reads like a definition
of knowing how: "It is enough to be able to recognize conformity and non-conformity to his
convention, and to be able to try to conform to it."
27 Even if my bias is more on Burge's side: "The stability of conventions is safeguarded not
only by enlightened self-interest, but by inertia, superstition, and ignorance." ("On Knowl-
edge and Convention", p. 253.)
28 Lewis' delightful picture of existing conventions; see Convention, pp. 42 and 92.
29 Convention, pp. 53 if., 59.
30 Convention, pp. 154--159. The proof is carried out for a primitive form of linguistic con-
vention, viz. signaling conventions, but Lewis conjectures that the result is generalizable (p.
159).
32 Actual in the sense of real rather than occurrent. Actual expectations are dispositions too.
On p. 9 of "Languages and Language", Lewis suggests: "Perhaps a negative version of
[common knowledge] would do the job: no one disbelieves that [the kernel holds], no one
believes that others disbelieve this, and so on." Applied to the above proof, this would yield
something like the following: I do not question that you might infer s upon observing that
I do u; I do not question that you recognize my desire to produce r, conditional upon s; I
do not question that you might recognize that I put up with producing r by doing u. -
Whether or not what follows from this would suffice in the context of re-modelled Gricean
conditions, as proposed, for instance, by Bennett (Linguistic Behaviour, p. 125) and Kem-
merling ("Utterer's Meaning Revisited"), I do not know.
34 The same assumption of actual beliefs (in contrast to one's being entitled to believe) is
used in Bennett's parallel argument in Linguistic Behaviour, p. 180: "[The speaker] utters S
intending to communicate P, and expecting to succeed only because he thinks that A will
think that he utters S with that intention; which means that he is relying upon A's recognizing
what his intention is ... ". (Italics mine.)
BIBLIOGRAPHY
Jonathan Bennett: 1976, Linguistic Behaviour, Cambridge University Press, Cambridge.
Tyler Burge: 1975, "On Knowledge and Convention", Philosophical Review 84, 249-255.
David E. Cooper: 1977, "Lewis on Our Knowledge of Conventions", Mind 86, 256-261.
Margaret Gilbert: 1983, "Agreements, Conventions, and Language", Synthese 54, 37~7.
Dale Jamieson: 1975, "David Lewis on Convention", Canadian Journal of Philosophy ~
73-81.
Andreas Kemmerling: 19"76, Konvention und sprachliche Kommunikation, Dissertation Miin-
chen.
Andreas Kemmerling: (In print), "Utterer's Meaning Revisited", in Richard Grandy and
Richard Warner (eds.), Philosophical Grounds of Rationality: Intentions, Categories, Ends,
Oxford U.P., Oxford.
David K. Lewis: 1969, Convention, Harvard U.P., Cambridge, Mass.
David K. Lewis: 1975, "Languages and Language", in K. Gunderson (ed.), Language, Mind,
and Knowledge (Minnesota Studies in the Philosophy of Science, VIII), Univ. of Minn. Pr.,
Minneapolis, Minnesota, 3-35.
David K. Lewis: 1976, "Convention: Reply to Jamieson", Canadian Journal of Philosophy
6, 113-120.
Eike von Savigny: 1976, "Listener-Oriented Versus Speaker-Oriented Analysis of Conven-
tional Meaning", American Philosophical Quarterly 13, 69-74.
Stephen Schiffer: 1972, Meaning, Clarendon Pr., Oxford.
Universitiit Bielefeld Received 26 October 1983

Abt. Philosophie
Postfach 8640
4800 Bielefeld 1
F.R.G.
C. ULISES MOULINES
THEORETICAL TERMS AND BRIDGE PRINCIPLES: A

CRITIQUE OF HEMPEL'S (SELF-)CRITICISMS
Hempel has significantly contributed to the discussion of the problem of

theoretical terms. This is a problem of the semantics and pragmatics of
science, i.e. it is a problem about the meaning and use of terms in science.
This essay contains some reflections on Hempel's views of the matter.
The first question we have to ask is this: What is the problem of theo-
retical terms? The second question is: Is this problem meaningful, i.e. is
there really a problem of theoretical terms? In his [1973] Hempel has given
a clear and succinct answer to the first question:
The sentences of a [scientific] theory can have objective empirical significance and can explain
empirical phenomena only if the theoretical terms they contain have clearly specifiable mean-
ings. Thus there arose the problem of characterizing those meanings and indicating how they
are assigned to the theoretical terms. This was one of the principal questions to which the
standard analysis was addressed; let us caJl it the meaning problem for theoretical expres-
sions l
As to the second question, Hempel's view has apparently undergone

radical changes. In his 'Theoretician's Dilemma' (Hempel [1958]), he not
only thought that the problem was meaningful but also that its analysis
provided a very important insight into the structure and functioning of
scientific theories: It was supposed to afford a precise and truthful account
of, among other things, the fact that science is able to explain and predict
phenomena in a way superior to common sense. 'The Theoretician's Di-
lemma' is the best exposition I know of the "standard" view on the nature
of theoretical terms and, by extension, of the structure of scientific theories.
On the other hand, in his [1973] Hempel seems to have changed his mind
completely. After criticizing in many ways the "standard construal" he
ends up with the pitiless words:
Hence at least one of the major problems to which the standard conception was addressed,
the problem of meaning specification for theoretical terms, rests on a mistaken presupposition
and thus requires no solution. 2
Hempel's negative conclusion and the accompanying selfcriticism, which

stand in the purest logical positivist tradition of deciding that the problem
Erkenntnis 22 (1985) 97-117. 0165-0106/85/0220-0097 $02.10

98 c. ULISES MOULINES
to which one oneself has devoted so many efforts for so many years is after
all nothing but nonsense, were already prepared in Hempel [1970]. It is to
this two (self-)critical essays that I shall devote the bulk of my reflections
in the first part of the present essay. In the second part, I develop my own
view of the matter.
What are the difficulties in the problem of theoretical terms that led
Hempel to reject it as meaningless? Why is it that the question of how the
meaning of theoretical terms, as opposed to observational ones, is deter-
mined comes to be a pseudo-question? The reason is not, that Hempel
converted to Quine's extensionalism and, by rejecting the notion of mean-
ing altogether, a fortiori rejected any question of meaning for theoretical
terms as "meaningless". For, as he points out quite explicitly,
The reason just given for this rejection does not hinge at all on the recently much debated
obscurities of the intensional notion of meaning, which have led Quine and other analytic
philosophers to reject that notion ... as hopelessly unclear. The considerations I have adduced
remain in force also for a purely extensional construal of the doctrine. 3
Nor is it that he became much impressed by the arguments of meaning-

variance theorists like Feyerabend that purport to show that, since all
terms of a theory change their meaning when the theory changes, all terms
are theoretical in this sense and therefore there is no specific problem of
theoretical terms. In fact, he makes quite clear that he takes his own argu-
ments as showing that Feyerabend's thesis is as untenable as the standard
construal. 4
It could be thought that Hempel's original motivation for departing
from the standard interpretation of the problem of theoretical terms is
intimately related to the alleged impossibility of finding a precise and con-
vincing characterization of the concept of an observational vocabulary or
observational basis: If it is not clear what an observational term is, then it
would not be clear either what its opposite, viz. a theoretical term, is. And
if it is not clear what a theoretical term is, then the problem of theoretical
terms vanishes as an unclear problem about unclear things.
Now, it is likely that Hempel, as so many others, was inspired to reflect
more critically on the notion of a theoretical term when the idea of an
absolute observational language made bankrupt in the epistemology of
the early sixties. However, I don't think that this is the profound reason
that led Hempel to dismiss the problem of theoretical terms. For, he him-
self substituted the notion of an antecedently available or pretheoretical
THEORETICAL TERMS 99
vocabulary for the notion of an observational language in his own char-

acterization of the standard construal s - rightly noting that replacement
of the idea of observationality by the notion of antecedency or pretheo-
reticity would not change much in the statement of the problem of theo-
retical terms. After all, Hempel was never quite so fond as other episte-
mologists of trying to convey a particularly privileged epistemic status to
the non-theoretical terms of a theory. In his systematic (and still positive)
exposition of the standard construal6 , he already characterizes the "basic
[non-theoretical] vocabulary" in a very flexible way: "VB consists of terms
which are antecedently understood. They might be observational terms,
in the somewhat vague sense explained earlier, but we need not insist on
this ". 7
Thus, the problem of theoretical terms does not depend for Hempel on
a narrow and epistemologically more or less questionable characterization
of an alleged observational vocabulary. 8 The problem may be restated
without having to resort to observational terms: When confronted with a
particular scientific theory, we encounter some terms which pose no spe-
cific semantical problem, since they are "antecedently understood" (how
this comes to be the case is another question for which Hempel does not
seem to have any definite answer); on the other hand, there are other terms
specific of that theory which do pose semantical problems, since they are
not understandable outside the theory in question. And the problem is
now: How is it that the theory provides (or determines or makes clear) the
meaning of those latter terms?
The standard construal's answer to this question is that this meaning is
provided by some of the theory's sentences. They are oftwo sorts: internal
axioms or basic laws on the one hand and correspondence rules or bridge
principles on the other. Hempel reserves the denomination "interpretative
sentences" as usual for the second class of sentences only; however, since
both kinds are supposed to act jointly to fix the meaning of theoretical
terms, I shall use "interpretative sentence" to cover both "bridge princi-
ples" and internal axioms as used for semantical purposes. The resulting
conception of the structure of a theory that comes from this view of the
way theoretical terms are determined can best be described by quoting
Hempel himself at length:
Prima facie, therefore, the formulation of a theory may be thought of as calling for statements
of two kinds; let us call them internal principles and bridge principles for short. The internal
principles serve to characterize the theoretical setting or the "theoretical scenario": they
specify the basic entities and processes posited by the theory, as well as the theoretical1aws
that are assumed to govern them. The bridge principles, on the other hand, indicate the ways
in which the scenario is linked to the previously examined phenomena which the theory is
intended to explain ...
The formulation of the internal principles will typically make use of a theoretical vocabulary
VT, i.e., a set of terms not employed in the earlier descriptions of, and generalizations about,
the empirical phenomena which T is to explain, but rather introduced specifically to char-
acterize the theoretical scenario and its laws. The bridge principles will evidently contain
both the terms of VT and those of the vocabulary used in formulating the original descriptions
of, and generalizations about, the phenomena for which the theory is to account. This vo-
cabulary will thus be available and understood before the introduction of the theory, and its
use will be governed by principles which, at least initially, are independent of the theory. Let
us refer to it as the pre-theoretical, or antecedent, vocabulary, VA, relative to the theory in
question. 9
This conception of a scientific theory is essentially the same as the one set
forth in Hempel [1958], p. 208 - the only essential difference being that
now Hempel thinks that it is fundamentally mistaken. The main reason
for his thinking so is neither extensionalism, nor meaning variance, nor
the fact that VA is historically and culturally relative. It is just the linguistic
approach to theoretical terms:
For the root of the difficulties we have noted lies in a misconception inherent in the construal
of the problem itself, namely, in the requirement of explicit linguistic specification, according
to which a philosophically adequate answer to the meaning problem must produce a set of
sentences which specify the meanings of the theoretical terms with the help of an antecedently
available empirical vocabulary ... and this general notion cannot be made good. 10
And the reason why Hempel thinks that the idea of a linguistic determi-
nation of theoretical terms through interpretative sentences is fundamen-
tally wrong is that it would make empirical theories true by convention -
which is a reductio ad absurdum for him. 11 He thinks that this idea comes
from a misled and untenable extrapolation of metamathematical theory
analysis of the Hilbert type to the philosophy of empirical science. 12
We may summarize Hempel's argument in the following way:
(A) A linguistic approach to the problem of theoretical terms implies
that the meaning of these terms in a given empirical theory T should be
rendered through a special class of sentences.
(B) These sentences can only be Ts axioms or Ts bridge principles to
some antecedently available vocabulary - the "or" not being exclusive.
(C) But this use of Ts axioms or bridge principles would make T true
by convention.
(D) Since T is an empirical theory, it cannot be true by convention.

Conclusions:
Therefore,
[1] the linguistic approach to the problem of theoretical terms is fun-
damentally wrong (Hempel: "a misconception inherent in the construal of
the problem itself, namely, in the requirement of explicit linguistic speci-
fications"13);
[2] in particular, there are no interpretative sentences for theoretical
terms (Hempel: "sentences with a special interpretative function ... There
are no such sentences" 14);
[3] hence, there is no point in sharply distinguishing between axioms and
bridge principles (Hempel: "it should be explicitly acknowledged now that
no precise criterion has been provided for distinguishing internal principles
from bridge principles ... The distinction is, thus, admittedly vague" 15);
[4] finally, then, there is no problem of theoretical terms (Hempel: "the
problem of meaning specification for theoretical terms rests on a mistaken
presupposition and thus requires no solution"16).
I think that conclusion [1] is true and moreover correct given the sort
of premises inherent to the linguistic approach; and it certainly has been
a significant gain for general philosophy of science that such a great classic
as Hempel came to state this conclusion as sharply as he did. I also think
that, cum grana salis, the intermediate conclusion [2] is also right - though
if we are not careful enough its wording may be misleading, as we shall
see later on. At any rate, I think that conclusions [3] and [4] are too hasty
and furthermore wrong in a very fundamental respect. In the remainder
of this paper I intend to explain why I think so.
Let's start with the last one. As a matter of fact, there is a meaningful
problem of theoretical terms, as Sneed has shown in his [1971]. The prob-
lem is not exactly the same as the one posed within the standard conception
of theories, but it is similar in spirit, and by analyzing it one understands
why the "great classics" rightly felt that there was indeed a problem with
theoretical terms - though their linguistic approach led them astray when
trying precisely to state what the problem was. It can even be said that
there is a meaning problem with theoretical terms if we admit the flexible
idea - while avoiding the intricacies of a general theory of meaning - that
the problem of determining the meaning of a scientific term is non-trivially
related, by implication or by "association", to the problem of determining
its extension. When analyzing an empirical theory with a minimal degree

of complexity - though not necessarily a metricized theory - we encounter
some terms, usually corresponding to the ones intuitively felt as "theoret-
ical" within that theory, whose extension cannot be determined at all un-
less the fundamental laws, and normally some of the more special laws as
well, are presupposed as true. In the particular case of metricized theories
this means that they contain metrical functions that cannot be measured
at all unless one presupposes the theory's laws. As Sneed has argued, this
leads to a kind of paradoxical situation. The paradox, however, is not
unsurmountable. It can be solved by abandoning the standard linguistic
approach and developing an alternative, and richer, concept of an empir-
ical theory. 1 7 That terms having the property just mentioned - and which
may justly be called "theoretical terms" - appear in almost all (or, perhaps,
even all) empirical theories with a reasonable degree of systematic power
has been substantiated by numerous reconstructions of real-life examples
of scientific theories in the last ten years - not only in physics but in other
disciplines as well. 1s
All this means that there is a sufficient amount of concrete reconstructed
material by now to support the following claims: First, there indeed are
theoretical terms in science in a welldefined sense. Second, the determi-
nation of their meaning (viz. extension) does pose a specific problem.
Third, the structure of this problem indicates that the linguistic approach
and the idea of interpretative sentences are on the wrong track. Therefore,
Hempel's conclusions [1] and [2] are admissible, but his further conclusion
[4] is not. Of course, those readers which are well acquainted with the
structuralist program for metascience and with the fact that I myself am
engaged in that program will not be very much surprised by my acceptance
of Hempel's [1] and [2], and rejection of Hempel's [4]. What is presumably
less well known is the reason someone engaged in the structuralist recon-
struction program might have for rejecting also Hempel's [3] and for re-
lating this rejection to the problem of theoretical terms. This reason comes
out as a corollary of the structuralist approach to intertheoretical relation-
ships as expounded in some current work within that program 19 • It is
impossible to develop a reasonably complete account of this issue here but
I shall try to outline the main ideas. By doing this, I shall also try to
convince Hempel that his rejection of the distinction between internal and
bridge principles is ill-founded.
I think the only two premises I need to convince Hempel of the rightness
of my argument are these:
(I) There are numerically distinct scientific theories; or, more exactly,
the term "scientic theory" is meaningful and it applies to certain numeri-
cally distinct things.
(II) Not all scientific theories are methodologically isolated from each
other.20 (In fact, we could plausibly strengthen this premise to the thesis
that all, or almost all, scientific theories have some kind of methodologi-
cally relevant relationship to some other scientific theories; but for the
present argument I only need the weaker version of the thesis.)
Now, if (I) is true, it follows that it makes sense to try to identify those
things called "scientific theories", that is, the ontological claim (I) about
the existence of particular things called "scientific theories" entails the
meaningfulness of the epistemological endeavor of identifying them. (If
you think there are centaurs, then it makes sense for you to try to identify
individual centaurs.) By the same token, the ontological conviction that
there are methodologically relevant relationships between different the-
ories implies that it is meaningful to try to find them. Now, my claim
against Hempel can in a nutshell be put this way: There is an essential tool
for performing the first epistemological task, namely internal axioms or
fundamental laws; and there is an essential tool for performing the second
epistemological task, namely bridge principles; moreover, generally speak-
ing these two kinds of tools cannot have the same logical form, so that it
makes sense to differentiate them; finally, the combination of the internal
axioms and the bridge principles plays an essential role in establishing the
distinction between the theoretical and the non-theoretical terms of a
theory.
I suspect that Hempel will tend to agree on premises (I) and (II), but
that he will have trouble in following the rest of the argument. Conse-
quently, I shall develop the line of the argument in more detail. Before
this, however, let me say one word on the rationale for (I) and (II).
Two objections may be raised against (I). Intuitively, they are closely
connected, though logically they are independent. The first is that theories
are, if anything at all, things that are continually changing so that there
is no way to identify them. The second objection is that theories have no
sharp boundaries, they are a diffuse sort of entity, so that we cannot tell
one from another. Both arguments are similar in kind and purpose, but
the first comes from a diachronic perspective while the second is synchro-
nic. Now, the first argument against the numerical identity of scientific
theories is not conclusive whereas the second is an extreme exaggeration
of a real situation. What the first argument essentially points to is the fact
that scientific theories are a genidentic sort of entity - which is true. But
from this we cannot infer that it is impossible to identify them or to dis-
tinguish one specimen from another of the same kind. Mter all, a human
being also is a genidentic entity, continuously changing21, which does not
preclude us to identify and distinguish individual people. It may be coun-
tered that the possibility of identifying and distinguishing human beings
or organisms in general in spite of their being genidentical entities comes
from the fact that, at any time, they are spatially well localized in a way
scientific theories, by being cultural products, are not. But then we may
take other genidentic cultural products that are well identifiable and dis-
tinguishable. Consider two different natural languages, say, English and
Spanish: They certainly have changed a lot since the times of Chaucer and
the Arcipreste de Hita, respectively; nevertheless, nobody doubts they exist
as two separate entities. In fact, I claim that scientific theories are very
similar to natural languages in many respects (though not in the sense
some classical philosophers of science used to say that "a theory is a lan-
guage"). If natural languages are identifiable and distinguishable, then I
see no reason why scientific theories, which normally are characterizable
in even more stringent terms than natural languages, should not also be
identifiable and distinguishable with at least the same degree of accuracy.
Generally speaking, there is no good reason for maintaining that geni-
dentical, culturally determined entities cannot have an identity of their
own.
As for the second objection, it is certainly true that theories usually have
borderline cases, that some concepts or laws may be common to different
theories, or that sometimes we may not be sure whether something is just
one theory, or many, or really only a part of a theory. But all this means
that scientific theories are empirically given things, and as so many other
empirically given things in this not too Platonic world, a certain degree of
fuzziness or of overlapping has to be admitted in their determination. If
we don't allow for this in the case of scientific theories, then we should not
allow it either in the case of many other sorts of fuzzy entities like organic
species, physiological and psychological states, illnesses, languages, na-
tions, institutions, economic systems, art styles, and so forth - which would
make the whole of the social sciences and a great deal of the natural
sciences just impossible. Any scientific analysis starts with an idealization
of its object of study, which means, among other things, to settle sharp
demarcation lines where there appears none. And I see no reason why we
should require higher standards of actual determinacy for the objects of
metascientific than for the objects of scientific analysis.
Therefore, I conclude that we may safely assume that there are such
things as distinct scientific theories - each one with its own personality, so
to speak. Of course, this does not mean that to identify them, to tell one
from another, to decide whether we are in front of one theory, or rather
of two, or perhaps only of a portion of one theory - that all this is an easy
matter. There are lots of things in the world that we may not be able to
identify, or to determine more than vaguely, even if we have good reason
to think they have a definite identity of their own. Scientific theories may
be among them. The hope of the philosopher of science is that they are
not; that is, he hopes that by adequately sharpening his tools of analysis
he will be able to identify theories with a reasonable degree of accuracy.
This is the first task a working program of logical reconstruction has got
to undertake. And this is precisely the epistemological task associated with
the ontological premise (I) - a task for which Hempel apparently does not
show much sympathy.
But before we go on to this issue, let us briefly consider the justification
for premise (II). Once we have accepted that there are different scientific
theories it is presumably still easier to accept that they are not theoretical
monads but that usually, though perhaps not always, they are mutually
related in some methodologically relevant way. That is, normally it is
"vital" for the ongoing life of a theory to be connected to some other
(clearly different) theories. 22 Instead of a general argument for this, it is
probably sufficient here to give some plausible examples. Whether or not
Maxwell's electrodynamics is just one theory or a whole family of similar
theories, and whether or not Newtonian mechanics is one theory or many,
it is just a metaempirical fact about Maxwell's electrodynamics that, while
being clearly distinguishable from Newtonian mechanics, from the very
beginning it (or its family members) became essentially associated to New-
tonian mechanics through a very important link (for electrodynamics): the
consideration of so-called Lorentz' forces. Each theory of the family
known as "phenomenological thermodynamics" is essentially linked to

some theory of fluid dynamics, for the very reason that the applicability
of any thermodynamic theory, as simple as it may be, presupposes the
notion of pressure, and this is provided by fluid dynamics. Whatever the
ultimate fate of the different versions (= theories) of decision "theory"
that are on the scientific market, they all must work with a metricizable
concept of belief, which, by having to satisfy KolmogorofI's axioms, im-
plies the presence of a link to the rather different theory of (subjective)
probability. I don't think there is any need to produce further examples.
The general point is this: If there are such things as theories at all, and if
they are supposed to work, then they will have to take some help from'
other individuals of the same sort. And this help is provided by some
special sorts of connections we shall from now on label "(intertheoretical)
links". Now, the same sort of passage from the ontological to the episte-
mological question is due here: Granted that there are links between the-
ories - how can we come to determine them?
The minimal reconstruction program I have in mind to undertake the
double epistemological task I have just outlined is, thus, one that allows
us to identify individual theories and individual links between theories.
Can this task be fulfilled in any reasonably precise way? It seems that
Hempel (and many others) strongly doubt it. I think, on the contrary, that
it can be done. And I also think that the apparatus of the structuralist
methodology is well suited to do this. However, I don't think that for this
task we need the whole battery of structuralist notions. The minimal pro-
gram I have in mind is but a small portion of the structuralist program -
a portion that could be used by other philosophers of science, even those
that wholeheartedly hate structuralism. Essentially, all we need is the idea
that to any given scientific theory a class of models can univocally be
associated. If we have identified the class of models of a theory, then we
have identified that theory in a quite precise way; and the easiest way to
identify a specific class of models is through the linguistic means of positing
a set of sentences called "axioms".
To avoid some misunderstandings that may easily be avoided at this
point let me make clear that there are two things that I am not saying
when I say that the. best way to identify a theory is through a class of
models and that the best way to identify this class is through some axioms.
First, I am not saying that a theory is just a class of models nor that the
class of models is the only essential thing to know about a theory. A theory
may be a kind of thing that is essentially constituted by many other com-
ponents besides a class of models. Actually, I think this is so in the great
majority of interesting cases. To know the class of models of a theory is
not to know all there is to know about the theory in question. All I am
saying is that to know the class of models is the most general and precise
way we have to identify a theory. An analogy may help here, if needed.
It is certainly not the case that to know the complete name of a person
together with her birth date and place is all there is to know about that
person. 'But as all policemen and statisticians know, this is a very good
means to identify her. There certainly are many other means to identify
a person, some of them providing a much more thorough knowledge of
her. But the "complete name - birth date - birth place" strategy is the best
in the sense of being most generally applicable and exact. It is the same
with the class-of-models strategy for scientific theories.
Secondly, I am not saying that the only way to identify a class of models
is through a set of axioms. Much less am I saying that a particular axiom
set is essential for a theory. It is well known that many different axiom sets
may all pick out the same class of structures as models for a given theory.
It is not very important which particular set of axioms we choose for
determining the class of models of a theory, so long as that set actually
determines the class we want. What is important, for practical reasons, is
that we know there is at least one such set. For theoretical reasons, not
even this is important. In principle, we could determine a class of models
by non-axiomatic means. And the class of models is all we need for the
theory's identification. But, of course, axioms are normally of a great help.
This is the reason why people that want to know which theory they are
confronted with with a sufficient amount of precision usually begin by
axiomatizing it in some way or another. This way can even be rather in-
formal, so long as it is precise enough to render just what we want. Thus,
in the last analysis, axioms, though theoretically dispensable, are practi-
cally essential to identify a theory - pace Hempe}23. This by no means
implies to fall back into the linguistic analysis of theories, which Hempel
and I agree was a wrong move in philosophy of science for quite a long
period. It only means that the linguistic device of choosing a set of for-
mulae as axioms for a given theory is useful to determine the non-linguistic
compounds of that theory. Language is not all there is in the world, not
even in the world of metascience, but it is a serviceable tool for finding out
the non-linguistic things of the world we are interested in.
The class of models for a theory normally is a proper subclass of a
"bigger" class of structures which are all of the same mathematical kind.
In more exact terms, this is the class of all structures of the same type as
the models in question. This notion may be generally defined by means of
the set-theoretic concept of a structure-species24 • But these technicalicities
are not important now. What is important here is that by means of the
notion of structure species we can "distill out" of a theory's own class of
models a super-class which may be called the "class of potential models"
or "possible realizations" of that theory and which contains all essential
information about the conceptual (set-theoretic) structure of the theory in
question. If M(T) symbolizes the class of models of T, then, by considering
the structure species of the elements of M(T), we may introduce a class
Mp(T), such that M(T) ~ Mp(T) and Mp(T) contains all systems T possibly
applies to - though it may fail to do so. Mp(T) determines so to speak the
space of possibilities of success for T.2S
It is plausible to think that, besides M(T), Mp(T) also belongs to the
"personal identity" of a given theory. Arguments for this view abound in
the structuralist literature. However, this is not the point now. I have not
introduced Mp(T) here as a tool for identifying particular theories, but as
a base for identifying links between theories. The intuition behind is this.
We should view the essential relationship between two given theories as
fixed by particular links as non-dependent on the respective classes of
models of either theory. To put it in linguistic terms, in any adequate
reconstruction of two theories and their essential interrelationships, the
formulae that express their links should not logically imply the axioms of
each theory. Actually, we should qualify the last statement and say: they
should not imply the fundamental laws of either theory, i.e. the so-called
"proper axioms" that determine the models (as opposed to the "improper"
or "structural axioms" which only determine the potential models). Other-
wise, from the knowledge of a link between two theories T and T' we could
already infer whether the fundamental laws of T and T' apply or not; that
is, a theory's most essential component for its own identification would be
dependent on the presence of a link to some other theory. This may be
plausible to assume for some cases of theories that are, from the very
beginning - so to speak, in their "essence" - subsidiary or "parasitic" with
respect to other theories. For example, it may well be argued that rigid
body mechanics has this sort of relationship to Newtonian particle me-
chanics - that you are ready to consider a rigid body as a Newtonian
particle system only if it satisfies the laws of momentum for rigid bodies
because the truth of these laws presupposes the truth of Newton's Second
and Third Laws. But this situation seems to be idiosyncratic to rigid body
mechanics and similar cases of theories which, by their very construction,
are supposed to be reducible (or equivalent) to some other, previously
established, theory. However, even here one could also argue against this
way of interpreting rigid body mechanics and similar examples. At any
rate, this possible situation is not to be viewed as the general case. If it
were the general case, then there would be no way by which any given
theory may "import" some essential information about its own applica-
bility that does not presuppose the truth of its fundamental laws. That is,
in such a case, empirical theories would be like totalitarian states that
allow the importation of outward information only under the condition
that it fits the official truth. In the last analysis, as will become clearer
towards the end of this paper, this would make independent checks for
any empirical theory impossible; that is, there would be no empirical theory
at all. And this is certainly not a consequence we are willing to buy. For
all we know about science, scientific theories are not ever self-satisfying
individuals.
All this means that a link between theories T and T cannot be subor-
dinated to M(1) and M(T'). Formally speaking, if I is any link between T
and T we are interested in, then in general we cannot define it as a relation
of the sort;
I ~ M(1) x M(T).
Of course, it could be the case that, by analyzing a particular link we find

out that it is of that sort. But this should be regarded as a contingent
matter. We should not presuppose of all links that they have this form.
On the other hand, a link between T and T must have something to do
with the conceptual structures of T and T. I take this to be an analytic
statement about links. Otherwise, we would not have any means to decide
whether a particular alleged link is really a link between Tand T. To do
so, we have to see that that link really relates some terms of T with some
terms of T. What this formally means is that a link 1 between T and T

should generally be defined as
1 £;; Mp(T) x Mp(T).
In linguistic terms this means the following. Let's call the particular
formula(e) that express(es) a link 1a theory Thas got to some other theory
(a) "bridge principle(s)" for 7'2 6 • Then, what we are saying here is that
when we formally reconstruct T we have to introduce Ts bridge princi-
ple(s) in such a way that they are logically dependent on at least some of
the structural axioms for Ts potential models. The converse is not true.
Ts structural axioms should, by themselves, imply no coordinating prin-
ciple to other theories. This, by the way, is what usually bothers people
when they are confronted with a set-theoretic axiomatization for the first
time. They complain the structural axioms say "almost nothing" about
the terms involved - in particular, they say nothing about the "physical"
or "empirical interpretation" of the terms. They fail to see that this is right
so in an adequate axiomatization, even in the case of empirical theories.
For what the so-called "physical interpretation" really amounts to is the
formulation of some links (considered essential) to other theories. For
example, when considering the "classic" set-theoretic axiomatization of
Newtonian particle mechanics 2 7, people find that the potential models of
this theory are structures of the form < P,T,s,mj> such that the only
structural axiom about, say, the term s just says that "s is a twice differ-
entiable vector function" but does not say that s is the position function.
To "interpret" s as the position function is not, however, the business of
any structural axiom for the models of mechanics; it is rather the statement
of a link of this theory to some underlying theory, like physical geometry
or geometrical optics. The bridge principle that, for example, links me-
chanics to geometry through the s-function implicitly entails the structural
axiom about s just quoted but not conversely.
To sum up. When formally reconstructing a scientific theory Ttwo kinds
of principles are needed (for "practical" if not for "theoretical" reasons):
internal principles or axioms to fix the class of models of T, which in turn
identifies T, and bridge principles to fix the links T has to other theories.
Further, these two kinds of principles can be distinguished sharply in a
careful reconstruction - both model-theoretically and syntactically: From
the model-theoretic point of view, the first kind of principles are those
associated with Mp(n and M(n; the second class are those associated
with some relations Ii on Mp(n x Mp(Ti); clearly, both kinds of things
are quite distinct. This difference may be translated into syntactic terms
(though this is not so important): The first kind of principles only contains
terms of the structures in Min; the second kind contains terms both of
Mp(n and some other Mp(Tj). If all this is correct, then Hempel's conclu-
sion [3] is false. On the other hand, the clear and distinct existence of tools
for fixing M(n and for fixing Ii for T does not in any way imply that either
the tools themselves or the theory they serve are "true by convention", if
this means that they will never be abandoned in the face of experience.
These tools are definitely not "analytic interpretative sentences". They may
become abandoned if the dynamics of scientific experience so requires.
Therefore, the fact that Hempel's conclusion [3] is false does not imply
that his [2] is also false, contrary to appearances. The reasons [2] and [3]
seem to be so closely connected that the abandonment of the one would
appear to entail the abandonment of the other is that usually the wit of
"bridge principles" and their sharp distinction from internal axioms is seen
only in their "interpretative" character. But this is a misinterpretation of
the role played by bridge principles between theories, as the previous dis-
cussion shows.
To all this Hempel may well retort that I just missed the point - that his
dissatisfaction with the idea of a sharply distinguished class of bridge prin-
ciples was in connection with the problem of theoretical terms, and not
with some other issues related to the reconstruction of theories. All right,
he might say, perhaps there are such things as well definable links between
theories - but what the hell has this to do with theoretical terms? Now, in
the final part of this paper I want to argue that the existence of bridge
principles and their sharp distinction from the axioms do playa role in the
distinction between theoretical and non-theoretical terms, though not in
a direct and simple manner. This also will provide some further clues on
the reasons why Hempel's [4] is wrong.
Let's come back for a moment to the Sneedian characterization of the
theoretical terms of a theory T. They are those terms specific of T in the
sense that the determination of their extension presupposes the applica-
bility of Ts axioms. We may describe this situation in somewhat more
general and exact terms. Let f be a fundamental term of T (i.e. f is not
definable through some other terms of n. If T is supposed to be a scientific
theory, then the determination of the extension of I in T will not be ar-

bitrary, or provided by mystic contemplation, or occult manipulations,
but there will be some well specifiable, intersubjective, repeatable patterns
of determination. In the case where I is a metrical concept, these patterns
are what is normally called "measurement methods". To include the more
general case, where metrical concepts are not necessarily involved, let us
speak of "methods of determination". These correspond to such things as
particular measurement devices together with some rules for operating
them. Ifwe want to say that it is within Tthat we are using certain methods
to determine f, then the conceptual structures representing those methods
will be of a type similar to, or compatible with, the one characteristic of
T; otherwise, we would not have any warrant for saying that we are meas-
uring (or determining) I within T. In the simplest case, the structures rep-
resenting those methods of determining I in T will fully coincide with the
structures characteristic of the conceptual structure of T, i.e. they will be
potential models of T.28 Let us call the set of all potential models of T
that are used as methods of determining I "MATJ". Then, Sneed's original
idea may be rendered more precisely through the following "criterion of
theoreticity":
(CT) lis T-theoretical iff M f [11 £; M[Tj.29
This is the model-theoretic explication of the intuition that a term is T-
theoretical iff its determination always presupposes the fundamental laws
of T. That is, "linguistically" speaking, among the conditions that fix the
f-determining structures in M f [11 the fundamental laws or axioms of T
will always appear. Normally (perhaps always) some special laws will be
added to the fundamental laws in order to fully determine aT-theoretical
termf30
If all fundamental terms of a theory T were T-theoretical, then, accord-
ing to (CT), all we would need to determine any term of T in any appli-
cation of T would be to pick out some appropriate element of M(T). This
element would be its method of determination. Linguistically, it would be
fixed through the structural axioms, the fundamental laws and probably
also some speciallaw(s) of the same T. However, since scientific theories
are not selfsatisfying individuals (pace Feyerabend), not all their terms will
be T-theoretical. In a scientific theory T we shall always encounter some
term g for which there is a method of determination within T that is not
a model of T. Let us call such terms T-non-theoretical 31. The methods of

determination for T-non-theoretical terms may still be conceptualized as
(full or partial) potential models of T, but, in general, they may not be
actual models of T. The methods of determination for T-non-theoretical
terms that are not actual models of T are those that allow for an independ-
ent check of the theory - those that tell whether the theory has gone astray
or not. They provide what is usually called "the empirical data" for T. Let
us symbolize one such method to determine a T -non-theoretical g by "n;'.
We have that ng determines the extension of g - if g is a metrical function
ng "measures" g - and that it is plausible to take ng E Mp[l1; but we also
have ng ¢ M[1]. That is, in this case only Ts structural axioms may play
a role in determining g, but the fundamental and special laws of T play no
role. Since the structural axioms are adInittedly very poor in content - how
are we supposed to determine g by means of ng at all?
It is here that intertheoreticallinks come in. The additional information
fixing ng to determine g is not supposed to be "pure observation", nor a
mystical and direct immersion into reality, nor any other atheoretical pro-
cess of that sort. It will also be theoretical information, but information
that comes from some other theory through an appropriate link. This link
connects the method of determination of gin T with some methodes) of
determination for some other function(s) in some T; and this (or these)
methodes) of determination in T; presumably even fulfill(s) T;'s funda-
mental laws and some of its special laws as well. Formally, this "linkeage"
condition for non-theoretical terms may be explicated as follows:
(LNT) For any T-non-dependent method of determination ng of a
T-non-theoretical term g there exist a theory T;, a link I;, and
a method of determination m; such that
(1) I; s;;; MiT) x MiT;)
(2) (ng,m;> E I;
(3) m; E M[T;]
(4) m; ¢ M[11
This "linkeage condition for non-theoretical terms" could further be re-

fined in order to take account of reducible theories and other complica-
tions of a kind similar to the ones that appear when formulating (eT) and
that have been dealt with in Balzer / Moulines [1980] (especially § IV).
However, I don't think it is necessary to enter into these sophistications
here. The intuitive plausibility of (LNT) - or a refined version of it - may

best be expounded by some examples: The non-theoreticity ofthe position
function in Newtonian particle mechanics is essentially associated to the
fact that position is linked to the length and angle functions of physical
geometry and there are geometric methods of determination of length and
angle that don't presuppose the laws of Newtonian mechanics; the non-
theoretical mole numbers of equilibrium thermodynamics are linked to
methods of determining molecular weights in stoichiometry; volume in any
thermodynamic or hydrodynamic theory is non-theoretical because it is
linked to length in physical geometry.
Thus, we may conclude that we can bind the distinction between theo-
retical and non-theoretical terms on the one side and the distinction be-
tween internal axioms and bridge principles on the other in the following
way. Internal axioms (particularly fundamental laws) are practically es-
sential to determine theoretical terms; bridge principles are practically es-
sential to determine non-theoretical terms. Some final remarks to avoid
misunderstandings about this conclusion are in order here.
First, I am not saying that all links serve the purpose of determining
non-theoretical terms. A theory T may have links that connect some of its
T -theoretical terms to the terms of other theories. A few examples of this
sort have already been detected. A conspicuous one - which triggered the
whole links idea within the structuralist program - is this. Internal energy
(U) in equilibrium thermodynamics is aT-theoretical function since the
determination of its value depends on assuming the validity of the fun-
damental equation of state for a thermodynamic system. Nevertheless U
is partially linked to a fundamental concept offluid dynamics, viz. pressure
P, through the important bridge principle:
oU
-P=-
oV
(where V is the volume of the system in question).32 Generally speaking,
a T-theoretical term may be linked to terms of other theories; if, however,
the full determination of that term is impossible unless Ts fundamental
laws are presupposed as true, the additional fact that it is linked to other
theories does not make it T-non-theoretical. Presumably, all theoretical
terms of all theories are linked in some way or other to terms of other
theories; but this does not eliminate their theoretical chara~ter.
Secondly, and most important, the condition (LNT) should not be

viewed as a criterion for T-non-theoreticity. The only criterion of T-non-
theoreticity I accept as generally valid is just the negation of the criterion
(CT) (or a more sophisticated version of it): a term is T-non-theoretical iff
it is not T-theoretical in the sense of (CT). The point of condition (LNT)
is to induce the formulation of the metaempirical hypothesis that all non-
-theoretical terms of any scientific theory are essentially linked to some
terms of other theories. To put it more cautiously, normally a T-non-the-
oretical term will satisfy condition (LNT). The rationale for the qualifi-
cation "normally" here is this. All non-theoretical terms of theories re-
constructed in detail so far satisfy (LNT). Moreover, all plausible candi-
dates for non-theoreticity in still unreconstructed, but more or less well
identifiable scientific theories seem to satisfy (LNT) toO.33 Hence, we could
define a "normal non-theoretical term" as one that satisfies not only the
negation of (CT) but also satisfies (LNT). And the question now is: Are
there "abnormal" non-theoretical terms in science? That is, are there some
scientific theories with terms that do not satisfy (CT) nor (LNT) either?
And, if there are such - how are they determined?
This, I think, is a significant question about general epistemology and
about the global structure of science. It is, one may say, the reappearance
of the problem of theoretical terms on a higher (global) level. For, if all
T-non-theoretical terms satisfy (LNT) for some (different) T i , then one
could well extend the usage of the expression "theoretical" and say that
they too are theoretical in the derived sense that there is always some
theory that fully determines them. Which would end up in a sort of a
coherentist view of the totality of science. And, if this is not true, if there
really are abnormal non-theoretical terms, then we may well ask what their
nature is, where they come from, and whether their function is to provide
the ultimate ground for the whole of science - the core idea so cherished
by fundamentalist philosophers since ages ago.
NOTES
1 Hempel [1973], pp. 367-368.

2 Hempel [1973], p. 377.
3 Hempel [1973], p. 370.
4 See Hempel [1973], p. 371.
5 See Hempel [1970], p. 143.
116 C. ULISES MOULINES
6 Hempel [1958], pp. 208-209.

7 Hempel [1958], pp. 208-209, my italics.
8 See also Hempel [1970], pp. 143-144.
9 Hempel [1970], pp. 142-143.

10 Hempel [1973], p. 376.
11 See Hempel [1970], pp. 151, 159f. and still clearer Hempel [1973], pp. 370,376.
12 See Hempel [1973], p. 369.
13 Hempel [1973], just quoted.
14 Hempel [1973], p. 377.
15 Hempel [1970], p. 161.
16 Hempel [1973], quoted above.

17 For more details on this understanding of the problem of theoretical terms see e.g. Sneed
[1971], pp. 27ff.; Stegmiiller [1973], pp. 63ff.; Balzer/Moulines [1980].
18 A succinct (though already incomplete) list of these examples may be found in Sneed's
introduction to the second edition of his work, Sneed [1979], p. XXII. Since then, still more
examples of theories containing theoretical terms in the sense of Sneed have been provided
in so different fields as hydrodynamics (Flematti [1984]), Marxian economics (G. de la Sienra
[1982]) and physiology (Miiller/Pilatus [1982]).
19 See Balzer/Moulines/Sneed [1983] and more specifically Moulines [1984].
20 A word of clarification for the expression "methodologically isolated" may be in order
here. By "theory T, is methodologically isolated from T 2 " I mean that T, is constructed,
used, applied, checked, revised, and so on, without taking care of T2 at all.
21 Unless it is believed that human beings are identical with permanent, unmodifiable souls
- but the sort of opponent I have in mind here is not one that believes that, at least when
he/she is philosophizing.
22 We may be able to strengthen the argument to support the view that this must always be
so; but this is not necessary for the present discussion.
23 Cf. his rather "anti-axiomatic" remarks in Hempel [1970], pp. 151f.
24 The notion of structure species in general comes from Bourbaki [1968], Ch. IV, § I. Its
use for the present model-theoretic purposes is a piece of current structuralist methodology
(though in somewhat modified terminology); see, e.g. Balzer [1982], pp. 273ff.
25 It is a still open question within the structuralist program whether Mp(T) should simply
be defined through the structure species of the models of T or rather be introduced as an
undefinable component of any theory, which, nevertheless, is closely connected to the cor-
responding structure species. The issue may be settled only after the intuitive consequences
are better understood and more real-life examples of empirical theories have been recon-
structed.
26 The present usage of this expression is reminiscent of Hempel's usage of it (see Hempel
[1970], p. 142), which, in tum, is inspired by previous usage by some philosophers of science
of the empiricist tradition. The conceptual differences are, however, to be noted. Whereas the
traditional "bridge principles" were meant to build a bridge from theory to pure experience,
or observation, or something of the kind, my intention is to understand bridge principles as
expressing intertheoretical relations, i.e. as bridges between theories. Curiously enough, Hem-
pel's own conception of bridge principles seems to oscillate between these two different in-
terpretations - as his introductory remarks and examples in Hempel [1970] seem to indicate.
27 Cf., e.g., Sneed [1971], p. 118.
28 The case where there is no full, but only partial coincidence between both kinds of struc-
tures is just a bit more complicated but can also be dealt with within the present framework.
For more details on this particular issue as well as on the general idea of reconstructing
measurement methods model-theoretically see Balzer/Moulines [1980], especially § II.
29 Cf. Ba1zer/Moulines [1980], p. 473. The rationale for this criterion, as well as the technical
difficulties that it implies, are fully expounded in that article.
30 Giihde has convincingly argued that this has always to be the case for theoretical functions
of physics. Cf. Giihde [1983], especially § 5.5.
31 Frequently, there will be methods of determination of a T -non-theoretical term g that are
models of T. But the point is that there should be at least one method of determination of
g that is not a model of T. That is, there should be at least one method of determining g that
does not presuppose Ts laws.
32 For more details on this example, cf. Moulines [1984].
33 This may have been the ultimate ground for Hanson, Feyerabend, and others, when they
put forth the hasty and ambiguous claim of the "theory-ladenness" of all scientific terms.
REFERENCES
Balzer, W.: 1982, Empirische Theorien: Modelle, Strukturen, Beispiele, Vieweg, Braun-
schweig/Wiesbaden.
Balzer, W. and Moulines, C. U.: 1980, 'On Theoreticity', Synthese 44,467--494.
Balzer, W., Moulines, C. U. and Sneed, J. D.: 1983, 'The Structure of Empirical Science:
Local and Global', in Proceedings of the 7th International Congress of Logic, Methodology
and Philosophy of Science, Salzburg.
Bourbaki, N.: 1968, Theory of Sets, Hermann and Addison-Wesley, Paris and New York.
Flematti, J.: 1984, 'A Logical Reconstruction of the Hydrodynamics ofIdea1 Fluids', Type-
script, Mexico.
Giihde, U.: 1983, T-Theoretizitiit und Holismus, Vieweg, Braunschweig/Wiesbaden.
Garcia de la Sienra, A.: 1982, 'The Basic Core of the Marxian Economic Theory', in Studies
in Contemporary Economics I, eds. W. Balzer, W. Spohn and W. Stegmiiller, Springer,
Berlin/Heidelberg.
Hempel, C. G.: 1958, 'The Theoretician's Dilemma: A Study in the Logic of Theory Con-
struction', in C. G. Hempel, 1965, Aspects of Scientific Explanation and Other Essays, The
Free Press, New York.
Hempel, C. G.: 1970, 'On the "Standard Conception" of Scientific Theories', in Minnesota
Studies in the Philosophy of Science, vol. 4, eds. M. Radner/S. Winokur, University of
Minnesota Press, Minneapolis.
Hempel, C. G.: 1973, 'The Meaning of Theoretical Terms: A Critique of the Standard Em-
piricist Construal', in Logic, Methodology and Philosophy of Science IV, eds. P. Suppes,
L. Henkin, A. Joja and Gr. C. Moisil, North-Holland/American Elsevier, Amsterdam/
London/New York.
Moulines, C. U.: 1984, 'Links, Loops, and the Global Structure of Science'. Philosophia
Naturalis, forthcoming.
Miiller, U., and Pilatus, S.: 1982, 'On Hodgkin and Huxley's Theory of Excitable Mem-
branes', Metamedicine 3, 193-208.
Sneed,J. D.: 1971, The Logical Structure of Mathematical Physics, 2nd revised edition, 1979,
Reidel, Dordrecht.
Stegmiiller, W.: 1973, Theorienstrukturen und Theoriendynamik, Springer, Berlin/Heidelberg.
Universitat Bielefeld Received 25 June 1984

Postfach 8640
4800 Bielefeld 1
F.R.G.
ANDREAS KAMLAH
ON REDUCTION OF THEORIES
1. INTRODUCTION
When Hempel and Oppenheim wrote their famous paper on explanation

in 1948 they formulated their concept of explanation in a very general way,
so that it includes both causal explanation and explanation of general
regularities as special cases. They write:
We shall not make it a necessary condition for a sound explanation, however, that the ex-
planans must contain at least one statement which is not a law; for to mention just one
reason, we would surely want to consider as an explanation the derivation of general reg-
ularities governing the motion of double stars from the laws of celestial mechanics, even
though all the statements in the explanans are general laws. (C. G. Hempel 1965, p. 248)
Thus they considered the explanation of regularities or laws of a less fun-

damental character as a deduction of these laws from others, since their
first condition of adequacy Rl says: "The explanandum must be a logical
consequence of the explanans". (p. 247) As nearly any detail of Hempel's
and Oppenheim's first model of explanation, also the claim that expla-
nations of laws and regularities have a deductive character has been chal-
lenged. This has been done most vigorously by P. Feyerabend in his first
article in the Minnesota Studies (1962). His main point was that in most
cases derivations of theories are only approximately valid. Thus the "de-
rived" secondary theories strictly speaking contradict the primary theories
from which they are obtained.
This critique is acknowledged in Hempel's extensive article 'Aspects of
Scientific Explanation':
It should be noted, however, that in the illustrations just mentumed, the theory invoked does
not, strictly speaking, imply the presumptive general laws to be explained; rather, it implies
that those laws hold only within a limited range, and even there, only approximately. (1965,
p.344)
In the last twenty years explanations of single events and of regularities

have been treated more and more as completely separate issues, and I
Erkenntnis 22 (1985) 119-142. 0165--{) 106/85/0220--0 11 9 $02.40
120 ANDREAS KAMLAH
think that the difficulties of the D-N-model can only be solved if one does
not try to find a model which embraces both kinds of explanation. So I
think that philosophers of science have learnt a lot in these two decades,
and that there has been a considerable progress since Hempel's important
papers. On the other hand some features of the concept of reduction which
have been already quite clear for Hempel got lost in the discussion. One
of the most important changes since 1964 was the introduction of set the-
oretical methods first by E. W. Adams (1959) and later by J. D. Sneed
instead of methods applying first order logic. But Sneed's logical recon-
struction for the concept of reduction lacks some of the features which
have already been clearly recognized by Hempel. (J. D. Sneed 1971, p.
216-234; W. Balzer, J. D. Sneed 1977 and 1978). In the subsequent dis-
cussion these features partly remained lost. For Hempel it was clear that
(Cl) reduction involves deduction
(C2) the laws which one wants to explain "hold only within a lim-
ited range"
(C3) they hold "only approximately".
So what one deduces is only something which comes very near to the
explanandum in a limited domain of phenomena. But even if this is so,
something is deduced from the primary theory or the explanans.
I think that these three conditions for reduction are characteristic for
this relation of scientific theories, laws, or regularities. Only the third one,
(C3), has been taken into account by philosophers applying the method
of Sneed to reduction in the seventies (C. U. Moulines 1976, 1980, 1981
and D. Mayr 1976, 1981a and 1981b). The first, (Cl), has only been em-
phasized quite recently by D. Pearce (1982a) while Stegmiiller claims the-
ories to be something different from propositions and thus not capable of
logical relations like implication. Stegmiiller believes that the "non-state-
ment view" of theories removes the conceptual muddle of the discussion,
between proponents and critics ofT. S. Kuhn's incommensurability thesis.
If reduction does not involve deduction it is imaginable that one theory
Til is reducible to T' though they are both incommensurable, and thus not
capable of any logical relation. He says: "Instead of arguing from non-
inferability to nonreducibility we will argue that an adequate concept of
reduction cannot be defined in terms of inference" (1973, p. 249; 1976, p.
ON REDUCTION OF THEORIES 121
216). In this paper I do not want to criticize Stegmiiller's claim. This has
already been done by D. Pearce in a quite convincing way. But I shall give
an account of reduction which does not agree with Stegmiiller's, and these
considerations automatically imply some objections against Stegmiiller's
view. I want to try in this paper also to be aware of the second feature,
(C2), and try to come to a more adequate explicans of reduction or ex-
planation of theories or laws which is at the same time nearer to Hempel's
ideas of twenty years ago, which were not so bad after all.
2. THE EXPLICANDUM OF THE TERM "REDUCTION"
In philosophy of science logical analysis of scientific theories and their

mutual relations and of scientific methods plays an extraordinary role.
Some philosophers have even claimed that philosophy of science is logical
analysis. The most important under the relations connecting different the-
ories in physics, the so called "intertheoretic relations", is the relation of
reduction. Reducing theories to more fundamental ones is a considerable
part of the work done by theoretical physicists, and when I myself was still
a theoretical physicist, I did nothing else than to reduce theories of phen-
omenological nuclear physics talking about the nucleus as a liquid drop
to a microsopic theory which talks about protons, neutrons and their in-
teracting forces. (A. Kamlah 1968, 1971, and J. Meyer, A. KamIah 1968).
For the logical analysis of such an important relation as the reduction
of theories we can apply R. Camap's method of rational explication of
scientific or methodological concepts. According to Camap we have first
to look for the explicandum of the concept, the term already used by
scientists. This explicandum has to be fitted by an explicans in an exact
formalized language. Camap does not demand that the explicandum and
the explicans apply to the same cases or objects, they may deviate con-
siderably from each other. In such cases however we could not claim any
more to have found a really adequate explicans for the explicandum given.
I have discussed elsewhere the problems raised by too much arbitrariness
in explication (A. Kamlah, 1980, p. 33 if.).
In any case the explications of philosophers of science during the last
decades have tried to come as near as possible to their explicanda, and
have not made much use of the freedom recommended by Camap. There-
fore we shall try to find an explicans as close as possible to the underlying
122 ANDREAS KAMLAH
intuition of the scientist. But where do we find the explicandum of the

concept of reduction? Theoretical physicists do not talk of reducing the-
ories; they do not use the same terminology as philosophers of science;
they will usually say that they derive one theory from the other. Such a
terminological discrepancy would be of no harm if physicists did nothing
more than just use another word for "reduction". The application of the
term derivation, however, seems to indicate that physicists make no sharp
distinction between mathematical deductions and physical reductions. I
do not think that they are unaware of this distinction. It is simply unim-
portant for the physicist. If a physicist puzzles mathematicians by one of
his "derivations" making much use of the sponge at the blackbord for
whiping out "small" terms in a sum, he will know very well that he is
doing something rather different from a purely mathematical derivation.
This physical derivation of one theory Til from another theory T' is the
explicandum for a precisely formulated relation of reduction. If the phy-
sicist says that he has derived Til from T', the philosopher of science will
say that the physicist has reduced Til to T'. And he will justify this ter-
minological change by pointing out that these "derivations" of the phy-
sicist are of course no derivations in the sense of the mathematician, and
that the label "reduction" will avoid a terminological muddle.
Such terminological discrepancies between physicists and philosophers
of science are not unusual. If a piece of evidence E confirms a theory T
the physicist will in many cases claim that E "proves" the theory T, using
an expression quite unacceptable for the philosopher. The physicist will
also say that the concept of mass is "defined" by Newton's third axiom
or that electrical charge is "defined" by Coulomb's law, unimpressed by
the objections of the philosopher of science.
The philosopher of science has to take these methodological claims of
the physicists rather seriously, even if it is impossible for him to accept
them without any revision. He is in a difficult situation. On the one hand
he cannot simply take over the often very crude methodological selfinter-
pretation of the scientists. On the other hand he wants to investigate what
the scientist really does, and how he is in part consciously and in part
unconsciously operating methodologically, and for these investigations the
utterances of scientists are his main source - of course supplemented by
the intuitions of the philosopher. Therefore physical derivations, physical
proofs and physical definitions have to be subject matter for the explica-
tions of the philosopher of science even if he is convinced that in all these

cases he is not dealing with proper derivations, proofs or definitions. He
has to try to find explications which apply in all cases in which the physicist
says that he derives, proves or defines something.
Now let us concentrate on the derivations of the physicists and ask for
their characteristic features. If physicists derive theories from others they
at first have always in their mind a certain range of possible intended
applications. If for example the ideal gas law p . V = R . T is derived from
van der Waals' gas law
this is always done for the case of high temperatures T and large volumes
V. Secondly they want only to show that the phenomena which the sec-
ondary or reduced theory Til claims to describe, are only reproduced to
a certain degree of accuracy by the theory which is in most cases identical
with the accuracy of available measurements. Thus we are led back to the
three features (C1-C3) of reduction formulated already by C. G. Hempel.
A (so called secondary) theory Til is derived physically from a (primary)
theory T' iff
(Cl) from some statements made by T' a claim C is logically de-
duced
(C2) C says that for a set of intended physical systems a claim C'
is true
(C3) C' says that the theory holds for these system with a certain
degree of approximation
If such a derivation is to count as an explanation of the secondary theory
Til there are additional conditions to be satisfied. Hempel makes it con-
vincingly clear that the premises of an explanation have to be true as much
as the explanandum. This is his condition (R4) (C. G. Hempel 1965, p.
248). To this condition we have still to add some others connecting the
languages of both theories. E. Nagel has pointed out that we cannot simply
derive one theory from the other if both are formulated in different lan-
guages (E. Nagel 1960, pp. 301-303). There must be a kind of dictionary
124 ANDREAS KAMLAH
which enables us to define the concepts of the secondary theory in terms

of the primary theoretical language. In many cases this relation between
both theoretical languages is an empirical hypothesis. Nagel refers to an
informative example: the interpretation of temperature as kinetic energy
of gas molecules. In section 6 I will formulate some conditions which ex-
press that the "dictionary" is correct.
Until now we have not yet dealt with the logical structure of theories
themselves, or with their reconstruction. But before details of the struc-
turalist account of theories are brought in we should first consider alter-
native explicanda for the concept of reduction.
3. AL TERNA TIVE EXPLICANDA
Physicists try to derive theories from other theories since they are inter-
ested in better explanations of the phenomena. In such cases they are look-
ing for a theory T' which explains more phenomena than a given theory
T" and explains them better. We are thus led to two different explicanda
for the reduction relation. The first of these two will be:
T' explains all phenomena which are explained by T".
This explicandum is very similar to the explicandum of the last section,
but different in an important respect. It does not say that T' explains any
possible set of phenomena E which is explainable by T". It talks only about
the phenomena really given in a certain historical situation. Let E be a
statement which describes all these given phenomena. Then the explan-
andum will lead to the statement
T" f-- E -+ T' f-- E.
That E may be derived both from T' and from T" however does not lead
to any logical relation between T' and T" themselves. It may just be due
to chance that T' f-- E and T" f-- E are both true. If there is to be a
structural relation between the theories T' and T" we have to demand that
in any possible world in which T' holds, any intended phenomenon of T"
is also explained by T'.
We see already how this important change has us brought nearer to the
explicandum of the last section.
Up to now I have not said anything about degrees of explanation. The
theory T' is not yet supposed to explain a piece of evidence E better than
Til. It is quite clear that physicists want to improve the explanation of
given phenomena by deriving a theory Til from a more fundamental, a
deeper or simpler theory T'. We are thus led to a second explicandum:
T' explains the given evidence E better than Til.
Or in the stronger version:
In any possible world in which T' holds, the totality of phe-

nomena E which is intended by Til to be accounted for is
explained in a better way by T'.
If we want to deal with this second explicandum we shall face serious

difficulties. Intuitively, we may know what is meant by a better and more
thorough explanation. But any attempt to reconstruct a degree of expla-
nation will lead us into the jungle of inductive logic. We cannot hope to
obtain useful results after a short investigation of our subject.
We may, therefore, see now the advantage of our original explicandum,
the physical derivation. Though physicists derive theories from other the-
ories since they are interested in the explanation of theories by more fun-
damental or deeper ones, physical derivations by themselves do not involve
any special inductive or explanational considerations which show the su-
periority of the primary over the secondary theories. One could equally
well derive simple theories from theories abounding with ad hoc hypothes-
es. Such a derivation may be uninteresting but is still a physical derivation.
4. THE CONCEPT OF A PHYSICAL THEORY
So far I have characterized t.he explicandum in informal terms. To obtain

an exact counterpart of this intuitively described concept we have now to
introduce a framework for the exact characterization of theories. For this
purpose a simplified version of Sneed's model theoretic scheme shall be
introduced similar to the scheme already applied by Adams. I personally
think that such a scheme is sufficient for the reconstruction of the reduction
relation. To give reasons for this opinion would exceed the scope of this
paper. I cannot present the arguments for it in a few pages.
126 ANDREAS KAMLAH
In modern mathematical axiomatizations one frequently starts with an

ordered set
x = (Db"" D k, It, ..., f,) of sets

which appear in the mathematical theory involved. In some preliminary
axioms the type of the functions f1 ... f, is stated, it is said in which way
from the D1 ... Dk or from sets constructed by their aid the arguments and
values of these functions are taken. By these preliminary axioms the type
of the structure is specified, while the theory in its full scope characterizes
a species of structures; it says which kind of structure - for example an
electromagnetic tensor field - has to be considered as a solution of Max-
well's equations. With W. Stegmiiller (1973, p. 124) and W. Balzer, J. D.
Sneed (1977, p. 198) we shall call the set of all structures having the same
type as the models of the theory M p , the set of possible models for the
theory. A structure x of this type or a possible model x E Mp is a model,
we say x E M, if it is in agreement with the theory. Thus M is the theory
predicate. To know M is equivalent to the knowledge of the theory. So
far we have only talked about abstract mathematical theories and our
reconstruction scheme is still lacking any connection with the real empir-
ical world. Physics however is an empirical science, and therefore we have
to introduce a concept which has the function of a bridge between the real
world and mathematical structures. This has already been done by Adams
who introduced a set Is;;:; Mp of "intended models" (which is not identical
with Sneed's so-called set) (Adams 1959, p. 261). I shall call this set W. It
is the set of all structures or possible models x E Mp which are descriptions
of real physical phenomena, systems or processes if the basic sets Dh ... ,
D k , and the functions flo ... , f, are subjected to the standard physical in-
terpretation of the theory. I want to call the set W the set of real descrip-
tions. W plays the same role as w, the description of the actual world in
customary semantics for modal logics. As long as we have not available
the symbol w for the actual world, we may interpret propositions of a
modal language like Dp or Op, i.e. say that p is necessary or possible, but
not that p is true. For the same reason we need the symbol W in our
scheme for theory reconstruction. We can say a lot about logical relations
of theories without it, but not that a theory is factually true. I have now
to explain how such an empirical claim can be made. In costumary se-
mantics of modal logics every proposition p is represented by the set lip II
of those possible worlds in which it is true. That p is true (in the actual
world) is expressed by W E lip II·
We have to find the analogous expression for the truth of a scientific
law in the real world. The physical processes which are described by pos-
sible models x E Mp for which x E Wholds are parts of the real world. We
may consider them as spatio-temporal parts of a total possible model w
embracing the whole space-time, for instance. But this would not be ap-
propriate if the basic sets Dl ... Dk of the possible models are not space,
time or space-time regions but rather sets of particles. Therefore I will
proceed in a more abstract way. I define a relation x 0 y (x is part of y)
in the following way:
(Db ... , Dk,fb ... , fi> 0 (Dl' ... , D",fl , .. . j!)
(DI) iff
Dl £ Dl and ... and Dk £ Die, and thefi agree with thefi in
the regions where their domains overlap.
In the special case that Db D2 are spatial or temporal regions, the relation
x 0 y means something like "x is a space-time part of y". Now I can
introduce the real world described in the language of a certain theory w
and identify the elements of W as parts of w:
XE W+-+xD wor
(D2)
W = {x I x 0 w}
Next we have to define truth for our reconstruction scheme. Take as an
example the proposition: p: "The mass of Jupiter is 318 times the mass of
the earth." I can represent this proposition by the set of all possible systems
or processes of celestial mechanics which are compatible with it. Certainly
a description of the double star (-g ursae maioris is compatible with it
whatever the values of the masses of its components may be. Also descrip-
tions of stars at any other place, no matter if they exist or not, are com-
patible with p.
Only a mechanical system at the same place with bodies of different
masses would be incompatible with p. We may call the set of possible
models which are compatible with p - as was done before - lip II. A prop-
osition is true if it is compatible with all parts of the real world x 0 w.
This seems to be quite evident. Thus p is true iff 1\ x (xD w -+ x E lip II)
128 ANDREAS KAMLAH
or iff W ~ lip II· So far we have specified how propositions can be repre-
sented by sets of possible models or descriptions of physical processes and
how we can express the truth of these propositions. Physical theories in
their mathematical parts are propositions or conjunctions of propositions;
they formulate conditions to be satisfied by physical processes. The de-
scriptions of those processes which are allowed by the theory are the
models x E M of the physical theory. In a set theoretical reconstruction
scheme as we use it for physics here, the mathematical part of the theory
is represented by its model set M. Thus we might consider a physical theory
as a triple
T = (Mp , M, W) with
M~ M p, W~ Mp
with the interpretation given to these symbols in the last pages. We are
tempted now to express the empirical truth of a theory T with the model
set M and the set of real descriptions as:
W~M.
This is perhaps E. W. Adams' account of physical theories (1959). I cannot

decide if we may interpret him like this, since his remarks about theories
are rather brief (see 1959, p. 260).
If we would accept this picture of physical theories which might have
been Adams' view of physical theories in 1959 we would miss one very
important aspect of it. This is simply not the way that theories are applied
to the world. If I write down the law of ideal gases P . V = R . T, I do not
claim that this will be true for any portion of matter. It is true only for
ideal gases. If I write down Ohm's law for conductors U = R . I, I do not
also claim that this is true for any physical body. It is just true for Ohmian
conductors. How do I know what ideal gases or Ohmian conductors are?
I know that under certain conditions hydrogen behaves like an ideal gas,
under more restricted conditions also carbon dioxide, and so on. We have
a practical knowledge of which substances behave under which conditions
as ideal gases and which do not. Thus P . V = R . T defines the concept of
an ideal gas and U = R . T defines the concept of an Ohmian conductor.
The physical knowledge about these species of objects is contained in the
list of kinds of substances to which the concept applies.
The theories of Ohm and of ideal gases are no exceptions. Maxwell's
equations will be invalid in space-time regions in which non-linear con-

ductors are present, such as transistors or non-linear magnetic materials
(like any sort of iron under the influence of a strong magnetic field).
Hooke's law will be invalid for very strong forces. And thus physicists
believe in general that there are certain limits for the validity and applic-
ability of a physical law. I want to call the set of possible models x E Mp
which are within the limits of applicability V, the domain of validity of the
theory. We thus obtain a more complicated scheme for the reconstruction
of theories as we have discussed before.
A physical theory T is an ordered set (Mp, M, W, V) with M s;;; M p,
W s;;; M p, V s;;; Mp and the interpretation given so far for M, W, V. The
formulation of empirical truth also will be more complicated for the theory
in the improved reconstruction. We have:
A physical theory T = (Mp , M, W, V) is true
iff
for all real physical processes of the kind T talks about, x E W,
it holds that: If x is in the domain of validity of T, x E V, x
will also have the mathematical structure prescribed by T, by
a model of the theory, x E M
or in terms of first order logic:
1\ x(x E W 1\ X E V -+ X E M)
or in set theoretical symbols:
WnVs;;;M
There is still something not quite correct in this description of the empirical
claim by a physical theory. In most cases physicists do not believe that
their theories are exactly true within their domain of validity V. They know
that a theory which describes the world exactly is not available, at least at
the present time, and that the theoretically predicted data can only ap-
proach the observed data within the limits of a certain degree of accuracy.
How can we account for this peculiarity of empirical claims? We may,
following G. Ludwig (1978, p. 50 fr.) and C. U. Moulines (1976), introduce
an uncertainty set U s;;; Mp x Mp. The set Uis a two-place relation on the
set of possible models M p , which coordinates to each possible model x a
neighborhood of possible models y which cannot be distinguished from x
130 ANDREAS KAMLAH
within a certain given degree of accuracy. For denoting different degrees

of accuracy we may introduce a parameter e and obtain for each value of
e another relation U•. It is reasonable to postulate that for the numerical
parameter e holds
e ~ e --+ U. ~ U.; x = y --+ <x, y) E U.
We can now express easily an empirical claim of a theory within the degree
of accuracy e:
/\X(XE W /\ XE V--+ Vy«x,y) E U. /\ YEM))
This formula can be read as follows:
For all real processes x E W holds: if x is within the domain of validity
V, then there is a process y indistinguishable from x within the degree of
accuracy e which is a model of the theory T, y E M.
For the sake of convenience we may introduce an abbreviation AlB;
which as already been used by R. Camap (1958, p. 114)
(D3) AlB = {x I Vy (y E B /\ <x, y) E A)}
Applying this new symbol we obtain for the empirical claims within the
degree of accuracy e:
W n V ~ U.IM
Thus our concept of a theory has still become a bit more complicated. A
physical theory T now has the structure of a <Mp , M, W, V, U.) with
M ~ Mp; W ~ Mp; V ~ Mp; U. ~ Mp x Mp.
The interpretation of this structure is the one just given.
5. THE PART-OF RELATION OF POSSIBLE MODELS
We have to discuss now some structural attributes of the part-of relation

o which we need in the next section. This relation has a certain structure
which can be described by some axioms. The type of this structure is as
follows:
X is a part-oj structure iff
1. X = <Mp , N, D)
2. N, 0 have the following types
N ~ pot (Mp); D ~ Mp x Mp
ON REDlJCTION OF THEORIES 131
3. for (Mp , N, D) the following axioms

Al-A4 are valid
These axioms are now to be introduced and discussed. The first two axioms
are simple, 0 is a transitive order with largest elements:
(AI) x D yAy 0 Z --+ x 0 z
(A2) x E Mp --+ Vy (y E Mp A X 0 Y A Iv (z E Mp A YD Z --+
y = z))
The next two axioms connect the part-of relation with the set N of physi-
cally relevant sets of possible models. These are those sets lip I which cor-
respond to propositions p. They contain descriptions of those physical
processes which are compatible with a proposition p. This interpretation
leads to certain further postulates:
(A3) A EN --+ (x 0 yAy E A --+ X E A)
The content of A3 is:
"If a physical process y is compatible with a proposition p (with A = I p II)
and contains another process x as its part, the latter process x is compatible
with p as well, x E lip I or x E A".
The next and last postulate is:
(A4) A E N A X E A --+ VY (y E A A X D y A
Iv (z E Mp A Y 0 Z --+ Y = z))
We read it as follows:
"If a physical process x is compatible with a proposition p with x E lip I
there will be a maximal process (which is only part of itself, and) which
is also compatible with p".
It is important to note here that the mathematical part m of a physical
theory is a proposition too. Thus the model set M = I m II describes the
set of physical processes which are consistent with m. Therefore we will
assume that MEN.
Maximal processes correspond to possible worlds. If a physical process
is compatible with a proposition p there must be a possible world to which
it belongs which is compatible with p as well. So much for the character-
ization of our part-of structure. The set of maximal processes (possible
worlds) we shall call M w:
(D4) Mw = {xix E Mp A I\y (y E Mp A X 0 Y --+ x = y)}
132 ANDREAS KAMLAH
We have now to prove in our little theory a theorem which we need for
the reconstruction of the reduction relation.
AeN A BeN-+
(Ll)
(Mw f"'I A ~ Mw f"'I B +-+- A ~ B)
The proof in one direction of the biconditional is trivial. We have to show
however that
Mw f"'I A ~ Mw f"'I B -+ A ~ B
This is done by application of D4, A3 and A4.
Let us assume that
(1) Xe A and
(2) M w f"'I A ~ M w f"'I B.
From A4 we may conclude that (1) yields with D4:
(3) There is a process y with y e Mw f"'I A and x 0 y.
From assumption (2) and from (3) we obtain:
(4) For y also holds: y e Mw f"'I B.
Finally applying A3 we get
(5) x e B.
Thus we have shown that (5) can be obtained from (1) and (2), Q.E.D.
6. REDUCTION
It is now quite easy to find an adequate expression for the reduction of a

theory T" to another theory T'. Usually one derives from the factual claim
of the primary theory T' that the secondary theory T" is true in a certain
domain of validity within certain limits of accuracy. That is exactly the
claim which can be made with T", and we have found a claim of this kind
in section 4. Therefore reduction of T" = (M~ , M", W", V", U") to
T' = (M~, M', W', V', U') means logical derivation of
V" f"'I WIt ~ U"/M" from V' f"'I W' ~ U'/M'
We have thus defined Hempel's condition (Cl) of the first section that
reduction involves deduction. Conditions (C2) and (C3) which say that the
laws which one wants to explain hold only within a limited range and are
only approximately valid, are already built in into our explanans of the
empirical claim.
But we have not yet attained our goal, we have only taken the right
course. First, T' and T" may be formulated in different languages. We can
derive the factual claim of T" from a kind of translation of the factual
claim of T' into the language ofT". Secondly we have to find an adequate
expression for the fact that one claim implies the other with necessity, i.e.
in any possible world, as has already been pointed out in section 3. The
first difficulty is solved by introducing a reduction relation p, the inverse
of which, p, maps any possible model x E M~ on a possible model
y E M~. Thus every description of a process in the secondary language of
T" can be represented by a description of a process in the primary lan-
guage of T'. The descriptions in the primary language can be more detailed
and contain predicates for attributes which are not describable in the sec-
ondary language. Therefore the relation p is a function, while p may be
not. There may be more than one possible model x E M~ and y E M~
which correspond to one possible model z E M~ by p(z, x) and p(z, y).
Thus P is a function:
This fact leads to a lemma which will become important later in this sec-
tion:
(L2) If P : M~ - M~, A ~ M~ and B ~ M~, then A ~ plB is
equivalent to piA ~ B
Proof Definition (D3) yields: A ~ PIB is equivalent to
(1) I\x Vy (x E A - p(Y, x) AyE B)
piA ~ B is equivalent to
(2) I\x I\Y (p(y, x) A X E A - Y E B)
Since p is a function, we may replace (1) by (2). and vice versa.
We may add without proof a third lemma which is a purely set theo-
retical consequence of definition (D3):
134 ANDREAS KAMLAH
(L3) If p £ M~ x M~, A £ M~, B £ M~ then

p / (A u B) = p/A u p/ B
The relation p has to be consistent with the part-of relation D. If x is the
description of a physical process and y the description of another physical
process in the same language, such that the former is part of the latter,
x D y, then the same relation has to obtain for these processes in another
language:
(CCI) x' 0 y' /\ p(x", x') /\ p(y", y') --+ x" 0 y"
(CC2) x" 0 y" /\ p(y", y') --+ Vx' (p(x", x') /\ x' 0 y')
(CC3) x" 0 y" /\ p(x", x') --+ Vy' (p(y", y') /\ x' 0 y')
These three conditions I call consistency conditions (CCI-CC3). The con-
sistency conditions imply that descriptions of maximal processes have to
be mapped on descriptions of the same kind.
x E M'w /\ p(y, x) --+ y EM':.,
These maximal processes are the possible worlds of modal logic under the
aspect of a certain physical theory. One of these possible worlds is the
actual world. The description of the actual world w' in the primary lan-
guage has to be translated by p into the description of the same actual
world w" in the secondary language.
(RC) p(w", w'), or what is the same: pew', w")
I want to call this condition the reality condition (RC) for the relation p.
(CCI-CC3) and (RC) correspond to the above mentioned conditions (see
section 2) which have been demanded by E. Nagel and state that the dic-
tionary between the primary and the secondary languages is correct.
Applying definitions (D2) and (D3) and Wand p we may derive from
(RC) and from p £ M~ x M~:
W' = p/W
So far we have dealt with our first task. We have found a set-theoretical
relation which connects the descriptions of processes in two different lan-
guages. We have now to find a formulation for condition (CI) for valid
reductions which says that the secondary factual claim is a necessary im-
plication of the primary factual claim.
That means:
In any possible world in which the primary claim is true the
secondary claim will be true too.
In order to handle them easier we write the factual claims of the theories
T' and T" in a slightly different way: We may, introducing the complement
A * of the set A, write instead of
Wn V s;;; U/M
the alternative expressions
W s;;; V* u U/M or {xix D w} S;;; V* u U/M or
/\x (x D w - X E V* u U/M)
Both claims are formulated in different languages. The possible worlds in

both languages are represented by their descriptions U E M'w and v E M':.,
which are related by the relation p. Therefore the condition (CI) finally
becomes
(Cl*) For any pair of descriptions U E M'wand v E M':., of a possible
world which are related by p(v, u) the factual claim made by
the primary theory T' for u implies the factual claim made by
the secondary theory T" for v.
I want to add to (CI *) an auxiliary condition (A) which will allow us to
arrive at a fairly simple expression for our explicans. It states that the
domain of validity of the secondary theory T" is included in the domain
of validity of the primary theory T'. Condition (A) is a very plausible one.
Nobody would expect that one can obtain from a theory laws which apply
to physical processes for which the original theory does not claim to apply.
Thus we may assume:
(A) /\x /\y (p(y, x) /\ Y E V" - X E V')
Or in a more concise notation:
(A) j5/V" S;;; V' or the equivalent condition V'* s;;; j5/V"*
(The equivalence is due to lemma (L2.» From these assumptions, (CI *)

and (A), we would like to derive a formula which expresses the most im-
portant result of the logical reconstruction in this paper:
136 ANDREAS KAMLAH
(I) V'* u U'/M' S;; p/(V"* u U"/M")

The formula says that if a physical process x is allowed by the primary
theory T' its reproduction p(x) is also allowed by the secondary theory
T". (It is clear that "x is allowed" means here "if x is in the domain of
validity, x is an approximate model".)
The set theoretical inclusion (I) replaces the inadequate inclusion
(B) M' S;; p/M"
of Adams, which - as 1 want to make clear in the last section - usually
does not hold in the cases considered to be reductions or physical deri-
vations by scientists. We will also see that Sneed's explication shares this
mistake with Adams' condition (B).
Unfortunately 1 was only able to prove formula (I) for the case in which
U' is the identity, i.e. for relations U' for which
I\x I\Y «(x, y) e U' ~ x = y)
is valid. 1 do not know if the proof can be extended to the general case.
But even the limited proof is of considerable value, since it gives some
evidence for (I) using very plausible assumptions. Neither Adams nor
Sneed made any attempt to derive the corresponding condition (B).
Thus the following theorem is formulated and proved:
(T) Let us presuppose for the theories T' = <M~, M', W', V',
U') and T" = <M;, M", W", V", U'), for the relation
p S;; M; x M~ and the part-of structures <M~, N', 0') and
<M;, N', 0") that the following conditions hold:
(a) M' e N' and M" e N"
(b) V'* S;; p/V"* which is (A)
(c) p: M~ -+ M;
(d) I\x I\Y
(xO'Y -+ p(x)O" p(y» which is (CCI)
(e) condition (C22) for M~, M; and p,
then the following two relations are equivalent:
(E) I\u I\v (u eM:" 1\ v eM:; 1\ p(v, u) -+
({xlxO'u} n V' s;; M' -+ {yIyO"v} n V" s;; U"/M"»
and
(1*) V'* u M' S;; p/(V"* u U"/M")
Proof· The implication (E) -+ (1*) is shown by an indirect proof. In order

to derive a contradiction we make the hypothesis that (1*) does not hold.
Then
(1) there is an x E M~ such that x E V'* u M' and
p(x) ¢ V"* u U"/M".
The possible model x is either an element of V'* or of M':
If x E V'*, we may obtain from (b) p(x) E V"*, and therefore
p(x) E V"* u U"/M" which contradicts assumption (1).
If x E M' then there is due to (a) and definition (D4) a description of a
possible world U EM:" with x 0' U and
(2) I\z (z 0' U -+ Z EM')
which implies
(3) I\z (z 0' U -+ Z E V'* u M')
From (3) we obtain using (E):
(4) N (y 0" p(u) -+ y E V"* u U"/M").
From condition (d) we obtain that
(5) p(x) 0" p(u)
and from (4) and (5) we get
(6) p(x) E V"* u U"/M"
which contradicts the assumption (1). Evidently we have derived from the
hypothesis (1) for both cases a contradiction, and thus established the
proof of (E) -+ (1*).
That (1*) implies (E) is shown more easily. Let us assume that
(7) {xix 0' u} !;; V'* u M'
Then we may derive from (1*) that

(8) {xix 0' u} !;; p/(V"* u U"/M")
which implies according to lemma (L2)
(9) p/{xlx 0' u} !;; V"* u U"/M"
138 ANDREAS KAMLAH
From (e) or (CC2) we obtain

(10) {yly 0" p(u)} S;;; p/{xlx 0' u}
and by application of this equation to (9):
(11) {yly 0" p(u)} S;;; V"* u U"/M"
The last line has been derived from (7) and (1*).
Therefore we have shown that (1*) implies «7) -+ (11)) which is equiv-
alent to (E). Thus the proof of theorem (T) is completed.
By application of the lemmas (L2) and (L3) and purely set-theoretical
theorems one may easily show that (E) and (A) are equivalent to (A) and
(E*) V" n p/M' S;;; U"/M".
This statement says that within the range of validity V" of the secondary
theory T" the translated model set M' of the primary theory T' is included
in the set of approximate models of the secondary theory T". Statement
(E*) looks much simpler than (E) and may be used instead of it.
We are now ready to state the results of our rational reconstruction. We
should first formulate a definition for the reduction of one theory to an-
other by a given reduction relation p. If we neglect that we have proved
theorem (T) only for a special case we may say that we can give three
nearly equivalent formulations for our definiens:
If T' = <M~, M', W', V', U) and

T" = <M~, M", WIt, V", U") are physical theories, then
p S;;; M~ x M~ reduces T" to T' iff
1. P is a function
2. conditions (CCI-CC3) hold
3. V" S;;; pV'
4a. W' n V' S;;; u' /M' implies with necessity that
WIt n V" S;;; u" /M" or alternatively
4b V'* u u' /M' s;;; p/( V" * u u" /M")
or
4c V" n p/U'/M' S;;; U"/M"
This is, however, not what has been discussed in section 2, where we dealt
with reducibility. Therefore we have still to define "T" is reducible to T'''.
We can do this now with the aid of the former definition:
ON RE"DUCTION OF THEORIES 139
A physical theory T' = <M~, M', W', V', U') is reducible to

a theory T" = <M~, M", W", V", U") iff the empirical
claims of T' and T" are true, there is a relation
p: M~ x M~ which reduces T" to T', and which satisfies (RC):
p/W' = W".
The truth of the empirical claims and of (RC) meet the requirements made
by Hempel and by Nagel, which were already discussed in section 2.
So far our task is completed, and we can start to compare our results
with the older ones obtained by E. W. Adams and J. D. Sneed.
7. ADAMS' AND SNEED'S ACCOUNT COMPARED WITH

THE PRESENT ONE
The goal of this paper was to show that the reduction relation is a quite
simple consequence of the formulations given to the factual claims of the
primary and the secondary theory. It is but an inclusion of two sets, the
sets of possible models which are allowed by both theories. There is noth-
ing mysterious about this relation which expresses that the factual claim
of the primary theory T' translated into the language of the secondary
theory T" implies logically the factual claim of T". This result was ob-
tained by paying attention to the explicandum of reduction, which is usu-
ally called "derivation of theories" by physicists.
My explicans differs considerably from those of J. D. Sneed and E. W.
Adams, and from those of philosophers who have tried to improve these
two. In order to discuss these differences I should first give a brief account
of Adams' and Sneed's reconstructions. Adams presents his rational re-
construction on three pages (1959, p. 259-261). It may be due to this very
short exposition that the meaning of Adams' "set of intended models" I
is not quite clear. Therefore we have to discuss two possible interpretations
of 1. The first, which seems to me to be more likely to agree with Adams'
intention, is in my terms
1= W.
(Elements of I are the descriptions of real physical processes.)
The second is
1= Wn V
140 ANDREAS KAMLAH
(The elements of I are the descriptions of real physical processes, within

the region in which the theory is considered to be valid.)
To these two interpretations correspond two different concepts of reduc-
tion. Adams' explicans of reduction is in terms of I and M for the primary
theory T' = (M', I') and the secondary theory Til = (Mil, !"):
{J/!" ~ I'; M' ~ {J/M"
(The relation p is the same as in the last section.) If we insert into these
two set-theoretical inclusions the two interpretations of I, we obtain for
the first interpretation
(la) W" ~ p/W'; (lb) p/M' ~ Mil
and for the second
(2a) W" n V" ~ p/(W' n V'); (2b) p/M' ~ Mil
Both versions of the logical reconstruction are objectionable.
I will discuss first of all the relation (lb) or (2b) between the model sets
M' and Mil. In many cases of reduction the mathematical part of the
secondary theory has models which contradict the mathematical parts of
the primary theory. These models, however, do not belong to the range of
validity of the secondary theory Til. The theory of ideal gases, for example,
has models for low temperatures and pressures which do not fit into van
der Waals' theory at all. But it is quite clear to every physicist that these
models do not belong to the range for which the theory of ideal gases is
applicable. Other examples are descriptions of mechanical processes in-
volving superluminous velocities which contradict the theory of relativity,
and light rays without dispersion in geometrical optics which contradict
wave optics. I have tried to replace (lb) or (2b) by
(E*) V" n p/M' ~ U"/M"
which does not show any more the deficiencies of (lb) or (2b). Another
point of criticism which has already been made in section 4 is that Adams'
concept of a theory does not account for the approximative character and
limited applicability of the factual claims of physical theories.
Sneed's concepts of physical theory and of reduction are much more
complicated than Adams'. But we may criticise them for the same reasons,
since Sneed's improvements do not change the mentioned shortcomings
of Adams' reconstruction. Sneed's explicans boils down to Adams' in the

second interpretation if we do not make the distinction between theoretical
and observational terms, and if the physical theories involved do not con-
tain any "constraints" in Sneed's terminology. Actually these simplifica-
tions should be justified. It would be necessary for me, however, to write
a second paper if I wanted to explain why I think that "constraints" are
completely superfluous and unnecessary for the reconstruction of physical
theories. The distinction between theoretical terms and observational
terms is meaningful in some sense, but I do not consider it to be important
for the investigation of the concept of reduction. Anyway, even if one tries
to replace Adams' reconstruction by the more complicated one of Sneed,
one has not yet repaired what is wrong with it. Sneed's concept inherits
the shortcomings of its predecessor which I have criticized.
Some of the deficiencies of Sneed's concept of a physical theory have
been removed by C. U. Moulines and D. Mayr. But I do not think that
their accounts are yet sufficient, as has been mentioned in section 1. As
long as Sneed's "set of intended applications" I is not split into the two
components Wand V, into the set of real descriptions and the domain of
validity, it remains essentially confused.
It goes without saying, however, that the ideas formulated in this paper
are inspired by Sneed's and that I am still very much indebted to what is
called "structuralism" by W. Stegmiiller. But from time to time it is ex-
tremely helpful to remind oneself how a discussion in philosophy of
science, as for example the one of reduction, has started some decades ago,
and to read the old papers written by E. Nagel, P. Feyerabend, and C. G.
Hempel. One will find there very useful and important hints for the present
study of the subject. I have already mentioned the conditions CI-C3 in
section 1. Notable, however, is also the careful attention payed to the
explicandum of reduction by these authors. The results of these old dis-
cussions should not get out of sight. We should try to remember the virtues
of "classical" philosophy of science to which C. G. Hempel has contri-
buted so much, and to apply to a scientific concept the most simple, most
suitable, and most economical logical reconstruction which we can
achieve.
REFERENCES
Adams, E. W.: 1959, 'The Foundation of Rigid Body Mechanics and the Derivation of its
142 ANDREAS KAMLAH
Laws from those of Particle Mechanics', in L. Henkin, P. Suppes and A. Tarski (eds.),
The Axiomatic Method, North-Holland Pub!. Comp., Amsterdam.
Balzer, W. and Sneed, J. D.: 1977, 'Generalized Net Structures of Empirical Theories, 1',
Studia Logica 36, 195-211.
Balzer, W. and Sneed, J. D.: 1978, 'Generalized Net Structures of Empirical Theories, II',
Studia Logica 27, 167-194.
Carnap, R.: 1958, Introduction to Symbolic Logic and its Applications, Dover, New York.
Hartkiimper, A. and Schmidt, H.-J. (eds.): 1981, Structure and Approximation in Physical
Theories, Plenum Pr., New York/London.
Hempel, C. G.: 1965, Aspects of Scientific Explanation and other Essays in the Philosophy of
Science, Free Press, New York.
Kamiah, A.: 1968, 'An Approximation for Rotation-Projected Expectation Values of the
Energy for Deformed Nuclei and a Derivation of the Cranking Variational Equation',
Zeitschrift fur Physik 216, 52--M.
Kamiah, A.: 1971, 'An Approximation for the Rotation-Vibration Transition Region', Nu-
clear Physics A 163, 166-192.
KamIah, A.: 1976, 'An Improved Definition of "Theoretical in a Given Theory"', Erkenntnis
10, 349-359.
Kamiah, A.: 1980, 'Wie arbeitet die analytische Wissenschaftstheorie?' Zeitschriftfiir allge-
meine Wissenschaftstheorie 11, 23-44.
KamIah, A. and Meyer, J.: 1968, 'Derivation of the Cranking Equations from the Ker-
man-Klein Formalism', Zeitschriftfur Physik 211, 293-303.
Ludwig, G.: 1978, Die Grundstrukturen einer physikalischen Theorie, Springer, Berlin/Hei-
delberg/New York.
Mayr, D.: 1976, 'Investigations of the Concept of Reduction, 1', Erkenntnis 10, 275-294.
Mayr, D.: 1981a, 'Investigations of the Concept of Reduction, II', Erkenntnis 16,109-129.
Mayr, D.: 1981b, 'Approximative Reduction by Completion of Empirical Uniformities', in
A. Hartkiimper and H. J. Schmidt (eds.) 1981, pp. 55-70.
Moulines, C. U.: 1976, 'Approximate Application of Empirical Theories: A General Expli-
cation', Erkenntnis 10, 201-227.
Moulines, C. U.: 1980, 'Intertheoretic Approximation: The Kepler-Newton Case', Synthese
45, 387-412.
Moulines, C. U.: 1981, 'A General ~heme for Intertheoretic Approximation', in A. Hart-
kiimper and H. J. Schmidt (eds.) 1981, pp. 123-146.
Nagel, E.: 1960, 'The Meaning of Reduction in the Natural Sciences', in A. Danto and S.
Morgenbesser, Philosophy of Science, World Pub!. Comp., New York, pp. 288-312.
Pearce, D.: 1982a, 'Logical Properties of the Structural Concept of Reduction', Erkenntnis
18, 307-333.
Pearce, D.: 1982b, 'Stegmiiller on Kuhn and Incommensurability', British Journal for the
Philosophy of Science 33, 389-396.
Sneed, J. D.: 1971, The Logical Structure of Theoretical Physics, Reidel, Dordrecht.
Stegmiiller, W.: 1973, Probleme und Resultate der Wissenschaftstheorie und Analytischen Phi-
losophie, Vo!. 2: Theorie und Erfahrung, Zweiter Halbband: Theorienstrukturen und Theo-
riendynamik, Springer, Berlin/Heidelberg/New York.
Stegmiiller, W.: 1976, The Structure and Dynamics of Theories, Springer, Berlin/Heidelberg/
New York.
Universitiit Osnabriick Manuscript received 15 June 1984

Fachbereich Kultur- u. Geowissenschaften
Postfach 4469, 4500 Osnabriick
F.R.G.
HILARY PUTNAM
REFLEXIVE REFLECTIONS
Today conferences on cognitive science regularly feature the presence of

philosophers in large numbers. These philosophers have very different
visions of the functioning of the brain/mind. Jerry Fodor and Daniel Den-
nett have represented l the different schools of thought with the aid of a
geographical metaphor: in the metaphor there is an "East Pole" which is
M.LT., and views which are less "propositional" or less "digital" (or both)
than M.LT. views are described as being more or less "West Coast" or
simply "West", according to the extent of the deviation from M.LT. Or-
thodoxy. Cute as this is, it is also seriously misleading. For M.LT. itself
is not a unified entity: if (as Dennett suggests) Jerry Fodor is the Pope of
the High Computationalist Church, Noam Chomsky shows alarming signs
of heresy! And Fodor himself has many different views, and the contro-
versies which separate him from, say, Kosslyn, are not at all the same as
the controversies which separate him from Searle, or as the controversies
which separate him from Dreyfus.
The first dogma of the High Computationalist Church, as represented
by Fodor, pictures the mind as "computing" in a formalized language with
a classical or "Fregian" quantificational structure. In addition, it believes
that standard belief-desire explanations (practical syllogisms, in the Aris-
totelian sense) correspond (more or less step by step) to reasonings that
take place in "Mentalese" (as the hypothetical "language of thought" is
called). But this second dogma of High Computationalism is independent
of the first dogma. Finally, Fodor's book The Language of Thought sug-
gested that the "contents" which are believed or desired by a human being
correspond to the same formulas in the language of thought no matter
what natural language the particular human believer-desirer happens to
speak. Again, this third dogma of High Computationalism, is independent
of dogmas one and two. It would be nice to report that the philosophers
at these conferences regularly succeed in distinguishing these three crucial
dogmas, but the sad fact is that usually the atmosphere is "all or none":
either the brain will compute in a way that satisfies all three dogmas, or
Erkenntnis 22 (1985) 143-153. 0165-0106/85/0220-0143 $01.10

144 HILARY PUTNAM
it will compute more like an "analog" device than a "digital" one. Or so

the disputants seem to believe.
Now, I am not myself terribly interested in guessing what future science
will discover. I do not believe that Peter Hempel regards it as the function
of a philosopher to become a heated partisan of one side or the other in
an ongoing scientific controversy. Here (in what I believe to be the spirit
of Hempel) I will rather try to see what philosophical bearing work by
computationalist psychologists and Artificial Intelligence researchers
might in principle have. (When I say "in principle", I don't, of course,
mean "independently of all possible changes in our conceptual scheme".
I have argued elsewhere that the philosopher does not have to choose
between claiming an apodictic status for his work and claiming merely to
be making empirical predictions or guesses.)
I want to begin by proving a theorem. The theorem says that if there is
a complete computational description of our own prescriptive competence
- a description of the way our minds ought to work, where the ought is
the ought of deductive logic or inductive logic - then we cannot come to
believe that that description is correct when our minds are in fact working
according to the description. By "inductive logic" I mean simply the theory
of non-demonstrative inference, or rather of demonstrative and non-de-
monstrative inference - for technical reasons, it is useful to treat deductive
logic as a part of "inductive logic". (Noam Chomsky objects to the notion
of "induction" - like Karl Popper he doesn't believe there is any such
thing, though not for Popper's reasons - but even Noam doesn't deny that
we have a capacity to reason non-demonstratively.) Like everything else,
this theorem can be viewed either optimistically or pessimistically. The
optimistic interpretation is: Isn't it wonderful! We always have the power
to go beyond any reasoning that we can survey and see to be sound. Reflexive
reflection cannot totally survey itself. The pessimistic interpretation is: How
sad!
The proof is not hard. For the mathematical case, consider the totality
of functions which are provably recursive according to M, where M is a
recursive description of our prescriptive mathematical competence. Add
a subroutine to M which lists separately the "index", or computational
description, of each partial recursive function that M ever proves to be
total (as M is allowed to run). Let [D be the "diagonal function" corre-
sponding to this effective list of total recursive functions (that is, the func-
REFLEXIVE REFLECTIONS 145
tion whose value at the argument n is one greater than the value of the
nth function in the list of provably recursive functions generated by M on
the argument n). A mathematician (even a beginner) can prove infinitely
many functions to be general recursive (total partial recursive) functions.
So if M is a reasonable simulation of prescriptive human mathematical
competence, there will be infinitely many functions that M can prove to
be general recursive, and hence infinitely many computational indices list-
ed by the subroutine we added to M. In this case, the diagonal function
[D will itself be total, as is easily seen. So, if we were able to prove by
intuitively correct mathematical reasoning that M is a sound proof pro-
cedure, we would also be able to prove that [D is a total recursive function.
But this proof will not itself be one which is captured by any proof in the
proof-scheme M, unless M is inconsistent! [If this proof were listed by M,
then [D would itself be a function listed by the subroutine (say, the kth).
Then, from the definition of [D, we would have [D(k) = [D(k) + I! Note that
this is a proof of the G6del Incompleteness Theorem.] Thus no formal-
ization of human mathematical proof ability can both be sound and such
that it is part of human mathematical proof ability to prove that soundness.
The more important case for our purposes is the case of our full pre-
scriptive competence: our competence to think thoughts that (from the
point of view of a normative notion of justification) are the ones we should
think on given evidence. In a paper2 I published back in 1963, I showed
that if the relation of confirmation in an inductive logic I is recursive, then,
given the computational description of that relation one can effectively
produce another inductive logic I' which is more adequate than I, in the
sense that by using I' one can discover every regularity that I would enable
one to disc~ver and additional regularities that I would not enable one to
discover, even if the data went on agreeing with them forever, besides.
(Subsequently a number of related theorems were proved by other workers
in the field of recursive inductive logics.) Again, the moral is that if I is
such that one could discover that I is sound by an intuitively correct (de-
monstrative or non-demonstrative) argument, then one would be justified
in using not only I but 1', and hence in regarding a certain regularity as
well confirmed (if the evidence supported it for a long time), even though
the prediction that the next instance one observes will conform to the
regularity is not "confirmed" in the sense ofl but only "confirmed" in the
sense of 1'. Since the conclusion that the regularity is one that we are
146 HILARY PUTNAM
justified in betting on is intuitively sound, and I does not register this

conclusion, then that already shows that. I is not a formalization of our
entire prescriptive inductive competence.
One might think to escape this problem by discovering someday that
the logic which formalizes our "inductive" (non-demonstrative as well as
demonstrative) reasoning is itself non-recursive. If the relation of being
confirmed by given evidence (or confirmed to a degree r by given evidence)
is not itself a recursive one, then the above theorem may be irrelevant (e.g.,
if the confirmation relation is recursive in the jump of the complete recur-
sively enumerable set, then the inductive logic may, in principle, be strong
enough to discover any regularity that is expressible as a Boolean com-
bination of recursively enumerable predicates 3 .) But I don't think that this
is really a significant "escape hatch".
To see why it isn't (apart from the difficulty of seeing what it would
mean psychologically to postulate that our "competence" is high enough
up in the arithmetic hierarchy to avoid the "diagonal argument referred
to), consider the following version of the Montague-Kaplan "Paradox of
the Knower": Let P be a mathematical description of our full (demon-
strative and non-demonstrative) prescriptive competence, and consider the
"Computational Liar", i.e. the sentence:
(/) There is no evidence on which acceptance o/the sentence (/) is
justified. 4
or rather the sentence of mathematics which is the Godelian equivalent of
(I). (It follows from GOdel's work that there is a sentence of mathematics
which is true if and only if P does not accept that very sentence on any
evidence, where P is any procedure itself definable in mathematics - not
necessarily a recursive procedures.
Now, suppose there is evidence e on which the acceptance of (I) is jus-
tified. Then (I) is false. And (I) is a sentence of pure mathematics. So there
is a situation in which we are prescriptively justified in believing a math-
ematically false sentence, if P really is a sound formalization of our pre-
scriptive competence! Many would consider this already to show that a
correct P cannot lead to the acceptance of (I) (e.g., the standard De Fi-
netti-Shimony requirements rule out assigning a degree of confirmation
less than 1 to a mathematical truth on any evidence). But let us go on for
a bit. Suppose there is evidence e on which the acceptance of (I)'s negation
is justified. Then, either there is also (presumably different) evidence on

which the acceptance of (I) is justified, or else (I)'s negation is a math-
ematically false sentence that P instructs us to accept on certain evidence.
(N.B. the negation of (I) asserts there is evidence on which the acceptance
of (I) is justified.) In the first case (there is different evidence on which the
acceptance of (I) is justified), we know by the previous argument that (I)
is a mathematically false sentence, and so there is again evidence - the
"different evidence" mentioned - on which P licenses the acceptance of a
mathematical falsehood, namely (I) itself. In the second case, we just noted
that (I)'s negation is a mathematical falsehood that P licenses us to accept
(an "omega inconsistency"). So, if P "converges" on the argument (I) -
that is, if the question "is the acceptance of (I) justified or not)" is one
to which there is ever an answer (relative to some evidence e or other)
which we are P-justified in giving - then there is evidence relative to which
P licenses us to accept a mathematical falsehood.
The point about this argument is that it is obvious. If there is anything
that I am "competent" to believe, anything I am "justified" in believing,
it is that P cannot converge on the argument (I) without licensing me to
believe what is mathematically false. So it certainly seems that the follow-
ing should be accepted as a criterion of adequacy on the acceptability of
any proposed formalization of our prescriptive competence:
(II) The acceptance of a formal procedure P as a formalization of
(part or all) of prescriptive inductive (demonstrative and non-
demonstrative) competence is only justified if one is justified in
believing that P does not converge on P's own Godel sentence
(I) as argument.
So if P is a description of a procedure the brain uses which is prescriptive
for human brains, then we cannot come to see that this procedure is pre-
scriptive for us as long as our brains function as Psays they should.
To spell this out: Suppose our brains are functioning in accordance with
P. We assume that is is rational to accept the above criterion of adequacy
(II). If we come to believe that P is a complete and correct description of
our own prescriptive competence, then, since we believe (II), we have to
believe that P does not converge on (I); but this is just to believe (I)! And,
since we believe that this belief is justified and that justified belief is for-
malized by P, we are committed to believing that P tells us to believe (I)
148 HILARY PUTNAM
- i.e., P converges on <n! But now we have a contradiction.

So a description of our "competence" which is complete and prescriptive
is certainly an impossible task for us (it might be an interesting task for
Alpha Centaurians. Unfortunately, if the Alpha Cenaurians figured out
what our prescriptive description is, either we could not understand it or
we would not be justified in trusting their statement!)
"COMPETENCE" AND PRESCRIPTIVE COMPETENCE
The natural reaction of a cognitive scientist to all this is "So What? No

one working on computational models of human reasoning aims at any-
thing like a complete description of how we ought to think. What does any
of this have to do with cognitive science?"
I certainly agree that this is the right reaction, or, rather, that it is right
as long as "cognitive science" is really pursued as empirical model building
and not as philosophy by other means. But, it is not at all clear how much
"cognitive science" really is empirical model building in any serious sense
as opposed to the making of claims about the philosophical significance
of various research programs. As soon as someone talks about "explaining
what makes thought thought" or "capturing the nature of thought" - and
one hears such talk constantly at "cognitive science" get-togethers - it is
perfectly legitimate to point out that one thing that "makes thought
thought"" as opposed to "cows mooing" (in Frege's phrase) is that
thought is sound or unsound, i.e., thought has normative epistemic prop-
erties. Perhaps these properties are not what "makes thought thought"
either (although I myself cannot believe that an account of "what makes
thought thought" which leaves us in the dark as to the nature and origin
of these properties could be correct). Perhaps it is rather semantic prop-
erties such as reference and truth that "make thought thought". But com-
puter models of the brain are certainly not going to give us an account of
reference or truth. Can a description of thought which by its own admission
does not aim at describing the difference between correct and incorrect
ways of thinking (in either an epistemic or a non-epistemic sense of "cor-
rect") have relevance to the philosophical problems that worried Frege or
the philosophical problem that worried Brentano?
It is important to remember that the Orhodox or East Pole version of
computationalism has taken the distinction between "competencee" and
"performance" to be the fundamental distinction on which cognitive

science is based. Now, Gilbert Harman and, at least some of the time (in
conversation) Noam Chomsky have suggested that our idealized or "com-
petence" description is a description of correct thinking in the normative
sense. 6 If they are right, then it follows from what has just been shown
that the aim of cognitive science is to describe something that it is in
principle impossible for us to describe completely. Some may not be both-
ered by this: I can sympathize with the sort of realist who is willing to
believe that either there are or there aren't infinitely many binary stars, for
example, even if it is beyond human competence to know the fact of this
matter. But the impossibility of knowing what our "competence descrip-
tion" is looks like a different sort of impossibility, an impossibility of
knowing a we-know-not-what. Tying the notion of "competencee" which
figures in descriptive cognitive science, the notion of a best idealization for
predictive and descriptive purposes, to the notion of competence which
figures in inductive logic does not look like a good way to go!
FODOR AND THE LANGUAGE OF THOUGHT
The next subject I want to discuss is an argument I employed a number

of years ago 7 against the idea of an innate Language of Thought. (The
connection between the subjects in this paper will become evident shortly,
I hope.) The Pope of High Church Computationalism - Jerry Fodor - has
declared that all concepts are innate (including, obviously, sparkplug). His
argument in The Language of Thought was that the brain must be thought
of as containing at least one agent (or "homunculus" in Dennett's widely
used terminology) who formulates such hypotheses as
(III) When Jones says "the sparkplug needs to be replaced", what he
says is true if and only if the sparkplug needs to be replaced.
The homunculus cannot have the ability to formulate arbitrary hy-
potheses of this kind unless its innate vocabulary (the vocabulary of
"Mentalese") includes every concept a human being could learn to under-
stand. So Fodor's incredibly strong version of the Innateness Hypothesis
- the thesis that the Oxford English Dictionary is innate - must be true.
Criticizing this argument, I argued that the brain need not think the way
I do. When I take the "intentional stance" towards my own brain, I may,
for example, attribute to it the knowledge that there are certain represen-
150 HILARY PUTNAM
tations in the left-eye stripes of areas 17 and 18, and certain representations
in the right-eye stripes of the same areas, and attribute to it certain infer-
ences to the existence of external objects with certain colors and shapes at
a certain distance from my bod.y, but it would be false to say "Hilary
Putnam knows that there are such and such representations stored in such
and such places in areas 17 and 18". Similarly, the homunculi in my brain
may formulate hypotheses of the following form, even if Hilary Putnam
does not:
(IV) If I want this body to speak like Jones, I should print out "The
sparkplug needs to be replaced" whenever I have neural inputs
of type blahblah.
and such hypotheses can be stated in a machine language much weaker

than the O.E.D. An analogy which it is useful to keep in mind in this
connection is the analogy to a Turing machine which proves theorems in
Zermelo-Fraenkel set theory. The program of such a machine can be stated
in machine languages much weaker than set theory.
When I published this argument, Fodor replied, fairly enough, that no
one has ever written down rules for using a language correctly which are
stated in such "procedural" terms. Of course he is right. But this is not
surprising. If we could state rules for using language without making any
normative blunders, then we could survey our own prescriptive com-
petence, and I have just argued that we know on GOdelian grounds that
we cannot do that. In fact, we have no reason to think we can even come
close to describing our own prescriptive competence. How close we can
come is, of course, an empirical question; but the Godelian results remind
us that, no matter how smart we may be, it is always possible that we are
not smart enough to answer the question "What makes you so smart?".
Of course, we may be able to write programs for "reasoning" which simu-
late the errors that people make on the tests designed by Tversky and
Kahneman! But anyone who thinks that such test results show "there isn't
much to human intelligence" should consider the well known difficulties
with simulating informal context-bound reasoning in Artificial Intelligence
(the "Frame Problem").
If we cannot solve the "Frame Problem", and we content ourselves with
a "menu" of rules for speaking which leads to "printing out" utterances
that are irrational, our simulated speaker will look "dumb"; if the utter-
ances are too obviously unwarranted, they will be linguistically "deviant".
We will not have "captured understanding". So it's Catch 22! Unless we
do something which is Utopian, Jerry Fodor can say "You haven't shown
that it is possible to understand a natural language by following rules that
are statable in machine language." It doesn't follow from this that what
is going on at the cause-effect level is Fodorian inferences (inferences of
the form (III) carried out with the aid of a language of thought in which
"all concepts are innate").
"COGNITIVE SCIENCE" AND INTENTIONALITY
I repeat that one thing that none of the programs in "cognitive science"
speak to is the problem of intentionality. Of course the philosophers who
write on issues in cognitive science do have theories of intentionality. Per-
haps the best known is Dennett's "intentional stance"theory. If this theory
is supposed to extend to semantics - if the idea is that sentences are always
true or false only relative to the "stance" of some interpreter or other -
then I don't see any way to even understand this without retreating to a
"redundancy" theory of truth and reference for the language of the last
interpreter in the chain. And such a theory seems to me inherently un-
stable. Redundancy theories, I claim, collapse into either magical theories
of reference or Linguistic Idealism. (This is a provocative remark.)
The reason that none of the computationalist programs speak to the
issue of intentionality, or the problem of the nature of understanding as
a deep philosophical problem, is that, as I remarked, the disputants use
notions like "competence" in ways that vacillate between modest descrip-
tive ambitions and total emptiness. As I have already pointed out, what
it means to speak of "the" competence of human beings in the prescriptive
sense is wholly unclear. I don't mean that the commonsense cenception of
an ability to tell sound from unsound reasoning to some extent is unclear;
I mean that the idea that this ability is really something like a "com-
petence" in the Chomskian sense, and thus something to be ideally repre-
sented by a computational description, is wholly unclear. Certainly a de-
scription of it would be beyond our powers. And even if those Alpha
Centaurians do have a description of it, we can restate the problem of
intentionality by asking "what single out their description of my brain as
152 HILARY PUTNAM
the competence description?" (As Kripke has emphasized 8 ).

Even if we waive the central issue, and even if we pretend that the Alpha
Centaurians do have that description D, the description of their prescrip-
tive competence isn't D but, say, D'. And the description of beings intel-
lectually powerful enough to find that out is D". And ... So it seems that
the description of the "competence" of an arbitrary physically possible
"intentional system" capable of understanding a language would be an
infinite disjunction D or D' or D" or ... and that no physically possible
intentional system would be able to survey (give a rule for writing down)
this disjunction. So the problem, if it is a problem, and it certainly feels
like a problem, "What do all instances of understanding have in com-
mon?" isn't addressed - isn't either "solved" or "dissolved". Anyone who
feels that this is the real issue will have to look to something other than
"Cognitive Science" - to metaphysics 9 or to Wittgensteinian therapy or to
the possibility of a new alternative in philosophy. His or her bafflement
won't be relieved by any amount of marvellous work in Artificial Intelli-
gence.
NOTES
1 Attributing this geographical metaphor to Fodor "several months ago in a heated dis-
cussion at MIT," Dennett writes,
He was equally ready, it turned out, to brand people at Brandeis or Sussex [in Boston]
as West Coast. He explained that just as when you're at the North Pole, moving away
from the pole in any direction is moving south, so moving away from MIT in any
direction is moving West. MIT is the East Pole and from a vantage point at the East
Pole, the inhabitants of Chicago, Pennsylvania, Sussex, and even Brandeis University
in Waltham are all distinctly Western in their appearance and manners. Boston has long
considered itself the Hub of the Universe; what Fodor has seen is that in cognitive
science the true center of the universe is across the Charles, but not so far upriver as the
wild and woolly ranchland of Harvard Square. ['The Logical Geography of Computa-
tional Approaches (A View from the East Pole), read at the MIT Sloan Conference,
May 19, 1984]
2 'Degree of Confirmation and Inductive Logic' pp. 270-292 of my Philosophical Papers,
vol I, Mathematics, Matter and Method (1975; second edition 1979.) First published in P. A.
Schilpp (ed.), The Philosophy of Rudolf Carnap, The Open Court Publishing Co., La Salle,
Ind., 1963.
3 The Boolean combinations of recursively enumerable predicates are exactly the k-trial
predicates (proved in my 'Trial and Error Predicates and the Solution to a Problem of Mos-
towski', Journal of Symbolic Logic 30, March 1965, 49-57). It is an easy theorem that the
k-trial predicates can be enumerated by a single trial-and-error predicate. The statement in
the text follows from the fact (proved in the same paper) that the trial-and-error predicates
are exactly the predicates in 1:2n2, i.e., the predicates recursive in the jump of the complete
r.e. set. For a general survey of results on the incompleteness of inductive logics belonging
to various classes in the arithmetic hierarchy, see Peter Kugel's 'Induction, Pure and Simple,'
in Information and ControllS, December 1977,276-336.
4 If the inductive logic P uses the notion of degree of confirmation rather than the notion
of acceptance, then one replaces "is justified" by "has instance confirmation greater than .5",
as in the paper cited in n. 2.
5 The relevant part of Godel's work is now generally referred to as the "Diagonal Lemma".
This says that if F(x) is anyone-place predicate definable in the language of a system of
mathematics which contains number theory, then there is a sentence of that very language
which is true if and only if the godel number of that sentence does not satisfy the predicate
F(x). In this form, the Diagonal Lemma was not, of course, stated in Godel's 1934 paper,
but all the techniques needed for the proof are present in that paper.
6 Harman makes this suggestion in 'Metaphysical Realism and Moral Relativism: Reflec-
tions on Hilary Putnam's Reason, Truth, and History.' Journal of Philosophy 79, 1982,568-
575.
7 In 'What is Innate and Why,' in Language and Learning, Massimo Piatelli (ed.), Harvard
University Press (1980), pp. 287-309.
8 Cf. his Wittgenstein on Rules and Private Language, Harvard University Press (1983).
9 Recent metaphysical ideas suggested in this connection are the idea that the "broad con-
tent" of a sentence depends on the state of affairs it "causally covaries with across possible
worlds" when the speaker is (in relevant respects) in an ideal epistic relation to his environ-
ment (Fodor in 'Narrow Content and Meaning Holism', unpublished), and the suggestion
that certain sets of external things are "elite", i.e., that it is in the nature of things that these
sets are the ones to which our natural kind terms have a singled-out correspondence (David
Lewis in 'New Work for a Theory of Universals', Austra/asian Journal of Philosophy 61,
1984,343-377). Notice that these suggestions are purely philosophical - they in no way depend
on computer modelling!
Received 9 June 1984
Opt. of Philosophy
Harvard University
Emerson Hall
Cambridge, MA 02138
U.S.A.
ANNETTE BAIER
EXPLAINING THE ACTIONS OF THE EXPLAINERS.
Explaining has been seen by many as the paradigm display of rationality

on our part, and rationality as our most distinctive attribute. To be an
explainer is, along the way, to be one who can generalize, infer, predict,
offer justifications, conform to the rules of evidence, and, in order to offer
more ambitious explanations, one must be a theorizer, and a theory evalu-
ator. But other things we do are equally complex and equally distinctive.
Such capacities as the capacity to compose, perform and interpret music,
to stage an opera, to design institutions and conform to their norms, to
write comedies and comedies about comedy writing, are not plausibly seen
as merely spinoffs from our capacity to become theorizers. The reverse
seems a likelier hypothesis, if we feel compelled to reduce some capacities
to others. We might well wonder which of the many things done by those
capable of offering explanations are the most wonder-provoking, and
most interestingly unlike the doings of volcanoes, whirlwinds, minerals,
plants, viruses, bacteria, insects, and most animals. Hempel believes that
it is our rationality, that property which, among its other workings, enables
us to give and improve explanations, and offer theories of explanation as
well as theories about other explananda, that entitles our actions to a
special explanation schema. I shall discuss his version of this rationality,
and try to see how much it does and doesn't explain about us, according
to his version of explanation.
Among the many things which only explainers do is to argue about the
nature of explanation. Why questions seem to be answered in a variety of
ways in a variety of contexts, and any attempt to reduce the variety to
some canonical essential form will provoke resistance among the users and
defenders of the forms which are most transformed by the reduction. Hem-
pel's Kantian account of explanation as subsumption under law has met
with most resistance from those who have attended to those after-the-
event-explanations we give of unlikely and unpredicted happenings. It is
not necessarily that these are not subsumable under law, but rather that
the laws do not appear to be what explains them. We seem to explain some
Erkenntnis 22 (1985) 155-173. 0165-0106/85/0220-0155 $01.90

156 ANNETTE BAIER
happenings by giving a narrative, a sequence of connected events made

possible but sometimes only improbable by the laws we know. Explana-
tions of events in human history seem to take this form, and since some
of the events in human history are intentional actions, so also do ordinary
explanations of some human actions. When Becket changes his lifestyle
on being made archbishop by Henry the Second, and surprises him by the
austerity and piety he assumes, this unpredicted tum of events and the
subsequent human actions leading up to Becket's murder are understand-
able, although few of them would naturally be described as what any ra-
tional agent in those circumstances would have done. The style of life
Becket adopted was quite appropriate for an archbishop, although a sur-
prising change of style from the bon vivant he had been before. What he
did when he excommunicated Eynsford despite the King's protest, and
refused to hand over to the civil powers a priest charged with debauchery
and murder, insisting on the church's privilege to try its priests, were
understandable assertions of the church power by its primate, but sur-
prising defiance of his former friend, companion, and benefactor. But such
surprising actions by rational persons are so familiar to us that it is sur-
prising that we retain our capacity for surprise. It is the norm for rational
persons to display their rational freedom in unexpected ways. When they
do, we seek to reconstruct their reasons, to understand how they came to
do the surprising things they did, and sometimes we succeed in getting an
explanation which satisfies us. Rarely will it show us that there was nothing
else for such a person in such circumstances to do, and we seem content
if we can be shown that what, say, Becket did was a reasonable thing to
have done in the circumstances. Had he cooperated with the King in curb-
ing the church's power, we would have understood that too, given his
friendship with the King, and his earlier apparent agreement with him
about the desirability of retrenching ecclesiastical privileges. But we know
that people do sometimes change their minds, change their allegiances,
change their goals and values, change the norms they endorse, so we are
not unduly perplexed by such happenings, provided we can discern some
motivation for them. We accept the limits of our power to predict the
paths of golfballs and, both for the same reasons and for the extra reasons
brought in by rationality, creative imagination, and freedom, of people's
lives. Like anything else subject to merely probabilistic laws, we are some-
times the subjects of improbable happenings, and since we have the power
EXPLAINING THE EXPLAINERS 157
to revise or reject the rules and policies we try to adhere to, we sometimes
do improbable things.
These considerations make it hard to see how we could ever get some-
thing like Becket's change of lifestyle on being appointed archbishop ex-
plained by Hempel's Schema R, for rational action. That schema is
A was in situation of type C.
A was a rational agent.
In a situation of type C any rational agent will do x.
Therefore A did x. (Aspects of Scientific Explanation, hereafter
Aspects, p. 473.)
The general regularity, subsumption under which explains the action,
here seems to be a general claim about what rational agents always do,
and a derived generalization about what rational agents do in type C situa-
tions. What is it that rational agents such as Becket always do? Hempel's
answer to this comes in various forms. They pursue their "total objec-
tives," where this is usually a matter of achieving some wanted end state
without violating norms they have adopted (Aspects, p. 465). But Becket
seems to have been adopting new norms when he changed his ways. He
switched from a worldly and hedonistic style of life to an ascetic other-
worldly one. Why? Hume, in trying to understand Becket's life and death,
finds pride and ambition to have been the driving forces, albeit "under the
disguise of sanctity and zeal for the interests of piety and religion" (David
Hume, History of England, Ch. VIII, pages on the year 1170). But why
this particular disguise? When Henry made him archbishop, he might have
satisfied his ambition by retaining rather than renouncing his position as
chancellor, and so exercised both ecclesiastical and civil power. Why did
he choose to separate himself from and oppose Henry, rather than to con-
tinue as his ally, and as his supporter in the attempt to re-confine eccle-
siastical privileges within their earlier bounds? Was it his pride which,
offended by his debt to Henry for such power as he had, dictated that he
take the path of separation and opposition rather than continued friend-
ship and cooperation? It is not clear that he could expect more power by
opposing Henry and consolidating his power as archbishop, than by com-
bining his power as chancellor with the power he would have had as arch-
bishop of a church less powerful than he in fact chose to aim to make it.
As Hume said (op. cit., writing of the year 1162) his appointment as arch-
158 ANNETTE BAIER
bishop "rendered him for life the second person in the kingdom, with some
pretensions of aspiring to be the first". He chose to exploit those preten-
sions, rather than to be second person as well as chancellor and friend of
the first. Hume makes his choice intelligible by pointing out how it satisfied
his ambition, and by mentioning his pride as well as his ambition, but all
that his account does is show us how it was as rational for Becket to do
what he did as it would have been for him to satisfy his pride and ambition
under the combined guises or disguises of archbishop and chancellor.
Hume retells the story of an episode from the days before Becket was
made archbishop, one that might be thought to explain why Becket, once
archbishop, renounced the chancellorship.
'One day, as the king and chancellor were riding together in the streets of London, they
observed a beggar who was shivering with cold. Would it not be very praiseworthy, said the
king, to give that poor man a cloak in this severe season? It would surely, replied the chan-
cellor; and you do well, Sir, to think of such good actions. Then he shall have one presently,
cried the king, and seizing the skirt of the chancellor's coat, began to pull it violently. The
chancellor defended himself for some time, and they had both of them tumbled off their
horses to the street when Becket, after a vehement struggle, let go of his coat; which the king
bestowed on the beggar.. .' (ibid.).
This public playful display of the way Henry saw Becket's duties as a
member of the clergy to combine with his privileges as chancellor may
have been taken by Becket as a warning as well as a public humiliation.
prudence as well as hurt pride may have dictated his refusal to try to
combine the powers of archbishop with those of chancellor, after this dem-
onstration of the precariousness of his dignity as chancellor. Still, we
scarcely get a tight argument of the form that any rational person who is
proud and ambitious, and whose pride and ambition had been both fur-
thered and mocked in the way Becket's had by Henry, would choose sep-
aration rather than alliance, given Becket's situation and opportunities.
We do understand Becket's motivation better, when we hear of this epi-
sode, but not because we see that it made his actual choice the only one
a rational person could make. If his norms could change to require mor-
tifi.cation of bodily appetities, to give his pride and ambition better scope,
then they might also have changed to require mortification of his pride,
to give his ambition wider scope. Even if we knew more of the secrets of
Becket's heart than we are ever likely to know, we would not necessarily
come to know that what he did was, for him, more rational than what he
could have done. Nor do we need to know this to understand his action
by coming to be acquainted with more details of his situation, his person-
ality, and his past.
Does Hempel think that, to explain as a rational action Becket's adop-
tion of a hostile attitude, we must show it to be the action any rational
agent would take, given Becket's situation, goals, and norms? No, for as
Hempel points out there may be no one course of action which is the
rational one, in his sense, under conditions of uncertainty (Aspects, pp.
468-70). Becket's choice was such a decision, but we may feel that even
had Becket been a firm adherent of maximax, or of maximin, or of some
other rule, this still would not settle whether he should indulge his pride
to the extent he did, on the interpretation adopted. It is not attitude to
risk or uncertainty which seems unfixed in Becket's case, as much as goals
and norms. Can we only be rational in Hempel's sense provided we achieve
stability in our goals and norms? 1 find some instability in Hempel's posi-
tion here. On the one hand he says "I will not impose the requirement that
there be "good reasons" for adopting the given end or norms: rationality
of an action will be understood in a strictly relative sense, as its suitability,
judged by the given information, for achieving the specified objective."
(Aspects, p. 465). This limits the use of Schema R to actions which do not
display a currently occurring change of mind about objectives and norms,
and it rules out explaining any such changes by schema R. On the other
hand Hempel characterizes rationality in terms of what presumably is
Ryle's concept of higher order dispositions (Asp~ts footnote 4), and it is
hard for me to see how Rylean higher order dispositiOllS can be prevented
from leading to the possibility. of alteration or revision of objectives and
norms as well as change or revision of beliefs and perhaps of attitudes to
risk and uncertainty. Hempel does not say that his concept of rationality
as the capacity for higher order response is Rylean, but he had earlier
(Aspects, p. 458) referred us to Ryle for details of what he, Hempel, means
by a disposition, and when he calls rationality a "broadly dispositional
trait" in the section where higher order dispositions are spoken of, he again
refers us to Ryle (Aspects, footnote 17). Ryle, like Hempel, was trying to
give an account of what distinctive features the human mind displays.
Ryle's aims were more descriptive than explanatory, nor did he choose the
term "rationality" for the cluster of higher level dispositions he found in
us. Nevertheless 1 think we should look to see what he meant by "higher
160 ANNETTE BAIER
order", to try to see if what Hempel calls rationality can both be higher
order, in Ryle's sense, and also relativised to goals and norms taken as
fixed.
Hempel says that the rationality of agents whose behaviour is to be
explained by Schema R is a "broadly dispositional trait" (Aspects, p. 472)
and that the relevant dispositions include higher order ones, "for the be-
liefs and ends in view in response to which, as it were, a rational agent acts
in a characteristic way are not manifest external stimuli, but rather, in
turn, broadly dispositional features of the agent" (Aspects, p. 473). Pre-
sumably this means that for merely intelligent behaviour, like a dog's or
an ape's, which manifests beliefs and ends in view rather than manifesting
a response to the agent's beliefs and ends in view, Schema R will be in-
appropriate. To explain the dog's behaviour in digging we will use, per-
haps, some weaker relative of R, let us call it "Schema I":
A was in circumstances C (craving a bone, "remembering"
where he buried one).
A is an intelligent agent.
In circumstances C any intelligent agent digs where he remem-
bers having buried a bone.
There may be a whole array of explanation types for simpler animals -
spiders may need "Schema N", for instinctive behaviour, and perhaps
heating systems will fall under a parallel "Schema S" for state maintaining
systems. The form of the explanation is in no way special to rational action,
nor even is the fact that the circumstances will always mention beliefs and
desires, since they will be needed also in explanations of intelligent animal
behaviour, and some primitive analogues of them may be needed for in-
stinctive behaviour. If we want to find out what is special about the ex-
planation of rational behaviour we will have to look at the particular sorts
of circumstances which need to be listed, given the force of "rational"
instead of "intelligent," "instinctive" or "state maintaining" in the prem-
isses. The circumstances will need to include the higher order beliefs, val-
ues, adopted strategies, and not just the knowledge and desires which we
would cite to explain the dog's behaviour.
How does Hempel's version of rationality, in terms of broad and higher
order dispositions, relate to Ryle's account of the sort of behaviour which
displays mind? Ryle distinguished between abilities such as the ability to
cycle, and the abilities to comment on, approve, criticize, teach, or ana-
lytically describe, cycling. The latter are all higher order abilities since they
are directed upon a lower order ability. Ryle devoted few words to reason
and rationality, contenting himself with intellect and intellectuality, and
it has been left for others, such as Hempel, to give a more or less Rylean
account of what a rational animal must be. Ryle distinguished the typically
intellectual activities of theorizing about and teaching other possibly non-
-intellectual displays of intelligence both from higher order yet non-intel-
lectual activities such as "heeding" how one was cycling, when the terrain
called for special care, and from the base level displays of intelligence, such
as learning how to cycle when given some training.
It seems clear that animals as well as human persons show intelligence,
in Ryle's sense, and they may show non-intellectual higher order capacities
too - the capacity to imitate, and to monitor their own and their offspring's
attempts at various tasks. They also seem to display the ability to train
their young. Are they rational in the sense Hempel thinks that we are, if
to be rational is to respond intelligently not just to intelligent performances
but to one's own beliefs and desires? The mother cat when she suckles her
kitten after being importuned by it can be seen as responding to the recog-
nized desire of the kitten and to its belief that it has come to the right
place to gratify that desire. We can treat cats as what Dennett calls "in-
tentional systems," attributing to them beliefs and desires, an we can also
attribute to them some sorts of responses to beliefs and desires. But then
again we might equally well describe the mother cat's response as a direct
response to the kitten's behaviour, rather than to the kitten's expressed
desire. Hempel contrasts response to beliefs and ends in view with response
to environmental conditions and external stimuli. Do the expressive and
purposive moves of one's fellows count as merely external stimuli? The cat
seems capable of responding to the intelligence-displaying and purpose-
displaying behaviour of other cats, just as we respond to the expressed but
unspoken purposes of our fellows. The cat displays nothing that we would
have no choice but to take as the manifestation of a response to its own
ends in view, and it is hard to see how any animal without linguistic expres-
sion of beliefs and ends in view could conclusively demonstrate such self-
-directed higher level attitudes. (Such activities as burying nuts for later
consumption, or preparing a nest, which might be seen as directed on
continuing and future desires, we dismiss as instinctive.) But if to be ra-
162 ANNETTE BAIER
tional and to behave in a way calling for Schema R is to display not merely
higher order dispositions, but self-directed ones and perhaps the special
case of those which are intellectual self-directed higher-order dispositions,
then most of our actions, as well as all the cat's, will not qualify as ex-
plananda requiring explanation by Schema R.
Ryle's themes, in The Concept of Mind, were that intelligence can be
displayed in non-intellectual tasks as well as in intellectual tasks, and that
self-directed higher order dispositions (such as that displayed in thinking
"sum res cogitans") are special cases of attention to intelligent perform-
ances, whose-ever they are. "At a certain stage the child disvovers the trick
of directing higher order acts upon his own lower order acts. Having been
separately victim and author of jokes, catechisms, criticisms and mimicries
in interpersonal dealings between others and himself, he finds out how to
play both roles at once." (Concept of Mind, (hereafter C.M.), p. 192).
"Cogito ergo sum" happens to be both an intellectual and a self directed
higher order performance, and Ryle's aim was to enable us to see that it
no more requires a ghostly agent than do all the other intelligent but non-
intellectual performances or other higher order and self-directed but non-
intellectual acts, such as self-ridicule or self-mimicry. "A person can, in-
deed, and must act sometimes as reporter upon his own doings, and some-
times as a prefect regulating his own conduct, but these higher order self-
dealings are only two out of innumerable brands, just as the corresponding
interpersonal dealings are only two out of innumerable brands" (C.M., p.
194). Ryle wanted to get our studied moves properly related to our unstu-
died moves, and our studied didactic and theorizing moves properly re-
lated to non-intellectual moves, as well as to relate higher order to lower
order acts, and self-directed to other-directed acts. He gives us no defini-
tion of rationality, but merely a series of spectra of more or less intelligent,
more or less studied, more or less intellectual, and of lower and higher
order acts. Indeed he derides the epistemologist's and moralist's concept
of Reason, and of that Conscience which is "just Reason talking in its
sabbatical tone of voice" (C.M., p. 315). "These internal lecturers are sup-
posed already to know, since they are competent to teach, the things which
their audience do not yet know. My Reason is, what I myself am not yet,
perfectly rational, and my Conscience is, what I am not yet, perfectly con-
scientious. They have not anything to learn. And if we asked 'Who taught
my Reason and who taught my Conscience the things that they have
learned and not forgotten?' we should perhaps be told of corresponding

instructors lodged inside their bosom" (ibid.). Reason is a mythical om-
niscient lecturer and theorizer in one's breast, and rationality is its wisdom.
Those of us who are unwilling to cede the terms "reason" and "ration-
ality" to the rationalists, against whom Ryle's devastating attack is mount-
ed, would probably want to say that to have the capacity for these Rylean
higher order didactic self-directed activities one must be rational, while to
have merely the capacity to benefit from training, as the kitten has, or even
the capacity to do the training, as the adult cat has, but to be without the
capacity for studied higher order acts, let alone intellectual ones, and to
have no capacity to "discover the trick" of turning higher order capacities
upon oneself, is not yet to be rational. But what of all the intermediate
possibilities? And what is the relationship between rationality and the ca-
pacity for higher order acts? Do we use Schema R to explain what the ape
does when it imitates the successful food washing moves of its more in-
ventive fellow? Do we use it for the human child reared by wolves who
can laugh at himself as well as at others but as yet has no language, nor
any chance to study? And, in our own case, where we believe the full range
of capacities to be present (if anything counts as the "full"range, given
that ridicule, criticism, endorsement, can go at indefinitely many levels)
but not constantly displayed, do we use Schema R for those of our actions
that do not display anything studied, higher order, or self-directed?
Hempel is well aware that much of our behaviour that we explain with
reason explanations exhibits something much less than a clear case of re-
sponse to our own recognized beliefs and desires. He discusses not merely
the ideal or limit case of someone like Bismarck responding to his own
clearly perceived various options - to make the unedited Ems telegram
public, not to make it public, to edit it before publicizing it, but also cases
where the agent is not fully aware of the desires he is indulging, censoring
or sublimating. Bismarck's all things considered endorsement of his own
preference for the public opinion-manipulating wily move to provoke
France to declare war on Prussia is a clear case of higher order response
to clearly seen aims and beliefs about effective strategies (although, had he
decided against editing or publicizing it, we would not need to invoke
different aims or norms to explain that alternative action - Bismarck might
have judged his editing move too risky). A different sort of reason, how-
ever, would be given by a Freudian for the less successful person's choice
164 ANNETTE BAIER
of a tricky subterfuge which gained her nothing that she could avow as a
wanted end, but rather led to her own exposure as a deceiver. The analyst
might cite, as an explaining reason, the woman's need to feel that she
controls what others believe, especially about herself, and a belief that
such control is better demonstrated by deceit than by openness, by self-
concealment than by self-revelation. This would-be wily person responds
to her own unconscious or not fully conscious ends by recognizing op-
portunities to pursue them. Schema R is to be used for explaining her
behaviour, as well as Bismarck's. So to be a rational agent, in Hempel's
sense, is to respond to beliefs and ends one mayor may not be able to
avow as one's own.
This considerably widens the field where a response counts as the re-
sponse any rational agent would make. A particularly obtuse person may
never get to the point of acknowledging the aims and companion beliefs
which Freudians would cite to explain her behaviour, so may never give
any evidence of capacity to respond to these particular beliefs and ends,
qua her own beliefs and ends. The "real reasons" for her behaviour will
not coincide with her reasons as she sees them. Real reasons, whether or
not they are avowed or even avowable by the agent, are what we will try
to cite when Schema R is used. Bismarck's avowed reasons are presumably
taken by Hempel as his real reasons - whether or not some deep unac-
knowledged personal desire to control the fate of nations lay behind his
desire to preserve and enhance Prussia's national honor by provoking a
war from which he expected Prussia to emerge victorious, he really did
want that war - it was his end-in-view, even if not his ultimate end, so the
reasons he gave for editing the telegram were real reasons, really explaining
his action. He acts in response to his acknowledged beliefs and ends in
view, whereas the less selfknowing agent acts in response to her unac-
knowledged beliefs and to ends not yet in clear view to her. The question
I raised was just what sort of higher order response to beliefs and ends
Hempel thinks necessary to demonstrate rationality. It cannot be con-
scious guidance by those ends and beliefs, implying as that does an ability
to recognize them as one's ends and beliefs. The person whose avowed
reasons are discrepant from her real reasons shows an ability to avow, but
not to recognize or acknowledge true reasons, and if she were clever
enough at selfdeception and unfortunate enough to have a motive to em-
ploy it, she might never acknowledge any of her real reasons. Should we
use Schema R for her, or is she not a rational agent? Should we use it to
explain her behaviour as long as she has shown some disposition to ac-
knowledge real reasons, although on this occasion the real reasons are
ones she never shows any ability to acknowledge? It is not yet clear how
extensive a disposition to what sort of response to her real beliefs and :t:eal
ends she must demonstrate to count as rational. Let us assume that it is
enough if she has displayed the ability to acknowledge some of her real
beliefs and ends, whether or not she shows that ability in relation to the
act to be explained by Schema R. Then among circumstances C we should
specify not only her real ends and beliefs, but also specify whether or not
she is aware of them. For what a person does when consciously guided by
her beliefs and ends will differ in many ways from what she does when she
needs to disguise them and their force. Reason shows one sort of cunning
when reaons are unacknowledged, a different sort when they are clearly
in view.
A question now arises about the scope of Schema R. If Freudians can
use it to explain action inappropriate relative to the agent's avowed ob-
jectives that is, however, appropriate relative to some postulated uncon-
scious objectives, then can it be invoked to explain all action which is less
than fully rational relative to avowed objectives? Can the postulation of
unavowed objectives serve to transform all apparently less than rational
action into rational action? Do we use Schema R to explain irrational and
weakwilled action, and action displaying incompetence or carelessness in
deliberation or execution, as well as to explain fully rational action? Of
course a necessary condition for doing so is that the resulting explanation
have some merits as an explanation - that the attribution to the agent of
beliefs and desires not avowed by that agent really would have some ex-
planatory power. Freudians, when they attribute unconscious objectives,
usually have independent evidence, from dream contents and the like, for
such attribution, so that the acceptability of their claims about a person's
unconscious goals do not rest solely on the explanatory gains of merely
postulating such goals. If we have no such independent evidence for una-
vowed goals, but attribute them merely to "rationalize" the agent's actual
behaviour over a stretch of time, then the acceptability of such attributions
will rest solely on their explanatory power and on the lack of better con-
firmed explanations. Our only reason to accept the statement attributing
such beliefs and desires to the agent will be its explanatory power. Other-
166 ANNETTE BAIER
wise we would simply have a groundless version of that part of the ex-
planans which takes the form "A was in situation of type C'. But is it
sufficient for use of Schema R that there is sufficient reason to attribute
to the agent a particular set of unavowed beliefs, goals and norms?
This will depend on the force of that other part of the explanans as-
serting A to be a rational agent. It must mean more than that guiding
beliefs and desires are plausibly attributed to the agent, otherwise all the
higher animals count as rational agents. It must mean less than that what
guides A's behaviour are consciously recognized beliefs and desires, other-
wise the schema cannot, as Hempel wants it to, cover Freudian explana-
tions. Do we or don't we invoke it to explain all less than rational behav-
iour on our part? If what it means is "A aspires to conform to the demands
of rationality," then this would indeed limit it to those agents capable of
such higher level aspiration, and it would have a role to play in explaining
their failures as well as their successes, but not always or often by dis-
playing what they did as the thing to do given these norms of rationality.
If we are to show that what A did, when he acted less than rationally, was
what any agent who aspired to be rational would do in circumstances C,
then circumstances C must contain those details of the limits of A's com-
petence or of his will to be rational which explain the partial failure of his
attempt at rationality. Freudian explanations purport to do this in one
special way - by citing aims competing with the avowed aims, norms com-
peting with the rational norms. There could be other reasons for our failure
to be as rational as we consciously want to be. We may sometimes be
singleminded but feebleminded, rather than, as the Freudians portray us,
strongminded but not singleminded. Inner conflict need not always be the
explanation of failure to conform to endorsed norms, be they the formal
norms of rationality or more substantive norms. We surely are imperfectly
rational beings and whatever schema fits our behaviour has to be a schema
to explain attempts at more than we usually succeed in doing.
Hempel, however, does not seem to want the force of "A is a rational
agent" to be "A accepts and tries to conform to the norms of rationality".
For when he explains why the explanans must contain the claim that A is
rational, he says "Now the information that agent A was in a situation of
kind C and that in this situation the rational thing to do was x, affords
reason for believing that it would have been rational for A to do x, but no
grounds for believing that A in fact did x. To justify this latter belief we
clearly need a further explanatory assumption namely that - at least at the

time in question - A was a rational agent and thus was disposed to do
whatever was rational in the circumstances" (Aspects, p. 471). Here "dis-
posed to" must mean more than "thinks he should, and aspires to", other-
wise we still would have, from the explanans, no way of deriving more
than that x was what A thought he should do, and made some efforts to
do. What Hempel later says about the "model of a consciously rational
agent", comparing its usefulness to other ideal concepts, such as that of
ideal gas behaviour, makes it quite clear that the norms of rationality cited
in Schema R are taken as norms which are in fact conformed to, on the
occasion in question, not ones making a difference by being on that oc-
casion unsuccessfully aspired to. "Very broadly speaking, the explanatory
model concept of the consciously rational action will be applicable in those
cases where the decision problem the agent seeks to solve is clearly struc-
tured and permits of a relatively simple solution, where the agent is suf-
ficiently intelligent to find the solution, and where the circumstances permit
careful deliberation free from disturbing influences" (Aspects, p. 482). This
means, I think, that Schema R can rarely be used of actual human action,
although it may explain some parts of human planning. Bismarck may
have acted on a rational plan, in this ideal sense, but Becket's decision
problem was less clearly structured - or was it that Becket's pride distract-
ed him from seeing the rational solution to the clearly enough structured
problem of how to maximize satisfaction of his ambition? Or did he in
fact solve it perfectly, achieving by martyrdom greater fame and glory than
Henry? As Hempel says, (Aspects, p. 478) in real life the notion of optimal
action relative to one's total set of objectives and beliefs is impossibly
obscure.
To be disposed to be rational, in Hempel's sense, seems to mean to be
about to succeed in doing what any perfectly rational agent in those cir-
cumstances would do, unless prevented by disturbing factors external to
the agent's competence, beliefs, and total objectives (conscious or, alter-
natively, unconscious). So this Hempelian rationality comes and goes, and,
when relative to conscious beliefs and objectives, is gone more often than
not. Thus Hempel can speak of a person being "a consciously rational
agent (at a certain time)" (Aspects, p. 479, my emphasis). Rationality
comes and goes, as we succeed or fail to conform to its norms. But, as we
usually speak, a rational agent is one with the capacity both to endorse
168 ANNETTE BAIER
the norms of rationality and to achieve some measure of success in living

up to them. The capacity may come and go (when one is mentally ill, it
may go) but it will not be gone whenever we act less than perfectly ac-
cording to the norms of rationality, (whatever they are - I shall later raise
a question about Hempel's version of them). There seem, then, at least
two reasons to question whether to be a rational agent is to do what is the
rational thing to do. First, there may be nothing which is the rational thing
to do, merely several things all counting as a rational thing to do, if ra-
tionality is the capacity of higher order response of a Rylean sort. Second-
ly, even when there is a unique answer to the question of what it is rational
for her to do, a rational agent may fail to see that answer, or having seen
it, fail to act on it, without thereby showing that she has lost her rational
capacity. What transpires will be different in nature from what would tran-
spire in a being who had no pretensions to doing what is seen to be ra-
tional. The failed cunning of reason shows itself differently in rational
animals than in non-rational animals.
The trouble, I think, lies in the way Hempel moves from rationality as
a normative concept to rationality as an explanatory concept. His move
is to let the explanatory concept apply only to those in whom, or those
times when, the norms of rationality are conformed to perfectly. He pos-
tulates the ideally rational agent, and lets Schema R apply only to such an
agent. But agents as we know them are not ephemeral beings but lasting
beings who have acquired and learned skills, which they display with vary-
ing success, and whose objectives make what sense they do only because
they are lasting beings, who know themselves as variable in performance,
capable of progress and regress when judged by the norms they endorse,
and capable of progressive or regressive change in such endorsements. We
cannot be agents at a certain time, only over time. If rationality is a prop-
erly characteristic of agents, and only derivatively of their actions, then it
too cannot be displayed in isolated episodes, but only in a pattern of ac-
tivities of forward and backward looking agents. Our actions vary in the
extent to which they display our rational capacities, so are more or less
imperfectly rational actions, but that does not mean that the higher level
capacities in which rationality consists come and go. We often need to
refer to our norms, including the norms of rationality, to explain our fail-
ures as well as our occasional successes. Schema R, however, does not help
us to explain our less than fully correct moves. The truth of that part of
the explanans which asserts A to be a rational agent is not independently

testable, if all it means is that A is on this occasion the locus of a visitation
by the spirit of perfect rationality, enabling A to do the rational thing
relative to the aims and beliefs attributed to A. It becomes trivially true
whenever the beliefs and aims attributed do make the action done the
rational thing to do. But it is (trivially) true only if "rational agent" means
"ordinary lasting agent having a passing moment of perfect rationality,"
while being false if we take rationality as a capacity for reliable displays
of such perfect rationality. The only capacity actual human agents have
is the capacity to make more or less rational plans, and to try to carry
them out in action. Their actual actions are seldom completely according
to plan, and their plans seldom perfectly rational. Schema R, then, seems
rarely to get application to our intentional consciously purposive actions.
It will apply more often to our prior planning, and it may more often
apply to our unconsciously purposive doings. But it fails to apply to our
only more or less successful attempts at consciously rational action.
If we cannot use Schema R for our actual actions, only for some of our
plans and for some of our unconsciously motivated failures to live ac-
cording to our avowed plans, norms and objectives, then we must use
some other schema to understand ourselves as we usually display ourselves
iIi intentional action. Schema R turns out to be more suitable for robots
than for us. It turns out also to be minimally different from my hypo-
thetical Schema N for successfully evolved insects, say, cockroaches.
Varying the previously quoted statement of Hempel's (Aspects, p.482) we
could say "The explanatory model of the instinctive action will be appli-
cable in those cases where the decision problem solved by the insect's
instinct is clearly structured and permits of a simple solution, and where
the insect is well enough adapted to have that solution wired in, and where
circumstances allow instinct to operate undisturbed." Such an ideal con-
cept will get application to insect actions more often than that of ideal
rationality will get application to human action. To explain our actions to
ourselves, we will do better with a schema for imperfectly self-conscious,
imperfectly rational, imperfectly singleminded animals, capable of revising
their beliefs, goals, and strategies, and capable of surprising one another
and themselves. This describes the higher animals as much as it describes
us, and I think that is what we need - a schema for ourselves which will
enable us to see how ones who act as we do could have come to do so,
given what it is rational to believe about our animal ancestry.
170 ANNETTE BAIER
A question I deferred earlier was that of the content of the norms of

rationality, those norms to which we aspire with varying degrees of success.
I shall give no positive answer to this hard question, but merely point out
a tension in Hempel's account. He speaks of rationality as requiring the
adoption of means appropriate to our total objectives or ends in view,
constrained by whatever other norms we espouse. He also speaks of maxi-
mizing expected utility. This instrumental conception of rationality seems
to me to fit uneasily with his version of scientific and so presumably ra-
tional explanation. To give a Hempelian explanation is to subsume under
law, not to suit means to ends. If, as rational explainers, we must see events
as conforming to law, if we must try to conform to the imperative "seek
for laws from which occurrent events can be deduced as instances," then
why, as rational agents intent on other activities than explanation, are we
free to seek merely for efficient means to our ends? There is a tension
between Hempel's Kantian account of rational explanation, and his
means-end relativist account of our other rational activities. If to be ra-
tional explainers we must ourselves conform to the laws or rules of the
explaining game, themselves dictating that we try to see everything as law
governed, then would we not expect many or all of our other rational
activities would also display such respect for laws and rules, rather than
mere unconstrained goal pursuit or mere maximization?
Two alternative reconciling moves are open to the Hempelian here. One
might try to show the concern with laws as a special case of means-end
rationality, or one might try to show means-end rationality as a special
case of Kantian law centered rationality. Hempel does the latter in as far
as he encourages us to subsume our own instrumental rationality under
descriptive law to understand it. But what of the practical norms we are
following if we do try to use Schema R, or some other Hempelian expla-
nation schema, to understand ourselves? Do such self-explaining moves
display Kantian or instrumental practical reason? Are we to have a teleo-
logical or a deontological account of the explaining game? Hempel, like
so many others, seems to want to accept a Kantian law-centered account
of theoretical reason, but to combine that with a means-end relativist ac-
count of practical reason. Explaining and its norms, which dictate law
discernment, would then have to be shown as suitable means to the ex-
plainers' total objectives. The Hempelian explainers' preoccupation with
laws will then derive from his special objectives, not from his rationality
as such, so Kantian reason will be absorbed into the instrumental reason

of those with a taste for law discernment.
Our self-approved rational actions are a subset of our actions as intel-
ligent and creative animals capable of speech and so some degree of self-
consciousness, and capable of other norm-constrained, inventive, and
higher order acts. Among our rational activities is explaining. Any ac-
ceptable explanatory schema for our actions ought to explain our explan-
atory moves. Schema R may be got to do so, if we take rationality as
adopting efficient means to achieve our objectives, and good covering law
explanations as among our objectives. To be a rational explainer will sim-
ply be to make the moves which can be expected to yield the sort of ex-
planations which one aims to get. But suppose we ask why we avow and
endorse this aim? Why do we, once we reflect on the matter, endorse the
objectives of improving our explanations and our understanding of their
structure and role? Is it rational, given our total objectives, to care about
explaining in the way we do? If these objectives include the protection of
redwoods, the continued preservation, creation and enjoyment of good
music, good novels, good jokes, good wine, as well as good explanations,
good science and nonlethal technology, then we would hope to see some
sort of coherence among our multiple goals, and some explicable place for
explaining and theorizing among them. But without its companion human
activities, explaining becomes a mysterious, inexplicable activity. The rath-
er rationalist conception of rationality which Hempel adopts, stressing as
it does the formal demands of consistency and maximization, and of ra-
tional explanation as subsumption under law, is one whose ancestors are
to be found in Descartes and Leibniz, ones which see reason as a divine
faculty, by which we approximate to a god's eye view of things. But if one
thing seems quite certain about divine activity it is that gods need do no
explaining whatever. Omnipotent creators and legislators understand their
creation and its laws without having to explain it to themselves - they
intend it, so know it without either observation or explanation. If we are
to understand ourselves as explainers, we had better not model our version
of the rationality which shows in our better explanations on any theolo-
gy-based version of our reason. Hempel shows no wish to base his concept
of rationality on any theological notions, nor is he concerned with the
intellectual ancestry of the concepts he relies on. I think that our usual
conceptions of deduction and of law are, however, concepts with a theo-
172 ANNETTE BAIER
logical ancestry, and our usual version of consistency and coherence in a

comprehensive system can be seen as an intellectua,l descendent of the
theologian's version of divine unity in an encompassing variety of attri-
butes. This ancestry can be ignored, but it tends to reassert itself.
It shows itself in a dualism between what is seen as our less than rational
animal habits, desires, means-end strategies, social propensities, customs,
artistic and musical tastes, and our more "divine" capacities for math-
ematics, explicit inference and reasoning, for law discernment, for scientific
theory, decision theory and other abstract systematizing activities. This
dualistic version of our nature has difficulty doing justice to the linkages
between, say, music and mathematics, or constitutional law and sq~ntific
law, between inference and associative thinking, explaining and explaining
away. I cannot here try to defend a more naturalistic Rylean conception
of our capacities against the rationalist conception, traces of which I find
in Hempel's concepts of explanation and of rationality, in tension there
with the more naturalistic and Wittgensteinian elements taken from Ryle.
The rationalist logic of logical empiricism never did combine very well
with the empiricism. To see ourselves as combining a quasi-divine or ex-
divine reason with other capacities of obvious animal origin is to see our-
selves as demi-gods in beasts' bodies, worse monsters than ghosts in ma-
chines. We need some account of human rationality which makes it neither
mere animal means-end intelligence, nor divine contemplation of an or-
derly creation, nor an uneasy alternation between these.
Hempel's account of explanation, and of its varieties, has and will con-
tinue to provoke reflective responses. My own response to it, in this essay,
has been less "rational" than the responses it more naturally invites, but
I hope not the less reflective nor the less appreciative for that. I have tried
to express some worries I have with the Hempelian account, worries which
may show more about the limitations of my rationality than the limitations
of the account whose implications I have tried to explore. At the end I
attempted to subject that account of the explanation of rational action to
the test of reflexivity - to ask if it can, in its own terms, explain itself. This
is to ask of it whether, like divine and rationalist mind, it can include itself
in its survey. "But, of course, if I am rejecting such a theological conception
of mind in favor of a more social and naturalistic one, why should this be
a reasonable demand to make of an account of explanation? It may not
be a reasonable demand. Still, if we are to exercise our capacity for higher
order response (whatever its ancestry), we seem bound to want to try to

turn accounts on themselves. It would, perhaps, be more accurate to say
that it is understandable that those of us who have no account of our own
to offer, but who are fascinated by and curious about the accounts of the
creative thinkers such as Hempel, should indulge in such parasitic reflec-
tions.
NOTE
* I am grateful for helpful comments on earlier drafts of this paper from Kurt Baier, Nicholas
Rescher, Arthur Ripstein and Patrick Maher.
Manuscript Received 14 May 1984
University of Pittsburgh
Dept. of Philosophy
Pittsburgh, PA 15260
U.S.A.
NORETT A KOER TGE
ON EXPLAINING BELIEFS·
INTRODUCTION
In explicating the concept of scientific explanation Hempel provided an

ideal in terms of which we can judge both philosophical theories of meth-
odology and contemporary scientific theories of natural and social phe-
nomena.
As an example of the first evaluation, it is instructive to ask whether
Lakatos' methodology of research programmes is the best way to achieve
Hempelian explanations. Are "hard core" propositions really testable
nomic generalizations in disguise or are they quite different from covering
laws? It is important that we not compartmentalize our theories of aims
and methods.
If wielded with tact and subtlety, the Hempelian model can also provide
a valuable critique of explanatory programs, especially in the social
sciences. For example, do structuralist approaches to folk tales generate
testable generalizations or do they merely point out patterns of formal
analogy? Is there a deductive (or strong inductive) relationship between
the premises of functional analyses and the detailed description of the
institution we wish to understand?
When I say the critique must be done with tact, I have in mind con-
siderations such as the following: Since the Hempelian model represents
the ideal aim of scientific inquiry, it is wise to construe each of the "re-
quirements" as desiderata which actual scientific accounts may approach
to a greater or less degree. For example, on many strict accounts of nom-
icity, in the early 17th century, neither Galileo's law of falling bodies nor
Kepler's laws of planetary motion would have counted as truly law-like
because at the time they were proposed, they referred essentially to par-
ticular objects, namely our earth and our solar system. Nevertheless, they
covered an enormous number of cases and in some sense were very close
to being law-like. Later, of course, they were derived from Newton's com-
pletely universal theory. It is also useful to think in terms of degrees or
Erkennitnis 22 (1985) 175--186 0165-0106/85/0220-0175 $01.20

176 NORETT A KOER TGE
distance from the ideal when applying the requirements of empirical test-
ability, maximal specificity and truth. l
In my experience with graduate students, there is initially a very strong
resistance to the prospect of evaluating social sciences, especially history,
in terms of the D-N and I-S models; but if one construes Hempel's theory
not as a harsh demarcation criterion, but as an approachable ideal, it
sounds more attractive.
However, there are still interesting questions which come up when one
looks at the types of explanation used in the history of science. My paper
deals with just one aspect of the problem of explaining beliefs.
BACKGROUND TO THE PROBLEM
Perhaps the best way of introducing the problem is to describe the reaction
of my history of science students to Hempel's account of historical expla-
nation:
We now understand the model one uses to explain why Caesar crossed the Rubicon or why
Admiral Tyron gave his fatal command, 2 but we're historians of ideas - we want to know
how to go about explaining why Aristotle ever canle to think that if one added wine dropwise
to water, the wine became water - or why some people became Copernicans and others
resisted. How do you explain people's ideas?
They were particularly interested in finding out whether the explanations

would involve both "externalist" and "internalist" considerations. I field-
ed this question in the typical philosopher's manner - by drawing a dis-
tinction. Namely, between the problem of explaining where new ideas
come from or how people came to entertain certain propositions - we
might call this the mutation problem - and the problem of explaining why
certain ideas are accepted, believed, taken seriously, etc. - we might call
this the survival problem.
In this essay I will not discuss the very thorny problem of trying to
explain mutations (or discovery). 3 Instead, I will concentrate on one aspect
of the survival, or acceptance problem.
It will be useful to introduce a rough schemata for explanations and to
review the Rationality Principle. My intention in this survey is to avoid
the controversial and problematic aspects of explaining actions and instead
to focus on the areas which we understand fairly well and on which there
is wide philosophical agreement.
ON EXPLAINING BELIEFS 177
Many scientific explanations can be represented diagramically in the

following way:
Transition
System in System in
State S1 State S2
Principle
For example, S1 might describe the initial positions and velocities of an

array of masses. Using Newtonian equations of motion, we can calculate
their positions and velocities at some later time.
We can also use explanatory inferences in connection with static situa-
tions:
System in
equilibrium; Equilibrium Value
Values of n-l of nth
variables given Principle variable
For example, if we are told that a balance is in horizontal equilibrium and

are given the values of the weights on it and all their distances from the
fulcrum but one, using the laws of statics we can calculate the position of
the last weight.
I have used the neutral term "principle" in these diagrams hoping that
it would be acceptable to instrumentalists who might consider Newton's
laws to be inference tickets, not claims, to non-Humean realists who think
that they describe necessary connections, and everyone in between.
Actions can be explained in a roughly similar way:
Agent in Rationality Agent

choice performs
situation Principle action
One describes the agent's problem situation, the set of options open to
him, how desirable he finds the outcomes of the various options, etc. We
then calculate which option is most appropriate for the agent and then
predict that he will indeed act on that option. There are interesting prob-
lems concerning the details of this model. For example, exactly which de-
cision principle do people use? Are they maximizers, satisfiers, minimaxers,
or what? We may also ask about the nomological force of whichever RP
we choose - when we make a rational choice is the outcome of our decision
determined in the same sense as are the motions of Newtonian particles?
There are also debates about how moral considerations should be intro-
duced into the decision matrix, but there is wide agreement on the general
features of the model. We know roughly what goes into the choice situa-
tion and how the decision is made.
Since the description of the initial choice situation includes a character-
ization of the situation as the agent sees it, the Rationality Principle (RP)
model can be used to explain the actions of religious fanatics, incompe-
tents, politicians, etc. In fact, it is sometimes remarked that it is just in
connection with the explanations of weird or "irrational" behaviour that
the RP is most helpful because it forces us to analyze the idiosyncracies
and unusual constraints operating in the choice situations involved in such
cases.
Thus, accompanying the RP model is a methodological claim to the
effect that all (or nearly all) actions can be explained this way.4 Even in
the case where agents appear simply to have made a mistake in their op-
timization calculations, RP theorists (such as economists) simply move to
a higher level and talk about the cost of spending too much time on trivial
decisions.
This is all very sweet (tout comprendre, tout pardonner), but just be-
cause of its universal applicability, the RP gives us no method of demar-
cating rational from irrational actions (here I am using "rational" in the
stronger ordinary language sense), nor of understanding why some people
sometimes have such distorted theories about their situations. Explana-
tions in terms of the RP model give us no deep answer to questions about
inappropriate behavior. For this we need a theory about beliefs.
COMPETING ACCOUNTS OF BELIEF FORMATION
Let us now try to construct a model for explaining beliefs. (Recall we are
really only discussing the problem of accepting readily available proposi-
tions. The problem of creating radically new ideas or even new languages
is not our concern here.) In rough outline, it might be as follows:
- ON EXPLAINING BELIEFS 179
Agent Consonance Agent

in doxic forms
situation Principle belief
So far, so good. But what are the details of the model? In particular, what
sorts of things are elements in the doxic situation? Here, unlike the situa-
tion regarding the explanation of actions, there are very fundamental dis-
agreements among philosophers.
Let us first sketch what I will call the "weight of evidence" approach.
To explain why Jones believes q, one would describe the experience, evi-
dence, theories, background knowledge of Jones which is relevant to q
(relevant in the confirmation theory sense), and show that this provides
reason to believe q. To be more specific, for a very simple case, a Bayesian
would use the following model:
Agent has prior Agent believes

probability q to degree '2.
Inference
p(q) = '1. and (where p(q,e) =
----~
receives '2)
Principle
evidence e
On this explanatory account, if two people's propositional attitudes to-

wards q differ, it is due either to the different evidence which they have
available or to the different manners in which they assign prior probabil-
ities. One could base a fairly optimistic philosophy on this model. Since
basic prior probabilities are inborn and have probably become fairly uni-
form through evolution, and since they are eventually swamped by the
evidence anyway, given world enough and time to pool our evidence, all
disagreements should vanish.
One might quarrel with various details of the Bayesian formulation -
some would argue that the credibility of a posit is not a probability - but
I think a qualitative version of the weight of evidence approach is a fairly
popular one, especially among internalist historians of science - the main
disagreement concerns how positivistic our construal of "evidence" should
be. Should the doxic situation of scientists include metaphysical and meth-
odological presuppositions?
Let us now turn to what I will call the "ideological" account of belief
formation. On this account, the description of the doxic situation must
include not only background beliefs, but also utilities, including psycho-
logical desires, political values, and the class interests of the individual.
The idea that the emotional evaluation which we place on the state of
affairs described by a proposition can influence our tendency to believe
that proposition is a very old one. Demosthenes, Third Olynthiac, 19 says:
" ... in such proposals the wish is father to the thought, and that is why
nothing is easier than self-deceit. For what each man wishes, that he also
believes to be true." Julius Caesar echoes this: " ... we readily believe what
we wish to be true."s
In more recent times, this basic idea has been spelled out in various
ways. Freud developed the theory of wish fulfillment and postulated psy-
chological mechanisms of denial, introjection, projection, etc., whereby
threatening propositions are changed into benign ones. Although we may
not be much impressed by the theory, some of the case studies describe
phenomena which cry out for further investigation and explanation. For
example, Geiger, in ldeologie and Wahrheit, describes a doctor who diag-
nosed his own case as bronchitis even in the face of the most obvious signs
oflung cancer - symptoms which he easily interpreted correctly in patients
other than himself.
Writers on ideology give numerous examples in which class interests are
held to influence beliefs. Many of their cases are not too convincing be-
cause it generally turns out that the evidence and received traditional
knowledge available to members of different classes varies as much as do
their interests. This leaves open the possibility that the differing beliefs can
be explained solely on weight-of-evidence grounds without invoking util-
ities.
It would obviously be of great interest to specify in some detail just how
desires affect beliefs - Do they lead to selective memory, to selective evi-
dence gathering, to distortions of our perceptions, or what? - and then to
test them experimentally. Experiments are very hard to carry out in this
area and also hard to interpret. For example, consider a classic experiment
which is said to demonstrate suggestibility. Subjects are shown two lines
which are really of equal length, stooges declare one line to be longer and
then the subjects follow suit, and let us assume that the subjects sincerely
believe that the designated line is longer. Is this phenomenon to be ex-
plained in terms of a wish to conform or is it perhaps a case in which the
subject has a theory that two heads are better than one and hence takes
the stooge's response as positive evidence that the designated line is longer?
One of the most interesting experiments in this area is the case in which
poor children are reported systematically to overestimate the size of coins,
but this effect has turned out to be non-reproducible, even though it has
been slow to disappear from psychology textbooks. (Does this mean that
psychologists are reluctant to drop their belief in the experiment because
they like it and \vish it were true?)
Forensic science provides a clear case in which the emotional state of
an observer directly affects the content of his or her beliefs. Victims of
violent crimes tend to overestimate the height and weight of their assail-
ants.6 (Note that this can hardly be a straightforward example of wish
fulfillment or denial.)
Although the evidence is not unambiguous, it seems likely that there is
at least some direct influence of fears, desires, interests on belief formation.
I will now propose a schema for explaining these phenomena.
PRAGMA TIC MODEL OF BELIEF FORMATION
According to my model, there are two determinants of acceptance. I begin

with the output from the weight-of-evidence schema, but then add on an
extra step:
Agent believes b Agent stores b

to degree r; Pragmatic principle in immediate
analysis of cost access belief
of error of Detachment bank
I will argue briefly for two claims:

(a) That the propositional contents of our beliefs are more readily ac-
cessible to us than the degrees to which we believe them.
(b) That utilities, namely costs of error, as well as weights of evidence
or probabilities enter into the detachment process.
182 NORETTA KOERTGE
First the claim that propositions in the belief bank are detached. Sup-
pose I am taking a Reader's Digest quiz. Given the question, "Is Mahler
later than Brahms?" I immediately answer "Yes." But if asked, "What
reasons do you have to believe he was later?" or "What odds would you
take on it?" much head-scratching ensues. I am even slowed down by the
question, "How sure are you?" "Well, pretty sure," I say.
Upon reflection, I can produce some evidence for my belief (Anna Mah-
ler had an affair with Kokoshka and I know he is a 'modern' painter) and
I can even settle on some odds. By contrast, the proposition "Mahler is
later than Brahms" simpliciter was instantly at my fingertips.
Since odds or weights of evidence are hard for the brain to calculate and
must always be revised as new evidence comes in, it seems to me that it
might make good evolutionary sense for an animal which depends on
quick reactions for survival to have readily available a simple information
bank. Juries and medical researchers can and should mull over the evi-
dence, but quiz kids, baseball referees and hunters need to make speedy
judgments.
Suppose I am correct in assuming that we have a working information
bank in which detached propositions are stored. How is the threshold
determined? How much weight of evidence must there be for a proposition
before it is stored? This could be done in various ways - for example, the
level might be constant and set at 0.5 or it could vary with personality
parameters, say how skeptical we are by nature, etc. I suggest that the level
is fixed by pragmatic considerations and that it varies roughly according
to the seriousness of the harmful consequences which are likely to occur
if a mistake is made.
A simple example of the type of process I have in mind is found in
jurisprudence. In British law, two different standards of proof are used. In
criminal cases (where a mistake could lead to the jailing of an innocent
person) members of the jury are instructed to be sure of their verdict "be-
yond reasonable doubt," while in civil cases they only need be convinced
"on the balance of probabilities" that the defendent is guilty. In addition,
there is fine structure within each category. For example, in discussing
certain kinds of fraud, Phipson's Manual of the Law of Evidence says:
The standard [of proof for such cases of fraud] is the civil one of preponderance of proba-
bilities, but what is "probable" depends upon the heinousness of what is alleged... 'in pro-
portion as the offence is grave, so ought the proof to be clear' ...
Legal commentators give two reasons for having variable standards of

proof. One is based on the dubious premise that people are inherently less
likely to commit serious crimes. 8 The other one is more in line with what
I have in mind: "What is reasonable doubt ... [should and does vary] in
practice according to ... the punishment which may be awarded."9
So I think that it is reasonable to expect that an organism would have
a relatively high, i.e., stringent weight-of-evidence threshold for proposi-
tions such as "This kind of mushroom is safe to eat" or "I am dying of
an incurable disease." In typical cases, nothing much is to be gained by
believing the proposition even if it's true, while making a mistake and
believing it when it is not true could lead to a good deal of unnecessary
unpleasantness. In short, the storage procedure is based on what statisti-
cians call an analysis of Types I and II error. I assume that this decision
as to which propositions are to be admitted into the belief bank is almost
entirely unconscious, and in some cases the threshold may be fixed at first
by inherited factors. For example, there may be good evolutionary reasons
why children have such conservative beliefs about what is good to eat.
This cost-of-error model differs somewhat from the familiar wish-is-
father-to-the-thought approach. In many cases we do tend to have lower
standards of acceptance for propositions which describe desirable states
of affairs and, conversely, are reluctant to believe statements which fore-
cast painful or fearful events. But this is not always the case.
If I am walking in a woodland frequented by rattlesnakes, the slightest
rustle ofleaves may be enough to make me believe that I glimpsed a snake.
Now if I am not a snake fancier, my credulity does not stem from wish-
fulfillment. However, if flight or defensive measures are called for, I (and
my adrenal glands) am wise to jump to the conclusion that a snake is there.
On the other hand, a herpetologist in search of specimens and not wanting
to waste time on false leads will be much slower to conclude that a snake
is around.
A cost-of-error analysis also partially explains why police are quick to
shoot suspected felons, while juries are (often) loathe to convict them. A
police person who fails to capture a possibly dangerous criminal has more
to lose than if a petty shoplifter evades arrest. However, jury members
who sentence someone to the death penalty have to live with the knowledge
that they might have been mistaken.
Now this automatic detachment process which I am postulating need
not lead to any loss of information or self-deception. One would hope that
Geiger's doctor, even though not accepting the proposition that he has
lung cancer, could nevertheless give a correct estimate of the probability
that he does. And if the utilities of his situation change - perhaps his family
and colleagues are upset at what they (mistakenly) consider to be self-
deception, or maybe some night school philosophy teacher conditions him
into feeling bad for having beliefs which are useful but not true, or he
reads a novel about heroism in the face of absurdity - then we would
expect his acceptance threshold and hence his belief to change. There are
two ways to change beliefs - one can change the available evidence or
change the cost-of-error.
It should be emphasized that the propositions stored in these instant-
access information banks are not our best candidates for knowledge
claims. They are often rather crude unqualified approximations, perhaps
only valid in rather provincial domains. They are only intended to be, as
Pete Seegar once said after unsuccessfully trying to tune his banjo, "good
enough for folk music" - good enough for the complicated business of
ordinary life. The selection of the type of proposition to be stored is also
influenced by pragmatic factors.
If our object is to find which of our beliefs are serious candidates for
being true, then these pragmatic elements must be stripped away. This is
an extremely complicated task - not only must I consider the extent to
which my acceptance of p was influenced by utilities; I must also find out
how much of the evidence for p came from the instant-access bank and
was itself influenced by my desires and interests.
In a sense then, Popper was right when he said that beliefs are of no
concern to the epistemologist. In a certain sense, beliefs are not knowledge
candidates but barriers to our attempts to improve our knowledge. But if
we want to understand and criticize ordinary human behavior, we must
study the formation and structure of beliefs. But what does my model
imply about the beliefs of scientists?
EXPLAINING SCIENTIFIC BELIEFS
On my model, the historian who wishes to explain why Priestley believed

in the phlogiston theory long after the publication of Lavoisier's textbook
must not only describe the evidence which Priestley saw as being relevant
to the theory, but must also give Priestley's assessment of the costs of being
wrong. But what might the nature of these costs be in the case of a pure
scientist (as opposed to a technologist)?
Of course, scientists are human and it may be that Priestley was swayed
by feelings of pride or nationalism. (After all, if it turned out that the gas
liberated by the calx of mercury were dephlogisticated air not oxygene, he,
Priestley, would get credit for the disvocery.)
But there are other costs involved which are more endemic to the scien-
tific process. Switching theories requires an enormous amount of institu-
tional re-tooling. Reagent bottles must be relabelled, textbooks and en-
cyclopedias rewritten, and one's entire working vocabulary and intuitive
thought processes revamped. It would be silly to make the change unless
there were very little hope that the phlogiston theory could be rescued.
Neither of the above costs were operative for Lavoisier, at least not after
the publication of his textbook. (One measure of the effort it takes to
undertake a major intellectual switch is the length of time from his original
experiments on the calx of mercury (November, 1774) as reported in his
Easter memoir of 1775 (which still speaks of oxygen as a purer form of
common air) to his final report to the French Academy in August, 1778).
Most interesting, however, are the differences in heuristic value of the
competing theories as viewed by Priestley and Lavoisier. On Lavoisier's
approach elements were (for the most part) chemicals which could be iso-
lated and weighed in the laboratory. Caloric was an exception, but even
it could be measured (with the ice calorimeter) and its effects described
quantitatively. This fact gave Lavoisier a very powerful tool for investi-
gating chemical reactions - namely, by keeping a quantitative check on
reactants and products. To stay with the phlogiston theory would be to
forego this promising general methodology.
Priestley was by no means opposed to the use of quantitative methods
in chemistry (it was his "goodness of air" assay which eventually led La-
voisier to his discovery). However, he was more skeptical ofthe possibility
of obtaining pure products of any kind in the laboratory, especially ele-
ments, and therefore mere weights of reactants and products seemed to
him to be of less value - more important was to monitor their qualitative
character. So Priestley saw little "opportunity loss" in rejecting the oxygen
theory.
The careful analyst may object at this point and claim that, in the end,
186 NORETTA KOERTGE
I have given a purely weight-of-evidence explanation of Priestley's adher-

ence to the phlogiston theory. I won't quarrel too much with this criticism.
It all depends on whether Priestley would have been able to articulate his
heuristic assessment at the time. And as I have stressed above, when we
move from the activity of explaining beliefs to criticizing and defending
them, it is crucial that as many tacit evaluations as possible be made ex-
plicit.
NOTES
• An early version of this essay was read at the 1974 meeting of the Philosophy of Science
Association. I benefited greatly from the points made in the discussion, particularly those of
Peter Hempel. This research was supported by the John Dewey Foundation.
1 An example of the dangers of being too rigid in applying philosophical norms is found
in Rosenberg's Sociobiology and the Preemption of Social Science. (See my review in Inquiry
2S (1982), 471-7.)
2 This fascinating example comes from John Watkins, 'Imperfect Rationality,' in R. Borger
and F. Cioffi (eds.), Explanation in the Behavioural Sciences, Cambridge University Press,
Cambridge, 1970.
3 See my 'Explaining Scientific Discovery,' PSA 1982, Vol. I, Philosophy of Science As-
sociation, East Lansing, pp. 14-28.
4 At least this is the case in Popper's account of such explanations. See my 'Popper's Meta-
physical Research Program for the Human Sciences,' Inquiry 18 (1975),437-42. A compar-
ison of Hempel's, Dray's and Popper's accounts of actions is given in my 'The Methodol-
ogical Status of Popper's Rationality Principle,' Theory and Decision 10 (1979), 83-95.
5 Quoted in W. Stark, Sociology of Knowledge, Routledge and Kegan Paul, London, 1958,
p.47.
6 A. Daniel Yarmey, The Psychology of Eyewitness Testimony, The Free Press, New York,
1979.
7 D. W. Elliott, Phipson's Manual of the Law of Evidence, lOth ed., Sweet and Maxwell,
London, 1972, p. 232.
8 Ibid., p. 232.
9 Ibid., p. 230. It should be noted that American law differs on this point. Although we
have two standards, one for civil and one for criminal cases, no further variation in standard
of proof is technically allowed. However, juries often seem to act more in accordance with
the British theory.
It is ironic to note that in police practice, the standard of proof is inverted. For example,
in Indiana, police are permitted to fire on a fugitive who is suspected of felony, but not on
someone suspected only of a misdemeanor. In this example, the more heinous the crime, the
less stringent the standard of proof.
Received 9 May 1984
Dept. of History and Philosophy of Sciences

Goodbody Hall 130
Indiana University
Bloomington, IN 47405
U.S.A.
PATRICK SUPPES
EXPLAINING THE UNPREDICTABLE*
It has been said - and I was among those saying it - that any theory of
explanation worth its salt should be able to make good predictions. If
good predictions could not be made, the explanation could hardly count
as serious. This is one more attempt at unification I now see as misplaced.
I want to examine some principled reasons why the thrust for predictability
was mistaken. I begin with the familiar sort of example of explanation, the
kind that occurs repeatedly in analyses of the past.
Hume's (1879) long and leisurely discussion of Charles I in his History
of England provides a number of excellent examples. Here is one in which
he is discussing Charles's decision to take action against the Scots in 1639.
So great was Charles's aversion to violent and sanguinary measures, and so strong his af-
fection to his native kingdom, that it is probable the contest in his breast would be nearly
equal between these laudable passions and his attachment to the hierarchy. The latter affec-
tion, however, prevailed for the time, and made him hasten those military preparations which
he had projected for subduing the refractory spirit of the Scottish nation. (History of England,
Volume V, p. 107)
Hume faces the standard difficulty of assessing attitudes and attachments

when there is any sort of complex issue at stake. Charles's conflict between
loyalty to Scotland and attachment to the religious hierarchy has the kind
of psychological instability that makes prediction impossible. But from
our perspective of looking back on the past, we are satisfied - at least
many of us are - by the explanation that Charles's religious ties and com-
mitments won out. By saying that we are satisfied I do not mean to suggest
any ultimate sense of satisfaction.
The central feature of this historical example, the instability of Charles's
conflicting feelings that are nearly equally matched, is the source of drama
in many important historical events, in the tensions surrounding private
choices of colleges to attend, careers to follow, and spouses to wed. The
importance of such conflict and instability in our lives is mirrored in the
importance they assume in the novels, plays, and movies that both express
and define our ways of feeling and talking.
Erkenntnis 22 (1985) 187-195. 0165-{)106/85/022G-OI87 $00.90

188 PATRICK SUPPES
This instability and unpredictability of human affairs are in no sense re-

stricted to conflicts of feeling. The vicissitudes of politics and war have
been recorded and analyzed since Thucydides. In that long tradition, al-
most without exception there have been sound attempts at explanation
but scarcely any attention given to what seemed to be the impossible task
of predicting the outcomes. There is, in fact, a general view of the matter
that is not correct in every detail but that expresses a major truth. Real
conflicts occur in human affairs when the outcomes are uncertain, because
the forces controlling them are unstable. One-sided battles that are known
in advance by all concerned parties to be such are the exception rather
than the rule. Napoleon thought he could conquer Russia, and at least
some of his generals believed him. Hitler and at least some of his generals
thought they could conquer the world. After the fact we can easily see how
foolish they were.
One view of the American adversarial system of justice is that only con-
flicts in the law that have unpredictable outcomes should reach the stage
of being tried in court. When the facts and the law are clear, early settle-
ment should be reached, because the clarity about the facts and the law
makes the situation stable and the outcome predictable. This rule does not
always work but it probably covers a substantial majority of instances of
legal conflict. Of course I do not want to overplay the argument for sta-
bility as the reason for settlement prior to trial. Just the expenses alone of
a trial push the parties for settlement even when the facts and the law
taken together do not provide a stable view of what the nature of a set-
tlement should be. All the same the stability of the facts and the law is an
important ingredient in many cases of conflict. The conflict goes nowhere
because of the sound advice of a good attorney who convinces his client
to control his anger and ignore his ruffled feathers. Major institutions of
our society are organized to a large extent to deal with the instability
generated by conflict. If the phenomena in question were predictable, much
of the need for the institutions would be eliminated. There may still be
Utopian social planners that dream of eliminating conflict and tension in
some ideally structured future society, but most of us are prepared to
accept conflict as part of the human condition and to work on ways to
minimize it locally without hope of eliminating it.
The difficulties of predicting outcomes in the kind of human situations
I have been describing are familiar. The complexity and subtlety of human
EXPLAINING THE UNPREDICTABLE 189
affairs are often singled out as features that make a science of human
behavior impossible, at least a predictive science in the way in which parts
of physics and chemistry are predictive. Moreover, given the absence of
powerful predictive methods, there are those who go on to say that the
behavior in question is not explainable. I have already indicated my dif-
ference from this view.
I now want to move on to my main point. There is, I claim, no major
conceptual difference between the problems of explaining the unpredict-
able in human affairs and in non-human affairs. There are, it is true, many
remarkable successes of prediction in the physical sciences of which we are
all aware, but these few successes of principled science making principled
predictions are, in many ways, misleading.
Let me begin my point with a couple of simple examples. Suppose we
balance a spoon on a knife edge. With a little steadiness of hand and
patience, this is something that any of us can do. The spoon comes ap-
proximately to rest, perhaps still oscillating a little up and down. Our
problem is to predict which way the spoon will fall, to the left or the right
side of the knife blade, when there is a slight disturbance from a passing
truck or some other source. In most such situations we are quite unable
to make successful predictions. If we conduct this experiment a hundred
times we will probably find it difficult to differentiate the sequence of out-
comes from that of the outcomes of flipping a fair coin a similar number
of times. Of course, in each of the particular cases we may be prepared to
offer a sound schematic explanation of why the spoon fell to the left or to
the right, but we have no serious powers of prediction.
Let me take as a second example one that I have now discussed on more
than one occasion. The example originates with Deborah Rosen and was
reported in my 1970 monograph on causality. A golfer makes a birdie by
accidentally hitting a limb of a tree at just the right angle. The birdie is
made by the ball's proceeding to go into the cup after hitting the limb of
the tree. This kind of example raises difficulties for probabilistic theories
of causality of the kind I have advocated. I do not want to go into the
difficulties at this point but rather to use this simple physical example as
a clear instance of having a good sense of explanation of the phenomenon,
but not having any powers of predicting it. A qualitative explanation is
that the exact angle at which the ball hit the limb of the tree deflected it
into the cup. We cover the difficulties of giving a quantitative explanation
190 PATRICK SUPPES
by the usual qualitative method of talking about the ball's hitting the limb
at "just the right angle." Of course, this is elliptical for "hitting the ball
at just the right angle, just the right velocity, and just the right spin." The
point is that we do not feel there is any mystery about the ball's hitting
the limb and then going into the cup. It is an event that we certainly did
not anticipate and could not have anticipated, but after it has occurred we
feel as comfortable as can be with our understanding of the event.
There is a principled way of describing our inability to predict the tra-
jectory of the golf ball. The trajectory observed with the end result of the
ball's going into the cup is a trajectory followed in an unstable environ-
ment. We cannot determine the values of parameters sufficiently precisely
to predict the golf ball will hit the limb of the tree at just the right angle
for bouncing into the cup because the right conditions of stability do not
obtain. To put it in a familiar way, very small errors in the measurement
of the initial conditions lead to significant variations in the trajectory -
here significant means going or not going into the cup. Correspondingly,
when intentions are pure and simple we can expect human behavior to be
stable and predictable, but as soon as major conflicts arise, e.g., of the sort
confronting Charles I, the knowledge of intentions is in and of itself of
little predictive help, though possibly of great explanatory help after the
fact. To put it in a summary way, Charles I facing the Scots and our golf
ball share a common important feature of instability.
Here is another simple example of a physical system of the sort much
studied under what is currently called chaos in classical dynamics. We have
a simple discrete deterministic system consisting of a ball being rotated
around the circumference of a fixed circle, each move "doubling" the last.
The only uncertainty is that we do not know the initial position with com-
plete precision. There is a small uncertainty not equal to zero in our knowl-
edge of the starting position of the ball. Then, although the motion of the
system is deterministic, with each iteration around the circumference of
the ball the initial uncertainty expands. More particularly, it is easy to
show that after n iterations the uncertainty will be 2" times the initial
uncertainty. So it is obvious that after a sufficient number of iterated moves
the initial uncertainty expands to fill the entire circumference of the circle,
and the location of the ball on the circle becomes completely unpredicta-
ble. 1 Though the system just described is slightly artificial, it is enormously
simple, with very limited degrees of freedom. It is an excellent example of
an unstable dynamical system - the instability coming from the fact that
a small uncertainty in the initial conditions produces an arbitrarily large
uncertainty in subsequent location.
The general principle that I am stressing in these analyses that stretch
from the mental conflicts of Charles I to simple rotating balls is the pres-
ence of instability as the central feature that makes prediction impossible.
Fortunately, after the events have occurred we can often give a reasonable
explanation.
I do not want to suggest that the absence of stability as such is the only
cause for failure of prediction. We can take the view that there is an ab-
sence of determinism itself as in probabilistic quantum phenomena and in
other domains as well. It will suffice here to consider some simple quantum
examples. Perhaps the best is that of radioactive decay. We cannot predict
when a particle will decay. We observe the uneven intervals between the
clicking of a Geiger counter. After the events of decay have occurred we
offer an "explanation", namely, we have a probabilistic law of decay.
There is no hope of making an exact prediction but we feel satisfied with
the explanation. Why are the intervals irregular? They are irregular be-
cause the phenomena are governed in a fundamental sense by a probabi-
listic decay law. Don't ask for a better explanation - none is possible.
I do not mean to suggest that instability and randomness are the only
causes of not being able to make predictions. I do suggest that they provide
principled explanations of why many phenomena are not predictable and
yet in one sense are explainable.
The point I want to emphasize is that instability is as present in purely
physical systems as it is in those we think of as characteristically human.
Our ability to explain but not predict human behavior is in the same gen-
eral category as our ability to explain but not predict many physical phe-
nomena. The underlying reasons for the inability to predict are the same.
The concept of instability which accounts for many of these failures is one
of the most neglected concepts in philosophy. We philosophers have as a
matter of practice put too much emphasis on the contrast between deter-
ministic and probabilistic phenomena. We have not emphasized enough
the salient differences between stable and unstable phenomena. One can
argue that the main sources of probabilistic or random behavior lie in
instability. We might even want to hold the speculative thesis that the
random behavior of quantum systems will itself in turn be explained by
192 PATRICK SUPPES
unstable behavior of classical dynamical systems. But whether this will

take place or not, much ordinary phenomena of randomness in the macro-
scopic world can best be accounted for in terms ofinstability. This is true
of the behavior of roulette wheels as much as it is of the turbulence of air
or the splash of a baby's bath.
A disturbing example of instability is to be found in the theory of popu-
lation growth. A reasonable hypothesis is that the rate of growth is pro-
portional to the current size of the population. The exponential solution
of this equation is unstable. This means that slight errors either in the
initial population count or in the constant of proportionality for the breed-
ing rate can cause large errors in prediction, quite apart from any other
influences that might disturb the correctness of the equation. 2
It is worth saying once more in somewhat more abstract terms the cen-
tral meaning of stability in the theory of dynamical systems. What I para-
phrase here is the classical Lyapunov condition for a stable solution. 3 The
idea is straightforward and already stated once intuitively. A solution is
Lyapunov-stable if two different trajectories keep arbitrarily close together
as they arise from different initial conditions provided the initial conditions
are sufficiently close. So the intuitive idea of stability is that a trajectory
can be known with any desired precision, given sufficiently small errors of
measurement in determining the initial conditions. This is exactly what is
not characteristic of instability. Very fine variations in the initial conditions
of a roulette wheel, not to speak of variations in its motion, produce very
large differences in outcome. Namely, in almost identical conditions we
have on one occasion a red and on another occasion a black outcome.
Now I am not suggesting that this exact idea of the stability of a dynamical
system can be applied to our analysis of the behavior of Charles I deciding
what to do with the Scots. There is, however, an underlying and robust
notion of stability that reflects the instability in his behavior in a faithful
way, just as much as it reflects the instability of a roulette wheel and its
resulting random behavior.
The general qualitative concept of stability is this. A process is stable if
it is not disturbed by causes of small magnitude. Thus, a chair is stable if
it cannot be easily pushed over. A political system is stable if it can with-
stand reasonably substantial shocks. A person is stable if he is not contin-
ually changing his views. More specifically, a person's belief in a given
proposition is stable if it can only be changed by very substantial new
evidence. We often say something similar about feelings. One of the fea-
tures of a stable personality is constancy of feeling. The Lyapunov formal
definition of stability can be put under this qualitative tent.
Some of the best and most sophisticated predictive science is about
well-defined stable systems, but here I am interested in the opposite story.
When a system is unstable we can predict its behavior very poorly. Yet in
many instances we can still have satisfactory explanations of behavior.
There are at least three kinds of explanation that may qualify as satisfac-
tory analyses of unpredictable behavior. The first and most satisfying
arises from having what is supported by prior evidence as a highly accurate
quantitative and deterministic theory of the phenomena in question.
Classical physics has constituted the most important collection of such
theories. In the golf-ball example discussed earlier we feel completely con-
fident that no new fundamental physical principles are needed to give an
account of the ball's surprising trajectory. It was and will remain hopeless
to accurately predict such trajectories with their salient but unexpected
qualitative properties. Our serenity, however, is principled. Classical mech-
anical systems that are unstable have unpredictable behavior but the physi-
cal principles that apply to them are just those that are highly successful
in predicting the behavior of stable systems. The explanatory extrapolation
from stable to unstable systems seems conceptually highly justified by prior
extensive experience. Moreover, in cases of importance, we can often es-
timate relevant parameters after the fact. Such estimates increase our con-
fidence in our explanatory powers. Ex post facto stress analyses of struc-
tural failures in airplanes, bridges, and buildings are good instances of
what I have in mind.
Application of fundamental theory or quantitative estimate of param-
eters seems out of the question in the second kind of explanation of un-
predictable behavior I consider. Here I have in mind familiar commonsense
psychological explanations of unpredictable behavior, exemplified in the
passage from Hume about Charles I. Consider, for instance, a standard
analysis of an election that was said to be "too close to call". A variety of
techniques are applied after the fact to explain the result: the bad weather
affected Democrats more than Republicans, the last-minute interview of
one candidate went badly, the rise in interest rates the past two weeks hurt
the Republican candidate, and so on and so on. Simple psychological hy-
potheses relate anyone of these explanatory conditions to the behavior of
194 PATRICK SUPPES
voters. Most of us have faith in at least some of these explanations, but

we have no illusion that they are derived from a fundamental theory of
political behavior. We also recognize the instability of the outcomes and
the consequent difficulty of prediction.
The third kind of explanation of unpredictable behavior does not appar-
ently depend on instability but on randomness. As has already been noted,
the random behavior of classical mechanical systems, roulette wheels, for
example, can be attributed to instability, but this is not the case for quan-
tum phenomena. In either case, however, the important point is that ex-
planation cannot go behind some basic probabilistic law that assigns a
probability distribution to the phenomena in question. We explain the
irregular pattern of radioactive decay or other data by the probability law
thought to govern the phenomena. Individual events, no matter how con-
trolled the environment, cannot be predicted with accuracy. Yet at a cer-
tain level we feel we have explained the phenomena.
Chaos, the original confusion in which all the elements were mixed to-
gether, was personified by the Greeks as the most ancient of the gods.
Now in the twentieth century, chaos has returned in force to attack that
citadel of order and harmony, classical mechanics. We have come to recog-
nize how rare and special are those physical systems whose behavior can
be predicted in detail. The naivete and hopes of earlier years will not re-
turn. For many phenomena in many domains there are principled reasons
to believe that we shall never be able to move from good explanations to
good predictions.
NOTES
• An earlier version of this article was given at the Pacific Division meeting of the American
Philosophical Association at a symposium on models of explanation, March 24, 1984. I am
indebted to Jens Erik Fenstad for a number of helpful co=ents on the earlier draft.
I A more technical description of such a simple deterministic description goes like this.
Instead of moving around a circle, we consider a first-order difference equation, which is a
mapping of the unit interval into itself:
X.+I = 2x. (mod 1),
where mod 1 means taking away the integer part so that X.+I lies in the unit interval. So if
Xl = 2/3, X2 = 1/3, X3 = 2/3, X4 = 1/3, etc., and if xl. = 2/3 + e, X2 = 1/3 + 2e, xa
= 2/3 + 4e, and in general
, _ {1/3 + 2·e (mod 1) for n even,
X n +l -
2/3 + 2"e (mod 1) for n odd,
the instability of this simple system is evident, for the initial difference in Xl and Xl> no matter
how small, grows exponentially.
dx
2 The differential equation expressing that growth of population dt is proportional to pres-
ent population is:

dx
dl = ax,
and the solution is
x = bea"
which is Lyapunov unstable, as defined in Note 3.
3 The classical Lyapunov condition for a system of ordinary differential equations is the
following, which formalizes the intuitive description in the text. Let a system of differential
equations
(1)
dXi
- = fi(xl> ... , x., t), i = I, ... , n
dt
be given. A solution Yi(t), i = I, ... , n of (I) with initial conditions Yi(/o) is a Lyapunov stable
solution if for any real number e > 0 there is a real number d > 0 such that for each solution
Xi(t), i = I, ... , n, if
IXi(to) - y;(/o)1 < d, i = I, ... , n

then
i = 1, ... , n
for all 1 ;:: to.
REFERENCES
Hume, D.: 1879, The History of Englandfrom the Invasion of Julius Caesar to the Abdication
of James the Second, 1688 (Vol. V), Harper, New York.
Suppes, P.: 1970, A Probabilistic Theory of Causality, North-Holland, Amsterdam.
Received 9 May 1984
Ventura Hall
Stanford University
Stanford, CA 94305
U.S.A.
RICHARD N. BOYD
THE LOGICIAN'S DILEMMA:
DEDUCTIVE LOGIC, INDUCTIVE INFERENCE

AND LOGICAL EMPIRICISM
It is ... difficult to escape the conclusion that when the

two ... opposing views on the cognitive status of theories
are each stated with some circumspection, each can as-
similate into its formulations not only the facts concerning
the primary subject matter explored by experimental in-
quiry but also all the relevant facts concerning the logic
and procedure of science. In brief, the opposition between
these views is a conflict over preferred modes of speech.
(Nagel, The Structure of Science, p. 152).
INTRODUCTION
One of the consistent and admirable features of the practice of logical

empiricist philosophers of science during the period in which logical em-
piricism dominated English-language philosophy of science was the
remarkable extent to which they subjected their own philosophical con-
ceptions to persistent criticism, especially with respect to the question of
their adequacy as an account of the (sound elements of) the actual practice
of scientists. During roughly the decade 1955-1965 it became clear, largely
as a result of investigations by logical empiricists (see, e.g., Feigl 1956;
Hempel 1954, 1958, 1965a) that there were serious problems to be solved
if logical empiricism was to prove adequate to the task of explicating all
of " ... the relevant facts concerning the logic and procedure of science."
In a number of papers (Boyd 1972, 1973, 1979, 1980, 1982, 1983, 1984a,
1984b) I have argued that in fact logical empiricism is inadequate to this
task and that a scientific realist conception of scientific knowledge is re-
quired in order to obtain an adequate account of even instrumental knowl-
edge in science. My strategy in those papers has been to try to show, about
every recognized feature of scientific methodology, that its epistemic reli-
ability (even with respect to instrumental knowledge) can neither be ex-
plained nor justified except from a realist perspective.
In the present essay my aim is to extend this criticism of logical empi-
ricism still further. When Nagel held that the empiricist perspective can
Erkenntnis 22 (1985) 197-252. 0165-0106/85/022Q.-{)197 $05.60
198 RICHARD N. BOYD
provide an account of the logic and procedures of science, it was no doubt

the inductive logic and procedures which were at issue. That logical em-
piricists should be able to provide an adequate account of deductive infer-
ences in science would have seemed beyond dispute. Nevertheless, what I
propose to do here is to deny that the consistent logical empiricist can give
an adequate account of the epistemological role of deductive inferences in
scientific inquiry. I will argue that the empiricist's conception of scientific
theories as deductively integrated formal systems for the prediction of ob-
servable phenomena is untenable from the perspective of consistent em-
piricism. The role which deductive inferences play in the inductive methods
of science (even as those methods have been "reconstructed" by logical
empiricists) is inexplicable for the consistent empiricist. In fact, scientific
theories are inductively integrated systems and the principles governing
their inductive integration are so distant from what a consistent empiricist
can accept that a scientific realist perspective is required to explain even
the role played by deductive inference in those inductive principles.
Having argued that the methods of science countenance patterns of in-
ductive integration of theories which the consistent logical empiricist can-
not reconstruct, I then raise the question whether logical empiricists were
nevertheless right to think of scientific theories as providing deductive (or
"inductive statistical") predictions whose experimental testing forms the
rational basis for their acceptance or rejection. I conclude that the answer
is "no". When we test theories by testing their predictions, it is typically
not their deductive (or "inductive statistical") predictions that we test, but
instead their inductive predictions obtained by inductive principles which
only the scientific realist can explain or justify. Moreover, the observations
which bear on the confirmation or disconfirmation of a theory are, by no
means, always tests of its predictions, deductive or inductive. Instead, the
actual logic of theory confirmation makes non-predicted observations rel-
evant to the testing of a theory via complex dialectical and inductive the-
oretical considerations which are entirely inexplicable from an empiricist
perspective. Only the scientific realist can explain either "the logic of con-
firmation" or even the methodological role of deductive logic!
THE LOGICIAN'S DILEMMA 199
1. EMPIRICISM AND INDUCTION
1.1 Empiricist Tentativeness Regarding Inductive Inferences
Scientific reasoning is, first and foremost, inductive reasoning: scientists

discover things about the world which go beyond just (what can be de-
duced from) the observational data upon which their investigations de-
pend. This is something which Hume taught all of us. Hume also taught
a conception of epistemology according to which one of its aims is the
articulation of the principles which underlie inductive inferences. 20th-cen-
tury empiricist successors to Hume - at least those in the dominant tradi-
tion of "logical empiricism" - have typically undertaken an especially
modest version of that task. Roughly, they have held that the epistemol-
ogist of science is to provide only an account of judgments of the "degree
of confirmation" which a theory receives on the basis of a given body of
observational evidence. Moreover, they have typically held that (subject
to a qualification regarding theories which are statistical rather than de-
terministic) observations constitute evidence for a theory only if they are
deductively predictable from the theory, together with appropriate "auxili-
ary hypotheses".
The principles which govern three other sorts of inductive inference do
not, on the view of typical logical empiricists, lie within the purview of the
epistemology of science. In the first place, there are the principles which
govern the invention of theories, the principles by which scientists are lead
to propose new theories to account for particular data or to resolve the-
oretical questions. Secondly, there are the principles which describe the
ways in which theoretical beliefs nondeductively confer plausibility on
other theoretical beliefs. Finally, there are the non-deductive principles
which lead scientists and others to expect certain observational findings,
and to find others surprising, as a result of their theoretical commitments.
The first two sorts of principles were treated by logical empiricists as be-
longing to the psychology, but not to the epistemology of science. Con-
siderations of the first sort reflect the "context of invention" rather than
the "context of confirmation" whereas considerations of the second sort
reflect the many "heuristic" principles which operate in the context of
invention.
In order to characterize the typical empiricist treatment of the third class
200 RICHARD N. BOYD
of inferential principles, we need to examine the necessary qualification to

the claim that logical empiricists took as potentially confirmatory only
those observations which are deductively predictable from the theory to be
tested (together, of course, with well confirmed auxiliary hypotheses).
Some scientific theories yield no deterministic predictions for some or all
of the observable phenomena to which they apply. To account for such
cases logical empiricists typically understood the experimentally relevant
predictions from a theory to include those indicated by deductive conse-
quences of the form p{o,E)=r, where the interpretation is that the prob-
ability of an occurrence of 0, given conditions E, is r. When a theory
(together with relevant auxiliary hypotheses) has such a deductive conse-
quence, there is the expectation that, in a suitably large random sample of
observations, the ratio of o's to Es will come to approximate r. To the
extent that this proves to be true, potentially confirmatory evidence for T
is provided; to the extent that the actual 0 to E ratio departs significantly
from r, there is potentially disconfirmatory evidence. .
Except in degenerate cases in which r is zero or one, statistical statements
of the form p{o,E)=r do not logically entail the expected experimental
results; it is logically consistent that T be true, that it entail that p{o,E) = r
and that the 0 to E ratio in actual observations be s, for any real number
s. Thus, the expected observational results whose experimental confirma-
tion supports, and whose experimental disconfirmation tends to discon-
firm, the theory T do not follow from T deductively. For this reason, sta-
tistical predictions of this sort, and especially explanations based on them,
are sometimes described by the term "inductive-statistical" (see, e.g., Hem-
pel 1965b). It is only in this sense that the logical empiricist tradition ac-
knowledges the legitimacy, or the relevance to theory confirmation, of
observational predictions inductively inferred from scientific theories. The
inference from T to the statistical prediction p{0 ,E) = r is supposed to be
deductive. The only inductive inference is that involved in the experimental
interpretation of the statistical prediction itself. In the case of statistical
predictions about observables, both E and 0 will have observational de-
scriptions, so we may characterize the empiricist conception by saying that
it typically acknowledges no legitimate principles of inference from theo-
retical premises except the principles of deductive logic. Legitimate prin-
ciples of inductive inference apply only in cases where the premises are
observational.
It will be instructive to examine the effect of this restriction on legitimate

inductive inferences by considering cases of plausible inductive inferences
from theoretical premises to observational conclusions which do not lie
within the scope of the eI!1piricist account. A class of such inferences is
suggested by empiricist treatments of the definition of probability state-
ments.
In the empiricist tradition, three different possible interpretations of
probability statements of the form p(o,E)=r were at various times (and
for various special cases) seriously discussed: (a) the expected frequency
interpretation, according to which what is affirmed is that r is the expected
ratio of o's to occurrences of E in the long run; (b) the propensity inter-
pretation according to which r reflects the propensity of Es to be associ-
ated with o's; and (c) the epistemic interpretation, according to which r
measures the evidential likelihood of 0 given E, or the evidential warrant
for expecting 0 given E (roughly, the "degree of confirmation" of 0 given
the evidence that E obtains). It appears that logical empiricists often held
that the choice from among these interpretations was largely irrelevant to
an understanding of the logic of confirmation of theories (see e.g., Hempel
I965b, Part 3). This is not to say that they agreed that any of these inter-
pretations would do as well as any other; certainly they did not. But they
appear to have held that the obvious and philosophically uncontroversial
differences between the various interpretations did not dictate the choice
of one over the others in the reconstruction of statistical predictions of the
sort we are discussing. That they should have thought this is an indication
of the exclusion of a large class of plausible inductive inferences from
consideration as scientifically legitimate.
Suppose that T is a scientific theory, and suppose that inductive infer-
ences from T at the theoretical level render plausible a theory T. Suppose
further that it is a deductive consequence of T (together with relevant
auxiliary hypotheses) thatp(o,E) = 1, where the probability in question has
a frequentist or propensity interpretation. Suppose, that is, that T' deter-
ministically predicts 0, given E. Suppose further that all of the theoretical
rivals to T which are also inductively supported by T or by other well-
established theories are such that they each entail that p(o,E) = O. Suppose
that we follow logical empiricists (as we probably should not) in holding
that degrees of inductive support are expressible as probabilities, and sup-
pose that the epistemic probability of T, given Tis, say 1/2. Under these
202 RICHARD N. BOYD
circumstances, it would be rational for someone who accepted T, and who

concurred in the relevant inferences, to assign a likelihood of 1/2 to 0,
given E; that is, to accept it that p(o,E) = 1/2, where the probability in
question is epistemic.
It is plain that if all of this is accepted at face value, we would have a
case for which the typical empiricist conception of the role of statistical
predictions in theory confirmation is inadequate. On the empiricist ac-
count, potential confirmatory evidence for T would be forthcoming if the
observed ratio of o's to Es in suitable experiments came to approximate
1/2. But this is false. Insofar as we take the observed ratio of o's to Es to
be epistemically relevant here (as we would if we took all of the inferences
involved in establishing the equation P(o,E) = 1/2 to be epistemically legit-
imate), the only circumstance in which potentially confirmatory evidence
would be forthcoming would be ones in which the observed ratio ap-
proached 1. An observed ratio approximating 1/2 would be prima facie
disconfirmatory, since this is an outcome predicted by none ofthe theories
which T recommends as inductively plausible.
In general, if the epistemic likelihood of T given Tis r, and T entails
that p(0 ,E) = r' then, subject to the other conditions stipulated above, it
would be rational for someone who accepted T and the relevant inferences
to accept the epistemic probability claim that p(o,E) = rr', but the observed
ratio of o's to Es which would be confirmatory would be one which ap-
proximates r'. It follows that, under the conditions envisioned, epistemic
probability statements are not interpretable as inductively justifying de-
terminate predictions about the expected ratio of o's to Es. If we have
inductively reasoned from T to the epistemic probability statement that
p(o,E) = s, then all this tells us is that the potentially confirmatory ratio of
o's to Es is some real number between sand 1. The claim that p(o,E)=s
does not represent an "inductive statistical" prediction about experimental
outcomes at all.
None of this is unobvious or controversial. Epistemic probability judg-
ments of the sort in question just do not behave like frequency or pro-
pensity judgments with respect to theory confirmation in cases like these.
Such judgments are however simply not among those which the empiricist
tradition counted as epistemically legitimate. In fact, the failure to coun-
tenance such judgments as legitimate is simply a consequence of the gen-
eral empiricist denial of the legitimacy of inductive inferences from theo-
retical premises to theoretical (or observational) conclusions. In order for

p(E,o) = 1/2 to be deducible from T in the case we first considered, it would
have to be possible to deduce from T that p(T1) = 1/2. That, in turn,
would be possible only if there were auxiliary hypotheses or epistemic
axioms available which asserted the legitimacy of inductive inferences from
theoretical premises to theoretical conclusions. But these are inferences
which empiricist philosophers of science typically treated as merely heu-
ristic features of the "context of invention" and not as legitimate features
of the "context of confirmation". Thus, quite independently of the doctrine
that their account of statistical prediction and theory confirmation was
itself largely independent of the question of the interpretation of proba-
bility statements, logical empiricists would be obliged by their understand-
ing of the boundary between heuristic and epistemic factors in scientific
practice to deny the epistemic legitimacy of inductive inferences from the-
ories to observational predictions of the sort we have been considering,
and to deny the relevance in theory confirmation (or disconfirmation) of
empirical tests of those predictions.
It must be emphasized that the failure of logical empiricists to counte-
nance inductive inferences of this sort does not, by itself, constitute a crit-
icism of their position. It might, for all I have so far said, be true that such
inferences are epistemically illegitimate and irrelevant to theory confir-
mation. Alternatively, it might be possible to accommodate such inferences
within a genuinely empiricist conception of scientific methodology. In fact,
for reasons I will presently indicate, neither possibility is a real one. For
the moment however, it will suffice to observe that - whatever their epis-
temic status - inferences of the sort in question seem to be crucial deter-
minants of the direction of scientific research. As Kuhn 1970 correctly
observes, the establishment of a new "paradigm" depends upon scientists
becoming convinced, on the basis of some important successes, that the
problem solutions recommended by the paradigm will generally prove suc-
cessful. It is clear that the ways in which paradigmatic theories "recom-
mend" new problem solutions involve just the sort of inductive inferences
from theoretical premises we have been discussing. Researchers immersed
in the paradigm frame both questions and solutions in ways suggested by
the "ontological" (that is the theoretical) picture presented by the para-
digm. There is no reason to believe that such recommendations all follow
deductively from the theoretical claims embodied in the paradigmatic pic-
204 RICHARD N. BOYD
ture of the world. Instead, scientists who accept a paradigm adopt the
"world view" of the paradigm: they accept as credible those problem so-
lutions which are inductively reasonable on the supposition that the on-
tological picture presented by the paradigm is true (or approximately true).
They will in solving new problems look for mechanisms, particles, forces,
etc. analogous to those postulated by the theories which constitute pre-
viously accepted solutions within the paradigm to analogous problems; the
relevant respects of analogy will themselves be paradigm-determined.
Reasoning by analogy is, of course, an inductive rather than a deductive
procedure so, whatever the appropriate philosophical reconstruction of
their reasoning might be, scientists who accept the methodological guid-
ance of a paradigm are engaging in reasoning whose form is that of in-
ductive inference from theoretical premises to theoretical conclusions.
1.2 The Source of Empiricist Reticence Regarding Theoretical Inductions
It is, of course, no mystery why logical empiricists were disinclined to treat

as epistemically (as opposed to "heuristically") legitimate the sort of in-
ductive inferences from theoretical premises to theoretical conclusions we
have been discussing. Scientific knowledge, on the empiricist conception,
extends no farther than to knowledge about observables; when we confirm
a theory we confirm only its (approximate) empirical adequacy. Thus the
theoretical premises of the inductive inferences in question do not reflect
knowledge about "theoretical entities" no matter how well confirmed are
the theories from which they are drawn. The inductive inferences, however,
depend upon the presupposition that they do. An example will illustrate
the point.
Suppose that we have confirmed a variety of different theories which
(taken at face value) postulate a variety of different theoretical entities all
sharing the unobservable property P. Suppose further that all of the the-
oretical entities postulated by those theories which are said to have P are
also said to have the equally unobservable property Q. Suppose that at
some later time an additional theory is proposed and is independently
well-confirmed. It postulates an additional theoretical entity with the prop-
erty P, but it is silent on whether that entity has or lacks the property Q.
Imagine that someone reasons inductively thus: From the fact that all of
the previously established theories have postulated entities satisfying the
rule (if Px then Qx) and from theoretical considerations which suggest
that the rule is "projectable", I conclude that probably every entity which
has P has Q. Therefore, probably the entity with P postulated by the more
recent theory also has Q. From this inductively derived assumption, to-
gether with the theories in question and relevant well-confirmed auxiliary
hypotheses, I deductively predict that the observation statement 0 is true.
Therefore I have good inductive warrant to believe o.
It will be plain that the partly inductive inference described here is ep-
istemically illegitimate on an empiricist conception of the limitations of
scientific knowledge. On such a conception, all we have confirmed in con-
firming the various theories employed in the inference is their empirical
reliability (or, better, their empirical reliability when integrated into the
existing "total science"). We have confirmed, at most, that the observa-
tional predictions deducible from those theories taken together with other
well-confirmed theories are (approximately) true. We have not confirmed
the existence either of the various unobservable entities postulated in those
theories, nor the existence of the various unobservable properties in ques-
tion nor, therefore, the particular claims about the properties of unob-
servables upon which the inductive step in the inference depends. Indeed,
empirically equivalent systems of theories exist in infinite number which
do not postulate the same (or structurally analogous) entities at all but
which embody - if the empiricist conception of scientific knowledge is right-
just the same empirical knowledge. The inductive step in the argument we
are considering would be legitimate only on the realist conception accord-
ing to which scientific knowledge extends to unobservable "theoretical ent-
ities" and their unobservable properties.
Once it is clear why it would be epistemically illegitimate to inductively
derive observational predictions from well-confirmed theories in this way,
it is also clear why the empiricist conception of theory-testing counten-
ances experimental tests only of deductive (although perhaps statistical)
predictions. In testing a proposed theory we test only its empirical ade-
quacy, and we do that by testing suitably representative examples of those
predictions whose overall truth determines how nearly empirically ad-
equate the theory is. These are just the theory's deductive observational
predictions. (I will henceforth use this terminology to refer as well to those
"inductive statistical" predictions acceptable to the traditional empiri-
cists.) To test those predictions of a theory which arise from inductive
206 RICHARD N. BOYD
inferences of the sort we have been discussing would be irrelevant to the

assessment of its empirical adequacy.
Thus neither in making predictions from already accepted theories nor
in the experimental assessment of proposed theories are inductive infer-
ences from theoretical statements to observational predictions epistemi-
cally legitimate. Exactly the same considerations dictate the conclusion
that inductive inferences from theoretical premises to purely theoretical
conclusions are epistemically illegitimate for the consistent empiricist.
Finally, consider the inductive processes by which scientists invent pos-
sible theories to account for observed phenomena. For all we have said,
the empiricists' reluctance to offer an epistemic rather than a purely psy-
chological or pragmatic account of such inferential processes might be
entirely unrelated to their reluctance to offer an epistemic account of in-
ductive inferences from theoretical premises. Indeed, historically the mo-
tivations for distinguishing the "context of invention" from the "context
of confirmation" were quite complex and probably cannot be reduced to
a single philosophical concern. Nevertheless, more recent work in the phi-
losophy of science (see e.g., Hanson 1958; Kuhn 1970; Boyd 1979, 1982,
1983, 1984a, 1984b) suggests that the plausibility considerations which
(both as a psychological matter and as a matter of scientific methodology
in practice) constrain and guide theory invention involve just the sort of
inductive inferences from theoretical premises which we are here discuss-
ing. Thus the consistent empiricist does in fact have just the same sort of
reasons to deny the epistemic legitimacy of inductive inferences of this sort
that she has to deny the legitimacy of other inductive inferences from
theoretical premises to theoretical or observational conclusions.
We may summarize by saying that genuine inductive generalizations
from theoretical premises are problematical for the consistent empiricist
and that, given the way science is actually done, the inductive principles
governing theory invention are problematical for just the same reason. It
should be emphasized that these conclusions are largely uncontroversial
at least insofar as traditional logical empiricism is concerned: a largely
pragmatic or purely psychological treatment of theoretical considerations
in science was an intended consequence of empiricist epistemology of
science. Empiricists correctly saw such a treatment oftheoretical consider-
ations as dictated by their verificationist skepticism concerning the possi-
bility of knowledge of unobservable "theoretical entities." On the empi-
ricist view, the only forms of inference which are epistemically (as opposed
to merely heuristically) justified are deductive inferences and the sorts of
inductive reasoning by which scientists assess the "degree of confirmation"
of an already proposed theory, given a body of observational data.
1.3 Problems with the Empiricist Conception
Critics of logical empiricism have doubted that so narrow a conception of
the epistemology and methods of science can provide an adequate account
of scientific rationality. The actual (and apparently rational) methods of
science seem to be permeated with just the sorts of theoretical reasoning
which the empiricist must dismiss as merely heuristic or merely pragmat-
ically justifiable (Hanson 1958; Kuhn 1970; Putnam 1975a, 1975b).
Against such criticisms it is always possible for the empiricist to reply that
the features of scientific practice which resist incorporation into an em-
piricist epistemology are really either purely conventional or purely prag-
matic rather than constitutive of epistemically rational empirical methods
in science - appearances to the contrary notwithstanding.
In a series of papers (Boyd 1972, 1973, 1979, 1980, 1982, 1983, 1984a,
1984b) I have, in effect, sought to show that this empiricist response is
wholly inadequate. I argue that no adequate account can be given of the
inductive inferences the empiricist does accept as epistemically legitimate
(viz., those involved in the experimental confirmation of the empirical ad-
equacy of proposed theories) without the adoption of a scientific realist
conception of scientific knowledge according to which knowledge ofunob-
servable "theoretical entities" is possible (indeed, actual) and according to
which the sorts of inductive inferences from theoretical premises we have
been discussing are (under appropriate conditions) epistemically legit-
imate. My argumentative strategy has been to examine a large number of
methodological practices which are apparently epistemically crucial to the
experimental assessment of the empirical adequacy of theories and to ar-
gue, about each, that the epistemic contribution which it makes to such
assessments can only be explained or justified from a realist perspective.
To a good first approximation, there is no epistemically legitimate prin-
ciple of inductive inference apt for assessing the· empirical adequacy of
theories for which an empiricist (as opposed to a realist) account is ad-
equate. Induction about observables is deeply problematical from an em-
piricist point of view.
208 RICHARD N. BOYD
One important feature of the recent critiques of logical empiricism has

been their systematicity: they employ a family of closely related and mu-
tually reinforcing criticisms of empiricist conceptions of confirmation, pro-
jectability, reference, the definitions of kinds, explanatory power, etc.
What I propose to do in the present essay is to show how existing criticisms
of logical empiricism can be extended to a critique of even the logical
empiricists' conception of the role of deductive inferences in science. I pro-
pose to show the following:
First, deductive inferences from theoretical premises are epistemically
illegitimate from the perspective of the consistent logical empiricist. Not
only can the empiricist not offer an adequate reconstruction of the central
inductive principles in science, she cannot provide an adequate account of
deductive inference either. No rational principles of scientific inference are
compatible with a consistent empiricism!
Second, the empiricist is anyway quite wrong in holding that experi-
mental testing of proposed theories is restricted to tests of their deductive
(or their "inductive statistical" in the empiricist sense) predictions about
observables. When, in testing a proposed theory, scientists test observa-
tional predictions of the theory, it is almost never the case that those pre-
dictions can be deduced (even in the "inductive statistical" sense) from the
theory in question (together with relevant auxiliary hypotheses). Instead,
it is almost always the case that, when scientists test a proposed theory by
testing observational predictions inferred from it, the inferences in ques-
tion involve the sort of inductive inferences at the theoretical level which
even logical empiricists agree that consistent logical empiricism cannot
accommodate.
Finally, it is almost never the case that all of the observations which
constitute confirmatory evidence for a proposed theory represent tests of
its observational predictions, deductive or otherwise. The empiricist con-
ception that it is only the empirical adequacy of scientific theories which
can be experimentally confirmed has naturally led to a picture of experi-
mental testing according to which the observational tests which confirm
a theory are "direct" tests of its empirical adequacy: that is, tests of the
observational predictions of the theory itself, whose truth would constitute
its empirical adequacy. This picture is mistaken in two quite different ways.
In the first place, when tests of a theory's observational predictions (de-
ductive or inductive) do playa crucial role in its confirmation (or discon-
firmation) there will almost always (always in the case of the sort of mature
sciences we usually think about in the philosophy of science) be a very
large number of other observations which have an evidential bearing on
the theory in question even though they do not in any sense test its ob-
servational predictions.
Moreover, as Darwin's work must by now have taught us, it is perfectly
possible for a theory to be supported entirely by observations almost none
of which it predicts and to be quite well-confirmed nevertheless. This is
inductive "inference to the best explanation" (Harman 1965) with a venge-
ance. On the logical empiricist conception, such inferences are utterly mys-
terious, but the appropriate realist treatment of the role of unpredicted
observations in more typical cases of theory confirmation extends natur-
ally to a treatment of the special case in which predicted observations are
wholly or largely absent. In both sorts of cases, the non-predicted but
evidentially relevant observations are connected to the theory being as-
sessed by just the sort of inductive inferences from theoretical premises
which are epistemically legitimate but which consistent empiricists must
reject.
I tum my attention first to the issue of deductive reasoning in science.
2. LOGICAL EMPIRICISM AND DEDUCTIVE LOGIC
2.1 Inductive Inference, Projectability, and the Deductive Integration of

Theories: The Problem Stated
What distinguishes the philosophical tradition we call "logical empiricism"

from earlier stages in the empiricist tradition is essentially the incorpora-
tion, into traditional empiricist philosophical analyses, of the resources of
modem: mathematical logic. Loose talk about ideas and the operations of
the mind on them, and equally loose and imprecise talk about the semantic
properties of actual languages, were to be replaced by a more precise ap-
proach which modeled scientific theories as deductive systems in a formal
language, and which treated methodological issues as questions about per-
missible inferences in such languages rather than as questions about per-
missible operations of the mind. Central to the whole logical empiricist
project is the conception of scientific theories as (at least in a suitable
210 RICHARD N. BOYD
idealization) deductively integrated formal systems. A good part of the

attractiveness of this conception arose from the conviction that - however
many intractable problems might arise in the formulation of principles of
inductive logic - modem mathematical logic had provided the contem-
porary empiricist with an entirely adequate characterization of the prin-
ciples of deductive inference, whose methodological appropriateness was,
in any event, beyond question. The legitimacy of deductive reasoning was
seen as straightforwardly a priori, as grounded in paradigm cases of ana-
lytic truths.
That the confidence of logical empiricists in the legitimacy, for a con-
sistent empiricist, of the rules of deductive logic might have been misplaced
can be seen by examining the role of those rules in empiricist treatments
of the problem of "projectability". As Goodman (1973) and Quine (1969)
use the term, projectability is thought of as a property of predicates or, by
a natural extension, of (otherwise) law-like sentences or statements. We
may usefully employ a more general formulation of the notion of project-
ability which is well suited to an examination of the epistemological prob-
lems facing the consistent empiricist. On the consistent empiricist's view,
a posteriori knowledge is limited to instrumental knowledge - to knowledge
about regularities in the behavior of "observables". Now, for any finite
body of observational data, there will be infinitely many different patterns
or regularities to which they all conform - infinitely many generalizations
about observables for which they might be taken as evidence. The problem
of projectability is the problem of which of the regularities or patterns are
"projectable": are such that the observational data in question potentially
confirm them as genuine regularities in the behavior of observable phe-
nomena. Philosophers have ordinarily formulated this question in terms
of the projectability or "law-likeness" of the terms or sentences used to
express the regularities or generalizations in question: projectable regular-
ities are just those whose instances are deductively predictable from pro-
jectable or law-like sentences. Typically it is understood that the project-
able or law-like sentences in question may themselves contain theoretical
as well as observational terminology. Since what we are discussing is the
legitimacy from the perspective of the consistent empiricist of appeal to
deductive inferences from theoretical premises, it will be useful to begin
with the more general characterization of projectability as a property of
patterns or regularities in observable data rather than with the more tra-
ditional formulation in terms of the deductive subsumption of such pat-

terns or regularities under projectable or law-like theories.
I have elsewhere argued that no adequate account of projectability judg-
ments, nor of judgments of "degrees of confirmation" which depend upon
them, can be given which does not reflect a "scientific realist" rather than
an empiricist conception of scientific epistemology (Boyd 1972, 1973, 1979,
1980, 1982, 1983, 1984a, 1984b). My general argument has been that con-
siderations based on previously accepted theories constrain projectability
and degree of confirmation judgments in such a way that the epistemic
reliability of such judgments (even with respect to the establishment of
instrumental knowledge) cannot be adequately explained unless one sees
such theory-dependent judgments as applications of previously acquired
(approximate) theoretical knowledge to determine the current projectabil-
ity of both theoretical and observational predicates. If we think of pro-
jectability judgments in the more abstract way suggested here, then the
theory-determined constraints on the content and structure of the theories
we take to be potentially confirmable will not be the only constraints which
determine our judgments of projectability of patterns or regularities in
observational data. The fact, if it is a fact as empiricists claim, that once
we have identified the projectable or law-like sentences we treat them as
deductively integrated formal systems apt for the prediction of observable
phenomena also contributes to the determination of the patterns or reg-
ularities in observational data which we take to be projectable. It is, after
all, the set of deductive observational predictions from a body of law-like
sentences which - according to the empiricist conception - represents a
projectable regularity in observational data. What I hope to establish in
the succeeding sections is that the apparently innocent appeal to deductive
inference in determining projectability is problematic for the consistent
empiricist. The inductive role of deductive inferences in science - the role
of deductive inferences in determining projectability judgments - cannot,
I shall argue, be adequately explained or justified from the anti-realist
perspective of consistent empiricism.
The arguments for this conclusion will perhaps be clearer if we first exam-
ine the corresponding issue of the legitimacy of the inductive integration
of theories. We have already seen thatprimajacie the consistent empiricist
cannot countenance as epistemically legitimate inductive inferences from
theoretical premises to theoretical conclusions, or inductive inferences
212 RICHARD N. BOYD
from theoretical premises to observational conclusions in which inductive

inferences to theoretical conclusions form intermediate steps. The reason
is simply that inductive inferences from theoretical premises - projecting
theoretical predicates of unobservable theoretical entities, where the pro-
jectability judgments involved are themselves determined by theories of
unobservables - would appear to make sense only if one were in a position
to know that the theoretical premises (together with the other theoretical
assumptions involved in the projectability judgments for theoretical predi-
cates) were (at least approximately) true. Since consistent logical empiri-
cists deny that such knowledge is possible, they apparently cannot accept
such inferences as epistemically legitimate. A certain sort of inductive in-
tegration of theories must be unacceptable to the consistent empiricist be-
cause it, in effect, presupposes a realist rather than an empiricist concep-
tion of the extent and limits of possible scientific knowledge.
The parallel between the issues of inductive and deductive integration
of theories will, I believe, be made clear if we examine a variation on the
previous example. Suppose that a scientist reports to us that she has ex-
perimentally tested a number of observational predictions derived from a
theory T. Some of those predictions, we may suppose, were obtained de-
ductively as the standard empiricist conception of theory-testing would
suggest, but others, she tells us, were obtained inductively from T together
with other previously accepted theories in the way described earlier. All
of these observational predictions, inductive as well as deductive, she re-
ports, were experimentally confirmed. She proposes to take these experi-
mental findings as evidence that future observational predictions obtained
deductively or inductively from T will be approximately true. [The example
is, if what I will say later in this paper is true, an utterly typical piece of
scientific reasoning.]
Here too, the consistent empiricist must, at least prima/acie, reject the
proposal insofar as evidence for the truth of future inductive predictions
from T is involved. There are, after all, infinitely many different possible
regularities in (past, present and) future observational data compatible
with the finitely many observations in question. To see as projectable the
regularity predicted by inductive inferences from T - and to see it as pro-
jectable because of its inductive connection to T - would appear to be
legitimate only if one took the evidence in question as evidence that Twas
approximately true (even when its subject matter is unobservable). Induc-
tive inferences at the theoretical level have, after all, realist presupposi-
tions. But, of course, the consistent empiricist is no more going to hold
that claims about unobservables can be confinued as approximately true
by successful tests of their inductive predictions about observables than she
is going to hold that such claims can be confinued by successful tests of
their deductive predictions. There is apparently no justification acceptable
to the consistent empiricist for making the inductive inference about the
inductive predictive reliability of Twhich our imaginary scientist proposes.
[I say "apparently" and "prima facie" here not just because I want to
acknowledge the logical possibility of an empiricist reply. In fact, I have
in mind a quite definite empiricist reply and it is discussed in Section 2.4.]
We have just seen that the consistent empiricist must prima facie reject
the proposal that scientific theories be thought of as inductively integrated
systems for the prediction of observable phenomena. In particular, the
consistent empiricist must reject the sorts of projectability judgments which
are entailed by the conception of scientific theories as inductively inte-
grated systems. The proposal to treat scientific theories as inductively inte-
grated is however exactly analogous to the actual empiricist proposal to
treat scientific theories as deductively integrated predictive instruments. In
particular, deductive inferences at the theoretical level playa role in de-
termining judgments of projectability for possible regularities in observa-
ble data just like that played by inductive inferences in the scheme for
inductive integration of theories we have just discussed. Scientists identify
certain sets of (partly) theoretical sentences as projectable or "law-like"
[In fact they do this by deeply theory-dependent inductive considerations
(Boyd 1979, 1982, 1983, 1984a, 1984b; van Fraassen 1980)]. They then
identify, as projectable regularities in observational data, those regularities
which are suitably related to projectable or law-like sentences. On the
empiricist conception the suitable relation is deductive prediction of in-
stances so, on the empiricist conception of scientific theories as deductively
integrated, deductive inferences at the theoretical level play an absolutely
crucial role in detenuining judgments of projectability for regularities in
data.
What I propose to do in the following sections is to make, against the
empiricists' appeal to deductive inferences in this context, a criticism ex-
actly analogous to that illustrated above against the conception that in-
ductive integration of theories can serve as a legitimate basis for project-
214 RICHARD N. BOYD
ability judgments from an empiricist point of view. I will ask what is so

special about deductive inferences which explains why it is the deductive
observational consequences of "law-like" or projectable sets of sentences
which should be taken to represent projectable regularities in observational
data. In the case of inductive inferences at the theoretical level we saw that
the answer apparently would have to be that it is the projectable sets of
sentences which are potentially confirmable as approximately true, and
that we should accept the inductive predictions about observables which
follow from well confirmed theories because inductive inferences tend to
preserve approximate truth. I shall argue that the role of deductive infer-
ences from theoretical premises in determining projectability judgments
about possible regularities in observable data must have exactly the same
sort of justification. Projectable or law-like sets of sentences are those
which are confirmable as approximately true by observational evidence;
when we have confirmed such a set of sentences we are warranted in be-
lieving that its deductive consequences about observables will be (approx-
imately) true precisely because deductive inferences are always truth pre-
serving and are, often enough, approximate-truth preserving as well. If
this is right, then the appeal to deductive inferences from theoretical prem-
ises in determining judgments of projectability is justified only if theoretical
premises are objects of possible (approximate) knowledge: only, that is, if
a realist rather than an empiricist conception of scientific knowledge is
true. Such an appeal to theoretical knowledge is illegitimate from the per-
spective of the consistent logical empiricism. Logical empiricism has no
right to the rules of (deductive) logic!
I shall argue for this conclusion at length in the following sections. For
the present I want to reply to four obvious objections to the argumentative
strategy which I have indicated here.
Objection One: The employment of the laws of deductive logic cannot
possibly be epistemically illegitimate because those laws are analytically
true: they follow from the very meanings of the logical words.
Reply: Suppose for the sake of argument that the laws of logic are ana-
lytically true. What's at issue isn't their truth but the role played by appeals
to deductive inferences from "law-like" or "projectable" sentences in em-
piricist accounts of the judgments which scientists make of the projecta-
bility of patterns or possible regularities in observational data. It does not
follow from the claim that the laws of logic are analytically true that the
empiricist's application of those laws in defining projectability of possible

regularities is consistent with empiricist epistemological conceptions.
Objection Two: But it's analytic that judgments of the sort you say the
empiricist cannot accept are legitimate.
Reply: If it is analytic that such judgments are legitimate, then consistent
empiricist conceptions of the epistemology of science cannot accommodate
analytic truths about justifiable inferences, and empiricism is analytically
refuted. But it's not analytic that judgments of the sort in question are
legitimate. In fact, it's pretty obvious that such judgments presuppose the
approximate truth of the received theories which form the background for
scientists' judgments of the law-likeness or projectability of sets of state-
ments (Boyd, op. cit.). It is certainly not analytic that background theories
of this sort are approximately true.
Objection Three: Even if it is true that employing deductive inferences
at the theoretical level in making projectability judgments presupposes the
truth (or approximate truth) of the theoretical premises, it does not follow
that consistent logical empiricists cannot accept the legitimacy of such
inferences. Following, e.g., Carnap 1950, they could hold that the basic
theoretical principles which form the premises for such deductive infer-
ences are analytic truths ("L-truths").
Reply: While it might be possible to maintain that some of the basic
theoretical principles in science are L-truths, it cannot be true that all
projectable or law-like statements are analytic. If they were, then all of
their observational predictions would have to be analytic, and genuinely
contingent regularities in observable data would never be projectable. But,
of course, they are.
Objection Four: There cannot be any problem from any point of view
with the practice of formulating projectable regularities in observational
data by treating them as the deductive consequences of sets of laws. Any
recursively enumerable and deductively closed set of observation sentences
can be thought of as the set of deductive consequences of a recursively
axiomatized theory, so thinking ofprojectable regularities as those whose
instances are deducible from theories places no limitation on the classes
of regularities which can be identified as projectable.
Reply: The objection is not that logical empiricists illegitimately employ
recursive axiomatizations to characterize the sets of observational predic-
tions which they take to be subject to confirmation. The objection is that
216 RICHARD N. BOYD
the consistent empiricist cannot explain why it is just those sets of predic-
tions which follow deductively from "law-like" theoretical axioms which
are the projectable sets of predictions. Not every recursively enumerable
and deductively closed set of observation sentences follow deductively
from a "projectable" or "law-like" set of theoretical axioms. That's the
whole point of the empiricist conceptions of projectability judgments.
2.2 "Unity of Science": The Diachronic Deductive Integration of Theories
Quite early on in the development of logical empiricist philosophy of

science Neurath (1931a, 1931b) and Carnap (1934) recognized an impor-
tant principle of scientific reasoning which is absolutely essential to the
plausibility of the empiricists' conception of scientific theories as deduc-
tively integrated formal systems. The principle, which Carnap called the
"unity of science" principle, affirms that, under appropriate conditions,
independently confirmed theories may be conjoined in the expectation that
their conjunction will prove (approximately) empirically adequate. As
both Neurath and Carnap recognized, some such principle is necessary in
order to provide an adequate account of the ways in which scientists apply
theories in practice, and also in order to account for the role of "auxiliary
hypotheses" in theory testing. The aim of the present section is to show
that consistent logical empiricists cannot explain or justify the unity of
science principle and, therefore, that they cannot account for the ways in
which deductive inferences contribute to the testing or application of scien-
tific theories.
According to the unity of science principle, when different scientific the-
ories have been independently confirmed by experimental evidence, the
evidential considerations which support each of the theories independently
should be thought of as (at least tentatively) confirming the (approximate)
instrumental reliability of their conjunction. The principle is quite striking
because it implies that if two theories Tl and Tz are independently con-
firmed, in the sense that no observational tests have been made of predic-
tions which follow from their conjunction but not from either theory in-
dependently (that is, if neither has served as an "auxiliary hypothesis" in
the testing of the other), it will still be the case that the observations which
confirm Tl and Tz independently will provide evidence for the instrumen-
tal reliability of their conjunction. The instrumental reliability of the COD-
junction of two theories can be established by experimental evidence which

does not involve the testing of any of the observational predictions peculiar
to that conjunction!
Both Camap and Neurath recognized the epistemic novelty of this prin-
ciple, and they recognized as well that a plausible explanation for its ap-
propriateness might seem to lie in the assumption, unacceptable to logical
empiricists, that there is unity in nature and that that unity is reflected in
the theoretical content of independently confirmed scientific theories. In-
deed, Camap and Neurath both seemed inclined to offer the unity of
science principle as a "rational reconstruction" of metaphysical material-
ism (see especially Camap 1934). In fact, they were right to see the unity
of science principle as closely connected to non-empiricist metaphysics. It
is by now a well-established criticism oflogical empiricism that it is unable
to account for the epistemic legitimacy of the unity of science principle
(Putnam 1975a; Boyd 1971, 1982, 1983, 1984a, 1984b). Roughly, the argu-
ment that the unity of science principle requires a non-empiricist, scientific
realist explication goes as follows:
What the unity of science principle requires is that current confirmation
of a scientific theory provide evidence that it will prove instrumentally
reliable when deductively integrated with theories which have yet to be
invented, much less confirmed, and with theories which, although they
have already been established, were unknown to the scientists whose work
confirmed the theory in question. Experimental tests of the observational
predictions obtained by synchronic deductive integration of a proposed
theory with currently available "auxiliary hypotheses" are supposed to
provide evidence for the truth of the observational predictions which will
result from subsequent diachronic deductive integration of the theory with
newly confirmed theories or from previously unanticipated conjoint ap-
plications of the theory together with previously confirmed ones. Appro-
priate testing of the predictions which result from the synchronic deductive
integration of a proposed theory with currently accepted "auxiliary hy-
potheses" confirms not only its synchronic instrumental reliability but its
diachronic instrumental reliability as well. The additional knowledge in-
volved - the knowledge over and above the synchronic instrumental reli-
ability which has been "directly" confirmed - is represented in the theo-
retical structure of the theory in question and is "extracted" by deductive
integration of the theory in question with the similarly represented "excess
218 RICHARD N. BOYD
knowledge" embodied in the theoretical structure of other independently

well-confirmed theories. Moreover, the extraction of this excess knowledge
is contingent on univocality for the theoretical terms occurring in the the-
ories in question. The factors involved in actual univocality judgments are
exactly those which conform to a naturalistic conception of judgments of
co-referentiality for theoretical terms (Boyd 1979, 1982). For the consistent
logical empiricist, who denies that theoretical claims about unobservable
phenomena can be experimentally confirmed, there can be no rational jus-
tification for this procedure for the extraction of excess knowledge em-
bodied in the theoretical structure of empirically confirmed theories. In-
deed, for the consistent empiricist is is difficult to see how it would be
possible to explain at all why the instrumental reliability of deductive sys-
tems which have not yet even been formulated can be confirmed by ex-
perimental tests of the observational predictions of quite different formal
systems which share with the former systems only an apparent unobserv-
able subject matter about which, according to the empiricist, no knowledge
is possible.
For the scientific realist, by contrast, the unity of science principle is
easily explicable. Approximate knowledge of unobservable theoretical ent-
ities is possible, and univocality judgments regarding theoretical terms are
just judgments of co-referentiality. The unity of science principle is simply
the principle that permits the deductive integration of independently es-
tablished fragments of knowledge about the same unobservable (as well
as observable) subject matter. It is rational to expect independently con-
firmed theories to prove instrumentally reliable when conjoined (provided
that their theoretical terms occur umambiguously in the conjunction) sim-
ply because the conjunction of approximately true theories will typically
yield approximately true deductive predictions about observables. Thus
the scientific realist can, as the consistent empiricist cannot, provide a satis-
factory justification for the unity of science principle which Camap and
other logical empiricists recognized as central to the role of deductive logic
in science.
It is true, of course, that if the realist conception of scientific knowledge
is right then the logical empiricists' formulation of the unity of science
principle is flawed. In the first place, when we think of two independently
confirmed theories to which the unity of science principle has been applied
we ought not to think of two theories whose conjoint deductive predictive
reliability has somehow been assessed although neither has been used as
an auxiliary hypothesis in deriving a tested deductive prediction from the
other. Instead, it would be more accurate to think of theories whose con-
joint observational predictions, inductive as well as deductive, have some-
how been shown to be reliable even though they have been independently
confirmed. We ought to think of the independence of their confirmation
as being defined by the fact that neither served as an auxiliary hypothesis
in the deductive or inductive derivation of one of the tested predictions of
the other. Finally, as I shall argue presently, some observational data
which tend to confirm a theory need not be predicted by the theory at all,
deductively or inductively, but may instead be determined to be eviden-
tially relevant by various inductive inferences at the theoretical level.
Therefore, we should think of theories as also being confirmed by con-
siderations of this sort and we should think of two theories as being in-
dependently confirmed just in case neither figured in the evidential con-
siderations favoring the other, either as a premise in a deductive or in-
ductive prediction, or as one of the theoretical considerations relevant to
the assessment of other sorts of evidence in its favor.
It should be clear that adding these complications to the conception of
the unity of science principle deepens rather than attenuates the difficulty
facing the consistent empiricist. The consistent empiricist can justify
neither the empiricist version of the unity of science principle, nor the more
inclusive version of that principle which the scientific realist would affirm.
Since the (empiricists' version of) the unity of science principle just is
the principle that scientific theories can be thought of as diachronically
deductively integrated formal systems, it follows, as promised, that the
consistent empiricist cannot account for the diachronic role of deductive
inferences in scientists' inductive inferences about observables. It will be
useful however to formulate the problem for the consistent empiricist ex-
plicitly in terms of the role of deductive inferences in the solution to the
problem of projectability for patterns or possible regularities in observa-
tional data. Putnam (l975a) formulates the criticism of logical positivism
we have been discussing by arguing that, on the empiricist conception,
acceptance of a theory T amounts merely to accepting that T leads to
successful predictions in the sense that the deductive predictions of Tare
true. He points out, against the empiricist, that the property " .. .leads to
successful predictions" does not have the formal properties of truth and
220 RICHARD N. BOYD
that, in particular, from the fact that Tl and Tz each lead to successful
predictions, it does not follow (as the unity of science principle would
require) that their conjunction leads to successful predictions. Imagine that
the empiricist accepts Putnam's point, but argues (correctly!) that Putnam
has mis-assessed the empiricists' conception of scientific knowledge. To
accept a theory T is to accept that T has the property" .. .leads to successful
predictions when conjoined with other well-confirmed theories (past, pres-
ent or future) which contain no theoretical terms which would occur am-
biguously in the relevant conjunction". This property, the empiricist in-
sists, does have the relevant property of truth: it is preserved under ap-
propriate conjunctions.
One especially clear way to see why the empiricist's reply to Putnam's
argument would prove inadequate is to examine the notion of projecta-
bility to which it is committed. The empiricist is committed by the unity
of science principle to holding that the projectable patterns in observa-
tional data are just those which correspond to the deductive observational
predictions of projectable or "law-like" sets of sentences when, as auxiliary
hypotheses, one is permitted to employ all other projectable or law-like
sets of sentences which should, at any time be experimentally confirmed,
subject only to the constraint of univocality for the relevant theoretical
terms. Projectability is defined in terms of the deductive closure of (suitably
disambiguated formulations of) future theoretical innovations, subject to
constraints of judgments of projectability or law-likeness for such inno-
vations, which are themselves determined by inductive theoretical consider-
ations determined by future as well as past theoretical achievements (Boyd
1979,1982,1983, 1984a, 1984b). Patterns are projectablejust in case their
instances are deducible from theoretical proposals which come to be ac-
cepted as theoretically plausible in the light of the growth of our theoretical
commitments.
For Hume, in effect, projectability was defined by the natural inclination
of the mind to form associations of ideas. Whatever the disadvantages of
this formulation, at least it saved Hume from, say, having to define pro-
jectability in terms of beliefs about insensible "real essences" or other ob-
jects of early empiricists' derision. By contrast, logical positivists are
obliged, given the unity of science principle, to define projectability pre-
cisely in terms of the deductive consequences of our theoretical speculations
about insensible real essences! Worse yet, they are obliged to define it in
terms of the deductive consequences of future as well as past speculations.

There is simply no reason to suppose that the deductive observational
consequences of ongoing theoretical speculations in science are especially
apt for confirmation (are indicative of the projectable regularities in ob-
servable phenomena) except on the realist assumption that such "specu-
lations" can be knowledgable, that they reflect growing (approximate)
knowledge of unobservable phenomena, knowledge which is "extracted"
by deductive inferences.
van Fraassen (1980) objects against the argument just rehearsed that in
fact scientists rarely accept the conjunction of independently confirmed
theories. Instead, van Fraassen insists, they typically correct such theories
prior to their conjoint application as predictive instruments. It is often,
but not always, true that they do this, but the fact that correction often
precedes deductive integration of accepted theories only compounds the
difficulties for the consistent logical positivist. Under many circumstances,
the independent confirmation of two theories makes it epistemically jus-
tified to believe that the observational predictions deduced from their con-
junction will be true; consistent empiricists cannot offer any satisfactory
explanation for this fact. In those cases in which revision prior to con-
junction is called for, the reason is that even well established theories are
typically only approximations to the truth about "theoretical" (as well as
observable) entities. When scientists revise theories prior to deductive in-
tegration with other theories, theoretical considerations - essentially in-
ductive inferences at the theoretical level - determine their judgments
about which revisions are likely to make the theories more closely ap-
proximate the truth about "theoretical entities" (as well as observable
ones). The revised premises are derived by inductive inferences from pre-
viously confirmed theoretical premises. No adequate account of such in-
ductive inferences at the theoretical level is possible for the consistent em-
piricist (Boyd 1984a).
What this case illustrates is a quite typical (perhaps universal) phenom-
enon: Logical empiricists felt obliged, given their denial of the possibility
of theoretical knowledge, to reconstruct a wide variety of theoretical in-
ferences in science as deductive. A great many of these inferences actually
involve inductive inferences at the theoretical level for which no adequate
empiricist justification is possible. In those cases in which the actual infer-
ences do fit the empiricist model- that is, in those cases in which deductive
222 RICHARD N. BOYD
inferences occur legitimately at the theoretical level - no adequate empi-

ricist account is possible of their legitimate role in inductive inferences
about observables. In those cases in which deductive inferences are legit-
imate only with respect to inductively derived premises, the consistent log-
ical empiricist can account neither for the initial inductive inferences at the
theoretical level nor for the subsequent deductive inferences. In cases in
which the theoretical inferences are essentially inductive, the empiricist has
no account to offer. I conclude that, at least in so far as the diachronic
integration of theories is concerned, deductive inferences from theoretical
premises are as problematical for the consistent empiricist as inductive
inferences from theoretical premises were always acknowledged to be: the
consistent empiricist can offer a satisfactory account of neither (but see
Section 2.4).
2.3 "Total Sciences": Synchronic Deductive Integration of Theories
It has been recognized for a long time (at least since Ayer 1946; see the
Preface to the Second Edition) that the unity of science principle - and its
corollary, the role of auxiliary hypotheses in theory testing - pose serious
difficulties for the traditional logical positivist, In the first place, they ren-
der impossible the initially attractive positivist project of dividing state-
ments into those which are testable in principle and those which are not
so testable because they make no predictions about observable phenom-
ena. With the adoption of new theories, and thus new potential auxiliary
hypotheses, theories without testable consequences could come to have
such consequences; theories have or lack testable observational conse-
quences only in particular theoretical environments.
Similar difficulties plague the logical empiricists' conception of empirical
equivalence of theories. The basic empiricist argument against the possi-
bility of knowledge of "theoretical entities" is that for any theory of such
entities there will be empirically equivalent alternative theories which pos-
tulate logically incompatible accounts of unobservable phenomena. But
empiricial equivalence is well-defined only with respect to a fixed stock of
"auxiliary hypotheses": previously equivalent theories can yield different
predictions about observable phenomena in consequence of the discovery
of new theories - new potential auxiliary hypotheses. It would thus appear
that theories can be at best empirically equivalent at a time; but this is not
enough for the empiricist's argument against scientific realism.
I have argued elsewhere (Boyd 1982, section 3.1; and especially Boyd
1983, Part 3) that the best response for the logical empiricist in the face of
these difficulties is to adopt a holistic approach of the sort often advanced
by Quine and Camap according to which the proper objects of epistemic
evaluation in science are not individual theories as we normally conceive
of them but, instead, "total sciences". On this conception, all of the ac-
cepted theories at a time ought to be thought of as a single deductively
integrated "total science" whose acceptability is determined by the accu-
racy of its deductive observational predictions. Since such total sciences
contain all of the currently available auxiliary hypotheses, one can without
difficulty define the relation of empirical equivalence between them. The
central epistemological argument of logical empiricism can then be restat-
ed: for any total science at a time there will be (infintely) many empirically
equivalent total sciences which postulate pair-wise logically incompatible
accounts of unobservable entities. Since crucial experiments to choose be-
tween empiricially equivalent total sciences are impossible, there can never
be at any time reason to accept the theoretical picture presented by the
currently accepted total science over the pictures presented by its infinitely
many empirically equivalent rivals. Thus knowledge of unobservable the-
oretical entities is impossible: scientific knowledge is limited to instrumen-
tal knowledge.
I have argued (in the papers just cited) that this "total science verifica-
tionism" cannot provide an adequate account of the instrumental relia-
bility of scientific practice and that it must be abandoned in favor of scien-
tific realism. Nevertheless it might even so provide the logical empiricist
with the resources necessary to defend, as legitimate from an empiricist
perspective, the principles of deductive logic. If each new "total science"
is thought of as an entirely new and already deductively integrated theory,
then there is no such phenomenon as the diachronic deductive integration
of theories to account for. There is only the phenomenon of the synchronic
deductive integration which defines individual total sciences. It is my pur-
pose in the present section to show that, in addition to its other difficulties,
this "total science" conception of the objects of epistemic evaluation in
science cannot provide, for the consistent empiricist, a justification of the
role of deductive inferences in science.
It might seem trivially evident that the total science approach cannot
work. It has been widely observed that the total body of scientific theories
224 RICHARD N. BOYD
accepted at anyone time is likely to be syntactically inconsistent, and thus

to predict all possible observations. Total sciences like that can't possibly
be the only legitimate objects of epistemic evaluation in science. In fact,
the problem of the inconsistency of total sciences is a special case of a
phenomenon mentioned in the previous section: the body of accepted the-
ories (in the usual sense) at anyone time contains theories which are only
approximately correct and whose respects of approximation to the truth
are such that they require revision before they are conjointly applied in
making successful observational predictions. Just as the consistent empi-
ricist who thinks in terms of the diachronic deductive integration of the-
ories cannot adequately account for the theory-dependent ways in which
the required revisions are made, it is also true that the consistent empiricist
lacks the resources to explain the inductive inferences at the theoretical
level which would be required to transform the total body of scientific
theories accepted at anyone time into a deductively integrated and pre-
dictively reliable total science.
These considerations - when taken together with a wide variety of other
examples of cases in which logical empiricism cannot consistently account
for basic features of scientific methodology (Boyd, operibus cit.) - consti-
tute, I believe, adequate reason for rejecting an empiricist conception of
scientific knowledge in favor of a scientific realist conception. But it seems
to me that the failure of logical empiricists to be able thus to provide an
entirely adequate account of total sciences should probably not be taken
to sustain the more specific charge that the consistent empiricist cannot
account for the synchronic deductive integration of scientific theories. If
we focus our attention solely on the question of the synchronic deductive
integration of theories - if, that is, we set to one side other difficulties
facing the consistent empiricist in reconstructing the methodology of
science - then it is doubtful that we should take the considerations just
rehearsed as decisive against the empiricist's appeal to principles of de-
ductive logic in reconstructing the "logic of confirmation". We might, for
example, consider the empiricist warranted in holding that the holistic
units of confirmation in science - while much larger than individual the-
ories as we ordinarily understand them - are nevertheless not so inclusive
as the sort of total sciences just considered.
In particular, the holist empiricist could, with some plausibility, main-
tain that the "total sciences" which are the proper objects of epistemic
evaluation in science consist, not of all the theories accepted at one time
considered as a deductively integrated formal system, but instead of very
large families of ordinary theories thought of as deductively integrated in
this way. Such a total science would have the property that for any ordi-
nary theory which it contains, it will also contain all of the simultaneously
accepted theories which scientists would in practice employ as auxiliary
hypotheses in making deductive observational predictions from it, but to-
tal sciences in this sense would not contain either pairs of syntactically
inconsistent theories, or pairs of theories which scientists would not or-
dinarily think appropriate for conjoint application in predicting observa-
ble phenomena.
If we understand total sciences in this way, the problem of syntactic
consistency and of revision of theories prior to conjoint application is
finessed (albeit by reference to unanalyzed judgments of scientists in actual
practice). Such total sciences are only synchronically deductively integra-
ted. When, in the history of science, a situation arises which we might
ordinarily describe by saying that two independently confirmed theories
which have previously not been conjointly applied are first employed in
making deductive observational predictions according to the unity of
science principle, the holist empiricist will instead describe the situation as
one in which available evidence has come to confirm a new total science:
one which contains, among other ordinary theories, the two theories in
question. There will be no diachronic integration of previously confirmed
theories. Each new total science is to be thought of as independently con-
firmed by all of the previously acquired observational evidence which we
would ordinarily describe as having confirmed its constituent theories (in
the ordinary sense of the term). There will be no such thing as the diach-
ronic deductive integration of scientific theories to confound the consistent
empiricist.
What I propose to do is to argue that the consistent empiricist cannot
even provide an adequate account of the synchronic deductive integration
of total sciences of the sort we are now considering. Even for "medium
sized" and safe total sciences, the consistent empiricist cannot explain why
the success of finitely many of the deductive observational predictions of
such a total science can provide reason to believe that the rest of its de-
ductive observational predictions are likely to be (approximately) true. In
this case, as in the case of the diachronic deductive integration of theories,
226 RICHARD N. BOYD
I shall argue that deductive inferences from theoretical premises play a

role in judgments of projectability inexplicable from a consistent empiricist
point of view.
It is important to recognize that, although the strategy of taking (some-
what less than) total sciences as the basic objects of epistemic evaluation
in science eliminates (at least apparently) the need to countenance diach-
ronic deductive integration of scientific theories, it cannot eliminate other
essentially diachronic features of scientific methodology. The holist em-
piricist must, of course, offer some account of projectability judgments.
Instead of taking possible patterns in observable phenomena to be pro-
jectable if their instances are deductively predicted by theories in the or-
dinary sense which are themselves projectible or law-like, the holist em-
piricist will, quite plausibly, take projectable patterns to be those whose
instances are deductively predictable from projectable total sciences (in the
more cautious version of that notion). We are in a position to say with
some confidence how projectability for individual theories or total sciences
is assessed in actual scientific practice. There is virtual consensus, across
wide variations in philosophical positions concerning the epistemology of
science, that theories or total sciences are projectable just in case they
represent plausible continuations of the existing theoretical tradition (see
Goodman 1973 on "entrenchment"; Quine 1969 on kinds that " .. .issue
from theory full-blown"; Kuhn 1970 on the paradigm-dependence ofprac-
tice in "normal science"; van Fraassen 1980 on " ... appeal to previously
constructed theories to guide the experimental inquiry"; Harman 1965 on
"inference to the best explanation"; Putnam 1975a, 1975b on the theory
dependence of kind definitions; Boyd op. cit.). To a very good first ap-
proximation, we accept as projectable those theories which represent theo-
retically plausible modifications of previously accepted theories in the light
of new data or new theoretical considerations (Boyd, op. cit.). Putting the
matter in terms of total sciences changes nothing essential: we count as
projectable those total sciences which represent modifications of immedi-
ately preceeding total sciences plausible in the light of the theoretical con-
ception of those previous total sciences and in the light of whatever new
data or new theoretical considerations we mean to respond to. Projecta-
bility judgments for total sciences are essentially diachronic: they depend
upon judgments of theoretical plausibility determined by immediately
prior total sciences. It will now, I think, be clear why the appeal to de-
ductive logic in the characterization of projectable patterns in observa-

tional data is problematical for the holist empiricist. The holist empiricist
proposes that - from among the infinitely many different patterns com-
patible with the already examined observational data - those patterns
should be considered as potentially confirmable which are deductively pre-
dictable from those possible total sciences which are theoretically plausible
in the light of the theoretical commitments embodied in the existing total
sciences. In other words (and this characterization reflects the current con-
sensus as well as any of the others given previously) the holist empiricist
must consider projectable just those patterns deductively predictable from
possible total sciences which are themselves supported by considerations
which have exactly the form of inductive inferences from the theoretical
claims embodied in the existing total sciences. No other description is pos-
sible of the form of scientists' reasoning in such matters. As van Fraassen
says,
...To accept a theory is to make a commitment, a commitment to the further confrontation
of new phenomena within the framework of that theory, a commitment to a research pro-
gramme, and a wager that all relevant phenomena can be accounted for without giving up
that theory. That is why someone who has accepted a certain theory, will henceforth answer
questions ex cathedra, or at least feel called upon to do so ... (van Fraassen 1980, page 88)
To accept a theory (or, on the model we are considering, a total science)

is to behave in one's methodological judgments about, e.g., projectability
just as one would if one believed the theory (or total science) to be (ap-
proximately) true.
Consistent empiricists do not believe that one can warrantedly believe
that the non-observational content of the theories or total sciences one
accepts are approximately true. They cannot accept inductive reasoning
from theoretical premises as epistemically legitimate. How then, we must
ask, could the consistent empiricist justify the practice of counting as po-
tentially confirmable only those patterns in possible observable data which
are the deductive consequences of total sciences which are supported by
what look for all the world like inductive inferences from currently ac-
cepted total sciences? The great virtue of deductive inferences is, after all,
that they are truth-preserving. Unless scientists have good reason to be-
lieve that currently accepted total sciences are (in their theoretical as well
as their observation statements) true (or approximately true) and that pos-
sible total sciences supported by (what look like) inductive inferences from
228 RICHARD N. BOYD
them are therefore also likely to be approximately true, why should they
believe that, from among the infinitely many possibilities, it is only the
patterns which reflect the deductive predictions of possible total sciences
so supported which are the legitimate candidates for experimental confir-
mation? The answer, I suggest, is that without accepting the realist con-
ception according to which previous total sciences embody knowledge of
unobservable phenomena, no justification can be given for the way in
which deductive inferences at the theoretical level function in determining
projectability judgments about possible patterns in observable data.
It might seem that the problem just posed for the empiricist is a problem
about the empiricist's account of projectability judgements, but not spe-
cifically a problem about the empiricist's appeal to deductive arguments
with theoretical premises. Appeal to deductive inferences might seem to
be the one unproblematical feature of the overall holist empiricist account
of how projectability judgments are made. In reply, I propose to suggest
two alternative conceptions of how projectability judgments might be
made, which differ from the holist empiricist's conception only with respect
to the rules by which observational predictions are to be obtained from
projectable total sciences. Each of these proposals is scientific nonsense
but, I claim, the consistent logical empiricist cannot justify our obviously
sound judgment that they are nonsensical. I call these proposals (actually
they are families of proposals) "projective timidity" and "projective an-
archism". The cogency of each of them (such as it is) rests upon the fact
that at any time in the history of science there will have been only finitely
many deductive observational predictions made and tested in the whole
history of science. Only finitely many predictions will have been deduced
from the currently accepted total science(s) and the same will have been
true for their predecessors. Here are the nonsensical methodological pro-
posals:
Projective Timidity: Let PA be the set of all observational statements
actually deduced from an accepted total science in the whole history of
science including, of course, those observation statements which have been
employed in stating such total sciences. PA will be finite.
For any possible total science, T, let 'O(n' denote the class of deductive
observational consequences of T. Now for any set S of observation sen-
tences, let RS be the "rule of inference" which takes any total science T
as a set of premises and which permits the "prediction" of just the obser-
vation sentences which lie in the deductive closure of the intersection of

0(1) with the union of S and PA. Let "RS(1)" denote the observational
predictions generated from T according to the rule of "inference" RS.
Three things are obvious about the inference rules RS:
(1) RS(1) will always be a sub-set of 0(1), and it will often be a
proper sub-set.
(2) For any total science T which has actually been accepted at
any time in the history of science, all of the tested predictions
of T will lie in RS(1), no matter what set S is chosen.
(3) Like the rules of deductive inference, the "rule of inference"
RS will, for any recursive set S, generate a recursively enu-
merable and deductively closed set of observational predic-
tions, given any finitely axiomatizable (or even recursively ax-
iomatizable) total science T.
Human achievements being always finite, we may assume that any ac-
tual total science is finitely axiomatizable. Now here is the methodological
proposal of the timid projectionist: instead of taking the success of the
observational deductive predictions of the currently accepted total science
to be evidence for its approximate empirical adequacy, for the approx-
imate truth of the sentences in 0(1), we should be more modest and take
such successes as evidence merely for the approximate truth of the sen-
tences in RS(1) for some recursive S such that RS(1) is a proper sub-set
of 0(1), preferably for an S such that the predictions in RS(1) are much
less bold and "daring" than the extravagant predictions embodied in 0(1).
Imagine that the timid projectionist cites as a justification for this pro-
posal the oft-repeated claim of philosophers in the empiricist tradition that
their evidential conception is to be preferred to that of scientific realists
because it is more modest: it countenances weaker conclusions about the
world on the basis of any significant body of observational evidence than
does the realists' conception of evidence (see, e.g., van Fraassen 1980). The
timid projectionist points out that timid projectionism has the same ad-
vantage over the traditional holist empiricist conception of evidence: given
any body of evidence the timid projectionist will draw a weaker set of
conclusions about the world than will the evidentially "incautious" em-
piricist. We may imagine that the timid projectionist takes the really dif-
ficult issue in the epistemology of science to be that of how "bold" a choice
of the recursive set S is sufficiently safe for legitimate science.
230 RICHARD N. BOYD
Note that the timid projectionist's proposal is simply a proposal about

how to define projectability for patterns in observational data, and that
the timid proposal is exactly analogous to the holist empiricist's proposal.
Each of them accepts the same standards for projectability of total
sciences. The only difference is that, where the holist empiricist appeals to
deductive inferences with theoretical premises in defining projectability for
patterns in observable data, the timid projectionist appeals to RS-infer-
ences, which are weaker. Note that RS-inferences agree with deductive
inferences for observation sentences; the set of RS-consequences of a total
science is a deductively closed set of observation sentences. RS-inference
agrees with deductive inference for just those sentences whose truth-values
the consistent empiricist thinks are subject to experimental confirmation.
Projective Anarchism: For each total science Twhich has ever been con-
sidered in the history of science to date, let PA(1) be the set of observation
statements actually deduced from T during the history of science to date.
PA(1) will always be finite. Each such total science will have been pre-
sented historically by the publication of finitely many theoretical (and ob-
servational) statements. Thus each total science actually considered in the
history of science will be finitely axiomatizable. Call a recursive function
j(x,y) an anarchistic pseudo-logic just in case whenever n is the Godel
number for a (suitably encoded) finite set of sentences, the I-place function
f (n,y) enumerates a deductively closed set of observation sentences con-
taining all of the pure observation sentences in the finite set, and just in
case j(n,y) and j(m,y) enumerate the same set whenever nand m encode
the same sets of sentences. Call such a pseudo-logic historically appropri-
ate just in case for every total science T actually considered in the history
of science, PA(1) is a sub-set of the range of j(k,y), where k encodes all
of the finitely many sentences by which Thas as a matter of historical fact
been presented.
The projective anarchist's proposal is that, instead of taking the cur-
rently available observational evidence to confirm the empirical adequacy
of the current total science T, we should instead choose some historically
appropriate pseudo-logic f, and take the available evidence to support the
approximate truth of the observational predictions which lie in the range
ofj(k,y), where k encodes the historically appropriate finite axiomatization
of the currently accepted total science. The proposal is interesting, of
course, just in case the pseudo-logic chosen is one of the infinitely many
such "logics" which are such that the range of f(k,y) is different from the
set of deductive observational predictions of T. The projective anarchist
points out (correctly) that every actually tested deductive observational
consequence of T (that is, every deductive observational consequence of
T which was recognized as such and was subjected to observational test)
is also anf-consequence of T, so that thef-wise instrumental reliability of
T has been tested by the same experimental tests which have thus far been
employed to assess Ts deductive instrumental reliability, with exactly the
same results.
It is clear that the projective anarchist's proposal amounts to an alter-
native conception of projectability for possible patterns in observational
data. The holist empiricist proposes that total sciences be assessed for their
"law-likeness" or projectability and that patterns be considered projecta-
ble just in case their instances are deducible from projectable total sciences.
The projective anarchist proposes instead that those patterns be considered
projectable whose instances are f-wise predictable from law-like or pro-
jectable total sciences. The standards for projectability of total sciences is
the same: all that is different is that deductive inferences from theoretical
premises are replaced by f-wise inferences in the anarchist's proposal. As
in the case of the restricted principles of logical inference proposed by the
timid projectionist,J-wise inference agrees with deductive inference on ob-
servation sentences, i.e., on those sentences whose truth-values the con-
sistent empiricist believes it is possible to establish experimentally.
Nothing could be clearer than that the two proposals just presented are
epistemically unsound. Timid projectionism would lead to our underesti-
mating the extent of our intellectual achievements in the realm of obser-
vational knowledge. Projective anarchism would lead us to accept obser-
vational predictions such that there is not the slightest justification for
supposing that they will prove approximately true. I deny that the con-
sistent empiricist can account for the inappropriateness of these two al-
ternative conceptions of projectability relative to the empiricist's actual
proposal to take as projectable the patterns in observational data deduc-
tively predictable from law-like or projectable total sciences. I do not deny
that the consistent empiricist can show that timid and anarchistic stan-
dards of projectability for possible patterns are epistemically arbitrary,
and that there is no reason to believe that either of the proposed standards
is epistemically justified. What I deny is that the consistent empiricist can
232 RICHARD N. BOYD
explain why the standard of projectability defined in terms of deductive

inferences from theoretical premises is any less arbitrary.
It is, of course, true that the empiricist's conception of projectability for
patterns in observable data is a special case of each of the two alternative
proposals; if S is the set of all observation statements, for example, then
the timid proposal coincides with the standard empiricist conception of
projectability for patterns. Similarly, ifthe pseudo-logic/is chosen so that
f(k,y) enumerates the deductive consequences of the finite set coded by k,
for all k, then anarchism will coincide with projective normalcy. The prob-
lem for the consistent empiricist is to provide some justification for think-
ing that these special cases are less arbitrary from an epistemic point of
view than their fellows. It is clear why the consistent empiricist faces great
difficulty here. Plainly the epistemic virtue of deductive inferences, which
really explains why projectability judgments defined by them are epistem-
ically privileged, is that they preserve truth. But consistent empiricists do
not believe that, in confirming total sciences, we confirm even the approx-
imate truth of the theoretical claims they embody. Thus they cannot jus-
tify, as the scientific realists can, the obviously correct preference for de-
ductive inferences in this case. The consistent empiricist, therefore, cannot
defend the empiricist conception of projectability which is defined in terms
of deductive inferences.
It should be emphasized that the consistent empiricist cannot reply, in
defense of the role assigned to deductive inferences from theoretical prem-
ises, that it is just such inferences which are dictated by the diachronic
unity of science principle; we have already seen that this principle cannot
itself be justified by consistent empiricism. Nor will it do for the empiricist
to insist that the recursively enumerable sets of observation sentences gen-
erated from total sciences by the deviant principles of inference under
discussion could as well be deduced from some finitely axiomatized theory
or other. This is indeed the case, but the problem for the empiricist is to
justify the procedure of counting as projectable the deductive observa-
tional consequences of (not just any but) just those total sciences which
meet ordinary standards of "law-likeness" or "projectability". The only
plausible justification for this procedure lies in the realist conception that
the projectability judgments dictated by the procedure represent inductive
inferences at the theoretical level from previously acquired (approximate)
theoretical knowledge. Only this (non-empiricist) conception can provide
an epistemic rationale for the empiricist's appeal to deductive inferences in

defining projectable patterns in observational data.
It might seem that, for the case of projective anarchy at least, the con-
sistent empiricist can identify two epistemic virtues which deductive infer-
ences possess andf-wise inferences lack (except, of course, for the special
case of pseudo-logics f which dictate inferences coincident with those dic-
tated by deductive logic). In the first place, deductive inferences do not
introduce logical inconsistencies in observational consequences when the
total science employed as a source of premises is itself logically consistent.
This is not true of all historically appropriate pseudo-logics. The consistent
empiricist might argue that, since inconsistent observation sentences can-
not be true, the inconsistency-avoiding properties of deductive logic pro-
vide an epistemic basis for preferring projectability judgments defined in
terms of deductive inferences from theoretical premises. The use of de-
ductive inferences provides a kind of insurance against the most spectacu-
lar sort of failure of predictive adequacy: the prediction of logically in-
consistent observable states of affairs.
It should be observed first that this argument, if successful, would not
provide any justification for accepting as epistemically justified the pro-
jectability judgments dictated by the standard empiricist conception. At
best, it would show that the anarchist proposals are even less justified from
the perspective of the consistent empiricist than is the empiricist proposal
to employ deductive logic in characterizing projectability for patterns in
observational data. Secondly, the use of deductive logic in defining pro-
jectability provides only a partial guarantee against inconsistency in the
deductive observational consequences of the total sciences one accepts. It
is sometimes possible to prove that a total science is itself consistent, and
to thus be assured that it will have no inconsistent deductive observation
consequences. But consistency is not a decidable property of total sciences,
so scientists must often go without this sort of assurance and "make do"
with the assurance provided by the evidence (which the consistent empi-
ricist cannot reconstruct) that methodologically acceptable total sciences
are likely to be approximately true (and thus at least consistent).
Now, for many cases of a total science T and a pseudo-logic J, it will be
possible to prove the consistency of the observationalf-wise consequences
of T. Where this is not possible, the anarchist will be, in so far as the
consistent empiricist is concerned, in just the same situation as the friend
234 RICHARD N. BOYD
of deductive logic for whom no consistency proof is available. Recall that

the consistent empiricist cannot reconstruct the assurance provided the
friend of deductive logic by the comforting evidence that the well-con-
finned total science is likely to be approximately true, so the consistent
empiricist cannot contrast the position of the anarchist with the position
of the friend of deductive logic who has such prima facie assurance of the
approximate truth (and thus the consistency) ofthe relevant observational
predictions.
Finally, note that, just as the consistent empiricist can hold that pro-
jectability judgments are revisable in the light of new evidence, the an-
archist can quite consistently recommend that one should revise one's
choice of pseudo-logic should inconsistencies in observational predictions
persist even though one has tried various alternatives from among the
"law-like" or "projectable" total sciences available. We know that this is
the appropriate solution only for the special case in which the new pseu-
do-logic chosen coincides with deductive logic, but the consistent empiri-
cist cannot explain why we are justified in that belief.
A second argument apparently available to the consistent empiricist is
this: The anarchist proposes to take available observational evidence to
support the f-wise predictive reliability of the currently accepted total
science rather than its deductive predictive reliability. But the fact that all
of the deductive observational predictions of the total science which have
actually been tested are also f-wise predictions does not, even prima facie,
provide reason for taking the available observational evidence to support
itsf-wise predictive reliability. After all, somef-wise predictions from the
total science in question will (in interesting cases) not be among its de-
ductive predictions and it might, for all the anarchist has said, be true that
observations already made ·refute some of these f-wise predictions of the
total science in question. The total science's f-wise predictive reliability
might thus have already been experimentally refuted under the circum-
stance envisioned for its confirmation by the anarchist.
It is true, of course, that under the circumstances envisioned observa-
tional data already collected at some time in the history of science might
have refuted the f-wise predictive reliability of the total science. It is sim-
ilarly true that, when all of the predictions actually deduced from a total
science have been confinned by experimental test, it may still be true that
data already collected refute some deductive observational prediction of
the total science; all that is required to make this possible is that a deduc-
tion of the relevant prediction not yet have been carried out. What is true
when a total science has been perfectly successful in passing experimental
tests is that all of the observations which were taken to be tests of the theory
have turned out in its favor. That provides no guarantee that other data
already collected will not come to be recognized as refuting the theory
when more of its deductive observational consequences are known.
The situation with respect to f-wise predictions is exactly the same.,
Under the circumstances envisioned, all of the observations which scien-
tists have taken as confirming the total science will have been both de-
ductive and f-wise consequences of the total science. It is true that some
as yet underivedf-wise predictions might be falsified by data already col-
lected, but this possibility does not distinguishf-wise predictions from de-
ductive ones. In neither case is there any gu~rantee that when the predictive
consequences of a total science are better understood it will not tum out
that previous collected data refute them. It is true, of course, that since the
patterns identified as projectable by the anarchist are not really projectable
(and are thus not really supported by the available evidence) the risk that
this will happen is greater for the generalizations which the anarchist
favors than for those favored by the friend of deductive logic. But, as we
have seen, the consistent empiricist cannot provide a justification for this
obviously correct claim.
It should be emphasized that the consistent empiricist cannot acceptably
reply at this point by insisting that one can provide pragmatic justifications
for defining projectability of patterns in terms of deductive inferences from
projectable or law-like theories rather than in terms of the deviant con-
ceptions of inference. It is true, of course, that there are a great many
pragmatic reasons for preferring the employment of deductive logic in
defining projectability. The principles of deductive inference are familiar;
they have already been (rationally) accepted for inferences involving ob-
servation sentences; the mathematical practice required for their applica-
tion in theoretical situations is already well developed, whereas we have
no developed practice in applying the various deviant logics proposed by
the timid and the anarchistic projectionists. It is also, perhaps, true that
the principles of deductive logic are analytic, so that the timid and anar-
chistic proposals require that we use logical terms in our theoretical (or
even partly theoretical) sentences with different meanings than those terms
236 RICHARD N. BOYD
possess in observation sentences; perhaps it would be more convenient to

keep the meaning oflogical terms constant throughout scientific usage. All
of these considerations - and no doubt more still - make the choice of
deductive principles of inference at the theoretical level vastly more con-
venient, and more tractable psychologically. No doubt all other things
being equal it would be more rational to prefer a definition of projectability
which referred to deductive inferences from theoretical premises than to
prefer one of the proposals for defining projectability offered by the timid
or the anarchistic projectionist.
The point of the present section is just that, for the consistent empiricist,
all other things are equal! There is, in particular, no reason to suppose that
the inductive generalizations about observables recommended by the pro-
jectability judgments of the friend of deductive logic will be any more likely
to be true than those generalizations recommended by the friends of the
various deviant "logics". The rational reasons available to the consistent
empiricist for preferring to employ deductive logic in defining projectabil-
ity of patterns do not include the resources for providing any epistemic
justification for assigning thisrole to deductive inferences. The consistent
empiricist has good rational reasons for wishing that it be so that the
methodological recommendations of the friend of deductive logic be the
epistemically reliable ones, but wishing does not make it so. The consistent
logical empiricist cannot account for the role of logic in scientific inquiry
(but see Section 2.4).
2.4 The Inductive Justification ofInduction: One Last Chance for Pragmatic
Justification of Methodological Principles
I argued in the last section that the consistent empiricist cannot rebut the
charge that she cannot provide an epistemological justification of the role
assigned by empiricists to deductive logic in determining projectability
judgments by offering to provide a pragmatic justification of the choice of
deductive rules of inference. The point at issue is important with respect
to our understanding, not only of the issue at hand, but also of other issues
which separate scientific realists from philosophers in the tradition of log-
ical empiricism. It is a characteristic feature of empiricist work in the phi-
losophy of science to treat as purely pragmatic those theory-dependent
features of scientific methodology in which the scientific realist sees an
ineliminable tacit commitment to a realist conception of scientific knowl-

edge (see, e.g., van Fraassen 1980 on "pragmatic virtues"). It is, I think,
obvious on careful reflection (see Boyd 1984a) that when methodological
principles playa genuinely epistemic role in scientific inquiry a purely prag-
matic justification of them does not suffice for the empiricist project of
showing that such principles do not require a realist explication. Indeed
so obvious is this point that I think that it may be a mistake to take
empiricist appeals to pragmatic justifications in such matters at face value.
I propose, instead, to offer a plausible and charitable reconstruction of
such appeals and to examine, in particular, its possible application to the
issue of the role of deductive logic in defining projectability for patterns
in observational data.
Here is the reconstruction. What I think is really going on when an
empiricist philosopher provides a purely pragmatic justification for some
methodological practice M (say, a practice which might seem to require
a realist justification) is this: The empiricist recognizes that the most ob-
vious epistemological justification for M is unacceptable to consistent em-
piricism. The empiricist proposes that initially our justification for em-
ploying M is the pragmatic justification provided. But the empiricist does
not (appearances to the contrary notwithstanding) take the initial prag-
matic justification for M to be the only justification available. After we
have been using M for a while, the empiricist reasons, we are able to make
an empirical assessment of M's reliability: we can see whether or not our
use of M has contributed to our success at obtaining instrumental knowl-
edge. If it has, then we have an additional justification for continuing to
use M: it has proved reliable in past scientific practice and we therefore
expect that it will prove reliable in the future as well.
On the reconstruction just offered, it must be noted, the overall empi-
ricist justification for M is not merely pragmatic. The purely pragmatic
considerations justify our initial choice of M as the method to try first, but
the subsequent additional justification is utterly epistemic. We justify the
continued application of M by an inductive inference about its epistemic
reliability; we have an inductive justification of an inductive practice and
if the inductive inference involved is sound, then we have good reason to
believe that the practice in question is conducive to the adoption of (ap-
proximately) true beliefs about observables (see Quine 1969, Goodman
1973). If this sort of inductive justification is available to the consistent
238 RICHARD N. BOYD
empiricist, then an adequate empiricist justification can be provided for

the practice M even though its justification initially seemed to require real-
ist premises. In the case of the methodological practices discussed in the
present essay - the practices involved in the diachronic and the synchronic
deductive integration of theories and the associated standards for project-
ability of patterns - the consistent empiricist might claim that these prac-
tices were initially justifiable only on pragmatic grounds but that the evi-
dent success (in inductive generalization about observables) which these
practices have made possible justifies us (epistemically) in continuing to
follow them. Since the conclusion of the inductive inference about induc-
tion - the conclusion that the methodological practices in question are
conducive to approximate knowledge about observables - does not itself
refer to unobservable phenomena, the empiricist could consistently point
to its empirical confirmation (rather than the tacitly realist justifications
offered by the realist) as the real epistemic justification for the practices in
question. Inductive inferences about inductive inferences would provide
the empiricist's justification for deductive inferences in theoretical science.
I have elsewhere discussed empiricist justifications of this sort for meth-
odological principles whose justification would seem to require the adop-
tion of a realist conception of scientific knowledge (Boyd 1983, Part 8;
1984a, Part III). I do not deny that in such cases the crucial inductive
inference from past methodological successes to the conclusion that the
methods in question are epistemically reliable is justified. I do argue, how-
ever, that the consistent empiricist cannot justify the inference. The reason
is this: in order to make the inductive inference about induction in ques-
tion, one must solve the problem of projectability for patterns in our scien-
tific practice. For any method M which prima facie has only a realist jus-
tification, but which has proven successful in the past, there will be infi-
nitely many quite different methods (which make pairwise different recom-
mendations about future practice) which are such that their recommenda-
tions coincide with those of M for all the actual cases in which M has been
applied. Thus the actual methodological success to date would have been
recommended by infinitely different possible "methodologies". In order to
decide that it is the method M whose epistemic reliability has been con-
firmed by the historical record, we need to justify the choice of M over all
of the other "methods" in question. We need, in other words, some prima
facie reason to think that it is M, rather than any of the alternative
methods, whose capacity to guide the search for (approximate observa-

tional) truth has been exhibited in our past successes. This reason cannot
be supplied by the inductive inference from past successes to the epistemic
reliability of M; the reason we seek is a prerequisite for that inference.
Instead, we must rely on the more ordinary tacitly realist reasons which
the consistent empiricist cannot accept. (For a fuller presentation of this
and closely related arguments regarding empiricist employment of induc-
tive justifications of inductive strategies see Boyd 1983, Part 8). The in-
ductive justification of methodological practices may well be very impor-
tant in both science and philosophy (indeed I think it is essential because
a priori justifications are never possible; see Boyd 1982, 1983, 1984a,
1984b). But if the argument just sketched is successful, inductive justifi-
cation of this sort cannot save the consistent empiricist from the challenge
of scientific realism.
If we now examine the particular case of the role of deductive inferences
from theoretical premises in determining judgments of projectability for
patterns in observational data, a natural variation on the strategy in ques-
tion permits a rebuttal to the empiricist's attempt to justify by induction
the methodological practices at issue. The point is easiest to see in the case
of the synchronic deductive integration of total sciences. All of the deviant
logics proposed by the timid or the anarchistic projectionist have the prop-
erty that the patterns they identify as projectable coincide with the patterns
diagnosed as projectable via the application of deductive logic in all cases
in which observational data have been taken to be evidentially relevant in
the history of science to date. Thus the actual predictive successes we have
had to date, which constitute evidence for the epistemic reliability of the
standard conception of projectability, would all have been obtained if we
had instead used anyone of the proposed deviant logics in defining pro-
jectability of patterns. In order to take our successes as evidence for the
epistemic reliability of the standard conception of projectability we need
reasons (reasons not supplied by the inductive inference we are consider-
ing) to attribute those successes to the standard conception rather than to
any of its rivals. We need, in short, to have a reason to believe prima facie
that the projectability judgments dictated by the standard conception and
not those dictated by the rival conceptions will prove most successful in
the future. If we have such reasons, then the actual successes we have
achieved will substantiate the prima facie case, but without such reasons
240 RICHARD N. BOYD
we cannot determine which of the infinitely many different conceptions of

projectability at issue is inductively supported by the successes of our prac-
tice to date.
Of course we have such reasons: they are the obvious scientific realist
reasons for preferring to define projectability of patterns in terms of the
full class of truth-preserving rules of logical inference in preference either
to some sub-class of these or to some principles of inference which are not
even truth-preserving. We are right - and these reasons show that we are
right - to take actual history of scientific practice as inductively justifying
our continued confidence in our standard procedures for assessing pro-
jectability. But the consistent empiricist cannot appeal to those reasons,
and the inductive justification of the role of deductive logic in determining
projectability judgments does not provide the consistent empiricist with
any way of justifying the appeal by empiricists to deductive inferences in
their account of projectability.
If we turn from the question of the synchronic deductive integration of
total sciences to the question of their diachronic deductive integration, the
difficulties facing the empiricist who appeals to inductive inferences about
inductive practices to justify the unity of science principle are essentially
the same. It is true that our successes in predicting the behavior of ob-
servable phenomena, which have depended in part on our accepting the
unity of science principle, provide inductive evidence for the epistemic
reliability of that principle. But there are infinitely many different princi-
ples for the integration of independently confirmed theories which would
have recommended just the same instances of deductive integration as the
unity of science principle for the finitely many actual cases in which it has
been applied, but which would dictate other patterns of integration in
other cases. Such alternative principles for diachronic integration can be
generated, for example, by specifying deviant "logics" which, like the de-
viant logics previously discussed, coincide with deductive logic in actual
cases and by proposing that independently confirmed theories be united,
not deductively, but via the inferences dictated by various ones of the
deviant logics. Likewise, one might propose the deductive integration of
independently confirmed theories but propose that judgments of univo-
cality for theoretical terms be defined by principles quite different from
the tacitly realistic standards actually employed, but by principles which
nevertheless yield the same judgments in all actual cases to date. However
one constructs these alternative methods for diachronic integration of the-

ories, the inference from actual historical successes to the conclusion that
the unity of science principle is epistemically reliable will depend on prior
prima facie reasons for taking the unity of science principle, and not some
of the others, to correctly characterize the reliability-making features of
our previous methodological practices. In this case too there are such
reasons, but they are just the realist reasons for thinking the unity of
science principle to be epistemically legitimate. These reasons are not avail-
able to the consistent empiricist, and the empiricist's attempt to employ
inductive reasoning about inductive methods to justify the unity of science
principle fails by empiricist standards.
We may therefore conclude that the reconstructed version of the em-
piricist's pragmatic justification for the role of deductive logic in defining
projectability fails to save the day for the consistent empiricist. Even the
empiricist who is prepared to offer an inductive justification for deductive
logic cannot do so in a way consistent with empiricist anti-realism. Con-
sistent logical empiricists must do without logic.
3. PREDICTIONS AND THEORY-TESTING
3.1 Are Deductive Observational Predictions What Scientists Test?
For the traditional empiricist, the testable predictions of a theory must be

its deductive predictions about observables (including the "inductive sta-
tistical" predictions which arise when suitable probability statements are
deducible from the theory in the way discussed in section 1.1). Only de-
ductive predictions of this sort appeared to empiricists to be epistemically
legitimate in the light of their verificationist commitments. We have seen
that, in fact, not even deductive predictions from theoretical premises are
legitimate for the consistent empiricist; defining projectability in terms of
such deductive predictions requires a distinctly realist justification. More-
over, the reason ~hy consistent empiricism cannot justify the role which
deductive inferences play in defining projectability (on the empiricit's own
account) is that projectability judgments actually reflect epistemically legit-
imate inductive inferences from (approximate) theoretical knowledge and
the consistent empiricist cannot account for the role which deductive in-
242 RICHARD N. BOYD
ferential steps play in such inductive inferences at the theoretical level. We

are forced to accept a scientific realist conception of scientific knowledge
according to which theoretical knowledge is possible and according to
which, therefore, inductive inferences from theoretical premises are epis-
tenlically legitimate. Thus neither of the traditional empiricist's reasons
for holding that it is the deductive predictions of theories which scientists
test are sound: it is neither true that inductive inferences from theoretical
premises are epistemically problematical nor is it true that empiricists can
consistently account for the role of deductive inferences from theoretical
premises in inductive inferences in science.
It is, therefore, reasonable to ask whether, even as a good first approx-
imation, we should accept the empiricist conception that, in testing and in
applying scientific theories, it is only their deductive observational conse-
quences which are relevant. If theoretical knowledge is obtained in science,
and if inductive inferences from theoretical premises are legitimate, then
might we not have a better conception of the actual practice of scientists
if we understood the application and testing of scientific theories to extend
to their inductive as well as their deductive observational consequences?
I think that the answer is almost certainly that only on such a conception
of the matter can we adequately understand the epistemology of actual
scientific practice. Scientific knowledge is in a very deep sense inductively
integrated, and a conception of theory testing or theory application which
does not reflect that fact is likely to miss important features of scientific
methodology. The depth of the inductive integration of scientific knowl-
edge does, however, pose a problem for the formulation of the view that
it is often the inductive rather than just the deductive consequences of
theories which we test or apply. Suppose that T is some theory and that
p is a prediction of T which is inductively derived from T together with
various previously accepted "auxiliary hypotheses". Since inductive infer-
ences from previously acquired theoretical knowledge will typically yield
further (approximate) knowledge, it will always be possible to portray the
situation as one in which the conditional "if T then p" is inductively de-
rived from background theoretical knowledge and is thus itself available
as an auxiliary hypothesis. On this understanding, of course, the prediction
p can be thought of as a deductive consequence of T when T is taken
together with well established auxiliary hypotheses. I propose to avoid this
difficulty by understanding by a theory or an auxiliary hypothesis some
relatively general theory of the sort which might be explicitly presented in

a textbook or a scientific paper, and to exclude from consideration in-
ductive generalizations from background theories which would be made
explicit, if at all, only in order to render deductive some particular obser-
vational prediction or other. I think that such a procedure is consonant
with the usual conceptions of theories and of auxiliary hypotheses, in-
cluding the conceptions of the logical empiricist whose model of theory-
testing I am proposing to challenge. In fact, I think that we can usefully
think of the contrasting class of "special purpose" inductive generaliz-
ations from established theories as providing a test for the claim that an
observational prediction from a theory T is an inductive, rather than a
deductive prediction from T. Suppose that, in order to render the predic-
tion of p from T deductive, we need to employ as premises statements
which are in content quite specific to the particular situation referred to
in p, which would not ordinarily be published as explicit theories, and
which we obtain by inductive inferences at the theoretical level from more
general statements which form parts of theories in the ordinary sense.
Under these circumstances, it is appropriate to think of p as being, in the
relevant sense, an inductive consequence of T (together with appropriate
auxiliary hypotheses) and not a deductive one.
If we adopt this conception of the matter, then in a great many cases in
which we test theories by testing their observational predictions and in
virtually every case in which we apply predictions of previously established
theories in practical applications, the predictions will be inductively rather
than deductively derived. It is a commonplace that when scientists conduct
experimental tests of theories they must ascertain that the experimental
design includes controls for possible artifacts - possible causal factors
which might influence the outcome of the experiment in ways which are
irrelevant to the intended test of the theory in question. Thus when a
scientists derives from a theory T the prediction that under observable
conditions C the observable outcome 0 will be obtained, she must - if
testing this prediction is to be relevant to the assessment of T - be confident
that circumstances satisfying the conditions C embody relevant controls
for experimental artifacts. Should this not be the case, the prediction "If
C then 0" will be inappropriately derived from T; rational acceptance of
Twould not dictate acceptance of the observational prediction in question.
Thus, if the scientist's derivation of the prediction in question from Tis
244 RICHARD N. BOYD
to be rendered deductive, she must be thought of as employing as a premise

some such statement as the following: "Under circumstances satisfying C,
the only causal factors which affect parameters Pt. ... Pn (within the rele-
vant error margins) are It. ... fk", where Pi ... , Pn are the observable
parameters whose values are specified in the predicted outcome 0, and
where 11, ... fk are all of the (typically unobservable) causal factors which
the scientist has actually taken account of in her derivation of the predic-
tion "If C then 0" from T. Now in any such case there are infinitely many
logically possible causal factors which might affect the parameters defining
the expected outcome O. In ascertaining that the conditions C embody
appropriate controls the scientist will however require controls for only a
small finite number of these possible artifacts: she will control for those
possible artifacts which are plausible in the light of currently accepted
theories. In all but the very simplest cases the relevant plausibility judg-
ments will represent inductive inferences from the relevant background
theories, and there will be only inductive reasons, given those premises, to
believe that no artifactual causal factor has been left uncontrolled for
(Boyd 1983, Part 2). Thus the crucial premise in the reconstructed deduc-
tive inference of "If C then 0" from T represents an inductive consequence
of the relevant background theories, with a content quite specific to the
particular observational prediction in question. Plainly the crucial premise
is not the sort of "theory" which would ordinarily be published as an
explicit theory (except as a part of a deductive reconstruction of the rele-
vant observational prediction). I conclude, therefore, that in the typical
case in which a scientist tests a theory by testing one of its observational
predictions, the prediction in question will have been inductively rather
than deductively derived from the theory in question together with relevant
auxiliary hypotheses.
The case is even stronger in the case of theory testing under naturally
occurring situations, or in the case of the application of previously estab-
lished theories to real-world situations. In such cases, when a prediction
of the form "If C then 0" is derived from a theory T the conditions C
will have been determined by the real world rather than by the design of
the scientists, so it will be even more difficult to confirm the "no artifacts"
premise required to make the inference deductive, and it will be corre-
spondingly less likely that this premise can be obtained deductively from
the relevant background theories. In these cases it is probably most ac-
curate to reconstruct the derivation of "If C then 0" as follows: One first
assumes that the observable conditions C obtain. One then (jbtains by an
inductive inference, employing various auxiliary hypotheses together (per-
haps) with T, the conclusion that some theoretical description C' also char-
acterizes the initial conditions. From the theoretical characterization C',
together with auxiliary hypotheses and (perhaps) T, one then derives, de-
ductively or inductively, the conclusion that observable conditions satis-
fying 0 will subsequently obtain. Finally, one obtains by conditionaliza-
tion the conclusion "If C then 0". The crucial point is that - whatever else
may be involved - there will almost certainly be an inductive inference
depending on theoretical premises by which the theoretical import of the
initial conditions C is assessed.
There is one further consideration which indicates clearly the role of
inductive considerations in the derivation of predictions from scientific
theories. If we are to find useful a model according to which such predic-
tions are obtained deductively then it must surely be the case that the
"rationally reconstructed" deductive inferences attributed to scientists by
that model must employ premises and involve intermediate conclusions
which scientists themselves could (at least to a good first approximation)
make explicit. The less "explicitable" are the considerations which scien-
tists employ in deriving predictions from theories, the less fruitful it will
be to rationally reconstruct their derivations as deductive inferences. At
least for those sciences which study complex systems (biology, the social
sciences, the earth sciences, the engineering sciences) and, perhaps, for the
others as well, it seems likely that many of the predictions which scientists
reliably make from theories depend upon considerations which, at the time
in question, no one could make explicit even to a good first approximation.
Consider predictions of the form "If C then 0" derived from a theory
in such a science, where the observable conditions C are naturally occur-
ring rather than contrived by the experimenter. As we have seen, this pre-
diction is legitimately derived from T only if there are reasons to think
that the conditions C embody adequate controls for the relevant sorts of
experimental artifacts. The judgment that this is so depends, in turn, large-
lyon considerations reflected in previously established theoretical knowl-
edge. Scientists employ their understanding of existing theories in making
judgments of the sort in question. If theoretical knowledge and theoretical
understanding were an entirely propositional matter - if what a scientist
246 RICHARD N. BOYD
understands about nature in virtue of having understood a body of the-

ories could be (to a good first approximation) captured by the theoretical
propositions which she could make explicit - then the judgments in ques-
tion would be explicitable in the required way. But, as Kuhn (1970) and
many others have observed, the understanding of fundamental theories or
"paradigms" involves more than such propositional understanding. It in-
volves a "feel" for materials and for experimental techniques which often
goes beyond anything that the relevant scientific community can at the
time make explicit. This sort of understanding depends crucially on explicit
theories but not in an entirely prqpositional way. There is every reason to
believe that, especially in cases iniiwpich a "feel" for very complex systems
in involved, scientist's judgments regarding the adequacy of experimental
controls will depend in fundamental ways on just this sort of non-prop-
ositional (but theory-determined) understanding. In such cases, to employ
a model for theory testing (or theory application) which treats the inference
from theory to observational prediction as deductive would be to miss an
extremely important fact about the role of trained judgment and of (partly)
non-propositional understanding in the testing and application of scientific
theories.
I conclude that, all things considered, the empiricist conception that it
is the deductive (rather than the inductive) observational predictions of
scientific theories which playa role in theory testing or theory application
is fundamentally mistaken.
3.2 Must Empirical Tests of a Theory Involve Tests of Its Observational

Predictions?
We have seen that in the typical case in which a scientist tests a theory by
testing one of its observational predictions the prediction in question will
have been derived from the theory (together with auxiliary hypotheses) by
inferences which are genuinely inductive rather than deductive. I now want
to suggest that the empiricist conception that we test theories by testing
their deductive observational predictions is wrong in yet another way.
There is a wide class of important cases in which observations are crucial
to the testing of a proposed theory even though they do not test its pre-
dictions, deductive or inductive. That this is so follows from dialectical
features of the way evidence for theories is assessed. I summarize below
the relevant features (for fuller discussions see Boyd 1972, 1973, 1982,
1983, 1984a, 1984b).
As we have seen, proposed theories are candidates for confirmation only
if they are projectable, and projectability is a matter- of theoretical plau-
sibility given currently accepted background theories. Plausibility judg-
ments are relevant to the assessment of projectability because they rep-
resent inductive inferences from previously acquired (approximate) theo-
retical knowledge. Projectable theories are just those which are supported
by inductive inferences of this sort; they are thus supported by theory-
mediated evidential considerations which rest ultimately on the body of
experimental evidence which supports the currently accepted theories. Our
preference for projectable theories is a preference for theories for which
there is already some empirical evidence.
When we assess the extent to which a proposed theory is supported by
observational evidence, considerations of projectability enter into our
methodology once again. There is, roughly speaking, one basic rule for
the assessment of "degrees of confirmation": a theory is significantly sup-
ported by empirical evidence only to the extent that that evidence favors
the theory over other projectable rival theories about the same phenomena.
The rationale for this principle is simple: it is the projectable rivals of the
proposed theory which (supposing that it itself is projectable) constitute
the other theories about the phenomena for which there is already some
empirical evidence. Our experimental investigations are aimed at choosing
from among the theories we already have some good reason to believe
true, and our testing procedures involve pitting these theories against one
another until (if all goes well) one (or, sometimes, a synthesis of several)
emerges as clearly the most likely in the light of the evidence we have
examined. This methodological procedure is epistemically reliable because
(and to the extent that) our previously confirmed background theories are
accurate enough to provide a good (inductive) guide with respect to the
question which theories might be (approximately) true.
As we have seen, we may think of the body of previously confirmed
theories as establishing principles of scientifically rational inductive
reasoning which, sometimes, dictate conclusions which we must accept,
given that we accept a particular theory. These principles may dictate in-
ductive inferences from theoretical premises to observational conclusions
of the sort discussed in the previous section and, of course, they may
248 RICHARD N. BOYD
dictate such inferences from theoretical premises to theoretical conclusions

as well. It is important to recognize that the same theory-dependent in-
ductive principles of scientific rationality also dictate more complex con-
straints on the dialectical development of our theoretical beliefs. Thus
scientific rationality may dictate that if one accepts or continues to accept
a certain theory T, given observational data D, then one must make other
methodological decisions in consequence; for example one might be
obliged to accept one or more of some family of additional theories or to
revise one's pre-existing theoretical conceptions in relatively specific ways.
Existing theoretical knowledge often sets sharp constraints on the meth-
odologically acceptable responses to new data or to new theoretical con-
siderations or problems. It is just these constraints which limit our Du-
hemian option to save any theory whatsoever from refutation by poten-
tially embarrassing data by changing our other theories in appropriate
ways.
Finally, these same dialectical principles of inductive inference provide
that the explanatory power of a theory should count as evidence in its
favor. Explanatory power cannot be adequately analyzed according to a
Humean deductive-nomological model (see e.g., Hempel 1965b) according
to which the explanatory power of a theory is merely a reflection of its
predictive capacity. Instead the explanatory power of a theory is measured
by the extent to which the causal claims it makes can be inductively inte-
grated into previously established (non-Humean) accounts of causal
powers and mechanisms (Boyd 1984b), and theories may often explain
phenomena which they do not predict.
In the light of these considerations, it will be plain why it is inappro-
priate to conceive of the evidence for a theory as always being provided
by tests of its observational predictions. In the first place, a theory is po-
tentially confirmable only if it is projectable and what it means for a theory
to be projectable is that it is evidentially supported by the myriad obser-
vations which support the various background theories which establish its
inductive plausibility; almost none of these observations will be predictible
from the theory itself on any reasonable construal of prediction.
Secondly, since the confirmation of a theory is possible only to the extent
that there is evidence that it is preferable to its projectable rivals, obser-
vational evidence which tends to disconfirm one or more rival theories will
count for the confirmation of a given theory whether or not the relevant
observations are predictable from that theory. It is true, of course, that

from a theory the falsity of its rivals is in some sense "predictable" but it
is by no means true that a given theory will predict the particular obser-
vations which constitute observational evidence against its rivals. Finally,
evidential support can be provided by showing that a theory explains
methodologically relevant phenomena. In many cases, the relevant obser-
vations are not predictable from the theory but are rather the observations
which support the other well-confirmed theories with which the theory in
question has been appropriately inductively integrated.
It might seem that even if the observations which support a theory are
often not predictable from it, disconfirmatory observations must be those
which disconfirm a prediction of the theory in question. In the face of
apparently disconfirming evidence, defenders of a given theory have a lim-
ited "Duhemian" option of modifying others of their theoretical commit-
ments in order to save the theory. Refutation of it theory occurs when the
available evidence, together with the theory-dependent rational constraints
on the dialectical development of theoretical beliefs, limit the defenders of
the theory to options which are themselves inductively untenable given the
best available theoretical knowledge. It is of course true that the obser-
vational evidence which thus backs a theory into an inductively untenable
comer may consist largely of observations which refute its observational
predictions, but this need not be the case. Observations which appear to
confirm a rival theory may tend to back the defenders of a theory into an
untenable comer even when they do not refute predictions of the theory
in question, for example, as may observations which tend to confirm a
broad theoretical conception with which the theory in question cannot
with inductive plausibility be integrated. Arguably the Darwinian refuta-
tion of special creationism and the refutation of vitalism in biology con-
formed to these latter patterns. In any event, disconfirmatory evidence
need not always consist of observed counterexamples to the predictions of
the theory disconfirmed.
I conclude that, whether we take the predictions to be deductive or
inductive, neither confirmatory nor disconfirmatory tests of a theory need
involve tests of its predictions. What is striking is that, by examining data
that neither directly confirm nor directly disconfirm a theory's observa-
tional predictions we can nevertheless obtain evidence that bears on the
question of the adequacy of those predictions. It is this phenomenon, and
250 RICHARD N. BOYD
the inductive and dialectical reasoning at the theoretical level which un-
derlies it, which the consistent empiricist cannot explain.
3.3 Premature Formalization: Concluding Realistic Postscript
Critics of the ways in which logical empiricists "formalized" scientific the-

ories as deductively integrated formal systems for making observational
predictions, and of the "solutions" to philosophical problems suggested
by such formalizations, have often described empiricist approaches to the
philosophy of science as involving "premature formalization". Recent
realist work in the philosophy of science, together with the considerations
developed here, strongly suggest that this characterization is much too
charitable. The problem is not that the formalization was premature, but
rather that it rested on untenable epistemological premises.
In the first place, the ability of logical empiricists to treat so many issues
in the philosophy of science "formally" rested in large part upon their
acceptance of the 20th-century version of the Humean account of causa-
tion according to which causal statements and (most if not all) statements
about explanation are to be "rationally reconstructed" solely in terms of
the deductive (perhaps statistical) subsumption of natural phenomena
under laws. These conceptions of causation and explanation rest entirely
on the verificationist conception that knowledge of unobservable causal
powers and mechanisms is impossible, and they are therefore untenable in
the light of the recognition that instrumental knowledge in science is in
fact derivative from the dialectical development of theoretical knowledge
and theory-dependent methods (Boyd 1984b). Thus much of the "for-
malization" of scientific issues by logical positivists rested upon a funda-
mentally Inistaken conception of the epistemology of science and a de-
rivative Inistake about the analysis of causal statements.
The issues discussed in the present essay indicate two other important
respects in which the "formalization" of scientific theories as deductively
integrated predictive formal systems reflects a very serious Inisunderstand-
ing of the epistemology of science. We have seen in the preceding section
that such a conception of theories makes it impossible to recognize the
fundamental ways in which inductive and dialectical processes of reason-
ing at the theoretical level determine the rational standards for the confir-
mation and disconfirmation of theories. An even more serious defect in
the conception of theories as deductively integrated formal systems may

be that it draws our attention away from the role played by inexplicit and
nonpropositional knowledge in theoretical understanding and in the the-
ory-dependent methods of science. It is a remarkable fact that we can (and
every graduate student in a science does) acquire and apply an understand-
ing of natural facts which, although dependent upon explicit theory, goes
considerably beyond what we have (at the time) the conceptual resources
to make explicit. We need, in any event, a naturalistic conception of scien-
tific knowing (Boyd 1982, 1983) and it is probable that a naturalistic ac-
count of this remarkable capacity will be one of the most important con-
tributions of the epistemology of science to epistemology in general.
In any event, the problem with positivist "formalization" was not that
it was premature. The problem was not the premature intrusion of math-
ematical precision into the philosophy of science but rather the intrusion
of mathematical techniques based on (what turned out to be) an entirely
inadequate conception of the epistemology of science. If the mathemati-
zation itself was at fault, the fault probably lay in the way in which math-
ematical precision can lend an air of scientific cogency to doctrines which
do not in fact survive close scientific investigation.
BIBLIOGRAPHY
Ayer, A. J.: 1946, Language, Truth and Logic, Dover, New York.
Boyd, R.: 1971, Realism and Scientific Epistemology, unpublished manuscript.
Boyd, R.: 1972, 'Determinism, Laws and Predictability in Principle,' Philosophy of Science
39.
Boyd, R.: 1973, 'Realism, Underdetermination and A Causal Theory of Evidence,' Nolls 7,
1-12.
Boyd, R.: 1979, 'Metaphor and Theory Change', in Methaphor and Thought, edited by An-
drew Ortony, Cambridge University Press, Cambridge.
Boyd,R.: 1980, 'Materialism Without Reductionism: What Physicalism does Not Entail', in
Readings in Philosophy of Psychology, Volume 1, edited by Ned Block, Harvard University
Press, Cambridge.
Boyd, R.: 1982, 'Scientific Realism and Naturalistic Epistemology', PSA 80, vol. 2, Philos-
ophy of Science Association, East Lansing.
Boyd, R.: 1983 'On the Current Status of the Issue of Scientific Realism', Erkenntnis 19,
45-90.
Boyd, R.: 1984a, 'Lex Orandi est Lex Credendi', in Churchland and Hooker (eds.), Images
of Science: Scientific Realism Versus Constructive Empiricism, University of Chicago Press,
Chicago.
Boyd, R.: 1984b, 'Observations, Explanatory Power and Simplicity: Towards A Non-Hu-
mean Account', in P. Achinstein and O. Hannaway (eds.), Experiment and Observation in
Modem Science, Bradford Books/MIT Press, Cambridge.
252 RICHARD N. BOYD
Carnap, R.: 1934, The Unity of Science, tr. by M. Black, Kegan Paul, Trench, Trubner and
Co., London.
Carnap, R.: 1950, 'Empiricism, Semantics and Ontology', Revue Intern. de Phil. 3, 000-000.
Feigl, H.: 1965, 'Some Major Issues and Developments in the Philosophy of Science of
Logical Empiricism,' in H. Feigl and M. Scriven (eds.), The Foundations of Science and
the Concepts of Psychology and Psychoanalysis, Minnesota Studies in the Philosophy of
Science, Vol. I, University of Minnesota Press, Mineapolis.
Goodman, N.: 1973, Fact Fiction and Forecast, 3rd edition, Bobbs-Merrill Co., Indianapolis
and New York.
Hanson, N. R.: 1958, Patterns of Discovery, Cambridge University Press, Cambridge.
Harman, G.: 1965, 'The Inference to the Best Explanation', Philosophical Review 74,88-95.
Hempel, C. G.: 1954, 'A Logical Appraisal of Operationism,' Scientific Monthly 79, 215-20.
Hempel, C. G.: 1958, 'The Theoretician's Dilemma: A Study in the Logic of Theory Con-
struction,' in H. Feigl, M. Scriven and G. Maxwell (eds.), Concepts, Theories, and the Mind
Body Problem, Minnesota Studies in the Philosophy of Science, Volume II. The University
of Minnesota Press, Minneapolis.
Hempel, C. G.: 1965a, 'Empiricist Criteria of Cognitive Significance: Problems and Changes',
in C. G. Hempel, Aspects of Scientific Explanation, The Free Press, New York.
Hempel, C. G.: 1965b, 'Aspects of Scientific Explanation', in C. G. Hempel, Aspects of
Scientific Explanation, The Free Press, New York.
Kuhn,T.: 1970, The Structure of Scientific Revolutions, University of Chicago Press, Chicago.
Neurath, 0.: 1931a, 'Physicalism: The Philosophy of the Viennese Circle,' Monist.
Neurath, 0.: 1931b, 'Physicalismus,' Scientia.
Putnam, H.: 1975a, 'Explanation and Reference', in Putnam, Mind, Language and Reality,
Cambridge University Press, Cambridge.
Putnam, H.: 1975b, 'The Meaning of "Meaning",' in Putnam, Mind, Language and Reality,
Cambridge University Press, Cambridge.
Quine, W. V. 0.: 1969, 'Natural Kinds', in Quine, W. V. 0., Ontological Relativity and Other
Essays, Columbia University Press, New York.
van Fraassen, B.: 1980, The Scientific Image, The Clarendon Press, Oxford.
Cornell University
The Sage School of Philosophy
Goldwin Smith Hall
Ithaca, NY 14853
U.S.A.
B. KANITSCHEIDER
EXPLANATION IN PHYSICAL COSMOLOGY:

ESSAY IN HONOR OF C. G. HEMPEL'S
EIGHTIETH BIRTHDAY
INTRODUCTION
Scientific explanation is a current theme in logical analysis since C.G.

Hempel published his monumental work on this topic. 1 Many ramifica-
tions, but also doubts on the universal applicability of the Hempel-
Oppenheim-scheme of deductive nomological explanation have been as-
serted. In the following we shall not be engaged in one of these logically
sophisticated subtleties. Instead we will focus our attention on the special
problem whether it is possibl~ to apply Hempel's classical analysis of
causal explanation if the domain in question is the largest possible physical
system, the universe. Finally, we will be concerned with the result of the
customary analysis of scientific explanation, that there are no final and no
complete explanations, in opposition to the recent attempts of mathemat-
ical physics to build up so called 'complete cosmological theories'.
I. THEORETICAL COSMOLOGY
Modem physical cosmology can be regarded as a piece of applied theory

of gravitation. In the standard case, Einstein's theory of gravitation is used
because it is now very well corroborated and seems to solve gravity prob-
lems better than its competitors. Since 1917, when Einstein found the first
cosmological solution of his general relativistic field equations, the so
called cylinder universe, the method to construct cosmological solutions
pertinent to different kinds of boundary conditions has not changed in
principle. 2
The central question concerns the discovery of exact solutions of the
field equations which fit as a good approximation to the large scale prop-
erties of the physical universe. These global properties consist in the con-
tent of matter and its flow on the one side and the corresponding overall
structure of space time on the other. These two key properties of the uni-
Erkenntnis 22 (1985) 253-263. 0165-0106/85/0220-0253 $01.10

254 B. KANITSCHEIDER
verse are, according to Einstein's theory of gravitation, not independent,

but intimately connected. Cosmology cannot be started in a purely em-
pirical fashion - we can not penetrate in a cumulative way more and more
the depths of space. Instead we have to anticipate some prior knowledge
on the global structure of space time, which might be validated subse-
quently. This does not mean that there is an absolutely uncontrollable
component in cosmology which can only be asserted in an a priori manner,
as some philosophers of science have claimed 3 , but that we have to start
with some element of prior knowledge that has to be vindicated afterwards
by independent methods. 4 As the reader might have guessed, it is the much
debated Copernican principle that we have in mind. Since Hermann Bon-
di's formulation of the democratic point of view concerning the large scale
structure of spacetime - in his own wording 'the earth is not in a central,
specially favoured position's - the Copernican principle is used as the afore
mentioned kind of prior selection principle to eliminate spatially non
homogeneous solutions of the field equations of gravitation. Bondi's for-
mulation states, that the position of a terrestrial observer is not in any way
specially distinguished, but it leaves open, whether there are different, per-
haps very remote, privileged locations in space. Stronger versions of the
democratic world view of spacetime are pointed out by Whittaker: 'It is
impossible to tell where one is in the universe'. 6 Perhaps the clearest way
to formulate the point in question is Harrison's location principle: 'It is
unlikely that we have special location in the universe'. 7 With this admix-
ture of non directly observable knowledge we can reduce the generally
quite complicated field equations to very simple differential equations.
They describe spacetimes whose space-like three-surfaces have the intrin-
sic property that all points on these surfaces are physically equivalent. This
special quality of the so called Friedman-Robertson-Walker spacetimes
can not be observed from our special point of view. Thus, homogeneity
is only partially corroborated. Empirically tested is the isotropy of the
extragalactic observations (e.g. X-ray, micro-wave background and the
distribution of galaxies beyond 100 Mpc.) If we use one of the principles
expressing that there are no distinguished locations in the universe and
that it is approximately (that means disregarding local irregularities) spher-
ically symmetrical about any point, then one can see immediately that all
the asymptotically flat spaces like the Schwarzschild-, Reissner-, Nord-
strl2lm- or Kerr-solutions cannot be applied within cosmology, because
EXPLANA TION IN PHYSICAL COSMOLOGY 255
their spherical symmetry shows up only when observed from special lo-
cations.
These are the reasons why in cosmology only global solutions of the
Friedman-type are used which possess the Robertson-Walker-metric with
constant curvature and a universal cosmic time parameter. Now, the high
symmetry of the Robertson-Walker spacetimes requires a special ideali-
zation concerning the content of matter. The conceptual place ofthe ideal-
izing assumptions for cosmic matter is called the model object. 'The model
object is hoped to constitute a profile of its referent'. 8 Seen from a
'bird's-eye view', we can treat galaxies as 'molecules' of a gas. In reality,
the 'molecules' have an internal structure but in this context these details
are omitted; the stellar character, the assemblage of the particles in the
cores of galaxies and the star associations, all these properties of condensed
matter are ignored, in the same way as classical mechanics treats whole
planets as mass points, ignoring the bulky properties with the only excep-
tion of inertial mass. Even the particular structure of the representational
matter can be disregarded in order to treat the cosmic stuff as a perfect
fluid; this model-object now bears three characteristics only, the 4-velocity
ua., the time dependent density of mass-energy p(t) and the pressure p(t)
which both depend, as indicated, on cosmic time. Any other possible fea-
ture of matter like shear, rotation or anisotropy-pressure is ignored in this
idealized model. But this approximation, very useful until today is not
beyond doubt for all time. It is a fallible assumption, and if the model of
a perfect fluid should prove to be too simple, it can be enriched with more
detailed pecularities of the cosmic matter.
Just to mention one case: if a Foucault-pendulum would reveal a meas-
urable relative rotation between the distant masses and the plane in which
the pendulum is swinging, we would be forced to enlarge the model-object
with a term describing absolute rotation. The dynamics of the universe
would not be any longer ruled by the simple Friedman equation. In the
same moment, when rotation is included in the stress-energy tensor, we
had to renounce with the property of cosmic time. There would not be any
longer a unique history of the cosmic evolution and surely the cosmolog-
ical description would be far more difficult than it is today. It is a nomo-
logically contingent feature of the special kind of universe we live in, that
it allows such a simple description, thereby corroborating an old philo-
sophical prepossession, namely the uniformity of nature.
II. DESCRIPTIVE COSMOLOGY
Theoretical cosmology as sketched above cannot be brought into contact

with the pertinent astrophysical data without further specifications. A
bridge has to be constructed between the theoretical Friedman equations
with their unobservable variables like curvature, density and pressure and
the astrophysically observable magnitudes like the Hubble and the dece-
leration parameter. This translation of the dynamic basic equation into
operational terms is by no means a trivial task. The sentences bridging the
theoretical and the empirical level are not definitions in a strict sense, but
semantical relations which state the kind of references of the theory in
empirical terms. If these empirical indicators of the theoretical parameters
are found which make the whole theory testable, then the descriptive aim
of cosmology can be faced. It consists in the seemingly very simple task of
finding two numbers H o, qo for the two empirical parameters H(t} and
q(t}. If the cosmological constant A.=O, such a pair H o, qo marks out a
concrete case in the class of uniform FRW models. If we knew the exact
numbers of this pair's today values, we could truly assert that the descrip-
tive problem of knowing the content of matter and the space time structure
of the entire universe would have been solved.
III. EXPLANATORY COSMOLOGY
Since the rise of scientific cosmology there was still lingering an argument
that draws its power from the apparently indubitable fact that the universe
is a unique system. It has been mostly taken in the way, that what we mean
by speaking of the universe is the most inclusive or complete physical
whole. This uniqueness of the universe, an undisputed logical fact until
recently, has led many authors of the philosophy of science to separate the
methodology of cosmology from the customary logic of science. The point
of difference concerns the existence of laws and theories: 'The phenomena
studied in ordinary branches of physics are of such a kind that a distinction
can readily be made between the common features to be found among a
variety of instances of some particular phenomenon, that is, what is ex-
pressed by the observationally established law of such a phenomenon as
contrasted with the differentiating individual characteristics that distin-
guish one instance or occurrence of the phenomenon from another.'9
To what these philosophers of science point is a kind of degeneration of

the concept of law which seems to occur, if there is only one instance to
which the law applies, in the case of cosmology if there is no plurality of
actual worlds. In one sense it is quite correct to focus on the difference
between a law which governs a set of galaxies, and points to a common
pattern in the ensemble like the spiral structure on the one side, and a
special lawlike feature of the universe as a whole like spatial compactness
on the other. But this conceptual difference does not mean that the char-
acteristics of the universe cannot be connected in a lawful way. Surely, a
global feature like a metrical property (e.g. curvature) or a topological one
(e.g. spatial closure) does not derive from repeatability, from an invariant
pattern in multiple cases (e.g. universes). But there is an intrauniversal
nomological connection between the cosmological key properties (e.g.
mass~ergy, density and the structure of space-time) in a way that sep-
arates a pure cosmographical from a cosmological approach.
Explanation surely depends on the existence of lawful patterns in the
pertinent domain of reality. They make sure that law formulas can be
found which interconnect the intrinsic magnitudes. Therefore the possi-
bility of explanatory cosmology does not depend on a plurality of worlds,
but on the logical connexity of the properties of the model applied. The
approximative character of the explanation is expressed by the fallible
status of the model. If an explanation proves to be unsuccessful, the next
step has to be to enlarge the model by a richer set of physical variables,
and if this strategy seems to be exhausted, the doubt will come down to
the theory itself. If necessary, an alternative theory of gravitation (e.g.
variable G-Theory, Scalar-Tensor Theory, Bimetric Theory) has to be
used with more physical degrees of freedom. Cosmologists themselves,
right from the beginning of the theoretical and empirical investigations in
the twentieth century, had no methodological qualms about applying the
concept of a causal explanation of cosmology. George Lemaitre concluded
his famous paper with the suggestive words: 'II resterait Ii se rendre compte
de la cause de l'expansion de l'univers'.lo At that time, almost nobody
dared to care about the question, why the universe is expanding, each
astronomer would have been deeply content, if he knew, how this gigantic
phenomenon could be described. Lemaitre uttered the idea, that perhaps
the work done by the pressure of the cosmic substratum might have some-
thing to do with the cause of the expansion. But as Eddington later
showed, this special proposal could not be taken seriously because there
could be expansion or contraction even in the case of zero pressure. Le-
maitre's model reduces to the static Einstein universe, if the pressure is set
to zero. The resulting s<K:alled cylinder universe is a world in equilibrium,
but this state is unstable. The slightest disturbance would induce suddenly
an expanding or contracting motion of cosmic matter. Although today
there is no empirical use of the static model - because the redshift of the
light of distant galaxies is almost unanimously interpreted as the result of
their receding motion - it sheds some interesting light on the type of pos-
sible explanatory answer to the question, why our universe is expanding
at all. Since we know that Einstein's cosmological solution of 1917 is the
only static one - de Sitter's solution of the same year is only static, if some
very special coordinates are used - and since this equilibrium state is un-
stable, the universe has to be necessarily dynamic if Einstein's theory of
gravitation is correct.
Here we see which type of explanatory answer is possible in cosmology:
A global trait of the universe can be understood by the combination of
spacetimes and matter constellations which is permitted by the local laws
of gravitation. Surely, the local field equations have to be tested by inde-
pendent observations which concern other solutions, in our case the central
symmetric spacetime of Schwarzschild. But this procedure is surely an
instance of the quite customary deductive nomological approach.
General relativity allows many solutions which lie, in accord with the
cosmological principle, within the scope of the permissible class of cos-
mological solutions depending on the values of parameters like A., p and
p. A few theoreticians, like Edward Milne, regarded the existence of these
alternatives as a big stumbling-block to a real solution of the cosmological
problem. He not only stressed the need for explanations in cosmology
altogether, but even claimed the uniqueness of the solution within one and
the same theory.
'There must be a unique solution to the question 'why?', applied to the
universe in respect of each feature'. 11 His approach did not permit any
'bifurcation of possibilities' which could be decided, as in relativistic
cosmology, only by empirical means, thereby renouncing of an answer to
the question, why the universe has this contingent trait. In his approach,
no alternatives are allowed, Milne's demand for a rigid theory of cosmo-
logy reminds of Geoffrey Chew's bootstrap principle that states: 'Nature
EXPLANATION IN PHYSICAL COSMOLOGY 259
is as it is, because this is the only possible nature consistent with itself.u
According to this holistic principle the different features of the universe
are so deeply concatenated that each property can be explained only by
means of all other qualities. Today, cosmological bootstrapping is at most
a research program, but it is possible to understand, how an explanation
founding only on the doctrine of internal relation works, if we think of
the principle of Mach. The inertia of any body is fully determined by the
distribution of the masses of all other bodies in the universe. If an explan-
atory question is asked: Why do terrestrial bodies resist accelerations in
the way they do, then the answer is, because the far-away masses cause
the local behavior of these objects.
There are, however, very few properties which can be explained by
means of internal relations and even the Mach principle is, in its episte-
mological status, not without doubt, but it strongly suggests that there are
possible explanations in cosmology, even if no allowance is made for a
whole set of universes.
IV. ANTHROPIC EXPLANATIONS
It is true, until the beginning seventies explanatory why-questions were

the exception rather than the rule. Astronomers had been so deeply in-
volved in getting some idea which of the FRW worlds are satisfactory
descriptions, that they could not approach the problem why just these
highly symmetrical models are successful. This much more fundamental
problem appeared on the scene only recently: 'Accepting the agreement
with observations we want to understand, why the laws of physics should
demand (rather than permit) a universe that is homogeneous and isotropic
to high accuracy on large scales'. 13 To answer this question, the customary
procedure is to start the universe, in an experiment of thought, with an
arbitrarily non-uniform set of initial conditions and to analyse its evolu-
tion up to the present day. If today's regular state of the world is guar-
anteed solely by the dynamic laws, in a manner totally independent of the
initial conditions, then we possess in a sense an explanation of the observed
symmetry of the universe, because any world, whatever its early state,
would be regular at very late times. But the investigation of the asymptotic
stability of the FRW worlds under the action of homogeneous anisotropic
pertubations delivered not this, but an entirely different result. As Collins
and Hawking could show in the case of the flat and hyperbolic models
(zero or positive binding energy), the property of isotropy is unstable, that
means that all open universes will become increasingly anisotropic at large
times 14, except the Euclidean model, which is extremely unlikely in the set
of all models.
This led to an alternative situation in the wording of Barrow and Tipler
who refined the analysis which 'tells us that either the Universe is young
~ instabilities having had insufficient time to grow remarkably as yet - or
that the universe is in a highly preferred, zero measure, binding--energy
state'.iS If we choose the second hom of the dilemma, a new explanatory
principle comes to the fore: the anthropic principle. The relative unlikeli-
hood is reduced, if an invocation of a large ensemble of worlds is made in
the way that only in a subset of these worlds with special advantageous
initial data sets galaxies and life supporting environments with heavy ele-
ments, necessary for biological evolution, could have arisen. Thus, if we
resort to man as an intelligent observer and regard his necessary physical
presuppositions, we are led to a universe with a very finely tuned expansion
rate. If it had been just a little bit too slow, the time scale would have been
too short, a too early recollapsing of the universe would have prevented
the start of the biological evolution. If the speed of the expansion had been
somewhat too fast, there could have been no density enhancement, the
statistical fluctuations could not have grown into condensed objects. An-
thropic explanations in cosmology can thus be paraphrased as follows:
The universe is as it is, because we are here. Of course, if it were otherwise,
there would not be anyone to ask this question, there would be no intel-
ligent life, the universe, perhaps filled only with gravitational radiation,
would not belong to the cognizable subset of all physically possible worlds.
Logically the next question to ask is:
Are anthropic explanations genuine explanations in the older sense? The
answer is straight-forward: Surely, they are not. The sentence: 'the isotro-
py of the universe is a consequence of our existence' cannot be interpreted
as a causal, deductive nomological explanation in accord with the expli-
cation of C.G. Hempel and Paul Oppenheim. The existence of intelligent
beings at late times cannot serve as explanans, wherefrom with the help
of physical laws, the isotropy might be deduced. Man is not the cause of
the symmetry of the universe. A regular and highly ordered universe is the
necessary, but surely not sufficient condition for the evolution of intelligent
organisms. Anthropic connections - not being themselves explanations

proper - can instead give hints to the construction of genuine explanations.
The informative content of anthropic connections is not zero. If there is
a coupling of certain invariant structures of the universe with the precon-
ditions of the existence of observers, then we should seek for new dynamic
evolutionary explanations. In this sense anthropic explanations are blanks
for dynamic explanations.
A very good example for such a situation is the above mentioned prob-
lem of the isotropy and, connected with it, the flatness of the universe.
There is, indeed, a mechanism available which makes it comprehensible,
how the conditions k=O or P = Pc or Q = 1 came about in a causal way.
The s~alled 'inflationary scenario', proposed by Alan Guth within the
grand unified theory (GUT), made it intelligible, why the universe finds
itself in such a preferred state of binding~nergy, without using the exist-
ence of observers. 16 According to the GUT, there has been an unbroken
symmetry of the three forces (strong, electromagnetic, and weak) at the
very early time of t = 1O- 3s sec., when the universe was very hot (T =
l0 28 K). In the following time, spontaneous breaking of the grand unifi-
cation symmetry and with it first-order phase transitions occurred. If we
assume that the universe had a prolonged period of exponential expansion
which is caused by the vacuum energy of a supercooled metastable state
of the phase transition, then such an epoch of fast de Sitter expansion can
explain the above mentioned flatness. Moreover, the inflationary scenario
supplies the solution to a number of outstanding problems in cosmology.
At first, it can be proved that at the end of the de Sitter phase the expansion
is almost automatically driven near to the critical rate, above which re-
collapse is unavoidable. Furthermore, the monopole-puzzle and the riddle
of uniformity come out as surplus-meaning without additional assump-
tions. 17
Here we can see that the dynamic explanation, which purveys an intrin-
sic mechanism to make today's flatness of the universe plausible, is ad-
vantageous in a twofold way: it is epistemologically· deeper and ontolog-
ically much more parsimonious. There is no need for the prodigality of the
many unknown and unknowable worlds and we can understand the
energy-state of the universe, one of its central properties, in a causal way.
The last example leads straight-forward to the limits of explanation in
cosmology at all. Ifmore and more key properties of the universe, coupling
constants, adjustable parameters, are deductively explained, the question

stands up how far this procedure can be driven. Theoretical physicists are
rather optimistical in this regard. Says for instance Stephen Hawking: 'I
think that the initial conditions of the universe are as suitable a subject for
scientific study and theory, as are the local physical laws' .18 If we could
subscribe to such a research program, the non-uniqueness of the initial
conditions and the arbitrariness of the local physical laws would only mir-
ror our present imperfect state of knowledge. Such a prospect of future
scientific activity does not necessarily clash with the famous theorem of
the theory of scientific explanation, namely that there are no final and no
complete explanations. 19 Even if there would be a mathematical proof that
our current unified theory of all interactions (perhaps supergravity) is com-
patible with only one set of initial conditions, it would still be open,
whether this temporarily 'final' formula could not be superseded by a fur-
ther 'final' formula, which provides an explanation still deeper. So science
will always be a matter of unending quest.
NOTES
1 C.G. Hempel: 1965, Aspects of Scientific Explanation, Macmillan, New York.

2 Cf. B. Kanitscheider: 1984, Cosmology. Theory and History in Philosophical Perspective,
Ph. Reclam, Stuttgart (in German).
3 K. Hubner: 1978, Kritik der wissenschaftlichen Vemunft, K. Alber, Freiburg, p. 265.
4 For an empirical evaluation of the global invariance of natural laws cf. A.D. Tubbs and
A.M. Wolfe: 1980, 'Evidence for the Large Uniformity of Physical Laws', Astr. Phys. Joum.
236 LIo5-LIo8
5 H. Bondi: 1969, Cosmology, Cambridge Univ. Press, Cambridge, p. 22.
6 E. Whittaker: 1958, From Euclid to Eddington, Dover, New York.
7 E. Harrison: 1981, Cosmology, Cambridge Univ. Press, Cambridge, p. 70.
8 M. Bunge: 1973, Method, Model and Matter, D. Reidel, Dordrecht, p. 95.
9 M.K. Munitz: 1962, 'The Logic of Cosmology', Brit. Journ. Phil. Sci. 13,36.
10 G. Lemaitre: 1927, 'Un univers homogene de masse constante et de rayon croissant,
rendant compte de la vitesse radiale des nebuleuses extragalactiques', Ann. de la Soc. Sci. de
Bruxelles 47A, 59.
11 E.A. Milne: 1952, Modern Cosmology and the Christian Idea of God, Clarendon Press,
Oxford, p. 49.
12 G.F. Chew: 1968, ' "Bootstrap": A Scientific Idea?' Science 161, 762.
13 Ch. Misner et al.: 1973, Gravitation, W. H. Freeman, San Francisco, p. 800.
14 B. Collins and S. Hawking: 1973, 'Why Is the Universe Isotropic?', Astrophys. Journ. 180,
317-334.
15 J.D. Barrow and F.J. Tipler: 1978, 'Eternity Is Unstable', Nature 276, 453.
16 A.H. Guth: 1981, 'Inflationary Universe: A Possible Solution to the Horizon and Flatness
Problems', Phys. Rev. D 23,2,347-356.
EXPLANATION IN PHYSICAL COSMOLOGY 263
17 S.W. Hawking and E.G. Moss: 'Supercooled Phase Transitions in the Very Early Uni-
verse', Phys. Lett. 110, Nr. I, 35.
18 S. Hawking: 1971, 'Is the End in Sight in Theoretical Physics', Cambridge, 1983, p. 3.
19 W. Stegmiiller: 1983, 'Problerne und Resultate der Wissenschaftstheorie und Analytischen
Philosophie. Band I: Erkliirung, Begrundung, Kausalitiit, 2. Aufi. Berlin, S. 150.
Zentrum fUr Philosophie und Grundlagen der Wissenschaft der lustus-Liebig-Universitiit

Giessen,
Otto Behagelstr. IOc, II O.G.,
6300 Giessen,
F.R.G.
NELSON GOODMAN
STATEMENTS AND PICTURES·
Although languages are obviously theory-laden, a language is on the face

of it quite different from a theory stated in the language. A language makes
no claims, is neutral as between opposing statements, accommodates var-
ied and conflicting theories. A theory makes claims that can be denied in
the same language and restated in other languages.
Thus Richard Rudner's argument to show that languages cannot be
distinguished from theories startles us; 1 but it cannot be lightly dismissed.
To understand a language is not merely to know what its terms are and
how to put them together properly, but to know also what its terms mean
or at least apply to. But if we know what "man" applies to and what
"mortal" applies to, we already know that "all men are mortal" is true.
The truth of the statement and the application of its terms are interdepen-
dent. To determine either is to determine the other and thus, the argument
runs, languages and theories cannot be distinguished from each other.
Yet plainly in some sense we may know a language - as I know English
- and still be in doubt about the truth of many statements in that language.
Indeed, only if I in some sense understand a language can I set out to
determine whether certain statements in it are true. How can we recognize
such obvious facts without reviving the spurious distinction, long ago dis-
credited by Quine, between the analytic and synthetic, between internal
and external questions? Our trouble here, I think, lies in the nature of
knowing a language. Knowing all about a language would include know-
ing the truth-value of every statement in it; but no one knows any language
that way. As a matter of fact, knowing all about anyone thing would
amount to knowing all about everything; to know all about the fallen
sparrow in Brooklyn is to know it fell just so many years, months, weeks,
days, hours, minutes, seconds after a certain Chinese mandarin drank his
ninety-second cup of tea in Soochow. But in any ordinary sense, knowing
is partial, and by knowing a term we mean something much less than
knowing everything that it applies to. I understand the term "green" in
that I know of some things that they are green, and of others that the are
Erkenntnis 22 (1985) 265-269. o165-{)106/85/0220-0265 $00.50
266 NELSON GOODMAN
not green, even though I do not know of still other things (because I have
not seen them, nobody has told me, I cannot make the inference from
other available information, or because they are borderline cases) whether
they are green or not. I may even know a term without knowing anything
it applies to if I know some of its extensional relationships to other terms
I know. And I know a language ifI know thus partially some of its terms.
Knowing a language in this way does not imply knowing the truth-value
of all statements in it; and so, quite within the bounds of extensionality,
languages are distinguished from statements and theories.
I want, then, to keep languages and symbol systems distinct from terms
and statements and other verbal and nonverbal symbols that make up
these languages and systems. Likewise, I think of works of art not, as
Rudner sometimes does, as symbol systems, but rather as symbols within
systems.
Rudner's main concern, however, is with certain relationships, especially
the relationship of contrast or conflict or incompatibility, between theories
or more generally between versions, that may be in the same language or
in different languages or in symbol systems that are not languages at all.
This is an important and difficult matter, much in the forefront of my own
current research.
Recently we have heard a good deal about theories that, rather than
being logically incompatible with each other, are 'incommensurable', so
that logical relationships do not obtain between them or are at least in-
determinable. "Incommensurable" has always seemed to me a rather in-
appropriate term; for incommensurable quantities are not incomparable.
The radius and circumference of a circle are incommensurable but their
lengths in terms of a common unit can be compared within any desired
degree of precision. I take it that the claimed disparity between theories,
even in the same language, is wider than that. However that may be, when
we take into account not only different theories, whether or not in the
same language, but also versions and visions in symbol systems that are
not languages and that admit of no statements or truth-values, we face
disparities of quite a different order.
Rudner takes his cue from my suggestion that while a musical score is
best construed as a denoting character like
" - - followed by - - followed by - - " etc.
STATEMENTS AND PICTURES 267
where the blanks are filled by descriptions of notes or chords, we might

alternatively interpret it as a statement like
"In John Doe's Opus 83 - - is followed by - - , which is followed by - - " etc.
So taken, the score has truth-value and stands to other scores in the usual
sentential relationships of compatibility, conflict, consequence. Rudner
proposes following the same course for all denotational symbols, and this
covers a good deal of territory including depictions. A picture of Babe
Ruth hitting a homer is taken as saying in effect that Babe Ruth hits a
homer; so that such pictures are also treated as statements. This leaves
nondenotational symbols - for instance, abstract paintings and music.
Rudner points out that exemplification and expression, like denotation,
are forms of reference, and suggests treating such works also as symbols
that make statements about what is referred to. 2
This is a considered and constructive proposal, worked out in some
detail, but it raises a number of questions. In the first place, while the
sentential correlate of a score seems more or less uniquely indicated, that
is less clear even for a verbal description. We readily tum "Babe Ruth
hitting a homer" into "Babe Ruth hits a homer"; but what do we do about
"the wrecked, yellow, 200 h.p. John Doe automobile"? A number of can-
didates, such as "John Doe's yellow 200 h.p. automobile is wrecked" and
"the wrecked yellow 200 h.p. automobile is John Doe's", seem to have
equal claims. We are even worse off when we come to pictures. What does
a picture of the Black Forest say? The trouble here is not merely that as
with the verbal description there are alternative sentential correlates but
that the picture, being in a syntactically undifferentiated symbol system,
does not resolve into separate word-like elements that make up sentence-
like sequences. Any correlation of picture with statement is thus much
more remote and arbitrary.
Rudner does not overlook such questions or dismiss them lightly. His
answer, as I understand it, is that each symbol or version in effect says
that it refers to what it does in fact refer to. The automobile description
thus says in effect "I (the string of words) denote the wrecked, yellow, 200
h.p. John Doe automobile"; the painting says "I depict the Black Forest";
and a spritely rondo says "I express spriteliness". Presumably, the auto-
mobile description would be true or right if and only if there is such an
automobile. But what could constitute a false or wrong picture of the
268 NELSON GOODMAN
Black Forest? Under this criterion, whatever depicts the Black Forest
counts as a true or right picture of it; for all the picture claims is that it
does depict the Black Forest. A picture that does not depict the Black
Forest, does not claim to do so, and thus cannot be counted as a false or
wrong picture of the Black Forest. To allow for wrong pictures, we would
have to interpret a picture of the Black Forest as saying something more
than that it depicts the Black Forest. But then we must ask "What more?';
we no longer have a clear general principle for unique correlation of a
statement with a picture.
These remarks by no means amount to comprehensive or conclusive
criticism of Rudner's interesting work-in-progress, but they may suggest
some of the reasons why, in approaching the same cluster of problems
from much the same base, I have taken almost the opposite course (Ways
of Worldmaking, VII). Rather than assimilating scores, descriptions, and
pictures to statements, I treat them all as nondeclarative: a score, descrip-
tion, picture, diagram, etc. may refer in various ways, or be a referential
symbol with empty reference, but does not state. This avoids the difficulties
outline above; but since nondeclarative versions have no truthvalue, we
must start from scratch in investigating the nature and standards of right-
ness and wrongness of such versions. Instead of appealing to truth, we
must seek a more general notion of rightness that may sometimes subsume
and sometimes compete with truth. That's tough.
Although, strictly speaking, pictures are not and do not make statements,
many pictures - especially but not exclusively serial and moving pictures
- do "tell stories", and telling stories surely seems close to making state-
ments. If a picture tells a story, why isn't it true or false?
An answer often given is that the picture and a text may tell the same
story, and that the picture is true or false only indirectly, only in that the
text is true or false. This is reasonable enough so long as we do not infer
that "telling the same story" implies that there is something called the
story that is not itself a version but is embodied in various versions. For
two texts to tell the same story is for them to be intertranslatable; and
translation is well-known to be nonunique and highly variable. What must
be preserved in translation depends on context and purpose; never is there
total preservation. But what is involved in translation between a text,
which is and makes a statement, and a picture, which does not?
STATEMENTS AND PICTURES 269
A picture, like a predicate, may denote certain events (or in the case Of
fiction, be a soandso-Iabel). When the predicates in a text denote those
same events 3 (or are also soandso-Iabels), the picture and the text are to
that extent intertranslatable; and the picture, though it makes no state-
ment, might be derivatively called true or false according as the text is.
But we must not forget that, strictly speaking, calling a picture true or
false is false.
NOTES
• This paper is an expanded version of Erkenntnis 12 (1978),176-179. The expanded version

has also appeared in N. Goodman, Of Mind and Other Matters, Harvard University Press,
Cambridge, Mass., 1984.
1 In 'Show or Tell: Incoherence among Symbol Systems', Erkenntnis 12 (1978), 129-151.
2 The reader of Rudner's paper must, by the way, watch for departures from the terminology
of Languages of Art. For instance, I have used 'referent' as a general term for anything
referred to, whether denoted, exemplified, or expressed, reserving 'compliant' for what is
denoted. Rudner uses 'compliant' as I use 'referent'.
3 Not that there is any unique criterion for considering two versions, verbal or pictorial, to
denote the same events.
Received 2 Apri11984
Dept. of Philosophy,
Harvard University,
Emerson Hall
Cambridge, MA 02138
U.S.A.
RAIMO TUOMELA
TRUTH AND BEST EXPLANATION
I. TRUTH AND EXPLANATION IN THE CONTEXT OF

SCIENTIFIC GROWTH
1. Nowadays it is not uncommon to find claims to the effect that the notion
of truth is ultimately an epistemic notion. Truth is accordingly said to
depend on one's viewpoint and background knowledge (or something like
that). I think that this is basically a correct idea. However, there remains
much to be done in clarifying what exactly is involved in the claimed ep-
istemic nature of truth. I have recently investigated this matter in the con-
text of developing and defending a version of epistemic (or 'internal') scien-
tific realism (see Tuomela, 1984). What I shall try to do below is somewhat
different. For I will take it for granted that truth indeed is epistemic and
then go on to investigate the relation between truth and (best) explanation.
It will not be my aim to claim that somehow (epistemic) truth and best
explanation are identical notions. But I do defend the claim that it is
reasonable to think that true theories and best-explaining theories co-
incide.
As I will be dealing with very broad topics below it is not always possible
to go deeply into details. Furthermore, I will have to rely on some of my
previous publications on these matters. But I will make an effort to sum-
marize briefly some of my previously defined key notions so as to make
the present paper tolerably self-contained. What will be new in the
present paper is a clarification of the notion of best explanation as well as
some detailed technical results concerning truth and explanation in certain
formal first-order languages.
2. In what follows I shall examine on a general level growth of knowledge
in terms of increasingly better explanations, having in mind the scientific
realist's point of view in particular. One central assumption in this point
of view is that sciences aim at giving evermore exact and truthlike pictures
of the world. According to the internal realism I espouse (see Tuomela,
1984), there is no conceptfree knowledge of the world; thus all inquiries
Erkenntnis 22 (1985) 271-299. 0165-0106/85/0220--0271 $02.90
272 RAIMO TUOMELA
into truth and explainability are to be tied to one conceptual scheme or

another. Truth is an epistemic notion, one which is relativized to a con-
ceptual scheme. We can accordingly claim, relative to a conceptual scheme,
that the best-explaining theory is a true theory, and even conversely. We
can thus claim that the following statements about a scientific theory T
are, vis-a-vis its conceptual scheme or language, equivalent (though per-
haps not merely for conceptual reasons):
(I) (a) T is (factually) true;
(b) T is epistemically true (or ideal);
(c) T is the best-explaining theory (at least asymptotically);
Of the theses in (1), (a) represents an unanalyzed, presystematic notion of
truth which is at least to some degree a correspondence notion (cf. the
classical saying Veritas est adequatio rei et intellectus). The factual ade-
quacy of a theory is understood to be some kind of adequate correspon-
dence between the theory and the portion of the world represented by it.
As I have argued, following Sellars (1963), (1968), such adequacy is ad-
equate picturing, and a theory cannot (at least ideally and asymptotically)
be factually true unless it pictures (the portion of) the world correctly (cf.
Tuomela, 1984b, Chapter 6, Section II). On the other hand I claim, and
show, in the mentioned work that the concept of truth is epistemic in that
it necessarily requires background knowledge. The equivalence of theses
(a) and (b) is in its essentials based on this fact. We can say that (a) and
(b) can be explicated in such a manner that they tum out to be equivalent,
partly on conceptual grounds, partly on the basis of general rationality
assumptions (cf. (17) and (22) below).
Let me lay special emphasis on the fact that the equivalence of (c) with
the other above-mentioned claims can only hold if T is a scientific theory.
Later I shall argue that (c) indeed can be related to (b) in the required
way. Asymptoticity in claim (a) refers to a situation in which the world
has in a sense been investigated exhaustively within the bounds of the
conceptual scheme employed.
Let us next have a closer look at the growth of knowledge encapsulated
by scientific theories. We consider a community, say A, of researchers who
act together with the general goal of finding out what the world is like.
Let the action, say U, performed by the community A be that of replacing
a predecessor theory Ti by a successor theory Ti + 1. How can such a social,
TRUTH AND BEST EXPLANATION 273
multi-agent action be justified? Assume that these agents first act within
the bounds of the theory T/ (e.g. Newtonian mechanics) in something like
the normal-scientific way described by Kuhn. The application of the theory
to some phenomena, say E1o ... ,E", proves to be impossible, however. We
shall then say that El)¢¢¢)E" are anomalies for our theory.
The scientific community A then produces a new theory candidate T j + 1
(e.g. Einstein's theory) which is able to explain those anomalies and which
'works' also in other respects at least as well as T j • We assume, in other
words, that T j + 1 explains 'correctively' both the success of T j and its an-
omalies. Accordingly, the rational community A forms the 'we-intention'
(group intention) to perform U, that is, to replace T/ by T j + h and executes
the intention. (Generally, a community A can be said to have such a we-
intention if all its members have it or if either most of its members have
it or at least its leaders have it; see Tuomela, 1984a, for the notion of
we-intention.) The execution of the we-intention takes place through the
performance by the members of A of their own component actions, which
together bring about U.
Next, I shall present a schema of practical inference which deals with
the transference of group intentions and especially their post factum jus-
tification in the situation just characterized (Rosenberg, 1980, p. 178, pre-
sents a rather smilar schema). We shall assume, then, that each member
of A can reason as follows, given that we allow A to survive changes in its
members, and given that 'we' here refers to a community with such an
identity:
(2) (i) We will have a theory which affords best explanatory ac-
commodation for those R-experiences, structured in terms of
it, with which we find ourselves.
(ii) We believe that our adopting theories which qualify as
successors whenever we find ourselves with R-experiences
anomalous vis-a-vis our extant (predecessor) theory facilitates
our having such best explanatory accommodation.
(iii) We will adopt a qualified successor theory whenever we
find ourselves with predecessor-anomalous R-experiences.
(iv) We believe that E 1o ... ,E" are R-experiences with which we
find ourselves but which cannot be afforded best explanatory
accommodation within our extant communally accepted
theory T j •
274 RAIMO TUOMELA
(v) We believe that T j + 1 (and it alone) qualifies as successor

to the predecessor theory T j with respect to Et. ... ,ElI •
(vi) We believe that replacing T j by T j + 1 is a case of adopting
a qualified successor schema upon finding ourselves with pre-
decessor-anomalous experiences.
(vii) Therefore: We will replace T; by T/+ 1.
Schema (2) is a conceptually valid one, if corrective explanation is under-
stood in the sense of analysis (15) to be given later.
Let us next examine the premisses of (2) more closely. As far as
premiss (i) is concerned there are reasons to think that the experiences
concerning the domain R have been conceptualized within the bounds of
the conceptual scheme ofthe said best-explaining theory. Best explanation
ideally involves, among other things, explanation which is free from an-
omalies. This notion will be clarified in Section III below. The notion of
contribution in premiss (ii) can be understood in part causally. Despite
this (ii) is close to being a conceptual it can be argued. This is because the
idea of contributing to the achieving of best explanation involves precisely
the elimination of anomalies. (Note, however, that each theory T j , so to
say, conceptualizes its own anomalies.)
There is no need here to provide a detailed commentary to premisses
(iii)-(v). Let me note, however, that premiss (ii), as understood here, in-
volves the notion of corrective explanation. This notion is important and
requires further elaboration, to be given later. But first we shall discuss
other topics, beginning with the notion or (ordinary) scientific explanation.
II. A PRAGMATIC ACCOUNT OF SCIENTIFIC EXPLANATION
I shall begin with the general idea of scientific explanation and accept as
my starting point the question-theoretic analysis I have presented else-
where (Tuomela, 1980; cf. also Tuomela, 1984a, Chapter 10). According
to this analysis scientific explanation can be viewed as communicative so-
cial action which consists of querying and answering. To obtain a brief
overview of this approach we shall examine the following tentative defi-
nition in which C refers to a particular explanatory context, P to relevant
'paradigmatic' background assumption (cf. Kuhn), and q to the oratio
obliqua form of a question (generally, a why-question):
(3) Agent A explains q scientifically to agent B by producing a

linguistic token utterance u in the context C relative to the
background assumptions P if and only if
(a) A believes that u is a scientific explanatory answer to the
question q (in the context C, relative to P) or at least that u
represents such an answer; and
(b) A produces u with the intention that producing it brings
about in an intended way that B understands q (in the context
C, relative to P) and that B's understanding q comes about,
in an intended way, through B's belief that u is a scientific
explanatory answer to q or at least represents such an answer
(in the context C, relative to P).
This analysis relies on the notion of a scientific explanatory answer (se-
answer, for short). We need to clarify it. What I shall call a complete
explanatory answer is a linguistic token utterance which (in a specific
sense) satisfies the (complete) presupposition of the explanation-seeking
question q (i.e. the items presupposed by the question), one which is under-
standable to B, and one which forms a nomological scientific explanatory
argument for q. Let me give a nomological scientific explanatory argument
for q. Let me give a simple illustration of this idea. Suppose somebody
asks the question 'Why did this object a expand?'. To answer this question
in our account by implicitly answering the following so-called complete
question corresponding to the above question: (?/) (a expanded for the
reason fl. Let us denote this complete question underlying our original
question by q. What you find within the parentheses is called the complete
presupposition of q. An answer to q will now be given in terms of the
following argument: i) a is a piece of copper, ii) a was heated, and iii) all
pieces of copper expand when heated; therefore iv) a expanded. Now the
statement 'iv) for the reason (that) i)Eii)fEiii)' counts as a (potential) com-
plete scientific explanatory answer to q (apart from the obvious
empirical simplifications and idealizations involved). Denoting the con-
junction i)Eii)Eiii) by 10 we can say that an se-answer serves to give a value
(viz.fo·to the question quantifier (?/».
Let me now reproduce my technical definition (from Tuomela, 1980) of
a complete scientific explanatory answer as follows:
(4) A linguistic token-statement u is a complete scientific explan-
276 RAIMO TUOMELA
atory answer to q in situation C, given P, if and only if

a) u is obtained from a complete presupposition (relative to C
and P), say (Ef)sif), corresponding to q (formalized by
(?f)sif) in its oratio recta form) by dropping the existential
quantifier and by substituting a constant, say /0, for f in it;
b) u is P-understandable in the context C; and
c) u constitutes a nomological argument, viz. and E-argument,
for q such that this argument has/o as its conjunction ofprem-
ises.
This definition embeds a Hempelian kind of idea of explanation as a logical
argument into a pragmatic context of scientific inquiry. There is no need
here to clarify in more detail the technical clause a) of (4), for it has no
direct bearing on our present topic. But clauses b) and c) clearly are rel-
evant. For when we want to compare the goodness of several explanatory
answers to one and the same question there are two factors in (4) relevant
to that, and they are the following. First and foremost, we have the notion
of P-understandability in b). In comparative terms, the more understand-
ing an se-answer produces the better it is. Thus, ceteris paribus, explana-
tory goodness covaries directly with the amount and kind of understanding
it produces. Secondly, we have the factor of goodness of argument related
to c). The notion of E-argument in this clause means a nomological argu-
ment. In the deductive case we are dealing with e-arguments and in the
inductive case with p-arguments. (For lack of space I can here only refer
the reader to my previous publications (e.g. Tuomela, 1973, 1976, 1980,
1981) for a characterization and discussion of these notions).
In any case, we can say that, ceteris paribus, explanatory goodness co-
varies with the degree of nomic expectability the E-argument in question
guarantees. In this sense deductive explanations are better than inductive
ones, ceteris paribus. (However, the matter of nomic expectability is tricky
- cf. low-probability explanations.)
What I call a scientific explanatory answer simpliciter consists of men-
tioning only a part of the E-argument in a complete scientific explanatory
answer (see Tuomela (1980), definition (9) for the exact notion). Thus, in
our earlier example, for instance, 'Because a was heated' would be a (po-
tential) se-answer simpliciter.
III. WHAT IS BEST EXPLANATION?
3. Let us now proceed to the situation our schema (2), concerned with
theory-change, deals with. There we have predecessor theory Tin L (cor-
responding to Ti+ 1 in (2)). In the general case we have perhaps only a
translation function tr connecting the two languages extensionally (see my
proposal for defining one in terms of general logic in Tuomela, 1984a,
Chapter 14).
We shall assume that we can specify our problem area - a set of ques-
tions - within each language. The number of problems or questions re-
quiring answer in principle may be less in L than in L', for typically the
successor language L' carves up the world in a finer and more informative
way than L (cf. the discussion of monadic conceptual enrichment in Ni-
iniluoto and Tuomela, 1973, and below in Section IV). In any case, we
translate L into L' by means of tr in the general case and thus assume this
much comparability. As defined in Tuomela (1984a), tr is truth-preserving
(viz., if a sentence is true in L, then its counterpart in L' will also be true)
and therefore our comparability assumption is nontrivial. We can also say
that this function tr expresses a principle of information retention.
Now we continue by trying to compare the explanatory power of T'
with that of tr(T) rather than T itself. Both T' and tr(T) are in L'. In order
to avoid the clumsy notation tr(T) we shall normally use the variable T
for any predecessor theory, be it in L' or not. Next we assume that we
have available a common problem area for our two theories to be com-
pared, and by a problem (anomaly or other problem) we here simply
mean a question stated within L'.
Let me first define a notion that we will have some use for below. Let
thus u and u' be se-answers to q. We shall assume from here on that the
language L' can be understood in a very broad sense so that it incorporates
paradigm- and context-relativity, viz. relativity to P and C; thus relativ-
izing our analysandum to L' will conveniently take care of these elements.
So we propose:
(5) u' is at least as good as u in answering q relative to P and C if
1) u' makes q at least as much P-understandable (in C) as u
does; and
2) u' constitutes an E-argument for q which is at least as good
as that constituted by u.
278 RAIMO TUOMELA
Note that (5) only states a sufficient condition. This is because some com-
pensatory interaction between 1) and 2) might affect necessity and thus
neither 1) nor 2) by itself qualifies as necessary for the analysandum. For
u' might be a worse E-argument than u but render q better P-understand-
able in a way which makes u and u' equally good as a whole.
Let us now say that theory T generates a scientific explanatory answer
u if T occurs as a premise in the E-argument that u constitutes. We then
go on to define, relative to a set of explanation-seeking questions Q (rep-
resenting anomalies and other problems), explanatory bettemess for two
theories, T and T', in the same language. Relating the present situation to
the context of scientific growth, we may take T' to be the successor theory
of a predecessor theory, whose translation into L' is just T. We then pro-
pose:
(6) T' is a better explanatory theory than T relative to the set Q
of explanation-seeking questions and to L' if and only if
a) for every q in Q, if T generates an se-answer u for q then
also T' generates an answer u' for q such that u' is at least as
good in answering q as u is;
b) for some q in Q, the se-answer u' generated by T' for q is
better than the se-answer u (if there is one) generated by T for
q.
In (6) we have the notion of an se-answer being better than another one.
How can it be elucidated? Obviously (5) is of some help here, for it displays
the factors that are relevant here. Consider thus the following rather
obious suggestion:
(7) u' is a better scientific explanatory answer to q than u relative
to L' if
a) u' makes q to a greater degree P-understandable than u does
and u' constitutes an at least as good E-argument for q as u
does,
or
b) u' constitutes a better E-argument for q than u does and u'
makes q P-understandable at least to the same degree as u
does.
Note that the analysans cannot quite be regarded as necessary for the
analysandum in (7). For suppose u' makes q P-understandable to a very

high degree but constitutes a little worse E-argument than u. Yet u' can
be regarded as better than u. (See (7*) below for a remedy).
Note that the notion of P-understandability is a fairly global notion. It
relates to a range of conceptual abilities (possessed by a person said to
P-understand why q) relating the question q to other questions (and an-
swers). On the other hand, the goodness of an E-argument is a more local
notion concerned with something directly relevant to q, viz. with direct
nomic reasons for q.
Going back to (6), note that it allows for anomalies in the case of both
theories, viz. it allows for unsolved problems and unanswered questions.
However, T' is required to solve all the problems that T is able to solve.
In other words, we have built explanation retention in this sense into our
system. But this may seem a strong requirement, which in any case would
need an appropriate defense. For it should be noted that betterness could
be achievable also by means of some compensatory mechanism. T' might
be overall better than T even if it fails in the case of some problems that
T succeeds in solving, if it still is better in the case of most problems. In
view of this we should perhaps in analogy with (7) drop the only if - part
of (6) - or else (6) is to be taken strictly as a stipulative definition.
If one prefers not to accept (6) as a stipulative definition the above idea
of compensation should be accounted for giving up the explanation reten-
tion property. We shall consider one way of doing it. Suppose thus that
the explanatory value of an se-answer be numerically measured so that we
can take EV(u) to stand for the numerical goodness-value of u as an ex-
planatory answer to u (relative to L'). Obviously the EV-function will have
to take into account both the goodness of the involved E-argument and
the degree of P-understandability u confers on q. How that could be done
I shall not try to investigate (I am even somewhat suspicious of the feas-
ibility of this project). In any case, were we given the EV-value of every
se-answer a theory T generates for Q we could measure the global explan-
atory value (GEV) of Trelative to Q(= qt. ... ,qm) by the (weighed) average
of the singular EV-values, viz. by
1 m
GEV(n = - L k i EV(Ui)
m i=l
280 RAIMO TUOMELA
where k i is the 'importance' weight related to the answer Ui and the prob-
lem qh which Ui is an answer to. Then we can define:
(6*) T' is a better explanatory theory than T relative to the set Q
of explanation-seeking questions and to L' if and only if
<
GEV(T) GEV(T').
I will not here discuss the approach (6*) further except for remarking that
it does not - for better or worse - satisfy the explanation retention prop-
erty, contrary to (6). It suffices for our purposes below to accept either (6)
or (6*) for our analysis of the notion of better-explaining theory.
In view of the above we may now consider replacing (7) by
(7*) u' is a better scientific explanatory answer to q than U relative
to L' if and only if
>
EV(u') EV(u).
Note that (7*) may be combined not only with (6*) but with (6) as well.
Given (6) or (6*), we can say that if a theory T' is a better explainer
than a theory T then clearly T' is a better problem-solver than T - prob-
lems understood in terms of questions. We can also say that in this case
T' is in a pragmatic sense more informative than T. Note, too, that our
(6) and (6*) define strict betterness: if T' is better than Tthen Tcannot be
better than T'.
We can now proceed to our definition of the notion of best-explaining
theory in terms of (6) (or, if you prefer, (6*»:
(8) T' is a best-explaining theory relative to Q in L' and only if for
all T, T' is a better explanatory theory than T relative to the
problem-set Q.
Ifwe define explanatory betterness in terms of (6) in (8), the best-explaining
theory T' will be unique. But if we use (6*) this theory T' need not be
unique, for there may be ties. Note that (8) idealizingly concerns all the
logically possible theories in L'. Therefore it can safely be assumed that
the best-explaining theory is capable of answering all the problems in Q
even when Q exhausts all the problems statable within L'. Indeed, in many
cases when discussing best-explaining theories we might want Q to be the
set of all problems in L'. That of course would amount to quantifying
universally over Q in (8) and would mean omitting relativization to Q.
The following definition gives a weaker notion of best explanation:

(9) T' is a best-explaining theory relative to Q in L'
if and only if
for all T, T is not a better explanatory theory than T' relative
to Q in L'.
(9), contrary to (8), allows that some other theories of L' are as good as
the best~explaining theory in the sense (5). While (8) and (9) both accept
the possibility of several best-explaining theories they thus differ in this
particular respect. We need not make a definite choice between (8) and (9)
for the purposes of this paper, even if I prefer the more clear-cut notion
(8).
As above, in (8) and (9) the relativity to language L' is meant to reflect
relativity to the 'paradigm' P (and, where appropriate, to the context C),
and thus we are working here with an extremely wide notion of language,
for convenience. Note that if we could find a theory T' which is a best-
explaining theory for every possible scientific language then we would have
a best-explaining theory in a language-independent and absolute sense.
But it seems impossible to make good sense of the quantifier 'for every
language' - the class of all languages is surely an open class due to the
ever-present possibility of conceptual innovation. On the other hand, we
can surely get rid of the explicit relativization to L' (and the implicitly
involved relativization to P and C) in our above definitions by quantifying
existentially over L' in their analysantia. There is no need to perform that
simple maneuver explicitly here.
Another issue about our definitions worth mentioning is that our no-
tions of better explanation and best explanation are non-epistemic and
non-doxastic. We could, alternatively or in addition, have formulated the
above definitions in terms of what the explainer believes or has good
reasons to believe about the goodness of the se-answers he is giving.
These kinds of definition would be doxastic and epistemic, respectively;
but it has not been necessary to formulate such counterpart definitions
explicitly above.
The above account of better and best explanation can also be applied
to corrective explanation - a notion to be defined by (15) below. Then we
must in effect deal with approximative E-arguments and that will take
place in terms of the notion of a theory approximating another theory.
282 RAIMO TUOMELA
4. Let us consider still another approach to analyzing best explanation.

Thus consider the following definition of the notions of the deductive ex-
planatory power (DEP), inductive explanatory power (IEP), and total ex-
planatory power (EP) of a theory T:
(10) a) DEP(T) = {Sf e(S,T)}
b) IEP(T) = {Sf p(S,T)}
c) EP(T) = DEP(T) u IEP(T)
a) defines the class of statements S for which T together with some auxili-
ary statements is capable of giving a deductive explanation in the sense of
providing an e-argument for S. Here S should be regarded as either a
fact-describing sentence or a law-describing sentence, suitably character-
ized.
In clause b) we deal with inductive rather than deductive arguments and
thus use the notion of p-argument, viz. (minimal) inductive argument.
Clause c) takes the union of the previously defined two classes to char-
acterize the total explanatory power of T. Note that we have not explicitly
relativized (10) to the problem set Q but that can easily be done when
needed (cf. below). (10) is implicitly relativized to the language L', as T
and T' both are in L'.
To comment on the connection between (6) (or (6*)) and (10), we first
note that in (10) we are dealing with explanatory arguments, which are
roughly of the kind'S because T (and the auxiliary assumptions)'. While
such arguments have the form of se-answers they are not yet se-answers,
for they are not concerned with the central notion of P-understandability
at all. Thus the notions defined by (6) and clause c) of (10) cannot be
equivalent in any sense. But these notions are not unconnected, as we shall
see below in some detail. Both are conected closely with explanatory uni-
fication, for example, as better explaining theories tend to be connected
with larger EP-classes. (Recall here Hempel's emphasis on unification, see
e.g. Hempel, 1965, pp. 343-345, and 1966, p. 83).
Consider now the following suggestion, claimed to be true in view of
what the goodness of E-arguments in (6) involves:
(11) If T' is a better explanatory theory than T (relative to Q and
L'), then EP(T) <EP(T').
I would claim that (11) is at least ideally the case, for T' cannot be better
than T (in the sense of (6» unless it explains all that T does and a little
more. (Cf. formula (7.6) of Tuomela, 1973, for a more detailed explication
of this idea.) The converse of (11) obviously does not hold true without
strong qualifications, as the class EP has been defined without reference
to the notion of P-understandability. But if we could assume T' and T to
be equal in the degree of P-understandability they yield, the converse of
(11) would also be true.
Another interesting property that we can require a better explaining
theory to have in view of (11) and what I have elsewhere (especially in
Tuomela, 1973) said of E-arguments is that T' should not be to a greater
degree an ad hoc theory than T is. By adhockery I here mean especially
so-called theoretical adhockery. The denial of such adhockery entails that
when a (mature) theory is extended and applied to explain phenomena in
a particular domain, its theoretical core content should not increase in that
expansion process (viz. the theory is not allowed to be substantially
changed merely in order to fit a new domain). (For a technical explication
of this see Tuomela, 1973, esp. formula (7.7).) Let me state this property
or criterion of adequacy for better-explaining theories somewhat vaguely
as follows:
(12) If T' is a better explanatory theory than T (relative to Q and
£'), then T' exhibits less theoretical adhockery than T does
when applied to a new domain.
While the above two desiderata for good explanatory theories relate
primarily to the properties of E-arguments, it should be obvious that the
property (11) is directly connected to the unifying power a theory has and
via this to the degree of understandability it yields concerning the problem
the question q expresses. But also (12) concerns the unifying power of a
theory: the more ad hoc a theory is, the less it obviously can unify in a
proper sense.
5. Let us now go on and discuss scientific understanding briefly. Under-

standing is an ability. If a person understands something, he is able to
make relevant inferences concerning the matter, discuss it intelligently
(ideally at least!) and informatively, and (in many cases) to manipulate the
world in a relevant way. In Tuomela (1980) I have spoken about the reason
pattern (or causal pattern) and the hermeneutic pattern of understanding;
284 RAIMO TUOMELA
these are understood to cover jointly the mentioned aspects of understand-

ing. Without going deeper into that here let me just emphasize that scien-
tific understanding involves fitting the explanandum rationally into a
broader nomological network of laws, theories, and facts, which often
involve underlying and 'deeper' theoretical scientific entities (cf. Tuomela,
1973, Chapter VII, on this). Thus some degree of (deeper) unification as
well as systemicity are always involved in scientific understanding.
Scientific understanding is a very complicated notion. While I have
above taken up some central features involved in it, a great many others
could have been mentioned and discussed. Thus, we might consider, for
example, such proposed criteria of adequacy for explanations as consiIi-
ence, simplicity, stability, originality, heuristic power, and a host of prag-
matic and strongly context-dependent factors and try to relate all these to
our present discussion. We shall not, however, do it here but concentrate
on an aspect that has not yet been properly emphasized, viz. truth.
To do that, let me start by restating the partial technical characterization

of paradigm- and context-dependent scientific understanding given in Tuo-
mela (1980);
(13) A person B (scientifically) understands q in the situation C,
given P only if
a) q is a sound question and B believes it is;
b) B has at least some acquaintance with and some (concep-
tual) mastery of the items mentioned in, or presupposed by,
q;
c) B knows an answer u to q that is correct in C, given P;
d) B has some acquaintance with and some (conceptual) mas-
tery of the items mentioned in, or presupposed by, the answer
u (or, if u is not a complete se-answer, in the complete se-
answer u* which corresponds to u).
The above analysis of understanding is related to actual rather than po-
tential understanding and this feature is reflected by the use of the phrase
'correct' in c). When we are discussing the best explanation of the world
in the fullest sense it is obviously actual explanation and actual under-
standing in something like the mentioned sense we must be concerned
with.
The best-explaining theory T' can be seen to be involved with actual

scientific understanding as follows. Suppose (13) is satisfied in the case of
some q. Then the se-answer u must be correct rather than, for instance,
merely believed to be correct. Thus if a child asks 'From where did our
new baby arrive?' the answer 'A stork brought her' might satisfy it, but it
does not make the child understand from where the baby arrived in the
sense of actual (scientific) understanding. This is because the answer is not
correct. Analogous examples would be the attempts to explain combustion
phenomena by reference to phlogiston or various happenings by reference
to God's will or Zeus' thunderbolts. For even if such examples could be
elaborated to become some kind of potential explanations (which can be
doubted), they would fail as actual explanations and hence as best expla-
nations simply because there are excellent reasons to think that there are
no such things as phlogiston, Zeus or God. So correctness must be required
of best explanations.
What is it for an se-answer to be correct? Obviously it must involve that
in the corresponding complete se-answer corresponding to that se-answer
the premisses of the involved E-argument must be correct. But an E-argu-
ment is nomological, viz. it involves a law or theory. Now surely the best
scientific understanding of the world must involve that the theory in ques-
tion here is the best-explaining theory T'.
Thus we have arrived at the requirement that the best-explaining theory
T' must be correct. Correctness again can be equated with truth in some
presystematic sense. But as I have argued elswhere (especially in Tuomela,
1984b), this presystematic sense makes truth epistemic and viewpoint-de-
pendent. The next step in the argument is to claim that the best-explaining
theory T' is not only epistemically true but true in a suitable correspond-
ence sense (viz. picturing) as well. We shall below in Section IV work with
theories formalizable within first-order predicate logic. Each such theory,
which is a conjunction of axioms (general sentences), can be transformed
into its distributive normal form (in the sense of Hintikka). As a conse-
quence each true theory, say T*, can be expressed as a disjunction of
constituents, Ci , which are maximally informative generalizations. Thus
T * is logically equivalent to a disjunction of such constituents such that
the true constituent, say C*, will be one of them. Within our present frame-
work it is true that epistemic truth (in the sense of (17) and (19) below)
involves the true constituent C* - and C* will be argued to be true in a
286 RAIMO TUOMELA
correspondence sense as well. It follows that the best-explaining theory T'

is and must be correct in a sense involving correspondence truth.
Another way of showing that correctness involves picturing or corre-
spondence truth would go via Sellars' notion of (correct) semantic assert-
ability. For the correctness involved in (13) would then be argued to be
just correct semantic assertability in the Sellarsian sense. Given that, we
can continue with the claim that the latter notion entails picturing truth
in the case of singular descriptive statements as well as the corresponding
constituents (see below (22) and Tuomela, 1984b, Chapter 6, for details).
We have above sketched two arguments to the effect that correct expla-
nations involve picturing truth and that best explanations must involve
correctness. Thus we have given reasons to think that in the claim
(14) T' is the best-explaining theory if and only if T' is true (in the
sense involving maximal informativeness)
the implication from left to right is true. How about the other implication?
Could T' be true without being the best-explaining theory? Suppose T' is
a logical tautology. Then it would be true (in a logical sense) but surely it
would not explain anything. But our (14) excludes logical tautologies from
the true sentences we are concerned with. (14) requires complete material
truth, for that is what the maximal informativeness amounts to.
But given this, we can regard also the implication from right to left in
(14) as true. Why? Let us consider the matter within our above framework,
for simplicity. We are then dealing only with constituents, for only they
are maximally informative in the required sense. But as only one of the
constituents can be true (in a given language), T'in (14) must be regarded
as just the true constituent C* (or a sentence logically equivalent with it).
We now have to argue that C* (or a sentence logically equivalent with
it).
We now have to argue that C* must be just the best-explaining theory.
How do we do that? Why on earth should we think that the maximally
informative true theory is the best-explaining theory? The theory C*, due
to its truth correctly reflects all the causal connections and invariances in
the world, relative to its conceptual framework (cf. (22) below). But, we
may ask, what else can the best explanation possibly do but that same
thing as well? Thus, if the best E-argument serves to make the explanan-
dum event (or regularity, as the case may be) nomically expectable that
must in part be in virtue of its being true, we may argue.

However, the matter is not so simple. We recall that best explanation
involves not only (best) E-argument but (best) P-understandabilityas well.
Could a maximally informative and true theory in a language incorpo-
rating the information that P and C represent fail to be maximally P-
understandable? It seems to me that if we are dealing with even moderately
rational scientists in our P-community we should answer this question
negatively. For rationality in this scientific context can plausibly be taken
to involve the use of reliable and self-corrective means for getting at correct
descriptions of the world and accordingly at correct se-answers (to Q-prob-
lems). It would then seem that at least in the long run rational scientists
possessing the true theory are bound to achieve the best achievable under-
standing of the world (recall (13) and our remarks related to it). So, with-
out here pursuing the matter further, I will assume - at least tentatively
- that the implication from right to left in (14) holds true.
Let me emphasize here again that both our notion of the best explaining
theory and our notion of (maximally informative) truth (epistemic truth
being conceptually the key notion) are relative to the paradigm P and the
context C, the former explicitly and the latter implicitly. But both notions
are to an equal degree viewpoint-dependent in the mentioned sense. As in
effect noted, there seems to be no way to get completely rid of this depend-
ence, nor should there be.
6. Let us now return to the schema (2) and to the notion of corrective
explanation in it. It has to do, first with the fact that the theory Ti+ 1
(corresponding to the better-explaining theory T' above) must be able to
give acceptable answers to all those why-questions which are connected
with the relevant Tranomalies Et. ... ,En and which are couched in the
language and conceptual scheme of Ti + 1. Secondly, Ti + 1 must also be able
to explain the success of Ti and hence to answer satisfactorily all the
why-questions answered acceptably by Ti • Recall nevertheless that both
theories formulate the questions from the point of view of their own con-
ceptual schemes. To put it somewhat crudely, theory T i + 1 will on the basis
of this be essentially superior to Ti as to its explanatory power over the
phenomena dealt with by Ti and the anomalies Et. . . .,En. (This claim holds
true at least if Ti + 1 does not create numerous - and serious - new anom-
alies relative to those of Ti.)
288 RAIMO TUOMELA
I have given an account of corrective explanation elsewhere (see Tuo-

mela, 1984, Chapter 14). Here I confine myself to sketch its basic features.
For simplicity, I shall leave out reference to the anomalies E 1o ••• ,En (as-
sumed to be included in the class Q of relevant problems) and assume that
T i + 1 manages to explain all the Ti-anomalies. Condition a)i) in the fol-
lowing analysis is to be understood to incorporate these simplifying as-
sumptions. As before, we shall assume that the explanation takes place in
or relative to the language or conceptual framework of the successor
theory T i + 1 • This relativity is left implicit in the following definition:
(15) Theory T i + 1 explains correctively theory Ti if and only if
a) there is a translation function tr from the conceptual scheme
(or language) of Ti into the conceptual scheme (or language)
of T i + 10 and an auxiliary hypothesis H and a theory S (both
in the language and conceptual scheme of T i + 1) such that
i) Ti + 1 jointly with H explains S, and
ii) S is an approximation of tr(Ti), where tr(Ti) is the trans-
lation of Ti into the conceptual scheme (or language) of T i + 1 ;
b) Ti + 1 is a better explanatory theory than T i.
Explanation in schema (15) means giving an explanatory argument (E-
argument), in the first place. An E-argument is not yet a scientific explan-
atory answer in the sense specified in definitions (3) and (4). But it is not
so far from being one. We could indeed have formulated clause a) in that
terminology and required an answer of the form'S because Ti + 1 and H'
to be an se-answer in clause a)i). In this way we can tie up schema (15)
with our earlier account of explanation. Given this connection, we can
safely proceed in terms of (15). (But, if necessary, when using (15) we may
assume that the so-called pragmatic factors in schema (3) bring about no
essential alterations in the situation and that we hence can confine our-
selves to the analysis of explanatory arguments.) Apart from some clari-
fying remarks, I shall not give a more detailed analysis of the contents of
schema (15). (See Tuomela, 1984a, Chapter 14, which specifically addresses
the difficult notion of translation.)
In clause b) reference to Q has been omitted but otherwise it should be
understood basically in the sense (6) (or (6*». However, there is the dif-
ference that (6) literally applies only to the comparison of T i + 1 with S. To
get to b) we must make additional moves. First, there is the move of
translating Ti into the language of T i+ 1. (Note that tr(Ti) corresponds to

the theory Tin (6) and (6*).) Secondly, there is the transition from tr(Ti)
to S. What these two moves involve in terms of explanatory power in
general is a difficult problem we cannot try to solve in this paper. (Note
that while clause a)i) may be taken to entail that Ti+ 1 is a better explan-
atory theory than S it does not without further argument entail clause b)
on the basis of our earlier definitions.) Still it seems reasonable to assume
that, given a truth-preserving notion of translation and a relatively strict
notion of approximation, it holds that if Ti + 1 is a better explanatory
theory than S (relative to Q and hence E b ... ,En) then it is also better than
Ti (relative to its conceptualization of the problems and anomalies).
Note, furthermore, T i + 1 may explain S only inductively. This together
with the mentioned leaps from Ti of tr(Ti) and from the latter to S, leaves
room for several kinds of deductive inconsistencies between T i+ 1 and Ti
as well as between T i + 1 and the various presuppositions of not only the
conceptual framework of Ti but that of T i+ 1. too. Thus, for instance, T i+ 1
may be able to correctively answer a question with a grossly false presup-
position by answering a question with a more truthlike presupposition
such that the former presupposition anyhow approximates the latter; cf.
the relation between Sand tr(Ti). But we shall not here make an attempt
to systematically botanize such inconsistencies and discrepancies.
Recall that our basic schema (2) involves corrective explanation. We can
now say in view of (2) and (15) that science progresses essentially via tran-
sitions to ever better-explaining theories about domains of inquiry. How-
ever, not even this condition gives a priori guarantee to the claim that
science ultimately ends up with the best-explaining theory (i.e., there is no
necessity that there be such a theory).
IV. INDUCTIVE LOGIC, EPISTEMIC TRUTH, AND BEST EXPLANATION
7. Given our account of best explanation and of comparative explanatory

power we shall now proceed to a somewhat technical discussion of epis-
temic truth and best explanation within the context of formal first-order
languages in an inductive context. We shall examine a slightly simplified
case in which several relevant factors are assumed to stay constant. Let us
consider a situation of growth of knowledge in which the language of the
theory Ti is L(A), and where A is a set of scientific predicates (here mon-
290 RAIMO TUOMELA
adic). Thus Ti only contains A-predicates. The language L(A) can in prin-
ciple take into account assumed meaning postulates and constitutive prin-
ciples by ruling out the sentences (and possibilities of classification) they
deny. We then assume that the language of the theory TiXl is L(A+{RD,
i.e. a language which has been obtained by adding to the language L(A) a
new monadic predicate R. We have here to do with growth of knowledge
based on simple enrichment of a conceptual scheme. (Note that there is
another simplifying feature here: the ontologies of the two theories are
essentially the same; the case of theories with differing ontologies is exam-
ined e.g. in Tuomela (l984a), Chapter 14.)
The theory Ti can thus be redescribed in the language L(A + {RD, and
hence be compared directly with T i + 1 • In other words the redescribed
theory Ti is in fact the theory S of condition a)i) in (15). We thus ground
the comparison of the two theories Ti + 1 and Ti on the comparison of the
theories Ti + 1 and T*. We can then apply to this case the system of in-
ductive logic developed by Hintikka, in the form it has been further de-
veloped in Niiniluoto and Tuomela (1973).
Each general sentence S of the language L(A+ {RD can now be repre-
sented as a disjunction of the constituents Ci of this language: --, S+-+C 1
v ... V Cn, for some selection of constituents Ct. ... ,Cn. If the number of
A-predicates in m-l, the number of the constituents of the richer language
is 2K , with K=2m. We can illustrate the situation by saying that our lan-
guage partitions the world into K 'cells' or types of individuals. Each con-
stituent is a strongest (maximally informative) possible description of such
a world, and the description says that in the world a certain number w of
the cells or types are exemplified, while the rest K - ware unexemplified.
Next, we shall assume that, relative to the research community A, we
have defined a distance measure dA for the sentences of the language
L(A + {RD such that O(dA ( 1. (See e.g. the possibilities presented in Niinil-
uoto, 1977, 1978, and Tuomela, 1978.) In this way we can further define
a kind of coherence measure for sentences. Thus, omitting reference to A,
we have:
(16) k(S,S') = I-d(S,S')
represents the degree of epistemic coherence or similarity between the sen-
tences Sand S' by help of the distance measure d(S,S'). This coherence
is in part based on purely formal, syntactic similarity, in part on epistem-
ically interpreted free parameters (see Tuomela (1978), (1980c); I shall dis-
cuss the epistemic interpretation of the measure d later). Sentences Sand
S' thus both represent potential or partial knowledge (from the point of
view of the community A). Once we have defined epistemic probabilities
for these sentences we can talk of degrees of knowledge. Pursuing this line,
we shall assume that we have defined, for the sentences of the language
under scrutiny, and in accordance with Hintikka's inductive logic, an ep-
istemic probability measure P A where the index A - which will often be
suppressed below - refers to relativization to the community A. PA now
represents degrees of knowledge. It satisfies the standard probability ax-
ioms as well as some symmetry and regularity principles (see e.g. Niinil-
uoto and Tuomela, 1973 and Hintikka and Niiniluoto, 1976). We shall
pay particular heed to the conditional probabilities P(C;je), which tell us
the epistemic probability of a constituent C j relative to some evidence e.
Here e is assumed to effect an unequivocal partitioning of the observed
individuals into types. As a general epistemological warrant for the meas-
ure P we can employ some suitable theory of knowledge, e.g. a causal one.
For lack of space we cannot here discuss this matter (see Tuomela, 1984b,
Chapter 9).
Given the above, we now reason as follows. The conceptual scheme
represented by the language L(A+{R}) is at bottom an epistemic system
which comprises the rules which govern human action (cf. the Sellarsian
world-language, language-language, and language-world rules) and the
practical inferences which ground or give a warrant for these rules. Part
of all this is structurally built into the language L(A+ {R}) itself as well as
in its constitutive postulates, part is taken into account by the probability
measure P A and the coherence measure k A •
It is not in fact altogether farfetched to think that men ideally aim at
maximizing their expected epistemic utilities - this is but a generally ac-
cepted rationality principle. Why could we not, then, think that epistem-
ically ideal truth is analyzed in this fashion? Isn't truth precisely what is,
ideally, expected to be epistemically best? The epistemic notion of truth,
I surmise, has to do with precisely this.
We can then give the following definition for the epistemic truth of a
sentence S in the language L(A+ {R}):
(17) S is epistemically true relative to a research community A and
the language L(A + {R}) if and only if S maximizes the expected
epistemic utilities of the community A.
292 RAIMO TUOMELA
Here we assume that the community A uses language L()" + {R}). We shall
make the further simple assumptions that the only epistemic utility of the
community A is epistemic coherence in the sense of measure kA and that
expected utilities are formed by help of the measure P A. Insofar as the
probability measure PAis at all an adequate measure of the degrees of
knowledge of the community A (and at the same time also, perhaps, of
the degrees of its rational, epistemica11y warranted beliefs), its maximum
value of course must stand for factual truth, in fact precisely picturing
truth (cf. Sellars, 1968, and Tuomela, 1984b). To put it in the Sellarsian
fashion, P A primarily takes into account the relationships or rules between
language and the world. On the other hand, the measure for coherence kA
takes into account the central epistemic feature of truth - its dependence
on background information (for discussion see Tuomela, 1984b, Chapter
6). Let us see where this assumption leads us.
We shall start with the formula for expected epistemic utilities, omitting
again the subindex A everywhere:
2K
(18) EU(S/en) = L P(Cden) k(S,C;)
;= 1
Here the subindex n refers to the quantity of 'observations' of individuals

contained by evidence e(n). The expected epistemic utility of a sentence S
is, then, calculated as follows. The epistemic coherence of sentence S, rel-
ative to the language L()"+ {R}), with each basic element or constituent
C;, is measured by help of k A • The higher is k(A)(S,C;), the greater is the
epistemic utility of the sentence. These utilities are then weighed by the
degree of knowledge P A of each constituent. This degree is understood to
be relativized to all the available evidence e(n) which can be fully repre-
sented in the language L()"l + {R}).
We now explicate definition (17) as follows:
(19) Sis epistemically true with respect to a research community A
and language L()..+{R}) if and only if S maximizes formula
(18), when n-+oo, for all generalizations of the language
L()..+{R}).
In Hintikka's inductive logic there is a unique constituent, Ce, such that,
when the amount of evidence grows without limit, viz. n-+oo, then
P( Celen) -+ 1. This constituent Ce says that the world is of the same type as
the asymptotic evidence: exactly those c kinds of individuals are exempli-
fied in the world as in the evidence. Here c expresses the said number of
kinds of individuals. We now have that
and
(21) EU(Celen)-+I, when n-+oo with a constant c.
A sentence say S*, which maximizes formula (18) and hence satisfies for-
mula (19) must therefore be logically equivalent with constituent Ce rela-
tive to asymptotic evidence. Furthermore, the following equation must
hold: k(S*,Cc) = 1. We could, as far as I can see, require more generally
even that k(S,Ce) = 1 only if S= Ce. Formulas (20) and (21) show, in a
sense, that in the asymptotic case the import of the 'internal' background
knowledge (but not of the 'external' background knowledge required for
language use) of a language system for an epistemically true theory in the
end vanishes - the crucial role will be allotted to its probability.
8. Although, in pursuing the above line of thought, we have had monadic

predicate logic in mind, the results (20) and (21) are valid for the whole
first-order predicate logic. They show in an intriguing way the possibility
of connecting asymptotically epistemic truth with picturing truth (i.e.,
when n -+ 00). For result (21) shows that constituent Ce is epistemically true
in the sense of definition (19) and that every epistemically true sentence S*
is logically equivalent with Ce, relative to asymptotic evidence. In the
asymptotic case the distance of an arbitrary sentence S from epistemic
truth is according to (20) given by k(S,Ce), i.e. the degree of epistemic
coherence of S with Ce.
We can now think that the truth of evidential sentences is defined as
picturing truth (see Tuomela, 1984b, Chapter 6). Thus the evidential sen-
tence-token 'a is of kind Q;' can be thought to be a Sellarsian verifying
token. Although evidential sentences in Hintikka's system in the end are
construed as existential and 'statistical' (e.g. 'There are p individuals of the
type Qn, the philosophical setting is not thereby changed in any essential
way. When the amount of evidence grows without limit, it is quite natural
to generalize the definition of picturing truth to cover also constituents in
294 RAIMO TUOMELA
this asymptotic case. For all that a constituent says is how the individuals
of the world are partitioned to different cells, when all this information in
fact exists in the singular evidence obtained through causal picturing. Al-
though I shall not here give any technical argument, I will assume that
unlimited (i.e. n-+ 00) singular description of the world through picturing
is a necessary and sufficient condition for the determination of the pictur-
ing truth of a constituent C;, and that Cc is true in the picturing sense if
and only if evidence in fact (asymptotically) exemplifies the c kinds of
individuals claimed by Cc to be exemplified.
As already said, we shall, on the other hand, define the epistemic truth
of constituents and other sentences by means of (19). Since (19) refers to
an asymptotic case, it follows that the notions of picturing truth and ep-
istemic truth in fact coincide in the asymptotic case. To avoid unnecessary
technical complications we shall assume that sentence S has been trans-
formed into its distributive normal form. Relying in part on this we claim
that k(S,Cc) = 1 only if S=Cc, so we get:
(22) S is epistemically true if and only if S is true in the picturing

sense, when n-+oo.
The results presented above give a rather satisfactory warrant for equiv-
alence of the claims (a) and (b) of hypothesis (1). All in all, we have claimed
that in the general case factual truth amounts precisely to epistemically
ideal truth, i.e., that l)a) is to be construed to be equivalent with claim
l)b). But analyzing l)b) with the help of formula (19) - which need not be
congenial to all supporters of epistemic truth - shows that at least for our
model languages (first-order languages) epistemically ideal truth turns out
to be, on the basis of some rationality and adequacy conditions, equivalent
with asymptotic picturing truth.
Let me note that we can measure degrees oftruthlikeness in the language
L(A. + {R}) irrespective of how we solve the problem of estimating epistemic
truth. This we can do simply by using formula (18). Thus the value
measures the degree of truthlikeness of an arbitrary sentence S, when S*(lJ

is epistemically true in the sense of (19). (S* is, however, often or usually
unknown in actual research - cf. the condition n-+ 00). When we investigate
a situation of growth of knowledge we can examine if the difference
in fact is smaller than the difference

EU(S*jen - EU(Ti*je n)
that is, if EU(Ti + 1 jen) >EU(Ti*jen) and if T i + 1 is thus closer to truth than
T*(z}. Note that when n-+oo, we are here according to (20) essentially
making comparisons between the degrees of coherence k(Ti +1,Cc) and
k(tr(Ti'Cc)).
I have above sketched a means by which the degrees of truthlikeness
and truth of successive theories can be studied, provided we can ground
the comparisons in the language or conceptual scheme of the successor
theory. In the general case this presupposes the specification of the trans-
lation function tr as well as solutions to problems related to the function.
The above discussion deals with linguistic enrichments (transitions from
the language L(A.) to the language L(A. + {RD), which are relatively simple
from the point of view of theory comparison. As said, for the general case
I shall refer to Tuomela (1984a), Chapter 14, where the problems of the
translation function and comparability are examined with the help of gen-
era110gic. We can add here that carrying out such a translation involves
clarifying the conceptual relations between languages (cf. (15)). On the
naturalistic side, so to say, there is the corresponding transition from pic-
turing within one linguistic system into picturing within another, successor
linguistic system. As I have clarified in Tuome1a (1984b), picturing is in a
sense tied to the ought-to-be-rules of a language (in the Sellarsian sense),
although it is describable, according to the view advanced, as non-inten-
tional though goal-directed action quite independently from the rules of
a (any) language. This opens a way to characterize, in principle at least,
the degrees of adequacy for describing the world, which is language-
independent. In a sense it enables a naturalistic means of justifying pro-
gress of knowledge which takes place on the conceptual side.
9. Let us now turn to the topic of best explanation. We claimed above that
the epistemically true theory in the language L(A. + {R}) will be precisely
Cc (or at least a theory which is logically equivalent with it). But this theory
may be claimed to also give the best explanation of the sentences of the
296 RAIMO TUOMELA
language L(A. + {RD. This holds at least on the condition that explanatory
power is measured by a means which satisfies the following likelihood-
condition, and selects, in this sense, the most informative (logically strong-
est) of the best theories. The likelihood-condition says that when we com-
pare two theories Tl and T2 for their respective explanatory powers rel-
ative to the sentences g of the language L(A.+ {RD (specifically, relative to
the sentences g(A.) which contain only A.-predicates and which represent
anomalies), and when the evidence or background knowledge comprises
the sentence e, T2 gives a better explanation of g than Tl if and only if
(24) P(g/e&T2) >P(g/e&T 1)
where P(g/e&T2) is the so-called likelihood-probability or the conditional

probability of g relative to evidence e and theory T;, i= 1,2. Here T2 can
be a successor theory which contains A. + (R)-predicates and Tl and pre-
decessor theory which only contains A.-predicates (or, to be more exact, Tl
is the counterpart theory, couched in the language L(A.+{RD of the suc-
cessor theory, of the predecessor theory originally presented in the lan-
guage L(A.). More generally, we may measure the average explanatory
power of T2 and Tl relative to all the observational statements (at least
generalizations) in the observational, or rather predecessor counterpart
language, L(A.).
Such measures of explanatory power (relative to sentences g(A.)) have
been examined e.g. in Niiniluoto and Tuomela (1973), Chapter 7. Applied
to Hintikka's inductive logic the measures of explanatory power which
satisfy the likelihood-condition give precisely the result that Ce is asymp-
totically the best-explaining theory (in the likelihood-sense) in the language
L(A.+ {RD. This of course is based on the result (21), which implies that
only the likelihood probability of the constituent Ce receives a non-zero
value when n-+oo. This being so the explanatory power of other theories
can be studied relative to the explanatory power of this constituent.
The notions of explanatory power and best explanation are, however,
relatively complex matters as we have seen. Our present investigations in
any case indicate that, at least for first-order languages and within induc-
tive contexts, the best-explaining theory will be or contain the most in-
formative true theory, viz. Ce• Note that our above treatment only depends
on the goodness of inductive E-arguments, viz. p-arguments. If we now
recall our definition (8) of best explanation we notice that we should also
consider matters related to P-understandability. We can, however, quickly

see that matters related to understandability cannot change the validity of
our above claim, for best explanation (in inductive contexts at least) in-
volves truth regardless of the degree of P-understandability the best-ex-
plaining theory confers upon the explanandum-statements.
But let us now consider the converse statement, viz. that the (maximally
informative) true theory must also be the best-explaining one. As we no-
ticed in our discussion of formula (14), this is a somewhat moot point. We
can of course trivially say that given that a theory T yields a maximal
amount of P-understandability, then if T is true it must also be the (or a)
best-explaining theory. For, as we have previously argued, the true theory
Cc will guarantee maximally good (inductive) E-arguments. But we can
say more, as we did before in Section III. A rational community of scien-
tists will at least in the long run come to understand the world better when
in possession of a true theory than when only a (perhaps grossly) false
theory is available to them, and this concerns both causal and hermeneutic
understanding in the sense of Section III and Tuomela (1980). In any case
the above discussion and results lend support to the truth of our hypothesis
(1) concerning the equivalence of best explanation with epistemic truth.
This equivalence is not of a purely conceptual nature. Rather, it hinges in
part on some rationality assumptions as well as on our views on expla-
nation.
If explanation and truthlikeness are thus tied up with the constituent Cc
of the language L(A. + {R}), and if the connection is preserved in the tran-
sition to the successor language, and so on, science does in fact get closer
to the truth - provided that we assume that scientists always act according
to the schema (2) (or some similar schema) and succeed in bringing about
ever better-explaining theories. The converse also holds: with the above
proviso, science proceeds towards ever better explanatory theories if it gets
closer to truth. There are also various kinds of formal results about, say,
when some successor theory gives a better explanation of empirical gen-
eralizations and anomalies than its predecessor, in the sense of formula
(15) (cf. Tuomela, 1973, Chapters VI and VII and Niiniluoto and Tuomela,
1973, Chapter 7).
In (2) and (15) we have chosen as the point of departure the requirement
of ever-increasing explanatory power. If the above connection is accepted,
we have as a consequence also ever-increasing truth-likeness. Note never-
298 RAIMO TUOMELA
theless that our assumptions make it in principle possible that science con-
tinues to grow forever and without limit, for it is possible that there always
are better and better languages and conceptual schemes.
Let us now consider (2) and ask why, then, the scientific community A
should infer and act in accordance with it. One general reply to the ques-
tion is that doing so is rational. In other words, it is rational to aim at best
and perhaps exhaustive explanations in science, although such explana-
tions might never be found, on the ground of ever possible conceptual
innovations. We may also say (in view of the connection between best
explanation and truth) that it is rational to hold that the aim of science is
to tell us what there is - scientia mensura. The means for striving for these
goals of truth and best explanation, presented in formula (2), can also be
considered rational, though this is no place to argue for it in detail. (Let
me here also refer to the relevant but quite different Kantian arguments
which Rosenberg (1980) brings to bear on his corresponding formula
which, in part, resembles (2).)
In nuce: my thesis, founded on the above results, concerning progress
in science is that science, rationally pursued, grows towards ever better-
explaining theories and, therefore, towards ever more truthlike theories.
As such this is not a factual claim concerning scientific progress, but rather
a rational-normative claim of, roughly, the form 'The scientific community
A ought to act rationally; and if it is rational it acts in accordance with
formula (2). As a result of this science grows towards ever more truthlike
theories'.
We have consequently run into the distinction normative (or norma-
tive-rational) .versus factual, for applying the schema of inference (2) to a
community A presupposes that A is rational. If A is not rational in the
sense specified, its theories may in fact develop in a quite different way.
Whether A de facto is so rational and whether science actually will develop
towards increasingly better-explaining and truthlike theories are naturally
problems which only the future development of science can solve.
REFERENCES
Hempel, C. G.: 1965, Aspects of Scientific Explanation, The Free Press, New York.
Hempel, C. G.: 1966, Philosophy of Natural Science, Prentice-Hall, Englewood Cliffs.
Hintikka, J, and Niiniluoto, I.: 1976, 'Axiomatic Foundation for the Theory of Inductive
Genera1ization', in PrzeIecki, M., Szaniawski, K., and Wojcicki, R. (eds.), Proceedings of
the Conference ofFormal Methods in the Methodology of the Empirical Sciences, D. Reidel,
Dordrecht, pp. 57-81.
Niiniluoto, I.: 1977, 'On the Truthlikeness of Generalizations', in Butts, R. and Hintikka, J.
(eds.), Basic Problems in Methodology and Linguistics, D. Reidel, Dordrecht, pp. 121-147.
Niiniluoto, I.: 1978, 'Truthlikeness: Comments on Recent Discussion', Synthese 38,281-330.
Niiniluoto, I. and Tuomela, R.: 1973, Theoretical Concepts and Hypothetico-Inductive Infer-
ence, D. Reidel, Dordrecht and Boston.
Rosenberg, J.: 1980, One World and Our Knowledge of It, D. Reidel, Dordrecht, London and
Boston.
Sellars, W.: 1963, Science, Perception and Reality, Routledge and Kegan Paul, London.
Sellars, W.: 1968, Science and Metaphysics, Routledge and Kegan Paul, London.
Tuomela, R.: 1973, Theoretical Concepts, Springer-Verlag, Vienna and New York.
Tuomela, R.: 1976, 'Morgan on Deductive Explanation: A Rejoinder', Journal of Philosoph-
ical Logic 5, 527-543.
Tuomela, R.: 1978, 'Theory-distance and Verisimilitude', Synthese 38, 213-246.
Tuomela, R.: 1980, 'Explaining Explaining', Erkenntnis 15, 211-243.
Tuomela, R.: 1981, 'Inductive Explanation', Synthese 48,257-294.
Tuomela, R.: 1984a, A Theory of Social Action, D. Reidel, Dordrecht, London and Boston.
Tuomela, R.: 1984b, Science, Action and Reality, D. Reidel, forthcoming.
Dept. of Philosophy
University of Helsinki
Unioninkatu 40B
00170 Helsinki
Finland
RAINER W. TRAPP
UTILITY THEORY AND PREFERENCE LOGIC
This paper consists of two parts. In the first one (which has two sections)
I begin by drawing attention to a fact that in my opinion has not been
sufficiently considered in constructing systems of preference logic (PL),
namely that the standard use of preference relations is such that the relata
of all preference comparisons have to be (at least) alternatives, in a sense
according to which altemativeness is more than mere incompatibility in
the actual world Wo but less than logical incompatibility or incompatibility
in all possible worlds Wi. I discuss some aspects of this assertion against
the background of utility theory (UT), and give, in the second section of
part I, some reasons backing my claim that it is true. In the second part
I make use of several conclusions from the first part, in discussing both
the acceptability of certain controversial rules and axioms suggested for
PL and the adequacy of what I hold to be the most appropriate type of
semantical groundwork given so far for PL. I also distinguish there be-
tween basic and nonbasic preference. The result of these reflections will be
that, though this semantical groundwork is provided by some hardly dis-
putable insights of UT, it is not, despite contrary claims, also adequate for
giving the truth conditions of standard preference statements. Moreover
I shall argue that the mere truth-functional structuring of preference relata
is, for several reasons, an insufficient formal equipment for an
appropriate logic of standard preference.
(Ia) I hold the following (= Thesis I) to be true: By the very nature of
preference no two relata of a preference relation should be considered to
be true in the same possible world Wi, all Wi (including the actual world
Wo itself) being elements in a certain class of possible worlds at most
minimally different from Woo
This requirement of altemativeness is at least made for what I consider
to be the standard meaning of preference comparisons, with which I ex-
clusively deal here. A standard preference statement 'x>-y', made by per-
Erkenntnis 22 (1985) 301-339. 0165-0106j85j022(H)301 $03.90
302 RAINER W. TRAPP
son a, shall be one to which corresponds a factual or counterfactual (ideal)

choice situation Se, such that person a strongly prefers x to y iff person a
in situation Se would rather choose state, course of events, etc., x to hap-
pen, i.e. a corresponding proposition to be true, than state, course of
events, etc., y to happen. (' >-' and '",' stand for strong preference and
indifference, defined as usual by weak preference '-':,' as the basic relation.)
The notion of choice already implies that exactly one of the two relata
will be true in person a's preferred world Wi. the other one being false in
Wi, and vice versa for the different world Wj to which Wi is preferred. In
short: getting one alternative implies renouncing the other one. However
mild this characterisation may appear - and therefore a temptation to
think that all types of preference relations distinguished in the literature
are special subcases of standard preference - it will nonetheless become
clear that certain properties of HP-conceptions ofPL (see Note 1), and its
semantics, presuppose a different notion of preference than the standard
one. Later it will be helpful to further distinguish between (1) basic or ideal
standard preference, involving only utility comparisons of two single pos-
sible worlds, and (2) non-basic or probabilistic standard preference, in-
volving more complex utility-probability comparisons between two dif-
ferent sets of possible worlds. It will also be helpful to make the distinction
between elementary and molecular preference, the latter involving at least
one truth-functionally complex relatum, the former only containing
atomic relata.
I shall not consider axiologically motivated subdivisions of standard
preference. One may, e.g., combine the standard preference relation with
a certain axiological criterion Ca and stipulate for a Ca-preference the fol-
lowing truth condition: 'A>-B' is true iff the intrinsic value of state A
according to criterion Ca is higher than that of state B. Chisholm/Sosa
(1966), for instance, assume only the pleasure ofliving beings to be intrins-
ically good; and on this hedonistic basis hold it to be true that it is better
that there be happy egrets than that there be stones, or that it is equally
good that there not be happy egrets and that there be stones, the quantity
of pleasure in both latter cases being zero. I consider the standard pref-
erence relation, in contradistinction to such special cases, to be strictly
value-neutral. Consequently a person could profess all possible preference
or indifference relations between the states mentioned without being blam-
able for making false statements.
UTILITY THEORY AND PREFERENCE LOGIC 303
Before giving thesis I a more exact setting by formulating an adequacy

condition for standard preference, in terms of possible worlds semantics,
let us sustain by a few informal considerations its claim that preference
relations are used in such a manner that they do not obtain between states
existing in the same world, if these states are described with appropriate
exactness. Having been confronted with views, according to which for no
kind of preference relation - be it standard in the sense given or not - is
incompatibility (in certain worlds) of the states compared a prerequisite,
I shall say a few more words on this.
Following von Wright I take generic states (or processes etc.) to be the
relata of preference relations. Accordingly, whenever things are preferred
to one another, as, e.g., in 'I prefer apples to oranges', this is to be taken
as an abbreviation for 'I prefer eating apples to eating oranges' (or what-
ever else may be meant), if only the relata are syntactically representable
by propositions. For since individual constants or class names (or what-
ever else be syntactically beyond the propositional status) are not bearers
of truth-values, all considerations as to the truth values of these relata
would be immaterial anyhow. Moreover, the use of truth-functional con-
nectives requires our interpretation.
It is not necessary for a person who prefers one state to another to be
in an actual decision situation. The choice situation may be purely coun-
terfactual- even involving fictitious alternatives, such as the queen of She-
ba being one's wife. Preferring one state to another basically just means
preferring one possible world of a certain kind to another, no matter how
likely the becoming actual of these worlds is (as judged by the person who
has a preference or somebody else), and no matter if one of these possible
worlds, or neither of them, already happens to be (part of) the actual
world. Let us have a look at the ordinary use of preference statements. It
surely makes sense to say
(1) 'I prefer Reagan being the president to Carter being the presi-
dent' (case T>-Fin Wo in 1983); or,
(2) 'I prefer Carter being the president to Reagan being the presi-
dent' (case F>T); or,
(3) 'I prefer Carter being the president to Goldwater being the
president' (case F>- F).
304 RAINER W. TRAPP
But, does it likewise make sense to say

(4) 'I prefer Reagan being the president of the U.S.A. to Andro-
pov being the party chief of the USSR'; or,
(5) 'I prefer more than 25° C to a blue sunny sky';
if in fact, Reagan is the president and Andropov is at the same time the
party chief and if, at the same time, the temperature is above 25° C and
the sky is blue and sunny? Of course, it would make sense if, instead of
(4) and (5), as they stand, one said, e.g., (4') 'I prefer Reagan being the
president of the U.S.A. (without Andropov being the party chief of the
USSR) to Andropov being the party chief of the USSR (without Reagan
being the president of the U.S.A.)' and (5') 'I prefer more than 25° C
(without a blue sunny sky) to a blue sunny sky (without the temperature
being higher than 25° C)'. But (4') and (5') are no longer, as (4) and (5)
are, examples of the T» T-case (in the real world Wo); but, under the
assumptions made, of the F>- F-case. (Remember that von Wright was
apparently so dissatisfied with the lack of alternativeness of the relata,
which caused him to admit also the T>- T-case for PL, that he chose to
interpret it as an F>-F-case 2 .)
The reason why (4) and (5) sound strange, whereas (1)-(3) do not, seems
to be that the standard notion of preference in fact requires this special
kind of alternativeness of the compared states that is claimed by thesis I
above. In order to bring this alternativeness to light, a careful description
of these states is indispensable. If someone professes to prefer a steak to
the beans which are both on the table before him, this seems, at first glance,
to be a clear-cut and unobjectionable case of T» T-preference. But what
is presumably meant by uttering such a preference is not that the state of
the steak being on the table is preferred to the state of the beans being on
the table, but that eating only the steak (or eating the steak first etc.) would
be preferred to eating only the beans (or eating them first), which pairs of
states (events, actions) are, as opposed to the mere co-existence of the
steaks and the beans, obviously pairs of alternatives.
To sum this up: it seems to be suitable for formulating the truth con-
ditions of standard preference assertions to introduce truth-values also for
certain different possible worlds. Since basic and non-basic standard pref-
erence, though sharing the aspect of alternativeness, differ in their com-
plexity, I shall give separate truth-conditions for each. (It will turn out
though that, from the viewpoint of UT, the former are a special case of
the latter.)
The idea that basic standard preference comparisons 'A~B', 'A>-B', or
'A'" B' are preference comparisons between two specific possible worlds
in which A and B, respectively, are true, formally yields somewhat like the
following truth condition:
Cb: <p(A>-B,Wo) = Tiff:
(1) <p«A ,.... -,B) 1\ (B --. -,A),Wo) = T; and
(2) Wa>- Wb
where',....' symbolizes conditional implication, and
Wa=df lW: R a-min (W,Wo) and Wb=df lW: R b-min (W,WO).3
In words, Cb means that a basic preference statement 'A>-B' in the actual
world Wo is true for a person x iff: (1) in x's belief4 (at least) the conditional
implications are true in Wo that if A were true in a certain possible world
Wa (i.e., if Wo changed to Wa), then B would be false in Wa and if B were
true in Wb (i.e., if Wo changed to Wb), then A would be false in Wb; and
(2) x prefers Wa to Wb; Wa and Wb are those possible worlds which bear
the relations R a- min and R b- min of a-minimal and b-minimal difference, re-
spectively, to the actual world Wo, i.e., differ from Wo only insofar as the
assumptions of A being true in Wa and of B being true in Wb require,
explicitly or implicitly, descriptions of Wo which are different from those
without these assumptions. This is the sense that Stalnaker and Thomason
give to the relation of minimal difference between worlds. Their seman tical
analysis of counterfactual conditional statements may also serve as a tool
for understanding the truth conditions of (1) in CbS. For those who prefer,
despite the objections brought forward against it, something like D. Lewis'
understanding of counterfactual conditionals, 6 the second conjunct 'B ,....
-,A' was inserted into condition (1). Approaches according to which the
law of contraposition also holds for counterfactual conditionals would
make this conjunct superfluous. In the special cases 'A>--,A' or '-,A>-A',
Cb would a fortiori be fulfilled, since logically incompatible states are even
incompatible in all possible worlds. So condition (1) of Cb is only a minimal
condition in that A and B must exclude each other at least by a
subjunctive conditional in the two worlds compared.
Note that a reformulation of C b in the object-language could not serve
306 RAINER W. TRAPP
as a conditioned contextual definition of the basic preference relation '>-'

itself, because, so understood, it would be circular.
The propositions substituting A and B in concrete preference statements
are to be taken as including the same time-constants in order to account
for the requirement that A and B, according to the condition of alterna-
tiveness, at least not be true at the same time.
The characteristic of Cb , in contradistinction to other approaches, is that
a basic standard preference comparison 'A >- B' is not considered as a com-
parison between two sets of worlds but only, as indicated by the descrip-
tion operator '1' in the definiens of Wa and W b, between exactly two
worlds.
This characteristic seems to me both intuitively adequate and of formal
advantage. For, firstly, there is among other senses a basic sense of a
person x' preferring A over B which does not ground on x' considering
several alternative possible worlds, in which A and B, respectively, may be
true (embedded), and which are only more or less likely to become real,
but according to which 'A>-B' is true in virtue of x' fancying just two
worlds Wa for A and Wb for B, assuming (rightly or not) full knowledge
of Wa's and Wb'S utility-relevant aspects and judging Wa's occurrence to
be better than Wb'S occurrence.
This assumption of the ideal knowledge of what will happen with cer-
tainty, if A and B, respectively, are chosen, might prompt one to call basic
preference also ideal preference in contradistinction to the more realistic
non-basic preference which copes with the fact that many preferences -
such as, e.g., those between actions under risk - do not only draw on mere
utility comparisons of worlds assumed to become real for sure, but on
probability-utility-considerations concerning several possible A-worlds and
B-worlds, respectively. So non-basic preference is also suitably termed
probabilistic preference. Secondly, considering only one world for each
basic preference relatum (instead of a set Sa of A-worlds and a set Sb of
B-worlds) spares one the problem of deciding which worlds of Sa and Sb
are to be compared. Is 'A>-B' true iff all A-worlds are better than all B-
worlds, or if some A-worlds are better than some B-worlds etc? Each of
these two, and further, possibilities have unwelcome consequences, the
first one, for example, that many propositions are incomparable and, what
is even worse, that 'A>-A' would not even be reflexive'. (Once we get to
non-basic preference, where we do not have to make only pure utility com-
parisons between the A-worlds A-Wi and B-worlds B-wj in Sa and Sb, we
can fall back upon UT-techniques of calculating expected utilities of dis-
junctive events y A-Wi and y B-wj (with all A-Wi E Sa and all B-Wj E Sb)
• J
and thus get along with the problem of how to consider the A-world and
B-world utilities in determining whether 'A>-B' is true. But this is some-
thing for section II.
One might ask here why I do not ground basic preference on this special
kind of truth-conditions on which von Wright based his particular notions
of preference in his (1963) and especially (1972)8. These notions are some-
where in-between basic and non-basic preference. With basic preference
they share the property of being non-probabilistic in that the
probabilities of the A- and B-worlds do not appear in the truth-conditions
for 'A>-B'. With non-basic preference they share the trait of being holistic
and non-ideal in the sense that A and B are embedded into several possible
A- and B-worlds, and that the truth of'A>-B' is based on the betterness-
relation between certain elements of these sets. The core of the answer to
this question is (beside principal objections against the too crude propo-
sitional framework of von Wright's and other systems ofPL to be exposed
later) already contained in the remarks in the last paragraph: basic pref-
erence as ideal preference needs no consideration of several worlds, and
non-basic preference should from the beginning be grounded on one of
the much more powerful versions of probabilistic semantics.
Moreover, von Wright's truth conditions also have some internal prob-
lems. To further justify my preference for a probabilistic PL-holism let me
briefly sketch them: Von Wright's holistic PL considers also the circum-
stances under which A, Bin 'A>-B' obtain by viewing A and B, respec-
tively, as embedded into certain possible worlds construed in a simple way
from a so-called state-space. A state-space Sn is a finite set of n logically
independent elementary states including those contained in A, B. From
each Sn one can construe 2n possible worlds representable by Carnapian
state-descriptions. Let A and B together contain m elements of Sn. The
remaining n-m elements can obtain or not in i possible combinations Ci
(1 ~ i ~ 2n-m). Each Ci is called a circumstance for A and B, respectively.
On this basis von Wright defines two kinds of circumstantial preference:
D 1 : A is preferred to B under the circumstances Ci iff every Ci-world
which is also an A-world but not a B-world is preferred to every Ci-world
which is also a B-world but not an A-world.
308 RAINER W. TRAPP
O 2 : A is preferred to B under Cj iff some Cj-world which is also an A-

world is preferred to some Cj-world which is also a B-world, and no Cj-
world which is a B-world is preferred to any Crworld which is an A-world.
Despite its advantage of trying (though not quite successfully) to account

for the requirement of alternativeness of A and B in comparing only
A 1\ -,B-worlds with B 1\ -,A-worlds and thus being more acceptable than
other conceptions of comparisons of two sets of worlds 9, 0 1 seems to me
counterintuitive and unacceptable. For there remain hardly any two states
that are comparable according to it.
Let S4 be the state-space {p,q,r,s}. D1 then makes a basic comparison
such as, e.g., 'p,;>q' is true iff the 4 conjuncts are true which are obtained
by asserting the preference of each of the 4 p 1\ - , q-worlds to its respective
4 q 1\ -,p-counterpart-worlds with identical Cj. One such conjunct is
'p 1\ - , q 1\ r 1\ - , s.':, q 1\ - , P 1\ r 1\ -,s'. A person's weak preference
for having the cholera (= p) to having cancer (= q) then would not be true,
since one presumably would not also weakly prefer, among other things,
the option of having the cholera (=p) and not having cancer (=-,q) and
there being a quick remedy against cancer (= r) and there being no remedy
against cholera (=-,s) to the option of having cancer (= q) and not having
the cholera (=-,p) and there being a quick remedy against cancer (= r)
and there being no remedy against the cholera (=-,s). Likewise, a con-
junct can be found that also makes q.':,p false (e.g.: q 1\ - , P 1\ -, r 1\ s
.':, P 1\ - , q 1\ - , r 1\ s) so that having cancer and having the cholera
cannot be compared relative to S4. It would be ad hoc and unacceptable
to exclude states as rand s f];Om state spaces for 'p,;>q' or 'q,;>p', since r
and s doubtless are highly utility relevant for the preference. Along these
lines, almost any pair of true (according to actual practice and also to
criterion C,,) preference statements 'p,;>q' and 'q,;>p' can be proved to be
false relative to some plausible Sn, so that for almost no two world states
p, q would there be a true preference assertion according to von Wright's
0 1•
Against O 2 one cannot object in this way. So, if! had to decide between
0 1 and D 2 , the choice would be clear. But I do not want to have just these
two alternatives, for if in defining the truth of' A';> B' one considers at all,
as holistic PL does, all A-worlds and all B-worlds, I think it more adequate
to consider also the probabilities of these worlds and to base the truth of
'A»B', e.g., on the fact that the A-world lottery ticket (i.e., the lottery
with all A-worlds as possible outcomes) has a higher expected utility than
the B-world lottery ticket. But this too is a story for Section II.
Let me nevertheless draw attention here to an obvious deficiency of Cb :
C b alone, though defining the truth conditions of atomic 10 basic prefer-
ence statements, does not suffice to define also the conditions for the logical
truth of truth-functional complexes of preference statements, notably im-
plications, because C b does not systematically define in what way the truth
conditions for one preference statement are automatically also those for
a different preference statement. Without a semantical basis determining
what characterises a preference-tautology, there is, however, no satisfac-
tory PL which is more than a logical system, the acceptability of whose
axioms rules and theorems is judged by intuition. The holistic conception
even provides the possibility for several versions of PL-semantics (which
will all turn out to have their difficulties); and I see no possibility of an
acceptable semantics without it. How could we do without a given state-
space that allows us to define possible worlds in which to embed preference
relata represented by only truth-functional language? So the notion of
state-space is highly useful for a satisfactory semantical basis of such
PL-systems. The question within the linguistic framework of these systems
then is how best to use this notion to semantically characterise preference
tautologies. Db for example, is not even appropriate for externally atomic
preference statements. He who nonetheless chooses it should indicate a
selection criterion for 'good' state-spaces which makes it possible to dis-
tinguish them from bad ones that destroy - as the state-space in the can-
cer-cholera-example - the comparability of world states which one wishes
to be comparable in any acceptable PL.
A first suggestion of a D 1 -adherent for such a selection criterion might
be the radical one of allowing only minimal state-spaces for each 'A»B',
'minimal' meaning that they only contain the union of elementary states
in A and B. But this, though avoiding our counterexample and all struc-
turally similar ones, is unacceptable because the circumstances which bear
on the truth or falsehood of 'A»B' for a person generally comprise more
than only the elementary states composing A and B. Any reasonable hol-
istic conception of PL should allow state-spaces larger than minimal.
Coming back to C b for another moment one might wonder why Wa and
Wb cannot be allowed to be more than minimally different from Woo The
3lO RAINER W. TRAPP
answer is that if they were, Cb would no longer be adequate: he who prefers

a possible world in which Kohl is chancellor of the F.R.G. (and therefore
Vogel is not) to a different possible world in which Vogel is chancellor
(and therefore Kohl is not), will most likely not do so if these possible
worlds are not, as for everything else, like the actual world but, say, also
differ insofar from it as Kohl is assumed to be the leader of the German
Communist party, everything else being as in the actual world. Likewise
even he who has a notorious preference for whisky over buttermilk may
change his preference if the possible worlds compared are also allowed to
differ from the actual world in that in them human physical nature is
assumed to be such that a little sip of whisky is enough to make a man
fall dead to the ground. Furthermore, more than minimal changes being
conceded, two states which in all minimally different worlds cannot coexist
may very well coexist. E.g., Kohl being chancellor and Vogel being chan-
cellor are alternatives only in such possible worlds in which the constitu-
tion of the F.R.G. is not changed such that two chancellors can simulta-
neously be heads of government. Usually, two logically compatible states
are alternatives to each other only relative to certain institutions, regula-
tions, causal laws, of whatever kind, etc., that induce counter/actual con-
ditionals of the kind 'If x were chancellor then y (x#y) would not be
chancellor'. That is why we need the semantical conditions for at least
counterfactual implication in the first conjunct of Cb • Of course, also log-
ically incompatible states, as a special subclass of alternative states -logical
incompatibility being the strongest form of alternativeness - can be put
into a preference relation. Material implication in the first conjunct of Cb
would not do. If already the actual falsity in Wo of at least one of two
propositions A and Bin Wo were sufficient for taking them as alternatives
and as suitable relata of a preference relation then the state of a Republican
being president (in 1983) could be preferred to New York not suffering
from criminality at the same time. What excludes them, despite the fact of
the TIF-combination of these propositions in W o, from being alternative
options in a choice situation, and thus suitable relata of a standard pref-
erence relation, and what explains the strangeness of such a preference
assertion, is the possibility of the coexistence of these states in possible
worlds only a-minimally or b-minimally different from Woo There is just
no reason to assume the conditionals 'If a Republican were president then
N.Y. would suffer from criminality' and 'If N.Y. did not suffer from crimi-
nality then no Republican would be president'.
Note, finally, that - contrary to what, e.g., Kanger or Saito syntactically

admit l l even for strict preference - our characterisation rules out the pos-
sibility of tautological and contradictory states being relata of any prefer-
ence relation. For, setting aside the lack of sense such preference-assertions
would make in ordinary language - logicians are often right in waiving
scruples about the ordinary-language acceptability of their results - neither
tautologies nor contradictions fulfill the conditions given in Cb•
He who - despite the arguments for alternativeness of states as a pre-
condition of the truth of a standard preference statement - insists on the
preferability of a state to a coexisting one must base his statement on a
totally different analysis of the notion of preference. If, for example,
while sitting at the dinner-table and enjoying a meal, one learns from the
radio that one's favourite football club has just won a match and utters
'I prefer this victory even to this excellent dinner!', or if in our example
some pages above one insists on preferring the steak to the beans, rejecting
all my above-mentioned attempts at interpreting this preference in an al-
ternativeness-conserving way, one must accept something like the follow-
ing as an analysis of the meaning of preference:
x»y iff (in a situation S in which x and yare co-experienced) the
degree of satisfaction derived from experiencing x contributes
more to the whole satisfaction in S than the degree of satis-
faction derived from experiencing y.
Certainly two (or more) coexisting and co-experienced states can contrib-
ute differently to a certain degree of satisfaction or dissatisfaction a person
feels. This analysis would, however, at least as I think people normally use
'preference statements', not hit the meaning of what I call standard pref-
. erence statements. Contributing more to a certain amount of felt satisfac-
tion (however this might be measured) than some other coexisting factor
does not seem to be the essence of preference. Laying the ground for this
kind of utility aggregation in a person P at a time t is rather a task for
multi-attribute-utility-theory, which, e.g., in its standard additive models,
surely would not consider the 'attributes' of a state inducing the additive
'attribute'-utilities as possible relata of preference relations. Also - on our
analysis - aspects or attributes of a state or of different states can be
brought into a standard preference relation; but to do this they first have
to be formulated as some kind of alternatives.
312 RAINER W. TRAPP
Last but not least: all important applications of preference relations in

the vast areas of utility measurement, individual decision theory, and social
choice theory are, as far as I can see, in silent accord with our condition
of alternativeness. Wherever weak preference orders are taken as the start-
ing point for constructing individual utili,ty functions unique up to certain
transformations and aggregating them in some way or other to get social
choice functions, these orders are defined over sets of alternatives of some
kind, which sets, varying with the approaches chosen, may contain prop-
ositions, states, events, social situations, conditional decisions, conse-
quences of actions or lotteries of such entities. Also the possible relevant
actions themselves in a decision situation, put pairwise into (non-basic)
preference relations according to their (expected) utility as determined by
the respective decision models, have to be alternatives. The notion of al-
ternativeness itself was more or less presupposed as sufficiently clear in all
these applications. Also L. Bergstrom 12 - who to my knowledge was the
first and, apart from some discussants of his theses, up to now the only
one to investigate the notion of alternativeness (restricted, however, to
pairs of actions) - made incompatibility a central requirement of alterna-
tiveness. Though he did not make use of possible - worlds - semantica1
talk in order to specify more closely for what classes of worlds at least
incompatibility has to hold, I yet think that the foregoing considerations
can be said to supplement and extend his analysis.
II
At first sight, the thesis of alternativeness of standard preference relata

seems to be blatantly false. For in many preference comparisons such as
'x"y~x', 'x~xvy', 'x~x=y' ('=' standing for the exclusive disjunc-
tion 'either or'), both relata are compatible in all possible worlds, or one
relatum even logically implies the other. Since thesis I is not meant to
recommend to reduce the expressive power of PL in such a way that pref-
erence comparisons as the ones given are not longer allowed, it follows
that what such comparisons actually want to convey must be transcribable
into a different notation which both complies with the requirement of al-
ternativeness and conserves the information of the second relatum having
less utility than the first one. (To avoid misunderstanding here, as in the
sequel when certain PL-formulae will be checked by the criterion of UT:
I do not at all intend to tum things upside down by giving priority to the
notion of utility over the notion of preference. But that utility differences
of two options are based among other things on the preference for one
over the other does not invalidate the thesis that once logical relations
between preference assertions are the topic (as in PL), UT turns out to be
an efficient criterion in judging the acceptability of these relations and the
semantics that justifies them).
An analysis of PL-formulae with truth-functionally compatible relata
along the lines of UT has the following properties (before coming to our
examples let us see this for a most general case): if the relation ';::' is so
axiomatically characterised that, for all relata A, B and quite apart from
their embedding into state-space generated worlds, the equivalence holds
A;::B iff u(A) ~ u(B) and therefore u(A)·I + u(B)-O ~ u(B)·I + u(A)·0,13
then one can view every weak preference for A over B as a preference for
a certain degenerate lottery ticket (A,I; B,O), i.e., 'either win A with prob-
ability p= 1 or win B with p=O' over a different one (B,I; A,O). Thus
viewed 'A;::B' boils down to preferring A and not-B to B and not-A, not-
withstanding the dispute ~J>L whether or not 'A;::B +-+ A /\ -,B ;::
B /\ -,A' and the formulae implied by it are acceptable for a system ofPL,
and on which semantical basis they are, and on which not. There is no
more possibility for dispute about the incompatibility or mutual exclu-
siveness 14 of A and B, at least in the worlds Wi in which they may occur.
Bringing in additive probabilities in this (admittedly somewhat artificial
way) already pre-supposes it.
Let us apply this idea to a special case of 'A;::B' such as 'x;::x~y'. It
is interpreted as being true for a person a iff that person prefers a degener-
ate lottery L1 with the outcomes either of getting x with probability lor,
with probability 0, the right to participate in a non-degenerate lottery
'either x with p or y with l-p' - p (O~p~ 1) being unknown to person a
- over a second degenerate lottery L2 with the outcomes of having, with
probability 1, the right to participate in that latter non-degenerate lottery
or, with probability 0, of getting x. Expressing this in terms ofUT, 'x>-x-
~y' would be true for a person a iff a would accept that
u(x)-I + u(x~y)·O ~ u(x~y)-1 + u(x)·O

i.e., since u(x~y) is interpreted as the expected utility of a's right to par-
ticipate in the lottery (x,p; y,I- p), iff a would accept that
314 RAINER W. TRAPP
u(x)·l + [U(X)·p + U(y)(1-p)]·O~ [U(X)-p + U(y)(1-p)}1 +

u(x)-O.
As formulated, this is in substance nothing but a utility theoretic tran-

scription of the idea, advanced in some systems ofPL, of conceiving 'A~B'
always as 'A 1\ -,B~B 1\ -,A', and thus of granting from the first some-
thing like alternativeness in its strongest form to the relata of all preference
relations. But the UT-transcription is superior to the PL-idea in an essen-
tial respect for, in contradistinction to it, it is generally applicable. In all
cases where one preference relatum follows logically from the other one
the PL-procedure fails. A preference statement (a) 'x 1\ y~ V y' cannot
reasonably be considered as equivalent to (b) 'x 1\ Y 1\ - , (x V y) ~ (x
v y) 1\ - , (x 1\ y)', i.e., as a weak preference for a contradictory state
over a non-contradictory one. Analysed as above in terms ofUT statement
(b), on the other hand, makes perfect sense. There is no contradiction
involved, but it just expresses that the utility of getting the sure option
'x I\y' but not the lottery ticket L1 (x 1\ y, P1; x 1\ - , y, P2; -, x 1\ y, I
- P1 - P2) is greater than, or equal to, the utility of getting L1 but not
the sure option. But also in cases in which no relatum is logically implied
by the other one the PL-procedure above is no generally acceptable way
of granting alternativeness. It is again the principle of intet:exchangeability
of equivalents that is at the root of the trouble. For it enforces that, e.g.,
(1) 'x 1\ - , (x $ y) >-- (x $ y) 1\ - , x' amounts to (2) 'x 1\ y >-- -, X 1\
y', whereas our UT-transcriptions do not permit a substitution correspon-
ding to this one.
This trait of UT once again turns out to be an advantage because, for
many x,y, it is grossly counterintuitive to hold that whoever has the pref-
erence expressed in (1) thereby does nothing else but having the preference
expressed in (2). If coming to a strange land I am suddenly apprehended
by people there and offered the choice between alternative 1 of either being
their king from time to on, but not having the right to participate in a
lottery with either my being king from to on or my being executed at to,
as possible outcomes, and alternative 2 of participating in the lottery but
not being their king, and I choose alternative 1, I yet would strongly deny
that this amounts to my strong preference for being king from to on and
being executed at to over my not being king but also being executed. I
would rather consider dying as their king, or as some beggar, as equally
unpleasant. (Let me mention already here that also on Rescher's semantics
in (1966) for one system of PL statements (1) and (2) would have exactly
the same truth conditions.) Holding the strong preference in (1) and
analysing this along the lines of UT sketched above makes, on the other
hand, perfect sense without committing one to the UT-inequality (2')
u(x "y» u(-,x "y) corresponding to (2) x "y>--,x "y. The UT-analysis
of preference (1) only commits one to u(x»u(y), which is doubtless ac-
ceptable if applied to my example, for I indeed prefer being king to being
executed, ceteris paribus.
One might suggest a radical way out of these difficulties with the exten-
sionality of propositional logic by restricting the interexchangeability of
equivalents in the relata of PL-formula. This restriction of extensionality
would have to be more severe than the one conceded by von Wright l5 •
For von Wright only restricts interexchangeability if the logically equiva-
lent formulae substituted contain propositional variables which the ones
replaced do not contain. This, however, is not the case in substituting
'x "y' for 'X" -'(x~y)' and '-,x "y' for '(x~y)" -,x'.
I would not recommend this radical solution: for there would arise dif-
ficulties elsewhere and, what counts most, there would not be much left
of a manageable logic at all. On the other hand, I do not know how to
heal, within the boundaries of propositional language, all these evils of PL
in a simple and straightforward way. (Some even more striking evils will
follow still later in this section.)
One essential point in transcribing PL formula into the different nota-
tion of UT is to show that truth-functionally compatible preference relata
turn out to fulfill my initial requirement of altemativeness; so the UT-
transcriptions of such PL-formulae must, in addition to their removing
difficulties of PL due to its extensionality, be interpretable as representing
the utility differences of options, which are always incompatible in a non-
empty set of worlds Wi including Woo For the UT-transcriptions of 'A >-B'
conceived as 'A" -,B>-B" -,A', I have just shown this. But even if this
PL-device of achieving incompatibility is not made use of, the UT-tran-
scriptions of 'A>-B' involve altemativeness of the options whose utilities
are compared. Let us first illustrate this by an example: whereas a PL-
formula such as (a) 'x~x~y' certainly has compatible relata, its UT-coun-
terpart (a') 'u(x);;::u(x)-p + u(y)(1-p)' (for allp in [0,1])16, interpreted as
a weak preference for a sure option over a lottery with unkown p, yields
even logical incompatibility of the relata. Logical incompatibility holds -
316 RAINER W. TRAPP
though 'x' and 'x~y' may be true in the same possible world - since there
is no possible world in which x is figured to be strictly determined to occur,
and yet at the same time not to be strictly determined to occur but, say,
because of the existence of a random generator, to occur with the variable
probability P,y occurring with 1-p instead of x. So two preference relata
that before our UT-analysis seemed to be compatible in all Wi turn out
in the end to be logically incompatible, because richer descriptions of the
relata are used. To say this leaves loopholes (for, literally, our UT-tran-
scriptions yield only certain sums of products, which, as such are only
functional terms, but not propositions that could be incompatible): in less
loose terms incompatibility is claimed for the descriptions of all pairs of
possible worlds Wj and W k the (expected) utility of which is indicated by
the numbers denoted by these functional terms.
I think this result is in accordance with our intuitive judgment of sure
options and lotteries of the given structure to be alternatives: being con-
fronted with the choice between the sure option of getting a gold coin
(= x) or a lottery ticket (with unknown p) - whose possible outcomes are
getting a gold coin or a silver coin (= y) - one may interpret this as being
confronted with the choice between a possible world in which there is no
random generator (or at least it is not used to determine whether x or y
occurs) and a different possible world incompatible with the first one in
which there is one (or made use of one).
Similarly one may cope with those cases of PL-expressions the relata of
which, being not only logically compatible but also allowing a one-direc-
tional logical implication, seem to be unredeemably inappropriate for
any PL accepting the principle of alternativeness of preference relata. For
the abovementioned standard way of achieving alternativeness cannot be
followed since it leads to at least one contradictory relatum. Take, e.g.,
'x I\YZ:x': interpreted (1) as 'x and yare weakly preferred to only x', i.e.,
as 'x I\YZ:x 1\ -,y' on a UT-analysis, it turns out to be a weak preference
for one sure option over a different one, logically incompatible with the
first; interpreted in a logically more natural way as (2) 'x 1\ y Z:
X 1\ (y~-,y)', it turns out to be a weak preference for the sure option
'x I\Y' over the lottery either 'x I\y' with p or 'x 1\ -,y' with I-p'. Take
another expression admissible in all systems ofPL: 'x l\yZ:xv y'. This may
also be interpreted as a weak preference for a sure option over a lottery,
yielding in UT-transcriptions:
u(xl\y) ~ U(XI\Y)·Pl + U(XI\-'Y)·PZ + u(-,xl\yHI-PI-PZ)17.

In short, once we get two logically incompatible sure options, or a sure
option and a lottery, or two different lotteries by UT-transcriptions of
seemingly compatible preference relata, we obtain even more than the al-
ternativeness needed: for - once lotteries are involved - actually possible
worlds with and without random generators or with different random gen-
erators are compared.
One may take this contrast between the truth-functional compatibility
of the state descriptions 'x' and 'x+t+y', or 'x I\y' and 'y', etc., and the
alternativeness of the possible worlds corresponding to the UT-reformu-
lations of these descriptions, as another indicator - beside unwelcome re-
sults stemming from the interexchangeability of logically equivalent relata
- that conceiving the relata of preference relations as mere truth-functional
schemata does not suffice to adequately represent what standard preference
statements actually intend to convey. What we need - to keep pace with UT
- is a less coarse formal representation of the relata of preference relations
than propositional logic offers.
Quite apart from this principle dissatisfaction with certain implications
of the extensionality of PL (which prompted me to deny general applica-
bility to bringing about alternativeness of the relata by conceiving 'A>-B'
as 'A 1\ -,B>-B 1\ -,A'), there are also different arguments against this
principle which had already been suggested by Hallden and von Wright
in the early days of pUs. Hansson, for example, argues that it leads to
counterintuitive consequences and therefore should be rejected as an ax-
iom for PL 19 . For, from
C1: x>-y +-+ x 1\ -,y >- -,x I\Y,
by substituting '-,y' and '-,x' for 'x' and 'y', respectively, thus getting
-,y>--,x +-+ -,y 1\ -,-,X >- -,-,y 1\ -,X, i.e.,
-,y>--,x +-+ x 1\ -,y >- y 1\ -,X,
one can easily deduce

C z: x>-y +-+ -,y>--,x.
C z, however, seems unacceptable to Hansson as a general theorem of PL

318 RAINER W. TRAPP
for the reason the following example makes clear: Suppose someone ( = S)
has bought some lottery tickets with two possible prizes P l and P 2 , the
first being of higher worth for S than the second. Then S reasonably will
prefer x (= winning the first prize Pl) to y (= winning some prize). Ac-
cording to C2 this would mean that S also prefers not winning any prize
to not winning Pl' Such a choice, however, would be irrational, for not
winning P l still leaves open the possibility of winning P 2 whereas
winning no prize at all excludes this. So, for Hansson, since C l leads to
C 2, both formulae are unacceptable.
Saito,20 referring to Hansson's example, points to the fact that in it y
logically implies x. Admitting that C l would indeed not be acceptable, for
there would be no sense in preferring x 1\ - , Y to -, X 1\ Y if x logically im-
plies y or vice versa, Saito tries to save C l and C 2 by requiring for both
that x and y not imply each other in either direction. I hold this restriction
to be too high a price for saving C l and C 2, for it would cut off formulae
from PL that make perfect sense if analysed in the light of UT, in which
latter tool I have more confidence than in PL. (The formulae of this kind
analysed above illustrate this.)
But, in arguing for or against certain PL-formulae such as C l or C 2, the
problem is no longer only to transcribe externally atomic PL-formulae (see
Note 10) into the richer language of UT in order to show that relata can
be regarded as alternatives, which truth-functionally they are not. The
problem is rather, what can count as a PL-tautology. Once logical relations
between PL-formulae are the subject, however, there will be wide agree-
ment that it is more promising to take up the question of an adequate
semantics (fixing the truth-conditions also for all externally molecular for-
mulae of a PL-system) than to judge case by case which of the formulae
a certain system contains as axioms or theorems are welcome and which
are not. The history of other philosophical logics abounds with evidence
for this. This does not mean that we can do away with all intuition, for
our willingness to accept a semantics for PL will depend on our conception
of the nature of preference. What is abandoned is only the casuistic ap-
plication of intuition to the formulae a PL-system puts out.
As I said before, criterion Cb of section I alone is not sufficient as a
semantical groundwork for PL. It does not permit one to say anything
about the truth-conditions of PL-implications. So something richer is
wanted in addition.
A first adequacy condition for such a semantical criterion is that like

Cb , it should account for the fact that generic states A,B in 'A>-B' are-
except for extreme deonto10gists - not intrinsically good whatever the cir-
cumstances, and therefore the possible consequences. States should be com-
pared while embedded into worlds that also contain utility-relevant con-
sequences of the states.
Moreover, in many cases of preference considerations a state A is not
only thought to be embedded into only one A-world held to be a sure
prospect - as basic preference assumes - but into several possible ones
depending on the assumptions of what co-obtains with A. So the second
condition suggests that we should consider non-basic preference compar-
isons between states as comparisons between certain sets of worlds in
which these states obtain. The point for non-basic PL-semantics, then, is
how, precisely, to compare these sets. Von Wright, as is sketched in section
I, restricts his holism to regarding circumstances in which the relata may
occur disregarding the probabilities of the worlds thus obtained. Conse-
quently, he does not base his truth conditions for PL on the expected
utilities or some similarly interpreted numerical value associated with each
set, but just on preference comparisons of certain specified ordered pairs
of worlds from these sets.
A more sophisticated version of PL-holism, according to which the se-
mantics for non-basic preference assertions should draw on UT-insights
and conceive preference comparisons as lottery ticket comparisons with
certain possible worlds as outcomes, has the advantage - apart from 21
objections raised against von Wright's ordinal holism other than those
made in section I - of providing straightforward truth-conditions for
PL-implications, and thus of defining the notion of PL-tautology. So UT-
guidance should be a third adequacy condition for PL-semantics.
One might suggest several variations of such a UT-guided semantics
(UTS). They will provide a framework for discussing Hansson's and other
arguments against the aforementioned theorems C 1 and C 2 as well as shed
some more light on my principal scepticism about the adequacy of mere
truth-functional structuring of preference relata. A version of UTS which
comes close to what Rescher suggests as a truth-condition for one kind of
preference relation appears, at first sight, highly attractive. 22
UTS 1 : 'A~B' is true iff the arithmetical mean of the utility values u..(Wj)
of all the n possible A-worlds Wj (i.e., worlds in which A is true) is equal
320 RAINER W. TRAPP
or higher than the arithmetical mean of the utility values Ub( w,,) of all the
m possible B-worlds, or in short:
1 n 1 m
A'/:,B iff - L u..(Wj) ~ - }yb(W,,).
n j=l m "=1
As in von Wright's holistic PL, the n A-worlds and m B-worlds are ele-
ments of a set of 2t possible worlds represented by state descriptions and
generated from a state-space S of t logically independent elementary states.
Whether A and B are atomic or molecular compounds of together i el-
ementary states Xl> ••• ,Xi, S must contain at least these i (i~2) elements.
State spaces Si-min that, relative to 'A'/:,B', do not contain more than these
i elements are called minimal. If, as in von Wright's holistic PL in (1972),
the A-worlds and B-worlds are to be embedded into 2r further possible
circumstances composed of the circumstance space C = {Xi+1, ••• ,Xi+r},
S for 'A»B' is non-minimal but the union of Si-min and C. Though UTS 1
is applicable for minimal and non-minimal state-spaces it is, of course,
more realistic to work with the latter. The range for the value of m, n in
UTS 1 is 1 '5.n,m'5.2t-l, because we exclude the extreme possibilities of A
and B, respectively, being true in no world at all or in all 2t worlds. As I
said in section I, tautologies as well as contradictions are not appropriate
preference relata since - apart from the lack of sense this would make
prima facie - they cannot stand in the relation of altemativeness to what-
ever second relatum.
UTS 1 easily permits us to define the notion of PL-entailment: a PL-
formulafl PL-entails a PL-formulaf2 iff the inequation gl> corresponding
tofl according to UTS 1 - for all possible assignments of utilities to worlds
-logically entails the inequation g2 corresponding to f2. 'Preference-tau-
tology' can thus be defined in terms of simple arithmetical truth. UTS 1
presupposes PL to be at least semi-extensional in the sense of von Wright
(see Note 15), for only the utility values of all A-worlds and B-worlds
respectively count; and such worlds are extensional entities representable
by a potentially infinite number of logically equivalent formulae contain-
ing the same set of basic propositions. Since the weighting factors! and
n
!m have the formal properties of probability values one may interpret the
two sums in UTS 1 as the expected utilities of two lottery tickets with all
the A-worlds and B-worlds as equiprobable possible outcomes. So UTS 1
fulfills all three adequacy conditions.
Both sides of C 1 and C 2 yield the same inequations for every state-space
from the minimal one {x,y} up: so C 1 and C 2 are valid according to UTS 1 •
This, however, no longer holds generally if molecular relata are taken for
A,B instead of the atomic relata x,y. Since Hansson rejects C 1 and C 2 even
for atomic relata, he must base his PL on a different semantical principle
than UTS 1 .
By considering the internal structure of x and y in Hansson's example,
one might also interpret it as an argument against a molecular rather than
atomic application of 'contraposition' for preferences (as von Wright calls
PL-formulae of the structure of C 1). Hansson's point is that, though win-
ning the first prize P 1 (=x) is better than winning some prize (=y), not
winning any prize is worse than not winning P 1 , which in a natural analy-
sis, making y molecular, yields his denying the truth of
(a) P 1 ';:-P 1 v P 2 - ,P1 A ,P2 ';:-,P 1 •
If one construes Hansson's example thus, his system of PL would, at least
so far as the exclusion of C 2 from PL-tautologies is concerned, be in accord
with UTSt, for (a) is, even for the minimal statespace {Pt,P 2 }, no PL-
tautology according to UTS 1 • Applying UTS 1 to the utilities u(w;)=u; of
the 4 possible worlds {W1: P 1 AP 2 ; W2: P 1 A,P2 ; W3: ,P1 AP2 ; W4:
,P1 A ,P2 } we get for (a)
H U1 + U2) > 1(U1 + U2 + U3) - U4 > HU3 + U4)
i.e.,
and this is not generally true but only holds for U4 = t(U1 + U2)' SO, on
the basis of UTSt, PL-contraposition is not valid for molecular relata A,B
and also C 1 would fail for unrestricted molecular A,B. In short, UTS 1
does not permit to get alternativeness into PL-formulae with compatible
relata by means of C 1.
Chisholm and Sosa present the following (often quoted) argument
against the validity of C 2 even for atomic relata 23 . Starting from the ax-
iological premise that only the states of pleasure of living beings are in-
322 RAINER W. TRAPP
trinsically valuable, they argue against PL-contraposition that although

the state consisting of there being happy egrets (h) is better than one that
consists of there being stones (s), that state that consists of there being no
stones (-,s) is no better or worse than one in which there are no happy
egrets (-,h). Even if one accepts their axiological basis this argument does
not seem convincing to me, for it totally neglects UT-rationality. Why,
even if only pleasure counts, should '-,s' be no better, or worse, than
'-,h'? For, if there are no stones, there mayor may not be happy egrets.
If the probability for there being some of the latter is greater than zero,
then this adds expected utility to the zero utility of there being no stones.
If, on the other hand, there are no happy egrets, the probability of there
also being stones may even be 1, yet it does not add - on the Chisholm/
Sosa-value-criterion - any expected utility to the (equally non-existing)
utility of there being no happy egrets. So '-,s' has more expected utility
for p(h) >0 than '-,h'. To put this less informally in terms ofUT plus the
C./S. value-criterion: the minimal state space {h,s} for 'h';:-s';:--,s';:--,h'
yields {Wl: hAS; W2: hA-'S; W3: -,hAS; W4: -,hA-'S}. The C./S. value-
criterion implies the following weak order of utility assignments to these
4 worlds: Ul = U2 > U3 == U4 = o.
That '-,s';:--,h' is true on the basis ofUTS 1 is clear anyway, since C./S.
presuppose 'h';:-s', and t(Ul + U2) > t(Ul + U3) implies 24 t(U2 + U4)
> t(U3 + U4). But, as informally argued above, one may even weaken the
truth conditions for 'h';:-s +-+ -,s';:--1h' considerably by pointing to the
fact that the C./S. utility order, given this implication, even comes out true
for variable probabilities of the h-, S-, non-s, and non-h-worlds, respec-
tively, and not only for equiprobability with p = t.
Let (Pi, 1 -Pi) with 0::;; i::;; 4 be the probability distributions that con-
secutively replace <t,t), in 'Ul·t + U2·t > Ul·t + U3·t implies U2·t
+ U4·t > U3·t + u4·f, where O::;;Pi::;; 1 is allowed except for the restric-
tion that P3 > O. Then 'h';:-s' would even remain true iff
Ul·Pl + U2·(l-Pl) > Ul·P2 + U3·(l-P2)
which because of Ul = U2 boils down to

Ul > Ul·P2 + u3·(I-P2).
For Ul > U3, as the C./S. criterion furthermore prescribes, this is even an
arithmetical truth so that 'h';:-s' comes out as a PL-tautology, under
UTSt, as well as under the weakened form of it I indicated, combined with

the C./S. criterion.
This arithmetical truth trivially implies the truth condition for
'-lS>--,h', namely
U2'P3 + u4·(1-P3) > U3'P4 + u4(1-P4)
since this inequation for U3 = U4 = 0 reduces to U2'P3 > 0 and thus is also
an arithmetical truth, if P3 > 0 and U2 > 0, as presupposed. So C 2 , combined
with the C./S. axiology, even prospers under more liberal truth conditions
than UTS 1 requires. Similarly one may argue against C./S.'s rejection of
C 1 • They object to the equivalence of 'x>-y' and 'x /\ -,y>-y /\ -'x', that
on their value-criterion there are cases in which the latter is true though
the former is false. E.g., it is better that there be stones and happy egrets
(x /\ -,y) than that there be no stones and no happy egrets (y /\ -,x); but
there being stones (x) is no better, or worse, than there being no happy
egrets (y). This latter assertion of indifference once again neglects the fact
that there being stones is equivalent to there being stones and either
happy egrets or not - which, if P > 0 for there being happy egrets, adds
expected utility to there being stones. If there are no happy egrets, how-
ever, there mayor may not be stones. In neither case will there be an
increase in expected utility. (I do not present the argument in formal detail
here.) So also C 1 survives the C./S. axiology, if UTSt, or as above, more
liberal truth conditions along the lines of UT, are accepted for PL 25 •
There is another interesting argument in PL-literature that shows both
(1) what pitfalls lurk not only for intuitions on certain PL-formulae but
also for strictly formal PL-deductions unaided by semantical UT-based
guidance, and also, once again (2), how badly merely truth-functional for-
mulae as preference relata (at least in some cases) fulfill their function of
adequately expressing what actually should be conveyed. Kanger (1980)
considers a rule of inference for PL, which - at least in a weaker version
R 1 ' containing '';;:' instead of '>-' in the consequent - indeed seems highly
plausible for any acceptable system of PL. 26
R 1: If A>-B then A>-B +# B>-B.
By means of this innocent-looking rule he then deduces a contradiction as
follows:
(1) x>-y premise;
324 RAINER W. TRAPP
(2) x>-x~y by Rl from (1);

(3) x~(x~y»x~y by Rl from (2);
(4) y>x~y from (3), since x~(x~y)-y;
(5) x~y>y by Rl from (1);
(6) -'(x~y>y) from (4), since '>-' is asymmetric.
So one must conclude that Rl is unacceptable, or that 'x>-y' is always

false (which would be absurd and destroy any PL before it has come out
of the shell), or that interchangeability of equivalents (as in line 4) is in-
admissible for PL. Kanger derives a contradiction also from the premise
'x>-y>-,y>-,x' and a rule which is just the weaker version R l ' of Rl
considered above. Since the core of the difficulty is essentially the same as
in the deduction given, and the second deduction is much longer, I analyse
only the first one here 27 •
A UT-analysis of these rules at once reveals why Rl and especially R l '
seem so attractive 28 . For if u(A»u(B) then in fact
u(A)~u(A~B)~u(B). As u(A~B) in UT is nothing but the expected
utility of a lottery ticket (A,p; B,I-p), this implication is even undeniable
- because if u(A»u(B) then u(A)~u(A)·p + u(B)(1-p)~u(B) is an ar-
ithmetical truth for all pin [0,1]. So it is in fact rational, if one strongly
prefers an option A to an option B, to weakly prefer also A to the right
of participation in any lottery which yields A with p and B with 1 - p, no
matter what value p might have - and to weakly prefer this lottery ticket
to B. (If p(A»O then Rl is even as plausible as R l ' is for all pin [0,1).)
At the same time this UT-transcription ofR l ' does not only corroborate
the plausibility of R l ' but also indicates a restriction of its application.
For in adding up probabilities of options one must make sure that they
are mutually exclusive. Whereas in line (2) of Kanger's deduction this re-
quirement is not violated - so that line (2) can be UT-transcribed - line
(3) no ,longer can, for 'x' and 'x~y' in 'x~(x~y)' for logical reasons
cannot be considered as describing mutually exclusive options. (This holds
in any case so long as PL describes these two options in mere truthfunc-
tionallanguage). So there is no plausibility in applying either of the rules
as done in line (3). More generally: whenever the options on both sides of
'~' are, for logical reasons, not mutually exclusive, Rl and R l ' are not
applicable. So the contradiction only arises because of an illegitimate use
of Rb as a UT-analysis reveals. If, however, an adherent of Rl and R t '
rejected my specific UT-reasons for considering the rules plausible and

furthermore permitted their unrestricted application, I would on the spot
withdraw my judgment that they are plausible!
Secondly this example once more offers an occasion to point to the
deficiency of mere truth-functional language within the preference relata.
For, if a more subtle language were used to express what line (3) of Kan-
ger's deduction is intended to convey, even this line would be in accord
with UT-considerations, and moreover, 'x' and 'X+i+y' would then be
genuine alternatives. What is line (3) actually intended to convey? I think
it is (under premise (1» the perfectly sound preference for a structurally
complex option lover a somewhat less complex option 2. Option 1 is: you
get the right to participate in a gamble with either the better outcome x
or with the outcome of participating in a second gamble with either a
(second chance for) x or with y as outcome. Option 2 as the right to
participate just once in the latter lottery is clearly worse than option 1
which offers two possibilities of gaining x. Once one has lost the first,
the second - which is identical with option 2 - still remains. Option 1,
however, is by no means adequately expressed by 'X+i+(X+i+y)', which ht
already made clear by the fact that this expression is logically equivalent
to 'y', whereas option 1 is a two-stage lottery, the utility of which is not at
all equal to that of the worse option Y' of premise I! As I put it
elsewhere: PL, insofar as it is (at least) semi-extensional, permitting the
exchange of equivalents in the relata, is a blind man bound to fall into all
kinds of traps if not led by the guidedog of UT. UTS 1 relying on UT
therefore is a step in the right direction though - as it stands - it will tum
out not to be totally satisfactory.
It is interesting to see that, though R1 and R 1' came out as truth-con-
serving under the 'direct' UT-analysis sketched ('direct' meaning that u(A),
u(B), and u(A+i+B) are the utilities of the sure options A, Band A+i+B and
not computed by averaging the utilities of all state-space generated A-, B
- and A+i+B - worlds, respectively, the situation is a bit different if R1 and
R 1 ' are checked by UTS 1 . For reasons of simplicity let us once again work
with only the minimal state space {A,B} and the set of worlds {W1: A /\ B;
W2: A /\ -,B; W3: -,A /\ B; W4: -,A /\ -,B} and only check R/:
R 1 ': If A>-B then A~A+i+B~B

326 RAINER W. TRAPP
yields
from whence we get, because of the transitivity of ' ~ " the arithmetical
truth, ifu2>u3 then U2~U3.
So R 1 ' eventually holds according to UTS 1 •
So far, so good. But what might be less welcome is that when we split
up R 1 ' into
If A';>-B then A?:,A+#B
and
If A,;>-B then A+#B?:,B,
and thus cannot make use of the additional premise of transitivity of ' ~ ,
within R 1 ' (a) and R 1 ' (b), the conjuncts do not come out true separately
under UTS 1 : for (a) 'If U2>U3 then U1 ~U3' is only true for U1 ~U2, and
(b) 'Ifu2>u3 then U2~U1' is only true for U3~U1.
In contradistinction to this at least at first sight counterintuitive result
- since everyone who accepts R 1 ' as plausible will also accept its logical
consequences R 1 ' (a) and R 1 ' (b) as plausible - the earlier direct UT-tran-
scription allows the splitting up of R 1 ' because both
(a') If u(A»u(B) then u(A)~u(A)·p+u(B)(l-p)
and
(b') If u(A) > u(B) then u(A)·p + u(B)(l-p) ~ u(B)
are arithmetical truths without any restriction. However, UTS 1 regains its
reputation (in this case), if one takes into account that what makes R1 and
R 1 ' so plausible is the interpretation 'if getting A (and nothing else) is
better than getting B (and nothing else) then .. .' rather than the interpre-
tation 'if getting A (and either B or not B) is better than getting B (and
either A or not A) then .. .'. The former interpretation is, however, formally
written (only for R1')
If A A -,B,;>-B A -,A then A A -,B?:,A+#B?:,B A-,A
comes out true under UTS 1 without the additional premise of the transi-
tivity of ' ~ " for
. U2+ U3
1fu2>U3 then U2 > - 2 - > U3
is an arithmetical truth that can also be split up in the way indicated so

that in this natural interpretation ofR 1 and R 1 ' the awkward result arrived
at above no longer comes up.
Having applied UTS 1 to analysing certain arguments from the PL-lit-
erature that are central for our topics of altemativeness and the relation
between PL and UT, we should now proceed to the more general problem
whether the formal traits of UTS 1 recommend it as a semantical founda-
tion for all non-basic preference comparisons. I already mentioned my
reservation about it, despite its fulfillment of the three conditions of ade-
quacy given above. Establishing this dissatisfaction calls for a closer look
at UTS 1 :
If, as was done above, one interprets the weighting factors! and ~ as
n m
probabilities, UTS 1 amounts to the definition:
A»B iff u(La) > u(Lb),
where u(La) and U(Lb) are the expected utilities of two lottery tickets with
all state-space generated A-worlds and B-worlds, respectively, as equi-
probable possible outcomes. Thus viewed the definiens is just a special
application of the well-known principle ofUT2 9 stating that, if AinAj=0
n
for all events Ai,A j and P( u Ai) =F 0, the expected utility of a disjunction
;=1
of i events A;, if Ai occurs with p(A i), is
"
i=1
In UTS 1 the Ai are replaced by the A-worlds and B-worlds, respectively.

(The principle itself does not presuppose anything about the nature of the
event-descriptions so that they do not have to be state-descriptions - or
328 RAINER W. TRAPP
their set-theoretical counterparts - gained by a given statespace.)

A first objection against UTS 1 now might be that it is not reasonable
to feign equiprobability for the possible outcomes of Ll and L 2 , if one is
in an actual decision situation confronted with the options A and Band
has evidence for a quite other than equal probability distribution over the
outcomes. Is not this suppression of utility relevant information? One
might meet this objection by suggesting a semantical criterion UTS 2 which
is like UTS 1 except that it requires the probabilities of the outcomes to be
fixed to the best of one's knowledge rather than to be principally con-
sidered as equal. 30
A second argument which I hold to be much more telling against UTS 1
might point to the lack of sense in attributing to 'possible' worlds - which
are, because of the incompatibility of two states in them, not even possible
- probabilities equal to those of worlds that may become real. Suppose
someone who has two errands to do prefers doing errand el first (= E 1)
rather than doing errand e2 first (= E2). Either option logically excludes
the other.
The most rigid solution here would be to argue that actions such as
these, and beyond them all alternative actions whatsoever (since all actions
contained in alternative-sets in decision situations pairwise even logically
exclude each other since they are time-identical), cannot be co-elements of
the same state-space, because the elements of a state-space have to be
logically independent of each other. So no preference comparisons involv-
ing elementary alternative actions x,y selected from an action-set for a
decision situation could fulfill the truth conditions of holistic PL, and thus
would all have to be excluded from being relata of '~', '>-' or '-'
in PL - a conclusion that would be extremely unwelcome, for preference
relations would, so to speak, have lost some of their best customers, and
would be left only with a severely reduced range of application. The only
exceptions would be preference comparisons of the special structure
'X>-,X' or vice versa if only one elementary action x is, beside possible
circumstances, in the respective state-space; x- and non-x-worlds can then
each be averaged.
But in many decision situations the action-sets contain more than the
contradictory actions X,I x. Preference relations among the n(n ~ 3) el-
ementary actions from action sets {x,y,z, .. .}, where each pair of elements
are only contrary negations of each other, simply cannot fulfill UTS 1 •
There are no x -, y -, z - .. , worlds, respectively, which one might compare,

for x,y,Z ... cannot be elements of the same state-space that could generate
these worlds.
Only actions, the logical incompatibility of which is due to the molecular
structure of their descriptions, could be compared on the basis of UTS 1
- and, moreover on the basis of any holistic semantics generating worlds
from state-spaces. So, instead of simply writing, e.g., 'x>-y' one would
have to write from the outset 'x A -,y A -'Z A -, '" >- yA -,x A -'Z A-'
... '. Only the whole conjunction of an elementary action and the negations
of all its co-members in the given alternative set would count as a proper
description of the action. But this would mean that the same action could
never occur in different action-sets! If only one co-action were changed,
left out, or added, all the other actions in the set would become different
actions and the utilities of different worlds would count according to
UTS 1 •
Another way to compare elementary actions x,y without blowing them
up to logically incompatible conjunctions would be to liberalise the re-
quirement that all preference relata must originate from the same state-
space. Instead one could allow that x-worlds and y-worlds generated by
two state-spaces SI and S2, respectively, can also be averaged if only the
circumstance space in SI and S2 is identical; or differently put, if SI and
S2 at most differ in that the former contains x instead of y.
If this way out, despite its ostensible plausibility and closeness to the
practice of several models of decision-theory, is rejected too and both
UTS 1 shall be applicable to preference comparisons between elementary
alternative actions and the requirement shall be stuck to that all preference
relata must stem from the same state space, there remains - as far as I can
see - only one further possibility: the alternative actions must be described
in such a way that they are logically independent and that time-identity
only entails empirical incompatibility in the same world. The minimal sta-
te-space {E 1 ,E2 } in the example given above would then yield the 4 worlds
{WI: El AE2; W2: El A -,E2; W3: -,E1 AE2; W4: -,E1 A -,E2} of which WI
clearly will not obtain, i.e., has probability O. And yet, UTS 1 forces us to
consider 'E1>-E2' as true iff t(UI +U2) > t(UI +U3) and in this context to
give WI a probability of t.
As the latter inequation comes up to 'U2 > U3', one might argue that
finally only the El A -,E2 - and E2 A -,E1 - worlds count, since the utility
330 RAINER W. TRAPP
of the empirically impossible El /\ E2 world can be eliminated on both

sides. But firstly the A /\ B world-utilities cannot be eliminated in that way
for all molecular relata A,B. And secondly why does one then not disregard
the A /\ B worlds from the start by changing UTS 1 in such a way to UTS 3
that in computing the truth-conditions for 'A>- B' only the A /\ -, B worlds
and B /\ -, A worlds get positive probabilities summing up to 1 in both
cases, whereas p( Wi) = 0 is stipulated for all A /\ B worlds Wi?
UTS 3 seems to gain even more attractiveness if one recalls that all pref-
erence relata, if only described with sufficient accuracy, are alternatives
that cannot coexist in any of the worlds into which either of them is embed-
ded and which is only minimally different from the actual world. But at
closer inspection UTS 3 reveals itself as a rather poor tool for catching the
gist of alternativeness for all cases of re1atum structures by purely exten-
sional means. It works for comparisons with two atomic relata A,B but
(as just seen for UTS 1 ) so do UTS 1 and UTS 2 for this simple case, since
the A /\ B world utility finally gets eliminated according to both. UTS 3
also works for molecular relata of certain symmetric structures. But it fails
for many molecular relata. E.g., it seems to me to be simply inadequate
to decree, as UTS 3 does, that 'A>-B' with 'L(A-+B)' - 'L' denoting logical
necessity - amounts to comparing a contradictory state with a different
state.
What sense does such a comparison - or one involving tautologies -
make? What utility does an impossible world have? Zero, as Rescher sug-
gests? On what kind of scale? If on an interval scale, why not I for fans
of contradictions? And yet the solution cannot simply be to forbid com-
parisons, for which 'L(A-+B)' holds, in order to save UTS 3 • For they do
make sense! 'x /\y>-xv y', for example - as analysed earlier by 'direct'
UT -transcription - is best conceived of as a preference for the sure option
'x /\y' over the lottery ticket (x /\y, Pi; x /\ -,y, P2; -,x /\y, I-Pi -P2).
So UTS 1 and UTS 2 , which also make sense of such comparisons - even
though the former rigidly requires equiprobabi1ity - come off better even-
tually.
But this does not only hold for comparisons in which one re1atum log-
ically implies the other. Consider 'xv y>--,y'. This is the assertion that
a certain three-outcome-10ttery ticket is better than a sure option accord-
ing to 'direct' UT-ana1ysis, or, working with the minimal state-space {x,y}
and UTS 1 or UTS 2 , that a certain three-outcome-10ttery ticket Ll is better
than a certain two-outcome-Iottery ticket L 2 • But can it reasonably be

considered as being true iff the expected utility of (x "Y,Pi;""'X "y, I-Pi)
is higher than the utility of·...., x " ...., y', as UTS 3 yields it due to extensional
computation? To me this is far away from the original preference for a
situation which offers the expected utility of Li but not that of L2 (or of
the sure option •....,y') over a situation that contains the converse offer. Of
course, UTS 3 could simply be stipulated as the truth-condition for PL-
formulae. The crucial point then would be the acceptability of semantical
criteria, as judged by at least the metacriterion of how well a criterion
copes with the requirements of alternativeness of relata and of UT-ground-
edness.
The crudeness of truth-functionality was detected as the root of all the
difficulties in combining both in a satisfactory way, and this brings us
almost to the end ofthe sad story ofPL and UT.1t seems indeed attractive
and quite straightforward to translate the UT-truth that x is better than
a lottery ticket with either x or y as outcomes into the PL-formula ·x>-
x+++y'; and, since the UT-truth can be formulated in an expanded way by
making explicit what it means, namely that a world Wi in which one gets
A but not the right to participate in the lottery is ceteris paribus better than
W2, where things are the other way round, it seems attractive to copy this
step in PL-Ianguage by the formulation ·x"...., (x+++y) >- (x+++y) " ....,x'.
But already here one is trapped by the consequences of truthfunctionality:
for Wi in the UT-reasoning is not simply a world in which ·x "y' is true,
and W2 is not simply a world in which •...., X" y' is true, as truthfunctionality
implies!
Let us recall a lesson that one can learn by an expanded version of this
example. It may help to justify my final judgment that present PL-systems,
though undeniably logics in a formal sense, (a) are not logics of preference
that specify in a systematic way how to conclude from the comparative
utility relation of states of structure A, B the utility relation of states of a
different structure A', B'; and (b) do so in a manner that formally rep-
resents the nature of rational, i.e., UT-guided, standard preference talk.
The essence of this lesson can be drawn from the following conversation:
Mr a in his favourite idiom of UT argues (= sequence 1): ·1 strongly
prefer x to y, or, more precisely, since 1 speak of alternatives, 1 strongly
prefer a world in which 1 get x but not y, other things being equal, to one
in which 1 get y but not x; thereforei 1 of course also at least weakly prefer
332 RAINER W. TRAPP
to get x for sure to the right of participating in any lottery with either x
or y as outcomes, or more precisely since 1 speak of alternatives, ... , and
therefore2 1 also weakly prefer the right to participate in any 2-stage lottery
with the outcomes 'either x for sure or the right of participating in a second
lottery with either x or y as outcomes' to the right to participate in that
simple lottery that is the second possible outcome of this 2-stage lottery,
or more precisely, since ...; therefore3 1 also wealdy prefer ... etc.
Mr b, having some knowledge of UT but preferring his native idiom
PL, interrupts Mr a - just when the latter once again raises his voice to
continue: 'Therefore999 1 also wealdy prefer .. .' - and says: 'Wonderful!
1 perfectly agree with you so far! But let me, before you come to speak on
the advantage of this 999-stage lottery, translate what you just said into
my language PL, so that your truths get the propagation they deserve also
in my country'. And he writes:
(1) x>-y, or more precisely, since ... : X"
IY>-Y" IX;
(2) therefore1: x;:: x +++Y, or more precisely, x " I (x +++ y) ;:: (x
+++ y) " I x;
(3) therefore2: x +++ (x +++ y) ;:: x +++ y, or more precisely ... (x +++
(x+++ y)) " I (x +++ y) ;:: (x +++ y) " I (x +++ (x +++ y)).
Before coming to line (4) Mr b turns to Mr a (who, also having some
knowledge of the PL-idiom already looks somewhat embarrassed) and
politely asks: 'I hope you don't mind my adding a little extra conclusion
that my idiom allows one to draw from line (3), though it did not occur
in your reasoning'. And, Mr a having nodded with a strange look in his
eyes, Mr b writes:
(4) Therefore: y>-x+++y ...
At that point, Mr a sadly turns his face from the paper Mr b writes on
and says: 'I fear you did not quite catch what 1 meant!' 1 cannot but agree
with Mr a here as firmly as Mr b agreed with Mr a's reasoning. The rules
R1 and R 1', as discussed by Kanger, are excellent examples of how PL
may set off with innocent and highly plausible starting points, often only
gained by throwing a look over the border to UT, and then, having made
itself independent of UT-reasoning and its restrictions by pressing the
UT-message into the linguistic Procrustean bed of (at least semi-) truth-
functionality31 and its rules, suddenly finds itself in the mess of counter-
intuitiveness or even contradiction. So truth-functionality, to stress it once

more, is the misery of PL. Propositional logic is -like predicate logic - too
crude a tool to convey what standard preference comparisons of some
complexity actually intend to convey. On the other hand, truth-function-
ality is in an essential sense also the glory of PL. For though it helps to
strip PL of its character of a logic of standard preference, it does impart
to it its character of an easily manageable logic. It is the very
identification of the utility values of certain state-space generated worlds,
the descriptions of which are extensionally equivalent, that so easily per-
mits us, together with a semantics such as UTS 1 or UTS 2 , to make infer-
ences from preference assertions with relata A,B to those with relata A',
B' of different logical structure. 'Direct' UT, which works with utilities u
and probabilities p directly attributed to propositions, and not with the
utilities and probabilities of clearly defined worlds which only indirectly
allow us to determine the utilities of propositions true in one world or
several, permits conclusions from preference assertions with A,B to those
with A', B' only to the extent that they are stipulated to correspond to
UT-transcriptions arithmetically true for all admissible u and p. Such
stipulations, as, e.g., the one we made frequent use of - namely that 'x+l+y'
corresponds to (x,p; y, I-p) - do, however, pave the way for only a very
limited set of PL-formulae such as 'x>--y--'x~X+l+y~y', which set hardly
deserves the name of a logic.
No conclusions that in some systems of PL may be drawn from pref-
erence comparisons involving only unnegated relata to comparisons with
relata that contain negations (without being logically equivalent with the
former relata), can be given a sense according to direct UT. If a certain
set of propositions {... x, y ... } has been weakly ordered and cardinalised
with the result that 'u(x) > u(y)" one cannot on this sole basis, even to-
gether with UT, say anything about whether or not 'u(-,y) > u(-,x)'; for
u(-,y) and u(-,x) are not defined without having recourse to a set of
worlds generated from at least a minimal state-space. One can only infer
that '-u(y» -u(x)': but this does not convey anything about negative
states. One may at best - if one wishes to - interpret it as the fact that, if
x is better than y, then also the justified regret at not getting y (represented
by '-u(y)') is ceteris paribus less intense than that at not getting x. The
reason why - even if! know, in a certain situation S, that I prefer eating
steak to eating sole - I do not have the slightest idea whether in that same
334 RAINER W. TRAPP
situation I therefore prefer also not eating sole to not eating steak is simply
my lack of information what not eating steak or sole, respectively, amounts
to in S ... to eating nothing at all; to eating a Hamburger; a dinner cooked
by Paul Bocuse; to washing the dishes at the time I otherwise would have
eaten; or what? Normally the informational content (in the Popperian
sense) of negated propositions is too small to put a decision maker in a
position to make direct preference comparisons between them.
This does not exclude the possibility that there may be situational
knowledge which gives sense to comparing negative relata. But this is so
only because behind the negative characterisations hide states that are
positively characterised and can be concluded from the situational pre-
suppositions. Contradictory negation must convey the information of con-
trary negation. If one knows that one must eat something in any case, and
that only steak or sole are being served, the question may be substantial
whether one prefers not to eat steak or not to eat sole. Without such
presuppositions, however, one should be cautious enough to ask some
questions before answering. For otherwise one may -like a Catholic in-
vited to a malicious atheist preference-logician's house on Friday and hav-
ing preferred not to eat steak to not eating sole - end up with a cutlet on
one's plate.
So direct UT can follow holistic PL in declaring, e.g., 'x~x~y' to be
a reasonable prefex:.ence if 'x»y', But once the question is whether there-
fore '-'(x~y)~-'x' is also true, their ways part. UT, without further
information, just does not know what it is not to participate in the lottery
(x, p; y, 1-p) and not to get x. For holistic PL (working with the minimal
state-space {x,y} this latter question is as clear as the former one and
amounts to whether '(XAY)V(-'XA-'Y)~(-'XAY)V(-'XA-'Y)' is true.
Without state-spaces and (at least semi-) extensionality there would be no
easily manageable fullfledged logic.
On the other hand, the problem of lack of content of negative propo-
sitions also appears in holistic PL at a different place. For, though state-
spaces define worlds, the determination of the utility values of these worlds
presupposes that the state descriptions into which these worlds are embed-
ded each have been assigned a utility value before 3l • But how can one
achieve this? How can one weakly order the 2n worlds generated from {Xl>
... , x n}? What rank can one give, e.g., '-'XI A -'Xl A ••• A -,xn' in the
order and how can one cardinalise it, be it by direct rating, v. Neumann/
Morgenstern or whatever other method? All x;'s may be described as pre-

cisely as possible; but without any situational presupposition one does not
know how to assign utilities to worlds that are conjunctions of exclusively
negated propositions. Utility value assignments require more information
than truth-value assignments. So there is one more problem for PL to solve.
The most radical solution might, as just indicated, lie in what one could
term the requirement of maximal enhancement of informational content for
negated propositions: each of the n negative conjuncts ',x;' must, wher-
ever it occurs in one of the state descriptions, be replaced by a positive
conjunct 'Xi·' which is a contrary negation of 'x;' and as such a sub-state
of the contradictory negation ''x;'. Less demanding (and indispensable,
I think) would be to require at least minimal enhancement of content by
replacing at least one ',Xi' by 'Xi·'. This would give each state description
at least one positive conjunct in which to anchor the utility assignment. I
shall leave the problem of negated states for PL at that. A careful inves-
tigation of its intricacies requires a paper of its own.
The foregoing considerations justify the summary of our account of the
relation of PL and UT by a half-way conciliatory verdict if only the for-
mer's claim as to its function is moved into place: both fruitfully cooperate
to a certain extent in as much as PL is based on one or other UT-semantics.
But P L clearly is not what its name makes it appear to be: a logic ofstandard
preference relations.
What PL based on UTS 1 or UTS 2 actually turns out to be is a calculus
of comparative value relations for propositions composed of logically in-
dependent atomic propositions. As such it has its merits and its range of
application. Since by the nature of its truth-conditions, which rely essen-
tially on extensionality, it does not appropriately cope with alternativeness
as a fundamental ingredient of standard preference. It should rather be
termed a calculus ofbetterness-relations between propositions. A real logic
of standard preference would require a much finer language than PL
offers.
It seems that PL, reduced to this calculus, has a much more limited use
than its authors and propagators think it has. For all elements of a sta-
te-space S= {Xl> ... ,x.. }, notwithstanding their logical independence, may
stand in all kinds of causal relations (necessary, sufficient, INUS 33 , or
whatever) to one another. Causal relations are even required for a fruitful
use of this model. What else but certain possible causal consequences of
336 RAINER W. TRAPP
a state Xi are (at least some of) the states Xi+h Xi+2, ... that co-obtain
with Xi in the possible worlds in which Xi is true?34 So, at least in normal,
i.e., consequentialist, use of the model, these states will neither be pairwise
stochastically independent of each other nor coexist, but rather weakly
orderable according to the beginning of the temporal intervals in which
they obtain. What, then, makes Xi useful or good to a smaller or greater
degree U(Xi) is, apart from its possible intrinsic goodness, just the sequence
of future consequences that it may have in each of the possible worlds it
belongs to. And yet all UTS-models allow us to determine the value u(A)
for each elementary or molecular state true in some world W;, be it the
nearest or the farthest in the future, by averaging the U(Wi), and also permit
us to calculate u(A), if A is a remote consequence of some other state in
the same world! But what sense does it make - contrary to the converse
procedure of determining the value of causal (co-) conditions in terms of
its consequences - to determine the utility of a state with view to the con-
ditions that causally (co-) determined it? Does one not, rather than look
backward, wish to determine the utility value of a consequence Xi of Xi in
terms of x/s possible intrinsic value plus that of the further consequences
it helps to bring about?
So it seems appropriate to restrict all UT-models in such a way that
only the temporally earliest states A (e.g., Xi and non-Xi) should be deter-
minable as for their utility by averaging the u(wi) of the Wi in which A is
true. This would come up to what most decision models do, which is to
determine the utility of actions in terms of their possible consequences and
not also of their causes. The greater generality of the UTS-calculations
(notably the realistic UTS 2 ) would consist in the earliest states of a world
not necessarily having to be actions. If not only Xj and non-Xj should be
compared, but all the actions of an action set {ah ... ,an }, then, as suggested
some pages above in a similar context, one should give up the requirement
that all Wj must stem from one and the same state-space S. Rather, one
should work with n state-spaces Sj= Cu{aj} which all have an identical
circumstance space C and only differ with regard to the aj. These actions
should always be the earliest states in the worlds they are possibly true in.
And so in general for non-standard holistic PL: preference comparisons
between states of whatever truthfunctional structure and whatever tem-
poral and causal position in the w,sequences as, e.g., Rescher describes
them on the basis of UTS b do (but not always) make sense in the con-
sequentialist application of preference relations; and this is certainly the

most interesting application.
NOTES
• My thanks are due to Uwe Steiner for carefully reading a draft of this paper.
1 Systems of PL working with a holistic and probabilistic ( == HP) semantica1 basis for PL as,
e.g., suggested by Rescher (1966) will, for reasons which will become clear later, at least
implicitly deny this requirement. But also other contributors to PL, such as Kanger, who do
not draw upon an HP-conception of PL, reject it. (Kanger did so in private conversation.)
Minimal difference of worlds is meant in the usual sense of this term current in conditional
semantics.
2 If 'p and q both obtain' a subject's preference for p over q 'must' - von Wright argues in
(1963), p. 24 'mean that he would rather lose q (and retainp) than lose p (and retain q).' So
what the subject actually does is to prefer'p 1\ ..., q' to •...,p 1\ q', which states do not co-obtain
but are actually both false ifp and q both obtain as presupposed. Evidently seeking something
like alternativeness, also for co-obtaining states, von Wright simply replaces the actual states
by two counterfactual possible ones as the relata for ';>'. But this is in fact a withdrawal of
his presupposition that p and q co-obtain and are the preference relata.
3 The condition for 'G' and • -' is like Cb with corresponding modifications in the definiens.
4 I could express this by writing' Tx' instead of' T'. A 'person' is, like in individual decision
theory, also any group of persons sharing a preference and thus speaking with one voice.
S See Stalnaker and Thomason (1970) and Stalnaker (1968).
6 D. Lewis (1973). See especially p. 35, where he denies the law of contraposition for coun-
terfactual conditionals.
7 F. von Kutschera, in (1976), pp. 123-24, considers von Wright's (1972) and other possi-
bilities of basing preference between propositions on a comparison of sets of possible worlds
and points to the difficulties this introduces. He, however, nowhere tries to regard only one
A- and B-world for at least that kind of preference that I call basic.
S See von Wright's (1963), pp. 29--30, and (1972), pp. 146-148. Rescher's holism in (1966),
esp. pp. 43-45, is a probabilistic extension of von Wright's.
9 Von Kutschera's definition s ofnon-probabilistic preference (see Note 7) include all worlds,
i.e., also A 1\ B-worlds, into the comparison sets; and also Rescher, in conceiving preference
as probabilistic, does this. The resulting drawback, as well as the reasons why working with
'A 1\ -,B' and 'B 1\ ...,A' is not sufficient to seize the gist of the altemativeness of 'A' and
'B', will become clear in Section II.
10 'Atomic' here refers only to the external structure of a preference statement as not being
a truth-functional part of a complex of at least two preference statements. The internal struc-
ture of each of its relata may be atomic or molecular.
11 See Kanger (1968) and Saito (1973).
12 See L. Bergstrom (1966), Chapter 2.
13 u(A) and u(B) are, as usual in UT, unique up to positive linear transformations. It does
not matter here whether the axioms for 'G' meet the standards of rigour of Fundamental
Measurement Theory (see Krantz et al. (1971» or whether one of the v. Neumann/Morgen-
stem-type-metrisations ofuti!ity, admitting probability values already in the axioms, is cho-
sen.
14 I do not distinguish here between treating these things in a set-theoretic or truth-functional
setting.
338 RAINER W. TRAPP
15 See von Wright (1963), p. 46, who - because of his restriction - calls PL a 'semi-extensional
calculus'. Harder restrictions than those considered would not even keep PL semi-extension-
al.
16 The UT-counterparts of strong preference for sure options over lotteries such as in (b)
x>-x<++y, Le., (b') u(x»u(x).p+u(y)(I-p) only hold for allp<1.
17 The UT-analysis of these latter two PL-formulae would agree with Rescher's truth-con-
ditions for these formulae in (1966), if equiprobability for the lottery outcomes were required.
But, as will become clear soon, the UT-analysis of PL-formula which I favour does not in
general amount to a Rescher semantics.
18 See Hallden (1957) and von Wright (1963).
19 See Hansson (1968), pp. 428-29.
20 See Saito (1973), pp. 387-88.
21 Rescher (1968), pp. 48-49, argues against the restrictedness of von Wright's purely ordinal
world comparisons.
22 Rescher (1968), see especially pp. 43-51.
23 Chisholm and Sosa (1966), p. 245.
24 As the treatment of Chisholm and Sosa's argument against C 2 demands only the con-
sideration of one direction of C2 , I disregard the other one, which, of course, is valid too on
UTS 1 •
25 In a second counterexample on p. 245 against C h 'x' logically implies 'y', which yields
a contradictory preference relatum 'x A ,y' according to C1 • I agree with this point against
C 10 not only because it does not make sense to compare a contradictory state with a pleasure
containing state, but for the general reason that contradictory states, not even fulfilling Cb of
section I, are not suitable relata for any system of PL. Remember that we used the resulting
contradictory relata, if x logically implies y and 'x>-y', as one principal argument against
the insufficiency of expressing alternativeness in the mere truth-functional language used in
the relata of C 1 •
26 See Kanger (1980), pp. 37-38. Actually, Kanger presented the following argument to me,
before its pUblication, in private conversation during his 1979/80 lecturing period in Frank-
furt/M. The rules Rl and R 1 ' were, as Kanger pointed out, extensively discussed on the
occasion of a congress in Oslo in 1982.
27 Both deductions are analysed in detail elsewhere. See Trapp (forthcoming).
28 I confine myself to a UT-look at R 1 •
29 This principle, from the beginning of UT with von Neumann/Morgenstern, belongs to the
core of utility measurement and the disciplines where the latter is needed. For example,
Jeffrey's central axiom (5-2) of his (1965) is nothing but a special case of it.
30 In Note 10 on p. 44 of his (1966) article Rescher considers this variation of UTS 1 •
31 This corresponds to von Wright's sense of 'semi-extensionality' (cf. Note 15).
32 Rescher, for one, simply presupposes this on p. 43 of (1966).
33 This term was introduced in Mackie's (1965) analysis of causality as defined thus: A is an
INUS condition of a result P iff, for some X and for some Y, «A and X) or Y) is a necessary
and sufficient condition of P, but neither A nor X is a sufficient condition of P.
34 The use which, e.g., von Kutschera in his most lucid and penetrating treatise (1982) makes
of this method of determining the utility (or, rather, 'goodness') of states, is clearly of this
kind. See esp. pp. 15-16.
REFERENCES
Aqvist, L., Chisholm, R. M. and Sosa, E.: 1968, 'Logics of Intrinsic Betterness and Value',
Noils 2, 253-70.
Bergstrom, L.: 1966, The Alternatives and Consequences of Actions, Almqvist and Wiksell,
Stockholm.
Chisholm, R. M. and Sosa, E.: 1966, 'On the Logic of 'Intrinsically Better',' American Phi-
losphical Quarterly 3,244-49.
Hallden, S.: 1957, The Logic of 'Better', vol. 2 of Theoria Library, Uppsala.
Hansson, B.: 1968, 'Fundamental Axioms for Preference Relations', Synthese 18, 423-42.
Jeffrey, R.: 1965, The Logic of Decision, McGraw-Hill, New York.
Kanger, S.: 1968, 'Preferenslogic', in Nio filosofiska studier tilliignade Konrad Marc-Wogau
(ed. H. Wennerberg), Uppsala (Xeroxed in English).
Kanger, S.: 1980, 'A Note on Preference-logic', in Philosophical Essays Dedicated to T.
Dahlquist, Uppsala University Publications, Uppsala.
von Kutschera, F.: 1976, EinjUhrung in die intensionale Semantik, de Gruyter, Berlin/New
York.
Rescher, N.: 1966, 'Semantic Foundations for the Logic of Preference', in The Logic of
Decision and Action (ed. N. Rescher), University of Pittsburgh Press, Pittsburgh, pp. 37-
79.
Saito, S.: 'Modality and Preference Relation', Notre Dame Journal of Formal Logic 14,
387-91.
Stalnaker, R. C.: 1968, 'A Theory of Conditionals', in Studies in Logical Theory (ed. N.
Rescher), Blackwell, Oxford.
Stalnaker, R. C. and Thomason, R. H.: 1970, 'A Semantic Analysis of Conditional Logic',
Theoria 36, 23-42.
Trapp, R. W.: forthcoming, 'Utility Theory as Guide Dog: Some Alleged Contradictions of
Preference Logic'.
Mackie, J. L.: 1965, 'Causes and Conditions', American Philos. Quarterly 2, 245-64.
von Wright, G. H.: 1963, The Logic of Preference, Edinburgh University Press, Edinburgh.
von Wright, G. H.: 1972, 'The Logic of Preference Reconsidered', Theory and Decision 3,
140-169.
Received 2 April1984
Dept. of Philosophy
Johann-Wolfgang-Goethe-University
Dante-Str. 4-6
6000 Frankfurt/M.
F.R.G.
RUDOLF HALLER
DER ERSTE WIENER KREIS
Die These, die ich zur Oberpriifung vorlege, ist die folgende: Es gab vor
der Griindung des sogenannten Wiener Kreises um Moritz Schlick einen
ersten Wiener Kreis mit Hans Hahn, Philipp Frank und Otto Neurath.
Dieser Kreis ist fiir die Bildung des Kreises um Moritz Schlick konstitutiv,
sodaB das Urteil gerechtfertigt werden kann, daB eigentlich Hans Hahn
den Wiener Kreis gegriindet hat. Um darauf aufmerksam zu machen, nen-
ne ich den einen den ersten und den anderen den zweiten Wiener Kreis.
Dem Historiker sind zunachst wie jedem Theoretiker bei der Bildung
seiner Hypothesen keine anderen Schranken geboten, als jene der
Verstandlichkeit. Solange seine Aussagen sinnvolle Annahmen sind, so-
lange hat er auch eine der notwendigen Bedingungen der Forschung erfii1lt
und Anspruch auf Priifung seiner Urteile. Nicht mehr beanspruche ich
zunachst fiir meine Oberlegungen.
Bevor ich diese These verteidige und begriinde, mochte ich sagen, daB
es nicht historische Neugierde ist, die eine Beschaftigung mit diesen Fragen
nahelegt oder gar erzwingt, obschon die neuerliche Belebung der Diskus-
sion um den Wiener Kreis auch ein Erkenntnisinteresse fiir die Herkunfts-
geschichte rechtfertigt. Es ist vielmehrdie Frage nach den philosophischen
und systematischen Divergenzen im bekannten Wiener Kreis, die eine Er-
klarung erfordert, die iiber die bloBe Beschreibung der Unterschiede bin-
ausgeht und dadurch den genetischen Aspekt ins Spiel bringt. Die Wie-
derbelebung der Diskussion um den logischen Positivismus geht ja nicht
auf die Auseinandersetzung mit den Fragen theoretischer und praktischer
Rationalitat zuriick, l die ja erst spat auftritt, sondem setzt ja bereits Ende
der Sechzigerjahre ein, um in der monumentalen Vienna Circle Collection
ihren deutlichsten Niederschlag zu finden. 2
Diese Wiederbelebung ist zunachst eine Parallelerscheinung mit dem
Aufkommen des Historismus in der neuen Wissenschaftstheorie. Insofem
dieser Historizismus gegen eine Standard-view wissenschaftlicher Theo-
Erkenntnis 22 (1985) 341-358 0165-0106/85/0220--341 $01.80

342 RUDOLF HALLER
rienbildung gerichtet war, die im Wiener Kreis vertreten sein sollte, hat
man die sog. 'received view' geme auch mit den Theorien des Wiener
Kreises identifiziert, wie ich sogleich betonen muB. Also ware es nur na-
heliegend gewesen, diese Ideen aufs neue aufzusuchen, ihre wahre Gestalt
kennen zu lemen. Denn in der Zwischenzeit war namlich, durch die rasche
Entwicklung der analytischen Wissenschaftstheorie, ein Klischeebild des
logischen Positivismus entstanden, das sich vomehmlich auf wenige Ve-
reinfachungen stiitzte. Nach diesem Bild hatten die Vertreter des Wiener
Kreises erstens eine reduktionistische Erkenntnistheorie vertreten, derzu-
folge alle Aussagen auf solche iiber unmittelbar Gegebenes zurUckgeflihrt
werden konnen, oder zumindest aus Elementarerlebnissen konstituierbar
waren; zweitens ware der logische Empirismus des Wiener Kreises allein
an logisch-analytischen und systematischen Problemen der Wissenschaften
interessiert und somit ahistorisch gewesen; und drittens hatte der Wiener
Kreis das Modell des kumulativen Fortschritts wissenschaftlicher Theorien
vertreten und damit eine unhaltbare Auffassung des Theorienwandels. Ich
habe schon bei verschiedenen Gelegenheiten darauf hingewiesen, daB die-
ses Bild des Wiener Kreises falsch ist und einer radikalen Korrektur be-
darf. 3 Die angeblichen Dogmen des logischen Empirismus wurden von
den logischen Empiristen selbst kritisiert, und einige Mitglieder des Wiener
Kreises haben nie Thesen vertreten, die mit den soeben herausgestellten
kontaminiert werden diirften. Dies herauszustellen, scheint mir allein
schon ein hinreichend starker Grund, sich mit der tatsiichlichen Geschichte
des Wiener Kreises aufs neue zu beschiiftigen. Ein weiterer Grund ist ein-
fach der, daB der logische Empirismus fUr lange Zeit nahezu die einzige
philosophische Theorie war, die das wissenschaftliche Weltbild unseres
Jahrhunderts in Logik, Mathematik, Physik, Biologie und Soziologie in
einer einheitlichen Perspektive zu verstehen trachtete und das fortge-
schrittenste Erkenntnismittel der Evolution, namlich wissenschaftliche
Methoden, zum zentralen Gegenstand philosophischer Analyse werden
lieB. Gegeniiber dieser Erkenntnisperspektive muBte eines der urspriing-
lichen Motive der neo-empiristischen Bewegung, die aufklarerische Anti-
~etaphysik, in den Hintergrund treten, obschon es auch die wissenschaft~
liche Weltauffassung zutiefst bestimmte. Das Vorbild Ernst Mach beein-
ftuBte eben nicht nur den bekannten Wiener Kreis, sondem erst recht den
'ersten'. So will ich denn von diesem sprechen.
DER ERSTE WIENER KREIS 343
II
In der Einleitung zu einer Sammlung seiner Aufsatze hat Philipp Frank

sehr deutlich und ausflihrlich beschrieben, wie der erste Wiener Kreis ent-
stand, welche Themen die Diskussionen beherrschten und welche haupt-
sachlichen Teilnehmer wir in unsere Geschichte aufzunehmen haben. 4 Der
Bericht von Frank geht davon aus, daB es urn die Jahre 1907 bis 1912 in
Wien einen Kreis von jungen Wissenschaftlern gab, die sich regelmaBlg -
im iibrigen wie der spatere Wiener Kreis donnerstags - trafen und deren
Interessen einerseits wesentlich durch methodologische und wissen-
schaftstheoretische Fragen, andererseits aber auch durch 'politische, histo-
rische und religiose Probleme' urnschrieben werden. Aus dieser Darstel-
lung wird klar, daB die wesentlichen Autoren, die die Einstellungen dieser
Gruppe von philosophisch interessierten Wissenschaftlern formten und
beeinfluBten, Ernst Mach sowie die franzosischen Wissenschaftstheoreti-
ker Pierre Duhem, Henri Poincare und Abel Rey waren. Diese Namen tau-
chen auch schon viel friiher als bei Frank in den Darstellungen auf, die
Otto Neurath der Geschichte und Vorgeschichte des Wiener Kreises wid-
mete. Und in der Tat sind es auch diese Namen, denen wir als Orientie-
rungshilfe bei der Suche nach den Leitlinien des ersten Wiener Kreises
immer wieder begegnen.
Man hat diese Richtung der Wissenschaftstheorie bei den Darstellungen
der Geschichte des logischen Empirismus zwar nie ganzlich ausgeklam-
mert, aber meistens nicht in gebiihrender Weise untersucht. Und erst recht
wurde nicht bemerkt, in welchem AusmaB die sogenannten franzosischen
Konventionalisten zusammen mit Mach und Popper-Lynkeus die Ideen
des ersten Wiener Kreises beeinfluBt haben. Es ist die Zeit des ersten Jahr-
zehnts unseres Jahrhunderts, in der die Hauptwerke Duhems und Poinca-
res erschienen, und es ist die gleichePeriode, in der die wichtigsten von
ihnen auch ins Deutsche iibertragen wurden. So erschien Ziel und Struktur
der physikalischen Theorien 1908 mit einem Vorwort von Ernst Mach in
der 'Obersetzung von Friedrich Adler bereits zwei Jahre nach der ersten
Buchveroffentlichung im Franzosischen. Die 'Obersetzung von Henri
Poincares Wissenschaft und Hypothese und Der Wert der Wissenschaft
waren ihr vorausgegangen. S Gleichfalls 1908 erschien Die Theorie der Phy~
sik bei den modernen Physikern von Abel Rey, in der deutschen 'Oberset-
zung von Rudolf Eisler.
344 RUDOLF HALLER
Ich glaube, man kann die Kemthese dieser franzosischen Konventiona-

listen durch das Bild, das Poincare von den verschiedenen Geometrien
gebracht hat, wiedergeben. 6 Poincare fiihrt dabei hypothetische Modelle
vor, die zeigen sollen, wie aus der verschiedenen Perspektive, die wir in
bezug auf die Dinge einnehmen, in sich widerspruchslose Gebilde resultie-
ren, die wir insofem selbst konstruieren, als es von unseren Entscheidun-
gen abhangt, welche Annahmen wir machen und welche wir akzeptieren.
Wir stellen den Rahmen her und fiigen ihm ein, woriiber wir iibereinkom-
men. Dieses Obereinkommen, d.h. das eigentlich konventionelle Moment,
ist freilich kein faktisches und auch kein Konsens einer idealen Diskus-
sionsrunde, sondem eine als vemiinftig angesehene Festsetzung, die nicht
allein durch Logik und Erfahrung begriindet und gerechtfertigt werden
kann.
Wie ist dieser Weg zwischen (oder, wenn man will, jenseits von) einem
aprioristischen Rationalismus und einem Empirismus moglich? Indem
man der Forschung nicht die Ziele zuschreibt, die ihr gewohnlich zuge-
schrieben werden, namlich eine moglichst umfassende und moglichst ge-
naue Beschreibung der Phanomene zu liefem und ihre grundlegenden Ge-
setze zu entdecken. Die Antwort, die Mach der traditionellen Auffassung
der Weltbeschreibung durch die Wissenschaften entgegenhaIt, riickt be-
reits das Wertkriterium der Konventionalisten in den Vordergrund. Wis-
senschaft ist nicht blofte Beschreibung, sondem okonomische, d.h. auf ein-
fache Elemente reduzierte Beschreibung der Phiinomene und ihrer Bezie-
hungen. 7 Die Phantasie, die uns den unbeschrankten Moglichkeitsraum
eroifnet, wird durch die Sparsamkeit rationaler Modelle in bezug auf einen
bestimmten Bereich von Gegenstiinden mehr und mehr eingeschrankt. Die
'Anpassung der Gedanken an die Tatsachen', die uns die Naturgeschichte
der Evolution vorfUhrte, wird durch die Theoriebildung fortgefUhrt und
vervollkomment. Freilich nie ohne den hohen Preis der Vereinfachung, der
Schematisierung und Idealiserung der Gegenstiinde.
Haufig wird uns, im iibrigen nicht nur von Gegnem des Empirismus,
ein falsches Bild der Lehren Machs geboten, so als ware er erstens ein
reiner Induktivist und zweitens ein reiner Phanomenalist. Beide Interpre-
tationen sind vollig unhaltbar und dienen nur der Verwirrung. Ohne jede
Zweideutigkeit hat Mach immer wieder das Einwirken des Okonomie-
prinzips in die Ordnung der Erfahrungen herausgestellt, und das heiSt
iibersetzt, das Einwirken von Verstandesoperationen auf die sinnlichen
Erlebnisse. Darum gibt es fiir Machs Wissenschaftstheorie sowenig eine

Prarogative der Beobachtung vor der Theorie, wie des subjektiven vor dem
objektiven Standpunkt. Mach teilt also mit den franzosischen Konventio-
nalisten den Standpunkt, daB unsere Erfahrungen nicht nur theoretisch
geordnet werden, sondern daB das geforderte Ideal der eindeutigen Be-
stimmtheit nur durch Entscheidungen moglich ist, die wir bei der Be-
schrankung unserer Phantasieprodukte ebenso walten lassen wie bei Ve-
reinfachungen im allgemeinen. Er stimmt auch Duhem darin zu, daB einem
'theoretischen Wert eine Unzahl von experimentellen Werten zugeordnet
werden kann', und daB schon daraus die bloB provisorische Geltung un-
serer Naturgesetze hervorgehe. 8
In der Tat war dies ja eine der wesentlichsten Erkenntnisse von Duhem,
daB einem Erfahrungsdatum mehrere theoretische Deutungen gegeben
werden konnen, und daB die Ubereinstimmung mit der Erfahrung - das
einzige Kriterium der Wahrheit fUr eine philosophische Theorie - durch
unsere Entscheidungen gepragt und bestimmt werden. Natiirlich verbleibt
zwischen dem Standpunkt Machs und dem der naturgemaBen Klassifi-
kation als Reflex der wirklichen Ordnung, wie ihn Duhem vertritt, eine
groBe Differenz, denn nirgends mochte Mach den metaphysischen Uber-
bau Duhems iibernehmen. Aber in der Charakterisierung der Theorien-
bildung, in der Einschatzung des konstruktiven Momentes der Abstrak-
tion, der Vereinfachung, scheint keine Differenz zwischen den beiden Den-
kern vorzuliegen.
Nun wird man nicht leugnen konnen, daB wir einige dieser Argumente
anerkennen miissen, wollen wir nicht das Wesen der wissenschaftlichen
Forschung verfalschen. So miissen wir zugeben, daB keine Tatsache ohne
eine Beschreibung als zutreffend oder nicht zutreffend behauptet werden
kann. Wir beziehen uns auf Tatsachen und Ereignisse, indem wir sie durch
sprachliche Zugriffe individuieren, sie als 'ein so und so' oder als 'der, die
oder das so und so' beschreiben. Und wie man weiB, konnen wir iiber jede
Tatsache eine unendliche Folge von Beschreibungen errichten. Und analog
laBt sich dann ein- und dieselbe Tatsache durch verschiedene theoretische
Satze erklaren. Fiir die praktischen Zwecke kiimmert uns diese Vielfalt
meist nicht sehr: wir wahlen die giinstigsten und einfachsten aus, je nach
den Bediirfnissen des betreffenden Falles. Aber wir wahlen und konnen
wahlen. Und in den theoretischen Disziplinen ist es nicht anders: nur kiim-
mert uns hier auch die Vielfalt der Moglichkeiten.
346 RUDOLF HALLER
Die entscheidenden Schritte, die Duhem den Machschen Uberlegungen

des Einfachheitsprinzips hinzufiigt, sind erstens, daB es zwischen zwei
Theorien, die zur Wahl der ErkHirung ein- und desselben Phanomens zur
Verfiigung stehen, kein experimentum crucis gibt, das iiber die Richtigkeit
oder Wahrheit einer beiden entscheiden konnte, und zweitens, daB die
Uberpriifung einer experimentellen Hypothese wie eines theoretischen Sat-
zes jeweils das ganze System von Beobachtungen wie von theoretischen
Satzen einander konfrontiert. 9 Dieser Holismus, der selbst theoretisch
schwer zu fassen ist, wirkt auf den ersten Wiener Kreis vielleicht am al-
lersHirksten. Und genau diesen Aspekt hat man in allen Darstellungen des
logischen Empirismus so gut wie iibergangen, wenn nicht iibersehen.
Es ware, wie gesagt, iibertrieben, zu meinen, daB Machs Elementenlehre
mit den Theorien der franzosischen Konventionalisten vollig iibereinge-
stimmt hatte. Vielmehr ist gerade aus der Spannung zwischen den wissen-
schaftstheoretischen Positionen, wie sie Poincare und Duhem auf der einen
und Mach auf der anderen Seite vertraten, das Biindel von Fragen er-
wachsen, das die Problemlage des ersten Wiener Kreises bestimmte.
Ich denke, daB die beiden franzosischen Wissenschaftstheoretiker vor-
nehmlich dadurch zu einer anderen Deutung der Naturgesetze als Mach
gelangten, weil sie an anderen Problemen interessiert waren und andere
Konstituentien wissenschaftlicher Theorienbildungen untersuchten. So
war der Ausgangspunkt von Poincares Argumenten zugunsten einer kon-
ventionalistischen Interpretation von Theorien der Vergleich der euklidi-
schen und nicht-euklidischen Geometrien. Da sich - den axiomatischen
Aufbau der einzelnen Theorien vorausgesetzt - verschiedene Axiomensys-
teme zur Beschreibung gleicher Gebilde verwenden lassen, muBte als die
wesentliche Frage jene nach dem Auswahlkriterium theoretischer Satze im
Vordergrund stehen. Und die Antwort war: wenn die Beschreibung eines
Phanomens in zwei Komponenten, eine theoretische und eine empirische,
zu zerlegen ist, und die erstere auf Festsetzungen beruht, dann kann die
Auswahl einer praferierten Hypothese nicht auf Kriterien der Zuordnung
und erst recht nicht der Ubereinstimmung beruhen. Wenn aber keine 10-
gischen und keine ontologischen Praferenzregeln lei tend sind, bleiben nur
pragmatische iibrig, wie z.B. Einfachheit und Bequemlichkeit. Ob ein geo-
metrisches System anderen vorgezogen wird, hangt sonach nicht nur da-
von ab, ob es hinreichend den Tatsachen 'angenahert' ist, sondern ob es
Eigenschaften aufweist, die eine Zuordnung zu beobachtbaren Pha-
nomenen eher ermoglichen als ein alternatives System und ob es sich dem
Insgesamt des verwendeten theoretischen Systems einfiigt. 1m Grunde se-
hen wir auch hier eine Konsequenz aus dem Machschen Gedanken der
Okonomie. Wie bei Mach die Vereinfachung oder Idealisierung eines theo-
retischen Satzes von dem pragmatischen Kriterium der 'Anpassung' ge-
lenkt sein solI, so heiBt es bei Duhem immer wieder: 'eine Unzahl ver-
schiedener theoretischer Tatsachen konnen als Ubersetzung derselben
praktischen Tatsache dienen'.10 Darum ermoglicht auch ein Widerspruch
im experimentellen Bereich nicht, eine physikalische Hypothese als falsch
oder wahr auszuzeichnen. Die einzige Oberpriifung einer Theorie, die Du-
hem als befriedigend erkHirt, ist der Vergleich 'des vollsHindigen Systems
der physikalischen Theorie mit der ganzen Gruppe experimenteller Tat-
sachen'.ll Es ist niemals moglich zu wissen, welche der Pramissen einer
vielschichtigen Theorie geandert werden muB, wenn ein experimenteller
Befund eine Theorie widerlegt. 12 Die Erklarung, die Poincare dafiir gibt,
ist einleuchtend. Der Kern seines Arguments besteht namlich in der These,
daB es durchaus der Fall sein konnte, daB die tatsachlichen Beziehungen
innerhalb eines bestimmten Bezugsbereichs durch eine Theorie richtig
erfaBt wiirden, daB aber ein Widerspruch 'nur in den Bildern liegt, deren
wir uns an Stelle der wirklichen Objekte bedient haben'P Hierin scheint
mir ein wesentlicher Aspekt der Theorienbildung iiberhaupt beriihrt, der
darin besteht, daB der Rahmen, innerhalb dessen theoretische Hypothesen
entworfen und prazisiert werden, der Rahmen einer Verstandlichkeitsan-
nahme ist, die den Gesamtzusammenhang der Phanomene eines bestimm-
ten Bereichs betrifft. In den meisten Fallen leitet uns hierbei zunachst die
natiirliche Weltauffassung, die Auffassung des Common sense und die vor-
gebildeten Pradikationsweisen der natiirlichen Sprache. Aber in vielen Be-
reichen 'verletzt' die wissenschaftliche Forschung ein solches Vorverstand-
nis und gelangt nicht zu einer Restitution des natiirlichen Weltbildes, aus
dem es zunachst seine Modelle bezieht. Darum ist die Trennung von Be-
deutung und Bezug oder - in Freges Terminologie - von Sinn und Bedeu-
tung fiir eine Rekonstruktion dieser Beziehungen fruchtbarer als die An-
wendung einer ungereinigten mengentheoretischen Terminologie. Darum
sollen, dem V orschlage von Duhem zufolge, nicht bloB die logischen Re-
geln, die mathematischen und physikalischen Prinzipien und Postulate bei
der Entscheidung iiber die Auswahl einer Hypothese angewendet und in
Betracht gezogen werden, sondern vor allem auch der gesunde Menschen-
348 RUDOLF HALLER
verstand (bon sense). Da es grundsatzlich unmoglich ist, eine Theorie aus

experimentellen oder Beobachtungsdaten, also allein induktiv aufzubauen,
weil -die Ableitung der theoretischen Satze immer auch andere theoretische
Satze voraussetzt oder impliziert, so muB jede Beobachtungstatsache auch
in bezug auf einen theoretischen Rahmen beurteilt werden. Darin liegt die
eigentliche Begriindung der holistischen Konzeption. Deshalb ist es sicher
eine Ubertreibung, wenn man behauptet, die Naturwissenschaft stelle fUr
den Konventionalisten 'kein Bild der Natur, sondem eine rein begriffiiche
Konstruktion' dar.14 Die Integration der begriffiichen Konstruktion ist
vielmehr geleitet durch Bilder, wie ich gerade erwahnt habe. Was vielmehr
verlangt wird - und eine solche Forderung findet man sowohl bei Poincare
wie bei Rey oder Duhem - ist, daB ein Gesetz, das als Definition des Ge-
genstandes, der gesetzartig ausgezeichnet wird, aufzufassen ist, immer
auch einen operativen Sinn haben muB. Mit anderen Worten, die Anwen-
dung des Gesetzes, etwa in MeBvorgangen, entscheidet iiber seine Brauch-
barkeit und Angemessenheit. Und richtigerweise haben die Konventio-
nalisten bereits sehr friih erkannt, daB nicht nur die Verifikation, sondem
auch die Falsifikation kein Auswahl- und daher auch kein Abgrenzungs-
kriterium bilden kann.
Wenn es jedoch in dem bedeutsamen Schritt der Theorienbildung, nam-
lich der Hypothesenauswahl, keinen logischen (deduktiven) und keinen
empirischen (induktiven) Zwang zu Bevorzugung gibt, wenn - wie in der
pragmatischen Wahrheitsauffassung - subjektive Einstellungen und Um-
stande kriterienbildend sind, dann muB der Willkiir in der Formation der
Theorien ein Platz eingeraumt werden. Gerade darin sehen viele die Gefahr
einer konventionalistischen Deutung wissenschaftlicher Forschung. 1m er-
sten Wiener Kreis und insbesondere in der Philo sophie Neuraths wurde
aber gerade dieser Gedanke als ein befreiender aufgenommen und beniitzt.
Warum?
Bevor ich zu dieser Frage im nachsten Abschnitt zuriickkehre, mochte
ich in aller Kiirze noch den dritten der friiher erwahnten Wissenschafts-
theoretiker behandeln, Abel Rey. Doch gilt diese Erinnerung nicht etwa
einem bahnbrechenden Denker, sondem jener Zusammenfassung, die in
besonderer Weise bei den Mitgliedem des ersten Wiener Kreises wirksam
war, vomehmlich aus dem Grunde, weil sie ein reiches Textmaterial aus
den Schriften zeitgenossischer Physiker mit einer klaren Disposition und
einer dezidierten Beurteilung verband. Viele der Themen und - was noch
erstaunlicher ist - viele jener Thesen, die im bekannten Wiener Kreis ver-
treten werden, finden sich bereits explizit im Werk Reys. Zunachst teilt er
mit Duhem die methodologische Pramisse, daB Wissenschaftstheorie ohne
Wissenschaftsgeschichte leer sei, weil der Wandel wissenschaftlicher Theo-
rien nur auf der Grundlage ihrer historischen Formen erfolgt und aus
diesen rekonstruiert werden kann. Zum zweiten vertritt Rey eine Theorie
der Einheit der Wissenschaften, die auch die Philo sophie in sich begreift
und ihre Aufgaben bestimmt.
Schon in dieser Darstellung aus dem Jahre 1905 begegnet uns der Ge-
gensatz zwischen einer kumulativen und einer nicht-kumulativen Interpre-
tation des Erkenntnisfortschrittes der exakten Wissenschaften in dem Ge-
wand des Gegensatzes des Mechanisten und des Konventionalisten. Und
Rey, der sich immer wieder von neuem mit den skeptischen Einwanden
gegen die Moglichkeit des tatsachlichen Fortschreitens physikalischer Er-
kenntnis auseinandersetzt, versucht zu zeigen, daB es ein- und derselbe
Bezugsbereich ist, ein- und dieselbe Referenz, auf den sich die beiden Meta-
theorien beziehen.
Der 'empirische Rationalismus', von dem Rey spricht, vereinigt in sich
die Machsche empiristische Anpassungstheorie einer Evolution der Er-
kenntnis mit der Idee eines neuen Rationalismus, dessen Grundlegung vor
allem in der Tatsache gesehen wird, daB objektive Erfahrung und Denken
sich wechselseitig bedingen oder, wie es wortlich heiBt, 'wechselseitig
Funktionen von einander' sind. 1sEine solche Konvergenz von zwei an-
tagonistischen Richtungen scheint schon deshalb notig, weil der Empiris-
mus, nach der Interpretation von Rey, 'notwendig zum Relativismus' fiih-
re. 16 Ein Grund hienur wird im Wesen der Hypothese selbst gesehen,
deren Revidierbarkeit zu ihrer Definition gehOrt. Alle Verallgemeinerun-
gen, d.h. alle induktiven Pramissen, sind durch Erfahrung gestiitzt und
konnen durch Erfahrung falsifiziert werden. 'Daraus folgt, daB sie nicht
mehr als definitiv und unveranderlich gelten; sie sind entwickIungsfahig,
wie Hypothesen, die, so fundamental sie auch sein mogen, sich immer
mehr prazisieren, vervollstandigen, verbessem. - Sie bleiben also standig
der Revision und Einschrankung offen.'17
Rey ist immer sehr kIar bemiiht, den historischen Punkt der Perspektive
anzugeben, an dem die modeme Naturwissenschaft angelegt ist, namlich
jenseits von Dogmatismus und Skeptizismus, den beiden Extremmoglich-
keiten, in die Rationalismus und Empirismus fiihren, oder abgleiten. Dar-
350 RUDOLF HALLER
urn sieht er im Strukturalismus, in einer Relationstheorie, die eigentliche

Aufgabe einer Metatheorie der exakten Wissenschaften. Rey bleibt inso-
fern Machianer, als er die Notwendigkeit, die sich in der Tatsache von
Beziehungen darzustellen scheint, auf Gewohnheiten zuriickfiihrt, die die
Evolution in uns ausgebildet habe. So werden das Rationale und das Em-
pirische als Momente der historischen Entwicklung begriffen.
Also auch von dieser Seite her waren die Mitglieder des ersten Wiener
Kreises sehr wohl mit den Problemen einer historischen Wissenschafts-
theorie bekannt, ein Faktum, das den Anti-Positivisten unserer Tage of-
fenkundig auch unbekannt geblieben ist.
III
Urn den Geist der Gruppe urn Hans Hahn wieder erstehen zu lassen, wird
es von Vorteil sein, vomehmlich nur jene Schriften in Betracht zu ziehen,
die der von mir behaupteten Abhangigkeit entstammen und auch var dem
Eintreffen Schlicks in Wien und var der Veroffentlichung von Wittgen-
steins logisch-philosophischer Abhandlung entstanden sind.
In einem grundlegenden Aufsatz iiber 'Die Bedeutung der physikali-
schen Erkenntnistheorie Machs fiir das Geistesleben der Gegenwart',18
stellt Ph. Frank, der 1912 auf die Lehrkanzel Einsteins nach Prag berufen
worden war, zunachst fest, daB der Hauptwert der Machschen Lehren
wesentlich darin bestiinde, das Gebaude der Physik nach auBen zu vertei-
digen, nicht aber darin, die physikalische Arbeit selbst zu befordem. Das
ist eine durchaus richtige Bemerkung iiber den Charakter wissenschafts-
oder grundlagentheoretischer Untersuchungen: Sie stellen keine Paralle-
luntersuchung zur objekttheoretischen Arbeit, sondem eine metawissen-
schaftliche Untersuchung der Forschung dar, die auch die erkenntnistheo-
retischen Fragen des jeweiligen Forschungsweges analysieren und klaren
sollte. Eben dieser Hauptwert der Machschen Lehre zeige sich am deut-
lichsten in dem Versuch, eine 'moglichst widerspruchsfreie Verbindung
zwischen Physik einerseits und der Physiologie und Psychologie anderer-
seits herzustellen', eine Verbindung, die es ermogliche, bei der Uberschrei-
tung der Grenzen der Disziplin nicht auch das Vokabular der Begriffe
vollig wechseln zu miissen. D.h. Frank sieht, nicht zu Unrecht, in dem
Versuch der Elementenlehre Machs, die Frank auch phanomenalistisch
nennt und deutet, einen Schritt zur Theorienvereinheitlichung, zur Ein-
heitswissenschaft. Urn es in dem 1ebendigen Bilde Machs zu sagen: 'Wie

das Blut, den Leib zu nahren, sich in zahllose Kapillaren teilt, urn dann
doch wieder im Herzen sich zu sammeln, so wird auch in der Wissenschaft
der Zukunft alles Wissen in einem einheitlichen Strom mehr und mehr
zusammenflieBen' .19 Dies ist der eine Grund der Bedeutung Machs fiir das
Geistes1eben der Gegenwart, den Frank hervorhebt. Der andere Grund
besteht, nach Frank, darin, daB Mach den Geist der Aufklarung im buch-
stablichen Sinne ins zwanzigste Jahrhundert hineingetragen habe, und sei-
ne Kritik an allen iiberfliissigen Wesenheiten es eigentlich gewesen sei, die
Mach auch an der Existenz von Atomen habe zweifeln lassen. Der Jiinger
Einsteins sieht natiirlich in Machs Kritik der Grundlagen der Newton-
schen Mechanik, in der Kritik an der Konzeption eines absoluten Raurnes,
ein Musterbeispiel jener Kritik aus dem Geiste des Okonomieprinzips, die
sich ja auch gegen die rein mechanistische Physik richtete. Und man wird
nicht umhin konnen, Frank zuzustimmen, wenn er die Machsche Skepsis
gegeniiber der Atomistik aus dessen Kritik der 'Hilfsbegriffe' des Materia-
lismus erklart. Indem namlich auch die urspriingliche Aufklarung verbun-
den war mit einer Theorie, deren 'HiIfsbegriffe', wie Materie und Atom,
ungereinigt waren, muBte der wahre aufkHirerische Geist auch diese Be-
griffe einer Kritik unterziehen. Dies war woh1 auch der Weg von Machs
groBartigen wissenschaftshistorischen Werken, von denen bier nur die
'Mechanik' und die friihe Schrift 'Die Gescbichte und Wurzel des Satzes
von der Erhaltung der Arbeit' (1872) erwahnt sei. Immer wieder hat Mach
eben die GrundbegrifJe der naturwissenschaftlichen Forschung, Raum,
Zeit, Substanz, Kausalitat u.a.m., einer Kritik unterzogen, die sie ihres
metaphysisch-mythologischen Mantels ebenso entkleiden, wie ihre Stel-
lung im Rahmen der naturwissenschaftlichen Theorien, d.h. der Psychop-
hysik, bestimmten sollte. Und wie spater der logische Empirismus beniitzt
auch Mach bereits ein sprachkritisches Argument zur Elimination der als
metaphysisch perhorreszierten Begriffe, namlich das Argument der Sinn-
losigkeit. Demzufolge ergibt die Anwendung eines pseudodeskriptiven
Ausdrucks in einem Urteil nicht einen falschen, sondem einen sinnlosen
Satz. Und aus dem gleichen Grunde kann auch schon die entsprechende
Frage als ein Scheinproblem entlarvt werden. Es ist vomehmlich der letz-
tere Aspekt, der Frank bei seiner Charakterisierung Machs als eines Auf-
klarungsphilosophen par excellence leitet: 'Die Aufgabe unseres Zeitalters
ist es nicht, die Aufklarung des 18. Jahrhunderts zu bekampfen, sondem
352 RUDOLF HALLER
ihr Werk fortzusetzen'.20 Dies also sind die beiden grundlegenden Ten-
denzen, die Frank Mach zuschreibt. Es sind aber auch die Postulate, die
fiir die wissenschaftliche Forschung als normativ gefordert werden. Und
Frank ist in diesem Sinne immer ein iiberzeugter Machianer geblieben.
Denn iiber seine wissenschaftstheoretische Position hinaus wird Mach
eben auch zum Herold der naturwissenschaftlichen Weltauffassung iiber-
haupt, wie sie Frank, Hahn und Neurath vertreten. Aufklarung und darum
Befreiung von der Metaphysik war und blieb das erste und ein wesentliches
Ziel der wissenschaftlichen Bemiihung.
Der Eindruck wird auch sogleich bestatigt, wenn man die Schriften Neu-
raths zu Rate zieht, insbesondere die umfangreichste Schrift, die er der
Geschichte des Wiener Kreises gewidmet hat. 21 Dort heiSt es: 'Die
Machsche und Einsteinsche Kritik der Newtonschen Physik und das neue
Denkgebaude, das daraus resultierte, haben in Wien eine ganz spezielle
Wirkung ausgeiibt. Philipp Frank stand schon als junger Physiker in di-
rektem Kontakt mit Mach und Einstein'. 22 Wie Frank hat Neurath die
Machsche Antimetaphysik als das scharfste Mittel zur Fortsetzung der
Aufklarung angesehen. Und auch in der Rechtfertigung der empiristischen
Grundeinstellung beruft sich Neurath durch aile Phasen seiner Entwick-
lung hindurch immer wieder auf Mach.23 Aber ebenso friih wie die An-
rufung der Autoritat Machs finden wir die Berufung auf den Konventio-
nalismus Poincares und Duhems, sowohl bei Frank wie bei Neurath. 24
Bereits in der friihen Arbeit von 1907 iiber das Kausaigesetz 2S formu-
Iiert Frank eine These, die er selbst expressis verbis als 'konventionali-
stisch' bezeichnet: 'Das Kausalgesetz,' - so heiSt es dort - 'das Fundament
jeder theoretischen Naturwissenschaft, laSt sich durch Erfahrung weder
bestatigen noch widerJegen, aber nicht aus dem Grunde, weil es eine a
priori denknotwendige Wahrheit ware, sondem weil es eine rein konven-
tionelle Festsetzung ist'. 26 Ais den realen Gehalt des Gesetzes schlagt
Frank dabei folgende Deutung vor: 'Wenn im Laufe der Zeit einmal auf
den Zustand A der Welt der Zustand B gefolgt ist, so folgt jedesmal, so
oft A eintritt, auch B darauf'.27 Mit anderen Worten, Frank akzeptiert
die Humesche Deutung der Kausalitat als Regularitat und postuliert ihre
Festsetzung als Forschungsregel der Naturwisschenschaft.
Das Zwei-Sprachen-Modell, das im Laufe der Entwicklung des logi-
schen Empirismus auf verschiedene Weise variiert wurde, findet sich schon
an dieser Stelle vorbereitet: Die beiden Sprachen werden namlich mit den
beiden Zweigen der Naturwissenschaft identifiziert. Wahrend die experi-

mentelle Naturwissenschaft 'die sinnlich gegebenen Eigenschaften der
Korper und deren Veranderung beschreibt, besteht die Aufgabe der theo-
retischen Naturwissenschaft darin, die Korper mit den fiktiven Eigenschaf-
ten auszustatten, die erst die Geltung des Kausalgesetzes sichern, die theo-
retische Naturwissenschaft ist nicht Forschung, sondern gewissermaBen
Umdichtung der Natur, sie ist Phantasiearbeit'. 28 Hier wird deutlich,
welche Gestalt die Kombination der Machschen Lehren mit jenen der
franzosischen Konventionalisten annimmt. Die freie Setzung der Annah-
men wird der Konzeption des kritischen Idealismus Kants insofern ver-
gleichbar, als beide einen vorausgesetzten Rahmen der Erfahrung behaup-
ten, nur daB die kantische Richtung diesen der Natur des Verstandes zu-
schreibt, wahrend Frank ihn als ein Erzeugnis der menschlichen Willkiir
behauptet. Die Machsche Grundidee, daB die Annahmen, Schematisie-
rungen und Idealisierungen letztlich Phantasieprodukte sind und Wissen-
schaft insofern der gleichen Quelle entstammte wie die iibrigen kulturellen
Leistungen des Menschen, wird mit dem Nachdruck, den die Konventio-
nalisten darauf legen, als ein logisches Willkiirprodukt identifiziert. Das
heiBt, die Natur als Gegenstand der theoretischen Wissenschaft 'ist nicht
die empirische Natur, sondern die fiktive Natur', die sonach nicht nur
Gegenstand menschlicher Neugierde, sondern Produkt des empirischen
Handelns ist. Eben darum ist sie fUr ihn restlos durchschaubar, erfaBbar.
Die konstruktiven Ziige der Theorie, deren Eindeutigkeit erzeugt und er-
zeugbar ist, spiegeln sich freilich nicht im Experiment, in der Erfahrung.
Darum ist der Spielraum der Erfahrung immer groBer als der Spielraurn
der Theorie. Die so erzeugten Weltbilder sind darum auch niemals wahr
oder falsch, sondern, je nach ihrer Gestalt, ihrer Struktur, komplizierter
oder einfacher. Und wie iiberhaupt im Konventionalismus tritt auch hier
das Prinzip der Einfachheit an die Stelle der Wahrheit.
In ahnlicher Weise spricht Wittgenstein im Traktat davon, daB wir im-
mer die Moglichkeit haben, eine beliebige Form einer Beschreibung zu
wahlen und daB die eine einfacher sein mag als die andere. Und er nennt
diese Formen 'Netze' oder 'Systeme' der Weltbeschreibung. Und das Kau-
salgesetz ist dann eben kein Gesetz, das sich auf das Gegenstandliche be-
zieht, sondern etwas, das sich darin zeigt, daB es Naturgesetze, also geset-
zartige Zusammenhange, gibt. 29 Und auch Wittgenstein scheint urn diese
Zeit die zufallige Aprioritat eines solchen Netzes als den Unterschied zur
354 RUDOLF HALLER
sonst ahnlichen kantischen Auffassung hervorheben zu miissen.

In seinem philosophischen Hauptwerk, das fUnfundzwanzig Jahre nach
dem Aufsatz iiber das Kausalgesetz erscheint, korrigiert Frank seinen
Hinweis auf Ahnlichkeiten mit der kantischen Position. Dort heiBt es nun:
'Wenn ich schon damals klar erkannt batte, daB die Lehren der soge-
nannten idealistische Philosophie, auch der transzendentalen, keine wis-
senschaftlichen Behauptungen sind, da sie weder tautologisch sind, noch
iiber die wirkliche Welt etwas aussagen, sondern Behauptungen iiber eine
jedem menschlichen Erkennen verborgene jenseitige Welt sind, so hatte ich
auf die Analogie zwischen der konventionalistischen und der kantischen
Auffassung der Kausalitat sicher weniger Gewicht gelegt. Die Analogie
besteht sicher und zwar darin, daB nach beiden Auffassungen das Kau-
salgesetz nichts iiber die empirische Welt aussagt, sondern iiber die Art,
wie der Mensch die Welt auffaBt'. Und dann fahrt Frank in der fUr den
eigentlichen Wiener Kreis typischen Auffassung fort: 'Es ist aber wichtig,
wenn man eine den wissenschaftlichen Fortschritt fordernde Stellung ein-
nehmen will, vor allem den Gegensatz zu der Kantischen Auffassung des
Kausalgesetzes hervorzuheben'. 30
Es ist an dieser Stelle nicht moglich, diesen Grunden des naheren nach-
zugehen, obschon der hauptsachliche Unterschied zwischen einer konven-
tionalistischen und der kantischen Interpretation fUr sich eine Untersu-
chung erforderte. Was fUr den gegenwartigen Zweck notwendig ist, wurde
bereits ausgesprochen, namlich, daB in der weiteren Entwicklung wohl eine
vermeintliche oder wirkliche Ahnlichkeit mit Kant korrigiert wird, nicht
aber der Kern der konventionalistischen These, daB namlich den allge-
meinen Gesetzen der Charakter von Tautologien zukommt, die einfach
sein und eine iibersichtliche Darstellung ermoglichen sollen. Darum wer-
den sie auch als dem jeweiligen Stande der Forschung adaptierbar und
somit als veranderbar bestimmt.
Neurath hat nun schon sehr fruh das Prinzip dieser Veranderung for-
muliert. Und obschon er weitgehend den Spuren von Duhem und Poincare
folgt, ist er darin m.E. urn einen wesentlichen Schritt weitergegangen als
die iibrigen Konventionalisten. Wohl hatten die beiden Franzosen imrner
wieder betont, daB ein Gesetz nicht eigentlich mit der Erfahrung kollidie-
ren konne, weshalb es - jedenfalls nach Duhem - auch kein experimentum
crucis geben kann. 31 Und wohl haben beide dennoch die Veranderungs-
moglichkeit eines jeden theoretischen Systems behauptet, das mit der Er-
fahrung nicht in Harmonie stiinde. Aber bei keinem der beiden Denker ist
die Dichotomie der beiden Moglichkeiten so klar ausgesprochen wie bei
Neurath, weshalb ich das Prinzip der Adaption im Theorienwandel auch
Neurath-Prinzip genannt habe. 32 Man kann dieses Prinzip nur begreifen,
wenn man die ganzheitliche Interpretation wissenschaftlicher Systeme be-
greift, die Neurath offenkundig mit allen Argumenten von Duhem iiber-
nimmt. Da es keine eindeutige Beschreibung von Erfahrungstatsachen in
bezug auf eine Theorie geben kann, ist die Auszeichnung des isolierten
Satzes als Falsifikationsinstanz abzulehnen. Ebenso besteht keine
Moglichkeit der Anwendung eines Kriteriurns zur Auswahl einer Hypo-
these in einem experimenturn crucis. Da zudem der ganzheitlichen Theo-
rienauffassung Duhemscher Provenienz nach ein wissenschaftliches Sys-
tem immer nur in toto akzeptiert oder abgelehnt werden kann, muBte die
Frage auftreten, wie die Abanderung denn moglich sei, bei einem imagi-
nierten oder tatsachlichen Konflikt von Erfahrungssatz und System. Und
die verblUffend selbstverstandliche Devise Neuraths besagt, daB entweder
das System geandert werden kann (das war die Meinung der franzosischen
Konventionalisten) oder der Satz selbst, weil beide den gleichen Rang hat-
ten. Interpretiert man dieses Prinzip durch den Theoriennaturalismus und
die Perspektive Quines, dann ergibt sich, daB die Anderungen an der Pe-
ripherie des Systems meistens die naherliegenden sind, daB aber letzten
Endes 90ch die Erfahrungssatze es auch sind, die als 'Eckpfeiler der Se-
manp.k' das System tragen helfen und ihnen daher der Vorrang einzurau-
men ist.
Wie bereits deutlich geworden ist, brachen diese Probleme nicht nur in
der beriihmten Protokollsatzdebatte der frUben DreiBigerjahre aufs neue
auf, sondem sie begleiten die Interpretation der Theorienstruktur bis auf
den heutigen Tag.
Ohne Zweifel aber hat die Herausbildung eines konventionalistischen
Empirismus in den Jahren des ersten Wiener Kreises dazu gefiihrt, daB
das Problem der Theorieninterpretation dort ansetzte, wo der theoretische
Standpunkt der Vertreter des ersten Wiener Kreises gelegen war.
Als dann Hans Hahn aufdem Umweg iiber Czemowitz und Bonn 1921
wieder nach Wien zuriickgekehrt war, traf sich die alte Gruppe, die auch
in ihren politischen Oberzeugungen einander h6chst nahestanden, aufs
neue. Und es istja inzwischen oft betont worden, wie der Kreis urn Schlick
sich im wesentlichen aus den Mitgliedem des mathematischen Seminars
von Hahn rekrutierte. 33
356 RUDOLF HALLER
Es ist meine Meinung, daB die philosophischen Spriinge im Wiener

Kreis, wie sie aus AniaB der Protokollsatzdebatte ausbrachen, nicht, wie
neuerdings geme suggeriert wird, politischen Fronten im Wiener Kreis
entsprangen (die gab es wohl von Anfang an), sondem der Wiederaufnah-
me der konventionalistischen Probleme, die durch die Diskussion des
Traktats und der damit verbundenen Erkenntnistheorie in den Hinter-
grund getreten waren. Aber iiber die Polarisierung der Wissenschaftsauf-
fassung im bekannten Wiener Kreis zu sprechen, war nicht die Absicht
dieser Abhandlung. 34 Es geniigt, wenn deutlich wird, was das Kemgeriist
der wissenschaftstheoretischen Auffassungen des ersten Wiener Kreises
ausmacht. Hat man dieses einmal erkannt, wird man es auch anderswo
wieder erkennen konnen.
ANMERKUNGEN
1 E. Mohn, Der logische Positivismus - Theorien und politische Praxis seiner Vertreter, Frank-
furt, 1977; R. Hegselmann, 'Otto Neurath - Empiristischer Aufkliirer und Sozialreformer',
in R. Hegselmann (Hrsg.), Otto Neurath, Wissenschaftliche Weltauffassung, Sozialismus und
logischer Empirismus" Frankfurt, 1979.
2 P. Achinstein & St. F. Barker (eds.), The Legacy of Logical Positivism, Baltimore, 1969; H.
L. Mulder, R. S. Cohen, B. McGuinness (eds.), Vienna Circle Collection, Dordrecht, 1973
fr. (bisher 15 Biinde).
3 Siehe O. Neurath, Gesammelte philosophische und methodologische Schriften, hrsg.v. R.
Haller und H. Rutte, Wien: 1981, Bd. 1, Vorwort, p. XIII; R. Haller, 'New Light on the
Vienna Circle', The Monist 65 (1982), 25-37; Ders., 'Bemerkungen zum Problem des ku-
mulativen Wissens', in K. Freisitzer und R. Haller (Hrsg.), Probleme des Erkenntnisfort-
schritts in den Wissenschaften, Wien: 1977, pp. 6-23; Ders., 'Das Neurath-Prinzip - Grund-
lagen und Foigerungen', in Arbeiterbildung in der Zwischenkriegszeit, hrsg.v. F. Stadler,
Wien: 1982, pp. 79-87.
4 Ph. Frank, Modern Science and its Philosophy, Collier Books, New York, 1961, p. 13.
S Man bedenke, daB die erste Obersetzung Duhems ins Englische (durch Ph. G. Wiener) erst
1954 folgt! Und die friihesten Arbeiten von J. Agassi, A. Griinbaum, N. R. Hanson, die sich
mit Duhem auseinandersetzten, erschienen gieichfalls erst Ende der Fiinfzigerjahre. In den
Darstellungen der Geschichte des Positivismus hat zumindest L. Kolakowski, Die Philosophie
des Positi~ismus, dt. Miinchen, 1971, den Beitrag des Konventionalismus in einem eigenen
Abschnitt gewiirdigt.
6 Siehe H. Poincare, Wissenschaft und Hypothese, dt. v. F. u. L. Lindemann, Leipzig, 19062 ;
siehe auch R. Haller, 'Als ob - Zu Stephan K6mers "Idealisierungen",' Grazer Philosophische
Studien 20 (1983), 117-27.
7 Vgl. R. Haller, 'Poetic Imagination and Economy. E. Mach as Theorist of Science', in J.
Agassi & R. S. Cohen (eds.), Scientific Philosophy Today, Dordrecht, 1981. Siehe auch A.
Rey, Die Theorie der Physik bei den modernen Physikern, Leipzig, 1908, besonders pp. 68-
109.
8Vgl. E. Mach, 'Vorwort' zur deutschen Ausgabe von P. Duhem, Ziel und Struktur der
physikalischen Theorien. Obersetzt v. F. Adler, Leipzig, 1908, neu hrsg.v. L. Schiifer, Ham-
burg, 1978, p. IV; siehe auch E. Mach, Erkenntnis WId [rrtum, Leipzig, 1905, p. 447 u.o.
9 Vgl. P. Duhem, op. cit., p. 296. Vgl. Abel Rey, Die Theorie der Physik bei den modernen
Physikem, pp. 122 fr. Vgl. A. Griinbaum, 'The Duhemian Argument', Philosophy of Science
27 (1960), 75-87; C. Giannoni, 'Quine, Griinbaum and the D-Thesis', Nolls 1 (1967), 283 fr.;
S. G. Harding (ed.), Can Theories Be Refuted? Essays on the Duhem-Quine-Thesis, Dordrecht,
1976 (Synthese Library, Vol. 81); W. Diederich, Konventionalitiit in der Physik, Berlin, 1974;
H. Schniidelbach, Erfahrung, Begriindung WId Reflexion, Frankfurt, 1971, pp. 165 fr.
10 P. Duhem, Ziel WId Struktur der physika/ischen Theorien, p. 175.
11 P. Duhem, ibid., p. 267.
12 H. Poincare, Wissenschaft WId Hypothese, p. 153.
13 H. Poincare, ibid., p. 164.
14 Vgl. K. Popper, Logik der Forschung, Tiibingen, 19662 , p. 48.
15 A. Rey, Die Theorie der Physik bei den modernen Physikem, dt. v. R. Eisler, Leipzig, 1908,
p. 363~
16 A. Rey, ibid., p. 308.
17 A. Rey, ibid., p. 319.
18 Ph. Frank, 'Die Bedeutung der physikalischen Erkenntnistheorie Machs fUr das Geistes-
leben der Gegenwart', in Die Naturwissenschaften (hrsg.v. A. Berliner und A. Piitter), 5. Jg.,
H. 5 (1917), pp. 65--72.
19 E. Mach, 'Ober den EinfiuB zufiilliger Umstiinde auf die Entwicklung von Erfindungen
und Entdeckungen', in: E. Mach, Populiir-wissenschaftliche Vorlesungen, Leipzig 1896, p.
277.
20 Ph. Frank, 'Die Bedeutung der physikalischen Erkenntnislehre', p. 71.
21 O. Neurath, Le developpement du Cerc/e de Vienne et l'avenir de l'Empirisme logique (1936),
dt. in O. Neurath, Gesammelte philosophische WId methodologische Schriften (kiinftig abge-
kiirzt G.S.), pp. 674-702.
22 Ebd., p. 695.
23 Siehe auch: 'Zwei Briefe von Otto Neurath an Ernst Mach', hrsg.v. R. Haller, (Hrsg.),
Grazer Philosophische Studien 10 (1980), 3-5. Wie mir erst nachtriiglich bekannt wird, sind
die Briefe bereits bei J. Thiele, Wissenschaftliche Kommunikation. Die Korrespondenz E.
Machs, Kastellaun, 1978, publiziert. Vgl. neuestens auch die materia1reiche, historische Dar-
stellung von F. S~dler, Vom Positivismus zur 'Wissenschaftlichen Weltauffassung', Wien,
1982, pp. 111-125.
24 Wie man dem Register der Neurath-Ausgabe der G.S. entnehmen kann, sind - von seinen
Zeitgenossen und ihm selbst abgesehen - Mach, Poincare und Duhem auch die am oftesten
zitierten Namen.
25 Ph. Frank, 'Kausalgesetz und Erfahrung', in Annalen der Naturphilosophie (hrsg. v. W.
Ostwald), Bd. 6 (1907), pp. 443-450.
26 Ebd., p. 444.
27 Ebd.; vgl. dazu R. Haller, 'Bedingungen, Voraussetzungen, Verursachungen', in Urteile
WId Ereignisse, Freiburg-Miinchen, pp. 76 fr.
28 Ph. Frank, 'Kausalgesetz und Erfahrung', p. 447.
29 Vgl. L. Wittgenstein, Tractatus logico-philosophicus, 6.32 bis 6.361.
30 Ph. Frank, Das Kausalgesetz WId seine Grenzen, Wien, 1932 (Schriften zur wissenschaft-
lichen Weltaufl"assung, hrsg. v. Ph. Frank und M. Schlick, Bd. 6), p. 242. Vgl. auch W.
Stegmiiller, 'Das Problem der Kausalitiit', in Probleme der Wissenschaftstheorie. Festschrift
for Victor Kraft, hrsg. v. E. Topitsch, Wien, 1960.
358 RUDOLF HALLER
31 Vgl. H. Poincare, Wissenschaft und Hypothese, pp. 107, 181 f, 138 ff, 153 u,o.; P. Duhem,
Ziel und Struktur der physikalischen Theorien, pp. 250 f, 342 f.
32 Vgl. R. Haller, 'Geschichte und wissenschaftliches System bei Otto Neurath', in H. Berghe1
et al. (eds.), Wittgenstein, der Wiener Kreis und der Kritische Rationalismus. Akten der 3. Int.
Wittgenstein Symposiums (1978), Wien, 1979, pp. 302 ff; R. Haller, 'Das Neurath-Prinzip
- Grund1agen und Fo1gerungen', a.a.O.
33 Vgl. den Nachruf Ph. Franks auf H. Hahn in Erkenntnis 4 (1934),315-316, wo Hahn als
der 'eigentliche Begriinder des "Wiener Kreises" , bezeichnet wird, sowie R. Haller, 'New
Light on the Vienna Circle', a.a.O. Ferner: K. Menger, 'Introduction' to Hans Hahn, Em-
piricism, Logic, and Mathematics, ed. B. McGuiness, Dordrecht, 1980, pp. IX f.
34 R. Haller, 'Zwei Arten von Erfahrungsbegriindung', in Grazer Philosophische Studien
16/17 (1982).
Received 20 March 1984
Karl-Franzens-Universitat Graz
Institut fUr Phi1osophie
Heinrichstrasse 26
A-8010 Graz
Osterreich
NICHOLAS RESCHER
ARE SYNOPTIC QUESTIONS ILLEGITIMATE?
Questions like 'Why is there anything at all?' 'Why are things-in-general

as they actually are?' 'Why is the law structure of the world as it is?' cannot
be answered within the standard causal framework. For causal explana-
tions need inputs: they are essentially transformational (rather than for-
mational pure and simple). They can address themselves to specific issues
distributively and seriatim, but not collectively and holistically. If we per-
sist in posing the sorts of global questions at issue, we cannot hope to
resolve them in orthodox causal terms. Does this mean that such questions
are improper?
On the rejectionist approach, the entire question of obtaining the (or a)
reason for the existence of things is simply dismissed as illegitimate. Even
to inquire into the existence of the entire universe is held to be somehow
illegitimate. It is just a mistake to ask for a causal explanation of existence
per se; the question should be abandoned as improper - as not representing
a legitimate issue. We are assured that in the light of closer scrutiny the
explanatory 'problem' vanishes as meaningless.
Dismissal of the problem as illegitimate is generally based on the idea
that the question at issue involves an illicit presupposition. It looks to
answers of the form 'Z is the (or an) explanation for the existence of
things.' Committed to this response-schema, the question has the thesis
'There is a ground for the existence of things - existence--in-general is the
sort of thing that has an explanation.' And this presumption - we are told
- might well be false. In principle its falsity could emerge in two ways:
1. on grounds of deep general principle inherent in the concep-
tual 'logic' of the situation, or
2. on grounds of a concrete doctrine of substantive metaphysics
or science that precludes the prospect of an answer - even as
quantum theory precludes the prospect of an answer to 'Why
did that atom of Califorium decay at that particular time?'
Let us begin by considering if the question of existence might be invali-
Erkenntnis 22 (1985) 359-363. 016HI06j85j0220-0359 SOO.50

360 NICHOLAS RESCHER
dated by considerations of the first sort and root in circumstances that lie
deep in the conceptual nature of things. Consider the following discussion
of e.G. Hempel's:
Why is there anything at all, rather than nothing? .. But what kind of an answer could be
appropriate? What seems to be wanted is an explanatory account which does not assume the
existence of something or other. But such an account, I would submit, is a logical impossi-
bility. For generally, when the question 'Why is it the case that A? is answered by 'Because
B is the case' ... [AJn answer to our riddle which made no assumptions about the existence of
anything cannot possibly provide adequate grounds ... The riddle has been constructed in a
manner that makes an answer logically impossible... 1
But this plausible line of argumentation has shortcomings. The most

serious of these is that it fails to distinguish appropriately between the
existence of things on the one hand and the obtaining offacts on the other, 2
and supplementarily also between specifically substantival facts regarding
existing things, and non substantival facts regarding states of affairs that
are not dependent on the operation of preexisting things.
We are confronted here with a principle of hypostatization to the effect
that the reason for anything must ultimately always inhere in the opera-
tions of things. And at this point we come to a prejudice as deep-rooted
as any in Western philosophy - the idea that things can only originate
from things, that nothing can come from nothing (ex nihilo nihil fit) in the
sense that no thing can emerge from a thingless condition. 3 Now this some-
what ambiguous principle is perfectly unproblematic when construed as
saying that if the existence of something real has a correct explanation at
all, then this explanation must pivot on something that is really and truly
so. Clearly, we cannot explain one fact without involving other facts to do
the explaining. But the principle becomes highly problematic when con-
strued in the manner of the precept that 'things must come from things',
that substances must inevitably be invoked to explain the existence of sub-
stances. For we now become committed to the thesis that everything in
nature has an efficient cause in some other natural thing that is its causal
source, its reason for being.
This stance is implicit in Hempel's argument. And it is explicit in much
of the philosophical tradition. Hume, for one, insists that there is no fea-
sible way in which an existential conclusion can be obtained from non-
existential premisses. 4 And the principle is also supported by philosophers
of a very different ilk on the other side of the channel - including Leibniz
himself, who writes:
SYNOPTIC QUESTIONS 361
The sufficient reason (of contingent existence)... must be outside this series of contingent
things, and must reside in a substance which is the cause of this series ... 5
Such a view amounts to a thesis of genetic homogeneity which says (on

analogy with the old but now rather obsolete principle th~t 'life must come
from life') that 'things must come from things', or 'stuff must come from
stuff', or 'substance must come from substance'.
What, after all, could be more plausible than the precept that only real
(existing) causes can have real (existing) effects? But despite its appeal,
this principle has its problems. It presupposes that there must be a type-
homogeneity between cause and effect on the lines of the ancient Greek
principle that 'like m~st come from like'. This highly dubious principle of
genetic homogeneity has taken hard knocks in the course of modem
science. Matter can come from energy, and living organisms from com-
plexes of inorganic molecules. If the principle fails with matter and life,
need it hold for substance as such? The claim that it does so would need
a very cogent defense. None has been forthcoming to date.
Is it indeed true that only things can engender things? Why need a
ground of change always inhere in a thing rather than in a nonsubstantival
'condition of things-in-general'? Must substance inevitably arise from sub-
stance? Even to state such a principle is in effect to challenge its credentials.
For why must the explanation of facts rest in the operation of things? To
be sure, fact--explanations must have inputs (all explanations must). Facts
must root in facts. But why thing--existential ones? A highly problematic
bit of metaphysics is involved here. Dogmas about explanatory homo-
geneity aside, there is no discernible reason why an existential fact cannot
be grounded in nonexistential ones, and the existence of substantial things
be explained on the basis of some nonsubstantival circumstance or prin-
ciple. Once we give up the principle of genetic homogeneity, and abandon
the idea that existing things must originate in existing things, we remove
the key prop of the idea that asking for an explanation of things in general
is a logically inappropriate demand. The footing of the rejectionist ap-
proach is gravely undermined.
There are, of course, other routes to rejectionism. One of them turns on
the thesis of Kant's First Antinomy that it is illegitimate to try to account
for the phenomenal universe as a whole (the entire Erscheinungswelt). Ex-
planation on this view is inherently partitive: phenomena can only be ac-
counted for in terms of other phenomena, so that it is in principle improper
362 NICHOLAS RESCHER
to ask for an account of phenomena-as-a-whole. The very idea of an

explanatory science of nature-as-a-whole is illegitimate. Yet this view is
deeply problematic. To all intents and purposes, science strives to explain
the age of the universe-as-a-whole, its structure, its volume, its laws, its
composition, etc. Why not then its existence as well? The decree that ex-
planatory discussion is necessarily partial and cannot deal with the whole
lacks plausibility. It seems a mere device for sidestepping embarrassingly
difficult questions.
Rejectionism is not a particularly appealing course. Any alternative to
rejectionism has the significant merit of retaining for rational inquiry and
investigation a question that would otherwise be abandoned. The question
of 'the reason why' behind existence is surely important. If there is any
possibility of getting an adequate answer - by hook or by crook - it seems
reasonable that we would very much like to have it. There is nothing
patently meaningless about this 'riddle of existence'. And it does not seem
to rest in any obvious way on any particularly problematic presupposition
- apart from the epistemically optimistic yet methodologically inevitable
idea that there are always reasons why things are as the are (the 'principle
of sufficient reason'). To dismiss the question as improper or illegitimate
is fruitless. Try as we will to put the question away, it comes back to haunt
us. 6
NOTES
1 Carl G. Hempel, 'Science Unlimited', The Annals of the Japan Associationfor Philosophy
of Science 14 (1973), pp. 187-202. (See p. 200.) Our italics.
2 Note too that the question of the existence of facts is a horse of a very different color
from that of the existence of thing. There being no things is undoubtedly a possible situation,
there being no facts is not (since if the situation were realized, this would itself constitute a
fact).
3 Aristotle taught that every change must emanate from a 'mover', Le., a substance whose
machinations provide the cause of change. This commitment to causal reification is at work
in much of the history of Western thought. That its pervasiveness is manifest at virtually
every juncture is clear from William Lane Craig's interesting study of The Cosmological
Argumentfrom Plato to Leibniz (London, 1980).
4 David Hume, Dialogues Concerning Natural Religion (ed. N.K. Smith), London, 1922,
p.189.
S G.W. Leibniz, 'Principles of Nature and of Grace', sect. 8, italics supplied. Compare St.
Thomas:
Of necessity, herefore anything in process of change is being change by something else.

(S.T., IA 2, 3).
SYNOPTIC QUESTIONS 363
The idea that only substances can produce changes goes back through Thomas' master,
Aristotle. In Plato and the Presocratics, the causal efficacy of principles is recognized (e.g.,
the love and strife of Empedocles).
6 For criticisms of ways of avoiding the question 'Why is there something rather than
nothing?' see Chap. III of William Rowe, The Cosmological Argument, Princeton, 1975. Cf.
also Donald R. Burrill (ed.), The Cosmological Argument, Garden City, 1967, esp. 'The Cos-
mological Argument' by Paul Edwards.
Received 12 February 1984
University of Pittsburgh,
1012 Cathedral,
Pittsburgh, PA 15260,
U.S.A.
W. K. ESSLER
ON DETERMINING DISPOSITIONS
Determining the qualitative, comparative and quantitative properties of

physical objects is widely regarded as the result of operations, i.e. of the
application of procedures to n-tuples (n)O) of elements of the universe.
Thus e.g., according to Hempel l , some x has less mass than some y if and
only if it is true that, when the objects are placed into opposite pans of a
balance in a vacuum, x rises and y sinks; and some mineral x is harder
than another one y if (and only if) a sharp point of x scratches a smooth
surface of y.
In analyzing the logical structure of procedures of this kind we need
open sentences A(xt. ... ,Xn,Zt. .. "zm) and B(xt. ... ,Xn,Zt. .. "zm), where the
Zt. •• ',Zm (m)O) may refer to pans, balances, vacua, sharp points, smooth
surfaces, etc.;2 sentence A describes the test condition and sentence B the
result. For the sake of simplicity, let n = m = 1. Then Camap's initial analy-
sis of this operational procedure 3 consists of a total definition of disposi-
tions:
(TDD) /\x[x E F +-+ /\z(A(x,z) -+ B(x,z))].
As Camap4 discovered later, this scheme, though satisfying the conditions

of total eliminability and noncreativity and being therefore formally cor-
rect, suffers serious disadvantages: for non-tested objects x, i.e. for objects
x satisfying ,V z A(x,z), the definiens and therefore also the definiendum
of (TDD) turns out to be true, whether or not this object has that disposi-
tion with regard to common - or common scientific - opinion.
In analyzing the procedure of operationalizing concepts Camap there-
fore suggested the following pair of reduction sentences:
(PRS) /\x/\z[ A(x,z) /\ B(x,z) -+ X E 1'] /\
/\x/\z[A(x,z) /\ IB(x,z) -+ IX E 1']
being equivalent to the bilateral reduction sentence:

(BRS) /\x/\z[A(x,z) -+ (x E F +-+ B(x,z)].
But this new scheme too brings about some difficulties, Let, e.g., F des-
Erkenntnis 22 (1985) 365-368. 0165--0106/85/0220-0365 $00.40
366 w. K. ESSLER
ignate the property of being magnetic and x some piece of soft iron being
within a magnetic field at time Zl but not at Z2; let A describe some test
situation realized at Zl as well as at Z2, and let B designate the correspond-
ing positive result. Then the statements A(X,Zl), A(X,Z2), B(X,Zl) and
-,B(X,Z2) tum out to be empirically true but entail, via (BRS), the con-
tradiction x E F 1\ - , X E F. Therefore neither (BRS) nor (PRS) are ad-
equate operational definitional schemes of concepts.
In analyzing the counterexamples to (TDD) we observe that for non-
tested objects not only I\z(A(x,z) -+ B(x,z» but also the contrary universal
implication I\z(A(x,z) -+ -, B(x,z» is true. But -, VzA(x,z) is logically
equivalent to the conjunction of these two universal implications therefore
expressing their content. Thus it is reasonable to restrict (TDD) to cases
where not both statements are true together, i.e. to introduce the concept
F by the following partial definition of a disposition:
(PDD) I\x[VzA(x,z) -+ [x E F +-+ I\z(A(x,z) -+ B(x,z»]].
This scheme satisfies the conditions of partial eliminability and of non-
creativity and may in addition be used to divide (BRS) into an analytic
and a synthetic part, since (BRS) is logically equivalent to a statement
(PDD) 1\ (GUA), where (GUA) formulates a general uniformity assump-
tion:
(GUA) I\x[Vz(A(x,z) 1\ B(x,z» -+ I\z(A(x,z) -+ B(x,z»].
Obviously, (GUA) reformulates the background presuppositions of classi-
cal mechanics that already one experiment being carefully realized and
ending with a positive result yields an empirical law. Therefore, if the
definiens of (PDD) is obtained via (GUA), then Vz(A(x,z) 1\ B(x,z» en-
tails x E F, and Vz(A(x,z) 1\ -,B(x,z» accordingly -,x E F.
But in spite of the fact that (PDD) is not synthetic it is too strong. For
the definiens of (PDD) may also be obtained without using (GUA), as the
following examples shows. Suppose that x designates some piece of soft
iron which was tested at some time z 1 with positive result without being
tested a second time in past, present or future, so that
(*) VZ1(A(x,Zl) 1\ B(X,Zl) 1\ -,VZ2(Z2 "# Zl 1\ A(X,Z2»)
is empirically true. Then VzA(x,z) as well as I\z(A(x,z) -+ B(x,z» are true
and therefore also x E F, asserting wrongly that this object is permanently
magnetic.
ON DETERMINING DISPOSITIONS 367
Therefore the realm of applications of (PDD) needs additional restric-

tions. Since (PDD) /\ (GUA) is logically equivalent to [(GUA) -+ (PDD)]
/\ (GUA), the restricted definition of disposition has the form:
(RDD) I\x[Vz(A(x,z) /\ B(x,z)) -+ I\z(A(x,z) -+ B(x,z))] -+
I\x[VzA(x,z) -+ [x E F +-+ I\z(A(x,z) -+ B(x,z))]].
Obviously, the condition (GUA) of (RDD) works as an extensional sub-
stitute of the statement asserting that the class of tests of some object of
the universe is necessarily uniform with regard to the result; for, even if in
some cases these classes may be of small cardinality, we usually take it for
granted that there exist other elements of our factual universe who were
tested at least k times, k being sufficiently larger. If we accept, at least as
an idealization, that the numbers of these classes though being finite, have
no upper bound, the condition (GUA) cannot be satisfied in a non-lawlike
manner6.
To see how (GUA) works, let us suppose that it is weakened by universal
instantiation, which leads to an assertion stronger than (R'bD):
I\x[[Vz(A(x,z) /\ B(x,z)) -+ I\z(A(x,z) -+ B(x,z))] -+
[VzA(x,z) -+ [x E F +-+ I\z(A(x,z) -+ B(x,z))]]].
Then, for some fixed x, its conditions
[Vz(A(x,z) /\ B(x,z)) -+ I\z(A(x,z) -+ B(x,z))] /\ VzA(x,z)
being logically equivalent to the contravalence
I\z(A(x,z) -+ B(x,z)) ~ I\z(A(x,z) -+ -,B(x,z))
is entailed again by (*) so that the uniformity of the class of tests with
regard to the results may be created by chance also here. Adding the con-
dition that there is some natural number k) 1 such that at least k tests were
realized, all of them leading to positive results, obviously does not solve
the problem, since these k results may also have happened by chance.
Moreover it causes new problems, since it need not be true for every ele-
ment of the universe that it is tested at least k times; thus this condition
restricts the domain of application of this stronger scheme without neces-
sity.
This suggests that the conditional definition (RDD) should be regarded
as an optimal scheme for introducing dispositional concepts as well as all
368 W. K. ESSLER
other kinds of concepts to be operationally characterized in science and

everyday life.
NOTES
1 Cf. C. G. Hempel, Fundamentals o/Concept Formation in Empirical Science, Chicago, 1952,

pp. 60 f.
2 If the property of an object is changing in time we represent it by a binary relation R
between objects and time points; if the property is permanent, one of the z/ will vary over
time points. Cf. also W. K. Essler and R. Trapp, 'Some Ways of Operationally Introducing
Dispositional Predicates with regard to Scientific and Ordinary Practice', Synthese 34 (1977),
371-396, and W. K. Essler, 'Some Remarks Concerning Partial Definitions in Empirical
Sciences', Pacific Philosophical Quarterly 61 (1980),455-462, cf. also W. K. Essler, Wissen-
scha/tstheorie I, 21982, §§ 19, 20.
3 Cf. R. Camap, Der logische Aujbau der Welt, Berlin, 1982, pp. 30-32, 201.
4 Cf. R. Camap, 'Testability and Meaning', Philosophy 0/ Science 3 (1936), 420-471 and 4
(1937),2-40, §§ 7-10.
S This example was brought to my attention by Dr. V. Rabinowicz.
6 On the contrary, if such an upper bound exists we need not care about what could happen
beyond the least upper bound, i.e. what could be the result of an impossible test.
Received 15 June 1984
Johann Wolfgang Goethe-Universitiit

Fachbereich Philosophie
Dantestr. 4-6
6000 Frankfurt/Main
F.R.G.
ULRICH BLAU
DIE LOGIK DER VNBESTIMMTHEITEN VND

PARADOXIEN
( KurzJassung 1 )
INHALT
1. Oberblick: Wahrheitswerte und Refiexionsstufen 369

1.1. Der fundierte Bereich 371
1.2. Zur logischen Form und den Paradoxien des Konditionals 377
1.3. Der unfundierte Bereich 382
2. Syntax von LR 395
3. Semantik von LR 398
4. Definitionen und Regeln 408
4.1. Standardjunktoren 414
4.2. Standardquantoren 415
4.3. Wahrheitswertkonstanten 417
4.4. Priijunktion 418
4.5. Syntaktische Priidikate 419
4.6. Semantische Satzpriidikate und entsprechende lunktoren 421
4.7. Semantische Priidikat- und Termpriidikate 423
4.8. Zahlen, Anzahlquantoren, anzahlpriisupponierende Quantoren 425
4.9. Kennzeichnung 428
4.10. Existenz, Denotation und kontingente Objektbezeichnungen 433
5. Die Losung der Paradoxien 436
5.1. Die Wahrheitsparadoxie (Eubulides, Lukasiewicz) 436
5.2. Die Erflillungsparadoxie (Grelling) 438
5.3. Die Bezeichnungsparadoxie (Konig, Berry, Finsler) 439
5.4. Die Mengenparadoxien 440
6. Transfinite Refiexionsstufen 451
Anmerkungen 454
Literatur 458
1. UBERBLICK: WAHRHEITSWERTE UND REFLEXIONSSTUFEN
Die hier vorgeschlagene Reflexionslogik LR hat vier bis fiinf Ziele, sechs
Wahrheitswerte und unendlich viele Refiexionsstufen. Sie hat zum Ziel:
1. die Analyse von Siitzen, die aus folgenden Griinden unbestimmt sind:
VI Vagheit
Erkenntnis 22 (1985) 369-459. 0165-0106/85/0220-0369 $09.10
370 ULRICH BLAU
V2 semantische Kategoriefehler (Grenzgebiet zwischen VI und V3)

V3 semantische Prasuppositionsfehler (Hauptbeispiel: Nichtreferenz)
V4 semantische Zirkel oder Regresse;
2. die Analyse der Begriffe wahr,falsch, unbestimmt,fundiert, paradox,
der Bezeichnungs- und Erfullungsrelation und anderer dieser Art;
3. die Losung der semantischen Paradoxien;
4. die Abschaffung der Metasprachenhierarchie: die informelle, unge-
stufte, selbstreferentielle Metasprache (natiirliche Sprache) soll Stuck fiir
Stiick in eine formale Objektsprache semantisch korrekt und syntaktisch
moglichst oberflachennah iibersetzt werden;
5* die Logik der Mengen und die Losung ihrer Paradoxien.
Das fiinfte Ziel wird auf alle Zeiten der Intuition vorbehalten bleiben,
aber formale Teillosungen sind moglich. Die intuitive Losung hat schon
Cantor gefunden und eine ausgezeichnete formale Teillosung ist das Stan-
dardsystem ZF. Das wird deutlicher am Ende, wenn die vier ersten Ziele
erreicht sind.
Diese sind erreichbar mit einer Hierarchie von Reflexionsstufen und
sechs Wahrheitswerten aufjeder Stufe: Neben W (wahr), F (jalsch) entste-
hen die Werte N (neutral) aus Griinden VI, V2, V3, femer 0 (offen) aus
GrUnden V4, femer zwei halboffene Werte W (nicht-wahr, aber offen, ob
falsch oder neutral) und F (nicht-falsch, aber offen, ob wahr oder neutral)
durch Konjunktion und Adjunktion von N und O. W und F hei13en D
(definit, bestimmt), die vier anderen Werte heiBen V (unbestimmt). W, F
und N heiBen IF (jundiert), die drei anderen Werte hei13en IU (unfundiert).
Trotz dieser Fiille ist LR eine konservative Logik 2 in folgendem Sinn:
L2 sei die klassische Junktoren- und Quantorenlogik 1. Stufe mit Identitat
und Russells Kennzeichnungstheorie 3 ; L3 sei die in Abschnitt 1.1. skizzierte
dreiwertige Logik; L2", und L3", seien L2 und L3 mit Vnendlichkeitsvo-
raussetzung. Dann sind L3 und L3", beziiglich Giiltigkeit, Folgerung, Er-
flillbarkeit konservative Erweiterungen von L2' und L2",; und LR ist auf
allen Stufen in denselben Hinsichten eine konservative Erweiterung von
L3",. (Die Vnendlichkeit kommt hinzu, da LR die Existenz aller formalen
Ausdriicke voraussetzt.) Der wesentliche Punkt wird natiirlich die Analyse
der unfundierten Satze sein, aber vorher soll etwas zu den fundierten gesagt
werden4 •
UNBESTIMMTHEITEN UND PARADOXIEN 371
1.1. Der Jundierte Bereich
Dieser Bereich der natiirlichen Sprache ist der gemeinsame Anwendungs-

bereich fUr L2 und L3, und die wahren Satze im Sinn von L2 sind genau
die wahren im Sinn von L3. Aber die falschen Satze im weiten Sinn von
L2 werden in L3 zerlegt in die falschen im engen Sinn und die neutralen.
VI, V2, V3 sind Falschheitsgrunde fUr L2, Neutralitatsgrunde fUr L3.
(Das umgangssprachliche "falsch" schwankt zwischen diesen und anderen
Sinnen.) Entsprechend wird die klassische Negation "," ("nicht wahr")
in L3 zerlegt in "-" ("nicht") und "W". Entsprechendes gilt fUr die Pra-
dikat-Extensionen: Jedes gewohnliche Elementarpradikat hat in beiden
Logiken denselben Positivbereich, aber sein Komplement wird in L3 zer-
legt in einen Negativ- und einen Neutralbereich. Der letztere enthalt alles,
worauf das Pradikat infolge von Vagheit oder kategorialer Beschrankung
weder zutrifft noch nicht zutrifft ("nicht" im Sinn von" - ".) Die Grenzen
zwischen Positiv/Neutral/Negativbereich sind scharf Das ist vollig un-
plausibel, wenn man an die Vagheitsbereiche von Pradikaten im Sinn des
Gegenwartsdeutschen denkt. Aber in einem gegebenen Kontext verschar-
fen sich die Grenzen: Der Sprecher wendet, aus welchen Grunden, Ab-
sichten, Riicksichten auch immer, das Pradikat auf bestimmte Objekte an,
auf andere nicht und bei manchen trifft er keine Entscheidung. Die Neu-
tralbereiche sollen genau diejenigen Objekte enthalten, bei denen der Spre-
cher im gegebenen Kontext, ware er allwissend, kein Vrteil abgeben wiirde.
Bine hilfreiche Fiktion, vor allem spater fUr LR, ist der allwissende Veri-
jikator V, der sich strikt an die kontextuell verscharften semantischen Re-
geln des Sprechers halt und alle Tatsachen ermitteln kann: Wahrheit und
Falschheit ist Verijizierbarkeit und Falsijizierbarkeit Jur VS. Diese i.f. oft
verwendeten Begriffe sind nie im empirischen Sinn, fUr uns, sondem im
absoluten Sinn, fUr V, zu verstehen. Betrachten wir Austins Beispiel:
(1) Frankreich ist sechseckig.

Der Satz, sollte man meinen, kommt Quines Ideal des "ewigen Satzes"
erfreulich nahe. Er ist syntaktisch eindeutig, Objektbezeichnung und Pra-
dikat sind (nahezu) kontextunabhangig, Referenzprobleme gibt es nicht,
das Pradikat ist exakt. Aber der Wahrheitswert? Der Franzosischlehrer
halt ihn fUr wahr, der Geographielehrer fUr Jalsch, der Mathematiklehrer
fUr kategorial falsch, also neutral, der Deutschlehrer fUr vage, also auch
372 ULRICH BLAU
neutral, und jeder hat recht - in seiner Kontextsprache K. Darunter verste-

hen wir ein Tripel <N, S, 1) mit einer natiirlichen Sprache N, einer Menge
S von Sprechem von N (nicht-leer, aber oft eine Einermenge) und einem
(oft kurzen) Zeitraum T, in dem fiir alle s E S
(a) der intendierte Objektbereich gleich bleibt,
(b) alle semantischen Regeln fiir die Elementarpdidikate von N gleich
bleiben und daher
(c) alle in L3 formalisierbaren Satze von N einen eindeutigen Wert W,
F oder N haben.
Das dreiwertige formale Gegenstiick zu Kist die formale Sprache von L3
zusammen mit einer Interpretation q> iiber dem Objektbereich D von K,
die jedem r-stelligen Pradikatparameter P ein Tripel <P:, P;, P~>, beste-
hend aus Positiv-, Negativ- und Neutralhereich von P bei q>, zuordnet, wo-
bei
P: u P; u P~ = Dr,
P: n P; = P: n P~ = P; n P~ = 0
Alle Objektbezeichnungen a werden in L3 letztlich als Kennzeichnungen
lxAx definiert, das Denotat lal." existiert genau dann bei q>, wenn Ax die
Einzigkeitsbedingung bei q> erftillt, und jeder elementare Satz B erhalt bei
q> einen Wahrheitswert IBI." = W, F oder N nach den folgenden Regeln:
alle lajl." existieren und

W
<Iall.", ..., larl.,,> E P:,
alle lajl." existieren und
F <=>
<Iall.", ... , larl.,,> E P;,
sonst, d.h. nicht alle lajl." existieren
N
I
oder <Iall.", ... , larl.,,> E P~.
W jlal." uild Ihl." existieren und sind identisch,

F <=> lal." und Ihl." existieren und sind verschieden,
N lal." oder Ihl." existiert nicht.
Eine Bemerkung zu den metasprachlichen Junktoren und Quantoren: ",.•..,"

("nicht"), "&" ("und"), "v" ("oder"), "=>" ("wenn-dann"), "_" ("genau
dann, wenn"), "'f/" (Allquantor), "3" (Existenzquantor). Sie konnen im
Sinn von L2, L3 oder LR verstanden werden, es macht keinen Unterschied;
denn dort, wo sie verwendet werden, treten die Unbestimmtheitsgriinde
Ul - U4 nicht auf. Das metasprachliche Identitatssymbol wird gelegentlich
zur Unterscheidung yom objektsprachlichen mit einem Punkt versehen:
" == ". Die objektsprachlichen logischen Symbole werden metaspracb,lich
meist autonym, als Namen ihrer selbst, verwendet.
Die objektsprachlichen logischen Grundsymbole von L3 sind: "w"
(Wahrheitsjunktor), "-" (starke Negation), "A" (Konjunktion), "/\"
(AI/quantor), "=" (Identitiit) und zur wahrheitsfunktionalen Vervollstan-
digung ein O-stelliger Neutralitiitsjunktor "No" mit der Regel 6 :
INoltp = N.
Einige Definitionen:
-,A - -WA schwache Negation
FA :=W-A Falschheitsjunktor
NA - -WA A -FA N eutralitiitsjunktor
DA - -NA Definitheits-
oder Bestimmtheitsjunktor
A v B:= -(-A A -B) Adjunktion
VxA -/\x-A Existenzquantor
Die l-stelligen Junktoren haben die Regeln:
A -A-,A WA FA NA DA
WFF W F F W
FWWFWFW
NNW F F W F
Warum wir dieselben Symbole als objektsprachliche Junktoren und me-
tasprachliche Wahrheitswertbezeichnungen verwenden, wird spater ver-
standlicher. "Wahr", "falsch" usw. sind semantisch und syntaktisch ziem-
lich vieldeutig. Um die Symbolik einfach zu halten, werden wir nur die
wesentlichen, d.h. die semantischen Unterschiede symbolisch festhalten.
Das objekt- und metasprachliche "w" driickt auf syntaktisch mehrdeutige
374 ULRICH BLAU
Weise denselben - im spateren Sinn absoluten, naiven, unreflektierten -

Wahrheitsbegriff aus.
Zur Erlauterung der I-stelligen Junktoren wollen wir Satz (1) ais neutral
betrachten. Welchen Wert hat die Negation?
(2) Frankreich ist nicht sechseckig.
Wer (2) ais neutral betrachtet, versteht "nicht" im starken Sinn:
(2') - Pa, P: sechseckig, a: Frankreich
Aber (2) kann auch so gemeint sein:
(2a) Frankreich ist selbstverstandlich nicht sechseckig.
(Das Pradikat ist kategorial fehi am Platz.)
(b) Frankreich ist genau genommen nicht sechseckig.
(Das Pradikat (!) ist hier zu vage.)
Dann ist "nicht" die schwache Negation:
(2") -,Pa
Neutralitatsgriinde VI, V2, V3 fUr A und -A sind Wahrheitsgriinde ffir
-,A. Das Beispiel (1), (2) weist daraufhin, daB VI, V2, V3 flieBend inein-
ander iibergehen; es ware wenig sinnvoll, den Wert N in die Werte "vage",
"kategorial faisch", "prasuppositional faisch" oder gar noch feiner zu zer-
Iegen. Man gewinnt nichts 7 und verliert vieles. - Satze der Art
(3) DaB Frankreich sechseckig ist, ist wahr WPa
DaB Frankreich sechseckig ist, ist faisch FPa
sind nach den obigen Regeln faisch, wenn (1) neutral ist. Das diirfte der
Intuition entsprechen, aber natiirlich ist bei dem vieldeutigen "faisch" Vor-
sicht geboten. Wenn der Mathematiklehrer sich so ausdriickt:
(4) DaB Frankreich sechseckig ware, ist schlicht und einfach
falsch, da geographische Gebilde ...
so meint er kategorial falsch, d.h.
(4') NPa
Zum Verstandnis von N ist es wichtig, semantische und pragmatische
Prasuppositionen auseinanderzuhalten. Ihr Vnterschied hat nichts mit
Kontextabhangigkeit zu tun, denn kontextabhangig sind beide. Semantik

analysiert Wahrheitsbedingungen von Satzen in K; Pragmatik analysiert
Angemessenheitsbedingungen von Satzen in K. Semantische Pdisupposi-
tionsfehler fiihren zur NeutraliHit in K; pragmatische Prasuppositionsfeh-
ler fiihren zur Unangemessenheit, Unverstandlichkeit, Dberfiiissigkeit in
K. Beispiel:
(5) Die Miinze ist echt.
(5') PzxQx, P: echt, Q: Miinze
Der Satz prasupponiert semantisch, daB es genau eine Miinze gibt; und
pragmatisch, daB ein Zweifel an ihrer Echtheit besteht oder berechtigt ist.
Ob berechtigt oder nicht, der Wahrheitswert von (5) bleibt davon unbe-
riihrt.
Aber ist die semantische Prasupposition von (5) nicht in jedem Fall
verletzt? Es gibt mehr als eine Miinze im Universum. Das stimmt. Trotz-
dem ist (5) in gewohnlichen Kontexten nicht neutral, denn die Aussage
lautet expliziter:
(5a) Die einzige Miinze, auf die ich jetzt referiere, ist echt.
Indikatoren wie "ich" und "jetzt" werden in dieser Kurzfassung nicht be-
handelt, sondern in die Parameteriibersetzung abgeschoben: Q ist genauer
zu lesen als "Miinze, auf die ich jetzt referiere". So viel zum Neutralitats-
grund U3.
Konjunktion, Adjunktion und die Quantoren sind in L2, L3, LR durch
die Prinzipien charakterisiert:
{W } { beide Konjunktionsglieder sind W,

P /\ A /\ B ist F <:> mindestens ein Konjunktionsglied ist F,
ist {WF } {mindestens ein Adjunktionsglied ist W,
Pv A v B <:> beide Adjunktionsglieder sind F,
{W } { alle Spezialisierungen sind W,
P /\ /\xAx ist F <:> mindestens eine Spezialisierung ist F,
ist {WF } { mindestens eine Spezialisierung ist W,
PV VxAx <:> alle Spezialisierungen sind F,
wobei mit "Spezialisierungen" Einsetzungen von Objektnamen gemeint
sind, die den Objekten des Bereichs bijektiv zugeordnet sein sollen 8 •
376 ULRICH BLAU
P /\, P V besagen unter Voraussetzung der Trivalenz:

R/\ 3 W F N RV3 W F N
W W F N W W W W
F F F F F W F N
N N F N N W N N
Vnd analog erganzen sich P1\, PV im dreiwertigen Fall:
I\xAx ist N keine Spezialisierung ist F, aber mindestens
¢>
eine ist N.
VxAx ist N ¢> keine Spezialisierung ist W, aber mindestens
eine ist N.
Diese Junktoren und Quantoren finden sich in vielen dreiwertigen Syste-
men. Interessant ist das Konditional "-+", denn an ihm wird deutlich, war-
urn die bekanntesten dreiwertigen Systeme (Lukasiewicz, Bochvar, Rei-
chenbach, Kleene) und viele andere gescheitert sind. Ihre Autoren haben
die Natur der Wahrheitswerte griindlich miBverstanden und einen 3. Wert
fUr "unbekannt", "unerkennbar", "unentscheidbar", "irrelevant", ~'sinn
los", "paradox" oder ahnliches einzufiihren versucht. Offenbar haben sie
Semantik mit Epistemologie oder Pragmatik verwechselt, also den allwis-
senden, interesselosen Verifikator mit uns. Aber fUr epistemisch oder prag-
matisch gefarbte Wahrheitswerte verhalten sich "und" und "oder" nicht
wahrhei tsfunktional:
(6) Casar hatte Blutgruppe B.
(7) Casar hatte nicht Blutgruppe B.
Wenn wir beiden Satzen den 3. Wert "unbekannt" geben, so hat ihre Kon-
junktion und Adjunktion nach R /\ 3, R V 3 ebenfalls diesen Wert. Tat-
sachlich ist uns '(6) und (7)' eindeutig alsfalsch bekannt und '(6) oder (7)'
eindeutig als wahr bekannt. 9 Ahnliches ergibt sich fUr pragmatisch geflirb-
te Wahrheitswerte: Tautologien und Kontradiktionen sind meistens unin-
teressant, irrelevant, iiberfiiissig, aber trotzdem wahr und falsch.
Die genannten Autoren und viele andere nahmen das Problem der
Wahrheitsfunktionalitat nicht so ernst und begingen ihren zweiten Fehler.
Sie nahmen die Nicht-Wahrheitsfunktionalitat des "wenn-dann" auch
nicht so ernst und wahlten irgendeinen halbwegs passenden dreiwertigen
Junktor als Konditional. Das ergab Systeme ohne erkennbare Anwen-

dung, die im allgemeinen die klassische Logik nicht konservieren. Ich gehe
auf sie nicht ein. In L2, L3, LR ist das Konditional so definiert:
A -+ B: = -, A v B. Das bedeutet im dreiwertigen Fall:
R-+3 W F N
W W F N
F W W W
N W W W
Dieser Junktor dient nicht zur semantisch korrekten Formalisierung von
"wenn-dann", sondern anderen eng verwandten Zwecken. Um dies zu zei-
gen, muB ich etwas zur logischen Form sagen.
1.2. Zur logischen Form und den Paradoxien des Konditionals
Ziel der logischen Formalisierung ist die Explikation der in/ormel/en 10-
gischen Folgerung und Giiltigkeit, nennen wir sie "IrK", DaB sie von K
= <N, S, 1) abhangt, nicht nur von N, zeigt etwa das Beispiel:
(1) Jede Aussage ist wahr oder faisch.
In vielen Kontexten ist "Aussage", "wahr" und "faisch" so definiert oder
prasupponiert, daB (1) stimmt; dann ist (1) nicht nur wahr, sondern in-
formell giiltig. 1m Kontext von L3 ist (1) falsch.
'Al' ... , An IrK B' und 'IrK B' sollen besagen, daB der allwissende V,
der die Aj und B im Sinn von K versteht, ohne Verwendung seines Tat-
sachenwissens, allein aufgrund seiner logischen Fahigkeiten, aus den An-
nahmen A j , bzw. ohne Annahmen, auf B schlieBen kann. Das ist natiirlich
keine Definition. Der Grundbegriff "IrK" ist nicht definierbar, sondern
nur schrittweise partiell explizierbar. Ware er vollig dunkel, so ware jede
formale Logik willkiirlich, ware er vollig klar, so ware sie iiberfiiissig. Eini-
ge weitere Begriffe:
Eine Formalisierung (in L2 bzw. L3) von A in Kist ein Satz A (in L2
bzw. L3) zusammen mit einer Dbersetzung TC seiner Parameter in Aus-
driicke syntaktisch gleichen Typs von N.
Eine entsprechende Interpretation <p fUr diese Formalisierung ist eine
solche iiber dem Objektbereich D von K, die den Parametern von A die
378 ULRICH BLAU
Extensionen der durch 1t zugeordneten Ausdriicke im Sinn von K gibt.

Eine Formalisierung A, 1t von A in K heiBt
stark genug, wenn fiir jedes entsprechende qJ gilt: IAltp = W II- KA;
schwach genug, wenn fiir jedes entsprechende qJ gilt: A II-K IAltp = W;
gleichstark, wenn sie stark und schwach genug ist, d.h. wenn fiir jedes
entsprechende qJ gilt: II-K (IAltp = W <=> A);
seman tisch korrekt, wenn fiir jedes entsprechende qJ gilt: II- K (IAltp
= Wahrheitswert von A).
Fiir L2-Formalisierungen ist "gleichstark" dasselbe wie "semantisch kor-

rekt"; fiir L3-Formalisierungen ist "gleichstark" schwacher10 • Verstiirkte
(bzw. abgeschwiichte) Formalisierungen sind solche, die stark genug, aber
nicht schwach genug (bzw. umgekehrt) sind. Nun zum Zusammenhang
von "~" und "wenn-dann". Diese Satzverkniipfung ist in den meisten
Kontexten nicht-wahrheitsfunktional und durch keinen 2-, 3- oder mehr-
wertigen Junktor semantisch korrekt, oder auch nur gleichstark, formali-
sierbar. Aber zumindest gilt fUr aIle iiblichen K und L = L2, L3:
(2) 'Wenn A, dann B' in K wird durch p ~ q, p: A, q: Bin L
schwach genug formalisiert.
Beweis: 'Wenn A, dann B' sei W in K. Dann ist ausgeschlossen, daB A in
K W, aber B in K nicht Wist (intuitive Semantik des "wenn-dann"). Also
ist in K entweder A nicht W oder B W. Dann ist bei jedem entsprechenden
qJ Ipltp =F W oder Iqltp = W und nach Def. von "~" in L2 und L3 ist
Ip ~ qltp = W, was zu zeigen war. Man beachte, daB jeder Schritt informell
zwingend, im Sinn von "II-K" ist.
Die semantischen Zusammenhange "und": " 1\ ", "oder": "v" sind be-
kanntlich enger. Fiir das gewohnliche (nicht-temporale, nicht-kausale
...) "und" und das nicht-ausschlieBende "oder" gilt in L = L2, L3:
(3) 'A und B', 'A oder B' in K wird durch p 1\ q, P V q, p: A, q: B
in L gleichstark formalisiert l l .
SchlieBlich noch etwas zur schwachen Negation:
(4) A in Kwird durch A, 1t in L schwach genug formalisiert <=> 'Es
ist nicht wahr, daB A' wird durch lA, 1t in L stark genug
formalisiert.
Das HiBt sich ahnlich wie (2) zeigen. Das niichste informelle Theorem
beantwortet eine ziemlich wichtige Frage, die selten gestellt und gewohn-
lich nicht beantwortet wird: Wie ist es moglich, einen informellen SchluB
formal zu beweisen?
T1 Anwendungstheorem fiir L = L2 und L3

Wenn (a) AI, ... , An II-L B und fUr j = 1, ... , n
(b) Aj in K durch Aj, n in L schwach genug formalisiert wird,
(c) B in K durch B, n in L stark genug formalisiert wird,
dann folgt AI, ... , An II-K B.
Beweis: Angenommen AI, ... , An. cp sei eine entsprechende Interpreta-
tion. Nach (b) folgt informell zwingend in K: IAjl'l' = W, fiir j = 1, ... , n.
Nach (a) folgt informell zwingend in K: IBI'I' = W. Nach (c) folgt informell
zwingend in K: B, was zu zeigen war.
Natiirlich ist dieser Beweis eben so informell wie die Grundrelation
"II-K". Wenn man ihn genauer analysiert, so verwendet er einige unpro-
blematische Eigenschaften dieser Relation (Transitivitiit, Modus Ponens,
Definitionsbeseitigung, Allspezialisierung) und femer:
Das ist ebenfalls unproblematisch, denn fiir L = L2 und L3 gibt es ada-

quate Kalkiile mit informell zwingendem Korrektheits- und Vollstiindig-
keitsbeweis; daher liiBt sich jede L-Folgerung in endlich vielen informell
zwingenden Schritten erkennen.
T1 zeigt, daB L2, L3 nur informell giiltige Schliisse liefem - sofern die
Formalisierungsvoraussetzungen (b), (c) erfUllt sind. Aber das ist nur in-
formell, kontextuell zu priifen. Daher bleibt die formale Logik auf die
informelle angewiesen; sie kann sie erhellen, aber nicht ersetzen. Zuriick
zum Thema.
(5) Wenn A, dann B.
Der Satz besagt in K gewohnlich:
(5a) Aus gewissen in K relevanten, nicht notwendig bekannten,
Tatsachen und GesetzmaBigkeiten und der Annahme A folgt
informell zwingend B.
Das ist vage und oft unentscheidbar; der Wahrheitswert von (5) in Kist
380 ULRICH BLAU
schwer zu fassen. Eher entscheidbar ist die Frage der informellen Giiltigkeit
von (5) in K:
(6) II-K Wenn A, dann B.
Das erscheint gleichbedeutend mit
(6a) A II-K B
d.h. der Bezug auf auBersprachliche Tatsachen und GesetzmaBigkeiten
entfallt. Bei dieser Auffassung von (6) folgt aus T1 unmittelbar
T2 Anwendungstheorem fur das Konditional in L = L2 und L3

Wenn (a) II-L A -+ B und
(b) A in K durch A, 1t in L schwach genug formalisiert wird,
(c) Bin K durch B, 1t in L stark genug formalisiert wird,
dann folgt II-K Wenn A, dann B.
(Denn analog zu L2 gilt in L3: (A II-L3 B) <=> (11- L3 A -+ B).) T2 erklart
die sog. Paradoxien des Konditionals. Vergleichen wir (7) und (8).
(7) P -+ P
p-+pvq
P/\q-+p
(p-+q) /\ p-+q
(p -+ q) -+ -,p v q
P /\ -, q -+ -, (p -+ q)
Diese L2- und L3-giiltigen Satze sind bei der Standardlesart "nicht",
"und", "oder", "wenn-dann" der Junktoren informell giiltig; die folgen-
den sind es nicht:
(
(8) -, (p -+ q) -+ p
-, (p -+ q) -+ -, q
-,p -+ (p -+ q)
q-+(p-+q)
-, (p -+ q) -+ (-, p -+ q)
-, (p -+ q) -+ (p -+ -, q)
Mit (2)-(4) erkennt man, daB die Standardlesarten von (7) die Formalisie-
rungsvoraussetzungen (b) und (c) erfilllen: sie miissen informell giiltig sein.
Die Standardlesarten von (8) verletzen diese'Voraussetzungen; man
braucht sich nicht zu wundem, daB sie informell nicht gelten. Das ist die
Losung der sog. Paradoxien des Konditionals, der bescheidensten aller
logischen Paradoxien.
(2) zeigt einen schwachen Zusammenhang zwischen "wenn-dann" und
"-+", der das L3-Konditional nicht eindeutig auszeichnet. (Es gibt 9 drei-
wertige Junktoren, die (2) erfiillen, und wenn man verlangt, daB die L2-
Giiltigkeit in L3 konserviert werden solI, so bleiben 4 Kandidaten fiir
"-+" iibrig). Die Auszeichnung des L3-Konditionals ergibt sich aus seiner
wichtigsten Anwendung: Beschriinkte Allsiitze
(9) Alle P's sind Q's
werden, vor allem in der Mathematik, ohne Prasuppositionen fiir die An-
zahl der P's, auch so ausgedriickt 12 :
(9a) Wenn x ein P ist, dann ist x ein Q.
Dieses "wenn-dann" ist strikt wahrheitsfunktional; (9) wird in L2 und L3
durch
(9') I\x (Px -+ Qx), P: P, Q: Q
semantisch korrekt formalisiert; kein anderer dreiwertiger Junktor ware
geeignet13. So viel in Kiirze zu L3. Der Vorteil gegeniiber den alteren
dreiwertigen Systemen liegt in der besseren semantischen Motivation der
Wahrheitswerte, der beiden Negationen, des Konditionals und in der Tat-
sache, daB fiir aIle Satze At. ... , An, B, die syntaktisch zu L2 gehoren, d.h.
als logische Konstanten nur -', /\, v, -+, - , 1\, V, =, I, das Existenz-
pradikat E! und Anzahlquantoren ~n!, ::;;n!, >n!, <n!, n! enthalten, das
Konservierungstheorem At. ... , An II-L2 B ~ At. ... , An II-L3 B gilt. Weitere
dreiwertige Junktoren "-", "==" (Bikonditional und Aquivalenz, die in
L2 zusammenfallen) und "j" (prasupponierende Konjunktion, die in L2
mit " /\ " zusammenfallt), betrachten wir spater im Rahmen der starkeren
Logik LR.
Diese entsteht aus L3 im wesentlichen durch Hinzunahme der Anfiih-
rungsfunktion '" ''', mit deren Hilfe alle Ausdriicke von LR in LR be-
zeichnet werden konnen, und durch Verallgemeinerung von "w" und "F".
Sie werden als Priidikate eingefUhrt und die entsprechenden Junktoren
werden definiert:
WA:= WrA" FA:= FrA',
382 ULRICH BLAU
denn bei "wahr" und "falsch" ist der eingebettete daj3-Satz in die direkte
Rede transformierbar salva veritate. Die Satzanfiihrungen rN in WrN,
prA' sind transparent: Man kann von auBen hineinquantifizieren, auf A
die iiblichen Substitutionsprinzipien anwenden und man kann die Anfiih-
rungszeichen einfach weglassen. Das ist eine Besonderheit von W, Fund
anderen mit ihnen definierbaren Pradikaten; im allgemeinen sind AnfUh-
rungen opak. Beispiel:
(10) "Px -+ Qx" ist eine mit "P" beginnende offene Formel
(10') RTx -+ Qx' T', Rxy: x isteinemit y beginnende offene Formel
(10) und (10') sind geschlossene Satze ohne freie Variablen; man kann die
Anfiihrungszeichen weder weglassen noch von auBen hineinquantifizie-
ren l 4, noch kann man "Px -+ Qx" durch die logisch aquivalente Formel
" I Px v Qx" ersetzen. Nicht einmal die Definitionsbeseitigung ist in
opaken Anfiihrungen zulassig.
Mit diesen Erweiterungen der formalen Ausdrucksmittel haben wir den
fundierten Bereich der Sprache iiberschritten und stehen vor dem Liigner.
1.3. Der unfundierte Bereich
(1) Satz (1) ist nicht wahr.

Was solI man damit anfangen? Die nachstliegenden Vorschlage, in natiir-
licher Reihenfolge, sind diese:
1. Man verbietet unfundierte Satze wie (I) syntaktisch, da sie die Objekt/
Metasprachenhierarchie verletzen lS • Dieses Verbot hat einschneidende
Folgen ftir die natiirliche Sprache; ihre Syntax wird unentscheidbar und
unerwiinscht eng. Beispiel:
(Ia) Der letzte Satz, den Philitas von Kos geauBert hat, ist nicht
wahr.
Vielleicht war seine letzte AuBerung der Liigner, an dem er angeblich in
Verzweiflung gestorben ist. Dann ware (Ia) unfundiert, also syntaktisch
unzulassig; wir werden es nie erfahren. Wer unfundierte Satze verbieten
will, sollte auch bedenken, daB sie nicht nur mit "wahr" und "falsch",
sondern mit samtlichen intentionalen Pradikaten zu bilden sind:
(1 b) a hat nur den Geburtstagswunsch, daB seine Frau ihm keinen

Geburtstagswunsch erfiillt.
(c) a behauptet nur, daB seine Frau ihm keine Behauptung glaubt.
(d) a hat nur die Absicht, daB seine Frau keine seiner Absichten
durchschaut.
(e) a erkennt nur, daB seine Frau nichts von sein~n Erkenntnissen
ahnt.
Aus (1 b) folgt eine ahnliche Paradoxie wie (1): a's Frau erfiillt seinen ein-
zigen Wunsch genau dann, wenn sie ihn nicht erfiillt. Aus (c)-{e) folgen
leere Regresse: Er behauptet, daB sie nicht glaubt, daB sie nicht glaubt,
daB sie nicht glaubt ... Und er hat die Absicht, daB sie nicht durchschaut,
daB er die Absicht hat, daB sie nicht durchschaut ... Wer (1) mit Hilfe der
Sprachstufenhierarchie verbietet, kommt kaum umhin, auch (1b-e) syn-
taktisch zu verbieten und analog zu den Satzen samtliche intentionalen
Objekte: Wiinsche, Behauptungen, Absichten, Erkenntnisse, Vorstellun-
gen, ... mit Stufen zu versehen. Eine soIche Syntax ware auBerst kiinstlich,
tatsachenabhangig, unentscheidbar und nicht einmal theoretisch wiin-
schenswert. Denn wenn sie jede semantische Selbstreferenz verbietet, so
verbietet sie den sprachlichen Ausdruck wesentlicher Teile unseres Den-
kens. Beispiel: die natiirlichen Zahlen 0, 1, 2, ... Wir verstehen sie, weil
wir das Symbol" ... " verstehen. Dieses vielleicht wichtigste logisch-mathe-
matische Symbol driickt selbstreferentielle Regeln aus; in diesem Fall:
(R) Addiere lund dann befolge den ersten Teil von (R).
Jedes nichttriviale Rechnerprogramm hat Schleifen und jede Unendlich-
keitsvorstellung beruht auf einer selbstreferentiellen Regel: Man betrachtet
ihre samtlichen Resultate, gleichsam von auBen, als ein Objekt. Dieser
Schritt von innen nach auBen, von der Befolgung der Regel zur Betrach-
tung, d.h. Verdinglichung, ihrer samtlichen Resultate als Menge, ist kiihn,
man kann ihn tun oder lassen, aber nicht auf kleinere Teilschritte redu-
zieren. Der wesentliche Teil der Mathematik beruht auf solchen innen-
auBen-Schritten, ebenso auch, wie ich zeigen mochte, die semantischen
und Mengenparadoxien; ebenso auch ihre Losung. Semantische Selbstre-
ferenz ist die gemeinsame Wurzel von sinnvollen Rekursionen, unendli-
chen Mengen, leeren Regressen, Paradoxien und ihren Losungen. Man
sollte sie nicht syntaktisch beschranken, sondem semantisch analysieren.
Aber wie?
384 ULRICH BLAU
2. Man gibt Satzen wie (1) den dritten Wahrheitswert (wie immer man ihn
nennt.) Dann ist Satz (1) nicht wahr, Aber das sagt er selbst, also ist er
wahr, also nicht wahr, also ...
3. Man nimmt weitere Wahrheitswerte hinzu (wieviele ist gleich.) Aber

wenn Satz (1) einen dieser neuen Werte hat, so ist er jedenfalls nicht wahr.
Und das sagt er selbst, also ist er wahr, also nicht wahr, also ...
4. Man laBt Satze ohne Wahrheitswert zu. Aber wenn Satz (1) keinen
Wahrheitswert hat, dann ist er nicht wahr, und das sagt er ja, also ist er
wahr, also ...
5. Man laBt neben eindeutig wahren und eindeutig falschen Satzen para~
doxe zu, die wahr und falsch zugleich sind. Aber wie in der Geschichte
yom Hasen und Igel, der Liigner ist schon wieder da:
(2) Satz (2) ist nicht eindeutig wahr.
Ebenso wie (1) kann (2) nicht eindeutig wahr und nicht eindeutig falsch
sein. Aber dann ist seine Behauptung eindeutig wahr, also ...
6. Man laBt beliebig viele (unendlich viele, iiberabzahlbar viele) Wahr-

heitswerte, darunter auch "wahr", zu und ordnet jedem Satz irgendeine
Teilmenge, einschlieBlich der totalen und leeren, zu - es niitzt alles nichts,
entweder hat (2) eindeutig die Wertmenge {wahr} oder aber nicht; in bei-
den Fallen folgt der Widerspruch. Man konnte nun allmahlich den Ein-
druck gewinnen, daB die Moglichkeiten, den Liigner strikt wahrheitsfunk-
tional zu kurieren, erschOpft sind. Daher empfiehlt sich vielleicht dieser
Vorschlag:
7. Man fUhrt eine intensionaie, evtl. partiell-wahrheitsfunktionale Nega-

tion ein und laBt zu, daB Satze wie (1) zugleich mit ihrer Negation wahr,
oder vielleicht beide falsch, sind. Das geschieht z.B. in Systemen der "pa-
rakonsistenten" und der "dialektischen" Logik, vgl. [10] und [31]. Leider
hat keins dieser Systeme eine Semantik, die diesen Namen verdient. Man
wiirde gem wissen, welche Satze bei Negation den Wahrheitswert behalten,
welche nicht, upd vor allem warum. Als Antwort erhalt man - nun ja,
einen Index fUr "mogliche Welten" oder "Situationen" mit ein paar for-
malen Ordnungseigenschaften. Dergleichen Semantik zu nennen, ist killin,

aber leider heute iiblich. Ich mochte etwas anderes vorschlagen, aber zuvor
eine Beobachtung mitteilen. Die meisten reagieren auf (1) diffus; der Wahr-
heitsbegriff scheint im Unfundierten seine Konturen zu verlieren. Wir hat-
ten vereinbart:
wahr = verifizierbar (im absoluten Sinn, fUr V)
Nun geschieht etwas Merkwiirdiges. Vergleichen wir (1) mit
(3) Satz (3) ist nicht verifizierbar.
Dieser Satz erscheint den meisten viel eher als wahr! Aber dann ware er
verifizierbar, also falsch, also nicht verifizierbar, also wahr, ... Trotzdem
bleibt der Eindruck, daB sich zwischen (1) und (3) etwas geandert hat:
unsere Reflexionsstufe. Verifizierbarkeit ist reflektierte Wahrheit, Wahrheit
von auBen. Mir scheint, daB der Betrachter unfundierter Satze bestimmte
Reflexionsstufen in einer festen Reihenfolge durchlauft, aber gewohnlich
nicht sieht, denn er blickt auf den Satzinhalt, nicht auf den eigenen Re-
flexionsprozeB. Sobald er ihn erkennt, verschwinden die Paradoxien. Bei-
spiel:
(4) Satz (4) ist unbestimmt.
Rein inhaltlich konnte man argumentieren:
(a) Satz (4) ist wahr, und daher unbestimmt - ebensogut auch:
(b) Satz (4) ist unbestimmt, und daher wahr.
Den natiirlichen ReflexionsprozeB gibt nur (b) wieder. Um den Wahrheits-
wert von Satz (4) zu erkennen, miissen wir ermitteln, ob er unbestimmt ist;
also ob es unbestimmt ist, ob er unbestimmt ist; also ob es ... Solange wir
blind, auf der O. Reflexionsstufe, dem Zirkel folgen, erkennen wir gar
nichts. Aber sobald wir erkennen, daB wir nichts erkennen werden, haben
wir den Zirkel schon durchbrochen: Wir erkennen auf der 1. Reflexions-
stufe, daB (4) auf der O. Stufe offen und daher unbestimmt ist. Aber das
sagt er selbst: Nun erkennen wir, daB (4) auf der 1. Stufe wahr ist. Aber
das sagt er nicht: Nun erkennen wir, daB (4) auf der 2. Stufe falsch ist.
Aber das sagt er nicht ... Insgesamt erkennen wir (auf Reflexionsstufe
co, die wir erst in Abschnitt 6 betrachten werden):
(4) ist 0°, Wl, F2, F3, F4, ...
386 ULRICH BLAU
und analog sehen wir, daB der Liigner (1) durch die endlichen Stufen os-
zilliert:
(1) ist 0°, Wl, F2, W3, F4, ...
Andere Mitglieder der selbstreferentiellen Familie:
"Dieser Satz ist wahr" ist 00, Fl, F2, F3, F4, .. .
"Dieser Satz ist falsch" ist 0°, Ft, W2, F3, W4, .. .
"Dieser Satz ist nicht falsch" ist 00, Wl, W2, W3, W4, .. .
"Dieser Satz ist nicht unbestimmt" ist 0°, pi, W2, W3, W4, .. .
Nichtunterscheidung der Stufen erzeugt Paradoxien, d.h. Siitze, die wahr
und falsch zugleich sind. Sie entstehen zwangsliiufig, da der Refiexionspro-
zeB auf keiner Stufe zur Ruhe kommt, und unmerklich, da die Stufen
durch keine sprachlichen Merkmale angezeigt werden; der Mitteilungswert
ist zu gering. Das ist die informelle ErkUirung der semantischen Parado-
xien.
Die formale Erklarung verwendet ein gestuftes, sich selbst refiektieren-
des Verifikationsverfahren, das in Abschnitt 3 definiert wird. Dieses Ver-
fahren zergliedert den gegebenen Satz baumformig nach seiner logischen
Struktur und endet, sobald aIle Aste den Boden der Tatsachen erreichen.
vn sei der Verifikator auf Stufe n des Verfahrens. Er kann die fundierten
Satze A der Stufe n verifizieren, falsifizieren oder aber neutralisieren; an-
ders ausgedriickt: er kann entweder W A oder FA oder aber - W A und
- FA verifizieren. Aber die unfundierten Satze B der Stufe n bleiben offen
oder halboffen fiir ihn: 1m offenen Fall kann er weder - WB noch - FB
verifizieren, in den halboffenen Fallen kann er nur - WB oder nur - FB
verifizieren. Jeder Versuch von vn, die semantische Liicke zu schlieBen,
miindet mit mindestens einem Ast entweder
- in einen Zirkel, oder
- in einen nicht-zirkularen RegreB, oder
- in eine, noch unentschiedene, Verifizierbarkeitsbehauptung hoherer
Stufe.
Nun erfolgt der Refiexionsschritt: Vn + l erkennt, daB B auf Stufe n un-
fundiert ist, d.h. einen Wert 0, Vi oder F hat. Zum genaueren Verstiindnis
des Prozesses miissen drei Wahrheits- und Falschheitsbegriffe unterschie-
den werden, die in LR wie folgt symbolisiert sind:
(a) W, F: unreflektierte (absolute, naive) Wahrheit und Falschheit;
(b) W*, F*: reflektierbare Wahrheit und Falschheit (Verijizierbarkeit und

Falsijizierbarkeit);
(c) wn, Fn: Wahrheit und Falschheit der Reflexionsstufe n (n-Verijizier-
barkeit und n-Falsijizierbarkeit).
Unsere natiirlichen Begriffe "wahr" und "falsch" sind ambig zwischen (a)
und (b); das ist gewohnlich nicht sichtbar, es zeigt sich nur an unfundierten
Satzen und nur aufhOheren Reflexionsstufen. Die Begriffe (c) kommen in
der natiirlichen Sprache nicht vor; es sind theoretische Begriffe zur Analyse
von (a) und (b) wie folgt:
yn sieht aIle vorangehenden Stufen m < n objektiv richtig, von auBen:
Er erkennt die Wahrheit oder Falschheit aller Satze wmA, FmA. Seine
eigene Stufe sieht er subjektiv richtig, von innen: WnA und FnA versteht
er absolut als W A und FA. Das ist, von auBen gesehen, nicht falsch, aber
unvollstandig: Er sieht nicht die semantischen Liicken seiner Stufe, son-
dern erst spater, von auBen, auf allen Stufen r > n. Diese hoheren Stufen
sieht vn noch gar nicht: w r A und prA sind offen fiir ihn.
Die reflektierbaren W*, F* andern ihre Bedeutung: Der nichtreflektie-
rende VO versteht sie als WO, FO, d.h. als absolute W, F. Der reflektierende
yn + 1 versteht sie als wn, F n, d.h. er reflektiert stets die unmittelbar voran-
gehende Stufe. (Auf den spater betrachteten Limesstufen Ot: reflektiert VIZ
unendlich viele unmittelbar vorangehende Stufen.)
Betrachten wir die obigen Beispiele. Der Liigner (1) hat zwei Lesarten
in LR:
(1') - Wa, a: ' - Wa'; unreflektierter Liigner
(1") - W*b, b: ' - W*b'; reflektierter Liigner
Bei entsprechender Interpretation <p, d.h. wenn laltp = ' - Wa' und Ibltp
= '- W*b', wird sich nach der spateren Semantik ergeben:
(1') ist 0°, 0 1 , 0 2, 0 3, 04, .. .
(1") ist 0°, WI, F2, W3, F4, .. .
Der unreflektierte Liigner ist nicht paradox. Gleichgiiltig wie hoch yn steht,
Wist immer das Wahrheitspradikat wn seiner Stufe, es wird nicht reflek-
tiert, (1') ist on, wie vr (r > n) riickblickend erkennt. - Der reflektierte
Liigner ist paradox; jeder Reflexionsschritt andert das letzte Resultat. Mir
scheint, daB die Ambiguitat (1'), (1") unse"re diffuse Intuition zu (1) erklart:
Es gibt einen unreflektierten, redundanten, rein affirmativen Gebrauch von
388 ULRICH BLAU
"wahr", bei dem sich die Stufe nicht erhoht (wahr von innen), und einen
refiektierten, nicht-redundanten Gebrauch, bei dem sich die Stufe erhoht
(wahr von auBen). 1st nun (I') oder (I") die natiirlichere Lesart von (1)?
Das ist schwer zu sagen, aber so viel ist sicher: Wer erkennt, daB er (1)
unrefiektiert versteht, der refiektiert schon - und gerat ziemlich sicher in
die andere Lesart: "Fiir mich ist (1) absolut nichtssagend, sinnlos, leer -
und daher natiirlich nicht wahr - und daher wahr ... " Betrachten wir zum
Vergleich die Variante des Liigners:
(3) Satz (3) ist nicht verifizierbar.
Wenn man "verifizierbar" ais immer wieder refiektierbar versteht, so hat
(3) die logische Form (I") und der Wert oszilliert. Aber diese Auffassung
ist vielleicht schon allzu refiektiert. Zunachst versteht man "nicht verifi-
zierbar" in (3) als "nicht unmittelbar verifizierbar". So verstanden hat (3)
die logische Form
(3') - WOa, a: '- WOa'
und die Werte 0°, WI, W2, W3, .. , Dieselben Werte hat auch
(2) Satz (2) ist nicht eindeutig wahr - im Sinn von:

(2a) Satz (2a) ist nicht auf allen endlichen Stufen wahr.
(2a') - /\x (Nx -+ WXa), a: '- /\x (Nx -+ WXa)' ,
mit dem spater definierten natiirlichen Zahlpradikat "N". Aber wie for-
malisiert man den schrankenlosen Liigner?
(5) Satz (5) ist auf keiner Stufe wahr.
Er bleibt auf allen Stufen offen und ist daher auf keiner Stufe wahr. Also
ist er wahr - aber auf welcher Stufe? Satze wie (5) treiben die Refiexion
iiber aIle endlichen Stufen, aIle transfiniten Stufen und - vieIleicht 1C5 - aIle
formalen Mengentheorien hinaus, immer weiter in Cantors absolutes Men-
genuniversum, das nie verlassen wird, nie Objekt, weder Menge noch
Klasse wird, iiber das nichts gesagt werden kann. Etwas mehr dariiber
wird in Abschnitt 5.4. gesagt, und anschlieBend, in Abschnitt 6, etwas
mehr zum schrankenlosen Liigner. Betrachten wir nun zwei nahe Ver-
wandte des Liigners.
(a) Die Erfiillungsparadoxie: Pradikate, die sich nicht selbst erfiillen, wie
"zweisilbig", "unverstandlich", "exakt", "Satz", nennt Grelling heterolo~

gisch und stellt die Frage, ob "heterologisch" heterologisch ist.
(b) Die Bezeichnungsparadoxie:
die kleinste natiirliche Zahl,

die durch keine Bezeichung
auf dieser Tafel bezeichnet
wird
k sei die gesamte Kennzeichnung auf der Tafel. Was, wenn iiberhaupt,
bezeichnet k? Wir stehen vor einem Trilemma: Wenn k nichts bezeichnet,
so miiBte k 0 bezeichnen. Wenn k 0 bezeichnet, so miiBte k 1 bezeichnen.
Wenn k eine Zahl > 0 bezeichnet, so miiBte k 0 bezeichnen. Das Trilemma
wird sich, ahnlich wie (a), durch Reftexionsstufenunterscheidung auftosen.
Zur Analyse von (a) benotigen wir die objektsprachliche Relation:
[die Objekte] Xlo ••• , Xr erfiillen [das Pradikat] y.
Es bietet sich an, stattdessen die Konverse einzufiihren:
y ist wahr flir (trifft zu auf, gilt flir) Xlo ••• , X r •
Die Konverse ist ein verallgemeinertes Wahrheitspriidikat: I-stellig an-

wendbar auf Satzbezeichnungen; r+ I-stellig anwendbar auf Bezeichnun-
gen r-stelliger Pradikate, gefolgt von r Objektbezeichnungen. Entspre-
chend werden wir formal Wi und Fi als Pradikate variabler Stellenzahl ~
1 einfiihren (i = leer, * oder n).
Ferner miissen wir zur Analyse von (a) in der Objektsprache iiber ob-
jektsprachliche Priidikate reden. Das sind Formeln mit freien Variablen,
deren Reihenfolge wir durch eine Argumentstellenmarkierung (Xl ••• Xr )
festlegen. Z.B. ist
(yx) VzPzxyx ein 2-stelliges Pradikat, falls y und X verschiedene Varia-
bIen sind.
Fiir (b) benotigen wir die objektsprachliche Bezeichnungs- oder Deno~
tationsrelation Ll i, mit denselben Indizes wie Wi und F i, denn das Denotat
unfundierter Bezeichnungen wie k hangt ebenfalls von der Reftexionsstufe
abo Ein etwas einfacheres Beispiel: k' sei die Kennzeichnung "der wahre
Satz auf t", wobei auf Tafel t nur der Liigner (1) steht. Bei reftektierter
Auffassung von (I) oszilliert das Denotat von k' zwischen Nichtexistenz
und (I).
390 ULRICH BLAU
Einen merkwiirdigen Status haben die TermanfUhrungen ra, in z:1 ira'b;

sie sind opak aber extensional. Man kann die AnfUhrungszeichen weder
weglassen noch bineinquantifizieren. Beispiel:
(6) "lXPxy" bezeichnet nichts
(6') -,VuJ rlxPxy' Z
(6') ist ein geschlossener Satz ohne freie Variablen und LR-giiltig (da im
Sinn der spateren Semantik offene Terme wie lXPxy nie etwas bezeichnen.)
Trotz der Nicht-Transparenz von ra, in Ai ra, b sind auf a die iiblichen
Substitutionsprinzipien anwendbar:
(7) "Franz" bezeichnet Franz.
Franz ist der anwesende Familienvater.
"Der anwesende Familienvater" bezeichnet Franz.
(7') A r a, a a: Franz
a = lX (Px 1\ Qx) P: anwesend
A [lX (Px 1\ Qx)] a Q: Familienvater
(8) "Der anwesende Familienvater" bezeichnet Franz.
Von den Anwesenden sind genau die Familienvaler kahl.
"Der anwesende Kahle" bezeichnet Franz.
(8') A rlX (px 1\ Qx)' a

Ax (px ~ (Qx +-+ Rx» R: kahl,
A r lX (Px 1\ Rx)' a
Schliisse dieser Art werden in LR giiltig sein. Mit wi, F i, Ai sind alle an-
deren bier betrachteten semantischen Begriffe definierbar. Zur Definition
der Paradoxie auf Stufe n ("wahr auf irgendeiner Stufe :::;; n und falsch auf
irgendeiner Stufe :::;; n") benotigen wir die natiirlichen Zahlen. Wir wahlen
ein LR-Symbol, den spater definierten Junktor "I", und identifizeren jede
Zahl n mit der Folge von n Strichen, also 0 mit dem leeren Ausdruck, I
mit "I" usw. Da wir voraussetzen, daB jedes LR-Universum alle Aus-
driicke (endlichen Symbolfolgen, einschlieBlich der leeren) enthalt, existie-
ren in jedem Universum die Zahlen und ihre Anfiihrungen als eindeutige
Bezeichnungen: "r 1" flir 0, "r(1" fiir I usw. Die iiblichen, metasprachlich
verwendeten arabischen Zahlbezeichnungen fUhren wir dann definitorisch
in die Objektsprache ein. Damit sind die numerischen Stufenindizes defi-
niert. Einen eigenen Status hat der Index ".", der flir VO die unreftektierte
Stufe 0, und fUr V" + 1 die refiektierte Stufe n bezeichnen solI. Er geh6rt
zur Kategorie der Indikatoren, wie "ich", "jetzt", "hier", "dies" usw. Ihre
Referenz wird entweder direkt, durch die kontextuellen Intentionen des
Sprechers bestimmt, oder indirekt, durch ein vom Sprecher bestimmtes
intentionales Subjekt, oder noch indirekter, bei weiterer Intentionsschach-
telung. Beispiel:
(9a) Du behauptest, daB ich 1iige.
(b) Du behauptest: "Ich liige".
(c) Du behauptest: "Er behauptet: "lch liige" ".
In (9a) bestimmt der Sprecher die Referenten von "du" und "ich"; in (b)
bestimmt er den Referenten von "du" und dieser den Referenten von "ich"
(sich selbst). In (c) bestimmt der Sprecher den Referenten von "du", dieser
den Referenten von "er" und dieser den Referenten von "ich". Vergleich-
bar mit (c) ist
(10) W*W*F*A, d.h. n. Def. w*rw*rF*rA"'.
Die Stufenindikatoren iterierter semantischer Pdidikate verhalten sich so
wie die Personenindikatoren iterierter intentionaler Pdidikate bei direkter
Rede. Wenn z.B. V 7 Satz (10) betrachtet, so bestimmt er den ersten Stem
von links als 6, dieser bestimmt den zweiten als 5 und dieser den dritten
als 4.
So viel zur Grundnotation von LR. A1s nachstes zeigen wir, warum
sechs Wahrheitswerte notwendig sind. Dazu zeigen wir, einfachheitshalber
nur fUr Stufe 0, daB die Anwendung der Junktoren W, F, /\, v auf un-
bestimmte Satze zur Unterscheidung von vier unbestimmten Werten
zwingt. DaB N und 0 unterschieden werden miissen, ist klar: W und F
erzeugen in Anwendung aufneutrale SatzeJalsche (vgl. Beispiel (3) in 1.1.),
aber in Anwendung auf offene Satze wie den Liigner (1) offene:
(11) DaB Satz (1) nicht wahr ist, ist wahr.
DaB Satz (1) nicht wahr ist, ist falsch.
Bei Betrachtung dieser Satze geraten wir auf Stufe 0 in denselben Zirkel
wie bei (1). Daher gilt
(12) A WA FA
N F F
o 0 0
392 VLRICH BLAU
Betrachten wir nun die Tafel t mit den Satzen Al-A6:
A1: Auf t stehen genau 6 Satze.

A2: Auf t stehen genau 7 Satze.
A3: Der 7. Satz auf t ist wahr.
A4: Der 6. und der 7. Satz auf t ist wahr.
As: Der 6. oder der 7. Satz auf t ist wahr.
A6: Der 6. Satz auf t ist wahr.
At. A 2 , A3 sind eindeutig W, F, N (prasuppositionsfehler). A6 ist ebenso

o wie der Liigner. Betrachten wir den Satz:
A4 = A6 1\ A3
MaBgeblich fiir seine Bewertung ist das obige Prinzip P 1\. VO stellt fest,
daB ~ nicht wahr sein kann, da A3 nicht wahr ist. Aber ob A4 falsch ist,
bleibt offen fiir ibn, denn A3 ist nicht falsifizierbar (da neutral) und beim
Versuch, A6 zu falsifizieren, gerat er in den ZirkeP7. Daher gilt auf Stufe
0:
(a) WA4 ist F
(b) FA4 ist 0
- Betrachten wir nun den Satz:
As = A6 V A3
VO stellt nach P v fest, daB As nicht falsch sein kann, da A3 nicht falsch
ist. Aber ob As wahr ist. bleibt offen fiir ihn, da A3 nicht verifizierbar ist
und der Versuch, A6 zu verifizieren, zum Zirkel fiihrt. Daher gilt auf Stufe
0:
(c) FAs ist F
(d) WAs ist 0
Nun folgt aus (12), daB A4 und As nicht 0 sind (nach (a) und (c», aber
auch nicht N sind (nach (b) und (d». Ferner folgt aus (a}-(d), daB A4 und
As weder W noch F sind und verschiedene Werte haben; wir nennen sie
Vi und F. Analog ergibt sich auf jeder Stufe n > 0 das
Spektrum der semantischen Priidikate:
Fundierte Hiilfte IFD
FnW
Unfundierte Hiilfte IUD

Fig. 1.
Der gewohnlich sichtbare Teil der Sprache ist IFO; wenn bisher und kiinftig
von "fundierten" Siitzen die Rede ist, sind stets die der Stufe 0 gemeint.
Sie beruhen auf auBersprachlichen Tatsachen und behalten ihren Wert W,
F oder N auf allen hOheren Stufen. Aber pathologische Siitze aus 1U0 kon-
nen auf hoheren Stufen in die obere Hiilfte P aufsteigen; sie beruhen dann
auf semantischen Tatsachen vorangehender Stufen. Die Siitze aus 0° kon-
nen beliebig oszillieren, die Siitze aus Vlo nur im Halbkreis FDW, die Siitze
aus po nur im Halbkreis PF. Aus diesen Einschriinkungen folgt, daB aIle
Paradoxien 0° sind. 1m iibrigen gibt es Siitze, die je nach Interpretation
in jeder beliebig regellosen, mit den zwei Einschriinkungen vertriiglichen
Weise oszillieren; man kommt mit keinem Teil der Hierarchie aus.18
Zum AbschluB dieses Oberblicks ein ftiichtiger Ausblick. Mir scheint,
daB nicht wenige philosophische Probleme durch unbemerkte oder
schlecht verstandene Reftexionsschritte erzeugt werden. Vergleichbare
Schritte sind:
394 ULRICH BLAU
von innen von aufJen

Standpunkt unreflektiert reflektiert
Objekte, Eigenschaften da konzipiert
Erkenntnis realistisch idealistisch
Wahrheit, Referenz Korrespondenz Kohiirenz
Sprache Ausdruckssystem Regelsystem
Bedeutung intentional konventional
Quantoren, } de re, de dicto,
intentionale Pradikate auf die Welt bezogen auf Ausdriicke bezogen
unfundierte Satze }
offen als offen erkennbar
und Bezeichnungen
Formale Logik Modelltheorie Beweistheorie
unbeschrankt, nicht erwei- beschriinkt, erweiterungsrli-
Mengenuniversum V
terungsfiihig hig
unbeschriinkte Gesamthei- Mengen im niichsten Uni-
auBerste Klassen in V
ten in V versum V'
Ein anderes Beispiel. Raurnzeitliche Objekte -+ Molekularstrukturen -+

Elektronenwolken -+ Partikelhaufen -+ Raum-Zeit-Gebiete -+ Mengen
von Zahlquadrupeln -+ formale Konstrukte aus der leeren Menge -+
formale Ausdriicke -+ Phoneme, Grapheme; d.h. raurnzeitliche Objekte
-+ ... 19 Schwerer zu verfolgen sind Reflexionsprozesse, die mit schwieri-
geren Objekten wie "ich" beginnen.
Es gibt mentale und formale, innovative und konservative, spontane
und mechanische Reflexionsprozesse, und die in LR sind besonders ein-
fach, formal, konservativ, mechanisch. AIle menschlichen Reflexionspro-
zesse haben etwas gemeinsam: Sie beginnen, bleiben und enden innen; es
gibt kein absolutes AuBen fiir uns. Wir setzen immer die einen oder an-
deren Objekte als real voraus, und unberiihrt von ihrem ZerfaIl bei kriti-
scher Reflexion bleibt unser letztes unreflektiertes Weltbild immer realis-
tisch, naiv; es solI die RealWit abbilden. Und wer glaubt, daB es in Wirk-
lichkeit keine, oder mehrere Realitaten gibt, der glaubt dies von der Wirk-
lichkeit. Das
Realitiitsprinzip: Es gibt genau eine Realitat
ist ein synthetisches Urteil a priori: weder beweisbar noch erfahrbar, aber
Vorbedingung jeder Logik und Erfahrung. Auf ihm beruht die Korre-
spondenzvorsteIlung der Wahrheit, auf dieser das Bivalenzprinzip, auf die-
sem die klassische Logik, die hier konservativ erweitert wird.
Zuriick zum Thema. Der unaufhebbare Abbildungsanspruch unseres
UNBESTIMMTHEITEN UND P ARADOXIEN 395
unreflektierten letzten Bildes verbindet sieh leieht mit einem Vollstiindig-

keitsanspruch. Das hat dramatische Folgen. Ein vollstandiges Bild muB
sieh selbst abbilden und ist unfundiert. Und wenn die Negation hinzu-
kommt ("Dieses Bild bildet sieh nieht selbst ab"), entstehen Paradoxien.
Sie werden bleiben, solange man das letzte Bild fiir (a) realistisch, (b) real
und (e) vollstandig halt, denn die Reflexion, ob man's will oder nieht, wird
fortschreiten und jedes letzte Bild im naehsten widerlegen. Die Paradoxien
versehwinden, sobald man den inneren Vollstandigkeitsansprueh aufgibt
und den ProzeB erkennt, der jedes Bild von aufJen - im giinstigen Fall
konservativ - erweitert20 •
Zur Losung der semantischen Paradoxien wird in LR die innere wahr-
heitsfunktionale Vollstandigkeit aufgegeben. Z.B. gibt es keinen Junktor
J mit der Regel:
J A ist {;:} - A ist {~:ht On

DaB A auf Stufe n offen ist, kann erst von auBen, auf hOheren Stufen r
> n, erkannt werden. Daher gibt es in LR einen Junktor on mit der Regel:
W'} -
onA ist {F'
{On
A ist nieht 0· (r > n)
LR ist wahrheitsfunktional vollstandig von aufJen: Fiir aIle k, -n ist ein k-

stelliger Junktor .r definierbar, der Art, daB .r(Al ... AJ fiir alle Bewer-
tungen der Ab ... , Ak auf Stufe n beliebig zugeordnete Werte konstant
auf allen Stufen r > n hat.
Zur Losung der Mengenparadoxien wird in den Standardsystemen eine
ahnliehe Vollstandigkeit aufgegeben: der innere AbschluB, die Allmenge.
Aber die spater erlauterte Intuition der kumulativen Mengenhierarehie
erzwingt fiir jedes korrekte Standardsystem einen ErweiterungsprozeB, der
jeden denkbaren AbschluB von aufJen herstellt. Der letzte AbsehluB ohne
AuBen, das Absolute, ist nieht denkbar.
2. SYNTAX VON LR
Die formale Spraehe21 L von LR hat als Grundsymbole

(a) die logischen Konstanten
"-", "A", "1\" , "w", "F" , "A" , "r ,
1" "=". ,
396 ULRICH BLAU
(b) den Indikator "*";

(c) abzahlbar unendlich viele Objektvariablen
(d) fiir jedes r ~ 1 abzablbar unendlich viele r-stellige Priidikat-

parameter "pr", "Qr", "B.''' "Pi ", "~ ", "B.~ ", "~ ", ...
und zusatzlich, zur Definition der Objekt- und Funktionsparameter, die
I-stelligen Pradikatparameter "~"', "p"', "g"', "~1'" ... und die r + I-stel-
ligen "f"', "~r"', "br"', "fi", ...
(e) Klammern "(", ")".
Neben diesen Grundsymbolen gehoren auch alle spater definierten Symbo-
Ie zur formalen Sprache L. Linkes und rechtes Aruuhrungszeichen sind
unselbstandige Teile eines Symbols. Da wir, wie iiblich, den angefiihrten
Ausdruck dazwischenstellen, entsteht eine Ambiguitat: 1st "r~uy1" die An-
fiihrung von "~", gefolgt von der Anfiihrung von "y" oder aber die An-
fiihrung von ,,~~r y"? Wir lassen nur die erste Antwort zu und schlieBen
Ausdriicke wie "~'ry" syntaktisch aus. Induktive Definition der Ausdrucke
von L:
I. Jedes Grundsymbol und jedes definierte Symbol ist ein Ausdruck.
2. Jede endliche Folge von Ausdriicken, einschlieBlich der leeren Folge,
ist ein Ausdruck.
3. 1st e ein Ausdruck, so ist r e, ein Ausdruck und heiBt Anfiihrung von
e.
Die Anfiihrungen von Ausdriicken und die Variablen sind die Terme in
Grundnotation von L, kurz: "in GN". (Man beachte, daB auch die Anfiih-
rungen von Ausdriicken mit definierten Symbolen Terme in GN sind.)
Induktive Definition der Formeln in GN:
1. h, ... , tr seien Terme in GN, s sei Term in GN, * oder leer, P sei r-
stelliger Pradikatparameter; dann sind
Formeln in GN. Dabei heiBt s, falls vorhanden, Reflexionsstufenindex.

2. A, B seien Formeln in GN, x sei Variable; dann sind
-A, (A 1\ B), AxA
Formeln in GN. Dabei heiBt Ax Allquantor uber x.
Ausdriicke der Gestalt (Xl ..• xr)A, mit r verschiedenen Variablen und
UNBESTIMMTHEITEN UND PARADOX lEN 397
einer Formel A in GN; heiBen r-stellige Priidikatformeln in GN; dabei

heiBt (Xl ... Xr) Argumentstellenmarkierung fiber Xl, ... , Xr. Die Terme,
Formeln und Priidikatformeln in GN heiBen wohlgeformte Ausdrucke in
GN. Als niichstes definieren wir die Bindung
(a) beliebiger Ausdriicke durch "opake" AnfUhrung,
(b) der Variablen durch Bindungszeichen, d.h. Quantoren oder Argu-
mentstellenmarkierungen.
e, et, e2 seien wohlgeformte Ausdriicke in GN. Induktive Definition der
transparenten Anfiihrung:
1. e enthalte aufterhalb aller Anfiihrungen eine Teilformel der Gestalt
wsrA\ psrA\ W"rAntl ... t., psrAntl ... t., wobei s Term oder * oder leer,
A Formel, N r-stellige Priidikatformel ist. Dann ist die Anfiihrung von A
bzw. N transparent in e.
2. 1st die Anfiihrung von e2 transparent in el und die Anfiihrung von
el transparent in e, so ist die Anf"lihrung von e2 transparent in e.
AIle in e nicht transparent vorkommenden AnfUhrungen heiBen opak
in e. (Man beachte, daB die AnfUhrung von e in Llsre't opak ist.)
In AxA und (Xl ... xr)A ist A der Bereich des voranstehenden Bindungs-
zeichens. Dieses birzdet alle Vorkommnisse von X bzw. Xj (j = 1, ... , r) in
A, die
1. nicht in einer AnfUhrung vorkommen, die in A opak ist,
2. nicht im Bereich eines Bindungszeichens von A iiber Xbzw. Xj stehen,
3. nicht Teil eines Bindungszeichens sind.
(Die gebundenen Variablen konnen an Argument- oder oberer Index-
stelle stehen.) Ein wohlgeformter Teilausdruck el von e heiBt frei in e,
wenn er
1. nicht in einer Anfiihrung vorkommt, die in e opak ist,
2. falls er Variable ist, nicht durch ein Bindungszeichen von e gebunden,
oder Teil von ihm ist.
(Man beachte, daB in re, der einzige freie Teilausdruck re, selbst ist,
wiihrend in W"re\ psre' Teilausdriicke von e frei sein konnen.) Bindungs-
zeichen Ax, (Xl ... Xr), die kein Vorkommnis von X bzw. von jedem Xt,
... , Xr binden, heiBen leer. Praktisch spielen sie keine Rolle, sie vereinfa-
chen nur die vorangehenden syntaktischen Regeln. Kiinftig beschriinken
wir uns auf Formeln und Priidikatformeln ohne leere Bindungszeichen.
Die Grundnotation von L wird durch die Definitionen von Abschnitt
4 der Gestalt e : = e' betriichtlich erweitert und der natiirlichen Sprache
398 ULRICH BLAU
angenahert. 1st e' Term, Formel oder r-stellige Pradikatformel von L, so

ist e ein definierter Ausdruck von L derselben Kategorie. Zu den definierten
Termen von L gehoren auch die kontext-definierten Kennzeichnungsterme
zxA an Argument- oder Indexstelle (nach DSz, DS' unten), mit denen wei-
tere Terme definiert werden (durch Da, Df, Dden, Oval unten). Die Bin-
dungsverhaItnisse iihertragen sich in offenkundiger Weise. Ein Beispiel:
(I) "1 "2 "3 der unhestimmte Satz auf der Tafel"3 hezeichnet
nichts" 2 ist unfundiert" 1 ist paradox.
(Die Indizes zeigen, welche AniUhrungszeichen zusammengehOren.) Mit
den spater definierten Pradikatkonstanten
I?*: paradox, IU*: unfundiert, U*: unbestimmt, ST*: Satz
und 1C = ':f2: steht auf, Q1: Tafel (auf die ich jetzt referiere), hat (1) die
logische Form:
(I') I?* rlU* r - V,!A* rz,! (ST*,! A U*,! A :f2'!~Ql,!)1 ,!11
Die heiden auBeren SatzanfUhrungen sind transparent in (1'), die innere
Kennzeichnungsanfiihrung ist opak in (I'), der Existenzquantor bindet nur
das letzte "'!", der erste Kennzeichner die drei folgenden "'!".
So viel zur Syntax. In nachsten Abschnitt erweitem wir die formale
Sprache L, je nach Interpretation qJ, zur (nicht prazisierten) Sprache L.,.
durch Hinzunahme von (nicht prazisierten) "Namen" fiir die
auBersprachlichen Objekte des Bereichs von qJ. Solche Namen mit logi-
scher Referenzgarantie gibt es weder in natiirlichen Satzen noch in ihren
LR-Strukturen, d.h. in Satzen von L. AIle natiirlichen Bezeichnungen
nicht-formaler Objekte sind kontingent: Ihre Referenz hangt von nicht-
logischen Tatsachen abo Die Objektnamen sind ein rein technisches Mittel,
das die iibliche, aber in LR unzweckmaBige, Variablenbelegung vermeidet
und die mogliche Liicke zwischen den in L bei qJ intendierten Objekten und
den abzahlbar vielen, in L bei qJ bezeichenbaren Objekten schlieBt.
3. SEMANTIK VON LR
E sei die Menge der Ausdriicke von L. E : = {re' lee E} sei die Menge
ihrer Anfiihrungen (also E ~ E). Ein Universum fiir LR ist eine Menge U
:::> E. D : = ~E heiBt aujJersprachlicher Objektbereich von U. Fiir jedes
auBersprachliche Objekt d E D setzen wir genau einen Objektnamen d E D

voraus. Ferner setzen wir voraus, daB kein Ausdruck, der Objektnamen
enthalt, in U vorkommt. a: = DuE sei die Menge der Grundbezeich-
a
nungen fiir U. Fiir jedes u E U sei ii E die zugeordnete Grundbezeichnung,
also d = d, e = r e,.
Eine Interpretation qJ iiber U ist (wie in L3) eine Funktion, die jedem
r-stelligen Pradikatparameter P ein Tripel (P:, P;, P~> zuordnet, wobei
~u~u~=~~n~=~n~=~n~=l~~~
D, D fiir gegebenes qJ eindeutig sind, bezeichnen wir sie auch als U"" a""
D"" D",.
Fiir jede Interpretation qJ erweitern wir nun die formale Sprache L zur
Sprache L",: Die Grundsymbole von L", sind die von Lund die Objektnamen
ED",. Die wohlgeformten Ausdrucke, d.h. die Terme, Formeln, Priidikat-
formeln von L", entstehen aus denen von L dadurch, daB 0 oder mehr freie
Variablen durch Objektnamen E D", ersetzt werden. (Demnach sind die
Anjuhrungen, d.h. die Terme der Gestalt re, von L"" nur die Anfiihrungen
von L.) Wohlgeformte Ausdriicke mit freien Variablen heiBen offen, an-
dernfalls geschlossen. Geschlossene Terme, Kennzeichnungsterme, For-
meln, Pradikatformeln von L bzw. L", heiBen Objektbezeichnungen, Kenn-
zeichnungen, Siitze, Priidikate von L bzw. L",. Die Ausdriicke von L bzw.
L", bezeichnen wir auch als reine Ausdrucke bzw. qJ-Ausdrucke. Wir ver-
wenden fiir sie dieselben Metavariablen, deren Liste wir noch einmal zu-
sammenstellen.
a, b, c Objektparameter
a,b, c Terme
d,a auBersprachliches Objekt und zugeordneter Ob-
jektname
e, r e, Ausdruck (evt. leer) und seine Anfiihrung
f, g, h Funktionsparameter
kategorial passender Stufenindex in Grundnotation:
leer, * oder Zahlanfiihrung rj ... f (im Sinn von LR)
j, k, I, m, n natiirliche Zahlen (im Sinn der Metasprache)
n arabische Bezeichnung der Zahl n im Dezimalsystem
ii Anfiihrung der Zahl n (im Sinn von LR, d.h. der Fol-
ge von n Strichen)
p, q, r positive natiirliche Zahlen (im Sinn der Metasprache)
400 ULRICH BLAU
s Stufenindex in Grundnotation: leer, * oder Term in

Grundnotation
t Term in Grundnotation: Objektname a, reine Aus-
drucksanfiihrung r e, oder Variable x
u,v,w auBersprachliche und sprachliche Objekte d, e E U"
ii, V, VI a,
ihre zugeordneten Grundbezeichnungen re, E U"
X,y, z Variablen
A,B,C Formeln
Ar, Br, cr r-stellige Pradikatformeln
Ax1",xr Formel, die mindestens Xl> ••• , Xr ein- oder mehrfach
in beliebiger Reihenfolge frei enthalt
Formel, die diese Variablen nicht notwendig frei ent-
halt.
Aal· ..ar } Formeln, die aus den vorangehenden durch Erset-
A[al" .ar ] zung alIer freien Xj durch aj entstehen
D" auBersprachlicher Objektbereich bei cp
Jj" Menge der zugeordneten Objektnamen
E Menge der reinen Ausdriicke
E MengeihrerAnfUhrungen
J,Q Junktoren, Quantoren
P,Q,R Pradikatparameter
S Symbole (Grundsymbole oder definierte)
U" = D" u E Universum bei cp
U" = Jj" u E Menge der Grundbezeichnungen fiir U".
Wir definieren nun induktiv nach n den semantischen Grundbegriff von

LR.
Eine ;- Verifikation fUr eine cp-Formel C ist ein Baum von cp-Formeln
mit dem Ursprung C, in dem jeder Ast endlich ist und jeder Punkt ent-
weder gemaB einer Definitionsbeseitigung DB oder einer Nachfolgerregel
NI-N18 einen, zwei oder unendlich viele Nachfolger hat oder gemaB einer
Endregel EI-E9 Endpunkt ist.
In NI-N8, EI-E7 sind je zwei Regeln zusammengefaBt; die Schemata
sind (a) unter AusschluB und (b) unter EinschluB der Symbole in den
eckigen Klammem zu verstehen.
DB Definitionsbeseitigung
A
B' wenn auf A keine Nachfolger- oder Endregel anwendbar ist und B
dadurch entsteht, daB in A der erste freie definierte Ausdruck von links
durch sein Definiens ersetzt wird.
lVachfolgerregeln
Nl [_] WirN N2 [_] FirN
[-] WiA [-] Wi-A
N3 [-] WirAojuI ... u r N4 [-] FirrAoUI ...u r

[ -] WiAuI ...u r [-] Wi-Aul ...Ur
fUr r-stellige cp-Pradikate N = (Xl ... Xr) Axl ... Xr
N5 [-] Aira'u N6 [-W]--A
[-] Wia = U [-W]A
N7 [-] W*A N8 [-] WiiA
[ -] W A, falls n = 0 [-]WA
bzw.
[ - ] wn -1, falls n > 0
N9 WA NIO -WA
A -A
NIO gilt fUr cp-Siitze A der Gestalten: [-] WirB\ [-] FirB\ [-] wirBoUI
- ... u- r, [ - ] LIAira1 u-
... u- r, [ - ] FirBo UI
Nll A/\B Nl2 -W-(A /\ B)
AlB -W-AI-W-B
N13 -(A /\ B) Nl4 -W (A /\ B)
-A bzw. -B -WA bzw. -WB
Nl5 I\xAx Nl6 -W-l\xAx
···IAul··· ••• 1-W-Au I···
fiir aIle U E U", fUr aIle U E U",
Nl7 -l\xAx Nl8 -Wl\xAx
-Au fur ein U E -WAu fiir ein U E
U", U",
402 ULRICH BLAU
Endregeln
El [-W-] Pu 1 •.•ft.. p+
E2 -[W] Pii 1 •••iir } wenn (u1. ...,0.) E {:~
E3 - W[ -] Pii 1 •••iir tp
E4 [ - W - ] ii = ii.
E5 -[W] ii = Y, wenn u #: V ist.
E6 - W[ - ]A, wenn A ein I'p-Satz WUV, PUV, WUVW1 W.,
piiVW1 ... W., LlUVw ist und ii keine Zahlanfiihrung ist.
E7 - W[ - ]A, wenn A ein I'p-Satz Wiy, piy, Wiy-w1 ... W., Piy-w1
••• Wr , Llivw ist und y keine Anfiihrung eines I'p-Satzes, bzw.
r-stelligen I'p-Pdidikats, bzw. einer reinen Objektbezeichnung
ist.
E8 WffiA, wenn m < n, A ein I'p-Satz und:-verifizierbar ist.
E9 - WmA, wenn m < n, A ein I'p-Satz und nicht :-verifizierbar

ist.
Diese Verifikationsregeln analysieren den gegebenen Satz nach seiner 10-
gischen Struktur gemaB den Prinzipien:
VPl niedere Stufe vor hOherer Stufe
VP2 Breite vor Tiefe
VP3 links vor rechts.
Ein weiteres Prinzip liegt der spateren Definition Dj
zugrunde 22 :
VP4 Prasuppositionen vor Behauptungen.
Aufjeden Satz ist hOchstens eine Regel anwendbar. DB und NI-NIO sind
lineare Regeln, die den einzigen unmittelbaren Nachfolger eindeutig he-
stimmen. NIl, N12 sind binare Verzweigungsregeln, die die heiden un-
mittelbaren Nachfolger eindeutig bestimmen. N13, N14 sind binare Aus-
wahlregeln fiir den einzigen unmittelbaren Nachfolger, N15, N16 sind
unendliche Verzweigungsregeln, die die" vielen unmittelbaren Nachfolger
eindeutig bestimmen, wobei" die Kardinalzahl von Utp ist, und N17, N18
sind unendliche Auswahlregeln fiir den einzigen unmittelbaren Nachfolger.
Zur Erlauterung der Regeln fiihren wir zwei informelle Begriffe ein.
Wenn B aus A durch Wegfall gewisser Symbole entsteht, so heiBen diese

(a) schwach redundant, wenn die Wahrheit von A intuitiv notwendig
und hinreichend fUr die Wahrheit von B ist;
(b) stark redundant, wenn A und B intuitiv notwendig denselben Wahr-
heitswert haben.
("intuitiv" bezieht sich auf die intendierte Bedeutung der logischen
Grundsymbole.) NI-N5 reduzieren Fi und ..1 i auf Wi und beseitigen die
stark redundanten Anfiihrungszeichen. N5 gilt aufgrund der syntaktischen
Regeln nur fiir -reine Objektbezeichnungen a, d.h. Anfiihrungen, Kenn-
zeichnungen, Objektparameter und funktionale Objektbezeichnungen oh-
ne Objektnamen: Ebenso wie in der natiirlichen Sprache konnen in LR
nur Objektbezeichnungen ohne fiktive Objektnamen etwas denotieren.
N6 beseitigt die stark redundante doppelte starke Negation. (Die dop-
pelte schwache Negation ist nur schwach redundant: A ist wahr genau
dann, wenn -,-,A wahr ist, aber fiir neutrales A ist -,-,A falsch.) N7
gibt dem reflektierbaren Wahrheitsbegriff die gewiinschte Bedeutung: fiir
den nicht-reflektierenden VO absolute Wahrheit, fiir den reflektierenden
p+l n-Verifizierbarkeit. Diese ist nach N8 absolute Wahrheit fur P.
N9 beseitigt den unreflektierten Wahrheitsjunktor. Er ist schwach re~
dundant: A ist wahr genau dann, wenn WA wahr ist. Aber er ist nicht
stark redundant: Fiir neutrales A ist WA falsch. NIO beruht darauf, daB
Siitze A der angegebenen Gestalten nie neutral sind, da die semantischen
Priidikate Wi, F i, ..1 i aIle Vagheiten, Kategorie- und Priisuppositionsver-
letzungen in Falschheitsgriinde verwandeln. Solche Siitze A sind entweder
wahr oder falsch oder offen; in Anwendung auf sie ist W stark redundant
und kann in - WA gestrichen werden. Die Verzweigungs- und Auswahl-
regeln Nll-N18 ergeben sich aus PI\ und PI\, wenn man "-W-" als
"nicht falsch" liest.
Betrachten wir nun die Endregeln. Jeder Ast einer ;-Verifikation muB
nach endlich vielen Schritten den Boden der Tatsachen beriihren. End-
punkte gemiiB EI-E3 driicken Tatsachen aus, die von der Kontextsprache,
also formal von <p, aber nicht von der Reflexionsstufe abhangen. End-
punkte gemiiB E4, E5 sind Ausdruck einer sehr allgemeinen Voraussetzung
der logischen Sprachanalyse, die mir ebenso fragwiirdig wie unvermeidlich
erscheint: der Voraussetzung des Universums als scharf individuierte Ge-
samtheit23 • (Fiir den Sprecher, von innen, ist U eine iiuBerste Klasse; fUr
den Verifikator, von auBen, ist U eine Menge; s. Abschnitt 5.4 unten.)
404 ULRICH BLAU
Endpunkte gemiiB E6-E9 driicken semantische Tatsachen aus; E6, E7

geben Kategoriepriisuppositionen der Grundpriidikate W, F, L1 an: Wenn
ihre Indizes keine Zahlen bezeichnen, wenn die Argumente von W, F weder
Siitze noch Pdidikate und entsprechend viele Objekte bezeichnen, wenn
das 1. Argument von L1 keine reine Objektbezeichnung bezeichnet, so ent-
stehen Siitze, die nicht wahr (" - W") und nicht falsch (" - W - "), sondern
neutral sind. .
Die Regeln E8, E9 iibersetzen, bei gegebenem ((J, die metasprachliche
[Nicht-]Verifizierbarkeit auf Stufe m unmittelbar als [- ]Wiii in die Ob-
jektsprache L'P aufhOheren Stufen n24. Mit den spiiteren Definitionen wer-
den weitere Begriffe iibersetzt und im Prinzip sind aIle iibersetzbar - die
Grenze, vor der wir hier halt machen, ist die metasprachliche Interpreta-
tionsfunktion.
Die vorangehenden Oberiegungen sollten die intuitive Korrektheit der
Verifikationsregeln zeigen. Ferner sind sie vollstiindig und daher intuitiv
adiiquat in folgendem Sinn. Fiir jeden ((J-Satz C gilt genau einer der 4
FiiIle:
(a) Auf C ist eine der Regeln DB, NI-NI8 anwendbar und C ist intuitiv
wahr auf Stufe n bei ((J genau dann, wenn der, bzw. die beiden, bzw. einer
der, bzw. die unendlich vielen, Nachfolger wahr ist (sind).
(b) Chat eine der Gestalten EI-E9 und ist intuitiv wahr auf Stufe n bei
((J genau dann, wenn die angegebene Bedingung erfiiIlt ist (bei E4 entrallt
sie).
(c) Chat einer der Gestalten W'A oder - W'A, wobei r > n. Dann
kann man C intuitiv als offen auf Stufe n bei ((J betrachten und dasselbe
wird sich fonnal ergeben.
(d) - W[ -] C ist ein Satz, der die Bedingung von E6 oder E7 erfiilit.
Dann ist C intuitiv neutral, da kategorial falsch, auf jeder Stufe, und das-
selbe wird sich fonnal ergeben. So viel rur Motivation der Semantik; als
niichstes einige metasprachliche Begriffe. l-stellige semantische Priidikate,
fiir ((J-Fonneln A, sind:
wn'P'.= {AlA ist ~-verifizierbar} ~-wahr
Fn'P - {AI-A E W~} ~-falsch
N°'P - {AI-WA, -W-A E W~} ~-neutral
W~:= {AI-WA E W~; -A, -W-A i W~} ~-nicht-wahr

:r; := {AI-W-A e WD; A, -WA ¢ W;} ;-nicht-falsch

0; := {AI-WA, -W-A ¢ W;} ;-offen
D; := W; u F; ;-definit (bestimmt)
u; := {AlA ¢ D;} ;-indefinit (unbestimmt)
IF; : = D; UN; ;-fundiert
IU; : = {AlA ¢ IF;} ;-unfundiert
I?!!~ : = U wmtp n U Fmtp
mSD mSD
!!-paradox
~
Dieselben Symbole, ohne unteren Interpretationsindex, werden spater fUr

die Objektsprache definiert. Der untere Index zeigt also stets die meta-
sprachliche Verwendung des Symbols an. Die metasprachliche Extensions-
funktion "I I;" definieren wir 1m I',O-Satze A und I',O-Pradikate N =
(Xl'" Xr) Axl .•. Xr total, fiir I',O-Objektbezeichnungen a nur partiell:
· _ das objektsprachliche Symbol S = W, F, N, Vi, F, 0, fiir

.- welches A e S;; Wahrheitswert von A auf Stufe n bei 1',0.25
· _ {<Ub ... , u.) e (Utp)r I AUl ... Ur e W;}; Positivbereich von
· - N auf Stufe n bei 1',0.
:= u, wenn a = U eW;,
:= { nicht existent, wenn nicht fiir genau ein u a = U e W;;
Denotat (Referent) von a auf Stufe n bei 1',0.
1st lei; fiir alle n gleich, bzw. fiir alle 1',0 gleich, bzw. fiir alle n und alle 1',0
gleich, so schreiben wir: leltp, bzw. lel D, bzw.lel.
Der Wahrheitsverlauf von I',O-Satz A bei 1',0 ist die Folge seiner Wahr-
heitswerte IAI~, IAI~, IAI~, ... auf allen endlichen Stufen. Der Wahrheits-
verlauf heiBt
konstant, wenn fiir alle n IAI; = IAI~,
konvergent ab Stufe m, wenn fiir alle n ~ m IAI; = IAI;' .
Reine Satze, die bei allen 1',0 denselben Wahrheitsverlauf haben, heiBen
logisch determiniert. Weitere semantische Begriffe sind wie iiblich definiert.
Fiir I',O-Satze A und I',O-Satzmengen M gilt:
1',0 erfiillt A (bzw. M) auf Stufe n: = A e W; (bzw. MeW;).

406 UNBESTIMMTHEITEN UND PARADOXIEN
Die nachsten Definitionen gelten nur fiir reine Satze A, B und reine Satz-
rnengen M.
A (bzw. Mist n-erfilllbar := 3qJ: A E W; (fiir alle A E M)
A (bzw. M) ist n-giiltig) := "i/qJ: A E W; (fiir alle A E M)
A und B sind n-bikonditional : = "i/ qJ: A E W; <=> B E W;
A und B sind n-iiquivalent "i/qJl: IAI; = IBI;
M n-impliziert A, oder
A ist n-Folgerung aus M := "i/qJ: M c: W; ~ A E W;
M n-priisupponiert A := "i/qJ: M c: D; ~ A E W;
M n-behauptet A : = M impliziert A, aber prasupponiert
nicht A auf Stufe n.
n-Giiltigkeit und -Folgerung symbolisieren wir durch "II--D", n-Bikondi-
tionalitat durch "-H-,D", n-Aquivalenz durch "== ~R" (mit unterem Index
zur Unterscheidung von der spater definierten objektsprachlichen Aqui-
valenz.) Einige Verallgemeinerungen dieser Begriffe:
A ist (allgemein) gUltig : = \fn, qJ: A E w; 26
A ist ~ n-giiltig := "i/r ~ n, qJ: A E ~ C'II--?:n A")
A ist endgiiltig : = 3n: A ist ~ n-giiltig ("II-- A")
Die endgiiltigen Satze bezeichnen wir auch als LR-Theoreme. Ahnlich kon-
nen die anderen Begriffe verallgemeinert werden, z.B.:
A und B sind (allgemein) bikonditional : = "i/n: A -II--D B
A und B sind (allgemein) aquivalent : = "i/n: A == ~R B
usw. Die Beweise der nachsten Metatheoreme sind wegen der Vielzahl der
Verifikationsregeln etwas lang, aber nicht schwierig; wir geben nur das
Beweisprinzip an. Urn zu zeigen:
(1) Aile A E W; haben die Eigenschaft P
geniigt es, zu zeigen:
(2) Fiir aIle ;-Verifikationen V und alle Punkte A von V gilt:
Wenn aIle unmittelbaren Nachfolger von A in V die Eigen-
schaft P haben, so auch A.
Denn angenommen, ein Ao E W; hiitte nicht die Eigenschaft P. V sei eine

;-Verifikation fUr Ao. Nach (2) hatte ein unmittelbarer Nachfolger Al von
Ao nicht die Eigenschaft P, ein unmittelbarer Nachfolger A2 von Al nicht
die Eigenschaft P usw. Dann hatte V einen unendlichen Ast A o, Alo A 2, ... ,
was n. Def. der ;-Verifikation nicht moglich ist. Wir bezeichnen (2) als
;-Verifikationsinduktion.
Satz I, Abschwiichungstheorem: A E W; => - W - A E W;

Beweis durch ;-Verifikationsinduktion.
Korol/ar 1.1: -A E W; => -WA E W; (nach N6)
Satz 2, Konsistenztheorem: A E W; => - WA ¢ W;

Beweis durch ;-Verifikationsinduktion.
Korol/ar 2.1: A E W; => -A ¢ W; (nach 1.1)
Satz 3, Sechswertigkeitstheorem: Bei jeder Interpretation ({) gehOrt jeder

({)-Satz auf jeder Stufe n zu genau einer der Mengen: W;, F;' N;' W;,
F;,O;.
Beweis: aus den Definitionen dieser Mengen mit den vorangehenden
Siitzen und Korollaren.
Satz 4, Stabilitiitstheorem: A E W~ => A E W;

Beweis durch ~-Verifikationsinduktion.
Korol/are 4.1.: A ist allgemein giiltig <=> A ist O-giiltig

4.2: Fiir S = W, F, N, D, IF gilt: S~ c: S;
4.3: Fiir S = 0, U, U..J gilt: S; c: S~
4.4' ~ c: W; u F; uN;
4.5: F~ c: F: u W; u N;
4.6: P; c: O~
4.7: P~ = P~ = 0
Beweis: aus den Definitionen.
Nach 4.2, 4.4, 4.5 hatjeder ({)-Satz bei ({) entweder einen konstanten Wahr-
heitsverlauf VI, V2, V3 oder einen, moglicherweise oszillierenden, Wahr-
heitsverlauf der Art V4, V5, V6:
408 ULRICH BLAU
VI WO, WI, W2, ...

V2 pO, PI, P2,
V3 N°, Nt, N2,
V4 WO, St, S2, wobei sn = Wn, pn oder N n ist
V5 po, St, S2, wobei sn = pn, wn oder Nn ist
V6 0°, St, S2, wobei sn irgendeiner der 6 Werte sein kann.
Alle diese WahrheitsverHiufe sind moglich; es gilt sogar
Satz 5: Es gibt reine Satze, die bei wechselnden Interpretationen jeden

moglichen Wahrheitsverlauf der Art VI-V6 haben.
Beweisgedanke: Ein besonders einfacher Satz dieser Art ist Vx (Px 1\
Wx). Man kann P in unterschiedlicher Weise als unendliche "Satzliste"
interpretieren und erhalt jeden noch so unregelmaBigen Wahrheitsverlauf
der Art VI-V6. - Die logisch determinierten Satze, deren Wahrheitsverlauf
nicht von der Interpretation ihrer Parameter abhangt, konnen in jedem
endlichen Anfangsabschnitt der Stufenhierarchie beliebig oszillieren, aber
im Unendlichen scheint der Verlauf stets periodisch zu sein. (Ich habe dazu
noch kein Resultat.)
4. DEFINITIONEN UND REGELN
Es folgt eine Liste von Definitionen, die spater erlautert werden. Zuvor die
Klammerkonventionen. l-stellige Junktoren, Quantoren, Argumentstellen-
markierungen und l-Operatoren binden enger als die 2-stelligen Junktoren,
und diese binden abnehmend eng in der Reihenfolge:
/\, v, ~, +-+, ==, /
Demnach iiberfliissige Klammern konnen weggelassen werden, ebenso
auBere Klammern und solche in Konjunktions- und Adjunktionsketten
(die nach links geklammert werden sollen.) In den Definientia, nicht in den
Definienda, werden diese Konventionen schon verwendet.
1. Standardjunktoren
DWj Wahrheitsjunktor
s sei Term in Grundnotation oder * oder leer.
WSA:= wsrA'
D-1 Schwache Negation

--,A:= -WA
Dv Adjunktion
(A v B):= -(-A /\ -B)
D--+ Konditional
(A --+ B) : = --, A v B
D+-+ Bikonditional
(A +-+ B) : = (A --+ B) /\ (B --+ A)
2. Standardquantoren
D /\ r-stelliger Allquantor
/\Xl ... xrA : = /\Xl ... /\xrA, fiir verschiedene Variablen
Xl> ... , Xr
DV r-stelliger Existenzquantor
VXl ... Xr:= -/\Xl ... xr-A
D /\' Beschriinkter Allquantor
/\Xl ... xr(A, B) : = /\Xl ... Xr (A --+ B)
DV' Beschriinkter Existenzquantor
VXl ... xr(A, B) : = VXl ... Xr (WA /\ B)
3. Wahrheitswertkonstanten
DWo Wahrheit
Wo:= /\x X = X
DFo Falschheit
Fo:= -Wo
DNo Neutralitiit
No := W r
,
DO o Offenheit
0 0 : = /\x (Wx, WX)
DWo Nicht-Wahrheit
Wo:= 0 0 /\ No
DFo Nicht-Falschheit
Fo:= 0 0 v No
4. Priijunktion (priisupponierende Konjunktion)
D/ (A/B) := (B --+ A) /\ (--,B --+ No) /\ (WB v --,B)

410 ULRICH BLAU
5. Syntaktische Priidikate
t sei Term in Grundnotation, d.h. Objektname, Anfiihrung oder Variable.

DST* Satz
ST*t := W( - W*t v -F*t)
DPR* r-stelliges Priidikat
PR; t : = WVXl ... Xr ( - W*tx 1 ••• Xr V - F*tXl ... x r), wobei
Xb ... , Xr :f= t
DRO* reine Objektbezeichnung
RO*t : = WVx (LI*tx v - LI*tx), wobei x :f= t
6. Semantische Satzpriidikate und entsprechende Junktoren
DD Defmit (Bestimmt)
DSt:= WSt v Ft
DU Indefinit (Unbestimmt)
USt:= -DSt
D~ }Veutral
~St := WSUt I ST*t
DW }Vicht-wahr
WSt : = FWt A USFT I ST*t
DF }Vicht-falsch
F't : = FFt A USWt I ST*t
DO Offen
0'1 : = USWt A USFt I ST*t
DIF Fundiert
1F't : = DSt v ~St
DIU Unfundiert
IUSt:= -Pt
DIP Paradox
1P't : = W S(OOt A (Wt v W* V xW"t) A (Ft v W* V xPt»
I ST*t
DSj l-stellige Junktoren
Analog zu DWj sei fiir S = F, ~, W, F, 0, D, U, IF, IU, IP
SSA := ssrA'
W, F, ~, W,"F, 0 heiBen in dieser Reihenfolge Wahrheitswerte
Vb ... , V 6 •
D= Aquivalenzpriidikat
t =S t' : = ~ t 1\ ~ t' v ... v ~ t 1\ ~ t' I ST·t 1\ ST·t'
D =j Aquivalenzjunktor
(A =S B) := rA' =srB'
7. Semantische Priidikat~ und Termpriidikate
DERF Erfiillung (mit variabler Stellenzahl ~ 2)

ERF'tl ... trt : = WSttl ... tr
DAUT Autologisch
AUT't : = WStt
DHET Heterologisch
HET't:= Ptt
DRF Referentiell
RFSt : = Vx LlStx,
wobei x ¥: s, t
DKRF Koreferentiell
KRF'tt' : = Vx (LI'tx 1\ LI't'x),
wobei x ¥: s, t, t'
DATN Autonym
A TNSt : = LI'tt
8. Zahlen, Anzahlquantoren und anzahlpriisupponierende Quantoren
Dn Zahlkonstanten
n, n+ 1, p, q seien die arabischen Ziffemfolgen, die im Dezimalsystem n,
n + 1, p, q bezeichnen; fi sei die Anfiihrung von n Strichen. Dann sei
n : = ii, fiir jede natiirliche Zahl n
D~p Mindestens p
~ Ixl ... x.A := VXl ... xrA,
~p+ IXl ••• xrA : = VXl ... Xr (Ax l ... Xr 1\ ~PYl ... Yr (AYl
... Yr 1\ (-Yl = Xl V ••• V -Yr = xr»),
wobei Yb ... , Yr verschieden sind und in AXl ... Xr nicht frei vorkommen.
D::s;;p Hochstens p
::S;;PXl ... xrA:= -~p+lxl ... xrA
D>p Mehr als p
>PXl ••. xrA:= ~p+ IXl ••• xrA
412 ULRICH BLAU
D<p Weniger als p

<PXl ... xrA:= - ~PXl ... xrA
Op Genaup
PXl ... xrA : = ~PXl ... xrA 1\ ~PXl'" xrA
In den 3 nachsten Definitionen sei e * ~, ~, >, < oder leer.
Oep! Verschiirfte Anzahlquantoren
ep!xl .. , xrA : = epXl .,' xrWA
Dep!' Beschriinkte Anzahlquantoren
ep! Xl", xr (A, B):= epxl , .. Xr (WA 1\ B)
DQcPI Beschriinkte anzahlpriisupponierende Quantoren
Qcp! Xl .. , Xr (A, B) : = QXl .. , Xr (A, B) / ep! Xl '.' xrA
fiir beliebige Quantoren Q == /\' V, eq, eq!
9. Kennzeichnung
OZ' Beschriinkter Quantor mit Einzigkeitspriisupposition

ZXl ... Xr (A, B) : = /\l! Xl ,., Xr (A, B)
OSz Kennzeichnungen an Argumentstellen
S sei ein Pradikatparameter oder eine Grundkonstante =, W, F, L1 oder
eine der in 5.-7, definierten Konstanten ST, , .. , ATN (aber nicht die in 10,
definierte Existenzkonstante E!), s sei wie bisher leer, * oder Term in
Grundnotation, r sei die Stellenzahl von S, S·. 'Yl .,. Yp .. entstehe durch
Anwendung von S· auf r Terme in Grundnotation, unter ihnen die p ver-
schiedenen Variablen Yt.,." YP (ein- oder mehrfach, p ~ r), die in s, AIXt.
,.,' ApXp nicht frei vorkommen sollen, S·. , (zxlAlXl) .,. (zxpApXp). , ent-
stehe aus SS . 'Yl ... Yp .. durch Ersetzung alIer Yj durch (zxjAjxj) (j = 1,
.. " p). Dann sei
S· .. (zxlAlXl) .. , (zxpApXp). , : =
ZYl , .. YP (WS (AIYl 1\ ... 1\ ApYp), S· . 'Yl , .. Yp .. )
OS' Kennzeichnung an Indexstelle
S sei eine Grundkonstante, W, F, L1 oder eine der 6.-7. definierten Kon-
stanten 0, .. " ATN oder die in 10. definierte Existenzkonstante E!; SYe
sei eine Formel, wobei Y in e und Ax nicht frei vorkommen soli. Dann sei
S·xAxe := zy(Ay, SYe)
UNBEST1MMTHEITEN UND PARADOXIEN 413
10. Existenz. Denotation und kontingente Objektbezeichungen
In den folgenden Definitionen seien a, at. ... , a r beliebige, aueh definierte,

Terme.
DEI Existenzpriidikat
Elsa:= WSa = a
Da Objektparameter
Fiir die l-stelligen Pradikatparameter a' = "3"', "g1 ", ... sei
a := lXa'x
(also "3" := "l~3'~", "31":= "l~3i~" usw.)
Df Funktionsparameter
Fiir die r + l-stelligen Pradikatparameter F' = "pm, "Pi", ... sei
F (a1 ... ar) : = lxF' a1 ... arX,
wobei X in at. ... , a r nieht frei vorkommt
(also "f(~1 ... ~r)" : = "zyf'~1 ... ~rY" USW.)
Dden Denotat
denea : = lX Lleax
wobei e leer, * oder ein beliebiger, aueh definierter, Term ist und x in e,
a nieht frei vorkommt; ebenso in der nachsten Definition.
Dval Wahrheitswert
valea : = lX (x = 'W'. /\ wea V X = rF /\ Fea v ... v x
= '0' /\ oea), mit 6 Adjunktionsgliedem.
Die Adaquatheit dieser Definitionen zeigen ihre semantischen Regeln,
die in den nachsten Abschnitten 4.1-4.10 angegeben sind. Die Beweise der
Regeln sind im allgemeinen einfaeh, aber zum Teil etwas langwierig und
wir verziehten bier darauf. Zuvor noeh die Regeln fiir Priidikatparameter
und Identitiit:
414 ULRICH BLAU
Diese Elementarsatzregeln fiir q>-Grundbezeichnungen fi (q>-Objektnamen

und Ausdrucksanfiihrungen) folgen unmittelbar aus EI-ES; fUr definierte
qrObjektbezeichungen a, deren Referenz von q> und manchmal auch von
der Reftexionsstufe abhiingt, sind die Regeln komp1izierter (vgl. RPa und
R=a in 4.10.)
4.1. Standardjunktoren
Aus DWj - D- folgen die Regeln:

R-, RW, R-, A WA -,A RA W F N VI F 0
W F W F W W F N Vi F 0
F W F W F F F F F F F
N N F W N N F N Vi N W
iN F F W Vi VI F iN VI W W
F iN 0 0 F F F N VI F 0
0 0 0 0 0 0 F VI VI 0 0
Rv W F N Vi F 0 R~ W F N VI F 0
W W W W W W W W W F N VI F 0
F W F N VI F 0 F W W W W W W
N W N N N F F N W W W W W W
Vi W iN N VI F 0 VI W W W W W W
F W F F F F F F W 0 F 0 F 0
0 W 0 F 0 F 0 0 W 0 F 0 F 0
R- W F N Vi F 0
W W F N VI F 0
F F W W W 0 0
N N W W W F F
VI VI W W W 0 0
F F 0 F 0 F 0
0 0 0 F 0 0 0
R /\ besagt z.B. fUr alle n, qJ und alle qJ-Siitze A, B:
IAI; = W, IBI; = W = IA /\ BI; = W,

IAI; = W, IBI; = F = IA /\ BI; = F usw.
Das obere Drittel der l-stelligen Tafeln und das linke obere Neuntel der
2-stelligen Tafeln enthiilt die Junktorenregeln von L2; "-" und "I" fal-
len bier zusammen, "W" wird zum redundanten Affirmator. Die obere
Hiilfte der l-stelligen Tafeln und das linke obere Viertel der 2-stelligen
Tafeln enthalt die Junktorenregeln von L3.
4.2. Standardquantoren
Ihre semantischen Regeln sind kompliziert; wir notieren nur die der 1-
stelligen unbeschriinkten Quantoren.
R/\ l/\xAxI; =
W <=> Vii IAiil; = W,
F <=> 3ii IAiil; = F,
N <=> 3ft IAftl; = N & 'Vii IAiil; = W, N oder F,
W <=> (3ft IAiil; = VI v (3ft IAftl; = N & 3ft IAftl; = 0» & '9'ii IAftl; ¥=
F,
F <=> 3ft IAft!; = F & Vii I Aft~ = W oder F,
o <=> 3ft IAft!; = 0 & 'Vft IAft~ = W, F oder O.
RVIVxAxl; =
W <=> 3ft IAftl; = W,
F <=> 'Vft IAftl; = F,
N <=> 3ft IAftl; = N & 'Vii I Aft~ = F, N oder VI,
VI <=> 3ft IAftl; = VI & 'Vii IAftl; = F oder W,
F <=> (3ft IAiil; = F v (3ft IAftl; = N & 3ft IAftl ; = 0» & 'Vii IAftl ; ¥=
W,
o <=> 3ft IAiil; = 0 & 'Vii IAft~ = F, W oder O.
416 ULRICH BLAU
Die meisten sprachlich vorkommenden Quantoren sind beschriinkt:

(1) AIle
Einige
Mindestens fiinf
Die beiden
Ax sind Bx.
Die Ralfte der
Viele
Wenige
Die meisten
Es erscheint natiirlich, allen Satzen dieser Art eine einheitliche logische
Struktur zu geben:
(1') Qx (Ax, Bx),
mit einem Quantor Q von Formelpaaren zu Formeln. Nur die vier ersten
Quantoren in (1) sind mit unbeschrankten Quantoren definierbar; die vier
anderen bleiben hier auBer Betracht27 • Mir scheint, daB sich Aussagen der
Art (1) im allgemeinen auf den Positivbereich von (x)Ax beschranken; ne-
gative und neutrale FaIle fiir dieses Pradikat haben keinen EinftuB auf den
Wahrheitswert von (1). Diese Auffassung liegt D/\' und DV' zugrunde.
Betrachten wir ein Beispiel. 1m Raum befinden sich 5 Personen: 3 schi-
zophrene Patienten und 2 maBig gestorte Arzte im Grenzbereich der Schi-
zophrenie. Die Parameteriibersetzung sei:
P: Person, P 1: Patient, P 2 , Arzt, Q: schizophren, R: im Raum.
Bei entsprechender Interpretation sind die Strukturen der folgenden Satze
auf allen Stufen W:
(2a) AIle Patienten im Raum sind schizophren.
/\x (P1X /\ Rx, Qx)
(b) AIle Schizophrenen im Raum sind Patienten.
/\x (Qx /\ Rx, P1x)
Die Strukturen der folgenden sind F:
(3a) AIle Personen im Raum sind eindeutig schizophren.
/\x (Px /\ Rx, WQx)
(b) Mindestens ein Schizophrener im Raum ist Arzt.
Vx (Qx /\ Rx, P 2 x)
Die Strukturen der fo1genden sind N:

(4a) Alle Personen im Raum sind schizophren.
Ax (Px /\ Rx, Qx)
(b) Mindestens ein Arzt im Raum ist schizophren.
Vx (P2X /\ Rx, Qx)
(3b) und (4b) zeigen, daB eine k1assisch gewohnte Aquiva1enz in LR - und
schon in L3 - entrallt: Die beiden Satze sind allgemein bikonditional; die
Wahrheit des einen ist notwendig und hinreichend fUr die Wahrheit des
anderen. Aber sie sind nicht allgemein iiquivalent, sondem konnen ver-
schiedene Werte haben. Diese1be Differenz, die hier aus Vagheitsgriinden
entstand, kann auch aus nicht-erfUllten Prasuppositionen entstehen. Bei-
spiel: Alle Anwesenden seien Junggesellen.
(5a) Mindestens einer, der alter als seine Frau ist, ist anwesend.
Vx (Qxf(x), Px)
(b) Mindestens ein Anwesender ist alter als seine Frau.
Vx (px, Qxf(x»
P: anwesend, Q: alter a1s, f: Frau von
Fur jeden, der alter a1s seine Frau ist, ist die Behauptung "anwesend"
falsch. Aber fUr jeden Anwesenden ist die Behauptung "alter als seine
Frau" neutral. Daher wird bei jeder entsprechenden Interpretation auf
jeder Stufe die Struktur von (5a) F, die von (b) N. Einige einfache Theo-
reme:
AXl ... Xr (A, B)
AXl ... Xr (A, -B) =LR - VXl ... Xr (A, B)
-AXl ... Xr (A, B) =LR VXl ... Xr (A, -B)
-AXl ... Xr (A, -B) =LR VXl ... Xr (A, B)
4.3. Wahrheitswertkonstanten
Diese Konstanten So kann man als O-stellige Junktoren, O-stellige Pradi-

katkonstanten oder Satzkonstanten betrachten. Sie sind so definiert, daB
sie bei jeder Interpretation auf jeder Stufe den Wert S haben.
RS o ISol = S, fiir S = W, F, N, W, P, o.
418 ULRICH BLAU
DaB 100 1= 0 sein sollte, erscheint zunachst absurd. Was konnte wahrer
sein als
(1) I\x CWx -+ Wx)
Aber dieser Satz sagt etwas iiber aile Wahrheiten aus, auch iiber sich selbst,
falls er wahr ware. Daher fiihrt jeder Verifikationsversuch zum Zirkel, wie
man leicht erkennt. Aber eine unabweisbare Intuition sagt uns, daB
(2) Alles Wahre ist wahr
ganz einfach wahr sein mufJ. Diese Intuition wird nicht durch (1) erfaBt
sondern durch
(2') I\x CW*x -+ W*x)
(2') ist auf Stufe 0 offen, aber auf allen hoheren Stufen giiltig, also end-
giiltig, d.h. LR-Theorem. Die iibrigen Fane von RS o sind klar: wr , ist
kategorial falsch, d.h. neutral bei allen n, cp und Wo, Fo entstehen durch
konjunktive bzw. adjunktive Verbindung des Offenen und Neutralen.
4.4. Priijunktion
A/B ist eine prasupponierende Konjunktion, zu lesen: "A, mit der Pra-
supposition B". Aus D/ folgt
R/ W F N Vi F 0 D.h
W W N N N 0 0 { IAI;' wenn B e W;
F F N N N 0 0 IA/BI; = N, wenn -,B eW;
N N N N N 0 0 0, sonst.
Vi iN N N N 0 0
F F N N N 0 0
0 0 N N N 0 0
Daraus folgt: A/B e D; =;. B e W;, d.h. A/B prasupponiert B auf allen
Stufen. Ferner gilt: 1- (A/B)I; = I( - A)/BI;, d.h. die starke Negation laBt
die Prasupposition unberiihrt. Mit diesem Junktor lassen sich Aussagen
in Behauptungen und Prasuppositionen aufspalten. Betrachten wir als Bei-
spiel die restriktive und attributive Lesart von Adjektiven, Adverbien und
Relativsatzen, wobei wir vorgreifend den Kennzeichnungsoperator und die
anzahlprasupponierenden Quantoren (mit unteren Indizes) verwenden.
(1) Der unzufriedene Horer ging.

Der Horer, der unzufrieden war, ging.
Beide Satze sind ambig. Wenn der Akzent auf "Unzufrieden" liegt, ist die
Bestimmung restriktiv und gehort zur Kennzeichnung:
(la') RIX (Px A -Qx), P: (war) Horer, Q: (war) zufrieden, R: ging.
Bei dieser Auffassung prasupponieren die Satze (I) intuitiv und formal,
daB es genau einen unzufriedenen Horer gab, und sie behaupten, daB er
ging. Aber "unzufrieden" kann, unbetont, auch attributiv verstanden wer-
den:
(1 b') RlxPxl - QIXPX
SO verstanden prasupponieren die Satze (I), daB es genau einen Horer gab
und daB er unzufrieden war; sie behaupten, daB er ging. Dieselbe Ambi-
guitat entsteht im Plural:
(2) Die unzufriedenen Horer gingen.
Die Horer, die unzufrieden waren, gingen.
Bei restriktiver Auffassung (a) prasupponieren die Satze, daB es minde-
stens zwei unzufriedene Horer gab; sie behaupten, daB alle unzufriedenen
Horer gingen. Bei attributiver Auffassung (b) prasupponieren die Satze,
daB es mindestens zwei Horer gab, und daB aile Horer unzufrieden waren;
sie behaupten, daB alle Horer gingen. Dasselbe ergibt sich formal:
(2a') 1\~2! x(px A -Qx, Rx)
(b') 1\~2! x(Px, Rx)ll\x (Px, -Qx).
Noch eine Bemerkung zum unfundierten Teil von R/. Die Spalten 5 und
6 sind durch Verifikationsprinzip
VP4 Prasuppositionen vor Behauptungen
motiviert. Solange die Frage der Wahrheit oder Nicht-Wahrheit der Pra-
supposition B offen ist, ist der Wert von AlB offen, selbst dann, wenn der
Wert von A determiniert ist.
4.5. Syntaktische Priidikate
Die Analyse der logischen Syntax gehort nicht zu den Zielen von LR, aber
420 ULRICH BLAU
die Begriffe "<p-Satz", "r-stelliges <p-Priidikat", "reine Objektbezeichnung"

sind auf semantischem Wege fiber die Kategoriepriisuppositionen der End-
regeln E6-E8 definierbar. Diese Definitionen sind auf allen hOheren Stufen
adaquat. Auf Stufe 0 fehlen im Positivbereich von ST* die O-offenen Siitze;
im Positivbereich von PR: die r-stelligen Priidikate, die nur O-offene Siitze
bilden; im Positivbereich von RO* die reinen Objektbezeichnungen, deren
Referenz O-offen ist:
{0)
RST*
{U ist ein <p-Satz A e O~,
IST*iiJ~ = W <=> U ist ein <p-Satz A ¢ O~,
F u ist kein <p-Satz.
W} {u ist ein <p-Satz,

ST*u 0 + 1 = { <=>
I I", F u ist kein <p-Satz.
I Im:
RPR:
r 0 \ r u ist ein <p-Priidikat (Xl ... Xr) AXl ... X., wobei
IPR:tiIO = <=> v"'.
all~ iit. ..., iir e Aiil ... ~r ~O~,
'" W u 1st em anderes r-stelliges <p-Pradikat,
F J u ist kein r-stelliges <p-Priidikat.
_ 0 +1 _ {W} {u ist ein r-stelliges <p-Priidikat,
I
IPR:UI", - F <=> u ist kein r-stelliges <p-Pradikat.
I
RRO*
1
IRO*-I o _
j 0 u ist :ine!eine_O~jekt~zeichnung a, wobei fUr
alle v e U", a - v eO""
u '" - W <=> u ist eine andere reine Objektbezeichnung,
l F l u ist keine reine Objektbezeichnung.
*_ 0+1 _ {W} {u ist eine reine Objektbezeichnung,

IRO ul", - F <=> u ist keine reine Objektbezeichnung.
Die Unvollstiindigkeit der syntaktischen Pradikate aufStufe 0 ist der Preis

ihrer semantischen Definition fiber den Wahrheitswert. Um sie syntak-
tisch, fiber die Form, schon auf Stufe 0 zu definieren, miiBte man LR
erweitem: entweder urn die Verkettungstheorie 28 oder gleich urn eine Men-
gentheorie. Das sp~itere System LRe, Abschnitt 6, ist fiir die bescheidenen
Zwecke der Syntax-Analyse schon ganz unn6tig stark.
4.6. Semantische Satzpriidikate und entsprechende Junktoren
Ahnlich wie die Grundkonstanten, W, F, L1 haben die definierten Pdidi-

katkonstanten D, ... , P, Kategoriepriisuppositionen fiir ihre Indizes und
Argumente.
Fur S = W, F, N, VI, F, 0, D, V, IF, IU, P gilt:
ISiivl", = N, wenn u keine Zahl (im Sinn von LR, d.h. Strich-
folge) oder v kein q>-Satz ist.
R==Kat Iv ==UWI", = N, wenn u keine Zahl (im Sinn von LR) oder v
oder w kein q>-Satz ist.
Die Kategorieprasuppositionen der Argumentstellen sind genau dann er-
fiillt, wenn v, Vi Anfiihrungen von q>-Satzen rN, rB' sind. Diese Anfiih-
rungen sind transparent in ssrA' und rA' ==s rB' und k6nnen nach DWj ,
DS j , D == j wegfallen, wodurch sich die Pradikatkonstanten in lunktoren
verwandeln. Fur die unreflektierten lunktoren auBer P gilt:
RS A WA FA NA VIA FA OA DA VA IFA IUA
W W F F F F F W F W F
F F W F F F F W F W F
N F F W F F F F W W F
Vi F 0 0 0 F F 0 0 0 0
F 0 F 0 F 0 F 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0
Die obere Ralfte dieser Rege1n, ihr dreiwertiger Teil, laBt sich so zusam-
menfassen:
Fiir S= W, ... , IU und q>-Satze A E IF; gilt:

ISAI; = {W, wenn A E S;,
F, sonst.
Mit diesen unreflektierten lunktoren sind also Aussagen uber fundierte
Satze der gegebenen Stufe adaquat formalisierbar: SA ist wahr oder falsch
422 ULRICH BLAU
auf Stufe n bei cp, je naehdem, ob A die metaspraehliehe Eigenschaft S;

hat oder nieht. Aber fiir Aussagen fiber unfundierte A sind die unreftek-
I
tierten Junktoren im allgemeinen ungeeignet, wie die untere HaIfte von RS
zeigt. Ahnliehes ergibt sieh fUr den unreftektierten Paradoxiejunktor:
I
RI?
W
F } <=>
A¢
A ot
e I?;, oder wmA eF;, fUr alle m ::s;; n,
oder FmA e F;, fUr alle m ::s;; n,
t0 sonst.
V" erkennt genau die Sitze, die bis zur Stufe n einschlieSlieh mindestens
einmal wahr und mindestens einmal falseh sind, als paradox, im absoluten
Sinn. 1m wesentliehen erkennt er aueh die Nieht-Paradoxien seiner Stufe
als nieht-paradox, im absoluten Sinn. Aber fUr gewisse unfundierte Nieht-
-Paradoxien seiner Stufe bleibt die Entseheidung offen. Fiirdie unreftek-
tierte Aquivalenz gilt:
R== W F N W F 0
W W F F F 0 0
F F W F 0 F 0
N F F W 0 0 0
VI F 0 0 0 0 0
F 0 F 0 0 0 0
0 0 0 0 0 0 0
Das linke obere Viertel, der dreiwertige Teil, liSt sieh so zusammenfassen.
R==3
Fiir cp-Sitze A, B e IF; gilt:
IA == BI; = { W, wenn IAI; = IBI;,
F, sonst.
RS, RI?, R == zeigen, daB semantische Aussagen fiber fundierte Sitze der
gegebenen Stufe mit den unreftektierten Junktoren adaquat formalisierbar
sind. Aber Aussagen fiber unfundierte Satze verlangen im allgemeinen
hOhere Reftexionsstufen und gestufte oder reftektierbare Junktoren, deren
Regeln wie folgt lauten:
RS, RI?, R== zeigen, daB semantische Aussagen iiber fundierte Satze der
gegebenen Stufe mit den unreflektierten Junktoren adaquat formaliserbar
sind. Aber Aussagen iiber unfundierte Satze verlangen im allgemeinen
hiihere Reflexionsstufen und gestufte oder reflektierbare Junktoren, deren
Regeln wie folgt lauten:
RS iii Fiir S W, F, N, Vi, P, 0, D. U, IF, IU, I? gilt:
=
W, wenn m < n und A ES:;,
ISiiiAID = { F, wenn m < n und A ¢S:;,
'P ISAI;, wenn m = n,
0, wenn m > n.
V D versteht die objektsprachlichen sm (m < n) bei qJ im Sinn der meta-

sprachlichen S:;. Die SD seiner eigenen Stufe versteht er absolut als S, die
Sf (r > n) versteht er gar nicht. Analog gilt:
R==iii
W, wenn m < n und IAI:; = IBI:;,
IA = iii BID = { F, wenn m < n und IAI: =F IBI:,
- 'P IA == BI;, wenn m = n,
0, wenn m > n.
Die reftektierbaren semantischen Begriffe S* versteht VO als absolute, un-

reflektierte S, und vn + 1 als SD:
RS* Fiir alle vorangehenden und spateren, undefinierten und de-

finierten Symbole S* gilt:
IS*el~ = ISeI~, oder beides existiert nicht,
IS*el; + 1 = ISDel; + 1, oder beides existiert nicht.
Die mogliche Nichtexistenz betrifft die Denotate von 'den* a' und 'val* a',
vgl. Abschnitt 4.10 unten.
4.7. Semantische Priidikat~ und Termpriidikate
AIle Pradikate ERF, ... , ATN haben Kategorieprasuppositionen. Z.B. gilt

fiir das Erfiillungspradikat:
RERFKat IERFiiVl" ,vrwl'P = N, wenn u keine Zahl (Strichfolge) oder w

kein r-stelliges qJ-Pradikat ist.
424 ULRICH BLAU
Wir geben die Kategorieregeln kunftig nicht an und beschranken uns auf
kategorial passende Stufenindizes i = leer oder iii (i = * fallt unter RS*
oben) und passende Argumente in Grundnotation. Fur r-stellige qJ-Pra-
dikate N = (Xl ... xr) Axl ... Xr gilt:
RERF
W, wenn m < n und (Ub ... ,ur) E IAr:,
IERFiii - - rAnl n = { F, wenn m < n und (Ub ... ,ur) ¢ IAr: ,
Ul·· .Ur '" IERF-Ul ... Ur
- rAnln"" wenn m -_ n,
0, wenn m > n.
I
Fur l-stellige qJ-Pradikate AI = (x)Ax gilt:
RAUT IAUTrAl'l~ = IWNAlll~.
W, wenn m < n und Al E IAII:,
IAUTmrAI'I~ = F, wenn m < n und Al ¢ IAI:,
IAUTrAIll~, wenn m = n,
0, wenn m > n.
RHET IHErAIll~ = IFNAl1l~.
W, wenn m < n und Al E I(x)-Axl;,
IHETmrAI'I~ = { F, wenn m < n und Al ¢ I(x)-Axl;,
IHErA 111~, wenn m = n,
0, wenn m > n.
Um ein Beispiel zu geben, zeigen wir, daB "l-stelliges Pradikat" autolo-
gisch ist.
If- ~m AUTir(x)PRix\ fiir i = leer, * oder ZahlanfUhrung m.
Beweis: Fur alle n ~ m und alle qJ gilt:
1. U = U E W~, fUr beliebiges U E U", E4
2. U = U ¢ o~ 1, Satz 3
3. '(x)x = x' ist ein l-stelliges Pradikat, wobei nicht fUr alle U E
(J U=UEon 2
4. '"
PRir(x)x = x' EW;'" 3, RPR*
5. '(x)PRix' ist ein l-stelliges Pradikat, wobei nicht fUr alle U E
U", PRiu E O~ 4
6. PRir(x)PRix' E W~ 5, RPR*
1. AUTir(x)PRix' E W~ 6, RAUT, RW, RS*.
Die nichsten Regeln fiir referentiell, koreferentiell, autonym gelten fUr

reine Objektbezeichnungen a, b, die definierte Symbole enthalten kannen.
RRF IRFira'l; = lElia!;
RKRF IKRFiralrb'l; = IWi a = bl;
RATN IATNira'l; = IWi a = [a]l;
Ein Beispiel fiir A TN steht auf der Tafel t: det lingste Ausdruck
auft
Dann gilt:
1. der lingste Ausdruck auf ~ = "der langste Ausdruck auf t"
2. "der lingste Ausdruck auf t" bezeichnet "der lingste Ausdruck auf t"
3. "der lingste Ausdruck auf t" ist autonym.
Dasselbe formal fiir die Tafel t': I
lXPX I
cp sei eine Interpretation, die Pals "Ausdruck maximaler Liinge auf t'"
interpretiert, d.h.
1. P: = {lxPx}. Dann sind die folgenden Satze eW;:
2. P[JxPx] 1, El
3. IlxPx 1, mit der spiteren Regel Rpl
4. lXPx = [JxPx] 2, 3, mit der spiteren Regel Rl =
5. ATN[zxPx] 4, RW, RATN.
4.8. Zahlen, Anzahlquantoren, anzahlpriisupponierende Quantoren
Fiir die Zahlkonstanten gilt:

Rn Inl = Zahl n im Sinn von LR, d.h. Folge von n Strichen.
Die Regeln der Anzahlquantoren gemiS D ~p - Dp sind kompliziert und
wir lassen sie aus. Stattdessen betrachten wir ein Beispiel, das eine kleine
Ambiguitit der natiirlichen Anzahlquantoren verrat. 1m gegebenen Kon-
text habe "Kontinent"
den Positivbereich P: = {Asien, Afrika, Amerika, Europa, Australien},
den Neutralbereich P~ = {Antarktis},
den Negativbereich P; = Rest des Universums von cpo
Nach den obigen Definitionen sind bei cp auf allen Stufen die Satze W:
(1) Es gibt mindestens 5 Kontinente. ~5xPx
426 ULRICH BLAU
(2) Es gibt hOchstens 6 Kontinente. =s;;6xPx

(3) Es gibt genau 5 klare FaIle von Kontinenten. 5xWPx
(4) Es gibt genau einen Grenzfall. IxNPx
Die folgenden sind F:
(5) Es gibt weniger als 5 Kontinente. <5xPx
(6) Es gibt mehr als 6 Kontinente. >6xPx
Und die foigenden sind N:
(7) Es gibt genau 5 Kontinente. 5xPx
(8) Es gibt mehr als 5 Kontinente. >5xPx
(9) Es gibt genau 6 Kontinente. 6xPx
(10) Es gibt weniger als 6 Kontinente. <6xPx
Die Werte W, F der Satze (I)-{6) entsprechen eindeutig der Intuition, aber
die Werte von (7)-{IO) sind intuitiv weniger deutlich. Angesichts der kon-
textuell unterstellten Vagheit von "Kontinent" konnte man sie als neutral
betrachten; dieser Auffassung entsprechen die Definitionen D ~p - Dp.
Aber dann entfallen einige klassisch giiltige Schliisse, z.B.
(11) ~5xPx 1f-L2 5xPx v >5xPx
Denn (1) ist W, aber (7) v (8) ist N. - Ebenso entfiillt
(12) =s;;6xPx If-L2 6xPx v <6xPx
Denn (2) ist W, aber (9) v (10) ist N.
Offenbar gibt es noch eine zweite Auffassung, bei der die Anzahlquan-
toren in (7)-{IO) sich auf den Positivbereich von "Kontinent", die 5 klaren
FaIle beschranken. So verstanden sind (7) und (10) W, (8) und (9) sind F
und die klassisch giiltigen Schliisse (11), (12) sind intuitiv giiltig. Dieser
zweiten Auffassung entspricht die Definition Dep! der verschiirften An-
zahlquantoren. Ihre Regeln lauten im I-stelligen Fall (der mehrstellige ist
vollig analog):
R~p!
fiir mindestens p u ist Au E W; ,
l~p!xAxl; =
{~}-{ fUr weniger als p u ist WAu E W;
sonst.
U 0;,
R=S;;p!
{n-\ fiir h6chstens p u ist W Au E W;

fUr mehr als p u ist Au E W;,
sonst.
U 0; ,
R>p!
fiir mehr als p u ist Au E W~ ,
l>p!xAxl~ = fUr h6chstens p u ist WAu E W~ U O~ ,
sonst.
R<p!
I! }-l
fUr weniger als p ii ist WAu E W~ U O~,
I I
l<p!xAxl~ = fiir mindestens p u ist Au E nW~,
sonst.
I
Rp!
W fUr genau p u ist Au E W~ und fUr kein
u ist Au E O~,
F <:;. fUr weniger als p u ist WAu E W~ u 0:
oder fiir mehr als p u ist Au E W~ ,
o sonst.
Fiir diese verscharften Anzahlquantoren wird die klassische Giiltigkeit,

Folgerung, Erfiillbarkeit in L3 und LR konserviert.
Ebenso wie die verscharften Anzahlquantoren und die beschrankten All-
und Existenzquantoren beziehen sich die beschriinkten Anzahlquantoren
und die anzahlpriisupponierenden Quantoren nach Dep!' und DQep! nur auf
den Positivbereich der Beschrankung AXl ... xr • Wir verzichten auf die
semantischen Regeln und betrachten einige Beispiele.
(13) Alle Kinder von Hans schlafen.
Beschrankte Allsatze auBerhalb der Mathematik und Logik prasuppo-
nieren im allgemeinen, daB der Bereich nicht-leer ist:
(13') ;\~l!x(Pxa, Qx), P: Kind von, a: Hans, Q: schlaft
Gew6hnlich wird die Prasupposition sogar starker sein; man wird (13) als
synonym betrachten mit
(14) Die Kinder von Hans schlafen.
Dann sind mindestens zwei Kinder prasupponiert:
(14') ;\~2!x(Pxa, Qx)
Ein Beispiel zur Schachtelung anzahlprasupponierender Quantoren:
428 ULRICH BLAU
(15) Hochstens drei der mehr als dreiBig Priiflinge haben beide
Aufgaben gelOst.
(lS') ::;;3!>30IX(PX, 1\2! y(Qy, Rxy»
P: Priifling, Q: Aufgabe, R: hat gelost
(15) und (lS') prasupponieren intuitiv und formal die Existenz von mehr
als dreiBig Priitlingen und genau zwei Aufgaben. Ihre starken Negationen
sind intuitiv und formal aquivalent mit
(l6) Mindestens vier der mehr als dreiBig Priiflinge haben beide
Aufgaben gelost.
(l6') ~4!>30IX(PX, 1\21 y(Qy, Rxy»
4.9. Kennzeichnung
Die Kennzeichnungstheorie gemaB Dl', DSl, DS' ist eine syntaktisch ver-
einfachte, semantisch verallgemeinerte und differenzierte Form vom Rus-
sells Theorie 29 • Sie erlaubt die Vnterscheidung von
Kl Existenzbehauptungen und -Prasuppositionen (wie in L3);
K2 geschachtelten einstelligen und simultanen mehrstelligen
Kennzeichnungen;
K3 verschiedenen Kennzeichnungsbereichen (wie bei Russell).
Bevor wir dies zeigen, geben wir die Regeln an. Fiir (lJ-Formeln A = AXl
... x" in denen Xl, ... , xr ein- oder mehrfach frei vorkommen, und B =
B[Xl ... xr], in denen Xl> ... , Xr nicht notwendig frei vorkommen, gilt:
Rl' Ilxl ... Xr (A, B)~ =
= j lB[Ul ... ur]I;, wenn I1!Xl' .. XrAXl ... xrl; = IAul' .. url; = W,
N , wenn I1!Xl' .. XrAXl ... xrl; = F,
o , wenn 11!Xl' .. XrAXl ... xrl; = o.
Vnter den Bedingungen von DSl gilt fUr i = leer, * oder Zahlanfiihrung:
RSl lSi .. (lxlAlXl)' .. (lxpApXp) .. ~ =
lSi •• U1
- -
••• U In
p •• qI'
= IWiAjUjl; = W, flir alle j = I, ... , p,

wenn II!xjWiAjXjl;
N, wenn II!xjWiAjXjl; = F, fiir ein j,
0, wenn II!xjWiAjxjl; fiir kein j F, aber flir ein jOist.
Wir halten einen Spezialfall fest, der am Ende von 4.7 verwendet wurde.
Rz= W 11!xAxl; = W und IAul; = W,
F 11!xAxl; = W und IAul; #: W,
IzxAx = iii; =
N II!xAxl; = F,
o 11!xAxl; = O.
Unter den Bedingungen von DS' gilt:

RS' IS·xAxel; =
ISiiiel;, wenn IzxAxl; = m (im Sinn von LR, dh.

Folge von m Strichen)
= N, wenn IzxAxl; existiert, aber keine LR-Zahl ist,
oder wenn 11!xAxl; = F,
0, wenn 11!xAxl; = O.
In allen diesen Regeln ist die Fallunterscheidung vollstiindig. Nun zum

Punkt KI, den Existenzbehauptungen und -Prasuppositionen. DSz und
RSz beziehen sich auf alle Pradikatsymbole S mit Ausnahme von E!, denn
das Existenzpradikat ist als einziges der hier betrachteten elementaren
Pradikate nicht existenzprasupponierend, sondem existenzbehauptend.
Wenn Nessie nicht existiert, sind die gewohnlichen Aussagen iiber sie -
daB sie lebt, ziirnt, mit sich identisch ist - neutral, aber "Nessie existiert"
ist falsch. Die Existenzpdisupposition der elementaren S #: E! iibertragt
sich auf manche komplexen Pradikate, auf manche nicht, und ob sie sich
iibertragt, ist nicht allgemein entscheidbar30 • Unter der starken Negation
bleibt sie erhalten, unter allen anderen oben definierten I-stelligen Junk-
430 ULRICH BLAU
toren geht sie verloren, insbesondere auch unter dem Wahrheitsjunktor

und der schwachen Negation (vgl. R-, R-" RS, Rill.) Andererseits ist
das Wahrheitspradikat, wie alle SauBer E!, an allen Argumentstellen exi-
stenzprasupponierend:
(I) Das l-steUige Pridikat auf der Tafel trifIt auf den Betrachter
zu.
(1') Wzx (PRix A PxzxQx) zxRx
P: steht auf, Q: Tafel, R: Betrachter
(1) prasupponiert intuitiv die Existenz der Tafel, des Pradikats auf der
Tafel und des Betrachters. Entsprechend prasupponiert (1') aufgrund von
RSz und der spateren Regel RE!
(1 a') E!zxQx A E!zx (PRix A PxzxQx) A E!zxRx
Nun zum Punkt K2, den ein- und mehrsteUigen Kennzeichnungen.
(2) Der Mann, der die Frau gepriigelt hat, war betrunken.
Der Satz ist mindestens 4-deutig, was seine Prasuppositionen anbelangt:
(a) Es gibt (im Kontext) genau eine Frau; femer gibt es genau
einen Mann, der sie gepriigelt hat.
(b) Es gibt (im Kontext) genau einen Mann, der genau eine Frau
gepriigelt hat.
(c) Es gibt (im Kontext) genau eine Frau, die von genau einem
Mann gepriigelt wurde.
(d) Es gibt (im Kontext) genau ein Paar (x,y) mit einem Mann
x und einer Frau y, wobei x y gepriigelt hat.
(a) ist logisch starker als (b), (c) und (d); (d) ist starker als (b) und (c); (b)
und (c) sind logisch unabhangig. Zur Formalisierung von (2) im Sinn von
(a), (b), (c) geniigen die iiblichen l-steUigen Kennzeichnungen, mit I alS
termbildendem Formeloperator:
(2a') Pzx (QIX A RxzyQ2Y)
P: war betrunken, Q1: Mann, Q2: Frau; R: hat gepriigelt
(2b') P,X (Q1X A l!y (Q2Y, Rxy» - oder aquivalent:
P,X (QIX A Rxzy (Q1Y A Rxy»
(2c') Pzx (Q1X A Rxzy (Q2Y A l!x (Q1X, Rxy»)
Die Formalisierung von (2) im Sinn von (d) mit den iiblichen l-stelligen
Kennzeichnungen ist moglich, aber sehr undurchsichtig:
(2d*) pzx (Q1X " Vy (Q2Y, Rxy) " l!y (Q2Y, Vx (Rxy»)
Einfacher, aber logisch aquivalent ist
(2d') IXy (Q1X " Q2Y " Rxy, Px)
Daher haben wir zum Zweck oberflachennaher Formalisierung den mehr-
stelligen Kennzeichner als Operator von Formelpaaren zu Formeln ein-
gefiihrt 31 . Nun zum Punkt K3. Russells Unterscheidung der Kennzeich-
nungsbereiche wird notwendig, sobald die Negation ins Spiel kommt. Bei-
spiel:
(3) Der Nachbar ist mir weder sympathisch noch unsympathisch,
(a) sondem gleichgiiltig (Vagheit)
(b) denn ich kenne ihn nicht (Vagheit-Kategoriefehler)
(c) denn ich habe keinen (prasuppositionsfehler).
In den meisten Kontexten prasupponiert (3) die Existenz des Nachbam,
aber es gibt Ausnahmen wie (c). Betrachten wir die Formalisierung:
(3') -,PzxQx " -, -PIXQX, oder LR-aquivalente Formen:
-, (PIXQX v - PIXQX), - WPzxQx " - FPzxQx, UPIXQX,
NPzxQx
P: ist mir sympathisch, Q: ist mein Nachbar
(Die Indikatoren lassen wir hier wie auch sonst unanalysiert.) In (3') steht
die Kennzeichnung im Bereich der schwachen Negation, die die Existenz-
prasupposition loscht. Daher ist (3') im Kontext (c) semantisch korrekt,
aber in Kontexten mit Existenzprasupposition wie (a) und (b) zu schwach.
Hier hat (3) die logisch starkere Struktur:
(3") IX (Qx, -,Px " -, -Px), oder LR-aquivalente Formen:
zx (Qx, -,(Px v -Px» usw.
Nun steht die schwache Negation im Bereich des Kennzeichnungsopera-
tors und die Existenzprasupposition bleibt erhalten. (3") ist wahr bei ({'
genau dann, wenn Q: eine Einermenge {d} ist und d E P~ also im Neu-
tralbereich von P liegt. Bereichsambiguitaten zwischen Kennzeichnungs-
und anderen Operatoren (Negation, Notwendigkeitsoperator, intentionale
432 ULRICH BLAU
Pradikate) sind seit langem bekannt. Das folgende Beispiel (4), (5) soll
zeigen, daB bei unfundierten Satzen merkwiirdige Bereichsambiguitaten
zwischen Kennzeichnungen und Kennzeichnungen hinzukommen.
Tafel t Tafel t'
Der Satz auf t ist nicht wahr formal: -W*zxPx
cp sei eine entsprechende Interpretation, d.h. P; {- W*ZXPX} Dann

oszilliert der Wert des Satzes auf t' in der bekannten Weise O~, W~ F~
W:, F:, ... Betrachten wir nun die folgende Aussage iiber t (d.h. formal
iiber t'):
(4) Der wahre Satz auf t ist der chinesische Satz auf t.
(4') zx (Px /\ W*x) = IX (Px /\ Qx), P: Satz auf t', Q: chinesisch
Das natiirliche Verstandnis von (4) folgt wohl dem "Breite-vor-Tiefe"-
Prinzip. Schon auf Stufe 0 erkennt man, daB kein chinesischer Satz auf t
steht; (4) ist infolge von Prasuppositionsverletzung neutral, auf allen Stu-
fen, und dasselbe ergibt sich fiir (4') bei entsprechendem cp, wenn P; n
Q; = 0, nach RSz. Um dies zu erreichen, wurde in DSz die simultane
Elimination aller Kennzeichnungen an den Argumentstellen von SS ver-
langt; keine Kennzeichnung wird der anderen iibergeordnet. - Dies ist die
natiirliche Lesart von (4), aberes gibt wohl noch eine andere. Durch starke
Betonung der linken Kennzeichnung kann das "Breite-vor-Tiefe"-Prinzip
auBer Kraft gesetzt und der Blick in die O-unfundierte Tiefe gezogen wer-
den:
(5) Der wahre Satz auf t ist der chinesische Satz auf t.
(5') zx (px /\ W*x, X = zx (px /\ Qx»
Nun ist die linke Prasupposition 1!x(Px /\ W*x) der rechten l!x(Px /\ Qx)
iibergeordnet und muB entschieden werden, bevor die Behauptung Ax (Px
/\ W*x, X = zx (Px /\ Qx» bewertet werden kann. Daher ist (5') bei
entsprechendem cp auf O. Stufe offen und auf allen hOheren Stufen neutral
(auf allen ungeraden Stufen, wei! die untergeordnete Prasupposition ver-
letzt ist, auf allen positiven geraden Stufen, weil die iibergeordnete Pra-
supposition verletzt ist.)32
4.10. Existenz, Denotation und kontingente Objektbezeichnungen
Fur das Existenzpriidikat und beliebige, auch definierte, Objektbezeich-

nungen a gilt:
!
RE!
IE!al~ ~ 1~ 1= la ~ al~ ~ ~ (vollstiindige

Fallunterscheidung)
W, wenn m < n und lal; existiert,

F, wenn m < n und lal; nicht existiert,
IE!al~, wenn m = n,
0, wenn m > n.
Die metasprachliche Aussage
(1) lal; existiert ("Das Denotat von a aufStufe m bei qJ existiert")
wird in LR bei qJ durch
(1') E!iiia
adiiquat ausgedruckt. Dies erkennt der Verifikator auf allen Stufen n >
m. 1m giinstigen Fall erkennt er es schon auf Stufe m, niimlich dann, wenn
a entweder auf Stufe m etwas bezeichnet oder in erkennbarer Weise nichts
bezeichnet (d.h. wenn la = al; = N).
Aber wenn auf Stufe m die Frage der Referenz von a offenbleibt, ist erst
auf Stufe m + 1 erkennbar, daB (1) und (1') falsch sind. Ahnliches ergibt
sich fUr die Denotationsrelation.
RL1 Fur reine Objektbezeichnungen a gilt:
W lal~ = u,
F lal~ existiert, aber ist =1= u, oder
Wa'ul~ =
IE!al~ = F,
o IE!al~ = O.
434 ULRICH BLAU
W, wenn m < n und lal;' = u,

F, wenn m < n und lal;' =1= u oder nieht existiert,
IJ'a' ill;, wenn m = n,
0, wennm > n.
Die metaspraehliehe Aussage
(2) lal;' = u ("a denotiert auf Stufe m bei qJ das Objekt u")
wird in LR bei qJ dureh
(2') Jiiiral ii
adaquat ausgedriiekt. RJ gilt nur fiir reine a, da die Syntax entspreehend

eingesehrankt ist. Ohne diese Einsehrankung ware z.B. der fragwiirdige
Satz allgemein giiItig:
(3) Jedes Objekt hat eine Bezeiehnung.
(3') !\xVyJyx
Dies gilt nur in abzahlbaren Universen, deren au13erspraehliehe Objekte
mit reinen Pradikaten eindeutig gekennzeiehnet werden konnen. Die
naehste Regel gibt an, ob und was die Kennzeiehnungen bezeiehnen.
Rz
= u = {u}
IzxAxl; { } - l(x)AxI; {
existiert nieht ist keine Einermenge.
Daraus fol81 fiir die Objektparameter a und die mit Funktionsparametem

f gebildeten Objektbezeiehnungen:
Ra
=u } = {u},
lal.. { - a': {
Rf
If(at . .. ar)l; {
= U} - I(x)f'at ... arxl;
{ = {u},
UNBESTIMMTHEITEN UNO PARAOOXIEN 435
Ob und was diese Bezeiehnungen denotieren, hangt im allgemeinen von n

und cp ab, bei den Objektparametem nur von cp. Sie entsprechen den sog.
Eigennamen, deren Referenz von der Kontextsprache, aber nicht der Re-
flexionsstufe abhangt (Nero, Nessie, New York) und die vollig flieBend in
echte Kennzeichnungen iibergehen:
(6) M iinchen
Griechenland
die Tiirkei
die Antarktis
der Nordpol
der nordliche Wendekreis
die norddeutsche Tiefebene
die bayerische Landeshauptstadt.
Die Funktionsparameter stehen fUr natiirliehe Funktionsausdriicke, die
bekanntlich fast immer partielle Funktionen ausdriicken. (Ausnahmen:
Identitat, mengentheoretische Funktionen und kiinstlich definierte Kon-
stanzfunktionen auf dem Universum.) Entsprechend ist I(Xl ... xry) f(Xl
... x r) = yl; fUr gegebene n, cp im allgemeinen nur eine partielle Funktion
auf (U",),. Die Funktionskonstanten fUr das Denotat und den Wahrheits-
wert haben die Regeln:
Rden Fiir reine Objektbezeichnungen a und m < n < r gilt:
Idenmra'l; = lal;', oder beide existieren nicht,
Ideniira'l; = Idenra'l; = lal;, oder aIle drei existieren nicht,
Idenrra'l; existiert nieht.
Rval Fiir cp-Satze A und m < n < r gilt:
Ival:rAlI; = IAI;',
Ival;rAlI~ = IvalrAlI~ = IAI~ } { IFn

o$>AE '"
Ival~rAlI~ und Ival rAll; existieren nicht Qj~,
Ival~rAll; existiert nicht.
Bei gegebenem cp driickt den iD auf allen Stufen ~ m die partiell definierte
Extensionsfunktion I I; der reinen Objektbezeichnungen aus und valiD
driickt auf allen Stufen > m die total definierte Extensionsfunktion I I;'
436 ULRICH BLAU
der fP-Satze aus. Die Extensionsfunktion der fP-Pradikate ist nicht aus-
driickbar, da LR keine Mengentheorie enthaIt.
Die vorangehenden Definitionen und Regeln geben einen Eindruck von
der Starke des Systems. Die meisten Regeln wurden nur fUr fP-Grundbe-
zeichnungen ii = a oder r e, formuliert; mit Hilfe von Substitutionsprin-
zipien, auf die ich nicht eingehe, lassen sie sich fUr beliebige Objektbezeich-
nungen a, b veraIIgemeinem. Z.B. gilt fUr die Pradikatparameter:
RPa W <Iall;, ... , larl;> E P: ,
F <Iall;, ..., larl;> E Pi,
N <Iall;, ... , larl;> E P~ oder
IPal' .. arl; = <=>
fiir ein j IElajl; = F (j = 1,..., r),
° fiir ein j IElajl; = ° und

fiir aile j IElajl; #: F.
Und fiir die Identitat gilt:

R=a
W lal; == Ibl;,
F lal; 'IF Ibl;, aber beide existieren,
N <=> IElal; = F oder IElb!; = F,
° IElal; = 0, IElb!; #: F oder

IElbl; = 0, lEla!; #: F.
5. DIE LOSUNG DER PARADOXIEN
Die eingangs skizzierten Losungen der semantischen Paradoxien lassen

sich in LR beweisen. Wir geben, ohne Beweis, einige Schemata objekt-
sprachlicher Theoreme zu den Paradoxien der Wahrheit, Erfiillung und
Bezeichnung an.
5.1. Die Wahrheitsparadoxie (Eubulides, Lukasiewicz)
Der reflektierte Liigner hat die Losung:

(WI) Wenn der einzige Satz auf der Tafellautet:
"Der Satz auf der Tafel ist nicht wahr"
und wenn "wahr" reflektiert wird, dann ist der Satz auf der
Tafel auf Stufe 0 offen, auf allen ungeraden Stufen wahr, auf
allen positiven geraden Stufen falsch, also paradox.
Die Aussage (WI) iiber aile endlichen Stufen gilt erst aufStufe m im nach-
sten System LR." Abschnitt 6. Aber in LR gilt fiir jedes einzelne n ~ 0:
II-~n+2 zxPx = r _ W*zxPx' -+ 00 ZxPx A W2n+l zxPx A F2n+2 zxPx A
!PZxPX
P: Satz auf der Tafel
Ahnlich paradox sind die reflektierten Verwandten des Liigners, die sich
als falsch, unbestimmt, unfundiert und offen bezeichnen:
(W2) 11-~2n+2 zxPx = rF*zxPx' -+ 0° zxPx A F2n+l ZXPx A
W2n +2 zxPx A !PZXPx

Und fUr S* = U*, U*, 0* gilt:
(W3) II-~n+2 zxPx = rs*ZxPx' -+ 00 zxPx A W 1zxPx A Fn+2 zxPx
A !PZXPx
Hingegen ist der Satz, der selbst paradox sein will, auf keiner Stufe wahr,
also nicht paradox:
(W4) II-~n+l zxPx = r!p*zxPx' -+ 0° zxPx A Fn+lZxPx A -!PzxPx
(Ebenso mit !P* statt !P.) Die unreBektierten AngehOrigen der selbstrefe-
rentiellen Familie sind allesamt nicht paradox. 1m Gegensatz zu (WI) gilt:
(WS) Wenn der einzige Satz auf der Tafellautet:
"Der Satz auf der Tafel ist nicht wahr"
und wenn "wahr" im absoluten, unreflektierten Sinn verstan-
den wird, dann ist der Satz auf der Tafel auf allen Stufen offen,
also nicht paradox.
II- ~n+ 1 zxPx = r - WzxPx' -+ ODZxPx A -!pD zxPx

Dasselbe gilt fiir aile unreBektierten semantischen Pradikate S und ihre
Negationen:
(W6) II-~n+l zxPx = r[_]szxPx' -+ onZXPx A -!pDZxPX
438 ULRICH BLAU
Man beachte, daB lP ii in (W5) und (W6) nicht durch IP ersetzt werden
kann: Der Verifikator erkennt, daB unreflektierte selbstreferentielle Satze
[ - ]SIXPx auf allen vorangehenden Stufen n offen und nicht paradox sind.
Aber daB sie auf seiner Stufe offen und nicht paradox sind, erkennt er
nicht - wiirde er es erkennen, so waren sie nicht offen und paradox.
Neben den Paradoxien der direkten Selbstreferenz gibt es uniibersehbar
viele aufgrund wechselseitiger Referenz, also indirekter Selbstreferenz. Ein
Beispiel ohne historische Gewahr:
(W7) Wenn Hegel iiber Schlegel nur sagt:
"Alles, was Schlegel iiber mich sagt, ist falsch",
und Schlegel iiber Hegel nur sagt:
"Alles, was Hegel iiber mich sagt, ist unfundiert",
dann ist beides paradox, Schlegel hat auf I. Stufe recht, Hegel
irrt sich auf den Stufen I und 2 und behiilt ab 3. Stufe endgiiltig
recht.
Wir unterstellen diesen Herren die reflektierte Verwendung von "falsch"
und "unfundiert". Dann gilt:
II-~n+3'xPabx = [/\X (pbax, F*x)] A ,xPbax = V\x (pabx,
ll.J*x)] -+ IPlxPabx A IPlxPbax A W1,XPbax A PzxPabx A
F2zxPabx A Wn+3lxPabx
Pxyz: x sagt iiber y z, a: Hegel, b: Schlegel
5.2. Die Erfiillungsparadoxie (Grelling)
1st "heterologisch" heterologisch oder autologisch? Beide Antworten sind

paradox, wenn man "heterologisch" und "autologisch" reflektiert.
(EI) 11-~2n+2 OOHET* r(x)HET*x' A F2n+l HET* r(x)HET*x' A
W2n+2HET* r(x)HET*x' A IP HET* r(x)HET*x'

Die positive Antwort oszilliert mit umgekehrtem Wahrheitsverlauf:
(E2) 11-~2n+2 OOAUT* r(x)HET*' A W2n+l AUT* r(x)HET*x' A
p2n+2 AUT* r(x)HET*x' A IP AUT* r(x)HET*x'

1st "autologisch" heterologisch oder autologisch? Die Frage hat auf allen
hoheren Stufen die eindeutige Antwort: "heterologisch".
(E3) If-~n+l OOHET* r(x)AUT*x' /\ W n+ 1 HET* r(x)AUT*x' /\
-\P HET* r(x)AUT*x'
(E4) If-~n+l OOAUT* r(x)AUT*x' /\ p+l AUT* r(x)AUT*x' /\
-\P AUT* r(x)AUT*x'
Die entsprechenden unreflektierten Aussagen sind wieder auf allen Stufen
offen, also nicht paradox:
(E5)
If-~n+l OiiHET r(x)HETx' /\ OiiAUT r(x)HETx' /\ onHET
r(x)AUTx' /\ OiiAUT r(x)AUTx'
5.3. Die Bezeichnungsparadoxie (Konig, Berry, Finsler)
Zur Analyse der unfundierten Kennzeichnung k, Abschnitt 1.3, erweitern

wir LR urn ein kleines Stiick Zahlentheorie. LR< hat zusatzlich die Kon-
stante "<" und die Endregeln:
E< 1 [-W - ] m < ii, wenn m <: n
E<2 -[W] m < ii, wenn m 2: n
E< 3 - W[ - ] ii < V, wenn ii oder V keine Zahlanfiihrung ist.
Dann ist definierbar:

DN Natiirliche Zahl
Nt: = t = 0 v WO < t
DJ.l J.l-Operator
J.lxAx:= zy (Ny /\ Ay /\ I\x (Nx /\ Ax, y < x v y = x)),
wobei y ¥ x.
Es folgt:
R<
Iii < vi =
1WI 1F <=>
u und v sind Zahlen (Strichfolgen) und u
U und v sind Zahlen und u ~ v,
<: v,
N u oder v ist keine Zahl.

R~
INul =
{W}
~
{ u ist eine Zahl (Strichfolge),
F u ist keine Zahl.
440 ULRICH BLAU
Rp. {= u} { u als kleinste Zahl,

lp.xAxl; existiert nieht <=> l(x)AxI; enthiUt keine Zahl.
IN LR< gilt:
(B) Wenn als einzige (vollstiindige) Bezeiehnung auf der Tafel

steht:
"die kleinste natiirliehe Zahl, die dureh keine Bezeiehnung
auf dieser Tafel bezeiehnet wird"
und wenn "bezeiehnet" reflektiert wird, dann ist die Bezeiehnung auf der
Tafel nieht-referentiell auf Stufe 0, sie bezeiehnet 0 auf allen ungeraden
Stufen und 1 auf allen positiven geraden.
11-0!:2n+2 ZxPx = [jty- Vx (Px, L1*xy)] -+ -RFozxPx 1\ L12n+l (zxPx)O 1\
L12n+2 (zxPx)1
Mit ..1 statt ..1* ist die Bezeiehnung auf allen Stufen nieht-referentiell, wie
der Verifikator riiekbliekend erkennt. Ahnlieh lasen sieh alle Paradoxien
der Wahrheit, Erfiillung und Bezeiehnung auf. Aber es gibt sehrankenlose
Liigner der Art:
(I) Dieser Satz ist auf allen Stufen nieht wahr.
(1') Ax (l\Ix, - W"zxPx), Px: x ist Satz (I')
Ihre Analyse erfordert transfinite Reflexionsstufen. Zuvor soll etwas zu
den Mengenparadoxien gesagt werden.
5.4. Die Mengenparadoxien 33
Die Intuition, die im folgenden verteidigt werden soll, ist so alt wie die
Mengentheorie; man kannte sie als offenen Mengenrealismus bezeiehnen.
Ieh mOchte zeigen, daB bei dieser Auffassung
1. alle Mengenparadoxien versehwinden,
2. die Standardsysteme Z, ZF, NBG, MKM und unabsehbar vie-
le Erweiterungen wahrseheinlieh korrekt sind.
Betraehten wir ZF (von Zermelo-Fraenkel, i.f. stets mit Fundierung und
Auswahl). Angenommen, es gibt ein Universum V, das durch ZF korrekt

beschrieben wird. Dann folgt, daB alle Objekte, d.h. Mengen, von V in
einer Hierarchie von Stufenmengen Va liegen. Diese Hierarchie wird aus
der leeren Menge durch Potenz- und Vereinigungsmengenbildung in einem
transfiniten, aber fundierten ProzeB erzeugt, der so weit reicht wie die Or-
dinalzahlen IX E V:
Vo := 0
V a + 1 := PV a
Va: = UV p, fUr Limeszahlen IX.
p<a
Fur aller Ordinalzahlen IX und alle Mengen x gilt in ZF:34

x E Va +-+ VP(P < IX /\ (x E Vp V X c V p»
In diesem Sinn ist V eine kumulative Hierarchie: Jede Stufenmenge enthalt
die Elemente und Teilmengen aller vorangehenden. (Vgl. Fig. 2.)
Zwei Fragen bleiben offen in ZF:(a) die Breite der transfiniten Stufenmen-
gen Va>(I)' denn das Potenzmengenaxiom verrat nicht, was aile Teilmengen
LJ +'
W
~~-----\V2 = {0. {0}}

;"\-.--/--- \V1 = {(1) }
0= rtJ = \VO
Fig. 2.
442 ULRICH BLAU
sind, und (b) die Hohe von V, d.h. die Anzahl der Va.. Formal, von ZF
her gesehen, ist V unterbestimmt, falls iiberhaupt vorhanden.
Intuitiv, vom Standpunkt des Mengenrealisten, nennen wir ihn R, sieht
die Sache anders aus. Sein V ist eine vollkommen bestimmte abstrakte
Realitat: die Gesamtheit aller formalen Objekte und Strukturen iiber-
haupt; ebenso vollstandig und eindeutig wie z.B. die winzige Teilstruktur
N.
Die in ZF offene Frage (a) kann R nicht beantworten, vielleicht nur aus
epistemischen Grunden, weil er die Losung des Kontinuumproblems nicht
kennt, vielleicht aber auch aus tieferen sprachlichen Grunden, weil er zwar
genau weiB, aber nicht genau sagen kann, was er mit allen Teilmengen
meint. Was er sagen kann, ist dies: Jede Menge x existiert in V als reales
Objekt, unabhangig von jeder Definition oder Konstruktion. Ebenso real
istjede Teilmenge y von x, gleichgiiltig, ob in irgendeinem Sinn konstruier-
bar; Px enthalt aile diese y.
Die in ZF offene Frage (b) hat fUr R keinen Sinn. Woran sollte er die
Hohe von V messen? V enthalt selbst aIle Ordinalzahlen und ist uner-
meBlich. (So scheint es immer. Ahnlich unermeBlich erschien uns einmal
die Zahlenreihe N, als wir sie zum ersten Mal verstanden.)
Die Elementrelation ist fUr Reine eindeutige scharfe Relation auf V,
daher muB jedes mengentheoretische Axiom eindeutig wahr oder falsch
sein. Wie steht R zu ZF? Die Axiome der Extensionalitat, Unendlichkeit,
Fundierung, Auswahl sind ihm absolut evident, ebenso die der Vereini-
gungs- und Potenzmenge. (1st x E Va., so ist Ux E Va. und Px E Va.+1.)
Nicht evident ist nur das Ersetzungsschema, zu dessen Rechtfertigung
spater mehr gesagt wird. Sollte in ZF tatsachlich ein Widerspruch auftre-
ten, so wiiBte R a priori, daB der Fehler im Ersetzungsschema stecken muj3;
er wiirde sich mit unerschiittertem Weltbild V auf das schwachere System
Z von Zermelo, mit Paarmenge und Aussonderungsschema zuriickziehen.
Z ist nun wirklich evident. (1st x EVa., Y E V p, so ist {x,y} E V max(..,p) + 1
und {z E x I Az} E V ... ) Ein Widerspruch in Z ist fiir R so undenkbar wie
ein Widerspruch in den Peanoaxiomen 1. Stufe. Leider ist Z sehr schwach,
es verlangt nur den verschwindend kleinen Anfangsabschnitt V w+w von
V, und R ware genotigt, von Fall zu Fall starkere Unendlichkeitsaxiome
zu fordern. Solange es reine GroBenforderungen sind, was oft schwer er-
kennbar ist, sind sie harmlos. R's V gewahrt lachelnd jede
GroBenforderung.
So weit so gut. Alle formalen Objekte sind in V, auBerhalb gibt es nichts.

Und nun kommt das Problem, der Kern aller Mengenparadoxien: R ver-
miftt V. Er vermiBt die auBersten Klassen, d.h. die vertikal unbeschrankten
Gesamtheiten, die in keiner Stufenmenge V IX Element sind, da sie Elemente
von zu hohem Rang enthalten (z.B. IX). Er vermiBt u.a. die Klasse R alIer
Mengen x ¢ x, die Klasse FND aller fundierten Mengen, die Klassen ON
und KN aller Ordinal- und Kardinalzahlen; vielleicht vermiBt er auch die
Kardinalzahl von V. Diese Klassen mu,P es geben, sonst batten die ge-
nannten Pradikate keine Extensionen und waren bedeutungslos. Aber aus
den evidenten Axiomen von Z folgt, daB diese Klassen keine Objekte in
V sein konnen. In NBG bzw. MKM kommen zu den Mengen die (pra-
dikativ bzw. impradikativ) definierbaren Klassen von Mengen als Objekte
2. Sorte hinzu, aber nun fehlen alle Klassen von Klassen! In allen Stan-
dardsystemen bleibt die Liicke, daher werden sie oft als ad hoc empfunden.
Manche Nicht-Standardtheoretiker empfinden die Liicke so schmerzlich,
daB sie die absolute, unfundierte, sich selbst enthaltende Allmenge zum
Objekt machen, urn das Reich der Formen von innen abzuschlieBen - stets
zum Preis von Prinzipien, die IUr R evident sind, u.a. Fundierung. (Die
bekanntesten Beispiele sind Frege und Quine; der eine endet im Wider-
spruch, der andere im Konventionalismus.) Mir scheint, fiir den reflektie-
renden Realisten liegt die Losung nicht im AbschluB, sondern im Erweite-
rungsproze'p der Standardtheorie:
(1) So : = ZF. Und fUr jedes IX > 0 sei
SIX := ZF + 'Es gibt ein gemeinsames Modell fUr alIe SII<IX'.
Dieser ErweiterungsprozeB ist IUr jeden Mengenrealisten zwangslaufig.
Wenn er die SII<IX akzeptiert, so muB er ihre semantische Konsistenz ak-
zeptieren, d.h. SIX. So gelangt er zu starkeren und starkeren Systemen. Um
diese Hierarchie zu prazisieren, werden wir (a) die formale Sprache, (b)
den Modellbegriff, (c) die Indizes IX prazisieren.
(a) Wir beschranken uns auf die Sprache L der Standardsysteme mit
den Symbolen: I , 1\,1\,E, Variablen und Klammern. Die Definitionen
fiir = und die iiblichen Konstanten sind als informelle Abkiirzungen zu
verstehen. Aus den reinen Ausdriicken von L entstehen die (fiktiven) Aus-
driicke von Lv daaurch, daB 0 oder mehr freie Variablen durch (fiktive)
Namen x der Mengen x E V ersetzt werden. Ausdriicke von Lv' die hOch-
444 ULRICH BLAU
stens Name fUr die Elemente von x enthalten, heiI3en x-Ausdrucke.

(b) Zum Modellbegriff. Wenn man fiir die Modelle niehts weiter fordert,
so reduziert sieh die semantische Konsistenzforderung in (1) auf die syn-
taktische Konsistenzforderung, und das ist viel weniger als R eigentlieh
meint. Ganz genau wird er natiirlieh nie sagen konnen, was er meint; so-
lange er konsistente Forderungen stellt, wird es nieht-intendierte, sogar
abzahlbare Lowenheim-Skolem-Modelle geben, aber das braueht ihn nieht
zu hindem, gewisse einleuehtende Forderungen zu stellen. Eine natiirliehe
Minimalforderung ist die, daB alle Modelle Stufenmengen VfJ sein sollen;
mehr solI zunaehst nieht verlangt werden.
(e) Zu den Indizes der S". Um S" angeben zu konnen, mussen wir IX
bezeiehnen konnen. Eine Menge x heiBe bezeichenbar, wenn
(2*) fiir eine reine L-Bezeiehnung a gilt: a = x.
Diese vorlaufige Definition wird·spater prazisiert; zunaehst einige Beispie-
Ie. Kleine Ordinalzahlen sind in natiirlieher Reihenfolge liiekenlos bezei-
ehenbar:
0,1 ... co, ro+ 1 ... co·2 .,. row ... 80, 80+ I ... 80.2 ... 85 ...
Irgendwo weiter oben werden die Bezeiehnungen weniger informativ. Z.B.
gibt es fiir die konstruktiven Ordinalzahlen35 noeh liiekenlose Bezeieh-
nungssysteme l:, aber fiir Bezeiehnungen a, bel: ist im allgemeinen nieht
entseheidbar, ob a eine groBere Ordinalzahl als b bezeiehnet. Irgendwo
weiter oben werden die Bezeiehnungen liiekenhaft, da es nur abzablbar
unendlieh viele gibt. ~1 ist bezeiehenbar, fast alle kleineren nieht. Wo liegt
die kleinste nieht bezeiehenbare Ordinalzahl? Und wie konnten wir sie soe-
ben bezeiehnen? Das ist die .mengentheoretische Bezeiehnungsparadoxie;
wir stellen sie etwas zuriiek.
Klar ist jedenfalls, daB formale Systeme S" der Spraehe L nur fiir be-
zeiehenbare IX existieren. Aber man kann eine Hierarehie von "fiktiven"
Systemen S" der Spraehe Lv' fiir alle IX E V, definieren. Das erfordert einige
Vorbereitungen.
g sei eine GOdelisierungsfunktion, die den Ausdriieken e von Lv Men-
gen x E V zuordnet:
1. Fiir jedes der abzahlbar unendlieh vielen Symbole s von L sei
g(s) eine positive natiirliehe Zahl.
2. Fiir jeden Mengennamen X sei g(x) : = <O,x).

3. Fiir jeden komplexen Ausdruck e = el ... ea , bestehend aus
L-Symbolen und Mengennamen ej, sei g(e) : = <g(el), ... ,
g(ea».
In der Sprache L konnen wir, bei Voraussetzung eines Standardsystems,
iiber ihre (gOdelisierte) Syntax reden, aber nicht iiber die Bedeutung ihrer
Ausdriicke bei Interpretation iiber 'V. (Sonst wiirde der Liigner mitreden.)
Aber wir konnen in dieser Sprache iiber die Bedeutung ihrer Ausdriicke
bei Interpretation iiber einer beliebigen Menge x E 'V reden. Genauer ge-
sagt sind unter Voraussetzung von ZF Funktionen EXT und VAL auf 'V
so definierbar, daB fUr aIle x E 'V, alle r-stelligen x-Pradikate A' und aIle
x-Satze A:
EXT(g(A'),x) = {<y, ... Yr) E xr I A'Yl ... Yr} (Extension von A' in x)
VAL(g(A),x) =
{ 0I} <=> A ist wahr in x (x ist Modell fiir A)
A ist falsch in x (x ist kein Modell fUr A)
Dann ist die Modellrelation definierbar:

x F y := I\z (z E y -+ VAL(z,x) = 1)
1st Meine Menge von x-Siitzen und y = {g(A) I A EM}, so ergibt sich:
x F y <=> x ist Modell fUr M
Wir benotigen ferner die Tatsache, daB zu jeder Formel Ay eine Funktion
G auf 'V so definiert werden kann, daB fiir alle x E 'V:
G(x) = g(AX)
Wir zeigen dies am Beispiel der Formell\z -,z E y. Diese Formel sieht,
polnisch geschrieben, so aus: I\z -, E Z y. Die g-Werte der Symbole 1\,
z, -', E seien Ph P2, P3, P4. Dann sei
G(x) : = <PhP2,P3,P4,P2,<O,X» = g(l\z -,z E x)
Fiir die gewiinschte Hierarchie (1) benotigen wir die "Erweiterungsfor-

mel":
VP ('V p F UWertbereich y)
446 ULRICH BLAU
AoY entstehe aus ihr durch Beseitigung aller definierten Symbole. Dann
gibt es eine Funktion GO, so daB:
Go(x) = g(Aox).
Ferner sei
K(x) : = {g(A) I A ist Axiom von ZF} u {Go(x)}
Nun definieren wir eine Hierarchie von Mengensystemen MIZ durch Re-
kursion iiber ON:
MIZ : = K(M I a) = {g(A) I A ist Axiom von ZF} u {g(Ao(M
I ~)}
Jedes MIZ enthiilt in godelisierter Form die ZF-Axiome und den weiteren
V
Lv-Satz Ao(M I Ii), also n. Def. P (V p 1= U Wertbereich (M t Ii)) oder
einfacher: Vp (Vp 1= Y<IZ
u_ MY). Dann sei
SIZ:= ZF + VP (Vp 1= Y<IZ

U_ MY)
SIZ ist ein "fiktives" System, dessen Zusatzaxiom den Mengennamen Ii ent-
halt. Aber fiir jede L-Bezeichnung a, fiir die wir beweisen konnen oder
behaupten wollen, daB ONa, erhalten wir einformales System der Sprache
L:
IHI sei die Hierarchie der SIZ. So ist nichts anderes als ZF, denn das Zusatz-
axiom ist trivial. Urn die Systeme auf der intendierten semantischen Ebene
zu erliiutern, flihren wir einige Begriffe ein. (Sie haben nur flir R einen Sinn
und sind nur dann korrekt definiert, wenn ZF korrekt ist.) Das intendierte
Modell, oder kurz, Universum flir SIZ sei die kleinste Stufenmenge der ku-
mulativen Mengenhierarchie, die SIZ erflillt. Dann ergibt sich eine kumu-
lative Hierarchie von Universen:
VO := 0
v<Z+ 1 : =das Universum flir SIZ
VIZ := U
P<IZ
VP, flir Limeszahlen a.
Jedes S.. behauptet bei intendierter Interpretation formal die Existenz von
V" und setzt informell sein eigenes Universum V.. + 1 voraus, dessen Exi-
stenz in den SII>" behauptet wird. k(a) sei die Kardinalzahl von V" und
mea) sei der untere Index, d.h. V" = Vm(IZ). Betraehten wird nun den An-
fang von IHI. SO hat das Universum V1. Von innen ist es ein unermeBlieh
hoher offener Triehter; es scheint das Absolute zu sein. Aber es ist nieht
das Absolute, sondern nur eine Etappe. Von auBen, im naehsten Univer-
sum V2, erkennen wir, daB V1 einen von innen unsiehtbaren AbschluB
hat, namIieh V1. In V2 ist V1 eine Stufenmenge V m(l). Und nun sehen wir
aueh die anderen in V1 vermiBten Objekte: Die Klassen R und FND in
V1 sind identisch mit V1, die Klasse ONfn Vt, ist die Menge m(1), die
Klasse KN in V1 ist die Menge k(1), beide sind Elemente von V2, und V2
ist eine Stufenmenge V m(2) im naehsten Universum V3. Jede Klasse von
innen ist eine Menge von aujJen. (Vgl. Fig. 3.)
m(3} \V 3
"i I 3
V =\V m (3)
m(2} \V 2
~ j 2
V =Wm (2)
V'
V3 m(1}~ j ,
\V =W m (1)
Fig. 3.
448 ULRICH BLAU
AIle Paradoxien verschwinden im Riickblick und wir verstehen, in wel-

chem Sinn die Frage der Hohe von V einen Sinn, und keinen Sinn hat.
Wir konnen buchstablich nichts iiber unser V sagen, ohne es in ein ver-
lassenes V .. zu verwandeln. V .. hat eine Hohe IX, aber unser V ist nie Objekt;
es ist der ii.uBerste Rahmen, so unfaBlich und im nachhinein so faBlich wie
die unreflektierte Reflexionsstufe. Dasselbe gilt fiir aIle ii.uBersten Klassen:
Sobald wir iiber sie prii.dizieren oder quantifizieren, Machen wir sie zu
Objekten im niichsten Universum. Ein Beispiel ist die Hierarchie IHI.
Ebenso wie V hat IHI einen von innen unsichtbaren Index: In V = V ..: ist
IHI = IHI .. : = {SII I P < IX}. IHI ist nie die ganze Hierarchie, ebensowenig wie
V je das Absolute ist. Ahnliches gilt fiir die obige Definition (2*) der be-
zeichenbaren Menge x. (2*) bezieht sich von innen auf unser V = V .. und
besagt, von auBen gesehen:
(2) Fiir eine reine L-Bezeichnung a gilt in V ..: a = i.
Die "innere" Definition (2*) ist in den Standardsystemen der Sprache L
nicht ausdriickbar - hier stoBen wir an die Grenze zwischen der klassischen
und der Reflexionslogik - aber die "iiuBere" Definition (2) ist ausdriickbar,
wenn wir ein groBere V II voraussetzen, das V .. als Element hat:
(2') Vy (y E V co /\ EXT(y, V ..) = {x}}
Betrachten wir nun die mengentheoretische Bezeichnungsparadoxie:
(3) die kleinste nicht bezeichenbare Ordinalzahl.
1m System LRe des niichsten Abschnitts, das seine Bezeichnungsrelation
Lt* enthii.lt, ist (3) formalisierbar:
(3') JtX(ONx /\ -, Vy Lt*yx)
und die Paradoxie lost sich, wie im letzten Abschnitt, durch Reflexions-
stufenunterscheidung, worauf wir nicht eingehen werden.
In den Standardsystemen der Sprache List die "innere" Bezeichnungs-
relation nicht definierbar, daher tritt die Paradoxie formal nicht auf. Aber
eine Art informelle Paradoxie entsteht fiir R, wenn er sein V fiir das Ab-
solute halt: R weiB, nach allen bisherigen Erfahrungen, daB jedes iiber-
haupt bezeichenbare mathematische Objekt in L bezeichenbarist. Ande-
rerseits weiB er, daB die Kennzeichnung (3) ein bestimmtes lXo E V bezeich-
net, das er nicht in L bezeichnen kann! Die Losung ist einfach: lXo hangt
von R's V = V IX abo ao ist nicht bezeichenbar in V IX' aber injedem groBeren
V p, in dem a durch ein a bezeichenbar ist:
(3") Jlx(ONx /\ IVy (y E Vw /\ EXT(y, Va) = {x}))
In V IX bezeichnet (3") entweder gar nichts oder aber nicht ao, je nachdem,
wie der Jl-Operator definiert ist, woraufwir nicht eingehen. (Eine Moglich-
keit ware JlxAx: = n
{x I ONx /\ Ax}.)
So viel zu den Mengenparadoxien. AbschlieBend einige Bemerkungen
zu den formalen Systemen sa, die aus den fiktiven SIX durch Ersetzung von
a durch a entstehen. Wenn sa konsistent ist, so ist sa + 1 nach Gode1s 2.
UnvollsHindigkeitstheorem starker. Z.B. gilt:
sa+l I- ON JlX (Vx F Ma)
In sa+l ist fiir mehr b beweisbar, daB ONb; in diesen Sb ist fiir mehr e
beweisbar, daB ONe, uSW. Die Gesamtheit der so entstehenden sa erweitert
sich fortlaufend. Aber es gibt noch starkere formale Systeme. Betrachten
wir das Axiom:
Mo:= l\aVP (Vp FUMY)

Y<IX
Mo ist intuitiv korrekt, falls ZF korrekt ist. Und ZF + Mo ist starker als
jedes obigen sa, da
M o, ONa I- VP (Vp FUMY),

y<a
also
ZF + M o, ONa I- sa.
Aber es gibt starkere Systeme als ZF + M o, denn auf dieser Basis konnen
wir eine neue Hierarchie 1HI0 definieren. Analog zu K, MIX, SIX sei
Ko(x):= {g(A) I A ist Axiom von ZF + Mo} U {Go(x)}

M~ := Ko(Mo I a)
S~ := ZF + M0 + VP(Vp F U Mb
y<1X
).
lHIo sei die Hierarchie der S~. Durch Ersetzung von a durch a entstehen
formale Systeme S~. Aber ZF + Ml ist noch starker, wobei
M 1: = U Mb
l\aVP (Vp Fy<1X ).
450 ULRICH BLAU
Auf der Basis ZF + M 1 konnen wir eine neue Hierarchie !HI1 und ein
starkeres Unendlichkeitsaxiom M2 definieren. So entstehen immer starkere
formale Systeme ZF + Ma. Aber ZF + M1.o ist noch starker, wobei
Und nun geht derselbe Mechanismus, nennen wir ihn M*, von yom los.
Ich weiB nicht, wie weit er reicht. Z.B. weiB ich nicht, ob auf diesem Wege
das Unerreichbarkeitsaxiom
I: = Es gibt eine unerreichbare Kardinalzahl
je entschieden werden kann. Wenn nicht, so miiBte R schlieBen, daB uner-
reichbare Kardinalzahlen entweder nicht existieren (was sehr merkwiirdig
ware) oder unvorstellbar groB sind (was sehr wahrscheinlich ware) und er
wiirde versuchen, die Modelle V p in !HI durch intuitiv plausible Forderun-
gen zu verstarken. In der Tendenz der Standardsysteme liegt die Forde-
rung, daB V p die Bilder aller Elemente unter allen (auBen existierenden)
Funktionen f enthaIt:
(4) I\x, f (x E VP /\ (f: V p - - V p) -+ f"x E V p)
Bei Hinzunabme dieser Forderung ist Sl aquivalent mit ZF + I; V1 ist
V;., fiir die kleinste unerreichbare Kardinalzahl A, und V2 enthaIt das
kleinste natiirliche Modell (V;., V H1) der Standardsysteme NBG und
MKM, wobei die Mengenvariablen iiber V;. und die Klassenvariablen iiber
V H 1 interpretiert werden. 36
Bei stii.rkeren Unendlichkeitsaxiomen als I verlaBt uns rasch die Intui-
tion; wir erkennen nicht ohne weiteres, ob es reine GroBenforderungen
sind. Vielleicht sind manche der heute umstrittenen Unendlichkeitsaxiome
- vielleicht sogar das Kontinuumproblem - entscheidbar in Systemen, die
der Erweiterungsmechanismus M*, evtl. mit Zusatzforderungen der Art
(4), erzeugt. Solche Entscheidungen waren fiir R, bei glaubwiirdigen Zu-
satzforderungen, ebenso glaubwiirdig wie das Basissystem ZF. Die Achil-
lesferse von ZF ist der Ersetzungsschema. 1st es korrekt? AIle bisherigen
Erfahrungen sprechen daflif. Besonders erwahnt sei ein Resultat von Scott
[32]. Unter Voraussetzung von ZF + Fundierung ist das Ersetzungs-
schema aquivalent mit dem Reflexionsschema:
(5) I\xVy (x c Y /\ I\Xl ... Xr (Xl E Y /\ ••• /\ Xr E Y ~

(AXl ... Xr - AYXl ... Xr)))
Dabei ist AYXl ... Xr die Relativierung von AXl ... Xr auf y; sie entsteht aus
Axl ... Xr durch Beseitigung aller definierten Symbole und anschlieBende
Beschrankung aller Quantoren auf y.
(5) besagt, daB es im Universum zujeder Formel A beliebig groBe Men-
gen y als "kleine Universen" gibt, in denen sich die A-Beziehungen ihrer
Elemente als AY-Beziehungen spiegeln. Kurz gesagt, das Universum weist
keine sprachlich faBbaren Singularitaten gegeniiber hinreichend groBen
Mengen auf. Das erscheint plausibel, denn das Universum ist aus demsel-
ben Stoff wie seine Mengen; es wird selbst zu diesem Stoff im nachsten
Universum. Warum sollte man die Schnitte im Absoluten nicht so ziehen
konnen, daB keine Singularitaten entstehen. Nach all dem ist ZF eine sehr
natiirliche, sehr wahrscheinlich korrekte Basis, aber die intuitive Losung
der Mengenparadoxien ist nicht an ZF sondern nur an Z gebunden. Sollte
man je einen Widerspruch in ZF entdecken, so wird man eine bescheide-
nere Basis fUr die groBe Reise ins Absolute finden.
Cantor, fUr den begrifHiche Mittel immer nur Mittel waren, hat die Un-
begreiflichkeit des Absoluten verstanden 37 • Frege, allzusehr auf begriff-
liche Mittel bedacht, hat das nicht verstanden 38 • Aber Zermelo hat es
verstanden. Spatestens 1908 waren die Mengenparadoxien endgiiltig ge-
lost, die Theorie eroberte die Mathematik und trotzdem wird noch heute
von ihren Paradoxien geredet. Die Technik ist dem Verstandnis davon-
gelaufen.
6. TRANSFINITE REFLEXIONSSTUFEN
Zuriick zu Abschnitt 5.3 und dem schrankenlosen Liigner:

(1) Dieser Satz ist auf allen Stufen nicht wahr.
(1') I\x (Nx, - WXIXPx), Px: x ist Satz (1')
1m Sinn von LR < verstanden beschrankt er sich auf aIle endlichen Stufen
und hat recht; er ist auf allen endlichen Stufen nicht wahr, sondern offen.
Daher gilt in LR < fiir jedes n:
(W8) II-~n+l IXPX = 'I\x (Nx, - wx IXPX), ~ OiilXPX /\ -WiiIXPX
452 ULRICH BLAU
Ebenso gilt fUr S = U, IIJ, 0:

(W9) Ir<!:n+l IxPx = Lt\x (Nx, sx IXPX)] -+ sn IXPX
Aile diese Liigner haben recht, aber nicht in LR<, sondero von auBen, auf
allen transfiniten Stufen im nachsten System LRe. Dies entsteht aus LR
durch Hinzunahme der Konstante "e" rlir die Elementrelation der reinen
Mengen. Eine LRe-Interpretation sei eine LR-Interpretation iiber dem
Standarduniversum VI = V m(l) von ZF, das wir in diesem Abschnitt als
existent (d.h. als Menge) voraussetzen.
AIle LRe-Interpretationen haben also dasselbe reine Mengenuniversum.
Damit geben wir, als iiberzeugte Mengenrealisten, die iibrigen Objekte auf
und identijizieren sie mit Mengen. (Das geschieht nur der Einfachheit hal-
ber; wer will, kann bei gewissen Modifikationen von ZF Nicht-Mengen
als Urelemente behalten.) AIle reinen Ausdriicke von LRe, einschlieBlich
ihrer Anfiihrungen, werden mit bestimmten n-Tupeln e V w identifiziert;
den iibrigen v e Vt, die keine reinen Ausdriicke sind, werden Namen v
: = (v, VI> zugeordnet. (Damit ist wie in LR gesichert, daB im Universum
VI kein Ausdruck vorkommt, der Namen enthaIt.) Jedes v e VI hat ent-
weder seine Anfiihrung oder seinen Namen als Grundbezeichnung v. Aus
den reinen Ausdriicken von LRe entstehen die qJ-Ausdriicke durch Erset-
zung von 0 oder mehr freien Variablen durch Namen. Nun verallgemei-
nero wir die Semantik. W* bedeute fiir den Verifikator V
(a) auf einer Nachfolgerstufe at: + I wie bisher: wahr auf der reflektier-
ten, d.h. der unmittelbar vorangehenden Stufe at:;
(b) auf einer Limesstufe at:: wahr auf unendlich vielen reflektierten, d.h.
unmittelbar vorangehenden Stufen P < at:.
Eine ~- Verijikation fiir qJ-Sat{C auf Stufe at: e VI ist ganz analog zur
;-Verifikation, Abschnitt 3, definiert, wobei die Regeln N7, N8, E6-EIO
wie folgt verallgemeinert werden und 2 Endregeln fiir e hinzukommen:
N7' [-] W*A
[ - ] W A, falls at: = 0
bzw. [-] Vx (ONx /\ x < ii, 1\Y(x ~ y /\ Y < ii, W"A)),
falls at: > 0
W* bedeutet also auf allen Stufen at: > 0:
(c) wahr bleibend ab einer vorangehenden Stufe x bis zur gegebenen
Stufe at: (ausschlieBlich).
Dies ist die gewiinschte Bedeutung (a), (b).

N8' [-] WiA
[-]WA
E6', E7', E8' entstehen aus E6, E7, E8 durch Ersetzung von "iii" durch
"II" .
E9' wPA, wenn p < (X und A ~ -verifizierbar ist.
E10' - WPA, wenn p < (X und A nicht~-verifizierbar ist.
Eel [ - W -] ii E Y, wenn u Element von v ist.
Ee2 - [W] ii E Y, wenn u nicht Element von v ist.
Die metasprachlichen Pradikate W;, F; usw., (X-Giiltigkeit, (X-Folgerung
usw. sind fiir LRe analog zu LR definiert. Z.B. besagt Ir-~"C, daB C E
W~, fiir ane LRe-Interpretationen qJ auf anen Stufen p ~ (X in VI. LRe ist
auf allen Stufen eine konservative Erweiterung bzgl. Giiltigkeit, Folge-
rung, Erfillibarkeit von LR. Zusatzlich gelten in LRe auf allen Stufen ane
ZF-Theoreme, femer auch alle uns unbekannten und ane in ZF unent-
scheidbaren wahren mengentheoretischen Aussagen iiber die Objekte E VI.
Von diesem Reichtum kennen wir leider nicht viel, aber die folgenden
einfachen Liigner-Theoreme lassen sich leicht beweisen. Der auf N be-
schrankte Liigner von (W8) ist auf allen transfiniten Stufen wahr:
(W10) Ir-~co zxPx = '/\x (Nx, -WX zxPx)' -+ WZXPX
Und allgemein istjeder aufirgendein (X E VI beschriinkte Liigner ab Stufe
(X wahr:
(Wll) Ir-~" zxPx = '/\x (x < eX, -WX zxPx)' -+ WZXPx39
Aber der schrankenlose transfinite Liigner ist auf allen Stufen offen, also
nicht wahr:
(W12) Ir-~"+l zxPx = '/\x (ONx, -WX zxPx)' -+ Oii zxPx "
_WiiZXPx
Hier wiederholt sich in LRe das Resultat (W8) von LR<. Und analog zu
(W9) gilt fiir S = U, IU, 0 in LRe:
(W13) Ir-~"+l zxPx = '/\x (ONx, SX zxPx)' -+ Si zxPx
Ane diese schrankenlosen Liigner haben recht - auf allen iiufteren Stufen
454 ULRICH BLAU
;:::: m(l), die wir im nachsten Universum \/2 finden. Und wenn wir - selbst
in \/3 - die LRe-Interpretationen iiber \/2 definieren, wird sich das Spiel
wiederholen.
Beobachten wir zum AbschluB, wie sich die ungestuften Liigner im
Transfiniten verhalten. Vom Unrefiektierten gibt es nichts Neues. Analog
zu (W 5) gilt:
(W14) 11-;':~+l,XPX r _ W lXPX' -+ 0" lXPX " _!FD ii lXPX
Und yom Refiektierten gibt es wenig Neues. Fiir alle Limeszahlen (X und
alle n ;:::: 0 gilt:
- (W15) 11-;':a+2n+1 lxPX r _ W*'XPX' -+ W~+2n lXPx " F~+2n+1
lXPX " !FD,XPX

Der refiektierte Liigner ist auf allen Limesstufen wahr, aber dann beginnt
er wieder zu oszillieren. Und ganz am Ende, nach allen Spriingen iiber alle
Stufen, springt er aus dem Universum \/p heraus auf die erste Limesstufe
m(p) im nachsten Universum \/P+ 1. Dort ist er wahr, aber sobald wir dies
erkennen, beginnt er wieder zu oszillieren: FM(P)+l, WM(P)+2, FM(P)+3, '"
Solange er nicht die Grenze des Absoluten erreicht, wird er keine Ruhe
finden.
ANMERKUNGEN
1 In dieser Kurzfassung sind fast alle Beweise und viele naheliegende Erweiterungen ausge-
lassen. Wichtiger war mir eine genaue Begriindung und Darstellung des Systems. Eine in
vielen Hinsichten erweiterte Fassung ist in Vorbereitung.
2 In meinen Augen ein Pleonasmus. Mir scheint, es gibt nur eine Logik, und da ihr klassischer
Teil bekannt ist, wird jede weitere Logik konservativ oder keine Logik sein.
3 In [34], Bd. I, S. 173 ff.
4 Genaueres zum fundierten Bereich und L3 enthiilt [2], [3], [4], [6]. Weitgehend iihnliche
Auffassungen vertritt Keenan, [19], [20], [21]. Eine noch friihere, viet zu wenig beachtete
Arbeit stammt von Smiley, [33].
5 Zugegeben, die Vorstellung, daB der epistemisch unendlich iiberlegene Verifikator dem
Sprecher noch semantisch folgen konnte, daB er iiberhaupt noch referieren und priidizieren,
d.h. Objekte konzipieren und klassifizieren konnte, ist fast unertriiglich naiv. Aber jeder, der
wahmimmt, spricht und Sprache analysiert, verhiilt sich naiv. Auf dem Wege
Realitiit -+ Wahmehmung -+ Sprache -+ Logik/Mengentheorie
zerstort jeder Schritt Zusammenhiinge und setzt Schnitte, die durch nichts zu rechtfertigen
sind, auBer durch ihre Unvermeidlichkeit. Daher brauchen wir nicht zu iiberlegen, ob, son-
dem nur wo wir Schnitte setzen wollen. Die hier vorgeschlagenen Schnitte zwischen Wahr-
heitswerten und Refiexionsstufen scheinen mir das absolute Minimum fUr die eingangs ge-
nannten Ziele zu sein.
6 No wird benotigt zur Definition der priisupponierenden Konjunktion "I", mit der die an-
zahlpriisupponierenden Quantoren und der Kennzeichnungsoperator definiert werden; ahn-
lich wie spater in LR.
7 Dies zeigen die sog. "Fuzzy"-Logiken, vgl. z.B. [35], die im wesentlichen aus der Verwechs-
lung von Wahrheitswerten mit Akzeptabilitiitsgraden entstanden sind. Vielleicht tragen sie
etwas bei zum Verstiindnis pragmatischer und dynamischer Aspekte der Sprache, aber in
logischer Hinsicht bieten sie nichts. Ich kenne keinen informell giiltigen SchluB, der mit einer
"Fuzzy"-Logik, und nicht besser ohne sie, beweisbar ware.
8 Naheres zu ihnen am Ende von Abschnitt 2.
9 Einfache Anfiihrungszeichen verwenden wir manchmal zur Quasianfiihrung im Sinn von
Quine. Seine Symbole "r", "1", die bei iterierter Anwendung vorteilhafter sind, verwenden
wir als objektsprachliche Anfiihrungszeichen in LR.
10 In L3 entspricht dem Unterschied zwischen gleichstarken und semantisch korrekten For-
malisierungen der Unterschied zwischen Bikonditional " ...." und Aquivalenz " == "; vgl. dazu
die Beispiele (3b)/(4b) und (5a)/(Sb) in Abschnitt 4.2.
11 Mir scheint sogar, semantisch korrekt. Aber darauf kommt es Lf. nicht an.
12 Beschriinkte Allsiitze auBerhalb der Mathematik und Logik haben im allgemeinen An-
zahlpriisuppositionen; vgl. Abschnitt 4.S.
13 Vgl. [2], S. 99 if.
14 ledenfalls nicht mit den iiblichen Objektvariablen, oder de-re-Variablen. Zur Analyse
nicht-extensionaler Konstruktionen scheint es mir zweckmiiBig, eine weitere Sorte von de-
dicto-Variablen einzufiihren, mit denen man auch in nicht-extensionale Bereiche, sogar An-
fiihrungen, hineinquantifizieren kann; vgl. Anmerkung 3S.
15 Eine vorsichtigere Form des I. Vorschlags wiirde unfundierte Siitze nicht als ungram-
matisch, sondern als logisch unanalysierbar brandmarken. Die folgenden Argumente richten
sich auch gegen die vorsichtigere Form. Z.B. wiirden wir nie erfahren, ob (la) logisch analy-
sierbar ist.
16 Vollig sicher ist das nicht. Es konnte eine sich selbst erweiternde Hierarchie von formalen
Mengentheorien geben, die das Absolute durchliiuft. Vgl. dazu den Mechanismus MI* in
Abschnitt 5.4.
17 VO geht nach dem Prinzip "Breite vor Tiefe" wie eine Turing-Suchmaschine vor, solange
er auf Stufe 0 bleibt. Mir scheint, daB der natiirliche Betrachter ebenso vorgeht, aber er hiilt
es nicht lange auf Stufe 0 aus.
18 Einige Bemerkungen zu verwandten Ansiitzen in der neueren Literatur. Wenn man die
FJN-Unterscheidung aufgibt und sich auf die Stufe 0 beschriinkt, so reduziert sich die Hie-
rarchie der Hexatomien auf die Trichotomie WOJFoW/OoW. Das sind die Wahrheitswerte
von Kripke [24]. Die gestuften wn, FD haben Ahnlichkeit mit den Wahrheitswerten von Burge
[S]. Die refiektierbaren, oszillierenden W*, F* finden sich iihnlich bei Gupta [17] und Herz-
berger [IS]. Meine Arbeit ist unabhiingig von diesen entstanden; um so mehr beeindruckt
mich die Konvergenz. Mir scheint aber, daB keiner dieser Autoren die Phiinomene ganz
erfaBt hat. Ihre LOsungen der semantischen Paradoxien sind mehr oder minder unk1ar, un-
vollstiindig und nicht, wie hier, objektsprachliche Theoreme; vgl. Abschnitte 5 und 6 unten.
19 So wird aus einem schonen runden Objekt, z.B. einer Tomate, ein trauriger Tintenfieck.
Nicht jeder theoretische Fortschritt ist ein praktischer Gewinn. Die beiden letzten forma1i-
stisch-nominalistischen Reduktionsschritte:
reine, d.h. aus 0 erzeugte, Mengen ..... forma1e Ausdriicke
..... konkrete Gebilde
erscheinen mir praktisch ebenso wertlos wie theoretisch blind. 1m Zerfall der konkreten Ob-
456 ULRICH BLAU
jekte unter dem Blick des Theoretikers bleiben die reinen Mengen erhalten, solange er Logik
und Mathematik verwendet. Reine Mengen sind reine Formen; da sie keine Substanz, keinen
Anteil am raumzeitlichen Universum haben, kannen sie nicht zerfallen. Das MiBtrauen gegen
sie stammt natiirlich nicht aus der empirischen Forschung, sondern aus der sog. Grundla-
genkrise der Mathematik, die es objektiv nie gegeben hat. Mehr dazu in Abschnitt 5.4.
20 Ein Wort zu den Altemativen zur Aufgabe von (c).
(a) Man kannte den Ko"ektheitsanspruch aufgeben und das Bild von Stufe zu Stufe innen
korrigieren. Etwas von dieser Art schlagt Gupta [17] zur LOsung der semantischen Parada-
xien vor. Mir scheint, daB rein logische Resultate nicht so einfach zu korrigieren sind. Welche
tiefere Evidenz konnte sie korrigieren? Guptas Evidenz ist sehr einfach: Jeder Satz, auch jeder
unfundierte, mu,P auf jeder Stufe wahr oder fa1sch sein. Warum muD er? Das ist der Voll-
stindigkeitsanspruch im unrefiektierten Bild L2.
(b) Man kannte die Realitiit des Bildes aufzugeben versuchen; das ist sehr schwierig. Ein
faszinierendes Beispiel ist Wittgensteins ziemlich hofrnungsloser Kampf mit dem Rea1ititsan-
spruch innerer und abstrakter Objekte.
21 Praktisch werden wir die formale Sprache nicht verwenden, sondem uns weiterhin mit
Metavariablen begniigen.
22 Vgl. dazu die Diskussion am Ende von Abschnitt 4.9.
23 Hat Casar in derselben Stadt gelebt wie der heutige Papst? Wann zeigen zwei Leute auf
dieselbe Wolke, wann haben sie denselben Gedanken, dasselbe Alter, denselben Beruf! Eine
"Fuzzification" der Existenz und Identitit ware eine Verschlimmbesserung; Stidte, Wolken,
Gedanken existieren nicht im Grad 0,649372381 und sie sind auch nicht in diesem Grade
identisch ("nicht" im schwachen Sinn.)
24 Die syntaktische Sprachstufenhierarchie wird also durch die semantische Refiexionsstu-
fenhierarchie ersetzt. Das kannte den Verdacht eines Etikettenwechsels hervorrufen; nichts
ware falscher. In L2 und der Sprachstufenhierarchie sind nur fundierte Satze formalisierbar,
deren Stufe man kennt. Der vielleicht ganz harmlose und fundierte obige Satz ist nicht for-
malisierbar:
(la) Der letzte Satz, den Philitas von Kos geauBert hat, ist nicht wahr.
Wir kennen seine Stufe nicht. In LR entfiillt diese Anwendungsbeschrankung, (la) ist ambig
zwischen
(la') - Wa (mit genauer analysierbarer Kennzeichnung a) und
(Ian) -W*a
Es wird nicht verlangt, daB wir die Refiexionsstufe des Sprechers kennen - falls er eine hat:
Bei fundierten satzen ist die Stufe gleich, bei oszillierenden satzen kann er kaum auf einer
Stufe halt machen.
25 Da wir die logischen Konstanten von LR metasprachlich autonym verwenden, schreiben
wir z.B. "IAI; = w" statt "IAI; = "w" ". Zur Korrektheit der Definition sei bemerkt, daB
IAI; nach dem spateren Satz 3 eindeUtig ist.
26 Wir flihren kein eigenes Symbol ein, da die Allgemeingiiltigkeit mit der O-Giiltigkeit "11- 0 "
zusammenfiillt.
27 Vgl. [6].
28 Vgl. [27], S. 283 fr.
29 V gl. [34], Bd. 1, S. 173 fr.
30 Vgl. [2], S. 210.
31 Mit ihm laBt sich auch das vieldiskutierte Bach-Peters-Paradox semantisch korrekt und
syntaktisch oberfl.achennah formalisieren, vgl. [1].
UNBESTIMMTHEITEN UND P ARADOXIEN 457
32 Die Kennzeiehnungstheorie von LR ist zwar besser als jede andere, die ieh kenne, aber
natiirlieh lost sie Dieht aile Kennzeiehnungsprobleme. 3 Probleme, deren LOsung ieh zu ken-
nen giaube:
(I) Das Klavier [d.h. dieses Klavier, d.h. das einzige Klavier, auf das ich jetzt
referiere] ist verstimmt.
(2) Das [durehschnittliehe, bzw. idealtypische] Klavier ist ein Hausinstrument.
(3) Die Kinder haben [gemeinsam] das Klavier aus dem Fenster geworfen.
(I) verlangt Indikatoren, (2) veriangt einen Abstraktionskennzeichner, (3) veriangt einen Kol-
lektionskennzeichner. Vorliufiges zu (I), (2), (3) enthilt [7], [4] und [3].
33 Dieser Abschnitt hat dureh hi1freiehe Kritik von M. Drommer, M. Lohner, A. Oberschelp
und vor aIlem von W. Pohlers einige vage Spekulationen und einen Fehler eingebiiBt.
34 1m folgenden iindem wir die Notation: "x", "y", "z" sind objekt- und metasprachliehe
Variablen fiir Mengen; "IX", "P", "y" sind spezie11e objekt- und metaspraehliehe Variablen
fUr Ordina1zah1en. Diese Anderung der Notation hat mit der Anderung des Gegenstandes
zu tun. Die semantischen Paradoxien sind ein Problem der formaien Sprache: Bisher haben
wir sie nur unwesentlieh verwendet und im wesentliehen iiber sie gesprochen. Dazu muBten
wir sie von der Metaspraehe unterscheiden. - Die Mengenparadoxien sind ein Problem der
realen. abstrakten Mengen: Um iiber sie zu sprechen, miissen wir wesentlieh die forma1e
Spraehe verwenden; die Unterscheidung von der Metasprache ist weniger wiehtig.
3S Hierarchien der Art (I) von arithmetischen Systemen S· mit konstruktiven 1% hat vor aIlem
Feferman [12] untersueht.
36 Vgi. Drake [11], S. 112 und 123.
37Unvergieiehlieh mehr als das "iiuBere" Paradoxieproblem der Darste11ungsmittel beunru-
higte ihn das "innere" Kontinuumproblem der dargeste11ten Sache. Man lese aueh noch
einmal die beriihmte Passage fiber das Absolute in [9], S. 205.
38 Mir scheint, daB in Freges Philosophie ein gewisser Zusammenhang besteht zwischen
(a) seinen absoluten Klassen, als Objekten mathematischer Priidikate,
(b) seinen objektiven Satzsinnen ("Gedanken"), als Objekten semantischer und intentio-
naler Pradikate ("wahr", "falsch", ... "giauben", "wissen", ...)
Beide sind innen-auBen-Zwitter. Ebenso wie man zum Verstiindnis der Mathematik Freges
absolute Klassen aufgeben muBte, wird man zum Verstandnis von Semantik und Intentio-
nalitiit friiher oder spiiter seine objektiven Gedanken aufgeben. Gedanken von innen, flir
den, der sie hat, sind Teil der Realitat; von auBen, fUr den, der sie betraehtet, sind sie Teil
eines Bildes. Zwischen "haben" und "betrachten" liegen dunkle Prozesse, in denen inten-
tionale Einste11ungen, Affirmation und Negation entstehen. Was nicht entsteht, sind scharf
individuierte Gedanken, Propositionen, Intensionen, wie immer man sie nennt und konzi-
piert. Mir scheint, daB diese Objekte fUr die logische Sprachanalyse ohne jeden Wert sind.
Um Dieht falsch verstanden zu werden: Wenn der Sprecher behauptet, was selten vorkommt:
(I) a hatte zwischen 9.15 und 9.41 genau sieben Gedanken
so soli der Sprachanalytiker ibm folgen:
(I') 7!x (Px, Qabcx)
P: Gedanke, QxYZX1: x hatte zwischen y und z Xl
b: 9.15, e: 9.41
Er soli die Ontologie des Sprechers Dieht korrigieren, da jede Dieht-forma1e Ontologie naiv
ist. Aber er selbst, als Theoretiker, sol1te solehe naiven Objekte wie Gedanken oder Propo-
sitionen Dieht zur Grundlage der Semantik machen.
458 ULRICH BLAU
Die bekannten logisch-linguistischen Schwierigkeiten mit der Intentionalitit entstehen aus

den innen-au.6en-Mischungen der Ontologie des intentionalen Subjekts mit der Sprecher-
Ontologie an den Argumentstellen intentionaler Prlidikate. Aber nach Aufgabe der Propo-
sitionen ist die Analyse nicht schwierig. Es geniigt, in der logischen Syntax
1. bei den Elementarprlidikaten de-re- und de-dicto-Argumentstellen durch Indizes zu un-
terscheiden,
2. de-re- und de-dicto- Variablen durch dieselben Indizes zu unterscheiden,
3. de-re-Variablen nur an de-re-Argumentstellen zuzulassen.
Die de-re-Argumentstellen sind referentiell, d.h. extensional und existenzpriisupponierend;
die de-dicto-Argumentstellen sind im aIlgemeinen weder das eine noch das andere. Die de-
re-Variablen beziehen sich auf die Objekte u des vom Sprecher intendierten Universums U;
ihre Quantorenregeln beziehen sich auf die Grundbezeichnungen ii E U. Die de-dicto-Varia-
bien beziehen sich moglicherweise auf gar nichts; ihre Quantorenregeln beziehen sich auf die
abzlihlbar unendlich vielen reinen Objektbezeichnungen, gleichgiiltig, ob und was sie bezeich-
nen. Mit diesen einfachen Mitteln losen sich die Probleme, an denen Quines Ralph [28],
Kripkes Pierre [25] und aile mir bekannten logisch-linguistischen Theorien der Intentionalitit
laborieren; vgl. [7].
39 Die LR,;-Theoreme (Wll)-{W15) gelten fUr aile Ordinalzahlen ex, die in \,1 bezeichenbar
sind, vgl. (2), S. 448.
LITERATUR
[I] Bach, E.: 'Problominalization', Linguistic Inquiry 1 (1970), 121-122.

[2] Blau, U.: Die dreiwertige Logik der Sprache. de Gruyter, Berlin, 1978.
[3] Blau, U.: 'Collective Objects', Theoretical Linguistics 8 (1981), 101-130.
[4] Blau, U.: 'Abstract Objects', Theoretical Linguistics 8 (1981), 131-144.
[5] Blau, U.: 'Vom Henker, vom Liigner und von ihrem Ende', Erkenntnis 19 (1983), 27-
44.
[6] Blau, U.: 'Three-valued Analysis of Precise, Vague, and Presupposing Quantifiers', in
Approaching Vagueness, (hg. v. T. Ballmer und M. Pinkal), North-Holland, Amsterdam,
1983, S. 79-129.
[7] Blau, U.: Intentionale Priidikate und das de-re/de-dicto-Problem, Unveroffentlichtes Vor-
lesungsskript, Miinchen, 1983.
[8] Burge, T.: 'Semantical Paradox'. Journal of Philosophy 76 (1979), 169-198.
[9] Cantor, G.: Geso.mmelte Abhandlungen (hg. v. E. Zermelo), Georg Olms, Hildesheim,
1962.
[10] da Costa, N. C. A.: 'On the Theory ofInconsistent Formal Systems', Notre Dame Jour-
nal of Formal Logic 15 (1974), 497-510.
[11] Drake, F. R.: Set Theory. An Introduction to Lorge Cardinals, North-Holland, Amster-
dam, 1974.
[12] Feferman, S.: 'Transfinite Recursive Progressions of Axiomatic Theories', Journal of
Symbolic Logic 27 (1962),259-316.
[13] Finsler, P.: 'Gibt es Widerspriiche in der Mathematik?' Jahresbericht dt. Mathemati-
ker-Vereinigung 34 (1925/26),143-155.
[14] Frege, G.: Grundgesetze der Arithmetik, Bd. 1 und 2. Hermann Pohle, Jena (1893/1903).
[15] Godel, K.: 'What is Cantor's Continuum Problem?', in Philosophy of Mathematics (hg.
v. P. Bencerraf und H. Putnam), Blackwell, Oxford, 1964, S. 258-273.
[16] Grelling, K. und Nelson, L.: 'Bemerkungen zu den Paradoxien von Russell und Bura-
Ii-Forti', Abhandlungen der Fries'schen Schule 2 (1907/08), 300-334.
(17) Gupta, A.: 'Truth and Paradox', Journal of Philosophical Logic 11 (1982), 1--60.
(18) Herzberger, H.: 'Notes on Naive Semantics', Journal of Philosophical Logic 11 (1982),
61-102.
(19) Keenan, E. L.: 'Quantifier Structures in English', Foundations of Language 7 (1971),
255-284.
(20) Keenan, E. L.: 'On Semantically Based Grammar', Linguistic Inquiry 3 (1972),413-461.
(21) Keenan, E. L.: 'Presupposition in Natural Logic', The Monist 57 (1973), 344-370.
(22) Kleene, S. c.: Introduction to Metamathematics, North-Holland, Amsterdam, 1952.
(23) Konig, J.: 'Ober die Grundlage der Mengenlehre und das Kontinuumproblern', Ma-
thematische Annalen 61 (1905), 156-160.
(24) Kripke, S. A.: 'Outline ofa Theory of Truth', Journal of Philosophy 72 (1975),690-716.
(25) Kripke, S. A.: 'A Puzzle about Belief', in 'Meaning and Use' (hg. v. A. Margalit), D.
Reidel, Dordrecht (1979), S. 234-283.
(26) Lukasiewicz, J.: Selected Works (h.g. v. L. Borkowski, North-Holland, Amsterdam,
1970.
[27] Quine, W. V. 0.: Mathematical Logic, Harvard University Press, Cambridge, 1951.
(28) Quine, W. V. 0.: 'Quantifiers and Propositional Attitudes', Journal of Philosophy 53
(1956), 177-187.
(29) Quine, W. V. 0.: Set Theory and its Logic. Harvard University Press, Cambridge, 1963.
(30) Reichenbach, H.: 'Ober die erkenntnistheoretische Problernlage und den Gebrauch einer
dreiwertigen Logik in der Quantenrnechanik', Zeitschriftfiir Naturforschung 6a (1951),
569-575.
(31) RoutIey, R.: 'Dialectical Logic, Semantics and Metamathematics', Erkenntnis 14 (1979),
301-331.
(32) Scott, D.: 'Axiomatizing Set Theory', in Proceedings of Symposia in Pure Mathematics
Bd. 13, Teil2: Axiomatic Set Theory, Amer. Math. Soc., Providence, 1974, S. 207-214.
(33) Smiley, T.: 'Sense without Denotation', Analysis 20 (1959/60), 125-135.
(34) Whitehead, A. N. und Russell, B.: Principia Mathematica, Bd 1-3, Cambridge University
Press, London 2 , 1925(27.
(35) Zadeh, L. A.: 'Fuzzy Logic and Approximate Reasoning', Synthese 30 (1975), 407-428.
(36) Zerrnelo, E.: 'Untersuchungen iiber die Grundlagen der Mengenlehre', Mathematische
Annalen 65 (1908), 261-281.
Manuscript Received 12 January 1984
Seminar fiir Philosophie,

Logik und Wissenschaftstheorie
Universitiit Miinchen
Ludwigstr. 31
8000 Miinchen 22
B.R.D.
HANS LENK
BEMERKUNGEN ZUR PRAGMATISCH-EPISTEMISCHEN

WENDE IN DER WISSENSCHAFTSTHEORETISCHEN
ANALYSE DER EREIGNISERKLARUNGEN
Carl Gustav Hempel, der Vater der modernen wissenschaftstheoretischen

Analyse der ErkUirung, unterschied "Erkliirung-verlangende" Warum-Fra-
gen (manchmal auch besser ErkUirung-heischende oder -suchende ge-
nannt) von "Begrundung-verlangenden oder epistemischen" (1977, 2f.)
Dementsprechend antworten ErkUirungen darauf, warum ein Phanomen
"in der angegebenen Weise auftritt und warum es physikalische Eigen-
schaften ... besitzt" (ebd). Bei Ereignissen wird eine solche "echte" Erkla-
rung normalerweise eine kausale Erklarung sein. Epistemische Warum-
Fragen suchen demgegenuber bloB nach Begriindungen dafiir, daB man
einen eine Eigenschaft oder ein Ereignis beschreibenden Satz annehmen,
glauben, erwarten darf: "Eine epistemische Warum-Frage setzt die Wahr-
heit der entsprechenden Aussage nicht voraus, sondern verlangt stattdes-
sen nach Grunden fUr die Annahme, daB sie wahr ist" (ebd.3).
Die neueren "pragmatisch-epistemischen" Ansatze in der Erklarungs-
theorie geben diese Unterscheidungen entweder prinzipiell oder prak-
tisch-faktisch einfach auf - aus underschiedlichen, einander erganzenden
Grunden. Ich werde zu zeigen versuchen, daB diese Griinde z.T. nicht
hinreichen, z.T. eine extreme Reaktion (namlich die vollige Nichtberuck-
sichtigung von echten Ursachenerklarungen oder deren Verdrangung aus
der Erklarungsdebatte), Einseitigkeiten der Analyse und terminologischer
Konventionen oder einen inadaquaten Modellintegrationismus zur Folge
haben, der die bisherige Orientierungsstrategie einfach umkehrt.
Hempels Hinweis, die Konstruktion eines nichtpragmatischen Modell-
begriffs der wissenschaftlichen Erklarung, "der sozusagen aus dem prag-
matischen abstrahiert wird und der einer Relativierung auf fragenstellende
Personen ebensowenig bedarf wie der mathematische Beweisbegriff', , sei
sinnvoll als idealisierende Modellkonstruktion, wird dabei ebensowenig
beachtet wie seine darin angedeutete Reservation: "Diese Modelle zu ver-
treten bedeutet deshalb weder, die pragmatische 'Dimension' der Erkla-
Erkenntnis 22 (1985) 461-473. 0165--0106/85/0220--0461 $01.30

462 HANS LENK
rung zu leugnen, noch, ihre Bedeutung herabzusetzen, natiirlich erst recht

Dicht, zu behaupten, daB man eine ErkUirung nur dann einleuchtend und
befriedigend finden kann, wenn sie einem der 'covering-Iaw'-Modelle ent-
spricht" (ebd. 144). Selbst wenn die semantisch-syntaktischen Rekonstruk-
tionsversuche fiir ein ideales Modell der deduktiv-nomologischen Erkla-
rung scheitern sollten, wie es aufgrund der Problematik der Mehrdeutig-
keit der Wissenssituation bei der induktiv-statistischen der Fall ist, konnte
es dennoch sinnvoll und heuristisch empfehlenswert sein, soweit wie mo-
glich rein logisch (syntaktisch-semantisch) vorzugehen und von einem vor-
eiligen Oberlaufen zur Totalpragmatik abzusehen. Dies diirfte die Inten-
tion der zitierten Hempelschen Stellungnahmen sein. Selbst wenn syste-
matisch kriterienorientiert kein vollstiindiger Erfolg moglich sein sollte,
ware heuristisch die logisch-semantische Rekonstruktion doch sinnvoll -
ebenso sinnvoll, wie es idealisierte Modellkonstruktionen und Gesetze in
der objektsprachlichen Theoriebildung der Wissenschaft selbst sind. Hist-
orisch gesehen, hat die Entwicklung der Diskussion von den idealisieren-
den Modellkonstruktionen erheblich profitiert, ja, der differenzierte Ent-
wicklungsstand und das hohe Niveau der Erklarungsdebatte waren ver-
mutlich ohne diese Rekonstruktionsbemiihungen Dicht moglich gewesen.
Natiirlich kann man auf eine pragmatische Einbettung der Instantiie-
rungen der idealisierten Modellrekonstruktionen nicht verzichten: Man
darf also die Bemiihungen urn die syntaktisch-semantische Modellkon-
struktion Dicht als einen volligen Verzicht aufjegliche Kontexteinbindung
auffassen. Angesichts der geschilderten Intention Hempels und dieser von
mir (schon 1972) angedeuteten Argumentation, ist es unverstandlich, wie
man "die Beharrung... auf einen kontextfreien Begriff der Ursachenerkla-
rung" in totaler Weise unterstellen und "schon deshalb unverstandlich"
finden kann (Schurz 1983,279).
N och unverstandlicher bleibt, wie man solche Versuche als "in gewisser
Weise anachronistisch" (ebd.) bezeichnen kann mit der Behauptung: "Der
Grund, warum zwischen Begriindungen und Erklarungen iiberhaupt un-
terschieden werden muBte, war ja gerade die Einsicht, daB Erkliirungen
im engen Sinne keinesfalls durch kontextfreie logische Argumente dar-
stellbar sind (!)". SolI gemeint sein, daB Begriindungen von vornherein
Dicht kontextfrei und unpragmatisch zu geben sind? Das aber ware trivial.
Sollten Argumentstrukturen von Ursachenerkliirungen, wie Schurz mit
Gardenfors behauptet, ebenfalls stets nur kontextgebunden explizierbar
EREIG NISERKLAR UNG EN 463
sein? Dann aber wiire die Notwendigkeit der Unterscheidung von Begriin-
dungen und Erkliirungen gar nicht auf dieses Merlanal zu griinden. Sollte
sie deshalb giinzlich fallen - im Gegensatz zur Intuition der Naturwissen-
schaftler und des Alltags?
Hintergrundwissen ist auch notig, um einzusehen, daB die Anderung der
Pendelliinge Ursache der Frequenziinderung ist, aber nicht umgekehrt
(ebd.) - dies ist unbestritten. Es hebt aber nicht den Unterschied zwischen
Begriindung und Ursachenerkliirung in bezug auf den Pragmatizitiitsgrad
giinzlich auf. (Das von Schurz (ebd. 279) angefiihrte Beispiel einer Aqui-
valenz zwischen Pendelliinge und Frequenz kann iibrigens nicht als Kau-
salgesetz gelten, da Kausalgezetze niemals Aquivalenzgesetze sind und als
solche nicht in kausalen Erkliirungen auftreten konnten; die Abschwii-
chung zu einem Gesetz formaler Implikation wiire nur in einer Richtung
fiir eine kausale Erkliirung, in der anderen nur fiir eine (Koexistenz-) Be-
griindung geeignet.) Sollte das pragmatische Hintergrundwissen bei der
Auszeichnung von Ursachenerkliirungen sich auf Gesichtspunkte des
durch Handelen direkt Veriinderbaren beziehen? (Pendelliingen konnen
wir direkt veriindem, Pendelfrequenzen nur indirekt - durch Anderung der
Pendelliinge). Diese Einsicht ist wohl richtig, betrifft aber nur die prag-
matische Gesamteinbettung der ErkUirung und deren faktische Anwen-
dung bzw. Kennzeichnung der Komponenten im Einzelfall, in ihr aber
notwendig die realen Grundlagen der Argumentstruktur. DaB pragma-
tische Kontextabhiingigkeit bei der Feststellung eine Rolle spielt, ob etwas
als ursiichlich anzusehen ist, bedeutet noch nicht, daB die Ursache-Wir-
kungs-Relation selbst giinZlich pragmatisiert, ins Pragmatische aufgelost
werden muB. Die Unterscheidung zwischen Begriindung und Erkliirung
wurde eher umgekehrt darauf gestiitzt, nach Moglichkeit Erkliirungen
eben zuniichst einmal modellmiiBig moglichst weitgehend unpragmatisch
zu explizieren oder zu rekonstruieren, und zudem auf die Verabredung,
daB wissenschaftliche Begriindungssystematisierungen nicht notwendig
vom Zutreffen des Explanandums ausgehen miissen. (Dies letztere ist auch
heute n'och bei Stegmiiller etwa der kennzeichnende Unterschied
zwischen "Begriindungen" und "Erkliirungen"). Der eben zitierte Autor
(Schurz 1983) beklagte selbst "die heute viel verbreitete resignative Ein-
stellung gegeniiber dem D-N-Erkliirungsmodell", die "auf eine falsche
Vermengung von Problemkategorien zuriickzufiihren" sei (ebd. 277). Im-
plizit nimmt er selbst auf" 'echte' Erkliirungen" Bezug (z.B. ebd. 279
u.a.m.).
464 HANS LENK
Urn nicht miBverstanden zu werden, mochte ich betonen, daB letztlich

die Beriicksichtigung pragmatischer Komponenten in der Wissenschafts-
theorie anzuerkennen ist und von mir (1975) auch gesehen und gefordert
wurde, allerdings bevor die "pragmatisch-epistemische Wende" (Stegmiil-
ler 1983, 940) eingelautet worden ist.
1m folgenden seien einige ausgewahlte Argumente der neuen Totalprag-
matiker der Erklarungstheorie diskutiert. Eine Folgerung wird sein, daB
eine anpasserische Manipulation mit dem Terminus "ErkHirung" vorge-
nommen worden ist und daB die Beziehung zwischen diesem Terminus
und un serer orientierenden Intuition, was eine "echte Erklarung" sei, in
Abhangigkeit von dieser Manipulation andersartig gedeutet oder besser:
vorgeschlagen wird.
Ausgehend von einem vieldiskutierten Beispiel von U. Blau, einem
SchluBschema, das in einer Deutung als "Pseudo-Erklarung", in einer an-
deren, "wenn auch nicht fUr eine 'Erklarung', so doch fUr eine korrekte
wissenschaftliche Begriindung" gehalten werden kann (Stegmiiller 1983,
919), versucht Schurz (1983, 269, 260 ff.) von den singularen Bedingungen
des Blauschen Beispiels wegzukommen und einen systematischen Beweis
dafUr zu liefem, "daB ein und dieselbe logische Struktur bei der einen
Interpretation zu Pseudoerklarungsfallen fUhren kann, die offensichtlich
absurd sind, wahrend sie bei einer anderen Interpretation zu absolut zu-
lassigen und sinnvollen Fallen von D-N-Erklarung fiihrt" (ebd. 269). Der
Autor macht ausschlieBlich die Kontextabhangigkeit und damit in gewis-
ser Weise doch wohl pragmatische Momente hierfUr verantwortlich und
fordert dementsprechend "die Beriicksichtigung epistemischer Hinter-
grundkontexte" in einer durchaus objektiven "logischen Pragmatik" (ebd.
274). Dabei wird besonders auf die "prognostische Relevanz" verwiesen
(auf die yom Wissen des Explanandums und dessen Bestatigung unab-
hangige Bestatigbarkeit der Explananspramissen), wobei zu Recht fest-
gestellt wird, daB bereits Ermittlung und Bestatigung nicht auf Verifika-
tion (im Sinne der Ableitbarkeit aus Basissatzen) einzuschranken sind
(ebd. 263, 273).
Schurz selbst gesteht allerdings auch logisch-semantisch zu kennzeich-
nende Merkmale der Erklarung zu wie die logisch-analytische Invarianz
von Pramissen und Konklusion in bezug auf logisch aquivalente Umfor-
mungen, Gesetzesadaquatheit und deduktive Relevanz ("aIle konjunktiven
Pramissenkomponenten sollen zur Deduktion des Explanandums bzw. sei-
EREIGNISERKLARUNGEN 465
ner konjunktiven Komponente wirklich notwendig sein; und keine dis-

junktive Explanandumkomponente darf durch ihre Negation ersetzbar
sein" (ebd. 263».
Ahnlich wie seinerzeit gegeniiber dem Beispiel von Blau (Verf. 1972,
52f.) kann man allerdings auch gegeniiber den vier Basistypen von Schurz
behaupten, daB es sich bei ihnen auch dort um Begriindungsfalle im Hem-
pelschen Sinne handelt, wo von Fallen "absolut akzeptabler Erklarung",
"giiltiger Erklarung", " 'echter' Erklarung" die Rede ist (ebd. 265, 279).
Beispielsweise stimmt es wohl nicht, daB "man das quantenmechanische
Beispiel. .. in jedem Fall nicht bloB als Begriindung, sondem auch als Er-
klarung im engen Sinne ansehen" "muB" (ebd. 280). Das Beispiel ist das
folgende:
"T: Der cheInische Symmetriezustand 1 fiihrt zu einer energetischen Er-
hOhung gegeniiber dem Grundzustand.
A: Die cheInische Verbindung a kann nur zum Symmetriezustand 1 oder
2 gehOren, und befindet sich im Grundzustand.
E: Daher befindet sich a im Symmetriezustand 2" (ebd. 266).

Entsprechend der Kopenhagener Interpretation der Quantenmechanik
muB dies ja wohl als Ausdruck unserer Kenntnis von Zustanden gedeutet
werden - ein typischer Fall fiir eine Begriindung, die angibt, warum wir
schlieBen, glauben, erwarten diirfen, wissen, daB etwas der Fall ist.
In der Tat konnte man meinen, die Kopenhagener Interpretation der
Quantenmechanik und die gesamte Debatte iiber die IndeterIniniertheit
der Messung quantenmechanischer Zustande ffihre dazu, daB hier nur
noch Ereignisbegriindungen vorgenommen werden konnen, obwohl auf der
Theorieebene zwischen den Zustandsvariablen sogar deterIninistische Zu-
sammenhange gegeben sind (Schrodinger-Gleichung). Offenbar werden
Kenntnisse von ZustandsmeBwerten als Begriindungsfaktoren genommen,
als "Vemunftgriinde" statt echter "Ursachen".
In der neuen Auflage seines Opus maximum Bd.1 (1983) wird Stegmiiller
nicht miide, eine "Abkoppelungsthese" oder "Abkoppelungsempfehlung"
derart zu propagieren, daB aus ZweckmaBigkeits- und Einfachheitsgriin-
den das Studium der pragmatischen "informativen Erklarungen" von dem
der "Kausalanalysen" (friiher kausale Erklarungen genannt) abgetrennt
466 HANS LENK
werden miisse (ebd. 9, 633fI., 954, 976, l005fI., 1028 u.a.). Eine gleichzeitige
Beriicksichtigung beider Problembereiche iibersteige die menschliche Leis-
tungs- und Integrationskraft (ebd). Dennoch miiBten letztlich die eher
wilIkiirlich getrennten verschiedenartigen Analysen wissenschaftstheore-
tisch und auch in der Praxis wieder koordiniert und zusammengefiihrt
werden: "Wer an beidem interessiert ist, und das sind diejenigen Personen,
die wir 'Kausalisten' nennen, ... muB beide Arten von Untersuchungen
anstelIen; sonst erreicht er nur die Halfte seines Zieles" (ebd. 954).
Der Hinweis auf die au13erordentliche Kompliziertheit des pragma-
tisch-epistemischen BegrifIs der "Wissenssituation" sowie der damit ver-
bundenen Analyse der Erklarung als der Reduktion von Oberraschungsin-
formation bzw. der ErhOhung von Oberzeugungs- und Glaubenswerten ist
natiirlich keine systematische, sondem eingestandenermaJ3en eine bloB aus
praktischen Oberlegungen motivierte Argumentation. Sie ist natiirlich sy-
stematisch gesehen unzureichend oder wilIkiirlich. 1m iibrigen gibt Steg-
miiller selbst das beste Beispiel, daB fallweise Menschen in der Lage sind,
in Personalunion beide Problembereiche zu iiberblicken bzw. zu integrie-
ren.
Man gewinnt den Eindruk, daB zugunsten der Analyse der epistemi-
schen Warum-Fragen bzw. der Begriindungen die erklarungheischenden
Warum-Fragen bzw. "echte Erklarungen" (Ursachenerklarungen) aus der
Erklarungsdebatte ausgeschieden, in eine eigene Analyse ("Kausalanaly-
se") verdrangt werden - einfach aus unsystematischen, praktischen GrUn-
den. Damit sind aber diese Fragen natiirlich nicht gelost. Es handelt sich
nur urn eine Beschrankung der Untersuchung auf einen Aspekt, namlich
den pragmatisch-epistemischen Gesichtspunkt der sogenannten "rein-in-
formativen" Erklarung. Diese Strategie erscheint wilIkiirlich, zumal man
mit StegmiilIer hervorheben muB, daB eine vollstandige Analyse von Er-
klarungsproblemen eben auch die KausalanalyseeinschlieJ3en miiBte. Der
bloJ3e Verweis auf die ZweckmaBigkeit der Abkopplungsthese ist natiirlich
kein systematisches Argument.
Mit all dem solI keineswegs geleugnet werden, daB eine epistemisch-
pragmatische Relativierung von Erklarungsargumenten sinnvoll und notig
ist (dies hat - wie oben zitiert - auch Hempel schon 1965 zugestanden).
Dies besagt aber wie erwahnt nicht, daBlogisch-semantische idealisierende
Rekonstruktionen und Analysen der kausalen Erklarung unniitz seien.
Was den letzteren Punkt betrifIt, so gesteht Stegmiiller dies implizit auch
wiederholt zu - ebenfaIls, daB der Kausalist die Abkopplungsthese kei-

neswegs akzeptieren muB, ja, daB man durchaus "den Dberzeugungswert
vergroBernde Griinde allein" nicht als hinreichende Erkliirungen anerken-
nen miisse, sondern weiterhin nach Seinsgriinden oder "wahren Ursachen"
suchen konne: "Eine solche Position kann eingenommen werden" (ebd.
1007). Der Kausalist wiirde nach Stegmiiller nur etwa die pragmatisch-
epistemischen ErkHirungen "(bloB) retrospektive Begriindung (oder: ex-
post-Begriindung)" nennen, die pragmatisch-epistemischen Begriindungen
vielleicht "prospektive Begrundungen (oder: ex-ante-Begriindungen)": Fiir je-
nen bedeute allesdas, "was innerhalbdes pragmatisch-epistemischen Rahmens
expliziertwird, ... ausschlieftlichverschiedeneFormen vonBegrundungen" (ebd.
1007).
In der Tat kann man fragen, warum der ErkHirungsbegriff nach der
pragmatisch-epistemischen Wende nicht mehr auf die "Kausalanalyse"
ausgedehnt werden solI oder, anders gewendet, warum "Erkliirungen" nur
noch pragmatisch-epistemische Begriindungen sein sollen. StegmiiIler
meint zwar, man gerate in eine "terminologische Schwierigkeit, wenn man
den Erklarungsfall im informativen Sinn... vom Begriindungsfall i.e.S. -
in welchem man noch gar nicht weiB, ob das zu Begriindende iiberhaupt
eintreffen wird oder nicht - abgrenzen solI", falls man "den Erklarungs-
begriff auf den kausalen Fall zu beschranken" wiinscht (ebd. 1017). Nun,
diese terminologische Schwierigkeit lieBe sich natiirlich leicht beheben -
und weniger miBverstandlich als bei der von den Begriindungspragmati-
kern vorgeschlagenen Variante - indem man etwa "wahre Ereignisbegriin-
dungen" von "potentiellen Begriindungen" unterscheidet. (Statt von wah-
ren Ereignisbegriindungen konnte man natiirlich auch von Begriindungen
eingetroifener Ereignisse oder in diesem Sinne von "zutreffenden Ereig-
nisbegriindungen" sprechen.)
Es scheint hier also eine eher irrefiihrende terminologische Entscheidung
vorzuliegen, wenn man von einer herkommlichen Intuition des Begriffs
"wissenschaftliche Erklarung" (wie sie etwa bei den Naturwissenschaftlern
vorherrscht) ausgeht.
Dberhaupt scheinen die terminologischen Konventionen ungliicklich zu
sein, wenn bei Erklarungen "Ursachenanalysen" ausgeschlossen werden
und im Gegensatz zu friiheren Ansichten Stegmiillers (was etwa den Be-
griindungscharakter der statistischen Systematisierung angeht) der Ter-
minus 'Erklarung' ausschlieBlich fUr den Dberzeugungswert vergroBernde
468 HANS LENK
Griinde reserviert wird. Hier scheint ein pragmatizistischer Obereifer am

Werke zu sein, der irgendwie anscheinend mit der allzu menschlichen Nei-
gung zusammenhangt, von einem Extrem ins andere zu fallen oder das
Kind mit dem Bade auszuschiitten.
In der Tat ist eine tiefgreifende Schwierigkeit bei den Rekonstruktionen
der Wissenschaftstheorie festzustellen: Einerseits werden Modelle rekon-
struiert - etwa eines der echten, alle partiellen Selbsterklarungen
ausschlieBenden D-N-Erklarungen - andererseits, werden diese Modelle
an Beispielen und an unserer vorherrschenden Intuition getestet. Die Sch-
wierigkeit besteht darin, daB die Intuition - abgesehen von extremen
Grenzfallen, etwa der Selbsterklarung oder absurder oder rein theoreti-
scher Erklarungen - keineswegs feststeht, sondem eine Menge mit ver-
flieBenden Grenzen und Zwischenlallen (vielleicht in beiderlei Sinn des
Wortes) ("fuzzy set") ausmacht. Die Intuitionen werden zum Teil durch
methodologische Rekonstruktionen der Vergangenheit erst gepragt; und
sie sind wiederum die MaBstabe fUr den Test der idealen Rekonstruktion.
Daraus folgt, daB gerade in diesen Grenzbereichen die begriindeten Fest-
legungen eine beachtliche Rolle spielen. Ob wir die bloSe ErhOhung eines
Glaubenswertes relativ zu einem Vergeleich von Wissenssituationen vor
Kenntnis der relevanten erklarenden Gesetze und Antecedensbedingungen
und nach ihr bzw. mit EinschluB des Explanandurns auch dann als Erkla-
rung ansehen, wenn die Eintrittswahrscheinlichkeit auSerst gering ist -
vielleicht unter 10% -, ist natiirlich in der Intuition hOchst urnstritten.
Eher scheint es sich hier bei der Idee der den Oberzeugungswert ver-
groBemden Aufgabe der Erklarungen bei Stegmiiller urn einen termino-
logischen Vorschlag zu handeln. Dieser Vorschlag ist nicht notwendig in-
tuitiv evident. Dies ergibt sich schon daraus, daB ein Befolgen der neuen
pragmatisch-epistemischen Vorschlage besonders im "epistemischen Mi-
nimalsinn" (ebd. 975) (das Explanandurn ist aufgrund der Theorien und
der Antecedensbedingungen eher als zutrefrend zu erwarten als ohne diese)
dazu fiihrt, daB es auch ErkHirungen gibt, die keine potentiellen Voraus-
sagen sind (ebd. 955, 983 u.a.) - namlich in dem Sinne, daB die erklarende
Kraft vergroBert werden kann, auch dann, wenn das Explanans noch nicht
fiir eine Prognose des Ereigniseintritts ausreicht, sondem allenfalls fiir eine
schwache probabilistische (bloB wahrscheinlichkeitserhOhende, aber die
Eintrittswahrscheinlichkeit nicht iiber 0,5 hinaustreibende) Argumenta-
tion. Dies wiirde ja bedeuten, daB etwa Hempel nach wie vor (1977, 40 fr.,
bes. 43 u.a.) keine adaquate Intuition von "Erklarung" habe. Mit anderen
Worten: Der Zusammenhang zwischen der priifenden intuitiven Instanz
einerseits und der durchaus auch aktiv funktionierenden methodologi-
schen Rekonstruktion von Erklarungsmodellen andererseits ist viel kom-
plizierter, als auch die neuen Epistemiker es sich vorstellen.
"Man kann", so kommentierte Kiittner brieflich, "doch wirklich mit
guten Griinden bezweifelen, daB eine befriedigende Erkliirnng vorliegt,
wenn das zu erklarende Ereignis nur mit 10% Wahrscheinlichkeit zu er-
warten war, mithin das Nichteintreffen fast sicher gewesen ware! Schurz
konnte das beispielsweise nicht akzeptieren, denn er fordert prognostische
Relevanz fiir erklarende Systematisierungen".
Die raffinierte Kasuistik der verschiedenen Wissenssituationen, die bei
pragmatisch-epistemischen Begriindungen und Erklarungen eine (auch
unterscheidende) Rolle spielen, kann nicht zu einer totalen Kasuistik und
pragmatisch idealisierenden Rekonstruktion aller gegebenen Intuitionen
fiihren, sondern diese Intuitionen werden durch die Modellkonstruktionen
eben selbst auch mitbeeinfluBt bzw. strukturiert oder suggeriert. Manch-
mal wird man bei den Verfeinerungen bis ins 25. Glied an Christian Mor-
genstern erinnert: "Wieviel Engel sitzen konnen auf der Spitze einer Nadel,
wolle dem dein Denken gonnen, Leser sonder Furcht und Tadel" - sprich:
"Wieviel Wissenssituationen sitzen konnen auf der Pointe eines Erkla-
rungsarguments, wolle dem dein Forschen widmen, werter Wissenschafts-
theoretiker, Student (oder Skribent,je nachdem)". Die Gefahr eines Scho-
lastizismus ist gerade auch beim Versuch, einen objektiven pragmatisch-
epistemischen Ansatz zur differenzierenden Kennzeichnung verschieden-
artiger Erklarungstypen unter EinschluB vielfaltiger unterschiedlicher Wis-
senssituationen und deren Dynamik zu entwickeln, nicht von der Hand zu
weisen. Geht es nicht eher darum, treffende Grundideen zu formulleren
und nur so weit zu formalisieren wie unbedingt notig? Doch dies ist ein
weites Feld, das nur beilaufig zum Thema dieses Beitrags gehOrt.
Statt hierauf einzugehen, m6chte ich noch einen weiteren kritischen
Gesichtspunkt einbringen, der ein Grundmotiv fiir die pragmatisch-epi-
stemische Wende darstellt - wenigstens nach Stemiillers Darstellung. Steg-
miiller meint zu Recht, Hempel zitierend, daB die herkommliche Wissen-
schaftstheorie sich zu stark am Modell des theoretischen und beweistheo-
retischen Vorgehens der Metamathematik (im weiteren Sinne) ausgerichtet
und dementsprechend die induktiv-statistischen Begriindungen (das "sta-
470 HANS LENK
tistische ErkUiren") als Erweiterungsfall der rein logisch-syntaktischen

Grundstruktur der deduktiv-nomologischen ErkHirung aufgefaBt habe.
Durch das Auftreten der Mehrdeutigkeit der statistischen Ereigniserkla-
rungen sei es wiederum Hempel gewesen, der mit seiner auf die statistische
Ereigniserklarung beziiglichen "Relativitatsthese die friihere empiristische
Oberzeugung, wonach das induktive Erklarungsmodell eine Verallgemei-
nerung des DN-Modells sei, zerstort" habe (ebd. 1021, u.a.). Hempel habe
damit "ganz entgegen seiner urspriinglichen Intention, selbst den Grund-
stock fUr die pragmatische Wende gelegt" (ebd. 7). Dies ist zweifellos rich-
tig und findet in vielfacher Weise seine Bestatigung in Hempels Schriften
(z.B. 1977, 111 ff., 118 ff.). Hinsichtlich der Inkommensurabilitat der Mo-
delle der induktiv-statistischen und der deduktiv-nomologischen Erkla-
rung ist dies ebenfalls richtig. Die epistemische Relativierung betrifft im
Fall statistischer Ereigniserklarung in der Tat bereits die Grundbegriffe;
und die Analogie zum deduktiv-nomologischen Fall ist nicht sehr iiber-
zeugend, das war sie aber doch wohl nie in vollstandiger Weise (Wahr-
scheinlichkeitsargumente waren nie deduktive Schliisse usw.). Die Frage
ist nur, ob man notwendigerweise, wie Stegmiiller vorschlagt, nun urnge-
kehit den Fall der deduktiv-nomologischen Erklarung als einen bloJ3en
Grenzfall einer Wahrscheinlichkeitserkliirung auffassen soll und das her-
kommliche Modell der logischen Subsurnption zu erklarender Ereignisse
unter Gesetze iiber Bord zu werfen hat. Stegmiiller sieht nur noch die
Moglichkeit, "diese unbefriedigende Situation" der Inkommensurabilitat
dadurch zu iiberwinden, indem er vorschlagt, "die Inkommensurabilitat
der beiden Hempelschen Erkliirungsbegriffe dadurch zu beseitigen, daB
auch der Begriff der D-N-Erklarung als solcher epistemisch relativiert"
und also auch die wahre D-N-Erkliirung abgeschafft wird (1983, 1022).
Man kann jedoch dagegen fragen, wieso diese Situation so "unbefrie-
digend" ist, warum man nun einfach den SpieB urndrehen solI und, wenn
die Induktion nicht zur Deduktion kommen will, eben die Deduktion zur
Induktion preBt. Warum miissen iiberhaupt induktiv-statistische und de-
duktiv-nomologische Erkliirungen noch iiber einen Kamm geschoren wer-
den? Warum muftman die Hempelsche Inkommensurabilitat iiberwinden,
da sie doch recht plausibel funktionierte? Oberwiegt die Ideologie eines
Modellintegrationismus, einer einheitlichen Ausrichtung nahezu urn jeden
Preis, so daB man den bei der statistischen Ereignisbegriindung sicherlich
sinnvollen und wegen der Mehrdeutigkeit notigen Pragmatizismus auf den
deduktiv-nomologischen Fall iibertragen muB, koste es, was es wolle - und

sei es jede Moglichkeit einer "wahren Erklarung?" Hat der Pragmatizis-
mus die Neigung, nur noch pragmatische Analysen zuzulassen, bier nicht
ins Extrem getrieben?
Mir scheint, daB man methodologisch bescheidener operieren konnte
und miiBte und aus den oben erwahnten Griinden wenigstens heuristisch
plausible Unterschiede weiterverfolgen sollte. Man darf nicht voreilig ei-
nem Gesamtintegrationismus im pragmatischen Sinne anheimfallen.
Ais ein seit iiber zehn Jahren fiir pragmatische Philosopbie Pladierender
mochte ich dallir eintreten, daB man nicht uniiberlegt von einem ins andere
Extrem springt und dabei plausible und traditionell bewahrte Unterschei-
dungen (wie die zwischen D-N-Erklarungen und statistischen Ereignis-
erkliirungen, sowie die zwischen Oberzeugungsbegriindungen und Ur-
sachenerklarungen) einfach verwischt.
In einer neueren (i. Dr.) Arbeit hat Kiittner eine immanente sowie eine
nicht-immanente Kritik an der pragmatizistischen Erklarungsidee der
Oberraschungsverminderung entwickelt. Er meint, "daB im Regelfall der
wissenschaftlichen deduktiv-nomologischen Erklarung der Glaubenswert
des Explanandums" (E) durch die Hinzufiigung des Explanans iiberhaupt
nicht beeinfluBt wird, mithin auch nicht erhOht werden kann". Es ist nicht
einzusehen, daB das Hinzufiigen des Explanans "die Oberzeugung einer
Person, daB E wahr ist, im mindesten beeinfluBt, denn diese Oberzeugung
hangt doch allein yom Vertrauen in die E feststellende Verfahrensweise
ab". "Die Idee der Oberraschungsverringerung hat nichts mit der Idee der
qeduktiven Erklarung als wissenschaftlicher Systematisierung zu tun".
Kiittner meint, Stegmiiller und Giirdenfors spezifizierten mit ihrem Begriff
der Oberraschungsverminderung eher "so etwas wie eine Aufk1iirung iiber
die einer bestimmten Person Po zu einer bestimmten Zeit To nicht verfiig-
baren Griinde fUr das Auftreten eines diese Person iiberraschenden Ereig-
nisses". Wissenschaftliche Erklarungen ~ gerade auch in den traditionellen
Auffassungen - wiirden weniger Oberraschung als eher ein Subsumierbar-
keitsdefizit abbauen. Sie "erhOhen nicht den Grad des Glaubens an das
Explanandum, sondem sind potentiell geeignet, den Glauben oder das Ver-
trauen an bzw. in die Theorie zu erhOhen". Oberraschung in wissenschaft-
lichen Zusammenhangen tritt eher auf, "wenn Sachverhalte, die in den
potentiellen Subsumtionsbereich einer Theorie fallen, tatsachlich nicht un-
ter sie subsumiert werden konnen". Kiittner bezieht demgegeniiber den
472 HANS LENK
"eigentlich pragmatischen Aspekt des Erklarens" auf die Angemessen-

heitsproblematik, die eine notwendig idealisierende Theorie bei der An-
wendung auf reale Gegenstiinde aufwirft: "Auf der Basis eines Hinter-
grundwissens urn die bisherigen erfolgreichen wie erfolglosen Anwendun-
gen einer Theorie aus dem vorliegenden Anwendungsfall" sei "der Glaube"
an deren "Angemessenheit zu begriinden, eventuell zu rechtfertigen".
Pragmatische Aspekte stellen sich also deutlich genug im Zusammen-
hang mit der Anwendung wissenschaftlicher Theorien, man braucht des-
halb nicht pragmatische Deutungen sogleich in aUe Grundbegriffe einzu-
bringen, vor allem nicht dort, wo dies nicht unbedingt notig ist. Die Idea-
lisierungsproblematik steUt sich n8mlich in der Wissenschaftstheorie, auch
in der Wissenschaftstheorie der Erklarungsargumente, ahnlich wie in den
Fachwissenschaften selbst. Auch dort wird nicht gleich alles epistemisiert
und pragmatisiert. Die quantentheoretische Diskussion diirfte eher in urn-
gekehrte Richtung weisen, indem epistemisch relativierte Argumente als
"echte .Erkliirungen" konstruiert werden.
Insgesamt ist festzustellen, daB die Beriicksichtigung pragmatischer und
epistemischer Aspekte wertvoll und wichtig, in vielen Zusammenhangen
(etwa beim statistischen Erklaren) unverzichtbar ist, daB aber eine Prag-
matisierung von allem und jedem nicht zweckmaBig ware und wichtige
Unterschiede verwischen diirfte und letztlich wohl zu einem pragmatisti-
schen erkenntnistheoretischen, wenn nicht gar ontologischen Idealismus
fiihren konnte. Dies aber haben die pragmatischen Epistemiker vielleicht
wohl doch nicht im Sinn.
LITERATUR
Girdenfors, P.: 1980, 'A Pragmatic Approach to Explanations', Philosophy of Science 47,
405-423.
Hempel, C.G.: 1977, Aspekte wissenschaftlicher Erkliirung, de Gruyter, Berlin/New York
(Original: Aspects of Scientific Explanation, The Free Press, New York 1965).
Kuttner, M.: 1983, 'G1aube, Wissen und die Pragmatik des Erklirens: Zur Kritik des An-
satzes von Girdenfors und Stegmiiller'. 1m Druck in: KongrefJakten des VllIten Interna-
tionalen Wittgensteinsymposions, Kirchberg, Ho1der-Pichier-Tempsky, Vienna.
Lenk, H.: 1972, Erkliirung - Prognose - Planung. Skizzen zu Brennpunktproblemen der Wis-
senschaftstheorie, Rombach, Freiburg.
Lenk, H.: 1975, Pragmatische Philosophie. Pliidoyers und Beispiele fUr eine praxisnahe Phi-
losophie WId Wissenschaftstheorie, Hoffmann und Campe, Hamburg.
Schurz, G.: 1983, Wissenschaftliche Erkliirung. Ansiitze zu einer logisch-pragmatischen Wis-

senschaftstheorie, Diss. Graz.
Stegmiiller, w.: 1983 Erkliirung - Begriindung - Kausalitiit, Springer, BerlinfHeidelbergfNew
York (zweite, erweiterte Aufiage von: Wissenschaftliche Erkliirung und Begriindung. Bd. I
von: Prob/eme und Remltate rkr Wissenschaftstheorie und Analytischen Philosophie, ebd.
1969).
MICHAEL KUTTNER
ZUR VERTEIDIGUNG EINIGER HEMPELSCHER THESEN

GEGEN KRITIKEN STEGMULLERS*
1. VORBEMERKUNGEN
Wenn ein Jubilar einen so auBerordentlich tiefen und nachhaltigen EinfluB

auf die modeme Analytische Philosophie genommen hat, wie das auf Carl
G. Hempel zweife1sfrei zutrifft, und einige der zum Teil vierzig Jahre zu-
riickreichenden Grundansichten des zu Ehrenden gerade jiingst durch
Uberlegungen Wolfgang Stegmiillers auf massive Kritik gestoBen sind,
dimn ist es sicherlich von betrachtlichem Interesse, die Haltbarkeit der
Kritiken zu priifen. 1m folgenden werde ich versuchen, einige der Hem-
pelschen Grundansichten zum Problemkreis der wissenschaftlichen Erkla-
rung und Vorhersage gegen die entsprechenden Kritiken Stegmiillers zu
verteidigen.
2. DIE ZU VERTEIDIGENDEN HEMPELSCHEN THESEN
Hempels Grundauffassungen zum genannten Problemkreis sind allseits

bekannt. Es geniigt daher, sie in (knappgehaltener) Thesenform wie folgt
wiederzugeben:
(I) Deduktiv-nomologische Erklarungen und Vorhersagen singularer
Ereignisse lassen sich rein syntaktisch-semantisch spezifizieren.
(II) Fiir erklarende und vorhersagende deduktiv-nomologische Argu-
mente laBt sich mindestens eine Teilthese der Strukturgleichheits-
these halten.
(III) Die Mehrdeutigkeit bei probabilistischen Einzelfallerklarungen
(sog. IS-Erklarungen) ist ein echtes Problem.
(IV) Probabilistische Ereigniserklarungen miissen auf genau eine Wis-
senssituation epistemisch relativiert werden.
(V) Probabilistische Ereigniserklarungen benotigen statistische Gesetze
mit hoher Wahrscheinlichkeit.
Hempel hat die ihm hier zugeschriebenen Thesen bekanntlich in meh-
reren Schriften deutlich differenzierter dargelegt. Fiir die weiteren Eror-
Erkenntnis 22 (1985) 475-484. 0165-m06j85j022<H>475 $01.00

476 MICHAEL KUTTNER
terungen ist das aber ohne Belang, denn schon diese sehr vorsichtigen und
liberalen Formulierungen will Stegmiiller so nicht akzeptieren. Beziiglich
der These (V) ist noch anzumerken, daB sie von Hempel als Reaktion auf
Kritiken Salnrons und Jeffreys seit langerem nicht mehr aufrecht erhalten
wird. Hempels Preisgabe wird als voreilig charakterisiert werden miissen.
3. DIE EINWANDE STEGMULLERS
Die genannten Hempelschen Thesen sind in der Vergangenheit jeweils

mehrfach von verschiedenen Autoren detaillierter Kritik unterzogen wor-
den. Wenn im folgenden gerade Stegmiillers Formulierungen herangezo-
gen werden, so hat das zwei Vorteile. Erstens hat Stegmiiller nahezu die
gesamte relevante Literatur hierzu aufgearbeitet. Zweitens fallen seine
Oberlegungen besonders pointiert aus. Sie lassen sich (wiederum sehr kom-
pakt) so darstellen:
Zu (I): Aufgrund des Beispielmusters von U. Blau (lind ahnlicher Sy-
stematisierungen) kann es keinen anomalienfreien Begriff der deduktiv-
nomologischen Einzelfallerklarung ohne wesentliche Bezugnahme auf
pragmatische Aspekte geben. (1983, S. 919 f.; so auch schon 1969)
Zu (II): Ob die Strukturgleichheitsthese oder eine ihrer Teilthesen zu
halten ist, hangt von Konventionen und von einer Entscheidung iiber den
Gebrauch der Bezeichnung "Vorhersage" abo (1983, S. 983)
Zu (III): Das Mehrdeutigkeitsproblem verschwindet automatisch, wenn
man weiB, daB das zu erklarende Ereignis tatsachlich vorliegt. Trifft man
diese Voraussetzung jedoch nicht, ist man gezwungen, von einer Ereignis-
erklarung auch dann zu sprechen, wenn das betreffende Ereignis iiber-
haupt nicht stattgefunden hat. (1983, S. 959)
Zu (IV): Probabilistische Ereigniserklarungen miissen auf genau drei
Wissenssituationen relativiert werden. (1983, S. 959 u. spater)
Zu (V): Probabilistische Ereigniserklarungen konnen auch in einem sog.
epistemischen Minimalsinn mit Wahrscheinlichkeiten kaum groBer null
gegeben werden. (1983, S. 972 u. spater)
4. ZURUCKWEISUNG DER EINW ANDE GEGEN THESE (I)
Alle diskutierten angeblichen Anomalien des DN-Erklarungsbegriffs las-

sen sich zuriickweisen, dajeweils ein VerstoB gegen mindestens eine Ad-
EINIGER HEMPELSCHER THE SEN 477
aquatheitsbedingung vorliegt. Mithin gibt es bis jetzt iiberhaupt keine

Gegeninstanz in Form eines Beispielmusters gegen die Idee der rein syn-
taktisch-semantisch spezifizierten DN-Erklarung (resp. -Vorhersage).
Das Blau-Beispiel ist, welche Deutung der Pradikate auch jeweils ge-
wahlt wird, immer total selbsterklarend:
Wegen
I\x(Fx-+, Gx) I\x(Fx-+, Gx)
Fa/\(Ga V Ha) (Fa 1\ Ga) V (Fa 1\ Ha)
- - - - - L-aqu.
Ha Ha
und
Fa 1\ Ga ist gesetzesunvertraglich
folgt: Die Wahrheit (oder: Akzeptanz) des Antezedens beruht allein auf
der Wahrheit (oder: Akzeptanz) des Teilsatzes Fa 1\ Ha. Aus Fa 1\ Ha
folgt aber das Explanandum. Wer diesen strengen Begriff der totalen
Selbsterklarung nicht akzeptieren mochte, muB immerhin zugeben, daB
dann in jedem Fall partielle Selbsterklarung vorliegt.
Auf die Unbrauchbarkeit des Blau-Beispiels fUr Zwecke des Abgehen-
miissens von rein syntaktisch-semantischen Spezifikationen des Erkla-
rungsbegriffs hatte schon Lenk (1972) hingewiesen. Dessen Dberlegungen,
wie auch die in Kiittner (1976), sind allerdings von Stegmillier nicht beach-
tet worden.
Stegmiiller Mtte andererseits auffallen miissen, daB gerade der von ihm
hinsichtlich der syntaktischen und semantischen Bedingungen favorisierte
Erkiarungsbegriff von Gardenfors in der von Stegmiiller selbst gegebenen
Form (1983, S. 946 ff.) durch die Bedingung
C5': Wenn {T,A} ein Explanans fUr E bildet, dann gibt es weder
Pradikate in E, die nicht in T vorkommen, noch Pradikate in
A, die nicht in T vorkommen
das Blau-Beispiel ausschlieBt, denn das einzige Pradikat H in E kommt ja

nicht im Gesetz vor.
478 MICHAEL KUTTNER
Das Schurz-Beispielspaar (Schurz 1983, S. 279)
Ax (Bx -+ (Lx ~ Fx» Ax (Bx -+ (Lx ~ Fx»

Ba A La Ba A Fa
Fa La
ist ebenfalls nicht geeignet, die angebliche Pragmatik oder (wie Schurz
einsichtiger formuliert:) Kontextabhiingigkeit zu stiitzen.
Fiir B . .. Pendel unter Normalbedingungen; L. .. Pendelhinge; F. .. Pen-
delfrequenz sei nach Schurz die rechte Systematisierung eine bloBe Be-
griindung, die linke hingegen eine Ursachenerklarung. Das mag wohl sein,
zeigt dann aber, daB das verwendete Kausalgesetz unzulassig stark for-
muliert ist. Als Kausalgesetz kommt dann nur die logisch schwachere
Form Ax(Bx-+(Lx-+Fx» in Betracht.
Wer an einem engen Begriff der Ursachenerklarung interessiert ist, muB
in einer Bedingung gewahrleisten, daB auch nUT Kausalgesetze fiir Erkla-
rungszwecke herangezogen werden, nicht aber irrelevante Verstiirkungen.
Wer in einem weiteren Begriff auch nicht-kausale (etwa Indikator- oder
Symptom-) Erklarungen zulaBt, akzeptiert auch das rechte Beispiel als
Erklarung. Von Pragmatik kann jedenfalls keine Rede sein.
Die angebliche Anomalie des John-Jones-Schwangerschaftsbeispiels ist
ebenfalls keine Gegeninstanz. Entweder wird innerhalb der theoretischen
Medizin das Gesetz "Wenn immer jemand die Pille nimmt, dann wird er
nicht schwanger" akzeptiert, dann handelt es sich zu recht urn eine kor-
rekte Erklarung; oder das Gesetz lautet "Wenn immer eine weibliche Per-
son die Pille nimmt, dann wird sie nicht schwanger", mit der Folge, daB
nicht einmal eine korrekte Deduktion auf John Jones' Nichtschwanger-
schaft vorliegt.
5. ZURUCKWEISUNG DER EINWANDE GEGEN THESE (II)
Eigentlich hiitte Stegmiiller ja behaupten miissen, daB die Strukturgleich-

heitsthese, mangels scharfer Bestimmbarkeit des Erklarungsbegriffes,
iiberhaupt nicht entschieden werden kann.
Hinge die Frage der Strukturgleichheit nur von Konventionen ab, ware
sie trivial. Es solI zunachst gezeigt werden, daB dies nicht der Fall ist.
Gleichgiiltig, llir welches Tripel von Konventionen (Stegmiiller 1983; S.

225, S. 228, S. 236) man sich entscheidet, die Strukturgleichheitsthese gilt
nicht:
/\x(Px-+Qx)
Pa A Qa
Qa
ist immer eine adaquate ErkUi.rung hingegen keine adaquate Vorhersage

(Kiittner 1983).
1m Erklarungsfall ist namlich zunachst das Explanandum gegeben. Die
Pramissenmenge wird nachtraglich hinzugezogen. Es ist dann unerheblich,
daB hierbei ein singulares Antezedens auftritt, welches den Explanandum-
Satz Qa iiberfliissigerweise noch einmal enthalt. {/\x(Px-+Qx), Pa A Qa}
ist ein korrektes Explanans fUr Qa, weil schon {/\x(Px-+Qx), Pal mit
einem konjunktionsabgeschwachten Antezedens korrekt ist. 1m Vorher-
sagefall bemerkt man jedoch, daB aus der Pramissenmenge nichts dedu-
zierbar ist, was man nicht schon durch das gegebene Antezedens weiB. Ein
prognostischer SchluB auf Qa kame dann einer totalen Selbstvorhersage
gleich. Die von Hempel favorisierte Teilthese, derzufolge jede korrekte
Erklarung bei Anderung der pragmatischen Umstande des Gegebenseins
eine korrekte Vorhersage ist, ist mithin unhaltbar; Stegmiillers Idee (S.
237), die Strukturgleichheitsthese konne endgiiltig verteidigt werden,
ebenso.
Bei geeigneter Wahl der Adaquatheitsbedingungen ist aber die umge-
kehrte Teilthese in einer der folgenden Formulierungen haltbar.
(i) Jede kausale Vorhersage ist ... eine kausale Erklarung.
(ii) Jede deduktive Vorhersage ist ... eine nicht notwendigerweise kausa-
Ie Erklarung.
(iii) Jede probabilistische Vorhersage ist ... eine probabilistische Erkla-
rung.
Kontrovers diirfte allenfalls (iii) sein. Hier hangt die Entscheidung tat-
sachlich davon ab, wie der Begriff "Vorhersage" gewahlt wird (Stegmiiller
1983; S. 983).
480 MICHAEL KUTTNER
6. ZURUCKWEISUNG DER EINWANDE GEGEN THESE (III)
Stegmiiller argurnentiert wie folgt: Hempel habe die epistemische Relati-

vierung der probabilistischen EreigniserkUirung falschlicherweise nur be-
ziiglich einer Wissenssituation vorgenommen. Hierdurch gerate er in das
folgende Dilemma: Entweder gehOrt das Explanandurn als Fakturn in die
Wissenssituation. Dann existiert keine Mehrdeutigkeit, da das Nichtein-
treten des Explanandurn nicht gleichzeitig Bestandteil dieser Wissenssitua-
tion sein kann. Oder das Explanandum gehort der Wissenssituation nicht
an. Dann gibt es probabilistische EreigniserkHirungen von Explananda,
die iiberhaupt nicht eingetroffen sind.
Geht man davon aus, daB im ErkHi.rungsfall das Explanandurn stets
gegeben ist, scheint das Hempelsche Mehrdeutigkeitsproblem nicht zu exi-
stieren.
Zunachst ist die Stegmiillersche Argumentation iiberhaupt etwas be-
fremdlich. Hempel hatte ja die epistemische Relativierung vorgenommen,
urn mit ihr das Mehrdeutigkeitsproblem zu beseitigen, wenigstens zu ent-
scharfen. Sollte dies gelungen sein, kann das Hempel doch nicht zum Vor-
wurf gemacht werden.
Nun hat Hempel aber das Mehrdeutigkeitsproblem zum AniaB genom-
men, einen fundamentalen, unbeseitigbaren Unterschied zwischen deduk-
tiven und probabilistischen Erklarungen zu behaupten. (Friiher hatte er
eher die Gemeinsamkeiten betont.) Fiir diesen Unterschied ist es relevant,
daB auch nach Einfiihrung einer (oder mehrerer) Wissenssituation(en) die
Mehrdeutigkeit bestehen bleibt. Auf einfache Weise kann man zeigen, daB
dies der Fall ist: Das Mehrdeutigkeitsproblem darf nicht so verstanden
werden, daB es zu jeder probabilistischen Ereigniserklarung eine konkur-
rierende gibt (bzw. geben kann), deren Pramissenmenge ebenfalls akzep-
tiert ist, die aber das Nicht-Bestehen des Ereignisses erklart.
Das Mehrdeutigkeitsproblem sollte m.E. statt dessen wie folgt formu-
liert werden: Das Vertrauen in eine probabilistische Erklarung eines Ereig-
nisses ist prinzipiell (d.h. unbeseitigbar) unvollkommen, weil es immer
moglich ist, daB auch die Negation des Ereignisses, ware sie aufgetreten,
durch eine ebenfalls akzeptierbare Pramissenmenge mit mindestens gleich
hoher Sicherheit hiitte erkliirt werden konnen.
Bei dieser Fassung schwindet der Einwand Stegmiillers vollig. Zugleich
ist auch klar, daB das Problem nicht dadurch gelost werden kann, indem
man mehr als eine Wissenssituation heranzieht. Insbesondere ist die Steg-
miillersche Oberlegung "Das Problem kann nur solange Bestand haben,
als ErkHirungen von miteinander unvertraglichen Satzen in Betracht ge-
zogen werden ... Das Hempelsche Problem verschwindet automatisch, so-
bald das Explanandurn als gewuBt vorausgesetzt wird ... " (1983; S. 985)
unter der Reformulierung des Mehrdeutigkeitsproblems nicht zutreffend.
Was bleibt ist: Probabilistische EreigniserkUirungen sind Verallgemei-
nerungen (im Sinne von Lockerungen) deduktiver ErkHirungen urn den
Preis eines durch die Mehrdeutigkeit eingehandelten Vertrauensschwun-
des.
7. ZURUCKWEISUNG DER EINWANDE GEGEN THESE (IV)
Eine wichtige Motivation, probabilistische Erklirungen auf mehr als eine

Wissenssituation zu beziehen, liegt fiir Stegmiiller in der Mehrdeutigkeits-
problematik. Akzeptiert man die eben gegebene Reformulierung, entfiiIlt
diese Motivation. Dariiber hinaus glaubt Stegmiiller, daB die "Oberlegun-
gen von Gardenfors (1980) hinsichtlich der mit Begriffen der informativen
Erklirung verkniipften Oberraschungsverminderung beim Erklarunghei-
schenden eine epistemische Relativierung auf drei Wissenssituationen
zwangslaufig mit sich bringen.
Nun ist aber der Aufbau der drei Wissenssituationen in hohem MaBe
kontraintuitiv (ausfiihrlich in Kiittner 1983b): Entweder nimmt man hin,
daB nur eine der Wissenssituationen real ist, die beiden anderen hingegen
fiktiv sind, oder der Erklarungsbegriff wird vallig trivial. Man mag sich
vorsteIlen, daB fiir eine gewisse Spielart der Begriffsfamilie "Erklarung"
ein Bezug auf fiktive Wissenssituationen durchaus angemessen ist. Ohne
Not wird man solches aber nicht fiir wissenschaftliche Erklirungen ver-
langen wollen. Folgerichtig hat dann auch Gardenfors (kiirzlich in einer
Diskussion) zugestanden, daB mit seinem Apparat kein Begriff der wis-
senschaftlichen, sondem eine Art von Alltagserklarung (genauer wohl: ein
Begriff der Oberraschungsverminderung in Alltagssituationen) spezifiziert
wird. Dies zeigen im iibrigen auch deutlich die von Gardenfors und Steg-
miiller gewahlten Illustrationsbeispiele.
Kritiker an Hempels epistemischer Bezugnahme auf nur eine Wissens-
situation im FaIle der wissenschaftlichen probabilistischen Ereigniserkla-
rung verfligen also bislang noch nicht iiber einen schlagkraftigen Beleg.
482 MICHAEL KUTTNER
8. ZURUCKWEISUNG DER EINWANDE GEGEN THESE (V)
Stegmiiller schlieBt sich hier an eine Oberlegung Jeffreys an, die auch Hem-
pel (1977; S. 100 u. spater) schlieBlich akzeptiert: Wie ein Ereignis zustande
kam, ist durch Hinweis auf den ZufallsprozeB, fiir den das in der Analyse
angefiihrte st~tistische Gesetz gilt, zu beantworten. Warum ein Ereignis
eintrat, ist durch Hinweis auf den Zufall zu beantworten.
Diese Auffassung, die ein Resultat der Intention ist, auch Ereignisse mit
niedriger Auftretenswahrscheinlichkeit als erklarbar anzusehen, ist m.E.
falsch (bzw.: unzweckmaBig). Es handelt sich namlich tatsachlich nur urn
eine, angesichts der Unbehebbarkeit der Mehrdeutigkeit, l'esignative Ver-
zichtslosung.
Man macht sich das am besten am Vorhersageproblem klar. Wenn eine
Pramissenmenge ein statistisches (besser: nichtdeterministisches) Gesetz
mit niedriger Wahrscheinlichkeit enthalt, dann ist keine Vorhersage des
Ereignisses moglich, in dem Sinne, daB das Ereignis rational zu erwarten
ist (so auch bei Schurz (1983», sondem nur die AuBerung, das Ereignis
weise eine entsprechend niedrige Eintreffenswahrscheinlichkeit auf. Wenn
eine Pramissenmenge ein statistisches Gesetz mit hoher Wahrscheinlichkeit
enthalt, dann ist das Ereignis vorhersagbar im Sinne des rationalen Er-
wartens.
Diesen Unterschied sollte man sich zunutze machen und von probabi-
listischer Ereignisvorhersage nur dann sprechen, wenn die Eintreffens-
wahrscheinlichkeit nahe eins ist, mindestens aber die Leibniz-Bedingung
erfiillt. Wenn hingegen nicht einmal die Leibniz-Bedingung erfiillt ist, dann
ist keine statistische Ereignisvorhersage moglich, sondem nur die nicht-
deduktive Ableitung mindestens eines Paares von moglichem Ereignis und
Eintreffenswahrscheinlichkeit. So liefert beispielsweise das (bevorstehende)
Werfen eines idealen Wiirfels keine Vorhersage irgendeines Augenzahl-
Resultates, sondem nur Paare (I; 1/6), (2; 1/6), ... Umgekehrt ist der Ver-
weis auf den idealen Wiirfel keine Erklarung dafiir, daB nach einem Wer-
fen des Wiirfels etwa die Augenzahl 3 aufgetreten ist.
Der durch Jeffrey nahegelegte Bezug auf den Zufall hat mit dem Pro-
blem nichts zu tun: Warum (urn ein bekanntes Beispiel aufzugreifen) Peter
Peterson katholisch und nicht protestantisch ist, hat nichts mit dem Zufall
zu tun. Das probabilistische "Gesetz", demzufolge fast alle Schweden
protestantisch sind, ist nicht ein Resultat eines Zufallsprozesses. (Mogli-
EINIGER HEMPELSCHER THESEN 483
cherweise fehlen hierin nur ein paar weitere Wenn-Komponentenbestand-

teile zur strikten Hypothese.) Die Jetrrey-Oberlegungen lassen sich sicher-
lich besser deuten als ErschlieBen der Wahrscheinlichkeit eines Ereignisses,
nicht als Vorhersage oder Erklarung des Ereignisses.
Stegmiiller insistiert nun noch auf dem Vorzug der Behandlung des
Problems mit Hilfe der drei Wissenssituationen anhand des Scriven-Bei-
spiels von der progressiven Paralyse Nietzsches. Aber auch dieses Beispiel
hilft nicht, den Ansatz zu retten. Es ist niimlich nur unzweckmaBigerweise
als probabilistische Systematisierung ausgewiesen worden. Da otrenbar
gilt, daB jeder, der progressive Paralyse bekam, vorher unbehandelte Sy-
philis gehabt haben muBte, ist die wegen zu niedriger Wahrscheinlichkeit
unbeantwortbare erklarungheischende Frage "Warum hat a progressive
Paralyse bekommen?" durch eine Wie~war~es~moglich~Frage zu ersetzen.
Die Antwort auf diese Frage kann sich auf das entsprechende determini-
stische Retrodiktionsgesetz stiitzen.
SUMMARY. The aim of this paper is to defend some ofe. G. Hempel's basic theses con-
cerning the logic of explanation and prediction against criticisms recently made by W. Steg-
millier. It is argued (very concisely) that
(I) there is no need for essentially pragmatic conditions in DN-arguments;
(II) only the structural identity sub-thesis "Every adequate prediction is ... an
adequate explanation" can be held instead of the one Hempel has in mind;
(III) the notion of the ambiguity of probabilistic explanations should be refor-
mulated;
(IV) there is no need for more than one knowledge situation;
(V) systematizations with low inductive probability do not give rational predic-
tions or explanations of singular events.
ANMERKUNG
.. Carl G. Hempel danke ich sehr fiir die kiirzliche Diskussion einiger Grundgedanken zu
den Abschnitten 4 und 7. Fehler und MiBverstindnisse gehen aber allein zu meinen Lasten.
LITERATUR
Gardenfors, P.: 1980, A Pragmatic Approach to Explanations, Philosophy of Science 47,

405-423.
Hempel, C. G.: 1977, Aspekte wissenschaftlicher Erkliirung, De Gruyter, Berlin, New York.
Kiittner, M.: 1976, Ein verbesserter deduktiv-nomologischer Erklarungsbegriff, Zeitschrift
flir allgemeine Wissenschaftstheorie 7 274-297.
Kuttner, M.: 1983, Logical Conditions of D-N Predictions, Structural Identity Thesis, and
Blau's Example, 7th International Congress of Logic, Methodology and Philosophy of
Science, Abstracts Section 6, Salzburg, pp. 118-121.
484 MICHAEIJ KUTTNER
Kiittner, M.: 1983, Glaube, Wissen Und die Pragmatik des Erkliirens: Zur Kritik des Ansatzes
von Giirdenfors und Stegmiiller, Proceedings of the 8th International Wittgenstein Sym-
posium, D. Reidel, Dordrecht (erscheint demnichst).
Lenk, H.: 1972, Erkltirung - Prognose - Planung, Rombach, Freiburg.
Schurz, G.: 1983, Wissenschaftliche Erklarung, Diss., Graz.
Stegmiiller, W.: 1983, Erklarung - Begriindung - Kausalittit, Springer, Berlin, Heidelberg,
New York.
Manuscript received 16 July 1984
Freie Universitit Berlin

Fachbereich Wirtschaftswissenschaft
Fachrichtung Wissenschaftstheorie
Garystrasse 6, D-looo Berlin 33
B.R.D.

Epistemology, Methodology, and Philosophy of Science

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Epistemology, Methodology, and Philosophy of Science

Uploaded by

Copyright:

Available Formats

EPISTEMOLOGY,

Essays in Honour of Carl G. Hempel

SPRlNGER-SClENCE+BUSINESS MEDIA, B.V.

Ali Rights Reserved

EPISTEMOLOGY, METHODOLOGY, AND

Essays in Honour of Carl G. Hempel on the Occasion

Professor C.G. Hempel (known to a host of admirers and friends as 'Peter'

terial 'Problems and Changes in the Empiricist Criterion of Significance'.

With love and admiration,

University of Frankfurt, Frankfurt am Main w. K. ESSLER

ON INDUCTIVE SUPPORT AND SOME RECENT TRICKS*

Inductive inference is the inference from particular cases to a general law,

Erkenntnis 22 (1985) 5-21 0165-{)106/85/0220-0005 $01.70

conscious expectations. It took a Humean analysis to trace back our basic

By straightforward induction we are lead to conclude that the oscillations

Pattern recognition, reasoning by analogy and inductive inference be-

sin x/x = n00 (

What distinguishes mathematics from empirical science is the availabil-

randomness of a data sequence obtained by tossing a fair coin. It signifies

An essential aspect of induction is that it provides the hypothesis in ques-

then the total win or loss is ::; S· L (1 - Pn) + Sand P i= 0 implies

The significance of simple inductive generalizations, such as "all ravens

The example from Newton's Principia is most striking, as it illustrates

Current statistical practice is a good source of illustrations of the in-

4. INDUCTIVE SUPPORT AS PROBABILITY INCREASE

In the course of induction new confirming evidence makes the hypothesis

do so let "x is P" be equivalent, by definition, to "x is the planet Jupiter

A large part of the Popper-Camap controversy of the fifty's was due to

Recently Popper and Miller have proposed a surprising proof which

and hI and h2 are probabilistically independent:

Moreover hI is the strongest statement which is logically implied either by

that b 2 = .J2g/d. Thus b 2 decreases with d. The increase of v with dis

• Research supported by a DFG grant.

Manuscript received 25 June 1984

Institute of Mathematics and Computer Science

INDUCTIVE INFERENCE IN THE LIMIT

There is an enormous and interesting theoretical literature 1 on inductive

Erkenntnis 22 (1985) 23-31. 0165-0106/85/0220-0023 $00.90

determining features of various satellite systems, or of the galaxy, or of

DEFINITION: An evidence frame for a theory is a countable collection

Some theories, of course, have only a countable number of elementarily

order language L, and add to L a denumerably infinite collection of in-

DEFINITION: The diagram, D(M), of a model M is the set of all quan-

We can recursively assign a natural number, a Godel number for ex-

DEFINITION: A presentation of a frame for T is a sequence S or ordered

DEFINITION: A machine, R, for a first order language L is a partial

Hempel's theory of confirmation provides the basis for an inductive

where i ranges over indices occurring in Sn.

The differences among these definitions can be illustrated by considering

of the extension of some relations may be deleted, so that every presen-

or on the theories themselves, save that they be first order. Satisfaction

a single expansion of the natural numbers. The important difference is that

Received 4 July 1984

2510 Cathedral of Learning

PROBABILITY AND LA WS*

tended to deal with. However they have in mind physical theories as is

After having recalled the paradoxes of Hempel and Goodman, Toraldo

families and considering an infinite number of attributes gives rise to two

4. In Hempel's opinion, his paradox is due to misguided intuition rather

iasson-Lindenbaum, we suppose that it is useless to observe butterflies in

(2) .n -,Qai) n (. n Qai)

in which Ne is the set of exceptions. When we cannot or do not want to

(3) (Z : (Z)r~ = i/N), i E ~N