Professional Documents
Culture Documents
Epistemology, Methodology, and Philosophy of Science
Epistemology, Methodology, and Philosophy of Science
METHODOLOGY, AND
PHILOSOPHY OF SCIENCE
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
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
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
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
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
2. PLAUSIBILITY
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
ON INDUCTIVE SUPPORT 11
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
3. SIMPLE INDUCTION
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
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
NOTES
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
ON INDUCTIVE SUPPORT 21
elementary algebra: p(h l ) . P(h2) - P(hl /\ h2) = (Xl - Y2) (X2 - Y2). This value is > 0,
unless Xl = YI or X2 = Y2.
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:
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.
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
INDUCTIVE INFERENCE 29
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.
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
© 1985 by D. Reidel Publishing Company
34 D. COSTANTINI
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
PROBABILITY AND LAWS 37
(1)
This is not the only type of deterministic law we can consider. Propositions
of the following type are also deterministic laws
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
f
+00
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.
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,
°
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
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.
where
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
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
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.
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-
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.
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-
SCIENTIFIC KNOWLEDGE 55
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.
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-
SCIENTIFIC KNOWLEDGE 57
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.
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
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.
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.
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
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
SCIENTIFIC KNOWLEDGE 67
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
SCIENTIFIC KNOWLEDGE 69
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.).
DIAGRAMS
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'.
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
SCIENTIFIC KNOWLEDGE 75
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.
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
FINAL REMARK
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.
Philosophisches Seminar
Universitiit Fribourg
CH-1700 Fribourg
Switzerland
EIKE VON SAVIGNY
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
2. NO RIGID PROBLEM-SOLVING
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
3. KNOWLEDGE OF CONVENTIONS
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
88 EIKE VON SA VIGNY
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
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
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
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
94 EIKE VON SA VIGNY
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.)
96 EIKE VON SA VIGNY
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).
31 Convention, p. 155.
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.
33 Convention, p. 155.
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.
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
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.
THEORETICAL TERMS 101
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
104 c. ULISES MOULINES
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-
THEORETICAL TERMS 105
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
106 c. ULISES MOULINES
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
108 c. ULISES MOULINES
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
THEORETICAL TERMS 109
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).
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
THEORETICAL TERMS 111
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
112 c. ULISES MOULINES
NOTES
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.
- 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.
THEORETICAL TERMS 117
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.
ON REDUCTION OF THEORIES
1. INTRODUCTION
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)
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.
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
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
ON REDUCTION OF THEORIES 125
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:
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.
1. X = <Mp , N, D)
2. N, 0 have the following types
N ~ pot (Mp); D ~ Mp x Mp
ON REDlJCTION OF THEORIES 131
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
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
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.
ON REDUCTION OF THEORIES 135
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)
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
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.
REFLEXIVE REFLECTIONS
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
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.
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").
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
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
REFLEXIVE REFLECTIONS 153
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!
Opt. of Philosophy
Harvard University
Emerson Hall
Cambridge, MA 02138
U.S.A.
ANNETTE BAIER
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
EXPLAINING THE EXPLAINERS 159
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
EXPLAINING THE EXPLAINERS 161
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
EXPLAINING THE EXPLAINERS 163
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
EXPLAINING THE EXPLAINERS 165
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
EXPLAINING THE EXPLAINERS 167
NOTE
* I am grateful for helpful comments on earlier drafts of this paper from Kurt Baier, Nicholas
Rescher, Arthur Ripstein and Patrick Maher.
University of Pittsburgh
Dept. of Philosophy
Pittsburgh, PA 15260
U.S.A.
NORETT A KOER TGE
ON EXPLAINING BELIEFS·
INTRODUCTION
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.
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?
Transition
System in System in
State S1 State S2
Principle
System in
equilibrium; Equilibrium Value
Values of n-l of nth
variables given Principle variable
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-
178 NORETT A KOER TGE
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.
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
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:
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
ON EXPLAINING BELIEFS 181
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.
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' ...
ON EXPLAINING BELIEFS 183
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?
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
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.
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)
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
EXPLAINING THE UNPREDICTABLE 191
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
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
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,
EXPLAINING THE UNPREDICTABLE 195
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-
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
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.
Ventura Hall
Stanford University
Stanford, CA 94305
U.S.A.
RICHARD N. BOYD
INTRODUCTION
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.
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
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
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.
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
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.
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
THE LOGICIAN'S DILEMMA 221
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
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
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-
THE LOGICIAN'S DILEMMA 229
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
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
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
THE LOGICIAN'S DILEMMA 237
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
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 LOGICIAN'S DILEMMA 247
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
the inductive and dialectical reasoning at the theoretical level which un-
derlies it, which the consistent empiricist cannot explain.
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
INTRODUCTION
I. THEORETICAL COSMOLOGY
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.
256 B. KANITSCHEIDER
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
EXPLANA TION IN PHYSICAL COSMOLOGY 257
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.
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
EXPLANA TION IN PHYSICAL COSMOLOGY 261
NOTES
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.
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
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
Received 2 Apri11984
Dept. of Philosophy,
Harvard University,
Emerson Hall
Cambridge, MA 02138
U.S.A.
RAIMO TUOMELA
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
© 1985 by D. Reidel Publishing Company
272 RAIMO TUOMELA
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
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):
TRUTH AND BEST EXPLANATION 275
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
TRUTH AND BEST EXPLANATION 279
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.
TRUTH AND BEST EXPLANATION 281
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.
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
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-
TRUTH AND BEST EXPLANATION 291
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
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.
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:
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
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)
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
TRUTH AND BEST EXPLANATION 299
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
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
© 1985 by D. Reidel Publishing Company
302 RAINER W. TRAPP
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
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
'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
II
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
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:
UTILITY THEORY AND PREFERENCE LOGIC 317
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.
UTILITY THEORY AND PREFERENCE LOGIC 319
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
UTILITY THEORY AND PREFERENCE LOGIC 321
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
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
UTILITY THEORY AND PREFERENCE LOGIC 323
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
UTILITY THEORY AND PREFERENCE LOGIC 327
comes out true under UTS 1 without the additional premise of the transi-
tivity of ' ~ " for
. U2+ U3
1fu2>U3 then U2 > - 2 - > U3
i=1
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-
UTILITY THEORY AND PREFERENCE LOGIC 333
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/
UTILITY THEORY AND PREFERENCE LOGIC 335
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-
UTILITY THEORY AND PREFERENCE LOGIC 337
• 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.
UTILITY THEORY AND PREFERENCE LOGIC 339
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
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-
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
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
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
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-
DER ERSTE WIENER KREIS 351
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
DER ERSTE WIENER KREIS 353
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
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
DER ERSTE WIENER KREIS 357
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).
Karl-Franzens-Universitat Graz
Institut fUr Phi1osophie
Heinrichstrasse 26
A-8010 Graz
Osterreich
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
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
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:
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.
University of Pittsburgh,
1012 Cathedral,
Pittsburgh, PA 15260,
U.S.A.
W. K. ESSLER
ON DETERMINING DISPOSITIONS
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
© 1985 by D. Reidel Publishing Company
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
NOTES
( KurzJassung 1 )
INHALT
Anmerkungen 454
Literatur 458
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
© 1985 by D. Reidel Publishing Company
370 ULRICH BLAU
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:
I
oder <Iall.", ... , larl.,,> E P~.
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
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
UNBESTIMMTHEITEN UND PARADOXIEN 377
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.
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
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?
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
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
UNBESTIMMTHEITEN UND PARADOXIEN 381
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. 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:
UNBESTIMMTHEITEN UND PARADOXIEN 383
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 ...
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 ...
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;
UNBESTIMMTHEITEN UND PARADOXIEN 387
"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
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
UNBESTIMMTHEITEN UND PARADOXIEN 389
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 •
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
FnW
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
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
W'} -
onA ist {F'
{On
A ist nieht 0· (r > n)
2. SYNTAX VON LR
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
UNBESTIMMTHEITEN UND PARADOXIEN 399
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
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
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.
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.
UNBESTIMMTHEITEN UND PARADOXIEN 407
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'
UNBESTIMMTHEITEN UND PARADOXIEN 409
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
5. Syntaktische Priidikate
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 •
UNBESTIMMTHEITEN UND PARADOXIEN 411
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'
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
9. Kennzeichnung
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
4.1. Standardjunktoren
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
UNBESTIMMTHEITEN UND PARADOXIEN 415
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
4.3. Wahrheitswertkonstanten
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.
UNBESTIMMTHEITEN UND PARADOX lEN 419
Die Analyse der logischen Syntax gehort nicht zu den Zielen von LR, aber
420 ULRICH BLAU
{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.
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.
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.
gentheorie. Das sp~itere System LRe, Abschnitt 6, ist fiir die bescheidenen
Zwecke der Syntax-Analyse schon ganz unn6tig stark.
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:
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:
UNBESTIMMTHEITEN UND PARADOXIEN 423
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.
Die mogliche Nichtexistenz betrifft die Denotate von 'den* a' und 'val* a',
vgl. Abschnitt 4.10 unten.
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*.
UNBESTIMMTHEITEN UND PARADOXIEN 425
R=S;;p!
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.
(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) .. ~ =
UNBESTIMMTHEITEN UND PARADOXIEN 429
lSi •• U1
- -
••• U In
p •• qI'
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.
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.
!
RE!
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
Rf
If(at . .. ar)l; {
= U} - I(x)f'at ... arxl;
{ = {u},
existiert nieht ist keine Einermenge.
UNBESTIMMTHEITEN UNO PARAOOXIEN 435
Ival:rAlI; = IAI;',
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),
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:
!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
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
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'
R<
Iii < vi =
1WI 1F <=>
u und v sind Zahlen (Strichfolgen) und u
U und v sind Zahlen und u ~ v,
<: v,
IN LR< gilt:
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.
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
UNBESTIMMTHEITEN UND PARADOXIEN 441
LJ +'
W
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.
UNBESTIMMTHEITEN UND PARADOXIEN 443
(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
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)
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
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.
UNBESTIMMTHEITEN UNO PARAOOXIEN 447
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
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:
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 ist intuitiv korrekt, falls ZF korrekt ist. Und ZF + Mo ist starker als
jedes obigen sa, da
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
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:
UNBESTIMMTHEITEN UND PARADOXIEN 451
6. TRANSFINITE REFLEXIONSSTUFEN
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).
UNBESTIMMTHEITEN UND PARADOXIEN 453
;:::: 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
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.
UNBESTIMMTHEITEN UND PARADOXIEN 455
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
LITERATUR
(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.
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
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
EREIGNISERKLARUNGEN 467
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
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.
EREIGNISERKLARUNGEN 473
1. VORBEMERKUNGEN
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.
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
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.
/\x(Px-+Qx)
Pa A Qa
Qa
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.
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
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
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.