You are on page 1of 374

FROM INSTRUMENTALISM TO

CONSTRUCTIVE REALISM

SYNTHESE LIBRARY
STUDIES IN EPISTEMOLOGY,
LOGIC, METHODOLOGY, AND PHILOSOPHY OF SCIENCE

Managing Editor:

JAAKKO HINTIKKA, Boston University


Editors:

DIRK VAN DALEN, University of Utrecht, The Netherlands


DONALD DAVIDSON, University of California, Berkeley
THEO A.F. KUIPERS, University of Groningen, The Netherlands
PATRICK SUPPES, Stanford University, California
JAN WOLENSKI, iagiellonian University. Krakow, Poland

VOLUME 287

FROM INSTRUMENTALISM
TO CONSTRUCTIVE REALISM
On Some Relations between Confirmation,
Empirical Progress, and Truth Approximation

by

THEO A.F. KUIPERS


University of Groningen, The Netherlands

SPRINGER-SCIENCE+BUSINESS MEDIA, B.Y.

A catalogue record for this book is available from the Library of Congress.

ISBN 978-90-481-5369-5
ISBN 978-94-017-1618-5 (eBook)
DOI 10.1007/978-94-017-1618-5

Printed on acid-free paper

Ali righ ts reserved


2000 Springer Science+Business Media Dordrecht
Originally published by K1uwer Academic Publishers in 2000
No part of the material protected by this copyright notice may be reproduced
or utilized in any form or by any means, electronic, mechanical,
including photocopying, recording or by any information storage and
retrieval system, without written permission from the copyright owners.

TABLE OF CONTENTS

FOREWORD

ix

CHAPTER 1: GENERAL INTRODUCTION: EPISTEMOLOGICAL


POSITIONS
1.1. Four perspectives on theories
1.2. The four main epistemological questions
1.3. The main epistemological and methodological claims
1.4. Preliminaries and a survey of cognitive structures

1
3
8
10

PART I: CONFIRMATION

CHAPTER 2:
2.1.
2.2.
2.3.

INTRODUCTION TO PART I

15

CONFIRMATION BY THE HD-METHOD


A qualitative theory of deductive confirmation
Ravens, emeralds, and other problems and solutions
Acceptance of hypotheses

17
21
27
38

CHAPTER 3: QUANTITATIVE CONFIRMATION, AND ITS


QUALITATIVE CONSEQUENCES
3.1 . Quantitative confirmation
3.2. Qualitative consequences
3.3. Acceptance criteria
Appendix 1: Corroboration as inclusive and impure
confirmation
Appendix 2: Comparison with standard analysis of the raven
paradox

43

CHAPTER 4: INDUCTIVE CONFIRMATION AND INDUCTIVE


LOGIC
4.1. Inductive confirmation
4.2. The continuum of inductive systems
4.3. Optimum inductive systems
4.4. Inductive analogy by similarity and proximity
4.5. Universal generalizations

73

44
55
65
68
70

74
77
79
80
83

vi

TABLE OF CONTENTS

PART II: EMPIRICAL PROGRESS

INTRODUCTION TO PART II
CHAPTER 5: SEPARATE EVALUATION OF THEORIES BY THE
HD-METHOD
5.1. HD-evaluation of a theory
5.2. Falsifying general hypotheses, statistical test implications,
and complicating factors
CHAPTER 6: EMPIRICAL PROGRESS AND PSEUDOSCIENCE
6.1. Comparative HD-evaluation of theories
6.2. Evaluation and falsification in the light of truth
approximation
6.3. Scientific and pseudoscientific dogmatism

91
93
95
103

111
111
120
126

PART III: BASIC TRUTH APPROXIMATION

INTRODUCTION TO PART III

137

CHAPTER 7:
7.1.
7.2.
7.3.
7.4.
7.5.

TRUTH LIKENESS AND TRUTH APPROXIMATION


Actual truthlikeness
Nomic truthlikeness
Actual and nomic truth approximation
Survey of bifurcations
Novel facts, crucial experiments, inference to the best
explanation, and descriptive research programs

139
142
146
154
165
166

CHAPTER 8:
8.1.
8.2.
8.3.

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS


Conceptual foundations of nomic truth approximation
Truthlikeness and the correspondence theory of truth
Explicating dialectical concepts

173
173
190
198

CHAPTER 9: EPISTEMOLOGICAL STRATIFICATION OF NOMIC


TRUTH APPROXIMATION
9.1. Theoretical and substantial nomic truth approximation
9.2. Referential truth approximation
9.3. Rules of inference, speculations, extensions, and
explanatory research programs
9.4. Epistemological positions reconsidered

208
209
219
228
236

TABLE OF CONTENTS

vii

PART IV: REFINED TRUTH APPROXIMATION

INTRODUCTION TO PART IV
CHAPTER 10: REFINEMENT OF NOMIC TRUTH
APPROXIMATION
10.1. Structurelikeness
10.2. Refined nomic truthlikeness and truth approximation
10.3. Foundations of refined nomic truth approximation
10.4. Application: idealization & concretization
to.5. Stratified refined nomic truth approximation

243
245
246
249
262
268
272

CHAPTER 11: EXAMPLES OF POTENTIAL TRUTH


APPROXIMATION
11.1. The old quantum theory
11.2. Capital structure theory

278

CHAPTER 12: QUANTITATIVE TRUTHLIKENESS AND TRUTH


APPROXIMATION
12.1. Quantitative actual truth likeness and truth approximation
12.2. Quantitative nomic truthlikeness
12.3. Quantitative nomic truth approximation

299

CHAPTER 13: CONCLUSION: CONSTRUCTIVE REALISM


13.1. Main conclusions
13.2. Three types of induction
13.3. Formation of observation terms
13.4. Direct applicability of terms
13.5. The metaphysical nature of scientific research
13.6. Portraits of real and fictitious scientists
13.7. Reference and ontology
13.8. Truth definitions and truth criteria
13.9. Metaphors

317
317
320
322
324
325
327
329
330
331

NOTES

334

REFERENCES

347

INDEX OF NAMES

355

INDEX OF SUBJECTS

359

278
288

301
302
308

FOREWORD

Over the years, I have been working on two prima facie rather different, if not
opposing, research programs, notably Carnap's confirmation theory and
Popper's truth approximation theory. However, I have always felt that they
were compatible, even smoothly synthesizable, for all empirical scientists use
confirmation intuitions, and many of them have truth approximation ideas.
Gradually it occurred to me that the glue between confirmation and truth
approximation was the instrumentalist or evaluation methodology, rather than
the falsificationist one. By separate and comparative evaluation of theories in
terms of their successes and problems, hence even if already falsified, the
evaluation methodology provides in theory and practice the straight route for
short-term empirical progress in science in the spirit of Laudan. Further analysis
showed that this sheds also new light on the long-term dynamics of science
and hence on the relation between the main epistemological positions, viz.,
instrumentalism, constructive empiricism, referential realism, and theory realism of a non-essentialist nature, here called constructive realism. Indeed, thanks
to the evaluation methodology, there are good, if not strong, reasons for all
three epistemological transitions "from instrumentalism to constructive
realism".
To be sure, the title of the book is ambiguous. In fact it covers (at least)
three interpretations. Firstly, the book gives an explication of the mentioned
and the intermediate epistemological positions. Secondly, it argues that the
successive transitions are plausible. Thirdly, it argues that this is largely due
to the instrumentalist rather than the falsificationist methodology. However,
to clearly distinguish between the instrumentalist methodology and the instrumentalist epistemological position, the former is here preferably called the
evaluation methodology. In the book there arises a clear picture of scientific
development, with a short-term and a long-term dynamics. In the former there
is a severely restricted role for confirmation and falsification, the dominant role
is played by (the aim of) empirical progress, and there are serious prospects
for observational, referential and theoretical truth approximation. Moreover,
the long-term dynamics is enabled by (observational, referential and theoretical)
inductive jumps, after 'sufficient confirmation', providing the means to enlarge
the observational vocabulary in order to investigate new domains of reality.
This book presents the synthesis of many pieces that were published in various
journals and books. Material of the following earlier publications or publications to appear has been used, with the kind permission of the publishers:
IX

FOREWORD

'The qualitative and quantitative success theory of confirmation" Part 1 and


2, to appear in Logique et Analyse. (Chapter 2, 3).
"The Carnap-Hintikka programme in inductive logic", in Knowledge and
Inquiry: Essays on laakko Hintikka's epistemology and philosophy of science, ed.
Matti Sintonen, Poznan Studies, Vol. 51, 1997, pp. 87-99. (Chapter 4).
"Explicating the falsificationist and the instrumentalist methodology by
decomposing the hypothetico-deductive method", in Cognitive patterns in science and common sense, eds. T. Kuipers and A.R. Mackor, Poznan Studies,
Vol. 45, Rodopi, Amsterdam, 1995, pp. 165-186. (Chapter 5).
"Naive and refined truth approximation", Synthese, 93, 1992, 299- 341.
(Chapter 7, 9, 10).
"The dual foundation of qualitative truth approximation", Erkenntnis, 47.2,
1997,145-179. (Section 8.1., Chapter 10).
"Truthlikeness and the correspondence theory of truth", Logic, Philosophy
of Science and Epistemology, Proc. 11th Wittgenstein Symp. 1986, Wenen, 1987,
171-176. (Section 8.2.).
"Structuralist explications of dialectics", in: Advances in scientific philosophy.
Essays in honour of Paul Weingartner on the occasion of the 60-th anniversary
of his birthday, eds. G. Schurz and G. Dorn, Poznan Studies, Vol. 24, Rodopi,
Amsterdam, 1991,295-312. (Section 8.3.).
"Comparative versus quantitative truthlikeness definitions. Reply to
Mormann", Erkenntnis, 47.2, 1997, 187- 192. (Chapter 10).
"Sommerfeld's Atombau: a case study in potential truth approximation",
together with H. Hettema, in Cognitive patterns in science and common sense,
eds T. Kuipers and A.R. Mackor, Poznan Studies, Vol. 45, Rodopi, Amsterdam,
1995, pp. 273-297. (Section 11.1.)
"Truth approximation by concretization in capital structure theory", together
with K. Cools and B. Hamminga, in: Idealization VI: Idealization in economics,
eds B. Hamminga and N.B. de Marchi, Poznan Studies, Vol. 38, Amsterdam,
Rodopi, 1994, pp. 205- 228. (Section 11.2.).
Many people have contributed over the years to my writing and teaching about
these subjects. To begin with the PhD-students who were engaged in related
research: Roberto Festa, Bert Hamminga, Hinne Hettema, Yao-Hua Tan, Henk
Zandvoort and Sjoerd Zwart. I profited a lot from interaction with them.
Moreover, I gratefully acknowledge the permission of the co-authors Hinne
Hettema, and Kees Cools and Bert Hamminga to use for CHAPTER 11 the
main parts of two earlier joint publications on applications of truthlikeness
ideas in physics and economics, respectively.
I would also like to thank many others who have critically commented upon
underlying research and previous versions of chapters: David Atkinson,
Wolfgang Balzer, Johan van Benthem, Anne Boomsma, Roger Cooke,
Domenico Costantini, Anton Derksen, Igor Douven, Job van Eck, Arthur Fine,
Kenneth Gemes, Carl Hempel, Johannes Heidema, Jaakko Hintikka, Richard

FOREWORD

Xl

Jeffrey, Otto Kardaun, Erik Krabbe, David Miller, Ivo Molenaar, Hans Mooij,
Thomas Mormann, Ulises Moulines, Wim Nieuwpoort, Ilkka Niiniluoto,
Leszek Nowak, Graham Oddie, David Pearce, Karl Popper, Hans Radder,
Hans Rott, Willem Schaafsma, Gerhard Schurz, Abner Shimony, Brian Skyrms,
Wolfgang Spohn, Peter Urbach, Nicolien Wierenga, Andrzej Wisniewski, Sandy
Zabell, Gerhard Zoubek, and Jan Zytkow.
The opportunity to write this book was provided by three important factors.
First, I got a sabbatical year (96/97) from my home university, the University
of Groningen, the Netherlands. Second, I was, after fourteen years, again invited
as a fellow for a year at the Dutch work-paradise for scholars, the Netherlands
Institute for Advanced Study of the Royal Dutch Academy of Sciences (NIAS,
in Wassenaar, near Leiden). The support of the staff was in various respects
very pleasant and efficient. I am especially grateful to Anne Simpson and Jane
Colling for editing my English. Third, and finally, three colleagues and friends
were willing to read and comment upon the entire manuscript: Roberto Festa,
Ilkka Niiniluoto, Andrzej Wisniewski, and some others were willing to comment
upon one or more chapters: Kenneth Gemes and Hans Radder. Of course, the
responsibility for any shortcomings is mine.
Theo A.F. Kuipers,
September 1999,
Groningen, The Netherlands

GENERAL INTRODUCTION:
EPISTEMOLOGICAL POSITIONS

Introduction
We will start by sketching in a systematic order the most important epistemological positions in the instrumentalism-realism debate, viz., instrumentalism,
constructive empiricism, referential realism and theory realism. They will be
ordered according to their answers to a number of leading questions, where
every next question presupposes an affirmative answer to the foregoing one.
The survey is restricted to the investigation of the natural world and hence to
the natural sciences. It should be stressed that several complications arise if
one wants to take the social and cultural world into account. However, the
present survey may well function as a point of departure for discussing epistemological positions in the social sciences and the humanities.
We will not only include the different answers to the standard questions
concerning true and false claims about the actual world, but also the most
plausible answers to questions concerning claims about the nomic world, that
is, about what is possible in the natural world. Moreover, we will include
answers to questions about 'actual' and 'nomic' truth approximation. At the
end we will give an indication of the implications of the results of the present
study of confirmation, empirical progress and truth approximation for the way
the epistemological positions are related. The main conclusions are the
following: there are good reasons for the instrumentalist to become a constructive empiricist; in his turn, in order to give deeper explanations of success
differences, the constructive empiricist is forced to become a referential realist;
in his turn, there are good reasons for the referential realist to become a theory
realist of a non-essentialist nature, here called a constructive realist.
1.1. FOUR PERSPECTIVES ON THEORIES

The core of the ongoing instrumentalism-realism debate is the nature of proper


theories, that is, theories using theoretical terms, or rather the attitude one
should have towards them. Proper theories arise from the two-level distinction
between observational and theoretical terms, as opposed to observational laws
and theories, which only use, by definition, observation terms. The resulting
two-level distinction between observational laws and proper theories gives rise

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

to a short-term dynamics in the development of scientific knowledge. That is,


as we will elaborate in Part I, observational hypotheses are tested and may be
falsified or 'sufficiently' confirmed such that they become accepted. On the
other hand, as we will elaborate in Part II, proper theories are, before and
after falsification, evaluated and compared in terms of the observational laws
they contradict or are able to successfully explain and predict. The survey of
positions will primarily be presented in terms of this short-term dynamics.
However, from time to time we will also take the long-term dynamics into
account generated by the transformation of proper theories into observation
theories and giving rise to a multi-level distinction according to which proper
theories may not only explain or predict a lower level observational law, but
also be presupposed by a higher level one. For the main positions the view on
the nature of proper theories implies also a view on the nature of the indicated
short-term and long-term dynamics.
For most of the positions it makes sense to distinguish four possible perspectives on the nature of theories in general. First, we may distinguish two versions
or modes of the natural world, the one called the actual world, that is, the
natural world of the past, the present and the future, the other called the nomic
world, that is, the natural world in the sense of all conceivable possibilities that
are physically possible. The actual world is of course in some sense part of the
nomic world. The two versions will lead to two modes of each position, that
is, the 'actualist' version, merely pertaining to the actual world, and the nomic
version, which primarily pertains to the nomic world. Second, one's primary
focus may be the truth-value of claims about the (actual or nomic) natural
world, or claims about truth approximation, that is, claims of the form that one
theory is closer to the (actual or nomic) truth than another.
In sum, there are four possible (primary) perspectives on the nature of
observational and proper theories, depicted in Table 1.1, in which we mention
some leading authors representing the realist tradition! . Below we will see that
anti-realist authors such as Van Fraassen usually subscribe to the standard,
actualist and truth-value oriented, perspective, restricted to observational theories, of course.
Table 1.1. Four (realist) perspectives on theories
Mode

Actual

Nomic

Standard/traditional
Peirce/ Niiniluoto

Giere
Popper

Focus
Truth-value
Truth approximation

Taking the variety of (versions of) positions seriously, implies that the notions
of 'true' and 'false' are assumed to have adapted specifications, that is, 'true'
and 'false' may refer to the actual or the nomic world and may concern
observationally, referentially, or theoretically true and false, respectively. The

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

same holds for the notion of 'the truth', but it should be stressed in advance
that it will always be specified in a domain-and-vocabulary relative way. Hence,
no language independent metaphysical or essentialist notion of 'THE TRUTH'
is assumed. With this in mind, let us start with the successive questions.
The first question really is not an epistemological question, but a preliminary
ontological question.
Question 0: Does a natural world that is independent of human
beings exist?

The question of whether there is a human-independent natural world is to


some extent ambiguous. If one thinks in terms of a conceptualized natural
world, that is, a world that in some way or other brings one or more conceptualizations with it, then one has to assume either a kind of essences underlying
the natural world, or one has to assume a non-human designer of the natural
world, or some human intervention. We take the question in the other sense,
that is, does a non-conceptualized natural world independent of human beings
exist? If one answers this question negatively one takes the position of (straightforward) ontological idealism. If one answers it positively we will speak of
ontological realism. It is certainly questionable (Rescher 1992) whether so-called
idealists, like Berkeley, really subscribed to ontological idealism. However this
may be, and this is our main argument, it is highly implausible and not taken
seriously by natural scientists. The positive answer may be plausible in the
actualist version, it is not evident in the nomic version. To speak meaningfully
about physical or, more generally, nomic possibilities seems to require at least
some conceptualization. However, in the same way, to say something specifically about the actual world presupposes some conceptualization. Hence, as
far as the role of conceptualizations is concerned, it can be argued that if one
believes in an unconceptualized actual world, which can be conceptualized in
different ways, one may equally well believe in an unconceptualized nomic
world, which can be conceptualized in at least as many ways. To be precise,
there are more, because in order to conceptualize the nomic world it is plausible
to take also dispositions into account which do not seem to be of any use in
characterizing the actual world. Apart from the comparable (epistemological!)
role of conceptualizations, it is clear that nomic ontological realism makes a
stronger ontological claim than actual ontological realism. In the following, it
will be clear when the nomic version is presupposed and when the actual
version suffices.
1.2. THE FOUR MAIN EPISTEMOLOGICAL QUESTIONS

The first really epistemological question is:


Question 1: Can we claim to possess true claims to knowledge about
the natural world?

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

Again it is important to first eliminate some ambiguities in this question. A


positive answer to this question does not presuppose that true claims can be
known to be true with certainty in some fundamental sense, nor that they can
be verified or can be recognized as true or can be obtained in some other way.
Moreover, a positive answer evidently does not presuppose that the true claims
the question refers to do not depend on a conceptualization, for the formulation
of claims requires a vocabulary. What is intended is only that the claim that
we can have good reasons for assuming that certain claims, by definition
phrased in a certain vocabulary, about the natural world are true in some
objective sense, whereas others are false. The negative answer amounts to the
position of epistemological relativism, the positive answer may be called epistemological realism, with an actual and a nomic version, depending on whether
all of these claims are just claims about the actual world or some about the
nomic world. As with the position of ontological idealism, the position of
epistemological relativism is hard to take seriously in some straightforward,
consistent way when dealing with scientific knowledge and knowledge development, despite the fact that it has played a dominant role in the history of
epistemology in the form of skepticism of one kind or another. The main forms
of skepticism are experiential skepticism, that is, skepticism with respect to
claims about sensory and introspective experiences and inductive skepticism,
that is, skepticism merely with respect to inductive extrapolations in the sense
of inductive predictions or inductive generalizations. Scientists can escape
skepticism by always taking the possibility of experiential illusions and inductive mistakes seriously. As a consequence, although skepticism still plays an
important role in philosophical curricula and, as a matter of fact, recently
became fashionable in certain, so-called postmodern, circles, it does not concern
the core issue between realism and instrumentalism.
The next question brings us to the heart of the distinction between observational and theoretical terms. Assuming a certain definition of observability and
hence of the border between observational and theoretical terms, the following
question arises:
Question 2: Can we claim to possess true claims to knowledge about
the natural world beyond what is observable?

In other words, the question is whether more than observational knowledge,


that is, knowledge in observation terms, is possible. A negative answer only
makes sense, of course, if the notion of observability is relatively fixed. Our
human observation possibilities might be extended, or just change, due to some
evolutionary or artificial change of our physiological abilities. Moreover, they
might be extended by accepting some observational laws, that enable the
definition of new concepts. However, in case of the negative answer, one has
to exclude the possibility of extension of the observable on the basis of the
acceptance of proper theories. In other words, the transformation process of

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

proper theories into observation theories, such as the atomic theory of matter
became for nuclear physics, has then to be conceived as merely a way of
speaking, giving rise to other kinds of as-if-behavior. A positive answer to the
present question amounts to so-called scientific realism, according to which
proper theories, or at least theoretical terms, have to be taken seriously. A
negative answer might be said to reflect observational realism or just empiricism.
As a matter of fact, there are two well-known types of the negative answer
to Q2. According to the first type, usually called instrumentalism, talking about
reference of theoretical terms does not make sense, let alone talking about true
or false (proper) theories. This way of talking reflects according to the instrumentalist a kind of category mistake by mistakenly extrapolating meaningful
terminology for the observational level to the theoretical level. The only function of proper theories is to provide good derivation instruments, that is, they
need to enable the derivation of as many true observational consequences as
possible and as few false observational consequences as possible. Hence, the
ultimate aim of the instrumentalist is the best derivation instrument, if any.
Well-known representatives of the instrumentalist position among philosophers
are Schlick (1938) and Toulmin (1953). Although Laudan (1977,1981) admits
that theories have truth-values, he is frequently called an instrumentalist
because of his methodology, according to which theories are not disqualified
by falsification as long as they cannot be replaced by better ones. Moreover,
although debatable, the physicist Bohr is a reputed instrumentalist, at least as
far as quantum mechanics is concerned. Notice that it is plausible to make the
distinction between an actual and a nomic version of instrumentalism depending on whether the relevant true and false observational consequences all
pertain to the actual world or at least some to the nomic world.
According to the second type of negative answer to Q2, called (constructive)
empiricism by its inventor and main proponent Van Fraassen (1980, 1989),
there is no category mistake, that is, the point is not whether or not theoretical
terms can refer and whether proper theories can be true or false. In fact, such
terms mayor may not refer and such theories are true or false, but the problem
is that we will never know this beyond reasonable doubt. Hence, what counts is
whether such theories are empirically adequate or inadequate or, to use our
favorite terminology, whether they are observationally true or false. Again
there are two versions, actual and nomic empiricism, depending on whether
the theories are supposed to deal all with the actual world or at least some
with the nomic world. Although Van Fraassen is clear about his non-nomic
intentions, the nomic analogue has some plausibility of its own. That is, it
makes perfect sense to leave room for observational dispositions, without taking
theoretical terms of other kinds seriously. In other words, if one conceives
dispositions in general, hence including observational dispositions, as theoretical terms, one may well reserve a special status for observational dispositions.
In both versions, it makes sense to talk about the observational truth in the
sense of the strongest true observational hypothesis about a certain domain of

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

the natural world within a certain vocabulary. Assuming that it is also possible
to make sense of the idea that (the observational theory following from) one
theory is closer to the observational truth than another, even convergence to
the observational truth is possible. As suggested, Van Fraassen is a strong
defender of empiricism in the traditional, that is, actualist and truth-value
oriented, perspective, where it should be remarked that his attitude is strongly
influenced by his interest in quantum mechanics. Although Van Fraassen
extrapolates this attitude to other proper theories, there are also scientists, in
fact there are many, who take advantage of the fact that there may be examples
of proper theories towards which an empiricist attitude is the best defensible
one, whereas there are other examples towards which a realist attitude is the
best defensible one. This gives rise to what Dorling (1992) has aptly called
local positivist versus realist disputes, as opposed to the global dispute about
whether it is a matter of yes or no for all proper theories at the same time. In
this respect the empiricist attitude is usually identified with a position in the
global dispute, the realist positions that follow usually leave room for local
empiricist deviations from the globally realist attitude, as a kind of default
heuristic rule.
As remarked already, for both types of empiricists, the long-term dynamics
in science, according to which proper theories transform into observation
theories, has to be seen as an as-if way of speaking. The question even arises
whether this is really a coherent way of deviating from scientific practice where
it seems totally accepted that the concept of observation is stretched to the
suggested theory-laden interpretation.
Hence, it is time to turn to the positive answer to Q2, that is, to the position
called scientific realism. Since the books by Hacking (1983) and Cartwright
(1983) there is a weaker version of realism than the traditional one, which is
suggested by the next question.

Question 3: Can we claim to possess true claims to knowledge about


the natural world beyond (what is observable and) reference claims
concerning theoretical terms?
Whereas Hacking and Cartwright, when answering this question in the negative
sense, primarily think of reference of entity terms, and call their position entity
realism, it is highly plausible to extrapolate that position to attribute terms, in
some plausible sense of reference, and speak of referential realism. 2 According
to referential realism, entity and attribute terms are intended to refer, and
frequently we have good reasons to assume that they do or do not refer. Again
it is possible to distinguish an actual and a nomic version, not only with respect
to observational consequences, but also with respect to theoretical terms. For
instance, when one takes the existence of atoms in the actual world seriously,
which goes beyond empiricism, it is also defensible to take the existence of
physically possible atoms seriously, even if they do not (yet) exist in the actual

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

world. In a sense this is just a definition of (existence in) the nomic world, as
encompassing the actual world. Moreover, in both versions it is possible that
one theory is observationally and referentially closer to the truth than another,
as soon as we assume, in addition to the previous assumptions for observational
truth approximation, that it is possible to define the idea that (the total
referential claim of) one theory can be closer to the referential truth than
another. Here, the referential truth is of course the strongest true referential
claim which can be made within a certain vocabulary about a certain domain.
However, since referentialists do not want to take theoretical induction seriously,
that is, deciding to further assume that a certain proper theory is true (see
further below), the transformation of proper theories into observation theories
is for them no more open than for empiricists, i.e., it is open only in some as-ifreading. Referential realism seems, however, more difficult to defend than
constructive empiricism, in particular when one takes the possibility of truth
approximation into account. That is, as long as one is only willing to think in
terms of true and false claims about theoretical terms when they are supposed
to refer, one may be inclined to hold that most of these claims, past and future
ones, are false. However, as soon as one conceives of sequences of such claims
that may approach the truth, it is hardly understandable that the truth would
not be a worthwhile target, at least in principle. Hence, let us turn to the
suggested stronger position.
The positive answer to Q3 brings us to so-called theoretical or theory realism,
in some version or another advocated by, for instance, Peirce (1934), Popper
(1963), and Niiniluoto (1987a)3. Theory realism shares with referential realism
the claim that theoretical terms are supposed to refer, and that, from time to
time, we have good reasons to assume that they refer, including the corresponding truth approximation claims. It adds to this the claim that theories are
claimed to be true, and that we have from time to time good reasons to further
assume that they are true, that is, to carry out a theoretical induction. Moreover,
proper theories can converge to the theoretical truth, that is, the strongest true
claim that can be made, in a given vocabulary, about a specific domain, again
leaving room for an actual and a nomic version. Although the truth to be
approached is again domain-and-vocabulary relative, this does not exclude, of
course, the possibility of comparison and translation of theories. Moreover,
theoretical induction is always a matter for the time being, a kind of temporal
default rule: as long as there is no counter-evidence, it is assumed to be true.
This default-assumption not only implies that the theoretical terms of the
theory then are assumed to refer, but also that the proper theory can from
then on be used as an observation theory. Hence, the transformation process
and the corresponding long-term dynamics are possible.
The last question to be considered is the following:
Question 4: Does there exist a correct or ideal conceptualization of
the natural world?

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

In contrast to the positive answers on the questions Q1 to Q3, a posItive


answer to the fourth question brings us to a position that is not purely
epistemologically built on the positive answer to QO (i.e., ontological realism),
viz., it amounts to an extreme kind of metaphysical realism which we like to
call essentialistic realism. The reason to call it this is, of course, that if there is
an ideal conceptualization then the natural world must have essences of a kind.
For instance, there must be natural kinds, not only in some pragmatic sense,
but in the sense of categories in which entities in the actual or the nomic world
perfectly fit. Philosophers of science like Boyd (1984) and Harre (1986) seem
to come close to this point of view. According to this extreme form of realism,
the challenge of science is to uncover the ideal conceptualization, that is, to
discover and extend the ideal vocabulary, on the basis of which perfect observational, referential and theoretical truths can be formulated. Refinement of this
vocabulary is not so much directed at more precise and deeper truths, but at
additional truths. Of course, it is possible to restrict the idea of an ideal
vocabulary to an observational vocabulary, but there do not seem to be
representatives of this kind of 'essentialistic empiricism'. It will also be clear
that there is again an actual and a nomic version, but if one is an essentialist,
the nomic version seems to be the most plausible one.
The negative answer to Q4 gives rise to what we call constructive realism. 4
The term was already used by Giere (1985) in more or less the same way. The
difference is that Giere does not take truth approximation into account. Peirce,
Popper and Niiniluoto, however, do take truth approximation into account.
Moreover, whereas Peirce and Niiniluoto focus on the actual version, Popper
and Giere seem to have primarily the nomic version in mind, without excluding
the actual version. In our view, the nomic version of constructive realism is
the best fit to scientific practice. The adjective 'constructive' is used for more
or less the same reason as it is used by Van Fraassen, in his case restricted to
the observational level. Vocabularies are constructed by the human mind,
guided by previous results. Of course, one set of terms may fit better than
another, in the sense that it produces, perhaps in cooperation with other related
vocabularies, more and/or more interesting truths about the domain than
another. The fruitfulness of alternative possibilities will usually be comparable,
at least in a practical sense, despite the possibility of fundamental incommensurability. There is however no reason to assume that there comes an end to the
improvement of vocabularies.
We summarize the preceding survey in Figure 1.1.
1.3 . THE MAIN EPISTEMOLOGICAL AND METHODOLOGICAL
CLAIMS

We consider instrumentalism, constructive empiricism, referential realism, and


constructive realism to be the main epistemological positions. With the emphasis on their nomic interpretation, they will be further characterized and compared in the light of the results of the analysis of confirmation, empirical

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS


00: independent natural world?
yes

no

~
no

theory realism

04: ideal conceptualization?


yes

ontological idealism

epistemological relativism
'experiential skepticism
'inductive skepticism
empiricism (observational realism)
, instrumentalism

scientific realism

, constructive empiricism

03: beyond reference?


yes

>

epistemological realism

02: beyond the observable?


yes

no

ontological realism

01: true claims about natural world?


yes

c:=::::::=>

no

no

:>

>

referential realism
=> entity realism
constructive realism

essentialistic realism

Figure 1.1. The main epistemological positions

progress and truth approximation in the rest of this book. The main results
will come available at the end of Chapter 2, 6, 9, and finally in Chapter 13. A
brief indication of the results is the following.
There are good reasons for the instrumentalist to become a constructive
empiricist; in his turn, in order to give deeper explanations of success differences,
the constructive empiricist is forced to become a referential realist; in his turn,
there are good reasons for the referential realist to become a theory realist.
The theory realist has good reasons to choose for constructive realism, since
there is no reason to assume that there are essences in the world. Notice that
the road to constructive realism amounts to a pragmatic argumentation for
this position, where the good reasons will mainly deal with the short-term and
the long-term dynamics generated by the nature of, and the relations between,
confirmation, empirical progress and truth approximation.
Besides these epistemological conclusions, there are some general methodological lessons to be drawn. There will appear to be good reasons for all positions
not to use the falsificationist but the instrumentalist or 'evaluation(ist), methodology. That is, the selection of theories should exclusively be guided by more
empirical success, even if the better theory has already been falsified. Hence, the
methodological role of falsifications will be strongly relativized. This does not at
all imply that we dispute Popper's claim that aiming at falsifiable theories is
characteristic for empirical science, on the contrary, only falsifiable theories can
obtain empirical success. Moreover, instead of denouncing the hypothetico-deductive method, the evaluation methodology amounts to a sophisticated application

10

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

of that method. As suggested, the evaluation methodology may also be called the
instrumentalist methodology, because the suggested methodology is usually associated with the instrumentalist epistemological position. The reason is, of course,
that from that position it is quite natural not to consider a theory as seriously
disqualified by mere falsification. However, since we will argue that that methodology is also very useful for the other positions, we want to terminologically separate
the instrumentalist methodology from the instrumentalist epistemological position, by calling the former the evaluation methodology, enabling to identify 'instrumentalism' with the latter.
We close this section with a warning. The suggested hierarchy of the heuristics
corresponding to the epistemological positions is, of course, not to be taken in
a dogmatic sense. That is, when one is unable to successfully use the constructive
realist heuristic, one should not stick to it, but try weaker heuristics, hence first
the referential realist, then the empiricist, and finally the instrumentalist heuristic. For, as with other kinds of heuristics, although not everything goes always,
pace (the suggestion of) Feyerabend, everything goes sometimes. Moreover,
after using a weaker heuristic, a stronger heuristic may become applicable at
a later stage: "reculer pour mieux sauter".

1.4. PRELIMINARIES AND A SURVEY OF COGNITIVE


STRUCTURES

The book is written in terms of three background assumptions. They concern


cognitive structures, research programs, and the structuralist view on theories.
These and many other topics are elaborated in a twin book, entitled Structures
in Science (Kuipers (SiS, consciously alluding to Ernest Nagel's seminal The
Structure of Science.
In our view, the main aim of philosophy of science is to uncover (cognitive)
structures in knowledge and knowledge development. The underlying assumption is, of course, that there is system in them. Research programs in the spirit
of Lakatos are a kind of macro-structures and can be classified into four idealtypes: descriptive, explanatory, design and explication research programs.
Although they have different goals, they have similar fine-structures, provided
they are centered around a strong idea. This book deals with some explication
programs, notably, around confirmation in the spirit of Carnap and Hempel,
empirical progress in the spirit of Laudan and truth approximation in the spirit
of Popper. Such explication programs may be seen as conceptual design programs. The subject of our explication activities concerns the nature of development in descriptive and explanatory research programs.
The following survey of the parts focuses on the main specific cognitive
structures that will be uncovered.
In Part I, entitled "Confirmation":
- The qualitative and quantitative conceptual systems of confirmation

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

11

and falsification of hypotheses, together with disconfirmation and verification (the 'deductive confirmation matrix' and the 'confirmation
square', respectively).
- Systems of inductive probability, representing inductive confirmation,
among which optimum systems, systems with analogy and systems with
non-zero probability for universal statements.
In Part II, entitled "Empirical progress":
- 'The evaluation report' of a theory in terms of general successes and
(individual) counter-examples, and a systematic survey of the factors
complicating theory testing and evaluation.
- The nature of comparative theory evaluation, giving rise to the 'rule of
success', characterizing 'empirical progress', prescribing to select the
theory, if any, that has so far proven to be the most successful one.
- The symmetric 'evaluation matrix' for the comparison of the relative
success of two theories.
In Part III, entitled "Basic truth approximation":
- The, domain and vocabulary relative, definitions of 'the actual truth'
and 'the nomic truth', the basic definitions of 'closer to the nomic truth'
or 'nomically more truthlike', the success theorem, according to which
'more truthlikeness' guarantees 'more successfulness'. With the consequence that scientists following the rule of success behave functionally
for truth approximation, whether they like it or not.
- The detailed reconstructions of intuitions of scientists and philosophers
about truth approximation, the correspondence theory of truth, and
dialectical patterns.
- The epistemological stratification of these results, including definitions
of reference, the referential truth, and referential truthlikeness, enabling
the detailed evaluation of the three transitions from instrumentalism to
constructive realism.
In Part IV, entitled "Refined truth approximation":
- The basic definitions do not take into account that theories are frequently improved by replacing mistaken models by other mistaken ones.
Roughly the same patterns arise as in Part III by refined definitions
that take this phenomenon into account.
- Specifically idealization and concretization become in this way functional for (potential) truth approximation, which will be illustrated by
three examples, two illustrating the specific pattern of 'double concretization', viz., the law of Van der Waals and capital structure theory, and
one illustrating the pattern of 'specification followed by concretization',
viz., the old quantum theory.
The book concludes with a sketch of the resulting favorite epistemological
position, viz., constructive realism, which may be seen as a background pattern

12

GENERAL INTRODUCTION: EPISTEMOLOGICAL POSITIONS

of thinking, guiding many, if not most, scientists when practicing the instrumentalist methodology.
This brings us to the use-value of cognitive structures. They always concern
informative patterns, which seem useful in one way or another. We briefly
indicate five kinds of possible use-value, with some particular references to
structures disentangled in this book
(a) they may provide the 'null hypothesis of ideal courses of events', which
can playa guiding role in social studies of science; e.g., the cognitive
justification of non-falsificationist behavior, following from the nature
of theory evaluation (Section 6.3.).
(b) they may clarify or even solve classical problems belonging to abstract
philosophy of science; e.g., the intralevel explication of the correspondence theory of truth (Section 8.2.).
(c) they may be useful as didactic instruments for writing advanced textbooks, leading to better understanding and remembrance; e.g., the
comparative evaluation matrix (Section 6.1.).
(d) they may playa heuristic role in research policy and even in science
policy; e.g., confirmation analysis suggests redirection of drugs testing
(Section 2.2.).
(e) last but not least, they may playa heuristic role in actual research; e.g.,
when explanatorily more successful, the theory realist may have good
reasons to hypothesize theoretical truth approximation despite some
extra observational problems (Chapter 9).5
A final remark about the technical means that will be used. Besides some
elementary logic and set theory, some of the main ideas of the structuralist
approach of theories will be used. Where relevant, starting in Part III, what is
strictly necessary will be explained. However, for detailed expositions the reader
is referred to (Balzer 1982) and (Balzer, Moulines and Sneed 1987). For a
general survey the reader may consult (Kuipers 1994), also to appear in
(Kuipers SiS).

PART I

CONFIRMATION

INTRODUCTION TO PART I

The first part is in many respects a systematic exposition of well-known ideas


on deductive, non-deductive, and inductive confirmation. However, we present
these ideas in a non-standard way and refine and revise several standard
solutions of problems associated with these ideas. This part consists of three
chapters, the first one dealing with qualitative (deductive) confirmation as
resulting from applying the hypothetico-deductive (HD-)method, and the
second one dealing with quantitative (deductive and non-deductive) confirmation and its qualitative consequences. The third one explicates the idea of
(quantitative) inductive confirmation and gives a brief survey of the main
systems of inductive confirmation in the Carnap-Hintikka tradition of so-called
inductive logic. The main non-standard aspect is the approach of confirmation
from the 'success perspective', according to which confirmation is equated with
evidential success, more specifically, with an increase of the plausibility of the
evidence on the basis of the hypothesis. Hence, in contrast to standard expositions, confirmation is not equated with an increase of the plausibility of the
hypothesis by the evidence. This is merely an additional aspect of confirmation
under appropriate conditions and epistemological assumptions.
Chapter 2 starts from the classical idea that testing of a hypothesis by the
HD-method aims at establishing the truth-value (of the observational consequences) of a hypothesis, and results in confirmation or falsification. A qualitative (classificatory and comparative) theory of deductive confirmation is
presented, as well as its (in details) non-standard solutions of Hempel's raven
paradoxes and Goodman's grue paradox. It is called a pure theory of confirmation because it is based on the principle that equally successful hypotheses
are equally confirmed by the relevant evidence, that is, independent of the
initial plausibility of the hypotheses.
In Chapter 3 the corresponding quantitative, i.c., probabilistic Bayesian,
theory of confirmation is presented, with a decomposition in deductive and
non-deductive confirmation. It is pure, and inclusive, hence non-standard, in
the sense that it leaves room for the confirmation of hypotheses with probability
zero. The qualitative theory of (general) confirmation, including the one for
deductive confirmation, generated by the quantitative theory, is also formulated.
In Appendix 1 it is argued that Popper's ideas about corroboration basically
amount to an inclusive and impure variant of the Bayesian approach to
confirmation. Appendix 2 elaborates some details of the quantitative treatment
of the raven paradoxes.
15

16

INTRODUCTION TO PART I

Deductive and non-deductive confirmation mayor may not have extrapolating or inductive features. The last chapter of this part (Chapter 4) deals with
the program of inductive confirmation or inductive logic, set up by Carnap
and significantly extended by Hintikka. It is a special explication program
within the Bayesian program of confirmation theory. It is governed by the idea
of extrapolating or learning from experience in a rational way, to be expressed
in terms of 'inductive probabilities' as opposed to 'non-inductive probabilities'.
In later chapters the role of falsification and confirmation will be relativized
in many respects. However, it will also become clear that they remain very
important for particular types of hypotheses, notably, for general observational
(conditional) hypotheses, and for several kinds of (testable) comparative
hypotheses, e.g., hypotheses claiming that one theory is more successful or
(observational, referential and theoretical) truthlike than another.
A recurring theme in this part, and the other ones, will be the localization
and comparison of the main standard epistemological positions, as they have
been presented in Chapter 1, viz., instrumentalism, constructive empiricism,
referential realism, and constructive (theory) realism.

2
CONFIRMATION BY THE HD-METHOD

Introduction
According to the leading expositions of the hypothetico-deductive
(HD-)method by Hempel (1966), Popper (1934/1959) and De Groot
(1961/1969), the aim of the HD-method is to answer the question whether a
hypothesis is true or false, that is, it is a method of testing. On closer inspection,
this formulation of the aim of the HD-method is not only laden with the
epistemological assumption of theory realism according to which it generally
makes sense to aim at true hypotheses, but it also mentions only one of the
realist aims. One other aim of the HD-method for the realist, an essentially
more refined aim, is to answer the question of which facts are explained by the
hypothesis and which facts are in conflict with it. In the line of this second
aim, we will show in Part II that the HD-method can also be used to evaluate
theories, separately and in comparison with other theories, among others, in
terms of their general successes and individual problems, where theories are
conceived as hypotheses of a strong kind. In Part III and IV we will argue
that the methodology of (comparative) HD-evaluation of theories is even
functional for truth approximation, the ultimate aim of the realist.
For the moment we will restrict attention to the HD-method as a method
of testing hypotheses. Though the realist has a clear aim with HD-testing, this
does not mean that HD-testing is only useful from that epistemological point
of view. Let us briefly review in the present connection the main other epistemological positions that were described in Chapter 1. Hypotheses mayor may
not use so-called 'theoretical terms', besides so-called 'observation terms'. What
is observational is not taken in some absolute, theory-free sense, but depends
greatly on the level of theoretical sophistication. Theoretical terms intended to
refer to something in the nomic world mayor may not in fact refer. For the
(constructive) empiricist the aim of HD-testing is to find out whether the
hypothesis is observationally true, i.e., has only true observational consequences, or is observationally or empirically adequate, to use Van Fraassen's
favorite expression. For the instrumentalist the aim of HD-testing is still more
liberal: is the hypothesis observationally true for all intended applications? The
referential realist, on the other hand, adds to the aim of the empiricist to find
out whether the hypothesis is referentially true, i.e., whether its referential
claims are correct. In contrast to the theory realist, he is not interested in the

17

18

CONFIRMATION BY THE HD-METHOD

question whether the theoretical claims, i.e., the claims using theoretical terms,
are true as well. Recall that claims may pertain to the actual world or to the
nomic world (of physical possibilities). However, this distinction will not play
an important role in the first part of this book.
Methodologies are ways of answering epistemological questions. It will turn
out that the method of HD-testing, the test methodology, is functional for
answering the truth question of all four epistemological positions. For this
reason, we will present the test methodology in fairly neutral terms, viz.,
plausibility, confirmation and falsification.
The expression 'the plausibility of a hypothesis' abbreviates the informal
qualification 'the plausibility, in the light of the background beliefs and the
evidence, that the hypothesis is true', where 'true' may be specified in one of
the four main senses: (1) observationally as far as the intended applications
are concerned, (2) observationally, in all possible respects, (3) and, moreover,
referentially, (4) and, even, theoretically. Admittedly, despite these possible
qualifications, the notion of 'plausibility' remains necessarily vague, but that is
what most scientists would be willing to subscribe to. 1 At the end of this
chapter we will further qualify the exposition for the four epistemological
positions when discussing the acceptance of hypotheses. When talking about
'the plausibility of certain evidence', we mean, of course, 'the (prior) plausibility
of the (observational) hypothesis that the test will result in the reported outcome'. Hence, here 'observationally true' and 'true' coincide by definition of
what can be considered as evidential statements.
Regarding the notions of 'confirmation' and 'falsification' the situation is
rather asymmetric. 'Falsification' of a hypothesis simply means that the evidence
entails that the hypothesis is observationally false, and hence also false in the
stronger senses. However, what 'confirmation' of a hypothesis precisely means,
is not so clear. The explication of the notion of 'confirmation' of a hypothesis
by certain evidence in terms of plausibility will be the main target of this and
the following chapter. It will be approached from the success perspective on
confirmation, equating confirmation with an increase of the plausibility of the
evidence on the basis of the hypothesis, and implying that the plausibility of
the hypothesis is increased by the evidence.
The variety of empirical hypotheses is large. To stress this we mention a
number of examples:
- Mozart was poisoned by Salieri,
- Dutch is more similar to English than to German,
- When people have bought something they have selective attention for
information justifying their choice,
- People tend to choose that action which maximizes their expected utility,
- The function of lungs is to supply oxygen to the organism,
- The average rainfall per year gradually increased in the 20th century,
- The universe originated from the big bang,

CONFIRMATION BY THE HD-METHOD

19

- Action is minus reaction,


- Dalton's theory of the atom. The last example will be used a couple of
times for illustrative purposes in this part of the book.
A general characterization of an empirical hypothesis is the following: an
empirical hypothesis is a tentative statement which, in some way or other, is
about the nomic world and which is testable. Statements about the nomic
world claim to tell something about the world, how it is or was (not), how it
will (not) be, how it can (not) be. They are supposed to have, or may acquire
a sufficiently clear meaning such that they are true or false in a sense which
has to be specified, preferably such that a false statement may, nevertheless, be
approximately true in some plausible sense. The above characterization of an
empirical hypothesis still leaves room for hypotheses with nonsensical or otherwise redundant additions. Instead of trying to exclude such additions, which
is not easy, it will become clear that such additions are relatively harmless.
A test for a hypothesis may be experimental or natural. That is, a test may
be an experiment, an active intervention in nature or culture, but it may also
concern the passive registration of what is or was the case, or what happens
or has happened. In the latter case of a so-called natural test the registration
may be a more or less complicated intervention, but is nevertheless supposed
to have no serious effect on the course of events of interest.
According to the HD-method a hypothesis H is tested by deriving test
implications from it, and checking, if possible, whether they are true or false.
Each test implication has to be formulated in terms that are considered to be
observation terms. A test implication mayor may not be of a general nature.
Usually, a test implication is of a conditional nature, if C then F (C -> F). Here
C denotes one or more initial conditions which can be, or have been realized,
by nature or artificially, i.e., by experiment. F denotes a potential fact (event
or state of affairs) predicted by Hand C. If C and F are of an individual nature,
F is called an individual test implication, and C -> F a conditional one. When
C is artificially realized, it is an experimental test, otherwise it is a natural test.
As is well-known, the basic logic of HD-testing can be represented by some
(valid) applications of Modus (Ponendo) Ponens (MP), where 'F' indicates
logical entailment:
Hl=l
H

HI=(C-+F)
H,C

Scheme 2.1.1.

Scheme 2.1.2.

It should be stressed that H /= I and H /=(C -> F) are supposed to be deductive


claims, i.e., claims of a logico-mathematical nature.
The remaining logic of hypothesis testing concerns the application of Modus
(Tollendo) Tollens (MT). Neglecting complications that may arise, if the test

20

CONFIRMATION BY THE HD-METHOD

implication is false, the hypothesis must be false, it has been falsified, for the
following arguments are deductively valid:
HFI
-,1

HF(C-+F)
C, -,F
-,H

Scheme 2.2.1.

Scheme 2.2.2.

When the test implication turns out to be true, the hypothesis has of course
not been (conclusively) verified, for the following arguments are invalid, indicated by '-/- /-':
HFI
I

HF(C-+F)
C,F

- j- j-

- j- j- j- j- j-

H?

H?

Scheme 2.3.).

Scheme 2.3.2.

Since the evidence (I or C&F) is compatible with H, H may still be true.


However, we can say more than that. Usually it is said that H has been
confirmed. It is important to note that such confirmation by the HD-method
is in the strong sense that H has obtained a success of a (conditional) deductive
nature; by entailing the evidence, H makes the evidence as plausible as possible.
This will be called the success perspective on conditional) deductive)
corifirmation.
Falsification and confirmation have many complications, e.g., due to auxiliary
hypotheses. We will deal with several complications, related to general and
individual test implications, at the end of Chapter 5. There is however a great
difference between falsification and confirmation. Whereas the 'logical grammar' of falsification is not very problematic, the grammar of confirmation, i.e.,
the explication of the concept of confirmation, has been a subject of much
dispute.
One of the most influential theories of confirmation is the so-called Bayesian
theory (Howson and Urbach 1989; Earman 1992), which is of a quantitative
nature. Unfortunately, quantitative formal methods necessarily have strong
arbitrary elements. However, a quantitative theory may have adequate qualitative features. The rest of this chapter deals with a qualitative theory of confirmation which is in agreement with the qualitative features of the corresponding
version of the Bayesian theory. The presentation and analysis of the Bayesian
theory, and the proof that it has these features, will be postponed to the next
chapter. Since the main ideas of Carnap and Hintikka about confirmation
concern a special form of the Bayesian theory (see Chapter 4), there is also
qualitative agreement with them. Finally, there is largely agreement with
Popper, for it will be argued in Appendix 1 of Chapter 3 that Popper's

CONFIRMATION BY THE HD-METHOD

21

qualitative ideas about confirmation (or corroboration, to use his favorite term)
are basically in agreement with the Bayesian approach.
However, there is at least one influential author about confirmation, viz.,
Glymour, whose ideas do not seem to be compatible with the qualitative theory
to be presented. In Note 11 to Subsection 2.2.3. we will briefly discuss Glymour's
project, and question it as a project of explication of the concept of confirmation, rather than of theoretical measurement, which it surely is.
Guided by the success perspective on confirmation, Section 2.1. gives an
encompassing qualitative theory of deductive confirmation by adding a comparative supplement to the classificatory basis of deductive confirmation. The
comparative supplement consists of two principles. In Section 2.2. several
classical problems of confirmation are dealt with. First, it is shown that the
theory has a plausible solution of Hempel's raven paradoxes, roughly, but not
in detail, in agreement with an influential Bayesian account. Second, it is shown
that the theory has a plausible, and instructive, solution of Goodman's problem
with 'grue' emeralds. Third, it is argued that further arguments against 'deductive confirmation' do not apply when conditional deductive confirmation is
also taken into account. Section 2.3. concludes with some remarks about the
problem of acceptance of well-confirmed hypotheses.
2.1. A QUALITATIVE THEORY OF DEDUCTIVE CONFIRMATION
Introduction

Traditionally, a qualitative theory of deductive confirmation is conceived as


merely a classificatory theory. However, we conceive it from the outset as a
combination of classificatory and comparative principles. In general, in this
book the term 'qualitative theory' should not be interpreted as referring to
merely a classificatory theory, but to a non-quantitative theory with classificatory and comparative aspects. We will use this terminology also for the theories
of theory evaluation and truth approximation to be presented in later chapters.
Regarding the qualitative theory of confirmation, the subject of this chapter,
we start with its classificatory principles.
2.1.1. Classificatory basis

The classificatory notions of evaluation of a hypothesis H by evidence E, are


generated by considering the four possible, mutually exclusive (except for some
extreme cases), deductive relations (1=) in a systematic order, assuming that H
is consistent and E is true. They are depicted in the following (Deductive)
Confirmation Matrix (Figure 2.1.). The names will first be elucidated. To speak
of falsification of H by E in the case H entails -, E, H 1= -, E, which is equivalent
to E 1= -, H, and of (conclusive) verification of H by E in the case -, H entails
-, E, -, H 1= -, E, which is equivalent to E 1= H, will be plausible enough. The
other names are suggested by the success perspective on confirmation, that is,

22

'H

CONFIRMATION BY THE HD-METHOD

E (true)

'E (false)

HI=E ('E/='H)

H/='E (EI='H)

Deductive Confirmation

Falsification

DC(E,H)

F(E,H)

'Ht=E ('E I=H)


Deductive Disconfirmation

'H /='E (E F=f1)


Verification

DD(E,H)

V(E,H)

Figure 2.1. The (Deductive) Confirmation Matrix

confirmation amounts to a success of a hypothesis. The first case, H FE, is a


paradigm case in which scientists speak of confirmation, viz., when E is a
(hypothetico-)deductive success of H, for that reason called (the principle of)
deductive (d-)confirmation. Note that only in this case E is a test implication
of H (which came true). Just as verification (of H) amounts to falsification of
--, H, the remaining case, --, H F E, amounts to a deductive success of --, H,
hence to deductive confirmation of --, H. However, speaking, by way of shorthand, precisely in this case of deductive disconfirmation seems also very plausible. Hence, whereas philosophers and scientists sometimes use the term
disconfirmation as an euphemism for falsification, we use the formal opportunity to qualitatively distinguish sharply between (deductive) disconfirmation
and falsification.
The Confirmation Matrix specifies the basic cognitive structure governing
hypothesis testing. However, although it gives the core of the classificatory part
of the qualitative success theory of confirmation, it does not yet reflect standard
scientific practice in detail. Recall that there occur as premises in hypotheticodeductive (HD-)prediction and deductive-nomological (DN-)explanation of
individual events, besides the hypothesis, the so-called 'initial conditions'. The
above definition of 'deductive confirmation', as confirmation by a deductive
success, directly suggests the notion of 'conditional deductive (cd-)confirmation',
i.e., confirmation of H by E, assuming a certain condition C, defined as follows:
H &C F E conditional Deductive Confirmation:

DC(E,H; C)

Here it is assumed that Hand C are logically independent (U(H, C, for


reasons becoming clear in Subsection 2.2.3., and that C does not entail E,

CONFIRMATION BY THE HD-METHOD

23

whereas C mayor may not be entailed by E. Hence, in view of the fact that
hypotheses and initial conditions usually are logically independent, successful
HD-prediction and ON-explanation of individual events form the paradigm
cases of cd-confirmation, for they report 'conditional deductive successes'.2
When dealing with specific examples, several expressions related to
cd-confirmation will be used with the following interpretation:

"E C-confirms H": E cd-confirms H on the condition (indicated by) C,


"E cd-confirms H": there is a (non-tautological) condition C, such that
E entails C and E C-confirms H,
"E is deductively (d-)neutral (evidence) for H": Hand E are logically
independent (i.e., none of the four deductive relations holds),
"E is conditional deductively (cd-)neutral (evidence) for H": E is d-neutral
for H and in addition E does not cd-confirm H nor -, H.
Note that 'cd-neutrality' trivially implies 'd-neutrality', but not vice versa.
Of course, 'conditionalization' of the other three deductive relations between
E and H, by just adding C as a premise, is also possible and more realistic,
leading to the corresponding conditional (Deductive) Confirmation Matrix.

2.1.2. Comparative supplement


As announced, we explicate 'confirmation by the HD-method', in this chapter,
as (conditional) deductive confirmation including some crucial comparative
principles. However, in order to do so it is useful to introduce first some main
principles of general confirmation, that is, confirmation of deductive and nondeductive nature. The success perspective on confirmation in general may be
explicated by the following definition of (general) confirmation:

SDC:

Success definition of cmifirmation:


E confirms H iff (E is a success of H in the sense that) H makes
E more plausible

This definition is satisfied by d-confirmation in the extreme sense that E is


made true, hence maximally plausible, by (the truth of) H. The second principle
deals with increasing the plausibility of hypotheses by confirming evidence:

RP P:

Reward principle of plausibility


E makes H more plausible 3 iff E confirms H

For d-confirmation the reward principle implies an increase of the plausibility


of the hypothesis.
Note that the two general principles imply the

PS:

Principle of symmetry
H makes E more plausible iff E makes H more plausible

Note, moreover, that the combination of the first two principles makes
comparative expressions of the form "E* confirms H* more than E confirms

24

CONFIRMATION BY THE HD-METHOD

H" essentially ambiguous. It can express that the plausibility increase of E* by


H* is higher than that of E by H, or that the plausibility increase of H* by E*
is higher than that of H by E. However, we exclude this ambiguity by adopting
as the third and last general principle, now of a comparative nature:

pes:

Principle of comparative symmetry


E* confirms H* (much) more than (as much as) E confirms H
iff
H* increases the plausibility of E* (much) more than (as much
as) H increases the plausibility of E
iff
E* increases the plausibility of H* (much) more than (as much
as) E increases the plausibility of H

For our purposes, the three principles SOC, RPP (together implying PS) and
PCS of general confirmation are sufficient. From now on in this chapter,
comparative confirmation claims will always pertain to (conditional) deductive
success and (conditional) deductive confirmation.
Now we are able to propose two comparative principles concerning
d-confirmation, one (P.l) for comparing two different pieces of evidence with
respect to the same hypothesis, and one (P.2) for comparing two different
hypotheses in the light of the same piece of evidence.
P.l
P.2

if E and E* d-confirm H then Ed-confirms H more than E* iff E


is less plausible than E* in the light of the background beliefs
if Ed-confirms Hand H* then Ed-confirms H* as much as H

To be sure, P.l and P.2 are rather vague, but we will see that they have
some plausible applications, called the special principles. Hence, if one finds
the general principles too vague, it is suggested that one primarily judges the
special principles.
The motivation of the plausibility principles is twofold. First, our claim is
that P.l and P.2 are roughly in agreement with scientific common sense
concerning confirmation by successes obtained from HO-tests. For P.l this is
obvious: less expected evidence has more 'confirmation value' than more
expected evidence. P.2 amounts to the claim that hypotheses should be equally
praised for the same success, which is pretty much in agreement with (at least
one version of) scientific common sense.
The second motivation of the principles P.I and P.2 will be postponed to
the next chapter, where we will show that they result, with some qualifications,
from certain quantitative confirmation considerations of a Bayesian nature
applied to HO-tests, Bayesian considerations for short, when 'more plausible'
is equated with 'higher probability'.
P.2 amounts to the claim that the 'amount of confirmation', more specifically,
the increase of plausibility by evidence E is independent of differences in initial
plausibility between Hand H* in the light of the background beliefs. In the

CONFIRMATION BY THE HD-METHOD

25

next chapter we will also deal with some quantitative theories of confirmation
for which holds that a more probable hypothesis profits more than a less
probable one. This may be seen as a methodological version of the so-called
Matthew-effect, according to which the rich profit more than the poor. It is
important to note, however, that P.2 does not deny that the resulting posterior
plausibility of a more plausible hypothesis is higher than that of a less plausible
one, but the increase does not depend on the initial plausibility. For this reason,
P.2 and the resulting theory are called neutral with respect to equally successful
hypotheses or, simply, pure4 , whereas theories having the Matthew-effect, or
the reverse effect, are called impure. The first type of impure theories are said
to favor plausible hypotheses, whereas theories of the second type favor implausible hypotheses. In Subsection 2.2.1 . of the next chapter, we will give an urnmodel, hence objective probabilistic, illustration and defence of P.2.
The condition 'in the light of the background beliefs' in P.l is in line with
the modern view (e.g., Sober 1988, p. 60, Sober 1995, p. 200) that confirmation
is a three place relation between evidence, hypothesis and background beliefs,
since without the latter it is frequently impossible to make differences between
the strength of confirmation claims. As a matter of fact, it would have been
better to include the background beliefs explicitly in all formal representations,
but we have refrained from doing so to make reading the formulas more easy.
This does not mean that background beliefs always playa role. For instance,
they do not playa role in the following two applications of the principles, that
is, the first two special principles (for non-equivalent E and E*):
S.l
S.2

if H 1= E 1= E* then Ed-confirms H more than E*


if H* 1= H 1= E then Ed-confirms H* as much as H

These special principles obviously are applications of the corresponding general


principles. For S.2 it is trivial and for S.l it is only necessary to assume that
(logically) weaker evidence is more plausible than stronger evidence. Hence,
they can be motivated in the same two ways as the general principles themselves:
indirectly, by showing later that they result from certain Bayesian considerations as soon as 'logically weaker' guarantees a higher prior probability, and
directly, by referring to scientific common sense around HD-tests. S.l states,
in line with the success perspective, that a (logically) stronger deductive success
of a hypothesis confirms that hypothesis more than a weaker success. S.2 states
that a logically stronger hypothesis is as much confirmed by a deductive success
as a weaker one which shares that success.
Since S.l and S.2 directly pertain to the rebuttal of two standard objections
in the literature against deductive confirmation, we will deal already now with
these objections, but postpone the treatment of some other ones to
Subsection 2.2.3. It is easily checked that deductive confirmation has the
so-called 'converse consequence property' with respect to the hypothesis (CCH) and the 'consequence property' with respect to the evidence (C-E). The first
property amounts to:

26

CONFIRMATION BY THE HD-METHOD

CC-H: converse consequence property with respect to H:


if Ed-confirms H then it also d-confirms any stronger H*
This property generates the 'irrelevant conjunction objection', i.e., CC-H has
the prima facie absurd consequence that if H F E then Ed-confirms H &H', for
any H', which is compatible with H, but not entailed by H (Hempel 1945jI965);
Glymour 1980ajb). From the cIassificatory-cum-comparative point of view, this
consequence is not at all absurd as soon as we are aware of all relevant
(qualitative) confirmation aspects, which we like to call the proper (conjunction)
connotation:
if
-

Ed-confirms H then
Ed-confirms H&H', for any H' compatible with H
even as much as H (due to S.2)
but E does not at all necessarily d-confirm H'
hence, the d-confirmation remains perfectly localizable

If one finds it strange that E confirms H&H' as much as H, it is important to


realize that the prior plausibility of H&H' will be less than that of H, and
hence, because of PCS and S.2, this will hold for the corresponding posterior
plausibilities. Hence, if H' is nonsensical or otherwise implausible, e.g., the
moon is made of cheese, H&H' will be implausible as well, a priori as well as
a posteriori. For this reason, nonsensical additions to empirical hypotheses are
relatively harmless.
The situation is to some extent similar for the standard objection against
the second property, C-E, the consequence property with respect to the evidence. This property amounts to:

C-E: consequence property with respect to the evidence:


if Ed-confirms H, then also any weaker E*
This property generates what might be called the "irrelevant disjunction objection", i.e., C-E has the prima facie absurd consequence that if H F E then EvE'
d-confirms H, for any E' (Grimes 1990). Again, from our point of view this
consequence is not at all absurd as soon as we are aware of all relevant
(qualitative) confirmation aspects, i.e., the proper (disjunction) connotation:
if Ed-confirms H then
- EVE' d-confirms H, for any E' compatible with E
- though (much) less (due to S.l)
- but E' does not at all necessarily d-confirm H
- hence, the d-confirmation remains perfectly localizable
It will be useful to conditionalize the general and special principles:

P.lc

if E C-confirms Hand E* C*-confirms H then E C-confirms H


more than E* C*-confirms H iff E* is, given C*, more plausible
than E, given C, in the light of the background beliefs

CONFIRMATION BY THE HD-METHOD

S.lc
P.2c
S.2c

27

if H FC -+ E FC* -+ E* then E C-confirms H more than E* C*confirms H


if E C-confirms Hand H* then E C-confirms H* as much as H
if H F C -+ E, H* F C -+ E and H* F H then E C-confirms H* as
much as H

Later we will introduce two other applications of the principles, more specifically of P.lc and P.2c, or, if you want, new special principles. They will provide
the coping-stones for the solution of the raven paradoxes and the grue problem.
In contrast to S.l and S.2, they presuppose background beliefs and concern
conditional (deductive) confirmation. Moreover, they deal with special types of
hypotheses, pieces of evidence and conditions. They will be indicated by
S#.lc(-ravens) and SQ.2c(-emeralds), respectively.
2.2. RAVENS, GRUE EMERALDS, AND OTHER PROBLEMS AND
SOLUTIONS

Introduction
In this section it will be shown how the qualitative theory of deductive confirmation resolves the famous paradoxes of confirmation presented by Hempel and
Goodman. Moreover, we will deal with the main types of criticism in the
literature against the purely classificatory theory of deductive confirmation,
that is, when a comparative supplement is absent.
In view of the length of the, relatively isolated, subsection (2.2.2.) about
Goodman's grue problem, the reader may well decide to skip that subsection
in first reading.

2.2.1. The raven paradoxes


Hempel (1945/ 1965) discovered two paradoxes about confirmation of the
hypothesis "all ravens are black" (RH) on the basis of two prima facie very
plausible conditions. Assuming that a black raven confirms RH (so-called
Nicod's criterion (NC) and that the logical formulation of RH may not matter
(equivalence condition (EC), Hempel derived not only that a non-black nonraven (the first paradox), but, even more counter-intuitive, also a black nonraven (the second paradox) confirms RH in the same sense as a black raven.
First, according to NC, a non-black non-raven confirms "all non-blacks objects
are non-ravens" and hence, according to EC, RH itself. Second, again according
to NC, a black non-raven confirms "all objects are non-ravens or black" and
hence, according to EC, RH itself.
It is easy to check that the previous section leads to the following classificatory results:
(1) a black raven, a non-black non-raven and a black non-raven are all
deductively (d-)neutral evidence for RH

28

CONFIRMATION BY THE HD-METHOD

(2) a black raven and a non-black non-raven both cd-confirm RH,


more specifically, a black raven on the condition of being a raven
and a non-black non-raven on the condition of being non-black
(3) a black non-raven is even cd-neutral evidence for RH, i.e., not only
just deductively, but even conditional deductively, for a black nonraven does not cd-confirm RH on the condition of being a nonraven, nor on the condition of being black.
All three results are in agreement with scientific common sense, provided the
following comparative claim can also be justified:
(4) a black raven cd-confirms RH (much) more than a non-black
non-raven
However, assuming that the background beliefs include or imply the
assumption:
A-ravens:

the number of ravens is much smaller than the number of


non-black objects

the desired result, i.e., (4), immediately follows from the third special principle,
viz., the following general application of P.lc. Though the symbols are suggested
by the raven example (e.g., an RB may represent a black raven, i.e., a raven
which is black), they can get any other interpretation. Let #R indicate the
number of R's, etc.
S#.lc(-ravens):

an RB R-confirms "all Rare B" more than an R8


8-confirms it iff the background beliefs imply that
#R<#8

S#.lc realizes in a precise sense the intuition that a black raven confirms "all
ravens are black" (much) more than a non-black non-raven. That S#.lc is an
application of P.lc can be shown as follows. If #R < #8 and "all Rare B" is
false, then the percentage of RB's among the R's is lower than the percentage
of R8's among the 8's, hence hitting among the R's at an R which is B is less
plausible (hence more surprising) than hitting among the 8's at a 8 which is
R. For example, even in the extreme case of just one non-black raven, the first
percentage is #R/(#R + 1), which is indeed smaller than the second percentage,
#8/(#8 + 1), if and only if #R is smaller than #8. In other words, the higher
the percentage of individuals with a certain trait in a population, the more it
is (to be) expected that it applies to an arbitrary member, where only the
comparison of percentages is relevant, and not the trait nor the (size of the)
population. 5 Hence, SH.lc can be motivated by referring directly to scientific
common sense, but also indirectly by showing later that it results from Bayesian
considerations when the evidence is assumed to be obtained by random sampling in the relevant universe. Of course, we speak of much more confirmation
in SH.lc when the background beliefs imply that #R is much smaller than #8
(#R #8), as in the case of A-ravens, and hence, if RH is false, the percentage

CONFIRMATION BY THE HD-METHOD

29

of RB's among the R's is (relatively speaking) much lower than the percentage
of RB's among the R's. Precisely because typical applications of S#.lc concern
such cases, it is defensible to call it a qualitative application and principle,
despite its explicit reference to numbers of individuals and the reference to
percentages in the motivation.6
In sum, cd-confirmation solves the raven paradox concerning black nonravens by (3) and the one concerning non-black non-ravens by (4), which is
guaranteed by applying P.lc, or its application S#.lc-ravens, to the background
assumption (A-ravens) that the number of ravens is much smaller than the
number of non-black objects.
There remains the question of what to think of Hempel's principles used to
derive the paradoxes of confirmation. It is clear that the equivalence condition
was not the problem, but Nicod's criterion that a black raven confirms RH
unconditionally. Whereas Nicod's criterion is usually renounced unconditionally, we may conclude that it is (only) right in a sophisticated sense: a black
raven is a case of cd-confirmation, viz., on the condition of being a raven.
2.2.2. Grue emeralds

The qualitative theory of deductive confirmation generates an instructive analysis of the other famous riddle of confirmation, i.e., the problem with 'grue'
emeralds, discovered by Goodman (1955). This problem is also called the grue
'paradox', for the same reason as one speaks about the raven paradoxes. They
both concern counter-intuitive consequences of certain principles of
confirmation.
The problem with grue emeralds is that a green emerald found before the
year 3000 seems to confirm not only the hypothesis that all emeralds are green
but also that all emeralds are 'grue', where grue is defined as the following
queer predicate: green if examined before 3000, and blue if not examined before
3000. Goodman's generally accepted account roughly is as follows: the predicate
'grue' is not weB-entrenched in predictively successful scientific generalizations,
hence, as it stands, it is below the mark of scientific respectability to be used
in generalizations that can be confirmed, i.e., to use Goodman's other favorite
expression, the 'grue-hypothesis' is not (yet) projectible. We will give a related,
but more detailed diagnosis of the problematic aspect. It may well be conceived
as a formal explication and justification of Goodman's informal account. It
can best be presented by using from time to time a formally similar, but less
queer, definition of 'grue', which is only applicable to living beings, say, eagles:
'being male and green, or being female and blue'. This wiB be called the 'gender'
reading, as opposed to the former 'temporal' reading.
Recall that we use the abbreviation: " .. . C-confirms ... " as a shorthand for
"... cd-confirms ... on the condition (indicated by) C". We add the abbreviations:
E: emerald/eagle (in this subsection not to be confused with 'evidence'), M:
(known to be) examined before 3000/male, M: not (known to be) examined
before 3OOO/female, G: green, B: blue, and Q: grue (queer), i.e., MG or MB. G

30

CONFIRMATION BY THE HD-METHOD

and B are supposed to be mutually exclusive, but they are not supposed to be
exhaustive.
We will first specify in detail to what extent 'green' and 'grue' are similar,
and show that additional assumptions are needed to create the intuitively
desirable asymmetry. More specifically, it will be shown that not only a strong,
but also a weak irrelevance assumption is suitable for this purpose. Both are
in line with the entrenchment analysis of Goodman.
The basic intuition

The basic intuition of Goodman is, of course, that, though a green emerald
investigated before 3000 confirms 'the green hypothesis' ("all E are G"), it does
not confirm 'the grue hypothesis' ("all E are Q"). It postulates an asymmetry
in confirmation behavior between 'green' and 'grue'. However, from the unconditional version of Nicod's criterion, 'Nicod-confirmation', we not only get
(1) an EMG Nicod-confirms "all E are G"

but also, as is easy to check,


(2) an EMG Nicod-confirms "all E are Q"
Hence Nicod's criterion excludes an account of the asymmetric intuition, which
shows an additional problematic feature of that criterion.
As may be expected, from the straightforward, that is, unconditional deductive point of view, both cases are invalid, for in both cases the evidence is
deductively neutral for the hypothesis. Hence, unconditional deductive confirmation also fails to account for the basic intuition, but it does not exclude
a 'conditional deductive' account. So, what about conditional deductive
confirmation?
Cd-confirmation, however, is also in conflict with the intuition, for it is easy
to check that the following classifications obtain:
(3) an EMG EM-confirms "all EM are G"
(4) an EMG EM-confirms "all E are G"
(5) an EMG EM-confirms "all E are Q"
Note first that (3) is an unproblematic formal analogue of the equally valid
claim that an EG E-confirms "all E are G", where E is replaced by EM. More
importantly, whereas (4) fits the basic intuition, (5) is in conflict with it.
However, before rejecting cd-confirmation because of (5), it is important to
study the validity of (5) in detail. To begin with, it is important to note that
the following claim is invalid:
(6*) an EMG EM-confirms "all EM are B"
The invalidity of (6*) amounts to the claim that a green object does not confirm
the hypothesis "all EM are B" on the condition that it is an emerald investigated
before 3000. In this light, we suggest that the problem with (5) derives from

CONFIRMATION BY THE HD-METHOD

31

the wrong impression that it implies (6*) and the intuition that it would indeed
be absurd if (6*) were to obtain. In other words, we take the invalidity of (6*)
as the proper interpretation of the confirmation-rejecting side of the basic
intuition, instead of the originally, but wrongly, suggested invalidity of (5).
It is interesting to see in more detail how (5), (6*), and (3) are related. Note
that the following equivalence holds:
(7) "all E are Q".;;. "all EM are G" & "all EM are B"

which makes clear, by the way, that


the temporal reading suggests, for in
of two descent hypotheses, viz., "all
eagles are blue". In view of (7), (5) is

"all E are Q" may be less strange than


the gender reading it is the conjunction
male eagles are green" and "all female
equivalent to:

(5) an EMG EM-confirms "all EM are G" & "all EM are B"
Now it is easy to see that (3), (5) and (6*) form an illustration of the conditional
version of the 'proper conjunction connotation' of Section 2.l. That is, whereas
cd-confirmation according to (3) transmits to a conjunction with some other
hypothesis according to (the equivalent version of) (5), and the latter confirmation is, according to S.2c, even as much as the former, no confirmation
transmits to the added conjunctive hypothesis, for (6*) is invalid.
The question why (6*) would be absurd if valid, however, remains interesting.
Is it, prima facie in line with Goodman's entrenchment considerations, because
it amounts to a surprising prediction across some clearly defined border (a
year, gender), breaking the continuity of nature? In this case, it would be
plausible to expect that the additional hypothesis suggested by continuity
considerations, viz., "all EM are G", is EM-confirmed by an EMG, since G is
well-entrenched, and the trouble with (5) would merely be caused by the queer,
non-entrenched character of Q. Or is it because the 'grue-induced' additional
hypothesis "all EM are B" reaches incautiously over a border that might be a
relevant distinction? In this case also the 'green-induced' additional hypothesis
should not be EM-confirmed by an EMG, and the usual, but wrong, assumption
that this is implied by (4), is brought to light by the queer predicate.
It is easy to check that the second option is the proper answer from the
perspective of cd-confirmation. That is,
(8*) an EMG EM-confirms "all EM are G"
is invalid, for the same reason as (6*): the antecedence EM of the hypothesis
'cannot be put to work' by the condition EM to derive G and B, respectively.
This suggests that (8*) can also provide an illustration of the proper conjunction
connotation. Note, for this purpose, that the following equivalence obtains:
(9) "all E are G".;;. "all EM are G" & "all EM are G"
and hence that (4) is equivalent to
(4) an EMG EM-confirms "all EM are G" & "all EM are G"

32

CONFIRMATION BY THE HD-METHOD

Accordingly, in spite of the validity of (3) and (4), (8*) is invalid, precisely for
the same reason that (6*) is invalid as opposed to (3) and (5), viz., being
another triple of instances of the proper conjunction connotation. Conditional
deductive confirmation (3) transmits to a conjunction with some other hypothesis (4) and this confirmation is as much as that of (3), according to S.2c, but
the confirmation does not transmit to the added conjunctive hypothesis (8*).
In sum, the prima facie absurdity of (5) has a hidden analogue in (4), which
is also due to improper connotations. The invalidity of (8*) is a formal blockade
against confirmation claims that cross a border that may be relevant:
cd-confirmation blockade: for all E, M and G, although an EMG

EM -confirms "all EM are G" (3) and even "all E are G" (4), it does
not EM-confirm "all EM are G" (8*).
Whereas the blockade may seem superfluous in the temporal reading, it is clear
that, for instance, in the gender reading it is highly plausible and desirable. A
green male eagle does not (cd-)confirm the hypothesis that all female eagles
are green. Awareness of the blockade is, for instance, expressed in the feminist
criticism of male oriented drug research. Results of testing drugs on male
subjects have frequently been extrapolated to women in an irresponsible way
(see e.g., Cotton 1990; Ray et al. 1993).
So far, however, the results of the conditional deductive perspective are
symmetric with respect to green and grue: (4) and (5) on the one hand, and
(6*) and (8*) on the other. Moreover, (3) is a common feature of both. Hence,
the question remains to account for the asymmetric basic intuition. The foregoing analysis shows that an additional assumption is needed to create an
asymmetric situation.
There are at least two possible ways. In the first way, a hypothesis is added
which removes the blockade, at the expense of the grue hypothesis. In the
second way, the blockade is not removed, but the grue hypothesis is downgraded, without excluding it.
Asymmetry by an extra assumption

According to the first way, we explicitly assume, as an extra hypothesis, that


the border is irrelevant for the kind of properties at stake, that is, that they
can be extrapolated across that border. In the temporal reading of M this
amounts to adding a strong irrelevance assumption of time for the color of
emeralds, formally:
SIA(-emeralds):

for all colors C, "all EM are C" implies "all E are C",
and hence "all EM are C", and vice versa.

This assumption may well be a background belief. The 'implication' in SIA is


stronger than just a material implication and weaker than a purely logical
implication. In particular, it may be an analytical (or semantical) implication.
However, it may also be a 'physical implication', in the sense of a physical

CONFIRMATION BY THE HD-METHOD

33

necessity. In the latter case, it may be based on the underlying background


belief that emeralds constitute a natural kind with respect to color. Note first
that SIA, in view of (7), excludes the grue hypothesis "all E are Q". Moreover,
an EMG now not only EM-confirms "all EM are G" (3), and hence also "all
E are G" (4), but even "all EM are G", i.e, the adapted version of (8*)
becomes valid:
(8-SIA)

an EMG (SIA & EM)-confirms "all EM are G"

This is in agreement with the intuition behind the grue problem that the
artificial time barrier will not change the color. However, according to the
cd-analysis, this is only guaranteed when we take this formally into account
by the auxiliary assumption SIA. In conjunction with SIA it is even guaranteed
that an EMG not only falsifies "all EM are B", and hence "all E are B", but
also "all EM are B", for the latter and SIA now imply "all E are B", and hence,
in view of (7), an EMG falsifies "all E are Q". Hence, instead of the invalid
confirmation claim (6*) that an EMG EM-confirms "all EM are B", we may
now even conclude that an EMG 'SIA-falsifies' that hypothesis in the plausible
sense that (EMG & SIA) is incompatible with the hypothesis:
(6-SIA)

an EMG SIA-falsifies "all EM are B"

Accordingly, one way to achieve an asymmetry between green and blue in


the line of the cd-perspective is by assuming SIA. This is very much in the
spirit of Goodman's entrenchment analysis in terms of so-called projectible
predicates7 However, whereas Goodman's notion remains rather vague, the
required irrelevance assumption to remove the surprising blockade highlighted
by the invalid (8*), and giving rise to the valid confirmation claim (8-SIA), is
crystal clear. At the same time, it allows the replacement of the appealing
blockade reported by the invalid (6*) even by the valid falsification claim
(6-SIA). Hence, we have obtained an asymmetric (cd-)explication of Goodman's
basic intuition and a dichotomous reading of his entrenchment analysis.
However, this is not satisfactory in all cases. In the gender reading, we do not
want to exclude relevance of sex for color, although we may have good reasons
to find irrelevance more likely than relevance. For this reason, it is plausible
to look for a possible refinement of the basic intuition, which also leads to
asymmetry, without totally excluding the grue hypothesis from the confirmation
game. For, as is clear from the gender reading of the grue hypothesis, there
may well be formally similar cases where the exclusion is highly debatable.
Asymmetry by refinement of the basic intuition

The following refinement of Goodman's basic intuition is (the qualitative


analogue of a quantitative refinement) inspired by Sober (1994). Reconsider
first:
(4) an EMG EM-confirms "all E are G"

34

CONFIRMATION BY THE HD-METHOD

(5) an EMG EM-confirms "all E are Q"


which are equivalent to
(4) an EMG EM-confirms "all EM are G" & "all EM are G"
(5) an EMG EM-confirms "all EM are G" & "all EM are B"
The prima facie version of the basic intuition stated that, whereas (4) is correct,
(5) is problematic. Above we have seen that (5) is not problematic, but also
that (6*) would be problematic, if valid, which it is not. This creates the room
for the following refinement of the basic intuition: a green emerald investigated
before 3000 may confirm "all E are Q" as well as "all E are G", even as much
as, but the resulting plausibility of the former is (remains) much lower than
that of the latter, i.e., in the relevant conditional sense:
(4&5) although an EMG EM-confirms "all E are Q" as much as "all
E are G", the resulting plausibility of the former is much less
than that of the latter
In order to justify this refined intuition we introduce the fourth special
principle, a general application of P.2c, with symbols suggesting the example,
but of course intended for general use (as in the case of SH.lc):
SQ.2c(-emeralds):

an EMG EM-confirms "all E are Q" as much as "all


E are G"

It is clear that SQ.2c is a straightforward application of P.2c, using the fact that
an EMG EM-confirms both hypotheses. Moreover, in the emerald version, it
is safe to assume as background belief the following weak irrelevance
assumption:

WIA(-emeralds):

for all colors C and C', C =f: C', "all E are C" is (much)
more plausible than the conjunction "all EM are C"
& "all EM are C'" (which is equivalent to "all E are
Q" when C = G and C' = B)

In view of the fact that (4) and (5) hold, we may apply the general principle
of comparative symmetry (PCS) and P.2c, or its application SQ.2c, to our
background belief WI A, which directly leads to the asymmetric cd-explication
of the refined intuition, that is, (4&5).
Note that SIA implies WIA as soon as we assume that SIA amounts to the
implication that grue-like hypothesis lack any plausibility, whereas green-like
hypothesis have at least some plausibility. Hence, in the light of SIA, (4&5)
provides an asymmetry additional to that between (6-SIA) and (8-SIA).
In the gender reading, however, only WIA may have some plausibility, but
not SIA. That is, it may well be that we would like to subscribe to the
background belief that the green hypothesis is (much) more plausible than the
grue hypothesis, without excluding the latter. The reason would be, of course,
that a systematic color difference between the sexes of a species regularly

CONFIRMATION BY THE HD-METHOD

35

occurs, though supposedly not as frequently as sex irrelevance for color. Even
in the temporal reading, WIA is defensible, and SIA not. It is surely the case
that, as far as we know, there are no types of stones that have changed color
at a certain moment in history. However, this does not exclude the possibility
that this might happen at a certain time for a certain type of stone, by some
cosmic event. To be sure, given what we know, any hypothesis which presupposes the color change is much less plausible than any hypothesis which
does not.
It is important to note that we cannot simply take as an assumption that
"all E are G" is more plausible than "all E are Q", that is, without reference
to background beliefs, but only with the motivation that the former generalization expresses more uniformity or continuity of nature than the latter. That is,
it may seem that adding "all EM are G" to the common generalization of "all
E are G" and "all E are Q", viz., "all EM are G", giving rise to "all E are G",
is more in line with that common generalization than adding "all EM are B",
giving rise to "all E are Q". If one thinks this way one does so because one
assumes that EMG is more similar to EMG than EMB. However, this 'uniformity argument' hinges upon the particular E/ M/G-language. As is well-known
from the discussion of the grue problem (and of the problem of language
dependence of definitions of verisimilitude, see (Zwart 1995 such arguments
are language dependent. More specifically, in the E/ M/G-Ianguage the suggested uniformity argument would imply that "all E are G" is, for example,
more plausible than "for all E: M iff G" . However, in the E/ M/ X-language,
with X =defM iff G, the suggested uniformity argument would imply the
opposite plausibility claim, viz., that "all E are X" is more plausible than "for
all E: M iff X". Hence, reference to background beliefs is unavoidable.
In sum, the basic intuition can be justified in terms of cd-confirmation in
two ways. In the first way, a strong assumption is added which excludes the
grue hypothesis. In the second way, the basic intuition is refined by downgrading the grue hypothesis, without excluding it. Both are very much in the spirit
of Goodman's entrenchment approach. Although Goodman's specific example
of grue primarily suggests the first, dichotomous way, since green is, and grue
is not, well-entrenched (in the temporal reading). However, his general exposition in terms of more or less entrenched predicates primarily suggests the
second, gradual way. As has been noted, the first way is a kind of extreme
version of the second. 8 Hence, the above analysis is highly congenial to
Goodman's informal account in terms of entrenchment. However, the
cd-analysis localizes in formal detail the symmetric point of departure for two
asymmetric explications, a stronger and a weaker one.
2.2.3. Objections to (conditional) deductive confirmation

In the literature several objections have been expressed to the very idea
of (conditional) deductive confirmation. Hence, our specific account of
(un-)conditional deductive confirmation is also subject to them. The reader is

36

CONFIRMATION BY THE HD-METHOD

invited to himself evaluate our rebuttals. We begin with unconditional deductive


confirmation. Two of the three standard objections against this idea have
already been dealt with in Subsection 2.1.2., viz., this type of confirmation is
transmitted to a stronger hypothesis (CC-H), and to weaker evidence (C-E),
leading to the two 'proper connotations' which also take comparative claims
into account.
An important remaining objection against unconditional deductive confirmation is that d-confirmation is not transmitted to a weaker hypothesis (lacks
the consequence property with respect to the hypothesis). Hempel (1945/1965)
and others point at the supposed intuition among scientists that confirmation
transmits to consequences of the hypothesis. From the success perspective on
deductive confirmation, i.e., deductive confirmation as deductive success, this
objection also disappears, for a consequence of H need not yield the success,
it even need not be co-responsible for the success. Hence, the straightforward
consequence property, viz., transmission of d-confirmation to a weakening of
the hypothesis, is invalid as well as implausible.
One might argue that a liberation of the property holds: d-confirmation of
H by E implies confirmation, not necessarily deductive, of any consequence of
H by E. However, if deductive confirmation were to have the liberated consequence property, any success, i.e., any established true consequence of a hypothesis would confirm any other consequence of it, even if the two consequences
use two non-overlapping (sub-)languages, and would hence not be recognized
as confirmationally related at all without the hypothesis. More generally,
together with the converse consequence property with respect to the hypothesis
(CC-H) of d-confirmation, the liberated property would have the absurd consequence that any two hypotheses with no overlapping vocabulary would always
be confirmationally related in the sense that any d-confirming evidence of the
one would be, via the conjunction of the two hypotheses, confirming evidence
of the other. In short, most scientists are well aware that a success usually is a
joint venture of most, if not all, of the components of the hypothesis. Hence, if
some scientists and philosophers nevertheless have also the intuition that
(deductive) confirmation transmits somehow to all consequences of the hypothesis, the foregoing analysis, dealing with deductive successes as the paradigm
cases of confirmation, provides an invitation to reconsider that connotation of
(deductive) confirmation seriously.9
As to the predictable objections against conditional deductive confirmation,
we start with a purely technical objection (Gemes 1990, and formally related,
Glymour 1980b). It is easy to check that the definition of DC(E, H; C), i.e.,
H &C F E, prima facie implies that, for any E and H, E cd-confirms H on a
condition that is entailed by E, hence true, viz., the condition H -+ E, i.e.,
DC(E, H; H -+ E), for trivially H &( H -+ E) FE. Hence, prima facie, any E
cd-confirms any H. However, we have added the requirement LI(H, C) to the
definition, i.e., the condition that C does not logically depend on H. Since not
only E but also.., H (trivially) entails the relevant condition H -+ E (.., H V E),

CONFIRMATION BY THE HD-METHOD

37

the purported trivialization of cd-confirmation has thus been excluded. Note


that the same component of U(H, C), that is, -, H may not entail C, is formally
required to prevent HO-prediction and ON-explanation of individual events
from similar trivialization. Of course, the other alternative is just to consider
these related trivial cases as illustrations of the trivialization phenomenon that
frequently occurs under logically extreme conditions. However, the other three
components of U(H, C) serve already to exclude more transparent improper
cases, viz., incompatible hypothesis and condition (H may not entail -, C), or
one of them being redundant (H may not entail C and C may not entail H),
so why not just include the fourth component?
Let us now turn to a prima facie more fundamental objection. First of all, it
is time to mention that the notion of cd-confirmation was already essentially
considered, and rejected, by Hempel (1945/ 1965) as a general criterion of
confirmation, under the heading 'prediction criterion'. Later, it was reconsidered
by Horwich (1983), and rejected as too narrow, which it of course is when
taken as pars pro toto, i.e., without taking unconditional confirmation, also
treated in this chapter, and non-deductive confirmation, treated in Chapter 3,
into account. Hempel argues that in interesting cases (he elaborates a case of
plane-polarized light) the plausible condition is already a universally quantified
statement, of which the acceptance presupposes, what he calls, a quasi-induction. Assuming moreover that the idea of quasi-induction presupposes the idea
of confirmation, he concludes that the predictive approach to confirmation
becomes circular. From our analysis it is clear, however, that not only unconditional deductive confirmation but also conditional deductive confirmation can,
without problems, be defined in general, and can be straightforwardly applied
to non-universal conditions. Hence, it may be true that the acceptance of a
universal condition requires criteria for the 'inductive jump' and it is even
plausible that these criteria will be phrased in terms of straightforward cases
of (conditional) deductive confirmation. However, it does not follow from this
that the analysis of (conditional) deductive confirmation is circular. lO
To conclude, we have argued that confirmation by the HO-method can be
explicated as (conditional) deductive confirmation, provided some comparative
principles are added to this classificatory point of departure. The resulting
theory not only leads to qualitative solutions of the raven paradoxes and the
grue problem, but can be defended against the objections normally raised
against the purely classificatory approach of qualitative confirmation as deductive confirmation.
As already suggested, many philosophers who are reserved about the possibility of an acceptable theory of deductive confirmation subscribe to (a version
of) the quantitative Bayesian theory of confirmation. As will be shown in the
next chapter, that theory perfectly leaves room for HO-testing of hypotheses
and a plausible specification of it implies the pure theory of confirmation
presented in this chapter. 11

38

CONFIRMATION BY THE HD-METHOD

2. 3. ACCEPTANCE OF HYPOTHESES

Let us, finally, briefly deal with the problem of the acceptance of hypotheses.
Explicating the idea of deductive confirmation of a hypothesis is one thing,
explicating the idea of being sufficiently confirmed to be accepted is another.
For the latter issue it is important to recognize that a hypothesis may be highly
confirmed, without having become very plausible, since it may have been very
implausible at the start. For acceptance we need something like 'being sufficiently confirmed to have become sufficiently plausible to be accepted'. If it
was already plausible at the start, the acquired confirmation may have been
not very important for this purpose. Hence, crucial is 'being sufficiently plausible
for being accepted'. Acceptance criteria may depend on the nature of the
hypothesis: is it of an individual or a general nature? does it contain theoretical
terms? etc. Moreover, they will depend on one's epistemological position, for
there are, of course, various types of being accepted, roughly corresponding to
the epistemological positions. To begin with the latter, for the realist acceptance
of a hypothesis amounts to accepting the hypothesis as literally true, including
its observational, referential and theoretical consequences. The referentialist
will drop the theoretical consequences, and the empiricist, in addition, the
referential consequences. Finally, the instrumentalist will drop the observational
consequences concerning non-intended applications. In all cases, acceptance of
a hypothesis means that it is to be added to the body of background beliefs,
of which the general status, of course, also depends on the relevant epistemological position.
Within each of the above mentioned types of accepting a hypothesis as true
we could also distinguish between at least five 'kinds of truth': true simpliciter,
approximately true, the truth, near to the truth, nearer to the truth than another
hypothesis. The above, epistemologically induced, qualifications were primarily
intended for the first and the third kind of truth, that is, 'true simpliciter' and
'the truth'. The fifth kind will be explicated for the various epistemological
types in great detail in later chapters. Although we have used, and will use,
informally expressions referring to the second and the fourth kind of truth,
such as 'approximately true' and 'near to the truth' themselves, we will not try
to give precise explications of them. It is clear that 'approximately true' would
need some conventional threshold for deviations from being true. Similarly,
'near to the truth' would need some threshold for being sufficiently near to the
truth. Of course, these five kinds have each two modes, the actual mode and
the nomic one. However, this will only become relevant in later chapters.
Turning to 'true simpliciter' and 'the truth', and assuming that the hypothesis
is of a general nature, with general test implications in different directions, and
using theoretical terms, the three positions beyond the instrumentalist one
(which we will further neglect as far as acceptance is concerned) have to make,
sooner or later, inductive jumps or, simply, inductions of at least four kinds 12 ,
that is, from a threshold plausibility for each of the distinguished epistemological senses to the corresponding type of truth. The empiricist has to make

CONFIRMATION BY THE HD-METHOD

39

elementary or first order inductive jumps, i.e., inductive generalizations (in


observation terms), and second order inductive jumps, generalizations of inductive generalizations. This distinction is not a very sharp one. Both are called
observational inductions, and may be aiming at 'observationally true simpliciter'
(weak) or 'the observational truth' (strong). The referentialist has to add referential inductions, and the realist, in addition, theoretical ones of a weak or even
strong nature. Hence, even if we neglect further refinements of referential and
theoretical inductive jumps, there are at least four different questions of explication and justification regarding 'true simpliciter' and four regarding 'the truth'.
To be sure, normally speaking, an inductive jump reflects a deductive fallacy.
The recent AI-literature has made it clear that deductive fallacies may be very
useful default rules of reasoning. The reader is referred to (Tan 1992) and
(Marek and Truszczynski 1993). Here, however, we prefer to go in another
direction. Unfortunately, there only seem to be debatable conditions for elementary inductive jumps, and similarly for the other three types. Be this as it may,
it will turn out to be fruitful to consider the non-elementary jumps in the
comparative perspective of empirical progress and truth approximation. In
Part II and III it will become clear that the conditions for accepting comparative
success and truth approximation claims are essentially of the same non-elementary nature. Hence, the assessment of empirical progress and truth approximation claims is fundamentally of the same nature as the assessment of certain
'truth simpliciter claims'.
Whether the inductive jump is of a separate or comparative nature and
although there are no undebatable conditions for 'being sufficiently confirmed
to have become sufficiently plausible to be accepted as true or even as the
truth', the two comparative principles of deductive confirmation are intuitively
appealing and very useful in guiding acceptance. For, assuming the three general
principles of confirmation, that is, the success definition (SOC), the reward
principle (RPC) and comparative symmetry (peS), and assuming some initial
plausibility of the relevant hypotheses, they provide the following guide lines.
The first principle tells us that we can increase the plausibility of a hypothesis
by producing 'deductive evidence' with decreasing initial plausibility, that is,
plausibility in the light of the background beliefs. The second principle tells us
that deductive evidence increases the plausibility of a weaker hypothesis as
much as that of a stronger one, and hence the initially more plausible weaker
hypothesis remains the more plausible one. Hence, whatever is considered as
'sufficiently plausible', the weaker of two equally confirmed hypotheses will
reach that threshold earlier than the stronger one. Moreover, assuming that
the stronger can be written as a conjunction of the weaker and an additional
conjunctive hypothesis, the latter's plausibility is only increased when it is
separately confirmed by the available evidence. This feature of (deductive)
confirmation safeguards the scientists against redundant or absurd additions.
The latter might suggest a direct plea for sticking to the aim of the constructive empiricist, that is, observational induction, for a theory cannot become

40

CONFIRMATION BY THE HD-METHOD

as plausible as its observational reduct, which shares, by definition, all its


deductive evidence. However, it is important to distinguish the different epistemological aims, with corresponding thresholds. Relative to the aim of the
theory realist, a weaker, but equally successful (unfalsified) theory is more
plausible than a stronger. Similarly for the referential realist and the constructive empiricist.
In this way we have reduced the space for the so-called under-determination
of theories by evidence. Two theories may be observationally equivalent (that
is, have the same observational consequences), and hence equally successful, as
long as the observational-theoretical distinction remains fixed . However, if the
initial plausibility of the two theories differs, e.g., because the one is just stronger
than the other, the posterior plausibility remains to differ, and hence, whatever
threshold, we will prefer the more plausible one. Of course, since plausibility
is basically determined by our background beliefs, this is an important second
way in which the background beliefs guide our preferences. However, there
remains the case where the background beliefs do not discriminate, that is, two
theories which are observationally equivalent and have the same (prior and
hence posterior) plausibility. Although duplication arguments, construing an
observationally equivalent theory, usually do not produce an equally plausible
alternative theory, it should be conceded that if the alternative is equally
plausible, we have no good reason for a preference. Hence, a theoretical
induction, that is, the provisional acceptance of a theory, as true or even as
the truth, should always be qualified by the addition 'or some other equally
plausible observationally equivalent alternative'. A similar qualification has to
be made in the case of referential induction. To be sure, the suggested qualifications are seldom more than of an academic nature since alternatives satisfying
the conditions are not available.

Concluding remarks
In the chapters to come, several new matters concerning hypothesis testing will
be dealt with. Chapter 3 mainly deals with a pure quantitative theory of
confirmation, briefly discusses the severity of tests. Chapter 5 will include a
detailed analysis of the derivation of test implications, and the complications
arising from, amongst others, auxiliary hypotheses. Moreover, testability will
be defended as a necessary condition for being an empirical hypothesis. At the
end of Chapter 7 we will have something to say on the nature and role of
novel facts, ad hoc repairments and crucial experiments.
In the course of these chapters it will become clear that the role offalsification
and confirmation has to be relativized in several respects. Prima facie falsification may be disputed in several ways, e.g., by questioning the description of
the counter-example or the truth of the auxiliary hypotheses needed to derive
the relevant test implication (Section 5.2.). More fundamentally, it will turn

CONFIRMATION BY THE HD-METHOD

41

out that 'being false', and 'being true' for that matter, is from the point of view
of empirical progress and truth approximation rather irrelevant, hence falsification will have to play a more modest role than frequently is assumed
(Section 6.2. and Chapter 7). We will also see that the realist may even claim
against the empiricist that one theory may be closer to the truth in the
encompassing theoretical sense than another, even though the first has some
counter-examples which are no counter-examples to the second (Chapter 9).
Similarly, the role of confirmation will be relativized along the same lines
and roughly at the same places. Since 'confirmation' has the connotation of
not yet being falsified, that is, the hypothesis may still be true, and since it will
turn out to make perfectly sense to continue the 'HD-evaluation' of a theory,
even though it has been falsified, the confirmation of a theory is not so
important, but the more general notion of obtaining (general) successes is very
important (Chapter 5 and 6). Prima facie successes may be disputed in similar
ways as prima facie falsification. Moreover, the obtainment of a success plays
a modest role similar to that of a counter-example. However, now the realist
cannot claim that a theory can be closer to the theoretical truth than another
despite the fact that the other has one or more extra successes.
Although the role of confirmation and falsification will be strongly relativized,
this does not mean that there is no need of a qualitative theory of deductive
confirmation and falsification, as developed in this chapter. On the contrary,
the notion of confirmation and falsification remain of crucial importance for
testing at least three types of hypotheses: (1) general test implications and
similar general observational hypotheses (Section 5.1.), (2) comparative success
hypotheses (Chapter 6), and (3) truth approximation hypotheses (Chapter 7,
9 and 10). Although it will turn out that testing truth approximation hypotheses
presupposes testing comparative success hypotheses and this on its turn testing
general test implications, it may, however, not be concluded that there is a
strong direct relation between confirmation and truth approximation. On the
contrary, as suggested before, it will turn out that there is no direct link between
'being true or false' and 'truth approximation'. This does not exclude that there
is some sophisticated link between confirmation and truth approximation, but
this will not be explored in this book (see Festa 1999a, Section 413 ) .
The three indicated remaining crucial roles of confirmation and falsification
highlight some main features of the short-term dynamics of science, that is, the
separate and comparative evaluation of theories, to be presented in Part II,
and the testing of truth approximation claims, to be presented in Part III and
IV Hence, we will have ample occasion to refer to the present chapter, but we
will only do so when it is particularly illuminating. However, as we have seen
in Section 2.3., there remains of course the possibility that, after a small or
large number of theory transitions along such evaluation lines, we have arrived
at a theory of which we may seriously conjecture that it is observationally,
referentially or even theoretically true (in a weak or strong sense, to be specified). Then confirmation and falsification guide again the choice between

42

CONFIRMATION BY THE HD-METHOD

rejection of that conjecture or acceptance, where the latter may concern observational, referential or theoretical induction, respectively, and may pave the
way for the long-term dynamics by providing the means for enlarging the set
of observation terms. In the process of sorting out theories with the same claim,
the background beliefs, determining the initial plausibility of evidence and
hypotheses, playa crucial role.
The main topics of the next two chapters concern a general quantitative
theory of confirmation and a survey of a coherent set of inductive specifications.
Since they will only playa marginal role in the remainder of this book, these
chapters could be skipped for the first reading, if one is not interested in the
very idea of quantitative confirmation.

3
QUANTITATIVE CONFIRMATION, AND ITS
QUALITATIVE CONSEQUENCES

Introduction
In the previous chapter, we have developed a qualitative (classificatory and
comparative) theory of deductive confirmation, guided by the success perspective. In this chapter we will present, in Section 3.1., the corresponding quantitative theory of confirmation, more specifically, the corresponding probabilistic
theory of confirmation of a Bayesian nature, with a decomposition in deductive
and non-deductive confirmation. It is again pure in the sense that all equally
successful hypotheses profit from their success to the same degree. It is inclusive
in the sense that it leaves room for confirmation of hypotheses with zero
probability (p-zero hypotheses). In Section 3.2. the resulting qualitative theory
of (general) confirmation, encompassing the qualitative theory of deductive
confirmation, will be indicated. Finally, in Section 3.3. we will briefly discuss
the acceptance of hypotheses in the light of quantitative confirmation. In
Appendix 1, it will be argued that Popper's quantitative theory of corroboration
amounts to an inclusive and impure Bayesian theory of confirmation. In
Appendix 2 the quantitative treatment of the raven paradoxes resulting from
our quantitative theory is compared in detail with an analysis in terms of the
standard Bayesian solution as presented by Horwich.
The quantitative approach to confirmation has a somewhat dubious character, since the assigned probabilities are, as a rule, largely artificial. Their main
purpose is to lead to adequate qualitative (classificatory and comparative)
judgments of confirmation. As far as deductive confirmation is concerned, we
have seen in the previous chapter that we do not need a quantitative approach
for that purpose. However, since to date no independent or direct qualitative
theory of general confirmation, or of non-deductive confirmation, has been
developed, a quantitative approach is required for that purpose. Such a dependent or indirect qualitative theory of general and non-deductive confirmation
will be presented in the second section.
Accordingly, we do not claim that the quantitative theory reflects quantitative
cognitive structures concerning confirmation. Instead, they should primarily
be conceived as quantitative explications of qualitative cognitive structures, to
be used only for their qualitative consequences. As will be argued, the justification of these qualitative consequences is at least as good as the justification of
43

44 QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES


the quantitative explications 'under ideal circumstances', that is, when the
probabilities make objective sense. Moreover, as in the qualitative case, it will
also become clear that there is not one 'language of quantitative confirmation',
but several, e.g., pure and impure ones, inclusive and non-inclusive ones. As
long as one uses the probability calculus, it does not matter which confirmation
language one chooses, the only important point is to always make clear which
one one has chosen. Although speaking of confirmation languages is hence
more appropriate, we will accept the current practice of speaking of confirmation theories.

3. 1. QUANTITATIVE CONFIRMATION

Introduction

In this section, a non-standard version will be presented of the so-called


Bayesian theory of confirmation, guided by the success perspective.
Quantitative confirmation will be decomposed into confirmation by a deductive
or a non-deductive success, or simply deductive and non-deductive confirmation. Both will be localized in the so-called Confirmation Square. The degree
of confirmation of a hypothesis by a piece of evidence will be equated with the
plausible degree of success, which happens to be equivalent to the ratio of the
posterior and prior probability when the latter is non-zero. The version of
Bayesianism is non-standard in two senses.l First, and foremost, it is inclusive
in the sense that it leaves room for a substantial degree of confirmation for 'pzero' hypotheses when they are confirmed. Second, it is pure in the sense that
equally successful hypotheses get the same degree of confirmation, irrespective
of their prior probability.
3.1.1. Non-deductive confirmation and the Confirmation Square

The four possible (unconditional) deductive relations between hypothesis and


evidence specified in the Confirmation Matrix in Subsection 2.1.1. of the previous chapter have somewhat weaker probabilistic versions, for which we propose
to use the same 'deductive' names.
Hf=E
Hf=-,E
-,Hf=E
-,Hf=-,E

=>
=>
=>
=>

p(E/H) = 1
p(E/H) = 0
p(E/ -, H) = 1
p(E/ -,H)=O

Deductive Confirmation:
Falsification:
Deductive Disconfirmation:
Verification:

DC(H, E)
F(H, E)
DD(H,E)
V(H, E)

Here we assume that there is some defensible probability function p, i.e., p may
well have subjective features, though then as much as possible in agreement
with objective information. In line with Bayesian philosophers of science
(Howson and Urbach 1989; Earman 1992), we will call p(E/H) and p(E/ -,H)
likelihoods.2

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES 45

A probabilistic theory of confirmation will be called Bayesian as soon as it


assumes, explicitly or implicitly, some prior distribution, that is, probability
values p(H) and p(-,H) = 1- p(H). As a rule, a prior distribution is already
assumed when one of the probabilities p(H), p(-,H), p(E) or p(Hj E) is used,
or both likelihoods p(Ej H) and p(Ej -,H).3 In the strict version, the Bayesian
theory prescribes to update the probability function on the basis of incoming
evidence E according to the standard definition of conditional probability,
viz., the probability of H on the condition E, p(Hj E) = p(H&E)jp(E) =
p(H)p(Ej H)j p(E). However, we will throughout assume the possibility of
'inductive jumps'.
According to the definition of conditional probability, p(Ej H) =
p(E&H)jp(H) is undefined when p(H) = O. However, this does not exclude that
p(Ej H) can be interpreted in this case 4 . For example, in case H entails E,
p(Ej H) is 1. Or consider the case that the hypotheses Hv for all possible values
v in [0, 1] for the probability of heads of a biased coin. Then p(Hv) = 0, but
p(Enj Hv) makes perfectly sense for a sequence En of outcomes of n throws,
viz., the corresponding binomial distribution. In this case, it is at most controversial for non-Bayesians whether and how p(Enj -,Hv) can be meaningfully
interpreted. For, in general, if p(H) = 0 then p(Ej -, H) = p(E), and in any
Bayesian approach it is assumed that p(E) can be assigned a value, whether
this is done in terms of the decomposition p(H)p(Ej H) + p( -, H)p(Ej -, H)
induced by H, hence p(Ej -, H), or in terms of some other decomposition. From
now on we will assume that both p(Ej H) and p(E), and hence p(Ej-,H), are
interpreted, even if p(H) = O. Similarly, there are cases where p(Hj E) can be
interpreted when p(E) = O. For instance, if E reports the specific value 0.2 of a
quantity X taking values in the [0, 1]-interval and H claims that the value of
a similar quantity Y will be below 0.5, it may well be reasonable to assign,
on the basis of the background beliefs, p(H) = p(Y < 0.5) = 0.5, p(Hj E) =
p(Y < O.5j X = 0.2) = 0.7 and p(E) = p(X = 0.2) = o.
Well then, by using the weaker 'likelihood versions', the four deductive
relations between Hand E can be depicted as the four sides of the unit square
of likelihood pairs <p(Ej H),p(Ej -,H), henceforth called the Confirmation
Square (CS), depicted in Figure 3.1.
The core of the 'quantitative success theory of confirmation' of a Bayesian
nature is completed by taking the interior of CS also into account. From the
success perspective, the conditions

< p(Ej H) Confirmation


C(H, E)
p(Ej H) < p(E) Oisconfirmation O(H, E)
p(E)

are the plausible criteria of confirmation and disconfirmation in general. The


first condition, p(E) < p(Ej H), will be called the S(uccess)-criterion 5 of
confirmation.
Note that the S-criterion coincides with the success definition of confirmation
in general (SOC) in the previous chapter, viz., H makes E more plausible, as

46

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

:t' 1
LfJ

-....;;..

Q.

Deductive Disconfirmation
Non-deductive
Disconfirmation

+=0

(j0

;;::

c:
0

+=0

c:
0

as
E
...
c:

0~

as

~,~

0
~

~;

!E-

O)

~'li
0":>

as
u.

>

+=0
0
;:,
"0

0)

Cl

Non-deductive
Confirmation

Verification

==> p(E/H)

Figure 3,1. The Confirmation Square (CS)

soon as we equate plausibility with probability, Of course, the criterion for 'no
confirmation' or neutral evidence is:

p(E) = p(E/H)

Neutral Evidence: NE(H, E)

To get a better view on extreme cases, represented by the sides of CS, and
of the non-extreme cases, represented by the interior, we formulate first equivalent criteria of confirmation, disconfirmation and neutral evidence.

p(E/ -, H) < p(E/H)

Confirmation:

C(H, E)

p(E/H) < p(E/ -, H)

Disconfirmation:

D(H, E)

p(E/ -,H) = p(E/ H)

Neutral Evidence: NE(H, E)

The depicted (/-) diagonal in CS represents neutral evidence, whereas everything


'below' the diagonal represents confirmation of one kind or another, and
everything 'above' it represents disconfirmation of one kind or another.
In the new formulation, the S-criterion for confirmation leaves clearly room
for the extreme cases of verification, p(E/ -, H) = 0, and deductive confirmation,
p(E/H) = 1. Similarly, the criterion for disconfirmation leaves room for the
extreme cases of falsification, p(E/H) = 0, and deductive disconfirmation,
p(E/-,H) = 1.
As a consequence, the region of the interior of CS right/below (left/above)
the diagonal typically represents non-extreme probabilistic successes of H
(-, H). These non-extreme cases represent the remaining intuitive cases of

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES 47

confirmation and disconfirmation, respectively. They will be called non-deductive:

o<p(Ej -,H) <p(Ej H) < 1


Non-deductive Confirmation:

NC(H, E)

O<p(Ej H)<p(Ej -,H)< 1

Non-deductive Disconfirmation:

ND(H, E)

Note that, as in the deductive case, non-deductive disconfirmation of H


amounts to non-deductive confirmation of -, H.
If one wants to set apart verification and falsification as extreme cases of
confirmation and disconfirmation, respectively, it is plausible to introduce the
notions of non-extreme or proper confirmation and disconfirmation:

o< p(Ej -, H) < p(Ej H)

Proper Confirmation:

PC(H, E)

0< p(EjH) < p(Ej -, H)

Proper Disconfirmation:

PD(H, E)

with some conceptually plausible consequences, in abbreviated form, indicating


subsets of CS by the relevant condition:
C(H, E)

= V(H, E) U PC(H, E) and

PC(H, E) = DC(H, E) U NC(H, E)

= F(H, E) U PD(H, E) and


PD(H, E) = DD(H, E) U ND(H, E)

D(H, E)

It is also fruitful to define conditional versions of non-deductive confirmation,


proper confirmation and confirmation in general:
O<p(Ej -,H&C)<p(Ej H&C) < 1

conditional Non-deductive Confirmation:

NC(H, E; C)

0< p(Ej -,H&C) < p(Ej H&C)

conditional Proper Confirmation:

PC(H,E; C)

p(Ej -,H&C) < p(Ej H&C)

conditional Confirmation:

C(H, E; C)

When supplemented with plausible definitions of conditional (deductive and


non-deductive) disconfirmation, each specific condition gives rise to its own
confirmation square, the conditional CS.
In sum, CS not only depicts falsification, verification and neutral evidence
but also suggests how to split proper confirmation and disconfirmation into
both a (basically qualitative) deductive subtype and a (fundamentally quantitative, at least so it seems) non-deductive subtype. This interpretation of the unit
square of likelihood pairs provides, as we will further illustrate, a quantitative
explication of the general idea of (qualitative) confirmation, that is, the basic

48

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

'cognitive structure' regarding confirmation that is implicitly used by empirical


scientists. However, since the required specific probabilities usually do not
correspond to anything in reality, neither in the object of study, nor in the
head of the scientist, consciously or unconsciously, they do not seem to directly
reflect a quantitative cognitive structure. However, one may argue that there
is something between a purely qualitative and a purely quantitative cognitive
structure, viz., by certain elicitation procedures one obtains interval assignments
of probabilities which may be interpreted as reflecting unconscious attitudes.
These interval assignments might obey a cognitive structure in terms of
intervals, but we will not pursue this possibility further.
Several aspects of CS will be treated in some detail. The following terminology will be very useful:

H is a p-zero hypothesis

p(H) = 0

H is a p-one hypothesis

p(H) = 1

H is a p-normal hypothesis

O<p(H)<1

H is a p-uncertain hypothesis

p(H) < 1

The present analysis provides in fact a decomposition of the standard Bayesian


theory of confirmation for p-normal hypotheses. Its criteria of confirmation
and neutrality read, respectively:

p(H) < p(H/ E)

p(H) = p(H/E)

(see e.g., Carnap 1963 2 , the new foreword to Carnap 1950, Horwich 1982,
Howson and Urbach 1989). This confirmation criterion, stating that the posterior probability is larger than the prior probability, may be said to be, not
success, but truth-value oriented. It will be called, more neutrally, the
F(orward-)criterion. It is in perfect agreement with the common sense idea,
expressed in the reward principle of plausibility (RPP) in the previous chapter,
that confirmation, normally, increases, or leads to the increase of, the probability of the hypothesis. Assuming that H is p-normal, the F -criterion is
equivalent to the S-criterion, C(H, E), as is easy to check. In view of the
"p(E/ , H) < p(E/H),,-version of the S-criterion, its decomposition of Bayesian
confirmation amounts to the following claim: assuming p-normality of H, the
F-criterion expressing Bayesian confirmation can be naturally decomposed
into three mutually exclusive and together exhaustive possibilities in
which the (equivalent) S-criterion can be satisfied: two extreme possibilities,
viz., verification (0 = p(E/ , H) < p(E/ H)) and deductive confirmation
(p(E/ I H) < p(E/ H) = 1), and the non-extreme possibility, viz., non-deductive
confirmation (O<p(E/ ,H)<p(E/ H)< 1).
The important difference is that the S-criterion is also non-trivially applicable
to p-zero hypotheses. Whereas the F-criterion makes all evidence neutral with
respect to p-zero hypotheses (for p(H) = 0 implies p(H/E) = 0), the S-criterion
leaves perfectly room for confirmation of such hypotheses. However, since

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

49

p(HjE) remains 0, the confirmation is, as it were, not rewarded in this case.
Note that the situation is different for p-one hypotheses. If p(H) = 1 then,
assuming that E and H are compatible, p(HjE) = p(H) and p(EjH) = p(E).
Hence, according to both criteria, p-one hypotheses cannot be confirmed. Note
in this connection also that, in contrast to the fact that the confirmation of a
p-normal hypothesis amounts to the disconfirmation of its negation, the confirmation of a p-zero hypothesis, according to the S-criterion, of course, does
not amount to the disconfirmation of its negation according to any of the two
criteria, which is easy to check. In view of the deviating behavior of the
S-criterion regarding p-zero hypotheses, the S-criterion will be called inclusive
and the F-criterion non-inclusive. Hence, although the inclusive and the noninclusive criteria are equivalent for the non-zero cases, they are incompatible
for the zero cases. As we will see in Appendix 1, Popper's approach (Popper
1959, 1963a, 1983) also presupposes the S-criterion, and hence is inclusive.
Inclusive behavior is very important in our opinion. Although there may be
good reasons (contra Popper, see Appendix (1) to assign sometimes non-zero
probabilities to genuine hypotheses, it also occurs that scientists would sometimes assign in advance zero probability to them and would nevertheless
concede that certain new evidence is in favor of them.
Whereas deductive confirmation has only one 'cause', the evidence is entailed
by the hypothesis, non-deductive confirmation may have different causes. In
the following we will restrict the attention to p-normal hypotheses and evidence.
As Salmon (1969) already pointed out in the context of the possibilities of an
inductive logic, a probability function may be such that E confirms H when H
partially entails E. Here 'partial entailment' essentially amounts to the claim
that the relative number of models in which E is true on the condition that H
is true is larger than the relative number of models in which E is true without
any condition. 6 For instance, a 'high outcome' (4, 5, or (6) with a fair die,
partially entails an even outcome, and vice versa. Both probabilistic criteria
lead to confirmation, since, e.g., p( 4v5v6j2v4v6) = 2/ 3 > 1/ 2 = p( 4v5v6). In a
'color language' with at least four colors, p will be such that the evidence that
a raven is black or white confirms the hypothesis that it is black or red. In
general, one may require that a probability function satisfies the principle of
partial entailment: if H partially entails E (..., E) then E confirms (disconfirms)
H. Fortunately, it seems that a probability function usually satisfies this principle. However, and this was Salmon's main message, it is not at all guaranteed
that such a function is such that E confirms H when H essentially is an
(inductive) extrapolation of E, notably from past to future instances of a certain
kind. For instance, one might like to have that the evidence that the first raven
is black confirms the hypothesis that the second raven is black as well. In
general, one may require that a probability function satisfies the principle of
extrapolation (or induction): if H extrapolates upon E (..., E) then E confirms
(disconfirms) H.7 In the next chapter we will study probability functions, e.g.,
Carnap's continuum of inductive methods, which satisfy both principles. Of

50

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

course, such functions are such that a hypothesis H which partially entails E
and extrapolates upon E is confirmed by E. In sum, we may distinguish at least
three causes or types of non-deductive confirmation: due to partial entailment,
which might be called 'partial (deductive) confirmation', due to extrapolation,
to be called 'inductive confirmation', and due to both factors. In Chapter 4 we
will introduce a third factor: analogy.
3.1.2. The ratio-degree of confirmation
Although the quantitative theory of confirmation presented thus far already
allows qualitative judgments of deductive and non-deductive confirmation, for
comparative purposes we also need a degree of confirmation. In the previous
chapter we have explicated 'confirmation' qualitatively as increase of plausibility of, in the first place, the evidence (SOC), and, in the second place, of the
hypothesis (RPP). In the present probabilistic context, it is plausible to identify
plausibility with probability, and hence, confirmation with increase of probability of the evidence, as we have noted, with the consequence, as far as pnormal hypotheses are concerned, that confirmation is rewarded by an increase
of the probability of the hypothesis.
There are many possibilities for defining a degree of confirmation, several
having some prima facie plausibility.s In the introduction we have already
remarked that, as long as one uses the probability calculus, it does not matter
very much which confirmation theory one chooses, and hence which degree of
confirmation, the only important point is to always make clear which one one
has chosen. In this section, we will restrict our attention to mainly one degree
of confirmation, viz., the ratio degree of confirmation, with some reference to
the standard and non-standard difference degree of confirmation. 9 Let us begin
by the latter, d(H, E) = defp(H/E) - p(H), that is, the difference between the
posterior and the prior probability of the hypothesis. From the success perspective, d'(H, E) = defp(E/ H) - p(E) is an at least as plausible difference measure
for it expresses in a way to what extent E is a success of H. Since they usually
give different values one has to choose between them.
The ratio degree of confirmation is usually presented as the ratio of the
posterior and the prior probability, p(H/E)/p(H). However, from the success
perspective, the ratio p(E/ H)/p(E) is at least as plausible as indicator of the
extent to which E is a success of H. The latter ratio may well be called the
amount or degree of success of H on the basis of E. Fortunately, now we do
not have to choose, for the two ratio measures are trivially equivalent, when
they are defined, hence we define:
r(H, E) = defp(H/E)/p(H)) = p(E/H)/p(E)
=

p(H&E)/(p(H)p(E))

to be called the r-degree or r-measure of success and confirmation. Note


that the first and the third ratio are not defined when p(H) = 0, and that the

QUANTITATIVE CONFIRMATION. AND ITS QUALITATIVE CONSEQUENCES

51

same holds for the second and the third ratio when p(E) = O. Since p is, as
a rule, not just an objective probability, both possibilities should not be
excluded beforehand. Recall that we have assumed that p(E/ H) can be
interpreted when p(H) = 0, and that p(H/ E) can be interpreted when p(E) =
O. Hence, r(H, E) is almost always defined, that is, it is always defined,
except when both p(E) and p(H) are zero, or when one of them is 0 such
that the corresponding conditional probability cannot be interpreted, possibilities that will further be disregarded.
In the following, we will evaluate the r-degree of confirmation in some
detail, partly in comparison with the d-degree and the d'-degree. To begin
with, being almost always defined need not be a positive feature, that
depends on the values that are assigned. For a first major advantage of r
over d and d' we study their extreme behavior. Note first that r has the
neutral value 1 and that d and d' both have the neutral value O. Higher
values indicate, of course, confirmation and lower values disconfirmation.
Let us see what happens under the extreme conditions that p(H) or p(E) is
zero. When p(H) = 0 d gets the neutral value. Hence d reflects the F-criterion
of confirmation, according to which a p-zero hypothesis is always neutrally
confirmed. That is, a hypothesis that is impossible according to p cannot be
confirmed or disconfirmed by evidence; all evidence is, by definition, neutral
for such hypotheses, a very strange situation indeed. Similarly, d' gets the
neutral value whenever p(E) = O. So, according to d' evidence that is
impossible according to p cannot confirm nor disconfirm a hypothesis, but
is always neutral. Note that in both cases, it would be less objectionable
when the degree of confirmation would not be defined. It is the assignment
of the neutral value which is conceptually unattractive.
It is easy to check that r may well assign a non-neutral value when either
p(H) or p(E) is zero (assuming that p(E/ H), respectively p(H/ E), can be
interpreted), and, as already remarked, it is undefined when p(H) = p(E) =
O. When p(H) and p(E) are both non-zero, r(H, E) reflects both the S- and
the F-criterion of confirmation, it reflects the S-criterion when p(H) = 0 and
the F -criterion when p(E) = O. Hence, we may say that the ratio-degree r
shows refined extreme behavior, whereas d and d ' show conceptually implausible extreme behavior.
To be sure, when p(H) = 0 and r(H, E) > 1, r(H, E) expresses confirmation
which is not rewarded, since p(H/ E) remains O. Note that r(H, E) equals
p(E/ H)/p(E/ i H)10 when p(H) = 0, since P(E) then equals p(E/ -,H). Similarly,
when p(E) = 0 and r(H, E) > 1, E is not recognized as confirming evidence,
since p(E/ H) remains O. In the case that r(H, E) > 1 and p(H) and p(E) are
both positive, E is recognized as confirming evidence of H, in the sense that
p(E/ H) has increased with respect to p(E) by the factor r(H, E) , whereas H is
rewarded for that success, in the sense that p(H/ E) has increased with respect
to p(H) by the same factor.
A second feature of the r-measure is its being a P-incremental measure l l in

52

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

the sense that it is (or can be written as) a function only of the probabilities
p(HjE) and p(H) which increases with increasing p(HjE) and decreases with
increasing p(H)12 . It may also be called an L-incremental measure in the sense
that it is (or can be written as) a function only of the likelihoods p(EjH) and
p(E) which increases with increasing p(EjH) and decreases with increasing
p(E). Note that d is also P-incremental, but not L-incremental, whereas d' is
L-incremental, but not P-incremental.
Next, the ratio of the r-degrees of confirmation of two hypotheses on the
basis of the same evidence, r(HI, E)/r(H2/E), just equals the ratio of the
likelihoods, p(E/ H1) jp(E/H2)Y This nicely fits the so-called likelihood ratio
approach in statistics to comparing two statistical hypotheses with each other,
assuming an underlying statistical model (see Note 10). Although d' is
L-incremental, it is not easily connectable to this statistical practice.
An important further difference between r and both d and d' is that r is
symmetric, that is, r(H, E) = r(E, H), whereas d and d' are asymmetric: d(H, E)
is unequal to d(E, H), in fact it is equal to d'(E, H), and similarly for d'.
Symmetry is particularly appealing in cases where the hypothesis is of the same
nature as the evidence. Consider, for example, the hypothesis (H) that the
outcome of a fair die will be even in relation to the evidence (E) that the
outcome is larger than 1 and the reverse situation that the evidence reports an
even die (E' = H), and the hypothesis (H' = E) states that the outcome will be
larger than 1. An asymmetric degree of confirmation may imply that E confirms
H more (or less) than E'( = H) confirms H'( = E), and d and d' do so. The
symmetry of r is, of course, directly related to the fact that r(H, E) can be seen
as a degree of mutual dependence between Hand E, since independence is
usually defined by the criterion p(H&E) = p(H)p(E)14.
Some special values of r(H, E) are relatively simple. For instance, r(H, E)
increases from 0, for falsification, via p(E/ H)/ [ 1 - p(H )p( -, E/H)] for deductive disconfirmation, to 1, for neutral (including tautological) evidence, from
which it increases further, via 1/p(E) for deductive confirmation, to 1/p(H),
for verification. The last value is, moreover, the maximum degree of
confirmation a hypothesis can get, viz., l / p(H) for verification, e.g., when
E = H. Note that this maximum is hypothesis specific, and that we have the
plausible extreme consequence that verification of a p-zero hypothesis
amounts to obtaining an infinite degree of confirmation. Similarly, 1/p(E) is
the maximum degree of confirmation certain E can provide for a hypothesis,
viz., by deductive confirmation, with the plausible extreme consequence that
the degree of confirmation in the case of deductive confirmation by p-zero
evidence is infinite.
3.1.3. Comparing and composing degrees of confirmation

Let us now turn to the comparative and composite behavior of r(H, E) by


presenting some trivial but crucial theorems, always assuming that H is puncertain (p(H) < I).

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES 53

We start by considering two pieces of evidence with respect to which a fixed


hypothesis is equally successful in the sense that they provide the hypothesis
with the same likelihood (e.g., 1 in the case of deductive confirmation):
Th.l

if p(E/H) = p(E*/H) > 0 then

r(H, E) > r(H, E*) iff p(E*) > p(E)


(iff p(H/E) > p(H/E*), if p(H) > 0)
Th.l states that, when H obtains the same likelihood from two pieces of
evidence, the degree of confirmation increases with decreasing prior probability
of the evidence or, if p(H) > 0, equivalently, with increasing posterior probability of the hypothesis. Hence, under the mentioned condition, according to
r(H, E), H gets 'richer' from less probable (more surprising) evidence, which
agrees with scientific common sense; we will call this the surprise bonus. Note
that, when p(H) = 0, this surprise bonus is not paid out in an increase of the
posterior probability, for that remains zero.
Let us now turn to fixed evidence and two hypotheses, which are equally
successful in the sense that they obtain the same likelihood from that evidence
(again, e.g., 1 in the case of deductive confirmation):
Th.2

if p(E/H) = p(E/H*) > 0 then

r(H, E) = r(H*, E)
and

p(H/E) > p(H*/E) iff p(H) > p(H*)


Th.2 shows in the first place that r(H, E) is a (hypothesis-) neutral 15 or pure
degree of confirmation, in the sense that two hypotheses which are equally
successful in the sense that they make the evidence equally plausible, obtain a
degree of confirmation which is independent of their prior probability. As noted
in the previous chapter (Note 4), this may be seen as a Humean feature. Th.2
states, moreover, that, assuming equal successfulness, the posterior probability
increases with increasing prior probability. Note that the first feature is in
sharp contrast to the 'impure' behavior of d(H, E). Since d(H, E) is equal to
p(H)(r(H, E) - 1), it has the Matthew-effect offavoring more plausible hypotheses, that is, it increases in the case of equal successfulness with the prior
probability. On the other hand, d' is easily seen to be pure.
Restricting attention to deductive confirmation and identifying plausibility
with probability, it follows directly from Th.l and Th.2 that quantitative
deductive confirmation, as measured by r(H, E), satisfies the qualitative principles of deductive confirmation P.l and P.2.
Let us also look at some specific cases that have been put forward in favor
of r(H, E) or d(H, E). Roberto Festa (1999a, p. 66) has suggested a version of
the following counter-intuitive case against d(H, E), and in favor of r(H, E),
when p(H) > O. Compare p(H/E) = 0.1 and p(H) = 0.0001 with p(H*/E) = 0.9

54 QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES


and p(H*) = 0.8. Although the respective differences are almost the same (~0.1
and 0.1, respectively) the first case of confirmation is intuitively much more
impressive than the second. It is easy to check that r(H, E) is in agreement
with this intuition (1000 and 9/ 8, respectively), which makes r(H, E) superior
to d(H, E). For a real-life (aircraft) example of a formally similar nature, see
(Schlesinger 1995, Section 4).
However, such specific intuitions may easily be countered by similar ones,
pointing in the opposite direction. Consider the following case against r(H, E)
and in favor of d(H, E), stemming from Eells and reported by Sober (1994).
In a slightly modified form, consider p(H/ E) = 0.9 and p(H) = 0.1 versus
p(H*/ E) = 0.001 and p(H*) = 0.00001. Though H may seem intuitively and
according to d(H, E) much more confirmed by Ethan H* (d(H, E) = 0.8 versus
d(H*, E) ~ 0.001), the r(H, E)-definition leads to the reverse conclusion (9
versus 1(0).
Accordingly, such examples make clear that our intuitions are confused and
that we can decide to reconsider our intuitions in the light of the fact that
there is something to choose, viz., principles we mayor may not want to
subscribe to.
So let us return to general properties of the r-measure. First we will consider
disjunctions of evidence and hypotheses. For the disjunction of two incompatible pieces of evidence, the r-degree of confirmation is the weighted sum of the
separate degrees of confirmation:
Th.3.l

if p(E&E') = 0 then
,p(E)
r(H, EVE) = ptE) + ptE') r(H, E)

p(E' )

+ ptE) + ptE') r(H, E )

Similarly, due to the symmetry of the r-degree with respect to E and H, the rdegree of confirmation of a disjunction of two incompatible hypotheses is the
weighted sum of the degrees of the disjuncts:
Th.3.2

if p(H&H') = 0 then
,
r(H V H, E)

p(H)

p(H')

= p(H) + p(H') r(H, E) + p(H) + p(H' ) r(H, E)

Let us now turn to conjunctions. Let E and E' be mutually independent pieces
of evidence in general and with respect to H. Then the degree of confirmation
provided by the conjunction is the product of the separate degrees:
Th.4.1

if p(E&E' ) = p(E)' ptE' ) and p(E&E'/ H) = p(E/ H)' p(E'/ H)


then r(H, E&E') = r(H, E) ' r(H, E ' )

Similarly, again due to the symmetry of the r-degree with respect to E and H,
for prior and posterior mutually independent hypotheses:
ThA.2

if p(H&H') = p(H) ' p(H') and p(H&H'/ E) = p(H/ E) ' p(H'/ E)


then r(H&H', E) = r(H, E)' r(H', E)

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES 55

Finally, let us consider the 'addition' of an irrelevant piece of evidence E',


defined by p(H/ E&E') = p(H/ E), or an irrelevant hypothesis, defined by
p(E/H&H') = p(E/H).
Th.5.1
Th.5.2

if p(H/ E&E') = p(H/E) then r(H, E&E') = r(H, E)


if p(E/H&H') = p(E/H) then r(H&H', E) = r(H, E)

In our opinion, the composite behavior of the r-measure, as expressed by


Theorems 3-5, is very plausible.
In sum, we conclude that r(H, E) is an attractive degree of confirmation. It
shows refined extreme behavior, it is incremental with respect to the probability
of the hypothesis as well as the evidence, it has hypothesis and evidence specific
maxima, it realizes the surprise bonus, it is pure in the sense of being neutral
with respect to equally successful hypotheses, independently from their prior
probabilities, and it has plausible composite behavior. 16 Since it implies the
qualitative principles of deductive confirmation, we may conclude from the
previous chapter, that it can deal with the standard objections to deductive
confirmation, with the raven paradoxes, and the grue problem. In the next
section, we will further evaluate r(H, E) with respect to qualitative consequences,
again partly in comparison with other candidates.
We conclude this section with three technical points. First, accepting r as
degree of confirmation, implies, of course, as explication of "E confirms H more
than E* confirms H*": r(H, E) > r(H*, E*). Second, as Milne (1995, 1996) has
rightly argued, a near relative to r, viz., log r(H, E), has some advantages over
r(H, E). E.g., its neutral value is O. Third, for completeness and later use, we
write down the conditional degree of confirmation corresponding to the unconditional one:

r(H, E; C) = deJp(H/E&C)/p(H/C) = p(E/H&C)/p(E/C)

= p(H&E/C)/(p(H/C)p(E/C))
It is easy to check that this conditional degree has similar properties to the
unconditional one.
The quantitative theory of confirmation based on r will be called the rtheory of confirmation, and those based on d- and d' will be called the d- and
the d'-theory, respectively.
3.2. QUALITATIVE CONSEQUENCES

Introduction
In this section it will first be argued in some more detail than in Subsection 3.1.3.
and 3.1.4. that the 'r-theory' restricted to deductive confirmation implies the
whole qualitative theory of deductive confirmation presented in Chapter 2. In
this connection it will be particularly illuminating to write out the quantitative
variant of the qualitative solution of the raven paradoxes. This example

56 QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES


illustrates, among other things, that the r-degree of confirmation can also be
interpreted as a degree of severity of tests, in particular of HD-tests, with
attractive qualitative consequences, Next, we will investigate to what extent a
corresponding qualitative theory of general and non-deductive confirmation
can be derived and defended,

32'], Derivation of the qualitative theory of deductive confirmation


The claim that the qualitative theory of deductive confirmation can be derived
from the quantitative theory amounts, of course, to the claim that deductive
confirmation is a subkind of quantitative confirmation that satisfies the comparative principles of deductive confirmation when plausibility is identified with
probability, We have already seen that deductive confirmation amounts to an
extreme kind of quantitative confirmation, due to the fact that H F E implies
that p(E/H) = L The corresponding r-degree of confirmation is l /p(E), which
exceeds 1, hence indicates confirmation, as soon as E is probabilistically uncertain, We have also concluded already, on the basis of Th'! and Th.2, that
quantitative confirmation respects the comparative principles P,1 and P,2,
when we identify plausibility with probability, Hence, quantitative confirmation
entails all principles of the qualitative theory of deductive confirmation, In
Subsection 3.2,4, we will review the extent to which it implies the general
principles of qualitative confirmation presented in the previous chapter, viz"
SDC, RPP, PS, and PCS,
In the previous chapter we have also alluded to the reverse perspective on
P,I/S'! and P ,2/S,2, that is, that they are made plausible by Bayesian considerations, For this purpose, it is important to note first that r(H, E) and d(H, E)
are both popular among Bayesians, Hence, since both measures support P,1
and S,!, these comparative postulates seem unproblematic for Bayesians,
Moreover, since r(H, E) is frequently suggested and used as an alternative to
the somewhat more current d(H, E), and since r(H, E) supports P.2 and S,2,
the latter comparative postulates are frequently implicitly assumed by
Bayesians, However, it should be conceded that d(H, E) is used at least as
frequently as r(H, E), Hence, for supporters of d, P,2 and S,2 will only become
acceptable as far as our general arguments in favor of r (above and below),
and those of others, such as Festa (1999a), Schlesinger ( 1995), and Milne ( 1995,
1996), are convincing for them,
In this respect, it is interesting to study the way the r- and the d-measure
deal with an irrelevant additional hypothesis H' in the case of deductive
confirmation of H by E, that is, when p(E/ H) = L Whereas d(H&H', E) becomes
smaller than d(H, E), by the factor p(H&H')/p(H), r(H&H', E) remains equal
to r(H, E), In general, if p(E/H&H') = p(E/H), the plausible condition for a,
relative to E in the face of H, irrelevant additional hypothesis H', we obtain
d(H&H', E) = (p(H&H')/ p(H)' d(H, E), whereas r(H&H', E) = r(H, E) (Th.5.2),
Hence, whereas r accounts for the irrelevance of H' in a straightforward way,

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES 57

d does so in a more complicated way, which mayor may not be conceived as


more sophisticated.
In order to question the latter suggestion we will conceive an 'objective case',
assuming that the degree of confirmation should give satisfactory answers in
cases where only objective probabilities are in the game, since our intuitions
may then be assumed to be as sharp as possible. Consider the following urnmodel of a two-step random experiment. A B-urn is an urn with precisely one
black ball, a BB-urn an urn with precisely two black balls and a BW -urn an
urn with one black and one white ball. First we randomly select an urn out of
a collection of 1 B-urn, 4 BB-urns and 5 BW-urns, hence with objective
probability IjlO, 2j 5, Ij2, respectively. Next, in the selected urn, we randomly
select balls with replacement. Suppose that the first n selections of the second
type lead to a black ball. It is easy to check that this type of evidence deductively
follows from, hence d-confirms, the hypothesis that the first selected urn is a
B-urn, H-B, as well as the hypothesis that it is a BB-urn, H-BB. Now the
question is whether this evidence (d-)confirms H-B more than H-BB. Since the
evidence differentiates in no way between the two hypotheses, the 'r-answer'
('as much as') seems the most plausible one, and not the 'd-answer' ('less than').
Similarly, consider the disjunctive hypothesis 'H-B or H-BB', being weaker
than its disjuncts, but nevertheless d-confirmed by the evidence. Again, the rclaim that it is as much confirmed as its disjuncts, seems more plausible than
the d-c1aim that it is more confirmed. To be sure, the prior and hence the
posterior probability of H-B is smaller than that of H-BB, and the latter, and
hence the former, is smaller than that of the disjunctive hypothesis 'H-B
or H-BB'.
Let us now consider the way the r- and the d-measure deal with irrelevant
disjunctive evidence E' in the case of deductive confirmation of H by E. To
avoid inessential complications, let us restrict attention to the case that E' is
incompatible with E. Then r(H, EVE') becomes smaller than r(H, E) by the
factor p(E)jp(E V E') = p(E)j ( p(E) + p(E' )), hence decreases with increasing
p(E'). Since d(H, E) = p(H)(r(H, E) - 1), d(H, EVE') decreases in a related
way. In general, if p(E'j H) = p(E'), the plausible condition for irrelevant, for
neutral, disjunctive evidence E', r(H, EVE') and d(H, EVE') both decrease
with increasing p(E'); in view ofTh. 3.1, the former does so in a more transparent
way than the latter. In sum, as was to be expected, rand d behave rather
similar with respect to irrelevant disjunctive evidence.
Combining the results for an irrelevant conjunctive hypothesis and an irrelevant disjunctive piece of evidence, we may conclude that the r-measure deals
with both in a plausible way.
Let us now turn to the special qualitative applications or principles of
Section 7.2. It will be useful to list first the relevant corollaries of Th.l and
Th.2 with respect to conditional deductive confirmation:
Th.lc

if p(EjH&C) = p(E* j H&C*) > 0 then

58

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

r(H, E; C) > r(H, E*; C*) iff p(E*/C*) > p(E/C)


(iff p(H/E&C) > p(H/ E*&C*), if p(H) > 0)
Th.2c

if p(E/ H&C) = p(E/ H*&C) > 0 then

r(H, E; C) = r(H*, E; C) = p(E/H&C)/p(E/ C) =


p(E/H*&C)/p(E/C)
and
p(H/E&C) > p(H*/ E&C) iff p(H/C) > p(H*/C)
(iff p(H) > p(H*) if p(C/H) = p(CfH*) > 0)
The condition "p(CfH) = p(CfH*) > 0" amounts, of course, to the claim that
the probability that C is, or will be, realized is independent of the hypothesis
under consideration. Note that the unconditional versions of Th.lc and Th.2c
arise by skipping C and C* in the formulas.
In the next subsection we will show that the special principle SII.lc, dealing
with a fixed hypothesis, e.g., the raven hypothesis, is realized by r(H, E) as a
special case of Th.lc if we are willing to express some relevant background
beliefs by the probabilistic assumption:
Ap-ravens

is based on random sampling in the relevant universe

Finally, it is easy to check that SQ.2c, dealing with fixed evidence, e.g., in the
emerald case, is trivially realized as a special case of (the first part of) Th.2c:
r("all E are Q", G; EM) = r("all E are G", G; EM)

= l/p(G/ EM)
If we are, moreover, willing to express the green/grue-case of the weak irrelevance assumption (WIA-emeralds) by the probabilistic assumption:

WIAp(-emeralds)

p("all E are G") > p("all E are Q")

it is easy to derive the probabilistic version of the refined intuition (4&5). Of


course, if p("all E are G") = 0, the refined intuition cannot be realized, but the
degree of confirmation of both hypotheses will remain l /p(G/ EM).
Since, assuming Ap-ravens, S".lc-ravens is realized by r(H, E), the raven
paradoxes are qualitatively solved in the same way as before, because all desired
results already followed qualitatively, assuming S".lc-ravens and A-ravens.
Similarly, in the light of the fact that SQ.2c-emeralds is realized by r(H, E), and
assuming WIAp-emeralds, the grue problem is qualitatively solved in the same
way as before, since all desired results already followed qualitatively, assuming
SQ.2c-emeralds and WIA-emeralds.17 As a matter of fact, Sober (1994) inspired
us to our proposal for the refinement of Goodman's basic intuition, viz., (4&5)
of Subsection 7.2.2. In fact, he derived from WIAp-emeralds the quantitative
counterpart of (an 'impure' version of) that refinement.
In sum, the 'r-theory' of confirmation can generate the qualitative theory of
deductive confirmation in the most encompassing way.

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

59

3.2.2, The raven paradoxes reconsidered


Whereas it is not interesting to write down the quantitative analysis of the
grue problem, it is instructive, also for later purposes, to spell out the quantitative solution of the raven paradoxes. Recall RH, the hypothesis that all ravens
are black. Table 3.1. introduces a matrix of numbers for the sizes of the four
cells constituting the relevant conceptual possibilities.
Table 3.1. Numbers of 'raven possibilities'

#R
#B
#8

Total

c
a+c

Total

b
d
b+d

a+b
c+d
a+b+ c +d

Of course, these numbers are assumed to be finite but further unknown.


There are only some comparative background beliefs. In particular, the assumption A-ravens stating that the number of ravens is much smaller than that of
non-black objects, which amounts to a + c c + d, and this is equivalent to
a d (and hence to a + b b + d). We assume, moreover, that a and bare
positive and, of course, that c is 0 if RH is true and positive if RH is false. All
results to be presented basically presuppose and use Ap-ravens, according to
which testing is random sampling in the relevant universe. In the following,
'sampling' is to be read as 'random sampling' and the explicit occurrence of 'c'
means that it 'resulted from' the condition that RH is false. Hence, from now
on c > O.
Recall that r(H, E) = p(E/H)/p(E)( = p(H/E)/p(H.18 Writing 'q' for p(RH),
which mayor may not be assumed to be positive, and using the appropriate
conditional versions whenever relevant (in which case C is, of course, supposed
to be neutral 'evidence' for RH, that is, p(RH/C) = p(RH) = q), the crucial
expression becomes:
__
&_C)~_________
_ ________~p~(E~/R_H
r( R H E' C)"
- p(RH/ C)p(E/ RH&C) + p(,RH/ C)p(E, RH&C)
1

q + (1 - q)p(E/ , RH&C)/p(E/ RH&C)

The results are as follows 19 :


a black raven, a non-black non-raven and a black non-raven,
resulting from sampling in the universe of objects (hence C tautologous), non-deductively confirm RH, all with the same r-value:
(a + b + c + d)/(a + b + qc + d) (e.g., for a black raven, via
p(BR/ RH) = a/(a + b + d) and p(BR/ , RH) = a/(a + b + c + d
(2p) a black raven resulting from sampling ravens cd-confirms RH

(lp)

60 QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

(3p)

(4p)

with r-value r(RH, BR; R) = (a + c)/(a + qc), via p(BR/ RH&R) =


1 and p(BR/ , RH&R) = a/(a + c), and a non-black non-raven
resulting from sampling non-black objects cd-confirms RH with
r-value r(RH, BR; B) = (c + d)/(qc + d) (similar calculation)
sampling black objects or non-ravens always leads to neutral
evidence, i.e., r-value 1, e.g., for black objects, via p(BR/ RH&B) =
p(BR/ , RH&B) = a/(a + b)
ad (2p): a black raven resulting from sampling ravens cd-confirms
RH much more than a non-black non-raven resulting from sampling non-black objects, for (a + c)/(a + qc)>> (c + d)/(qc + d) iff
a d, where the latter condition follows from A-r~vens.

Note that all r-values, except those in (3p), exceed 1, and that this remains the
case when p(RH) = q = O. It is easy to check that (4p) essentially amounts to
a special case of Th.lc, realizing S#.lc-ravens.
It should be noted that this analysis deviates somewhat from the more or
less standard Bayesian solutions of the paradoxes. In the light of the many
references to Horwich (1982), he may be supposed to have given the best
version. In Appendix 2 (based on (Kuipers, to appear)) we argue that our
solution, though highly similar, has some advantages compared to that of
Horwich.

3.2.3. The severity of tests


The raven example provides a nice illustration of the fact that, at least in the
case of a HD-test, the degree of confirmation can also be conceived as the
'degree of severity' of the test. When H entails E, r(H, E) expresses the degree
of success or the degree of confirmation H has obtained or can obtain from an
experiment that results in E or non-E. In the latter reading, it expresses a
potential degree of success. The smaller p(E), the more success H can obtain,
but the less probable it will obtain this success. Both aspects are crucial for
the intuition of the severity of tests. The more severe a test is for a hypothesis,
the less probable that the hypothesis will pass the test, but the more success is
obtained when it passes the test. More specifically, the degree of severity of a
HD-test is according to Popper (1959, Appendix *ix, 1963a, Addendum 2,
1983, Section 32) (a measure increasing with) the probability that the test leads
to falsification or a counter-example, or, to quote Popper (1983, p.247), "the
improbability of the prediction measures the severity of the test". This specification amounts to taking p(, E) as the degree of severity, or some function
increasing with p(,E), as do Popper's proposals in Addendum 2 of (Popper
1963a). One of Popper's proposals for the degree of severity of a HD-test is
the r-value 1/p(E), where it is again important that p(E) is calculated before
the experiment is performed, or at least before its result is known. Like Popper,
we see no reason not to generalize this definition to p(E/H)/p(E)20 for non(conditionally) deductive tests. However, it is primarily HD-tests where we

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

61

seem to have specific qualitative severity intuitions, some of which will be


studied now.
First, rephrasing the results (1p)-( 4p) concerning the raven paradoxes in the
previous subsection in severity terms, the analysis explains and justifies, essentially in a deductive way, why scientists prefer, if possible, random testing of
ravens, that is, randomly looking among ravens to see whether they are black,
and, in general, choose that way of conditional random testing among the ones
that are possible, which is the most severe.
Second, a standard objection to Bayesian theories of confirmation in general
is the so-called 'problem of old evidence'. If we know already that E is true,
and then find out that H entails E, the question arises whether E still confirms
H. The problem is, of course, that our up to date probability function will be
such that p(E) = 1, and hence p(Ej H) = p(E) = 1 and p(Hj E) = p(H). Hence,
the r-degree then leads to the neutral value 1. This reflects the intuition that
there is no severe test involved any longer. Despite this severity diagnosis, E
does nevertheless represent a success or confirming evidence of the degree
1jp'(E), where p' refers to the probability function before E became known or,
similarly, the probability function based on the background knowledge minus
E. The latter, counterfactual, defence is more or less standard among Bayesians
(Howson and Urbach 1989; Earman 1992)21, but the additional severity diagnosis is not.
Third, there are two other intuitions associated with severity. The first one
is the famous 'diminishing returns' intuition of Popper: "There is something
like a law of diminishing returns from repeated tests" (Popper 1963a, p. 240).
It expresses the idea that the returns and hence the severity of repeated tests
decreases in one way or another. Let us look at the raven example. Let Rn
represent n random drawings (with replacement) of ravens and let Bn indicate
that these n ravens are black. Of course we have that p(BnjRH&Rn) = 1.
Moreover, replacing p(RH) again by q, p(Bnj Rn) = p(RH). p(BnjRH&Rn)
+ p( -, RH). p(Bnj-, RH&Rn) = q + (1 - q )p(Bnj-, RH&Rn). Suppose first that
q > 0. Assuming that -, RH implies that there is a positive probability (1 - q)
of drawing a non-black raven, p(Bnj-, RH&Rn) = q" goes to 0, for increasing
n, and hence r(RH, Bn; Rn) will increase to 1jq. Hence, r(RH, Bn + 1;
Rn + 1) - r(RH, Bn; Rn) has to go to 0, that is, the additional returns or degree
of confirmation obtained by an extra (successful) test goes to 0. In terms
of severity, the additional degree of severity by an extra test goes to 0. A
similar diminishing effect arises, of course, when we consider the ratio
r(RH, Bn + 1; Rn + 1)jr(RH, Bn; Rn), which will go to 1. However, if q = the
situation is different: r(RH, Bn; Rn) = 1j p(Bnj -, RH&Rn) = I j q". Hence, r
increases without limit, so does the extra returnsj confirmationj severity
1jq"+1_1jq"=(1jq"+1)(1-q), whereas the ratio r(RH,Bn+1;Rn+1)j
r(RH, Bn; Rn) remains constant (1 j q). In sum, the r-measure perfectly reflects
the intuition of diminishing returns, assuming that q = p( RH) is positive.
The last severity intuition to be considered, may be called the 'superiority

62 QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

of new tests', that is, the idea that a new test is more severe than a mere
repetition. It is a specific instance of the more general 'variety of evidence'
intuition. However, it appears to be not easy to give a rigorous explication
and proof of the general intuition (Earman 1992, p. 77-79). But for the special
case, it is plausible to build an objective probabilistic model which realizes the
intuition under fairly general conditions. The set-up is a direct adaptation of
an old proposal for the severity of test (Kuipers 1983). Let us start by making
the intuition as precise as possible. Suppose that we can distinguish types of
(HD-)test-conditions and their tokens by means other than severity considerations. E.g., ravens from different regions, and individual drawings from a region.
Any sequence of tokens can then be represented as a sequence of Nand M,
where N indicates a token of a new type, i.e., a new test-condition, and M a
token of the foregoing type, i.e., a mere repetition. Each test can result in a
success B or failure non-B. Any test sequence starts, of course, with N. Suppose
further that any such sequence X n, of length n, is probabilistic with respect to
the outcome sequence it generates. Note that RH is still supposed to imply
that all n-sequences result in Bn, and hence that one non-B in the outcome
sequence pertains to a falsification of RH. A plausible interpretation of the
intuition now is that the severity of a X nN-sequence is higher than that of a
XnM-sequence, that is:
1

q + (1 - q)p(Bn + 1/ ..., RH&XnN)

>

q + (1 - q)p(Bn + 1/ ..., RH&XnM)

which is equivalent to:

p(Bn + 1/ ..., RH&XnN) < p(Bn + 1/ ..., RH&XnM)


to be called the superiority condition.
The remaining question is whether this condition holds under general
assumptions. For this purpose, we will construct an urn-model which reflects
the ideas of new tests and mere repetitions. Suppose there is a reservoir with
an unknown finite number of urns, each containing an unknown finite number
of balls, which number mayor may not differ from urn to urn. Our hypothesis
to be tested states that all balls in the reservoir, and hence in each urn, are
black. Restricting our attention to random selections of urns and balls with
replacement, the possibilities for probabilistic test sequences are as follows:
start by a random selection of an urn, draw randomly and successively a
number of balls out of that urn with replacement, replace the urn and start
over again, with the same or a different number of ball selections out of the
next urn. It turns out to be non-trivial (see Kuipers 1983, 219- 220) to prove
the superiority condition assuming one very plausible condition, viz., if RH is
false, the ratio of black balls may not be the same in all urns.
In sum, in the case of deductive confirmation, the ratio degree of confirmation
may well be conceived as the degree of severity of the HD-test giving rise to

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES 63

the confirmation, for it satisfies the main qualitative features associated to


current severity intuitions. 22

3.2.4. Qualitative non-deductive confirmation


The plausible question now arises whether it is possible to give a qualitative
explication of non-deductive confirmation, that is, an explication of non-deductive confirmation in terms of 'plausibility'. It will be easier to concentrate first
on confirmation in general, or general confirmation, after which non-deductive
confirmation can be identified with general confirmation of a non-deductive
nature.
But first we will check whether and to what extent the r-theory realizes the
success definition of general confirmation and the further general principles
that were presented in Subsection 2.1.2. For this purpose, we have to replace
'E confirms H' by 'r(H, E) > l' and 'plausibility' by 'probability'. According to
the success definition of confirmation (SOC), we should have that r(H, E) > 1 iff
p(E/H) > p(E). This holds, by definition, whenever p(E) > O. According to the
reward principle of plausibility (RPP) we should have that p(H/ E) > p(H) iff
r(H, E) > 1, which holds whenever p(H) > 0.23 The joint consequence, that is,
the principle of symmetry (PS), p(H/E) > p(H) iff p(E/H) > p(E), holds whenever
both p(H) and p(E) are positive. Finally, r(H, E) realizes the principle of comparative symmetry (peS) trivially by its two-sided definition, where the inequalities
only hold as far as the relevant prior probabilities are non-zero and the relevant
conditional probabilities can be interpreted under 'p-zero conditions'.
For the remaining comparative principles we first state trivial generalizations
of Th.1 and Th.2:
If 0 < p(H) < 1 and p(E*/ H) = ;;:: > p(E/H) > 0 then

Th.lG

r(H, E*)
r(H, E)

---'-=

p(E*/ H) p(E)
p(H/E*)
--=
p(E/H) p(E*)
p(H/E)

p(E)
> >-p(E*)

Th.l G suggests
P.1G

(a) If H makes E* as plausible as E then E* confirms H as much


as E if (and only if) E* is as plausible as E
(b) If H makes E* at least as plausible as E then E* confirms
H at least as much as E if E* is at most as plausible as E
(c) If H makes E* more plausible than E then E* confirms H
more than E if E* is less plausible than E

' '--

Th.2 can be generalized to:


Th.2G If 0 < p(H) < 1 and p(E/H*)

= ;;::

> p(E/H) > 0 then

--- =

r(H*, E)
r(H, E)

p(E/H*)
= > > 1 and
p(E/H)
-

p(H*/ E)
p(H/E)

p(E/H*) p(H*)
-p(E/H) p(H)

::...:....---'---' =

p(H*)
p(H)

> > --

64

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

Th.2G suggests
P.2G

(a) If H* makes E as plausible as H then E confirms H* as much


as H (with the consequence that the relative plausibility of
H* with respect to H remains the same)
(b) If H* makes E at least as plausible as H then E confirms
H* at least as much as H (with the consequence that the
relative plausibility of H* with respect to H remains at least
the same)

(c) If H* makes E more plausible than H then E confirms H*


more than H (with the consequence that the relative plausibility of H* with respect to H increases)
It is not difficult to check that P.I and P.2, and hence S.1 and S.2, are subcases

of P.1G and P.2G, respectively. Since the special principles dealing with ravens
and emeralds concerned specific types of (conditional) deductive confirmation,
we do not need to generalize them.
In sum, the qualitative explication of general confirmation can be given by
SDC, RPP, PCS, P.1G and P.2G. As announced, it is now plausible to define
general confirmation of a non-deductive nature simply as non-deductive general
confirmation.
Although it is apparently possible to give qualitative explications of general
and non-deductive confirmation, we do not claim that these explications are
independent of the corresponding quantitative explications. In particular, we
would certainly not have arrived at P.1G and P.2G without the quantitative
detour. To be sure, this is a claim about the discovery of these principles; we
do not want to exclude that they can be justified by purely non-quantitative
considerations.
Similarly, although it is possible to suggest that the two (qualitative) proper
connotations formulated for deductive confirmation can be extrapolated to
general and non-deductive confirmation, we would only subscribe to them, at
least for the time being, to the extent that their quantitative analogues hold.
However, in this respect the quantitative situation turns out to be rather
complicated, hence, its re-translation in qualitative terms becomes even more
so. Fortunately, the proper connotations looked for do not belong to the core
of a qualitative theory of general confirmation.
Accordingly, although there is no direct, intuitively appealing, qualitative
explication of general and non-deductive confirmation, beyond the principles
SDC, RPP, and PCS, there is a plausible qualitative explication of their core
features in the sense that it can be derived via the quantitative explication. In
other words, we have an indirectly, more specifically, quantitatively justified
qualitative explication of general and non-deductive confirmation.
It is important to argue that its justification is at least as strong as the
justification of the quantitative explication 'under ideal circumstances', that is,
when the probabilities make objective sense. At first sight, it may seem that we

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

65

have to take into account our relativization of the quantitative explication by


emphasizing and criticizing the artificial character of most of the probabilities.
However, this is not the correct evaluation. As soon as we agree that the
quantitative explication is the right one in cases where these probabilities make
(objective) sense 24 , the qualitative consequences are justified in general, since
they are not laden with artificial probabilities. In other words, the justification
of the qualitative explication is at least as strong as the justification of the
quantitative explication in cases were the relevant probabilities make sense.
An interesting question is to what extent the d-theory, counting d(H, E) > 0
as confirmation, leads to another explication of general and non-deductive
confirmation. It is almost evident that d realizes SOC, RPP, PS, PCS, and
P.IG for p-normal hypotheses, that it does not leave room for confirmation of
p-zero hypothesis, that is, it is non-inclusive, and finally that it is impure in the
sense that it is in conflict with P.2G. More specifically, the d-theory favors
more probable hypotheses among equally successful ones. In other words, the
d-theory gives rise to an alternative explication of deductive, general and nondeductive confirmation. We leave the question of whether the two resulting
sets of principles should be considered as expressing the 'robust qualitative
features' of two different concepts of confirmation or simply as two different
ways in which the intuitive concept of confirmation can be modelled, as an
open problem to the reader.

3.3. ACCEPTANCE CRITERIA

Finally, let us briefly discuss the acceptance of hypotheses in the light of


quantitative confirmation. The subject of probabilistic rules of acceptance has
received much attention in the last decades. For a lucid survey of 'cognitive
decision theory', see (Festa 1999b). As Festa documents, there is a strong
tendency to take 'cognitive utilities', such as information content and distance
from the truth into consideration. However, in our set-up, we only need rules
of acceptance in the traditional sense of rules for 'inductive jumps', that is,
rules that use conditions for acceptance that may be assumed to give good
reasons for believing that the hypothesis is true simpliciter. The story of theory
evaluation and truth approximation, the short-term dynamics, to be told in
the next chapters only presupposes such traditional rules of acceptance, in
particular, for general observational hypotheses (first order observational
inductions) and for comparative hypotheses, comparing success (comparative
second order observational inductions) and truth approximation claims of
theories (comparative referential and theoretical inductions). For the long-term
dynamics, the acceptance of separate hypotheses as true is crucial, notably for
extending the sphere of the observable. In Section 2.3. we have already discussed
the role of deductive confirmation in the acceptance of hypotheses as presumably true. For confirmation in general the role remains roughly the same, so we

66 QUANTITATIVE CONFIRMATION. AND ITS QUALITATIVE CONSEQUENCES


will not repeat this in every detail. However, it remains interesting to see what
precise role the degree of confirmation plays in this context.
Let us first note that the acceptance of H (as true) on the basis of E is a
'non-Bayesian move', that is, assuming that E does not verify H, i.e., p(H/E) < 1,
acceptance of H amounts to the replacement of p(H/ E) by the new prior
probability p'(H) = 1, at least for the time being. It is also plausible to think
that, if p(H) > 0, and if there is something like a threshold for acceptance, this
threshold is independent of p(H). That is, there is assumed some number e
1/2), such that H is accepted when p(H/ E) > 1 - e. As is well-known, the
suggested rule immediately leads to the lottery-paradox: for a sufficiently large
lottery, one may know that there is just one winning ticket, and at the same
time, have to accept for each of the tickets that it is not the winning one, hence
that there is no winning ticket. However, the paradox is based on a priori
reasoning, hence evidence and posterior probabilities do not playa role in it.
So let us assume that the (standard and non-standard) Bayesian approach to
confirmation can formally be combined with a sophisticated kind of nonBayesian high probability rule of acceptance, in which evidence plays a role,
the threshold is independent of specific prior probabilities and the 'posterior
analogue' of the lottery-paradox is avoided. That this is possible for a very
special class of (double-)inductive probability functions defined for monadic
predicate languages has been shown by Hintikka and Hilpinen (see Hintikka
1966 and Hintikka and Hilpinen 1966) and was extended to other classes
of inductive probability functions for such languages by Kuipers (1978,
Section 8.6.. However, a save, that is, lottery-paradox avoiding, general rule
of acceptance, also applicable for non-monadic predicate languages, is still
lacking, but see, for instance, (Pollock 1990) for some interesting attempts.
The question now is, what role does the degree of confirmation play in such
rules? The answer is, of course, none, for the degree of confirmation amounts
to an expression of the increase of the probability of the hypothesis, in a pure
or an impure form, and not to the resulting posterior probability, which is
simply calculated by Bayes' rule. Hence, whether we construe the degree of
confirmation in one way or another, it does not matter for the acceptance of
a hypothesis as 'true simpliciter'.
To be sure, it may well be that our confirmation intuitions are laden with a
mixture of the suggested two aspects, that is, the increase and the resulting
probability. From the 'pure' point of view, we will say that a more probable
hypothesis is not more confirmed by the same deductive success than a less
probable one, it just gets a higher posterior probability, and hence will earlier
pass the threshold for acceptance. Specifically, the conjunction intuition and
the refined grue intuition are done justice by the fact that the corresponding
strange hypotheses will obtain at most a small non-zero posterior probability,
never high enough to pass the threshold, as long as we do not find rock-blocks
of cheese on the moon or color changing emeralds.
Accordingly, the Bayesian approach to posterior probabilities guarantees

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES 67

that different definitions of the degree of confirmation will not lead to differences
in acceptance behavior, as long as the resulting posterior probabilities are
crucial for the rules of acceptance.
However, p-zero hypotheses will not get accepted in this way, since their
posterior probability remains zero. So, let us see what role the r-degree of
confirmation might play in acceptance rules for p-zero hypotheses. We have
already remarked that, although it may perfectly make sense to assign nonzero probabilities to genuine hypotheses, it nevertheless occurs that scientists
would initially have assigned zero probability to certain hypotheses, of which
they are nevertheless willing to say that they later have come across confirming
evidence for them, and even that they have later decided to accept them. Now
one may argue that this can be reconstructed in an 'as if' way in standard
terms: if the scientist would have assigned at least such and such a (positive)
prior value, the posterior value would have passed the threshold. To calculate
this minimal prior value, both d(H, E) and r(H, E) would be suitable. However,
only r(H, E) is a degree for which this 'as if' degree would be the same as the
'original' degree, for r(H, E) does not explicitly depend on p(H)25. In contrast
to this feature, the original d-degree of confirmation assumes its neutral value O.
If one follows this path, it is also plausible to look for a general acceptance
criterion that does justice to the, in most cases, relative arbitrariness of the
prior distribution. Let us, for that purpose, first assume that for cases of
objective probability one decided to take as the acceptance threshold 1 - e,
with 0 < e < 1/2. One plausible criterion now seems to be the r-degree of
confirmation that is required for the transition from p(H) = e to p(HI E) 2 1 - e,
that is, r(H, E) = (1 - e)/e. The suggested criterion can be used for p-normal as
well as p-zero hypotheses. However, as Jeffrey (1975) rightly remarks, most
genuine scientific hypothesis not only start with very low initial probability,
but will remain to have a low posterior probability. Hence, if e is very small,
they may not pass the threshold. However, passing the threshold is essentially
independently defined from p(H). For deductive confirmation, it is easily
checked to amount to the condition p(E) < e/( 1 - e). Hence, for somebody for
whom p(H) = e, deductive success E should be almost as surprising as H itself.
Whether the criterion is useful in other cases has still to be studied.

Concluding remarks
The possibility of a quantitative, i.e., probabilistic, theory of confirmation is
one thing; its status and relevance is another. Although probabilistic reasoning
is certainly practiced by scientists, it is also clear that specific probabilities
usually do not playa role in that reasoning. Hence, in the best instrumentalist
traditions, as remarked before, the required probabilities in a quantitative
account correspond, as a rule, to nothing in reality, i.e., neither in the world
that is studied, nor in the head of the scientist. They simply provide a possibility
of deriving the qualitative features of scientific reasoning.

68 QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES


If our reservations amounted to the claim that the quantitative accounts are
not yet perfect and have still to be improved, it would be plausible to call them
tentative explanations and even justifications of the corresponding kinds of
qualitative reasoning. However, nothing of that kind seems to be the case.
Hence, it remains questionable to what extent these formal accounts can be
said to reveal quantitative cognitive structures that underlie scientific reasoning.
The situation would change in an interesting way if the r-theory itself, or some
alternative quantitative theory, could be given a justification. In particular, we
do not exclude that such a justification could be given in terms of functionality
for truth approximation (see Festa 1999a and 1999b). However, although the
following chapters deal with truth approximation, they will not touch the
problem of such a justification.
In the meantime, we may only conclude that the r-theory should primarily
be conceived as a quantitative explication of a qualitative cognitive structure,
to be used only for its qualitative consequences. As has been argued, the
justification of these qualitative consequences is at least as good as the justification of the quantitative explication 'under ideal circumstances', that is, when
the probabilities make objective sense.
APPENDIX 1
CORROBORATION AS INCLUSIVE AND IMPURE
CONFIRMATION

As is well-known, Popper preferred to talk about '(degree of) corroboration',


instead of '(degree of) confirmation', but the question is whether his views
essentially deviate from the Bayesian approach. Jeffrey (1975) argued already
that this is not the case. In this appendix, we will more specifically argue that
Popper's quantitative theory of corroboration amounts to an inclusive and
impure Bayesian theory.
Popper's main expositions about corroboration can be found in (Popper
1959, 1963a, 1983), where Section 32 of Popper (1983) summarizes his main
ideas. He formulates six conditions of adequacy for a quantitative degree of
corroboration, here denoted by c(H, E), which he conceives as the best proposal.
We will first list (the core of) these conditions, in which it is important to
realize that p(,H) and p(,E) may be conceived as measures for the (amount
of) empirical content of Hand E, respectively.
(i) -15,c(H,E)5,p(,H)5,1
(ii) -1 = c(H&, H, E) = c(H" H) 5, c(H, E)
5, c(H, H) 5, p( , H) 5, 1
(iii) c(H y ,H,E)=O
(iv) if H entails E and E* and if peE) < p(*)
then c(H, E*) < c(H, E)
(v) if (H* entails H such that) 0 < p(H*) < p(H) < 1
then c(H, H) < c(H*, H*)26

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES 69

(vi) if (H* entails H such that) 0 ~ p(H*) < p(H) < 1


and p(EjH*) ~ p(EjH) then c(H*, E) < c(H, E)
The following definition is the simplest one fulfilling these six conditions Popper
has found.

c(H, E) =

p(EjH) - p(E)
p(,H)p(EjH) + p(E)

Note that c(H,H)=p(,H) and that c(H,E)=p(,E)/(p(,H)+p(E in the


case of 'deductive corroboration', that is, when H entails E.
Note first that c(H, E) is inclusive, in the sense that it can assign substantial
values when p(H) = 0 and p(EjH) can nevertheless be interpreted. In that case,
c(H,E) amounts to (p(EjH)-p(E j ,H j(p(EjH) + p(Ej ,H, which reduces
to p( I Ej I H )j[ 1 + p(Ej I H)] in the deductive case. The inclusiveness of
c(H, E) is important in view of a specific dispute of Popper with the Bayesian
approach as far as it assigns non-zero probabilities to genuine universal hypotheses. However, several authors have argued that Popper's arguments against
p-normal hypotheses (Popper 1959, Appendix *vii and *viii) fail, e.g., Earman
(1992, Section 4.3), Howson and Urbach (1989, Section 11.c) and Kuipers
(1978, Ch. 6).27
It is easy to check that c(H, E) satisfies the qualitative principle P.l of
deductive confirmation rephrased in terms of deductive (d-)corroboration:
P.lcor

if E and E* d-corroborate H then Ed-corroborates H more


than E* iff E* is more plausible than E in the light of the
background beliefs

Surprisingly enough 28 , it satisfies the (rephrased) impure alternative to P .2


favoring more probable hypotheses when equally successful:
P .2Icor

if Ed-corroborates Hand H* then Ed-corroborates H more


than H* iff H is more plausible than H* in the light of the
background beliefs

To study the qualitative features of 'general' corroboration, we will look in


some detail at the conditions (i)-(vi). The first three conditions deal with
quantitative special values, and are mainly conventional, except that (i) and
(ii) together require that, for each H, c(H, H) is the maximum value, by Popper
called the degree of corroborability, which should not exceed p( I H) . Although
Popper restricts (iv) to d-corroboration, giving rise to P.lcor, his definition of
c(H, E) satisfies the generalization of P.lcor to
P.l Gcor

if H makes E at least as plausible as E* and if E is less


plausible than E* in the light of the background beliefs, then
E corroborates H more than E*

which corresponds to P.lG. Condition (v) amounts, in combination with (i) and

70 QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES


(ii), to the idea that a less probable hypothesis should be able to get a higher
maximum degree of corroboration than a more plausible one. The Bayesian
measures d(H, E) and r(H, E) also satisfy this idea, where d(H, E) has the same
maximum value, viz., d(H, H) = p(, H), while that of r(H, E), r(H, H), equals
I j p(H). Finally, condition (vi) amounts to the qualitative idea
P.2IGcor

if H makes E at least as plausible as H* and if H is more


plausible than H* in the light of the background beliefs,
then E corroborates H more than H*

Recall that the d-degree of confirmation was impure as well, more specifically,
also favoring plausible hypotheses. Hence, it is no surprise that it also satisfies
the analogues of (vi), P.2lcor and P.2IGcor.
The foregoing comparison suffices to support the claim that the resulting
qualitative theory of Popper roughly corresponds to the impure qualitative
Bayesian theory based on d(H, E) for p-normal hypotheses. However, in contrast to d(H, E), c(H, E) is inclusive.

APPENDIX 2
COMPARISON WITH STANDARD ANALYSIS OF THE RAVEN
PARADOX

As suggested in Subsection 3.2.3., there is a more or less standard Bayesian


solution of the paradoxes of which Horwich (1982) may be assumed to have
given the best version 29 . Hence, we will compare our solution with his. After
criticizing Mackie's account (Mackie 1963) and pointing out that an unconditional approach will miss conditional connotations of the (first) raven paradox,
he introduces the idea of conditional sampling, and obtains roughly the same
confirmation claims as reported in (2p) and (4p) . However, as we will explain,
his precise quantitative versions of (2p) and (4p) are wrong, and he misses the
core of (lp) and (3p). The latter shortcoming is mainly due to overlooking a
plausible relation between the two relevant matrices. The former shortcoming
is due to a systematic mistake in the intended calculation of r-values.
There is one difference in Horwich's approach which is not really important.
He does not interpret the matrix as a survey of the sizes of the cells, but as
reporting subjective probabilities of randomly selecting a member of the cells.
It is easy to transform our matrix in this sense by dividing all numbers by their
sum (a + b + c + d), without changing any of the results.
The first serious difference is the following. We use the matrix for two
purposes, with c> 0 for calculating p(Ej , RH&C)-values, to be called the
,RH-matrix, and with c = 0 for calculating p(Ej RH&C)-values, to be called
the RH-matrix. The latter provides the numerator of r(RH,E; C) =
p(Ej RH&C)jp(Ej C) , whereas its denominator p(Ej C) is calculated by

QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES

(*)

71

p(E/C) = p(RH/ C)p(E/ RH&C)


+ p( -, RH/C)p(E/ -, RH&C)
= q . p(E/ RH&C) + (1 - q )p(E/ -, RH&C)

Recall that the simplification derives from the plausible assumption that

p(RH/C) = p(RH) = q, that is, C is neutral evidence for H.

Horwich, instead, interprets the a/b/c/d/-matrix as directly prepared for


p(E/ C)-values, to be called the HOR-matrix. Hence, he leaves open whether c
is zero or positive. This is problematic, however, for some of his general
conclusions (see below) only hold if one assumes that c in the HOR-matrix is
positive, hence that RH is false, hence that it is in fact the -, RH-matrix. As a
consequence, though he intends to directly base the p(E/ C)-value occurring in
r(RH, E; C) on the HOR-matrix, he bases it in fact on the -, RH-matrix.
However, a genuine Bayesian approach requires to calculate p(E)-values on
the basis of (*), and hence on both the RH- and the -, HR-matrix, to be called
the proper calculation.
The second main difference is that Horwich introduces an independent
RH-matrix for calculating p(E/ HR&C), to be indicated as the HOR-RH-matrix,
with IX, p, Y = 0, b, representing the relevant probabilities under the assumption
that RH is true. Horwich waves away the idea of a relation between his two
matrices. However, in his probability interpretation of the a/b/c/d-matrix, it
seems rather plausible to take IX, p, and b proportional to a, b, and d, respectively,
such that they add to 1. This corresponds to identifying the non-zero values
in our RH-matrix with a, b, and d in the size interpretation of our -, RHmatrix, as we in fact did. Why should the relative probabilities for two cells
differ depending on whether a third cell is empty or non-empty? When two
matrices are related in the suggested way they will be said to be tuned. Hence,
our RH- and -, RH-matrix are tuned and our results (lp) - (4p) directly follow
from the proper calculation presupposing these tuned matrices.
Let us call (r-)values tuned when they are based on two tuned matrices,
otherwise they are called untuned. Due to the improper calculation, Horwich
calculates in fact untuned values for the ratio p(E/ HR&C)/p(E/ -, HR&C), to
be called the ro-value (with index 0, for it corresponds to the r-ratio when
p(RH) = q = 0), instead of the intended untuned r + -values, p(E/ RH&C)/ p(E/C),
based on some q > O. Of course, when the calculated values are tuned one gets
tuned ro-values instead of tuned r + -values.
Now we can sum up Horwich's deviations from his intended proper Bayesian
treatment of the raven paradoxes. By the improper calculation, Horwich got
the wrong r-values of (2p) and (4p); he got in fact the ro-values, ro(RH, BR; R) =
(a + c)/a and ro(RH, HR; H) = (c + d)/d, which he did not intend, for he apparently assumes throughout that q is non-zero. Note that the intended qualitative
results of (2p) and (4p), i.e., (2) and (4) of Subsection 2.2.1. of the preceding
chapter, only follow when c is positive, for if c = 0, both ro-values are 1. Hence,
contrary to his suggestion of using the HOR-matrix (in which c mayor may

72 QUANTITATIVE CONFIRMATION, AND ITS QUALITATIVE CONSEQUENCES


not be 0), Horwich uses in fact the -, RH-matrix (with c > 0) for the (improper)
calculation of p(E/C).
Moreover, by not assuming tuned matrices, he missed the results (lp) and
(3p) . Regarding (lp), he does not even calculate the values for unconditional
confirmation corresponding to the uniform one we obtained, viz., (a + b + c + d)
/ (a + b + qc + d). The improper calculation would have given, assuming
the size-interpretation of all numbers, ~F/a, f3F/b, and bF/d, with F =
(a + b + c + d)/(rx + f3 + b), for a black raven, a non-black non-raven and a
black non-raven, respectively. If these values are not tuned they differ from
each other. If they are tuned, they assume a uniform value, but the wrong one,
viz., the corresponding ro-value (a + b + c + d)/(a + b + d) . Finally, regarding
(3p), Horwich does calculate the (ro-)values for sampling non-ravens and black
objects corresponding to the uniform one we obtained, viz., 1. If these values
are not tuned, they differ for all four possible results, a non-raven that is black
or non-black, and a black object that is a raven or a non-raven. However, if
they are tuned, this gives, more or less by accident, the right uniform value 1,
reported in (3p), for the '0- and the r + -value for neutral confirmation are both
equal to 1.

INDUCTIVE CONFIRMATION AND


INDUCTIVE LOGIC

Introduction
In the previous two chapters we elaborated the ideas of qualitative and quantitative theories of confirmation along the following lines. According to the
hypothetico-deductive method a theory is tested by examining its implications.
As will be further elaborated in the next chapter, the result of an individual
test of a general hypothesis stated in observation terms, and hence of a general
test implication of a theory, can be positive, negative or neutral. If it is neutral
the test was not well devised, if it is negative, the hypothesis, and hence the
theory, has been falsified. Qualitative confirmation theory primarily aims at
further explicating the intuitive notions of neutral and positive test results.
Some paradoxical features discovered by Hempel (1945/1965) and some queer
predicates defined by Goodman (1955) show that this is not an easy task. In
Chapter 2 we have presented a qualitative theory (with comparative component) which could deal with these problems. Quantitative, more specifically,
probabilistic theories of confirmation aim at explicating the idea of confirmation
as increasing probability, odds, or some other measure, due to new evidence.
In Chapter 3 we have seen that one of the Bayesian theories of confirmation,
the r-theory, is satisfactory in that it generates all main features of the qualitative theory.
In this chapter we will deal with a special explication program within the
Bayesian program, viz., the Carnap-Hintikka program of inductive logic, governed by the idea of learning from experience in a rational way. Carnap (1950)
introduced this perspective and pointed confirmation theory toward the search
for a suitable notion of logical or inductive probability. Generally speaking,
such probabilities combine indifference properties with inductive properties.
For this reason, probabilistic confirmation theory in this style is also called
inductive probability theory or, simply, inductive logic.
Carnap (1952) initiated the program more specifically by designing a continuum of systems for individual hypotheses. It has guided the search for optimum
systems and for systems that take analogy into account.
Carnapian systems, however, assign zero probability to universal hypotheses,
which excludes 'quantitative' confirmation of them. laakko Hintikka was the
first to reconsider their confirmation. His way of using Carnap's continuum

73

74

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

for this purpose has set the stage for a whole spectrum of inductive systems of
this type. l
The Carnap-Hintikka program of inductive logic clearly has considerable
internal dynamics; as will be indicated, its results can also be applied in several
directions. However, since Popper and Miller have questioned the very possibility of inductive probabilities, we will start in Section 4.1. with making clear
that a probability function may be inductive by leaving room for inductive
confirmation. In Section 4.2. Carnap's continuum of inductive methods, and
its generalization by Stegmiiller, will be presented. In Section 4.3. Festa's
proposal is sketched for estimating the (generalized) optimum method.
Section 4.4. deals with the approaches of the present author and Skyrms, and
hints upon one by Festa, related to that of Skyrms, and of di Maio, to designing
Carnapian-like systems that take into account considerations of analogy by
similarity and proximity.
Section 4.5. presents Hintikka's first approach to universal hypotheses, a
plausible generalization, Hintikka's second approach, together with Niiniluoto,
and their mutual relations. The section also hints at the extension of such
systems to polyadic predicates by Tuomela and to infinitely many (monadic)
predicates, e.g., in terms of partitions by Zabell. We will conclude by indicating
some of the main directions of application of the CarnaJrHintikka program.
The CarnaJrHintikka program may be seen as a special version of the
standard (i.e., forward) Bayesian approach to quantitative confirmation and
inductive inference. See Howson and Urbach (1989) for a handbook on the
Bayesian approach of confirmation of deterministic and statistical hypotheses,
paying some attention to the CarnaJrHintikka version and to the classical
statistical approach. There are also non-Bayesian approaches, from Popper's
degrees of corroboration (see Chapter 3) to L.1. Cohen's 'baconian probabilities'
(Cohen 1977/1991) and the degrees of confidence of orthodox statistics. Cohen
and Hesse (1980) present several quantitative and qualitative perspectives.
4.1. INDUCTIVE CONFIRMATION

We start by showing the very possibility of inductive confirmation. Popper and


Miller (1983) claimed to prove that inductive probabilities and, hence, inductive
confirmation and inductive logic are impossible. They first noted the trivial
fact that a hypothesis H may be decomposed in terms of evidence E by the
logical equivalence H <;> (E V H )&(E -+ H). Since the first conjunct is the strongest common consequence of Hand E this may be called the deductive conjunct.
The second conjunct represents as it were the remaining hypothesis, that is,
the hypothesis as far as it goes beyond the evidence, to be called the inductive
conjunct. Secondly, they noted the further trivial fact of the probability calculus
that p(E ..... HIE) is always smaller than p(E ..... H)(the proof is left as an exercise).
Assuming that a genuine inductive probability function should leave room for
the opposite possibility, that is, for confirmation of the inductive component,

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

75

in the sense defined above, this second fact seems dramatic. Their paper
generated a lot of responses. The impossibility 'proof' of inductive probabilities
is an illustration of the fact that a prima facie plausible explication in terms of
logical consequences needs to be reconsidered in the face of clear cases, i.c.,
inductive probabilities. In Section 8.1. we will show in detail that Popper was
also unlucky with explicating the idea of verisimilitude in terms of logical
consequences and that an alternative explication is feasible. In this chapter we
will present the main lines of some paradigm examples of inductive probabilities
and the corresponding general explication of such probabilities.
The explication of what is typical of inductive as opposed to non-inductive
probabilities presupposed in the Carnap-Hintikka program simply is the satisfaction of the principle of positive instantial relevance, also called the principle
of instantial confirmation or the principle of inductive confirmation:
(lC)
where, phrased in terms of individuals ai' a2 , . , R is a possible property of
them and en reports the relevant properties of the first n (;~ 0) individuals.
Moreover, it may generally be assumed that p(R(an+1)/en) = p(R(a n+ 2)/en ), with
the consequence that (IC) becomes equivalent to
p{R{an +2)/en &R(a n + d) > p(R(a n +2)/e n )

It is beyond doubt that (lC) extrapolates information contained in the evidence


to the hypothesis in a certain way: an extra occurrence of an 'R-individual'
increases the probability that the next individual is an R-individual. The qualitative version of (IC) might be phrased similarly: an extra occurrence of an 'Rindividual' increases the plausibility (or expectation) that the next individual is
an R-individual. As we will see, it is perfectly possible to construct consistent
probability functions with the property (lC), despite the fact that also in this
case the probability of the relevant 'remaining hypotheses' in the sense of
Popper and Miller decreases, e.g., p(e n -+ R{a n+ d /en ) < p(e n-+ R(a n + d)
In the context of a number of mutually exclusive and proper universal
generalizations, a probability function may be inductive for two different
reasons, and a combination of them. In this case we have
p(R(an+d/en) = r..iP(H;)p(en&R(a n+ l)/Hi)/p(e n)

generated by a prior distribution p(H i ) and the likelihood functions P(.jHi) ' If
the prior distribution is non-inductive, that is, the improper universal hypothesis
is the only one that gets a non-zero probability, hence 1, the system may
nevertheless be inductive due to the fact that the only relevant likelihood
function is inductive. Carnap-systems amount to such systems. On the other
hand, if the prior distribution assigns non-zero probabilities to all relevant
genuine universal hypotheses the phenomenon of inductive confirmation occurs
already even if the likelihoods are themselves non-inductive. This second combination is more or less the standard Bayesian approach. Hence, Carnap deviates

76

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

from this standard by a non-inductive prior distribution and an inductive


likelihood function. As we will see, Hintikka deviates from the standard
approach by assuming both an inductive prior distribution and inductive
likelihoods (of a Carnapian nature). Of course, Popper basically prefers a purely
non-inductive approach: neither inductive priors nor inductive likelihoods.
Before we deal with the question whether there is a plausible degree of
inductive confirmation, we want to indicate to what extent an inductive probability function is inductive and to what extent non-inductive. A plausible
measure of the 'inductive influence' in e.g., p(E) is its difference with the
corresponding non-inductive probability function, to be called the inductive
influence or the inductive difference in p(E). Hence, the question is what that
non-inductive function is. This function has already been formulated in a very
general sense by Kemeny in his "A logical measure function" (Kemeny 1953),
which we will indicate, like him, by m. It amounts in fact to the probabilistic
version of the idea of 'partial entailment' of Salmon (1969). For finite, interpreted languages, with finitely many models, say M, m(S) of a sentence S is defined
as the number of models of that sentence N(S) divided by M, i.e., N(S)/M . In
many cases it is plausible how to extrapolate the definition to an infinite
language 2 Of course, truly 'infinite sentences', hence genuine universal hypotheses, will get the m-value O.
The difference measure now leads to p(E) - m(E) for the inductive influence
in p(E). For the inductive influence in p(R(a n+ d l en) we get p(R(a n+ d l en)
- m(R(a n+ 1 )/en), where m(R(a. + 1 )/e.) is, as a direct consequence of the definition, equal to m(R(an + d , and it may be assumed that p(R(a.+d=m(R(an+d).
Assuming that R is one of a number of mutually exclusive and together
exhaustive predicates, it is plausible to try to define a measure for the total
inductive influence in the prediction of the property of the next individual. It
is of course not possible just to sum up the differences for that would always
simply lead to 1. The sum of the squares of the differences is the plausible
alternative, which amounts to :Ep2(R(a n+ d i e.) - :Em2(R(a n+ d), which equals 0
when p = m, as it should be.
Similarly, the inductive influence in p(H), for a genuine universal hypothesis
H , is p(H) - m(H) = p(H). Moreover, the total inductive influence in the prior
distribution may be equated with the sum of the 'proper priors', which equals
of course 1 minus the 'improper prior', which amounts to 0 in the case of a
non-inductive prior distribution, again as it should be. In the case of an
inductive prior distribution and inductive likelihoods there are two types of
inductive influence, via the priors and via the likelihoods. Such functions may
be called 'double-inductive'.
The inductive influence in p(EI H) and p(HI E) can be measured in an analogous way by p(EI H) - m(EI H) and p(HI E) - m(HI E), where the latter equals
of course p(HI E) for genuine universal hypotheses, since m(HI E) will be O.
Let us finally consider the inductive influences in the degree of confirmation.
Recall that if p(H) > 0, e.g., in the case of inductive priors (standard Bayesian

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

77

or Hintikka), the two criteria of general confirmation coincide: p(H/ E) > p(H)
(the standard forward (F-)criterion) iff p(E/H) > prE) (the backward or success
(S-)criterion), in which case the ratio degree of confirmation r(H, E) may either
be defined as p(H/E)/p(H) or, equivalently, p(E/H)/p(E). However, if p(H) = 0,
e.g., in the case of non-inductive priors (Popper and Carnap), the two criteria
deviate; only the S-criterion allows confirmation. In this case the degree of
confirmation was identified with the corresponding ratio, i.e., p(E/H)/p(E),
which is well-defined, assuming that p(E/ H) can be interpreted.
Let us indicate the degree of confirmation of H by E according to p by
r p(H, E), and that according to m by r m(H, E). In view of the ratio structure of
the two degrees it is now plausible to take the ratio of the former to the latter
as the degree of inductive confirmation, rj(H, E) = r p(H, E)/rm(H, E).3 An
attractive consequence of this definition is that the degree of inductive confirmation is simply equal to the degree of confirmation (according to p) when
there is no non-inductive confirmation. E.g., assuming (IC), the (backward =
forward) degree of inductive confirmation of R(a.+d bye. according
to p is p(R(a. + d /e.)/ p(R(a. + d. This holds of course also for the corresponding conditional degrees of confirmation: rj(R(a. + 2), R(a. + 1); e.) =
r p(R(a.+ 2 ), R(a. + 1); e.) = p(R(a. + 2)/e.&R(a.+ 1)/p(R(a. + 2)/e.).
The proposed (conditional) degree of inductive confirmation seems rich
enough to express the inductive influences in cases of confirmation when p(H)
and m(H) are both non-zero or both O. In the first case, the F- and S-criterion
coincide for both p and m, and the corresponding (conditional) ratio measures
are uniquely determined. In the second case, only the success criterion can
apply, and the corresponding ratio measures can do their job. However, the
situation is different if p(H) is non-zero and m(H) = 04 , which typically applies
in case of an inductive prior, e.g., standard Bayesian or Hintikka-systems. In
Section 4.5. we will explicate how forward confirmation then deviates from
success confirmation.
In the rest of this chapter we will only indicate in a few number of cases the
resulting degree of confirmation simpliciter and the degree of inductive confirmation, in particular when they lead to prima facie surprising results.

4.2 . THE CONTINUUM OF INDUCTIVE SYSTEMS s

Mainly by his The Continuum of Inductive Methods (1952) Carnap started a


fruitful research program centering around the famous 2-continuum 6 . The
probability systems in this program can be described in terms of individuals
and observation predicates or in terms of trials and observable outcomes. The
latter way of presentation will be used here in an informal way. Moreover, we
will presuppose an objective probability process, although the systems to be
presented can be applied in other situations as well.
Consider a hidden wheel of fortune. You are only told, truthfully, that it has

78

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

precisely four colored segments, BLUE, GREEN, RED, and YELLOW, without further information about the relative size of the segments. So you do not
know the objective probabilities. What you subsequently learn are only the
outcomes of successive trials. Given the sequence of outcomes en of the first n
trials, your task is to assign reasonable probabilities, p(R/e.), to the hypothesis
that the next trial will result in, for example, RED.7
There are several ways of introducing the ),-continuum, but the basic idea
behind it is that it reflects gradually learning from experience. According to
Carnap's favorite approach p(R/e.) should depend only on n and the number
of occurrences of RED thus far, nR . Hence, it is indifferent to the color
distribution among the other n - nR trials. This is called the principle of
restricted relevance. More specifically, Carnap wanted p(R/e.) to be a special
weighted mean of the observed relative frequency nR/n and the (reasonable)
initial probability 1/4. This turns out to leave room for a continuum of
(C-)systems, the ),-continuum, 0 <), < 00, for four outcomes:

p(R/e.) = (nR + )./4 )/(n + ).) = (n/(n + ).)) . (nR/n) + (J./(n + },))( 1/4)
Note that the weights n/(n + ).) and )J(n + ).) add up to 1 such that for increasing
n the former gradually increases from 0 to 1 and hence the latter from 1 to O.
Moreover, the larger the value of ), the slower the first increases, at the expense
of the second, i.e., the slower one is willing to learn from experience.
C-systems have several attractive indifference, confirmation and convergence
properties (for a systematic treatment see (Kuipers 1978)). The most important
are these.

- initial indifference, p(R/eo ) = p(R) = 1/4,


- order indifference or exchangeability: the resulting prior probability, p(e.),
for e., does not depend on the order of the results of the trials, i.e.,
p(e.) = p(e:) for any permutation e: of e.; this property implies
- conditional order indifference or conditional exchangeability, i.e.,
p(RB/e.) = p(BR/e.), where the first probability concerns the hypothesis
that, given e., first RED and then BLUE will occur, and the second the
reverse order;
- instantial confirmation: if e. is followed by RED this is favorable for
RED, i.e., p(R/e.R) > p(R/e.), hence such that the corresponding conditional degree of (purely) inductive confirmation p(R/enR)/p(R/e.) exceeds
1. Moreover, the inductive influence in p(R/e.) amounts to p(R/en) - 1/4,
which is equal to (n/(n + J.))(nR /n - 1/4).
- instantial convergence: p(R/e.) approaches nR /n for increasing n.
C-systems satisfy moreover:

- universal-instance confirmation, i.e., suppose for example that RED and


BLUE are the only colors that have occurred according to en, then
p(R VB/e.) < p(R V B/e.R V B). The name 'universal-instance confirmation' is plausible in view of the fact that the first probability

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

79

concerns the hypothesis that, given en, the next result will also be RED
or BLUE, i.e., will be in agreement with the universal hypothesis that
all future results are RED or BLUE, and the second probability concerns
the same hypothesis for the next trial, but now not only given en, but
also that it was followed by RED or BLUE, i.e., a result in agreement
with the universal hypothesis,
- universal-instance convergence, i.e., p(R V B/en ) approaches 1 for increasing n if only RED and BLUE remain occurring according to en'
Though C-systems have the properties of universal-instance confirmation
and convergence, forward confirmation of, and convergence to, universal
hypotheses are excluded. The reason is that C-systems in fact assign zero prior
probability to all universal generalizations, for instance to the hypothesis that
all results will be RED. Of course, this is desirable in the described situation,
but if you were only told that there are at most four colored segments, you
would like to leave room for this possibility. Success confirmation of universal
hypotheses is, of course, not excluded. For simplicity, we will deal with it in
Section 4.5.
Stegmliller (1973) generalized C-systems to (GC-)systems, in which the uniform initial probability 1/4 is replaced by arbitrary non-zero values, e.g., p(R),
leading to the special values:
p(R/en )

= (nR + p(R) . ),)/(n + ).)

Apart from the intended weakening of restricted relevance all properties of


C-systems continue to apply.
Johnson (1932) anticipated C-systems already in 1932, including their axiomatic foundation . He introduced C-systems as a natural representation of symmetric so-called Dirichlet distributions, which are well-known in Bayesian
statistics. Zabell (1982) rediscovered this and showed that non-symmetric versions are GC-systems. For a detailed analysis of the relation between
GC-systems and Dirichlet distributions, see Festa (1993). Skyrms ( 1993a) points
out that the analogue of C- and GC-systems for a continuum of possible values
can be found in the statistical literature.

4.3. OPTIMUM INDUCTIVE SYSTEMS

Carnap (1952) proved that for certain kinds of objective probability processes,
such as a wheel of fortune, there is an optimal value of A, depending on the
objective probabilities, in the sense that the average mistake may be expected
to be lowest for this value. Surprisingly, this optimal value is independent of
n. In fact he proved the existence of such an optimal C-system for any multinomial process, i.e., sequences of independent experiments with constant objective
probabilities qi for a fixed finite number of k outcomes. The relevant optimal

80

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

value of A. is given by:


A.... (C) = (1 - r.qf)/((r.qf) - 11k)

Of course, in actual research, where we do not know the objective probabilities,


this optimal value cannot be calculated. Carnap did not raise the question of
a reasonable estimate of the optimal value for a specific objective process, for
he saw the problem of selecting a value of ;. primarily as a choice that each
scientist had to make in general. However, the question of a reasonable estimate
has attracted the attention of later researchers.
In reaction to some 'hyperempiricist' estimations of the optimal value of 8
(Kuipers 1986), solely based on the available evidence, Festa (1993)8 proposed
basing the estimate on 'contextual' knowledge of similar processes in nature.
For example, in ecology one may know the relative frequencies of certain
species in different habitats before one starts the investigation of a new habitat.
Festa first generalizes the proof of the existence of an optimal value of ;. to the
classes of GC-systems with the same initial probabilities Yi , leading to the
optimal value:
A. ... (GC) = (l - r.qf)/(r.(qi - ('i)2)

which is easily checked to reduce to ;.... (C) when all fi equal 11k.
Then he formulates for such a class two solutions of the estimation problem,
which turn out to be essentially equivalent. Moreover, one of these solutions
relates the research area of inductive logic to that of truth approximation: the
optimum solution may be expected to be the most efficient way of approaching
the objective or true probabilities. As a matter of fact, the present probabilistic
context provides an example of truth approximation where a quantitative
approach is not only possible, but even plausible (See the end of Subsection 12.3.2. for the precise link).
Unfortunately, wheels of fortune do not constitute a technological (let alone
a biological) kind, for which you can use information about previously investigated instances. But if you had knowledge of a random sample of all existing
wheels of fortune, Festa's approach would work on the average for a new,
randomly drawn, one.

4.4. INDUCTIVE ANALOGY BY SIMILARITY AND PROXIMITY

Carnap struggled with the question of how to include analogy considerations


into inductive probabilities. His important distinction between two kinds of
analogy - by similarity and by proximity - was posthumously published
(Carnap 1980).
Suppose you find GREEN more similar to BLUE than to RED. Carnap's
intuition of analogy by similarity, here called Carnap-analogy, is that, for
instance, the posterior probability of GREEN should be higher if, other things

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

81

being equal, some occurrences of RED are replaced by occurrences of BLUE.


The background intuition, not valid for artificial wheels of fortune, is that
similarity in one respect (e.g., colors) will go together with similarities in other
respects (e.g., objective probabilities). However, explicating what precise principle of analogy one wants to fulfil, has turned out to be very difficult.
Niiniluoto (1988) discusses, among others, proposals for inductive analogy
by similarity by D. Costantini, T. Kuipers, I. Niiniluoto and W. Spohn. See
also (Welch 1999) for a useful survey. Here we restrict ourselves to the 'virtual
approach' in (Kuipers 1984b, 1988). It goes in terms of virtual trials, which is
a way of determining how analogy influence is to be distributed after n trials.
To present the main idea, it is convenient to start from another heuristic for
introducing C-systems. A special value p(R/en ) can be seen as the ratio of nR
real plus ,1/4 virtual occurrences of RED and n real plus ;, virtual trials.
Similarly, we may introduce virtual occurrences and trials, accounting for
analogy influences. The number of virtual analogy occurrences O:R(en ) obtained
by RED depends on the particular number of occurrences of the other colors
and the distances of them to RED, expressing their dissimilarity. To put it in
a nutshell for the case indicated above, BLUE gets more analogy credits than
RED from an occurrence of GREEN. The (total) number of virtual analogy
trials o:(n) determines how much analogy influence in total is to be distributed
after n trials. The resulting special value is:

peR/en) = (nR + O:R(en ) + ;';4 )/(n + o:(n) + i.)


For reasonable restrictions on o:(n) and relations between CiR(e n) and the distances of RED to the other outcomes, the reader is referred to (Kuipers 1984b).
It is interesting to note, and easy to check, that Carnap's favorite way of
presenting C-systems can now also be extended to the special values accounting
for analogy. They can be rewritten as the weighted mean of the observed
relative frequency, the initial probability and an analogy factor.
The resulting systems of virtual analogy have the same confirmation and
convergence properties as C-systems. Moreover, they satisfy a principle of
analogy, called virtual analogy, that is in general more plausible than Carnap's
principle indicated above. They only satisfy Carnap-analogy, indicated above,
under special conditions. Roughly, the new principle of virtual analogy says
that replacement of a previous occurrence of BLUE by GREEN makes less
difference for the probability that the next trial will be BLUE, than replacement
of a previous occurrence of RED by GREEN makes for the probability that
the next trial will be RED.
The resulting conditional degree of inductive confirmation with analogical
influences is straightforwardly defined. It is also plausible to define the resulting
analogy influence in p(R/en) as peR/en) - pdR/en), where Pc refers to the corresponding Carnapian system. Moreover, peR/en) may of course be written as
[1/4] + [pdR/en) - 1/4] + [peR/en) - pdR/en)] , that is, as the sum of the initial
probability, the inductive influence and the (resulting) analogy influence.

82

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

Unfortunately, systems of virtual analogy are not exchangeable: the order of


results makes a difference, e.g., the initial and posterior probability of the
sequence RB differs in general from that of BR. There are, however, exchangeable systems with some kind of analogy by similarity. For example, Skyrms
(1993b) has sketched a Bayesian approach that uses information about possible
gravitational bias of a symmetric wheel of fortune to determine the influence
of analogy. To be more precise, Skyrms's basic idea is the following. Let (only
in this paragraph) the wheel of fortune be symmetric, i.e., let the four segments
be of equal size, with the possibility that it is gravitationally biased in favor of
one color, the most strongly at the expense of the opposite color and moderately
at the expense of the remaining two. Analogy influences may now be objectively
based on the resulting four possibilities and implemented in a Bayesian conditionalized probability system along the following lines. Assuming that we know
the relative positions of the colored segments, but do not know in favor of
which color the possible bias is, suggests four metahypotheses, to be implemented in a Bayesian conditionalized probability system by imposing for each
conditional system a GC-system with suitable initial values in accordance with
the relevant metahypothesis. For example, the value 0.5 for the favored color,
0.1 for the opposite one and 0.2 for the two other colors. Starting with equal
prior probabilities for the metahypotheses, it is evident that the resulting
'hyperCarnapian' system realizes influences of analogy by similarity of sorts.
What precise kind of analogy properties these systems have is an interesting question. Festa (1996) shows that the general class of symmetric
hyperCarnapian systems, with equally many metahypotheses as possible outcomes (k), suggested by Skyrms (Sk-systems), may start with satisfying the
conditions of Carnap- and virtual analogy, but may lose it after a certain
number of trials. Both are proven to be the case for the specific system proposed
by Skyrms. However, Festa also proves that a very special class of
hyperCarnapian (HC-)systems, viz., those with only three outcomes, generally
satisfy Carnap-analogy. Unfortunately, the further prospects for HC-systems
to generally satisfy Carnap- or virtual analogy for all relevant triples of outcomes are as yet unclear.
It is also possible to construct a completely different kind of exchangeable
system that have analogy by similarity influences of another sort (Kuipers
1984b). They are obtained by assuming the right kind of prior distribution in
systems leaving room for universal generalizations, to be dealt with in the next
section. Again, though there evidently are analogy effects in the intended sense,
it is difficult to catch these effects in general analogy properties.
Another exchangeable approach is due to Oi Maio (1995). She weakens
restricted relevance by a combination of a plausible suggestion by Carnap
(1980) with a newly invented, also plausible, principle in such a way that it
leads to exchangeable systems, most of which are sensitive to analogy by
similarity. However, the resulting systems still are everything but transparent.
Future research has to make clear whether there is any relation, or at least

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

83

compatibility, between the three indicated exchangeable approaches. However,


so far they seem to have one feature in common: unlike the virtual approach,
the analogy effects of the exchangeable approaches are not easy to determine
or even to express. Such features may be considered as a high price for the
property of exchangeability, but it may also be argued that the latter property
deserves a high price. Besides pragmatic reasons, viz., it makes systems easier
to handle in certain respects, the fundamental reason in favor of exchangeability
is of course that the paradigmatic underlying objective probability processes
are exchangeable, such as (biased or unbiased) wheels of fortune with constant
probabilities.
Carnap also set the stage for analogy by proximity: differences may arise due
to the number of trials between different occurrences of the same outcome.
This may concern occurrences in the past, i.e., in the evidence, or in the future,
i.e., in the hypothesis. Both are studied in (Kuipers 1988). For past influences
it is again possible to use the idea of virtual occurrences. For future proximity
influences a general approach is still missing, but the problem also redirects
the attention to the question of how to approach a so-called Markov chain.
Approaching it piecemeal by using C-systems as component systems is the
plausible idea (Kuipers 1988), which has been elaborated by Skyrms (1991a)
in Bayesian terms and derived by Zabell (1995) by an extension of Johnson's
(1932) axiomatic way of deriving C-systems.

4.5. UNIVERSAL GENERALIZATIONS 9

Carnap was well aware that C-systems do not leave room for non-analytic
universal generalizations. In fact, apart from one type of system in the previous
section, all systems presented thus far have this problem. Although tolerant of
other views, Carnap himself was inclined to downplay, for this and other
reasons, the theoretical importance of universal statements. However, as already
mentioned, if you are only told that the wheel of fortune has at most four
colored segments, you might want to leave room for statements claiming for
example that all results will be RED.
Hintikka (1966) took up the problem of universal statements. His basic idea
for such a system, here called H-system, is a Bayesian conditionalized probability system of the following double-inductive kind: (1) a prior distribution
of non-zero probabilities for (the outcome spaces of) mixed (existential and
universal) statements, called constituents, stating that all and only the colors
so-and-so will occur in the long run, (2) for each of these constituents an
appropriate C-system as conditional system, (3) the constituents in (1) and (2)
dealing with sets of colors of the same size (w) get the same prior values, p(R ",),
and the corresponding conditional C-systems not only have the same initial
values (l/w) for all relevant outcomes, but also the same A-values, A"" (4)
Bayes's theorem for conditionalizing posterior probabilities is applied. This

84

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

leads, amongst others, to tractable formulas for the special values, such as
p(Rje n), and for the posterior values p(H w/en).
Hintikka, more specifically, proposed a two-dimensional system, the so-called
a-A-continuum, by taking a uniform value of ;, for all conditional systems and
the prior distribution proportional to the probability that a (not to be confused
with a in the previous section) virtual individuals or trials are, according to
the C-system for all possible colors, compatible with the relevant constituent.
Here a is a finite number to be chosen by the researcher. The larger it is chosen,
the lower the probability of all constituents based on fewer outcomes than
possible.
However, as shown in (Kuipers 1978), H-systems in general have already,
besides the instantial, and universal-instance, confirmation and convergence
properties, the desired property of (forward) universal confirmation, i.e., the
probability of a not yet falsified universal statement increases with another
instance of it. For example, assuming that only RED occurred according to en'
p(HR /en) is smaller than p(HR /enR), where HR indicates the constituent that
only RED will occur in the long run. H-systems have, moreover, universal
convergence, i.e., the probability of the strongest not yet falsified universal
statement converges to o. For example, p(HR jen) approaches 1 for increasing n
as long as only RED continues to occur. For the special case of H-systems
belonging to Hintikka's a-I,-continuum holds that, for increasing parameter a,
universal confirmation is smaller and universal convergence slower.
The conditional degree of (purely) inductive confirmation corresponding to
instantial confirmation is straightforwardly and uniquely defined for H-systems.
However, the conditional degree of (inductive) confirmation corresponding to
universal confirmation has two sides. Let us first consider the success side
(story). Since p(R/HRe n) = 1, assuming that en reports n times R, the conditional
degree of (backward and deductive) confirmation of H R by R, given en' according to p, denoted by rp(HR' R; en), is p(RjHRen)/p(R/en) = l/p(R/en). The corresponding degree for the m-function is 1/m(R/en) = 4. Since nR = np(R/en) will
be, due to 'recurring' instantial confirmation, larger than 1/4, rj(H R, R; en) =
m(Rjen)/p(Rjen) = (1 /4 )/p(R/e n) becomes smaller than 1. Although both degrees
report confirmation, both cases even concern in fact deductive confirmation, it
may be surprising at first sight that the 'p-degree' is smaller than the 'm-degree',
but on second thoughts it is not. An inductive probability function invests as
it were its inductive attitude towards uniform evidence a priory, with the
consequence that the explanatory success of a universal hypothesis is downgraded when it is repeatedly exemplified. On the other hand, the non-inductive
function m acknowledges each new success of a universal hypothesis as equally
successful as the previous ones. In other words, since both degrees of confirmation may also be seen as degrees of success, it is plausible that this degree is
larger according to the m-function than according to an inductive p-function, which by definition anticipates relatively uniform evidence lO Note
that the success confirmation of H R by R, given n times R, occurs of course

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

85

also when p is a C-system. In that case, ri(H R, R; en) becomes equal to


(1/4)(n + A)/(n + ),/4), which is again smaller than 1.
Returning to H-systems, with an inductive prior distribution, it may be
expected that in terms of forward confirmation there is some kind of (purely)
inductive confirmation of universal hypotheses. We can make this explicit by
introducing the degree of forward confirmationfr(H, E) = p(H/E)/p(H), which
equals the degree of (backward) confirmation r(H, E) when p(H) 0, and which
we define as 1 when p(H) = 0. In line with this, the degree of inductive forward
confirmationfri(H, E) is defined asfrp(H, E)/frm(H, E). This degree equals the
earlier defined degree of inductive confirmation when p(H) and m(H) are both
0, it equals 1 when both are 0, and it equals fr p(H, E) = r p(H, E) when
p(H) > = m(H). In the last case, the forward confirmation is purely inductive,
as was to be expected. Of course, conditional forward degrees are defined in a
similar way as conditional backward degrees. Returning to H-systems, the
resulting conditional forward degree of (purely) inductive confirmation of
HR by R, given n times R, that is, fri(H R, R; en) = frp(HR' R; en), equals
p(H R/enR)/p(H R/e n). The claim that this degree exceeds 1 is trivially equivalent
to the definition of forward universal confirmation in this case, and which is
not difficult to prove for H-systems.
An easily definable and transparent special class of H-systems, SH-systems,
has become particularly interesting from a finitistic point of view, viz., the
subclass of H-systems based on an arbitrary prior distribution and with conditional C-systems with ). proportional to the corresponding number of possible
outcomes. Note that, although SH-systems may have Hintikka's originally
proposed prior distribution, they nevertheless do not belong to the C(-I.-continuum in that case, due to the non-uniform character of ;. in SH-systems.
SH-systems became interesting due to a prima facie completely different type
of system introduced by Hintikka and Niiniluoto (1976/1980). Kemeny (1963)
showed already that for three or more outcomes C-systems could be axiomatically introduced by imposing the indifference properties of exchangeability and
(positive) restricted relevance. By weakening restricted relevance in some way
Stegmiiller extended this to the already mentioned GC-systems (see
Section 4.2.).11 Hintikka and Niiniluoto showed that exchangeability and
another weakening of restricted relevance, according to which p(R/e n ) may not
only depend on nand nR but also on the number of different colors reported
in en, leads to a very interesting class of systems. Such NH-systems are determined by 'finite' principles and parameters, i.e., principles and parameters
related to finite sequences of outcomes.
Now, the theorem can be proved (Kuipers 1978) that the classes of NHand SH-systems coincide. The consequence is that SH-systems can be axiomatically introduced as NH-systems. After calculating the corresponding prior
distribution, the attractive confirmation and convergence properties and many
other characteristics can be studied by the transparent SH-representation of
these systems.

"*

86

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

H-systems, and hence SH/ NH-systems, can be generalized by using plausibly


related initial probabilities for the conditional systems. Such GH- (GSH/GNH-)
systems keep all properties, except some further weakening of restricted relevance for the conditional C-systems and the NH-systemsY Tuomela (1966)
showed how Hintikka's approach to polyadic predicates or ordered outcomes
can be extended. Whether the same idea can be used to extend H-systems and
even GH-systems to include polyadic predicates still needs to be investigated.
There is a plausible 'delabeled' reformulation of H-systems, hence of SH/NHsystems, which can be extended to the very interesting case of an unknown
denumerable number of possible outcomes (Kuipers 1978). Presenting the
delabeling in terms of exchangeable partitions, Zabell has studied this extension
by principles, in strength between those of C- and NH -systems, leading to an
interesting class of systems with three parameters (Zabell 1996b), anticipated
in (Zabell 1992).

Concluding remarks
In contrast to Lakatos's claim (Lakatos 1968) the Carnap-Hintikka program
in inductive logic, combines a strong internal dynamics with applications in
several directions. The strong internal dynamics will be clear from the foregoing.
Let us finally indicate some of its intended and unintended applications. Systems
of inductive probability were primarily intended for explicating confirmation
in terms of increasing probability. As for probabilistic Bayesian confirmation
theory in general, one may define a quantitative degree of confirmation in
various ways (see Chapter 3), but we argued for the ratio measure. However
this may be, we still have to address the question whether we need and are
justified to use inductive probabilities for confirmation purposes. For it is one
thing to claim that they can be used to describe or explicate a kind of supposedly
rational behavior towards hypotheses in relation to evidence, this does not
answer the question whether we are justified in using them and whether we
really need them. To begin with the latter, it is clear that the success theory of
confirmation (based on the S-criterion and the ratio degree) applied to the
non-inductive probability function m is essentially sufficient to evaluate the
confirmation of H by E. This approach may be called Popperian, although
Popper's own proposal for the degree of corroboration (!) is somewhat different,
and, as we have seen in Appendix 1 of Chapter 3, surprisingly enough, such
that it entails the Matthew-effect. The question whether we are justified in
using inductive probabilities brings us back to the Principle of the Uniformity
of Nature (PUN). If PUN is true in some sophisticated sense, e.g., in the sense
that there are laws of nature of one kind or another, the use of inductive
probabilities makes sense. However, justifying PUN is another story. Hume
was obviously right in claiming that PUN cannot be justified in a non-circular
way. Nevertheless, all living beings happen to have made, in the process of

INDUCTIVE CONFIRMATION AND INDUCTIVE LOGIC

87

biological evolution, the inductive jump to PUN for everyday purposes with
success. Hence, so far PUN was rewarding indeed. We may also conclude from
this fact that something like a high probability rule of acceptance is defensible.
A remaining question then is how it is possible that Popper, who denounces
PUN but believes in laws of nature (Popper 1972, p. 99), apparently is able to
circumvent the use of inductive probabilities. The point is, of course, that every
time a Popperian scientists makes the, relatively great, jump from merely a
hypothesis to a hypothesis which is, for the time being, accepted as true, the
leap is just not prepared by inductive 'mini-leaps'. That Popper refused to call
the big jumps he needed inductive is ironical.
Let us finally turn to other than confirmation applications of inductive
probabilities. Carnap (1971a) and Stegmiiller (1973) stressed that they can be
used in decision making, and Skyrms (1991b) applies them in game theory by
letting players update their beliefs with C-systems. Costantini et al. (1982) use
them in a rational reconstruction of elementary particle statistics. Festa (1993)
suggests several areas of empirical science where optimum inductive systems
can be used. Welch (1997) indicates how systems with virtual analogy can be
used for conceptual analysis in general and for solving conflicts about ethical
concepts in particular. Finally, for universal hypotheses, Hintikka (1966),
Hilpinen (1968)13 (see also Hintikka and Hilpinen 1966) and Pietarinen 1972)14
use systems of inductive probability to formulate rules of acceptance, i.e., rules
for inductive jumps, the task of inductive logic in the classical sense. As we
noted already in Section 3.3., their main proposal for the IX-},-continuum
successfully avoids the lottery-paradox, and could be extended to another class
of systems of inductive probability, in fact to all SH/ NH-systems (Kuipers
1978, Section 8.6.). However, they remain restricted to monadic predicate
languages. As will be indicated in Chapter 12, Niiniluoto (1987a) uses such
systems to estimate degrees of verisimilitude, hence giving rise to a second
fusion of inductive logic and the truth approximation program. IS
With closing this chapter we also close the part about confirmation. Recall
what we stated already at the end of Chapter 2. Although the role of confirmation (and falsification) will be strongly relativized, there is a strong need for
a sophisticated account of confirmation. The reason is that the notions of
confirmation and falsification remain of crucial importance for testing at least
three types of hypotheses, viz., general test implications, comparative success
hypotheses, and truth approximation hypotheses. In the next part we turn to
the first two types of hypotheses and the role of confirmation and falsification
in the separate and comparative evaluation of them.

PART II

EMPIRICAL PROGRESS

INTRODUCTION TO PART II

Confirmation of a hypothesis has the connotation that the hypothesis has not
yet been falsified. Whatever the truth claim associated with a hypothesis, as
soon as it has been falsified, the plausibility (or probability) that it is true in
the epistemologically preferred sense is and remains nihil. Hence, from the
forward perspective, the plausibility of the hypothesis cannot increase. Similarly,
from the success perspective, the plausibility of evidence which includes falsifying evidence, on the condition of the hypothesis, is and remains nihil, and
hence cannot increase. In this part we will elaborate how the evaluation of
theories can nevertheless be proceeded after falsification.
In Chapter 5 the attention is directed at the more sophisticated qualitative
HD-evaluation of the merits of theories, in terms of successes and counterexamples, obtained by testing test implications of theories. The resulting evaluation report leads to three interesting models of separate HD-evaluation of
theories. Special attention is paid to the many factors that complicate
HD-evaluation and, roughly for the same reasons, HD-testing.
In Chapter 6 it is pointed out that the evaluation report resulting from
separate evaluation naturally leads to the comparative evaluation of theories,
with the crucial notion of 'more successfulness', which in its turn suggests 'the
rule of success', which marks empirical progress. It will be argued that this
'instrumentalist' or 'evaluation methodology', by denying a dramatic role for
falsification, and even leaving room for some dogmatism, is methodologically
superior to the 'falsificationist methodology', which assigns a theory eliminative
role to falsification. Moreover, the former methodology will be argued to have
better perspectives for being functional for truth approximation than the latter.
Finally, the analysis sheds new light on the distinction between science and
pseudoscience.

91

5
SEPARATE EVALUATION OF THEORIES BY THE
HD-METHOD

Introduction
In Chapter 2 we gave an exposition of HD-testing, the HD-method of testing
hypotheses, with emphasis on the corresponding explication of confirmation.
HD-testing attempts to give an answer to one of the questions one may be
interested in, the truth question, which may be qualified according to the
relevant epistemological position l . However, the (theory) realist, for instance,
is not only interested in the truth question, but also in some other questions.
To begin with, there is the more refined question of which (individual or
general) facts 2 the hypothesis explains (its explanatory successes) and with
which facts it is in conflict (its failures), the success question for short. We will
show in this chapter that the HD-method can also be used in such a way that
it is functional for (partially) answering this question. This method is called
HD-evaluation, and uses HD-testing. Since the realist ultimately aims to
approach the strongest true hypothesis, if any, i.e., the (theoretical-cum-observational) truth about the subject matter, the plausible third aim of the HD-method
is to help answer the question of how far a hypothesis is from the truth, the
truth approximation question. Here the truth will be taken in a relatively modest
sense, viz., relative to a given domain and conceptual frame. In the next chapter
we will make plausible and in Part III prove that HD-evaluation is also
functional for answering the truth approximation question.
As we will indicate in a moment, the other epistemological positions are
guided by two related, but more modest success and truth approximation
questions, and we will show later that the HD-method is also functional for
answering these related questions. But first we will articulate the realist viewpoint in some more detail. For the realist, a hypothesis is a statement that may
be true or false, and it may explain a number of facts. A theory will here be
conceived as a hypothesis of a general nature claiming that it is the strongest
true hypothesis, i.e., the truth, about the chosen domain (subject matter) within
the chosen conceptual frame (generated by a vocabulary). This claim implies,
of course, that a theory claims to explain all relevant facts. Hence, a theory
may not only be true or false, it may also explain more or fewer facts, and it
may even be more or less near the truth. To be sure, presenting the realist
notion of a theory as the indicated special kind of hypothesis is to some extent
93

94

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

a matter of choice, which will turn out to be very useful. The same holds for
adapted versions of the other epistemological positions.
Let us briefly look at the relevant questions from the other main epistemological viewpoints, repeating the relevant version of the truth question. The constructive empiricist is interested in the question of whether the theory is
empirically adequate or observationally true, i.e., whether the observational
theory implied by the full theory is true, in the refined success question what
its true observational consequences and its observational failures are, and in
the question of how far the implied observational theory is from the strongest
true observational hypothesis, the observational truth. The referential realist is,
in addition, interested in the truth of the reference claims of the theory and
how far it is from the strongest true reference claim, the referential truth. The
instrumentalist phrases the first question of the empiricist more liberally: for
what (sub-)domain is it observationally true? He retains the success question
of the empiricist. Finally, he will reformulate the third question: to what extent
is it the best (and hence the most widely applicable) derivation instrument?
The method of HD-evaluation will turn out, in this and the following chapter,
to be a direct way to answer the success question and, in later chapters, an
indirect way to answer the truth approximation question, in both cases for all
four epistemological positions. The relevant chapters will primarily be presented
in relatively neutral terminology, with specific remarks relating to the various
positions. In this chapter, the success question will be presented in terms of
successes and counter-examples 3 : what are the potential successes and counterexamples of the theory?
In sum, two, related, ways of applying the HD-method to theories can be
distinguished. The first one is HD-testing, which aims to answer the truth
question. However, as soon as the theory is falsified, the realist of a falsificationist nature, i.e., advocating exclusively the method of HD-testing, sees this as
a disqualification of any prima facie explanatory success. The reason is that
genuine explanation is supposed to presuppose the truth of the theory. Hence,
from the realist-falsificationist point of view a falsified theory has to be given
up and one has to look for a new one.
However, the second method to be distinguished, HD-evaluation, remains to
take falsified theories seriously. It aims at answering the success question, the
evaluation of a theory in terms of its successes and counter-examples (problems)
(Laudan 1977). For the (non-falsificationist) realist, successes are explanatory
successes and, when evaluating a theory, they are counted as such, even if
the theory is known to be false. It is important to note that the term
'(HD-)evaluation' refers to the evaluation in terms of successes and counterexamples, and not in terms of truth approximation, despite the fact that the
method of HD-evaluation will nevertheless turn out to be functional for truth
approximation. Hence, the method of HD-evaluation can be used meaningfully
without any explicit interest in truth approximation and without even any
substantial commitment to a particular epistemological position stronger than
instrumen tali sm. 4

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

95

We have already seen, in Chapter 2, that the HD-method can be used for
testing, but what is the relation of the HD-method to evaluation? Recall that,
roughly speaking, the HD-method prescribes to derive test implications and
to test them. In Section 5.1., it is shown that a decomposition of the HD-method
applied to theories naturally leads to an explication of the method of separate
HD-evaluation, using HD-testing, even in terms of three models. Among other
things, it will turn out that HD-evaluation is effective and efficient in answering
the success question.
In Section 5.2., so-called falsifying general facts will first be analyzed. Then
the decomposition of the HD-method will be adapted for statistical test implications. Finally, it is shown that the decomposition suggests a systematic presentation of the different factors that complicate the straightforward application of
the HD-methods of testing and evaluation.
In the next chapter we will use the separate HD-evaluation of theories for
the comparative HD-evaluation of them. Strictly speaking, only Section 5.1. of
this chapter is required for that purpose.
5.1. HD-EVALUATION OF A THEORY

Introduction
The core of the HD-method for the evaluation of theories amounts to deriving
from the theory in question, say X, General Test Implication (GTI's) and
subsequently (HD-)testing them. For every GTI I holds that testing leads
sooner or later either to a counter-example of I, and hence a counter-example
of X, or to the (revocable) acceptance of I: a success of X . A counter-example
implies, of course, the falsification of I and X. A success minimally means a
'derivational success'; it depends on the circumstances whether it is a predictive
success and it depends on one's epistemological beliefs whether or not one
speaks of an explanatory success.
However this may be, from the point of view of evaluation falsification is,
although an interesting fact, no reason to stop the evaluation of the theory.
One will derive and test new test implications. The result of such a systematic
application of the HD-method is a (time relative) evaluation report of X,
consisting of registered counter-examples and successes.
Now, it turns out to be very clarifying to write out in detail what is implicitly
well-known from the work of Hempel and Popper, viz., that the HD-method
applied to theories is essentially a stratified, two-step method, based on a
macro- and a micro-argument, with much room for complications. In the
already indicated macro-step, one derives GTI's from the theory. In their turn,
such GTI's are tested by deriving from them, in the micro-step, with the help
of suitable initial conditions, testable individual statements, called Individual
Test Implications (ITI's).
In this section we will deal with the macro-argument, the micro-argument,

96

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

and their combination into three models of (separate) evaluation. In the second
section special attention will be paid to so-called falsifying general hypotheses,
to statistical test implications, and to complications of testing and evaluation.

5.1.1. The macro H D-argument


As indicated, a theory X is tested and evaluated by deriving General Test
Implications (GTI's) from it, and testing these separately. Each GTI deals with
a certain domain specified in observation terms. A GTI is general in the sense
that the domain is not related to a specific (object or) system and place and
time. In at least one of these respects, the domain is supposed to be general,
but not necessarily in the universal sense. That is, the domain may be restricted
within some boundaries, e.g., to all systems of a certain type, to all places in a
region, to all times in an interval. s Moreover, within these boundaries, it may
concern all actual cases or it may concern all possibilities in reality, all nomic
possibilities. If the GTI is true, one speaks in the first case about an accidental
general fact, and in the second case about a lawlike or nomic general fact, or
simply a law.

Example: A famous example of a GTI is Dalton's derivation of what


became known as the law of multiple proportions from his theory of
the atom. The internal and bridge principles of Dalton's theory were
the following, in abbreviated form: IPl: atoms are indivisible, unchangeable, hence undestroyable, small material particles. IP2: atoms are
grouped into molecules, and they may regroup into other molecules.
BPl: pure substances consist of one type of molecules. BP2: chemical
reactions come down to systematic regrouping of the molecules of a
substance.
The suggested GTI, i.e., the law of multiple proportions, says that
when two different elements unite into two different compounds, the
different proportions bear a simple numerical relation to one another.
Note that Proust's law of definite proportions, stating that compounds
always decompose into components with constant weight ratios, can
also be reconstructed as a GTI of Dalton's theory. However, in fact this
law was not only one of Dalton's starting points for his theory, but, as
we will see, is also useful in testing the multiple proportions GTI.
A GTI is assumed to be testable in principle and in practice. It is testable in
principle if it is formulated in observation terms. To be testable in practice too,
several specific conditions, depending on the context, will have to be satisfied.
A GTI is an implication or conditional statement in two senses. First, in the
sense that it claims that for all cases in its domain, satisfying certain Initial
Conditions (IC), a certain other individual fact applies. The conditionally
claimed fact can be, like the initial conditions, of a (simple or compound)
deterministic or statistical nature.

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

97

In sum, a GTI is formally of the form :


I: for all x in D [if C(x) then F(x)]

that is, for all x in the domain D, satisfying the initial conditions C(x), the fact
F(x) is 'predicted'.

The second sense in which a GTI is an implication directly relates to the


HD-method. It is assumed to be an implication of some theory, i.c., X, i.e., to
follow from this X (and auxiliary hypotheses, see below) by some logicomathematical derivation. Scheme 5.1. represents this, what we call, macro
HD-argument. (LMC indicates the suggested Logico-Mathematical Claim and
MP indicates Modus Ponens.) Of course, it is a subtype of the MP-argument
scheme (Scheme 2.1.1.) given in Chapter 2. Though not indicated in the scheme,
it is assumed that X is necessary for the derivation, i.e., I is not a logicomathematically provable truth.
theory: X
LMC: if X then I
_ _ _ _ _ _ _ MP

GTI: I
Scheme 5.1. The macro H D-argument

5.1.2. Individual problems and general successes

Let us now concentrate on the results of testing GTI's. When testing a GTI of
a theory, we are interested in its truth-value, hence we use the test terminology.
Successive testing of a particular GTI I will lead to two mutually exclusive
results. The one possibility is that sooner or later we get falsification of I by
coming across a falsifying instance or counter-example of I, i.e., some Xo in D
such that C(xo) and not-F(xo), where the latter conjunction may be called a
falsifying combined (individual) fact.
Assuming that LMC is correct, a counter-example of I is, strictly speaking,
also a counter-example of X, falsifying X, for not only can not-I be derived
from the falsifying combined fact, but also not-X by Modus Tollens. Hence,
from the point of view of testing, it is plausible to speak also of falsification of
the theory. However, it will frequently be useful in this chapter, and perhaps
more in line with the evaluation terminology, to call the counter-example less
dramatically a negative instance, and further to speak of a negative (combined)
individual fact, or simply an individual problem of X.
The alternative possibility is that, despite variations in members of D and
ways in which C can be satisfied, all our attempts to falsify I fail, i.e., lead to
the predicted results. The conclusion attached to repeated success of I is of
course that I is established as true, i.e., as a general (reproducible) fact.
Now one usually calls the acceptance of I as true at the same time a
confirmation or corroboration of X , and the realist will want to add that I has

98

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

been explained by X and hence that it is an explanatory success of X. However,


this terminology is laden with the assumption that X has not yet been falsified,
via another general test implication, or even that it is accepted as true. To
block these connotations, it will be useful for evaluative purposes to call I a
positive general fact or simply a (general) success of X.
It may well be that certain GTI's of X have already been tested long before
X was taken into consideration. The corresponding individual problems and
general successes have to be included in the evaluation report of X (see below).
Registered problems and successes are (partial) answers to the success question: what are the potential successes and problems of the theory? Hence,
testing GTI's derived according to the macro HD-argument is effective in
answering this question. Moreover, it is efficient, for it will never lead to
irrelevant, neutral results. Neutral results only come into the picture when we
take the comparative evaluation of two or more theories into consideration
(see the next chapter). In sum, evaluating a theory along the lines of the macro
HD-argument is effective and efficient for answering the success question,
whatever the epistemological beliefs of the scientist.
We call the list of partial answers to the success question, which are available
at a certain moment t, the evaluation report of X at t, consisting of the following
two components:

the set of individual problems,


i.e., established counter-examples of GTI's of X,
the set of general successes,
i.e., the established GTI's of X, that is, general facts derivable from X.
Hence, the goal of separate theory evaluation can be explicated as aiming at
such an evaluation report.
Note that the two components of the report concern matters of a different
nature. An individual problem specifies (a combined statement concerning) an
individual item of the domain. A general success is a general statement considered to be true.
Note also that when we derive a general test implication from a theory, this
is essentially a prediction of a certain general fact. If this predicted general fact
turns out to be true, this general fact is called a (general) predictive success. If
the general fact was established before it was derived from the theory, it is
sometimes called a postdictive success.
Of course, the realist will call a predictive as well as a postdictive (individual
or general) success also an explanatory success and refer to the famous symmetry between explanation and prediction. 6 The only difference between explanation and prediction concerns the question of whether the relevant fact was
established as a fact before or after the derivation. This allows the realist to
interchange explanation and prediction, depending merely on whether the
supposed fact has, or has not yet, been established as such.

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

99

5.1.3. The micro HD-argument

Let us have a closer look at the testing of a general test implication, the microstep of the HD-method. To study the testing of GTI's in detail, it is plausible
to widen the perspective to the evaluative point of view on GTI's and to neglect
their derivability from X.
Let us call a statement satisfying all conditions for being a GTI, except its
derivability from some given theory, a General Testable Conditional (GTC).
Let G be such a statement, which of course remains of the same form as a
GTI:
G: for all x in D [if C(x) then F(x)]
To evaluate G we derive from G for some Xo in D and suitable initial conditions
(IC), viz., C(xo), the predicted individual fact or individual prediction F(xo),
i.e., an Individual Test Implication (ITI). It is an individual prediction in the
sense that it concerns a specific statement about an individual item in the
domain, as do the relevant initial conditions. It is a prediction in the sense that
the outcome is assumed not to be known beforehand. Hence, talking about a
prediction does not imply that the fact itself should occur later than the
prediction, only that the establishment of the fact has to occur later (leaving
room for so-called retrodictions).
What is predicted, i.e., F(xo), is usually called an (individual) effect or event.
Both are, in general, misleading. It may concern a retrodiction (even relative
to the initial conditions), and hence it may be a cause. And it may concern a
state of affairs instead of an event. For these reasons we have chosen the neutral
term (predicted) individual 'fact'.
In Scheme 5.2. the micro-reasoning of the HD-method, the micro
HD-argument, is represented, where UI indicates Universal Instantiation.
G: for all x in D [if C(x) then F(x)]
Xo in D
_______________________ UI

if C(x o ) then F(xo)


Ie: C(xo)

_______________________ MP
ITI: F(x o)
Scheme 5.2. The micro H D-argument

We suggest speaking in general of strictly individual applications of the micro


HD-argument when the items concern a specific object, time and place.
Example: Proust's law of definite proportions is in fact a GTe. Assuming
that we have determined the weight ratio of two elements in one particular sample of a given compound (initial condition), we can derive the

100

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

prediction (ITI) that the weight ratio in another particular sample is


the same.
With respect to Dalton's proper GTI of his theory of the atom, that
is, the one concerning multiple proportions, individual experiments can
be reconstructed as starting from a given weight proportion of the
elements in a sample of one of the compounds and predicting a 'simple'
multiple, or inverse multiple, proportion in a sample of the other. In
this case the testing of the multiple proportions hypothesis is strictly
individual.
It is clear that in the present example, knowledge of relevant cases of
Proust's law of definite proportions, one of Dalton's starting points, is
sufficient to check whether there is a mUltiple proportion relation
between the weight ratios. More precisely, given the definite proportion
of the elements in one compound (initial condition), the GTI predicts
the ITI that the definite proportion in the other bears a simple relation
to the former. In this case Ie and ITI are not related to a specific time
and place, and hence this way of testing the multiple proportions hypothesis is not strictly individual.
5.1.4. Individual problems and individual successes

The specific prediction posed by the individual test implication can come true
or proven to be false. If the specific prediction turns out to be false, then,
assuming that the initial conditions were indeed satisfied, the hypothesis G has
been falsified. The combined individual fact "C(xo) & not-F(xo)", "Co & notFo" for short, may be called a falsifying individual fact and Xo a falsifying
instance or counter-example of G. It will again be useful for evaluative purposes
to speak of a negative instance, and a negative (combined) individual fact or
simply an individual problem of G.
If the specific prediction posed by the individual test implication turns out
to be true, we get the combined individual fact "Co & F0'" which is not only
compatible with (the truth of) G, but is, not in full, but partially derivable from
G in the following sense (implying partial entailment by G, in the sense of
Section 3.1.). One of its conjuncts can be derived from G, given the other.
Again, one may be inclined to talk about confirmation 7 or even about explanation. However, given that we do not want to exclude that G has already been
falsified, we prefer again the neutral (evaluation) terminology: Xo is called a
positive instance and the combined individual fact "Co & F 0" a positive (combined) individual fact or simply an individual success of G.
It is easy to check that the same story can be told about "not-Co & not-Fo"
for some Xo in D, by replacing the role of "Co" by that of "not-Fo", the new
initial conditions, and the role of "Fo" by that of "not-Co", the new individual
test implication. The crucial point is that "if Co then F 0" is logically equivalent
to "if not-Fo then not-Co". By consequence, Xo is a positive instance satisfying
"not-Co & not-Fo", being a positive individual fact or an individual success of G.

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

101

The remaining question concerns how to evaluate the fourth and last combined individual fact "not-Co & Fo" concerning some Xo in D. Of course, this
fact is compatible with G, but none of the components is derivable from G and
the other component. Hence, the fourth combined fact cannot be partially
derived from G. Or, to put it differently, none of its components, taken as
initial condition, can lead to a negative instance, whereas this is the case for
(precisely) one of the components in the two cases of partially derivable facts.
Hence the terms neutral instance and neutral (combined) individual fact or
neutral result are the proper qualifications.
Consequently, the evaluation report of GTe's has, like the evaluation reports
of theories, two sides; one for problems and the other for successes. Again, they
form partial answers to the success question now raised by the GTe. However,
here the two sides list entities of the same kind: negative or positive instances
or individual facts, that is, individual problems and individual successes,
respectively.
It is again clear that the micro HD-argument for a GTC G is effective and
efficient for making its evaluation report: each test of G either leads to a positive
instance, and hence to an increase of G's individual successes, or it leads to a
negative instance, and hence to an increase of G's individual problems. It does
not result in neutral instances. Note that it is crucial for this analysis that
GTe's have, by definition, a conditional character.
What we have described above is the micro HD-argument for evaluating a
GTC. When we restrict attention to establishing its truth-value, and hence stop
with the first counter-example, it is the (micro) HD-argument for testing the
GTC.

5.1.5. Models of H D-evaluation and H D-testing of a theory


Applications of the macro- or micro-argument, or a combination of both, will
be called applications of the HD-method. In this subsection, we construct three
interesting models of (separate) HD-evaluation of a theory.
Concatenation of the macro and micro HD-argument gives the full argument
for theory evaluation in terms of individual combined facts: individual initial
conditions and individual conclusions, leading to individual problems and
individual successes.
Instead of the two-step concatenated account, theory evaluation can also be
presented completely in terms of contracted HD-evaluation, without the intermediate GTI's, directly leading to individual problems and individual successes.
In the contracted argument, the premises are the theory X and the initial
conditions C(xo), and the conclusion, i.e., IT!, is F(x o). The transition is based
on logico-mathematical argumentation. It is easy to check that it is always
possible to (re)introduce the intermediate level of a GTI.
Any application of the HD-method (concatenated or contracted) leading to
an evaluation report with individual problems and individual successes will be
called an application of the micro-model of HD-evaluation. It is clear that

102

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

application of the micro-model is possible for all kinds of general hypotheses,


from GTC's to theories with proper theoretical terms.
However, as far as theories which are not just GTC's, are concerned, the
macro-step also suggests the model of asymmetric HD-evaluation of a theory,
leading to an evaluation report with individual problems and general successes.
In that case, GTI's are derived in the macro-step, and only tested, not evaluated,
in the micro-step. Hence, individual successes not fitting into an already established general derivable fact, are not registered as successes. Note that it directly
follows from the contracted argument that an individual problem can always
be reconstructed as a counter-example of a GTI of the theory.
In the micro-model of HD-evaluation of theories, in particular when contraction is used, the intermediate general successes of theories, i.e., positive
general facts, may disappear from the picture. However, in scientific practice,
these intermediate results frequently play an important role. The individual
successes of theories are summarized, as far as possible, in positive general
facts. These general successes relativize the dramatic role of falsification via
other general test implications. As we will see in the next chapter, they form a
natural unit of merit for theory comparison, together with conflicting individual
facts, as the unit of problems.
In later chapters, the model of asymmetric HD-evaluation wiII playa dominant role. The results it reports will then usually be called counter-examples
and (general) successes.
In the next section, we study the possibility that individual problems can be
summarized in general problems, that is, in negative general facts. Hence, there
is also room for a macro-model of HD-evaluation, where the evaluation report
lists, besides general successes, general problems as well. In this case, all individual successes and individual problems are left out of the picture, as long as
they do not fit into an established general fact derivable from or in conflict
with the theory.
Note that there is also the possibility of a fourth model of HD-evaluation
of an asymmetric nature, with individual successes and general problems, but
as far as we can see, it does not playa role in scientific practice.
The three interesting models of HD-evaluation of theories can be ordered
by increasing refinement: the macro-model, the asymmetric model and the
micro-model. Of course, in models where all individual problems are taken
into consideration, they mayor may not be summarized as far as possible in
general problems. The same holds for individual successes.
Finally, it is plausible to characterize models of H D-testing of theories as the
application of one of the three models up to the point of registering the first
individual or general problem, that is, a counter-example or a falsifying general
fact, respectively. It is clear that the corresponding test reports now provide
partial answers, not only to the truth question, but also to the success question,
up to falsification. For the realist this implies, among other things, a restriction
to successes that may be genuine explanatory merits. Moreover, it will be clear

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

103

that the models of HD-testing produce the test reports in an effective and
efficient way for the same reasons as the models of HD-evaluation do:
HD-testing leads to successes or problems, and not to neutral results. As
already suggested, the exclusive interest in HD-testing of theories will be called
the (naive) falsificationisf perspective or method.
Table 5.1. Methodological categories of theory evaluation and testing
Problems

Successes

Individual

individual problem
negative instance
negative individual fact
counter-example

individual success
positive instance
positive individual fact

General

general problem
negative general fact
falsifying general fact

general success
positive general fact

Table 5.1. summarizes the four methodologically relevant categories and their
terminological variants. It is easy to read off the four possible models of
HD-evaluation and HD-testing.
5.2. FALSIFYING GENERAL HYPOTHESES, STATISTICAL TEST
IMPLICATIONS, AND COMPLICATING FACTORS
Introduction

In this section we will first deal with so-called falsifying general hypotheses,
that is, general problems, summarizing individual problems, i.e., counter-examples. Then we will show that the main lines of the analysis of testing and
evaluation also apply when the test implications are of a statistical nature.
Finally, we will deal with all kinds of complications of testing and evaluation,
giving occasion to dogmatic strategies and suggesting a refined scheme of
HD-argumentation.
5.2.1. Falsifying general hypotheses and general problems
In this subsection we pay attention to a methodological issue that plays an
important role in the methodology of Popper and others: so-called falsifying
general hypotheses. In contrast to their dramatic role in HD-testing of theories,
in HD-evaluation they play the more modest role of a general problem which
summarizes individual problems on the minus side of the evaluation report of
a theory.
Let us return to the evaluation or testing of a general testable conditional
(GTC) G. Finding a partially implied instance of G does, of course, not imply
that G is true. Repetitions of the individual tests, varying the different ways in
which the initial conditions can be realized, are necessary to make it plausible

104

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

that G is true. To conclude that it is true implies making an inductive jump,


at least in the case that G has some infinite aspect. But as soon as one makes
this jump, one has established a new general fact, i.c., G. If G is in fact a GTI
of X, it becomes a positive general fact of X, i.e., a general success of X, dealt
with before.
If we are only interested in the truth-value of G, its conclusive falsification
stops the relevance of its further testing. From this test perspective, repeating
similar individual tests a number of times, after one falsification, only serves
the purpose of making sure that the initial conditions and the falsifying outcome
have been realized properly at least once.
However, further evaluation of a conclusively falsified GTC makes sense for
several other reasons, in particular, when it is a test implication of some theory.
The results of repeated tests may suggest an interesting division of counterexamples and individual successes of G, and hence one or more alternative
general hypotheses. This may lead to the establishment of other general facts
which are also relevant for theory comparison. In particular, it may turn out
that an alternative GTC, with the same domain:
G*: for all x in D [if C*(x) then F*(x)]

can be established, with C* implying C and F* implying not-F, such that each
partially derivable individual fact of G* of type "C*&F*" is a falsifying instance
of G. In that case, G* may be called a falsifying or negative general fact for G.
When G is a GTI of X, G* may also be called a lower level falsifying (general)
hypothesis, to use Popper's phrase, now contradicting not only G but also X.
However, more in line with the terminology of evaluation, we call it a negative
general fact or general problem of X.
Example: An example is the law of combining volumes (Gay-Lussac).
Apart from by Dalton himself, it was generally considered as a problem
for Dalton's version of the atomic theory. It can easily be reconstructed
as a general problem in the technical sense defined above. As is wellknown, Avogadro turned the tables on the atomic theory by a fundamental change in order to cope with this problem.

A technical alternative to the structure of a falsifying general fact which was


described above is the following. A G* may have been established for which
not-C* implies C and not-F* implies not-F. Each partially derivable individual
fact of this G* of type "not-C* & not-F*" is a falsifying instance of G. Hence,
this G* would also be a falsifying general fact for G, and a falsifying general
fact, or simply a general problem for X, if G is a GTI of X.
Note that in both described cases G* and G form a contradiction as soon
as C*, respectively not-F*, can be satisfied in the domain.
Of course, when positive and negative instances of G occur, repeated tests
may lead not only to negative general facts, but also to the establishment of

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

105

weaker variants of G. In particular, consider the GTe


G': for all x in D [if C(x) then F'(x)]

with C implying C and F implying F', then G implies G' and every negative
instance of G' is a negative instance of G. When G' becomes established one
may call it a general fact derivable from G, and hence a general success of X
if G is a GTI of X.
As soon as, and as long as, all negative individual facts of a theory can be
summarized in negative general facts, the individual problems in the evaluation
report can be replaced by the corresponding general problems. In this way we
get on both sides of the record the same kind of conceptual entities, viz., general
facts, forming the ingredients of the macro-model of HD-evaluation of theories.
Hence, the foregoing exposition concerning general facts amounts to an additional illustration of the fact that general hypotheses and the like can play an
important intermediate role in the evaluation of theories. Skipping the intermediate notions of general successes and general problems would hide the fact
that they make the evaluation of theories very efficient, in theory and practice.
Instead of confronting theories with all previously or subsequently established
combined individual facts, it is possible to restrict the confrontation as much
as possible to a confrontation with old or new summarizing general facts.
5.2.2. Statistical test implications

The presentation thus far may have suggested that the analysis only applies to
hypotheses and their test implications as far as they are of a deterministic and
non-comparative nature. In this subsection we will present the adapted main
lines for statistical test implications, first of a non-comparative and then of a
comparative nature. In the literature concerning statistical testing, one can find
all kinds of variants and details.8 Here it is only necessary to show that
statistical general and individual test implications can essentially be tested in
a similar way to non-statistical ones.
In the non-comparative form, a typical case is that the theory, e.g., Mendel's
theory entails a probabilistic GTI of the following abstract form: in domain D
the probability of feature F on the condition C satisfies a certain probability
distribution p (e.g., binomial or normal). The sample version, the proper
Statistical GTI, and the corresponding Ie and ITI are then respectively of the
following form:
GTI: for all rx (0 < rx < 1) and for all random samples s of (sufficiently
large, determined by the distribution) size n of individuals from domain
D satisfying condition C there are a and b (0 < a < b < 1) such that the
probability that the ratio of individuals satisfying F is in the region [0, a]
does not exceed rx/ 2 and similar for the region [b, 1

Ie: sample s is a random sample of (sufficiently large) size n of individuals

106

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

from D satisfying C
ITI: the probability that the ratio in s of individuals satisfying F is in
R(a, n) = defeO, a] U [b, I] does not exceed :x

If a is small enough, according to our taste, we may, by a final non-deductive


jump, hypothesize the non-probabilistic ITI:

ITInp:

the ratio in s of individuals satisfying F will be outside R(a, n)

In classical statistics a test of this kind is called a significance test with significance level a (standard values of IX are 0,005 and 0,001) and critical region
R(a, n). Moreover, the abstract GTI is called the null hypothesis and the classical
decision rule prescribes to reject it when ITInp turns out to be false and not to
reject it when ITInp comes true. However, from our perspective it is plausible
to categorize the first test result merely as a negative sample result or a
(statistical) 'counter-sample' and the second as a positive one. Of course, strictly
speaking, a counter-sample does not falsify the GTI, let alone the theory.
Moreover, from repeated positive results one may inductively jump to the
conclusion of a general statistical success GTI.
Statistical test implications are frequently of a comparative nature. This
holds in particular when they derive from causal hypotheses. The reason is
that such hypotheses are essentially double hypotheses, one for the case that
the relevant (supposedly) causal factor is present and another for the case that
it is absent. Typical examples are the causal hypotheses governing drug testing.
An adapted standard significance test for a supposedly normally distributed
feature F on condition C in domain D1 and D2, with the same variance,
focusses on the (null) hypothesis that their expectation values are the same. It
is called the 't-test or Student-test' (after W.S. Gossett who wrote under the
name 'Student') and the resulting proper Statistical GTI, and corresponding
IC, ITI and ITInp are now respectively of the following form:
GTI(comp): for all a (0 < IX < 1) and for all sufficiently large (random)
samples sl and s2 of size nl and n2 of individuals from domain Dl and
D2, respectively, satisfying condition C there are a and b (0 < a < b < 1)
such that the probability that a certain function (the t-statistic) of the
difference between the respective ratios of individuals satisfying F is in
the region [0, a] does not exceed al2 and similar for the region [b, 1]10
IC(comp): sample s1 and s2 are random samples of (sufficiently large)
sizes n1 and n2 of individuals from Dl and D2, respectively, satisfying C
ITI(comp): the probability that the value of the t-statistic is in R(a, n) =
[0, aJ U [b, 1] does not exceed a

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

107

Again, if rx is small enough, according to our taste, we may, by a final nondeductive jump, hypothesize the non-probabilistic ITI:
ITInp(comp):

the value of the t-statistic will be outside R(rx, n)

This can again be described in classical terms of significance level and rejection
conditions. However, in our perspective it is again more plausible to use the
more cautious terminology of positive and negative test results for the null
hypothesis, which are, of course, negative and positive test results for the
primary (e.g., causal) hypothesis at stake.
5.2.3. Factors complicating HD-testing and -evaluation

According to the idealized versions of HD-testing and -evaluation presented


so far there are only cases of evident success or failure. However, as is wellknown, several factors complicate the application of the HD-method. Let us
approach them first from the falsificationist perspective. Given the fact that
scientists usually believe that their theory is (approximately) true, they have,
on the basis of these factors, developed strategies to avoid the conclusion of
falsification. The important point of these dogmatic or conservative strategies
is that they may rightly save the theory from falsification, because the relevant
factor may really be the cause of the seeming falsification. Although the recognition of a problem for a theory is more dramatic from the falsificationist
perspective, when evaluating a theory one may also have good reasons for
trying to avoid a problem.
We distinguish five complicating factors, each leading to a standard saving
strategy. They show in detail, among other things, how Lakatos' methodology
of research programs (Lakatos 1970, 1978), saving the hard core, can be
defended and effected. Though perhaps less frequently, the same factors may
also be used, rightly or wrongly, as point of impact for contesting some success.
In this case, there is one additional factor. All six factors concern suppositions
in the concatenated macro and micro HD-argument. We do not claim originality with these factors as such; most of them have been mentioned by Lakatos
and have been anticipated by Hempel, Popper and others. However, the
following systematic survey and localization of them is made possible by the
decomposition of the macro and micro HD-argument. It is left to the reader
to identify examples of the factors, e.g., in the history of the theory of the atom.
Dogmatic strategies

(1) The derivation of a general test implication from the theory is usually
impossible without invoking explicitly or implicitly one or more auxiliary
hypotheses, i.e., hypotheses which do not form a substantial part of the theory
under test, but are nevertheless required to derive the test implication. An
important type of such auxiliary hypotheses are specification hypotheses, that

108

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

is, hypotheses that specify particular constraints for the values of certain (theoretical or non-theoretical) quantities in the specific kind of cases concerned.
Hence, to avoid falsification, one can challenge an auxiliary hypothesis. A
particular problem of auxiliary hypotheses may be that they are too idealized;
they may need concretization.
(2) The derivation ofthe general test implication presupposes that the logicomathematical claim can convincingly be proven. One may challenge this. A
successful challenge may, of course, lead to tracing implicit auxiliary hypotheses,
i.e., new instances of the first factor, which can be questioned as indicated
under the first point.
(3) Test implications have to be formulated in observation terms. However,
at present, there is almost general agreement that pure observation terms do
not exist. All observation terms are laden with hypotheses and theories. New,
higher order, observation terms are frequently defined on the basis of more
elementary ones and certain additional presuppositions, which provide e.g., the
relevant existence and uniqueness conditions. Such presuppositions form the
bridge between more and less theory-laden observation terms. These presuppositions belong to the so-called background knowledge; they form the underlying
hypotheses and theories that are taken for granted. We call them the observation
presuppositions. The theory to be evaluated should itself not be an observation
presupposition, i.e., the observation terms relative to the theory should not be
laden with that theory itself, only with other ones. The relevant strategy now
is to challenge an observation presupposition. Ultimately, this may bring us
back to the observation language of a lay-person who is instructed by the
experimenter to 'materially realize' (Radder 1988) an experiment.
(4) A general test implication specifies initial (test) conditions. They have
actually to be realized in order to conclude that the individual test implication
must be (come) true. One may challenge the claim that these initial conditions
were actually fulfilled. One important reason for repeating an experiment a
number of times is to make it sufficiently sure that these conditions have at
least once been realized. But if another outcome than the one predicted systematically occurs, one may defend the idea that there are structural causes preventing the fulfillment of the intended initial conditions in the way one is trying to
install them.
(5) Whether an outcome is or is not in agreement with the predicted outcome
is usually not a straightforward matter. This is particularly the case when the
observation terms have vague boundaries or are quantitative, or when the
predicted effects and the observation data are of a statistical nature. In all such
cases, the question is whether the actual outcome is approximately equal to
the predicted outcome. To decide this we need a (previously chosen) decision
criterion. In the case of statistical individual test implications, several statistical
decision criteria have been standardized. Although a statistical decision criterion also concerns an approximation question, we propose to reserve the term
approximation decision criterion for the case of vague or quantitative observation concepts. In general, so the strategy goes, a decision criterion mayor may

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

109

not be adequate in a particular case, i.e., it mayor may not lead to the correct
decision in that case.
(6) Finally, in order to conclude that the theory has acquired a new general
success, the relevant general test implication has first to be established on the
basis of repeated tests. As a rule, this requires an inductive jump or inductive
generalization, which may, of course, always be contested as unjustified. Note
that a similar step is involved in the establishment of a general problem, but
we neglect it here.
To be sure, the localization of these factors need not always be as unambiguous as suggested by the previous exposition, although we claim to have identified the main occurrences. However this may be, the consequence of the first
five factors (auxiliary hypotheses, logico-mathematical claims, observation presuppositions, initial conditions, and decision criteria) is that a negative outcome
of a test of a theory only points unambiguously in the direction of falsification
when it may be assumed that the auxiliary hypotheses and the observation
presuppositions are (approximately) true, that the logico-mathematical claim
is valid, that the initial conditions were indeed realized and that the used
decision criteria were adequate in the particular case. Hence, a beloved theory
can be protected from threatening falsification by challenging one or more of
these suppositions.

Refined H D-scheme
In the subjoined, refined schematization of the concatenated HD-test arguments
(Scheme 5.3.) the five plus one vulnerable factors or weak spots in the argument
have been made explicit and emphasized by question marks. As in the foregoing
Theory under test: X
Auxiliary Hypotheses: A ?1?
LMC: if X and A then GTI ?2?
____________________________ MP
GTI: for all x in D [if C(x) then F(x)]
Observation presuppositions: CjC*, Fj F* ?3?

__________________________ SEE
GTI*: for all x in D [if C*(x) then F*(x)
Xo in D
Initial conditions: C*(xo) ?4?
___________________ VI

+ MP

ITI*: F*(x o)
Data from repeated tests
_________________ Decision Criteria ?5?
either sooner or later
a counter-example
of GTI*, leading to the
conclusion not-GTI*

or only positive

instances of GTI*,
suggestive interference of GTI*
by Inductive Generalization ?6?

Scheme 5.3. The refined macro + micro H D-argument

110

SEPARATE EVALUATION OF THEORIES BY THE HD-METHOD

exposition, the indication in the scheme of the different types of weak spots is
restricted to their main occurrences in the argument. SEE indicates Substitution
of presupposed Empirically Equivalent terms (in particular, C by C* and F by
F*). We contract the application of Universal Instantation (UI) followed by
Modus Ponens (MP).
Neglecting the complicating factors, the left-hand case in Scheme 5.3. at the
bottom results in falsification of GTI * and hence of GTI, leading to falsification
of X. Under the same idealizations, the right-hand case results first in (the
implied part ITJ* of) individual successes of GTJ* (and indirectly of GTJ),
and then, after the suggested inductive generalization, in the general success
GTJ* (and hence GTI) of X.

Concluding remarks
If the truth question regarding a certain theory is the guiding question, most
results of this chapter, e.g., the decomposition of the HD-method, the evaluation
report and the survey of complications, are only interesting as long as the
theory has not been falsified. However, if one is also, or primarily, interested
in the success question the results remain interesting after falsification. In the
next chapter we will show how this kind of separate HD-evaluation can be
put to work in comparing the success of theories. Amongst others, this will
explain and even justify non-falsificationist behavior, including certain kinds
of dogmatic behavior.

6
EMPIRICAL PROGRESS AND PSEUDOSCIENCE

Introduction
In this chapter we will extend the analysis of the previous chapter to the
comparison of theories, giving rise to a definition of empirical progress and a
sophisticated distinction between scientific and pseudoscientific behavior.
In Section 6.1. we will first describe the main line of theory comparison that
forms the natural extension of separate HD-evaluation to comparative
HD-evaluation. Moreover, we will introduce the rule of success for theory
selection suggested by comparative HD-evaluation, leading to an encompassing
evaluation methodology of instrumentalist flavor. This methodology can be
seen as the core method for the assessment of claims to empirical progress.
In Section 6.2. we will compare the evaluation methodology with the three
methods distinguished by Lakatos (1970, 1978): the naive and sophisticated
falsificationist method and the method of research programs, favored by
Lakatos. We will make it plausible that the evaluation methodology resembles
the sophisticated falsificationist methodology the most and that it may well be
more efficient for truth approximation than the naive falsificationist method.
In Section 6.3. we will argue that the, in some way dogmatic, method of
research programs may be a responsible way of truth approximation, as
opposed to pseudoscientific dogmatic behavior.

6.1. COMPARATIVE HD-EVALUATION OF THEORIES

Introduction

The presented analysis of separate HD-evaluation has important consequences


for theory comparison and theory selection. The momentary evaluation report
of a theory generated by the macro-step immediately suggests a plausible way
of comparing the success of theories, of further testing the comparative hypothesis that a more successful theory will remain more successful and, finally, the
rule of theory selection, prescribing to adopt it, for the time being, if it has so
far proven to be more successful.
The suggested comparison and rule of selection will be based on the asymmetric model of evaluation in terms of general successes and individual problems.
However, it will also be shown that the symmetric approach, in terms of either

111

112

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

individual or general successes and problems, leads to an illuminating symmetric evaluation matrix, with corresponding rules of selection.

6.1.1. Theory comparison


A central question for methodology is what makes a new theory better than
an old one. The intuitive answer for the new theory being as good as the old
is plausible enough. The new theory has at least to save the established strengths
of the old one and not to add new weaknesses on the basis of the former tests.
In principle, we can choose any combination of individual or general successes
and problems to measure strengths and weaknesses. However, the combination
of general successes and individual problems, i.e., the two results of the asymmetric model of (separate) HD-evaluation, is the most attractive. First, this
combination seems the closest to actual practice and, second, it turns out to
be the most suitable one for a direct link with questions of truth approximation.
For these reasons we will first deal with this alternative and come back to the
two symmetric alternatives in a separate subsection (6.1.4.).
Given the present choice, the following definition is the obvious formal
interpretation of the idea of (prima facie) progress, i.e., increasing success:
Theory Y is (at time t) at least as successful as (more successful than or
better than) theory X iff (at t)
- the set of individual problems of Y forms a subset of that of X
- the set of general successes of X forms a subset of that of Y
(- in at least one case the relevant subset is a proper subset)1.2
The definition presupposes, of course, that for every registered (individual)
problem of one theory, it has been checked whether or not it is also a problem
for the other, and similarly for (general) successes. In later chapters the first
clause might be called the 'internal' clause and the second the 'external' clause,
for reasons that will become clear in the next chapter. However, the name
'instantial' clause for the first one is more appealing and relatively neutral.
From the realist perspective it is plausible to call the second clause the 'explanatory' clause. From other epistemological perspectives one may choose another,
perhaps more neutral name, such as, the general success clause.
It is also obvious how one should define, in similar terms to those above,
the general notion of 'the most successful theory thus far among the available
alternatives' or, simply, 'the best (available) theory'.
It should be stressed that the diagnosis that Y is more successful than X
does not guarantee that this remains the case. It is a prima facie diagnosis
based only on thus far established facts, and new evidence may change the
comparative judgment. But, assuming that established facts are not called into
question, it is easy to check that it cannot become the other way around, i.e.,
that X becomes more successful than Y in the light of old and new evidence.

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

113

For, whatever happens, X has extra individual problems or Y has extra general
successes.
It should be conceded that it will frequently not be possible to establish the
comparative claim, let alone that one theory is more successful than all its
available alternatives. The reason is that these definitions do not guarantee a
constant linear ordering, but only an evidence-dependent partial ordering of
the relevant theories. Of course, one may interpret this as a challenge for
refinements, e.g., by introducing different concepts of 'relatively maximal' successful theories or by a quantitative approach.

6.1.2. Testing the comparative success hypothesis


Be this as it may, we have defined enough notions to introduce our explication
of the core of HD-evaluation in the following heuristic principles. The first
principle will be self-evident: as long as there is no best theory, one may
continue the separate HD-evaluation of all available theories in order to explore
the domain further in terms of general facts to be accounted for and individual
problems to be overcome by an overall better theory. We will concentrate on
the second principle, applicable in the case that one theory is more successful
than another one, and hence in the case that one theory is the best.
Suppose theory Y is at t more successful than theory X. This is not yet a
sufficient reason to prefer Y in some substantival sense. That would be a case
of 'instant rationality'. However, when Y is at a certain moment more successful
than X this does suggest the following comparative success hypothesis:
CSH:

Y (is and) will remain more successful than X

CSH is an interesting hypothesis, even if Y is already falsified. Given that Y is


known to be more successful than X at t, CSH amounts at t to two components,
one about problems, the other about successes:
CSH-P:
CSH-S:

all individual problems of Yare individual problems of X


all general successes of X are general successes of Y

where 'all' is to be read as 'all past and future'.


Although there may occasionally be restrictions of a fundamental or practical
nature, these two components will usually concern testable generalizations.
Hence, testing CSH requires application of the micro HD-argument. Following
CSH-P, we may derive a GTI from Y that does not follow from X, and test
it. When we get a counter-example of this GTI, and hence an individual
problem of Y, this mayor may not be an individual problem of X. If it is not,
we have falsified CSH-P.
Alternatively, following CSH-S, we may derive a GTI from X which cannot
be derived from Y, and test it. If it becomes accepted this means falsification
of CSH-S. Of course, in both cases, the opposite test result confirms the

114

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

corresponding comparative sub-hypothesis, and hence CSH, and hence


increases the registered success difference. In the following we will call (these
two ways of) testing CSH for obvious reasons comparative HD-evaluation.3

6.1.3. The rule of success


The plausible rule of theory selection is now the following:

Rule of success (RS)


When Y has so far proven to be more successful than X, i.e., when CSH
has been 'sufficiently confirmed' to be accepted as true, eliminate X in
favor of Y, at least for the time being
RS does not speak of 'remaining more successful', for that would imply the
presupposition that the CSH could be completely verified (when true). Hence
we speak of 'so far proven to be more successful' in the sense that CSH has
been 'sufficiently confirmed' to be accepted as true, that is, CSH is accepted as
a (twofold) inductive generalization. When CSH is 'sufficiently confirmed' may
be a matter of dispute. However this may be, the acceptance of CHS and
consequent application of RS is the core idea of empirical progress, a new
theory that is better than an old one. RS may even be considered as the (fallible)
criterion and hallmark of scientific rationality, acceptable for the empiricist as
well as for the realist. 4
As soon as CSH is (supposed to be) true, the relevance offurther comparative
HD-evaluation gets lost. Applying RS, i.e., selecting the more successful theory,
then means the following, whether or not it already has individual problems.
One may concentrate on the further separate HD-evaluation of the selected
theory, or one may concentrate on the attempt to invent new interesting
competitors, that is, competitors that are at least as successful as the selected
one.
Given the tension between increasing the domain of a theory and increasing
its (general observational) success, it is not an easy task to find such interesting
competitors. The search for such competitors cannot, of course, be guided by
prescriptive rules, like RS, but there certainly are heuristic principles of which
it is easy to see that they stimulate new applications of RS. Let us start by
explicitly stating the two suggested principles leading to RS. First, the principle
of separate H D-evaluation (PSE) "Aim via general test implications at establishing new laws which can be derived from your theory (general successes) or,
equivalently, aim at new negative instances (individual problems) of your
theory". Secondly, the principle of comparative HD-evaluation (PCE) "Aim at
HD-testing of the comparative success hypothesis, when that hypothesis has
not yet been convincingly falsified".5
As already suggested, RS presupposes previous application of PSE and PCE.
But also some additional heuristic principles, though not necessary, may promote the application of RS. To begin with, the principle of content (PC) "Aim

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

115

at success preserving, strengthening or, pace Popper, weakening of your theory".


A stronger theory is likely to introduce new individual problems but gain new
general successes. If the latter arise and the former do not materialize, RS can
be applied. Something similar applies to a weaker theory. It may solve problems
without losing successes. We would also like to mention the principle ofdialectics
(PD) for two theories that escape RS because of divided success "Aim at a
success preserving synthesis of two RS-escaping theories". In Section 8.3. we
will explicate a number of dialectical notions in this direction. Of course, there
may come a point at which further attempts to improve a theory and hence
to new applications of RS are given up.
In sum, the asymmetric model of HD-evaluation of theories naturally suggests the definition of 'more successful', the comparative success hypothesis,
testing it, i.e., comparative HD-evaluation, and the rule of success (RS) as the
cornerstone of empirical progress. Separate and comparative HD-evaluation
provide the right ingredients for applying first the definition of 'more successful'
and, after sufficient tests, that of RS, respectively. In short, separate and comparative HD-evaluation are functional for RS, and HD-testing evidently is functional for both types of HD-evaluation. The method of HD-evaluation of
theories combined with RS and the principles stimulating the application of
RS might well be called the instrumentalist methodology. In particular, it may
be seen as a free interpretation or explication of Laudan's problem solving
model (Laudan 1977) which is generally conceived as a paradigm specification
of the idea of an instrumentalist methodology. However, it will be called, more
neutrally, the evaluation methodology. It will be said that this methodology is
governed by RS. The claim is that this methodology governs the short-term
dynamics of science, more specifically, the internal and competitive development
of research programs.
Note that the evaluation methodology demonstrates continued interest in a
falsified theory. The reasons behind it are easy to conceive. First, it is perfectly
possible that the theory nevertheless passes other general test implications,
leading to the establishment of new general successes. Second, even new tests
leading to new individual problems are very useful, because they have to be
overcome by a new theory. Hence, at least as long as no better theory has
been invented, it remains useful to evaluate the old theory further in order to
reach a better understanding of its strengths and weaknesses.
6.1.4. Symmetric theory comparison

The symmetric models of separate HD-evaluation, i.e., the micro- and the
macro-model, suggest a somewhat different approach to theory comparison.
Although these approaches do not seem to be in use to the extent of the
asymmetric one, and can only indirectly be related to truth approximation,
they lead to a very illuminating (comparative) evaluation matrix.
Let us first examine in more detail precisely what we want to list in the three
types of evaluation reports corresponding to the three models. From the present

116

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

perspective of theory comparison, this is very important. A better theory has


to be at least as successful as the old one, and this suggests general conditions
of adequacy for the definitions of a success, of a problem and of a neutral
result. A success is typically something to be retained by a better theory, a
better theory is supposed not to introduce a new problem, and, finally, neutral
results should remain neutral or become successes. Note that the definition of
'at least as successful' in Subsection 6.1.1. just summarizes the first two of these
conditions in terms of individual problems and general successes, and leaves
neutral results out of the picture (see below). We will now see that the general
conditions of adequacy are very instructive for the specification of what precisely can be listed as successes and problems.
The notions of general successes and general problems are not problematic.
A better theory retains general successes as general test implications, and does
not give rise to new general test implications of which testing leads to the
establishment of new general problems. Furthermore, the notion of a neutral
general fact does not create problems. The notions of individual successes,
individual problems and neutral individual facts are also not problematic as
long as we list them in terms of positive, negative and neutral instances,
respectively. A better theory keeps the positive instances as such; it does not
lead to new negative instances, and neutral instances may remain neutral or
become positive. However, if we want to list individual successes and/or individual problems in terms of statements, the situation becomes more complicated.
That, and how, it is possible is presented below, but the reader might prefer to
go immediately to the evaluation matrix, which may be read in both ways.

Individual successes and problems in statement form


Recall that "Co & Fo" was called a 'positive (combined) individual fact' of the
GTe "if C then F" and hence of theory X if this GTe is a GTI of X. Recall
also that the conjunction "Co & F 0" was not derivable from G or X, but only
partially: Fo was derivable from G and hence of X, given Co. But do we want
to retain that F0 is, given Co, derivable from a supposedly better theory Y? As
long as we have not established this G, or another GTI for which "Co & Fo"
is also a positive individual fact, the answer is: no! For it may well be that
there is no better theory that can bridge the gap between Co and F o. How
then should we specify in statement terms what we want to retain? In the light
of the foregoing analysis, the plausible answer is the generalization of the idea
of partial derivability: F0 is a positive individual fact for X iff there are initial
conditions ct, not necessarily equivalent to Co, such that X and ct together
(and not separately) imply Fo. Of course, only the strongest positive individual
facts need to be listed and the whole story can be repeated for positive combined
individual facts of the form "not-F o & not-Co".
Let us now look at "Co & not-Fo", a 'negative (combined) individual fact' of
the GTe "if C then F" and hence of theory X if this GTe is a GTI of X. The
following definition is now suggested by the analysis of positive individual

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

117

facts. Not-Fo is a negative individual fact for X iff there are initial conditions
ct, not necessarily equivalent to Co, such that X and Ct together (and not
separately) imply F o. Note first that this definition automatically implies that
the relevant q is also a negative individual fact for X. Negative individual
facts typically come in pairs, and a new theory should not introduce such
new pairs.
What happens if a new theory is better in the sense that it loses some
individual facts as negative facts. Let not-Fo and Co be a pair of negative
individual facts for X, and let not-Fo be not a negative individual fact for Y.
This does not imply that Co is not a negative individual fact for Y, for it might
come into conflict with Y and some other individual fact than not-Fo. Hence,
although negative individual facts come in pairs, they do not need to lose that
status together with regard to some other theory.
That not-Fo is not a negative individual fact for Y also does not imply that
it is a positive individual fact for Y in the sense defined above. This suggests
the plausible definition of a neutral individual fact for Y: a fact which is neither
positive nor negative for Y. Note that if F 0 is a positive individual fact for Y,
due to Co, then this information alone suggests that Co is a neutral fact for Y.
But it may well be that other facts than F0 and/or more information about Ys
consequences lead to another status of Co.
The evaluation matrix

Let us now look more specifically at the symmetric micro-model, counting in


terms of individual problems, successes and neutral results, that is, negative,
positive and neutral instances or (statements of) individual facts. Hence, in
total, we get for two theories a matrix (Table 6.1.) of nine combinations of
possible instances or individual facts . In order to make the matrix useful for
the macro-model also, we present it in terms of facts. For the moment, these
facts are to be interpreted as individual facts. The entries give the status of a
fact with respect to the indicated theories.
Table 6.1. The (comparative) evaluation matrix

Negative
Neutral
Positive

Negative

X
Neutral

B4: 0
B8: +
B9: +

B2: B5:0
B7: +

Positive
Bl:-

B3: B6: 0

From the perspective of Y the boxes Bl / B2/ B3 represent unfavorable facts


(indicated by '- '), B4/ BS/ B6 (comparatively neutral or) indifferent facts (0),
and B7/ B8/B9 favorable facts (+). The numbering of the boxes, anticipating a
possible quantitative use, was determined by three considerations: increasing

118

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

number for increasingly favorable results for Y, symmetry with respect to the
\-diagonal, and increasing number for increasingly positive indifferent facts. 6
It is now highly plausible to define the idea that Y is more successful than
X in the light of the available facts as follows: there are no unfavorable facts
and there are some favorable facts, that is, Bl /2/ 3 should be empty, and at
least one of B7/8/9 non-empty. This immediately suggests adapted versions of
the comparative success hypothesis and the rule of success.
It is also clear that we obtain macro-versions of the matrix, the notion of
comparative success, the comparative success hypothesis and the rule of success
by replacing individual facts by general facts. A general fact may be a general
success, a general problem or a neutral general fact for a theory. Note that
combinations with individual and general facts are also possible. 7
In all these variants, the situation of being more successful will again be rare.
However, it is certainly not excluded. In Chapter 11 we will see, for instance,
that the theories of the atom developed by Rutherford, Bohr and Sommerfeld
can be ordered in terms of general facts according to the symmetric definition.
Another set of examples of this kind provides Table 6.2. (from: Panofsky and
Phillips 1962 2 , p. 282) below, representing the records in the face of 13 general
experimental facts of the special theory of relativity (STR) and six alternative
electrodynamic theories, viz., three versions of the ether theory and three
emission theories.
Table 6.2. Comparison of experimental record of seven electrodynamic theories
Light propagation experiments

>

!!

.!!

Ul

J:

'"

." .
.'"" ~.. ...." "" f.
.. fi '"
..8e -!"

.. ,g ..'"
u.
..
.,"
"0
i
,g "
!
. .. >." .... " ...,
0

<:)

I!

.!!

.."
.
" ,.. ...
. ".. ~
C
~
,g
.
.
". jj "
~.
..,

<
IE
0

Theory

>
0
u

.!!

"-

Ether

Emission

theories

J:

theories

'ii

Ii

Experiments from other fields

"

I!

"
'"
=
c

.....

"C

:;;

.!!

0.

g
C
c

,.:;
0

CD

~
>

E
c

c
.!!
~

0-

!!'

.lI

.S;

.c

'!'

>
0
E
E

.2

..

!!'

<!

.!!

a:

..'"
ec

>

.c

c;

..

0-

..,.!!0 lii
".5E
Z
::I C

c
s .!!C>.
8.'"
_c
"... 5~
e
~

Stationary ether,
no contraction

Stationary ether,
Lorentz contraction

Ether attached to
ponderable bodies

0
0

0
0

Original source

Ballistic

New source

Special theory of relativity

0
0

Legend: A: agreement, D: disagreement, N: not applicable.

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

119

As is easy to check, STR is more successful than any of the other ones, in
fact it is maximally successful as far as the 13 experimental facts are concerned.
Moreover, Lorentz's contraction version of the (stationary) ether theory is more
successful than the contractionless version. Similarly, the ballistic version of
the emission theory is more successful than the other two. However, it also
clear that many combinations lead to divided results. For instance, Lorentz's
theory is more successful in certain respects (e.g., De Sitter's spectroscopic
binaries) than the ballistic theory, but less successful in other respects (e.g., the
Kennedy-Thorndike experiments).
In the present approach it is plausible to define, in general, one type of
divided success as a liberal version of more successfulness. Y is almost more
successful than X if there are, besides some favorable facts and (possibly) some
indifferent facts, some unfavorable facts, but only of the B3-type, provided
there are (favorable) B8- or B9-facts or the number of B3-facts is (much) smaller
than that of their antipodes, that is, B7-facts. The provision clause guarantees
that it remains an asymmetric relation. Crucial is the special treatment of
B3-facts. They correspond to what is called Kuhn-loss: the new theory seems
no longer to retain a success of the old one. The idea behind the suggested
relatively undramatic nature is that further investigation may show that a
B3-fact turns out to be a success after all, perhaps by adding some additional
(non-problematic) hypothesis. In this case it becomes an (indifferent) B6-fact.
Hence, the presence of B3-facts is first of all an invitation to further research.
If this is without success, such a B3-fact becomes a case of recognized Kuhnloss. To be sure, when it concerns a general fact of a nomic nature it is more
impressive than when it concerns some general or individual fact that may be
conceived as 'accidental'. Unfortunately, Table 6.2. does not contain an example
of an almost more successful theory.
Cases of divided success may also be approached by some
(quasi-)quantitative weighing offacts. Something like the following quantitative
evaluation matrix (Table 6.3) is directly suggested by the same considerations
that governed the number ordering of the boxes.
It is easy to calculate that all qualitative (i.e., Table 6.1. induced) success
orderings of electrodynamic theories to which Table 6.2. gave rise, remain on
the basis of Table 6.3. in tact (which is not automatically the case). Moreover,
we then get of course a linear ordering, with Lorentz's theory on the second
position after STR far ahead of the other alternatives. Of course, one may
Table 6.3. The quantitative (comparative) evaluation matrix
X

Negative
Neutral
Positive

Negative

Neutral

Positive

B4: -1/ -1
B8: + 3/ -3
B9: +4/ -4

B2: -3/ +3
B5: 0/0
B7: + 2/ -2

BI : -4/ +4
B3: -2/ +2
B6: +1 / + 1

120

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

further refine such orderings by assigning different basic weights to the different
facts, to be multiplied by the relative weights specified in the matrix of Table 6.3.
Note that the qualitative and the quantitative versions of the evaluation
matrix can be seen as explications of some core aspects of Laudan's (1977)
problem-solving model of scientific progress, at least as far as empirical problems and their solutions are concerned.
Let us briefly consider the possible role of simplicity. In (Kuipers 1993, SiS)
we show that the examples of theory comparison presented by Thagard (1992)
to evaluate the computer program ECHO can well be treated by the comparative evaluation matrix (CEM), that is, ECHO and CEM are equally successful
with respect to these examples. However, ECHO is in many respects more
complicated than CEM. One respect is the fact that CEM only uses success
considerations, whereas ECHO applies simultaneously success and simplicity
considerations, even in such a way that, in theory, success may be sacrificed to
simplicity. Of course, as long as two theories are equally successful, we may
add as a supplement to the rule of success, that simplicity considerations
provide good, pragmatic, reasons to prefer the more simple ones, e.g., on the
meta-level CEM is to be preferable to ECHO. 9
This applies, by definition, always in case of observationally equivalent
theories. Similar to the case of acceptance (Section 2.3.), simplicity considerations and, more generally, our background beliefs, may determine which of
two (perhaps already falsified) theories is more plausible (in the sense that it
is closer to the theoretical truth, the theory realist will add). Background beliefs
may determine in this way our preference among observationally equivalent
(falsified) theories; of course, these beliefs do no longer serve this purpose if
they diagnose the two theories as equally plausible.
Be this as it may, returning to just momentarily equally successful theories,
we do not see good reasons to apply simplicity considerations as long as
success criteria lead to an ordering of two theories. Lack of good reasons seems
to hold at least for deterministic theories. In particular, in the case of such
theories there does not seem to be any link between simplicity and truth
approximation. However, from recent publications (see Sober 1998) for an
overview) it seems to be a link between simplicity and (observational) truth
approximation. If so, simplicity may well be counted as a kind of success.
6.2. EVALUATION AND FALSIFICATION IN THE LIGHT OF
TRUTH APPROXIMATION

Introduction
Whereas the method of HD-testing and HD-evaluation, and hence the evaluation methodology, have a falsificationist flavor, each with its own aim, they
are certainly not naive in the sense in which Popper's methodology has sometimes been construed. Naive falsificationism in the sense described by Lakatos

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

121

(1970, 1978) roughly amounts to applying HD-testing for purposes of theory


evaluation and elimination. Its core feature then becomes to further discard
(convincingly) falsified theories. Lakatos has also construed a sophisticated
version of falsificationism such that, when comparing theories, he takes their
'unrefuted content' into account, which suggests leaving falsified theories in the
game. Moreover, Lakatos has proposed a 'methodology of research programs',
which operates in a sophisticated falsificationist way. However, it works in
such a way that it postpones the recognition of falsifications of the 'hard core
theory' as long as it is possible to roll off the causes of falsification dogmatically
onto auxiliary hypotheses or background theories.
It will be argued that HD-evaluation can be seen as an explication of
sophisticated falsificationism, leaving room for a dogmatic, research program
specification. For easy reference, we will simply use the terms 'falsificationist
methodology' and 'falsificationism' in the naive sense, except when otherwise
stated.
Now it can already be made plausible by a suggestive picture (1) that the
falsificationist and the evaluation methodology may be functional for truth
approximation, and (2) that the latter non-falsificationist methodology, ironically enough, is much more efficient for that purpose. The basic proof of both
claims, in terms of the structuralist explication of truthlikeness, will be given
in the next chapter.
We like to make these claims about truth approximation already now, for it
enables us to sketch the main methodological consequence, which may stimulate the reader to 'consume' Part III and IV. This consequence is that a new
explanation, even justification, can be given for the observation of Kuhn,
Lakatos and others that there is quite a discrepancy between falsificationist
(methodological) theory and non-falsificationist practice. This will be elaborated in the next section.
6.2.1. The falsificationisr methodology

As suggested, we call a methodology a falsificationist methodology, when it


includes the claim that a conclusively falsified theory has become essentially
useless, and rules of theory selection are hence essentially restricted to not-yetfalsified theories. For example, in all standard probabilistic methodologies,
including the Bayesian, all falsified theories are thrown on the scrap-heap of
'theories with zero posterior probability', and theory selection is something
between 'theories with non-zero posterior probability'. Realist epistemologists
too distance themselves from falsified theories, as far as they consider the truth
question as the main question. Although empiricist methodology is usually not
presented as falsificationist, it prescribes to conclude essentially the same in the
case of falsification: a falsified theory has passed the fundamental border of
observational adequacy for which scientists are supposed to strive. Of course,
for all approaches it is possible to contest the supposition that we have to do

122

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

with a genuine falsification, because of all kinds of complications dealt with in


Subsection 5.2.3.
We get the technical core of the falsificationist methodology, based on the
asymmetric model of separate HO-testing of theories introduced in
Subsection 5.1.5., by restricting the interesting hypothesis (CSH) and the rule
(RS) to not yet falsified theories, indicated by rCSH and rRS, respectively, and
by adding the
Rule of Elimination (RE)
When a theory has been convincingly falsified, elimination should follow,
and one should look for a new theory

Hence, whereas the evaluation methodology is governed by the unrestricted


RS, the falsificationist methodology is governed by the rRS and RE (rRS&RE).
Let us first make some further comparative remarks of a methodological
nature. Of course, according to both methodologies, the application of their
respective rules (RS versus rRS&RE) has to be aimed at. As already suggested,
RS presupposes the unrestricted application of the separate and comparative
principles of HO-evaluation (PSE and PCE). Moreover, we suggested some
other specific heuristic principles that might stimulate the application of RS,
viz., the (unrestricted) application of the principle of content (PC) and the
principle of dialectics (PO). Aiming at the unrestricted application of RS and
hence of PCE, PC, PO and perhaps some additional principles can be summarized by the general heuristic principle:
Principle of Improvement (of theories) (PI)
Aim at a more successful theory, and successive application of RS

The term 'improvement' may be somewhat misleading, since the more successful
theory may be either an improvement of an older theory, e.g., within the same
research program, or really a new theory.
On the falsificationist side, rRS and RE both only presuppose the application
of all mentioned principles restricted to not yet falsified theories, for as soon
as we have obtained in this way a (convincingly) falsified theory, it is put out
of the game by RE. In other words, the falsificationist methodology, governed
by rRS&RE, presupposes the restricted application of PSE, indicated by rPSE,
and the restricted version of PI, indicated by rPI. If one does not yet dispose
of an unfalsified theory, to apply rPSE, one has to invent one.
It is also important to note that the application of RE and of PSE, whether
the latter is restricted or not, presupposes that the relevant theory is testable,
or falsifiable or confirmable:
Principle of Testability (PT)
Aim at theories that can be tested, and hence evaluated, in the sense

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

123

that test implications can be derived which can be tested for their truthvalue by way of observation
Hence, the relativization of the methodological role of falsification, inherent in
the evaluation methodology, should not be construed as a plea to drop falsifiability as a criterion for being an empirical theory. On the contrary, empirical
theories are supposed to be able to score successes, to be precise, general
successes. This evidently presupposes falsifiability. However, we prefer the
neutral term 'testability'to, for falsifiability and confirmability obviously are
two sides of the same coin. This observation is seriously in conflict with
Popper's critique on aiming at confirmation. Note also that Popper's plea to
give priority to testing the most unexpected test implications can equally well
be conceived as aiming at as high confirmation as possible, for surviving such
tests gives, of course, the highest success value.
In sum, the evaluation methodology, governed by RS, can now be summarized by PI, presupposing PSE and PT, whereas the falsificationist methodology, governed by rRS&RE, amounts to rPI&RE, both presupposing rPSE
and PT.
Though Popper primarily promoted that theories should be falsifiable, i.e.,
testable (PT), and that they should be tested, i.e., evaluated in our sense (PSE)
and improved in our sense (PI), one frequently suggests that he promoted the
more inclusive falsification principle, including RE, and hence restricted PSE
and PI to rPSE and rPI. However, the number of times that Popper seems to
plea for RE, or for the combination rPI&RE, is negligible compared to the
number of times that he pleads for the unrestricted PI. In this respect, it is
important to stress that Popper uses the expression '(principle or rule of)
elimination', almost always in the sense of 'elimination of error', and this is
precisely what PI amounts to. Hence, as Lakatos has suggested, Popper is
most of the time a sophisticated falsificationist. However this may be, it is at
least evident that rPI&RE, due to RE, does not use all opportunities for
empirical progress, whereas PI does. Hence, RE is not useful for empirical
progress in the sense of RS (and, as we will see, not for truth approximation
either). The only justification of RE that remains is a pragmatic one. If one
wants to use a theory to design a certain product or process and if it is
important in that context to avoid risks, it is plausible to apply RE when it is
possible.
But besides retarding empirical progress in the sense suggested, it is also
plausible to think that RE affects the prospects for truth approximation. A
striking feature of PI in this respect is that the question of whether the more
successful theory is false or not does not playa role at all. That is, the more
successful theory may well be false, provided all its counter-examples are also
counter-examples of the old theory. In the next chapter it will be proven that
RS, and hence PI, are not only functional (to say 'effective' would have here
too strong connotations) for approaching the truth in a precise sense, whatever

124

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

the truth precisely is, but that they are even efficient in doing so. On the other
hand, it will be shown that rRS&RE, and hence rPI&RE, is also functional for
truth approximation, due to rRS, but very inefficient, due to PE. The reason
is that RE prescribes that when a theory encounters a counter-example one
always has to look for a new theory that is compatible with the data thus far.
A short cut to the truth of a theory with many (types of) counter-examples,
via theories with fewer ones, is excluded. To be sure, the falsificationist methodology, including the comparative part, is functional and efficient in searching
for an answer to the leading question of testing a theory, viz., its truth-value.
To put it somewhat more generally and dramatically, first, something like
the cunning of reason is operative in scientific research: the evaluation and the
falsificationist methodologies, though no guarantee for truth approximation,
are both functional for truth approximation in a weaker but precise sense.
Hence, realists may claim that they approach the truth by using the falsificationist method, at least as a rule. However, more surprisingly, if one applies the
evaluation methodology, one comes, as a rule, closer to the truth, whether one
likes it or not.
Last but not least, the irony is that the cunning of reason works more
efficiently when the evaluation methodology is applied than when the falsificationist methodology is applied. The reason is that the falsificationist allows
himself, as it were, to be distracted by something which turns out to be
irrelevant for approaching the truth, viz., that the theory is false.
The proof of these claims starts from the asymmetric model of HD-evaluation
and is based on the structuralist theory of truthlikeness and plausible interpretations of individual problems and general successes. The main theorem,
called the success theorem, to be presented in the next chapter, states that
being the closest to the truth among the available theories guarantees that it
will be the most successful theory. To extend the proof to the two symmetric
models merely requires some transformation of individual successes and general
problems. By way of an easy lemma, it follows that the result of a crucial
experiment is always functional , though again, is no guarantee, for truth
approximation.

6.2.2. A suggestive picture


It is easy to convince the reader in statement terms that the claims of functionality and efficiency for truth approximation can be right. The set of statements
(more precisely, the set of classes of equivalent statements, also called (the set
of) propositions) that can be formulated within a certain vocabulary can be
partially ordered according to logical strength, see Figure 6.1. The tautology
and the contradiction form in this ordering two unique extremes. Between
these extremes there is for every strength more than one (non-equivalent)
statement. The number first increases and then decreases. A diamond-shape is
formed. For an arbitrary statement it holds that the set of statements following
from it, when ordered by strength, has roughly the same diamond structure.

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

.c:
.....

125

tautology

0)

Q)

.....en
L..

0)

en
co
Q)
L..

o
C

the truth
false
contrad iction

Figure 6.1. The landscape of truth

Naturally, this also applies to the set of statements following from the truth
(recall, the strongest true statement), which is, of course, precisely the set of
true statements. This is the case when the set of true statements are taken in
isolation. However, the closed diamond of true statements in Figure 6.1. only
arises after sufficient reshuffling of the true and false statements of the same
strength (i.e., horizontally). But the points that will be made do not depend on
this graphic simplification.
The following possibilities are easily read off from Figure 6.1, which may be
said to represent the landscape of the truth. A false theory may well be (much)
closer to the truth than a true one. And, although the question of whether a
theory is true or false is relevant for the question of whether the theory coincides
with the truth, the first question is irrelevant for its distance to the truth as
long as the theory in question does not exactly coincide with the truth. Finally,
it follows immediately from the landscape of the truth that it is possible to use
a theory development strategy, such as idealization and concretization, that
leads via a whole chain of false theories to the truth.
It is clear that all these possibilities exist due to the plausible explication of
'the truth' as the strongest true statement. However, it is also important to

126

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

note that the three indicated possibilities do not presuppose that the truth can
be recognized as such, nor that the truth is verifiable, let alone that it can be
established with certainty. What is needed is only that the truth gives recognizable signals, without making their source derivable from them.
These remarks are easy to combine with a literally geographical analogy for
truth approximation in general and the possibility of the irony of the cunning
of reason in particular. To find in the Netherlands the most south-east spot at
the same level as N.A.P. (Normaal Amsterdam Peil(=level, assuming some
very plausible arrangements, there must be precisely one such spot; there is no
reason to try with spasmodic efforts to start and remain in areas not below
N.A.P ..

6.3 . SCIENTIFIC AND PSEUDOSCIENTIFIC DOGMATISM

Introduction

As announced, we have made these claims about truth approximation plausible


in order to sketch the main methodological consequence of them. They enable
a new explanation, even justification, of the observation of Kuhn, Lakatos and
others that there is quite a discrepancy between falsificationist (methodological)
theory and non-falsificationist practice. Straightforward truth approximation
may be seen as the primary, conscious or unconscious, motive for non-falsificationist behavior is. Dogmatic behavior, in the sense of working within a research
program, is only a secondary motive for non-falsificationist behavior. Whatever
the main motive, as long as such behavior is directed at theory improvement
within the program, it can be distinguished from pseudoscientific behavior.
6.3.1 . Falsificationist theory and non-falsificationist practice

"Scientists do not practice what they, and philosophers of science, preach" is


an observation constituting one of the most important findings in empirical
studies of science. Popper, or at least his naive interpreters and followers, still
thought to have good reasons to prescribe that scientists behave non-dogmatically. Falsificationism was proposed as a specific methodology: theories have
to be tested deductively, and in case of convincing falsification, elimination of
the theory should follow.
However, Kuhn and Lakatos showed that in practice it is seldom that one
observes consistent falsificationist behavior. Appealing to the so-called DuhemQuine thesis, Lakatos has argued convincingly that dogmatic behavior can be
justified to a certain extent. As we have illustrated in detail in Subsection 5.2.3,
deriving test implications from a theory usually requires all kinds of additional
assumptions which can be blamed for the falsification. This enables one to
retain the 'hard core' of the theory, giving rise to the methodology of (scientific)
research programs. Opposed to this methodological justification of dogmatic

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

127

behavior, there is Kuhn's social explanation of non-falsificationist behavior.


He relativizes the importance of any method whatever. Science is a social
process and the so-called scientific method is more or less irrelevant dressing.
The similarity between Kuhn and Lakatos is that both call attention to the
open discrepancy between falsificationist (methodological) theory and nonfalsificationist practice.
There seems to be no reason to assume that there is just one general
explanation for all cases of non-falsificationist behavior. The different explanations of Kuhn and Lakatos can be very plausible in concrete cases and there
is equally no reason to exclude that in still other cases other explanations have
to be preferred.
However, it is possible to indicate a hierarchy of explanations. When a
philosophical-methodological explanation is possible, it has priority over a
purely social-psychological explanation. Moreover, one philosophical-methodological explanation may be more general than another.
A good reason to suppose that a more general explanation than the one of
Lakatos is possible is provided by what is perhaps the most striking case of
non-falsificationist behavior: the theory development strategy elaborated by
Nowak and others, called "idealization and concretization". According to this
strategy, one starts with a theory neglecting some factors of which one knows
that they are relevant. Hence, one knows beforehand that the theory, suggested
by strong idealization, is false. To note that, one does not need experimental
testing. The same ('to be born refuted') holds for all concretizations of the
theory that one obtains from successive accounting for, up to that point,
neglected factors. Only when one assumes that all factors that are supposed to
be relevant, have in fact been taken into account does it make sense not to
exclude the possibility that the theory is true. One may say, of course, that the
hypothesis that a certain factor is negligible is an auxiliary hypothesis that can
be blamed for the falsification. But the problem is that the falsification is not
at all a surprising result for which one has to find a cause. In other words,
there need not be an inclination to dogmatism with respect to a certain theory
one wants to save.
The strategies of theory development of Lakatos and Nowak are in fact
special cases of the evaluation methodology which can, but need not, go
together. Hence, if we can give a justification for the evaluation methodology,
this covers the rational explanation of the non-falsificationist aspect of both
strategies.
In our view the evaluation methodology, including its non-falsificationist
feature, is practiced because, as we have already indicated and will further
elaborate in the next chapter, it happens to be functional and efficient for truth
approximation. That is, scientists behave as if they are aiming at the truth
concerning some subject matter, and as if they know that straightforward
elimination of falsified theories is not the best way to achieve that goal.

128

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

6.3.2. Responsible dogmatic behavior

As mentioned, the evaluation methodology can be seen as an explication of


Lakatos' sophisticated falsificationism. The basic common feature is that theory
evaluation is primarily a matter of comparative evaluation, leaving falsified
theories in the game as long as there are no better alternatives. Our analysis
adds to this a justification in terms of the claim that this procedure, though
no guarantee for truth approximation, is functional for truth approximation,
without detours. There are also some important differences. They concern the
fact that Lakatos imposes two extra conditions on theory elimination, and
hence progress. First, testing the comparative success hypothesis should not
only show that the more successful theory remains more successful in the light
of these facts, but also that it has to score new extra successes ('novel facts').
In other words, empirical progress not only requires additional explanatory or
derivational success, but even additional predictive or retrodictive successY
Secondly, Lakatos ultimately proposes to use sophisticated falsificationist
means in a 'methodology of (scientific) research programs'. That is, he considers
something like RS primarily acceptable for theory transitions within a research
program, where a research program not only presupposes some vocabulary,
but also some hard core principles that should remain untouched, that is, one
should follow the dogmatic strategy.
From our analysis it will follow that these extra conditions are not necessarily
functional for truth approximation. However, concerning the first extra requirement, it can be made plausible on logical grounds that a theory closer to the
truth than another will almost always be able to score 'new extra successes'.
This not only makes it understandable that Popper and Lakatos could show
that in the history of science this usually occurs in the case of important theory
transitions, but also that they have taken it for a necessary feature.
Concerning the second extra requirement of Lakatos, sticking to the hard
core of a research program, we have already suggested that this is indeed an
important specific, but not obligatory, theory development strategy. As a matter
of fact, the evaluation methodology can easily be integrated in a 'methodology
of research programs'. Theories within a research program are separately
evaluated by the HD-method, if necessary, with recourse to dogmatic strategies,
and theory transitions can be made by applying RS on the basis of comparative
HD-evaluation.
Our methodological portrait also leaves room for a well-conceived transition
of one research program to another. It is possible that a new research program
turns out to 'have a theory' of which the corresponding observational theory
is more successful than that corresponding to the best known theory of an old
program. Of course, this is always a momentary judgement, because, as Lakatos
has stressed, the old program can later turn out to leave room for a theory
that surpasses the success of the best known theory belonging to the new
program. However this may be, as Lakatos has argued, something like a

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

129

principle of improvement of research programs is plausible: if it is difficult to


make empirical progress within a program, aim at a more successful program,
whether by repairing the old one or by replacing the old one by a new one.
There remains the justification of dogmatic behavior in the sense of a research
program. Recall that, according to Lakatos, and we like to agree with him,
scientists develop a research program on the basis of some main idea, in the
form of one or more core hypotheses. They do so by continually thinking of
new auxiliary hypotheses that form together with the core hypothesis a better
total theory. By successfully doing so, there arises ideally a succession of ever
better theories with the same hard core. Hence, in the development of a research
program, the principle of improvement (PI) is applied while saving the central
dogmas. For this reason, such research frequently has the character of presupposing or applying these dogmas, instead of evaluating them. Such research is
nevertheless governed by PI, and hence functional for empirical progress and
even truth approximation. Accordingly, dogmatic behavior may be responsible
scientific behavior when it is combined with PI.
Although it is difficult to find a specific statement to this effect, it seems that
Lakatos implicitly assumed that the hard core of a program can only be
sustained with success during a long period of time when it is true. For
otherwise it is difficult to explain why Lakatos did not account for the regularly
occurring phenomenon that the hard core does not appear to be so hard as it
seemed, and is adapted. However, in the light of the foregoing, it appears that
improving the 'hard' core, while sticking to the vocabulary, may be equally
good for empirical progress and truth approximation. When this also does not
work, it is time to look for completely different dogmas. 12 Whether one appreciates it or not, young researchers are frequently quicker of the mark to search
out new dogmas than are established scientists.
In sum, we may say that scientific dogmatic behavior satisfies a concretized
version of PI:
Principle of improvement guided by research programs (PIRP)

One should primarily aim at progress within a research program, i.e.,


aim at a better theory while keeping the hard core of the program in
tact. If, and only if, this does not work out, try to adapt the hard core,
while leaving the vocabulary in tact. If, and only if, this is also unsuccessful, look for another program with better perspectives on progress
Note first that PIRP is so formulated that it contains something like a principle
of improvement of programs, where the latter is in the first instance guided by
a vocabulary. Since PIRP is a special version of PI, it is also functional for
truth approximation. From the work of Kuhn and Lakatos, it may be concluded
that PIRP has been very successful in the history of science.
Descriptive research programs frequently do not have a hard core proper,
at least not a specific one, but only a core vocabulary. If one wants to talk

130

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

about a hard core in this case, it is the general idea that the core vocabulary
generates, by itself or in extended versions, an interesting restriction of what is
possible in reality. This suggests that there is, between the general PI and the
special version PIRP, a broader special version in between, dealing with dogmatically sticking to 'core vocabularies':
Principle of improvement guided by core vocabularies (PICV)

One should primarily aim at progress within a core vocabulary, i.e., aim
at a better theory while keeping the core vocabulary in tact; if, and only
if, this does not work, look for another vocabulary, which mayor may
not be very different
Again, PICV is so formulated that it includes something like a principle of
improvement of vocabularies. Moreover, since PICV is a special version of PI
it is also functional for truth approximation. PICV has been very successful in
the history of science, in particular where descriptive research is concerned.
The history of 'descriptive (or inductive, see below) thermodynamics', dealing
with the search for relations between volume, pressure and temperature, and
the history of 'descriptive chemistry', dealing with chemical reactions, provide
cases in point.
6.3.3. Pseudoscience

From the foregoing it follows that dogmatically dealing with theories has to
be qualified as unscientific when one apparently does not aim at applying PIRP
or PICV. Moreover, it is plausible to characterize pseudoscience as the combination of scientific pretensions and the neglect of PI, in particular its dogmatic
versions PIRP and PICV. This characterization can be seen as an, in some
respects, improved version of that in the introduction of Lakatos (1970, 1978).
The standard examples of pseudoscience, such as astrology, graphology, homeopathy, parapsychology, creationism, and ufology, satisfy these conditions. 13
In all these cases, it is not only quite clear that central dogmas are the point
of departure of unscientific research and application, but it is also rather easy
to indicate how they could become the point of departure of serious research.
Such research, however, is seldom started.
We do not, of course, claim that within the sphere of academic research
pseudoscientific behavior does not occur, but it takes place less than outside
that sphere. Marxist economics, psychoanalytic psychology, and evolutionary
biology increasingly seem to follow the general rules and principles of scientific
research. However, claims of this kind are highly controversial, as the works
of Biaug (1980) and Grunbaum (1984) illustrate for marxist economics and
psychoanalytic theory, respectively.
An interesting question concerns how theology and philosophy should be
evaluated in this respect. So-called systematic theology certainly has scientific
pretensions, but usually no empirical ones. Nevertheless, it has, directly or

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

131

indirectly, pretensions regarding empirical reality. Hence, the question of


whether theology has more resemblance to science than to pseudoscience is
not easy to answer. In view of the unclear nature of the pretensions, this
question cannot be separated from other aspects of scientific evaluation of
other theological claims to knowledge, such as the following ones. To what
extent are they in conflict with empirical scientific insights, and what would a
straightforward empirical scientific approach to these claims, and their persistence, look like? So-called religious studies direct their attention primarily at
the description of other than Western religious cultures, but it is plausible also
to take into account in this connection systematic theology and to direct
attention to the evaluation and explanation of religious belief claims.
The three general principles, PT, PSE and PI, including its program version,
do m.m. also apply to philosophy. Of course, in this case, testability and
evaluation do not now refer to empirical matters, as far as there are no empirical
pretensions. Successes of conceptual theories, for example, should be interpreted
as satisfactorily treated cases and realized general conditions of adequacy.
Counter-examples become problems arising from cases and aspects that are
evidently wrongly dealt with. As a consequence, there is also in philosophy
much room for making progress in the form of responsible dogmatism, and
good philosophy can rather easily be distinguished from pseudo- or paraphilosophy. In the case of pseudo-philosophy, philosophical pretensions are
usually combined with unscientific dogmatism. Several forms of exegesis of
great philosophers belong to it. Instead of trying to develop or replace research
programs from earlier philosophers, an attempt is made to preserve all statements of 'the master'. The previous and the present chapter may be seen as the
result of an attempt to improve some of the conceptual aspects of the work of
Popper and Lakatos, instead of defending them at all costs.

Final remarks: the context of evaluation


Although it may be conceded that the scientific method does not exist, this
does not yet imply that any method works, as Feyerabend (1975) suggested
with his slogan "anything goes". It is more realistic to start with the distinction
(to be specified pragmatically) between two aspects of scientific research, viz.,
invention and testing of theories. This distinction became known as the Context
of Discovery versus the Context of Justification.
For the Context of Discovery it may well be the case that almost all conceivable methods, from inductive generalization to sleeping another night, work in
certain cases: "anything goes sometimes".
Within the Cbntext of Justification, a universal method may also be lacking.
However, pace Glymour, the HD-method certainly is a dominant method.
Unfortunately, the term 'Context of Justification', whether or not specified in
a falsificationist way, suggests, like the terms 'confirmation' and 'corroboration',

132

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

that the truth or falsity of a theory is the sole interest. Our analysis of the
HD-method makes it clear that it would be much more adequate to speak of
the Context of Evaluation. The term 'evaluation' would refer, in the first place,
to the separate and comparative HD-evaluation of theories in terms of successes
and problems. As we have indicated, it may even refer to the further evaluation
of their relative merits in approaching the truth, or at least the observational
truth.
As a consequence, the foregoing may not only be interpreted as a direct plea
for restricting the test methodology to cases where our only interest is the truth
question. As soon as we are interested in the success question, then the evaluation methodology is more adequate. That methodology can, moreover, be
justified, surprisingly enough, in terms of truth approximation.
Let us, finally, pay some more attention to the pro's and con's of our strict
comparative approach. As has been stressed already, there are few theories
that can be ordered in terms of 'more successfulness' according to our strict
definition. The same holds for our basic and refined orderings in terms of 'more
truthlikeness' in the following chapters. Hence, the limited applicability of our
comparative notions might be seen as a serious shortcoming, supporting a
more liberal comparative or even quantitative approach. To be sure, such
liberalizations are very welcome as far as they are realistic concretizations. We
introduced already the idea of 'almost more successful', which has a plausible
analogue in terms of 'almost more truthlike', which we will, however, not
pursue in further detail. In Chapter 10 we will introduce the refined notions of
'more successfulness' and 'more truthlikeness'. The basic idea behind the (further) refined notion of more successfulness is that one counter-example may
nevertheless be better than another. Although it could be presented already
now, for it is a matter of empirical success evaluation, leading to a refined
version of the rule of success, it is technically more suitable to present it after
the presentation of the refined notion of 'more truthlikeness'. In Chapter 10
we will also question the usefulness of quantitative liberalizations of'successfulness' and 'truthlikeness', mainly because they need real-valued distances
between models, which are very unrealistic in most scientific contexts. Hence,
the applicability of liberal notions may well be laden with arbitrariness. For
this reason, we want to focus on unproblematic cases, guaranteeing that we
get the bottom line of progress and rationality. In Chapter 11 we will deal with
the succession of theories of the atom of Rutherford, Bohr and Sommerfeld
and argue that it is a sequence of increasing success and, potentially, even of
increasing truthlikeness, both in the strict (refined) sense. Hence, although the
strict approach may not have too many examples, it has impressive examples.
Moreover, it is even more important that the strict strategy does not lead to
void or almost void methodological principles. If there is divided success
between theories, the Principle of Improvement amounts, more specifically, to
the recommendation to try to apply the Principle of Dialectics: "Aim at a
success preserving synthesis of the two RS-escaping theories", of course, with

EMPIRICAL PROGRESS AND PSEUDOSCIENCE

133

a plausible program-bound version. Similarly, for truth approximation aims: if


there is reason to suppose that two theories cannot be ordered in terms of
'more truthlikeness' in the strict sense, the challenge is to construe a theory
which is more truthlike than both. In sum, the restricted applicability of strict
notions of comparative success and truthlikeness does not exclude the possibility to formulate clear challenges in cases where they do not apply, on the
contrary.

PART III

BASIC TRUTH APPROXIMATION

INTRODUCTION TO PART III

This part of the book introduces and analyzes the theory of naive or basic
truth approximation, and its relation to empirical progress and confirmation,
first for epistemologically unstratified theories, later for stratified ones.
In Chapter 7 the qualitative idea of truthlikeness will be introduced, more
specifically the idea that one description can be closer or more similar to the
truth than another, called 'actual truthlikeness', and the idea that one theory
can be closer to the truth than another, called 'nomic truthlikeness'. In the first
case, the truth concerns 'the actual truth', that is the truth about the actual
world, or the actualized possibility, as can be expressed within a given vocabulary. In the second case it concerns 'the nomic truth', i.e., the strongest true
hypothesis, assumed to exist according to the 'nomic postulate', about what
are the physical or nomic possibilities, called 'the nomic world', restricted to a
given domain and, again, as far as can be expressed within a given vocabulary.
Besides indicating some plausible ways of approaching the actual truth, it will
be argued that the evaluation methodology is effective and efficient for nomic
truth approximation. In this and the following chapter, the 'basic' explication
of nomic truthlikeness and truth approximation will be at stake, which, though
appealing to scientific and philosophical common sense, has no real-life scientific examples as yet. A preview will be given of more realistic bifurcations of
actual and nomic truth approximation to be elaborated in later chapters.
Chapter 7 closes with explicating the role of novel facts, crucial experiments,
inference to the best explanation, and the idea of inductive research programs
in the light of truth approximation.
Chapter 8 begins by arguing that ' basic' nomic truthlikeness and the corresponding methodology have plausible conceptual foundations, of which the
dual foundation will be the most appealing to scientific common sense: 'more
truth like' amounts to 'more true consequences and more correct models',
whereas 'more successful' amounts to 'more established true consequences, i.e.,
successes, and fewer established incorrect models, i.e., counter-examples'. The
theme of conceptual foundations will recur in later chapters, and the adapted
dual foundation will remain superior to the two uniform foundations, i.e., one
solely in the terms of models and the other solely in terms of consequences.
Next, it will be argued that actual and basic nomic truthlikeness suggest a nonstandard, viz., intralevel instead of interlevel, explication of the main intuitions
governing the so-called correspondence theory of truth. Finally, it will be made
137

138

INTRODUCTION TO PART III

clear that the presented cognitive structures suggest logical, methodological


and ontological explications of some main dialectical concepts, viz., dialectical
negation, double negation, and the triad of thesis-antithesis-synthesis.
Chapter 9 introduces the first major sophistication, the stratification arising
from the (changing) distinction between observational and theoretical terms,
leading in the first place to the distinction between observational and theoretical
truth approximation. In the second place it leads to the idea of 'the referential
truth', i.e., the truth about which terms of a vocabulary refer, and which do
not, where reference is precisely defined on the basis of the nomic postulate.
In the third place, this enables the definition of the referential claim of a theory
and hence of the idea of one theory being closer to the referential truth than
another. Fourth, and finally, it suggests the idea of 'the substantial truth', i.e.,
the strongest true hypothesis of the sub-vocabulary of referring terms, and
hence of 'substantial truth approximation'. Besides presenting several elaborations, speculations and extensions, it will be argued that the evaluation methodology is not only functional for observational truth approximation, but also,
though less straightforwardly, for theoretical and substantial truth approximation, leading to a further relativization of falsification. However, the prospects
of referential truth approximation by the evaluation methodology are the least
impressive. Hence, according to the resulting evaluation of epistemological
positions, the instrumentalist can better become a (nomic kind of) constructive
empiricist, who may, even in the face of experimental evidence from various
kinds, stubbornly continue to refuse to become a referential realist, who in his
turn can well become a theory realist.
For the sake of overview, we indicate already here the main lines of the next
part, called 'refined truth approximation'. It starts with another sophistication
of the basic approach: it accounts for the fact that progress is frequently made
by new theories that introduce new mistakes, something which is excluded
according to basic truth approximation. This refinement will allow some reallife illustrations of (potential) truth approximation, one from physics and
another from economics. Moreover, in that part it will be shown that there are
also quantitative versions of refined truth approximation, based upon distances
between structures.
Recall that in the last chapter of the book, which can largely be read
independently of the refinements of Part IV, we will draw the main lines of the
resulting favorite epistemological position suggested by the analysis in this
book, viz., constructive realism.

7
TRUTHLIKENESS AND TRUTH APPROXIMATION

Introduction
When Karl Popper published in 1963 in Chapter 10 of Conjectures and
Refutations his definition of 'closer to the truth' this was an important intellectual event, but not a shocking one. Everybody could react by saying that the
definition was as it should be, and even that it could have been expected. For
plausible the definition was indeed: a theory is closer to the truth than another
if the true consequences of the first include those of the second and the false
consequences of the second include those of the first.
About ten years later the event of 1963 became shocking with retrospective
effect when David Miller (1974) and Pavel Tichy (1974) independently proved
that a false theory, in the sense of a theory with at least one false consequence,
could according to Popper's definition never be closer to the truth than
another one.
With this proof they demolished the definition, for it could not do justice to
the presupposed nature of most of the culminating points in the history of
science. New theories, such as Einstein's theory, though presumably false, are
more successful than their predecessors, such as Newton's theory, just because
they are closer to the truth. In other words, the greater success of new theories
cannot be explained simply in terms of the truth of new theories and the falsity
of their predecessors, but should be explained in terms of their decreasing
distance from the truth: that a sequence of presumably false theories is converging to the truth should explain their increasing success. Of course, 'the truth'
should then be interpreted as the unknown strongest true hypothesis about
the relevant domain of phenomena within a certain vocabulary.
Miller and Tichy unchained with their proof in the beginning beside signs
of deception mainly skeptical remarks like "only the intuitive idea is important,
fortunately", "it shows that you can't solve philosophical problems by formal
means" and last but not least "it is the punishment for freely speaking about
the truth" 1.
However, after some time, Miller, Tichy and other philosophers recovered
from the fright and started to develop alternative definitions of 'closer to the
truth'. Today there can be distinguished at least five approaches to verisimilitude
or truthlikeness, as the subject is called more and more. This name refers to
the fact that the distance of a theory to the true theory gradually became
139

140

TRUTHLIKENESS AND TRUTH APPROXIMATION

reinterpreted as the similarity of the theory with the true theory. The core of
the problem is the explication of the similarity of theories. The concept of
converging to the truth then follows almost automatically.
Four approaches are represented by the following authors and publications:
Niiniluoto (1987a/ b), Oddie (1986, 1987a/ b), Schurz and Weingartner (1987),
Brink and Heidema (1987). The fifth approach, called the structuralist
approach, was first published in its naive or basic form in 1982 (Kuipers 1982b).
It was independently introduced by Miller (1978) for a very special case. The
structuralist definition is here presented in its full basic, stratified and refined
forms (this and the following chapters), leaning heavily on (Kuipers 1982b,
1984a, 1987b, 1992a/b). (Kuipers 1987a) contains a parade of the approaches,
except that of Brink and Heidema. Niiniluoto (1987a/ b) and Oddie (1986)
present versions of the so-called similarity approach. For recent, detailed, and
comparative, studies, see (Kieseppa 1996) and (Zwart 1998).
All approaches have in common that they succeed in avoiding the problem
of Popper's definition. Hence, a sequence of false theories converging to the
truth is technically possible. This enables the formulation of the claim that
historical sequences of theories constitute, as a rule, sequences of theories
converging to the truth. However, this does not yet imply that all these
approaches can give a clear justification for the already suggested basic intuition
that the increasing success of successive theories can be explained in terms of
increasing truthlikeness of them. Justification of this intuition may be called
the challenge of Larry Laudan (1981). According to Laudan, realists usually
assume this connection, but they should prove it. Two of the five approaches
explicitly claim to do so: Niiniluoto's 'quantitative likeness' approach and our
structuralist approach.
Like Popper's point of departure, all approaches refrain from a purely
metaphysical conception of 'the truth' and agree upon some kind of moderate
metaphysical realism. It is assumed that 'the truth' about a certain part or
aspect of reality is on the one hand determined by the previously chosen
vocabulary, and within these conceptual boundaries it is determined by the
nature of reality itself. The vocabulary determines what aspects of reality can
be expressed. It is evident that there are for each domain several vocabularies,
leading to equally many truths about that domain. In contrast, extreme metaphysical realism may be supposed to claim that reality carries with it its ideal
vocabulary.
Moderate realism does not need to degenerate into conceptual relativism, as
long as it is assumed that there may be all kinds of overlap, connections and
constraints between vocabularies and that such vocabularies may be put
together into more encompassing vocabularies. In other words, fundamental
incommensurability between paradigms, research programs, conceptual frameworks and the like, need not be assumed as the rule, and it is even possible to
reject it on formal grounds (as will be argued in Subsection 9.3.3.). Moreover,
even if fundamental incommensurability is sometimes unavoidable, this does
not yet exclude practical commensurability.

TRUTHLIKENESS AND TRUTH APPROXIMATION

141

This form of realism is highly defensible for the natural sciences: it implies
that theories are true or false and leaves room for the possibility that theoretical
terms pretend to refer to something in reality. However, this type of realism is
difficult to defend for the social sciences, because the conceptual representation
of social reality as well as social reality itself are both human constructions.
Although there may be assumed to be a human-independent natural world,
speaking of a human-independent social world is a contradiction.
The approaches differ technically. In the first place, many varieties of using
syntactic and semantic means occur. It is only the structuralist approach that
refrains, by definition, from any explicit role of (the sentences of the relevant)
language (although it can be reconstructed in model-theoretic terms, see
Section 8.1.). In the second place, one may focus immediately on a quantitative
notion of truth likeness, as e.g., Niiniluoto, or claim that the problem of qualitative or comparative truthlikeness should have priority, as is for instance plausible in the structuralist approach. Strictly speaking, any comparative approach
is concerned with the notion of 'more' or 'increasing truthlikeness', but we will
frequently just talk about 'truthlikeness' when misunderstandings are excluded.
Another difference is of a more fundamental nature. One may start from the
idea that there is essentially one type of (conceptually relative) truth: the truth
about the actual world, and hence one problem of truthlikeness. On the other
hand, one may claim, as in the structuralist approach, that there are essentially
two problems of truthlikeness. The first is engaged with explicating the idea
that one description is more similar to the true description than another
description, the problem of actual truthlikeness; the second is engaged with
explicating the idea that one theory is more similar to the true theory about
what is possible in reality, i.e., what is nomically possible, than another theory,
the problem of nomic truthlikeness 2 .

Whatever the relative problems and merits of the post-Popper approaches,


they not only don't have the drawback of Popper's, most of them display the
feature of language dependence. This feature was discovered by David Miller
and is conceived by him (Miller 1990) and others as a serious problem for any
explication of truthlikeness. It amounts to the fact that a description or theory
may be closer to the truth than another such that after an ingenious transformation into another language, the ordering is reversed. The second theory has
become closer to the truth than the first. However, whether it is really a
problem is a matter of serious debate (see e.g., Niiniluoto 1987a; Mormann
1988; Zwart 1995, 1998). That is, at least up to now, most definitions are
essentially language dependent, but the question is how this should be evaluated. Is the transformation a genuine translation and is the resulting reversal
of order dramatic, unavoidable or even desirable? Whatever the conclusion is,
the same conclusion then has to be drawn for the idea that one theory is more
successful than another, for it follows immediately from analogues to Miller's
argument that this notion, if explicated along similar lines, is equally language
dependent. Hence, language dependence is not specific for truth likeness, but

142

TRUTHLIKENESS AND TRUTH APPROXIMATION

extends to other comparative notions used by philosophers and scientists of


all sorts. Oddie (1986, p. 158) illustrates the same point with some other notions
and we are strongly inclined to agree with his summary statement: "An argument which purports to show that the notions of accuracy, truthlikeness,
structure, change, sameness of state, confirmation and disconfirmation, are all
spurious ( ... ) must harbor a defect somewhere."
For this reason, if the debate would end in the conclusion that language
dependence is a serious problem, this is devastating for scientific practice in
general. To be sure, we are more inclined to the opposite view. In the light of
successful scientific practice, it is more probable that language dependence is
at most an academic problem. However this may be, the basic form of our
explication of nomic truthlikeness, including the stratified version of Chapter 9,
is language independent, but the refined version, to be introduced in Chapter 10,
as well as the explication of actual truth likeness, to be introduced in this
chapter, is language dependent.
Section 7.1. and 7.2. deal with the logical problems of truth likeness, where
it is assumed that the truth is known, and the only question is how to define
'more actual truthlikeness' and 'more nomic truthlikeness', respectively.
Section 7.3. deals with the relation between truthlikeness and success, with
the Success Theorem as core, giving rise to the Forward Theorem, and leading
to the conclusion that the evaluation methodology explicated in Part II, governed by the rule of success, is functional for truth approximation 3 .
Section 7.4. presents a survey of ramifications and refinements of both
truthlikeness definitions.
Section 7.5. first reconsiders the comparative methodological conclusion of
the truth approximation analysis anticipated at the end of Chapter 6, viz., the
superiority of the evaluation methodology in comparison with the falsificationist methodology. Then we deal with novel facts and crucial experiments from
the truth approximation perspective. Next we argue that so-called 'inference
to the best explanation' has to be seriously revised. We conclude with an
explication of the idea of inductive (descriptive) research programs from this
perspective.
7. 1. ACTUAL TRUTH LIKENESS

Introduction

In this section we introduce some structuralist notions and then specify the
definition of 'actual truthlikeness' for structure descriptions, in particular for
those of a propositional nature.
7.1.1. Actual truthlikeness

We start with the problem of actual truthlikeness. Consider the following


electrical circuit (see Figure 7.1.).

TRUTHLIKENESS AND TRUTH APPROXIMATION

143

pO

(~rl

'-!--

p1

"'p2

,,",

p4

p3

Figure 7.1. An electrical circuit

Let Pi for 1::;; i::;; 4 indicate that switch i is on (+-+) and -, Pi that it is off (t).
Let Po( -, Po) indicate that the bulb lights (does not light). It is assumed that
the bulb is not defective and that there is enough voltage. A possible state of
the circuit can be represented by a conjunction of negated and un-negated p;'s.
It is clear that there is just one true description of the actual world, i.e., state,
of the circuit as it is depicted, PO&Pl & -, P2 &P3 &P4, according to the standard
propositional representation. Hence, the example nicely illustrates, among
others, that we consider 'the actual world' primarily as something partial and
local, i.e., one or more aspects of a small part of the actual universe. However,
it need not be restricted to a momentary state, it may also concern an actual
trajectory of states. In sum, the actual world is the actual world in a certain
context.
For the general case, let there be given a domain D of natural phenomena
(states, situations, systems) to be investigated. D is supposed to be circumscribed
by some informal, intensional description and may be called the primitive set
of intended applications. Let there also be given a vocabulary V, designed to
characterize D. V is some kind of set-theoretic vocabulary, i.e., a vocabulary
specifiable in set-theoretic terms. One type concerns ordered sets of elementary
possibilities represented by so-called elementary propositions, as in the case of
the circuit example (propositional vocabularies). Another type concerns ordered
sets of terms (or components), indicating domain-sets, as well as properties,
relations and functions defined on them (first order vocabularies). A vocabulary
V gives rise to a set Mp(V), or simply Mp, of potential models or conceptual
possibilities4 . Mp is also called the conceptual frame, designed for D. It may be
assumed that Mp is, technically speaking, a set of structures of a certain
(similarity) type; below a characterization will be given of first order structures
of the same type. In practice Mp will be the conceptual frame of a research
program for D. Note that V has to contain a proper sub-vocabulary, the
vocabulary of the domain, that enables fixing D in an unambiguous way, viz.,
as a subset of the set of conceptual possibilities generated by that subvocabulary.
The confrontation of a particular situation or state of affairs or system in D,

144

TRUTH LIKENESS AND TRUTH APPROXIMATION

the actual world so to speak, with Mp, is assumed to generate just one correct
representation or one true description, indicated by t.
The problem of actual truth likeness amounts to the explication of the idea
that an arbitrary description, i.e., an arbitrary conceptual possibility, is (more
similar or) closer to the true description than another one. As suggested by the
example, we will first concentrate on propositional descriptions.
7.1.2. Truthlikeness of propositional descriptions

Propositional descriptions, also called propositional constituents, are generated


by a fixed propositional vocabulary, that is, a finite set of elementary propositions: EP = {pI, p2, ... , pn} indicating elementary possibilities. Think of the
conceptually possible states of the electric circuit, where an elementary proposition corresponds to the claim that a certain elementary possibility holds, e.g
that a switch is on or that a bulb lights. A constituent is normally defined as
an arbitrary conjunction of the elementary propositions, each of them
un-negated (indicated by '+ ') or negated, leading to the following conceptual
possibilities: (+ / "'1 ) pI & ( + / "'1 ) p2 & ... & ( + / "'1 ) pn. However, assuming
that the conceptual frame Mp is restricted to the constituents of a fixed set EP,
it is easy to see that a constituent can also be represented by the set of (for
example) its un-negated elementary propositions. Let x, y, z, etc. indicate subsets
of EP representing constituents in this way, and t represents the relevant subset
for the actual world.
An arbitrary description x is, more precisely than before, defined as the
constituent x together with the claim "x = t", and it is called true or false when
its claim is true or false, respectively. Notice that there is, as desirable, just one
true description, viz., t, also called the (actual) truth.
Now the basic intuition behind more truthlikeness of descriptions can be
expressed by a simple example in standard propositional terms: pl& "'1 p2 is
more similar to "'1pl&"'1p2 than pl&p2. It is easy to check that the following
definition of (more) truthlikeness of descriptions is a generalization of that idea
in non-standard terms. Description y is at least as similar (close) to the truth
(t) as description x, indicated by sp(x, y,t) if the following two conditions are
satisfied:
(Oi)
(Oii)

y-tisasubsetofx-t
t - y is a subset of t - x

There is also a strict version, requiring that in at least one case the subset is a
proper subset or, equivalently, that sp(y, x,t) does not obtain at the same time.
This strict version is represented in Figure 7.2.
It is easy to check that the conjunction of the two clauses is equivalent to
the claim that the symmetric difference y L1 t between y and t, defined as
(y - t) U (t - y), has to be a (proper) subset of the symmetric difference between
x and t.

145

TRUTH LIKENESS AND TRUTH APPROXIMATION

EP
X

-----

8
'<

3*
_

,/"

6 ,~

~~4 ~"~

r-~\

7/~
/

//

,/)

~J

/-~
.

'----- t
Figure 7.2. Actual truthlikeness of propositional descriptions: y is more similar to t than x: sets 1
and 5 empty, sets 3 or 7 non-empty

Returning to the standard notation for the circuit example, it is easy to


check that PO&'Pl&'P2&P3&P4 is, according to the given definition, closer
to the true description of the depicted state PO&Pl &, P2&P3&P4 than
PO&'PI &'P2&P3&'P4 '
The definition transforms into a general definition of likeness of constituents,
sp(x, y, z), by replacing 'description' by 'constituent' and the true description 't'
by the variable constituent z.
But let us return to truthlikeness of descriptions. When the claim of a
description is true or false with respect to a particular elementary proposition,
that proposition is called an (actual) 'match' or 'mistake' of the description,
respectively. Hence, the members of x - t and t - x are the mistakes of description x, and the members of x n t and x n t are its matches ('x indicates the
complement of x, i.e., EP - x) are its matches. Now it is easy to check that
description y is at least as close to the truth as description x if and only if the
set of matches of the former includes those of the latter, or, equivalently, if and
only if the set of mistakes of the former is a subset of that of the latter. 5
There is a plausible quantitative variant: the distance d(x, t) between constituents x and t is 1t - xl + 1x - t I. This definition has already been proposed by
Tichy (1974) (in terms of the size of the symmetric difference between x and
t). However, we will postpone further attention to quantitative truth likeness
to Chapter 12.

7.1.3. Structurelikeness and truth likeness of structure descriptions


Propositional descriptions are special kinds of descriptions. In general, a
description can always be represented as a set-theoretic structure, a structure
description, with the claim that it is equal to the true description. To fit this

146

TRUTH LIKENESS AND TRUTH APPROXIMATION

perspective, a propositional constituent x can be represented more formally as


the structure (EP, x).
We will not aim at a general definition of (non-trivial) truthlikeness of
structure descriptions, or more generally of structurelikeness, for a precise
definition will have to depend on the specific nature of the conceptual possibilities. We will only give some additional examples.
But there is an extreme kind of structurelikeness for which a general definition
is possible. Trivial structurelikeness, indicated by triv(x, y, z), is generally
defined by x = y = z.
It is plausible how to formulate the idea of structurelikeness for first order
structures. Such a structure is an ordered set of domain-sets Dj and relations
(including properties and functions) R j defined on one or more of them. Two
structures are of the same (similarity) type6 when they have the same number of
domain-sets and relations and when the 'corresponding' relations are always
defined in the same way on the 'corresponding' domain-sets, e.g., Rl is always a
function from Dl onto D2 . An example of a similarity type is, for instance, the
class of all structures of the form (D, R), where R is a binary relation on the
domain-set D. Let Mp consist of structures ( ... Dj ... ; ... R j . ) of a fixed similarity type.
First order structurelikeness sf(x, y, z), y is at least as similar to Z as x, is now
defined by the requirement that the corresponding domains are the same (Dj(x) =
Dj(y) = Dj(z) for all i) and that the corresponding relations are related as follows:
Rj(z) - Rj(Y) is a subset of Rj(z) - Rj(x) and Rj(Y) - Rj(z) is a subset of
Rj(x) - Rj(z) for allj. Note that this is equivalent to the claim that the symmetric
difference between Rj(Y) and Rj(z) is a subset of that between Rj(x) and Rj(z) for
all j. Note also that three different first order structures x, y and z, related by
sf(x, y, z), can only be realized at different moments, due to the fact that Sf (x, y, z)
requires that they have the same domain-sets.
Let us also indicate some examples of elementary real number structures specifying one or more ordered real numbers. Formally, the domain-set is the set of real
numbers, or the relevant Cartesian product with two or more 'dimensions', but
we will leave the domain-set implicit. Real numbers of the same dimension will
be indicated by numbered x's, etc. For one dimension it is plausible to define
structurelikeness by "xl ~ x2 ~ x3 or x3 ~ x2 ~ xl", indicated by sl(xl, x2, x3).
For two dimensions one possible definition reads "sl(xl, x2, x3) and
sl(yl, y2, y3)", indicated by s2xl, yl), (x2, y2), (x3, y3.
Of course, actual truthlikeness amounts to the structurelikeness claim that
structure description y is, in the relevant sense, at least as (or more) similar to
the true structure description t as (than) structure description x.
7.2. NOMIC TRUTH LIKENESS

Introduction
The idea of nomic truthlikeness is based on the Nomic Postulate to be presented
first. Following this, an outline is given of the structuralist approach to theories.

TRUTHLIKENESS AND TRUTH APPROXIMATION

147

Then we present the so-called basic definition of '(more) nomic truthlikeness',


in which all mistakes of a theory are treated equally. We conclude with some
formal properties of basic nomic truthlikeness.

7.2.1. The Nomic Postulate


Recall the circuit example (Figure 7.1.). It is clear that there is not only just
one true description of the actual state, but also just one true theory in the
sense of a characterization of the one and only set of all nomically possible
states of the circuit. This true theory can be represented by the propositional
formula: Po ~ (((PI &P2)VP3)&P4). Our new subject is the likeness of theories to
the true theory. We start with a number of preparations.
Recall that D is the domain of natural phenomena, V the vocabulary designed
to characterize D, and Mp the set of conceptual possibilities, i.e., the (conceptual)
frame, generated by V. The confrontation of D with Mp, i.e., D seen through
Mp, is assumed to generate a unique, time-independent subset Mp(D) = T of
all physical or nomic possibilities, i.e., the set of Mp-representations of the
nomically possible members of D, also called the Mp-set of intended applications. Although the assumption of such a unique, time-independent set is
somewhat metaphysical, it is, due to its vocabulary or conceptually relative
nature, metaphysical of a moderate kind. The assumption will be called the
Nomic Postulate associated with <D, V) or <D, Mp). As a consequence, Mp - T
contains the relevant nomic impossibilities (or virtual possibilities). As a rule,
T is unknown, or even the great unknown, and is hence the target of theorydirected research in the domain. However, until further notice, we will deal
with logical questions and assume that T is known. 7
T is not only Mp-dependent, that is, conceptually relative, T also depends
on reality through D. However, it does not represent 'the actual world', i.e.,
some actual state, situation or system, but the metaphysical notion of 'the set
of nomically possible worlds' (as far as D, seen through Mp, is concerned). For
that reason, the present type of realism may be called nomic realism in contrast
to actual(ist) realism, i.e., realism about the actual world. Note that from this
nomic perspective, an experiment is the realization of a nomic possibility,
having a unique true description within the conceptual frame.
We will also assume that T can be characterized in one way or another in
terms of the given vocabulary (without or with higher order constructions).
This stronger postulate, which might be called the 'characterizability or nomological postulate'S may well be false. To account for this possibility one has to
decide in advance whether one wants to aim at the smallest characterizable
superset of T or at the largest characterizable subset of T. Though the first
alternative seems in general more plausible than the second, it is in both cases
not guaranteed that the suggested limit exists. If there is no limit, it always
remains possible to improve the approximation of the supposedly uncharacterizable T. Of course, the theories with which we try to characterize T determine,

148

TRUTHLIKENESS AND TRUTH APPROXIMATION

by definition, characterizable sets of conceptual possibilities. In the following


we will pay no further attention to violations of the characterizability postulate.
As suggested, one may equate 'nomically possible' with 'physically possible'.
However, in several contexts, it makes good sense to distinguish kinds of nomic
possibilities, as e.g., suggested by the following sequence of 'lower' to 'higher'
levels: the physical, chemical, biological, psychological, cultural-socio-economical level. Being nomically possible at a higher level then implies being nomically
possible at a lower level, but not the converse. Another kind of nomic possibilities concerns the idea of nomically possible states of an artifact, assuming that
it remains intact, as illustrated by the electrical circuit. As we all know from
daily experience, the intactness assumption amounts to a severe restriction to
the physically possible, including broken, states of an artifact.
Let us now make more precise the articulation of (general) hypotheses and
theories, as already suggested in the introduction to the previous chapter, to
begin with theories. A (general) theory is any combination of a subset X of
Mp and the (strong) claim that T is equal to X, and will be briefly indicated
by 'theory X' or just 'X'. Members of X are called models of theory X. Theory
X might be called true or false when its claim 'T = X' is true or false, respectively. However, we will not do so in general, but only use one of the consequences of that definition. According to this definition there would be only
one true theory, viz., theory T itself. For this reason, T may be called 'the true
theory' or 'the nomic truth' or even 'the truth'.
Restricting attention also to general hypotheses, such a hypothesis is defined
as the combination of a subset X of Mp and the (weak) claim that T is a subset
of X (i.e., all nomic possibilities satisfy the conditions of X). Hypothesis X is
true or false when its claim 'T <;; X' is true or false, respectively. Members of
X are now also called models of hypothesis X. A true (general) hypothesis is
also called a law.
If we speak of a true or a false theory, without further qualification, we mean
a theory which is true or false 'as a hypothesis', despite the fact that the strong
claim apparently is supposed to be associated with it. If Y is a subset of X it
is easy to check that the claim of hypothesis Y implies the claim of hypothesis
X. In that case, hypothesis Y is also said to imply hypothesis X, in agreement
with the standard model-theoretic usage according to which a statement Sl
logically implies the statement S2 (or S2 is a logical consequence of Sl) if and
only if the models of Sl form a subset of those of S2. If hypothesis Y implies
hypothesis X, i.e., Y is a subset of X, it will not only be said that hypothesis
Y(X) is stronger (weaker) than hypothesis X( Y), but also that theory Y(X) is
stronger (weaker) than theory X(Y), although the full claims of these theories
are mutually incompatible as soon as Y is a proper subset of X. Now it is
evident that the true theory T is the strongest true hypothesis or strongest law,
for T is stronger than any other true (general) hypothesis.
If hypothesis or theory Y entails a true hypothesis X, it may be said to
explain it, where explanation then is used in the liberal realist sense of entailment, irrespective of whether Y is known to be false. However, T, being the

TRUTHLIKENESS AND TRUTH APPROXIMATION

149

strongest law, explains all other laws even in the non-liberal realist sense. In
the following we will use the neutral entailment terminology. Moreover, all
hypotheses and theories are supposed to deal with the same domain. 9
A law of nature is traditionally understood to be a true impossibility statement, e.g., a perpetuum mobile is (nomically) impossible. It is important to
note that a hypothesis X in our sense is in fact a domain-relative version of
such a, potentially true, impossibility claim, viz., phenomena that, given Mp,
would have to be represented by a member of Mp - X are claimed to be nomic
impossibilities. In case hypothesis X is true, i.e., when we may speak of law X,
this claim is true. Note further that, when hypothesis X, whether true or false,
fails to recognize a nomic impossibility x as such, i.e., when x belongs to X - T,
this means that it fails to entail the law to the effect that x and similar
conceptual possibilities are nomic impossibilities.
It is clear that the notions of hypothesis and theory are here used in a specific
sense, not related to theoretical terms. Here, both a hypothesis and a theory
mayor may not have theoretical terms of itself, i.e., they may both be proper
or merely observational hypotheses and theories in the sense of using or not
using theoretical terms. In Chapter 9 we will consider stratified hypotheses and
theories. They are designed for observational hypotheses and theories on the
observational level and proper hypotheses and theories using the theoretical
level. The present chapter essentially treats all non-logico-mathematical terms
as if they were observational or, alternatively, as if we are in an omniscient
position as far as the application of terms is concerned. The crucial difference
between a hypothesis and a theory in the sense used here is that the claim of
hypothesis X is just one of the two claims of the corresponding theory X.
The 'problem state' of theory X is depicted in Figure 7.3. Theory X can
make two kinds of mistakes. The members of T - X, if any, are called (T-)internal

Mp(V)

TnX

Figure 7.3. The problem state of X: its matches Tn X (internal) and Tn X (external), and mistakes
T-X and X - T

150

TRUTH LIKENESS AND TRUTH APPROXIMATION

mistakes: nomic possibilities that are excluded by X; in other words, they are
the realizable counter-examples, or wrongly missing models of X . On the other
hand, the members of X - T, if any, may be called (T-)external mistakes of X:
nomic impossibilities that are not excluded by X, that is, wrongly admitted
models, also called mistaken models. Note that the external mistakes form, by
definition, kinds of counter-examples that cannot be realized. The set of all
mistakes of theory X is the union of these two sets T - X and X - T, which is
technically called the symmetric difference between X and T, indicated by X L'I T.
A theory not only makes mistakes, but also makes matches. Tn X represents
the (T-)internal matches: nomic possibilities that are recognized as such by X
or, in other words, they are the realizable examples of X or the rightly admitted
or correct models of X. Let X indicate the complement of X with respect to
Mp, i.e., Mp - X. The members of Tn X are the (T-)external matches: nomic
impossibilities that are rightly excluded by X. The external matches are kinds
of examples that cannot be realized. The union of the two sets of matches of
theory X is of course equal to the complement of the total set of mistakes, viz.,
Mp - (T L'I X).
Note that all mistakes and matches are ultimate in the sense that they need
not have been established. Established mistakes and matches will later come
into the picture.
Note that the terminology of (T-)internal and (T-)external matches and
mistakes is not laden with the Nomic Postulate. It leaves room for another
type of target set T.
Table 7.1. presents the matches and mistakes of a theory in a matrix.
Table 7.1. The matrix of matches and mistakes of theory X

Internal
External
Total (union)

Mistakes

Matches

Total (union)

T-X
X-T
T t1 X

TnX

Mp-(T t1 X)

Mp

fnx

7.2.2. The basic definition


From now on X, Y, etc. refer to theories or just to the sets X and Y, depending
on the context. When the hypotheses X and Yare intended it will be explicitly
mentioned.
The basic definition of truthlikeness states that theory Y is at least as similar
(close) to the truth (T) as theory X, or informally Y is more truthlike than X,
indicated by MTL(X, Y, T), and briefly paraphrased by 'Y is at least as truthlike
as X', if the following two conditions are satisfied:

TRUTHLIKENESS AND TRUTH APPROXIMATION

(Bi)
(Bii)

151

Y - T is subset of X - T (i.e., X nTis subset of Yn T)


T - Y is subset of T - X (i.e., X nTis subset of Y n T)

To begin with the second clause (Bii), it says that the internal mistakes of X
include those of Y, or equivalently, that all internal matches (correct models)
of X are internal matches (correct models) of Y. Hence (Bii) can be read as a
claim about all nomic possibilities: for all nomic possibilities x, if x is not a
model of Y then it is not a model of X or, equivalently, if x is a model of X
then it is a model of Y. For this reason (Bii) will be called the (T-)internal
clause. The first clause (Bi) states that the external mistakes (mistaken models)
of X include those of Y, or equivalently, that all external matches of X are
external matches of Y. Hence (Bi) can be read as a claim about all nomic
impossibilities, for which reason it is called the (T-)external clause. Note that
(Bi) and (Bii) together are equivalent to the claim that the mistakes of Y
(Y L1 T) form a subset of those of X (X L1 T).
By MTL + (X, Y, T) we indicate that Y is more truthlike than X 'in the strict
sense' that the mistakes of Y form a proper subset of those of X, that is, Y is
at least as truthlike as X, but not the reverse. If the reverse holds as well, then
they are 'equally truthlike'. If none is at least as truthlike as the other, they are
'incomparable' (in truthlikeness). Here and later the strong verbal expressions
'closer to' or 'more similar to' will however also be used to refer to the
corresponding weak notion. When the strict notion is meant, it will be explicitly
stated by adding 'in the strict sense'.10 The strict version is depicted in Figure 7.4.
(in which Mp is not explicitly indicated). Note that Figure 7.4., including the
conditions, is formally equivalent to Figure 7.2.

Mp(V)

Figure 7.4. Nomic truth likeness of theories: Y is more similar (close) to T than X: 1 and 5 empty,
and 3 or 7 non-empty

At first sight one might think that it is plausible to add, as a kind of


preliminary condition, that Y should be in strength between X and T, to
be called the strong boundary condition (hence for the finite case:

152

TRUTHLIKENESS AND TRUTH APPROXIMATION

min(IXI, 1TI) ~I YI ~ max(IXI,1 TJ). However, when X is weaker than T this


would exclude all cases of further weakening, followed by strengthening l l ; when
X is stronger than T it would exclude all cases of further strengthening, followed
by weakening. The given definition does not exclude this, but puts natural
limits to weakening and strengthening: no weakening beyond the union of X
and T, and no strengthening beyond the intersection of X and T. In terms of
strength this amounts to 1X n TI ~ 1YI ~ 1XU TI for the finite case. In Part IV
we will see that this condition, to be called the boundary condition, implied
by the basic definition, makes much sense as an extra condition for refined
truthlikeness.
It requires some easy tabulator work to check that, for instance, in a circuit
where Po ~ PI &P2&P3 is the true theory, Po ~ PI &P2 is according to the given
definition closer to the truth than Po ~ PI' in agreement with what we intuitively
would like to have.
Some equivalent formulations of MTL are instructive. Very useful are (the
numbers refer to Figure 7.4.):
(Bi)'
(Bii)'

Y - X U T = 5 is empty
n T - Y = 1 is empty

The first condition now reveals that Y makes no extra external mistakes, and
the second discloses that Y makes no extra internal mistakes.
It is important to note that improving a theory X in the sense of finding a
theory Y such that MTL + (X, Y, T) is not an easy task, due to the fact that
both components are counteracting. This can be nicely illustrated by considering e.g., just weakening of theory X: if Y is weaker than X (which was
defined as: X S Y), then Y contains at least all nomic possibilities contained
by X, but X misses at least all nomic impossibilities Y misses. Of course,
strengthening a theory leads to the opposite tension.
The external clause (Bi) has a very illuminating equivalent version on the
level of sets of sets, to be called the second level (of sets), as opposed to the
first level of sets. Let Q(X) indicate the set of supersets of X, that is, the set of
subsets of Mp which include X, which might be called the co-powerset of XI2.
Q(X) represents the set of hypotheses following from theory X, for, as is easy
to check, all hypothesis-claims associated with members of Q(X) follow from
the theory-claim of X. Recall that a true hypothesis corresponds to a set
including T, hence to a member of Q(T). Now it is not difficult to prove the
'bridge-theorem' that the external clause (Bi), i.e., Y - Ts X - T, is equivalent
to the condition that all true consequences of theory X are also (true) consequences of theory Y, i.e., formally:
(BiC)

Q(X) n Q(T)

Q(Y) n Q(T)

To prove this, start from (BiC) and note first that Q(X) n Q(T) = Q(X U T).
Hence (BiC) is equivalent to the claim: Q(X U T) is a subset of Q(YU T). This
is on the first level equivalent to the claim: Y U T is a subset of X U T. In its

TRUTH LIKENESS AND TRUTH APPROXIMATION

153

turn, this is trivially equivalent to the claim that Y - T is a subset of X - T,


i.e., (Bi). It is important to note that we have not claimed that X - T and
Q(X) n Q(T) are equivalent, but only that these sets and their Y-versions lead
to equivalent comparative statements.
In view of our interpretation of T, Q(X) n Q(T) amounts to the set of laws
entailed by X, to be called the nomological or, simply, law matches. Q(T) - Q(X)
amounts to the set of laws not entailed by X, to be called the law mistakes.
Hence (BiC) reflects our intuitions concerning improving theories in an ultimate
sense: ideally speaking, a new theory should entail at least all laws (established
or not) entailed by the old theory. Hence (Bi) can be called the law version of
the external clause, or just the law clause. The difficulty of improving a theory
can now be expressed by noting that, if Y is weaker than X, then Y contains
at least all nomic possibilities which X contains, but X entails at least all laws
entailed by Y. Again, strengthening a theory leads to the opposite tension. Of
course, we might call the members of Q(X) - Q(T) 'non-law' mistakes of X
and members of Q(X) n Q(T) 'non-law' matches, although these notions do not
seem to correspond to notions of scientific common sense.
Note that much of the foregoing was formulated in terms of (un-)nomic
possibilities and laws, deriving from the leading interpretation of T as the set
of nomic possibilities. The main points can be rephrased in an appropriate
terminology when T is the target set in some other sense. E.g., a member of
Q(X) n Q(T) remains a true consequence of X, that is, a consequence of X
that is true with respect to the set of target possibilities T. Table 7.2. extends
Table 7.1. to all categories of mistakes and matches of theory X on both levels,
with some new general names between brackets, which will be used in Chapter 8.
Table 7.2. Epistemological categories of theory X
Mistakes/mismatches

Model level

internal

external

Consequence
level

non-law

law

Matches

T-X

xnT

internal mistakes
(wrongly missing models)

internal matches
(correct models)

X-T

xnf

external mistakes
(mistaken models)

external matches
(rightly missing models)

Q(X)-Q(T)

non-law mistakes
(false consequences)

Q(X)nQ(T)
non-law matches
(false non-consequences)

Q(T)- Q(X)

Q(X)nQ(T)

law mistakes
(true non-consequences)

law matches
(true consequences)

7.2.3. Some formal properties


Any definition of the binary relation of truthlikeness between theories X and
Y can be seen as a special case of a ternary relation of theorylikeness

154

TRUTHLIKENESS AND TRUTH APPROXIMATION

MTL(X, Y, Z) between theories X, Y and Z by replacing the fixed true theory


T by the variable theory Z . Then we get the general definition: MTL(X, Y, Z)
= defZ - Y s; Z - X and Y - Z s; X - Z, with plausible equivalent
formulations.
MTL has several interesting properties. We list the main ones, starting with
those which together imply the possibility of convergence to the truth is
possible.

reflexivity:
antisymmetry:
symmetry:
transitivity:

MTL(X, X , Y) (left) MTL(X, Y, Y) (right)


MTL(X, Y, Z) and MTL(Y, X, Z) imply X = Y (left)
MTL(X, Y, Z) and MTL(X, Z, Y) imply Y = Z (right)
MTL(X, Y, Z) implies MTL(Z, Y, X) (central)
e.g., left: if MTL(W, X, Z) and MTL(X, Y, Z)
then MTL(W, Y, Z)

Hence, from left reflexivity, left antisymmetry and left transitivity it follows
that MTL(X, Y, Z) is for fixed Z a partial ordering of theories. As a consequence, a sequence of theories converging to the truth is perfectly possible.
Some other interesting properties are:

centeredness:
centering:
specularity:
concentricity:
context neutrality:

MTL(X,X,X)
if MTL(X, Y, X) then X = Y
if MTL(X, Y, Z) then MTL(X, f, Z)
if X s; Y s; Z then MTL (X, Y, Z) and
MTL(Z, Y, X)
if X, Yand Z are subsets of Mp and Mp is itself
a subset of a larger set of conceptual possibilities
Mp*, then MTL(X, Y, Z) implies MTL*(X, Y, Z)

7.3 . ACTUAL AND NOMIC TRUTH APPROXIMATION

Introduction
Thus far, we have been studying truth approximation from the ideal perspective
that we know the true theory or true description. Of course, in practice, this
is exactly what we do not know. In this section we will study truth approximation from the perspective of observational success of descriptions and theories,
and methodological rules based on success differences. We start with (propositional) descriptions. Then the idea of problems and successes of theories is
explicated in structuralist terms. This is followed by the Success Theorem,
according to which success dominance of a theory can be explained by the
hypothesis that it is closer to the truth than its competitor(s). This leads to the
conclusion that the evaluation methodology explicated in Part II, governed by
the rule of success, is functional for truth approximation.

TRUTHLIKENESS AND TRUTH APPROXIMATION

155

7.3.1 . Success of propositional descriptions

Let us assume a propositional frame and suppose that we do not know the
(relevant whole) actual truth. At successive stages, an increasing part of the
truth may be assumed to become known, in some way or other, in terms of
which the success and the problems of a certain description will have to be
defined, as well as the comparative judgment that one description is more
successful than another. The latter may guide the choice between descriptions.
Let us also suppose that 'actual descriptions' are not presented in a piecemeal
way, but in total, that is to say, as far as the relevant elementary propositions
are concerned. Think of the descriptions resulting from one or more researchers
or mechanical description devices. To evaluate such descriptions we assume
the idealization of an infallible metaposition.
Let p and n indicate the (mutually exclusive) sets of elementary propositions
of which it has been established at a certain time 13 that their truth-value is
'true' (positive) or 'false' (negative), respectively. Together p and n are called
the available data at a certain time. In the course of time both sets can of
course only grow, not shrink. Recall that descriptions were represented by the
set of un-negated elementary propositions. Consider now (description) x. The
union of p n x and n n (EP - x) = n - x indicates the set of elementary propositions about which it has been established that x makes correct elementary
claims, called the established matches, and the union of p - x and
n - (EP - x) = n n x indicates the set of elementary propositions about which
it has been established that x makes elementary mistakes, called the established
mistakes. Together they constitute the descriptive success of x .
Now it is plausible to define that (description) y is at least as successful at a
certain time as x if and only if the set of established matches of y includes that
of x, or, equivalently, if and only if the set of established mistakes of y is a
proper subset of that of x. Or, formally, using the latter version, description y
is at least as successful as description x if and only if
(Di)-n n n y is a subset of n n x
(Dii)-p p - y is a subset of p - x
The strong version for 'more successful' is obtained by requiring at least one
proper subset relation. The strong version is depicted in Figure 7.5., where the
actual truth t is indicated by an interrupted line, to stress that it is unknown.
Notice that the 'success' of a description is completely symmetric with respect
to true and false elementary propositions. When we can establish that such a
proposition is true when it is true, we can also establish that it is false when it
is false, and vice versa. For this reason we call the present notion of success
symmetric. It is interesting to note here already that the primary notion of
success of theories is an asymmetric one, essentially due to the semantic
impossibility of realizing nomic impossibilities.
The plausible methodological rule prescribes, of course, to favor the more

156

TRUTHLIKENESS AND TRUTH APPROXIMATION

EP

n
x

/' t
Figure 7.5. Increasing success of propositional descriptions: y is more successful than x, relative to
data pin: sets 1.1 and 5.1 empty and sets 3.1 or 7.1 non-empty

(most) successful description, if any. That this rule is functional for truth
approximation is easy to see. The hypothesis that the more (most) successful
description is closer to the actual truth can explain its success dominance, and,
in view of the extra success, it is already impossible that the alternative(s) is
(are) closer to the actual truth.
To be sure, all claims presuppose the auxiliary hypothesis that p and n are
correct, which amounts to the assumption that p is a subset of t and n of EP - t.
It should be noted, however, that even if these conditions are satisfied, there
is no guarantee of truth approximation. Additional evidence may destroy the
success dominance, although a complete turn of the tables is impossible.
It will be clear that similar definitions of successfulness of structure description of a non-propositional nature can be construed, and that, moreover,
refinements are possible to include the success of 'partial' descriptions.
7.3.2. Nomic explication of methodological categories

Recall that as far as theories are concerned, we have dealt up to now with the
logical problem of defining nomic truthlikeness, assuming that T, the set of
nomic possibilities, is at our disposal. In actual scientific practice we don't
know T; it is the target of our theoretical and experimental efforts. Before we
turn our attention to methodological rules guiding these efforts, it is fruitful to
explicate the idea that one theory is more successful than another and to show
that this can be explained by the hypothesis that the first theory is more similar
to the truth than the second.
Recall that in the propositional form of the actual case, the (actual) truth

TRUTH LIKENESS AND TRUTH APPROXIMATION

157

could be gradually established by establishing the truth-value of more and


more elementary propositions in a correct way. This process was seen to be
essentially symmetric between true and false propositions. The important
difference between the actual and the nomic case arises from the fact that
coming to know part of the nomic truth is normally something asymmetrical:
you can establish that a certain conceptual possibility is nomically possible by
experimentally realizing this possibility, but you cannot establish that a certain
conceptual possibility is nomically impossible in a direct way, for you cannot
realize nomic impossibilities. The standard (partially) indirect way to circumvent this problem is by establishing on the one hand nomic possibilities by
realizing them, and establishing on their basis (observational) laws on the
other. As we have seen in the preceding chapter, this is precisely what the
separate HD-evaluation of theories amounts to. That is, theories are evaluated
in terms of their capacity to respect the realized possibilities, i.e., to avoid
counter-examples, and to entail the observational laws, i.e., to have general
successes. 14
The problems and successes of a theory will have to be expressed in terms
of the data to be accounted for. The data at a certain moment t (not to be
confused with the actual truth!) can be represented as follows. Let R(t) indicate
the set of realized possibilities up to t, i.e., the accepted instances 15 , which have
to be admitted. Note that there may be more than one realized possibility at
the same time, before or at t, with plausible restrictions for overlapping domains.
Up to t there will also be some accepted general hypotheses, the (explicitly)
accepted laws, which have to be accounted for. On the basis of them, the
strongest accepted law to be accounted for is the general hypothesis S(t)
associated with the intersection of the sets constituting the accepted laws. Of
course, S(t) is, via the laws constituting it, in some way or other based on R(t);
minimally we may assume that R(t) is not in conflict with S(t), that is, R(t) is
a subset of S(t). In the following, however, we shall need the much stronger
correct data (CD-)hypothesis R(t) T S(t), guaranteeing that R(t) only contains nomic possibilities, and that hypothesis S(t) only excludes nomic impossibilities. It is thus true as a hypothesis and may hence rightly be called a law.
Henceforth we will assume the CD-hypothesis. R(t) may now be called the set
of established nomic possibilities, and Q(S(t)) the set of established laws and S(t)
the strongest established law.
It is plausible to call R( t) - X the set of established internal mistakes of
X and X n R(t) the set of established internal matches of X. Moreover,
Q(S(t)) - Q(X) may be called the set of established law mistakes, and
Q(X) n Q(S(t)) the set of established law matches. In a similar way we can

identify established external mistakes and matches, as well as established nonlaw mistakes and matches (see Table 7.2.)
Assuming R(t) and S(t), it is now easy to give explications of the notions of
individual problems and general successes of a theory X at time t we met in
Part II concerning the HD-evaluation of theories. The set of individual problems,

158

TRUTHLIKENESS AND TRUTH APPROXIMATION

of X at t, is equated with the established internal mistakes. However, it is


important to note that the set of established internal matches not only includes
the set of individual successes of X at t, but also the neutral instances of X at
t. Hence, an established internal match only refers to an established nomic
possibility which is compatible with the theory, whereas an individual success
refers to a nomic possibility which is partially derivable from it.
Similarly, the set of general successes, of X at t, is equated with the set of
established law matches. Again it is important to note that the set of established
law mistakes not only includes the set of general problems of X at t, but also
the set of established laws which are neutral for X, i.e., established laws that
are compatible with X, but not entailed by X, the neutral laws of X, for short.
Hence, an established law mistake just refers to an established law which is
not entailed by X, but which mayor may not be compatible with X, whereas
a general problem refers to an established law which is incompatible with X.
Table 7.3. gives a survey of the explication of methodological categories, again
with some alternative terms, which are not laden with the Nomic Postulate,
between brackets and which will be clarified in the next chapter.
Table 7.3. Methodological categories of theory X

Established mistakes
Model level

internal

Established matches

R-X

xnR

est. internal mistake


(counter-example)

est. internal match


(example)

individual problem

individual success
or neutral instance

external

X-S

est. external mistake


(est. mistaken model)

Consequence level

non-law

law

xns

est. external match


(est. rightly missing model)

Q(X)-Q(S)

Q(X)nQ(S)

est. non-law mistake


(est. false consequence)

est. non-law match


(est. false non-consequence)

Q(S)-Q(X)

Q(X)nQ(S)

est. law mistake


(est. true non-consequence)

est. law match


(est. true consequence)

general problem
or neutral law

general success

7.3.3. Success Theorem


For comparative judgements of the success of theories we will now explicate
the individual problem clause of Chapter 6, in terms of established nomic
possibilities, i.e., R(t). Theory Y is (T-)internally or instantially at least as
successful as X if and only if the individual problems, i.e., the established

TRUTHLIKENESS AND TRUTH APPROXIMATION

159

internal mistakes, of Y form a subset of those of X, that is, Y has no extra


individual problems, i.e., established internal mistakes, or, equivalently, the
established internal matches of X form a subset of those of Y. Formally:
(Bii)-R

R(t) - Y ; R(t) - X

(=)

XnR(t)- Y= 1.1 =4J


X

(=)

n R(t) ; yn R(t)

This explication of the individual problem clause can be read as ranging over
the established nomic possibilities: for all z in R( t), if z is a model of X then it
is a model of Y. Hence it will be called the established internal success clause
or the instantial (success) clause.
On the other side, for the general success clause we have two options for
explication, one on the first and one on the second level, leading again to two
equivalent comparative statements. To begin with the second level, the level of
consequences, theory Y is (T-)externally or explanatorily at least as successful
as X if and only if the general successes, i.e., the established law matches, of X
form a subset of those of Y, that is, X has no extra general successes, i.e.,
established law matches, or, equivalently, the established law mistakes of Y
form a subset of those of X. Formally we get
(BiC)-S Q(X)nQ(s(t;Q(Y)nQ(S(t
Q(X)nQ(S(t-Q(Y)=4J

(=)

(=)

Q(S(t - Q(Y) ; Q(S(t - Q(X)

On the first level, the level of sets, this is equivalent to the condition that the
established external matches of X form a subset of those of Y, that is, X has
no extra established external matches, that is, Y has no extra established
external mistakes, or, equivalently, the established external mistakes of Y form
a subset of those of X. Formally,
(Bi)-S

S(t)nX;S(t)n

(=)

Y-XUS(t)=5.l=4J

(=)

Y-S(t);X -S(t)

The same line of formal reasoning as in the case of the equivalence of (Bi) and
(BiC) leads to the conclusion that (Bi)-S and (BiC)-S are equivalent. The first
level explication of the general success clause, i.e., (Bi)-S, can be read as ranging
over the established nomic impossibilities: for all z in S(t), if z is not a model
of X then it is not a model of Y. Hence it will be called the established external
success clause. Its second level explication, (BiC)-S, ranges over established
laws and may hence also be called explanatory (success) clause.

160

TRUTHLIKENESS AND TRUTH APPROXIMATION

The conjunction of the established internal and law clause forms the general
definition of the statement that one theory is at a certain time at least as
successful as another, relative to the data R(t)jS(t). More informally we say
that Y is more successful than X , relative to R(t)jS(t), indicated by
MSF(X, Y, R(t)/S(t)) . We obtain the strict version MSF + (X , Y, R(t)/S(t))
when in at least one of the cases proper subsets are concerned. This strict
version is depicted on the first level in Figure 7.6. (in which T is indicated by
an interrupted circle to stress that it is unknown).

Mp(V)
y

Figure 7.6. Increasing success of theories: Y is more successful than X, relative to data R(tl/S(t):
sets 1.1 and 5.1 empty, sets 3.1 or 7.1 non-empty

Note that Figures 7.5. and 7.6. are formally equivalent, when we compare
n(t) with Mp - S(t) . However, in view of the fundamental methodologically
different nature of these sets, this is only a formal similarity.
Now it is easy to prove the following crucial theorem:

Success Theorem:
If theory Y is at least as similar to the nomic truth T as X and if the
data are correct then Y (always) remains at least as successful as X (i.e.,
if MTL(X, Y, T) and CD(R(t) , S(t) then MSF(X, Y, R(t}/S(t)))
From this theorem immediately follows the corollary that success dominance
of Y over X in the sense that Y is at least as successful as X can be derived
from, and hence explained by, the following hypotheses: the truth approximation
(TA -)hypothesis, Y is at least as similar to the nomic truth T as X, and the
auxiliary correct data (CD- )hypothesis 16.
All notions in the theorem and the corollary have been explicated, and the
proof is, on the first level, only a matter of elementary set-theoretical manipulation, as will be clear from the following presentation of the theorem as an
argument:

TRUTH LIKENESS AND TRUTH APPROXIMATION

(Bi)

Y-Tr;;.X-T

(Bii)

161

T-yr;;.T-X

(TA-hypothesis)
Tr;;. S(t)

R(t) r;;. T

(CD-hypothesis)
(Bi)-S

Y - S(t)r;;.X-S(t)

(Bii)-R

R(t)-yr;;.R(t)-X

(success dominance)
As a rule, a new theory will introduce some new individual problems and/or
will not include all general successes of the former theory. The idea is that the
relative merits can now be explained on the basis of a detailed analysis of the
relative 'position' to the truth. However, for such cases a general theorem is
obviously not possible.
The Success Theorem makes clear that and how empirical progress is possible
within a conceptual frame Mp for a domain D. It is important to note that the
specific TA-hypothesis presupposes the Nomic Postulate of the research program that <D, Mp) indeed generates a unique, time-independent set T of nomic
possibilities. The Nomic Postulate creates, as it were, the possibility that there
may occur theories closer to the truth than others and that if theories are more
successful than others, it may be (but need not be) for that reason. In other
words, although each specific example of empirical progress is explained on
the basis of the corresponding specific TA-hypothesis, assuming the
CD-hypothesis, the general possibility of empirical progress within a research
program is explained on the basis of the Nomic Postulate associated with

<D, Mp).
Two successive generalizations bring us to the explanation of the success of
the natural sciences in general. First, the Nomic Postulate is true for all possible
conceptual frames Mp with respect to the natural domain D. Second, the Nomic
Postulate is true for all frames for all natural domains. We do not claim that
these generalizations don't have exceptions. If they are true in the majority of
cases, they serve their purpose.
7.3.4. Methodological consequences
Let us return to one particular combination <D, Mp) and the corresponding
Nomic Postulate, and let us spell out some methodological consequences of
the Success Theorem. Recall the Comparative Success Hypothesis (CSH) and
the Rule of Success (RS), introduced in Chapter 6:
CSH:
RS:

Y (is and) remains more successful than X


If CSH has been sufficiently confirmed to be accepted as true,
i.e., if Y has so far proven to be more successful than X, then
eliminate X, in favor of Y, for the time being.

For an instrumentalist, CSH and RS is itself sufficiently interesting, but the

162

TRUTHLIKENESS AND TRUTH APPROXIMATION

(theory-)realist will only appreciate it for its possible relation to truth approximation, whereas the empiricist and the referential realist will have intermediate
interests.
The Success Theorem shows that RS is functional for approaching the truth
in the following sense. Assuming correct data, the theorem suggests that the
fact that 'Y has so far proven to be more successful than X' may well be the
consequence of the fact that Y is closer to the truth than X. For the theorem
enables the attachment of three conclusions to the fact that Y has so far proven
to be more successful than X; conclusions which are independent of what
exactly 'the nomic truth' is:
- first, it is still possible that Y is closer to the truth than X, a possibility
which, when conceived as hypothesis, the TA-hypothesis, according to
the Success Theorem, would explain the greater success,
- second, it is impossible that Y is further from the truth than X (and
hence X closer to the truth than Y), for otherwise, so teaches the Success
Theorem, Y could not be more successful,
- third, it is also possible that Y is neither closer nor further from the
truth than X, in which case, however, another specific explanation has
to be given for the fact that Y has so far proven to be more successful.
Hence we may conclude that, though 'so far proven to be more successful' does
not guarantee that the theory is closer to the truth, it provides good reasons
to make this plausible. And this is increasingly the case, the more the number
and variation of tests of the comparative success hypothesis increase. It is in
this sense that we interpret the claim that RS is functional for truth approximation: the longer the success dominance lasts, the more plausible that this is the
effect of being closer to the truthY We may at least informally summarize the
situation by saying that the truth is an attractor when RS is systematically
applied.
In view of the way in which the evaluation methodology is governed by RS,
this methodology is, in general, functional for truth approximation. We would
like to spell out this claim in more detail. Recall that the separate and comparative HD-evaluation of theories was functional for applying RS in the sense that
they precisely provide the ingredients for the application of RS. Recall moreover
that HD-testing of hypotheses is functional for HD-evaluation of theories
entailing them. Hence, we get a transitive sequence of functional steps for truth
approximation, depicted in Scheme 7.1.
HD-testing of hypotheses
..... separate HD-evaluation of theories
---> comparative HD-evaluation of theories
---> Rule of Success (RS)
---> Truth Approximation (TA)
Scheme 7.1. Functional steps for truth approximation

TRUTHLIKENESS AND TRUTH APPROXIMATION

163

Consequently, from the point of view of truth approximation, RS can be


justified as a prescriptive rule, and HD-testing and HD-evaluation as its drive
mechanisms. Intuitive versions of the rule and the two methods are usually
seen as the hall-mark of scientific rationality. The analysis of their truth approximating cooperation can be conceived as an explication of what many scientists
are inclined to think, and others are inclined to doubt. To be sure, the understanding is not relevant for the practice. That the practice is functional for
truth approximation may be conceived as the cunning of reason in science.
It is important to stress once more that RS does not guarantee that the more
successful theory is closer to the truth. As long as one does not dispose of
explicit knowledge of T, it is impossible to have a rule of success that can
guarantee that the more successful theory is closer to the truth. As we will see
at the end of this chapter, there is only one (near) exception to this claim:
purely inductive research.
Another way to summarize the above findings is the following. The
TA-hypothesis, claiming that one theory is at least as close to the truth as
another, is a perfect example of an empirically testable comparative hypothesis.
The Success Theorem says that TA-hypothesis implies, and hence explains, that
the first theory will always be at least as successful as the second.
In terms of an application of HD-testing, the Success Theorem amounts to
the following claim: the TA-hypothesis has the following two general comparative test implications (assuming a strong, but plausible, auxiliary hypothesis):
all general successes of X are general successes of Y
all individual problems of Yare individual problems of X
Note that these are precisely the two components of the comparative success
hypothesis (CSH). Hence, when Y is at least as successful as X, the further
HD-evaluation, i.e., the further testing of CSH, can indirectly be seen as further
HD-testing of the TA-hypothesis. When doing so, the latter hypothesis can be
falsified, or it can be used to explain newly obtained success dominance.
Note that the strong TA-hypothesis that Y is closer to the truth than X in
the strict sense trivially implies the (weak) TA-hypothesis. Of course, when Y
is more successful than X, the strong TA-hypothesis can also be further tested
by testing its test implications. Given the fact that Y has already some extra
success, it is sufficient to further test the two general test implications of the
weaker TA-hypothesis, for together with the extra success, it implies the
strong version.
As long as two theories are equally successful, it is interesting to question
what further test implications the strong TA-hypothesis has compared to the
weak one, i.e., what extra test implication follows from the additional component hypothesis. It is something like: "in the long run the one closer to the
truth will show some extra success compared to the other, provided the two
theories are not observationally equivalent".18 For this and other reasons, it is

164

TRUTHLIKENESS AND TRUTH APPROXIMATION

a weak prediction. To be precise, it is a weak prediction amongst other things


in the sense that the additional component hypothesis can not be confirmed
or falsified, but it can only be verified or disconfirmed in the sense of Chapter 2.
Recall that we noted in Chapter 6 that the application of the prescriptive
rule RS can be stimulated by several heuristic principles, viz., the principle of
separate HD-evaluation (PSE), the principle of comparative HD-evaluation
(PCE), the principle of content (PC), and, finally, the principle of dialectics (PD).
Of course, we may now conclude that all these principles belonging to the
evaluation methodology are indirectly functional for truth approximation.
The Success Theorem is not only attractive from the realist point of view, it
is also instructive for weaker epistemological positions, even for the instrumentalist. The Success Theorem implies that a theory which is closer to the truth
is also a better derivation instrument for successes, and that the truth (the true
theory) is the best derivation instrument. From this not only follows indirectly
the self-evident fact that RS and HD-evaluation are functional for approaching
the best derivation instrument, but also that the heuristic of the realist, provided
by the idea that each conceptual frame has a unique strongest true hypothesis,
i.e., the truth, may also be of help to the instrumentalist. Intermediate considerations apply both to the empiricist and to the referential realist.

7.3.5. Forward Theorem


If we assume an observational vocabulary, or, alternatively, if we assume an
omniscient point of view with respect to all terms, we can even go further.
Given the proof of the Success Theorem, the following theorem is now easy
to prove 19 :

Forward Theorem
If the vocabulary is observational and if CSH, which speaks of remaining
more successful, is true, this implies the TA-hypothesis, that is, if Y is
not closer to the nomic truth than X, (further) testing of CSH will
sooner or later lead to an extra counter-example of Y or to an extra
success of X.

In other words, for observational theories holds that 'so far proven to be more
successful' can only be explained by the TA-hypothesis (Success Theorem) or
by assuming that the comparative success hypothesis has not yet been sufficiently tested (Forward Theorem).
Note that it follows from the Success Theorem and the Forward Theorem,
that CSH is even equivalent to the TA-hypothesis for observational vocabularies, assuming in addition that all members of T can in fact be realized, and
that T can in fact be established as a true hypothesis. This equivalence directly
follows from the fact that T contains all possibilities that can possibly be
realized and is the strongest (observational!) law that can possibly be
established.

TRUTHLIKENESS AND TRUTH APPROXIMATION

165

In Chapter 9 we will elaborate the relevance of the Forward Theorem for


the case that vocabulary contains theoretical terms. The theorem can then be
applied to the observational sub-vocabulary, leading to the conclusion that the
hypothesis of being closer to the observational nomic truth, i.e., the nomic
truth of the observational sub-vocabulary, is equivalent to CSH. From this
equivalence it follows that the instrumentalist cognitive aim of the best derivation instrument coincides in principle with the (most far-reaching) aim of the
empiricist, viz., observational truth approximation.
7.4 . SURVEY OF BIFURCATIONS

In Section 7.1. we introduced the idea of comparative actual truthlikeness,


defined the trivial form and gave some examples of non-trivial form. Then we
introduced in Section 7.2. the so-called basic definition of comparative nomic
truthlikeness. It is clear that this definition is at least informally speaking based
on the corresponding trivial notion of actual truthlikeness: every mistake of a
theory is counted as equally bad. In other words, a theory cannot come closer
to the true theory by just replacing some mistakes by lesser mistakes. This
feature of the basic definition explains why it is not easy to give real-life
examples of basic truthlikeness (and successfulness for that matter), since
improvement of scientific theories usually is a matter of improving mistaken
models, e.g., by the theory development strategy of idealization and
concretization.
In Chapter 10 we will show how a refined version of nomic truthlikeness,
including concretization and corresponding versions of successfulness, can be
based on non-trivial actual truthlikeness, with the basic definition as extreme
case. Hence, in that chapter the approach will become more realistic for
scientific practice. In Chapter 11 we will illustrate this by two examples.
Besides qualitative or, more specifically, comparative notions, trivial and
non-trivial quantitative definitions of actual truthlikeness can be given, as well
as basic and refined quantitative definitions of nomic truthlikeness, which are
related in a similar way. They will be given in Chapter 12, together with
corresponding quantitative notions of success. However, such quantitative
notions do not seem to reflect scientific practice.
Scheme 7.2. gives a survey of all mentioned bifurcations. A bar indicates that
the relevant notion of nomic truthlikeness is based on the notion of actual
truthlikeness above.
The actual and the nomic truth will be called (the two) types of 'the truth'
and, similarly, actual and nomic truthlikeness and truth approximation will be
called (the two) types of truthlikeness and truth approximation.
In this survey, the bifurcation with the (relative) distinction between observational and theoretical structures has not yet been included. In particular, for
nomic truthlikeness this distinction is rather important, as will be elaborated
in Chapter 9. Applying the Nomic Postulate, the epistemological stratification

166

TRUTHLIKENESS AND TRUTH APPROXIMATION

nomic truthlikeness
comparative
basic

quantitative
refined

trivial

non-trivial

basic

refined

trivial

non-trivial

comparative

quantitative
actual truthlikeness

Scheme 7.2. Bifurcations of (unstratified) truthlikeness

gives rise to three variants of 'the nomic truth', and hence of nomic truthlikeness
and nomic truth approximation: the observational truth, the theoretical truth,
and, in a sense in between, the substantial truth. Here 'the observational truth'
stands for the strongest true claim that can be made with the observation terms
of the vocabulary about all nomic possibilities, whereas 'the theoretical truth'
amounts to the strongest true claim that can be made with the (observational
and) theoretical terms of the vocabulary about all nomic possibilities. It turns
out to be possible to define, on the basis of the theoretical truth, which
theoretical terms really refer and which do not, to be summarized as 'the
referential truth'. This referential truth gives rise to the third kind of nomic
truth, the one in between, viz., 'the substantial truth', defined as the strongest
true claim in observational and referring theoretical terms about all nomic
possibilities. In Chapter 9 we will present the basic versions of stratified truth
approximation, and in Chapter to, after the refinement of the basic definition
for unstratified theories, we will deal with stratified versions of refined truth
approximation.
The next chapter, viz., Chapter 8, will still be restricted to comparative actual
and basic nomic truthlikeness, and will shed more light on the foundations of
these definitions and their usefulness for explicating the idea of a correspondence of truth and several intuitions governing dialectical reasoning.
In all following chapters we will bring in matters of success and methodology,
whenever relevant. We have already done this in the previous section for some
fundamental aspects of the given definitions. In the final section of this chapter
we add a couple of additional logical and methodological consequences.
7.5. NOVEL FACTS, CRUCIAL EXPERIMENTS , INFERENCE TO
THE BEST EXPLANATION, AND DESCRIPTIVE RESEARCH
PROGRAMS

Introduction
In Section 7.3. we have already drawn the conclusion from the Success Theorem
and the Forward Theorem that RS, and hence the evaluation methodology,
was functional for truth approximation. And we may add that the unrestricted

TRUTH LIKENESS AND TRUTH APPROXIMATION

167

version of RS is also efficient in approaching that goal, as all its components


are relevant for that purpose. Moreover, as already anticipated at the end of
the previous chapter, the falsificationist combination of the restricted version
of RS, rRS, i.e., RS restricted to unfalsified theories, and the Rule of Elimination
(RE), though also functional for truth approximation, due to rRS, is clearly
less efficient, due to RE. The reason is that RE renounces a theory simply
because it is false, whereas the falsity of a theory is not at all relevant for its
distance from the truth. Note that this conclusion holds irrespective of the
possible complications that might arise due to epistemological stratification of
the theories. In sum, the evaluation methodology is superior to the falsificationist methodology with regard to truth approximation. Hence, the standard
explanations of the fact that scientists do not attribute a dramatic impact to
falsifications, in terms of defensible dogmatic strategies, or in terms of social
factors, are superfluous. The non-falsificationist practice can be justified in
terms of truth approximation.
In sum, we saw in the previous chapter that, in contrast to the falsificationist
methodology, the evaluation methodology does not degrade false theories from
further evaluation and competition. From the truth approximation perspective
developed in this chapter, it is clear that the evaluation methodology is, precisely
for this reason, the most plausible, for the question of whether a theory is true
or false is essentially inessential. It is tempting to summarize the main reason
for this latter claim by stating that a false theory may well be closer to the
truth than a true theory. However, it should be conceded that neither basic
nor stratified qualitative formal explications of closer-to-the-truth (Chapter 9)
permit such a judgement, nor the converse judgment that a true, but not the
strongest, theory may be closer to the truth than a false one. This is due simply
to the strict character of comparative explications, leading to partial orderings
of theories. However, it is evident that basic and refined quantitative explications based on numbers or distance functions will make such judgements
perfectly possible. Be that as it may, the restriction to unfalsified theories in
the falsificationist approach to theory evaluation and selection is indefensible
from the truth approximation point of view. Note that this objection also
applies to all standard probabilistic approaches, notably the Bayesian. They
are falsificationist by virtue of downgrading all falsified theories on the uniform
level of zero posterior probability, and putting them aside from further evaluation, let alone from selection.
We have also anticipated in the previous chapter already the conclusion that
the evaluation methodology can be seen as an explication of Lakatos' sophisticated falsification ism, leaving room for dogmatic, research program strategies.
We have shown in Section 7.3. that the evaluation methodology can be given
a justification in the basic terms of truth approximation. For the specification
to a Lakatosian dogmatic strategy it is also easy to show its functionality for
truth approximation in basic terms. Let C indicate the hard core hypothesis.
Let H* be an improved version of the auxiliary hypothesis H, in the sense that

168

TRUTHLIKENESS AND TRUTH APPROXIMATION

C&H* is more successful than C&H. This success difference can be explained
by assuming that C is true as a dogma and that H* on its own, however far
it may be from the truth, is nevertheless closer to the truth than H on its own.
For under these assumptions, in structuralist terms, it follows immediately that
C&H* is also closer to the truth than C&H. Of course, using the Success
Theorem, this consequence explains the success difference.
There remain the diverging point about novel facts and the interesting
question concerning to what extent Lakatos' analysis of crucial experiments
can be upheld. We will first evaluate the emphasis of Popper and Lakatos put
on novel facts from the truth approximation perspective. Then it will be shown
that the Lakatos' relativization of crucial experiments directly follows from
that perspective.
We will close this chapter by considering the ideas of 'inference to the best
explanation' and 'descriptive, in particular, inductive research programs' in the
light of truth approximation.
7.5.1. Novel facts and ad hoc repairs

Suppose that our favorite theory has been falsified. Now it is possible that a wellconceived change of the theory leads to a new theory which is not falsified by the
counter-example of the old. As Popper has stressed, in scientific practice it is
considered to be very important that such a new theory not only avoids the
problem of the old one, in which case it is just an ad hoc repair, but that it also
leads to new test implications which could not be derived from the old theory and
which turn out to pass the corresponding tests. Popper and Lakatos even thought
that this extra success, predicted novel facts, was the litmus test for whether or not
the new theory is possibly closer to the (relevant) truth than the old one.
From our analysis it immediately follows that it is formally possible that a
new theory is closer to the truth than the old one while it only corrects the
individual problems of the old one, without extra general success. Suppose that
X is a subset of T and let I be the general test implication of X (hence X is
subset of 1) which has been falsified. Suppose now that Y is such that Y n I =
X and that Y is, like X, a subset of T. Under these conditions Y is closer to
the truth than X, only by losing known individual problems of Y, without
unintended general success. For under the specified conditions, X must be a
subset of Y, and hence theory X implies all general hypotheses following from
theory Y, hence all general successes of Y.
A similar case is possible for theories X and Y containing T as subset. Let
X fail to imply the established law Land let Y = X n L, then Y is again closer
to the truth than X, with only L as extra general success, and no unintended
loss of individual problems.
These special cases can be summarized as follows: if the theory under evaluation happens to be stronger or weaker than the true theory, the suggested ad
hoc repairs will bring one closer to the truth, without unexpected extra success.
However, if the theory is not simply stronger or weaker than the true theory,

TRUTH LIKENESS AND TRUTH APPROXIMATION

169

then the suggested ad hoc changes of the theory will almost inevitably lead to
new predictions of extra success, some of them coming true or false when the
repaired theory is or is not closer to the truth. It is not even excluded that
there are plausible general conditions under which 'almost inevitably' can be
replaced by 'inevitably'. However this may be, if comparative HD-evaluation
of a new theory is in favor of that theory, then, depending on the test route,
this either means an unexpected extra individual problem of the old theory, or
an unexpected extra general success of the new theory, where what is unexpected
is, of course, determined by the old theory. In sum, ad hoc repair of a theory
will seldom be a real improvement without unexpected extra success. In other
words, comparative HD-evaluation of an ad hoc repair will either lead to
unexpected extra successes of the new theory or extra successes of the old
theory that could have been, but were not, explicitly expected before. Hence,
besides some qualifications, the intuitions of Popper and Lakatos with respect
to ad hoc repairs and novel facts are largely justified. Instead of a ban on ad
hoc changes, they can be allowed, provided they are subjected to comparative
HD-testing with the original theory.
7.5.2. Crucial experiments

We have seen before that the HD-testing can be applied to the comparative
hypothesis "theory Y is closer to the truth than theory X". The comparative
hypothesis is suggested when one theory is more successful than another. Let us
now look from the truth approximation perspective to the situation that two
theories are equally successful, and hence to the idea of a so-called crucial test.
So let the two theories concerned, say X and Y, be equally successful before
the crucial test. We mayor may not assume, in addition, that both theories
have not yet been falsified. The methodological side of the idea of a crucial
test amounts to the following. First, a crucial test typically is supposed to be
a repeatable experiment, hence it concerns general test implications. More
specifically, the idea is to derive from X a general test implication /(X) of the
form "always when C then F" and from Y J(Y) of the form: "always when C
then not-F". Let us further assume that it follows from our background knowledge that one of these general testable conditionals has to hold and that it is
possible to test them with C as initial condition.
Under these conditions it is excluded that the two theories will remain
equally successful, for the experiments will force us to accept either J(X) or
/(Y) . Moreover, if the experiments force us to accept for instance J(Y), this not
only implies that Y is more successful than X as far as general successes are
concerned, but also that Y is more successful than X as far as individual
problems are concerned. The reason is that J(Y), starting as a falsifying general
hypothesis of X in the sense of Popper, has become a falsifying general fact of
X. Every investigated 'C-case' apparently resulted in not-F, making their combination not only in agreement with J(Y)'s prediction, but also making it a

170

TRUTHLIKENESS AND TRUTH APPROXIMATION

falsifying instance of /(X) (and hence of X). /(Y) summarizes and, pace Popper,
inductively generalizes these falsifying instances.
The assumption that the tests can and will have C as initial test condition
is important. If, due to practical constraints, the initial test conditions have to
be not-F and F, respectively, the situation is not that asymmetric. For it is not
difficult to check that in that case every successful test result for the one will
merely be a neutral result for the other, and not a falsifying one.
So much for the methodological aspects of a crucial test. What about its
consequences for truth approximation? Of course, at least all conclusions we
have drawn from the comparative statement that one theory is more successful
than another follow: X cannot be closer to the truth than Y, and Y can still be
closer to the truth than X. Moreover, new experiments (related to an old or a
new GTI of X or Y) may destroy the success dominance. This cannot destroy
the conclusion that X is not closer to the truth than Y, but it destroys the
conclusion that Y could still be closer to the truth than X. As a consequence,
a crucial experiment may temporarily lead to better truth approximation
perspectives for one theory compared to the other, but these perspectives may
well become destroyed. To be sure, the reverse perspective cannot arise, except
when the outcome of the crucial experiments are put into question or when
new considerations about auxiliary hypotheses lead to the conclusion that the
supposed falsifying general fact is, in fact, a general success.
For the case that both X and Y had not yet been falsified before the crucial
test, the following non-comparative conclusions can be added to the above
truth approximation conclusions. X is false, and Y may still be true. Moreover,
Y may later become falsified as well, but the conclusion that X is false can
only be withdrawn by reconsidering data or auxiliary hypotheses.
In several respects the present analysis is in accordance with Lakatos' analyses of crucial experiments, in which the temporary character and the revisability
of the conclusions is generally accepted. Our analysis adds to this that the
conclusions can be stated unproblematically in terms of (perspectives on) truth
and truth approximation, and can be generalized to falsified theories. The latter
point is very important as long as the theories under consideration must be
assumed to be 'born refuted', for instance due to unavoidable idealizations.
7.5.3. Inference to the best explanation

An interesting consequence of the truth approximation analysis concerns


so-called 'inference to the best explanation'. We will argue that it has to be
replaced by 'inference to the best theory'. Let us first define the two crucial
notions: 'the best theory' is the best, i.e., the most successful, theory among the
available theories, assuming that there is one, whereas 'the best explanation' is
the best theory which has not yet been falsified. The standard idea of 'inference
to the best explanation'20 amounts to the following:

TRUTHLIKENESS AND TRUTH APPROXIMATION

171

Inference to the best explanation (IBE):

If an explanation has so far proven to be the best one among the


available theories, then (choose it, i.e., apply rRS and) conclude, for the
time being, that it is true (i.e., that it is a true hypothesis)
IBE is problematic for at least three reasons. First, it does not say anything
about falsified theories. It is restricted to the case that the best theory has not
yet been falsified, hence individual problems are not taken into account. Second,
it couples a non-comparative conclusion, being true, to a comparative premise,
being the best unfalsified theory. Third, and finally, it is difficult to see how it
could be justified in terms of general considerations concerning the true/ falsedistinction.
Consider the plausible alternative, suggested by RS and the Forward
Theorem:
Inference to the best theory (IBT) :

If a theory has so far proven to be the best one among the available
theories, then (choose it, i.e., apply RS and) conclude, for the time being,
that it is the closest to the nomic truth T of the available theories
IBT does not have the three shortcomings of IBE. It applies to unfalsified as
well as to falsified theories. It couples a comparative conclusion, being the
closest to the truth, to a comparative premise, being the best theory. Last but
not least, it is directly justifiable in terms of truth approximation, viz., by the
Forward Theorem. To be precise, the comparative hypothesis that the best
theory remains the best, a hypothesis that apparently seems to be true, implies
that the best theory is the closest to the truth.
Hence, IBT can be seen as a, for good reasons, severely corrected version of
IBE. Let us summarize the differences. IBE is restricted to the case that the
best theory has not yet been falsified, whereas IBT applies in general in the
case that there is a best theory, already falsified or not. Moreover, IBE is,
unlike IBT, a standard rule of inference in the sense that the conclusion
is supposed to be true when the premises are. That is, according to IBE the
best unfalsified theory is supposed to be true, whereas IBT only infers that the
best theory is the closest to the truth. As a consequence, even if the best theory
is not falsified, IBT does not conclude that this theory is true, it still leaves
perfect room for the possibility that it is false. Finally, whereas it is difficult to
imagine a justification for IBE, IBT has a straightforward justification in terms
of truth approximation.
7.5.4. Descriptive and inductive research programs

The truth approximation analysis also gives the opportunity to consider the
nature of descriptive research programs. A descriptive research program uses
an observational conceptual frame, and may either exclusively aim at one or

172

TRUTHLIKENESS AND TRUTH APPROXIMATION

more true descriptions (as for example in most historiography), or it may also
aim at the true (observational) theory in the following specific way. Aiming at
the true theory is carried out exclusively by establishing observational laws.
Given that this requires (observational) inductive jumps (of the first and the
second kind), it is plausible to call such programs inductive research programs.21
It is easy to see that such programs 'approach the truth by induction'. For the
establishment of observational laws, the micro-step of the HD-method may be
applied, resulting in true descriptions which either falsify the relevant general
observational hypothesis or are partially derivable from it. According to the
basic definition, assuming that accepted observational laws are true, any newly
accepted observational law guarantees a step in the direction of the true theory.
For if S(t) and S(t') indicate at time t and t' later than t, the strongest accepted
law S(t') is closer to T than S(t). Hence, inductive research programs are
relatively safe strategies of truth approximation: as far the inductive jumps
happen to lead to true accepted laws, the approach not only makes truth
approximation plausible, it even guarantees it.
Note that inductive logic of the Hintikka variant ('induction by enumeration
and elimination'), see Chapter 6, in which generalizations are taken into account
up to the moment that they are falsified, is tailored to such inductive programs.
Explication of the nature of explanatory programs can best be postponed to
Chapter 9, which deals with stratified theories.

Concluding remark
At the end of Chapter 6 we argued for replacing the term Context of Justification
by Context of Evaluation, in view of the fact that the evaluation of theories
also deals with falsified theories, i.e., theories that certainly cannot be justified.
In this chapter we have shown that the term Context of Evaluation would also
be more appropriate in the light of the evaluation of truth approximation
claims of theories.
The next chapter will strengthen this claim for we will see that the intuitive
foundations of the presented definitions, as well as related intuitions concerning
the correspondence theory of truth and principles of dialectical reasoning, all
deal equally easily with false and falsified theories as with true and not - yetfalsified theories.

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

Introduction
In this chapter we present some topics concerning truthlikeness and truth
approximation in terms of intuitions that play, at least in our opinion, an
important role in scientific and philosophical theorizing. In Section 8.1. we will
spell out in model-theoretical terms that the basic definition of 'more truthlike'
in the nomic sense not only has an intuitively appealing 'model foundation',
but also, at least partially, a conceptually plausible 'consequence foundation'.
Moreover, combining the relevant parts of both leads to a very appealing 'dual
foundation', the more so since the relevant methodological notions, viz., 'more
successful' and its ingredients provided by HD-evaluation, can be given a
similar dual foundation. In Section 8.2. we will argue that the definition of
basic truthlikeness can be reinterpreted as an intra-level explication, for the
actual as well as the nomic perspective, of the idea of a correspondence theory
of truth. In contrast to the usual inter-level reading, the intra-level reading
makes straightforward sense of the intuitions that 'true/false' is a matter. of
corresponding or not corresponding to the facts and that 'closer to the truth'
is a matter of better corresponding to the facts. Finally, in Section 8.3. we will
show that the basic definitions of 'more truthlike' and 'more successful' lead to
straightforward explications of several dialectical notions, such as 'dialectical
negation', 'double negation', the triad 'thesis- antithesis-synthesis', and 'the
absolute'. In fact, logical, methodological and ontological explications of such
notions can be given for the actual as well as for the nomic perspective.
The reader is warned that the second and third section are rather controversial, a fact which will also become clear from the way of presentation. They
might well be skipped for the first reading, since only a few references to it will
be made later.

8. 1. CONCEPTUAL FOUNDATIONS OF NOMIC TRUTH


APPROXIMATION

Introduction

In the previous chapter we introduced, among other things, the basic definition
of nomic truthlikeness and its methodological consequences. In this section we
173

174

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

study the optimal conceptual foundations of the definition and its consequences.
In fact, it is the start, to be continued in later chapters, of a systematic search
for intuitively appealing conceptual foundations of ways of truth approximation
by the HD-method in terms of consequences and models of theories. In each
case we first consider the crucial notions 'more truthlike', 'more successful' and
results of HD-evaluation, separately. Secondly, we look for a coherent foundation, that is, a foundation that is essentially the same for all three notions.
There are three types of foundations of a definition, viz., a 'model foundation',
purely in terms of models, a 'consequence foundation', purely in terms of
consequences and a mixed or 'dual foundation', partly in terms of models,
partly in terms of consequences.
It will turn out that the dual foundation is the superior one: it can be given
for all distinguished qualitative ways of nomic truth approximation (basic and
refined, and observational, referential and theoretical) and it is in all cases
conceptually plausible and intuitively appealing. The model, as well as the
consequence foundation, turn out to have specific shortcomings, the first with
respect to the methodological notions, the second with respect to the refined
notions. Since stratified and refined truth approximation will be dealt with in
Chapter 9 and 10 these conclusions can only be drawn, by way of summary,
at the end of Chapter 10.
Recall that the basic explication of truthlikeness (and of successfulness)
presuppose a certain fixed vocabulary generating a fixed set of conceptually
possible structures, the conceptual frame. Hence, there is no dubious metaphysics involved in talking about 'the truth'. Like many others, Popper assumes,
moreover, that 'the truth' or 'the true theory' concerns one unique structure,
viz., the structure representing the actual world. However, we shall not adopt
this assumption. Although Popper is formally interested in the problem of
'actual truthlikeness', he seems informally more interested in the problem of
'nomic truthlikeness', to use our terminology. He wants to define the idea that
one theory is closer to the truth than another, and theories, in the natural
sciences at least, deal with what is nomically (im)possible. That is, the truth
does not so much concern the actual state of a system, but the set of nomically
possible states of a system or some other set of intended applications. However
this may be in various particular contexts, in our approach 'the truth' is
technically equated with some target set of structures, whereas some terminology will suggest, though will strictly not presuppose, the nomic interpretation
of that target set. That is, although we will continue to speak of nomic
truthlikeness and nomic truth approximation in this chapter, we will for the
rest mainly use terminology that is not laden with the Nomic Postulate. The
reader is referred to Tables 7.2. and 7.3. of the previous chapter, in which the
terminology to be introduced in this chapter has -already been mentioned
between brackets. With this terminology we want to stress that the analysis of
'non-actual' truthlikeness and truth approximation is useful for any target set
T of conceptual possibilities.

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

175

Recall also that a (general) hypothesis, by definition, characterizes a subset


of structures and that it does so by logico-mathematical means of one kind or
other. It is true if this subset includes the target set, and false otherwise, i.e.,
when there are counter-examples, of one kind or another, to the claim that the
hypothesis 'covers' all target structures. If the target set itself, i.e., 'the truth',
can be captured by a hypothesis, it may rightly be called 'the most informative,
or strongest, true hypothesis' of the context.
It is also important to stress again that we focus on a pure or 'qualitative'
definition of 'more truthlike', i.e., a definition not exploiting quantitative measures, such as numbers and real-valued distances. In this respect it is important
to note that by saying that one theory has 'more' so-and-so consequences or
models than another means that the set of so-and-so consequences or models
of the first (properly) includes that of the second; the same holds good for
'fewer'. Thus the notions of 'more' and 'fewer' are here to be understood as
referring to the qualitative relation of set-inclusion rather than to a quantitative
comparison. Similarly, 'closer to the truth' is not intended as a quantitative
notion, but is identified with the qualitative 'more truthlike'. For convenience,
the expression 'more truthlike' (or 'closer to the truth') is only used in the strict
sense when this is explicitly stated. If not, it is used in the weak sense of 'at
least as truthlike'. All remarks in this paragraph apply m.m. to qualitative
definitions of 'more successful'.
In this section we deal only with basic truthlikeness, successfulness, and their
relation to HD-results. In Chapter 9 we will see that nothing essential changes
for epistemologically stratified theories. Moreover, in Chapter 10 we will deal
with the corresponding refined notions and close with a general survey of
results and conclusions concerning foundations.
Although the main definitions are usually presented in their structuralist
version, as was done in the previous chapter, the logical, i.e., model-theoretical,
versions of the basic definition, and of its Popperian ancestor, in terms of sets
of models, consequences, and consequence-classes, form the best starting point
for foundational purposes. In what follows, 'logical' will always refer to such
model-theoretical notions.
8.1.1. Popper:S consequence-definition of 'more truth like ,

Although Popper's definition of 'more truthlike', viz., more true and fewer false
consequences (Popper 1963a), was intuitively very appealing, it nevertheless
turned out, as we have stated already, to be rather problematic. However, as
a stepping-stone towards the foundations of the basic definition, it is very
instructive to first disentangle Popper's proposal in logical terms.
Initially we shall focus on a first order (vocabulary generating a first order)
language L and, for convenience, even more specifically on finitely axiomatizable theories of that language, and hence on theories that can each be conceived
as the set oflogical consequences of one single sentence. However, the restriction
to finitely axiomatizable theories, including the assumption that the target set

176

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

of structures is elementary in the technical sense (see below), is inessential for


our purposes.
Let Sent(L) indicate the set of sentences and Str(L) the set of structures of
L, i.e., ordered tuples of sets and relations defined on them, in terms of which
L can be interpreted. Following Tarski's truth definition, in each structure every
sentence of L acquires a definite truth-value ('true' or 'false'). When IX is true
in a structure the latter is called a model of IX. Let the target set of structures
be elementary in the sense that there is a sentence e such that the target set
coincides with the set of models of Hence, both and its set of models may
be called 'the truth' or 'the true theory'. Popper technically assumes in addition
that 'the truth' concerns a unique (up to an isomorphism) target structure, viz.,
the structure representing the actual world. Let us call this the uts-assumption.
Hence, Popper is, formally speaking, dealing with approaching the actual truth.
However, since he seems primarily interested in the extent to which theories
succeed in approaching that truth, it is plausible to compare Popper's definition
with the basic definition, with a special assumption about the target set of
models, viz., the uts-assumption. The uts-assumption implies that is complete
in the sense that for every sentence IX, IX or ..., IX follows from e. Whenever the
uts-assumption, or just the completeness assumption, is to playa role, this will
be announced.
Popper's definition of the expression "1/1 is closer (or more similar) to ethan
4/' can, by way of reconstruction, be understood as being based on the following
underlying 'positive' intuition (PI) and 'negative' intuition (NI):

e.

PI:
NI:

all good parts of f/! are (good) parts of 1/1, in the sense that, for all
parts of the truth, if f/! has one, 1/1 has it as well
all bad parts of 1/1 are (bad) parts of f/!, in the sense that, for all
non-parts of the truth, if 1/1 has one, f/! has it as well

Both intuitions can be rephrased in terms of wrongly, respectively rightly,


missing parts. E.g., PI is equivalent to all "wrongly missing parts of 1/1 are
(wrongly) missing parts of f'.!
What the parts and non-parts of a theory are has still to be specified, but a
good part of a theory is always interpreted as a part which is also a part of
the truth and a bad part as a part which is not. To be sure, Popper did not
introduce the above intuitions explicitly, but presented his truthlikeness definition immediately in terms of sentential interpretations of these intuitions, more
specifically in terms of true and false consequences:
C.PI
C.NI

all true consequences of f/! are (true) consequences of 1/1


all false consequences of 1/1 are (false) consequences of f/!

For the purpose of presenting a formal translation of this definition we shall


use the standard (model-theoretic) definition of "IX is a consequence of {3" or

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

177

"fJ entails

a": "a is true in every model of fJ" (notation: fJ Fa). Cn(a) indicates
the set of consequences of c(. A true consequence of 'J. is a consequence of a
which is at the same time a consequence of 8, and hence true in all target
structures. Hence, Cn(a) n Cn(8) indicates the set of true consequences (the
truth content) of a. A consequence of a which is not a consequence of 8 is false
in the sense that it is false in some target structure. Hence, Cn(a) - Cn(8)
indicates the set of false consequences (the falsity content).
Consequently, the formal translations of C.PI and C.NI are as follows:

(it

Cn(~) n Cn(8) ~ Cn(I/!)(n Cn(8))

(iit

Cn(I/!) - Cn(8) ~ Cn(~)( -Cn(8))

or
Cn(I/!) ~ Cn(~) U Cn(8)

Recall that, strictly speaking, such a definition concerns the weak notion: "I/!
is at least as close to 8 as ", leading to the strict notion by imposing, in
addition, the requirement that the reverse claim "~ is at least as close to 8 as
I/!" does not hold at the same time.
Almost equally well-known as Popper's plausible definition of 1963, reproduced above, is the knock-down argument against it that came forward in
1974, and which has already been mentioned in the introduction of the previous
chapter. Miller (1974) and Tichy (1974) independently proved a theorem showing that the definition had a property of which even Popper immediately
conceded that it was highly undesirable. The theorem presupposes the utsassumption or at least the assumption that the truth is complete. It excludes
that a false theory, i.e., a theory having at least one false consequence, can be
closer to the truth than another one in the strict sense. Intuitively, however, it
is desirable that a false theory can be closer to the truth than another one. For
example, even if Einstein's theory is false, it must be possible that it is closer
to the truth than Newton's theory. Hence, in the light of the MillerjTichytheorem one has to conclude that Popper's explication is inadequate.
8.1.2. A related way of introducing the basic definition

Though the conceptual foundations of Popper's definition are intuitively


appealing, they are apparently defective. Before we formally introduce the
model version of the basic definition and show that its conceptual foundations
are not defective, we first derive two consecutive transformations of Popper's
definition. The second transformation will amount to the claim that Popper's
first clause (it is equivalent to the formalization of the model-interpretation
of the negative intuition NI, viz.:
M.NI

all mistaken models of I/! are (mistaken) models of ~

178

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

whereas his second clause (iit is much stronger than the formalization of the
model-interpretation of the positive intuition PI, viz.:
M.PI

all correct models of are (correct) models of tjJ

This is particularly interesting because the basic definition will turn out to be
based on M.NI and M.PI.
Let us start with the transformations. The standard logical theorems
"tl E Cn(tl)" (trivial) and "Cn(tl V fJ) = Cn(tl) n Cn(fJ)" (non-trivial) immediately
lead to a translation of Popper's definition in terms of elementary consequence
claims. Popper's first clause (it, the formalization of CPI, is equivalent to:

(jt

tjJl=v8

For the model-translation of such a claim we need some direct consequences


of Tarski's truth definition: Mod(tl V fJ) = Mod(tl) U Mod(fJ), Mod(tl 1\ fJ) =
Mod(tl) n Mod(fJ), Mod( -, tl) = Mod(tl) = Str(L) - Mod(tl). Here Mod(tl) indicates the set of models of tl, or the model-set of tl.
It is easy to check that (jt is equivalent to "Mod(tjJ) Mod() U Mod(8)"
and hence to the formalization of the claim that tjJ has fewer mistaken models
than :
(kt Mod(tjJ) - Mod(8) Mod()( - Mod(8))
Hence, (kt is the direct formalization of a model-interpretation of the negative
intuition NI, i.e., M.NI (see above). Recall that (kt is indirectly, viz., via (it,
also a formal translation of the consequence-interpretation of PI (CPI). Hence,
the consequence-interpretation of PI (CPI) is equivalent to the model-interpretation of NI (M.PI).
Now let us look at Popper's second clause (iit, the formalization of CNI.
It is equivalent to
(jjt either l=tjJ or 81=tjJ, or both
In its turn, it is equivalent to
(kkt either Mod() Mod(tjJ) or Mod(8) ; Mod(tjJ), or both
for which there is no equally appealing informal reading. In particular, it is
not the formalization of the claim that tjJ has more correct models than , i.e.,
it is not the (direct) formalization of the model-interpretation of the positive
intuition PI, i.e., M.PI (see above).
The formal translation of M.PI, interpreting a correct model of , of course,
as a member of Mod()) n Mod(8), is:
(kk)

Mod()

n Mod(8) ; Mod(tjJ)(n Mod(8))

It is easy to see that each of the two disjuncts of (kkt separately imply (kk),
with the consequence that (kkt is (much) stronger than (kk), and hence that

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

179

the consequence-interpretation of NI (CNI) is (much) stronger than the modelinterpretation of PI (M.PI). The other way of easily seeing this is based on the
logical transformation of (kk) in terms of sets of consequences, which is, together
with the intermediate logical transformation, given by:
(kk)

(jj)

~/\(}I=I/J

(ii)

Cn(I/J)()Cn(-,(})~Cn(~)

Since (iit is equivalent to Cn(I/J) () Cn(}) ~ Cn(~) it is (much) stronger than (ii),
for Cn( -, (}) evidently is a proper subset of (}).
In the light of the foregoing we may conclude that the model-interpretation
M.PI and M.NI of the intuitions PI and NI is not only intuitively appealing,
but may be non-defective. This suggests the basic definition of 'more truthlike'
for finitely axiomatizable theories in terms of models: "I/J is closer to () than iff

(k)
(kk)

Mod(I/J) - Mod(}) ~ Mod(~)


Mod(~) () Mod(}) ~ Mod(I/J)

Whereas (k) is just equivalent to (kt, (kk) is a motivated weakening of (kkt


We call this the model-formulation of (the logical version of) the basic
defini tion. 2
For survey purposes we repeat the logical transformations of (kt and (kk),
now omitting, in view of the equivalence of (k) and (kt, the superscript P.
(k)

(kk)

(j)

I/JI=~ V ()

(jj)

~/\(}I=I/J

(i)
(ii)

Cn(~) () Cn(}) ~ Cn(I/J


Cn(I/J)()Cn(-,(})~Cn(~)

8.1.3. Comparison

Table 8.1. depicts the relations between the two (formalized) interpretations of
the two intuitions, and demonstrates that Popper's definition is stronger than
the basic definition.
It is also illuminating to draw figures that represent the two set versions of
both definitions. Figure 8.1. represents Popper's definition. Figure 8.l .a. represents the sets of consequences of ~, I/J and () as subsets of Sent(L) or, more
precisely, of the set of equivalence classes of logically equivalent sentences (i.e.,
Table 8.1. Comparison of Popper's definition with the basic definition as consequence versus model
interpretation of the same intuitions
Popper's definition:

basic definition:

conseq uenceinterpretation

modelinterpretation

PI-+C.PI=
NI -+ C.NI =

=M.NI-NI
(ii/jj/ kkt ~ (ii/jj/kk) =

4=

M.PI - PI

180

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

Sent(L)

Cn(

Str(L)

Cn(lp)

Cn(8)

Mod(8)

a: consequences

b: models

Figure 8.1. Popper's definition: (a) consequences, (b) models

propositions in the technical sense) of L. Figure 8.l.b. represents the modelsets of </I, '" and () as subsets of Str(L) or, more precisely, the set of equivalence
classes of isomorphic structures of L. When the definition applies, this means
that certain areas are empty on logical or conceptual grounds. A blackened
area indicates that the area is empty due to (it in Figure 8.l.a. or equivalently
due to (kt in Figure 8.1.b. In Figure 8.l.a. double shading indicates the empty
area due to (iit Its equivalent (kkt in terms of models is indicated in
Figure 8.l.b. by horizontal and vertical shading of the two areas of which at
least one should be empty. Note that this implies that their double shaded
intersection should be empty anyhow.
Figure 8.2. represents the basic definition, using the same conventions.
Sent(L)

Str(L)

Cn(-rJ)

><""7f'-;r;r-M~od(11')

Cn(8)

a: consequences
Figure 8.2. Basic definition: (a) consequences, (b) models

Mod(e)
b: models

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

181

Figure 8.2.a. depicts the consequence version of the basic definition, i.e., (i) and
(ii). Note that all the theories have one point in common, viz., the tautology.
Figure 8.2.b. depicts the model version, i.e., (k) and (kk). Note the plausible
perfect similarity between the Figure 8.2.b and Figure 8.l.a. and the dissimilarity
between Figure 8.2.a. and Figure 8.l.b.
From Figures 8.l.a. and 8.2.a. it is easy to see that (iit is much stronger
than (ii), for the double shaded area in the first is clearly a superset of the one
in the second. Similarly, from Figures 8.l.b. and 8.2.b. it is easy to see that
(kkt is much stronger than (kk), for (kk) only requires that the double shaded
area is empty, whereas (kkt adds to this that at least one of the two areas
shaded only once is also empty.
8.1.4. The knock-down argument and its blockade

It will now be argued that and why the arguments of Miller and Tichy cannot

be used to knock down the weaker, basic definition, based on M.NI and M.PI.
For that purpose it is necessary to set up a decomposed version of the
Miller/Tichy-theorem and the proof in terms of consequences, hence in particular Figure 8.l.a. and 8.2.a. may help the reader in following the arguments.
Lemma 1:

If (ii

t and t/J is false then </I F t/J.

Proof: The lemma-conditions amount to showing that Cn(t/J) n Cn(fJ) - Cn(</I)

is empty. Assume that some ex belongs to that set. Given that t/J is false, it
follows from (iit that there must be some /3 in (Cn(</I) n Cn(t/J - Cn(fJ). This
implies that cx&/3 is a false consequence of t/J, but not of </I, which is excluded
by (iit. Hence, we may conclude that t/J is a weakening of 1/1. Qed.
Lemma 2: If (it, fJ is complete and

t/J is false then t/J F</I

t/J, which in view of (it amounts


to the assumption that there is some y belonging to Cn(</I) - (Cn(t/J) U Cn(fJ.
That t/J is false enables taking a /3 in Cn(t/J) - Cn(8). From the completeness of
8 it follows that -, /3 has to be a consequence of 8. Hence, -, /3 V Y is a true
consequence of </I and not of t/J, contrary to (it (for if -, /3 V Y were a consequence of t/J, y would, contrary to our assumption, be a consequence of t/J, in
view of the fact that /3 is assumed to be a consequence of t/J). Hence, 1/1 must
be a weakening of t/J. Qed.
Proof: Assume that </I is not a weakening of

Combining the two lemmas we get:


Miller/ Tichy-theorem:

If (it and (iit, and 8 is complete and

t/J false then </I and t/J are equivalent 3

From the proofs of the two lemmas it is clear that the consequence-interpretation of NI, i.e., CNI, hence (iit, is crucial for the proof and hence for the

182

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

problem. To be precise, according to Lemma 1, proving that the possibility of


obtaining a false theory with a smaller, but non-empty, set offalse consequences,
i.e., the strict version of (iit, can only be fulfilled by weakening, needs no
further condition, whereas, according to Lemma 2, proving that the possibility
of obtaining a false theory with a larger set of true consequences, i.e., the strict
version of (it, can only be fulfilled by strengthening, requires essentially the
completeness assumption of the true theory. Hence, (iit is in a sense much
stronger, viz., more restrictive, than (it Without doubt, Popper thought that
the first possibility could be fulfilled in other ways than, of all means, by
weakening. However, whatever alternative definition we may propose, Lemma 1
does not leave room for a definition which can fulfil this possibility, that is,
a false theory having fewer false consequences than another has to be a
weakening.
Now it is easy to check that the adapted potential version of Lemma 1, and
hence of the Miller/Tichy-theorem, is blocked when (iit is replaced by (ii).
Two steps in the proof of Lemma 1 are now blocked. First, though the falsity
of 1/1 implies that 1/1 has some false consequence /3, (ii) does not imply that this
/3 has to be a consequence of ~. Second, if it nevertheless is a consequence of
~, its conjunction with a true consequence C{ of 1/1 not shared by ~ is a false
consequence of 1/1 not shared by ~, which is not excluded by (ii). Hence, if the
uts-assumption applies too, (ii) and (it leave room for the case that 1/1 is false
and nevertheless not equivalent to ~, in other words, that 1/1, although false, is
closer to the truth than ~ in the strict sense. To be sure, Lemma 2 then requires
that ~ is false and that 1/1 needs to be a strengthening of ~.
In the light of the foregoing we may conclude that the model-interpretation
M.PI and M.NI of the intuitions PI and NI is not only intuitively appealing,
but also non-defective. This legitimates the basic definition of 'more truthlike',
at least for finitely axiomatizable theories. Moreover, C.NI is problematic, but
C.PI is not.
8.1.5. Intuitive foundations of the basic definition

Assuming that the basic definition is the adequate one for finitely axiomatizable
theories, it is tempting to claim that it is clearly the model-interpretation of
the positive and the negative intuition (M.PI and M.NI) which leads to the
adequate definition, and hence provides its proper intuitive foundation.
However, there is another interesting interpretation of the situation. In view of
the equivalence of the formal translations of M.NI and C.PI, the basic definition
may also be seen to be based on the combination of the model- and the
consequence-interpretation of the positive intuition, that is:
C.PI (i)
M.PI (kk)

all true consequences of ~ are (true) consequences of 1/1


all correct models of ~ are (correct) models of 1/1

We call this the dual foundation of the basic definition, roughly: more true

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

183

consequences and more correct models. In contrast, the combination of M.PI


(kk) and M.NI (k) is called its model foundation, roughly: more correct and
fewer mistaken models. Both foundations seem to represent the intuitions of
scientists. However, this is certainly also the case for Popper's consequenceinterpretation of the positive and negative intuition, roughly: more true and
fewer false consequences. Hence, it is the negative intuition that turns out to
be dubious. From this perspective, the dual foundation, which only uses the
positive intuition, is preferable to the model foundation. However this may be,
the model and the dual foundation are both intuitively appealing and conceptually plausible. From now on, for convenience, we will call conceptual foundations of a certain type and purpose, if available at all, plausible when they can
be shown to have a conceptually sound basis. We call them intuitive when they
are moreover intuitively appealing, i.e., seem to express 'scientific common
sense'. Hence, the model and the dual foundation are both intuitive foundations
of the basic definition.
There does not seem to be an intuitive foundation of the basic definition
purely in terms of (sets of) consequences. For, in view of the fact that members
of Cn(l/I) n Cn( -, 8) do not seem to have a current name, the relevant clause
(ii) does not have an intuitively plausible reading as an application of the
negative intuition to consequences. However, (ii) can be given a non-intuitive,
but nevertheless plausible reading as an application of the negative intuition
to consequences. From the trivial fact that Cn( -,8) is deductively closed while
all its non-tautological members are false, it follows that all its members
transmit their falsehood to their non-tautological consequences, in the same
way as members of Cn(8) transmit their truth to their consequences. 4 It is
therefore plausible to call the members of Cn( -,8) 'strongly false', and the
members of Cn(l/I) n Cn(-, 8) 'strongly false consequences of 1/1'. Hence, (ii) can
be read as the application of the negative intuition to strongly false consequences, 1/1 has fewer strongly false consequences than 4>, or more precisely:
C.NI*

(ii) all strongly false consequences of 1/1 are (strongly false)


consequences of 4>

Though this reading is plausible, it is certainly not intuitively appealing, hence,


although C.PI and C.NI* together lead to the basic definition, their combination can hardly be called an intuitive foundation for it. In sum, though there
is also a plausible consequence foundation of the basic definition, this is not an
intuitive foundation. The intuitively plausible consequence foundation for being
closer to the truth, i.e., Popper's C.PI and C.NI, has been shown by Miller and
Tichy to fail.
Lack of space prevents us from presenting the simple and plausible generalization of the basic definition and its foundations to cover theories of a first
order language which are not finitely axiomatizable. Consequence sets of sentences can simply be replaced by deductively closed sets of sentences, and

184

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

model-sets of sentences can be replaced by sets of models of sets of sentences.


Hence, the above restriction to finitely axiomatizable theories was inessential,
including the assumption that the target set of structures was elementary.5
It will turn out in the next chapter that the foregoing model and dual
foundations designed for unstratified theories can easily be adapted for epistemologically stratified theories. Moreover, it will be shown there that the
refined definition of truthlikeness can be based on a concretization of the model
foundation, as well as on a concretization of the dual foundation. Since the
idea of refined truthlikeness can most easily be presented in structuralist terms,
it is desirable to conclude this section with a reformulation of the structuralist
version of the basic definition and its adapted foundations.

8.1.6. The structuralist version of the basic definition


From the structuralist perspective, one does not need to be explicitly concerned
with the language, as long as this is not required by one's questions of interest.
All we need to assume is a set of structures of a certain similarity type, providing
the set of potential models or conceptual possibilities Mp, which mayor may
not be the set of structures of a first order language. Members of a subset X
of Mp are called structuralist (s-)models of X. Let subset T, 'the truth', of Mp
be the target set of research. Recall that the interpretation of this set may be
that it represents the singleton set of the actual world, the set of nomic
possibilities, or some other set of (one or more) intended applications.
However this may be, analogous to the logical definition of truth likeness,
the following structuralist model-interpretation of the two basic intuitions NI
and PI is plausible as the definition of "Y is at least as close to T as X"
(MTL(X, Y, T)):
(Bi)
(Bii)

Y - Tr:;X(or:Yr:;TUX)
X n Tr:; Y (or: X n Tr:; Y)

It requires some argumentation to see that this structuralist definition can


also be conceived as resulting from the dual interpretation of the positive
intuition. Recall the definition of Q(X), the 'co-powerset of X', for an arbitrary
set of structures X:

Q(X):

the set of supersets of X (being subsets of Mp)

If Mp is the set of structures of a certain first order language L, i.e., Mp =


Str(L), and if X = Mod(<p) then Q(X) includes the set of models for each logical
consequence of <p. However, in general, Q(X) not only represents all first order
consequences of X, but all other set-theoretically characterizable consequences
of X. For simplicity, the set of s-models corresponding to an arbitrary consequence is itself called a consequence, in this section more specifically a structuralist (s-)consequence, in order to hint at non-first order consequences. In other

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

185

words, Y is an s-consequence of X if and only if X ~ Y. We call Q(X) the


structuralist (s-)consequence-class of X. Note that Q(.) has some nice properties
similar to properties of logical consequence-classes: X ~ Y<> Q(Y) ~ Q(X) and
Q(X) n Q(Y) = Q(X U Y).
It is easily verified that clauses (Bi) and (Bii) are equivalent to the following
consequence-class formulations in structuralist terms:
(Bi)
(Bii)

<>
<>

Q(X U T) ~ Q(Y)
Q(Y);Q(XnT)

<>
<>

Q(X) n Q(T) ~ Q(Y) n Q(T)


Q(Y) n Q(T) ; Q(X) n Q(T)

Note first that the equivalent of (Bi) on the right side corresponds to (BiC) of
the previous chapter. Note also that a member of Q(X U T) is an s-consequence
of X and T, i.e., a true s-consequence of X. Hence, (Bi) roughly states that Y
has more true s-consequences than X. Hence, the structuralist version of the
basic definition can also, in view of the Q-reading of (Bi), and the straightforward reading of (Bii), be based on the dual foundation, i.e., the consequenceand the model-interpretation of the positive intuition:
all true s-consequences of X are (true) s-consequences of Y
and
all correct s-models of X are (correct) s-models of Y
Or, roughly, "Y is closer to the truth than X" iff it has more true s-consequences
models and more correct s-models.
Again there is also a plausible, though non-intuitive, consequence foundation.
As mentioned already, (Bii) is equivalent to
Q(Y) n Q(T)

Q(X) n Q(T)

Moreover, it is easy to check that non-empty members of Q( Y) n Q(T) may


be interpreted as 'strongly false s-consequences' of Y in the same sense as the
members of Cn(,O). Hence, (Bi) and (Bii) may be interpreted as
all true s-consequences of X are (true) s-consequences of Y
and
all strongly false s-consequences of Yare (strongly false) s-consequences
of X
Note, finally, that the structuralist definition has, on the one hand, a larger
domain than the logical definition, for it also applies to non-first order theories.
On the other hand, for the subdomain of first order theories it is stronger, for
it also takes non-first order consequences into account.

186

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

In sum, so far it has been shown, starting from Popper's failing consequencedefinition of 'more truthlike', viz., more true and fewer false consequences, that
a reinterpretation of Popper's underlying intuitions in terms of models leads
to the perfectly viable so-called basic definition, viz., more correct and fewer
mistaken models. It has also been argued that this intuitive 'model foundation'
of the basic definition can be replaced by a 'dual foundation', which is also
intuitively appealing: more true consequences and more correct models.
Moreover, there is, on second analysis, also a plausible 'consequence foundation', which is, however, not intuitively appealing: more true and fewer strongly
false consequences. Finally, similar foundations are possible for the structuralist
version of the basic definition.
8.1.7. Basic methodology

The given foundational analysis invites for extension into at least the following
directions. What foundations can be given for related definitions? ( 1) The basic
definition of 'more successful' in terms of successes and problems, (2) the basic
definition of 'closer to the observational truth' and 'observationally more successful', in the case of a distinction between theoretical and observational
components, (3) the refined definitions of 'closer to the truth' and 'more
successful, and their stratified versions, (4) the quantitative definitions of truthlikeness and success, based on numbers of models, or distances between them.
It may seem plausible that in some of these cases there can also be given
intuitive model and dual foundations and that, at most, a non-intuitive consequence foundation can be provided. In this subsection we will investigate these
possibilities for the first, methodological question. In the next chapters we will
pay attention to the other questions.
Recall that we have shown in detail, in Chapter 6 and 7, that the hypotheticodeductive (HD- )method is functional for basic truth approximation when combined with the rule of success (RS). This rule presupposes an explication of the
idea that one theory is more successful than another in a similarly basic way
to the basic definition of greater truthlikeness. The rule prescribes to adopt the
more successful theory. The core of the functionality proof was two-fold. First,
it was shown that HD-evaluation of separate theories provides the ingredients
for comparing the success of two theories in the basic sense. Second, it was
proved that 'being more truthlike in the basic sense' implies 'being at least as
successful in the basic sense', the Success Theorem.
Now it will be argued that there is also an intuitive dual foundation of 'being
more successful' in the basic sense, and, moreover, that this foundation is
strongly suggested by the method of HD-evaluation. Combining the relevant
observations, the ultimate conclusion can be that the evaluation methodology
is a method of basic truth approximation having an intuitive dual foundation
in terms of successes and problems. In contrast to this attractive dual foundation, however, now there does not seem to be a plausible, let alone an intuitively

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

187

appealing, model foundation. On the other hand, there is a plausible consequence foundation of HD-results and of 'more successful' in the basic sense.
The presentation will be in structuralist terms to ensure that the continuity
between the basic and refined approaches can be as large as possible in
Chapter to.
Let us briefly review the HD-method, in the sense of separate and comparative HD-evaluation, presented in Part II. The core of HD-evaluation amounts
to deriving from the theory in question, say X, (general) test implications and
to subsequently testing them. A test implication amounts to a superset I of X
with the claim "T!;; I". Testing this claim leads sooner or later either to a
counter-example of this claim, and hence to an (individual) problem of X, or
to the (revocable) acceptance of (the claim of) I: a (general) success of X.
Although a counter-example implies the falsification of X, this is, according to
the evaluation methodology, no reason to stop the HD-evaluation of X.
Systematic HD-evaluation of X will lead to a (time relative) evaluation report
of X, consisting of registered problems and successes. These two types of
HD-results directly specify the intuitive dual foundation of the HD-results.
Recall also that a success essentially concerns the idea of a law, when Tis
interpreted as the set of nomic possibilities. For, if T is a subset of I, the latter
says something about all nomic possibilities, viz., that they satisfy the conditions
specified in I. In this subsection some 'law-terminology' will suggest, but not
presuppose, this interpretation.
A success of one theory mayor may not be a success of another theory. The
same holds for problems. Successes and problems provide together the basis
for success comparison of theories. For comparative judgements it is easy to
have a common representation of all registered successes and problems of some
theories, or still more generally, of all established nomic possibilities and all
established laws. Let R(t) indicate the set of established nomic possibilities at
time t and S(t) the intersection of all established laws. Hence S(t) is the strongest
law established, and all its consequences are laws. Assuming that mistakes have
not been made, R(t) is a subset of T and S(t) a superset of T. The evaluation
report of X can now be summarized by R(t) - X, indicating the set of all
problems of X, reinterpreted as 'established wrongly missing models', and by
Q(S(t n Q(X), indicating the set of all successes of X, reinterpreted as 'established laws' or, more neutrally, 'established true consequences'. In sum, R( t) - X
and Q(S(t n Q(X) summarize the results of HD-evaluation of theory X and
reflect the intuitively appealing dual foundation of HD-results. These
HD-results can be compared with the results of another theory, which may
lead to theory choice on the basis of the appropriate rule of success.
Before we go to the actual comparison, it is interesting to investigate whether
there are, besides (individual) problems and (general) successes, other plausible
'units of comparison', enabling a uniform model or consequence foundation of
the HD-results.
The model foundation of HD-results requires, besides the idea of an individual problem, a model substitute of a general success of a theory. It is easy to

188

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

check that a general success is the complement of a set of established rightly


missing models. Hence, the central notion is an established rightly missing
model of a theory, or an established impossibility match, as it was called in
the previous chapter. However, to be established as a rightly missing model is
essentially a comparative claim, concerning a non-model, with respect to all
target structures. This notion not only seems to be alien in scientific practice,
it is also conceptually very complicated. Essentially, the same holds for its
counterpart, the notion of an established mistaken model, or an established
impossibility mistake. Hence, the model foundation of HD-results, though
available, is not plausible, let alone intuitively appealing.
The consequence foundation of HD-results requires, besides the idea of a
(general) success, a consequence substitute of an individual problem, i.e., a
counter-example of a theory. Now it is easy to check that, in structuralist
terms, the complement of any set of established counter-examples, including a
single one, is an established strongly false consequence of the theory. Though
not intuitively appealing, it may be called conceptually plausible, in particular
because to be established as a strongly false consequence 'only' requires that
its complement set is a set of established counter-examples. This makes the
concept of an established strongly false consequence at least less complicated
than that of an established rightly missing model, discussed above. The same
holds for its counterpart, the notion of an established strongly false (rightly)
missing consequence. Hence, the suggested consequence foundation of
HD-results may be called plausible, though not intuitively appealing.

8.1.8. The rule of success


Crucial for rules of success is the explication of the comparative notion 'being
at least as successful'. We will take the explication suggested by the basic
definition according to its model version as a point of departure. Each of the
resulting two clauses will be followed by its equivalent in terms of the relevant
ingredients provided by the HD-method.
Clause (Bi) suggests:
(Bi)-S

Y-S(t)~X

all established mistaken models of Yare models of X


which reads In terms of (general) successes, i.e., established true general
consequences:
Q(X) n Q(S(t)) ~ Q(Y)
all successes of X are successes of Y

Calling established correct models, examples, clause (Bii) suggests:

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

(Bii)-R

189

XnR(t)S; y

all examples of X are examples of Y


or in terms of counter-examples ('established wrongly missing models', i.e.,
individual problems):
R(t)- ys;R(t)-X

all counter-examples of Yare counter-examples of X 6


Both clauses have two other equivalent versions, the first one in terms of
models, the second in terms of consequences. However, as suggested in the
previous subsection, comparison of theories usually takes place in terms of
general successes and individual problems, that is, the direct products of
HD-evaluation. Hence, the intuitively most plausible basic definition of 'more
successful' is strongly suggested by the method of HD-evaluation and amounts
to the second version of the two above mentioned clauses, roughly speaking:
more successes and fewer counter-examples.7 Compared to this particular form
of the dual foundation of 'being more successful', i.e., in terms of consequences
and (missing) models, the possible uniform foundations do not seem to be
intuitively appealing. To be sure, these uniform foundations can be specified,
but the consequence foundation has, at most, some conceptual plausibility,
whereas the model foundation is not even plausible. s
All this suggests that the dual foundation is the only coherent foundation of
nomic truth approximation, at least for the basic perspective. To subscribe to
this claim the link between 'more successful' and 'more truthlike' will briefly
be summarized. When a theory is at a certain moment more successful than
another this is not yet a sufficient reason to prefer the first one in some
substantial sense. That would be a case of 'instant rationality'. However, when
Y is at a certain moment more successful than X, this does suggest the
comparative success hypothesis (CSH) that Y will remain more successful than
X. It is important to realize that this is an interesting hypothesis, even if Y has
already been falsified. It is this comparative hypothesis which is, indirectly, the
object of further evaluation of the two theories. When the results of this
comparative evaluation justify it, theory selection will take place by the
following conditional elimination rule:
Rule of Success (RS):

If Y has so far proven to be more successful than X,


i.e., CSH has been sufficiently confirmed to be accepted
as true, then eliminate X in favor of Y

RS doesn't require that Y '(always) remains more successful', for that would
rest on the presupposition that CSH could be completely verified (when true).
Hence we only demand that CSH has been sufficiently confirmed, which is, of
course, a matter of debate, for which reason the choice is always temporary.
In Chapter 7 we have argued in great detail that the basic evaluation

190

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

methodology, i.e., the combination of HD-evaluation and RS, is functional for


basic truth approximation due to the Success Theorem which states that 'closer
to the truth' guarantees 'at least as successful', both taken in the basic sense.
In the light of the foregoing, we may now conclude, in addition, that the basic
evaluation methodology has an intuitive dual foundation in terms of consequences, viz., successes, and missing models, viz., counter-examples.
According to the resulting dual foundation of basic truth approximation,
HD-evaluation provides successes (established true consequences) and counterexamples (established wrongly missing models) of theories. Such HD-results
may support the tentative conclusion that one theory remains more successful
than another in the basic sense of having more successes and fewer counterexamples. If so, this provides good reasons for believing that the more successful
theory is also more truthlike in the basic sense of having more true consequences
and more correct models.
Table 8.2. Foundations of basic truth approximation
Foundations
Model

Dual

Consequence

More truth like

basic

intuitive

intuitive

plausible

More successful

basic

available

intuitive

plausible

HD-results

available

intuitive

plausible

Table 8.2. summarizes the foundational conclusions about the basic notions.
Recall that if conceptual foundations of a certain type and purpose are available
at all, they are called plausible when they can be shown to have a conceptually
sound basis. They are called intuitive when they are moreover intuitively
appealing, i.e., seem to express 'scientific common sense'.

8.2. TRUTH LIKENESS AND THE CORRESPONDENCE THEORY


OF TRUTH

Introduction
In this section it will be argued that it is possible to conceive the definition of
actual and nomic truthlikeness as intralevel explications of the correspondence
theory of truth. In contrast to interlevel attempts, this intralevel explication
not only realizes in a transparent way that 'true' and 'false' amount to corresponding to the facts and not corresponding to the facts, respectively, but also
that 'more truthlike' amounts to better corresponding to the facts.

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

191

8.2.1. Interlevel and intralevel interpretations of correspondence

For the empirical sciences it is plausible to distinguish one concrete and two
abstract levels. Level C, the object level, is the level of the nomic world of
which the basic units consist of (almost) unconceptualized systems, state of
affairs and events. On level I, the first level of representation, the basic units
are set-theoretic structures; they are used to represent members, parts or aspects
of level C. Finally, on level II, the second level of representation, the basic
units are the sentences of formal languages; they can be used to formalize level
I, and they can be interpreted on level I. Scheme 8.1. depicts the three levels.
II

sentences

formalization / interpretation

structures (models)
representation

1
C systems
Scheme 8.1. Levels of representation

Whatever 'truth' is, it is clear that 'correspondence' in 'the correspondence


theory of truth', in view of the three levels, may be an interlevel or an intralevel
affair. Moreover, it can be conceded that at first sight an interlevel interpretation
of 'correspondence' seems the most plausible, and it is well known that Tarski's
so-called semantic definition of truth (SOT) is frequently claimed to explicate
such an interlevel conception (between II on the one hand, and I or C on
the other).
In its modem form SOT is based on the notion of a model of a formal
language. Leaving technical details aside, a model M for language L associates
with each sentence s of L a statement or interpreted sentence i(s, M) about the
structure generated by M, the interpretation of sin M. Now, SOT is a recursive
procedure, using assignments that do or do not satisfy atomic and complex
formulas, such that s gets the truth-value 'true' in M if and only if (it is true
that) i(s, M), i.e., if and only if the corresponding statement (interpreted sentence) is true. Hence, SOT defines a formal truth-value for sentences relative
to a model in terms of an absolute, albeit non-formal, truth-value of statements
about the structure generated by the model. Moreover, it is clear that SOT is
in this modem form an interlevel matter between I and II.
Hodges (1986) gives an illuminating account of the differences between
Tarski's original paper of 1935 (Tarski 1935/ 1956) (unfortunately, Hodges does
not even mention Tarski's influential informal version of 1944! (Tarksi 1944
and the modem version of SOT, occurring in Tarski's papers, according to
Hodges, for the first time in (Tarski and Vaught 1957). The main difference is

192

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

that Tarski deals in his 1935-paper only with 'Frege-Peano languages', which
areJormalized languages with afixed interpretation for all non-logical constants,
and not with modern (first order) Jormallanguages, of which it is characteristic
that they have non-logical uninterpreted constants, for which a model provides
an interpretation.
We would like to add two remarks to Hodges's clarifying differential diagnosis. In line with the type of language, Tarski assumes in his original papers a
context where there is only one world or, more carefully, one model or structure
representing that world, described by the formalized language, of which every
sentence gets a definite truth-value. This is indeed quite different from the
present-day flexible, and very fruitful, use of SOT for formal languages or, to
use Hodges' way of phrasing, 'truth in a structure' is one of the few important
scientific inventions of an indexical kind.
Surprisingly enough, Hodges does not pay attention to the fact that Tarski
wrongly suggests that his technical definition can be directly generalized to the
empirical sciences. To be precise, Tarski suggests in the informal parts of both
mentioned technical papers, but mostly in the informal version of 1944, that
he is dealing not only with (pure) mathematics, but also with (formalizable)
empirical sciences. However, he only gives the definition in full detail for what
he calls 'the language of the calculus of classes' (with a fixed interpretation for
'inclusion'!). In this context the target is not only one world, but an abstract
one, or at least a non-material one, viz., the universe of classes. Having this
mathematical paradigm in mind might well explain why Tarski considers only
one target world, as well as why he does not make the distinction corresponding
to our distinction between concrete level C and abstract level I. This may be
defensible, even unavoidable, for pure mathematics; it is certainly an avoidable
confusion for the empirical sciences.
Tarski claimed first of all that SOT explicates Aristotle's dictum:
"To say of what is that it is not, or of what is not that it is, is false, while to
say of what is that it is, or of what is not that it is not, is true", or in terms of
a famous instance of Tarski's material condition of adequacy: "The sentence
'snow is white' is true if, and only if, snow is white". This condition will here
be called the condition oj co-ordination. The claim that SOT realizes this
condition has never been seriously disputed, and it will be clear from our
description of SOT above that we can easily concede this point, although only
as far as coordination between levels I and II is concerned.
For us it is more important that Tarski, though hesitatingly, and later
Carnap, Stegmiiller and Popper, without hesitation, claimed that SOT provides
in addition an (interlevel) explication of the idea of a correspondence theory
of truth. This claim, however, has been seriously disputed. For surveys of
criticism one may consult e.g., O'Connor (1975), Puntel (1978), and Kirkham
(1992). Here we will evaluate the claim in terms of what we consider to be the
two basic correspondence intuitions:

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

(CI)

193

'>1 is true (false)" if and only if


'>1 corresponds (does not correspond) to 'the facts'"

(C2) "B is more similar (closer) to the truth than A" if and only if
"B corresponds better to 'the facts' than A"
Assuming that A and B are sentences on level II and that 'the facts' are
localized on level I or level C, it is clear that SOT does not explicate intuition
(C2), at least we do not see any point of contact, let alone an indication, in
the direction of (C2). If we accept this, the fact that SOT may seem to explicate
intuition (CI) may well be due to the fact that it explicates and satisfies the
coordination condition. These considerations equally apply to the original
version of SOT.
One might object to this evaluation that the interlevel correspondence claim
associated with SOT has to be restricted to atomic sentences, as the 'snow is
white'-example, used as an instance of correspondence, may suggest. However,
in this way the correspondence claim is not only unusually reduced, but the
nature of the correspondence is left even more in the dark.
The natural question now is of course whether there are other candidates
for an interlevel correspondence theory of truth, explicating at least intuitions
(CI) and (C2). If there was a fully elaborated picture theory of truth (between
level II and I or C, or even between I and C), this would be a plausible
candidate. However, neither Wittgenstein's formulation of it, nor a modern
version of it, (Oddie 1987a), are easy to understand, let alone to evaluate with
respect to our question. Hence, let us turn to possible intralevel explications
of (CI) and (C2).
Truthlikeness theories (TL-theories) are primarily designed to explicate the
expression '~ is closer to the truth T than B" in an intralevel way, i.e., A, B
and T are assumed to be of the same level, I or II. TL-theories usually also
have a plausible definition of'~ is true (false)" in terms of T. Hence, it is clear
that TL-theories can, at the same time, be conceived as explications of the
correspondence intuitions (C 1) and (C2) as soon as one is prepared to localize
'the Jacts ' at the same level as A and B, of course by identifying them with 'the
truth'. In this way 'the facts' do not refer to unconceptualized, but to conceptualized facts. This brings us to the main claims of this section.
Claim 1:

The correspondence intuitions have to be interpreted in an


intralevel way, and TL-theories explicate them in principle.

Claim 2: lnterlevel interpretation of the correspondence intuitions is a


tempting misinterpretation and an easy confusion due to the
undeniable distinguishability of levels.

Notably, Popper fell explicitly a victim of this confusion. Being a champion


of an interlevel interpretation, he nevertheless presented his famous (defective)

194

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

intralevel definition of 'closer to the truth' also as a definition of 'better


correspondence to the facts' (Popper 1963a).
As we have seen at the beginning of the previous chapter, several new
definitions have been proposed, all circumventing the problem raised by
Popper's definition in one way or another. The following classification of the
new definitions gives a summary of their main characteristics:
Level-II or 'sentential' TL-theories (basic unit: sentences)
- based on logical (syntactic or semantic) consequences: Popper (1963a),
Newton - Smith (1981), Schurz/Weingartner (1987), Brink/Heidema
(1987)
- based on similarities between sentences: Tichy (1976, 1978), Oddie
(1986), Hilpinen (1976), Niiniluoto (1987a, partly)
Level-lor structuralist TL-theories (basic unit: structures)
- conceiving the truth as one structure: Miller (1978), Niiniluoto
(1987a, partly)
- differentiating between actual and nomic truthlikeness: Kuipers
(Chapter 7 of this book), Niiniluoto (1987a, partly)

As mentioned before, most of these approaches are presented and further


documented by their respective defenders in (Kuipers 1987a). It is clear from
the previous section that Level-I TL-theories, when presented in terms of
models, may fully or partly be reproduced as Level-II TL-theories, and vice
versa. Despite the variety of theories, our last and main claim to be defended
in this section will not come as a surprise.
Claim 3:

Of all intralevel interpretations of the correspondence intuitions, at


least as far as TL-theories are concerned, the structuralist TL-theory
differentiating between actual and nomic truthlikeness is the most
adequate. In particular, its basic, qualitative versions for actual and
nomic truth show in detail how false descriptions and theories lack
(full) 'correspondence to the facts' and how 'being closer to the truth'
increases the correspondence to the facts.

The specific part of this claim will be illustrated in the next subsection.
8.2.2. Actual and nomic truthlikeness as correspondence theory of truth

We start with an explicit first-order structuralist representation of a variant of


the electrical circuit of the previous chapter. The vocabulary is restricted to
two domain terms, viz., 'switch' and 'bulb', and two predicate terms, viz., 'being
on' and 'lighting', hence giving rise to the following set of conceptual
possibilities.
We repeat the definitions of actual and nomic truthlikeness. Actual truth likeness: let t (t E Mp) indicate the structure representing the actual state of the

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

195

Structures:
switches:
bulbs:
on-switches:
lighting bulbs:

x = <S, B, C, L)
S = {sl , s2, s3}
B= {bl , b2}

Mp

CS

LB

s2

Figure 8.3. Another electrical circuit

circuit (hence, according to Figure 8.3., the structure with C = {s2, s3} and L =
{b2}: the true description (representation) or the actual truth. The main
TL-question is: how to define "description y is closer to t than description x"?
Nomic truth likeness: let T (T Mp) indicate the set of structures representing
the nomically possible states of the circuit (hence, roughly: a bulb lights if and
only if it is in a connected path): the true theory or the nomic truth. That there
exists such a set is guaranteed by the Nomic Postulate. The main TL-question
is now: how to define "Theory Y, claiming that the set of structures Y is T, is
closer to T than theory X"?
In order to support Claim 3 we can confine ourselves to the comparative
notions of non-trivial actual truthlikeness and basic nomic truthlikeness.
Figure 8.4. and Figure 8.5. illustrate the actual as well as the nomic story. We
start with actual truthlikeness.
To begin with, the actual truth t is identified with 'the (actual) facts'. Let us
further assume that all conceptual possibilities have the same domains D j , but
that the relations R j defined on them may vary. (In Figure 8.4 and 8.5 F(Rj)
indicates the field of Rj, i.e., the Cartesian product of the relevant domains of
Rj). Recall that the claim "x = t" is associated with 'description x'. The following
actualist explications of the correspondence intuitions are now plausible:
(CI-D)

description x is true (false) if and only if


x does (not) correspond to t, defined by:
x = t (x *- t), i.e., (not) for all i: Rj(x) LI Rj(t) = 0

(C2-D)

y is closer to t than x if and only if


y corresponds better to t than x, non-trivially defined by:
for all i: Rj(y) LI Rj(t) Rj(x) LI Rj(t), and at least once a proper
inclusion

Figure 8.4. illustrates the false version of (CI-D), and Figure 8.5. illustrates
(C2-D), assuming in both cases that R j is one of the crucial relations. The
figures show clearly that and how '( better) correspondence' is explicated as a

196

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS


F(Rj)/Mp(V)

Figure 8.4. One or both *-areas are non-empty, representing that x does not correspond to
(assuming that R j is crucial), respectively, that X does not correspond to T

F(Rj)/Mp(V)
Rj(y)/y

Figure 8,5. One or both *-areas are non-empty, representing that y corresponds better to t than x
(assuming that R j is crucial), respectively, that Y corresponds better to T than X

matter of (more) overlap between sets (representing relations). (That the areas
have equal shape and size has no particular relevance.)
The above explication of 'closer to the actual truth' is said to be non-trivial,
as opposed to the trivial explication, according to which y is only closer to t
than x, if y = t and x # y.
Let us now turn to the nomic case. Now we identify the nomic truth T with
'the (nomic) facts'. It is important to notice once more the crucial difference
between the differentiated theory and all other TL-theories: we do not assume,
in a theory-directed context, that 'the truth' concerns one single structure
(representing the actual possibility/world) or, on the sentential level II, one
complete theory in the standard sense. Recall that the claim "X = T" is associated with 'theory X'. A weak notion of 'true/ false' arises when a theory is called
'true/false' when the 'hypothesis-claim' that T is a subset of X is true/false. A
strong notion arises when we take as criterion that the full claim of the theory,
"X = T", is true/false. In this strong interpretation, the explications become
similar to the actual ones:
(Cl-T)

theory X is true (false) if and only if


X does (not) correspond to T, defined by:
X = T (X # T), i.e., X LJ T = 0 (# 0)

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

(C2-T)

197

Y is closer to T than X if and only if


Y corresponds better to T than X, 'basically' defined by:
Y LJ T eX LJ T (proper inclusion!)

Note that Figure 8.4. also illustrates the false version of (CI-T), and that
Figure 8.5. also illustrates (C2-T), again showing that '( better) correspondence'
is a matter of (more) overlap of sets, now sets of structures.
Recall that the above explication of 'closer to the nomic truth' is said to be
basic, and is implicitly based on trivial actual truthlikeness, as opposed to
refined explications, which are based on non-trivial forms of actual truthlikeness. One gets Miller's TL-theory from basic nomic truthlikeness if one assumes,
in addition, that T only contains one structure, representing the actual world.
Coming back to the main topic of this section, the question is: what are we
to think of the presented intralevel explication of the correspondence intuitions?
To start with, we would like to conclude that, if it is a satisfactory explication,
then any notion of which the basic explication can be given in terms of the
set-theoretic difference is a kind of intralevel correspondence notion. But should
the general nature of an explication be an objection?
The main question seems to be this. The explication is question-begging for
it presupposes the statements "t is the true description" and "T is the true
theory", and the interlevel correspondence (possibly pictorial) is hidden in
them. However, our tentative answer is that the meaning of these statements
is only this: t and T are the (hypothetical) results of a faultless application of
the relevant representation conventions, given by (variants of) Tarski's truth
definition. Of course, if one wishes to formalize, and hence to work with two
abstract levels, it is desirable to have the representations on both levels such
that they mutually satisfy the coordination condition, i.e., Tarski's material
condition of adequacy or Aristotle's dictum. But this coordination is only a
matter of easy convention, where theoretical terms do not seem to form an
additional problem, as opposed to claims of (lack of) intralevel correspondence
and better correspondence. The latter claims form the substantial part of the
empirical sciences, where 'the facts' or 'the truth' form the unknown. Of course,
we may not expect to be able to establish such objective truths beyond any
reasonable doubt. We will, at most, be able to find intersubjectively applicable
truth criteria that lead to intersubjective truths, our best candidates for objective truths.
For the natural sciences at least, we would like to defend this. In the social
sciences it is doubtful whether it makes sense to talk about 'the facts/the truth',
in particular the nomic ones, even if we fix the conceptual means beforehand,
as has been presupposed throughout this section.
Referring to the survey of bifurcations of actual and nomic truthlikeness
given in the previous chapter, at the end of the next chapter we will briefly
come back, as suggested earlier, to the correspondence theory of truth in the
context of epistemologically stratified theories (Chapter 9). Moreover, it will

198

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

be noted in Chapter 10 that the given analysis of the correspondence theory


of truth, as far as the nomic truth is concerned, can easily be adapted for the
refined version of nomic truthlikeness.
8.3. EXPLICATING DIALECTICAL CONCEPTS

Introduction

In this section we will explicate a number of related dialectical concepts:


'dialectical negation', 'double negation', 'thesis-anti thesis-synthesis', and 'the
absolute'. We do not claim to have the last word on these concepts; on the
contrary, the only aim is to give the most basic elaboration of a new perspective
on them. In doing so, we have two purposes in mind. On the one hand, we
want to show to analytic philosophers that it is possible to make good sense
of dialectical phrases. In this respect our position is, as far as formal aspects
are concerned, quite opposite to the famous attack on dialectics by Popper
(1963b). On the other hand, we hope that our exercises have some appeal to
formally interested dialecticians such that they become motivated to undertake
refined explications that serve their purposes better. This second, limited purpose also explains why we will not attempt to support our explications by
citing quotations of classical dialectical philosophers. For the present, the only
important thing is to have some elementary idea of the kinds of intuitions that
seem to be involved.
The new perspective on dialectical concepts is again provided by the structuralist approach to the problem of truthlikeness. As has become clear, the
structuralist theory of truthlikeness has generated a number of unexpected and
unintended results. The most important ones are:
- a justification of fundamental methodological rules, in particular the
rule to choose the most successful theory (Subsection 7.3.3.),
- an explanation of the success of the natural sciences in terms of truth
approximation (Subsection 7.3.3.),
- an explanation and justification of the non-falsificationist behavior of
scientists (Section 6.2., 6.3. and 7.3.),
- a corrective explication of so-called 'inference to the best explanation'
(Subsection 7.5.3.),
- an explication of Popper's bad luck, i.e., the convincing failure of his at
first sight very plausible definition of truthlikeness (Section 8.1.),
- an explication of the correspondence theory of truth as an intralevel
intuition (Section 8.2.).
However, the most unintended result is an explication of dialectical concepts,
and even that is suggested by the structuralist theory of truthlikeness.
As we saw, the structuralist point of view applied to the problem of truthlikeness immediately leads to the recognition that there are essentially two problems

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

199

of truthlikeness: actual truthlikeness, related to the actual world, represented


by one particular structure, and nomic truthlikeness, related to the set of nomic
possibilities, represented by a set of structures. Hence, there is an actual and a
nomic perspective on dialectical concepts. It turns out to be possible to explicate
the main dialectical concepts in both perspectives in three different ways.
More specifically, in both perspectives there are formally resembling
(epistemo-)logical readings, as well as formally resembling ontological readings
of the notions. First we deal with those of the actual perspective, then with
those of the nomic perspective. Moreover, in both perspectives there are, in
addition, interesting methodological readings which mutually differ in a fascinating way.
8.3.1. Actual dialectics

Recall the problem of actual truthlikeness: what does (can) it mean that one
description is closer, or more similar, to the true description of the actual world
than another description. Here we will restrict our attention to the solution of
that problem, dealt with in Section 7.1., for propositional structures, or propositional constituents. These constituents are conjunctions of negated and
un-negated elementary propositions belonging to a fixed finite set: EP =
{pI, p2, ... , pn}. A constituent can be represented by the set of its un-negated
elementary propositions. Let x, y, etc. indicate subsets of EP representing
constituents in this way.
When the claim of a description x, viz., "x = t", where t represents the true
description, is true or false with respect to a particular elementary proposition,
that proposition is called a 'match' or 'mistake' of the description, respectively.
Hence, the set of matches of x is (x n t) U (x n i), and the set of mistakes is
(x - t) U (t - x). The definition of y is closer to the truth t than x, (Di) and
(Dii), Subsection 7.1.2., amounted to the claim that the set of matches of y
properly includes that of x or, equivalently, that the set of mistakes of y is a
proper subset of that of x.
8.3.1.1. Dialectical approach of the actual truth: logical version

Let us start with one of the main dialectical concepts. The idea of dialectical
negation seems to be composed of at least three partial intuitions: contradicting,
incorporating and superseding what is negated. Note that 'superseding what is
negated' implies that what is negated has to be false or at least cannot be the
whole truth. This implication might be considered as a separate intuition, but
we will neglect it further because it is implied by the three basic intuitions. We
will now specify the basic intuitions in terms of statements. When statement
S* is a dialectical negation of S, S* contradicts S, it incorporates what was
good of S and supersedes S in this respect.
Descriptions, as interpreted above, are statements, due to their claims. Now
it is easy to check that when y is closer to the truth t than x, they (i.e., their

200

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

claims) contradict each other and y is an improvement of x in the sense that


y incorporates all matches of x and that y supersedes x by having some extra
matches (where x still makes mistakes). As a consequence, 'closer to the actual
truth' is one plausible explication of dialectical negation.
One possible objection to this explication is that there will be, in general,
more than one description closer to the truth than a given one, whereas
dialectical negation may have the connotation of uniqueness and that, for this
reason, the explication is more appropriate for the related idea of dialectical
correspondence. However this may be, it is important to note that most descriptions are not closer to the truth than a given one and that there is room for
supplementary intuitions. Moreover, in the context of refined truthlikeness,
Chapter to, we will encounter a possibility for a more sophisticated explication
of 'dialectical correspondence'.
In the line of the foregoing, it is now also plausible to explicate the idea of
double negation as the combination of two successive actual dialectical negations, leading to a stepwise improvement of descriptions: x ~ y ~ Z, i.e., from x
to Z with y as intermediate. It is easy to check that dialectical negation is an
asymmetric and transitive relation. As a result, the present explication of double
negation of the triple (x, y, z) could also be interpreted as an asymmetric
explication of the famous triad: thesis, antithesis, synthesis, for the triple
(y, x, z) would not satisfy the double negation conditions.
In our opinion, the following symmetric explication of thesis-antithesis-synthesis is more plausible. Suppose that x and yare incomparable in the sense that
neither of the two is closer to the truth than the other, implying not only that
they contradict each other, but also that both have 'pros and cons' with respect
to each other. Suppose further that z is closer to the truth than, and hence a
dialectical negation of, x as well as y. Consequently, z incorporates all common
and non-common merits of the two such that it improves upon both. This
seems the core of the triad: the synthesis contradicts, incorporates and supersedes thesis and antithesis, which contradict each other as equals. Hence, the
present explication is symmetric between x and y, for (x, y, z) satisfies the
conditions if and only if (y, x, z) satisfies them, apart from a possible implicit
time order.
At the end of this 'first dialectical round' it is tempting to qualify 't' as the
absolute, although it is relative to the conceptual frame. A metaphysical realist,
who assumes that there is, for some or for all scientific contexts something like
an ideal or optimal conceptual frame, with corresponding actual truth, the
absolute is really absolute when Mp is generated by the relevant optimal
conceptual frame. In that case, the explication of 'the absolute' would realize
the absolute connotation it certainly has. Of course, the burden of proof is on
the metaphysical realist.
Notice, for later purposes, that the explications of truthlikeness and dialectics
of descriptions can be relativized by replacing t by some z, and hence by the
conditional statement: suppose that z is the true description, i.e., suppose z =
t, then y is closer to the truth than x or y is a dialectical negation of x iff ....

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

201

Given the restriction of the foregoing explication to knowledge claims about


the actual world, this explication may be called epistemological dialectics with
respect to the actual world. However, to evaluate the fundamental statement
that one description is closer to the actual truth than another presupposes that
we know the true description. Hence, the explication is primarily a logical
exercise, for in scientific practice we do not know the true description. When
we are describing the actual world, e.g., the conditions and outcome of an
experiment, we are aiming at the true description of the relevant actual world.
In Subsection 8.3.3.1. we will introduce a related alternative to the present
epistemo-logical explication which has not only logical but also methodological
relevance.
8.3.1.2. Dialectical development of the actual world

Dialectics is said to apply not only to knowledge claims (above and below)
and concepts (below), but also to reality, in particular to the development of
the actual world, which might be called ontological dialectics of the actual
world. Of course, if the actual world changes, the actual truth changes as well.
It is now plausible to focus on successive actual states of the world and to
attach indices, indicating successive moments of time, to the momentary true
representation t, leading to: t1, t2, ....
We begin by relativizing all judgments to an arbitrary later state. In the light
of t3, t2 is a dialectical negation of t1 iff t2 is more similar to t3 than t1. In
the light of t4, t3 is a double negation of t1 with t2 as intermediate iff t2 is
more similar to t4 than t1 and t3 is more similar to t4 than t2. In the light of
t4, <t1, t2, t3) is a (symmetric) thesis-antithesis-synthesis-triad iff t3 is more
similar to t4 than t 1 as well as t2, whereas t2 is not more similar to t4 than t1,
nor the converse.
These qualifications of developments acquire an absolute character of a
strongly metaphysical nature when there is supposed to be something like a
final state, the absolute, in the light of which all judgements are made, even if
one does not believe in an ideal conceptual frame. It gets a still stronger
absolute character when it is supposed, in addition, that the conceptual frame
is ideal.
8.3.2. Nomic dialectics

Recall that the second problem of truthlikeness concerned the idea that one
theory is closer, or more similar, to the true theory about what is nomically
possible than another theory. Unlike the situation in the actual perspective,
there was in the nomic perspective no reason to restrict the exposition to a
particular kind of structure. We only have to assume the Nomic Postulate,
that is, for a given domain and Mp, there is a unique subset T of nomic
possibilities. X, Y, etc. indicate again subsets of Mp.
When the claim of a theory X, viz., "X = T" is true or false with respect to

202

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

a particular conceptual possibility, this possibility is called a (nomic) 'match'


or 'mistake' of that theory, respectively. It is again easy to check that in this
terminology, theory Y is closer to the truth than X, according to the basic
definition (Bi) and (Bii) of Subsection 7.2.2., iff the set of matches of Y properly
includes the set of matches of X or, equivalently, iff the set of mistakes of Y is
a proper subset of that of X.
8.3.2.1. Dialectical approach of the nomic truth: logical version

The explication of dialectical notions in the nomic perspective, now in terms


of nomic matches and mistakes, is completely analogous to the epistemological
explication of dialectical concepts in the actualist perspective. We merely state
the results: Y is a (nomic) dialectical negation of X iff Y is closer to the truth
than X, Z is a double negation of X with Y as intermediate iff Y is closer to
the truth than X and Z is closer to the truth than Y. Z is a synthesis of the
thesis X and the antithesis Y iff Z is closer to the truth than X as well as y,
but Y is not closer to the truth than X, nor the converse. Finally, T is the
conceptually relative, or even the conceptually ideal, absolute.
If dialecticians suggest that the scientific enterprise is a dialectical process,
then the just-presented nomic notions may be more appropriate to characterize
this process than the actualist notions. It is also in this perspective that the
dialectics of statements can be seen to be also a dialectics of concepts. For a
theory X is supposed to be specified as a subset of conceptual possibilities:
those members of Mp that satisfy certain conditions. In other words, X can
also be conceived as the extension of a set-theoretic predicate and hence as the
extension of a concept. Consequently, the dialectics of theories can also be
conceived as a dialectics of concepts.
Notice again for later purposes that the explication of truthlikeness and
dialectics of theories could have been relativized by replacing T by some Z,
and hence by the conditional statement: suppose that Z is the true theory, i.e.,
suppose Z = T, then ... etc.
The explications are again (epistemo-)logical in the sense that the evaluation
of the crucial statement that one theory is closer to the (nomic) truth than
another presupposes that one knows the true theory, which is far from scientific
practice. However, in Subsection 8.3.3.2. we will show that there is again a
methodologically relevant alternative.
8.3.2.2. Dialectical development of the nomic world

Formally, it is no problem to formulate analogous to the actual case an


ontological variant of the epistemological explication of the nomic perspective.
Again, when the nomic world changes, the nomic truth changes as well. Suppose
that T, i.e., the set of nomic possibilities, changes in time, leading to the sequence
Tt, T2, ... . of subsets of Mp.
Again, we begin by relativizing all judgments to an arbitrary later moment.

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

203

In the light of T3, T2 is a dialectical negation of Tl iff T2 is more similar to


T3 than Tl. In the light of T4, T3 is a double negation of Tl with T2 as
intermediate iff T2 is more similar to T4 than Tl and T3 is more similar to
T4 than T2. In the light of T4, <Tl, T2, T3) is a thesis-antithesis-synthesistriad iff T3 is more similar to T4 than Tl as well as T2, whereas T2 is not
more similar to T4 than Tl, nor the converse.
These qualifications of developments acquire again an absolute, conceptually
relative, or even conceptually ideal, character when there is supposed to be
something like a final set of nomic possibilities, the absolute, in the light of
which all judgements are made.
An important problem with both ontological readings in the nomic perspective is that they presuppose that the set of nomic possibilities changes. If one
equates 'nomically possible' with 'physically possible' this implies the assumption that the laws of nature may change, a very problematic assumption indeed.
However, as mentioned before, in several contexts it makes good sense to
distinguish perspectives of nomic possibility, as e.g., suggested by the following
sequence of 'lower' to 'higher' levels: the physical, chemical, biological, psychological, cultural-socio-economical level. Being nomically possible at a higher
level then implies being nomically possible at a lower level, but not the converse.
Another example concerns the idea of nomically possible states of an artifact,
e.g., the electrical circuit in your home, assuming that it remains intact, which
means a severe restriction to its physically possible, including broken, states.
Assuming such a hierarchy of levels of nomic possibilities, it may make sense
to assume that the set of nomic possibilities at some higher level may change
without the nomic possibilities at the lower levels changing. For this line of
thought, it is only necessary that in stable periods the boundaries between the
different levels of nomic possibility can be defined more or less sharply.
8.3.3. Dialectical methodology

The epistemological explications of truthlikeness and dialectics of descriptions


and theories were a priori or logical in the sense that the relevant judgements
presupposed knowledge of the (whole) actual and the nomic truth, respectively.
For both perspectives, there are methodological alternatives which do not
presuppose knowledge of the whole truth. At successive stages, an increasing
part of the truth is supposed to become known, in terms of which the success
and the problems of a description or theory are defined, as well as the comparative judgment that one description or theory is more successful than another.
Recall that the comparative success judgements are methodologically very
relevant, for they guide the choice between descriptions and theories, respectively. Hence, the corresponding methodological explications of dialectics may
even be more appropriate for the scientific process.
8.3.3.1. Dialectical approach of the actual truth: methodological version

For the actual perspective, we will again restrict our attention to propositional
descriptions. Recall that explicit descriptions were supposed to be given in total

204

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

and that we assume the idealization of an infallible meta-position with respect


to the data in order to evaluate descriptions. Recall also that the definition of
"y is more successful at a certain time than x" amounted to the claim that
n n y is a subset of n n x and p - y is subset of p - x, and at least once a proper
subset. Here p and n indicate the (mutually exclusive) sets of elementary
propositions of which it has been established at the given time that its truthvalue is 'true' (positive) or 'false' (negative), respectively. Together p and n form
the available data at the time. Recall, finally, that the 'success' of a description
is a symmetric notion in the sense that it is completely symmetric with respect
to true and false elementary propositions.
Of course, we now suggest explicating the dialectical concepts in terms of
the notion 'more successful'. In this way, we obtain historized explications of
these concepts, for their applicability becomes data and hence time relative. In
the light of the data pin at a certain time, y is at that time a dialectical negation
of x iff y is more successful than x at that time. It is easy to check also that,
according to this data-relative comparative statement, y contradicts x and
incorporates and supersedes what is good of it. The main consequences of the
data-relative character are that y does not always need to have superseded (the
success of) x before the given time (it may have had equal success for some
time, but not less), and that y does not need to be able to incorporate the
future success of x. All this results from the fact that the qualification 'more
successful' simply depends on the (growing) momentary data-sets.
A data-relative double negation is again a succession of two data-relative
dialectical negations, where the moment of the first may not now be later than
that of the second. For it is not at all guaranteed that the third description
was more successful than the first at the first moment. In the otherwise plausible
data-relative (symmetric) explication of the thesis-antithesis-synthesis-triad,
the relevant moments have to succeed each other in the same order.
In the present methodological version of the actualist perspective there is no
explicit place for the absolute. But the unknown actual truth t functions as the
implicit goal: a data-relative dialectical negation of a description has a chance
to be a dialectical negation in the (epistemo-)logical sense. But it need not be,
with the consequence that a data-relative dialectical negation need not have
to retain this status in the light of new evidence. However, the situation changes
if the data- relative dialectical negation is indeed closer to the truth, and hence
is also a dialectical negation in the logical sense. If so, it always had to be at
least as successful and it will remain more successful; hence it will retain its
status as a data-relative dialectical negation in the light of all possible new
evidence.
Of course, if the conceptual frame is supposed to be ideal, the implicit goal
t is the absolute in a conceptually non-relative sense.
8.3.3.2. Dialectical approach of the nomic truth: methodological version

Far from being known, the nomic truth T is, in theory directed empirical
sciences, the 'great unknown'. We have seen in the previous chapter that

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

205

theories are primarily evaluated in an asymmetric way, i.e., in terms of their


capacity to respect the realized possibilities R(t) and to entail the established
empirical laws, hence the strongest one S(t). Hence, R(t) and S(t) now form
the data with which a theory is confronted at time t and we again assume the
idealization that we are infallible as far as the data are concerned, i.e., that
actual mistakes have not been made and that all accepted laws are true: hence,
R(t) is a subset of T and T of S(t), whatever T precisely is. In line with this,
we assume the idealization that R(t) can only grow in time and that S(t) can
only shrink.
Recall the definition that Y is more successful at time t than X: R(t) - Y is
a subset of R(t) - X and Q(S(t)) - Q(Y) is a subset of Q(S(t)) - Q(X), and at
least one proper subset relation. The first clause could be paraphrased as the
claim that the individual problems of Y form a proper subset of those of X.
The second clause amounts to the claim that all established laws not entailed
by Yare also not entailed by X.
We now suggest explicating the dialectical concepts in terms of the nomic
notion of 'more successful'. The verbal similarity with the corresponding actualist explications hides the important difference due to the asymmetric character
of the present notion of 'more successful'. In this way we get again historized,
i.e., data and hence time relative, explications of these concepts. In the light of
R(t)/ S(t), Y is a dialectical negation of X iff Y is more successful than X at t.
It is easy to check also that according to this data-relative comparative statement Y contradicts X and incorporates and supersedes its success. The main
consequences of the data-relative character are that Y does not always need
to have superseded (the success of) X before t (it might have been equal for
some time, but not less), and that Y does not need to be able to incorporate
the future success of X.
A data-relative double negation is again a succession of two data-relative
dialectical negations, where the moment of the first may not now be later than
that of the second. For it is not at all guaranteed that the third theory was
already more successful than the first at the first moment. In the otherwise
plausible data-relative (symmetric) explication of the thesis-antithesis-synthesistriad the relevant moments have to succeed each other in the same order.
Recall that the principle of dialectics (PO) of Subsection 6.1.3. suggested aiming
at a, success preserving, synthesis of two theories with divided success, that is,
in the present terms, a synthesis that is a dialectical negation of a thesis and
an antithesis.
In the present methodological version of the nomic perspective there is no
explicit place for the absolute. But the unknown nomic truth T now functions
as the implicit goal: a data-relative dialectical negation of a theory has a good
chance of being a dialectical negation in the (epistemo-)logical sense. But it
need not be, with the consequence that a data-relative dialectical negation need
not have to retain this status. However, as is easy to check, if the data-relative
dialectical negation is indeed closer to the truth, and hence is also a dialectical

206

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

negation in the logical sense, then it has always to be at least as successful and
it will remain more successful and hence retain its status as data-relative
dialectical negation in the light of all possible new evidence.
Again, when the conceptual frame is supposed to be ideal, the implicit goal
T is the absolute in a conceptually non-relative sense.

Concluding remarks
We start with some remarks about our explication of dialectics, leading to
some general points. Since the structuralist theory of truthlikeness can be made
more sophisticated in several ways, in particular epistemological stratification
and refinement (Chapter 9 and 10), such sophistication may well lead to
refinements of the proposed explications of dialectical concepts. In Chapter 10
we will argue that (structuralistically represented) 'concretization' can be seen
as a special, refined kind of dialectical negation or correspondence, in harmony
with the corresponding claims of Izabella Nowakowa ( 1974) and Leszek Nowak
(1977). However, what has been presented will be enough for formally interested
dialecticians to decide whether the structuralist theory of truthlikeness can help
them to clarify their concepts.
One objection we can predict already. Although, as mentioned, the dialectics
of theories can be reinterpreted as the dialectics of concepts, dialecticians will
argue that the dialectics of concepts must involve more, in particular the
development of what we called the conceptual possibilities constituting the
conceptual space, which we kept constant. As a matter of fact Thagard (1982)
deals with this problem from the general structuralist point of view (not
specifically related to truthlikeness). The fact that we had several technical
hesitations about Thagard's paper inspired and motivated us very much to
design explications within a fixed conceptual space.
Of course, in general, assuming a fixed conceptual frame is a strong idealization, about which we would like to make some closing remarks. T was introduced as the set of nomic possibilities. In standard structuralist terms, this set
is one possible specification of the (unstratified, basic version of the) so-called
set of intended applications or intended domain. T is a subset of Mp and hence
represents the intended domain within the conceptual means of Mp, i.e., it is
the intended domain 'as seen through Mp'. It is evident that Mp is man-made,
hence T, being Mp-dependent, is man-made. Hence we subscribe to a fundamental form of conceptual relativity and constructivism. But this need not imply
an extreme form of relativism: claims of theories and hypotheses are objectively
true or false, for their truth or falsehood depends on the natural world.
On the other hand, we see that the objective character of claims does not
imply that the intended domain and the conceptual frame (Mp), and hence the
nomic truth T, are fixed beforehand, and that the only task that remains is to
formulate a subset X of Mp leading to a true theory claim. As a matter of fact,

INTUITIONS OF SCIENTISTS AND PHILOSOPHERS

207

in practice, the determination of the intended domain, the conceptual space,


and a subset X is a complicated interaction process, guided by the desire to
formulate informative and true theories and hypotheses. This process evidently
has a dialectical character in the common-sense meaning. Unfortunately, it
seems difficult to discern general patterns in this interaction process without
making some important idealizations, such as fixing the conceptual framework.
However, in the next chapter we will see that it is possible to define the idea
that one vocabulary is better in referring to theoretical entities and properties
regarding a certain domain than another. Hence, variable conceptual frames
can be taken into account. In the next chapter and the final chapter, we will
come back to the possibility of an ideal vocabulary for a certain domain.
Finally, variable domains can also be taken into account, where the main
changes concern extensions and restrictions. We will not study this issue, but
see (Zwart 1998, Ch. 2-4) for some illuminating elaborations in this connection,
among other things, the way in which strengthening/ weakening of a theory
and extending/reducing its domain interact.

EPISTEMOLOGICAL STRATIFICATION OF NOMIC


TRUTH APPROXIMATION

Introduction
Thus far it might seem that our conceptually relative point of departure leads
to an extreme form of relativistic (nomic) realism. However, this would only
be the case if we excluded constraints between the truths generated by different
vocabularies for the same domain. In this chapter we will deal with the relation
between an observational and a theoretical (cum observational) level, generated
by an observational and an encompassing theoretical vocabulary for the same
domain. Here the distinction between observational and theoretical components is, of course, assumed to be not of the classical, absolute form but of a
sophisticated, relative kind. In particular, the distinction is supposed to reflect
as much as possible what scientists consider as observable versus nonobservable/theoretical, including the fact that the distinction changes when
theories are accepted and become observation theories. However, here we
assume a temporarily fixed distinction. It will generate not only different kinds
of (basic) truthlikeness, but also several relations between them.
Besides the intralevel Nomic Postulate, we will introduce two interlevel
postulates, the Truth Projection Postulate and the Fiction Postulate. 1 On their
basis it will be possible to prove the Projection Theorem stating that, under
certain conditions, more truthlikeness on the theoretical(-cum-observational)
level is projected on the observational level. In other words, more theoretical
truthlikeness implies more observational truthlikeness. Together with the old
Success Theorem, according to which more observational truthlikeness implies
more successfulness, we get the corollary that, under the same extra conditions,
more theoretical truthlikeness implies more successfulness. In the reverse direction we get, by a combination of the old Forward Theorem and the new
Upward Theorem, suggested by the Projection Theorem, that RS and hence
HD-evaluation is functional for theoretical truth approximation.
We will first present the Projection and the Upward Theorem in the suggested
straightforward sense. It will turn out to be possible to define, on the basis of
the Nomic Postulate applied to a given vocabulary and domain, which theoretical terms really refer and which do not. The two new theorems can then also
be derived in two steps. One between observational truthlikeness and truthlikeness on the level of referring terms, called 'substantial truthlikeness', and the
other between substantial truthlikeness and theoretical truthlikeness.

208

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

209

As suggested, the two new postulates enable a plausible definition of reference


of theoretical terms, leading to plausible definitions of the referential truth, the
referential claim of a theory, and the relation of more referential truthlikeness
between two theories. Though there are strong relations between successfulness
and observational, substantial and theoretical truthlikeness, there is not such
a strong relation between any of these, on the one hand, and referential
truthlikeness on the other. For this reason, besides theoretical arguments,
experimental arguments in favor of reference claims will be discussed.
Section 9.1. deals with theoretical and substantial truthlikeness and truth
approximation. In Section 9.2. we will define referential truthlikeness and the
prospects for referential truth approximation, including the role of experimental
criteria for accepting and rejecting specific reference claims. In Section 9.3. we
will extend the corrective analysis of inference to the best explanation, introduced in Section 7.4. Moreover, we will formulate speculations about commensurability and ideal languages, and extrapolate the analysis of Section 9.1. and
9.2. to the comparison of vocabularies and research programs. We will explicate
the idea of theoretical or explanatory research programs. In Section 9.4. we
will review the main epistemological positions and conclusions. Finally, we will
briefly extend the analysis of intuitions of scientists and philosophers presented
in Chapter 8 and recapitulate the main reason for a further refinement of the
analysis, in order to cover real life scientific examples.
It is important to stress in advance that, like in Chapter 6-8, and unlike in
Chapter 5, we will neglect all complications that derive from auxiliary hypotheses and the like which playa role in the derivation of consequences of theories.
Though such complications frequently playa vital role, they do not seem to
affect the main lines of stratified nomic truth approximation.
9.1. THEORETICAL AND SUBSTANTIAL NOMIC TRUTH
APPROXIMATION

Introduction

Although the basic definition of nomic truthlikeness can be applied to theories


within any vocabulary, claims to truth approximation of the unknown nomic
truth corresponding to that vocabulary and the given domain have to relate
truthlikeness to observational success in one way or another. In this respect,
Chapter 7 only dealt with observational truth approximation. In particular,
the Success Theorem disclosed that being closer to the observational truth
guarantees being at least as successful, and the Forward Theorem revealed that
'so far proven to be more successful' supports being closer to the observational truth.
In the present context of theoretical vocabularies and sub-vocabularies, the
projection function n will playa crucial role. It is the function from the set of
conceptual possibilities Mp(V) of some vocabulary V onto that set Mp(V' ) of

210

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

a sub-vocabulary V' in the sense that, for all x in Mp(V), n(x) is the result of
dropping the extra components in x and the structure clauses using them.
Technically speaking, this means that n(x) is a substructure of x belonging to
Mp(V').2

In the context of n, it will always be clear what V and V' are. <v, Mp(V) is
also called a (conceptual) level, and a superlevel of <V', Mp(V'), whereas the
latter is called a sublevel of the former. For a subset X of Mp( V), nX indicates
the set of projections of all members of X, being a subset of Mp(V'), and will
be called the V'-projection of X. Conversely, for a subset X of Mp(V'), n- I X
is the subset of members of Mp(V) with projections in X and is called the Vreproduction of X. Note that n-lnX is always a superset of X, and mayor
may not coincide with X. Both n{x} = {n(x)} and n(x) itself are called the V'projection of x, and n- 1 {x} the V-reproduction of x (note that n-I(x) has not
yet been defined, hence it might be equated with n - 1 {x}).
9.1.1. Theoretical truth approximation

Let there be given a domain of interest D. We will first study the relation
between an observational (0- ) level and a theoretical (t- ) level for D. In the next
subsection we will take the intermediate, referential (r-)level of really referring
terms into account. Let Vt indicate the richest vocabulary of non-logicomathematical terms that playa role in one of the theories involved, that is, all
relevant observational and theoretical terms which are in consideration. Let
Vo indicate the subset of Vt of observation terms.
Let Mp(Vo) and Mp(Vt) indicate the set of conceptual possibilities on the 0and t-Ievel, respectively. Applying the Nomic Postulate on both levels, these
sets generate, in combination with D, unique, time-independent subsets of
nomic possibilities on each level, called the observational (nomic) truth To and
the theoretical (nomic) truth Tt, respectively. The term 'nomic' will from now
on be omitted. Moreover, in this chapter truthlikeness of theories on any level
will always refer to the basic notion defined in Chapter 7.
We will assume that nTt = To, the Truth Projection (TP- ) Postulate (from
the t-Ievel to the o-level), indicated by TPP(t, 0). That nTt is a subset of To
is a semantic fact, for if a conceptual possibility on the t-Ievel is nomically
possible, it will remain nomically possible when one skips some components.
The motivation for the substantial side of the TP-Postulate, viz., To is a subset
of nTt, will be postponed till the introduction of the intermediate referential
level. Although we do not see how TPP(t, 0) could be violated, we will nevertheless mention TPP(t,o) in proof sketches when the substantial side is
presupposed.
The important question is, of course, whether not only the truth but also
truthlikeness on the t-Ievel is projected on the o-level, that is, assuming that X
and Yare subsets of Mp(Vt), does MTL(X, Y, Tt), i.e., Y is more truthlike than
X or, more precisely, Yis at least as close to Tt as X, imply MTL(nX, nY, nTt),

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

211

Mp(Vo)
empty ?

Figure 9.1. Theoretical truthlikeness (shaded areas Mp(Vt) empty) implies external observational
truthlikeness (shaded a rea Mp(Vo) empty), but not internal (?-area may be non-empty)

and hence, using the TP-Postulate, MTL(nX, nY, To)? Figure 9.1. depicts the
formal situation.
On the 'external side' we have to prove that the external (law) clause of
truthlikeness on the t-level implies that on the o-level. Hence, using the (Bi)'version of the external clause, to be proved is that the emptiness of Y - X U Tt
implies that of nY -nX U nTt, which immediately follows from the fact that n
is a function. Moreover, the semantic side of the TP-Postulate guarantees that
nY -nX U To is empty when nY -nX U nTt is.
A similar unconditional proof is impossible for the 'internal side', essentially
due to the many-one character of the projection function. We will first characterize the gap and then specify a sufficient condition to fill that gap. Suppose that
the internal clause of "Y is at least as close to Tt as X" holds, that is, according
to (Bii)', suppose that X n Tt - Y is empty. What we would like to prove is
that nX n To -nYis empty. This would immediately follow from the emptiness
of nX n nTt -nY, by the substantial side of the TP-Postulate, i.e., To is a
subset of nTt. Hence, from this perspective, it remains to be proved that
nX n nTt - n Y is empty.
What is the nature of possible elements of nX n nTt - nY? (Bii)' excludes
that they are projections deriving from X n Tt - Y. Since members of
X n Tt - Y would be extra 'theoretically correct' models of X , we might say
that X has no extra observationally correct models deriving from theoretically
correct models. If nX n nTt -nY is nevertheless non-empty, its members must

212

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

derive from extra theoretically incorrect models of X, i.e., members of


X - Tt U Y, which are accidentally observationally correct. More precisely, let
x in X - Tt be called accidentally observationally (0- )correct when n(x) belongs
to nTt, hence to nX n nTt. If nX n nTt - n Y is non-empty, and assuming
(Bii)', it must be due to accidentally o-correct members of X not belonging to
Y. Hence, necessary and sufficient to guarantee projection, to fill the gap, is to
assume that all accidentally o-correct members of X are also (accidentally
o-correct) members of Y. But how problematic is it for Y to miss such accidentally o-correct models? It is not problematic at all; quite the contrary. If Y
misses an accidentally o-correct model of X, this is progress on the theoretical
level. It is even so that these considerations directly lead to a further relativization of falsification. Let n(x) for x in X be a member of nTt, but let x not
belong to Y, hence n(x) is a potential counter-example to nY (hence to Y).
Now a defender of Y may well claim that an established extra counter-example
to n Y (hence to Y) is not very telling, because it may well be an accidentally
o-correct model of X, even such that Y can nevertheless be closer to Tt than
X. In other words, Y may well be closer to Tt than X even if n Y is only
externally and not internally closer to nTt than nX.
Let us nevertheless look at a sufficient condition to fill the gap. Suppose that
there are counter-examples to the claim that nX n rrTt -nY is empty. Let C
belong to rrX n nTt - n Y. Now it is not difficult to derive that there have to
be x in X - YU Tt and z in Tt - X U Y such that n(x) = n(z) = c. This <x, z>tuple provides a counter-example to the following substantial claim:
for all x in X - Tt and z in Tt - X: if n(x) = n(z)
then there is y in X n Tt such that n(x) = n(y) = n(z)
For trivial reasons we may replace, in this claim, 'X - Tt' by 'X' and 'Yt - X'
by 'Yt'. Hence, it can be reduced to the claim that nX n nTt is a subset of
n(X n Tt) and, therefore, together with the logico-mathematical fact that the
converse inclusion is automatically the case, we get the claim: n(X n Tt) =
nX n nTt. Hence, a general, Y-independent way, of excluding counter-examples
to the claim that nX n nTt - rr Y is empty, is by assuming that X is relatively
correct in the sense that it satisfies the condition
RC(X)

n(X

n Tt) = nX n nTt

which will sometimes be indicated by RC(X, t, 0), to indicate the relevant levels.
Note that RC(X) has been defined such that it is independent of the
TP-Postulate. The name of the assumption is based on the fact that it precisely
guarantees that (relative to rrTt) correct models of nX can be extended to
(relative to Tt) correct models of X. In other words, it guarantees that X has
theoretically correct reproductions of all its o-correct models. How strong is
RC(X)? To be sure, it means something like: X is on the right track, with the

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

213

important relativization, as far as Vt - Vo is concerned. But apart from this, it


is not a very strong condition. First, RC(X) has now been defined for the
projection of theory X from the t-level on the o-level, RC(X, t, 0). Later it will
be argued that its 'intermediate' projection on the level of referring terms is
trivially relatively correct in the relevant sense. Second, it is important to note
that the antecedence of the substantial part of RC(X) requires n(x) = n(z) for
x in X and z in Tt, that is, it requires a perfect fit between a model at the
o-level and the relevant projected nomic possibility of the t-level. The occurrence of such perfectly fitting observational pairs will, in general, be rather
exceptional. (Of course, it is even more unlikely to have perfectly fitting theoretical models.) But if X has no perfectly fitting observational models, RC(X) is
not only trivial, but the internal observational clause itself is as well. Third, as
soon as we take refined versions of truth approximation into account (see the
next chapter) perfect fit even is more or less excluded. Hence, RC(X) is not a
very strong assumption about X, apart from the fact that it is on the right
track as far as Vt - Vo is concerned.
In sum, assuming the TP-Postulate, we have proved the
Projection Theorem:

if Y is at least as close to the theoretical truth Tt as X, and if X is


relatively correct, then n Y is at least as close to the observational truth
To as nX,
formally: MTL(X, Y, Tt) and RC(X) imply MTL(nX, nY, To)
Recall that the Success Theorem amounts, in the present context, to the
claim that "nY is always at least as successful as nX" (MSF(X, Y, R jS)), we no
longer make the time-relative nature of Rand S explicit) is guaranteed by "n Y
being closer to To than nX". Hence we obtain the following corollary from
the combination of the Projection Theorem and the Success Theorem.
Success Corollary:

if Y is at least as close to Tt as X, and if RC(X), then Y is always at


least as successful as X
formally, MTL(X, Y, Tt) and RC(X) imply MSF(X, Y, R /S)
Hence, more successfulness not only d-confirms more observational truthlikeness but, on the condition that the less successful theory is relatively correct,
it also d-confirms more theoretical truthlikeness, according to the theory of
deductive confirmation exposed in Chapter 23.
Recall also that the Success Theorem suggested the rationale behind RS,
and hence behind HD-evaluation, viz., the Forward Theorem. According to
this theorem, the hypothesis that Y always remains more successful than X
implies that Y is closer to the observational truth. The consequence is that the
fact that a theory has so far proven to be more successful is, if not due to

214

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

observational truth approximation, due to a lucky evaluation history thus far;


a consequence which applies to success dominance in general. Now it is easy
to prove, in addition, the following
Upward Theorem:
if nY is closer to To than nX then y* is closer to Tt than X*, where

X* =deJX - {x E X - Yjn(x) E nyn To}


y* =deJX n Y

Note that both X* and y* are strengthenings of their corresponding departure


theory. Note also that X* might become equal to Y* , but this will not be the
rule. When they are unequal, the Upward Theorem amounts to the claim that
(assuming TPP(t,
an example of theoretical truth approximation can always
be construed starting from a case of observational truth approximation.
This theorem leads, together with the Forward Theorem, to the claim that
RS, and hence HD-evaluation, are functional for approaching the observational
truth, which in its turn is functional for approaching the theoretical truth. The
'forward step' of this claim has already been elucidated. Suppose now that we
conclude in the first step that we have a case of observational truth approximation. The 'upward step' requires additional elucidation. As far as the
TP-Postulate and RC(X) hold, it follows from the Projection Theorem that
the theoretical truth approximation hypothesis (TA-hypothesis on the t-level')
can explain the observational truth approximation claim, and can hence explain
the success dominance in a deeper way. To be precise, if we have a counterexample to the (internal clause of the) claim that n Y is closer to To than nX,
there are three mutually exclusive possibilities: it may be (or better, generate)
a counter-example to the (internal clause of the) claim that Y is closer to Tt
than X, or to the TP-Postulate, or to the claim that X is Relatively Correct.
In the latter case, the only shortcoming of Y, if it is a shortcoming at all, is not
to take over an accidentally o-correct model of X. Moreover, from the Upward
Theorem it can be concluded that if the observational truth approximation
claim is true, it implies, together with the TP-Postulate, that the corresponding
theoretical truth approximation claim is true or at least that a related true
theoretical truth approximation claim can be construed, whatever the observational truth precisely is. However, it is only possible to specify this theoretical
claim if we assume knowledge of the observational truth.
The theorems hang together so that the constructible theoretical truth
approximation claim MTL(X*, Y*, Tt) not only follows from the original
observational truth approximation MTL(nX, nY, To), but also entails the
adapted observational truth approximation claim MTL(nX*, nY*, To) and
hence entails that n y* is more successful than nX* (MSF(nX*, n Y*, R jS). To
obtain the adapted observational truth approximation claim, it is not even
necessary to impose RC(X*), for the definition of X* happens to be such that
it is automatically relatively correct.

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

215

9.1 .2. Substantial truth approximation and the innocent role of fictions

Now we will introduce the intermediate level of terms that really refer to
something real concerning the domain, called the substantial or referential
(r-) level. For term, in Vt, let Vt'indicate Vt - {,} and let n project Mp(Vt)
onto Mp(Vt'). Then, is said to refer according to Tt, or simply really refers or
Tt-refers, if and only if Tt, being on formal grounds a subset of n- 1 nTt, is a
proper subset of n- 1 nTt. That is, Tt is a proper subset of the reproduction of
the projection of Tt. This defining clause, the Tt-reference clause, is not yet to
be interpreted as a reference criterion. It amounts to the claim that, makes a
difference or plays a substantial role in the theoretical truth Tt in the sense
that Tt is constrained by , in the sense defined. With this definition, we formally
operationalize the idea that there is in reality some (type of) item which plays
the role , is ascribed to by Tt. 4
The definition still needs some qualification. It seems adequate primarily for
theoretical terms purporting to refer to attributes (properties, relations, functions); attribute terms for short. However, for theoretical terms purporting to
refer to classes of entities, used as domain-sets for attribute terms - entity terms
for short - this definition does not seem to work. However, it is plausible to
say that a (theoretical) entity term in Vt Tt-refers if and only if there is at least
one (theoretical) attribute term that Tt-refers and that uses the entity term as
(one of) its domain-set(s). The consequence is that an entity term does not Ttrefer if there are no attribute terms using it as domain-set. However, in this
case it is difficult to see how that theoretical domain could playa substantial
role in Tt, which is precisely the reason for the detour via attribute terms. If
this detour is not possible, the theoretical entities hang in the air as unconstrained entities, not distinguishable from genuine fictitious entities, at least
not with the means provided by D and Vt.
Note that the very possibility to define reference of entity terms on the basis
of the definition of reference of attribute terms is a good reason to extend the
idea of 'entity realism' to 'referential realism', as has already been suggested in
Chapter 1.
The given definition of reference is explicitly Tt-relative. We might abstract
from this by the definition that a term, refers in an absolute sense if there is
at least one domain and if , belongs to at least one vocabulary such that ,
refers relative to the resulting theoretical truth.
For the moment we define the referential vocabulary Vr(Tt) or simply Vr,
called the referential truth, giving rise to the referential (r- ) level, as the union
of Vo and those members of Vt that Tt-refer. Hence, by definition, Vo is a
subset of Vr and Vr is a subset of Vt. So we assume that it has already been
established that the observation terms refer in some other context. They may
or may not refer in the present context, i.e., they mayor may not Tt-refer.
Let Mp(Vr) indicate the set of conceptual possibilities on the r-level. By
applying the Nomic Postulate on this level, Mp(Vr) generates, in combination

216

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

with D, a unique, time-independent subset of nomic possibilities, called the


substantial truth Tr. We will first concentrate on substantial truth approximation and its relation to observational and theoretical truth approximation. In
the next section we deal with the topic of referential truth approximation.
In the present context of logical analysis we may assume to know Tt and
hence Vr and Tr. Let us first consider the TP-Postulate. Our claim now is, of
course, that Tr equals the Vr-projection of Tt, i.e., TPP(t, r), and that To equals
the Yo-projection of Tr, i.e., TPP(r, 0). Due to the transitivity of the projection
function we regain TPP(t, 0) used in the previous subsection. For this reason,
the delayed motivation of (the substantial part of) TPP(t, 0) will be given by
motivating TPP(t, r) and TPP(r, 0) separately.
Let us first consider the relation between the r- and the o-level. Hence, let n
refer, till further notice, to the projection of Mp(Vr) onto Mp(Vo). As in the
case ofTPP(t, 0), that nTr is a subset of To is a semantic fact, for if a conceptual
possibility on the r-level is nomically possible, it will remain nomically possible
when one skips some components. However, the substantial claim that To is
a subset of nTr is now also plausible, for how could To leave room for an
'observational' possibility without leaving room for at least one reproduction
of it on the r-level? In other words, how could To leave room for more
observational possibilities than only those descending from the nomic possibilities on the level of all really referring terms taken into consideration? As before,
although we do not see how the substantial part of TPP(r, 0) could be violated,
we will nevertheless mention TPP(r, 0) in proof sketches when its substantial
side is presupposed.
As earlier explained, RC(X, t, 0) is not a very strong assumption. The same
arguments can be used for the claim that the RC(X, r, 0) for subset X of Mp(Vr)
is not very strong either.
The result of these considerations is that it is possible and plausible to
reproduce the additional relativization of falsification, the Projection Theorem,
the Success Corollary and the Upward Theorem, now between the r- and the
o-level. We will not repeat the further relativization of falsification. For the rest
we obtain in this way:
r/o Projection Theorem:

if Y is at least as close to the substantial truth Tr as X , and if X is


relatively correct, then n Y is at least as close to the observational truth
To as nX,
formally : MTL(X, Y, Tr) and RC(X, r, 0) imply MTL(nX, nY, To)
and, using the Success Theorem,
r/o Success Corollary:

if Y is at least as close to Tr as X, and if RC(X, r, 0), then Y is always


at least as successful as X
formally, MTL(X, Y, Tr) and RC(X, r, 0) imply MSF(X, Y, R/S)

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

217

and, finally,
ojr Upward Theorem:
if n Y is closer to To than nX then y* is closer to Tr than X*, where
X* =defX - {x E X - Yjn(x) E nyn To}
Y*=def xn Y

On the basis of these results we can motivate the claim that RS, and hence
HD-evaluation, are functional for substantial truth approximation in formally
the same way as we motivated that they are functional for theoretical truth
approximation.
Let us now turn to the relation between the r-level and the t-level, and hence
the relation between substantial and theoretical truth approximation. Till further notice, n now refers to the projection of Mp(Vt) onto Mp(Vr). Recall that
the terms in Vt - Vr are supposed not to Tt-refer in the sense that they do not
curtail the nomic possibilities on the t-level. Formally, it is assumed that none
of these terms satisfies the Tt-reference clause. In other words, as far as Tt is
concerned, they are fictions. Hence, we may assume that Tt is the Vt-reproduction of Tr, i.e., Tt = n - I Tr, which will be called the Fiction (F -) Postulate
(FP(t, r)). This name is somewhat misleading, for the mentioned condition is
only a necessary feature of fictions. That Tt = n - 1 Tr is not a sufficient condition
for letting Vt - Vr concern fictions is due to the fact that one or more extra
terms may not narrow down the number of nomic possibilities in co-production
with Vr, but there may be other referring terms, outside Vt, that would do.
However this may be, note that Vr could have been defined as the smallest set
V' between Vo and Vt such that its truth T' satisfies the fiction postulate
relative to Tt.
Now it is important to note and easy to check that the F-Postulate implies
the TP-Postulate between Tt and Tr, i.e., n- I Tt = Tr, indicated by TPP(t, r).
Hence, in combination with the motivation of TPP(r, 0) we may conclude that
the motivation for TPP(t, 0) has been completed.
As already explained, RC(X, t, 0) for a subset X of Mp(Vt) and RC(X, r, 0)
for a subset of Mp(Vr) are not very strong. Since Tt is, due to the F-Postulate,
as large as possible, RC(X, t, r) for some subset X of Mp(Vt) becomes even
trivial. That is, every theory will satisfy it, as is not difficult to check. Moreover,
FP(t,o) guarantees that for subset X of Mp(Vt) RC(X, t, 0) iff RC(nX, r, 0),
for its Vr-projection nX.
In sum we may now derive the unconditional
tjr Projection Theorem:
if Y is at least as close to the theoretical truth Tt as X then n Y is at
least as close to the substantial truth Tr as nX,
formally: MTL(X, Y, Tt) implies MTL(nX, nY, Tr)

218

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

It is easy to check that, using RC(X, t, 0) iff RC(nX,r, 0), the t/r and the rio
Projection Theorem together imply, by a kind of transitivity, the straightforward (t/o) Projection Theorem.
Of course, we can also derive

r/t Upward Theorem:


if nY is closer to To than nX then y* is closer to Tt than X*, where
X* =defX - {x E X - Y/n(x) E nyn Tr}
y* =defX n Y
This theorem amounts to the claim that an example of theoretical truth
approximation can be constructed from an example of substantial truth approximation, assuming FP(t, r). Again, it is not difficult to check that the combination of the o/r and the r/t Upward Theorem implies, with some qualifications,
the straightforward Upward Theorem.
Moreover, concatenation of the forward and the two upward theorems
implies that RS, and hence HD-evaluation, are functional for approaching the
observational truth, which in its turn is functional for approaching the substantial truth, which in its turn is functional for theoretical truth approximation.
The forward step and the first upward step have already been elucidated.
Suppose now that we conclude in the first two steps, conditionally, that we
have a case of substantial truth approximation. For the third step, i.e., the
second upward step, it follows from the t/r Projection Theorem, assuming
FP(t, r), that the theoretical truth approximation hypothesis (TA-hypothesis
on the t-level') can 'explain' the substantial truth approximation claim, and
hence the observational truth approximation claim, and hence the success
dominance. However, the three indicated explanatory claims are not serious
cases of explanation, for they use non-referring terms. However, due to the
F-Postulate, these terms do not playa substantial role in making these claims
true, for which reason these claims are relatively innocent. From the r/ t Upward
Theorem it can be concluded that the substantial truth approximation claim,
together with F-Postulate, in principle enables making such an innocent theoretical truth approximation claim, whatever the substantial truth precisely is.
However, analogous to the first upward step, it is only possible to specify this
theoretical claim if we assume knowledge of the substantial truth.
The theorems hang together so that the constructible theoretical truth
approximation claim MTL(X* , Y*, Tt) not only follows from the original
observational truth approximation MTL(nX, nY, To), but also entails the
adapted substantial truth approximation claim, which entails the adapted
observational truth approximation claim, which entails the adapted more successfulness claim. To obtain the adapted substantial and observational truth
approximation claims, it is again not necessary to impose RC(X*), for the
definition of X* happens to be such that it is automatically relatively correct.

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

219

From this section we may conclude that there is a strong relation between
more observational success and the three thus-far considered kinds of truthlikeness: observational, substantial and theoretical. Empirical progress supports
all three kinds of corresponding truth approximation claims in a clearly defined
way. The interesting remaining question is whether empirical progress supports,
in addition, the corresponding referential truth approximation claim.

9.2. REFERENTIAL TRUTH APPROXIMATION

Introduction

In this section we will first give a precise definition of the referential claim of
a theory and the idea that one theory is closer to the referential truth than
another. It will turn out that theoretical truthlikeness does not imply referential
truthlikeness, with the consequence that empirical progress only weakly supports referential truth approximation. For this reason we will look for other
kinds of theoretical and experimental arguments in favor of referential truth
approximation claims.
9.2.1. Referential truth likeness

Let us now turn to referential truthlikeness and truth approximation. Recall


the definition based on Tt of the referential truth Vr(Tt), or simply Vr, as the
vocabulary between Vo and Vt of referring terms. In view of the F-Postulate,
it is automatically guaranteed that Vr = Vr(Tr).
Now it is crucial to note that, whereas Tt determines which terms really or
Tt-refer and which do not, every theory implies in a similar way one or more
specific Tt-reference claims. For theory X, being a subset of Mp(Vt) with the
claim "X = Tt", implicitly claims that r in Vt - Vo Tt-refers if r plays a substantial role in X, that is, if X, being on formal grounds a subset of n- l nX (where
n projects Mp(Vt) again onto Mp(Vt - {r}, is a proper subset of n- l nX.
As in the case of the definition of Tt-reference, we have again to qualify this
definition because it only seems appropriate for theoretical attribute terms. For
specific Tt-reference claims for theoretical entity terms we may again use the
detour via specific Tt-reference claims for attribute terms: if r in Vt - Vo is an
entity term it Tt-refers according to X iff there is at least one attribute term,
using that entity term as (one of) its domain-set(s), which is Tt-referring
according to X.
Let us indicate the subset of Vt consisting of Vo united with the set of
theoretical terms that Tt-refer according to X by Vr(X). Hence, Vo is a subset
of Vr(X), which is a subset of Vt. However, we do not know how Vr(X) is
situated with respect to Vr. We call the conjunctive claim of X that all r in Vo
refer, that all r in Vr(X) - Vo Tt-refer and that all r in Vt - Vr(X) do not Ttrefer the referential claim of X, and it will be indicated by RC-X5. In this way,

220

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

the referential truth becomes the referential claim of Tt, hence the strongest
true referential claim about Vt.
For subsets X and Y of Mp(Vt) we define, of course, the comparative claim
"theory Y is at least as close to the referential truth Vr as X", or "Y is
referentially at least as close to the truth as X", if RC- Y is at least as close to
the referential truth as RC-X in the sense of propositional actual truthlikeness,
defined in Chapter 7, ranging over all elementary propositions of the form: "r
refers" for all r in Vo and "r Tt-refers" for all r in Vt - Yo. In terms of sets this
amounts to:
for all r in Vr if r belongs to Vr(X) then it belongs to Vr(Y)
(Vr () Vr(X) s; Vr( Y))

for all r in Vt - Vr if r belongs to Vr(Y) then it belongs to Vr(X)


(Vr(Y) - Vr

S;

Vr(X))

The interesting question concerns what the relation is between more referential and more theoretical truthlikeness. Whereas it is not plausible to expect
that the first implies the second, the reverse may seem plausible, in particular
because the theoretical claim of a theory "X = Tt" implies the referential claim
"Vr(X) = Vr(Tt)". However, a proof for the suggested conjecture is not possible,
for interesting reasons. Suppose that MTL(X, Y, Tt). What we would like to
prove amounts to the following: all r in Vr( Y) - Vr(X) belong to Vr and all r
in Vr(X) - Vr(Y) do not belong to Yr. Let us suppose that r in Vr(Y) - Vr(X)
does not belong to Yr. Hence, Y wrongly claims that it does, whereas X rightly
claims that it does not. However, X's claim may well be based on a wrong
aspect of its theoretical claim, whereas Y may not yet be so good that its
theoretical claim implies the reference claim for the right reasons. It is important
to note that the suggested proof already fails in the case of just one theoretical
term. Moreover, a restriction to theories which are true-as-hypothesis does not
alter the situation. Similar arguments can be given for the possibility that r in
Vr(X) - Vr(Y) may belong to Yr.
It may seem interesting to enquire whether the relation holds when we only
look at the referential level. Does substantial truthlikeness imply referential
truthlikeness as far as referring terms are concerned? Of course, for subsets X
and Y of Mp(Vt) and n indicating the projection on Mp(Vr), "nY is at least as
close to the referential truth Vr as nX" amounts to (as far as Vr-terms are
concerned ):
for all r

In

(Vr () Vr(nX)

Vr if r belongs to Vr(nX) then it belongs to Vr(n Y)


Vr(n Y))

S;

As the reader may check by proof-attempts, it remains impossible to prove the


suggested adapted version of the original idea, essentially for the same reason.
Although both conjectures do not hold strictly, it is nevertheless the case

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

221

that being closer to the theoretical truth or to the substantial truth provide
good arguments for supposing that the relevant theory is closer to the referential
truth on the corresponding level. For as can be learned from both types of
proof-attempts, the suggested entailments are only violated when a theory
bases its referential claim on wrong reasons in precisely the right direction,
which would be rather exceptional because it is rather artificial. 6
In the meantime, it is a plausible suggestion that referential truthlikeness on
the theoretical level does imply referential truthlikeness on the referential level.
However, this is also not the case, for similar reasons. The suggested (t-r)
projection of referential truthlikeness would be easy to prove if Vr(nX) =
Vr n Vr(X) held in general, but this is not the case. It is only guaranteed that
Vr(nX) is a subset of Vr n Vr(X), a subset relation which is even more or less
trivial. But there may be a term in Vr n Vr(X) which does not belong to Vr(nX).
That is, there may be a Vr-term about which X, apparently rightly, claims that
it Tt-refers, whereas nX no longer claims that it Tt-refers. This occurs when
the claim of X essentially uses non- Tt-referring terms which have been dropped
in nX. In other words, a theory may have a true specific reference claim for
the wrong reasons, in the particular sense of using wrong, non-referring, other
means. Due to this possibility, referential truth approximation on the theoretical
level need not be projected on the referential level. In particular, the suggested
transition from the claim "theory Y is referentially at least as close to the truth
as X" to the claim "n Y is referentially at least as close to the truth as nX" is
invalid iff Y loses a true specific reference claim based on wrong reasons, while
X does not lose that claim, apparently, since X based it on better grounds.
Since the suggested projection claim is invalid, it is plausible to focus,
henceforth, on referential truthlikeness on the theoretical level, where, so to
speak, all specific reference claims are taken equally seriously.
9.2.2. Referential truth approximation and the assessment of reference claims

What are the prospects for referential truth approximation on the basis of
more observational success (empirical progress) or even observational truthlikeness? Given that the kind of proofs required for this purpose are formally
similar to those searched for in the previous cases, there do not seem to be
interesting theorems in this respect. At least, we did not find any.
In this subsection we will discuss in general what arguments can be given
for specific and general, separate and comparative, reference claims. First we
will deal with theoretical arguments, followed by experimental and then combined ones. Then we will deal with the consequences of the acceptance of
specific reference claims together with experimental and/or theoretical criteria
for applying them, that is, when a shift in the Observable/Unobservable
Distinction has taken place. This is called an OUD-shift by (Douven 1996).
Before we start with theoretical arguments that can be used to accept or
reject such reference claims, it is important to make clear that we presuppose
a sharp distinction between the question of whether a term refers and whether

222

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

a term, as used by different theories and theorists, refers to the same item.
There is the famous debate (Laudan 1981, Hacking 1983, Radder 1988) on
whether the notion 'electron', as used by Lorentz, Bohr, Schrodinger, Dirac
and in modern quantum electrodynamics, refers to the same item. Such 'coreference claims' are certainly interesting, but they essentially presuppose that
all these terms refer anyhow, and suggest that straightforward reference claims
can be evaluated.
As we have shown, starting from the Nomic Postulate, it is possible to define
reference and reference claims in such a way that we do not have to specify in
detail what role the hypothesized item is supposed to play. To be precise, what
we need to be able to specify is the formal character of a term, that is, whether
it is an entity term, supposed to refer to a (kind of) entity, or an attribute
term, supposed to refer to a property, relation or function . For an attribute
term, it will be possible to specify what the relevant (sub-)domains are. To
define Tt-reference, in the relevant abstract sense, this is essentially enough.
To define that a certain theory implies a certain Tt-reference claim, we need
to specify the theory itself, which implies in principle a specification of the role
the item is supposed to play. Hence, the reference claim of a theory is usually
conceived as the claim that the term refers and that the relevant item plays
such and such a role. As we have seen, our definition leaves room for the
possibility that the primary claim is true and that the additional 'role-claim' is
false, and hence that the reference claim can be taken in its pure form .
As far as co-reference claims are about theories using (part of) the same
vocabulary, it is clear that it is now possible to say that these theories refer to
the same item, without attributing to it the same role. They refer to the same
item if the term Tt-refers and each of them attributes to it a somewhat different,
and probably mistaken, role. For example, let us suppose, for a while, that
Sommerfeld's version of the 'old quantum theory' of the one-electron-atom
corresponds to the theoretical truth and hence implies the referential truth. In
that case, Rutherford's and Bohr's versions were referring to the same electrons,
but ascribing different roles to them. To be sure, to extend such claims in a
responsible way to theories using fundamentally different vocabularies is not
an easy task, and we will not go into that issue further.
Let us now start with the separate evaluation of the (general) referential
claim of a theory X, "Vr(X) = Vr" , RC-X. For accepting it, it is necessary to
argue for all component claims. For rejecting it, it is sufficient to argue against
one. Of course, it may well be possible to argue in favor of the total claim by
an argument in favor of all claims at the same time. Similarly, it may be
possible to argue against the total claim in one stroke. Such arguments will be
called holistic.
Let us first suppose that the evidence is such that X has not yet been falsified
and that it explains all established laws, i.e., R s; rrX s; S. This observational
success of X, indicated by OS-X, deductively confirms the observational claim
"rrX = To" of X (OC-X), for that entails OS-x. Since, the theoretical claim

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

223

"X = Tt" of X, TC-X, implies OC-X, TC-X is also deductively confirmed by


that success 7 Since the referential claim "Vr(X) = Vr(Tt)" of X, RC-X, is
entailed by TC-X, one might think that RC-X is now also confirmed. However,

we have seen in Chapter 2 that deductive confirmation is not transmitted to


consequences. Since neither OC-X nor (the weaker) OS-X is a deductive
consequence of RC-X, RC-X is not deductively confirmed by OS-X. But we
would say in many cases, though the conditions under which it could occur
may be difficult to specify, that RC-X and OC-X and even OS-X strengthen
each other. Hence, in such cases, we would say that OS-X confirms RC-X in
a non-deductive sense. And although we may have been inclined to reject RCX before coming to know OS-X, we may become inclined to accept it, the
more OS-X "approaches" OC-X, and hence the more we become inclined to
accept OC-x.
This qualitative analysis can be reconstructed as a version of an impressive
Bayesian argument presented by Dorling (1992) for local realist conversions
of (local) positivists, as he calls instrumentalists and empiricists alike. Given
that TC-X entails RC-X as well as OC-X, and hence OS-X, Table 9.1. specifies
the comparative relations for any probability function p (the last row will be
elucidated later):
Table 9.1. Probabilistic comparison of claims
p(TC-X)

<

<
p(RC-X)

<
p(RCr-X)

p(TC-X/OS-X)

<

<
<?

p(RC-X/OS-X)

<
< ? p(RCr -X/OS-X)

p(TC-X/OC-X)

<
< ry

p(RC-X/OC-X)

<
< ? p(RCr-X/OC-X)

Now it may well be that p(RC-X), and hence p(TC-X), is (much) smaller
than 1/2, whereas p(TC-X)/OS-X), and hence p(RC/OS-X), becomes larger
than 1/2. The latter may, in particular, happen when OS-X has become such
that p(OC-X/OS-X) approaches I, for p(TC-X/OS-X), which is equal to
p(OC-X /OS-X) p(TC-X/OC-X)

+ p(non-OC-X/OS-X) p(TC-X/non-OC-X)
then approaches p(TC-X/OC-X), which may indeed well be larger than 1/2.
In confirmation terms, note first that OC-X (deductively) confirms TC-X
and OS-X. Note further that the first joint assumption, i.e., p(RC-X)
< 1/2 < p(TC-X/OS-X), is also a sufficient, but by no means a necessary,
condition for p(RC-X) < p(RC-X/OS-X), i.e., OS-X non-deductively confirms
RC-X (Section 3.1.). Similarly, the second joint assumption, i.e., p(RC-X)
< 1/2 < p(TC-X/OC-X), is also a sufficient, but not a necessary, condition for
p(RC-X) < p(RC-X/OC-X), i.e., OC-X non-deductively confirms RC-X.

224

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

The suggested transition itself amounts, in Dorling's terms, to the claim that
a positivist attitude towards the referential claim of X, and a fortiori to its
theoretical claim, may well transform into a realist attitude towards its theoretical claim, and a fortiori towards its referential claim. It should be noted that
Dorling in fact deals with weaker claims. In our terms, they are (almost)
equivalent to: Tt is a subset of X, Vr(X) is a subset of Vr(Tt), To is a subset
of nX, and To is a subset of nX However, the formal argument is essentially
the same. Dorling illustrates the historical reality of the suggested conversion
with many examples, notably Dalton's theory of the atom. To be sure, Dorling
also notes that the conversion may also move in the reverse direction, for
various reasons.
Of course, as indicated in the last row of Table 9.1., the same argument may
be applied to each specific reference claim of X with respect to r in Vt( - Yo),
indicated by RCr-X Whether we deal with the general claim RC-X or a
specific claim RCr-X, the argument essentially uses their entailment by TC-X,
and its entailment of OC-X and hence OS-X. Since these concern whole
theories, these theoretical arguments are of a holistic nature. It may well be
that there are theoretical arguments for specific reference claims, but we do
not know them.
Let us now suppose that X has been falsified by the evidence and hence that
we have to conclude that OS-X is false and hence that OC-X and hence TCX are also false. Now the qualitative and the quantitative argument collapse.
In quantitative terms, this means that the relevant posterior probabilities
p(OC-X/not-OS-X) and p(TC-X/not-OS-X) become (instead of becoming
larger than 1/2). To be sure, it does not imply that p(RC-X/not-OS-X) or
p(RC-X/not-OC-X) also become 0, for the referential claim itself has not been
falsified. However, the evidence also does not confirm it, to say the least.
Finally, suppose nX is not a subset of S and hence that X does not entail
all established laws. In that case, we also have to conclude that OS-X has been
falsified, with the same consequences as before.
Returning to the case of favorable evidence, confirming the referential claim,
it is worthwhile discussing the well-known challenge of Laudan (1981) that
there have been in the history of science many examples of very successful
theories, that nevertheless use, according to our present lights, terms that do
not refer. From our analysis thus far it is clear that this is perfectly possible.
That is, it is perfectly possible that RC-X, and hence TC-X are false, but that
nevertheless TC-X implies OC-X and hence OS-X Hence, TC-X may be false
and nevertheless be confirmed by the evidence, without having a true referential claim.
Let us now turn to the evaluation of the comparative claim that theory Y
is closer to the referential truth Vr than theory X To accept such a comparative
claim, it is necessary to argue for all component claims of Y about which X
disagrees, and to reject it, it is sufficient to argue against one such claim (e.g.,
if Y claims that r Tt-refers then argue that it does). But again there may be

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

225

holistic arguments in favor of all deviation claims at the same time, or against
the total deviation claim in one stroke.
Recall that MSF(X, Y, R/S) indicates that Y is observationally more successful than X, and that MTL(nX, nY, To) and MTL(X, Y, Tt) indicate that Y is
observationally and theoretically closer to the truth than X, respectively. Let
MTL(Vr(X), Vr(Y), Vr(Tt, or simply MR(X, Y, Tt), indicate that Y is referentially closer to the truth than X. We further assume that X is relatively
correct. Now, due to the Success Theorem, MSF(X, Y, R/S) d-confirms
MTL(nX, nY, To) which, due to the Projection Theorem, d-confirms
MTL(X, Y, Tt) on the condition that X is relatively correct (RC(X. By
transitivity, MSF(X, Y, R/S) also d-confirms MTL(X, Y, Tt) on that condition.
Moreover, as a rule, we will also assume that MR(X, Y, Tt) and
MTL(nX, nY, To) and even MSF(X, Y, R/ S) strengthen each other, though
there is no deductive relation. In other words, MTL(nX, nY, To) and
MSF(X, Y, R/S) non-deductively confirm MR(X, Y, Tt).
Thus far the situation is formally the same as for the separate claims.
However, since MTL(X, Y, Tt) does not entail MR(X, Y, Tt), we can not
transplant Dorling's argument that, starting from an inclination to reject
MR(X, Y, Tt), and hence MTL(X, Y, Tt), we may well be inclined to accept
MTL(X, Y, Tt) and hence MR(X, Y, Tt) . Of course, the fact that the quantitative argument is no longer valid does not exclude that the crucial condition
for a 'comparative referential conversion', viz., p(MR(X, Y, Tt < 1/ 2 <
p(MR(X, Y, Tt)/MSF(X, Y, R/S, applies. In qualitative terms, the evidence
may well provide good arguments for the comparative referential conversion.
More generally, despite the fact that there is no strong supportive relation
between empirical progress and referential truth approximation, a kind of
default rule seems defensible: if Y has so far proven to be more successful than
X, then conclude, for the time being, that Y is at least as close to the referential
truth as X, except when there is evidence to the contrary. We will return to
such a rule in Subsection 9.3.1.
Suppose now that MSF(X, Y, R/S) is supposed to (non-deductively) confirm
MR(X, Y, Tt) such that we tend to accept it. What does the acceptance of
MR(X, Y, Tt) mean for the specific reference claims of X and Y? As far as
terms about which X and Y agree are concerned, nothing is accepted. As far
as their claims differ, the claim of Y is accepted, and hence the opposite one
of X is rejected. More precisely, for all , in Vr( Y) - Vr(X) the claim is that,
Tt-refers (and hence belongs to Vr), and for all , in Vr(X) - Vr( Y) the claim is
that, does not Tt-refer. Since the successes and failures of both theories as a
whole are at stake in MSF(X, Y, R/S) the theoretical arguments for these
specific claims of Y opposing those of X are again of a holistic nature.
Let us look at a similar argument to the one of Laudan in relation to the
comparative referential claim. Suppose that there are historical examples of
pairs of theories <X, Y) for which holds that at the time Y was more successful
than X, whereas according to our present standards X was referentially closer

226

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

to the truth than Y. Again, this is perfectly possible, for using referring vocabulary by no means implies formulating very successful theories with it. The real
challenge of the Laudan type is hence to find pairs of theories <X, Y ) that
satisfy the two mentioned conditions, such that X is, moreover, the best known
theory of those using the same vocabulary as X, or more precisely, having the
same referential claim as X. Though we cannot formally exclude that such
pairs exist, it seems highly unlikely as long as nobody has provided one.
In sum we may conclude that for acceptance and rejection of specific reference
claims on theoretical grounds, it is important to relate the reference claim to
the theoretical truth Tt corresponding to a certain domain D and a certain
theoretical vocabulary Vt. Then we may take as pragmatic necessary and
sufficient criterion that a term r Tt-refers if and only if it is supposed to refer
according to the best available 'theory in Vt' in the sense defined above: for
attribute terms, Tt is a proper subset of the reproduction of its projection on
Vt - {r} , and for entity terms, there are Tt-referring attribute terms using the
entity terms as domain-set. To be sure, the best available theory may be
mistaken.
Probably because there is no clear-cut deductive relation between maximal
successfulness and the referential claim, let alone between being more successful
and having more referential truth likeness, the suggested arguments may not,
however, be convincing. There is an ongoing debate about criteria of reference
in the literature. As Hacking (1983) and Cartwright (1983) have tried to argue,
it may not be important nor even desirable that theoretical claims are true; it
is certainly desirable that our theoretical terms refer. Happily enough, they are
not too pessimistic about criteria for accepting or rejecting such reference
claims. This debate concentrates on experimental criteria, to which we now
turn our attention.
Hacking's manipulation criterion is perhaps the most known and defensible.
In a somewhat broader form than he presents it himself, it amounts to the
following: an entity term and the relevant attribute terms refer as soon as we
regularly do experiments in which we manipulate these entities using the
various properties attributed to them in order to interfere in other more
hypothetical parts of nature. Hacking and Cartwright, defending 'entity realism',
focus on the reference of entity terms, or on the reality of theoretical entities,
and restrict the reference criterion for them to the causal properties attributed
to them. To be sure, the manipulation criterion is a (pragmatic, hence fallible)
sufficient condition. It can only be used for accepting reference claims, not for
rejecting them. Radder (1988,148- 153, 1996,73-92), defending a strong form
of 'referential realism', beyond theory realism, (see Chapter 1), has developed
another, but to some extent comparable, manipulation criterion which also
covers attribute terms, and which he uses for the formulation of a co-reference
criterion. Of course, when speaking of reference of attribute terms, it is always
reference via entities having the corresponding attributes.

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

227

There are other experimental arguments. For instance, there is Hacking's


argument of the grid (Hacking 1985), according to which what we seem to see
through a microscope is veridical. Hence the corresponding terms can be
accepted as observation terms, and hence as referring terms. Similarly, it seems
that Douven's criterion for the truthfulness of an instrument (Douven 1996),
implies that the corresponding terms are observational and hence do refer.
In sum, for the examination of a certain reference claim we may look at
theoretical arguments, and hence at theories according to which the term refers,
or at experimental arguments. All such arguments may lead to an application
of the famous triangulation argument. The more a term successfully functions
in theories with partly different vocabularies and the more experimental procedures suggest its non-fictional character, the more likely it is that the term
refers. That is, different 'success reasons' for reference strengthen each other.
To be sure, it is difficult to make such intuitive arguments formally convincing.
However, there is no more reason to be pessimistic in this respect than about
the presupposition of the referential skeptic, according to whom it will always
be possible to invent 'empirically equivalent theories', i.e., alternative theories
with the same observational consequences, which hence can explain the same
variety of successes and success differences. It should be conceded that inventing
such theories can not be excluded. In the same way, the skeptic can always tell
a story that explains the seemingly veridical nature of our experimental procedures, without them really being veridical. We have to admit, however, that
(natural) scientists, after a certain period of hesitation, make the inductive jump
to the conclusion that a certain term refers, that is, sooner or later, they practice
referential induction.
If we accept that a certain term refers, this may lead to a shift of the
Observable/Unobservable-Distinction, an OUD-shift (Douven 1996) in at least
two different ways. Either we accept a theory in which that term occurs in
such a way that it determines observational conditions for the (non-)application
of the term. From then on, the theory is called an observation theory, and the
term an observation term. Or we accept an experimental procedure for applying
the term, and hence accepting it as an observation term. Of course, there may
be various triangulation possibilities, in the sense that both ways may be
possible and each may even have more than one alternative. The strength of
the triangulation argument then is that all these ways of determination always
lead to essentially the same conclusion. We have already alluded to this kind
of shift in Chapter 1.
Here it should be mentioned that an interesting test can be attached to the
rationality of RS, which is a version of a test suggested by Douven (1996) and
which we would like to call the Douven-test. If most of our RS-choices on the
basis of the old OUD, remain in tact after the OUD-shift, it suggests that RS
is not only in theory, but also in practice, very fruitful for truth approximation.

228

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

9.3 . RULES OF INFERENCE, SPECULATIONS, EXTENSIONS, AND


EXPLANATORY RESEARCH PROGRAMS

Introduction

We will first extend the corrective analysis of inference to the best explanation,
introduced in Section 7.4. Moreover, we will formulate speculations about
commensurability and ideal languages, and extrapolate the analysis of
Section 9.1. and 9.2. to the comparison of vocabularies and research programs.
We will also explicate the notion of theoretical or explanatory research program.
9.3.1. Inference to the best theory

The foregoing analysis suggests also a stratification of 'inference to the best


theory', as presented in Chapter 7. Recall that RS and the Forward Theorem
suggested the rule to provisionally conclude that the best theory is the closest
to the truth:
Inference to the best theory (lBT):
If a theory has so far proven to be the best one among the available
theories, then (choose it, i.e., apply RS and) conclude, for the time being,
that it is the closest to the nomic truth T of the available theories

Now it is plausible to formulate stratified versions of IBT: Inference to the best


theory on the observational/referential/theoretical level (as the closest to the
relevant truth):
IBT -i: If a theory on level i has so far proven to be the best one among
the available theories on that level, then (choose it, i.e., apply
RS and) conclude, for the time being, that it is the closest to
the nomic truth on level i of the available theories on that level
Of course, for the referential and the theoretical level we not only need to
appeal to the Forward Theorem but also to the relevant Upward Theorems,
which leaves room for the possibility that the relevant truth approximation
claims only apply to strengthened theories. However, we will not spell this out
in detail.
Recall the severe corrections IBT contains in comparison to what is standardly called 'Inference to the best explanation (as true)' (IBE), where the best
explanation is conceived as the best not yet falsified theory. Again we can
distinguish three versions.
IBE-i:

If an explanation on level i has so far proven to be the best one


among the available theories on that level, then (choose it, i.e.,
apply rRS and) conclude, for the time being, that it is true (that
the truth on level i is a subset of that explanation)

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

229

The corrections of IBT-i compared to IBE-i are the following: unlike IBE-i,
IBT-i is not restricted to unfalsified theories, and IBT -i amounts to a comparative conclusion attached to a comparative premise, whereas IBE-i amounts to
an absolute conclusion to a comparative premise. Finally, it is directly justifiable
in terms of truth approximation, viz., by the Forward Theorem and the relevant
Upward Theorems, whereas a justification of IBE-i for any i is difficult to
imagine.
IBT-o may well be acceptable for the constructive empiricist as far as he is
willing to take the notion of the nomic truth on the observational level seriously,
which does not seem to be the case for Van Fraassen (1980). Moreover, it is
not easy to see why an instrumentalist could hesitate to subscribe to it. IBT-r
on the referential level will be acceptable for the (theory) realist. Whether IBT-t
is also acceptable simply depends on whether one is willing to speak of true
or false claims when there may be fictions involved. To be sure, the truth on
that level does not essentially use fictional terms, for it is merely a reproduction
of the referential truth.
What about the referential realist? He or she will be willing to subscribe to
IBT-0, but it is unlikely that somebody like Cartwright (1983) will subscribe
to IBT-r, let alone to IBT-t. On the basis of our analysis, she might conclude
that IBT-r is defensible and hence innocent, but she will see it nevertheless as
unimportant. For the referential realist, the crucial question concerns, of course,
whether there is something like 'inference to the most likely cause' (Cartwright
1983), or in our terminology, and more generally, 'inference to the best
(total) referential claim'. In the light of our analysis, we may formulate it
as follows:
Inference to the best referential claim as the closest to the referential truth

(IBRC-CRT)
If a theory in Vt has so far proven to be the best one among the available
theories, then (choose it, i.e., apply RS and) conclude, for the time being,
that its referential claim is the closest to the referential truth
Recall that if Y is the best theory in Vt, then IBRC-CRT suggests concluding
that Vr(Y) is closer to Vr = Vr(Tt) = Vr(Tr) than Vr(X), as defined in
Subsection 9.2.1., for any alternative theory X. Note that it does not seem to
make sense to formulate IBRC-CRT for the theories on the r-Ievel, for then
we have already assumed prior knowledge of which terms refer and which
do not.
Our previous analysis suggests that IBRC-CRT is plausible, though its
conclusion is certainly, like the IBT-i conclusions, not compelling. However, it
is without doubt more plausible than the straightforwardly generalized rule
suggested by Cartwright's 'inference to the most likely cause', where we may
or may not leave room for already falsified theories:

230

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

Inference to the best referential claim as true (IBRC-T)


If a theory/explanation in Vt has so far proven to be the best one among
the available theories, then (choose it, i.e., apply (r)RS and) conclude,
for the time being, that its referential claim is true

There is only one exception to the superiority of IBRC-CRT over IBRC-T (in
the best theory version). They coincide when there is only one theoretical term
at stake, which is Tt-referring according to the best theory.
9.3.2. Generality speculations

The TP-Postulate is the first (conceptual) constraint leading away from absolute
relativism of vocabularies. The F-Postulate is another one. In this subsection
we will formulate some other notions that make it possible to explore still
stronger non-relativistic positions. Some of the notions that will be defined
(e.g., exhaustiveness) are related to (some version of) the logical notion of
so-called Ramsey-eliminability, but we will not explore these relations here.
A vocabulary V for domain D, generating Mp(V) and a unique subset of
nomic possibilities T(V, D), guaranteed by the Nomic Postulate, is a (proper
conceptual) level for D if T'(V ', D) = T(V, D) for all vocabularies V' such that
Mp(V') includes Mp(V). In particular, V' might contain a relation, while V has
a function that happens to be sufficient to characterize T. Hence, when functions
are postulated it is always possible to enlarge the type of structure to a proper
level by replacing that function by a relation of the appropriate kind. For this
reason we may assume that Yo, Vt and Vr in the previous subsections generated
proper levels. From now on, vocabularies generate proper levels for D, fulfilling,
in line with the previous sections, minimally the following relation.
Vt is a superlevel of Vo (and Vo a sublevel of Vt) if Vo is a subset of Vt such
that nT(Vt, D) = T(Vo, D), the Truth Projection Postulate, where n is the projection function from Mp(Vt) onto Mp(Vo).
Recall that theory X (X ~ Mp( Vt reproduces theory X 0 (X 0 ~ Mp( Vo if
n- 1XO = def {x E Mp(Vt)/n{x) E Xo} = X, which implies that nX = Xo, due to
the fact that nn - 1 X 0 = X 0 holds in general.
It is now not difficult to prove the following commensurability claim: for any
two levels there is at least one common superlevel (possibly trivially defined
by 'concatenation' of tuples of components) on which all theories can be
reproduced and, hence, can be compared in principle. In Subsection 9.3.3. we
will show how this works in some more detail.
We say that Vt - V' consists of T'-fictions if Tt = T(Vt, D) is the
(Vt-)reproduction of T' = T(V', D), the Fiction Postulate. The complement of
the F-Postulate for Vt - V' is the Reference Postulate for Vt - V': Tt is a proper
subset of the (Vt-)reproduction of T(V", D) of any genuine sublevel V" of Vt
being equal to or a superlevel of V'. Under this condition, all terms in Vt - V '
Tt-refer. Vt is said to (Tt-)refer if all its (theoretical) members Tt-refer.
Vt is exhaustive for D if it has no referring superlevel, i.e., there is no super level

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

231

Vt' such that Vt' - Vt satisfies the Reference Postulate. Note that if Vt and Vt'
are both exhaustive, then one has to be a subset or a sublevel of the other.
Finally, Vt is optimal if Vt is exhaustive for D and if it Tt-refers, i.e., in suggestive
terms, if Tt = T(Vt, D) is the whole, and nothing but, the (nomic) truth about D.
In the present perspective the classical ideal language assumption (ILA) can
be formulated as follows: for every domain there is an optimal conceptual level.
This seems to be the strongest non-relativistic position. It amounts to extreme
metaphysical realism. Of course, ILA presupposes the Nomic Postulate. Hence
for the social sciences it is at least as problematic as that postulate itself is for
the social sciences. But ILA is certainly also problematic for the natural sciences.
Although ILA may regularly be fruitful as a heuristic principle, the following
conflicting principle seems to be better defensible (e.g by Popper) as a guideline:
although there may exist referring levels, there are no exhaustive levels, hence,
no optimal levels. In other words, for every level one can find a referring
superlevel. This heuristic principle may be called the refinement principle. It is
the fundamental assumption, for instance, underlying idealization and concretization. Every fortunate application of this principle not only leads to new types
of empirical success and hence empirical progress, but also, together with the
Nomic Postulate, to a deeper explanation of successes and success differences.
Since we subscribe to this principle, it is clear that we have to construct our
language. Besides this aspect, the name 'constructive realism' for our favorite
position is supposed to refer as well to (nomic) realism, that is, the intralevel
Nomic Postulate and the two interlevel postulates: the TP- and the F-Postulate.
9.3.3. Extended truthlikeness comparison

Up to now, we have restricted truthlikeness comparisons to theories using the


same vocabulary. Let us now turn our attention to truth likeness comparisons
taking different theoretical vocabularies into account. We may compare not
only theories, but also research programs and even vocabularies. Of course,
we have to ensure that the relevant domains are, or are made, the same. When
one wants to compare the theory, vocabulary or research program of motion
of Aristotle with that of Newton, the initial domains are not the same. Hence,
either one has to restrict Aristotle's domain to motion as 'change of place', or
one has to supplement Newton's theory with theories about the other kinds of
motion Aristotle had in mind with his general characterization of motion as
'transition from potential to actual'.
Hence, from now on we assume that there is a fixed domain D involved.
Characterizing the domain in an intersubjective way presupposes that there
are accepted observation terms. As far as this requires some commensurability
problems, there are two options. One is to be still more careful by restricting
the use of observation terms as much as possible. The other is to accept some
fundamental incommensurability combined with pragmatical commensurability. In this respect it is important to note that there is in the history of science
very much 'conceptual continuity', to use Radder's (Radder 1988, 1996) favorite

232

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

term for pragmatic commensurability, whereas the cases of genuine 'conceptual


discontinuity' are relatively rare. Of course, we may add in due course new,
but previously accepted, observation terms when needed for some reason or
other. Since theories can be neutrally reproduced with additional terms, we
may assume that there is not only a fixed D but also a fixed set of observation
terms Yo.
We start with the truthlikeness comparison of theoretical vocabularies, and
the improvement of vocabularies with respect to truthlikeness. An arbitrary
vocabulary Vt, being the union of Vo and the set of theoretical terms Vt - Yo,
generates, according to the Nomic Postulate in combination with D, not only
its own To and Tt, but also its own Vr and Tr, such that Vr(Tt) = Vr(rrTt) =
Vr(Tr) = Yr. For vocabulary comparison we will assume a common observational vocabulary, obtained with or without some manipulation. Let Vtl2
indicate Vtl U Vt2, i.e., the united vocabulary of Vtl and Vt2. Let Tt12, Vr12,
and Tr12 indicate the theoretical, referential, and substantial truth corresponding to (D, VtI2), respectively. In view of the nature of Tt12 and Vr12, it must
be the case that Vrl and Vr2, and hence their union, are subsets of Vrl2.
However, even their union may be a proper subset of Vr12, since the
co-production of Vtl and Vt2 may make the reference of some terms active,
whereas their reference is not active with Vtl and Vt2 alone.
In view of this situation, we propose defining the claim that vocabulary Vt2
is at least as close to the referential truth Vrl2 as Vtl as follows:
Vrl

is a subset of Vr2 (which is a subset of Vr12)

Vrl2 - Vrl U Vr2

is a subset of Vt2

Vt2 - Vrl2

is a subset of Vtl (- Vr12)

The first clause guarantees that Vt2 does not loose Ttl-referring terms. The
second clause guarantees that the union of Vtl and Vt2 creates no extra
referring terms that belong to Vt I but not to Vt2. The third clause guarantees
that the, according to Tt12, non-referring terms of Vt2 already belong to Vtl.
It is easy to check that this definition is such that Vtl2 is always at least as
close to Vr12 as Vtl. Hence, a fusion of vocabularies cannot become referentially
worse. Moreover, when no new referring terms are created, i.e., when
Vrl2 - Vrl U Vr2 is empty, the second clause vanishes, and the third clause
reduces to the claim that Vt2 - Vr2 is a subset of Vtl.
Turning to the truth likeness comparison of theories, and hence to the
improvement of theories with respect to truthlikeness, we will formally elaborate
the commensurability claim mentioned in the previous section. Note first that
a theory X can be more precisely represented by (D(X), Vo(X), Vt(X), X),
where X is a subset of Mp(Vt(X)) and where Vt(X) generates, in combination
with D(X), the theoretical truth Tt(X), the referential truth Vr(X) and the
substantial truth Tr(X). Let, in line with the commensurability claim, Vt(X, Y)
indicate the united vocabularies of Vt(X) and Vt(Y) and Tt(X, Y), Vr(X, Y),

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

233

and Tr(X, Y) the corresponding theoretical, referential and substantial truth.


Now it is plausible to introduce the 'standard interpretation' of the idea that
Y is closer to the truth than X : Vt( Y) is at least as close to Vr(X, Y) as Vt(X),
and Y" is closer to Tr(X , Y) than X" . Here e.g., X" is the Vr(X, Y)-reproduction of the Vr(X)-projection of X. Recall that the first clause implies that
Vt(X, Y) is at least as close to Vr(X, Y) as Vt(Y). This standard interpretation
can even be acceptable for instrumentalists, for it only requires speaking of
true and false on a referential level, whereas it leaves open which, if any, nonobservational terms refer. Note that, when X and Y have the same vocabulary,
the standard interpretation reduces to: rr Y is closer to Tr than rrX.
The s/o Projection Theorem guarantees the projection to all 'lower' levels,
as far as the out-distanced theory is relatively correct, and the functionality
claim with respect to the evaluation methodology can simply be summarized
as 'functional for truth approximation'. The sit Upward Theorem guarantees that y* is closer to Tt(X, Y) than X*, where X* and y* correspond to
(definable strengthenings of) X " and Y", respectively.
Let us finally turn to the truthlikeness comparison of research programs,
and hence to the improvement of research programs with respect to truthlikeness. A program P = <D, Yo, Vt, C) has an observational vocabulary Vo and
an (encompassing) theoretical one Vt, and a hard core of 'program possibilities',
a fixed subset C of Mp(Vt), also called the basic theory. All theories of a
program correspond to subsets of C. A program generates, like a vocabulary
in combination with D, its own To, Tt, Vr, Tr. Note that the theoretical and
substantial truth of a program merely remain subsets of Mp(Vt) and Mp(Vr),
respectively. Hence, it may well be that these truths do not need to overlap
with, let alone to be a subset of, C or its projection on Yr. The 'standard
interpretation' of program P2 is closer to the truth than PI becomes, of course,
the claim that Vt2 is closer to the referential truth Vr12 than Vtl and that C2
is closer to the truth Ttl2 than Cl.
9.3.4. Extended separate and comparative success evaluation

Besides the success evaluation of theories, we will discuss the evaluation of


vocabularies and research programs. As in the case of truthlikeness comparison,
the success evaluation will concern the merits of the epistemological categories
with respect to a certain domain D, assuming a fixed observational vocabulary yo.
Let us first consider the separate evaluation of theories, vocabularies and
research programs. Recall the evaluation report of a theory in terms of (general)
successes and counter-examples, specifying its observational success. In view
of the fact that a theory implied reference claims, we now add the list of
accepted and rejected reference claims, specifying its referential success. Recall
that such acceptance or rejection can be based on (a combination of) experimental and theoretical considerations. A diehard instrumentalist or empiricist

234

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

may refuse to accept such claims. However this may be, the full evaluation
report of a theory accounts for observational and referential successes.
More or less analogously, we can introduce the notion of an evaluation
report of a vocabulary. We first deal with the evaluation report of Va. It
mentions the accepted observational generalizations and theories and the
accepted sequences of observational theories (# observation theories) of
increasing (observational) success. Note that such a report is holistic in the
sense that it judges Va as a whole. Similarly, assuming that there is associated
with each theoretical term of a vocabulary Vt the reference claim that the term
Tt-refers, the evaluation report of Vt mentions (in addition) the accepted and
rejected reference claims and, on the level of accepted reference claims, the
accepted hypotheses and theories, and the accepted sequences of theories of
increasing (observational) success. Again, all this is a holistic evaluation of Vt.
In sum, the evaluation report of a vocabulary accounts for observational,
referential and theoretical successes.
Finally, the evaluation report of a research program, refines the evaluation
report of the corresponding vocabulary by indicating whether or not an
accepted hypothesis or sequence of theories respects the core theory of the
program, all on the accepted referential level.
Let us now turn to the success comparison of theories, vocabularies and programs. For success comparison of theories it is, of course, not necessary that
they share the same theoretical vocabulary, let alone that they belong to the
same research program. We only have to assume that they share the domain
and the observational vocabulary, which mayor may not require some manipulation. This assumption is enough for the possibility that one theory is not
only observationally more successful than another (as defined before) but also
referentially more successful, in the sense that all accepted reference claims of
the second theory are (accepted) reference claims of the first, whereas all rejected
reference claims of the first are (rejected) reference claims of the second.
A vocabulary Vt2 is observationally at least as successful as Vt I if all accepted
observational generalizations and theories and all accepted sequences of observational theories of increasing (observational) success of Vtl also belong to
Vt2. Similarly, Vt2 is, in addition, referentially at least as successful as Vtl if
(in addition) all accepted reference claims of VtI2 (hence including possible
extra claims that were only accepted due to the fusion of vocabularies) are
reference claims of Vt2 and all rejected reference claims of Vt2 are reference
claims of Vtl. Finally, Vt2 is theoretically at least as successful as Vtl if, on
the level of accepted reference claims, the accepted hypotheses and theories,
and the accepted sequences of theories of increasing (observational) success of
Vt I also belong to Vt2.
Finally, program P2 is observationally/referentially/theoretically at least as
successful as PI if the corresponding vocabulary Vt(P2) is at least as successful
as Vt(Pl) in the relevant sense and if C(P2) is at least as successful as C(Pl)

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

235

in the relevant sense and if all violations of C(P2) of the successes of Vt(P2)
and Vt(Pl) are also violations of C(Pl).
The extended comparative success evaluation of theories, programs and vocabularies can be used to refine the general principle of improvement of theories
(PI), and the special principles of improvement, viz., improvement guided by
a research program (PIRP), or just by a (core) vocabulary (PICV). Recall that
PIRP and PIC V include something like a principle of improvement of programs
and a principle of improvement of vocabularies, respectively (where the
improvement of programs is, in the first instance, guided by a vocabulary).
Hence, these principles will then also be refined.
In Subsection 9.3.1. we studied rules of inference restricted not only to a fixed
domain, but also to a fixed vocabulary. It is tempting to try to extend IBRCCRT to different vocabularies and research programs for the same domain.
Above we have suggested definitions for the claim that one vocabulary is more
successful than another. Hence, it is clear that something like 'inference to the
best vocabulary as the closest to the (united) referential truth' can be formulated,
but we will not elaborate such a definition. Similarly, since we have suggested
definitions for the claim that one research program is more successful than
another, something like 'inference to the best program as the closest to the
truth' seems possible.
9.3.5. Explanatory research programs

One other question can now be answered. Recall that we explicated at the end
of Chapter 7 the nature of descriptive and, more specifically, inductive research
programs, and that we had to postpone the explication of the nature of
explanatory or theoretical programs to the present chapter dealing with stratified theories. An explanatory program mayor may not use a theoretical
vocabulary. Even (nomic) empiricists can agree that it is directed at establishing
the true observational theory. If there are theoretical terms involved, the
referential realists will add that it is also directed at establishing the referential
truth. The theory realist will add to this that it is even directed at establishing
the theoretical truth. To be sure, aiming at the referential truth and the theoretical truth implies aiming at the substantial truth. Scientists working within such
a program will do so by proposing theories respecting the hard core as long
as possible, but hopefully not at any price. They will HD-evaluate these theories
separately and comparatively. RS directs theory choice. Although that rule is
demonstrably functional for all four distinguished kinds of truth approximation,
it cannot guarantee, even assuming correct data, a step in the direction of the
relevant truth. Though the basic notions of successfulness and truthlikeness
are sufficient to give the above characterization of the typical features of
explanatory research programs, they usually presuppose refined means of comparison, which will be motivated soon and elaborated in the next chapter.

236

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

9.4 . EPISTEMOLOGICAL POSITIONS RECONSIDERED

Now we are in a position to review the main epistemological positions. Our


free interpretations of the four main epistemological positions can be better
characterized and compared than before. In Section 7.3. we could already
conclude, from the Forward Theorem, that the instrumentalist is better off
replacing his aim, viz., the best derivation instrument, by the (constructive)
empiricist's aim, in our nomic interpretation, the observational truth, that is,
the strongest true observational hypothesis. From the foregoing, it not only
follows that this goal also serves unwillingly the goal of approaching the
substantial, the theoretical, and the referential truth, but also that, like (referential and theory-) realists, empiricists are better off exchanging their essentially
falsificationist methodology for the evaluation methodology, in order to avoid
irrelevant retardations. Here, the evaluation methodology is our explication of
the instrumentalist methodology, which seems to be the dominant methodology
in scientific practice.
We saw in Chapter 6 that the empiricist version of our methodological
portrait suggests the transition of one research program (with a hard or semihard core) to another when the new research program turns out to 'have a
theory' of which the corresponding observational theory (on the basis of past
tests) more successful than the projection of the best known theory of an old
program. Whatever future test results, assuming that the old test results are
not disputed, we may now conclude that the program transition is functional
for approaching the observational truth, for the theory of the new program
can still be closer to the observational truth than the original best theory of
the old program, and the reverse is impossible.
Empiricists and instrumentalists may learn from the Projection Theorems
that the realist heuristic, suggested by the ideas of reference and theoretical
truth, may also be useful for their respective purposes: the observational truth
and the best derivation instrument. Put still stronger, in the same way as the
instrumentalist has sooner or later to raise the empiricist question of how to
explain the surprising differences in success, it subsequently becomes unavoidable for the empiricist to take a theoretical level into consideration, because
observational truth approximation itself is, in general, surprising and requires
an explanation. The truth approximation hypothesis on the referential or
theoretical level can function as an explanatory hypothesis for this purpose.
Depending on its explanatory success with respect to success differences, the
theoretical or substantial truth approximation hypothesis can be made plausible
by that success.
Referential realism is situated at an intermediate position between constructive empiricism and theory realism. The referential realist believes that
there can be good reasons to assume that particular theoretical terms refer,
but none to aim at true theories, let alone to conclude that one has presumably
succeeded in this. A good reason for that position corresponds with the earlier

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

237

conclusion that being true of a theory is not very important in the context of
designing better and better theories. A bad reason, however, is rejecting the
idea to try to approach the substantial or theoretical truth. For, whether the
referential realist likes it or not, in view of the Forward and the Upward
Theorems, the application of RS is functional for truth approximation on the
corresponding level. In this light, it seems wiser to use, without reservation,
the heuristic of the theory realist: the aim is to approach the strongest true
theory. This not only enables an explanation of observational truth approximation, but also gives rise to deeper explanations of success differences.
Empiricists and referential realists who are not convinced by these arguments
in favor of the theory realist heuristic, but subscribe to the view, as most of
them seem to do, that (terms do or do not refer and) that hypotheses are true
or false, whether we can test such claims in a straightforward sense or not,
should realize the following. If one subscribes, in addition, to the view that
HD-testing of a hypothesis pertains to the truth question of that hypothesis in
one way or another, and hence that it is functional for answering this question,
it immediately follows that comparative HD-evaluation is functional, not only
for observational, but also for theoretical and substantial truth approximation.
The reason is that comparative HD-evaluation is, or at least can be reconstructed as, testing the relevant comparative truth approximation hypothesis.
To deny this, one has to claim that HD-testing is not at all relevant, let alone
informative, about the truth question of the theoretical surplus claim of the
hypothesis under consideration. This is really far from scientific practice, in
particular, when one realizes that time and again theories have been accepted
as true on the basis of the available evidence. Amongst other things, scientists
extend in this way their observational power. Of course, all this does not
exclude that later research forced the conclusion that such theories were, after
all, false.
The foregoing considerations essentially presuppose occasions where RS can
be applied. However, the applicability of RS is, in general, exceptional, for
success is usually divided. For this reason, it is important to note that the
realist heuristic has other significant advantages over the empiricist and referential realist position. This is particularly the case when theory Y is explanatorily
better than X, in the sense that the general successes of Y exceed those of X,
whereas Y has some extra counter-examples. As has been explained, such extra
counter-examples of Y may always be relativized. That they are no counterexample of X may only be accidental, i.e., it may be due to accidentally
observationally correct models of X. In that case, and only in that case, Y may
still be closer to the theoretical truth than X, even though n Y is only externally,
but not internally, closer to the observational truth than nX. This further
relativization of the role of falsification within the evaluation methodology is
not only unavailable for the empiricist; it is also unavailable for the referential realist.
In sum, there are good reasons for the instrumentalist to become constructive

238

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

empiricist; in his turn, in order to give deeper explanations of success differences,


the constructive empiricist is forced to become referential realist; in his turn,
there are good reasons for the referential realist to become a theory realist.
Besides this, there are good reasons for all of them not to be falsificationist
but 'evaluationist'. As we have seen in Chapter 6, the latter does not at all
imply that we dispute Popper's claim that aiming at falsifiable theories is
characteristic for empirical science, on the contrary, we support it, for only
falsifiable theories can obtain empirical success. But the methodological role
of falsifications is relativized in important respects. The suggested hierarchy of
heuristics is, of course, not to be taken in a dogmatic sense, that is, when one
is unable to successfully use the theory realist heuristic, one should not stick
to it, but try weaker heuristics, hence first the referential realist, then the
empiricist, and finally the instrumentalist heuristics. For, like in other kinds of
heuristics, although not everything goes, in the sense that not everything always
goes, pace (the suggestion of) Feyerabend, everything goes sometimes.
Moreover, after using a weaker heuristic, a stronger heuristic may become
applicable in a later stage: "reculer pour mieux sauter".

Concluding remarks
We conclude by briefly extending the analysis of the foundational, correspondence, and dialectical intuitions scientists and philosophers of Chapter 8 to
stratified truthlikeness and truth approximation. Moreover, we will recapitulate
the main reason for a further refinement of the analysis, in order to cover reallife scientific examples, in particular examples based on idealization and
concretization.
In Section 8.1. concerning foundations, there remained the question of
whether a dual foundation can be given for 'closer to the observational truth'.
Given that the basic definitions of observational, substantial and theoretical
truthlikeness all have a formally similar internal and an external clause, they
can all obtain a dual foundation, for the external clause can always be transformed into a plausible clause in terms of sets of sets of conceptual possibilities,
representing laws and general hypotheses on the relevant level. For the notion
of referential truthlikeness, the notion of a dual or a uniform (model or consequence) foundation does not make sense. It is essentially an application of the
definition of actual truthlikeness for propositions. To be sure, this definition
uses the same underlying intuition in terms of good and bad parts as that from
which Popper apparently started the study of truthlikeness, but then with an
unfortunate specification.
In Section 8.2. we gave intralevel explications of correspondence intuitions
concerning actual and nomic truth and truthlikeness. It is not difficult to see
that the actualist explications can only be applied to referential truth and
referential truthlikeness, whereas the nomic explications can be applied to

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

239

observational, theoretical and substantial truth and truthlikeness. However, it


would need further investigation to find out whether the interlevel coordination
condition meets serious problems when theoretical terms are involved, but we
do not yet see them.
Similarly, the epistemological explications of dialectical concepts, given in
Section 8.3., can be applied to the relevant kinds of truth and truthlikeness
distinguished in this chapter. Moreover, the methodological explications can
be applied to the relevant kinds of truth approximation.
In Chapter 6 we argued that the evaluation methodology is a refined and
corrected version of Lakatos' methodology. Theories within the vocabulary of
a research program are evaluated by separate HD-evaluation and theory
transitions can be made on the basis of comparative HD-evaluation. However,
we also saw that the dogmatic strategy of a research program was not the only
good reason to leave falsified theories in the game. From the truth approximation perspective, we can conclude that all this is functional for it, though no
guarantee. Moreover, we saw that Lakatos' view on crucial experiments basically follows from this point of view, but that his ban on ad hoc repairs is
superfluous.
A good reason to suppose that Lakatos' theory development strategy is not
the only way of justifying non-falsificationist practice is provided by what is
perhaps the most striking case of this practice: the theory development strategy
elaborated by Krajewski (1977) and Nowak (1980), called 'idealization and
concretization'. According to this strategy, one starts with a theory neglecting
some factors of which one knows that they are relevant. Hence, one knows
beforehand that the theory, suggested by strong idealization, is false. To note
that, one does not need experimental testing. The same ('to be born refuted')
holds for all concretizations of the theory that one obtains from successive
accounting for, up to that point, neglected factors. Only when one assumes
that all relevant factors of the vocabulary have been taken into account does
it make sense not to exclude that the theory is true. Following Lakatos, one
may say, of course, that the hypothesis that a certain factor is negligible is an
auxiliary hypothesis that can be blamed for the falsification. But the problem
is that the falsification is not at all a surprising result for which one has to find
a cause. In other words, there need not be an inclination to dogmatism with
respect to a certain theory one wants to save.
Consequently, the non-falsificationist strategy of idealization and concretization does not at all presuppose a division between a hard core and auxiliary
hypotheses, but neither does it exclude it. In other words, the strategies of
theory development of Lakatos and Nowak can, but need not, go together.
Both are specific strategies fitting very well into the evaluation methodology.
In both cases, it follows automatically that they are functional for truth approximation as far as they succeed in generating more and more successful theories.
The strategies of Lakatos and Nowak illustrate the general message that the
evaluation methodology and the realist epistemology can go nicely together.

240

STRATIFICATION OF NOMIC TRUTH APPROXIMATION

In Chapter 7 we saw that for an idealized version of Lakatos' dogmatic


strategy, it is possible to show its functionality for truth approximation in basic
terms. The argument given there can essentially be reproduced within the
stratified perspective. To be sure, in many examples of Lakatos, the change of
auxiliary hypotheses has aspects of concretization. However, the basic definition
of truthlikeness does not yet allow an account of truth approximation by
concretization. Such an account requires a refined notion of truthlikeness. The
refined notion and the way in which idealization and concretization become
functional for truth approximation will be spelled out in detail in the first
chapter of Part IV, Chapter 10. Among other things, we will see there that
analogous results for stratified refined truth approximation can be obtained to
the ones presented in this section about stratified basic truth approximation.
It is still to be investigated whether the stratified analysis can also be extended
to quantitative forms of basic and refined truthlikeness.
In Part IV we will also present, in Chapter 11, some real life examples of
theories illustrating the refined approach. Moreover, in Chapter 12, we will
investigate the possibilities for a quantitative approach to (unstratified) truth
approximation.
Finally, in the last chapter of the book (13), after a survey of the main
conclusions of the four parts, we will further articulate our favorite epistemological position, called constructive realism. However, detailed acquaintance with
Part IV is not strictly necessary for reading this last chapter. For that purpose,
Part II and Chapter 7 and 9 of Part Ill, and hence their main conclusions
summarized in the first section of Chapter 13, are the most important ones to
appreciate that last chapter.

PART IV

REFINED TRUTH APPROXIMATION

INTRODUCTION TO PART IV

In this last part we will complete the study of truthlikeness and truth approximation by introducing a second major sophistication accounting for a fundamental feature of most theory improvement, viz., new theories introduce new
mistakes, but mistakes that are in some way less problematic than the mistakes
they replace.
Chapter 10 introduces this refinement in a qualitative way by taking into
account that one incorrect model may be more similar, or 'more structurelike'
to a target model than another. This leads to refined versions of nomic truth likeness and truth approximation, with adapted conceptual foundations. It is
argued, and illustrated by the Law of Van der Waals, that the frequently and
variously applied method of 'idealization and successive, in particular, double
concretization' is a special kind of (potential) refined nomic truth approximation. Combining both sophistications, one obtains explications of stratified
refined nomic truthlikeness and truth approximation.
Chapter 11 illustrates by two sophisticated examples that the final analysis
pertains to real-life, theory-oriented, empirical science. The first example shows
that the successive theories of the atom, called 'the old quantum theory', viz.,
the theories of Rutherford, Bohr, and Sommerfeld, are such that Bohr's theory
is closer to Sommerfeld's than Rutherford's. Here, Bohr's theory is a (quantum)
specialization of Rutherford's theory, whereas Sommerfeld's is a (relativistic)
concretization of Bohr's theory. This guarantees that the nomic truth, if not
caught by the theory of Sommerfeld itself, could have been a concretization of
the latter. In both cases, Sommerfeld would have come closer to the truth than
Bohr and Rutherford. The second example illustrates a non-empirical use of
the idealization and concretization methodology, viz., aiming at (approaching)
a provable interesting truth. In particular, it is shown that the theory of the
capital structure of firms of Modigliani and Miller is closer to a provable
interesting truth than the original theory of Kraus and Litzenberger, of which
the former is a double concretization.
In Chapter 12 the prospects for quantitative versions of actual and nomic
truth likeness are investigated. In the nomic refined case there are essentially
two different ways of corresponding quantitative truth approximation, a nonprobabilistic one, in the line of the qualitative evaluation methodology, and a
probabilistic one, in which the truthlikeness of theories is estimated on the
basis of a suitable probability function. As already stressed in Chapter 3,
243

244

INTRODUCTION TO PART IV

probabilistic methodological reasoning, notably about confirmation, is already


rather artificial, although it is to some extent used by scientists. However,
quantitatively measuring the distance between theories and between their successes is in most cases even more artificial and, moreover, rare. Hence, the
quantitative accounts of (non-) probabilistic truth approximation are presented
with many reservations.
As we announced already in the introduction to the previous part, in the
final Chapter 13, the main lines of the resulting epistemological position suggested by the analysis in this book are drawn, viz., constructive realism. It is a
conceptually relative, nomic version of theory realism, accounting for objective
truths. They can be approached by an intersubjective method, i.c., the evaluation methodology for theories, in which the role of (truth-)testing of hypotheses
primarily concerns testing test implications of theories. This final chapter of
the book can largely be read independently of the present part, since the refined
perspective does not lead to any relativization of our conclusions about the
epistemological positions and their relation as presented at the end of the
previous part.

10
REFINEMENT OF NOMIC TRUTH
APPROXIMATION

Introduction
In Chapter 7 we presented the basic definition of nomic truthlikeness 1 of
theories in terms of the structuralist approach to theories. It surpassed first of
all the main problem of Popper's original definition, which was inadequate,
for it did not leave room for false theories, i.e., theories having actual mistakes,
i.e., real counter-examples. The basic definition was, moreover, attractive in
other conceptual, logical and methodological respects, as we have shown in
Part III. Moreover, for true theories, the basic definition captures 'scientific
common sense' and practice. Improving a true theory, in a cautious way,
amounts to strengthening it without making it false, i.e., without loosing correct
models. Hence, in perfect agreement with the basic definition, improving a true
theory is primarily a matter of dropping mistaken models, without introducing
new ones.
However, for improving false theories, the basic definition is rather naive.
Although that definition, in contrast to Popper's, leaves room for improving a
false theory by another false theory, 'basic improvement' can only be a matter
of dropping mistaken models and adding correct ones, or, in short, a matter
of replacing mistaken models by correct ones. Hence, improvement by replacing
mistaken models by other mistaken models, but better ones, is out of the
question: all mistaken models are equally bad. This certainly is not in agreement
with scientific practice and common sense. For this reason, the basic definition
does not have many real life scientific examples. In cases of scientific progress
a false theory is usually replaced by another false theory, but a better one, in
the sense that mistaken models are replaced by models which are also mistaken,
but less so. A paradigmatic case is the concretization of an idealized theory.
The indicated problem with the basic definition becomes particularly telling
when the truth is assumed to be complete. As Oddie (1981) has aptly remarked,
it then is child's play to approach the truth when we have a false theory at our
disposal, for then (as is easy to check by consulting Lemma 2 and (ii) of
Section 8.1.) we only have to strengthen it 2 . Although this is rather plausible
from the basic perspective, strengthening may well be a matter of dropping
relatively less mistaken models, in which case the child's play should be blocked
from the refined perspective of better and worse models.

245

246

REFINEMENT OF NOMIC TRUTH APPROXIMATION

A refined definition of truthlikeness should hence account for, not only, real
counter-examples but also for the fact that one mistaken model may be more
similar to a required model than another. For then there is room for improving
a theory by introducing new, but less mistaken models. Of course, a refined
definition should reduce to the basic definition under the relevant assumptions.
Finally, it should retain the attractive logical and methodological features of
the basic definition.
The structuralist approach to nomic truth likeness is particularly useful for
the suggested refinement, for it is plausible to base it on a postulated underlying
notion of structurelikeness. We start in Section 10.1. with some general constraints for the notion of structurelikeness, and hence for the notion of truthlikeness of structure descriptions. A number of specific examples, already indicated
in Chapter 7 in the context of actual truthlikeness and truth approximation,
are now studied in some more detail. In Section 10.2. the refined definition of
truthlikeness of theories is presented, based on the notion of structurelikeness,
using a sophisticated version of the conceptual justification for the basic definition. It will be shown that refined versions of all merits of the basic definition
follow. Section 10.3 shows that a refined version of the dual foundation of
truthlikeness and truth approximation is possible. In Section 10.4 it is pointed
out that 'idealization and concretization' is a special kind of potentially refined
truth approximation. This is illustrated by Van der Waals's theory of ideal
gases. Moreover, it is indicated how idealization and concretization can function
as a strategy in validity research around 'interesting theorems'. Finally, in
Section 10.5. we show that stratification of refined truthlikeness and truth
approximation is perfectly possible, leading to refined versions of the Projection
and Upward Theorems
10. 1. STRUCTURELIKENESS

Introduction

In Chapter 7 we studied the problem of actual truthlikeness and more generally


structurelikeness, with special emphasis on propositional structures. Recall that
we assumed, given a domain (D) and a vocabulary (V), and in addition to the
Nomic Postulate claiming that there is just one true theory about that domain
in that vocabulary, that every particular situation or state of affairs (of a system)
in the domain has just one correct representation or one true description. As a
consequence, a unique, true description within the vocabulary is associated
with each experiment, i.e., with each realization of a nomic possibility. Recall
also that the conceptual possibilities, the members of Mp(V), generated by the
vocabulary are the potential descriptions, which are, from the structuralist
point of view, structure descriptions, or simply structures. In this section we
will mainly deal with the general features of (more) actual truthlikeness of
descriptions, that is (more) structurelikeness.

REFINEMENT OF NOMIC TRUTH APPROXIMATION

247

10.1.1. Definitions and properties

Let x, y, z indicate structures in Mp(V) and s(x, y, z) indicate that y is at least


as similar (close) to z as x. The true structure of the context is indicated by t.
Recall that a general definition of structurelikeness is implausible, for a precise
definition will have to depend on the specific nature of the conceptual possibilities. Hence, we will concentrate on some general notions and some minimum
and other properties, and illustrate them with examples.
When s(x, y, z) y is said to be between, or to be an intermediate of, x and z.
Structurelikeness is not generally assumed to be symmetric: s(x, y, z) does not
generally imply s(z, y, x). As a consequence, being in between or an intermediate
may be a directed notion: if y is between x and z in the sense of s(x, y, z), this
does not yet imply that y is also between z and x in that sense of structurelikeness.
Structures x and z are said to be connected or related, r(x, z), if and only if
there is a y such that s(x, y, z). It also follows that r is not by definition
symmetric, i.e., r(x, z) does not automatically imply r(z, x). But r(x, z) is already
guaranteed by s(x, x, z) or s(x, z, z). Hence, the basic idea behind r(x, z) is not
the existence of a proper intermediate, but only that x and z have at least so
much in common that it makes sense to talk about (proper and improper)
intermediates: in other words, they may also be said to be comparable. For
instance, all pairs of propositional structures, as in the case of the circuit
example, will be comparable if Mp contains only structures constituted by one
set of elementary propositions. But as soon as structures based on subsets of
this set are also taken into consideration, then not all pairs are comparable
any longer. For example, p&q and - p& - q have an intermediate, e.g., p& - q,
but p&q and p don't have. The case of concretization below (Subsection 10.4.2.)
provides another example of this situation, e.g., not every Van der Waals gas
model is a concretization of every ideal gas model (they may deal with different
sets of states and/or different numbers of moles). Hence, they don't need to
have an intermediate. In general, a minimal condition for comparability seems
to be that the two structures have the same domain-sets.
It would have been possible to introduce r(x, z) as a primitive term. But we
have not done so because it always seems possible to read r(x, z) directly off
from the relevant specific definition of s(x, y, z).
In the next section we will see that basic nomic truthlikeness as defined in
Chapter 7 is in fact based on trivial structurelikeness, indicated by t(x, y, z),
defined by x = y = z. (The symbol '( is used in this text as a time variable, and
to indicate trivial structurelikeness and 'the descriptive truth', but confusions
need not arise).
Now it immediately follows that in the case of trivial structurelikeness two
structures are only related or comparable when they are equal. Hence, basic
nomic truthlikeness is based on the idea that two different structures are never
comparable.

248

REFINEMENT OF NOMIC TRUTH APPROXIMATION

Let us assume some plausible properties of the notion of structurelikeness.


s is centered iff sex, x, x) and centering iff sex, y, x) implies x = y. s is said to be
conditionally left/right reflexive iff sex, y, z) implies all kinds of left and right
reflexivity, i.e., sex, x, y), sex, x, z), sty, y, z) and sex, y, y), sex, z, z), sty, z, z),
respectively. Together these properties are called the minimal s-conditions. 3 The
first two s-conditions will be plausible enough. Together they amount to: y is
at least as similar to x as x itself if and only if x = y. The third s-condition
may need some elucidation. At first sight, it may seem already unproblematic
to unconditionally assume sex, x, y) etc .. However, as suggested before, it is
attractive to leave room for incomparable structures and the condition sex, y, z)
merely guarantees that not all pairs become comparable.
Note that rex, z) now implies s(x, x, z) and s(x, z, z). Note also that being
centered implies that r is reflexive, but the converse does not hold.
s is called symmetric when s(x, y, z) implies s(z, y, x) and antisymmetric when
sex, y, z) and s(z, y, x) imply x = y = z, in which case centering trivially follows.
If s is symmetric then r is symmetric as well, and if s is antisymmetric then r
is antisymmetric. Note that the converse implications do not hold.
There are many ways in which s can be transitive, e.g., left transitivity:
sew, x, z) and sex, y, z) imply sew, y, z). However, none of these ways implies that
r is transitive, as a laborious survey makes clear. Nor does the transitivity of
r imply any of these ways. Moreover, and this is even more important to
note, r can be transitive without assuming that the middle term is the intermediate: if rex, y) and r(y, z) then rex, z) may generally be the case, without implying
that rex, z) is due to sex, y, z).
As far as r is concerned, it is useful to state in sum that r may well be an
equivalence relation or a partial ordering, without strong implications for s. In
the case of an equivalence relation, comparability generates equivalence classes
of comparable structures. In the case of a partial ordering, directed sequences
of comparable structures arise.
In Section IDA. we will see that concretization provides a good example of
anti symmetric structurelikeness, generating a partial ordering of comparable
structures. In the present section we will restrict attention to some symmetric
examples of structurelikeness, generating in all cases comparability as an equivalence relation, sometimes in the trivial sense that all structures are comparable.

10.1.2. Examples of symmetric structurelikeness


A typical symmetric example is the case of likeness of propositional structures,
already defined in Section 7.1., and based on a fixed set of elementary propositions. Recall that it realized the intuition that p& -, q is more similar to -, p& -, q
than p&q. It is easy to check that all propositional structures (generated by a
fixed number of elementary propositions) are comparable and that s satisfies
not only the minimal s-conditions of being centered, centering and left/right
reflexivity, but also of symmetry: if sex, y, z) then s(z, y, x).
It is just a technical exercise to generalize the qualitative definition to

REFINEMENT OF NOMIC TRUTH APPROXIMATION

249

propositional constituents based on different sets of elementary propositions.


Comparability of structures then coincides with being based on the same set
of elementary propositions.
We have also already introduced likeness of first order structures as follows.
Let Mp(V) consist of structures <Dj .. ; .. R j ) of a fixed similarity type, with D;'s
as domain-sets and the R/s as relations defined on one or more of them. First
order structurelikeness sr(x, y, z) was defined by the requirement that the corresponding domains are the same (Dj(x) = Dj(y) = Dj(z) for all i) and that the
corresponding relations are related as follows: the symmetric difference between
Rj(Y) and Rj(z) is a subset of that between Rj(x) and Rj(z) for all j. Now it is
easy to check that s satisfies the minimal s-conditions, that it is symmetric, and
that the corresponding notion of comparability between two structures amounts
to having the same domain-sets, i.e., the first requirement.
Note that three different first order structures x, y and z, related by s(x, y, z),
can only be realized at different moments, due to the fact that s(x, y, z) requires
that they have the same domain.
We conclude this section by reconsidering the examples of elementary real
number structures specifying one or more ordered real numbers. Real numbers
of 'the same dimension' are indicated by numbered x's, etc..
For one dimension, one plausible definition of structurelikeness was
"xl :s; x2 :s; x3 or x3 :s; x2 :s; x I ", indicated by SI (x I, x2, x3). For two dimensions,
one possible definition was "sl(xl, x2, x3) and sl(yl, y2, y3)", indicated by
s2xl, yl), <x2, y2), <x3, y3). It is now easy to check that both definitions
satisfy the minimal s-conditions and symmetry, and that all structures are
comparable.
The latter property, and only that one, is not shared by another possible
definition of structurelikeness for two dimensions, viz., "SI (x I, x2, x3) and y I =
y2 = y3". In that case, only 'horizontal pairs' are comparable. Restricted comparability of real number structures mayor may not make sense, depending on
the nature of the properties the real numbers are supposed to represent.
Note that we would have obtained anti symmetric, directed variants of real
number structurelikeness if we had started for one dimension not by
"xl :s; x2:s; x3 or x3:s; x2:s; xl" but simply by "xl :s; x2:s; X3".4

10.2. REFINED NOMIC TRUTH LIKENESS AND TRUTH


APPROXIMATION

Introduction
It is clear that the basic definition of nomic truthlikeness of theories does not

exploit the idea that one structure may be more similar to a second than a
third, i.e., the idea of an underlying notion of structurelikeness. Let us assume
that there is such an underlying ternary relation of structurelikeness s(x, y, z),

250

REFINEMENT OF NOMIC TRUTH APPROXIMATION

stating that y is at least as similar to z as x, and satisfying the minimal sconditions introduced in Section 10.1.: being centered, centering (together:
s(x, y, x) if and only if x = y) and conditional left and right reflexivity (s(x, y, z)
implies e.g., s(x, x, y) and s(y, z, z)). Recall finally that x and z are said to be
related or comparable, indicated by r(x, z), iff there is a y between x and z, that
is, iff there is a y such that s(x, y, z). Subset X of Mp(V) is convex when it
contains all y for which there are x and z in x such that s(x, y, z).
In this chapter 'models' and 'consequences' are always s-models and sconsequences in the sense of Section 8.1. That is, if X and Yare subsets of
Mp(V) then x in X is a model of X and Y is a consequence of X iff X is a
subset of Y.

10.2.1. Refined truth likeness and theory likeness


The first refined clause to be presented is a refinement of the internal clause (Bii):
(Rii)

for all x in X and z in T


if r(x, z) then there is y in Y such that s(x, y, z)

This clause, which may be plausible in itself, will be elucidated in the next
section. It is easy to check that (Rii) implies (Bii), due to the first minimal scondition, and that, for the same reason, it even reduces to (Bii) when X is
true, i.e., when T is a subset of X. Hence, it is a strengthening of the corresponding basic clause for false theories. Moreover, it is easy to see that it nicely
reduces to the basic clause when the underlying structurelikeness is trivial, i.e.,
when s(x, y, z) iff x = y = Z.5
The second refined clause, the refined external clause, is not a strengthening
but a weakening of the corresponding basic one, which required that
Y - (X U T) was empty. As suggested before, in the context of (non-trivial)
structurelikeness, for improving false theories, it is plausible to leave room for
members of Y - (X U T), i.e., extra mistaken models of Y.
(Ri) for all y in Y -(XU T)
there are x in X - T and z in T - X such that s(x, y, z)
(Ri) requires that the extra mistakes are between X - T and T - X, that is, all
y in Y - X U Thave to be between a member of X - T and one of T - X. This
clause guarantees, as it were, that Y is moving up from X to T, without detour.
Note that (Ri) reduces to the basic clause (Bi) when X is true, for in that case,
there cannot be members in T - X, hence there is no room for extra mistaken
models of Y. Note, moreover, that (Ri) reduces to (Bi) when the underlying
structurelikeness is trivial.
In sum, the resulting definition of refined nomic truthlikeness is the conjunction (Ri)&(Rii), indicated by MTL,(X, Y, T). The strict version
MTL, + (X, Y, T) is again defined by imposing in addition that MTL,(Y, X, T)

REFINEMENT OF NOMIC TRUTH APPROXIMATION

251

does not obtain. Moreover, both definitions reduce to the corresponding basic
truthlikeness definitions when X is true and/or when s is trivial. It is easy to
check that (Ri) and (Rii), like (Bi) and (Bii), can be read as dealing with all
non-nomic possibilities and all nomic possibilities, respectively. E.g., (Ri) may
start as follows: "for all y not in T, if y belongs to Y - X, then .... " Hence,
their names, external and internal clause, respectively, are again plausible. In
the next section we deal with the intuitive foundations of this definition.
Figure 10.1. represents the resulting situation, assuming that being on a
horizontal line corresponds to being related. Arrows indicate existential claims.
Whereas x', y' and z' are only supposed to be on a horizontal line in X, Y, and
T, respectively, x, y and z are supposed to belong more particularly to X - T,
Y - (X U T), and T - X, respectively. Note that the emptiness of X n T - Y,
corresponding to 1, see below, has already been built into the figure.

Mp(V)

(Ri)

(Rii)

Figure 10.1. Refined truthlikeness: Y is closer to the truth T in the refined sense than X

We will present some of the nice general properties of refined truthlikeness.


For this purpose we jump to the general definition of refined theorylikeness
which is obtained from the formal version of the refined definition of truthlikeness by replacing the true theory Tby the arbitrary theory Z. It will turn out
that MTL, satisfies almost all properties which have been listed as properties
of MTL, with some qualifications, in particular for symmetry. Numbered sets
refer to Figure 10.2., which gives the relevant partition of Mp(V).
The suffiCient condition property (SC-property) is illuminating: if sets
X n Z - Y = 1 and Y - X U Z = 5 and Z - X U Y = 2 and/or X - Y U Z = 3
are empty, then MTL,(X, Y, Z). From this property immediately follow the
following properties. Concentricity: if X is a subset of Y and Y of Z or if Z is

252

REFINEMENT OF NOMIC TRUTH APPROXIMATION

Mp(V)

(
z

Figure 10.2. X, Y, Z and their partition of Mp(V)

a subset of Yand Y of X, then MTL,(X, Y, Z), with the immediate consequence


that MTL, is centered, i.e., MTL,(X, X, X). Moreover, concentricity implies
(unconditional) left and right reflexivity: MTL,(X, X, Y) and MTL,(X, Y, Y),
respectively. Hence, all theories are comparable in the sense that for all X and
Z, there is Y such that MTL,(X, Y, Z). It is also easy to prove that MTL,
satisfies centering, i.e., if MTL,(X, Y, X) then X = Y As a consequence, MTL,
satisfies, like MTL, the three properties that were caIled the minimal s-conditions for the underlying notion of structurelikeness. That these likeness-notions
share these minimal formal properties is plausible and desirable: theory likeness
may weIl function as structurelikeness for truthlikeness of 'theories about
theories'. Such theories are represented by sets of sets of structures. 6
Another interesting property MTL, shares with MTL is context neutrality:
MTL, depends only on the sets X, Yand T More precisely, let X, Y and Z
be subsets of Mp. If Mp itself is a subset of an extended set of conceptual
possibilities Mp* and if s* is an extension of s (with r* and T* based on s*),
then MTL,(X, Y, Z) if and only if MTL, *(X, Y, Z). The proof uses the fact
that conditional reflexivity of s already guarantees for all x and z in Mp that
rex, z) if and only if r*(x, z). Hence, theorylikeness is not disturbed by enlargement or diminution of the set of conceptual possibilities, as long as the theories
themselves are not changed.
Now we turn our attention to (anti- ) symmetry: first the central versions.
Whereas basic theorylikeness was triviaIIy symmetric, it now depends on the
specific nature of the underlying notion of structurelikeness whether or not the
ternary relation of theorylikeness is symmetric in the sense that MTL,(X, Y, Z)
implies MTL,(Z, Y, X) . If it is not symmetric it may be antisymmetric: if
MTL,(X, Y, Z) and MTL,(Z, Y, X) then X = Y = Z, in which case centering
(MTL,(X, Y, X) implies X = Y) immediately follows.
It is easy to check that MTL,(X, Y, Z) is symmetric when it is based on
symmetric structurelikeness. In other words, the refined definition guarantees
symmetry transport from the level of structures to the level of theories. However,
antisymmetry transport from the level of structures to that of theories is not

REFINEMENT OF NOMIC TRUTH APPROXIMATION

253

guaranteed by the refined definition. In Section 10.4. we will see that theorylikeness based on (antisymmetric) concretization of structures provides an antisymmetric example.
Turning to non-central symmetry notions, left antisymmetry is, in view of
the possibility of sequences of theories straightforwardly converging to the
truth, the most interesting notion. Under certain, strong, conditions theorylikeness is left antisymmetric: viz., when structurelikeness is (de- )composable,
defined by s(x, y, z) if and only if r(x, y) and r(y, z), and when X and Yare
convex, then MTLr(X, Y, Z) and MTLr(Y, X, Z) imply X = Y. That s is decomposable in the sense that s(x, y, z) implies r(x, y) and r(y, z) follows immediately
from conditional reflexivity and the definition of r. That s is composable in the
sense that r(x, y) and r(y, z) together imply s(x, y, z) is a substantial condition.
But, as we will see in Section 10.4., it is trivially satisfied by the ternary relation
of concretization. Since the proof of the claimed left antisymmetry seems,
relatively speaking, the most difficult one in the present section, we will give it.

Proof First, from MTLr(X, Y, Z) it follows that X n Z - Y is empty


and from MTLr(Y, X, Z) that yn Z - X is empty. Suppose now that
Y - X U Z is non-empty and that y belongs to it. Then, according to
(Ri) applied to MTLr(X, Y, Z), there are x in X - Z and z in Z - X
such that s(x, y, z). Hence r(x, y) and r(y, z) now follow by decomposability. In view of the latter, (Rii) applied to MTLr(Y, X, Z), means that
there must be x' in X such that s(y, x', z). Hence we have r(y, x') by
decomposability. In combination with r(x, y) we obtain by composability
that s(x, y, x'). Since x and x' are both in X, the assumed convexity of
X prescribes that y belongs to X as well, contrary to our point of
departure. A similar argument shows that X - Y U Z is empty. In sum,
X= Y. Qed.
MTL satisfied all kinds of transitivity, of which, again in view of truth convergent sequences, left transitivity is the most important. It is not difficult to
prove that left transitivity of structurelikeness is transported to refined theorylikeness, that is, it guarantees: MTLr(W, X, Z) and MTLr(X, Y, Z) imply
MTLr(W, Y, Z). Other, but similar, transitivity results follow easily.
Combining the properties of left reflexivity, left antisymmetry and left transitivity we see that MTLr(X, Y, Z) is a partial ordering of convex theories for
fixed Z if s(x, y, z) is (de-)composable. Accordingly, under these conditions, a
sequence of convex theories converging to the truth is perfectly possible, e.g.,
by successive concretization, as we will see in Section 10.4.
Now it is also easy to check that if s (satisfies the minimal s-conditions and)
is (de-)composable then s is left reflexive, antisymmetric and transitive, and
hence forms a partial ordering for a fixed third element. Hence the result might
be summarized by saying that if s is (de-)composable, the partial ordering
generated by it is transported from the level of structures to that of theories.

254

REFINEMENT OF NOMIC TRUTH APPROXIMATION

There is one property of MTL which is not at all shared by MTL r, viz.,
specularity: MTLr(X, Y, Z) does not generally imply MTLr(X, Y, Z). This is
directly related to the fact that MTL deals with internal and external mistakes
in essentially the same way, whereas MTLr introduces a fundamental asymmetry between these kinds of mistakes.7
10.2.2. Objections and their evaluation

Let us consider a famous objection to the basic definition presented by Graham


Oddie (1981). He formulated this objection in a discussion about David Miller's
original version of the basic definition in which the true theory is identified
with the one-element set {t} containing the unique structure t representing the
actual world. Oddie noted that it would be child's play to replace a theory
which is, in our terminology, false as hypothesis, in the sense of a theory not
containing t, by one which is closer to the truth, viz., by just strengthening the
theory. In our 'nomic possibility version' of the truth, i.e., the nomic truth, and
the basic definition, this objection can still be made when a theory does not
contain any nomic possibility: if X n T = tP and Y is a strengthening, i.e., a
subset, of X, then MTL(X, Y, T) . However, as is easy to see, according to
MTLr this child's play is excluded, in particular by (Ri). By strengthening a
theory, mistaken models may be dropped which are necessary for an intermediate between a mistaken model and an actual mistake of the original theory.
This would not be a problem if such, in this respect problematic, strengthenings
could be distinguished from non-problematic strengthenings; but this is, of
course, impossible without knowing T
It should be noted that the refined definition here presented is a slightly
revised version of the one published in (Kuipers 1992a/ b). The latter still left
room for refined, but not basic, improvement of true theories that are nonconvex. By requiring just "z in T", instead of "z in T - X", the 1992-version of
(Ri) left room for such cases. For that reason, several of the properties referred
to in Subsection lO.2.1. could not, generally, be claimed to apply beyond the
domain of convex theories. For instance, the reducibility of the 1992-version
of (Ri) to (Bi) fails for true but non-convex X . However, I1kka Kieseppa
(1994/ 1996) (see also Kuipers 1997a, pp. 162- 3) presented some counter-intuitive cases, which convincingly showed, at least in our opinion, that reducing
to (Bi) for all true theories X (including non-convex ones) should be a condition
of adequacy for this refined clause, at least for its most cautious version.
According to Zwart (1998, Subsection 3.5.3.3), however, it is a serious drawback, precisely because it allows only truth approximation of a true theory by
strengthening it. E.g., according to our definition {, p&q, p&q} is not closer to
the truth {p&q} than {,p&,q, p&q} ; similarly, {2, 3} is not closer to the truth
{3} than {l, 3}. These counter-intuitive 'non-examples' show that a liberalization of the clause as far as true theories are concerned is welcome, however, if
possible, without the price of Kieseppa's counter-intuitive cases.
At first sight, the following objection to the refined definition may also seem

255

REFINEMENT OF NOMIC TRUTH APPROXIMATION

convincing. It was brought to our attention by Thomas Mormann, on the basis


of which a discussion arose (Mormann 1997; Kuipers 1997b). Denote the
standard metric of ~2 (where ~ is the set of real numbers) by d and define
Sd(X, y, z) by d(y, z) ::; d(x, z). Then Sd satisfies the minimal s-conditions. We will
define some theories consisting of (unions of) unit squares, leading to counterintuitive results. A unit square will be indicated by the vertex with the highest
co-ordinates. So the (1, 2)-square is the unit square determined by the interval
[0, I] on the X -axis and by [I, 2] on the Y-axis. Now let, case 1, T be the
<3,3)-square, X the (1, I)-square and Y the union of the (1,1)-, (5,1)-,
(1,5)- and (5,5)-squares. See the left side of Figure 10.3. (in which
co-ordinates are left out). Hence X is a proper subset of Y, and T does not
overlap with Y, but lies nevertheless in the midst of the four Y-squares, one of
which is X. Now it follows that X is closer to the truth than Y in the basic
sense, but Y is as close to the truth as X in the refined sense. The primary
counter-intuitive aspect is, of course, that Y is much weaker than X and T,
whereas the latter two are of equal strength. Hence, Y is in this respect a
fundamental drawback in approaching T, starting from X. We readily agree
that this is a somewhat counter-intuitive case.

0
~

y
T

Figure 10.3. Case 1 on the left. Case I with T blown up on the right

A first reaction might be that the type of similarity relation Sd seems, due to
its purely quantitative background, not in the spirit of our qualitative approach.
The refined claim is blocked as soon as we impose a plausible qualitative
restriction on Sd leading to the relation defined in Subsection 10.1.2., indicated
by S2. That is, let (x2, y2) be at least as similar to (x3, y3) as (x 1, y1) iff
xl::; x2::; x3 or x3 ::; x2::; xl and y1 ::; y2 ::; y3 or y3::; y2::; y1. It is easy to
check that this condition implies that sdxl, yl), (x2, y2), (x3, y3 holds, but
that the counter-intuitive case 1 is now excluded.
However, S2 has similar counter-intuitive cases. Case 2: let T be the (unit)
(1, 5)-square, and let X be the (1, 1)-square. Then consider the rectangular
Y formed by the [0, I)-interval on the X-axis and the [0, 3]-interval on the
Y-axis, hence the union of the (1,1)-, the (1,2)- and the (1, 3)-square. See
the left side of Figure 10.4. (again without co-ordinates). Again, X is closer to
T than Y in the basic sense, whereas Y is (not only as close to, but even) closer

256

REFINEMENT OF NOMIC TRUTH APPROXIMATION

Figure 10.4. Case 2 on the left. Case 2 with T blown up on the right

to T than X in the refined sense. This may seem again counter-intuitive since
Y is a strong weakenting of X.
Hence, the cause of the trouble is not Sd as such, but that 'quantitativelyinduced' similarity relations seem to permit, in general, easy ways of truth
approximation by weakening of a theory to a theory which is much weaker
than T itself. We may well exclude such counter-intuitive cases by the extra
condition that Y should be in strength between X and T: the strong boundary
condition. However, although it would be effective for excluding all convincing counter-examples to follow in this section, as already indicated in
Subsection 8.2.2., this might well exclude interesting cases of weakening followed
by strengthening or vice versa 8 .
To prevent such cases and to get a systematic connection between the basic
and refined approaches, it is more plausible to impose a somewhat weaker
condition. That is, the content of Y should be not larger than the content of
the union of X and T and not smaller than the content of the intersection of
X and T. Assuming that there is a quantitative measure m of the content of
theories, this boundary condition amounts to m(X n T).::;; m(Y)'::;; m(X U T) .
Note that this is a direct implication of the basic definition of comparative
truthlikeness: X n T <:; Y <:; X U T, since any measure function respects, by definition, set-theoretic inclusion.
In the context of the two counter-examples it is plausible to measure the
content in unit squares, which leads for case 1 to m(X) = 1, m(Y) = 4, m(T) = 1,
m(X n T) = 0 and m(X U T) = 2, and hence to violation of the condition; for
case 2 it leads to m(X) = 1, m(Y) = 3, m(T) = 1, m(X n T) = 0 and m(X U T) = 2,
hence, again violated by the condition. In sum, adding the boundary condition
to the refined definition excludes the two counter-intuitive cases. Hence, this
applies to the strong version, rejected for being too strong.
It is plausible to require that quantitative definitions of basic and refined
truthlikeness satisfy the boundary condition, and one mayor may not, depending on one's purposes, impose it for qualitative definitions. It is important to
note that the boundary condition is automatically fulfilled by a paradigmatic
case of refined truth approximation, that is, double concretization, to be dealt
with in Section 10.4. As we will see, when Y is a concretization of X, and T of

257

REFINEMENT OF NOMIC TRUTH APPROXIMATION

Y, then Y is not only closer to T than X in the refined sense, but Y also satisfies
the boundary condition, simply due to the one-many character of the concretization relation (including of a one-one relation as an extreme case). We only
have to assume that X and T do not overlap, which is the normal situation
for concretization. Similarly for the other paradigmatic pattern, specialization,
followed by (non-overlapping) concretization. However, wherever the boundary
condition makes a difference, a quantitative operationalization is required.
Mormann (1997) has objected to this extra condition that one would get
back similar counter-examples as before, by 'blowing up' T For example, if
you blow up T in case 1 to the square with sides of 3 units of length, keeping
the center the same, see the right side of Figure 10.3., you get m(X) = 1, m(Y) =
4, m(T) = 9, m(X n T) = and m(X U T) = m(X) + m(T) = 10, hence obeying
the boundary condition. However, now it is highly doubtful whether everybody
remains to see it as a counter-intuitive case of the claim that Y is closer to T
than X . Note that Y now even satisfies the strong boundary condition.
Similarly, if you blow up T in case 2 to the rectangle formed by the [0, 1]interval on the X-axis and the [4, 7]-interval on the Y-axis, see the right side
of Figure lOA., you get m(X) = 1, m(Y) = 3, m(T) = 3, m(X n T) = and
m(X U T) = m(X) + m(T) = 4, hence, also obeying the boundary condition,
again even the strong one. However, again you get a case which is not so
obviously counter-intuitive.
It remains to concede that we introduce by the boundary condition a
quantitative condition that seems to depart from the qualitative approach.
However, the fundamental difference between a refined qualitative and a refined
quantitative approach (Chapter 12) is that the latter presupposes a distance
function between structures. Below we will expose our reservations about such
distances. But first we want to elaborate the claim that the qualitative approach
should not be conceived as an approach that wants to exclude the use of
quantitative aspects of the relevant structures. Second, we like to stress that
although it will not always be easy to operationalize the required content
measure, the prospects are not so bad as Mormann seems to suggest.
That we do not want to exclude the use of quantitative features is clear from
our favorite, indeed paradigmatic type and examples of refined comparative
truthlikeness based on concretization, viz., the case where comparative structurelikeness amounts to 'double concretization', e.g., the law of Van der Waals
(Section lOA.), or to 'specialization, followed by concretization', e.g., the old
quantum theory (Section 11.1). We consider such examples as testcases for a
refined definition. Although concretization almost by definition exploits quantitative features of structures, it does not at all presuppose distances between
structures.
Regarding the operationalization of the content measure, it is clear that in
a purely set-theoretic context, with finite sets, it is plausible to take the number
of elements in the relevant sets. In the two-dimensional geometric context of
the two examples it was plausible to take the areas of the relevant sets. As

258

REFINEMENT OF NOMIC TRUTH APPROXIMATION

Mormann rightly remarks, the area suggestion for the two-dimensional case
does not work for rational numbers. However, we are not so sure as Mormann
that it is in many or most infinite cases impossible to define some reasonable
content measure of theories. On the contrary: many measure-functions, in the
technical sense of measure theory, on the set of so-called measurable subsets
of the relevant set of structures seem to do the job. Hence, in general, the
prospects for such a measure are less problematic than Mormann suggests.
Let us therefore turn to the fundamental problem of a quantitative approach,
the use of a distance function between structures. Although Niiniluoto's proposals for definitions of quantitative truthlikeness based on distances between
structures (to be presented in Chapter 12) are impressive, the problem is that
apart from some exceptional cases, see below, there usually is nothing like a
natural real-valued distance function between the structures of scientific theories, let alone something like a quantitative comparison of theories based on
such a distance-function. And even in cases where it is technically easy to define
a distance function, as in some paradigm examples of the structuralist approach,
e.g., between the potential models of classical particle mechanics (CPM), as far
as they share the same domain and time interval, such a distance function
seems never to be used by scientists. The main reason obviously is that as soon
as there are two or more real-valued functions involved indicating quite different
quantities, and hence expressed in quite different units of measurement, any
'overall' distance function has equally many fundamentally arbitrary aspects.
For we have to add in one way or another functions with values in meters (m)
for position, kilograms (kg) for mass and newtons (kg.m/sec 2 ) for force. One
should not be misled by the fact that scientists frequently compare models
quantitatively, even in terms of distances. However, what they do in such cases
is quantitatively comparing one (type of) function, hence one aspect of such
models, but that is not at stake. If we want to compare e.g., classical particle
mechanics with special or general relativity theory we have to take full potential
models into account.
As we already hinted upon, there are exceptions to the rule that distances
between structures are out of order, for instance, in the case the theories to be
compared just amount to probability vectors. Roberto Festa (1993, 1995) has
convincingly shown that in such cases truth approximation can take a quantitative form, even such that systems of inductive probability may be more or less
optimal as strategies of truth approximation. However, it will also be argued
that this primarily concerns improper cases of nomic truth approximation, that
is, it amounts to refined actual truth approximation. Be this as it may,
Niiniluoto (1998, p. 12) is certainly right when he suggests that there may be
examples of intuitive truth approximation that require a quantitative approach.
He gives two graphical examples, of which the first is more convincing than
the second. A number version of the first example is the claim that
{<O, 1), (100, I)} is closer to {(3, 1), (100,0)} than {<O, 2), (100, 2)}, which
is not captured by the refined definition, but it also does not allow the opposite

REFINEMENT OF NOMIC TRUTH APPROXIMATION

259

claim: to get it right, it may well necessary to use a quantitative truth likeness
measure based on real-valued distances between structures. A number version
of his second example is the claim that {I 0 1, 199} is closer to the truth
{l00,200} than {l01, 150, 199}, whereas the refined definition also applies in
the other direction, which is not excluded by the boundary condition (although
it is excluded by the strong version). This example is less convincing, since it
may well be excluded by some defensible strengthening of (Ri). However this
may be, occasional need of a quantitative approach is one thing, the feasibility
of a meaningful quantitative approach in all cases is another.
Accordingly, in general, we cannot count on conceptually plausible distance
functions for proper cases of nomic truth approximation. For this reason, an
incommensurability reason of sorts, we focus on qualitative or, more precisely
directly-comparative definitions of truthlikeness, which do not essentially depend
on distances between structures. 9 Such definitions may be criticized for not
rightly classifying clear positive cases, leaving room for defensible liberalizations; as long as they do not wrongly classify clear negative cases as positive
cases, they are defensible as a cautious point of departure.
In this respect, Zwart (1998, Ch. (6) has proposed an interesting way to
combine, what he calls, the content and the likeness approach, for propositional
languages. Whereas our basic and refined definition both are 'truth-value
dependent', that is, only a true (false) theory can be closer to the truth than
another true (false) theory, his definition of refined verisimilitude leaves room
for the possibility that a false theory is closer to the truth than a true one:
e.g., {p&q&-'r} (false) is closer to the truth {p&q&r} than {pvqvr} (true).
Unfortunately, his proposal is technically rather complex and it seems unlikely
that it can straightforwardly be generalized to e.g., first order languages.
However, using his main intuitions for a liberalization of our refined definition
seems feasible.
For the time being, the refined definition, together with the boundary condition, is useful as a sufficient condition. Recall that the boundary condition is
implied by the basic definition and that it excludes the foregoing counterintuitive cases, suggested by Mormann, of prima facie truth approximation by
strong weakening of the theory. After all, the decisive objection to the basic
definition was not a matter of allowing truth approximation by relatively weak
theories, but a matter of, starting from a false theory, excluding truth approximation by another false theory introducing 'less mistaken' models. By the
suggested combination of criteria, we bring two desiderata together, with the
consequence that the counter-intuitive examples are excluded, whereas the 'T
blown up cases' are not, without having to exclude structurelikeness notions
defined with the aid of distance-functions between structures, such as Sd and
S2. To be sure, the comparative approach becomes in this way far from merely
qualitative. What really matters is that the combination of criteria does still
not depend on distances between structures, for the required notion of structurelikeness may well be defined without using such (overall) distances. What surely

260

REFINEMENT OF NOMIC TRUTH APPROXIMATION

is required in this combined approach is a quantitative measure of strength or


content. But we argued already that the prospects for such a measure are less
problematic than Mormann has suggested.
Let us turn to a final, 'external' reason against depending on real-valued
distances. As we have already pointed out for the basic case (Section 7.3.) and
will point out in Subsection 10.2.3. for the refined case, there is a direct
connection between basic and refined comparative truthlikeness and the comparative use of the HD-method. In our opinion, the basic and refined versions
of the Rule of Success reflect the hard core of methodological reasoning, and
we simply do not believe that comparative methodology requires distances
between structures. This is, of course, not to say that it may not make use of
distances, probably in the line of Niiniluoto's proposals, when they are available.
However, from Niiniluoto's point of departure it seems very difficult to
reconstruct and justify comparative methodology, whereas it is given into the
bargain of our comparative approach to truth approximation.
In the rest of this chapter and the next one we will pay almost no further
attention to the boundary condition, since it is satisfied in the paradigmatic
cases. However, we will come back to it in the chapter on quantitative truthlikeness and truth approximation.
10.2.3. Refined truth approximation

An important question is whether a plausible refined definition of 'more successful' can also be given, such that the corresponding rule of success is functional
for approaching the truth in the refined sense. The answer to this question is
positive; to show this, it is crucial to prove the Refined Success Theorem which
states that the adapted TA(truth approximation)-hypothesis and the
CD (correct data)-hypothesis can explain that one theory is more successful
than another in the refined sense.
Recall that R indicates the set of accepted instances at a certain time (not
made explicit) and S the strongest accepted law and that the CD-hypothesis
guarantees that R is a subset of T and T of S. We will first give the refined
definition of "theory Y is, relative to simply R jS, at least as successful as theory
X", indicated by MSFr(X, Y, R/S), and then paraphrase the clauses:
(Ri)-S

for all y in Y - (X U S)
there are x in X - Sand z in S - X such that s(x, y, z)

(Rii)-R

for all x in X and z in R


if r(x, z) then there is y in Y such that s(x, y, z)

Note first the strong analogy between this definition and the refined definition
of truthlikeness. T has been replaced by Sand R, respectively, in a systematic
way. The second clause says that Y represents R at least as well as X. Hence,
Y may be said to represent R at least as well as X. The first clause states that

REFINEMENT OF NOMIC TRUTH APPROXIMATION

261

for every established extra external mistake of Y, there is an established external


mistake of X which is at least as serious with respect to some model of S.
It is easy to check that the second refined success clause (Rii)-R is stronger
than the basic one (Bii)-R, hence established extra internal mistakes remain
forbidden for Y, formally: X n R - Y = I/J. What (Rii)-R precisely adds to (Bii)-R
will be specified in the next section. On the other hand, (Ri)-S is substantially
weaker than the corresponding basic clause (Bi )-S. Established external mistakes of Yare now not forbidden, but they have to lie between an external
mistake of X with respect to S and a model of S. Finally, it is easy to check
that the two refined success clauses reduce to the corresponding basic ones
when s is assumed to be trivial.
Assuming correct data, i.e., R is a subset of T and T a subset of S, and
assuming that S is convex, the two refined success clauses follow trivially from
the corresponding MTLr-clauses. Hence, the Refined Success Theorem holds,
provided we assume that S is convex. 10 Accordingly, analogous to the basic
case, being more successful in the refined sense can be explained by, hence
deductively confirms, the refined truth approximation hypothesis, assuming
correct data and convex S. Note that R, fortunately, does not need to be convex
for this result.
Again it is important to realize that the refined TA-hypothesis can be further
tested by realizing new nomic possibilities and/or establishing new laws. That
is, it is a decent empirical hypothesis.
In the same way as in the basic case, the explanation of the possibility of
specific progress presupposes the Nomic Postulate that <D, Mp) indeed generates the vocabulary-relative, but otherwise unique, time-independent set T of
nomic possibilities. It is easy to reformulate the two successive generalizations
of the Nomic Postulate which were required for the basic explanation of the
success of the natural sciences in general, leading to a refined version of that
general explanation.
From the Refined Success Theorem it follows that the refined rule of success
Refined Rule of Success (RS r): when Y has so far proven to be more
successful than X, in the refined sense, eliminate X in favor of Y, for the
time being

is again functional for approaching the truth in the sense that the chosen theory
may still be closer to the truth (which would explain its being at least as
successful) and that the rejected theory cannot be closer to the truth (for
otherwise it would not be less successful). The Refined Forward Theorem now
states that, if the vocabulary is observational and if the refined version of the
Comparative Success Hypothesis (CSH r) "Y remains more successful in the
refined sense than X" is true, then "Y is closer to To in the refined sense than
X". In other words, if Y is not closer to To than X, (further) testing of CSH r
will sooner or later lead to falsification of CSH r. In Subsection 10.3.2. it will

262

REFINEMENT OF NOMIC TRUTH APPROXIMATION

become clear that for this falsification it is, in contrast to the falsification of
the basic version of CSH, now not enough that Y acquires an extra counterexample or X an extra success, but the counter-example should be 'redundant'
and the success should be 'relevant', in the sense defined in that section.
The argumentation that RS, is functional for refined truth approximation
can now be extended on the basis of the Refined Forward Theorem along
similar lines as in the basic case. It is also easy to check that the adapted
versions of the heuristic principles discussed in the basic case (Subsection 7.3.4.),
viz., the Principles of Separate and Comparative Evaluation, the Principle of
Content and the Principle of Dialectics, are in their turn functional for the
Refined Rule of Success, in the sense that they stimulate new applications of
the latter.

10.3. FOUNDATIONS OF REFINED NOMIC TRUTH


APPROXIMATION

Introduction

Recall that (Bi) could be interpreted as "all mistaken models of Yare (mistaken)
models of X" or as "all true consequences of X are (true) consequences of Y".
On the other hand, (Bii) could be interpreted as "all correct models of X are
(correct) models of Y" or as "all strongly false consequences of Y are (strongly
false) consequences of X". From these 2 x 2 alternatives it was possible to
construe two uniform foundations, a model and a consequence one, and formally two, but only one very plausible dual foundation of the basic definition.
It will turn out that in the refined case too it is possible to give a uniform
model foundation, but a consequence foundation does not seem possible.
However, a dual foundation is again possible, and it can now be extended to
the refined methodology, which makes it once more the most appealing one.
10.3.1. Intuitive foundations of refined truth likeness

It is plausible to call a model x of X a relevant model if there is a z in T with


which it can be compared, i.e., for which r(x, z). Hence, we may interpret (Rii)
as realizing the intuition:

M.PIr

all relevant models of X are improved by (relevant) models of Y

where 'improved' is to be interpreted in the weak sense that Y has a model


that is at least as similar to a target structure, i.e., a model of T, as a given
model of X.
This intuition certainly is a concretization of the original positive model
intuition, and it reduces to it when s is just trivial structurelikeness. For in that
case a relevant model needs to be a correct model, and a correct model cannot,

REFINEMENT OF NOMIC TRUTH APPROXIMATION

263

of course, be improved, it can only be retained. Hence, we then get back: all
correct models of X are (correct) models of Y.
(Ri) can, similarly, be seen as a concretization of the model-interpretation
of the negative intuition. It is plausible to call a new mistaken model of Y
redundant if it does not improve upon a mistaken model of X with respect to
some still missing target structure, i.e., a member of T - X. In this way, (Ri)
explicates the intuition:
M.NIr

new mistaken models of Yare not redundant

Hence, in sum, the refined definition has an intuitive model foundation, roughly:
improving relevant models (M.PIr) while avoiding redundant models (M.Nlr).
As mentioned above, the first clause (Bi) of the basic definition could also
be interpreted as explicating the intuition "all true consequences of X are (true)
consequences of Y". It will be shown that, on the level of sets of structures,
(Ri) has a similar, conceptually attractive, interpretation. It is plausible to look
for a transition to the claim "all relevant consequences of X are improved by
Y" in the weak sense that they are retained if not strengthened by Y. Whatever
the meaning of 'relevant consequence', this will amount to the claim that "all
relevant consequences of X are consequences of Y". Hence, the remaining
question is: what are 'relevant consequences'?
Now (the set of models of) a true consequence of X, i.e., a set of which T
as well as X are subsets, mayor may not exclude mistaken models of Y that
improve mistaken models of X. If it does, it cannot be a consequence of Y.
Hence it is plausible to take only true consequences of X into consideration
that do not exclude, hence include, such mistaken models excluded by X but
improving mistaken models of X.
To elucidate this idea, let us first define the set of all such structures, to be
called 'the bridge between X and T':
B(X, T) = {y not in XU TI
there are x in X - T and z in T - X such that s(x, y, z)}

It is easy to check that (Ri) is on the level of sets equivalent to: Y is a subset
of X U B(X, T) U T, and hence, on the level of sets of sets, to

(RiC)

Q(X U B(X, T) U T) is a subset of Q(Y)

Now it is plausible to define "Z is a relevant (true!) consequence of X" by the


requirement that XU B(X, T) U T is a subset of Z, for then it is a true consequence of X not excluding useful mistaken models (This is a technical definition
of a relevant consequence, not related to so-called 'relevance logic'!).
Note that it immediately follows that XU B(X, T) U T is the strongest relevant consequence of X . Hence, (RiC) can be interpreted as explicating not only
M.Nlr, via (Ri), but also the following refined positive intuition:

264

REFINEMENT OF NOMIC TRUTH APPROXIMATION

C.PIr

all relevant consequences of X are consequences of Y

For this is precisely what (RiC) states in a formal manner. Another way of
reading (RiC) is that Ys moving up from X to T without detour means that
it goes without losing relevant consequences of X by going over the bridge
B(X, T)Y
Hence, from the consequence-analysis of the second clause, we may conclude
that the definition of refined truthlikeness, based on (Rii) and (Ri), or, equivalently, (RiC), may also be paraphrased by "improving relevant models (M.PIr)
while saving relevant consequences (C.PIr)", that is, we have not only an
intuitively appealing refined version of the model foundation of truthlikeness,
but also of the dual foundation.
We did not find any, let alone plausible, let alone intuitively appealing,
phrasing of (Rii) in terms of strongly false consequences. Hence, a refined
version of the consequence foundation for truthlikeness does not even seem
to exist.
In sum, we dealt with the fact that the basic definition of 'more truthlike' is
naive because it only leaves room for the improvement of a false theory by
dropping mistaken models and adding correct ones, whereas in scientific practice, false theories are frequently improved by replacing mistaken models by
other mistaken ones, which are, however, less mistaken. It was shown that the
structuralist-refined definition of ' more truthlike' can be based on an, intuitively
appealing, concretization of the model foundation: improving relevant models
while avoiding redundant models, as well as on a similarly appealing concretization of the dual foundation: improving relevant models while saving relevant
consequences. However, a refined consequence foundation could not be
delivered.
10.3.2. Intuitive foundations of refined methodology

Recall that we argued in Section 8.1. that the basic methodology of truth
approximation, viz., HD-evaluation combined with the basic rule of success
for theory selection, had an intuitive dual foundation in terms of successes and
counter-examples, which is directly suggested by the HD-method. That is, the
methodologically crucial notion 'more successful' is based on the intuition:
more successes and fewer counter-examples. Though a model foundation could
also be given, it was hardly plausible. On the other hand, a plausible consequence foundation could be given for HD-results and the basic rule of success.
From the refined perspective it is first of all important to note, contrary to
what one might expect, that the results of HD-evaluation, and hence their
foundation, remain the same. That is, refinement of methodology is not a
matter of refinement of HD-results, but a matter of refinement of the comparison, and hence of the rule of success.
Hence, let us review the three possible foundations of HD-results. According

REFINEMENT OF NOMIC TRUTH APPROXIMATION

265

to the intuitive dual foundation, separate HD-evaluation leads to counterexamples and successes. According to the model foundation of HD-results,
HD-evaluation should lead to, besides counter-examples, established rightly
missing models. Though the notion of an established rightly missing model is
available, it is not plausible, let alone intuitively appealing. Finally, according
to the consequence foundation of HD-results, HD-evaluation leads to, besides
successes, established strongly false consequences. Though not intuitively
appealing, the latter notion is conceptually plausible, because to be established
as a strongly false consequence 'only' requires that its complement set is a set
of established counter-examples.
For the foundations of refined methodology, recall the definition of 'being
at least as successful' in the refined sense:
(Ri)-S

for all y in Y - (X U S)
there are x in X - Sand z in S - X such that sex, y, z)

(Rii)-R

for all x in X and z in R


if rex, z) then there is y in Y such that sex, y, z)

and the corresponding rule of success:

Refined Rule of Success (RS,): when Y has so far proven to be more


successful than X, in the refined sense, eliminate X in favor of Y, for the
time being
What about the intuitive foundation of RS, and its precise relation to the
HD-method? To begin with, it is plausible that we can formulate (Ri)-S and
(Rii)-R not only in terms of models, but in three other technical ways, viz.,
corresponding to missing models, consequences and missing consequences.
However, our analysis of basic methodology strongly suggests that our methodological intuitions are in terms of successes and counter-examples. Hence, the
main challenge is to find intuitive formulations of the consequence version of
(Ri)-S and the counter-example (missing model) version of (Rii)-R. To be
precise, in view of C.PIr, it is plausible to expect that (Ri)-S can be interpreted as:
all relevant successes of X are successes of Y
and (Rii)-R, assuming some kind of transposition of M.PIr, as:
all counter-examples of Yare counter-examples of X accommodated by Y
For the first expectation it is plausible to consider now 'the bridge between X
and S', i.e., B(X, S), of course, defined as {y not in XU SI there are x in X - S
and z in S - X such that sex, y, z)}. It is easy to check that (Ri)-S is, on the

266

REFINEMENT OF NOMIC TRUTH APPROXIMATION

level of sets, equivalent to: Y is a subset of XU B(X, S) U S. It is plausible to


define "Z is a relevant success of X" by the requirement that XU B(X, S) U S
is a subset of Z, for then, assuming T to be a subset of S, it is a true consequence
of X not excluding useful mistaken models. Note that it immediately follows
that XU B(X, S) US is the strongest relevant success of X. Hence, (Ri)-S
amounts to the claim that Y is a subset of the strongest relevant success, which
can directly be interpreted, as expected, as explicating the idea that all relevant
successes of X are successes of Y.
To study the second expectation, let us first note that (Rii)-R is, due to the
minimal s-condition that s(x, z, z) if r(x, z), trivially equivalent to
for all x in X and z in R - Y
if r(x, z) then there is y in Y such that s(x, y, z)
Assume that z is in R - Y, hence, assuming in addition that R is a subset of T,
z is a counter-example of Y. What we have to prove is two-fold: z has to be a
counter-example of X, i.e., a member of R - X, and z has to be accommodated
by Y, i.e., Y has to improve upon X with respect to z (in the assumed weak
sense of structurelikeness). The first part trivially follows from the minimal sconditions. 12 The second part directly follows from (Rii )-R in the sense that Y
has, for any model of X relevant for the counter-example z of Y, a model that
is at least as similar to z. Hence, (Rii )-R may, as expected, be interpreted as
explicating the idea that all counter-examples of Yare counter-examples of X
accommodated by Y.
Hence we may conclude that 'more successful' in the refined sense, and hence
RS" have an intuitive dual foundation in terms of successes and counterexamples: accommodating counter-examples while saving relevant successes.
Since 'more successful' in the refined sense is defined by (Ri)-S and (Rii)-R,
that is, in terms of models, there is also a uniform model foundation available,
but it is hardly possible to call it plausible. The reason is that the plausibility
of (Ri)-S, if any, completely derives from its being suggested by (Ri). For a
uniform consequence foundation we would need a version of (Rii)-R in terms
of 'strongly false consequences', but constructing such a version seems even
more difficult than one for (Rii), which we tried without success. Hence, we
have no consequence foundation available.
As expected, the refined methodological perspective comes in only when we
start comparing successes and counter-examples of two theories. The refined
perspective does not play a role in separate HD-evaluation as such. That
method provides the ingredients for basic or refined theory comparison.
In sum, refined methodology, consisting of the HD-evaluation and RS" is
not only functional for refined truth approximation, but both components have
also intuitive dual foundations: HD-evaluation in terms of consequences, viz.,
successes, and missing models, viz., counter-examples, and RS" and its crucial

REFINEMENT OF NOMIC TRUTH APPROXIMATION

267

notion 'more successful', in terms of saving relevant successes and accommodating counter-examples. Finally, according to the resulting dual foundation of
'refined truth approximation', 'more truthlike' can be taken in the refined sense
of saving relevant consequences, while improving relevant models. In this
way we obtain a dual account of qualitative truth approximation by the
HD-method, which is very close to scientific common sense.

10.3.3. Extended summary of foundations

The new results of our foundational investigations will be summarized by an


extended version of the Table 8.2. at the end of Section 8.1. Recall that if
conceptual foundations of a certain type and purpose are available at all, they
are called plausible when they can be shown to have a conceptually sound
basis. They are called intuitive when they are, moreover, intuitively appealing,
i.e., seem to express 'scientific common sense'. The survey in Table 10.1. below
concerns the whole network of related, basic and refined, qualitative, definitions
of 'more truthlike', 'more successful' and HD-results.
Table 10.1. Foundations of basic and refined truth approximation

Foundations
Model

Dual

Consequence

More truthlike

basic
refined

intuitive
intuitive

intuitive
intuitive

plausible
impossible?

More successful

basic
refined

available
available

intuitive
intuitive

plausible
impossible?

available

intuitive

plausible

HD-results

The dual foundation of qualitative basic and refined truth approximation is


evidently superior to the other two. Although the model foundation can also
be extended to the methodological notions, this is not in a very plausible way.
And although the consequence foundation of basic truth approximation is
plausible, it does not seem extendable to refined truth approximation.
It is not difficult to check that it is also possible to give a refined, again
intralevel, version of the basic explication of the intuitions underlying the
correspondence theory of truth presented in Section 8.2. Similarly, refined
versions of the basic explications of the dialectical concepts, presented in
Section 8.3., are not difficult to give. In the next section we will confine ourselves
to the explication of the dialectical notion of correspondence, for which there
does not seem to be a basic explication at all.

268

REFINEMENT OF NOMIC TRUTH APPROXIMATION

10.4 . APPLICATION : IDEALIZATION AND CONCRETIZATION

Introduction
Idealization and concretization is an important strategy for program development and there is a the related strategy of guidance by an interesting theorem
(see Kuipers, SiS). In this section we will study these strategies in some detail
from the point of view of truth approximation. From the general exposition of
refined truthlikeness it then trivially follows that concretization of theories can
be a truth approximation strategy. This will be illustrated by the transition of
the theory of ideal gases to that of Van der Waals. Then we will outline how
concretization is also an important strategy in the investigation of the domain
of validity of an interesting theorem and, in particular, whether it is true for
the actual or even the nomicallY possible worlds.

10.4.1 . Truth approximation by (idealization and) concretization


Concretization or factualization, as it has been presented by the Polish philosophers Wladisla w Krajewski (1977) and Leszek Nowak (1980), is basically a
relation between real-valued functions. Hence, let us assume that the conceptual
structures to be considered contain one or more real-valued functions, with or
without one or more real constants. For what follows, it does not directly
matter whether the structures considered belong to Mp(Vo), Mp(Vr), or Mp(Vt) .
Structure y is called a concretization of x and x an idealization of y, indicated
by con(x, y), if y transforms, directly or by a limit procedure, into x when one
or more constants occurring in y assume the value O. It is easy to see that it
is a necessary condition for con (x, y) that x and y have the same domain-sets.
Moreover, it is easy to check that con is reflexive, antisymmetric and transitive.
In a subsection to follow, the example of a Van der Waals gas model is
presented as a concretization of an ideal gas model.
Concretization is primarily a binary relation, but for our purposes, we need
the plausible ternary version leading to a concretization triple: ct(x, y, z) if and
only if con(x, y) and con(y, z). We will assume ct as the underlying notion of
structurelikeness. The relation of relatedness based on ct is easily seen to be
equivalent to con. Note that we have here a clear example in which relatedness
is not symmetric, but directed. Note also that ct is trivially decomposable.
Truth-/theory-likeness based on this ternary relation will be indicated by
MTLct(X, Y, T) and MTLc,(X, Y, Z), respectively. It is easy to check that ct
satisfies the minimal s-conditions of being centered, centering and conditional
left and right reflexivity. Moreover, it is antisymmetric (central, left and right)
and it satisfies all conceivable kinds of transitivity, e.g., left: if ct(w, x, z) and
ct(x, y, z,) then ct(w, y, z).
Our next task is to define the binary relation of concretization between
theories. Again we will do this as weakly as possible: Y is a concretization of
X and X an idealization of Y, indicated by CON (X, Y), if and only if all

REFINEMENT OF NOMIC TRUTH APPROXIMATION

269

members of X have a concretization in Y and all members of Y have an


idealization in X. At first sight, one might think that the second clause should
be strengthened to: and all members of Y have a unique idealization in X.
However, this would exclude e.g., 'inclusive' concretization triples <X, Y, Z)
with X as subset of Yand Y of Z and CON (X, Y) and CON(Y, Z).
It is trivial that CON is reflexive and transitive. However, contrary to what
one might expect, it need not be antisymmetric. But sufficient for anti symmetry
of CON (X, Y) is that X and Yare convex (i.e., closed for intermediates). Now
convexity amounts to: if con(x, y) and con(y, z), i.e., ct(x, y, z), and x and z in
X, then y in X.
The ternary relation of concretization of theories we define again as weakly
as possible: CT(X, Y, Z) if and only if CON(X, Y) and CON(Y, Z). It is easy
to check that CT has the properties of being centered, centering for convex
sets, conditional left and right reflexivity, antisymmetry (central, left, right) for
convex sets and all conceivable forms of transitivity. As a consequence,
CT(X, Y, Z) is for fixed Z a partial ordering as far as convex theories are
concerned.
The main question is whether or under what conditions CT(X, Y, Z) implies
MTLct(X, Y, Z). It turns out that some conditions have to be added to guarantee this implication, but there are some alternative possibilities. We are, of
course, primarily interested in conditions on X and/or Y or their combination,
for in the crucial case we do not dispose of Z, i.e., T One sufficient combination
of conditions is the following: Y should be convex as well as mediating, the
latter condition being defined as: if z is a concretization of x and if x has a
concretization in Y and z an idealization in Y, then Y also provides an intermediate for x and z; or, more formally, if con(x, z) and if there are y and y' in Y
such that con(x, y) and con(y', z), then there is y" in Y such that con (x, y") and
con(y", z) (i.e., ct(x, y", z.
Note that both conditions only concern Y. Although being mediating is a
more specific property than convexity, it is not a very restrictive condition in
the present context. Note also that it follows that any X can be an idealized
starting point for successive concretization. However, the starting point X will
usually even be closed for idealizations in the sense that if x in X and con(x', x),
then x' in X. It is easy to check that this trivially implies that X is convex and
mediating.
Let us formally state the main claim: it is (easily) possible to prove the
following Double Concretization (DC- ) Theorem: if CT(X, Y, Z) and if Y is
convex and mediating, then MTLct(X, Y, Z). In words: the intermediate theory
of a concretization triple is closer to the third than the first, assuming that it
is convex and mediating.
We may define stronger versions of concretization triples such as
CT*(X, Y, Z) = CT(X, Y, Z) and Y convex and mediating or even
CT**(X, Y, Z) = CT*(X, Y, Z) and X and Z also convex. According to the
DC-Theorem, both are special kinds of more theorylikeness. Moreover, it has

270

REFINEMENT OF NOMIC TRUTH APPROXIMATION

already been mentioned that CT(X, Y, Z) is antisymmetric (in the central sense)
as soon as the three sets are convex; hence CT** is an antisymmetric special
type of theorylikenes.
A direct consequence of the DC-Theorem is that, if theory Y is a concretization of theory X, if Y is convex and mediating, and if the true set of nomic
possibilities T is a concretization of Y, then Y is closer to the truth than X.
This may be called the Truth Approximation by Double Concretization
(TADC-)Corollary - a major goal of this section - viz., to show that and in
what sense concretization may be a form of truth approximation. All conditions
for truth approximation can be checked, except, of course, the crucial heuristic
hypothesis that T is a concretization of Y.
Let us first link the TADC-corollary with the explication of dialectical
concepts presented in Section 8.3. Recall that basic truthlikeness MTL(X, Y, T)
could be considered as the (nomic epistemological) explication of the idea of
dialectical negation, and that several other dialectical notions could be explicated in a similar way. Hence, it is plausible to claim that MTL,(X, Y, T) can
be seen as the refined explication of dialectical negation, and that the other
explications can be refined analogously. However, in Section 8.3. we could not
yet explicate the idea of 'dialectical correspondence' (Nowakowa 1974, 1994).
Since dialectical correspondence has certainly to do with concretization it is
now plausible to explicate the idea that Y dialectically corresponds with X by
MTLct(X, Y, T). The TADC-corollary specifies the conditions under which a
concretization Y of X dialectically corresponds with X and hence is closer to
the truth than X.
Returning to the TADC-corollary in general, to obtain 'good reasons' to
assume that the required heuristic hypothesis that T is a concretization of Y
is true, it is important that the concretization has some type of (necessarily
insufficient) justification, of a theoretical or empirical nature, suggesting that
the account of the new factor is in the proper direction. In this respect, it is
plausible to speak of theoretical and/or empirical concretization. The famous
case of Van der Waals to be presented in the next subsection evidently is a
case of theoretical concretization, followed by empirical support. The same
holds true for Sommerfeld's concretization of the 'old quantum theory' to be
presented in the next chapter.
10.4.2. Application to gas models

The transition from the theory of ideal gases to Van der Waals's theory of
gases has frequently been presented as a paradigmatic case of concretization.
The challenge of any sophisticated theory of truthlikeness hence is to show
that this transition can be a case of truth approximation.
For this purpose, we start by formulating the relevant models in elementary
structuralist terms. (S, n,P, V, T) is a potential gas model (PGM) if and only if
S represents a set of thermal states of n moles of a gas and P, V and Tare

REFINEMENT OF NOMIC TRUTH APPROXIMATION

271

real-valued functions defined on S representing pressure, volume and (empirical


absolute) temperature, respectively.
Specific gas models are PGM's satisfying an additional condition. The ideal
gas models (IGM) satisfy in addition P(s)V(s) = nRT(s) for all s in S, or simply
PV = nRT, where R is the so-called ideal gas constant. For gas models with
mutual attraction (GMa) there is a non-negative real (number) constant a,
within a certain fixed interval, such that (P + (n 2a/ V2 V = nRT For gas models
with non-zero volume of molecules (GMb) there is a non-negative real constant
b, within a certain fixed interval, such that P(V- nb) = nRT Finally, in the case
of Van der Waals gas models (WGM) there are non-negative real constants a
and b, within the previously mentioned two intervals, such that (P + (n 2a/V2
(V- nb) = nRT
Note first that it is a necessaria condition for con (x, y) (x and y in PGM)
that Sx = Sy. Note also that IGM/GMa/GMb/ WGM have been defined such
that they are all convex and mediating.
It is easy to check that IGM, GMa and WGM as well as IGM, GMa and
WGM constitute a concretization triple: an element of WGM transforms into
an element of GMa/GMb by substituting the value 0 for b and a, respectively.
The resulting elements of GMa and GMb transform into elements of IGM by
substituting 0 for a and b, respectively.
Due to the DC-Theorem, it follows that GMa and GMb are both closer to
WGM than IGM. As a consequence, if WGM represented the true set of
nomically possible gases, GMa and GMb would be closer to the truth than
IGM. Finally, and most importantly, the TADC-Corollary guarantees that
WGM is closer to the truth than IGM, assuming the heuristic hypothesis that
the true set of nomically possible gases is on its turn a concretization of WGM.

10.4.3. Validity research


Scientific research is not always directed at describing the actual world or
characterizing the set of nomically possible worlds. It may also primarily aim
at proving interesting theorems for certain conceptual possibilities. Hamminga
(1983) showed this for neo-classical economicsY
Let a certain D and V, hence Mp(V), have been chosen, let T indicate the
(unknown) subset of nomic possibilities, and let R indicate the (equally known)
subset (of T) of realized (nomic) possibilities, possibly containing just one
element, the actual possibility.
Let IT indicate an 'interesting theorem', that is some insightful claim, of
which it is interesting to know whether it is true for the nomic possibilities, or
at least the realized possibilities. Let Val(IT), or simply Val, indicate the set of
conceptual possibilities for which IT can be proved. Val is called the domain
of (provable) validity of IT, and it is assumed to be not yet explicitly
characterized.
A frequently occurring type of scientific progress is the following. Suppose
that it was earlier proved that IT holds for X, i.e., that X is a subset of Val.

272

REFINEMENT OF NOMIC TRUTH APPROXIMATION

The new result is that Y, which includes X, is also, like X, included in Val. Due
to concentricity of the basic and refined theorylikeness notions, it follows in
this case that MTL/ MTLr(X, Y, Val). The ultimate purpose of this type of
research was to find out whether T, or at least R, are subsets of Val. Of course,
the larger Val has been proven to be, as in the described case, the greater the
chance, informally speaking, that R or even T are subsets of Val. However,
simply enlarging the proven domain of validity does not necessarily go in the
direction of Rand T For this purpose, concretization is the standard strategy.
Let it first have been shown that X is a subset of Val, and later that a
concretization Y of X (CON(X, Y) , X need not be a subset of Y) is also a
subset of Val. It then trivially follows that MTLr(X, X U Y, Val). If, moreover,
Y is convex and mediating, it follows from the heuristic hypothesis that Y is a
concretization of T (CON(Y, T)), using the DC-Theorem, that MTLr(X, Y, T) .
Hence, we have proved IT for a set Y which is more similar to T than X,
which increases the chance that IT holds for T, ipso facto for R.
A complex form of validity research concerns the case that IT is not fixed,
but that realistic factors are successively accounted for. Formally, this is also
a form of concretization. IT2 is called a concretization of ITt if Val(IT2) =
Val2 is a concretization of Val (ITt) = Vall.
Now suppose that ITt is proven for X. The relevant heuristic strategy is to
look for a concretization Y of X and a concretization IT2 of ITt such that
IT2 can be proved for Y. The heuristic hypotheses are that T is a concretization
of Y and that there is a concretization IT* of IT2 such that IT* holds for T
and hence for R. This makes sense because, if Y and IT2 are convex and
mediating, it not only follows that Y is closer to T than X, but also that Val2
is closer to Val* than Vall. Hence, in this case we are not only on the way to
T but also to IT*.
The concretization of the theory and corresponding theorem of Modigliani
and Miller concerning the capital structure of firms by Kraus and Litzenberg
turns out to be a perfect example of this kind of approximation of a provable
interesting truth (Cools, Hamminga and Kuipers 1994). It will be presented in
the next chapter.

10.5. STRATIFIED REFINED NOMIC TRUTH APPROXIMATION

Introduction

We will now investigate how the refined definition works out for theories that
are stratified in terms of a distinction between theoretical and (relatively) nontheoretical or observational components. The main question is again whether
and to what extent truthlikeness on the theoretical level (including theoretical
and observational components) is preserved on the observational level.
Recall that we distinguished three vocabularies: Vo for the observation terms,

REFINEMENT OF NOMIC TRUTH APPROXIMATION

273

Vr for Vo plus the referring theoretical terms, and Vt for Vr plus all nonreferring theoretical terms. By the Nomic Postulate these vocabularies were
supposed to generate subsets To, Tr and Tt of their corresponding sets of
conceptual possibilities, representing the corresponding sets of nomic
possi bili ties.
We concentrate on the question of whether truthlikeness is projected from
the t-level onto the o-Ievel. Now s, refers to structurelikeness on the t-level and
not to trivial structurelikeness. Let re indicate the projection function from
Mp(Vt) onto Mp(Vo). Recall also that and how the Truth Projection Postulate
(TPP(t,
telling that To = reTt was made plausible. It was first of all supposed
to hold between the r-Ievel and the o-Ievel (TPP(r, t. Between the t- and
o-Ievel it was a direct consequence of the Fiction Reproduction Postulate
(FRP(t, r). The combination implies the overarching TPP(t,o). It will be
assumed from now on. Recall, finally, that it is plausible to assume that
R ; To ; S, the CD-hypothesis, which will not be further mentioned.
In the refined set-up, it is now also plausible to assume throughout that s
satisfies the s-Projection Postulate: s,(x, y, z) implies so(re(x), re(y), re(z. Note,
contrary to what one might think at first sight, that the projection reX of a
convex set X need not be convex, nor the other way around.
Let us start again with the external clause. Let X and Y be subsets of Mp(Vt)
and let them satisfy (Ri) with regard to Tt, i.e.,

(Ri),

for all y in Y - (X U Tt)


there are x in X - Tt and z in Tt - X such that s,(x, y, z)

It is easy to check that this clause is projected on the o-Ievel, assuming TPP(t, 0)
and assuming that X and To is convex.

Proof: We have to prove that if Yo is in reY - X U To then there is Xo


in reX - To and Zo in To - reX such that so(xo, Yo, zo). There must be a
yin Y - X U Tt such that re(y) = yo. (Ri), guarantees that there are x in
X - Tt and z in Tt - X such that s,(x, y, z), and hence so(re(x), re(y), re(z.
If re(x) in reX n To then, due to the convexity of To, re(y) = Yo in To
contrary to our starting assumption. Hence, re(x) belongs to reX - To.
In a similar way it can be argued that re(z) belongs to To - rex' Qed.
And again, now using the relevant side of the Refined Success Theorem, the
external or explanatory success clause also follows: re Y is at least as successful
with respect to S as reX, provided S is convex.
Let us now consider the internal clause and recall that we only obtained
projection in the basic approach by assuming (TPP(t, 0) and) the technical
condition RC(X) that X was 'relatively correct'. In the refined case, straightforward projection of the internal clause is also invalid. However, there is now
not only a refined version of the relative correctness assumption for X, but

274

REFINEMENT OF NOMIC TRUTH APPROXIMATION

there are also two interesting stronger sufficient extra conditions, i.e., conditions
which guarantee the projection of the internal clause. We will consider all three
conditions in the order of decreasing strength, and we will argue in passing
that the further relativization of falsification we noted in the stratified basic
case can be extrapolated to the stratified refined case.
Let X and Y satisfy (Rii) with regard to Tt:
(Rii)t

for all x in X and z in Tt


if rt(x, z) then there is y in Y such that St(x, y, z)

and what we are eager to prove is:


(Rii)o

for all Xo in nX and Zo in To if ro(x o, zo) then


there is Yo in nY such that so(x o, Yo, zo)

Due to the nature of n, this is equivalent to


for all x in X and z in Tt if ro(n(x), n(z then
there is y in Y such that so(n(x), n(y), n(z
Let us first realize what violation of (Rii)o means in the face of (Rii)p which is
the easiest to formulate when we conceive structurelikeness as the concretization
relation: X has 'accidental' observational idealizations of members of Tt upon
which Y does not improve. Hence, a necessary and sufficient condition to fill
the gap is to assume that X does not have such accidental observational
idealizations. In the case of violation, the relevant members of Tt are internal
mistakes X and Y, hence potential counter-examples, which are, on the observationallevel, better approached by X than by Y, but not on the theoretical level.
As a consequence, like in the stratified basic case, the stratified refined case
leads again to a further relativization of falsification, for 'more serious' counterexamples of Y may well be claimed to be accidentally less serious for X, while
Y is nevertheless on the theoretical level, i.e., taking 'everything' into account,
closer to Tt than X.
It is easy to check that (Rii)o directly follows from (Rii)t and the assumption
that all Vt-structures are comparable:
(AI)

for all x and z in Mp(Vt) rt(x, z)

Note that (AI), which trivially implies that all observational structures are also
comparable, is a rather strong condition. But there may well be cases where it
is satisfied: for instance, the case of propositional structures based on fixed
finite sets of elementary propositions formulated in observational and theoretical terms.

REFINEMENT OF NOMIC TRUTH APPROXIMATION

275

A weaker sufficient condition (in combination with (Rii),) assumes that


observational structurelikeness guarantees structurelikeness on the t-level:
(A2) for all x and z in Mp(Vt) if ro(n(x), n(z)) then r,(x, z)
Although (A2) is weaker than (AI), it is also a strong condition. Because
comparability presupposes at least that the domain-sets are the same, (A2)
excludes observationally comparable structures with different domain-sets of
theoretical terms.
A still weaker sufficient condition leaves this possibility open. But, whereas
(Ai) and (A2) were general conditions, the weakest sufficient condition is
specific in the sense that it imposes a restriction on one of the theories in
question. It amounts to the refined version of the condition we introduced for
the case of stratified basic truthlikeness. As is easy to check, (Rii)o is also
guaranteed by (Rii), if X is relatively correct in the refined sense, indicated by
RC,(X, t, 0) or simply RC,(X):
(A3) = RC,((X)

for all x in X and z in Tt if ro(n(x), n(z)) then there


are x' in X and z' in Tt such that n(x') = n(x) and
n(z') = n(z) and r,(x', z')

Note that (A3) reduces to the basic version RC(X) when ro and r, both are
based on trivial structurelikeness. In Chapter 9 we argued that RC(X) was not
a very restrictive condition on X. However, it is much less clear how restrictive
RC,(X) in fact is. In the case of concretization, it amounts to the claim that X
has no 'accidental' observational idealizations of members of Y. That is, a
theoretical idealization corresponds to any observational idealization. As in
the basic case, this at least means that X is on the right track, but now it is
much more far-reaching, since the relation of (observational) concretization
applies, of course, much more frequently than that of perfect observational fit.
It is interesting, also for the basic version, to consider the nature of (A3) in
some detail. Let RC,(X, Z) indicate that X, being of the t-level, is relatively
correct for Z with respect to the underlying o-level in the sense defined by
(A3). As we have seen in Subsection 10.1.1., it may well be that r is either an
equivalence relation or a partial ordering. It is not difficult to prove that
RC,(X, Z) is, in this case, also an equivalence relation or a partial ordering on
subsets of Mp(Vt), respectively. In the equivalence case, the theories considered
may all be relatively correct with respect to each other, in which case all or
none of them belong to the equivalence class associated with Tt. In the case
of a partial ordering, the theories considered may form an ordered sequence
which mayor may not end with Tt, but many of such paths will end in Tt. It
is easy to check that the basic general version of (A3), based on trivial structurelikeness, RC(X, Z) is an equivalence relation.
Although (A3) reduces to RC(X) when s is trivial, the two stronger sufficient
conditions, (Ai) and (A2) have no interesting analogue in the basic approach.

276

REFINEMENT OF NOMIC TRUTH APPROXIMATION

For, recall that trivial structurelikeness implies that two different structures are
always incomparable. As a consequence, (Ai) is excluded as soon as Mp(Vt)
contains more than one element and (A2) reduces to the condition that n is a
one-one function, which is, of course, an improper extreme case of stratification.
If the internal clause is projected for To, it follows also directly that n Y is
at least as successful with respect to R as nX. Hence, if Y is, with respect to
the internal clause, at least as close to Tt as X and if (Ai) or (A2) or (A3)
hold, then n Y is at least as successful with respect to R as nX.
In sum, the Projection Theorem can be refined: projection of refined truthlikeness on the level of theoretical terms is again guaranteed under some interesting
conditions, with the relevant success consequence as a corollary in combination
with the Refined Success Theorem.
It is evident that a similar story can be told about the relation between the
r- and the o-level, and also, but more trivially, about the relation between the
t- and the r-Ievel. As in the stratified basic case, the sufficient condition for
projection, RC,(X), is trivial between the t- and the r-Ievel, since Tt contains
all conceivable expansions of Tr.
Again we also have a refined version of the Upward Theorem, now stating
that if n Y is closer to To in the refined sense than nX then y* is closer to Tt
in the refined sense than X*, where
X* =dejX - {x EX - Y/n(x) E nyn To}
U {XEX - YjlVzEn-'To r(x,z)=>3YEXn Ys(x,y,z)}
and y* = dejX n Y Hence, y* is the same as in the basic case, whereas X* is
a further strengthening of its corresponding version in the basic case.
In view of the similarity between the basic and the refined version of the
Success/ Forward and Projection/ Upward Theorems we can argue that
HD-evaluation and RS, are functional for observational, substantial and theoretical truth approximation in essentially the same way as in the basic case.
Since the refinement introduced in this chapter is basically a comparative
matter, the definitions of Tt-reference and of the (total) referential claim of a
theory remain unchanged, and hence the analysis of referential truth approximation is not affected. In sum, the results of refinement and epistemological
stratification can be integrated.

Concluding remarks
In this chapter we have presented a conceptually plausible definition of refined
truthlikeness of theories based on an underlying notion of structurelikeness.
Taking into account the assumed fixed character of the vocabulary (hence of
the set of conceptual possibilities), it allows minimally the conclusion that
conceptually relative but otherwise objective truth approximation is possible
in a sophisticated sense, for example by concretization. Moreover, it justifies

REFINEMENT OF NOMIC TRUTH APPROXIMATION

277

the corresponding, intersubjectively applicable, methodological rule to choose,


whenever possible, the more successful of two theories.
In sum, together with the basic explication, our explication of the idea of
truthlikeness is coherent and conceptually attractive and fruitful. It is tempting
to mention one other fruit, the plausible explication of the everyday expression
"the truth lies in between": in terms of structurelikeness, when the true description t is concerned: s(x, t, z)&s(z, t, x); and in terms of basic or refined theorylikeness when the true theory T is concerned: MTL(X, T, Z)&MTL(Z, T, X) and
MTL,(X, T, Z)&MTL,(Z, T, X), respectively.
Moreover, as we have seen in this and the foregoing chapters, our basic and
refined explications allowed basic and refined explanations of the global success
of (natural) science by assuming that vocabularies for natural domains contain,
as a rule, a unique, time-independent subset of nomic possibilities. For this
so-called Nomic Postulate is sufficient to explain local progress in success on
the basis of (vocabulary-relative) truth approximation. We showed that, assuming the Truth Projection Postulate and the s-Projection Postulate and the
Fiction Postulate, this is even the case, under a certain condition, when a
distinction is made between theoretical and non-theoretical terms, for under
that condition, basic and refined truthlikeness are projected on the observational level. As a consequence, the notions of basic and refined truthlikeness
relativize, to some extent, incommensurability claims between theories formulated within different vocabularies. Hence, our analysis of truthlikeness clearly
paves the way for fundamental, or at least pragmatic, commensurability of
related theories.
Finally, it should be noted that our elaboration of the refined perspective
did not suggest any relativization of our conclusions about the epistemological
positions and their relation as presented in the end of the previous part. On
the contrary, since all theorems relating the positions have refined versions the
conclusions also apply in the refined perspective. Hence, in view of the real life
character of the refined approach, to be further illustrated in the next chapter,
the epistemological and methodological diagnosis in Section 9.4. may now be
claimed to apply to the natural sciences.

11
EXAMPLES OF POTENTIAL TR UTH
APPROXIMATION

Introduction
In the previous chapter we showed that the refined definition has real life
scientific examples, viz., by pointing out that the Law of van der Waals provides
a perfect case of (refined) potential truth approximation, in particular by
(double) concretization. Recall that a sequence of three theories is called a case
of potential truth approximation when the second is closer to the third than
the first is to the third. Two other scientific examples of such sequences have
been studied as well. In (Hettema and Kuipers 1995) it is demonstrated that
Sommerfeld's reconstruction of the sequence consisting of the theories of
Rutherford, of Bohr and his own theory, i.e., the successive stages of what now
is called 'the old quantum theory', is also a case of potential truth approximation. In the previous chapter (Section 10.4.) we explained that the technical
definition of basic and refined truth approximation can also be directed at
so-called provable interesting truths. In (Cools, Hamminga and Kuipers 1994)
it is shown that the theory of the capital structure of firms of Modigliani and
Miller is closer to a provable interesting truth than the original theory of Kraus
and Litzenberger.
In this chapter both examples will be presented as far as the formal aspects
are concerned. It should be stressed that these examples could only be analyzed
with the help of experts in the field and that the following text is heavily based
on the formal sections of the two papers mentioned. We have only added some
remarks about reference.
In sum, we will show that some important examples of successive theory
transition can be reconstructed as cases of potential truth approximation. This
illustrates that the refined definition is apparently in accordance with the basic
heuristics of truth approximation: steps forward in the direction of the truth
are made by improving theories according to shared scientific standards.
11.1. THE OLD QUANTUM THEORY

Introduction

This section presents a case of potential truth approximation in physics 1 : the


sequence of relevant theories of the atom of Rutherford, Bohr, and Sommerfeld

278

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

279

(the 'old' quantum theory of the atom), in the form as given in Sommerfeld's
Atombau und Spektrallinien. In particular, it will first be shown in structuralist
terms that Bohr's theory is a specialization of Rutherford's theory and that
Sommerfeld's relativistic theory in its turn is a concretization of Bohr's theory.
Then it will be shown, in general, that a specialization followed by a concretization may well be a case of refined truth approximation, which would explain
the increase of explanatory success of the successive theories.
Since its first appearance in 1919, Sommerfeld's Atombau und Spektrallinien
(Atombau for short) served as a comprehensive treatise on the structure of
atoms and spectral lines, and, as such, was sometimes called the 'Bible of
atomic theory'. It was even to survive the birth of quantum mechanics in 1926,
and the book, though supplied with a Wellenmechanischer Ergiinzungsband,
served its purpose well into the thirties. It was revised many a time after its
original publication and first translated into English in 1923. The book was
originally written at a time when, according to Lakatos' reconstruction of the
'old' quantum theory as a research program (Lakatos 1970, 1978), the old
quantum theory was on the verge of entering its degenerative phase. In Lakatos'
reconstruction, Bohr's program had been progressive since its inception in
1913, started to degenerate around 1920, and was finally replaced by the new
quantum theory of Schrodinger and Heisenberg in 1926.2
This points out that, from the perspective of the historian of science, the
Atombau is placed in a peculiar position: it is not very often that a statement
of theoretical principles, particularly when written in such a clear style as that
of the Atombau, survives the replacement of those principles by many years.
Part of this may be due to the fact that, in some parts, Sommerfeld's book
reads like a history of science itself, and has thus remained of interest to
practicing scientists long after the rejection of its theoretical basis. But perhaps
more important in this respect was that, in some parts of atomic theory at
least, the explanations offered by the Atombau were immensely successful in
the explanation of empirical facts and remained so even after the construction
of quantum mechanics.
A very important aim in the development of theories of the atom after the
success of the Bohr (1913a/b) theory of the atom was the explanation of atomic
spectra. Balmer's law, predicting the lines of the hydrogen spectrum, had
already been explained by Bohr's theory. Starting from Bohr's postulates,
Sommerfeld's achievement was twofold. Firstly, he analyzed the theory of
spectra in the presence of perturbing fields in great detail, thus providing an
analysis of the Stark effect in the case of electric and the Zeeman effect in the
case of magnetic fields. Furthermore, he convincingly explained the observed
fine-structure in the Balmer lines by blending the quantum principles pioneered
by Bohr with the theory of special relativity. Our discussion will focus on the
second aspect.
Within the context of truth approximation, a large amount of empirical
success is supposed to be indicative of 'being close to the truth'. In this section

280

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

we will show that the logical structure of the Atombau can well be described
in terms of refined truth approximation. We will start by a self-contained
presentation of the necessary formalizations, followed by the truthlikeness
analysis and some conclusions. For a sketch of the historical and scientific
background the reader is referred to the original paper.
11.1.1 . Structuralist reconstruction of the relevant theories

In this section we present set theoretical formalizations of four one-electron


atomic theories: the one of Rutherford, the crucial one of Bohr, and two of
Sommerfeld. Our presentation follows Sommerfeld's didactic presentation in
the Atombau quite closely, in that we first specify a one-electron atom, then a
set of quantum conditions, after which Bohr's theory of the atom is constructed
as a special case of the more general Sommerfeld theory. From Sommerfeld's
analysis, it was immediately clear that Bohr-except for reasons of simplicityhad no good reasons to restrict attention to circular orbits. Hence, we follow
Sommerfeld in his elliptic reconstruction of the theories of Rutherford and Bohr.
We see a 'Rutherford' model as describing a classical theory of the atom
where the electrons move in elliptical orbits around a central nucleus with
charge Z. The 'Bohr' and 'Sommerfeld' theories are semi-classical, where the
Bohr theory allows only circular orbits and the two Sommerfeld theories, one
relativistic, the other non-relativistic, allow elliptic orbits.
We start with the so-called potential models (Mp) for one-electron atoms.
DEF. 1: x = (T, r, rp, E) is a one-electron atom (x

Mp(A-l)) iff

(1) T is a subset of: a time interval;


(2) r is a function from T to IR1 +: the distance of the electron

to the nucleus(-center);
(3) rp is a function from T to [0,2nJ: the angular position of

the electron;
(4) E is a function from T to IR1 +: the total energy of the atom.

(Ad. (2) and 3), time-derivatives of rand rp are indicated by a


dot on top)
Note first that the vocabulary consists of, besides the (relatively speaking)
observational time interval domain-set, three evidently theoretical terms, viz.,
the radial distance and angular position of the electron and the total energy
of the atom. The three theories to be considered will be such that these terms
are claimed to refer; whether they actually do is another story, beyond potential
truth approximation. Moreover, since all three theories pose these terms, the
question of whether there is referential progress in this respect is outside our
scope. The only extra theoretical terms that will be introduced are the quantum
numbers postulated by the theories of Bohr and Sommerfeld.
To be sure, the theories presuppose the notions of atom, nucleus and electron,

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

281

and they will use the notions of the charge of the nucleus ('atom without the
electron'), the (rest) mass and charge of an electron, Planck's constant and the
velocity of light. In the study of the structure of the atom, all these terms were
already supposed to refer, and the (rest) mass and charge of an electron and a
nucleus, Planck's constant and the velocity of light, were even measured, hence
conceived as observable, at the time.
We will now present the models (M) of the four theories of the atom, each
time followed by one or two remarks. Each formalized theory allows one set
of 'energy-states', each set characterized by an appropriate set of quantum
numbers. The sets differ for each theory we discuss. Atomic spectra are predicted
by taking the energy difference for two different states in a set and dividing
this difference by the constant of Planck. 'Explanation' obtains when the
so - predicted expression for the frequencies entails the empirical law to be
explained.
We start with Rutherford's theory of the (one-electron) atom: RA
DEF. 2: x is a Rutherford atom (x

M(RA)) iff

(1) x E Mp(A- l)

(2) (m/e: mass/charge of electron; Ze: the charge of the nucleus


for some integer Z)
m
2

E(t) = - [f2

Ze 2
r

+ r2tj12] - -

(i)

(3) E(t) is constant in time

Note that (i) specifies the total energy in terms of the kinetic energy, further
subdivided into a 'radial' and an 'angular' component and the potential energy,
the latter as given by Coulomb's law.
The following theorem, essentially a variant of Kepler's law of the elliptic
orbit of planets, was given by Sommerfeld for the electron in the atom.

Theorem 1: in a Rutherford atom the electron 'describes' an elliptic


orbit, with the nucleus in one of the focal points. More precisely, there
are positive real numbers a and e such that
(j)

We now jump to the crucial theory of Bohr (BA), introducing the quantization hypothesis. Since our present formalization most easily admits
Sommerfeld's formula for the quantization hypothesis, we will use Sommerfeld's
formulation, and derive Bohr's quantization hypothesis from it. Sommerfeld's

282

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

quantization hypothesis, adapted to our special case of an electron in an elliptic


orbit, can be given in the form

Quantization hypothesis (QH): there are integers nr, nq> such that (h is
Planck's constant)
m

LX> r dr = nrh

which quantizes the distance to the nucleus and the angle, so both variables
characterize together the elliptic orbit of the electron. We give the formalization
of Sommerfeld's non-relativistic version of Bohr's theory in the following:
DEF. 3: x is a Bohr atom (x E M(BA)) iff
(1) xEM(RA)

(2) QH
Now it is possible to prove the following:

Theorem 2: a Bohr atom is such that


E(t)=-

2n 2 mZ 2 e4 [
1
]
h2
(nq>+nr)2

(ii)

As an aside, it should be noted that the Bohr theory of the atom under this
construction is not equivalent to the 1913 Bohr theory of the atom, where only
circular orbits are allowed. For the plausible reconstruction of the authentic
Bohr theory as a specialization of the one above, we refer to the original paper.
It is important to note that M(BA) is a proper subset of M(RA).
Conceptually, this means that Bohr's theory is a so-called specialization of
Rutherford's theory. More precisely, it is a theoretical core specialization in
the sense that it concerns the models and hence the core of the theory, and the
quantum numbers n" and nr are extra theoretical terms.
Before we define Sommerfeld's relativistic theory of the one-electron atom,
we first introduce the relativistic version of Rutherford's theory: rRA.
DEF. 4: x is a relativistic Rutherford atom (x E M(rRA)) iff
(1) x E Mp(A- 1);
(2) (e: velocity of light; mo : rest mass of the electron)

E(t) = e2 mo

(hfJ2 -1) _
1-

Ze
r

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

283

with

(3) E(t) is constant.

Sommerfeld showed that, as for the orbits of planets, the relativistic orbit is
only approximately an ellipse (more precisely, it is an ellipse with an advancing
perihelion):
Theorem 3: in a relativistic Rutherford atom the electron describes a
precessing ellipse (relativistic Kepler orbit), to be precise, there are a
and e such that

with

where P", is the moment of momentum mv, the areal constant. As e goes to
infinity, it is easily seen that yapproaches 1 (see (Sommerfeld 1919, p.254.
Sommerfeld's final goal was the quantization of this relativistic Rutherford
atom, or in our terms, Sommerfeld's theory of the atom: SA.
DEF. 5: x is a Sommerfeld atom (x E M(SA iff

(1) x E M(rRA);
(2) QH.

It can be proven that


Theorem 4: a Sommerfeld atom is such that
E(t)=m o e2

with

(J,

{ [ 1+

(J,2Z2

(nr + -In; -

J-

1/2

-1 }

(ii )r.1

(J,2 Z2)2

defined as the fine-structure factor


2ne 2
(J,=--

he

In view of the fact that M(SA) is a proper subset of M(rRA), Sommerfeld's

284

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

theory is a specialization of the relativistic version of Rutherford's theory, hence


again leading to Sommerfeld's precessing ellipses. Moreover, now it is possible
to prove the following
Limit theorem: if c goes to infinity then (i)rel> (ii)rel and (j)rel reduce to
(i), (ii) and (j), respectively, for m = mo.

The proofs of the reduction of, in particular, the two characteristic formulas
(i)rel and (ii)rel require sophisticated mathematical approximation. Be this as it
may, we may conclude that the relativistic models are concretizations of the
respective non-relativistic models, in the technical sense that the former reduce
to the latter by a limit procedure. Applying the relevant technical definition of
concretization of theories (see Section 10.4.), we may conclude, in view of the
fact that every non-relativistic Rutherford-/ Bohr-atom has a relativistic concretization, that the relativistic Rutherford theory rRA is a concretization of the
(non-relativistic) Rutherford theory RA, and also, specifically, that the
Sommerfeld theory SA is a concretization of the Bohr theory BA.
Figure 11.1. depicts the formal situation of the classes of models of the
theories (indicated by the names of the theories).

Mp(A-1)

RA
==>

..... > concretization


exclusion by specialization

Figure 11.1. Relations between RA, rRA, BA, SA

<

In the next subsection we will prove that the sequence RA, BA,SA), due
to these formal relations, is a case of potential (refined) truth approximation.
11.1.2. Truth approximation analysis

In view of the fact that the present example is not a case of two successive
concretizations, the Double Concretization (DC-)theorem presented in the

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

285

previous chapter (Section 10.4.) can not be applied. We need some adapted
version which is suitable for specialization, followed by concretization. It turns
out to be possible to use much of the concretization apparatus as developed
in Section 10.4. In its terms we need some additional technical preparations.
Let us define, for a subset Y of X, when CON(X, Z), Z(Y) as the restriction
of Z to concretizations of members of Y:
Z(Y) = {z E Z j there is y E Y con(y, z)}

It is easy to check that CON(Y, Z(Y.


Let con+(x, y) indicate "con(x, y) and x"# y", hence, exclude the extreme
reflexive case. For a specific kind of idealization (-procedure) it is by definition
the case that it is uniquely determined in the following sense: if con + (x, y) and
con + (x', y) then x = x'.
Now the following Specialization-Concretization (SC-)Theorem is easy to
prove: if Y is a (proper) subset of X, X and Z are non-overlapping such that
CON(X, Z)3 then MTLcr(X, Y, Z(Y.
Proof:

Ad. (Ri) Trivial, for Y - X U Z(Y) is empty, which follows directly from
the fact that Y is a subset of X.
Ad. (Rii) Let z E Z(Y), X E X and r(x, z), that is con(x, z), and, due to
the resulting empty intersection of Z( Y) and X, even con + (x, z). From
z E Z(Y) it follows that there must be y E Y such that con(y, z) and even
con+(y', z), due to the resulting empty intersection of Z(Y) and Y. Hence,
due to the unique determination of specific idealizations, x = y and
hence x belongs to Y. Together with con(x, x) we now get ct(x, x, z)
with x E Y.
In words, the theorem says in fact that a specialization of a theory, followed
by a concretization that amounts to the relevant restriction, i.e., specialization,
of a non-overlapping concretization of the original theory, is a case of potential
refined truth approximation. It also suggests the Truth Approximation by
Specializationj Concretization (TASC-)Corollary, which states that if theory Y
is a specialization of theory X and if the true set of nomic possibilities T equals
the relevant specialization of a non-overlapping concretization of X, then Y is
closer to the truth than X. The assumption that there is such a T, for the
domain of one electron atoms and the frame Mp(A-l), is the Nomic Postulate.
The crucial heuristic hypothesis for this truth approximation claim is that T
equals the relevant specialization of a non-overlapping concretization of X.
Note that, unlike the case of double concretization, there are no further technical conditions on X or Y in the SC-Theorem and hence the TASC-corollary.
It is easy to check that the four theories formalized in the previous section,
viz., RA, BA, rRA, SA, satisfy the conditions of the theorem, not only in the
informal sense, but also in the specified formal sense:

286

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

BA is a (proper) subset, hence a specialization of RA


rRA is a non-overlapping concretization of RA
SA is rRA restricted to concretizations of members of BA
Hence we may conclude from the SC-Theorem that MTLcr(RA, BA, SA), that
is, Bohr's theory is closer to Sommerfeld's theory than is Rutherford's theory.
As a consequence, the sequence is a case of potential refined truth approximation: as long as it is/was not excluded that SA is/was the true theory of the
(one-electron) atom, BA could be a step in the direction of the truth.
At first sight, it may come as a surprise to learn that the alternative sequence
<RA, rRA, SA), briefly 'concretization followed by specialization', is not a case
of potential truth approximation. The reason for this is instructive. Whereas it
is not difficult to prove clause (Rii) for the alternative sequence, it is impossible
to do so for clause (Ri). As we have seen, clause (Ri) requires in fact that all
mistaken, extra models of the intermediate theory should improve upon members of the starting theory; it should not include models that are redundant for
that purpose.
It is important to note that confronted with a historical sequence <X, Y, Z)
reflecting concretization followed by specialization, it will always be possible
to formulate a theory y* such that <X, Y*, Z) is a sequence of specialization
followed by concretization, and hence a case of potential truth approximation.
In this way, the concretization is restricted to the relevant specialization of the
original theory. Hence, the historical order of concretization and specialization
is not essential for potential truth approximation.
Recall the (refined) Success Theorem according to which closer to the truth
than another one will always be explanatorily at least as successful as the latter,
due to the Projection Theorem, unconditionally, even if it introduces new
theoretical terms. As a consequence, the fact that a theory is explanatorily
more successful than another can always be explained by the hypothesis that
it is closer to the truth. Recall now from the introduction that, in contrast to
RA, BA could explain the Balmer spectral lines, and SA, in addition, their finestructure. Also, the Sommerfeld atom was more successful in the description
of the Stark and Zeeman effects. Accordingly, in view of these explanatory
results and the fact that BA did not lose any explanatory success of RA, and
SA not of BA, BA was explanatorily more successful than RA and SA explanatorily more successful than BA. From the Success and Projection Theorem it
follows that these cases of empirical progress can be explained by the hypothesis
that BA and SA are closer to the truth than RA and BA, respectively. From
the additional fact that < RA, BA, SA) is a case of potential truth approximation, it follows that SA can still be the truth, i.e., the true theory about the
atom. Note that we focused above on the external, law, side of the success
comparison. The internal clause, comparing the seriousness of observational
counter-examples is left out of the picture, since this has not been systematically

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

287

reported by Sommerfeld or others, which may illustrate the 'further relativization' of falsification .
As has been explained in Chapter 9, it is difficult to attach referential truth
approximation to empirical progress, not only since referential truth approximation does not guarantee empirical progress, but, more importantly, since
neither theoretical nor substantial truth approximation imply referential truth
approximation. Hence, although the increase of success was deductive evidence
for the relevant truth approximation hypotheses, it was, at most, non-deductive
evidence for the claim that the quantum numbers refer.
It is interesting to note that the fine-structure can be obtained by concretization of the Balmer lines, and hence that the fine-structure spectrum reduces to
the Balmer spectrum when the velocity of light goes to infinity. Now it is also
plausible to think that many (if not all) consequences of a concretization of a
theory are concretizations of consequences of that theory. If it concerns (approximately) true observable consequences, then they are explanatory successes. To
illustrate the general point, we have included Theorem 1 and 3 and the reduction
claim of (j)rcl to (j) in the formalization. In view of the fact that the orbit of
an electron is unobservable, Theorem 1 (elliptic orbits for RA- and hence
BA-atoms) and 3 (precessing ellipse for rRA- and hence SA-atoms) and the
reduction claim that a precessing ellipse reduces to an ellipse, are simply
theoretical claims.
11 .1.3. Conclusion

We have shown that the progress in theories of the atom, as given in


Sommerfeld's Atombau, can be reconstructed as a case of refined potential truth
approximation. More specifically, within the conceptual space spanned by the
specialization / concretization directions represented in Figure 11.1., SA embodies both the quantum specialization and the relativistic concretization of RA,
whereas BA only embodies the quantum specialization of RA. We argued,
moreover, that this could explain that BA was (explanatorily) more successful
than RA, under the assumption that SA is 'the true theory', and that this could
explain that SA was more successful than BA, under the assumption that SA
is closer to the true theory than BA. However, the empirical progress is at
most non-deductive evidence for the reference claim that BA and SA add to
RA, viz., that the quantum numbers refer.
We do not know whether SA is (observationally) true, let alone whether it
yields 'the true (observational) theory' in the sense ofthe strongest true (observational) theory that can be formulated within the conceptual framework of the
old quantum theory. That is, we do not know of false consequences or unexplained phenomena that can be formulated within that framework. However,
the new quantum theory was started for other reasons than such problems (see
the first note to this section).
In the light of our conclusions, one may wish to review some of Lakatos'
statements on the development of the old quantum theory. In particular,

288

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

Lakatos' reconstruction of the constructive phase of the old theory of the atom
can be refined on many points, but his reconstruction of the 'degenerative'
phase is also open to criticism (Radder 1982). Unfortunately, Lakatos' reconstruction largely bypasses the role of Sommerfeld's Atombau, for it could have
profited from a closer consideration of it.
Lakatos' assessment of the situation around 1917 can be summarized in the
contention that the old quantum theory had its major portion of success. In
Lakatos' words, Bohr then 'got worried' (Lakatos 1970, p. 150) about the fate
of the 'old' quantum theory and as a consequence, had passed the initiative to
Sommerfeld. This, in our opinion, is oversimplified. Though one can only
surmise, the main reason for the further development of the theory by
Sommerfeld may have been his somewhat earlier adoption of the new formulation of the quantum conditions, which greatly simplified the theoretical progress. Secondly, Bohr, instead of 'getting worried' about the fate of his theory,
was led into a more rigorous formulation of his theory which he felt could
now no longer be postponed (Darrigol 1992, p. 100), since his theory had
gained a more widespread acceptance and its foundations should therefore be
clear. Our assessment of the situation from Bohr's viewpoint is that (i) from
1913 on Bohr saw his theory as a step towards a more final theory of the atom
but in no way as 'the' theory of the atom; and (ii) anticipating this, he tried to
determine which parts of his theory could be taken over into a more 'final'
theory of the atom and which could not.
Sommerfeld's 1919 Atombau, by its clear exposition of the theoretical principles of the 'old' quantum theory and its detailed consideration of the consequences of these principles, more or less precisely fit it into this intellectual
niche. Therefore, Sommerfeld's work is strongly complementary to Bohr's in
this phase of the development of the quantum theory. The Atombau, by its
clear exposition of the then current status of this rapidly developing field, was
accessible to both theorists and experimenters alike. In the subsequent editions,
the book kept up surprisingly well with the latest experimental and theoretical
developments. As such, it provided an indispensable basis for the training of
the younger generation of theorists who developed and applied quantum
mechanics after 1926, starting from both the achievements of both Bohr and
Sommerfeld. 4
11.2. CAPITAL STR UCTURE THEORY

Introduction

Recall that we indicated in Subsection 10.4.3. that the concretization apparatus


could deal with elementary and complex forms of concretization in the context
of validity research. The main point of the present section is to show under
what conditions and in what sense the theory of Kraus and Litzenberger (1973)
about the capital structure of firms is a concretization of the theory of

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

289

Modigliani and Miller that brings us closer to a provable interesting truth


about the actual financing behavior of firms.
In the first subsection a structuralist reconstruction is given of the relevant
theories. We refer to the original paper for an informal exposition of the crucial
results in capital structure theory. The second subsection explains in some
detail what the example illustrates: economists attempt to approach interesting
truths by concretization, and they may well succeed in this.
11 .2.1. Structuralist reconstruction of the relevant theories

This section displays the logical skeleton of "A state-preference model of optimal
financial leverage" by Kraus and Litzenberger (1973). As usual in theoretical
economics, the strategy of model construction is focused on an interesting
theorem (Hamminga 1983). Here, what is to be derived is a theorem on how
the value of (debt plus equity of) a firm depends on the proportion of debt in
that value. Modigliani and Miller (1958/ 1963) used a plausible model to arrive
at implausible, and therefore intriguing theorems: the proportion of debt has
no influence at all on the value of the firm if corporate tax is nonexistent, and
should be 100% if corporate tax exists. The model of Kraus and Litzenberger,
based on a 'state preference' approach, could explain the existence of an optimal
solution between 0% and 100%.
The 'state preference' approach consists of assuming a set r of all conceptually
possible end-of - period 'states of the world'. In every single possible state of
the world j there is a fixed and given return X fj for every firm f, a given tax
rate ~ and given bankruptcy cost CJj (C fj = 0 for firms fthat survive in state j).
It is a one period analysis. The return (earnings before interest and taxes)
Xfj of firm f in state j is identical to the end-of-period market value of firm f
If you knew what state j would be realized, you would have no problems of
uncertainty. Hence, the problem of uncertainty consists solely of investors not
knowing what state will be realized, and having different probability beliefs
concerning j E J.
A 'market expectation' about the states j E J is negotiated among investors
by assuming tradeable 'primitive securities' OJ, which can be thought of as
'lottery tickets' yielding 1$ if state j occurs and nothing if j' oF j occurs. The
equilibrium price Pj of OJ could be derived from (I) probability beliefs of
investors concerning state j (2) their time preference for holding money, (3) their
risk aversion and (4) their utility functions. Kraus and Litzenberger do not
perform this derivation, but take Pj' j E J, as given, exogenous variables. Any
other security bought by an investor can now be identified with a definite
number of lottery tickets OJ for every state j (since investors are assumed to
know the return of every security in every state).
Now we are ready to specify what structure a thing x should have in order
to be a conceptual possibility in the language that Kraus and Litzenberger
have chosen for their theory.

290

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

DEF.l: x = (1, F, '1', II, D, X, T, C, P, Y, Z, B, S, V) is a potential debt

equity market system with corporate tax and bankruptcy cost


E DETCp) iff

(x

(1) The symbols J, F, 'I' denote domain sets with the following
meanings:
J: the set of states of the world
F: the set of firms in the market system
'1': the set of primitive securities
(2) The symbols II, D, X, T, C, P, Y, Z, B, S, V denote functions, of which the meaning follows after the second ":":
II: J -+ '1': the primitive security IIj that has a return of 1$
if state j occurs, and 0$ in any state j' #- j.
D: F -+~;: the debt of firm f, a promise to pay a fixed
amount D f' irrespective of the state that occurs, a nonnegative real number
X: F x J -+ ~ x: the return of f in state j, the end of period
value of f, a real number XJj' negative in case of
bankruptcy.
T: J -+ ~~.11: the tax rate over X in state j, a real number
~ in the closed interval [0, 1].
C: F x J -+ ~~ : the bankruptcy cost Cf j of f in state j, zero
if the firm survives in state j.
(3) With the help of the primary functions II, D, X, T, C the
secondary functions P, Y, Z, B, S, V are defined:
P: 'I' -+~; yields the price Pj of primitive security II j
Y: F x J -+~; is the return to holders of debt Df of every
firm f in state j, where
Df for D f 5, X Jj (the firm 'survived')
{
YJj:= XJj-CJj for Df>XJj

(the firm is bankrupt)


Z: F x J -+~; is the return to the holders of the equities
of f in state j, where

Z fj =

XJj(1-~) + ~Df -

Df for D f 5, XJj

(the firm 'survived')

o for D f

> X Jj (the firm is bankrupt)

(4) The last three secondary functions have again a different


logical role:

291

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

B: ~; ---->~;: the present market value B(D,) of the debt


of firm f
S: ~; ---->~;: the present market value S(D,) of the equities
of firm f
V: ~; ---->~;: the present total value V(D,) of firm f, where
V(DI) : == B(D,) + S(D I )
Since Band S have only their domains and ranges specified, nothing is yet
said about how they could be determined from the variables introduced earlier.
This is all we do to specify the set DETCp. A thing x is a potential structure
of the Kraus and Litzenberger theory, or a potential model iff it satisfies all
requirements mentioned so far, and hence BI' SI and V, can be, in an
x E DETCp, any non-negative real number.
The set of models of the theory differs from the set of potential models by a
specification of how the present market value of debt and equity is determined.
This yields a subset DETC c DETCp of structures, the presentation of which
will be interrupted twice:

DEF.2: x is a debt equity market system with corporate tax and bankruptcy cost (x E DETC) iff
(1) x is a DETCp
(2) P is such that primitive securities in 'I' (OJ, j
market equilibrium prices Pj (j E J)5

J) have

For convenient notation of the requirements on the functions Band S, Kraus


and Litzenberger (re)number the states jl ,j2, ... , j. in J for every firm in climbing
order of the known return XJj in each state. So we have, for every f:
Xf! ~ Xf2 ~ ... ~ XJj ~ ... ~ XI" Now we can formulate clause 3)
(3) B is such that for all firms

B(D,) =

YJjPj

j=1

DI

Pj iff DI is such that D,~Xf!

j=1
k- I

j=1

j=1

(XJj - C Jj)Pj + DI

Pj iff D, is such that

j=k

3k> 1 (XI .k - 1 <D,<XId


(XJj - CJj)Pj iff DI is such that DI> XI.

292

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

The first of these three functions is meant for the case in which debt D f is such
that the firm survives (will not go bankrupt) in every possible state. The second
function is meant for the cases where debt is such that the firm is bankrupt in
state j = 1, .. ., k - 1 and survives in the remaining states. The third function is
meant for the case in which debt is such that in all states the firm will be
bankrupt.
The present market value of debt is thus reduced to the present market value
of primitive securities. The same is done for the equity of f:
(4) S is such that for all

S(D f ) =

ZJjPj

j=1
n

j=1

[XJj(1-1])+ 1]Df -D f JPj

iff Df is such that Df

Xf!

j=k

[XJj(l-1])+ 1]Df -D f JPj

iff D j is such that 3 k > 1 (Xf .k- I < Df < X fk )

o iff Df

is such that Df > X f

The meaning of the tripartition is again bankruptcy in no/some/all possible


end-of-period states.
This ends the list of additional requirements that a structure x E DETCp
must meet in order to be an element of DETC c DETCp. The list is the result
of a strategy meant to find the class of structures for which the following
theorem T-DETC holds:

Theorem T-DETC: in all structures x E DETC, for all firms f, if 1] is


identical in all states j (dropping subscript j) ,
V(D f ) = V(O)

+ T B(D f ) -

(1- T)

o iff Df is such that Df ~ Xf!


k-I
I CJjPj iff Df is such that
j=1
n

j=1

CJjPj iff Df is such that Df>Xfn

Figure 11.2. depicts an example of a shape that V(D f) could take for some firm

f The slopes of the line segments are determined by T and Pj (Kraus and

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

293

V(Df)

MAX

J,

~~

V(O)

I
I
I

Xf1

I
I

I
I
I

X12 X13

I
I
I

X14

X1n

01

Figure 11.2. Typical shape implied by T-DETC

Litzenberger 1973, p.916), and every time D J passes the value at which it
equals some state's total return of the firm, XIi' the bankruptcy costs of that
state kick down the value of the function V(D J).
T-DETC implies (1) that there can be a unique optimal amount of debt DJ ,
(2) this optimum amount of debt can be 'interior', that is, larger than zero and
smaller than the total return of the firm in the firm's most lucrative state (XJ .).
Figure 11.2. is drawn as an example of this possibility. This possibility arises
as a result of the introduction of bankruptcy cost in the model. If bankruptcy
costs are removed from the structures in DETC and DETCp, a poorer set of
structures, DET(p), results 6, where x E DET(p) is called (potential) debt equity
market system with corporate tax. This brings us back to the original paradoxical
Modigliani-Miller theorem on the optimal amount of corporate debt in the
case of the existence of corporate tax; setting CJj = 0 in T-DETC immediately
yields T-DET:
Theorem T-DET: in all structures x E DET, for all firms f, if ~ is identical
in all states j then V(D J ) = V(O) + T B(D J )

Figure 11.3. depicts the typical shape of V(D J ). T-DET implies, like T-DETC,
the possible existence of a unique optimal amount of Debt DJ but if it exists,
it is a 'corner solution': the more debt, the higher the value of the firm, therefore
100% debt financing is optimal.
If, finally, tax ~ is removed from the structures 7, we obtain DE(p), where
V(DfJ

V(O)

Xf1

X12 X13

X14

Figure 11.3. Typical shape implied by T-DET

01

294

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

X E DE(p) is called a (potential) debt equity market system. This yields the initial
Modigliani-Miller Theorem T-DE.

Theorem T-DE: in all structures

XE

DE, for all firms/, V(D J ) = V(O)

This means that the debt-equity ratio is irrelevant to the value of a firm. The
graph becomes a horizontal line.
History would have been nice and simple if Modigliani and Miller had first
discovered T-DE in the Kraus and Litzenberger framework presented above,
then had introduced tax to arrive at T-DET, and finally that Kraus and
Litzenberger enriched the theory with bankruptcy cost to arrive at T-DETC.
However, Modigliani and Miller did not use the state preference framework.
For what this example is meant to illustrate below, we do not need to go into
the logical intricacies of reducing Modigliani and Miller's method of introducing market forces to the method used by Kraus and Litzenberger.
11.2.2. Truth approximation analysis

We now apply the concretization apparatus to the case of capital structure


theory. We will present the applications in three rounds. In the first application
round we concentrate on the relations between the three theories, conceived
as sets of models. In the second round we focus on interesting theorems and
their domains of validity. Finally, in the third round we will characterize the
heuristic strategy and hypotheses of the researchers which are typical for the
present case of truth approximation. Only in the last round is the application
of the DC-theorem of substantial importance, but we will also mention its
application in the previous rounds. Since the extra terms, viz., corporate tax
and bankruptcy costs, are just observation terms, there is nothing like referential
truth approximation at stake.
Concretization of theories

Recall from Subsection 11.2.1. that DE, DET, and DETC are subsets of the
set DETCp of potential Debt Equity market system with corporate Tax and
bankruptcy Costs satisfying the extra conditions 2, 3 and 4 of DEF.2. Moreover,
in the case of DE-models, the bankruptcy cost function C and the corporate
tax function T are uniformly zero, whereas in the case of DET-models, only C
is uniformly zero, and in the case of DETC-models, neither T nor Care
uniformly zero. Note that by trivial consequence DE, DET and DETC are
mutually non-overlapping.
Let z be a DETC-model. It is not difficult to check that there is a (unique)
DET-model y and a (unique) DE-model x such that con (x, y) and con(y, z),
and hence ct(x, y, z). From this it is a small step to prove directly
MTLct(DE, DET, DETC), i.e., DET is closer to DETC than DE on the basis
of concretization, represented in Figure 11.4. This result is also easy to obtain

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

295

DETCp

DET

DETC

DET is closer (more similar) to DETC than DE


Figure 11.4. Concretization of theories

via the DC-theorem from the fact that DET is convex and mediating and the
combination of CON(DE, DET) and CON(DET, DETC).
Concretization of interesting theorems

In the second application round we introduce the interesting theorems in


Hamminga's sense (Subsection 11.2.1.) and their domain of validity. Recall that
theorems T-DE, T-DET, and T-DETC were associated with DE, DET and
DETC, respectively. Let the consequents of T-DE, T-DET and T-DETC be
indicated by CT-DE, CT-DET and CT-DETC, respectively. Let VCT-DE,
VCT-DET and VCT-DETC indicate the subsets of DEp, DETp and DETCp
for which these consequents are respectively provable, to be called their respective domains of validity. (For technical reasons we restrict VCT-DE and VCTDET to DEp and DETp, respectively, assuring that the three domains of
validity under consideration are mutually non-overlapping; extension to
DETCp (only) meets some technical complications.)
The first crucial result of Modigliani and Miller can now be rephrased as
the proof that DE is a subset of VCT-DE, i.e., CT-DE is provable for
DE-models8 . Their second result amounts to the proof that DET is a subset
of VCT-DET. Finally, the crucial result of Kraus and Litzenberg is the proof
that DETC is a subset of VCT-DETC.
Let us call in general, assuming some set of potential models Mp, a theorem
T* a concretization of T if and only if the domain of validity of T* in Mp is a
concretization of that of T, i.e., CON(VT, VT*). This relation between two theorems may well be provable without the above mentioned type of subset proofs,
as in the present case. That is, it is easy to check that CON(VCT-DE, VCT-DET)
and CON(VCT-DET, VCT-DETC), and hence, by the DC-theorem and noting
that VCT-DET is convex and mediating, that VCT-DET is closer to VCT-DETC
than VCT-DE on the basis of concretization. This is depicted in Figure 11.5.

296

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

DETCp

VCT-DET is closer to VCT-DETC than VeT-DE


Figure 11.5. Concretization of interesting theorems

Again, the latter result, i.e., MTLct(VCT-DE, VCT-DET, VCT-DETC), is also


easy to prove directly. Recall that we have already arrived to the conclusion
MTLct(DE, DET, DETC) in the first round.
The heuristic strategy of Kraus and Litzenberg

Why does concretization make sense? What further goals were Kraus and
Litzenberg aiming at? For the third, and final, round let us start from the basic
result of Modigliani and Miller to the effect that DE is a subset of VCT-DE.
We will omit further reference to the intermediate result concerning (T-)DET
and use as many verbal formulations as possible.
The more or less explicit heuristic strategy of Kraus and Litzenberg which
brought them to (T-)DETC can be characterized as follows. Look for DE*
and CT-DEI such that DE* is a convex and mediating concretization of DE,
CT-DE* is a convex and mediating (interesting) concretization of CT-DE and
DE* is a subset of VCT-DE*. This strategy is depicted in Figure 11.6.
DETCp

EP
DE-" VCT-DE* is closer to EP" VCT-EP than DE /I VCT-DE
Figure 11.6. Heuristic strategy: look for DE* and VCT-DE* such that ...

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

297

The motivation for this strategy can be given in terms of an unknown set of
(potential) economically possible systems EP (EPp), including, of course, the
actual ones EA. To be precise, the following heuristic hypotheses form 'the
good reasons' for this strategy.

Hi For the domain of firms and the set of potential economically


possible systems DETCp, there is a unique subset EP characterizing the
economically possible systems, the Nomic Postulate, and a unique subset
EA of EP characterizing the economically actual systems,
H2 all relevant sets can be reformulated as subsets of DETCp,
H3 EP is a concretization of DE*,
H4 there is an interesting theorem CT-EP such that
(a) VCT-EP is a concretization of VCT-DE*
(b) CT-EP is provably true for EP-systems,
and hence provably true for AE-systems.
If these hypotheses are true and if DE* and CT-DE* are in accordance with
the explicit heuristic goals, we may conclude that DE* is closer to EP than
DE, and that VCT-DE* is closer to VCT-EP than VCT-DE. In both cases the
application of the DC-theorem is crucial, as long as we do not dispose of EP
and CT-EP.
The four hypotheses mayor may not be true. There does not seem to be a
reason why they are unlikely to be true, except for the fact that economic
reality is infinitely complex. This objection however can be conceded by assuming only that the hypotheses are approximately true in some intuitive sense.
The other crucial point is H3: EP will only be a concretization of DE* if
the relevant factor(s) introduced in the concretization of DE to DE* have been
accounted for in an adequate way. Of course, it is impossible to evaluate the
adequacy of the concretization of DE to DE* in a conclusive way. Hence, we
have to rely on good reasons, which may in general be of an empirical and/or
theoretical nature, leading to empirical and/or theoretical concretization. The
example of Van der Waals is a clear case of theoretical concretization followed
by empirical support. For Kraus and Litzenberger, the reasons for engaging in
concretization were empirical doubts about the validity of the original M&Mresult for real world situations in general. The reasons for believing in the
adequacy of the concretization were only of a theoretical economic nature.
11.2.3. Conclusion

We may summarize the foregoing analysis of an example from economics as


follows. By concretizing a set of models (i.e., a theory) and a provable interesting
theorem about them, we try to approximate a provable interesting truth about

298

EXAMPLES OF POTENTIAL TRUTH APPROXIMATION

economically possible, and hence about economically actual, systems. The


heuristic hypotheses guaranteeing the success of this attempt may well be true,
provided that the concretization of the theory is not an empirically empty
gesture. Besides being a (convex and mediating) concretization in the logicomathematical sense, it has to be plausible on theoretical and/or empirical
grounds.

12
QUANTITATIVE TRUTH LIKENESS AND TRUTH
APPROXIMATION

Introduction
In Subsection 10.2.2. we have elaborated our scepsis about the idea of quantitative truth approximation, mainly because it presupposes real-valued distances
between structures. Since such distances are not presupposed by quantitative
confirmation we are even more skeptical about quantitative approaches to
truth approximation than to confirmation. However, we will nevertheless present some main lines of such approaches. We will first deal, in Section 12.1.,
with quantitative actual truthlikeness and truth approximation, which is relatively unproblematic when there is a plausible distance function on Mp. For
the basic nomic case, to be dealt with in Subsection 12.2.1. and part of
Subsection 12.3.1., the situation is also relatively clear, but the refined nomic
case faces us with essentially two types of problems. First, for the idea of
quantitative truthlikeness, discussed in Subsection 12.2.2., we have to find a
plausible distance function between two theories, and hence between two
subsets of Mp, based on a distance function on Mp. Even guided by some
plausible principles, our best refined proposal is not completely satisfactory.
Second, we have to choose between two essentially different ways of measuring
the quantitative success of theories: a non-probabilistic and a probabilistic way.
In the first case, presented in Subsection 12.3.1., the success, or better the
failure, of a theory in meeting the available evidence is quantitatively expressed
in a similar way as the chosen distance function between two theories. The
second case, presented in Subsection 12.3.2., arises from an adaptation of
Niiniluoto's plausible idea of estimating the distance of a theory from the
complete, hence actual, truth on the basis of the available evidence. It will be
shown that there is, in the case that probabilities as well as distances are
plausible, a very sophisticated way of doing so, where ideas of Carnap, Hintikka
and Festa are combined, leading to 'double' truth approximation. However, it
will also be argued that this primarily concerns improper cases of nomic truth
approximation, that is, it amounts to refined actual truth approximation in
two respects.
The announced results for quantitative nomic truthlikeness and truth approximation are restricted to non-stratified theories based on a finite Mp. Since
they are not yet impressive, we refrain from including considerations dealing

299

300

QUANTITATIVE TRUTHLIKENESS AND TRUTH APPROXIMATION

with stratification, and we will only make a few remarks about extension to
an infinite Mp. We will conclude with some remarks about the intuitive foundation of the quantitative proposals.
In this chapter we will frequently assume or define a distance function or
metric d on Mp(V) = Mp or its powerset P(Mp). For a survey of definitions
and examples of distance functions, see (Niiniluoto 1987a, Ch. 1). We will use
the following general definitions: d: X x X -+ is a proper distance function (df)
on the arbitrary set X, and <X, d) a metric space iff
(01) d(x, y) ~ 0
(02)

d(x, x) = 0

(03)

d(x, y) = d(y, x)

(04)

d(x, y) :s; d(x, z) + d(z, y)

(05)

d(x, y) = 0 only if x = y

(symmetry)
(triangle inequality)

As weaker versions, we distinguish a deletion of one or more of the properties


(03), (04) and (05), leading to the qualifications 'quasi', 'semi' and pseudo'
for the pure forms, respectively. Hence, d is a quasi semi df if it satisfies
(01),(02) and (05), and quasi semi pseudo df if it satisfies only (01) and
(02), i.e., the very minimum indeed, which we will freely call a distance function.
The space <X, d) is qualified similarly, e.g., it is a pseudometric space, when d
is a pseudo df on X. In Table 12.1. we get the following elementary kinds:
Table 12.1. Elementary kinds of distance functions

Proper
Pseudo
Semi
Quasi

Dl

D2

D3

D4

D5

+
+
+
+

+
+
+
+

+
+
+

+
+

+
+

In the context of a distance function d on Mp, the ternary relation of


structurelikeness relation Sd(X, y, z) on Mp x Mp x Mp, Y is at least as similar
to z as x, is defined by d(y, z) :s; d(x, z), also expressed by "y is at least as close
to z as x". It is easy to check that Sd satisfies (the combination of) the first two
of the minimal s-conditions: for all x and y sAx, y, x) iff x = y, as soon as Sd
satisfies (05), besides (01) and (02). Moreover, for all x and y we have
sAx, x, y) and sAx, y, y). This implies two things. First, for all pairs <x, z) holds
r(x, z), i.e., all pairs are (trivially) comparable in the sense that there is a y such
that sAx, y, Z).l Secondly, it implies that Sd (again trivially) satisfies the third
minimal s-condition, all kinds of left and right reflexivity are guaranteed by
Sd(X, y, z), e.g., Sd(X, x, y). In sum, (01), (02) and (05) guarantee that Sd satisfies
the minimal s-conditions.
Finally, if d and d' are distance functions such that d' is an order-preserving

QUANTITATIVE TRUTH LIKENESS AND TRUTH APPROXIMATION

301

transformation of d, in the sense that d(x, y) ~ d(u, v) iff d'(x, y) ~ d'(u, v), then
d' and d agree on all comparative judgments of the form "sAx, y, z) iff Sd'(X, y, z)".
12.1. QUANTITATIVE ACTUAL TRUTH LIKENESS AND TRUTH
APPROXIMATION

Defining and assessing quantitative actual truthlikeness and truth approximation requires definitions of distance functions and success measures which are
specific for the relevant kind of structures. Recall that we argued that there
could not be a general definition of more actual truthlikeness, for the general
notion of more structurelikeness sex, y, z) can not be defined in a general way,
but will have to depend on the type of structures considered. The same holds
for a quantitative notion of structurelikeness or distance between two structures,
and ipso facto for quantitative actual truthlikeness. There is one exception; the
trivial distance function is of course generally defined by d,(x, y) = 1(0) iff x =
y (x #- y), which is a proper df
For non-trivial distance functions, we briefly reconsider the different types
of structures dealt with in Chapter 7- 10 and will give some hints for definitions.
We start with propositional structures, based on a finite set EP of elementary
propositions. Recall that we represented a propositional constituent by the
subset of the set of elementary propositions containing precisely its un-negated
conjuncts. In this representation, (description) y was said to be at least as
similar to the actual truth t as x if t - y is a subset of t - x and y - t a subset
of x - t, hence y LJ t is a subset of x LJ t.
There is a plausible quantitative variant: the distance dps(x, t) between constituents x and t is the proper df It - x I + It - x I = It LJ x I This definition has
already been proposed by Tichy (1974). It is an example of the so-called
Hamming distance or Clifford measure (Niiniluoto 1987a). It is easy to check
that this definition satisfies the connection principle, i.e., "y is at least as similar
to t as x" in the comparative sense, implies that dps(y, t) ~ dps(x, t).
It is easy to extrapolate this distance to an appropriate quantitative measure
of success. Recall that the data p and n indicate the (mutually exclusive) sets
of elementary propositions of which it has been established at a certain time
that their truth-value is 'true' (positive) or 'false' (negative), respectively. Recall
also that y is at least as successful at a certain time as x if and only if n n y is
a subset of n n x and p - y is a subset of p - x. The quantitative variant for x
is the number of established mismatches of x: Ip - xl + In n xl, which never
exceeds p + n, and increases from 0 to dps(x, t) when the number of elementary
propositions for which the truth-value has been established increases.
Alternatively, we may define the success as the number of matches:
Ip n x I + In - x I, which also never exceeds Ip I + In I and increases from 0 to
IEPI - dps(x, t). Note that this definition satisfies the adapted connection principle. However, whereas the qualitative notion satisfies the success principle,
i.e., 'at least as similar' implies 'at least as successful', this is no longer the case

302

QUANTITATIVE TRUTH LIKENESS AND TRUTH APPROXIMATION

for the quantitative definition. But when we assume that the elementary propositions about which the data deal have been randomly selected, it is not difficult
to prove that 'at least as similar' implies that the expected size of the success
is at least as large. We call this the expected success principle.
For first order structures, it is not always clear how to define a distance
function, let alone a quantitative measure of success. However, in certain cases,
there are plausible definitions. For instance, in the case of a language with one
monadic predicate and a fixed domain a symmetric difference definition suggests
itself again. For definitions of distances between first-order structures, based
on their first order theories, see (Niiniluoto 1987a, Section 10.3.).
On the other hand, in the case of real-valued structures there are usually
one or more well-known and plausible distance (from the truth) functions, with
related quantitative definitions of the success of a structure description. We
mention just one example. Let the structures (D,f) concern a real-valued
functionf on a fixed domain D, and letft indicate the true function. Then one
plausible definition for dD,f), (D,ft) is the square root of the sum of the
squares of the differences. A similar definition for the success takes the sum
restricted to the subdomain of individuals for which the true value has been
established.
Thus far, we have only mentioned 'absolute' distance and success functions.
Usually it is easy to see how the functions can be 'normalized', i.e., by what
fixed number the values have to be divided to obtain 1 as the maximum
distance. In numeric cases, this number is the total number of possibilities; in
other cases other 'normalization factors' are plausible.

12.2. QUANTITATIVE NOMIC TRUTH LIKENESS

Introduction

The case of quantitative nomic truthlikeness and truth approximation is much


more difficult. In this section we will deal with basic and refined definitions of
truthlikeness. In the third section we will deal with two essentially different
ways of a quantitative approach to truth approximation. We will restrict the
attention to non-stratified theories, and assume that Mp is finite, except when
otherwise stated.
12.2.1. Basic quantitative truthlikeness

The distance definition corresponding to the basic comparative definition is in


terms of the number of mistakes. That is, we define the basic distance of X from
T as the weighted sum of a (T-)internal and a (T-)external component:
Db(X, T) = D~(X, T)

+ D;;(X, T) =

ylT - XI

+ y'IX -

TI

where y and y' in [0, 1] are weights for the two different kinds of mistakes. Of

QUANTITATIVE TRUTHLIKENESS AND TRUTH APPROXIMATION

303

course, y' may be equal to (1 - y), or, alternatively, y' = y = 1 may hold. In the
last case, the definition is technically analogous to the definition of the quantitative distance between propositional constituents given above, and hence also
an example of a Hamming or Clifford distance.
It is easy to check that Db is a proper df and that it satisfies the connection
principle: MTL(X, Y, T) guarantees Db(Y, T) =:; Db(X, T), which also holds for
the corresponding internal and external clauses separately.
Since Db leaves room for different weights of the two types of mistakes 2, it
is plausible to be more careful in the definition of 'more truthlike'. Let us say
that Y is at least as close to T as X when Y is internally as well as externally
at least as close to T as X.
It is important to note that the condition that Y is internally /externally at
least as close to T as X guarantees that Y is not stronger/weaker than
XnT/XUT, in the (combined) sense that IXnTI=:;IYI=:;IXUTI. We will
say that Db satisfies the boundary principle. The boundary condition, i.e.,
IX n TI =:; IYI =:; IX UTI. is a direct quantitative generalization of comparative
truthlikeness: X n T ; Y ; XU T. When X and T do not overlap, as in many
cases of concretization, then the boundary condition reduces to
o=:; IYI =:; IXI + ITI- The latter condition will, in general, be called the weak
boundary condition, giving rise to the weak boundary principle.
Note that the result of replacing in Db(X, T) y and y' by y/(y + y') and
y'/(y + y'), respectively, leads to an order preserving distance function (see the
Introduction). Hence, in this and in all cases below where two separate weights
appear, the function can be transformed into order preserving functions by
replacing y' by (1 - y).
There is also a plausible 'averaged' version of Db:
D:(X, T) = Dt(X, T) + D:e(X, T) = ylT - XI/ITI

D:

+ y'IX - TI/IXI

i satisfies the (internal versions of the) connection and


It is trivial that
boundary principle, since the dividing factor ITI remains constant in the comparison. D:e only satisfies the connection principle with the full assumption
that Y is (externally as well as internally) comparatively at least as close to T
as X. Moreover, it does not even satisfy the boundary principle under a
similarly strong assumption. In sum, the averaged version is not very attractive.
The assumption that Mp is finite covers the case of propositional structures,
provided EP is finite as well. It covers, in addition, the case of a finite partition
or a finite number of mutually exclusive and together exhaustive kinds. The
corresponding structures are then of the form <{a}, Kl, ... , Kn), such that
precisely one of the Ki's is {a}, whereas the others are empty. We will come
back to this example when dealing with refined distances. Since these two types
of examples essentially exhaust all possibilities for finite Mp, the latter assumption is rather restrictive.
For the infinite case, it is plausible to assume a so-called measure function
m on (the measurable members of) P(Mp), i.e., a real-valued function, of which

304

QUANTITATIVE TRUTH LIKENESS AND TRUTH APPROXIMATION

the normalization, i.e., division by m(Mp), has formally the properties of a


probability measure. For measurable X, Y, and T we then define:
Dbm(X, T) =defym(T - X)

+ y'm(X -

T)

Since there may be non-empty sets with measure 0, D bm may well be a pseudo
df It is easy to check that Dbm satisfies the connection principle and the relevant
version of the boundary principle, i.e., with m(X n T) :::;; m( Y) :::;; m(X U T) as
boundary condition.
There is again a plausible way of 'averaging' Dbm(X, T), viz.,
Dtm(X, T)

= ym(T -

X}/m(T)

+ im(X -

T)/m(X)

with similar less attractive properties as in the case of Dt .


Note that normalization, by dividing by m(Mp), does not make essential
differences in both cases. Note, moreover, that we regain the results for the
finite case if we then assume that m corresponds to the number of elements.
In sum, Db is a plausible explication of basic quantitative truthlikeness for
finite Mp, and D bm for infinite Mp.
12.2.2. Refined quantitative truthlikeness

The challenge of a refined quantitative explication of truthlikeness is a distance


function between subsets of Mp, based on a (non-trivial) proper distance
function between the members of Mp, called the lower level df
When D is based on (a special type of) d we will write Dd , when relevant.
As a rule, we will distinguish between a weighted (T-)internal and (T-)external
distance, with the parameters y, y' in [0, 1] as weights, i.e.,
Dd(X, T)

= yD~(X, T) + y'D:;(X, T)

The internal distance will always have to do with T, but the external distance
mayor may not be directly related to Mp - T Again, we only say that Y is
quantitatively at least as close to T as X when Y is internally as well as
externally at least as close to T as X.
What kind of df we want is one thing; that there will be other desiderata is
another. The following additional desiderata seem plausible, at least at first
sight:
connection principle:

MTLsd(X, Y, T) => Dd(Y, T):::;; Dd(X, T)

boundary principle:

DAY, T):::;; DAX, T) => IX

reduction principle:

Dd,(Y, T):::;; Dd,(X, T) => Db(Y, T):::;; Db(X, T)

n TI:::;; IYI:::;; IX UTI

We have already seen that Db satisfies the first two principles, whereas the third
is, of course, not relevant for Do. The connection principle states that 'comparatively more truthlike', MTLsAX, Y, T), guarantees 'quantitatively closer to the
truth', DAY, T):::;; DAX, T). (Recall that Sd is s based on d.) The boundary
principle maintains that 'quantitatively closer to the truth' is only possible for

QUANTITATIVE TRUTHLIKENESS AND TRUTH APPROXIMATION

305

not too strong and not too weak theories, where the boundaries are suggested
by the quantitative boundaries of basic comparative truthlikeness. Finally, the
reduction principle states that when the lower level distance function is the
trivial one, the 'distance from the truth' becomes the basic one.
The desiderata may turn out to be realizable or not. Note that the principles
may be decomposed into an internal and an external principle, when the
definition itself can be split up in that way. When relevant and not mentioned
otherwise, they will be supposed to be imposed and satisfied in this split way.
Note also that the boundary principle is a principle which can simply be added
to a provisional definition when not yet satisfied by that definition.
Let us now review a couple of prima facie plausible refined definitions. Since
our experience with 'average' versions is not very impressive, we neglect them.
Let us first look at Niiniluoto's main proposal for a truth likeness function
and see whether it is useful for our purposes. Niiniluoto takes the truth to be
a complete theory, hence, in our representation, he wants to define the distance
between a subset X of Mp and 'the true structure' t, on the basis of an
underlying df For finite Mp, his favorite proposal is the min-sum-function
(Niiniluoto 1987a, p. 216, (44)):
DN(X, t)

= ydmin(t, X) + y'dsum(t, X)

where d.um(t, X) is defined as LXE xd(t, X)/ L XE Mpd(t, x). Moreover, Niiniluoto
assumes that d is balanced in the sense that LXE Mpd(t, x) = (1 /2)IMpl, hence
dsum(t, X) amounts to 2LxEXd(t, x)/ IMpl. Hence, the second term, although
simply called 'sum', is a kind of relativized sum of distances. Niiniluoto defines,
in general, the verisimilitude of X as 1 - D(X, t). He makes clear that DN is, in
many respects, very satisfactory for his particular conception of the verisimilitude problem. However, for our purposes, DN is not (yet) useful, for we consider
the nomic truth as a set, viz., T. Recall, in general, that Niiniluoto and others
conceive truth approximation as a matter of approaching 'elements by sets',
whereas we distinguish actual truth approximation as a matter of approaching
elements by elements, and nomic truth approximation as a matter of
approaching sets by sets.
Niiniluoto (1987a, p. 381 /2) rightly suggested that, under certain conditions,
our problem can be rephrased in terms of nomic constituents (see Zwart 1998,
Section 2.6.) for a qualitative elaboration of this idea). When Mp is finite, "T =
X" is, informally speaking, the finite conjunction of IMp I elementary claims:
for every x in X it says that it is nomically possible and for every x not in X
that it is nomically impossible. Its qualitative distance from the truth can be
expressed by the set of elementary claims that are false, which is, of course, a
symmetric difference, and the quantitative distance by the size of that set.
Hence, the resulting quantitative distance is equal to the most simple basic
distance function Db(X, T) = IT - X I + IX - TI (with y' = y = 1), since there is
a 1-1 function between T L'l X and the false elementary claims. Hence, by
rephrasing our problem in terms of nomic constituents, we reach a new level

306

QUANTITATIVE TRUTH LIKENESS AND TRUTH APPROXIMATION

of something like propositional constituents, such that the nomic truth becomes
complete, and hence a kind of actual truth. Although Niiniluoto's min-sum
proposal now leads to a proposal for the distance of a set of nomic constituents
to the truth, he has to presuppose for such a proposal a distance function
between constituents. As we have seen, the most plausible first proposal essentially amounts to our basic distance function. Hence, Niiniluoto's approach
does not yet give a refinement of this basic function, the ultimate topic of
this section.
The most plausible suggestion for our purposes is the following
Dl(X, T) = ~'E r DN(X, t) = Y~'E rdmin(t, X) + y'~tE rdsum{t, X)

However, as a potential distance function, it is rather strange. To begin with,


the second term is not a proper T-external term. Moreover, due to this second
term, it does not satisfy Dl (X, X) = 0, i.e., one of the two minimum conditions
for a df, viz., (D2). Finally, due to the first term, it is not symmetric. Hence, it
is not very attractive.
Several other constructions are possible, but no interesting results follow.
E.g., Festa (1986) has defined a plausible distance function for an interval of
real numbers to a particular real number, essentially governed by intuitions
similar to those of Niiniluoto, amongst which is conceiving the truth as a point.
Again, we do not know of a plausible way to generalize this definition to a
distance function between two intervals, notwithstanding some plausible conditions for such a definition mentioned by Festa (1986, p.305). Similarly,
Kieseppa (1996) has extended Niiniluoto's approach to 'multidimensional,
quantitative cognitive problems', with the same general features.
So let us start all over again. The most simple, non-decomposable, definition
is a function which is sometimes used as distance measure between sets,
D2(X, T)

= min {d(x, z)/x E X

& Z E T}

It is easy to check that it is, at most, a pseudo semi distance function, for (D4)
and (DS) will normally not be satisfied. Moreover, it is not satisfactory in the
light of the desiderata, since none of them is fulfilled.
A decomposable function arises by taking the weighted sum of the minimal
and the maximal mistake:
D3(X, T)

= Y min {d(x, z)/x E X &

ZE

T}

+ y' max {d(x, z)/x E X & Z E T}


However, it is not even a distance function in the minimal sense, for (D2) is
not satisfied in general (D3(X, X) is non-zero, as a rule).
Another possibility is to restrict the attention to genuine mistakes:
D4(X, T)

= y min {d(x, z)/x E X & Z E T - X}


+ y' max {d(x, z)/x E X - T & Z E T}

QUANTITATIVE TRUTHLIKENESS AND TRUTH APPROXIMATION

307

One may also take, in both cases, the maximal mistakes, or the minimal ones:
Ds(X, T) = y max {d(x, z)/x E X & Z E T - X}

+ y' max {d(x, z)/x EX D6(X, T) = y min {d(x, z)/x E X &

+ y' min {d(x, z)/x E X

T & Z E T}
E T - X}

- T & Z E T}

But none of them give really attractive results.


Let us now try a summation version, in particular of D6(X, T):
D7(X, T) = y~z E Tdmin(z, X)

+ y'~x

X dmin(x, T)

Here, the minimum distance between e.g., Z and X, dmin(z, X), is defined as the
minimum of d(z, x) for all x in X. D7 is discussed by Niiniluoto (1987a, p. 246)
in the general form of a distance function between statements. Intuitively, D7
is very attractive for our purposes. It is a weighted sum of the sums of the
minimal internal and external mistakes, respectively. D7 is even a proper
distance function, which is easy to check, except for the triangle inequality.
Moreover, it satisfies the reduction principle, which is a direct consequence of
the fact that e.g., dmin(z, X) for trivial d amounts to 1 when Z : X and 0 when
Z E X, and hence ~z E Td min (z, X) amounts to 1
T - Z I.
However, the other principles are not valid in general. The boundary principle
is not valid in general, but, as suggested earlier, this might be imposed as an
extra condition. Hence, more important is the fact that only the internal side
of the connection principle holds unconditionally: when (Rii) applies to
<X, Y, T) then D~(Y, T) ~ D~(X, T). On the other hand, if (Ri) applies to
<X, Y, T) then D;(Y, T) ~ D;(X, T) iff
~YEY-XuTdmin(Y, T) ~ ~XE x- YuTdmin(x, T).

Moreover, if we strengthen (Ri) to:


(Ri-Q)

there is a 1-1 function from Y - X U T to X - Y U T such that


for all y in Y - X U T there is Z in T - X such that s(f(y), y, z)

DHY, T) ~ D;(X, T) is guaranteed in general.

This strong version of (Ri) is trivially satisfied in the case of basic truthlikeness. However, it need not be satisfied in other cases, e.g., in the context of
concretization. For consider the gas model example of Subsection 10.4.2. It is
easy to check that, for instance, the concretization triple of (sets of) gas models
<JGM, GMa, WGM) does not satisfy (Ri-Q), simply due to the fact that the
set GMa is (much) larger than IGM. As a consequence, it may well be the case
that all plausible lower level distance functions in this case, if any, are such
that D;(GMa, WGM) is larger than D;(IGM, WGM), notwithstanding the
fact that GMa is closer to WGM than IGM, not only according to our

308

QUANTITATIVE TRUTH LIKENESS AND TRUTH APPROXIMATION

qualitative definition of refined truthlikeness, but also according to a generally


accepted informal judgement.
There is, however, a weaker sufficient condition for special types of cases
where the weak boundary condition (0 ~ I YI ~ IXI + IT!) applies, as in the case
of double concretization, e.g the Van der Waals case above, i.e.,
O<IGMal<IIGMI+IWGMI, and the capital structure theory, treated
in Section 11.2., provided we counterfactually assume that these sets are
finite. The target inequality D~ (Y, T) ~ D~(X, T) can be rephrased by
IYldmean(Y, T) ~ IXldmean(X, T), where e.g., dmean(X, T) indicates the mean distance of the members of X to T, i.e., r. xE xdmin(X, T)/ IXI . Now it is easy to
check that, assuming that the weak boundary condition holds,
dmean{Y, T) ~ IXI /(IXI + ITI) x dmean{X, T) is a sufficient condition for the
target inequality. In particular, in the case of double concretization, assuming
a distance function, this does not seem to be an unreasonable condition. Given
the fact that the 'internal inequality' is, in the case of double concretization
rather unproblematic, D7 is satisfactory for double concretization. 3
What about specialization-followed-by-concretization, the formal structure
of the 'old quantum theory' dealt with in Section 11.1? In this case, not only
the boundary condition, but even the external inequality itself is trivial, assuming again finite sets4. Hence, also for this kind of theory succession D7 is
satisfactory.
In sum, D7 is, if not completely satisfactory, an acceptable proposal for finite
Mp.5 Under appropriate conditions, it will be possible to extend D7 (and the
(weak) boundary condition) to a denumerably infinite as well as to a nondenumerable set Mp, but we will not spell this out.
12.3 . QUANTITATIVE NOMIC TRUTH APPROXIMATION

Introduction

When dealing with comparative truth likeness, it was plausible to evaluate the
truth approximation hypothesis, Y closer to the truth than X , in terms of a
comparative notion of more successfulness. In the present, quantitative context,
we have essentially two different ways of evaluation. In the line of the comparative case, we may define quantitative notions of success, corresponding to the
relevant quantitative notion of truthlikeness. Or, for that matter, a failure
function 6 corresponding to the relevant distance from the truth. A further
desideratum then is, of course, that 'closer to the truth' guarantees a lower
'failure value' or at least a lower 'expectation value' of that value, assuming
that the data have arisen from random experimentation. However, the suggested
notion of quantitative failure is essentially non-probabilistic. It is the subject
of the first subsection.
We also have now the possibility of taking into account the posterior probability of the theories which have not yet been falsified . With the help of them,

QUANTITATIVE TRUTHLIKENESS AND TRUTH APPROXIMATION

309

and a distance from the truth definition, we can calculate the 'estimated distance
from the truth' for each theory. Hence, this is a straightforwardly probabilistic
evaluation of the merits of a theory in approaching the truth.
For the case of a complete truth, Niiniluoto (1987a) has propagated and
developed this idea convincingly. That is, apart from the fact that the assumption that theories are aiming at the complete truth about the actual world, it
is a very plausible idea, which we will adapt for approaching the nomic truth.
In both cases we will restrict the attention to finite Mp.
12.3.1. Non-probabilistic (quantitative) truth approximation

For non-probabilistic truth approximation, we have to define a notion of


success in terms of the data Rj S which is suitable for the relevant quantitative
distance from the truth.
We start with the basic notion. For the basic failure function we take:
Fb(X, Rj S) = Fb(X, R) + Fb(X, S) = ylR - XI + y'IX - SI

which does not exceed y IR 1+ y' 151, and increases from 0 to Db(X, T) .
As in the case of success related to the actual truth, only the expectation
version of the success principle holds: if Db(Y, T) ~ Db(X, T) then the
(internal/external) failure value of Y may be expected to be not higher than
the (internaljexternal) failure value of X as far as R can be conceived as the
result of random selection of IR I members of T and S as the result of random
selection of IMp - SI members of Mp - T. It is easy to check that Fb(X, Rj S)
satisfies the relevant version of the connection principle, viz., if Y is qualitatively
more successful than X , MS(X, Y, Rj S), then Y has a lower, or at least not a
higher, failure value, Fb(Y, RjS) ~ Fb(X, RjS) . Note finally that Fb(X, Rj S) satisfies the relevant version of the boundary principle, viz., Fd ( Y, Rj S) ~ FAX, Rj S)
implies IX n RI ~ IYI ~ IX U SI.
Let us now turn to refined notions of success Fd , based on an underlying
proper distance function d. As in the case of distance from the truth, we will
assume, when relevant, weighted versions, of which the components can now
be indicated by the relevant data:
FAX, Rj S) = yFAX, R) + y'FAX, S)

As desiderata we have, of course, the relevant versions of the desiderata for


the distance from the truth, viz.,
F-connection principle:
MSFsAX, Y, Rj S) = FAY, Rj S) ~ FAX, Rj S)
F-boundary principle:

310

QUANTITATIVE TRUTH LIKENESS AND TRUTH APPROXIMATION

F -reduction principle:
Fdt(Y, RjS) S; Fdr(X, RjS) => Fb(Y, RjS) S; Fb(X, RjS)

However, we now also impose the


expected success (ES-) principle:
Dd(Y, T)

S;

DAX, T) => EV(Fd(Y, RjS S; EV(FAX, RjS

This principle maintains that 'closer to the truth', though not guaranteeing a
lower failure value, leads to a lower expected value (EV) of the failure value,
where the expected value is based on an appropriate random way of obtaining
Rand S. Recall that the basic definition of failure satisfies the F -connection,
the F-boundary and the ES - principle. The F-reduction principle is, as in the
case of the basic distance from the truth, not relevant for basic success.
Recall that
D7 (X, T) = YL z ET dmin(z, X)

+ y'LXE Xdmin(X, T)

was, relatively speaking, the best 'distance from the truth' function. It suggests
the following failure function:

YLzERdmin(Z, X)+ y'LxEXdmin(X, S)

It is easily seen to satisfy the F -reduction principle and the F -boundary principle
can again simply be imposed. Moreover, the internal or R-side satisfies the
F -connection principle, which is easy to check. Moreover, it satisfies the
ES-principle.

Proof We have to prove that D~(X, T) ~ D~(Y, T), that is,


LZE Tdmin(z, X) ~ L z E Tdmin(z, Y) implies EV(F7 (X, R ~ EV(F7(Y, R.
Let T(r) indicate {RjR <:; T, IRI = r}. Then EV(F7(X, R =
LRET(r)F7(X,R)fIT(r)I=LRET(r)LZERdmin(Z,X)jIT(r)l. It expresses the
average distance from X of 'subsets R of T of size r', and hence 'the
expected failure of X with respect to random variable R', where R is
the random variable indicating the random selection of a subset of T
of size r. Now it is not difficult to check that LR E T(r)L ZE Rdmin(z, X)
must be proportional to L z E Td min (z, X), for every Z in T occurs in an
equal number of members of T(r), hence dmin(z, X) is summed up an
equal number of times for all Z in T. Qed.

However, the S-side does not unconditionally satisfy the F -connection and the
ES-principle. But under certain conditions it does. For the S-side of
the F-connection principle, essentially the same thing holds as for the external
side of the 'D-'connection principle. If (Ri)-S applies to (X, Y, S> then
F7(Y,S)S;F7(X,S) iff LYEY-Xvsdmin(y,S)S;LxEX-yvsdmin(X,S). Moreover, if
we strengthen (Ri)-S to:

QUANTITATIVE TRUTH LIKENESS AND TRUTH APPROXIMATION

(Ri-Q)-S

311

there is a 1- 1 function from Y - X U S to X - Y U S such


that for all y in Y - X U S there is z in S - X such that
s(f(y), y, z)

F7 (Y, S)::;; F7 (X, S) is guaranteed in general. This strong version of (Ri)-S is

trivially satisfied in the case of basic truthlikeness. But it need not be satisfied
in other cases, in particular in the case of concretization. For D7 we illustrated
this with the gas model example of Subsection 10.4.2. This illustration can now
formally be simply reproduced, whereas WGM is now conceived not as the
truth, but 'merely' as an established law. But again, there is another sufficient condition for 'successful' double concretization. If the adapted version
of the weak boundary condition applies, i.e., 0::;; IYI ::;; ISI then
dmean(Y, S)::;; IXI /(IXI + lSI) x dmean(X, S) is a sufficient condition for the target
external inequality.
Turning now to the S-side of the ES-prin