Introduction To The Philosophy of Science First Part
Introduction To The Philosophy of Science First Part
INTRODUCTION
TO THE PHILOSOPHY
OF SCIENCE
A Text by Members of the Department
of the History a n d Philosophy o f Science
of the University o f Pittsburgh
Clark Glymour
J. E. McGuire
Kenneth F. Schqffner
To the giants
On whose shoulders we stand
12 3 4 5 6 7 8 9
The paper used in this publication meets the minimum requirements of American National Standard
for Information SciencesPermanence of Paper for Printed Library Materials, ANSI Z39.48-1984.
CONTENTS
PREFACE
INTRODUCTION
2
THE CONFIRMATION OF SCIENTIFIC HYPOTHESES
42
3
REALISM AND THE NATURE OF THEORIES
104
4
SCIENTIFIC CHANGE: PERSPECTIVES AND PROPOSALS
132
179
iii
6
DETERMINISM IN THE PHYSICAL SCIENCES
232
7
PHILOSOPHY OF BIOLOGY
269
8
PHILOSOPHY OF MEDICINE
310
9
PHILOSOPHY OF PSYCHOLOGY
346
10
ANDROID EPISTEMOLOGY; COMPUTATION, ARTIFICIAL INTELLIGENCE,
AND THE PHILOSOPHY OF SCIENCE
364
11
PHILOSOPHY OF THE SOCIAL SCIENCES
404
BIBLIOGRAPHY
426
INDEX
447
iv
Contents
PREFACE
The philosophy of science has become such a far-flung and specialized enterprise that
no one person is in a position to write an authoritative survey of the field, even at an
introductory level. This is an especially unfortunate situation in view of the felt need
for an up-to-date textbook in this field. Our solution is drawn on the combined
expertise of members of the History and Philosophy of Science Department of the
University of Pittsburgh: John Earman, Clark Glymour, James G. Lennox, J. E.
McGuire, Peter Machamer, John D. Norton, Merrilee H. Salmon, Wesley C. Salmon,
and Kenneth F. Schaffner.* Although individual names are attached to chapters, this
volume is the result of a cooperative effort, and to the extent that it succeeds the credit
is to be equally divided, except for additional measures for Merrilee Salmon, who
conceived the project and prodded us into action, and for John Norton, who drew the
figures for Chapters 2, 5, 6, and 9.
The primary audience for our text consists of upper-level undergraduates and
beginning graduate students. While it is possible to teach philosophy of science to
first and second year students, our experience is that students will benefit more from
such a course if they have already mastered some college-level science and /or history
of science. We have attempted to reach a compromise between a presentation that is
accessible to a wide audience and one that makes contact with current research. This
is by no means an easy compromise to achieve, but texts that do not aspire to it are
not worth the candle.
The volume contains more material than can be comfortably covered in one
semester and certainly more than can be covered in a ten-week quarter. The instructor
is thus presented with a number of choices. Those interested mainly in scientific
methodology may want to devote the entire course to the general topics in the
V
philosophy of science that are explored in Part One, Chapters 1-4. Alternatively, the
first half of the course could be devoted to a survey of the material in Chapters 1-4
and the second to a sampling of the foundational problems in various physical,
biological, behavioral and social sciences that are discussed in Chapters 5-11. Or the
text could be used as the basis of a two-course sequence.
Each chapter contains a list of suggested readings and study questions. Some of
the questions are appropriate for test questions; others can be used as topics for
writing assignments.
We have tried to keep technical notation to a minimum, but sometimes its use
is unavoidable. Symbols are defined and their use is explained with examples in the
contexts in which they first occur. When symbols occur in more than one chapter, first
occurrences are noted in the index.
We want to express our deepest thanks to Philip Kitcher, Universiy of California, San Diego, who read the entire manuscript at an early stage and offered extensive
and helpful comments on every chapter. We are also grateful to James E. Roper,
Michigan State University, Arnold Wilson, University of Cincinnati, and Michael J.
Zenzen, Rensselaer Polytechnic Institute, for their critical reading of the manuscript
and their good advice. John Beatty, University of Minnesota, Harry Corwin, University of Pittsburgh, and David Kelly, Institute for Objectivist Studies, offered useful
suggestions for Chapter 7. Rob Pennock, University of Texas at Austin, offered
important suggestions regarding Chapter 1. Kevin Kelly, Carnegie-Mellon University, provided useful suggestions and help with the pictures in Chapters 3 and 10.
David Hillman gave valuable help in checking and assembling the bibliography, and
Judith Meiksin provided excellent editorial assistance. Madeline Larson compiled the
index. The Faculty Proficiency Enhancement Program, University of Pittsburgh,
provided the computer system on which the text, graphics, and index were assembled. Special thanks are due to the students in several Honors Philosophy of Science
courses at the University of Pittsburgh. They served as willing guinea pigs for our
efforts to produce a suitable text for undergraduates, and their lively participation in
class discussions helped immensely in refining our pedagogical strategies.
John Earman
vi
Preface
INTRODUCTION
such as the question just raised about the nature of scientific explanation. Some other
questions studied by philosophers of science are as follows:
What are the aims of science?
What is the role of observations and experiments in obtaining scientific knowledge?
How do scientists justify their claims? What is a scientific proof?
What is a scientific law?
Are there methods for making scientific discoveries?
How does scientific knowledge advance and grow?
How do the historical and cultural settings in which scientific work occurs affect
the content and quality of such work?
Does science employ or require a special language?
Science itself is made up of many subdisciplines: physics, astronomy, chemistry, biology, psychology, sociology, anthropology, and medicine, to name a few.
The presence of so many different fields within science raises interesting questions
about what it means to be a science and whether a single method is common to all
sciences. Philosophy of science thus addresses also the following sorts of questions:
Is it possible to give a general account of scientific methodology, or are there
different methods and forms of explanation for various branches of science?
How do physical, biological, and social sciences differ from one another?
Can some sciences be reduced to others?
Finally, philosophy of science is concerned with specific issues that arise in
connection with particular fields of science. For example, while experimentation
plays a major role in some sciences, in others, such as astronomy, it does not. Some
other discipline-specific questions are these:
Does the existence of free will pose a special problem for a science of human
behavior?
Is medicine more an art than a science?
Are statistical techniques useful in anthropology, where sample sizes are very
small?
All of the questions raised above are complex and difficult, so it should come
as no surprise that the opinions of philosophers of science (and scientists in their
philosophical moments) on these topics vary considerably. In the twentieth century,
two disparate approaches have been dominant. The earlier tradition, developed by
logical positivists (members of the Vienna Circle) and logical empiricists (a similar
group from Berlin), set rigorous standards for the conduct of philosophy of science,
as close to those of science itself as the subject matter would allow. These philosophers and scientists attempted to provide logical analyses of the nature of scientific
2
Introduction
concepts, the relation between evidence and theory, and the nature of scientific
explanation. In their desire to be precise, they made extensive use of the language and
techniques of symbolic logic. Despite many differences in points of view, the logical
positivists and logical empiricists generally were concerned with emphasizing such
distinctions as
the demarcation between scientific knowledge and other types of knowledge,
the difference between facts and values,
the difference between the language used to state observations and that used to
refer to theoretical entities, and
the difference between how theories are discovered and how they are justified.
Logical empiricists and logical positivists were also concerned with establishing clear
meanings for all the terms used in science. Some approached this problem by searching for a verifiability criterion of meaning while others, particularly scientists themselves, tried to formulate operational definitions of scientific terms. These efforts
were closely related to their concern with providing a solid foundation for scientific
theorizing by linking it firmly to an observational basis. Although they believed that
justification rather than discovery was the proper concern of science, they shared an
optimism about the ability of science to provide genuine knowledge of the features of
an independently existing world.
At the time of World War Two, many of these philosophers left Europe for
England and the United States where their works have significantly affected the
development of philosophy of science in English-speaking countries. Even at the
level of undergraduate education, their influence has been important. Carl G. Hempel,
who came to America from Berlin, for example, has literally defined the philosophy
of the natural sciences for generations of students who first learned about the subject
from his introductory text, Philosophy of Natural Science (1966). The power and
broad influence of the general approach outlined by Hempel in this work justifies
calling it "the standard view" of philosophy of science.
During the past twenty-five years, however, many criticisms have been raised
against perceived faults of the standard view. (Indeed, Hempel himself has criticized
some of its features.) A major objection is that the standard view fails to take account
of the bearing of history of science on the philosophy of science. Critics of the
standard view cite Thomas Kuhn's Structure of Scientific Revolutions (1962, 1970),
which argues that most scientific textbooks ignore history and distort the real nature
of progress in science by presenting it as a series of accumulations of new discoveries
that straightforwardly build on and add to knowledge already attained. Kuhn draws
attention to the revolutionary character of scienceits replacement of outworn theories by newer ones that are so different from the old that the two do not share the
same problems or even a common language. He also draws attention to the "irrational' ' aspects of changes in science, that is to say, those features of scientific change
that cannot be accounted for entirely in terms of scientists' allegiance to "facts" and
logic. Kuhn argues that only a refusal to take seriously the history of science could
account for the gross distortion presented in scientific textbooks.
Introduction
Appealing to Kuhn's account of science, critics of the standard view of philosophy of science say that it embodies and promotes an ahistorical view of scientific
activity by emphasizing the logical characteristics of science while ignoring the
cultural context of scientific activity, which strongly influences the style of the enterprise and the content of its results. Furthermore, critics say, failure to take account
of the rhetorical features of scientific discourse can only lead to a distorted notion of
how science really works. The values of society and of individual practitioners of
science, they say, influence not only the choice of problems and the amount of effort
devoted to their solution, but also the interpretation of the results. They maintain that
so-called facts can only be grasped through theories, which are the creations of
members of a specific culture, and are never completely free of the values and
aspirations of that culture.
Both the standard view and that of its critics have merits and shortcomings.
Both views are likewise too complex to state succinctly without distortion and oversimplification; the above brief synopsis is intended only to introduce the reader to the
subject. The ensuing chapters will survey many aspects of the dispute and will
examine the reasons offered in support of the various positions.
The approach to the philosophy of science exemplified in this work does not fall
neatly into either of the two main categories briefly outlined. The authors of this text
are all members of a Department of History and Philosophy of Science. The marriage
between history and philosophy in the Department is not merely one of convenience
between philosophers and historians each of whom happens to be concerned with
science. Instead, the Department was founded because the members believe that the
study of the philosophy of science must be informed by an understanding of the
historical and social context of science, as well as by a grasp of the workings of
science itself. At the same time, the general approach of this book disavows the
extreme forms of relativism and skepticism that characterize some of the more strident critics of the standard view.
Part One of this book takes up topics requisite for any adequate introduction to
the philosophy of science: Explanation; Induction and Confirmation; Realism and the
Nature of Scientific Theories; and Scientific Change: Perspectives and Proposals.
These four chapters outline and discuss fundamental issues in philosophy of science
and form the foundation for discussions in the remaining chapters of the book. In Part
One, the reader is introduced to the pertinent history of the topics discussed as well
as to the vocabulary, techniques, and most important issues in contemporary philosophy of science. The intention of the authors in each case is to presume no prior
knowledge of philosophy of science, but to lead the reader to an appreciation of some
of the knottiest problems that concern contemporary philosophers of science. In the
first chapter, "Scientific Explanation," Wesley C. Salmon discusses the elements
involved in the special kind of understanding of our world and what takes place within
it that is provided by the various sciences. In the second chapter, "The Confirmation
of Scientific Hypotheses," John Earman and Wesley C. Salmon deal with questions
concerning the relationship between empirical evidence and scientific hypotheses,
laws, and theories. In the course of the discussion they consider the nature of inductive reasoning and the meanings of the concept of probability. Chapter 3, by Clark
Glymour, considers the major traditional arguments against literal belief in the claims
4
Introduction
Introduction
One
SCIENTIFIC EXPLANATION
Wesley C. Salmon
The fruits of science are many and various. When science is first mentioned, many
people think immediately of high technology. Such items as computers, nuclear energy , genetic engineering, and high-temperature superconductors are likely to be included. These are the fruits of applied science. Evaluating the benefits, hazards, and
costs of such technological developments often leads to lively and impassioned debate.
In this chapter, however, we are going to focus on a different aspect of science,
namely, the intellectual understanding it gives us of the world we live in. This is the
fruit of pure science, and it is one we highly prize. All of us frequently ask the
question "Why?" in order to achieve some degree of understanding with regard to
various phenomena. This seems to be an expression of a natural human curiosity.
Why, during a total lunar eclipse, does the moon take on a coppery color instead of
just becoming dark when the earth passes between it and the sun? Because the earth's
atmosphere acts like a prism, diffracting the sunlight passing through it in such a way
that the light in the red region of the spectrum falls upon the lunar surface. This is a
rough sketch of a scientific explanation of that phenomenon, and it imparts at least
some degree of scientific understanding.
Our task in this chapter is to try to say with some precision just what scientific
explanation consists in. Before we embark on that enterprise, however, some preliminary points of clarification are in order.
1.1 EXPLANATION VS. CONFIRMATION
The first step in clarifying the notion of scientific explanation is to draw a sharp
distinction between explaining why a particular phenomenon occurs and giving rea7
sons for believing that it occurs. My reason for believing that the moon turns coppery
during total eclipse is that I have observed it with my own eyes. I can also appeal to
the testimony of other observers. That is how the proposition that the moon turns
coppery during a total eclipse is confirmed,1 and it is entirely different from explaining why it happens. Consider another example. According to contemporary cosmology all of the distant galaxies are receding from us at high velocities. The evidence
for this is the fact that the light from them is shifted toward the red end of the
spectrum; such evidence confirms the statement that the other galaxies are moving
away from our galaxy (the Milky Way). The fact that there is such a red shift does
not explain why the galaxies are moving in that way; instead, the fact that they are
receding explainsin terms of the Doppler effect2why the light is shifted toward
the red end of the spectrum. The explanation of the recession lies in the "big bang"
with which our universe began several billion years ago; this is what makes all of the
galaxies recede from one another and, consequently, makes all of the others move
away from us.
1.2 OTHER KINDS OF EXPLANATION
Another preliminary step in clarifying the notion of scientific explanation is to recognize that there are many different kinds of explanation in addition to those we
classify as scientific. For example, we often encounter explanations of how to do
somethinghow to use a new kitchen gadget, or how to find a certain address in a
strange city. There are, in addition, explanations of whatwhat an unfamiliar word
means, or what is wrong with an automobile. While many, if not all, scientific
explanations can be requested by means of why-questions, requests for explanations
of these other sorts would not normally be phrased in why-questions; instead, howto-questions and wto-questions would be natural.
Still other types of explanation exist. Someone might ask for an explanation of
the meaning of a painting or a poem; such a request calls for an artistic interpretation.
Or, someone might ask for an explanation of a mathematical proof; an appropriate
response would be to fill in additional steps to show how one gets from one step to
another in the original demonstration. Neither of these qualifies as scientific explanation. Also excluded from our domain of scientific explanation are explanations of
formal facts of pure mathematics, such as the infinitude of the set of prime numbers.
We are concerned only with explanation in the empirical sciences.
As we understand the concept of scientific explanation, such an explanation is
an attempt to render understandable or intelligible some particular event (such as the
1986 accident at the Chernobyl nuclear facility) or some general fact (such as the
copper color of the moon during total eclipse) by appealing to other particular and/or
general facts drawn from one or more branches of empirical science. This formulation
1
Scientific Explanation
is not meant as a definition, because such terms as "understandable" and "intelligible' ' are as much in need of clarification as is the term ' 'explanation.'' But it should
serve as a rough indication of what we are driving at.
In pointing out the distinction between scientific explanations and explanations
of other types we do not mean to disparage the others. The aim is only to emphasize
the fact that the word "explanation" is extremely broadit applies to a great many
different things. We simply want to be clear on the type of explanation with which our
discussion is concerned.
"How did large mammals get to New Zealand?" The answer is that they came in
boatsthe first were humans, and humans brought other large mammals. Or, consider the question, "How is genetic information transmitted from parents to offspring?" The answer to this question involves the structure of the DNA molecule and
the genetic code.
In this chapter we will not try to argue one way or other on the issue of whether
all scientific explanations can appropriately be requested by means of why-questions.
We will leave open the possibility that some explanations cannot suitably be requested by why-questions.
1.4 SOME MATTERS OF TERMINOLOGY
As we will see, one influential philosophical account of explanation regards all bona
fide scientific explanations as arguments. An argument is simply a set of statements,
10
Scientific Explanation
one of which is singled out as the conclusion of the argument. The remaining members of the set are premises. There may be one or more premises; no fixed number of
premises is required. 3 The premises provide support for the conclusion.
All logically correct arguments fall into two types, deductive and inductive, and
these types differ fundamentally from one another. For purposes of this chapter (and
later chapters as well) we need a reasonably precise characterization of them. Four
characteristics are important for our discussion.
DEDUCTION
1. In a valid deductive argument, all of
the content of the conclusion is present,
at least implicitly, in the premises. Deduction is nonampliative.
2. If the premises are true, the conclusion
must be true. Valid deduction is necessarily truth-preserving.
3. If new premises are added to a valid
deductive argument (and none of the
original premises is changed or deleted)
the argument remains valid. Deduction is
erosion-proof.
4. Deductive validity is an all-or-nothing
matter; validity does not come in degrees,
An argument is totally valid or it is invalid.
INDUCTION
1. Induction is ampliative. The conclusion of an inductive argument has content
that goes beyond the content of its premises.
2. A correct inductive argument may
have true premises and a false conclusion. Induction is not necessarily truthpreserving.
3. New premises may completely undermine a strong inductive argument. Induction is not erosion-proof.
Argument (1) is obviously a valid deduction. When we have said that all humans are
mortal, we have already said that Socrates is mortal, given that Socrates is human.
Thus, it is nonampliative. Because it is nonampliative, it is necessarily truthpreserving. Since nothing is said by the conclusion that is not already stated by the
premises, what the conclusion says must be true if what the premises assert is true.
Moreover, the argument remains nonampliative, and hence, necessarily truthpreserving, if new premisesfor example, "Xantippe is human"are added. You
cannot make a valid deduction invalid just by adding premises. Finally, the premises
3
Because of certain logical technicalities, there are valid deductive arguments that have no premises at
all, but arguments of this sort will not be involved in our discussion.
Scientific Explanation
11
support the conclusion totally, not just to some degree; to accept the premises and
reject the conclusion would be outright self-contradiction.
(2)
This argument is obviously ampliative; the premise refers only to ravens that have
been observed, while the conclusion makes a statement about all ravens, observed or
unobserved. It is not necessarily truth-preserving. Quite possibly there is, was, or will
beat some place or timea white raven, or one of a different color. It is not
erosion-proof; the observation of one non-black raven would undermine it completely. And its strength is a matter of degree. If only a few ravens in one limited
environment had been observed, the premise would not support the conclusion very
strongly; if vast numbers of ravens have been observed under a wide variety of
circumstances, the support would be much stronger. But in neither case would the
conclusion be necessitated by the premise.
Deductive validity and inductive correctness do not hinge on the truth of the
premises or the conclusion of the argument. A valid deduction may have true premises and a true conclusion, one or more false premises and a false conclusion, and
one or more false premises and a true conclusion. 4 When we say that valid deduction
is necessarily truth-preserving, we mean that the conclusion would have to be true if
the premises were true. Thus there cannot be a valid deduction with true premises and
a false conclusion. Where correct inductive arguments are concerned, since they are
not necessarily truth-preserving, any combination of truth values of premises and
conclusion is possible. What we would like to say is that, if the premises are true (and
embody all relevant knowledge), the conclusion is probable. As we will see in
Chapter 2, however, many profound difficulties arise in attempting to support this
claim about inductive arguments.
We have chosen very simpleindeed, apparently trivialexamples in order to
illustrate the basic concepts. In actual science, of course, the arguments are much
more complex. Most of the deductive arguments found in serious scientific contexts
are mathematical derivations, and these can be extremely complicated. Nevertheless,
the basic fact remains that all of them fulfill the four characteristics listed above.
Although deep and interesting problems arise in the philosophy of mathematics, they
are not our primary concern in this book. Our attention is focused on the empirical
sciences, which, as we argue in Chapter 2, necessarily involve induction. In that
chapter we encounter much more complex and interesting inductive arguments.
The familiar slogan, "Garbage in, garbage out," does not accurately characterize deductive argu-
ments.
12
Scientific Explanation
scientists since then. Nevertheless, many other philosophers and scientists have maintained that science must "stick to the facts," and consequently can answer only
questions about what but not about why. To understand "the why of things," they
felt, it is necessary to appeal to theology or metaphysics. Science can describe natural
phenomena and predict future occurrences, but it cannot furnish explanations. This
attitude was particularly prevalent in the early decades of the twentieth century. Since
it is based upon certain misconceptions regarding scientific explanation, we need to
say a bit about it.
It is natural enough, when attempting to find out why a person did something,
to seek a conscious (or perhaps unconscious) motive. For example, to the question,
"Why did you buy that book?" a satisfactory answer might run, "Because I wanted
to read an amusing novel, and I have read several other novels by the same author,
all of which I found amusing.'' This type of explanation is satisfying because we can
put ourselves in the place of the subject and understand how such motivation works.
The concept of understanding is critical in this context, for it signifies empathy. If we
yearn for that kind of empathetic understanding of nonhuman phenomena, we have to
look elsewhere for motivation or purpose. One immediate suggestion is to make the
source of purpose supernatural. Thus, prior to Darwin, the variety of species of living
things was explained by special creationthat is, God's will. Another manifestation
of the same viewpointheld by some, but not all, vitalistswas the notion that
behind all living phenomena there is a vital force or entelechy directing what goes on.
These entitiesentelechies and vital forcesare not open to empirical investigation.
The insistence that all aspects of nature be explained in human terms is known
as anthropomorphism. The suppositioncommon before the rise of modern
sciencethat the universe is a cozy little place, created for our benefit, with humans
at its center, is an anthropomorphic conception. The doctrines of special creation and
some forms of vitalism are anthropomorphic. So-called "creation science" is anthropomorphic. Teleoiogical explanation of nonhuman phenomena in terms of
human-like purposes is anthropomorphic. 5
Many philosophers and scientists rejected the appeal to anthropomorphic and
teleoiogical explanations as an appeal to hypotheses that could not, even in principle,
be investigated by empirical science. If this is what is needed for explanation, they
said, we want no part of it. Science is simply not concerned with explaining natural
phenomena; anyone who wants explanations will have to look for them outside of
science. Such scientists and philosophers were eager to make clear that scientific
knowledge does not rest on nonempirical metaphysical principles.
Not all philosophers were willing to forgo the claim that science provides
explanations of natural phenomena. Karl R. Popper (1935), Carl G. Hempel (1948),
R. B. Braithwaite (1953), and Ernest Nagel (1961) published important works in
which they maintained that there are such things as legitimate scientific explanations,
and that such explanations can be provided without going beyond the bounds of
empirical science. They attempted to provide precise characterizations of scientific
5
As James Lennox points out in Chapter 7, teleoiogical explanations are anthropomorphic only if they
appeal to human-like purposes. In evolutionary biologyand other scientific domains as wellthere are
teleoiogical explanations that are not anthropomorphic.
Scientific Explanation
13
explanation, and they were, to a very large degree, in agreement with respect to the
core of the account. The line of thought they pursued grew into a theory that enjoyed
a great deal of acceptance among philosophers of science. We will discuss it at length
in later sections of this chapter.
1.7 DOES EXPLANATION INVOLVE REDUCTION TO THE
FAMILIAR?
It has sometimes been asserted that explanation consists in reducing the mysterious or
unfamiliar to that which is familiar. Before Newton, for example, comets were
regarded as mysterious and fearsome objects. Even among educated people, the
appearance of a comet signified impending disaster, for example, earthquakes, floods,
famines, or epidemic diseases. Newton showed that comets could be understood as
planet-like objects that travel around the sun in highly eccentric orbits. For that
reason, any given comet spends most of its time far from the sun and well beyond the
range of human observation. When one appeared it was a surprise. But when we
learned that they behave very much as the familiar planets do, their behavior was
explained, and they were no longer objects of dread.
Appealing as the notion of reduction of the unfamiliar to the familiar may be,
it is not a satisfactory characterization of scientific explanation. The point can best be
made in terms of a famous puzzle known as Olbers's paradoxwhich is named after
a nineteenth-century astronomer but was actually formulated by Edmund Halley in
1720why is the sky dark at night? Nothing could be more familiar than the darkness
of the night sky. But Halley and later astronomers realized that, if Newton's conception of the universe were correct, then the whole night sky should shine as brightly
as the noonday sun. The question of how to explain the darkness of the sky at night
is extremely difficult, and there may be no answer generally accepted by the experts.
Among the serious explanations that have been offered, however, appeal is made to
such esoteric facts as the non-Euclidean character of space or the mean free path of
photons in space. In this case, and in many others as well, a familiar phenomenon is
explained by reference to facts that are very unfamiliar indeed.
I suspect that a deep connection exists between the anthropomorphic conception
of explanation and the thesis that explanation consists in reduction of the unfamiliar
to the familiar. The type of explanation with which we are best acquainted is that in
which human action is explained in terms of conscious purposes. If it is possible to
explain the phenomena of physics or biology in terms of attempting to realize a goal,
that is a striking case of reduction to the familiar. A problem with this approach is,
of course, that a great deal of the progress in scientific understanding has resulted in
the elimination, not the injection, of purposes.
1.8 THE DEDUCTIVE-NOMOLOGICAL PATTERN OF SCIENTIFIC
EXPLANATION
In a classic 1948 paper, Carl G. Hempel and Paul Oppenheim formulated, with great
precision, one pattern of scientific explanation that is central to all discussions of the
subject. It is known as the deductive-nomological (D-N) model of scientific explana14
Scientific Explanation
tion. Stated very simply, an explanation of this type explains by subsuming its
explanandum-fact under a general law. This can best be appreciated by looking at an
example.
A figure skater with arms outstretched stands balanced on one skate. Propelling
herself with her other skate she begins to rotate slowly. She stops propelling herself,
but she continues to rotate slowly for a few moments. Suddenlywithout propelling
herself again and without being propelled by any external object, such as another
skatershe begins spinning very rapidly. Why? Because she drew her arms in close
to her body, thus concentrating her total body mass closer to the axis of rotation.
Because of the law of conservation of angular momentum, her rate of rotation had to
increase to compensate for her more compact body configuration.
More technically, the angular momentum of an object is the product of its
angular velocity (rate of rotation) and its moment of inertia. The moment of inertia
depends upon the mass of the object and the average distance of the mass from the
axis of rotation; for a fixed mass, the moment of inertia is smaller the more compactly
the mass is distributed about the axis of rotation. The law of conservation of angular
momentum says that the angular momentum of a body that is not being propelled or
retarded by external forces does not change; hence, since the moment of inertia is
decreased, the rate of rotation must increase to keep the value of the product constant.6
According to Hempel and Oppenheim, an explanation of the foregoing sort is to
be viewed as a deductive argument. It can be set out more formally as follows:
(3)
The angular momentum of any body (whose rate of rotation is not being increased or decreased by external forces) remains constant.
The skater is not interacting with any external object in such a way as to alter her angular
velocity.
The skater is rotating (her angular momentum is not zero).
The skater reduces her moment of inertia by drawing her arms in close to her body.
The skater's rate of rotation increases.
In this example we may ignore the friction of the skate on the ice, and the friction of the skater's body
in the surrounding air.
Scientific Explanation
15
1. The explanandum must be a logical consequence of the explanans; that is, the
explanation must be a valid deductive argument.
2. The explanans must contain at least one general law, and it must actually be
required for the derivation of the explanandum; in other words, if the law or
laws were deleted, without adding any new premises, the argument would no
longer be valid.
3. The explanans must have empirical content; it must be capable, at least in
principle, of test by experiment or observation.
4. The sentences constituting the explanans must be true.
These conditions are evidently fulfilled by our example. The first three are classified
as logical conditions of adequacy; the fourth is empirical. An argument that fulfills all
four conditions is an explanation (for emphasis we sometimes say ''true explanation"). An argument that fulfills the first three conditions, without necessarily fulfilling the fourth, is called a potential explanation. It is an argument that would be an
explanation if its premises were true.7
According to Hempel and Oppenheim, it is possible to have D-N explanations,
not only of particular occurrences as in argument (3), but also of general laws. For
example, in the context of Newtonian mechanics, it is possible to set up the following
argument:
(4) F = ma (Newton's second law).
For every action there is an equal and opposite reaction (Newton's third law).
In every interaction, the total linear momentum of the system of interacting bodies remains
constant (law of conservation of linear momentum).
This argument is valid, and among its premises are statements of general laws. There
are no statements of antecedent conditions, but that is not a problem since the
conditions of adequacy do not require them. Because we are not concerned to explain
any particular facts, no premises regarding particular facts are needed. Both premises
in the explanans are obviously testable, for they have been tested countless times.
Thus, argument (4) fulfills the logical conditions of adequacy, and consequently, it
qualifies as a potential explanation. Strictly speaking, it does not qualify as a true
explanation, for we do not consider Newton's laws of motion literally true, but in
many contexts they can be taken as correct because they provide extremely accurate
approximations to the truth.
Although Hempel and Oppenheim discussed both deductive explanations of
particular facts and deductive explanations of general laws, they offered a precise
characterization only of the former, but not of the latter. They declined to attempt to
provide a characterization of explanations of general laws because of a problem they
recognized but did not know how to solve. Consider Kepler's laws of planetary
motion K and Boyle's law of gases B. If, on the one hand, we conjoin the two to form
7
Hempel and Oppenheim provide, in addition to these conditions of adequacy, a precise technical
definition of "explanation." In this book we will not deal with these technicalities.
16
Scientific Explanation
a law K . B, we can obviously deduce K from it. But this could not be regarded as
an explanation of K, for it is only a pointless derivation of K from itself. On the other
hand, the derivation of K from Newton's laws of motion and gravitation constitutes
an extremely illuminating explanation of Kepler's laws. Hempel and Oppenheim
themselves confessed that they were unable to provide any criterion to distinguish the
pointless pseudoexplanations from the genuine explanations of laws (see Hempel and
Oppenheim 1948 as reprinted in Hempel 1965b, 273, f.n. 33).
Hempel and Oppenheim envisioned two types of D-N explanation, though they
were able to provide an account of only one of them. In addition, they remarked that
other types of explanation are to be found in the sciences, namely, explanations that
appeal, not to universal generalizations, but to statistical laws instead (ibid., 250251).
Table 1.1 shows the four kinds of explanations to which Hempel and Oppenheim called attention; they furnished an account only for the type found in the upper
left-hand box. Some years later Hempel (1962) offered an account of the I-S pattern
in the lower left-hand box. In Hempel (1965b) he treated both the I-S and the D-S
patterns. In 1948, Hempel and Oppenheim were looking forward to the time when
theories of explanation dealing with all four boxes would be available.
TABLE 1.1
^^^^Explanada
Laws
^^^^^
Particular Facts
General Regularities
Universal
Laws
D-N
Deductive-Nomological
D-N
Deductive-Nomological
Statistical
Laws
I-S
Inductive-Statistical
D-S
Deductive-Statistical
Hempel and Oppenheim emphasized the crucial role played by laws in scientific
explanation; in fact, the D-N pattern is often called the covering-law model. As we
will see, laws play a central part in other conceptions of scientific explanation as well.
Roughly speaking, a law is a regularity that holds throughout the universe, at all
places and all times. A law-statement is simply a statement to the effect that such a
regularity exists. A problem arises immediately. Some regularities appear to be
lawful and others do not. Consider some examples of laws:
(i) All gases, kept in closed containers of fixed size, exert greater pressure when
heated.
(ii) In all closed systems the quantity of energy remains constant.
(iii) No signals travel faster than light.
Contrast these with the following:
Scientific Explanation
17
If the occurrence of the kilogram in (vi) seems to make reference to a particular objectthe international prototype kilogram kept at the international bureau of standardsthe problem can easily be circumvented
by defining mass in terms of atomic mass units.
18
Scientific Explanation
conditionals. Even if we agree that nowhere in the entire history of the universe
past, present, or futuredoes there exist a gold sphere of mass greater than 100,000
kilograms, we would not be justified in claiming that it is impossible to fabricate a
gold sphere of such mass. I once made a rough calculation of the amount of gold in
the oceans of the earth, and it came to about 1,000,000 kilograms. If an incredibly
rich prince were determined to impress a woman passionately devoted to golden
spheres it would be physically possible for him to extract a little more than 100,000
kilograms from the sea to create a sphere that massive.
Statement (vi) also lacks the capacity to support counterfactual conditionals.
We would not be justified in concluding that, if two golden hemispheres, each of
50,001 kilogram mass, were put together, they would not form a golden sphere of
mass greater than 100,000 kilograms. To appreciate the force of this point, consider
the following statement:
(vii) No enriched uranium sphere has a mass greater than 100,000 kilograms.
This is a lawful generalization, because the critical mass for a nuclear chain reaction
is just a few kilograms. If 100,000 kilograms of enriched uranium were to be assembled, we would have a gigantic nuclear explosion. No comparable catastrophe
would ensue, as far as we know, if a golden sphere of the same mass were put
together.
Philosophers have often claimed that we can distinguish true generalizations
that are lawful from those that are accidental. Even if we grant the truth of (vi), we
must conclude that it is an accidental generalization. Moreover, they have maintained
that among universal generalizations, regardless of truth, it is possible to distinguish
lawlike generalizations from those that are not lawlike. A lawlike generalization is
one that has all of the qualifications for being a law except, perhaps, being true.
It is relatively easy to point to the characteristic of statements (iv) and (v) that
makes them nonlawlike, namely, that they make reference to particular objects,
persons, events, places, or times. The nonlawlike character of statement (vi) is harder
to diagnose. One obvious suggestion is to apply the criteria of supporting counterfactual and/or modal statements. We have seen that (vi) fails on that score. The
problem with that approach is that it runs a serious risk of turning out to be circular.
Consider statement (ii). Why do we consider it physically impossible to build a
perpetual motion machine (of the first type)? Because to do so would violate a law of
nature, namely (ii). Consider statement (vi). Why do we consider it physically
possible to fabricate a golden sphere whose mass exceeds 100,000 kilograms? Because to do so would not violate a law of nature. It appears that the question of what
modal statements to accept hinges on the question of what regularities qualify as laws
of nature.
A similar point applies to the support of counterfactual conditionals. Consider
statement (i). Given a container of gas that is not being heated, we can say that, if it
were to be heated, it would exert increased pressure on the walls of its container
sufficient in many cases to burst the container. (I learned my lesson on this as a Boy
Scout heating an unopened can of beans in a camp fire.) The reason that we can make
such a counterfactual claim is that we can infer from statement (i) what would
Scientific Explanation
19
happen, and (i) states a law of nature. Similarly, from (iii) we can deduce that if
something travels faster than light it is not a signalthat is, it cannot transmit
information. You might think that this is vacuous because, as the theory of relativity
tells us, nothing can travel faster than light. However, this opinion is incorrect.
Shadows and various other kinds of' 'things" can easily be shown to travel faster than
light. We can legitimately conclude that, if something does travel faster than light, it
is not functioning as a signal, because (iii) is, indeed, a law of nature.
What are the fundamental differences between statement (vi) on the one hand
and statements (i)-(iii) and (vii) on the other? The main difference seems to be that
(i)-(iii) and (vii) are all deeply embedded in well-developed scientific theories, and
that they have been, directly or indirectly, extensively tested. This means that (i)-(iii)
and (vii) have a very different status within our body of scientific knowledge than do
(iv)-(vi). The question remains, however, whether the regularities described by
(i)-(iii) and (vii) have a different status in the physical universe than do (iv)-(vi).
At the very beginning of this chapter, we considered the explanation of the fact
that the moon assumes a coppery hue during total eclipse. This is a regularity found
in nature, but is it a lawful regularity? Is the statement, "The moon turns a coppery
color during total eclipses," a law-statement? The immediate temptation is to respond
in the negative, for the statement makes an explicit reference to a particular object,
namely, our moon. But if we reject that statement as a lawful generalization, it would
seem necessary to reject Kepler's laws of planetary motion as well, for they make
explicit reference to our solar system. Galileo's law of falling bodies would also have
to go, for it refers to things falling near the surface of the earth. It would be unreasonable to disqualify all of them as laws.
We can, instead, make a distinction between basic and derived laws. Kepler's
laws and Galileo's law can be derived from Newton's laws of motion and gravitation,
in conjunction with descriptions of the solar system and the bodies that make it up.
Newton's laws are completely general and make no reference to any particular person, object, event, place, or time. The statement about the color of the moon during
total eclipse can be derived from the laws of optics in conjunction with a description
of the earth's atmosphere and the configuration of the sun, moon, and earth when an
eclipse occurs. The statement about the color of the moon can also be taken as a
derivative law. But what about statements (iv) and (v)? The color of the apples in my
refrigerator can in no way be derived from basic laws of nature in conjunction with
a description of the refrigerator. No matter how fond I may be of golden delicious
apples, there is no physical impossibility of a red delicious getting into my refrigerator. Similarly, there are no laws of nature from which, in conjunction with descriptions of the Apache and their baskets, it would be possible to derive that they can only
be made by women.
Scientific Explanation
The counterexample was devised by Sylvain Bromberger, but to the best of my knowledge he never
published it.
Scientific Explanation
21
tion of the sun is a crucial causal factor in the relation between the height of the
flagpole and the length of the shadow, the flagpole and its shadow play no causal role
in the position of the sun in the sky.
CE-2. The barometer and the storm. Given a sharp drop in the reading on a
properly functioning barometer, we can predict that a storm will shortly occur.
Nevertheless, the reading on the barometer does not explain the storm. A sharp drop
in atmospheric pressure, which is registered on the barometer, explains both the storm
and the barometric reading.
The moral: Many times we find two effects of a common cause that are correlated with one another. In such cases we do not explain one effect by means of the
other. The point is illustrated also by diseases. A given illness may have many
different symptoms. The disease explains the symptoms; one symptom does not
explain another.
CE-3. A solar eclipse. From the present positions of the earth, moon, and
sun, using laws of celestial mechanics, astronomers can predict a future total eclipse
of the sun. After the eclipse has occurred, the very same data, laws, and calculations
provide a legitimate D-N explanation of the eclipse. So far, so good. However, using
the same laws and the same positions of the earth, moon, and sun, astronomers can
retrodict the previous occurrence of a solar eclipse. The argument by which this
retrodiction is made fulfills the requirements for a D-N explanation just as fully as
does the prediction of the eclipse. Nevertheless, most of us would say that, while it
is possible to explain an eclipse in terms of antecedent conditions, it is not possible
to explain an eclipse in terms of subsequent conditions.
The moral: We invoke earlier conditions to explain subsequent facts; we do not
invoke later conditions to explain earlier facts. The reason for this asymmetry seems
to lie in the fact that causes, which have explanatory import, precede their effects
they do not follow their effects.
CE-4. The man and the pilL A man explains his failure to become pregnant
during the past year on the ground that he has regularly consumed his wife's birth
control pills, and that any man who regularly takes oral contraceptives will avoid
getting pregnant.
The moral: This example shows that it is possible to construct valid deductive
arguments with true premises in which some fact asserted by the premises is actually
irrelevant. Since men do not get pregnant regardless, the fact that this man took birth
control pills is irrelevant. Nevertheless, it conforms to the D-N pattern.
Counterexamples CE-l-CE-4 are all cases in which an argument that fulfills the
Hempel-Oppenheim requirements manifestly fails to constitute a bona fide explanation. They were designed to show that these requirements are too weak to sort out the
illegitimate explanations. A natural suggestion would be to strengthen them in ways
that would rule out counterexamples of these kinds. For example, CE-1 and CE-2
could be disqualified if we stipulated that the antecedent conditions cited in the
explanans must be causes of the explanandum. CE-3 could be eliminated by insisting
that the so-called antecedent conditions must actually obtain prior to the explanandum. And CE-4 could be ruled out by stipulating that the antecedent conditions must
22
Scientific Explanation
23
of C from the atmosphere; wood that has been cut cannot do so. Here is the D-S
explanation:
(5)
Every C14 atom (that is not exposed to external radiation10) has a probability of x/i of
disintegrating within any period of 5730 years.
In any large collection of C 14 atoms (that are not exposed to external radiation) approximately
three-fourths will very probably decay within 11,460 years.
Almost all cases of streptococcus infection clear up quickly after the administration of penicillin.
Jane Jones had a streptococcus infection.
10
This qualification is required to assure that the disintegration is spontaneous and not induced by
external radiation.
11
As James Lennox remarks in Chapter 7 on philosophy of biology, Darwin's principle of natural
selection is an example of a statistical law.
24
Scientific Explanation
This explanation is an argument that has three premises (the explanans); the first
premise states a statistical regularitya statistical lawwhile the other two state
antecedent conditions. The conclusion (the explanandum) states the fact-to-beexplained. However, a crucial difference exists between explanations (3) and (6):
D-N explanations subsume the events to be explained deductively, while I-S explanations subsume them inductively. The single line separating the premises from the
conclusion in (3) signifies a relation of deductive entailment between the premises
and conclusion. The double line in (6) represents a relationship of inductive support,
and the attached variable r stands for the strength of that support. This strength of
support may be expressed exactly, as a numerical value of a probability, or vaguely,
by means of such phrases as "very probably" or "almost certainly."
An explanation of either of these two kinds can be described as an argument to
the effect that the event to be explained was to be expected by virtue of certain
explanatory facts. In a D-N explanation, the event to be explained is deductively
certain, given the explanatory facts; in an I-S explanation the event to be explained
has high inductive probability relative to the explanatory facts. This feature of expectability is closely related to the explanation-prediction symmetry thesis for explanations of particular facts. According to this thesis any acceptable explanation of a
particular fact is an argument, deductive or inductive, that could have been used to
predict the fact in question if the facts stated in the explanans had been available prior
to its occurrence. I2 As we shall see, this symmetry thesis met with serious opposition.
Hempel was not by any means the only philosopher in the early 1960s to notice
that statistical explanations play a significant role in modern science. He was, however, the first to present a detailed account of the nature of statistical explanation, and
the first to bring out a fundamental problem concerning statistical explanations of
particular facts. The case of Jane Jones and her quick recovery can be used as an
illustration. It is well known that certain strains of the streptococcus bacterium are
penicillin-resistant, and if Jones's infection were of that type, the probability of her
quick recovery after treatment with penicillin would be small. We could, in fact, set
up the following inductive argument:
(7) Almost no cases of penicillin-resistant streptococcus infection clear up quickly after the
administration of penicillin.
Jane Jones had a penicillin-resistant streptococcus infection.
Jane Jones received treatment with penicillin.
Jane Jones did not recover quickly.
The remarkable fact about arguments (6) and (7) is that their premises are
mutually compatiblethey could all be true. Nevertheless, their conclusions contra12
This thesis was advanced for D-N explanation in Hempel-Oppenheim (1948, 249), and reiterated,
with some qualifications, for D-N and I-S explanations in Hempel (1965a, Sections 2.4, 3.5).
Scientific Explanation
25
diet one another. This is a situation that can never occur with deductive arguments.
Given two valid deductions with incompatible conclusions, their premises must also
be incompatible. Thus, the problem that has arisen in connection with I-S explanations has no analog in D-N explanations. Hempel called this the problem of ambiguity
of I-S explanation.
The source of the problem of ambiguity is a simple and fundamental difference
between universal laws and statistical laws. Given the proposition that all A are B, it
follows immediately that all things that are both A and C are B. If all humans are mortal,
then all people who are over six feet tall are mortal. However, even if almost all humans
who are alive now will be alive five years from now, it does not follow that almost all
living humans with advanced cases of pancreatic cancer will be alive five years hence.
As we noted in Section 1.5, there is a parallel fact about arguments. Given a valid
deductive argument, the argument will remain valid if additional premises are supplied,
as long as none of the original premises is taken away. Deduction is erosion-proof.
Given a strong inductive argumentone that supports its conclusion with a high degree
of probabilitythe addition of one more premise may undermine it completely. For
centuries Europeans had a great body of inductive evidence to support the proposition
that all swans are white, but one true report of a black swan in Australia completely
refuted that conclusion. Induction is not erosion-proof.
Hempel sought to resolve the problem of ambiguity by means of his requirement of maximal specificity (RMS). It is extremely tricky to state RMS with precision,
but the basic idea is fairly simple. In constructing I-S explanations we must include
all relevant knowledge we have that would have been available, in principle, prior to
the explanandum-fact. If the information that Jones's infection is of the penicillinresistant type is available to us, argument (6) would not qualify as an acceptable I-S
explanation.13
In Section 1.8 we stated Hempel and Oppenheim's four conditions of adequacy
for D-N explanations. We can now generalize these conditions so that they apply both
to D-N and I-S explanations as follows:
1. The explanation must be an argument having correct (deductive or inductive)
logical form.
2. The explanans must contain at least one general law (universal or statistical),
and this law must actually be required for the derivation of the explanandum.
3. The explanans must have empirical content; it must be capable, at least in
principle, of test by experiment or observation.
4. The sentences constituting the explanans must be true.
5. The explanation must satisfy the requirement of maximal specificity. 14
13
Nor would (6) qualify as an acceptable I-S explanation if we had found that Jones's infection was of
the non-penicillin-resistant variety, for the probability of quick recovery among people with that type of
infection is different from the probability of quick recovery among those who have an unspecified type of
streptococcus infection.
14
D-N explanations of particular facts automatically satisfy this requirement. If all A are B, the probability that an A is a B is one. Under those circumstances, the probability that an A which is also a C is a B is
also one. Therefore, no partition of A is relevant to B.
26
Scientific Explanation
We noticed in Section 1.10 that major criticisms of the D-N pattern of scientific
explanation can be posed by means of well-known counterexamples. The same situation arises in connection with the I-S pattern. Consider the following:
CE-6. Psychotherapy. Suppose that Bruce Brown has a troublesome neurotic
symptom. He undergoes psychotherapy and his symptom disappears. Can we explain
his recovery in terms of the treatment he has undergone? We could set out the
following inductive argument, in analogy with argument (6):
(8)
Most people who have a neurotic symptom of type N and who undergo psychotherapy
experience relief from that symptom.
Bruce Brown had a symptom of type N and he underwent psychotherapy.
Bruce Brown experienced relief from his symptom.
Before attempting to evaluate this proffered explanation we should take account of the
fact that there is a fairly high spontaneous remission ratethat is, many people who
suffer from that sort of symptom get better regardless of treatment. No matter how
large the number r, if the rate of recovery for people who undergo psychotherapy is
no larger than the spontaneous remission rate, it would be a mistake to consider
argument (8) a legitimate explanation. A high probability is not sufficient for a correct
explanation. If, however, the number r is not very large, but is greater than the
spontaneous remission rate, the fact that the patient underwent psychotherapy has at
least some degree of explanatory force. A high probability is not necessary for a
sound explanation.
Another example reinforces the same point.
CE-7. Vitamin C and the common cold* 15 Suppose someone were to claim
that large doses of vitamin C would produce rapid cures for the common cold. To
ascertain the efficacy of vitamin C in producing rapid recovery from colds, we should
note, it is not sufficient to establish that most people recover quickly; most colds
disappear within a few days regardless of treatment. What is required is a double15
Around the time Hempel was working out his theory of I-S explanation, Linus Pauling's claims about
the value of massive doses of vitamin C in the prevention of common colds was receiving a great deal of
attention. Although Pauling made no claims about the ability of vitamin C to cure colds, it occurred to me that
a fictitious example of this sort could be concocted.
Scientific Explanation
27
blind controlled experiment in which the rate of quick recovery for those who take
vitamin C is compared with the rate of quick recovery for those who receive only a
placebo. If there is a significant difference in the probability of quick recovery for
those who take vitamin C and for those who do not, we may conclude that vitamin
C has some degree of causal efficacy in lessening the duration of colds. If, however,
there is no difference between the two groups, then it would be a mistake to try to
explain a person's quick recovery from a cold by constructing an argument analogous
to (6) in which that result is attributed to treatment with vitamin C.
The moral. CE-6 and CE-7 call attention to the same point as CE-4 (the man and
the pill). All of them show that something must be done to exclude irrelevancies from
scientific explanations. If the rate of pregnancy among men who consume oral contraceptives is the same as for men who do not, then the use of birth control pills is
causally and explanatorily irrelevant to pregnancy among males. Likewise, if the rate
of relief from neurotic symptoms is the same for those who undergo psychotherapy
as it is for those who do not, then psychotherapy is causally and explanatorily
irrelevant to the relief from neurotic symptoms. Again, if the rate of rapid recovery
from common colds is the same for those who do and those who do not take massive
doses of vitamin C, then consumption of massive doses of vitamin C is causally and
explanatorily irrelevant to rapid recovery from colds.17 Hempel's requirement of
maximal specificity was designed to insure that all relevant information (of a suitable
sort) is included in I-S explanations. What is needed in addition is a requirement
insuring that only relevant information is included in D-N or I-S explanations.
CE-8. Syphilis and paresis. Paresis is a form of tertiary syphilis which can be
contracted only by people who go through the primary, secondary, and latent forms
of syphilis without treatment with penicillin. If one should ask why a particular
person suffers from paresis, a correct answer is that he or she was a victim of
untreated latent syphilis. Nevertheless, only a small proportion of those with untreated latent syphilisabout 25%actually contract paresis. Given a randomly
selected member of the class of victims of untreated latent syphilis, one should predict
that that person will not develop paresis.
The moral: there are legitimate I-S explanations in which the explanans does not
render the explanandum highly probable. CE-8 responds to the explanation-prediction
symmetry thesisthe claim that an explanation is an argument of such a sort that it
could have been used to predict the explanandum if it had been available prior to the
16
In a controlled experiment there are two groups of subjects, the experimental group and the control
group. These groups should be as similar to one another as possible. The members of the experimental group
receive the substance being tested, vitamin C. The members of the control group receive a placebo, that is, an
inert substance such as a sugar pill that is known to have no effect on the common cold. In a blind experiment
the subjects do not know whether they are receiving vitamin C or the placebo. This is important, for if the
subjects knew which treatment they were receiving, the power of suggestion might skew the results. An
experiment is double-blind if neither the person who hands out the pills nor the subjects know which subject is
getting which type of pill. If the experiment is not double-blind, the person administering the pills might, in spite
of every effort not to, convey some hint to the subject.
17
It should be carefully noted that I am claiming neither that psychotherapy is irrelevant to remission of
neurotic symptoms nor that vitamin C is irrelevant to rate of recovery from colds. I am saying that that is the
point at issue so far as I-S explanation is concerned.
28
Scientific Explanation
When we look at an I-S explanation such as (6), there is a strong temptation to regard
it as incomplete. It may, to be sure, incorporate all of the relevant knowledge we
happen to possess. Nevertheless, we may feel, it is altogether possible that medical
science will discover enough about streptococcus infections and about penicillin
treatment to be able to determine precisely which individuals with strep infections
will recover quickly upon treatment with penicillin and which individuals will not.
When that degree of knowledge is available we will not have to settle for I-S explanations of rapid recoveries from strep infections; we will be able to provide D-N
explanations instead. Similar remarks can also be made about several of the
counterexamplesin particular, examples CE-6-CE-9.
Consider CE-8, the syphilis-paresis example. As remarked above, with our
present state of knowledge we can predict that about 25% of all victims of untreated
latent syphilis contract paresis, but we do not know how to distinguish those who will
develop paresis from those who will not. Suppose Sam Smith develops paresis. At
this stage of our knowledge the best we can do by way of an I-S explanation of
Smith's paresis is the following:
(9)
This could not be accepted as an I-S explanation because of the weakness of the
relation of inductive support.
Suppose that further research on the causes of paresis reveals a factor in the
bloodcall it the P-factorwhich enables us to pick out, with fair reliabilitysay
Scientific Explanation
29
95%those who will develop paresis. Given that Smith has the P-factor, we can
construct the following argument:
(10)
95% of all victims of untreated latent syphilis who have the P-factor develop paresis.
Smith had untreated latent syphilis.
Smith had the P-factor.
[95]
Smith developed paresis.
In the knowledge situation just described, this would count as a pretty good I-S
explanation, for 0.95 is fairly close to 1.
Let us now suppose further that additional medical research reveals that, among
those victims of untreated latent syphilis who have the P-factor, those whose spinal
fluid contains another factor Q invariably develop paresis. Given that information,
and the fact that Smith has the g-factor, we can set up the following explanation:
(11)
All victims of untreated latent syphilis who have the P-factor and the g-factor develop
paresis.
Smith had untreated latent syphilis.
Smith had the P-factor.
Smith had the g-factor.
Smith developed paresis.
If the suppositions about the P-factor and the g-factor were true, this argument would
qualify as a correct D-N explanation. We accepted (10) as a correct explanation of
Smith's paresis only because we were lacking the information that enabled us to set
up (11).
Determinism is the doctrine that says that everything that happens in our universe is completely determined by prior conditions.18 If this thesis is correct, then
each and every event in the history of the universepast, present, or futureis, in
principle, deductively explainable. If determinism is true, then every sound I-S explanation is merely an incomplete D-N explanation. Under these circumstances, the
I-S pattern is not really a stand-alone type of explanation; all fully correct explanations fit the D-N pattern. The lower left-hand box of Table 1.1 would be empty. This
does not mean that I-S explanationsthat is, incomplete D-N explanationsare
useless, only that they are incomplete.
Is determinism true? We will not take a stand on that issue in this chapter.
Modern physicsquantum mechanics in particularseems to offer strong reasons to
believe that determinism is false, but not everyone agrees with this interpretation.
However, we will take the position that indeterminism may be true, and see what the
consequences are with respect to statistical explanation.
According to most physicists and philosophers of physics, the spontaneous
disintegration of the nucleus of an atom of a radioactive substance is a genuinely
indeterministic happening. Radioactive decay is governed by laws, but they are
18
30
[r]
In this example, r differs from 1 by an incredibly tiny margin, but is not literally equal
to 1. In a world that is not deterministic, I-S explanations that are not merely incomplete D-N explanations can be formulated.
1.14 THE STATISTICAL RELEVANCE (S-R) MODEL OF
EXPLANATION
According to the received view, scientific explanations are arguments; each type of
explanation in Table 1.1 is some type of argument satisfying certain conditions. For
this reason, we can classify the received view as an inferential conception of scientific
explanation. Because of certain difficulties, associated primarily with I-S explanation,
another pattern for statistical explanations of particular occurrences was developed. A
fundamental feature of this model of explanation is that it does not construe explanations as arguments.
One of the earliest objections to the I-S pattern of explanationas shown by
CE-6 (psychotherapy) and CE-7 (vitamin C and the common cold)is that statistical
relevance rather than high probability is the crucial relationship in statistical explanations. Statistical relevance involves a relationship between two different probabilities. Consider the psychotherapy example. Bruce Brown is a member of the class of
people who have a neurotic symptom of type N. Within that class, regardless of what
the person does in the way of treatment or nontreatment, there is a certain probability
of relief from the symptom (R). That is the prior probability of recovery; let us
symbolize it as "Pr(/?/A0" Then there is a probability of recovery in the class of
19
This qualification is required to assure that the disintegrations have been spontaneous and not induced
by external radiation.
Scientific Explanation
31
people with that symptom who undergo psychotherapy (F); it can be symbolized as
"Pr(RfN.P)." If
Pr(RIN.P) > Pr(R/N)
then psychotherapy is positively relevant to recovery, and if
Pr(R/N.P)
<Pr(R/N)
S.C.W
-s.c.w
S.C.-W
-S.CrW
S .-C .W
~S.~C.W
S.~C.~W
~S.~C.~W
Scientific Explanation
Items 1-3 constitute the explanans; the explanandum is Green's heart attack. Moreover, no restrictions are placed on the size of the probabilitiesthey can be high,
middling, or low. All that is required is that these probabilities differ from one another
in various ways, because we are centrally concerned with relations of statistical
relevance.
Although the S-R pattern of scientific explanation provides some improvements
over the I-S model, it suffers from a fundamental inadequacy. It focuses on statistical
relevance rather than causal relevance. It may, as a result, tend to foster a confusion
of causes and correlations. In the vitamin C example, for instance, we want a controlled experiment to find out whether taking massive doses of vitamin C is causally
relevant to quick recovery from colds. We attempt to find out whether taking vitamin
C is statistically relevant to rapid relief because the statistical relevance relation is
evidence regarding the presence or absence of causal relevance. It is causal relevance
that has genuine explanatory import. The same remark applies to other examples as
well. In the psychotherapy example we try to find out whether such treatment is
statistically relevant to relief from neurotic symptoms in order to tell whether it is
causally relevant. In the case of the heart attack, many clinical studies have tried to
find statistical relevance relations as a basis for determining what is causally relevant
to the occurrence of serious heart attacks.
We have been looking at the development of the received view, and at some of the
criticisms that have been leveled against it. The strongest intuitive appeal of that view
comes much more from explanations of laws than from explanations of particular
facts. One great example is the Newtonian synthesis. Prior to Newton we had a
miscellaneous collection of laws including Kepler's three laws of planetary motion
and Galileo's laws of falling objects, inertia, projectile motion, and pendulums. By
invoking three simple laws of motion and one law of gravitation, Newton was able to
explain these lawsand in some cases correct them. In addition, he was able to
explain many other regularities, such as the behavior of comets and tides, as well.
Later on, the molecular-kinetic theory provided a Newtonian explanation of many
laws pertaining to gases. Quite possibly the most important feature of the Newtonian
synthesis was the extent to which it systematized our knowledge of the physical world
by subsuming all sorts of regularities under a small number of very simple laws.
Another excellent historical example is the explanation of light by subsumption under
Maxwell's theory of electromagnetic radiation.
The watchword in these beautiful historical examples is unification. A large
number of specific regularities are unified in one theory with a small number of
assumptions or postulates. This theme was elaborated by Michael Friedman (1974)
who asserted that our comprehension of the universe is increased as the number of
independently acceptable assumptions we require is reduced. I would be inclined to
add that this sort of systematic unification of our scientific knowledge provides a
comprehensive world picture or worldview. This, I think, represents one major aspect
of scientific explanationit is the notion that we understand what goes on in the
Scientific Explanation
33
world if we can fit it into a comprehensive worldview. As Friedman points out, this
is a global conception of explanation. The value of explanation lies in fitting things
into a universal pattern, or a pattern that covers major segments of the universe.20
As we look at many of the criticisms that have been directed against the received
view, it becomes clear that causality is a major focus. Scriven offered his ink stain
example, CE-5, to support the claim that finding the explanation amounts, in many
cases, simply to finding the causes. This is clearly explanation on a very local level.
All we need to do, according to Scriven, is to get a handle on events in an extremely
limited spacetime region that led up, causally, to the stain on the carpet, and we have
adequate understanding of that particular fact. In this connection, we should also
recall CE-1 and CE-2. In the first of these we sought a local causal explanation for the
length of a shadow, and in the second we wanted a causal explanation for a particular
storm. Closely related noncausal ''explanations" were patently unacceptable. In such
cases as the Chernobyl accident and the Challenger space-shuttle explosion we also
seek causal explanations, partly in order to try to avoid such tragedies in the future.
Scientific explanation has its practical as well as its purely intellectual value.
It often happens, when we try to find causal explanations for various occurrences, that we have to appeal to entities that are not directly observable with the
unaided human senses. For example, to understand AIDS (Acquired Immunodeficiency Syndrome), we must deal with viruses and cells. To understand the transmission of traits from parents to offspring, we become involved with the structure of the
DNA molecule. To explain a large range of phenomena associated with the nuclear
accident at Three Mile Island, we must deal with atoms and subatomic particles.
When we try to construct causal explanations we are attempting to discover the
mechanismsoften hidden mechanismsthat bring about the facts we seek to understand. The search for causal explanations, and the associated attempt to expose the
hidden workings of nature, represent a second grand tradition regarding scientific
explanation. We can refer to it as the causal-mechanical tradition.
Having contrasted the two major traditions, we should call attention to an
important respect in which they overlap. When the search for hidden mechanisms is
successful, the result is often to reveal a small number of basic mechanisms that
underlie wide ranges of phenomena. The explanation of diverse phenomena in terms
of the same mechanisms constitutes theoretical unification. For instance, the kineticmolecular theory of gases unified thermodynamic phenomena with Newtonian particle mechanics. The discovery of the double-helical structure of DNA, for another
example, produced a major unification of biology and chemistry.
Each of the two grand traditions faces certain fundamental problems. The tradition of explanation as unificationassociated with the received viewstill faces
the problem concerning explanations of laws that was pointed out in 1948 by Hempel
and Oppenheim. It was never solved by Hempel in any of his subsequent work on
scientific explanation. If the technical details of Friedman's theory of unification were
satisfactory, it would provide a solution to that problem. Unfortunately, it appears to
encounter serious technical difficulties (see Kitcher 1976 and Salmon 1989).
20
The unification approach has been dramatically extended and improved by Philip Kitcher (1976, 1981,
and 1989).
34
Scientific Explanation
Scientific Explanation
35
and semantical terms alone. Pragmatic considerations were not dealt with. Hempel's
later characterization of the other types of explanations were given mainly in syntactical and semantical terms, although I-S explanations are, as we noted, relativized
to knowledge situations. Knowledge situations are aspects of the human contexts in
which explanations are sought and given. Such contexts have other aspects as well.
One way to look at the pragmatic dimensions of explanation is to start with the
question by which an explanation is sought. In Section 1.3 we touched briefly on this
matter. We noted that many, if not all, explanations can properly be requested by
explanation-seeking why-questions. In many cases, the first pragmatic step is to
clarify the question being asked; often the sentence uttered by the questioner depends
upon contextual clues for its interpretation. As Bas van Fraassen, one of the most
important contributors to the study of the pragmatics of explanation, has shown, the
emphasis with which a speaker poses a question may play a crucial role in determining just what question is being asked. He goes to the Biblical story of the Garden of
Eden to illustrate. Consider the following three questions:
(i) Why did Adam eat the apple?
(ii) Why did Adam eat the apple?
(iii) Why did Adam eat the apple?
Although the words are the sameand in the same orderin each, they pose three
very different questions. This can be shown by considering what van Fraassen calls
the contrast class. Sentence (i) asks why Adam ate the apple instead of a pear, a
banana, or a pomegranate. Sentence (ii) asks why Adam, instead of Eve, the serpent,
or a goat, ate the apple. Sentence (iii) asks why Adam ate the apple instead of
throwing it away, feeding it to a goat, or hiding it somewhere. Unless we become
clear on the question being asked, we can hardly expect to furnish appropriate answers.
Another pragmatic feature of explanation concerns the knowledge and intellectual ability of the person or group requesting the explanation. On the one hand, there
is usually no point in including in an explanation matters that are obvious to all
concerned. Returning to (3)our prime example of a D-N explanation of a particular
factone person requesting an explanation of the sudden dramatic increase in the
skater's rate of rotation might have been well aware of the fact that she drew her arms
in close to her body, but unfamiliar with the law of conservation of angular momentum. For this questioner, knowledge of the law of conservation of angular momentum
is required in order to understand the explanandum-fact. Another person might have
been fully aware of the law of conservation of angular momentum, but failed to notice
what the skater did with her arms. This person needs to be informed of the skater's
arm maneuver. Still another person might have noticed the arm maneuver, and might
also be aware of the law of conservation of angular momentum, but failed to notice
that this law applies to the skater's movement. This person needs to be shown how
to apply the law in the case in question.
On the other hand, there is no point in including material in an explanation that
is beyond the listeners' ability to comprehend. To most schoolchildren, for example,
36
Scientific Explanation
an explanation of the darkness of the night sky that made reference to the nonEuclidean structure of space or the mean free path of a photon would be inappropriate. Many of the explanations we encounter in real-life situations are incomplete on
account of the explainer's view of the background knowledge of the audience.
A further pragmatic consideration concerns the interests of the audience. A
scientist giving an explanation of a serious accident to a congressional investigating
committee may tell the members of Congress far more than they want to know about
the scientific details. In learning why an airplane crashed, the committee might be
very interested to find that it was because of an accumulation of ice on the wing, but
totally bored by the scientific reason why ice-accumulations cause airplanes to crash.
Peter Railton (1981) has offered a distinction that helps considerably in understanding the role of pragmatics in scientific explanation. First, he introduces the
notion of an ideal explanatory text. An ideal explanatory text contains all of the facts
and all of the laws that are relevant to the explanandum-fact. It details all of the causal
connections among those facts and all of the hidden mechanisms. In most cases the
ideal explanatory text is huge and complex. Consider, for example, an explanation of
an automobile accident. The full details of such items as the behavior of both drivers,
the operations of both autos, the condition of the highway surface, the dirt on the
windshields, and the weather, would be unbelievably complicated. That does not
really matter, for the ideal explanatory text is seldom, if ever, spelled out fully. What
is important is to have the ability to illuminate portions of the ideal text as they are
wanted or needed. When we do provide knowledge to fill in some aspect of the ideal
text we are furnishing explanatory information.
A request for a scientific explanation of a given fact is almost alwaysif not
literally alwaysa request, not for the ideal explanatory text, but for explanatory
information. The ideal text contains all of the facts and laws pertaining to the
explanandum-fact. These are the completely objective and nonpragmatic aspects of
the explanation. If explanatory information is to count as legitimate it must correspond to the objective features of the ideal text. The ideal text determines what is
relevant to the explanandum-fact. Since, however, we cannot provide the whole ideal
text, nor do we want to, a selection of information to be supplied must be made. This
depends on the knowledge and interests of those requesting and those furnishing
explanations. The information that satisfies the request in terms of the interests and
knowledge of the audience is salient information. The pragmatics of explanation
determines saliencethat is, what aspects of the ideal explanatory text are appropriate for an explanation in a particular context.
1.17 CONCLUSION
Several years ago, a friend and colleaguewhom I will call the friendly physicist
was sitting on a jet airplane awaiting takeoff. Directly across the aisle was a young
boy holding a helium-filled balloon by a string. In an effort to pique the child's
curiosity, the friendly physicist asked him what he thought the balloon would do
when the plane accelerated for takeoff. After a moment's thought the boy said that it
would move toward the back of the plane. The friendly physicist replied that he
Scientific Explanation
37
thought it would move toward the front of the cabin. Several adults in the vicinity
became interested in the conversation, and they insisted that the friendly physicist was
wrong. A flight attendant offered to wager a miniature bottle of Scotch that he was
mistakena bet that he was quite willing to accept. Soon thereafter the plane accelerated, the balloon moved forward, and the friendly physicist enjoyed a free drink.23
Why did the balloon move toward the front of the cabin? Two explanations can
be offered, both of which are correct. First, one can tell a story about the behavior of
the molecules that made up the air in the cabin, explaining how the rear wall collided
with nearby molecules when it began its forward motion, thus creating a pressure gradient from the back to the front of the cabin. This pressure gradient imposed an unbalanced force on the back side of the balloon, causing it to move forward with respect
to the walls of the cabin.24 Second, one can cite an extremely general physical
principleEinstein's principle of equivalenceaccording to which an acceleration is
physically equivalent, from the standpoint of the occupants of the cabin, to a gravitational field. Since helium-filled balloons tend to rise in the atmosphere in the earth's
gravitational field, they will move forward when the airplane accelerates, reacting just
as they would if a massive object were suddenly placed behind the rear wall.
The first of these explanations is causal-mechanical. It appeals to unobservable
entities, describing the causal processes and causal interactions involved in the explanandum phenomenon. When we are made aware of these explanatory facts we
understand how the phenomenon came about. This is the kind of explanation that
advocates of the causal-mechanical tradition find congenial. The second explanation
illustrates the unification approach. By appealing to an extremely general physical
principle, it shows how this odd little occurrence fits into the universal scheme of
things. It does not refer to the detailed mechanisms. This explanation provides a
different kind of understanding of the same fact.
Which of these explanations is correct? Both are. Both of them are embedded
in the ideal explanatory text. Each of them furnishes valuable explanatory information. It would be a serious error to suppose that any phenomenon has only one
explanation. It is a mistake, I believe, to ask for the explanation of any occurrence.
Each of these explanations confers a kind of scientific understanding. Pragmatic
considerations might dictate the choice of one rather than the other in a given context.
For example, the explanation in terms of the equivalence principle would be unsuitable for a ten-year-old child. The same explanation might be just right in an undergraduate physics course. But both are bona fide explanations.
As we noted in Section 1.10, the 1948 Hempel-Oppenheim essay attracted
almost no attention for about a decade after its publication. Around 1959 it became
the focus of intense controversy, much of it stemming from those who saw causality
as central to scientific explanation. The subsequent thirty years have seen a strong
opposition between the advocates of the received view and the proponents of causal
explanation. Each of the two major approaches has evolved considerably during this
23
This little story was previously published in Salmon (1980). I did not offer an explanation of the
phenomenon in that article.
24
Objects that are denser than air do not move toward the front of the cabin because the pressure
difference is insufficient to overcome their inertia.
38
Scientific Explanation
periodindeed, they have developed to the point that they can peacefully coexist as
two distinct aspects of scientific explanation. Scientific understanding is, after all, a
complicated affair; we should not be surprised to learn that it has many different
aspects. Exposing underlying mechanisms and fitting phenomena into comprehensive
pictures of the world seem to constitute two important aspects. Moreover, as remarked above, we should remember that these two types of understanding frequently
overlap. When we find that the same mechanisms underlie diverse types of natural
phenomena this ipso facto constitutes a theoretical unification.
On one basic thesis there is nearly complete consensus. Recall that in the early
decades of the twentieth century many scientists and philosophers denied that there
can be any such thing as scientific explanation. Explanation is to be found, according
to this view, only in the realms of theology and metaphysics. At present it seems
virtually unanimously agreed that, however it may be explicated, there is such a thing
as scientific explanation. Science can provide deep understanding of our world. We
do not need to appeal to supernatural agencies to achieve understanding. Equally
importantly, we can contrast the objectively based explanations of contemporary
science with the pseudounderstanding offered by such flagrantly unscientific approaches as astrology, creation science, and Scientology. These are points worth
remembering in an age of rampant pseudoscience.
QUESTIONS
1. Must every scientific explanation contain a law of nature? According to philosophers who
support "the received view" the answer is affirmative. Other philosophers have answered in the
negative. Discuss critically the arguments pro and con. Give your own answer, supported by
reasons.
2. Are there any inductive or statistical explanations of particular facts? In their classic 1948 paper
Hempel and Oppenheim say that there are such explanations, but do not offer any explication of
their nature. Later attempts to work out the details ran into many difficulties. Discuss these
problems and say whether you think they are insuperable. Give your reasons.
3. According to the explanation-prediction symmetry thesis, every satisfactory scientific explanation could (in some suitable context) serve as a scientific prediction, and every scientific prediction could (in some suitable context) serve as a scientific explanation. Critically discuss both
parts of this symmetry thesis. Give your reasons for accepting or rejecting each part.
4. Are there any fundamental differences between explanations in the natural sciences and explanations in the social sciences? (See Merrilee H. Salmon's chapter on philosophy of the social
sciences.) Are there basic differences between human behavior and the behavior of other kinds
of physical objects that make one kind more amenable to explanation than the other? Is explanation of human behavior that involves conscious deliberation and free choice possible? Discuss
critically.
5. In this chapter it was suggested that "No gold sphere has a mass greater than 100,000 kg" is not
a lawlike statement, whereas "No enriched uranium sphere has a mass greater than 100,000 kg"
is a lawlike statement. Discuss the distinction between lawlike and accidental generalizations.
Explain as clearly as possible why one is lawlike and the other is not.
6. Discuss the role of causality in scientific explanation. Do all legitimate scientific explanations
make reference to causal relations? Is causality essentially irrelevant to scientific explanation?
Are some good explanations causal and other good explanations not? Discuss critically.
Scientific Explanation
39
7. Choose an actual example of a scientific explanation from a magazine such as Scientific American, Science, Nature, American Scientist, or from a textbook you have used in a science course.
Give a concise summary of this explanation, and analyze it in terms of the models (such as D-N,
I-S, D-S, S-R) and concepts (such as covering law, causal-mechanical, unification) introduced
in this chapter. Evaluate the explanation in terms of these models and/or concepts.
8. In Section 1.9 it was claimed that
(i) All gases, kept in closed containers of fixed size, exert greater pressure when heated
is a general statement, whereas
(v) All Apache basketry is made by women
is not completely general because it refers specifically to a particular group of people. But, it
might be objected, (i) refers to physical objects of a specific type, namely, gases in closed
containers, so it is not completely general either. Moreover, (v) is a general statement about the
Apache. Discuss this objection. Hint: Statement (i) can be formulated as follows: "If anything
is a gas in a closed container that is heated, it will expand." But: Statement (v) can likewise be
reformulated as follows: "If anything is an Apache basket, it was made by a woman." Is there
a fundamental logical difference between the two statements as reformulated?
SUGGESTED READINGS
(1974), "Explanation and Scientific Understanding," Journal of Philosophy
71: 5-19. Reprinted in Pitt (1988). This is the original statement of the unification approach to
scientific explanation. Although brief highly technical parts appear near the end, the article
contains interesting general discussion of the basic issues.
HEMPEL, CARL G. (1942), "The Function of General Laws in History," Journal of Philosophy 39:
35^48. Reprinted in Hempel (1965b). Hempel's original essay on explanation in history.
. (1959), "The Logic of Functional Analysis," in Llewellyn Gross (ed.), Symposium on
Sociological Theory. New York: Harper & Row, pp. 271-307. Reprinted in Hempel (1965b).
Hempel's original essay on functional explanation.
. (1962), "Explanation in Science and in History," in Robert G. Colodny (ed.), Frontiers
of Science and Philosophy. Pittsburgh: University of Pittsburgh Press, pp. 7-33. A lucid and
highly readable brief presentation of Hempel's basic views.
. (1965a), "Aspects of Scientific Explanation," in Hempel (1965b), pp. 331-496. Hempel's magisterial comprehensive monograph on scientific explanation.
. (1965b), Aspects of Scientific Explanation and Other Essays in the Philosophy of Science.
New York: The Free Press. Part 4 contains four classic articles on scientific explanation.
. (1966), Philosophy ofNatural Science. Englewood Cliffs, NJ: Prentice-Hall. Chapters 5-8
offer an extremely elementary and highly readable introduction to Hempel's views.
HEMPEL, CARL G. and PAUL OPPENHEIM (1948), "Studies in the Logic of Explanation," Philosophy of Science 15: 135-175. Reprinted in Hempel (1965b) and in Pitt (1988). This is the
modern classic on scientific explanation. Parts 1 and 3 are especially important.
HUMPHREYS, PAUL (1981), "Aleatory Explanation," Synthese 48: 225-232. An introductory
account of a new approach to statistical explanation.
JEFFREY, RICHARD C. (1969), "Statistical Explanation vs. Statistical Inference," in Nicholas
Rescher (ed.), Essays in Honor of Carl G. Hempel. Dordrecht: Reidel, pp. 104-113. Reprinted
in Wesley C. Salmon and others (1970) Statistical Explanation and Statistical Relevance. PittsFRIEDMAN, MICHAEL
40
Scientific Explanation
burgh: University of Pittsburgh Press. This philosophical gem explicitly raises the question of
whether explanations are arguments.
PITT, JOSEPH C. (ed.) (1988), Theories of Explanation. New York: Oxford University Press. An
anthology containing a number of important articles.
RAILTON, PETER (1981), "Probability, Explanation, and Information," Synthese 48: 233-256. A
clear and careful elementary exposition of a mechanistic approach to probabilistic explanation.
SALMON, WESLEY C. (1978), "Why Ask, 'Why'?An Inquiry Concerning Scientific Explanation, '' Proceedings and Addresses of the American Philosophical Association 51: 683-705. An
introductory discussion of explanation by means of unobservable entities.
. (1982), "Comets, Pollen, and Dreams: Some Reflections on Scientific Explanation," in
Robert McLaughlin (ed.), What? Where? When? Why? Dordrecht: Reidel, pp. 155-178. A
popularized discussion of some of the basic issues concerning scientific explanation.
. (1990), Four Decades of Scientific Explanation. Minneapolis: University of Minnesota
Press. A historical survey of philosophical developments since the classic Hempel-Oppenheim
(1948) essay.
SCRIVEN, MICHAEL (1959), "Explanation and Prediction in Evolutionary Theory," Science 130:
477-482. One of the first strong challenges to Hempel's account.
. (1962), "Explanations, Predictions, and Laws," in Herbert Feigl and Grover Maxwell
(eds.), Minnesota Studies in the Philosophy ofScience. Vol. 3, Scientific Explanation, Spaceand
Time. Minneapolis: University of Minnesota Press, pp. 170-230. Another strong attack on the
received view of scientific explanation.
VAN FRAASSEN, BAS C. (1980), The Scientific Image. Oxford: Clarendon Press. Chapter 5 presents
an influential treatment of the pragmatics of explanation.
WRIGHT, LARRY (1976), Teleological Explanations. Berkeley and Los Angeles: University of
California Press. An important treatment of teleological and functional explanation.
Scientific Explanation
41
Two
T H E CONFIRMATION
OF SCIENTIFIC HYPOTHESES
The physical, biological, and behavioral sciences are all empirical. This means that
their assertions must ultimately face the test of observation. Some scientific statements face the observational evidence directly; for example, "All swans are white,"
was supported by many observations of European swans, all of which were white, but
it was refuted by the observation of black swans in Australia. Other scientific statements confront the observational evidence in indirect ways; for instance, "Every
proton contains three quarks," can be checked observationally only by looking at the
results of exceedingly complex experiments. Innumerable cases, of course, fall between these two extremes.
Human beings are medium-sized objects; we are much larger than atoms and
much smaller than galaxies. Our environment is full of other medium-sized things
for example, insects, frisbees, automobiles, and skyscrapers. These can be observed
with normal unaided human senses. Other things, such as microbes, are too small to
be seen directly; in these cases we can use instruments of observation
microscopesto extend our powers of observation. Similarly, telescopes are extensions of our senses that enable us to see things that are too far away to be observed
directly. Our senses of hearing and touch can also be enhanced by various kinds of
instruments. Ordinary eyeglassesin contrast to microscopes and telescopesare
not extensions of normal human senses; they are devices that provide more normal
sight for those whose vision is somewhat impaired.
An observation that correctly reveals the featuressuch as size, shape, color,
and textureof what we are observing is called veridical. Observations that are not
veridical are illusory. Among the illusory observations are hallucinations, afterimages, optical illusions, and experiences that occur in dreams. Philosophical arguments
going back to antiquity show that we cannot be absolutely certain that our direct
observations are veridical. It is impossible to prove conclusively, for example, that
any given observation is not a dream experience. That point must be conceded. We
can, however, adopt the attitude that our observations of ordinary middle-sized physical objects are reasonably reliable, and that, even though we cannot achieve certainty, we can take measures to check on the veridicality of our observations and
make corrections as required (see Chapter 4 for further discussion of the topics of
skepticism and antirealism).
We can make a rough and ready distinction among three kinds of entities: (i)
those that can be observed directly with normal unaided human senses; (ii) those that
can be observed only indirectly by using some instrument that extends the normal
human senses; and (iii) those that cannot be observed either directly or indirectly,
whose existence and nature can be established only by some sort of theoretical
inference. We do not claim that these distinctions are precise; that will not matter for
our subsequent discussion. We say much more about category (iii) and the kinds of
inferences that are involved as this chapter develops.
Our scientific languages should also be noted to contain terms of two types. We
The Confirmation of Scientific Hypotheses
43
As we have seen, science contains some statements that are reports of direct observation, and others that are not. When we ask how statements of this latter type are to
meet the test of experience, the answer often given is the hypothetico-deductive (H-D)
method; indeed, the H-D method is sometimes offered as the method of scientific
inference. We must examine its logic.
The term hypothesis can appropriately be applied to any statement that is intended for evaluation in terms of its consequences. The idea is to articulate some
statement, particular or general, from which observational consquences can be drawn.
An observational consequence is a statementone that might be true or might be
falsewhose truth or falsity can be established by making observations. These observational consequences are then checked by observation to determine whether they
are true or false. If the observational consequence turns out to be true, that is said to
confirm the hypothesis to some degree. If it turns out to be false, that is said to
disconfirm the hypothesis.
Let us begin by taking a look at the H-D testing of hypotheses having the form
of universal generalizations. For a very simple example, consider Boyle's law of
44
gases, which says that, for any gas kept at a constant temperature T, the pressure P
is inversely proportional to the volume V,1 that is,
p x V = constant (at constant T).
This implies, for example, that doubling the pressure on a gas will reduce its volume
by a half. Suppose we have a sample of gas in a cylinder with a movable piston, and
that the pressure of the gas is equal to the pressure exerted by the atmosphereabout
15 pounds per square inch. It occupies a certain volume, say, 1 cubic foot. We now
apply an additional pressure of 1 atmosphere, making the total pressure 2 atmospheres. The volume of the gas decreases to Vi cubic foot. This constitutes a
hypothetico-deductive confirmation of Boyle's law. It can be schematized as follows:
(1)
Argument (1) is a valid deduction. The first premise is the hypothesis that is being
tested, namely, Boyle's law. It should be carefully noted, however, that Boyle's law
is not the only premise of this argument. From the hypothesis alone it is impossible
to deduce any observational prediction; other premises are required. The following
four premises state the initial conditions under which the test is performed. The
conclusion is the observational prediction that is derived from the hypothesis and the
initial conditions. Since the temperature, pressure, and volume can be directly measured, let us assume for the moment that we need have no serious doubts about the
truth of the statements of initial conditions. The argument can be schematized as
follows:
(2) H (test hypothesis)
/ (initial conditions)
O (observational prediction)
This relationship does not hold for temperatures and pressures close to the point at which the gas in
question condenses into a liquid or solid state.
The Confirmation of Scientific Hypotheses
45
conclusion is a valid deduction but the argument from the conclusion to the premises
is not. If it has any merit at all, it must be as an inductive argument.
Let us reconstruct the argument from the observational prediction to the hypothesis as follows:
(3)
The
The
The
The
The
No one would seriously suppose that (3) establishes Boyle's law conclusively, or
even that it renders the law highly probable. At best, it provides a tiny bit of inductive
support. If we want to provide solid inductive support for Boyle's law it is necessary
to make repeated tests of this gas, at the same temperature, for different pressures and
volumes, and to make other tests at other temperatures. In addition, other kinds of
gases must be tested in a similar manner.
In one respect, at least, our treatment of the test of Boyle's law has been
oversimplified. In carrying out the test we do not directly observesay by feeling the
containerthat the initial and final temperatures of the gas are the same. Some type
of thermometer is used; what we observe directly is not the temperature of the gas but
the reading on the thermometer. We are therefore relying on an auxiliary hypothesis
to the effect that the thermometer is a reliable instrument for the measurement of
temperature. On the basis of an additional hypothesis of this sort we claim that we can
observe the temperature indirectly. Similarly, we do not observe the pressures directly, by feeling the force against our hands; instead, we use some sort of pressure
gauge. Again, we need an auxiliary hypothesis stating that the instrument is a reliable
indicator.
The need for auxiliary hypotheses is not peculiar to the example we have
chosen. In the vast majority of casesif not in every caseauxiliary hypotheses are
required. In biological and medical experiments, for example, microscopes of various
types are employedfrom the simple optical type to the tunneling scanning electron
microscope, each of which requires a different set of auxiliary hypotheses. Likewise,
in astronomical work telescopesrefracting and reflecting optical, infrared, radio,
X-ray, as well as cameras are used. The optical theory of the telescope and the
chemical theory of photographic emulsions are therefore required as auxiliary hypotheses. In sophisticated physical experiments using particle accelerators, an elaborate set of auxiliary hypotheses concerning the operation of all of the various sorts
of equipment is needed. In view of this fact, schema (2) should be expanded:
(4) H (test hypothesis)
A (auxiliary hypotheses)
/ (initial conditions)
O (observational prediction)
46
Up to this point we have considered the case in which the observational prediction turns out to be true. The question arises, what if the observational prediction
happens to be false? To deal with this case we need a different example.
At the beginning of the nineteenth century a serious controversy existed about
the nature of light. Two major hypotheses were in contention. According to one
theory light consists of tiny particles; according to the other, light consists of waves.
If the corpuscular theory is true, a circular object such as a coin or ball bearing, if
brightly illuminated, will cast a uniformly dark circular shadow. The following H-D
test was performed:
(5)
Surprisingly, when the experiment was performed, it turned out that the shadow had
a bright spot in its center. Thus, the result of the test was negative; the observational
prediction was false.
Argument (5) is a valid deduction; accordingly, if its premises are true its
conclusion must also be true. But the conclusion is not true. Hence, at least one of the
premises must be false. Since the second premise was known to be true on the basis
of direct observation, the first premisethe corpuscular hypothesismust be false.
We have examined two examples of H-D tests of hypotheses. In the first,
Boyle's law, the outcome was positivethe observational prediction was found to be
true. We saw that, even assuming the truth of the other premises in argument (1), the
positive outcome could, at best, lend a small bit of support to the hypothesis. In the
second, the corpuscular theory of light, the outcome was negativethe observational
prediction was found to be false. In that case, assuming the truth of the other premise,
the hypothesis was conclusively refuted.
The negative outcome of an H-D test is often less straightforward than the
example just discussed. For example, astronomers who used Newtonian mechanics
to predict the orbit of the planet Uranus found that their observational predictions
were incorrect. In their calculations they had, of course, taken account only of the
gravitational influences of the planets that were known at the time. Instead of taking the negative result of the H-D test as a refutation of Newtonian mechanics, they
postulated the existence of another planet that had not previously been observed.
That planet, Neptune, was observed shortly thereafter. An auxiliary hypothesis concerning the constitution of the solar system was rejected rather than Newtonian
mechanics.
It is interesting to compare the Uranus example with that of Mercury. Mercury
also moves in a path that differs from the orbit calculated on the basis of Newtonian
mechanics. This irregularity, however, could not be successfully explained by postulating another planet, though this strategy was tried. As it turned out, the perturbation of Mercury's orbit became one of three primary pieces of evidence supporting
Einstein's general theory of relativitythe theory that has replaced Newtonian me2
Except when they pass from one medium (e.g., air) to another medium (e.g., glass or water).
47
chanics in the twentieth century. The moral is that negative outcomes of H-D tests
sometimes do, and sometimes do not, result in the refutation of the test hypothesis.
Since auxiliary hypotheses are almost always present in H-D tests, we must face the
possibility that an auxiliary hypothesis, rather than the test hypothesis, is responsible
for the negative outcome.
2.3 PROBLEMS WITH THE HYPOTHETICO-DEDUCTIVE METHOD
The H-D method has two serious shortcomings that must be taken into account. The
first of these might well be called the problem of alternative hypotheses. Let us
reconsider the case of Boyle's law. If we represent that law graphically, it says that
a plot of pressures against volumes is a smooth curve, as shown in Figure 2.1.
The result of the test, schematized in argument (1), is that we have two points
(indicated by arrows) on this curveone corresponding to a pressure of 1 atmosphere
and a volume of 1 cubic foot, the other corresponding to a pressure of 2 atmospheres
and a volume of 2 cubic foot. While these two points conform to the solid curve
shown in the figure, they agree with infinitely many other curves as wellfor example, the dashed straight line through those two points. If we perform another test,
with a pressure of 3 atmospheres, we will find that it yields a volume of V3 cubic foot.
This is incompatible with the straight line curve, but the three points we now have are
still compatible with infinitely many curves, such as the dotted one, that go through
these three. Obviously we can make only afinitenumber of tests; thus, it is clear that,
no matter how many tests we make, the results will be compatible with infinitely
many different curves.
2.5
2.0
t Volume
(cu.ft.)
10
0.5
0
0
0.5
1.0
1.5
2.0
Pressure
(atm.)
w
^
2.5
3.0
3.5
Figure 2.1
48
The best known alternative to the H-D method is an account of qualitative confirmation developed by Carl G. Hempel (1945). The leading idea of Hempel's approach
is that hypotheses are confirmed by their positive instances. Although seemingly
simple and straightforward, this intuitive idea turns out to be difficult to pin down.
Consider, for example, Nicod's attempt to explicate the idea for universal conditionals; for example:
H: (x) (Rx D Bx) (All ravens are black).
(The symbol "(*)" is the so-called universal quantifier, which can be read, "for
every object x"\ " D " is the sign of material implication, which can be read very
roughly "if. . .then. . . .") Although this statement is too simpleminded to qualify
as a serious scientific hypothesis, the logical considerations that will be raised apply
to all universal generalizations in science, no matter how sophisticatedsee Section
49
Such examples might lead one to try to build the equivalence condition into the definition of Nicodconfirmation along the following lines:
(N') E Nicod-confirms H just in case there is an H' such that \~ H = H' and such that E implies that
the objects mentioned satisfy both the antecedent and consequent of H'.
But as the following example due to Hempel shows, (N') leads to confirmation where it is not wanted in the case
of multiply quantified hypotheses. Consider
H: (x) (y) Rxy
H'\ (x) (y)[~(Rxy . Ryx) D (Rxy~ Ryx)]
E: Rab. ~ Rba
E implies that the pair a, b satisfies both the antecedent and the consequent of H', and H' is logically equivalent
to H. So by (N') E Nicod-confirms H. But this is an unacceptable result since E contradicts H.
50
Special consequence condition: If E confirms H and HY- H' then E confirms H'.
Consistency condition: If E confirms H and also confirms H' then H and H' are
logically consistent.
As a result, he rejects
Converse consequence condition: If E confirms H and H' \H then E confirms H'.
For to accept the converse consequence condition along with the entailment and
special consequence conditions would lead to the disastrous result that any E confirms
any H. (Proof of this statement is one of the exercises at the end of this chapter.) Note
that the H-D account satisfies the converse consequence condition but neither the
special consequence condition nor the consistency condition.
Hempel provided a definition of confirmation that satisfies all of his adequacy
conditions. The key idea of his definition is that of the development, dev/H), of a
hypothesis H for a set / of individuals. Intuitively, devf(H) is what H says about a
domain that contains exactly the individuals of/. Formally, universal quantifiers are
replaced by conjunctions and existential quantifiers are replaced by disjunctions. For
example, let / = {a, b}, and take
H: (x) Bx (Everything is beautiful)
then
devj(H) = Ba.Bb.
Or take
Hr: (3x) Rx (Something is rotten)
then
deVjiH') =
RavRb.
(The wedge " v " symbolizes the inclusive disjunction; it means "and/or"that is,
"one, or the other, or both.") Or take
H": (JC) (3y) Lxy (Everybody loves somebody);
then
devj{H") = (Laa V Lab).(Zi>a V Lbb). 4
Using this notion we can now state the main definitions:
4
In formulas like H" that have mixed quantifiers, we proceed in two steps, working from the inside out.
In the first step we replace the existential quantifier by a disjunction, which yields
(x) (Lxa v Lxb).
In the next step we replace the universal quantifier with a conjunction, which yields devffl').
The Confirmation of Scientific Hypotheses
51
=
=
=
=
X
Y + Z
Y + X
Z
The Os are supposed to be observable quantities while the Xs and Fs are theoretical.
For purposes of a concrete example, suppose that we have samples of four
different gases in separate containers. All of the containers have the same volume,
and they are at the same pressure and temperature. According to Avogadro's law,
then, each sample contains the same number of molecules. Observable quantities
Ox~04 are simply the weights of the four samples:
Ox = 28 g, 02 = 44 g, 03 = 44 g, 04 = 28 g.
Our hypotheses say
H1: The first sample consists solely of molecular nitrogenN2molecular weight
28; X is the weight of a mole of N 2 (28 g).
H2: The second sample consists of carbon dioxideC02molecular weight 44; Y
is the weight of a mole of atomic oxygen O (16 g), Z is the weight of a mole
of carbon monoxide CO (28 g).
H3: The third sample consists of nitrous oxideN20molecular weight 44; Y is
the weight of a mole of atomic oxygen O (16 g) and X is the weight of a mole
of molecular nitrogen (28 g).
H4: The fourth sample consists of carbon monoxideCOmolecular weight 28; Z
is the weight of a mole of CO (28 g).
(The integral values for atomic and molecular weights are not precisely correct,
but they furnish a good approximation for this example.)
To show how Hx can be bootstrap-confirmed relative to the other three hypotheses, suppose that an experiment has determined values Ot9 02, 03, 04, for the
observables. From the values for 02 and 04 we can, using H2 and H4, compute values
for Y + Z and forZ. Together these determine a value for Y. Then from the value for
03 we can, using H3, compute a value for Y + X. Then from these latter two values
we get a value for X. Finally, we compare this computed value for X with the
The Confirmation of Scientific Hypotheses
53
observed value for Ox. If they are equal, H1 is confirmed. Although this simple
example may seem a bit contrived, it is in principle similar to the kinds of measurements and reasoning actually used by chemists in the nineteenth century to establish
molecular and atomic weights.
If we want the bootstrap procedure to constitute a test in the sense that it carries
with it the potential for falsification, then we should also require that there are
possible values for the observables such that, using these values and the very same
bootstrap calculations that led to a confirmatory instance, values for the theoretical
quantities are produced that contradict the hypothesis in question. This requirement is
met in the present example.
In Glymour's original formalization of the bootstrapping idea, macho bootstrapping was allowed; that is, in deducing instances of H, it was allowed that H itself
could be used as an auxiliary assumption. To illustrate, consider again the earlier
example of the perfect gas law P(ressure) X V(olume) = K X T(emperature), and
suppose P, V, T to be observable quantities while the gas constant K is theoretical.
We proceed to bootstrap-test this law relative to itself by measuring the observables
on two different occasions and then comparing the values kx and k2 for K deduced
from the law itself and the two sets of observation values pl9 vls tx and p2, v2, t2.
However, macho bootstrapping can lead to unwanted results, and in any case it may
be unnecessary since, for instance, in the gas law example it is possible to analyze the
logic of the test without using the very hypothesis being tested as an auxiliary
assumption in the bootstrap calculation (see Edidin 1983 and van Fraassen 1983).
These and other questions about bootstrap testing are currently under discussion in the
philosophy journals. (The original account of bootstrapping, Glymour 1980, is open
to various counterexamples discussed in Christensen 1983; see also Glymour 1983.)
Let us now return to Hempel's account of confirmation to ask whether it is too
liberal. Two reasons for giving a positive answer are contained in the following
paradoxes.
Paradox of the ravens. Consider again the hypothesis that all ravens are
black: (x) (Rx D Bx). Which of the following evidence statements Hempel-confirm
the ravens hypothesis?
Ex\ Ral.Bal
E2: ~ Ra2
E3: Ba3
E4: ~ Ra4.~ Ba4
E5: ~ Ra5.Ba5
E6:Ra^.- Ba6
The answer is that Ex-Es all confirm the hypothesis. Only the evidence E6 that refutes
the hypothesis fails to confirm it. The indoor ornithology of some of these Hempelconfirmation relationsthe confirmation of the ravens hypothesis, say, by the evidence that an individual is a piece of white chalkhas seemed to many to be too easy
to be true.
Goodman's paradox. If anything seems safe in this area it is that the evidence
Ra.Ba that a is a black raven confirms the ravens hypothesis (x) (Rx D Bx). But on
54
Hempel's approach nothing rides on the interpretation of the predicates Rx and Bx.
Thus, Hempel confirmation would still obtain if we interpreted Bx to mean that x is
blite, where "blite" is so defined that an object is blite if it is examined on or before
December 31, 2000, and is black or else is examined afterwards and found to be
white. Thus, by the special consequence condition, the evidence that a is a black
raven confirms the prediction that if b is a raven examined after 2000, it will be white,
which is counterintuitive to say the least.
Puzzles of the sort just mentionedinvolving blite ravens and grue emeralds (an
object is grue if it is examined on or before December 31, 2000 and is green, or it is
examined thereafter and is blue)were presented in Nelson Goodman (1955) under
the rubric of the new riddle of induction. Goodman sought the basis of our apparent
willingness to generalize inductively with respect to such predicates as "black,"
"white," "green," and "blue," but not with respect to "blite" and "grue." To
mark this distinction he spoke of projectible predicates and unprojectible predicates;
and he supposed that there are predicates of each of these types. The problem is to find
grounds for deciding which are which.
There is, however, a difficulty that is both historically and logically prior. In his
Treatise of Human Nature ([1739-1740] 1978) and his Enquiry Concerning Human
Understanding (1748) David Hume called into serious question the thesis that we
have any logical or rational basis for any inductive generalizationsthat is, for
considering any predicate to be projectible.
Hume divided all reasoning into two types, reasoning concerning relations of
ideas and reasoning concerning matters of fact and existence. All of the deductive
arguments of pure mathematics and logic fall into the first category; they are unproblematic. In modern terminology we say that they are necessarily truth-preserving
because they are nonampliative (see Chapter 1, Section 1.5). If the premises of any
such argument are true its conclusion must also be true because the conclusion says
nothing that was not said, at least implicitly, by the premises.
Not all scientific reasoning belongs to the first category. Whenever we make
inferences from observed facts to the unobserved we are clearly reasoning
ampliativelythat is, the content of the conclusion goes beyond the content of the
premises. When we predict future occurrences, when we retrodict past occurrences,
when we make inferences about what is happening elsewhere, and when we establish
generalizations that apply to all times and places we are engaged in reasoning concerning matters of fact and existence. In connection with reasoning of the second type
Hume directly poses the question: What is the foundation of our inferences from the
observed to the unobserved? He readily concludes that such reasoning is based upon
relations of cause ancf effect. WTien we see lightning nearby (cause) we infer that the
sound of thunder (effect) will ensue. When we see human footprints in the sand
The Confirmation of Scientific Hypotheses
55
(effect) we infer that a person recently walked there (cause). When we hear a knock
and a familiar voice saying "Anybody home?" (effect) we infer the presence of a
friend (cause) outside the door.
The next question arises automatically: How can we establish knowledge of the
cause-effect relations to which we appeal in making inferences from the observed to
the unobserved? Hume canvasses several possibilities. Do we have a priori knowledge of causal relations? Can we look at an effect and deduce the nature of the cause?
He answers emphatically in the negative. For a person who has had no experience of
diamonds or of icewhich are very similar in appearancethere is no way to infer
that intense heat and pressure can produce the former but would destroy the latter.
Observing the effect, we have no way to deduce the cause. Likewise, for a person
who has had no experience of fire or snow, there is no way to infer that the former
will feel hot while the latter will feel cold. Observing the cause, we have no way to
deduce the effect. All of our knowledge of causal relations must, Hume argues, be
based upon experience.
When one event causes another event, we might suppose that three factors are
presentnamely, the cause, the effect, and the causal connection between them.
However, in scrutinizing such situations Hume fails to find the third itemthe causal
connection itself. Suppose that one billiard ball lies at rest on a table while another
moves rapidly toward it. They collide. The ball that was at rest begins to move. What
we observe, Hume notes, is the initial motion of the one ball and its collision with the
other. We observe the subsequent motion of the other. This is, he says, as perfect a
case of cause and effect as we will ever see. We notice three things about the
situation. The first is temporal priority; the cause comes before the effect. The second
is spatiotemporal proximity; the cause and effect are close together in space and time.
The third is constant conjunction; if we repeat the experiment many times we find that
the result is just the same as thefirsttime. The ball that was at rest always moves away
after the collision.
Our great familiarity with situations similar to the case of the billiard balls may
give us the impression that "it stands to reason" that the moving ball will produce
motion in the one at rest, but Hume is careful to point out that a priori reasoning cannot
support any such conclusion. We can, without contradiction, imagine many possibilities: When they collide the two balls might vanish in a puff of smoke; the moving ball
might jump right over the one at rest; or the ball that is initially at rest might remain
fixed while the moving ball returns in the direction from which it came. Moreover, no
matter how closely we examine the situation, the thing we cannot see, Hume maintains,
is the causal connection itselfthe "secret power" by which the cause brings about
the effect. If we observe two events in spatiotemporal proximity, one of which follows
right after the other, just once, we cannot tell whether it is a mere coincidence or a
genuine causal connection. Hans Reichenbach reported an incident that occurred in a
theater in California as he was watching a movie. Just as a large explosion was depicted
on the screen the theatre began to tremble. An individual'sfirstinstinct was to link them
as cause and effect, but, in fact, by sheer coincidence, a minor earthquake occurred at
precisely that moment. Returning to Hume's billiard ball example, on the first observation of such a collision we would not know whether the motion of the ball originally
at rest occurred by coincidence or as a result of the collision with the moving ball. It
56
is only after repeated observations of such events that we are warranted in concluding
that a genuine causal relation exists. This fact shows that the causal connection itself
is not an observable feature of the situation. If it were an observable feature we would
not need to observe repetitions of the sequence of events, for we would be able to
observe it in the first instance.5
What, then, is the basis for our judgements about causal relations? Hume
answers that it is a matter of custom or habit. We observe, on one occasion, an event
of type C and observe that it is followed by an event of type E. On another occasion
we observe a similar event of type C followed by a similar event of type E. This
happens repeatedly. Thereafter, when we notice an event of type C we expect that it
will be followed by an event of type E. This is merely a fact about human psychology;
we form a habit, we become conditioned to expect E whenever C occurs. There is no
logical necessity in all of this.
Indeed, Hume uncovered a logical circle. We began by asking for the basis on
which inferences from the observed to the unobserved are founded. The answer was
that all such reasoning is based upon relations of cause and effect. We then asked how
we can establish knowledge of cause-effect relations. The answer was that we
assumeor psychologically anticipatethat future cases of events of type C will be
followed by events of type E, just as in past cases events of type C were followed by
events of type E. In other words, we assume that nature is uniformthat the future
will be like the pastthat regularities that have been observed to hold up to now will
continue to hold in the future.
But what reason do we have for supposing that nature is uniform? If you say
that nature's uniformity has been established on the basis of past observations, then
to suppose that it will continue to be uniform is simply to suppose that the future will
be like the past. That is flagrantly circular reasoning. If you say that science proceeds
on the presumption that nature is uniform, and that science has been extremely
successful in predicting future occurrences, Hume's retort is the same. To assume that
future scientific endeavors will succeed because science has a record of past success
is, again, to suppose that the future will be like the past. Furthermore, Hume points
out, it is entirely possible that nature will not be uniform in the futurethat the future
need not be like the pastfor we can consistently imagine all sorts of other possibilities. There is no contradiction in supposing that, at some future time, a substance
resembling snow should fall from the heavens, but that it would feel like fire and taste
like salt. There is no contradiction in supposing that the sun will not rise tomorrow
morning. One can consistently imagine that a lead ball, released from the hand,
would rise rather than fall. We do not expect such outlandish occurrences, but that is
a result of our psychological makeup. It is not a matter of logic.
5
The English philosopher John Locke had claimed that in one sort of situation we do observe the actual
power of one event to bring about another, namely, in cases in which a person has a volition or desire to perform
some act and does so as a result. We might, for example, wish to raise our arms, and then do so. According to
Locke we would be aware of our power to produce motion in a part of our body. Hume gave careful consideration to Locke's claim, and argued that it is incorrect. He points out, among other things, that there is a
complex relationship of which we are not directly awareinvolving transmission of impulses along nerves and
the contractions of various musclesbetween the volition originating in the brain and the actual motion of the
arm. Hume's critique effectively cut the ground from under Locke's claim.
The Confirmation of Scientific Hypotheses
57
Max Black and R. B. Braithwaite both argued that inductive justifications of induction could escape
circularity. The arguments of Black are criticized in detail in Salmon (1967, 12-17); Braithwaite's arguments
are open to analogous criticism.
The Confirmation of Scientific Hypotheses
59
reasonable (two classic statements of this view are given by Ayer 1956, 71-75 and
Strawson 1952, Chapter 9). The problem vanishes when we achieve a clear understanding of such terms as "evidence" and "rationality."
The foregoing argument is often reinforced by another consideration. Suppose
someone continues to demand a justification for the fundamental principles of induction, for example, that past regularities can be projected into the future. The question
then becomes, to what principle may we appeal in order to supply any such justification? Since the basic principles of inductive reasoning, like those of deductive
reasoning, are ultimate, it is impossible to find anything more basic in terms of which
to formulate a justification. Thus, the demand for justification of our most basic
principles is misplaced, for such principles define the concept of justification itself.
In spite of its popular appeal among philosophers, this attempt to dispose of
Hume's problem of justification of induction is open to serious objection. It can be
formulated in terms of a useful distinction, drawn by Herbert Feigl (1950), between
two kinds of justificationvalidation and vindication. A validation of a principle
consists in a derivation of that principle from other, more basic, principles that we
accept. For example, we often try to validate moral and/or legal principles. Some
people argue that abortion is wrong, and should be outlawed, because it is wrong to
take human life (except in certain extreme circumstances) and human life begins at
the time of conception. Others (in America) argue, by appealing to certain rights they
take to be guaranteed by the Constitution of the United States, that abortion should be
permitted. What counts as a validation for any individual obviously depends upon the
fundamental principles that person adopts.
Validation also occurs in mathematics and logic. The derivation of the
Pythagorean theorem from the postulates of Euclidean geometry constitutes a good
mathematical example. In logic, the inference rule modus tollens
(6)
p^q
~ q
~
(7)
P
q
and contraposition
(pDq)
(~qD~p).7
A less trivial example in deductive logic is the validation of the rule of conditional proof by means of
the deduction theorem. The deduction theorem shows that any conclusion that can be established by means of
conditional proof can be derived using standard basic deductive rules without appeal to conditional proof.
Conditional proof greatly simplifies many derivations, but it does not allow the derivation of any conclusion that
cannot be derived without it.
60
61
accept; an inference is rejected if it violates a rule we are unwilling to amend. The process
of justification is the delicate one of making mutual adjustments between rules and accepted
inferences; and in the agreement achieved lies the only justification needed for either. (1955,
67, italics in the original)
He continues:
All this applies equally well to induction. An inductive inference, too, is justified by
conformity to general rules, and a general rule by conformity to accepted inductive inferences. Predictions are justified if they conform to valid canons of induction; and the canons
are valid if they accurately codify accepted inductive practice. (Ibid.)
(8)
9
P
A priori rule: regardless of the makeup of the observed sample, infer that approximately lh of all marbles in the urn are red. (The fraction Vz is chosen because three
colors occur in the total population of marbles in the urn.)
Counterinductive rule: if mln of the marbles in the sample are red, infer that
approximately (n m)in of the marbles in the urn are red.
Certain characteristics of these rules can be established by general arguments.
The counterinductive rule is so called because it uses observed evidence in a negative
way. If we observe the proportion of red marbles in a sample, this rule instructs us
to project that the proportion of red in the whole population is approximately equal to
the proportion that are not red in the sample. Use of this rule would rapidly land us
in an outright contradiction. Suppose, for the sake of simplicity, that our observed
sample contains Vs red, xh yellow, and xh blue. Using the counterinductive rule for
each of the colors would yield the conclusion that % of the marbles in the urn are red,
and % of the marbles in the urn are yellow, and % of the marbles in the urn are blue.
This is logically impossible; clearly, the counterinductive rule is unsatisfactory.
Suppose we use the a priori rule. Then, even if 98 percent of our observed
sample were red, 1 percent yellow, and 1 percent blue, the rule would direct us to
ignore that empirical evidence and infer that only about xh of the marbles in the urn
are red. Because the a priori rule makes observation irrelevant to prediction, it, too,
should be rejected.
The rule of induction by enumeration does not have either of the foregoing
defects, and it has some virtues. One virtue is that if it is used persistently on larger
and larger samples, it must eventually yield inferences that are approximately correct.
If we are unlucky, and begin by drawing unrepresentative samples, it will take a long
time to start giving accurate results; if we are lucky and draw mainly representative
samples, the accurate results will come much sooner. (Some philosophers have
derived considerable comfort from the fact that the vast majority of samples that could
be drawn are very nearly representative. See Williams 1947.)
Obviously manyindeed, infinitely manypossible rules exist for making inductive inferences. The problem of deciding which of these rules to use is complicated and difficult. We have seen, nevertheless, that general considerations can be
brought to bear on the choice. It is not just a matter of consulting our intuitions
regarding the acceptability or nonacceptability of particular inductive inferences. This
is not to deny, however, that intuitive consideration of particular inferences has a
great deal of heuristic value.
Although we have been skeptical about Goodman's success in dismissing the
old riddle of induction, we must remark on the importance of his new riddle. First,
Hume never explicitly took account of the fact that some forms of constant conjunction do not give rise to habits of expectation. Such Goodmanian predicates as "blite"
and "grue" call attention vividly to this point. Second, Goodman's examples provide
another way of showing that there can be no noncircular justification of induction by
means of a uniformity principle. There are many uniformities, and the question of
which ones will extend into the future is the problem of induction all over again.
4. Deductivism.
One influential philosopher, Sir Karl Popper, has attacked
Hume's problem by denying that science involves any use of induction. He takes
The Confirmation of Scientific Hypotheses
63
Hume to have proved decisively that induction cannot be justified, and he concludes
that scienceif it is to be a rational enterprisemust do without it. The only logic
of science, he maintains, is deduction.
Popper characterizes the method of science as trial and error, as conjecture and
refutation. The scientist formulates bold explanatory hypotheses, and then subjects
them to severe testing. This test procedure is very much like hypothetico-deductive
testing, but there is an absolutely crucial difference. According to the H-D theory,
when the observational prediction turns out to be true, that confirms the hypothesis to
some degree. Popper denies that there is any such thing as confirmation. If, however,
the observational prediction turns out to be false, modus tollens can be used to
conclude deductively that some premise is false. If we are confident of the initial
conditions and auxiliary hypotheses, then we reject the hypothesis. The hypothesis
was a conjecture; the test provided a refutation. Hypotheses that are refuted must be
rejected.
If a bold hypothesis is subjected to severe testing and is not refuted, it is said to
be corroborated. Popper emphatically denies that corroboration is any brand of
confirmation. H-D theorists regard confirmation as a process that increases to some
degree the probability of the hypothesis and, by implication, the probability that the
hypothesis will yield correct predictions. Corroboration, in contrast, says nothing
whatever about the future predictive success of the hypothesis; it is, instead, a report
exclusively on the past performance of the hypothesis. The corroboration-rating is a
statement of the past success of the hypothesis as an explanatory theory. The corroboration report is not contaminated with any inductive elements.
Even if we were to grant Popper's dubious claim that theoretical science is
concerned only with explanation, and not with prediction, it would be necessary to
recognize that we use scientific knowledge in making practical decisions. If we wish
to put an artificial satellite into an orbit around the earth, we use Newtonian mechanics to compute the trajectory, and we confidently expect the satellite to perform as
predicted. An inductivist would claim that we base such expectations on the fact that,
within certain well-defined limits, Newtonian mechanics is a well-confirmed theory.
Popper maintains that, for purposes of practical prediction, using well-corroborated
theories is advisable, for nothing could be more rational.
The crucial question is, however, whether anything could be less rational than
to use the corroboration-rating of a theory as a basis for choosing it for predictive
purposes. Recalling that Popper has emphatically stated that the corroboration-rating
refers only to past performance, and not to future performance, the corroborationrating would seem to be totally irrelevant to the predictive virtues of the theory. The
use of highly corroborated theories for prediction has no greater claim to rationality
than do the predictions of fortune-tellers or sheer blind guessing. The price for
banishing all inductive elements from science is to render science useless for prediction and practical decision making (see Salmon 1981).
5. Pragmatic vindication. Reichenbach fully accepted Hume's conclusion
about the impossibility of proving that nature is uniform. He agreed that we have no
way of knowing whether past uniformities will extend into the future. He recognized
that, for all we can know, every inductive inference we make in the future may lead
64
to a false prediction. Nevertheless, he attempted to construct a practical decisiontheoretic justification for the use of induction.
Given our inability to know whether nature is uniform, we can consider what
happens in either case. Hume showed convincingly that, if nature is uniform, inductive reasoning will work very well, whereas, if nature is not uniform, inductive
reasoning will fail. This much is pretty easy to see. Reichenbach suggested, however,
that we should consider other options besides the use of induction for purposes of
trying to predict the future. Suppose we try consulting a crystal gaze to get our
predictions. We cannot say a priori that we will get correct predictions, even if nature
turns out to be uniform, but we cannot say a priori that we won't. We just don't know.
Let us set up a chart:
TABLE 2.1
We use induction
We don't use induction
Nature is uniform
Success
Success or Failure
The crucial entry is in the lower right-hand box. What if nature is not uniform and we
do not use induction? One possibility is simply not to make any predictions at all;
whether nature is uniform or not, that obviously does not result in successful predictions. Another possibility is that we adopt a noninductive method such as crystal
gazing. Any methodincluding wild guessingmay yield a true prediction once in
a while by chance, whether nature is uniform or not. But suppose that crystal gazing
were to work consistently. Then, that would be an important uniformity, and it could
be established inductivelythat is, on the basis of the observed record of the crystal
gazer in making successful predictions we could infer inductively that crystal gazing
will be successful in making correct predictions in the future. Thus, if crystal gazing
can produce consistent successful predictions so can the use of induction. What has
just been said about crystal gazing obviously applies to any noninductive method.
Reichenbach therefore concluded that if any method will succeed consistently, then
induction will succeed consistently. The same conclusion can be reformulated (by
contraposition) as follows: If induction does not work, then no other method will
work. We therefore have everything to gain and nothing to loseso far as predicting
the future is concernedby adopting the inductive method. No other method can
make an analogous claim. Reichenbach's argument is an attempt at vindication of
induction. He is trying to show thateven acknowledging Hume's skeptical
argumentsinduction is better suited to the goal of predicting the future than any
other methods that might be adopted.
Although Reichenbach's pragmatic justification may seem promising at first
glance, it does face serious difficulties on closer inspection. The greatest problem
with the foregoing formulation is that it suffers from severe vagueness. What do we
mean by speaking of the uniformity of nature? Nature is not completely uniform;
things do change. At the same timeup to the present at any ratenature has
exhibited certain kinds of uniformity. What degree of uniformity do we need in order
for the argument to succeed? We should be much more precise on this point. Likewise, when we spoke about noninductive methods we did not carefully survey all of
The Confirmation of Scientific Hypotheses
65
the available options. When the argument is tightened sufficiently, it turns out, it does
not vindicate just one rule of inductive inference; instead, it equally justifies an
infinite class of rules. Serious effortsup to this timeto find a satisfactory basis for
selecting a unique rule have been unsuccessful, (the technical details are discussed in
Salmon 1967, Chapter 6).
Where do things stand now250 years after the publication of Hume's Treatise
of Human Naturewith respect to the problem we have inherited from him? Although many ingenious attempts have been made to solve or dissolve it there is still
no consensus. It still stands as an item of "unfinished business" for philosophy of
science (see Salmon 1978a). The problem may, perhaps, best be summarized by a
passage from Hume himself:
Let the course of things be allowed hitherto ever so regular, that alone, without some new
argument or inference, proves not that for the future it will continue so. In vain do you
pretend to have learned the nature of bodies from your past experience. Their secret nature,
and consequently all their effects and influence, may change without any change in their
sensible qualities. This happens sometimes, and with regard to some objects: Why may it not
happen always, and with regard to all objects? What logic, what process or argument secures
you against this supposition? My practice, you say, refutes my doubts. But you mistake the
purport of my question. As an agent, I am quite satisfied in the point; but as a philosopher
. . . I want to learn the foundation of this inference. (1748, Section 4)
Our discussion up to this point has been carried on without the aid of a powerful
toolthe calculus of probability. The time has come to invoke it. The defects of the
qualitative approaches to confirmation discussed in Sections 2.3 and 2.4 suggest that
an adequate account of the confirmation of scientific statements must resort to quantitative or probabilistic methods. In support of this suggestion, recall that we have
already come across the concept of probability in the discussion of the qualitative
approaches. In our discussion of the H-D method, for instance, we encountered the
concept of probability in at least two ways. First, noting that a positive result of an
H-D test does not conclusively establish a hypothesis, we remarked that it might
render the hypothesis a little more probable than it was before the test. Second, in
dealing with the problem of statistical hypotheses, we saw that only probabilistic
observational predictions can be derived from such test hypotheses. In order to pursue
our investigation of the issues that have been raised we must take a closer look at the
concept or concepts of probability.
66
The modern theory of probability had its origins in the seventeenth century.
Legend has it that a famous gentleman, the Chevalier de Mere, posed some questions
about games of chance to the philosopher-mathematician Blaise Pascal. Pascal communicated the problems to the mathematician Pierre de Fermat, and that was how it
all began. Be that as it may, the serious study of mathematical probability theory
began around 1660, and Pascal and Fermat, along with Christian Huygens, played
crucial roles in that development, (for an historical account see Hacking 1975 and
Stigler 1986).
In order to introduce the theory of probability, we take probability to be a relationship between events of two different typesfor example, between tossing a standard die and getting a six, or drawing from a standard bridge deck and getting a king.
We designate probabilities by means of the following notation:
Pr(B\A) is the probability of a result of the type B given an event of the type A.
If A is a toss of a standard die and B is getting a three, then ' 'Pr{B IA)" stands for the
probability of getting a three if you toss a standard die. As the theory of probability
is seen today, all of the elementary rules of probability can be derived from a few
simple axioms. The meanings of these axioms and rules can be made intuitively
obvious by citing examples from games of chance that use such devices as cards and
dice. After some elementary features of the mathematical calculus of probability have
been introduced in this section, we look in the following section at a variety of
interpretations of probability that have been proposed.
AXIOMS (BASIC RULES)
Axiom (rule) 1: Every probability is a unique real number between zero and one
inclusive; that is,
O <Pr(B\A) < 1.
Axiom (rule) 2: If A logically entails B, then Pr(B\A) = 1.
Definition: Events of types B and C are mutually exclusive if it is impossible for both
B and C to happen on any given occasion. Thus, for example, on any draw from a
standard deck, drawing a heart and drawing a spade are mutually exclusive, for no
card is both a heart and a spade.
Axiom (rule) 3: If B and C are mutually exclusive, then
Pr(B V CIA) = Pr(B\A) + Pr(C\A).
This axiom is also known as the special addition rule.
Example: The probability of drawing a heart or a spade equals the probability of
drawing a heart plus the probability of drawing a spade.
Axiom (rule) 4: The probability of a joint occurrencethat is, of a conjunction of
B and Cis equal to the probability of the first multiplied by the probability of the
second given that the first has occurred:
Pr(B.C\A) = Pr(B\A) x Pr(C\A.B).
The Confirmation of Scientific Hypotheses
67
From the four axioms (basic rules) just stated, several other rules are easy to
derive that are extremely useful in calculating probabilities. First, we need a definition:
Definition: The events B and C are independent if and only if
Pr{C\A.B) = Pr(C\A).
When the events B and C are independent of one another, the multiplication rule
(axiom 4) takes on a very simple form:
Rule 5: If B and C are independent, given A, then
Pr(B.C\A) = Pr(B\A) X Pr{C\A).
This rule is known as the special multiplication rule. (Proofs, sketches of proofs,
and other technical items will be placed in boxes. They can be omitted on first
reading.)
Proof of Rule 5: Substitute Pr(C\A) for Pr(B.C\A) in Axiom 4.
Example: What is the probability of getting double 6 ("boxcars") when a standard
pair of dice is thrown? Since the outcomes on the two dice are independent, and the
probability of 6 on each die is 1/6, the probability of double 6 is
1/6 x 1/6 -
1/36
Example: What is the probability of drawing two spades on two consecutive draws
when the drawing is done with replacement? The probability of getting a spade on
the first draw is 13/52 = 1/4. After the first card is drawn, whether it is a spade or
not, it is put back in the deck and the deck is reshuffled. Then the second card is
drawn. Because of the replacement, the outcome of the second draw is independent
of the outcome of the first draw. Therefore, the probability of getting a spade on the
second draw is just the same as it was on the first draw. Thus, the probability of
getting two spades on two consecutive draws is
1/4 x 1/4 = 1/16
68
Pr(B\A).
Pr(B.C\A).
This is the general addition rule. Unlike Rule 3, this rule applies to outcomes B and
C even if they are not mutually exclusive.
Example: What is the probability of getting a spade or a face card in a draw from a
standard deck? These two alternatives are not mutually exclusive, for there are three
8
This example is closely related to one of the problems posed by the Chevalier de Mere\ How many
tosses of a pair of dice, he asked, are required to have at least afifty-fiftychance of getting at least one double
6? It seems that a common opinion among gamblers at the time was that 24 tosses would be sufficient. The
Chevalier doubted that answer, and it turned out that he was right. One needs 25 tosses to have at least a
fifty-fifty chance.
The Confirmation of Scientific Hypotheses
cardsking, queen, and jack of spadesthat are both face cards and spades. Since
there are 12 face cards and 13 spades, the probability of a spade or a face card is
12/52 + 13/52 - 3/52 = 22/52
It is easy to see why this rule has the form that it does. If B and C are not
mutually exclusive, then some outcomes may be both B and C. Any such items will
be counted twiceonce when we count the Z?s and again when we count the Cs. (In
the foregoing example, the king of spades is counted once as a face card and again
as a spade. The same goes for the queen and jack of spades.) Thus, we must subtract
the number of items that are both B and C, in order that they be counted only once.
How to prove Rule 7. First, we note that the class of things that are B
or C in the inclusive sense consists of those things that are B. C or ~ B. C
orB. C, where these latter three classes are mutually exclusive. Thus,
Rule 3 can be applied, giving
Pr(B V CIA) = Pr{B.C\A) + P r ( ~ B.CIA) +
Pr(B.~C\A).
Rule 4 is applied to each of the three terms on the right-hand side, and
then Rule 6 is used to get rid of the negations inside of the parentheses.
A bit of simple algebra yields Rule 7.
Rule 8: Pr(C\A) = Pr(B\A) x Pr(C\A.B) + Pr{~ B\A) X Pr{C\A~
B).
Pr{B\A) X Pr(C\A.B )
^
>Pr{B\A) X Pr(C\A.B)
_
" Pr(B\A) X Pr(C\A.B) + />r(~IA) X Pr(CIA.~fl)
provided that Pr(C\A) i= 0. The fact that these two forms are equivalent follows
immediately from the rule of total probability (Rule 8), which shows that the denominators of the right-hand sides are equal to one another.
Rule 9 is known as Bayes's rule; it has extremely important applications. For
purposes of illustration, however, let us go back to the trivial example of the frisbee
factory that was used to illustrate the rule of total probability.
Example: Suppose we have chosen a frisbee at random from the day's production
and it turns out to be defective. We did not see which machine produced it. What
is the probabilityPr(B\A . C)that it was produced by the new machine? Bayes's rule gives the answer:
0.8 x 0.01
0.8 X 0.01 + 0.2 X 0.02
0.008
0.012 = 2/3
The really important fact about Bayes's rule is that it tells us a great deal about
the confirmation of hypotheses. The frisbee example illustrates this point. We have a
frisbee produced at this factory (A) that turns out, on inspection, to be defective (C),
and we wonder whether it was produced (caused) by the new machine (B). In other
The Confirmation of Scientific Hypotheses
71
Proof of Bayes's rule: Bayes's rule has two forms as given above; we
show how to prove both. We begin by writing Rule 4 twice; in the
second case we interchange B and C.
Pr(B.C\A) = Pr(B\A) x Pr(C\A.B)
Pr(C.B\A) = Pr(C\A) X Pr{B\AJC)
Since the class B. C is obviously identical to the class C.B we can equate
the right-hand sides of the two equations:
Pr(C\A) X Pr(B\A.C) = Pr(B\A ) X Pr(C\A.B)
Assuming that P(C\A) = 0, we divide both sides by that quantity:
Pr(B\A.C)
Pr(B\A) X Pr(C\A.B)
Pr{C\A)
This is the first form. Using Rule 8, the rule of total probability, we
replace the denominator, yielding the second form:
Pr(B\A.C) =
Pr(B\A) X Pr(C\A.B)
Pr(B\A) X Pr(C\A.B) + Pr(~B\A)
Pr{C\A~B)
words, we are evaluating the hypothesis that the new machine produced this defective
frisbee. As we have just seen, the probability is 2/3.
Inasmuch as we are changing our viewpoint from talking about types of objects
and events A, B, C, . . . to talking about hypotheses, let us make a small change in
notation to help in the transition. Instead of using " A " to stand for the day's production of frisbees, we shall use "K" to stand for our background knowledge about
the situation in that factory. Instead of using " B " to stand for the frisbees produced
by the new machine B, we shall use " H " to stand for the hypothesis that a given
frisbee was produced by machine B. And instead of using ' ' C ' to stand for defective
frisbees, we shall use " " ' to stand for the evidence that the given frisbee is defective.
Now Bayes's rule reads as follows:
Rule 9: Pr{H\K.E) =
Pr(H\K) X Pr(E\K.H)
Pr(H\K) X Pr{E\K.H) + Pr{~H\K) X Pr(E\K.~H)
Changing the letters in the formula (always replacing the same old letter for the same
new letter) obviously makes no difference to the significance of the rule. If the axioms
72
are rewritten making the same changes in variables, Rule 9 would follow from them
in exactly the same way. And inasmuch as we are still talking about probabilities
albeit the probabilities of hypotheses instead of the probabilities of eventswe still
need the same rules.
We can now think of the probability expressions that occur in Bayes's rule in the
following terms:
Pr(H\K) is the prior probability of hypothesis H just on the basis of our background
knowledge K without taking into account the specific new evidence E. (In our
example, it is the probability that a given frisbee was produced by machine B.)
Pr(~H\K) is the prior probability that our hypothesis H is false. (In our example,
it is the probability that a given frisbee was produced by machine ~B.) Notice that
H and ~H must exhaust all of the possibilities.
By the negation rule (Rule 6), these two prior probabilities must add up to 1; hence,
if one of them is known the other can immediately be calculated.
Pr(E\K.H) is the probability that evidence E would obtain given the truth of hypothesis H in addition to our background knowledge K. (In our example, it is the
probability that a particular frisbee is defective, given that it was produced by
machine B.) This probability is known as a likelihood.
Pr(E\K. ~H) is the probability that evidence E would obtain if our hypothesis H is
false. (In our example, it is the probability that a particular frisbee is defective if it
was not produced by machine B.) This probability is also a likelihood.
The two likelihoodsin sharp contrast to the prior probabilitiesare independent of
one another. Given only the value of one of them, it is impossible to calculate the
value of the other.
Pr{E\K) is the probability that our evidence E would obtain, regardless of whether
hypothesis H is true or false. (In our example, it is the probability that a given
frisbee is defective, regardless of which machine produced it.) This probability is
often called the expectedness of the evidence.9
Pr(H\K . E) is the probability of our hypothesis, judged in terms of our background
knowledge K and the specific evidence E. It is known as the posterior probability.
This is the probability we are trying to ascertain. (In our example, it is the probability that the frisbee was produced by the new machine. Since the posterior
probability ofH is different from the prior probability of H, the fact that the frisbee
is defective is evidence relevant to that hypothesis.)
9
Expectedness is the opposite of surprisingness. If the expectedness of the evidence is small the
evidence is surprising. Since the expectedness occurs in the denominator of the fraction, the smaller the
expectedness, the greater the value of the fraction. Surprising evidence confirms hypotheses more than evidence
that is to be expected regardless of the hypothesis.
The Confirmation of Scientific Hypotheses
73
Notice that, although the likelihood of a defective product is twice as great for the old
machine (0.02) as for the new (0.01), the posterior probability that a defective
frisbee was produced by the new machine (2/3) is twice as great as the probability that
it was produced by the old one (1/3).
In Section 2.9 we return to the problem of assigning probabilities to hypotheses,
which is the main subject of this chapter.
2.8 THE MEANING OF PROBABILITY
74
Compare the coin example with the following from modern physics.
Suppose you have two helium-4 atoms in a box. Each one has a fiftyfifty chance of being in the left-hand side of the box at any given time.
What is the probability of both atoms being in the left-hand side at a
particular time? The answer is 1/3. Since the two atoms are in principle
indistinguishableunlike the coins, which are obviously distinguishablewe cannot regard atom # 1 in the left-hand side and atom # 2 in
the right-hand side as a case distinct from atom # 1 in the right-hand side
and atom # 2 in the left-hand side. Indeed, it does not even make sense
to talk about atom # 1 and atom # 2 since we have no way, even in
principle, of telling which is which.
Suppose, for example, that we examine a coin very carefully and find that it is
perfectly symmetrical. Any reason one might give to suppose it will come up heads
can be matched by an equally good reason to suppose it will land tails up. We say that
the two sides are equally possible, and we conclude that the probability of heads is
1/2. If, however, we toss the coin a large number of times and find that it lands heads
up in about 3/4 of all tosses and tails up in about 1/4 of all tosses, we do have good
reason to prefer one outcome to the other, so we would not declare them equally
possible. The basic idea behind the principle of indifference is this: when we have no
reason to consider one outcome more probable than another, we should not arbitrarily choose one outcome to favor over another. This seems like a sound principle
of probabilistic reasoning.
There is, however, a profound difficulty connected with the principle of indifference; its use can lead to outright inconsistency. The problem is that it can be
applied in different ways to the same situation, yielding incompatible values for a
particular probability. Again, consider an example, namely, the case of Joe, the
sloppy bartender. When a customer orders a 3:1 martini (3 parts of gin to 1 part of
dry vermouth), Joe may mix anything from a 2:1 to a 4:1 martini, and there is no
further information to tell us where in that range the mix may lie. According to the
principle of indifference, then, we may say that there is a fifty-fifty chance that the
mix will be between 2:1 and 3:1, and an equal chance that it will be between 3:1 and
4:1. Fair enough. But there is another way to look at the same situation. A 2:1 martini
contains 1/3 vermouth, and a 4:1 martini contains 1/5 vermouth. Since we have no
further information about the proportion of vermouth we can apply the principle of
indifference once more. Since 1/3 = 20/60 and 1/5 = 12/60, we can say that there
is a fifty-fifty chance that the proportion of vermouth is between 20/60 and 16/60 and
an equal chance that it is between 16/60 and 12/60. So far, so good?
Unfortunately, no. We have just contradicted ourselves. A 3:1 martini contains
25 percent vermouth, which is equal to 15/60, not 16/60. The principle of indifference
has told us both that there is a fifty-fifty chance that the proportion of vermouth is
between 20/60 and 16/60, and also that there is a fifty-fifty chance that it is between
75
20/60 and 15/60. The situation is shown graphically in Figure 2.2. As the graph
shows, the same result occurs for those who prefer their martinis drier; the numbers
are, however, not as easy to handle.
We must recall, at this point, our first axiom, which states, in part, that the
probability of a given outcome under specified conditions is a unique real number. As
we have just seen, the classical interpretation of probability does not furnish unique
results; we have just found two different probabilities for the same outcome. Thus, it
turns out, the classical interpretation is not an admissible interpretation of probability.
You might be tempted to think the case of the sloppy bartender is an isolated and
inconsequential fictitious example. Nothing could be farther from the truth. This
example illustrates a broad range of cases in which the principle of indifference leads
to contradiction. The source of the difficulty lies in the fact that we have two
quantitiesthe ratio of gin to vermouth and the proportion of vermouththat are
interdefinable; if you know one you can calculate the other. However, as Figure 2.2
clearly shows, the definitional relation is not linear; the graph is not a straight line. We
can state generally: Whenever there is a nonlinear definitional relationship between
two quantities, the principle of indifference can lead to a similar contradiction. To
convince yourself of this point, work out the details of another example. Suppose
there is a square piece of metal inside of a closed box. You cannot see it. But you are
told that its area is somewhere between 1 square inch and 4 square inches, but nothing
else is known about the area. First apply the principle of indifference to the area of
the square, and then apply it to the length of the side which is, of course, directly
6:1 h
Ratio
gin to
vermouth 5:1 f~
4:1
*^
Equal
intervals
3:1
2:1 >*- - 2 : 1 15/60
1:1
0:1
12/60
i
J
0/60
|
20/60
i
\k 4^
20/60
10/60
\ /
Unequal
intervals
30/60
40/60
.Fractional parts
vermouth in martini
Figure 2.2
The Confirmation of Scientific Hypotheses
ascertainable from the area. (For another example, involving a car on a racetrack, see
Salmon 1967, 66-67.)
Although the classical interpretation fails to provide a satisfactory basic definition of the probability concept, that does not mean that the idea of the ratio of
favorable to equiprobable possible outcomes is useless. The trouble lies with the
principle of indifference, and its aim of transforming ignorance of probabilities into
values of probabilities. However, in situations in which we have positive knowledge
that we are dealing with alternatives that have equal probabilities, the strategy of
counting equiprobable favorable cases and forming the ratio of favorable to equiprobable possible cases is often handy for facilitating computations.
2. The frequency interpretation. The frequency interpretation has a venerable history, going all the way back to Aristotle (4th century B.C.), who said that the
probable is that which happens often. It was first elaborated with precision and in
detail by the English logician John Venn (1866, [1888] 1962). The basic idea is easily
illustrated. Consider an ordinary coin that is being flipped in the standard way. As it
is flipped repeatedly a sequence of outcomes is generated:
H T H T T T H H T T H T T T T H T H T T T H H H H . . . 10
We can associate with this sequence of results a sequence of relative frequencies
that is, the proportion of tosses that have resulted in heads up to a given point in the
sequenceas follows:
1/1, 1/2, 2/3, 2/4, 2/5, 2/6, 3/7, 4/8, 4/9, 4/10, 5/11, 5/12, 5/13/ 5/14, 5/15,
6/16, 6/17, 7/18, 7/19, 7/20, 7/21, 8/22, 9/23, 10/24, 11/25, . . .
The denominator in each fraction represents the number of tosses made up to that
point; the numerator represents the number of heads up to that point. We could, of
course, continue flipping the coin, recording the results, and tabulating the associated
relative frequencies. We are reasonably convinced that this coin is fair and that it was
flipped in an unbiased manner. Thus, we believe that the probability of heads is 1/2.
If that belief is correct, then, as the number of tosses increases, the relative frequencies will become and remain close to 1/2. The situation is shown graphically in Figure
2.3. There is no particular number of tosses at which the fraction of heads is and
remains precisely 1/2; indeed, in an odd number of tosses the ratio cannot possibly
equal 1/2. Moreover, if, at some point in the sequence, the relative frequency does
equal precisely 1/2, it will necessarily differ from that value on the next flip. Instead
of saying that the relative frequency must equal 1/2 in any particular number of
throws, we say that it approaches 1/2 in the long run.
Although we know that no coin can ever be flipped an infinite number of times,
it is useful, as a mathematical idealization, to think in terms of a. potentially infinite
sequence of tosses. That is, we imagine that, no matter how many throws have been
10
These are the results of 25 flips made in an actual trial by the authors.
77
1
Relative
frequency
of heads
1/2
0
Figure 2.3
made, it is still possible to make more; that is, there is no particular finite number N
at which point the sequence of tosses is considered complete. Then we can say that
the limit of the sequence of relative frequencies equals the probability; this is the
meaning of the statement that the probability of a particular sort of occurrence is, by
definition, its long run relative frequency.
What is the meaning of the phrase "limit of the relative frequency"? lMfl,f2,
/ 3 , . . . be the successive terms of the sequence of relative frequencies. In the example above,/! = l , / 2 = 1/2, f3 = 2/3, and so on. Suppose that/? is the limit of
the relative frequency. This means that the values of/ become and remain arbitrarily
close to p as n becomes larger and larger. More precisely, let 8 be any small number
greater than 0. Then, there exists some finite integer N such that, for any n > N9fn
does not differ from p by more than 8.
Many objections have been lodged against the frequency interpretation of probability. One of the least significant is that mentioned above, namely, the finitude of
all actual sequences of events, at least within the scope of human experience. The
reason this does not carry much weight is the fact that science is full of similar sorts
of idealizations. In applying geometry to the physical world we deal with ideal
straight lines and perfect circles. In using the infinitesimal calculus we assume that
certain quantitiessuch as electric chargecan vary continuously, when we know
that they are actually discrete. Such practices carry no danger provided we are clearly
aware of the idealizations we are using. Dealing with infinite sequences is technically
easier than dealing with finite sequences having huge numbers of members.
A much more serious problem arises when we ask how we are supposed to
ascertain the values of these limiting frequencies. It seems that we observe some
limited portion of such a sequence and then extrapolate on the basis of what has been
observed. We may not want to judge the probability of heads for a certain coin on the
basis of 25 flips, but we might well be willing to do so on the basis of several hundred.
Nevertheless, there are several logical problems with this procedure. First, no matter
how many flips we have observed, it is always possible for a long run of heads to
occur that would raise the relative frequency of heads well above 1/2. Similarly, a
long run of future tails could reduce the relative frequency far below 1/2.
Another way to see the same point is this. Suppose that, for each n, mln is the
fraction of heads to tosses as of the nth toss. Suppose also that/ does have the
78
limiting value p. Let a and b be any two fixed positive integers where a ^ b. If we
add the constant a to every value of m and the constant b to every value of n, the
resulting sequence (m + a)i(n + b) will converge to the very same value p. That
means that you could attach any sequence of b tosses, a of which are heads, to the
beginning of your sequence, without changing the limiting value of the relative
frequency. Moreover, you can chop off any finite number b of members, a of which
are heads, from the beginning of your sequence without changing the limiting frequency/'. As m and n get very large, the addition or subtraction of fixed numbers a
and b has less and less effect on the value of the fraction. This seems to mean that the
observed relative frequency in any finite sample is irrelevant to the limiting frequency. How, then, are we supposed to find out what these limiting frequencies
probabilitiesare?
It would seem that things could not get much worse for the frequency interpretation of probability, but they do. For any sequence, such as our sequence of coin
tosses, there is no guarantee that any limit of the relative frequency even exists. It is
logically possible that long runs of heads followed by longer runs of tails followed by
still longer runs of heads, and so on, might make the relative frequency of heads
fluctuate between widely separated extremes throughout the infinite remainder of the
sequence. If no limit exists there is no such thing as the probability of a head when
this coin is tossed.
In spite of these difficulties, the frequency concept of probability seems to be
used widely in the sciences. In Chapter 1, for instance, we mentioned the spontaneous decay of C14 atoms, commenting that the half-life is 5730 years. That is
the rate at which atoms of this type have decayed in the past; we confidently predict
that they will continue to do so. The relative frequency of disintegration of C14
atoms within 5730 years is 1/2. This type of example is of considerable interest to
archaeologists, physicists, and geophysicists. In the biological sciences it has been
noted, for example, that there is a very stable excess of human male births over
human female births, and that is expected to continue. Social scientists note, however, that human females, on average, live longer than human males. This frequency is also extrapolated.
It is easy to prove that the frequency interpretation satisfies the axioms of
probability laid down in the preceding section. This interpretation is, therefore,
admissible. Its main difficulty lies in the area of ascertainability. How are we to
establish values of probabilities of this sort? This question again raises Hume's
problem of justification of induction.
A further problem remains. Probabilities of the frequency variety are used in
two ways. On the one hand, they appear in statistical laws, such as the law of
radioactive decay of unstable species of nuclei. On the other hand, they are often
applied in making predictions of single events, or finite classes of events. Pollsters,
for example, predict outcomes of single elections on the basis of interviews with
samples of voters. If, however, probability is defined as a limiting frequency in a
potentially infinite sequence of events, it does not seem to make any sense to talk
about probabilities of single occurrences. The problem of the single case raises a
problem about the applicability of the frequency interpretation of probability.
Before we leave the frequency interpretation a word of caution is in order. The
The Confirmation of Scientific Hypotheses
79
frequency interpretation and the classical interpretations are completely different from
one another, and they should not be confused. When the classical interpretation refers
to possible outcomes and favorable outcomes it is referring to types or classes of
eventsfor example, the class of all cases in which heads comes up is one possible
outcome; the class of cases in which tails comes up is one other possible outcome. In
this example there are only two possible outcomes. These classesnot their
membersare what you count for purposes of the classical interpretation. In the
frequency interpretation, it is the members of these classes that are counted. If
the coin is tossed a large number of times there are many heads and many tails. In the
frequency interpretation, the numbers of items of which ratios are formed keep
changing as the number of individual events increases. In the classical interpretation,
the probability does not depend in any way on how many heads or tails actually occur.
3. The propensity interpretation. The propensity interpretation is a relatively
recent innovation in the theory of probability. Although suggested earlier, particularly
by Charles Saunders Peirce, it was first clearly articulated by Popper (1957b, 1960).
It was introduced specifically to deal with the problem of the single case.
The sort of situation Popper originally envisaged was a potentially infinite
sequence of tosses of a loaded die that was biased in such a way that side 6 had a
probability of 1/4. The limiting frequency of 6 in this sequence is, of course, 1/4.
Suppose, however, that three of the tosses were not made with the biased die, but
rather with a fair die. Whatever the outcomes of these three throws, they would have
no effect on the limiting frequency. Nevertheless, Popper maintained, we surely want
to say that the probability of 6 on those three tosses was 1/6not 1/4. Popper argued
that the appropriate way to deal with such cases is to associate the probability with the
chance setup that produces the outcome, rather than to define it in terms of the
sequence of outcomes themselves. Thus, he claims, each time the fair die is thrown,
the mechanismconsisting of the die and the throwerhas a causal tendency or
propensity of 1/6 to produce the outcome 6. Similarly, each time the loaded die is
tossed, the mechanism has a propensity of 1/4 to produce the outcome 6.
Although this idea of propensityprobabilistic causal tendencyis important
and valuable, it does not provide an admissible interpretation of the probability
calculus. This can easily be seen in terms of the case of the frisbee factory introduced
in the preceding section. That example, we recall, consisted of two machines, each
of which had a certain propensity or tendency to produce defective frisbees. For the
new machine the propensity was 0.01; for the old machine it was 0.02. Using the rule
of total probability we calculated the propensity of the factory to produce faulty
frisbees; it was 0.012. So far, so good.
The problem arises in connection with Bayes's rule. Having picked a defective frisbee at random from the day's production, we asked for the probability that
it was produced by the new machine; the answer was 2/3. This is a perfectly legitimate probability, but it cannot be construed as a propensity. It makes no sense
to say that this frisbee has a propensity of 2/3 to have been produced by the new
machine. Either it was produced by the new machine or by the old. It does not
have a tendency of 1/3 to have been produced by the old machine and a tendency
of 2/3 to have been produced by the new one. The basic point is that causes pre80
cede their effects and causes produce their effects, even if the causal relationship
has probabilistic aspects. We can speak meaningfully of the causal tendency of a
machine to produce a faulty product. Effects do not produce their causes. It does
not make sense to talk about the causal tendency of the effect to have been produced by one cause or another.
Bayes's rule enables us to compute what are sometimes called inverse probabilities,.Whereas the rule of total probability enables us to calculate the forward
probability of an effect, given suitable information about antecedent causal factors,
Bayes's rule allows us to compute the inverse probability that a given effect was
produced by a particular cause. These inverse probabilities are an integral part of the
mathematical calculus of probability, but no propensities correspond to them. For this
reason the propensity interpretation is not an admissible interpretation of the probability calculus.
4. The subjective interpretation. Both the frequency interpretation and the
propensity interpretation are regarded by their proponents as types of physical probabilities. They are objective features of the real world. But probability seems to many
philosophers and mathematicians to have a subjective side as well. This aspect has
something to do with the degree of conviction with which an individual believes in
one proposition or another. For instance, Mary Smith is sure that it will be cold in
Montana next winterthat is, in some place in that state the temperature will fall
below 50 degrees Fahrenheit between 21 December and 21 March. Her subjective
probability for this event is extremely close to 1. Also, she disbelieves completely
that Antarctica will be hot any time during its summerthat is, she is sure that the
temperature will not rise above 100 degrees Fahrenheit between 21 December and 21
March. Her subjective probability for real heat in Antarctica in summer is very close
to 0. She neither believes in rain in Pittsburgh tomorrow, nor disbelieves in rain in
Pittsburgh tomorrow; her conviction for either one of these alternatives is just as
strong as for the other. Her subjective probability for rain tomorrow in Pittsburgh is
just about 1/2. As she runs through the various propositions in which she might
believe or disbelieve she finds a range of degrees of conviction spanning the whole
scale from 0 to 1. Other people will, of course, have different degrees of conviction
in these same propositions.
It is easy to see immediately that subjective degrees of commitment do not
provide an admissible interpretation of the probability calculus. Take a simple example. Many people believe that the probability of getting a 6 with a fair die is 1/6,
and that the outcomes of successive tosses are independent of one another. They also
believe that we have afifty-fiftychance of getting 6 at least once in three throws. As
we saw in the previous section, however, that probability is significantly below 1/2.
Therefore, the preceding set of degrees of conviction violate the mathematical calculus of probability. Of course, not everyone makes that particular mistake, but
extensive empirical research has shown that most of us do make various kinds of
mistakes in dealing with probabilities. In general, a given individual's degrees of
conviction fail to satisfy the mathematical calculus.
5. Personal probabilities. What if there were a person whose degrees of
conviction did not violate the probability calculus? That person's subjective probaThe Confirmation of Scientific Hypotheses
81
rnmK.E,
PrW[K)
PrimKH)
Pr(E\K-H)
= 4/103 0.04.
2/3 - 0.67.
83
1/2 x 1
:
- 4/5 = 0.80.
1 / 2 X 1 + 1/2 x 1/4
After ten heads, he has
1/2 X 1
1 / 2 x 1 + 1/2 x 1/1024
These calculations show two things. First, they show how Bayes's rule can be
used to ascertain the probability of a hypothesis if we have values for the prior
probabilities. If we employ personal probabilities the prior probabilities become
available. They are simply a person's degrees of conviction in the hypothesis prior to
receipt of the observational evidence. In this kind of example the likelihoods can be
calculated from assumptions we share concerning the behavior of fair and two-headed
coins.
Second, these calculations illustrate a phenomenon known as washing out of the
priors or swamping of the priors. Notice that we did two sets of calculationsone for
John and one for Wes. We started with widely divergent degrees of conviction in the
hypothesis; Wes's was 1/2 and John's was 1/100. As the evidence accumulated our
degrees of conviction became closer and closer. After ten heads, Wes's degree of
conviction is approximately 0.99 and John's is approximately 0.91. As more heads
occur our agreement becomes even stronger. This illustrates a general feature of
Bayes's rule. Suppose there are two people with differing prior probabilitiesas far
apart as you like provided neither has an extreme value of 0 or 1. Then, if they agree
on the likelihoods and if they share the same observational evidence, their posterior
probabilities will get closer and closer together as the evidence accumulates. The
influence of the prior probabilities on the posterior probabilities decreases as more
evidence becomes available. This phenomenon of washing out of the priors should
help to ease the worry we might have about appealing to admittedly subjective
degrees of conviction in our evaluations of scientific hypotheses.
Still, profound problems are associated with the personalistic interpretation of
probability. The only restriction imposed by this interpretation on the values of probabilities is that they be coherentthat they satisfy the rules of mathematical probability. This is a very weak constraint. If we look at the rules of probability we note that,
with a couple of trivial exceptions, the mathematical calculus of probability does not
by itself furnish us with any values of probabilities. The exceptions are that a logically
necessary proposition must have probability 1 and a contradiction must have probability 0. In all other cases, the rules of probability enable us to calculate some probability values from others. You plug in some probability values, turn the crank, and
others come out. This means that there need be little contact between our personal
probabilities and what goes on in the external world. For example, it is possible for a
person to have a degree of conviction of 9/10 that the next toss of a coin will result in
heads even though the coin has been tossed hundreds of times and has come up tails
on the vast majority of these tosses. By suitably adjusting one's other probabilities one
can have such personal probabilities as these without becoming incoherent. If our probabilities are to represent reasonable degrees of conviction some stronger restrictions
surely appear to be needed.
84
Fa.Fb.Fc
Fa.Fb.~Fc
Fa.-Fb.Fc
-Fa.Fb.Fc
5. Fa.~Fb.~Fc
6. ~Fa.Fb.~~Fc
7. ~Fa.~Fb.Fc
8. ~Fa.~Fb.~Fc
Any consistent statement that we can form in this miniature language can be expressed by means of these state descriptions. For example, (x)Fx, which says that
every ball is red, is equivalent to state description 1. The statement (3x)Fx, which
says that at least one ball is red, is equivalent to the disjunction of state descriptions
1-7; that is, it says that either state description 1 or 2 or 3 or 4 or 5 or 6 or 7 is true.
Fa is equivalent to the disjunction of state descriptions 1, 2, 3, and 5. Fa.Fb is
equivalent to the disjunction of state descriptions 1 and 2. If we agree to admitjust
for the sake of conveniencethat there can be disjunctions with only one term, we
The Confirmation of Scientific Hypotheses
85
can say that every consistent statement is equivalent to some disjunction of state
descriptions. The state descriptions in any such disjunction constitute the range of
that statement. A contradictory statement is equivalent to the denial of all eight of the
state descriptions. Its range is empty.
In the following discussion, H is any statement that is being taken as a hypothesis and E any statement that is being taken as evidence. In this discussion any
consistent statement that can be formulated in our language can serve as a statement
of evidence, and any statementconsistent or inconsistentcan serve as a hypothesis. Now, consider the hypothesis (Bx)Fx and evidence Fc. Clearly this evidence deductively entails this hypothesis; if the third ball is red at least one must
be red. If we look at the ranges of this evidence and this hypothesis, we see that
the range of Fc (state descriptions 1, 3, 4, 7) is entirely included in the range of
(3x)Fx (state descriptions 1-7). This situation always holds. If one statement entails another, the range of the first is included within the range of the second. This
means that every possible state of the universe in which the first is true is a possible
state of the universe in which the second is true. If two statements have identical
ranges, they are logically equivalent, and each one entails the other. If two statements are logically incompatible with one another, their ranges do not overlap at
allthat is, there is no possible state of the universe in which they can both be
true. We see, then, that deductive relationships can be represented as relationships
among the ranges of the statements involved.
Let us now turn to inductive relationships. Consider the hypothesis (x)Fx and
the evidence Fa. This evidence obviously does not entail the hypothesis, but it seems
reasonable to suppose that it provides some degree of inductive support or confirmation. The range of the evidence (1, 2, 3, 5) is not completely included in the range
of the hypothesis (1), but it does overlap that rangethe two ranges have state
description 1 in common. What we need is a way of expressing the idea of confirmation in terms of the overlapping of ranges. When we take any statement E as
evidence, we are accepting it as true; in so doing we are ruling out all possible states
of the universe that are incompatible with the evidence E. Having ruled out all of
those, we want to know to what degree the possible states in which the evidence holds
true are possible states in which the hypothesis also holds true. This can be expressed
in the form of a ratio, range (Zs./fVrange (), and this is the basic idea behind the
concept of degree of confirmation.
Consider the range of (x)Fx; this hypothesis holds in one state description out
of eight. If, however, we learn that Fa is true, we rule out four of the state descriptions, leaving only four as possibilities. Now the hypothesis holds in one out of four.
If we now discover that Fb is also true, our combined evidence Fa.Fb holds in only
two state descriptions, and our hypothesis holds in one of the two. It looks reasonable
to say that our hypothesis had a probability of 1/8 on the basis of no evidence, a
probability of 1/4 on the basis of the first bit of evidence, and a probability of 1/2 on
the two pieces of evidence. (This suggestion was offered by Wittgenstein 1922). But
appearances are deceiving in this case.
If we were to adopt this suggestion as it stands, Carnap realized, we would
rule out altogether the possibility of learning from experience; we would have no
86
basis at all for predicting future occurrences. Consider, instead of (x)Fx, the hypothesis Fc. By itself, this hypothesis holds in four (1, 3, 4, 7) out of eight state
descriptions. Suppose we find as evidence that Fa. The range of this evidence is
four state descriptions (1, 2, 3, 5), and the hypothesis holds in two of them. But
4/8 = 2/4, so the evidence has done nothing to support the hypothesis. Moreover,
if we learn that Fb is true our new evidence is Fa.Fb, which holds in two state
descriptions (1, 2), and our hypothesis holds in only one of them, giving us a ratio
of 1/2. Hence, according to this way of defining confirmation, what we observe in
the past and present has no bearing on what will occur in the future. This is an
unacceptable consequence. When we examined the hypothesis (x)Fx in the preceding paragraph we appeared to be achieving genuine confirmation, but that was
not happening at all. The hypothesis (x)Fx simply states that a, b, and c all have
property F. When we find out by observing the first ball that it is red, we have
simply reduced the predictive content of h. At first it predicted the color of three
balls; after we examine the first ball it predicts the color of only two balls. After
we observe the second ball, the hypothesis predicts the color of only one ball. If
we were to examine the third ball and find it to be red, our hypothesis would have
no predictive content at all. Instead of confirming our hypothesis we were actually
simply reducing its predictive import.
In order to get around the foregoing difficulty, Carnap proposed a different way
of evaluating the ranges of statements. The method adopted by Wittgenstein amounts
to assigning equal weights to all of the state descriptions. Carnap suggested assigning
unequal weights on the following basis. Let us take another look at our list of state
descriptions in Table 2.2:
TABLE 2.2
State Description
Weight
Structure Description
Weight
1. Fa.Fb.Fc
1/4
A11F
1/4
2.Fa.Fb.~Fc
3.Fa.~Fb.Fc
4. ~Fa.Fb.Fc
1/12
1/12
1/12
2 F, 1 ~ F
1/4
5.Fa.~Fb~Fc
6. ~Fa.Fb.~Fc
1. ~Fa.~Fb.Fc
1/12
1/12
1/12
1 F, 2 ~F
1/4
1/4
NoF
1/4
8.
~Fa.-Fb.~Fc
Carnap noticed that state descriptions 2, 3, and 4 make similar statements about
our miniature universe; they say that two entities have property F and one lacks it.
Taken together, they describe a certain structure. They differ from one another in
identifying the ball that is not red, but Carnap suggests that that is a secondary
consideration. Similarly, state descriptions 5, 6, and 7, taken together describe a
certain structure, namely, a universe in which one individual has property F and two
lack it. Again, they differ in identifying the object that has this property. In contrast,
87
state description 1, all by itself, describes a particular structure, namely, all threeentities have property F. Similarly, state description 8 describes the structure in which
no object has that property.
Having identified the structure descriptions, Carnap proceeds to assign equal
weights to them (each gets 1/4); he then assigns equal weights to the state descriptions within each structure description. The resulting system of weights is
shown above. These weights are then used as a measure of the ranges of statements;11 this system of measures is called m*. A confirmation function c* is defined as follows:12
c*(H\E) = m*(H.E)/m*(E).
To see how it works, let us reconsider the hypothesis Fc in the light of different
bits of evidence. First, the range of Fc consists of state description 1, which has
weight 1/4, and 3,4, and 7, each of which has weight 1/12. The sum of all of them
is 1/2; that is, the probability of our hypothesis before we have any evidence. Now,
we find that Fa; its measure is 1/2. The range of Fa.Fc is state descriptions 1 and
3, whose weights are, respectively, 1/4 and 1/12, for a total of 1/3. We can now
calculate the degree of confirmation of our hypothesis on this evidence:
c*(H\E) = m*(E.H)/m*(E) = 1/3 + 1/2 = 2/3.
Carrying out the same sort of calculation for evidence Fa.Fb we find that our hypothesis has degree of confirmation 3/4. If, however, our first bit of evidence had
been Fa, the degree of confirmation of our hypothesis would have been 1/3. If
our second bit of evidence had been ~~Fb, that would have reduced its degree of
confirmation to 1/4. The confirmation function c* seems to do the right sorts of
things. When the evidence is what we normally consider to be positive, the degree
of confirmation goes up. When the evidence is what we usually take to be negative,
the degree of confirmation goes down. Clearly, c* allows for learning from experience.
A serious philosophical problem arises, however. Once we start playing the
game of assigning weights to state descriptions, we face a huge plethora of possibilities. In setting up the machinery of state descriptions and weights, Carnap
demands only that the weights for all of the state descriptions add up to 1, and that
each state description have a weight greater than 0. These conditions are sufficient
to guarantee an admissible interpretation of the probability calculus. Carnap recognized the obvious fact that infinitely many confirmation functions satisfying this
basic requirement are possible. The question is how to make an appropriate choice.
It can easily be shown that choosing a confirmation function is precisely the same
as assigning prior probabilities to all of the hypotheses that can be stated in the
given language.
Consider the following possibility for a measure function:
11
The measure of the range of any statement H can be identified with the prior probability of that
statement in the absence of any background knowledge K. It is an a priori prior probability.
12
Wittgenstein's measure function assigns the weight Vfe to each state description; the confirmation
function based upon it is designated c|.
88
TABLE 2.3
State Description
Weight
Structure Description
Weight
1. Fa.Fb.Fc
1/20
A11F
1/20
2.
3.
4.
3/20
3/20
3/20
2 F, 1 -F
9/20
5.
Fa.-Fb.-Fc
6. Fa.Fb.~Fc
1.
-Fa-Fb.Fc
3/20
3/20
3/20
\F,2~F
9/20
8.
1/20
NoF
1/20
Fa.Fb-Fc
Fa.-Fb.Fc
-Fa.Fb.Fc
~Fa.-Fb.-Fc
(The idea of a confirmation function of this type was given in Burks 1953; the
philosophical issues are further discussed in Burks 1977, Chapter 3.) This method of
weighting, which may be designated m , yields a confirmation function C, which
is a sort of counterinductive method. Whereas m* places higher weights on the first
and last state descriptions, which are state descriptions for universes with a great deal
of uniformity (either every object has the property, or none has it), m places lower
weights on descriptions of uniform universes. Like c*, c allows for ''learning from
experience,'' but it is a funny kind of anti-inductive ' 'learning.'' Before we reject m
out of hand, however, we should ask ourselves if we have any a priori guarantee that
our universe is uniform. Can we select a suitable confirmation function without being
totally arbitrary about it? This is the basic problem with the logical interpretation of
probability.
We now turn to the task of illustrating how the probabilistic apparatus developed
above can be used to illuminate various issues concerning the confirmation of scientific statements. Bayes's theorem (Rule 9) will appear again and again in these
illustrations, justifying the appellation of Bayesian confirmation theory.
Various ways are available to connect the probabilistic concept of confirmation
back to the qualitative concept, but perhaps the most widely followed route utilizes
an incremental notion of confirmation: E confirms H relative to the background
knowledge K just in case the addition of to A' raises the probability of//, that is,
Pr(H\E.K) > ) Pr(H\K).13 Hempel's study of instance confirmation in terms of a
13
Sometimes, when we say that a hypothesis has been confirmed, we mean that it has been rendered
highly probable by the evidence. This is a high probability or absolute concept of confirmation, and it should
be carefully distinguished from the incremental concept now under discussion (see Carnap 1962, Salmon 1973,
and Salmon 1975). Salmon (1973) is the most elementary discussion.
The Confirmation of Scientific Hypotheses
89
two-place relation can be taken to be directed at the special case where K contains no
information. Alternatively, we can suppose that K has been absorbed into the probability function in the sense that Pr(K) = l, 14 in which case the condition for
incremental confirmation reduces to Pr(H\E) > Pr{H). (The unconditional probability Pr(H) can be understood as the conditional probability Pr(H\T), where T is a
vacuous statement, for example, a tautology. The axioms of Section 2.7 apply only
to conditional probabilities.)
It is easy to see that on the incremental version of confirmation, Hempel's
consistency condition is violated as is
Conjunction condition: If E confirms H and also H' then E confirms H.H'.
It takes a bit more work to construct a counterexample to the special consequence
condition. (This example is taken from Carnap 1950 and Salmon 1975, the latter of
which contains a detailed discussion of Hempel's adequacy conditions in the light of
the incremental notion of confirmation.) Towards this end take the background knowledge to contain the following information. Ten players participate in a chess tournament in Pittsburgh; some are locals, some are from out of town; some are juniors,
some are seniors; and some are men (Af), some are women (W). Their distribution is
given by
TABLE 2.4
Locals
Out-of-towners
Juniors
M, W, W
M,M
Seniors
M, M
W, W, W
And finally, each player initially has an equal chance of winning. Now consider the
hypotheses H: an out-of-towner wins, and H'\ a senior wins, and the evidence E: a
woman wins. We find that
Pr(H\E) = 3/5 > Pr(H) = 1/2
so E confirms H. But
Pr{HvH'\E)
So E does not confirm H V H'; in fact E confirms ~(H v H') and so disconfirms
HvH' even though H V H' is a consequence of H.
The upshot is that on the incremental conception of confirmation, Hempel's
adequacy conditions and, hence, his definition of qualitative confirmation, are inadequate. However, his adequacy conditions fare better on the high probability conception of confirmation according to which E confirms H relative to K just in case
Pr{H\E.K) > r, where r is some number greater than 0.5. But this notion of
14
As would be the case if learning from experience is modeled as change of probability function through
conditionalization; that is, when K is learned, Prold is placed by Prnew ( ) = Prold ( | K). From this point of
view, Bayes's theorem (Rule 9) describes how probability changes when a new fact is learned.
90
confirmation cannot be what Hempel has in mind; for he wants to say that the
observation of a single black raven (E) confirms the hypothesis that all ravens are
black (//), although for typical K, Pr{H\E.K) will surely not be as great as 0.5. Thus,
in what follows we continue to work with the incremental concept.
The probabilistic approach to confirmation coupled with a simple application of
Bayes's theorem also serves to reveal a kernel of truth in the H-D method. Suppose
that the following conditions hold:
(i) H, K^E;
Condition (i) is the basic H-D condition. Conditions (ii) and (iii) say that neither//
nor E is known on the basis of the background information K to be almost surely false
or almost surely true. Then on the incremental conception it follows, as the H-D
methodology would have it, that E confirms H on the basis of K. By Bayes's theorem
Pr{H\K)
since by (i),
Pr(E\H.K) = 1.
It then follows from (ii) and (iii) that
Pr(H\E.K) > Pr{H\K).
Notice also that the smaller Pr{E\K) is, the greater the incremental confirmation
afforded by E. This helps to ground the intuition that "surprising" evidence gives
better confirmational value. However, this observation is really double-edged as will
be seen in Section 2.10.
The Bayesian analysis also affords a means of handling a disquieting feature of
the H-D method, sometimes called the problem of irrelevant conjunction. If the H-D
condition (i) holds for //, then it also holds for H.X where X is anything you like,
including conjuncts to which E is intuitively irrelevant. In one sense the problem is
mirrored in the Bayesian approach, for assuming that 1 > Pr(H.X\K) > 0, it follows
that E incrementally confirms H.X. But since the special consequence condition does
not hold in the Bayesian approach, we cannot infer that E confirms the consequence
X of H.X. Moreover, under the H-D condition (i), the incremental confirmation of a
hypothesis is directly proportional to its prior probability. Since
Pr{H\K) > Pr(H.X\K), with strict inequality holding in typical cases, the incremental confirmation for H will be greater than for H.X.
Bayesian methods are flexible enough to overcome various of the shortcomings
of Hempel's account. Nothing, for example, prevents the explication of confirmation
in terms of a ZV-function which allows observational evidence to boost the probability
of theoretical hypotheses. In addition the Bayesian approach illuminates the paradoxes of the ravens and Goodman's paradox.
In the case of the ravens paradox we may grant that the evidence that the
individual a is a piece of white chalk can confirm the hypothesis that "All ravens are
black" since, to put it crudely, this evidence exhausts part of the content of the
The Confirmation of Scientific Hypotheses
91
hypothesis. Nevertheless, as Suppes (1966) has noted, if we are interested in subjecting the hypothesis to a sharp test, it may be preferable to do outdoor ornithology
and sample from the class of ravens rather than sampling from the class of nonblack
things. Let a denote a randomly chosen object and let
Pr(Ra.Ba) = pu
Pr(-Ra.Ba) = p39
Pr(Ra.-Ba) = p2
Pr(~Ra.~Ba) = p4.
Then
Pr{~Ba\Ra) = p2 + (px + p2)
Pr(Ral-Ba) = p2(p2 + p4)
Thus, Pr(~Ba\Ra) > Pr{Ra\~Bd) just in case p4 > px. In our world it certainly
seems true that/?4 > px. Thus, Suppes concludes that sampling ravens is more likely
to produce a counterinstance to the ravens hypothesis than is sampling the class of
nonblack things.
There are two problems here. The first is that it is not clear how the last
statement follows since a was supposed to be an object drawn at random from the
universe at large. With that understanding, how does it follow that Pr{~Ba \Ra) is the
probability that an object drawn at random from the class of ravens is nonblack?
Second, it is the anti-inductivists such as Popper (see item 4 in Section 2.8 above and
2.10 below) who are concerned with attempts to falsify hypotheses. It would seem
that the Bayesian should concentrate on strategies that enhance absolute and incremental probabilities. An approach due to Gaifman (1979) and Horwich (1982) combines both of these points.
Let us make it part of the background information K that a is an object drawn
at random from the class of ravens while b is an object drawn at random from the class
of nonblack things. Then an application of Bayes's theorem shows that
Pr(HlRa.Ba.K) >
Pr(H\~Rb.~Bb.K)
just in case
1 > Pr(~Rb\K) > Pr(Ba\K).
To explore the meaning of the latter inequality, use the principle of total probability
to find that
Pr(Ba\K) = Pr(Ba\H.K) Pr(H\K) + Pr(Ba\~H.K) Pr(~H\K)
= Pr(H\K) + Pr(Ba\~H.K) Pr(~H\K)
and that
Pr(~Rb\K) - Pr(H\K) + Pr(~Rb\~H.K)
Pr(~H\K).
> Pr(Ba\~H.K),
or
Pr(~Ba\~H.K)
92
which is presumably true in our universe. For supposing that some ravens are nonblack, a random sample from the class of ravens is more apt to produce such a bird
than is a random sample from the class of nonblack things since the class of nonblack
things is much larger than the class of ravens. Thus, under the assumption of the
stated sampling procedures, the evidence Ra.Ba does raise the probability of the
ravens hypothesis more than the evidence ~Rb.~Bb does. The reason for this is
precisely the differential propensities of the two sampling procedures to produce
counterexamples, as Suppes originally suggested.
The Bayesian analysis also casts light on the problems of induction, old and
new, Humean and Goodmanian. Russell (1948) formulated two categories of induction by enumeration:
Induction by simple enumeration is the following principle: "Given a number n of a's which
have been found to be p's, and no a which has been found to be not a p, then the two
statements: (a) 'the next a will be a p,' (b) 'all a's are p's,' both have a probability which
increases as n increases, and approaches certainty as a limit as n approaches infinity."
I shall call (a) "particular induction" and (b) "general induction." (1948, 401)
Between Russell's "particular induction" and his "general induction" we can interpolate another type, as the following definitions show (note that Russell's " a " and
" P " refer to properties, not to individual things):
Def. Relative to K, the predicate " P " is weakly projectible over the sequence of
individuals al9 a2, just in case15
lim Pr(Pan+l\Pal
Pan.K) = 1.
Def Relative to K, "P" is strongly projectible over al9 a2, ... just in case
lim Pr(Pan+l
n, m >
Pan+m\ Pa,
Pan.K) = 1.
(The notation
lim indicates the limit as m and n both tend to infinity in any manner
m, n 0
you like.) A sufficient condition for both weak and strong probability is that the
general hypothesis H: (i)Pat receives a nonzero prior probability. To see that it is
sufficient for weak projectibility, we follow Jeffreys's (1957) proof. By Bayes's
theorem
Pr(H\Pa
Pr(HlPa
>
_
~ Pr(Pai\K)
P.
Pa
+1
JH K) ~
Pr(Pai
Pan+,\H.K) Pr{H\K)
PriPa,
Pan+1\K)
Pr(H\K)
Pr(Pa2\Pai.K)- ... -PriPa^^Pa^
Pan-K)
15
Equation lim xn = L means that, for any real number e > 0, there is an integer N > 0
rto
such that, for all n > N, I xn L I < e.
The Confirmation of Scientific Hypotheses
93
Unless Pr(Pan+l\Pal
Pan.K) goes to 1 as n > , the denominator on the
right-hand side of the second equality will eventually become less than Pr{H\K),
contradicting the truth of probability that the left-hand side is no greater than 1.
The posit that
(?) Pr{[{i)Pat\K\ > 0
is not necessary for weak projectibility. Carnap's systems of inductive logic (see item
6 in Section 2.8 above) are relevant examples since in these systems (P) fails in a
universe with an infinite number of individuals although weak projectibility can hold
in these systems.16 But if we impose the requirement of countable additivity
(CA) lim Pr(Pat
Pan.K] = 1.
(Russell 1948 lays down a number of empirical postulates he thought were necessary
for induction to work. From the present point of view these postulates can be interpreted as being directed to the question of which universal hypotheses should be given
nonzero priors.)
Humean skeptics who regiment their beliefs according to the axioms of probability cannot remain skeptical about the next instance or the universal generalization
in the face of ever-increasing positive instances (and no negative instances) unless
they assign a zero prior to the universal generalization. But
Pr[(i)Pai\K] = 0
implies that
Pr[(3i) ~Pat\K\ = 1,
which says that there is certainty that a counterinstance exists, which does not seem
like a very skeptical attitude.
16
A nonzero prior for the general hypothesis is a necessary condition for strong projectibility but not for
weak projectibility. The point can be illustrated by using de Finetti's representation theorem, which says that if
P is exchangeable over aXt a2, . . . (which means roughly that the probability does not depend on the order)
then:
Pr(Pai.Pa2
where p,(6) is a uniquely determined measure on the unit interval 0 < 6 < 1. For the uniform measure d|x(6)
= d(0) we have
Pr(Pan+l\Pax
Pan. K) = n + \in + 2
and
Pr (Pan + 1
94
Pan+m\Pax
Pan.K) = m + \in + m + 1.
The Confirmation of Scientific Hypotheses
Note also that the above results on instance induction hold whether "P" is a
normal or a Goodmanized predicatefor example, they hold just as well for P*at
which is defined as
[(i <= 2000).PaJ V Hi > 2000).-Pa,-)],
where Pat means that at is purple. But this fact just goes to show how weak the results
are; in particular, they hold only in the limit as n and they give no information
about how rapidly the limit is approached.
Another way to bring out the weakness is to note that (P) does not guarantee
even a weak form of Hume projectibility.
Def Relative to K, "P" is weakly Hume projectible over the doubly infinite
sequence . . . , a_2, a_x, a0, al7 a2, . . . just in case for any n,
l\m Pr{Pan\Pan_x
Pan_k . K) = 1.
k>
(To illustrate the difference between the Humean and non-Humean versions of projectibility, let Pan mean that the sun rises on day n. The non-Humean form of
projectibility requires that if you see the sun rise on day 1, on day 2, and so on, then
for any e > 0 there will come a day N when your probability that the sun will rise on
day N + I will be at least 1 e. By contrast, Hume projectibility requires that if
you saw the sun rise yesterday, the day before yesterday, and so on into the past, then
eventually your confidence that the sun will rise tomorrow approaches certainty.)
If (P) were sufficient for Hume projectibility we could assign nonzero priors to
both {i)Pat and (i)P*ai9 with the result that as the past instances accumulate, the
probabilities for Pa200i and for P*a200l both approach 1, which is a contradiction.
A sufficient condition for Hume projectibility is exchangeability.
Def Relative to K, "P" is exchangeable for Pr over the ats just in case for any n
and m
Pr(Pan
Pan
+ m\K)
= Pr(Pan.
Pan.
+ m.\K)
where indicates that either P or its negation may be chosen and [av] is any
permutation of the ats in which all but a finite number are left fixed. Should we then
use a /V-function for which the predicate "purple" is exchangeable rather than the
Goodmanized version of "purple"? Bayesianism per se does not give the answer
anymore than it gives the answer to who will win the presidential election in the year
2000. But it does permit us to identify the assumptions needed to guarantee the
validity of one form or another of induction.
Having touted the virtues of the Bayesian approach to confirmation, it is now
only fair to acknowledge that it is subject to some serious challenges. If it can rise to
these challenges, it becomes all the more attractive.
2.10 CHALLENGES TO BAYESIANISM
1. Nonzero priors. Popper (1959) claims that "in an infinite universe . . . the
probability of any (non-tautological) universal law will be zero.'' If Popper were right
The Confirmation of Scientific Hypotheses
95
Pan\K).
GO
and that
(E)
Then except for the uninteresting case that Pr(Pan\K) = 1 for each n, it follows that
lim Pr(Pai
Pan\K) = 0
ft 00
Pan. K) = 1.
But, Popper urges, what is sauce for the goose is sauce for the gander. For we can do
the same for a Goodmanized P*, and from the limit statements we can conclude that
for some r > 0.5 there is a sufficiently large N such that for any N' > N, the
probabilities for Pa t and for P* a t are both greater than r, which is a contradiction
for appropriately chosen P*. But the reasoning here is fallacious and there is in
fact no contradiction lurking in Jeffreys's limit theorem since the convergence is
not supposed to be uniform over different predicatesindeed, Popper's reasoning
shows that it cannot be.
96
Of course, none of this helps with the difficult questions of which hypotheses
should be assigned nonzero priors and how large the priors should be. The example
from item 5 in Section 2.8 above suggests that the latter question can be ignored to
some extent since the accumulation of evidence tends to swamp differences in priors
and force merger of posterior opinion. Some powerful results from advanced probability theory show that such merger takes place in a very general setting (on this
matter see Gaifman and Snir 1982).
2. Probabilification vs. inductive support. Popper and Miller (1983) have argued that even if it is conceded that universal hypotheses may have nonzero priors and
thus can be probabilified further and further by the accumulation of positive evidence,
the increase in probability cannot be equated with genuine inductive support. This contention is based on the application of two lemmas from the probability calculus:
Lemma l.Pr(~H\E.K) x Pr(~E\K) = Pr(H V ~E\K) - Pr(H v
~E\E.K).
Pan.
Intuitively, the part of// that goes beyond this evidence is (/) [ (i > n) ... Pat] and
Pan).
not the Popper-Miller (i)Pa1 V (Pal
Gillies (1986) restated the Popper-Miller argument using a measure of inductive
support based on the incremental model of confirmation: (leaving aside K) the support
given by E to H is S{H, E) = Pr(H\E) - Pr(Z/). We can then show that
Lemma 3. S(H, E) = S(H V E, E) 4- S(H V ~E, E).
The Confirmation of Scientific Hypotheses
97
Gillies suggested that S(H V EE, ) be identified as the deductive support given
H by E and S(H V ~E, E) as the inductive support. And as we have already seen,
in the interesting cases the latter is negative. Dunn and Hellman (1986) responded by
dualizing. Hypothesis H is logically equivalent to (H.E) v (H.E) and S(H, E)
S(H.E, E) + S(H. *~E, E). Identify the second component as the deductive countersupport. Since this is negative, any positive support must be contributed by the first
component which is a measure of the nondeductive support.
3. The problem of old evidence. In the Bayesian identification of the valid
kernel of the H-D method we assumed that Pr(E\K) < 1, that is, there was some
surprise to the evidence E. But this is often not the case in important historical
examples. When Einstein proposed his general theory of relativity (H) at the close of
1915 the anomalous advance of the perihelion of Mercury (E) was old news, that is,
Pr(E\K) = 1. Thus, Pr{H\E.K) = Pr(H\K), and so on the incremental conception
of confirmation, Mercury's perihelion does not confirm Einstein's theory, a result that
flies in the face of the fact that the resolution of the perihelion problem was widely
regarded as one of the major triumphs of general relativity. Of course, one could seek
to explain the triumph in nonconfirmational terms, but that would be a desperate
move.
Garber (1983) and Jeffrey (1983) have suggested that Bayesianism be given a
more human face. Actual Bayesian agents are not logically omniscient, and Einstein
for all his genius was no exception. When he proposed his general theory he did not
initially know that it did in fact resolve the perihelion anomaly, and he had to go
through an elaborate derivation to show that it did indeed entail the missing 43" of arc
per century. Actual flesh and blood scientists learn not only empirical facts but
logicomathematical facts as well, and if we take the new evidence to consist in such
facts we can hope to preserve the incremental model of confirmation. To illustrate, let
us make the following assumptions about Einstein's degrees of belief in 1915:
(a) Pr(H\K) > 0 (Einstein assigned a nonzero prior to his general theory.)
(b) Pr(E\K) = 1 (The perihelion advance was old evidence.)
(c) Pr(H h- E\K) < 1 (Einstein was not logically omniscient and did not invent his
theory so as to guarantee that it entailed the 43".)
E)\K] = 1 (Einstein knew that his theory entailed a
(d) Pr[(H \- E)v (H \
definite result for the perihhelion motion.)
(e) Pr[H.(Hh-~E)\K] = Pr[H.(H\- ~E).~E\K\ (Constrainton interpreting has
logical implication.)
From (a)-(e) it can be shown that Pr[H\(H \-E).K\. > Pr(H\K). So learning that his
theory entailed the happy result served to increase Einstein's confidence in the theory.
Although the Garber-Jeffrey approach does have the virtue of making Bayesian
agents more human and, therefore, more realistic, it avoids the question of whether
the perihelion phenomena did in fact confirm the general theory of relativity in favor
of focusing on Einstein's personal psychology. Nor is it adequate to dismiss this
98
concern with the remark that the personalist form of Bayesianism is concerned precisely with psychology of particular agents, for even if we are concerned principally
with Einstein himself, the above calculations seem to miss the mark. We now believe
that for Einstein in 1915 the perihelion phenomena provided a strong confirmation of
his general theory. And contrary to what the Garber-Jeffrey approach would suggest,
we would not change our minds if historians of science discovered a manuscript
showing that as Einstein was writing down his field equations he saw in a flash of
t E or alternatively that he consciously constructed his
mathematical insight that H
field equations so as to guarantee that they entailed E. "Did E confirm H for Einstein?" and "Did learning that H\ E increase Einstein's confidence in / / ? " are two
distinct questions with possibly different answers. (In addition, the fact that agents are
allowed to assign Pr (H \- E\K)<\ means that the Dutch book justification for the
probability axioms has to be abandoned. This is anathema for orthodox Bayesian
personalists who identify with the betting quotient definition of probability.)
A different approach to the problem of old evidence is to apply the incremental
model of confirmation to the counterfactual degrees of belief that would have obtained had E not been known. Readers are invited to explore the prospects and
problems of this approach for themselves. (For further discussion of the problem of
old evidence, see Howson 1985, Eells 1985, and van Fraassen 1988.)
2.11 CONCLUSION
The topic of this chapter has been the logic of science. We have been trying to
characterize and understand the patterns of inference that are considered legitimate in
establishing scientific resultsin particular, in providing support for the hypotheses
that become part of the corpus of one science or another. We began by examining
some extremely simple and basic modes of reasoningthe hypothetico-deductive
method, instance confirmation, and induction by enumeration. Certainly (pace Popper) all of them are frequently employed in actual scientific work.
We findboth in contemporary science and in the history of sciencethat
scientists do advance hypotheses from which (with the aid of initial conditions and
auxiliary hypotheses) they deduce observational predictions. The test of Einstein's
theory of relativity in terms of the bending of starlight passing close to the sun during
a total solar eclipse is an oft-cited example. Others were given in this chapter.
Whether the example is as complex as general relativity or as simple as Boyle's law,
the logical problems are the same. Although the H-D method contains a valid
kernelas shown by Bayes's ruleit must be considered a serious oversimplification
of what actually is involved in scientific confirmation. Indeed, Bayes's rule itself
seems to offer a schema far more adequate than the H-D method. Butas we have
seenit, too, is open to serious objections (such as the problem of old evidence).
When we looked at Hempel's theory of instance confirmation, we discussed an
example that has been widely cited in the philosophical literaturenamely, the generalization "All ravens are black." If this is a scientific generalization, it is certainly
at a low level, but it is not scientifically irrelevant. More complex examples raise the
same logical problems. At present, practicing scientists are concerned withand
The Confirmation of Scientific Hypotheses
99
excited bysuch generalizations as, "All substances having the chemical structure
given by the formula YBa2Cu 307 are superconductors at 70 kelvins." As if indoor
ornithology weren't bad enough, we see, by Hempel's analysis, that we can confirm
this latter-day generalization by observing black crows. It seems that observations by
birdwatchers can confirm hypotheses of solid state physics. (We realize that birdlovers would disapprove of the kind of test that would need to be performed to
establish that a raven is not a superconductor at 70K.) We have also noted, however,
the extreme limitations of the kind of evidence that can be gathered in any such
fashion.
Although induction by enumeration is used to establish universal generalizations, its most conspicuous use in contemporary science is connected with statistical
generalizations. An early example is found in Rutherford's counting of the frequencies with which alpha particles bombarding a gold foil were scattered backward (more
or less in the direction from which they came). The counting of instances led to a
statistical hypothesis attributing stable frequencies to such events. A more recent
exampleemploying highly sophisticated experimentsinvolves the detection of
neutrinos emitted by the sun. Physicists are puzzled by the fact that they are detecting
a much smaller frequency than current theory predicts. (Obviously probabilities of the
type characterized as frequencies are involved in examples of the sort mentioned
here.) In each of these cases an inductive extrapolation is drawn from observed
frequencies. In our examination of induction by enumeration, however, we have
found that it is plagued by Hume's old riddle and Goodman's new one.
One development of overwhelming importance in twentieth-century philosophy
of science has been the widespread questioning of whether there is any such thing as
a logic of science. Thomas Kuhn's influential work, The Structure of Scientific
Revolutions (1962, 1970), asserted that the choice of scientific theories (or hypotheses) involves factors that go beyond observation and logicincluding judgement,
persuasion, and various psychological and sociological influences. There is, however, a strong possibility that, when he wrote about going beyond the bounds of
observation and logic, the kind of logic he had in mind was the highly inadequate H-D
schema, (see Salmon 1989 for an extended discussion of this question, and for an
analysis of Kuhn's views in the light of Bayes's rule). The issues raised by the
Kuhnian approach to philosophy of science are discussed at length in Chapter 4 of this
book.
Among the problems we have discussed there areobviouslymany to which
we do not have adequate solutions. Profound philosophical difficulties remain. But
the deep and extensive work done by twentieth-century philosophers of science in
these areas has cast a good deal of light on the nature of the problems. It is an area
in which important research is currently going on and in which significant new results
are to be expected.
DISCUSSION QUESTIONS
1. Select a science with which you are familiar and find a case in which a hypothesis or theory is
taken to be confirmed by some item of evidence. Try to characterize the relationship between the
100
evidence and hypothesis or theory confirmed in terms of the schemas discussed here. If none of
them is applicable, can you find a new schema that is?
2. If the prior probability of every universal hypothesis is zero how would you have to rate the
probability of the statement that unicorns (at least one) exist? Explain your answer.
3. Show that accepting the combination of the entailment condition, the special consequence
condition, and the converse consequence condition (see Section 2.4) entails that any E confirms
any//.
4. Consider a population that consists of all of the adult population of some particular district. We
want to test the hypothesis that all voters are literate,
(x)(Vx Z) Lx),
which is, of course, equivalent to
(x)(~ LxD ~ Vx).
Suppose that approximately 75 percent of the population are literate voters, approximately 15
percent are literate nonvoters, approximately 5 percent are illiterate nonvoters, and approximately 5 percent are illiterate votersbut this does not preclude the possibility that no voters are
illiterate. Would it be best to sample the class of voters or the class of illiterate people? Explain
your answer. (This example is given in Suppes 1966, 201.)
5. Goodman's examples challenge the idea that hypotheses are confirmed by their instances.
Goodman holds that the distinction between those hypotheses that are and those that are not
projectable on the basis of their instances is to be drawn in terms of entrenchment. Predicates
become entrenched as antecedents or consequents by playing those roles in universal conditionals that are actually projected. Call a hypothesis admissible just in case it has some positive
instances, no negative instances, and is not exhausted. Say that H overrrides H' just in case H
and H' conflict, H is admissible and is better entrenched than H' (i.e., has a better entrenched
antecedent and equally well entrenched consequent or vice versa), and H is not in conflict with
some still better entrenched admissible hypothesis. Critically discuss the idea that// is projectable on the basis of its positive instances just in case it is admissible but not overridden.
6. Show that
H: (x) (3 y) Rxy.(x) - Rxx.(x) (y) (z) [(Rxy.Ryz) D Rxz]
cannot be Hempel-confirmed by any consistent E.
7. It is often assumed in philosophy of science that if one is going to represent numerically the
degree to which evidence E supports hypothesis H with respect to background B, then the
numbers so produced P(H\E.B) must obey the probability calculus. What are the prospects of alternative calculi? (Hint: Consider each of the axioms in turn and ask under what
circumstances each axiom could be violated in the context of a confirmation theory. What
alternative axiom might you choose?)
8. If Bayes's rule is taken as a schema for confirmation of scientific hypotheses, it is necessary to
decide on an interpretation of probability that is suitable for that context. It is especially crucial
to think about how the prior probabilities are to be understood. Discuss this problem in the light
of the admissible interpretations offered in this chapter.
9. William Tell gave his young cousin Wesley a two-week intensive archery course. At its completion, William tested Wes's skill by asking him to shoot arrows at a round target, ten feet in
radius with a centered bull's-eye, five feet in radius.
"You have learned no control at all," scolded William after the test. "Of those arrows
that hit the target, five are within five feet of dead center and five more between five and ten feet
from dead center." "Not so," replied Wes, who had been distracted from archery practice by
The Confirmation of Scientific Hypotheses
101
his newfound love of geometry. "That five out of ten arrows on the target hit the bull's-eye
shows I do have control. The bullseye is only one quarter the total area of the target."
Adjudicate this dispute in the light of the issues raised in the chapter. Note that an
alternative form of Bayes's rule which applies when one considers the relative confirmation
accrued by two hypotheses Hx and H2 by evidence E with respect to background B is:
Pr(Hx\E.B) __ Pr(E \H^B) Pr{Hx\B)
Pr(H2\E.B) " Pr{E\H2.B) * Pr(H2\B)
10. Let {Hlt H2, . . . , Hn} be a set of competing hypotheses. Say that E selectively Hempelconfirms some # , just in case it Hempel-confirms Hs but fails to confirm the alternative Hs. Use
this notion of selective confirmation to discuss the relative confirmatory powers of black ravens
versus nonblack nonravens for alternative hypotheses about the color of ravens.
11. Prove Lemmas 1,2, and 3 of Section 2.10.
12. Discuss the prospects of resolving the problem of old evidence by using counterfactual degrees
of belief, that is, the degrees of belief that would have obtained had the evidence E not been
known.
13. Work out the details of the following example, which was mentioned in Section 2.8. There is
a square piece of metal in a closed box. You cannot see it. But you are told that its area is
somewhere between 1 square inch and 4 square inches. Show how the use of the principle of
indifference can lead to conflicting probability values.
14. Suppose there is a chest with two drawers. In each drawer are two coins; one drawer contains
two gold coins, the other contains one gold coin and one silver coin. A coin will be drawn from
one of these drawers. Suppose, further, that you know (without appealing to the principle of
indifference) that each drawer has an equal chance of being chosen for the draw, and that,
within each drawer, each coin has an equal chance of being chosen. When the coin is drawn
it turns out to be gold. What is the probability that the other coin in the same drawer is gold?
Explain how you arrived at your answer.
15. Discuss the problem of ascertaining limits of relative frequencies on the basis of observed
frequencies in initial sections of sequences of events. This topic is especially suitable for those
who have studied David Hume's problem regarding the justification of inductive inference in
Part II of this chapter.
16. When scientists are considering new hypotheses they often appeal to plausibility arguments. As
a possible justification for this procedure, it has been suggested that plausibility arguments are
attempts at establishing prior probabilities. Discuss this suggestion, using concrete illustrations
from the history of science or contemporary science.
17. Analyze the bootstrap confirmation of the perfect gas law in such a way that no "macho"
bootstrapping is used, that is, the gas law itself is not used as an auxiliary to deduce instances
of itself.
SUGGESTED READINGS
(1980), Theory and Evidence. Princeton: Princeton University Press. This book,
which is rather technical in parts, contains the original presentation of bootstrap confirmation.
GOODMAN, NELSON (1955), Fact, Fiction, and Forecast. 1st ed. Cambridge, MA: Harvard University Press. This twentieth-century classic is now in its 4th edition. Chapter 3 contains Goodman's dissolution of "the old riddle of induction" and presents his "new riddle of induction"
the grue-bleen paradox. Chapter 4 gives Goodman's solution of the new riddle in terms of
projectibility.
GLYMOUR, CLARK
102
HEMPEL, CARL
(1966), Philosophy of Natural Science. Englewood Cliffs, NJ: Prentice-Hall. Chapters 2-4
provide an extremely elementary and readable introduction to the concept of scientific confirmation.
HUME, DAVID (1748), An Enquiry Concerning Human Understanding. Many editions available.
This is the philosophical classic on the problem of induction, and it is highly readable. Sections
47 deal with induction, causality, probability, necessary connection, and the uniformity of
nature.
POPPER, KARL R. (1972), ' 'Conjectural Knowledge: My Solution of the Problem of Induction,'' in
Objective Knowledge. Oxford: Clarendon Press, pp. 1-31. This is an introductory presentation of
his deductivist point of view.
SALMON, WESLEY C. (1967), The Foundations of Scientific Inference. Pittsburgh: University of
Pittsburgh. This book provides an introductory, but moderately thorough, survey of many of the
issues connected with confirmation, probability, and induction.
STRAWSON, P. F. (1952), Introduction to Logical Theory. London: Methuen. Chapter 9 presents the
ordinary language dissolution of the problem of induction.
103
Three
REALISM
AND THE NATURE
OF THEORIE S
Clark
Glymour
When we read about the results of scientific discoveries, we usually understand these
accounts as descriptions of nature. If we pick up an issue of Scientific American, for
example, we may find an article about the way in which some feature of the immune
system works, or about the origin of the planets. It seems strangely trivial to note that
we understand such articles to be telling us about the immune system, or how the
planets came to be. What else would they be about except what their words say they
are about? The same applies to articles in scientific journals and books, and to
scientific lectures and even to conversations on scientific matters. Obviously they are
about what they seem to be talking about.
On many occasions throughout the history of modern science this apparently obvious view of what scientific papers, books, lectures and conversations are about has
been emphatically denied. Many of the naysayers have been among the greatest of
scientists, or have attached their denials to great scientific works. The preface to Copernicus's great work, De Revolutionibus (1952), was not written by Copernicus himself, but it announced that despite what the words in the book appeared to mean,
Copernicus's theory was not really about how the planets move; instead, so the preface
claimed, the book merely presented a mathematical device for computing the positions
of the planets on the celestial sphere. Copernican theory, according to the preface, was
an instrument, not a description. In the 1830s different methods of determining the
atomic masses of the elements gave different and conflicting results. Methods using
chemical analogies gave different masses from methods using the law of Dulong and
Petit which determine atomic masses from heat capacities. The latter methods gave still
different masses from methods that used vapor density measurements. Jean Marie Dumas, the leading French chemist of the time, concluded that a different atomic mass
104
should apply according to how the atomic mass is determined. Thus, said Dumas, atoms don't have a single mass, they have a mass associated with vapor density measurements, another mass associated with heat capacity measurements, and so on.
Contrary to appearances, in Dumas's view science does not show us how to measure
one and the same property of things in different ways; instead each kind of measurement operation has a distinct property. In the twentieth century, after the theory of
relativity had replaced the Newtonian concept of mass with two distinct notionsrest
mass and proper massPercy Bridgeman (1927), a distinguished physicist, proposed
that every distinct physical operation determines a distinct property, a view he called
operationalism. Late in the nineteenth century many consistent alternatives to Euclidean geometry had been developed, and Henri Poincare, one of the greatest mathematicians and mathematical physicists of the time, realized that by various changes in
physical dynamics any of these alternative geometries could be made consistent with
scientific observations. Poincare argued that the geometry of space is not something the
world forces upon us; instead, we force geometry on the world. We, in effect, adopt
the convention that we will measure things in such a way, and formulate our physics
in such a way, that Euclidean geometry is true.
Many, many other examples like these could be given. The question of just how
scientific theories should be understood becomes a burning issue whenever it seems
that the content of our theories is uncertain, and especially whenever it seems that
science is faced with alternative theories that could equally account for all possible
observations. The threat that the content of science might in some respects be underdetermined by all evidence we might ever have often brings about proposals by
scientists and by philosophers to reinterpret the meaning of that content.
In this chapter we consider the evolution of modern philosophical and scientific
debates over scientific realism. These debates concern both the nature and content of
scientific theories, and whether we can have any real justification for believing the
claims of science.
3.1 METAPHYSICAL SKEPTICISM AND INDUCTIVE SKEPTICISM
In the seventeenth century, mechanical explanations flourished in chemistry, in physics, in the study of gases, and in many other areas of science. The emerging new
sciences tried to explain phenomena by the motions, masses, shapes and collisions of
component bodiesthe atoms or "corpuscles" that were thought to make up all
matter. English philosophers of science in the seventeenth and eighteenth centuries
tried to make philosophical sense of the great revolution in the sciences that had taken
hold in the seventeenth century. One of the most immediate philosophical problems
was the difference between the world we experience and the world described by
seventeenth-century mechanical theories of nature. The world of our experience is
filled with colors, tastes, smells, sounds, heats, and more that does not appear among
the basic properties in Newtonian mechanical science. One fundamental philosophical question therefore had to do with the connection between the world we experience and the purely mechanical world that the new sciences postulated: How are the
colors, tastes, odors, and heats we observe in the world produced by mechanical
actions, and how can we know how this comes about?
Realism and the Nature of Theories
105
you see, hear, taste, smell and feel is unreal; your sensations are produced by stimulating your sensory nerves according to a plan that is executed by a sophisticated
computer. No matter what you may observe in the course of your life, your observations will be consistent both with the hypothesis that you are a brain in a vat and
with the hypothesis that you are not.
To such metaphysical perplexities about science and knowledge, David Hume
added another in the eighteenth century. Hume' s problem of induction has already been
considered (see Chapter 2). Hume phrased it as a problem about our knowledge of any
necessary connection between cause and effect. He argued that we do not perceive any
necessary connection between a cause and its effect, and therefore the idea of necessary
connection cannot come from our experience of the external world. Hume proposed
that the sourceand therefore the contentof the idea of necessary connection is our
observation within ourselves of the force of habit that leads us to expect a familiar effect
when we observe a familiar cause. From this analysis, Hume argued that neither experience nor reason can guarantee that future examples of any familiar cause will be
followed by the familiar effect. For all that reason or experience can establish, bread
will not nourish people tomorrow, or tomorrow daytime will not be followed by night.
Hume framed his argument in terms that were appropriate to the "idea" idea
and to the philosophical language of the eighteenth century, but his inductive skepticism is almost as old as philosophy. A related point is made very clearly, and in
some respects more clearly than in Hume, by Plato (1900-1903) in his dialogue, The
Meno. Plato's protagonists, Socrates and Meno, try to establish a general, universal
hypothesis by looking for examples or counterexamples. The general hypotheses that
Plato considers in the dialogue are proposed conditions for virtue, but Plato's philosophical point applies as well to any general scientific hypotheses, for example
hypotheses about the melting point of water.
Imaginably, all pure water could melt at one and the same temperature, or
different samples of water could melt at different temperatures. It is perfectly consistent to imagine a world in which before this time all samples of water melted at 0
degrees celsius, but after this time water will melt at 10 degrees celsius. We trust that
we don't actually live in such a world, but we can imagine one well enough. Suppose
we conjecture on the basis of our experience that all pure water melts at 0 degrees
celsius and suppose in fact we are right about the actual world we inhabit. In the
future we will therefore continue to be correct about the melting point of water. A
modern version of Plato's question is this: Even if we are correct right now about
what the melting point of water is for all cases everywhere, in the past and in the
future, how can we know right now that we are correct?
We can understand the logical structure of Hume's and Plato's problems if we
imagine a game a scientist must play against a demon bent on deceiving the scientist.
We will suppose the two players can live forever. The demon has available a collection of possible worlds. Both the demon and the scientist know which worlds are
in the collection. The scientist and the demon agree on a proposition that is true in
some of these possible worlds and false in others. The proposition might be "bread
always nourishes" or "the melting point of water is zero degrees celsius" or any
other proposition of interest. The demon gets to choose a possible world, and he gives
the scientist facts about that world one at a time in any order the demon pleases. But
Realism and the Nature of Theories
107
the demon is not allowed to lie about the facts and is not allowed to withhold any
relevant fact; every fact about the world the demon has chosen must eventually be put
in evidence for the scientist. The scientist must guess whether or not the proposition
is true in the world the demon has chosen. If he wishes the scientist can change his
guess whenever a new fact is received.
What does it mean to win this game? Winning is coming to know. For Plato
(and, less clearly, for Hume) the scientist knows the truth or falsity of the proposition
if and only if no matter which world the demon chooses, after some finite number of
facts have been received, the scientist can announce the conjecture that will always
henceforth be given about the proposition, and that conjecture is correct. For Plato,
knowledge requires true belief acquired by a reliable method that permits the knower
to know that she knows.
The simple but devastating logical fact is that except for the very trivial propositions or very restricted sets of alternative worlds, the scientist can never win this
sort of game. The conditions Plato and Hume require for knowledge can almost never
be met unless the alternative circumstances are restricted a priori, a supposition Plato
endorsed and Hume denied.
In one form or another the problems of metaphysical skepticism and inductive
skepticism have been central to theories of knowledge and philosophies of science
throughout the history of philosophy, and the same remains true of philosophy of
science today. The two problems are alike in calling into question our ability to have
scientific knowledge by considering circumstances in which alternative hypotheses
are consistent with the evidence we might have. The problems differ in detail.
Inductive skepticism trades on the fact that we arefinitebeings who only have a finite
amount of data available at any time. Metaphysical skepticism trades on the assumption of a separation between the kinds of things we have evidence about and the kinds
of things we wish to know about; even if we had infinite knowledge of things of the
first kind, it would not suffice to determine the truth about things of the second kind.
These problems are at the bottom of many disputes over the nature of scientific
claims, the structure of scientific theories, and the appropriateness of belief in the
claims of modern science.
the world we experience we have tables, chairs, houses and other people; in the world
we experience, bodies move (approximately) according to the law of inertia, events
occur one after another, things have shapes, the properties of figures and shapes are
described by geometrical principles, and familiar causes produce familiar effects.
This is the world science tries to describe, not the world of things in themselves.
In an important sense, Kant rejected the notion that what we perceive are our own
ideas. We perceive things in the world of experience, not our own ideas. InKant'sview
it is true enough that the structure of our experience depends in part on the structure of
our minds. The objects we see, touch, hear, feel or smell are in some respects constituted by our own minds: The unknowable things in themselves together with our
minds produce the world of experience. But exactly what it is to be an object of
experiencean ordinary chair for exampleis to be an entity constructed in this way.
While this is one way out of metaphysical skepticism, what about inductive
skepticism? Kant had an ingenious solution there as well, and it is much more
intricate and requires a further look at Kant's philosophical project. Kant asked the
following sorts of questions: In order for there to be any experience of things in space
and time and any experience of events occurring one after the other, what conditions
are necessary? We can interpret this sort of question psychologicallyhow must the
mind work in order for one to have experience of the sort we do?or logicallywhat
principles are logically presupposed by the assumption that we have experience at all?
(Kant's words suggest both a psychological and a logical reading of his project, and
scholars have argued about the matter ever since.) Kant called these sorts of questions
transcendental, and to this day philosophers call "transcendental" any arguments
that try to establish that something must be true because it is (or is claimed to be) a
necessary condition for human knowledge or for human experience.
What separates transcendental arguments from ordinary empirical arguments of
the kind given in the sciences is not entirely clear. Some empirical arguments take as
premises particular facts of our experience (for example, that particular samples of
sodium melted at particular temperatures) and argue to generalizations (for example,
that all pure sodium melts within a specific interval of temperatures). Some scientific
arguments take general features of human experience (for example, that the motions
of the planets satisfy Kepler's second law) and argue to deeper general laws (such as
the law of gravitation). "Transcendental" arguments seem to be distinguished only
by the fact that their premises about experience are very, very general (and therefore
often banal), and their conclusions are not supposed merely to be the best explanation
of the premises or the best scientific conjecture, but demonstrably necessary in order
for the premises to be true.
Kant claimed that the way the mind forms experience guarantees that objects of
experience must have certain features. Objects are perceived in space, and Kant held
that the "empirical intuition" of spacethat is, seeing particular things in space
requires that the principles of Euclidean geometry truly describe spatial relations.
Events occur in time, and Kant further held that the "intuition" of time requires that
events occur in a linear, serial order like the integers. Euclidean geometry and
arithmetic are not themselves spatial or temporal intuitions; they are what Kant called
the forms of sensible intuitions of space and time. Indeed, Kant held that because of
these forms of intuition, we can have a priori knowledge that geometry and arithmetic
Realism and the Nature of Theories
109
will be true in the world we experiencethat is, knowledge not derived from any
particular experiences. Kant thought our knowledge of the truth of propositions of
arithmetic and geometry was especially puzzling. These propositions appear to be
certain and not to be based on experiment of any kind, but their truth did not seem to
Kant to be "analytic" either. Although 7 + 5 = 12, the concept of seven plus five,
Kant maintained, does not contain the concept of twelve, and hence we could not
come to know that 7 + 5 = 12 merely by analyzing the concept of seven plus five.
The doctrine that arithmetic is the pure form of the intuition of time seemed to explain
how we could know such "synthetic" claims a priori.
In addition to these ' 'pure forms of intuition,'' Kant held that certain ' 'pure concepts of the understanding" necessarily apply to any judgements we make about the
objects of experience. Kant called these concepts ' 'pure'' because he held that they too
are not obtained from any experience; instead they form experience, or they are preconditions of experience of objects. Kant's "pure concepts of the understanding" were
taken from the logical categories commonly used in his day (and, indeed, Kant himself
wrote a treatise on logic). For our purposes the most important concept of the understanding is that of causality. The upshot in psychological terms is that through a process
that is not entirely clear, the understanding guarantees that objects of experience are
subject to causal regularities. Kant did not hold that the understanding generally reveals
a priori which possible causal laws are true; that is up to science to determine from
experience. The understanding only guarantees that if objects are experienced they and
their properties must be subject to some causal regularities or other. But Kant did argue
that the understanding also guarantees a priori the truth of many principles of Newtonian physics, for example the law of inertia.
Kant's system seemed to provide an answer to most of the philosophical problems that beset eighteenth-century science. The doubts of metaphysical skepticism
were pushed aside as of no relevance to the aims of science or the concerns of
humans. The problem of inductive skepticism appeared to be solved. Furthermore,
Kant's system justified the feelingmore common in the eighteenth century than
todaythat geometry, arithmetic, and the fundamental principles of physical kinematics and dynamics are too certain to be based on generalizations from experience
and too full of content to be merely disguised definitions.
The influence Kant's system had among scientists and philosophers is hard for
us to appreciate today. One example will have to serve by way of illustration. Josiah
Willard Gibbs was the first great American-born theoretical physicist, one of the
figures most responsible for forming that branch of physics called statistical thermodynamics, which attempts to understand phenomena of heat through the application
of probability theory to the mechanics of systems of particles. In the preface to his
seminal work, Elementary Principles of Statistical Mechanics (1960), Gibbs wrote
that his aim was to reduce the phenomena of heat to regularities of mechanics that
have an a priori foundation.
3.3 SKEPTICISM AND THE ANTINOMIES OF REASON
meaningful scientific question we might pose about the world of experience can be
settled by scientific investigation. Hume's problem may be solved by Kant's picture
of the role of the concept of causation, but the logical problem about learning from
experience that Plato posed doesn't seem to be solved at all. Kant himself argued that
questions arise that cannot be solved empirically or rationally. Kant ([1787] 1929,
384-484) described four ' 'antinomies of pure reason,'' two of which seem to be clear
scientific questions about the world of experience:
1. Does the world have a beginning in time and is space unlimited?
2. Are objects infinitely divisible?
Kant held that these are intelligible questions about the world of experience, and
that no experience or a priori reasoning could settle them. He tried to establish as
much by giving an a priori proof in each case that the answer is "yes" and in each
case another equally good a priori proof that the answer is "no." Kant's "proofs,"
like most of his attempts at demonstrations, are dubious and equivocal, but a rigorous
logical sense exists in which he was correct that these questions cannot be settled by
experience.
Recall the scientist and the demon. Suppose the demon may choose from among
different possible worlds in which some particular objecta banana, for example
may be divided into exactly two pieces, a world in which it may be divided into three
pieces, a world in which it may be divided into four pieces, and so on for every
number. Let us also include a world in which the banana can be divided without end.
The second question, about infinite divisibility, has a "yes" answer only in the last
of these possible worlds. Let us suppose the scientist can do experiments to try to
divide the banana or a part of it. One experimental method may not succeed at
dividing an object, while another experimental method will work. If the object is
divisible, let us assume that an experimental method available to the scientist will
divide it, although the scientist can only discover which method, if any, works by
trying the method experimentally. Let us further assume that the number of methods
of division that might be tried is unlimited. The demon picks one of the possible
worlds and in that world the banana has some definite divisibility. The scientist tries
experimental methods to divide the banana or its parts, and after each experiment the
scientist can conjecture whether the banana is infinitely divisible.
Can the scientist win this game? Not if we use Plato's criterion for winning,
namely, that the scientist must not only get the right answer to the question but she
must eventually know when she had got the right answer. Indeed, the scientist cannot
win this game even if we loosen Plato's conception of knowledge and do not require
that the scientist eventually know when she is correct. Suppose we change things as
follows: The scientist wins if no matter what world the demon chooses, the scientist
eventually reaches a point at which she guesses the correct answer to the second
question in that world, and continues to guess the correct answer in that world ever
after. The scientist does not have to be able to say when she has reached the point of
convergence to the truth. A moment's reflection will show that many games exist that
a scientist can win in this way that cannot be won if the scientist is held to Plato's
standard. For example, the scientist could now win a game in which the question was
Realism and the Nature of Theories
111
whether all samples of water melt at zero degrees celsius. But the scientist cannot win
the game on the question of infinite divisibility even with this more charitable criterion of success. In a precise sense, Kant was right that the answer to the question is
beyond all possible experiences.
Kant's picture therefore leaves open important practical issues about which
scientific questions can and which cannot be settled by which methods. Before and
after Kant those issues sparked debates over scientific realism.
3.4 LOGICAL POSITIVISM AND THE LINGUISTIC TURN
Most philosophers and many scientists educated in the late nineteenth and early
twentieth centuries were taught Kantian ideas, but by the 1920s the most daring and
sophisticated young philosophers of science had come to reject the Kantian picture.
The partial disintegration of the Kantian solution was due to several factors. One was
the scientific rejection of principles of geometry and physics that Kant had claimed to
prove are synthetic a priori truths. The middle and late nineteenth centuries saw the
development of non-Euclidean geometries that vied with Kant's Euclidean geometry
as descriptions of the structure of space. By the 1920s the general theory of relativity
had led a good many scientists to thinkcontrary to Poincarethat Euclidean geometry is not true of physical space, and had also led scientists to reject the law of
inertia as Newton and Kant understood it. Such results inevitably led scientifically
informed philosophers to wonder about Kant's arguments: How good could Kant's
demonstrations be if they were refuted by the advance of science?
Another factor in the demise of Kant's system was the development of modern
logic at the hands of Gottlob Frege (1972), David Hilbert and others. In 1879 Frege
had revolutionized logic, laying the foundations for a theory of mathematical proof
that permitted the mathematical study of the structure of mathematics itself. By the
1920s Hilbert, Bertrand Russell and others had developed Frege's ideas into a rich
and powerful tool for analyzing language and arguments in mathematics, physics and
even philosophy. From the perspective provided by this new tool, Kant's arguments
didn't look very sound. Rather than rigorous demonstrations, they seemed to be
equivocal, full of gaps, and well short of proving their claims. Kant's response to
Hume, for example, seemed inadequate when one moved from talk of justifying a
vague principle of causality to questions about how universal laws of nature could be
known on the basis of any finite amount of evidence. Kant was left with a sort of
picture but without convincing arguments for its details. The Kantian system was
replaced by a new philosophical enterprise: unfolding the logic of science.
Frege's logical theory focused on formal languages, which were presented as
abstract but precise mathematical objects, as definite as the system of natural numbers. Frege's greatest achievement was to show how a variety of mathematical theories could be represented in such formal languages and how the notion of a proof
could be characterized as a precise mathematical property of certain sets of sequences
of symbols in such a language. Early twentieth-century logicians clarified Frege's
accomplishment, formalized still more powerful languages and other mathematical
theories, and introduced a mathematical representation of how language describes
various kinds of facts.
112
The new subject of model theory that emerged concerned the representation of
how language describes the actual world or any possible world. Associated with any
formal language were an infinity of set-theoretic structures called models. Each model
consisted of a set (the domain of the model) and special subsets of the domain, or
special subsets of pairs or triples or /z-tuples of members of the domain. The domain
of a model was a mathematical representation of the objects in some possible world,
and the subsets, and sets of pairs, triples, and so on, of members of the domain were
mathematical representations of properties or relations among things in that possible
world. The idea, roughly, is that a property corresponds to the set of things that have
that property in each possible world. ' 'Brown'' corresponds in each possible world to
the set of brown things in that world.
Parts of the formal language were associated with parts of any model for the
language. Certain symbols in the language named members of the domain, while
other symbols denoted properties or relationsthat is, subsets or sets of pairs or
triples and so on of members of the domain. Other symbols in the formal language
were used as variables that range over members of the domain. Under such an
association between symbols of an abstract formal language, on the one hand, and
features of a model, on the other hand, it was possible to define in explicit mathematical terms what it means for a sentence in the formal language to be true (or false)
in the model.
With this apparatus, one could study the expressive power of various formalized
theories entirely mathematically; one could study the conditions under which theories
could (or could not) be described by a finite collection of sentences; one could study
which properties of collections of models could (or could not) be characterized by any
possible theory in a formal language; one could study whether formalized theories
were complete or incomplete in the sense that they left no, or some, claims about their
models indeterminate; and we could study which properties of formal theories could
be determined by algorithmsthe way the long division algorithm of elementary
arithmetic determines the result of dividing one number by another. By 1930 Kurt
Godel, then a young logician, had proved that a reformulation of Frege's logic by
Hilbert and Ackermann is complete, meaning that every sentence of a kind of formal
language that is true in every model associated with the language is also a sentence
that can be proved by their rules of proof.
The logical revolution of the late nineteenth and early twentieth centuries led to
revolutionary developments in the foundations of mathematics and eventually to the
creation of the contemporary subject of computer science. Along the way it also
changed philosophy of science, and that is what interests us here.
Two leaders of this new logical perspective were Bertrand Russell in England
and Rudolf Carnap in Austria. Russell had originally been a Kantianin fact his
doctoral thesis was a defense of the Kantian view of geometry in the face of the
development of non-Euclidean geometries. Russell, however, subsequently almost
fully abandoned the Kantian perspective; he corresponded with Frege and made
important contributions to the development of logical theory. Carnap, while he made
no original contributions to logic, studied with Frege and fully absorbed the new
logical methods, and like Russell almost entirely abandoned the Kantian perspective.
Russell and especially Carnap saw in Frege's mathematical reconstruction of the idea
Realism and the Nature of Theories
113
of proof a model for all of philosophy. Philosophy would use the tools of modern
logic to reconstruct all of the notions of scientific methodologythe notion of a
scientific theory, of testing, confirmation, explanation, prediction, and more. Rather
than talking in Kantian terms about "synthesis" or in Lockean terms about "ideas"
philosophy would talk about language, about symbols, their mathematical relations
and their meanings.
This new logical perspective led to a program to replace Kantian epistemology
with other solutions to the problems of skepticism, especially metaphysical skepticism. One of the first efforts of the new perspective was to attempt to replace the
Kantian synthesis of the world of experience by a logical construction of the world
of experience. The idea was something like Locke's except that rather than talking
vaguely about combining ideas, certain parts of a formal language would be assumed
to describe the deliverances of experience, and then new terms would be introduced
into the language through formal, logical definitions that reduced the new terms to the
terms denoting immediate experiences (see Chapter 4, Sections 4.2 and 4.3). On this
view, the external world is whatever satisfies these logical constructions, and the
structure of definitions guarantees that what satisfies the constructions, if anything,
must be sets of immediate experiences, or sets of sets of immediate experiences, or
sets of sets of sets of immediate experiences, and so on. In this way any intelligible
talk of the objects of science, the external world, or anything else is reduced to talk
about possible experiences. Metaphysical skepticism was avoided in something of the
same way it was avoided in Kant, and the possibility of empirical knowledge was
explained. Russell sketched such an account and around 1925 Carnap attempted to
carry it out in detail in a book he entitled The Logical Structure of the World (1969).
A similar viewpoint was developed near the same time by the American philosopher
and logician C. I. Lewis in his Mind and the World Order ([1929] 1956).
Kant's system provided a kind of explanation of mathematical knowledge. How
did the new logical turn of the 1920s explain our knowledge of arithmetic and
geometry? Not by any attempt to establish that some mathematical knowledge is
synthetic a priori. The new philosophers of science unanimously rejected Kant's
claims about the source and character of mathematical knowledge; instead, several
other alternatives were pursued. One idea was logicism, the proposal that just as
claims about the world can be reduced by appropriate definitions to claims entirely
about experience, so claims in mathematics can be reduced by appropriate definitions
to claims about nothing and everything. By appropriate definitions of mathematical
objects and operations, mathematical claims were to be reduced to complex logical
truths. Logical truths are true of everything, true in every possible world, and they
can be established by purely logical demonstrations. In this way the certainty and a
priori character of mathematics would be accounted for, but only by making mathematics in a sense empty of content. This proposal was carried out in considerable
detail by Bertrand Russell and Alfred North Whitehead (1925). A second idea is that
mathematical knowledge is tacitly conditional and axiomatic. Many mathematicians
had shown how to give rigorous sets of axioms for various fundamental mathematical
theories. In the nineteenth century Giuseppe Peano had given a set of axioms for
arithmetic, and later Hilbert ([1909] 1971) gave a rigorous set of axioms for geometry. The axiomatic idea about mathematical knowledge was that we don't in fact
114
know a priori that the axioms of arithmetic are true of anything; what we know is only
that if the axioms of arithmetic are true of a system then certain other claims must
necessarily also be true. We don't know a priori, for example, that 1 + 1 = 2 is true
of any particular system, and sometimes it isn't: 1 volume of water mixed with 1
volume of alcohol does not make two volumes of anything because the two liquids are
mutually soluble. What we know a priori is that // Peano's axioms describe any
system of quantities, then 1 + 1 = 2 is true of them. The same is true of geometry.
As Einstein put the point, insofar as geometry is about experience it is not certain, and
insofar as geometry is certain, it is not about experience.
The philosophical movement Carnap represented came to be known as "logical
positivism." That term was applied mostly by German and Austrian philosophers to
their point of view, but as we have already noted the perspective was shared by others
in England and the United States. The logical-linguistic approach to scientific knowledge underwent a rapid evolution, caused largely by various technical difficulties
encountered by attempts at logical reconstruction of human knowledge. One of the
very first difficulties was the failure of Carnap's attempt to show how the external
world could be constructed logically from descriptions of experiences. Carnap found
that he could construct classes reducible to a simple relation between terms denoting
experiences, and that some of these classes seem reasonable reconstructions of sensory modalities such as color and sound. But he did not find a satisfactory way to
interpret the notion of a physical object or of places and times as such reducible
classes. Metaphysical skepticism seemed to threaten once more; fortunately Carnap
and the other logical positivists had a stick with which to beat it down.
3.5 THE VERIFIABILITY PRINCIPLE OF MEANING
Locke had claimed that we can only think about ideas, and he held that the source of
ideas must be either sensation or internal observation of the operations of our own
minds. In a fashion not entirely welcome to Locke, this point of view contains a
solution to metaphysical skepticism, since on Locke's view we cannot really think of
objects in themselves, only of ideas. What we cannot think of we cannot intelligibly
talk of. So when we ask whether the world in itself could be made of nonentities
without primary qualities, we are not, on this view, really saying anything intelligible. For metaphysical skepticism and other perplexities, the logical positivists proposed a still more radical solution: the verifiability principle.
The verifiability principle holds that a claim is meaningful if and only if it
could be verified, that is, if and only if some possible set of observations exists
that, were they to be made, would establish the truth of the claim (see Chapter 4,
Section 4.4). A claim that is actually false can be verifiable in this sense. It is false
that I weigh less than 200 pounds, but a possible sequence of observations exists
that were they to be made would establish that I do weigh less than 200 pounds.
Because I don't weigh less than 200 pounds, no such observations will in fact be
made, but they are possible.
By embracing the verifiability principle, the logical positivists were able to
dismiss metaphysical skepticism and metaphysical speculation as sheer nonsense, as
Realism and the Nature of Theories
115
literally talk with no meaning, for metaphysical skepticism only gets started by
attempting to describe some circumstance whose truth or falsity allegedly is underdetermined by all possible evidence. According to the verifiability principle, no such
circumstance can ever be described; words do not have the power to express what is
meaningless. (For an example of the application of the principle in a scientific
context, see Chapter 5, Section 5.1).
With the verifiability principle available, the failure of Catnap's attempt to show
how to construct descriptions of the world from descriptions of experience did not
herald the rebirth of metaphysical skepticism, and the principle is still held today by
a few philosophers. But the verifiability principle seemed to create as many problems
as it solved. The problem, once more, was Plato's.
Consider whether the claim that the melting point of water is zero degrees
celsius can be verified. If "verify" means to establish the truth in such a way that
we need not reserve the right to change our minds, then Plato's criterion for success in the knowledge game seems to apply: To verify the proposition there must
exist a finite sequence of observations such that upon making those observations we
can be certain that the proposition is true. So the observations must necessitate the
truth of the proposition, or in other words, the demon must not have available a
possible world in which the observations are true and the proposition in question is
false. However, we cannot verify in this way any general scientific law. We cannot, for example, use observations of the melting point of water to verify that water
melts at zero degrees celsius because it is logically possible (even if we do not
believe it) that in the future water will melt at 10 degrees celsius. Thus, according
to the verifiability principle, understood in this way, every scientific law is meaningless. The verifiability principle saves the world from metaphysics, but at the
price of losing the world!
Some, but not many, philosophers have been willing to accept this conclusion.
A number of philosophers educated at Oxford endorse it (Dummett 1978, Wright
1980.) Stephen Toulmin, a prominent philosopher of science, proposed for example
that scientific theories are not really claims about the world, they are "inference
tickets" that lead us to predict new observations from old observations. Toulmin's
idea is a form of instrumentalism, much like the view proposed by Osiander, the
author of the Copernican preface. In recent decades, however, most people concluded
that the verifiability principle is not acceptable. That conclusion is easy to draw when
we reflect that no independent arguments for the principle derived from any empirical
analysis of how language actually functions. The verifiability principle seems simply
to have been a convenient dogma for discrediting metaphysical perplexities. The
perplexities won.
3.6 CONFIRMATION, MEANING AND THE "STANDARD
CONCEPTION" OF THEORIES
With the failure of attempts to show how the world could be viewed as a logical
construction from simple descriptions of experience, and the rejection of the verifiability principle, the problem of separating science from metaphysics began to seem
all the more difficult. Kant had solved the problem of metaphysical skepticism by
116
confining the claims of empirical science to the world of experience. The first attempts of the new logical methods to replace the Kantian system with a clearer, more
rigorous, more "scientific" understanding of knowledge resulted either in a failure to
establish how the world of experience could be known at all, or else a failure to
separate the claims of science from those of metaphysical speculation, with all of their
underdetermination. To this dilemma philosophers of science in the 1930s responded
by using the notions of testing and confirmation to limit meaning. In the end that
proved not to help matters much.
One clever solution to the problem of metaphysical skepticism was proposed by
Hans Reichenbach (1938): Even if they are expressed differently, two claims mean
the same thing if every possible piece of evidence for (or against) one of the claims
is equally strong evidence for (or against) the other claim, and vice versa. The usual
examples of metaphysical skepticism arise when there appear to be two or more
claims that are indistinguishable in this way by any possible evidence. On Reichenbach's proposal, making a case for metaphysical skepticism by articulating two
claims that are indistinguishable by any possible evidence is actually making a case
that there is no difference in the claims: There is nothing between them to be in doubt
about. Reichenbach viewed confirmation as a matter of probability relations: Evidence confirms a hypothesis by raising its probability in accordance with Bayes' rule
(see Chapter 2). In his view, two theories say the same thing if they would receive
the same probability on every possible observation.
As with the verifiability principle no particular argument arose for this proposal
founded on how language is used, but important technical difficulties did appear.
Reichenbach had proposed an outline of a characterization of synonymy: A is synonymous with B just if A and B receive the same confirmation on every piece of
evidence. Since this synonymy relation depends on an initial distribution of probabilities, there is no way to be sure just what synonymy relation Reichenbach had in
mind. Suppose we separate from a formal language a special part or sub-language in
which observations are to be reported, and suppose in Reichenbach's spirit we say
that two theories are synonymous if and only if they entail exactly the same collection
of sentences in this observation language. Frege's theory of proof, then, no longer
applies, and indeed no effective theory of proof is possible within such a language.
So a comparatively clear variant of Reichenbach's proposal would destroy one of the
very pillars of the logical movement.
Reichenbach, as well as Hermann Weyl ([ 1949] 1963), the great mathematician
and mathematical physicist of the early part of this century, Carnap and Reichenbach's (and Hilbert's) student, Carl Hempel, developed another perspective. The
perspective presented in this chapter follows Carnap's, but all of these philosophers
developed very closely related ideas.
In Carnap's picture we can observe ordinary objects and their ordinary properties; we are not separated from the world by a veil of ideas. However, a great deal
exists that we cannot observe: events inaccessible in space-time, things too small, and
differences in properties too slight for our senses. Science is concerned with both the
observed and the unobserved, and indeed even the unobserva&fe. We describe what
we observe in language and we state our scientific theories and hypotheses in language as well. Both aspects of language can be formalized. Scientific theories and
Realism and the Nature of Theories
117
their claims about the world can be represented in a formal language of any of the
kinds Frege, Russell, Hilbert and other logicians developed. The difference between
that part of the formal language which represents claims about observations and the
more theoretical parts of the formal language are chiefly differences in vocabulary.
Many of the symbols that represent the language of science' 'photon,'' ' 'molecular
bond," "gene," and so onare not symbols that denote observed things or properties or events.
By making observations and conducting experiments we can obtain descriptions of particular facts; these descriptions are represented in a formal language as
singular sentences in the vocabulary of observation terms. How such facts might be
used as evidence for or against generalizations that are expressed purely in observational terms seems clear enough. If the generalization is, for example, that all
water melts at zero degrees celsius, then observations of water melting at zero degrees celsius count for the generalization, at least until we find a sample of water
that does not. Carnap (1936) proposed that singular sentences describing observations stand in a relation of confirmation to general claims. The sentence, "This
sample of water melted at zero degrees celsius," confirms "All water melts at zero
degrees celsius." But how are we to confirm generalizations that are not limited to
observation terms? How, for instance, are we to confirm that the atomic weight of
oxygen is greater than the atomic weight of hydrogen? Carnap's answer to this
fundamental question was a form of conventionalism in which the conventions
specify meanings. Theoretical terms, Carnap suggested, are literally meaningless
until some conditions or rules are given that specify what observable circumstance
would count as determining instances of the terms. Carnap called such rules meaning postulates. This principle is weaker than the verifiability principle; Carnap did
not require that meaningful sentences be verifiable, he required only that they be
composed of predicates for which meaning postulates exist (see Chapter 4, Sections
4.3 and 4.4).
Meaning postulates look very much like other generalizations. Although Carnap proposed some restricted logical forms for meaning postulates, what is special
about them is their role and their justification. Their role is both to establish the
meaning of theoretical terms and to permit the confirmation of theoretical claims.
They permit confirmation of theoretical claims by enabling us to infer instances of
theoretical generalizations from singular observation statements. If, for instance,
the law of Dulong and Petit were taken as a meaning postulate, then we could use
it to infer values of atomic weights from measurements of heat capacities. In this
way we could confirm the claim that the atomic weight of oxygen is greater than
the atomic weight of hydrogen. The justification of meaning postulates is that they
are simply stipulations, conventions about how we will use theoretical terms. They
are therefore, in Carnap's terminology, analytic truths, while the claims of science
that are not either truths of logic or consequences of meaning postulates are synthetic claims.
Contemporary commentators sometimes refer to variants of this conception of
the structure of scientific theories as the "standard conception." Its elements are
the assumption that scientific language can be formalized, a division between observation terms and theoretical terms, a set of meaning postulates or stipulations
118
relating the two sorts of terms, and an analysis of confirmation in terms of relations
among sentences in the same vocabulary. The standard conception is the closest
modern philosophy of science ever came to a logical successor to the Kantian picture. The problems of metaphysical skepticism are solved in the standard conception by noting that the terms in metaphysical disputes lack adequate meaning
postulates to permit the confirmation of disputed metaphysical claims. A picture of
how scientific knowledge is possible emerges: Meaning postulates constrain scientific language, and observations generate observational reports that confirm or disconfirm scientific hypotheses.
3.7 REALISM AND THE LIMITS OF SCIENTIFIC KNOWLEDGE
The standard conception set the framework for logicians to investigate one of the
fundamental Kantian questions: What are the limits of possible scientific knowledge?
The questions at issue are exactly those that concerned Kant, but now the logical
conception permitted logicians to find answers that have real demonstrations rather
than Kant's intuitive but inconclusive arguments. This work was begun around 1960
by Hilary Putnam (1965), and has been continued since by a number of logicians and
philosophers.
One way to understand such questions is to consider again the knowledge
games an imaginary scientist might win against a demon, assuming the standard
conception. The demon and the scientist agree on a division between observation
and nonobservation terms, and on a set of meaning postulates, and the demon has
a set of possible worlds from which to choose. In each of these possible worlds the
meaning postulates must be true, but everything else may vary between worlds.
Whatever world the demon chooses, the demon must give the scientist the observable facts of that world in some order, and every observable fact must eventually
be given. The scientist must conjecture whether some claim is true or false. We can
adopt Plato's criterion of success for the scientist or the weaker criterion suggested
by Kant's antinomies of reason. We know that the scientist cannot win in any
interesting case if we adopt Plato's criterion of success, in which no matter what
the demon's choice the scientist's conjectures must eventually converge to the truth
and the scientist must know when she has converged to the truth. But what if we
grant that the scientist wins even if the scientist may not know when she had converged to the truth? Then the question of knowledge is more interesting. In that
case, assuming a formal language of the kind developed by Hilbert and Ackermann, whether the scientist can win depends entirely on the logical properties of
the claim in question and the logical properties of the meaning postulates. The
scientist can win just if the meaning postulates entail that the claim to be decided
is true if and only if purely observational claims of a special logical form are true.
With the aid of the meaning postulates the claim under investigation must be reducible to a claim of the form, "For every x, y, . . . , z, there exists a u, w, . . . ,
v, such that so and so," where the "so and so" contains only observational terms
and does not contain any expressions (called quantifiers) such as "for every" or
"there exists." Furthermore, for the scientist to win, the meaning postulates must
Realism and the Nature of Theories
119
also entail that the claim under investigation is reducible to another purely observational claim with the quantifiers reversed, that is, to a claim of the form, "There
exists an x, y, . . . , z such that for all u, w, . . . , v, so and so." 1
Using the standard conception logicians have been able to obtain similar results
for a wide range of alternative formal languages and for problems in which the
scientist must discover an entire theory rather than settle a preestablished question.
Such mathematical analyses can be obtained even when the characterization of possible observation sentences is not by vocabulary, when all the evidence the scientist
will see is incomplete, and even when the observation sentences are not limited to
statements of particular facts. If we want to go to the trouble, such logical results
enable us to say precisely what questions are or are not beyond human knowledge
given a definite logical system, a set of meaning postulates, and a characterization of
which sentences can be decided directly by observation.
Given the elements of the standard conception, logical analyses of the limits of
knowledge give rather pessimistic results: Relative to the logical complexity of the
observation sentences, any claim of science that could be known even in the limit
must be equivalent to sentences with specially restricted quantifier structure. In addition, of course, what cannot be known in the long run cannot be known in the short
run either. On plausible accounts of what we can actually observe, much of what we
think we know we could not possibly know. Kant's problem of the antinomies of
reason finds a modern form.
Even without detailed logical analyses, by looking at particular scientific theories and enterprises many philosophers sensed that much of scientific belief would
prove on close examination to be underdetermined by observation, not just in the
short run but even in the long run. Arguments of these sorts were developed in
different ways for features of geometry and physics by Reichenbach ([1928] 1957),
Salmon (1975), Griinbaum (1973), Glymour (1977), Malament (1977b) and others,
and for sciences that depend on deciphering meanings by Quine (1960) and Davidson
(1984). These difficulties have been met with at least three different responses.
One response to the apparently strict limitations on the possibility of scientific
1
The following procedure will determine in the limit that "There exists an x such that for all v F(;t,y)"
is true (if it is) and (if it is not) will otherwise not converge to any hypothesis: Let x0, xx . . . xn . . . be an
enumeration of all of the objects that could be a value v of x for which "for all y F(v,y)" is true. Start by
conjecturing yes to the claim that' 'There exists an x such that for all y F(j>c,y)" so long as the data are consistent
with "for all y F(x 0 ,y)." If a datum is received that contradicts "for all y F(x0, y)" conjecture no to the
question. Thereafter, conjecture yes provided all evidence so far received is consistent with "for all y F(x1 ,y)''
until this is contradicted, at which point no is conjectured and the procedure moves to "for all y F(.*2,y)" and
so on. If' 'There exists an x such that for all y F(x,y)" is true, then the procedure eventually finds an xn for which
"for all y F(xn,y) is never contradicted by the data and so the procedure says yes ever after. If "There exists
an x such that for all y F(jt,y)" is false, then for each xn "for all y F(x,y)" is eventually contradicted by the
data, and the procedure changes conjectures infinitely often.
If the denial of "There exists an x such that for all y F(x,y)" is logically equivalent to a "There exists
an x, for all y such and such" formula, then a similar procedure can be applied. The two procedures, one for
the original sentence and one for its denial, can be dovetailed into a single procedure that converges to yes if
' 'There exists an x such that for all y F(jt,y)" is true and converge to no if the denial of it is true. It can be shown
that a sentence not equivalent to a sentence of the form "there exists an x, for all y, such and such" cannot have
its truth verified in the limit.
120
121
away from asking questions about the limits or foundations of scientific knowledge,
for the product of science in this view is not so much knowledge of general propositions but as an understanding of systems of models and how to embed various
classes of phenomena within those models. On this conception we may accept and use
a scientific theory, but see no sense or point to believing the theory. Differing versions
of this approach have been developed by a number of philosophers of science including Suppes, van Fraassen (1980), Suppe (1989), Sneed (1971), Giere (1988) and
others.
The semantic conception certainly captures one important aspect of how theoretical conceptions are deployed in science. It is not, however, clear that the semantic
conception solves any of the fundamental traditional problems facing philosophy of
science. In some cases, such as cosmology, there may be no difference between
saying that a theory is true and saying that it is true of some system. Again, it is
plausible enough to treat Newtonian dynamics as a predicate of systems, or as a class
of models that vary in details, but it seems somehow less plausible to treat the atomic
theory in chemistry that way. More importantly, even on the semantic conception,
one of the aims of science is to predict the course of phenomena, and corresponding
to the limitations of knowledge there are limitations of reliable prediction. In various
settings precise theorems can be demonstrated about such limitations, even in the
limit of arbitrarily increasing evidence.
A third response to underdetermination, advocated by Russell (1948), Quine
(1969b) and several other philosophers is the conception of naturalized epistemology.
Granted that starting from nothing but our observations, without any "background
knowledge" about how the world works, we could come to know only a very
restricted body of claims about the world. Granted that in games with the demon, if
the scientist does not know beforehand that the collection of possible worlds the
demon might choose is very restricted, then the scientist can only win (even in the
more generous sense that does not require that the scientist know when he is right) for
a very limited set of questions. Nonetheless, because of how we are in fact constructed biologically and socially, we do not start inquiry utterly ignorant. We have
evolved to favor certain behaviors and to organize our sensations in particular ways.
Unless hindered in some serious way, infants rapidly learn to identify and reidentify
objects, and they learn that objects continue to exist when unobserved; infants have
available almost from birth some simple facts about size, distance and perspective.
And so on. Society provides us with a great deal of belief about the world, about other
people and about social relations. (Of course some of what society leads us to believe
is erroneous.) Assuming as background the beliefs thus bestowed on us, we can study
our own perceptual and cognitive apparatus and our own social structures to discover
how beliefs are formed and to determine the reliability (and unreliability) of the
processes of human cognition. Such inquiries lead us to devices and procedures to
improve the way we acquire new beliefs, and they may even lead us to modify or
abandon some of the beliefs common to our culture. Assuming the beliefsor most
of the beliefsto which we are disposed by nature and culture, we proceed to expand
scientific knowledge. Our faith is that in the game we play against nature's demon,
the demon has available only those possible worlds in which our background beliefs
are true.
122
Naturalized epistemology is a program for snatching victory from defeat. Metaphysical skepticism wins, true enough, but in practice we are not and could not be in
the circumstance of trying to learn from pure ignorance. Assuming the circumstances
we cannot help but believe we are in, the powers of science to expand our knowledge
are increasednature's demon has fewer ways to deceive us. Logical and statistical
investigations of the possibilities and limitations of scientific inquiry remain valuable
critical tools for determining, relative to our background knowledge, which scientific
projects and programs can hope to solve which questions. When epistemology is
naturalized, normative philosophy of science remains a most important enterprise.
3.8 MEANING, OBSERVATION AND HOLISM
In the last twenty years the standard conception and its solutions to the problems of
metaphysical skepticism and justification of scientific belief have come under heavy
criticism. One criticism of the standard conception is that there appears to be no
objective feature to tell us which scientific claims are meaning postulates. No mark
of scientific hypotheses tells us that they are meaning postulates that are not subject
to empirical disconfirmation. Even a principle that is explicitly introduced as a stipulation about how a new term will be used may, as observations accumulate, come
to be abandoned in favor of other claims that historically were confirmed by assuming
and using the putative meaning postulate. As Quine put it, any claim can be held true
come what may.
Curiously, both Reichenbach and Carnap had anticipated part of this criticism.
They had viewed the selection of particular claims as meaning stipulations as a
somewhat arbitrary part of the logical reconstruction of science. Their view seems to
have been that in order to understand how scientific inference and argument works at
any moment, some claims must be understood to be functioning as meaning postulates at that time. The meaning postulates may change over time, but these changes
are themselves alterations in the meanings of scientific terms, not alterations in the
content of scientific claims. The picture for which Quine argued was subtly but
importantly different. At any moment any accepted scientific claim might be used to
justify any other accepted scientific claim; except for rare usages, such as abbreviations, there is not even at a given time any distinction between the two.
Quine's criticisms generated two lines of thought, both premised on the assumption that he was correct that no sentences function simply as meaning postulates.
One line of thought tried to save the achievements of the standard conception by
explaining how scientific claims could be tested and confirmed even without meaning
postulates. The other focused on meaning rather than evidence. Without meaning
postulates, how do theoretical terms get their meaning? What determines the conditions under which theoretical terms correctly apply from the conditions under which
they do not correctly apply? One answer went roughly as follows: A theoretical term
is generally introduced in order to describe the cause of some aspect of a body of
experimental or observational phenomena. The theoretical term then denotes whatever it is (if it is anything at all) that plays that causal role for that aspect of the
phenomena. It may be that the aspects of several phenomena for which a term is
Realism and the Nature of Theories
123
introduced have several causes, in which case the theoretical term may "partially
denote" each of these features rather than one of them (Field 1973). A more radical
view was championed by Quine himself: A theoretical term gets its meaning from the
entire body of scientific beliefs in which it is embedded. This view, often called
meaning holism, further denies any distinct basis for meaning as against belief. It has
the consequence that if beliefs about a collection of sentences change, so do the
meanings of the terms in the sentences. This doctrine, or rather the phenomenon it
allegesthat whenever scientific opinion changes the meanings of all scientific terms
changeis sometimes called meaning variance,
Just as no clear mark seemed to separate meaning postulates from other scientific claims, so on reflection no clear mark seemed to separate what is observable from
what is not observable, and certainly no distinguishing mark seemed suitable for
founding a theory of meaning. Depending on details of context and scientific belief,
almost any sentence in the language of our theories might serve as a report of
observations. For example, a scientist who believes that the law of Dulong and Petit
is approximately correct might make measurements with a calorimeter and report, "It
was observed that the atomic weight of hydrogen in our samples is less than the
atomic weight of oxygen in our samples." A physicist checking photographs of
particles from an accelerator might say, "A Z particle was observed." In practice we
judge the correctness of reports of observations by whether the circumstance claimed
to be observed really did occur, and by whether the observers were situated so that
they could have distinguished the circumstance from others, for example, from the
absence of a Z particle in the picture. Neither of these requirements is met only by
sentences in some special "observational" vocabulary. Moreover, how observations
are reported isas Locke notedvery sensitive to the beliefs of the observer. A
scientist who holds one theory may honestly describe experimental observations
differently from the way the outcomes of the same experiment are honestly described
by a scientist who holds a competing theory. However, because of the doctrine of
meaning holism, no neutral formulation exists of what has been observed. Observations are unavoidably theory laden.
Sometimes the freshest lines of thought in a discipline come from those outside
it who sense the larger issues without being entangled in the detailed arguments of
specialists. The attack on the standard conception was led by Thomas Kuhn (1970),
who at the time had studied physics and the history of science, but not philosophy.
Kuhn argued through historical examples that occasionally radical breaks arise in
scientific tradition when practitioners follow different lines of research using different
standards of argument and making different inferences from observations. Kuhn
claimed that the practitioners separated by such a break literally do not understand
each othertheir scientific languages are incommensurableand they share no methods for resolving their disputes. Between sides in scientific revolutions, success is
determined by rhetoric and politics, not by evidence. Moreover, science does not
accumulate truths as time passes. When a scientific break occurs, all of the results
accumulated by the previous tradition are rejected; either they are dismissed altogether by the new tradition (or paradigm as Kuhn called it) or else they are given a
new meaning.
Much of the philosophy of science in the last twenty years has centered around
124
arguments over whether to abandon the standard conception entirely, or to save its
essential elements by arguing, for example, that the doctrines of meaning holism and
meaning variance are false exaggerations, that characterizations of observations exist
that are neutral between competing theories, and that Kuhn's and others' historical
accounts are incorrect. Rather than entering into the details of these debates (see
Chapter 4), let us consider some responses to abandoning the standard conception.
3.9 HOLISM, RELATIVISM AND SOCIETY
One of the first consequences of rejecting the standard conception is a problem about
the relevance of evidence. Abandoning meaning postulates meant that confirmation
relations involving theoretical terms could no longer be specified for an entire language of theoretical and observational terms. Instead, philosophers of science attempted to characterize confirmation relations that are relative to a theory. Different
theories will generate different confirmations of hypotheses from the same set of
observation statements. This line of thought was undercut by its fundamental inability
to resolve the issues about the limits of knowledge, the very questions that had
motivated the entire philosophical development leading to the standard conception.
One could, for example, analyze logically which claims a scientist could discover in
the limit assuming a theory, and the answer was essentially the same as in the
standard conception but with the entire theory serving in place of a system of meaning
postulates. Of course the truth of entire theories is not guaranteed by stipulations
about meaning, so how can we know which theories to use? If no theories are
assumed then scientists could only win games in which the claim to be decided had
the appropriate logical form and no theoretical vocabulary.
Once we abandon the view of the standard conception that the relevance of
evidence can be localized, that an experimental outcome bears for or against particular hypotheses but not others, scientific claims appear to become ever more underdetermined. This was the conclusion of Quine (1961), and of Pierre Duhem ([1906]
1954), a distinguished physical chemist who wrote before the standard conception
had been articulated. Duhem claimed that in any modern physical experiment virtually the whole of physics is involved; the design of the apparatus and interpretation
of its output may involve not only the hypothesis supposedly to be tested, but also
principles of optics, thermodynamics, electronics and hence electromagnetism, and
so on. If the experiment does not give the expected result, the hypothesis to be tested
might be blamed, but from a logical point of view any of these other physical
principles might be faulted instead. The much disputed claim that no principled way
exists to localize the bearing of evidence is often called the Duhem-Quine thesis.
One use of the Duhem-Quine thesis is for what philosophers of science call the
pessimistic induction, an argument to the effect that the only reasonable conclusion we
can have about the present claims of science is that they are false. The argument
(Laudan, 1984) is as follows: Every theory we can name in the history of science is,
in retrospect, erroneous in some respect. The Newtonian theory of gravitation is incorrect, as is the classical theory of electromagnetism, Dalton's atomic theory, classical physical optics, the special theory of relativity, the Bohr theory of the atom, and
Realism and the Nature of Theories
125
so on. The errors of these theories may not matter for most practical purposes, but from
a contemporary point of view they are all, strictly, false theories. Since all theories in
history have been false, the argument continues, we should conclude that the methods
of science do not generate true theories; hence our present scientific theories, which
were obtained by the same methods, are false as well.
It is important to see that whatever plausibility the pessimistic induction has
depends on rejecting the standard conception of theories and invoking the DuhemQuine thesis or something very much like it. When we look back at historical cases
of false theories, we note that parts of the theory are false, but other parts seem less
so. Dalton's atomic theory, for example, contained claims that heat is a fluid, "caloric," and that all atoms have the same volume. In retrospect, these claims seem
false. Dalton also held that all atoms of the same element have the same weight,
which also seems false, but less so, and in fact nearly true. Dalton also claimed that
all molecules of a compound contain the same number of atoms of each element, and
that seems even today entirely correct. Just as we informally separate parts of Dalton's theory as more and less correct, we also informally weigh how sufficient his
evidence was for these various parts of the theory. In retrospect Dalton's evidence for
the caloric theory and for the hypothesis about atoms seems insubstantial, and the
evidence for the constitution of molecules and the sameness of weights seems better.
These separations make some sort of sense in the standard conception, but they are
nonsensical on the holistic view of evidence.
Another response to the decline of the standard conception sentiments is represented by the sociology of knowledge movement. In its strongest form (indeed,
sometimes called the strong program) the view advocated by some sociologists and
cultural anthropologists is that the content of science is entirely an artifact of elaborate
social customs whose function is often hidden from scientific practitioners themselves. Scientists, on this view, do not really discover anything, although they may
think they do. Instead science is an elaborate and expensive social system for deciding
what to say, how to talk about the world, and for making social decisions about
technical matters (whether to build a nuclear power plant, for example). The scientific
community makes such decisions essentially on political grounds; some people have
more authority, more influence and more power than others, and so long as they retain
these advantages the scientific conceptions they hold are deferred to. There is no
normative difference whatsoever between the claims of science and the claims of any
religion or religious group; scientific claims have no more warrant, no more justification, no greater rationale, than the claims of Islamic or Christian sects, or than
flat-Earth theories or astrology.
Few philosophers of science have much sympathy with the strong program, but
its viewpoint should be at least be correctly understood. The position is not (or at least
need not be) that the only facts are social facts. Undoubtedly a great many facts exist
that are not social, but exactly because science is a social enterprise it is claimed to
be incapable of giving any warrant that its claims are correct.
The doctrines of meaning variance and meaning holism combined with the
rejection of the linguistic observational-theoretical distinction result in a view in
which the success of science becomes very mysterious: Meaning changes with belief,
and the doctrine of meaning variance applies as much to reports of observations as to
126
other claims in science. Scientists who hold different opinions do not share the same
meanings, even if their differences appear to be remote from the subject under
discussion. Competing scientific theories, it seems, cannot be tested against a common standard of observations since what the observations mean depends on which
theory is believed. No means seem to demonstrate when scientific inquiry can and
cannot reliably lead to the truth. When both meaning and the world of experience vary
with the beliefs or conceptual scheme of the observer, the great philosophical issues
about the possibility of knowledge cannot be answered because they cannot sensibly
be formulated. Philosophy of science (and some would say, philosophy generally)
comes to an impasse in which there is nothing that can be done; there is nothing for
philosophy or for philosophy of science to discover about how to reliably acquire
knowledge of the world, or about the limits of such knowledge. This is exactly the
conclusion reached by some prominent philosophers, Rorty (1979) for example, who
have followed and accepted the attacks on the standard conception, and who recommend giving up the pursuit of philosophical questions about knowledge.
Separately from the arguments over the standard conception of theories, modern
quantum mechanics has given some philosophers of science reason to think that we
cannot acquire any knowledge of the world independently of variable features of
ourselves. We cannot because there is no such world.
The quantum theory includes both dynamical variables such as position, momentum, time and energy, and states of a physical system. The theory does not,
however, include any physical states in which all dynamical variables have precise
values. Any state in which position is perfectly precise, for example, is a state in
which momentum is completely indeterminate. The quantum theory restricts how
precisely any allowed state can specify both position and momentum and similarly
both time and energy. These restrictions are the famous Heisenberg Uncertainty
relations. (See Chapter 6, Sections 6.11-6.16 for in-depth discussion.)
A natural response to the quantum theory is to think that microscopic systems
have precise values for all of their dynamical variables but the theory simply is unable
to determine these quantities precisely. While this view has its supporters, a great deal
of empirical and mathematical research has led many physicists and philosophers of
science to think it is false. Instead, systems literally do not have a simultaneous
precise value for position and momentum; however, if we conduct an experiment to
measure position we find one, and if we conduct a measurement to measure momentum we find one. What properties the world exhibits depends on what we ask of it.
Niels Bohr called this phenomenon ''complementarity,'' and the conclusion he
drew was a revision of Kant's perspective. In the psychological version of the Kantian
picture is a fixed but unknowable world in itself and a fixed "us," and the world we
experience is determined by the two together. In Bohr's picture, an unknowable and
undescribable world in itself exists, as well as "us," but the " u s " is not fixed. We
can ask one set of questions about the world and get a coherent set of answers, or we
can ask any of many other sets of questions about the world and in each case get
Realism and the Nature of Theories
127
coherent answers, but the coherent answers to the several sets of questions cannot be
fitted together into a single coherent picture of the world. Changing the experiments
we conduct is like changing conceptual schemes or paradigms: we experience a
different world. Just as no world of experience combines different conceptual
schemes, no reality we can experience (even indirectly through our experiments)
combines precise position and precise momentum.
3.11 CONCLUSION: REALISM, RELATIVISM AND PHILOSOPHY
OF SCIENCE
The most immediate connection between philosophy of science and the rest of philosophy is through issues about the limits to knowledge and the character and possibilities for justified belief. For a brief while a consensus among the best philosophers
of science pursuing the logical program provided a clear framework for exploring
these issues and obtaining interesting answers. That consensus has disappeared in the
closing years of the century. The community of philosophers of science is fragmented
among those who regard the general perspective of the standard conception of theories as correctalthough most would modify it in some important wayand those
who reject it for various alternatives. The most radical, and at least in one sense the
most interesting, fragment argues that because meaning varies with belief and conceptual scheme, and because no linguistic characterization of the observable exists,
the epistemological questions that have motivated the enterprise of philosophy of
science are unanswerable. Locke's picture, Kant's and Carnap's all ask what can be
known by science assuming some thingssuch as meaningsare fixed. The radicals,
however, claim that those things are not fixed, and so the questions have no answer.
Many, perhaps most, philosophers of science reject this relativist view, and the
debates over it and how best to mend or replace the standard conception remain
central to philosophy of science today. Suppose, however, that we accept the relativist view entirely: meaning and truth vary in some unknown way with aspects of
belief or with the experiments we choose to conduct; we observe a world but the
world we observe depends on features of our belief, culture, or experiments. Surprisingly, even in this radical picture questions about scientific realism and the limits
of scientific knowledge still make sense and have answers; the puzzles of Kant's
antinomies still survive. There is information about how best to conduct inquiry that
philosophy can still seek after and find.
Even if the relativist picture were correct, philosophers (and others) can investigate how meaning depends on belief, and what aspects of sense, reference and experience are altered by what changes in belief, custom or culture. Perhaps more
importantly, relativism is not subjectivism; that the world of experience depends in part
on variable features of us does not mean that we can have any world of experience we
wish. What we experience depends on us and on something not usKant's things in
themselves for lack of a better term. The world in itself and our conceptual scheme
together determine the world we experience, and thus what questions we can and cannot answer. If the world we experience depends on some feature of us and our
communitycall that feature our conceptual schemethen there are logical facts
about what questions can and cannot be answered in a given conceptual scheme, and
128
about which questions can be answered by changing conceptual schemes. Formal models of inquiry exist in which certain questions can be settled in the limitscientists can
beat the demonbut only if scientists are free to alter conceptual schemes as need be.
In these models for some issues successful discovery requires scientific revolutions.
Even if relativism were true, the limits of knowledge could be investigated and better
and worse ways to conduct inquiry would arise. Philosophy of science endures through
relativism.2
DISCUSSION QUESTIONS
1. Suppose someone knew that two incompatible theories, call them A and B, are underdetermined
by all possible evidence of a certain kind, E: Any possible evidence of kind E is known to be
consistent with A and with B. Is there any interesting point to claiming that, nonetheless,
evidence of kind E confirms A rather than Bl If so, what point, and why is it interesting? Is there
any interesting sense or point to claiming that, nonetheless, evidence of kind E would justify
such a person in believing A rather than Bl What goals might be served by having a shared
confirmation relation that prefers one underdetermined hypothesis to another?
2. One theory about the notion of truth is that a claim that a sentence is true does nothing more than
reassert the sentence. According to this view, sometimes called the redundancy theory of truth,
the sentence, " 'The sky is blue' is true" simply asserts that the sky is blue. The claim that
''What Sam believes is true" simply asserts the otherwise unknown set of propositions believed
by Sam.
One advocate of the redundancy theory asks the question: Why is it practically useful to
believe the truth? Does this question even make sense according to the redundancy theory? Why
or why not? What expressions using the word "truth" are difficult to account for with a
redundancy conception?
3. Suppose after careful logical and psychological study it is found that the array of ordinary beliefs
(about ordinary objects, spatial and temporal relations, causal connections, and such) needed as
background knowledge for reliable scientific inquiry cannot themselves be reliably acquired
from any possible array of facts about elementary experiences such as those available to an
infant. What, in that case, should we say about our scientific claims and our ordinary beliefs
about the world? Would they be knowledge, dogma, what?
4. Suppose two methods of inquiry are exactly alike so far as their convenience and reliability are
concerned, except that method 1 will find out that A is true if in fact A is true, but method 2 will
not. Suppose you are convinced that A is not true. Is there any reason for you to prefer method
1 to method 2?
5. What could be meant by the claim that two scientists who speak the same natural language and
work in the same discipline literally do not understand one another's claims? What does understanding another person's claims require? Is it sufficient, for example, to be able to describe what
the other would say about any case or circumstance? Could there be good evidence that historical
figures in science who ascribe to different theories or ' 'paradigms'' literally could not understand
one another? What would such evidence be like?
6. Social approaches to epistemology emphasize the value of procedures for obtaining consensus
rather than the value of reliable procedures of inquiry. What ways of organizing rewards,
communicating opinions, and making resources available would further the goal of obtaining a
2
The general perspective of this chapterand the bananaowe a great deal to Kevin Kelly.
129
consensus of true opinions? Is science so organized? What evidence is there that the enterprise
of science has goals other than consensus regardless of truth value?
7. If you were in fact a brain in a vat all of whose experiences are illusions produced by computercontrolled stimulations of your nerve endings, what would your words * 'brain in a vat'' signify?
8. Could an ideal scientific theory, that gave the best possible explanation of all possible evidence,
nonetheless be false?
SUGGESTED READINGS
Some important sources from the philosophical tradition can be found in
M. (ed.) (1977), Classics of Western Philosophy. Indianapolis: Hackett.
HAMILTON, EDITH and HUNTINGTON CAIRNS (eds.) (1961), The Collected Dialogues of Plato,
Including the Letters. Translated by Lane Cooper and others. New York: Pantheon Books.
KANT, IMMANUEL ([1787] 1865), Critique of Pure Reason. Translated by Norman Kemp Smith.
New York: St. Martin's Press.
CAHN, STEVEN
Many important papers on the logical revolution and its applications to philosophy of science will
be found in
A. J. (ed.), Logical Positivism. New York: The Free Press.
VAN HEIJENOORT, JEAN (1967), From Frege to Godel: A Source Book in Mathematical Logic.
Cambridge, MA: Harvard University Press.
AYER,
Work on the logical investigation of discovery begins with Putnam's seminal paper and continues
to this day:
and KEVIN KELLY (forthcoming), Logic, Computation and Discovery. Cambridge, England: Cambridge University Press.
KELLY, KEVIN (forthcoming), The Logic of Reliable Inquiry.
OSHERSON, DANIEL N., MICHAEL STOB, and SCOTT WEINSTEIN (1986), Systems That Learn: An
Introduction to Learning Theory for Cognitive and Computer Scientists. Cambridge, MA:
Bradford/MIT Press.
PUTNAM, HILARY (1965), 'Trial and Error Predicates and the Solution to a Problem of Mostowski,
The Journal of Symbolic Logic 30: 49-57.
GLYMOUR, CLARK
Press.
(1980), Science Without Numbers: A Defence of Nominalism. Princeton: Princeton
University Press.
FINE, ARTHUR (1986), The* Shaky Game: Einstein, Realism, and the Quantum Theory. Chicago:
University of Chicago Press.
FIELD, HARTRY
130
GOODMAN, NELSON
PUTNAM, HILARY
Press.
VAN FRAASSEN, BAS
131