Stud. Hist. Phil. Sci.

33 (2002) 511533
www.elsevier.com/locate/shpsa

The complete Duhemian underdetermination

argument: scientific language and practice
Karen Merikangas Darling
Department of Philosophy, Northwestern University, 1818 Hinman Avenue, Evanston, IL 60208-1315,
USA

Received 2 May 2001; received in revised form 26 September 2001

Abstract

Current discussion of scientific realism and antirealism often cites Pierre Duhems argument
for the underdetermination of theory choice by evidence. Participants draw on an account of
his underdetermination thesis that is familiar, but incomplete. The purpose of this article is to
complete the familiar account. I argue that a closer look at Duhems The aim and structure
of physical theory (1914) suggests that the rationale for his underdetermination thesis comes
from his philosophy of scientific language. I explore how an understanding of physical laws
as symbolic is meant to support the thesis. In the course of my argument, I point out that
Duhemian underdetermination is not meta-practical but grounded in the practice of science,
specifically in the scientists use of instruments and measurement techniques. Measurement has
a significant limitation, according to Duhem: it always involves approximation and a degree of
experimental error. Consequently, it cannot overcome the gap between the ordinary, concrete
language of observation and the (abstract and symbolic) mathematical language of science.
Moreover, Duhem argues that the use of instruments in experiment invokes whole groups of
theories. I contend that, ultimately, this reliance on auxiliary assumptionswhich makes poss-
ible the use of instrumentsis the foundation of his thesis and that recognizing this completes
the familiar account of his underdetermination argument.
2002 Elsevier Science Ltd. All rights reserved.

Keywords: Duhem; Suppes; Underdetermination; Measurement; Physical theory

E-mail address: [email protected] (K. M. Darling).

0039-3681/02/$ - see front matter 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 3 9 - 3 6 8 1 ( 0 2 ) 0 0 0 2 7 - 4
512 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

1. Introduction

Some think that physical theory cant approach the world and meet it, cheek to
cheek. A mediating layer, they believe, necessarily separates the two. This belief
has motivated a family of arguments that intend to show the underdetermination of
theory choice by evidence. These arguments share the following conclusion: evi-
dence is, strictly speaking, incapable of dictating theory or hypothesis choice. Pierre
Duhem is the author of one well known version, which is often cited in current
discussions of realism and antirealism. From the observation that in physics the test-
ing of a particular theory or hypothesis necessarily makes use of auxiliary assump-
tions, he aims to demonstrate that the two methods that might logically determine
theory choicefalsification and verificationare insufficient. If this is the case, then
(presumably) pragmatic or other concerns enter into the scientific communitys
choice of which theory or hypothesis to accept. This idea is particularly attractive
to antirealists because it suggests that, properly understood, scientific theories are
not candidates for true depictions of the way the world is; rather, they are at best
one empirically adequate choice. Antirealists, then, may well appeal to Duhems
thesis in defense of their interpretation of science, for it appears to undermine realist
claims to the truth of scientific theories and the reality of the objects of scientific
investigation.
But appearance can be deceiving.1 The standard depiction of Duhems argument
lays out two conclusions. First, scientists cannot logically determine, from the results
of an experiment, whether the failure of an expected outcome to obtain mandates
the rejection of the hypothesis or the auxiliary assumptions. Second, scientists are
unable to confirm the hypothesis conclusively because (a) a disjunction could not
be constructed wherein all possible competing hypotheses save one could be elimin-
ated, and (b) no empirical generalization arises directly, and hence perfectly, from
observation. This characterization is familiar, but incomplete. To evaluate contem-
porary uses of Duhems thesis, we need to complete the picture. That is to say, we
need to look more closely at what drives the argument. What, on his view, allows
for the gap between theory and world?
It would be natural to suppose that the basis for Duhems belief in this gap would
be some practical aspect of his experience as a scientist. After all, Duhem was princi-
pally a physicisthe is reported to have said about returning to Paris as a historian:
I am a physicist. Paris will obtain me only as such, if I ever should return there.2
One approach to studying his underdetermination thesis could, then, begin with his
scientific writings. For example, they might show that Duhems thesis is supported
by his experience of needing essential idealizations in his scientific work.3 Another
option presents itself, however, and I will pursue it here.

1
Needham (1998, 2000) provides a nicely balanced treatment of realism and antirealism in Duhem.
2
This is quoted in Jaki (1984), p. 181. Incidentally, Duhem never did.
3
Mark Wilson offered an argument along these lines in Mechanics at the breakfast table, a talk
given at Northwestern University, November 10, 2000, and at the APA Central Division Meeting in
Minneapolis, May 5, 2001.
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 513

My approach is to locate the rationale for Duhems underdetermination thesis by

turning to where it is most fully elaborated, namely his major philosophical work
The aim and structure of physical theory.4 (This is a work that incorporates much
of his earlier writing in the philosophy of science.) There the rationale for underdeter-
mination seems to come from his philosophy of scientific language. That is to say,
Duhem provides an argument which, I take it, is intended to show that the under-
determination thesis is a consequence of the nature of scientific language, specifically
its mathematical character. In the sections that follow I will examine this philosophi-
cal argument and, in the process, demonstrate that Duhems underdetermination the-
sis is not ultimately meta-practical but does indeed rest on his experience as a practis-
ing scientist.
In Section 2 I discuss the Duhemian underdetermination argument as it is usually
presented. This is necessary preparation for Section 3 where I argue that the familiar
account is incompleteit loses sight of a crucial premise of the argument. What I
shall call the familiar underdetermination argument fails to consider Duhems con-
ception of physics. On his view, physics has two principal features: (1) it is a math-
ematical science, and (2) this is made possible by the use of tools and procedures
for measuring. (As I will explain below, measurement enables theoretical terms to
be represented by numbers.) By examining how physical theory employs numerical
symbols and how its development depends upon measurement, we will avoid the
unfortunate consequences of neglecting Duhems understanding of physics. For one,
we will be able to chart his account of how physical theory is applied and thereby
introduce a new voice into contemporary discussion of that topic. But, more
importantly, we will be able to locate the source of Duhemian underdetermination.

2. The familiar Duhemian underdetermination argument

The familiar formulation of Duhems underdetermination thesis relies on Part 2,

Chapter 6 of The aim and structure of physical theory. Although those who appropri-
ate his arguments do not usually mention this, he opens that chapter with a proviso:
he is only concerned with whether a particular type of theory is confirmed or weak-
ened by facts. For Duhem, the type of theory at issue is mathematical and makes
use of symbolic representations. It is the kind of theory that, to use his famous
phrase, is impossible to leave outside the laboratory door (Duhem, 1991, p. 182).
He calls this physical theory; it is the domain of the physicist.
When the physicist performs an experiment to test a theory or a certain theoretical
proposition, Duhem argues that the physicist does not follow the method of falsifi-
cation. Duhem defines falsification by answering the question How will the physicist
demonstrate the inaccuracy of a theory or hypothesis?. His response is: From the
proposition under indictment [the physicist] will derive the prediction of an experi-

4
References to this work, the 1991 English translation of the second edition, will appear parenthetically
in the text.
514 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

mental fact; he will bring into existence the conditions under which this fact should
be produced; if the predicted fact is not produced, the proposition which served as
the basis of the prediction will be irremediably condemned (p. 184). The objection
he then raises to this experimental method is that the conditions under which it
functions are much more complicated than is supposed in what we have just said (p.
185). In point of fact, he explains, the physicist makes use of auxiliary assumptions to
derive the predicted phenomenon and to establish whether it should be, and whether
it is, produced in the experimental situation. Thus the failure of the expected result
to obtain could be explained by an error in the theoretical proposition used to gener-
ate the prediction or in another that was needed either to design or institute the
experiment, or interpret its results. In short, it is impossible, according to Duhem,
to compare the deductive consequences of the theory, such as its predictions for the
results of an experiment, to an experimental fact in isolation from theory.
Duhems conclusions on this matter have been succinctly formulated by Philip
Quinn as the separability and falsifiability theses. The separability thesis states: no
single or individual theoretical hypothesis by itself has any observational conse-
quences (Quinn, 1974, p. 37). The falsifiability thesis, as we have seen, follows
from this. It reads: no single theoretical hypothesis can be conclusively falsified by
any observations (Quinn, 1974, p. 37). Duhem supplements his argument against
falsification with two against verification.
He argues that neither direct nor indirect verification can logically determine
theory choice. On his view, crucial experiments cannot be used to demonstrate the
truth of a theory indirectly. Crucial experiments were meant to decide between rival
theoretical hypotheses by condemning all competitors save one, and thereby con-
firming that one. Duhems objection to this procedure, according to the familiar
account, is independent of the separability and falsifiability theses I outlined above.5
He denies that a disjunction could be constructed wherein all possibilities but one
could be eliminated, and his claim is based on the following practical consideration:
the physicist is never sure he has exhausted all the imaginable assumptions (Duhem,
1991, p. 190). Therefore, on Duhems view, in physics indirect verification via cru-
cial experiment is impossible.
Furthermore, he objects to the conclusive verification of a theoretical hypothesis
by direct demonstration, which would involve establishing the truth of a proposition
by the sole use of those two intellectual operations called induction and generaliza-
tion (p. 190). Duhem does not proceed in the usual way, by noting the logical
problem with induction. Rather, he again invokes the role that theory plays in
science.6 He explains that the scientist presupposes a whole group of hypotheses in

5
See Quinn (1974), pp. 3943. I will show below that this account neglects an argument in Aim and
structure that does appeal to these two theses.
6
Duhem wrote at a time when an instrumentalist interpretation of scientific theories was already
fashionable and a mistrust of scientific explanation was pervasive. In France, theory was being devalued.
Duhem desired to assert its cognitive value without an appeal to metaphysical dogmatism. As a result,
he criticized the positivist conception of science with arguments that rely on the role of theory. For more
on this, see Maiocchi (1990).
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 515

order to translate laws that bear almost directly on observation (for example, Keplers
laws) into symbolic laws (for example, Newtons law of universal gravitation). In
other words, he concludes that the choice of a particular symbolic law must be
justified by more than empirical generalization or induction on the experimental law.
He defends the need for a choice by arguing that no experimental law is exact but
only approximate, and is therefore susceptible to an infinity of distinct symbolic
translations (p. 199). How we are to understand this will be addressed in the next
section. For now, it is sufficient to note that, on Duhems view, the choice between
alternative symbolic laws, like the choice involved in falsification, can only be made
with reference to factors external to logic. This signals the failure to determine theory
or hypothesis choice by direct verification.
Duhems arguments against verification and falsification, taken together, develop
his underdetermination thesis, and usually they are presented as above. Nevertheless,
this familiar account is unfinished and partial. It is incomplete because the assump-
tions underlying the thesis are lost in the background and unanalyzed. In the next
section I will depict Duhems three arguments schematically in order to locate their
assumptions, which I will then explain.

3. The complete Duhemian underdetermination argument

It is telling that when Duhems arguments against falsification and (direct and
indirect) verification are cited today his initial proviso is left unanalyzed. Duhem is
only concerned with a mathematical or abstract and symbolic physics. His emphasis
on approximation and experimental error, as characteristic of physics, is overlooked.
Thus, the first and, I will argue, most important element of his underdetermination
argument is neglected. What I aim to demonstrate in this section is that his concep-
tion of physics is indispensable to understanding his underdetermination thesis. I will
first find and then investigate the assumptions of Duhems tripartite argument (3.1).
Spelling out how he supports them, I argue, involves answering the question What
does Duhem understand the physicist to do when she sets out to construct a physical
theory?. By examining Duhems understanding of mathematical physics (3.2),
mathematical deduction (3.3) and the structure of physical theories (3.4), we will
complete the familiar account of Duhemian underdetermination.

3.1. Assumptions

As I have explained, Duhems underdetermination argument consists of three

parts. Duhem argues against (i) falsification, (ii) indirect verification (or crucial
experiment), and (iii) direct verification. In this subsection I locate the assumptions
of these arguments; for once they are brought to the fore, we shall be able to see
which questions must be addressed in order to fully understand Duhemian under-
determination.
The familiar argument against falsification (i) begins with the claim that the physi-
cist who carries out an experiment, or gives a report of one, implicitly recognizes
516 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

the accuracy of a whole group of theories (Duhem, 1991, p. 183). Duhem asks his
readers to accept this principle and see what consequences we may deduce from
it (p. 183). The consequences are as follows:

1. The physicist does not follow the standard account of how to demonstrate
the inaccuracy of a proposition.7
2. The physicist, having invoked a whole group of theories to perform the
experiment, is unable to falsify a single theoretical hypothesis.

Clearly, the assumption at work in this argument is the principle that says Experi-
ment in physics requires the recognition of a whole theoretical group. Or, somewhat
differently, the argument opens with the separability thesis, and then deduces the
falsifiability thesis. From either formulation, it is apparent that the argument against
falsification relies on Duhems understanding of the nature of physical theory and
its role in the practice of science. Specifically, to evaluate the argument we need to
understand why Duhem thinks physical theory is the kind of theory that is imposs-
ible to leave outside the laboratory door (p. 182).
Usually, the argument against indirect verification (ii) is simply put: the physicist
cannot enumerate all the hypotheses that could possibly account for the phenomena
in question, so experimental contradiction cannot eliminate all but one. Duhem
argues that, for example, Foucaults crucial experiment concerning the nature of light
does not constitute a strict dilemma, for it might be the case that light is neither a
particle nor a wave. It is important to note that this familiar argument is given after
Duhem asks us to admit for a moment that in each of these systems [i.e. Newtons
optics and Huygenss optics] everything is compelled to be necessary by strict logic,
except a single hypothesis, because that would let us admit that the facts, in con-
demning one of the two systems, condemn once and for all the single doubtful
assumption it contains (p. 189). This caveat is a denial of the separability and falsi-
fiability theses. If we accept them, then, according to Duhem, we are faced with
another argument against indirect verification. (See p. 198.)
This other argument downplays the decisiveness of crucial experiment by com-
menting on the nature and role of physical theory. Duhem maintains that the physicist
does not really judge between two single or individual hypothesesfor example,
about the speed of lightusing only the results of an experiment. In his view, the
most the results can tell the physicist is which set of theoriesNewtons or Huyg-
enssbetter accords with the facts. For given the separability and falsifiability the-
ses, none of the alternative hypotheses could individually be deemed inaccurate.
What follows is that the physicist cannot demonstrate the truth of an isolated prop-
osition by means of a crucial experiment. For our purposes, this is the more interest-

7
The standard account has four steps: (1) a formula is deduced from fundamental hypotheses of the
theory; (2) conditions of an experiment are plugged into the formula; (3) the calculation associated with
the formula is performed, generating theoretical predictions for the result of the experiment; and (4) the
prediction is compared to the actual result of the experiment for assessment.
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 517

ing argument against indirect verification and the one that will be pursued, even
though Duhem did not present it as his final case against crucial experiment.
Duhems case against direct verification (iii) most obviously assumes a particular
appreciation of the structure of physical theory. Schematically, a sketch of the argu-
ment looks like this:

A. To translate experimental law into symbolic law, the physicist must invoke
a whole group of theories.
B. In addition, because experimental laws are approximate, they are susceptible
to an infinite number of distinct symbolic translations.
C. Therefore, there is a choice of symbolic law and it is justified by more than
empirical generalization or induction on the experimental law(s).

This brief sketch of the argument does seem to motivate the need for extra-logical
factors in deciding on a particular symbolic law. This can be made clear by taking
up the following questions: what does Duhem mean when he says that physical
laws are symbolic? Why does translation into this form require theory? What are
experimental laws? And why are they approximate? The purpose of the next three
subsections is to set us up to answer these questions (in 3.5), for they lie at the
heart not only of this argument, but of the others as well. Duhems particular under-
standing of the nature and role of physical theory is the assumption supporting all
three arguments. Therefore, appreciating his position will enable us not only to com-
plete the standard or familiar portrayal of his underdetermination thesis but also to
evaluate it en bloc.

3.2. Mathematical physics

According to Duhem, theoretical physics is mathematical physics. In a mathemat-

ical physics all theoretical concepts are capable of being represented by numbers.
Physical attributes are signified by a numerical symbol. This characterization of
physical theory introduces the expression symbolic, which Duhem subsequently
attaches to facts, experimental laws and, of course, symbolic laws. It is possible
to signify an attribute by a numerical symbol (such as 5) when the attribute is
a magnitude.
The length of a line, for example, is a magnitude, firstly, because the relations of
equality and inequality are applicable to different line lengths. We know this by
comparing them. Because we have specified a procedure for comparison, by which
we define equality and inequality of lengths, we can recognize properties of equality
and inequality, such as: two lengths equal to the same length are equal to each other.
Thus, we can express that the lengths of A and B are equal using the arithmetical
symbol = by writing A B or that A is greater than B by writing A B or
B A.
Secondly, length is a magnitude because the joining of several lengths can be
represented by addition. Duhem asks us to place several lengths end to endwe
should add in a straight lineand observe that we obtain a new length S which
518 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

is greater than each of the component lengths A, B, and C (p. 108). We also notice
that S does not change if we change the order in which we put the components end
to end; neither does it change if we replace some of the component lengths (e.g., B
and C) by the length obtained by putting them end to end (pp. 108109). On
Duhems view, these observations permit us to employ the addition sign (+) to rep-
resent placing A, B and C end to end in a straight line to form S. We can also express
the commutative and associative laws for addition. Furthermore, because placing n
lengths equal to A end to end will form S, S can be represented by n A. Duhem
recognizes this as the starting point for the measurement of lengths (p. 109).
Using a standard of length, say the meter (which is the length given to us under
very specific conditions by a certain metal bar deposited in the International Bureau
of Weights and Measuresp. 109), we can reproduce a length by laying out n
lengths equal to a meter end to end. Or, we can divide the meter into q equal seg-
ments, perform the same procedure to get p segments, and then call our length p/q
meters. In short, any rational number x followed by our standard will represent
any length.
Now, taking together the algebraic operations we defined above, namely, the pos-
tulates of arithmetic, and the results of measurement, Duhem grants us the power
to represent operations performed with lengths, volumes, angles, times, and so on
(p. 110). All of these magnitudes are capable of being signified by a numerical
symbol and situated in algebraic equations. Nevertheless, Duhem has not yet captured
the whole of physics mathematically. In Aristotelian terms magnitudes are quantities,
and we have not yet spoken of their counterpart, qualities.8 Qualities are different
from quantities primarily in that addition, and the sort of measurement that relies on
it, does not apply to them. There is no directly applicable additive joining operation.
The Cartesian move is to assert that at bottom everything is quantity and thus to
admit only the quantitative into physical theory. Alternatively, Duhem believes (1)
that physics cannot decide whether the real properties of things (as opposed to their
observable appearances) are quantitative or qualitative, and (2) that the purely quali-
tative character of a notion is not opposed to the use of numbers to symbolize its
various states (p. 114). Therefore, Duhem sets out a procedure for numerically rep-
resenting the intensities of a single quality. We can affix a label and number, so
to speak, to each of these various intensities, registering the same number in two
circumstances where the same quality is found with the same intensity, and ident-
ifying a second case, where the quality considered is more intense than in the first
case, by a second number greater than the first (pp. 115116). For example, substi-
tute for the quality of heat a numerical symbol, its temperature. Successful substi-
tution relies on taking into account the transitive relations concerning equality and
inequality (i.e. if A is as warm as B and B is as warm as C, then A is as warm as
C and if A is warmer than B and B is warmer than C, then A is warmer than C).
But the information available by knowing, for example, that the numbers 10, 20 and

8
To put this somewhat differently, that is, in modern terminology, Duhem has accounted for extensive,
but not yet intensive, quantities.
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 519

100 represent three temperatures is very incomplete. We have full information only
when we join it to concrete knowledge of the standard which represents the unit
(p. 117). Concrete knowledge of the standard, in turn, entails knowing a concrete
procedure for obtaining the measurement scale. This scale gives physical meaning
to the algebraic propositions that indirectly involvevia a quantitative effect of the
qualitythe qualitative intensities.
So far I have discussed Duhems remarks on measurement. They compose the
first part of his vision of physics, wherein the physicist picks out primary properties
qualitative or quantitativeand correlates them with numerical symbols.9 To put this
somewhat differently, when physicists set out to construct a physical theory (p.
132), Duhem understands them first to designate some observable properties as pri-
mary. These they then represent numerically (i.e. by algebraic or geometric sym-
bolsp. 132). We have seen how this works through the choice of a standard and
measuring technique or of a scale which allows us to substitute numbers for the
various intensities of a quality. Measurement procedures are performed so that the
physicist can then reason about the properties using the abstract and symbolic langu-
age of mathematics. Later, I will discuss the important implications of this
(principally in 3.3).
For now, let us return to what physicists do, according to Duhem, when they
construct physical theory. We must establish relations, he states, among the
algebraic or geometric symbols representing the primary properties (p. 132). Duhem,
in other words, is referring to the statement of hypotheses. The underdetermination
argument is presented in this context, for it is meant to explain, in part, how physi-
cists come to choose their hypotheses. However, rather than turn directly to this
second stage, Duhem proceeds to the third, the mathematical development of the
theory. Mathematical deduction, he explains, intends to teach us that on the
strength of the fundamental hypotheses of the theory the coming together of such
and such circumstances will entail such and such consequences (p. 132). It is sig-
nificant that Duhem examines the third operation physicists perform in building a
theory before the second. His explanation is that doing this will reveal the structure
of physical theory, as well as its weaknesses. And this, he claims, it is indispensable
to know before we investigate how hypothesesthe principles used for deduction
are determined (p. 132). This claim supports the one I made at the outset, namely,
that Duhems understanding of the nature and role of physical theory drives his
underdetermination argument. For that reason, let us now turn to mathematical
deduction.

3.3. Mathematical deduction

According to Duhem, the mathematical development of physical theory requires

a dictionary. Two translations are necessary to connect observable facts and the

9
Owing to his anti-metaphysical stance, Duhem does not regard primary qualities as by nature elemen-
tary. Instead, on his view, they are so classified provisionally; primary qualities are the results of our best
efforts at reduction to date. For more on this, see Part 2, Ch. 3 of The aim and structure of physical theory.
520 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

theorys mathematical development. First, in order to be introduced into calculations,

the relevant concrete circumstances of an experiment must be translated into numeri-
cal values. Then, to verify the results the theory predicts for the experiment, a
translation exercise must transform a numerical value into a reading formulated in
experimental language (p. 133). The dictionary that makes possible these two trans-
lations is the method of measurement. Measurement translates the concrete experi-
mental circumstances into a number, and by reference to the same measurement
procedure the theory-predicted result, also a number, will be made to correspond to
a certain observable fact or instrument reading. That predicted reading could then
be checked empirically.
But, Duhem warns, translation is treacherous; for there is never complete equiv-
alence between something and its translated version (p. 133). In the context of physi-
cal theory, this introduces a distinction between two types of facts. A theoretical fact
is that set of mathematical data through which a concrete fact is replaced in the
reasoning and calculations of the theorist (p. 133). In contrast is a practical fact.
The latter lacks the precision or exactness and ideality of the former. If a theoretical
fact states the temperature of a certain body, the body is geometrically defined
(having sides without thickness, and so on) and an unambiguous temperature corre-
sponds to each point. The practical fact translated by this, however, concerns a con-
crete or non-ideal body with vague edges and dimensions. The temperature it gives
is also vague in the sense that we cannot declare, for example, that this temperature
is strictly equal to 10 (p. 134). Instead, the practical fact only asserts, explains
Duhem, that the difference between this temperature and 10 does not exceed a
certain fraction of a degree depending on the precision of our thermometric methods
(p. 134).
The distinction between theoretical and practical facts has the following notable
consequence: the same practical fact may be translated into an infinity of different
theoretical facts. There is indetermination, he argues, at the level of facts. Because
measuring instruments are only accurate to a certain degree or within a certain margin
of error, an infinity of distinct and incompatible theoretical facts will correspond to
one and the same practical fact. In other words, a practical fact is not translated by
a single theoretical fact but by a kind of bundle including an infinity of different
theoretical facts (pp. 134135). This bundles range extends only to the limit of
error of the measurement. Duhem unmistakably makes use of measuring instruments
to formulate his point about translation. This might suggest to some that Duhem
postulates perfect instruments or techniques as the solution for indetermination.10 On
this reading, if physicists used perfect instruments then a practical fact would stand
in a fixed relation to its corresponding theoretical fact. Duhem, however, seems to
suggest that there never will be such perfection when he quite clearly insists that

10
Cartwright, Cat, Fleck and Uebel (1996) come close to endorsing this view. They write: Duhems
account of scientific testing offers some hope for bridging the logical gap between concrete facts and the
deductive system of theoretical hypotheses. With more and more refined techniques we can zero in on
the correct theoretical description to associate with an everyday experience (p. 212). See also Uebel
(19981999, p. 82).
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 521

the limits of error never become so narrow that they vanish (p. 135).11 This, he
claims, has serious consequences for the mathematical development of physical
theory.
The theory develops mathematically when the conditions of an experiment are
plugged into a formula that is deduced from fundamental hypotheses of the theory.
In performing the calculation associated with the formula, the physicist generates a
theory-driven prediction for the results of the experiment. This prediction can then
be checked empirically against the actual results of the experiment in order to evalu-
ate the mathematically deduced formula. What this means, on Duhems view, is that
the testing of the formula will necessarily involve two translations as well as two
bundles of theoretical facts.
The first bundle is the result of the first translation, which involves putting into
numerical form the practical facts of the concrete experimental conditions. Next, the
calculation is performed using these experimental parameters (i.e. using the first
bundle) and a second bundle of theoretical facts ensues. This second bundle is
intended to stand for the result of the experiment but it must first be translated into
concrete language or practical facts, for only then shall we know truly the result
assigned to our experiment by our theory (p. 136). The result will be truly known
because, via the second translation, expected instrument readings will be substituted
for the theoretical predictions. Yet even so, the results of this second translation will
not be able to test the mathematically deduced formula unless they comprise one
practical fact. The translation could yield a single practical fact, according to Duhem,
if the theory-driven predictions never differed by more than the degree of sensitivity
of the measuring instrument. Alternatively, it may happen that we obtain not a single
practical fact but several practical facts which the sensitivity of our instruments will
allow us to distinguish (pp. 136137). If the latter were the case, then the prediction
would correspond to more than one possible observation and, consequently, it would
be useless for testing the formula in question.
The issue then is: can the second bundle of theoretical facts be made to correspond
to only one practical fact? In some cases, Duhem argues, sufficient tightening of the
bundle of theoretical facts will result from increasing the degree of sensitivity of (a)
the apparatus used to observe the results of the experiment and (b) the means of
measurement used in the first translation from practical to theoretical facts. In those
cases, such tightening would guarantee the desired correlation with one practical
fact. However, he continues, this is not always so. Duhem provides two examples
of mathematical deductions that defy having their results represented by a single
practical fact. The first concerns the trajectories or geodesic lines of points on certain
surfaces. The second is similar but specifically addresses the problem of the stability

11
Furthermore, even if there were perfect instruments or techniques, this would not guarantee a fixed
relation between theoretical and practical facts and an end to indetermination. As I will explain below,
Duhem observes that in some cases, however precise and minute are the instruments by which the
experimental conditions will be translated into numbers, the second translation from theoretical to practi-
cal language will still correlate an infinity of different practical results with practically determined experi-
mental conditions (p. 139).
522 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

of the solar system.12 Duhem claims that these examples show that the tightening
of the first bundle of theoretical facts (intended to depict the experimental situation)
will not prevent the second bundle (the calculated results) from correlating with an
infinity of practical facts. Hence, on Duhems view, there are some cases in which
indetermination is never completely overcome in practice and, as a result, some
mathematical deductions are of no use for the physicist. What is more, Duhem argues
that the physicist always works with a mathematics of the approximate, in which
the theoretical consequence of an approximately true proposition must be approxi-
mately exact (and the range of these two approximations must be delimited) in order
to be useful.
In his discussion of the nature of experiment, Duhem adds to this list of conse-
quences. He spells out in more detail indeterminations role in the development of
experimental laws. Experiment, he says, has two parts. In the first place, it consists
in the observation of certain facts; in order to make this observation it suffices for
you to be attentive and alert enough with your senses (p. 145). This contrasts with
the second part, which requires an interpretation of the observed facts. To interpret
the observed facts it is necessary to know the accepted theories and to know how
to apply them, in short, to be a physicist (p. 145). Examples of interpretation include
reading off the level of mercury in a thermometer as a temperature or the motions
of a spot of light on a ruler as a coils resistance. The observed or concrete facts
(for example, that the mercury is level with a certain line-mark) are set in opposition
to the abstract symbols (for example, the degree of temperature) to which they are
connected by theory. Duhem explains: An experiment in physics is the precise
observation of phenomena accompanied by an interpretation of these phenomena;
this interpretation substitutes for the concrete data really gathered by observation
abstract and symbolic representations which correspond to them by virtue of the
theories admitted by the observer (p. 147). In other words, using an assortment of
theories, the physicist correlates instrument readings with numerical symbols that
represent intensities of significant qualities.
Above (3.2) we saw the physicist, through the choice of a standard and measuring
technique or of a scale, represent primary qualities or quantities numerically in order
to have a mathematical physics. Now we see Duhem unpack this notion of represen-
tation. He makes it clear that the ability to represent or correlate a concrete fact with
a theoretical one depends on the acceptance of various theories, say, of optics or
hydrostatics, because these theories justify and make possible the use of certain
instruments. And these instruments are needed to give a value to the abstract and
symbolic expression (such as pressure, volume, temperature) at issue. Furthermore,
because he takes the stated results of an experiment in physics to relate these same
abstract and symbolic ideas, Duhem claims that they too are marked by theoretical
interpretation and hence the involvement of an assortment of theories. Today the

12
The first example describes a certain chaotic system with sensitive initial conditions, and Duhem
was right to conclude that in such a system widely different predictions will result from a patch, as small
as one could want, of initial conditions. Presumably he thought (again correctly) that the solar system
was another example of a chaotic system.
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 523

theory-ladenness of phenomena and experimental results is well established, but its

implications for the structure of physical theory are not.13

3.4. Structure of physical theories

Before we examine in detail Duhems understanding of the structure of theories,

three preliminary remarks are in order.
First, the abstract and symbolic language of physics, as Duhem describes it, is
not a technical language, conventionally established to express a specific and con-
crete object or procedure. Mass, volume and pressure, for example, are abstract con-
cepts like those of a technical language; they are generalized from and realized in
concrete cases. However, they are also symbolic and as such their meaning is estab-
lished with the aid of instruments and measurements, which in turn rely on an
assortment of theories (p. 166). Unlike commonsense laws (for example, all men are
mortal), laws that feature abstract and symbolic terms are, according to Duhem, not
the sort of abstractions that emerge spontaneously from concrete reality (p. 167).
Instead, to attach meaning to them requires knowledge of the physical theories in
which they are embedded. If one knows the relevant theories, then one can translate
statements that make use of mass and the like into facts about how to measure or
manipulate it. The difference, then, between a technical language and the language
of physics is that the latter can translate its peculiar terms into facts in a variety
(even an infinity) of ways, whereas a technical language expresses a specific oper-
ation performed on very specific objects (p. 149). For example, a ship captain speaks
a technical language to his sailors when he says All hands, tackle the halyard and
bowlines everywhere!, because the sailors can carry out his order in only one way
(p. 148). The flexibility in translation that physics allows accounts for situations in
which, for instance, an effect is said to be present even though a typical measurement
procedure indicates the opposite.
Flexibility in translation, however, is constrained by physical theory. Having pre-
supposed the theory, the physicist in effect condenses an infinity of different mean-
ings for an abstract and symbolic proposition into a few. This few, however, cannot
be condensed into one, according to Duhem: any single abstract and symbolic for-
mula cannot be the exact equivalent of a concrete situation. This is because, he
argues, Between an abstract symbol and a concrete fact there may be a correspon-
dence, but there cannot be complete parity; the abstract symbol cannot be the
adequate representation of the concrete fact, the concrete fact cannot be the exact
realization of the abstract symbol (p. 152).
To put this somewhat differently, the correspondence between practical and theor-
etical facts is not determinate. For every experiment there is a degree of what Duhem
calls symbolic indeterminationit is the degree of approximation or limit of error
of the experiment in questionand it cannot be eliminated because scientific instru-
ments cannot uniquely assign a value or number to correspond to, say, a volume or

13
On the theory-ladenness of phenomena (in contrast to data) see Bogen & Woodward (1988).
524 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

pressure of a gas. The failure of determination has been treated above (3.3), and
my second preliminary remark is simply a reminder of this twofold disparity.
The last of my preliminary remarks concerns Duhems understanding of the struc-
ture of scientific theories. As do more modern discussions on this topic, The aim
and structure of physical theory sets out to provide an analysis of theories. Note,
however: Duhem proceeds from the standpoint of the physicist creating or con-
structing a physical theory. He intends to depict the physicists labor and from his
portrayal pick out the structure of theories. This resists the so-called Received View
and is more in line with its alternatives.14 In what follows I aim to explicate Duhems
view by considering it in relation to Patrick Suppess well known hierarchy of mod-
els.
Like Duhem, Suppes is concerned to examine what it takes to assess the adequacy
of an empirical theory. That is, they both wonder what is involved in comparing
actual experimental results with results predicted for an experiment by theory. Suppes
proposes that a hierarchy of models mediates between actual experiments and their
results (i.e. data), on the one hand, and physical theory on the other.15 Although I do
think the ins and outs of his account of the creation of canonical data is informative, I
find the notion of a hierarchy unhelpful here, and so I drop that aspect of Suppess
characterization from my discussion. For me, a one-dimensional depiction is not
sufficient to evoke the complexity of the interrelationships between terms and the
ultimate convergence of two of them for final assessment. I think it is more useful
to see, as in Fig. 1, a data path and a theory path leading to a single point (depicted
as a circle) where canonical data and theoretical predictions can finally meet and
be compared.
Additionally, Suppes makes use of the language of set theory to define the notion

Fig. 1. Adaptation of Suppess Hierarchy.

14
For a thorough treatment of the Received View and some alternatives, see Suppe (1977). More recent
alternatives include Giere (1999) and Morgan & Morrison (1999).
15
See Suppes (1962, 1967). For a careful discussion of the former see Suppe (1977).
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 525

of a model.16 It is beyond the scope of this paper to defend or question this prefer-
ence, for the debate pertaining to this issue has a long history.17 At any rate, I think
Suppess basic thesis can be explained without the meaning of model and the pre-
cise nature of the theory/model relation being spelled out. In what follows I take it
upon myself to explain his insight using plain language.18
Let us look again at Fig. 1. It highlights two paths, one emanating from theory
and the other leading up to and away from data. They represent Suppess claim that
the comparison of theoretical predictions and experimental results involves models
of different logical type. The following two considerations are intended to support
this: often terms that feature in the theory have no direct observable analogue; and
often models of the theory contain continuous functions or infinite sequences despite
the fact that the experimental data do not (Suppes, 1962, p. 253). The issue, therefore,
is what can be said about each type of model. Suppes chose to characterize these
types by explaining their purpose and role in preparing for confirmation/
disconfirmation.
On the theory path, we begin with a physical theory and then extract from it those
principles or conditions relevant to the class of experiments we are considering.
Suppes calls the limited theory the theory of the experiment, perhaps because the
winnowing process is guided by experimental considerations and its purpose is to
limit the possible data-sets obtainable to those for our given experiment. As a model
of the theory, these data-sets are the theoretical predictions that will be compared
in the end to the experimental results.19
Turning to the data path we encounter the actual experimental setup. Explicit
principles of experimental design and unstated ceteris paribus conditions are taken
into account in the building and running of the experiment. The actual data produced,
however, is not ready to be compared with the theorys predictions. First it must be
corrected or transformed into canonical data, so that it becomes, in Suppess termin-
ology, a model of the data.20 For example, if the actual data comes from frictional
surfaces and the theory-derived possible data is about frictionless surfaces, correction
procedures must be employed to account for the difference. Which correction pro-
cedures are applicable for a given experiment is determined by what Fred Suppe,

16
Specifically, he writes, a model of a theory may be defined as a possible realization in which all
valid sentences of the theory are satisfied, and a possible realization of the theory is an entity of the
appropriate set-theoretical structure (Suppes, 1962, p. 252).
17
See Suppe (1977) and, for an assessment of developments in the work on the structure of scientific
theories approximately thirty years after the conference that gave rise to it, see Suppe (2000), da Costa &
French (2000) and Huggett (2000).
18
My discussion of Suppes is drawn from Suppes (1962, 1967). I will also bring in Suppe (1977) at
points, which I suspect will introduce anothers (closely related) insight as well.
19
As this description should suggest, Fig. 1 includes both theories and models, and roughly indicates
the relations between them. Theories are boxed without fill and the boxes of models are filled with gray.
Data is without a box to indicate that it rests in a separate category; it is neither theory nor model but
the product of the actual experimental setup.
20
He uses this language in Suppes (1962). In Suppes (1967) he seems to denote this same thing by
theory of experiment.
526 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

loosely following Suppes, calls the theory of the data. Basically, the theory of the
data is what specifies how the actual data should be modified so as to be in terms
of the possible data given by the theory.
Once the two are in similar terms, a comparison can be made between them.
According to Suppes, the comparison involves statistical tests for goodness of fit,
as well as other considerations (such as homogeneity, stationarity and order). In the
end, what we learn from all this is that Suppess insight is in having seen that the
relata being compared are not the extremes: actual experimental results and physical
theory. Rather, we ask if a model of the data is sufficiently similar to a model of
the theory.
So, how does Duhem envision the structure of scientific theories? Above we saw
Suppes complicate theory structure by introducing a hierarchy of models and theories
to account for the analysis of empirical theories. Duhem complicates theory structure
for the same reason. Specifically, he complicates it in order to correct the standard
view of how physical theory develops by mathematical deduction. He supplements
the following four basic steps:

1. A formula is deduced from the fundamental hypotheses of the physical

theory.
2. The conditions of an experiment are plugged into the formula.
3. A calculation associated with the formula is performed, generating a theor-
etical prediction for the result of the experiment.
4. The actual results of the experiment are compared to the prediction.

As we have seen (3.3), Duhem adds to this two translations. The first puts the
practical facts of the concrete experimental conditions into a useful form and the
second translates the theoretical prediction into expected instrument readings or
observations. So we now have:

1a. The conditions of an experiment are translated into numerical form.

3a. The prediction for the results of the experiment is translated into the concrete
language of practical facts.

The first translation (1a) is also in Suppess picture, but the second (3a) isnt. As
far as I can tell, this second translation is a move unique to Duhem. But even this
picture, Duhem notes, is still too simplistic: what goes on in performing the experi-
ment needs investigation. In physics, experiment involves instruments, and Duhem
argues that using these instruments necessarily appeals to physical theory. For theory,
he says, helps the experimenter to distinguish noise from signal. It also allows her
to formulate, for instance, an instrument panels display as a reading of pressure,
current, temperature, and so on.
Take, for example, an ion gauge, which is used for measuring very low pressures.
See Fig. 2, below. A current is run through a filament and that causes electrons to
be ejected from the filament by thermionic emission (of current i). A wire grid
surrounds, but is not in electrical contact with, a straight wire called the collector.
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 527

Fig. 2. Ion gauge schematic.

A bias or voltage is applied between the grid and ground, which causes the electrons
emitted from the filament to accelerate toward the collector and ionize the local gas.
This ionized gas carries a small current (i+) from the collector to ground. The value
of that current depends, in the ideal case, on only two factors: the type of gas and
the amount of gas (the pressure). Thus, to determine the pressure of the gas, the
experimenter measures i+by a process that for our purposes will remain unana-
lyzedand uses a simple formula to convert it into a pressure reading. The for-
mula is:
i+ ipC
where p is the pressure of the gas and C is the sensitivity. C is a constant that is
chosen based on the type of gas involved and how both currents are measured.
Derived from thermodynamics and electromagnetism, this formula is used to cal-
culate pressure given values for i+, i and C. What Duhem highlights is that values
like these depend on characteristics of the particular experimental setup. These let-
ters [in our case, i+, i and C] are replaced by the numerical values suitable for the
instrument used and for the laboratory in which it is (p. 155). For example, in the
case above, if C were normalized for nitrogen, measuring the pressure of hydrogen
would require multiplying C by a factor of 2.2. What this suggests to Duhem is that
the scientist is occupied with two very distinct representations of the instrument on
which he is working (p. 155). One is the concrete instrument situated in a certain
lab and the other is a schematic model of the same (p. 156). The latter is con-
structed with the aid of symbols supplied by theories, which means that the terms
needed to depict it are theoretical (p. 156). It exists in the physicists reason and is
used for reasoning.
To say that the physicist reasons on the schematic instrument is just to say that
her calculations, which are used to translate the display reading into (say) a pressure,
contain its idealized conditions. Because of this, the physicist is often concerned to
complicate the schematic instrument in order to better represent reality. According
528 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

to Duhem, This passage from a certain schematic instrument to another which better
symbolizes the concrete instrument is essentially the operation that the word correc-
tion designates in physics (p. 157). The corrections themselves are, he explains,
these transformations that we make on the immediate data of the experiment (p.
156). At first glance, this last quotation sounds a lot like Suppes, who transforms
raw data into canonical data by means of correction procedures. But Duhem is saying
something different. In short, he locates correction as occurring after the theoretical
prediction is compared to the observed experimental data.
According to Duhem, after an experiment is performed the physicist observes
certain practical facts, such as the height difference of the mercury between the two
sides of the manometer is 7 cm. This observed practical fact must be translated by
means of a formula, or more generally a model, to become a statement about press-
ure. In this case the standard formula is:
p rgh
where p is the pressure of the fluid being measured, r is the density of mercury, g
is a gravitational constant, and h is the difference in height of the mercury columns.
Duhems point is that other terms could be added to this equation to make it more
accurate. He phrases this generally when he speaks of corrections being made to the
schematic model so that it will represent the actual or concrete situation (reality)
as closely as possible. If the model were not corrected, the stated result of the experi-
ment (for example, the atmospheric pressure is such and such) would be insufficiently
precise. So corrections, on Duhems view, are complications. The new schematic
model of the manometer, for example, could add another term to account for tem-
perature variation (p. 157).
Suppes, on the other hand, seems to view correction procedures as a way to get
messy real-world data more in line with the idealized conditions that were assumed
in the theoretical calculations.21 He understands scientists to simplify raw data
for example, by subtracting the influence friction would have on an inclined plane
experimentin order to get canonical data; for the latter is more likely to bear to
a theoretical model a satisfactory goodness of fit relation. Suppess concern with
how a predicted theoretical model could be compared with actual data led him to
conclude that the data must be modified or corrected prior to the comparison. For
the comparison to be fruitful the data should be, he argued, a model of the data,
that is, in a form that might match a model of the theory.
Alternatively, Duhems concern with the same initial comparisonbetween a
theoretical prediction and actual dataled him to argue (for 3a) that the theoretical
prediction should be translated into an observable practical fact, for instance the
needle on the instrument panel points to the number 12. If the many unique theoreti-
cal facts generated by a calculation could be translated into just one practical fact
like that, then the experimenter could look at her instrument panel and judge the
hypothesis in question. However, Duhem was aware that, as it stands here, his sol-

21
This, at least, is how Suppe reads him (Suppe, 1977, pp. 107108).
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 529

ution is incomplete. That is why he analyzed the result of an experiment in physics.

Knowing that the needle does point to 12 as predicted does not advance the develop-
ment of physical theory. Rather, the result must be stated again using theoretical
terms or abstract and symbolic expressions, that is, using the words pressure, tem-
perature, density, mass, and so on. Thus, unlike Suppess picture, correction comes
into play after the comparison, in this translation between practical and theoretical
facts.22 In other words, Duhem introduces a third translation to put practically stated
results into theoretical language. On his view, this third translation adds a final step
to the mathematical development of physical theory, namely:

5. The practically stated results of the comparison are translated into useful
theoretical language.

Theoretical language is useful because it alone allows for the growth of physical
theory. Theory grows when abstract and symbolic expressions are shown to be con-
nected in new, general ways.

3.5. Answers

We are now in a position to answer the questions with which I began this section
namely, what does Duhem mean when he says that physical laws are symbolic?
Why does translation into this form require theory? What are experimental laws?
And why are they approximate?
To begin, an experimental law is a series of theoretical facts connected in various
ways. Duhem gives as an example of one the following proposition: the electro-
motive force of a certain gas battery increases by so many volts when the pressure
is increased by so many atmospheres (p. 148). As we saw at the outset, he also
identifies Keplers laws as experimental laws. In contrast, physical laws, on his view,
are more general; they unite a number of experimental laws under one formula. This
nesting relationship explains why experimental results must be translated into
abstract and symbolic expressions for physical theory to grow.
It also explains why physical laws are symbolic. They are comprised of symbolic
expressions such as force and mass, united in very general formulas. Furthermore,
now we see the origin of the indetermination of experimental results, and hence the
experimental laws which unite them, and ultimately the symbolic laws which unite
them. Indeterminacy emerges in translation because there cannot be an exact equival-
ence between a practical fact and a theoretical one, for the former has vague and
uncertain boundaries while the latters are precise and rigorous. Instruments cannot
capture the concrete situation with a unique numerical value, only a bundle of values,
so there is always a degree of experimental error involved. Recognizing this explains

22
I suspect that Duhems analysis of experiment could supplement Suppess as a description of scientific
practice; however, it is beyond the scope of this article to investigate the possibility. It is also too tangential
to my present purpose to examine nineteenth-century science to see if it, perhaps, fits Duhems model.
530 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

why experimental (and symbolic) laws are approximate. Simply put: they are
approximate because they are made up of symbolic expressions which (due to
translation) always exhibit a degree of symbolic indetermination equal to the
degree of approximation of the experiment in question (p. 169). In any case, what-
ever correspondence there is between practical and theoretical facts and whatever
meaning is assigned to symbolic expressions is established with instruments and
measurement techniques and these invoke a whole group of theories.
Articulating these relationships completes the portrayal of Duhems underdeterm-
ination argument with which we began, for they support the inferences that lead to
his conclusion. Specifically, Duhems argument against falsification depends on the
physicists inability to leave theory outside of the lab, for the physicists use of
theory makes it impossible to know whether to reject the hypothesis in question or
the auxiliary assumptions. Likewise, Duhem argues that indirect verification (i.e. via
crucial experiment) fails because whole groups of theories are needed to perform
experiments; hence, a single hypothesis cannot be left standing after experiment,
only a whole group of theories. In short, both of these attempts to determine theory
or hypothesis choice logically are unsuccessful, on his view, because experiment
requires auxiliary assumptions or an act of faith in a whole group of theories (p.
183).
Why theory is necessary for experiment is made somewhat more explicit in
Duhems argument against direct verification. In physics, he argues, the choice of a
particular law is justified by more than empirical generalization or induction on a
less general law, for a whole group of theories is required to translate from the less
general to the more general law and ultimately a choice must be made among sym-
bolic laws without . . . being guided by experiment at all (p. 199). So, why theories
are needed for translation between experimental laws (also called observational or
phenomenal laws) and more general symbolic laws, and between practical and theor-
etical facts, is key. Duhems underdetermination argument rests on the idea that the
realization and interpretation of no matter what experiment in physics imply adher-
ence to a whole set of theoretical propositions (p. 200). Now that we have studied
his understanding of the structure of scientific theories, we are in a position to exam-
ine this idea in detail. See Fig. 3, below.
This illustration diagrams how experiment in physics proceeds, according to
Duhem.23 From it and from what I have said already, we can see that a whole
set of theoretical propositions enters into experiment whenever an instrument is
manipulated, in order to make its use possible. And instruments are used to design
and execute the test of the hypothesis in question; to translate (i) the concrete experi-
mental situation into theoretically useful experimental parameters, (ii) the theoretical
predictions into expected observations (i.e. practically stated predictions), and (iii)
the practically stated results into useful symbolic constructions which give us the
result of the experiment; and, lastly, they are used to give meaning to the theoretical

23
The shaded items are described or given in practical terms, those without fill in theoretical terms,
and again the circle indicates the ultimate point of comparison.
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 531

Fig. 3. Flowchart of Duhems Hierarchy.

terms of the hypotheses, the theoretical prediction(s), and the result of the experi-
ment. What Duhem noticed is that instruments are indispensable to every aspect of
experiment and, moreover, that their use requires an untold number of theoretical
propositions. Interestingly, this straightforward observation is the foundation for his
contentious underdetermination thesis.

4. Concluding remarks

Putting Duhemian underdetermination aside for the moment, the basic or standard
underdetermination thesis is just a bit of logic and hence meta-practical. The main
reason for accepting it is a belief in the existence of algorithms for generating empiri-
cally equivalent rivals to any theory.24 What Duhem has shown is that underdeterm-
ination cannot be avoided in practice. The language of science is mathematical,
abstract and symbolic, and necessarily is not determinate in relation to the practical
facts it represents. Physicists must perform translations between practical and theor-
etical facts through the use of instruments and measurement techniques and necessar-
ily this invokes whole groups of theories. Consequently, the argument goes, neither
verification nor falsification can logically determine theory choice. In my treatment

24
A candidate for one such algorithm is to construct a rival theory T1 by accepting the empirical
consequences of any theory T but denying the existence of the theoretical entities postulated by T.
532 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533

of the complete Duhemian underdetermination argument, I hope to have shown

that Duhems thesis relies on his philosophy of scientific language and the practice
of science. If I am right, then for Duhem the gap between theory and world is the
result of the difference between the mathematical language of science and the ordi-
nary language of observation; it is narrowed yet never bridged by the instruments
and measurement techniques used by scientists. I think we can now see why contem-
porary appeals to Duhemian underdetermination should not interpret it as a purely
logical thesis with obvious implications. Even if Duhem is right and scientists do
encounter underdetermination every day in their routine activities, this does not
straightforwardly entail any grand conclusions about realism or antirealism. In short:
those who appeal to Duhems work should not neglect the background, rich in the
practice of science, that Duhem, himself a scientist, brought to bear in his philosophy.
For, as I have argued, Duhems often cited philosophical thesis of underdetermination
rests on his own observation that instruments are ubiquitous in mathematical
sciences.

Acknowledgements

I owe special thanks to Arthur Fine for his valuable advice and encouragement.
Thanks also to Mathias Frisch, Axel Mueller, Eric Winsberg and an anonymous
reviewer for helpful comments on earlier drafts of this article.

References

Bogen, J., & Woodward, J. (1988). Saving the phenomena. The Philosophical Review, 97(3), 303352.
Cartwright, N., Cat, J., Fleck, L., & Uebel, T. E. (1996). Otto Neurath: Philosophy between science and
politics. Cambridge: Cambridge University Press.
da Costa, N., & French, S. (2000). Models, theories and structures: Thirty years on. Philosophy of Science,
67(Proceedings), S116S127.
Duhem, P. (1991). The aim and structure of physical theory (2nd ed) (P. Wiener, Trans.). Princeton, NJ:
Princeton University Press. First published 1914 as La theorie physique: Son objet, sa structure.
Giere, R. N. (1999). Science without laws. Chicago: University of Chicago Press.
Huggett, N. (2000). Local philosophies of science. Philosophy of Science, 67(Proceedings), S128S137.
Jaki, S. L. (1984). Uneasy genius: The life and work of Pierre Duhem. The Hague: Martinus Nijhoff.
Maiocchi, R. (1990). Pierre Duhems The aim and structure of physical theory: A book against conven-
tionalism. Synthese, 83, 385400.
Morgan, M. S., & Morrison, M. (1999). Models as mediators: Perspectives on natural and social science.
Cambridge: Cambridge University Press.
Needham, P. (1998). Duhems physicalism. Studies in History and Philosophy of Science, 29A, 3362.
Needham, P. (2000). Duhem and Quine. Dialectica, 54(2), 109132.
Quinn, P. L. (1974). What Duhem really meant. In R. S. Cohen, & M. W. Wartofsky (Eds.), Methodo-
logical and historical essays in the natural and social sciences. Boston: Reidel.
Suppe, F. (2000). Understanding scientific theories: An assessment of developments, 19691998. Philo-
sophy of Science, 67(Proceedings), S102S115.
Suppe, F. (Ed.) (1977). The structure of scientific theories (2nd ed.). Urbana, IL: University of Illinois
Press.
 K.M. Darling / Stud. Hist. Phil. Sci. 33 (2002) 511533 533

Suppes, P. (1962). Models of data. In E. Nagel, P. Suppes, & A. Tarski (Eds.), Logic, methodology and
philosophy of science: Proceedings of the International Congress (pp. 252261). Stanford, CA: Stan-
ford University Press.
Suppes, P. (1967). What is a scientific theory? In S. Morgenbesser (Ed.), Philosophy of science today.
New York: Basic Books.
Uebel, T. E. (19981999). Fact, hypothesis and convention in Poincare and Duhem. Philosophia Scientiae,
3(2), 7595.