Approximate Truth and Scientific Realism

Author(s): Thomas Weston

Source: Philosophy of Science, Vol. 59, No. 1 (Mar., 1992), pp. 53-74
Published by: University of Chicago Press on behalf of the Philosophy of Science Association
Stable URL: http://www.jstor.org/stable/188119
Accessed: 09-11-2015 05:10 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

University of Chicago Press and Philosophy of Science Association are collaborating with JSTOR to digitize, preserve and
extend access to Philosophy of Science.

http://www.jstor.org

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM*

THOMAS WESTONtt
Department of Philosophy
San Diego State University

This paper describes a theory of accuracy or approximate truth and applies it

to problems in the realist interpretation of scientific theories. It argues not only
that realism requires approximate truth, but that an adequate theory of approx-
imation also presupposes some elements of a realist interpretation of theories.
The paper distinguishes approximate truth from vagueness, probability and
verisimilitude, and applies it to problems of confirmation and deduction from
inaccurate premises. Basic results are cited, but details appear elsewhere. Ob-
jections are surveyed, including arguments by Miller, Laymon, and Laudan.
Comparison is made with Niiniluoto's theory of verisimilitude, and the utility
of his theory for realism assessed.

1. Introduction. It has been widely recognized that realism in philos-

ophy of science cannot be defended without a notion of approximate or
near truth. Boyd (1973, 1984, 1985a), Putnam (1975b), and Smart (1968),
each explicitly include such a notion in their various formulations of sci-
entific realism. Other advocates of realism use concepts like "relative
truth"(Leplin 1984, 215) or "partialknowledge" (Trigg 1980, 196) which
are most readily understood in terms of approximate truth. On the other
side, antirealists Kuhn (1970) and Laudan (1977, 1981) argue that we
now have no clear notion of approximate truth which is suitable for de-
fense of realism, and we are not likely to have one.
This paper develops the theory of approximate truth in two directions:
It describes a technical proposal, the details of which have appeared else-
where (Weston 1987), which I believe captures the notion of approximate
truth, at least to a first approximation. It also argues that an assertion of
the approximate truth of a particular theory can only be formulated cor-
rectly with the aid of information about the specific subject matter of that
theory. Consequently, a "theory neutral" or nonempirical explanation of
approach to the truth, which has been the goal assumed by many realists
and their critics, is unattainable. I argue, however, that realism can do
splendidly with nonneutral approximate truth.
*Received January 1988; revised March 1990.
tThanks to R. Boyd, A. Garfinkel, H. Hertz, R. Thomason, S. Weissman and anon-
ymous referees for advice and criticism.
:Send reprint requests to the author, Department of Philosophy, College of Arts and
Letters, San Diego State University, San Diego, CA 92182-0303, USA.

Philosophyof Science, 59 (1992) pp. 53-74.

Copyright? 1992 by the Philosophyof Science Association.

53

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
54 THOMAS WESTON

In clarifying the concept of approximate truth, let us distinguish it from

related notions which occur in some of the same contexts. Approximate
truth is not probability, degree of confirmation, or degree of rational be-
lief. Available evidence may make it certain, for example, that the value
of the physical quantity x lies in an interval (c - e,c + e), so that the
equation "x = c" is certain to be "approximately true within e" since x
and c differ by no more than E, although we may be uncertain as to where
exactly x lies within the interval. The same example helps distinguish
approximate truth from truth of a vague statement, since we need not
suppose that equality has borderline cases or other features of vague pred-
icates in order to regard "x = c" as approximately true in this case. It is
also possible for a statement to be vague but true without approximation,
as in "John is rather tall", said of a rather tall person.
Approximate truth is not completeness or comprehensiveness. A state-
ment will count as approximately true to the degree that it is accurate in
whatever it asserts; omission of information that would be included in a
fuller account will not make a statement any less approximately true, only
less informative than it might be. A trivial true statement about some
complex subject can be perfect in approximate truth, although low in
comprehensiveness. Our sense of "approximatetruth"is thus close to that
of "accuracy".
The distinction among the various ways in which a statement or theory
could be deficient in (correct) information has particular importance for
the problem of formulating scientific realism. A characteristic set of prob-
lems arises if we try, for example, to explain realism in terms of truth.
Let us suppose that we have available a realist semantics, that is, a
correspondence explanation of truth. We might then try to describe a real-
ist attitude toward a currently accepted theory T of some subject matter
as asserting one or more of the following:

(1) that the theory T is true,

(2) that the theory T is probably true, or
(3) that available evidence indicates that the theory T is true.
None of these alternatives is likely to be defensible. The history of
science provides abundant evidence for what Newton-Smith calls the
"dismal induction": Theories which have gained and deserved acceptance
have almost always turned out to be false. Past experience would indicate
that even if a theory is sufficiently supported to warrant acceptance, the
probability that it is precisely true is roughly zero.
Each of the formulations (1), (2), and (3) fails to be a defensible for-
mulation of realism because accepting it has the effect of endorsing the
currently accepted theory as true or probably true even when its literal

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 55

truth is very unlikely. It is easy enough, however, to avoid this objection

by reformulating, say, (3) as follows:
(4) Available evidence indicates that the theory is approximatelytrue.
Notice that reformulating (3) in terms of comprehensiveness would not
help provide a more plausible version of it. The claim that
(4') the available evidence indicates that the theory is the (nearly)
comprehensive truth about its subject matter
is much harder to justify than (4) since it is easier to show that a theory
is nearly accurate in what it actually asserts than to show that it contains
nearly everything to be said about the subject. Comprehensiveness is thus
not helpful in confronting the "dismal induction". It is, of course, a cru-
cial concept for describing other aspects of the growth of science since
we should not count a field as progressing very far if it merely accu-
mulates accurate but trivial statements.
Probability or probability surrogates like "available evidence" also do
not avoid the dismal induction. It is seldom reasonable to believe of a
complex theory that its truth is made probable by the available evidence.
The most that could reasonably be said is that its approximate truth is
probable. Both the concepts of probability or confirmation and the con-
cept of approximation are needed to produce a defensible assessment of
real theories on actual evidence.
Although statement (4) is only a fragment of a formulation of realism,
it will already be too strong for some who call their views realist. Popper
(1962), for example, describes his position as realist, but he is not pre-
pared to agree that we can have evidence for the approximate truth of
theories. To do so is to commit the sin of "inductivism" just as surely
as to maintain that we can have evidence for their truth. Unlike Popper's
view, the brand of realism I wish to amplify and support is unregenerately
inductivist. As formulated by Boyd (1973), it maintains that experimental
evidence for a theory is evidence that those causal relations it describes
operate to produce the regularities in observable phenomena which the
theory predicts and, further, that "the particularcausal relations in ques-
tion explain the predicted regularities in the behavior of observable phe-
nomena" (p. 1). Evidently this view requires a specifically realist account
of cause since Humean accounts that take causation to be conformity with
observable regularities would make it incoherent. Boyd (1985b) describes
such a realist, non-Humean view. When necessary to avoid confusion, I
will call this position "strong realism".
It would be a happy accident if two or more of the disparate concepts
of accuracy, comprehensiveness, vagueness or probability could be mod-
eled or clarified with the same technical apparatusof logic. My conten-

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
56 THOMAS WESTON

tion is that this accident does not happen, and that technical devices de-
signed to accountfor vagueness, for example, would give the wrong answers
if applied to approximate truth.
One simple condition, which I will call the Accuracy Principle (AP)
distinguishes approximate truth from the other notions:

(AP) A true statement is accurate or approximately true according

to any method of measuring error and within any allowable
limits of error.

This principle separates approximate truth from each of the other no-
tions mentioned since a true statement is required to be accurate, but it
might be improbable (on available evidence), or vague, or far from com-
prehensive. In particular, theories of verisimilitude, which usually com-
bine treatments of accuracy with comprehensiveness, will not satisy AP.
Niiniluoto's theory of verisimilitude, for example, gives a treatment of
equations similar to that given below, but regards the statement that x is
in the interval (a,b), where a < b, as a positive distance from the truth,
even if x is in fact the midpoint of the interval (Niiniluoto 1982, 221).
Given Niiniluoto's intentions, this makes sense since locating x some-
where in the interval does not tell the comprehensive truth about x.
Since we are construing approximate truth as accuracy, it is easy to
see that AP should be true. If we speak, even figuratively, of "distance"
or "nearness"to the truth, then the truthmust be zero distance from itself.
We take "the truth" here not to be some grand comprehensive and final
account, but simply in the modest sense of making no errors of com-
mission. With truth understood in this usual way, a true statement is zero
distance from the truth, so it is true within any assigned limits since zero
is less than or equal to every distance.
Condition AP was defended (in effect) by Hilpinen (1976) and used in
Weston (1977, 1987). Weston (1987) has shown that under modest as-
sumptions, no truth-functionalmultivaluedlogic satisfies AP. I will sketch
below a non-truth-functionaltheory of degrees of approximate truth that
satisfies AP and has other pleasing properties.

2. Inaccuracy and Deduction. In addition to problems about the prob-

ability of theories, the dismal induction has perplexing consequences con-
cerning deductive inference from accepted theories. To illustrate, con-
sider a deduction of the ellipticity of the orbit of Mars of the sort given
in physics texts. To derive our conclusion, we need Newton's laws, in-
cluding universal gravitation, boundary conditions such as the existence
of a sun of specified mass, and suitable values of mass, position and
velocity of Mars at some specified time, and so forth. We also need a

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 57

critical condition which may not always be stated, that gravitational force
due to any other bodies in the universe is zero.
We need to know that other forces are zero because, if they are not,
the deduction is hopeless. The presence of any other body which is grav-
itationally interacting with Mars and the sun yields an instance of the
famous "three-body problem", whose exact solution is possible only in
cases which never occur in the solar system (lyanaga and Kawada 1977,
1262). The absence of such an exact solution implies, however, that a
valid argument for the ellipticity of Mars's orbit cannot be constructed
unless it has at least one false premise: either an assertion that the grav-
itational effects of all but two bodies are zero, or a (strictly false) ap-
proximation for a three-body case.
Perversely, it is appropriatethat ellipticity cannot be deduced from true
premises, for the good reason that Mars's orbit is not elliptical. The orbit
is close to ellipticity in a certain sense, but strictly speaking, "The orbit
of Mars is an ellipse" is a false statement. The textbook example turns
out to be a deduction of a falsehood by means of falsehoods. That this
result is not exceptional but typical may be seen by casting a perfection-
ist's eye over the other examples in natural science texts, or indeed, over
specialists' research reports in the journals.
The appearance of absurdity in deducing falsehoods from falsehoods
is easily dispelled by means of approximate truth: While it is false that
the other planets have no gravitational effect on Mars, and false as well
that its orbit is elliptical, both of these statements are, in some sense,
approximately true. If we can show that the inference patterns used here
will yield approximately true conclusions from approximately true prem-
ises, we will avoid pointlessness in applying deductive logic to false the-
ories. Subject to later refinement, we use the term approximately valid
for inferences which preserve approximate truth.
The accuracy principle implies that some valid arguments will not be
approximately valid. Consider the modus ponens argument:
P D 0, 0, therefore 0.

Suppose that P is false although near the truth, so that by some method
of measuring inaccuracy, 0 is a positive distance a from the truth. If
"D" is understood in its classical sense (as are all logical symbols in this
paper), P D 0 must then be true by falsity of antecedent. Since ' D 0
is true, the accuracy principle implies that its distance from the truth is
0, which is necessarily less than a. Thus we have shown that both prem-
ises of a modus ponens can be made true within a without placing any
constraints at all on the conclusion 0, which may be as far from the truth
as one pleases. Thus no constraints on the degree of accuracy of the prem-

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
58 THOMAS WESTON

ises can guarantee any accuracy in the conclusion, unless the premises
are actually true.
It is necessary, therefore, to investigate which valid inferences are ap-
proximately valid. So far, there are no general treatments of approximate
validity, although the beginnings are in Weston (1987). There are sub-
disciplines of various sciences, however, which study concepts of ap-
proximate validity appropriateto their own areas. These concepts include
well-posedness in differential equations (Mikhlin 1970), stability (LaSalle
and Lefchetz 1961), structural stability (Thom 1975), similitude (Kline
1965), robust statistical inference (Huber 1972), and others (Davis 1975).
Two features are common to most of these treatments. First, there is
a distinctionin all of these cases between problemsthat have "well-behaved"
(or robust or well-posed) solutions, and those that do not. The distinctive
property of the well-behaved problem is the continuity condition that if
sufficiently accurate "data" for the problem are given, a sufficiently ac-
curate solution will result. In the Mars-orbit case, for example, the po-
sition, mass, and velocity of Mars and the sun, and so forth are the data
of the problem. The continuity condition would be satisfied if nearly ac-
curate values of this data yield a nearly correct orbit.
The second notable feature is that there are a variety of ways of mea-
suring accuracy, and whether a problem has well-behaved solutions de-
pends on the method chosen. We will see the importance of these two
points below. We turn first to several preliminaries needed to define ap-
proximate truth.

3. Approximate Truth: The Basic Idea. The general pattern of the ac-
count of approximate truth described here is that a statement is approx-
imately true under a given interpretationif it is actually true under some
"nearby"interpretation.This much is in agreement with Hilpinin's (1976)
earlier account, but my proposal differs from his by being considerably
more specific in several respects.
Distance between "nearby" interpretationsis to be reckoned by count-
ing distances between the values assumed by the causally significant mag-
nitudes mentioned in the statements to be assessed. Moreover, the as-
sessment of these magnitudes is to be made via a metric whose choice
is to be justified on empirical grounds, and will typically be derived from
existing scientific practice in the relevant discipline.
In the case of the ideal gas law, pv = kt, for example, we will count
this as approximately true within 8 if each of the "errorterms" dp, Av,
At, and Ak are at most 8 according to a suitable metric, and if
(p + Ap)(v + Av) = (k + Ak)(t + At)
is true. This supposes, of course, that the quantities p, v, k, and t are

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 59

causally significant magnitudes of the thermodynamics of gases, an em-

pirical, but noncontroversial fact.
A detailed justification of the definition of approximate truth is found
in Weston (1987); here we note only the features we need. We assume
that the constants and function symbols of our language are divided into
those with denotations assumed to be perfectly accurate, such as math-
ematical functions and constants, and those with designations that may
be approximate. These later terms designate quantities, that is, properties
with amounts or degrees which are measured by real numbers with di-
mensions and units of measurement. Some examples of quantities are
mass, temperature, and electrical charge. Fundamental terms of most
quantitatively formulated scientific theories are naturally represented as
quantities and translated to the predicate calculus as function symbols or
individual constants. Laws will usually take the form of mathematical
equations among the quantities.
The semantics of approximate truth sketched here allows approxima-
tion to occur only in the relations between quantity terms and the quan-
tities they denote, not in the denotations of nonquantity terms or exten-
sions of predicates.1Consequently, statementslike "thatbody is a planet",
which contain no quantity terms, will be approximately true just in case
they are true, while "that planet is 108 miles from the sun" will be ap-
proximately true if the distance of the planet from the sun is nearly 108
miles from the sun, with "nearness" measured in an appropriateway.
It is important that there is no one appropriatemeasure of nearness or
error, and allowance for alternatemeasures of error is essential-so much
so that different measures must be used even within a single example. In
the Mars-orbit case, the statement that nearly accurate data yield a nearly
accurate orbit already refers to at least two methods of measuring, one
for the data, and one for the orbit. In the case of the data, the most natural
measures of accuracy are those obtained by simple arithmetic. Thus if
the true velocity of Mars at some instant is, say, 2.5 x 105 km/hr, the
accuracy of the false statement that that velocity is 2.6 x 105 m/sec may
be reported by stating that it is off by 105 km/hr or by 4 percent, or in
other ways.
Assessing the accuracy of a "nearly" correct orbit is far less straight-
forward since we cannot compare two curves in space by simply sub-
tracting them. There are in fact a variety of incompatible measures of the
'The present theory thus omits several ways in which a statement might be approximately
true. It does not construe "all" as "almost all" in analogy with the measure-theoretic con-
cept of "almost everywhere". This concept of approximate truth is not necessary for the
cases described here, but may be needed in others. See Adams (1974). A second possibility
is that predicates with approximatelycorrect extensions should be considered. Adams (1982)
reports results for systems of this sort.

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
60 THOMAS WESTON

Figure 1.

"nearness" of two curves and the choice of the right one for a specific
purpose is a complex matter.
Figure 1 (after Thom 1975) illustrates one of many difficulties which
can arise. Let us try to determine which of the curves B and C is more
similar to curve A. The average distance between A and B is smaller than
that between A and C, but the shape of C is much closer to that of A. If
we must choose between curves B and C in order to make a numerical
prediction of A, it would be wise to select B as more similar to A. For
other purposes, however, C's greater qualitative similarity to A might
lead us to select it instead. We should not expect, therefore, to choose a
priori between B and C, but leave this choice to be determined by the
subject matter concerned.
Mathematically speaking, we will only require that the method of mea-
suring error be a function S, defined on pairs of values of quantities (i.e.,
on pairs of real numbers) and yielding magnitudes of error which obey
the usual conditions for metrics. Such a function will be called a sense
of approximation. We have already referred above to two common senses
of approximation: absolute difference, and relative error, functions of x
and y defined respectively as

Ix - Y\ and 2Ix - yl/lx + yl.

4. Defining Approximate Truth. Let us choose a first-order language

L, an interpretation I of that language, a division of the constants and
function symbols into quantity and nonquantity symbols, and an appro-
priate sense of approximation. The basic idea of the definition of ap-

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 61

proximate truth is that a statement will be approximately true under an

interpretationI if there is an interpretationJ which is "near"I, and under
which it is actually true.
We call two interpretations of the language similar if they have the
same universe and give the same interpretationto all predicates and non-
quantity terms. For terms interpretedby the similar structures, we define
the distance between the denotations of two constants r and rTas the
distance is given by the sense of approximation, applied to the denotations
that T and XThave under their respective interpretations. If r and T are
function symbols, then the distance will be that between the graphs of r
and TTunder their respective interpretations. Similar interpretationsI and
J are a-close if for no quantity symbol (constant or function symbol) is
the distance between the denotations assigned to it by I and by J greater
than a.
We define approximate truth as follows: 0 is approximately true with
respect to a given interpretationI, sense of approximation S, and within
error limit a if and only if there is an interpretationJ which is a-close
to I such that (P is true with respect to J. A set of sentences {,P,0,,. . .}
is simultaneously approximately true within a if and only if there is a J
which is a-close to I, and such that each of ', 0, Q2, . . . is true with
respect to J. Theory T is approximately true within a if and only if all
axioms of some axiomitization of T are simultaneously approximately
true within a.
If the interpretationand sense of approximationare understood, we will
write " = V" to indicate that P is true, and "c,ta" to indicate that P is
approximately true within a. In general, approximatetruthof two or more
statements will not result in their simultaneous approximate truth. For
example, there are cases in which both K=, 0 and 1=, -~p, but it can
never happen that =a,(P&~c').
The argument:
P, 0, . . ., therefore D2
is called approximately valid if and only if for every interpretationI,
every sense of approximation S and every error limit a, if t=aP, and t,0,
and . . . then 1a,2.
All valid arguments with a single premise are approximately valid in
this sense, but most multipremise arguments are not, as the earlier dis-
cussion of modus ponens illustrates. Many such multipremise arguments
may still preserve approximate truth in various weaker senses. The tran-
sitivity of identity argument:
a = b and b = c, therefore, a = c
is not approximately valid but approximate truth of the premises within

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
62 THOMAS WESTON

a will guarantee the approximate truth of the conclusion within at most

two times a, regardless of the sense of approximation. The precise factor
by which the inference inflates the error will depend on which sense of
approximation is chosen, and on the denotations of "a", "b", and "c".
Arguments which preserve some but perhaps not all of the accuracy in
their premises will be called continuous. Arguments can also preserve
approximate truth by compensating for inaccuracy in one premise by a
kind of super-accuracy called stability in another premise (see Weston
1987).
Many valid inferences are clearly not approximately valid, although
they may preserve approximate truth to some lesser degree. This state of
affairs is what should be expected when the well-behaved problem is taken
as a model since some problems fail to be well-behaved only because
they cannot meet the continuity condition, although their solutions are
still derived by deductively valid reasoning. Hence there must be valid
arguments which are not approximately valid, even for interpretationsand
senses of approximation that are naturaland importantfor scientific prac-
tice.

5. An Example. This section illustrates what it means to say that one

physically interesting statement is approximately true, an especially well-
chewed example from Special Relativity. We employ the first-order lan-
guage with the one-place quantity symbols "m", "p", and "v" giving,
according to our chosen interpretation, the mass, momentum, and veloc-
ity, respectively, of each particle in some chosen portion of the universe.
We also include the mathematical symbols "1", "+ ", "-", ", "/",
and "V" with their usual meaning. For simplicity, we measure velocity
and momentum in one dimension only.
We can express the relations among momentum, mass and velocity as
the "classical" physicist believed them to be as follows:
p(x) = m(x) . v(x). (5)
As sense of approximation we specify the previously defined relative
error metric. We suppose, for the sake of the example, that
' - v(x)
p(x) = m(x) v(x)/V[1 . v(x)] (6)
is true, as would be dictated by Special Relativity if velocity is measured
in units of the speed of light. Figure 2 shows the least a such that (5) is
approximately true within a as a function of the maximum velocity of
any particle in the chosen portion of the universe. The lower line shows
the smallest a, calculated on the assumption that (6) is true. This is, of
course, an unreasonably strong assumption, and approximate truth would
have little use if we could only have evidence that a statement is ap-

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 63

30%

20%

10%

0%

0% 40% 80%
Figure 2. Error in (5) as a function of velocity, measured in units of the speed of light.
The lower line is the error assuming (6) is true. The upper line is the error in (5) assuming
that (6) is true within 10%.

proximately true if we knew some other statement which is precisely true.

The upper line in figure 2 calculates an upper bound on a for (5) on the
assumption that (6) is approximately true within 10 percent.2

6. Causal Constraints on Senses of Approximation. Thus far, we have

only required that a sense of approximation provide a metric on the uni-
verse of the interpretationI. This condition is satisfied by some rather
bizarre functions, and is not sufficient to give sensible results. We will
2The values shown in figure 2 are calculated as follows: Let I[rr,-r] be the denotation
of term rTin I when the variables in 7r are interpreted by the sequence o-. With relative
error as the sense of approximation, let
/3 = l.u.b. over o of 3\/V{I['p(x)' ,r]}/{I['m(x)' ,o] .I['v(x)' ,a]}].
If (6) is true and no particle has velocity greater than, say, 80 percent of the speed of
light, P will be finite. Calculation shows that
p-(x) = m(x) v(x) if and only ifa - 2f3 - 1l/(/3+ 1).
Statement (6) implies that /3 = 1/6V/{l - I['v(x)',cr]2}J, which is used to compute
the a's as a function of v-(i.e., of I['v(x)',or]-which is plotted as the lower line in
figure 2.
If (6) is not true, but is approximately true within T, /3 will be larger. If 8 = (2 + r)/
(2 - 7) then 83= 8/[6/{1 -82 2I['v(x)',o]12}l. For r = 10 percent, the resulting a's are
plotted as the upper line in figure 2.

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
64 THOMAS WESTON

see that nonmathematical conditions that make reference to the subject

matter to be described must be fulfilled.
Some criteria for selection of a sense of approximation SA are already
implicit in our earlier discussion of two intended applications of approx-
imate truth: (a) easing the problem of giving evidence for a theory by
requiring evidence only for its proximity to the truth, and (b) making
application of a theory whose known or suspected inaccuracies we wish
to ignore. These aims are likely to conflict because a theory's approxi-
mate truth will be easy to confirm to the extent that large departuresfrom
the truth will still count as approximately true. Large errors, on the other
hand, may destroy the theory's usefulness.
It is easy to see that if our aim is (a), we ought to use a SA which
counts as small only those deviations from the truth which have small
effects. This is so because the realist, or more precisely, the strong realist,
maintains that evidence for a theory is evidence for the approximate ac-
curacy of its asserted causal relations. Thus if evidence for the approx-
imate truth of a theory is to be evidence for its approximate truth as the
realist understands that theory, then the errorsthat the SA counts as small
must actually have small effects within the network of interacting entities
the theory describes. It is not necessary for a realist theory of confir-
mation that all departures which a SA counts as small should actually
have small effects, but only that effects be small within the subject matter
of the theory in question. For example, in population genetics, errors in
gene frequency counted small by an appropriate SA should have small
effects on selection pressures, relative fitness, and other causally signif-
icant magnitudes of population genetics. We need not require that soci-
ological effects-like influencing popular views about race or disease-
be small, except insofar as such effects indicate errors in the magnitudes
of population genetics proper.
Clear as it may be that this vaguely stated "small effects" condition
must be met in order to further aim (a), the confirmation of a theory
understood realistically, realists also want to argue that a "small effects"
condition must be met for aim (b). To claim this involves rejection of
instrumentalist views that a theory may be satisfactorily applied without
any suspicion that its asserted causal relations actually obtain. A critique
of this view is beyond the scope of this paper, but is developed in detail
in Boyd (1973, 1981, and 1984).
We should note that not only philosophers defend causal constraints on
approximation, but that specialists in approximation theory also argue,
for example, that the topic of approximation "lies between physical the-
ory and mathematics" (Kline 1965, vii), and that equations should be
formulated so that the "size of terms represents the size of the relevant
physical effect" (ibid., 97), allowing small terms to be dropped as "in-

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 65

significant" (ibid., 99; see remarksby Birkhoff and Synge quoted in Lay-
mon 1985, 153).
It is clear, to realists at least, that to use approximate truth for confir-
mation-realist aim (a)-or for deductive applications of theories-real-
ist aim (b)-any inaccuracy that a SA counts as small should have small
effects. This causal condition cannot be guaranteed by purely mathe-
matical conditions on SA's which can be applied to any theory, regardless
of subject matter. In particular, the absolute difference and relative error
SA's used in earlier examples cannot work equally well for other cases.
To see why not, consider the example of statistical thermodynamics.
For thermodynamic systems of any complexity, individual quantity values
of momentum, energy, and so forth figure only in averages computed
over many particles or long times. Only average values have causal sig-
nificance in the theory, but these averages can only be described by mak-
ing reference to the quantity values of individual molecules. Thus the
theory must be expressed in a vocabulary including quantity symbols for
quantities of individual molecules, for example k(x,t) for the kinetic en-
ergy of molecule x at time t, so that the averages may be computed from
them. Using a SA like absolute difference would count two systems of
molecules X and Y as nearby (i.e., a-close) when the position and mo-
mentum of each molecule in X differs from that of a corresponding mol-
ecule in Y by an amount no greater than a. Even if X and Y are composed
of the same molecules in a gas, but at slightly different times, a is likely
to be enormous, while thermodynamic laws would imply that the ther-
modynamic properties of X and Y are almost the same. A SA that does
not drastically overestimate errors can be constructed by averaging over
all molecules and some appropriateperiod of time. Knowledge of at least
the general class of systems under study might be needed to select the
method of averaging (see, e.g., Soo 1962, 374-380 for examples), but
the fact that averaging is necessary at all, and the form most appropriate
for it, is already a consequence of physical principles, not solely math-
ematical ones.
Knowledge that a particularSA is appropriatefor evaluating statements
about a subject matter requires that we have some knowledge of the sub-
ject matter of the theory. Thus we must know some part of what a theory
might tell us before we can say with any precision exactly what it means
to claim that the theory is approximatelytrue. Since the informationneeded
to select a SA is generally much less than that contained in the theory to
be evaluated with it, we will seldom risk entering the vicious circle of
needing to know something is true before we know how to say it. Specific
cases might arise, however, where appeal to approximatetruthmight really
involve some kind of vicious circle.
Fortunately, the information required for correct formulation will often

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
66 THOMAS WESTON

be in the intersection of the assertions by rival theories in the same dis-

cipline. The case of momentum according to classical mechanics and to
Special Relativity is a case of this sort. More complex cases, however,
arise where conflicting theoretical commitments lead directly to contro-
versy over what standard to use as a measure of accuracy. MacKenzie
(1978), for example, argues that the turn-of-the-century controversy be-
tween statisticians Pearson and Yule over how to measure degree of as-
sociation between nominal variables is a case of this sort. The debate on
this technical issue was particularly fierce because its outcome had po-
tential consequences for genetics and associated eugenic social views.3

7. Laymon's Objection. It is important to distinguish between the

sometimes fanciful assertions which describe the model that inspires a
calculation, and the approximate truth of statements actually used in de-
rivations and calculations. The model may be full of massless strings and
frictionless surfaces, but what the derivations usually need is for it to be
approximately true that the string has zero mass, or approximately true
that the friction is zero, and so on. The distinction between the fanciful
description and the approximations inspired by it is particularlyimportant
in considering Ronald Laymon's argument that in cases importantfor the
confirmationof various physical theories, there is "no clear sense in which
we can assign a metrical distance from the truth" (Laymon 1985, 155).
As examples of supposedly intractable cases, Laymon lists "the as-
sumption of Newton and Schwarzschild that the universe contains only
one massy body" (ibid., 154), and the point mass assumption in the clas-
sical theory of the electron. Laymon is right that if we ask how nearly
true is the assumption that the universe has only one stationary body in
it, for example, we are not going to find convincing answers. If one is
trying to calculate the gravitational field near a nearly spherical, slowly
rotating star, however, it is reasonable to expect that the consequences
concerning that field which are derived from the Schwarzschild model by
approximately valid or continuous reasoning will be approximately true.
We need evidence, of course, that small or very distant masses will pro-
duce only small perturbations in the derived field, but even when
Schwarzschild (circa 1916) wrote, his readers had reasons to believe this
on physical grounds since it would be a consequence of Newtonian phys-
ics or Special Relativity.
Laymon is concerned that if the results are derived by methods not
3The choice to ignore units of measurement in this paper has the effect that equal nu-
merical differences between quantities are treated as equally significant in the calculation
of error. This simplification may conflict with the small effects condition since errors of
equal numerical magnitude in different quantities need not be equally significant. One
remedy for this problem is to assess errors as n-tuples of errors of n fundamental quantities
with units, for example, errors 3 = (a grams, r seconds).

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 67

known to be reliable in advance, but which must themselves be validated

by producing approximately true results, then the fact that approximately
true results are derived does not add much to the confirmation of the
theory. There is a problem here, but it is a well-known one that provides
no specific objection to approximate truth. Deriving testable conse-
quences from a theory usually requires auxiliary hypotheses, and giving
evidence for the truth of these hypotheses which is independent of the
derived result may not be easy. See Putnam (1975a) for an extensive
discussion of this problem.

8. Laudan's Objection. There is another important application of the

idea that having approximately true consequences may be sufficient for
successful application of false statements, which are either not approxi-
mately true or whose degree of approximationcannot be assessed. Laudan
(1981) argues that realist claims to explain why successful science is suc-
cessful fail because "subtle fluid" theories like those for caloric and phlo-
giston were successful, but cannot be approximately true since their key
theoretical terms do not refer. Part of the explanation of the successes of
these theories is surely that they had some important consequences that
are approximately true, and these consequences led to the successful ap-
plications.
In the phlogiston theory, for example, combustion is supposed to be a
chemical reaction in which phlogiston is released when a substance con-
taining it is heated, instead of oxygen being absorbed. The true conse-
quence of this explanation is that combustion is a chemical reaction, and
this idea guided eighteenth-century chemists to analyze the products of
combustion to discover, for example, "dephlogisticated air" (oxygen) or
"phlogisticatedair" (nitrogen). Notice that many eighteenth-centuryterms
formed from "phlogisticated" do refer, even though "phlogiston" does
not, so that Cavendish, although a phlogiston theorist, could express
Lavoisier's ideas in his own terms as the true statement "water consists
of inflammable air [hydrogen] united to dephlogisticated air" (Cavendish
1766, 152).

9. Miller's Objection. Problems like those with the choice of an ap-

propriate SA can also arise in the choice of primitive vocabulary of a
theory. Using arguments along this line, David Miller (1975) maintains
that the assessment of a false theory's accuracy is not possible.
Briefly stated, Miller's argument is that if false theory T1 is more ac-
curate than false theory T2 in the values it assigns to certain quantities,
then there will often be other quantities, definable from the quantities
mentioned in T1 and T2, about which T2 will be more accurate than T1.
He concludes that comparison of relative accuracy of theories is possible

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
68 THOMAS WESTON

only with respect to a given set of quantities, and that such judgements
can be reversed by choosing formulations in terms of other, equally suit-
able quantities.
Miller bases his argument on the method of measuring accuracy of an
equation by taking the absolute difference between the left and right sides.
Weston (1987) shows this method to be unsatisfactory since pairs of ob-
viously equivalent statements, such as real number equations a = b and
lOOa = 100b, have different degrees of inaccuracy. Miller uses a similar
technique of error inflation to construct quantities about which a false
theory will yield inaccuracies as large as desired. These quantities will
not have any causally significant role in the subject matter of the theory.
If the theory is about the mechanics of motion, for example, it may have
small errors in the values it assigns to primitives like kinetic energy K or
potential energy P, but give large errors in a "quantity" like 4093K -
789342P.
Although Miller's techniques for measuring accuracy are inadequate,
his point can still be made another way. If we add a symbol for one of
his cooked-up quantities to the vocabulary of the language, the assess-
ment of a-closeness for interpretations of the theory would include the
distance between denotations of this symbol within the interpretations.
This device would allow arbitrarilylarge alterations in the degree of ap-
proximate truth of statements of the theory.
Clearly, then, we must establish an appropriatevocabulary for a state-
ment or theory in order to assess its approximate truth, even when other
choices of vocabulary yield a theory which is (in a sense we will not
elaborate here) mathematically equivalent. Miller argues that we cannot
measure accuracy, because to do so requires us to regard some magni-
tudes as "more important, more significant, more fundamental" (1975,
184) than others, a view incompatible with his "critical or falsificationist
standpoint" (ibid.).
From the point of view of the (strong) realist, part of the aim of science
is to identify the causally significant entities and relations within the sub-
ject matter of a given theory, so that it is only reasonable that a theory
be formulated in terms of these entities for assessment of approximate
truth. Since Miller denies that this is possible, I will briefly describe some
examples that show that it has physical significance to distinguish be-
tween what is real and what is a (mathematically equivalent) fiction.
According to classical mechanics, for example, the centrifugal and
Coriolis "forces" which appear in rotating coordinate systems are fic-
tional forces. These "forces" appear only as the result of choosing a non-
inertial coordinate system; there is, in fact, nothing pushing or pulling as
the "forces"are supposed to do. Classically speaking, it makes good sense
to prefer inertial formulations of the laws, despite their mathematical

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 69

equivalence with noninertial ones, because the former contain no terms

for nonexistent "forces".
The General Theory of Relativity (GR) modifies the distinction be-
tween inertial and noninertial coordinate systems, but that theory itself
provides an analogous case of a more fundamental equivalent formula-
tion. According to the usual interpretationof GR, gravity is a fictional
force, which only appears when spacetime is described in the wrong ge-
ometry, but other interpretationsregard gravity as a real force (Weinberg
1972, viii, 289). Boyd (1973) shows the potentially different empirical
consequences of mathematically equivalent versions, one with real grav-
ity and the "wrong" geometry, the other with fictional gravity and the
right geometry.

10. Comprehensiveness and Verisimilitude. Notwithstanding the im-

portance of approximate truth in a realist account of the development of
science, other concepts are surely required as well. Some way to express
growth of content or information in successive theories is needed, and
will be essentially different from approximate truth-since a tautology is
perfectly accurate-but presumably devoid of information. Theories of
verisimilitude have usually aimed to explain such a notion of content or
information, or to combine it with accuracy. This section explains briefly
why I think the prospects are very remote for a logical theory of com-
prehensiveness which is useful for realism.
The usual approach to comprehensiveness has been to suppose the ex-
istence of a most complete theory and then to devise a metric that assesses
the distance between a statement or theory in question and that fullest
truth. Realism is not committed to the existence of such a grand theory,
even for limited subject matters, since it does not imply that humans have
the capacity even to formulate a perfect description of reality.
If the existence of such a super theory were granted, however, it would
not be reasonable to assume that it could be framed in the vocabulary
now available. Thus verisimilitude theories have generally had to make
do with assumed knowledge of a "most correct" theory in some given
vocabulary.4
Acknowledging these difficulties, Ilkka Niiniluoto has developed a
framework of verisimilitude measures. Let us give a sketch of his ap-
4If "theory" means something that a working scientist (as opposed to a logician) might
recognize as a theory, then there need not even be a most accurate theory, much less a
most comprehensive one. For example, Fejer's theorem (Phillips 1971, 313) guarantees
that an astronomical theory employing sufficiently many epicycles can approximate any
continuous periodic orbit as closely as desired. Only for very special orbit shapes, how-
ever, will a finite sum of circular motions give a perfectly accurate representation. In the
remaining cases, including the case of an elliptical orbit, there will be no finite set of
statements in the language of epicycles plus arithmetic that describes the orbit accurately.

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
70 THOMAS WESTON

proach and show its dim prospects for a realist explanation of compre-
hensiveness. Since Niiniluoto's philosophical framework is a version of
realism (although a weaker one than that assumed here), and since his
work on truthlikeness is the most thorough in the literature, poor pros-
pects for his theory extend to the whole project of verisimilitude.
To apply Niiniluoto's framework, we must first pick a language and a
"cognitive problem" in that language. Specifying a cognitive problem
requires (1) a set of mutually exclusive and jointly exhaustive statements
Oi of potential complete answers to some question, (2) a metric A(0i, Oj)
on this set, and (3) designation of one of these answers 0* as the most
complete true answer, called the "target". An answer 0 to the problem
is a disjunction 0 = 0 V . . .V k of one or more of the potential
complete answers.
It is not obvious how to use the metric A to specify what it means for
an answer 0 to be close to the target, or closer than some other answer
0'. Niiniluoto's proposal is to give a metric A, that depends on A and
two real parameters T and 7r, 0 - T, Tr - 1. The metric AS is a weighted
average of two terms, Amin and Asum,and T and' 7 are their respective
weights. The term Amin is intended to express degree of inaccuracy, and
for some choices of the metric A it does so in a manner essentially the
same as approximate truth.5Other A's give results incompatible with ap-
proximate truth. The term Asumis intended to measure information defi-
ciency of a problem answer 0, so is a maximum if 0 is a tautology,
which expresses no information. For answer 0 = 01 V . . . V 0K, Asum
is the sum (or an integral, if the Oi form a continuum) of the distances
A(0i, ,):

ZAsum(0,0) = Ai(i,0*)/N

where N is an appropriatenormalizing constant to make 0 - Asum- 1.

To apply this definition, Niiniluoto needs a way of expressing a state-
ment as a disjunction of mutually exclusive basic answers. For the ma-
jority of the cases he studies (Niiniluoto 1982, 1984, 1987), he does this
with Q-predicates and state- or structure-descriptionstaken from Carnap's
inductive logic, and their generalizations in Hintikka's disjunctive normal
form theorem. This strategy introduces the familiar host of problems that
5In order for Niiniluoto's Amin to correspond to approximate truth, the cognitive problem
should take as target a statement that some structure I is the most correct interpretation of
L, and as possible answers all statements declaring some structure J similar to I to be the
correct interpretation. Distance between statements may be defined as distance between
corresponding structures, as in the discussion of a-closeness in the text. In Niiniluoto
(1987; section 11.4), Niiniluoto's definition of a statement's "applying approximately" to
a structure would have the same effect as my definition of approximate truth if the metric
Aminwere used to explicate closeness, although he does not develop this idea.

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 71

accompany these devices in Carap's work (see Nagel 1963), and several
especially troubling difficulties result. We mention a few.
For approximate truth, it is easy to prove that adding to the vocabulary
of a language L will not affect the approximate truth of statements not
containing the new vocabulary. In contrast A, behaves badly under ex-
tension of the language: The A1 of a statement in the old vocabulary will
nearly always change-increasing for some vocabulary additions, de-
creasing for others. It can also easily happen that a statement 0 is closer
to the truth than statement O' (via A,) before additions to the vocabu-
lary, but O' is closer than 0 afterwards.
The upshot is that A, can only give an assessment of comprehensive-
ness useful in describing progress toward the whole truthif the vocabulary
of the cognitive problem is already known to be complete in the sense
that no further additions will help give a fuller description of the subject
matter. It is doubtful, to say the least, that we ever have strong evidence
that this is the case. Lacking such evidence we will be unable to use A,
to assess comprehensiveness since part of what is being asked in the
question "how comprehensive is this theory?" is precisely whether the
vocabulary will need further supplementation to present the truth more
fully.
Acknowledging these difficulties, Niiniluoto maintains that truthlike-
ness of a statement "is not a measure of its distance from the 'whole
truth', but from a chosen target" (1987, 449), and expanding the language
will change the target. It is hard to see, however, how he can grant this
and still maintain that the concept of truthlikeness is "crucially important
for developing a realist view of science" (ibid., 471) if it does not in fact
allow assessment of comprehensiveness. Worse yet, Niiniluoto does not
seem to be able to give even one realistic application of Asum,the com-
prehensiveness term in A,, to an example from the history of science.
The examples in Niiniluoto (1987) divide roughly into two categories:
(a) those that use an approach similar to approximate truth, and (b) those
which use the apparatus of Q-predicates. Cases of the first sort include
some assessments of actual or candidatephysical laws, (e.g., Snell's Law).
These examples either do not use the comprehensiveness measure Asumat
all, or use it only to assess a vague statement against a precise target
(e.g., Niiniluoto 1987, 391).
Cases of type (b), which do use Asum,are exclusively toy examples,
which illustrate the technical apparatusof verisimilitude, but are not real
applications of it. One example describes a problem of classification of
the sex and age at death of a skeleton of homo erectus in terms of the
five predicates: male, female, child, young adult, older adult (ibid., 293).
Niiniluoto calculates the distance of various incorrect classifications from
the supposed correct classification of female child, subject to some ar-

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
72 THOMAS WESTON

bitrary assumptions about distances among the various predicates. Even

if these assumptions had some plausible basis, however, it is obvious that
a five-predicate vocabulary is not adequate for a comprehensive descrip-
tion of the skeleton or of the individualwhose skeleton is was. It is equally
obvious that expansion of the five-predicate vocabulary to a more realistic
(but still incomplete) vocabulary can alter the verisimilitude measures by
arbitrarilylarge amounts. Thus we do not have an example of compre-
hensiveness, nor an example that might be expanded into a real appli-
cation. Unfortunately, this is typical of the type (b) cases.
Despite the limitations of type (b) examples, a still greater obstacle
arises in applying the Carnap-Hintikka apparatus to polyadic theories,
which are the only kind capable of modeling a real scientific theory.
Carnap's Q-predicate apparatus does not work for such theories, and
Hintikka's disjunctive normal form applies to polyadic theories only in a
relativized form, yielding a normal form that depends on a numerical
parameter called the depth. Roughly speaking, the depth specifies how
many entities will have their relationships represented in the normal form
(Hintikka 1970). Except in cases where we know that some specific small
number of entities are all those involved-surely a very unusual case if
we are treating an entire theory or a major application of it-the choice
of a depth d will not be dictated by any empirical properties of a theory
or its intended application. However, both the verisimilitude values that
A, assigns to a theory and the relative ranking of theories by verisimili-
tude that A, gives will depend on this essentially arbitrarychoice of depth.
Niiniluoto describes the state of his research as having solved the log-
ical problem of verisimilitude, but not having solved the "pragmaticprob-
lem" of how to apply the logical theory to actual cases (1987, 263; chap.
13.4). A more plausible assessment is that his approach makes it unlikely
that application to real cases is possible, and that is a good reason to
doubt that the "logical problem" has in fact been solved. In any case, we
can not use truthlikeness as a tool for interpreting the development of
science if it cannot be applied to real cases in the history of science.6
None of these problems are attributableto errors or inadequate devel-
opment of Niiniluoto's program, which has been very thorough in fact
(see Niiniluoto 1987). The fundamental problem is that distance from the
chosen target of a cognitive problem is no substitute for distance from
the truly comprehensive theory. Such a grand theory need not exist, how-
6Space does not permit discussing Niiniluoto's view that the choice of a measure of
similarity is "pragmatic", and depends on "our cognitive interests" (Niiniluoto 1987; chap.
13.4). My view is that the choice of similarity measure for a given subject matter must
be primarily empirical, rather than pragmatic, and that the empirical constraints on an
acceptable measure would include a small effects condition, as in the choice of a sense
of approximation. Niiniluoto's pragmatic substitute for a small effects criterion is the "cog-
nitive seriousness" (ibid., 315) of errors.

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
 APPROXIMATE TRUTH AND SCIENTIFIC REALISM 73

ever, and the supposition that it does is not a modest idealization from
scientific practice.

11. Conclusion. Without a satisfactory logical theory of comprehen-

siveness, the logical tools for describing scientific progress are incom-
plete. Approximate truth does not do everything we want in explaining
what progress is, but it does have the very important applications de-
scribed above, especially in the clarification and defense of realism, and
it survives the criticisms made by Miller, Laudan and Laymon.
We have also seen that important elements of (strong) realism are re-
quired to give a coherent explanation of approximate truth, specifically
that causal criteria must be used to select the right vocabulary and ac-
curacy measure for a theory.
This interdependence between realism and approximate truth is not a
defect, but a conclusion useful for the realist. To the extent that the pres-
ent theory of approximate truth should prove successful, the realist is in
a position to explain why the widespread use of approximation, ideali-
zation and modeling in science is not mere wishful thinking, but will,
under conditions we can partly specify, lead to nearly true conclusions.
Clearly an instrumentalist will not be justified in using a theory of ap-
proximate truth that depends on strong realism, and I believe that the
arguments in this paper show great obstacles to a purely pragmatic ac-
count of approximate truth. Thus, subject to further development-tech-
nical and philosophical-of the theory of approximate truth, approxi-
mation can be added to the list of features of science that realists can
explain much more satisfactorily than instrumentalists.
REFERENCES
Adams, E. (1974), "The Logic of 'Almost All'", Journal of Philosophical Logic 3: 3-
17.
. (1982), "Approximate Generalizations and Their Idealization", in P. Asquith and
T. Nickles (eds.), PSA 1982, vol. 1. East Lansing: Philosophy of Science Associa-
tion, pp. 199-207.
Boyd, R. (1973), "Realism, Underdetermination, and a Causal Theory of Evidence", Nous
7: 1-12.
. (1981), "Scientific Realism and Naturalistic Epistemology", in P. D. Asquith and
R. N. Giere (eds.), PSA 1980, vol. 2. East Lansing: Philosophy of Science Asso-
ciation, pp. 613-662.
. (1984), "The Current Status of Scientific Realism", in J. Leplin (ed.), Scientific
Realism. Berkeley and Los Angeles: University of California, pp. 41-82.
. (1985a), "Lex Orandi est Lex Credendi", in P. M. Churchland and C. A. Hooker,
(eds.), Images of Science. Chicago: University of Chicago Press, pp. 3-34.
. (1985b), "Observations,ExplanatoryPower, and Simplicity: Toward a Non-Humean
Account", in P. Achinstein and 0. Hannaway (eds.), Observation, Experiment, and
Hypothesis in Modern Physical Science. Cambridge, MA: MIT Press, pp. 47-94.
Cavendish, H. ([1766] 1952), "Three Papers, Containing Experiments on Factitious Air",
in H. M. Leicester and H. S. Klickstein, A Source Book in Chemistry, 1400-1900.
Cambridge, MA: Harvard University Press, pp. 134-153.

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions
74 THOMAS WESTON

Davis, P. J. (1975), Interpolation and Approximation. New York: Dover.

Hilpinen, R. (1976), "Approximate Truth and Truthlikeness", in M. Przelecki, K.
Szaniawski, and R. Wojcicki, Formal Methods in the Methodology of the Empirical
Sciences. Dordrecht: Reidel, pp. 19-42.
Hintikka, J. (1970), "Surface Information and Depth Information", in J. Hintikka and P.
Suppes, Information and Inference. Dordrecht: Reidel, pp. 263-297.
Huber, P. J. (1972), "Robust Statistics", Annals of Mathematical Statistics 43: 1041-
1067.
Iyanaga, S. and Y. Kawada (1977), Encyclopedic Dictionary of Mathematics. Cambridge,
MA: MIT Press.
Kline, S. J. (1965), Similitude and Approximation Theory. New York, Academic Press.
Kuhn, T. S. (1970), The Structure of Scientific Revolutions. 2d ed. Chicago: University
of Chicago.
LaSalle, J. and S. Lefchetz (1961), Stability by Liapunov's Direct Method. New York:
Academic Press.
Laudan, L. (1977), Progress and Its Problems: Towards a Theory of Scientific Growth.
Berkeley and Los Angeles: University of California.
. (1981), "A Confutation of Convergent Realism", Philosophy of Science 48: 19-
49.
Laymon, R. (1985), "Idealization and the Testing of Theories by Experimentation", in P.
Achinstein and 0. Hannaway (eds.), Observation, Experiment, and Hypothesis in
Modern Physical Science. Cambridge, MA: MIT Press, pp. 147-173.
Leplin, J. (1984), "Truthand Scientific Progress", in Scientific Realism. Berkeley and Los
Angeles: University of California, pp. 193-217.
Mackenzie, D. (1978), "Statistical Theory and Social Interests: A Case-Study", Social
Studies of Science 8: 35-83.
Mikhlin, S. (1970), MathematicalPhysics: An Advanced Course. Amsterdam:North-Holland.
Miller, D. (1975), "The Accuracy of Predictions", Synthese 30: 159-191.
Nagel, E. (1963), "Carnap'sTheory of Induction", in P. A. Schilpp (ed.), The Philosophy
of Rudolf Carnap. LaSalle: Open Court, pp. 785-825.
Niiniluoto, I. (1982), "Truthlikeness for Quantitative Statements", in P. Asquith and T.
Nickles (eds.), PSA 1982, vol. 1. East Lansing: Philosophy of Science Association,
pp. 208-215.
(1984), Is Science Progressive? Dordrecht: Reidel.
. (1987), Truthlikeness. Dordrecht: Reidel.
Phillips, E. R. (1971), An Introduction to Analysis and Integration Theory. Scranton, PA:
Intext Educational Publishers.
Popper, K. (1962), Conjectures and Refutations. London: Basic Books.
Putnam, H. (1975a), "The 'Corroboration' of Theories", in Philosophical Papers. Vol.
1, Mathematics, Matter and Method. Cambridge, England: Cambridge University
Press, pp. 250-269.
. (1975b), "How Not to Talk about Meaning", in Philosophical Papers. Vol. 2,
Mind, Language, and Reality. Cambridge, England: Cambridge University Press, pp.
117-131.
Smart, J. J. C. (1968), Between Science and Philosophy. New York: Random House.
Soo, S. L. (1962), Analytical Thermodynamics. Englewood Cliffs: Prentice-Hall.
Thom, R. (1975), Structural Stability and Morphogenesis: An Outline of a General The-
ory. Reading, MA: Addison-Wesley.
Trigg, R. (1980), Reality at Risk: A Defence of Realism in Philosophy and the Sciences.
Brighton: Harvester Press.
Weinberg, S. (1972), Gravitation and Cosmology. New York: Wiley.
Weston, T. (1977), Abstract for "Approximate Truth", Journal of Symbolic Logic 42:
157-158.
. (1987), "Approximate Truth", Journal of Philosophical Logic 16: 203-227.

This content downloaded from 141.233.160.21 on Mon, 09 Nov 2015 05:10:05 UTC
All use subject to JSTOR Terms and Conditions