Okasha, S. - Theory Choice and Social Choice - Kuhn Versus Arrow PDF

Theory Choice and Social Choice:
Kuhn versus Arrow

Samir Okasha
University of Bristol
[email protected]
Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Kuhn’s famous thesis that there is ‘no unique algorithm’ for choosing between rival
scientific theories is analysed using the machinery of social choice theory. It is
shown that the problem of theory choice as posed by Kuhn is formally identical
to a standard social choice problem. This suggests that analogues of well-known
results from the social choice literature, such as Arrow’s impossibility theorem, may
apply to theory choice. If an analogue of Arrow’s theorem does hold for theory
choice this would refute Kuhn’s thesis, but it would also pose a threat to the
rationality of science, a threat that is if anything more worrying than that posed
by Kuhn. Various possible ‘escape routes’ from Arrow’s impossibility result are
examined, in particular Amartya Sen’s idea of ‘enriching the informational basis’.
It is shown that Sen’s idea can be applied to the problem of theory choice in
science. This in turn sheds light on two well-known approaches to inductive
inference in philosophy of science: Bayesianism and statistical model selection.
1. Introduction
In the Postscript to The Structure of Scientific Revolutions, Thomas
Kuhn famously argued that there is ‘no neutral algorithm for theory
choice’ in science. Kuhn allowed that scientists might have good
reasons for choosing one theory over its rivals, citing ‘accuracy, sim-
plicity, fruitfulness, and so on’ as examples of such reasons, but he
insisted that they fall short of providing an algorithm (Kuhn 1969,
p. 199). Even if two scientists agree on the features that a good theory
should have, they will not necessarily be led to make the same choices,
Kuhn argued, for they may weight the features differently. For ex-
ample, the two scientists might agree that accuracy and simplicity
are both important theoretical virtues, but disagree about their relative
importance, and thus be led to choose different theories. Neither can
be called irrational, Kuhn insisted, and neither was necessarily acting
unscientifically.
The idea that there is no algorithm for theory choice met with a
favourable response from Kuhn’s critics, even among those
Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

doi:10.1093/mind/fzr010 Advance Access publication 31 March 2011
84 Theory Choice and Social Choice: Kuhn versus Arrow
unsympathetic to other aspects of his philosophy, and seems to be

regarded as fairly uncontroversial.1 In part, this is probably because
Kuhn’s ‘no algorithm’ claim tallies well with the widely-held view that
Carnap-style inductive logic is an impossible dream. Carnap’s aim was
precisely to devise an algorithm for inductive reasoning — a rule that,
given a set of hypotheses and a body of data as input, would tell us
which hypothesis is best confirmed by the data. The perceived failure
of Carnap’s project, which Kuhn himself alluded to in a later essay,2
lent credence to his claim that there is no algorithm for theory choice.
My aim in this paper is to explore Kuhn’s ‘no algorithm’ thesis from

a new angle, by drawing on some ideas from social choice theory,
in particular Arrow’s famous impossibility theorem (Arrow 1951).
Social choice theory studies the problem of how a society should
choose between a set of ‘social alternatives’, given that the individuals
in society may have different preferences over those alternatives.
I show that the problem of theory choice, as described by Kuhn,
is formally identically to a standard social choice problem. This
raises an interesting possibility: namely, that we might be able to
use formal results from social choice theory, such as Arrow’s theorem,
to assess Kuhn’s claim that there is no algorithm for theory choice.
That is my primary goal in this paper; a secondary goal is to illustrate
more generally how techniques from theoretical economics can be
applied to problems in epistemology.
The structure of this paper is as follows. Section 2 examines
and clarifies Kuhn’s ‘no algorithm’ argument and its philosophical
consequences. Section 3 introduces social choice theory and gives a
non-technical exposition of Arrow’s theorem. Section 4 shows how
the problem of theory choice can be cast in a social choice framework,
and asks whether an analogue of Arrow’s theorem applies to theory
choice. Section 5 examines possible ‘escape routes’ from Arrow’s the-
orem. Section 6 discusses the most promising escape route, namely
Amartya Sen’s ‘informational basis’ approach. Section 7 applies Sen’s
approach to the problem of theory choice. Section 8 shows how the
foregoing results shed light on two well-known approaches to induct-
ive inference in philosophy of science: Bayesianism, and statistical
model selection. Section 9 draws together the pieces and concludes.
1
See for example the discussion in Newton-Smith 1981.
2
In that essay Kuhn wrote ‘most philosophers of science would … I think, now regard the
sort of algorithm which has traditionally been sought as a not quite attainable ideal’ (Kuhn
1977a, p. 328).

Samir Okasha 85
2. Kuhn’s ‘no algorithm’ argument: clarifications

Kuhn developed his ‘no algorithm’ argument most thoroughly in a
1977 essay entitled ‘Objectivity, Value Judgment, and Theory Choice’.
In that essay, he identifies five criteria that provide ‘the shared basis
for theory choice’, namely accuracy, consistency, scope, simplicity,
and fruitfulness (Kuhn 1977a, p. 321). These five, he says, are ‘the
standard criteria for evaluating the adequacy of a theory ’, widely
agreed on by mainstream philosophers of science. Kuhn has no quar-
rel with the standard view that these criteria play a key role in scientific
theory choice; indeed, he regards them as partially constitutive of what

science is. However, he argues, using examples from the history of
science, that the criteria fail to uniquely determine theory choice, for
two reasons. Firstly, the criteria are ambiguous — it may be unclear
which of two theories is simpler, for example. In some respects
Copernicus’ theory was simpler than Ptolemy ’s, Kuhn says, but in
others it was not. Secondly, there is the problem of how to appropri-
ately weight the criteria when they pull in different directions. How
should simplicity be traded off against accuracy and scope, for ex-
ample? Kuhn says that ‘no progress’ has been made towards solving
this problem (Kuhn 1977a, p. 329).
It is this second problem — weighting — that will be the focus of
our attention here, so the first problem — ambiguity — will be
ignored. In any case, the first problem arguably collapses into the
second. Consider Kuhn’s own example. He says that Copernicus’
theory was simpler than Ptolemy ’s in that it invoked more parsimo-
nious mathematics, but was no simpler in that the computational
labour required to predict planetary positions was the same for
both. If this is correct, then simplicity, in this example, needs to be
sub-divided into two criteria: mathematical parsimony and computa-
tional ease, neither of which is ambiguous. Of course, this raises the
question of how the two types of simplicity should be weighted, which
is more important, etc. But that is just an instance of the second
problem. It seems, therefore, that the ambiguity of Kuhn’s five criteria
provides no principled reason to doubt the existence of an algorithm
for theory choice. Disambiguation can always be carried out by
sub-dividing an ambiguous criterion, though this exacerbates the
weighting problem. So the latter is more fundamental.
Suppose it is true that there is no algorithm for theory choice. What
follows? The conclusion Kuhn drew is that value judgements play an
inevitable role in theory choice, and thus that the ideal of ‘objectivity ’,

as that notion was understood by traditional philosophers of science,

is unattainable. This does not mean that theory choice is irrational,
Kuhn stressed, or that ‘anything goes’, but rather that the traditional
conception of rationality is too demanding. Two scientists, on the
basis of the same empirical evidence, could arrive at different theories
without either of them being irrational, Kuhn argues. So the ‘no al-
gorithm’ argument does not undermine the rationality of science, he
thinks, but rather forces us to a more realistic conception of what
rational theory choice is like. Whether or not we accept this, Kuhn
is surely right that the ‘no algorithm’ argument has important epis-

temological consequences.
Importantly, when Kuhn says there is no algorithm for theory
choice, he means that there is no unique algorithm. His point is
that given the five criteria — simplicity, accuracy, fruitfulness,
etc. — which all parties agree form the shared basis for theory
choice, there are many conceivable algorithms that one could con-
struct, and no obvious way of choosing between them. (Thus Kuhn
talks about ‘the algorithms of different individuals’ in a scientific
community, while insisting that there is no such thing as ‘the algo-
rithm of objective choice’ (Kuhn 1977a, p. 328).) So the problem is
that there are too many algorithms for theory choice, each perfectly
acceptable, and no way of singling out the ‘right’ one. This is why
Kuhn sometimes expresses his thesis by saying there is no neutral
algorithm for theory choice (Kuhn 1969, p. 199).
My point in stressing this is that there is another, quite different
thing that might be meant by saying there is ‘no algorithm’ for theory
choice. One might mean that there is no acceptable algorithm at all,
not that there are too many. It might be the case that there is no way
of constructing an algorithm, based on the five criteria, which meets
minimal standards of acceptability. But this is not Kuhn’s claim.
Rather, he thinks that there are many acceptable algorithms, each of
which weights the five criteria differently, and each of which is as
rationally defensible as any other.
Another way to capture the distinction is as follows. Kuhn will
presumably agree that there are certain minimal standards which
any acceptable algorithm for theory choice must meet. For example,
if theory T1 scores higher than T2 on each of his five criteria, then an
algorithm that selects T2 over T1 is obviously unacceptable — no ra-
tional scientist could use it. So these minimal standards, whatever
exactly they are, enable us to rule out some possible algorithms for
theory choice. But on Kuhn’s view, many algorithms will still

Samir Okasha 87
remain — and there is no way of narrowing down the choice to a

single one. An alternative view, however, is that no algorithms at all
will remain after those that fail to meet the minimal standards have
been discarded. These are quite different views, and are mutually in-
compatible, though each might be expressed by saying that there is ‘no
algorithm’ for theory choice.
The distinction between these two views might be thought inconse-
quential, for both imply that there is no unique algorithm for theory
choice, though for different reasons. And if Kuhn is right that the notion
of rationality of traditional philosophy of science presumes the exist-

ence of a unique algorithm, then there is a threat to that notion in either
case. This may be so; but it is still surely important to know why there is
no unique algorithm for theory choice, if there is not. Is it because there
are many acceptable algorithms and no good way of choosing between
them, or because there are no acceptable algorithms? The epis-
temological consequences might be similar in either case, but the dis-
tinction is a real one and will play an important role in what follows.
The idea that theory choice is based on multiple criteria that may
pull in different directions is not unique to Kuhn’s philosophy of
science. Rather, it is common to diverse philosophical views on how
scientific inference works. For example, proponents of ‘inference to
the best explanation’ cite multiple factors, such as simplicity, unifying
power, and scope, which enter into the assessment of how good a
candidate explanation is (Thagard 1978, Lipton 1990). Bayesians
argue that the choice between rival theories depends on how well
they score on two different criteria — prior probability and likeli-
hood — which can conflict (Howson and Urbach 1992, Earman
1992).3 Finally, proponents of ‘statistical model selection’ argue that
the choice between rival hypotheses again depends on two factors —
fit-with-the-data and simplicity — which typically do pull in different
directions (Forster and Sober 1994, Forster 2001). So the issues to be
discussed below have a relevance that extends beyond an assessment of
Kuhn’s ‘no algorithm’ thesis.
3. Social choice theory and Arrow’s impossibility theorem

Social choice theory deals with the problem of aggregating individuals’
preferences, over a set of alternatives, into a single ‘social preference’.
3
The point being that P(H1/e) > P(H2/e) if and only if P(H1).P(e/H1) > P(H2).P(e/H2). See
Sect. 8 below.

Table 1: A profile of preference orders for four individuals

John 1. Cinema 2. School 3. Hospital 4. Airport
Mary 1. Hospital 2. Cinema 3. Airport 4. School
Jane 1. School 2. Airport 3. Hospital 4. Cinema
Peter 1. Cinema 2. Airport 3. Hospital 4. School
For example, suppose a given society has four alternatives: building a

school, a hospital, an airport, or a cinema, only one of which can be

chosen. Each individual in society is assumed to have a weak preference
order over the alternatives, that is, a ranking of the alternatives from
best to worst, with ties permitted. Formally, a weak preference order is
a binary relation that is transitive, reflexive, and complete. The pref-
erence order of the i th individual will be denoted Ri, so ‘xRiy ’ means
that the ith individual weakly prefers alternative x to alternative y, that
is, she does not strictly prefer y to x. From the weak preference relation
Ri, we can define a corresponding relation of strict preference Pi, and
of indifference Ii.4
Suppose that for each of the n individuals in society, we know
their preference order over the four alternatives. We can encode all
this information in a single profile, denoted <R1, … , Rn>, which is
simply a list, or vector, of preference orders, one for each individual.
An example of a profile is contained in Table 1 above. Given this
information, we would then like to be able to construct a single
social preference order R, which ranks the four alternatives in terms
of how good they are ‘for society as a whole’.5 On any reasonable
ethical view, the social preference order should depend somehow on
the preference orders of the individuals in society, but how exactly?
In his seminal 1951 work, Kenneth Arrow devised a novel way to
study this problem. Arrow’s idea was to consider a function from
profiles of individual preference orders to a social preference order,
and then ask what conditions the function should satisfy. I will call
such a function a ‘social choice rule’.6 The rule takes as input a profile
of preference orders, and yields as output a single social preference
4
Thus xPiy iff xRiy and it is not the case that yRi x, while xIiy iff xRiy and yRi x.
5
A social preference order is an object of the same sort as an individual preference order,
i.e. a transitive, reflexive, complete binary relation over the alternatives.
6
Arrow himself used the expression ‘social welfare function’, but this is often used in
another sense today.

Samir Okasha 89
order. In other words, given the preferences of all the members of

society over the alternatives, the rule tells us what the social preference
order should be. A social choice rule is thus a kind of algorithm for
making social choices, based on information about individuals’
preferences.
Arrow proposed four conditions that any reasonable social
choice rule should satisfy: unrestricted domain (U), weak Pareto
(P), independence of irrelevant alternatives (I), and non-dictatorship
(N). He then proved, remarkably, that there exists no social choice
rule that satisfies all four conditions, so long as there are at least three

social alternatives. That is the content of his famous ‘impossibility
theorem’. If we agree with Arrow that all four conditions are reason-
able ones, his result spells bad news for the possibility of making
coherent social decisions.
Condition U says that the domain of the social choice rule is the set
of all possible profiles. This means that whatever the preferences of the
individuals in society, the rule must output a social preference order,
that is, there is no a priori restriction on the preferences that individ-
uals are allowed to have. In many applications of Arrow’s framework,
this condition is extremely natural. For example, if the ‘alternatives’
are candidates in an election, then condition U says that voters can
rank the candidates however they like, and the rule must still output a
social preference, that is, an election result. This is obviously reason-
able. In other applications, restricting the domain of the social choice
rule may make sense; for example, if the alternatives are different ways
of dividing up society ’s resources, then the assumption that individ-
uals prefer more to less, ceteris paribus, suggests a natural domain
restriction. But in general, condition U is well-motivated.
Condition P — weak Pareto — says that if all individuals in society
strictly prefer alternative x to y, then society should also prefer x to y,
that is, the social preference order must rank x above y. This seems
indisputable: if everyone would rather have a cinema to a
swimming-pool, then ‘cinema’ should obviously be higher than
‘swimming pool’ in the social ranking. This captures the intuitive
idea that social choices must reflect what the members of society
want; so if they all want the same thing, that is what should be chosen.
Condition N — non-dictatorship — says that there cannot be an
individual who is such that whenever he or she strictly prefers alter-
native x to y, so does society. Such an individual would be a dicta-
tor — their preferences would automatically override those of all other
members of society. The existence of a dictator is clearly undesirable,

as it conflicts with basic democratic ideals. So condition N seems

unexceptionable.
Condition I — independence of irrelevant alternatives — is at the
crux of Arrow’s argument, and is slightly trickier than the others. It
says that the social choice between alternatives x and y can only
depend on individuals’ preferences between x and y — not on their
preferences over other alternatives. More precisely, consider two pro-
files of individual preference orders <R1, … , Rn> and <R91, … , R9n>,
such that for every individual i, xRiy if and only if xR9iy, that is, each
individual’s preference for x over y is the same in the two profiles.

Condition I then says that the social choice rule, when applied to
both profiles, must yield the same social preference for x over y,
that is, xRy if and only if xR9y. Any differences between the two pro-
files are irrelevant to the social choice between x and y, according to
condition I, since the two profiles are identical in the only respect that
matters.
The intuitive force of condition I can be seen by considering an
election in which voters must rank three candidates, Labour, Tory,
and Liberal, in order of preference. Various different ways of aggre-
gating the individual preferences into a single social preference are
conceivable. Condition I imposes a requirement on acceptable aggre-
gation schemes — it says that the social preference between the Labour
and Tory candidates, for example, can depend only on the individuals’
preferences between Labour and Tory. This is highly intuitive — in
order to determine whether the Labour or Tory candidate is socially
preferable, surely the individuals’ preferences involving the Liberal
candidate should not matter. If you know of each individual whether
they prefer the Labour to the Tory candidate (or are indifferent), then
you know everything that is relevant to determining the social choice
between these two candidates, according to condition I. Arrow wrote
that his four conditions ‘taken together, express the doctrines of citi-
zens’ sovereignty and rationality in a very general form’ (Arrow 1951,
p. 31). Condition I is in fact somewhat controversial, as discussed
below, but nonetheless Arrow’s four conditions arguably represent
quite reasonable constraints on a social choice rule. But, remarkably,
Arrow proved that all four cannot be simultaneously satisfied, so long
as there are at least three alternatives; equivalently, any social choice
rule that satisfies conditions U, I, and P must be a dictatorship of one
individual.
Arrow could have expressed his impossibility result by saying that
there is ‘no algorithm’ for social choice that meets certain reasonable

Samir Okasha 91
conditions. This way of expressing Arrow’s theorem immediately sug-

gests a comparison with Kuhn’s views on theory choice, to which I
now turn.
4. Theory choice cast as a social choice problem

The problem of theory choice, as formulated by Kuhn, hinges on the
fact that there are multiple desiderata that we want our theories to
satisfy — simplicity, accuracy, scope, etc. A theory that scores well on
one desideratum might score badly on another, hence the weighting

problem that Kuhn discusses. This problem may seem quite different
to the social choice problem as formulated by Arrow, but in fact the
two share a common structure. The key to seeing this is to regard each
criterion of theory choice as an ‘individual’, with their own ‘preference
order’ over the alternative theories. This may sound odd, but can be
easily explained.
Take, for example, simplicity. Let us assume that simplicity can be
defined reasonably precisely, enough to permit pair-wise comparisons
between the theories that we wish to choose between.7 Then, we can
define a binary relation ‘is at least as simple as’, on the set of alterna-
tive theories, which will be a weak ordering, that is, reflexive, transi-
tive, and complete. Let us do the same for accuracy, scope, and the
other Kuhnian criteria. From a formal point of view, each criterion is
then analogous to an individual in Arrow’s set-up. Just as each indi-
vidual rank-orders the social alternatives according to how much they
like them, so each criterion rank-orders the alternative theories ac-
cording to how well they satisfy it. So each of Kuhn’s criteria corres-
ponds to an individual in Arrow’s framework, and the alternative
theories correspond to the social alternatives.
It might be objected that for some criteria of theory choice, the
binary relation will not be complete. Take for example scope.
Plausibly, one might take a theory ’s ‘scope’ to be its total set of logical
consequences, and the relation ‘T1 has at least as much scope as T2’ to
mean that T2’s consequence class is a subset of T1’s. But this relation,
though reflexive and transitive, need not be complete, for the conse-
quence classes of a pair of theories may be non-nested, that is, the
theories may be non-comparable for scope. Though this is a valid
point, ‘scope’ is arguably the only one of Kuhn’s five criteria that it
7
This assumption is not unproblematic, but the problems it raises are orthogonal to those
under discussion here.

affects. (In the case of simplicity, for example, it is plausible that for
any two theories, either one is simpler than the other or they are
equally simple, i.e. ‘is at least as simple as’ is complete.) So the com-
pleteness assumption can be justified as a reasonable idealization.
After all, the assumption that individuals’ preference relations are
complete is also an idealization.8
The next step is to consider a ‘theory choice rule’, defined by direct
analogy with an Arrovian social choice rule. Given a profile of weak
orders, one for each criterion of theory choice, a theory choice rule
yields a single ordering of the alternative theories. So, for example,

suppose we have four alternative theories, and three criteria: simpli-
city, accuracy, and scope. By assumption, we know how to rank-order
the theories by each criterion. We feed this information into the
theory choice rule, which then outputs an ‘overall’ ranking of the
theories, from best to worst. Formally, the theory choice rule is
defined in exactly the same way as Arrow’s social choice rule.
Next, let us ask whether Arrow’s four conditions apply to the theory
choice rule. Condition U (unrestricted domain) seems unexception-
able — however the theories are ranked by the various criteria, the rule
must be able to yield an overall ranking. There should be no a priori
restriction on the permissible rankings that are fed into the rule.
Such a restriction might make sense if there is an intrinsic trade-off
(or correlation) between two of the criteria. For example, if greater
simplicity always involves a sacrifice of accuracy, then the simplicity
rank-ordering will be the inverse of the accuracy rank-ordering. This
will rule out some possible inputs to the rule, which implies a natural
domain restriction. But unless we have specific reason to think such
trade-offs must always obtain, condition U seems reasonable.
Condition P (weak Pareto) seems undeniable. If theory T1 does
better than theory T2 by each of Kuhn’s criteria, that is, it is simpler
and more accurate and more fruitful etc., then it must surely be
preferred overall. This seems as obvious as its analogue for social
choice. What about condition N (non-dictatorship)? It says that
8
There is in fact a technical trick to get around the problem. Suppose R is reflexive and
transitive but incomplete. We can then extend R to a complete relation R*, by stipulating that
for any two objects x and y that are not related by R, xI*y, i.e. neither xR*y nor yR*x. The
relation R* will then be reflexive and complete, but non-transitive; however, it will be
quasi-transitive (which means that P*, the corresponding strict preference relation, is transitive,
but I* is not) (Sen 1969). Arrow’s theorem will then apply; for the theorem does not in fact
require that the individual preference orders be fully transitive — quasi-transitivity is enough.
(By contrast, the full transitivity of the social preference order is essential to the theorem.)

Samir Okasha 93
there is no one criterion such that if T1 is ranked above T2 by that

criterion, then T1 is automatically above T2 in the overall ranking. This
condition makes good sense, so long as we agree that all the criteria
are relevant to theory choice. Violation of the condition would mean
that one criterion, for example simplicity, was regarded as so import-
ant that a less simple theory would never be preferred to a more simple
one, however highly it scored on the other criteria.
What about condition I (independence of irrelevant alternatives)?
It says that the overall ranking of T1 and T2 should depend only on
how the criteria rank T1 and T2, not on how they rank other theories.

So, for example, suppose we have three criteria, simplicity, accuracy,
and scope, and two theories. Suppose T1 is simpler than T2, T2 is more
accurate than T1, and T1 has greater scope than T2. Condition I says
that this is all the information that is relevant to the overall ranking of
T1 versus T2; so if in this case the theory choice rule ranks T1 above T2
(for example), then it must rank T1 above T2 in every relevantly similar
case, that is, every case where T1 is simpler than T2, T2 more accurate
than T1, and T1 greater in scope than T2. As with the social choice rule,
this condition has strong intuitive appeal, capturing the idea that ra-
tional theory choice should not depend on irrelevant factors.
If we agree that U, P, N, and I are conditions on reasonable theory
choice, then it is obvious that an Arrovian impossibility result applies.
So long as there are at least three alternative theories, there exists no
theory choice rule that satisfies all four conditions. This spells bad
news for the possibility of making ‘rational’ theory choices.
One might naturally express this impossibility result by saying that
there can be ‘no algorithm’ for rational theory choice. This sounds
similar to what Kuhn said, but recall the discussion of section 2. As we
saw, Kuhn meant that there is no unique algorithm; he argued that the
multiple criteria for theory choice could be combined into a decision
rule in many ways, and there is no good way of choosing between
them. So the problem according to Kuhn is that there are too many
algorithms. But the Arrow impossibility result implies the opposite —
there is no algorithm for theory choice that meets reasonable condi-
tions. Where Kuhn saw an embarrassment of riches, Arrow tells us
that there is nothing at all.
Both Kuhn’s view and the Arrow-inspired view imply, obviously,
that there is no single algorithm for theory choice, over three or more
alternatives, which is rationally acceptable. This conclusion conflicts
with the traditional ideal of rationality, associated with Carnap, ac-
cording to which two rational agents with the same ‘total evidence’

must end up in the same epistemic state. In the context of theory

choice, the Carnapian ideal implies that if two scientists agree on
what the relevant criteria are, and agree about how well each theory
performs on each criterion, then they should agree on how to rank the
theories. On Kuhn’s view, such agreement is unattainable — the two
scientists may weight the criteria using different algorithms, and there
is no saying which is rationally correct. On the Arrovian view, agree-
ment is also unattainable, but for a very different reason, namely that
no algorithm meets minimal standards of rationality.
Despite both implying that the traditional ideal of rationality

cannot be met, the Kuhnian and Arrovian views are diametrically
opposed. Moreover, it makes a big difference which is right. If
Kuhn is right that the problem is too many algorithms, two options
suggest themselves. First, we might seek further conditions that any
acceptable algorithm must satisfy, to narrow down the pool. Second,
we might try to liberalize the notion of rationality, and argue that two
scientists could both count as rational despite employing different
algorithms for theory choice.9 But if the Arrovian view is right,
then neither of these options holds any promise and our epistemo-
logical predicament is correspondingly more serious. Put differently,
Kuhn makes rational theory choice look difficult, at least if we cleave
to a certain conception of rationality, but Arrow makes it look out-
right impossible.
5. Possible escape routes

Since the Arrow-style impossibility result threatens the rationality of
theory choice, and thus of science, it would be nice if there were a way
out. Various possibilities suggest themselves. One is simply that many
real cases of theory choice are binary, that is, involve just two alter-
natives. It is striking that Kuhn’s own examples tend to be binary —
geocentrism versus heliocentrism, oxygen versus phlogiston, relativity
versus classical mechanics. And of course Arrow’s impossibility result
only holds if there are three or more alternatives. With just two alter-
natives to choose between, numerous algorithms become possible that
9
This idea was common in the post-Kuhn literature; see for example McMullin 1993. It is
suggested by certain remarks by Kuhn himself, in the Postscript to Structure of Scientific
Revolutions, where he says that his aim is not to show the irrationality of science, but
rather to arrive at a more realistic view of what scientific rationality involves.

Samir Okasha 95
satisfy conditions U, P, I, and N.10 Does this reconcile Kuhn with

Arrow?
I think the answer is ‘no’, though it is certainly interesting that the
distinction between binary and non-binary choice becomes significant
once theory choice is formulated in the social choice framework.
(By contrast, on most standard philosophical approaches to theory
choice, such as Bayesianism, it is of no particular relevance whether
the choice is binary or not.) Not all theory choice in science is binary,
even if the large-scale paradigm shifts that Kuhn is interested in typ-
ically are. More mundane cases are often not. Think for example of

climate change science, where researchers compare the merits of nu-
merous models of climate change, not just two. More generally, in
many branches of science a typical problem might involve choosing
between three candidate explanations of an observed correlation be-
tween two variables x and y: (i) x causes y; (ii) y causes x; (iii) x and y
are joint effects of a common cause. Or consider statistical estimation,
where a researcher might want to estimate the value of a real-valued
parameter in the unit interval; the alternatives that must be chosen
between are uncountably many. So focusing exclusively on binary
choice, as a way of trying to avoid the Arrovian predicament, is at
odds with scientific practice.
Another possibility for reconciling Arrow with Kuhn is simply to
reject one or more of Arrow’s conditions. Kuhn was sceptical (in some
moods) about the existence of ‘trans-paradigmatic’ criteria of ration-
ality, which are universally binding on scientists across all eras;
perhaps, then, he would argue that an acceptable algorithm for ra-
tional theory choice need not respect Arrow’s conditions. Given the
intuitiveness of those conditions, this option does not seem especially
plausible, but one possible argument against condition N (non-
dictatorship) is worth briefly discussing.
In the context of theory choice, condition N says that no criterion
can be a dictator, that is, can be such that whenever x ranks above y by
that criterion, then x ranks above y overall. However, a strong em-
piricist might well hold that the criterion of ‘fit-with-the-data’ should
be a dictator. Empiricists in philosophy of science have long argued
that criteria such as simplicity are of lesser importance than
fit-with-the-data, and should only be invoked, if at all, where the
10
For example, ‘majority rule’, which says that alternative x is preferred to y just if it is
ranked higher by more individuals (or criteria), is one such algorithm. With three or more
alternatives, majority rule may lead the social preference to be intransitive, so does not even
meet the definition of a social choice rule.

data cannot decide between two theories. Similarly, in a discussion of

Kuhn’s five criteria of theory choice, McMullin (1993) argued that
‘accuracy ’ held a special role, for it is an end in itself while the
others are only valuable in so far as they are reliable indicators of
accuracy. So from an empiricist perspective, not all of Kuhn’s criteria
are equal.
Importantly, dictatorship of ‘accuracy ’ (or ‘fit-with-the-data’) need
not mean that the other criteria play no role at all in theory choice.
For recall the definition of a dictator: a criterion (or individual)
whose strict preference for x over y always leads x to be ranked

higher than y overall. By contrast, a strong dictator is a criterion (or
individual) whose preference for x over y, strict or weak, always be-
comes the overall preference. A strong dictatorship of ‘fit with the
data’ would be an extreme form of empiricism — that refused to
invoke extra-empirical criteria of theory choice even to break ties
between pairs of theories that fit the data equally well. But an ordinary
(not strong) dictatorship of ‘fit with the data’ could use criteria such
as simplicity to break ties, that is, to settle cases where the dictator is
indifferent. This is known as a ‘serial’ or ‘lexicographic’ dictatorship,
and represents a more moderate form of empiricism.
Accepting a serial dictatorship of ‘fit with the data’ is in principle a
way out of the impossibility result, since this theory choice rule does
satisfy conditions I, P, and U. However, even if one accepts the under-
lying empiricist motivation, there are two problems with this solution.
Firstly, to make it work, a complete lexicographic hierarchy of all the
criteria of theory choice must be established, that is, a specification of
the order in which they should be applied to break ties. If there are
only two criteria, for example fit-with-the-data and simplicity, then
this is not a problem, but if there are more than two there is a prob-
lem. For it is quite unclear how the hierarchy should be generated.
Should simplicity or scope be invoked first, when fit-with-the-data
cannot separate a pair of theories? Secondly and more importantly,
a dictatorship of ‘fit with the data’, even serial, seems unattractive
when we take account of the fact that our data invariably contain
‘noise’. If our data were noise-free, always preferring a theory that
fitted the data better would make sense. But with noisy data, perfect
fit is not always desirable, as emphasized in the model-selection lit-
erature (Forster and Sober 1994). This ‘problem of over-fitting’, as it is
known, constitutes a strong reason not to relax condition N in the
manner mooted above, even if we are empiricists.

Samir Okasha 97
Finally, two further possible escape routes, well-known among

social choice theorists, are worth briefly noting, though neither
holds much promise.11 The first is to modify the goal. Instead of
trying to rank-order the alternatives, as in Arrow’s formulation of
the problem, suppose we instead try to pick the best. More precisely,
we seek a ‘choice function’ which tells us, for any subset of the alter-
natives, which is (or are) the best. (In the scientific case this has some
plausibility, as the problem of theory choice is often formulated as the
problem of which theory to ‘accept’.) This is a weaker goal than
Arrow’s, since a social preference relation entails the existence of a

choice function but not vice versa, thus holding some promise of an
escape from the impossibility result. However, it turns out that if the
choice function is required to satisfy certain quite reasonable proper-
ties, then analogues of Arrow’s impossibility result re-emerge
(Austen-Smith and Banks 1999). This escape route is thus thought
unpromising by most social choice theorists, and seems equally un-
promising as applied to theory choice.
The second option is domain restriction, i.e. dropping condition U.
It is well known that with a restricted domain, there may exist social
choice rules that satisfy conditions P, I, and N. For certain applica-
tions of the social choice apparatus, ‘natural’ domain restrictions
suggest themselves, though not for others. As noted in section 4, in
the theory choice case, a natural domain restriction would apply if two
of the criteria of theory choice exhibit an intrinsic trade-off (or cor-
relation) — for example, if a gain in simplicity always means a loss of
accuracy. Then, certain profiles would be impossible, and could be
legitimately excluded from the domain of the theory choice rule.
However, that such trade-offs always exist does not seem very plaus-
ible; and anyway there is no guarantee that the resulting domain re-
striction would be of the right sort to alleviate the Arrovian
impossibility.
6. Sen’s ‘informational basis’ approach

I turn now to what is arguably the most attractive ‘escape route’ from
Arrow, namely Amartya Sen’s idea of using an ‘enriched informational
11
A third possible escape route is to relax condition I, a move defended by some authors,
e.g. Fleurbaey and Maniquet (2008), in the context of choosing allocations for an economy.
However it seems unlikely that the rationale for relaxing I, in such contexts, transposes to the
context of theory choice discussed here.

basis’ (Sen 1970, 1977, 1986).12 Sen observes that the information Arrow
uses as input to his social choice rule, namely a profile of individual
preference orders, is quite meagre. This is for two reasons. Firstly,
preference orders are ‘purely ordinal’ — they contain no information
about intensity of preference. If an individual prefers x to y to z, this
tells us nothing about whether their preference for x over y is greater
or less than their preference for y over z. Secondly, preference orders
do not permit interpersonal comparisons. From a profile of individual
preference orders, statements such as ‘in alternative x, individual 1 is
better off than individual 2’ cannot be deduced.

To remedy these problems, Sen suggests that we start not with
a profile of preference orders, but rather of utility functions, one for
each individual in society. An individual’s utility function assigns
a real number to each alternative, which reflects how much utility
that alternative would bring them. Let ui denote the utility function
of the i th individual; let <u1, … , un> denote a profile of utility func-
tions. An individual’s utility function is required to represent their
preference order, in the sense that xRiy iff ui(x) ! ui(y), for all alter-
natives x and y. (Recall that ‘xRiy ’ means that individual i weakly
prefers x to y.) Note that if ui represents Ri, then any increasing trans-
formation of ui will also represent Ri. Thus there is a many–one rela-
tion between utility functions and the preference orders that they
represent.
Next, Sen introduces the concept of a social welfare functional
(SWFL). This is a function that takes as input a profile of utility
functions, and yields as output a social ranking of the alternatives.
An SWFL is analogous to an Arrovian social choice rule, in that both
yield the same output; however, the former takes a profile of utility
functions, rather than preference orders, as input. Potentially, this
allows more information to be taken into account.
Analogues of Arrow’s four conditions can now be imposed on the
SWFL. The analogue of U says that the domain of the SWFL is the
set of all possible profiles of utility functions, that is, individuals
can have whatever utility functions they please. The analogue of P
says that if everyone gets more utility in alternative x than in y,
then x is socially preferred to y. The analogue of N says that there
can be no individual such that whenever they get more utility from x
than y, then x is socially preferred to y. The analogue of I, known as
12
Sen’s approach has been further developed by numerous authors; for good overviews see
Gaertner 2006, Roemer 1997, and Bossert and Weymark 2004.

Samir Okasha 99
‘independence of irrelevant utilities’, says that the social preference

between x and y must depend only on individuals’ utilities in x and
y. These conditions on the SWFL will be denoted U9, P9, I9, and N9;
they are motivated by arguments similar to those that motivate the
Arrovian originals.
One might think that an analogue of Arrow’s impossibility result
will now apply, that is, that no SWFL can satisfy conditions U9, P9, I9,
and N9. However, this is not correct. Arrow’s impossibility result can
be derived in Sen’s framework, but it requires an additional condition,
capturing the fact that Arrow uses purely ordinal, non-interpersonally

comparable information. To see how this informational assumption
can be captured, consider a profile of utility functions <u1, … , un>.
Now suppose each of the n individuals applies an increasing trans-
formation to their utility function, yielding a new profile <v1, … , vn>.
(Different individuals may apply different transformations.) On
Arrow’s assumption, the two profiles contain exactly the same infor-
mation — since the transformed utility functions represent the very
same preferences. So Arrow will argue that the SWFL should yield the
same social ranking when applied to the two profiles. This condition is
called ‘invariance with respect to ordinal, non-comparable informa-
tion’ or ONC. Arrow’s theorem can now be stated in Sen’s framework:
for three or more social alternatives, no SWFL can satisfy conditions
ONC, U9, P9, I9, and N9.
If the ONC condition is imposed on the SWFL, this implies that
interpersonal comparison of utility is deemed impossible (or mean-
ingless). To see why, suppose that in profile <u1, … , un>, individual
1 gets more utility than individual 2 in a given alternative x, that is,
u1(x) > u2(x). But this inequality is not necessarily preserved if the
individuals apply different positive transformations to their utility
functions. So in the transformed profile <v1, … , vn>, it need not be
true that v1(x) > v2(x). Therefore, if the two profiles are treated as
informationally equivalent, as the ONC condition demands, it follows
that interpersonal comparisons cannot be made.
The natural next step is to ask what happens if the ONC condition
is relaxed. There are two ways it can be relaxed: (i) drop the assump-
tion that utility is purely ordinal; (ii) permit interpersonal compari-
sons. To effect (i), we restrict the transformations that can be applied
to a given utility function; to effect (ii), we cease to allow individuals
to choose their own transformations independently of others. Let us
take (i) first. Instead of ordinal utility, we might hold that utility is
measured on a cardinal scale, so only positive linear transformations,

of the form vi = aui + b, a > 0, are held to preserve information.13

In effect, this means that an individual’s utility function contains in-
formation about the intensity of their preferences, so utility differences
become meaningful. Alternatively, we might hold that utility is mea-
sured on a ratio scale, so only transformations of the form vi = aui,
a > 0, are held to preserve information. This means that the utility
scale has a natural zero point, so utility ratios become meaningful.14
Finally, we might hold that utility is measured on an absolute scale,
that is, only the identity transformation preserves information. This
means that actual utility numbers are meaningful.

Once a scale for utility has been chosen — ordinal, cardinal, ratio,
or absolute — a decision about interpersonal comparability is neces-
sary. If utility is non-comparable, then each individual can apply a
transformation (from the permissible class) independently of others.
If utility is fully comparable, then each individual must apply the same
transformation. Depending on the utility scale, a form of partial com-
parability may also be possible. With cardinal utility, if utility is unit
comparable, then individuals’ positive linear transformations must all
have the same slope, but can have different intercepts.
Numerous alternatives to Arrow’s ONC condition are now possible.
They include: cardinal-scale utility with no comparability (CNC);
cardinal-scale utility with full comparability (CFC); ratio-scale utility
with full comparability (RFC); ratio-scale utility with no comparabil-
ity (RNC); and absolute-scale utility with full comparability (AFC).
In effect, each of these conditions partitions the set of all profiles of
utility functions into equivalence classes of ‘informationally equiva-
lent’ profiles, and requires that the SWFL yield the same social ranking
for all the profiles in a given equivalence class. ONC is the strongest
condition — for the classes of profiles that it treats as informationally
equivalent are very large, and thus the restriction on the SWFL con-
siderable. By contrast, AFC is the weakest condition — it places each
profile into a singleton class of its own, which implies no restriction
on the SWFL. This illustrates a general moral: the richer the informa-
tional basis, that is, the finer the partition of the profiles into equiva-
lence classes, the weaker the resulting condition on the SWFL.
Sen now asks: What happens if we retain the four Arrovian condi-
tions U9, P9, I9, and N9, but replace ONC with a weaker condition?
Can the impossibility result be avoided? The answer is that
13
Temperature in Celsius or Fahrenheit is measured on a cardinal, or interval, scale.
14
Length in centimetres is an example of a quantity measured on a ratio scale.

Samir Okasha 101
impossibility can be avoided, but only if some interpersonal compari-

son is allowed. Replacing ONC with CNC — that is, moving from
ordinal to cardinal utility — is no help on its own. The same is true
of moving to ratio-scaled utility (RNC). However, if ONC is replaced
with CFC, RFC, or AFC, then Arrovian impossibility is avoided. There
do exist social welfare functionals that satisfy Arrow’s four conditions,
plus one of these alternatives to ONC.
(The case of ratio-scale non-comparability (RNC) merits further
discussion, for a reason that will become clear. Although replacing
ONC with RNC does not avoid Arrovian impossibility, it does if all

utilities are required to be non-negative (Tsui and Weymark 1997). If all
utilities are non-negative, there do exist social welfare functionals that
satisfy RNC and Arrow’s four conditions. Also, note that RNC, despite
its name, does permit a limited sort of interpersonal utility compari-
son.15 With RNC, percentage increases in utility can be meaningfully
compared, that is, statements such as ‘in moving from alternative x to
y, individual 1’s percentage gain is greater than individual 2’s’, are
meaningful (Fishburn 1987). It is easy to verify that the truth-value
of this statement will be unaltered if the two individuals apply differ-
ent ratio-scale transforms to their utility functions.)
Sen’s analysis raises two issues. First, how can interpersonal com-
parisons of utility be made? Second, once such comparisons are
allowed, how large is the class of SWFLs that satisfy the Arrow con-
ditions? Can further conditions be found that narrow down the per-
missible SWFLs to a single one? There is an extensive literature on
both these points, but space does not permit them to be explored
here.16 For the moment, the point to note is just this: Sen’s work
demonstrates clearly that Arrow’s impossibility result is in large part
a consequence of the impoverished information he feeds into his
social choice rule. Enriching the informational basis, while retaining
Arrow’s four conditions — now understood as conditions on the
social welfare functional, rather than the social choice rule — is suffi-
cient to avoid the impossibility.17 In short, given enough information,
reasonable social choices can be made.
15
For this reason, some authors prefer the label RSM (ratio-scale measurability) for what
I am calling RNC.
16
On the first issue, see the papers in Elster and Roemer 1991; on the latter, see Gaertner
2006, Roemer 1997, or Bossert and Weymark 2004.
17
In a way, Sen’s approach involves a rejection of Arrow’s condition I (independence of
irrelevant alternatives). For Arrow’s condition I is equivalent to the conjunction of ONC and

7. Theory choice: the informational basis

Let us return to theory choice and apply the morals of the previous
section. Recall that we defined a theory choice rule, by direct analogy
with Arrow’s social choice rule, as a function that takes as input a
profile of weak orders, one for each criterion of theory choice, and
outputs an ‘overall ranking’ of the alternative theories. Just as Sen
replaced Arrow’s social choice rule with a social welfare functional,
so we need to replace our theory choice rule with a ‘theory choice
functional’. So instead of starting with a profile of ‘preference orders’,
one for each criterion of theory choice, we start with a profile of

‘utility functions’, that is, real-valued representations of those
orders. In principle, this allows an enrichment of the informational
basis.18
The natural next question is: Which profiles should be treated as
informationally equivalent, that is, which invariance condition should
be imposed on the theory choice functional? To address this question,
we need to consider both measurement scales and ‘inter-criterion’
comparability. Let us take them in turn. For some criteria of theory
choice an ordinal scale might be appropriate. Kuhn’s criterion of
‘fruitfulness’ is an example. Conceivably, one could order a set of
theories by how fruitful they are, but it is hard to believe that differ-
ences in fruitfulness can be compared; a statement such as ‘the differ-
ence in fruitfulness between T1 and T2 exceeds the difference between
T2 and T3’ hardly seems meaningful. If this is right, then the
real-valued ‘utility ’ function that represents the fruitfulness preference
order is merely ordinal — any increasing transformation can be
applied to it without loss of information.
However, for other criteria of theory choice, we can go beyond
ordinal measurement, at least in certain contexts. Take, for example,
fit-with-the-data (or ‘accuracy ’), and suppose that the context is linear
regression analysis. The usual measure of how well a hypothesis fits the
data in linear regression is its ‘sum of squares’ (SOS) score.19 The
appropriate type of measurement scale for SOS scores depends on
condition I9 (independence of irrelevant utilities). Retaining I9 while rejecting ONC thus

abandons the letter of Arrow’s original independence condition, while retaining its spirit.
18
This application of Sen’s framework to the problem of theory choice illustrates a point
made by Kelsey (1987), namely that the functions used in an SWFL need not necessarily be
interpreted as utility functions.
19
The SOS of a hypothesis is the sum of the squared distance of each data point from the
hypothesis’s prediction.

Samir Okasha 103
the dependent variable in the regression model. If, for example, that
variable is length, which is a ratio-scale measurable quantity, then
the SOS scores will also be ratio-scale measurable. Therefore, the
real-valued ‘utility ’ function that represents the ‘fit-with-the-data’
preference order will be ratio-scaled, thus multiplication by a posi-
tive constant is the only information-preserving transformation.
Statements such as ‘T1 fits the data three times as well as T2’ will be
meaningful.
To take another example of how we can often go beyond ordinal
information, consider simplicity. In certain contexts, such as statistical

model selection, the simplicity of a hypothesis is taken to be the
number of free parameters it contains. Thus for example, the hypoth-
esis ‘y = ax + b’ is simpler than ‘y = ax 2 + bx + c’ because the former
contains two free parameters, the latter three. So in this case, simpli-
city is measured on an absolute scale — the actual numbers are mean-
ingful, so only the identity transformation preserves information.
Similarly, in a Bayesian context, a prior probability distribution over
a set of hypotheses is a case of absolute measurability — the actual
numbers assigned are meaningful. So in both these cases, we have
much more than ordinal information.
This suggests that the question of what measurement scales are
appropriate for criteria of theory choice does not have a simple
answer. Different scales may be appropriate for different criteria,
and may depend on the inferential techniques that we are using. It
may be that for the ‘large scale’ theory choices that Kuhn was inter-
ested in, ordinal comparisons are all that can be achieved. But it seems
clear that in other more humdrum cases, particularly where the prob-
lem may be formulated statistically, we may have much more than
ordinal information at our disposal.
Finally, note that the situation for theory choice is more compli-
cated than for social choice. In social choice, one normally assumes a
single type of measurement scale for all utility functions. It would
make little sense to suggest that individual 1’s utility function was
ordinal, individual 2’s cardinal. But the analogous situation for
theory choice makes good sense. It might well be that fruitfulness,
for example, is merely ordinal but that fit-with-the-data is ratio-scale
measurable.
What about inter-criterion comparisons, the analogue of interper-
sonal comparisons? One might think that such comparisons are un-
likely. Take, for example, the statement: ‘the difference in simplicity
between T1 and T2 exceeds the difference in accuracy between T3 and

T4’. It is hard to see what the basis for such a judgement might be. It is
harder still to see how comparisons of levels, rather than differences,
could be made — this would permit statements such as ‘the accuracy of
T1 is less than the simplicity of T2’, which sound even odder. Since
inter-criterion comparability is needed to avoid the impossibility result,
as we know, the prospects for escaping the Arrovian predicament by
enriching the informational basis of theory choice may seem dim.
However, this is overly pessimistic for two reasons. Firstly, note that
if all criteria are absolutely measurable, then interpersonal compar-
ability follows immediately. If the ‘utility ’ functions that represent the

simplicity and accuracy orderings cannot be transformed without loss
of information, then statements such as ‘the accuracy of T1 is less than
the simplicity of T2’ automatically become meaningful. (Crucially,
‘meaningful’ here has a technical sense, i.e. invariance under the per-
missible transformations; it does not mean that there would be any
particular purpose in uttering the statement in question.)20 Since, as
argued above, absolute measurability may be appropriate for some
criteria of theory choice in some contexts, inter-criterion comparabil-
ity should not be dismissed out of hand.
Secondly, recall the discussion of ratio-scale measurability in section
6. If the criteria of theory choice are each measured on their own
ratio-scale (i.e. RNC), then this: (i) permits a limited form of
inter-criterion comparability, and (ii) avoids Arrovian impossibility
so long as all ‘utilities’ are non-negative. Ratio-scale measurability is
fairly plausible in certain inferential contexts. Consider ‘scope’, for
example. If differences in scope can be compared, and if in addition
there is a natural zero point — that is, it makes sense to talk about a
theory with zero scope — then scope is ratio-scale measurable. This
does not seem altogether implausible, for some criteria in some infer-
ential contexts. If both scope and accuracy (say) are ratio-scale meas-
urable, each with their own scale, then this permits a limited form of
inter-criterion comparability: percentage increases in scope may be
compared with percentage increases in accuracy. (So statements
such as ‘T1 has 10% less scope than T2, but is 15% more accurate’
can be made.) As regards point (ii), the restriction to non-negative
‘utilities’ seems unproblematic; if ‘scope’ has a natural zero point, why
demand that the theory choice functional be able to deal with profiles
20
It is also important to see that inter-criterion comparability does not require that the two
criteria be measured in the same units, for the comparison in question is a comparison of real
numbers, not of the quantities that they represent.

Samir Okasha 105
in which some theories (per impossible) have negative scope? So there

is a potential escape route from Arrow here too.
To sum up, Sen’s work, transposed to the theory choice case, tells us
that there do exist theory choice functionals that satisfy Arrow’s four
conditions, so long as the ONC condition is replaced in favour of
one that permits inter-criterion comparison. This prompts the ques-
tion of what replacement of ONC (if any) is appropriate, that is, which
profiles of ‘utility ’ functions should be treated as informationally
equivalent. There is no simple answer to this question. However, in
some cases absolute measurement will be appropriate, implying that

ONC should be replaced with AFC; this permits the impossibility
result to be avoided. In other cases ratio-scale measurement will be
appropriate, which also permits the impossibility result to be avoided.
The general moral is that enriching the informational basis of theory
choice does permit an escape from Arrow; though which enrichments
are defensible must be answered on a case-by-case basis.
Where does this leave us vis-à-vis Kuhn’s ‘no algorithm’ thesis?
If we can escape the Arrovian predicament by enlarging the infor-
mational basis, as described above, we will end up with many the-
ory choice functionals that meet our reasonableness constraints. For
replacing ONC with an alternative condition (such as AFC), while
retaining Arrow’s four conditions, does not narrow down the class of
permissible theory choice functionals to a single one. So we escape
Arrow’s predicament only to enter Kuhn’s: many acceptable algo-
rithms, and no way to select between them. To escape both predica-
ments, we need reasonableness conditions that are satisfied by exactly
one algorithm. In the social choice literature, researchers have mana-
ged to identify conditions that uniquely pick out particular social
welfare functionals, such as the utilitarian SWFL, Rawlsian maximin,
and others; but it is doubtful whether the analogues of these condi-
tions, transposed to the theory choice case, would be defensible.
(By contrast, the analogues of Arrow’s conditions are certainly defens-
ible.) Therefore, in the theory choice case, escaping Arrovian impos-
sibility by enriching the informational basis seems to lead us straight
to Kuhn’s ‘no algorithm’ thesis.
8. Illustrations: Bayesianism and statistical model selection

The previous section’s main conclusion — that Sen’s escape route
from Arrow does apply to theory choice — can be illustrated by

considering the orthodox Bayesian approach to scientific inference.

Suppose we have a body of evidence E, and five rival hypotheses
{T1, … , T5} that are pair-wise exclusive. On the Bayesian view, we
use two criteria to choose between the hypotheses: prior probability
P(Ti), and likelihood, P(E/Ti). (Likelihood can be thought of as a
measure of the ‘fit’ between evidence and hypothesis, prior probability
a measure of the antecedent plausibility of a hypothesis before the
evidence.) Clearly, there are many possible ways of combining these
two criteria into a decision rule; but Bayesians argue that the right way
to do it is to multiply the prior by the likelihood, that is, to consider

the quantity [P(Ti) " P(E/Ti)]. The theory with the highest value of
this quantity is the most deserving of our credence, according to
Bayesians; and, more generally, this quantity can be used to generate
an overall ranking of the theories, from best to worst.
Bayesians have some sophisticated arguments for why this is the
right way to combine the two criteria, but they need not detain us.
For the moment, we want to relate Bayesianism to our foregoing
discussion. To do this, simply think of P(Ti) and P(E/Ti) as ‘utility ’
functions, both of which assign a real number to each of the five
theories. The ordered pair <P(Ti), P(E/Ti)> is then a profile, corres-
ponding to a profile of utility functions in a two-person society, in
Sen’s framework. Now consider the function which maps the set of
profiles onto the ranking generated by the quantity [P(Ti) " P(E/Ti)].
This corresponds to a social welfare functional in Sen’s framework, or
what we called a ‘theory choice functional’ in section 7. The functional
takes as input the prior probability and likelihood of each theory, and
yields as output a ranking of the theories, from best to worst. Let us
call this the ‘Bayesian theory choice functional’, or BCF.
Now let us ask: Does the BCF satisfy the four Arrovian conditions?
Consider firstly condition P9, weak Pareto. It easy to see that P9 is
satisfied: if theory T1 has a higher prior and a higher likelihood than
theory T2, that is, P(T1) > P(T2) and P(E/T1) > P(E/T2), then T1 will
obviously be ranked higher than T2 by the BCF. Condition I9, the
independence of irrelevant ‘utilities’ condition, is also satisfied:
whether T1 or T2 is ranked higher by the BCF is entirely determined
by the priors and likelihoods of those two theories; no other informa-
tion is relevant. Finally, condition N9, non-dictatorship, is also satis-
fied. Neither criterion (prior or likelihood) is able to dictate over the
other — it is not true that if T1 has a higher prior than T2 then it must
be ranked higher, and similarly for likelihood.

Samir Okasha 107
What about condition U9, unrestricted domain? This says that the
domain of the theory choice functional must be the set of all possible
profiles, that is, pairs of real-valued functions. Clearly the BCF does
not satisfy this condition, for both of the functions that we feed into it,
P(Ti) and P(E/Ti), can only take P on values in the unit interval [0,1];
moreover, it is required that P(Ti) # 1. Thus there are two restric-
tions on the permissible values of the functions we feed into the
Bayesian theory choice functional. So condition U9 is not satisfied,
whereas conditions P9, I9, and N9 are.
Mindful of Arrow’s theorem, one might think that it is because the

BCF has a restricted domain, and so violates condition U9, that it can
satisfy P9, I9, and N9. But this is not correct. Recall that within Sen’s
framework, the derivation of Arrow’s result requires the ONC condi-
tion, in addition to conditions U9, P9, I9, and N9. The ONC condition
says that any two profiles are informationally equivalent, hence should
be mapped to the same ranking, if one is derivable from the other by
applying increasing transformations to the functions in the profile.
But the BCF does not satisfy this condition; on the contrary, it is
quite possible to have two profiles <P(Ti), P(E/Ti)> and <Q(Ti),
Q(E/Ti)>, where the two prior functions P(Ti) and Q(Ti) rank
the theories identically, and the two likelihood functions P(E/Ti)
and Q(E/Ti) also rank them identically, and yet the overall rankings
generated by the BCF are different in the two cases.
This prompts the question: Which measurability/comparability
assumption is appropriate for the Bayesian theory choice functional?
Since probabilities are measured on an absolute scale, the answer
is clear: absolute full comparability (AFC). Given a profile <P(Ti),
P(E/Ti)>, applying any transformation to it other than the identity
transformation will alter its informational content — for the actual
probability numbers are meaningful.21 So the prior probability func-
tion contains much more than the merely ordinal information that
gives rise to Arrovian impossibility; the same is true of the likelihood
function.
Therefore, the Bayesian theory choice functional violates two of
the conditions that are required to generate Arrow’s impossibility
result in Sen’s framework — U9 and ONC. The fact that the BCF
has a restricted domain is not crucial — what matters is the fact that
21
Probabilities are measured on an absolute scale modulo the convention that the sure event
has probability 1. If this convention were relaxed, i.e. if we chose some other positive number
for the probability of the sure event, then probability would become ratio-scale measurable.

it uses more than ordinal non-comparable information. For even with

the domain restriction appropriate to the BCF, if the ONC condition
were imposed, then the P9, I9, and N9 conditions would be jointly
unsatisfiable (see Appendix for proof.) The fact that the BCF satisfies
the latter three conditions is therefore attributable to the richness of
the information fed into it, rather than to its restricted domain.
This illustrates how Sen’s escape route from Arrow does apply to
theory choice. The Bayesian theory choice functional constitutes a
kind of algorithm for theory choice, and does satisfy the Arrovian
conditions (other than U9); this is possible because the appropriate

measurability/comparability assumption is AFC, rather than ONC.
The BCF is by no means the only theory choice functional that satisfies
AFC and the Arrovian conditions; so there is potential for a Kuhnian
‘no unique algorithm’ argument. However, such an argument would
have to counter the Bayesians’ argument for why their theory choice
functional is the ‘correct’ one. This important issue cannot be ad-
dressed here. My aim has only been to illustrate how in theory
choice, enriching the informational basis permits an escape from
Arrovian impossibility, just as it does in social choice.
A second illustration of this point is provided by a quite different
approach to scientific inference, namely statistical model selection.
This approach is a sophisticated variant of what philosophers call
the ‘curve-fitting’ problem, or inferring the functional relation be-
tween two variables from finite data. However, unlike in more sim-
plified discussions of curve-fitting, it is assumed that the data are
‘noisy ’, that is, the data points are affected by measurement error.
The aim is to choose between alternative hypotheses about the func-
tional relation between two variables x and y, from the noisy data.
Two criteria are used: simplicity and fit-with-the-data. The two will
often exhibit a trade-off: improving fit means sacrificing simplicity.
Clearly, we have here the ingredients for a Kuhnian ‘no unique algo-
rithm’ claim — there are many conceivable ways of combining fit and
simplicity into a single decision rule.
In a typical model-selection problem, we start with a number of
families of hypotheses. For example, one family consists of all hypo-
theses of the form y = a + bx, a, b 2 R; let us call this family LIN.
Another family, PAR, consists of all hypotheses of the form
y = a + bx + cx 2; and a third, EXP, consists of all hypotheses of the
form y = ax. The real numbers a, b, and c are called free parameters;
and the simplicity of a hypothesis is defined as the number of free
parameters in its family. Thus each hypothesis in LIN is simpler than

Samir Okasha 109
each in PAR. Next, one finds the best-fitting hypothesis in each family,
denoted L(LIN), L(PAR), and L(EXP) respectively; the criterion of
best fit is highest likelihood, where ‘likelihood’ has its customary stat-
istical meaning (roughly, the probability of observing the actual data
if the hypothesis were true.)
In the statistical literature, there exist various suggestions for how to
combine simplicity and fit into a decision rule (Forster 2001); such a
rule would allow us to choose between the three hypotheses L(LIN),
L(PAR), and L(EXP), in the above example. One of the best-known is
the Akaike criterion, which says that we should choose the hypothesis

with the highest Akaike score; the Akaike score of a hypothesis H
is defined as [log-likelihood H – k], where k is the number of free
parameters of the family to which the hypothesis belongs. Therefore,
the better a hypothesis fits the data the higher its Akaike score;
however, there is a penalty for complexity, that is, for having lots
of free parameters. Thus Akaike’s criterion combines simplicity and
fit-with-the-data into a single algorithm for hypothesis choice, which
ranks the hypotheses from best to worst. Obviously, there are many
other conceivable algorithms, but there exist sophisticated arguments
for why Akaike’s is the ‘correct’ one.
To relate statistical model selection to social choice, think of
log-likelihood Hi and ki as ‘utility functions’, each of which assigns
a number to each hypothesis reflecting, respectively, its fit and its
simplicity. Then consider the set of all ordered pairs (profiles) of
the form <log-likelihood Hi, ki> — these are all the possible combin-
ations of fit with simplicity. Then consider the function from this set
to the ranking of hypotheses generated by the Akaike score. Formally,
this function corresponds to the social choice functional of section 6
and the theory choice functional of section 7. Let us call it the ‘Akaike
choice functional’.
Now we can ask: Does the Akaike choice functional satisfy the
Arrovian conditions? It is easy to see that conditions P9, I9, and N9
are all satisfied, by an argument parallel to the one given above for the
Bayesian choice functional. But condition U9 (unrestricted domain) is
not satisfied — for the log-likelihood function only takes negative
values and the free parameter function only takes positive integer
values, which implies a domain restriction. What about the ONC
(ordinal non-comparability) assumption? This is not satisfied either.
It is not the case that if one profile can be got from another by
applying increasing transformations to the two ‘utility ’ functions,
that the Akaike choice functional will necessarily map them to the

same ranking. Again, the appropriate assumption, in lieu of ONC, is

absolute full comparability (AFC); since the actual numbers assigned
by the two functions are meaningful, any transformations will change
their informational content. As with the Bayesian choice functional, it
is because the Akaike choice functional does not satisfy ONC, rather
than because of its restricted domain, that it is able to satisfy condi-
tions P9, I9, and N9.22
Again, the point of this is not to defend the Akaike criterion in
particular, but rather to illustrate how Sen’s moral — that enriching
the informational basis can avoid Arrovian impossibility — applies to

theory choice. The information we feed as input into the Akaike
choice functional is far more than merely ordinal, which explains
why it satisfies the Arrovian conditions (other than U9). Of course
there are many alternative functionals, besides Akaike’s, that will also
satisfy those conditions, so again, escaping the Arrovian predicament
may land us in the Kuhnian one. However, there are also arguments
for why the Akaike criterion is the uniquely ‘correct’ one; so it may be
that both predicaments can be avoided. A proper assessment of this
issue cannot be undertaken here.
9. Conclusion
Although Kuhn’s ‘no algorithm’ thesis is quite widely accepted in
philosophy of science, there have been few attempts to subject it to
serious scrutiny. To remedy this situation, I have used the machinery
of social choice theory, and tried to relate Kuhn’s thesis to Arrow’s
famous impossibility theorem. Though superficially similar to Kuhn’s,
Arrow’s conclusion that there is ‘no algorithm’ for social choice is in
fact quite different. For Kuhn’s claim is that there are many algo-
rithms, all equally acceptable, while Arrow’s claim is that no algorithm
meets minimum standards of acceptability.
By identifying Kuhn’s five criteria with Arrow’s individuals, the
theory choice problem was seen to have the same structure as a stand-
ard social choice problem. Moreover, Arrow’s four conditions seem as
defensible for theory choice as they are for social choice, which raises
the spectre of an Arrovian impossibility result for theory choice.
22
An argument parallel to the one given in the Appendix shows that with the domain
restriction appropriate to the Akaike choice functional, conditions P9, I9, and N9 are jointly
unsatisfiable if ONC is assumed. Thus it is the violation of ONC, not the domain restriction,
that permits the satisfaction of conditions P9, I9, and N9.

Samir Okasha 111
Such a result would constitute a refutation of Kuhn’s thesis, but would

also pose a threat for the rationality of science; a threat that if anything
is more worrying than that posed by Kuhn.
A number of possible ‘ways out’ of the impossibility, while re-
maining within Arrow’s framework, were canvassed; these included
confining attention to binary theory choices and accepting a serial
dictatorship of ‘accuracy ’. More promising was Sen’s idea of
moving to a different framework by enriching the informational
basis, that is, going beyond the ordinal, non-comparable information
that Arrow starts with. Applying Sen’s idea to theory choice raised

difficult questions about the appropriate measurability/comparability
assumptions; however, we showed by example that two well-known
approaches to theory choice, Bayesianism and statistical model selec-
tion, avoid Arrovian impossibility precisely by incorporating more
than ordinal, non-comparable information.
Finally, what then of Kuhn’s ‘no algorithm’ thesis? Is it correct?
No simple answer to this question emerges from the foregoing ana-
lysis. It may be that we can avoid Arrovian impossibility only by
opening the door to many different algorithms, thus vindicating
Kuhn; but we should not rule out the possibility of finding additional
rationality constraints which considerably narrow down the acceptable
ones, possibly even to uniqueness. Although my analysis does not
provide a definitive resolution, I hope to have identified the sorts of
consideration that are relevant to determining whether Kuhn’s thesis
is correct.23
Appendix
Consider a finite set X of pair-wise exclusive theories {T1, … , Tn}. We
assume X is a partition of logical space.24 Let Y be the set of all
orderings of X.
There are two ‘individuals’, each with a real-valued ‘utility ’ func-
tion over X, denoted u1 and u2 respectively. A profile (ordered pair) of
utility functions, one for each individual, is denoted <u1, u2>.
23
Thanks to audiences at Bristol, Cambridge, Konstanz, Madrid, and Cardiff, where ver-
sions of this paper were presented; to Elliott Sober, Marcel Weber, Hannes Leitgeb, Armin
Schulz, and Kit Patrick for comments and discussion; and in particular to an anonymous
referee for Mind for extremely detailed comments. This work was supported by AHRC grant
AH/F017502/1.
24
Taking X to be a partition makes the second half of the proof easier, as it allows us to
P P
assume that u1(Ti) = 1 as opposed to merely u1(Ti) # 1, but could easily be relaxed.

Let D be the set of all possible profiles.

Let R $ D be the set of all profiles satisfying the condition: for all
individuals i, and all x 2 X, 0 # ui(x) # 1, i.e. both utility functions can
only take values in [0, 1].
Let TP$ R be the set of all profiles satisfying the additional con-
dition: u1(Ti) = 1, i.e. the utilities that individual 1 assigns to the
theories sum to 1.
So we have T $ R $ D (see diagram).
D

R
By Arrow’s theorem (in Sen’s framework), there exists no function

f : D=Y satisfying the conditions P9, I9, N9, and ONC.
We wish to show that there is no function f 9 : T=Y satisfying the
conditions P9, I9, N9, and ONC.
We show firstly that there is no function g : R=Y satisfying the
1. Suppose for reductio that there exists a function g : R=Y satisfying
Let S $ R be the set of all profiles satisfying the condition: for all
individuals i, and all x 2 X, 0 < ui(x) < 1.
Since g satisfies P9, I9, N9, and ONC, and since g has domain R and
S $ R, then there exists a function g 9: S=Y satisfying the conditions
P9, I9, N9, and ONC.
Let t : R=(0,1) be any increasing function mapping the real number
line onto the open unit interval, e.g. t(x) = [x–2]–1 if x < 0, t(x) = [x+1]/
[x+2] if x ! 0.
Let h be the result of applying the transformation ui(x)=t[ui(x)]
to each individual’s utility function in each profile <u1, u2> in
D. Therefore h: D=S maps each profile in D onto a profile in S.
The function g 9 o h (‘g 9 after h’) then has domain D and range Y.
Since g 9 satisfies conditions P9, I9, N9, and ONC, and since t is
increasing, it follows that g 9 o h satisfies conditions P9, I9, N9, and

Samir Okasha 113
ONC too. But by Arrow’s theorem no function with domain D satis-

fies these conditions.
Therefore there is no function g: R=Y satisfying conditions P9, I9,
N9, and ONC.
2. Suppose for reductio that there exists a function f 9: T=Y satisfy-
ing the conditions P9, I9, N9, and ONC.
Let r: [0,1] = [0,1] be a function mapping individual 1’s utility func-
tion u1(Ti) onto [u1(Ti) / [(u1(T1) + u1(T2) +, … , + u1(Tn)]]. (Applying
r to individualP1’s utility function has the effect of normalizing it,
to ensure that u1(Ti) = 1.)

Let h9 be the result of applying the transformation <u1, u2> =
<r(u1), u2> to each profile in R. Therefore h9: R=T maps each profile
in R onto a profile in T.
The function f 9 o h9 (‘f 9 after h9’) then has domain R and range Y.
Since f 9 satisfies conditions P9, N9, and ONC, and since r is
increasing, it follows that f 9 o h9 satisfies conditions P9, N9, and
ONC too.
To show that f 9 o h9 satisfies condition I9, consider two profiles
<u1, u2> and <v1, v2> in R that coincide over the theories T1 and T2,
i.e. u1(T1) = v1(T1), u2(T1) = v2(T1), and u1(T2) = v1(T2), u2(T2) = v2(T2).
[Note: it does not follow that <r(u1), u2> and <r(v1), v2> must also
coincide over T1 and T2.]
Choose any profile <r9(u1), u2> in T such that:
(i) r9(u1) and r(u1) order the theories in the same way; and
(ii) r9(u1)(T1) = r(v1)(T1), and r9(u1)(T2) = r(v1)(T2).
(There must be such a profile, since r(u1) and r(v1) order T1 and T2 in
the same way; for all other theories Ti, choose r9(u1)(Ti) so as to
produce the same order as r(u1).)
Now, the profiles <r9(u1), u2> and <r(v1), v2> coincide over T1
and T2, so f 9 must map them onto orderings of the set X in which
T1 and T2 are ranked identically, by condition I9.
Since r9(u1) and r(u1) order the theories in the same way, f 9 must
map <r9(u1), u2> and <r(u1), u2> onto the same ordering of the set
X, by condition ONC.
Therefore, f 9 must map <r(u1), u2> and <r(v1), v2> onto orderings
of the set X in which T1 and T2 are ranked identically.
Therefore, f 9 o h9 maps <u1, u2> and <v1, v2> onto orderings of
the set X in which T1 and T2 are ranked identically, i.e. f 9 o h9 satisfies
condition I9.

So f 9 o h9 has domain R and satisfies conditions P9, I’, N’, and ONC.
However, by part 1 above, no function with domain R satisfies those
conditions.
Therefore, there is no function f 9: T=Y satisfying the conditions P9,
I9, N9, and ONC.
References
Arrow, Kenneth 1951: Social Choice and Individual Values. New York:

John Wiley.
Arrow, K. and M. Instiligator (eds) 1986: Handbook of Mathematical
Economics. North-Holland: Elsevier.
Austen-Smith, David and Jeffrey S. Banks 1999: Positive Political
Theory 1: Collective Preference. Ann Arbor: University of
Michigan Press.
Barbera, S., P. Hammond, and C. Seidl (eds) 2004: Handbook of
Utility Theory, volume 2. Dordrecht: Kluwer.
Bossert, Walter and John A. Weymark 2004: ‘Utility in Social
Choice’. In Barbera, Hammond, and Seidl 2004, pp. 1099–177.
Earman, John 1992: Bayes or Bust? Cambridge, MA: MIT Press.
Elster, John and John E. Roemer (eds) 1991: Interpersonal Com-
parisons of Well-Being. Cambridge: Cambridge University Press.
Fishburn, Peter C. 1987: Interprofile Conditions and Impossibility.
London: Routledge.
Fleurbaey, Marc and Francois Maniquet 2008: ‘Utilitarianism versus
Fairness in Welfare Economics’. In Fleurbaey, Salles, and
Weymark 2008, pp. 263–80.
Fleurbaey, M., M. Salles, and J. A. Weymark (eds) 2008: Justice,
Political Liberalism and Utilitarianism. Cambridge: Cambridge
University Press.
Forster, Malcolm R. 2001: ‘The New Science of Simplicity ’. In
Zellner, Keuzenkamp, and McAleer 2001, pp. 83–119.
Forster, Malcolm R. and Elliott Sober 1994: ‘How to Tell When
Simpler, More Unified or Less Ad Hoc Theories Will Provide
More Accurate Predictions’. British Journal for the Philosophy of
Science, 45, pp. 1–35.
Gaertner, Wulf 2006: A Primer in Social Choice Theory. Oxford:
Oxford University Press.
Horwich, P. (ed.) 1993: World Changes: Thomas Kuhn and the Nature
of Science. Cambridge, MA: MIT Press.

Samir Okasha 115
Howson, Colin and Peter Urbach 1992: Scientific Reasoning: The

Bayesian Approach. La Salle: Open Court.
Kelsey, David 1987: ‘The Role of Information in Social Welfare
Judgements’. Oxford Economic Papers, 39, pp. 301–17.
Kuhn, Thomas 1969: The Structure of Scientific Revolutions. Chicago:
University of Chicago Press.
—— 1977a: ‘Objectivity, Value Judgment, and Theory Choice’, in his
1977b, pp. 320–39.
—— 1977b: The Essential Tension. Chicago: University of Chicago
Press.

Lipton, Peter 1990: Inference to the Best Explanation. London:
Routledge.
McMullin, Ernan 1993: ‘Rationality and Paradigm Change in
Science’. In Horwich 1993, pp. 55–78.
Newton-Smith, William H. 1981: The Rationality of Science. London:
Routledge.
Roemer, John 1997: Theories of Distributive Justice. Oxford: Oxford
University Press.
Sen, Amartya 1969: ‘Quasi-Transitivity, Rational Choice and
Collective Decisions’. Review of Economic Studies, 36, pp. 381–94.
—— 1970: Collective Choice and Social Welfare. San Francisco:
Holden-Day.
—— 1977: ‘On Weights and Measures: Informational Constraints in
Social Welfare Analysis’. Econometrica, 45, pp. 1539–72.
—— 1986: ‘Social Choice Theory ’. In Arrow and Instiligator 1986,
pp. 1073–181.
Thagard, Paul 1978: ‘The Best Explanation: Criteria for Theory
Choice’. Journal of Philosophy, 75, pp. 76–92.
Tsui, Kai-yuen and John A. Weymark 1997: ‘Social Welfare
Orderings for Ratio Scale Measurable Utilities’. Economic
Theory, 10, pp. 241–56.
Zellner, A., H. Keuzenkamp, and M. McAleer (eds) 2001: Simplicity,
Inference and Modelling. Cambridge: Cambridge University Press.

Okasha, S. - Theory Choice and Social Choice - Kuhn Versus Arrow PDF

Uploaded by

Copyright:

Available Formats

Okasha, S. - Theory Choice and Social Choice - Kuhn Versus Arrow PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Okasha, S. - Theory Choice and Social Choice - Kuhn Versus Arrow PDF

Uploaded by

Copyright:

Available Formats

Theory Choice and Social Choice:

Kuhn versus Arrow

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

unsympathetic to other aspects of his philosophy, and seems to be

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

2. Kuhn’s ‘no algorithm’ argument: clarifications

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

as that notion was understood by traditional philosophers of science,

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

remain — and there is no way of narrowing down the choice to a

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

3. Social choice theory and Arrow’s impossibility theorem

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

Table 1: A profile of preference orders for four individuals

For example, suppose a given society has four alternatives: building a

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

order. In other words, given the preferences of all the members of

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

as it conflicts with basic democratic ideals. So condition N seems

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

conditions. This way of expressing Arrow’s theorem immediately sug-

4. Theory choice cast as a social choice problem

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

there is no one criterion such that if T1 is ranked above T2 by that

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

must end up in the same epistemic state. In the context of theory

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

5. Possible escape routes

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

satisfy conditions U, P, I, and N.10 Does this reconcile Kuhn with

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

data cannot decide between two theories. Similarly, in a discussion of

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

Finally, two further possible escape routes, well-known among

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

6. Sen’s ‘informational basis’ approach

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

‘independence of irrelevant utilities’, says that the social preference

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

of the form vi = aui + b, a > 0, are held to preserve information.13

Downloaded from http://mind.oxfordjournals.org/ at Duke University on May 6, 2012

Mind, Vol. 120 . 477 . January 2011 ! Okasha 2011

impossibility can be avoided, but only if some interpersonal compari-