Binary Choice Probabilities on Mixture Sets

Matthew Ryan
Auckland University of Technology

November 25, 2014

Abstract

Experimental evidence suggests that choice behaviour has a stochastic element.

Much of this evidence is based on studying choices between lotteries –choice under
risk. Binary choice probabilities admit a strong utility representation (SUR) if there
is a utility function such that the probability of choosing option A over option B
is a strictly increasing function of the utility di¤erence between A and B. Debreu
(1958) obtained a simple set of su¢ cient conditions on binary choice probabilities
for the existence of a SUR. More recently, Dagsvik (2008) considered binary choices
between lotteries and provided axiomatic foundations for a SUR in which the un-
derlying utility function is linear (i.e., conforms to expected utility). Our paper
strengthens and generalises Dagsvik’s result. We show that one of Dagsvik’s ax-
ioms can be weakened, and we extend his analysis to encompass choices between
uncertain prospects, as well as various non-linear speci…cations of utility.

1 Introduction
Experiments on risky choice reveal consistent violations of expected utility (EU) theory,
but also robust evidence of stochastic choice in repeated trials. Theoretical responses have
mainly focussed on the former, resulting in a range of alternatives to EU. However, since
the early 1990s increasing attention has been paid to the latter. Indeed, by 1995 John
Hey was prepared to advance the following tentative hypothesis:

“[O]ne can explain experimental anlayses of decision making under risk better
(and simpler) as EU plus noise –rather than through some higher level func-
tional –as long as one speci…es the noise appropriately.”(Hey, 1995, p.640)

1
 This hypothesis is a challenge to econometricians, but also to decision theorists, and
it has revived interest (amongst economists) in models of stochastic choice.
In support of Hey’s hypothesis, consider the evidence of Schmidt and Neugebauer
(2007). They presented 24 subjects with the same 28 binary choice problems on three
consecutive days. The binary choices consisted of pairs of lotteries, constructed so as
to contain 28 pairs of binary choices that re‡ect common consequence e¤ects, common
ratio e¤ects or a mixture of the two. Schmidt and Neugebauer observed that many
subjects made “inconsistent” choices for at least some of these choice problems (i.e.,
they did not make the same choice on all three presentations of the problem), and, more
importantly, that the majority of the violations of expected utility were observed amongst
these “inconsistent choosers”. This evidence is suggestive of noisy choice behaviour, but
of choice behaviour that may be guided by expected utility nonetheless.
Several recent papers attempt to identify the most suitable model for accommodating
both randomness in choice and the possibility of deviations from EU, drawing inspiration
from the psychophysics literature. A strong utility model treats choice as fundamentally
random, with choice probabilities in binary comparisons determined by the di¤erence in
the utility-stimulus of each option. A strong utility model thus involves a utility function
that determines the strength of stimuli and a distribution function that converts utility
di¤erences into choice probabilities.
In the context of binary choices between lotteries, it is natural to ask whether an
expected utility – or some other – functional form is consistent with the data, for some
suitable choice of distribution function. What are necessary and su¢ cient conditions
on binary choice probabilities for a strong utility representation (SUR) of this form?
The necessary conditions provide testable restrictions on observed probabilities; a set of
su¢ cient conditions provides some basis for assessing the theoretical plausibility of the
model.
At present, there is limited understanding of the axiomatic foundations of strong
utility models for choice under risk or uncertainty. Dagsvik (2008) and Blavattsky (2008)
are important recent contributions. Both of these authors study stochastic choice with
expected utility stimuli.1 To the best of my knowledge, no axiomatic treatments exist for
binary stochastic choice with non-expected utility stimuli, or for binary choices between
uncertain (as opposed to risky) prospects. The present paper …lls some of this gap. It
also re…nes the set of su¢ cient conditions provided by Dagsvik (2008) for the EU case.
1
Mention should also be made of Gul and Pesendorfer (2006), who axiomatise random expected utility,
a close relative of the Fechnerian models.

2
2 Binary stochastic choice
Debreu (1958) proved a famous representation theorem for binary stochastic choice. The
primitives of the theory are a pair (A; P ), where A is an arbitrary set of alternatives and
P is a binary choice probability (BCP). The latter is a mapping

P :A A ! [0; 1]

satisfying
P (a; b) = 1 P (b; a) (1)
for all a; b 2 A.
The quantity P (a; b) is the probability with which the decision-maker selects a when
given the choice between a or b (with abstention not being an option). This interpretation
is behaviourally meaningful only if a 6= b, but it is traditional to de…ne P on the entire
Cartesian product A A for convenience. An immediate implication of (1) is that
1
P (a; a) =
2
for all a 2 A.
Given P , it is natural to impute the following binary relation on A: for any a; b 2 A:
1
a%b , P (a; b) P (b; a) , P (a; b) (2)
2
where the second equivalence follows from (1). In other words, a is “weakly preferred”to
b i¤ the decision-maker chooses a over b at least half of the time.
Debreu provides su¢ cient conditions for the existence of a utility function u : A ! R
such that
P (a; b) P (c; d) i¤ u (a) u (b) u (c) u (d) (3)
Choice probabilities are thus determined by utility di¤erences, in the Fechnerian tradition
of psychophysics (Falmagne, 2002).2 Following Marschak (1960) we say that P has a
strong utility representation when (3) obtains for some u.3
2 ^ on A
There is a closely related literature in which the primitive P is replaced by a weak order % A,
called a di¤ erence relation, and su¢ cient conditions are sought for a utility-di¤ erence representation:
^ (c; d)
(a; b) % , u (a) u (b) u (c) u (d) .

See Köbberling (2006) for a recent example and a summary of the previous literature.
3
The literature is not consistent in the use of terminology. Some authors (e.g., Luce and Suppes, 1965)
use “strong utility” for a representation in which (3) is only required to hold when P (a; b) 2= f0; 1g and
P (c; d) 2
= f0; 1g.

3
 Debreu’s su¢ cient conditions comprise two axioms, the …rst of which is sometimes
called the quadruple condition (Davidson and Marschak, 1959):

Axiom D1 For all a; b; a0 ; b0 2 A:

P (a; b) P (a0 ; b0 ) i¤ P (a; a0 ) P (b; b0 )

The quadruple condition is a close relative of the following:

Strong Stochastic Transitivity (SST) For all a; b; c 2 A, if

1
min fP (a; b) ; P (b; c)g
2
then
P (a; c) max fP (a; b) ; P (b; c)g .

It is well known that Axiom D1 implies SST, but not conversely.4 Davidson and
Marschak (1959, p.240) prove that SST is equivalent to the following weak substitutability
condition:5 for any a; b; c 2 A,
1
P (a; b) ) P (a; c) P (b; c) (4)
2
The extra strength of Axiom D1 comes at the cost of some intuitive appeal. The SST
condition has a familiar and transparent logic, while the quadruple condition is arguably
less compelling from a normative point of view.
The second of Debreu’s (1958) axioms is a solvability condition:

Axiom D2 For all a; b; c 2 A and all 2 (0; 1)

P (a; b) P (a; c) ) P (a; e) = for some e 2 A

4
See Luce and Suppes (1965, Theorem 39). Example 1, described below, also shows that SST does
not imply the quadruple condition.
5
If we replace “)” with “,” in (4) we obtain the substitutability condition (Tversky and Russo,
1969), which is stronger (as Example 1 below illustrates). Note that Blavatskyy’s (2008) Axiom 5, which
he calls interchangeability, is implied by weak substitutability –it corresponds to (4) with the inequalities
replaced by equalities. Since Blavatskyy’s axioms include SST (Axiom 2), it follows that his Axiom 5 is
redundant.

4
 Debreu proves that Axioms D1–D2 su¢ ce for a strong utility representation.6
It is important to note that Axiom D1 cannot be replaced by the weaker SST without
jeopardising Debreu’s result.7

Example 1. Let A be a compact, convex subset of some Euclidean space and let u^ :
1 3
A ! R be a continuous function with u^ (A) = [u; u] [0; 1] and ;
4 4
(u; u).
Now de…ne P : A A ! [0; 1] as follows:
8 3 1 3
>
> if [1 + u^ (a) u^ (b)]
>
>
4 2 4
>
>
<
1 1 1
P (a; b) = if [1 + u^ (a) u^ (b)]
>
>
4 2 4
>
>
>
>
: 1
2
[1 + u^ (a) u^ (b)] otherwise

It is straightforward to check that P satis…es (1) and Axiom D2. It also satis…es
weak substitutability (and hence SST):
1 1 1
P (a; b) , [1 + u^ (a) u^ (b)]
2 2 2

, u^ (a) u^ (b)

1 1
, [1 + u^ (a) u^ (c)] [1 + u^ (b) u^ (c)]
2 2

) P (a; c) P (b; c) .
6
Scott (1964) showed that D1 alone is not su¢ cient for a SUR, though it is clearly necessary. Axiom
D2 is not necessary, as the following example illustrates: A = [0; 1] and
8
3 1
>
> + (a b) if a > b
>
>
4 4
>
>
<
P (a; b) = 1
if a = b
>
>
2
>
>
>
>
: 1 1
4 + 4 (a b) if a < b

Then P (a; b) P (c; d) i¤ a b c d but solvability is violated: for example, take any a; b; c with
1 1
c > a > b and any 2 4; 2
.
7
Köbberling (2006, Theorem 1) shows that the QC, which is called the strong crossover property when
translated into the language of di¤erence relations, can be weakened while preserving the SUR. However,
one of her axioms, weak separability, implies the substitutability property, which is stronger than SST.

5
 3
However, P does not satisfy Axiom D1. Let a; b; c 2 A be such that u^ (a) > u^ (b) > 4
and u^ (c) = 41 . Then

1 1
P (a; b) = [1 + u^ (a) u^ (b)] > = P (c; c)
2 2
but P (a; c) = P (b; c) = 34 . Since Axiom D1 is clearly a necessary condition for a
representation of the form (3), no such representation exists for this example.

One consequence of Axiom D2 is that A must be suitably rich. A familiar context

that satis…es this richness condition is choice between risky prospects. For example, A
may be the unit simplex in Rn , interpreted as the set of lotteries over a …xed outcome set
X = fx1 ; :::; xn g. Dagsvik (2008) proves a strong utility representation theorem for this
context. More precisely, he provides su¢ cient conditions for a SUR in which the utility
function u : A ! R is linear. Let us call this a strong expected utility representation
(SEUR).8 Dagsvik’s su¢ cient conditions for a SEUR augment Debreu’s axioms with two
more, which we discuss below (Section 4.1).
Dagsvik’s result is an important re…nement of Debreu’s theorem for binary choices
between risky prospects. The purpose of the present paper is to provide further such
re…nements, encompassing various non-expected utility models of choice under risk or un-
certainty. We also show that Dagsvik’s su¢ cient conditions for a SEUR may be weakened:
the quadruple condition can be replaced by SST.

3 Mixture sets
Our central representation result (Corollary 0) requires only that A is a mixture set (Her-
stein and Milnor, 1953).9 Given a; b 2 A and 2 [0; 1], we write a b for the -mixture of
a and b, where a1b = a. For example, if a; b 2 A Rn then

a b = a + (1 )b

under the standard mixture operation on Rn .

There are many axiom systems for preferences over mixture sets, leading to various
classes of utility functions. The best known, of course, are the axiom systems for expected
8
Luce and Suppes (1965, p.360) use the same appellation to refer to a representation that obeys (3)
whenever P (a; b) 2
= f0; 1g and P (c; d) 2
= f0; 1g, and u is of the expected utility form.
9
However, mixture sets can be discrete, so the richness condition implied by solvability will implicitly
exclude such sets.

6
utility and the associated class of (mixture-)linear utility functions, but a wide range of
non-expected utility classes have also been axiomatised for speci…c mixture-set contexts.
It is therefore natural to seek probabilistic extensions of these expected and non-expected
utility models, leading to strong utility representations in which the utility function u :
A ! R falls within the relevant class.
To determine suitable classes of utility functions for this purpose, let us introduce the
following concept:

De…nition 1 Given some M A we say that u : A ! R is M -linear if u (M ) = u (A)

and
u (a b) = u (a) + (1 ) u (b)

for any a 2 A, any b 2 M and any 2 [0; 1].

If M = A then M -linearity is (mixture-)linearity simpliciter. However, if M is a

proper subset of A, then M -linearity is a weaker property of u. Importantly, several non-
EU models (for preferences over A) have M -linear representations for some M A (see
Section 4).
Given M A, if P has a strong utility representation with respect to an M -linear
utility function u we say that P has a strong M -linear utility representation. The following
is clearly necessary for such a representation to exist:

Strong M -Independence For any a; b; c; d 2 A, any e 2 M and any 2 (0; 1),

P (a; b) P (c; d) ) P (a e; b e) P (c e; d e) .

Note that Strong M -Independence implies Strong M 0 -Independence for any M 0 M.

Strong M -Independence is a weakening of Dagsvik’s (2008) Strong Independence axiom:

Strong Independence For all a; b; a0 ; b0 ; c 2 A and all 2 (0; 1)

P (a; b) P (a0 ; b0 ) ) P (a c; b c) P (a0 c; b0 c) .

Strong Independence implies Strong M -Independence for any M A.

Our main representation result is a corollary (Corollary 0) to the following theorem.

Theorem 0 Let be an interval in R (not necessarily bounded) containing 0 in its

interior. Suppose : ! [0; 1] satis…es the following conditions for any
0 0
2 [0; 1] and any x; y; x ; y ; z 2 :

7
 1
(i) (x; y) = 2
i¤ x = y
(ii) (x; y) + (y; x) = 1
(iii) (x; y) = (x0 ; y 0 ) implies

( x + (1 ) z; y + (1 ) z) = ( x0 + (1 ) z; y 0 + (1 ) z)

(iv) is non-decreasing (respectively, non-increasing) in its …rst (respectively, sec-

ond) argument
(v) is continuous in each argument

Then
(x; y) (x0 ; y 0 ) i¤ x y x0 y0 (5)

for any x; y; x0 ; y 0 2 .

The proof of Theorem 0 may be found in Appendix A.

To interpret Theorem 0, suppose we have a representation u : A ! R for the preference
relation % de…ned by (2). Let = u (A). Assuming 0 2 int ( ) is without loss of
generality (WLOG) provided preferences are non-trivial ( is not a singleton). Now re-
calibrate P in terms of utility pairs by de…ning a mapping : ! [0; 1] as follows:
(x; y) = P (a; b) for any a; b 2 A with u (a) = x and u (b) = y. Provided this mapping is
well-de…ned and satis…es (i)–(v), Theorem 0 ensures that (3) holds. Corollary 0 establishes
su¢ cient conditions for these provisos to be met. In particular, Strong M -Independence
ensures (iii).10

Corollary 0 Let A be a mixture set and M A, and let P satisfy SST, D2 and Strong
M -Independence. If u : A ! R is M -linear and represents % de…ned by (2), then
(3) holds for any a; b; c; d 2 A.
10
Note that the function induced by Example 1 violates property (iii). Suppose, for example, that
1
u = 1 and u = 0. Let x = 1, x0 = 2 and y = z = 0, and note that

3
(x; 0) = (x0 ; 0) =
4
while
1 3 5 1 0
x; 0 = 6= = x ;0 .
2 4 8 2

8
 Corollary 0 is the main result of the paper. It is proved in Appendix A.
Corollary 0 suggests the following “recipe”for producing strong M -linear utility rep-
resentation theorems.

Step I. Choose a set of axiomatic restrictions on preferences over A (i.e., % A A)

that are su¢ cient for an M -linear utility representation. Translate these axioms
into the corresponding restrictions on P using (2).

Step II. Add SST, D2 and Strong M -Independence, then apply Corollary 0.

In most cases, the axioms required at Step II already imply certain of the restrictions
needed for Step I, allowing for some economies to be achieved.
Of course, for this recipe to be useful from a normative point of view, Strong M -
Independence must be a reasonable restriction on P . If we assume –as seems natural –
that the binary relation % de…ned by (2) orders the (utility) stimuli experienced by the
agent, then we may argue as follows.
The axioms imposed at Step I embody an implicit argument that % should possess an
M -linear representation. If so, then % satis…es the following condition (which might be
called M -Independence): for any a; b 2 A, any c 2 M and any 2 (0; 1),

a%b ) a c % b c.

In other words, if option a generates a stimulus at least as great as option b, then this
remains true after mixing with a common element from M . For example, it may be that
elements from M do not (or should not) “interfere” with other stimuli under mixing, so
options are (or should be) compared by …rst excluding common elements mixed in from
M . By this logic –which need not be compelling in all circumstances –we have reason to
expect that choice probabilities will (or should) be una¤ected by mixing with a common
element from M .

4 Applications
The following sections apply our recipe to prove strong M -linear utility representation
theorems for a variety of speci…cations of A, M and the mixture operation a b. These
provide selective illustrations only –other applications are no doubt possible and poten-
tially of interest.

9
 We …rst consider the case M = A, so u is (mixture-)linear simpliciter. When A is the
unit simplex in Rn and the usual mixture operation is assumed, we are in the domain of
Dagsvik (2008). Our recipe therefore provides an alternative axiomatisation, and proof, of
the strong expected utility representation. If A is a set of acts of the Anscombe-Aumann
(AA) variety, we obtain a stochastic extension of subjective expected utility.
We then consider three examples in which M is a proper subset of A, obtaining
stochastic extensions of dual theory (DT), Choquet expected utility (CEU) and maxmin
expected utility (MEU).
In proving these results, much of the work is in extracting economies at Step I from
the axioms needed at Step II. More direct applications of the recipe would yield shorter
proofs, but longer lists of axioms in the statements of propositions.
For future reference, we de…ne the following natural strengthening of Axiom D2, which
will be useful in the sequel:

Mixture Solvability (MS) For all a; b; c 2 A and all 2 (0; 1)

P (a; b) P (a; c) ) P (a; b c) = for some 2 [0; 1]

The MS condition, together with Strong M -Independence, ensures that the binary
relation (2) possesses su¢ cient “continuity”to make Step I in the examples below.11
Note that if A in Example 1 is a mixture set and u^ is (mixture-)linear, then MS is
satis…ed. It follows that strengthening D2 to MS does not, on its own, allow us to weaken
the quadruple condition to SST in Debreu’s theorem. This observation is signi…cant, as
our representation theorems will use SST rather than the stronger quadruple condition.

4.1 Strong expected utility

Let us …rst recall Dagsvik’s (2008) representation theorem.12
In addition to Axioms D1, D2 and Strong Independence, Dagsvik also requires:

Axiom D3 [Archimedean Property] For all a; b; c 2 A, if

1
P (a; b) > > P (c; b)
2
11
See Lemma 1.
12
Blavatskyy (2008) proves a closely related result, using di¤erent axioms. However, Blavatskyy’s
representation is of the form
P (a; b) = f [u (a) u (b)]
for some non-decreasing function f : R ! R, so it is not a strong utility representation. In particular, it
is easy to check that Example 1 satis…es all of Blavatskyy’s axioms when u ^ is linear.

10
 then there exist ; 2 (0; 1) such that
1
P (a c; b) > > P (a c; b) .
2

Theorem D [Dagsvik, 2008, Theorems 2 and 4] If A is the unit simplex in Rn and

P satis…es D1-D3 and Strong Independence, then P has a strong expected utility
representation.

Corollary 0 may be applied to give a variation on Theorem D; one which is applicable

to any mixture set. It uses SST in place of the (stronger) quadruple condition (Axiom
D1). Note that several economies are possible at Step I. The restrictions on P imposed
by transitivity and mixture-independence of the binary relation % are guaranteed by SST
and Stong Independence respectively. We invoke MS and Strong Independence to ensure
that P also satis…es the restrictions implied by a standard continuity condition on %
(Lemma 1).

Proposition 1 If A is a mixture set and P satis…es SST, MS and Strong Independence,

then P has a strong expected utility representation.

Proof. Let % be the binary relation de…ned by (2). For Step I it su¢ ces to prove that
% has an A-linear (i.e., linear) representation.
It is obvious that % is complete. It is transitive by SST. Setting a0 = b0 in the de…nition
of Strong Independence, we see that % inherits the following independence property: for
all a; b; c 2 A and all 2 (0; 1)

a%b ) a c%b c (6)

Finally, % satis…es a standard continuity property.

Lemma 1 For any a; b; c 2 A, the sets

f 2 [0; 1] j a b % cg

and
f 2 [0; 1] j c % a bg

are closed.

11
Proof. The basic idea of the proof is simple. First, we use (6) to show that preferences
are monotone in . Second, MS implies that preferences exhibit solvability with respect
to . Finally, we observe that monotonicity plus solvability implies continuity in .
Fix some a; b; c 2 A and consider the set

f 2 [0; 1] j c % a bg (7)

It is WLOG to assume a % b. Then

a b%b

for any 2 (0; 1) by (6). Hence, applying (6) once more:

a b % b (a b) , a b % a [ (1 )] b
0
for any 2 (0; 1). In other words, if is in (7) then so is any < . It remains to
exclude the possibility that (7) is of the form [0; ) for some > 0.
Let f 1
m gm=1 [0; ) be a convergent sequence with limit . Suppose a mb % c for
each m, while c a b. Then
1
P (c; a b) > P (c; a m b)
2
for each m. Let
1
x2 P (c; a b) ; .
2
Given some m 2 f1; 2; :::g, MS ensures that there exists 2 (0; 1) such that

x = P (c; (a b) (a m b)) = P (c; a [ + (1 ) m ] b) .

1
Since + (1 ) m 2 [0; ) and x > 2
we have the desired contradiction.
The same sort of argument (mutatis mutandis) shows that the set

f 2 [0; 1] j a b % c g

is closed.

By Theorem 1 of Fishburn (1982, Chapter 2), % possesses a linear representation, u.

The result now follows by Corollary 0. In particular, Strong Independence implies Strong
A-Independence.

Proposition 1 provides a more “compact” axiomatisation of strong expected utility

than Theorem D, having one fewer axiom (D3 is not required). However, it uses a

12
strengthened version of solvability (MS rather than D2). On the other hand, SST is
weaker than QC, so the relationship between the two sets of axioms is not transparent.
Appendix B o¤ers some clari…cation. It proves a stronger version of Theorem D,
called Theorem D , in which SST replaces Axiom D1 (QC). Proposition 1 and Theorem
D show that MS is equivalent to the conjunction of D2 and D3, given SST and Strong
Independence.
The proof of Theorem D follows our usual recipe, but a more elaborate argument
is required at Stage I, since we cannot use MS to establish the required continuity of %
(Lemma 1). Instead, the argument leans more heavily on Strong Independence, which, it
turns out, imposes restrictions on % over and above the independence condition (6).

4.2 Preferences over acts

For the remaining applications we assume that A has the following structure: there is
a set S (the “states”) and a mixture set C (the “consequences”) and A is the set of all
functions from S to C. The elements of A are thus acts. We adopt the standard mixing
operation on A: given a; b 2 A and 2 [0; 1], the function a b 2 A maps state s 2 S to
c
a (s) b (s) 2 C. We also use A to denote the set of constant functions in A, and take
the usual notational liberty of identifying Ac with C.

4.2.1 Subjective expected utility

Consider an Anscombe-Aumann environment in which S is …nite and C is the unit simplex

in Rn , interpreted as the set of roulette lotteries over a …xed set of n prizes. We apply
the usual mixing operation on C.
It will also be convenient to de…ne
( )
X
(S) = : S ! [0; 1] (s) = 1 .
s

A utility function u : A ! R has the SEU structure if it takes the form

X
u (a) = (s) v (a (s)) (8)
s2S

for some linear function v : C ! R and some 2 (S). Note that u in (8) is A-linear,
but not every A-linear function has the form (8). If (3) holds with u of the SEU form, we
say that P has a strong SEU representation.

13
 We may apply our recipe to obtain a strong SEU representation theorem. Several
economies are possible in specifying the axioms. Given the axioms introduced at Step II,
we need only add the following to make Step I:13

Stochastic State Monotonicity For any a; b 2 A, if

1
P (a (s) ; b (s))
2
1
for every s 2 S then P (a; b) 2
.

Proposition 2 If P satis…es SST, MS, Stochastic State Monotonicity and Strong Inde-
pendence, then P has a strong SEU representation.

Proof. By Corollary 0, it su¢ ces to show that % has a SEU representation.

Observe that Ac is a mixture set and the restriction of P to Ac Ac satis…es SST,
Strong Independence and MS. From the proof of Proposition 1, we therefore deduce that
the restriction of % to Ac is a weak order satisfying (6) and (7), and hence possesses
a linear representation. Denote the latter by v : C ! R and let = v (C). Then
is a closed interval. Given Stochastic State Monotonicity, the result is trivial if is a
singleton, so assume otherwise.
S
Consider the following binary relation on the set of state-utility vectors: x y
if a % b for some a; b 2 A with xs = v (a (s)) and ys = v (b (s)) for each s 2 S. This binary
relation is well-de…ned by Stochastic State Monotonicity: for any a; b 2 A

a (s) % b (s) for all s 2 S ) a%b (9)

Let > denote the asymmetric part of .

It is straightforward to verify (using SST and Strong Independence) that is a weak
order that satis…es

x y ) x + (1 )z y + (1 )z

S
for all x; y; z 2 and all 2 (0; 1). To prove that has a linear representation, it
su¢ ces (see the proof of Proposition 1) to show:

S
Lemma 2 The following sets are closed for all x; y; z 2 :

f 2 [0; 1] j x + (1 )y zg
13
Recall that we identify the consequence p 2 C with the constant function in Ac that maps each state
to p.

14
 and
f 2 [0; 1] j z x + (1 )yg.

S S
Proof. De…ne P : ! [0; 1] as follows: P (x; y) = P (a; b) if there exists a; b 2 A
with xs = v (a (s)) and ys = v (b (s)) for each s 2 S. This is well-de…ned, since weak
substitutability implies P (a; b) = P (a0 ; b0 ) whenever a a0 and b b0 , and the state
monotonicity property (9) means that a a0 whenever a (s) a0 (s) for all s 2 S. Thus,
1
x y i¤ P (x; y) 2
.
Note that P inherits the MS property from P : for any x; y; z 2 , any 2 (0; 1) and
any a; b; c 2 A with xs = v (a (s)), ys = v (b (s)) and zs = v (c (s)) for each s 2 S:

P (x; y) P (x; z) ) P (a; b) P (a; c)

) P (a; b c) = for some 2 [0; 1]
) P (x; y + (1 ) z) =

where the …nal inequality uses the linearity of v.

Lemma 2 now follows by the same argument as we used to prove Lemma 1.

Thus, has a linear representation: there exists : S ! R such that

X X
x y , (s) xs (s) ys
s s

S
for any x; y 2 . Using (9) and the fact that is a non-degenerate interval, it is easily
veri…ed that (s) 0 for all s and strictly so for at least one. We may therefore normalise
so it lies in (S).

4.2.2 Choquet expected utility

An important generalisation of SEU, within the AA environment, is the Choquet expected

utility (CEU) model of Schmeidler (1989). If there exists a utility function u : A ! R of
the CEU form (de…ned below) such that (3) holds for any a; b; c; d 2 A, we say that P
has a strong CEU representation.
We continue to work the Anscombe-Aumann speci…cation of A – and the related
mixture operations – from the previous section. However, to discuss CEU we need a
couple of additional de…nitions.
Mappings f : S ! R and g : S ! R are comonotonic if there do not exist states
s; s 2 S with f (s) > g (s) and g (s0 ) > f (s0 ). Analogously, we say that acts a; b 2 A
0

15
are comonotonic if there do not exist states s; s0 2 S with a (s) b (s) and b (s0 ) a (s0 ).
The notion of comonotonicity is central to the CEU model.
A capacity on S is a mapping ! : 2S ! [0; 1] that satis…es ! (;) = 0, ! (S) = 1 and
! (A) ! (B) whenever A B. Capacities are non-additive generalisations of probability
measures.
A utility function u : A ! R has the CEU form if there exists a linear function
v : C ! R and a capacity ! : 2S ! [0; 1] such that

Z0 Z1
u (a) = [! (fs 2 S j v (a (s)) > z g) 1] dz + ! (fs 2 S j v (a (s)) > z g) dz
1 0

We may write this more compactly using the Choquet integral (Denneberg, 1994):
Z
u (a) = (v a) d! (10)

If ! is additive, and hence a probability measure, then (10) is just the usual expected
value of v a with respect to !.
Note that if preferences over A have CEU representation (10), then acts a; b 2 A are
comonotonic i¤ the mappings v a and v b are comonotonic.
It is well known (Denneberg, 1994) that the Choquet integral is homogeneous and
comonotonically additive. That is, given any two comonotonic functions f and g, any
> 0 and any capacity !: Z Z
f d! = f d!.

and Z Z Z
(f + g) d! = f d! + g d!.

It follows, using the linearity of v, that

u (a b) = u (a) + (1 ) u (b)

whenever a; b 2 A are comonotonic. Since b 2 Ac is comonotonic with any a 2 A, the

CEU utility function (10) is Ac -linear.14
14
One may verify that u (Ac ) = u (A) as follows. The function

F (z) = ! (fs 2 S j v (a (s)) z g)

is the cumulative distribution function for a random variable whose range is contained in v (C) and (10)
is the expected value of this random variable. It follows that u (A) v (C) = u (Ac ).

16
 Strong Ac -Independence is therefore necessary for the existence of a strong CEU rep-
resentation. However, Strong Ac -Independence, together with SST, MS and Stochastic
State Monotonicity, do not su¢ ce.15 We need the following additional condition to com-
plete Step I:16

Stochastic Comonotonic Independence For any pairwise comonotonic a; b; c 2 A

and any 2 (0; 1),
1 1
P (a; b) ) P (a c; b c) .
2 2
Proposition 3 If P satis…es SST, MS, Strong Ac -Independence, Stochastic State Monotonic-
ity and Stochastic Comonotonic Independence, then P has a strong CEU represen-
tation.

Proof. We …rst show that % de…ned by (2) has a CEU representation. Our argument
follows the structure in Ryan (2009; Section 5).
As in the proof of Proposition 2, we …rst consider the restriction of P to Ac Ac and
the associated restriction of % to Ac . By the same argument as in the earlier proof,17 %
(restricted to Ac ) has a linear representation, which we denote v : C ! R. As before, we
let = v (C), a closed interval. The case of a singleton being trivial, it is excluded
henceforth.
De…ne the binary relation on S
as before: x y if a % b for some a; b 2 A with
xs = v (a (s)) and ys = v (b (s)) for each s 2 S.
For each permutation : S ! f1; 2; :::; jSjg we de…ne the associated (convex) comonotonic-
ity region C ( ) as in Ryan (2009): x 2 C ( ) i¤

x (1) x (2) x (jSj) .

Let denote restricted to C ( ). Strong Comonotonic Independence ensures that

each satis…es

x y ) x + (1 )z y + (1 )z
15
Consider preferences % A A which have a maxmin expected utility representation u : A ! R (see
Section 4.2.4) but no CEU representation. Let Z = u (A) u (A) and let P (a; b) = g (u (a) u (b)) for
some continuous and strictly increasing function g : Z ! [0; 1] with g (z) + g ( z) = 1. Then P satis…es
SST, MS, Strong Ac -Independence and Stochastic State Monotonicity, but it does not have a strong CEU
representation.
16
Stochastic Comonotonic Independence is the restriction on P imposed by comonotonic independence
(Schmeidler, 1989) of the binary relation % de…ned by (2).
17
In particular, Strong Ac -Independence of P ensures that the restriction of P to Ac Ac satis…es
Strong Independence.

17
for all x; y; z 2 C ( ) and all 2 [0; 1]. We may therefore follow the same logic as in the
proof of Proposition 2 to show that each has a linear representation. That is, for each
permutation , there exists 2 (S) such that
X X
x y , (s) xs (s) ys (11)
s s

for any x; y 2 C ( ). The argument in Ryan (2009, pp.345-7) now shows that there exists
a capacity ! : 2S ! [0; 1] such that x y i¤
Z0 Z1
[! (fs 2 S j xs > z g) 1] dz + ! (fs 2 S j xs > z g) dz
1 0

Z0 Z1
[! (fs 2 S j ys > z g) 1] + ! (fs 2 S j ys > z g) dz
1 0

This gives the desired CEU representation for %.

To complete the proof, apply Corollary 0.

4.2.3 Dual theory

The Anscombe-Aumann environment plays a crucial role in our strong CEU representation
result. We needed the linearity of v : C ! R to establish the Ac -linearity of u : A ! R.
Our arguments would not su¢ ce in a Savage environment.
For the same reason, obtaining a SUR for the rank-dependent expected utility (RDEU)
model under risk is problematic. However, the special case of RDEU known as the Dual
Theory (DT) model (Yaari, 1987) is amenable to the application of our recipe.
In the following, we exploit the close relationship between CEU and RDEU, originally
pointed out by Wakker (1990). However, we start from a Savage-type environment, with
C = R (and the usual mixture operation on R). We retain the assumption that S is
…nite, but endow S with a probability measure p so that acts embody risk rather than
uncertainty.
Note that this structure induces a rather limited range of distributions on C. It
is therefore conventional to work with a richer set of states (such as [0; 1] endowed with
Lebesgue measure) when studying risk. We could have done so here, but the more restric-
tive environment allows us to leverage o¤ our e¤orts in the previous section to expedite
the analysis. The reader is referred to Ryan (2009, Appendix 5 of the Supplementary
Material) for the required elaborations.

18
 A utility function u : A ! R has the DT form if there exists a non-decreasing and
surjective function : [0; 1] ! [0; 1] such that
Z
u (a) = a d( p) (12)

where integration is in the sense of Choquet.18 We observe, recalling the analysis of the
CEU model, that (12) is Ac -linear.
The key to our result is the following strengthening of Stochastic State Monotonicity:

Stochastic Monotonicity For any a; b 2 A, if the distribution on C induced by a …rst-

order stochastically dominates the distribution induced by b, then
1
P (a; b) .
2
Proposition 4 If P satis…es SST, MS, Stochastic Monotonicity, Strong Ac -Independence
and Stochastic Comonotonic Independence, then P has a strong utility representa-
tion with respect to some u of the DT form (12).

Proof. We …rst show that % de…ned by (2) has a DT representation. This binary relation
shares the same properties as in the proof of Proposition 3. By the argument in that
proof, % therefore possesses a CEU representation
Z
u (a) = a d!

for some capacity !. By Stochastic Monotonicity, ! = p for some non-decreasing and

surjective mapping : [0; 1] ! [0; 1] (Wakker, 1990).
c
Since u is A -linear, we may apply Corollary 0 to complete the proof.

4.2.4 Maxmin expected utility

For our …nal example, we return to the Anscombe-Aumann environment.

A utility function u : A ! R has the maxmin expected utility (MEU) form if there
exists a linear function v : C ! R and a closed and convex set P (S) such that
X
u (a) = min (s) v (a (s)) (13)
2P
s2S

(Gilboa and Schmeidler, 1989). Note that (13) is Ac -linear.

To obtain a SUR with u of the MEU form, we shall require:
18
It is easily checked that p is a capacity.

19
Stochastic Uncertainty Aversion For any a; b 2 A and any 2 (0; 1),
1 1
P (a; b) = ) P (a b; b) .
2 2

Proposition 5 If P satis…es SST, MS, Stochastic State Monotonicity, Stochastic Uncer-

tainty Aversion and Strong Ac -Independence, then P has a strong utility represen-
tation with respect to some u of the MEU form (13).

Proof. As for the proof of Proposition 2, we deduce that the restriction of % to Ac

possesses a linear representation, v : C ! R. We let = v (C) and exclude the trivial
case in which is a singleton. We may therefore assume, WLOG, that contains 0 in
S
its interior. The binary relation on is de…ned in the usual way, with asymmetric
part > and symmetric part = .
S
We shall also identify with the set of constant vectors in .
S
Then is a weak order that satis…es the following conditions for any x; y 2 , any
k2 and any 2 (0; 1):

x y ) x + (1 )k y + (1 )k (14)

x= y ) x + (1 )y x (15)

Furthermore, Stochastic State Monotonicity implies that satis…es the following strict
S 0
monotonicity condition: for any x; y 2 and any k; k 2 ,

x y ) x y (16)

and
k > k0 ) k > k0 (17)

We next prove that there exists a closed and convex set P (S) such that is
represented by
X
u (x) = min (s) xs (18)
2P
s2S

The result will then follow by Corollary 0. The following lemmata establish the required
representation (18).

S
Lemma 3 For any x 2 , there exists a unique k 2 such that x = k.

20
Proof. De…ne P as in the proof of Lemma 2, and note that it satis…es mixture solvability
by the same argument as before. Let x = mins2S xs and x = maxs2S xs . Then (16) implies
1
P (x; x) P (x; x) .
2
Mixture solvability ensures that x = x + (1 ) x for some 2 [0; 1]. Uniqueness
follows by (17).

In view of Lemma 3, for each x 2 we let k x denote the unique k 2 satisfying

x = k.

S
Lemma 4 For any x; y 2 , any k 2 and any 2 (0; 1):

x> y ) x + (1 )k > y + (1 )k

Proof. If x > y then k x + (1 ) k > k y + (1 ) k. Since

x + (1 )k = k x + (1 )k

and
y + (1 )k = k y + (1 )k
by (14), the result follows.

Combining this result with (14) we have:

S
Corollary 1 For any x; y 2 , any k 2 and any 2 (0; 1):

x y , x + (1 )k y + (1 )k (19)

Lemma 5 There is a closed, convex cone K RS such that

S S
x2 x 0 = K\ .

Moreover, for any k 2 ,

S S
x2 x k = (K + fkg) \ .

Proof. Lemma 2 and Corollary 2 in Ryan (2009) ensure the existence of a suitable convex
cone K. It remains to show that K is closed, for which it su¢ ces to show that the set
S
x2 x 0 (20)

21
is closed, since 0 lies in the interior of .
Let fxm g1
m=1
S
be a sequence with limit x^ 2 S
such that xm 0 for all m and
0 > x^. Since the set (20) is convex, x^ must be a boundary point. Let k^ > 0 be from the
interior of . Then k^ 2 S is in the interior of (20) and we have
1
P (0; x^) > P 0; k^ .
2

Given 2 1
2
;P 0; k^ there exists some 2 (0; 1) such that

P 0; x^ k^ = .

But x^ k^ must lie in the set (20), which is a contradiction.

From Lemma 5 and the fact that RS+ K the desired representation (18) follows by
the argument on p.338 of Ryan (2009).
This completes the proof of Proposition 5.

5 Concluding remarks
Despite its simple –and hence restrictive –form, the strong utility model continues to be a
workhorse for the empirical analysis of choice behaviour. Schmidt and Neugebauer (2007),
for example, use it to test the hypothesis of Hey (1995) mentioned in the Introduction.19
Much of the empirical literature is based on analysing choices between alternatives
that embody risk (experimental data) or uncertainty (natural data), and non-EU forms
for utility are often assumed or tested. The present paper provides su¢ cient conditions
for the validity of several such models.
In fact, our conditions are also necessary for the versions of strong utility that are
typically encountered in empirical implementations. These assume that

P (a; b) = Pr [X u (a) u (b)] (21)

for some symmetric, zero-mean random variable X with a continuous distribution func-
tion. The existence of a SUR already implies that

P (a; b) = [u (a) u (b)] (22)

19
In fact, their empirical framework is somewhat less restrictive than the strong utility model, but the
added generality is irrelevant to the tests of their core hypotheses.

22
for some strictly increasing function : Z ! [0; 1] satisfying (z) + ( z) = 1, where
20
Z = u (A) u (A). If satis…es (Z) = [0; 1], then is the distribution function for
some zero-mean, symmetric random variable.21 Inspection of Propositions 1–5 reveals
that MS is the only axiom which is not necessary for (22). However, it is necessary if
is assumed to be continuous, as is typically done in applied work.
Numerous authors (e.g., Butler, Isoni and Loomes, 2012) have observed that the struc-
ture imposed on BCP’s by (21) may be unduly restrictive for describing observed choice
behaviour, and many studies allow the distribution of the random variable X to vary with
aspects of the binary choice problem. This is an empirical challenge to the axiomatic foun-
dations of model (22), but it is nonetheless useful to render these foundations transparent.
We can then better understand the nature of the empirical violations and how best to
modify the axioms to accommodate them.
Of course, our axiomatisations are restricted to M -linear utility classes and therefore
exclude important examples such as RDEU (apart from the DT sub-class).22 Filling these
gaps is an important item for the future research agenda.

References
Anscombe, F.J. and Aumann, R.J. (1963) “A De…nition of Subjective Probability,”
Annals of Mathematical Statistics, 34, 199–205.
Blavatskyy, P.R. (2008), “Stochastic Utility Theorem,”Journal of Mathematical Eco-
nomics 44, 1049–1056.
Butler, D., A. Isoni and G. Loomes (2012), “Testing the ‘Standard’Model of Sto-
chastic Choice under Risk,”Journal of Risk and Uncertainty 45, 191–213.
Dagsvik, J.K. (2008), “Axiomatization of Stochastic Models for Choice Under Uncer-
tainty,”Mathematical Social Sciences 55, 341–370.
Davidson, D. and J. Marschak. (1959), “Experimental Tests of a Stochastic Deci-
sion Theory,” in C.W. Churchman and P. Ratoosh (eds), Measurement: De…nitions and
Theories. John Wiley and Sons: New York, NY.
20
Note that u (A) is a compact interval for all the models studied here, so Z is a compact interval
centred at 0.
21
Even if (Z) is a proper subset of [0; 1], we can extend , symmetrically, to some Z containing Z
such that Z = [0; 1].
22
Using a strong utility framework with Normally distributed “errors”, Hey and Orme (1994) found
that RDEU provided the best …t to their data. They also tested DT, which performed poorly.

23
Debreu, G. (1958), “Stochastic Choice and Cardinal Utility,”Econometrica 26(3), 440–
444.
Denneberg, D. (1994), Non-Additive Measure and Integral. Kluwer: Dordrecht, The
Netherlands.
Falmagne, J.C. (2002), Elements of Psychophysical Theory. Oxford University Press.
Fishburn, P.J. (1982), The Foundations of Expected Utility. D. Reidel Publishing:
Dordrecht.
Fishburn, P.J. (1998), “Stochastic Utility,”in S. Barberà, P.J. Hammond and C. Seidl
(eds), Handbook of Utility Theory. Kluwer Academic: Dordrecht.
Gilboa, I. and D. Schmeidler (1989), “Maxmin Expected Utility with Non-Unique
Prior,”Journal of Mathematical Economics 18, 141–153.
Gul, F. and W. Pesendorfer (2006), “Random Expected Utility,” Econometrica
74(1), 121–146.
Herstein, I.N.. and J. Milnor (1953), “An Axiomatic Approach to Measurable Util-
ity,”Econometrica 21, 291–297.
Hey, J.D. (1995), “Experimental Investigations of Errors in Decision Making under
Risk,”European Economic Review 39, 633–640.
Hey, J.D. and C. Orme (1995), “Investigating Generalizations of Expected Utility
Theory Using Experimental Data,”Econometrica 62(6), 1291–1326.
Kobberling, V. (2006), “Strength of Preference and Cardinal Utility,”Economic Theory
27, 375–391.
Luce, R.. and P. Suppes (1965), “Preference, Utility and Subjective Probability,” in
R.D. Luce, R.B. Bush and E. Galanter (eds), Handbook of Mathematical Psychology,
Volume III. John Wiley and Sons: New York.
Marschak, J. (1960), “Binary Choice Constraints and Random Utility Indicators,” in
K.J. Arrow, S. Karlin and P. Suppes (eds), Mathematical Methods in the Social Sciences.
Stanford University Press: Stanford, CA.
Ryan, M.J. (2009), “Generalizations of SEU: A Geometric Tour of Some Non-Standard
Models,”Oxford Economic Papers 61, 327–354.
Schmeidler, D. (1989), “Subjective Probability and Expected Utility without Additiv-
ity,”Econometrica 57(3), 571–587.

24
Schmidt, U. and T. Neugebauer (2007), “Testing Expected Utility in the Presence
of Errors,”Economic Journal 117, 470–485.

Scott, D. (1964), “Measurement Structures and Linear Inequalities,”Journal of Math-

ematical Psychology 1, 233–247.

Suppes, P. (1961), “Behavioristic Foundations of Utility,”Econometrica 29, 186–202.

Suppes, P. and M. Winet (1955), “An Axiomatisation of Utility Based on the Notion
of Utility Di¤erences,”Management Science 1(3/4), 259–270.

Tversky, A. and J.E. Russo (1969), “Substitutability and Similarity in Binary Choices,”
Journal of Mathematical Psychology 6, 1–12.

Wakker, P.P. (1990), “Under Stochastic Dominance Choquet-Expected Utility and

Anticipated Utility are Identical,”Theory and Decision 29, 119–132.

Yaari, M. (1987), “The Dual Theory of Choice under Risk,” Econometrica 55(1), 95–
115.

Appendices

A Proofs of Theorem 0 and Corollary 0

A.1 Proof of Theorem 0
To prove Theorem 0, suppose that : ! [0; 1] satis…es (i)–(v). Let us de…ne

= f(x; y) 2 jx y = 0g

for ease of reference. Property (i) says that is the level curve corresponding to = 21 .
We need to show that all the other level curves are parallel to this one.
First observe that we may strengthen the monotonicity condition (iv).23

Lemma A The function is strictly increasing (respectively, decreasing) in its …rst (re-
spectively, second) argument.
23 1
Since (x; y) 2 i¤ x y from (i), Lemma 0 implies that exhibits the following substitutability
property:
1
(x; y) i¤ (x; z) (y; z) .
2

25
Proof. Suppose x > x0 but (x; y) = (x0 ; y). Properties (i) and (iv) imply that x0 > y
or y > x. Assume the former (i.e., y > x > x0 ), the latter case being handled similarly.
Thus
1
(x; y) = (x0 ; y) > (23)
2
from (i) and (iv).
De…ne x0 = x, x1 = x0 and
x1 y
= ,
x0 y
so x1 = x0 y. For each k 2 f2; 3; :::g let

k
xk = xk 1
y = x0 y.

Since 2 (0; 1) we have xk ! y as k ! 1. By (v) and (i), we therefore have

1
lim xk ; y = (y; y) = (24)
k!1 2
Since (x0 ; y) = (x1 ; y), successive applications of (iii) give

xk ; y = (x; y)

for all k, so (23) and (24) deliver the required contradiction.

Now suppose, contrary to what we seek to show, that there exist x; y; x0 ; y 0 2 with

x y = x0 y 0 6= 0

but
(x; y) > (x0 ; y 0 ) .

It su¢ ces to consider the case24

x y = x0 y 0 < 0.

Thus, (x; y) and (x0 ; y 0 ) are distinct points lying above and on a line parallel to it. See
Figure 1 (which assumes x0 > x).
24
The other case follows directly, since (x; y) > (x0 ; y 0 ) implies (y; x) < (y 0 ; x0 ) by virtue of (ii).

26
 Figure 1

From (i) and Lemma A we have

1
(x0 ; y 0 ) < (x; y) < (y; y) = = (x0 ; x0 ) .
2
Therefore, using (v) and Lemma A, we have (x; y) = (x0 ; y 00 ) for some y 00 2 (x0 ; y 0 ).
Since the line joining (x; y) and (x0 ; y 00 ) is not parallel to we have

(x0 ; y 00 ) = (x z; y z) (25)

for some z 2 R and some 2 (0; 1). Figure 2 illustrates (again for the case x < x0 ). We
may assume that z 2 ; otherwise, we can use (iii) to contract (x; y) and (x0 ; y 00 ) towards
(0; 0) until this condition is satis…ed.25
Starting from (25), repeated applications of (iii) give that
1 n n
> (x; y) = (x z; y z) for each n 2 f1; 2; :::g (26)
2
n n
Since (x z; y z) ! (z; z) as n ! 1 and (z; z) = 12 , we shall show that (26) contradicts
(v).
Given Lemma A and the fact that x < y, (26) can only hold if z > y > x (as in Figure
2) or z < x < y. Consider the former case, the latter being handled similarly. Then
1
(x; z) < (x; y) < = (z; z)
2
25
Recall that 0 2 int ( ).

27
and (v) ensures that (x; y) = (x z; z) for some some 2 (0; 1). Hence, by (iv),

(x; y) < (x z; z) < (x z; y z)

for all < . This contradicts (26).

We have therefore shown that is constant on f(x; y) 2 j x y = kg for any
…xed k 2 R. From this fact and Lemma A we deduce (5). This completes the proof of
Theorem 0.

Figure 2

A.2 Proof of Corollary 0

Let = u (M ) = u (A). Since u is M -linear, is an interval in R. The result is trivial if
is a singleton, so we assume otherwise henceforth. It is without loss of generality to sup-
pose that 0 is contained in the interior of –if not, apply a suitable a¢ ne transformation
to u.
The weak substitutability condition (4), together with (1), imply that
1
u (a) = u (b) , P (a; b) = ) P (a; c) = P (b; c) , P (c; a) = P (c; b)
2
for any a; b; c 2 A. It follows that there exists a function : ! [0; 1] such that
P (a; b) = (u (a) ; u (b)) for any a; b 2 A. It su¢ ces to show that satis…es properties
(i)–(v) of Theorem 0.

28
 Property (i) is immediate from the de…nition of u. Property (ii) follows from (1).
Strong M -Independence and the M -linearity of u imply

(x; y) = (x0 ; y 0 ) ) (x z; y z) = (x0 z; y 0 z)

for any x; y; x0 ; y 0 2 and any z 2 u (M ). Since u (M ) = u (A) = , property (iii) holds.

Property (iv) follows from SST via the weak substitutability property (4).
Finally, to establish (v) note that inherits solvability: for any x; y; z 2 and any
q 2 [0; 1], if (z; x) q (z; y) then (z; w) = q for some w 2 . From solvability (of
) and weak monotonicity –property (iv) –we easily deduce that is continuous in its
second argument. Continuity in its …rst argument therefore follows from (ii).

B Strengthening Theorem D
The following theorem strengthens Theorem D.

Theorem D If A is the unit simplex in Rn and P satis…es SST, D2, D3 and Strong
Independence, then P has a strong expected utility representation.

Proof. The proof is the same as for Proposition 1, except that we can no longer appeal
to Lemma 1. Instead, we show that % possesses a linear representation by an alternative
route.
As per the proof of Proposition 1, we may assume that % is a weak order satisfying

a%b ) a c%b c (27)

for all a; b; c 2 A and all 2 (0; 1). Since the result is obvious if % is trivial (i.e., a b
for all a; b 2 A), let us assume otherwise. From D3, we also know that % satis…es the
following Archimedean property: for any a; b; c 2 A

a b c ) a c b a c (28)

for some ; 2 (0; 1). It therefore su¢ ces (Fishburn, 1982, Chapter 2) to establish that

a b ) a c b c (29)

for all a; b; c 2 A and all 2 (0; 1).

Lemma B.1 Condition (29) holds on the interior of A (that is, for any a; b; c 2 A\Rn++ ).

29
Proof. Suppose a; b; c 2 A \ Rn++ with a b and a c b c.26 That is,

1 1
P (a; b) > and P (a c; b c) = (30)
2 2
We claim that
1 1
P (a; b) P (d; e) ) P (d c; e c) = (31)
2 2
for any d; e 2 A. To see why, observe that Strong Independence and
1
P (d; e) = P (c; c)
2
give
1
.
P (d c; e c)
2
The reverse inequality follows by applying Strong Independence to P (a; b) P (d; e) and
using (30).
We next show that d = a b and e = b satisfy the antecedent in (31) for any 2 [0; 1].
The cases 2 f0; 1g are trivial so we focus on 2 (0; 1).
Since
1
P (a; b) > = P (b; b) = P (a; a)
2
we have
1
P (a b; b) (32)
2
and
1
P (a; a b) = P (a; b (1 ) a)
2
by Strong Independence. From the latter inequality and weak substitutability (SST) we
have
P (a; b) P (a b; b) (33)

Thus, from (31)–(33):

1
P ((a b) c; b c) = (34)
2
for any 2 [0; 1].
26
Note that b c a c is ruled out by (27) and a b.

30
 Figure 3

From (34) we obtain a linear segment of strictly positive length,27 parallel to the line
joining a and b, which forms part of an indi¤erence curve for %. Since a and b are in
the interior of the simplex, we can use the existence of this linear segment, together with
transitivity of % and the independence property (27), to deduce that the line segment
joining a and b must also be part of an indi¤erence curve. Figure 3 illustrates the required
construction (for the case n = 3). By moving point z along the segment from x to y we
sweep out an indi¤erence curve joining a to b.28 This is the desired contradiction, since
a b.

Since A \ Rn++ is a mixture set (under the usual mixing operation for Rn ), it follows
that % possesses a linear representation on A \ Rn++ . Let u be such a representation.
Observe that D3 is not required for the proof of Lemma B.1 – Strong Independence
does all the work. We now use D3 to extend the linear representation to the boundary of
the simplex.
27
Recall that a b, so a 6= b.
28
Let a^ = a c and ^b = b c. The points x and y are chosen such that a = x a
^ and b = y ^b for some
2 (0; 1). The independence property (27) implies that u a ^ b^ (x y) is constant in 2 [0; 1] for
any 2 [0; 1].

31
 First, u has a unique linear extension to A. We use u to denote the extended function,
as no confusion should arise. Let a be a boundary point of A. There are two possibilities:
either (i) u (a) = u (b) for some b 2 A \ Rn++ or else (ii) the face of the simplex containing
a is part of a contour of u.
For case (i) we show that a b. It follows that a b for any boundary point a and
any interior point on the same utility contour as a. If, for example, a b then we can
…nd some c 2 A \ Rn++ with u (c) < u (b) = u (a). Hence a b c, but u (a c) < u (b)
for all 2 (0; 1) by the linearity of u. Since a c 2 A \ Rn++ for any 2 (0; 1), we have
b a c for all 2 (0; 1), which contradicts the Archimedean property (28). Assuming
b a leads similarly to contradiction.
Now consider case (ii). Thus, u is constant on the face of the simplex containing a.
Suppose, in particular, that u (a) > u (b) for all b 2 A \ Rn++ . The alternative scenario,
in which u (a) < u (b) for all b 2 A \ Rn++ , may be handled similarly. We will show that
a b for any b on the same face as a, and a b for any b 2 A not on the same face as a.
Combining with case (i), this shows that % is represented by u on all of A.
Suppose b 2 A with u (a) = u (b), so b lies on the same face of the simplex as a. If
a b then a b c for any c 2 A \ Rn++ . Since b a c 2 A \ Rn++ for any 2 (0; 1),
we deduce a contradiction to (28). Assuming b a leads similarly to contradiction.
Finally, let b 2 A with u (a) > u (b). Assume, contrary to what we seek to show, that
b % a. Since u (a) > u (b) we may choose some c; d 2 A \ Rn++ with

u (a) > u (d) > u (c) > u (b) .

Therefore, d c b % a and

u (d a) = u (d) + (1 ) u (a) > u (d) > u (c)

It follows that d a c for all 2 (0; 1). This contradicts the Archimedean property (28).
This completes the proof of Theorem D .

We conjecture, but have not been able to prove, that Theorem D could be strength-
ened further, by dropping D3. This would yield a result that implies both Proposition 1
and Theorem D .

32