DIVISION OF THE HUMANITIES AND SOCIAL SCIENCES

CALIFORNIA INSTITUTE OF TECHNOLOGY

PASADENA, CALIFORNIA 91125
AMBIGUITY MADE PRECISE: A COMPARATIVE FOUNDATION
Paolo Ghirardato
California Institute of Technology
Massimo Marinacci
Universit`a di Torino
1891
C
A
L
I
F
O
R
N
I
A

I
N
S
T
I
T
UT E

O
F

T
E
C
H
N
O
L
O
G
Y
SOCIAL SCIENCE WORKING PAPER 1026
October 1997
Revised: December 2000
Ambiguity Made Precise: A Comparative
Foundation
Paolo Ghirardato Massimo Marinacci
Abstract
The theory of subjective expected utility (SEU) has been recently extended to allow
ambiguity to matter for choice. We propose a notion of absolute ambiguity aversion
by building on a notion of comparative ambiguity aversion. We characterize it for a
preference model which encompasses some of the most popular models in the literature.
We next build on these ideas to provide a denition of unambiguous act and event,
and show the characterization of the latter. As an illustration, we consider the classical
Ellsberg 3-color urn problem and nd that the notions developed in the paper provide
intuitive answers.
Journal of Economic Literature Classication Number: D81
JEL classication numbers: D81
Key words: Ambiguity Aversion, Cardinal Risk Aversion, Choquet Expected Utility,
Maxmin Expected Utility, Biseparable Preferences
Ambiguity Made Precise: A Comparative
Foundation

Paolo Ghirardato Massimo Marinacci

Introduction
In this paper we propose and characterize a formal denition of ambiguity aversion for
a class of preference models which encompasses the most popular models developed to
allow ambiguity attitude in decision making. Using this notion, we dene and characterize
ambiguity of events for ambiguity averse or loving preferences. Our analysis is based on
a fully subjective framework with no extraneous devices (like a roulette wheel, or a rich
set of exogenously unambiguous events). This yields a denition that can be fruitfully
used with any preference in the mentioned class, though it imposes a limitation in the
denitions ability of distinguishing real ambiguity aversion from other behavioral traits
that have been observed experimentally.
The subjective expected utility (SEU) theory of decision making under uncertainty of
Savage [25] is rmly established as the choice-theoretic underpinning of modern economic
theory. However, such success has well known costs: SEUs simple and powerful represen-
tation is often violated by actual behavior, and it imposes unwanted restrictions. In par-
ticular, Ellsbergs [7] famous thought experiment (see Section 5) convincingly shows that
SEU cannot take into account the possibility that the information a decision maker (DM)
has about some relevant uncertain event is vague or imprecise, and that such ambiguity
aects her behavior. Ellsberg observed that ambiguity aected his nonexperimental
subjects in a consistent fashion: Most of them preferred to bet on unambiguous rather
than ambiguous events. Furthermore, he found that even when shown the inconsistency
of their behavior with SEU, the subjects stood their ground because it seems to them
the sensible way to behave. This attitude has later been named ambiguity aversion,

An earlier version of this paper was circulated with the title Ambiguity Made Precise: A Com-
parative Foundation and Some Implications. We thank Kim Border, Eddie Dekel, Itzhak Gilboa, Tony
Kwasnica, Antonio Rangel, David Schmeidler, audiences at Caltech, Johns Hopkins, Northwestern, NYU,
Rochester, UC-Irvine, Universite Paris I, the TARK VII-Summer Micro Conference (Northwestern, July
1998), the 1999 RUD Workshop, and especially Simon Grant, Peter Klibano, Biung-Ghi Ju, Peter
Wakker, and an anonymous referee for helpful comments and discussion. Our greatest debt of gratitude
is however to Larry Epstein, who sparked our interest on this subject with his paper [8], and stimulated
it with many discussions. Marinacci gratefully acknowledges the nancial support of MURST.
and has received ample experimental conrmation.
1
Savage was well aware of this limit
of SEU, for he wrote that
[...] There seem to be some probability relations about which we feel rela-
tively sure as compared with others. [...] The notion of sure and unsure
introduced here is vague, and my complaint is precisely that neither the the-
ory of personal probability, as it is developed in this book, nor any other
device known to me renders the notion less vague. [25, pp. 5758 of the 1972
edition]
In the wake of Ellsbergs contribution, extensions of SEU have been developed al-
lowing ambiguity, and the DMs attitude towards it, to play a role in her choices. Two
methods for extending SEU have established themselves as the standards of this litera-
ture. The rst, originally proposed in Schmeidler [26], is to allow the DMs beliefs on
the state space to be represented by non-additive probabilities, called capacities, and her
preferences by Choquet integrals (which are just standard integrals when integrated with
respect to additive probabilities). For this reason, this generalization is called the theory
of Choquet expected utility (CEU) maximization. The second, axiomatized by Gilboa
and Schmeidler [15], allows the DMs beliefs to be represented by multiple probabilities,
and represents her preferences by the maximin on the set of the expected utilities. This
generalization is thus called the maxmin expected utility (MEU) theory. Here we use
the general class of preferences with ambiguity attitudes developed in our [13]. These
orderings, that we call biseparable preferences, are all those such that the ranking of con-
sequences can be represented by a state-independent cardinal utility u, and the ranking
of bets on events by u and a unique numerical function (a capacity) .
2
The latter repre-
sents the DMs willingness to bet; i.e., (A) is roughly the number of euros she is willing
to exchange for a bet that pays 1 euro if event A obtains and 0 euros otherwise. The
only restriction imposed on the ranking of non-binary acts is a mild dominance condition.
CEU and MEU are special cases of biseparable preferences, where is respectively the
DMs non-additive belief and the lower envelope of her multiple probabilities.
An important reason for the lasting success of SEU theory is the elegant theory of the
measurement of risk aversion developed from the seminal contributions of de Finetti [6],
Arrow [2] and Pratt [24]. Unlike risk aversion, ambiguity aversion is yet without a fully
general formalization, one that does not require extraneous devices and applies to most if
not all the existing models of ambiguity averse behavior. This paper attempts to ll this
gap: We propose a denition of ambiguity aversion and show its formal characterization
in the general decision-theoretic framework of Savage, whose only restriction is a richness
condition on the set of consequences. Our denition is behavioral ; that is, it only requires
observation of the DMs preferences on acts in this fully subjective setting. However, the
1
Other widespread names are uncertainty aversion and aversion to Knightian uncertainty. We
like to use uncertainty in its common meaning of any situation in which the consequences of the DMs
possible actions are not known at the time of choice.
2
A bet on an event is any binary act in which a better payo (win) is received when the event
obtains.
2
denition works as well (indeed better, see Proposition 11) in the Anscombe-Aumann
framework, a special case of Savages framework which presumes the existence of an
auxiliary device with known probabilities.
Decision models with ambiguity averse preferences the objects of increasing attention
by economists and political scientists interested in explaining phenomena at odds with
SEU. For example, they have been used to explain the existence of incomplete contracts
(Mukerji [22]), the existence of substantial volatility in stock markets (Epstein and Wang
[9], Hansen, Sargent and Tallarini [16]), or selective abstention in political elections (Ghi-
rardato and Katz [12]). We hope that the characterization provided here will turn out to
be useful for the applications of models of ambiguity aversion, as that of risk aversion
was for the applications of SEU. More concretely, we hope that it will help to understand
the predictive dierences of risk and ambiguity attitudes.
To understand our denition, it is helpful to go back to the characterization of risk
aversion in the SEU model. The following approach to dening risk aversion was inspired
by Yaari [30]. Given a state space S, let T denote a collection of acts, maps from S
into R (e.g., monetary payos). Dene a comparative notion of risk aversion for SEU
preferences as follows: Say that
2
is more risk averse than
1
if they have identical
beliefs and the following implications hold for every riskless (i.e., constant) act x and
every risky act f:
x
1
f x
2
f (1)
x ~
1
f x ~
2
f (2)
(where ~ is the asymmetric component of ). Identity of beliefs is required to avoid
possible confusions between dierences in risk attitudes and in beliefs (cf. Yaari [30,
p.317]). We can use this comparative ranking to obtain an absolute notion of risk aversion
by calling some DMs for instance expected value maximizers risk neutral, and by
then calling risk averse those DMs who are more risk averse than risk neutrals. As it is
well known, this comparatively founded notion has the usual characterization. Like the
traditional direct denition of risk aversion, it is fully behavioral in the sense dened
above. However, its interpretation is based on two primitive assumptions. First, constant
acts are intuitively riskless. Second, expected value maximization intuitively reects risk
neutral behavior, so that it can be used as our benchmark for measuring risk aversion.
In this paper, we follow the example of Epstein [8] in giving a comparative foundation
to ambiguity attitude: We start from a more ambiguity averse than... ranking and then
establish a benchmark, thus obtaining an absolute denition of ambiguity aversion.
Analogously to Yaaris, our more ambiguity averse... relation is based on the following
intuitive consideration: If a DM prefers an unambiguous (resp. ambiguous) act to an
ambiguous (resp. unambiguous) one, a more (resp. less) ambiguity averse one will do
the same. This is natural, but it raises the obvious question of which acts should be used
as the unambiguous acts for this ranking. Depending on the decision problem the DM
is facing and on her information, there might be dierent sets of obviously unambiguous
acts; i.e., acts that we are condent that any DM perceives as unambiguous. It seems
3
intuitive to us that in any well-formulated problem, the constant acts will be in this set.
Hence, we make our rst primitive assumption: Constant acts are the only acts that
are obviously unambiguous in any problem, since other acts may not be perceived as
unambiguous by some DM in some state of information. This assumption implies that a
preference (not necessarily SEU)
2
is more ambiguity averse than
1
whenever Eqs. (1)
and (2) hold. However, the following example casts some doubts as to the intuitive appeal
of such denition:
Example 1 Consider an (Ellsberg) urn containing balls of two colors: Black and Red.
Two DMs are facing this urn, and they have no information on its composition. The rst
DM has SEU preferences
1
, with a utility function on the set of consequences R given
by u
1
(x) = x, and beliefs on the state space of ball extractions S = B, R given by

1
(B) =
1
2
and
1
(R) =
1
2
.
The second DM also has SEU preferences, and identical beliefs: Her preference
2
is
represented by u
2
(x) =

x and
2
=
1
. Both (1) and (2) hold, but it is quite clear that
this is due to dierences in the DMs risk attitudes, and not in their ambiguity attitudes:
They both apparently disregard the ambiguity in their information.
Given a biseparable preference, call cardinal risk attitude the psychological trait described
by the utility function u what explains any dierences in the choices over bets of two
biseparable preferences with the same willingness to bet . The problem with the example
is that the two DMs have dierent cardinal risk attitude. To avoid confusions of this sort,
our comparative ambiguity ranking uses Eqs. (1) and (2) only on pairs which satisfy a
behavioral condition, called cardinal symmetry, that implies that two DMs have identical
u. As it only looks at each DMs preferences over bets on one event (which may be
dierent across DMs), cardinal symmetry does not impose any restriction on the DMs
relative ambiguity attitudes.
Having thus constructed the comparative ambiguity ranking, we next choose a bench-
mark against which to measure ambiguity aversion. It seems generally agreed that SEU
preferences are intuitively ambiguity neutral. We use SEU preferences as benchmarks
because we posit our second primitive assumption that they are the only ones that
are obviously ambiguity neutral in any decision problem and in any situation. Thus,
ambiguity averse is any preference relation for which there is a SEU preference less
ambiguity averse than . Ambiguity love and (endogenous) neutrality are dened in
the obvious way.
The main results in the paper present the characterization of these notions of ambi-
guity attitude for biseparable preferences. The characterization of ambiguity neutrality
is simply stated: A preference is ambiguity neutral if and only if it has a SEU represen-
tation. That is, the only preferences which are endogenously ambiguity neutral are SEU.
The general characterization of ambiguity aversion (resp. love) implies in particular that
a preference is ambiguity averse (resp. loving) only if its willingness to bet is pointwise
4
dominated by (resp. pointwise dominates) a probability. In the CEU case, the converse
is also true: A CEU preference is ambiguity averse if and only if its belief (which is equal
to ) is dominated by a probability; i.e., it has a non-empty core. On the other hand,
all MEU preferences are ambiguity averse, as it is intuitive. As to comparative ambiguity
aversion, we nd that if
2
is more ambiguity averse than
1
then
1

2
. That is, a
less ambiguity averse DM will have uniformly higher willingness to bet. The latter con-
dition is also sucient for CEU preferences, whereas for MEU preferences containment
of the sets of probabilities is necessary and sucient for relative ambiguity.
We next briey turn to the issue of dening ambiguity itself. A behavioral notion
of unambiguous act follows naturally from our earlier analysis: Say that an act is un-
ambiguous if an ambiguity averse (or loving) DM evaluates it in an ambiguity neutral
fashion. The unambiguous events are those that unambiguous acts depend upon. We
obtain the following simple characterization of the set of unambiguous events for bisep-
arable preferences: For an ambiguity averse (or loving) DM with willingness to bet ,
event A is unambiguous if and only if (A) +(A
c
) = 1. (A more extensive discussion of
ambiguity is contained in the companion [14].)
Finally, as an application of the previous analysis, we consider the classical Ellsberg
problem with a 3-color urn. We show that the theory delivers the intuitive answers, once
the information provided to the DM is correctly incorporated.
It is important to underscore from the outset two important limitations of the notions
of ambiguity attitude we propose. The rst limitation is that while the comparative
foundation makes our absolute notion notion behavioral, in the sense dened above,
it also makes it computationally demanding. A more satisfactory denition would be
one which is more direct: It can be veried by observing a smaller subset of the DMs
preference relation. While we conjecture that it may be possible to construct such a
denition obtaining the same characterization as the one proposed here we leave
its development to future work.
Our comparative notion is more direct, thus less amenable to this criticism. However,
it is in turn limited by the requirement of the identity of cardinal risk attitude. The
absolute notion is not, as it conceptually builds on the comparison of the DM with an
idealized version of herself, identical to her in all traits but her ambiguity aversion.
The second limitation stems from the fact that no extraneous devices are used in this
paper. An advantage of this is that our notions apply to any decision problem under
uncertainty, and our results to any biseparable preference. However, such wide scope
carries costs: Our notion of ambiguity aversion comprises behavioral traits that may
not be due to ambiguity like probabilistic risk aversion, the tendency of discounting
objective probabilities that has been observed in many experiments on decision making
under risk (including the celebrated Allais paradox). Thus, one may consider it more
appropriate to use a dierent name for what is measured here, like chance aversion or
extended ambiguity aversion.
5
The reason for our choice of terminology is that we see a ranking of conceptual impor-
tance between ambiguity aversion/love and other departures from SEU maximization. As
we argued above using Savages words, the presence of ambiguity provides a normatively
compelling reason for violating SEU. We do not feel that other documented reasons are
similarly compelling. Moreover, we hold (see below and Subsection 6.3) that extraneous
devices say, a rich set of exogenously unambiguous events are required for ascer-
taining the reason of a given departure. Thus, when these devices are not available
say, because the set of unambiguous events is not rich enough we prefer to attribute
a departure to the reasons we nd normatively more compelling. However, the reader
is warned, so that he/she may choose to give a dierent name to the phenomenon we
formally describe.
The Related Literature
The problem of dening ambiguity and ambiguity aversion is discussed in a number of
earlier papers. The closest to ours in spirit and generality is Epstein [8], the rst paper to
develop a notion of absolute ambiguity aversion from a comparative foundation.
3
As we
discuss in more detail in Subsection 6.3, the comparative notion and benchmarks he uses
are dierent from ours. Epsteins objective is to provide a more precise measurement of
ambiguity attitude than the one we attempt here; in particular, to lter out probabilistic
risk aversion. For this reason, he assumes that in the absence of ambiguity a DMs pref-
erences are probabilistically sophisticated in the sense of Machina and Schmeidler [20].
However, we argue that for its conclusions to conform with intuition, Epsteins approach
requires an extraneous device: a rich set of acts which are exogenously established to be
unambiguous, much larger than the set of the constants that we use. Thus, the higher
accuracy of his approach limits its applicability vis ` a vis our cruder but less demanding
approach.
The most widely known and accepted denition of absolute ambiguity aversion is
that proposed by Schmeidler in his seminal CEU model [26]. Employing an Anscombe-
Aumann framework, he denes ambiguity aversion as the preference for objective mix-
tures of acts, and he shows that for CEU preferences this notion is characterized by the
convexity of the capacity representing the DMs beliefs. While the intuition behind this
denition is certainly compelling, Schmeidlers axiom captures more than our notion of
ambiguity aversion. It gives rise to ambiguity averse behavior, but it entails additional
structure that does not seem to be related to ambiguity aversion (see Example 25).
Doubts about the relation of convexity to ambiguity aversion in the CEU case are also
raised by Epstein [8], but he concludes that they are completely unrelated (see Section 5
for a discussion).
There are other interesting papers dealing with ambiguity and ambiguity aversion. In
a nite setting, Kelsey and Nandeibam [18] propose a notion of comparative ambiguity for
3
There are earlier papers that use a comparative approach for studying ambiguity attitude, but they
do not use it as a basis for dening absolute notions. E.g., Tversky and Wakker [27].
6
the CEU and MEU models similar to ours and obtain a similar characterization, as well as
an additional characterization in the CEU case. Unlike us, they do not consider absolute
ambiguity attitude, and they do not discuss the issue of the distinction of cardinal risk
and ambiguity attitude. Montesano and Giovannoni [21] notice a connection between
absolute ambiguity aversion in the CEU model and nonemptiness of the core, but they
base themselves purely on intuitive considerations on Ellsbergs example. Chateauneuf
and Tallon [4] present an intuitive necessary and sucient condition for non-emptiness of
the core of CEU preferences in an Anscombe-Aumann framework. Zhang [31], Nehring
[23], and Epstein and Zhang [10] propose dierent denitions of unambiguous event and
act. Fishburn [11] characterizes axiomatically a primitive notion of ambiguity.
Organization
The structure of the paper is as follows. Section 1 provides the necessary denitions and
set-up. Section 2 introduces the notions of ambiguity aversion. The cardinal symmetry
condition is introduced in Subsection 2.1, and the comparative and absolute denitions
in 2.2. Section 3 presents the characterization results. Section 4 contains the notions
of unambiguous act and event, and the characterization of the latter. In Section 5, we
go back to the Ellsberg urn and show the implications of our results for that example.
Section 6 discusses the key aspects of our approach, in particular, the choices of the
comparative ambiguity ranking and the benchmark for dening ambiguity neutrality; it
thus provides a more detailed comparison with Epsteins [8] approach. The Appendices
contain the proofs and some technical material.
1 Set-Up and Preliminaries
The general set-up of Savage [25] is the following. There is a set S of states of the world,
an algebra of subsets of S, and a set X of consequences. The choice set T is the set
of all nite-valued acts f : S X which are measurable w.r.t. . With the customary
abuse of notation, for x X we dene x T to be the constant act x(s) = x for all
s S, so that X T. Given A , we denote by x Ay the binary act (bet) f T
such that f(s) = x for s A, and f(s) = y for s / A.
Our denitions require that the DMs preferences be represented by a weak order
on T: a complete and transitive binary relation , with asymmetric (resp. symmetric)
component ~ (resp. ). The weak order is called nontrivial if there are f, g T
such that f ~ g. We henceforth call preference relation any nontrivial weak order on
T.
A functional V : T R is a representation of if for every f, g T, f g if and
only if V (f) V (g). A representation V is called: monotonic if f(s) g(s) for every
s S implies V (f) V (g); nontrivial if V (f) > V (g) for some f, g T.
7
While the denitions apply to any preference relation, our results require a little more
structure, provided by a general decision model introduced in Ghirardato and Marinacci
[13]. To present it, we need the following notion of nontrivial event: Given a preference
relation , A is essential for if for some x, y X, we have x ~ xAy ~ y.
Denition 2 Let be a binary relation. We say that a representation V : T R of
is canonical if it is nontrivial monotonic and there exists a set-function : [0, 1]
such that, letting u(x) V (x) for all x X, for all consequences x y and all events
A,
V (xAy) = u(x) (A) +u(y) (1 (A)). (3)
A relation is called a biseparable preference if it admits a canonical representation,
and moreover such representation is unique up to a positive ane transformation when
has at least one essential event.
Clearly, a biseparable preference is a preference relation. If V is a canonical representation
of , then u is a cardinal state-independent representation of the DMs preferences over
consequences, hence we call it his canonical utility index. Moreover, for all x ~ y
and all events A, B we have xAy xBy if and only if (A) (B). Thus,
represents the DMs willingness to bet (likelihood relation) on events. Moreover, is
easily shown to be a capacity a set-function normalized and monotonic w.r.t. set
inclusion so that V evaluates binary acts by taking the Choquet expectation of u with
respect to .
4
However, the DMs preferences over non-binary acts are not constrained
to a specic functional form.
To understand the rationale of the clause relating to essential events, rst observe
that for any with a canonical representation with willingness to bet , an event A is
essential if and only if 0 < (A) < 1. Thus, there are no essential events i (A) is either
0 or 1 for every A; that is, the DM behaves as if he does not judge any bet to be uncertain,
and his canonical utility index is ordinal. In such a case, the DMs cardinal risk attitude
is then intuitively not dened: without an uncertain event there is no risk. On the other
hand, it can be shown [13, Theorem 4] that cardinal risk attitude is characterized by
a cardinal property of the canonical utility index, its concavity. Hence the additional
requirement in Denition 2 guarantees that when there is some uncertain event cardinal
risk aversion is well dened.
As the dierences in two DMs cardinal risk attitude might play a role in the choices
in Eqs. (1) and (2), it is useful to identify the situation in which these attitudes are
dened: Say that preference relations
1
and
2
have essential events if there are
events A
1
, A
2
such that for each i = 1, 2, A
i
is essential for
i
.
To avoid repetitions, the following lists all the assumptions on the structure of the
decision problem and on the DMs preferences that are tacitly assumed in all results in
the paper:
4
See Appendix A for the denition of capacities, Choquet integrals, and some of their properties.
8
Structural Assumption X is a connected and separable topological space (e.g., a con-
vex subset of R
n
with the usual topology). Every biseparable preference on T has a
continuous canonical utility function.
A full axiomatic characterization of the biseparable preferences satisfying the Structural
Assumption is provided in our [13].
1.1 Some Examples of Biseparable Preferences
As mentioned above, the biseparable preference model is very general. In fact, it contains
most of the known preference models that obtain a separation between cardinal (state-
independent) utility and willingness to bet. We now illustrate this claim by showing some
examples of decision models which under mild additional restrictions (e.g., the Structural
Assumption) belong to the biseparable class. (More examples and details are found in
[13].)
(i) A binary relation on T is a CEU ordering if there exist a cardinal utility index
u on X and a capacity on (S, ) such that can be represented by the functional
V : T R dened by the following equation:
V (f) =

S
u(f()) d, (4)
where the integral is taken in the sense of Choquet (notice that it is nite because
each act in T is nite-valued). The functional V is immediately seen to be a
canonical representation of , and = is its willingness to bet. An important
subclass of CEU orderings are the SEU orderings, which correspond to the special
case in which is a probability measure, i.e., a nitely additive capacity. See
Wakker [28] for an axiomatization of CEU and SEU preferences (satisfying the
Structural Assumption) in the Savage setting.
(ii) Let denote the set of all the probability measures on (S, ). A binary relation
on T is a MEU ordering if there exist a cardinal utility index u and a unique
non-empty, (weak

)-compact and convex set C such that can be represented

by the functional V : T R dened by the following equation:
V (f) = min
PC

S
u(f(s)) P(ds). (5)
SEU also corresponds to the special case of MEU in which C = P for some
probability measure P. If we now let for any A ,
P(A) = min
PC
P(A), (6)
we see that P is an exact capacity. While in general V (f) is not equal to the
Choquet integral of u(f) with respect to P, this is the case for binary acts f. This
9
shows that V is a canonical representation of , with willingness to bet = P. See
Casadesus-Masanell, Klibano and Ozdenoren [3] for an axiomatization of MEU
preferences (satisfying the Structural Assumption) in the Savage setting.
More generally, consider an -MEU preference which assigns some weight to both
the worst-case and best-case scenarios. Formally, there is a cardinal utility u, a set
or probabilities C, and [0, 1], such that is represented by
V (f) =

min
PC

S
u(f(s)) P(ds) + (1 ) max
PC

S
u(f(s)) P(ds)

.
This includes the case of a maximax DM, who has 0. V is canonical, so that
is biseparable, with given by (A) = min
PC
P(A) + (1 ) max
PC
P(A),
for A .
(iii) Consider a binary relation constructed as follows: There is a cardinal utility u,
a probability P and a number [0, 1] such that is represented by
V (f) (1 )

S
u(f(s)) P(ds) + (u f),
where
(u f) sup

S
u(g(s)) P(ds) : g T binary, u(g(s)) u(f(s)) for all s S

.
describes a DM who behaves as if he was maximizing SEU when choosing among
binary acts, but not when comparing more complex acts. The higher the parameter
, the farther the preference relation is from SEU on non-binary acts. V is mono-
tonic and it satises Eq. (3) with = P, so that it is a canonical representation of
.
1.2 The Anscombe-Aumann Case
The Anscombe-Aumann framework is a widely used special case of our framework in
which the consequences have an objective feature: X is also a convex subset of a vector
space. For instance, X is the set of all the lotteries on a set of prizes if the DM has
access to an objective independent randomizing device. In this framework, it is natural
to consider the following variant of the biseparable preference model where for every
f, g T and [0, 1], f +(1)g denotes the act which pays f(s)+(1)g(s) X
for every s S.
Denition 3 A canonical representation V of a preference relation is constant lin-
ear (c-linear for short) if V (f +(1)x) = V (f)+(1)V (x) for all binary f T,
x X, and [0, 1]. A relation is called a c-linearly biseparable preference if it
admits a c-linear canonical representation.
10
Again, an axiomatic characterization of this model is found in [13]. It generalizes the SEU
model of Anscombe and Aumann [1] and many non-EU extensions that followed, like the
CEU and MEU models of Schmeidler [26] and Gilboa and Schmeidler [15] respectively.
In fact, a c-linearly biseparable preference behaves in a SEU fashion over the set X
of the constant acts, but it is almost unconstrained over non-binary acts. (C-linearity
guarantees the cardinality of V and hence u.)
All the results in this paper are immediately translated to this class of preferences, in
particular to the CEU and MEU models in the Anscombe-Aumann framework mentioned
above. Indeed, as we show in Proposition 11 below, in this case removing cardinal risk
aversion is much easier than in the more general framework we use.
2 The Denitions
As anticipated in the Introduction, the point of departure of our search for an extended
notion of ambiguity aversion is the following partial order on preference relations:
Denition 4 Let
1
and
2
be two preference relations. We say that
2
is more
uncertainty averse than
1
if: For all x X and f T, both
x
1
f x
2
f (7)
and
x ~
1
f x ~
2
f. (8)
This order has the advantage of making the weakest prejudgment on which acts are
intuitively unambiguous: The constants. However, Example 1 illustrates that it does
not discriminate between cardinal risk attitude and ambiguity attitude: DMs 1 and 2
are intuitively both ambiguity neutral, but 1 is more cardinal risk averse, and hence more
uncertainty averse than 2. The problem is that constant acts are neutral with respect
to ambiguity and with respect to cardinal risk. Given that our objective is comparing
ambiguity attitudes, we thus need to nd ways to coarsen the ranking above, so as to
identify which part is due to dierences in cardinal risk attitude and which is due to
dierences in ambiguity attitude.
2.1 Filtering Cardinal Risk Attitude
While the factorization just described can be achieved easily if we impose more struc-
ture on the decision framework (see, e.g., the discussion in Subsection 6.3), we present a
method for separating cardinal risk and ambiguity attitude which is only based on pref-
erences, does not employ extraneous devices, and obtains the result for all biseparable
preferences. Moreover, this approach does not impose any restrictions on the two DMs
11
beliefs (and hence on their relative ambiguity attitude), a problem that all alternatives
share. The key step is coarsening comparative uncertainty aversion by adding the fol-
lowing restriction on which pairs of preferences are to be compared (we write x, y ~ z
as a short-hand for x ~ z and y ~ z, and similarly for ):
Denition 5 Two preference relations
1
and
2
are cardinally symmetric if for
any pair (A
1
, A
2
) such that each A
i
is essential for
i
, i = 1, 2, and any
v

, v

, w

, w

X such that v

1
v

and w

2
w

we have:
If there are x, y X such that v

~
1
x, y, w

~
2
x, y, and
v

A
1
x
1
v

A
1
y and w

A
2
x
2
w

A
2
y, (9)
then for every x

, y

X such that v

~
1
x

, y

, w

~
2
x

, y

we have
v

A
1
x

1
v

A
1
y

A
2
x

2
w

A
2
y

. (10)
Symmetrically, if there are x, y X such that v

1
x, y, w

2
x, y, and
xA
1
v

1
y A
1
v

and xA
2
w

2
y A
2
w

, (11)
then for every x

, y

X such that v

1
x

, y

, w

2
x

, y

we have
x

A
1
v

1
y

A
1
v

A
2
w

2
y

A
2
w

. (12)
This condition is inspired by the utility construction technique used in the axiomatiza-
tions of additive conjoint measurement in, e.g., Krantz et al. [19] and Wakker [28].
A few remarks are in order: First, cardinal symmetry holds vacuously for any pair
of preferences which do not have essential events. Second, cardinal symmetry does not
impose restrictions on the DMs relative ambiguity attitudes. In fact, for all acts ranked
by
i
, the consequence obtained if A
i
is always strictly better than that obtained if A
c
i
,
so that all acts are bets on the same event A
i
. Intuitively, a DMs ambiguity attitude
aects these bets symmetrically, so that his preferences do not convey any information
about it. Moreover, cardinal symmetry does not constrain the DMs relative condence
on A
1
and A
2
, since the win (or loss) payos can be dierent for the two DMs.
On the other hand, it does unsurprisingly restrict their relative cardinal risk attitudes.
To better understand the relative restrictions implied by cardinal symmetry, assume that
consequences are monetary payos and that both DMs like more money to less. Suppose
that, when betting on events (A
1
, A
2
), (9) holds for some loss payos x and y and win
payos v

~
1
v

and w

~
2
w

respectively. This says that exchanging v

for v

as the
prize for A
1
, and w

for w

as the prize for A

2
, can for both DMs be traded o with a
reduction in loss from x to y. Suppose that when the initial loss is x

< x,
1
is willing
to trade o the increase in win with a reduction in loss to y

, but
2
accepts reducing
loss only to y

> y

(that is, w

A
2
x

~
2
w

A
2
y

, in violation of (10)). That is, as the

12
amount of the low payo decreases,
2
becomes more sensible to dierences in payos
than
1
. Such diversity of behavior that we intuitively attribute to dierences in the
DMs risk attitude is ruled out by cardinal symmetry, which requires that the two
DMs consistently agree on the acceptable tradeo for improving their win payo, and
similarly for the loss payo. It is important to stress that this discussion makes sense
only when both DMs are faced with nontrivial uncertainty (i.e., they are both betting on
essential events). Thus, we do not use trade-o to mean certain substitution; rather,
substitution in the context of an uncertain prospect.
To see how cardinal symmetry is used to show that two biseparable preferences have
the same cardinal risk attitude, assume rst that the two relations are ordinally equiv-
alent: for every x, y X, x
1
y x
2
y. When that is the case, cardinal symmetry
holds if and only if their canonical utility indices are positive ane transformations of
each other. In order to simplify the statements, we write u
1
u
2
to denote such equality
of indices.
Proposition 6 Suppose that
1
and
2
are ordinally equivalent biseparable preferences
which have essential events. Then
1
and
2
are cardinally symmetric if and only if
their canonical utility indices satisfy u
1
u
2
.
The intuition of the proof (see Appendix B) can be quickly grasped by rewriting, say,
Eqs. (9) and (10) in terms of the canonical representations to nd that for every x, y, x

, y

X,
u
1
(x) u
1
(y) = u
1
(x

) u
1
(y

) u
2
(x) u
2
(y) = u
2
(x

) u
2
(y

).
Notice however that this does not imply that the preferences are identical on binary acts:
The DMs beliefs on events could be totally dierent.
The comparative notion of ambiguity aversion we propose in the next subsection
checks comparative uncertainty aversion in preferences with the same cardinal risk atti-
tude. Clearly, it would be nicer to have a comparative notion that ranks also preferences
without the same cardinal risk attitude. In Subsection 6.1, we discuss how to extend our
notion to deal with these cases. This extension requires the exact measurement of the
two preferences canonical utility indices, and is thus less behavioral than the one we
just anticipated.
Finally, we remark that a symmetric exercise to that performed here is to coarsen
comparative uncertainty aversion so as to rank preferences by their cardinal risk aversion
only. In [13] it is shown that for biseparable preferences such ranking is represented
by the ordering of canonical utilities by their relative concavity, thus generalizing the
standard result.
13
2.2 Comparative and Absolute Ambiguity Aversion
Having thus prepared the ground, our comparative notion of ambiguity is immediately
stated:
Denition 7 Let
1
and
2
be two preference relations. We say that
2
is more
ambiguity averse than
1
whenever both the following conditions hold:
(A)
2
is more uncertainty averse than
1
;
(B)
1
and
2
are cardinally symmetric.
Thus, we restrict our attention to pairs which are cardinally symmetric. As explained
earlier, when one DMs preference does not have an essential event, cardinal risk aversion
does not play a role in that DMs choices, so that we do not need to remove it from the
picture.
Remark 8 So far, we have tacitly assumed that cardinal risk and ambiguity attitude
completely characterize biseparable preferences. Indeed, the validity of this can be easily
veried by observing that if two such preferences are as uncertainty averse as each other
(that is,
1
is more uncertainty averse than
2
, and vice versa), they are identical.
We nally come to the absolute denition of ambiguity aversion and love. Let be a
preference relation on T with a SEU representation.
5
As we observed in the Introduction,
these relations intuitively embody ambiguity neutrality. We propose to use them as the
benchmark for dening ambiguity aversion. Of course, one could intuitively hold that
the SEU ones are not the only relations embodying ambiguity neutrality, and thus prefer
using a wider set of benchmarks. This alternative route is discussed in Subsection 6.3
below.
Denition 9 A preference relation is ambiguity averse (loving) if there exists a
SEU preference relation which is less (more) ambiguity averse than . It is ambiguity
neutral if it is both ambiguity averse and ambiguity loving.
If is a SEU preference which is less ambiguity averse than , we call it a benchmark
preference for . We denote by 1() the set of all benchmark preferences for . That
is,
1() T T : is SEU and is more ambiguity averse than .
Each benchmark preference 1() induces a probability measure P on , so a natural
twin of 1() is the set of the benchmark measures:
/() = P : P represents , for 1().
Using this notation, Denition 9 can be rewritten as follows: is ambiguity averse if
either 1() ,= , or /() ,= .
5
We use the symbols (and >) to denote SEU weak (and strict) preferences.
14
3 The Characterizations
We now characterize the notions of comparative and absolute ambiguity aversion dened
in the previous section for the general case of biseparable preferences, and the important
subcases of CEU and MEU preferences. To start, we use Proposition 6 and the observa-
tion that the canonical utility index of a preference with no essential events is ordinal,
to show that if two preferences are biseparable and they are ranked by Denition 7, they
have the same canonical utility index:
Theorem 10 Suppose that
1
and
2
are biseparable preferences, and that
2
is more
ambiguity averse than
1
. Then u
1
u
2
.
Checking cardinal symmetry is clearly not a trivial task, but for an important subclass
of preference relations the c-linearly biseparable preferences in an Anscombe-Aumann
setting it is implied by comparative uncertainty aversion. In fact, under c-linearity,
ordinal equivalence easily implies cardinal symmetry, so that we get:
Proposition 11 Suppose that X is a convex subset of a vector space, and that
1
and

2
are c-linearly biseparable preferences.
2
is more ambiguity averse than
1
if and
only if
2
is more uncertainty averse than
1
.
Therefore, in this case Denition 4 can be directly used as our denition of comparative
ambiguity attitude.
3.1 Absolute Ambiguity Aversion
We rst characterize absolute ambiguity aversion for a general biseparable preference .
Suppose that V is a canonical representation of , with canonical utility u. We let
T()

P :

S
u(f(s)) P(ds) V (f) for all f T

.
That is, T(), which depends only on V , is the set of beliefs inducing preferences which
assign (weakly) higher expected utility to every act f. These preferences exhaust the set
of the benchmarks of :
Theorem 12 Let be a biseparable preference. Then, /() = T(). In particular,
is ambiguity averse if and only if T() ,= .
Let be the capacity associated with the canonical representation V . It is immediate to
see that if P T(), then P . Thus, non-emptiness of the core of (the set of the
probabilities that dominate pointwise, that we denote (()) is necessary for to be
ambiguity averse. In Subsection 3.2 it is shown to be not sucient in general.
15
Turn now to the characterization of ambiguity aversion for the popular CEU and MEU
models. Suppose rst that is a CEU preference relation represented by the capacity
, and let (() denote s possibly empty core. It is shown that T() = ((), so that
the following result which also provides a novel decision-theoretic interpretation of the
core as the set of all the benchmark measures follows as a corollary of Theorem 12.
Corollary 13 Suppose that is a CEU preference relation, represented by capacity .
Then (() = /(). In particular, is ambiguity averse if and only if (() ,= .
Thus, the core of an ambiguity averse capacity is equal to the set of its benchmark
measures, and the ambiguity averse capacities are those with a non-empty core, called
balanced. A classical result (see, e.g., Kannai [17]) thus provides an internal characteri-
zation of ambiguity aversion in the CEU case: Letting 1
A
denote the characteristic func-
tion of A , a capacity reects ambiguity aversion if and only if for all
1
, . . . ,
n
0
and all A
1
, . . . , A
n
such that

n
i=1

i
1
A
i
1
S
, we have

n
i=1

i
(A
i
) 1. As convex
capacities are balanced, but not conversely, the corollary motivates our claim that con-
vexity does not characterize our notion of ambiguity aversion. This point is illustrated by
Example 25 below, which presents a capacity that intuitively reects ambiguity aversion
but is not convex.
On the other hand, given a MEU preference relation with set of priors C, it is
shown that T() = C. Thus, Theorem 12 implies that any MEU preference is ambiguity
averse (as it is intuitive) and, more interestingly, that the set C can be interpreted as the
set of the benchmark measures for .
Corollary 14 Suppose that is a MEU preference relation, represented by the set of
probabilities C. Then C = /(), so that is ambiguity averse.
As to ambiguity love, reversing the proof of Theorem 12 shows that for any biseparable
preference, ambiguity love is characterized by nonemptiness of the set
c()

P :

S
u(f(s)) P(ds) V (f) for all f T

.
In particular, a CEU preference with capacity is ambiguity loving if and only if the
set of probabilities dominated by is non-empty. As for MEU preferences: None is
ambiguity loving. Conversely, any maximax EU preference is ambiguity loving, with
c() = C.
Finally, we look at ambiguity neutrality. Since we started with an informal intuition
of SEU preferences as reecting neutrality to ambiguity, an important consistency check
on our analysis is to verify that they are ambiguity neutral in the formal sense. This is
the case:
Proposition 15 Let be a biseparable preference. Then is ambiguity neutral if and
only if it is a SEU preference relation.
16
3.2 Comparative Ambiguity Aversion
We conclude the section with the characterization of comparative ambiguity aversion.
The general result on comparative ambiguity, an immediate consequence of Theorem 12,
is stated as follows (where
1
and
2
represent the willingness to bet of
1
and
2
respectively):
Proposition 16 Let
1
and
2
be two biseparable preferences. If
2
is more ambiguity
averse than
1
, then
1

2
, T(
1
) T(
2
), c(
1
) c(
2
) and u
1
u
2
.
Thus, relative ambiguity implies containment of the sets T() and c() (clearly in
opposite directions), and dominance of the willingness to bet . Of course, the propo-
sition lacks a converse, and thus it does not oer a full characterization. As we argue
below, biseparable preferences seem to have too little structure for obtaining a general
characterization result.
Things are dierent if we restrict our attention to specic models. For instance, the
next result characterizes comparative ambiguity for the CEU and MEU models:
Theorem 17 Let
1
and
2
be biseparable preferences, with canonical utilities u
1
and
u
2
respectively.
(i) Suppose that
1
and
2
are CEU, with respective capacities
1
and
2
. Then
2
is
more ambiguity averse than
1
if and only if
1

2
and u
1
u
2
.
(ii) Suppose that
1
is MEU, with set of probabilities C
1
. Then
2
is more ambiguity
averse than
1
if and only if C
1
= T(
1
) T(
2
) and u
1
u
2
.
Observe that part (ii) of the proposition does more than characterize comparative am-
biguity for MEU preferences, as it applies to any biseparable
2
. For instance, it is
immediate to notice that one can characterize absolute ambiguity aversion using that
result and the fact that if
1
is a SEU preference relation with beliefs P, then C
1
= P.
Also, a symmetric result to (ii) holds: If
2
is maximax EU, it is more ambiguity averse
than
1
i C
2
= c(
2
) c(
1
).
Remark 18 Proposition 17 can be used to explain the apparent incongruence of the
characterization of comparative risk aversion in SEU (in the sense of Yaari [30]) and
of comparative ambiguity aversion in CEU: Convexity of seems to be the natural
counterpart of concavity of u, but it is not. This is due to the dierent uniqueness
properties of utility functions and capacities. A SEU
2
is more risk averse than a SEU

1
i for every common normalization of the utilities, we have u
2
(x) u(x) inside the
interval of normalization. Since any normalization is allowed, u
2
must then be a concave
transformation of u
1
. In the case of capacities only one normalization is allowed, so we
only have
1

2
.
17
It is not dicult to show that the necessary conditions of Proposition 16 are not suf-
cient if taken one by one. For instance, there are pairs of MEU (resp. CEU) preferences

1
and
2
such that
1

2
(resp. ((
1
) = T(
1
) T(
2
) = ((
2
)) does not entail
that
2
is more ambiguity averse than
1
.
Example 19 Let S = s
1
, s
2
, s
3
, the power set of S. Consider the probabilities P,
Q and R dened by P = [1/2, 0, 1/2], Q = [0, 1, 0] and R = [1/2, 1/2, 0]. Let C
1
and C
2
respectively be the closed convex hull of P, Q and P, Q, R. Then,
1
= P
1
= P
2
=
2
,
but C
2
, C
1
, and indeed by Prop. 17 the MEU preference
2
inducing C
2
is more
ambiguity averse than the MEU preference
1
inducing C
1
.
Consider next a capacity such that (A) = 1/3 for any A ,= , S, and a probability
P

equal to 1/3 on each singleton. Then (() = P

, so that is balanced, but not

exact (for instance, P

(s
1
, s
2
) = 2/3 > 1/3 = (s
1
, s
2
)). We have (() ((P

) but
, P

, and by Prop. 17 the CEU preference inducing is not more ambiguity averse
than that inducing P

. In contrast, P

is exact, and we have both ((P

) (() and
P

.
These examples illustrate two conceptual observations. The rst (anticipated in Sub-
section 3.1) is that non-emptiness of the core of is not sucient for absolute ambiguity
aversion: A probability can dominate without being a benchmark measure for . Un-
surprisingly, in general the capacity does not completely describe the DMs ambiguity
attitude. The second observation is that, while T() does characterize the DMs ab-
solute ambiguity aversion, it is also an incomplete description of the DMs ambiguity
attitude: There can be preferences
1
and
2
strictly ranked by comparative ambiguity
even though T(
1
) = T(
2
).
To better appreciate the diculty of obtaining a general suciency result for bisepa-
rable preferences, we now present an example in which all the necessary conditions hold
but the comparative ranking does not obtain.
Example 20 For a general S and (but see the restriction on P below), consider two
preference relations
1
and
2
which behave according to example (iii) of biseparable
preference in Section 1. Both have identical P and u (which ranges in a nondegenerate
interval of R), with the following restriction on P: There are at least three disjoint events
in , A
1
, A
2
and A
3
such that P(A
i
) > 0 for i = 1, 2, 3 (otherwise both preferences are
indistinguishable from SEU preferences with utility u and beliefs P). Their parameters
are dierent, in particular
2
>
1
> 0. Clearly
1
=
2
= P and u
1
u
2
. It is also
immediate to verify that, under the assumption on P, T(
1
) = T(
2
) = P and
c(
1
) = c(
2
) = , so that both preferences are (strictly) ambiguity averse. However,

1
is not more ambiguity averse than
2
(nor are
1
and
2
equal, which would follow
from two applications of the converse). Indeed, the parameter measures comparative
ambiguity for these preferences, so that
2
is more ambiguity averse than
1
.
18
4 Unambiguous Acts and Events
Let be an ambiguity averse or loving preference relation. Even though the preference
relation has a strict ambiguity attitude, it may nevertheless behave in an ambiguity neu-
tral fashion with respect to some subclass of acts and events, that we may like to consider
unambiguous. The purpose of this section is to identify the class of the unambiguous
acts and the related class of unambiguous events, and to present a characterization of the
latter for biseparable (in particular CEU and MEU) preference relations. We henceforth
focus on ambiguity averse preference relations, but it is easy to see that all the results in
this section can be shown for ambiguity loving preferences. A more extensive discussion
of the behavioral denition of ambiguity for events and acts is found in our [14].
In view of our results so far, the natural approach in dening the class of unambiguous
events of a preference relation is to x a a benchmark 1(), and to consider the
subset of all the acts in T over which is as ambiguity averse as . Intuitively, ambiguity
is a property that the DM attaches to partitions of events, so that nonconstant acts which
generate the same partition should be consistently deemed either both ambiguous or both
unambiguous. Hence, we consider as truly unambiguous only the acts which belong to
the set dened below.
Denition 21 Given a preference relation and 1(), the set of -unambiguous
acts, denoted H

, is the largest subset of T satisfying the following two conditions:

6
(A) For every x X and every f H

, and agree on the ranking of f and x.

(B) For every f H

and every g T, if g
1
(x) : x X f
1
(x) : x X,
then g H

.
Given a preference relation , for any f T denote by
f
the collection of all the
upper pre-image sets of f, that is,

f
= s : f (s) x : x X . (13)
Since any benchmark 1() is ordinally equivalent to , for any act f T the upper
pre-images of f with respect to and coincide: for all x X, s : f(s) x = s :
f(s) x. The set

of the -unambiguous events is thus naturally dened to

be the collection of all sets of upper pre-images of the acts in H

. That is,

fH

f
.
It is immediate to observe that if A

, then for every x, y X the binary act xAy

belongs to H

. This implies that A

c

(that is,

is closed w.r.t. complements).

6
Such set is well-dened since it is trivially true that the union of any collection of sets satisfying
(A) and (B) below also satises the two conditions.
19
We now present the characterization of the set

. This turns out to be quite simple

and intuitive: It is the subset of the events over which the capacity representing s
willingness to bet is complement-additive (sometimes called symmetric):
Proposition 22 Let be an ambiguity averse biseparable preference with willingness to
bet . Then for every 1(), the set

satises:

= A : (A) +(A
c
) = 1 . (14)
It immediately follows from the proposition that the choice of the specic benchmark
does not change the resulting set of events. In light of this, we henceforth call =

the set of unambiguous events for .

The consequences of the proposition for the CEU and MEU models are clear: Just
substitute or P for . In particular, when is a MEU preference with a set of prob-
abilities C, it can be further shown that is the set of events on which all probabilities
agree:
= A : (A) = P(A) for all P C .
It is also interesting to observe that is in general not an algebra. This is intuitive, as
the intersection of unambiguous events could be ambiguous.
7
As to the set of unambiguous acts H

, it can also be seen to be independent of the

choice of benchmark. In general, the only way to ascertain which acts are unambiguous is
to construct the set H

. However, for MEU preferences and for CEU preferences whose

capacity is exact (the lower envelope of its core), the set H

is the set of all the acts which

are measurable with respect to the events in . Therefore, in these cases characterizes
the set of unambiguous acts as well. (All these results are proved in [14].)
5 Back to Ellsberg
We now illustrate our results using the classical Ellsberg urn. The urn contains 90 balls
of three colors: red, blue and yellow. The DM knows that there are 30 red balls and that
the other 60 balls are either blue or yellow. However, he does not know their relative
proportion. The state space for an extraction from the urn is S = B, R, Y . Given the
nature of his information, it is natural to assume that the DMs preference relation
will be such that its set of unambiguous events satises , R, B, Y , S.
In particular, assume that the DMs preference relation is CEU and it induces the
capacity . To reect the fact that R and B, Y form an unambiguous partition, we
7
See Zhang [31] for a compelling urn example in which this happens.
20
know from the previous section that if the DM is ambiguity averse (or loving) must
satisfy
(R) +(B, Y ) = 1. (15)
Also, because of the symmetry of the information that the DM is given, it is natural to
assume that
(B) = (Y ) and (B, R) = (R, Y ) . (16)
We rst show that, if the ambiguity restriction (15) is imposed, ambiguity aversion is
not compatible with the following beliefs, which induce behavior that would on intuitive
grounds be considered ambiguity loving:
(R) < (B) = (Y ) ;
(B, Y ) < (B, R) = (R, Y ) .
(17)
Proposition 23 No ambiguity averse CEU preference relation such that its set of un-
ambiguous events contains R, B, Y can agree with the ranking (17).
In his paper on ambiguity aversion, Epstein [8] also discusses the Ellsberg urn, and he
presents a convex capacity compatible with ambiguity loving in his sense (see Subsection
6.3 for a brief review), which satises the conditions in (17). This is the capacity
1
dened by

1
(R) =
1
12
,
1
(B, Y ) =
1
3
,

1
(B) =
1
(Y ) =
1
6
,
1
(B, R) =
1
(R, Y ) =
1
2
.
He thus concludes that convexity of beliefs does not imply ambiguity aversion for CEU
preferences (it is also not implied, in his denition).
We know from Corollary 13 that convexity implies ambiguity aversion in our sense.
Proposition 23 helps clarifying why this example does not conict with the intuition
developed earlier: In fact,
1
does embody ambiguity aversion in our sense, but it does
not reect the usual presumption that R and B, Y are seen as unambiguous events.
If it did, it would have to satisfy (15), which is not the case (it cannot be, since convex
capacities are balanced). For us, the DM with beliefs
1
does not perceive R and
B, Y as unambiguous. Of course, then it is not clear in which sense the conditions in
(17) should intuitively embody ambiguity loving behavior.
Going back to the example, we would say that the DMs preferences intuitively reect
ambiguity aversion if the reverse inequalities held:
(R) (B) = (Y ) ;
(B, Y ) (B, R) = (R, Y ) .
(18)
21
We now show that the notion of ambiguity aversion proposed earlier characterizes this
intuitive ranking when, besides the obvious symmetry restrictions in (16), we strengthen
the requirement in (15) in the following natural way:
(R) =
1
3
and (B, Y ) =
2
3
. (19)
Proposition 24 Let be a CEU preference relation such that its representing capacity
satises the equalities (16) and (19). Then is ambiguity averse if and only if agrees
with the ranking (18).
In closing our discussion of Ellsbergs problem, we provide further backing for our
belief that convexity is not necessary for ambiguity aversion. Here is a capacity which is
not convex, and still makes the typical Ellsberg choices.
Example 25 Consider the capacity
2
dened by (19) and

2
(B) =
2
(Y ) =
7
24
,
2
(B, R) =
2
(R, Y ) =
1
2
.
This capacity satises (18), so that it reects ambiguity aversion both formally and
intuitively, but it is not superadditive, let alone convex.
6 Discussion
In this section we discuss some of the choices we have made in the previous sections. First
we briey discuss how the comparative ambiguity ranking can be extended to preferences
with dierent cardinal risk attitude. Then we discuss in more detail how the unambiguous
acts described in Section 4 can be used in the comparative ranking, and why we chose
SEU preferences as benchmarks.
6.1 Comparative Ambiguity and Equality of Cardinal Risk At-
titude
As we observed earlier, our comparative ambiguity aversion notion cannot compare bisep-
arable preferences with dierent canonical utility indices. Of course, the characterization
results of Section 3 can be used to qualitatively compare two preference by ambiguity:
For instance, we can look at two CEU preferences and compare their willingness to bet,
or we can use utility functions to compare two SEU preferences by risk aversion, even if
they do not have the same beliefs.
However, when dealing with biseparable preferences, it is easy to apply the intuition
of our comparative ranking to compare preferences which do not have the same canonical
utility. This requires eliciting the canonical utility indices rst, and then using acts and
22
constants that are utility equivalents in Eqs. (7) and (8).
8
The ranking thus obtained is
very general (it does not even entail ordinal equivalence), but it yields mutatis mutandis
the same characterization results that we obtained with the more restrictive one. For
instance: is ambiguity averse i T() ,= , and CEU (MEU) preference
2
is more
ambiguity averse than CEU (MEU) preference
1
i
1

2
(C
1
C
2
) (but of course
in general u
1
, u
2
). Nonetheless, this ranking requires the full elicitation of the DMs
canonical utility indices, and is thus operationally more complex than that in Denition 7.
6.2 Using Unambiguous Acts in the Comparative Ranking
One of the intuitive assumptions that our analysis builds on is that constant acts are
primitively unambiguous: That is, we assume that every DM perceives constants as
unambiguous. No other acts are unambiguous in this primitive sense. However, one
could argue that it is natural to use in the comparative ranking also those acts which are
revealed to be deemed unambiguous by both DMs, even if they are not constant.
Suppose that is an ambiguity averse biseparable preference, and let H

) be its
set of unambiguous acts (events), as dened in Section 4. It is possible to see [14] that
for every 1() and every h H

and f T, we have
h f h f and h > f h ~ f. (20)
That is, all benchmarks according to Denition 7 satisfy the stronger comparative ranking
suggested above. Conversely, it is obvious that if and a SEU preference are cardinally
symmetric and satisfy (20), they satisfy Denition 7. Thus, modifying Denition 7 to
have (20) in part (A) does not change the set of the ambiguity averse preferences.
6.3 A More General Benchmark
We chose SEU maximization as the benchmark representing ambiguity neutrality. While
few would disagree that SEU preferences are ambiguity neutral (in a primitive, non-
formal sense), some readers may nd that the result of Proposition 15 that SEU maxi-
mization characterizes ambiguity neutrality does not agree with their intuition of what
constitutes ambiguity neutral behavior. In particular, they might feel that we should also
classify as ambiguity neutral any non-SEU preference whose likelihood relation can still
be represented by a probability measure. This would clearly be the case if we let such
preferences be benchmarks for our comparative ambiguity notion. Here we explain why
we have not followed that route, and the consequences of this choice for the interpretation
of our notions.
8
For any pair of biseparable preferences which have essential events, this elicitation can be done
without extraneous devices by using the tradeo method briey outlined in Appendix B.
23
The non-SEU preferences in question are those that are probabilistically sophisticated
(PS) in the sense of Machina and Schmeidler [20]. For example, consider a CEU pref-
erence whose willingness to bet is = g(P) for some probability measure P and
distortion function g; that is, an increasing g : [0, 1] [0, 1] such that g(0) = 0 and
g(1) = 1. Such is PS since its ranking of bets (likelihood relation) is represented by the
probability P, but it is not SEU if g is dierent from the identity function. According to
the point of view suggested above, such is ambiguity neutral; it should thus be used as
a benchmark in characterizing ambiguity aversion. Moreover, if we used PS preferences
as benchmarks it might be possible to avoid attributing to ambiguity aversion the eects
of probabilistic risk aversion. However, go back to the ambiguous urn of Example 1 and
consider the following:
Example 1 (continued) In the framework of Example 1, consider a third DM with
CEU preferences
3
, with canonical utility u(x) = x and willingness to bet dened by

3
(B) =
1
4
and
3
(R) =
1
4
.
It is immediate to verify that according to Denition 7, DM 3 is more ambiguity averse
than DM 1 (who is SEU), so that he is ambiguity averse in our sense. That seems
quite natural, since he is willing to invest less in bets on the ball extractions. With PS
benchmarks, we conclude that both DMs are ambiguity neutral, since their willingness
to bet are ordinally equivalent to the probability
1
(
3
= g(
1
) for any distortion g such
that g(1/2) = 1/4), so that both are PS. Hence, DM 3s behavior is only due to his
probabilistic risk aversion. Yet, it seems that the fact that DM 3 is only willing to bet
1/4 utils on any color may at least in part be due to the ambiguity of the urn and his
possible ambiguity aversion.
This example is not the only case in which using PS benchmarks yields counterintu-
itive conclusions. When the state space is nite, if we use PS preferences as benchmarks
we nd that almost every CEU preference inducing a strictly positive on a nite state
space is both ambiguity averse and loving. Thus, a large set of preferences are shown to
be ambiguity neutral. Including, as the following example illustrates, many preferences
which are not PS.
Example 26 Suppose that two DMs are faced with the following decision problem.
There are two urns, both containing 100 balls, either red or black. The DMs are told
that Urn I contains at least 40 balls of each color, while Urn II contains at least 10
balls of each color. One ball will be extracted from each urn. Thus, the state space is
S = Rr, Rb, Br, Bb, where the upper (lower) case letter stands for the color of the
ball extracted from Urn I (II). Suppose that both DMs have CEU preferences
1
and

2
, with respective willingness to bet
1
and
2
. Using obvious notation, suppose that

1
(b) =
1
(r) = 0.1 and
1
(B) =
1
(R) = 0.4, that
1
(s) = 0.04 for each singleton s, and
for every other event
1
is obtained by additivity. According to Denition 9, DM 1 is
24
strictly ambiguity averse. In contrast, with PS benchmarks the result mentioned above
shows that he is ambiguity neutral.
Let
2
be as follows:
2
(b) =
2
(r) = 0.9 and
2
(B) =
2
(R) = 0.6,
2
(s) = 0.54 for
each singleton s,
2
(A) = 0.92 for each A Rr Bb, Rb Br, and
2
(A) = 0.95 for
each ternary set. According to Denition 9, DM 2 is ambiguity loving, but if we use PS
benchmarks we conclude that she is ambiguity neutral. Both conclusions go against our
intuition. Moreover, since both
1
and
2
are not ordinally equivalent to a probability,

1
and
2
are not PS.
The foregoing discussion shows some of the diculties that may arise if we use PS,
rather than SEU, preferences as benchmarks with our comparative ambiguity aversion
notion: We end up attributing too much explanatory power to probabilistic risk aversion.
Instead, with SEU benchmarks we overemphasize the role of ambiguity aversion. Is it
possible to remove probabilistic risk attitude from the picture, as we did for cardinal risk
attitude?
9
6.3.1 Removing Probabilistic Risk Aversion
Suppose that there is a subset c of acts which are universally accepted as unambiguous,
in the sense that we are sure that a DMs choices among these acts are unaected by his
ambiguity attitude. Then, if c (and the associated set of unambiguous events, denoted
) is suciently rich, we can discriminate between probabilistic risk and ambiguity aver-
sion. For instance, modify Example 1 by assuming the availability of an unambiguous
randomizing device, so that each state describes the result of the device as well. Now,
nd a set A of results of the device (obviously, here is the family of all such sets) which
is as likely as R(ed) and then check if B(lack) is as likely as A
c
. If it is, the DM behaves
identically when faced with (equally likely) ambiguous and unambiguous events, so that
all the non-additivity of
3
on B, R must be due to his probabilistic risk aversion. His
preferences are also PS on the extended problem. If it is not, then DM 3s behavior is
aected by ambiguity, and his preferences are not PS on the extended problem. The
point is that in the presence of a suciently rich , a DM whose preferences are PS
is treating ambiguous and unambiguous events symmetrically, and is hence intuitively
ambiguity neutral. Therefore, in such a case we would expect PS preferences to be found
ambiguity neutral. This is not the case in the original version of Example 1, since a rich
set of unambiguous events is missing.
More generally, consider a biseparable preference which is not PS overall, but is
PS when comparing only unambiguous acts. That is, the DM behaves as if he forms
a probability P on the set , and calculates his willingness to bet on these events by
means of a distortion function g which only reects his probabilistic risk attitude. As
we did in controlling for cardinal risk attitude, we want to use as benchmarks for only
those PS preferences that with a small abuse of notation we also denote which
9
We thank Peter Klibano for his substantial help in developing the ensuing discussion.
25
have the same probabilistic risk attitude; e.g., those biseparable preferences which share
g as distortion function. Interestingly, it turns out that if the set c is rich enough, any
PS preference satisfying Eq. (20) for all h c has this property. This is exactly the
approach followed by Epstein [8] in his work on ambiguity aversion: He assumes the
existence of a suitably rich set of unambiguous events,
10
denes c as the set of all the
-measurable acts, and uses Eq. (20) with h c as his comparative ambiguity notion.
His choice of benchmark are PS preferences.
This approach attains the objective of ltering the eects of probabilistic risk atti-
tude from our absolute ambiguity notion. It thus yields a ner assessment of the DMs
ambiguity attitude. However, the foregoing discussion has illustrated that a crucial ingre-
dient to this ltration is the existence of a set of unambiguous acts which is suciently
rich: If it is too poor (e.g., it contains only the constants, as in Example 26), we may use
benchmarks whose probabilistic risk attitude is dierent from the DMs. This may cause
Epsteins approach to reach counterintuitive conclusions, as illustrated in the previous
examples.
The main problem we have with this approach is that we nd it undesirable to base
our measurement of ambiguity attitude on an exogenous notion of ambiguity, especially
in view of the richness requisite. It seems that in many cases of interest the obvious
set of unambiguous acts does not satisfy such requisite; e.g., Ellsbergs example. Our
objective is to develop a notion of ambiguity attitude which is based on the weakest set
of primitive requisites (like the two assumptions stated in the Introduction), even though
this has a cost in terms of the purity of the interpretation of the behavioral feature we
measure.
Epstein and Zhang [10] propose a behavioral foundation to the notion of ambiguity,
so that the existence of a rich set c can be objectively veried, solving the problem
mentioned above. In [14] we present an example which suggests that their behavioral
notion can lead to counterintuitive conclusions (in that case, an intuitively ambiguous
event is found unambiguous). More generally, we see the following problem with this
enterprise: There may be events which are unambiguous (resp. ambiguous) with
respect to which the DM nonetheless behaves in an ambiguity non-neutral (resp. neutral)
fashion. Consider a DM who listens to a weather forecast stated as a probabilistic
judgement. If the DM does not consider the specic source reliable, he might express a
willingness to bet which is a distortion of this judgement, while being probabilistic risk
neutral. Alternatively, he may nd the source reliable, hence perceive no ambiguity, but
be probabilistically risk averse. A preference-based notion of ambiguity must be able to
distinguish between these two cases, classifying the relevant events ambiguous in the rst
case and unambiguous in the second. And this without using any auxiliary information.
Considering moreover that the set of veriably unambiguous events must be rich, we
are skeptical that this feat is possible: The problem is that the Savage set-up does not
10
The richness condition is: For every F E in and A such that A is as likely as E, there is
B A in such that B is as likely as F. Epstein remarks that richness of is not required for some of
his results.
26
provide us with enough instruments; it is too abstract.
6.3.2 Summing Up
We have argued that what motivates using PS (rather than SEU) preferences as bench-
marks is the objective of discriminating between probabilistic risk aversion and ambiguity
attitude. We have shown that this requires a rich set of veriably unambiguous events,
and briey reviewed our doubts about the possibility of providing a behavioral foundation
to this veriable ambiguity notion in a general subjective setting without extraneous
devices. In contrast, the analysis in this paper shows that there are no such problems in
using SEU benchmarks to identify an extended notion of ambiguity attitude, which can
be disentangled from cardinal risk attitude using only behavioral data and no extraneous
devices. Though it does not distinguish real ambiguity and probabilistic risk attitudes,
we think that this extended ambiguity attitude is worthwhile, especially because of its
wider applicability.
27
Appendix A Capacities and Choquet Integrals
A set-function on (S, ) is called a capacity if it is monotone and normalized. That
is: if for A, B , A B, then (A) (B); () = 0 and (S) = 1. A capacity
is called a probability measure if it is nitely additive: (A B) = (A) + (B)
for all A disjoint from B. It is called convex if for every pair A, B , we have
(A B) (A) +(B) (A B).
The core of a capacity is the (possibly empty) set (() of all the probability
measures on (S, ) which dominate it, that is,
(() P : P , P(A) (A) for all A .
Following the usage in Cooperative Game Theory (e.g., Kannai [17]), all capacities with
nonempty core are called balanced. A capacity is called exact it is balanced and it
is equal to the lower envelope of its core (i.e., for all A , (A) = min
PC()
P(A)).
Convex implies exact, which in turn implies balanced, but the converse implications are
all false.
The notion of integral used for capacities is the Choquet integral, due to Choquet
[5]. For a given -measurable function : S R, the Choquet integral of with respect
to a capacity is dened as:

S
d =

0
(s S : (s) )d +

[1 (s S : (s) )]d (21)

where the r.h.s. is a Riemann integral (which is well dened because is monotone).
When is additive, (21) becomes a standard (additive) integral. In general it is seen
to be monotonic, positive homogeneous and comonotonic additive: If , : S R
are non-negative and comonotonic, then

( + ) d =

d +

d. Two functions
, : S R are called comonotonic if there are no s, s

S such that (s) > (s

)
and (s) < (s

).
Appendix B Cardinal Symmetry and Biseparable
Preferences
In this Appendix, we prove Proposition 6. In order to make the proof as clear as possible,
we rst explain the notion of standard sequence, and then show how the latter can be
used to prove the proposition.
B.1 Standard Sequences
Consider a DM whose preferences have a canonical representation V , with canonical
utility index u, willingness to bet , and an essential event A . Fix a pair of conse-
quences v

~ v

, and consider x
0
X such that x
0
~ v

. If there is an x X such that

28
xAv

~ x
0
Av

, then by (3) and the convexity of the range of u, there is x

1
X such
that
x
1
Av

x
0
Av

. (22)
It is easy to verify that x
1
~ x
0
: If x
0
x
1
held, by monotonicity and biseparability,
we would have x
0
Av

x
1
Av

and x
1
Av

~ x
1
Av

. This yields x
0
Av

~ x
1
Av

, a
contradiction. Assuming that there is an x X such that x Av

~ x
1
Av

, as above we
can nd x
2
X such that
x
2
Av

x
1
Av

. (23)
Again, x
2
~ x
1
. We can use the representation V to check that the equivalences in (22)
and (23) translate to
u(x
1
) u(x
0
) =
1 (A)
(A)
(u(v

) u(v

)) = u(x
2
) u(x
1
), (24)
that is, the three points x
0
, x
1
, x
2
, are equidistant in u. Proceeding in this fashion we
can construct a sequence of points x
0
, x
1
, x
2
, . . . all evenly spaced in utility. Such
sequence we call an increasing standard sequence with base x
0
, carrier A and
mesh (v

, v

). (Notice that the distance in utility between the points in the sequence is
proportional to the distance in utility between v

and v

, which is used as the measuring

rod.)
Analogously, we can construct a decreasing standard sequence with base x
0
, carrier
A and mesh (v

, v

) where v

~ x
0
. This will be a sequence starting again from x
0
, but
now moving in the direction of decreasing utility: For every n 0, v

Ax
n+1
v

Ax
n
.
Henceforth, we call a standard sequence w.r.t. (x
0
, A) any sequence x
0
, x
1
, x
2
, . . .
such that x
0
= x
0
, and there is a pair of points (above or below x
0
) which provides the
mesh for obtaining x
0
, x
1
, x
2
, . . . as a decreasing/increasing standard sequence with
carrier A.
It is simple to see how having xed an essential event A, and a base x
0
which is
non-extremal in the ordering on X (i.e., there are y, z X such that y ~ x
0
~ z)
standard sequences can be used to measure the canonical utility index u of a biseparable
preference (extending the scope of the method proposed by Wakker and Denee [29]):
One just needs to construct (increasing and decreasing) standard sequences with base
x
0
and ner and ner mesh. In what follows we use standard sequences and cardinal
symmetry to show that equality of the u
i
, i = 1, 2, can be veried without eliciting them.
B.2 Equality of Utilities: Proof of Proposition 6
The proof of Proposition 6 builds on two lemmas. The rst lemma, whose simple proof we
omit, shows the following: Suppose that a pair of biseparable preferences are cardinally
29
symmetric, then for xed non-extremal x
0
and essential events A
1
and A
2
, the sets
of the standard sequences (with respect to (x
0
, A
1
) and (x
0
, A
2
) respectively) of the
orderings are nested into each other. Stating this lemma requires some terminology
and notation: Given a standard sequence x
n
for preference relation
i
, we say that a
sequence y
m
X is a renement of x
n
if it is itself a standard sequence, and it is
such that y
m
= x
n
whenever m = kn for some k N. Two canonical utility indices are
subject to a common normalization if they take identical values on two consequences
x, y X such that x ~
i
y for both i. Finally, for the rest of this section: For each
i = 1, 2, the carrier of any standard sequence for
i
is a xed essential event A
i
, and
SQ(
i
, x
0
) X denotes the set of the points belonging to some standard sequence of
i
with base x
0
and carrier A
i
.
Lemma 27 Suppose that
1
,
2
are as assumed in Proposition 6. Fix a non-extremal
x
0
X. If
1
and
2
are cardinally symmetric, then the following holds: Either every
standard sequence for ordering
1
is a renement of a standard sequence for
2
, or every
standard sequence for ordering
2
is a renement of a standard sequence for
1
. Hence,
SQ(
1
, x
0
) = SQ(
2
, x
0
) SQ(x
0
).
The second lemma shows that, because of cardinal symmetry, the result holds on
SQ(x
0
):
Lemma 28 Suppose that
1
,
2
are as assumed in Proposition 6. If
1
and
2
are
cardinally symmetric, then for any non-extremal x
0
X and any common normalization
of the two indices, u
1
(x) = u
2
(x) for every x SQ(x
0
).
Proof : Fix a non-extremal x
0
. Suppose that x belongs to an increasing standard sequence
for
i
, x
n
. Since the relations are cardinally symmetric, by Lemma 27 it is w.l.o.g.
(taking renements if necessary) to take the sequence to be standard for both orderings.
That is, there are v

, v

, w

, w

X such that v

~
1
v

, w

~
2
w

and for n 0,
x
n+1
A
1
v

1
x
n
A
1
v

,
and analogously for
2
(with w replacing v). Moreover, there is n 0 such that x = x
n
.
Choose x
m
for some m > n, and take positive ane transformations of the two canonical
utility functions so as to obtain u
1
(x
0
) = u
2
(x
0
) = 0 and u
1
(x
m
) = u
2
(x
m
) = 1. All points
in the sequence are evenly spaced for both preferences (cf. Eq. (24)). Hence we have
u
1
(x
n
) = u
2
(x
n
) = n/m. The case in which x belongs to a decreasing standard sequence is
treated symmetrically. Finally, we have the immediate observation that if u
1
(x) = u
2
(x)
for one common normalization, the equality holds for every common normalization.
Proof of Proposition 6: The if part follows immediately from the canonical represen-
tation. We now prove the only if. Start by xing a non-extremal x
0
and adding a
constant to both indices, so that u
1
(x
0
) = u
2
(x
0
) = 0. Suppose that (after this transfor-
mation) there is x X such that u
1
(x) ,= u
2
(x). By relabelling if necessary, assume that
30
u
1
(x) = > = u
2
(x). There are dierent cases to consider, depending on where and
are located.
Suppose rst that 0. Choose v

X such that x
0
~
1
v

and further transform

the utilities so that u
1
(v

) = u
2
(v

) = 1, to obtain u
1
(x) = >

= u
2
(x). Choose
> 0 such that

> . By the connectedness of the range of each u
i
and Lemma 27,
there are v

, w

X such that (v

, v

) and (w

, v

) generate the same standard sequence

x
n
and
u
1
(x
n+1
) u
1
(x
n
) = u
2
(x
n+1
) u
2
(x
n
) < .
So the length of the utility interval between each element in the increasing standard
sequence is smaller than the distance between and

. We also proved in Lemma 28
that for each element in the standard sequence, we have equality of the utilities (since
we imposed a common normalization). Hence there must be n 0 such that u
1
(x
n
) =
u
2
(x
n
) = (

, ). We then have
u
1
(x
n
) > u
1
(x) x
n
~
1
x and u
2
(x
n
) < u
2
(x) x
n

2
x,
which contradicts the assumption of ordinal equivalence.
The case in which 0 is treated symmetrically. If, nally, > 0 > then,
using an argument similar to the one just presented, one can nd x X such that
u
1
( x) = u
2
( x) (0, ) and obtain a similar contradiction. This shows that u
1
(x) = u
2
(x)
for every x X.
Appendix C Proofs for Sections 3 to 5
C.1 Section 3
Proof of Theorem 10: We rst state without proof an immediate result:
Lemma 29 Two preference relations
1
and
2
satisfying Eqs. (7) and (8) are ordinally
equivalent.
Given this lemma, if
1
and
2
have essential events the result follows immediately
from Proposition 6. If, say, relation
i
does not have essential events, any ordinal
transformations of u
i
is still a canonical utility. Since the two preferences are ordinally
equivalent by the lemma, it is then w.l.o.g. to use u
j
(j ,= i) to represent both of them.
Proof of Theorem 12: We rst prove that T() /(). Given a canonical representa-
tion V of with canonical utility u, suppose that P T(), and consider the relation
31
induced by P and u. We want to show that is more ambiguity averse than . Since
P T(),

u(f) dP V (f) for all f T, so that for every x X and f T,

u(x)

S
u(f(s)) P(ds) =V (x) V (f),
where the implication follows from the denition u(x) = V (x) for all x X. This
proves that (7) holds. Similarly one shows the validity of (8). Part (B) of Denition 7 is
immediate: If and have essential events, then the result follows from Proposition 6.
Hence 1(), or in other words P /().
We now prove the opposite inclusion T() /(). Suppose that P /(). Let
be the benchmark preference corresponding to P, and let u

be the canonical utility

index of . Since is a benchmark for , we have for every x X and f T,
u

(x)

S
u

(f(s)) P(ds) =u(x) V (f), (25)

and the same with strict inequality. We have to show that P T(). By Theorem 10,
it is w.l.o.g. to take u = u

. Hence, (25) implies that

u(f) dP V (f) for all f T,

and so P T().
Proof of Corollary 13: By Theorem 12, /() = T(). Let P T(). For every A
and x

~ x

, consider the act f = x

Ax

. Normalizing u(x

) = 1 and u(x

) = 0, we
have
P(A) =

S
u(f(s)) P(ds)

S
u(f(s)) (ds) = (A),
and so P ((). This implies T() ((). The converse inclusion is trivial, since
P (() implies

u(f) dP

u(f) d for all f T.

Proof of Corollary 14: We are done if we show that for all f, g T,
f g min
PD()

S
u(f(s)) P(ds) min
PD()

S
u(g(s)) P(ds). (26)
This follows from the fact that there exists a unique weak

-compact and convex set C

representing . T() is clearly weak

-compact (so that the minimum in (26) is well

dened) and convex. Hence, if (26) holds C = T(), and by Theorem 12, T() =
/().
To prove (26), suppose there are f, g T such that
min
PC

u(f) dP min
PC

u(g) dP and min

PD()

u(f) dP < min

PD()

u(g) dP.
Let P

arg min

S
u(f(s)) P(ds) : P T(). Since C T(), we have:
min
PC

S
u(f(s)) P(ds)

S
u(f(s)) P

(ds) < min

PD()

S
u(g(s)) P(ds) min
PC

S
u(g(s)) P(ds),
32
a contradiction. Similarly, one shows that there cannot be f, g T such that the
preference based on T() prefers weakly f to g, while g ~ f. This shows that Eq. (26)
holds, concluding the proof.
Proof of Proposition 15: That every SEU preference is ambiguity neutral follows imme-
diately from two applications of Theorem 13. As for the converse: If is both ambiguity
averse and ambiguity loving, there are a SEU preference relation
1
(represented by prob-
ability P
1
) such that is more ambiguity averse than
1
, and a SEU preference relation

2
(represented by probability P
2
) which is more ambiguity averse than . Applying
Denition 7 twice, we obtain that for every f T and x X,
x
1
f x
2
f and x >
1
f x >
2
f.
We show that
1
and
2
are cardinally symmetric. This requires rst showing that
if
2
has an essential event, so must . Suppose that A is essential for
2
, so
that for some x ~ y (remember that and
1
and
2
are all ordinally equivalent),
x >
2
xAy >
2
y. Using the contrapositive of (7), we then have xAy ~ y. Since
2
is
a SEU preference, A
c
is also
2
-essential, similarly implying xA
c
y ~ y. Now, suppose
that has no essential event. Because of the preferences we just derived, we must have
both x xAy and x xA
c
y. This is impossible since
1
1(), for the contrapositive
of (8) then yields xAy
1
x, which implies P
1
(A) = 1, and xA
c
y
1
x, which implies
P
1
(A) = 0. This gives us a contradiction, so that must have an essential event if

2
does. Hence,
2
and have essential events, and they are cardinally symmetric by
assumption. Similarly one shows that
1
and have essential events and are cardinally
symmetric. It is now immediate to check that these facts imply that
1
and
2
are
cardinally symmetric. We thus conclude that
2
is more ambiguity averse than
1
.
Mimicking the last part of the proof of Theorem 12, we then show that then P
1
P
2
,
which immediately implies P
1
= P
2
, so that
1
=
2
. Thus is both more and less
ambiguity averse than , which immediately implies =.
Proof of Theorem 17: Part (i) follows immediately along the lines of the proofs of Theo-
rem 12 and Corollary 13. As for part (ii), it is similarly immediate to show that if
2
is
more ambiguity averse than
1
, then C
1
T(
2
) and u
1
u
2
. We show the converse.
Let V
1
and V
2
denote the canonical representations of
1
and
2
, and w.l.o.g. assume
that u
1
= u
2
= u. Then C
1
T(
2
) implies that for every f T and every P C
1
,
V
2
(f)

u(f) dP. Hence, using the fact that

1
is MEU, we nd
V
2
(f) min
PC
1

S
u(f(s)) P(ds) = V
1
(f),
which immediately yields the desired result.
33
C.2 Section 4
Proof of Proposition 22: Let 1() and set A : (A) + (A
c
) = 1. If
A

for all x X we have

u(x) = P(A) u(x) = (A)
u(x) = P(A
c
) u(x) = (A
c
),
and so (A) = P(A) and (A
c
) = P(A
c
). This implies that A , so that

.
Now, if A we have
(A) = P(A) and (A
c
) = P(A
c
). (27)
In order to show that A

, we need to show that any act measurable w.r.t. the

partition A, A
c
is in H

. This follows from (27), as for every x, y X we have

V (xAy) = V

(xAy). Thus

, which concludes the proof.

C.3 Section 5
Proof of Proposition 23: Suppose, to the contrary, that agrees with (17). If Eq. (15)
holds then P(R) = (R) and P(B, Y ) = (B, Y ) for all P ((), so that we have
P (B, Y ) = (B, Y ) < (B, R) P(B, R).
In turn, this implies P(Y ) < P(R), yielding (Y ) P(Y ) < P(R) = (R). Hence
(Y ) < (R), contradicting (17).
Proof of Proposition 24: Every which satises (18) is such that (() ,= . For, the
measure P such that P(R) = P(B) = P(Y ) = 1/3 belongs to ((). This proves that all
preferences satisfying (18) are ambiguity averse.
As to the converse, let be ambiguity averse, i.e. (() ,= . Let P ((). Assume
rst that (B) = (Y ) > (R). Since P(B) (B) and P(Y ) (Y ),
P(B) +P(R) +P(Y ) > (B) +(R) +(Y ) > 1,
a contradiction. Assume now (B, Y ) < (B, R) = (R, Y ). This implies P(B, Y ) <
P(B, R) and P(B, Y ) < P(R, Y ), so that P(Y ) < P(R), P(B) < P(R), and P(B) +
P(R) +P(Y ) < 1, a contradiction.
34
References
[1] Frank J. Anscombe and Robert J. Aumann. A denition of subjective probability.
Annals of Mathematical Statistics, 34:199205, 1963.
[2] Kenneth J. Arrow. The theory of risk aversion. In Essays in the Theory of Risk-
Bearing, chapter 3. North-Holland, Amsterdam, 1974. (Part of the Yri o Jahnssonin
Sa atio lectures in Helsinki, 1965).
[3] Ramon Casadesus-Masanell, Peter Klibano, and Emre Ozdenoren. Maxmin ex-
pected utility over Savage acts with a set of priors. Journal of Economic Theory,
92:3365, 2000.
[4] Alain Chateauneuf and Jean-Marc Tallon. Diversication, convex preferences and
non-empty core. Mimeo, Universite Paris I, July 1998.
[5] Gustave Choquet. Theory of capacities. Annales de lInstitut Fourier (Grenoble),
5:131295, 1953.
[6] Bruno de Finetti. Sulla preferibilit` a. Giornale degli Economisti e Annali di Econo-
mia, 6:327, 1952.
[7] Daniel Ellsberg. Risk, ambiguity, and the Savage axioms. Quarterly Journal of
Economics, 75:643669, 1961.
[8] Larry G. Epstein. A denition of uncertainty aversion. Review of Economic Studies,
66:579608, 1999.
[9] Larry G. Epstein and Tan Wang. Intertemporal asset pricing under Knightian un-
certainty. Econometrica, 62:283322, 1994.
[10] Larry G. Epstein and Jiankang Zhang. Subjective probabilities on subjectively un-
ambiguous events. Econometrica, forthcoming.
[11] Peter C. Fishburn. The axioms and algebra of ambiguity. Theory and Decision,
34:119137, 1993.
[12] Paolo Ghirardato and Jonathan N. Katz. Indecision theory: Explaining selective ab-
stention in multiple elections. Social Science Working Paper 1106, Caltech, Novem-
ber 2000. http://masada.hss.caltech.edu/paolo/voting.pdf.
[13] Paolo Ghirardato and Massimo Marinacci. Risk, ambiguity,
and the separation of utility and beliefs. Social Science Work-
ing Paper 1085, Caltech, March 2000. (Revised: August 2000);
http://masada.hss.caltech.edu/paolo/canonica4wp.pdf.
[14] Paolo Ghirardato and Massimo Marinacci. A subjective denition of ambiguity.
Work in progress, Caltech and Universit` a di Torino, 2000.
35
[15] Itzhak Gilboa and David Schmeidler. Maxmin expected utility with a non-unique
prior. Journal of Mathematical Economics, 18:141153, 1989.
[16] Lars P. Hansen, Thomas Sargent, and Thomas D. Tallarini. Robust permanent
income and pricing. Review of Economic Studies, 66:873907, 1999.
[17] Yakar Kannai. The core and balancedness. In Robert J. Aumann and Sergiu Hart,
editors, Handbook of Game Theory, pages 355395. NorthHolland, Amsterdam,
1992.
[18] David Kelsey and Shasikanta Nandeibam. On the measurement of uncertainty aver-
sion. Mimeo, University of Birmingham, September 1996.
[19] David H. Krantz, R. Duncan Luce, Patrick Suppes, and Amos Tversky. Foundations
of Measurement: Additive and Polynomial Representations, volume 1. Academic
Press, San Diego, 1971.
[20] Mark J. Machina and David Schmeidler. A more robust denition of subjective
probability. Econometrica, 60:745780, 1992.
[21] Aldo Montesano and Francesco Giovannoni. Uncertainty aversion and aversion to
increasing uncertainty. Theory and Decision, 41:133148, 1996.
[22] Sujoy Mukerji. Ambiguity aversion and incompleteness of contractual form. Amer-
ican Economic Review, 88:12071231, 1998.
[23] Klaus Nehring. Capacities and probabilistic beliefs: A precarious coexistence. Math-
emathical Social Sciences, 38:197213, 1999.
[24] John W. Pratt. Risk aversion in the small and in the large. Econometrica, 32:122
136, 1964.
[25] Leonard J. Savage. The Foundations of Statistics. Wiley, New York, 1954.
[26] David Schmeidler. Subjective probability and expected utility without additivity.
Econometrica, 57:571587, 1989.
[27] Amos Tversky and Peter P. Wakker. Risk attitudes and decision weights. Econo-
metrica, 63:12551280, 1995.
[28] Peter P. Wakker. Additive Representations of Preferences. Kluwer, Dordrecht, 1989.
[29] Peter P. Wakker and Daniel Denee. Eliciting von Neumann-Morgenstern utilitites
when probabilities are distorted or unknown. Management Science, 42:11311150,
1996.
[30] Menachem E. Yaari. Some remarks on measures of risk aversion and on their uses.
Journal of Economic Theory, 1:315329, 1969.
[31] Jiankang Zhang. Subjective ambiguity, probability and capacity. Mimeo, University
of Toronto, October 1996.
36