The Geneva Papers on Risk and Insurance Theory, 26: 117–137, 2001
c 2001 The Geneva Association
On Microfoundations of the Dual Theory of Choice
SERGEI GURIEV sguriev@nes.ru
New Economic School, CEFIR and CEPR, Nakhimovsky pr. 47, Moscow 117418, Russia
Received July 20, 2000; Revised June 29, 2001
Abstract
We show that Yaari’s dual theory of choice under risk may be derived as an indirect utility when a risk-neutral
agent faces financial imperfections.We consider an agent that maximizes expected discounted cash flows under a
bid-ask spread in the credit market. It turns out that the agent evaluates lotteries as if she were maximizing Yaari’s
dual utility function. We also generalize the dual theory of choice for unbounded lotteries.
Key words: dual theory of choice, financial imperfections
JEL classification No.: D81, G11
1. Introduction
The dual theory of choice under risk (DTC)was introduced inYaari [1987] in order to resolve
certain theoretical problems with expected utility. One of the most important properties of
the latter is that risk aversion is equivalent to diminishing marginal utility. Yaari argued
that risk aversion and diminishing marginal utility are ‘horses of different colors’ and put
forward a theory which allowed for linear risk averse utilities. The linearity property makes
modelling behavior of risk-averse agents much simpler. This is why DTC has become a
popular tool for testing robustness of various economic models that have been so far analyzed
only in the expected utility framework. Demers and Demers [1990] apply DTC to firms’
production decisions. Hadar and Seo [1995] examine portfolio choice and diversification in
DTC. Doherty and Eeckhoudt [1995] study optimal insurance, Epstein and Zin [1990] use
DTC as the simple utility with first-order risk aversion to offer an explanation of the equity
premium. Volij [1999] studies revenue equivalence in auction with bidders that maximize
DTC utility. Schmidt [1998] applies DTC to principal-agent problems.
Despite being a handy tool in theoretical models, the dual theory of choice has not done
well in experimental studies. Both Harless and Camerer [1994] and Hey and Orme [1994]
showed that DTC performs rather badly not only in absolute terms but also relative to
other non-expected utility theories as well as relative to the expected utility theory. This
may not be surprising since the experiments were carried out on individuals. One should
expect individuals to have diminishing marginal utility. On the other hand, firms and banks,
especially in the long-run should have linear objective function. In the meanwhile, there
118 SERGEI GURIEV
exists a substantial empirical evidence that firms and even banks are risk-averse [see Rose,
1989; Davidson et al., 1992; Park and Antonovitz, 1992 etc.].
In the expected utility theory, risk-aversion of firms can be explained by market imperfections
[Greenwald and Stiglitz, 1990]. If credit markets are imperfect, a firm prefers
certainty: a mean-preserving spread of project payoffs decreases expected profits after interest.
Even if the firm is orginally risk-neutral and maximizes expected cash flows, market
imperfections may make it risk-averse.
The goal of our paper is to provide similar microfoundations for the dual theory of choice.
We also consider a firm that is initially risk neutral. The firm evaluates lotteries (or project
portfolios) according to the amount of expected discounted cash flows that can be obtained
with those lotteries. We assume that the firm faces bid-ask spread in the financial market,
i.e. the interest rate on loans is higher than the interest rate on deposits. The innovation of
the paper is to assume that the firm anticipates the future need for borrowing if the returns
are low and saving extra funds if the returns are high. Therefore firm can adjust its financial
position before realization of stochastic payoffs. We derive indirect utility as a function of
the distribution of returns and show that the indirect utility belongs to a certain subset of
DTC utilities.
Another contribution of the paper is a generalization of DTC for unbounded random
variables. Although infinite payments are not likely to occur in real world, in models they
are quite common (e.g. normal distribution).We obtain representation for random variables
with finite means. Our representation form turns out to be similar to one introduced in Roell
[1987].
The paper is organized as follows. In Section 2 we introduce notation and obtain representation
of DTC for unbounded lotteries. In Section 3, we show that an agent that faces a
bid-ask spread in the financial market, evaluates lotteries as if she were a DTC utility maximizer.
In Section 4 we check whether this result also holds for an arbitrary (not necessarily
piecewise constant) interest rate schedule and show that it does not. Section 5 concludes.
The Appendix contains proofs.
2. Dual theory of choice revisited
In this Section we extend the dual theory of choice to the case of unbounded random
variables with finite means. In addition to original Yaari’s axioms, we assume finiteness of
certainty equivalents. Then we obtain a representation result, and characterize risk aversion.
2.1. Notation
We shall consider a preference relation  over real-valued random variables (‘lotteries’) X.
We consider a probability space (,A,P) where  is the set of states of nature, A is a σ-
algebra on,P is a probability measure onA.Arandom variable is a real-valued function of
state of nature X :  −→ . For each random variable we introduce a cumulative distribution
function FX (x) = Prob{X ≤ x}. The distribution functions are right-continuous, nondecreasing
and map [−∞,+∞] onto [0, 1].We will also use inverse distribution functions
(IDF) (p) = F−1(p) = sup{x : p ∈ ˆF (x)}, where ˆF (x) = {p : F(x − 0) ≤ p ≤ F(x)}
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 119
is the set-valued distribution function. According to this definition, a point (p, x) belongs
to the graph of (p) if and only if the point (x, p) belongs to the graph of F(x). (p) is
a non-decreasing function that maps [0, 1] onto [−∞,+∞].
We will consider all random variables with finite means so that |
 +∞
−∞ xdF(x)| =
|
 1
0 (p)dp|<∞. Then (p) ∈ L1(0, 1) and F(x) ∈ L1(−∞,+∞). Denote M a set of all
non-decreasing upper semi-continuous functions that belong to L1(0, 1).1 Then M includes
IDFs of all random variables with finite means. M is a semi-linear space. Indeed, for any
1,2 ∈ M and α ≥0 we have 1 +2 ∈ M and α1 ∈ M. There is a zero element in
M : 0(p)≡0 for all p ∈ [0, 1]. Indeed, 0 =0 ·  and 0 += for any H ∈ M.
We do not require (0) or (1) to be finite and therefore allow for unbounded random
variables. The L1 norm is very convenient for dealing with IDFs: the distance between
two distribution functions
 +∞
−∞ |F1(x)− F2(x)|dx in terms of space L1(−∞,+∞) coincides
with distance between corresponding IDFs 1 −2=
 1
0
|1(p)−2(p)|dp
in terms of L1(0, 1).
We shall also use the concept of comonotonicity.
Definition 1: Two random variables X and Y are said to be comonotonic if (X(ω) −
Y (ω))(X(ω

) − Y (ω

)) ≥ 0 holds for any pair of states of nature ω,ω
.
We will denote C(b) a degenerate random variable that takes real value b with probability
1 : C(b) ≡ b.
2.2. Dual theory of choice for unbounded lotteries
We use Yaari’s axioms i.e. the axioms of expected utility theory where the von Neuman-
Morgenstern independence axiom is replaced with Yaari’s dual independence axiom. Then
we introduce an additional condition that allows us to deal with unbounded random
variables.
Axiom 1 (Neutrality (A1)): If inverse distribution functions of two random variables X
and Y coincide (i.e. X (·) = Y (·)) then X ∼ Y.2
This means that instead of dealing with random variables we can simply define preference
relation on M. By definition, X (·) = Y (·) is equivalent to FX (·) = FY (·). Hereinafter,
we shall understand X (·) = Y (·) as equality almost everywhere (i.e. maybe except for
a set of measure zero).
Axiom 2 (Complete weak order (A 2)): Preference  is a complete weak order.
Axiom 3 (Continuity (A 3)): The preference relation is continuous on M in terms of
L1 norm i.e. if X  Y then there exists ε>0 such that for all Z with finite means
Z −Y <ε implies X  Z.
Axiom 4 (Monotonicity (A 4)): If X (p) ≥ Y (p) for every p ∈ [0, 1] then X  Y .
The monotonicity axiom A4 is equivalent to conventional monotonicity in terms of firstorder
stochastic dominance (FOSD). Indeed,X (·) ≥ Y (·) is equivalent to the well-known
FOSD condition FX (·) ≤ FY (·).
120 SERGEI GURIEV
Axiom 5 (Dual independence (A 5)): If random variables X, Y and Z are pairwise comonotonic
then for every real number α ∈ [0, 1], X  Y implies αX+(1−α)Z  αY+(1−α)Z.
The dual independence axiom introduces linearity with regard to payoffs like von
Neuman-Morgenstern independence axiom makes sure that utility is linear with regard
to probabilities. We will recurrently refer to this duality between payoffs and probabilities
throughout the paper. Notice that A5 imposes linearity only upon the comonotonic lotteries
(i.e. ones that cannot be used as a hedge against each other); otherwise the preferences
would have to be risk neutral.
Axiom 6 (Finite certainty equivalent (A 6)): For each random variable with finite mean X,
there exist such (finite) real values a and A that C(a) ≺ X ≺ C(A).
The first five axioms are precisely the ones introduced by Yaari [1987]. If the preference
were defined on the set of uniformly bounded random variables, axioms A1–A5
would be sufficient to characterize dual theory theory of choice. However, since our goal
is to study utilities that can also compare unbounded lotteries, we need to introduce an
additional axiom A6 which states that each lottery (with a finite mean) has a finite certainty
equivalent. In the original Yaari’s setting A6 is automatically satisfied: a =0 and
A=1. Notice that A6 does not require a and A to be the same for all lotteries. The axiom
A6 is sufficient to extend the dual theory of choice to the case of unbounded lotteries.
In order to obtain the representation result, we first describe the certainty equivalent
functional.
Proposition 1: If A1–A6 hold, there exists a functional U that assigns a real number
(‘certainty equivalent’) to any random variable (with a finite mean) such that X  Y if
and only if U(X) ≥ U(Y ). For any real a holds U(C(a)) = a. If random variables X and
Y are comonotonic, U(X + Y ) = U(X) + U(Y ), where U(·) is the certainty equivalence
functional. Also, for any random variable X and any real α ≥ 0 U(αX) = αU(X).
All Proofs are provided in the Appendix unless stated otherwise.
Let us also define the certainty equivalent functional for inverse distribution functions.
Definition 2: Consider a preference relation  that satisfies axioms A1–A6. A functional
V(·) that assigns a real number to any function  ∈ M is said to be a certainty equivalent
for inverse distribution functions if for every random variable X with finite mean V(X ) =
U(X).
According to A1, the definition is consistent: if two variables have the same IDF then
they are equivalent, hence the value of certainty equivalent is the same. Therefore X  Y
is equivalent to V(X ) ≥ V(Y ).
Lemma 1: Preference relation satisfies A1–A6 if and only if the corresponding certainty
equivalent functional defined on IDFs is a linear continuous functional on M i.e. for all
inverse distribution functions 1,2 ∈ M and non-negative real α:
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 121
• V(α1) = αV(1).
• V(1 + 2) = V(1) + V(2).
• If 1(p) ≡ 1 for all p then V(1) = 1.
Comment. We could have also replaced ‘for any non-negative α’ with ‘for any real α’
and the statement would still be the same because there is no negative α such that both 1
and α1 belong to M (inverse distribution functions must be non-decreasing).
Now we can easily prove the representation result.
Theorem 1: Preference relation satisfies A1–A6 if and only if the corresponding certainty
equivalent functional defined on IDFs can be represented in the following form:
V() =
 1
0
h(p)(p) dp (1)
where h(p) ∈ L∞[0, 1], h(p) ≥ 0 and
 1
0 h(p) dp = 1.
The idea of the proof is to apply a well-known result of functional analysis that the space
of linear continuous functionals is isomorphic to the conjugate space.3 The weight function
h(p) must belong to L∞
(0, 1) because we allow for all random variables with finite means.
If we allowed only for random variables with both finite mean and variance we would have
h(p) ∈ L2(0, 1) instead (L2(0, 1) is the conjugate space for L2(0, 1)).
What is the economic interpretation of (1)? The representation (1) emphasizes that Yaari
utility is a linear rank-dependent utility, i.e. it maximizes a weighted average of payments
in different states of nature with weights being a function of the rank of each state. Indeed,
since (p) is a monotonic function, the states are ranked in the order of increasing payoffs,
and p is precisely the rank of each state. The weight function h(p) shows how much weight
is given to the relatively good states and to the relatively bad ones.
Remark: If FX (x) is continuous then utility may also be represented in terms of cumulative
distribution function
U(X) =
 +∞
−∞
h(FX (ξ ))ξdFX (ξ ). (2)
Note that by construction this representation gives utility which is properly defined for
all random variables with finite means including the unbounded ones. In the meantime,
Yaari’s representation
U =
 +∞
−∞
g(1 − FX (ξ )) dξ (3)
is defined only for random variables that are uniformly bounded. Indeed, if X is unbounded
then
 +∞
−∞
g(1 − FX (ξ )) dξ =
 0
−∞
g(1 − FX (ξ )) dξ +
 +∞
0
g(1 − FX (ξ )) dξ
122 SERGEI GURIEV
is finite only if g(0) = g(1) = 0. But monotonicity axiom requires that g is non-decreasing.
Thus the only functional (3) defined for unbounded random variables is the trivial one g ≡ 0.
For uniformly bounded random variables, though, integration by parts converts (2) to
Yaari’s form and back (which has been done in Roell [1987] and Demers and Demers
[1990]. Indeed, suppose that we only consider lotteries with X ∈ [0, 1] with probability 1.
Then
 1
0
g(1 − FX (ξ )) dξ = g(0) +
 1
0
g
(1 − FX (ξ ))ξdFX (ξ ).
Taking
h(p) = g
(1 − p) (4)
we obtain (2).
2.3. Risk-aversion
Similarly to Yaari’s characterization of risk aversion in DTC [Yaari, 1986, 1987], we shall
determine conditions on h(·) for utility (1) to be risk-averse.
In order to define risk aversion, we use Rotschild-Stiglitz concept of mean-preserving
spread. Consider arbitrary random variable X (with finite expected value) and an uncorrelated
noise ξ(E(ξ | X)=0). Then X +ξ is a mean-preserving spread of X and therefore
risk-averse agents should prefer X to X +ξ .
Definition 3: Utility U(·) is said to be risk-averse if U(X) ≥ U(X + ξ) for all X and ξ
such that E(ξ | X)=0.
It is well known that if neutrality axiom A1 holds, such definition of risk-aversion can
be re-written in terms of distribution functions. Utility functional is risk averse if for any
random variables X and Y with the same mean EX = EY the following is true: if inequality
 a
−∞
FX (x) dx ≤
 a
−∞
FY (y) dy (5)
holds for any real a then U(X) ≤ U(Y ).4
Let us reformulate risk aversion in terms of IDF. Condition (5) is equivalent to
X (p) − a ≤ Y (p) − a. (6)
Theorem 2: The utility functional (1) is risk averse if and only if the weight function h(p)
is non-increasing.
This result is perfectly consistent with Yaari’s characterization of risk aversion: the utility
is risk averse whenever the distortion function g(·) is convex [Yaari, 1986, 1987]. Since for
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 123
uniformly bounded random variables the weight function is related to the first derivative of
g (see (4)), the convexity of g(·) is equivalent to monotonicity of h(·).
The result is very intuitive. Indeed, take a lottery X and consider a mean preserving
spread Y , that pays the same on average, but pays less in bad states and more in good states.
The utility functional is risk averse U(X) ≥ U(Y ) if and only if the loss from getting less in
the bad states is greater than the benefit from getting more in the good states. This implies
that risk aversion is equivalent to larger weights for bad states and smaller weights for good
states. Hence, the weight function h(p) must be non-increasing.
3. Microfoundations of dual theory of choice
In this Section we provide ‘microfoundations’ for the dual theory of choice. In this paper,
microfoundations are understood as a model with rational choice that explains why an agent
who is originally risk-neutral may end up having a risk averse dual utility. In the expected
utility world, the story may go as follows: consider an agent that maximizes expected
income and faces financial market imperfections, i.e. the costs of borrowing are higher than
interest rate on savings. In this situation the agent would prefer less risky alternatives. This
framework may also explain risk loving preferences. Indeed, consider a situation where the
agent is not punished at all for negative income. Under limited liability, firm’s owners get
zero income if firms’ profit is negative but get all the profit whenever it is positive. Then
the owners would prefer more risky alternatives: in the good states their return is very high
while in the bad state the punishment is bounded from below. In this Section we build a
similar model for DTC preferences.
3.1. The basic model
We consider a simple two-period model of a risk-neutral agent that faces a bid-ask spread in
the credit market. Namely, the interest rate the agent pays on her loans is Rl while she can
only save at the risk-free rate Rs < Rl . The agent is neutral to risk and maximizes expected
discounted withdrawn earnings in periods 0 and 1 with the discount rate . The agent
has constant marginal utility so that the model below is applicable to firms rather than to
households.5
In order to rule out trivial cases, we assume (all rates are gross)
Rs <  < Rl . (7)
The agent is to evaluate a lottery (e.g. a risky project) that pays X in the period 1.
The distribution function of payoffs F(·) is such that the expected value of X is finite:
|EX| = |
 +∞
−∞ xdFX (x)| = |
 1
0 X (p) dp|<∞.
The agent chooses to withdraw C in the first period. If X >C then the agent saves
S = X −C and therefore the second-period payoff will be Rs(X − C). If X <C then the
agent will have to borrow L = C − X and will have to pay Rl (C − X) in the second period.
124 SERGEI GURIEV
The expected present discounted value of the project is6
U(X) =
 +∞
−∞
dFX (x){C + 
−1(Rs [x − C]+ − Rl [C − x]+)}
=
 1
0
dp{C + 
−1(Rs [X (p) − C]+ − Rl [C − X (p)]+)} (8)
It is natural to assume that the agent chooses C to maximize (8). To describe the choice of
C, we need to specify, however, whether the agent chooses C knowing the realization of X
or at least its distribution. We shall calculate the agent’s evaluation of the project U(X) for
three scenarios.
First, we will consider the case of extreme flexibility where the first period spending
decision C is taken after the realization of X is observed (timeline (a) in figure 1). The
opposite extreme is to assume that the agent cannot vary C at all (timeline (b) in figure 1).
Suppose that C = C∗ is exogenously set by someone else (or by the agent herself but before
she even learns FX (·)). The third case is intermediate: the agent can vary C but has to take
this decision before X is observed (timeline (c) in figure 1). In this case C cannot depend
on X but can depend on the distribution of X.
Figure 1. Agent’s evaluation of a lottery X depends on timeline. Under timeline (a) the agent chooses first-period
withdrawals C after observing realization of X. In this case she is risk-neutral and prefers a lottery with higher EX.
In the case (b) where the agent evaluates the lottery after C is set, her preferences are described by a risk-averse
expected utility. If the agent chooses first-period withdrawals C before knowing realization of X (case (c)) she is
risk-averse and maximizes a Yaari utility (9).
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 125
In the first case (timeline (a)), the solution is straightforward: take C(X) = X. Indeed,
(7) implies that saving pays off too little, so that the agent prefers consumption to saving;
borrowing is too costly, so that borrowing to consume is not rational. SubstitutingC(X)= X
into (8), we obtain U = EX—agent remains risk-neutral.
The second case (timeline (b)) is very different. SubstitutingC = C∗ into (8) we obtain an
expected utility similar to those discussed in Greenwald and Stiglitz [1990], Eeckhoudt et al.
[1997]. Indeed, agent gets expected value of a concave piecewise-linear utility function with
a kink at X =C∗. The agent maximizes Eu(X) where u(X)= Rs [X −C∗]+ − Rl [C∗ − X]+.
Eeckhoudt et al. [1997] analyze such preferences in detail and showthat they have first-order
risk aversion.
The innovation of our paper is to study the third, intermediate case (timeline (c)) where
the choice C depends on distribution but not on realization of X. In this case, the agent’s
evaluation of the lottery is
U(X) = max
C

C − Rl

 C
−∞
(C − x) dFX (x) + Rs

 ∞
C
(x − C) dFX (x)

. (9)
Proposition 2: The solution C to maximization problem (9) satisfies7
FX (C − 0) ≤ ( − Rs)/(Rl − Rs) ≤ FX (C). (10)
If FX (·) is continuous, (10) takes the form
FX (C) = ( − Rs)/(Rl − Rs) (11)
Proof trivially follows from the first- and second-order conditions.
Formulas (10) and (11) implies C =X (p∗
), where p∗ =(− Rs)/(Rl − Rs). The firstperiod
spending is simply the value of inverse distribution function of the project’s payoff
taken at a given point (− Rs)/(Rl − Rs). Therefore if project’s payoff increases by a
dollar in all states of nature, C also goes up by one dollar. If payoffs in all states double,
C doubles as well. If there are two comonotonic projects, the corresponding values of the
first-period consumption add up. Thus, the first period consumption C =X (p∗
) is a linear
functional defined on the lotteries so there is little wonder than overall evaluation of the
project U(·) is a Yaari functional.
Proposition 3: Formula (9) defines a DTC utility functional (1) with the weight function
h(p) =

RB if p < p∗
RG if p ≥ p∗ (12)
(see figure 2), where RG = Rs/, RB = Rs/, p∗ =(1− RG)/(RB − RG).
The formula (9) describes the simplest possible risk-averse DTC functional. The agent
maximizes a weighted average of payoffs with higher weights RB for bad states of nature
(p < p∗
) and lower weights RG for good states of nature. The only weight function that
126 SERGEI GURIEV
Figure 2. The simplest risk averse DTC utility (12).
is simpler than (12) is the one which takes only one value in all states h(p) ≡ 1, but the
corresponding utility is risk neutral.
The statement opposite to Proposition 3 is also true. For an arbitrary risk-averse DTC
utility functional (1) such that h(p) takes only two values, there exist such Rl and Rs that
this functional can be derived as the value function in (9). This class of utilities is the set
of simplest possible risk-averse rank-dependent utilities that are defined as follows: the
states are ranked in the order of payoffs and divided into the set of the ‘good’ ones and
that of the ‘bad’ ones. The agent’s payoff of getting a dollar in a good state is RG, while
that of getting a dollar in a bad state is RB > RG. Such utilities are fully characterized by
two parameters. The first one is the cutoff p∗ between the ‘good’ states of nature and the
‘bad’ ones. The other parameter is the relative difference of utility of getting a dollar in
a good state and that of getting a dollar in a bad state r = RG/RB. Using normalization
1=
 1
0 h(p)dp = RB p∗ +(1− p∗
)RG we obtain (12).
Let us compare the results for the three scenarios. If the agent choosesC after she learns X,
(case (a)) she adjusts C accordingly and remains risk neutral U =EX. If she can only adjust
C before she learns X (case (c)), then she takes a risk that she would need to borrow in some
states and save in others. Since financial market is imperfect, she gets less than EX.However,
by adjusing C, the agent minimizes the losses due to financial market imperfections given
the distribution. If she cannot adjust C at all (case (b)) then the evaluation decreases even
further away from the expected value EX. Since C =C∗ is fixed, the utility is less than in
case (c) (unless C∗ luckily coinsides with the optimal choice X (p∗
) which is not a generic
case).
The cases (b) and (c) are dual in terms ofYaari’s duality between payoffs and probabilities.
In both cases, the agent calculates a weighted average of payoffs, the weights are the same
but the concept of weighing is different. In both cases the agent assigns high weight RB
to each dollar received in a bad state and low weight RG to each dollar received in a good
state. The difference is in the definition of the good vs. bad states. In the dual utility case
(c) the cutoff is based on probability: the bad state is defined as the state with p < p∗ (and
the states with p > p∗ are good ones). In the expected utility case (b) the cutoff is bases on
payoffs: bad states are the ones with X < C∗. The two functionals are equal to each other
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 127
whenever C∗ =X (p∗
) (or p∗ = FX (C∗
)) but this coincidence occurs only for a very small
subset of lottery space.
Besides the capital market imperfections, the analysis above can also be applied to an
agent operating under a piece-wise linear tax schedule. Consider an agent that faces, positive
tax rate for positive net cash flows and zero tax for negative cash flows (‘no-loss-offset’
rule). The model would identical to the one above. The discontinuity in the tax rate makes
effective RG lower: RG = (1 − τ)RB, where τ is the tax rate for positive profits.
The class of the simplest DTC utilities (12) is narrow but it contains preferences that
are consistent with Allais paradox [Allais, 1979]. One can easily check that the utility
(12) is consistent with common consequence effect (as defined in Allais [1979], Starmer
[2000] whenever RG <1/1.39, 0.89RB +0.5RG >1, RB >40RG. Such RB and RG apparently
exist. Also, utility (12) is consistent with common ratio effect if p∗ ∈ (1/4, 3/4)
and RB >1.25, e.g. RB =3/2, RG =1/2, p∗ =1/2.
3.2. The dynamic extension
The model (c) in the previous Subsection may not seem robust to extending the horizon
beyond the period t =1. It turns out, however, that one can set the model in the infinite
horizon framework and still obtain the Yaari utility (12). Suppose that the agent receives
uncertain payoffs Xt at each period t. The payoffs Xt are independent identically distributed
random variables with distribution function FX (·) known to the agent. The expected value
of X is finite: |
 +∞
−∞ xdFX (x)|<∞.
At each period t, the agent can hold non-interest bearing cash Mt and invest in risk-free
bonds St with gross interest rate Rs . The agent can also borrow Lt at a gross interest rate
Rl . The bonds have one-period maturity and the loans must be repaid next period, too. In
addition to (7), we assume that the return on bonds is greater than one on money
Rs > 1. (13)
The agent maximizes expected discounted cash flows Ct :
v = E
∞
t=0

−tCt , (14)
by choosing Mt+1 ≥ 0, Lt+1 ≥ 0, St+1 ≥ 0 and Ct under the following constraint
Mt+1 = Mt − Rl Lt + Rs St + Lt+1 − St+1 + Xt − Ct (15)
initial conditions
M0 = M ≥ 0, L0 = L ≥ 0, S0 = S ≥ 0,
and the no-Ponzi-game condition
Prob

lim
t→∞

−t Lt = 0

= 1. (16)
128 SERGEI GURIEV
The agent chooses whithdrawals Ct before realization of Xt and therefore can only rely
upon information on state variables Mt , Lt , St . In the meanwhile, the choice of Mt+1, Lt+1,
St+1 is made after revelation of Xt and may depend upon Xt as well as on Mt , Lt , St. We
do not impose non-negativity constraint on Ct . If the agent faces low Mt , St low expected
value of Xt or high indebtedness Lt , then she has to plan losses (or try to attract additional
capital e.g. via issue of new equity or sales of assets).
Proposition 4: If (7), (13) hold, the stochastic programming problem (14)–(16) has a
solution. The value of the objective function (14) as a function of the initial conditions
satisfies the Bellman equation
v(M, L, S) = max
C

C + 
−1EX max
M,L,S≥0
v(M
, L
, S
)

(17)
where the inside maximum is taken subject to M = M − Rl L + Rs S + L − S + X − C
and the random variable X has a c.d.f. FX (·).
Proposition 5: For both stochastic programming problem (14)–(16) and dynamic programming
problem (17) there exists the same unique solution which is as follows. The
Bellman function is:
v(M, L, S)= M − Rl L + Rs S +U(X)/( − 1)
where U(X) is given by (9). The control variables are
M =0, L =[ ¯C − X]+, S =[X − ¯C ]+,C = ¯C + M − Rl L + Rs S, (18)
where ¯C is determined from condition
FX ( ¯C −0) ≤ (− Rs)/(Rl − Rs) ≤ FX ( ¯C ). (19)
If FX (·) is continuous, (19) is becomes
FX ( ¯C )=(− Rs)/(Rl − Rs ).
Thus, the Bellman function in a model with infinite horizon is the sum of combination
of state variables M − Rl L + Rs S and the present value of getting DTC utility (9) every
period ad infinitum.
4. Evaluation of lotteries under non-linear financial contracts
A natural question emerges how broad is the class of DTC functionals that can be derived as
an indirect utility of an agent that faces financial imperfections. Formally,we can ask whether
some other DTC utilities can be derived if we introduce arbitrary schedules of interest rates
as functions of amount invested/borrowed. Suppose that the gross interest payments are
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 129
I (ξ) where ξ is the net savings (negative if borrowings),  is the gross discount rate and
I (ξ ) is the interest rate schedule normalized by discount factor. In the previous Subsection
we considered the simplest case of bid-ask spread I (ξ ) = ¯ I0(ξ ; RB, RG) where
¯ I0(ξ ; RB, RG) =

RGξ if ξ ≥ 0.
RBξ if ξ < 0
(20)
Firms (as well as households) often face interest rates that not only differ for positive
and negative net financial position but also depend on amounts saved or borrowed. Let us
extend our analysis to the case of arbitrary I (ξ ). The expected discounted earnings are U =
C +
 +∞
−∞ I (x −C)dFX (x). Again, the timeline is crucial. If the agent chooses C knowing
X (similar to (a) in figure 1) then the agent remains risk-neutral U = EX+maxξ {I (ξ )−ξ }.
If C cannot be varied at all (case (b) in figure 1), then agent evaluates the projects according
to an expected utility functional. The agent maximizes expectation of non-linear utility
function
u(x) = I (x − C). (21)
Apparently, such an agent is risk-averse whenever I (ξ ) is concave.
If agent chooses C before observing X (case (c)) then the expected discounted earnings
are
U(X) = max
C

C +
 1
0
I (X (p) − C) dp

, (22)
where X (·) is the IDF of X. We already know that for a particular case (20) with
RG <1< RB the utility (22) is a Yaari one with the weight function (12). The problem
that we will address now is whether there exist other interest rate schedules that generate
Yaari’s utility. The answer to this question is negative. It turns out that no other DTC
functional can be derived as an indirect utility in the maximization problem (22).
Theorem 3: Let I (ξ ) be continuous and differentiable on (−∞,+∞) except maybe a
countable set. Let us also assume that there exist ( finite or infinite) limits R+ = limξ→∞
I (ξ )/ξ and R− = limξ→−∞ I (ξ )/ξ . Then utility (22) is a DTC one (1) if and only if
R+ ≤ 1 ≤ R− and there exists a real number b such that I (ξ ) = ¯ Ib(ξ ; R−, R+) where
¯ Ib(ξ ; R−, R+) =

b + R+(ξ − b) if ξ ≥ b
b + R−(ξ − b) if ξ < b.
(23)
The Theorem complements the results obtained in the Subsection 3.1, where we showed
that all risk-averse DTC utilities (1) with a weight function that takes only two values RB
and RG can be derived as preferences of a risk-neutral agent that faces a bid-ask spread
in the interest rates. Theorem 3 essentially says that no other DTC utility can be derived
in such a model. No interest rate schedule that satisfies the technical conditions of the
130 SERGEI GURIEV
Theorem can result in a DTC utility except the two-rate interest schedule (23) which is
essentially equivalent to the bid-ask spread (20) we have already studied. Althouth (23)
seems to be more general than (20) (which is only a partial case at b = 0), the DTC utility
obtained by substituting (23) into (22) does not depend on b. Indeed, ¯ Ib(ξ ; R−, R+) =
b + ¯ I0(ξ − b; R−, R+). Hence, substituing ˜C = C + b, we obtain
U(X) = max
C
C +b+
 1
0
I (X (p)−C −b)dp = max
˜C
˜C
+
 1
0
I (X (p)− ˜C ) dp
which is the same DTC utility as in the Subsection 3.1 with RG = R+ and RB = R−.
It is also worth emphasizing that only risk-averse utilities have microfoundations. The
DTC utilities with the weight function (12) and RG > RB cannot be derived in the form (9).
Whenever R+ > R−, there is no finite solution to the maximization problem in (22).8 This
is again very different from the expected utility case: (21) implies that for every expected
utility functional there exist an interest rate schedule. In particular, every risk loving expected
utility can be generated by a convex I (·) that can be found from (21).
5. Conclusions
The main contribution of the paper is to show that a certain class of dual utilities can be
obtained as an indirect utility of a risk-neutral agent that faces a bid-ask spread in the
credit market. These utilities are the simplest risk-averse rank-dependent utilities which are
parameterized by two numbers: the cutoff point that separates the ‘good’ and the ‘bad’ states
of nature and the relative penalty for being in a bad state. The model is simple: the agent
has to take the decision on the first-period withdrawals before she learns the realization of
stochastic returns. If she wants to withdraw too much today, she has to borrow which will
result in high interest payments and therefore low consumption tomorrow. If she withdraws
too little, she would have to save and get returns tomorrow at a relatively low deposit rate. It
is the agent’s ability to adjust the first-period withdrawal knowing the distribution of payoffs
but not the actual realization that makes her evaluation of the lottery a risk-averse dual utility
functional. If the agent were not able to adjust, her preferences would be described by a
risk-averse expected utility. Vice versa, if the agent were able to adjust the withdrawals after
observing the realization, she would remain risk-neutral.
We also show that no DTC utility outside this particular class has such microfoundations
even if we allow for arbitrary non-linear financial contracts. Only simplest risk-averse
rank-dependent utilities can be obtained as a solution to an optimization problem. This
result reveals a striking difference between expected utility and dual theory of choice. Any
expected utility can be derived as preferences of an agent under financial imperfections in
the model where agent cannot vary the first-period consumption. In the world where agent
adjusts her first-period consumption, things are very different. Under arbitrary non-linear
interest schedules, the agent still has a constant absolute risk aversion, but not necessarily
a constant relative risk aversion. It turns out that the only interest rate schedule that results
in a DTC utility is the bid-ask spread studied above where the agent pays constant interest
rate on loans and gets a constant though lower rate on deposits.
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 131
Another contribution of the paper is generalization of the dual theory of choice for
unbounded lotteries with finite means. The original Yaari’s formulation cannot be extended
to the unbounded case. The representation form that we get is rather similar to that of Roell
[1987] though the latter was also obtained for uniformly bounded lotteries.
We believe that our results justify the growing popularity of DTC as an appropriate
(and very simple) tool for analysis of firms’ decision-making under risk. Being tested on
individuals, DTC has performed rather poorly. However, it has not yet been empirically
tested on firms or banks. There is some evidence that firms’ and banks’ are risk-averse.
Rose [1989], Davidson et al. [1992], Park and Antonovitz [1992] and many other authors
prove that firms indeed seek insurance and diversification. On the other hand, only riskaverse
expected utility has been tested on firms; it remains unknown whether DTC would
perform better or worse in such tests.
Appendix
Proofs
Proof of Proposition 1. Let us first prove that the certainty equivalent exists. Take an arbitrary
random variable X. According to A6, there exist such real a, A thatC(a) ≺ X ≺ C(A).
Continuity (A3) and monotonicity (A4) imply that there exists such real ξ ∈ (a, A) that
C(ξ )∼ X. Hence, for each X we have found the certainty equivalent ξ =U(X). By definition,
if a random variable takes the same value in all states of nature, its certainty equivalent
equals this value: U(C(ξ ))=ξ .
Proving U(αX)=αU(X) is trivial. Since X ∼ C(U(X)), A5 implies αX ∼ C(αU(X))
for every α ∈ [0, 1] (take Y =C(U(X)) and Z =C(0)). Hence U(αX)=αU(X) for every
α ∈ [0, 1]. Now take arbitrary X and α>1. Since U( 1
α αX)= 1
αU(αX) (indeed, 1
α <1),
we obtain U(αX)=αU(X).
The last step is to prove that certainty equivalence is additive. Let us first prove that for any
real b, U(X +b) = U(X)+b. A5 implies that since X ∼ C(U(X)), then X+b
2
∼ C(U(X))+b
2
(take Y = C(U(X), Z = C(b), α = 1/2). Hence U( X+b
2 ) = U( C(U(X))+b
2 ) = C(U(X))+b
2 .
Using U(αX) = αU(X) we obtain U(X + b) = U(X) + b. Now consider arbitrary
comonotonic X and Y . Again, X ∼ C(U(X)) implies X+Y
2
∼ C(U(X))+Y
2 . Hence U(X +
Y ) = U(C(U(X)) + Y ). We just proved U(C(U(X)) + Y ) = U(X) + U(Y ), therefore
U(X + Y ) = U(X) + U(Y ). ✷
Proof of Lemma 1. First if Y = αX then Y = αX and if X and Y are comonotonic
then X+Y = X + Y . So whenever certainty equivalence functional defined on IDFs is
linear the preference relation satisfies A1–A6.
To prove the ‘only if’ part we need to use neutrality axiom A1. Consider arbitrary
1,2 ∈ M, 2 = α1. Take any random variable X such that X = 1 and then
consider random variable αX. By definition 2 = αX . Then for any random variable Y
such that Y = 2 = αX neutrality requires U(Y ) = U(αX) = αU(X).
Similarly, consider 1,2,3 ∈ M : 3 = 1 + 2. Then take any Z : Z = 3
and define X and Y in the following way: for every real z ∈ [3(0),3(1)] find ˆ p(z) such
132 SERGEI GURIEV
that 3( ˆ p) = z. Then X takes value 1( ˆ p(z)) and Y takes value 2( ˆ p(z)) whenever Z
takes value z. Thus we have obtained comonotonic X, Y such that X = 1, Y = 2
and X + Y = Z. For them U(X + Y ) = U(Z) = U(X) +U(Y ). Due to A1, this will also
be true for all X, Y, Z such that X = 1, Y = 2, Z = 3. ✷
Proof of Theorem 1. The utility functional V is defined and linear on M. Let us extend
it onto the whole space L1[0, 1]. For any function g ∈ L1(0, 1) there exists a representation
g = limn→∞ gn, where gn = 1n
− 2n
, and 1n
,2n
∈ M. Indeed, the set of
continuous functions is dense in L1(0, 1); the set of polynomials is dense in the set of continuous
functions, and every polynomial can be represented as a difference of two monotonic
functions.
Define V(g) = limn→∞ V(1n
)−V(2n
). Due to linearity of V over M there is no ambiguity:
V(g) does not depend upon representation of g. Let g = limn→∞ gn = limn→∞ ˜ gn,
where gn = 1n
−2n
, and ˜ gn = ˜
1n
− ˜
2n
. Then gn − ˜gn = 1n
−2n
− ˜
1n
+ ˜
2n
 → 0.
Grouping the terms and using continuity of V over M we obtain V(1n
+ ˜
2n
) − V(˜
1n
+
2n
) → 0. Then applying linearity of V on M and re-grouping terms back we get (V(1n
)−
V(2n
)) − (V(˜
1n
)−V(˜
2n
)) → 0. Thus, every utility built upon a preference relation that
satisfies A1–A6 generates a linear continuous functional on L1[0, 1]. It is well known [see
Yosida, 1978] that the set of all such functionals is isomorphic to conjugate space L∞[0, 1]
i.e. each utility functional can be represented as
 1
0 h(p)(p)dp where h ∈ L∞[0, 1].
Monotonicity implies h(p) ≥  0 and certainty equivalence axiom requires normalization 1
0 h(p)dp = 1. ✷
Proof of Theorem 2. The proof basically follows Yaari [1987]. First let us consider two
random variables with IDF 1(p) and 2(p) such that
 1
0 1(p)dp =
 1
0 2(p)dp (equal
means) and for any a holds 1(p) − a2(p) − a. Consider (p) = 1(p) − 2(p).
Let us now prove that K(q) =
 q
0 (p)dp ≥ 0 for any q ∈ [0, 1].
Then let us divide [0, 1] into intervals (pk , pk+1) on which (p) has same sign ((p)
changes sign at least once). Then for every sign change point pk there exist ak such that
1,2(p) ≥ ak whenever p > pk and 1,2(p) ≤ ak whenever p < pk . Then using the risk
aversion condition we obtain 0 ≥ 1(p)−ak−2(p)−ak = 2(
 pk
0 ((1(p)−ak)−
(2(p) − ak))dp), i.e. K(pk) ≥ 0. Similarly, K(pk+1) ≥ 0. Furthermore, for an arbitrary
p ∈ (pk , pk+1) we have either K(pk) ≤ K(p) ≤ K(pk+1) (if (p) is non-negative on
(pk , pk+1)) or K(pk+1 ≤ K(p) ≤ K(pk) (if (p) is non-positive on (pk , pk+1)).
Now when we have proved that K(q) ≥ 0 for all q ∈ [0, 1], we may integrate by
parts: V(1) − V(2) =
 1
0 h(p)(p)dp =
 1
0 h(p)dK(p) = h(p)K(p)|10
−
 1
0 K(p)
dh(p)≥0.
The non-integral term is zero since K(0) = K(1) = 0. The integral term is non-positive
(since h(p) is monotonic, dh(p) is a non-positive measure and since K(p) is an absolutely
continuous function, the integral exists and is non-positive).
Let us now assume that there is a set P ⊂ [0, 1] of non-zero measure on which a risk
averse h(p) is strictly increasing. Let us divide the set into two subsets of equal measure:
P = P1 ⊕ P2,

P1
dp =

P2
dp, such that p1 ≤ p2 for all p1 ∈ P1, p2 ∈ P2. Consider two
random variables ξ and η : ξ (p) = η(p) for all p / ∈ P, ξ (p) = 1/2 for p ∈ P and
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 133
η(p) = 0 if p ∈ P1 and Hη(p) = 1 if p2 ∈ P2. Then ξ (p) − a ≤ η(p) − a
for any real a. Risk aversion implies U(ξ ) ≥ U(η). Hence 12

P h(p)dp ≥

P2
 h(p)dp, or
P1
h(p)dp ≥

P2
h(p)dp. This means that h(p) can not be strictly increasing on P. ✷
Proof of Proposition 3. For a given distribution, let us calculate the value of (9) substituting
C = X (p∗
):
U(X) = X (p∗
)− RB
 P∗
0
(X (p∗
)−X (p)) dp + RG
 1
p∗
(X (p)−X (p∗
)) dp
=
 p∗
0
RBX (p) dp +
 1
p∗
RGX (p) dp =
 1
0
h(p)X (p)dp.
where the weight function h(·) is described by (12). ✷
Proof of Proposition 4. The partial sum of (14) is as follows.
T
t=τ
τ−tCt = {Mτ − Rl Lτ + Rs Sτ} +
T
t=τ
τ−t Xt
−( − 1)
T
t=τ+1
τ−tMt − (Rl − )
T
t=τ+1
τ−t Lt
−( − Rs)
T
t=τ+1
τ−t St + τ−T LT+1
−τ−T (MT+1 + ST+1)
Under assumptions about interest rates (7) and no-Ponzi-game condition (16) expected
value of

T
t=τ τ−tCt+1 is bounded from above by Mt − Rl Lt + Rs St + const. Therefore
the problem has a solution and the Bellman function is well-defined. Let us define a
state Zt = {Mt , Lt , St }. Then the agent’s strategy is a quadruple of Borel functions  =
{ ˜M (t, Z, X), ˜L (t, Z, X), ˜S (t, Z, X), ˜C(t, Z)}, where ˜M (t, Z, X), ˜L (t, Z, X), ˜S(t, Z, X)
are non-negative and for all M, L, S and X holds ˜M (t, {M, L, S}, X) = M + X − Rl L +
Rs S + ˜L (t, {M, L, S}, X) − ˜S (t, {M, L, S}, X) − ˜C (t, {M, L, S}). The strategy defines a
Markov process Zt+1 = { ˜M (t, Zt , Xt ), ˜L(t, Zt , Xt ), ˜S(t, Zt , Xt )} = G(t, Zt , Xt ) and a
series of random variables Ct = ˜C (t, Zt ). Therefore for each strategy  and initial conditions
Z one can calculate the expected net present value v(τ, Z) = E{
T
t=τ τ−tCt |Zτ =
Z}. We have just shown that v is bounded and the maximization problem max J(τ, Z)
is well-defined. The Kolmogorov equation for v(t, Z) is as follows:
v(t, Z) = ˜C (t, Z) + 
−1 J(t + 1, G(t, Z, X)).
134 SERGEI GURIEV
By definition the optimal strategy ˆ satisfies vˆ(t, Z) ≥ v(t, Z) for all t, Z and .
Therefore it also satisfies the Bellman equation
J(t, Z) = sup
(t,·)
˜C
(t, Z) + 
−1v(t + 1, G(t, Z, X)).
Apparently, time index t can be omitted. Substituting for Z and , we obtain (17). ✷
Proof of Proposition 5. In order to solve the Bellman equation, we shall introduce C =
C − (M − Rl L + Rs S). The Bellman equation can then be re-written as
v(M, L, S) = M − Rl L + Rs S + max


 + 
−1EX max
M
,L
,S≥0
M=L−S+X−
v(M
, L
, S
)

.
The Bellman function is therefore linear:
v(M, L, S) = M − Rl L + Rs S + ( − 1)
−1U,
where
U(X) = max
¯C

¯C
+ 
−1EX max
L
,S≥0
L−S+X− ¯C≥0
L − S + X − ¯C − Rl L + Rs S

.
According to (7), the expression L − S + Rl L + Rs S increases whenever L and S
decrease by the same amount. Therefore L = [ ¯C − X]+ and S = [X − ¯C ]+. Then
U(X) = max
¯C

¯C
− RB
 ¯C
−∞
( ¯C − x)dFX (x) + RG
 ∞
¯C(
x
−
¯C
)dFX (x)

,
where RB = Rl/>1 and RG = Rs/<1. The first- and second-order conditions imply
(19). ✷
Proof of Theorem 3. First, let us prove that the maximization problem (22) has a finite
solution only if R+ ≤ 1 ≤ R−. Indeed, if R+ > 1, the agent would choose C = ∞and
would get an infinite utility, while if R− < 1, infinite utility is achieved by taking C = −∞.
All DTC utilities (1) have both constant absolute and constant relative risk-aversion.
We shall prove now that in order for (22) having a constant absolute and constant relative
risk-aversion, the interest payment schedule I (ξ ) must be represented in the form (23).
It is clear that (22) has constant absolute risk aversion: for any real b
V( + b) = b = max
C

C − b +
 1
0
I ((p) − (C − b))dp

= b + V().
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 135
Let us now check whether (22) has constant relative risk aversion. For any realα > 0 we
should have α
−1V(α) = V(). Apparently,
α
−1V(α) = max
C

C +
 1
0
α
−1 I (α((p) − C)) dp

. (24)
First, we shall consider I (ξ ) such that for any real non-negative α holds I (ξ) = αI (ξ ).
This implies
I (ξ ) =

R+ξ if ξ ≥ 0
R−ξ if ξ < 0.
If R+ < R− then it is the DTC utility derived in the Subsection 3.1 (it is a particular case
of (23) for a = 0). If R+ = R− = 1 then it is the case of risk-neutrality.
Now we shall see what happens if for some α > 0 and ξ we have I (αξ )  = αI (ξ ). Take
this α and introduce J (ξ ) = α
−1 I (αξ ). Then (24) takes the form
α
−1V(α) = max
C

C +
 1
0
J((p) − C) dp

. (25)
Let CI () and CJ () be solutions of the maximization problems in (22) and (25), correspondingly.
Notice that since R+ ≤ 1 ≤ R− < ∞,CJ () is finite even at α→∞.
Assuming constant relative risk aversion α
−1V(α) = V(), we obtain
CJ () +
 1
0
J((p) − CJ ()) dp = CI () +
 1
0
I ((p) − CI ()) dp.
This equality should hold for any . Consider ˜ sufficiently close to . According to the
envelope theorem,
V( ˜ ) − V() =
 1
0
I 
((p) − CI ())δ(p)dp + O(δ2)
α
−1V(α ˜ ) − α
−1V(α) =
 1
0
J 
((p) − CJ ())δ(p)dp + O(δ2)
where δ = ˜ −  is small. Hence,
 1
0 J 
((p) − CJ ())δ(p)dp =
 1
0 I 
((p) −
CI ())δ(p) dp + O(δ2). Since the set of all possible deviations δ(p) is sufficiently
rich we should have J 
(ξ ) = I 
(ξ + b) almost everywhere. Here b = CJ () − CI ().
The condition J 
(ξ ) = I 
(ξ + b) implies J (ξ ) = J (0) + I (ξ + b) − I (b). Substituting
into α
−1V(α) = V() we obtain J (0) = I (b) − b. Therefore J (ξ ) = I (ξ + b) − b.
Thus, utility (22) has constant relative risk aversion only if for every α > 0 there exists
a real b such that J (ξ ) = α
−1 I (αξ ) = I (ξ + b) − b for all ξ . Let us see what happens if
136 SERGEI GURIEV
α→∞. For a given η > 0
η lim
αξ→+∞
I (αη)
αη
= ηR+ = I (η + b) − b.
Thus, I (η+b) = b+ R+η. Similarly, forη < 0 we have I (η+b) = b+ R−η. Substituting
ξ = η + b, we obtain the formula (23).
We have proved that utility (22) can be a DTC one only if the interest payment schedule
is (23). Let us check whether the condition is also (23) sufficient. Let us take arbitrary b
and R+ ≤ 1 ≤ R−. Then making straightforward calculations one can show that the utility
(22) is the one introduced in the Subsection 3.1, i.e. the utility (1) with the weight function
(12) where RG = R+ and RB = R−. ✷
Acknowledgments
The author is grateful to Drew Fudenberg, Paul Klein, Victor Polterovitch, Igor Pospelov,
Menahem Yaari, two anonymous referees, participants of Winter European Meeting of the
Econometric Society at Liblice, European Meeting of the Econometric Society at Berlin
and seminar participants at Computing Center and CEMI, Russian Academy of Science and
the Central Bank of Russia. Financial support from USIA, IREX and Russian Foundation
of Fundamental Research is acknowledged.
Notes
1. Strictly speaking, L1 is a space of classes of functions which may differ on a set of measure zero. According
to our definition of IDF, for each class we take the representative of the class that is upper semi-continuous or,
which is the same, right semi-continuous. However, the choice of a particular representative function within
the class does not matter.
2. Hereinafter FX (·) and X (·) are cumulative distribution function and inverse distribution function of random
variable X.
3. Yaari’s proof is different. Its idea is to ‘lay Neuman-Morgenstern result on its side’ [Yaari, 1987].
4. As shown in Rotchild and Stiglitz [1970] for such X and Y there exists a noise ξ that meets conditions of
Definition 3 so that Y and X + ξ have the same distribution.
5. Throughout the paper we stick to the conventional choice-under-risk setting: an individual evaluates lotteries
to maximize consumption. On the other hand, the model also describes a firm. For the firm’s case,  would be
internal rate of return (or cost of internal finance), while Rl would be cost of external finance, and Rs outside
investment opportunities.
6. Hereinafter ξ+ = max{ξ, 0}.
7. Hereinafter F(x − 0) = limx→x,x
<x F(x
).
8. The risk-loving utility functionals with RG > RB are similar to the Pangloss value functional in Krugman
[1998].
References
ALLAIS, M. [1979]: in Expected Utility Hypotheses and the Allais Paradox, M. Allais and O. Hagen (Eds.),
Reidel, Dordrecht.
DAVIDSON, W., CROSS, M., and THORNTON, J. [1992]: “Corporate Demand for Insurance: Some Empirical
and Theoretical Results,” Journal of Financial Services Research, 6, 61–72.
ON MICROFOUNDATIONS OF THE DUAL THEORY OF CHOICE 137
DEMERS, F. and DEMERS, M. [1990]: “Price Uncertainty, the Competitive Firm and the Dual Theory of Choice
under Risk,” European Economic Review, 34, 1181–1199.
DOHERTY, N. and EECKHOUDT, L. [1995]: “Optimal Insurance Without Expected Utility: The Dual Theory
and the Linearity of Insurance Contracts,” Journal of Risk and Uncertainty, 10, 157–179.
EECKHOUDT, L., GOLLIER, C., and SCHLESINGER, H. [1997]: “The No-Loss Offset Provision and the
Attitude towards Risk of a Risk-Neutral Firm,” Journal of Public Economics, 65, 207–217.
EPSTEIN, L. and ZIN, S. [1990]: “First-Order Aversion and the Equity Premium Puzzle,” Journal of Monetary
Economics, 26, 387–407.
FISHBURN, P. [1994]: “Utility and Subjective Probability,” in Handbook of Game Theory with Economic Applications,
Vol. 2, North Holland Elsevier, Amsterdam, 1397–1436.
GREENWALD, B. and STIGLITZ, J. [1990]: “Asymmetric Information and the NewTheory of the Firm: Financial
Constraints and Risk Behavior,” American Economic Review, Papers and Proceedings, 80, 160–165.
HADAR, J. and SEO, TAE KUN [1995]: “Asset Diversification in Yaari’s Dual Theory,” European Economic
Review, 39, 1171–1180.
HARLESS, D. and CAMERER, C. [1994]: “The Predictive Utility of Generalized Expected Utility Theories,”
Econometrica, 62, 1251–1289.
HEY, J. and ORME, C. [1994]: “Investigating Generalization of Expected Utility Theory using Experimental
Data,” Econometrica, 62, 1291–1326.
KRUGMAN, P. [1998]: “What Happened to Asia?,” Mimeo, M.I.T.
PARK, T. and ANTONOVITZ, F. [1992]: “Econometric Tests of Firm Decision Making under Uncertainty,”
Southern Economic Journal, 58, 593–609.
ROELL, A. [1987]: “Risk Aversion in Quiggin and Yaari’s Rank-Order Model of Choice under Uncertainty,”
Economic Journal, 97, 143–159.
ROSE, P.S. [1989]: “Diversification of the Banking Firm,” Financial Review, 24, 251–280.
ROTSCHILD, M. and STIGLITZ, J.E. [1970]: “Increasing Risk: I. A Definition,” Journal of Economic Theory,
2, 225–243.
SCHMIDT, U. [1998]: “Moral Hazard and First-Order Risk-Aversion,” Mimeo, University of Kiel.
STARMER,C. [2000]: “The Hunt for a Descriptive Theory of Choice Under Risk,” Journal of Economic Literature,
38(2), 332–384.
VOLIJ, O. [1999]: “Utility Equivalence in Sealed Bid Auctions and the Dual Theory of Choice Under Risk,”
Mimeo, Hebrew University.
YAARI, M. [1986]: “Univariate and Multivariate Comparisons of Risk Aversion: A New Approach,” in Essays in
Honor ofKenneth J. Arrow,W.P. Heller,R. Starr, and D. Starrett (Eds.), Cambridge University Press, Cambridge.
YAARI, M. [1987]. The Dual Theory of Choice under Risk, Econometerica, 55, 95–115.
YOSIDA, K. [1978]: Functional Analysis. Berlin: Springer-Verlag.