Sumit Joshi

Microeconomic Theory

Expected Utility

1 Introduction
In consumer theory under certainty, it was assumed that prices, income,
and all other parameters influencing consumer choice were known with cer-
tainty. Further, the outcome resulting from any consumption or investment
decision by the consumer was deterministic. However, in the real world, a
consumer must make choices in an environment involving either uncertainty
about some parameter influencing the decision problem (for example, the
price of an asset that a consumer wants to purchase tomorrow may not be
known with certainty today, future income required to finance a current pur-
chase may be random) or there may be uncertainty about the consequences
stemming from an action (the returns from choosing a portfolio of stocks
may not be known with certainty today due to fluctuations in the value of
the stocks). In this chapter, we will formalize the choices confronting a con-
sumer in an environment of uncertainty by means of simple and compound
lotteries. We will then develop the foundations of expected utility theory in
order to choose among lotteries.

2 Lotteries
The first step in describing the behavior of the consumer under uncertainty
is to identify the objects of choice facing the consumer. For example,
the objects of choice when studying the behavior of the consumer under
certainty were consumption bundles representing non-negative quantities of
the goods under consideration. Under uncertainty, the objects of choice
are lotteries (or gambles). Lotteries are probability distributions over
outcomes (or consequences). The set of outcomes can be a discrete set
or a continuous interval. In the discrete case, outcomes can take a finite
or countably infinite number of values while in the continuous case they can
take any value on the real line (−∞ +∞) or some subset of it (i.e. they

1
can assume an uncountably infinite number of values).1 We will concentrate
here on the discrete case and assume for simplicity that there are a finite
number of outcomes given by the set  = {1  2    }. At this stage
there is no restriction on the outcomes other than that they are discrete.
They could be monetary rewards, wealth levels, or consumption bundles.

In the discrete case, lotteries can be represented as probability trees. We

start with a simple or single stage lottery which has the probability tree
shown in Figure 1. In a simple lottery, all uncertainty resolves in one stage.

x1
p1
x2
p2
p3 x3

...
Chance
Node ...
pn
xn
Probabilities
Outcome

Figure 1: A Simple Lottery

Here 1  2    are the possible outcomes and 1  2  P

 are the proba-
bilities associated with these outcomes where  ≥ 0 and =1  = 1. The
lottery originates at a node called the chance node. Emanating from the
chance node are branches representing the probabilities of the outcomes.
Associated with each branch is a terminal node representing the outcome.
The simple lottery can be represented by the notation:
 = (1 ◦ 1 ; 2 ◦ 2 ; ;  ◦  ) (1)
Note that an outcome  can itself be thought of as a lottery — a lottery that
yields  with probability one. In the above notation, it can be written as:
 = (1 ◦ 0; 2 ◦ 0; ;  ◦ 1 ;  ◦ 0) (2)
1
A set of objects is countably infinite if the elements of the set can be put in a one-to-
one correspondence with the set of positive integers; otherwise, the set is uncountably
infinite.

2
Such a lottery that concentrates all probability mass on a single outcome is
called a degenerate lottery. From now on  will denote both an outcome
as well as the degenerate lottery,  , that gives  with probability 1. The
set of all simple lotteries on  is denoted as L . Also, since each outcome
is a degenerate simple lottery,  ⊂ L .

3 Probabilities over Outcomes

The probabilities of the various outcomes can be objective in the sense that
they represent events that have been observed repeatedly in the past and
so a consumer can calculate the relative frequencies with which these events
have occurred. There may be cases, however, where events are unique and
have not been observed in the past and therefore an objective probability
measure cannot be calculated. In this case, consumers can assign subjec-
tive probabilities to events based on their perception of the likelihood of
that event occurring. It is quite possible that different individuals may as-
sign different probabilities to the same event based on differences in their
subjective perceptions about the likelihood of the event.

The distinction between objective and subjective probabilities was high-

lighted by F. Knight who used it to distinguish between risk and uncer-
tainty. Risk describes those situations where the probabilities of occurrence
of various events are known or can be calculated objectively by computing
relative frequencies based on past observations of the event. Uncertainty,
on the other hand, describes those situations where probabilities of occur-
rence of various events are not known and, because the events have not
been observed in the past, cannot be computed objectively. They can only
be determined subjectively.

In this chapter we will work with objectively determined probabilities over

the set of outcomes.

4 Preferences over Simple Lotteries

The preferences of an individual over the set of simple lotteries is expressed
by a binary relation º on L . Given any two simple lotteries 1  2 ∈ L ,
1 º 2 means that 1 is at least as good as 2 

3
Preferences º are assumed to satisfy three fundamental axioms. These are
also referred to as the rationality axioms.

Assumption 1 (Reflexivity): Given any 1 ∈ L , 1 º 1 .

In other words, each simple lottery is at least as good as itself.

Assumption 2 (Completeness): Given any two simple lotteries 1  2 ∈

L , either 1 º 2 or 2 º 1 , or both.

Completeness requires that an individual possesses all relevant information

needed to express a preference ranking over every pair of simple lotteries in
L . Note that reflexivity is a special case of completeness.

Assumption 3 (Transitivity): For simple lotteries 1  2  3 ∈ L , if

1 º 2 and 2 º 3 , then 1 º 3 .

Transitivity requires that an individual should make consistent choices.

We will discuss transitivity in more detail later.

Recall that each outcome in the finite set  = {1  2    } can be
thought of as a degenerate lottery. Then under completeness and transi-
tivity of º, we can order the finite number of degenerate lotteries in 
from best to worst according to º.2 We will assume that the outcomes are
indexed such that:
1 º 2 º · · · º −1 º  (3)
It will be convenient to denote the best outcome, 1 , as  and the worst
outcome,  , as  . Therefore  º 2 º · · · º −1 º  . In addition, to
rule out indifference throughout, we will assume that:

Â (4)
2
Start with 1 and 2 . From completeness, either 1 º 2 or 2 º 1 . Suppose
1 º 2 . Let us call 1 the “winner” of this pairwise comparison. We now compare 1
with 3 to determine the winner. (If both 1 º 2 and 2 º 1 hold, then arbitrarily pick
either as the winner.) By transitivity, the winner of 1 with 3 will be at least as good
as 2 as well. Next compare the winner of 1 with 3 with 4 . Continue inductively to
determine the winner in a pairwise comparison in this way until the final comparison is
made with  . The final winner is then the best in the finite set . After removing the
best lottery from , the reasoning can be repeated to determine the second-best element
in . Continuing in this way allows all elements in  to be ranked from best to worst.

4
Indifference between two simple lotteries is denoted by ∼. Given any two
lotteries 1  2 ∈ L , 1 ∼ 2 is interpreted as 1 is indifferent to 2 or
both lotteries are equally good. The indifference relation ∼ can be induced
from the preference relation º as follows:

1 ∼ 2 ⇐⇒ 1 º 2 and 2 º 1 (5)

The properties of the indifference relation can be deduced from the assump-
tions imposed on º. In particular, if º satisfies reflexivity, completeness and
transitivity, then the indifference relation ∼ satisfies reflexivity, transitivity
and symmetry (if 1 ∼ 2 , then 2 ∼ 1 ).3

Strict preference for one simple lottery over another is denoted by Â. For
1  2 ∈ L , 1 Â 2 is interpreted as 1 is strictly preferred to 2 . Strict
preference can be induced from the preference relation º as follows:

1 Â 2 ⇐⇒ 1 º 2 and 2 ² 1 (6)

The properties of the strict preference relation can be inferred from the
assumptions imposed on º. In particular, if º satisfies reflexivity, com-
pleteness and transitivity, then strict preference  satisfies irreflexivity
(1 ¨ 1 ) and transitivity.4

Remark: We started with the preference relation º as the primitive and

then induced ∼ and Â. It is also possible to start with the strict prefer-
ence order  as the primitive. The starting assumptions imposed on strict
preference  are those of asymmetry (for no pair of lotteries is it possible
that 1 Â 2 and 2 Â 1 ) and negative transitivity (if 1 Â 2 , then
3
From reflexivity of %, 1 % 1 and 1 % 1 and it follows that 1 ∼ 1 . Next
consider transitivity of indifference, i.e. let 1 ∼ 2 and 2 ∼ 3 . Then:

1 ∼ 2 ⇐⇒ 1 % 2 and 2 % 1
2 ∼ 3 ⇐⇒ 2 % 3 and 3 % 2

Now from the RHS and the transitivity of % it follows that 1 % 3 and 3 % 1 and
therefore 1 ∼ 3 . The symmetry of ∼ follows similarly.
4
Irreflexivity and transitivity of  can be shown by contradiction. If 1  1 for some
lottery, then by definition 1 º 1 and 1 ² 1 . But 1 ² 1 contradicts reflexivity of º.
Next consider transitivity. Suppose 1 Â 2 , 2 Â 3 but 1 ¨ 3 . Note that 1 ¨ 3
is equivalent to 3 º 1 from completeness of º. Also note that 1 Â 2 is equivalent to
1 º 2 and 2 ² 1 . Now combining 3 º 1 and 1 º 2 yields 3 º 2 from the
transitivity of º. But 3 º 2 contradicts the starting hypothesis that 2 Â 3 .

5
for any third lottery 3 , either 1 Â 3 or 3 Â 2 or both)5 . We can then
induce the preference orders º and ∼ as follows:

1 º 2 ⇐⇒ 2 ¨ 1
1 ∼ 2 ⇐⇒ 1 ¨ 2 and 2 ¨ 1

The two approaches are equivalent. If we start with  satisfying asymme-

try and negative transitivity, then we can induce the preference relation º
satisfying completeness and transtivity. If we start with º satisfying com-
pleteness and transitivity, then we can induce  satisfying asymmetry and
negative transitivity.6 ¥

We close this section with a discussion of transitivity of preferences. Con-

sider what would happen if the preference relation º was intransitive. Sup-
pose therefore that 1 º 2 , 2 º 3 , but 1 ² 3 (i.e. 3 Â 1 ). Suppose
the individual has the lottery 1 . The individual would then pay some small
sum   0 to get 3 instead of 1 . Since 2 is at least as good as 3 , the in-
dividual would once again be willing to pay a small amount 0 ≥ 0 to get 2
for 3 . Finally, since 1 is at least as good as 2 , the consumer would pay
00 ≥ 0 to get 1 for 2 . After foregoing  + 0 + 00  0 the consumer is left
with the same lottery 1 as before. This money-pumping argument can
be repeated continuously till the individual is bankrupt. An individual with
intransitive preferences is open to this form of systematic manipulation.

Another view of transitivity is that it ensures that consumers make con-

sistent (or non-contradictory) choices. Consider once again the example
where the individual’s preferences are intransitive: 1 º 2 , 2 º 3 , but
1 ² 3 . In a direct choice between 1 and 3 , the individual would prefer
3 . Now consider an indirect choice between these two lotteries by way of an
intermediate lottery 2 . Starting with 3 , the individual would be willing
to substitute 3 with 2 , and then 2 with 1 . But this would contradict
the earlier direct choice made by the individual.

5
The statements “ ⇒ ” and “  ⇒   ” are logically equivalent and the
latter statement is called the contrapositive of the former. It can be verified that negative
transitivity of  is the contrapositive of the transitivity of º.
6
We have already seen that negative transitivity of  is equivalent to transitivity of º.
Similarly, asymmetry of  is equivalent to completeness of º (“for no pair of lotteries 1
and 2 is it possible that 1 Â 2 and 2 Â 1 ” is equivalent to “1 ¨ 2 or 2 ¨ 1
for all pair of lotteries 1 and 2 ”. But note that 1 ¨ 2 or 2 ¨ 1 is the same as
2 º 1 or 1 º 2 or both.

6
5 Compound Lotteries
The next step is to develop more complex lotteries. A compound lottery
or multistage lottery is one in which the outcomes themselves are lotteries
(called sublotteries). That is, unlike the simple lottery in which all un-
certainty is resolved in one stage, the uncertainty in a compound lottery is
resolved in multiple stages. For instance, the structure of a two-stage lottery
is shown in Figure 2.

q x

p y
1-q

1-p r w

1-r z

Figure 2: A Two-Stage Compound Lottery

The uncertainty in a compound lottery resolves in successive stages rather

than simultaneously. It is generally assumed that the process of resolution
of uncertainty does not require an economically significant amount of time
and the individual has no economic decisions to undertake in the interim
between lotteries. Therefore what matters to an individual in a lottery are
the effective probabilities of each outcome. Formally:

Assumption 4 (Reduction of Compound Lotteries): An individual

is indifferent between a compound lottery and its probabilistically equivalent
simple lottery.

Under this assumption, the two-stage lottery in Figure 2 is indifferent to the

following probabilistically equivalent simple lottery of Figure 3:

7
 x
pq

p(1-q) y

(1-p)r

(1-p)(1-r) w

Figure 3: Reduction to Equivalent Simple Lottery

The reduction of compound lotteries axiom is a great simplification because

it helps to extend preferences from simple lotteries to compound lotteries.7
In order to determine which of two given compound lotteries is preferred,
simply reduce them to their probabilistically equivalent simple lotteries and
use the preference ranking on the set of simple lotteries.

Remark: While we will only work with simple lotteries, for the sake of
completeness we describe how to formally construct compound lotteries.
This is accomplished recursively as follows. Let L1 = L . Let L2 denote the
set of two-stage compound lottories which are probability distributions
P over
simple lotteries in L . Therefore for  ≥ 0  = 1 2   and =1  = 1:
L2 = {(1 ◦ 1 ; 2 ◦ 2 ; ;  ◦  ) :  ∈ L1   = 1 2  }
Continuing inductively in this manner we have the set of ( + 1)-stage com-
pound lotteries that are probability distributions over -stage compound
lotteries:
L+1 = {(1 ◦ 1 ; 2 ◦ 2 ; ;  ◦  ) :  ∈ L   = 1 2  }

=1 L denotes the set of all simple and compound lotteries. ¥

Then L = ∪∞ 

7
The reduction of compound lotteries assumption is violated when an individual has a
preference for early or late resolution of uncertainty. This has been formally examined by
Kreps and Porteus (1978).

8
6 The Continuity Assumption
The next assumption is a technical restriction on preferences called conti-
nuity.

Assumption 5 (Continuity): Given any simple lotteries 1  2  3 ∈ L ,

the following sets are closed :

{ ∈ [0 1] : (1 ◦ ; 2 ◦ (1 − )) º 3 }

{ ∈ [0 1] : (1 ◦ ; 2 ◦ (1 − )) ¹ 3 }

A set  is closed if the limit of any convergent sequence in  also belongs

to . Continuity thus captures the idea that small changes in probabilities
do not influence the preference ranking between lotteries. Thus for example
if 1 º 3 , then for  close to 1, the compound lottery 1 + (1 − )2 º
3 . The continuity and rationality assumptions play an important role in
Theorem 1 of the next section.

7 Utility and Expected Utility Representation of

Preferences
The objective now is to obtain a numerical representation of the preferences
º on the set of simple lotteries using a utility function  . Therefore an
individual’s pairwise ranking over the entire set of simple lotteries can be
conveniently summarized by a mathematical function. Moreover, the exis-
tence of such a mathematical function makes the analysis amicable to tools
such as calculus.

Definition 1 A real-valued utility function  whose domain is L is a rep-

resentation of the preference ordering º if for any pair of simple lotteries
1 ,2 ∈ L :
1 º 2 ⇔  (1 ) ≥ (2 ) (7)

Note that using the definitions of indifference ∼ and strict preference  it

follows that:8
8
Suppose that 1 ∼ 2 . Then 1 º 2 (i.e.  (1 ) ≥ (2 )) and 2 º 1 (i.e.
(2 ) ≥  (1 )). Putting it together, (1 ) =  (2 ). The representation for strict
preference follows similarly.

9
 1 ∼ 2 ⇔  (1 ) =  (2 ) (8)
1 Â 2 ⇔  (1 )   (2 ) (9)

The existence of such a utility function follows from a general result first
established by Debreu (1954)9 :

Theorem 1 The preference relation º on a set  is representable by a

continuous utility function  :  → R if and only if º is reflexive, complete,
transitive and continuous.

Since we have assumed that preferences º on the set L satisfy reflexivity,

completeness, transitivity and continuity, it follows that there exists a func-
tion  : L → R representing º. We will be interested in a utility function
 representing preferences º on the set of simple lotteries that has a par-
ticular form called the expected utility representation. It is defined as
follows:

Definition 2 The utility function  on the set of simple lotteries has an

expected utility representation if there exists a function  on the set of
outcomes, , so that given any lottery  = (1 ◦ 1 ; 2 ◦ 2 ;   ◦  ), the
utility from the lottery  () can be written as:

X
 () =  ( ) = 1 (1 ) + 2 (2 ) + · · · +  ( ) (10)
=1

Here  is sometimes called the Bernoulli utility function over out-

comes. It reflects the consumer’s preferences over the set of outcomes or
consequences. The function  is called the von Neumann - Morgen-
stern (vNM) utility function over lotteries. It reflects the consumer’s
preference ranking over the set of simple lotteries.

Remark:  and  coincide for degenerate lotteries. If  denotes the

degenerate lottery that gives outcome  for sure, then:

 ( ) = ( ) (11)

9
Proofs are also provided in Bowen (1968) and Barten and Böhm (1982).

10
Therefore many authors do not distinguish between the two utility functions
and refer to  as the vNM utility function over outcomes and  as the vNM
utility function over lotteries. Further:

X
 () =   ( ) = 1  (1 ) + 2  (2 ) + · · · +   ( ) (12)
=1
i.e. the same notation is used when referring to the utility from the outcome
or from the lottery. Both (10) and (12) are equivalent representations of
expected utility. We will use either form according to convenience. ¥

In order to obtain an expected utility representation of preferences we need

an additional assumption that is discussed in the next section.

8 The Independence Assumption

This key assumption behind the expected utility representation is the fol-
lowing independence, or substitution, assumption.

Assumption 6 (Independence): Let 1  2  ∗ ∈ L and for any 0 

  1 consider the compound lotteries, 1 = (1 ◦ ; ∗ ◦ (1 − )) and
2 = (2 ◦ ; ∗ ◦ (1 − )) (shown in Figure 4). Then 1 º 2 if and only if
1 º 2 .

S1 S2
p
p

1-p
1-p

L1 L* L2 L*

Figure 4: Independence or Substitution Assumption

11
The independence assumption is based on the observation that the events
that yield the sublottery  ,  = 1 2 and ∗ in the lotteries above are mutu-
ally exclusive. Therefore the two sublotteries are not “consumed” together.
Since they are non-complementary, we can analyze preferences over sublot-
teries  in the upper branch of the probability tree separately from the
outcome corresponding to the mutually exclusive lower branch of the prob-
ability tree. If 1 is at least as good as 2 , then irrespective of the common
sublottery in the lower branch, 1 is at least as good as 2 . Another way of
stating this assumption is that one lottery is at least as good as another if
and only if the realizations of the former lottery are at least as good as real-
izations of the latter, and these realizations occur with the same probability.
A detailed discussion is available in Machina (1987).

Remark: Suppose 1 ∼ 2 . This is equivalent to 1 º 2 and 2 º

1 . From the independence assumption, these are respectively equivalent to
1 º 2 and 2 º 1 . Therefore 1 ∼ 2 is equivalent to 1 ∼ 2 . Similarly
from the independence assumption, 1 Â 2 is equivalent to 1 Â 2 . ¥

9 Implications of Continuity and Independence

Continuity and independence have some important implications that are
useful in deriving the expected utility representation. These are as follows:

Implication 1 (Boundedness of the set L ): We saw that the set  is

bounded from below by the worst outcome,  , and bounded from above
by the best outcome, . From the independence assumption it now follows
that the set of simple lotteries L is also bounded from below by  and
bounded from above by . To see this, consider any  ∈ L where  =
( ◦ 1 ; 2 ◦ 2 ; ; −1 ◦ −1 ;  ◦  ). By definition,  º  ,  = 2 3  .
Using the independence assumption on each branch one at a time:

 = ( ◦ 1 ;  ◦ 2 ;  ◦ 3 ; ;  ◦ −1 ;  ◦  )
º ( ◦ 1 ; 2 ◦ 2 ;  ◦ 3 ; ;  ◦ −1 ;  ◦  )
º ( ◦ 1 ; 2 ◦ 2 ; 3 ◦ 3; ;  ◦ −1 ;  ◦  )
······························
º ( ◦ 1 ; 2 ◦ 2 ; 3 ◦ 3; ; −1 ◦ −1 ;  ◦  )
º ( ◦ 1 ; 2 ◦ 2 ; 3 ◦ 3; ; −1 ◦ −1 ;  ◦  )
= 

12
Similarly it follows that  º  for any  ∈ L . Therefore  º  º  for
all  ∈ L .

Implication 2 (Simple lotteries as mixtures of  and  ): Implication

1 and the continuity assumption implies the following: For any simple lottery
 ∈  , there exists a probability 0 ≤  ≤ 1 such that  is indifferent to
a simple lottery that yields  with probability  and  with probability
1 −  .10 (This simple lottery is shown in Figure 5 below):

pL

1-pL
W

Figure 5: Continuity Assumption

From Implication 1 each  ∈ L is bounded from below and above by 

and  respectively. Now we invoke the continuity assumption. Consider the
following sets:

1 = { ∈ [0 1] : ( ◦ ;  ◦ (1 − )) º }
2 = { ∈ [0 1] : ( ◦ ;  ◦ (1 − )) ¹ }

Both 1 and 2 are closed from continuity. By definition of the best and
worst outcomes, 1 ∈ 1 and 0 ∈ 2 and therefore these sets are non-empty.
Further, since preferences º are complete, each  ∈ [0 1] belongs to either
1 or 2 . Therefore 1 ∪ 2 = [0 1]. Since [0 1] is connected11 , it follows
10
Some authors use this implication as the definition of continuity to simplify the expo-
sition.
11
A set  is disconnected if  = 1 ∪ 2  where 1 and 2 are non-empty, and
1 ∩ 2 = ∅. A set is connected if it is not disconnected. Any interval in the real line
is connected.

13
that 1 ∩ 2 6= ∅ and thus there exists some  ∈ 1 ∩ 2 .12 By definition
of the sets 1 and 2 :

 ∈ 1 =⇒ ( ◦  ;  ◦ (1 −  )) º 
 ∈ 2 =⇒ ( ◦  ;  ◦ (1 −  )) ¹ 

and it therefore follows that ( ◦  ;  ◦ (1 −  )) ∼ . ¥

Implication 3 (Monotonicity): The next implication of the independence

assumption captures the idea that “more is preferred to less.” It states that
if two simple lotteries give the same outcomes,  and  , then the lottery
that gives a higher probability of  is strictly preferred. Formally, consider
any two simple lotteries 1 = ( ◦;  ◦(1−)) and 2 = ( ◦;  ◦(1−))
(shown in Figure 6 below). Then 1 Â 2 if and only if   .13

B B
p
q

1-q
1-p

L1 W L2 W

Figure 6: 1 Â 2 if and only if   

First of all note that since  Â  , then from the independence assumption
it follows that for 0    1:

 = ( ◦ ;  ◦ (1 − ))
 ( ◦ ;  ◦ (1 − )) (13)
−
(⇐=) Suppose   . Let  = 1− . Consider Figure 7:
12
There is no assertion at this point that  is unique. Uniqueness will follow later as
a consequence of the monotonicity result.
13
Equivalently, 1 º 2 if and only if  ≥ .

14
 q B
B B
p r r W
1-q
(indifferent to) (Strictly preferred
q B from independence q B
1-p 1-r and (12) to) 1-r
W W
1-q W 1-q

Figure 7: Monotonicity

Figure 7 shows that the lottery on the left, 1 , is indifferent to the one in
the middle (from the reduction of compound lotteries assumption) which in
turn is strictly preferred to the lottery on the right (which is simply 2 from
the reduction of compound lotteries assumption). We have therefore shown
that if   , then 1 Â 2 .

(=⇒) We next establish the converse, i.e. if 1 Â 2 , then   . This

can be proven by establishing the contrapositive:  ≤  implies 2 º 1 .
If  = , then clearly 2 ∼ 1 . If   , then the same argument as for
establishing (⇐=) shows that 2 Â 1 .

These implications help establish the expected utility theorem of the next
section.

10 Expected Utility Theorem

The main result of this chapter is the following:

Theorem 2 Suppose preferences º on the set of simple lotteries L satisfy

the rationality assumptions, continuity, and independence. Then for any
simple lottery  ∈ L where  = (1 ◦ 1 ; 2 ◦ 2 ; ; −1 ◦ −1 ;  ◦  ),
the
P preference relation º has an expected utility representation, () =
=1  ( ).

15
We will give a simple constructive graphical proof of this result for the case
where a simple lottery  has only two possible outcomes: an outcome  with
probability  and the outcome  with probability 1 − , where  Â . This
lottery is shown in Figure 8 below. The proof is easily generalized to any
simple lottery on the set .

x
p

1-p

L y

Figure 8: A Simple Lottery with Two Outcomes

We will now construct a utility function  over outcomes (or equivalently,

degenerate lotteries). Let:

( ) = 0 () = 1 (14)

This fixes the utility numbers for the best and worst outcomes.

px B py B

1-px 1-py W
W

Degenerate lottery y
Degenerate lottery x

Figure 9: Outcomes   as mixtures of  

16
Consider the outcome  which is indifferent to the simple lottery on the left
in Figure 9 from Implication 2. We will assign a utility number to  in the
following way:
() =  (15)
Note that the utility number is unique.14 Moreover, ( ) ≤ () ≤ ().
This rule of assigning utility numbers preserves the consumer’s ranking of
outcomes. For example, suppose outcome  is preferred to . From Impli-
cation 2, there exists a probability 0 ≤  ≤ 1 such that  is indifferent
to the simple lottery on the right in Figure 9. From the rule of assigning
probabilities:
() =  (16)
From Implication 3,  Â  is equivalent to    . Therefore, ()  ()
and the utility function preserves the preference ranking over outcomes.

We can now construct the von Neumann - Morgenstern utility function over
simple lotteries. Consider the lottery in Figure 8. Using the independence
assumption, the lottery in Figure 8 is indifferent to the one in Figure 10:

px B

p W
1-px

1-p

Figure 10: Replacing  with Indifferent Simple Lottery

Invoking independence once again, the lottery in Figure 10 is indifferent to

the one in Figure 11:
14
This follows from Implication 3 (monotonicity). If there were two probabilities, 0 
 , for which  was indifferent to a mix of  and  , then from monotonicity  Â 
contradicting irreflexivity of Â.

17
 px B

p
1-px W

1-p py B

1-py W

Figure 11: Replacing  with Indifferent Simple Lottery

From the reduction of compound lotteries axiom, the lottery in Figure 11 is

indifferent to the one in Figure 12:

ppx + (1-p)py

p(1-px) + (1-p)(1-py)
W

Figure 12: Reduction of Compound Lottery

We have shown that the simple lottery  of Figure 8 is indifferent to the one
in Figure 12. According to our rule of assigning utility numbers, any lottery
that is indifferent to some mix of  and  is assigned a utility number equal
to the probability of the outcome . Therefore the utility number assigned

18
to  is:

 () =  + (1 − )

= () + (1 − )() (17)

and we have derived an expected utility form for the vNM utility function.

To complete the argument, we now have to show that the expected utility
form is a numerical representation of preferences, i.e. if  º 0 , then
 () ≥  (0 ). Noting that  () =  + (1 − ) , the lottery in Figure
11 can be written as:

 ∼ ( ◦  ();  ◦ (1 −  ()))

Now any lottery 0 can be similarly written as:

0 ∼ ( ◦  (0 );  ◦ (1 −  (0 )))

If  º 0 , then from transitivity of º we have that:

( ◦  ();  ◦ (1 −  ())) ∼ 
º 0
∼ ( ◦  (0 );  ◦ (1 −  (0 )))

and therefore:

( ◦  ();  ◦ (1 −  ())) º ( ◦ (0 );  ◦ (1 −  (0 )))

It now follows from Implication 3 (monotonicity) that the lottery on the

LHS is at least as good as the lottery on the right if and only if:

 () ≥  (0 )

and we have shown that (17) is a representation of preferences º.

11 Ordinal versus Cardinal Preferences

In consumer theory under certainty, the ordinal nature of preferences im-
plied that the preference ranking was invariant to an increasing monotonic
transformation of the utility function. Therefore, if  was a utility function

19
on the set of consumption bundles, then  = ( ), 0  0, was an equiva-
lent utility function reflecting the same preference ranking. The implication
was that the utility numbers allotted to consumption bundles by themselves
had no connotation of utility or satisfaction but were just a means of re-
flecting the consumer’s ranking of bundles. More than one utility function
allotting a different set of numbers to consumption bundles could equally
well suffice to represent preferences as long as they preserved the underlying
ranking of consumption bundles.

Under uncertainty, the situation is a little different. The expected utility

representation is not invariant to any increasing monotonic transformation.
The expected utility form is only preserved by increasing linear (or affine)
transformations of the Bernoulli and vNM utility functions.

Definition 3 If  is any function, then  is an increasing linear (or affine)

transformation of  if:

 =  +  0 (18)

It will be convenient at this point to write expected utility in the form given
by (12). We will now prove:

Theorem 3 Suppose the utility function  gives an expected utility repre-

sentation of preferences º on the set of simple lotteries. The utility function
 on the set of simple lotteries also gives an expected utility representation
of preferences º if and only if there exists constants   0 and  such that:

 () =  () +   ∈ L (19)

(⇐=) Suppose there exists constants   0 and  such that  () =  ()+.
Consider any simple lottery,  = (1 ◦ 1 ; 2 ◦ 2 ; ; −1 ◦ −1 ;  ◦  ).
Since  has the expected utility form:

X 
X
 () =   ( ) = 1  (1 ) + 2  (2 ) + · · · +   ( )  = 1
=1 =1

Then it follows that:

20
  () =  () + 
"  # 
X X
=    ( ) +  
=1 =1

X
=  [ ( ) + ]
=1

X
=   ( ) (20)
=1

Moreover, if 1 º 2 , then  (1 ) ≥  (2 ) and it follows that  (1 ) =

 (1 ) +  ≥  (2 ) +  =  (2 ). Therefore  represents the same
preferences as  .

(=⇒) Consider any  ∈ L . Since from Implication 1,  º  º  , and 

represents preferences º:

 () ≥  () ≥  ( )

Therefore  () can be expressed as a weighted average of  () and  ( ),

i.e. it follows that there exists a constant 0 ≤   ≤ 1 such that:

 () =    () + (1 −   ) ( ) (21)

and since  ()   ( ), we have:

 () −  ( )
 = (22)
 () −  ( )

Since  has an expected utility representation, it follows from (21) that:

 ( ◦   ;  ◦ (1 −   )) =    () + (1 −   ) ( ) =  ()

and therefore:
 ∼ ( ◦   ;  ◦ (1 −   ))
Since  represents the same preferences, it follows that:

 () =  ( ◦   ;  ◦ (1 −   ))

21
and since  also has an expected utility representation:

 () =    () + (1 −   ) ( )
=  ( ) +   [ () −  ( )]
∙ ¸
 () −  ( )
=  ( ) + [ () −  ( )]
() −  ( )
∙ ¸ ∙ ¸
 () −  ( )  () −  ( )
=  () +  ( ) −  ( )
 () −  ( )  () −  ( )
=  () + 

where   are constants since  ( )  ()  ( ) and  () are constants.

Further, since  Â  , and  and  represent preferences,  ()   ( )
and  ()   ( ). Therefore, as required:

 () −  ( )
= 0
 () −  ( )

This proves the theorem. Another way of stating this result is that the
expected utility form is unique up to an increasing linear transformation.
The fact that the expected utility form is unique only up to an increasing
linear transformation means that the expected utility property is cardinal
in nature. The utility numbers attached to outcomes and lotteries - and in
particular the differences between the numbers - have significance and cannot
be changed in an arbitrary way without destroying the ranking implied by
the expected utility form. For example, suppose there are four possible
outcomes,    and  and the utility numbers attached to them are such
that:
() − ()  () − () (23)
A simple rearrangement of the above inequality gives the equivalent condi-
tion that:
1 1 1 1
() + ()  () + () (24)
2 2 2 2
Now consider the two simple lotteries shown in Figure 13. Note that 1 Â 2
since:
1 1 1 1
 (1 ) = () + ()  () + () =  (2 ) (25)
2 2 2 2

22
 x y
1/2
1/2

1/2
1/2

L1 z L2 w

Figure 13: Comparing Simple Lotteries

Any increasing linear transformation of  will preserve this ranking. How-

ever, any arbitrary monotonic transformation of , such as  = ln , may
yield:
() − ()  () − ()
If we now use an expected utility representation with respect to the function
 then:
1 1 1 1
 (1 ) = () + ()  () + () =  (2 )
2 2 2 2
and we have reversed the ranking of the two lotteries obtained through .
Therefore ranking of preferences according to expected utility can only be
preserved by increasing linear transformations of utility.

Remark: It is important to see that a linear transformation of  is required

only to preserve the expected utility form. However, once we have a ranking
of simple lotteries by the vNM utility function  , any increasing monotonic
transformation of  will give the same ranking of simple lotteries. For
example, comparing the two lotteries in Figure 13, we know that  (1 ) 
 (2 ). It follows that exp  , log  and  2 will give the same ranking of 1
and 2 . For example:
 (1 ) = [ (1 )]2
∙ ¸2
1 1
= () + ()
2 2
∙ ¸2
1 1
 () + ()
2 2
= [ (2 )]2
=  (2 )

23
Of course, the new vNM utility function  no longer has an expected utility
representation. But if all that we are interested in is preserving the prefer-
ence ranking of simple lotteries under  , then this ranking is respected by
any arbitrary monotonic increasing function  as well. ¥

12 Distribution Functions
In formal analysis, a lottery can be represented by a distribution function
on the set of outcomes. The advantage of letting the objects of choice under
uncertainty be a distribution function is that it permits the outcomes to be
either discrete or continuous. When the outcomes are discrete, the set of all
possible outcomes is either finite or countably infinite; when the outcomes
are continuous, the set of all possible outcomes is uncountably infinite.

Let  be a random variable which now represents monetary outcomes. Let

 denote the probability of an outcome. The distribution function  of the
random variable  is defined for any real number  by:

 () = Pr{ ≤ } (26)

In words,  () is the probability that the random variable  takes a value
that is less than or equal to the real number .

A random variable is discrete if it can take a finite or countably infinite

number of values. For discrete random variables, the probability mass
function, (), is defined by:

() = Pr{ = } (27)

Therefore, for discrete random variables, the distribution function can be

written as: X
 () = () (28)
≤

If the set of values  = {1  2    }, then  (1 ) = 0,  ( ) = 1, and 

is non-decreasing.

Example 1: Consider the following probability mass function:

 0 1 2 4
1 1 1
() 0 4 2 4

24
Then the distribution function corresponding to the above probability mass
function is given by:

 0 1 2 4
1 3
 () 0 4 4 1

Note that  (0) = 0,  (4) = 1, and  is a non-decreasing function. ¥

A random variable  is said to be continuous if there exists a probability

density function,  (), on the interval [ ], −∞      ∞, such that
for any  ⊂ [ ]: Z
Pr( ∈ ) =  () (29)

Therefore, for continuous random variables, the distribution function corre-
sponding to the probability density function can be written as:
Z 
 () =  ()  ≤  ≤  (30)


For a continuous random variable, given the distribution function  , the

probability density function  can be recovered by differentiating  :

 0 () =  () (31)

Note that  () = 0,  () = 1, and  is a non-decreasing function.

Example 2: Consider a random variable, , which is uniformly distributed

over the interval [ ],   , and has a probability density function given
by: ½ 1
 () = −       (32)
0 ∈ [ ]
This is called the uniform probability density function. The uniform dis-
tribution function corresponding to this probability density function gives
the probability that  takes a value less than or equal to any given :
⎧
⎨ 0 
−
 () =   ∈ [ ] (33)
⎩ −
1 

It can be verified that  () = 0,  () = 1, and  0 () =  ()  0 on [ ].
¥

25
The expected value or mean of the random variable  can be defined in
the discrete case as: X
 [] = () (34)
:()0

provided the summation exists. In the continuous case:

Z  Z 
 [] =  () =  0 () (35)
 

provided the integral exists. Note that the expected value P

can be interpreted
as a degenerate lottery which yields the sure outcome :()0 () in
R
the discrete case and   0 () in the continuous case.

We have already seen the representation of expected utility in the discrete

case. For any Bernoulli utility function  over outcomes and probability
mass function , expected utility is represented as:
X
 () =  [] = ()() (36)
:()0

In the continuous case, for density function  (or equivalently distribution

function  ) and Bernoulli utility function , expected utility is given by:
Z  Z 
 ( ) =  [] = () () = () 0 () (37)
 

provided the integral exists.

References
[1] A. Barten and V. Böhm, “Consumer Theory”. In Handbook of Mathe-
matical Economics, K. Arrow and M. Intrilligator (eds), (1982), 382-429,
North Holland.

[2] R. Bowen, “A New Proof of a Theorem in Utility Theory”, International

Economic Review, 9 (1968), 374.

[3] G. Debreu, “Representation of a Preference Ordering by a Numerical

Function”, Decision Processes, R.M. Thrall et al (eds), (1954), 159-165,
John Wiley.

26
[4] D.M. Kreps and E.L. Porteus, “Temporal Resolution of Uncertainty and
Dynamic Choice Theory”, Econometrica, 46 (1978), 185-200.

[5] M. Machina, “Choice Under Uncertainty: Problems Solved and Un-

solved”, Journal of Economic Perspectives, 1 (1987), 121-154.

27