Decision Making Under Uncertainty

Econ 2100 Fall 2015

Lecture 11, October 7

Outline
1 Decision Making Under Uncertainty
2 Convex Consumption Set and Independence
3 Mixture Space Theorem
4 Preferences Over Lotteries
5 von-Neuman & Morgenstern Expected Utility
 Decision Making Under Uncertainty: Prelude
So far, consumption has been an here and now matter.
Preference orderings compare alternatives available for immediate use.
Next, the decision maker must choose now among items that will be
consumed in the future.
This is reasonably straightforward when the future is known with certainty, one
only needs to worry about discounting (future consumption may not be as
valuable as current consumption).
Things are more complicated (and interesting) when there is uncertainty about
what will happen.
Then, future consumption depends on current choices through the way in which
uncertainty is resolved.
We will typically think of a future consumption vector as a random variable:
only one of many possibilities (states of the world) will occur, but
an exhaustive list of all these possibilities (the state space) describes possible
future consumption.
Preferences compare (now) alternatives that will be consumed in the future.
We will see classic results on utility functions representing preferences. These
results are more about getting a specic useful functional form than
establishing existence.
 Decision Making Under Uncertainty: Prelude
Focus on two periods: today and tomorrow.
Decisions are made today but items are consumed tomorrow.

There is a way to make this t into the existing model. Here is how.
Make a list of all the states of the world that are possible.
Assume a good can be dened by its physical description (e.g. umbrella) as well
as the state in which that good is consumed (e.g. tomorrow is a rainy day).
Thus, an umbrella tomorrow when it rains is a good; an ubmrella tomorrow
when it is sunny is another good.
Do this for all "physical goods" and all "states" to get the consumption space.
A consumption vector species quantities of all these goods.
1 umbrella tomorrow when it rains, 0 ubmrella tomorrow when it is sunny,
and so on.
If one assumes a preference relation over this consumption space satises
completeness, transitivity, and continuity, then there exist a utility function
representing it.
Wether or not goods are deterministic or random does not change that theory.
This approach, although entirely satisfactory from a formal standpoint, is
abstract and hard to use in applications.
For example, it hides how decision making may depend on stateslikelihoods.
If rain is very likely, DM should want many umbrella tomorrow when it rains.
 Lotteries and Probabilities
A box contains green and red balls. One ball will be drawn from this urn.

Green Red

G = 40 R = 60

G 4 R 6
Prob(G ) = G +R = 10 Prob(R) = G +R = 10

Consider these lotteries:

A: you win $10 if a Green ball is drawn and zero otherwise
B: you win $10 if a Red ball is drawn and zero otherwise
Would you rather have A or B? Why?
Is the expected value of a lottery a reasonable utility function? Does it
describe how much one is willing to pay for it?
 Expected Values and Probabilities
St. Petersburg Paradox
Toss a fair coin until the rst Head obtains.
consider a gamble that pays $2t if the rst head appears on the t-th toss
$2 if H on the 1st toss, $4 if H on the 2sd , $8 if H on the 3rd ,...
What is the value (utility) of this gamble? Would the expected gains work?
1 1 1
2 + 4 + 8 + ::: = 1
2 4 8
is anyone willing to pay that much for this lottery?
As a solution, Bernoulli suggested using log 2t to evaluate it:
1 1 1
log 2 + log 4 + log 8 + ::: a nite number
2 4 8
This solution works for this gamble... but how about a gamble that pays e 2t
for H on the t -th toss?
1 1 1
log e 2 + log e 4 + log e 8 + ::: = 1
2 4 8
we are back at the paradox.
Expected gains do not seeem to capture how we feel about it.
 Expected Utility (imprecisely described)
Suppose there is only one good: call it money.
The state space consists of S mutually exclusive states of the world;
a generic element is denoted s 2 S .
Let =( 1 ; :::; S) be a probability distribution over S:
X
1 s 0 and s =1
s 2S
A consumption bundle x = x1 ; :::; xS is a random variable.
the agent consumes the amount xs if and only if state s happens.
Consumption bundles are evaluated with the expected value of their utility
XS
U (x) = s u (xs ) :
s =1
This function separatesprobability from utility of consumption.
The utility function U ( ) is linear in probabilities and utils of consequences.
If we let u (x ) = (u(x1 ); :::; u(xS )), we can write this as u (x ) .
Decision making is connected to the theory of probability developed in
mathematics and statistics.
An advantage of this approach is that we can use the rules of probability theory
to evaluate how information about dierent events enters the decision making
process (this is handy for game theory and information economics).
 Convex Consumption Space
To obtain representations similar to the one in the previous slide, we need
more assumption than completeness, transitivity, and continuity.
This is intuitive since the utility function has a very special functional form.
We also neet to modify the space over which consumption is dened.

Consumption Set is Convex

The consumption set is a convex subset of Rn denoted .

Convexity means that if x; y 2 , then

if x; y 2 , then x + (1 )y 2 for all 2 (0; 1):
Start with preferences over an abstract convex space.
Then, add more structure to this space to get more specic results.

Denition
A binary relation % on is:
complete if, for all x; y 2 , x % y or y % x, or both;
transitive if, for all x; y ; z 2 , x % y and y % z imply x % z;
 Archimedean Axiom
Since the consumption space is convex, one can use a weaker version of
continuity.

Denition
A binary relation % on is Archimedean if, for all ; ; 2 ,
8
< 9 2 (0; 1) such that + (1 )
=) and
:
9 2 (0; 1) such that + (1 )

Exercise
Show that if % is continuous then it is Archimedean.

Exercise
Let = R and let % on R dened by the utility function
8
< 1 if > 0
U( ) = 0 if = 0 :
:
1 if < 0
Verify that % is Archimedean but not continuous.
 Independence

A crucial new assumption yields additive separability of the representation.

Denition
A binary relation % on satises independence if, for all x; y ; z 2 and 2 (0; 1),
x % y , x + (1 )z % y + (1 )z:

Example
Suppose = R2 and % dened by
x % y if and only if x12 + x22 y12 + y22
this says: x is weakly preferred to y whenever the norm of x is weakly larger than
the norm of y .
Then (4; 0) (0; 4), but
1 1 1 1
(4; 0) + (2; 0) = (3; 0) (1; 2) = (0; 4) + (2; 0)
2 2 2 2
So % is not independent.
 Consequences of Independence

Proposition
If % satises independence its indierence classes are convex.

Proof.
Suppose % is independent. We need to show that
) + (1 ) ; 8 2 [0; 1]
If , clearly % . Thus, by independence, for all 2 [0; 1],
% ) + (1 ) % + (1 ) ) % + (1 ) :
If , clearly % . Thus, by independence, for all 2 [0; 1],
% ) + (1 ) % + (1 ) ) + (1 ) % :
Therefore we get + (1 ) for all 2 [0; 1].
The same logic shows that + (1 ) .
Therefore the indierence classes are convex.
 Characterization of Independence

The following provides an alternate characterization of independence, which is

sometimes useful in proofs.

Exercise
Prove that a binary relation on is independent if and only if, for all ; ; 2 ,
and 2 (0; 1),
, + (1 ) + (1 )
and
, + (1 ) + (1 )
 Linear and A ne Functions

Denition
A function f : ! R is a ne if, for all ; 2 and 2 [0; 1]
f( + (1 ) ) = f ( ) + (1 )f ( ):

Denition
A function f : Rn ! R is linear if, for all ; 2 Rn and ; 2 R,
f( + ) = f ( ) + f ( ):

Exercise
Prove that a function f : Rn ! R is a ne if and only if g ( ) = f ( ) f (0n ) is
linear.
 Mixture Space Theorem
Theorem (Mixture Space Theorem, Herstein and Milnor)
A binary relation % on (a convex subset of Rn ) is complete, transitive,
independent and Archimedean if and only if there exists an a ne function
U : ! R such that
% ,U( ) U( )

Moreover, if U : ! R represents %, then U 0 : ! R also represents % if and

only if there exist real numbers a > 0 and b such that U 0 ( ) = aU( ) + b for all
2 .

The rst part of the statement says that the preference are represented by an
a ne utility function.
The second that the representation is unique up to linear transformations.
Remarks
This holds for any convex subset of an arbitrary vector space.
The utility function U( ) is cardinal and not just ordinal as before (Why?).
The Mixture Space Theorem asserts that there exists some a ne
representation, not that all representations are a ne.
 Mixture Space Theorem
Theorem (Mixture Space Theorem, Herstein and Milnor)
A binary relation % on (a convex subset of Rn ) is complete, transitive,
independent and Archimedean if and only if there exists an a ne function
U : ! R representating %:
% ,U( ) U( )
Moreover, if U : ! R is an a ne representation of %, then U 0 : ! R is an
a ne representation of % if and only if there exist real numbers a > 0 and b such
that U 0 ( ) = aU( ) + b for all 2 .

As with Debreus utility representation proof, the main step is to nd the

unique number that is indierent to a given 2 .
The main dierence is that the assumptions now imply an a ne utility function
(rather than continuous one as in Debreus theorem).
Proof: Roees class.
This theorem will yield an expected utility representation when used on special
convex consumption sets.
 Preferences Over Lotteries
Von Neumann and Morgenstern (1947)
Let X be a nite set of size n and X be the space of probabilities on X .

Enumerate X = fx1 ; x2 ;(
: : : ; xn g, and let (xi ) = i so that )
Xn
X = 2 Rn : i = 1 and i 0; 8i
i =1
X is a convex subset of the vector space Rn .
An element of X identies a lottery over the elements of X (the prizes).
One can write a lottery also as ( 1 ; x1 ; 2 ; x2 ; : : : ; n ; x n ).
The degenerate lottery that yields some outcome x with certainty is called a
Dirac lottery on x and denoted x ;
xk is the unit vector in the direction k : xk = (0 1 ; : : : ; 0 k 1 ; 1 k ; 0 k +1 ; : : : ; 0 n ).
The set of Dirac lotteries is f x : x 2 X g X , and it constitutes the extreme
points of X (what are the extreme points?).

% is dened over X

Remark
A preference relation ranks probability distributions over a nite set of objects.
Since the set of prizes is xed, the decision makers preference order is over lotteries.
 Lotteries
If X = fx1 ; x2 ; x3 g, a typical lottery =( 1 ; x1 ; 2 ; x2 ; 3 ; x3 ) is described
using an event tree:

Then x2 , the degenerate lottery which yields x2 with certainty, is:

The space X assumes all uncertainty is resolved at one point in time; it does
not allow for compound lotteries (lotteries over lotteries).
This domain restriction can be justied by introducing a reduction of
compound lotteriesassumption as to reduce every compound lottery to a single
lottery in X .
 Compound Lotteries
The convex combination + (1 ) might be interpreted as the compound
lottery ( ; ; 1 ; ) which yields with prob. and with prob. 1 .
If compounded correctly, this yields the same probabilities on consequences as
+ (1 ) :

One could assume the decision maker compounds correctly, but we cannot
state this assumption since compound lotteries are not part of the primitives.
Suppose Z is a nite subset of X . A lottery 2 Z is a compound lottery,
because it is a lottery over lotteries.
If one takes a 2 Z then is an element of X , a lottery over X .
 Independence and Lotteries
0
independence: for all ; ; 2 X and 2 (0; 1),
% 0
, + (1 ) % 0
+ (1 )

Hence:

The decision maker cares only about paths which dier.

This is a normativejustication for the independence axiom on X.
 Expected Utility
Things we already know
Under completeness, transitivity and continuity, there exists a continuous utility
function representing the preferences.
When X is convex, one can replace continuity with the archimedean axiom and add
independence, and show that the utility function is a ne.

Under the extra structure given by X , the representation theorem identies a

function v : X ! R such that
X Xn
U( ) = (x)v (x) = i v (xi )
x 2X i =1
represents %.
DM weights the utility of each outcome by the probability of receiving that
outcome.
Since the probability distribution over is given, the theorem identies via
preferences the functional form of U( ) and the function v ( ).

Remark
The function v : X ! R yields a vector v 2 Rn by letting vi = v (xi ).
The expected utility formula is the dot product of two vectors (v and ) in Rn .
 Expected Utility Theorem
Theorem (Expected Utility Theorem, von Neumann and Morgenstern 1947)
Let X be the set of all probability distributions on a nite set X . The preference
relation % on X is complete, transitive, independent and Archimedean if and only
if there exists a function v : X ! R such that
X
U( ) = v (x) (x)
x 2X
is a representation of %. This representation is unique up to a ne transformations.

U represents % means X X
% , U( ) U( ) , v (x) (x) v (x) (x):
P x 2X x 2X
Uniqueness means: U 0 ( ) = x v 0 (x) (x) also represents % if and only if
there exist a > 0 and b 2 R such that v 0 (x) = av (x) + b for all x 2 X .
The function v is called von Neumann & Morgenstern utility index or Bernoulli
utility function.
Remark
v : X ! R is not a utility representation of %; the domain of v is X , which is
not equal to X . The utility index v is a component of the utility
representation U, which is dened on X (the correct domain).
 Econ 2100 Fall 2015

Problem Set 6
Due 12 October, Monday, at the beginning of class
1. Assume the function f : R2 ! R is supermodular in (x; q) and show that supermodularity is equiv-
alent to the property that

q0 > q implies f (x; q 0 ) f (x; q) is non-decreasing in x:

2. Without using Milgrom and Shannons theorem, prove that if f : R2 ! R is supermodular in (x; q),
then x (q) arg max f (x; q) is non-decreasing.

3. Consider the following extension of the monotone comparative statics results allowing for constrained
optimization. Suppose f : Rn Rm ! R is quasi-supermodular in x and satises the single crossing
property. Suppose ' : Rm ! Rn is a correspondence such that '(q 0 ) s '(q) for every q 0 > q. Set

x (q) = arg max f (x; q)

x2'(q)

and assume jx (q)j = 1 for each q. Prove that x (q 0 ) x (q) for every q 0 > q.

4. Suppose a rm purchases two inputs x1 and x2 at unit costs w1 and w2 to produce a single product
with production function f (x1 ; x2 ) to sell at a unit market price of p. Its prot is therefore:

(x1 ; x2 ; p; w1 ; w2 ) = pf (x1 ; x2 ) w 1 x1 w 2 x2 :
2
Suppose f is C 2 and @x@ 1fx2 > 0, i.e. x1 and x2 are complements. Assume that x (p; w1 ; w2 ) =
arg max (x1 ; x2 ; p; w1 ; w2 ) is unique. Use monotone comparative statics to prove that if p0 p,
0 0 0 0 0
w1 w1 , and w2 w2 , then x (p ; w1 ; w2 ) x (p; w1 ; w2 ). (Hint: you will have to do some change
of variables.)

5. Show that the set of Dirac lotteries, f x : x 2 Xg X, constitutes the extreme points of the
convex set X (An extreme point of a convex set A is a point x 2 A with the property that if
x = y + (1 )z with y; z 2 A and 2 [0; 1], then y = x and/or z = x).

6. Suppose X is nite and % is an independent preference relation on X. Prove that if there exist
; 2 X such that , then there exist x; y 2 X such that x y.

7. Suppose X is nite. Prove that if % is an independent preference relation, there exist x; y 2 X such
that x % % y for all 2 X.

8. Suppose Z is a nite subset of X and % is an independent and Archimedean preference relation on

Z. A lottery 2 Z is a compound lottery, because it is a lottery over lotteries. Check that the
following statement is false. Suppose ; ; + (1 ) ; + (1 ) 2 Z; then % implies
+(1 ) % +(1 ) .