von Neumann & Morgenstern Expected Utility

Econ 2100 Fall 2015

Lecture 12, October 12

Outline
1 Von Neumann & Morgentstern Expected UtilityTheorem
2 Objective Probabilities?
3 Anscombe & Aumann Acts
4 State-Dependent Expected Utility
5 Midterm (Todays material will not be included)
 From last class: Expected Utility Theorem
X = fx1 ; x2 ; : : : ; xn g is a nite set.
X is the space of all probabilities on X :
Xn
X = f 2 Rn : i = 1 and i 0; 8ig;
i =1

Theorem (Expected Utility Theorem, von Neumann and Morgenstern 1947)

Let X be the set of all probabilities 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
% , v (x) (x) v (x) (x):
x 2X x 2X
 Necessity and Uniqueness in vN&Ms Expected
Utility Theorem
Question 1, Problem Set 8
1 Necessity (() part of vN&Ms Expected Utility TheoremP
If there exists a vNM index v : X ! R such that u( ) = x 2X v (x) (x) is a
utility representation of %, then % is independent and Archimedean.
2 Uniqueness part
P of vNMs Expected Utility Theorem
Let U( )P= x v (x) (x) be a utility representation of %. Then,
U 0 ( ) = x v 0 (x) (x) is also representation of % 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 .

We will see two proofs of su ciency. Both use the mixture space theorem, so
all we need to prove is that the a ne function is the expected utility function.
The resason for the second proof is to connect the theorem to properties of
the space of linear functions.
This maybe mysterious now, but could be important if one is interested in
weakining this result.
 Su ciency of vNMs Expected Utility Theorem
Proof.
Su ciency ()) of vN&Ms Expected Utility Theorem
Enumerate X = fx1 ; x2 ; : : : ; xn g; observe that X is a convex subset of Rn .
The Mixture Space Theorem implies existence of an a ne utility
representation U : X ! R.
Let v (xi ) = U( xi ) (the utility of the Dirac lottery on xi ).
This pins down the utility value for each prize.
Pick some 2 X and denote i = (xi ).
Pn Pn
Verify that = i =1 (xi ) xi = i =1 i xi

this follows because xi is the unit vector pointing in the i -th dimension.
Pn
Since U is a ne, each 0, and i =1 i = 1, we know (Q2, PS 8) that
i
Xn n
X
U( ) = U( i xi ) = i U( xi )
i =1 i =1
By construction, this implies
n
X X
U( ) = i v (xi ) = (x)v (x):
i =1 x 2X
 Riesz Representation Theorem
Duality between linear functions and vectors

Lemma
A function f : Rn P
! R is linear if and only if there exists a unique vector v 2 Rn
n
such that f (x) = i =1 vi xi = v x.

This says there is a duality between linear functions and the vector space over
which they are dened: every vector denes a linear function via the dot
product, and every linear function is identied with some vector.
We proved a version of this result in math camp.
 Another Proof of Su ciency of vNMs Theorem
Proof.
Su ciency ()) of vN&Ms Expected Utility Theorem (Again)
Let X = fx1 ; x2 ; : : : ; xn g. Observe that X is a convex subset of Rn .
The Mixture Space Theorem says there exists U : X ! R is an a ne
representation of %.
Any a ne function from X to R can be extended to another a ne function
U : Rn ! R such that U( ) = U( ) for all 2 .
WLOG, assume U is linear by subtracting the constant U(0n ) (see last class).
By the previous Lemma, there exists a vector v 2 Rn such that for all z 2 Rn
Xn
U(z) = v z = vi zi :
i =1
Dene v (xi ) = vi and let (xi ) = i for all 2 X.
Then, for any 2 X,
n
X X
U( ) = U( ) = v (xi ) (xi ) = (x)v (x):
i =1 x 2X
 Pessimistic Expected Utility
Example
Consider this alternative representation of %:
U( ) = minfv (x) : (x) > 0g:
DM evaluates each lottery by the worst element on the support. If there is a
non zero chance, no matter how small, to get a terrible outcome the decision
maker will evaluate the lottery as if she was going to receive that outcome for
sure.
This is extreme pessimism: if something bad is possible, it will happen.
This is sometimes called innite risk-aversion (we will talk about risk-aversion
soon).

Question 3, Problem Set 8

Suppose there exists a vNM utility index v : X ! R such that the following is a
utility representation for %:
U( ) = minfv (x) : (x) > 0g:
Prove or disprove the following: (a) % is independent; (b) % is Archimedean; and
(c) minfv 0 (x) : (x) > 0g is a representation of % if and only if v 0 = av + b for
some a > 0; b 2 R.
 Expected Utility with Innitely Many Prizes
Allowing X to be innite requires some tricky functional analysis. The real line
is the space of consequences, and the equivalent of X is the set of measures
over that.
Let X = R and R be the set of density functions on X with nite variance.

Theorem
The preference relation % on R is complete, transitive, independent, and
continuous if and only if there exists a square-integrable index v : R ! R such that
Z
U(f ) = v (x)f (x)dx
is a representation of %.

One needs the stronger continuity assumption so that the representation

produced by the Mixture Space Theorem is continuous (more than just a ne).
R contains probability measures with a density and nite variance. This is
unsatisfying:
density functions represent a small subset of all measures; for example, the
degenerate measures are excluded.
 Objective Probabilities
In von Neumann and Morgensterns expected utility representation preferences
rank lotteries.
Probabilities and outcomes are therefore known primitives of the model.

Theorem
The preference relation % on X is 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 %.

In most cases, however, probabilities are not a given; they are in the mind of
the decision maker and so they dier across individuals.
To account for this, we want a model that lets the likelihood of events be in
the mind of the decision maker.
A DMs prefence relation should reveal her probability distribution over events
as well as her utility for consequences.
We need a more general consumption space to achieve this.
 Anscombe and Aumann Structure
= f1; 2; : : : ; mg is a nite set of states, with generic element s 2 .
X = fx1 ; : : : ; xn g is a nite set of outcomes, with a generic element x 2 X .
H = ( X ) is the space of all functions from to X .
this is a convex subset of the space of functions from to Rn .

AnscombeAumann acts
An AnscombeAumann act h 2 H is a function h : ! X;
AA acts assign a lottery (an element of X ) to each state (an element of ).

Let hs = h(s) 2 X , and denote hs (x) = [h(s)](x) 2 [0; 1].

This is the probability of x conditional on s, given the act h: hs (x ) = Pr(x js; h).

Notation
A 2 X denotes the constant act f : ! X s.t.f (s) = for all s 2 .
n n
Given H (R ) , if f ; g : ! R , then the function f + g : ! Rn is
dened by [ f + g ](s) = f (s) + g (s).
This denition is crucial: Archimedean and Indepdence axioms sum acts.
Summing is all about the objective lotteries (not about the horse race).
 Horse Lotteries and Roulette Lotteries
There are two kinds of sources of randomness: states of the world ( ) and
lotteries over consequences ( X ).
How can we think about them?
Interpretation of H
Elements of X are bets on an objective roulette
outcomesprobabilities are objectively determined (everyone agrees on them).
A state of the world represent the event that a specic horse named s wins
a race among the eld of horses
the decision maker subjectively assesses each horses strength (dierent DMs
can evaluate each horse dierently).
The theorys aim is to identify the decision makers personal assessment of the
probability that horse s will win the race using her preferences.
To perform this identication, we set the payout on horse s equal to a lottery
that depends on the outcome of a roulette spin.
So, rst the horses run the race, and second the roulette is spun.
The roulettes payo depends on which horse wins.

Next, equivalent descriptions of H.

 Three mathematical representations of H

First. The original mathematical interpretation where H = ( X ) .

Suppose = fs1 ; s2 ; s3 g and X = fx1 ; x2 ; x3 g.

Then a particular h : ! X would be the following:
h(s1 ) = (0:3; 0:2; 0:5)
h(s2 ) = (0:4; 0:6; 0)
h(s3 ) = (0; 1; 0)
 Three mathematical representations of H
Second. H as a set of compound lotteries.

The subjective rst stage lottery is over which state s 2 obtains, and the
objective second stage lottery (conditional on s) is over which x 2 X obtains.
These compound lotteries can be written as probability trees.
For example: h is
 Three mathematical representations of H

Third. H is the set of weakly positive m n matrices where

Xn
hi ;j = 1 for each i = 1; : : : m:
j =1

Then the h on the previous slides can be represented as

2 3
Pr(x = x1 ) Pr(x = x2 ) Pr(x = x3 )
6s1 0:3 0:2 0:5 7
6 7
6
h = 6s2 0:4 0:6 0 7:
7
4s3 0 1 0 5

We can then write hs (x) = hs ;x ; for example hs1 (x2 ) = h1;2 = 0:2.
 Objective Lotteries Are Anscombe-Aumann Acts
Notation
One can identify X as a subset of H.
The probability distributions on X are acts that, regardless of the state, give the
same lottery over outcomes.
These are usually called constant acts, and the set of all constant acts is
denoted Hc (this subset of H is isomorphic to X )
X Hc = ff 2 H : f (s) = f (s 0 ) for all s 2 g

One can also identify X as a subset of H.

Elements of X are Dirac lotteries (degenerate probability distributions) in X ,
denoted x .
Thus, X is a subset of Hc dened as follows
X = ff 2 H : f (s) = f (s 0 ) 8 s 2 ; and f (s) = x for some x 2 X g

Anscombe-Aumann acts generalize von Neumann and Morgensterns setting.

Without extra assumptions, we can use the Mixture Space Theorem on H.

 State Dependent Expected Utility

Theorem
The preference relation % on H is complete, transitive, independent, and
Archimedean if and only if there exists a set of vNM indices v1 ; : : : ; vm : X ! R
such that XX
U(h) = vs (x)hs (x)
s x
is a utility representation.

This follows from the Mixture Space Theorem applied to H (a convex set).
The utility of a consequence depends on the state in which it obtains (vs (x)
depends on s ): state-dependent additive representation.
The su ciency proof is like vNM: there exists a linear function U : Rm n ! R
representing %, and this linear function is uniquely determined by a vector in
Rm n .
 State Dependent Expected Utility
Theorem
The preference relation % on H is complete, transitive, independent, and
Archimedean if and only if there exists a set of vNM indices v1 ; : : : ; vm : X ! R
such that XX
U(h) = vs (x)hs (x)
s x
is a utility representation.

Can we identify a probability distribution over ?

If we could, DM would take the subjectiveexpectation with respect to some
probability distribution over of a state-dependent vNM index vs :
U(h) = E (Ehs (vs )):
In other words, one would like the utility function to be
" #
X X
U(h) = (s) vs (x)hs (x) ;
s x
In the last exercise of Problem Set 8, you will show that such a cannot be
uniquely identied.
 State Dependent Expected Utility: Discussion
More precisely, suppose we have
" #
X X
(s) vs (x)hs (x)
s x
where is a probability distribution over .
Then for any 0 2 such that 0 (s) > 0 for all s 2 , there exist indices
0 0
v1 ; : : : vm : X ! R such that
" #
X X
0
(s) vs0 (x)hs (x) :
s x

Remark
One cannot identify probabilities using state dependent expected utility.
Any representation that has ; vs ( ) is equivalent to a representation that uses
0
; vs0 ( ).

One cannot identify a unique probability from preference under the

assumptions of the state-dependent expected utility theorem.
To pin down a probability distribution one needs a state-independent
representation.
 Econ 2100 Fall 2015

Problem Set 7
Due 20 October, Tuesday, at the beginning of class
1. Prove the necessity and uniqueness parts of von Neumann & Morgenstern Expected Utility Theorem.
More precisely, prove the following.
P
(a) If there exists a von Neumann & Morgenstern index v : X ! R such that u( ) = x2X v(x) (x)
is a utility representation of %, then % is complete, transitive, independent and Archimedean.
P P
(b) Let U ( ) = x v(x) (x) be a utility representation of %. Then, U 0 ( ) = x v 0 (x) (x) is also
representation of % 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.

2. Prove
Pm that a function f : ! R is a ne if and only if, for all 1; 2; : : : ; m 0 such that
k=1 k = 1 and
(1)
; (2)
; : : : ; (m) 2 ,1
!
X
m X
m
(k) (k)
f k = kf ( ):
k=1 k=1

3. The purpose of this exercise is to illustrate how expected utility theory allows us to make consistent
decisions. Suppose that a safety agency is thinking of establishing a criterion under which an area
prone to ooding should be evacuated. The probability of ooding is 1%. There are four possible
outcomes: (A): no evacuation is necessary, and none is performed; (B) an evacuation is performed
that is unnecessary; (C) an evacuation is performed that is necessary; (D) no evacuation is per-
formed, and a ood causes a disaster.
Suppose that the agency is indierent between the sure outcome B and the lottery of A with prob-
ability p and D with probability 1 p, and between the sure outcome C and the lottery of A with
probability q and D with probability 1 q. Suppose also that it prefers A to D and that p 2 (0; 1)
and q 2 (0; 1). Assume that the conditions of the expected utility theorem are satised.

(a) Construct a utility function of the expected utility form for the agency. (Normalize utility so
that u(A) = 1 and u(D) = 0)
(b) Consider two dierent policy criteria:
Criterion 1: This criterion will result in an evacuation in 90% of the cases in which ooding
will occur and an unnecessary evacuation in 10% of the cases in which no ooding occurs.
Criterion 2: This criterion is more conservative. It will result in an evacuation in 95% of the
cases in which ooding will occur and an unnecessary evacuation in 15% of the cases in which
no ooding occurs.
First, derive the probability distributions over the four outcomes under these two criteria. Then,
by using the utility function from your answer to above, decide which criterion the agency would
prefer.
1 (k)
T h e e le m e nts a re in d e x e d by p a re nth e siz e d su p e rsc rip ts to avo id c o n fu sio n w ith th e su b sc rip ts w h ich d e n o te d im e n sio n s.

1
4. Suppose there exists a vNM utility index v : X ! R such that the following is a utility representation
for %:
U ( ) = minfv(x) : (x) > 0g:
Prove or disprove the following: (a) % is independent; (b) % is Archimedean; and (c) minfv 0 (x) :
(x) > 0g is a representation of % if and only if v 0 = av + b for some a > 0; b 2 R.

5. Suppose " #
X X
U (h) = (s) vs (x)hs (x)
s x
0 0
Show that for any probability distribution 2 such that (s) > 0 for each s 2 , there exist
0
vNM indices v10 ; : : : vm : X ! R such that
" #
X X
0
U (h) = (s) vs0 (x)hs (x) :
s x