Vous êtes sur la page 1sur 20

Anscombe & Aumann Expected Utility

Econ 2100 Fall 2015

Lecture 13, October 20

Outline
1 State Independence
2 Subjective Expected Utility
3 Qualitative Probabilities
4 Paradoxes
Anscombe and Aumann Structure

= f1; 2; : : : ; mg is a nite set of states, with generic element s 2 .


X is a nite set of size n of consequences with a generic element x 2 X .
X is the set of all probability distributions over X .
Elements of reprent subjectively perceived randomness, while elements of
X represent objectively perceived randomness.
H = ( X ) is the space of all functions from to X.
An act is a function h : ! X that assigns a lottery in X to each s 2 .
Let h s = h(s ) 2 X , and denote h s (x ) = [h(s )](x ) 2 [0; 1].
This is the probability of x conditional on s , given the act h: (Pr(x js ; h))

The set of constant acts is


Hc = ff 2 H : f (s) = f (s 0 ) for all s 2 g
If f ; g 2 H, then the function f + g : ! X is dened by
[ f + g ](s) = f (s) + g (s):
Draw a picture to make sure you see how this works.
Subjective Expected Utility (SEU): Idea
Starting from preferences, identify a probability distribution 2 and a
utility index v : X ! R such that a utility
" representation # of these preferences is
X X
U(h) = (s) v (x)hs (x)
s2 x 2X
equivalently: " # " #
X X X X
f %g , (s) v (x)fs (x) (s) v (x)gs (x)
s2 x 2X s2 x 2X

Some Accounting Details


U(h) is the (subjective) expected value of v given act h since
2 3
" # Pr(s ) Pr(x js ;h)
X X X X z}|{ z }| {
6 7
(s) v (x)hs (x) = v (x) 4 (s) hs (x) 5 :
s x x s
| {z }
Pr(x )
P
s (s)hs (x ) is the total or unconditional subjective probability of x under the
function h, denoted Pr(x ).
P
Therefore Pr(x ) = s Pr(s )Pr(x js ; h), since (s )h s (x ) = Pr(s )Pr(x js ; h).
Last Class: 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. the utility representation is
XX
U(h) = vs (x)hs (x)
s x

Is this a subjectiveexpectation, with respect to some probability distribution


over , of a state-dependent vNM index vs ?
U(h) = E (Ehs (vs )):
No, because in this model such a cannot be uniquely identied.
Any representation that has ; vs ( ) is equivalent to a representation that uses
some 0 ; vs0 ( ).
To pin down a probability distribution one needs a state-independent
representation.
Null States
Notation
Given an act h 2 ( X ) , a state s 2 , and a lottery 2 X , dene the new act
(h s ; ) : ! X by (h s ; ) = (h1 ; : : : ; hs 1 ; ; hs +1 ; : : : ; hm ). So
if t = s
[(h s ; )](t) = ;
h(t) if t 6= s

(h s ; ) replaces hs (the lottery that act h assigns to state s) with the lottery
; the remainder of h stays the same.

With this notation, one can identify states the decision maker never cares
about; they should be irrelevant to the representation.

Denitions
A state s 2 is null if, for all h 2 ( X ) and ; 2 X , (h s; ) (h s; ).
A state s 2 is non-null if it is not null, i.e. if there exist h 2 ( X ) and
; 2 X such that (h s ; ) (h s ; ).

A state is null if it never aects rankings.


State Independence
Denition
The binary relation % on H is state-independent if, for all non-null states s; t 2 ,
for all acts h; g 2 H, and for all lotteries ; 2 X ,
(h s; ) % (h s; ) ) (g t ; ) % (g t ; ):

In words, the ranking of roulette lotteries does not depend on the state in
which they obtain.
This only needs to hold in states the decision maker cares about.
Implies the utility index over consequences is the same across states.

Exercise
Let = frain; shineg and X = fUmbrella; Sunglassesg. The decision makers
preferences satisfy:
( U ; S ) ( U ; U ) (she strictly prefers getting an umbrella during rain and
sunglasses during shine to getting an umbrella for sure) and
( U ; S ) ( S ; S ) (she strictly prefers getting an umbrella in rain and
sunglasses in shine to getting sunglasses for sure).
Verify that these preferences violate state-independence.
Null States and State Independence
Proposition
Fix a preference relation % on H. Dene %hs on X by
%hs if and only if (h s; ) % (h s; )
Then, %hs is a preference relation on X and the following holds:
1 If % is state-independent, then %hs = %gt for all non-null states s; t 2 and
acts h; g 2 ( X ) .
h
2 If s is null, then s for all h 2 ( X ) .

What is %hs ?
A preference order that ranks roulette lotteries in a particular state s given a
particular Anscombe-Aumann act h.
1. means this preference is the same in all non-null states.
2. says that if a state is null this preference order ranks everything as
indierent.
The proof of 1. and 2. is left as Exercise, we will proof that %hs is a preference
next.
Null States and State Independence
Fix a preference relation % on H. Dene %hs on X by
%hs if and only if (h s; ) % (h s; )
Then %hs is a preference relation on X

Proof.
To show that %hs is a preference we use the fact that the original preference % is
complete and transitive to prove that %hs also has these properties.
Completeness:
Pick arbitrary ; 2 X , then (h s ; ); (h s ; ) 2 H;
by completeness of %, either (h s ; ) % (h s ; ) or (h s ; ) % (h s ; ) or both;
hence, either %hs or %hs or both. Therefore %hs is complete.
Transitivity:
Pick ; ; 2 X such that %hs , and %hs .
By construction (h s ; ), (h s ; ) , and (h s ; ) are all in H.
By denition, (h s ; ) % (h s ; ) and (h s ; ) % (h s ; ); and
by transitivity of %, we have (h s ; ) % (h s ; ).
We conclude that %hs , and therefore %hs is transitive.
Monotonicity
Denition
The binary relation % on ( X ) is monotonic if, for all h; g 2 ( X ) ,
hs % gs ; 8s 2 ) h % g:

Recall that hs also denotes the constant function which assigns the lottery
hs 2 X to all states t 2 .

An act preferred in each state must be preferred overall.


One can think of monotonicity as state by state dominance.

Exercise
Prove that if a preference relation % on H is state-independent, then % is
monotonic.

Exercise
Prove that if % is monotonic, then %hs is equal to % restricted to constant acts
(this means %hs if and only if ( ; ; :::; ) % ( ; ; :::; )).
Subjective Expected Utility Theorem
Theorem (Expected Utility Theorem, Anscombe and Aumann)
A preference relation % on H is complete, transitive, independent, Archimedean,
and state-independent if and only if there exists a vNM index v : X ! R and a
probability 2 such that
" #
X X
U(h) = (s) v (x)hs (x)
s x
is a utility representation of %.
Moreover, this representation is unique up to a ne transformations provided at
least two acts can be ranked strictly.

Therefore: " # " #


X X X X
h%g , (s) v (x)hs (x) (s) v (x)gs (x)
s x s x
The second part says: if there exist h; g 2 P
H such that
P h g and U(h)
dened above represents %, then U 0 (h) = s 0 (s) [ x v 0 (x)hs (x)] also
represents % if and only if 0 = and v 0 = av + b for some a > 0 and b 2 R.
Remark
The probability distribution is unique.
Subjective Expected Utility Theorem
Theorem (Expected Utility Theorem, Anscombe and Aumann)
A preference relation % on H is independent, Archimedean, and state-independent
if and only if there exists a vNM index v : X ! R and a probability 2 such
that 2 3
" # Pr(s ) Pr(x js )
X X X X z}|{ z }| {
6 7
U(h) = (s) v (x)hs (x) = v (x) 4 (s) hs (x) 5
s x x s
| {z }
Pr(x )

is a utility representation of %. This representation is unique up to a ne


transformations.

Preferences identify two things:


the utility index over consequences v : X ! R and
the probability distribution over states 2 .
Dierent preferences may imply dierent beliefs on .
von NeumannMorgensterns Theorem only identies the utility index v .

For this reason, this is sometimes called Subjective Expected Utility.


Anscombe & Aumann Proof
We will not do it in detail, but here is a breakdown of what would happen.

Proof.
First, convert each act h to a vector x 2 [0; 1]m , by assigning each dimension
s 2 f1; :::; mg a vNM expected utility for hs
P
this gives x v (x )hs (x ) in each state s;
monotonicity, a consequence of state-independence, is essential for this step.
Then, construct an independent and Archimedean preference relation on
[0; 1]m and the measure is the dual of the a ne utility on [0; 1]m .
The necessity and uniqueness parts are as usual.

Notice that this argument involves using the mixture space theorem (or vNM
theorem) twice.
P
the rst time to nd x v (x )hs (x ) in each state s;
the second to nd the measure as the dual of the a ne utility on [0; 1]m .
Savage Subjective Expected Utility
Anscombe & Aumann assume the existence of an objective randomizing device
(preferences are dened over functions from states to objective lotteries).
X is a convex space and one can apply the mixture space theorem.

Savages theory instead considers all uncertainty as subjective: no probabilities


are assumed a priori; all probabilities are identied by preferences over basic
objects.
This comes at the cost of higher mathematical complexity (see KrepsTheory
of Choice).
is the state space, and X is a nite set of (sure) consequences.
Savage acts are elements of X , the space of all functions from to X .
The decision maker has preferences % over these acts.
The expected utility representation is something like
X Z
U(f ) = v (x) (f 1 (x)) = v fd = E [v f ];
x 2X
where f is any act, v : X ! R is a vNM utility index on consequences, and
is a probability measure on .
Utility and probability are fully endogenous.
Probability Distributions

Economists think of preference as primitive, and derive subjective probability


as a piece of a subjects utility function.
Some statisticians think about probabilities the same way economists think
about utility functions.

Denition
A (nitely additive) probability measure on is a function : 2 ! [0; 1] such
that:
1 (X ) = 1;
2 If A \ B = ;, then (A [ B) = (A) + (B).

They usually assume that is also countably additive, which is very


technically useful.
Qualitative Probabilities

Denition
A binary relation on 2 is a qualitative probability if:
1 is complete and transitive;
2 A ;, for all A ;
a
3 ;;
4 A B if and only if A [ C B [ C , for all A; B; C such that
A \ C = B \ C = ;.
a By A B , we mean A B and not B A, i.e. is the asymmetric component of .

Think of this as a binary relation expressing the idea of more likely than.
This is a personal characteristic; it varies from individual to individual and
can be use to infer the existence of personal probabilities.
Qualitative Probabilities and Probability Measures
The relationship between qualitative probability relations and probability
measures is similar to the one between preference orders and utility functions.

Denition
The probability measure represents the binary relation on 2 if:
(A) (B) if and only if A B:

Result
All binary relations that are represented by a probability measure are
qualitative probabilities.

The converse is false: not all qualitative probabilities can be represented by a


probability measure.
A stronger version of additivity is required to guarantee representation by a
probability measure (see Fishbrun 1986 in Statistical Science).

Our theorems about preferences infer the existence of probabilities. Their


theorem about probabilities infer the existence of a preference relation.
Allais Paradox
The decision maker must choose between the following objective lotteries.

and between
One reasonable prefrence may rank A B and D C.
Any such preferences violates independence:

Consider the following lotteries E and F :

0:11A + :89A = A B = 0:11E + :89A

0:11A + :89F = C D = 0:11E + 0:89F

But independence says for all f ; g ; h 2 H and 2 (0; 1),


f % g , f + (1 )h % g + (1 )h:
Ellseberg Paradox
An urn contains 90 balls. 30 balls are red (r ) and 60 are either blue (b) or yellow
(y ). One ball will be drawn at random. Consider the following bets.

and
Many prefer A to B and C over D. Is this consistent with expected utility?
By expected utility, there exists a (subjective) probability distribution over
fr ; y ; bg.
From A B and C D we conclude:
(r )U($100) + (1 (r )) U($0) > (b)U($100) + (1 (b)) U($0)
| {z } | {z }
(b)+ (y ) (r )+ (y )

) (r ) > (b)
and
( (b) + (y ))U($100) + (r )U($0) > ( (r ) + (y ))U($100) + (b)U($0)
) (r ) < (b)

This contradicts the hypothesis of consistent subjective beliefs.


(Knightian) Uncertainty vs. Risk
Risk Knightian Uncertainty (Ambiguity)
B R B R

B = 30 R = 70 20 B 50 50 R 80

B+R=100 B+R=100
3 7
P(B) = 10 P(R) = 10 P(B) =? P(R) =?

Some theories that account for Uncertainty.

Gilboa & Schmeidler: Maxmin Expected Utility


X X
x % y , min (s)u(xs ) min (s)u(ys )
2C 2C
s s
Truman Bewley: Expected Utility with Sets
X X
x %y , (s)u(xs ) (s)u(ys ) for all 2C
s s
Problem Set 8 (Incomplete)
Due Monday 26 October
1 Suppose = fs1 ; s2 ; s3 g and X = fx1 ; x2 ; x3 g. The preference relation % is independent,
Archimedean, and state-independent. The following are known: x1 x2 x3 and
( x1 ; x2 ; x3 ) ( x1 ; (0; 21 ; 12 ); (0; 12 ; 12 )) (( 23 ; 13 ; 0); ( 23 ; 31 ; 0); x3 ). Find the decision makers
subjective belief .
2 Suppose = fs1 ; s2 ; s3 g and X = fx1 ; x2 g. Assume % is a preference relation on
Anscombe-Aumann acts, ( X ) , which admits an Anscombe-Aumann 2 expected 3 utility
(hs1 (x1 ); hs1 (x2 ))
representation. Denote each AnscombeAumann act as h = 4(hs2 (x1 ); hs2 (x2 ))5. The
(hs3 (x1 ); hs3 (x2 ))
following are known:
2 3 2 3 2 3 2 3 2 3 2 3
(1; 0) (0; 1) (1; 0) (0; 1) (1; 0) (0; 1)
4(1; 0)5 4(0; 1)5 ; 4(0; 1)5 % 4(1; 0)5 ; 4(0; 1)5 4(0; 1)5 :
(1; 0) (0; 1) (0; 1) (0; 1) (0; 1) (1; 0)
What is the set of beliefs in that is consistent with these data (if the decision maker
follows AnsombeAumann expected utility)? Is a unique belief identied, i.e. is the identied
set a singleton? What if the middle weak preference were an indierence relation?
3 Suppose that % is an independent and Archimedean preference relation on (X ). Explain in
as much detail as you can how (X ) diers from ( X ) (the latter is the space over which
preferences were dened in class). Explain why one can apply the mixture space theorem to
% dened over (X ) and describe the representation one would obtain? How is it dierent
from the state dependent expected utility representation we discussed in class? Can you think
of an assumption that would make preference over (X ) similar to preference over ( X ) ?