Tutorial on von Neumann-Morgenstern Theorem
Preferences
Let be a set. A relation on a set is a subset of (the set of pairs of elements from ). That is, is a relation on means that . It is often convenient to write for . To simplify our notation we write when .
An example of a relation on the set is
We are interested in using relations on a set to represent a decision maker's preferences over the objects in .
We are typically not interested in any relation on , but rather relations that satisfy certain properties. The key properties that we study are:
is reflexive provided that for all , .
is irreflexive provided that for all , it is not the case that .
is complete provided that for all , or (or both).
is transitive provided that for all , if and then .
To illustrate, in the box below enter a relation. Find three different relations:
A relation that is complete but not transitive.
A relation that is transitive but not complete.
A relation that is both transitive and complete.
Let be a set and . We say that is a rational preference relation provided that is complete and transitive.
Example: Let and . Then, is complete and transitive, so is a rational preference relation on . This relation describes a decision maker that:
is indifferent between and ;
strictly prefers over ; and
strictly prefers over .
Given a rational preference relation on , we define the following:
Strict Preference: is the relation where for all , if, and only if, and .
Indifference: is the relation where for all , if, and only if, and .
We sometimes use to denote a preference relation instead of . Then, is the strict preference relation generated by and is the indifference relation generated by .
Utility
A utility on a set is a function , where is the set of real numbers.
Representing a Preference Ordering: Suppose that is a preference ordering. We say that is representable by a utility function provided that there is a such that for all , iff .
For example, suppose that and is the rational preference relation where:
, and
.
Find two utility functions that represent this preference relation.
This section can be skipped on a first reading. It contains some of the mathematical details.
The key observation is that every rational preference relation is representable by a utility function.
Utility Theorem. Suppose that is a finite set. A relation is a rational preference ordering if, and only if, is representable by a utility function.
Proof. We leave it to the reader to show that if is representable by a utility function, then is transitive and complete.
We prove the following: For all , any preference relation on a set of size is representable by a utility function . The proof is by induction on the size of the set of objects . The base case is when . In this case, for some object . If is a transitive and complete ordering on , then . Then, (any real number would work here) clearly represents . The induction hypothesis is: if , then any preference ordering on is representable. Suppose that and is a preference ordering on . Then, for some object , where . Note that the restriction\footnote{If is a relation on and , then is the {\bf restriction of to } provided for all , iff .} of to , denoted , is a preference ordering on . By the induction hypothesis, is representable by a utility function . We will show how to extend to a utility function that represents . For all , let . For the object (the unique object in but not in ), there are four cases:
for all . Let .
for all . Let .
for some . Let .
There are such that . Let .
Then, it is straightforward to show that represents (the details are left to the reader).
Remark. The above proof can be extended to relations on infinite sets . However, additional technical assumptions are needed. It is beyond the scope of this article to discuss these technicalities here.
It is not hard to see that if a preference relation is representable by a utility function, then it is representable by {\em infinitely} many different utility functions. To make this more precise, say that a function is {\bf monotone} provided implies .
Lemma. Suppose that is representable by a utility function and is a monotone function. Then, also represents .
Proof. The proof is immediate from the definitions: Suppose that . Then, iff (since represents ) iff (since is monotone).
Lotteries
Suppose that is a finite set. A probability on is a function such that . If , then . (Note: There are a number of mathematical details about probability measures that we are glossing over here. Our discussion in this section is greatly simplified since we assume that the set of objects is finite.) In the remainder of this section, elements of are called prizes, or outcomes.
Lottery. Suppose that is a set of elements from . A lottery on is denoted as follows: where each and .
We have defined lotteries for any subset of a fixed set . Without loss of generality, we can restrict attention to all lotteries on . For instance, suppose that and is a lottery on . That is, . This lottery can be trivially extended to a lottery over as follows:
Suppose that is a finite set. Let be the set of lotteries on (we often write instead of to simplify notation). There are two technical issues that need to be addressed.
First of all, we can identify elements with lotteries . Thus, we may abuse notation and say that " is contained in ".
Second, we will need the notion of a compound lottery. Suppose that are lotteries. Then, is the compound lottery, where .
We are interested in decision makers that have preferences over the set of lotteries (on some fixed set ). We denote the decision maker's preferences on by the relation . That is, given two lotteries and , we assume that a decision maker has one of the following opinions:
The decision maker strict prefers over (denoted ).
The decision maker strict prefers over (denoted )
The decision maker is indifferent between and (denoted ).
The decision maker cannot compare and .
Assuming that the decision maker's preferences relation is complete rules out case number 4. One way to compare lotteries is to assign a number of each lottery and rank the lotteries according to the number assigned to them. That is, for a set of lotteries and a function assigning a real number of teach lottery, define the relation as follows: for all , if, and only if, .
One way to define a function for the set of lotteries on is to start with a utility function on the set of prizes and then for each define as some function combining the probabilities in with the utility of the prizes in .
For example, let be a utility function on . The following are examples of different ways to assign a value to lotteries:
is the function where for each ,
is the function where for each ,
is the function where for each ,
is the function where for each ,
is the function where for each ,
is the function where for each ,
is the function where for each ,
We can also define functions that treat sure-things different than other lotteries: Let be the function where:
ParseError: KaTeX parse error: Unexpected end of input in a macro argument, expected '}' at end of input: …(z) &\mbox{ if L[z:1]z\in XParseError: KaTeX parse error: Expected 'EOF', got '}' at position 2: }̲\\ \sum_{i=1}^n…L=[x_1:p_1,\ldots, x_n:p_n]ParseError: KaTeX parse error: Expected 'EOF', got '}' at position 42: …o probabilities}̲ \end{cases}
Note that function is the expected utility of lotteries with respect to the utility function . The key property of the expected utility valuation function is that it is linear:
A function is linear provided that for all , .
The functions and are all linear.
To illustrate the above functions, let and consider the utility function where , and . Enter the probability for the prizes to see the different values associated with the lottery.
It is also useful to compare the above 8 value functions based on the value that assign to the lottery as ranges from 0 to 1.
The von Neumann-Morgenstern Theorem characterizes the relations on the set of lotteries that can be represented by linear utility functions. Suppose that is a set, is the set of lotteries on , and is a relation on . von Neumann and Morgenstern consider the following 4 axioms.
The first axiom is the standard assumption that rational preferences are complete and transitive:
Preference: The relation is a complete and transitive relation on .
The next two axioms play a central role von Neumann-Morgenstern theorem.
Independence For all and , if, and only if, .
Continuity For all , if , then there exists such that .
Both axioms have been criticized as rationality principles. For example, consider the continuity axiom. Consider the prizes "win $1000", "win $100" and "get hit by a car". Clearly, it is natural to assume that a decision make would have the preference . Now, Continuity implies that there is some number such that . Thus, the decision maker would strictly prefer a lottery in which there is some non-zero chance of getting hit by a car to a lottery in which gives a guaranteed payoff of $100. Arguably, many people would not hold such a preference no matter how small the chance is of getting hit by a car. Here we bracket this are related philosophical discussions about the above axioms and focus on what follows from the axioms. The first observation is a straightforward consequence of Independence and the assumption that the preference ordering is complete.
Lemma Suppose that is a preference relation on satisfying the Independence axiom. For all lotteries and real numbers , if , then .
The second observation is that decision makers prefers lotteries in which there is a better chance of winning a more preferred prize.
Lemma If is a preference relation on satisfying Compound Lotteries and Independence, then for all lotteries , if , and , then .
The last axiom concerns compound lotteries:
Compound Lotteries: Suppose that is a compound lottery, where for each , we have . Then,
The axiom means that decision makers do not get any utility from the "thrill of gambling". That is, what matters to the decision maker is how likely she is to receive prizes that she prefers. For example, suppose that and . Then, the decision maker is assumed to be indifferent between the compound lottery and the simple lottery .
The von Neumann-Morgentern Representation Theorem A binary relation on satisfies Preference, Independence, Continuity and Compound Lotteries iff is representable by a linear utility function .
Moreover, represents iff there exists real numbers and such that . (" is unique up to linear transformations.")
One consequence of the von Neumann-Morgenstern Theorem is that a preference ordering generated by a value function that is not linear must violate at least one of the von Neumann-Morgenstern axioms. For instance, consider the value function . This function is not linear: (Recall that we identify each item in with the sure-thing lottery.)
This means that the preference ordering generated by must violate at least one the 4 axioms used in the von Neumann-Morgenstern Theorem.
Since is represented by a utility function on the lotteries (i.e., it assigns a real number to every lottery), the Utility Theorem shows that is complete and transitive. It is also easy to see that satisfies Compound Lotteries.
To show that violates the Independence Axiom, we must find three lotteries , , and a such that but .
To show that violates the Continuity Axiom, we must find three lotteries , , and a such that but there is no such that .
To illustrate, let , and . We have , and note in th graph below that for , .
For the above lotteries: and , find a lottery such that:
There is no such that .