How to have
rational preferences
A lottery is a list of possible outcomes with a known chance of each. Suppose all you can say is which of two lotteries you would rather have. John von Neumann and Oskar Morgenstern showed that if those preferences obey four rules, then they can be summarized by one number for each outcome, its utility, and you always prefer the lottery whose utility is higher on average.
The four rules are: you can always compare two lotteries; your preferences do not go round in a circle; no outcome is infinitely better or worse than another; and adding the same side possibility to two lotteries does not change which you prefer. Four figures come first, to build intuition. The proof follows.
A gamble can be used as a ruler
Take a worst prize, $0, and a best prize, $100. To find out how much you value $40, compare two offers: $40 for certain, or a gamble that pays $100 with some chance and $0 otherwise. If the chance is tiny you take the $40. If it is nearly certain you take the gamble. Somewhere in between is a chance at which you cannot choose. That chance is your utility for $40, on a scale where $0 is 0 and $100 is 1.
Set the four sliders to your own break-even chances. There are no right answers.
The chart plots your four answers. The dashed line is what someone would answer who valued each dollar the same. A curve above the line means you would give up some money on average to avoid a gamble.
The rule that makes this measurement possible is called continuity: for any prize between the worst and the best, there is some chance at which the gamble is exactly as good as the prize.
If your preferences cycle, you can be charged for standing still
Suppose you would rather have a week at the beach than a week in the city, a week in the mountains than at the beach, and a week in the city than in the mountains. Each preference sounds harmless. Together they form a circle. Someone can then offer you, for a small fee each time, a swap to the trip you prefer over the one you hold.
Each arrow points from a trip to the one you like better. The filled circle is the trip you hold now. Each swap costs $1.
The rule that prevents this is transitivity: if you prefer the first thing to the second and the second to the third, you prefer the first to the third. With completeness, which says you can always compare, it puts every lottery in a single order from worst to best.
Mixing in the same side possibility cannot change which you prefer
With three prizes, every lottery is a point in a triangle. The corners are the three prizes for certain. Moving right adds chance of the worst prize, and moving up adds chance of the best. Lines join lotteries you find equally good.
Here A is $3,000 for certain and B is an 80% chance of $4,000. Now water both down: with some chance you get A or B as before, and otherwise you get nothing. That slides both points toward the "$0 for certain" corner, giving A′ and B′. The independence rule says the watering down is the same for both, so it cannot change which one you prefer.
| offer | average utility |
|---|
Thin lines join equally good lotteries, and better is toward the upper left. Independence forces these lines to be straight and parallel. The segment from A′ to B′ is the segment from A to B shrunk toward the right-hand corner, so it stays parallel to the original and crosses the lines the same way.
The first slider is the utility from Figure 1, here for $3,000 on a scale where $0 is 0 and $4,000 is 1. It sets the tilt of the lines. Whatever the tilt, the answer for A against B and the answer for A′ against B′ always match.
Many people's choices break the independence rule
Here are the two comparisons from Figure 3 as plain questions. Choose in each, going by what you would actually pick.
Each bar shows the values of "utility for $3,000" that would explain one of your choices. Expected utility needs a single value that explains both.
Question 2 is Question 1 watered down: C is a 25% chance of A, and D is a 25% chance of B. Independence says your answers should match. Choosing A and then D is common. These two questions are the version Daniel Kahneman and Amos Tversky used in 1979, a simplified form of a problem Maurice Allais posed in 1953. Whether that shows people choose badly or that the rule asks too much is still argued.
The theorem in four steps
Fix a finite set of outcomes. A lottery gives outcome with probability . Write for " is at least as good as ," for strictly better and for equally good. For , the mixture is the lottery that gives with probability .
What is assumed
- Completeness. For any : or .
- Transitivity. If and then .
- Continuity. If then for some strictly between 0 and 1.
- Independence. For any and any : exactly when .
Independence as stated carries over to and . By completeness and transitivity the outcomes can be ranked, so there is a best outcome and a worst outcome . If every lottery is equally good and there is nothing to prove, so take . Write for the reference gamble .
1. A better chance of the best is better
Claim: exactly when . Mixing both sides of with gives for . For , put . Then , and mixing both sides of with gives
The reverse direction follows because the order is complete.
2. Every outcome has a place on the reference scale
For each outcome there is exactly one number in with
If take , and if take . Otherwise and continuity supplies the number. Step 1 makes it unique, because two different reference gambles are never equally good. This is the measurement in Figure 1.
3. Replace each outcome by its reference gamble
A lottery is a mixture of its outcomes. By independence, swapping one outcome for something equally good leaves the lottery equally good. Swap each in turn for . The result pays or and nothing else, and its total chance of is . So
4. Compare
By step 3 and transitivity, comparing with is comparing two reference gambles, and step 1 says the one with the larger chance of is better:
That is expected utility. Conversely, preferences defined by comparing for any function satisfy all four rules, as a direct check shows.
How unique is u?
Suppose also represents the same preferences in this way. Since , the two must have equal expected , so . That is
So utility is fixed once you choose its zero and its unit, and not before.
What the theorem does and does not say
- Utility is a summary of choices. The theorem does not say anyone feels utility or consciously maximizes it. It says that preferences obeying the rules can be described that way.
- Utility is not money. Figure 1 lets the curve bend. The Dutch book argument assumed each dollar is valued the same; replacing dollars by utility removes that assumption.
- The zero and the unit are free. Because of that, the theorem gives no way to compare one person's utilities with another's.
- The chances are given. The lotteries come with known probabilities. Leonard Savage's 1954 theory, and a simpler version by Anscombe and Aumann in 1963, derive the probabilities and the utilities together from preferences alone.
- Independence is the rule people break. Figure 4 is one example. Since about 1980 there have been theories that weaken it.
Sources:
- J. von Neumann and O. Morgenstern, Theory of Games and Economic Behavior. The axiomatic proof appears in the second edition (1947).
- M. Allais (1953), for the choice problems that violate independence.
- L. J. Savage, The Foundations of Statistics (1954).
- F. J. Anscombe and R. J. Aumann, "A definition of subjective probability," Annals of Mathematical Statistics 34 (1963), pp. 199–205.
- D. Kahneman and A. Tversky, "Prospect theory: an analysis of decision under risk," Econometrica 47 (1979), pp. 263–291. The two questions in Figure 4 are their Problems 3 and 4.