The Dutch book argument

Price your beliefs badly
and you lose, whatever happens.

Treat each of your degrees of belief as a price: what you would pay for a bet that returns $1 if you turn out to be right. If those prices break the rules of probability, a bookie can offer you a set of bets, each one fair by your own lights, that together lose you money in every possible outcome. That set of bets is called a Dutch book. If your prices obey the rules, no such set exists.

This is the classic argument that rational degrees of belief must be probabilities. Three figures you can play with come first, to build intuition. The math follows.

Figure 1 · A credence is a price

One ticket, two sides

The ticket below pays $1 if it rains tomorrow. Set your credence that it will rain. That number is the price at which you are equally happy to buy the ticket or to sell it.

TICKET
Pays $1.00 if it rains tomorrow
Your price:
it rainsit stays dryaverage by your credence
You buy it
You sell it

Either side can win or lose. What makes the price yours is the last column: weighted by your own credence, neither side has an edge.

Figure 2 · The book

Three horses, three prices, one sure loss

Three horses race and exactly one will win. Set your credence that each one wins. The bookie then looks only at your three prices and picks the trades.

If the winner isyour net

The bar stacks your three prices against the $1 that the winning ticket pays. Any gap between the end of the bar and the $1 line is money you lose in every outcome.

Each bet looked fair to you on its own. The loss comes from the prices not fitting together: you are valuing "some horse wins," which is certain and worth exactly $1, at something other than $1.

Figure 3 · Why coherent prices are safe

Now you are the bookie. Try to book a coherent bettor.

This bettor's prices add up to exactly $1. Choose how many tickets on each horse to sell to the bettor (right of center) or buy from the bettor (left of center). The bars show the bettor's net for each possible winner. Your goal is to make all three bars red.

The dashed line is the average of the three bars, weighted by the bettor's own prices. With coherent prices it sits at $0 no matter what you do, so the bars can never all be below it. Tick the box and the line comes loose: now you can drag it, and all three bars, below zero.

The math

Incoherent prices can be booked. Coherent prices cannot.

Setup

Your credence is your price for a ticket that pays $1 if is true and nothing otherwise. You are willing to buy or sell any number of tickets at that price.

Buy one ticket. You pay up front. If is true, the ticket pays $1, so you end up ahead by . If is false, the ticket pays nothing, so you are down the you paid. So:

Here is 1 when is true and 0 when it is false.

Sell one ticket. Now suppose you are on the other side of the same trade. You collect up front and must pay $1 if is true. Your gain equals the buyer's loss, .

Any number of tickets. Buying tickets repeats the one-ticket purchase times, so your net gain is . Selling tickets repeats the one-ticket sale times, so your net gain is . The two differ only in sign.

So record a bet as one signed number, the stake : set when you buy tickets and when you sell . Then one formula covers both:

Whichever side you take, your net gain from the bet is .

A Dutch book is a set of such bets whose total is negative however the world turns out.

If your prices break a rule, a Dutch book can be made against you

Take outcomes , exactly one of which will happen. The bookie chooses which trades to offer. You accept each one, because each is at your own price and so looks fair to you. All the bookie needs to do is add up your prices.

Your prices total more than 1. The bookie sells you one ticket on each outcome. You pay every price, and exactly one ticket pays $1. So in every outcome your net is

Your prices total less than 1. The bookie buys one ticket on each outcome from you. You collect every price, and you must pay $1 on the one ticket that wins. So in every outcome your net is

Either way you lose for certain, by exactly the amount your prices miss 1. Only prices that total exactly 1 give the bookie no opening. The other rules fail the same way:

Rule, and how it is brokenThe bookie has youYou gainYou loseNet, in every outcome

broken:
buy a ticket on at most
the payout, if is true

the price
at most

broken:
sell a ticket on the certain event
the price

the payout, which is certain

for that cannot both happen
broken:
buy , buy , sell " or "
the price of the ticket you sold

the prices of the two you bought

prices a bet on that is refunded if is false
broken:
buy the refundable bet, buy tickets on , sell " and "
the price of the ticket you sold, plus back in every outcome

the prices of the two bets you bought

Gain is money coming in across all the trades, loss is money going out, and net is gain minus loss. In the last two rows, a $1 payout you receive is always matched by a $1 payout you owe, so those cancel and are left out. In the last row you get back whatever happens: as winnings on the tickets if is true, and as the refund if it is false. If an inequality points the other way, reverse every trade.

If your prices obey the rules of probability, there is no Dutch book

Suppose your prices obey all four rules. Then every "net" entry in the table above is zero, so those trades gain the bookie nothing. But a bookie is free to try any combination of bets, so we need an argument that covers them all at once.

Weigh the outcomes by your own credences. List the possible outcomes , the complete ways things could turn out, exactly one of which will happen. For a three-horse race there are three: "Aster wins," "Bolt wins," and "Comet wins." If the bets concern rain and wind tomorrow there are four: rain with wind, rain without wind, wind without rain, and neither. An event such as "it rains" is then true in some of the outcomes (here the first two) and false in the rest.

Because your prices obey the rules, the credences are never negative, they add up to 1, and is the sum of over the outcomes in which is true.

One bet. Take a bet with stake on . Exactly one outcome will actually happen, and the bet has a definite net gain in each case. If the outcome that happens is one where is true, the net gain is . If it is one where is false, the net gain is . These are the Setup formula with equal to 1 in the first case and 0 in the second. For each outcome, multiply the net gain you would get if it happens by your credence that it happens, and add them up:

The first step splits the possible outcomes into those where is true and those where it is false. The second pulls out the net gain, which is the same for every outcome within a group. The third replaces the two sums: the credences of the outcomes where is true add up to , and the rest add up to , because all of them together add up to 1.

So for every single bet, the credence-weighted sum of its net gains is zero. The fourth rule makes the same true of the refundable bets.

Any set of bets. The net gain of a set of bets is the sum of the net gains of its members, so its credence-weighted sum is zero as well.

No sure loss. Now suppose some set of bets lost money in every outcome. Each of its net gains would be negative. The weights are never negative and add up to 1, so the weighted sum would be negative too. But we just showed it is zero. So no set of bets loses in every outcome.

Fine print

What the argument assumes

The argument goes back to Frank Ramsey (published in 1931) and Bruno de Finetti (1937). Everything above concerns your credences at a single time. A separate argument, devised by David Lewis and first published, with credit to him, by Paul Teller in 1973, is more contested. It uses a sequence of bets to argue that you should also update by conditioning on what you learn, replacing with once you learn .