Dutch-Book Arguments
A Dutch book is a set of bets, each of which an agent regards as fair or favourable, that together guarantee a loss whatever happens. The Dutch-book argument for probabilism is that an agent whose credences violate the probability axioms is susceptible to one, and that this susceptibility convicts the credences of irrationality.
It is the oldest argument for the Bayesian position — Ramsey states it in a sentence, de Finetti develops it — and the most frequently taught. It is also the most frequently criticised, because the defect it identifies is in the first instance practical, and the conclusion is supposed to be epistemic. Whether that gap can be closed is the main question of this page.
The setup
Assume an agent's credence in determines the odds at which they will bet on it: with credence , they regard as fair a bet costing that pays if is true and nothing otherwise, and are willing to take either side. The stake may be positive or negative — buying or selling the bet.
The argument is then a construction. Suppose the agent's credences violate an axiom; produce a set of bets each fair by their own lights; show the net payoff is negative in every state of the world.
Violating normalisation. Suppose for a logical truth. The agent will sell a bet on for that pays . Since is certainly true, they pay out and lose with certainty.
Violating additivity. Suppose and are incompatible and the agent has but . Buy bets on and on for each, total ; sell a bet on for . If neither obtains, the agent gains and loses . If exactly one obtains — and both cannot — they gain and lose . Either way the net is .
Violating non-negativity. A negative credence means paying someone to take a bet that can only pay out in one's own favour, which loses in every state.
The construction generalises: any violation of the axioms admits such a book.
The converse theorem
The argument would be incomplete with only this half, since it would show that incoherence is sufficient for vulnerability without showing coherence is safe. The converse Dutch-book theorem supplies the other direction: if the agent's credences do satisfy the axioms, no Dutch book can be made against them.
The proof is straightforward. If credences are probabilities, then for any finite set of bets the expected net payoff computed with those probabilities equals the sum of the individual expectations, each of which is zero at fair odds. A book guaranteeing a loss would have negative payoff in every state and hence negative expectation, contradiction.
Together the two directions give a clean biconditional: credences are immune to Dutch books if and only if they are probabilities. That is a real result, and it is what makes the argument more than a collection of examples.
Two extensions are worth noting. The result covers only finitely many bets, so it establishes finite additivity and not countable additivity — extending it requires allowing an agent to accept infinitely many bets simultaneously, which is a substantial further assumption and one de Finetti rejected. And a Czech book (or "semi-Dutch book") argument strengthens the conclusion slightly: incoherent agents are not merely exposed to sure loss but forgo sure gains.
The diachronic argument
Lewis and Teller extended the argument to the dynamic norm. An agent who announces in advance that on learning they will adopt some credence other than can be offered a set of bets, some before and some after the learning, that guarantees a loss.
The construction exploits the announced discrepancy: the bookie sells bets now at the current rates and buys them back later at the announced rates, pocketing the difference. This is taken to show that conditionalization is the uniquely rational update rule.
The diachronic version is considerably weaker than the synchronic one, for three reasons.
- It requires the bookie to know the agent's future update policy, which the agent must therefore have announced or be transparently disposed to follow. An agent who updates unpredictably is not exploitable in this way, and it is odd if unpredictability is thereby rational.
- Changing one's mind is not obviously irrational. The argument treats any deviation from the plan as a defect, but an agent who realises their prior was badly chosen and revises it is doing something we would ordinarily commend. The argument cannot distinguish improvement from incoherence.
- It presupposes that the agent's evidence comes in the form of a proposition learned with certainty, which is the idealisation Jeffrey conditionalization exists to relax.
Objections
The pragmatic/epistemic gap
The central objection. A Dutch book shows that certain credences lead to bad bets. But rationality of belief is not the same as prudence in gambling, and the argument seems to establish the wrong kind of defect.
The point can be sharpened. Suppose an agent is incoherent but will never be offered a bet — a hermit, or someone who refuses to gamble on principle. The argument's premises give no purchase, yet their credences seem just as defective. Conversely, an agent with impeccably coherent credences who is bad at arithmetic under time pressure may lose money without any epistemic failing. Vulnerability to loss and epistemic defectiveness come apart in both directions.
The depragmatised reading
The most influential response, due to Skyrms, Armendt, Christensen, and Howson and Urbach, is that the money is a dramatic device. What the Dutch book really reveals is an inconsistency in evaluation: the agent assigns different values to the same set of payoffs depending on how it is described. In the additivity example, the agent values the package at and the identical package at , and these are the same bet under two descriptions.
On this reading the defect is not that one would lose money but that one's evaluations are internally inconsistent, which is an epistemic failing regardless of whether any bet is offered. The betting scenario is a way of making the inconsistency visible, in the way a truth table makes a logical inconsistency visible.
This is the strongest version of the argument. It has a cost: once the pragmatic element is removed, the argument no longer explains why consistency of evaluation is required, and it starts to look like it presupposes what it was meant to prove. If the objection to incoherence is just that it is incoherent, we have a restatement rather than an argument.
The betting interpretation
The bridge from credence to betting rate is not free.
- Diminishing marginal utility. A bet at stake measures belief only if utility is linear in the payoff, which for money it is not. Using utils instead assumes a utility function, which on the Savage approach is derived jointly with probability — so the argument cannot presuppose it without circularity.
- Stakes affect beliefs. Being offered a bet is evidence: an eager bookmaker suggests they know something. A rational agent's credence may change on being offered the bet, undermining the elicitation.
- Non-doxastic attitudes to gambling. Someone with moral objections to betting, or who enjoys risk for its own sake, will not bet at their credences.
These are objections to the operationalisation, not to probabilism itself, and the depragmatised reading avoids most of them.
What the argument establishes
A fair summary. The Dutch-book theorem and its converse are genuine mathematical results, and they establish an exact correspondence between coherence and immunity to sure loss. What they do not establish, without additional premises, is that incoherence is an epistemic failing. Getting there requires either accepting that practical exploitability indicates epistemic defect — which is contestable — or adopting the depragmatised reading, which sacrifices some of the argument's independent force.
This is why the accuracy-based argument matters. It reaches the same conclusion from purely epistemic premises: credences aim at truth, and incoherent credences are guaranteed to be less accurate than a coherent alternative. Two arguments from independent starting points converging on the same conclusion is stronger evidence than either alone, and the accuracy argument is not vulnerable to any of the objections in this section.
Where this sits
Dutch books are the historical foundation of subjective Bayesianism and the reason coherence rather than truth-tracking became the position's central norm. The argument's limitations shaped the field: the pragmatic/epistemic gap motivated the accuracy program, the finite restriction left countable additivity unsettled, and the weakness of the diachronic version is why conditionalization still needs independent defence.