Probability as a Measure of Belief

May 2026 · 12 min read

So far we have motivated probability through repeated experiments and relative frequencies. But many real probability statements have nothing to do with repetition: “The chance it rains in San Francisco tomorrow is $30\%$”, “The probability the Dodgers win the World Series is $1/4$”, “The probability that life exists on Europa is $1/100$”. There is no “long-run” in any of these — tomorrow is one event, the World Series happens once, Europa exists once.

The subjective (or Bayesian) interpretation handles these statements: $P(E)$ is a rational agent's degree of belief in $E$, expressed as a number in $[0, 1]$. The remarkable fact — and the punchline of this post — is that any coherent system of beliefs must obey the same three axioms. The mathematics doesn't care which interpretation you choose.

Degrees of Belief, Operationalised

How would we even measure a degree of belief? The standard trick is to imagine you are willing to bet on the event. Suppose I offer you a contract that pays $\$1$ if event $E$ occurs and $\$0$ otherwise. The maximum price you are willing to pay for that contract is your fair price for $E$, and we identify it with $P(E)$.

• If you'd pay $\$0.30$ for the rain ticket but no more, your $P(\text{rain}) = 0.30$.
• If you'd pay $\$1$ to win $\$1$ if today is a weekday (and today is Monday), $P(\text{weekday}) = 1$ — you're certain.
• If you'd pay $\$0$ for a contract that pays $\$1$ if $0 = 1$, then $P(\emptyset) = 0$.

This betting interpretation is operational: it gives us a procedure for eliciting probabilities from a person, even when there is no “long run” to appeal to.

The Dutch-Book Argument

An agent's beliefs are coherent if no opponent can construct a finite set of bets at the agent's stated prices that guarantees the agent loses money no matter what happens. Such a money-pump is called a Dutch book. The amazing theorem due to de Finetti (1937) says:

So if your beliefs deviate from these axioms in any way, a sufficiently clever opponent can extract free money from you. In that sense the axioms are forced upon any rational agent.

A Concrete Dutch Book

Suppose your beliefs about a coin toss are $P(H) = 0.6$ and $P(T) = 0.6$. These are non-negative and individually in $[0, 1]$, but they violate Axiom 2 because $H$ and $T$ are disjoint and exhaustive yet $P(H) + P(T) = 1.2 \ne 1$.

An opponent uses your prices to do the following:

1. Sells you the “heads” ticket for $\$0.60$ (you'll receive $\$1$ if heads, $\$0$ otherwise).
2. Sells you the “tails” ticket for $\$0.60$ (you'll receive $\$1$ if tails, $\$0$ otherwise).

You paid $\$1.20$. The coin lands heads or tails, so exactly one ticket pays $\$1$. You always net $-\$0.20$. The opponent pockets $\$0.20$ guaranteed. Dutch book.

Now suppose instead your beliefs were $P(H) = 0.4$ and $P(T) = 0.4$ (sum $0.8 < 1$).

1. Buys both tickets from you for $\$0.40$ each.
2. Pays you $\$0.80$ total. Then collects $\$1$ from you no matter the outcome.

Net loss to you: $\$0.20$ guaranteed. Same kind of Dutch book in reverse.

The only set of prices for $\{H, T\}$ that no opponent can exploit is $P(H) + P(T) = 1$ — i.e.\ Axiom 2. The argument generalises: pricing failures of any axiom open you up to a Dutch book.

Visual: When Beliefs Are Incoherent, You Lose

Step through the heads-priced-at-0.6, tails-priced-at-0.6 example to see how the opponent extracts free money on every outcome.

Frequentist vs Bayesian, Reconciled

The frequentist interpretation says $P(E)$ is the long-run relative frequency of $E$. The subjectivist says $P(E)$ is a degree of belief, calibrated by willingness to bet. They sound philosophically different but both produce functions satisfying the same three axioms, so all theorems we prove (Bayes' rule, the central limit theorem, the law of large numbers, etc.) apply to both.

You can even view the law of large numbers as the bridge: if the experiment really can be repeated under identical conditions and beliefs are calibrated, the long-run frequency of $E$ converges to $P(E)$ with probability $1$. The subjective and frequentist values agree in the limit.

Updating Beliefs: A Preview of Conditioning

The real power of the subjective view is that it tells us how beliefs should change as evidence comes in. If $E$ is your hypothesis and $D$ is observed data, your updated belief should be the conditional probability

$\displaystyle P(E \mid D) = \frac{P(E \cap D)}{P(D)}.$

This is Bayes' rule, the topic of Ross's Chapter 3. Once you accept that beliefs obey the axioms, Bayes' rule follows automatically as the unique consistent way to update them. Anthropic, Google, and modern ML systems are built on this recipe.

When You Choose the Probabilities Yourself

In real applications you rarely “know” $P$ analytically. You make modelling choices. Three useful checks:

1. Coherence. Do your assigned probabilities satisfy the axioms? If not, fix them — your future inferences will contradict each other otherwise.
2. Calibration. Among events you've assigned probability $p$, do they actually occur about $100p\%$ of the time? Calibrated beliefs match the data.
3. Sensitivity. If small changes to $P$ produce big changes to your decision, you may be over-confident. Robust decisions tolerate prior uncertainty.

Common Mistakes

• Treating subjective probability as “less rigorous”. The mathematics is identical to the frequentist mathematics; only the interpretation differs.
• Reporting probabilities that don't sum to $1$ (e.g.\ "70% chance of rain, 70% chance of sun"). That's a textbook incoherent prior.
• Mistaking calibration for accuracy. A weatherman who always predicts $30\%$ rain may be perfectly calibrated yet give no information about which days will be rainy.
• Forgetting that subjective probabilities can be updated with data via Bayes' rule. Rational beliefs are not static.

Practice

1. Construct a Dutch book against an agent with $P(A) = 0.5$, $P(B) = 0.5$, $P(A \cup B) = 0.6$ for events with $A \cap B = \emptyset$. Identify which axiom is violated.
2. Suppose your subjective probabilities are $P(\text{rain}) = 0.4$ and $P(\text{rain or snow}) = 0.3$. Why is this incoherent? What does the axiomatic fix look like?
3. An agent with prior $P(\text{coin biased toward heads}) = 0.6$ observes a head. Use Bayes' rule to update the prior, given that the conditional probability of a head under "biased toward heads" is $0.7$ and under "fair" is $0.5$.
4. Why does the Dutch-book theorem only force finite additivity, not countable additivity? (Hint: a Dutch book is a finite collection of bets.)
5. Read about proper scoring rules (Brier score, log score) and explain why they incentivise reporting your true subjective probabilities.

Series Recap

Across this 7-part series we built probability from the ground up:

1. Series Overview — what's coming, why axioms matter.
2. Sample Spaces and Events — the set-theoretic vocabulary every later post depends on.
3. The Three Axioms — non-negativity, normalization, countable additivity, motivated by relative frequency.
4. Simple Propositions & Inclusion–Exclusion — the small list of consequences that all later proofs reuse.
5. Equally Likely Outcomes — the classical formula and its tight bond with combinatorial counting.
6. Continuity of Probability — limits of monotone event sequences, the dynamic face of countable additivity.
7. Probability as a Measure of Belief — the Dutch-book argument: rational beliefs must obey the same axioms.

From here Ross moves to conditional probability, independence, and Bayes' rule (Chapter 3) — the next sub-series in this Math blog.