Two readings of probability

Over two hundred days, a forecaster announced a 30% chance of rain, and it rained on fifty-four of them. Is the announcement a claim about how often rain follows such a forecast, or a statement of how strongly the forecaster believed it would rain? Two readings of probability answer that question, and the arithmetic does not choose between them.

Are probabilities frequencies or beliefs? A probability can be read as the long-run share of trials in which an event occurs, or as the strength of a belief in a proposition. The first reading is called frequentist, and the second is called Bayesian [S:5:717]. Each reading licenses some questions and forbids others, and the difference matters most when the event happens once. !Each square is one forecast day, and the same 200 days are read twice: the announced 0.30 predicts 60 wet days, and the record shows 54, a frequency of 0.27. Calibration is the match between the announced number and the observed share over many such blocks. Status: synthetic.

The frequentist reading ties a probability to a repeatable process. The probability of an event is the fraction of trials in which the event occurred, so the statement is a claim about a long sequence rather than about a single case [S:5:717]. A coin, a die and a clinical trial fit this reading, because each admits many repetitions under similar conditions.

The Bayesian reading ties a probability to a state of knowledge. A probability is a degree of belief in a proposition, and the proposition may be a one-off: that a particular number is prime, that a defendant is guilty, that it rains tomorrow [S:5:717]. The reading is broader, and it pays for the breadth with a dependence on prior belief that the frequentist reading avoids.

Worked example. A forecaster issues a 30% chance of rain on two hundred separate days. Under the frequency reading, the number predicts the share of wet days in that set, so the expected count, the number of wet days you would predict before looking, is 0.30 × 200 = 60.

$$0.30\times 200=60$$

That line says: a 30% chance, applied to two hundred days, predicts sixty wet days. [F01]

The record shows fifty-four wet days rather than sixty. The observed frequency is 54/200, which equals 0.27. A frequentist reads the record as a measurement of the forecaster's long-run rate and finds it slightly below the announced 30%. A Bayesian reads the same record as evidence about the forecaster's calibration, the match between announced probabilities and observed frequencies over many cases.

The two readings agree about the arithmetic and disagree about what the arithmetic is about. Both use the sum rule, both use conditioning, and both accept Bayes' rule, which follows from the definition of conditional probability rather than from a new assumption. Neither reading changes the answer to a calculation once the inputs are fixed. They disagree about which inputs are meaningful, and about which events admit a number at all, so a disagreement about vocabulary can be a disagreement about whether there is a number at all.

The Bayesian reading attaches the probability to the reasoner's information, so new information changes the number. The frequentist reading attaches it to the repeatable process, so a change of information changes the number only by changing the reference class, the group of cases over which the frequency is counted. The two readings do not share the claim that the number is a function of information alone.

Tomorrow will be one day, and it will either rain or it will not. A forecast of 30% is not reporting the share of rainy tomorrows, because there is only one tomorrow. The number still does work: it changes what you carry and what you insure.

The frequentist reading has no long-run frequency to offer for a unique event. It cannot assign one to the statement that a particular number is prime, because the number does not get drawn again [S:5:717]. The Bayesian reading assigns a belief to that statement without difficulty, and the belief changes as evidence about the number arrives.

Take a hundred-digit number from a cryptography key. The number is either prime or it is not, so no repetition of this number is available. A frequentist can estimate the density of primes near it, meaning how many numbers in that range are prime, and quote that estimate as a statement about the class of numbers rather than about this one. The textbook lists the density estimate as one possible assignment and then takes the position that a probability for this particular number is nonsense, because the number is either prime or not [S:5:717]. A Bayesian assigns the belief directly and revises it when a primality test returns evidence.

The Bayesian reading has trouble of its own. Two people with different priors can reach different posteriors from the same evidence, and the reading supplies no single correct prior [S:5:717]. That is the qualification flagged in section 11.1: identical information and the same starting beliefs force the same number, and different priors are different inputs. Jaynes answers that the assignment describes a state of knowledge rather than a physical quantity, so agreement requires shared information rather than shared intuition [S:29:58].

In practice the two readings converge when the data are plentiful. A long sequence of trials pins down the frequency, and a likelihood that is sharp enough overwhelms a reasonable range of priors. They diverge when the data are scarce or the event is unique, and that is where the choice of reading shows up in the answer.

The idea has a cost, because neither reading is free. The frequentist must find repetitions that the world may not supply, and the Bayesian must defend a prior that the evidence may not settle. The cost is worth paying, since the alternative is to reason without a number at all, and this chapter has shown what the numbers can do.

The formal statement. The frequentist reading is often introduced through a limit of relative frequency:

$$P(\text{event})=\underset{n\to \infty }{\lim }\frac{\text{count of the event in }n\text{ trials}}{n}.$$

That line is von Mises' interpretation, not the definition used in modern statistics [S:48:29]. The definition is Kolmogorov's axioms, on which probability is a measure on a probability space [S:62:41]. Long-run frequency statements are theorems derived under independent, identically distributed sampling, and the limit can fail to exist for a particular sequence [S:42:103]. The Bayesian reading instead calls probability a degree of belief in a proposition, constrained by the sum and product rules [S:5:717].

The derivation. The two readings do not produce different arithmetic, but they do not share every axiom. Both can be set on a probability space. The shared core is non-negativity, normalisation, and finite additivity. Countable additivity is an extra axiom of Kolmogorov's system that not every Bayesian accepts: de Finetti keeps finite additivity and rejects some countable extensions [S:29:13]. The difference is the domain of the function. The frequentist attaches probability to repeatable events, while the Bayesian attaches it to propositions, including one-off ones [S:5:717]. Once the domain and the inputs are fixed, every calculation agrees. The readings disagree about which inputs are admissible.

The edge case. Take a one-off proposition, such as the statement that a particular large number is prime. The frequentist has no repetition, so the limit definition does not apply, and the textbook treats a probability here as nonsense [S:5:717]. The Bayesian assigns one and revises it as tests return evidence. The dispute is not about arithmetic. It is about whether the argument of the belief function is the right kind of object.

A second edge case. Calibration. A forecaster who announces 0.3 on many days can be checked: among those days, about 30% should be wet [S:35:316]. Calibration is a frequency property of a belief-shaped output, and it is the bridge between the readings. It can fail while every internal calculation is coherent, because calibration is a claim about the world rather than about the rules.

What the core hid. The core said the arithmetic is shared and the readings differ over which events carry a probability. That is correct. It hid that the frequentist interpretation needs a limit that may not exist, and that the Bayesian definition needs a prior that the evidence may not settle. It also hid that calibration is a third criterion, available to either reading.

Where the analogy breaks. The core used no analogy in this section. A metaphor would hide the real split: the Bayesian reading covers more propositions, and the frequentist reading makes a stronger claim about each proposition it covers.

Original sources. The frequentist limit appears as one interpretation in the probabilistic-artificial-intelligence text [S:60:12]. The collectives programme of von Mises, and the later work by Kolmogorov, are described in the conformal prediction introduction [S:48:29]. De Finetti's Bayesian reading is attributed in the same probabilistic-AI text [S:60:12]. The axiomatic core of the frequentist reading is Kolmogorov's [S:62:41].

An open question. The frequency definition is circular, and its reference class is undefined. Hutter states both objections: the limit may be anything, and saying that it equals one half with high probability presupposes probability [S:42:103]. A doctor who counts the disease rate among similar patients runs out of similar patients once every known feature is used [S:42:104]. Evidence that would settle a particular case: a reference class fixed before the data are seen, with a stated rule for membership, or a decision rule whose answer does not depend on the choice.

A frontier result. Conformal prediction attaches a coverage guarantee to each prediction set under an exchangeability assumption, and it grew out of the same debate about collectives and random sequences [S:48:30]. It gives a frequency guarantee for a single prediction problem, which is the case the classical limit definition could not reach. The guarantee is marginal: it holds on average over the calibration and test draws [S:48:4].

The readings also report different objects. A frequentist reports a confidence interval, whose coverage is a property of the procedure across repeated samples. A Bayesian reports a posterior interval, whose statement is about the parameter given the data and the prior [S:62:114]. The credible interval is defined at [S:62:101], which also says that it is not the same as a confidence interval. The two intervals can coincide numerically and still mean different things.

What to read next. Read the philosophy section in Mathematics for Computer Science [S:5:717], the first chapter of the probabilistic-AI text for both readings [S:60:12], Hutter's slides for the objections [S:42:103], and the conformal introduction for the modern use of exchangeability [S:48:30].