Interpretations of Probability: An Overview

An interpretation of probability answers the question: what has to be the case for "$P(A)=x$" to be true? The answers on offer are strikingly diverse — a ratio of cases, a relation between propositions, a frequency in a sequence, a disposition of a physical setup, a coherent betting rate, a term in the best systematisation of the world — and each was developed to solve a real problem that its predecessors failed on.

This page is the map. It states each interpretation in capsule form, sets out the criteria of adequacy the individual pages apply, and compares the positions along the dimensions that actually discriminate between them. The deep-dives live in the interpretations folder; the point here is to see the logical geography, and in particular to see which disagreements are genuine rivalries and which are differences of subject matter.

Criteria of adequacy

Before the positions, the tests. Hájek's list, in the form used throughout this section, gives five things an interpretation should deliver. No interpretation scores well on all five, which is why the field is still open.

1. Admissibility — the interpreted quantity must actually satisfy the axioms. If it fails additivity or normalisation, it is not an interpretation of probability but a change of subject.
2. Ascertainability — there must be some method, at least in principle, of finding out what the probabilities are. An interpretation on which probabilities are real but permanently unknowable is epistemically idle.
3. Applicability — the interpretation must explain why probability is useful in the ways it demonstrably is: for inference, for prediction, for decision, and in the sciences. This is where most interpretations do their real competing.
4. Single-case applicability — it should make sense of the probability of a particular event, not just of a class. Physics and ordinary life both traffic in single-case claims.
5. Explanation of the calculus — an interpretation should make it intelligible that probabilities combine as the axioms say, rather than merely stipulating that they do.

A sixth constraint, less often listed but doing much of the work in contemporary debates, is guidance: the interpretation should explain why probabilities are relevant to what an agent should believe and do. This is the demand the Principal Principle codifies, and objective interpretations that cannot meet it are in trouble regardless of their metaphysical merits.

The positions in capsule

Classical

Probability is the ratio of favourable to equipossible cases. Laplace's formulation, and the historical starting point: the probability of an even roll is $3/6$ because three of six equally possible faces are even.

Its strength is that it is immediately applicable to the symmetric setups that motivated the subject. Its difficulties are severe and were recognised early. "Equipossible" either means "equiprobable" — in which case the definition is circular — or means something else that must be specified. It has no application to biased coins, to continuous quantities, or to any case where symmetry is absent. And the Principle of Indifference that underwrites it generates contradictions when possibilities can be carved more than one way, as Bertrand's paradox shows. Full treatment: classical probability.

Logical

Probability is a degree of partial entailment between propositions: $P(h\mid e)$ measures the extent to which $e$ logically supports $h$, with deductive entailment as the limiting case. Keynes, Carnap, and W. E. Johnson developed the program, whose ambition was to make inductive inference a branch of logic and to fix rational credence uniquely given evidence.

The program is the most systematic attempt at objectivity on the epistemic side, and its failure is instructive rather than embarrassing. The confirmation function depends on the language chosen, so Goodman's grue supplies a rival with equal formal credentials; Carnap could not select a unique member of his own continuum of inductive methods on logical grounds; and universal generalisations receive probability zero over infinite domains, making confirmation of a law impossible. Its descendants are objective Bayesianism and maximum-entropy methods. Full treatment: logical probability and inductive logic.

Frequentist

Probability is relative frequency: either the actual frequency in a finite class, or the limit of the relative frequency in an infinite sequence. Venn, von Mises, and Reichenbach are the principal developers; von Mises added a randomness condition — the frequency must be invariant under admissible place selections — to distinguish genuine collectives from merely regular sequences.

Frequentism scores best on ascertainability and has an obvious claim to explain why probability matters in statistics and the sciences. Its problems are the ones every treatment rehearses: finite frequentism makes probabilities depend on how many trials happen to occur and rules out irrational values; limiting frequentism appeals to sequences that do not exist and makes the limit depend on the ordering; and the reference-class problem means no individual event has a probability except relative to a class the theory does not select. It fails criterion 4 outright. Full treatment: frequentism.

Propensity

Probability is a physical disposition — a tendency of a chance setup to produce an outcome. Popper introduced the view partly to make sense of single-case quantum probabilities, where there is no repeated trial to take frequencies over.

Propensity is the natural home for objective single-case chance, and it connects probability to the broader metaphysics of dispositions and powers. Its liabilities are that the notion is close to primitive — critics say "propensity" names the problem rather than solving it — and that propensities may not satisfy the axioms. Humphreys' paradox presses this: probabilities invert freely by Bayes' theorem, so from $P(\text{effect}\mid \text{cause})$ one obtains $P(\text{cause}\mid \text{effect})$; but a causal tendency does not run backwards, so if propensities are causal they are not probabilities. Full treatment: propensity interpretations.

Subjective Bayesian

Probability is an agent's degree of belief, constrained only by coherence. Ramsey, de Finetti, and Savage developed it; degrees of belief are elicited from preferences or betting behaviour, and the only rational requirement is that they satisfy the axioms — enforced by Dutch-book arguments showing that incoherent agents accept sets of bets guaranteeing loss.

The position scores superbly on applicability and single-case applicability, since any proposition whatever can have a credence, and it gives probability a clear role in inference and decision. Its notorious cost is permissiveness: two agents with the same evidence may hold wildly different coherent credences, and nothing in the theory faults either. Its defenders answer with convergence and merging results; its critics reply that these are asymptotic and that science appears to involve probabilities not hostage to anyone's prior. Full treatment: subjective Bayesianism.

Objective Bayesian

Probability is the degree of belief that the evidence rationally warrants — credence, as with subjectivism, but with additional constraints beyond coherence: symmetry, invariance, maximum entropy, and calibration to known chances. Jaynes is the best-known advocate; Cox's theorem is often cited as showing that any measure of plausibility meeting weak structural desiderata must obey the probability calculus.

The view inherits the ambitions of the logical interpretation with the machinery scaled back from a full inductive logic to constraints on priors. Whether the constraints can be specified without arbitrariness — and whether maximum entropy avoids the language-dependence that sank Carnap's program — is the open question. Full treatment: objective Bayesianism.

Humean best-system

Probability is a chance appearing in the best system: the axiomatisation of world history that optimally balances simplicity, strength, and fit, where fit measures how probable the actual course of events is according to the system. Lewis's proposal, refined by Loewer and others.

It is the leading way to have objective chance within a broadly naturalistic metaphysics that admits no irreducible modal facts, and it connects directly to the best-system account of laws of nature. Its difficulties are that chances then depend on the whole of history, including the future, which generates the undermining problem and the resulting tension with the Principal Principle. Full treatment: Humean best-system chance.

Comparison

Two structural observations

The columns are not all rivals. Frequentism, propensity, and best-system chance answer the metaphysical question; subjective and objective Bayesianism answer the epistemic one. It is perfectly consistent to be a propensity theorist about chance and a subjectivist about belief — indeed that combination is common — with the Principal Principle joining the halves. Treating the seven positions as competitors for a single office is the most frequent structural error in introductory presentations, and it makes several positions look weaker than they are by faulting them for not answering a question they never addressed.

Pluralism is a live conclusion, not a failure of nerve. A century of attempted reductions has produced no interpretation that meets all six criteria. One reasonable response is that "probability" is not univocal: physical chance, evidential support, and rational credence are distinct notions sharing a calculus because the calculus is weak enough to accommodate all three. This is Hájek's position, and its cost is the obligation to explain what unifies them — an obligation the monist does not incur.

Where this leaves us

The interpretations divide first by what they are theories of — the world or the mind — and only then by their internal machinery. Reading the table by column rather than by row shows where the real work remains: no objective interpretation gives a satisfying account of guidance, and no epistemic interpretation gives a satisfying account of objectivity.

Before the individual positions can be assessed, one further question has to be faced, and it cuts across all of them. Every interpretation assumes that some probability model is the right one for a given situation. But sample spaces are chosen rather than discovered, and rival choices give incompatible answers. The next page takes up how a formal measure comes to represent a real setup — the problem that Bertrand's paradox makes vivid and that indifference-based interpretations founder on.