Philosophy of Probability

Reference notes on what probability means. The mathematics is settled: since Kolmogorov, a probability is a measure of total mass one, and the calculus that follows is developed in full in the Probability section. What the mathematics conspicuously does not supply is an interpretation. The axioms fix how probabilities combine once a measure is given; they say nothing about whether the numbers report frequencies in a sequence, physical tendencies of an experimental setup, degrees of rational belief, logical relations between propositions, or features of the best systematisation of the world's history. Nor do they say which measure a real situation calls for, or why anyone facing an uncertain future should care about the answer.

These pages take up the questions the mathematics leaves open. Is there objective chance — a genuine worldly magnitude — or is all probability an artifact of our ignorance? What are degrees of belief, and why should they obey the probability axioms rather than some other calculus? When a physical chance is known, why must rational confidence match it? Can probability solve, or even soften, the problem of induction? Does statistical dependence tell us anything about causation? And what are the probabilities in quantum mechanics and statistical mechanics probabilities of?

The distinctive feature of the section is that the formal results are already available and are cited rather than re-derived. The strong law of large numbers, de Finetti's representation theorem, the failure of countable additivity in the finitely additive setting, Martin-Löf randomness, and Gleason's theorem are all theorems with precise statements, and each is regularly enlisted in a philosophical argument that it does not by itself support. Keeping the theorem and the thesis apart is the main discipline these pages try to impose.

The treatment is philosophical — arguments, distinctions, and the structure of positions — rather than a course in probability or statistics. Estimation, hypothesis testing, and statistical practice are not covered.

Status. All twelve folders are written. Further work would be refinement and cross-linking rather than new pages.

Prerequisites

Nothing beyond the ability to read $P(A\mid B)$ is strictly required. Where a page leans on a real theorem it links to the statement in the mathematics section rather than assuming it. Readers who want the formal underpinning first should take the Kolmogorov axioms, random variables, and conditional expectation in that order; readers who want the philosophy can start here and follow the links backwards when a result is invoked.

The division of labour between the two sections:

Contents

The spine

• What Is the Philosophy of Probability? — the field defined: the three master questions (the metaphysical, the epistemic, and the application question), the map of interpretations, why the axioms are neutral among them, and what the section is not.
• Chance and Credence — the organising distinction: objective chance as a putative feature of the world versus credence as an agent's graded confidence. Same calculus, different subject matter, and the coordination problem of saying how the two are related.
• Interpretations of Probability: An Overview — the comparative map: what each interpretation takes probability to be, how it handles the single case, what it says about priors, and its characteristic failure. Includes the criteria of adequacy the later pages apply.
• Models, Measures, and Applications — how a formal measure acquires empirical content. Sample spaces are chosen, not found; Bertrand's paradox as a problem about model selection rather than a defect in the axioms.

Interpretations

What probability is. The first two are historical and largely superseded, but their failure modes recur throughout; the middle two are theories of objective chance; the last three divide the epistemic and Humean ground between them.

• Classical Probability — Laplace's ratio of equipossible cases and the Principle of Indifference; the circularity charge; the wine–water and cube-factory paradoxes; and the vindication of physical symmetry by the method of arbitrary functions.
• Logical Probability and Inductive Logic — probability as partial entailment, its kinship with logicism, the Keynes–Carnap program, and the five results that broke it: language dependence, the $\lambda$-continuum, zero probability for universal laws, Putnam's diagonal argument, and indifference.
• Frequentism — finite and limiting relative frequency, von Mises collectives and place selection, and the objections: the single case, the reference class, order dependence, and why the law of large numbers cannot supply the definition.
• Propensity Interpretations — chance as a physical disposition of a setup; long-run versus single-case versions; Humphreys' paradox and the asymmetry of causal tendency; the problem of deterministic worlds.
• Subjective Bayesianism — credence as coherent betting rate; Ramsey and Savage representation theorems; what de Finetti's theorem and merging-of-opinions do and do not establish; permissiveness as the standing cost.
• Objective Bayesianism — the three norms of probability, calibration, and equivocation; maximum entropy and transformation groups; Cox's theorem and its regularity assumptions.
• Humean Best-System Chance — chance as a term in the best summary of the mosaic; simplicity, strength, and fit; deterministic chance as a positive result; undermining, the Big Bad Bug, and the New Principle.

Rational credence

How degrees of belief ought to behave. The first three pages argue for the Bayesian norms; the last three examine where they strain.

• Probabilism — the thesis that rational credences obey the axioms; whether beliefs come in degrees at all; finite versus countable additivity; regularity; permissivism and uniqueness.
• Dutch-Book Arguments — the construction and its converse theorem, the diachronic extension, and the pragmatic/epistemic gap that motivates the depragmatised reading.
• Accuracy and Scoring Rules — the Brier score, strict propriety, and Joyce's dominance theorem; why coherence is convexity; the accuracy-first program.
• Conditionalization and Reflection — the dynamic norm, Jeffrey's generalisation, the martingale connection, and the two places the framework strains: zero priors and genuinely new hypotheses.
• Priors and Indifference — the unsolved core: washing out and its three qualifications, invariance and maximum entropy, improper priors, dogmatism, and simplicity.
• Imprecise Probability — credal sets and lower previsions, Ellsberg and ambiguity, and the costs: dilation and decision paralysis.

Chance and laws

What objective chance is for, and how it behaves.

• The Principal Principle — the bridge between chance and credence; the admissibility proviso and why it resists definition; the Big Bad Bug and the New Principle.
• Determinism and Chance — determinism versus predictability; three routes to deterministic chance, including the method of arbitrary functions; and why chance cannot supply agential control.
• Chance and Explanation — the failure of the high-probability requirement in both directions; statistical relevance; whether an irreducible chance can explain an individual outcome.
• Chance over Time — time-indexing, the martingale property, the transition problem, and what the asymmetry of chance does and does not show about the openness of the future.
• Typicality and Probability — "almost all" versus "probably"; the measure problem; the Past Hypothesis; Bohmian quantum equilibrium and Boltzmann brains.

Induction and confirmation

What probability can and cannot do for the theory of evidence.

• The Problem of Induction — Hume's dilemma and Goodman's new riddle; Reichenbach's pragmatic vindication; why Bayesian conditionalization systematises induction without justifying it; Williams–Stove and the no-free-lunch results.
• Bayesian Confirmation — confirmation as probability raising, incremental versus absolute, the competing measures, and the successes: the ravens, diverse evidence, severity, and Duhem–Quine.
• Likelihood and Evidential Support — the law of likelihood, Bayes factors and the Ockham factor, the likelihood principle and the stopping-rule dispute, and Mayo's severity as the error-statistical rival.
• Old Evidence and New Theories — why known evidence cannot confirm, Garber's learning-the-entailment repair, the harder problem of hypotheses that did not exist, and prediction versus accommodation.
• Underdetermination, Simplicity, and Model Choice — three grades of underdetermination, overfitting as the strongest argument that simplicity is truth-conducive, and the language-relativity that remains.
• Testimony and Disagreement — the likelihood-ratio model, why correlated witnesses cap the evidence, peer disagreement, and the impossibility results for pooling credences.

Probability and causation

What statistical dependence can establish about causal structure.

• Probabilistic Causation — probability raising, why it is neither necessary nor sufficient, contextual unanimity, and the type/token distinction.
• Screening Off and Common Causes — Reichenbach's principle, conjunctive forks and colliders, the causal Markov and faithfulness conditions, and the quantum failure via Bell.
• Simpson's Paradox — reversal under aggregation; confounder, mediator, and collider as three roles the statistics cannot distinguish; the Berkeley admissions case.
• Causal Models and Intervention — DAGs and structural equations, the do-operator and graph surgery, the back-door criterion and the ladder of causation, and causal discovery.
• Correlation, Prediction, and Explanation — Berkson's paradox and collider bias, why predictive accuracy is not causal understanding, and what convergent evidence can establish.

Puzzles and pressure tests

Six cases where an apparently determinate probability question turns out to depend on a modelling choice.

• Bertrand's Paradox and Model Selection — the three answers, Jaynes's invariance resolution and its limits, and the general structure the other puzzles share.
• Conditioning on Null Events — the Borel–Kolmogorov paradox, why conditional expectation is defined only up to null sets, and how the notation $P(A\mid X=x)$ hides a choice of disintegrating variable.
• The Reference-Class Problem — why no statistical rule selects a class, how the difficulty varies by interpretation, and its consequences in law, insurance, and medicine.
• The Lottery and Preface Paradoxes — the Lockean thesis, conjunction closure, and consistency cannot all hold; the responses, including Leitgeb's stability account.
• Sleeping Beauty — halfers and thirders, why the framework has no agreed extension to self-locating belief, and why the betting argument shows less than it appears.
• Anthropic and Self-Locating Reasoning — SSA and SIA, Doomsday, the presumptuous philosopher, fine-tuning's measure problem, and Boltzmann brains as the one sound application.

Scientific case studies

• Quantum Chance — Gleason's theorem and what it does not fix; collapse, Bohmian, Everettian, and QBist readings; what Bell settles about the structure of quantum probability.
• Probability in Statistical Mechanics — Gibbsian and Boltzmannian frameworks, Loschmidt and Zermelo, the Statistical Postulate and the Past Hypothesis, and four readings of the probabilities.
• Randomness: Processes and Individual Outcomes — the process/product distinction, the Levin–Schnorr convergence of incompressibility, unpredictability, and typicality, and pseudorandomness.
• Free Will and Chance — why the luck objection follows from what a chance is, and what remains available to the libertarian.

History, method, and figures

• A History of the Philosophy of Probability — Pascal to the present, keeping to arguments still in play; why formalisation settled less than expected.
• Method: What Formal Results Can Establish — five kinds of result, eight recurring invalid inferences, the role of idealisation, and the suppressed model-relativity clause.
• Key Figures — a compact reference to the principal contributors and the claim each is cited for.

Planned

The section is complete as planned. Possible extensions are noted in the roadmap at work/philosophy-of-probability-roadmap.md: a general philosophy-of-causation section, a decision-theory section, a philosophy-of-statistics folder, and a plain-language companion under general/.

How the spine fits together

graph TD
  I[1 What Is the Philosophy of Probability?]
  CC[2 Chance and Credence]
  IO[3 Interpretations Overview]
  MA[4 Models and Applications]
  I --> CC
  CC --> IO
  I --> MA
  CC -.->|"objective side"| OBJ[chance, laws, science]
  CC -.->|"epistemic side"| EPI[credence, confirmation]
  IO --> OBJ
  IO --> EPI
  MA -.-> PUZ[puzzles: Bertrand, reference classes]

Read the four spine pages in order. Chance and Credence is the page the rest of the section depends on most heavily: nearly every dispute below turns out to be either a disagreement about which of the two notions is fundamental, or a confusion between them.

Scope

• No probability theory. Measure-theoretic probability is developed in math/11-probability and cited here. These pages state theorems; they do not prove them.
• No statistics. Estimation, testing, model fitting, and experimental design are a different subject. Where statistical method raises a philosophical issue — the likelihood principle, stopping rules, the reference class — the issue is treated and the method is not taught.
• No general decision theory. Expected utility, risk, and the causal/evidential dispute appear only where they bear on whether credences must be probabilities.
• No general theory of causation. The causal pages ask what probability can and cannot establish about causal structure; counterfactual, process, and powers theories appear as contrasts, with the ontology treated in Ontology.
• No preferred interpretation. Each position is given its strongest formulation and its hardest objection. Where these notes take a stand, it is on the narrow and largely negative claim that several familiar inferences from formal results to interpretive conclusions are invalid.

Relation to the rest of the book

Probability is already in use across these notes, generally as an undefined primitive, and part of the point of the section is to give those uses a home:

• Free will turns on whether indeterminism — the existence of non-trivial chances — helps or hurts agency, and the luck objection is a claim about what an objective chance can and cannot underwrite.
• Laws of nature in the best-system account are probabilistic laws, and the account needs a story about what their chances are and why credence should track them.
• Philosophy of religion runs its cumulative case in explicitly Bayesian terms, and both the fine-tuning argument and the treatment of testimony for miracles depend on contested assumptions about priors and reference classes.
• The thermodynamic arrow rests on a statistical postulate whose status — chance, typicality, or ignorance — is left open there and taken up here.
• The philosophy of Bell's theorem already works with local response probabilities and Reichenbach's common-cause principle, which are exactly the notions the causation pages analyse.