The Material Theory of Induction
The material theory of induction is John D. Norton’s account of inductive inference, developed in The Material Theory of Induction (BSPS Open / University of Calgary Press, 2021; open access) and extended in The Large-Scale Structure of Inductive Inference (2024). Its central claim is that the warrant for an inductive inference is always a fact about the subject matter — a local, “material” truth — and never a universal formal schema: “Inductive inferences are warranted by facts not by formal schema… the warrant is ‘material’ and not formal.” 1 The Lakatos Award selectors summarized the two books as offering “sweeping but compelling arguments for two theses: (i) there can be no content-neutral, formal, universally applicable rules or procedures for justifiable inductive inference and (ii) real inductive inferences in science are nevertheless often justified, with the key work in the justification being performed by local ‘material’ facts.” 2
Norton’s own framing is constructive rather than skeptical: the theory “does not seek to undermine inductive inference; it seeks to save it” by relocating “the warrant of inductive inferences from rules to facts.” 3
The core claim
Facts play a dual role in inference: a factual statement can serve both as a premise and as the warrant authorizing the step to a conclusion — much as “if A then B” does in deductive logic. What distinguishes inductive inference is ampliation — its conclusion asserts more than its premises contain — so the premises alone can never warrant it; something extra must restrict the situations in which the conclusion holds, and that something is a factual truth about the domain. The theory’s slogan is “All induction is local”: every inductive schema is confined to the domain of the facts that warrant it, and the millennia-long search for a universal inductive logic — the “formal approach” — has failed. 1 4
The foundational argument (Ch. 2)
The most compact argument (Ch. 2, §2.5) answers the question of what powers induction: “It is powered by facts.” Its first premise is that inductive inference is ampliative; deductive inference, by contrast, merely restates what the premises already contain (the warrant “lies fully within the premises”). The second premise is that nothing in an ampliative inference’s premises can supply the missing warrant: a prediction that the next winter will be snowy, or that all radium chloride shares one crystal family, is “viable only in certain worlds” — hospitable ones. The warrant must be a background fact (such as Haüy’s principle of crystal isomorphism) that makes the world hospitable to that inference. Schemas may be reused, but only where their warranting facts obtain: “All induction is local” is “the second major tenet of a material theory of induction.” 4
Illustration: crystal forms and Curie’s radium (Ch. 1)
Norton’s flagship illustration is Marie Curie’s inference from a few grains of radium–barium chloride to the crystallographic properties of all radium — an inference of breathtaking scope from a single sample. A Bayesian reconstruction fails to capture it (she “did not take a large n limit”; the analysis imports external inductive content in its description of the evidence). The material analysis succeeds: already-established facts — the close chemical analogy of barium and radium, and the well-established isomorphism between chemical and crystallographic similarity — supplied the warrant. We can generalize the crystal family of a substance from one sample because Haüy’s principle holds; we cannot generalize, say, the sample’s size, because no comparable fact restricts size. 1
“Why Not Bayes”: the boundary of Bayesian universality (Ch. 10)
Chapter 10 argues that Bayesian analysis — taken as a universal logic of induction — fails at a discernible boundary. Norton separates its sound uses from its imperial claim: “I have no quarrel with the application of Bayesian analysis in specific cases… My concern is the claim that Bayesian analysis is universally applicable to all systems, where it can supply the one, true logic of inductive inference.” 5
The chapter’s central weapon is completely neutral support (§§10.7–10.9). Neutrality (from the principle of indifference) is shown to require, of any completely uninformative background: [tautology | Ω] = 1; [any contingent proposition | Ω] = I; [contradiction | Ω] = 0 (equation (5); the book’s equations print the tautology symbol as Ω — some running-text passages print it as ‘W’). Applied to von Mises’ wine-and-water case (two mixtures whose ratios x, y are each known only to lie in 0.5–2), the neutrality requirements and the structure of the events collide: the analysis forces [A₁ | Ω] = [A₂ | Ω] = [A₃ | Ω] = I and P(A₁ | Ω) = 1/3, while the same event’s decomposition as B₂ ∨ B₃ forces 2/3 — an outright contradiction (1/3 = P(A₁ | Ω) = P(B₂ ∨ B₃ | Ω) = P(B₂ | Ω) + P(B₃ | Ω) = 2/3). Neutral support — the device meant to ground “objective” priors — cannot be delivered. 5
The chapter generalizes: retreats to imprecise probability do not escape the boundary (§§10.6, 10.10); “all proofs of the necessity of probabilities are circular” (§10.11, with illustrations in §10.12); and the Dutch book argument’s notion of a “fair bet” is itself circular, its regresses familiar from the justification trilemma (§10.13). The conclusion presses the problem of the priors — priors “antecedent to the consideration of evidence” introduce “an arbitrariness into the analysis that has been the bane of all forms of Bayesianism”; objective Bayesians’ distinguished-prior proposals (such as Jaynes’ maximum entropy principle) are among the failures. 5
The regress problem (Epilog)
If warrants are facts, and those facts are themselves supported by further inductive inferences, a regress threatens — “the analog of Hume’s problem of induction, but now played out in the material theory.” The Epilog acknowledges it (“There’s something like a regress lurking about. It is something I should address”) and promises a fuller treatment elsewhere (see Norton 2014, “A Material Dissolution of the Problem of Induction,” Synthese 191: 671–90) — a thread the 2024 sequel takes up. 6
Reception
- The Material Theory of Induction was discussed in a Metascience symposium (2022; commentaries by Mousa Mohammadian, William Peden, Elay Shech; author’s response); the sequel in a Journal for General Philosophy of Science symposium (2025) and a Studies in History and Philosophy of Science book forum (replies to Bolinska, Kao, Scholl). 7
- The two books won the 2026 Lakatos Award, announced by LSE on 14 September 2026 — praised as “a work of exceptional scholarly rigour” whose writing is “unusually clear and readable, indeed almost uniquely so.” The prize lecture at LSE is to come. 2
Related pages
- john-d-norton — author; career, empiricism, and the 2026 Lakatos Award
- bayesian-apologetics — the critique’s most-watched application domain
- subjectivity-of-priors — the priors problem, pressed by the neutral-support result
- post-hoc-probability-fallacy — complementary failure mode of probability arguments
- fine-tuning-argument — where indifferent priors over universes are invoked
- reasoning-methods — where induction sits among the modes of inference
- kolmogorov-probability-axioms — the calculus whose universal applicability the chapter bounds
- david-hume — Hume’s problem, recast in material form