Objective Bayesianism

Objective Bayesianism keeps the Bayesian apparatus — credence, coherence, conditionalization — and adds what subjectivism refuses: further constraints that narrow the range of rationally permissible credences, ideally to one. It is the heir of the logical interpretation with the ambitions scaled back. Where Carnap wanted a complete inductive logic assigning a unique $P(h\mid e)$ for every pair of propositions in a formal language, the objective Bayesian wants only that evidence should substantially constrain priors, and is content if the constraints leave some latitude.

The motivation is the permissiveness objection. If coherence is the sole requirement, an agent may hold any prior at all, and the appearance that science delivers the probability of a hypothesis is an illusion. Objective Bayesians take that appearance seriously.

This page sets out the constraints in the standard three-norm form, the maximum-entropy machinery that implements them, Cox's theorem, and the objections — including the question, pressed by critics, of whether the position is an interpretation of probability at all.

Three norms

Jon Williamson's formulation is the clearest statement of the position's structure. Rational credence is governed by three requirements, in increasing order of controversy.

1. Probability. Credences must satisfy the axioms. Shared with subjectivism, and justified by the same Dutch-book and accuracy arguments.
2. Calibration. Credences must respect what is known about physical chances and observed frequencies. If the agent knows the chance of heads is $0.7$, credence should be $0.7$. This is the Principal Principle appearing as a constraint on priors rather than as a separate bridge; more generally, credence should lie within the range of chance values compatible with the evidence.
3. Equivocation. Subject to 1 and 2, credences should be as non-committal as possible — spread as evenly as the constraints allow. This is the descendant of the Principle of Indifference, and it is where all the difficulty lies.

The first two norms are widely accepted; a subjectivist who accepts the Principal Principle already has both. The third is the distinctive commitment, and it is what makes the position objective rather than merely chance-respecting.

Maximum entropy

The technical implementation of equivocation is the maximum entropy principle, developed by Jaynes: among all distributions consistent with the constraints, choose the one maximising the Shannon entropy

$$H(P)=-\sum _{i}P(x_i)\log P(x_i)$$

The rationale is information-theoretic. Entropy measures uncertainty, so maximising it is choosing the distribution that assumes least beyond what is known. Any lower-entropy distribution encodes information the agent does not possess.

The method has real successes. With no constraints it returns the uniform distribution, recovering indifference in the cases where indifference is right. Given a known mean it returns the exponential family; given a known mean and variance on the real line it returns the Gaussian — which is a satisfying result, since it explains the ubiquity of the normal distribution as reflecting minimal commitment beyond first and second moments rather than a fact about nature. And in statistical mechanics the maximum-entropy distribution subject to a known expected energy is the Boltzmann distribution, which lets Jaynes present the whole apparatus of equilibrium statistical mechanics as an exercise in inference rather than in dynamics.

Transformation groups extend the method to continuous cases, and this is the sharper tool. If a problem is invariant under a group of transformations, the prior should be invariant too, which often determines it uniquely. For a scale parameter — a quantity with no natural unit — invariance under $x\mapsto cx$ forces the Jeffreys prior $p(x)\propto 1/x$, not the uniform prior. This is a genuine advance over the classical Principle of Indifference, because it replaces "no reason to discriminate" with an explicit and checkable symmetry claim, and it is Jaynes's proposed resolution of Bertrand's paradox.

Cox's theorem

Objective Bayesians frequently appeal to Cox's theorem as showing that the probability calculus is not one option among many but the unique consistent extension of deductive logic to graded belief.

The theorem assumes that a degree of plausibility $(A\mid B)$ is a real number, that it depends only on the propositions involved, and that it satisfies structural conditions: the plausibility of $\neg A$ is a function of that of $A$; the plausibility of $A\wedge B$ is a function of that of $B$ and of $A$ given $B$; and equivalent derivation routes agree. From these, plus regularity conditions, the plausibility measure is shown to be isomorphic to a probability measure.

If sound, this answers the "explains the calculus" criterion better than anything else on offer, since it derives the axioms from desiderata about reasoning rather than about betting. The caveats are that the regularity assumptions are doing real work — Halpern produced finite counterexamples showing the result fails without continuity or differentiability conditions — and that assuming plausibility is a real number already excludes the imprecise alternatives, which is close to assuming the conclusion.

Objections

Equivocation inherits the paradoxes. Maximum entropy is defined relative to a partition or a parameterisation, and different choices give different answers. In the continuous case the entropy functional is not invariant under reparameterisation at all; the repair — relative entropy with respect to an invariant reference measure — requires that measure to be specified, which is the original problem in new clothes. The transformation-group method works only where a genuine symmetry group is available and is silent otherwise, and the choice of which invariances to demand can itself be contested. The cube factory and wine–water paradoxes are not solved so much as pushed up a level.

Is objectivity even desirable? Subjectivists argue that equivocation is not a rational requirement but an arbitrary preference for the middle. If evidence genuinely underdetermines the credence, holding an extreme-but-coherent credence is not a mistake; it is merely unusual. The debate here is the uniqueness thesis versus permissivism, which is a general question in epistemology about whether evidence ever permits more than one rational response.

It may not be an interpretation of probability. The sharpest structural objection. Objective Bayesianism tells us how credences ought to be constrained — it is a normative epistemology. It does not obviously tell us what probability is. Its calibration norm presupposes a notion of physical chance supplied by some other interpretation (propensity, best-system, or frequency), so as a theory of what probability means it is either incomplete or parasitic. Defenders reply that the position is a theory of epistemic probability specifically, and that expecting one theory to cover both sides of the chance/credence divide is the mistake.

Maximum entropy can conflict with conditionalization. The two updating rules are not always consistent: updating by conditionalization and re-maximising entropy on new constraints can give different results, as the Judy Benjamin problem shows. Since both are supposed to be requirements of rationality, this is a genuine tension, and it forces a choice about which rule governs which kind of learning.

Assessment

The honest summary is that objective Bayesianism scores well wherever the constraints are determinate and inherits the classical paradoxes wherever they are not. Its practical success — maximum-entropy methods work, Jeffreys priors are standard, and the framework underwrites a great deal of applied inference — is a point in its favour that no amount of philosophical difficulty about parameterisation quite erases.

Where this sits

Objective Bayesianism occupies the middle of the epistemic side: more constrained than subjectivism, less ambitious than the logical interpretation. It is best understood as the surviving fragment of Carnap's program — keeping the aim of evidence-driven, agent-independent credence while abandoning the claim that a unique confirmation function is derivable from logic alone.

Whether it constitutes a distinct interpretation or a normative supplement to subjectivism is a fair question, and the answer affects how the table in the overview should be read. Its calibration norm points at the one remaining position: an account of the objective chances that credence is supposed to track, consistent with a broadly Humean metaphysics. That is best-system chance, the last of the interpretations.