Standard Bayesian conditionalization assumes an idealization rarely met in practice: that evidence arrives with certainty. When an agent observes proposition E, she sets P(E) = 1 and updates all other beliefs via P'(H) = P(H|E). But observation is seldom so crisp. A dim glance at a shirt in candlelight shifts one's confidence that it is blue without settling the matter.

Richard Jeffrey's probability kinematics, introduced in The Logic of Decision (1965), generalizes conditionalization to accommodate this reality. Rather than requiring some proposition to become certain, Jeffrey's rule permits updating when experience directly shifts probabilities across a partition of possibilities, without any cell attaining probability one.

The formal machinery is elegant, but its philosophical stakes are substantial. Jeffrey's framework forces us to confront what evidence really is, whether belief change admits a determinate calculus when inputs are themselves uncertain, and whether the order in which we process ambiguous observations should influence our final credences. These questions matter for any formal theory of rational agency, from AI systems reasoning under sensor noise to human agents integrating fallible perceptual reports.

Let {E₁, E₂, ..., E_n} be a partition of the sample space, and suppose an experience directly modifies the agent's probabilities over this partition from prior values P(E_i) to posterior values P'(E_i), where the P'(E_i) sum to unity but none need equal one. Jeffrey's rule prescribes updating any hypothesis H via: P'(H) = Σ_i P(H|E_i) · P'(E_i).

This is a weighted average of the conditional probabilities P(H|E_i), with weights supplied by the new marginal probabilities across the partition. When some E_k receives probability one, the formula collapses to standard conditionalization on E_k, recovering the classical Bayesian case as a limit.

The rule respects a natural desideratum: it reallocates probability mass across the partition according to the new distribution while preserving how each cell would distribute credence over the remainder of the algebra. The shape of the conditional beliefs is retained; only the mixture weights change.

Consider the candlelit shirt. Let E₁ = blue, E₂ = green, E₃ = black. A prior of (0.4, 0.4, 0.2) might shift under observation to (0.7, 0.25, 0.05). The agent then updates her beliefs about downstream propositions—whether the shirt was recently purchased, whom it belongs to—by mixing conditional probabilities according to these new weights.

Notably, Jeffrey resists identifying the input to updating with any proposition in the agent's language. The experience itself is not represented as a learned sentence but as a direct redistribution over the partition. This propositional deficit is philosophically consequential: it decouples belief revision from linguistic evidence.

Takeaway

Bayesian updating need not require certainty as input. When experience merely reshuffles probability across a partition, Jeffrey's rule provides a principled extension that preserves the structural role of conditional beliefs.

Jeffrey's rule is not merely a computational convenience—it presupposes a substantive assumption known as rigidity: for each cell E_i of the partition, the conditional probability P(H|E_i) is preserved across the update, so that P'(H|E_i) = P(H|E_i) for all H.

Rigidity encodes the claim that the experience bears on H only through its bearing on the partition. Once one conditions on which E_i obtains, the experience adds nothing. This is the probabilistic analog of screening off, familiar from causal graphical models: the partition acts as a sufficient statistic for the evidential impact of the observation.

The philosophical justification is subtle. Rigidity fails when the experience carries information about H that is not mediated by the partition. If a partial glimpse of the shirt also reveals information about its fabric texture—information relevant to H = recently laundered—then P'(H|blue) may legitimately differ from P(H|blue). The partition was too coarse to capture the evidential structure.

This suggests a methodological principle: the choice of partition is itself an epistemic act. A well-chosen partition renders rigidity plausible; a poorly chosen one violates it. Formal epistemology thus inherits the problem of specifying evidential granularity, a matter Diaconis and Zabell (1982) analyzed by characterizing when a probability shift admits a Jeffrey-style representation.

Field's variant, which parameterizes updates by input likelihood ratios rather than posterior marginals, offers an alternative that some argue better captures the intuitive notion of evidential force. Yet Field's approach and Jeffrey's coincide precisely when rigidity holds, revealing rigidity as the pivotal assumption around which the entire framework turns.

Takeaway

Rigidity is the hidden hinge of uncertain updating: it demands that our partition of possibilities be fine enough to fully mediate the evidential impact of experience. Choosing this partition is itself a philosophical commitment.

Standard conditionalization is commutative: updating on E₁ then E₂ yields the same posterior as updating on E₂ then E₁, since both amount to conditioning on their conjunction. Jeffrey conditionalization, notoriously, lacks this property in general.

The failure was crisply demonstrated by Levi and later analyzed by Field and by Wagner. If two successive experiences directly shift probabilities over different (or even the same) partition to distinct marginal values, the resulting posterior generally depends on the order of updates. Reversing the sequence produces a different final credence over H.

This is not a computational quirk but a philosophical embarrassment. If rational belief revision is a function of prior credences and evidence received, and if evidence has no intrinsic temporal ordering beyond arrival, then order-dependence suggests either that the framework is incomplete or that something about the evidence itself must include its history.

Wagner's diagnosis is instructive: commutativity fails because Jeffrey inputs specify posterior probabilities, which are themselves prior-dependent, rather than specifying the evidential force of the experience in a prior-independent manner. Reformulations using Bayes factors—likelihood ratios that encode the intrinsic diagnostic weight of an observation—restore commutativity.

This resolution comes at a cost: it requires representing evidence not by what one's beliefs became but by how the world's data compares across hypotheses. It aligns formal epistemology with likelihoodist and information-theoretic traditions, where evidential strength is measured in bits or log-likelihood ratios—quantities invariant under order of receipt.

Takeaway

When evidence is specified by its output on our beliefs rather than by its intrinsic diagnostic weight, the sequence of experience contaminates the destination. True evidence should be a vector, not a landing point.

Probability kinematics extends Bayesian rationality into the domain where actual epistemic agents dwell: a world of uncertain perception, ambiguous testimony, and noisy signals. Jeffrey's rule preserves the mathematical spirit of conditionalization while relaxing its idealized precondition of certain evidence.

Yet the framework's difficulties—the specification of appropriate partitions, the substantive commitment of rigidity, the failure of commutativity—are not defects to be regretted but discoveries. They reveal that uncertain evidence is a more structured notion than it first appears, demanding formal representations that separate diagnostic force from posterior effect.

The lesson generalizes. Every formal theory of rational belief change must grapple with what evidence is, not merely with how it operates. Jeffrey's kinematics gives us both a working tool and a diagnostic: it shows where the concept of evidence itself requires further refinement.