Belief Revision¶

Incorporate newly accepted information into a logically represented belief state while restoring consistency and preserving as much of the prior commitment structure as a declared minimal-change policy warrants.

Version: v2 · 2026-09-06 · History
Domain-specific #: 1366
Origin domain: computer science
Subdomain: knowledge representation
Aliases: AGM belief revision, Belief change by revision

Core Idea¶

Belief Revision is the formal knowledge-representation operation that incorporates a newly accepted proposition into an existing belief set or epistemic state while resolving contradiction and preserving prior commitments according to a rational minimal-change policy. If a deductively closed belief set K is revised by proposition phi, the result is written K * phi. Success normally requires the result to include phi; consistency requires a consistent result when phi is consistent; preservation principles constrain which old consequences are lost. Alchourrón, Gärdenfors, and Makinson's 1985 paper establishes the canonical AGM theory through revision and partial-meet contraction functions.^[1]

The problem is nontrivial because inconsistency under classical consequence is explosive and several repairs may restore consistency. Suppose K entails C from commitments A, B, and A and B implies C, while new accepted information is not C. Revision cannot simply append not C; it must relinquish at least one supporting commitment. A selection or plausibility ordering decides which coherent alternatives remain closest to the prior state. Katsuno and Mendelzon give a model-theoretic characterization in which AGM-style revision selects minimal models of the new information under a faithful preorder associated with the old theory.^[2]

Revision differs from expansion, contraction, and update. Expansion adds information without repairing inconsistency. Contraction gives up a proposition without accepting its negation. Revision accepts a proposition that may conflict with the old state. Knowledge-base update, in the Katsuno–Mendelzon distinction, represents a changing world: old beliefs describe an earlier state and new information describes the present, so each old possible world can evolve. Revision instead treats old and new claims as concerning the same underlying situation and changes which prior description is credible.^[3] This domain-specific update must not be confused with Bayesian Updating in the encyclopedia.

Belief Revision is deliberately distinct from Bayesian Updating. Bayesian updating transforms graded probabilities by conditioning or likelihood combination under a fixed event algebra and probabilistic model. AGM revision operates canonically on logically closed sets or related qualitative epistemic states, prioritizes consistency and minimal loss, and handles newly accepted propositions that may contradict prior commitments. Probabilistic belief change and qualitative revision can be related through rankings or conditionalization, but neither is an alias of the other. A Bayesian posterior can redistribute degrees without retracting any proposition held at full acceptance; an AGM operator can change categorical membership without assigning numerical credences.

The candidate survives exact-surface review because the accepted Bayesian Updating node has the one-liner Update beliefs with evidence, a broad surface that could attract the phrase belief revision, but its structural core is probabilistic prior–likelihood–posterior transformation. The reopened node retains AGM belief-set or epistemic-state change, success, consistency, closure, extensionality, minimal change, contraction and revision relations, and the revision-versus-world-update distinction. State and State Transition is the strict parent: the operation maps one represented epistemic condition to another under an input and rationality policy. An exact-surface registry boundary is queued; no registry mutation is applied.

Structural Signature¶

The prior epistemic state. A belief set, base, theory, plausibility ordering, or related representation records current commitments.
The new accepted information. A proposition or information item is taken as input with a declared reliability status.
The conflict test. Logical consequence reveals whether simple expansion would violate consistency.
The revision operator. A rule maps the prior state and input to a successor state.
The success condition. The new proposition belongs to or is entailed by the revised state under standard formulations.
The consistency condition. A consistent input should not force an inconsistent successor under the intended logic.
The minimal-change policy. Entrenchment, remainder selection, distance, or plausibility order determines retained commitments.
The closure and equivalence policy. Consequence and logically equivalent inputs are treated under declared postulates.
The operator family. Expansion, contraction, revision, and dynamic-world update remain distinct.
The audit explanation. Lost beliefs and selected alternatives can be traced to the policy rather than hidden in a posterior number.

What It Is Not¶

Not Bayesian Updating. AGM revision is qualitative logical change, not necessarily numerical conditioning.
Not simple expansion. Contradiction can require relinquishing old commitments.
Not contraction. Contraction removes commitment without necessarily accepting a contrary proposition.
Not knowledge-base update in the technical KM sense. Update models a changed world rather than corrected information about one situation.
Not arbitrary database editing. Rationality and minimal-change constraints govern the successor state.
Not psychological attitude change by default. The formal operator need not model every human cognitive mechanism.
Not Revisionism. Historical reinterpretation is an unrelated prime surface.

Scope of Application¶

Belief Revision is literal when a represented epistemic state incorporates accepted information, repairs conflict, and selects retained commitments through an explicit qualitative rationality or minimal-change policy.

Knowledge representation. Maintaining logically coherent agent theories after new inputs.
Artificial intelligence. Designing rational qualitative belief-change operators.
Database integrity. Distinguishing correction of beliefs from update of a changing world.
Nonmonotonic reasoning. Retracting consequences when premises lose standing.
Epistemic logic. Representing contraction, revision, and iterated change.
Ontology maintenance. Repairing inconsistent axiom sets under priorities.
Theory change. Comparing entrenchment and remainder-selection policies.
Hybrid uncertainty models. Relating qualitative rankings to probabilistic representations without collapse.

Clarity¶

A clear revision problem states the base logic, whether the state is a deductively closed belief set or finite base, the new formula, input reliability, operator notation, rationality postulates, minimal-change representation, and treatment of inconsistency. It distinguishes same-world correction from changed-world update and qualitative acceptance from probability revision. An implementation gives the ordering, distance, entrenchment, or selection function rather than appealing to similarity informally. Iterated revision requires an epistemic state richer than a bare consequence set if future priorities matter. The surface update is qualified because database, dynamic-world, and Bayesian uses are structurally different.

Manages Complexity¶

Belief revision compresses a combinatorial repair problem into an operator constrained by postulates and a preference representation. Instead of enumerating every arbitrary deletion, it preserves theories closest to the prior state while enforcing success and consistency. The AGM framework also exposes impossibility and underdetermination: postulates characterize rational families but do not uniquely choose one policy. Complexity returns through large model spaces, inconsistent bases, nonclassical logics, iterated inputs, source reliability, and relevance. The abstraction manages this by separating logical requirements from the extra ordering that carries domain priorities.

Abstract Reasoning¶

Represent the current belief set or richer epistemic state in a declared logic.
Receive the new proposition and determine whether it is to be accepted categorically.
Test whether simple expansion preserves consistency and other integrity conditions.
If conflict occurs, generate or characterize the maximal coherent ways to relinquish old commitments.
Apply an entrenchment, remainder-selection, distance, or plausibility policy.
Construct the successor state that includes the input and preserves preferred prior content.
Close the result under the declared consequence relation if the representation requires closure.
Check success, consistency, extensionality, inclusion, and relevant recovery-style conditions.
Distinguish revision from dynamic-world update and from graded probabilistic change.
Record which commitments were lost and why, then prepare the state for possible iteration.

Knowledge Transfer¶

Belief Revision transfers a general constrained-state-repair pattern: accept a binding new condition, locate conflict, preserve a maximal or preferred part of the prior state, and expose the policy that decides among repairs. The pattern applies to configuration management, ontology repair, legal-rule maintenance, and requirements change. The logical accent remains important: equivalence, consequence closure, categorical acceptance, and postulates differ from numerical Bayesian learning or ordinary document versioning. Minimal change is not self-defining; every transferred use must declare its distance or priority structure.

Examples¶

Canonical¶

Let a belief set contain A, B, and if A and B then C. New trusted information is not C. Expansion would be inconsistent. A revision policy might preserve A and the conditional while dropping B, or preserve A and B while dropping the conditional; AGM postulates alone do not select one without an entrenchment or remainder-selection policy. The revised set contains not C, remains consistent, and explains its retained commitments through the selected ordering.^[1]

Mapped back: prior logical theory + contradictory accepted proposition → alternative coherent remainders → minimal-change selection → consistent successor theory.

Applied / In Practice¶

A system believes that either satellite A or B is still in orbit and receives trusted information that A is not. If the information corrects beliefs about one unchanged situation, revision can infer that B remains in orbit. If time has passed and the satellites may have landed, dynamic-world update need not preserve that conclusion. Katsuno and Mendelzon use such contrasts to show that revision and update can produce different successor models even from similar formulas.^[3]

Mapped back: same formulas + different temporal interpretation → revision or world update operator → different minimal-change semantics → different warranted conclusion.

Structural Tensions¶

Success vs. prior retention. Accepting new information can require abandoning established consequences. Diagnostic: Which commitment is least entrenched and why?
Consistency vs. relevance. A repair can be consistent while deleting unrelated beliefs. Diagnostic: Does the policy preserve independent language fragments?
Postulates vs. unique procedure. Rational constraints allow many operators. Diagnostic: What extra ordering chooses among remainders?
Same-world revision vs. changed-world update. Formula pairs underdetermine temporal interpretation. Diagnostic: Do old and new claims describe one state or successive states?
Qualitative acceptance vs. graded credence. Sets and probabilities encode different commitments. Diagnostic: Is a Bayesian posterior being mistaken for AGM membership?
Single revision vs. iteration. A bare result may lose priority information needed later. Diagnostic: What epistemic-state structure persists across rounds?
Logical equivalence vs. syntactic provenance. Equivalent formulas can have different source histories. Diagnostic: Does the operator revise theories, bases, or source-labeled items?

Structural–Framed Character¶

The structure is prior epistemic state, accepted proposition, conflict, revision operator, success, consistency, minimal change, closure, and successor state. The frame is the underlying logic, set-versus-base representation, entrenchment or distance, source priorities, temporal interpretation, and iteration policy. Changing from propositions to an ontology can preserve the structure; replacing categorical logical change with prior–likelihood multiplication changes the abstraction.

Structural Core vs. Domain Accent¶

The transferable core is state + binding new condition + conflict → priority-governed repair → minimally changed valid successor. The domain accent is belief sets, logical consequence, AGM postulates, contraction and revision functions, faithful preorders, knowledge bases, and the revision/update distinction. Remove the accent and State and State Transition remains; retain it and Belief Revision is autonomous.

State and State Transition is the strict parent by specialization. A revision operator maps a represented epistemic state and accepted input to a successor state under consistency and minimal-change constraints. The parent is broader and does not prescribe beliefs, logic, or AGM rationality. Bayesian Updating is a neighboring probabilistic specialization, not the parent and not an alias.

The prospective workspace queue contains one strict upward edge to prime:state_and_state_transition. No live DAG mutation is authorized.

Relationships to Other Abstractions¶

Current abstraction Belief Revision Domain-specific

Parents (1) — more general patterns this builds on

Belief Revision is a kind of State and State Transition Prime

State and State Transition is the strict parent by specialization.

Hierarchy path (1) — routes to 1 parentless root

Belief Revision → State and State Transition → Phase Space

Neighborhood in Abstraction Space¶

Belief Revision sits in a sparse region of the domain-specific corpus (91^st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Bayesian Updating. Graded probabilistic transformation from prior and likelihood to posterior.
Expansion. Addition without conflict repair.
Contraction. Giving up a proposition without accepting its negation.
Belief Update. In KM terminology, change representing evolution of the world.
Belief Formation. Transition from noncommitment to holding a proposition, not the full consistency-repair theory.
Minimal Modification Principle. A broader accepted prime concerning preservation under alternative-scenario construction.
Revisionism. Historical reinterpretation rather than formal epistemic change.

References¶

[1] Carlos E. Alchourrón, Peter Gärdenfors, and David Makinson, On the Logic of Theory Change: Partial Meet Contraction and Revision Functions, Journal of Symbolic Logic 50, no. 2 (1985): 510–530, https://doi.org/10.2307/2274239. registry ↩a ↩b

[2] Hirofumi Katsuno and Alberto O. Mendelzon, Propositional Knowledge Base Revision and Minimal Change, Artificial Intelligence 52, no. 3 (1991): 263–294, https://doi.org/10.1016/0004-3702(91)90069-V. registry ↩

[3] Hirofumi Katsuno and Alberto O. Mendelzon, On the Difference Between Updating a Knowledge Base and Revising It, in Peter Gärdenfors, ed., Belief Revision, Cambridge Tracts in Theoretical Computer Science 29 (Cambridge University Press, 1992): 183–203, https://doi.org/10.1017/CBO9780511526664.007. registry ↩a ↩b