AGM Belief Revision Theory

Updated 22 June 2026

AGM Belief Revision Theory is a formal framework that defines rational belief change through expansion, contraction, and revision operations based on minimal change principles.
The theory employs deductively closed belief sets and total preorders on models to represent syntactic and semantic foundations, ensuring consistency and minimal loss during updates.
Extensions of AGM span modal logics, non-classical frameworks, and multi-agent systems, highlighting its broad applications in knowledge representation and automated reasoning.

Alchourrón–Gärdenfors–Makinson (AGM) Belief Revision Theory provides a rigorous formal foundation for reasoning about rational belief change under new information. Originating in the context of classical propositional logic, the AGM paradigm is characterized by a system of postulates that define the allowable ways to expand, contract, or revise belief sets in response to evidence. Over the past decades, this theory has been extended, semantically characterized, and generalized to broader classes of logical systems, richer epistemic models, and new domains, while retaining the “minimal change” ideal as its organizing principle.

1. Syntactic and Semantic Foundations

The basic AGM framework formalizes belief states as deductively closed, consistent sets of formulas—typically in classical propositional or first-order logic. Formally, a belief set $K$ in language $\mathcal{L}$ with consequence relation $\mathit{Cn}$ satisfies $K = \mathit{Cn}(K)$ and $K \nvdash \bot$ (Voorbraak, 2013). The theory is equipped with three core operations:

Expansion $K+\varphi$ : Classical expansion by a new formula, defined as $K+\varphi = \mathit{Cn}(K \cup \{\varphi\})$ .
Contraction $K-\varphi$ : Removal of just enough to prevent entailment of $\varphi$ , with minimal loss otherwise.
Revision $K*\varphi$ : Incorporating potentially contradicting information $\mathcal{L}$ 0 by performing minimal contraction by $\mathcal{L}$ 1, then expanding by $\mathcal{L}$ 2 (Levi identity: $\mathcal{L}$ 3).

The postulates for these operations enforce closure, success, inclusion/minimality, and extensionality properties, guaranteeing rational and consistent management of information (Voorbraak, 2013, Falakh et al., 2021, Falakh et al., 2021).

Table: Key AGM Revision Postulates (for belief set $\mathcal{L}$ 4 and formula $\mathcal{L}$ 5)

Label	Schematic Statement	Intuitive Function
(*1)	$\mathcal{L}$ 6	Closure
(*2)	$\mathcal{L}$ 7	Success
(*3)	$\mathcal{L}$ 8	Inclusion/Minimality
(*4)	$\mathcal{L}$ 9	Vacuity
(*5)	$\mathit{Cn}$ 0	Extensionality

Semantically, AGM revision was connected to plausibility orderings on possible worlds: every belief set $\mathit{Cn}$ 1 induces a total preorder $\mathit{Cn}$ 2 on models (interpretations), and revision by $\mathit{Cn}$ 3 selects the minimal $\mathit{Cn}$ 4-worlds: $\mathit{Cn}$ 5\varphi $\mathit{Cn}$ 6 (Falakh et al., 2021, Falakh et al., 2021, Bonanno, 2023).

2. Semantic Representation and Generalizations

The classic finite propositional case (Katsuno–Mendelzon) guarantees that every AGM operator is determined by a faithful assignment of total preorders to belief states (Falakh et al., 2021, Falakh et al., 2021). More recent generalizations extend this paradigm:

Tarskian and arbitrary monotonic logics: AGM-revision can be represented using assignments from belief bases to total (but not necessarily transitive) relations on models, dubbed min-friendly assignments (Falakh et al., 2021, Falakh et al., 2021, Aiguier et al., 2015).
Min-friendly relations: These are total, min-complete (guaranteeing minima exist), and min-retractive (ensuring minimality propagates downwards). They coincide with preorders (transitive total relations) when certain loop-freeness or disjunction properties hold in the base logic (Falakh et al., 2021).
Critical loops: In non-classical logics (e.g., Horn logic), cycles may exist that prevent transitive (preorder) representations; in such cases, AGM revision can only be modeled by more general min-friendly but non-transitive assignments (Falakh et al., 2021, Falakh et al., 2021).

A key result: The AGM postulates (in generalized base logic) are equivalent to the existence of a min-friendly faithful assignment $\mathit{Cn}$ 7 such that, for all bases $\mathit{Cn}$ 8, the revised model set is minimal: $\mathit{Cn}$ 9. (Falakh et al., 2021, Falakh et al., 2021)

Modal logics, notably those with belief ( $K = \mathit{Cn}(K)$ 0), necessity ( $K = \mathit{Cn}(K)$ 1), and Ramsey-style conditionals ( $K = \mathit{Cn}(K)$ 2), have been shown to admit fine-grained axiomatizations of the AGM postulates (Bonanno, 20 Feb 2025, Bonanno, 26 Feb 2026, Bonanno, 2023). In this setting:

Agents are modeled with Kripke frames $K = \mathit{Cn}(K)$ 3, where $K = \mathit{Cn}(K)$ 4 is a belief accessibility and $K = \mathit{Cn}(K)$ 5 is a Lewis selection function for conditionals.
Revision by input $K = \mathit{Cn}(K)$ 6 yields the revised belief set $K = \mathit{Cn}(K)$ 7 as those $K = \mathit{Cn}(K)$ 8 such that all $K = \mathit{Cn}(K)$ 9-closest worlds (from all $K \nvdash \bot$ 0-accessible states) satisfy $K \nvdash \bot$ 1.
Each AGM postulate corresponds to a modal axiom in this language, and the class of Kripke–Lewis frames for AGM revision is axiomatized by frame conditions mirroring the postulates (Bonanno, 2023, Bonanno, 20 Feb 2025).

Furthermore, this approach reveals that AGM revision can be seen as a special case of the more general KM belief update, with distinction determined by two strengthened frame conditions and a unique axiom regarding “unsurprising information” (revision when $K \nvdash \bot$ 2 was not disbelieved) (Bonanno, 26 Feb 2026).

4. Extensions and Variants

The AGM paradigm has been robustly extended across several axes:

Belief bases and prioritized revision: Dropping deductive closure, base revision is characterized by operators on finite or arbitrary sentence sets, generalized to non-classical logics and dynamic belief bases, with corresponding representation theorems (Falakh et al., 2021, Souza et al., 2019).
Non-classical and paraconsistent logics: In paraconsistent (e.g., PAC) and multi-valued settings, modified AGM postulates are adapted to handle inconsistencies and relativized priorities, with preservation of closure, minimality, and a weakened recovery (Ebrahimi, 2017, Borges et al., 2019).
Probabilistic belief states: Probabilistic analogues distinguish between expansion by constraining (removing non-supporting distributions) vs. conditioning (likelihood reweighting); the former aligns more closely with AGM expansion (Voorbraak, 2013).
Dynamic/iterated revision: Modal languages such as Dynamic Preference Logic and belief algebra frameworks internalize AGM/DP postulates for complex, iterated revision, with unique operators ensured by added constraints (Souza et al., 2021, Meng et al., 10 May 2025).
Multi-agent and source-sensitive settings: Extensions to multi-agent epistemic models (Kripke models for sets of agents, epistemic planning) generalize the AGM postulates to multi-agent belief dynamics and manage source reliability via epistemic-reliability functions (Thielscher et al., 4 May 2026, Ebrahimi, 2017).
Intentions, goals, and temporal dimensions: AGM-style postulates support joint revision of temporal beliefs and intention databases, extending the paradigm to planning and dynamic intention management (Zee et al., 2016).
Filtered and partial (non-universal success) revision: Handling non-credible or only "allowable" information via filtered belief revision replaces or weakens the success axiom, reflecting resistance to unreliable sources (Bonanno, 2019).

5. Methodological and Practical Implications

Advances in AGM theory provide a foundation for belief change in:

Knowledge representation: Uniform model-theoretic semantics for revision in description logics, hybrid non-classical logics, and ontology evolution (Aiguier et al., 2015).
Database and preference revision: Generalizations to database belief management (separating strong/weak beliefs, intention coherence) and preference revision (ordering of direct comparisons) situate AGM theory as a unifying pattern for rational update across information systems (Haret et al., 2021, Zee et al., 2016).
Automated reasoning and AI planning: Iterated revision, belief base organization, and multi-agent extensions facilitate applications in dynamic reasoning, agent systems, and planning environments (Souza et al., 2019, Thielscher et al., 4 May 2026).

6. Conceptual Boundaries, Open Problems, and Directions

Current research identifies sharp boundaries for the AGM framework:

Critical loops and non-preorder representability: Failure of loop-freeness prohibits preorder-based representations, especially in logics lacking expressive disjunction (Falakh et al., 2021, Falakh et al., 2021).
Expressive completeness and iterated revision: Despite robust single-shot characterizations, complete modal axiom systems for iterated AGM/DP change, especially for more expressive languages, remain open (Souza et al., 2021, Bonanno, 20 Feb 2025).
Partial and imprecise belief structures: Accept-desirability models and similar frameworks generalize AGM, but introduce phenomena such as dilation that disrupt certain postulates—prompting ongoing investigation into robust AGM-like frameworks for imprecise probabilities and quantum logics (Coussement et al., 22 Dec 2025).
Computation and implementation: The complexity of constructing min-friendly assignments or minimal-model selection in description or modal logics is an area of active research (Falakh et al., 2021, Aiguier et al., 2015).

Recent directions include algorithmic realization of unique revision operators in belief algebras, dynamic modeling of latent beliefs, and the development of iterated, source-sensitive, and preference-guided belief update protocols for practical systems (Meng et al., 10 May 2025, Arisaka, 2015, Haret et al., 2021).

References

(Voorbraak, 2013) Probabilistic Belief Change: Expansion, Conditioning and Constraining
(Falakh et al., 2021) A General Katsuno-Mendelzon-Style Characterization of AGM Belief Base Revision for Arbitrary Monotonic Logics
(Falakh et al., 2021) AGM Belief Revision, Semantically
(Bonanno, 2023) A Kripke-Lewis semantics for belief update and belief revision
(Bonanno, 20 Feb 2025) A modal logic translation of the AGM axioms for belief revision
(Ebrahimi, 2017) Source-Sensitive Belief Change
(Coussement et al., 22 Dec 2025) Conditioning Accept-Desirability models in the context of AGM-like belief change
(Souza et al., 2021) Dynamic Preference Logic meets Iterated Belief Change: Representation Results and Postulates Characterization
(Souza et al., 2019) Iterated Belief Base Revision: A Dynamic Epistemic Logic Approach
(Bonanno, 26 Feb 2026) The logic of KM belief update is contained in the logic of AGM belief revision
(Aiguier et al., 2015) Belief Revision, Minimal Change and Relaxation: A General Framework based on Satisfaction Systems, and Applications to Description Logics
(Zee et al., 2016) AGM-Style Revision of Beliefs and Intentions from a Database Perspective (Preliminary Version)
(Arisaka, 2015) How do you revise your belief set with %$;@*?
(Borges et al., 2019) Belief revision and 3-valued logics: Characterization of 19,683 belief change operators
(Haret et al., 2021) An AGM Approach to Revising Preferences
(Thielscher et al., 4 May 2026) A Study of Belief Revision Postulates in Multi-Agent Systems (Extended Version)
(Bonanno, 2019) Credible Information, Allowable Information and Belief Revision -- Extended Abstract