Introduction
While the logic of how agents add new information—known as belief revision—has been studied extensively for decades, the logic of how they sequentially remove information, or iterated belief contraction, has been largely overlooked. This paper addresses this gap by proposing "Principled Factored Insertion," a new formal requirement for iterated contraction designed to ensure agents do not maintain "zombie beliefs" that have lost their logical justification.
The AGM Foundation and the Iteration Problem
In the field of Belief Change, the gold standard is the AGM framework, named after Alchourrón, Gärdenfors, and Makinson. AGM defines how a rational agent should modify their knowledge base ($K$) when faced with new information ($x$) through a set of eight core postulates.
However, the AGM framework was originally designed for "one-shot" changes. It provides little guidance on what happens when an agent undergoes a sequence of changes, such as contracting belief $x$ and then subsequently contracting belief $y$. Most research in iterated change has focused on revision, leaving contraction—the act of retracting a belief while remaining rational—neglected. Previous attempts to bridge this gap, such as Hans Rott’s "Qualified Insertion," often fail intuitive tests. For instance, if an agent believes both "Argentina will win" and "Brazil will win" on independent grounds, and is then forced to stop believing both, they should not logically continue to believe the disjunction "Either Argentina or Brazil will win." Existing models often fail to catch this nuance, allowing these unsupported "zombie beliefs" to persist.
Methods: Deriving Principles from Logic
The authors derive their approach from first principles by refining existing logical axioms through several steps:
- Critiquing Qualified Insertion: Using counter-examples like the FIFA World Cup scenario, the authors demonstrate that previous proposals are too permissive.
- Factoring: They utilize the "Factoring" property, which relates the contraction of a conjunction ($x \wedge y$) to the contraction of its individual components ($x$ and $y$).
- Principled Factored Insertion: This novel contribution breaks down sequential contraction into three "trigger conditions" based on the logical relationship between the first belief removed ($x$) and the second ($y$).
- Semantic Modeling: To validate these principles, the authors employ Lexicographic State Contraction. This involves representing an agent's mental state as a "plausibility ordering" of possible worlds. When a belief is contracted, the "best" worlds where that belief is false are promoted to the top of the ordering.
Key Insights and Results
The paper identifies a deep structural symmetry between adding and removing beliefs. Specifically, the authors identify Observations 3 and 4, which act as the contraction-based analogues to the famous Darwiche-Pearl (DP) postulates for iterated revision.
Key results include:
- Formalization of Principled Factored Insertion: A new principle governing rational behavior during sequential contractions.
- Validation via Lexicographic Contraction: Theorem 2 proves that any contraction operation generated via Lexicographic State Contraction automatically satisfies the new "Principled Factored Insertion" property.
- Refinement of Existing Models: The work demonstrates exactly where previous models fall short and how to "qualify" them to be more robust against the "Zombie Belief" problem.
Limitations and Future Work
The authors acknowledge that while "Principled Factored Insertion" is a necessary condition for rational iterated contraction, it is not yet a sufficient one. A system could satisfy this principle and still behave irrationally in other contexts. The principle is currently "too weak to completely characterise" Lexicographic Contraction. Future research aims to identify the additional postulates required to fully capture the logic of lexicographic state changes, potentially leading to a complete axiomatic characterization of iterated belief contraction.
Summary updated