When is an example a counterexample?

Eric Pacuit, Arthur Paul Pedersen, Jan-Willem RomeijnTARK 2013

Abstract

The AGM theory of belief revision is a cornerstone of formal epistemology, yet it faces persistent criticism regarding its application to iterated revision. Prominent "counterexamples" suggest that the theory's postulates lead to unintuitive outcomes in simple scenarios. This paper argues that these conflicts often arise from a failure to explicitly model meta-information, such as source reliability and independence. By integrating nonstandard Bayesian models with qualitative logic, the authors demonstrate that AGM theory can be reconciled with intuitive reasoning.

Introduction: The Limits of AGM Theory

The AGM theory (Alchourrón, Gärdenfors, and Makinson) provides a rigorous framework for how a rational agent should update their beliefs when encountering new information. While highly successful for single revisions, the theory becomes controversial when applied to sequences of changes—known as iterated belief revision.

Logicians like Robert Stalnaker have proposed puzzles that seemingly invalidate core postulates of iterated revision, such as I1 and I2. These postulates govern how an agent's conditional beliefs should persist or change. The central question of this research is methodological: are these "counterexamples" genuine flaws in the logic, or are they simply misapplications where the model fails to capture the full context of the scenario?

Methods: Modeling Meta-Information

To address this, the authors re-examine Stalnaker’s coin-toss puzzle using a multi-layered modeling approach. The puzzle involves two coins in opaque boxes and reports from various witnesses whose reliability and independence are key factors.

Heuristic Plausibility and Iterated Revision

The authors first use "Grove’s spheres"—plausibility orderings of possible worlds—to visualize how different types of updates (like "radical upgrades") affect an agent's belief state. The following figure illustrates the transition of an agent's mental model through a sequence of reports in the coin puzzle.

Iterated belief revision process in Stalnaker's coin puzzle, showing transitions between models M0 through M3.

Bayesian Modeling and Nonstandard Probability

To bridge the gap between qualitative logical jumps and quantitative probability, the authors employ nonstandard analysis. By assigning infinitesimal probabilities (represented by hyperreal numbers) to certain events, they create a "regular" probability measure where every world has some weight, however small. This allows Bayesian conditioning to mimic the behavior of AGM revision operators precisely.

Bayesian odds and evidence probabilities across time steps t=0 to t=3 using nonstandard analysis to represent infinitesimal weights.

Key Insights: The Meta-Information Gap

The primary finding is that intuition often relies on "meta-information" that is omitted from basic logical models. For instance, the perceived reliability of a witness or the independence of two reports significantly influences how a rational agent should update their beliefs. If the model is too "naive," the theory appears to fail.

The authors introduce the CAR (Coarsening At Random) condition from statistics as a diagnostic tool. If a model's failure stems from ignoring how information is simplified or "coarsened" before reaching the agent, it is a misapplication of the theory rather than a counterexample to the logic itself.

Key Results and Conclusion

The authors successfully provide a model for Stalnaker’s puzzle that complies with the AGM postulates while still yielding the intuitively "correct" answer. Furthermore, they prove that for any finite algebra, a belief revision operator satisfies the AGM postulates if and only if it can be represented by a regular nonstandard probability measure.

This mathematical equivalence suggests that the AGM theory remains a robust foundation for rational belief change, provided that the models are sufficiently detailed to include all relevant meta-information. Future work aims to further formalize the CAR litmus test to systematically distinguish between theoretical flaws and modeling errors.

Summary updated

BibTeX
@inproceedings{pacuit2013when,
  author = {Eric Pacuit and Arthur Paul Pedersen and Jan-Willem Romeijn},
  title = {{When is an example a counterexample?}},
  booktitle = {TARK},
  year = {2013},
}
Figure