Documentation

Linglib.Phenomena.DefaultReasoning.Studies.DarwichePearl1997

Darwiche & Pearl (1997): On the Logic of Iterated Belief Revision #

@cite{darwiche-pearl-1997}

AGM belief revision constrains how a single revision should update an agent's beliefs, but says nothing about how the agent's disposition to revise should change. @cite{darwiche-pearl-1997} show that AGM- compatible revision operators can behave pathologically under iteration, and propose four additional postulates C1–C4 that rule out these pathologies.

Representation Theorem (Theorem 13) #

C1–C4 are equivalent to conditions CR1–CR4 on the faithful assignment (total pre-order) that represents an epistemic state:

Counterexamples (Tables A.1–A.4) #

For each CR_i, there exists an AGM-compatible revision operator that violates CR_i while satisfying the other three. This shows the four conditions are logically independent — none is derivable from the rest.

Bridge to Ranking Functions #

Core.Logic.Ranking.ranking_satisfies_C1..C4 proves that Spohn's A,α-conditionalization satisfies all four postulates. The counterexamples here use non-ranking revision operators — arbitrary total pre-order transformations that respect AGM success but violate the D&P constraints.

Linguistic Connection #

Ranking functions → PlausibilityOrderNormalityOrder → Kratzer's ordering sources for modals/conditionals. The D&P postulates constrain how modal bases evolve under discourse update — without them, an agent's conditional beliefs can shift arbitrarily between utterances. Dynamic semantics (DRT/DPL context update) is iterated revision: each sentence revises the common ground, and C1–C4 ensure that the ordering of live possibilities evolves rationally.

Four-element world type for counterexamples.

Instances For
    Equations
    • One or more equations did not get rendered due to their size.
    Instances For
      Equations
      • One or more equations did not get rendered due to their size.
      @[reducible]

      Convert a ranking function to its induced normality ordering. le w v ↔ κ(w) ≤ κ(v). Defined directly (not via toPlausibilityOrder) so that le reduces for native_decide.

      Equations
      • One or more equations did not get rendered due to their size.
      Instances For
        @[reducible]

        AGM success: all rank-0 worlds in the posterior satisfy μ.

        Equations
        Instances For

          Prior: w1=0, w2=1, w3=2, w4=3. Revise by μ = {w2,w3,w4}. Posterior: w1=1, w2=0, w3=2, w4=1. Within μ, the ordering of w3 vs w4 flips (2<3 → 2>1).

          Equations
          • One or more equations did not get rendered due to their size.
          Instances For
            Equations
            • One or more equations did not get rendered due to their size.
            Instances For

              Prior: w1=0, w2=1, w3=1, w4=2. Revise by μ = {w3,w4}. Posterior: w1=2, w2=1, w3=0, w4=1. Within ¬μ = {w1,w2}, the ordering flips (0<1 → 2>1).

              Equations
              • One or more equations did not get rendered due to their size.
              Instances For
                Equations
                • One or more equations did not get rendered due to their size.
                Instances For

                  Prior: w1=0, w2=1, w3=2, w4=3. Revise by μ = {w1,w2}. Posterior: w1=0, w2=2, w3=2, w4=3. w2 ∈ μ, w3 ∈ ¬μ: prior 1 < 2 (strict), posterior 2 ≤ 2 (not strict).

                  Equations
                  • One or more equations did not get rendered due to their size.
                  Instances For
                    Equations
                    • One or more equations did not get rendered due to their size.
                    Instances For

                      Prior: w1=0, w2=1, w3=1, w4=2. Revise by μ = {w1,w2}. Posterior: w1=0, w2=2, w3=1, w4=3. w2 ∈ μ, w3 ∈ ¬μ: prior 1 ≤ 1, but posterior 2 > 1.

                      Equations
                      • One or more equations did not get rendered due to their size.
                      Instances For
                        Equations
                        • One or more equations did not get rendered due to their size.
                        Instances For

                          The four conditions are logically independent: for each CR_i, there exists an AGM-compatible revision that violates CR_i alone.

                          This is the content of @cite{darwiche-pearl-1997}, Appendix A.

                          @cite{darwiche-pearl-1997}, Theorem 17: Spohn's ranking conditioning satisfies C1–C4 (equivalently CR1–CR4). The proofs are in Core.Logic.Ranking:

                          - `ranking_satisfies_C1` — C1 holds for `conditionα` with any α, β
                          - `ranking_satisfies_C2` — C2 holds for `conditionα` with any α, β
                          - `ranking_satisfies_C3` — C3 holds for canonical `revise`
                          - `ranking_satisfies_C4` — C4 holds for canonical `revise`
                          
                          Together with the counterexamples above, this shows that ranking
                          conditioning is the *tightest* well-behaved revision operator:
                          it satisfies all four independence conditions that AGM alone
                          leaves unconstrained.
                          
                          The chain to linguistics:
                          ```
                          RankingFunction.revise
                            → satisfies C1–C4 (this file's counterexamples show AGM alone doesn't)
                            → PlausibilityOrder (via toPlausibilityOrder)
                            → NormalityOrder (via toNormalityOrder)
                            → Kratzer ordering sources (via fromProps)
                            → modal/conditional semantics
                          ```