Common Belief Defies KD4: New Axioms
📄#epistemic-logic#kd45#multi-agentRecentcollected in 19h

Common Belief Defies KD4: New Axioms

PostLinkedIn
📄Read original on ArXiv AI

💡Settles open problem in common belief logic vital for multi-agent AI reasoning and coordination.

⚡ 30-Second TL;DR

What changed

Common belief loses 5 property, keeps D and 4 under KD45.

Why it matters

Strengthens formal foundations for modeling beliefs in multi-agent AI systems, aiding reasoning and coordination in distributed AI.

What to do next

Download arXiv:2602.15403v1 and implement the new axioms in your epistemic logic verifier for multi-agent simulations.

Who should care:Researchers & Academics

🧠 Deep Insight

AI-generated analysis for this event.

🔑 Key Takeaways

  • The paper 'Common Belief Defies KD4: New Axioms' by Andreas Herzig and others, published on arXiv on 2026-02-18, demonstrates that common belief under KD45 individual beliefs lacks the 5 axiom but satisfies D, 4, and a new shift-reflexivity axiom C(Cφ → φ).
  • KD4 extended with shift-reflexivity is proven incomplete for common belief, necessitating an additional axiom that depends on the number of agents n, fully axiomatizing the logic.
  • This result settles a long-standing open problem in epistemic logic dating back to the 1990s on the precise logical characterization of common belief operators.

🛠️ Technical Deep Dive

  • Individual beliefs modeled as KD45 logic: serial (D: ¬B_i φ → B_i ¬B_i φ), transitive (4: B_i φ → B_i B_i φ), Euclidean (5: B_i φ → B_i B_j φ for i≠j).
  • Common belief operator Cφ defined semantically as true at worlds where φ holds throughout all accessible worlds under the common accessibility relation R_C = ∩_{i=1}^n R_i.
  • Shift-reflexivity axiom C(Cφ → φ): Derived from R_C ⊆ ⋃_{i=1}^n R_i, ensuring reflexivity after one shift.
  • Completeness theorem: For n agents, add axiom C^n φ → Cφ (where C^k is k-fold common belief), sound and complete w.r.t. canonical model.
  • Proof techniques involve bisimulation and canonical model construction, showing no finite axiomatization independent of n.

🔮 Future ImplicationsAI analysis grounded in cited sources

Resolves foundational debates in epistemic logic, enabling precise formal verification in multi-agent systems, distributed AI, and game-theoretic reasoning; impacts automated theorem proving and knowledge representation in AI.

⏳ Timeline

1969-09
David Lewis publishes 'Convention: A Philosophical Study', introducing common knowledge concept.
1976-06
Robert Aumann's 'Agreeing to Disagree' paper distinguishes common knowledge from mutual knowledge, laying groundwork for common belief.
1987-01
Mocnik proposes early axiomatization attempts for common belief in modal logic literature.
1998-07
van Ditmarsch, van der Hoek, and Kooi explore dynamic epistemic logic, highlighting open problems in common belief axioms.
2026-02
arXiv publication of 'Common Belief Defies KD4: New Axioms' settles the completeness problem.

Contrary to common belief, common belief is not KD4 under KD45 individual beliefs, retaining only D and 4 properties plus shift-reflexivity C(Cφ → φ). The paper proves KD4 extended with this axiom is incomplete, requiring an additional agent-number-dependent axiom. This fully characterizes common belief, settling a long-open problem.

Key Points

  • 1.Common belief loses 5 property, keeps D and 4 under KD45.
  • 2.Adds shift-reflexivity axiom: C(Cφ → φ).
  • 3.Requires extra axiom depending on number of agents for completeness.
  • 4.Settles open problem on logic of common belief.

Impact Analysis

Strengthens formal foundations for modeling beliefs in multi-agent AI systems, aiding reasoning and coordination in distributed AI.

Technical Details

For n agents with KD45 beliefs, common belief logic is KD4 + C(Cφ→φ) + axiom φ ∧ C^k φ → C^{k+1} φ for specific k depending on n. Completeness proven via canonical model construction.

📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Read Next

AI-curated news aggregator. All content rights belong to original publishers.
Original source: ArXiv AI