[Paper] On the Semantics of Primary Cause in Hybrid Dynamic Domains

Source: arXiv - 2602.14994v1

Overview

The paper tackles a classic AI problem—identifying the actual cause of an observed effect—but does so in a setting where the world changes both discretely (e.g., pressing a button) and continuously (e.g., a robot arm moving). By extending the well‑known Situation Calculus to handle hybrid (discrete + continuous) dynamics, the authors propose two mathematically rigorous definitions of primary cause and prove they are equivalent. This bridges a gap between philosophical notions of causation and the needs of modern, cyber‑physical systems.

Key Contributions

• Hybrid Temporal Situation Calculus (HTSC): An extension of the classic Situation Calculus that natively represents both instantaneous actions and continuous processes.
• Two Primary‑Cause Definitions:
  1. A foundational definition based on counterfactual histories.
  2. A contribution‑based definition that quantifies how much a candidate cause “contributes” to an effect, enabling a modified “but‑for” test.
• Equivalence Proof: Formal theorem showing the two definitions coincide, giving developers flexibility in which formulation to use.
• Intuitive Property Verification: Demonstrates that the definitions satisfy desirable causality properties (e.g., minimality, relevance) in hybrid domains.

Methodology

1. Modeling Hybrid Worlds: The authors encode hybrid dynamics as situations (states after a sequence of actions) enriched with temporal fluents that evolve continuously according to differential equations.
2. Counterfactual Construction: For any candidate cause, they generate alternative histories where that cause is omitted or altered, then compare the resulting trajectories.
3. Contribution Metric: They define a numeric “contribution” that captures the difference in the effect’s value between the actual history and the counterfactual one.
4. Formal Proofs: Using the axioms of HTSC, they prove that the counterfactual‑based and contribution‑based definitions yield the same set of primary causes.

The approach stays within first‑order logic augmented with real‑valued functions, making it amenable to existing automated reasoning tools.

Results & Findings

• Equivalence Established: The two definitions are mathematically identical, meaning developers can pick the more intuitive “but‑for” style or the quantitative contribution style without losing correctness.
• Property Satisfaction: The definitions respect key intuitions such as causal minimality (no superfluous causes) and temporal relevance (only events that actually affect the effect’s timeline are considered).
• Illustrative Scenarios: The paper walks through classic hybrid examples (e.g., a thermostat controlling temperature, a robot navigating while its battery drains) and shows how the definitions correctly pinpoint primary causes.

Practical Implications

• Debugging Cyber‑Physical Systems: Engineers can automatically trace why a safety violation occurred (e.g., a drone crashed) by querying the HTSC model for primary causes, even when the failure involves intertwined discrete commands and continuous dynamics.
• Explainable AI for Robotics: The contribution‑based definition yields a numeric “impact score,” which can be presented to operators as a clear explanation (“the sudden brake command contributed 0.73 to the collision”).
• Policy Verification: Autonomous vehicle policies often involve continuous control laws plus discrete decision points. HTSC can be used to verify that a policy’s intended causes (e.g., lane‑change triggers) are indeed the primary drivers of observed outcomes.
• Tool Integration: Because the formalism stays within first‑order logic with real arithmetic, it can be plugged into existing theorem provers (e.g., Z3) or model checkers, enabling automated causality analysis pipelines.

Limitations & Future Work

• Scalability: The current proofs and examples are modest in size; applying HTSC to large‑scale systems may hit performance bottlenecks in reasoning engines.
• Learning the Model: The framework assumes a hand‑crafted hybrid action theory. Future research could explore automated extraction of HTSC models from sensor data or code.
• Probabilistic Extensions: Real‑world systems often involve stochastic noise. Extending the definitions to probabilistic hybrid domains is an open direction.
• User‑Friendly Tooling: The authors note the need for higher‑level APIs or DSLs to make HTSC accessible to developers without deep logical expertise.

---

Bottom line: By giving a solid, dual‑definition foundation for primary causation in hybrid dynamic worlds, this work equips AI developers, robotics engineers, and safety analysts with a rigorous yet practical lens for answering “what really caused that outcome?”—a question that’s becoming ever more critical as software increasingly controls physical processes.

Authors

• Shakil M. Khan
• Asim Mehmood
• Sandra Zilles

Paper Information

• arXiv ID: 2602.14994v1
• Categories: cs.AI
• Published: February 16, 2026
• PDF: Download PDF