Causal reasoning has historically been fragmented across different disciplines. Philosophers debate definitions of "actual causation" to determine which specific event caused an outcome. Statisticians and epidemiologists use Structural Causal Models (SCMs) or potential outcomes to measure average treatment effects across populations. Meanwhile, computer scientists build distinct logical frameworks to handle the side effects of actions or to verify the correctness of concurrent software systems.
This fragmentation creates a fundamental challenge: how can we compare competing theories, or build AI systems capable of reasoning across these disparate domains?
The core insight driving recent research is that these seemingly incompatible paradigms can be unified. By elevating the mathematical abstraction—using parameterized logics, modal causal bases, or measure-theoretic spaces—we can view diverse causal theories not as rivals, but as specific configurations or special cases of a single underlying structure.
Parameterizing Actual Causation
Consider a classic philosophical puzzle: a short circuit occurs, a fire starts, and a building burns down. Did the short circuit cause the fire, or was it the presence of oxygen? What if a lightning strike was about to hit the building a millisecond later anyway (a scenario known as preemption)?
To handle these edge cases, philosophers and computer scientists have proposed numerous competing definitions of actual causation. Instead of forcing a choice between them, Beckers et al. (2015, 2016) demonstrate that these diverse definitions can be unified within a single mathematical framework using Causal Probabilistic (CP) logic.
The mechanism behind this unification is parameterization. The authors show that competing historical definitions differ primarily in how they filter out irrelevant background laws and how they "lock" intrinsic causal mechanisms in place during counterfactual testing. By treating these choices as adjustable parameters within a single logical structure, different philosophical theories can be instantiated simply by changing the parameter configuration. This parameterization proves that these definitions are not fundamentally incompatible, but are instead different ways of querying the same underlying causal story.
Unifying Logic and Intervention
In artificial intelligence, causal queries typically fall into two distinct camps:
- Logic-based abductive explanations: Finding a sufficient set of conditions that guarantees an outcome.
- Intervention-based actual causes: Evaluating counterfactual necessity by hypothetically changing a variable and observing the effect.
de Lima et al. (2024) show that these two schools of thought are actually different query types over a shared propositional structure called a "causal base." Instead of maintaining separate systems for logical sufficiency and counterfactual necessity, a single system can evaluate both. Similarly, Lorini (2023) introduces a rule-based modal language that provides a unified logical syntax to express necessity, possibility, interventions, and counterfactuals simultaneously, mapping their formal interrelations.
┌────────────────────────────────────────┐
│ Shared Causal Base │
│ (Unified Propositional/Modal Logic) │
└───────────────────┬────────────────────┘
│
┌────────────────┴────────────────┐
▼ ▼
Abductive Query Interventional Query
(Logical Sufficiency) (Counterfactual Necessity)
This logical unification also resolves long-standing problems in reasoning about actions. For instance, an action can have indirect consequences (ramifications) or can be blocked by certain conditions (qualifications). McCain et al. (1995) demonstrate that ramifications and qualifications are not distinct logical problems but are different manifestations of the same underlying causal rules. When these rules contain cyclic dependencies, Lin et al. (2011) show that a uniform "double minimization" technique can compile both acyclic and cyclic side effects into standard axioms without requiring specialized modeling strategies for loops. Going further back, Geffner (1990) establishes that diverse semantics for nonmonotonic logic programming can be unified under a single causal explanation framework that minimizes unexplained abnormalities.
Elevating Causality to Measure Theory
In statistical causal inference, a deep divide has persisted between Structural Causal Models (SCMs), which rely on deterministic functional equations, and the Potential Outcomes framework, which focuses on counterfactual random variables. SCMs provide clear graphical structures but struggle with continuous-time processes or non-linear feedback loops.
To bridge this gap, Park et al. (2023) introduce a measure-theoretic axiomatization of causality built on "causal kernels." By shifting the foundational language from algebraic equations to measure-theoretic probability, they create a superset framework. SCMs and potential outcomes emerge naturally as restricted applications of this more universal language, which seamlessly accommodates cycles and continuous time.
Furthermore, this measure-theoretic shift changes what we can analyze. Traditional causal models require defining coarse, discrete variables (e.g., "Drug Dosage"). Park et al. (2025) extend this to "causal spaces" based on event sets and sigma-algebras. They prove that standard metrics like the Average Treatment Effect (ATE) are mathematically recoverable as special cases within this broader framework. This allows researchers to calculate fine-grained causal effects in complex, high-dimensional domains like computer vision or natural language processing, where data cannot easily be sliced into simple variables.
Unifying Time and Concurrency
Causality becomes even more complex when events unfold over time or occur concurrently across distributed systems.
To unify temporal verification and causal analysis, Finkbeiner et al. (2023) extend temporal logic with counterfactual operators ($QPTL_{cf}$). This creates a single mathematical language capable of translating and evaluating disparate definitions of actual or temporal causality over infinite sequences of events, making it possible to pinpoint the exact causes of system failures.
| Domain | Fragmented Approaches | Unifying Mechanism |
|---|---|---|
| Philosophy | Counterfactual vs. Production theories of actual cause | Parameterized CP-Logic (Beckers et al. 2016) |
| AI & Logic | Abductive explanation vs. Interventional necessity | Propositional Causal Bases (de Lima et al. 2024) |
| Statistics | SCMs vs. Potential Outcomes | Measure-Theoretic Kernels (Park et al. 2023) |
| Formal Methods | Process calculi vs. Petri nets vs. AI definitions | Language-Theoretic Frameworks (Dimitrova et al. 2019) |
For concurrent reactive systems, Dimitrova et al. (2019) present a language-theoretic framework that integrates disparate causality definitions from both AI and formal methods into a single mathematical structure, allowing practitioners to evaluate necessary and sufficient causes side-by-side. This aligns with work by Künnemann et al. (2019), which integrates process calculi and Petri nets to model control flow explicitly, resolving tricky preemption scenarios where the order of concurrent events dictates the causal outcome. Finally, Xu (2025) formalizes the classical concept of "efficient causation" as an invariant interface, demonstrating that a single core inferential structure can bridge diverse reasoning types from syllogistic logic to counterfactual analysis.
By shifting from ad-hoc, domain-specific definitions to these unified mathematical foundations, the field is moving toward a rigorous, general science of causality—one where logic, statistics, and computation speak the same language.
Key Foundational Work
A Causal Theory of Ramifications and Qualifications (1995) — contributed 2 concept elements to this direction.
Causal Theories for Nonmonotonic Reasoning (1990) — contributed 1 concept elements to this direction.