Presented at AAAI 2015, this research provides a formal bridge between two major but historically separate approaches to causality: Judea Pearl’s structural causal models (SCMs), widely used in statistics and social sciences, and the "causal calculus," a nonmonotonic logic developed within the artificial intelligence community for reasoning about actions and change.
Introduction
For decades, causal reasoning has been divided into two distinct methodologies. One camp, led by Judea Pearl, utilizes Structural Causal Models—mathematical equations that describe how variables determine one another. The other camp, rooted in AI research, employs specialized logics to reason about how specific actions alter the state of the world.
The primary challenge in unifying these fields is that Pearl’s equations are not merely mathematical equalities; they possess a directional, causal meaning where the left side represents the effect and the right side represents the cause. Classical logic struggles to capture this inherent asymmetry. This paper aims to resolve this disconnect by demonstrating that Pearl’s causality can be fully expressed within a logical setting without sacrificing its unique properties.
Methodology: The Causal Calculus
The authors utilize the Causal Calculus (developed by McCain, Turner, and Lifschitz) as their primary formal tool. This is a nonmonotonic logic, meaning that the introduction of new information can invalidate previous conclusions—a necessary feature for reasoning about interventions and changing environments.
To bridge the gap, the authors developed several key mapping techniques:
- Mapping Equations to Rules: They translate structural equations (e.g., $V = f(U)$) into logical causal rules of the form $F \Rightarrow A$ ("F causes A"). This preserves the directional nature of the original model.
- Representing Exogenous Variables: Background variables not explained by the model are represented using "self-causing" rules ($U \Rightarrow U$). This allows these variables to take any value while ensuring they are formally accounted for within the logical theory.
- Defining Submodels and Interventions: Pearl’s $do$-operator (forcing a variable to a specific value) is shown to be logically equivalent to forming subtheories—specifically, replacing existing rules in a causal theory with new constants.
- Functional Completion: The authors use a technique called "completion" to transform complex causal rules into classical logical formulas. This allows them to prove that the solutions to Pearl's equations are identical to the "exact models" of the corresponding logical theory.
Key Insights
The Asymmetry of Causation
In standard algebra, $x = y$ is equivalent to $y = x$. However, in Pearl’s SCMs, the relationship is one-way. The paper demonstrates that the causal rule $A \Rightarrow B$ in logic perfectly captures this asymmetry, ensuring that the logic respects the flow of influence from cause to effect.
Nonmonotonicity and the Principle of Causality
The paper argues that for a logic to represent Pearl's worlds accurately, it must be nonmonotonic. It operates on the principle that "everything that is true must have a cause." If a proposition lacks a causal explanation within the model, it cannot be part of a valid "causal world."
Modularity and Flexibility
A hallmark of Pearl’s models is modularity—the ability to change one mechanism (one equation) without affecting others. The authors prove that their logical representation preserves this modularity, making it "intervention-equivalent" to Pearl’s original models. Furthermore, the logical framework is more flexible than standard SCMs, as it can naturally handle cyclic dependencies and models with multiple (or zero) solutions.
Formal Results
The paper establishes several core theorems that solidify this logical foundation:
- Theorems 3 & 4: These prove that for both propositional and first-order cases, the "causal worlds" (mathematical solutions) of a Pearl model are exactly the same as the "exact models" (logical solutions) of the corresponding causal theory.
- Theorems 5 & 6: These establish a correspondence between "causal equivalence" in logic and "intervention equivalence" in Pearl’s theory. If two causal theories are logically equivalent, they will behave identically under any possible intervention.
Limitations and Future Work
While the paper successfully unifies these frameworks, the authors identify areas for further exploration. The computational complexity of finding minimal sets of "parents" for modular rules can be intensive. Additionally, while the paper lays the groundwork for interventions, it only briefly addresses counterfactuals (reasoning about "what might have been").
The authors suggest that the logical generality of the causal calculus could eventually allow for more complex counterfactual reasoning than Pearl’s original theory—such as handling disjunctive antecedents (e.g., "If it had been sunny OR I had an umbrella..."). Finally, while the logic accommodates cycles, the practical implications of cyclic causal models in this specific logical setting require deeper investigation.
Summary updated