Counterfactual Identifiability Criterion
Abstract
Counterfactual recovery from all single-world marginals is exactly constancy on coupling fibers, and complete Boolean counterfactuals fail this criterion.
Theorem 1.1 (Counterfactual identifiability is marginal-fiber constancy).
Proof. Machine-checked in Lean as D5/S3/ConceptDynamics/Interventions/CounterfactualIdentifiabilityCriterion.counterfactual_identifiable_iff_constant_on_fiber (✓ std3). ∎
Source. Repository-derived.
Commentary.
Let a marginal map send each coupling to its complete family of single-world data. A target is recoverable from those data exactly when equal marginal families force equal target values.
The coupling fiber over a data value is constructed as the preimage of that value. Thus the equality-kernel condition states constancy of the target on every coupling fiber.
Theorem 1.2 (Single-world identifiability is constancy on coupling fibers).
Proof. Machine-checked in Lean as D5/S3/ConceptDynamics/Interventions/CounterfactualIdentifiabilityCriterion.boolean_counterfactual_identifiable_iff_constant_on_coupling_fibers (✓ std3). ∎
Source. Repository-derived.
Commentary.
The observable record of a deterministic Boolean joint model is the family of outcome-count marginals indexed by intervention. A target Q is recoverable from that record exactly when any two models in the same explicitly represented coupling fiber have the same Q-value.
This specializes the general factorization criterion: a target factors through an observable map exactly when it is constant on every fiber. Nonemptiness of the target type permits the factor map to be extended to observable records outside the map’s image.
Lemma 1.3 (The complete counterfactual varies within one coupling fiber).
Proof. Machine-checked in Lean as D5/S3/ConceptDynamics/Interventions/CounterfactualIdentifiabilityCriterion.boolean_counterfactual_varies_on_coupling_fiber (✓ std3). ∎
Source. Repository-derived.
Commentary.
Two deterministic Boolean joint models have the same marginal outcome counts under every intervention, so they occupy a single fiber of the all-single-world-marginals map.
Their complete unit-level counterfactual tables nevertheless differ. The observable fiber therefore contains a concrete variation of the counterfactual target.
Lemma 1.4 (The complete Boolean counterfactual is not identifiable).
Proof. Machine-checked in Lean as D5/S3/ConceptDynamics/Interventions/CounterfactualIdentifiabilityCriterion.boolean_counterfactual_not_identifiable (✓ std3). ∎
Source. Repository-derived.
Commentary.
If the complete unit-level counterfactual table could be recovered from all single-world intervention marginals, the fiber criterion would make it constant on every coupling fiber.
The two-model fiber witness has identical observable marginals but different counterfactual tables, contradicting that required constancy and ruling out every such recovery map.
References
- Truth anchor:
D5/S3/ConceptDynamics/Interventions/CounterfactualIdentifiabilityCriterion.boolean_counterfactual_identifiable_iff_constant_on_coupling_fibers - Truth anchor:
D5/S3/ConceptDynamics/Interventions/CounterfactualIdentifiabilityCriterion.boolean_counterfactual_not_identifiable - Truth anchor:
D5/S3/ConceptDynamics/Interventions/CounterfactualIdentifiabilityCriterion.boolean_counterfactual_varies_on_coupling_fiber - Truth anchor:
D5/S3/ConceptDynamics/Interventions/CounterfactualIdentifiabilityCriterion.counterfactual_identifiable_iff_constant_on_fiber - Dependency: D5/S3/ConceptDynamics/Interventions/InterventionCounterfactualSeparation