Characterize when witness-set sparsity makes PCI sampling efficient

Determine whether the presence of causally inert witness variables generically makes Monte Carlo sampling under the Probabilistic Causal Impact framework substantially more efficient than exhaustive witness-set enumeration, or whether this benefit occurs only in causal graphs that are sparse relative to the tested intervention.

Background

Probabilistic Causal Impact replaces the existential search over witness sets used in actual-causality analysis with Monte Carlo sampling from a distribution over witness sets. The paper observes that, in the worst case, a single revealing witness set among exponentially many candidates may require as many samples to find as exhaustive enumeration would require evaluations.

The authors note that causally inert witnesses are effectively redundant: pinning or releasing them does not alter the counterfactual outcome, so revealing witness sets may occur in neighborhoods rather than as isolated configurations. This can increase their sampling mass and improve practical efficiency. However, the paper does not establish whether this rescue is a general phenomenon or depends on graph sparsity and intervention structure.

References

Whether this rescues sampling generically, or only in graphs sparse relative to the tested intervention, is a sharper open question than we have characterised here.

— A Computationally Feasible Framework for Causal Probabilistic Explanation  (2609.04177 - Urbaniak et al., 3 Sep 2026) in Section 7, Discussion and Conclusions, paragraph beginning “Set against that favourable comparison”