Iterative Identification Closure: Amplifying Causal Identifiability in Linear SEMs

Researchers Ziyi Ding and Xiao-Ping Zhang have introduced Iterative Identification Closure (IIC), a novel framework that significantly advances the field of causal AI. The work addresses a fundamental limitation in causal discovery: the Half-Trek Criterion (HTC), the primary graphical tool for determining if causal effects can be identified from observational data, leaves 15-23% of effects in moderate graphs as "inconclusive." IIC breaks this deadlock by decoupling the identification process into two phases. First, it uses any available external information—like instrumental variables, interventions, or prior knowledge—as a seed to identify an initial set of causal edges. Second, and most crucially, it employs a novel "Reduced HTC propagation" that iteratively substitutes these known coefficients back into the system, reducing its complexity and unlocking the identification of previously indeterminate edges.

This iterative feedback mechanism is the core innovation, a capability absent from all prior graphical criteria. The researchers proved the soundness of this approach with a new theoretical result, the Reduced HTC Theorem, which ensures the mathematical validity of the propagation step. Exhaustive testing on all graphs with up to 5 nodes (covering 134,144 edges) confirmed 100% precision with zero false positives. When combined with seed information, IIC reduced the HTC identification gap by over 80%. The propagation gain was substantial, with a gamma factor of ~4x, meaning a small amount of initial information could be amplified to identify nearly the entire causal graph, far outperforming prior methods that lacked iterative feedback.