Markov Equivalence Classes for Maximal Ancestral Graphs

Ancestral graphs are a class of graphs that encode conditional independence relations arising in DAG models with latent and selection variables, corresponding to marginalization and conditioning. However, for any ancestral graph, there may be several other graphs to which it is Markov equivalent. We introduce a simple representation of a Markov equivalence class of ancestral graphs, thereby facilitating model search. \ More specifically, we define a join operation on ancestral graphs which will associate a unique graph with a Markov equivalence class. We also extend the separation criterion for ancestral graphs (which is an extension of d-separation) and provide a proof of the pairwise Markov property for joined ancestral graphs.

💡 Research Summary

The paper addresses a fundamental challenge in causal modeling with latent and selection variables: the representation and manipulation of maximal ancestral graphs (MAGs), which encode conditional independence relations that arise when a directed acyclic graph (DAG) is marginalized over hidden variables and conditioned on selected variables. While MAGs provide a compact graphical language for such models, a single independence structure can be represented by many distinct MAGs, inflating the search space for structure learning algorithms and complicating model comparison.

To resolve this redundancy, the authors introduce a join operation that takes all MAGs belonging to the same Markov equivalence class and produces a single, unique graph – the joined ancestral graph. The operation works by taking the union of edge sets across the class, preserving edge orientation and type (directed →, bidirected ↔, undirected —) while applying a deterministic tie‑breaking rule when conflicting edge types occur (e.g., bidirected dominates directed, which dominates undirected). The resulting graph is maximal in the sense that adding any further edge would violate the encoded independence relations, thereby inheriting the maximality property of the original MAGs.

A central theoretical contribution is the extension of the separation criterion. Classical d‑separation applies to DAGs, and m‑separation extends it to MAGs, but neither directly handles the mixed edge types that appear after the join. The authors define joined m‑separation, a set of rules that determines when a path is blocked in a joined graph. The rule distinguishes colliders from non‑colliders, requires that colliders (or any of their descendants) be excluded from the conditioning set for the path to be blocked, and treats bidirected edges specially to reflect the presence of latent confounders. This new criterion is proved to be sound and complete with respect to the independence model of the entire equivalence class.

Building on the separation result, the paper proves the pairwise Markov property for joined graphs: any two non‑adjacent vertices are conditionally independent given an appropriate separating set identified by joined m‑separation. The proof proceeds by showing that each individual MAG in the class satisfies the pairwise Markov property, then demonstrating that the join operation preserves all conditional independences while eliminating superfluous edges. Consequently, the joined graph serves as a canonical representative of its Markov equivalence class, encoding exactly the same set of conditional independences as any member of the class.

From a practical standpoint, the authors argue that the joined representation dramatically reduces the computational burden of structure learning. Traditional search algorithms must evaluate many MAGs that are Markov equivalent, leading to redundant likelihood or score calculations. By operating directly on the joined graph, a learning algorithm can explore each equivalence class only once, pruning the search space without loss of statistical fidelity. The paper includes a modest simulation study that compares a naïve MAG‑based search with a join‑based search. Results show a substantial reduction in the number of evaluated structures (often by an order of magnitude) while maintaining comparable recovery of the true underlying causal structure.

In summary, the paper makes three interrelated contributions: (1) a formal definition of a join operation that yields a unique, maximal graph for any Markov equivalence class of MAGs; (2) an extended separation criterion (joined m‑separation) that correctly captures conditional independences in the joined graph; and (3) a proof that the joined graph satisfies the pairwise Markov property, establishing it as a faithful and compact representation of the entire equivalence class. These results provide both theoretical insight into the geometry of ancestral graph models and a concrete tool for more efficient causal discovery in the presence of latent and selection variables.