A Transformational Characterization of Markov Equivalence for Directed Acyclic Graphs with Latent Variables

Different directed acyclic graphs (DAGs) may be Markov equivalent in the sense that they entail the same conditional independence relations among the observed variables. Chickering (1995) provided a transformational characterization of Markov equivalence for DAGs (with no latent variables), which is useful in deriving properties shared by Markov equivalent DAGs, and, with certain generalization, is needed to prove the asymptotic correctness of a search procedure over Markov equivalence classes, known as the GES algorithm. For DAG models with latent variables, maximal ancestral graphs (MAGs) provide a neat representation that facilitates model search. However, no transformational characterization – analogous to Chickering’s – of Markov equivalent MAGs is yet available. This paper establishes such a characterization for directed MAGs, which we expect will have similar uses as it does for DAGs.

💡 Research Summary

The paper addresses a fundamental gap in causal discovery theory: while the transformational characterization of Markov equivalence for directed acyclic graphs (DAGs) without latent variables is well‑established through Chickering’s 1995 results, an analogous framework for models that include latent variables has been missing. Latent variables render the observed‑variable independence structure more complex, and maximal ancestral graphs (MAGs) have become the standard representation for such settings because they can encode directed, bidirected, and undirected edges that capture direct causation, hidden common causes, and selection bias respectively.

The authors first review the classic DAG case. Chickering showed that two DAGs are Markov equivalent if and only if they share the same skeleton (the underlying undirected graph) and the same set of unshielded colliders (v‑structures). Moreover, he introduced a set of elementary operations—covered edge reversals and edge additions/removals—that can transform any DAG in an equivalence class into any other while preserving the Markov properties. These operations underpin the correctness proof of the Greedy Equivalence Search (GES) algorithm, which navigates the space of equivalence classes efficiently.

Turning to MAGs, the paper notes that existing criteria for MAG equivalence (identical skeleton, identical unshielded colliders, and identical discriminating paths) are purely declarative; they do not provide a constructive way to move between equivalent graphs. To fill this void, the authors define two elementary transformations that respect the ancestral and m‑separation properties of MAGs:

1. Covered Directed Edge Flip – If a directed edge X→Y is “covered” (i.e., the parent sets of X and Y, excluding each other, are identical), its orientation may be reversed to Y→X. This operation mirrors the covered edge reversal in DAGs but must be checked against the presence of bidirected and undirected edges to ensure that no new cycles or violations of ancestry are introduced.

2. Covered Bidirected Edge Replacement – If a bidirected edge X↔Y is covered, it can be replaced by a pair of opposite directed edges (X→Y and Y→X) or, conversely, a pair of opposite directed edges that form a covered configuration can be collapsed into a single bidirected edge. The authors prove that this replacement preserves all conditional independencies encoded by the original MAG.

The central theorem states that two MAGs are Markov equivalent if and only if there exists a finite sequence of the above covered flips and replacements that transforms one into the other. Conversely, any sequence of such operations applied to a MAG yields another MAG that is Markov equivalent to the original. The proof proceeds by showing that each elementary operation maintains the m‑separation relations, the ancestral property (no directed cycles), and the maximality condition (no additional edges can be added without violating ancestralness). The authors also demonstrate that any difference between two equivalent MAGs can be resolved by a series of local covered modifications, thereby establishing completeness of the transformation set.

Having established the transformational characterization, the paper explores its algorithmic implications. In particular, the authors outline how a MAG‑adapted version of GES (often called GES‑MAG) can use the covered flips and replacements as its basic move set. Because each move is guaranteed to stay within the same equivalence class, the algorithm can focus on scoring equivalence classes rather than individual graphs, dramatically reducing the number of required conditional independence tests. The paper also discusses integration with constraint‑based methods such as the Fast Causal Inference (FCI) algorithm, where the new transformations can be employed to navigate the partially oriented output of FCI more systematically.

The discussion concludes with a candid assessment of limitations and future work. The current transformation set is tailored to maximal ancestral graphs; extending it to more general mixed graphs, to graphs with partially observed variables, or to settings with cycles (e.g., structural equation models with feedback) remains an open challenge. Moreover, while the authors prove existence of a transformation sequence, they do not yet provide bounds on the minimal number of steps required, nor do they analyze the computational complexity of finding an optimal sequence. Addressing these questions could lead to more efficient search strategies and deeper theoretical insight into the geometry of MAG equivalence classes.

In summary, the paper delivers the first transformational characterization of Markov equivalence for directed MAGs, mirroring Chickering’s seminal work for DAGs. This contribution fills a crucial theoretical gap, offers a foundation for provably correct search algorithms in the presence of latent variables, and opens several promising avenues for further research in causal structure learning.