Stable mixed graphs

In this paper, we study classes of graphs with three types of edges that capture the modified independence structure of a directed acyclic graph (DAG) after marginalisation over unobserved variables and conditioning on selection variables using the $m$-separation criterion. These include MC, summary, and ancestral graphs. As a modification of MC graphs, we define the class of ribbonless graphs (RGs) that permits the use of the $m$-separation criterion. RGs contain summary and ancestral graphs as subclasses, and each RG can be generated by a DAG after marginalisation and conditioning. We derive simple algorithms to generate RGs, from given DAGs or RGs, and also to generate summary and ancestral graphs in a simple way by further extension of the RG-generating algorithm. This enables us to develop a parallel theory on these three classes and to study the relationships between them as well as the use of each class.

💡 Research Summary

The paper addresses a fundamental limitation of directed acyclic graphs (DAGs) in representing conditional independence when latent (unobserved) variables are marginalized out or selection variables are conditioned on. While DAGs are the cornerstone of many statistical and causal models, they are not stable under these operations: the independence structure among the remaining observed variables can no longer be captured by a DAG. To overcome this, the authors examine three existing families of “stable mixed graphs” – MC graphs (MCGs), summary graphs (SGs), and ancestral graphs (AGs) – all of which employ three edge types (directed arrows, bidirected arcs, and undirected lines) to encode the effects of marginalisation and conditioning.

The main contribution is the introduction of ribbon‑less graphs (RGs), a new subclass of loop‑less mixed graphs (LMGs). An RG is defined by forbidding a specific configuration called a “ribbon”: a V‑shaped collider where the two outer nodes are not directly connected and the inner collider node either participates in an undirected line or belongs to a directed‑cycle. By eliminating ribbons and loops, the authors guarantee that the standard m‑separation criterion (originally defined for ancestral graphs) can be applied directly to RGs, yielding a unique independence model (J_m(G)).

Key theoretical results include:

1. Stability: For any DAG (G) and any disjoint marginalisation set (M) and conditioning set (C), there exists an RG (H) such that (J_m(H) = \alpha(J_m(G); M, C)), where (\alpha) denotes the formal marginalisation/conditioning operator on independence models. Thus RGs are closed under these operations, unlike DAGs.
2. Algorithmic Generation: The paper provides polynomial‑time algorithms that (a) transform a given RG (or directly a DAG) into the RG representing the desired marginalised/conditioned model, (b) extend this procedure to produce the corresponding summary graph and ancestral graph. The algorithms work locally: they add edges to eliminate ribbons, delete loops, and preserve arrowheads at collider endpoints.
3. Equivalence of Operations: The authors prove that splitting the marginalisation or conditioning set into two subsets and applying the transformations sequentially yields the same final graph as applying them jointly. This property ensures that the graphical representation is independent of the order of operations.
4. Inclusion Relationships: RGs form a superset that contains SGs and AGs as proper subclasses; undirected graphs (UGs) and bidirected graphs (BGs) are further special cases. Conversely, DAGs are not stable and cannot be recovered after marginalisation or conditioning without moving to a larger class.
5. Practical Implications: By providing a unified framework where a single separation criterion (m‑separation) suffices for all three stable mixed graph families, the work simplifies causal inference with latent variables and selection bias. Researchers can start from a DAG, apply the RG‑generation algorithm to obtain a graph that faithfully represents the post‑marginalisation independence structure, and then optionally convert it to a summary or ancestral graph depending on the analysis needs (e.g., focusing on conditional independences vs. ancestral relations).

The paper’s structure proceeds systematically: after reviewing basic graph‑theoretic notions and formalising independence models, it defines RGs and the m‑separation rule, formalises marginalisation and conditioning on independence models, and establishes the stability concept. Subsequent sections detail the three graph families, present the generation algorithms, prove correctness, and explore the relationships among them. An appendix contains rigorous proofs of all lemmas and theorems.

In summary, the authors deliver a coherent theoretical and algorithmic toolkit that extends the expressive power of graphical models beyond DAGs while preserving a single, intuitive separation criterion. This advancement has direct relevance for fields such as econometrics, social science, and artificial intelligence, where hidden confounders and selection mechanisms are ubiquitous and where robust, graph‑based representations of conditional independence are essential.