In classical causal inference, the Directed Acyclic Graph (DAG) is the gold standard. It allows us to use d-separation—a simple set of rules for tracing paths—to determine which variables are conditionally independent. However, real-world systems like biological pathways, economic markets, and climate systems are rarely acyclic; they are full of feedback loops.
When cycles are introduced, the elegant logic of d-separation breaks down. Recent research has focused on extending these foundations to Directed Mixed Graphs (DMGs) and Structural Causal Models (SCMs) that incorporate feedback, latent confounders, and selection bias.
The Failure of d-Separation in Cycles
In an acyclic graph, d-separation perfectly captures the relationship between the graph's topology and the statistical dependencies in the data. But as Katie Everett demonstrates in Cycles in Causal Learning, cyclic structures can create "information cancellations." A feedback loop can result in zero mutual information, making variables appear statistically independent even when they are functionally dependent.
Standard d-separation fails here because it cannot account for how non-linear feedback loops propagate information. If we rely on it, we risk missing critical causal links or incorrectly predicting the effects of interventions.
Beyond Paths: Sigma-Separation and HEDGes
To fix this, researchers have developed more robust separation criteria. The most prominent is $\sigma$-separation (sigma-separation). Unlike d-separation, which looks at paths, $\sigma$-separation is designed to handle the "unique solvability" of cyclic equations.
Patrick Forré and Joris M. Mooij have been central to this shift. In Markov Properties for Graphical Models with Cycles and Latent Variables, they introduce HEDGes (Directed Graphs with Hyperedges) and a refined path-blocking criterion. This criterion behaves exactly like d-separation in acyclic graphs but remains theoretically sound when cycles are present, correctly identifying independencies in non-linear systems where traditional methods fail.
From Graph Tracing to Algebraic Relations
While $\sigma$-separation solves the problem for many graphs, some researchers argue that we need a more fundamental mathematical language. Jean-Philippe Chancelier and Michel De Lara propose shifting from "path-tracing" to the language of binary relations.
In Conditional Separation as a Binary Relation, they generalize d-separation to cyclic and even infinite graphs by treating it as an algebraic property. Their follow-up work on Information Dependency Models uses "topological closure" to verify independence. If the functional ancestors of two sets of variables don't overlap (disjoint closures), the variables are independent. This provides a unified framework that handles cycles and context-specific dependencies without needing a graph at all.
The Challenge of Faithfulness
Even with better separation criteria, discovering these structures from data remains difficult. A core assumption in causal discovery is Faithfulness: the idea that every statistical independence in the data corresponds to a separation in the graph.
Jonas Wahl and Jakob Runge show in Foundations of causal discovery on groups of variables that while the Markov property (the graph implies independence) usually holds, Faithfulness often breaks down in cyclic systems or when variables are grouped into "macro" variables. This suggests that in cyclic domains, we cannot rely on independence tests alone; we must use higher-order statistics, such as the non-Gaussianity approaches explored by Marina Garrote-López and colleagues, to distinguish between equivalent cyclic structures.