Most causal discovery algorithms rely on a fundamental simplification: the world is a Directed Acyclic Graph (DAG). In a DAG, influence flows in one direction, and you can never return to a starting point by following the arrows. This assumption is mathematically convenient—it ensures that joint probability distributions are well-defined and simplifies the search space for "what causes what."
However, nature rarely respects this constraint. Biological systems rely on feedback loops to maintain homeostasis, and economic markets operate through reciprocal influences. When we force these systems into the "procrustean bed" of acyclicity, we lose the very mechanisms that define their behavior. A new wave of research is moving beyond DAGs to model the cyclic reality of complex systems.
The Conflict Between Homeostasis and Acyclicity
In biology, cycles are not just common; they are essential. A protein might inhibit the very gene that produces it, creating a negative feedback loop that keeps concentrations stable. As noted by Karen Sachs and Munther A. Dahleh in Learning Cyclic Signaling Pathway Structures While Minimizing Data Requirements, traditional Bayesian networks fail here because their mathematical structure cannot represent self-regulating systems.
When an algorithm constrained to DAGs encounters a cycle, it typically does one of two things: it either omits one of the edges to break the loop or identifies a "spurious" direction of causality. This is particularly problematic in fields like neuroimaging. Eli Shlizerman and Rahul Biswas demonstrate in Statistical Perspective on Functional and Causal Neural Connectomics that the feedback loops and self-connections fundamental to neural networks are invisible to standard DAG-based models, leading to inaccurate "wiring diagrams" of the brain.
Cycles as Steady States
If we allow $A \to B$ and $B \to A$, how do we define the value of $A$? In a DAG, we simply look at the parents. In a cycle, the variables are mutually dependent.
The solution often involves treating the system as being at equilibrium. Joris M. Mooij, Bernhard Schölkopf, and colleagues address this in Theoretical Aspects of Cyclic Structural Causal Models. They propose that cyclic Structural Causal Models (SCMs) are valid if they possess "unique solvability"—meaning that for any given noise or external intervention, there is a unique set of variable values that satisfies all equations in the loop. This allows us to treat a cycle as a fixed-point problem where the variables have settled into a stable state.
Resolving Cycles Through Time
Another way to understand cycles is to realize they are often an artifact of how we measure time. If we could observe a system at an infinite frequency, a cycle $A \rightleftarrows B$ would reveal itself as a sequence: $A_t \to B_{t+1} \to A_{t+2}$.
Alexander G. Reisach and Sebastian Weichwald argue in The Case for Time in Causal DAGs that the "acyclicity" assumption is only rigorous when variables are tied to specific time points. When we use "snapshot" data (aggregating over time), we collapse these temporal steps into a single cyclic graph.
This perspective is crucial for high-speed systems. In iCITRIS: Causal Representation Learning for Instantaneous Temporal Effects, Phillip Lippe and Sara Magliacane show that when effects happen faster than the sampling rate, we must model these "instantaneous" interactions carefully to avoid misidentifying the causal structure.
Modern Inference for Cyclic Graphs
Searching for the "best" cyclic graph is even harder than searching for the best DAG because the search space is no longer restricted by a topological sort. Recent work has turned to probabilistic generative models to navigate this complexity.
One promising direction is the use of Generative Flow Networks (GFlowNets). In DynGFN: Towards Bayesian Inference of Gene Regulatory Networks with GFlowNets, Lazar Atanackovic and Alexander Tong use GFlowNets to sample from the posterior distribution of cyclic graphs. Unlike traditional methods that search for a single optimal graph, this approach captures the uncertainty of the regulatory relationships, which is vital when biological data is noisy and multiple different cyclic structures might explain the same observations.
Similarly, Wenqian Li and Shengyu Zhu demonstrate in GFlowCausal that by relaxing the acyclicity constraint, generative models can more effectively discover causal structures in domains like economics where reciprocal relationships are the norm rather than the exception.
Why It Works: The Power of Interventions
The key to identifying cycles—and distinguishing them from simple correlations—is intervention. If we have a cycle $A \to B \to A$, and we manually fix the value of $A$ (a "knockout" in genetics), the link from $B$ back to $A$ is effectively broken. By observing how the system responds to these perturbations, we can "unroll" the cycle.
Mathieu Chevalley and Arash Mehrjou emphasize in A large-scale benchmark for network inference from single-cell perturbation data that algorithms permitting cycles are significantly more accurate at capturing underlying regulatory mechanisms than those constrained by acyclicity. By embracing the complexity of feedback loops, these models move us closer to a truly functional understanding of the world's most intricate systems.