Most foundational frameworks for causal inference rely on a convenient assumption: causality is a one-way street. In Directed Acyclic Graphs (DAGs), information flows exclusively from causes to effects without ever looping back. This acyclic assumption makes the mathematics clean, allowing us to easily order variables and determine conditional independencies.
However, the real world is full of loops. In biology, genes regulate one another through reciprocal feedback pathways. In economics, supply and demand perpetually dictate each other's trajectories. When variables mutually influence one another, traditional DAG-based frameworks break down. We can no its longer define a simple topological ordering, and standard algebraic tools fail to resolve the recursive dependencies.
To model these complex systems, researchers have developed cyclic causal modeling—a paradigm shift that re-imagines causal graphs not as static, one-way flowcharts, but as systems operating in equilibrium.
The Equilibrium View of Cycles
How can a causal model contain a loop without triggering an infinite mathematical regress? The answer lies in separating the fast, fine-grained temporal dynamics from the high-level system behavior.
When a physical system governed by differential equations reaches a steady state, its continuous feedback loops settle into a stationary equilibrium. As shown by Joris M. Mooij et al. in From Ordinary Differential Equations to Structural Causal Models: the deterministic case, cyclic structural equation models naturally emerge as valid abstractions of these underlying acyclic, time-dependent dynamical systems. Instead of tracking every micro-step through time, we can model the system at its stationary state using simultaneous equations.
To ensure this abstraction is mathematically sound, Paul K. Rubenstein et al. formalize these relationships in Causal Consistency of Structural Equation Models. They prove that high-level cyclic representations can remain perfectly consistent with low-level physical realities, accurately predicting the outcomes of interventions even when the underlying temporal details are omitted.
Alternative Mathematical Representations
Representing feedback loops through standard structural equations can introduce severe algebraic constraints, often failing to yield unique solutions. To bypass these solvability issues, researchers have introduced more flexible mathematical representations that do not rely on rigid algebraic equations.
- Measure-Theoretic Kernels: Rather than forcing unique algebraic solutions, Junhyung Park et al. propose a framework in A Measure-Theoretic Axiomatisation of Causality that models causal relationships through interventional probability kernels. By treating cyclic feedback loops as stable equilibria of these kernels, they avoid the algebraic pitfalls of traditional structural equations.
- Probability Factorization: Moving away from graphs entirely, Gecia Bravo Hermsdorff et al. introduce the Interventional Factor Model (IFM) in Intervention Generalization: A View from Factor Graph Models. By focusing purely on how probability distributions factorize, IFMs remain completely agnostic to feedback loops and hidden variables, offering a robust alternative for predicting novel combinations of interventions.
- Implicit Flows: For complex, nonlinear systems, Muralikrishnna Guruswamy Sethuraman et al. utilize implicit flows in Differentiable Cyclic Causal Discovery Under Unmeasured Confounders. This approach maps observed variables to exogenous noise through an invertible equilibrium state, capturing intricate feedback loops without requiring explicit temporal modeling.
Tracking Information Flow Through Loops
In standard DAGs, we use a graphical criterion called d-separation to determine whether two variables are conditionally independent. If all paths between two nodes are blocked by a set of conditioning variables, they cannot share statistical information. However, when paths can loop back on themselves, d-separation fails to correctly identify these independencies.
To resolve this, Patrick Forré et al. expand the graphical toolkit in Causal Calculus in the Presence of Cycles, Latent Confounders and Selection Bias. They discuss generalized separation criteria, such as $\sigma$-separation, which accurately capture how information propagates through directed graphs containing cycles. As further explored by Jonas Wahl et al. in Foundations of causal discovery on groups of variables, these generalized criteria allow researchers to formally derive conditional independence statements even when feedback loops and latent confounders complicate the system.
Moreover, identifying specific causal parameters in a cyclic graph requires a rethinking of the graph's components. In the presence of cycles, the authors of Causal Calculus in the Presence of Cycles, Latent Confounders and Selection Bias show that we must aggregate nodes into consolidated "districts" to account for bidirectional information flow. This structural adjustment allows classical identification algorithms to operate safely within cyclic environments.
The Necessity of Interventions
Can we discover these cyclic structures from purely observational data? In acyclic models, observational data is often enough to narrow down the true causal structure to a set of statistically equivalent graphs (a Markov equivalence class). In cyclic models, however, the task is significantly harder because feedback loops obscure the direction of cause and effect.
As Jan-Christian Hütter demonstrates in Estimation Rates for Sparse Linear Cyclic Causal Models, cyclic causal models strictly require interventional data to achieve identifiability and resolve feedback loops.
To understand why, imagine a thermostat regulating room temperature: the temperature dictates the thermostat's state, and the thermostat dictates the temperature. Observational data merely shows a static correlation between the two. Only by actively intervening—such as forcing the thermostat to a specific setting or introducing an external heat source—can we break the feedback loop and isolate the individual directional effects.
Key Foundational Work
A Probabilistic Network of Predicates (1992) — contributed 1 concept elements to this direction.
A Discovery Algorithm for Directed Cyclic Graphs (1996) — contributed 1 concept elements to this direction.