Most causal discovery algorithms operate under a convenient but restrictive "fine print" assumption: the world is a Directed Acyclic Graph (DAG). In a DAG, influence flows in one direction, like a waterfall. If variable $A$ causes $B$, then $B$ cannot, under any circumstances, cause $A$.
However, the real world is rarely so linear. In biology, gene regulatory networks rely on feedback loops to maintain homeostasis. In economics, supply affects price, which in turn affects supply. When we force these "non-recursive" systems into acyclic models, we don't just lose detail—we produce fundamentally incorrect maps of reality. Recent research is shifting toward non-recursive causal models, developing the mathematical and statistical tools to identify and reason about systems where everything might influence everything else.
The Recursive Divide
The formal "rules of the game" for causal reasoning were largely established by Joseph Halpern in Axiomatizing Causal Reasoning. Halpern distinguishes between recursive systems (where variables can be ordered such that each depends only on predecessors) and non-recursive systems (where equations may have cyclic dependencies).
In a recursive model, the effect of an intervention is easy to calculate because the "ripples" of a change never come back to haunt the source. In non-recursive systems, an intervention on $A$ might change $B$, which then circles back to further modify $A$ until the system reaches a new equilibrium. Halpern’s work provides the logical completeness for these complex systems, but it also highlights a steep cost: the computational complexity of determining "actual causality" in non-recursive models can be significantly higher than in their acyclic counterparts The Computational Complexity of Structure-Based Causality.
Overcoming the Gaussian Hurdle
If cycles are so common, why have we avoided them? The primary reason is identifiability. In a linear system with Gaussian noise, it is mathematically impossible to tell if $A \to B$ or $B \to A$ just by looking at correlations. When you add cycles ($A \rightleftharpoons B$), the problem becomes even more "underdetermined."
To break this stalemate, researchers leverage two key properties: non-Gaussianity and sparsity.
- Non-Gaussianity: Yan Zeng and Shohei Shimizu utilize the fact that real-world data is rarely "bell-curved." In Causal Discovery with Multi-Domain LiNGAM for Latent Factors, they show that non-Gaussian distributions provide enough "statistical shape" to orient edges even when hidden factors and multiple domains are involved.
- Structural Sparsity: Even if data is Gaussian, we can identify cyclic structures if the connections are "sparse" enough. Ignavier Ng and Kun Zhang proved in On the Identifiability of Sparse ICA without Assuming Non-Gaussianity that a condition called "structural variability" allows us to recover the true mixing matrix (and thus the causal structure) using only simple second-order statistics. Essentially, if the "map" of connections is sufficiently simple, there is only one way the variables could have generated the observed correlations.
Learning Cycles from Messy Data
Moving from theory to practice requires algorithms that can handle the "missingness" of real-world data. Muralikrishnna Guruswamy Sethuraman and Faramarz Fekri address this in Learning Cyclic Causal Models from Incomplete Data, creating frameworks that don't break down when some variables are unobserved or datasets are incomplete.
This is particularly vital in longitudinal studies (studies over time). Gia H. Ngo introduced GLACIAL GLACIAL: Granger and Learning-based Causality Analysis for Longitudinal Imaging Studies, which combines Granger causality with neural networks to find reciprocal influences in sparse medical imaging data. By allowing for cycles, GLACIAL can model how different biomarkers in the brain might reinforce or inhibit one another over years of patient follow-ups.
Beyond Discovery: Generative Cycles
The push for non-acyclic modeling isn't just about discovery; it's also about generation. Generative Flow Networks (GFlowNets) were originally restricted to DAGs, meaning they couldn't model environments where actions are reversible or states can be revisited (like chemical reactions or physical navigation).
Leo Maxime Brunswic and colleagues extended this frontier in A Theory of Non-Acyclic Generative Flow Networks. By introducing "stable" loss functions, they prevent models from getting trapped in infinite loops during training, allowing GFlowNets to sample from complex, cyclic state spaces.
Why It Works: The Intuition of Equilibrium
The core intuition behind non-recursive modeling is that a cycle is not just a "loop" but an equilibrium state. When we observe a system with feedback, we are often observing the stable point where all mutual influences have balanced out.
By treating causal discovery as a search for the structural equations that satisfy these equilibrium conditions—rather than just a search for a linear ordering—researchers like Niklas Pfister and Jonas Peters are enabling us to model the world as it actually is: a web of reciprocal influences rather than a one-way street Identifiability of Sparse Causal Effects using Instrumental Variables.