Cycle-Consistent Causal Modeling

• Cycle-consistent causal modeling is a framework that formalizes feedback and equilibrium phenomena in systems with cyclic dependencies using directed graphs and joint distributions.
• It employs specialized techniques including constraint-based discovery, relational acyclification, and non-Gaussian SEMs to capture mutual influences that traditional DAG models overlook.
• The approach has broad implications for quantum, physical, and relational systems, though challenges like overlapping cycles and model robustness remain.

Cycle-Consistent Causal Modeling formalizes, identifies, and exploits feedback in causal systems represented as directed cycles in graphical models. Unlike acyclic causal discovery frameworks, cycle-consistent (or cyclic) models are essential for describing equilibrium phenomena, mutual influence, and certain physical, quantum, or relational systems where standard DAG assumptions fail. Developments across constraint-based, algebraic, information-theoretic, quantum, and relational frameworks provide a rigorous foundation for cycle-consistent modeling, with specialized identifiability, algorithmic, and interpretational results.

1. Foundations and Definitions

A cycle-consistent causal model is typically a pair $(G, P)$, where $G$ is a directed graph that may contain cycles, and $P$ is a joint distribution on observed variables $X_1, ..., X_n$ compatible with $G$ in the sense of the directed Markov property or its generalizations. Compatibility ensures that every d-separation in $G$ implies a corresponding conditional independence in $P$; faithfulness is sometimes assumed but not always required, particularly in frameworks allowing fine-tuning or non-classical effects (Vilasini et al., 2021).

Cyclic causal models arise from equilibrium or fixed-point equations:

$X = f(X, E),$

with $f$ possibly depending on $X$ in a mutually entangled fashion. Unique solvability via global contraction (i.e., for some $G$0, $G$1) guarantees well-posedness for both factual and interventional queries in general cyclic structural causal models (SCMs) (Saha et al., 28 Oct 2025).

In relational domains, cycle-consistent causal modeling leverages relational schemas $G$2 and instantiates dependencies $G$3 along relational paths, leading naturally to highly structured, potentially cyclic ground graphs (Ahsan et al., 2022).

2. Information-Theoretic and Algebraic Properties

Cycle-consistent (self-referentially factorizing) distributions possess subtle and often counterintuitive information-theoretic signatures. For a collection $G$4, the factorization

$G$5

(“cycle product”) induces the following phenomena (Everett et al., 2020):

• Zero Marginal Mutual Information: For $G$6, any such distribution has $G$7, i.e., all variables are mutually independent at the observational level, even when interventions reveal nontrivial feedback.
• Statistical Indistinguishability: Passive, observational dependence tests (mutual information, correlation, HSIC, etc.) fail to detect cycles, leading to “observer’s paradox.”
• Mechanism Coupling: The joint distribution constrains conditional mechanisms globally, violating the “independent mechanisms” assumption.
• Reversal Symmetry: Reverse-cycle factorization holds, suggesting indistinguishability between forward and backward time in equilibrium.

These results underscore a fundamental barrier to constraint-based discovery in general cyclic models.

3. Constraint-Based Discovery and Relational Acyclification

Constraint-based algorithms such as PC or FCI rely fundamentally on acyclicity. Relational and cyclic generalizations introduce specialized operations and assumptions:

• Relational Acyclification: For any cyclic relational causal model (RCM), there exists a “relational acyclification” $G$8—a DAG with a bounded hop threshold $G$9—such that the same set of observable conditional independencies is preserved. This enables reduction of cyclic discovery to acyclic discovery under appropriate “$P$0-faithfulness” and relational acyclification assumptions (Ahsan et al., 2022).
• Soundness and Completeness: Under these conditions, constraint-based relational discovery (RCD algorithm) is both sound and Markov-complete for cyclic RCMs; the DPAG output recovers all ancestral relations, arrowhead, and tail completions, and distinguishes Markov equivalence classes.

Empirical work demonstrates that $P$1-separation-based discovery (as opposed to classical d-separation) perfectly recovers cyclic relational structures in synthetic and real-world datasets (Ahsan et al., 2022).

4. Causal Discovery in Linear Cyclic and Non-Gaussian Models

Linear non-Gaussian SEMs with cycles admit unique identification properties unattainable in the Gaussian case. Several frameworks address cycle-consistent discovery in these models:

• Equilibrium Equation: $P$2 with $P$3 possibly cyclic (spectral norm $P$4 for stability), $P$5 independent, non-Gaussian.
• Identifiability: ICA/ISA recovers the LiNG-equivalence class (up to certain row-permutation and scaling)—in the acyclic case uniquely, in the cyclic case up to equivalence transformations (Dai et al., 2024).
• Block-Topological Ordering: For cycle-disjoint graphs, quadratic (2×2) and cubic (3×3) low-order moment polynomial constraints can localize source cycles, whose effects are then regressed out recursively to yield a causal order among cycles (Drton et al., 14 Jul 2025).
• Efficiency: The cycle discovery and regression algorithms are consistent under mild moment conditions and scale cubically with the number of variables (Drton et al., 14 Jul 2025).
• Practical Considerations: Approaches assume non-zero third cumulants, i.i.d. samples, and faithfulness; overlapping cycles remain challenging (Dai et al., 2024).

Local ISA-based and regression-based methods recover Markov blanket structure and edge weights exactly in both cyclic and acyclic regimes, with consistent empirical superiority over acyclic-only methods (Dai et al., 2024).

5. Counterfactuals and Interventions in Cyclic SCMs

Classical counterfactual inference relies on acyclicity for the well-posedness of abduction–action–prediction procedures. Cycle-consistent SCMs extend these principles:

• Unique Solvability: Global contraction ensures the unique fixed-point solution for any (possibly cyclic) set of structural equations under both factual and post-interventional (soft, shift–scale) transformations (Saha et al., 28 Oct 2025).
• Twin-SCM Construction: To define unit-level counterfactuals, a coupled twin system of the form $P$6 is constructed, with the intervention applied only to the second copy.
• Cycle-Consistent Counterfactuals: The distribution of the counterfactual variables produced via the twin-SCM coincides with that obtained from classical abduction–action–prediction when unique solvability holds, extending the soundness of counterfactual inference to cyclic domains (Saha et al., 28 Oct 2025).

A worked example with linear feedback illustrates that population and individual-level counterfactual distributions match exactly, validating the theory.

6. Quantum and Fine-Tuned Causal Cycles

Cycle-consistency also arises in quantum and post-quantum scenarios, often with non-classical or fine-tuned mechanisms (Vanrietvelde et al., 2022, Vilasini et al., 2021):

• Quantum Routed-Graphs: Consistent quantum processes with cyclic causal order can be constructed using directed routed graphs with Boolean route-matrices. Two key conditions—bi-univocality and weak-loops on the branch graph—guarantee logical consistency (no “grandfather paradox”). Cyclic feedback in these models manifests as classical “which-branch” feedback and never as paradoxical quantum feedback (Vanrietvelde et al., 2022).
• Fine-Tuned Cycles: Classical and quantum models may admit cycles that are “fine-tuned”—i.e., compatible with observed independencies but undetectable via observational tests. A taxonomy of “affects causal loops” (ACLs) distinguishes operationally detectable from hidden causal cycles, and compatibility with spacetime structures is analyzed to prevent signaling paradoxes (Vilasini et al., 2021).

These frameworks are crucial for modeling processes in quantum information, cryptographic protocols, and violations of causal inequalities.

7. Algorithmic Frameworks and Empirical Evaluation

Algorithmic work achieves consistent, scalable inference for a large class of cycle-consistent models:

Framework	Model Class	Key Conditions	Identifiability	Notes
Cyclic RCD (Ahsan et al., 2022)	Relational CRCMs	$P$7-faithful, acyclification	Markov-complete	Handles feedback in relational data
CCI (Strobl, 2018)	Linear SEM (CLS)	Faithful, invertible	MAAG correct	Handles cycles, latents, selection
Block-Order (Drton et al., 14 Jul 2025)	LiNG, disjoint cycles	Non-Gaussian, cycle-disjoint	Block-topological, consistent	Polynomial time
ISA/ISA-Local (Dai et al., 2024)	LiNG, general cycles	Non-Gaussian	Local equivalence class	Exact recovery on Markov blanket

Empirical validation consistently demonstrates that cycle-consistent approaches outperform acyclic-limited algorithms on synthetic and real datasets, particularly in the presence of feedback loops, latent variables, and selection bias (Ahsan et al., 2022, Drton et al., 14 Jul 2025, Dai et al., 2024, Strobl, 2018).

8. Open Problems and Future Directions

Open challenges include scalable learning with overlapping cycles, robustness to model mis-specification, non-linear cyclic dependencies, and integration with domain knowledge and soft constraints. Connections to indefinite causal order (quantum), stability of causal embeddings in spacetime, and causality in general probabilistic theories represent active research frontiers (Vanrietvelde et al., 2022, Vilasini et al., 2021).

Cycle-consistent causal modeling offers a comprehensive, theoretically grounded paradigm for discovering, interpreting, and manipulating feedback-driven processes, with rigorous guarantees across a diverse array of structural, statistical, and logical contexts.