Unique Solvability and Contractivity in Cyclic Causal Models

This direction is merged into another direction: acyclicity constraint. Its last version stays readable and citable here.

Version 1 of 110 papers18 conceptsfrozen 2026-10-04

Most causal inference frameworks rely on the assumption of acyclicity—the idea that cause and effect flow in a one-way street, represented by a Directed Acyclic Graph (DAG). However, real-world systems in biology, economics, and ecology are rarely so simple. They are full of feedback loops where variable $A$ influences $B$, which in turn influences $A$.

While cycles are intuitive in nature, they are mathematically treacherous. A system with feedback can easily become "unsolvable," leading to infinite loops, multiple conflicting equilibria, or no solution at all. Recent research has focused on making these cyclic systems as mathematically rigorous and learnable as their acyclic counterparts by enforcing a property called unique solvability through contractive mappings.

The Problem of Feedback Loops

In a standard DAG, you can determine the value of any variable by simply following the arrows from the "root" causes downward. In a cyclic Structural Causal Model (SCM), this is impossible because variables are defined in terms of each other. For example: $$X = f(Y, \epsilon_X)$$ $$Y = g(X, \epsilon_Y)$$

If $f$ and $g$ are arbitrary functions, there might be many pairs of $(X, Y)$ that satisfy these equations for a given set of noise $(\epsilon_X, \epsilon_Y)$, or perhaps none at all. This ambiguity breaks the fundamental requirements for causal reasoning: if we can't determine a unique state for the system, we cannot calculate likelihoods, perform interventions, or reason about counterfactuals.

Unique Solvability as a Foundation

To fix this, Jonas Peters, Stephan Bongers, Bernhard Schölkopf, and Joris M. Mooij established the theoretical groundwork in Theoretical Aspects of Cyclic Structural Causal Models. They identified unique solvability as the critical property that allows cyclic models to behave predictably.

A cyclic model is uniquely solvable if, for every possible input of exogenous noise, the system of equations yields exactly one output for the internal variables. When a model satisfies this, it obeys the "directed global Markov property," meaning we can use many of the same statistical tools developed for DAGs to analyze systems with feedback.

Ensuring Stability via Contractivity

How do we actually build a model that is guaranteed to be uniquely solvable? The most effective mechanism identified in recent literature is the global contraction condition.

As explored by Saptarshi Saha and colleagues in Cyclic Counterfactuals under Shift–Scale Interventions, if the causal mechanisms are "contractions"—meaning they reduce the distance between points—the system is guaranteed to settle into a single, stable equilibrium point. This is a direct application of the Banach Fixed-Point Theorem: if you repeatedly apply a contractive function in a loop, you will always converge to a unique fixed point regardless of where you start.

By enforcing contractivity (such as $\ell_p$-contraction constraints), researchers ensure that feedback loops do not amplify small perturbations into divergent, infinite behavior. This stability is what makes the system "well-behaved" for counterfactual reasoning.

Differentiable Discovery of Cycles

The shift from theory to practice involves using these constraints to learn cyclic structures directly from data. Traditional causal discovery algorithms often fail when cycles are present because the likelihood functions become mathematically intractable.

Muralikrishnna Guruswamy Sethuraman and Faramarz Fekri have pioneered the use of contractive residual flows to solve this. In works like Learning Cyclic Causal Models from Incomplete Data and MissNODAG, they use neural networks to represent causal mechanisms but constrain the weights to ensure contractivity.

This architectural choice has two massive benefits:

  1. Invertibility: It ensures the mapping between the observed data and the underlying noise is invertible, which is required for calculating the exact log-likelihood of the data.
  2. Differentiability: It allows the entire causal graph—including the cycles—to be learned using standard gradient-based optimization (backpropagation).

By combining these contractive neural networks with frameworks like SCOUT (SCOUT: Cyclic Causal Discovery Under Soft Interventions with Unknown Targets), it is now possible to recover complex, non-linear feedback loops even when the data is messy, incomplete, or comes from interventions where the exact target is unknown. This moves causal discovery out of the "toy-world" of acyclic assumptions and into the complex, interconnected reality of natural systems.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_unique_solvability_and_contractivity_in__7ecf3439,
  title = {Unique Solvability and Contractivity in Cyclic Causal Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unique-solvability-and-contractivity-in-cyclic-causal-models/rdr_7ecf343937d64f8791dc4cd47f4eb28a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_unique_solvability_and_contractivity_in__7ecf3439_v1,
  title = {Unique Solvability and Contractivity in Cyclic Causal Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unique-solvability-and-contractivity-in-cyclic-causal-models/rdr_7ecf343937d64f8791dc4cd47f4eb28a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}