Causal Discovery with Feedback Loops

462 concepts321 papers1 proposal

Most foundational frameworks in causal inference rely on a comforting geometric assumption: causality flows in one direction, never looping back on itself. This is the Directed Acyclic Graph (DAG) assumption. In a DAG, if variable $A$ causes variable $B$, then $B$ cannot cause $A$, either directly or through a chain of intermediate variables.

DAGs are mathematically elegant. They allow us to factorize joint probability distributions cleanly, use d-separation to read off conditional independencies, and calculate the effects of actions using Pearl's do-calculus. However, the real world rarely conforms to this strict one-way flow.

The Reality of Reciprocal Causation

In many domains, variables operate in continuous, dynamic loops where feedback is a fundamental feature rather than an exception:

  • Biology: In gene regulatory networks, Gene $A$ might activate Gene $B$, which over time triggers a suppressor mechanism that turns off Gene $A$.
  • Economics: Higher prices reduce consumer demand, but reduced demand eventually forces prices down.
  • Engineering and Manufacturing: Industrial systems rely on continuous feedback loops where sensors adjust actuators, and actuators immediately alter the sensor readings, as highlighted by Johannes Huegle et al.

When we force an acyclicity constraint onto these systems, our causal discovery algorithms break down. As noted by Charles K. Assaad et al. and David S. Watson et al., forcing a cyclic system into an acyclic framework means we fail to model reciprocal relationships accurately, leading to incorrect structural conclusions and flawed policy interventions.

The Illusion of Simultaneity

Why do cycles appear in our data in the first place? Often, it is an artifact of temporal aggregation.

Imagine a high-frequency trading system where Event $A$ triggers Event $B$ within milliseconds, and $B$ loops back to influence $A$ a millisecond later. If our sensors or data collection tools only record snapshots every second, these distinct sequential events get compressed into the same observation window.

To the algorithm, the interactions look instantaneous and bidirectional. Jie Qiao et al. and Phillip Lippe et al. point out that this "instantaneous effect" illusion is a major hurdle for standard causal representation learning. Similarly, Andreas Gerhardus argues that assuming contemporaneous acyclicity—the idea that no feedback loops occur within a single time step—severely limits our ability to model real-world time series where variables exert mutual, rapid influence on one another.

When we strip away the acyclicity assumption, classic guarantees for identifying causal parameters disappear. For instance, Bryant Chen et al. and Amin Jaber et al. show that traditional completeness results and auxiliary variable methods struggle to maintain mathematical validity when recursive, loop-free structures are replaced by cyclic ones.

How to Discover and Model Cycles

If the acyclicity constraint is a barrier, how do we bypass it? Researchers have developed several distinct mechanisms to discover and reason about cyclic causal structures from data.

1. Extending Graphical Separation Criteria

The first question is whether our graphical tools still work when arrows loop back. Can we still use conditional independence to infer structure?

Judea Pearl et al. provided a foundational answer by proving that the d-separation criterion remains a valid test for reading off conditional independencies in cyclic graphs, provided the variables are discrete. This means that even when feedback loops are present, the statistical footprint of separation in the graph still corresponds to independence in the data distribution.

2. Exploiting Non-Gaussian Noise Asymmetry

In purely linear, Gaussian systems, the direction of causality is statistically indistinguishable; $A \rightarrow B$ looks identical to $B \rightarrow A$. However, if the background noise driving the system is non-Gaussian, an asymmetry emerges. The residual errors from a linear regression will only be independent of the predictor in the true causal direction.

Joris M. Mooij et al. extended this intuition to cyclic additive noise models. They demonstrated that under non-Gaussian noise, the exact cyclic structure is uniquely identifiable from steady-state observational data.

Because mapping a global cyclic network can be computationally explosive, Haoyue Dai et al. leveraged Independent Subspace Analysis to perform local causal discovery. This allows researchers to isolate and identify the cyclic neighbors and feedback strengths of a single target variable without needing to reconstruct the entire massive system.

3. SAT-Based Constraint Solving

When a system contains both feedback loops and unmeasured common causes (latent confounders), the space of possible graphs becomes vast and difficult to navigate using traditional greedy search algorithms.

To solve this, Antti Hyttinen et al. formulated causal discovery as a Boolean satisfiability (SAT) problem. By encoding d-separation constraints directly into a logical framework, a SAT solver can systematically deduce whether a cyclic edge is present, absent, or underdetermined by the data, providing a rigorous way to handle cycles and hidden variables simultaneously.

4. Interventional Potential Functions

If we have loops, how do we mathematically define what happens when we step in and change a variable (an intervention)?

Mark Schmidt et al. addressed this by introducing Directed Cyclic Graphical Models (DCGs). Instead of relying on the standard directed factorization that breaks down in cycles, DCGs utilize "interventional potential functions" combined with global normalization. This formulation bridges the gap between directed graphs and undirected models, allowing the system to handle feedback loops naturally while preserving clean, mathematically rigorous interventional semantics.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_acyclicity_constraint_3aeb030e,
  title = {acyclicity constraint},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/acyclicity-constraint/rdr_3aeb030ea5fd4085b8580035d70255db},
  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_acyclicity_constraint_3aeb030e_v1,
  title = {acyclicity constraint},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/acyclicity-constraint/rdr_3aeb030ea5fd4085b8580035d70255db/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake