Causal Discovery for Non-Linear and Non-Gaussian Systems

This direction is merged into another direction: non-linear causal discovery. Its last version stays readable and citable here.

Version 1 of 118 papers18 conceptsfrozen 2026-10-04

Causal discovery is often described as the "holy grail" of data science. As Jack Kuipers and Giusi Moffa argue in The interventional Bayesian Gaussian equivalent score for Bayesian causal inference with unknown soft interventions, we rarely want to know just that two variables are correlated; we want to know if intervening on one will change the other.

However, most classical algorithms for uncovering these relationships rely on a "Linear-Gaussian" simplification: they assume that effects are linear combinations of causes and that noise follows a bell curve. While mathematically convenient, this rarely matches the messy, non-linear, and skewed reality of biological or physical systems. A growing body of research is moving beyond these constraints, developing methods that can identify causal structures in the presence of complex dynamics and non-standard noise.

From Rigid Tests to Bayesian Probabilities

Traditional "constraint-based" methods rely on a series of statistical independence tests. If the data is limited or the signal is weak, these tests often fail. Honghua Dai 0001, Kevin B. Korb, and their colleagues demonstrated in A Study of Causal Discovery With Weak Links and Small Samples that Bayesian approaches—specifically Minimum Message Length (MML)—outperform these rigid tests.

The intuition here is to treat causal discovery as a compression problem. A model that accurately reflects the true causal structure will be the most "efficient" way to describe the data. By scoring models based on their information content rather than binary significance tests, Bayesian methods can recover accurate graphs from much smaller, noisier samples.

To scale this to high-dimensional data, Yashas Annadani, Alexandre Lacoste, and others introduced Variational Causal Networks. Instead of searching for a single "best" graph, they use neural networks to learn a distribution over possible graphs, allowing the system to express uncertainty about the causal structure.

Leveraging Generative Models and Transport

If we cannot assume Gaussian noise, how do we model the relationship between variables? Recent work looks toward generative modeling and geometry.

In Diffusion Based Causal Representation Learning, Amir Mohammad Karimi Mamaghan and Stefan Bauer utilize diffusion models to uncover latent causal variables. Unlike standard autoencoders that provide point estimates, diffusion models learn the "score function" of the data distribution. This allows the model to capture "infinite-dimensional" trajectories, making it much easier to distinguish between causal directions in complex, non-linear systems.

Similarly, Luca Ganassali and Negar Kiyavash propose using Optimal Transport in Learning Causal Graphs via Monotone Triangular Transport Maps. By mapping complex distributions onto simpler ones using monotone triangular maps, they create a "distribution-agnostic" framework. This works because the geometry of how one distribution transforms into another reveals the underlying conditional dependencies, regardless of whether the noise is Gaussian or the relationships are linear.

Strength in Numbers and Targeted Reduction

Sometimes, a single dataset isn't enough to resolve causal ambiguity. Xinshi Chen and Le Song show in Multi-task Learning of Order-Consistent Causal Graphs that learning multiple related causal graphs simultaneously can resolve "non-identifiable" cases. If we assume that different tasks share a consistent causal ordering, the joint signal helps prune away incorrect structures that might look plausible in a single task.

When the system is too complex to map entirely, researchers use Targeted Causal Reduction (TCR). As Armin Kekić and Bernhard Schölkopf describe in Targeted Reduction of Causal Models, this involves distilling high-dimensional simulations into a few interpretable causal factors. By focusing only on a specific "target" phenomenon, TCR bypasses the need to model every non-linear interaction in a massive system.

Making Discovery Practical

To move these theoretical advances into the hands of scientists, the field is focusing on automation and reliability.

  • Automation: Konstantina Biza and Sofia Triantafillou developed ETIA, an end-to-end pipeline that automates feature selection and model tuning, specifically aiming to handle the non-linearities that break older tools.
  • Power Analysis: Erich Kummerfeld and Sisi Ma introduced the first formal framework for Power Analysis for Causal Discovery, allowing researchers to calculate exactly how much data is required to find a causal link given a specific effect size.
  • Parameter Tuning: Makoto Yamada and Masashi Sugiyama developed Least-Squares Independence Regression (LSIR), which uses mutual information estimators to automatically tune non-linear models, removing the need for manual "best guesses."

By replacing the rigid assumptions of the past with flexible Bayesian models, generative score functions, and automated pipelines, these researchers are making it possible to discover the true causal engines driving complex real-world data.

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_causal_discovery_for_non_linear_and_non__4640c750,
  title = {Causal Discovery for Non-Linear and Non-Gaussian Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-for-non-linear-and-non-gaussian-systems/rdr_4640c750677f415885006015c96cd31d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_discovery_for_non_linear_and_non__4640c750_v1,
  title = {Causal Discovery for Non-Linear and Non-Gaussian Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-for-non-linear-and-non-gaussian-systems/rdr_4640c750677f415885006015c96cd31d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}