Identifying Causal Direction via Non-Gaussianity and Nonlinearity

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 268 papers77 conceptsfrozen 2026-10-04

In classical statistics, the mantra "correlation is not causation" is often treated as a dead end. If we observe two correlated variables, $X$ and $Y$, standard tools cannot tell us if $X \to Y$, $Y \to X$, or if a hidden third variable causes both. This is because, in a world of linear relationships and Gaussian (normal) noise, the statistical signature of $X$ causing $Y$ is identical to $Y$ causing $X$.

However, a transformative body of research has shown that this "observational equivalence" is a special case of the Gaussian distribution, not a fundamental law of nature. By leveraging non-Gaussian noise and nonlinear functional forms, researchers can now mathematically identify the direction of causality from purely observational data.

The Gaussian Barrier to Identifiability

To understand why this works, we must first understand why it usually fails. In a Linear Gaussian model, we might assume $Y = \beta X + \epsilon$, where $\epsilon$ is Gaussian noise independent of $X$. If $X$ is also Gaussian, the joint distribution of $(X, Y)$ is a symmetric bivariate normal distribution.

As noted by Patrik O. Hoyer and Antti Hyttinen in Bayesian Discovery of Linear Acyclic Causal Models, this symmetry makes the causal direction "unidentifiable." You can flip the equation to $X = \gamma Y + \epsilon'$ and find a new Gaussian noise term $\epsilon'$ that is perfectly independent of $Y$. Statistically, the two directions are indistinguishable.

Breaking Symmetry with Non-Gaussianity

The breakthrough came with the realization that real-world noise is rarely perfectly Gaussian. When the noise distribution is skewed or heavy-tailed, the symmetry breaks.

In the LiNGAM (Linear Non-Gaussian Acyclic Model) framework, Patrik O. Hoyer and colleagues demonstrated that if the noise terms are non-Gaussian, the true causal direction is the only one where the inferred noise remains independent of the predictor. If you try to model the effect as the cause, the resulting "noise" will be statistically dependent on the "predictor." Aapo Hyvarinen further simplified this in Pairwise Measures of Causal Direction in Linear Non-Gaussian Acyclic Models by using simple likelihood ratios and cumulants to detect this independence, outperforming complex iterative algorithms.

Nonlinearity as a Blessing

While nonlinearity is often viewed as a complication in modeling, Jonas Peters, Bernhard Schölkopf, and others argued in Nonlinear causal discovery with additive noise models that it is actually a "blessing" for causal discovery.

In an Additive Noise Model (ANM), we assume $Y = f(X) + \epsilon$. If $f$ is nonlinear, there is generally no way to write $X = g(Y) + \epsilon'$ such that $\epsilon'$ is independent of $Y$. The "wiggliness" of the nonlinear function $f$ creates a specific pattern in the joint distribution that cannot be mirrored in the reverse direction. This allows researchers to identify the "arrow of time" or the flow of causality even when the noise is Gaussian, provided the relationship itself is not linear.

Handling Hidden Variables and Measurement Error

In practice, we rarely measure every relevant variable. Hidden "latent" variables can create spurious correlations. Feng Xie has pioneered methods like LaHME and LLCS-AD to identify causal hierarchies among these unobserved factors. In Identification of Linear Non-Gaussian Latent Hierarchical Structure, Xie shows that non-Gaussianity allows us to "see" through the latent layers, identifying structures even when hidden variables have no direct observed children.

Similarly, measurement error—noise in the sensors themselves—usually biases causal estimates. Kun Zhang and Peter Spirtes addressed this in Causal Discovery with Linear Non-Gaussian Models under Measurement Error, proving that higher-order statistics (beyond simple covariance) can recover the true causal strength even when the data is "fuzzy."

Multi-View and Non-Stationary Data

Recent work has even found ways to bypass the non-Gaussian requirement. Ambroise Heurtebise and Aapo Hyvarinen introduced LiMVAM in Identifiable Multi-View Causal Discovery Without Non-Gaussianity. They show that if you have multiple "views" of the same system (e.g., different subjects reacting to the same stimulus), the cross-view correlations provide enough constraints to identify the causal graph, even if every variable is perfectly Gaussian.

Similarly, Biwei Huang and Kun Zhang argue in Causal Discovery and Forecasting in Nonstationary Environments that non-stationarity—data that changes over time—is a feature, not a bug. These changes act like natural experiments, providing the necessary "jitter" to distinguish the stable causal mechanism from the changing background noise.

Summary of Intuition

The core idea across these 62 papers is that causal directionality leaves a footprint. - In Linear Non-Gaussian models, the footprint is the independence of the noise term. - In Nonlinear models, the footprint is the structural asymmetry of the function. - In Multi-view or Non-stationary models, the footprint is the consistency of the mechanism across different contexts.

By moving beyond second-order statistics (correlation and covariance) and embracing the "messiness" of real-world data—its non-normality, its nonlinearity, and its fluctuations—we can finally distinguish the cause from the effect.

Papers in this version

Showing 60 of 68 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 96 papers
  2. Version 1 · 2026-10-04 · 68 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_causal_identifiability_de397d98,
  title = {causal identifiability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-identifiability/rdr_de397d98372d45208365895b4f7d407d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_identifying_causal_direction_via_non_gau_de397d98_v1,
  title = {Identifying Causal Direction via Non-Gaussianity and Nonlinearity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/identifying-causal-direction-via-non-gaussianity-and-nonlinearity/rdr_de397d98372d45208365895b4f7d407d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}