Causal Discovery via Heteroscedastic Location Scale Models

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

Version 1 of 276 papers114 conceptsfrozen 2026-10-04

In the search for causal relationships—determining whether $X$ causes $Y$ or $Y$ causes $X$—researchers have traditionally relied on Additive Noise Models (ANMs). These models assume that an effect is simply the sum of a function of its cause and some independent "noise" ($Y = f(X) + \epsilon$). Crucially, they often assume homoscedasticity: the idea that the variance of the noise $\epsilon$ remains constant regardless of the value of the cause $X$.

However, real-world data is rarely so uniform. In many systems, the "jitter" or uncertainty of an effect changes depending on the magnitude of the cause—a phenomenon known as heteroscedasticity. For example, the variance in a person's spending habits (effect) typically increases as their income (cause) grows. Recent research has shifted from viewing this varying noise as a nuisance to treating it as a powerful signal for identifying causal direction.

The Signal in the Variance

The core intuition behind this research direction is that the "shape" of noise provides a fingerprint of causality. In a Location-Scale Noise Model (LSNM), the effect is modeled not just by a mean function (location), but also by a scaling function that modulates the noise based on the cause: $$Y = f(X) + \sigma(X) \cdot \epsilon$$ As Alexander Immer, Alexander Marx, and colleagues demonstrate in On the Identifiability and Estimation of Causal Location-Scale Noise Models, this structure makes the causal direction identifiable.

Why does this work? If $X \to Y$ is the true direction, the residuals of $Y$ after accounting for $X$ will be independent of $X$ once they are properly scaled by $\sigma(X)$. However, if we try to model the relationship in reverse ($Y \to X$), the resulting residuals will almost always remain dependent on $Y$ in a complex, non-linear way. This asymmetry allows algorithms to "see" the flow of causation by finding the direction where the noise can be most cleanly decomposed into a function of the input.

Moving Beyond Averages with Quantiles

Traditional methods often focus on the conditional mean of the data. But heteroscedasticity lives in the "spread" of the distribution. To capture this, Natasa Tagasovska and her team developed bQCD (Bivariate Quantile Causal Discovery) in Distinguishing Cause from Effect Using Quantiles: Bivariate Quantile Causal Discovery.

Instead of looking at the average relationship, bQCD examines different "slices" or quantiles of the data. By analyzing how well these quantiles can be compressed or modeled, the method can identify causes even when the noise interacts with the causal variable in complex ways. Similarly, Yikang Chen and Dehui Du propose using the Quantile Partial Effect (QPE) in Causal Discovery via Quantile Partial Effect, proving that the causal direction is identifiable because the "shape" of the relationship is inherently simpler in the true causal direction.

Concomitant Estimation and Robustness

One major challenge in causal discovery is that we often don't know the noise levels in advance. If an algorithm assumes all variables have equal noise (homoscedasticity) when they don't, it may hallucinate "spurious edges"—false causal links—simply because it mistakes high noise for causal influence.

To solve this, Gonzalo Mateos and Seyed Saman Saboksayr introduced CoLiDE (Concomitant Linear DAG Estimation) in CoLiDE: Concomitant Linear DAG Estimation. Their approach "concomitantly" estimates the graph structure and the noise scales for each node simultaneously. By allowing each variable to have its own noise profile, the model becomes robust to sensors or data sources with varying reliability.

Handling Messy Real-World Data

The research has recently expanded to handle even more complex noise profiles:

By embracing the complexity of noise rather than smoothing it away, these researchers have turned heteroscedasticity from a statistical hurdle into a primary tool for uncovering the hidden causal architecture of the world.

Papers in this version

Showing 60 of 76 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 75 papers
  2. Version 1 · 2026-10-04 · 76 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_heteroscedastic_noise_modeling_homosceda_e86cedc8,
  title = {heteroscedastic noise modeling (homoscedasticity assumption)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/heteroscedastic-noise-modeling-homoscedasticity-assumption/rdr_e86cedc8c852462092f07a5167830b55},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_discovery_via_heteroscedastic_loc_e86cedc8_v1,
  title = {Causal Discovery via Heteroscedastic Location Scale Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-via-heteroscedastic-location-scale-models/rdr_e86cedc8c852462092f07a5167830b55/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}