📊 statistics

Causal Discovery in Structural VAR Models Under Equal Noise Variance

This paper addresses the challenge of causal discovery in linear Gaussian structural VAR models with equal noise variance by establishing a theoretical framework for observational equivalence and proposing ENVAR, a sparsity-based algorithm that identifies a representative causal graph within the equivalence class, validated on both synthetic and fMRI data.

Original authors: SeyedSina Seyedi HasanAbadi, Fahimeh Arab, Erfan Nozari, AmirEmad Ghassami

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: SeyedSina Seyedi HasanAbadi, Fahimeh Arab, Erfan Nozari, AmirEmad Ghassami

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out how a group of friends influences each other's moods throughout the day. You have a video recording (the data) of them talking, but the camera is slow. It only takes a picture once every minute, even though their conversations and reactions happen in split seconds.

This is the core problem the paper addresses: Causal Discovery in Structural VAR Models.

Here is a breakdown of the paper's ideas using simple analogies:

1. The Problem: The "Slow Camera" Effect

In fields like neuroscience (specifically fMRI brain scans), scientists want to know which part of the brain causes activity in another part.

The Reality: Neurons fire and talk to each other in milliseconds.
The Data: The scanner only takes a "snapshot" every second or so.
The Confusion: Because the camera is so slow, two things might happen within the same snapshot:
1. Lagged Effects: Brain A influenced Brain B from the previous second (easy to spot).
2. Contemporaneous Effects: Brain A and Brain B influenced each other instantly within that same second (hard to spot).

If you try to draw a map of who influences whom, you might get it wrong because the "instant" influences look like a tangled web, and you can't tell who started the conversation.

2. The Special Rule: "Equal Noise"

The authors introduce a specific rule to help solve this puzzle: Equal Noise Variance.

The Analogy: Imagine every friend in the group has a slightly different "background noise" (like a fan humming in the background) that affects their mood. Usually, these fans are different sizes and volumes, making it impossible to tell who is talking to whom.
The Assumption: The authors assume that all these background fans are exactly the same volume and type.
Why it helps: If you know the "static" is identical for everyone, you can subtract it out more easily to see the actual conversations.

3. The Big Surprise: One Map, Many Possibilities

In many previous studies, if you had this "equal noise" rule, you could find the one true map of who influences whom.

The Paper's Discovery: In this time-series setting (where things happen over time), the "equal noise" rule does not give you a single, unique map.
The Analogy: Imagine you have a Rubik's Cube. You can twist and rotate the cube (change the perspective), and the colors on the outside (the data you see) look exactly the same, but the internal structure of the cube has changed.
The Result: There isn't just one "correct" map. Instead, there is a whole family of maps (called an Observational Equivalence Class) that all produce the exact same video footage. You cannot tell them apart just by looking at the data.

4. The Solution: The "ENVAR" Procedure

Since we can't find the one true map, the authors propose a new method called ENVAR (Equal-Noise VAR).

The Strategy: Instead of hunting for the single "true" map, ENVAR looks for the simplest map within that family of possibilities.
The Analogy: Imagine you have a suitcase full of different outfits that all look the same from a distance. You can't tell which one is the "real" one, but you decide to pick the one with the fewest buttons and pockets (the sparsest one) because it's the most elegant and likely to be the intended design.
How it works: The math searches through all the possible "rotated" maps and picks the one where the fewest connections exist. This is based on the idea that nature usually prefers simple explanations over complex ones.

5. Testing the Idea

The authors tested their method in two ways:

Synthetic Data: They created fake brain data where they knew the "true" answer. They showed that their method (ENVAR) found a map much closer to the truth than other existing methods.
Real fMRI Data: They applied it to real brain scans from people doing a motor task (moving their fingers).
- The Result: The map they produced highlighted the parts of the brain known to control movement (like the motor cortex) in a way that matched what scientists already know about human biology.

Summary

The paper says: "When looking at slow-motion brain data, assuming all background noise is equal doesn't give us a single answer. Instead, it gives us a group of possible answers. Our new tool, ENVAR, finds the simplest, most logical map within that group, and it works better than current tools at finding the right connections."

Technical Summary: Causal Discovery in Structural VAR Models Under Equal Noise Variance

Problem Statement
Causal discovery from multivariate time series is particularly challenging when causal effects occur both across time (lagged) and within the same sampling interval (contemporaneous). This issue is critical in domains like neuroscience, where sampling rates (e.g., fMRI) are often coarse relative to underlying dynamics, necessitating models that capture within-bin interactions. While linear Gaussian Structural Vector Autoregressive (SVAR) models are powerful for this, standard identification techniques often fail in this setting.

Unlike cross-sectional Structural Equation Models (SEMs) where an equal noise variance assumption combined with a Directed Acyclic Graph (DAG) assumption yields point identification of a unique causal graph, the time-series setting with contemporaneous effects does not generally yield a unique graph. In SVAR models, contemporaneous effects can form directed cycles, and multiple distinct structural parameterizations can induce the exact same stationary observed process law. Consequently, the target of causal discovery shifts from identifying a single graph to identifying an observational equivalence class.

Methodology
The paper proposes a framework for causal discovery in linear Gaussian SVAR(1) models under the assumption of equal noise variance (structural noise terms are mutually independent with a common variance $\sigma^2$ ).

Observational Equivalence Characterization:
The authors prove that two admissible SVAR models $M = (A_0, A_1, \sigma)$ and $M' = (A'_0, A'_1, \sigma')$ are observationally equivalent (inducing the same stationary Gaussian process) if and only if their structural parameters are related by an orthogonal transformation and a global positive scale. Specifically, there exists an orthogonal matrix $Q \in O(p)$ and a scalar $c > 0$ such that:
$B' = cQB, \quad A'_1 = cQA_1, \quad \sigma' = c\sigma$
where $B = I - A_0$ . This characterization implies that the structural graph is not point-identified; rather, the data identifies a class of models related by these transformations.
Observational Alignment Discrepancy:
To evaluate model recovery in this setting, the authors introduce the observational alignment discrepancy ( $\Delta^{\text{obs}}_{\text{align}}$ ). Unlike standard Frobenius errors, this metric measures the distance between a candidate model and the entire observational equivalence class of the ground truth. It is defined as the infimum of the reconstruction error over all valid orthogonal transformations and global scalings. A closed-form solution is derived using Singular Value Decomposition (SVD), allowing for efficient evaluation of structural recovery without needing to match a specific arbitrary representative.
ENVAR Algorithm:
The authors propose ENVAR (Equal-Noise VAR causal discovery), a sparsity-based procedure to select a preferred representative from the equivalence class. The algorithm proceeds in three steps:
- Reduced-Form Estimation: Estimate the reduced-form transition matrix $\Phi$ and residual covariance $\Sigma_u$ via Ordinary Least Squares (OLS).
- Canonical Representative Construction: Construct a base structural representation using the Cholesky factor of the empirical precision matrix.
- Sparse Search: Search over the orthogonal manifold and global scale to find a structural representative that minimizes the $\ell_1$ norm of the off-diagonal entries of $A_0$ (contemporaneous effects) and all entries of $A_1$ (lagged effects), subject to the normalization constraint $\text{diag}(I - A_0) = 1$ . This is solved using non-convex optimization on the orthogonal manifold (via PyTorch's orthogonal parameterization).

Key Results
The methodology was evaluated on synthetic data and real fMRI data from the Human Connectome Project (HCP).

Synthetic Data: ENVAR was compared against EqVarGDS, VARLiNGAM, and DYNOTEARS. ENVAR achieved the lowest scale-free observational alignment discrepancy across various graph dimensions ( $p$ ), indicating it successfully recovers models closest to the true equivalence class. It also demonstrated higher Pearson correlation with ground-truth reduced-form parameters and structural matrices compared to baselines. Notably, ENVAR showed robustness to violations of the equal-variance assumption, maintaining superior performance even when noise variances varied moderately across nodes.
Real fMRI Data: Applied to HCP motor task data, ENVAR recovered causal graphs where nodal centralities (in-degree, out-degree, net flow) aligned remarkably well with known neuroscientific activation maps. NeuroSynth analysis confirmed that the decoded cognitive terms associated with the learned graph structures were consistent with motor control literature.

Significance and Claims
The paper claims that under equal noise variance, the structural graph in SVAR models is generally not point-identified due to the presence of contemporaneous cycles and the invariance of the observed law under orthogonal transformations. Therefore, the appropriate target for causal discovery is the observational equivalence class, not a single graph.

The significance of this work lies in:

Providing an exact characterization of this equivalence class for time-series data, distinguishing it from the cross-sectional DAG setting.
Introducing an equivalence-aware error metric (observational alignment discrepancy) that correctly evaluates structural recovery by accounting for the inherent non-identifiability.
Proposing ENVAR, a practical algorithm that leverages sparsity to select a normalized representative from the equivalence class, demonstrating improved accuracy over state-of-the-art methods in both synthetic and real-world neuroimaging scenarios.

The authors note limitations, including the reliance on linearity, stationarity, and the absence of unobserved confounders, as well as the use of non-convex optimization which may converge to local minima. Future work is suggested to address non-stationarities and higher-order dynamics.