Sample Complexity of Causal Identification with Temporal Heterogeneity

Recovering a unique causal graph from observational data is an ill-posed problem because multiple generating mechanisms can lead to the same observational distribution. This problem becomes solvable only by exploiting specific structural or distributional assumptions. While recent work has separately utilized time-series dynamics or multi-environment heterogeneity to constrain this problem, we integrate both as complementary sources of heterogeneity. This integration yields unified necessary identifiability conditions and enables a rigorous analysis of the statistical limits of recovery under thin versus heavy-tailed noise. In particular, temporal structure is shown to effectively substitute for missing environmental diversity, possibly achieving identifiability even under insufficient heterogeneity. Extending this analysis to heavy-tailed (Student’s t) distributions, we demonstrate that while geometric identifiability conditions remain invariant, the sample complexity diverges significantly from the Gaussian baseline. Explicit information-theoretic bounds quantify this cost of robustness, establishing the fundamental limits of covariance-based causal graph recovery methods in realistic non-stationary systems. This work shifts the focus from whether causal structure is identifiable to whether it is statistically recoverable in practice.

💡 Research Summary

The paper tackles the fundamental challenge of recovering a unique causal graph from observational data, a problem that is ill‑posed because many distinct structural causal models can generate the same joint distribution. Traditional approaches either exploit temporal precedence in time‑series data or leverage distributional shifts across multiple environments (regimes). Both strategies, however, rely on the assumption that the source of heterogeneity—temporal or environmental—is sufficiently rich to break the symmetry of the Markov equivalence class.

The authors propose a unified framework that simultaneously harnesses two complementary sources of heterogeneity: (i) temporal heteroskedasticity, i.e., time‑varying variances within a single environment, and (ii) multi‑environment variance scaling, i.e., different diagonal covariance matrices across distinct regimes. The underlying data‑generating process is a linear structural equation model (SEM) with instantaneous effects: Xₜ = BᵀXₜ + εₜ, where B encodes the directed acyclic graph (DAG) and εₜ are independent structural noises. Across environments u = 1,…,k, the noise covariance Σ(u) = diag(σ²₁,u,…,σ²_d,u) varies, while the causal mechanism B remains invariant.

The first major contribution is a set of identifiability conditions that remain valid even when environmental heterogeneity is rank‑deficient. Specifically, if each environment provides at most rank‑r variance information (r < d), then a temporal window of length T ≥ ⌈d/r⌉ is necessary and sufficient for any second‑order‑moment‑based method to uniquely recover the DAG. This result (Theorem 4.2) rests on two mild assumptions: (a) Temporal Faithfulness of the Jacobian of the mixing function at the mean of the latent process, and (b) Distinct Variance Profiles, meaning that each variable’s variance trajectory across time or regimes is unique. The bound depends solely on the geometric rank of the variance shifts and is invariant to the tail behavior of the noise distribution.

The second contribution extends the analysis to heavy‑tailed noise modeled by a multivariate Student’s‑t distribution with ν > 4 degrees of freedom. The authors derive closed‑form expressions for fourth‑order mixed moments (Proposition 4.3) and show that the covariance estimator’s variance is inflated by a factor (1 + 3/(ν‑4)) relative to the Gaussian case. Consequently, the sample complexity required to estimate the covariance matrix within a relative error ε with confidence 1 − δ scales as N(ν) ≥ Cδ·(1 + 3/(ν‑4))/ε² (Proposition 4.4). This yields a clear quantitative penalty: as ν decreases (heavier tails), the number of samples needed grows, potentially doubling the Gaussian baseline for modest ν. Importantly, the geometric identifiability condition (T ≥ ⌈d/r⌉) remains unchanged; only the statistical cost of detecting variance shifts increases.

From an algorithmic perspective, the paper highlights that standard covariance‑based causal discovery methods (e.g., ICA‑based approaches, PCMCI) are statistically inefficient under heavy‑tailed noise unless the sample size satisfies the derived lower bound. Robust alternatives—such as M‑estimators, trimming, or methods that exploit higher‑order moments—are necessary to approach the theoretical limits. Moreover, the temporal window length T must be chosen to satisfy the rank condition, which may require longer observation periods in practice.

Overall, the work shifts the focus from “is the causal structure identifiable?” to “is it statistically recoverable given realistic non‑stationary, heavy‑tailed data?”. By integrating temporal and multi‑environment heterogeneity, providing unified identifiability theorems, and quantifying the heavy‑tail sample‑complexity penalty through information‑theoretic bounds, the paper offers a comprehensive theoretical foundation for causal discovery in modern, non‑Gaussian, non‑stationary systems.