Causal Discovery via Distributional Invariance

• Distributional-invariance-based causal discovery is a paradigm that identifies stable causal mechanisms by exploiting invariant statistical relationships across varying environments.
• It employs techniques like differentiable invariance constraints, synthetic downsampling, and risk minimization to distinguish true causal edges from spurious dependencies.
• Empirical studies demonstrate improved accuracy, reduced false discovery rates, and enhanced scalability in both linear and nonlinear structural equation models.

Distributional-invariance-based causal discovery refers to a paradigm in which the identification of causal structure leverages invariance properties of conditional or marginal distributions across multiple environments, interventions, or transformations of observed data. The core premise is that true causal mechanisms remain stable—i.e., their functional form or statistical relationships do not change—across shifts in (i) the distribution of exogenous noise, (ii) intervened components, or (iii) observed marginal distributions, while spurious associations typically lack such invariance. Modern frameworks operationalize this intuition either through optimization constraints, test statistics, or explicit regularization terms in a causal search objective.

1. Foundational Principles

Distributional invariance in causal discovery is rooted in the assumption that the structural equations determining the value of each system variable remain fixed across environments, whereas the distribution of noise components or the distribution of observed variables may shift. Let $X_1, \ldots, X_d$ be the observed variables and $\mathcal{E} = \{1, \ldots, E\}$ the index set of environments or domains. The data from each environment $e$ consists of samples $X^{(e)} \in \mathbb{R}^d$ generated by a structural equation model (SEM) of the form: $X_j = F_j(\mathrm{Pa}(X_j)) + z_j^{(e)},$ with the key invariance property that $\{F_j\}_{j=1}^d$ are fixed across $e$, but the noise distributions $z_j^{(e)}$ may vary in variance, functional form, or even support (Wang et al., 2022, Montagna et al., 13 May 2026).

Causal edges $X_i \to X_j$ are defined as environment-invariant when the relationship remains unchanged under all plausible environment-dependent perturbations. More formally, for any candidate DAG $G$ and parameter set $\mathcal{E} = \{1, \ldots, E\}$0, the invariance criterion asserts that the optimal conditional distribution $\mathcal{E} = \{1, \ldots, E\}$1 (as fit in each environment $\mathcal{E} = \{1, \ldots, E\}$2) must be invariant to $\mathcal{E} = \{1, \ldots, E\}$3 for the true structure, but not necessarily for spurious edges (Wang et al., 2022, Nguyen et al., 3 Feb 2026).

2. Identifiability via Invariance

The major theoretical breakthrough enabled by distributional invariance is the identifiable recovery of causal graphs in scenarios (particularly, non-linear, non-Gaussian settings) where purely observational data are fundamentally limited by Markov equivalence.

• Finite Environments, Nonlinear SCMs: Under acyclicity and invariance assumptions, only two auxiliary environments with generic shifts in noise suffice to identify both the DAG and SEM mechanisms up to permissible ambiguities (component-wise invertible transformations of noise) (Montagna et al., 13 May 2026). Faithfulness between the observed joint and the underlying graph structure is assumed.
• Linear Gaussian SCMs: For additive linear models, identifiability of the DAG is guaranteed when, for each non-source node, its noise variance varies across at least two environments while variances for all other nodes remain fixed between those environments (Wang et al., 2022). Lemmas in this setting show that only the true graph and parameters minimize the joint loss and penalty enforcing invariance.
• Mixed Graph Models: Invariance properties under interventions refine the Markov equivalence class (MEC) to an interventional MEC (iMEC), enabling unique identification in the presence of latent confounders or selection bias (Solus, 2019).
• Single-environment identifiability: In some GLM settings with known noise dispersion, joint conditions of Pearson risk invariance and likelihood maximization suffice for identification—even without explicit multi-environment data (Polinelli et al., 2024).

3. Model Formulations and Computational Strategies

The exploitation of distributional invariance is realized through a diverse toolkit, with the central motif being the penalization or enforcement of invariance constraints:

• Differentiable Invariant Causal Discovery (DICD): Optimizes over structure matrix $\mathcal{E} = \{1, \ldots, E\}$4 and SEM parameters $\mathcal{E} = \{1, \ldots, E\}$5, minimizing reconstruction loss plus a penalty measuring the environment-wise violation of parameter invariance. The acyclicity constraint is relaxed via differentiable smoothness (e.g., NOTEARS trace exponential method). The invariance penalty enforces vanishing gradients with respect to edge rescaling parameters $\mathcal{E} = \{1, \ldots, E\}$6 in each environment (Wang et al., 2022).
• GLIDE Algorithm: Uses observational data to synthetically generate several perturbed datasets with altered marginals of candidate parent sets. For each effect $\mathcal{E} = \{1, \ldots, E\}$7, tests whether the conditional distribution $\mathcal{E} = \{1, \ldots, E\}$8 is invariant to such changes—holding only for $\mathcal{E} = \{1, \ldots, E\}$9 being the true parent set. The test statistic is a sample variance across environments, with parent sets identified via minimization over plausible Markov Blanket-based cliques, ensuring scalability and avoiding exponential subset enumeration (Nguyen et al., 3 Feb 2026).
• NegDRO and Group-DRO: Invariance is enforced by solving a minimax optimization where one seeks predictors with equal risk across all environments. The negative-weight DRO (“NegDRO”) formulation allows for negative environment weights, which, while breaking convexity, provably identifies the unique causal model under minimal heterogeneity conditions (Wang et al., 2024).
• Information-Theoretic Linear Programming Approaches: Formalize the invariance constraint between observational and interventional distributions as a set of linear equalities, transforming causal discovery into a KL-minimization LP, tractable in small discrete settings (Gmeiner, 2020).

4. Weak, Strong, and Mixed Invariance Criteria

Methods vary in how stringently the invariance property is enforced:

• Strong Invariance: Demands exact equality of conditionals or mechanism parameters across environments or interventions (e.g., $e$0 for all $e$1) (Nguyen et al., 3 Feb 2026).
• Weak or Partial Invariance: Enforces that only some feature of the latent variable (e.g., marginal distribution, support, variance) remains invariant, permitting imperfection, e.g., only a block of coordinates or a subset of environment shifts (Ahuja et al., 2023).
• Nonparametric Invariance: The mechanism is invariant not only to environment but also under arbitrary reparametrizations (e.g., monotonic bijections of $e$2 or $e$3), yielding methods robust to marginal transformations (Jørgensen et al., 2020).

Table: Main invariance-based causal discovery approaches

Method	Environment requirement	Invariance property
DICD (Wang et al., 2022)	Observed domains ($e$4)	SEM parameter invariance
GLIDE (Nguyen et al., 3 Feb 2026)	Synthetic downsampling	$e$5 invariance across marginals
NegDRO (Wang et al., 2024)	$e$6 environments, additive	Equal risk across environments
CD-NOD (Huang et al., 2019)	Surrogate domain variable $e$7	Local module invariance/independent change
Causal de Finetti (Guo et al., 2022)	Exchangeable multi-environment	CI structure lifting to ICM
Info-theoretic LP (Gmeiner, 2020)	Observational+interventional	KL-minimizing, $e$8

5. Experimental Validation and Empirical Results

Comparative studies extensively benchmark distributional-invariance-based methods against state-of-the-art alternatives—note, for example:

• DICD reduces structural Hamming distance (SHD) by up to 36% and consistently outperforms NOTEARS, DAG-GNN, and regression invariance methods, especially in suppressing spurious edges in both linear and nonlinear settings (Wang et al., 2022).
• GLIDE achieves order-of-magnitude speedups (up to 25×) over NOTEARS and PC, while matching or improving accuracy and lowering the false discovery rate, even on graphs of $e$9 nodes (Nguyen et al., 3 Feb 2026).
• NegDRO scales to $X^{(e)} \in \mathbb{R}^d$0 predictors, in contrast to $X^{(e)} \in \mathbb{R}^d$1 for subset search, and has empirical risk decay as $X^{(e)} \in \mathbb{R}^d$2, with robustness in settings with limited interventions (Wang et al., 2024).
• Causal de Finetti empirically verifies that exchangeable data under ICM admits identifiability of the full DAG, outperforming PC, GES, and ICP in both bivariate and multivariate settings (Guo et al., 2022).
• Nonparametric MQV-based inference exhibits robustness to reparametrization and does not degrade on real (CEP) data if full bijection-marginalization is enforced (Jørgensen et al., 2020).

6. Extensions, Limitations, and Open Questions

• Sufficiency of invariance: In many settings, distributional invariance is a necessary but not sufficient condition for causal directions, especially in the presence of hidden confounding, insufficiently rich interventions or in near-degenerate structural regimes (Gmeiner, 2020).
• Number of environments: Two environments are often sufficient for identifiability in nonlinear acyclic SCMs, but detection power saturates at moderate environment counts (Montagna et al., 13 May 2026). The efficacy of specific invariance penalties in single-environment settings relies on strong model class assumptions (e.g., GLM with known dispersion (Polinelli et al., 2024)).
• Types of shifts: Most results rest on generic noise-variance shifts or mechanism perturbations; non-generic or collinear interventions can result in theoretical non-identifiability unless explicitly dealt with by more sophisticated test designs (Montagna et al., 13 May 2026).
• Computational scaling: Design of polynomial-time algorithms underpins progress beyond the exponential bottlenecks of subset enumeration inherited from classical invariant prediction or constraint-based methods (Wang et al., 2024, Nguyen et al., 3 Feb 2026).
• Extensions and future work: Generalization to continuous environment parameters, learning environment partitions, and applications to high-dimensional domains (graph, vision) represent active research directions (Qiu et al., 23 Oct 2025, Yao et al., 2024). The use of invariance in unsupervised latent discovery and joint causal representation learning is gaining prominence (Ahuja et al., 2023, Yao et al., 2024).

7. Relationship to Other Paradigms and Synthesis

Distributional-invariance-based causal discovery connects the graphical independence-based tradition (e.g., constraint-based algorithms), interventional methods, and invariant risk minimization (IRM) into a unified framework wherein environment-induced shifts, either observed or synthetically generated, act as "natural experiments." The invariant conditional or risk property can be viewed as a statistical signature of causality, and by operationalizing this property via optimization, subset search, or nonparametric statistical tests, these methods can overcome the limitations of learning from purely i.i.d. data and approach full identifiability in fundamentally non-identifiable regimes (Wang et al., 2022, Guo et al., 2022, Montagna et al., 13 May 2026). Theoretical results and empirical validations have established these approaches as state-of-the-art in both high-dimensional and nonparametric settings, though certain open problems remain in settings of weak interventions or single-domain data.