Causal Graph Recovery

• Causal graph recovery is a process that reconstructs the underlying directed acyclic graph (DAG) from observational or interventional data using advanced statistical and computational methods.
• It leverages constraint-based, score-based, and convex relaxation approaches, exploiting properties like non-Gaussianity and equal-variance assumptions for accurate structure identification.
• This methodology is critical in fields such as biomedicine, neuroscience, and social sciences, addressing challenges like latent confounding, missing data, and subsampling.

Causal graph recovery is the process of reconstructing the structure of a causal graphical model—typically a directed acyclic graph (DAG), possibly with latent or mixed edges—from observational or interventional data. Its central objective is to identify the set of directed (and possibly bidirected) edges representing the underlying causal mechanisms among a collection of variables. This task is foundational for causal inference, enabling not just prediction but also intervention and counterfactual reasoning across diverse application domains such as biomedicine, neuroscience, social sciences, and genomics. Causal graph recovery encompasses a wide spectrum of statistical, algorithmic, and computational advances, ranging from classical constraint-based and score-based algorithms to recent developments exploiting invariance, distributional or spectral properties, latent structure identifiability, regularization, and the integration of domain knowledge or external information sources.

1. Identifiability in Linear and Nonlinear Structural Equation Models

A central theme in causal graph recovery is the identifiability of the underlying DAG from the observed data distribution. In the linear acyclic structural equation model (SEM) framework, the classic result asserts that, under independent non-Gaussian errors, the entire DAG is identifiable from observational data (LiNGAM and extensions). Further, when error variances are equal across nodes, the DAG is also uniquely identifiable by minimizing total least-squares risk: for any candidate DAG, the sum of nodewise linear regression residual variances attains its minimum (exactly $p \sigma^2$) if and only if the DAG contains all true edges. This property underpins consistent and Bayesian score-based methods that explicitly exploit the equal-variance structure for DAG recovery, achieving provably consistent graph selection as sample size increases (Chaudhuri et al., 18 Sep 2025).

In the presence of latent confounding, the recoverability of the full DAG structure is typically lost; nonetheless, identifiability may be restored in specific classes of mixed graphs. For example, under a bow-free acyclic path diagram (BAP) with non-Gaussian errors, all direct and bidirected edges (representing direct effects and unobserved confounders, respectively) can be uniquely recovered from observational moments up to order $K\geq 3$. Polynomials of higher-order cumulants supply the critical information that breaks Markov equivalence, which is otherwise a hard barrier under Gaussian error assumptions (Wang et al., 2020). Under only Gaussianity and equal error variances, Cholesky factorization of the population covariance—coupled with greedy or convex-penalized search—yields consistent recovery of both the node ordering and edge structure, even enabling the detection and partial isolation of latent variables (Cai et al., 2023).

The table below organizes these core identifiability scenarios. The columns "Setting," "Key Assumption," and "Exact Recovery Possible?" summarize structurally diverse regimes.

Setting	Key Assumption	Exact Recovery Possible?
Linear SEM, Gaussian	Markov + Faithfulness	No (Markov equivalence)
Linear SEM, Equal-Variance	Homoskedastic errors	Yes (Chaudhuri et al., 18 Sep 2025)
Linear SEM, Non-Gaussian	Non-Gaussian errors	Yes (LiNGAM)
Linear SEM + latent conf	Bow-free + non-Gaussian	Yes, to BAP (Wang et al., 2020)
Nonlinear SEM (additive)	Non-Gaussian noise + ANM	Yes (ANM, see refs)

2. Methodological Advances: Algorithms and Theoretical Guarantees

Causal graph recovery algorithms are broadly classified as constraint-based, score-based, and hybrid methods, as well as newer approaches based on convex relaxations, invariance, or information theory.

• Constraint-Based: These algorithms (e.g., PC, FCI, ICD) use conditional independence tests to iteratively prune and orient edges, exploiting the Markov and faithfulness assumptions. Extensions such as Iterative Causal Discovery (ICD) tie conditioning set size to the graph distance and reduce high-order CI-test complexity, offering significant computational and statistical savings compared to classical FCI, especially in the presence of latent variables or selection bias (Rohekar et al., 2020).
• Score-Based: Methods relying on DAG scoring (e.g., BIC, Bayesian g-priors) search the space of candidate graphs. Under equal-variance linear SEM, the minimum population residual risk property is directly linked to the true DAG, and Bayesian structure learning yields consistent DAG selection (Chaudhuri et al., 18 Sep 2025).
• Cholesky/Convex Factorization: Under linearity and equal-variance Gaussian noise, Cholesky-based algorithms provide greedy and convex frameworks for structure recovery, and extend with $\ell_1$ penalties and block-finding procedures to the setting with latent variables (Cai et al., 2023).
• Higher-Order Moment and Mixture Approaches: Methods based on non-Gaussianity (e.g., BANG) and mixture-oracle reduction allow exact recovery of mixed graphs or latent DAGs and exploit polynomial cumulant structure or mixture-component factorization (Wang et al., 2020, Kivva et al., 2021).
• Invariance and Distributional Methods: Novel approaches use environment-driven or artificial interventions (downsampling or sample reweighting) to detect causal directions by checking invariance of $P(Y|X)$ under changes to $P(X)$, dramatically improving scalability and precision in the large-$d$ regime (Nguyen et al., 3 Feb 2026).
• Information-Theoretic Compression: Directed information and copula-based multi-information provide a margin-free, semi-parametric approach to causal segmentation and edge assignment, applicable to both time series and bipartite graph structure (Wieczorek et al., 2016).

The mathematical and algorithmic guarantees of these methods are sharp. For example, the finite-sample risk bound for model averaging of candidate DAGs with edge thresholding (DAGgr) is oracle-optimal up to logarithmic factors, and theoretical results establish both acyclicity preservation and strong edge-selection consistency under mild conditions (Wu et al., 18 May 2026).

3. Robustness to Latent Confounding, Missing Data, and Subsampling

Latent confounding, missingness mechanisms, and measurement subsampling are pervasive in real data, often impeding naive causal recovery.

• Latent Confounding: Mixed graph models (BAPs, ADMGs) incorporating bidirected edges represent both direct and confounded relationships. In the presence of non-Gaussianity, exact recovery of both directed and bidirected edges is achievable. For general identifiability in latent-factor time-series settings, rational identifiability criteria (e.g., latent-factor half-trek) on the algebraic spectrum encode when and how direct and indirect effects are estimable from observed data (Wang et al., 2020, Reiter et al., 2024).
• Missing Data: Formal integration of missingness graphs and extensions of score-based search or semiparametric weighting efficiently propagate MCAR, MAR, or MNAR assumptions throughout the discovery process. Modifications such as bootstrap-SEM and inverse-probability weighted scores yield consistency guarantees and enable constrained search informed by domain expertise. Empirical results confirm the improvement in out-of-sample fit and clinical plausibility when explicitly modeling non-random missingness (Zanga et al., 2023).
• Subsampling and Measurement Frequency: Undersampled time series (e.g., fMRI data) can confound the mapping between observed and latent causal graphs. Constraint optimization formulations (ASP, RnR) encode subsampling consistency and acyclicity at the strongly connected component (SCC) meta-graph level, returning equivalence classes of compatible solutions for expert adjudication, showing greater precision, recall, and robustness compared to standard baselines (Abavisani et al., 10 Jun 2025).

4. Incorporating Auxiliary Knowledge and Modern LLM-Based Approaches

Recent trends include the assimilation of external knowledge—literature, expert constraints, or curated retrievals—into the graph recovery pipeline.

• Retrieval-Augmented Generation (RAG) + LLMs: The LACR framework integrates retrieval from biomedical literature with structured LLM prompting, extracting noisily labeled association and orientation "votes." Aggregation and verification mechanisms yield undirected and oriented causal skeletons with competitive precision and recall on benchmark networks; LACR is also sensitive to new scientific evidence via vector database updates, supporting continual refinement as the field evolves (Zhang et al., 2024).
• Domain Integration: Hard or soft constraints informed by biological plausibility, anatomical maps, or experimental design can be encoded into search or optimization frameworks, ensuring expert-relevant edge inclusion/exclusion and enhancing validity in sensitive applications such as gene regulatory network recovery from Perturb-seq data (Park et al., 5 Jan 2026).

5. Applications, Empirical Evaluation, and Distance Metrics

Causal graph recovery is validated both via structural metrics (structural Hamming distance, edge precision/recall, F1, etc.) and by evaluation on downstream effect-estimation tasks.

• Benchmarks and Real Data: Leading methods are evaluated on protein-signaling networks, simulated gene networks, ecological systems, and complex time-series from neuroimaging and clinical studies. Novel evaluation metrics such as Fixing Identification Distance (FID) quantify discrepancy between candidate and reference ADMGs in terms of interventional effect estimands, offering a more meaningful assessment in the presence of latent confounding compared to traditional edit-based distances (Li et al., 28 Oct 2025).
• Model Averaging and Aggregation: Ensemble approaches such as DAGgr exploit out-of-sample predictive likelihood to weight candidate DAGs, yielding stable, acyclic, and high-fidelity aggregate graphs that match or surpass the best single method, enhancing stability and interpretability (Wu et al., 18 May 2026).
• Interventional Optimization: When interventions are costly or constrained (e.g., gene knockouts), adaptive search strategies optimize experimental design to orient a full DAG given node-dependent intervention costs, achieving polylogarithmic optimality relative to the worst-case Markov equivalence class (Choo et al., 2023).

6. Future Perspectives and Open Challenges

Causal graph recovery continues to be shaped by advances in statistical theory, optimization, and machine learning.

• Extending Identifiability: Ongoing efforts target relaxation of equal-variance and linearity assumptions, integration of heavy-tailed or non-parametric noise, and model misspecification robustness.
• Scalability: Quadratic or even sub-quadratic algorithms based on invariance, moments, or spectral properties have achieved practical recovery for networks with hundreds to thousands of nodes (Nguyen et al., 3 Feb 2026).
• Adaptive, Continual, and Hybrid Inference: The fusion of literature, experiment, and data-driven inference—augmented by LLMs and constraint optimization—signals a shift toward hybrid, continuously updatable causal graph recovery systems responsive to domain advances and emerging data streams (Zhang et al., 2024).
• Evaluation Paradigms: The introduction of estimand-based and effect-level distances reframes methodological progress towards more semantically meaningful, intervention-aware assessments, particularly in the regime of latent confounding or partial observability (Li et al., 28 Oct 2025).

The trajectory of causal graph recovery is thus driven by integrative frameworks combining statistical rigor, computational scalability, theoretical guarantees, and contextual domain knowledge to realize accurate, interpretable, and actionable causal models in complex, high-dimensional settings.