CausalStruct: Unified Causal Framework

• CausalStruct is a unifying framework that encompasses diverse formalisms, from DAGs and latent graphs to quantum and categorical causal models.
• It systematically separates causal content from mere correlations, enabling robust interventions, counterfactual analysis, and insights into structural invariants.
• Its versatile representations support applications in statistical discovery, deep learning disentanglement, temporal modeling, and 3D scene generation while addressing identifiability challenges.

In the literature surveyed here, “CausalStruct” functions as an umbrella label for a family of formalisms that represent, learn, manipulate, or interpret causal structure across machine learning, statistics, quantum theory, category theory, and mathematical physics. Depending on the setting, causal structure appears as a directed acyclic graph (DAG), a maximal ancestral graph (MAG), a partial ancestral graph (PAG), a process matrix, a causal set, a simplicial or categorical object, or a field of projective hypersurfaces. What unifies these uses is the attempt to separate causal content from purely correlational description, to support interventions and counterfactuals, and to make explicit which aspects of a model are invariant under structural transformations or changes of environment (Squires et al., 2022, Kissinger et al., 2017, Wuthrich, 2012, Makhmali, 2017).

1. Core representations of causal structure

A recurring starting point is the structural causal model. In the graph-theoretic setting, a DAG $G$ on nodes $[p] = \{1,\dots,p\}$ encodes direct causes via directed edges, and a Markovian structural causal model associates each variable with a structural assignment $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$, with mutually independent exogenous noises. Under this setup, the induced joint distribution factorizes according to the graph, and causal discovery asks which graph, or equivalence class of graphs, is compatible with observational or interventional data. When latent confounding is allowed, DAGs are replaced by MAGs and their equivalence classes by PAGs, with m-separation replacing d-separation (Squires et al., 2022).

A second recurring representation is the latent causal graph used in representation learning. In "CausalVAE" the latent variables are split into independent exogenous factors $u$ and endogenous causal factors $z$, related by a structural causal model implemented as a Causal Layer. The adjacency matrix $A \in \mathbb{R}^{d \times d}$ encodes edges among latent concepts, and acyclicity is enforced by constraints such as $h(A)=\operatorname{tr}(\exp(A)) - d = 0$ or related NOTEARS-style variants. This turns disentanglement from an independence-only problem into one of structured causal disentanglement (Yang et al., 2020).

A third representation treats causal structure as a constraint on allowable probability models rather than as a single sharp graph. "Structural Causal Models Are (Solvable by) Credal Networks" shows that endogenous observations induce linear constraints on the probabilities of the exogenous variables, yielding an exact mapping from a structural causal model into a credal network. In identifiable cases this returns sharp values; in unidentifiable cases it returns exact bounds as intervals (Zaffalon et al., 2020).

Representation	Role	Representative sources
DAG, CPDAG, MAG, PAG	Direct causes, equivalence classes, latent confounding	(Squires et al., 2022, Heinze-Deml et al., 2017, Chen et al., 2021)
Latent causal graph with Causal Layer	Causal disentanglement and counterfactual generation	(Yang et al., 2020)
Credal network	Exact bounds for interventional and counterfactual queries	(Zaffalon et al., 2020)
Process matrix, causal category, Caus$[\mathcal{C}]$	Quantum and higher-order causal constraints	(Feix et al., 2016, Coecke et al., 2011, Kissinger et al., 2017)
Causal set or geometric field	Physical and structural interpretations of causality	(Wuthrich, 2012, Makhmali, 2017)

2. Causal structure learning from data

In statistical causal discovery, the central issue is not merely graph estimation but identifiability under explicit assumptions. Under faithfulness, observational data identifies a Markov equivalence class rather than a unique DAG; this class is represented by a CPDAG. Interventional data refines the equivalence class further. The review "Causal Structure Learning: a Combinatorial Perspective" emphasizes two combinatorial themes: the structure of the search space over causal graphs and the structure of equivalence classes, including CPDAGs for DAGs and PAGs for latent-variable models. It also records the super-exponential growth of DAGs and the role of covered edge reversals, imsets, and minimal I-MAPs in navigating this space (Squires et al., 2022).

Algorithmically, the surveyed toolbox includes constraint-based, score-based, hybrid, and invariance-based methods. "Causal Structure Learning" discusses PC and rankPC for causally sufficient observational settings, FCI and RFCI for latent confounding, GES and GIES for score-based learning over equivalence classes, MMHC as a hybrid procedure, LiNGAM for non-Gaussian linear models, and BACKSHIFT for unknown shift interventions and cyclic systems. The same survey stresses that algorithm choice is inseparable from assumptions about latent confounding, acyclicity, intervention type, noise model, and sparsity (Heinze-Deml et al., 2017).

High-dimensional latent-variable discovery motivates more localized notions of structure. "Causal Structural Learning Via Local Graphs" introduces local FCI (lFCI), which replaces degree-based sparsity by a local-separation property. This allows hub nodes while retaining polynomial complexity $O(p^{\eta+2})$ under the stated assumptions, and yields PAGs in the presence of latent confounders and selection bias. The same work makes explicit that local separators may fail in dense graphs and that completeness depends on a local discriminating paths assumption (Chen et al., 2021).

Nonstationary data provide a different route to structure. "Causal Discovery from Changes" formalizes local mechanism changes through transition sequences and proves that, for focal variables, change-based equivalence classes refine standard Markov equivalence by fixing parent sets. The same framework derives order graphs and marked order graphs from patterns of detected marginal changes, while also warning that influentiality failures, latent confounding, or simultaneous changes weaken identifiability (Tian et al., 2013).

Recommendation systems supply a case where interventions are neither absent nor known. "Causal Structure Learning with Recommendation System" models RS feedback as a mixture of a causal mechanism and an unknown intervention mechanism generated by the recommender. The graph is learned through an augmented Lagrangian objective with acyclicity constraint $$h(\Gamma)=\operatorname{tr}(\exp(\sigma(\Gamma))) - d$$, and empirical comparisons are reported on Amazon Electronics, Walmart Electronics, and synthetic data. A notable feature of this formulation is that the unknown intervention is not treated as nuisance noise but as a competing mechanism to be learned jointly with the causal graph (Xu et al., 2022).

3. Latent, temporal, and dynamic variants

Structured latent-variable models use causal structure to go beyond factorized priors. In "CausalVAE: Structured Causal Disentanglement in Variational Autoencoder," the generative process takes independent exogenous variables $[p] = \{1,\dots,p\}$0, maps them through a structural causal model into endogenous causal latents $[p] = \{1,\dots,p\}$1, and then decodes to observations. The framework supports supervision for semantic alignment, learns a DAG over latent concepts, analyzes identifiability under additive-noise assumptions and supervision, and generates counterfactual data through do-operations on latent factors. The paper’s formulation therefore connects representation learning, causal graph learning, and counterfactual generation within a single VAE-based architecture (Yang et al., 2020).

Temporal event data require further extensions because causal structure may include contemporaneous effects. "Structural Hawkes Processes for Learning Causal Structure from Discrete-Time Event Sequences" introduces discrete-time Hawkes models with an explicit instantaneous term $[p] = \{1,\dots,p\}$2, alongside lagged effects $[p] = \{1,\dots,p\}$3. The paper argues that instantaneous effect is a blessing rather than a curse: in low-resolution data it breaks symmetries that make lag-only Granger-causality approaches unidentifiable. Learning proceeds by a sparse likelihood objective with an MM algorithm and hill-climbing structure search (Qiao et al., 2023).

Longitudinal systems whose qualitative regimes change over time motivate path-dependent structure. "Path Dependent Structural Equation Models" defines PDSEMs as families of state-specific DAGs or CDAGs indexed by discrete states, together with state-transition mechanisms. The key point is that interventions may counterfactually alter which state-specific graph applies later, so the temporal evolution of the system itself becomes path-dependent. Identification is given by state-specific truncated g-formulas in the fully observed case and by nested Markov factorizations under hidden variables (Srinivasan et al., 2020).

A related but graph-theoretic notion of temporal evolution appears in "Flow of dynamical causal structures with an application to correlations." There, a flow is a rooted directed graph whose nodes are causal structures obtained by sequential source reductions, while the superflow is a parameter-free outer approximation computed from the initial structure alone. The paper proves that if all leaves of a flow are trivial, then the corresponding process produces causal correlations only, and notes that the superflow may contain unrealizable transitions (Baumeler et al., 2024).

Testing structural priors rather than discovering them from scratch is another dynamic use of CausalStruct. "Causal Structural Hypothesis Testing and Data Generation Models" encodes a hypothesized graph as a masking layer inside node-wise neural networks and compares alternative structural priors by deliberate out-of-distribution splits. Its variational extension, CSVHT, adds a latent-variable layer and an ELBO-style objective. The paper’s empirical message is that out-of-distribution generalization error can function as a proxy for structural-prior hypothesis testing, and that lossy hypotheses are reliably worse than leaky ones (Jiang et al., 2022).

4. Quantum and categorical formulations

Quantum causal structure generalizes graph-based causality in two distinct directions: one retains definite order but allows quantum resources, while the other studies higher-order processes where order itself can become indefinite. "Quantum superpositions of 'common-cause' and 'direct-cause' causal structures" studies the first direction in the process-matrix formalism with fixed order $[p] = \{1,\dots,p\}$4. It distinguishes direct-cause processes, common-cause processes, their classical convex hull, and coherent superpositions between the two. Nonclassicality is certified by an SDP witness, and a concrete qubit realization yields robustness values $[p] = \{1,\dots,p\}$5, $[p] = \{1,\dots,p\}$6, and $[p] = \{1,\dots,p\}$7 under progressively restricted measurements. The same paper emphasizes that this resource is distinct from the quantum switch, because it superposes types of causation rather than causal orders (Feix et al., 2016).

The second direction appears in the sectorial-constraint analysis of the quantum switch. "Causal structure in the presence of sectorial constraints, with application to the quantum switch" introduces sectorizations of Hilbert spaces and routed maps that respect sector-dependent block structure. Within this framework, several intuitions about no-influence are shown to be equivalent, quantum causal structures can be represented by directed graphs, and a fine-graining becomes possible in which individual sectors of a system are the causal relata. Applied to photonic implementations of the quantum switch, the conclusion is that the coarse-grained causal structure is cyclic whereas the fine-grained causal structure is acyclic; the experiments therefore realize indefinite causal order only in a weak sense (Ormrod et al., 2022).

Category theory supplies a more abstract semantics. "Causal categories: relativistically interacting processes" identifies causal dependencies with topological connectedness in a symmetric monoidal graphical calculus, shows that correlations force terminality of the tensor unit, and argues that well-definedness of global states forces the monoidal product to be only partially defined. The resulting causal category, or caucat, encodes causal structure by allowing tensor products only for spacelike-separated objects and by treating connectedness as information flow (Coecke et al., 2011).

"A categorical semantics for causal structure" develops a related but distinct construction, Caus$[p] = \{1,\dots,p\}$8, from a precausal category. Its objects are pairs $[p] = \{1,\dots,p\}$9 of systems and admissible state-sets, and the resulting category is *-autonomous rather than compact closed. This makes it possible to type non-signalling processes, semi-causal processes, quantum $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$0-combs, and the higher-order families $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$1. In this semantics, non-signalling bipartite processes have type $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$2, whereas all causal bipartite processes inhabit the corresponding par type. The same framework accommodates process matrices and the quantum switch as higher-order causal processes (Kissinger et al., 2017).

5. Physical and geometric interpretations

Outside machine learning and quantum information, CausalStruct also names explicitly structural conceptions of causality in fundamental physics. In causal set theory, a causal set is the relational structure $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$3 with $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$4 a partial order satisfying local finitarity. "The structure of causal sets" argues that the physically relevant content of a model is its isomorphism class rather than its labels, and uses causal sets as a paradigmatic case for a structuralist interpretation of physical theory. The same paper treats discrete general covariance, automorphisms, faithful embedding into continuum spacetime, and Poisson sprinkling as central to the relation between the discrete order and relativistic spacetime (Wuthrich, 2012).

A differential-geometric meaning of causal structure appears in "Differential Geometric Aspects of Causal Structures." There, a causal structure on an $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$5-manifold is a field of tangentially non-degenerate projective hypersurfaces in the projectivized tangent bundle. Cartan’s method produces an $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$6-structure on an appropriate principal bundle, identifies the essential local invariants as the Fubini cubic form and the Wsf curvature, and relates the theory to parabolic geometries of type $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$7, $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$8, and, in dimension three, $X_i = f_i(X_{\mathrm{Pa}_G(i)}, \epsilon_i)$9. Special subclasses include lifts of pseudo-conformal structures and causal structures with vanishing Wsf curvature (Makhmali, 2017).

A still more abstract synthesis is provided by "A Layered Architecture for Universal Causality." UCLA organizes causality into four categorical layers: interventions via the simplicial category $u$0, causal models as graph- or string-diagram categories, data as Set-valued functors and Grothendieck categories of elements, and evaluation via nerves and homotopy colimits. In this formulation, causal inference between layers is cast as a lifting problem, and causal effect is defined as the homotopy colimit of the nerve of the category of elements (Mahadevan, 2022).

6. Applications, evaluation, and recurring limitations

Recent applications extend CausalStruct into domains where causal structure is not only learned or interpreted but operationally deployed. "Navigating the Impact of Structured Output Format on LLMs through the Compass of Causal Inference" reduces the effect of structured output on LLM generation to five candidate causal structures: collider without m-bias, collider with m-bias, single cause from instruction, single cause from output format, and independence. Across seven public and one developed reasoning tasks, the study reports no causal impact in 43 out of 48 GPT-4o scenarios, with the remaining cases split among multifaceted collider structures and single-cause cases (Yuan et al., 26 Sep 2025).

In 3D generation, "Causal Reasoning Elicits Controllable 3D Scene Generation" uses CausalStruct to build causal scene graphs from text or text-plus-image inputs, infer causal order for object placement, refine uncertain edges by intervention, and then adjust object position and scale with a PID controller. The representation is 3D Gaussian Splatting, and optimization uses object-level and scene-level Score Distillation Sampling. The reported effect is improved logical coherence, realistic spatial interactions, and controllable editing relative to 3DGS-based baselines (Chen et al., 18 Sep 2025).

Across these literatures, several limitations recur. In graph-based learning, identifiability depends on assumptions such as faithfulness, sparsity, additive noise, local-separation properties, or the availability of interventions (Squires et al., 2022, Chen et al., 2021). In latent and generative models, the learned structure may depend sensitively on supervision anchors, SCM assumptions, and hyperparameter balancing between reconstruction, sparsity, and acyclicity penalties (Yang et al., 2020, Xu et al., 2022). In temporal and event-sequence models, stationarity, stability, or finite-state assumptions constrain applicability, and outer approximations such as superflows may contain unrealizable branches (Qiao et al., 2023, Baumeler et al., 2024). In quantum and categorical settings, the admissible local operations may be restricted by sectorial constraints, and strong indefinite causal order may disappear under finer causal analysis (Ormrod et al., 2022). These limitations do not negate the utility of CausalStruct; they instead show that causal structure is always defined relative to a modeling language, an intervention regime, and a notion of admissible transformation.

Taken together, the surveyed work suggests that CausalStruct is less a single theory than a common research program: to encode causal constraints explicitly enough that interventions, counterfactuals, equivalence classes, and structural invariants can be reasoned about inside the model rather than imposed after the fact. This suggests a unifying methodological principle rather than a universal formalism. Whether the setting is a VAE with latent DAGs, a PAG under latent confounding, a routed quantum process, a causal set, or a projective hypersurface field, the same question recurs: which part of a system’s behavior is fixed by causal structure itself, and which part is contingent on parameterization, representation, or observation regime.