Dynamic Causal Structure

• Dynamic causal structure is the formal representation of time-evolving causal relationships in multivariate systems, capturing both instantaneous and lagged dependencies.
• It encompasses mathematical frameworks such as DSCMs, SDCMs, dynamic Bayesian networks, and nonparametric models to address temporal variations and latent confounding.
• The topic is applied across fields like biology, robotics, climate science, and economics, leveraging online tracking and adaptive algorithms for robust causal discovery.

Dynamic causal structure refers to the formal representation, identification, and estimation of temporally evolving causal relationships within multivariate systems, either in discrete or continuous time, and either in the presence or absence of instantaneous, lagged, or non-stationary dependencies. Dynamic causal models generalize static causal graphs (such as SCMs and DAGs) to account for temporally indexed mechanisms or edge sets, allowing the causal structure itself to vary as a function of time, system state, or exogenous factors. The recovery of dynamic causal structure is critical for correct causal inference in systems where interactions evolve—such as biological, physical, economic, or engineered systems—and underpins methods dealing with dynamical system identification, time-series causal discovery, and learning from streaming or non-stationary environments.

1. Mathematical Frameworks for Dynamic Causal Structure

Research on dynamic causal structure encompasses several formal frameworks, each generalizing the notion of a causal graphical model:

• Dynamic Structural Causal Models (DSCMs): The endogenous variables are stochastic processes (e.g., functions of time), with structural equations of the form $X_{v}^\mathcal{I} = f_{v^{\mathcal{I}}}(X_V, X_W)$ for evaluation block $\mathcal{I}$ and process $v$, allowing for cycles and latent confounding (Boeken et al., 2024). The dependency graph is a directed mixed graph (DMG). DSCMs support perfect interventions on trajectories or time-points, and their intervention calculus extends Pearl's do-calculus by replacing d-separation with σ-separation.
• Structural Dynamical Causal Models (SDCMs): These models represent dynamical systems as collections of (possibly stochastic) processes governed by random differential (or difference) equations of arbitrary order, with causal semantics provided through structured system dynamics and explicit intervention rules (Bongers et al., 2018). The causal graph extends to include derivatives and exogenous processes.
• Discrete-time Approaches: Time-varying structural equation models (SEMs) and dynamic Bayesian networks (DBNs) are augmented so that at each time index $t$, the causal graph $G_t$ is a DAG which may vary, and edge-weights or structures may be parameterized by basis expansions, nonparametric functions, or neural networks (Wang et al., 11 Jan 2025, Zhang et al., 2024).
• Nonparametric Dynamic Structures with Latent Variables: Nonparametric SEMs with colored latent blocks $Z_t$ driving the observed measurements $X_t$, under identifiable 3-measurement conditions, can capture arbitrarily complex, time-evolving, and functionally nonparametric dependencies (Fu et al., 21 Jan 2025). Functional equivalence principles link these to nonlinear ICA identifiability.

2. System Identification and Causal Discovery Methodologies

Dynamic causal structure discovery utilizes specialized algorithms for data-driven learning of temporally evolving mechanisms:

• Score-based and Variational Methods: Approaches such as SC3D (Das et al., 2 Feb 2026), LOCAL (Zhang et al., 2024), and dynamic basis expansion methods (Wang et al., 11 Jan 2025) employ empirical likelihoods (often with sparsity and acyclicity penalties) optimized under differentiable or combinatorial constraints, exploiting properties such as group-lasso screening, low-rank dynamic parameterization, and spectral or log-determinant acyclicity for efficiency and identifiability.
• Continuous-time and Difference-based Methods: CaDyT (Tagliapietra et al., 16 Dec 2025) applies GP priors to difference-based approximations of derivatives—robust to irregular sampling and milder in modeling assumptions—coupled with MDL/Algorithmic Markov Condition-guided greedy search for optimal dynamic DAGs, with guarantees of consistency and identifiability in the RKHS setting.
• Frequency-domain and Wiener Projection Techniques: For linear, stationary systems, FFT-based causal structure recovery leverages the phase response of partial Wiener coefficients to reconstruct dynamic graphs across frequencies, reducing complexity from combinatorial PC-type searches to O(n)-scale regressions (Veedu et al., 2023).
• Latent System Identification: For systems with latent processes, variational autoencoder or normalizing-flow-based estimators (e.g., NCDL) can nonparametrically recover both the latent Markov process and instantaneous/lagged observed graphs under minimal constraints (Fu et al., 21 Jan 2025).
• Online Tracking of Structure: Streaming algorithms (FOFCI/OFCI) maintain a current local Markov equivalence class (PAG) using recursive statistics and change-detection (via pooled Mahalanobis distances), with relearning triggered adaptively to accommodate abrupt structural shifts (Kocacoban et al., 2019).

3. Constraints, Acyclicity, and Model Stability

Enforcing valid dynamic causal graphs requires specific mathematical and algorithmic considerations:

• Acyclicity across time: For discrete models, requirement of acyclicity at each time instant (instantaneous graph) is typically enforced via algebraic constraints such as spectral radius (SC3D), orientation matrices (LOCAL), norm-scaled log-det penalties (DyCausal), or by parametrizing the graph within a finite set of basis anchors linearly interpolated over time (DyCausal) (Das et al., 2 Feb 2026, Yang et al., 26 Feb 2026, Zhang et al., 2024).
• Causal faithfulness and identifiability: Guarantees that true parent sets are not dropped are established under standard faithfulness and model well-specification assumptions, with explicit theoretical results (e.g., Theorem 3.1 in SC3D; consistency and score gap in CaDyT; monoblock identifiability in NCDL) (Das et al., 2 Feb 2026, Tagliapietra et al., 16 Dec 2025, Fu et al., 21 Jan 2025).
• Stability and efficiency: Advanced optimization schemes address gradient stability even under fast-evolving and high-dimensional graphs by introducing scaling or spectral constraint strategies (e.g., norm-scaled log-det, spectral penalty annealing), amortized over windows, and by parameter compression (low-rank factorization in LOCAL, anchor-based interpolation in DyCausal) (Yang et al., 26 Feb 2026, Zhang et al., 2024).

4. Applications, Performance, and Empirical Validation

Dynamic causal structure modeling and identification have been validated across diverse scientific and engineering domains:

• Physical and Biological Systems: Recovery of dynamic structures in Lorenz-96, Lotka-Volterra, and chaotic or nonlinear high-dimensional networks, where state-of-the-art methods (e.g., SC3D, DyCausal, CaDyT) exhibit significant performance improvements in structural Hamming distance, AUROC, and F1 scores, particularly under non-stationarity and with time-varying weights (Das et al., 2 Feb 2026, Yang et al., 26 Feb 2026, Tagliapietra et al., 16 Dec 2025).
• Robotics and Control: Dynamic structure-distribution informed planning models lead to superior robustness and adaptability to corrupted sensory input and rapid environmental regime shifts in both simulated and real robotic platforms (e.g., CADY and related architectures) (Murillo-Gonzalez et al., 8 Aug 2025, Baumann et al., 2020).
• Climate and Environmental Science: Nonparametric dynamic causal structure identification has led to interpretable recovery of both observed and unobserved drivers of climate phenomena and complex spatiotemporal feedbacks, achieving high overlap with physically meaningful variables (Fu et al., 21 Jan 2025).
• Socioeconomic and Epidemiological Analysis: Estimation of dynamic causal effects (e.g., time-varying policy efficacy in COVID transmission) is enabled via basis-expansion SVAR frameworks, providing interpretable decompositions of direct and mediated effect pathways over time (Wang et al., 11 Jan 2025).

5. Non-classical and Foundational Aspects

Dynamic causal structure encompasses situations where the causal order itself is indefinite or fundamentally variable:

• Quantum Gravity and Indefinite Causality: In quantum gravity regimes, causal structure is dynamic and possibly nonseparable; mathematically, equivalence transformations enable the use of standard entropic and correlation measures (from causally-biased theory) via data-representation corrections, thus extending quantum Shannon theory to causally nonseparable processes (Gyongyosi, 2016).
• Quantum Causal Processes: The process-matrix and supermap formalisms generalize classical causal order to quantum settings, with the dynamics of causal structure described hierarchically by higher-order maps. Continuous and reversible transformations preserve definite causal order, while non-reversible (e.g., quantum switch) can alter causal ordering, but physical realization remains unresolved (Castro-Ruiz et al., 2017).
• Dynamic Event Structures: In concurrency and distributed systems theory, event-structure models have been generalized to allow for both the addition and removal of causal dependencies at runtime by event occurrence, dramatically increasing expressive power for modeling adaptive, self-modifying workflows or protocols (Arbach et al., 2018, Arbach et al., 2015).

6. Theoretical and Computational Properties

Extensive theoretical and algorithmic analyses establish essential properties:

• Expressiveness and Subsumption: Dynamic-causality event structures (DCESs, DPESs) strictly subsume classical prime, flow, and bundle event structures, being able to encode reconfigurable causality; however, they are incomparable to resolvable-conflict event structures (Arbach et al., 2015, Arbach et al., 2018).
• Computational Complexity: Exact configuration reachability and model checking remain in PSPACE in dynamic event structures. For dynamic graph recovery in time-series, advances such as FFT-based algorithms, low-rank modeling, and anchor-based interpolation yield orders-of-magnitude speedup over classical PC or batch FCI algorithms, scaling efficiently to high-dimensional settings (Veedu et al., 2023, Zhang et al., 2024, Yang et al., 26 Feb 2026).
• Online Adaptivity: Online learning frameworks efficiently track non-stationary and possibly confounded dynamic graphs under computational, storage, and sample efficiency constraints, providing rapid adaptation to changing environments without full recomputation (Kocacoban et al., 2019).

7. Challenges and Future Directions

Key open problems and research frontiers in dynamic causal structure include:

• Scalability under Partial Observability: Extension of dynamical identifiability results to large-scale systems with latent confounding and unobserved variables remains non-trivial, especially outside the nonparametric or GP-regularized regime.
• Quantum and Indefinite-order Generalization: The extension of flow-theoretic or operator-algebraic dynamic causal models to general quantum frameworks remains unresolved, with complications due to non-commutativity and operational trace constraints (Baumeler et al., 2024, Castro-Ruiz et al., 2017).
• Integration of Causal Structure Uncertainty: Techniques that propagate epistemic uncertainty over entire structure distributions into planning and control have empirically improved robustness and sample efficiency, but theoretical characterization and extensions to continuous or quantum processes are fertile areas for further study (Murillo-Gonzalez et al., 8 Aug 2025).
• Non-equilibrium and Transient Behavior: Current DSCM and SDCM frameworks often focus on asymptotic (steady-state) behavior; formalization of transient, adaptive, or regime-switching causal structure, especially in the presence of exogenous shocks, requires further development (Rubenstein et al., 2016, Boeken et al., 2024).

Dynamic causal structure is foundational for correctly modeling, identifying, and controlling systems in which the pattern and strength of interactions evolve in time—enabling principled treatment of non-stationary, adaptive, or otherwise temporally complex phenomena across the sciences and engineering.