Abstract
Structural Causal Models (SCMs) are widely used in causal modelling, but how they relate to other modelling tools is poorly understood. In this paper we provide a novel perspective on the relationship between Ordinary Differential Equations and Structural Causal Models. We show how, under certain conditions, the asymptotic behaviour of an Ordinary Differential Equation under non-constant interventions can be modelled using Dynamic Structural Causal Models. In contrast to earlier work, we study not only the effect of interventions on equilibrium states; rather, we model asymptotic behaviour that is dynamic under interventions that vary in time, and include as a special case the study of static equilibria.
Introduction
Bridging the gap between low-level physical descriptions and high-level causal reasoning is a central challenge in machine learning. Ordinary Differential Equations (ODEs) provide a precise language for deterministic physical systems, describing instantaneous changes. Conversely, Structural Causal Models (SCMs) offer a framework for reasoning about interventions and counterfactuals. While previous work has linked these two by deriving SCMs from the static equilibria of ODEs, many real-world systems—from biological oscillators to economic cycles—never reach a constant state. This paper introduces Dynamic Structural Causal Models (DSCMs) to capture the time-varying asymptotic behavior of such systems.
The Challenge of Dynamic Equilibria
Traditional SCM derivations from ODEs rely on the assumption that the system settles into a static equilibrium. If a system is described by $\dot{x} = f(x, u)$, a static equilibrium is a point $x^*$ where $f(x^*, u) = 0$. However, many systems exhibit periodic or quasi-periodic behavior. For example, a damped mass-spring system driven by an external force will eventually oscillate at the frequency of that force.
To model this, we need to extend the concept of an intervention. In standard Pearlian do-calculus, an intervention $do(X=c)$ clamps a variable to a constant. In the dynamic setting, we consider Time-Dependent Perfect Interventions, where a variable is forced to follow a specific trajectory $\zeta(t)$, denoted as $do(X(t) = \zeta(t))$.
Dynamic Structural Causal Models (DSCMs)
The authors propose DSCMs as an abstraction that integrates out transient dynamics. In a DSCM, the variables are not scalar values at a point in time, but entire trajectories over a time interval.
Structural Dynamic Stability
A key requirement for this abstraction is Structural Dynamic Stability. A system is dynamically stable if, for any initial condition, it converges to a unique trajectory. Structural dynamic stability extends this: the system must remain stable even when arbitrary subsets of variables are subjected to interventions from a predefined set of trajectories.
As shown in the figure above, a physical system like a mass-spring chain (a) can be represented by a causal graph $\mathcal{D}$ (b). Under an intervention on $X_1$, the graph is modified to $\mathcal{D}_{do(X_1=\zeta_1)}$ (c), where the incoming edges to $X_1$ are removed, reflecting the external control over its trajectory.
Key Insights: Abstraction and Commutativity
One of the most significant theoretical contributions is the proof that interventions commute with the abstraction from ODEs to DSCMs. Specifically, if one intervenes on an ODE and then observes its long-term behavior, the result is identical to intervening directly on the derived DSCM. This justifies using DSCMs for high-level causal reasoning without needing to simulate the underlying differential equations at every time step.
Furthermore, DSCMs avoid the computational pitfalls of discretized time-series models like Dynamic Bayesian Networks (DBNs). By mapping parent trajectories directly to child trajectories, the model focuses purely on the steady-state causal relationships.
Results and Case Study
The framework was validated using a coupled mass-spring system. When subjected to periodic forcing, the system's state converges to a dynamic equilibrium.
The plot above illustrates this convergence. Regardless of the initial transient behavior (the high-variance part of the plot for $t < 40$), the system eventually settles into a stable oscillatory pattern. The DSCM successfully captures the relationship between the amplitudes and phases of these oscillations. The authors derived closed-form structural equations for this system, demonstrating that the framework is mathematically tractable for non-trivial physical models.
Limitations and Future Outlook
While DSCMs provide a powerful abstraction, they currently have several constraints:
- Uniqueness of Attractors: The theory assumes a single global attractor. Systems with multiple stable states or chaotic behavior are not yet covered.
- Deterministic Nature: The current work focuses on deterministic ODEs. Extending this to stochastic differential equations is a necessary step for handling real-world noise.
- Learning from Data: This paper establishes the mathematical foundation. Future research is needed to develop algorithms that can learn DSCM structures and parameters directly from observed time-series data.
By moving beyond static equilibria, DSCMs open the door to causal analysis in a much broader class of dynamical systems, providing a rigorous link between the physics of motion and the logic of causation.
Summary updated