Causal-Dynamical Organisation Under Intervention

• Causal-dynamical organisation under intervention is defined by the structured repertoire of responses a system exhibits when interventions actively reconfigure its local causal mechanisms.
• It spans diverse frameworks—from semi-Markovian Bayesian networks to differential kinetic and action models—emphasizing modular mechanisms and local intervention effects.
• The topic offers practical insights into identifying, testing, and recovering causal organization through rigorous interventional semantics and dynamic response analysis.

Causal-dynamical organisation under intervention denotes a family of closely related ideas in which a system is identified not merely by an observational distribution or a static dependency pattern, but by the structured repertoire of responses it exhibits under intervention. Across semi-Markovian causal Bayesian networks, structural causal dynamical models, causal kinetic models, state-space action models, and intervention logics, the common commitment is modularity: a system is represented as a collection of local mechanisms, and interventions reveal or reconfigure organization by replacing some mechanisms, fixing some processes, or altering which information is transmitted between components (Acharya et al., 2018, Bongers et al., 2018, Peters et al., 2020, Cohen, 2022).

1. Representational forms of causal-dynamical organisation

Several distinct formalisms instantiate this idea.

In semi-Markovian causal Bayesian networks with latent confounding, the basic representational object is a causal Bayesian network

$$\mathcal{M}=\left\langle \mathbf V,\mathbf U,G,\{[V_i\mid \bm{\pi}(V_i)]\},\{[U_i\mid \bm{\pi}(U_i)]\}\right\rangle,$$

where hidden common causes are represented by bidirected edges and induce a decomposition into c-components. In this setting, causal equality is defined interventionally: two models are equal iff they agree on all interventional distributions, and they are $\epsilon$-far iff there exists some intervention under which their induced distributions differ by more than $\epsilon$ in total variation distance (Acharya et al., 2018). This makes “organization” a property of the family of all interventional kernels rather than of a single observational law.

In continuous-time causal modeling, organization is encoded by local differential mechanisms and initial conditions. A deterministic causal kinetic model is specified by ODEs

$\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$

and a stochastic causal kinetic model by SDEs

$\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$

Here causal structure is not just a graph: it is the parent sets, the local mechanisms $f^k$ and $h^k$, the initial values, and the coupled law of time evolution they induce (Peters et al., 2020). Structural Causal Dynamical Models formalize the same idea in a general stochastic-process setting by replacing static SCM variables with endogenous stochastic processes and their derivatives, while keeping the SCM principle that each component has its own autonomous mechanism (Bongers et al., 2018).

A third line of work grounds organization in admissible transformations of state space. In the embodied framework of action models, the primitive intervention-like object is a map

$do_X(a):X\to X,$

typically induced by running a policy, option, or skill, together with a process map $proc_Y:X\to Y$ and outcome maps

$outcome_Y^a = proc_Y\,do_X(a).$

Variables arise only after factorizing the outcome space $\epsilon$0, and mechanisms are defined as invariant predictors that survive families of later actions (Cohen, 2022). This suggests a broader notion of causal-dynamical organization in which the relevant modular units are not only variables and equations but also stable action-relative transformations.

Framework	Basic object	Organizational unit
Semi-Markovian CBN	Interventional family on a known graph	c-components and local kernels
Causal kinetic / SCDM	Processes, derivatives, ODE/RDE/SDE mechanisms	Component-wise autonomous mechanisms
Action model	State transformations and outcome maps	Invariant predictors across actions

2. Intervention semantics and modularity

The canonical intervention semantics is Pearl’s $\epsilon$1. In the semi-Markovian Bayesian-network setting, for $\epsilon$2 and assignment $\epsilon$3, the intervention $\epsilon$4 forces the target variables and modifies the generative mechanism rather than conditioning on an event. The induced distribution is written

$\epsilon$5

This distinction is fundamental: observational equivalence does not imply causal equivalence, because two models may agree observationally yet disagree under some intervention (Acharya et al., 2018).

In continuous time, interventions are likewise defined as mechanism replacements. For deterministic causal kinetic models one may intervene by replacing an initial condition,

$\epsilon$6

or by replacing a differential law,

$\epsilon$7

For stochastic causal kinetic models one may similarly replace

$\epsilon$8

The same modularity principle appears in Structural Causal Dynamical Models: a stochastic perfect intervention $\epsilon$9 replaces the targeted endogenous processes $\epsilon$0 by externally assigned processes $\epsilon$1, while all untargeted component mechanisms remain invariant (Peters et al., 2020, Bongers et al., 2018).

Other frameworks refine or generalize this semantics. In causal teams, the interventionist counterfactual operator is evaluated by first constructing the post-intervention causal team $\epsilon$2, so that

$\epsilon$3

whereas selective implication

$\epsilon$4

is only observational restriction and does not alter graph or equations (Barbero et al., 2019). In “info intervention”, by contrast, the structural equation itself is preserved while the information passed from an intervened variable to its descendants is replaced: $\epsilon$5 The paper explicitly contrasts this with Pearlian surgery: $\epsilon$6-intervention intervenes the causal mechanisms, whereas info intervention intervenes the input/output information of causal mechanisms (Heyang et al., 2019). A physically oriented critique goes further and argues that perfect “atomic” or “surgical” interventions are physically impossible, because real interventions require measurement and feedback in open systems and are constrained by thermodynamics in both the classical and quantum cases (Milburn et al., 2018).

3. Local mechanisms, decomposition, and compositional structure

A recurring theme is that global organization becomes tractable only when it decomposes into local modules.

In semi-Markovian Bayesian networks, the decisive structural unit is the c-component, a maximal set of observables linked by bidirected paths. If $\epsilon$7, then under intervention the global distribution factorizes as

$\epsilon$8

A further localization result states that if $\epsilon$9, then

$\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$0

Thus interventions outside a module matter only through interventions on its observable parents. The organizational content of the model is therefore a library of local interventional kernels $\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$1 plus a composition rule that reassembles them into global interventional behavior (Acharya et al., 2018).

In Structural Causal Dynamical Models, decomposition is likewise local but takes a different form. The endogenous process $\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$2 is partitioned into components $\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$3, and the full derivative tuple $\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$4 belongs to the same causal component rather than to separate autonomous variables. The associated graph therefore uses clusters of derivative nodes per component, with cross-cluster directed edges for genuine functional parentage and within-cluster dashed edges for derivative or integration relations (Bongers et al., 2018). This preserves the intuition that organization is attached to subsystems, not to isolated derivatives.

In the embodied action framework, compositionality appears at the level of actions and invariant determinations. Actions compose sequentially,

$\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$5

and mechanisms are represented by determinations such as

$\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$6

that remain invariant under later actions $\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$7,

$\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$8

The paper’s phrase “mechanism as an invariant predictor” formalizes organization as those predictive relations that survive transformation families; interventions then probe the boundaries of those invariances (Cohen, 2022). This suggests that causal-dynamical organization can be understood either as modular factorization of interventional kernels or as persistence of local determination maps under structured transformation.

4. Identifiability, testing, and recovery of organization

One major research program asks when interventionally defined organization can be tested or learned efficiently.

For semi-Markovian Bayesian networks on a known graph with bounded in-degree $\dot{x}_t^k := f^k(x_t^{\PA(k)}), \qquad x_0^k := \xi_0^k,$9, bounded c-component size $\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$0, and alphabet size $\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$1, causal two-sample testing admits an algorithm using

$\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$2

interventions on each network and

$\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$3

samples per intervention. For learning on known $\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$4, there is an improper learner using

$\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$5

interventions and returning an oracle $\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$6 whose response to every intervention is within $\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$7 in total variation. The key technical tool is a subadditivity inequality for squared Hellinger distance: if every local c-component kernel is Hellinger-close, then every global interventional distribution is Hellinger-close, with

$\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$8

The same paper proves a matching lower bound: $\mathrm{d}X_t^k := f^k(X_t^{\PA(k)})\mathrm{d}t + h^k(X_t^{\PA(k)})\mathrm{d}W_t^k, \qquad X_0^k := \xi_0^k.$9 interventions are necessary in general, even adaptively (Acharya et al., 2018).

Restricted dynamical classes can be more sharply identifiable. In directional chain-reaction systems modeled by a directed tree, blocking interventions $f^k$0 reveal downstream organization through

$f^k$1

This yields exact graph identification by transitive reduction of the ancestor matrix, with a minimal estimator whose false-positive probability decays exponentially and whose sample complexity is logarithmic in the number of objects. Empirically, the reported main finding is that 1–2 blocking interventions per object suffice for reliable exact recovery across six physics-based chain-reaction environments (Panayiotou et al., 23 Mar 2026).

For mixtures of latent DAG-governed regimes, observational inseparability is no longer a reliable indicator of direct causal adjacency. The identification target becomes the set of true edges, meaning edges present in at least one component DAG. The paper establishes matching necessary and sufficient intervention-size bounds: in the worst case, deciding whether $f^k$2 requires and suffices to use an intervention of size

$f^k$3

and for mixtures of directed trees the corresponding quantity is $f^k$4, where $f^k$5 is the number of components. The adaptive CADIM algorithm then recovers all true edges using $f^k$6 interventions, with the excess over optimal intervention size controlled by the cyclic complexity number $f^k$7 (Varıcı et al., 2024).

Interventional notions can sometimes be reconstructed from observational trajectories alone. “Interventional Dynamical Causality” defines causation in terms of whether perturbing $f^k$8 changes $f^k$9, then uses delay embedding to derive

$h^k$0

for infinitesimal perturbations, and quantifies the retained intervention information by

$h^k$1

This does not use actual interventions, but the paper argues that it reconstructs an intervention-relevant causal notion from passive time series. It reports strong performance on simulated systems and on C. elegans, COVID-19 transmission networks in Japan, and circadian gene regulation (Shi et al., 2024).

For contagious outcomes, identification becomes explicitly dynamical because interference is activated only after infection. In the two-person partnership model, infection time is decomposed as

$h^k$2

and controlled as well as natural contagion, susceptibility, and infectiousness effects are identified under stated exchangeability, consistency, and positivity assumptions. The key insight is that contagion induces a regime change: before first infection, there is no within-pair interference; after infection, a transmission pathway becomes active (Cai et al., 2019).

5. Temporal propagation, equilibrium, and intervention optimization

Not all treatments of interventional organization are explicitly temporal, but several are.

Continuous-time causal kinetic models insist that causal analysis should often target full trajectories rather than only equilibria. The framework was introduced partly because interventions in differential-equation systems are most naturally formulated as differential equations themselves, and because transient behavior, oscillations, and rates of convergence may differ even when equilibria coincide (Peters et al., 2020). Structural Causal Dynamical Models complement this by giving conditions under which equilibrating dynamical systems reduce to equilibrium SCMs. If an SCDM is steady and a solution equilibrates, then the equilibrium state $h^k$3 solves the equilibrated SCM $h^k$4; moreover, for steady stochastic perfect interventions,

$h^k$5

so intervention and equilibration commute (Bongers et al., 2018). This provides a precise bridge between trajectory-level organization and static equilibrium causal semantics.

A discrete-time analogue is developed for stochastic processes governed by

$h^k$6

with interventions beginning at time $h^k$7. The paper defines a time-indexed causal effect

$h^k$8

and shows, for stable VAR($h^k$9) processes, that a long-run transformed process can be represented by a linear SCM that may be cyclic and confounded. This yields a practical “causal VAR” framework in which additive interventions preserve the transition operator, whereas forcing interventions modify it and may even destabilize the system (Cinquini et al., 2024).

A separate line of work studies how to choose interventions sequentially. Dynamic Causal Bayesian Optimization defines, at each time $do_X(a):X\to X,$0,

$do_X(a):X\to X,$1

and uses a dynamic causal GP prior whose mean is derived from a time-operator theorem. The framework is explicitly myopic—optimizing the current target $do_X(a):X\to X,$2 conditional on already chosen interventions—yet it formalizes how past interventions reshape later intervention-response surfaces through the rolled-out causal graph (Aglietti et al., 2021).

Graph-coupled causal Bayesian optimization pushes this further by coupling different intervention effects through shared identifiable parameters $do_X(a):X\to X,$3. If

$do_X(a):X\to X,$4

then uncertainty transfer across interventions is induced by the kernel

$do_X(a):X\to X,$5

whose rank is bounded by $do_X(a):X\to X,$6 rather than by the size of the intervention menu. The resulting regret bound separates optimization error, causal-estimation error, and intervention-family choice error (Javidian, 31 May 2026). This suggests that intervention-response organization itself may have a low-dimensional latent structure even when the admissible intervention family is combinatorially large.

6. Domains, applications, and limitations

The empirical and methodological scope of this literature is broad. Chain-reaction systems treat organization as monotone suppression structure over binary activation events (Panayiotou et al., 23 Mar 2026). Embodied AI treats actions, options, and skills as state-space transformations and asks when they qualify as surgical interventions at an abstract level (Cohen, 2022). Personalized psychological treatment proposals combine causal discovery from intensive longitudinal data, conversion of causal effects into a nonlinear dynamic Markov process, and simulation of intervention trajectories to choose treatment focus for individual patients (Waldorp et al., 8 Jun 2026). Observational recovery of intervention-relevant organization has been demonstrated on neural connectomes of C. elegans, COVID-19 transmission networks in Japan, and circadian gene systems (Shi et al., 2024).

Several limitations recur. Many formal guarantees require known graphs, finite discrete alphabets, bounded in-degree, bounded c-component size, or identifiability of the queried interventional distributions (Acharya et al., 2018, Javidian, 31 May 2026). Continuous-time causal kinetic models are introduced as a starting point and explicitly leave many issues open, including adjustment results, do-calculus, effects of hidden variables, and structure learning (Peters et al., 2020). The personalized treatment framework is explicitly methodological and simulation-based; its nonlinear dynamics are not fitted to data, and uncertainty quantification is limited (Waldorp et al., 8 Jun 2026). In the embodied framework, all actions are state transformations, but not all actions count as surgical interventions; surgicality is relational and not given by a complete necessary-and-sufficient theorem (Cohen, 2022).

Two misconceptions are addressed repeatedly. First, observational equivalence is not causal equivalence: interventional distinguishability is strictly richer than observational distinguishability, especially under latent confounding (Acharya et al., 2018). Second, “dynamical” is not always synonymous with explicit temporal dynamics. In some papers the relevant dynamic aspect is behavior under manipulation rather than time-indexed evolution; this is explicit in work that defines organization through families of interventional distributions or action-induced transformations rather than through differential equations (Acharya et al., 2018, Cohen, 2022).

A final controversy concerns the idealization of intervention itself. The physical analysis of classical and quantum interventions argues that perfect Pearlian atomic or surgical interventions are physically impossible: actual interventions require measurement and feedback in open systems and are constrained by thermodynamics, uncertainty, and entanglement (Milburn et al., 2018). This suggests that causal-dynamical organisation under intervention is often an approximate and resource-dependent notion. Even so, the literature converges on a robust core thesis: causal organization is most faithfully characterized by the modular mechanisms that persist across contexts, and intervention is the operation that reveals, tests, or selectively reconfigures those mechanisms.