Trajectory-Aware Node Contributions and the Limits of Static Controllability

Technical Summary: Trajectory-Aware Node Contributions and the Limits of Static Controllability

Problem Statement
A recurring challenge in data mining for complex, time-evolving networks is identifying which individual nodes (variables) are dynamically consequential—specifically, which components' perturbations propagate most strongly through a fitted system. Existing approaches fall into two categories, both of which have limitations for empirically estimated systems:

1. Static Network Measures: Graph-theoretic centralities (e.g., degree, betweenness, PageRank) characterize structural position within a fixed adjacency matrix but fail to capture dynamical influence. Two systems with identical topology but different dynamical laws yield identical centralities, even if one node is highly influential in one system and inert in the other.
2. Static Controllability Measures: Metrics like the controllability Gramian, average controllability, and modal controllability quantify dynamical influence but rely on the assumption that the system is linear and time-invariant (LTI), governed by $\dot{x} = Ax + Bu$. However, systems estimated from data (e.g., via neural vector autoregressions, neural ODEs, or nonlinear state-space models) are typically nonlinear and time-varying. Their local dynamics depend on the state, and governing relationships may drift. Despite this, controllability-based quantities are often applied under the LTI approximation without validation of its validity.

The core question addressed is: When do node rankings implied by static controllability remain reliable, and when do they fail for nonlinear, time-varying systems?

Methodology: Emergent Contribution ($E_j$)
The authors define Emergent Contribution ($E_j$), a finite-horizon measure of a node's dynamical leverage. It is constructed from the Jacobians of any differentiable estimated model, making it estimator-agnostic.

• Definition: For a discrete-time system $x_{t+1} = f_t(x_t) + \epsilon_t$, the local causal influence is the Jacobian $A_t(x) = \frac{\partial f_t}{\partial x}$. The state-transition matrix $\Phi^\tau_t(x)$ is the ordered product of Jacobians along a trajectory. The emergent contribution of component $j$ over a horizon $T$ is the accumulated metric-weighted energy of its impulse response:
  $$E_j(t, x; T) = \sum_{\tau=0}^{T-1} (\Phi^\tau_t(x) e_j)^\top M (\Phi^\tau_t(x) e_j)$$
  where $e_j$ is the perturbation vector for node $j$, and $M$ is a positive-definite metric (default $M=I$ on standardized components).
• Computation: The measure propagates a unit perturbation through the sequence of Jacobians along a simulated or observed trajectory, accumulating the squared norm of the resulting state vectors. It requires only that the underlying estimator be differentiable.
• Gauge and Scale: The authors emphasize that $E_j$ is scale-dependent. They propose a "gauge" choice of computing $E_j$ on standardized components with $M=I$ to ensure comparability.
• Handling Lags: For models with $K$ lags, the authors utilize a companion-form Jacobian ($nK \times nK$) to correctly represent multi-step propagation, rather than using the leading lag-1 block alone.

Key Contributions

1. Formalization of a Trajectory-Aware Measure: The paper formalizes $E_j$ as an interpretable, estimator-agnostic node-level measure for empirically estimated nonlinear, time-varying systems. It is not a new dynamical quantity per se (it is finite-horizon impulse-response energy) but a new operational framework for its application.
2. Theoretical Reduction to LTI: The authors prove that in the Linear Time-Invariant (LTI) limit, $E_j$ reduces exactly to the average controllability measure (the trace of the node-specific controllability Gramian). This establishes $E_j$ as a generalization of a well-established quantity rather than an unrelated concept.
3. Characterization of Divergence: Through synthetic experiments, the paper maps the conditions under which $E_j$ diverges from static controllability. The divergence is driven not merely by nonlinearity, but by the interaction of nonlinearity with persistent regime structure, particularly persistent sign reversal in pathways.
4. Empirical Diagnostic Framework: The authors place five real-world estimated systems (political, economic, financial, environmental) into a synthetic phase space to diagnose when $E_j$ provides information beyond static controllability.

Results

• Synthetic Validation (Phase Diagram):
  • Agreement: Under static or smoothly drifting dynamics, $E_j$ and average controllability agree closely and both track ground-truth contribution.
  • Divergence: Divergence emerges under persistent regime switching. It is strongest under persistent sign reversal (where a pathway's effect flips sign between regimes). In these cases, static controllability (based on an averaged Jacobian) fails because opposing effects cancel out, underestimating the node's true leverage. $E_j$, by tracking the realized trajectory, retains the contribution of each regime.
  • Limits: At extreme perturbation amplitudes, the local linearization underlying $E_j$ degrades, and both measures fail to track ground truth, identifying the boundary of validity for local-linearization approaches.
• Real-World Applications:
  • Five domains (Macro-finance, Democracy, Realized Volatility, Air Quality, Economic Development) were analyzed using neural vector autoregressions.
  • Low Departure: Economic development, air quality, and realized volatility showed negligible departure from the realized-Jacobian baseline, indicating their dynamics are effectively static or smoothly varying over the horizon.
  • Moderate/High Departure: Macro-finance and the Democracy panel (V-Dem) showed larger departures. The Democracy panel, in particular, exhibited a robust variance–leverage dissociation: nodes with high temporal variance (e.g., "Elected officials," "Suffrage") had low dynamical leverage, while nodes with moderate variance (e.g., "Freedom of association," "Freedom of expression") had high leverage. This structure was invisible to static centralities and variance-based summaries.

Significance and Claims
The paper claims to provide a practical, estimator-agnostic diagnostic for determining when the computational cost of trajectory-aware analysis is justified.

• Descriptive, Not Normative: The authors explicitly state that $E_j$ measures dynamical leverage (how perturbations propagate), not intrinsic "importance" or value. A high-leverage node may be a load-bearing component, a fragile amplifier, or a relay; the measure is descriptive.
• Decision Rule: The paper offers a decision rule based on the phase diagram:
  • If dynamics are static or smoothly drifting, static controllability suffices.
  • If dynamics exhibit persistent regime changes (especially sign reversal), $E_j$ provides necessary information that static controllability misses.
  • If perturbations are extreme, even $E_j$ may lose fidelity.
• Operational Gap: The work fills a gap in the literature where flexible nonlinear models are used for estimation, but interpretation tools remain stuck in linear or static paradigms. It provides a method to extract node-level dynamical leverage directly from fitted models without requiring full nonlinear controllability analysis or reachability certificates.

The authors conclude that emergent contribution extends the trusted linear answer (average controllability) to regimes where the dynamics demand more, specifically when persistent regime-dependent variation exists within a range where local linearization remains faithful.