Overcoming the Linear Time Invariant Assumption

26 concepts28 papers1 proposal

In classical control theory, network science, and time-series analysis, the default assumption is almost always linearity. Whether modeling a power grid, predicting financial markets, or tracking a drone, researchers frequently rely on Linear Time-Invariant (LTI) systems or linear stochastic differential equations.

The reasons for this are practical: linear systems are mathematically elegant. They allow us to use matrix algebra, compute exact closed-form solutions, and guarantee stability using eigenvalues. When a system is non-linear, the standard engineering workaround is linearization—zooming in close to a specific equilibrium point and using a Taylor expansion to treat the local behavior as linear.

However, a growing body of research highlights a critical bottleneck: real-world systems rarely stay close to equilibrium, and the assumption of global linearity frequently breaks down. When applied to complex biological networks, chaotic physical processes, or evolving social systems, the linear dynamics assumption introduces a severe modeling gap.

Why and Where Linearity Fails

To understand why the linear assumption fails, we must look at the gap between local approximations and global realities. Linearization works well for a system undergoing minor, slow fluctuations. But as soon as a system experiences rapid or large-scale state changes, the linear model loses its predictive power.

1. Global Transitions vs. Local Equilibria

In physical control systems, such as high-speed drones or automated infrastructure, states change too quickly for local linear models to keep up. As noted by Inaltekin et al. (2017), control frameworks built on linear assumptions fail to capture the true dynamics of systems undergoing rapid, large-scale state transitions. Similarly, topology learning algorithms designed to map out network structures struggle when a system operates far from its linear equilibrium points, a limitation highlighted by Doddi et al. (2022).

2. Hard Physical Constraints

Real-world environments are full of non-linear phenomena that cannot be smoothed away, such as friction, actuator saturation, and heavy-tailed noise. Classical system identification methods often assume linear time-invariance, but as Rajendran et al. (2024) point out, these frameworks fail to account for practical constraints like saturation. Furthermore, Mediano et al. (2017) emphasize that relying on linear models and idealized Gaussian noise simplifications overlooks the exotic informational behaviors and heavy-tailed noise distributions found in natural environments.

3. Non-Linear Signal Propagation in Networks

In complex networks—such as gene regulatory pathways, financial portfolios, or social networks—the relationship between variables is rarely a static linear combination. Jalali et al. (2012) argue that linear stochastic dynamics limit our ability to model state-dependent or non-linear feedback mechanisms common in biology and finance. When structural influence models rely strictly on linear equations, they fail to capture how real-world physical or social variables truly interact (Qin et al. (2023)). This constraint also binds quantitative finance; Mehra et al. (2014) note that modeling asset returns as linear dynamical systems prevents allocation mechanisms from capturing complex, non-linear market interactions.

The Breakdown of Network Controllability

The limitations of the linear assumption become particularly acute in network controllability—the study of how to guide a complex network from any initial state to a desired target state using a minimal set of "driver nodes."

Most foundational metrics in this field, such as control centrality and structural controllability, assume underlying LTI dynamics. Liu et al. (2012) observe that while systems can be linearized around specific equilibria, these approximations fail to capture global dynamic phenomena in complex biological or social systems.

Linear Assumption:      [Driver Node] ---> (Linear Propagation) ---> [Target State] (Predictable)
Non-Linear Reality:     [Driver Node] ---> (State-Dependent Feedback) -> [Target State] (Chaotic/Saturated)

This reliance on linearity distorts how we understand network behavior in several ways:

  • Misclassifying Control Modes: Liu et al. (2013) discovered that networks naturally fall into distinct "centralized" or "distributed" control modes based on their structure. However, the authors caution that inherent non-linear interactions could fundamentally alter these identified control modes and node classifications.
  • Distorting Signal Propagation: In his work on network control capacity, Albert-László Barabási demonstrates that non-linearities significantly alter how signals propagate through a network, changing how driver nodes influence the overall system state (Barabási (2013)).
  • Failing Under Network Growth and Mobility: Real networks are rarely static or linear. Zhang et al. (2019) show that the principles governing how driver nodes evolve in growing networks may not hold when dynamics are non-linear. Similarly, state estimation frameworks fail when forced to handle both non-linear behaviors and mobile, dynamic network structures (Mitra et al. (2020)).

As Li et al. (2023) and Wu-Yan et al. (2020) summarize, the widespread reliance on linear discrete-time models limits the generalizability of controllability metrics to real-world environments where node states and topology interact dynamically.

Moving Beyond Linearity

To bridge this gap, researchers are developing frameworks that preserve mathematical rigor without relying on the crutch of global linearity.

Continuous-Time Differential Equations

Instead of forcing time-series data into static linear combinations, newer methods embrace continuous, non-linear trajectories. For instance, Bellot et al. (2021) developed Neural Continuous Synthetic Controls (NC-SC). By utilizing controlled differential equations (CDEs), this approach natively models non-linear and time-evolving dynamics in complex economic and biological processes, moving past the rigid assumptions of traditional linear baselines. A similar shift toward flexible temporal reasoning was early championed by Dagum et al. (1993), whose Dynamic Network Models bypassed strict linearity and normality assumptions to handle evolving relationships over time.

Geometric and Manifold Learning

When dealing with highly non-linear or chaotic behaviors, viewing data through a flat Euclidean lens fails. Zhang et al. (2016) addressed this by treating temporal sequences as points on a curved Riemannian manifold rather than raw Euclidean vectors. This geometric approach allows for the efficient comparison and classification of complex, non-linear sequences (like human actions) where traditional linear regressive models (such as Hankel matrices) fall short.

Data-Driven and Localized Control

Another alternative is to abandon the goal of identifying a global mathematical model altogether, focusing instead on localized, data-driven predictions.

  • One-Step-Ahead Greedy Identification: Rather than trying to plan for long horizons using a fragile global linear model, Blanke et al. (2022) demonstrated that a simple "greedy" algorithm maximizing information gain just one step ahead can successfully identify system parameters without needing the underlying mechanisms to naturally generalize across non-linear state transitions.
  • Trajectory Prediction: In a similar vein, Liu et al. (2022) introduced a data-driven self-triggered control framework. By leveraging historical data to predict future trajectories directly, the system can safely control complex processes without ever needing to enforce or identify a strict, global linear model.

By shifting from rigid global assumptions to flexible, localized, and geometrically aware models, modern system theory is slowly unlocking the ability to analyze and control real-world systems exactly as they are: complex, dynamic, and fundamentally non-linear.

Key Foundational Work

Control Centrality and Hierarchical Structure in Complex Networks (2012) — contributed 1 concept elements to this direction.
Forecasting Sleep Apnea with Dynamic Network Models (1993) — contributed 1 concept elements to this direction.
Learning the Dependence Graph of Time Series with Latent Factors (2012) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_linear_dynamics_assumption_f0ffab84,
  title = {linear dynamics assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/linear-dynamics-assumption/rdr_f0ffab84f665495c9a35cf8891f8e6f4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_linear_dynamics_assumption_f0ffab84_v1,
  title = {linear dynamics assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/linear-dynamics-assumption/rdr_f0ffab84f665495c9a35cf8891f8e6f4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake