Learning Linearized Models from Nonlinear Systems under Initialization Constraints with Finite Data

Lei Xin, Baike She, Qi Dou, George Chiu, Shreyas Sundaram

Introduction

System identification is a cornerstone of control engineering, providing the mathematical models necessary for designing robust controllers. While classical methods often assume infinite data or perfectly linear systems, modern research has shifted toward "finite-sample guarantees"—quantifying the data needed for a specific level of accuracy. However, a significant gap remains: real-world systems are inherently nonlinear, and physical hardware often imposes strict constraints on where experiments can be initialized. This paper addresses the challenge of learning a linearized model from a nonlinear system under such initialization constraints using finite, noisy data.

The Challenge of Nonlinearity and Constraints

Most existing finite-sample guarantees rely on data collected from a single, long trajectory driven by random noise. In a perfectly linear system, this approach is highly effective. However, in a nonlinear system, random inputs eventually push the state far from the desired operating point. As the system drifts, higher-order nonlinearities are excited, which degrades the quality of the linear approximation.

Furthermore, many identification methods assume "global" initialization—the ability to start the system at any arbitrary point in the state space. In practice, safety limits and physical boundaries restrict experiments to a "feasible region." This paper asks how we can accurately learn linearized dynamics when we are restricted to a small neighborhood around an operating point and only have access to finite, noisy observations.

Proposed Methodology: Deterministic Multi-Trajectory Sampling

To overcome these hurdles, the authors propose a shift from single long trajectories to a multiple short trajectories approach. The core of their method is a deterministic data acquisition strategy. Instead of relying on random walks, the algorithm selects initial state-input pairs within a user-defined feasible region, defined by a center point $m$ and a perturbation magnitude $q$.

By resetting the system after every single time step, the algorithm ensures the state never drifts far from the reference point. This "power of resets" keeps the higher-order Taylor expansion terms (the nonlinearities) negligible, ensuring the linear approximation remains valid. Once the one-step transition data is collected, the linear system matrices $A$ and $B$ are estimated using regularized least squares (Ridge regression), which provides a safety net against high noise or strong nonlinearity.

Theoretical Analysis and the Fundamental Trade-off

A key contribution of the paper is the derivation of a rigorous finite-sample upper bound on the estimation error. This bound reveals a fundamental trade-off governed by the perturbation magnitude $q$:

  1. Noise Sensitivity: A small $q$ makes the system more susceptible to process noise, as the signal-to-noise ratio is low.
  2. Nonlinearity Bias: A large $q$ improves the signal-to-noise ratio but pushes the system into regions where nonlinear effects are stronger, introducing a "bias" error caused by the remainder terms of the Taylor expansion.
Error plateauing in the standard single-trajectory approach due to nonlinearity.

As shown in Figure 2, the standard single-trajectory approach fails in nonlinear settings because the error plateaus regardless of the number of samples $N$. This occurs because the trajectory inevitably explores regions where the linear model is no longer a valid approximation of the underlying dynamics.

Empirical Validation

The authors validated their theoretical findings using simulated environments, including a discretized nonlinear pendulum and a synthetic polynomial system.

Estimation error and theoretical bounds for different perturbation magnitudes q.

Figure 1 demonstrates the convergence of the proposed method. For different values of $q$, the estimation error decreases as the number of experiments $N$ increases, following the predicted $\mathcal{O}(1/N^{1/4})$ rate. The plot also highlights the $q$ trade-off: while smaller $q$ values (e.g., $q=0.6$) eventually reach lower errors, they may start with higher errors due to noise sensitivity at low $N$. In contrast, the standard single-trajectory method's error plateaued at around 1.0 in similar tests.

Conclusion and Future Work

The proposed deterministic multi-trajectory approach provides a robust framework for system identification under physical constraints. By leveraging resets and deterministic sampling, it avoids the pitfalls of trajectory drift in nonlinear systems. While the current theoretical guarantees rely on the ability to reset the system to exact deterministic initial conditions, empirical tests showed the algorithm remains robust even with small Gaussian noise in the initialization process.

Future work will focus on developing algorithms that require less physical precision for setting initial conditions and investigating the impact of measurement noise on the theoretical learning bounds.

Summary updated

BibTeX
@article{xin2025learning,
  author = {Lei Xin and Baike She and Qi Dou and George Chiu and Shreyas Sundaram},
  title = {{Learning Linearized Models from Nonlinear Systems under Initialization Constraints with Finite Data}},
  journal = {Automatica},
  year = {2025},
  doi = {10.1016/j.automatica.2025.112478},
}
Figure