Turnpike properties in nonlinear system identification

Julian D. Schiller, Matthias A. MüllerarXiv 2026 · arXiv

TL;DR

When training nonlinear models like recurrent neural networks (RNNs), fixing the initial state is a common computational shortcut, but it risks introducing optimization errors. This paper adapts "turnpike theory" from optimal control to prove that under mild stability and convexity conditions, the outputs of these computationally tractable, fixed-initial-state setups stay remarkably close to the ideal, free-initial-state solutions. They establish fundamental mathematical equivalences between this turnpike behavior, value function coercivity, and system dissipativity.

What Problem Does This Paper Address?

System identification focuses on learning mathematical models of dynamical systems from input-output data. A gold standard approach is Simulation Error Minimization (SEM), which optimizes model parameters by minimizing multi-step prediction errors over a long time horizon. While excellent for ensuring long-term model accuracy, SEM is computationally brutal: calculating gradients requires simulating the entire model forward and backward through massive datasets.

To make SEM practical—especially when training deep learning models like RNNs—practitioners often use shortcuts: they break the data into shorter subsequences and simply fix the initial state (e.g., setting it to zero or a random value) rather than optimizing it.

However, this raises a critical theoretical question: How much does fixing the initial state degrade the quality of our learned model? The authors tackle this gap by analyzing whether the optimal output sequences of the easier, fixed-initial-state problem converge toward the ideal, unconstrained SEM solutions as the dataset grows.

Methods

The authors answer this question by adapting turnpike theory—a concept from economics and optimal control which states that optimal trajectories spend most of their time resting near a specific steady state or reference trajectory (the "turnpike").

Because neural networks and black-box models suffer from non-uniqueness (different parameter combinations can produce identical outputs), the authors introduce an output-based turnpike property. Instead of forcing states or parameters to match, they require the output sequences of the constrained problem to stay close to the closest optimal output sequence of the unconstrained problem.

The core methodology establishes a three-way mathematical equivalence. Under a mild "cost reachability" condition, they prove that the cumulative output turnpike property is entirely equivalent to:

  1. Coercivity of the value function: Ensuring the normalized cost gap grows predictably with output discrepancy.
  2. Strict dissipativity: A tailored system-theoretic property showing that the system structurally dissipates energy (or error) along its optimal paths, formulated using time-varying storage functions and supply rates.

They also introduce a weaker variant called the cardinality turnpike property (which limits how often a trajectory can wander away from the turnpike, rather than bounding the total integrated deviation) and map out exactly when these two definitions align.

Datasets

The paper focuses on fundamental control theory and does not utilize standard empirical benchmarks. Instead, the authors validate their mathematical framework using a synthetic dataset generated by a simulated nonlinear system:

  • Scalar Elman-type Recurrent Neural Network: A classic RNN cell governed by $x^+ = \tanh(\theta x + u)$ and $y = x$.
  • Data Generation: Generated using a true parameter $\theta_{\text{true}} = -0.4$, an initial state $x_0^{\text{true}} = 0.6$, and a sinusoidal input trajectory $U^{\text{d}}$. No measurement noise was added to isolate the exact geometric impact of fixing the initial state.

Key Insights and Ideas

  • Shifting from States to Outputs: Traditional turnpike theory tracks states and controls. By shifting the metric entirely to outputs, the authors made the theory compatible with modern, overparameterized black-box models where multiple internal states can yield identical external behavior.
  • The "Forgetful" Advantage: The authors show that incremental output stability—the property where a model exponentially "forgets" its initial state over time—is the primary engine behind turnpike behavior. If a model is stable in this sense, the penalty of choosing a wrong initial state vanishes rapidly, theoretically justifying the fixed-initial-state shortcut.
  • Frequency vs. Magnitude (Cardinality vs. Cumulative): The authors demonstrate a subtle but crucial distinction: a system can possess a cardinality turnpike (wandering away from the ideal path only a few times) but still fail spectacularly if those few deviations are infinitely large. They prove that bounding the output space (compactness) is the mathematical bridge required to make these definitions equivalent.

Key Results

  • Theoretical Equivalence (Theorem 1): The authors successfully proved that cumulative turnpike behavior, value function coercivity, and strict dissipativity are mathematically equivalent expressions of the same underlying system structure.
  • Parameter Convergence: In the numerical RNN example, as the dataset size $N$ increased from 2 to 20,000 steps, the parameter $\theta'$ learned via the fixed-initial-state shortcut converged directly from an inaccurate $0.152$ to the exact true value of $-0.400$.
  • Bounded Cumulative Error: Despite increasing the dataset size by four orders of magnitude ($N=2$ to $N=20,000$), the total cumulative output discrepancy ($E_N$) remained strictly bounded, flattening out at approximately $1.685$. This empirically confirmed the presence of the cumulative turnpike property.
Dataset Size ($N$) Learned Parameter ($\theta'$) Scaled Value-Function Gap ($\frac{V'(\bar{x})-V^*}{M(N)} \cdot 10^2$) Cumulative Output Discrepancy ($E_N$)
2 0.152 0.003 2.587
20 -0.322 2.464 1.641
200 -0.383 3.447 1.675
2,000 -0.398 3.665 1.684
20,000 -0.400 3.691 1.685

Limitations and Future Work

  • Convexity Constraints: The sufficient conditions derived for turnpike behavior rely heavily on the uniform strict convexity of the stage cost (such as Mean Squared Error). It remains unclear how these guarantees hold up under non-convex loss functions common in advanced machine learning.
  • Toy Model Scale: While mathematically rigorous, the numerical verification was restricted to a simple, scalar, single-unit RNN. Testing these properties on deep, multi-layer architectures is an open challenge.
  • Advanced Architectures: The authors note that future work should extend this analysis to more complex SEM variants, such as setups that use learned encoders to guess the initial state from past data, to find relaxed conditions where turnpike behavior can still be rigorously guaranteed.

Summary updated

Abstract

We analyze the problem of learning general discrete-time nonlinear state-space models using the simulation error minimization (SEM) method. In this setting, model parameters are typically learned by minimizing the mismatch between simulated and measured outputs over a training dataset, or shorter subsequences extracted from it. Specifically, we study the cumulative output turnpike property of the underlying SEM optimization problem, which requires optimal output sequences emanating from a fixed initial state to approach and remain close to an optimal output sequence of the corresponding SEM problem with free initial state. In the presence of non-unique optimal output sequences---as may arise, for instance, in fully black-box system identification using neural networks---the property is formulated with respect to the closest such sequence. Turnpike behavior is generally desirable in practice, as it provides a theoretical justification for employing computationally more tractable SEM formulations with fixed initial states while ensuring that their optimal output sequences remain close to unconstrained optimal ones. Under a mild reachability condition, we establish equivalence between the cumulative turnpike property, coercivity of the value function, and a tailored notion of strict dissipativity. We additionally introduce a cardinality turnpike property and show that it is strictly weaker than the cumulative notion. Finally, we establish sufficient conditions for turnpike behavior based on incremental output stability, convexity of the stage cost, and a suitable optimality condition, and illustrate the theory by means of a numerical example.
BibTeX
@misc{schiller2026turnpike,
  author = {Julian D. Schiller and Matthias A. Müller},
  title = {{Turnpike properties in nonlinear system identification}},
  year = {2026},
  eprint = {2609.02071},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.02071},
}
Figure