Extrapolating learned dynamics across bifurcations based on generalized synchronization

Lingfeng Chen, Masanobu InubushiarXiv 2026 · arXiv

TL;DR

Researchers have developed a mathematically grounded method to predict how complex, chaotic systems behave when their underlying parameters change—even into regimes where no training data exists. By linearly blending just two pretrained "readout" weights in a Reservoir Computing model, they can accurately reconstruct unseen bifurcations and chaotic attractors with rigorous error bounds.

What Problem Does This Paper Address?

In nonlinear science, systems often undergo bifurcations: sudden qualitative changes in behavior (like shifting from a steady state to a chaotic one) triggered by small changes in a parameter. Predicting these transitions is a "holy grail" for scientific machine learning, especially in fields like fluid turbulence where experimental data is only available for a narrow range of conditions.

While Reservoir Computing (RC)—a type of recurrent neural network where only the output layer is trained—has shown promise in mimicking chaotic dynamics, we haven't fully understood why or how it can extrapolate to unseen parameter values. This paper aims to provide a theoretical foundation for this capability, moving beyond empirical "black box" observations to rigorous mathematical proofs.

Methods

The authors leverage the framework of Generalized Synchronization (GS). In RC, the "reservoir" (the internal network) synchronizes with the input signal. The authors focus on the ideal readout, a mathematical map that perfectly translates the reservoir's internal state back into the target system's state.

The core innovation is a surprisingly simple linear combination formula:

  1. Train two separate readout weights ($W_{\mu_a}$ and $W_{\mu_b}$) for the same fixed reservoir at two different parameter points ($\mu_a$ and $\mu_b$).
  2. Interpolate/Extrapolate to a new parameter value by blending these weights: $W_\theta = (1 - \theta)W_{\mu_a} + \theta W_{\mu_b}$.
  3. Analyze the error using classical Lagrange interpolation theory, which allows them to derive rigorous bounds on how much the reconstructed dynamics might deviate from the truth.
Figure 1: Concept of readout interpolation and extrapolation with a linear combination of pre- trained readout weights Wθ = (1 −θ)Wµa + θWµb. The same fixed reservoir is used throughout; only the readout weights change. Weights Wµa and Wµb are trained at µa and µb. The mixing coefficient θ gives interpolation for 0 ≤θ ≤1 and extrapolation otherwise. The lower axis uses the Lorenz parameter µ = ρ, with µa = 23, µb = 24, and comparison value ˜µ(2) = 25 under the parameter-correspondence ansatz in Eq. (11).

Datasets

The method was validated on three classic benchmarks in dynamical systems:

  • Saddle-node normal form: A basic model for the sudden appearance or disappearance of fixed points.
  • Supercritical pitchfork normal form: A model for a system splitting from one stable state into two.
  • The Lorenz System: The iconic "butterfly" attractor, used to test the method's ability to handle high-dimensional chaotic dynamics. For training, they used data at $\rho = 23$ and $24$, then extrapolated to predict behavior at $\rho = 25$ and beyond.

Key Insights and Ideas

  • Readout Linearity: The most striking insight is that while the underlying dynamics are highly nonlinear and chaotic, the ideal readout map often varies smoothly with the system parameters. This smoothness is what allows simple linear interpolation of weights to work so effectively.
  • Reconstructing the Unseen: The model doesn't just replicate what it saw; it can reconstruct unstable fixed points even if it was only trained on stable trajectories. This suggests the reservoir captures the "global" structure of the phase space, not just the local path it was shown.
  • The Identity Theorem Link: The authors propose a fascinating link to complex analysis. If the readout error vanishes over a small range of parameters, the Identity Theorem for analytic functions suggests it should theoretically vanish over the entire connected parameter space.

Key Results

The researchers demonstrated that their linear weight-blending method accurately tracks bifurcations and chaotic statistics:

  • Fixed-Point Accuracy: In the Lorenz system, the error in the reconstructed fixed-point position followed a predictable parabolic curve between training points, exactly as predicted by their Lagrange error bounds.
  • Chaotic Extrapolation: The model successfully predicted the chaotic "butterfly" attractor at $\rho = 25$ (an unseen value). The predicted marginal probability distributions of the system states showed high agreement with the ground truth, measured via Kullback–Leibler (KL) divergence.
  • Bifurcation Tracking: In scalar models, the method correctly identified the loss of stability and the emergence of new branches that were entirely absent from the training data.
Reconstruction of Lorenz dynamics. (a) shows the error bound, (b) compares the predicted vs. reference chaotic trajectories, and (d) shows the KL divergence across a range of parameters.

Limitations and Future Work

  • The Affine Ansatz: The method assumes a simple linear relationship between the mixing coefficient $\theta$ and the physical parameter $\mu$. While this worked for the tested systems, more complex bifurcations might require a nonlinear mapping.
  • The "Butterfly Effect": Because the systems are chaotic, the model cannot predict a specific individual trajectory far into the future (due to sensitivity to initial conditions). The authors suggest combining this method with data assimilation to keep long-term predictions on track.
  • Conjectural Theory: The link to the identity theorem for analytic functions remains "conjectural" for finite training data and requires further formal proof in a machine learning context.

Summary updated

Abstract

Predicting dynamics across bifurcations by extrapolating trained models is crucial in diverse fields of physics. Generalized synchronization provides a framework for understanding dynamical reconstruction by neural networks, especially in the context of reservoir computing. Building on this framework, we derive a formula combining pretrained readouts to reconstruct unseen chaotic dynamics. Using classical Lagrange interpolation theory, we derive rigorous bounds on readout errors and conjugacy residuals. Numerical examples demonstrate accurate prediction of chaotic dynamics across bifurcations.
BibTeX
@misc{chen2026extrapolating,
  author = {Lingfeng Chen and Masanobu Inubushi},
  title = {{Extrapolating learned dynamics across bifurcations based on generalized synchronization}},
  year = {2026},
  eprint = {2610.34471},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.34471},
}
Figure