Identification of Structural Model for Chaotic Systems

Technical Summary: Identification of Structural Model for Chaotic Systems

Problem Statement
The paper addresses the challenge of structural identification for dynamical systems exhibiting chaotic dynamics based on one-dimensional time series data. The authors note that reconstructing a dynamical system from a single scalar realization is an ill-posed problem; infinitely many distinct systems can generate signals with similar statistical properties. Existing global reconstruction methods often suffer from several limitations: they frequently rely on arbitrary choices of nonlinearities that fail for real-world systems, produce evolution equations that are cumbersome and difficult to use, or are only validated on simple, low-dimensional model systems where the result is known in advance. Furthermore, processing experimental data introduces noise, where sequential differentiation (often used for phase trajectory restoration) amplifies high-order noise components, and standard methods struggle with heterogeneous signals containing alternating fast and slow motion regions.

Methodology
The authors propose a constructive method for structural identification that leverages the symmetry properties of the reconstructed attractor to determine the form of the underlying model. The methodology proceeds through the following stages:

1. Theoretical Foundation (Symmetry and Reduction):

   • The approach is grounded in the theory of symmetry groups and the classification of dynamical systems up to equivalence relations (diffeomorphisms or homeomorphisms).
   • The authors utilize the concept of the central invariant manifold. They prove that for systems with symmetries, the dynamics can be reduced to a central manifold where the system takes a specific, simplified form.
   • Theorem 1 & 2 establish the existence of center-stable and center-unstable manifolds in the neighborhood of an equilibrium point, demonstrating that the symmetry group of the equilibrium is defined regardless of the specific choice of these manifolds.
   • Theorem 3 shows that through a smooth transformation, the system can be locally reduced to a form where the evolution is governed by a linear part and a nonlinear function determined by the symmetry transformation. For discrete systems, this results in a form $x(t+1) = Ax(t) + \vartheta(x)$, where $\vartheta$ is defined by the symmetry group constructed on the reconstructed attractor.

2. Algorithmic Implementation:

   • Fragment Extraction: The algorithm automatically searches for symmetries (translation, rotation, stretching, compression) within local sections of the phase trajectory. It identifies "markers" (e.g., extrema or evenly distributed points) to partition the trajectory into fragments.
   • Normalization: To compare fragments, a normalization procedure is applied to make them invariant to translation, rotation, and scaling. This involves:
     • Identifying an "axis" (the longest segment between two points).
     • Translating the fragment to the origin.
     • Rotating the fragment to align with coordinate axes.
     • Scaling the fragment so its main axis length equals 1.
     • Representing the fragment as a matrix in homogeneous coordinates.
   • Symmetry Evaluation: Normalized fragments are converted into descriptors using the Discrete Fourier Transform (DFT). The distance between two fragments is calculated by comparing their spectral components, weighted by empirical discount factors to prioritize low-frequency (structural) features over high-frequency noise.
   • Optimization: The problem of selecting the optimal set of non-overlapping, symmetric fragments from the time series is treated as a combinatorial optimization problem. The authors employ a genetic algorithm to find the sequence of fragments that maximizes symmetry while satisfying length constraints.

3. Model Construction:

   • Once the symmetry group is identified via the genetic algorithm, the structure of the evolution equations is constrained to the form derived from the central manifold reduction.
   • The specific coefficients of the affine system (linear part) and the nonlinear terms are identified using the method of least squares.

Key Contributions and Results

• Symmetry-Based Identification: The paper introduces a method where the identification of the system's structure is driven by the detection of geometric symmetries in the reconstructed attractor, rather than arbitrary function selection.
• Geometric Reduction: The authors provide a theoretical proof that systems with symmetries can be reduced to a central invariant manifold, simplifying the identification problem to determining the form of the nonlinear function on this manifold.
• Algorithmic Framework: A complete algorithm is presented, combining geometric normalization, Fourier-based similarity metrics, and evolutionary optimization to handle noisy, real-world data.
• Experimental Validation:
  • Rössler System: The method was tested on a simulated Rössler system. The algorithm successfully identified symmetries in the attractor, with the genetic algorithm converging to a solution with a fragment similarity of 0.3917.
  • Traffic Network Data: The method was applied to experimental data from a traffic network. The algorithm identified a winning solution with a fragment similarity of 0.576. The resulting model, constructed using the identified symmetry structure and least squares, successfully reproduced the dynamics of the original system, as demonstrated in the comparison of time series and attractors.

Significance and Claims
The authors claim that their developed algorithm allows for the determination of observed symmetries in an attractor, which in turn specifies the form of the minimal system model. By utilizing the symmetry group, the method avoids the ambiguity of arbitrary nonlinear function selection. The paper asserts that this approach is effective for real systems, as demonstrated by the successful reconstruction of models from both simulated and experimental data. The results suggest that reconstructed attractors serve as a rich source of information not only for dynamics but also for identifying the structural basis of the models. The work positions itself as a constructive method for structural identification that bridges the gap between theoretical symmetry analysis and practical application to chaotic systems.