Data-driven spectral analysis for coordinative structures in periodic systems with unknown and redundant dynamics

Keisuke Fujii, Naoya Takeishi, Benio Kibushi, Motoki Kouzaki, Yoshinobu KawaharabioRxiv 2019 · DOI

TL;DR

This paper introduces a novel data-driven spectral analysis method, Hankel Dynamic Mode Decomposition (Hankel DMD), to uncover the underlying dynamical properties of complex biological movements like human walking. By treating these movements as nonlinear limit-cycle oscillations, the method extracts not just the structure but also the temporal dynamics (frequency and phase) of coordinative movements, offering a deeper understanding of biological systems with unknown dynamics.

What Problem Does This Paper Address?

Living organisms, particularly in their motor control, exhibit remarkable flexibility and adaptability despite being composed of numerous components. This "redundancy problem" (also known as Bernstein's degrees of freedom problem) has long puzzled researchers. While traditional methods like Principal Component Analysis (PCA) can identify low-dimensional "coordinative structures" in movements, they often fail to capture the inherent dynamical properties, such as frequency and phase evolution. This is a significant limitation because these dynamical aspects are crucial for understanding how biological systems adapt and respond. The paper specifically targets periodic biological systems, like human locomotion, where the governing equations are not fully understood, making data-driven approaches essential.

Methods

The core of this paper is the application of Hankel Dynamic Mode Decomposition (Hankel DMD). This method is an extension of Dynamic Mode Decomposition (DMD), which is itself a powerful tool for extracting spectral properties (eigenvalues and eigenfunctions) of dynamical systems from data.

Here's a breakdown of the key methodological aspects:

  • Operator-Theoretic Spectral Analysis: The authors frame the problem within the context of Koopman operator theory. This theory allows for the linear analysis of nonlinear dynamical systems by lifting the problem to an infinite-dimensional function space. The Koopman operator captures the system's time evolution, and its spectral properties (eigenvalues and eigenfunctions) directly correspond to observable dynamics like frequencies and coherent structures.
  • Hankel DMD: Unlike standard DMD, Hankel DMD utilizes delay-embedding techniques by constructing Hankel matrices from the time-series data. This is theoretically shown to yield the Koopman eigenvalues and eigenfunctions more directly, especially for periodic systems. The paper explores both "row-type" and "column-type" Hankel DMD.
  • Focus on Dynamics: The key innovation is the emphasis on extracting dynamical properties. While standard dimensionality reduction methods focus on spatial structures, Hankel DMD aims to reveal temporal characteristics like frequency and phase, which are crucial for understanding how coordinative structures evolve over time and adapt to different conditions (e.g., walking speed).
  • Phase Reduction: The method also enables a form of "phase reduction," where a complex, high-dimensional limit-cycle oscillation is approximated by a simpler phase oscillator. This simplifies analysis and provides insights into the system's phase dynamics.

Datasets

The primary dataset used in this study is:

  • Human Locomotion Data: Kinematic data (intersegmental angles) from ten healthy male participants walking on a treadmill at 14 different controlled speeds, ranging from 2.0 to 8.0 km/h, plus a preferred walking speed. This data was collected using a 3D optical motion capture system.

The authors also used simulation data for validation:

  • Double Pendulum Model: A simplified double pendulum model with known governing equations and analytically derived eigenfrequencies.
  • Walking Model Simulation: A five-link neural oscillator control model simulating human locomotion, with explicit governing equations but unknown frequency modes.

Key Insights and Ideas

  • Beyond Structure: Dynamics Matter: The paper strongly argues that understanding biological movement requires analyzing its dynamical properties (frequency, phase) in addition to its spatial structure. Conventional methods often miss this crucial aspect.
  • Hankel DMD for Biological Systems: Hankel DMD proves to be a powerful tool for extracting these dynamical properties from complex, noisy biological data, even when the underlying governing equations are unknown.
  • Harmonic Frequencies in Gait: Human walking exhibits not just a fundamental gait frequency but also several harmonic frequencies, which Hankel DMD can effectively identify. These harmonics may arise from the complex interactions within the multi-link system and with the environment, rather than just measurement noise.
  • Speed-Dependent Phase Evolution: The phase of coordinative structures evolves differently depending on walking speed. This suggests that the system dynamically adjusts its phase relationships to accommodate varying locomotion speeds.
  • Bridging Data-Driven and Dynamical Systems: The approach effectively bridges the gap between purely data-driven methods and the theoretical framework of nonlinear dynamical systems, offering a unified perspective.

Key Results

  • Extraction of Harmonic Frequencies: Hankel DMD successfully extracted the first five harmonic frequencies of human gait, outperforming other DMD variants in terms of accuracy and completeness.
  • Speed-Dependent Phase Dynamics: The study revealed speed-dependent time-evolving behaviors of the phase on conventional low-dimensional coordinative structures, visualized through Koopman eigenfunctions. This indicates how the system's phase dynamics adapt to different walking speeds.
  • Validation with Simulations: The Hankel DMD approach was validated on both a double pendulum model and a walking model simulation, demonstrating its ability to accurately recover known frequencies and dynamics.
  • Comparison with SVD: While SVD-based methods showed good reconstruction error, they failed to extract the underlying dynamical properties, highlighting the advantage of Hankel DMD for this purpose.

Limitations and Future Work

  • Idealized Assumptions: The study assumes that walking dynamics can be approximated as a limit cycle. However, real-world locomotion can involve rhythm shifts and perturbations, which are not perfectly captured by this assumption.
  • Mechanical Interpretation: While the study identifies harmonic frequencies and phase dynamics, providing strict mechanical interpretations for these phenomena requires further investigation.
  • Control Applications: The authors suggest future work could involve applying this understanding to control problems using forward simulations, potentially integrating neuro-musculo-skeletal models with neural inputs and environmental feedback.
  • Generalizability: Further research is needed to explore the application of this method to other types of locomotion and to investigate variations within and among participants.
  • Parameter Selection: While the paper discusses methods for selecting Hankel DMD parameters (p and m), it acknowledges that cross-validation criteria might conflict with the goal of avoiding fitting to high-frequency dynamics. The chosen parameters were based on convergence and avoiding overfitting.

Summary updated

Abstract

Abstract Living organisms dynamically and flexibly operate a great number of components. As one of such redundant control mechanisms, low-dimensional coordinative structures among multiple components have been investigated. However, structures extracted from the conventional statistical dimensionality reduction methods do not reflect dynamical properties in principle. Here we regard coordinative structures in biological periodic systems with unknown and redundant dynamics as a nonlinear limit-cycle oscillation, and apply a data-driven operator-theoretic spectral analysis, which obtains dynamical properties of coordinative structures such as frequency and phase from the estimated eigenvalues and eigenfunctions of a composition operator. First, from intersegmental angles during human walking, we extracted the speed-independent harmonics of gait frequency. Second, we discovered the speed-dependent time-evolving behaviors of the phase on the conventional low-dimensional structures by estimating the eigenfunctions. Our approach contributes to the understanding of biological periodic phenomena with unknown and redundant dynamics from the perspective of nonlinear dynamical systems.
BibTeX
@misc{fujii2019data,
  author = {Keisuke Fujii and Naoya Takeishi and Benio Kibushi and Motoki Kouzaki and Yoshinobu Kawahara},
  title = {{Data-driven spectral analysis for coordinative structures in periodic systems with unknown and redundant dynamics}},
  year = {2019},
  doi = {10.1101/511642},
  howpublished = {bioRxiv},
}
Figure