Inverse problems for dynamic patterns in coupled oscillator networks: When larger networks are simpler

1. Problem Statement

The paper addresses the inverse problem of reconstructing the parameters and interaction functions of large-scale dynamical systems, specifically networks of coupled phase oscillators.

• Context: While forward modeling (predicting behavior from known parameters) is well-established, inferring system parameters (e.g., coupling strength, interaction range, phase lag) from observational data is challenging.
• Challenges: Traditional data-driven methods often require massive datasets, high-resolution time derivatives, and complete knowledge of all oscillator trajectories. These methods become computationally prohibitive as the network size ($N$) increases and are sensitive to noise and partial observations.
• Goal: To develop a method that can reconstruct model parameters (natural frequency $\omega$, phase lag $\alpha$, and coupling function $G(x)$) from a single observation of a complex dynamic pattern (specifically a chimera state) using only time-averaged quantities, without needing trajectory derivatives or full state knowledge.

2. Methodology

The proposed approach leverages the mean-field limit and statistical equilibrium relations (SERs) derived from the thermodynamic limit ($N \to \infty$).

A. Theoretical Foundation

The authors focus on the Kuramoto-Battogtokh system of nonlocally coupled phase oscillators:
$$\frac{d\theta_j}{dt} = \omega - \frac{2\pi}{N} \sum_{k=1}^N G(x_j - x_k) \sin(\theta_j - \theta_k + \alpha)$$
In the thermodynamic limit, the system's statistical equilibrium is described by a probability density function $\rho(\theta, x, t)$. For stationary coherence-incoherence patterns (chimera states), the system admits an analytical solution on the Ott-Antonsen manifold.

B. Derivation of Statistical Equilibrium Relations (SERs)

By analyzing the continuity equation and the self-consistency of the stationary solution, the authors derive exact algebraic relations linking time-averaged observables to system parameters. These relations hold for the infinite system limit and remain highly accurate for large finite $N$.

Key derived relations for a stationary chimera state include:

1. Frequency-Order Parameter Relation:
   $$\frac{\omega - \Omega_j}{\omega - \Omega} = \frac{2|\zeta_j|^2}{1 + |\zeta_j|^2}$$
   Where $\Omega_j$ is the effective frequency of oscillator $j$, $\Omega$ is the global frequency, and $\zeta_j$ is the local order parameter.
2. Coupling Relation:
   $$\text{Re}\left(\frac{\xi_j}{\zeta_j}\right) = \frac{2}{1 + |\zeta_j|^2}$$
   Where $\xi_j$ involves the coupling function $G(x)$ and the phase lag $\beta = \pi/2 - \alpha$.

C. The Reconstruction Algorithm

The algorithm operates in three main steps:

1. Data Reduction: Instead of storing full trajectories, the method computes only $O(N)$ time-averaged quantities for each oscillator:
   • Effective frequency: $\Omega_j = \langle \dot{\theta}_j \rangle$
   • Local order parameter: $\zeta_j = \langle e^{i\theta_j} \frac{\bar{Z}}{|Z|} \rangle$ (where $Z$ is the global order parameter).
   • Note: Derivatives are not required; only phase values are needed.
2. Parameter Estimation:
   • $\omega$ and $\Omega$: Determined via linear regression on the frequency-order relation (Eq. 7).
   • Phase Lag ($\beta$): Found by minimizing a cost functional $J_{incoh}(\beta)$ derived from the incoherent part of the chimera state (where $|\zeta_j| < 1$).
   • Coupling Function ($G(x)$): Once $\beta$ is known, the Fourier coefficients of $G(x)$ are found by solving a linear system derived from the remaining SERs.
3. Robustness Features:
   • Partial Data: The method works with a random subset of oscillators (e.g., 25% of $N$) by replacing global sums with subset averages.
   • Noise Insensitivity: Since it relies on time averaging (a low-pass filter), it is robust to unbiased measurement noise.
   • Protophase Handling: It can reconstruct parameters even if the observed data is a "protophase" (a transformed version of the physical phase) without knowing the transformation function.

3. Key Contributions

• Paradigm Shift: Demonstrates that for large networks, inverse problems become simpler due to the emergence of statistical equilibrium relations, contrasting with the complexity of small-system inference.
• Non-Invasive & Fast: The method requires no perturbation of the system, no derivative calculations, and has a computational cost of $O(N^2 M)$ (where $M \ll N$ is the number of Fourier modes), significantly lower than standard matrix inversion ($O(N^3)$).
• Robustness: Proven effective for:
  • Unevenly sampled or noisy data.
  • Partial observations (only a fraction of oscillators observed).
  • Different coupling topologies (top-hat, exponential, cosine).
• Generalizability: The framework is extended to all-to-all coupled networks (Kuramoto-Sakaguchi) and random networks, showing that SERs can be adapted to various topologies.

4. Results

The method was validated numerically on the Kuramoto-Battogtokh model with $N$ up to 8192 oscillators:

• Accuracy:
  • Phase Lag ($\beta$): Reconstructed with absolute error $< 0.0015$.
  • Natural Frequency ($\omega$): Reconstructed with error $< 0.0008$.
  • Coupling Function ($G(x)$): Fourier coefficients were recovered with high precision, allowing clear distinction between top-hat, exponential, and cosine coupling types.
• Finite-Size Effects: The SERs were shown to be accurate for $N > 1000$ and averaging times $T > 1000$, behaving similarly to thermodynamic laws in finite particle systems.
• Partial Data: Successfully reconstructed parameters using only 25% of the oscillators in an $N=8192$ network.
• Protophases: The algorithm successfully handled data where the physical phase was unknown and replaced by a protophase, provided the coherent frequency could be identified.

5. Significance

• Bridging Scales: The paper provides a rigorous mathematical link between microscopic oscillator dynamics and macroscopic statistical observables, turning a high-dimensional inverse problem into a tractable algebraic one.
• Practical Application: The method is highly relevant for real-world systems where full state measurement is impossible (e.g., neuroscience, power grids, biological rhythms). It allows researchers to infer interaction ranges and coupling strengths from limited, noisy, or partial data.
• Theoretical Insight: It highlights that "larger networks are simpler" in the context of inverse problems because the statistical regularity (SERs) dominates the microscopic fluctuations, providing a robust "fingerprint" of the system parameters.
• Future Directions: The authors suggest this approach can be extended to non-stationary patterns (traveling/breathing chimeras), higher-order interactions, and other complex systems like neural networks and power grids.