TL;DR
This paper introduces a novel approach to identify parameters in chaotic dynamical systems by fitting their invariant measures (long-term statistical behavior) using optimal transport theory. This method overcomes limitations of traditional trajectory-based approaches, especially when data is noisy or sparse, by leveraging the power of optimal transport to compare probability distributions.
What Problem Does This Paper Address?
Identifying the parameters of complex, chaotic dynamical systems is crucial in many scientific and engineering fields. Traditional methods often rely on fitting simulated trajectories to observed data. However, chaotic systems are highly sensitive to initial conditions, making trajectory-based fitting unreliable when data is noisy, sampled infrequently, or when time derivatives are difficult to estimate. This paper addresses this challenge by proposing a method that focuses on the system's invariant measure – its long-term statistical distribution – which is more robust to these data limitations.
Methods
The core of the proposed method lies in using optimal transport (OT) theory, specifically the Wasserstein distance, as a cost function to compare the invariant measure of a model (parameterized by $\theta$) with the observed invariant measure.
- Invariant Measure Approximation: Instead of directly simulating trajectories, the paper uses a PDE-based approach to approximate the invariant measure. This involves solving a stationary continuity equation, which is then regularized using a teleportation technique (similar to Google's PageRank) to ensure a unique and differentiable solution. This regularized solution, $\rho_\epsilon(\theta)$, serves as the model's invariant measure.
- Optimal Transport Cost: The Wasserstein distance ($W_p$) is employed as the objective function to quantify the difference between the model's invariant measure $\rho_\epsilon(\theta)$ and the target invariant measure $\rho^*$.
- Parameter Identification: The parameter identification problem is framed as minimizing this Wasserstein distance with respect to the model parameters $\theta$. This is achieved using gradient-based optimization algorithms.
- Gradient Calculation: A key contribution is the derivation of gradient formulas for the Wasserstein distance with respect to the model parameters. The paper explores two methods: an implicit function theorem approach and an adjoint-state method, with the latter being more computationally efficient.
Datasets
The paper evaluates its method on three well-known chaotic systems:
- Lorenz System: A classic model for atmospheric convection.
- Rössler System: A simpler chaotic system with a single-lobe attractor.
- Chen System: Known for its double-scroll chaotic attractor.
Additionally, a modified "Arctan Lorenz System" is used to test the method's ability to handle nonlinear parameter dependencies. The "reference data" (target invariant measure) for these systems is generated either by solving the PDE surrogate model or by accumulating histograms from direct numerical simulations (DNS) of the ODEs, with and without added noise.
Key Insights and Ideas
- Invariant Measures as Robust Signatures: Focusing on invariant measures, rather than trajectories, provides a more stable and robust way to identify parameters in chaotic systems, especially when data is imperfect.
- Optimal Transport for Distributional Comparison: The Wasserstein distance offers a powerful tool to compare probability distributions, capturing both local and global differences, making it suitable for comparing invariant measures.
- PDE-Based Surrogate Model: Using a PDE (stationary continuity equation) to represent the invariant measure, regularized with teleportation, provides a differentiable forward model that is amenable to gradient-based optimization.
- Differentiability is Key: Rigorously establishing the differentiability of the optimal transport cost with respect to model parameters is crucial for enabling efficient gradient-based optimization.
Key Results
- The proposed method successfully identifies parameters for the Lorenz, Rössler, and Chen systems.
- The method demonstrates robustness to noise (both intrinsic and extrinsic) and works with subsampled trajectory data, mimicking realistic scenarios.
- Coordinate gradient descent is shown to be effective for multi-parameter inference, mitigating issues of parameter interdependencies.
- The paper provides theoretical analysis on the differentiability of the optimal transport cost, guaranteeing the applicability of gradient-based methods.
Limitations and Future Work
- Computational Cost: Solving the forward PDE problem repetitively for optimization can be computationally expensive, especially for high-dimensional state spaces.
- Mesh Size Dependence: The accuracy of the PDE solver is dependent on mesh size, and numerical diffusion can introduce model discrepancy.
- Data Requirements: While the method is more robust to sparse sampling, it still requires access to the full state-space dynamics and a parameterized representation of the system.
- Future Directions:
- Exploring other objective functions beyond the $W_2$ metric.
- Developing more efficient numerical schemes and adaptive mesh refinement strategies.
- Extending the framework to black-box parameter estimation from time-delay-embedded coordinates.
- Investigating the use of neural networks to approximate dynamical systems and their derivatives.
Summary updated