Reproduce “On the relationship between Koopman operator approximations and neural ordinary differential equations for data-driven time-evolution predictions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jake Buzhardt, C. Ricardo Constante-Amores, Michael D. Graham
Researchers from the University of Wisconsin-Madison formally demonstrated that Koopman operator approximations, when combined with a state-space projection at each timestep, are mathematically equivalent to neural ordinary differential equations. This approach provided accurate short-term predictions (e.g., <2% error for Lorenz system) and robustly reproduced long-term statistics and rare event forecasts for chaotic dynamical systems.
Key insights
- EDMD-DL with state-space projection at each timestep fundamentally reintroduces nonlinearity, making it functionally equivalent to a neural network representation of the nonlinear discrete-time flow map on the state space.
- Purely linear Koopman approximations, which do not incorporate state-space projection, are inadequate for capturing the dynamics of chaotic systems.
- The identified neural network structure inherent in projected EDMD-DL provides a blueprint for an 'EDMD-structured neural ODE' that can enhance model design by integrating domain-inspired architectural constraints.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.