Dynamical Disentanglement via Local Flow Decomposition

High-dimensional time series collected from real-world systems often contain mixtures of multiple dynamical processes. Delay-coordinate methods can reconstruct the underlying dynamical state up to a diffeomorphism from observed measurements, but such reconstructions may inherit entanglement, making them difficult to interpret and model. We propose an unsupervised dynamical disentanglement method that uses local flow prediction and a contrastive objective to learn low-dimensional components with approximately independent, self-contained dynamics. We evaluate the method on synthetic mixtures of nonlinear dynamical systems and on real high-dimensional recordings. The results suggest that the proposed method addresses a useful class of decomposition problems and provides a practical tool for analyzing mixed nonlinear time series from a new angle.