Transfer Operators from Optimal Transport Plans for Coherent Set Detection

Péter Koltai, Johannes von Lindheim, Sebastian Neumayer, Gabriele SteidlarXiv 2020 · DOI · OpenReview

Abstract

The topic of this study lies in the intersection of two fields. One is related with analyzing transport phenomena in complicated flows. For this purpose, we use so-called coherent sets: non-dispersing, possibly moving regions in the flow's domain. The other is concerned with reconstructing a flow field from observing its action on a measure, which we address by optimal transport. We show that the framework of optimal transport is well suited for delivering the formal requirements on which a coherent-set analysis can be based on. The necessary noise-robustness requirement of coherence can be matched by the computationally efficient concept of unbalanced regularized optimal transport. Moreover, the applied regularization can be interpreted as an optimal way of retrieving the full dynamics given the extremely restricted information of an initial and a final distribution of particles moving according to Brownian motion.

Introduction: Tracking the Unlabeled

In the study of fluid dynamics and dynamical systems, researchers often look for "coherent sets"—regions of a flow that resist dispersion and stay together over time, like the Great Red Spot on Jupiter or a persistent weather front. Traditionally, identifying these sets requires knowing the exact trajectories of individual particles (Lagrangian data).

However, in many real-world scenarios, such as satellite meteorology or cell imaging, we only have access to "Eulerian" data: snapshots of density fields at discrete times. Because we lose the "identity" of the particles between snapshots, calculating how much a region dispersed becomes a significant challenge. This paper bridges the gap between Optimal Transport (OT) and Dynamical Systems theory to reconstruct underlying dynamics from unlabeled distributions.

Snapshots of a particle system at initial and final time steps, demonstrating the challenge of identifying coherent sets from unlabeled density data.

Methodology

The core innovation is the construction of a Transfer Operator ($L$) directly from an Optimal Transport plan ($\pi$). The authors utilize several key concepts to achieve this:

  1. Regularized Optimal Transport: Instead of standard OT, the authors use entropic regularization (the Sinkhorn algorithm). This introduces a small amount of diffusion into the transport plan, which the authors physically justify via the "Schrödinger Connection": regularized OT represents the most likely path for a cloud of particles undergoing independent Brownian motion.
  2. Unbalanced OT (UOT): Real-world data often involves mass loss (e.g., rain) or gain. UOT relaxes the strict requirement that total mass must be conserved, preventing the algorithm from forcing unrealistic transport links between distant, unrelated regions.
  3. Spectral Segmentation: Once the transfer operator is built from the OT plan, a Singular Value Decomposition (SVD) is performed. The second singular vector acts as a classifier, naturally partitioning the domain into coherent sets.

Experimental Validation

The authors validated their approach on a mix of real-world and synthetic datasets:

  • Precipitation Data: Using radar-derived densities over Germany, the Unbalanced OT approach successfully separated distinct cloud systems, even as mass was lost to rainfall.
  • 3-Well Potential: In a stochastic differential equation model, the method correctly identified three "wells" as coherent sets, even under conditions of strong particle mixing.
  • Molecular Dynamics: In a simulation of 200 particles interacting via Lennart-Jones potentials, the OT-based operator's cluster labels agreed with ground-truth particle identities in 84% of cases, despite the algorithm never seeing the identities.

Key Insights and Discussion

A primary takeaway is that regularization is physical. In this context, the entropic regularization parameter $\epsilon$ directly corresponds to the diffusion of the particles. Furthermore, the study demonstrates that mass conservation can be a burden in geophysical applications; Unbalanced OT is essential for tracking structures like clouds that grow and shrink over time.

While the method shows high accuracy, it remains sensitive to the choice of parameters $\epsilon$ and the imbalance parameter $\kappa$. Future work may involve extending this framework to "multiple measure pairs"—using a sequence of many snapshots to build a more robust concatenated transfer operator for highly chaotic systems.

Full Citation: Koltai, P., von Lindheim, J., Neumayer, S., & Steidl, G. (2021). Transfer Operators from Optimal Transport Plans for Coherent Set Detection. arXiv preprint arXiv:2006.16085v2.

Summary updated

Abstract

The topic of this study lies in the intersection of two fields. One is related with analyzing transport phenomena in complicated flows.For this purpose, we use so-called coherent sets: non-dispersing, possibly moving regions in the flow's domain. The other is concerned with reconstructing a flow field from observing its action on a measure, which we address by optimal transport. We show that the framework of optimal transport is well suited for delivering the formal requirements on which a coherent-set analysis can be based on. The necessary noise-robustness requirement of coherence can be matched by the computationally efficient concept of unbalanced regularized optimal transport. Moreover, the applied regularization can be interpreted as an optimal way of retrieving the full dynamics given the extremely restricted information of an initial and a final distribution of particles moving according to Brownian motion.
BibTeX
@misc{koltai2020transfer,
  author = {Péter Koltai and Johannes von Lindheim and Sebastian Neumayer and Gabriele Steidl},
  title = {{Transfer Operators from Optimal Transport Plans for Coherent Set Detection}},
  year = {2020},
  eprint = {2006.16085},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2006.16085},
}
Figure