Diffeomorphic Latent Mappings for Dynamical Systems

18 concepts12 papers1 proposal

Modeling complex, high-dimensional dynamical systems—such as chaotic fluid flows, biological oscillations, or robotic maneuvers—presents a fundamental tension. On one hand, deep learning offers the flexibility to capture intricate patterns from raw data. On the other hand, classical control theory and physics require rigorous guarantees like stability, energy conservation, and linearity that standard "black-box" neural networks often violate.

A growing body of research resolves this tension by using invertible neural networks (INNs) to learn a diffeomorphism: a mapping that is both smooth and bijectively invertible. By transforming messy, real-world observations into a structured latent space, researchers can define simple, "primitive" dynamics that are mathematically guaranteed to behave well, then map those behaviors back to the original space without losing their structural properties.

Decoupling Geometry from Dynamics

The core intuition behind this approach is the separation of spatial complexity from temporal evolution. In a standard Neural ODE, the network must learn the entire vector field of a system at once. In contrast, Zhi et al. propose "morphing" complex target dynamics into a simple base system. By using an INN to learn a coordinate transformation, the model offloads the spatial complexity (the "shape" of the trajectories) to the mapping, while the temporal dynamics are handled by a simple, tractable ODE.

This decoupling allows for provable stability. For instance, Takeishi et al. map high-dimensional data into a latent space where the dynamics are restricted to geometric primitives like limit cycles or line attractors. Because the mapping is a diffeomorphism, the stability properties defined in the latent space are preserved when the data is mapped back to the high-dimensional manifold. Similarly, Sochopoulos et al. use this mechanism to ensure that robot motions generated from raw images always reach their goal safely.

Linearization and Control

A major application of these mappings is feedback linearization. In control theory, many powerful tools like the Linear Quadratic Regulator (LQR) only work for linear systems. Beier et al. demonstrate that INNs can learn the necessary diffeomorphism to transform nonlinear system states into a latent Linear Time-Invariant (LTI) model.

The choice of an invertible architecture is critical here. As Goswami et al. argue, a valid coordinate transformation in control must be a diffeomorphism to ensure that the original state can be recovered from the linearized representation. Standard autoencoders do not strictly enforce this, but INNs—often implemented via coupling layers—provide the architectural constraint necessary to satisfy this mathematical requirement. This approach is further extended by Wang et al., who use normalizing flows to map nonlinear stochastic systems into a latent linear Gaussian space, enabling efficient Bayesian filtering.

Solving the Inverse Observation Problem

Diffeomorphic mappings also address the challenge of data assimilation, where one must estimate the full state of a system from sparse, noisy observations. Traditional methods involve solving difficult non-convex optimization problems.

Instead, Frerix et al. and the authors of VIVID propose learning an "inverse observation operator." By training a neural network to map sparse measurements back into the high-dimensional physical state space, they provide a global reconstruction that guides optimization toward physically consistent solutions. This concept is formalized in the TorchDA framework, which integrates these learned inverse mappings into standard Kalman filtering and variational workflows.

Preserving Physical and Topological Truths

When the underlying physics of a system are known (e.g., energy conservation), diffeomorphic mappings provide a way to enforce these constraints even when the "canonical" coordinates are unknown. Lim et al. show that by integrating INNs into Neural ODEs, one can map observed states to a latent space where physical inductive biases can be applied more effectively.

Furthermore, these mappings preserve the "topology" of the system. Moriel et al. introduce Smooth Prototype Equivalences (SPE) to bridge the gap between noisy experimental data and clean analytical models. Because the mapping is smooth and invertible, it ensures that features like fixed points or oscillations in the data correspond directly to those in the theoretical prototype. This topological consistency is also what allows Duan et al. to discover closed-form ODEs from partial observations, as the INN recovers the hidden dimensions of the state space required for the dynamics to make sense.

Key Foundational Work

Towards Data-driven LQR with Koopmanizing Flows⋆ (2022) — contributed 2 concept elements to this direction.
Learning Efficient and Robust Ordinary Differential Equations via Invertible Neural Networks (2022) — contributed 2 concept elements to this direction.
Variational Data Assimilation with a Learned Inverse Observation Operator (2021) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_diffeomorphic_latent_mapping_8ace1d01,
  title = {diffeomorphic latent mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/diffeomorphic-latent-mapping/rdr_8ace1d0157a646048d38ce2e3d0749e9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_diffeomorphic_latent_mapping_8ace1d01_v1,
  title = {diffeomorphic latent mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/diffeomorphic-latent-mapping/rdr_8ace1d0157a646048d38ce2e3d0749e9/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake