[Paper] Deep Embedded Multiplicative DMD for Algebra-Preserving Koopman Learning
Source: arXiv - 2606.05131v1
Overview
The paper introduces Deep Embedded Multiplicative Dynamic Mode Decomposition (DeepMDMD), a new framework that blends deep‑learning flexibility with the algebraic rigor of Koopman operator theory. By learning a latent representation of the dynamics and enforcing the exact Koopman product rule, the method builds compact, dynamically coherent dictionaries that can predict complex nonlinear systems far more reliably than existing approaches.
Key Contributions
- Hybrid learning scheme that simultaneously discovers latent coordinates and imposes the Koopman multiplication identity as a hard constraint.
- Alternating training algorithm: an exact multiplicative operator update followed by a differentiable clustering step that groups latent states into “cells” where Koopman closure holds.
- Finite‑dimensional transition map whose non‑zero eigenvalues lie on the unit circle, guaranteeing stability and reducing spectral pollution.
- Demonstrated scalability on high‑dimensional fluid‑dynamics benchmarks (e.g., a 158 k‑dimensional cylinder wake and a noisy Re = 20 000 lid‑driven cavity).
- Empirical evidence that DeepMDMD yields far more compact dictionaries, preserves coherent structures, and delivers stable long‑time forecasts even under severe measurement noise.
Methodology
- Latent embedding – A neural encoder maps raw state snapshots (x_t) into a low‑dimensional latent vector (z_t).
- Latent partitioning (clustering) – The latent space is split into discrete cells. Within each cell the Koopman operator is assumed to act linearly.
- Multiplicative DMD update – For each cell, the method solves an exact multiplicative dynamic‑mode‑decomposition problem that enforces the Koopman product rule
[ \mathcal{K}(f\cdot g)=\mathcal{K}f\cdot \mathcal{K}g, ]
ensuring that the learned operator respects the underlying algebra. - Alternating optimization – The encoder weights are updated via back‑propagation while the multiplicative operator is recomputed analytically. This loop continues until convergence.
- Prediction & decoding – Future latent states are propagated with the learned finite‑dimensional Koopman matrix, then a decoder network maps them back to the physical state space.
The key idea is that the coordinates are learned flexibly (deep nets) but the algebra they must obey is fixed, preventing the model from drifting into spurious, non‑physical representations.
Results & Findings
| Benchmark | Dim. | Baseline (MDMD) | DeepMDMD | Highlights |
|---|---|---|---|---|
| Hamiltonian oscillator | 2 | Spectral leakage, unstable forecasts | Eigenvalues on unit circle, exact energy preservation | Long‑time stability |
| Chaotic Lorenz‑96 | 40 | Over‑complete dictionary, noisy spectra | Compact dictionary, clearer continuous‑spectrum structure | Reduced spectral pollution |
| Cylinder wake (PIV) | 158 624 | Lost vortex shedding patterns after a few steps | Preserved coherent vortex structures for >100 ms | Scalable to >10⁵ dimensions |
| Noisy lid‑driven cavity (Re = 20 k) | 120 000 | Divergent predictions, blurred flow features | Robust to sensor noise, accurate statistical moments | Real‑world turbulence handling |
Across all tests, DeepMDMD required orders of magnitude fewer basis functions than geometric MDMD while delivering more accurate eigenvalue spectra and stable long‑term predictions.
Practical Implications
- Model‑Based Control – Engineers can embed the learned Koopman matrix into model‑predictive controllers, gaining linear‑system tools (e.g., LQR) for inherently nonlinear processes such as flow control or robotics.
- Reduced‑Order Modeling (ROM) – The compact dictionaries act as high‑fidelity ROMs that can be deployed on edge devices or in real‑time simulation loops.
- Data‑Driven Diagnostics – Spectral information (eigenvalues on the unit circle) provides a quick sanity check for stability, useful for monitoring critical infrastructure (e.g., power‑grid dynamics).
- Noise‑Robust Forecasting – Because the algebraic constraint filters out spurious modes, the method is well‑suited for sensor‑rich environments where measurements are noisy (e.g., aerospace telemetry).
- Plug‑and‑Play Pipelines – DeepMDMD can be wrapped around existing DMD/Koopman libraries (PyKoopman, DMDpy) with minimal code changes, letting developers experiment without rebuilding the entire pipeline.
Limitations & Future Work
- Latent dimensionality selection remains heuristic; automated rank‑determination could further streamline deployment.
- The current clustering step assumes a hard partition of latent space, which may struggle with smoothly varying dynamics; soft‑assignment or manifold‑aware clustering is a promising direction.
- Scaling to billions of snapshots will require distributed training and memory‑efficient operator updates.
- Extending the framework to control inputs (Koopman with actuation) and parameter‑varying systems is left for future research.
Overall, DeepMDMD offers a compelling recipe: let deep nets discover where to look, then force the how of Koopman algebra to keep the model honest. This blend of flexibility and structure could become a standard tool in the data‑driven dynamics toolbox.
Authors
- Kelan Gray
- Finlay Brown
- Nicolas Boullé
- Matthew J. Colbrook
Paper Information
- arXiv ID: 2606.05131v1
- Categories: cs.LG, math.DS, math.NA, math.OC, math.SP
- Published: June 3, 2026
- PDF: Download PDF