Spatiotemporal Pattern Extraction via Dynamic Mode Decomposition

35 concepts33 papers

Complex systems—ranging from the turbulent flow of fluids to the rhythmic firing of neurons—often appear chaotic and high-dimensional. Traditional analysis tools like the Fourier Transform excel at finding frequencies but ignore spatial structure, while Principal Component Analysis (PCA) identifies spatial patterns but discards temporal order. Dynamic Mode Decomposition (DMD) bridges this gap by extracting coherent spatiotemporal structures that evolve linearly in time.

The Linear Approximation of Nonlinearity

At its core, DMD treats a sequence of data snapshots as the output of a linear dynamical system. If we represent the state of a system at time $t$ as a vector $x_t$, DMD seeks a linear operator $A$ such that:

$$x_{t+1} \approx Ax_t$$

Even when the underlying physics are nonlinear, Nachbin et al. explain that DMD can identify the best-fit linear transition that maps one state to the next. By performing an eigendecomposition on this learned operator $A$, the method yields DMD modes: spatial patterns that each oscillate at a specific frequency and grow or decay at a specific rate.

This approach is powerful because it provides a "temporally-aware basis." As Grinspun et al. demonstrate in fluid simulations, this basis allows researchers to predict future states through simple matrix multiplication, capturing the evolution of the fluid rather than just its static geometry.

Beyond Fourier: Capturing Damping and Growth

While Fourier analysis assumes signals are composed of eternal sines and cosines, real-world systems are often dissipative or unstable. DMD’s eigenvalues provide "complex energies" that describe not just the frequency of a mode, but its damping rate.

In quantum physics, Sakata et al. used this property to isolate weak, heavily damped oscillations in semiconductors that standard Fourier methods missed. Similarly, in neuroscience, Casorso et al. applied DMD to fMRI data to calculate "damping times," revealing how long specific brain networks persist before fading—a metric that correlates with individual behavioral traits.

Scaling to High-Dimensional Data

A major challenge in spatiotemporal analysis is the sheer volume of data. To handle high-resolution video or large-scale neural recordings, researchers have developed variants that prioritize efficiency:

  • Randomized DMD: Donovan et al. introduced a randomized approach (rDMD) that uses matrix approximations to separate moving objects from backgrounds in real-time, outperforming traditional computer vision methods in speed.
  • Sparse Sensing: In industrial settings like HVAC control, Kramer et al. showed that DMD can reconstruct complex flow regimes using only a few strategically placed sensors, identifying the "fingerprint" of the system's dynamics from minimal input.
  • Hankel-DMD: To capture more complex, non-linear trajectories—such as pedestrians moving through a crowd—Jiang et al. utilized Hankel matrices (stacking delayed versions of the data) to lift the dynamics into a higher-dimensional space where they become linear and predictable.

Applications Across Scales

The versatility of DMD stems from its ability to find order in "messy" biological and physical data. Brunton et al. pioneered the treatment of large-scale neural activity as a "flowing fluid," using DMD to extract sleep spindle networks and motor maps. This perspective helps explain how biological diversity—or "neural heterogeneity"—actually aids information processing. As Wu et al. argue, DMD can quantify how well a network's internal modes align with external sensory inputs, ensuring robust representation even in noisy environments.

Even in the evaluation of modern AI, the method provides new rigors. ABBAHADDOU et al. developed a DMD-based metric to detect "mode collapse" in generative models, ensuring that synthetic time series capture the full geometric and temporal diversity of real-world data. By decomposing complex signals into interpretable, linear components, DMD remains a fundamental tool for understanding the rhythms of the natural and synthetic world.

Key Foundational Work

Extracting spatial–temporal coherent patterns in large-scale neural recordings using dynamic mode decomposition (2015) — contributed 1 concept elements to this direction.
Multi-resolution Dynamic Mode Decomposition for Foreground/Background Separation and Object Tracking (2015) — contributed 1 concept elements to this direction.
A unified approach to network anomaly detection (2014) — contributed 1 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_dynamic_mode_decomposition_e2fc03fd,
  title = {dynamic mode decomposition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-mode-decomposition/rdr_e2fc03fdefa44b6d895a2681726ea34f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_dynamic_mode_decomposition_e2fc03fd_v1,
  title = {dynamic mode decomposition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-mode-decomposition/rdr_e2fc03fdefa44b6d895a2681726ea34f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake