Must the original system be linear?

No, but the basic fitted operator is linear in the chosen measurements and has approximation limits.

Does a mode prove a physical cause?

No. Physical interpretation needs independent evidence.

Is a growing eigenvalue always a real instability?

No. It may reflect a local trend, noise, limited observations or model mismatch.

AI Glossary / Statistics and probability / Advanced

Dynamic mode decomposition

Definition

Dynamic mode decomposition, usually abbreviated DMD, is a data-driven method for representing observed evolution through spatial patterns and associated temporal rates. In a standard formulation, it fits a linear mapping between paired snapshots and studies that mapping's eigenvalues and eigenvectors. The resulting modes can describe growth, decay and oscillation in the measured data. DMD can be applied to observations from nonlinear systems, but a fitted finite-dimensional linear model does not automatically provide an exact global description of the underlying nonlinear dynamics.

What it means

A snapshot is a vector of measurements taken at one point in a sequence, such as a spatial field or several sensor readings. Rather than only finding directions of large variation, DMD asks how measurements evolve from one snapshot to the next. An eigenvalue describes a mode's step-to-step multiplier, while the corresponding mode describes its measurement-space pattern. This makes the method useful for separating coherent temporal behavior. Interpretation depends on the observations, sampling interval, rank choice and algorithm variant used.

How it works

Arrange paired snapshots as columns of X and their successors as columns of Y. A basic least-squares operator is A = Y X-plus, where X-plus is the Moore-Penrose pseudoinverse. Practical algorithms often use a truncated singular-value decomposition of X to form a smaller operator whose eigenstructure can be computed efficiently. Reconstruct suitable DMD modes in the original measurement space and estimate amplitudes from the data. For evenly spaced snapshots, an eigenvalue's magnitude indicates growth or decay per step, and its angle encodes oscillation. Connections with Koopman operators require attention to observables and approximation conditions; exact and projected DMD variants should not be conflated.

Three practical examples

Sensor analysis: identify recurring spatial patterns with distinct temporal behavior in a measured process.
Simulation review: summarize dominant coherent evolution from a sequence of high-dimensional solution snapshots.
Model reduction research: evaluate whether a smaller modal representation reconstructs and predicts measurements adequately on unseen sequences.

Think of it this way

Think of describing a moving picture as several patterns, each changing according to its own multiplier and rhythm. Adding the patterns can reconstruct the observed sequence. This is richer than a still-image summary, but the patterns are mathematical components and do not automatically correspond to separate physical causes.

Worked example

Consider three hypothetical two-dimensional snapshots one time unit apart: x0 = (1, 1), x1 = (2, 0.5) and x2 = (4, 0.25). Form X with columns x0 and x1, and Y with columns x1 and x2. X has determinant 1 x 0.5 - 2 x 1 = -1.5, so it is invertible. Multiplying Y by X inverse gives A with diagonal entries two and 0.5 and zero off-diagonal entries. Its eigenvectors are the coordinate directions, and its eigenvalues are two and 0.5. With initial amplitudes one, the first mode doubles while the second halves each step. The fitted next snapshot is (8, 0.125). Direct matrix multiplication verifies both observed transitions exactly. This deliberately noise-free linear example demonstrates the calculation; exact reconstruction of two transitions is not evidence that an arbitrary real process will continue doubling or halving.

Illustrative case study

Fictional case: Ash Process Lab applies DMD to regularly sampled temperature fields from a controlled experiment. Engineers compare reconstructed fields with held-out snapshots and inspect whether modes remain stable when the analysis rank changes. When an actuator setting changes during collection, they consider a method that explicitly represents control inputs rather than attributing every change to autonomous dynamics. This invented case illustrates validation and model scope. It does not claim that a visually coherent mode establishes a physical mechanism or that one decomposition guarantees useful control decisions.

Common misconceptions and mistakes

Do not confuse DMD modes with principal components: the objectives and associated temporal interpretation differ. Mode amplitudes and ordering do not have a universal meaning independent of normalization and fitting choices. Measurement noise can bias standard estimates, and rank truncation can remove important behavior or retain noise. Irregular sampling needs a method suited to the actual time intervals. Oscillation frequencies can be aliased when sampling is too slow. A real nonlinear system may require richer observables or local approximations, and external forcing can invalidate an autonomous-model interpretation.

Business use

Businesses working with dense sensor or simulation data can use DMD as an exploratory and reduced-modeling tool. Define whether the goal is interpretation, reconstruction or prediction, because success on one does not guarantee success on the others. Preserve sampling intervals, preprocessing and rank settings, and compare results with meaningful baselines. Evaluate uncertainty and sensitivity before using modes in operational decisions. When inputs or interventions drive the process, incorporate that information explicitly. A compact set of modes can simplify a dataset while still missing rare but operationally significant events.

Check your understanding

What does eigenvalue 0.5 mean in the worked example?

Answer: The associated modal contribution halves each sampled step. It describes the fitted step interval, so a different interval changes the interpretation of a per-step multiplier.

Related terms

Eigenvalue
Singular value decomposition
Principal component analysis
Koopman operator
reduced-order model
snapshot matrix
dynamic mode decomposition with control

Learn next

compare reconstruction and out-of-sample prediction, then examine how noise-aware variants, richer observables and explicit control inputs change the assumptions behind a modal model.

Sources

Sources: Tu, Rowley, Luchtenburg, Brunton and Kutz, On Dynamic Mode Decomposition: Theory and Applications: https://arxiv.org/abs/1312.0041 ; author-hosted published paper: https://faculty.washington.edu/sbrunton/papers/TuRoLuBrKu2014.pdf ; Proctor, Brunton and Kutz, Dynamic Mode Decomposition with Control: https://arxiv.org/abs/1409.6358 .

Previous: Dynamic Markov compression Next: Dynamic model