Research Radar — 2026-10-22

Methodology Review: Dynamic mode decomposition: separate patterns by how they evolve

English · About 10 minutes

Main algorithm checked in full author manuscript; application and Koopman scope claims limited to primary publisher abstract/introduction. No empirical omics validation is claimed.

The intuition

Imagine watching two lamps through a translucent screen. One fades rapidly; the other changes slowly. Every photograph mixes both lamps, and a bright pattern need not be the most persistent one. The question is not just which photographs resemble each other, but which patterns obey a reproducible step-to-step rule. This invented scene is a useful way to separate compression from prediction: a compact description of the present is not automatically a model of the future.

Dynamic mode decomposition, or DMD, fits a linear map between paired snapshots and studies its eigenvectors. A mode is a pattern that the fitted map rescales rather than changes in shape; its eigenvalue specifies that rescaling. Several modes can coexist in one observation. The method is attractive when measurements are high-dimensional but only a limited set of temporal patterns is needed. Its fitted matrix is a description of the supplied pairs, not an automatic identification of the mechanism that generated them. [R1]

Think of the pairing as part of the measurement, not a clerical choice. A before-and-after pair from one interval asks a different question from two unrelated samples placed next to each other. Sorting specimens by a convenient score does not manufacture observations of the same system evolving. This distinction will matter more for a proposed omics use than whether the matrix calculation is fast.

A small worked example

Here is an original, noiseless example with two measured coordinates and equally spaced observations. Let \(x_0=(2,1)^\top\), \(x_1=(1,1)^\top\) and \(x_2=(0.5,1)^\top\), where the subscript counts time steps and the superscript \(\top\) turns a row listing into a column vector. The first coordinate halves each step, while the second stays constant. Use the first two snapshots as inputs and the last two as their matched outputs. This example is deliberately algebraic; neither coordinate represents a particular gene.

The input and output matrices, named \(X\) and \(Y\), and the fitted transition matrix \(A\) are

The inverse exists because the determinant of the input matrix is one. Multiplication checks both pairs exactly. In this small example ordinary inversion suffices; real rectangular datasets require the pseudoinverse described below.

The two modes are \(\phi_1=(1,0)^\top\) and \(\phi_2=(0,1)^\top\), with eigenvalues \(\lambda_1=0.5\) and \(\lambda_2=1\). The initial amplitudes are \(b_1=2\) and \(b_2=1\). After \(k\) steps the reconstructed vector is \(x_k=2(0.5)^k\phi_1+\phi_2\). Thus the next prediction is \(x_3=(0.25,1)^\top\). Each symbol has an operational meaning: a mode says where the pattern lives, an amplitude says how much is initially present, and an eigenvalue says what happens to it per step.

Now perform two paper-and-pencil checks. First, double the first coordinate in every snapshot. Its initial amplitude doubles, but its halving factor does not change. Second, replace only the last snapshot by a different value. With just two independent input columns, an exact map can still be fitted, yet its next prediction can change sharply. Zero training error was guaranteed by this tiny setup, not earned by extensive evidence. A fourth, withheld snapshot is more informative about prediction than admiring the exact fit.

The necessary mathematics

For \(m\) paired observations with \(n\) coordinates, collect input columns in \(X\in\mathbb{R}^{n\times m}\) and matched output columns in \(Y\in\mathbb{R}^{n\times m}\). The minimum-norm least-squares choice is \(A=YX^{+}\), where \(X^{+}\) denotes the Moore–Penrose pseudoinverse. It minimizes \(\lVert Y-AX\rVert_F\); the subscript \(F\) means the square root of the sum of squared matrix entries. Among multiple minimizers it chooses the one with smallest such norm. DMD then examines \(A\phi=\lambda\phi\), with nonzero vector \(\phi\) and scalar eigenvalue \(\lambda\). [R1]

Computation can avoid explicitly storing the large transition matrix. With the reduced singular value decomposition \(X=U\Sigma V^\top\), columns of \(U\) and \(V\) are orthonormal singular vectors and the diagonal entries of \(\Sigma\) are retained nonzero singular values. Form \(\widetilde A=U^\top YV\Sigma^{-1}\) and diagonalize this smaller matrix. A reduced eigenvector \(w\) obeys \(\widetilde A w=\lambda w\). For a nonzero eigenvalue, \(YV\Sigma^{-1}w\) gives an exact DMD eigenvector up to nonzero scaling. Deliberate rank truncation changes the fitted problem; it is not merely a faster identical computation. [R1]

Assumptions and failure modes

A practical checklist should begin with the target interval, coordinate system and treatment of external inputs. In the toy calculation, one step has a fixed meaning. Pooling different elapsed times while calling each pair one step would change the question. Likewise, mixing an untreated trajectory with a strongly forced trajectory without representing that distinction can make a single transition rule hard to interpret. These are design cautions for the proposed use, not claims that a successful numerical decomposition has diagnosed the cause of a failure.

Measurement error deserves its own check. In the displayed reduced-operator formula, division by a very small retained singular value can magnify a small perturbation in the measured data. This is a direct conditioning consequence of the inverse, not a claim that every noisy dataset fails. Discarding weak singular directions can reduce that sensitivity, but also removes information and changes predictions. Compare ranks using held-out transitions rather than choosing the rank solely for a visually attractive reconstruction. [R1]

The Koopman connection is useful but easy to overstate. An operator on functions of a nonlinear state can be linear even when the state dynamics are not. That does not guarantee a finite, exact linear model containing the original state coordinates. Brunton and colleagues examine invariant observable subspaces and show important limitations for global state-inclusive representations, notably with multiple fixed points or more complicated attractors. A fitted matrix and a mathematical closure theorem are different objects. [R3]

For a first implementation, keep a temporal holdout, compare against persistence of the most recent observation, and perturb the chosen rank and measurement scaling. Inspect residuals coordinate by coordinate rather than reporting only one aggregate error. Ask whether a slow-looking mode survives a different observation window. These are proposed diagnostics, not a calibrated hypothesis test or a guarantee of biological meaning. In particular, forecasting accuracy alone would not demonstrate that a mode is a causal regulatory program.

Verified applications

Schmid and colleagues report applying DMD to helium-jet Schlieren images and time-resolved particle-image-velocimetry measurements of unforced and harmonically forced jets. These are established experimental-fluid applications with actual ordered measurements, rather than a biological analogy inferred from the title. The inspected publisher abstract supports that application statement, but not a claim of universal noise resistance or superiority over alternative decompositions. Its online publication date is August 2010; the volume appeared in 2011. [R2]

Possible transfers — speculative

A possible omics experiment would use repeated population-level measurements from independently replicated cultures at controlled intervals, with a small, predefined set of molecular summaries. Fit on some trajectories and forecast the next measurement on held-out cultures. Compare with a constant predictor and a separate trend for each summary before seeking interpretable modes. This is a speculative transfer, not a validated omics application established by these three readings. Destructive sampling would require explicit assumptions that matched cultures represent comparable evolving populations.

Do not turn a single-cell pseudotime ordering into a literal movie of one cell. Cross-sectional heterogeneity, changing population composition and uncertain alignment can all produce apparent transitions without observed longitudinal pairs. A defensible pilot would keep those uncertainties visible and restrict its claim to the measured population summaries. If the intended question is an intervention effect, collect an appropriate intervention design; DMD cannot supply that design retrospectively.

What to try and what to read next

Start by reproducing the two-coordinate calculation and withholding the fourth snapshot. Then deliberately corrupt the pairing and explain what changed. The central habit is to write down what a transition means before fitting it. Read Tu and colleagues for the algorithm, Schmid and colleagues for experimental context, and Brunton and colleagues for the limits of finite linear representations. A useful result is not merely a colorful mode: it is a clearly defined prediction problem whose assumptions and failures remain inspectable.

Reading list

• [R1] Primary reading: On dynamic mode decomposition: Theory and applications. Jonathan H. Tu, Clarence W. Rowley, Dirk M. Luchtenburg, Steven L. Brunton and J. Nathan Kutz (2014). Main reading: sections 2.2–2.3 explain exact DMD, the reduced operator and linear consistency. The 2013 author manuscript and 2014 journal paper are one work.

• [R2] Related reading: Applications of the dynamic mode decomposition. P. J. Schmid, L. Li, M. P. Juniper and O. Pust (2010). Application reading: see what a physical snapshot and coherent mode mean in measured flow fields. Abstract-level inspection does not establish quantitative robustness.

• [R3] Related reading: Koopman Invariant Subspaces and Finite Linear Representations of Nonlinear Dynamical Systems for Control. Steven L. Brunton, Bingni W. Brunton, Joshua L. Proctor and J. Nathan Kutz (2016). Boundary reading: distinguish a fitted DMD matrix from an exact finite Koopman representation of nonlinear dynamics.