Fourier Motion Operator: Spectral Methods

• Fourier Motion Operator is a family of methods that represent motion through Fourier-domain techniques such as phase modulation, spectral coefficients, and graph harmonics.
• It integrates models including neural operators, phase-based transforms, and graph Fourier analysis to extract, compensate, and predict motion fields.
• The approach emphasizes operator learning, harmonic decomposition, and spectral compression to enhance applications in imaging, seismology, and computational dynamics.

Searching arXiv for recent and foundational papers relevant to “Fourier Motion Operator,” including Fourier-domain motion representations, operator-learning formulations, and closely related terminology. “Fourier Motion Operator” does not denote a single standardized method in the literature surveyed here. Instead, it refers to a family of Fourier-domain operators that represent, predict, compensate, or analyze motion by acting on trajectories, spatiotemporal fields, or motion-derived signals in the frequency domain. In different research areas, closely related constructs include Fourier Neural Operators that map static media to dynamic motion histories, phase-ramp operators that convert translation into Fourier phase modulation, scalar spectral descriptors of motion regularity, and graph Fourier decompositions of collective motion (Lehmann et al., 2023, Mosbach et al., 2021, Li et al., 2023, Jones et al., 2015, Schultz et al., 2021).

1. Conceptual scope

Across applied mathematics, scientific machine learning, imaging, and signal processing, a Fourier motion operator is best understood as an operator whose internal representation of motion is spectral rather than purely spatial or temporal. The common mechanism is that motion is encoded by Fourier coefficients, Fourier phases, Fourier multipliers, or graph Fourier harmonics, and the operator either maps one field to another, extracts a motion descriptor, or compensates motion-induced phase changes.

Interpretation	Core map	Representative papers
Neural-operator surrogate	static field $\mapsto$ dynamic motion history	(Lehmann et al., 2023, Lee et al., 23 Jul 2025)
Phase-domain motion transform	translation $\mapsto$ Fourier phase modulation	(Egidi et al., 2024, Li et al., 2023, Bian et al., 2016)
Spectral motion descriptor	trajectory $\mapsto$ scalar regularity/coherence	(Jones et al., 2015)
Graph-spectral motion analysis	swarm signal $\mapsto$ harmonic coefficients	(Schultz et al., 2021)
Abstract operator-theoretic construction	Fourier transform composed with algebraic or pseudodifferential operators	(Perez-Leija et al., 2015, Triebel, 2022, Katsnelson et al., 2009)

This variety suggests that the term is most precise when qualified by domain. In seismology and stochastic dynamics it denotes an operator-learning map between function spaces; in imaging it denotes a phase operator induced by translation; in biomedical motion analysis it denotes a compressed spectral statistic; and in graph signal processing it denotes a Laplacian-induced harmonic decomposition of collective motion.

2. Core spectral principles

The most recurrent principle is the Fourier shift property: spatial translation leaves Fourier magnitude unchanged and appears as a phase factor. In phase-based motion magnification, if $I_2(x)=I_1(x+\delta)$, then

$$\hat I_2(\omega)=e^{2\pi \iota \delta^T\cdot \omega}\hat I_1(\omega),$$

and motion magnification is implemented by amplifying that phase term before inverse transformation (Egidi et al., 2024). In single-pixel imaging of moving objects, the same principle yields the motion-compensation operator

$$\tilde{I}_{corr}(k_x,k_y)=\tilde{I_i}(k_x,k_y)\exp\!\left[-2\pi j(k_x\Delta x_i+k_y\Delta y_i)\right],$$

which maps Fourier measurements acquired in a moving frame back to a common reference frame (Li et al., 2023). In motion-corrected Fourier ptychography, per-measurement sample motion similarly acts as a multiplicative phase ramp

$e^{j(k_x\Delta x+k_y\Delta y)}$

on each pupil-windowed sub-spectrum (Bian et al., 2016).

A second principle is spectral operator learning. In Fourier Neural Operators, a latent field is updated by a Fourier multiplier together with a pointwise linear term:

$$v_{\ell+1}=\sigma\left(\mathcal F^{-1}(R_\ell\cdot \mathcal F(v_\ell))+W_\ell v_\ell\right).$$

Here the learned coefficients $R_\ell$ parameterize a kernel operator in Fourier space, while $\mapsto$0 acts pointwise (Lehmann et al., 2023). This turns motion prediction into approximation of an operator between function spaces rather than a finite-dimensional regression tied to one mesh.

A third principle is harmonic decomposition on non-Euclidean domains. In graph Fourier analysis of collective motion, the graph Laplacian

$\mapsto$1

or its normalized variant induces an eigendecomposition

$\mapsto$2

and the graph Fourier transform is

$\mapsto$3

Motion signals such as headings, positions, and phases are then analyzed through their harmonic coefficients, with low graph frequencies encoding topology-respecting collective structure (Schultz et al., 2021).

A fourth principle is spectral compression of motion into reduced descriptors. In pancreatic SBRT, the superior-inferior tumor trajectory is transformed into the frequency domain and compressed into Spectral Coherence,

$\mapsto$4

the fraction of Fourier bins whose amplitudes fall below a threshold (Jones et al., 2015). This is not a motion propagator; it is a Fourier-domain operator that maps a trajectory to a scalar regularity measure.

3. Neural-operator formulations for spatiotemporal motion

A particularly strong realization of the concept appears in operator-learning surrogates for PDE-driven dynamics. In 3D seismology, a Fourier Neural Operator in a U-shaped variant (UNO) learns an operator

$\mapsto$5

where $\mapsto$6 is a static 3D geological field and $\mapsto$7 is the three-component surface ground-velocity field over space and time (Lehmann et al., 2023). The input is a $\mapsto$8 voxelized geology, the output is surface velocity on a $\mapsto$9 grid over 128 time steps, and time is handled by converting the encoder’s depth axis into the decoder’s temporal axis. The architecture has 8 Fourier layers, 87 million parameters, and is trained on 30,000 SEM3D simulations. It is explicitly one-shot rather than autoregressive. Phase fidelity is stronger than amplitude fidelity, discrepancy increases above about $\mapsto$0 Hz, Peak Ground Velocity is underestimated, and the method is strongest for wave-arrival timing, coherent large-scale motion patterns, and longer-period or moderate-period motion metrics (Lehmann et al., 2023).

Mirror-padded Fourier Neural Operators extend this logic to non-Markovian stochastic trajectories on finite intervals. MFNO inserts mirror padding

$\mapsto$1

before Fourier processing and truncation back to $\mapsto$2 afterward (Lee et al., 23 Jul 2025). This makes non-periodic stochastic paths compatible with Fourier processing on a torus. The paper proves that MFNOs can approximate solutions of path-dependent stochastic differential equations and Lipschitz transformations of fractional Brownian motions to arbitrary accuracy, using Wong–Zakai approximations as the bridge from stochastic dynamics to continuous path-space operators. Empirically, MFNO exhibits strong resolution generalization and generates sample paths faster than classical numerical schemes (Lee et al., 23 Jul 2025).

Two extensions broaden the operator-learning picture. COMPOL augments per-process FNO branches with recurrent or attention-based latent aggregation to model coupled multi-physics outputs, reporting a two to three-fold improvement in predictive performance on biological systems, fluid mechanics, and multiphase flow (Li et al., 28 Jan 2025). Iso-FNO constrains Fourier kernels by discrete isotropy, enforcing

$\mapsto$3

which reduces parameters by up to a factor of 16 in 2D and 96 in 3D and improves test performance on isotropic PDE data (Staddon, 4 May 2026). These are not motion operators in a narrow kinematic sense, but they show how Fourier operators can be specialized for dynamical field evolution.

4. Imaging, video, and computational photography

In video prediction, the operative Fourier motion object is the phase-difference transformation

$\mapsto$4

where $\mapsto$5 are frame Fourier coefficients (Mosbach et al., 2021). This transformation represents motion as phase change rather than pixel displacement. The method then derives explicit translation parameters, builds higher-order transformations, and uses relational object motion in a parent-child graph. On synthetic data it reports sharper predictions than a pixel-space GRU baseline, with 14.0K parameters versus 26.0M for the baseline, and substantially lower MSE across 2-object and 3-object prediction tasks (Mosbach et al., 2021).

Phase-based motion magnification uses the ratio

$\mapsto$6

to isolate translation-induced phase, then amplifies it by $\mapsto$7 so that the reconstructed displacement becomes $\mapsto$8 (Egidi et al., 2024). The implementation described is pairwise and global rather than the full temporal bandpass pipeline associated with earlier phase-based magnification literature. It works best for global translation, synthetic grayscale data, and sub-pixel motion, and it does not yet address local spatially varying motion (Egidi et al., 2024).

Single-pixel imaging of moving objects uses low-frequency Fourier patterns both for localization and for image acquisition. With six tracking patterns and three imaging patterns per motion frame, and a 20,000 Hz DMD refresh rate, the paper reports a maximum achievable frame rate of approximately 2222 Hz (Li et al., 2023). Object translation is inferred from the phase of low-frequency coefficients, and motion-compensated Fourier coefficients are assembled for FFT reconstruction of both images and edges. The method is tailored to pure translation of small moving objects and is explicitly unsuitable for rotation or deformation (Li et al., 2023).

Motion-corrected Fourier ptychography expresses unknown sample motion as a phase ramp on each sub-spectrum and adaptively updates the object spectrum from low spatial frequencies toward high spatial frequencies. The large overlap between adjacent sub-spectra guides motion recovery, and the reported method can correct unknown sample motion with standard deviation up to 10% of the field-of-view scale (Bian et al., 2016).

5. Reduced descriptors and structured motion analysis

Not all Fourier motion operators generate trajectories. Some compress motion into diagnostically or structurally informative spectral summaries. In pancreatic SBRT, daily tumor motion is reconstructed from CBCT projections and Fourier analyzed in the superior-inferior direction only. The resulting Spectral Coherence is correlated with mean trajectory error with $\mapsto$9 and correlation coefficient $\mapsto$0, compared with $\mapsto$1 when motion magnitude itself is used as the predictor (Jones et al., 2015). The threshold is selected as $\mapsto$2 cm, roughly the 98th percentile of Fourier-transform values. The approach supports an adaptive margin protocol in which the SI margin is increased to 7 mm for the lower quartile of mean SC and decreased to 3 mm for the upper quartile, with reported improvements in D90, D95, and D99 (Jones et al., 2015). This is a scalar motion regularity operator rather than a field predictor.

Graph Fourier analysis offers another reduced but structured view. Collective motion is represented as a graph signal on a swarm graph, and harmonics derived from the Laplacian reveal hidden structure beyond swarm-averaged order parameters. For connected graphs, the first harmonic of a heading signal corresponds to the standard alignment mode, while subsequent low harmonics capture milling and annular circulation; for disconnected swarms, each component has its own local DC mode (Schultz et al., 2021). The framework is also used with doubled-angle signals such as $\mapsto$3 to expose counter-rotating structure, and with position-phase graphs in swarmalators to relate phase organization to spatial topology (Schultz et al., 2021). This suggests a broader notion of motion operator in which the Fourier basis is induced by relational geometry rather than Euclidean translation.

A further stochastic-process extension appears in operator fractional Lévy motion, whose real harmonizable class uses the kernel

$\mapsto$4

with $\mapsto$5 (Boniece et al., 2021). Here the Fourier-domain kernel defines a non-Gaussian multivariate motion process with covariance operator self-similarity and stationary increments. The paper shows that moving-average and harmonizable constructions generally differ outside the Gaussian case, even when they share covariance structure (Boniece et al., 2021).

6. Abstract operator-theoretic interpretations and recurring limitations

In a more abstract sense, Fourier motion operators also arise as algebraic or analytic operators related to phase, translation, and spectral compression. In finite-dimensional quantum optics, the Pegg–Barnett phase operator is obtained by applying the discrete Fourier transform to the finite number operator,

$\mapsto$6

and its exponential is the cyclic shift operator

$\mapsto$7

a genuine finite translation operator on a cyclic space (Perez-Leija et al., 2015). In functional analysis, the truncated Fourier operator

$\mapsto$8

compresses Fourier action to a measurable set $\mapsto$9 and serves as a canonical restricted Fourier operator model (Katsnelson et al., 2009). More generally, Fourier operators of the form

$I_2(x)=I_1(x+\delta)$0

compose the Fourier transform with a pseudodifferential operator and are studied through compactness and spectral theory on Besov and Triebel–Lizorkin scales (Triebel, 2022).

These abstract formulations clarify a common misconception. A Fourier motion operator is not necessarily a generic Fourier parameterization of motion trajectories, and it is not necessarily a universal motion-analysis module detached from modeling assumptions. The cited literature imposes strong structural conditions. Seismological neural operators fix source location, mechanism, and discretization regime (Lehmann et al., 2023). SBRT Spectral Coherence analyzes the superior-inferior component only and does not disentangle amplitude variation, baseline shifts, and period variation (Jones et al., 2015). Phase-based magnification, single-pixel tracking, and Fourier ptychography model translation; they do not directly handle rotation, deformation, or strongly local motion without further machinery (Egidi et al., 2024, Li et al., 2023, Bian et al., 2016). Graph Fourier motion analysis depends materially on graph construction, Laplacian choice, and per-time-step spectral decomposition (Schultz et al., 2021). MFNO theory relies on mirror padding, piecewise-linear driving approximations, and regularity assumptions on path-dependent coefficients (Lee et al., 23 Jul 2025).

Taken together, these works indicate that “Fourier Motion Operator” is best treated as a family resemblance term. Its unifying feature is that motion is represented or manipulated through Fourier-domain structure—classical, learned, graph-based, stochastic, or operator-theoretic—but its exact meaning depends on whether the target object is a trajectory, a PDE solution operator, a phase correction, a scalar coherence statistic, or a harmonic decomposition of collective behavior.