Time-Frequency Mode Decomposition

• TFMD is an adaptive signal decomposition method that represents nonstationary signals as localized oscillatory modes in the time-frequency plane.
• It employs diverse methodologies—including STFT segmentation, sparse dictionary learning, and multichannel clustering—to extract coherent oscillatory components.
• TFMD enables precise mode separation and reconstruction, facilitating accurate analysis and interpretation of complex signal structures.

Searching arXiv for recent and foundational papers on Time-Frequency Mode Decomposition and closely related adaptive mode decomposition methods. Time-Frequency Mode Decomposition (TFMD) denotes a family of adaptive signal decomposition methods for nonstationary and multicomponent data in which a signal is represented through modes that are localized in time and frequency rather than through a single global basis. In the narrowest sense, TFMD has been introduced as an STFT-based, non-iterative framework in which modes are defined as contiguous high-energy regions of a spectrogram and reconstructed by inverse STFT masking (Zhou et al., 16 Jul 2025). In a broader literature, closely related formulations include sparse AM–FM decompositions over nonlinear dictionaries (Hou et al., 2012), segmentwise nonlinear Fourier fitting for instantaneous frequency and amplitude extraction (Shea et al., 2021), and sparse time-frequency clustering based on random features (Richardson et al., 2022). Taken together, these works show that TFMD is less a single algorithm than a methodological class centered on adaptive mode extraction in the time-frequency plane.

1. Modal concepts and signal models

A recurring starting point is the AM–FM representation

$f(t)=\sum_{k=1}^M a_k(t)\cos\theta_k(t),$

with $\theta_k'(t)\ge 0$, where $a_k(t)$ is the amplitude envelope and $\theta_k'(t)$ is the instantaneous frequency (Hou et al., 2012). In this formulation, TFMD seeks a sparse set of oscillatory components whose amplitudes and frequencies are smoother than the carrier $\cos\theta_k(t)$, so that each component corresponds to a coherent trajectory in the time-frequency plane.

A closely related model used in nonstationary Fourier fitting is

$z(t)=\mu(t)+\sum_{l=1}^{L} A_l(t)\cos(\phi_l(t)),$

where $\mu(t)$ is a nonperiodic instantaneous mean, $A_l(t)$ is a time-varying amplitude, and $\phi_l(t)$ is a time-varying phase (Shea et al., 2021). Here the instantaneous frequency of a mode is identified with the segmentwise fitted frequency trajectory, while amplitude is recovered from local sine/cosine coefficients. This makes the mode notion explicitly local and windowed.

The direct TFMD formulation introduced later replaces analytic mode laws by a geometric definition: a mode is a structurally coherent, contiguous, high-energy region in the STFT plane (Zhou et al., 16 Jul 2025). Reconstruction then proceeds from masks rather than from an explicit phase law. By contrast, sparse random mode decomposition assumes that the support of a time-frequency coefficient field $\alpha(\omega,\tau)$ is small for fixed $\theta_k'(t)\ge 0$0, so that modes become clusters in a sparse set of active time-frequency atoms (Richardson et al., 2022).

Mode definitions also broaden in waveform-adaptive settings. Multiresolution mode decomposition models a single physical oscillation as a multiresolution intrinsic mode function (MIMF),

$\theta_k'(t)\ge 0$1

so that one mode may occupy a family of harmonically related time-frequency curves rather than a single narrow ridge (Yang, 2017). This suggests that TFMD is best understood as a continuum ranging from narrowband ridge extraction to morphology-aware decomposition.

2. Principal mathematical archetypes

The literature realizes TFMD through several distinct constructions.

Family	Representative formulation	Typical output
Sparse nonlinear dictionaries	$\theta_k'(t)\ge 0$2 with sparse pursuit over adaptive dictionaries (Hou et al., 2012, Hou et al., 2013)	Modes, amplitudes, $\theta_k'(t)\ge 0$3
Segmentwise local fitting	Sliding-window nonlinear Fourier fitting with warm starts (Shea et al., 2021)	Segmentwise frequencies and amplitudes
Sparse time-frequency support clustering	$\theta_k'(t)\ge 0$4-fit of randomized short-time atoms, then DBSCAN on active support (Richardson et al., 2022)	One reconstructed mode per cluster
Spectrogram segmentation	Thresholded connected components of an STFT magnitude image (Zhou et al., 16 Jul 2025)	Binary masks and ISTFT modes

In sparse nonlinear dictionary approaches, the decomposition is posed as a nonlinear $\theta_k'(t)\ge 0$5 or $\theta_k'(t)\ge 0$6 problem over atoms of the form $\theta_k'(t)\ge 0$7. One formulation minimizes the number of modes subject to

$\theta_k'(t)\ge 0$8

and solves the problem by nonlinear matching pursuit with Gauss–Newton phase updates; for periodic data, an FFT-based implementation achieves $\theta_k'(t)\ge 0$9 complexity (Hou et al., 2012). A related dictionary-learning variant builds phase-adapted wavelet dictionaries $a_k(t)$0 and solves

$a_k(t)$1

using an Augmented Lagrangian Multiplier method accelerated by fast wavelet transforms (Hou et al., 2013).

In segmentwise local fitting, the signal is treated as locally representable by a few stationary Fourier modes, estimated repeatedly on overlapping windows. For a full series $a_k(t)$2, the local model is

$a_k(t)$3

with frequencies initialized from FFT peaks and refined by alternating least-squares coefficient updates and gradient-descent frequency updates (Shea et al., 2021). This produces explicit local mode parameters rather than a dense TFR.

Sparse time-frequency clustering methods start from a continuous STFT-like representation, approximate it by a randomized set of localized atoms

$a_k(t)$4

learn sparse coefficients by basis pursuit denoising or LASSO, and reconstruct one mode per DBSCAN cluster in the active $a_k(t)$5 cloud (Richardson et al., 2022). Spectrogram-segmentation TFMD instead treats the STFT magnitude as an image and applies smoothing, thresholding, connected-component labeling, and mask-based reconstruction (Zhou et al., 16 Jul 2025).

3. TFMD in the strict STFT-segmentation sense

The named TFMD method introduced in 2025 uses the STFT as a front end, smooths the nonnegative-frequency magnitude spectrogram with a $a_k(t)$6 averaging kernel, thresholds it by

$a_k(t)$7

with $a_k(t)$8, and then identifies candidate modes by 8-connected connected-component labeling and size-based filtering with $a_k(t)$9 and $\theta_k'(t)$0 (Zhou et al., 16 Jul 2025). The retained masks are mirrored across the negative frequencies, applied to the original complex STFT, and each masked component is reconstructed by ISTFT. Because the number of retained connected components is $\theta_k'(t)$1, the method estimates the number of modes automatically rather than requiring it a priori.

This pipeline makes the operational definition of a mode entirely geometric: a mode is a contiguous high-energy spectrogram region. The full procedure is non-iterative, and the paper gives the dominant complexity as

$\theta_k'(t)$2

with the main costs coming from the STFT and the $\theta_k'(t)$3 ISTFT reconstructions (Zhou et al., 16 Jul 2025). The image-processing stages are comparatively light.

The method is validated on six synthetic cases—two frequency-separated chirps, two sinusoidal FM signals, a four-component mixture with chirp, tone, time-limited FM, and transient AM burst, a low-frequency chirp plus AM tone, a seven-component generalized nonlinear signal with severe overlap, and two stationary tones—and it correctly identified the number of modes in all six cases. Total relative reconstruction errors were reported between $\theta_k'(t)$4 and $\theta_k'(t)$5, and in five of the six test cases TFMD achieved the highest output SNR at 5 dB input SNR among the compared methods (Zhou et al., 16 Jul 2025). A footbridge application further extracted four distinct modes whose dominant peaks aligned closely with the first four known natural frequencies of the Dowling Hall Footbridge.

The same paper also states the principal limitations of this strict TFMD formulation: severely overlapping but non-crossing modes may merge into one mask; weak components may fall below threshold; performance remains tied to the STFT time-frequency tradeoff; binary masks introduce an intrinsic reconstruction error floor; and parameter choice, though limited, still matters (Zhou et al., 16 Jul 2025).

4. One-dimensional TFMD-style algorithms beyond spectrogram segmentation

A major neighboring line is nonstationary Fourier mode decomposition. The method fits a small number of Fourier atoms in each short overlapping segment, stacks the local parameter estimates into matrices $\theta_k'(t)$6 and $\theta_k'(t)$7, and recovers each mode’s instantaneous amplitude by

$\theta_k'(t)$8

Because it imposes no explicit smoothness penalty or bandwidth constraint, abrupt changes in instantaneous frequency, amplitude, or mean can be represented as jumps between adjacent windows. The same absence of regularization, however, means that noise propagates directly into the parameter trajectories, and the method depends strongly on window size and on FFT-based initialization in a nonconvex optimization (Shea et al., 2021).

A different STFT-based family is ETFR-MD, designed for close, overlapping, or crossing instantaneous frequencies under low SNR. It estimates initial ridges through a path-optimization problem with three penalties—one based on STFT magnitude ranking, one enforcing continuity between adjacent time indices, and one penalizing curvature through adjacent segment gradients—then refines the IFs by kernel phase averaging and reconstructs each mode from STFT coefficients restricted to the enhanced ridge locations (Zhang et al., 2020). The paper derives a sufficient interference condition

$\theta_k'(t)$9

and a window-length bound

$\cos\theta_k(t)$0

showing explicitly how separation, chirp rate, and averaging length interact (Zhang et al., 2020).

Empirical Fourier Decomposition occupies a different position. It performs adaptive segmentation of the Fourier spectrum by selecting the $\cos\theta_k(t)$1 strongest spectral peaks, placing boundaries at local minima between them, and reconstructing each mode with an ideal filter bank. Its aim is to avoid the transition-band mode mixing of EWT and the low-to-high versus high-to-low inconsistency of FDM (Zhou et al., 2020). The method requires the number of modes to be predefined, but the paper reports that it yields accurate and consistent decomposition results for multiple non-stationary modes and closely-spaced modes and that it has the highest computational efficiency among the compared methods (Zhou et al., 2020).

The IMFogram takes yet another route. Starting from IMFs obtained by Iterative Filtering or Fast Iterative Filtering, it assigns each IMF a local amplitude and a local frequency on time windows and accumulates them into a mode-based time-frequency matrix through

$\cos\theta_k(t)$2

For a class of piecewise stationary multicomponent signals and $\cos\theta_k(t)$3, the Hadamard square of the IMFogram converges to the spectrogram representation (Cicone et al., 2022). This gives a rigorous bridge between adaptive mode decomposition and a classical TFR.

5. Multichannel, spatiotemporal, and morphology-aware extensions

In multivariate settings, TFMD-like ideas often appear as shared-frequency or shared-phase decompositions. Multivariate variational mode decomposition (MVMD) decomposes

$\cos\theta_k(t)$4

into channel-wise mode components $\cos\theta_k(t)$5 that share a common center frequency $\cos\theta_k(t)$6. The center-frequency update aggregates spectral energy across channels,

$\cos\theta_k(t)$7

which enforces global multichannel mode alignment. In simulations for time-varying phase synchronization, MVMD outperformed BEMD and na-MEMD by yielding narrower confidence intervals, better separation of nearby frequencies, and more accurate state reconstruction (Honari et al., 2022).

Spatiotemporal Intrinsic Mode Decomposition (STIMD) extends IMF-constrained decomposition to mixed multichannel signals through

$\cos\theta_k(t)$8

where columns of $\cos\theta_k(t)$9 are spatial modes and rows of $z(t)=\mu(t)+\sum_{l=1}^{L} A_l(t)\cos(\phi_l(t)),$0 are temporal modes constrained to satisfy IMF-like models $z(t)=\mu(t)+\sum_{l=1}^{L} A_l(t)\cos(\phi_l(t)),$1 (Hirsh et al., 2018). Because each temporal mode admits a meaningful Hilbert transform, STIMD supplies both source separation and Hilbert-spectrum analysis, and it is also designed to enable future-state prediction.

Reduced-order variational mode decomposition (RVMD) moves TFMD into spatiotemporal field analysis. It assumes

$z(t)=\mu(t)+\sum_{l=1}^{L} A_l(t)\cos(\phi_l(t)),$2

penalizes the demodulated bandwidth of each temporal coefficient $z(t)=\mu(t)+\sum_{l=1}^{L} A_l(t)\cos(\phi_l(t)),$3, and solves the resulting problem by block coordinate descent (Liao et al., 2022). The paper shows that RVMD can be reduced into proper orthogonal decomposition or discrete Fourier transform at particular parameter settings, so it explicitly interpolates between low-rank energetic structure and narrowband harmonic structure.

Morphology-aware decompositions go further by weakening the assumption that one mode corresponds to one narrow ridge. Multiresolution mode decomposition uses MIMFs to represent oscillations with time-varying waveform shape, and the paper argues that a single physical component may generate a family of harmonically related TF structures that should be treated as one structured mode rather than several unrelated ones (Yang, 2017). Robust Modal Decomposition (RMD) combines trajectory-Gram embedding with bandwidth regularization, leading to the generalized eigenvalue problem

$z(t)=\mu(t)+\sum_{l=1}^{L} A_l(t)\cos(\phi_l(t)),$4

so that modes are ranked not only by trajectory-space variance but also by differential-energy smoothness (Hao et al., 27 Oct 2025). A plausible implication is that recent TFMD-adjacent work is increasingly hybrid, combining subspace geometry, bandwidth constraints, and adaptive reconstruction.

6. Strengths, limitations, and recurrent controversies

Across the literature, TFMD methods share several strengths. They seek interpretable oscillatory components rather than dense transform coefficients; they often separate drifting means or trends from oscillatory content; and many of them remain effective on nonstationary signals that violate the assumptions of global Fourier analysis. Nonstationary Fourier mode decomposition is explicitly advantageous for abrupt changes because it imposes no smoothness preference (Shea et al., 2021). Spectrogram-segmentation TFMD removes the need to preset the number of modes (Zhou et al., 16 Jul 2025). MVMD improves multichannel phase analysis by enforcing common modal centers across channels (Honari et al., 2022).

Their limitations are equally recurrent. Many methods still require a user-specified number of modes or harmonics—this is explicit for NFMD, EFD, MVMD, RVMD, and MMD—and parameter choice remains central even when the formalism is otherwise adaptive (Shea et al., 2021). Several objectives are nonconvex and depend on good initialization, especially in sparse nonlinear dictionary methods and segmentwise parametric fitting (Hou et al., 2012). Window size, lag length, filter length, or bandwidth penalty control crucial tradeoffs among resolution, stability, and interpretability (Zhang et al., 2020).

A persistent controversy concerns what qualifies as a meaningful mode. One line of work requires narrowband or nearly monochromatic components so that instantaneous phase and frequency are physically interpretable. Another permits one physical mode to occupy several harmonically related ridges or to have a time-varying waveform (Yang, 2017). This suggests that “mode” in TFMD is not a universal primitive but a model-dependent construct.

A second recurring misconception is that good reconstruction alone is sufficient. The analysis of noisy EMD shows why this is false: under additive noise, decomposition can generate transition modes containing both signal and noise, and once this happens the resulting instantaneous frequencies become unreliable even if substantial signal energy is preserved (Kaslovsky et al., 2010). The same paper emphasizes that mode purity is more important than mere reconstruction for meaningful local frequency estimation. In modern TFMD terms, a decomposition must therefore be judged not only by error norms but also by whether its modes support physically interpretable time-frequency trajectories.

Overall, the literature supports a broad but technically coherent view of TFMD. It encompasses spectrogram segmentation, adaptive AM–FM pursuit, sparse time-frequency clustering, local nonlinear Fourier fitting, multichannel bandwidth-constrained decompositions, and morphology-aware mode bundles. The unifying theme is the extraction of a small number of adaptive components whose time-varying frequency content is sufficiently structured to support analysis, reconstruction, and, in some frameworks, direct physical interpretation (Zhou et al., 16 Jul 2025).