Fast Anti-Noise Mode Decomposition

Updated 10 July 2026

Anti-noise fast mode decomposition is a research direction that extracts nearly monochromatic oscillatory modes and trends from nonstationary data, minimizing noise interference.
It integrates diverse techniques—such as fast intrinsic methods, FFT-based iterative filtering, and time-frequency ridge extraction—to improve noise resilience without repeated sifting.
The framework is applicable in multichannel signal analysis and dynamic mode decomposition, offering robust, efficient performance even under severe noise conditions.

Anti-noise fast mode decomposition denotes a research direction in adaptive signal and modal analysis concerned with extracting oscillatory modes, trends, or coherent dynamical components from nonstationary and multicomponent data while limiting degradation from noise and reducing the computational burden of repeated sifting or large-scale optimization. In the literature considered here, the term does not correspond to a single canonical algorithm; it spans fast intrinsic decomposition schemes based on integral constraints, FFT-based iterative filtering, time-frequency ridge and segmentation methods, latent-variable and variational multichannel decompositions, and several noise-aware dynamic mode decomposition frameworks (Lu, 2011, Cicone et al., 2019, Zhou et al., 16 Jul 2025, Weiner et al., 2024).

1. Conceptual basis and diagnostic foundations

A central motivation for anti-noise mode decomposition is the failure of classical empirical mode decomposition (EMD) in noisy settings. For instantaneous frequency to be meaningful, each extracted mode must be nearly monochromatic, but this condition is not guaranteed by EMD and can fail badly under noise. The paper "Noise Corruption of Empirical Mode Decomposition and Its Effect on Instantaneous Frequency" identifies the extraction of modes containing both signal and noise—so-called transition modes—as the mechanism that destroys reliable instantaneous-frequency estimation; it further links this behavior to the filter-bank-like action of EMD on pure noise, spectral leak between modes, and the phase of the underlying signal (Kaslovsky et al., 2010).

A second foundation is the critique of classical EMD-style sifting. "Equivalent Effect Function and Fast Intrinsic Mode Decomposition" lists four EMD problems: the spline envelope is not the real envelope and can overshoot, the stopping condition affects results, repetitive sifting is time consuming, and riding waves on steep edges can be missed. These points are important because they explain why later anti-noise fast decomposers often remove or weaken dependence on raw extrema interpolation, stop-rule tuning, and repeated local-mean subtraction (Lu, 2011).

These diagnostic studies imply different robustness targets for later methods. A plausible implication is that an anti-noise fast decomposer should avoid transition-mode extraction, reduce reliance on noisy extrema, and impose some form of smoothing, dynamic consistency, or structural prior before mode estimation.

Family	Representative formulations	Core mechanism
Integral or fast intrinsic decomposition	FastIMD (Lu, 2011)	Integral matching, no sifting
Iterative filtering	MvFIF (Cicone et al., 2019)	FFT-based filtering with shared multichannel scale
Time-frequency decomposition	NMD (Iatsenko et al., 2012), ETFR-MD (Zhang et al., 2020), TFMD (Zhou et al., 16 Jul 2025)	Ridge extraction, surrogate testing, or spectrogram segmentation
Variational or latent multivariate decomposition	SJMD/SMJMD (Nazari et al., 11 Apr 2025), VLMD (Morante et al., 23 May 2025)	Narrow-band regularization, sparse residuals, latent structure
Noise-robust DMD	fbDMD/tlsDMD (Dawson et al., 2015), KFDMD (1804.00143), OCDMD (Weiner et al., 2024), weak-DMD (Bennett et al., 15 Apr 2026), CR-DMD (Nakamura et al., 16 Jan 2026)	Bias correction, explicit noise modeling, weak projection, robust reduction

2. Fast intrinsic and iterative-filtering approaches

Fast intrinsic mode decomposition in the narrow sense is represented by the Equivalent Effect Function (EEF) framework. For a signal $f(t)$ , FastIMD defines the accumulated integral $F(t)=\int_{t_0}^{t} f(x)\,dx$ , chooses control points $t_i$ , constructs an integral spline $E(t)$ such that $E(t_i)=F(t_i)$ , and then defines the trend as $e(t)=E'(t)$ . The fluctuation or IMF-like component is $d(t)=f(t)-e(t)$ , and it has zero integral at the control points. The algorithm proceeds by detecting inflexion points from derivative extrema, forming a piecewise-linear trend estimate, selecting control points from extrema of the detrended residual, fitting the integral spline, differentiating once, and recursing on the extracted trend until the number of extrema drops to $2$. The paper explicitly characterizes the method as “fast and reliable as there is no sifting or complicated calculation involved,” and notes that EEF-based sampling can act as “kind of low pass filtering” (Lu, 2011).

FastIMD is not presented as a formal denoising theory. The same paper states that it does not provide a formal anti-noise theorem, denoising benchmark, or explicit robustness study with additive noise. Its anti-noise relevance is therefore indirect: integration before interpolation attenuates high-frequency perturbations, spline fitting in the integral domain is less local-extrema-sensitive than envelope interpolation, and eliminating repeated sifting reduces opportunities for oversifting or artifact accumulation. The paper also reports that quintic spline interpolation for $E(t)$ gave the best experimental results for intrinsic mode decomposition, which further emphasizes smooth trend construction rather than local pointwise fitting (Lu, 2011).

A more explicitly fast and multivariate line is Multivariate Fast Iterative Filtering (MvFIF). MvFIF represents a multichannel signal as $s\in\mathbb{R}^{n\times m}$ , defines the inter-sample rotation angle

$F(t)=\int_{t_0}^{t} f(x)\,dx$ 0

and chooses a single filter length $F(t)=\int_{t_0}^{t} f(x)\,dx$ 1 as the double average distance between subsequent extrema of $F(t)=\int_{t_0}^{t} f(x)\,dx$ 2. This shared $F(t)=\int_{t_0}^{t} f(x)\,dx$ 3 is then used in FFT-based iterative filtering across all channels, which enforces frequency alignment of the extracted multivariate modes. The paper proves finite-time convergence to a prescribed tolerance, states complexity $F(t)=\int_{t_0}^{t} f(x)\,dx$ 4, and reports that MvFIF is robust to noise perturbation, produces a quasi-dyadic filterbank on white Gaussian noise, and remains stable when the number of channels increases considerably (Cicone et al., 2019).

MvFIF is not formulated as an explicit denoiser, but its empirical anti-noise behavior is substantial. On four-channel white Gaussian noise, the paper reports a quasi-dyadic filterbank and good frequency alignment across channels. On additive Gaussian-noise tests with SNR pairs $F(t)=\int_{t_0}^{t} f(x)\,dx$ 5, $F(t)=\int_{t_0}^{t} f(x)\,dx$ 6, $F(t)=\int_{t_0}^{t} f(x)\,dx$ 7, and $F(t)=\int_{t_0}^{t} f(x)\,dx$ 8 dB, it reports that MvFIF was “perfectly able” to reconstruct the clean oscillatory components and trends while separating noise contributions. The same study gives strong runtime evidence for the “fast” label: for 100 realizations of four-channel white Gaussian noise with 1000 samples, MvFIF took $F(t)=\int_{t_0}^{t} f(x)\,dx$ 9 s whereas MEMD took $t_i$ 0 s (Cicone et al., 2019).

3. Time-frequency and surrogate-based decompositions

A major anti-noise branch works directly in the time-frequency plane. Nonlinear Mode Decomposition (NMD) decomposes a signal $t_i$ 1 into physically meaningful nonlinear modes $t_i$ 2, where each nonlinear mode is a phase-locked harmonic family

$t_i$ 3

Its robustness comes from combining WFT or WT ridge extraction, adaptive resolution selection, harmonic consistency testing, and surrogate-data tests that distinguish deterministic oscillations from random activity. NMD uses time-shifted surrogates to validate harmonics and FT surrogates to decide when the residual should be regarded as noise, so it explicitly stops before decomposing pure noise into meaningless components. The paper presents NMD as a noise-robust adaptive decomposition method and reports qualitative and quantitative superiority over EMD, EEMD, Karhunen–Loève expansion, and ICA in the examples studied (Iatsenko et al., 2012).

Enhanced Time-Frequency Representation and Mode Decomposition (ETFR-MD) is more specialized: it targets multicomponent FM signals with closely spaced, spectrally overlapped, or crossing instantaneous frequencies under low SNR. The method models the noisy signal as

$t_i$ 4

extracts initial IF curves by path optimization with energy, continuity, and curvature penalties, then applies a low-complexity kernel phase averaging (KPA) enhancement scheme derived from the sinusoidal time-frequency distribution without explicitly computing the STFD. Its computationally cheap parts are the KPA enhancement and STFT-based IA extraction, with stated complexities $t_i$ 5 for KPA and $t_i$ 6 for STFT, while the initial IF search is the heavier stage. The method is explicitly motivated by anti-noise mode separation rather than generic denoising, and its central contribution is stable IF and IA extraction in adverse environments (Zhang et al., 2020).

A more recent non-iterative formulation is time-frequency mode decomposition (TFMD). TFMD starts from the hypothesis that constituent modes form contiguous high-energy regions in the spectrogram whereas noise produces diffuse, fragmented, low-coherence regions. It computes an STFT, smooths the non-negative-frequency magnitude spectrogram with a $t_i$ 7 rectangular mean filter, thresholds it with

$t_i$ 8

using $t_i$ 9, then performs 8-connected component labeling and size filtering with $E(t)$ 0 and $E(t)$ 1. Each surviving connected component becomes a binary mask for inverse-STFT reconstruction. The paper reports total complexity $E(t)$ 2, automatic mode-number estimation in all six clean synthetic cases, and particularly strong performance in high-noise conditions: for input SNRs below roughly $E(t)$ 3 dB, the output SNR is generally higher than the input SNR, while at higher input SNR the method reaches an intrinsic reconstruction-error floor (Zhou et al., 16 Jul 2025).

Taken together, these time-frequency methods show three distinct anti-noise strategies: surrogate-based discrimination of deterministic structure, enhancement of ridge geometry before amplitude recovery, and spectrogram-domain segmentation that removes small fragmented components likely due to noise. A plausible synthesis is that they shift the decomposition problem from repeated signal-domain sifting toward explicit discrimination of coherent time-frequency structure.

4. Variational, latent, and jump-aware multivariate decompositions

Variational Latent Mode Decomposition (VLMD) addresses anti-noise fast mode decomposition in the multichannel setting by moving the decomposition from channel space into a lower-dimensional latent space. It assumes

$E(t)$ 4

where $E(t)$ 5 is a sparse coefficient matrix, $E(t)$ 6, and each latent mode is an AM-FM oscillation centered around a common $E(t)$ 7. The variational objective combines reconstruction fidelity, $E(t)$ 8-sparsity on $E(t)$ 9, and VMD-style narrow-band frequency regularization, solved with an ADMM-style alternating scheme and latent-space Fourier-domain updates. The paper states that VLMD is robust to noise, less sensitive to overestimating $E(t_i)=F(t_i)$ 0, and “almost twice as fast as MVMD” on average; it further reports that MVMD deteriorates rapidly as noise increases, whereas VLMD remains reliable in the synthetic scenarios considered (Morante et al., 23 May 2025).

Successive Jump and Mode Decomposition (SJMD) and its multivariate extension (SMJMD) incorporate a different robustness mechanism: they jointly model AM-FM modes, jump components, and a residual that explicitly contains yet-to-be-extracted modes and noise. At each extraction step the signal is written as

$E(t_i)=F(t_i)$ 1

and the objective combines narrow-band mode extraction, residual separation around the target center frequency, a sparsity-promoting penalty on $E(t_i)=F(t_i)$ 2, and quadratic data fidelity. The method does not require the number of modes $E(t_i)=F(t_i)$ 3 in advance; instead it extracts one mode at a time, updates the jump component by ADMM, and stops when the incremental energy of a new mode falls below threshold. The paper argues that this yields both computational and robustness gains, because noise and distortion can remain in the residual rather than being forced into oscillatory modes or jumps (Nazari et al., 11 Apr 2025).

SMJMD also exploits shared center frequency $E(t_i)=F(t_i)$ 4 across channels, which enforces frequency alignment. In synthetic multichannel tests with additive white Gaussian noise, the paper reports correlation coefficients $E(t_i)=F(t_i)$ 5, $E(t_i)=F(t_i)$ 6, and $E(t_i)=F(t_i)$ 7 for SMJMD at noise levels $E(t_i)=F(t_i)$ 8, $E(t_i)=F(t_i)$ 9, and $e(t)=E'(t)$ 0, respectively, all above the corresponding MJMD values. In an ECG experiment with electrode-motion artifact and resulting SNR of $e(t)=E'(t)$ 1 dB, SJMD achieved CC $e(t)=E'(t)$ 2 and MSE $e(t)=E'(t)$ 3 for EDR recovery, and CC $e(t)=E'(t)$ 4 and MSE $e(t)=E'(t)$ 5 for jump recovery; the same study reports elapsed times of $e(t)=E'(t)$ 6 s for SJMD and $e(t)=E'(t)$ 7 s for JMD on the reported setup (Nazari et al., 11 Apr 2025).

These variational and latent approaches differ from fast filtering and TFR segmentation by explicitly parameterizing what should count as shared structure, residual structure, or jump structure. This suggests that anti-noise performance improves when the decomposition space itself is structured to absorb nuisance variability rather than merely smoothing it away.

In the DMD literature, anti-noise mode decomposition begins with the observation that standard DMD is analytically biased by sensor noise. "Characterizing and correcting for the effect of sensor noise in the dynamic mode decomposition" shows that standard DMD under additive sensor noise is biased toward spurious decay and proposes three corrections: a direct noise-corrected DMD using known noise properties, forward-backward DMD (fbDMD), and a total-least-squares-inspired algorithm (tlsDMD). The paper further shows that the bias does not vanish by merely adding more snapshots, and that fbDMD and tlsDMD are the most practical defaults when noise statistics are unknown (Dawson et al., 2015).

A complementary survey evaluates practical anti-noise tools inserted before or within DMD. It experimentally tests RPCA with ADM, RPCA with inexact ALM, and TLS-DMD on three PDE-generated datasets, and concludes that there is no universally best anti-noise method across all datasets and SNR conditions. Inexact ALM is described as “very fast” for solving RPCA relative to some alternatives, but in the reported tests it often produces very low-rank outputs that can over-filter the true dynamics, whereas TLS-DMD is best on the FitzHugh–Nagumo dataset and ADM-RPCA is strongest on parts of the NLSE regime (Chowdhury et al., 2021).

Kalman Filter DMD (KFDMD) reframes DMD as sequential parameter estimation: the entries of $e(t)=E'(t)$ 8 are treated as hidden states and updated by a Kalman filter, with a fast implementation based on truncated POD. The paper states that KFDMD can estimate eigenmodes more precisely than standard DMD or tlsDMD under severe noise if the nature of the observation noise is known, while tlsDMD works better at low and medium noise levels; it also emphasizes that KFDMD naturally extends to time-varying systems because process noise in the parameter dynamics lets it track changes in $e(t)=E'(t)$ 9 online (1804.00143).

Optimized consistent DMD (OCDMD) pushes robustness further by jointly learning DMD eigendynamics and an explicit noise matrix, while enforcing both forward and backward predictive consistency over multiple time horizons. In reduced POD coordinates the observed snapshots are modeled as $d(t)=f(t)-e(t)$ 0, and the optimization minimizes combined forward and backward multi-step prediction losses over $d(t)=f(t)-e(t)$ 1. On laminar flow past a cylinder, the paper reports that more than $d(t)=f(t)-e(t)$ 2 of the randomly generated white noise is identified correctly if $d(t)=f(t)-e(t)$ 3, that vanilla DMD reconstruction error is close to $d(t)=f(t)-e(t)$ 4 under the reported noisy benchmark, and that OCDMD keeps reconstruction error below $d(t)=f(t)-e(t)$ 5 even at the lowest SNR, albeit with increased computational cost (Weiner et al., 2024).

Weak-DMD attacks noise through weak projection rather than explicit noise variables. It reconstructs the measured time signal in a trial space,

$d(t)=f(t)-e(t)$ 6

forms weak DMD matrices $d(t)=f(t)-e(t)$ 7 by integration against test functions, and thereby moves differentiation away from the noisy data and onto smooth basis functions. The paper states that weak-DMD filters noise and precludes timestep considerations, but also notes that optimized-DMD is generally faster, so its advantage is robustness rather than computational speed (Bennett et al., 15 Apr 2026).

Comprehensive Robust DMD (CR-DMD) extends robustification to the entire DMD pipeline, from mode extraction to dimensional reduction. It first decomposes the observation $d(t)=f(t)-e(t)$ 8 as $d(t)=f(t)-e(t)$ 9 by solving a convex preprocessing problem with spatio-temporal total variation, an $2$0-ball for sparse corruption, and an $2$1-ball for dense noise, then applies standard DMD to the denoised data, and finally solves a second convex problem for sparse, observation-referenced amplitude estimation. Both stages are solved by a preconditioned primal-dual splitting method. On cylinder wake and channel-flow datasets corrupted by mixed noise, the paper reports that CR-DMD consistently outperforms the robust DMD baselines used in terms of mode accuracy and fidelity of low-dimensional representations (Nakamura et al., 16 Jan 2026).

A related but differently motivated branch is High Order DMD (HODMD) with Kernel Density Spectrum. HODMD uses delay embedding,

$2$2

to capture temporal complexity exceeding spatial complexity, and KDS converts sparse modal estimates into a spectrum-like object. The paper presents HODMD as a high-resolution modal/spectral decomposition tool for damped, noisy, transient mechanical vibrations and reports frequency errors below $2$3 Hz in the eight-mode synthetic benchmark; however, it does not position HODMD as faster than FFT in raw computation, only as faster in the sense of extracting interpretable modal content from short noisy records (Tuor et al., 2023).

6. Comparative assessment, limitations, and research outlook

The literature does not support a single universal meaning of “anti-noise fast mode decomposition.” Instead, “fast” refers to different algorithmic decisions in different subfields: no repeated sifting in FastIMD, FFT-diagonalized filtering in MvFIF, non-iterative spectrogram segmentation in TFMD, lower-dimensional latent optimization in VLMD, or reduced-coordinate implementations in DMD variants. Likewise, “anti-noise” may mean transition-mode avoidance, structural smoothing in the integral domain, residual modeling, surrogate rejection of random activity, connected-component pruning in the spectrogram, explicit identification of a noise matrix, or convex separation of dense and sparse corruption (Lu, 2011, Cicone et al., 2019, Zhou et al., 16 Jul 2025, Weiner et al., 2024).

Several recurrent limitations are equally clear. Some fast decomposers are only indirectly anti-noise: FastIMD offers plausible robustness through integral-domain smoothing but no formal additive-noise benchmark (Lu, 2011). Some noise-robust decomposers are not actually fast in computational terms: weak-DMD, OCDMD, and CR-DMD all trade additional optimization or projection machinery for stability (Weiner et al., 2024, Bennett et al., 15 Apr 2026, Nakamura et al., 16 Jan 2026). Some strong multivariate methods depend on shared latent or frequency structure across channels, so their advantage may diminish when such structure is absent, as VLMD itself notes through the reduction to conventional MMD when $2$4 (Morante et al., 23 May 2025). Time-frequency segmentation and ridge methods depend on TF separability and parameter choice; TFMD may merge severely overlapping but non-crossing modes, and ETFR-MD pays a computational price for initial IF path estimation (Zhou et al., 16 Jul 2025, Zhang et al., 2020).

A common misconception is that robustness and denoising are identical. Several methods surveyed here are primarily decomposition engines rather than denoisers in the strict sense. MvFIF is best characterized as a fast, adaptive multivariate decomposition with strong empirical robustness to noise, not as an explicit denoising criterion (Cicone et al., 2019). The same qualification applies to HODMD and, in a different way, to ETFR-MD, whose focus is accurate IF and IA recovery under noise rather than minimax denoising (Tuor et al., 2023, Zhang et al., 2020).

Taken together, these works suggest a set of design principles for future anti-noise fast mode decomposition. Robust methods tend to postpone hard decisions until after smoothing, averaging, or projection; they exploit shared structure across time, channels, or prediction horizons; they retain an explicit residual so that unexplained variability is not forced into the modes; and they include a stopping or selection mechanism that prevents decomposition of pure noise. This synthesis is inferential rather than a statement of any single paper, but it is consistent with the progression from EMD noise diagnostics to modern latent, variational, time-frequency, and DMD-based frameworks (Kaslovsky et al., 2010, Iatsenko et al., 2012, Nazari et al., 11 Apr 2025, Nakamura et al., 16 Jan 2026).