Fast Intrinsic Mode Decomposition (FIMD)

• Fast Intrinsic Mode Decomposition (FIMD) is a set of adaptive techniques that extract intrinsic mode functions from nonstationary, nonlinear signals with reduced iterative sifting compared to classical EMD.
• One approach employs a one-pass sawtooth-transform to directly compute the IMF, while another uses iterative residue refinement with median-adjusted control points and cubic splines.
• An EEF-based formulation further refines trend and fluctuation extraction via integral-preserving spline derivatives, broadening FIMD’s application to multidimensional data analysis.

Searching arXiv for Fast Intrinsic Mode Decomposition and closely related mode-decomposition papers. Fast Intrinsic Mode Decomposition (FIMD) denotes a set of intrinsic-mode extraction methods developed within the Hilbert–Huang Transform (HHT) framework for nonlinear and non-stationary data. Across the arXiv record, the term is associated with at least three closely related formulations: a one-pass sawtooth-transform method for direct IMF extraction (0710.3170), a fast convergent iterative method that refines a residue through median-adjusted control points and cubic splines (0808.2827), and an Equivalent Effect Function (EEF)-based FastIMD method in which the trend and fluctuation are obtained from spline derivatives constrained by integral values on control points (Lu, 2011). All three formulations are presented as alternatives to classical Empirical Mode Decomposition (EMD), with the shared objective of reducing the cost, subjectivity, and spline-envelope pathologies associated with repetitive sifting.

1. Position within the HHT and IMF framework

FIMD is rooted in the HHT view that a signal should first be decomposed into intrinsic mode functions (IMFs) and a final residual trend, after which each IMF can be subjected to Hilbert analysis. In the time-series formulation, the Hilbert transform of $X(t)$ is written as

$$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$

with analytic signal

$Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$

amplitude and phase

$$a(t)=\sqrt{X^2(t)+Y^2(t)}, \qquad \theta(t)=\arctan\!\frac{Y(t)}{X(t)},$$

and instantaneous frequency

$\omega(t)=\frac{d\theta(t)}{dt}.$

The IMF conditions restated in the FIMD literature are the standard HHT conditions: over the whole data set, the number of zero crossings and the number of extrema must be equal or differ by at most one, and at every point the mean of the upper envelope and lower envelope must be zero (0710.3170).

The principal target of critique is classical EMD. In the conventional sifting loop, one identifies local extrema, interpolates upper and lower envelopes, averages the envelopes, subtracts the mean, and repeats until a stopping criterion is met. The FIMD papers argue that spline envelopes are not true envelopes in general, that cubic splines can overshoot, that different stopping conditions produce different IMFs, and that repetitive sifting is slow; the 2008 formulation further adds that EMD may fail to separate small oscillations riding on steep edges (0808.2827). The 2011 EEF paper positions FastIMD in the same problem setting for nonlinear, nonstationary data and explicitly frames it as an answer to weaknesses of classical EMD, especially repeated sifting (Lu, 2011).

In this sense, FIMD is not a departure from IMF-based decomposition, but a re-engineering of how the local mean or residue is constructed. The unifying idea is to preserve the adaptive, data-driven spirit of EMD while replacing the standard maxima/minima-envelope sifting loop with a more direct or faster mechanism.

2. Sawtooth-transform FIMD

The earliest explicit arXiv formulation of FIMD is the sawtooth-transform method of 2007, which constructs an IMF in one pass rather than through repeated sifting (0710.3170). Its core idea is to transform the original data function into a piecewise linear sawtooth function, construct the upper envelope by connecting maxima and the lower envelope by connecting minima with straight line segments in the sawtooth space, compute the IMF there as the difference between the sawtooth function and the mean of the upper and lower envelopes, and then reverse-transform the result into the original data space.

If the original signal has extrema at $t_0,t_1,\dots,t_{m-1}$ with values $E(t_i)$, the sawtooth function is defined segmentwise by

$$s(t)=E(t_i)+\frac{E(t_{i+1})-E(t_i)}{t_{i+1}-t_i}(t-t_i), \qquad 0\le i<m-1.$$

The horizontal coordinate is then warped so that the original monotone segment between consecutive extrema is mapped onto the corresponding straight segment: $$u(t)=t_i+ \frac{f(t)-E(t_i)}{E(t_{i+1})-E(t_i)}(t_{i+1}-t_i), \qquad t_i \le t \le t_{i+1}, \quad 0\le i<m-1,$$ with $s(u)=x(t)$. In sawtooth space, the upper and lower envelopes are linear interpolants through maxima and minima,

$$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$0

$$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$1

the residue is

$$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$2

and the IMF is

$$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$3

The outputs in original time are then obtained by composition with the coordinate warp, such as $$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$4 and $$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$5 (0710.3170).

The paper gives substantial implementation detail. It treats flat maxima as a flat maxima ceiling and flat minima as a flat minima floor, and it provides four endpoint extension strategies: even extension, odd extension, cyclic extension, and trend extension. The decomposition stops when the current residual has fewer than two extrema, so there is no iterative sifting stopping criterion for each IMF. An alternative method based on sawtooth function expansion is also presented; it is described as less efficient for 1D data than the main transform method, but potentially easier to extend to 2D or higher dimensions (0710.3170).

The reported evidence is qualitative but concrete. The implementation was written in the D programming language. The examples include original data and last residue, a first IMF with 210 extrema, a second IMF with 85 extrema, a third IMF with 33 extrema, and a fourth IMF with 12 extrema. The paper’s conclusion is that the method processes the data in one pass to obtain a unique IMF component without the time consuming repetitive sifting process of EMD (0710.3170). This suggests that the sawtooth-transform formulation should be understood as the most direct realization of the “fast” claim: it removes inner-loop sifting entirely.

3. Residue-centered iterative FIMD

The 2008 reformulation retains the same objective but changes the mechanism. Instead of transforming the signal into sawtooth space, it iteratively adjusts control points on the data function corresponding to the extrema of a refining IMF, computes control points of the residue function as the median of the straight line segments passing through the data control points, constructs the residue as the cubic spline function of the median points, and defines the refining IMF as the difference between the data function and the improved residue function (0808.2827).

The initialization is distinctive. The initial residue function is constructed as the straight line segments passing through the extrema of the first derivative of the data function; the corresponding points on the original data are inflection points of $$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$6. With initial residue $$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$7, the initial IMF estimate is

$$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$8

At each refinement stage, extrema of the current IMF are found at times $$Y(t) = -\, P \left[\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{X(\tau)}{t-\tau}\,d\tau\right],$$9, and the corresponding points on the original data,

$Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$0

become control points. A turning direction is defined from consecutive control-point segments by a vector product,

$Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$1

and when neighboring turning directions change sign, the updated control ordinate is replaced by a median point computed from the original control point and the straight line joining neighboring control points. If no such change occurs, the control ordinate is unchanged. The improved residue is then the cubic spline through the median points, and the refining IMF is again the signal minus that residue (0808.2827).

The stopping criteria are explicitly practical rather than theorem-driven: maximal iteration count, the maximum point distance on successive residue functions falling below a given limit, or that distance having stopped decreasing. The paper argues that satisfactory results can often be obtained in only two iterations, that FIMD builds only one cubic spline per iteration while EMD typically builds two splines per sift, and that the resulting IMF and residue shapes remain stable even with excessive iterations. It further claims reduced spline overshoot impact because “the control points have doubled on the residue than on the envelopes” (0808.2827).

A second major contribution of the 2008 paper is filtering. The proposed filtering method marks extrema on an IMF according to jump time on the increasing or decreasing edges of IMF cycles or according to amplitude limits, groups contiguous marked extrema into disjoint marked lists, reconstructs the part to be filtered out with cubic splines built on locally defined control points, and iterates until no extrema remain marked or the maximum pointwise difference between successive passed functions is below a threshold. The blocked function is then the difference between the original input data and the filtered data. The paper gives an explicit example with the condition “jump time less than 20,” for which riding waves around $Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$2 and between $Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$3 and $Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$4 are filtered out. It also states that the demo program was coded in MS Visual C# 2008 and distributed under a BSD open source license (0808.2827).

The empirical evidence remains qualitative but more detailed than in the 2007 paper. On randomly generated simulated data, the decomposition yields six IMF components plus a final residue, with extrema counts and iteration numbers reported as IMF 1: 147 extrema, 3 iterations; IMF 2: 76 extrema, 2 iterations; IMF 3: 37 extrema, 8 iterations; IMF 4: 19 extrema, 7 iterations; IMF 5: 9 extrema, 4 iterations; and IMF 6: 3 extrema, 5 iterations. A comparison of two initialization schemes shows that derivative-based initialization separates a riding wave around $Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$5, whereas direct initialization from the data function does not (0808.2827). A plausible implication is that the 2008 method should be viewed not merely as a speed improvement over the 2007 method, but as an attempt to improve mode sensitivity on steep edges while avoiding the sawtooth method’s reported leakage of high-frequency ripples into the residue.

4. Equivalent Effect Function and FastIMD

The 2011 paper recasts FastIMD around the Equivalent Effect Function (EEF), which is defined as having the identical integral values on the control points of the original time series data (Lu, 2011). Operationally, the EEF is obtained from the derivative of the spline function passing through the integral values on the control points. By choosing control points with different criteria, the EEF can be used to find the intrinsic mode function (IMF, fluctuation) and the residue (trend), to fit the curve of the original data function, and to take samples on original data with equivalent effect (Lu, 2011).

This formulation retains the FIMD objective of avoiding the repetitive sifting cycle of classical EMD, but the local mean is no longer introduced through sawtooth geometry or median-adjusted residue splines. Instead, it is encoded by an integral-equivalence constraint on selected control points. In the paper’s framing, the distinction between fluctuation and trend is therefore governed by how control points are chosen and how the spline derivative reproduces integral values. This suggests a shift from purely extrema-envelope geometry toward an integral-preserving spline formalism.

The reported applications are broader than in the earlier 1D papers. The paper calculates trend and fluctuation on real stock historical data on different time scales. It also introduces a new approach to extend the EEF to 2D intrinsic mode decomposition to resolve the inter slice non continuity problem, and presents photo image decomposition examples (Lu, 2011). In the context of the FIMD literature, this is the clearest transition from scalar time series toward multidimensional decomposition.

5. Multidimensional extensions and related fast intrinsic-mode frameworks

Later work broadens the technical landscape around FIMD without always using the exact acronym. In flow analysis, the 2023 paper on “fast adaptive multivariate empirical mode decomposition” (FA-MVEMD) is presented as a fast, multidimensional, multivariate extension of EMD and as the closest equivalent in that paper to a fast intrinsic-mode decomposition method for multidimensional flow-field data (Souza et al., 2023). The decomposition is written as

$Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$6

and the study uses a bivariate input in which the first channel is the physical pressure field and the second channel is white noise with 10% of the original data power. In the SD7003 dynamic stall case, 5 IMFs were extracted; in the NACA0012 transitional case, 3 IMFs were extracted. For each IMF and each spatial location, the analytic signal

$Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$7

is computed pointwise, with instantaneous frequency $Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$8. The paper concludes that the combination of a multidimensional EMD with the Hilbert transform provides modes with superior spatial support when compared to mrDMD and condenses a larger amount of information within a single IMF (Souza et al., 2023). Relative to the 2011 EEF-based 2D extension, FA-MVEMD represents a different multidimensional route: multivariate noise-aided decomposition with mode alignment rather than integral-equivalent spline construction.

Other later methods pursue similar practical goals through different computational mechanisms. SRMD, or Sparse Random Mode Decomposition, is explicitly described as a fast, sparse, randomized time-frequency alternative to classical intrinsic mode decomposition methods, not as a variant of EMD in the strict sifting sense (Richardson et al., 2022). It approximates the signal by a sparse linear combination of random localized sinusoids,

$Z(t)=X(t)+iY(t)=a(t)e^{i\theta(t)},$9

with $$a(t)=\sqrt{X^2(t)+Y^2(t)}, \qquad \theta(t)=\arctan\!\frac{Y(t)}{X(t)},$$0, $$a(t)=\sqrt{X^2(t)+Y^2(t)}, \qquad \theta(t)=\arctan\!\frac{Y(t)}{X(t)},$$1, and $$a(t)=\sqrt{X^2(t)+Y^2(t)}, \qquad \theta(t)=\arctan\!\frac{Y(t)}{X(t)},$$2, estimates sparse coefficients by basis pursuit denoising or LASSO, forms the support set $$a(t)=\sqrt{X^2(t)+Y^2(t)}, \qquad \theta(t)=\arctan\!\frac{Y(t)}{X(t)},$$3, clusters the active atoms with DBSCAN, and reconstructs each cluster as one mode. The paper is relevant to FIMD because it pursues the same practical goal—separating a signal into physically meaningful oscillatory components or “intrinsic modes”—but defines modes through sparse time-frequency support segmentation rather than through envelope interpolation or residue refinement (Richardson et al., 2022).

By contrast, the 2024 paper on orthogonal mode decomposition does not mention FIMD at all and cannot be summarized as an explicit FIMD variant (Li et al., 2024). It instead defines mode extraction for finite discrete signals as orthogonal projection of the interpolation function space onto low-frequency or narrow-band subspaces, with a mode characterized as a maximal narrow-band projection whose intrinsic instantaneous frequency remains strictly positive or strictly negative over the full time segment. For FIMD research, it is therefore best understood as a theoretical alternative to spline/sifting-based intrinsic mode extraction (Li et al., 2024).

6. Conceptual interpretation, limitations, and recurring misconceptions

A common misconception is that FIMD names a single canonical algorithm. The arXiv record does not support that view. The 2007 paper presents a one-pass sawtooth-transform method (0710.3170); the 2008 paper presents a fast convergent iterative method centered on median-adjusted residue splines and derivative-based initialization (0808.2827); and the 2011 paper presents FastIMD through the Equivalent Effect Function and integral-preserving spline derivatives (Lu, 2011). The continuity lies in the target—fast extraction of IMF-like fluctuation and residue-like trend—not in a single invariant computational pipeline.

A second misconception is that “fast” always denotes benchmarked asymptotic superiority. In the FIMD papers, the claim is primarily structural. The sawtooth-transform method is fast because it processes the data in one pass and eliminates the time consuming repetitive sifting process of EMD (0710.3170). The residue-centered iterative method is fast because it converges in few iterations, uses one spline instead of two, and offers an easier stop condition through residue-shape stability (0808.2827). The EEF paper presents FastIMD as a new approach intended to address weaknesses of EMD, especially repeated sifting, but the evidence is again methodological and application-based rather than a formal runtime table (Lu, 2011). This suggests that FIMD should be interpreted as a design philosophy of direct or rapidly convergent IMF extraction, not as a single complexity class.

A third issue concerns the meaning of “intrinsic mode.” In the classical FIMD papers, the concept remains close to Huang’s IMF criteria of extrema/zero-crossing balance and zero local envelope mean (0710.3170). In SRMD, by contrast, a mode is a cluster of active coefficients in a sharpened sparse time-frequency representation (Richardson et al., 2022). In orthogonal mode decomposition, a mode is a maximal narrow-band projection with sign-definite intrinsic instantaneous frequency (Li et al., 2024). The terminology is therefore shared, but the operative definitions are not identical.

The limitations reported across the literature are correspondingly method-specific. The sawtooth-transform formulation relies heavily on extrema detection, boundary extension, and treatment of flat regions; the 2007 paper explicitly provides even, odd, cyclic, and trend extension rules (0710.3170). The 2008 residue-centered method depends on extrema of the first derivative, careful endpoint extension, and a local geometric median-point rule, while also acknowledging that revealing all riding waves is not always equivalent to producing the preferred denoised representation (0808.2827). In multidimensional FA-MVEMD, the papers note that IMFs may contain spurious positive-pressure oscillations adjacent to a vortex and that each IMF and the residue has the same size as the original dataset, increasing total data volume (Souza et al., 2023). For later non-FIMD alternatives, the differences are conceptual as much as technical: SRMD abandons envelope symmetry in favor of sparse TF clustering, whereas orthogonal mode decomposition broadens the notion of a mode beyond standard IMF intuition (Richardson et al., 2022); (Li et al., 2024).

Taken together, these works place FIMD within a broader history of adaptive mode decomposition: first as a direct replacement for EMD sifting, then as a residue-centered fast iterative framework, then as an integral-equivalent spline method, and finally as one reference point among multidimensional, sparse, and projection-based intrinsic-mode extraction paradigms.