Empirical Wavelet Transform: Adaptive Analysis

• Empirical Wavelet Transform is a data-driven time-frequency analysis tool that adapts filter banks to a signal’s Fourier spectrum for precise mode decomposition.
• It employs adaptive frequency partitioning and smooth filter construction to form a tight frame, ensuring exact reconstruction and resilient mode isolation.
• EWT offers superior performance over traditional methods by providing robust, energy-preserving decomposition for nonstationary and multicomponent signals.

The Empirical Wavelet Transform (EWT) is a fully data-driven, adaptive time-frequency analysis framework in which the frequency partition and associated wavelet filter bank are directly tailored to the spectral content of the analyzed signal. Unlike classical wavelet transforms based on fixed dyadic or predetermined frequency tilings, EWT constructs a set of bandpass and lowpass filters whose supports are determined by the actual modes observed in the signal’s Fourier spectrum. The resulting transform provides an adaptive tight frame with exact reconstruction, robust mode isolation, and favorable theoretical properties for both 1D and multidimensional signals. EWT has been influential for mode decomposition, time–frequency analysis, and adaptive processing of real-world nonstationary and multicomponent signals (Gilles, 2024).

1. Signal Model and Frequency Partitioning

EWT operates under the hypothesis that an observed real-valued signal $f(t)$ admits a decomposition

$f(t) = \sum_{k=0}^N f_k(t),$

where each mode $f_k(t)$ is an amplitude-modulated–frequency-modulated (AM–FM) component with compact or localized Fourier support—essentially a generalization of Intrinsic Mode Functions (IMFs) from Empirical Mode Decomposition (EMD) (Gilles, 2024).

Adaptive frequency band segmentation is the first, critical step:

• Compute the magnitude spectrum $F(\omega) = |\widehat{f}(\omega)|$ for $\omega \in [0,\pi]$.
• Detect all local maxima $\{\xi_m\}_{m=1}^M$ in $F(\omega)$.
• Sort maxima by descending $F(\xi_m)$ and select the largest $N-1$ to define $N$ desired modes (user or algorithm selected).
• Sort the selected frequencies in ascending order $\{\eta_1 < \eta_2 < \cdots < \eta_{N-1}\}$, set $\eta_0=0$, $\eta_N=\pi$.
• Define band boundaries: $\omega_k = \frac{1}{2}(\eta_{k-1} + \eta_k)$ for $k=0,\ldots,N$ so that $[0, \pi] = \bigcup_{n=1}^{N} \Lambda_n$, $\Lambda_n = [\omega_{n-1},\omega_n]$.

This partition directly reflects the spectral features of $f$ rather than adhering to a predefined scale arrangement.

2. Construction of the Adaptive Filter Bank

For each detected frequency boundary $\omega_n$, a smooth transition width $\tau_n = \gamma \omega_n$ (with $0<\gamma<1$) is set. The frequency-domain filters are designed using a smooth $C^k([0,1])$ ramp polynomial

$\beta(x) = x^4(35 - 84x + 70x^2 - 20x^3)$

satisfying $\beta(0)=0$, $\beta(1)=1$, and $\beta(x) + \beta(1-x) = 1$.

Empirical scaling and wavelet functions:

• Scaling (lowpass):

$$\widehat{\varphi}_1(\omega) = \begin{cases} 1, & |\omega| \le (1-\gamma)\omega_1, \ \cos\left[\frac{\pi}{2} \beta\left(\frac{|\omega|-(1-\gamma)\omega_1}{2\gamma\omega_1}\right)\right], & (1-\gamma)\omega_1 \le |\omega| \le (1+\gamma)\omega_1, \ 0, & \text{otherwise}. \end{cases}$$

• Wavelets (for $n=1,\ldots,N-1$):

$$\widehat{\psi}_{n}(\omega) = \begin{cases} 1, & (1+\gamma)\omega_n \le |\omega| \le (1-\gamma)\omega_{n+1}, \ \sin\left[\frac{\pi}{2} \beta\left(\frac{|\omega|-(1-\gamma)\omega_{n}}{2\gamma\omega_{n}}\right)\right], & (1-\gamma)\omega_{n} \le |\omega| \le (1+\gamma)\omega_{n}, \ \cos\left[\frac{\pi}{2} \beta\left(\frac{|\omega|-(1-\gamma)\omega_{n+1}}{2\gamma\omega_{n+1}}\right)\right], & (1-\gamma)\omega_{n+1} \le |\omega| \le (1+\gamma)\omega_{n+1}, \ 0, & \text{otherwise}. \end{cases}$$

Inverse Fourier transforms yield the time-domain scaling and wavelet functions.

A crucial theoretical guarantee is the Parseval-tight-frame property, provided $\gamma$ is sufficiently small:

$$|\widehat{\varphi}_1(\omega + 2k\pi)|^2 + \sum_{n=1}^N |\widehat{\psi}_n(\omega + 2k\pi)|^2 \equiv 1, \quad \forall \omega,\, \forall k\in\mathbb{Z}$$

if

$$\gamma < \min_{0\leq n < N} \frac{\omega_{n+1} - \omega_n}{\omega_{n+1} + \omega_n}$$

(Gilles, 2024).

3. Algorithmic Implementation and Computational Complexity

The EWT workflow is algorithmically structured as follows:

1. Compute $F[\omega] = FFT(f)$ and detect all local maxima $\{\xi_m\}$ of $|F[\omega]|$.
2. Select and sort $N-1$ significant maxima as boundary markers.
3. Define the adaptive band boundaries $\{\omega_k\}$.
4. Design filters $\widehat{\varphi}_1(\omega)$, $\widehat{\psi}_n(\omega)$ using the analytic formulas above.
5. Analysis: Compute mode coefficients as

$$c_0[t] = IDFT(F[\omega] \,\widehat{\varphi}_1(\omega))$$

$$c_n[t] = IDFT(F[\omega] \,\widehat{\psi}_n(\omega)),\;\forall n=1,\ldots,N$$

1. Synthesis: Reconstruct

$$\tilde{f}[t] = (c_0 * \varphi_1)[t] + \sum_{n=1}^{N} (c_n * \psi_n)[t]$$

which is exact given the tight frame property.

Computational costs:

• FFT/IFFT: $O(L\log L)$ for length-$L$ signals.
• Local maxima detection and sorting: $O(L)$.
• Filter bank multiplication: $N \times O(L)$. The dominant cost is the FFT and $N$ filter applications per decomposition (Gilles, 2024).

4. Comparison to Empirical Mode Decomposition and Other Approaches

The EWT differs fundamentally from EMD, which applies a nonlinear and iterative sifting procedure to extract IMFs in the time domain. EWT, in contrast, provides:

• Linearity: The decomposition and reconstruction are linear and exact due to the tight frame structure.
• Stability: Small perturbations in $f$ yield proportionally bounded perturbations in coefficients, while EMD shows sensitivity to noise (necessitating modifications like Ensemble EMD).
• Reconstruction Guarantees: EWT's frame guarantees closed-form, norm-preserving reconstruction, unlike EMD, where exact inversion lacks theoretical backing.
• Empirical Mode Separation: EWT exhibits better separation of synthetic and real-world multicomponent signals, while EMD can suffer from mode splitting in the presence of overlapping spectral content (Gilles, 2024).

5. Extensions to Multidimensional and Partition-Generalized EWT

EWT extends naturally to higher-dimensional signals by data-adaptive partitioning of the multidimensional Fourier domain. Several frameworks exist:

• Tensor EWT: Apply 1D EWT along each axis and construct 2D (or higher) tensor-product filter banks (Gilles et al., 2024, Ren et al., 2020).
• Radial and Curvelet-Type EWT: Use pseudo-polar FFTs and adaptive band partitioning in both scale and orientation, yielding empirical versions of Littlewood–Paley wavelets, ridgelets, and curvelets (Gilles et al., 2024).
• General Multidimensional Formulation: Following (Lucas et al., 2024), spectral domains are partitioned into connected regions $\{\Omega_n\}$, filters are constructed by diffeomorphic warping of a mother wavelet’s Fourier support, and frame conditions ensure stable decomposition and reconstruction.

Partitions may be generated by:

• Watershed segmentation of the log-magnitude spectrum (empirical watershed wavelet transform, EWWT) (Hurat et al., 2024).
• Voronoi diagrams for enhanced regularity (empirical Voronoi wavelets) (Gilles, 2024).
• Demons registration or other registration schemes to map empirical harmonic regions to canonical shapes for robust filter construction (Lucas et al., 2024).

The stability and perfect reconstruction are preserved provided the filter bank yields appropriately bounded frame norms.

6. Applications, Advantages, and Limitations

EWT has been demonstrated for:

• Extraction of AM–FM modes from synthetic and real signals, including ECG and seismic data (Gilles, 2024).
• Time–frequency analysis via Hilbert transform on each mode.
• Feature extraction and regularization in image analysis, including texture segmentation and image deconvolution (Hurat et al., 2024, Lucas et al., 2024, Gilles, 2024).
• Flow decomposition in experimental and simulated fluid dynamics (Ren et al., 2020).

Advantages:

• Resolves nonstationary, multicomponent signals with non-dyadic spectral structures.
• The tight-frame guarantees enable exact, robust, and energy-preserving decomposition.
• Adaptation to complex spectra leads to sparser representations in many tasks compared to fixed wavelet decompositions.

Limitations:

• Band-boundary detection can fail when modes overlap spectrally or have similar magnitudes.
• Choice of number of bands $N$ and transition parameter $\gamma$ requires careful selection.
• For highly non-stationary or crossing-frequency signals, EMD may outperform EWT in mode separation.

Open challenges include robust automation of frequency partitioning, extension to non-Euclidean domains, and further theoretical development for irregular or redundant frame constructions (Gilles, 2024).

---

Reference Table: Key Theoretical Properties

Aspect	EWT	EMD
Linearity	Yes (linear filter bank)	No (nonlinear)
Frame property	Tight frame with $\ell^2$-norm	Not a frame
Reconstruction	Exact (with explicit formula)	Practical, not formalized
Stability	Bounded; perturbation-resilient	Sensitive to noise
Partition	Data-driven (spectral maxima)	Data-driven (signals)

---

References:

• J. Gilles, “Empirical Wavelet Transform,” IEEE Trans. Signal Processing, 2013 (Gilles, 2024)
• J. Gilles et al., “2D Empirical Transforms ...,” 2024 (Gilles et al., 2024)
• C.-G. Lucas & J. Gilles, “Multidimensional empirical wavelet transform,” 2024 (Lucas et al., 2024)
• J. Gilles, “Continuous empirical wavelets systems,” 2024 (Gilles, 2024)
• J. Gilles, “Empirical Voronoi Wavelets,” 2024 (Gilles, 2024)
• J. Gilles, “The Empirical Watershed Wavelet,” 2024 (Hurat et al., 2024)
• G. Ren et al., “Image-based flow decomposition using empirical wavelet transform,” 2020 (Ren et al., 2020)
• X. Wang et al., “Compressive Sensing Empirical Wavelet Transform ...,” 2025 (Liu et al., 14 Feb 2025)