Variational Mode Decomposition (VMD)

• Variational Mode Decomposition (VMD) is a non-recursive signal decomposition technique that isolates spectrally compact intrinsic modes with explicit frequency tuning.
• The framework formulates a constrained variational optimization solved via ADMM, ensuring precise reconstruction fidelity and robustness against noise and mode-mixing.
• VMD's adaptability to multichannel and spatiotemporal data has enhanced denoising, forecasting, and modal analysis across diverse practical applications.

Variational Mode Decomposition (VMD) is a non-recursive, bandwidth-penalized signal decomposition framework for extracting a prescribed number of spectrally compact intrinsic mode functions (IMFs) from nonstationary data. Unlike recursive approaches such as Empirical Mode Decomposition (EMD), VMD formulates the decomposition as a constrained variational optimization, providing rigorous control over reconstruction fidelity and bandwidth of the resulting modes. The algorithm permits direct frequency-domain updates and explicit tuning of its hyperparameters, making it robust against noise, mode-mixing, and end effects. VMD has become foundational in modal analysis, denoising, forecasting, remote sensing, and multivariate signal decomposition.

1. Mathematical Formulation and Objective Function

VMD seeks $K$ band-limited modes $u_k(t)$, each concentrated around an unknown center frequency $\omega_k$, such that the modes sum to reconstruct the original input $y(t)$ or $f(t)$: $$\min_{\{u_k\}, \{\omega_k\}} \sum_{k=1}^K \|\partial_t[(\delta(t) + j/\pi t) * u_k(t)] e^{-j \omega_k t}\|_2^2$$ subject to

$y(t)=\sum_{k=1}^K u_k(t)$

where $(\delta(t) + j/\pi t) * u_k(t)$ is the analytic signal (via Hilbert transform), and the penalty $\alpha$ governs compactness in the spectral domain around each $\omega_k$ (Ohmichi, 2023).

The full augmented Lagrangian formulation includes quadratic penalties and a Lagrange multiplier $\lambda(t)$: $$\mathcal{L}(\{u_k\}, \{\omega_k\}, \lambda) = \alpha \sum_{k=1}^K \|\partial_t[(\delta + j/\pi t)*u_k] e^{-j\omega_k t}\|_2^2 + \|y - \sum_k u_k\|_2^2 + \langle \lambda, y - \sum_k u_k \rangle$$ This saddle-point construction enables efficient solution via the Alternating Direction Method of Multipliers (ADMM).

2. ADMM Algorithm and Solver Structure

ADMM is the canonical optimization scheme for VMD. For each iteration, the solver alternates:

• Mode update in the Fourier domain:

$$\widehat u_k^{n+1}(\omega) = \frac{\widehat y(\omega) - \sum_{i\neq k} \widehat u_i^n(\omega) + \tfrac{1}{2} \widehat \lambda^n(\omega)}{1 + 2\alpha(\omega-\omega_k^n)^2}$$

• Center frequency update:

$$\omega_k^{n+1} = \frac{\int_0^\infty \omega |\widehat u_k^{n+1}(\omega)|^2 d\omega}{\int_0^\infty |\widehat u_k^{n+1}(\omega)|^2 d\omega}$$

• Dual variable update:

$$\widehat \lambda^{n+1}(\omega) = \widehat \lambda^{n}(\omega) + \tau (\widehat y(\omega) - \sum_k \widehat u_k^{n+1}(\omega))$$

Iterations continue until relative change across modes drops below a small threshold (e.g., $\epsilon=10^{-6}$) (Ohmichi, 2023).

Parameter selection is crucial: the number of modes $K$ and penalty $\alpha$ are chosen to ensure physically meaningful mode separation and adequate reconstruction fidelity. Large $\alpha$ values enforce narrowly-banded, strictly periodic modes; small $\alpha$ allow broader, more nonstationary components. Initialization impacts convergence and stability.

3. Role of Parameters: Number of Modes and Bandwidth Penalty

• Number of modes $K$ directly controls the rank of decomposition. Too few modes amalgamate distinct phenomena; too many introduce spurious narrow-band artifacts with diluted interpretability.
• Bandwidth penalty $\alpha$ tunes spectral compactness. Large $\alpha$ yields sharply periodic, DFT-like atoms; small $\alpha$ allows broad-band, flexible, more nonstationary modal patterns.
• Guideline: Adjust $K$ incrementally and monitor $E = \|y - \sum_k u_k\|_2^2 / \|y\|_2^2$ for adequate reconstruction but retain physically interpretable modes (Ohmichi, 2023, Liao et al., 2022).

4. Multivariate and High-Dimensional Extensions

VMD generalizes to multivariate (MVMD) and spatiotemporal settings:

• MVMD operates on $C$-channel signals, enforcing a common set of center frequencies $\omega_k$ across all channels for each mode. The multichannel objective minimizes

$$\sum_{k=1}^K \sum_{c=1}^C \|\partial_t [u_{k,c}^+(t) e^{-j\omega_k t}]\|_2^2 \quad \text{with} \quad \sum_k u_{k,c}(t) = x_c(t)\:\forall c$$

yielding joint multivariate modes with tight channel alignment (Rehman et al., 2019).

• Spatiotemporal data (VMD-NCS): Dimensionality reduction with Proper Orthogonal Decomposition (POD) precedes multivariate VMD on the coefficient time-series. Reconstruction recovers "intrinsic coherent structures" (ICS) $\psi_k(x,t)$ whose shapes can evolve in both space and time (Ohmichi, 2023).
• Extensions also accommodate windowed (STVMD) and dynamic frequency (window-indexed $\omega_{k,\tau}$) settings, improving adaptability for highly nonstationary signals (Jia et al., 15 Jan 2025).

5. Key Applications and Comparative Evaluation

VMD is extensively used in:

Application Area	Key Benefit	Representative Work [arXiv id]
Nonstationary modal analysis	Coherent structure retrieval in transient flows, spatiotemporal adaption	(Ohmichi, 2023, Liao et al., 2022)
Denoising	Robustness to noise and mode-mixing, end-effect resilience	(Xu et al., 2017, Naveed et al., 2020)
Time-series forecasting	Improved predictability, volatility reduction, leakage-free decomposition	(Putra et al., 2024, Feng et al., 18 Sep 2025, Li, 18 Aug 2025)
Heartbeat/radar signal	Adaptive selection; sub-band isolation; automated K, $\alpha$ optimization	(Gu et al., 16 Feb 2025)
Harmonic/interharmonic detection	Automated band count selection via fractal box dimension	(Yuhang et al., 2024)

Comparisons with EMD demonstrate VMD's mathematical rigor, avoidance of mode mixing, superior convergence, and cleaner feature extraction, especially in noisy and nonstationary contexts (Li, 18 Aug 2025, Xu et al., 2017, Naveed et al., 2020).

6. Automated Hyperparameter Selection and Algorithmic Enhancements

Recent advances address unsupervised determination of $K$ and initial $\omega_k$:

• Globally convergent convex optimization for $K$ and $\omega_k$: By identifying "support baseline" of the spectral function $|\hat f(\omega)|$ via boundary-value solution of a fourth-order ODE and separating significant spectral regions, intrinsic mode count and frequencies are computed prior to VMD (Zhong et al., 4 Jan 2026). This guarantees termination to a global optimum.
• Fractal box dimension criterion: For power harmonics/interharmonics, sweeping $K$ and selecting decomposition where minimum FBD is achieved yields near-optimal layer determination (Yuhang et al., 2024).
• Bilevel Stackelberg-inspired optimization: In forecasting, leader/follower dynamics tune $K$ and $\alpha$ via forecastability and mutual information metrics (Feng et al., 18 Sep 2025).

7. Computational Complexity, Implementation, and Limit Cases

VMD is efficient for moderate $K$ and sample sizes, with each ADMM iteration requiring $O(KN\log N)$ time for FFTs and $O(KN)$ for updates. Typical convergence is reached within $10^2$–$10^3$ iterations. Initialization (random or spectral windows) and frequency grid selection affect performance and convergence speed (Ohmichi, 2023).

• Limit cases:
  • As $\alpha \rightarrow 0$, VMD approaches Proper Orthogonal Decomposition (POD).
  • As $\alpha \rightarrow \infty$ and high $K$, each mode narrows to a single frequency, mimicking DFT (Liao et al., 2022).

Implementations require only FFT routines, frequency-domain algebra, and looped ADMM updates. Empirical studies confirm significant improvements in signal reconstruction error, forecasting accuracy, and physical interpretability across application domains.

---

VMD defines a mathematically rigorous paradigm for nonstationary signal decomposition, with broad extensions to multichannel, spatiotemporal, and adaptive settings. The explicit variational structure underpins its robustness, computational tractability, and suitability for both theoretical investigation and practical engineering use (Ohmichi, 2023, Liao et al., 2022, Jia et al., 15 Jan 2025, Gu et al., 16 Feb 2025, Putra et al., 2024, Feng et al., 18 Sep 2025, Zhong et al., 4 Jan 2026).