Robust Modal Decomposition (RMD)

• Robust Modal Decomposition (RMD) is a framework that extracts physically meaningful modal components from noisy and corrupted time series and spatio-temporal data.
• It integrates variance maximization, bandwidth penalization, and robust estimators to effectively suppress noise and spurious modes.
• RMD improves signal-to-noise ratios and stability in applications like ECG analysis, radar sensing, and nonlinear system diagnostics.

Robust Modal Decomposition (RMD) is a general framework for extracting physically meaningful modal components from noisy or corrupted time series and spatio-temporal data. By integrating variance maximization, physical constraints (bandwidth control), and algorithmic robustness, RMD overcomes key weaknesses of classical empirical, variational, and spectral decomposition techniques. The RMD paradigm encompasses recent advancements in Hankel trajectory analysis, robust dynamic mode decomposition, penalization strategies, and complex-valued formulations, with particular utility in time-frequency analysis, nonlinear system diagnostics, and phase-resolved sensing.

1. Foundational Principles

RMD originated from the need to combine interpretability, noise resistance, and spectral control in modal decomposition for time series and high-dimensional observations. Traditional algorithms fall into two broad categories: optimization-based approaches like Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD), and spectral/linear algebraic methods like Singular Spectrum Analysis (SSA)/Principal Component Analysis (PCA). EMD and VMD tend to generate spurious modes due to lack of physical constraints, while SSA rapidly degrades under noise or strong nonlinearity due to its sensitivity to high-frequency disturbances and lack of explicit bandwidth or smoothness control (Hao et al., 27 Oct 2025).

The core innovation of RMD lies in regularizing the modal decomposition process—either by penalizing "rough" (broadband or noisy) modes directly in the optimization or by embedding robust estimators and constraints into the spectral extraction pipeline. This is implemented by mapping input signals into trajectory matrices (Hankel embeddings), constructing Gram matrices that encapsulate global autocorrelation, and solving generalized eigenvalue problems with explicit bandwidth control to enforce smooth, physically interpretable modes.

2. Mathematical and Algorithmic Construction

Let $x[n],~n=1,\dots,N$ be a real (or complex) signal. The RMD methodology first forms a $K$-dimensional time-delay Hankel matrix $X \in \mathbb{F}^{K \times L}$ (with $L = N-K+1$, $\mathbb{F}=\mathbb{R}$ or $\mathbb{C}$), defined as: $X_{i,j} = x[i+j-1]$ The Gram matrix

$$G = X^T X \in \mathbb{R}^{L \times L} \quad \text{or} \quad X X^T \in \mathbb{R}^{K \times K}$$

captures the global structure.

A difference operator $D\in \mathbb{F}^{(K-1)\times K}$ is constructed for discrete derivative regularization, yielding $R=D^T D$, which acts as a bandwidth penalty. RMD seeks vectors $K$0 solving: $K$1 with bandwidth penalty $K$2. The stationarity condition yields the generalized eigenproblem: $K$3 for real-valued RMD (Hao et al., 27 Oct 2025), and with transpositions replaced by Hermitian conjugates for complex RMD formulations (Hao et al., 3 Nov 2025). Modal reconstructions are performed by computing $K$4, followed by diagonal averaging ("hankelization") to recover the time series for each mode.

Advanced RMD variants extend this backbone by (i) incorporating clustering/merging of similar eigenvectors (based on similarity threshold $K$5) to reject spurious or split modes; (ii) adapting to complex-valued observational data (CRMD); and (iii) supporting higher-order difference penalties for further regularization of nonsinusoidal or nonstationary modes.

3. Robustness and Noise Rejection Mechanisms

RMD achieves robust mode extraction primarily through two mathematically synergistic features:

• Bandwidth penalization: This suppresses high-frequency (broadband) noise, as modes with large roughness $K$6 experience eigenvalue shrinkage $K$7, thereby naturally discriminating against noise-dominant modes (Hao et al., 27 Oct 2025).
• Global phase-space structure: By mapping the signal into its trajectory Gram matrix, RMD exploits global, rather than purely local, autocorrelation—which is less sensitive to isolated outliers.

Perturbation theory indicates that, relative to classical SSA/PCA, RMD reduces the variance of noise contamination (by a factor $K$8) while introducing minimal bias for narrowband physical signals. Theoretical results guarantee uniqueness (up to phase), stability under small perturbations, and effective outlier rejection. Complex extensions (CRMD) apply the same principle to data with phase structure, naturally combining amplitude and quadrature smoothness for electromagnetic, radar, or analytic signal decomposition scenarios (Hao et al., 3 Nov 2025).

4. Comparison with Related Modal Decomposition Techniques

Method	Physical Constraints	Bandwidth Control	Noise Resistance	Ad hoc Parameters
EMD	No (sifting, empirical)	None	Low	Mode count/stop
VMD	Yes (narrowband variational)	Strong	Moderate	Penalty/center
SSA	Yes (phase autocorrelation)	None	Low (under noise)	None
RMD	Yes (phase + bandwidth)	Tunable	High	$K$9, $X \in \mathbb{F}^{K \times L}$0

RMD combines the interpretability and global dynamics preservation of SSA/PCA with the explicit bandwidth narrowing of VMD, but without requiring iterative sifting or delicate initialization of center frequencies. Empirical benchmarks indicate that in scenarios with heavy noise or nonlinearity (SNR as low as –15 dB), RMD outperforms both VMD and SSA: for instance, it cleanly separates three superimposed sinusoids with less than 0.2 Hz frequency shift, where alternatives either miss or distort key modes (Hao et al., 27 Oct 2025).

In application to nonlinear AM/FM signals and physiological data (ECG, respiration from radar echoes), RMD demonstrates enhanced mode isolation, improved SNR (by 4–8 dB versus SSA), and reduced spurious component rates (<5%, as opposed to 20%+ for EMD/VMD) (Hao et al., 27 Oct 2025).

5. Extensions: Complex, Multivariate, and Domain-Specific RMD

Complex Robust Modal Decomposition (CRMD) generalizes all real-valued algebra to the complex domain, employing Hermitian Gram matrices and difference penalties. The mode extraction and regularization pipeline remains unchanged in structure but operates on analytic or quadrature-rich signals. Experimental validation on IQ radar, bivariate vibration, radio-frequency identification, and WiFi Channel State Information demonstrates large improvements in both modal SNR and classification accuracy compared to unconstrained variants or real-valued band-reducing methods (Hao et al., 3 Nov 2025).

Further, the RMD principle underpins modern robust DMD (Dynamic Mode Decomposition) frameworks, both via penalty-augmented fitting for various noise models (e.g., mixed/dense/sparse/multiplicative) (Abolmasoumi et al., 2021, Nakamura et al., 16 Jan 2026, Lee et al., 2021), and as a preprocessing step in robust principal component analysis pipelines (RPCA), which decompose the observed data into low-rank coherent structure and sparse/dense corruptions, stabilizing downstream modal extraction (Scherl et al., 2019).

Residual DMD, described in (Colbrook et al., 2022), complements RMD by equipping standard DMD/EDMD algorithms with rigorous, data-driven residual checks—filtering out unphysical, spuriously isolated eigenmodes, and providing guaranteed error bounds and convergence rates.

6. Computational Considerations and Practical Implementation

Computational complexity is dominated by the formation and eigendecomposition of the Gram matrix. For a time series of length $X \in \mathbb{F}^{K \times L}$1 and embedding dimension $X \in \mathbb{F}^{K \times L}$2 (typically $X \in \mathbb{F}^{K \times L}$3 in practice), the total computational cost is $X \in \mathbb{F}^{K \times L}$4. Memory requirements scale with $X \in \mathbb{F}^{K \times L}$5 and $X \in \mathbb{F}^{K \times L}$6 (Hao et al., 27 Oct 2025, Hao et al., 3 Nov 2025).

No alternating direction methods or outer optimization loops are required; the problem reduces to a generalized Hermitian (or symmetric) eigenproblem, solvable by standard dense linear algebra routines once parameters are fixed. Parameter selection for $X \in \mathbb{F}^{K \times L}$7, $X \in \mathbb{F}^{K \times L}$8, and mode similarity threshold $X \in \mathbb{F}^{K \times L}$9 can benefit from cross-validation or automated selection (e.g., via ML or Bayesian model selection). Mode clustering post-processing assists in rejecting repeated roots and consolidating split spectral branches.

CRMD and multivariate/complex extensions do not substantially increase algorithmic cost, as the principal objects grow in field (real to complex), not in size. Bandpass or higher-order difference penalties can be systematically incorporated by augmenting $L = N-K+1$0 accordingly.

Accelerated solvers and memory-efficient strategies remain areas of active investigation for embedded or real-time deployment (e.g., radar DSP), as do extensions to jointly learning optimal regularization in datastream or adaptive settings (Hao et al., 27 Oct 2025).

7. Limitations, Applications, and Ongoing Developments

RMD is highly effective for signals comprising underlying smooth, narrowband, or physically structured components embedded in heavy noise or exhibiting wide SNR ranges. Its main limitations currently pertain to:

• Sensitivity to parameter selection (embedding dimension $L = N-K+1$1, penalty $L = N-K+1$2, clustering $L = N-K+1$3), though cross-validation offers practical mitigation.
• $L = N-K+1$4 complexity, motivating future research on fast/approximate eigensolvers for long signals or large-scale sensor arrays.
• Out-of-sample and online mode tracking, which are at present not natively supported in the baseline framework.

In applied domains, RMD has demonstrated superior performance in physiological signal analysis (ECG, PCG, radar vital signs), mechanical fault detection (bearing data), high-noise communications, and nonlinear AM/FM synthesis. It enables interpretable decoding of phase and amplitude in complex-valued sensing (CRMD), robust detection in radar/RF/CSI-based systems, and improved automation in spectrum tracking. Bandpass-variant and domain-specific extensions are straightforward and actively pursued for embedded and high-throughput scenarios (Hao et al., 27 Oct 2025, Hao et al., 3 Nov 2025).

---

Selected References:

• "Rmd: Robust Modal Decomposition with Constrained Bandwidth" (Hao et al., 27 Oct 2025)
• "CRMD: Complex Robust Modal Decomposition" (Hao et al., 3 Nov 2025)
• "Robust Dynamic Mode Decomposition" (Abolmasoumi et al., 2021)
• "Comprehensive Robust Dynamic Mode Decomposition from Mode Extraction to Dimensional Reduction" (Nakamura et al., 16 Jan 2026)
• "Robust Principal Component Analysis for Modal Decomposition of Corrupt Fluid Flows" (Scherl et al., 2019)
• "Residual Dynamic Mode Decomposition: Robust and verified Koopmanism" (Colbrook et al., 2022)
• "An Optimized Dynamic Mode Decomposition Model Robust to Multiplicative Noise" (Lee et al., 2021)