Enhanced TFR-MD: Robust Time-Frequency Decomposition

• Enhanced TFR-MD is a comprehensive framework that unifies adaptive time–frequency representation with mode decomposition to analyze nonstationary, multimodal signals.
• The method employs advanced techniques like STFT-based IF extraction, kernel-based phase averaging, and convex-regularized updates to robustly separate overlapping oscillatory components.
• Empirical results show that ETFR-MD achieves sharper time–frequency concentration and superior noise robustness compared to classical approaches such as EMD, SST, and VMD.

Enhanced Time–Frequency Representation and Mode Decomposition (ETFR-MD) unifies advanced time–frequency analysis with robust mode decomposition for nonstationary, multimode, and low signal-to-noise ratio (SNR) scenarios. ETFR-MD builds on and extends classical linear TFRs (e.g., STFT), empirical mode decomposition, multiresolution and sparse random methods, and convex-regularized mode decompositions, delivering adaptive recovery and separation of time-varying oscillatory components in complex signals. The framework provides principled solutions for instantaneous frequency (IF) and amplitude (IA) extraction, accurate reconstruction of overlapping/crossing modes, and adaptivity to intra-wave modulation and shape diversity.

1. Conceptual Foundations and Multiresolution Modeling

The theoretical heart of ETFR-MD is the use of generalized signal models that go beyond simple AM–FM representations. A signal $f:[0,1]\to\mathbb R$ is expressed as a superposition of multiresolution intrinsic mode functions (MIMFs), each with an expansion

$$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$

where $\phi(t)$ is a smooth, strictly increasing phase function, $\{a_n\}, \{b_n\}$ are multiresolution coefficients, and $\{s_{cn}\}, \{s_{sn}\}$ are $2\pi$-periodic shape functions (Yang, 2017). This representation captures not only the primary oscillatory structure but also time-dependent waveform modifications, enabling the modeling and extraction of intra-wave modulation and non-sinusoidal wave shapes.

The explicit use of periodic shape functions and multiple harmonics allows sharp energy concentration in the time–frequency plane, yielding enhanced TFRs compared to classical approaches where energy is often smeared due to wave-shape variability.

2. Algorithmic Implementations and Core Procedures

A typical ETFR-MD pipeline integrates several advanced signal processing procedures, which can be specialized or hybridized according to the application domain:

A. Short-Time Fourier Transform (STFT) and Path-Based IF Estimation

ETFR-MD is often built on the STFT,

$$X(t, \omega) = \int_{-\infty}^\infty x(\tau) w(\tau - t) e^{-j\omega\tau} d\tau,$$

providing the basic TFR. For multi-component signals with close or crossing IFs, ETFR-MD formulates IF extraction as a penalized path optimization in the $(t, \omega)$ plane, balancing STFT magnitude, continuity, and curvature terms to robustly trace true ridges—even under spectral overlap or low SNR. The output is a set of IF trajectories, refined iteratively (Zhang et al., 2020).

B. Mode Enhancement, Separation, and Amplitude Recovery

To further denoise and disambiguate the initially extracted IFs, ETFR-MD deploys kernel-based phase-averaging (KPA) on encoded analytic signals. For each IF, a quadratic kernel accumulates phase information to maximize concentration along the ridge, averaging to suppress cross-mode leakage. Instantaneous amplitudes are estimated via a synchro-extracting operator (SEO) that projects STFT energy along these refined ridges, yielding invertible TFRs and mode reconstructions.

C. Recursive Multiresolution and Gauss–Seidel Updates

In the multiresolution setting, each mode is identified by sequentially demodulating along $\phi_k(t)$ (phase unwinding), folding samples to exploit periodic shape symmetries, and performing partition-based regression to estimate shape functions. The Gauss–Seidel updating ensures that each newly estimated component is subtracted immediately from the residual, improving separation and convergence (Yang, 2017).

D. Sparse Random Feature Approaches and Clustering

Alternative ETFR-MD methods, such as Sparse Random Mode Decomposition (SRMD), generate sparse approximations to the spectrogram using random time–frequency windowed atoms and solve for sparse coefficients via $\ell_1$-regularized problems (e.g., LASSO). The localization of nonzero atoms in the $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$0 plane facilitates density-based clustering (DBSCAN) for automatic mode separation, with support for extensions such as chirp atoms, adaptive window sizing, and hybridization with variational mode decomposition (Richardson et al., 2022).

E. Convex-regularized DMD Extensions

Extensions to Dynamic Mode Decomposition (DMD) with time-varying amplitudes allow each spatial mode to have an adaptively weighted activation envelope, regulated by $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$1 (sparsity) and $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$2 (smoothness) penalties. The resulting convex optimization yields mode decompositions and TFRs that are sensitive to transient and intermittent activity, something not feasible with standard stationary DMD (Tanaka et al., 14 Aug 2025).

3. Comparison to Classical and Competing Frameworks

ETFR-MD unifies and surpasses limitations of prior model classes:

• Fixed-template time–frequency methods (STFT, wavelets) are limited by the Heisenberg uncertainty, suffering from spectral smearing when components are non-sinusoidal or strongly nonstationary.
• Empirical Mode Decomposition (EMD) and its multivariate extensions assume AM–FM structure but cannot capture intra-wave shape modulation, often leaving oscillatory residuals.
• Synchrosqueezing (SST) and variants reassign energy in the TFR but, for non-sinusoidal components, can exhibit energy spreading across multiple ridges and fail to offer invertible or directly reconstructible decompositions in challenging cases.
• Naive sparse and random approaches can miss closely-spaced or crossing modes without specialized path extraction or enhancement.

ETFR-MD, especially the multiresolution and sparse random variants, achieves sharper TFR ridges, superior mode separability (including at IF crossings), and recovers physical intra-wave content. It accommodates signals with minimal separation, strong nonlinearity, and low SNR, where classical and even some modern methods fail (Zhang et al., 2020, Yang, 2017, Richardson et al., 2022).

4. Theoretical Guarantees and Practical Performance

Key mathematical properties and empirical results are established:

• Identifiability: Successful decomposition requires phase functions $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$3 to satisfy well-differentiation/separation conditions (minimum distance in IF trajectories), with explicit coupling bounds ensuring convergence in Gauss–Seidel or regression-based updates (Yang, 2017).
• Regression risk: Partition-based regression error is $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$4 for $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$5 samples and bin size $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$6.
• Convergence: Under suitable bounds (e.g., $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$7), mode updates converge linearly to the ground-truth up to the regression noise floor.
• Robustness to noise: Additive noise is averaged out by regression, yielding reliable convergence as the sample count grows.

Performance metrics include mean square error (MSE), output SNR, whiteness tests, ridge compactness, and mode reconstruction errors. Empirically, ETFR-MD outperforms SST, SET, VMD, and EMD across benchmarks featuring closely-spaced/crossing IFs, heavy noise, and real-world signals such as ECG, PPG, and gravitational waveforms (Zhang et al., 2020, Yang, 2017, Richardson et al., 2022).

5. Extensions to Multidimensional, Multivariate, and Transient Analysis

ETFR-MD frameworks are adaptable to multidimensional and multivariate data:

• Multivariate EMD (FA-MVEMD): Decomposes $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$8 into spatially robust IMFs with Hilbert spectral analysis, outperforming multi-resolution DMD (mrDMD) in localization, information condensation, and fidelity for transient/intermittent flows in computational fluid dynamics applications (Souza et al., 2023).
• Spatiotemporal Intrinsic Mode Decomposition (STIMD): Factors sensor matrices as $$f(t) = \sum_{n=-N/2}^{N/2-1} a_n \cos(2\pi n\phi(t)) s_{cn}(2\pi N\phi(t)) + \sum_{n=-N/2}^{N/2-1} b_n \sin(2\pi n\phi(t)) s_{sn}(2\pi N\phi(t)),$$9, constraining the temporal modes $\phi(t)$0 to be IMFs while the spatial modes $\phi(t)$1 capture correlated structure. This permits meaningfully sharpened Hilbert spectra and nonlinear future-state prediction, improving mode interpretability for neuroscience and astrophysical data (Hirsh et al., 2018).
• Sparse and smooth DMD for transients: Incorporates regularization to extract dynamically significant mode activations during non-steady regimes, revealing intermittent mode activity that is invisible to stationary DMD (Tanaka et al., 14 Aug 2025).

6. Applications and Benchmarking

ETFR-MD has been validated across a spectrum of signals and domains:

• Biomedical signals: Separation of cardiac and respiratory components in PPG, clean heartbeat isolation in ECG with artifacts, extraction of QRS variability.
• Fluid mechanics: Vortex-shedding, dynamic stall and vortex pairing in airfoils, with spatial modes unambiguously mapping to coherent structures and their time-varying frequencies (Souza et al., 2023, Tanaka et al., 14 Aug 2025).
• Astrophysics: Gravitational-wave chirp extraction with high TFR sharpness and separation from instrumental noise (Hirsh et al., 2018).
• Music and speech: Clear separation and clustering of modes under strong downsampling and noise (Richardson et al., 2022).
• General nonstationary signals: Robust demixing and recovery for signals with multiple coexisting, crossing, or highly modulated components.

A concise summary table of method properties follows:

Method	Adaptive TFR	Handles Closely-Spaced IFs	Intra-wave Modulation	SNR Robustness	Reconstruction
STFT/Wavelet	–	–	–	–	+
EMD/VMD	+	+	–	+	±
SST/SET	+	±	–	±	±
ETFR-MD (MIMF/KPA)	+	+	+	+	+
SRMD (Sparse Rand.)	+	+	±	+	+

7. Practical Considerations and Limitations

ETFR-MD methods are typically nonparametric, requiring no explicit analytic FM law modeling. Window lengths, path-penalty parameters, and sparsity/smoothness weights are set empirically or via cross-validation. The computational complexity is dominated by STFT and convex optimization (random features or $\phi(t)$2 solvers, regression, or Gauss–Seidel inner loops), generally scaling as $\phi(t)$3 for TFRs and $\phi(t)$4 for random feature approaches, with robust convergence under moderate sample rates.

Potential limitations include sensitivity to hyperparameter/tolerance choice, possible need for coarse phase initialization in spatiotemporal variants, and computational overhead in ultra-high-dimensional settings. In practice, ETFR-MD can be combined with or replace ridge-extraction, cluster-based, or variational algorithms, supporting hybrid and domain-specific workflows.

---

Enhanced Time–Frequency Representation and Mode Decomposition provides a comprehensive, mathematically grounded, and empirically verified suite of methods, allowing precise, interpretable, and robust decomposition and spectral analysis of the most challenging nonstationary, multimode signals, outperforming both classical and many recent alternatives in a wide range of practical scenarios (Zhang et al., 2020, Yang, 2017, Richardson et al., 2022, Souza et al., 2023, Hirsh et al., 2018, Tanaka et al., 14 Aug 2025).