Abstract
The finite STFT Synchrosqueezing transform is a time-frequency analysis method that can decompose finite complex signals into time-varying oscillatory components. This representation is sparse and invertible, allowing recovery of the original signal. The STFT Synchrosqueezing transform on finite dimensional signals has the advantage of an efficient matrix representation. This article defines the finite STFT Synchrosqueezing transform and describes some properties of this transform. We compare the finite STFT and the finite STFT Synchrosqueezing transform by applying these transform to a set of signals.
Introduction: Sharper Signals through Finite Synchrosqueezing
Time-frequency analysis is a cornerstone of signal processing, yet it often suffers from a "foggy window" effect. Standard tools like the Short-Time Fourier Transform (STFT) spread a signal's energy across the time-frequency plane, creating a blurry representation that obscures the underlying instantaneous frequency. While the Synchrosqueezing Transform (SST) was developed to sharpen this image by reassigning energy, most existing theory focuses on continuous signals.
In this work, Mohammadpour et al. bridge the gap by introducing a mathematically rigorous Finite Synchrosqueezing Transform based on the STFT. This discrete version is not only sharper and sparser than the standard STFT but also fully invertible and computationally efficient, making it ideal for digital signal processing.
The Challenge of Time-Frequency Resolution
The fundamental challenge in time-frequency analysis is the trade-off between time and frequency resolution, governed by the uncertainty principle. When analyzing an oscillatory signal—such as a bird song or a seismic wave—the STFT smears the energy of each component over several frequency bins.
While continuous SST can "squeeze" this energy back into a sharp line, real-world data is discrete and finite. Applying continuous theory to discrete data often involves messy approximations. The authors address this by building a version of SST from the ground up using the language of finite-dimensional Hilbert spaces and numerical linear algebra.
Methodology: The Finite STST
The "Finite STST" is constructed through three main steps:
- Discrete STFT: A standard discrete STFT is performed on a cyclic group $\mathbb{Z}_N$, providing a matrix-friendly representation.
- Instantaneous Frequency (IF) Estimation: For every point in the time-frequency plane, the "true" local frequency is estimated by calculating the complex logarithm of the ratio of STFT values at adjacent time steps. A rounding function then maps this to the discrete frequency grid.
- Synchrosqueezing (Reassignment): Energy is moved (or "squeezed") from the original STFT bins to the estimated IF bins. If multiple bins point to the same IF, their energies are summed.
Crucially, this transform is invertible. By summing the squeezed coefficients over a small frequency band around the estimated IF, individual oscillatory components of the original signal can be perfectly recovered.
Experimental Results and Key Insights
The algorithm was evaluated using synthetic test signals, including single-component chirps and multi-component signals with interlacing frequencies.
Key findings include:
- Visual Clarity: The Finite SST (
ISynSTFTI) produced significantly thinner, more accurate lines for the instantaneous frequency compared to the "blurry" standard STFT (ISTFTI). - Mathematical Rigor: The authors provide a stability theorem proving that if noise is bounded, the error in IF estimation remains small.
- Sparsity: The "squeezing" process naturally leads to a sparse representation where energy is concentrated only where the signal actually exists.
- Noise Robustness: Even with significant noise, the Finite SST successfully identified underlying signal components that were difficult to interpret in the standard STFT.
Conclusion
The Finite Synchrosqueezing Transform offers a powerful, discrete alternative to traditional time-frequency methods. By providing a sparse, high-resolution, and invertible representation, it paves the way for more precise analysis of complex oscillatory signals in fields like geophysics, climate modeling, and medical signal processing.
Summary updated