Abstract
This article presents a two-stage topological algorithm for recovering an estimate of a quasiperiodic function from a set of noisy measurements. The first stage of the algorithm is a topological phase estimator, which detects the quasiperiodic structure of the function without placing additional restrictions on the function. By respecting this phase estimate, the algorithm avoids creating distortion even when it uses a large number of samples for the estimate of the function.
Introduction
Recovering signals that are "almost" periodic but subject to unknown warping—such as musical vibrato, frequency-modulating chirps, or ocean wave textures—presents a significant challenge in signal processing. When these quasiperiodic signals are corrupted by noise, standard linear filters often fail. Because the signal lacks a constant spectral shape, traditional filtering techniques typically blur internal structures or fail to accurately recover the signal's amplitude.
This paper introduces the Quasiperiodic Low Pass Filter (QPLPF), a novel two-stage algorithm designed to denoise these signals by exploiting their hidden geometric structure. Unlike traditional methods that average over temporal or spatial neighborhoods, the QPLPF identifies the underlying topological "phase space" to average across structurally similar points, preserving the signal's true shape and amplitude.
Methodology: The Topological Approach
The QPLPF operates in two distinct stages to ensure high-fidelity signal recovery:
- Topological Estimation (Phase Space Identification): Instead of treating the signal as a simple sequence, the algorithm employs delay immersion. By comparing the signal to translated copies of itself, it maps the data into a higher-dimensional space. A distance-based graph is then constructed to connect points that are topologically similar, effectively identifying points representing the same "phase" of the underlying warped wave.
- Neighborhood Averaging (Filtering): Denoising is performed by averaging the values of the $S$ nearest neighbors as defined by the topological graph. This "class averaging" approach ensures that even points far apart in time or space can be averaged if they are structurally identical.
The authors provide a rigorous mathematical framework, proving that smooth functions possess "quasiperiodic factorizations." This ensures a Zero Distortion Guarantee: in a noiseless scenario with an appropriately chosen neighborhood size, the QPLPF recovers the original signal perfectly, as the distortion from mismatched filtering (Stage 1 error) vanishes.
Experimental Results
The efficacy of the QPLPF was evaluated across synthetic and real-world datasets, demonstrating significant advantages over traditional adaptive filters.
2D Image Denoising
The algorithm was tested on simulated 2D quasiperiodic images. As shown in Figure 5, the QPLPF effectively removes noise while maintaining the integrity of the rippling patterns, which would typically be blurred by standard spatial filters.
1D LFM Chirp and Amplitude Stability
On a discretized Linear Frequency Modulated (LFM) chirp with additive white Gaussian noise, the QPLPF significantly outperformed standard adaptive variable-bandwidth filters. While traditional filters struggle to maintain a stable amplitude envelope when tracking changing frequencies, the QPLPF recovers the amplitude reliably by respecting the signal's underlying geometry.
Real-World Maritime SAR Data
The filter was also applied to a $100 \times 100$ pixel Synthetic Aperture Radar (SAR) image of ocean swells. The QPLPF successfully removed speckle noise and enhanced wave contrast. Furthermore, analysis of the 2D Fast Fourier Transform (FFT) showed that the filter removed horizontal streak artifacts, clarifying the spatial frequency and direction of the swells.
Discussion and Limitations
While the QPLPF offers superior performance, particularly at low Signal-to-Noise Ratios (SNR), it faces certain practical constraints. The current 2D implementation is computationally intensive, limiting extensive statistical validation to 1D signals. Additionally, performance depends on the neighborhood size ($S$); a trade-off exists between minimizing structural distortion (Stage 1 error) and maximizing noise reduction (Stage 2 error).
Future work may focus on optimizing the algorithm for large-scale applications, such as cryo-electron microscopy, leveraging the universal quasiperiodic factorization framework established here.
Summary updated