Abstract
Convolutional Dictionary Learning (CDL) exploits translation invariance, while Non-negative Matrix Factorization (NMF) variants exploit repeating frequency patterns in spectrograms. This paper establishes a formal mathematical bridge between these two paradigms. We demonstrate that a signal admitting a Convolutional Sparse Coding (CSC) decomposition also possesses time-frequency synthesis coefficients factorizable via semi-NMF or complex-NMF. Experimental comparisons on synthetic and audio signals validate these theoretical links and highlight the practical trade-offs of each approach.
Introduction
A fundamental goal in signal processing is the decomposition of complex signals into elementary "building blocks." Two dominant frameworks have emerged:
- Convolutional Sparse Coding (CSC): Operates in the time domain, modeling signals as a sum of shifted "atoms" (short waveforms). This naturally captures translation invariance.
- Non-negative Matrix Factorization (NMF): Operates in the time-frequency domain (spectrograms), decomposing them into spectral templates and temporal activations.
While both aim to identify recurring structures, they have historically occupied different mathematical spaces. This research formalizes the link between them, proving that CSC is essentially equivalent to a specific form of rank-1 factorization in the complex time-frequency plane.
Theoretical Framework: Bridging Time and Frequency
The authors utilize the Gabor Transform (a specific Short-Time Fourier Transform) to bridge these domains.
The Equivalence Theorem
The core contribution is Theorem 1, which proves that a signal formed by the convolution of a filter $d$ and an activation $z$ can be perfectly represented in the time-frequency domain as a rank-1 factorization. The synthesis coefficients $\alpha$ decompose into a frequency component (derived from the filter) and a time component (the activation).
Optimization and Algorithms
Corollary 1 extends this equivalence to the learning process, showing that time-domain dictionary learning is equivalent to semi-NMF or complex-NMF optimization in the time-frequency domain. To implement this, the authors propose Algorithm 1, a hybrid approach using projected gradient descent in the time-frequency domain with a semi-NMF subroutine.
The Role of Redundancy and Phase
A critical insight is that this equivalence strictly holds only under "maximum redundancy"—where the transform has a hop size of 1 sample. As overlap decreases, the models diverge. Furthermore, because the STFT is complex-valued, standard NMF (which ignores phase) is insufficient; Semi-NMF or Complex-NMF must be used to maintain the mathematical link.
Experimental Evaluation
The theoretical findings were tested on two distinct signal types:
- Stationary Signals (Piano): A 15-second recording of four piano notes played in various combinations.
- Non-Stationary Signals (Synthetic Chirps): Signals with time-varying frequency content, designed to test the limits of spectral template-based methods.
Key Results and Discussion
Experimental results confirmed that both CSC and the proposed NMF-based Algorithm 1 can successfully identify stationary motifs like piano notes. However, performance varies significantly with non-stationary signals and varying levels of redundancy.
Performance Insights
- Stationarity: For the piano dataset, both methods performed well, though the NMF approach showed sensitivity to overlap; lower redundancy often led to split motifs or artifacts.
- Non-Stationarity: CSC provided superior separation for synthetic chirps. Semi-NMF struggled with these time-varying structures unless very large window sizes and high overlaps (80%+) were employed.
- Robustness: Interestingly, Itakura-Saito NMF (IS-NMF) proved more robust than semi-NMF at lower redundancy levels (e.g., 50% overlap), even though it lacks the direct mathematical equivalence to the time-domain convolutional model.
Conclusion and Future Directions
This work provides a rigorous mathematical foundation for the long-suspected link between CSC and NMF. While the equivalence is theoretically sound, practical implementation faces challenges. High-redundancy STFTs required for equivalence lead to significant computational overhead. Future research will explore the relationship between CSC and Low-Rank Time-Frequency Synthesis (LRTFS) and investigate how window size relative to atom length affects robustness.
Summary updated