TL;DR
The Wasserstein-Fourier (WF) distance offers a novel way to compare stationary time series by measuring how their spectral energy "moves" across frequencies. By applying Optimal Transport (Wasserstein distance) to normalized power spectral densities, the authors create a metric that is robust to frequency shifts and differing spectral supports, outperforming traditional "vertical" distances in classification, interpolation, and dimensionality reduction tasks.
What Problem Does This Paper Address?
Traditional methods for comparing time series in the spectral domain—like Euclidean distance, Kullback-Leibler (KL) divergence, or Itakura-Saito (IS) distance—are "vertical" metrics. They compare the power of two signals at the exact same frequency.
This approach has a major flaw: if two signals have similar shapes but are slightly shifted in frequency (e.g., two bird calls at different pitches), vertical distances will see them as completely different because their "peaks" don't overlap. Furthermore, these metrics often require the signals to share the same "support" (the same set of frequencies), making them brittle when dealing with real-world noise or varying acquisition conditions. The authors aim to provide a "horizontal" distance that quantifies the displacement of energy, effectively capturing how much one spectrum must be "warped" to match another.
Methods
The core of the paper is the Wasserstein-Fourier (WF) distance. The process involves:
- Normalization: Converting a signal's Power Spectral Density (PSD) into a Normalized PSD (NPSD), which treats the distribution of energy like a probability distribution. This makes the metric invariant to the signal's total power and time shifts.
- Optimal Transport (OT): Calculating the $W_2$ Wasserstein distance between these NPSDs.
- Efficiency: Because NPSDs are one-dimensional distributions, the Wasserstein distance has a closed-form solution using inverse cumulative distribution functions (CDFs), making it computationally efficient (linear complexity).
The authors extend this distance into three frameworks:
- Geodesic Interpolation: Creating "in-between" signals that smoothly transition frequencies rather than just fading one signal out and the other in.
- Principal Geodesic Analysis (PGA): A version of PCA designed for the non-linear Wasserstein space, allowing for dimensionality reduction that respects the geometry of spectral energy.
- Spectral Classifiers: Equipping standard models like K-Nearest Neighbors (KNN) and Logistic Regression with the WF distance.
Datasets
The authors validate their approach across several diverse datasets:
- Synthetic Signals: Cosine and sinc pulses with random frequencies to test dimensionality reduction and separability.
- C. elegans Behavioral Database: 1D time series representing the movements of wild-type worms, used for data augmentation/interpolation.
- Fungi Melt Curves: rDNA internal transcribed spacer (ITS) region curves for 18 fungal species.
- UrbanSound8k: A large dataset of urban sounds (sirens, drilling, etc.). The authors specifically focused on discriminating "gun shots" from other classes.
- GunPoint: Part of the UCR archive, containing hand trajectories of actors either drawing a gun or just pointing.
Key Insights and Ideas
- Horizontal vs. Vertical: The most intuitive "aha" moment is the distinction between vertical and horizontal distances. A vertical distance measures the difference in height at a fixed point; a horizontal distance measures the distance energy has to travel to reach a new point.
- Equivalence Classes: By focusing on NPSDs, the WF distance naturally creates an "equivalence class" for time series. Signals that are shifted in time or scaled in magnitude are treated as identical, which is often desirable in pattern recognition where the "shape" of the oscillation matters more than when it started.
- Sparse Interpolation: Unlike Euclidean interpolation, which creates "ghosting" (a blurry superposition of two signals), WF interpolation moves the energy peaks. If you interpolate between a 10Hz signal and a 20Hz signal, the WF midpoint is a 15Hz signal, whereas the Euclidean midpoint is a weak 10Hz signal mixed with a weak 20Hz signal.
Key Results
- Classification Accuracy: On the UrbanSound8k dataset, the WF-based KNN classifier was the top performer in 4 out of 9 tasks. An aggregated model combining Euclidean, KL, and WF distances outperformed all single-distance models, suggesting WF captures unique features others miss.
- Linear Separability: In the synthetic tests, PGA using the WF distance was able to linearly separate cosine signals from sinc signals, a task where standard Functional PCA (FPCA) failed.
- Robustness: The WF distance proved far more effective for the GunPoint dataset (body motion), where the "Young vs. Old" and "Draw vs. Point" classifications showed tighter error bars and higher accuracy compared to Euclidean or IS distances.
Limitations and Future Work
- Phase and Magnitude Loss: By design, the WF distance ignores phase information and absolute magnitude. While this provides invariance, it means the distance cannot distinguish between signals that have the same energy distribution but different temporal structures (e.g., a chirp vs. stationary noise with the same average spectrum).
- Computational Overhead: While 1D Wasserstein is fast, the authors acknowledge that WF still takes longer to compute than Euclidean distance (e.g., 3.7 seconds vs 0.1 seconds for KNN training in their tests).
- Future Directions: The authors suggest exploring "unbalanced" optimal transport to compare signals without the need for normalization, and developing new training procedures for Gaussian Processes based on the geodesic paths discovered here.
Summary updated