Abstract
We provide statistical analysis methods for samples of curves when the image but not the parametrisation of the curves is of interest. A parametrisation invariant analysis can be based on the elastic distance of the curves modulo warping, but existing methods have limitations in common realistic settings where curves are irregularly and potentially sparsely observed. We provide methods and algorithms to approximate the elastic distance for such curves via interpreting them as polygons. Moreover, we propose to use spline curves for modelling smooth or polygonal Fréchet means of open or closed curves with respect to the elastic distance and show identifiability of the spline model modulo warping. We illustrate the use of our methods for elastic mean and distance computation by application to two datasets. The first application clusters sparsely sampled GPS tracks based on the elastic distance and computes smooth means for each cluster to find new paths on Tempelhof field in Berlin. The second classifies irregularly sampled handwritten spirals of Parkinson's patients and controls based on the elastic distance to a mean spiral curve computed using our approach. All developed methods are implemented in the \texttt{R}-package \texttt{elasdics} and evaluated in simulations.
Introduction
Functional Data Analysis (FDA) typically focuses on analyzing data that can be represented as functions or curves. In many applications, the primary interest lies in the geometric shape—the image—of the curve rather than its specific temporal or spatial parametrization. For instance, in handwriting analysis or GPS tracking, the speed at which a path is traversed (the "phase") is often a nuisance variable, while the path itself (the "amplitude") contains the relevant information.
Elastic analysis provides a framework to handle this by using the Fisher-Rao Riemannian metric, which allows for "warping" or re-parametrizing curves to align their shapes. However, standard elastic methods often assume dense, regularly sampled data. This paper addresses the common real-world scenario where data is sparse, irregular, or "messy," providing a robust framework for computing distances and means in these challenging conditions.
Methodology
The proposed framework leverages the Square-Root Velocity (SRV) transform, which maps curves into a space where the elastic distance corresponds to the standard $L_2$ distance. To handle sparsity, the authors introduce two key innovations: polygonal interpretation and SRV splines.
Polygonal Interpretation and Alignment
Rather than assuming a continuous smooth function, the method treats discretely observed points as vertices of a polygon with constant speed between them. This discretization transforms the infinite-dimensional warping problem into a finite optimization task.
To find the optimal alignment between two such polygons, the authors developed a Coordinate-Wise Optimization (CWO) algorithm. Unlike Dynamic Programming (DP), which is restricted to fixed grids, CWO offers the flexibility needed to handle the irregular spacing inherent in sparse datasets.
SRV Splines for Fréchet Means
The Fréchet mean represents the "average" shape of a set of curves. To ensure the mean is both smooth and representative without overfitting to the noise of sparse observations, the authors model the mean using linear splines at the SRV level. This approach provides a natural regularization, ensuring the resulting mean curve is parsimonious. Crucially, the paper provides theoretical proofs that these spline models are identifiable modulo warping, ensuring a unique and well-defined representation for shapes in higher dimensions.
Key Insights and Applications
The core strength of this approach is its ability to separate shape from timing. By aligning curves elastically, researchers can isolate the "pure" shape of the data.
GPS Track Analysis
The researchers applied their methods to 55 GPS tracks from Berlin's Tempelhof Field. These tracks, containing only 15 to 45 waypoints each, were successfully clustered into distinct paths. The resulting smooth means provided accurate representations of common walking routes, demonstrating the method's utility for automated mapping and urban planning.
Parkinson’s Disease Diagnostics
In a medical context, the framework was used to analyze handwritten Archimedes spirals. By downsampling the data to 10% density, the authors simulated the sparse conditions often found in remote monitoring. They discovered that while the shape of the spiral is a useful diagnostic tool, the warping function—the adjustment needed to align a patient's tremor-filled drawing to a mean spiral—offered significant additional information about the patient's motor control.
Results
The proposed methods demonstrated superior performance in both simulations and real-world applications:
- Alignment Efficiency: In sparse data simulations ($m=30$ points), the CWO algorithm reduced average elastic distances by 33% for open curves and 26% for closed curves compared to traditional DP methods.
- Classification Accuracy: The elastic distance-based approach achieved a 92.5% accuracy in distinguishing Parkinson’s patients from healthy controls, a significant improvement over the 80% accuracy achieved by standard $L_2$ methods.
Conclusion
This work provides a comprehensive toolkit for the elastic analysis of sparse and irregularly sampled curves. By combining polygonal representations with spline-based means, the authors offer a mathematically sound and practically applicable solution for "low-resolution" functional data. The implementation of these methods in the R package elasdics makes this robust framework accessible to the scientific community for applications ranging from geoinformatics to medical diagnostics.
Summary updated