Paper review · October 28, 2026
Synthetic Data Augmentation for Functional Data: Wasserstein Risk Bounds and Diffusion-Based Conditional Generation
The authors augment curve data with a class-conditional functional diffusion generator and derive upper and lower risk bounds based on the Wasserstein-1 distance, yielding improved classification and regression performance and a criterion for choosing the number of synthetic samples.
Problem
Functional data in the form of curves is costly to collect, so sample sizes of tens to hundreds are common. Existing theory on synthetic data augmentation assumes finite-dimensional data and relies on the total variation distance, which provides little information in function spaces.
Method
Synthetic curves are produced with a class-conditional functional diffusion generator, and the guarantees are stated in terms of the Wasserstein-1 distance. For stable downstream estimators, the risk of training on synthetic curves is bounded above and below by the risk of training on the same number of real curves plus a generation-error term. From this, the authors derive the optimal number of synthetic samples and the conditions under which synthetic data outperforms real data. For class-imbalanced settings, they bound the excess risk of training on real and synthetic curves together without sample splitting, and use this bound to give a criterion for allocating synthetic curves.
Results
- The authors report that, in applications to temperature, hyperspectral and gait curve data, augmentation improved performance on classification, function-to-scalar regression and function-to-function regression.
- In class-imbalanced settings, they report that the largest gains came from allocating synthetic curves mainly to minority classes.
- Theoretically, they derived the optimal number of synthetic samples and the conditions under which synthetic data becomes more advantageous than real data.
Limitations
This analysis is based on the abstract. The input text contained only the title and author information, so specific numbers, experimental setup and comparison baselines could not be checked. The theoretical guarantees assume a "stable" downstream estimator, and the applications are limited to three types of data.
Why it matters
In fields with little curve data, such as sensor time series, spectra and biosignals, it gives a theoretical basis for deciding how much synthetic data to add, and to which classes, when expanding data with generative models.
Key terms
- Functional data
- Data in which each observation is not a vector of numbers but an entire curve over time, wavelength or a similar variable.
- Wasserstein-1 distance
- A distance that measures the difference between two distributions as the minimum cost of moving one distribution onto the other. It allows meaningful comparisons even in function spaces.
- Total variation distance
- The largest difference in the probabilities two distributions assign to the same event. In continuous function spaces it is almost always close to its maximum value, so it carries little information.
Original
Authors: Yuhao Qi, Wenlin Dai, Xuehu Zhu · Posted 2026-10-25 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.