Fast Linear Regression Method Estimates Wasserstein Distances with Unprecedented Efficiency
Researchers have introduced a novel, highly efficient method for estimating the computationally intensive Wasserstein distance between multiple pairs of probability distributions. The technique, detailed in a new arXiv preprint, leverages a fast linear regression model trained on sliced Wasserstein (SW) distances to predict the true Wasserstein distance, bypassing the need for costly optimal transport calculations for every new pair. By combining both lower-bound (standard SW) and upper-bound (lifted SW) predictors, the model achieves high accuracy from a remarkably small number of training examples, offering a significant speed-up for applications in machine learning and data analysis.
Bridging the Gap with Sliced Wasserstein Predictors
The core innovation of this work is framing the estimation of the full Wasserstein distance as a supervised learning problem. Instead of computing the distance directly—a process with cubic complexity—the method learns a linear relationship between easily computed SW distances and the target Wasserstein metric. The researchers propose two model variants: an unconstrained linear model with a straightforward least-squares solution and a more parsimonious constrained model that uses only half the parameters while maintaining performance. Once trained on a curated set of distribution pairs, the model can predict distances for new pairs via a simple, efficient linear combination of their SW values.
Empirical Validation Across Diverse Domains
The method's efficacy was rigorously tested across a suite of challenging tasks. Empirical validation covered Gaussian mixture models, point-cloud classification, and visualizations in Wasserstein space for 3D data. The model was evaluated on high-dimensional real-world datasets including MNIST point clouds, ShapeNetV2 3D models, MERFISH Cell Niches in spatial transcriptomics, and scRNA-seq data. In all cases, the regression-based estimator provided a superior approximation of the Wasserstein distance compared to the current state-of-the-art embedding model, Wasserstein Wormhole, especially in low-data regimes where sample efficiency is critical.
Accelerating Existing Frameworks with RG-Wormhole
The utility of this fast estimator extends beyond standalone distance calculation. The researchers demonstrated that it can be integrated into existing pipelines to drastically reduce training time. By employing their regression model to approximate distances during the learning phase, they accelerated the training of the Wasserstein Wormhole framework, resulting in a new, more efficient model dubbed RG-Wormhole. This synergy shows the method's potential as a plug-in component for enhancing the scalability of other optimal transport-based algorithms.
Why This Matters: Key Takeaways
- Computational Breakthrough: This method transforms the expensive calculation of Wasserstein distances into a fast prediction task, enabling its use in large-scale, iterative applications like hyperparameter tuning or online learning.
- Sample Efficiency: The linear models achieve high accuracy after training on only a small number of distribution pairs, making them practical for domains where data or computational budget is limited.
- Broad Applicability: Successful validation on diverse data types—from image-derived point clouds to complex biological data—proves the estimator's robustness and general utility across machine learning and scientific computing.
- Synergistic Potential: The creation of RG-Wormhole illustrates how this estimator can be used not just in isolation, but to speed up the training of other advanced models that rely on optimal transport, multiplying its impact.