When we want to find the "average" of several data points, our first instinct is to use the arithmetic mean. In a Euclidean space, this works perfectly: the average of two points is simply the point halfway between them. However, when our data points are themselves probability distributions—such as images, point clouds, or sensor noise profiles—this simple averaging fails catastrophically.
If you average two images of a digit "1"—one shifted to the left and one to the right—the arithmetic mean (pixel-wise average) results in a blurry image with two faint "ghost" ones. It doesn't represent a "1" at all. To solve this, researchers use the Wasserstein Barycenter, a geometric average that respects the underlying spatial structure of the data by "transporting" mass rather than just blending intensities.
The Geometry of Optimal Transport
The core problem with Euclidean averaging is that it ignores the ground metric of the space. As noted by Emmanuel Boissard and Thibault Le Gouic in Distribution's template estimate with Wasserstein metrics, traditional averaging fails to capture the "structural essence" of distributions subject to geometric transformations.
The Wasserstein distance (or Optimal Transport distance) measures the minimum "work" required to transform one distribution into another. A Wasserstein Barycenter is the distribution that minimizes the sum of squared Wasserstein distances to a set of input distributions. Instead of blurring features, this approach moves them. In the case of the shifted "1"s, the Wasserstein barycenter would be a single, sharp "1" located in the center.
Overcoming the Computational Bottleneck
While conceptually elegant, computing these barycenters is notoriously difficult. The problem is often non-smooth and high-dimensional. Early methods struggled with "blurring" artifacts caused by entropic regularization, a challenge addressed by Edward Chien and Justin Solomon in Stochastic Wasserstein Barycenters, who proposed a stochastic algorithm to recover sharp geometric structures.
Recent breakthroughs have focused on different computational paradigms:
- Decentralized Networks: Cesar A Uribe and Darina Dvinskikh developed algorithms for agents to reach a consensus on a global barycenter using only local interactions, achieving optimal convergence rates in Distributed Computation of Wasserstein Barycenters Over Networks.
- Neural Networks: To handle high-dimensional data like images, Jiaojiao Fan and Yongxin Chen introduced Scalable Computations of Wasserstein Barycenter via Input Convex Neural Networks, which allows for the computation of barycenters for any combination of weights in a single training session.
- Functional Approaches: Zebang Shen proposed Sinkhorn Barycenter via Functional Gradient Descent, optimizing over function spaces to solve problems in 100 dimensions—previously thought to be computationally impossible.
Applications: From Bayesian Inference to Image Processing
The ability to find a geometrically faithful average has transformative implications across fields. In Bayesian statistics, the WASP (Wasserstein Posterior) framework by Sanvesh Srivastava and David B. Dunson allows for "divide-and-conquer" inference. By running parallel MCMC on data subsets and merging them via Wasserstein barycenters, they maintain the structural integrity of the posterior that simple averaging would destroy WASP: Scalable Bayes via barycenters of subset posteriors.
Similarly, the "analysis problem"—finding the weights needed to reconstruct a target distribution from a set of references—was solved by Matthew Werenski and colleagues in Measure Estimation in the Barycentric Coding Model, enabling fast and consistent image inpainting and feature interpolation.
By treating distributions as points in a curved geometric space rather than flat vectors, Wasserstein barycenters provide a mathematically rigorous way to aggregate information without losing the very features that make the data meaningful.