Wasserstein Barycenters for Geometry Preserving Averaging

22 concepts19 papers1 proposal

Imagine you have two images: one contains a sharp circle on the left side of the frame, and the other contains an identical circle on the right. If you compute the standard arithmetic (Euclidean) average of these two images, the result is not a circle in the center. Instead, you get two faint, blurry "ghost" circles.

This phenomenon highlights a fundamental flaw in traditional data aggregation: Euclidean averaging operates point-wise or pixel-by-pixel. It averages intensities at fixed coordinates rather than accounting for the underlying spatial structure of the data. As a result, when aggregating probability distributions, shapes, or sensor heatmaps, standard linear interpolation destroys geometric features, creating blurred artifacts or physically nonsensical multi-modal distributions.

To preserve the structural integrity of data during aggregation, researchers have turned to optimal transport and Wasserstein barycenters. By treating data as probability measures and averaging them based on the geometry of moving mass rather than blending local intensities, this framework provides a mathematically rigorous way to compute structure-aware averages.

The Geometry of Moving Mass

The core limitation of Euclidean averaging is that it is geometry-agnostic. Whether averaging shifted images, complex curves representing physical structures (Sebastian Kurtek et al.), or statistical dependence structures (Gautier Marti et al.), point-wise blending fails because it does not align features.

Optimal transport resolves this by introducing a metric space where the distance between two distributions is the minimum cost required to transport the probability mass of one into the other. A Wasserstein barycenter is defined as the Fréchet mean within this space—the distribution that minimizes the sum of squared Wasserstein distances to a set of input distributions.

Instead of fading pixel intensities in and out, a Wasserstein barycenter physically slides mass across the domain. Returning to the example of the two circles, an optimal transport average smoothly moves the mass from the left and right sides to meet precisely in the middle, maintaining a single, perfectly sharp circle. As demonstrated by Giulia Luise et al. and Jinjin Chi et al., this transport-based approach ensures that distinct spatial patterns and modes are shifted rather than merged into a generic cloud.

Overcoming the Computational Bottleneck

Historically, the primary obstacle to using Wasserstein barycenters was their extreme computational cost. Finding the optimal transport plan traditionally requires solving large-scale linear programs, which quickly becomes intractable for high-dimensional data or continuous distributions.

A foundational breakthrough by marco cuturi et al. introduced entropic regularization to the optimal transport problem. By adding an entropy penalty to the objective, the optimization problem becomes strictly convex and can be solved rapidly via Sinkhorn iterations. This formulation transformed a computationally heavy linear programming task into a fast, GPU-friendly matrix scaling problem, making geometric data analysis practical at scale.

However, entropic regularization introduces a trade-off: the entropy term acts as a smoothing operator, which can reintroduce blurring artifacts into the barycenter. To maintain absolute sharpness, several optimization strategies have emerged:

  • Second-Order Precision: Dongdong Ge et al. demonstrated that second-order interior-point methods can maintain quadratic convergence and high numerical stability, recovering perfectly sharp geometric features where first-order entropic methods lose precision.
  • Continuous Parameterization: To avoid fixing a discrete support grid in advance, Jiaojiao Fan et al. leveraged Input Convex Neural Networks (ICNNs) to parameterize continuous transport maps, allowing the model to learn barycenters for arbitrary weight combinations in a single training session.
  • Functional Optimization: In Sinkhorn Barycenter via Functional Gradient Descent, the authors optimized over a space of functions within a Reproducing Kernel Hilbert Space (RKHS), bypassing discrete grids entirely and allowing barycentric computations to scale linearly with dimension.

Convergence on Curved Manifolds

Because the space of probability measures equipped with the Wasserstein metric is a curved Riemannian manifold, the barycenter objective is inherently non-convex. This curvature makes the global convergence of standard optimization techniques difficult to guarantee theoretically, even though first-order methods perform exceptionally well in practice.

This theoretical gap was bridged by Sinho Chewi et al. for the case of Gaussian distributions (known as the Bures-Wasserstein setting). The authors proved that despite the non-convexity of the space, gradient descent and stochastic gradient descent converge globally to the true barycenter because the objective satisfies a special geometric condition called the Polyak-Łojasiewicz (PL) inequality.

Expanding on this, Jason Altschuler et al. established dimension-free convergence rates for these Riemannian gradient descent algorithms. This guarantee ensures that the computational complexity of finding geometric averages does not explode exponentially as the dimensionality of the data increases, providing a firm theoretical foundation for high-dimensional applications.

Applications and Extensions

The ability to compute structure-preserving averages has unlocked new capabilities across decentralized systems, statistics, and generative modeling.

Scalable Bayesian Inference

In massive data regimes, running Markov Chain Monte Carlo (MCMC) on a single machine is often impossible. The WASP (Wasserstein Posterior) framework proposed by Sanvesh Srivastava et al. solves this via a "divide-and-conquer" strategy. It runs parallel MCMC chains on smaller data subsets and merges the resulting subset posteriors using Wasserstein barycenters. Because the barycenter respects the geometric structure of the probability measures, it avoids the loss of critical structural features and modes that occurs with standard arithmetic pooling.

Decentralized Network Consensus

In distributed systems, agents often need to reach an agreement on a common state or distribution using only local communication. While consensus algorithms are well-understood for Euclidean vectors, Adrian N. Bishop and Cesar A Uribe et al. extended this to probability spaces. They proved that a network of agents can reach a consensus on a global Wasserstein barycenter over time-varying and intermittently connected networks, ensuring that the final agreed-upon distribution retains a sharp, structurally coherent representation of the collective data.

Robust Latent Structure Extraction

When input distributions are corrupted by noise or are poorly aligned, standard geometric blending can still suffer from minor smearing. To isolate the true underlying signal, Elsa Cazelles et al. introduced weak barycenters based on optimal weak transport.

Euclidean Average:      [ Distribution A ] + [ Distribution B ] ---> [ Blurred / Bimodal Artifact ]
Wasserstein Barycenter: [ Distribution A ] -> (Move Mass) -> [ Distribution B ] ---> [ Sharp Geometric Mean ]
Weak Barycenter:        [ Noisy Inputs ] -> (Convex Concentration) -> [ Sharp Latent Structure ]

By leveraging the principle of convex order concentration, weak barycenters seek the most concentrated measure that satisfies transport constraints. Instead of simply blending the inputs, this method filters out noise and outliers to extract the sharp, latent geometric structure common to the entire population.

Key Foundational Work

Fast Computation of Wasserstein Barycenters (2014) — contributed 2 concept elements to this direction.
WASP: Scalable Bayes via barycenters of subset posteriors (2015) — contributed 1 concept elements to this direction.
Analysis of AneuRisk65 data: Elastic shape registration of curves (2014) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 19 papers
  2. Version 1 · 2026-10-04 · 32 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_geometric_distribution_averaging_3abce794,
  title = {geometric distribution averaging},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-distribution-averaging/rdr_3abce79405cb4563a6c6e32f1797dbea},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_geometric_distribution_averaging_3abce794_v2,
  title = {geometric distribution averaging},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-distribution-averaging/rdr_3abce79405cb4563a6c6e32f1797dbea/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake