Wasserstein Barycenter Fusion

• Wasserstein barycenter-based fusion is a geometric framework that aggregates multiple probability measures into a central distribution using the 2-Wasserstein metric.
• It employs support-adaptive stochastic and primal-dual algorithms to optimize the barycenter computation while preserving sharp structural details.
• Applications range from image fusion and Bayesian inference to decentralized sensor integration, demonstrating its significant practical impact.

Wasserstein barycenter-based fusion denotes a class of methodologies for aggregating multiple probability measures—potentially with different supports, structural properties, and statistical characteristics—into a “central” distribution under the geometry of optimal transport, specifically the 2-Wasserstein metric. These algorithms are foundational in geometric statistics, scalable Bayesian inference, high-dimensional data summarization, decentralized information fusion, and model aggregation. This entry reviews core mathematical principles, algorithmic frameworks, recent advances, and pivotal applications, with emphasis on precise technical formulations and convergence properties.

1. Mathematical Foundations and Problem Statement

Given a finite collection of probability measures $\{\mu_j\}_{j=1}^N$ on a metric space $(\mathcal{X}, d)$ (typically $\mathbb{R}^d$), and nonnegative weights $\{\lambda_j\}$ (with $\sum_j \lambda_j = 1$), the 2-Wasserstein barycenter is defined as

$$\nu^* = \arg\min_{\nu \in \mathcal{P}_2(\mathcal{X})} \sum_{j=1}^N \lambda_j W_2^2(\nu, \mu_j),$$

where $W_2$ is the 2-Wasserstein distance, and $\mathcal{P}_2(\mathcal{X})$ denotes the set of Borel probability measures with finite second moment. Existence and uniqueness are guaranteed under mild regularity (at least one $\mu_j$ absolutely continuous, compact support, etc.) (Álvarez-Esteban et al., 2015, Srivastava et al., 2015).

The barycenter is the Fréchet mean under the Wasserstein geometry and encapsulates a globally optimal compromise between the input measures in terms of transport cost.

2. Unregularized, Support-Adaptive Stochastic Algorithms

Classic approaches to Wasserstein barycenter computation—such as fixed-grid or entropically regularized methods—either restrict the barycenter support or introduce bias via smoothing. An alternative support-adaptive approach is the stochastic algorithm of (Claici et al., 2018), which directly optimizes over the positions of atomic support points $\{x^i\}_{i=1}^m$ representing the barycenter as a uniform empirical measure: $(\mathcal{X}, d)$0 The algorithm alternates:

• Dual ascent: for each $(\mathcal{X}, d)$1, maximize the semi-discrete Kantorovich dual over vectors $(\mathcal{X}, d)$2.
• Support "snap": move each support point $(\mathcal{X}, d)$3 toward the barycenter (weighted mean) of its power cell assignments under all measures, using Monte Carlo to estimate relevant integrals.

Update rules: $(\mathcal{X}, d)$4 where $(\mathcal{X}, d)$5, $(\mathcal{X}, d)$6 are Monte Carlo estimates of cell masses and barycenters.

Key features:

• No entropic regularization: sharp support, structure-preserving barycenters.
• Adaptive: support points move to reflect the true barycenter’s geometry (edges, manifolds, mixture modes).
• Convergence guarantees: local minima to stationary points; rate $(\mathcal{X}, d)$7 for the best $(\mathcal{X}, d)$8-point approximation in $(\mathcal{X}, d)$9 (Claici et al., 2018).

3. Minimax Optimization and Advanced Primal-Dual Flows

Recent algorithms approach the barycenter as a nonconvex-concave saddle point problem, alternating between optimal transport on primal (measure) and dual (potential) variables. The WDHA algorithm (Kim et al., 24 Jan 2025) exemplifies this for discrete densities on fixed grids:

• Kantorovich dual step: update potentials $\mathbb{R}^d$0 in Sobolev geometry ($\mathbb{R}^d$1), using gradient ascent and convex projection ($\mathbb{R}^d$2).
• Primal (barycenter) step: perform Wasserstein gradient descent on $\mathbb{R}^d$3 using the averaged potential gradient.

Formally, the update at iteration $\mathbb{R}^d$4 is: $\mathbb{R}^d$5 where $\mathbb{R}^d$6.

The WDHA yields near-linear computational complexity ($\mathbb{R}^d$7 per iteration for $\mathbb{R}^d$8 grid points), scaling to high-dimensional images, with convergence proven under weak regularity (Kim et al., 24 Jan 2025).

4. Statistical and Bayesian Fusion

Wasserstein barycenter-based fusion is foundational in large-scale Bayesian inference and model aggregation. In the "divide-and-conquer" WASP approach (Srivastava et al., 2015):

1. Data are partitioned, subset posteriors are sampled in parallel.
2. Empirical barycenters of subset posteriors are computed via a linear program, using cost matrices evaluated over pooled sample grids.
3. The fusion accuracy is controlled by posterior contraction rates; under regularity, the error to the full-data posterior decays as $\mathbb{R}^d$9 for Gaussian models, and nearly the optimal $\{\lambda_j\}$0 rate more generally.

This method is agnostic to the parametrization and supports streaming, large-scale, and parallel computation, outperforming classical consensus-MC and semiparametric density-product fusion methods (Srivastava et al., 2015).

5. Algorithmic and Practical Considerations

The implementation of Wasserstein barycenter-based fusion is conditioned by measure type (continuous vs discrete), dimensionality, regularization, and computational cost.

Sampling and Monte Carlo

• For general continuous measures, all integrals (cell masses, barycenter locations) are evaluated via Monte Carlo, with batch sizes $\{\lambda_j\}$1 typically yielding stable performance (Claici et al., 2018).

Initialization

• Support points for semi-discrete or grid-based methods can be initialized by k-means++ on pooled data, uniform grids, or random samples from mixture models.

Parallelization

• Dual updates on each measure $\{\lambda_j\}$2 are independent and can be parallelized (multithreaded or GPU) (Claici et al., 2018).
• In distributed settings, displacement-interpolation-based protocols achieve consensus on the barycenter without central coordination (Cisneros-Velarde et al., 2020).

Complexity

• Adaptive (e.g., WDHA (Kim et al., 24 Jan 2025)): $\{\lambda_j\}$3 per iteration for $\{\lambda_j\}$4 grid points.
• Classic LP or Sinkhorn: $\{\lambda_j\}$5 or worse for $\{\lambda_j\}$6 measures of support $\{\lambda_j\}$7, but regularization (entropic or projection-robust (Huang et al., 2021)) and stochastic approximations mitigate cost and curse of dimensionality.

Regularization and Robustness

• Absence of an entropic term recovers sharper structures but slows convergence and may require more careful step-size control.
• Entropic (Schrödinger) regularization accelerates algorithms but blurs sharp features and limits resolution (Li et al., 4 Feb 2025).

6. Applications: Fusion in Images, Bayesian Posteriors, and Beyond

• Image fusion and morphing: Wasserstein barycenters interpolate between images or modalities, preserving shape, edges, and low-dimensional support (Claici et al., 2018, Kim et al., 24 Jan 2025). Applications include generating super-samples, blue-noise point sets, and multimodal mixtures.
• Bayesian large-scale inference: WASP implements posterior fusion for tractable inference on data splits (Srivastava et al., 2015, Li et al., 4 Feb 2025).
• Distributed sensor fusion: Algorithms transmit only summary statistics (e.g., means, covariances for Gaussians), scale in node count under asynchronous updates, and require minimal coordination to agree on barycenter fusion rules (Cisneros-Velarde et al., 2020).
• Emergent models: Neural network model fusion via barycenters aligns weights and achieves linear mode connectivity (Akash et al., 2022). Shape, graph, and merge-tree analyses leverage generalized barycenter metrics to summarize and cluster ensemble data (Pont et al., 2021).

7. Comparative Analysis and Extensions

Method	Support Adaptivity	Regularization	Computational Regime
Stochastic (1802)	Adaptive	None	Continuous empirical
WDHA (2501)	Grid/discrete	None	Multi-modal, high-dim
WASP (1508)	Discrete/empirical	None	Bayesian posteriors
Entropic (Schrödinger)	Any (tree, grid)	Entropy	High-d, smooth barycenters
Model fusion (2210)	Layerwise	Optional (GW, entropic)	Deep models

• Support-adaptive (unregularized) algorithms admit sharper geometric features, precisely recover barycenters with singular or manifold-supported measures, but are more susceptible to local minima and are costly in very high dimensions.
• Regularized (entropic/Schrödinger) and projection-robust approaches trade exactness for accelerated convergence and robustness to high dimensionality, at the cost of blurring features and introducing bias.
• Decentralized and parallel methods (pairwise interpolation, decentralized Sinkhorn) extend barycenter fusion to large, asynchronous networks (Cisneros-Velarde et al., 2020, Baheri et al., 18 Sep 2025).

• (Claici et al., 2018): Stochastic Wasserstein Barycenters.
• (Kim et al., 24 Jan 2025): Optimal Transport Barycenter via Nonconvex-Concave Minimax Optimization.
• (Srivastava et al., 2015): Scalable Bayes via Barycenter in Wasserstein Space.
• (Li et al., 4 Feb 2025): Multimarginal Schrödinger Barycenter.
• (Cisneros-Velarde et al., 2020): Distributed Wasserstein Barycenters via Displacement Interpolation.
• (Akash et al., 2022): Wasserstein Barycenter-based Model Fusion and Linear Mode Connectivity of Neural Networks.
• (Pont et al., 2021): Wasserstein Distances, Geodesics and Barycenters of Merge Trees.

Wasserstein barycenter-based fusion thus provides a flexible, geometrically principled, and theoretically grounded mechanism for the aggregation and fusion of probability measures, with rapidly developing algorithmic solutions for diverse scientific and engineering domains.