📊 statistics

Robust Wasserstein barycenter

This paper proposes the robust Wasserstein barycenter (RWB) to overcome the sensitivity to outliers and moment assumptions of the classical Wasserstein barycenter, establishing its theoretical guarantees and demonstrating its superior robustness through extensive numerical experiments on simulated and real-world data.

Original authors: Zixiong Cheng, Hang Liu

Published 2026-03-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zixiong Cheng, Hang Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding the "Average" When Things Go Wrong

Imagine you are a chef trying to create the perfect "average" soup based on recipes from 100 different restaurants.

In the world of statistics and machine learning, this "average soup" is called a Barycenter. It's a way to find the central point or the "typical" version of a group of things (like images, stock prices, or data distributions).

For a long time, mathematicians have used a specific tool called the Wasserstein Barycenter to do this. It's like a very smart, geometric ruler that measures how different two soups are and finds the one that is closest to all of them.

The Problem:
The classic "smart ruler" has a major flaw: it is easily tricked by noise.

If 99 restaurants make a delicious tomato soup, but one restaurant accidentally drops a whole bucket of hot sauce into their pot (an "outlier"), the classic ruler gets confused. Because it tries to be mathematically perfect, it thinks, "Wow, that hot sauce is a huge difference! I must move the average soup recipe way over to the spicy side to accommodate it."

The result? Your "average" soup becomes inedible because of one bad apple.

The Solution: The "Robust" Barycenter

The authors of this paper, Zixiong Cheng and Hang Liu, invented a new tool called the Robust Wasserstein Barycenter (RWB).

Think of the RWB as a smart filter or a skeptical chef.

The Classic Ruler: "I see a bucket of hot sauce! I must adjust the average recipe by 50%!"
The Robust Ruler (RWB): "I see a bucket of hot sauce. That looks like a mistake or a glitch. I'm going to ignore anything that is too extreme. I'll only listen to the differences that make sense."

They achieve this by using a technique called truncation. Imagine a speed limit sign. If you are driving at 60 mph, the fine is proportional to your speed. But if you are driving at 200 mph, the RWB says, "Okay, 200 mph is crazy, but for the sake of calculating the average speed of traffic, let's just treat it as if you were driving at the speed limit." It caps the influence of extreme outliers so they don't ruin the whole calculation.

Why This Matters (The "Why Should I Care?")

The paper proves three main things:

It Exists and Works: They mathematically proved that this new "Robust Average" always exists and that if you keep adding more data, it will eventually settle on the true, correct average (even if some data is messy).
It Doesn't Need Perfect Data: The old method required data to be "well-behaved" (mathematically speaking, it needed finite moments). The new method works even if the data is chaotic, heavy-tailed, or full of wild outliers.
It's Faster in High Dimensions: Calculating these averages for complex data (like high-resolution images) is usually incredibly slow and hard (the "Curse of Dimensionality"). The authors found a way to let the computer figure out where to look for the average, rather than forcing it to check every single possible spot. This makes the calculation much more efficient.

The Experiments: Putting It to the Test

The authors tested their new tool in two main ways:

1. The Image Test (The Blurry vs. Sharp Photo)

Scenario: They took 200 pictures of nested ellipses (like a target). They added random "noise" dots (outliers) to the top-right corner of every picture.
Classic Result: When they averaged the pictures, the classic method produced a blurry image where the shape was dragged toward the noise dots.
Robust Result: The RWB produced a sharp, clean image that ignored the noise dots entirely. It kept the shape true.

2. The Stock Market Test (The Crazy Investor)

Scenario: They looked at stock market closing prices. Most days, prices move slightly. But sometimes, a stock crashes or spikes wildly (outliers).
Classic Result: The average stock trend calculated by the old method was skewed by these crazy days.
Robust Result: The RWB ignored the crazy spikes and gave an average trend that looked much more like the "clean" market data (the days without the crashes).

The Takeaway

In a world full of messy data, fake news, sensor glitches, and financial crashes, the Robust Wasserstein Barycenter is like a wise elder who knows how to separate the signal from the noise.

Instead of letting one crazy data point drag the whole group off course, it stands firm, filters out the extremes, and finds the true center. This makes it a powerful new tool for everything from medical imaging to financial analysis, ensuring that our "averages" actually represent reality, not just the outliers.

1. Problem Statement

The Wasserstein barycenter (WB) is a powerful tool in statistics and machine learning for averaging probability distributions, widely used in tasks like image processing, clustering, and Bayesian inference. However, the classical WB suffers from two fundamental limitations:

Sensitivity to Outliers: The classical WB is highly sensitive to data contamination and extreme values (outliers), which can drastically shift the resulting barycenter.
Moment Assumptions: The existence and consistency of the classical WB rely on the assumption that the input probability distributions have finite first or second moments. This restricts its applicability to heavy-tailed distributions (e.g., $t$ -distributions with low degrees of freedom) where moments may be undefined or infinite.

The paper aims to address these limitations by proposing a Robust Wasserstein Barycenter (RWB) that maintains the geometric benefits of the Wasserstein distance while ensuring robustness against outliers and removing strict moment requirements.

2. Methodology

A. Robust Optimal Transport (ROT) Foundation

The authors build upon the concept of Robust Optimal Transport, which modifies the cost function of the classical Wasserstein distance.

Truncated Cost Function: Instead of using the raw distance $d(x, y)$ , the method introduces a regularization parameter $\lambda > 0$ to truncate the cost:
$c^{(\lambda)}(x, y) = \min\{d(x, y), \lambda\}$
Robust Wasserstein Distance ( $W^{(\lambda)}_p$ ): The $p$ $p$ -th order robust Wasserstein distance is defined as the optimal transport cost using the truncated metric raised to the power $p$ $p$ :
$W^{(\lambda)}_p(\mu, \nu) := \left( \inf_{\gamma \in \Pi(\mu, \nu)} \int_{X \times X} [c^{(\lambda)}(x, y)]^p d\gamma \right)^{1/p}$
- Key Property: Unlike the classical $W_p$ , $W^{(\lambda)}_p$ does not require finite moment assumptions on the probability measures because the cost function is bounded by $\lambda$ .

B. Definition of Robust Wasserstein Barycenter (RWB)

The RWB is defined as the Fréchet mean of a set of probability measures under the robust Wasserstein metric. For a set of distributions $\{\mu_i\}_{i=1}^n$ with weights $w_i$ :
$\nu^{(\lambda)}_{p,k} := \arg\min_{\xi \in \mathcal{W}^{(\lambda)}_p(X)} \sum_{i=1}^n w_i \left[ W^{(\lambda)}_p(\mu_i, \xi) \right]^k$
In practice, $k=p$ is typically used.

C. Theoretical Guarantees

The paper establishes rigorous theoretical foundations:

Existence: The functional to be minimized is bounded and lower semi-continuous. Using Prokhorov's theorem, the authors prove that a minimizer (the RWB) always exists for any random probability measure, even without moment assumptions.
Consistency: The sequence of empirical RWBs converges to the true population RWB as the sample size increases, provided the input measures converge in the robust Wasserstein metric.
Support Characterization: The authors prove that if input distributions have finite supports, the RWB also has a finite support. Specifically, the size of the support is bounded by $\sum S_i + 1 - n$ (where $S_i$ is the support size of input $i$ ).

D. Computational Algorithm: Free-Support Method

Computing Wasserstein barycenters is often NP-hard due to the "curse of dimensionality" when using fixed-grid (fixed-support) methods like Iterative Bregman Projections (IBP).

Free-Support Approach: The authors propose a free-support algorithm that optimizes both the probability masses and the locations of the support points.
Fixed-Point Iteration: The algorithm alternates between:
1. Computing optimal transport plans (masses) for a fixed set of support points.
2. Updating the support point locations using a fixed-point update rule derived from the optimality conditions (minimizing the weighted sum of truncated costs).
Advantage: This approach avoids the exponential growth of grid points in high dimensions and leverages the sparsity of the optimal solution.

3. Key Contributions

Novel Concept: Introduction of the Robust Wasserstein Barycenter (RWB), which generalizes the classical WB to handle outliers and heavy-tailed distributions without moment assumptions.
Theoretical Proof: Establishment of existence and consistency guarantees for the RWB, filling a theoretical gap in robust optimal transport literature.
Algorithmic Innovation: Development of a free-support algorithm for computing the RWB. This method overcomes the curse of dimensionality inherent in fixed-grid methods and improves computational efficiency in high-dimensional scenarios (e.g., image processing).
Support Size Characterization: Providing a theoretical upper bound on the support size of the RWB, which aids in algorithm design.

4. Experimental Results

The authors validated the RWB through extensive simulations and real-world data analysis:

Image Processing (Elliptical Images):
- Setup: 200 random nested ellipse images with added outlier points in the corners.
- Result: The classical WB boundaries became blurred and shifted toward the outliers. The RWB maintained sharp boundaries and showed no significant deviation, demonstrating superior robustness.
- Observation: The free-support variant of the algorithm produced visually sharper results than the fixed-support IBP.
Numerical Simulations (Contamination & Heavy Tails):
- Contamination: In datasets with Gaussian signals contaminated by outliers, the RWB remained closer to the true barycenter as the contamination ratio increased (0% to 25%).
- Heavy Tails: Using $t$ -distributions (where second moments are undefined for low degrees of freedom), the classical $W_2$ barycenter failed or performed poorly, while the RWB successfully converged and outperformed the WB.
- Parameter Sensitivity: The parameter $\lambda$ controls the trade-off between robustness and information loss. An optimal $\lambda$ (e.g., $\lambda=30$ in specific simulations) was found to minimize the distance to the true barycenter.
Real-World Data (Stock Market Analysis):
- Dataset: Daily log-returns of stock closing prices, which contain extreme outliers.
- Result: The RWB calculated on the full dataset (including outliers) was significantly closer to the "clean" barycenter (calculated without outliers) than the classical WB was. This confirms the RWB's ability to filter out market noise and identify the central trend of asset returns.

5. Significance

This paper makes a significant contribution to the fields of optimal transport and robust statistics:

Practical Applicability: It provides a viable solution for averaging distributions in real-world scenarios where data is often noisy, contaminated, or heavy-tailed, rendering classical methods unreliable.
Theoretical Advancement: By removing the finite moment assumption, the framework extends the applicability of Wasserstein barycenters to a broader class of probability measures.
Computational Efficiency: The proposed free-support algorithm offers a scalable solution for high-dimensional problems, addressing a major bottleneck in current Wasserstein barycenter implementations.
Future Directions: The authors suggest exploring entropy-regularized versions of the RWB and integrating it with deep learning models, opening new avenues for robust machine learning applications.