DataTerminology.comNEW
Also known as: Earth Mover’s Distance
A metric between probability distributions defined by the minimal optimal transport cost to morph one distribution into another.
DESCRIPTION
Wasserstein distance measures how much ‘mass’ must move to transform one distribution into another, grounded in optimal transport theory. It provides meaningful geometry for distributions with disjoint supports, aiding robust learning. In machine learning, WGANs use Wasserstein distances for stable training. Different orders (e.g., W1, W2) define alternative cost structures. Computational methods include linear programming, Sinkhorn regularization, and entropic approximations.
ALTERNATIVE DEFINITIONS
Wikipedia
In mathematics, the Wasserstein metric, also known as the earth mover’s distance, is a distance between probability distributions on a given metric space.
Connects the term with optimal transport and metric spaces.
Encyclopedia of Mathematics
The Wasserstein metric defines a distance between probability measures on a metric space based on optimal transportation.
Formal mathematical framing as distance between probability measures.
RELATED TERMS
Optimal Transport Probability Distribution WGAN Metric Space Loss Function Coupling Monge–Kantorovich Problem Dataset Shift
COMPETING TERMS
Kullback–Leibler Divergence Total Variation Distance
KNOWLEDGE AREAS
TARGET AUDIENCE
EXTERNAL SOURCES
- Wikipedia - Wasserstein metric
- POT (Python Optimal Transport) - Optimal transport library docs
- arXiv - WGAN paper using Wasserstein
- MIT - Optimal transport book online
TOP QUESTIONS
Why is Wasserstein Distance useful in ML?
It provides meaningful gradients and stable training for generative models, notably WGANs.
Category: Relevancy
What differentiates Wasserstein from KL divergence?
Wasserstein remains finite when supports don’t overlap, capturing geometry via transport cost; KL can be infinite.
Category: Distinction
What are common mistakes using Wasserstein Distance?
Ignoring computation cost; misconfiguring regularization (Sinkhorn) can bias results.
Category: Mistakes
How is Wasserstein Distance computed efficiently?
Use Sinkhorn iterations with entropic regularization or specialized solvers in POT.
Category: Requirements
What risks exist with Wasserstein Distance?
High computational cost and sensitivity to cost matrix definition can mislead conclusions.
Category: Risk
Where is Wasserstein Distance applied?
Generative modeling, domain adaptation, distribution shift analysis, and computer graphics.
Category: Application
What are Wasserstein orders?
W1, W2 refer to different cost formulations, often L1 vs. squared Euclidean transport cost.
Category: Overview
How to validate Wasserstein computations?
Test on toy distributions with known distances and check convergence of iterative solvers.
Category: Validation
What is a striking analogy for Wasserstein Distance?
Moving piles of earth to match another landscape—all movement costs matter.
Category: Example