← Back to Topics

KLD / JSD / Wasserstein

Overview

You have two distributions P and Q. These metrics measure: "How different are P and Q?"

Metric Comparison

Loading diagram...

Summary

KLD: Selfish — direction matters. NOT symmetric.

JSD: Fair version of KLD. Symmetric.

Wasserstein: Effort to reshape P into Q. Symmetric.

KLD — Gaussian

P = N(μ₁, σ₁²)    Q = N(μ₂, σ₂²)

KLD(P||Q) = log(σ₂/σ₁) + [σ₁² + (μ₁-μ₂)²]/(2σ₂²) - ½

To get KLD(Q||P): swap ALL 1s and 2s

Wasserstein Distance

W₂ = √[(μ₁-μ₂)² + (σ₁-σ₂)²]

= √[center difference² + spread difference²]

JSD Shortcut

For Gaussians: "M = ½(P+Q) is a mixture. Closed form not possible. JSD(P||Q) = JSD(Q||P) by symmetry." — Full marks, no calculation needed.

View in LibraryTry in Sandbox