L2 Distance was Giving Me Wrong Answers. Here’s the Metric That Fixed it.

2026-05-16 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Advanced, medium

Summary

The article addresses the limitations of L2 distance for comparing complex data distributions, specifically in the context of audio fingerprints for OmniPulse. It explains that L2 distance, which measures "are these numbers similar?", is inadequate for comparing distributions of energy across wavelet scales and time, as it fails to account for structural differences. The author introduces Wasserstein distance, particularly Sliced-Wasserstein (SW₁), as a superior metric for measuring the "work" required to transform one distribution into another. Sliced-Wasserstein overcomes the O(N³) computational complexity of exact Wasserstein by projecting high-dimensional data onto random 1D directions, computing 1D Wasserstein distance (O(N log N)), and averaging the results. This technique achieves a practical O(L × N log N) complexity, making it viable for large datasets. The implementation details for a Rust library `sliced-wasserstein` are provided, along with correctness guarantees and real-world test results demonstrating its ability to capture physically coherent signal structure in audio fingerprints.

Key takeaway

For AI Scientists and Research Scientists working with data that represents distributions or point clouds, such as audio fingerprints, LiDAR scans, or document embeddings, you should consider adopting Sliced-Wasserstein distance instead of L2. This metric provides a geometrically correct measure of similarity, ensuring that your models capture meaningful structural differences and leading to more accurate retrieval and analysis, as demonstrated by its application in OmniPulse's HNSW index.

Key insights

Sliced-Wasserstein distance effectively compares complex data distributions by measuring transformation work, outperforming L2 distance.

Principles

• L2 distance is insufficient for comparing structural similarity in distributions.
• Wasserstein distance quantifies the "work" to transform one distribution into another.
• Slicing enables efficient approximation of high-dimensional Wasserstein distance.

Method

Sliced-Wasserstein projects high-dimensional distributions onto multiple random 1D lines, computes 1D Wasserstein distance for each projection, and averages these distances to estimate the true Wasserstein distance.

In practice

• Use `sliced-wasserstein` crate for distribution comparisons.
• Configure `n_projections` for accuracy vs. speed trade-off.
• Set `seed` for deterministic distance calculations.

Topics

• L2 Distance Limitations
• Wasserstein Distance
• Sliced-Wasserstein
• Audio Fingerprints
• Wavelet Scattering Transform

Best for: AI Scientist, Research Scientist, Machine Learning Engineer, AI Engineer, Data Scientist

Related on AIssential

• Wasserstein Distance is Just Moving Dirt — Wasserstein distance measures the minimum work to transform one distribution into another, providing useful gradients where classical metrics fail.
• Distance-Matrix Wasserstein Statistics for Scalable Gromov--Wasserstein Learning — DMW offers a scalable, provable relaxation of Gromov--Wasserstein for comparing structural data via sampled distance matrices.
• skwdro: a library for Wasserstein distributionally robust machine learning — skwdro simplifies robust ML model training using Wasserstein distributionally robust optimization with PyTorch and scikit-learn compatibility.
• $L^2$ over Wasserstein: Statistical Analysis for Optimal Transport — The $L^{2}$ over Wasserstein space unifies optimal transport under statistical uncertainty, extending its geometric and flow properties to random measures.
• I-SAFE: Wasserstein Coherence Metrics for Structural Auditing of Scientific AI Models — Scientific AI model performance benchmarks are insufficient to confirm alignment with domain-relevant structural behavior.
• Canonical Variates in Wasserstein Metric Space — Dimension reduction in Wasserstein metric space via Fisher's ratio significantly improves classification of distributional data instances.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.