L2 distance, also known as Euclidean distance, is a fundamental metric in mathematics and machine learning that quantifies the geometric distance between two points or vectors in multi-dimensional space. In vector similarity applications, L2 distance measures how far apart two vectors are, with smaller values indicating greater similarity between the vectors. This metric is widely used in machine learning, data analysis, and vector database systems for comparing embeddings and performing nearest neighbor searches ¹⁾.

The L2 distance between two vectors u and v in n-dimensional space is calculated using the formula:

d(u,v) = √(Σ(uᵢ - vᵢ)²)

where the distance is the square root of the sum of squared differences across all dimensions. This calculation represents the length of the shortest straight-line path connecting two points in Euclidean space, making it geometrically intuitive and computationally well-understood ²⁾.

The metric derives from the Pythagorean theorem extended to multiple dimensions. In two dimensions, L2 distance reduces to the familiar formula c = √(a² + b²), while higher-dimensional spaces follow the same principle. The metric is always non-negative, equals zero only when vectors are identical, and satisfies the triangle inequality property required of valid distance metrics.

L2 distance serves as a primary similarity metric in modern vector databases and embedding systems. PostgreSQL's pgvector extension, widely used for storing and querying vector embeddings, implements L2 distance as one of its supported distance operators ³⁾.

In machine learning applications, L2 distance enables nearest neighbor search, where systems identify the most similar vectors to a query vector by finding those with the smallest L2 distances. This approach powers semantic search, recommendation systems, and similarity-based retrieval. When combined with approximate nearest neighbor algorithms like hierarchical navigable small-world (HNSW) graphs, L2 distance enables efficient similarity searches across large-scale vector collections containing millions or billions of embeddings.

L2 distance differs from other common similarity metrics used in vector comparison. Cosine similarity measures the angle between vectors regardless of their magnitude, making it useful for comparing normalized embeddings where direction matters more than scale. Manhattan distance (L1 norm) sums absolute differences rather than squared differences, offering faster computation in some scenarios but less geometric interpretability.

Dot product, while not technically a distance metric, provides a similarity measure proportional to vector magnitude and angle alignment. For embeddings normalized to unit length, dot product approximates cosine similarity. The choice between L2 distance, cosine similarity, and other metrics depends on the embedding model's characteristics, the dimensionality of the vectors, and specific application requirements.

Computing L2 distance involves floating-point arithmetic operations across all dimensions of compared vectors. For high-dimensional embeddings—such as those produced by modern language models or vision models with hundreds or thousands of dimensions—direct distance computation scales linearly with dimensionality. This computational cost motivates the use of approximate search algorithms and index structures that reduce the number of explicit distance calculations required to identify nearest neighbors.

Specialized hardware implementations and vectorized CPU operations can accelerate L2 distance calculations. When embedding vectors are stored in standardized formats and memory-aligned, modern processors compute distances efficiently across multiple vector pairs simultaneously. This efficiency explains the widespread adoption of L2 distance in production vector database systems handling real-time similarity queries.

L2 distance exhibits the “curse of dimensionality”—in very high-dimensional spaces, distance distributions become increasingly uniform, reducing the discriminative power of the metric. Vectors of different magnitudes may produce large L2 distances even if conceptually similar, making L2 distance sensitive to embedding scale and normalization.

The metric assumes Euclidean geometry, which may not reflect the actual similarity structure of data in all domains. Embeddings from neural networks often operate in non-Euclidean manifolds where Euclidean distance may not capture semantic relationships optimally. Applications requiring robustness to outliers may benefit from alternative metrics like L1 distance, which exhibits greater robustness to extreme values than squared-difference-based L2.

• Inner Product Distance Metric
• Cosine Similarity
• Hamming Distance
• Jaccard Distance