Concentration of Measure

Source: Paul Levy, Problemes concrets d’analyse fonctionnelle, 1951; Michel Talagrand, Publications Mathematiques de l’IHES 81, 1995; Vitali Milman, 1988 Institution: Multiple

Finding

On a high-dimensional sphere S^n, as n increases, the surface area concentrates overwhelmingly near the equator relative to any chosen pole. The fraction of surface area farther than epsilon from the equator decreases exponentially with n. In dimensions encountered in neural network hidden states (n = 768, 2304, 4096), virtually all surface area lies within a thin band around the equatorial hyperplane. This is a theorem of measure theory, not an approximation. The Talagrand concentration inequality extends the result to any Lipschitz function on high-dimensional product spaces.

Pattern Mapping

Proportion — The concentration is exactly exponential in dimension, not faster or slower. The mathematical bound is tight. The relationship between dimension and concentration is precise.

Honesty — The theorem does not explain why epistemic properties concentrate near a linear boundary in language models; it explains why it is not surprising that they do. The distinction matters. The mathematical result is honest about its scope.

Connections

Status

Established mathematics. See Ledoux, The Concentration of Measure Phenomenon (2001). The connection to neural network representations is this project’s structural interpretation.


The mapping to the five properties is this project’s structural interpretation.