Normalized State Decomposition How a half a century Formula Reveals That Neural Networks Are Navigators, Not Reasoners

We present a unified mathematical framework showing that a single formula, sd = ± |vd| / ||v||, underlying the mathematical structure of neural network normalization for half a century, provides a complete and deterministic account of how neural networks encode and process information. We prove that normalized networks decompose representations into a normalized direction vector on a unit hypersphere and a stability/confidence scalar, with the signed normalized magnitude isolating each dimension's contribution to the entire vocabulary simultaneously. We demonstrate that neural networks operate as high-dimensional navigators: the LM Head weights define a fixed coordinate system on the hypersphere, and forward propagation traces a trajectory through this space, with each of the D hidden dimensions acting as a control axis that pushes or pulls probability mass across all V vocabulary tokens. We discover that neuron specialization operates at two levels—symbol format detection (numbers, code, punctuation) and co-occurrence pattern detection (syntactic structures like "animal + is + a") —revealing that what looks like "concept recognition" is actually format pattern matching. This framework provides the mathematical foundation for Flash Interpretability (Davies, 2025), explaining why direct LM Head weight decoding works and why "dog neurons" are not concept detectors but co-occurrence pattern detectors. The framework extends universally across all architectures (Transformers, MoE, ViTs, Autoencoders, Neural ODEs) and resolves interpretability's core mysteries: polysemanticity arises from shared format detection, superposition from limited dimensional capacity, and the linear representation hypothesis from hyperspherical geometry.