park-2023-riesz-map-unifies-embedding-unembedding

IN premise — summaries/2026/08/24/park-2023-linear-representation-s3-inner-product-for-language-model.md

Created 2026-08-24T17:11:03+00:00

Under the causal inner product, the Riesz map γ̄ ↦ ⟨γ̄, ·⟩_C sends each unembedding representation γ̄_W exactly to its embedding representation λ̄_W; in the transformed space with A = M^{1/2}, embeddings and unembeddings are literally equal (g̃_W = l̃_W) and the Euclidean inner product becomes the causal one.

Summary

What were traditionally tracked as two separate objects — a layer's embedding and its unembedding — are actually the same quantity once you measure them with the right geometric ruler. This means the system can treat them as a single unified entity rather than maintaining parallel copies, simplifying analysis and removing a whole class of inconsistency bugs that arise when the two drift apart.

Dependents

These beliefs depend on this one: