covariance-geometry-as-canonical-tool
IN derived (depth 1)
Created 2026-08-25T03:00:40+00:00 · Reviewed 2026-08-25T04:28:09+00:00
Independent lines of work (ROME's key-space projection and Park's unembedding whitening) converge on using empirical second-moment matrices to define the "correct" inner product for reasoning about transformer representations.
Summary
Two separate lines of work, one on minimally disruptive knowledge editing and one on measuring semantic similarity between words, independently arrived at the same geometric tool: the inverse covariance matrix of transformer representations defines the natural ruler for measuring distance and angle in that space. This convergence suggests the tool is not an ad hoc convenience but a principled, canonical choice, giving the system a single well-justified way to reason about similarity, projection, and disruption in representation space.
Justifications
This belief has 2 justifications — it is IN if any one holds.
SL — Both ROME (C = KKᵀ from hidden-state samples) and Park (Cov(γ)⁻¹ for whitening) independently derive the canonical geometry from data statistics, suggesting a shared inductive principle: the data distribution defines the natural metric for representation-space reasoning.
Antecedents (all must be IN):
- IN rome-c-gram-matrix-definition — In ROME's derivation, C = KKᵀ is the Gram matrix of the key set, and its inverse C⁻¹ projects the new key into the subspace orthogonal to existing keys to minimize disruption.
SL — Both ROME (C = KKᵀ from hidden-state samples) and Park (Cov(γ)⁻¹ for whitening) independently derive the canonical geometry from data statistics, suggesting a shared inductive principle: the data distribution defines the natural metric for representation-space reasoning.
Antecedents (all must be IN):
- IN park-2023-causal-inner-product-closed-form — The causal inner product in Park et al. (2023) has the closed-form expression ⟨γ̄, γ̄'⟩_C = γ̄^⊤ Cov(γ)^{-1} γ̄', where γ is the unembedding vector of a vocabulary-uniform word, making the inverse vocabulary-covariance matrix serve as the metric (analogous to Mahalanobis distance).
Dependents
These beliefs depend on this one:
- OUT covariance-geometry-unifies-analysis-and-editing — The mathematically principled framework for both interpreting (SAE feature extraction, Park polytope analysis) and modifying (ROME rank-one edits) LLM representations is second-moment covariance geometry applied to the residual stream, since C = KKᵀ whitening defines the canonical coordinate system in which all three operations become linear algebra on the same substrate.
- IN evaluation-geometry-convergence — The convergence of embedding evaluation (cosine → Spearman correlation in SBERT/MTEB) and internal geometry analysis (covariance whitening in ROME, polytope geometry in Park) on the same second-moment structure reveals that standard benchmark evaluation is measuring the same geometric quantity that governs internal feature organization—evaluation and analysis are two readouts of one space.
- OUT superposition-necessitates-covariance-whitening — Over-complete superposition in the shared residual-stream substrate is the precise structural condition that necessitates covariance/whitening (second-moment projection) as the canonical tool for isolating individual features and performing targeted rank-one edits; without superposition, raw Euclidean geometry would suffice and the entire covariance-geometry framework would be unnecessary.