Executive Summary
This paper proves that structural entanglement is a mathematical theorem, not an empirical observation. When k concepts are encoded in d dimensions with d much greater than k, concentration of measure on the high-dimensional sphere ensures that every informative direction carries all concepts, with entanglement intensity approaching a constant c greater than 1 as d/k approaches infinity. The proof proceeds through three lemmas: concentration on the sphere (no direction can be concept-pure), Johnson-Lindenstrauss preservation (random projection preserves entanglement above a dimensional threshold), and PCA disentanglement (variance-based projection escapes the concentration effect).
Four corollaries follow from the theorem. First, superlinear amplification: adding a (k+1)th concept amplifies interference among existing concepts by more than a factor of 1, matching the empirical 2x ratio. Second, concept-type independence: EI depends on d, k, and cardinalities, not semantic content. Third, the specialist bound: a model tracking k_i much less than k_total concepts has EI bounded by f(d/k_i), making compositional architecture a geometric necessity rather than a design preference. Fourth, the alignment implication: surgical concept editing is geometrically limited because no direction can be modified without affecting all encoded concepts.
The theorem positions against three adjacent literatures: Mueller et al. (2025), who find one-to-many feature-concept relationships but not all-to-all entanglement; Erogullari et al. (2025), who treat entanglement as a correctable training artifact; and the LLM unlearning literature, which treats forget-retain entanglement as a problem to solve. The theorem shows that entanglement is not a training artifact, a correctable deficiency, or a problem — it is the geometry.
Key Findings
- Theorem proved: Structural entanglement follows from concentration of measure — it is geometric, not learned
- Three lemmas: Concentration on sphere (exponential purity bound), JL preservation (dimensional threshold at 32r), PCA disentanglement (variance-based escape)
- Specialist bound: Compositional architecture is a geometric necessity when k_total is large, not a design preference
- Alignment implication: Surgical concept editing is geometrically limited — no direction modification preserves other concepts
- Zero predictions falsified: All theorem predictions confirmed across 4 architectures, 2 concept families, all tested conditions
- Positions against adjacent work: Distinguishes from one-to-many (Mueller), correctable artifact (Erogullari), and solvable problem (unlearning) perspectives
Key References
- McEntire (2026) — Entangled Directions (AI-25): discovers the damage matrix measurement tool
- McEntire (2026) — Structural Entanglement (AI-26): eight-experiment empirical validation
- Mueller et al. (2025) — From Isolation to Entanglement: one-to-many feature-concept relationships (arXiv:2512.15134)
- Erogullari et al. (2025) — Post-Hoc Concept Disentanglement via CAVs (arXiv:2503.05522)
- Liu et al. (2026) — EGUP: LLM unlearning with entanglement-guided utility preservation (arXiv:2508.20443)
- Raginsky and Sason (2014) — Concentration of measure inequalities in information theory (arXiv:1212.4663)