Executive Summary
This paper proves that concept entanglement in neural network activation spaces is a mathematical consequence of high-dimensional encoding. When k concepts are encoded in d dimensions with d much greater than k, every direction in the informative subspace carries all k concepts, with entanglement intensity converging to a constant c greater than 1 as d/k approaches infinity. The proof rests on three lemmas: concentration of measure on the sphere (no informative direction can be concept-pure), Johnson-Lindenstrauss preservation (random projection to at least 32r dimensions preserves entanglement), and PCA disentanglement (variance-based projection reduces it). For the experimental setting (largest concept subspace 3, informative rank 7), the main theorem's bound gives EI of at least 1.33, consistent with the observed 1.4–1.6.
Four corollaries follow. First, superlinear amplification: adding a concept amplifies interference among the existing concepts rather than only adding its own, consistent with the observed ~2x ratio for three concepts. Second, concept-type independence: EI depends on d, k, and concept cardinalities, not semantic content. Third, the specialist bound: a model tracking k_i much less than k concepts has EI bounded by f(d/k_i), a geometric argument for compositional over monolithic architecture. Fourth, the alignment implication: surgical concept editing is geometrically limited, because removing any informative direction damages other concepts.
The paper positions the result against three adjacent literatures: Mueller et al. (2025), who find one-to-many feature-concept relationships; Erogullari et al. (2025), who treat entanglement as correctable via orthogonalization; and Liu et al. (2025), who navigate it in unlearning via loss reweighting. The theorem places a floor on entanglement for a given informative subspace rank r. It does not constrain the model's choice of r: companion work finds that fine-tuning drove EI to zero in Qwen-32B across 8 seeds, so the paper treats entanglement as diagnostic of how a model compresses its training data.
Key Findings
- Theorem proved: For a given informative subspace rank, entanglement follows from concentration of measure
- Three lemmas: Concentration on the sphere (exponential tail bound), JL preservation (threshold at m of at least 32r), PCA disentanglement (variance-based escape)
- Specialist bound: A geometric argument for compositional over monolithic architecture
- Alignment implication: Surgical concept editing is geometrically limited — removing an informative direction damages other concepts
- Empirical confirmation: Eight experiments across three architectures (GPT-2 to Qwen-7B) confirm every prediction; an adversarial relabeling test shows the cardinality condition is empirically necessary
- Not a fixed limit: The bound does not constrain the model's choice of rank; fine-tuning drove EI to zero in Qwen-32B
Key References
- McEntire (2026) — Entangled Directions (AI-25)
- McEntire (2026) — Universal Entanglement in the Informative Subspace (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 (arXiv:2503.05522)
- Liu et al. (2025) — Entanglement-Guided Unlearning with Proxy Constraint (arXiv:2508.20443)
- Johnson and Lindenstrauss (1984) — Extensions of Lipschitz Mappings into a Hilbert Space (Contemporary Mathematics, 26)