Angel J. Gallego and Román Orús (2022), SN Computer Science 3, 140 — ARXIV-1708.01525
Key takeaways
Read in full; see NOTE-718.
- Parse as coarse-graining. Each MERGE discards its constituents’ identities and keeps one label for the next level. As a probability model, each MERGE is a 3-index tensor and a parse is a tree tensor network; with one tensor everywhere it is a PCFG.
- Correlated factorization. MERGE’s output is a function of its input (the converse fails), so fixing the words fixes every internal index. A parsed sentence’s probability is a product of one coefficient per node, computed in O(n).
- Correlations follow the tree. Correlation decays exponentially in tree distance. That is exponential in word separation for a sequential parse and polynomial for a balanced one. The claim that languages on average have power-law correlations rests on an asserted average over parse shapes, and is close to Lin and Tegmark’s earlier result, which the paper cites but does not engage.
- Quantum state and perplexity bound. Square roots of the coefficients give a tree tensor-network state preparable by two-body gates. Its entanglement gives a classical lower bound on a word block’s perplexity.
- Renormalization in a thin sense. No rescaling, fixed points or flows. The authors themselves qualify irreversibility: speakers keep the discarded detail for a while.
Standing in the record
Each record reads it for its own question. The anthology reads it for which architectures suit language. This record filed it for QUESTION-026. Reading answered the three checks the filing set.
- QUESTION-026. The paper fixes a whole sentence’s syntactic probability from its parts and the parse, which has the shape the question asks about. But its observable is syntactic, not semantic or context-indexed, and the authors set semantics aside. It offers neither the construction nor the separation the question names. If it leans either way, it leans by omission to the second branch: its composition stops before the proposition.
- CLAIM-098. The resemblance fails. MERGE is many-to-one, and Appendix A ties phase impenetrability to RG irreversibility. That irreversibility runs from parts to whole inside one structure, and the paper has no transport between media. A deterministic many-to-one map is the trivial case of the claim’s semigroup. The authors also say the brain does not fully erase what MERGE discards.
- Lin and Tegmark. The anthology’s caution holds. The power-law result restates theirs less rigorously.
Two nearer bearings turned up in reading. The coarse-graining keeps what the grammar’s next step needs, a target in CLAIM-152’s sense, which the paper illustrates without saying. Its verdict on neural networks is an architecture-only answer to QUESTION-060, of the kind THEORY-197 warns against; it is ML practice and the anthology’s reading.
Recommendation. The link to this record’s questions was by resemblance, and reading does not make it more. Its substance is the anthology’s question. The reader recommends the owner consider setting it Rejected here, with the reason “read; bears on no question here beyond illustration; held by the anthology as ANTH-LIT-695”. It stays Active until the owner decides.
linguistics comes first, as the paper’s subject. information-theory is
for its correlation, entropy and perplexity results. anthology-candidate
is last, because an anthology topic holds it.