Contribution
No new result. The paper is a four-page review that puts three literatures side by side and proposes that they meet in one geometry: hyperbolic random graphs and their navigability (Krioukov, Boguñá and colleagues), Zipf’s law as a mark of statistical criticality and of maximally informative representations (Mora & Bialek; Schwab, Nemenman & Mehta; Aitchison, Corradi & Latham; Cubero et al.), and the hyperbolic geometry its author’s lab had found in odour statistics and odour perception (Zhou, Smith & Sharpee 2018). After it, the hypothesis that neural representations are hyperbolic has a citable statement; it was tested in neural recordings three years later (Zhang et al. 2022, LIT-876).
Key insight
Hierarchy means exponentially many states at each further level; a low-dimensional hyperbolic space holds exponentially many well-separated points within a given radius; so a system that must represent a hierarchy in few dimensions would be expected to use a hyperbolic metric. The paper’s last paragraph puts it as a reconciliation: neural and behavioural data lie on low-dimensional manifolds, behaviour has exponentially many hierarchically organised states, and hyperbolic space of low dimension holds both.
Assumptions
- That a shared exponential is a shared geometry. The Zipf argument needs states-with-energy and points-with-radius to be the same kind of thing. Nothing in the paper says what the “states” of a circuit are as points of a space, or what metric relates them.
- That Zipf’s law marks maximal informativeness. Taken from Cubero et al. and Schwab et al. as cited, with no statement of the conditions under which it holds.
- That 3D is privileged. It rests on three observations of 3D (two olfactory, one from tree visualisation) and on a rigidity theorem.
Key results
The paper reports no new analysis. What it contains, section by section:
- Opening survey. Hyperbolic geometry has been used for vision (Luneburg 1947; Gallant et al. on hyperbolic gratings in V4), olfaction, touch, colour metrics, ML embeddings (Ganea, Bécigneul & Hofmann, Hyperbolic Neural Networks, NeurIPS 2018), phylogenetic trees and Internet routing.
- Fig. 1. The half-plane and disk models; in the disk, nodes are sampled with density growing exponentially with radius and degree falling exponentially with it, which gives a power-law degree distribution (from Krioukov et al.). This is the one derived link in the paper between hyperbolic geometry and a power law, and it concerns node degrees, not state probabilities.
- Zipf and criticality. P(s) ∝ 1/r(s) (Eq. 1); with energy the log-improbability, E = log r (Eq. 2, sign written loosely), so the number of states grows as e^E. Zipf’s law arises in systems whose internal state is narrowly distributed for fixed latent variables but shifts strongly with them (refs. 24–25), and in the most informative representations of those variables (ref. 26). These “echo” the navigability of hyperbolic networks (refs. 27–30).
- Fig. 2 (from ref. 4). Fruit volatiles from tomato and strawberry, placed by the correlation of their concentrations across samples, lie near the surface of a 3D Poincaré ball after a topological test excluded Euclidean and spherical geometry and hyperbolic non-metric MDS placed them. Odorants correlated with human pleasantness ratings cluster in one region. The paper reads this as order in the representation where the early olfactory code (random projections to the mushroom body) looks random, and notes that the “potato-chip” surfaces of earlier perceptual studies are consistent with hyperboloid sections.
- Three dimensions. 3D appears in odour statistics, odour perception and tree visualisation; Mostow rigidity is offered as a reason 3D may be the lowest dimension robust to noise.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Hyperbolic geometry should apply broadly to neural and other biological circuits | conjecture | the survey; analogy to complex networks |
| C2 | Zipf’s law is a signature of hyperbolic geometry | weak: shown to share an exponential count of states, not a geometry | Eqs. 1–2 |
| C3 | Networks with hyperbolic geometry are maximally responsive to perturbations | asserted in the abstract; in the text only the Zipf-informativeness results (refs. 24–26) and navigability (refs. 27–30) are set side by side | citations |
| C4 | Fruit-odour statistics are best described by a 3D hyperbolic space with a pleasantness topography | moderate, from ref. 4; not new here | Fig. 2 |
| C5 | 3D hyperbolic representations may be the lowest-dimensional ones robust to noise | speculation; the cited theorem does not address noise | Mostow (ref. 41) |
| C6 | Hyperbolic coordinates may reveal topography elsewhere in the nervous system | conjecture | Fig. 2 by analogy |
Method
Argument by juxtaposition of cited results, one two-line derivation (Eq. 2) and one reproduced figure.
Concepts
- Zipf’s law: P(s) ∝ 1/r(s), probability inverse to rank; here generalised from words to the states of any system.
- energy: E = −log P(s), as in the statistical-mechanics reading of criticality; Zipf’s law is then a density of states e^E, an entropy equal to the energy.
- navigability: greedy routing with only local connectivity and coordinates finds near-shortest paths, as in hyperbolic networks.
- Mostow rigidity: a finite-volume hyperbolic manifold of dimension ≥ 3 is determined up to isometry by its fundamental group, so it admits no continuous deformation of its hyperbolic structure; surfaces (2D) do.
Connections
- Zhou, Smith & Sharpee 2018 (LIT-872, filed in the same part of the batch) is this paper’s ref. 4 and the source of Fig. 2, its only data.
- Zhang, Rich, Lee & Sharpee 2022 (LIT-876, NOTE-668, THEORY-192) is the programme’s test in neural activity. It also takes 3D as the dimension, and its confound (hyperbolic metric against an exponential spread of scales) is the same one Eq. 2 glosses here: an exponential count of states is a property of the distribution and does not fix a metric.
- Krioukov et al. 2010 (LIT-865, NOTE-680, THEORY-188) is ref. 8, the source of Fig. 1 and of the tree-approximation claim. There the power law is derived, from two exponentials in the radius; here the same shape is asserted for state probabilities without a model.
- Ganea, Bécigneul & Hofmann 2018: ref. 9 is their Hyperbolic Neural Networks (NeurIPS), not the entailment-cones paper held as LIT-866.
- The record’s other Zipf reading, NOTE-243 on LIT-270, makes the same point from the other side: Zipf-like curves arise in text without meaning, so a Zipf curve alone does not identify a mechanism.
Bearing on the record
- QUESTION-025. Little, and nothing toward either answer it names. The paper’s one relevant idea is the general one already in the question’s notes from Sala et al. and Krioukov et al.: a hierarchy needs exponentially many states, and hyperbolic space holds them in low dimension, so a hierarchy may live in a radial coordinate and distances rather than in linear directions. It adds a reason to think word statistics are a case in point, since Zipf’s law originated with word frequencies and the paper names “word distribution” with olfaction as places where its link applies. That reason does not survive reading: a Zipfian unigram distribution says nothing about the structure of co-occurrence among attributes, which is what THEORY-185’s additivity premise and the question concern.
- CLAIM-119 (communicative categories form a lattice rather than a hierarchy). Only by contrast: everything this paper argues for is a property of trees, and a lattice with shared attributes and cross-cutting constraints is exactly what a hyperbolic tree-approximation does not favour. It neither supports nor tests the claim.
- No THEORY: the paper’s claims are conjectures, and its one finding belongs to ref. 4. No CLAIM rests on it.
Limitations
- The Zipf-to-hyperbolic step is not an argument for hyperbolic geometry. Eq. 2 holds for any rank distribution: P ∝ r^(−a) gives a state count e^(E/a). More broadly, any system with extensive entropy has exponentially many states. What makes Zipf special, the unit slope (entropy equal to energy), is not used in the link to geometry, so the argument would make every power-law system hyperbolic.
- States are never placed in a space. The geometry is claimed for “networks”, but the Zipf states are configurations of a system, and no metric on them is proposed.
- The rigidity argument does not apply. Mostow’s theorem concerns the hyperbolic structure of finite-volume manifolds with a given fundamental group, holds in every dimension from 3 up, and says nothing about the stability of a finite point embedding under noise. The cited ref. 43 is about complex hyperbolic geometry.
- The abstract’s “maximally responsive” claim is not argued in the body beyond setting two cited literatures side by side.
- The evidence is the author’s own, one figure from one prior paper, plus older perceptual studies reinterpreted.
Open questions
- Is there a model in which a hyperbolic latent metric produces Zipf’s law in the frequencies of a circuit’s states, with the unit exponent and not just a power law?
- Does any test of “3D is most robust” exist that compares noisy embeddings across hyperbolic dimensions, rather than invoking rigidity?
- In a co-occurrence embedding, does Zipfian frequency go with radial position (frequent, general words near the centre), as the hub-near-the- centre picture of Fig. 1 would suggest by analogy?