Contribution
Hyperbolic geometry had been proposed for neural circuits (Sharpee 2019) and found in olfactory stimuli and human odour perception (Zhou, Smith & Sharpee 2018), but not in recorded neural activity. This paper applies Giusti et al.’s clique topology to CA1 place-cell correlations and finds their Betti curves consistent with a 3D hyperbolic geometry and not with low-dimensional Euclidean ones. It adds a dynamic finding, the radius growing with log exploration time, and an efficiency argument: the exponential field-size distribution that a hyperbolic representation implies carries more positional information, and the optimal radius for CA1’s cell count matches the one measured.
Key insight
A population of place cells whose field sizes span scales, with exponentially more small fields than large ones, is a discretised tree: large fields contain smaller ones, and a tree embeds with low distortion in hyperbolic space. Read as points, the neurons sit in a ball whose radius is the log of the ratio between the largest and smallest scales represented. Adding small fields as the animal learns pushes the boundary out, and since their number grows as e^R, the radius grows only as the log of the time spent.
Assumptions
- Correlation is similarity. C_ij is the peak-normalised cross-correlogram integrated over ±1 s, read as degree of place-field overlap (after Hampson et al. 1996). Only its rank order is used.
- Uniform sampling. Model points are drawn uniformly in each geometry: uniform in the unit cube for Euclidean; uniform angle and radial density ∝ sinh^(d−1)(r) for the native hyperbolic ball of curvature −1 and radius R_max. Neurons are assumed to be a uniform sample of the representation.
- Noise. 5% multiplicative Gaussian noise on model distances, none on data.
- Cell selection. Putative pyramidal cells firing 0.1–7 Hz; inhibitory and unidentified units excluded; cells silent over 30 min excluded; for CRCNS sessions only the last two thirds of each session.
- One field per neuron for the field-size estimator (Eq. 1, Bayesian estimate of ζ with fields below 25 cm excluded); the topological method does not assume it.
- Fisher model. Independent Poisson neurons, 2D Gaussian fields with random orientation, widths i.i.d. exponential (or uniform, log-normal matched in mean, and variance for log-normal).
Key results
- Geometry (Fig. 2; Extended Data Figs. 1–5). On three animals on the novel 48-m track (34–41 active cells), two 2.5-m-track sessions and seven box sessions (CRCNS hc-3), experimental Betti curves fall within the spread of 3D hyperbolic samples on integrated Betti value and L1 distance, and outside that of Euclidean cubes; time-shifted spike trains fall outside the hyperbolic spread. 3D fits better than the three other hyperbolic dimensions compared (χ² statistic, ANOVA, Tukey). Restricting to the most spatially informative 75% of cells does not change this. Radii 10.5–15.5.
- Not anatomical. Field size on the 48-m track is not correlated with septotemporal or proximodistal position within dorsal CA1, so this is not the known dorsoventral gradient of field size.
- Growth with experience (Fig. 3; Extended Data Figs. 7–10). Box radius rises with log exploration time over days (r = 0.85, P = 3·10⁻⁶), fitted by I = log(1 + T/t₀) + (T/t₀) log(1 + t₀/T). On the 48-m track the radius of each 1-m segment, from field sizes, rises with time per metre (r = 0.50), and with entropy of the track’s turns, but the entropy effect vanishes once speed is controlled (partial r = 0.18, P = 0.24). After 4 h of sleep the radius is below the extrapolated value: the growth is with experience, not elapsed time.
- Decoding (Fig. 4a–c). Median decoding error falls exponentially with spikes used; the exponent rises over days and correlates with radius (r = 0.70).
- Efficiency (Figs. 4d–g, 5). Exponential widths beat uniform and log-normal ones on Fisher information at every N simulated; the optimal R grows as log N; extrapolated to the number of active CA1 neurons it matches the most-familiar-box radius. Same with gamma-distributed field counts per cell.
- Log-normal reconciliation (Fig. 6). Truncating small fields and pooling exponentials of different rates make exponential samples fit a log-normal (simulation only).
- Undersampling (Extended Data Fig. 9). Under about a minute of data the topological radius estimate is biased upward; this is why the seconds-scale analysis uses field sizes instead.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | CA1 correlation rank-orders match uniform 3D hyperbolic samples and not uniform low-dimensional Euclidean ones | moderate (12 sessions, 3 data sets, shuffle control) | Fig. 2; ED Figs. 1–5 |
| C2 | The hyperbolic radius grows with log exploration time, as maximal acquirable information does | moderate (correlational; log fit and Eq. 2 fit both shown) | Fig. 3; ED Figs. 7, 8, 10 |
| C3 | The growth tracks experience, not speed, area or elapsed time | moderate | ED Fig. 8 |
| C4 | Exponentially distributed field sizes maximise positional Fisher information among the three families tried | strong within the model; three families only | Figs. 4e, 5 |
| C5 | The measured radius is the optimum for CA1’s neuron count | weak (extrapolation of a simulated log N curve over four orders of magnitude) | Fig. 4g, 5b |
| C6 | The representation is hierarchical in a tree-like way | illustrative; no tree is inferred from data | Fig. 1a–b |
| C7 | Spikes from different neurons should be order-dependent because hyperbolic addition is non-commutative | conjecture, untested | Discussion |
Method
The geometric inference is Giusti et al.’s: Betti curves of the order complex of a similarity matrix, compared with the curves of matrices built from points sampled in candidate geometries, 300 samples per candidate, with P values from where the data fall in the sampled distribution. Radius is chosen by maximising the product of four such P values over R in {5, 5.5, …, 24}, with 75% subsampling of cells repeated 100 times. The second estimator fits an exponential to place-field sizes and reads the rate as curvature. Efficiency is argued by simulated Fisher information.
Concepts
- hyperbolic radius R_max: the radius of the ball at curvature −1; equivalently the inverse curvature at fixed radius. The paper calls it the “size” of the representation.
- Betti curve β_m(ρ): number of m-dimensional cycles (not bounded by cliques) in the graph keeping the top fraction ρ of pairs.
- integrated Betti value: ∫ β_m(ρ) dρ.
- temporal familiarity: time spent per metre of track, the inverse of mean speed over the fields in a segment.
Connections
- Krioukov et al. (2010), Hyperbolic Geometry of Complex Networks, filed and read in the same batch (LIT-865, NOTE-680): cited here (ref. 13, with Gromov) for a tree-like network being a mesh over a hidden hyperbolic geometry, and Boguñá, Papadopoulos & Krioukov (2010) for greedy routing. The node-radius ↔ log-degree map there corresponds to the field-size ↔ radius map here: in both, radial position is a log-scale variable and uniform sampling in the ball gives an exponential tail. That is my connection; the paper draws only the routing one.
- Sala et al. (2018), filed in the same batch (LIT-874): its subject is the precision cost of embedding trees in hyperbolic space; this paper embeds no tree, so the tradeoffs there are not engaged.
- In the record’s neuroscience holdings nothing else measures the geometry of a population code; LIT-633 (ripple-mediated co-firing) is the nearest by method, using pairwise co-firing, and does not ask about geometry.
Bearing on the record
- QUESTION-025. Little. The question asks whether attributes that imply or exclude one another in co-occurrence statistics still give linear attribute directions and a non-Boolean concept lattice. This paper is about a spatial code, and its hierarchy is containment of place fields by scale, not implication among attributes. It offers one analogy and no answer: nested fields are nested extents, a partial order by inclusion, and here such an order is said to be represented with a hyperbolic metric rather than by linear directions in a flat space. If that carried over, a hierarchical attribute structure would show up as radial depth plus angle rather than as a sum of attribute vectors, which is the opposite of THEORY-185’s premise. The paper does not test anything of the kind, and it does not satisfy either kind of answer QUESTION-025 names.
- It is the source of THEORY-192.
- No CLAIM in the record rests on biological representation geometry, so it grounds none; no instruction for ML practice, nothing for the anthology.
Limitations
- Metric and distribution are confounded. The paper says so itself: model Betti curves differ “through (1) different distribution of points … and (2) different distance metrics”. The Euclidean null is a uniform cube; no Euclidean null with an exponential (multiscale) distribution of a size coordinate was tried, nor any other non-uniform Euclidean or spherical model. So C1 shows “uniform hyperbolic, not uniform Euclidean”, and the step to “the code’s metric is hyperbolic” is not isolated from “field sizes are exponential”. The paper’s own Fig. 1b construction (centre in the plane, size as height) is the half-space picture in which those two coincide.
- Few cells. 34–56 cells per session for a 3D geometry with one free radius; the radius estimate is biased under about a minute of data.
- No tree recovered. “Hierarchical” rests on the construction in Fig. 1b and on the exponential field-size distribution, not on any inferred branching structure in the data.
- Correlational dynamics. Radius growth versus time is a regression on sorted sessions; the log fit and the information formula are both fitted, with t₀ free.
- Efficiency is model-bound. The Fisher comparison is against two alternative families, under independent Poisson neurons, and the match to CA1’s cell count extrapolates a simulated curve far beyond the simulated range.
- Dorsal CA1 only, rats only, a few environments.
Open questions
- Does a Euclidean model with exponentially distributed field sizes, or a model with field centres and scales sampled independently, produce the same Betti curves? If so, C1 is a finding about scale distribution rather than about metric.
- Does the order-dependence of spike integration predicted from non-commutative hyperbolic addition (C7) appear in downstream readout?
- Does the same signature hold in ventral CA1, entorhinal cortex, or non-spatial hippocampal codes, where the authors expect a broader hierarchy?