nucleation

LIT-876: Hippocampal spatial representations exhibit a hyperbolic geometry that expands with experience

3 min read

Properties4
StatusActive — read 2026-10-09 (NOTE-668); worth reading as, on its own account, the first evidence that neural responses, not stimuli, carry a hyperbolic geometry. Pairwise spike-train correlations of a few dozen active dorsal CA1 cells per session (34 to 56 in the sessions shown), in rats on a novel 48-m track, a 2.5-m track and a 1.8-m box, give Betti curves (clique topology, invariant to any monotone transform of the correlations) that match points sampled uniformly from a 3D hyperbolic ball and fail to match uniform samples from Euclidean cubes of low dimension or shuffled spikes. The fitted radius (10.5–15.5 in units of inverse curvature) grows with the logarithm of exploration time, on scales from seconds to days, as the maximal information acquirable does. Place-field sizes are close to exponential, as the model requires, and in simulation exponentially distributed field sizes give more Fisher information than uniform or log-normal ones, with an optimal radius growing as log N that extrapolates to the radius measured. The "hierarchy" is one of nested place-field sizes; no tree is recovered from the data, and the only null geometry tested is Euclidean with uniform sampling.
Tagsneuroscience, mathematics, information-theory
Cited as “paper” byNOTE-668 — Hippocampal spatial representations exhibit a hyperbolic geometry
Cited as “source” byTHEORY-192 — In twelve sessions of rat dorsal CA1 from three data sets, pairwise correlations among place cells have the Betti-curve signature of points sampled uniformly from a three-dimensional hyperbolic ball, not of a low-dimensional Euclidean cube, and the fitted radius rose with the logarithm of exploration time, THEORY-618 — Neural circuits that must represent hierarchically organised states in few dimensions use a hyperbolic geometry, three-dimensional where it has been measured, because hyperbolic space holds exponentially many well-separated states within a small radius; the argument from Zipf's law to this geometry does not hold, and the evidence is two Betti-curve studies tested only against uniform Euclidean nulls

Graph View

Huanqiu Zhang, P. Dylan Rich, Albert K. Lee and Tatyana O. Sharpee (2022 online; 2023 issue), Nature Neuroscience 26(1):131–139 — DOI-10.1038/s41593-022-01212-4, open access at PubMed Central (PMC9829541)

Key takeaways

  • The geometry test. Distances between neurons are their negated pairwise spike correlations. Clique topology (Giusti et al. 2015) thresholds the matrix at every edge density and counts 1-, 2- and 3-cycles, so the Betti curves depend only on the rank order of the correlations. The experimental curves match uniform samples (300 per model) from a 3D hyperbolic ball, are rejected for uniform samples from Euclidean cubes of low dimension (up to 10 in the robustness check), and are rejected for the same cells after time-shifting each spike train. 3D fits better than other hyperbolic dimensions in the six sessions compared (two-way ANOVA).
  • The hierarchy is one of scale. The paper’s tree is built by putting neurons with larger place fields higher and linking a field to the larger fields that contain it. With one field per neuron, field sizes in a hyperbolic representation are near exponential, p(s) ≈ ζ e^(−ζs), and the measured sizes on the 48-m track fit an exponential (χ² P = 0.85). The tree is an illustration; no tree is inferred from the data.
  • The radius grows with experience. Across days in a 1.8-m box the radius rises with log exploration time (r = 0.85), and the paper fits the formula for the maximal information acquirable in time T. On the 48-m track, within seconds, radius estimated from field sizes rises with time spent per metre. Running speed, area covered and time away from the environment do not account for it. A mechanistic reading: if a small field forms after a fixed dwell t₀ and small fields lie at the edge, whose count grows as e^R, then R ≈ log(T/t₀).
  • Why it would be efficient. In a model of independent Poisson neurons with 2D Gaussian fields, exponentially distributed field widths give more positional Fisher information than uniform or log-normal ones of the same mean, and for N neurons there is an optimal radius growing as log N. Extrapolated to 30–40% of the 320,000–490,000 CA1 pyramidal cells, it lands on the radius measured in the most familiar box sessions. A gamma-Poisson model of multiple fields per cell gives the same.
  • Log-normal field sizes are not counter-evidence, the paper argues: discarding small fields, and pooling exponentials of different rates, turns exponential samples into near log-normal histograms (simulation).

Standing in the record

Filed at the owner’s request on 2026-10-09, in a batch of six works on hierarchy and hyperbolic geometry asked for after the record opened QUESTION-025: Cagnetta et al.’s random hierarchy model, Krioukov et al. (2010), Sala et al. (2018), Lin et al. (2023), this paper, and Yang et al. (2023). The owner named it as “biological evidence for hyperbolically organized neural representations”.

Read on 2026-10-09 (NOTE-668). The reading is the source of THEORY-192, the record’s statement of the finding and its limits. It cites Krioukov et al. (2010), filed in the same batch, for the claim that a tree-like network is a mesh over a hidden hyperbolic geometry.

It is not a reading for the anthology: no machine-learning practice is involved, and its subject is a biological measurement.

Backlinks