Pith. sign in

REVIEW 6 major objections 5 minor 2 cited by

A group-theoretic framework for machine learning in hyperbolic spaces

T0 review · 6 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims hyperbolic ML can get rigorous answers to its four basic questions—mean, computation, sampling, and parameter estimation—in hyperbolic balls.

desk verdict A coherent group-theoretic framework with several correctable algebraic errors; the current version is not rigorous enough to accept, but the higher-dimensional extensions are worth a revision. read the letter →

arxiv 2501.06934 v1 pith:6I6M44PP submitted 2025-01-12 cs.LG

Vladimir Jaćimović This is my paper · ORCID
classification cs.LG MSC 51M1053B2062F1068T07
keywords hyperbolicmachinelearningconformalbarycenterMöbiusdistributionsPoincaréballBergmangradientdescentmaximumlikelihoodgroup
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to give hyperbolic machine learning the same basic toolkit that Euclidean machine learning takes for granted. It claims that for hyperbolic balls one can define a principled mean (a conformal barycenter), a family of probability distributions (Möbius distributions) that is invariant under isometries, and efficient algorithms for computing both the barycenter and maximum-likelihood estimates. If the framework is right, hyperbolic latent spaces could use these objects the way Euclidean models use the arithmetic mean and the Gaussian, replacing ad-hoc exponential-map and projected-gradient recipes. The payoff would be compact hyperbolic models whose geometry, statistics, and optimization are mutually consistent.

What carries the argument

The load-bearing object is the group of Möbius (disc-preserving linear-fractional) transformations of the hyperbolic ball and its infinitesimal generators. The barycenter is characterized through this group: a configuration is balanced when its center is zero, and the conformal barycenter is the point whose generating Möbius transformation sends the configuration to a balanced one. The computational machinery is the Poincaré swarm, an ODE whose vector field is asserted to be a time-dependent linear combination of the three infinitesimal generators of the Möbius group; a Lie-group theorem then implies the swarm evolves by Möbius transformations, and a hyperbolic-gradient check identifies the evolving parameter with gradient descent on the potential $H$. The same pattern—generators of the isometry group, a swarm ODE, and a gradient-flow identity—is repeated for Poincaré balls using Lorentz boosts and rotations and for Bergman balls using holomorphic automorphisms.

What would settle it

Compute $\frac{d}{dt}\big|_{t=0} \frac{t-z}{1-tz}$; it equals $1-z^2$, whereas Section 3.2 states the infinitesimal generator $v_2 = z^2 - 1$. Substituting the paper's $h_1,h_2$ into (20) and comparing with (17) then shows directly whether the linear-combination claim used to prove Theorem 5 is correct.

Watch

Extended reading notes

Core claim

The central discovery is that the four basic questions facing any metric-space machine-learning setup—how to average points, how to compute that average, how to sample random points, and how to fit parameters—have rigorous answers when the metric space is a hyperbolic ball. For a configuration of points in the Poincaré disc, the paper defines the conformal barycenter as the unique point $a$ such that the Möbius-transformed configuration is balanced, equivalently the unique minimizer of the potential $H(a) = -\sum_i \log \frac{(1-|a|^2)(1-|z_i|^2)}{|1-\bar{a}z_i|^2}$; this barycenter is conformally invariant. It then introduces Poincaré swarms, dynamical systems whose trajectories evolve by Möbius group actions and which are hyperbolic gradient flows for $H$, so that running the swarm computes the barycenter. The paper also introduces the Möbius family of densities $p(z;a,s) \propto ((1-|z|^2)(1-|a|^2)/|1-\bar{a}z|^2)^s$, shows the family is invariant under the isometry group, gives an explicit inversion-sampling scheme, and proves that maximum likelihood separates into a barycenter problem for $a$ and a one-dimensional convex problem for the concentration $s$. These constructions extend to higher-dimensional Poincaré balls and to the non-equivalent Bergman balls in even dimensions.

Load-bearing premise

Every swarm algorithm for computing barycenters depends on the assertion that the right-hand side of the swarm ODE is exactly a linear combination of the Möbius group's infinitesimal generators with the stated coefficients; if that assertion fails, the proof that the swarm computes the barycenter collapses.

Editorial extensions

If this is right

  • Barycenters in hyperbolic balls are conformally invariant: applying any isometry to a configuration moves the mean by that same isometry, which is the equivariance property required of a mean in metric learning.
  • The Möbius family gives hyperbolic ML a parametric probability model with a closed-form sampling recipe and a maximum-likelihood procedure that reduces to computing the barycenter and solving a one-dimensional convex problem.
  • Poincaré swarms give a gradient-descent algorithm that respects the hyperbolic metric, so barycenter computation does not require projecting Euclidean updates back into the ball.
  • In even dimensions, the same framework works for both Poincaré and Bergman balls, and the paper argues the choice between them makes no qualitative difference for statistics or computation.
  • The framework can support more elaborate pipelines, such as expectation-maximization, variational inference, and normalizing flows in hyperbolic latent spaces.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: if the swarm proof holds, the same Lie-group-plus-gradient-flow template should produce barycenter algorithms on other homogeneous spaces, such as spheres or the hyperboloid model, by writing their isometry generators explicitly.
  • Editorial extension: one concrete test of the framework's practical value is to benchmark the Möbius family against the wrapped normal distribution in a hyperbolic variational autoencoder; the paper does not report such experiments, but they would directly probe whether the sampling and MLE properties translate into better latent-space modeling.
  • Editorial extension: conformal invariance suggests that equivariant neural network layers could be designed around the Möbius group rather than around tangent-space operations, but the paper only sketches this direction.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. The paper proposes a group-theoretic framework for machine learning on hyperbolic balls. In the Poincaré disc it defines a conformal barycenter as the point whose associated Möbius transformation balances a configuration, introduces 'Poincaré swarms' whose dynamics are claimed to evolve by the Möbius group and to implement hyperbolic gradient descent, and defines a conformally invariant family of 'Möbius' probability distributions with exact sampling and maximum likelihood estimation. These constructions are then extended to higher-dimensional Poincaré and Bergman balls. The stated contribution is to give rigorous answers to four methodological questions: definition and computation of a mean, sampling, and parameter estimation.

Significance. If the mathematical claims were correct, the framework would offer a principled and computationally useful alternative to current hyperbolic ML tools: a group-invariant notion of mean, a gradient-flow algorithm, and a tractable statistical family with exact sampling and closed-form MLE. The paper also usefully connects hyperbolic ML to the Kuramoto oscillator literature and to complex hyperbolic geometry. However, the manuscript as written contains several algebraic errors in the central derivations: the metric tensor in Definition 1 is not the standard Poincaré metric, an infinitesimal generator in Section 3.2 has the wrong sign, the sampling inversion in Eq. (28) is incorrect, and the normalizing constant in Eq. (42) has the wrong power of π. In addition, core existence and uniqueness results are imported from the author's own unpublished preprints rather than proved. The intended contributions are therefore not established in this version.

major comments (6)
  1. [Section 2.2, Definition 1] The metric defined in Definition 1, g_x(u,v)=⟨u,v⟩/(1-|x|^2), is not the standard Poincaré ball metric. The standard metric is 4⟨u,v⟩/(1-|x|^2)^2 (up to a constant factor) and has constant sectional curvature -1; the metric as written is conformal to Euclidean with conformal factor (1-|x|^2)^{-1} and does not have constant negative curvature. This is not a harmless convention issue: the distance formula (5)-(6), the hyperbolic measure (4), and the hyperbolic gradient formula (23) used throughout the paper are all compatible with the square-denominator metric, not with the displayed one. The definition must be corrected and the curvature- and gradient-dependent claims rechecked.
  2. [Section 3.2, Eq. (20)] The infinitesimal generator v2 is miscomputed. For m2(t)=(t-z)/(1-tz), one obtains m2'(0)=1-z^2, not z^2-1. With the printed v2=z^2-1 and the stated h1,h2, the combination h1(z^2-1)+h2(iz^2+i) equals (K/2N)(Σ_k ζ_k) z^2 - (K/2N)(Σ_k \bar ζ_k), whereas the right-hand side of Eq. (17) is -(K/2N)(Σ_k \bar ζ_k) z^2 + (K/2N)Σ_k ζ_k. The coefficients of z^2 and the constant term are interchanged, so Eq. (17) is not a linear combination of the claimed generators. Replacing v2 by 1-z^2 would make the identity correct, but as written the proof of Theorem 5 and hence the swarm-based algorithm are unsupported.
  3. [Section 4.2, Eqs. (27)-(28)] The cumulative distribution function for the radial coordinate is incorrect. From the integral in Eq. (27), the CDF is ∫0^{2π}∫0^b (s-1)/π (1-r^2)^{s-2} r dr dφ = 1-(1-b^2)^{s-1}, not 1-(1-√b)^{s-1}. The inversion in Eq. (28) should therefore be |z| = sqrt(1-(1-κ)^{1/(s-1)}), not (1-(1-κ)^{1/(s-1)})^2. As printed, the sampling scheme generates points from a different radial distribution, so the answer to question iii) is not valid.
  4. [Section 6.1, Eq. (42)] The normalizing constant for the Möbius family on the Poincaré ball has the wrong power of π. The correct constant for the density with respect to dΛ(x) in Eq. (4) is π^{-d/2} Γ(1+s-d/2)/Γ(1+s-d) (up to a metric normalization), not π^{d/2} times that ratio. As a check, at d=2 Eq. (42) gives π(s-1)(1-|z|^2)^s, while the two-dimensional family in Eq. (26) has density (s-1)/π (1-|z|^2)^s; the two disagree by a factor of π^2. This error propagates into the likelihood (45) and the MLE formulas.
  5. [Sections 3.1 and 4.1 (Theorems 1, 4, Proposition 9)] Load-bearing statements are imported from the author's own preprints [47] and [57] without proof: the existence and uniqueness of the Möbius transformation balancing a configuration (Theorem 1), geodesic convexity and uniqueness of the minimum of H (Theorem 4), and the conformal invariance of the family (Proposition 9). The manuscript says 'We will omit the proofs, as they are provided therein.' Since these results constitute the definition and existence of the barycenter and the definition of the statistical family, the paper is not self-contained. For a paper whose stated aim is to give 'detailed and rigorous answers,' these results should either be proved here or the manuscript should state clearly that it builds on and assumes them.
  6. [Section 5.1, proof of Theorem 13] The gradient computation in the proof of Theorem 13 contains sign errors. Between the second and third displayed lines, the numerator changes from a|y_i-a|^2 - a(1-|a|^2) + y_i(1-|a|^2) to a|y_i-a|^2 + (y_i-a)(1-|a|^2), and the Möbius transformation h_a in Eq. (7) has numerator proportional to (a-x) rather than (y_i-a). The subsequent substitution that replaces h_a(x_i(0)) by x_i in the hyperbolic gradient is also not justified, so the identification of Eq. (36) as the gradient flow for H_d is not established.
minor comments (5)
  1. [Section 2.3, Definition 2] The index n is used interchangeably with the complex dimension m in the definitions of b(z)_ij and K(z,w); the notation should be made consistent.
  2. [Section 2.2, Eq. (10)] The Jacobian formula uses n for the dimension, which is inconsistent with the rest of the section using d; it should be d throughout.
  3. [Section 3.1, Theorem 1 vs Definition 4] Theorem 1 asserts uniqueness 'up to a rotation,' while Definition 4 asserts a unique point a; the role of the residual rotation in the definition of the barycenter should be clarified.
  4. [Section 3.2, Eq. (19)] The transformation ga(z)=(z-a)/(1-\bar a z) in Eq. (19) differs from the convention in Eq. (1) by a minus sign; this is likely absorbed by the eiθ factor, but the convention should be stated explicitly to avoid confusion.
  5. [Section 6.1, after Eq. (42)] The transitivity statement says 'h(a1) = h(a2)'; this should read 'h(a1) = a2' for a transformation h mapping μ1 to μ2.

Circularity Check

2 steps flagged · score 4.0 of 10

Central barycenter existence/uniqueness and the proposed density family are imported from the author's own prior work [47] and [57], so the framework's foundations are load-bearing self-citations; the MLE derivations and higher-dimensional extensions are independent.

  1. uniqueness imported from authors [Section 3.1, Theorem 1 and Theorem 4; used again in Sections 3.2, 4.3, 5.1, 5.2]
    "We proceed with several facts and notions that have been introduced in [47]. We will omit the proofs, as they are provided therein. Theorem 1. Let {z1,...,zN} be a configuration of points in B2. Then, there exists a unique (up to a rotation) Möbius transformation ga ∈ G2, such that the configuration {ga(z1),...,ga(zN)} is balanced. ... Theorem 4. [47] The function (16) is geodesically convex in B2 and hence has a unique minimum in B2. This minimum is conformal barycenter of the configuration {z1,...,zN}."

    The paper's answer to question i)—existence and uniqueness of the mean—is exactly the content of Theorem 1 and Theorem 4, both attributed to the author's own prior paper [47] and not proved here. The conformal barycenter is defined through this imported uniqueness, and every later claim that the minimizer of H is the barycenter, including the MLE statement 'ˆa is precisely the conformal barycenter of points z1,...,zN' in Section 4.3 and the ball generalizations in Sections 5.1–5.2, inherits this same self-citation. The central mathematical object is thus not derived in the present paper but is taken as an input from a same-author source, making the 'rigorous answer to i)' a citation rather than an independent derivation.

  2. self citation load bearing [Section 4.1, Eq. (26) and Proposition 9; also Section 6.1–6.2]
    "We consider probability distributions on B2 defined by densities of the form (see [57]) p(x;a,s) = (s-1)/π * ((1-|z|^2)(1-|a|^2)/|1-az̄|^2)^s. ... We denote this family of probability distributions by Moeb2(a,s) and refer to these distributions as Möbius distributions in hyperbolic disc."

    The paper presents the Möbius family as the 'novel family' answering question iii), but the density is taken verbatim from the author's own prior paper [57] and no derivation from the group-theoretic framework is supplied. Proposition 9, the conformal invariance that makes the family usable for the stated ML tasks, is asserted without proof. The subsequent random-generation and MLE algorithms operate on this imported ansatz, so the statistical model is an input from a same-author citation rather than a consequence derived in the paper. This is load-bearing because the statistical sections' conclusions all presuppose that this particular family is the right one.

full rationale

The paper does not exhibit a fitted parameter renamed as a prediction, nor does it define a quantity in terms of the very quantity it claims to derive. The main mathematical mechanism—showing that Poincaré swarms (17) evolve by Möbius actions and that the resulting ODE is hyperbolic gradient flow—is a genuine derivation using Lie-group generators and the gradient-flow condition; those parts are not circular, even if the generator computation in Section 3.2 appears algebraically incorrect (a correctness concern, not a circularity concern). The circularity burden is concentrated in the foundations: the conformal barycenter's existence, uniqueness, and minimizing-potential characterization are imported from the same-author paper [47], and the Möbius density family is imported from the same-author paper [57]. These are load-bearing self-citations because the central definitions and the statistical model are not reproved or rederived here. However, the paper does add independent content: the explicit swarm dynamics, the MLE formula manipulations, the higher-dimensional Poincaré/Bergman extensions, and the sampling algorithms go beyond mere restatement. Thus the paper is partially circular but not wholly so; the score of 4 reflects substantial load-bearing self-citation with genuinely independent technical work around it.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No parameters are fitted; a and s are model parameters. The framework rests on the author's own unproved prior theorems, on standard Lie group machinery, and on a non-standard metric definition. No new physical entities are postulated; the Möbius distributions and swarms are mathematical constructs.

assumptions (4)
  • domain assumption Theorem 1 and Theorem 4 from [47]: existence and uniqueness of a balanced Möbius image, and geodesic convexity of the potential H(a).
    Section 3.1 states 'We will omit the proofs, as they are provided therein [47]'. These theorems define the conformal barycenter and are load-bearing for every later algorithm.
  • standard math Theorem 6 from [48]: trajectories of a time-dependent linear combination of infinitesimal generators evolve by group action.
    Section 3.2 invokes this result to claim Poincaré swarms evolve under the Möbius group.
  • ad hoc to paper The Poincaré ball metric as written in Definition 1, with denominator (1-|x|^2), defines a constant negative curvature manifold.
    Section 2.2, Definition 1: the standard metric has denominator (1-|x|^2)^2; the paper does not flag this deviation, yet all subsequent distance and gradient formulas use the standard metric.
  • standard math Bergman balls have constant negative sectional curvature, per [43].
    Section 2.3 relies on this standard result to treat Bergman balls as hyperbolic models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A group-theoretic framework for machine learning in hyperbolic spaces." pith.science (2026). https://pith.science/paper/6I6M44PP

@misc{pith2026250106934,
  author       = {Pith},
  title        = {Pith review of: A group-theoretic framework for machine learning in hyperbolic spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6I6M44PP}},
  note         = {Machine review of arXiv:2501.06934}
}
read the original abstract

Embedding the data in hyperbolic spaces can preserve complex relationships in very few dimensions, thus enabling compact models and improving efficiency of machine learning (ML) algorithms. The underlying idea is that hyperbolic representations can prevent the loss of important structural information for certain ubiquitous types of data. However, further advances in hyperbolic ML require more principled mathematical approaches and adequate geometric methods. The present study aims at enhancing mathematical foundations of hyperbolic ML by combining group-theoretic and conformal-geometric arguments with optimization and statistical techniques. Precisely, we introduce the notion of the mean (barycenter) and the novel family of probability distributions on hyperbolic balls. We further propose efficient optimization algorithms for computation of the barycenter and for maximum likelihood estimation. One can build upon basic concepts presented here in order to design more demanding algorithms and implement hyperbolic deep learning pipelines.

Figures

Figures reproduced from arXiv: 2501.06934 by the authors.

Figure 1
Figure 1. This Figure illustrates balanced configurations c [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Barycenters in the Poincaré disc. Remark 2. Results exposed in this Section are to a great extent inspired by geometric investigations of Kuramoto ensembles of coupled phase oscillators [49, 50]. Phase oscillators are described by their phases (angles) only, while amplitudes are neglected. Therefore, Kuramoto model with N oscillators can be treated as dynamical system on the N-torus S 1 × · · · × S 1 . In the contin… view at source ↗
Figure 3
Figure 3. Random samples in the Poincaré disc with parameter [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Random samples from the Poincare ball B 3 for parameter values: a) a = (0, 0, 0), s = 3; b) a = (0, 0, 0), s = 5; c) a = (0.9, 0, 0), s = 3; d) a = (0.9, 0, 0), s = 5. Holomorphic barycenter is holomorphicaly invariant, meaning that if a is holomorphic barycenter of th…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Cartan Networks: Group theoretical Hyperbolic Deep Learning

    cs.LG 2025-05 reject novelty 6.0 of 10

    Cartan networks compose solvable-group homomorphisms with isometries to define hyperbolic layers, and the paper reports competitive benchmark performance.

  2. Clustering in hyperbolic balls

    cs.LG 2025-01 reject novelty 4.0 of 10

    K-means and EM clustering for points in Poincaré hyperbolic balls are defined via conformal barycenters and Möbius distributions, with synthetic experiments in 2D and 3D.

Reference graph

Works this paper leans on

60 extracted references · 51 canonical work pages · cited by 2 Pith papers

  1. [47]

    & Kalaj, D

    Jacimovic, V . & Kalaj, D. Conformal and holomorphic barycenters in hyperbolic balls. arXiv:2410.02257. (2024)

    arXiv 2024

  2. [57]

    & Markovi ´c, M

    Ja ´cimovi´c, V . & Markovi ´c, M. Conformally Natural Families of Probability Distribu tions on Hyperbolic Disc with a View on Geometric Deep Learning. arXiv:2407.16733. (2024)

    arXiv 2024

  3. [1]

    &Boguná, M

    Krioukov, D., Papadopoulos, F., Kitsak, M., V ahdat, A. &Boguná, M. Hyperbolic geometry of complex networks. Physical Review E. 82, 036106 (2010)

    work page 2010

  4. [2]

    & Can nistraci, C

    Muscoloni, A., Thomas, J., Ciucci, S., Bianconi, G. & Can nistraci, C. Machine learning meets complex networks via coalescent embedding in the hyperbolic space. Nature Communications. 8, 1615 (2017)

    work page 2017

  5. [3]

    & V eli ˇckovi´c, P

    Bronstein, M., Bruna, J., Cohen, T. & V eli ˇckovi´c, P . Geometric deep learning: Grids, groups, graphs, geode sics, and gauges. arXiv:2104.13478. (2021)

    arXiv 2021

  6. [4]

    & Ganea, O

    Tifrea, A., Bécigneul, G. & Ganea, O. Poincaré GloV e: Hyp erbolic Word Embeddings. International Conference on Learning Representations. (2019)

    work page 2019

  7. [5]

    & Kiela, D

    Nickel, M. & Kiela, D. Learning continuous hierarchies i n the Lorentz model of hyperbolic geometry. Proceed- ings of the 35th International Conference on Machine Learni ng. pp. 3779-3788 (2018)

    work page 2018

  8. [6]

    & Wilson, B

    Leimeister, M. & Wilson, B. Skip-gram word embeddings in hyperbolic space. arXiv:1809.01498. (2018)

    arXiv 2018

Show all 60 references
  1. [7]

    & Y eung, S

    Mettes, P ., Ghadimi Atigh, M., Keller-Ressel, M., Gu, J. & Y eung, S. Hyperbolic deep learning in computer vision: A survey. International Journal of Computer Vision . 132 pp. 3484-3508 (2024)

    2024

  2. [8]

    & Lempitsky, V

    Khrulkov, V ., Mirvakhabova, L., Ustinova, E., Oseledet s, I. & Lempitsky, V . Hyperbolic image embeddings. Proceedings of the IEEE/CVF Conference on Computer Vision a nd Pattern Recognition. pp. 6418-6428 (2020)

    2020

  3. [9]

    Hyperbolic matrix factorization improves prediction of drug-target associations

    Poleksic, A. Hyperbolic matrix factorization improves prediction of drug-target associations. Scientific Reports . 13, 959 (2023)

    2023

  4. [10]

    & Lio, P

    Corso, G., Ying, R., Pándy, M., V eli ˇckovi´c, P ., Leskovec, J. & Lio, P . Neural distance embeddings for biological sequences. Proceedings of the 35th International Conference on Neural Information Processing Systems . pp. 18539-18551 (2021)

    2021

  5. [11]

    & V argas, S

    Chamberlain, B., Hardwick, S., Wardrope, D., Dzogang, F., Daolio, F. & V argas, S. Scalable hyperbolic recom- mender systems. arXiv:1902.08648. (2019)

    2019 arXiv

  6. [12]

    & Wang, Z

    Li, A., Y ang, B., Huo, H., Chen, H., Xu, G. & Wang, Z. Hyper bolic neural collaborative recommender. IEEE Transactions on Knowledge and Data Engineering . 35, 9114-9127 (2022)

    2022

  7. [13]

    Chami, I., Wolf, A., Juan, D., Sala, F., Ravi, S. & Ré, C. L ow-dimensional hyperbolic knowledge graph embed- dings. arXiv:2005.00545. (2020)

    2020 arXiv

  8. [14]

    & Pantazis, D

    Baker, C., Suárez-Méndez, I., Smith, G., Marsh, E., Fun ke, M., Mosher, J., Maestú, F., Xu, M. & Pantazis, D. Hyperbolic graph embedding of MEG brain networks to study br ain alterations in individuals with subjective cognitive decline. IEEE Journal of Biomedical and Health In...

    2024

  9. [15]

    & Serrano, M

    Allard, A. & Serrano, M. Navigable maps of structural br ain networks across species. PLoS Computational Biology. 16, e1007584 (2020)

    2020

  10. [16]

    & Balay, M

    Joseph, A., Francis, N. & Balay, M. Hyperbolic Brain Rep resentations. arXiv:2409.12990. (2024)

    2024 arXiv

  11. [17]

    & Hunt, J

    Cetin, E., Chamberlain, B., Bronstein, M. & Hunt, J. Hyp erbolic deep reinforcement learning. International Conference on Learning Representations. (2023)

    2023

  12. [18]

    , Miller, D

    Bogatskiy, A., Anderson, B., Offermann, J., Roussi, M. , Miller, D. & Kondor, R. Lorentz group equivariant neural network for particle physics. Proceedings of the 37th International Conference on Machin e Learning. pp. 992-1002 (2020)

    2020

  13. [19]

    & others Hyperbolic Attention Networks

    Gulcehre, C., Denil, M., Malinowski, M., Razavi, A., Pa scanu, R., Hermann, K., Battaglia, P ., Bapst, V ., Raposo, D., Santoro, A. & others Hyperbolic Attention Networks. International Conference on Learning Representations. (2018)

    2018

  14. [20]

    & Hofmann, T

    Ganea, O., Bécigneul, G. & Hofmann, T. Hyperbolic neura l networks. Proceedings of the 32nd International Conference on Neural Information Processing Systems . pp. 5350-5360 (2018)

    2018

  15. [21]

    & Nakayama, H

    Takeuchi, J., Nishida, N. & Nakayama, H. Neural network s in a product of hyperbolic spaces. Proceedings of the 2022 Conference of the North American Chapter of the Asso ciation for Computational Linguistics: Human Language T echnologies: Student Research W orkshop. pp. 211-221 (2022)

    2022

  16. [22]

    & Welling, M

    Cohen, T. & Welling, M. Group equivariant convolutiona l networks. Proceedings of the 33rd International Con- ference on Machine Learning . pp. 2990-2999 (2016)

    2016

  17. [23]

    & Kiela, D

    Nickel, M. & Kiela, D. Poincaré embeddings for learning hierarchical representations. Proceedings of the 31st International Conference on Neural Information Processin g Systems. pp. 6341-6350 (2017)

    2017

  18. [24]

    & Teh, Y .Continuous hierarchical representations with Poincaré variational auto-encoders

    Mathieu, E., Lan, C., Maddison, C., Tomioka, R. & Teh, Y .Continuous hierarchical representations with Poincaré variational auto-encoders. Proceedings of the 33rd International Conference on Neural Information Processing Systems. pp. 12565-12576 (2019)

    2019

  19. [25]

    Hyperbolic trigonometry and its application in the Poincaré ball model of hyperbolic geometry

    Ungar, A. Hyperbolic trigonometry and its application in the Poincaré ball model of hyperbolic geometry. Com- puters & Mathematics with Applications . 41, 135-147 (2001)

    2001

  20. [26]

    & Koyama, M

    Nagano, Y ., Y amaguchi, S., Fujita, Y . & Koyama, M. A wrap ped normal distribution on hyperbolic space for gradient-based learning. Proceedings of the 36th International Conference on Machin e Learning. pp. 4693-4702 (2019)

    2019

  21. [27]

    & Kim, D

    Cho, S., Lee, J., Park, J. & Kim, D. A rotated hyperbolic w rapped normal distribution for hierarchical represen- tation learning. Proceedings of the 36th International Conference on Neural Information Processing Systems. pp. 17831-17843 (2022)

    2022

  22. [28]

    & Zhao, G

    Peng, W ., V aranka, T., Mostafa, A., Shi, H. & Zhao, G. Hyp erbolic deep neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence . 44, 10023-10044 (2021)

    2021

  23. [29]

    & Welling, M

    Cohen, T. & Welling, M. Learning the irreducible repres entations of commutative Lie groups. Proceedings of the 31st International Conference on Machine Learning . pp. 1755-1763 (2014)

    2014

  24. [30]

    Lie group machine learning and Gibbs den sity on Poincaré unit disk from Souriau Lie groups thermodynamics and SU(1,1) coadjoint orbits

    Barbaresco, F. Lie group machine learning and Gibbs den sity on Poincaré unit disk from Souriau Lie groups thermodynamics and SU(1,1) coadjoint orbits. Geometric Science of Information: 4th International Confe rence. pp. 157-170 (2019)

    2019

  25. [31]

    & Hauberg, S

    Arvanitidis, G., Hansen, L. & Hauberg, S. Latent Space O ddity: on the Curvature of Deep Generative Models. International Conference on Learning Representations . (2018)

    2018

  26. [32]

    & Duin, R

    Wilson, R., Hancock, E., Pekalska, E. & Duin, R. Spheric al and hyperbolic embeddings of data. IEEE Transac- tions on Pattern Analysis and Machine Intelligence . 36, 2255-2269 (2014)

    2014

  27. [33]

    & Calder, C

    Smith, A., Asta, D. & Calder, C. The geometry of continuo us latent space models for network data. Statistical Science: A Review Journal of the Institute of Mathematical S tatistics. 34, 428 (2019)

    2019

  28. [34]

    & Hamilt on, W

    Bose, J., Smofsky, A., Liao, R., Panangaden, P . & Hamilt on, W . Latent variable modelling with hyperbolic nor- malizing flows. Proceedings of the 37th International Conference on Machin e Learning. pp. 1045-1055 (2020)

    2020

  29. [35]

    & Strapasson, J

    Costa, S., Santos, S. & Strapasson, J. Fisher informati on distance: A geometrical reading. Discrete Applied Mathematics. 197 pp. 59-69 (2015)

    2015

  30. [36]

    & Mettes, P

    Atigh, M., Keller-Ressel, M. & Mettes, P . Hyperbolic Bu semann learning with ideal prototypes. Proceedings of the 35th International Conference on Neural Information Pr ocessing Systems. pp. 103-115 (2021)

    2021

  31. [37]

    Chami, I., Gu, A., Nguyen, D. & Ré, C. HoroPCA: Hyperboli c dimensionality reduction via horospherical projections. Proceedings of the 38th International Conference on Machin e Learning. pp. 1419-1429 (2021) 21 A group-theoretic framework for machine learning in hyperb olic sp...

    2021

  32. [38]

    & Ostermeier, A

    Hansen, N. & Ostermeier, A. Completely derandomized se lf-adaptation in evolution strategies. Evolutionary Computation. 9, 159-195 (2001)

    2001

  33. [39]

    Harmonic and subharmonic function theory on t he hyperbolic ball

    Stoll, M. Harmonic and subharmonic function theory on t he hyperbolic ball. (Cambridge University Press,2016)

    2016

  34. [40]

    Notes on complex hyperbolic geometry

    Parker, J. Notes on complex hyperbolic geometry. (http s://www.icts.res.in/sites/default/files/NCHG.pdf), Ac - cessed: 11 January 2025

    2025

  35. [41]

    Möbius Transformations in Several Dimensi ons

    Ahlfors, L. Möbius Transformations in Several Dimensi ons. (University of Minnesota,1981)

    1981

  36. [42]

    Spaces of holomorphic functions in the unit ball

    Zhu, K. Spaces of holomorphic functions in the unit ball . (Springer,2005)

    2005

  37. [43]

    Huang, X. & Li, S. Bergman metrics with constant holomor phic sectional curvatures. arXiv:2302.13456v2. (2024)

    2024 arXiv

  38. [44]

    & Y ang, L

    Arnaudon, M., Barbaresco, F. & Y ang, L. Medians and mean s in Riemannian geometry: Existence, uniqueness and computation. F . Nielsen, R. Bhatia (Eds.) Matrix Information Geometry. pp. 169-197 (2012)

    2012

  39. [45]

    Convex Analysis and Optimization in Hadamard Spaces

    Bacak, M. Convex Analysis and Optimization in Hadamard Spaces. (De Gruyter,2014)

    2014

  40. [46]

    & D e Sa, C

    Lou, A., Katsman, I., Jiang, Q., Belongie, S., Lim, S. & D e Sa, C. Differentiating through the Fréchet mean. Proceedings of the 37th International Conference on Machin e Learning. pp. 6393-6403 (2020)

    2020

  41. [48]

    Applications of Lie groups to differential eq uations

    Olver, P . Applications of Lie groups to differential eq uations. (Springer,1993)

    1993

  42. [49]

    Self-entrainment of a population of coupl ed nonlinear oscillators

    Kuramoto, Y . Self-entrainment of a population of coupl ed nonlinear oscillators. H. Araki (Eds.) International Symposium on Mathematical Problems in Theoretical Physics . pp. 420-422 (1975)

    1975

  43. [50]

    & Strogatz, S

    Marvel, S., Mirollo, R. & Strogatz, S. Identical phase o scillators with global sinusoidal coupling evolve by Möbius group action. Chaos: An Interdisciplinary Journal of Nonlinear Science . 19 (2009)

    2009

  44. [51]

    & Strogatz, S

    Watanabe, S. & Strogatz, S. Constants of motion for supe rconducting Josephson arrays. Physica D: Nonlinear Phenomena. 74, 197-253 (1994)

    1994

  45. [52]

    & Mirollo, R

    Chen, B., Engelbrecht, J. & Mirollo, R. Hyperbolic geom etry of Kuramoto oscillator networks. Journal of Physics A: Mathematical and Theoretical . 50, 355101 (2017)

    2017

  46. [53]

    Computing the Douady–Earle extension using Kuramoto o scillators

    Ja ´cimovi´c, V . Computing the Douady–Earle extension using Kuramoto o scillators. Analysis and Mathematical Physics. 9 pp. 523-529 (2019)

    2019

  47. [54]

    & Earle, C

    Douady, A. & Earle, C. Conformally natural extension of homeomorphisms of the circle. Acta Mathematica. 157 pp. 23-48 (1986)

    1986

  48. [55]

    & Strogatz, S

    Lipton, M., Mirollo, R. & Strogatz, S. The Kuramoto mode l on a sphere: Explaining its low-dimensional dynam- ics with group theory and hyperbolic geometry. Chaos: An Interdisciplinary Journal of Nonlinear Science . 31 (2021)

    2021

  49. [56]

    Consensus on nonlinear spaces

    Sepulchre, R. Consensus on nonlinear spaces. Annual Reviews in Control. 35, 56-64 (2011)

    2011

  50. [58]

    The Möbius distribution on the disc

    Jones, M. The Möbius distribution on the disc. Annals of the Institute of Statistical Mathematics . 56 pp. 733-742 (2004)

    2004

  51. [59]

    Sala, F., De Sa, C., Gu, A. & Ré, C. Representation tradeo ffs for hyperbolic embeddings. Proceedings of the 35th International Conference on Machine Learning . pp. 4460-4469 (2018)

    2018

  52. [60]

    & Zemel, R

    Law, M., Liao, R., Snell, J. & Zemel, R. Lorentzian dista nce learning for hyperbolic representations. Proceedings of the 36th International Conference on Machine Learning . pp. 3672-3681 (2019) 22

    2019

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.