Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

Navigation through Non-Compact Symmetric Spaces: a mathematical perspective on Cartan Neural Networks

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a neural network's nonlinearity can be supplied entirely by geometry: each layer is a non-compact symmetric space and each layer map is a covariant composition of a solvable-group homomorphism and an isometry, with…

desk verdict A serious but over-advertised framework: the master layer map and worked examples are genuinely new, yet the claimed full generality rests on an unreviewed companion preprint and a false curvature blanket statement. read the letter →

arxiv 2507.16871 v1 pith:FI5CM67T submitted 2025-07-22 cs.LG hep-th

Pietro Giuseppe Fré , Federico Milanesio , Guido Sanguinetti , Matteo Santoro This is my paper
classification cs.LGhep-th MSC 17B3022E4653C3568T07
keywords Cartanneuralnetworksnon-compactsymmetricspacessolvableLiegroupscovariancehyperbolicMaurer-CartanformsgeometricdeeplearningTits-Satakeclasses
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 establish that the nonlinearity of a neural network can come entirely from the geometry of non-compact symmetric spaces, with no pointwise activation functions. Each hidden layer is a symmetric space $U/H$, and each transition between layers is a covariant map built from a linear homomorphism of solvable Lie algebras followed by an isometry of the target layer; the learned parameters are the homomorphism and the isometry parameters, so they have intrinsic geometric meaning. The central explicit construction is the master formula (3.59) for the layer map, which in the rank-one hyperbolic case reduces to the closed form $h(Y)=(Y_1, WY_2+(1-e^{-Y_1})b)$. If the construction is sound, the geometric interpretability lost in classical Euclidean networks is restored, and the same mechanism extends beyond hyperbolic spaces to higher-rank symmetric spaces.

What carries the argument

The load-bearing object is the metric equivalence between every non-compact symmetric space $U/H$ and a solvable Lie group $S_{U/H}$ (Theorem 3.2 plus the canonical triangular embedding of Statement 3.1, both imported from the companion foundational paper [2]). On top of this, the argument uses Maurer-Cartan one-forms: a linear map $W$ from one solvable Lie algebra to another is a homomorphism exactly when it pulls the target Maurer-Cartan equations back to the source ones, and solving $E^i = W^i_\alpha \varepsilon^\alpha$ gives the nonlinear coordinate map $\Phi[W|\cdot]$. The isometry group of the target layer then supplies the final factor in the master formula (3.59), which combines the solvable homomorphism, the parametrization $\Sigma$, and an adjoint rotation by $g(\Psi)$ into one covariant layer map.

What would settle it

Compute the pullback of the target Maurer-Cartan forms for one explicit transition, say from $H^{1+q}$ to $H^{1+q'}$ using a generic matrix $W$, and check whether equation (4.22) solves equations (4.20)-(4.21); a single admissible $W$ whose pullback violates those equations would falsify the claimed general form. Alternatively, exhibit one non-compact symmetric space in the pseudo-orthogonal series for which the solvable subgroup $S_{U/H}$ is not metrically equivalent to $U/H$, which would remove the foundation of the master formula (3.59).

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the map from the $i$-th to the $(i+1)$-th layer of a Cartan neural network can be written as $$\hat K_i = \mathrm{Adj}_{g(\Psi_{i+1})} \circ \varpi_{i+1}^{-1} \circ \mho_{W_i} \circ \varpi_i,$$ where $\varpi$ identifies the symmetric space with its solvable Lie group, $\mho_{W_i}$ is the solvable-group homomorphism induced by a linear map $W_i$ of solvable Lie algebras, and $\mathrm{Adj}_{g(\Psi_{i+1})}$ is an isometry of the target symmetric space. This composition is covariant by construction: every ingredient is defined intrinsically, without a preferred coordinate basis. In the $r=1$ hyperbolic case the construction yields the closed form $h(Y)=(Y_1, WY_2+(1-e^{-Y_1})b)$, so the bias $b$ enters through an exponential factor and the entire nonlinearity is geometric. The paper therefore claims that pointwise activation functions are not needed: covariance is built in, the learned parameters are a Lie algebra homomorphism and target-space isometries, and classification can be rephrased as logistic regression on the oriented geodesic distance from a separator submanifold.

Load-bearing premise

The construction rests on the claim that every space used as a hidden layer can be identified with a solvable group of symmetries of the same dimension in a way that preserves distances; if that identification fails for any of the spaces, the layer maps as written do not exist.

Editorial extensions

If this is right

  • The layer maps are covariant, so the nonlinearity and the learned parameters do not depend on a choice of coordinate basis in any layer.
  • Each learned matrix $W$ is a genuine homomorphism of solvable Lie algebras, giving every parameter an intrinsic geometric role rather than an ad hoc one.
  • In the rank-one hyperbolic case the whole layer transition is the closed formula $h(Y)=(Y_1, WY_2+(1-e^{-Y_1})b)$, an explicit and cheaply computable map in which the bias enters exponentially.
  • The construction applies beyond hyperbolic spaces: higher-rank spaces such as $SL(N,\mathbb{R})/SO(N)$ provide several Cartan coordinates and polynomial nonlinearities of increasing degree, and the layer maps may even switch Tits-Satake universality classes.
  • Classification can be implemented as logistic regression on the oriented distance from a separator that is a conjugate of a lower-rank symmetric submanifold, with softmax for many classes.

Reading between the lines

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

  • Beyond the paper: if the construction is numerically stable, the universal-approximation question shifts from choosing activation functions to the expressivity of solvable-Lie-algebra homomorphisms; a natural conjecture is that shallow $SL(N,\mathbb{R})/SO(N)$ layers approximate continuous functions on compacta as $N$ grows.
  • Beyond the paper: because $W$ is an honest Lie algebra homomorphism, row-reduced forms of trained $W$ matrices should reveal which coordinate types (Cartan versus nilpotent) survive from layer to layer, providing the post-hoc interpretability tool the paper lists as future work.
  • Beyond the paper: one testable prediction is that trained classification boundaries in the last hyperbolic layer coincide, up to isometry, with copies of the lower-rank submanifold $M_{1,q-1}$; checking this against fitted networks would separate the geometric hypothesis from generic separator fitting.
  • Beyond the paper: comparing the $r=1$ map (4.22) with a same-parameter-count classical multilayer perceptron on data with known hierarchy would test whether the solvable-coordinate grading actually matches hierarchical structure, as the geometric distance-uniqueness motivation suggests.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a geometric reformulation of neural-network layer maps using non-compact symmetric spaces U/H, viewed through their solvable Lie group models. Its central object is the master formula (3.59), which expresses a layer transition as a target-space isometry composed with the inverse solvable parametrization, a solvable-group homomorphism induced by a Lie algebra map W_i, and the source parametrization. For the r=1 hyperbolic case the formula reduces to the explicit map h(Y) = (Y_1, W Y_2 + (1 - e^{-Y_1}) b) in eq. (4.22). The paper also discusses Tits-Satake class-switching homomorphisms, with worked examples in Appendices A and B, and develops a geometric account of classification through separators, oriented distance, and logistic/softmax maps. The presentation is framed as the mathematical companion to the authors' twin paper arXiv:2505.24353 on Cartan Networks.

Significance. If the master formula is valid, it offers a parameterization-invariant, activation-free layer transition, replacing pointwise nonlinearities with solvable Lie group homomorphisms and isometries. The r=1 computation is concrete and checkable, and its derivation in Appendix A is consistent with the Maurer-Cartan consistency conditions. The class-switching examples in Appendix B are also explicit and useful as computational templates. However, the full-generality claim is not established: it depends on an imported theorem from the authors' own preprint and on an embedding statement that explicitly covers only part of the Cartan classification, while the r>=2 separator construction is deferred to a forthcoming paper. The most solid contributions are the r=1 formula and the method for computing solvable Lie algebra homomorphisms, not the claimed full universality over all non-compact symmetric spaces.

major comments (3)
  1. [§3.5, Eq. (3.59), property (b)] The claim that eq. (3.59) is 'fully general within the class of chosen manifolds, namely the non-compact symmetric spaces' is not supported by the material presented. Theorem 3.2 is quoted from the authors' preprint [2] and is not proved here; Statement 3.1 explicitly covers only real sections of types b_l, c_l, d_l with q even, while type a_l is treated separately in Section 5 only for the Borel subalgebra of sl(N,R), and the exceptional algebras e6, e7, e8, f4, g2 are not addressed. Since the maps ϖ_i and ℧_{W_i} depend on the solvable-group equivalence and the triangular embedding, the existence of the layer maps for all non-compact symmetric spaces is exactly as strong as these unproved inputs. I recommend either proving or citing a complete standard reference for the solvable-group equivalence and triangular embedding for every classical and exceptional family, or restricting the statement of (3.59) to the pseudo-orthogonal families (3.3) for which the construction is actually carried out.
  2. [§1.2 (after Eq. (1.1))] The statement that all non-compact symmetric spaces are hyperbolic because they are Cartan-Hadamard manifolds with everywhere negative sectional curvature is incorrect. Rank-two and higher non-compact symmetric spaces, for example SL(3,R)/SO(3), contain flat totally geodesic subspaces, so some sectional curvatures vanish; the correct invariant is non-positive sectional curvature. The later arguments in the paper only require the Cartan-Hadamard property, so the claim can be repaired by replacing 'everywhere negative' with 'non-positive' and by not using 'hyperbolic' as a synonym for all non-compact symmetric spaces.
  3. [§6.3, Definition 6.1; §4 after Eq. (4.1)] The construction of separators for r >= 2 is asserted without proof: the text states that for r >= 2 separators are codimension-one homogeneous spaces that are neither symmetric nor totally geodesic, and refers to the forthcoming paper [40]. Since Definition 6.1 and the logistic/softmax formulas (6.16) and (6.24) require a parameterized separator S[θ] and an oriented distance δ̂, the classification generalization to r >= 2 is conditional on [40]. The manuscript should either provide the existence and geometric-property proof for such separators or explicitly state that the classification section applies only to the r=1 case.
minor comments (6)
  1. [§3.4, Eqs. (3.41)-(3.43)] The linear map W is first named Σ, and the same symbol Σ is then used for the parametrization map from the solvable Lie algebra to the solvable group; this notational conflict makes the text confusing and should be resolved.
  2. [§3.3.2, Eq. (3.38)] The spelling 'Cholewsky-Crout' should be 'Cholesky-Crout'.
  3. [§2.1] The word 'perception' appears where 'perceptron' is meant in several places; these should be corrected.
  4. [§4.2.2, after Eq. (4.22)] The displayed dimension of the matrix W is ambiguous: the block form [1 0; 0 W] indicates that the lower-right block is separated from the first row and column, but the text writes W in R^{(1+q_{i+1})×(1+q_i)} without this clarification.
  5. [§6.3, Definition 6.1] The conditions M = M_+ ∪ M_- and M_+ ∩ M_- = S are inconsistent unless closures are intended; the definition should specify whether M_± are open halves or closed halves of the manifold.
  6. [Appendices C and D] The probability-theory and historical appendices are clearly written but contain mostly textbook material; they could be shortened or moved to a supplementary document without affecting the main argument.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the layer-map derivation is a formal composition of standard Lie-theoretic constructions; imported self-cited theorems are independent support, not inputs to the claimed result.

full rationale

The central map (3.59) is not a restatement of an input: K_i is defined as a composition of (i) the Cholesky-type coordinate map ϖ, (ii) the solvable-group homomorphism ℧_W obtained by solving Maurer-Cartan equations with a linear algebra map W, and (iii) an isometry Adj_{g(Ψ)}. The r=1 reduction (4.22) is obtained in Appendix A by imposing the pullback Maurer-Cartan equations (A.1)-(A.5) and solving for coordinates, so W and b are free parameters characterizing genuine homomorphisms rather than fitted values disguised as predictions. The main caveat is external-support, not circular: Theorem 3.2 (Th. 1.1 of [2]) and Statement 3.1 are quoted from the authors' own preprint [2] rather than proved here; Statement 3.1 is explicitly limited to real forms b_l, c_l, d_l, with a_l handled separately in Section 5 and exceptional types not treated, while Section 3.5(b) claims full generality over non-compact symmetric spaces. This is a scope/completeness gap that should be addressed by a proof or a restricted statement. It is not a circularity because the imported metric equivalence is a parameter-free classical result with independent antecedents in the literature (e.g., Alekseevsky's normal solvable groups, Helgason), and the embedding statement is a concrete linear-algebra construction that does not assume the layer-map formula. The paper contains no fitted-input-called-prediction: no numerical experiment is performed here, and the only empirical claims are explicitly delegated to the twin paper [11].

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities or fitted constants. Its central claim depends on the imported solvable-group equivalence from [2], the standard lift from Lie algebra to Lie group homomorphisms, and on two unproven or overly broad auxiliary assertions about negative curvature and higher-rank separators.

assumptions (5)
  • domain assumption Every noncompact symmetric space U/H is metrically equivalent to a solvable Lie group S_U/H.
    Imported as Theorem 3.2 from the authors' own preprint [2] and used throughout, especially in defining the master layer map in equation (3.59).
  • domain assumption The canonical triangular embedding U into sl(N,R) exists, with Solv(U/H) realized by upper triangular matrices.
    Statement 3.1 is imported from [2] and underpins the coordinate parameterizations and the computability of the maps.
  • standard math A Lie algebra homomorphism lifts uniquely to a Lie group homomorphism.
    Invoked as Lemma 3.1 from Hall's textbook, and used to define the map from solvable Lie algebra homomorphisms to layer maps.
  • ad hoc to paper All noncompact symmetric spaces have everywhere negative sectional curvature.
    Asserted in Section 1.2 to justify calling all noncompact symmetric spaces hyperbolic. This is false for higher-rank spaces, which contain flat subspaces, and the claim is not needed for the main algebraic construction.
  • ad hoc to paper For r>=2, separators exist as codimension-one homogeneous spaces that are neither symmetric nor totally geodesic.
    Stated in Section 4 and needed for the classification discussion in Section 6, but no proof is given and the construction is deferred to the forthcoming paper [40].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Navigation through Non-Compact Symmetric Spaces: a mathematical perspective on Cartan Neural Networks." pith.science (2026). https://pith.science/paper/FI5CM67T

@misc{pith2026250716871,
  author       = {Pith},
  title        = {Pith review of: Navigation through Non-Compact Symmetric Spaces: a mathematical perspective on Cartan Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FI5CM67T}},
  note         = {Machine review of arXiv:2507.16871}
}
read the original abstract

Recent work has identified non-compact symmetric spaces U/H as a promising class of homogeneous manifolds to develop a geometrically consistent theory of neural networks. An initial implementation of these concepts has been presented in a twin paper under the moniker of Cartan Neural Networks, showing both the feasibility and the performance of these geometric concepts in a machine learning context. The current paper expands on the mathematical structures underpinning Cartan Neural Networks, detailing the geometric properties of the layers and how the maps between layers interact with such structures to make Cartan Neural Networks covariant and geometrically interpretable. Together, these twin papers constitute a first step towards a fully geometrically interpretable theory of neural networks exploiting group-theoretic structures

Figures

Figures reproduced from arXiv: 2507.16871 by the authors.

Figure 1
Figure 1. Simplified conceptual diagram of a perceptron or neuron. The neuron receives an input composed of a [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Conceptual scheme of the Multi-Layer Perceptron, original name for what was later named a Deep Neural [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. The Dynkin diagram of aℓ type is done in the following way. We consider the ℓ + 1–dimensional Euclidean space R ℓ+1 and let ϵ1, . . . , ϵℓ+1 denote the unit vectors along the ℓ + 1 axes: ϵ1 =   1 0 . . . . . . 0   , ϵ2 =   0 1 0 . . . 0   . . . ϵℓ+1 =   0 0 . . . . . . 1   (5.4) Given the vector v = ϵ1 + ϵ2 + · · · + ϵℓ+1 we define I ⊂ R ℓ+1 to be the ℓ + 1 –dime… view at source ↗

Discussion (0). Sign in 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. General Properties of the Thermo-Metric for CV event manifolds and the magnetization combinatorial scheme

    hep-th 2026-07 conditional novelty 6.0 of 10

    Freezing contiguous magnetic fields on CV thermo-metrics produces universal curved M_n|reg manifolds that embed via a_{2n-1} Cartan metrics and send geodesics to hypercube vertices/face centers by initial slope.

  2. Thermodynamics a la Souriau on K\"ahler Non Compact Symmetric Spaces for Cartan Neural Networks

    cs.IT 2025-12 conditional novelty 6.0 of 10

    On non-compact symmetric spaces supporting Souriau-type Gibbs distributions are exactly the Kähler ones; their convergence domain is the U-adjoint orbit of a positivity chamber in the compact Cartan subalgebra of H — ...

Reference graph

Works this paper leans on

57 extracted references · 54 canonical work pages · cited by 2 Pith papers

  1. [2]

    The Paint Group Tits Satake Theory of Hyperbolic Symmetric Spaces: the distance function, paint invariants and discrete subgroups,

    U. Bruzzo, P. G. Fr´ e, and M. Trigiante, “The Paint Group Tits Satake Theory of Hyperbolic Symmetric Spaces: the distance function, paint invariants and discrete subgroups,” arXiv, 2025. arXiv:2503.07626 [math.DG]

    arXiv 2025

  2. [40]

    Tessellation Groups, Harmonic Analysis on Hyperbolic U /H manifolds and the Heat Kernel in view of Cartan Convolutional Neural Networks,

    P. G. Fr´ e, F. Milanesio, M. Oyarzo, M. Santoro, and M. Trigiante, “Tessellation Groups, Harmonic Analysis on Hyperbolic U /H manifolds and the Heat Kernel in view of Cartan Convolutional Neural Networks,” arXiv (to appear), 2025

    work page 2025

  3. [1]

    Expressivity of deep neural networks,

    I. G¨ uhring, M. Raslan, and G. Kutyniok, “Expressivity of deep neural networks,” in Mathematical Aspects of Deep Learning, p. 149–199, Cambridge University Press, 2022

    work page 2022

  4. [3]

    Helgason, Differential geometry and symmetric spaces

    S. Helgason, Differential geometry and symmetric spaces . Academic press, 1962

    work page 1962

  5. [4]

    Die Zusammensetzung der stetigen endlichen Transformationsgruppen,

    W. Killing, “Die Zusammensetzung der stetigen endlichen Transformationsgruppen,” Math. Ann., vol. 33, no. 1, pp. 1–48, 1888

    work page

  6. [5]

    ¨Uber die einfachen Transformationsgruppen,

    ´E. Cartan, “ ¨Uber die einfachen Transformationsgruppen,” Ber. Verh. K¨ onigl. S¨ achs. Ges. Wiss. Leipzig, Math.- Phys. Kl. , vol. 45, pp. 395–420, 1893

    work page

  7. [6]

    P. G. Fr´ e,A Conceptual History of Symmetry from Plato to Special Geometries . Springer, 2018

    work page 2018

  8. [7]

    Lectures on special K¨ ahler geometry and electric - magnetic duality rotations,

    P. G. Fr´ e, “Lectures on special K¨ ahler geometry and electric - magnetic duality rotations,”Nucl. Phys. B Proc. Suppl., vol. 45, pp. 59–114, 1996

    work page 1996

Show all 57 references
  1. [8]

    Classification of quaternionic spaces with a transitive solvable group of motions,

    D. Alekseevsky, “Classification of quaternionic spaces with a transitive solvable group of motions,” Math. USSR Izvestija, vol. 9, pp. 297–339, 1975

    1975

  2. [9]

    Alekseevskian spaces,

    V. Cort´ es, “Alekseevskian spaces,”Diff. Geom. Appl. , vol. 6, pp. 129–168, 1996

    1996

  3. [10]

    Polyvector superPoincar´ e algebras,

    D. V. Alekseevsky, V. Cortes, C. Devchand, and A. Van Proeyen, “Polyvector superPoincar´ e algebras,”Commun. Math. Phys. , vol. 253, pp. 385–422, 2004

    2004

  4. [11]

    Cartan Networks: Group theoretical Hyperbolic Deep Learning,

    P. G. Fr´ e, F. Milanesio, M. Santoro, and G. Sanguinetti, “Cartan Networks: Group theoretical Hyperbolic Deep Learning,” arXiv, 2025. arXiv:2505.24353 [cs.LG]

    2025 arXiv

  5. [12]

    Poincar´ e Embeddings for Learning Hierarchical Representations,

    M. Nickel and D. Kiela, “Poincar´ e Embeddings for Learning Hierarchical Representations,” in Adv. Neural Inf. Process. Syst., vol. 30, Curran Associates, Inc., 2017

    2017

  6. [13]

    Hyperbolic Neural Networks,

    O.-E. Ganea, G. Becigneul, and T. Hofmann, “Hyperbolic Neural Networks,” in Adv. Neural Inf. Process. Syst. , vol. 31, Curran Associates, Inc., 2018

    2018

  7. [14]

    Poincar´ e maps for analyzing complex hierarchies in single-cell data,

    A. Klimovskaia, D. Lopez-Paz, L. Bottou, and M. Nickel, “Poincar´ e maps for analyzing complex hierarchies in single-cell data,” Nat. Commun. , vol. 11, no. 1, p. 2966, 2020

    2020

  8. [15]

    Hyperbolic Deep Neural Networks: A Survey,

    W. Peng, T. Varanka, A. Mostafa, H. Shi, and G. Zhao, “Hyperbolic Deep Neural Networks: A Survey,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 44, no. 12, pp. 10023–10044, 2022

    2022

  9. [16]

    Fully Hyperbolic Convolutional Neural Networks for Computer Vision,

    A. Bdeir, K. Schwethelm, and N. Landwehr, “Fully Hyperbolic Convolutional Neural Networks for Computer Vision,” in Int. Conf. Learn. Represent. , 2024

    2024

  10. [17]

    Hyperbolic neural networks++,

    R. Shimizu, Y. Mukuta, and T. Harada, “Hyperbolic neural networks++,” in Int. Conf. Learn. Represent., 2021

    2021

  11. [18]

    Fully hyperbolic neural networks,

    W. Chen, X. Han, Y. Lin, H. Zhao, Z. Liu, P. Li, M. Sun, and J. Zhou, “Fully hyperbolic neural networks,” in Proc. of ACL 2022 , pp. 5672–5686, ACL, 2022

    2022

  12. [19]

    A. A. Ungar, A Gyrovector Space Approach to Hyperbolic Geometry. Synth. Lect. Math. Stat., Morgan & Claypool, 2009. 57

    2009

  13. [20]

    Hyperbolic Convolutional Neural Networks,

    A. Skliar and M. Weiler, “Hyperbolic Convolutional Neural Networks,” arXiv, 2023. arXiv:2308.15639 [cs.LG]

    2023 arXiv

  14. [21]

    On the universal statistical consistency of expansive hyperbolic deep convolutional neural networks,

    S. Ghosh, K. Bose, and S. Das, “On the universal statistical consistency of expansive hyperbolic deep convolutional neural networks,” arXiv, 2024. arXiv:2411.10128 [stat.ML]

    2024 arXiv

  15. [22]

    Hyperbolic Graph Convolutional Neural Networks,

    I. Chami, Z. Ying, C. R´ e, and J. Leskovec, “Hyperbolic Graph Convolutional Neural Networks,” in Adv. Neural Inf. Process. Syst., vol. 32, Curran Associates, Inc., 2019

    2019

  16. [23]

    Hyperbolic attention networks,

    C. Gulcehre, M. Denil, M. Malinowski, A. Razavi, R. Pascanu, K. M. Hermann, P. Battaglia, V. Bapst, D. Raposo, A. Santoro, and N. de Freitas, “Hyperbolic attention networks,” in Int. Conf. Learn. Represent. , 2019

    2019

  17. [24]

    Castellani, R

    L. Castellani, R. D’Auria, and P. G. Fr´ e, Supergravity and superstrings: A Geometric perspective. Vol. 1,2,3. World Scientific, 1991

    1991

  18. [25]

    Gilmore, Lie groups, Lie algebras, and some of their applications

    R. Gilmore, Lie groups, Lie algebras, and some of their applications . Dover Publications, 2016. Renewed edition. First published by John Wiley & Sons in 1974

    2016

  19. [26]

    P. G. Fr´ e,Gravity, a Geometrical Course , vol. 1,2. Springer Science & Business Media, 2012

    2012

  20. [27]

    A logical calculus of the ideas immanent in nervous activity,

    W. S. McCulloch and W. Pitts, “A logical calculus of the ideas immanent in nervous activity,” Bull. Math. Biophys., vol. 5, no. 4, pp. 115–133, 1943

    1943

  21. [28]

    The perceptron: A probabilistic model for information storage and organization in the brain,

    F. Rosenblatt, “The perceptron: A probabilistic model for information storage and organization in the brain,” Psychol. Rev., vol. 65, no. 6, pp. 386–408, 1958

    1958

  22. [29]

    Approximation by superpositions of a sigmoidal function,

    G. Cybenko, “Approximation by superpositions of a sigmoidal function,” Math. Control Signals Syst. , vol. 2(4), pp. 303–314, 1989

    1989

  23. [30]

    Approximation by ridge functions and neural networks with one hidden layer,

    C. K. Chui and X. Li, “Approximation by ridge functions and neural networks with one hidden layer,” J. Approx. Theory, vol. 70(2), pp. 131–141, 1992

    1992

  24. [31]

    On the approximate realization of continuous mappings by neural networks,

    K. Funahashi, “On the approximate realization of continuous mappings by neural networks,” Neural Netw. , vol. 2(3), pp. 183–192, 1989

    1989

  25. [32]

    Approximation capabilities of multilayer feedforward networks,

    K. Hornik, “Approximation capabilities of multilayer feedforward networks,” Neural Netw., vol. 4(2), pp. 251–257, 1991

    1991

  26. [33]

    Construction of neural nets using the radon transform,

    S. Carroll and B. Dickinson, “Construction of neural nets using the radon transform,” International 1989 Joint Conference on Neural Networks , vol. 1, pp. 607–611, 1989

    1989

  27. [34]

    Multilayer feedforward networks with a nonpolynomial activation function can approximate any function,

    M. Leshno, V. Y. Lin, A. Pinkus, and S. S., “Multilayer feedforward networks with a nonpolynomial activation function can approximate any function,” Neural Netw., vol. 6(6), pp. 861 – 867, 1993

    1993

  28. [35]

    Semisimple Lie Algebras,

    J. E. Humphreys, “Semisimple Lie Algebras,” in Introduction to Lie Algebras and Representation Theory, pp. 15– 41, Springer, 1972

    1972

  29. [36]

    P. G. Fr´ e,Discrete, Finite and Lie Groups . De Gruyter, 2023

    2023

  30. [37]

    Extremal multicenter black holes: nilpotent orbits and Tits Satake universality classes,

    P. G. Fr´ e and A. S. Sorin, “Extremal multicenter black holes: nilpotent orbits and Tits Satake universality classes,” J. High Energy Phys. , vol. 2013, no. 1, 2013

    2013

  31. [38]

    Black Hole Nilpotent Orbits and Tits Satake Universality Classes,

    P. G. Fr´ e, A. S. Sorin, and M. Trigiante, “Black Hole Nilpotent Orbits and Tits Satake Universality Classes,” arXiv, 2011. arXiv:1107.5986 [hep-th]. 58

    2011 arXiv

  32. [39]

    P. G. Fr´ e,Advances in Geometry and Lie Algebras from Supergravity . Theor. Math. Phys., Springer, 2018

    2018

  33. [41]

    Cosmic billiards with painted walls in non-maximal supergravities: a worked out example,

    P. G. Fr´ e, F. Gargiulo, and K. Rulik, “Cosmic billiards with painted walls in non-maximal supergravities: a worked out example,” Nucl. Phys. B , vol. 737, no. 1, pp. 1–48, 2006

    2006

  34. [42]

    Tits–Satake projections of homogeneous special geometries,

    P. G. Fr´ e, F. Gargiulo, J. Rosseel, K. Rulik, M. Trigiante, and A. Van Proeyen, “Tits–Satake projections of homogeneous special geometries,” Class. Quantum Gravity , vol. 24, no. 1, pp. 27–78, 2006

    2006

  35. [43]

    B. C. Hall, Lie Groups, Lie Algebras, and Representations: An Elementary Introduction , vol. 222 of Grad. Texts Math. Springer, 2015

    2015

  36. [44]

    Cosmological backgrounds of superstring theory and solvable algebras: Oxidation and branes,

    P. Fr´ e, V. Gili, F. Gargiulo, A. S. Sorin, K. Rulik, and M. Trigiante, “Cosmological backgrounds of superstring theory and solvable algebras: Oxidation and branes,” Nucl. Phys. B , vol. 685, pp. 3–64, 2004

    2004

  37. [45]

    Integrability of supergravity billiards and the generalized Toda lattice equation,

    P. G. Fr´ e and A. S. Sorin, “Integrability of supergravity billiards and the generalized Toda lattice equation,” Nucl. Phys. B , vol. 733, pp. 334–355, 2006

    2006

  38. [46]

    The Weyl group and asymptotics: All supergravity billiards have a closed form general integral,

    P. G. Fr´ e and A. S. Sorin, “The Weyl group and asymptotics: All supergravity billiards have a closed form general integral,” Nucl. Phys. B , vol. 815, pp. 430–494, 2009

    2009

  39. [47]

    Training Stochastic Model Recognition Algorithms as Networks can Lead to Maximum Mutual Infor- mation Estimation of Parameters,

    J. Bridle, “Training Stochastic Model Recognition Algorithms as Networks can Lead to Maximum Mutual Infor- mation Estimation of Parameters,” in Adv. Neural Inf. Process. Syst. , vol. 2, Morgan-Kaufmann, 1989

    1989

  40. [48]

    Bishop, Pattern Recognition and Machine Learning

    C. Bishop, Pattern Recognition and Machine Learning . Springer, 2006

    2006

  41. [49]

    K. P. Murphy, Machine Learning: A Probabilistic Perspective . MIT Press, 2012

    2012

  42. [50]

    Gareth, D

    J. Gareth, D. Witten, T. Hastie, and R. Tibshirani, An Introduction to Statistical Learning . Springer, 2021

    2021

  43. [51]

    L. B. Koralov and Y. G. Sinai, Theory of Probability and Random Processes . Springer, 2007

    2007

  44. [52]

    Contact Geometry, Measurement and Thermodynamics,

    V. V. Lychagin, “Contact Geometry, Measurement and Thermodynamics,” in Nonlinear PDEs, their geometry and applications. Proceedings of Wisla 18 Summer School , pp. 3–52, Springer Nature, 2019

    2019

  45. [53]

    Fano, Metodi Matematici della Meccanica Quantistica

    G. Fano, Metodi Matematici della Meccanica Quantistica . Zanichelli, 1967

    1967

  46. [54]

    Gradient-based learning applied to document recognition,

    Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998

    1998

  47. [55]

    R-R scalars, U-duality and solvable Lie algebras,

    L. Andrianopoli, R. D’Auria, S. Ferrara, P. G. Fr´ e, and M. Trigiante, “R-R scalars, U-duality and solvable Lie algebras,” Nucl. Phys. B , vol. 496, no. 3, pp. 617–629, 1997

    1997

  48. [56]

    Solvable Lie algebras in type IIA, type IIB and M-theories,

    L. Andrianopoli, R. D’Auria, S. Ferrara, P. Fr´ e, R. Minasian, and M. Trigiante, “Solvable Lie algebras in type IIA, type IIB and M-theories,” Nucl. Phys. B , vol. 493, no. 1-2, pp. 249–277, 1997

    1997

  49. [57]

    Trigiante, Dualities in supergravity and solvable Lie algebras

    M. Trigiante, Dualities in supergravity and solvable Lie algebras . PhD thesis, Swansea U., 1998. 59

    1998

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.