Cartan Networks: Group theoretical Hyperbolic Deep Learning

Federico Milanesio, Matteo Santoro, Pietro Giuseppe Frè, Guido Sanguinetti · 2025

Read Cartan Networks: Group theoretical Hyperbolic Deep Learning by Federico Milanesio, Matteo Santoro, Pietro Giuseppe Frè, Guido Sanguinetti (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-24353.txt.

Introduction

Federico Milanesio and co-authors build a new class of deep learning nets that work in curved space, in a paper from 2025. Curved space suits tree like data, since it holds vast room near its edge. Past nets used the shape side of that space. The team adds its group side, since the space can also be read as a smooth group with a distance rule.

Their nets swap group maps that keep class with warps that keep distance, in turns. That blend forms a fresh net family with clean math roots. Scores on standard test sets look strong for a first build. The authors see the work as the seed of a wider class of curved space nets for ranked and nested data.

Catalog record

CREATORFederico Milanesio, Matteo Santoro, Pietro Giuseppe Frè, Guido Sanguinetti
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-24353

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Cartan Networks: Group theoretical Hyperbolic Deep Learning Federico Milanesio, Matteo Santoro, Pietro Giuseppe Frè, Guido Sanguinetti Abstract Hyperbolic deep learning leverages the metric properties of hyperbolic spaces to develop efficient and informative embeddings of hierarchical data. Here, we focus on the solvable group structure of hyperbolic spaces, which follows naturally from their construction as symmetric spaces. This dual nature of Lie group and Riemannian manifold allows us to propose a new class of hyperbolic deep learning algorithms where group homomorphisms are interleaved with metric-preserving diffeomorphisms. The resulting algorithms, which we call Cartan networks , show promising results on various benchmark data sets and open the way to a novel class of hyperbolic deep learning architectures. 1 Introduction The concept of distance is the core of machine learning and pattern recognition. While much classical machine learning can be recast as learning distances directly from data (e.g. [ 1 ] ), recent developments have pointed out that common data structures, such as trees and graphs, cannot easily be accommodated within Euclidean spaces, requiring a more radical rethink of the geometry of data spaces. In this context, the n -dimensional hyperbolic space \mathbb{H}^{n} has received significant attention as a suitable space in which to embed hierarchically structured data [ 2 ] , spurring a productive line of research combining hyperbolic geometry with various deep learning architectures [ 3 , 4 , 5 , 6 , 7 , 8 , 9 ] . These so-called hyperbolic neural networks have found applications in fields as diverse as neuroscience [ 10 ] , single-cell transcriptomics [ 11 ] , and recommender systems [ 12 ] . Geometrically \mathbb{H}^{n} is a n -dimensional hyperboloid, namely the quadric locus \sum_{i=1}^{n}X_{i}^{2}-X_{n+1}^{2}=-1 in \mathbb{R}^{n+1} . It is also a coset manifold , namely the quotient of a Lie Group modulus a maximal Lie subgroup, \mathbb{H}^{n}\simeq\mathrm{SO}(1,n)/\mathrm{SO}(n) , and more specifically a symmetric space . The study and classification of symmetric spaces is one of the monumental achievements of the French mathematician Èlie Cartan [ 13 , 14 , 15 , 16 ] . The non-compact symmetric spaces are all metrically equivalent to a corresponding solvable Lie group S of the same dimension, a mathematical result that was discovered and developed in the context of Supergravity Theory [ 17 , 18 , 19 , 20 , 21 , 22 ] , and amply reviewed and systematically reorganized for machine learning applications in [ 23 ] . This result, to our knowledge not known so far in the machine learning literature, has significant algorithmic consequences, which we explore in this work and will develop in a more general theoretical paper in preparation [ 24 ] . The dual nature of group and Riemannian manifold of the hyperbolic space \mathbb{H}^{n} enables us to construct a deep learning framework based entirely on intrinsic geometric operations, where group homomorphisms are interleaved with metric-preserving diffeomorphisms in creating a powerful function approximation machine. Importantly, the nonlinearities naturally arising from group-theoretic exponential and logarithmic maps give flexibility to the framework, which achieves promising results on benchmark data sets when compared with similar-sized standard deep learning architectures. The main contributions of this work are as follows: • We highlight the metric equivalence of the hyperbolic space with a solvable Lie group to exploit the group structure as a tool in architecture design. • We propose a new deep learning architecture where each layer is a solvable Lie group S_{i} and where the map from layer i to layer i+1 can be represented as a combination of homomorphisms from the solvable Lie group S_{i} to the next one S_{i+1} and the isometries of S_{i+1} . The construction is general for any symmetric space, and we implement it for the hyperbolic space \mathbb{H}^{n} . • We extensively benchmark these architectures on real and synthetic datasets, showing competitive or better performance w.r.t. Euclidean and standard hyperbolic neural networks. 1.1 Previous literature Early works in hyperbolic deep learning focused on hyperbolic embeddings for hierarchical data. [ 2 ] introduced Poincaré embeddings , showing superior hierarchical representation compared to Euclidean embeddings. Ganea et al. ( [ 3 ] ) and subsequent works [ 4 , 5 , 6 , 9 ] , extended hyperbolic geometry to deep learning by developing hyperbolic neural networks , using Möbius operations [ 25 ] . Various generalizations of hyperbolic networks have been explored. Convolutional networks [ 26 , 27 ] , graph neural networks [ 7 ] , and attention mechanisms [ 8 ] hyperbolic variants were introduced to handle different datasets. Lie groups and Lie algebras are often studied in deep learning for their equivariance properties [ 28 , 29 , 30 ] . Architectures based on semisimple Lie algebras have been introduced under the name Lie Neurons [ 31 ] , focusing on making these networks adjoint-equivariant. The notion that \mathbb{H}^{n} is isometric to a Lie group was explored in the context of probability distributions and Frechét means [ 32 ] . However, the isometry between symmetric spaces and solvable groups was not highlighted in full generality, and the knowledge was never applied to the study of deep learning architectures. 2 Theoretical preliminaries We will assume basic knowledge of Lie groups (see App. A for a brief introduction). Solvable groups and Cartan subalgebras. Definition 2.1 (Subalgebra commutator) . Let \mathfrak{h_{1},h_{2}} be two subalgebras of \mathfrak{g} . Their commutator subalgebra is [\mathfrak{h}_{1},\mathfrak{h_{2}}]:=\{[h_{1},h_{2}]\in\mathfrak{g}\>|\>h_{1}\in\mathfrak{h_{1}},\>h_{2}\in\mathfrak{h}_{2}\} (1) where [\>\cdot\>,\>\cdot\>] denotes the Lie bracket of the algebra. Definition 2.2 (Derived series) . Let \mathfrak{g} be a Lie algebra. Its derived series is the series \mathfrak{g}^{(0)}=\mathfrak{g},\quad\quad\mathfrak{g}^{(n+1)}=[\mathfrak{g}^{(n)},\mathfrak{g}^{(n)}]\quad\forall n\in\mathbb{N}_{+} (2) The derived series is a decreasing sequence of ideals in the algebra. Definition 2.3 (Solvable algebras) . A Lie algebra \mathfrak{g} is solvable if its derived series is eventually 0, that is to say, if \exists n\in\mathbb{N}\quad\mathrm{s.t.}\quad\mathfrak{g}^{(n)}=0 A Lie group is solvable if its Lie Algebra is solvable. In practice, solvable groups are best understood in terms of their matrix representation. In fact, Theorem 2.1 (Lie’s theorem [ 33 ] ) . Let \mathfrak{g} be a solvable subalgebra of the general linear group \mathfrak{gl}_{V} . Then there exists a basis of V with respect to which \mathfrak{g} is made of upper triangular matrices. This theorem shows we can think of solvable groups as upper-triangular matrix Lie groups. Definition 2.4 (Cartan subalgebras) . Let \mathfrak{g}\subseteq\mathfrak{gl}_{n}(\mathbb{R}) be a matrix Lie algebra consisting of upper triangular matrices. Its Cartan subalgebra is the subspace of diagonal matrices. Symmetric spaces. Let G be a Lie group and H a normal subgroup, \mathfrak{g} and \mathfrak{h} the corresponding Lie algebras. A coset manifold \mathrm{G/H} is a symmetric space if and only if there is an orthogonal decomposition of \mathfrak{g} , as a vector space, as follows: \mathfrak{g}\,=\,\mathfrak{h}\oplus\mathfrak{m}\quad;\quad\left\{\begin{array}[]{lcr}\left[\mathfrak{h}\,,\,\mathfrak{h}\right]&\subset&\mathfrak{h}\\ \left[\mathfrak{h}\,,\,\mathfrak{m}\right]&\subset&\mathfrak{m}\\ \left[\mathfrak{m}\,,\,\mathfrak{m}\right]&\subset&\mathfrak{h}\\ \end{array}\right. (3) One interesting class of non-compact symmetric spaces is given by \mathcal{M}^{[r,r+p]}=\dfrac{\text{SO}(r,r+p)}{\text{SO}(r)\times\text{SO}(r+p)},\>\>r>0,\,p\geq 0 (4) This family of manifolds is easily tractable thanks to the metric equivalence between these and an appropriate solvable Lie group , studied in the context of theoretical physics in [ 23 ] , \mathcal{M}^{[r,p]}\simeq\mathrm{Exp}\left[{\mathrm{Solv}_{[r,p]}}\right] where we denote \mathrm{Solv}_{[r,p]} the solvable Lie algebra of the solvable Lie subgroup S_{[r,p]}\subset\mathrm{SO}(r,r+p) with r Cartan generators. For r=1 we realize the hyperbolic space \mathbb{H}^{p-1}\simeq\mathcal{M}^{[1,\,1+p]} (where \simeq denotes a metric equivalence). Solvable coordinates of

Read the complete text