Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.03169

Cup and Cap Topological Neural Network

Marta Niedostatek·Ferran Hernandez Caralt·Runyue Wang·Federica Baccini·Lorenzo Giambagli·Pietro Lió·Ginestra Bianconi
View original source ↗Read source document ↗Export metadata ↓

Abstract

Topological Deep Learning (TDL) is designed to learn features associated with higher-dimensional simplices including not only nodes but also edges, triangles and so on. However, most TDL approaches, being based on boundary operators and Hodge Laplacians, have the limitation that features and signals cannot be lifted or lowered across more than one dimension per layer. To overcome this limitation, in this work, we propose the adoption of the cup and cap products. Specifically, we formulate the Cup and Cap Topological Neural Network (CCNN), a topological deep learning architecture designed to learn node-based variables (or 0-cochains) by taking into account their many-body interactions (e.g. triangles) present in the data. We validate CCNN by assessing its performance on the TopoBench datasets, revealing its competitiveness with respect to other simplicial complex neural networks.

Explore related subjects

cond-mat.dis-nn

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Marta Niedostatek, Ferran Hernandez Caralt, Runyue Wang, Federica Baccini, Lorenzo Giambagli, Pietro Lió, Ginestra Bianconi. 2026-10-02. Cup and Cap Topological Neural Network. https://arxiv.org/abs/2610.03169

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Disorder-Tailored Delocalization

We derive disorder fields tailored by the details of a choice of a delocalized wave function. We first investigate the unidirectional Hatano-Nelson chain and its localization properties under $M$-base diagonal disorder with variable weights. The spectrum forms loops in the complex plane and the loop parameter is a good quantum number similar to a momentum. All eigenstates are subexponentially `localized', i.e. the logarithm of the absolute value of the wave function performs a random walk in space, and are characterized by a corresponding length scale $ξ_{sel}$ as shown in 1998 by Silvestrov for the general Hatano-Nelson chain. For $M=2$ real-valued binary disorder with equal weights the model was solved in Zeitschrift für Naturforschung A 81 421, yielding Cassini oval spectral loops and a diverging subexponential localization length for two eigenstates and for disorder weaker than a critical value set by the hopping strength. When the disorder field for any $M$ and arbitrary weights is confined to circles in the complex plane with radius equal to the hopping strength, the circle center will belong to the spectrum and to one of the spectral loops, and host a plane-wave-like eigenstate with diverging $ξ_{sel}$. We generalize to tailoring on-site disorder for a given eigenstate at a given energy, for any lattice dimension, and any hopping field (both ordered and disordered), for Hermitian and non-Hermitian systems. We exemplify by constructing a one-dimensional chain with Anderson-localized eigenstates hosting a completely delocalized one with random phases. The localization length diverges as $1/|E|^{2/3}$ upon approaching the delocalized state. Our method can be used for the systematic construction of disorder fields which host predefined eigenstates with arbitrary properties.

cond-mat.dis-nn↗
arXiv2026

Graphicality of power-law and double power-law degree sequences

The graphicality problem -- whether or not a sequence of integers can be used to create a simple graph -- is a key question in network theory and combinatorics, with many important practical applications. In this work, we study the graphicality of degree sequences distributed as a power-law with a size-dependent cutoff and as a double power-law with a size-dependent crossover. We combine the application of exact sufficient conditions for graphicality with heuristic conditions for nongraphicality which allow us to elucidate the physical reasons why some sequences are not graphical. For single power-laws we recover the known phase-diagram, we highlight the subtle interplay of distinct mechanisms violating graphicality and we explain why the infinite-size limit behavior is in some cases very far from being observed for finite sequences. For double power-laws we derive the graphicality of infinite sequences for all possible values of the degree exponents $γ_1$ and $γ_2$, uncovering a rich phase-diagram and pointing out the existence of six qualitatively distinct ways graphicality can be violated. The validity of theoretical arguments is supported by extensive numerical analysis.

cond-mat.dis-nn↗
arXiv2026

Hyperbolic lattices with mass disorder: Phases and phase transitions

Nearest-neighbor (NN) tight-binding models (TBMs) on plaquette-centered $\{ 10,3\}$ (Schläfli symbol), $\{ 8,3\}$, and $\{ 8,4\}$ hyperbolic lattices on a Poincaré disk with open boundary conditions display a vanishing, a finite, and a diverging density of states (DOS) near zero energy (band center), respectively, yielding a Dirac liquid, a Fermi liquid, and a flat band. Emergent bipartite nature of these lattices within the framework of NN-TBMs, allows us to scrutinize the impact of mass disorder on the electronic states and DOS therein. From extensive numerical calculations of the average and typical DOS using the kernel polynomial method, we show that hyperbolic Dirac liquid remains stable against weak mass disorder, undergoing a semimetal-to-metal quantum phase transition at moderate disorder, followed by an Anderson metal-to-insulator transition at even stronger disorder. The remaining two systems display only the latter transition. Critical exponents near all these transitions are found to be close to the ones mediated by on-site potential disorder.

cond-mat.dis-nn↗