Cup and Cap Topological Neural Network

arXiv:2610.03169 · cond-mat.dis-nn · Submitted 2026-10-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "Cup and Cap Topological Neural Network".

Mira: Here's the gist of the paper. 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 The Cup and Cap Topological Neural Network (CCNN), a topological deep learning architecture designed to learn nodebased variables (or 0-cochains) by taking into account their many-body interactions (e.g.

Kai: triangles) present in the data, overcomes the limitation of existing TDL approaches that cannot lift or lower features across more than one dimension per layer The core problem addressed Most Topological Deep Learning (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 This restriction prevents these architectures from lifting node features up to triangles The main contribution of this work is the proposal of the Cup and Cap Topological Neural Network (CCNN), which uses cup and cap products instead of exclusively boundary operators to allow for lifting and lowering features across an arbitrary number of dimensions per layer Mathematical framework for topological operations The paper defines simplicial complexes using simplices as basis elements, where a q-cochain is expressed as a linear combination of these simplices The boundary operator maps q-chains to (q - 1)-chains, and the coboundary operator maps a generic q-cochain into a (q + 1)-cochain The Hodge Laplacians are defined as Lq = L up q + L down q, where L up q = Bq+1B⊤q+1 and L down q = B⊤q Bq Mechanism of signal transformation The cup product acts as a lifting operator, allowing the lift of 0-cochains to higher-dimensional cochains, such as a 2-cochain from a 0-cochain using an element like ψˆ = ˆξ ⌣ B⊤1 θˆ The cap product functions as a lowering operator, transforming cochains into lower-dimensional chains, exemplified by constructing a 0-chain η from a 1-cochain and an adjacency tensor Architectural design of CCNN The CCNN architecture consists of L layers, where the recursive equation for the cochains θˆl∈C0 and their associated chains θl in the forward pass is given by a complex system involving learnable matrices Wlq, Wˆ lq, Dl, Dˆ lq, and D˜ lq Two variants are proposed: CCNN-η and CCNN-ω Experimental validation and results The CCNN is validated on the TopoBench datasets, revealing its competitiveness with respect to other simplicial complex neural networks The experiments show that CCNN-η and CCNN-ω outperform other simplicial complex architectures on 8 out of 20 tested datasets This provides evidence of the advantages of using the cup and cap products to leverage the topological structure of the data The CCNN demonstrates that TDL architectures based on cup and cap products can overcome the limitation imposed by boundary operators, enabling node-based predictions by exploiting higher-order interactions like triangles The work showcases the potential of using the cup and cap products for Topological Deep Learning Author Contributions G.B.

Mira: designed the project, M.N.

Kai: and F.H.C.

Title and authors: Mira: Now that we understand the architecture, let's look at what the paper actually summarizes about this Cup and Cap Topological Neural Network. It boils down to this: they are proposing a network that uses cup and cap products instead of relying solely on boundary operators to learn features associated with higher-dimensional simplices.

Kai: So, instead of just mapping nodes to something simpler, like edges or triangles in a single step, they are using these products to allow the network to lift node signals up to these higher dimensions across multiple layers.

Lev: Can you tell me more about that lifting process? How do the cup and cap products actually function mathematically in this context? It sounds like a sophisticated operation.

Mira: The paper defines how a q-cochain, which is just a linear function mapping chains to real numbers, can be expressed as a combination of basis elements associated with simplices. This means the cochains are naturally representing the features tied to those geometric shapes.

Kai: They then introduce the boundary operator and coboundary operator, which are standard tools in algebraic topology, but they show that these alone can only move features by one dimension per layer.

Mira: The key contribution is proposing that by using cup and cap products instead of *exclusively* the boundary operators, you get this mechanism to lift or lower features across an arbitrary number of dimensions per layer.

Lev: So if I'm visualizing this, it’m like they’re giving the network a multi-dimensional ladder instead of just a single step up and down for every topological feature.

Kai: That’s a good way to think about it. The cup product acts as the lifting operator, allowing you to take that zero-cochain—the node feature—and generate something higher dimensional, like a two-cochain from it.

Mira: And the cap product does the opposite, acting as a lowering operator, which lets you transform those higher-dimensional cochains back down into lower dimensions using things like adjacency tensors.

Lev: So the paper is showing that by combining these two products in this specific way within their architecture, they can achieve a much more flexible signal transformation capability than traditional TDL models.

Kai: It’s about incorporating the full topological structure into the prediction of node features because those higher-order interactions are directly encoded in those cochains.

Mira: And that means the network is not just looking at neighbors; it's looking at triangles, and potentially even larger structures, to make its decision at each layer.

Lev: That makes sense from a system modeling standpoint. If you’re modeling something like a physical material or a social network, those higher-order correlations are often what drive the dynamics, so having the network explicitly model them is valuable.

Kai: So in short, the paper summarizes that CCNN overcomes the one-dimension limitation by using cup and cap products to enable arbitrary dimensional lifting and lowering of features.

The paper's summary: Lev: Okay, moving on to what they say are the actual improvements this architecture offers over what came before. It sounds like they’re focused on the capabilities that were missing in previous TDL models.

Mira: The main improvement is definitely overcoming that restriction where you could only lift node features up to triangles with traditional boundary operator methods. CCNN allows for lifting them across any number of dimensions per layer, which is a significant structural gain.

Kai: And they also emphasize the ability to model higher-order interactions more effectively than older models because the cup and cap products naturally allow the network to incorporate those features into its learning process.

Mira: They point out that by using these products, you get a natural framework for lifting and lowering cochains, which means you can incorporate the full topological structure of the simplicial complex directly into how node features are predicted.

Lev: From an engineering standpoint, this flexibility is key because it means we aren't limited to a very shallow or very deep network structure just to try and capture more topological information; we can choose the depth that suits the task.

Kai: And they also mention their architectural design, specifically the recursive equations for cochains involving learnable matrices like W q, W hat q, D q, and others. This shows how they translate these abstract algebraic concepts into a concrete neural network structure.

Mira: They also propose two specific variants, CCNN-eta and CCNN-omega. The ablation study on this architecture suggests that choosing the combination of these terms allows for the greatest flexibility because the network can learn which of those specific terms are actually relevant for its particular task and adjust its weights accordingly.

Lev: That sounds like a powerful way to manage complexity; instead of having a fixed set of operations, you let the learning process prune or emphasize certain topological connections based on what's useful.

Kai: So it’s about giving the network control over which parts of the topological structure it prioritizes without having to manually define those dependencies beforehand.

Mira: It moves us toward a more robust way to handle data where those higher-order correlations are important, like in social networks or physical systems, because these interactions are often what define the behavior at the node level.

Lev: If this works as well as reported on the TopoBench datasets, it suggests that this approach is genuinely competitive for many of the problems we see in complex system analysis.

Kai: It is competitive, and that’s what they show on those twenty tested datasets where CCNN-eta and CCNN-omega outperformed other simplicial complex architectures.

The paper's improvements: Mira: So, wrapping up the Cup and Cap Topological Neural Network paper, it demonstrates that by using cup and cap products successfully, we can overcome the limitation imposed by boundary operators in traditional topological deep learning models.

Kai: It shows that this approach enables node-based predictions by exploiting those higher-order interactions like triangles directly within the network structure.

Lev: So, the big picture here is that CCNN provides a mechanism for lifting and lowering features across an arbitrary number of dimensions per layer, which was a key hurdle in existing architectures.

Mira: It opens up a new way to model complex systems by allowing us to explicitly account for the full topological structure of the simplicial complex into our node predictions.

Kai: The potential is that this method can be used across many domains where understanding multi-dimensional relationships is crucial, from materials science to network analysis.

Lev: I just think it’s a strong piece of work because they managed to marry advanced algebraic topology with deep learning in a way that addresses a known constraint directly.

Mira: It's definitely a direction we should be watching as we look at how these concepts integrate further into more complex AI models.

Kai: We'll keep an eye on the next papers, but for now, this CCNN paper shows that using cup and cap products can indeed unlock new possibilities for topological deep learning.

Conclusion: Mira: So that’s what we’ve seen in the Cup and Cap Topological Neural Network paper. It shows how using cup and cap products instead of just boundary operators lets you lift node features across any number of dimensions per layer, which is a big step for topological deep learning.

Kai: Right, and it means those higher-order interactions, like triangles, are actually being used to predict things at the node level. I’m curious about the actual hardware setup they used to test this fidelity on.

Lev: Well, they validated it on TopoBench datasets and CCNN eta and CCNN omega performed better than other architectures on eight out of twenty tested datasets. That’s a solid experimental validation showing it works beyond just theory.

Mira: It really proves that you can build a framework for lifting and lowering cochains naturally using these products, which means you can incorporate the whole topological structure into your node feature prediction process without making assumptions about fixed layer depths.

Kai: But what does that mean practically? If we have a complex system, like a chemical reaction or a social network, this suggests we can build models that capture those multi-dimensional relationships more effectively than the old methods did.

Lev: From an error correction standpoint, the flexibility they built in means you don’t have to over-engineer your entire architecture to get higher dimensional features; you can learn which specific terms are useful for the task and adjust them.

Mira: Exactly, it gives us a more natural framework for incorporating those high-order interactions that are often driving the dynamics at the node level in those complex systems we study.

Kai: It’s interesting how they handle the math—the cup product as a lifting operator, and cap product as a lowering operator—it’s just very clean conceptually once you get past all that notation.

Lev: Yeah, it's elegant because it addresses that one-dimension limitation head-on by changing the fundamental mathematical operation used for feature transformation.

Mira: And that’s what we have on this one—the Cup and Cap Topological Neural Network—it moves us toward more powerful node-based predictions by exploiting those higher-order interactions like triangles.

Kai: It shows that using cup and cap products can really unlock new possibilities for topological deep learning.

Lev: Next up, we’re looking at how this kind of topological structure might apply to analyzing complex fluid dynamics, specifically looking at activity and competing length scales in those anomalous core-softened fluids.

Queen Mary University of London · University of Cambridge · Sapienza University of Rome · Freie Universät Berlin

cond-mat.dis-nn

Submitted: 2026-10-02

Updated: 2026-11-05

Comments: 9 pages, 1 table + appendices

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Topological Deep Learning (TDL) is designed to learn features associated with higher-dimensional simplices including not only nodes but also edges…

Key concepts

Topological Deep Learning (TDL)
A field that learns features associated with higher-dimensional simplices in data, including nodes, edges, triangles, and so on. It aims to capture the complex topological structure of the data.
Cup Product
An operator used in CCNN that acts as a lifting mechanism. It allows 0-cochains (node features) to be lifted up to higher-dimensional cochains, such as triangles, enabling feature learning across multiple dimensions per layer.
Cap Product
An operator in CCNN that functions as a lowering operator. It transforms higher-dimensional cochains into lower-dimensional chains, which is useful for constructing node features from more complex topological structures.
Simplicial Complex
A mathematical structure used to represent data where simplices are the basis elements. These elements include nodes (0-simplices), edges (1-simplices), triangles (2-simplices), and higher-dimensional shapes, forming the full structure of the data.

Terminology

Summary

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 The Cup and Cap Topological Neural Network (CCNN), a topological deep learning architecture designed to learn nodebased variables (or 0-cochains) by taking into account their many-body interactions (e. p. 0g. triangles) present in the data, overcomes the limitation of existing TDL approaches that cannot lift or lower features across more than one dimension per layer

The core problem addressed

Most Topological Deep Learning (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 This restriction prevents these architectures from lifting node features up to triangles The main contribution of this work is the proposal of the Cup and Cap Topological Neural Network (CCNN), which uses cup and cap products instead of exclusively boundary operators to allow for lifting and lowering features across an arbitrary number of dimensions per layer

Mathematical framework for topological operations

The paper defines simplicial complexes using simplices as basis elements, where a q-cochain is expressed as a linear combination of these simplices The boundary operator maps q-chains to (q - 1)-chains, and the coboundary operator maps a generic q-cochain into a (q + 1)-cochain The Hodge Laplacians are defined as Lq = L up q + L down q, where L up q = Bq+1B⊤q+1 and L down q = B⊤q Bq

Mechanism of signal transformation

The cup product acts as a lifting operator, allowing the lift of 0-cochains to higher-dimensional cochains, such as a 2-cochain from a 0-cochain using an element like ψˆ = ˆξ ⌣ B⊤1 θˆ The cap product functions as a lowering operator, transforming cochains into lower-dimensional chains, exemplified by constructing a 0-chain η from a 1-cochain and an adjacency tensor

Architectural design of CCNN

The CCNN architecture consists of L layers, where the recursive equation for the cochains θˆl∈C0 and their associated chains θl in the forward pass is given by a complex system involving learnable matrices Wlq, Wˆ lq, Dl, Dˆ lq, and D˜ lq Two variants are proposed: CCNN-η and CCNN-ω

Experimental validation and results

The CCNN is validated on the TopoBench datasets, revealing its competitiveness with respect to other simplicial complex neural networks The experiments show that CCNN-η and CCNN-ω outperform other simplicial complex architectures on 8 out of 20 tested datasets This provides evidence of the advantages of using the cup and cap products to leverage the topological structure of the data

Conclusion

The CCNN demonstrates that TDL architectures based on cup and cap products can overcome the limitation imposed by boundary operators, enabling node-based predictions by exploiting higher-order interactions like triangles The work showcases the potential of using the cup and cap products for Topological Deep Learning

Author Contributions

G.B. designed the project, M. p. 9N. and F.H.C. led the methodological development and execution of the project The authors contributed through discussions, feedback, interpretation of results and writing of the manuscript

Acknowledgments

G.B. acknowledges interesting discussions with V. p. 9 Nanda, S. Barbarossa and L. p. 1 Di Nino M.N was funded by the UKRI/BBSRC Collaborative Training Partnership in AI for Drug Discovery and Queen Mary University of London This research utilised Queen Mary’s Apocrita HPC facility, supported by QMUL Research-IT H. p. 9

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Higher order feature lifting capability: The Cup and Cap Neural Network (CCNN) can lift node features across an arbitrary number of dimensions per layer, overcoming a limitation where traditional boundary operator-based models cannot be used to lift node features up to triangles.

  2. Enhanced relational modeling for complex systems: The CCNN learns node features by taking into account the higher-order interactions among the nodes, enabling it to model real-world systems like social networks where the presence of triangles and cliques strongly biases the opinion dynamics occurring at the node level.

  3. Flexible topological signal processing: By using cup and cap products, the CCNN enables a natural framework for lifting and lowering cochains, allowing the system to incorporate the full topological structure of the simplicial complex into the prediction of node features.

  4. Robust architecture selection: The ablation study suggests that choosing the A + B + C term combination allows for the greatest flexibility as the neural network can learn which of the A, B and C terms are relevant for the specific task and adjust weights accordingly.

  5. Competitive performance across domains: The CCNN demonstrates competitiveness with respect to other simplicial complex neural networks and outperforms them on several datasets, establishing it as a genuine option for prediction tasks.

Sources

Related papers