Simplicial Message Passing

• Simplicial Message Passing is a paradigm that extends traditional message-passing to simplicial complexes, enabling feature learning over vertices, edges, triangles, and higher-order faces.
• It aggregates multiway relations using boundary, coboundary, and neighborhood interactions, offering increased expressivity and overcoming limitations of standard graph neural networks.
• Empirical studies in molecular property prediction, geometric modeling, and graph isomorphism demonstrate its effectiveness with significant error reductions and state-of-the-art performance.

Simplicial Message Passing (SMP) is a principled mathematical and computational paradigm that extends message-passing neural networks (MPNNs) from graphs to the higher-order categorical structures known as simplicial complexes. By enabling neural feature learning on vertices, edges, triangles, and higher-order faces simultaneously, SMP is foundational for topological deep learning, offering compelling advances in both expressivity and symmetry for modeling data with complex multibody relations, geometric structure, or nontrivial topology. SMP has catalyzed rapid progress in molecular property prediction, geometric deep learning, graph isomorphism testing, and beyond.

1. Mathematical Foundations: Simplicial Complexes and Local Adjacencies

A simplicial complex $\mathcal{K}$ is a family of nonempty subsets of a finite set $V$ (vertices) closed under inclusion. A $k$-simplex $\sigma \in \mathcal{K}$ is a set of $k+1$ vertices, representing combinatorial $k$-dimensional faces (vertices, edges, triangles, tetrahedra, etc.). Simplicial complexes generalize graphs by modeling higher-order relations while maintaining a clear combinatorial structure.

Central to SMP are canonical local relations between simplices:

• The boundary of $\sigma$ ($\partial \sigma$) is the set of all $(k-1)$-faces of $\sigma$.
• The coboundary ($V$0) is the set of all $V$1-simplices containing $V$2.
• Additional adjacencies include lower and upper neighborhoods, often needed for maximal expressive power (Bodnar et al., 2021).

Orientation and the associated signed boundary and coboundary operators are critical, appearing as matrices $V$3 and directly connecting SMP to combinatorial Hodge theory.

2. Simplicial Message Passing: Mechanisms and Variants

A general SMP layer assigns feature vectors $V$4 to each simplex $V$5 at layer $V$6, with updates based on multiway aggregations across the simplicial complex. The canonical message-passing protocol is:

• For each $V$7-simplex and relation type $V$8 (e.g., boundary, coboundary), aggregate transformed features from neighboring simplices:

$V$9

where $k$0 is typically a learnable multilayer perceptron (MLP), and $k$1 encodes invariants such as geometric quantities.

• Update each feature as

$k$2

where the update function can exploit orientation, simplex dimension, and geometric invariants (Kovač et al., 2024, Eijkelboom et al., 2023).

• For oriented complexes, equivariance is achieved by appropriately signing the messages and architectural choices in activation functions (Bodnar et al., 2021).

Several SMP architectures have emerged:

• MPSN (Message Passing Simplicial Networks): Four-channel updates from boundary, coboundary, lower- and upper-adjacency (Bodnar et al., 2021).
• EMPSN/EMPCN: E(n)-equivariant geometric schemes aggregating invariants (e.g., area, volume, dihedral angle) for maximal geometric expressivity (Eijkelboom et al., 2023, Kovač et al., 2024).
• Geometry-aware SMP: Incorporate vertex coordinates directly via geometric color refinements, enabling complete geometric discrimination as formalized by the GSWL framework (Wang et al., 7 May 2026).
• Sheafified SMP: Generalizes Laplacian-based diffusion to "data-aware" sheaf Laplacians, overcoming degeneracies and enabling arbitrary subsimplex-class separation (sheaf cohomology) (Hume et al., 27 Sep 2025).

3. Expressivity: Beyond Graphs and the 1-WL Barrier

SMPs overcome the fundamental pairwise limitations of graph neural networks. Theoretically, the Simplicial Weisfeiler–Lehman (SWL) procedure characterizes the discrimination power of SMPs:

• SWL refines classical 1-WL and is strictly more powerful, capable of separating non-WL-distinguishable pairs such as strongly regular graphs with identical degree distributions but distinct triangle counts.
• SWL is proven not weaker than 3-WL on clique complexes (Bodnar et al., 2021).
• Geometry-aware SMPs, characterized by the GSWL test, can distinguish geometric embeddings invisible to combinatorial schemes (Wang et al., 7 May 2026).
• Sheafified SMPs extend expressivity further: any classification task on $k$3-simplices can be implemented by suitably designed sheaf Laplacians and message-passing schemes (Hume et al., 27 Sep 2025).

From a functional viewpoint, MPSNs are provably more expressive in terms of piecewise-linear region count than either ordinary GNNs or spectral SCNNs (Bodnar et al., 2021).

4. Hierarchical and Geometric Generalizations

Modern SMP schemes exploit full hierarchical and geometric information:

• Hierarchical messaging: Information propagates across simplex dimensions—e.g., triangle features update edge and vertex features—enabling direct modeling of multiway relations and enabling hierarchical cross-order aggregation (Lan et al., 2023, Hajij et al., 2021).
• E(n)-equivariance: Protocols such as EMPCN and Clifford group-equivariant SMPs are constructed to be equivariant with respect to Euclidean isometries (rotations, reflections, translations) (Kovač et al., 2024, Liu et al., 2024).
• Geometric invariants: Areas, volumes, and dihedral angles are used as arguments in message functions, making the scheme attuned to the true geometry of the domain (Kovač et al., 2024, Eijkelboom et al., 2023, Wang et al., 7 May 2026).
• Sheafification: By parametrizing stalks and restriction maps as data-driven objects, sheafified SMPs allow the explicit, local control of inductive bias (e.g., capturing heterophily) and overcome the cohomological collapse present in constant sheaf Laplacians (Hume et al., 27 Sep 2025).

5. Computational Efficiency and Architectures

While a full $k$4-order SMP is potentially combinatorially expensive—the number of $k$5-simplices can scale as $k$6—architectural innovations mitigate these costs:

• Decoupled designs use a fully connected GNN backbone for 0-simplices, only adding top-dimensional simplex-to-vertex interactions, reducing cost to $k$7 (e.g., decoupled EMPCNs) (Kovač et al., 2024).
• Dimension-sharing: Clifford SMPs achieve efficiency by sharing message networks across simplex dimensions with type identifiers (Liu et al., 2024).
• Sparse lifts: Clique and Vietoris–Rips lifts can be restricted by radius or maximal order to control the number of simplices (Eijkelboom et al., 2023, Lan et al., 2023).
• Pseudocode structures for typical SMP layers, including boundary/coboundary list precomputation and batch-processing routines, appear in (Kovač et al., 2024, Hajij et al., 2021).

6. Empirical Validation: Applications and Ablations

SMP schemes have demonstrated superior or state-of-the-art results across several domains:

• Chemical property prediction: SMPNN outperforms MPNN, GCN, D-MPNN on multiple chemical property regression benchmarks (e.g., QM9, OCHEM) by leveraging up to order-2 simplices; ablation confirms 20--30% error reduction when higher-order faces are included (Lan et al., 2023).
• Geometric trajectory and motion capture: EMPSN and Clifford SMPN outperform conventional EGNNs on N-body and human pose prediction by integrating high-dimensional simplex features and geometric equivariance (Liu et al., 2024, Kovač et al., 2024).
• Graph isomorphism and classification: SMP architectures distinguish strongly regular graph pairs undetected by standard GNNs (Bodnar et al., 2021).
• Oversquashing and relational bottlenecks: Relational SMP models demonstrate reduced sensitivity decay and better performance under rewiring mitigations, analogous to classical GNNs (Taha et al., 6 Jun 2025).

The following table summarizes select empirical results:

Domain	SMP Type	Improvement	Reference
Chemical regression	SMPNN vs. MPNN	MAE↓ by 10--30%, SOTA achieved	(Lan et al., 2023)
Geometric prediction	EMPCN vs. EGNN	MSE↓ from 0.0071 to 0.0046	(Kovač et al., 2024)
SRG Isomorphism	MPSN vs. GIN	Near zero error vs. complete failure	(Bodnar et al., 2021)
N-body (5 particles)	Clifford SMPN vs. EMPSN	MSE↓ from 0.007 to 0.002	(Liu et al., 2024)
Deformation classification	Geometry-aware SMP	100% vs. 25% (chance) accuracy	(Wang et al., 7 May 2026)

Ablation studies confirm: the necessity of higher-order simplices for learned tasks, the importance of geometric signals, the superiority of coface-inclusive message-passing over boundary-only variants, and the effectiveness of data-aware sheaf Laplacians in avoiding over-smoothing and cohomological degenerate collapse (Hume et al., 27 Sep 2025, Kovač et al., 2024).

7. Open Challenges and Future Directions

Outstanding issues for simplicial message passing include:

• Scalability: Managing combinatorial explosion at high simplex orders remains an active area, with research into sparse, adaptive, or learnable pruning strategies (Lan et al., 2023).
• Dynamic and heterogeneous architectures: Most current frameworks require all faces up to maximal order; less restrictive or dynamically evolving architectures are an open field (Papillon et al., 2023).
• Expressiveness vs. inductive bias: Characterizing the precise relation between SMP variants and $k$8-WL expressivity (especially in the geometric setting) remains a priority (Wang et al., 7 May 2026, Eijkelboom et al., 2023).
• Sheaf-theoretic generalizations: The expressive capacity of sheafified SMPs and their codification of locality and data-aware coupling is still being charted (Hume et al., 27 Sep 2025).
• Oversquashing: Theoretical understanding and algorithmic mitigation (e.g., curvature-inspired rewiring) demand further exploration in higher-order settings (Taha et al., 6 Jun 2025).
• Unified notation and benchmarks: Recent surveys emphasize the lack of standardized frameworks and suggest the need for a coherent, cross-methodological comparison across topological deep learning (Papillon et al., 2023).

In summary, simplicial message passing provides a mathematically rigorous, combinatorially expressive, and empirically validated extension of graph neural networks, crucial for learning on data with complex higher-order or geometric structure. Ongoing advances in equivariant, geometric, sheaf-theoretic, and computationally efficient formulations continue to broaden its impact and scope in machine learning and related fields.