Directed Sheaf Laplacians

• Directed sheaf Laplacians are operators defined on cellular sheaves over directed graphs, hypergraphs, and simplicial complexes that incorporate local linear constraints via restriction maps.
• They are constructed using cochain complexes and coboundary operators, capturing asymmetric diffusion and intricate network dynamics through tailored block matrix representations.
• These operators support advanced applications such as sheaf neural networks and higher-order diffusion models, providing robust tools for analyzing complex, directional, and heterophilic networks.

A directed sheaf Laplacian is a mathematical operator that generalizes the familiar graph Laplacian to the setting of cellular sheaves over directed graphs, directed hypergraphs, and higher-order structures such as directed simplicial complexes. Unlike standard Laplacians, directed sheaf Laplacians encode not only the adjacency structure and edge directions but also incorporate arbitrary local linear constraints on data via restriction maps (parallel transport) between the vector spaces (“stalks”) attached to nodes, edges, or higher-order cells. This flexible formalism supports nonconstant, asymmetric, and higher-dimensional relations, enabling faithful modeling of directed, signed, or heterophilic networks and complex diffusion dynamics in higher-order topologies (Hansen et al., 2020, Gong et al., 2024, Ribeiro et al., 1 Jul 2025, Mule et al., 6 Oct 2025, Fioresi et al., 24 Feb 2026).

1. Sheaf Structures on Directed Graphs and Hypergraphs

A cellular sheaf on a directed graph $G=(V,E)$ assigns:

• a vector space (the stalk) $F(v)$ to each vertex $v$,
• a vector space $F(e)$ to each edge $e:u\to v$,
• two linear restriction maps $F_{u\to e}:F(u)\to F(e)$ and $F_{v\to e}:F(v)\to F(e)$ for the tail and head.

For directed hypergraphs, the analogous assignment uses complex stalks and phases to encode directionality, with each hyperedge endowed with tail and head sets. For directed simplicial complexes, the setup further extends with unitary parallel transport matrices assigned to inclusions of simplices, supporting higher-order directionality and intricate combinatorics (Mule et al., 6 Oct 2025, Gong et al., 2024).

In all cases, the collection of stalks and restriction maps determines the morphisms of the sheaf, which directly shape the Laplacian and its spectrum.

2. Definition and Construction of Directed Sheaf Laplacians

The core construction proceeds via the cellular (Čech) cochain complex:

• $C^0(G;F) = \bigoplus_{v\in V} F(v)$ (0-cochains: vertex signal assignments),
• $C^1(G;F) = \bigoplus_{e\in E} F(e)$ (1-cochains: edge-signal assignments), with the coboundary operator $\delta: C^0 \to C^1$ defined for $F(v)$0 by

$F(v)$1

encoding how features “transport” along each directed edge.

The directed sheaf Laplacian is then

$F(v)$2

where $F(v)$3 is the adjoint with respect to chosen inner products (Hansen et al., 2020, Fioresi et al., 24 Feb 2026).

For directed graphs, alternative nonsymmetric “out-degree” and “in-degree” variants are defined:

• $F(v)$4 focuses on outgoing message passing,
• $F(v)$5 on incoming, each assembled from stalk-wise expressions involving the restriction maps for outgoing or incoming edges (Ribeiro et al., 1 Jul 2025).

In directed hypergraphs, complex phase factors encode head/tail roles via $F(v)$6, yielding possibly complex Hermitian Laplacians (Mule et al., 6 Oct 2025). For directed simplicial complexes, the “connection Laplacian” incorporates U(1) or U(2) phase matrices, generalizing the magnetic Laplacian and supporting higher-order diffusion (Gong et al., 2024).

3. Algebraic Properties and Spectral Theory

Directed sheaf Laplacians inherit several foundational properties from the undirected sheaf and classical Laplacian settings:

• Symmetry and Positive Semidefiniteness: $F(v)$7 is always symmetric (Hermitian) and positive semidefinite, ensuring real nonnegative spectrum, provided compatible inner products are chosen (Hansen et al., 2020, Fioresi et al., 24 Feb 2026, Mule et al., 6 Oct 2025).
• Kernel Characterization: $F(v)$8 equals the space of global sections—assignments to nodes that are locally consistent across all restrictions. In the case of $F(v)$9 and $v$0, the nullspaces encode “out-consensus” or “in-consensus” relations, generalizing the notion of harmonic functions on graphs (Ribeiro et al., 1 Jul 2025).
• Spectral Interpretation: The spectrum of a directed sheaf Laplacian quantifies the degree of local inconsistency and governs the relaxation rates of sheaf-diffusion dynamics (Hansen et al., 2020, Gong et al., 2024).
• Generalization: Setting all restriction maps to the identity and all stalks to $v$1 (or $v$2) reduces the framework to classical directed and undirected Laplacians, the magnetic Laplacian, or their hypergraph analogues (Mule et al., 6 Oct 2025).

4. Block Structure and Matrix Representations

The Laplacian operator can be represented as a block matrix indexed by nodes (and, in higher-order cases, higher-dimensional cells):

$v$3

$v$4

or, in incidence notation, $v$5. In directed hypergraphs, the presence of complex scalars in the incidence matrix produces Hermitian, potentially non-real operators, essential for encoding non-reversible diffusions and phase winding (frustration) (Mule et al., 6 Oct 2025, Gong et al., 2024).

A summary table of Laplacian types is below:

Context	Laplacian Formula	Key Properties
Directed graphs (sheaf)	$v$6	Symmetric, PSD, real
Out-degree (dir. sheaf)	$v$7 from $v$8	In general, nonsymmetric
Dir. hypergraphs (sheaf)	$v$9	Hermitian, complex
Dir. simplicial complexes	$F(e)$0	Hermitian, complex

5. Encoding Directionality and Frustration

Direction and sign are incorporated via asymmetric restriction maps. For each $F(e)$1, picking $F(e)$2, $F(e)$3 with $F(e)$4 imposes asymmetry. In higher-order or hypergraph settings, complex phases $F(e)$5 distinguish head versus tail participation, critically affecting the spectral and diffusion properties of the operator (Hansen et al., 2020, Mule et al., 6 Oct 2025).

Nontrivial loops of parallel transport that do not multiply to the identity (i.e., do not trivialize) manifest as “frustration” in diffusion, leading to slow relaxations, nontrivial phase winding, and breakdown of Hodge decomposition into gradient and curl flows in higher-order Laplacians (Gong et al., 2024).

6. Applications and Network Learning Architectures

Directed sheaf Laplacians form the analytic backbone of recent message passing and spectral learning frameworks:

• Sheaf Neural Networks: Use the sheaf Laplacian for feature diffusion, outperforming GCNs in asymmetric/signed data (Hansen et al., 2020).
• Cooperative Sheaf Neural Networks (CSNNs): Exploit out/in-degree variants to enable selective, directionally aware diffusion, expanding the receptive field and mitigating oversquashing in heterophilic graphs (Ribeiro et al., 1 Jul 2025).
• Directional Sheaf Hypergraph Networks (DSHN): Apply the directed hypergraph sheaf Laplacian as the diffusion kernel for node signal learning, yielding state-of-the-art results in directed and undirected hypergraph scenarios (Mule et al., 6 Oct 2025).
• Higher-Order Laplacians: Enable the study of multi-body diffusion and topological frustration, relevant for complex systems beyond pairwise interactions (Gong et al., 2024).

7. Mathematical Generalizations and Connections

The sheaf Laplacian admits alternative characterizations:

• Noncommutative Geometry: By equipping a sheaf with discrete parallel transport (flat connection), the resulting Bochner Laplacian $F(e)$6 coincides (up to a sign) with the sheaf Laplacian, linking sheaf-theoretic and noncommutative differential calculi (Fioresi et al., 24 Feb 2026).
• Spectral Theory: The normalized directed sheaf Laplacian is always Hermitian, diagonalizable, and bounded in $F(e)$7; its Dirichlet form gives a precise account of diffusion energy (Mule et al., 6 Oct 2025).
• Reduction to Classical Operators: The directed sheaf Laplacian recovers the classical Laplacian, magnetic Laplacian, and several existing graph/hypergraph Laplacian schemes as special cases depending on the choice of the sheaf and restriction maps (Mule et al., 6 Oct 2025).

The unified perspective of directed sheaf Laplacians thus provides a rigorous and versatile analytic toolset for modern data- and topology-driven network science.