Persistent Sheaf Laplacians

• Persistent Sheaf Laplacians are novel operators that integrate sheaf theory with multiscale Laplacians to refine persistent homology.
• They capture both topological invariants via zero modes and geometric features via the nonzero spectrum, enhancing data analysis.
• Their framework supports data fusion in bioinformatics, image analysis, and network science through scalable spectral algorithms.

A persistent sheaf Laplacian is an operator-theoretic refinement of persistent homology, integrating sheaf theory with multiscale Laplacians to encode both the topology and heterogeneous data of a filtered complex. Unlike standard persistent Laplacians, persistent sheaf Laplacians (PSLs) enrich the spectrum to capture not only topological invariants (via kernels/zero modes) but also multiscale geometric, physical, or chemical features (via nonzero spectrum), with applications in bioinformatics, image analysis, and network science. Formally, PSLs generalize classical Hodge Laplacians to the setting of cellular sheaves and their filtrations, providing a flexible algebraic-topological framework for data analysis (Wei et al., 2021, Wei et al., 2023, Hayes et al., 12 Feb 2025, Ren et al., 18 Jan 2026, Jones et al., 15 Aug 2025).

1. Mathematical Structure and Core Definitions

A cellular sheaf on a finite (simplicial or CW) complex $X$ consists of a family of stalks $\mathcal{S}(\sigma)$ (finite-dimensional vector spaces over $\mathbb{R}$) assigned to each simplex (cell) $\sigma\in X$, together with linear restriction maps $$\mathcal{S}_{\sigma\leq\tau}:\mathcal{S}(\sigma)\to\mathcal{S}(\tau)$$ for all face inclusions, satisfying functoriality: for any $\rho\leq\sigma\leq\tau$,

$$\mathcal{S}_{\rho\leq\tau} = \mathcal{S}_{\sigma\leq\tau} \circ \mathcal{S}_{\rho\leq\sigma},\qquad \mathcal{S}_{\sigma\leq\sigma} = \mathrm{id}.$$

The $q$-th sheaf cochain group is

$$C^q_{\mathcal{S}}(X) = \bigoplus_{\substack{\sigma\in X \ \dim\sigma=q}} \mathcal{S}(\sigma)$$

with the coboundary map

$$d^q|_{\mathcal{S}(\sigma)} = \sum_{\substack{\sigma\leq\tau \ \dim\tau=q+1}} [\sigma:\tau]\, \mathcal{S}_{\sigma\leq\tau},$$

where $\mathcal{S}(\sigma)$0 are orientation- or incidence-induced signs.

With standard inner products on stalks and the induced Euclidean structure on cochain groups, the adjoint $\mathcal{S}(\sigma)$1 is defined. The (cellular) sheaf Laplacian is given by

$\mathcal{S}(\sigma)$2

whose kernel recovers the $\mathcal{S}(\sigma)$3-th sheaf cohomology $\mathcal{S}(\sigma)$4 (Wei et al., 2021, Jones et al., 15 Aug 2025, Hayes et al., 12 Feb 2025, Wei et al., 2023).

2. Persistent Sheaf Laplacian Construction

Given a filtered complex $\mathcal{S}(\sigma)$5 and a cellular sheaf $\mathcal{S}(\sigma)$6 on $\mathcal{S}(\sigma)$7 (with restrictions $\mathcal{S}(\sigma)$8), cochain inclusions $\mathcal{S}(\sigma)$9 are induced for $\mathbb{R}$0. The persistent coboundary $\mathbb{R}$1 is obtained by restriction and projection. The $\mathbb{R}$2-th persistent sheaf Laplacian is then

$\mathbb{R}$3

which encodes the evolution of cohomology classes over $\mathbb{R}$4 and tracks sheaf-consistent structures surviving from $\mathbb{R}$5 to $\mathbb{R}$6. The dimension of the kernel $\mathbb{R}$7 is the $\mathbb{R}$8-th persistent sheaf Betti number, i.e., the rank of persistent cohomology (Ren et al., 18 Jan 2026, Wei et al., 2021, Jones et al., 15 Aug 2025, Wei et al., 2023).

3. Spectral Interpretation: Harmonic and Non-Harmonic Modes

The spectrum of $\mathbb{R}$9 or $\sigma\in X$0 naturally decomposes:

• Harmonic spectrum: Zero eigenvalues correspond to (persistent) sheaf cohomology classes surviving between intervals; the count recovers persistent Betti numbers and persistent barcodes in $\sigma\in X$1-th degree.
• Non-harmonic spectrum: Strictly positive eigenvalues quantify geometric and homotopic features—such as local stiffness, energetic barriers, or coherent deformation modes—that are invisible to persistent homology. Small nonzero eigenvalues indicate globally consistent features, while large values correspond to highly localized fluctuations or structural changes (Wei et al., 2023, Hayes et al., 23 Oct 2025, Hayes et al., 12 Feb 2025, Ren et al., 18 Jan 2026).

This spectral enrichment enables PSLs to distinguish homotopy-type, geometric phase transitions, and data-driven transitions in heterogeneous datasets.

4. Data Fusion and Encoding Heterogeneous Information

A major distinction of PSLs is the ability to directly incorporate heterogeneous, cell- or simplex-wise labels into the sheaf structure. For instance, in molecular systems:

• Stalks are $\sigma\in X$2 (or $\sigma\in X$3) and restriction maps on edges encode weighted functions: e.g., for vertices $\sigma\in X$4, use $\sigma\in X$5 for partial charge $\sigma\in X$6 and distance $\sigma\in X$7.
• More generally, restrictions can be

$\sigma\in X$8

for a global weight function $\sigma\in X$9, allowing explicit encoding of chemical, physical, or multi-modal signal data (Ren et al., 18 Jan 2026, Wei et al., 2023, Wei et al., 2021).

This mechanism enables PSLs to natively model structure-function relationships, parameter fields, or context-sensitive constraints across scales.

5. Algorithmic Realization and Computational Aspects

The computation of persistent sheaf Laplacians follows this general pipeline (Ren et al., 18 Jan 2026, Jones et al., 15 Aug 2025, Wei et al., 2021):

1. Input: Point cloud or network data, potentially with heterogeneous, cellwise labels.
2. Filtration: Construct a nested sequence of complexes (e.g., Vietoris–Rips, Alpha, or clique complexes) indexed by a filtration parameter (geometric, statistical, or purely algebraic).
3. Sheaf assignment: Define stalks and restriction maps (potentially vector-valued; possibly using data-driven kernels).
4. Coboundary assembly: Build block-sparse coboundary matrices for each filtration step.
5. Persistent (cross-scale) Laplacian construction: For each pair $$\mathcal{S}_{\sigma\leq\tau}:\mathcal{S}(\sigma)\to\mathcal{S}(\tau)$$0, project and conjugate to compute $$\mathcal{S}_{\sigma\leq\tau}:\mathcal{S}(\sigma)\to\mathcal{S}(\tau)$$1.
6. Spectral analysis: Use sparse or block eigensolvers to extract Betti numbers and summarize nonzero spectra (e.g., spectral gap, moment statistics, trace heat kernels).
7. Feature export: Barcodes and spectral summaries can feed into machine learning or statistical pipelines, enabling interpretable multiscale data analysis.

High-performance implementations such as PETLS provide C++ and Python interfaces, leveraging tools such as GUDHI and ARPACK and supporting efficient handling of high-dimensional, sparse complexes (Jones et al., 15 Aug 2025).

6. Applications in Data Analysis, Computational Biology, and Machine Learning

Persistent sheaf Laplacians enable novel data representations and learning approaches across diverse domains:

Application Area	Role of PSL	Reference
Protein structure/function	Physicochemical-aware spectral features for stability, flexibility, and solubility prediction (partial charge integration, B-factor regression)	(Ren et al., 18 Jan 2026, Hayes et al., 12 Feb 2025, Hayes et al., 23 Oct 2025)
Image analysis	Multiscale, multidimensional spectral encoding for image classification, robust under varying feature dimension	(Wang et al., 16 Feb 2026)
Single-cell transcriptomics	Local relational, multiscale spectral features for robust cell representation	(Wang et al., 27 Mar 2026)
Graph deep learning	Differentiable, persistent local homology sheaf Laplacians enable topologically expressive, data-fusion-aware neural architectures	(Cesa et al., 2023)

Beyond these, PSLs have been employed in sensor fusion, network anomaly detection, and as general unifying frameworks for heterogeneous data integration (Wei et al., 2021, Wei et al., 2023).

7. Theoretical Properties and Future Directions

Persistent sheaf Laplacians satisfy several foundational results (Wei et al., 2023, Wei et al., 2021):

• Persistent Hodge theorem: $$\mathcal{S}_{\sigma\leq\tau}:\mathcal{S}(\sigma)\to\mathcal{S}(\tau)$$2 recovers im$$\mathcal{S}_{\sigma\leq\tau}:\mathcal{S}(\sigma)\to\mathcal{S}(\tau)$$3, aligning PSL zero-modes with persistent cohomology.
• Algebraic stability: If restriction maps change by at most $$\mathcal{S}_{\sigma\leq\tau}:\mathcal{S}(\sigma)\to\mathcal{S}(\tau)$$4 in operator norm, every spectrum moves by at most $$\mathcal{S}_{\sigma\leq\tau}:\mathcal{S}(\sigma)\to\mathcal{S}(\tau)$$5—providing perturbation-robustness analogous to classical persistent homology stability.
• Bridging discrete and continuous: Sheaf-theoretic PSLs parallel manifold Hodge theory with twisted coefficients and connect to de Rham–Hodge settings.

Open directions include stability analysis for higher-dimensional and nonlinear sheaves, scalable algorithms for large sensor networks and biomolecular ensembles, advances in persistent sheaf Dirac operators, and new machine-learning architectures that natively exploit sheaf-valued representations and their spectra (Wei et al., 2021, Wei et al., 2023, Jones et al., 15 Aug 2025).

---

Persistent sheaf Laplacians represent a major advance in topological and spectral data analysis, simultaneously generalizing persistent homology and Laplacians, and providing rich, interpretable, multiscale invariants for the fusion of topology with heterogeneous data (Wei et al., 2021, Ren et al., 18 Jan 2026, Jones et al., 15 Aug 2025).