Persistent Sheaf Laplacian (PSL)

• Persistent Sheaf Laplacian (PSL) is an operator-valued framework in TDA that fuses cellular sheaves with multiscale spectral analysis to capture topological and physical features in complex systems.
• It tracks both harmonic and non-harmonic spectral components over filtrations, revealing persistent sheaf cohomology and geometric insights applicable in biomolecular and network analysis.
• PSLs employ sparse matrix algorithms and efficient eigenanalysis to integrate heterogeneous, cell-wise data, enhancing predictions in biophysics, omics, and machine learning.

A Persistent Sheaf Laplacian (PSL) is an operator-valued framework in topological data analysis (TDA) that fuses the functorial, local-to-global encoding of cellular sheaves with the multiscale spectral tools of persistent Laplacians. PSLs generalize classical persistent Laplacians by encoding not only topological relationships but also heterogeneous, cell-wise data—such as physical, chemical, or functional labels—directly into the Laplacian spectrum. By tracking the evolution of both harmonic (kernel) and non-harmonic (positive spectrum) features across a filtration of simplicial complexes decorated with sheaves, PSLs provide a unifying, information-rich, and robust characterization of complex systems, ranging from biomolecular structures to networks and omics data (Hayes et al., 12 Feb 2025, Hayes et al., 23 Oct 2025, Wei et al., 2021, Wei et al., 2023, Cottrell et al., 29 Sep 2025, Ren et al., 18 Jan 2026).

1. Algebraic and Topological Structure

A PSL operates on a cellular sheaf $\mathscr{F}$ over a finite simplicial complex $K$. The sheaf assigns to each simplex $\sigma \subset K$ a finite-dimensional real (or possibly more general) vector space $\mathscr{F}(\sigma)$, called the stalk, and to each face inclusion $\sigma \leq \tau$ a restriction map $$\mathscr{F}_{\sigma \leq \tau}: \mathscr{F}(\sigma) \to \mathscr{F}(\tau)$$ with functorial composition, i.e., $$\mathscr{F}_{\rho \leq \tau} = \mathscr{F}_{\sigma \leq \tau} \circ \mathscr{F}_{\rho \leq \sigma}$$ whenever $\rho \leq \sigma \leq \tau$. Each stalk is equipped with an inner product, uniquely extending to the direct sum over each cochain degree.

The cochain complex associated to $\mathscr{F}$ consists of vector spaces $$C^p(K; \mathscr{F}) = \bigoplus_{\sigma \in K_p}\mathscr{F}(\sigma)$$ and coboundary operators $d^p$ defined using combinatorial incidences weighted by the restriction maps. Explicitly, on each stalk summand,

$$(d^p\omega)_\tau = \sum_{\sigma \leq \tau} [\sigma:\tau]\ \mathscr{F}_{\sigma \leq \tau}(\omega_\sigma),$$

where $[\sigma:\tau]$ is the signed incidence number. The sheaf Laplacian in degree $p$ is

$\Delta^p = (d^p)^*d^p + d^{p-1}(d^{p-1})^*,$

which decomposes the cochain space into harmonic (kernel) and non-harmonic (positive spectrum) components (Wei et al., 2021, Hayes et al., 12 Feb 2025).

2. Persistence: Multiscale Spectral Theory

To capture multiscale structure, PSLs are constructed over a filtration $\{K_{t}\}_{t \geq 0}$ of simplicial complexes (e.g., via Vietoris–Rips or alpha complexes). At each scale $t$, the sheaf $\mathscr{F}_t$ is restricted to $K_t$, and the corresponding Laplacian $\Delta^p_{K_t}$. The persistent sheaf Laplacian between scales $s \leq t$ is defined by projecting coboundary operators and adjoints through the inclusion $K_s \hookrightarrow K_t$ and assembling the operator on $C^p(K_s; \mathscr{F}_s)$. Formally,

$$\Delta^p_{s,t} = (d^p_{s,t})^*d^p_{s,t} + d^{p-1}_s(d^{p-1}_s)^*,$$

where $d^p_{s,t}$ is the projection of the coboundary at scale $t$ to cochains at scale $s$ (Wei et al., 2023, Wei et al., 2021).

The zero modes of $\Delta^p_{s,t}$ are isomorphic to the image of the sheaf cohomology $$H^p(K_s; \mathscr{F}_s) \to H^p(K_t; \mathscr{F}_t)$$, yielding persistent sheaf Betti numbers. Positive eigenvalues quantify geometric obstructions and encode “almost persistent” classes, thus extending the classical persistent homology barcode (Wei et al., 2023).

3. Construction of Sheaf Data and Physical Encoding

A defining feature of PSLs is the explicit encoding of heterogeneous, physically or chemically meaningful data at the level of stalks and restriction maps. In molecular applications, stalks can be set as $\mathscr{F}(v_i) = \mathbb{R}$ for each atom or residue $v_i$, and edge restriction maps

$\mathscr{F}_{v_i \leq e_{ij}} = q_j / r_{ij},$

where $q_j$ is a physical parameter (e.g., partial charge) and $r_{ij}$ is the interatomic distance. For higher cells, the construction extends by composition, allowing complex, element-specific, or direction-sensitive informational fusion. This sheaf-encoding enables PSLs to integrate geometric, topological, and high-fidelity physicochemical information, which cannot be fused into combinatorial or persistent Laplacians (Ren et al., 18 Jan 2026, Hayes et al., 12 Feb 2025).

4. Spectral Interpretation and Data Analysis

The PSL spectrum at each filtration step exhibits two principal components:

• Harmonic part (zero eigenvalues): Encodes persistent sheaf cohomology; multiplicities recover persistent Betti numbers.
• Non-harmonic spectrum (positive eigenvalues): Sensitive to both topological and physical inhomogeneities, describing rigidity modes, bending, twisting, or “near obstructions” to persistence.

Tracking the flow of both spectral components across filtration yields multiscale, localized, and physically interpretable features. In applications such as protein flexibility, low non-harmonic eigenvalues indicate soft (“floppy”) modes correlating with high B-factors, while high eigenvalues indicate stiffness (Hayes et al., 12 Feb 2025, Wei et al., 2021). Feature extraction for machine learning typically involves assembling summary statistics (counts of zero modes, minimal/maximal/mean eigenvalues) over multiple scales (Hayes et al., 23 Oct 2025, Ren et al., 18 Jan 2026).

5. Computational Algorithms and Scaling

Construction and diagonalization of PSLs rely on sparse matrix algorithms:

• Matrix Assembly: Coboundary matrices are assembled using combinatorial incidences and restriction maps per stalk and face relation. For molecular data or networks, the cochain dimensions and sparsity are governed by the local neighborhood size and stalk dimension.
• Eigenanalysis: Only a truncated spectrum is required (e.g., first $K=50$ nontrivial modes). Sparse eigensolvers such as ARPACK or Lanczos are standard, with scaling $O(K \cdot \text{nnz})$ where $\text{nnz}$ is the number of nonzero entries—typically linear in system size for fixed neighborhood degree.
• Software: Public implementations (e.g., PETLS) provide modular interfaces for constructing PSLs on arbitrary filtrations and user-supplied sheaf data. PETLS leverages Gudhi SimplexTree data structures and C++/Python wrappers, supporting functionality for domain scientists and data analysts (Jones et al., 15 Aug 2025).

PSLs are empirically 5–10× more computationally intensive than classical combinatorial Laplacians, but exploit identical optimization strategies (Schur complements, null-space reduction) to achieve practical performance (Jones et al., 15 Aug 2025).

6. Applications in Biophysics, Omics, and Machine Learning

PSLs have been deployed in a range of scientific domains:

• Protein Flexibility: PSL-based B-factor prediction achieves a 32% increase in predictive accuracy over classical Gaussian Network Models on a dataset of 364 proteins (PCC: 0.588 vs. 0.444), with further improvement using integrated feature sets and regression methods (Hayes et al., 12 Feb 2025).
• Protein-Nucleic Acid Complexes: PSL yields up to a 21% improvement in Pearson correlation for B-factor prediction compared to GNM and mFRI on RNA-protein complexes, demonstrating robustness to biomolecular heterogeneity (Hayes et al., 23 Oct 2025).
• Mutation Impact: Changes in PSL spectra upon simulated mutation encode local and global stability or solubility differences, enabling automated feature construction for deep learning predictors such as SheafLapNet (Ren et al., 18 Jan 2026).
• Gene Regulatory and PPI Networks: PSLs identify functionally significant genes by their local topological impact, outperforming ordinary persistent homology and integrating gene expression data at the spectral level (Cottrell et al., 29 Sep 2025).
• Graph and Neural Architectures: Integration of PSLs into graph convolutional frameworks augments non-isotropic message passing and localizes topological signals, improving expressivity in heterophilous and geometric graph learning tasks (Cesa et al., 2023).

7. Theoretical Properties, Stability, and Limitations

PSLs inherit several stability and functoriality guarantees:

• Hodge-Type Theorem: The kernel of the PSL operator recovers the persistent sheaf cohomology (and thus, the barcode structure), ensuring that no topological information is lost relative to standard persistent homology (Wei et al., 2023).
• Stability: Both harmonic and non-harmonic spectra are continuous (Lipschitz, in certain metrics) under small perturbations of the filtration or sheaf parameters, ensuring robustness in data-driven analyses (Wei et al., 2021, Cottrell et al., 29 Sep 2025).
• Generality: By varying stalk dimension and restriction maps, PSLs subsume combinatorial Laplacians, standard persistent Laplacians, and enable encoding of arbitrary multivariate data; they also admit construction over generalized complexes (flag, hypergraph, digraph, etc.) (Wei et al., 2023).
• Computational Limitations: Bottlenecks persist for large-scale, high-dimensional filtrations, as spectral solvers scale poorly with increased stalk dimension and simplex number. Interpretation of high-frequency eigenmodes may present challenges in certain complex datasets (Wei et al., 2023).

A plausible implication of the cited data is that PSLs provide a mechanism for simultaneous multiscale, topologically aware, and physically grounded feature extraction, delivering both interpretability and generalizability for machine learning pipelines in structured scientific data (Ren et al., 18 Jan 2026, Hayes et al., 23 Oct 2025, Hayes et al., 12 Feb 2025).