Persistent Laplacians in Topological Data Analysis

Updated 23 February 2026

Persistent Laplacians are spectral operators that generalize persistent homology by encoding both topological invariants and geometric features through eigenvalues.
They decompose into up and down components, allowing for exact recovery of Betti numbers and stable eigenvalue computations via techniques like the Schur complement.
Their versatile framework applies to various structures including simplicial complexes, hypergraphs, and sheaf models, with significant implications in biomolecular analysis and data science.

A persistent Laplacian is a spectral, operator-theoretic generalization of persistent homology that encodes both the topological invariants of a filtration and supplementary geometric, combinatorial, or metric-sensitive features through its eigenvalues. The theory originated from foundational constructions on simplicial complexes and has been extended to a broad spectrum of algebraic-topological structures including hypergraphs, directed flag complexes, manifold discretizations, and cellular sheaves. Persistent Laplacians rigorously refine the summary power of persistent homology, retaining barcodes as their harmonic spectra while enabling the stable extraction of additional geometric information from the evolution of the full spectrum across a filtration.

1. Mathematical Foundations

1.1 Classical and Persistent Laplacian

Given a filtered chain complex of real vector spaces $\{C_*(t)\}$ with boundary operators $\partial_k(t): C_k(t) \to C_{k-1}(t)$ , the $k$ -th combinatorial (Hodge) Laplacian is

$\Delta_k(t) = \partial_{k+1}(t) \partial_{k+1}(t)^T + \partial_k(t)^T \partial_k(t)$

where a standard inner product (simplices orthonormal) is assumed. $\ker \Delta_k(t) \cong H_k(C_*(t))$ recovers the homology, and the zero-eigenvalue multiplicity yields the Betti number.

Persistent Laplacians incorporate scale by considering pairs, sequences, or families of complexes with inclusions. For $a \leq b$ in the filtration parameter, the $(a,b)$ -persistent Laplacian on $k$ -chains is defined by

$\Delta_k^{a,b} = \partial_{k+1}^{a,b} (\partial_{k+1}^{a,b})^* + \partial_k(a)^* \partial_k(a)$

where $\partial_{k+1}^{a,b}$ restricts $(k+1)$ -chains in $C_{k+1}(b)$ to those whose boundaries land in $C_k(a)$ (Jones et al., 15 Aug 2025, Mémoli et al., 2020). This framework generalizes to

path complexes
directed flag complexes
hypergraphs/hyperdigraphs
(cellular) sheaves
de Rham complexes on manifolds (Wei et al., 2023, Su et al., 2024, Wolf et al., 24 Sep 2025) Each setting involves appropriate chain/cochain groups and boundary/coboundary operators, and the persistent Laplacian always decomposes into up and down pieces.

1.2 Harmonic Spaces and Persistent Betti Numbers

A general principle, the Persistent Hodge Theorem, asserts that the dimension of $\ker(\Delta_k^{a,b})$ exactly recovers the $k$ -th persistent Betti number, i.e., the image of $H_k(a) \to H_k(b)$ (Mémoli et al., 2020, Jung et al., 5 Dec 2025, Hayes et al., 23 Oct 2025). Zero-eigenvalues of the Laplacian correspond to topological invariants (persistent features), whereas nonzero spectra capture finer geometric information.

2. Persistent Laplacian Variants and Generalizations

Persistent Laplacians appear in several distinct algebraic-topological contexts, summarized below:

Framework	Chain Complexes	Persistent Boundary/Restriction
Simplicial	$C_q(X_i), C_{q+1}(X_j)$	$\partial_{q+1}$ restricted to relevant domain
Path complex	$(\Omega_q(X_i), \Omega_{q+1}(X_j))$	Path boundary in allowed subspace
Flag complex (digraphs)	Cliques as chains	Ordered boundary respecting orientation
Hypergraph/Hyperdigraph	Infimum chain groups	Restricted by directed or undirected structure
Cellular sheaf	$\bigoplus_\sigma \mathscr{F}(\sigma)$	Coboundary with sheaf restriction maps
Manifolds (de Rham/Hodge)	$\Omega^k(M_l), \Omega^k(M_{l+p})$	Harmonic extension/restriction between subdomains
Interaction topology	Tensor products of subcomplex chains	Differential induced via Leibniz rule

All instances utilize the central formula: $\Delta = d_{k+1}^{a,b} (d_{k+1}^{a,b})^* + (d_k^a)^* d_k^a$ with problem-specific structure on $d$ and the ambient chain spaces (Wei et al., 2023, Liu et al., 2024, Chen et al., 2023).

3. Principal Theoretical Results

3.1 Stability and Monotonicity

Persistent Laplacians satisfy strong algebraic stability properties (Liu et al., 2023, Anh et al., 26 Jun 2025, Wei et al., 2023, Wolf et al., 24 Sep 2025). Specifically:

Kernel stability: The barcode (persistent Betti numbers) is $d_I$ -stable with respect to sup-norm perturbations of the filtration (Liu et al., 2024).
Eigenvalue stability: The spectrum evolves continuously; under $\delta$ -changes in filtration, all eigenvalues of $\Delta_k^{a,b}$ change by at most $O(\delta)$ (Weyl’s inequality applies due to self-adjointness) (Liu et al., 2024, Anh et al., 26 Jun 2025).
Monotonicity: Up-persistent spectra are monotone non-decreasing with respect to the inclusion, and down-persistent spectra are monotone non-increasing, under filtration advancement (Mémoli et al., 2020, Wolf et al., 24 Sep 2025).
Lipschitz eigenvalue bound: The smallest perturbation due to a one-simplex insertion is uniformly controlled (by twice the norm of the boundary vector) for the up-persistent Laplacian eigenvalues (Anh et al., 26 Jun 2025).

3.2 Schur Complement and Efficient Computation

A Schur-complement interpretation underpins both the algorithmics and structural understanding. The up-persistent Laplacian is the Schur complement of the full Laplacian matrix onto the subset of chains present at an earlier filtration step (Mémoli et al., 2020, Gülen et al., 2023). This leads to efficient algorithms for both Betti numbers and spectral computation, and is also foundational for extensions to simplicial maps (Gülen et al., 2023).

4. Algorithmic and Computational Frameworks

Computing persistent Laplacians and their spectra requires:

Sparse assembly of boundary and Laplacian matrices at filtration steps, exploiting filtration-aware incremental updates and basis ordering (Jones et al., 15 Aug 2025).
Schur complement and column reduction for assembling up-persistent pieces (Jones et al., 15 Aug 2025, Mémoli et al., 2020).
Numerical eigendecomposition: typically via sparse Lanczos or Arnoldi routines for a small subset of eigenpairs, as eigenvalue computation costs dominate runtime for large complexes (Jones et al., 15 Aug 2025).
Warm-start strategies and parallelization, especially for filtration sweeps or parameterized complexes.
Open-source software implementations such as PETLS, supporting a unified interface for simplicial, flag, Dowker, and sheaf complexes, and specialized algorithms for each structure (Jones et al., 15 Aug 2025).

5. Applications and Extended Frameworks

Persistent Laplacians have been generalized to, and demonstrated utility in, several domains:

5.1 Interaction Topology

Persistent interaction Laplacians are constructed on n-interaction complexes (subcomplex covers), enabling the study of element-specific or subsystem-specific multiscale topology in heterogeneous data, such as molecular systems. The induced spectra and barcodes are sensitive to localized interactions missed by global topology (Liu et al., 2024).

5.2 Hyperdigraphs, Directed Flags

Persistent Laplacians have been developed for hyperdigraphs and directed flag complexes, incorporating asymmetric and higher-arity relational structure. The spectra quantify cycles and higher-order motifs unique to the directed or hypergraph context (Chen et al., 2023, Jones et al., 2023).

5.3 Manifold-Valued Data (de Rham-Hodge)

For volumetric or manifold-valued data, persistent Laplacians are realized as discretized de Rham–Hodge Laplacians on Eulerian grids, maintaining well-conditioned numerical discretizations across scales and enabling spectral-persistence analysis of sublevel-set filtrations (Su et al., 2024).

5.4 Sheaf-Theoretic and Multi-Parameter Settings

Persistent sheaf Laplacians capture multiscale topology in data with fused geometric and functional labels, with applications including protein flexibility prediction and robust image analysis (Hayes et al., 23 Oct 2025, Wang et al., 16 Feb 2026, Wei et al., 2021). The framework extends naturally to multi-dimensional filtrations and cosheaf settings (Wei et al., 2021, Wang et al., 16 Feb 2026).

6. Comparative Features, Strengths, and Limitations

Feature	Persistent Laplacians	Persistent Homology
Invariant recovered	Barcode (zero spectrum)	Barcode (explicit)
Additional features	Multiscale metric/geometric spectral	None (topological only)
Local/global control	Local with interaction and element specificity, e.g., in interaction Laplacians	Only global (whole-complex)
Spectral monotonicity	Up/down-parts monotone, full spectrum may fail monotonicity (Wolf et al., 24 Sep 2025)	Barcodes monotone
Algorithmic cost	Matrix assembly + eigensolver dominated	Boundary matrix reduction
Limitations	Complexity increases with interaction order and for general sheaf/cosheaf settings;	No geometric sensitivity

Benefits: Element-specific analysis (interaction Laplacians), geometric sensitivity of nonzero spectra, stable vectorizations (e.g., persistent Laplacian diagrams/images), computational feasibility via reduced matrix size for interaction models (Liu et al., 2024, Jung et al., 5 Dec 2025).
Drawbacks: Potential to lose global topology (in interaction settings), combinatorial explosion for high interaction order, and need for sophisticated numerical handling of nontrivial domains (e.g., sheaf/cosheaf structures, manifold discretizations).

7. Empirical Results and Data-Driven Insights

In biomolecular science, persistent Laplacians, and specifically the smallest nonzero eigenvalues (“spectral gaps”), correlate with physical properties such as isomer enthalpy differences and protein–ligand binding affinities (Liu et al., 2024, Su et al., 2024).
Spectral features from persistent Laplacians enhance machine learning models for protein mutational effect prediction and outperform persistence-barcode-only features (Wee et al., 2023, Wei et al., 2023).
Spectral summaries, such as persistent Laplacian diagrams and images, extend stable vectorization to spectral descriptors, permitting finer discrimination among combinatorially similar but geometrically distinct datasets (Jung et al., 5 Dec 2025).

Persistent Laplacians constitute a robust, extensible, and computationally accessible framework that marries classical persistent homology with spectral invariants reflecting topology, geometry, and metric structure. Their variants encompass a wide array of algebraic-topological models, and their stability and monotonicity properties, together with algorithmic innovations and empirical successes, establish them as foundational operators for modern multiscale data analysis (Jones et al., 15 Aug 2025, Liu et al., 2024, Wei et al., 2023, Wolf et al., 24 Sep 2025, Su et al., 2024).