---
title: Persistent Topological Laplacians
url: https://www.emergentmind.com/topics/persistent-topological-laplacians
type: topic
---

# Persistent Topological Laplacians

Searching arXiv for recent and foundational papers on persistent topological Laplacians.
Persistent topological Laplacians, also called persistent Laplacians, are Laplace-type operators defined on filtered topological objects so that both topological information and geometric information are tracked across scales. In the simplicial setting, they extend the combinatorial Hodge Laplacian from a single simplicial complex to a filtration or to a pair of complexes \(K\subseteq L\), with the multiplicity of the zero eigenvalue recovering persistent Betti numbers and the nonzero spectrum encoding additional multiscale geometric information [2508.11560]. More generally, recent work places them in a broad operator-theoretic framework of Hilbert complexes, encompassing discrete combinatorial settings, cellular and cosheaf constructions, and de Rham complexes of smooth manifolds [2509.20220]. Across this literature, persistent topological Laplacians function as spectral refinements of persistent homology: they retain the harmonic information corresponding to persistence while adding non-harmonic spectral data that reflect diffusion, stiffness, connectivity, and homotopic shape evolution [2302.03902].

## 1. Definitions and mathematical setting

A standard starting point is a finite simplicial complex \(K\) on a vertex set \(V\), together with a filtration
\[
\{K(r)\}_{r\ge 0},\qquad K(r)\subseteq K(s)\text{ for }r\le s,
\]
with finitely many strict inclusions. Vietoris–Rips and Čech filtrations from point clouds are standard examples [2506.21352]. On each \(K(r)\), one forms chain groups over \(\mathbb R\),
\[
C_k(K(r)) := \mathbb R^{\{\text{$k$-simplices of }K(r)\}},
\]
and boundary operators
\[
\partial_k : C_k(K(r))\to C_{k-1}(K(r)),
\]
given on an oriented \(k\)-simplex by
\[
\partial_k [v_0\cdots v_k] = \sum_{i=0}^k (-1)^i [v_0\cdots \widehat{v_i}\cdots v_k].
\]
For a single complex, the combinatorial \(k\)-Hodge Laplacian is
\[
L_k = \partial_{k+1}\partial_{k+1}^T + \partial_k^T\partial_k,
\]
with the up and down parts
\[
L_k^{\mathrm{up}}=\partial_{k+1}\partial_{k+1}^T,\qquad
L_k^{\mathrm{down}}=\partial_k^T\partial_k
\]
[2506.21352].

One widely used persistent construction for a filtration pair \(a\le b\) defines
\[
C_{n+1}^{a,b} = \{\sigma \in C_{n+1}^b \mid d_{n+1}^b(\sigma) \in i_n^{a,b}(C_n^a)\},
\]
restricts the boundary by
\[
d_{n+1}^{a,b} := d_{n+1}^b\big|_{C_{n+1}^{a,b}} : C_{n+1}^{a,b}\to C_n^a,
\]
and sets
\[
\Delta_n^{a,b}
= d_{n+1}^{a,b}(d_{n+1}^{a,b})^* + (d_n^a)^*d_n^a
\]
[2508.11560]. In this formulation, \(\ker \Delta_n^{a,b}\cong H_n^{a,b}(K;\mathbb R)\), so the multiplicity of the zero eigenvalue equals the persistent Betti number \(\beta_n^{a,b}\) [2508.11560].

Another influential formulation emphasizes a morphism of chain complexes \(f:V\to W\) and defines a Laplacian associated to the morphism,
\[
\Theta_f := \{x\in W \mid dx \in f(V)\},\qquad
\Delta_f = f^\ast d \iota \iota^\ast d^\ast f + d^\ast d,
\]
where \(\iota:\Theta_f\hookrightarrow W\) is inclusion. For a persistence differential graded inner product space \(\mathcal S:(\mathbb R,\le)\to \mathbf{DGI}\), the \((a,b)\)-persistent Laplacian is
\[
\Delta^{a,b}_\mathcal{S}
= f_{a\to b}^\ast d \iota \iota^\ast d^\ast f_{a\to b} + d^\ast d,
\]
and its kernel is naturally isomorphic to the \((a,b)\)-persistent harmonic space and hence to \((a,b)\)-persistent homology [2302.03902].

A more recent operator-theoretic generalization begins with a Hilbert complex \(P=(C^P_\bullet,d^P_\bullet)\) and defines up-, down-, and full chain Laplacians by quadratic forms. For an inclusion \(P\subset Q\), it introduces a persistent chain space
\[
C_{k+1}^{P,Q} := \overline{\{\,c\in \dom(d_{k+1}^Q) : d_{k+1}^Q(c) \in \iota_k(C^P_k)\,\}},
\]
a persistent differential
\[
d_{k+1}^{P,Q}: C_{k+1}^{P,Q} \to C_k^P,
\]
and persistent Laplacians
\[
\Delta_{+,k}^{P,Q} := d_{k+1}^{P,Q}(d_{k+1}^{P,Q})^*,\qquad
\Delta_{-,k}^{P,Q} := (d_k^P)^*d_k^P,
\]
together with a full persistent Laplacian induced by the quadratic form
\[
\delta_k^{P,Q}[u] = \|d_k^P u\|^2 + \|(d_{k+1}^{P,Q})^*u\|^2
\]
[2509.20220]. This framework covers finite simplicial complexes, cellular or cosheaf chain complexes, and de Rham complexes of manifolds [2509.20220].

## 2. Hodge-theoretic meaning and relation to persistence

Persistent topological Laplacians are designed so that their harmonic sector recovers persistence. In the simplicial pair construction, the persistent Hodge decomposition takes the form
\[
C_n^a
=\mathrm{im}(d_n^a)^* \;\oplus\; \ker \Delta_n^{a,b} \;\oplus\; \mathrm{im}(d_{n+1}^{a,b}),
\]
with
\[
\ker \Delta_n^{a,b}\cong H_n^{a,b}(K;\mathbb R)
\]
[2508.11560]. In the categorical formulation, there is a natural isomorphism between the harmonic-space functor and the homology functor, and the persistent Hodge decomposition identifies the kernel of \(\Delta_\mathcal S^{a,b}\) with persistent homology [2302.03902]. In the operator-theoretic framework, a generalized Hodge theorem gives
\[
\ker(\Delta_k^P)\cong \overline{H_k}(P)
\]
for a Hilbert complex, and under a closed-range hypothesis one obtains
\[
\ker(\Delta_k^{P,Q}) \cong \im(\iota_*:\overline{H_k}(P)\to\overline{H_k}(Q))
\]
for persistent Laplacians [2509.20220].

This spectral viewpoint refines classical persistent homology. Persistent homology tracks the birth and death of homology classes across a filtration and is summarized by barcodes or persistence diagrams [2508.11560]. Persistent topological Laplacians recover that information through the multiplicity of zero eigenvalues, but their positive eigenvalues and eigenvectors encode additional geometric and combinatorial structure, including diffusion scales, higher-order connectivity, stiffness, and multiscale geometric information beyond persistent homology [2508.11560]. The literature repeatedly describes this distinction as the difference between purely topological persistence and a topological–spectral description that also captures homotopic shape evolution [1912.04135, 2301.10865].

A recurring misconception is that persistent Laplacians merely repackage barcodes. The theory does not support that reduction. The harmonic spectra reproduce persistent Betti numbers, but the non-harmonic spectra contain information not present in barcodes, including distinctions between geometrically different but homotopy-equivalent structures and quantitative sensitivity to local structural change [1912.04135, 2301.10865]. Another misconception is that the “full” persistent Laplacian is always the most natural object. Recent generalized theory shows that the classical full persistent Laplacian can fail to satisfy monotonicity and stability, whereas the up- and down-persistent Laplacians satisfy these properties individually, and their nonzero spectra fully determine the nonzero spectrum of the full Laplacian [2509.20220].

## 3. Main constructions and variants

The literature now contains a family of persistent topological Laplacians tailored to different data modalities.

The foundational simplicial constructions include persistent combinatorial Laplacians and persistent spectral graph formulations on Vietoris–Rips or related filtrations of point clouds [1912.04135, 2012.02808]. In these models, a filtration produces a family of simplicial complexes and associated Laplacian matrices whose harmonic spectra recover persistent homology and whose non-harmonic spectra are used for data analysis, modeling, and prediction [1912.04135].

Persistent sheaf Laplacians replace ordinary chain or cochain complexes by sheaf cochain complexes, allowing local labels or physical quantities to be fused into the operator through stalks and restriction maps [2112.10906]. In the point-cloud constructions described there, the spectra of persistent sheaf Laplacians encode both geometrical and non-geometrical information, and the theory is presented as a way to fuse different types of data [2112.10906]. Persistent interaction Laplacians further localize the construction to specified interacting subsystems by replacing ordinary chains with interaction chains built from tuples of simplices with non-empty intersection [2404.11799].

Directed and asymmetric data motivate additional variants. Persistent path Laplacians are built from path homology on digraphs, with harmonic spectra recovering persistent path homology and non-harmonic spectra revealing the homotopic shape evolution of data during filtration [2206.03991]. Persistent directed flag Laplacians extend the framework to directed flag complexes, where simplices correspond to directed cliques, yielding a persistent directed flag Laplacian suited to directed networks and asymmetric data [2312.02099]. The algebraic-stability framework of Laplacian trees also applies to real-valued functions on digraphs via path homology [2302.03902].

For data on manifolds or volumetric domains, persistent de Rham–Hodge Laplacians, or persistent Hodge Laplacians, are defined in an Eulerian representation using structure-preserving Cartesian grids [2408.00220]. This approach avoids the numerical inconsistency caused by remeshing the involving manifolds in the Lagrangian representation and yields discrete Laplacians whose kernels approximate topological invariants while nonzero spectra track geometric evolution [2408.00220].

A concise summary of major variants is useful:

| Variant | Underlying object | Distinguishing feature |
|---|---|---|
| Persistent Laplacian / persistent combinatorial Laplacian | Filtered simplicial complexes | Kernel recovers persistent homology; nonzero spectrum captures multiscale geometry [2012.02808] |
| Persistent sheaf Laplacian | Cellular sheaves on filtered complexes | Encodes geometrical and non-geometrical information [2112.10906] |
| Persistent path Laplacian | Path complexes of digraphs | Harmonic spectra recover persistent path homology [2206.03991] |
| Persistent directed flag Laplacian | Directed flag complexes | Distinct analysis of directed higher-order structure [2312.02099] |
| Persistent interaction Laplacian | Interaction chain complexes | Emphasizes specified interacting elements or subsystems [2404.11799] |
| Persistent de Rham–Hodge Laplacian | Filtrations of manifolds | Eulerian, grid-based manifold topological learning [2408.00220] |

This range of constructions suggests that “persistent topological Laplacian” is best understood as an umbrella term for persistent Laplacian-type operators on filtered chain or cochain complexes, rather than a single matrix formula. A plausible implication is that the common structure is Hodge-theoretic rather than simplicial: persistence is encoded by restricting differentials across scales, and harmonic spaces recover persistent (co)homology.

## 4. Spectral properties, stability, and perturbation theory

A central line of work studies the robustness of persistent Laplacian spectra. Algebraic stability for persistent Laplacians was established by organizing all Laplacians persisting from a parameter into “Laplacian trees” and proving that the interleaving distance of persistence Laplacian trees equals the interleaving distance of the underlying persistence differential graded inner product spaces:
\[
d_I(\mathcal{L}_\mathcal{S},\mathcal{L}_\mathcal{T}) = d_I(\mathcal{S},\mathcal{T})
\]
[2302.03902]. For non-decreasing real-valued functions \(f,g:K\to\mathbb R\) on a fixed simplicial complex, this yields
\[
d_I(\mathcal L^f,\mathcal L^g)\le \|f-g\|_\infty
\]
[2302.03902]. Similar bounds are given for weighted digraphs [2302.03902].

A different and more local notion of robustness appears in the recent one-simplex perturbation theorem for up-persistent Laplacians. If \(K' = K\cup\{\sigma\}\), where \(\sigma\) is a single \(k\)-simplex, then for every radius \(r\ge 0\) and every index \(j\),
\[
\bigl|\lambda_j^{K'}(r)-\lambda_j^{K}(r)\bigr| \le 2\,\|\partial\sigma\|_{2}.
\]
This bound is uniform in the filtration scale and independent of the size of the complex [2506.21352]. The paper describes the result as the first eigenvalue-level robustness guarantee for spectral topological data analysis and notes that the constant \(2\) is optimal for rank-one Hermitian perturbations [2506.21352]. The proof exploits a rank-one decomposition
\[
\Delta_{k}^{K'}(r) =
\begin{bmatrix}
\Delta_k^K(r) & 0\\
0 & 0
\end{bmatrix}
+ uu^*
\]
with \(u=\partial\sigma\), and combines matrix interlacing with Weyl’s inequality [2506.21352].

Generalized persistent Laplacian theory sharpens the discussion of spectral behavior. In that framework, the up- and down-persistent Laplacians satisfy monotonicity and stability individually, while the full persistent Laplacian may fail both properties [2509.20220]. The up- and down-persistent spectral counting functions are 1-Lipschitz with respect to the filtration interleaving distance:
\[
d_I(\lambda_{+,k,q}^P,\lambda_{+,k,q}^Q) \le d_I(P,Q),\qquad
d_I(\lambda_{-,k,q}^P,\lambda_{-,k,q}^Q) \le d_I(P,Q)
\]
[2509.20220]. At the same time, the nonzero spectrum of the full Laplacian is the multiset union of the nonzero spectra of the up and down parts, both in finite-dimensional and infinite-dimensional settings [2509.20220]. This suggests that analyzing up and down components separately is not merely technically convenient but spectrally complete.

The generalized theory also clarifies a controversy sometimes implicit in earlier work: whether one should privilege the full persistent Laplacian because it resembles the ordinary Hodge Laplacian. The newer results indicate that full persistent Laplacians can fail monotonicity and stability, whereas up- and down-persistent Laplacians are individually preferable and sufficient for spectral analysis [2509.20220]. This does not invalidate earlier full-Laplacian constructions; it changes the theoretical emphasis.

## 5. Algorithms, software, and computational practice

Computing persistent topological Laplacians is dominated by constructing the persistent up-part. PETLS, a C++ library with Python bindings titled “PErsistent Topological Laplacian Software,” implements existing and new algorithms for these computations and interfaces with simplicial, alpha, directed flag, Dowker, and cellular Sheaf complexes [2508.11560]. Its default pipeline stores a global boundary matrix \(B_n\) for the final filtration level, realizes \(B_n^a\) as submatrices, computes down-Laplacians as
\[
\Delta_{n,\mathrm{down}}^{a} = (B_n^a)^T B_n^a,
\]
and constructs \(\Delta_{n,\mathrm{up}}^{a,b}\) using a chosen algorithm, with the Schur complement method as the default [2508.11560].

The Schur complement method begins with the full up-Laplacian at scale \(b\),
\[
\Delta_{n,\mathrm{up}}^b = B_{n+1}^b(B_{n+1}^b)^T,
\]
block-partitions it as
\[
\Delta_{n,\mathrm{up}}^b =
\begin{pmatrix}
A & B\\
C & D
\end{pmatrix},
\]
and then computes the persistent up-Laplacian by the Schur complement
\[
\Delta_{n,\mathrm{up}}^{a,b} = A - BD^{-1}C
\]
[2508.11560]. Other reviewed approaches include a gauge or projection method, a reduction plus explicit restriction algorithm, and a specialized algorithm for non-branching complexes [2508.11560].

PETLS also introduces several implementation ideas motivated by spectral computation. In the top dimension, it can use a “flipped” Laplacian principle: since the nonzero eigenvalues of \(AB\) and \(BA\) coincide, PETLS can compute the nonzero spectrum of a large top-dimensional down-Laplacian by diagonalizing a smaller up-Laplacian, which can give \(>2\times\) speedups on the top-dimensional spectral computation [2508.11560]. It also supports reduction to the orthogonal complement of the kernel, using a basis of persistent homology representatives to compress \(\Delta_n^{a,b}\) to a smaller positive definite block [2508.11560].

The software literature emphasizes that PTL eigenvalue problems are atypical: they often have high multiplicities of zero, very sparse but structured matrices, and problem sizes where both dense and sparse solvers are plausible [2508.11560]. PETLS therefore provides multiple backends, including Eigen’s `SelfAdjointEigenSolver`, the Spectra library, and SciPy/LAPACK via Python bindings [2508.11560]. This suggests that algorithm selection is not secondary. A plausible implication is that the practical behavior of persistent Laplacians depends as much on spectral linear algebra choices as on the topological construction.

## 6. Applications, feature engineering, and current directions

Persistent topological Laplacians have been used in biology, physics, and machine learning [2506.21352]. In protein and molecular applications, persistent Laplacian features have been used for SARS-CoV-2 variant analysis, protein engineering, protein–ligand binding, and virus-like particle stoichiometry [2301.10865, 2306.04658, 2507.21417]. In one SARS-CoV-2 study, persistent topological Laplacians were used to examine mutation-induced structural changes of receptor-binding domain complexes and to analyze binding-induced structural changes across variants, with the non-harmonic spectra detecting structural differences that persistent Betti numbers alone could not [2301.10865]. In protein engineering, persistent Laplacians are described as making structure-based embeddings a superb option in AI-assisted directed evolution and AI-assisted protein engineering, alongside persistent path, sheaf, hypergraph, and hyperdigraph Laplacians [2306.04658].

A large-scale recent application to virus-like particles constructs Vietoris–Rips and Alpha filtrations on element-specific atom sets from VLP asymmetric units, computes 0D persistent Laplacians and 1D/2D harmonic information, and assembles a 1044-dimensional feature vector from harmonic and non-harmonic summaries [2507.21417]. That work reports performance on VLP200 and VLP706 datasets and attributes the predictive signal to the joint use of harmonic and non-harmonic spectra [2507.21417].

Machine-learning pipelines based on full spectra face a “varying length” problem and high dimensionality. A recent alternative distills persistent Laplacians into three mathematically grounded invariants: Betti numbers, the spectral gap, and analytic torsion [2606.16990]. For a persistent Laplacian \(\Delta_q^{X,Y}\) with persistent Betti number \(m_q\), the spectral gap is
\[
\gamma_q^{X,Y} = \lambda_{q,m_q+1}^{X,Y},
\]
and the torsion-style summary uses
\[
\log\det_0(\Delta_q^{X,Y}) = \sum_{\lambda>0}\log\lambda,\qquad
\log T(X,Y)=\sum_{q\ge 0}(-1)^{q+1}q\,\log\det_0(\Delta_q^{X,Y})
\]
[2606.16990]. On MNIST, QM-3D, and SKEMPI WT, this reduced feature space is reported to capture the essential predictive signal of the full spectrum, while reducing training time and memory [2606.16990]. This suggests that persistent Laplacian feature engineering may be moving toward principled low-dimensional spectral summaries rather than raw eigenvalue lists.

The broader software and theory landscape indicates several open directions. Weighted complexes, multiple simplex insertions, down-persistent versions in full detail, map-induced perturbations, and structure-aware norms are all identified as natural extensions of current perturbation theory [2506.21352]. PETLS points toward integrated algorithms that combine persistent homology computation with PTL kernel reduction, scalable eigensolvers tailored to large nullspaces, and support for higher-dimensional sheaf stalks [2508.11560]. Generalized persistent Laplacian theory suggests a shift in emphasis from full Laplacians to up- and down-components as the robust and sufficient spectral invariants [2509.20220].

Taken together, the literature portrays persistent topological Laplacians as a family of Hodge-theoretic operators for filtrations, with kernels that recover persistence and positive spectra that encode multiscale geometry. Their development now spans simplicial complexes, digraphs, sheaves, interaction complexes, manifolds, and generalized Hilbert complexes; their computational realization is supported by dedicated software; and their current theoretical frontier concerns stability, perturbation bounds, and principled spectral summaries for learning tasks [2302.03902, 2506.21352, 2508.11560, 2509.20220].

Source: https://www.emergentmind.com/topics/persistent-topological-laplacians