Multiscale Laplacian Graph Kernels

• Multiscale Laplacian graph kernels are a family of graph-dependent kernels that use the Laplacian's spectral properties to encode multiscale structure.
• They employ recursive subgraph extraction and spectral filtering techniques to combine local, intermediate, and global graph features.
• These methods enhance tasks like graph comparison, classification, and semi-supervised learning while balancing model expressiveness with computational scalability.

Searching arXiv for relevant papers on multiscale Laplacian graph kernels and closely related Laplacian-kernel frameworks. Multiscale Laplacian graph kernels are graph-dependent kernels and spectral filters that use graph Laplacians across several spatial, spectral, or hierarchical resolutions. In one prominent formulation, they are graph-level kernels built by recursively applying a Laplacian-based kernel to nested subgraphs; in a broader later usage, they denote kernels or covariance operators obtained from multiscale functions of a graph Laplacian, including heat kernels, resolvents, polynomial spectral filters, warped kernels, and transductive Gaussian process covariances (Kondor et al., 2016). Across these formulations, the central idea is that Laplacian structure encodes local and global organization simultaneously: low-frequency modes reflect coarse community or manifold structure, while higher-frequency modes encode increasingly local variation (Zhi et al., 2022).

1. Conceptual basis and scope

The motivation for multiscale Laplacian graph kernels is the mismatch between purely local and purely global graph similarity notions. Local kernels typically aggregate small substructures or neighborhood patterns, whereas spectral or diffusion-based kernels depend on eigenvalues or eigenvectors of operators such as the adjacency or Laplacian. Real graphs, especially molecular graphs, often exhibit structure at multiple different scales: local labels and bonds, intermediate functional groups, and overall scaffold or shape. The multiscale Laplacian viewpoint addresses this by making kernels sensitive to structure at a range of different scales rather than at a single neighborhood radius or a single diffusion time (Kondor et al., 2016).

Laplacian-based multiscale structure can be described in several equivalent languages. In the spectral language, a graph Laplacian $L$ or $L_{\mathrm{sym}}$ has eigenvectors ordered by graph frequency: lower eigenvalues correspond to smooth, global modes, and higher eigenvalues correspond to oscillatory, local modes. In the vertex-domain language, powers of the Laplacian mix information over progressively longer paths, so a polynomial in $L$ combines $0$-hop, $1$-hop, up to $d$-hop interactions with learnable weights. In a diffusion language, a family such as $e^{-tL}$ or $(I+tL)^{-p}$ provides scale control through $t$, with small $t$ emphasizing local structure and larger $L_{\mathrm{sym}}$0 emphasizing coarse structure (Zhi et al., 2022).

The phrase also extends beyond graph classification. Later work uses multiscale Laplacian constructions for semi-supervised learning, Gaussian processes on graphs, manifold regularization, graph comparison across different graph sizes, and multiscale spectral embeddings that remain consistent across resolutions. This broader usage suggests that “multiscale Laplacian graph kernel” denotes a family of constructions unified less by a single formula than by a common mechanism: graph geometry is encoded through one or more Laplacian operators, and scale is introduced through hierarchy, spectral filtering, diffusion time, polynomial degree, or multiresolution graph refinement (Merkurjev et al., 2021).

2. Feature Space Laplacian Graph kernel

The 2016 formulation introduces the Multiscale Laplacian Graph kernel through a single-scale building block, the Feature Space Laplacian Graph kernel. Let $L_{\mathrm{sym}}$1 be a weighted, undirected graph with adjacency matrix $L_{\mathrm{sym}}$2, degree matrix $L_{\mathrm{sym}}$3, and combinatorial Laplacian

$L_{\mathrm{sym}}$4

To ensure invertibility, the construction uses the regularized Laplacian

$L_{\mathrm{sym}}$5

A symmetric positive semidefinite base kernel $L_{\mathrm{sym}}$6 on vertices induces an RKHS feature map $L_{\mathrm{sym}}$7, or, in the explicit-feature setting, a feature matrix $L_{\mathrm{sym}}$8 with entries $L_{\mathrm{sym}}$9 (Kondor et al., 2016).

The starting point is a Gaussian Markov random field over graph vertices, with precision matrix $L$0 and covariance $L$1. Passing from vertex variables to feature-space variables yields the covariance

$L$2

With feature-space regularization $L$3, define

$L$4

For two graphs $L$5 and $L$6, the Feature Space Laplacian Graph kernel is the Bhattacharyya kernel between the zero-mean Gaussians with covariances $L$7 and $L$8:

$L$9

Because it is a Bhattacharyya kernel, it is positive semidefinite; because it depends on $0$0, it is invariant to vertex relabeling (Kondor et al., 2016).

A key part of the construction is kernelization. If vertex features are implicit, one forms the joint Gram matrix over the vertices of the two graphs using the base kernel $0$1, diagonalizes that Gram matrix, and projects each graph into the common span of vertex RKHS features. Writing the resulting projected matrices as $0$2 and $0$3, one obtains

$0$4

and then applies the same Bhattacharyya formula. This gives a graph kernel induced from an arbitrary positive semidefinite vertex kernel, including one-hot label kernels and other local invariant feature kernels (Kondor et al., 2016).

3. Recursive multiscale construction on nested subgraphs

The multiscale step is recursive. For each vertex $0$5 in a graph $0$6, define a nested sequence of neighborhoods

$0$7

and let $0$8 be the induced subgraph on $0$9. A simple choice is a hierarchy of $1$0-hop balls, with neighborhood size growing roughly exponentially with the level. Using these induced subgraphs, the construction defines Multiscale Laplacian Subgraph kernels recursively (Kondor et al., 2016).

At the first level,

$1$1

At higher levels,

$1$2

The top-level Multiscale Laplacian Graph kernel between full graphs is then

$1$3

Each scale uses the previous scale’s kernel as a new base kernel, so the hierarchy lifts vertex-level information into subgraph-level information and ultimately into graph-level similarity (Kondor et al., 2016).

Naïve recursion is expensive, so the original work introduces a randomized projection procedure similar to the Nyström method, but for RKHS operators. A subsample of landmark vertices is chosen, the corresponding landmark Gram matrix is diagonalized, and each graph is projected into a low-dimensional feature space. If $1$4 is the projected feature matrix for graph $1$5, then the approximating covariance is

$1$6

and the same Bhattacharyya expression is evaluated on these reduced matrices. Combined with caching of subgraph-kernel evaluations, this converts a potentially prohibitive recursive computation into a tractable dataset-level approximation (Kondor et al., 2016).

This recursive definition clarifies one sense of “multiscale.” The multiscale behavior does not arise from varying a single diffusion time alone; it is created by repeatedly replacing vertices with subgraphs, then comparing those subgraphs with a Laplacian-sensitive kernel. The result is simultaneously sensitive to local labels, to the organization of larger neighborhoods, and to global graph geometry (Kondor et al., 2016).

4. Spectral-filter and semi-supervised formulations

A broader later line of work formulates multiscale Laplacian graph kernels directly as functions of the graph Laplacian. In the transductive Gaussian process framework, the graph and node features are treated as two coupled Hilbert spaces with regularizers $1$7 and $1$8, leading to the node covariance

$1$9

Here $d$0 is a feature-space kernel and $d$1 is a graph spectral regularizer. A flexible multiscale choice is

$d$2

which yields a learnable polynomial spectral filter while keeping $d$3 and the covariance nonsingular. When features are absent, this framework recovers classical graph-only kernels such as the regularized Laplacian, diffusion or heat kernel, $d$4-step random walk family, and fractional or Matérn-type kernels on graphs (Zhi et al., 2022).

The multiscale interpretation is explicit in this spectral form. A polynomial

$d$5

corresponds in operator form to

$d$6

Since $d$7 mixes information across paths of length up to $d$8, the resulting kernel aggregates local and increasingly global interactions. Heat kernels and resolvent kernels become special cases: the heat kernel is a pure low-pass filter with a single scale parameter $d$9, while the regularized Laplacian has transfer function $e^{-tL}$0. The polynomial form differs by learning the weights assigned to distinct propagation scales rather than fixing them a priori (Zhi et al., 2022).

Multiscale Laplacian Learning introduces a different multiscale operator,

$e^{-tL}$1

where each $e^{-tL}$2 is built from a scale-dependent affinity matrix

$e^{-tL}$3

These Hermite-polynomial modulated Gaussian affinities produce several graph Laplacians at different scales, which are then aggregated. In the Multikernel Manifold Learning method, this operator enters a warped kernel

$e^{-tL}$4

with $e^{-tL}$5 in Algorithm 1. In the Multiscale MBO method, the same operator appears as an implicit diffusion filter through the resolvent

$e^{-tL}$6

which is the backward-Euler spectral filter associated with $e^{-tL}$7 (Merkurjev et al., 2021).

These formulations share a common principle: multiscale Laplacian kernels are not restricted to a single kernel family. They may arise as recursively lifted Bhattacharyya kernels, as inverses of summed regularizers, as warped kernels in manifold regularization, or as resolvent-based diffusion operators. What makes them multiscale is the explicit control of several graph scales through hierarchy, polynomial powers, Hermite-Gaussian filtrations, or spectral transfer functions (Zhi et al., 2022).

5. Graph comparison, embeddings, and cross-graph correspondence

For graph comparison across different graph sizes, one line of work uses Laplacian eigenvectors as multiscale embeddings. The Embedded Laplacian Discrepancy represents each graph by the first $e^{-tL}$8 nontrivial Laplacian eigenvectors and, for each eigenvector index $e^{-tL}$9, forms the sign-symmetrized empirical measure

$(I+tL)^{-p}$0

The discrepancy between graphs is then

$(I+tL)^{-p}$1

This construction removes eigenvector sign ambiguity, is a pseudo-metric under simple spectra, and is invariant under graph isomorphism. For non-simple spectra, the method perturbs the adjacency matrix with a small random symmetric hollow perturbation to split multiplicities, and stability is justified by Weyl’s inequality and Davis–Kahan perturbation theory. An RBF kernel built from $(I+tL)^{-p}$2 is widely usable but not universally positive semidefinite, whereas the axis-wise kernel mean embedding construction

$(I+tL)^{-p}$3

is positive semidefinite by construction (Tam et al., 2022).

Heat Kernel Coupling addresses a different comparison problem: constructing multimodal spectral geometry on weighted graphs of different size without vertex-wise bijective correspondence. Given corresponding probe functions $(I+tL)^{-p}$4 and $(I+tL)^{-p}$5, weak coupling requires

$(I+tL)^{-p}$6

The paper optimizes modified Laplacians so that these projected heat operators agree across several times while remaining close to the original Laplacians. In the special case of equal vertex and edge sets with $(I+tL)^{-p}$7, the limit $(I+tL)^{-p}$8 yields Laplacian averaging,

$(I+tL)^{-p}$9

showing that Laplacian averaging is a limit case of heat-kernel coupling rather than a separate principle (Bronstein et al., 2013).

A related multiscale issue arises when the same dataset is represented by graphs at several resolutions. Laplacian eigenvector cascading uses warm starts to accelerate eigenvector computation across scales and, more importantly, aligns eigenspaces across scales. Within clusters of nearly repeated eigenvalues, the method projects transferred eigenvectors into the current eigenspace and orthonormalizes them; equivalently, it solves an orthogonal Procrustes problem inside each clustered eigenspace. This removes arbitrary sign flips and basis rotations, producing consistent spectral coordinates and suggesting a route to cross-scale kernels that remain stable under graph refinement (Mike et al., 2018).

6. Computation, theoretical analysis, and empirical behavior

Computation is a defining constraint of multiscale Laplacian kernels. The original recursive MLG construction is made feasible by randomized projection and dataset-level linearization; the transductive Gaussian process kernel requires an eigendecomposition of the Laplacian and a dense inverse of $t$0, which is cubic in the number of nodes and therefore limited to moderate graph sizes; and exact optimal-transport or eigenvector-based graph-comparison methods are often dominated by spectral computation (Kondor et al., 2016). For large-scale matrix-function kernels $t$1, block Krylov subspace methods provide a different route: after constructing a single block Krylov basis for $t$2, one can approximate

$t$3

for many scales $t$4. Among the methods studied, classical block Lanczos preserves symmetric positive definiteness of Gram matrices when $t$5 is positive on the Laplacian spectrum; global block Lanczos preserves symmetry but not necessarily positive definiteness for small subspace dimension; sequential Lanczos may lose symmetry; Chebyshev approximation is symmetric but not necessarily positive semidefinite; and the squared-Chebyshev variant enforces positive semidefiniteness by approximating $t$6 and squaring the result (Erb, 2023).

Recent analysis also clarifies the continuum limit behind scale selection. For i.i.d. samples on a closed manifold, the empirical Gaussian graph Laplacian

$t$7

is analyzed as an approximation to the manifold operator $t$8, where $t$9 is the Laplace–Beltrami operator. The resulting non-asymptotic bounds show that the multiscale parameter $t$0 creates a bias–variance tradeoff: stochastic error increases as $t$1, while approximation bias between the Euclidean Gaussian kernel and the manifold heat kernel decreases as $t$2 (Wahl, 2024). A complementary RKHS-based approach, Kernelized Diffusion Maps, replaces local averaging by an estimator built from covariance and derivative-reproducing operators in an RKHS. Under a source condition that the first eigenfunctions of the diffusion operator lie in the RKHS, it yields dimension-free operator-norm convergence, with the paper recommending $t$3 as the bias–variance balance and using Nyström subsampling or random Fourier features for scalability (Pillaud-Vivien et al., 2023).

Empirical results differ by task but consistently show that multiscale Laplacian constructions are useful when single-scale or single-view kernels are too rigid. For graph classification, the original MLG kernel reported $t$4 on MUTAG, $t$5 on PTC, $t$6 on ENZYMES, and $t$7 on PROTEINS; on NCI1 and NCI109 it was below Weisfeiler–Lehman but above other non-WL baselines (Kondor et al., 2016). In low-label semi-supervised learning, the transductive Gaussian process with a degree-$t$8 Laplacian polynomial was top or competitive on multiple real graphs, and TGGP-X often obtained the best reported accuracies, including $t$9 on Texas, $L_{\mathrm{sym}}$00 on Cornell, $L_{\mathrm{sym}}$01 on Wisconsin, and $L_{\mathrm{sym}}$02 on Citeseer (Zhi et al., 2022). In Multiscale Laplacian Learning, MMBO achieved $L_{\mathrm{sym}}$03 PRBEP on WebKB page+link and MML reached $L_{\mathrm{sym}}$04 on USPST with only $L_{\mathrm{sym}}$05 labels, while two or three scales were reported to be typically sufficient for significant gains over the single-scale case (Merkurjev et al., 2021).

Several recurring limitations follow from these results. Over-smoothing remains a risk when the chosen spectral filter is too strongly low-pass, especially on low-homophily graphs; positive definiteness is not automatic for every distance-substitution kernel or every approximate matrix-function method; and cubic eigendecomposition or inversion costs still limit exact transductive or recursive formulations. A plausible implication is that contemporary work on multiscale Laplacian graph kernels is best understood as a trade-off among three objectives: expressive multiscale geometry, positive semidefinite kernel structure, and scalable approximation.