Deformation Laplacian: Concepts and Applications

• Deformation Laplacian is a family of operators that use Laplacian coordinates to drive shape deformations in meshes, graphs, and point clouds.
• It employs least-squares formulations and regularization techniques to propagate control constraints while preserving local geometric details.
• Its applications span point cloud completion, non-rigid registration, spectral analysis, and dynamic deformations, showcasing its versatility in computational geometry.

Deformation Laplacian denotes a family of Laplace-governed deformation frameworks rather than a single universally fixed operator. In geometry processing, “Deformation Laplacian” or “Laplacian deformation” refers to methods that edit or reconstruct shapes by minimizing a quadratic Laplacian energy under positional constraints. In more recent work, the same conceptual pattern appears in graph-based point cloud completion, learned shape operators for image deformation, proxy-free Gaussian-splat editing, and in spectral geometry as the study of how Laplace operators and their eigenobjects vary when domains, metrics, or branch-point configurations are deformed (Shi et al., 2021, Lee et al., 2022, Dabrowski, 2017).

1. Classical Laplacian deformation framework

In the classical graphics formulation, the discrete Laplace operator on a mesh or graph with vertices $V=\{v_1,\dots,v_n\}$ and edges $E$ is a linear operator $L\in\mathbb{R}^{n\times n}$ acting on vertex positions $X\in\mathbb{R}^{n\times 3}$ by

$(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$

with $L=D-W$, where $W=(w_{ij})$ is a weighted adjacency matrix and $D$ is the diagonal degree matrix. Two standard choices are the uniform, or umbrella, Laplacian and the cotangent Laplacian on triangulated surfaces (Shi et al., 2021).

The central object is the Laplacian coordinate

$\delta_i=(LX)_i,$

which encodes local differential shape. These coordinates are translation invariant and sensitive to local shape and detail. Laplacian deformation therefore preserves local differential structure rather than raw vertex positions. In the least-squares formulation, one prescribes a set $C$ of constrained or control vertices with target positions $E$0, and seeks deformed positions $E$1 minimizing a Laplacian term together with a constraint term: $E$2 For each coordinate, this yields the normal equations

$E$3

This framework fixes the canonical meaning of deformation Laplacian in geometry processing: a Laplace-type operator propagates the effect of handles or control points while regularizing the deformation by preserving local shape. The later graph, point-cloud, and learned formulations are best understood as extensions, relaxations, or simulations of this least-squares paradigm.

2. Graph- and mesh-based realizations

A direct graph realization appears in point cloud completion. “Graph-Guided Deformation for Point Cloud Completion” recasts completion as a mesh-like Laplacian deformation problem on a point cloud graph: a coarse generated point set $E$4 provides supporting points, a subset of observed input points $E$5 provides controlling points, and the combined set $E$6 is equipped with a $E$7-nearest-neighbor graph whose umbrella Laplacian is

$E$8

The method introduces a shape-preserving Laplacian loss together with reconstruction and control-point matching losses,

$E$9

and uses a GCN with a G-ResNet design to update point positions while recomputing adjacency at every forward pass. The paper explicitly states that it “transform[s] the point cloud completion problem into a traditional mesh deformation task” and reports that removing the deformation network degrades Chamfer distance on ShapeNet from $L\in\mathbb{R}^{n\times n}$0 to $L\in\mathbb{R}^{n\times n}$1 (Shi et al., 2021).

A mesh-based realization appears in non-rigid registration. “Laplacian ICP for Progressive Registration of 3D Human Head Meshes” uses the cotangent approximation to the Laplace–Beltrami operator as a deformation regularizer for progressive coarse-to-fine ICP. With template vertices $L\in\mathbb{R}^{n\times n}$2 and cotangent Laplacian $L\in\mathbb{R}^{n\times n}$3, the regularization energy is

$L\in\mathbb{R}^{n\times n}$4

which, under a small-deformation-per-iteration assumption, is approximated by

$L\in\mathbb{R}^{n\times n}$5

This produces a sparse linear least-squares system that regularizes incremental displacement fields rather than per-vertex affine transforms. On the Headspace benchmark, the method achieves comparable annotation-transfer performance to a per-vertex affine-constraint variant while being about $L\in\mathbb{R}^{n\times n}$6 faster, reported as $L\in\mathbb{R}^{n\times n}$7 seconds versus approximately $L\in\mathbb{R}^{n\times n}$8 minutes on a laptop (Pears et al., 2023).

These two cases illustrate the principal discrete variants of the deformation Laplacian idea. In one, the Laplacian is synthesized on a changing point-cloud graph and learned through GCN message passing; in the other, it is the classical cotangent Laplace–Beltrami operator on fixed mesh connectivity. In both, deformation is regularized by preserving local Laplacian structure rather than by enforcing rigid correspondences alone.

3. Learned operators, biharmonic energies, and Gaussian splats

A learned deformation Laplacian appears in single-image manipulation. “Pop-Out Motion: 3D-Aware Image Deformation via Learning the Shape Laplacian” predicts a discrete shape Laplacian on a volumetric point-cloud representation reconstructed from a single image, rather than deriving it directly from noisy or non-manifold geometry. The learned operator $L\in\mathbb{R}^{n\times n}$9 drives bounded biharmonic weights through the quadratic form

$X\in\mathbb{R}^{n\times 3}$0

and handle-based deformation is then obtained from these weights. The paper argues that reconstruction alone is insufficient because of topology errors and reports more accurate deformation weights than alternative methods based on mesh reconstruction and point-cloud Laplacian methods (Lee et al., 2022).

A proxy-free variant appears for Gaussian splats. “Proxy-Free Gaussian Splats Deformation with Splat-Based Surface Estimation” constructs a surface-aware splat graph directly on anisotropic Gaussian primitives. Neighboring splats are defined by intersections of occupancy regions rather than by center-to-center distance alone, geodesic neighborhoods are computed on this graph, and a cotan-like Laplacian is then used inside standard ARAP and bounded biharmonic weights pipelines. For BBW, the quadratic energy is written with

$X\in\mathbb{R}^{n\times 3}$1

where $X\in\mathbb{R}^{n\times 3}$2 is the splat Laplacian and $X\in\mathbb{R}^{n\times 3}$3 a mass matrix. The method is evaluated on $X\in\mathbb{R}^{n\times 3}$4 challenging objects from ShapeNet, Objaverse, and Sketchfab, as well as NeRF-Synthetic, and is presented as a direct transplantation of Laplacian deformation from meshes to Gaussian-splat representations (Kim et al., 24 Nov 2025).

A useful clarification is provided by dynamic Gaussian-splatting work. “Laplacian Analysis Meets Dynamics Modelling: Gaussian Splatting for 4D Reconstruction” uses the term Laplacian in two different senses: a Fourier-like temporal basis

$X\in\mathbb{R}^{n\times 3}$5

for motion representation, and a multi-scale image Laplacian pyramid loss

$X\in\mathbb{R}^{n\times 3}$6

The paper explicitly states that there is no graph Laplacian, mesh Laplacian, or deformation Laplacian on Gaussians in the traditional geometry-processing sense (Zhou et al., 7 Aug 2025).

Taken together, these works show two divergent tendencies in the recent literature. One tendency preserves the classical operator-based meaning and adapts it to new representations by learning weights or by constructing new adjacency structures. The other uses “Laplacian” spectrally or multiscale, without introducing a spatial deformation operator of the classical kind.

4. Spectral variation under domain and metric deformation

In spectral geometry, deformation Laplacian often means the Laplacian on a varying domain or varying metric. “A localized boundary deformation which splits the spectrum of the Laplacian” studies Dirichlet, Neumann, and Robin Laplacians on a bounded Lipschitz domain $X\in\mathbb{R}^{n\times 3}$7 under localized boundary diffeomorphisms $X\in\mathbb{R}^{n\times 3}$8, with deformed domains $X\in\mathbb{R}^{n\times 3}$9. Using a Hadamard-type derivative formula and spectral stability estimates, the paper proves that for any $(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$0 and any boundary point $(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$1, one can construct a Lipschitz domain $(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$2 with

$(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$3

such that all eigenvalues of the Laplacian on $(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$4 are simple (Dabrowski, 2017).

A stability result of a different kind appears for Neumann eigenfunctions. “The Stability of the First Neumann Laplacian Eigenfunction Under Domain Deformations and Applications” studies a diffeomorphism $(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$5 and compares the first nontrivial Neumann eigenfunction $(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$6 on $(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$7 with the eigenbasis $(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$8 on a reference domain $(LX)_i=\sum_{j\in\mathcal{N}(i)} w_{ij}(x_i-x_j),$9. If the singular values of $L=D-W$0 lie in $L=D-W$1, then

$L=D-W$2

On the rectangle $L=D-W$3, the paper obtains

$L=D-W$4

showing that the first Neumann eigenfunction remains essentially one-dimensional on a tall thin domain under small deformation (Marshall, 2017).

A broader perturbative framework is developed in “The spectrum of the Laplacian on forms”. For self-adjoint nonnegative operators, the paper proves a generalized Weyl criterion and applies it to the Laplacian on $L=D-W$5-forms under continuous metric perturbations. If two complete metrics $L=D-W$6 and $L=D-W$7 are $L=D-W$8-close, then for each $L=D-W$9 the spectra and essential spectra of the corresponding Hodge Laplacians are close in pointed Gromov–Hausdorff sense; on compact manifolds this recovers continuity of eigenvalues under metric deformation (Charalambous et al., 2018).

These results shift the meaning of deformation Laplacian from a regularizer used to deform geometry to a spectral object whose eigenvalues and eigenfunctions are themselves tracked under deformation of the underlying geometric data.

5. Waveguides, conifolds, and metric-measure deformation properties

In thin-waveguide analysis, geometric deformation changes the operator directly. “On the spectrum and weakly effective operator for Dirichlet Laplacian in thin deformed tubes” studies a family of thin tubes whose cross-sections are multiplied by a scalar deformation $W=(w_{ij})$0 with a single global maximum. After subtracting the leading transverse energy $W=(w_{ij})$1 and rescaling, the eigenvalues converge to those of the one-dimensional harmonic oscillator

$W=(w_{ij})$2

and the paper emphasizes that this asymptotic behavior is not influenced by curvature, torsion, or twisting (Oliveira et al., 2011).

A related waveguide problem is treated in the magnetic setting. “Magnetic Dirichlet Laplacian on deformed waveguides” studies

$W=(w_{ij})$3

with a compactly supported magnetic field. The paper proves that for sufficiently small deformation amplitude $W=(w_{ij})$4, the discrete spectrum below the threshold $W=(w_{ij})$5 is empty, while the essential spectrum remains

$W=(w_{ij})$6

It also shows that, in the non-magnetic case, the discrete spectrum below $W=(w_{ij})$7 is non-empty under an explicit condition on $W=(w_{ij})$8 and $W=(w_{ij})$9, thereby isolating the stabilizing role of the magnetic field (Alpay et al., 14 Apr 2026).

In exceptional holonomy, the spectrum on the link of a cone controls deformation theory. “Deformation theory of $D$0 conifolds” studies asymptotically conical and conically singular $D$1-manifolds and shows that, for generic asymptotic rates $D$2 in the AC case, the moduli space is smooth, while for generic $D$3 in the AC case and for generic positive rates in the CS case the deformation theory is in general obstructed. The obstruction spaces are described explicitly in terms of the spectrum of the Laplacian on the links of the cones (Karigiannis et al., 2012).

In metric-measure geometry, the term acquires yet another analytic meaning. “A note on Laplacian bounds, deformation properties and isoperimetric sets in metric measure spaces” shows that an upper bound on the Laplacian of squared distance functions,

$D$4

implies a deformation inequality for finite-perimeter sets,

$D$5

and consequently yields the deformation property used to prove that every isoperimetric set in an essentially non-branching $D$6 space has an open representative (Pasqualetto et al., 18 Mar 2025).

These examples show that, in continuum analysis, deformation Laplacian frequently means an operator whose asymptotics, coercivity, or obstruction theory are governed by how geometry changes at infinity, near singularities, or under weighted perturbations.

6. Rigidity, classification, and scope

A strong rigidity statement appears in “Deformation rigidity for $D$7 eigensections”. For a configuration $D$8 of $D$9 branch points on $\delta_i=(LX)_i,$0, the paper studies the Laplacian on a flat real line bundle $\delta_i=(LX)_i,$1 with monodromy $\delta_i=(LX)_i,$2 around each branch point. Eigensections have local expansions

$\delta_i=(LX)_i,$3

and a critical eigensection is defined by vanishing trace $\delta_i=(LX)_i,$4. The main theorem states that every minimal non-degenerate critical eigensection is deformation rigid: any sufficiently small deformation of the configuration that still admits a critical eigensection must come from an $\delta_i=(LX)_i,$5-rotation (Haydys et al., 18 Apr 2026).

This rigidity result clarifies a recurrent structural theme across the literature. Whether one studies classical Laplacian surface editing, graph-guided point-cloud deformation, spectral simplicity under localized boundary perturbation, or $\delta_i=(LX)_i,$6 conifold moduli, the essential question is how a Laplace-type operator encodes what may deform freely and what is constrained by local differential structure, global topology, or symmetry.

A common misconception is that deformation Laplacian always names a specific graph or mesh matrix. The recent literature does not support that identification. In geometry processing it usually denotes a quadratic regularizer based on Laplacian coordinates; in learned settings it may denote a predicted graph operator or a data-driven deformation prior; in spectral geometry it may refer to the Laplacian on a deformed domain or metric; and in some works the term “Laplacian” is only spectral or multiscale rather than operator-theoretic in the classical sense (Zhou et al., 7 Aug 2025).

This suggests that deformation Laplacian is best treated as a technical family of Laplace-centered deformation paradigms. What unifies them is not a single formula, but the role of a Laplace-type structure in propagating constraints, preserving local differential information, characterizing admissible deformations, or controlling the spectral response of a geometric object under perturbation.