Laplace Harmonic Eigenmaps

• Laplace harmonic eigenmaps are coordinate systems derived from low-frequency Laplacian eigenfunctions, providing intrinsic, geometry-adapted embeddings in both manifold and graph settings.
• They employ methods from the Laplace–Beltrami framework and discrete FEM techniques to capture global and localized geometric structures across smooth and sampled data.
• Advanced variants integrate spectral filtering, harmonic mapping to ellipsoids, and Hodge theoretic extensions to address challenges such as sign ambiguity, localization, and scale selection.

Laplace harmonic eigenmaps are coordinate constructions derived from eigenfunctions of Laplace-type operators. In the literature considered here, the expression spans two closely related uses. In manifold learning and spectral geometry, low-frequency eigenfunctions of the Laplace–Beltrami operator or of a graph Laplacian are used as intrinsic coordinates, so that sampled eigenvectors act as discrete manifold harmonics (Wahl, 2024). In geometric analysis, the term is also used explicitly for harmonic maps from a surface into an ellipsoid $\mathcal E_\Lambda\subset \mathbb R^n$ whose coordinate functions become Laplace eigenfunctions for an associated conformal metric (Petrides, 14 Aug 2025). Across both uses, the common principle is that Laplacian eigenmodes furnish geometry-adapted coordinates, embeddings, or map components.

1. Spectral foundations

The common analytic starting point is the Laplace–Beltrami eigenproblem on a compact manifold $M$, possibly with boundary,

$-\Delta_M \phi = \lambda \phi,$

with an admissible boundary condition $$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$. This produces nonnegative eigenvalues

$0 \le \lambda_1 \le \lambda_2 \le \cdots$

and eigenfunctions that form an orthonormal basis for $L^2(M)$. In this sense, manifold harmonics are the geometric analogue of Fourier modes (Beckman et al., 20 May 2026).

On a compact Riemannian manifold $X$, the Laplace–Beltrami operator may be written

$\Delta_X=-\mathrm{div}_X(\nabla_X),$

and the associated Dirichlet energy is

$$\mathcal E_S(f)=\langle \nabla_X f,\nabla_X f\rangle_{L^2(TX)}=\langle f,\Delta f\rangle_{L^2(X)}.$$

The classical eigenbasis is characterized variationally as the smoothest orthonormal basis: $$\min_{\psi_1,\dots,\psi_k}\ \sum_{i=1}^k \mathcal E_S(\psi_i) \quad \text{s.t.}\quad \langle \psi_i,\psi_j\rangle_{L^2(X)}=\delta_{ij}.$$ This makes low eigenmodes the globally smoothest orthonormal coordinates available to the geometry (Melzi et al., 2017).

A plausible synthesis is that “Laplace harmonic eigenmap” is not a single algorithmic object but a family of constructions induced by these intrinsic harmonics. The main distinctions concern whether the operator is continuous or discrete, scalar or Hodge-theoretic, global or localized, and whether the target is Euclidean space, a sphere, an ellipsoid, or a space of differential forms.

2. Classical eigenmap constructions on manifolds and graphs

In the standard spectral-embedding picture, points are assigned coordinates by evaluating a small number of low-frequency eigenfunctions. In the discrete manifold-learning setting, this is the familiar construction

$M$0

which uses the first nontrivial eigenvectors as embedding coordinates (Melzi et al., 2017). On graphs, the generalized eigenproblem takes the form

$M$1

and after diagonal rescaling $M$2,

$M$3

These discrete eigenvectors are treated as analogues of Laplace–Beltrami eigenfunctions (Beckman et al., 20 May 2026).

On triangulated surfaces, a standard discretization uses the FEM cotangent Laplacian. With stiffness matrix $M$4 and area matrix $M$5, the discrete eigenproblem is

$M$6

with area-weighted inner product $M$7. This is the canonical mesh-based route from smooth manifold harmonics to computable eigen-coordinates (Melzi et al., 2017).

The diffusion-geometric interpretation is equally central. The heat kernel admits the spectral expansion

$M$8

and the diffusion map embedding is

$M$9

At $-\Delta_M \phi = \lambda \phi,$0, this reduces to the raw Laplace–Beltrami eigenspace embedding. The Euclidean distance in these coordinates approximates diffusion distance, which explains why low eigenmodes capture intrinsic large-scale geometry (Shtern et al., 2013).

Recent analysis makes this harmonic interpretation precise. For i.i.d. samples uniformly distributed on a closed manifold $-\Delta_M \phi = \lambda \phi,$1, non-asymptotic error bounds show that the eigenvalues and eigenspaces of the empirical Gaussian graph Laplacian are close to those of the Laplace–Beltrami operator. The same work identifies the heat kernel of $-\Delta_M \phi = \lambda \phi,$2 as a reproducing-kernel feature map and the heat semigroup as the corresponding covariance operator, so graph eigenvectors can be interpreted as estimators of sampled manifold harmonics (Wahl, 2024).

3. Localized, filtered, and adaptive variants

A basic limitation of classical Laplacian eigenbases is global support: low modes encode the entire manifold and are inefficient for isolating local structure. Localized manifold harmonics address this by replacing the standard Laplacian with a modified operator

$-\Delta_M \phi = \lambda \phi,$3

and solving

$-\Delta_M \phi = \lambda \phi,$4

The diagonal potential term localizes the basis on a prescribed region, while the low-rank projector term enforces approximate orthogonality to an already chosen low-frequency Laplacian subspace. The resulting mixed coordinate system

$-\Delta_M \phi = \lambda \phi,$5

acts as a global-plus-local harmonic eigenmap, refining coarse geometry with region-aware spectral detail (Melzi et al., 2017).

A broader generalization uses spectral filtering rather than basis truncation. For a positive filter $-\Delta_M \phi = \lambda \phi,$6, the spectral kernel

$-\Delta_M \phi = \lambda \phi,$7

and spectral distance

$-\Delta_M \phi = \lambda \phi,$8

define filtered harmonic embeddings that subsume harmonic or commute-time, diffusion, biharmonic, and wave constructions. In this viewpoint, the entire Laplacian basis is retained and reweighted rather than truncated (Patanè, 2019).

Low Distortion Local Eigenmaps pushes locality further. It computes a global pool of graph Laplacian eigenvectors, estimates local inner products of eigenfunction gradients, and selects different subsets of eigenvectors in different neighborhoods to construct low-distortion local views $-\Delta_M \phi = \lambda \phi,$9. These local charts are then registered by Procrustes analysis. The method can embed closed and non-orientable manifolds into their intrinsic dimension by tearing them apart, which directly targets a failure mode of global spectral embeddings on manifolds with nontrivial topology or severe global distortion (Kohli et al., 2021).

4. Harmonic maps to ellipsoids and extremal metrics

In geometric analysis, a Laplace harmonic eigenmap is a harmonic map

$$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$0

where the target ellipsoid is

$$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$1

The defining constraint is

$$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$2

and the harmonic map equation becomes

$$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$3

Equivalently, each coordinate solves

$$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$4

The ellipsoid is natural because it encodes the weighted quadratic identity satisfied by a family of eigenfunctions with possibly different eigenvalues (Petrides, 14 Aug 2025).

After conformally changing the metric by

$$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$5

the coordinate functions $$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$6 become Laplace eigenfunctions associated with the eigenvalues $$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$7. The sphere case $$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$8 reduces to the classical harmonic map/eigenmap picture. Analytically, the equation can be rewritten in Rivière form

$$g\!\left(\phi,\frac{\partial \phi}{\partial n}\right)=0$$9

with

$0 \le \lambda_1 \le \lambda_2 \le \cdots$0

and this structure underlies dimension-free $0 \le \lambda_1 \le \lambda_2 \le \cdots$1-regularity and $0 \le \lambda_1 \le \lambda_2 \le \cdots$2 gradient bounds under small normal-energy assumptions (Petrides, 14 Aug 2025).

The same circle of ideas appears in extremal metric theory. For a closed $0 \le \lambda_1 \le \lambda_2 \le \cdots$3-manifold, a conformally extremal metric for the normalized Laplace eigenvalue corresponds to a non-degenerate $0 \le \lambda_1 \le \lambda_2 \le \cdots$4-harmonic map

$0 \le \lambda_1 \le \lambda_2 \le \cdots$5

with

$0 \le \lambda_1 \le \lambda_2 \le \cdots$6

Among all metrics, the corresponding objects are minimal immersions into spheres. Thus, in higher dimensions, ordinary harmonic maps are replaced by $0 \le \lambda_1 \le \lambda_2 \le \cdots$7-harmonic sphere maps, which become genuine eigenmaps after the canonical conformal normalization (Karpukhin et al., 2021).

5. Computation and numerical approximation

On smooth embedded surfaces, Laplace–Beltrami eigenpairs can be computed without global parametrization by the Closest Point Method. The surface problem

$0 \le \lambda_1 \le \lambda_2 \le \cdots$8

is replaced by a regularized embedding-space eigenproblem

$0 \le \lambda_1 \le \lambda_2 \le \cdots$9

on a narrow Cartesian band around the surface. For $L^2(M)$0, the paper establishes a one-to-one correspondence between the surface eigenproblem and the regularized embedding problem. After discretization, the key matrix is

$L^2(M)$1

where $L^2(M)$2 is a Cartesian finite-difference Laplacian and $L^2(M)$3 is the closest-point extension matrix (Macdonald et al., 2011).

Once eigenfunctions are available, the manifold harmonic transform expresses sampled data by

$L^2(M)$4

A butterfly factorization compresses the dense matrix $L^2(M)$5 by hierarchical low-rank approximations, so that manifold-harmonic synthesis and related inverse operations become feasible at scales where explicit dense storage is prohibitive. The same compression applies to graph-Laplacian eigenvectors in Laplacian eigenmaps (Beckman et al., 20 May 2026).

When explicit eigendecomposition is undesirable, Laplacian spectral kernels and distances can be approximated without direct spectral truncation. A Padé–Chebyshev rational approximation of $L^2(M)$6 reduces the computation of filtered kernels and distances to sparse, symmetric, and well-conditioned linear systems associated with inhomogeneous Laplace equations. This replaces repeated eigenpair computation by iterative sparse solves while preserving the operator-theoretic structure of the embedding (Patanè, 2019).

6. Topological, Hodge-theoretic, and directed extensions

Scalar Laplacian eigenmaps admit higher-order analogues based on differential forms. Hodge Diffusion Maps replace the scalar Laplace–Beltrami operator by the $L^2(M)$7-th Hodge Laplacian

$L^2(M)$8

approximate the exterior derivative from point samples, and form the discrete Hodge Laplacian

$L^2(M)$9

The embedding then uses eigenvectors of $X$0. Because

$X$1

these constructions capture $X$2-dimensional holes and higher-order topology rather than only scalar diffusion structure (Gomez et al., 10 Apr 2025).

A more rigorous version of this extension is given by empirical Hodge Laplacians on differential forms. There, symmetrized empirical operators converge in probability to the smooth Hodge Laplacian, extending the scalar Belkin–Niyogi framework from functions to forms. The resulting theory recovers harmonic forms, the cohomology ring $X$3, the second fundamental form, the curvature tensor, and Pontryagin classes and numbers from sampled data (Lê, 21 May 2026).

Topological enrichment can also replace real-valued coordinates by circle-valued ones. Persistent cohomology identifies a class in $X$4, harmonic smoothing solves

$X$5

and the resulting harmonic cocycle $X$6 satisfies

$X$7

Integrating $X$8 yields a coordinate in $X$9. This construction is complementary to Laplacian eigenmaps precisely when the intrinsic coordinate is periodic rather than real-valued (0905.4887).

Directed networks require yet another generalization. Magnetic eigenmaps replace the ordinary graph Laplacian by the Hermitian magnetic Laplacian

$\Delta_X=-\mathrm{div}_X(\nabla_X),$0

so that low-energy complex eigenvectors encode edge orientation, directed cycles, and flow structure. In that setting, the relevant harmonic objects are no longer real scalar graph functions but complex sections twisted by a $\Delta_X=-\mathrm{div}_X(\nabla_X),$1 connection (Fanuel et al., 2016).

7. Ambiguities, stability, and open limitations

A recurrent limitation is that independently computed eigenfunctions are often incompatible across datasets or shapes. For non-rigid shapes, low-order Laplace–Beltrami eigenfunctions suffer from sign ambiguity, permutation ambiguity, multiplicity-induced basis nonuniqueness, and high-frequency instability. Matching eigenspaces across shapes therefore requires additional machinery; one successful strategy minimizes discrepancies in third-order moments and augments them with gradient-based invariants involving

$\Delta_X=-\mathrm{div}_X(\nabla_X),$2

to resolve antisymmetric sign ambiguities (Shtern et al., 2013).

Scale selection is a second major instability. In data-based constructions, approximation of the target eigenmaps depends crucially on the scaling parameter $\Delta_X=-\mathrm{div}_X(\nabla_X),$3: if $\Delta_X=-\mathrm{div}_X(\nabla_X),$4 is too small or too large, the approximation is inaccurate or completely breaks down. Explicitly solvable models on intervals, weighted one-dimensional spaces, squares, tori, spheres, and the Sierpinski gasket show that the $\Delta_X=-\mathrm{div}_X(\nabla_X),$5 optimal for recovering low eigenspaces may differ from the $\Delta_X=-\mathrm{div}_X(\nabla_X),$6 optimal for pointwise operator approximation, and that singular spaces require geometry-adapted scaling exponents rather than the manifold value $\Delta_X=-\mathrm{div}_X(\nabla_X),$7 (Akwei et al., 2024).

These limitations also clarify a common misconception. Laplace harmonic eigenmaps are not a single universally fixed embedding with unique coordinates. In the cited literature, they form a family of constructions ranging from low-frequency graph embeddings, through localized and filtered manifold harmonics, to ellipsoid-valued harmonic maps and Hodge-theoretic extensions. A plausible unifying statement is that all of them treat Laplacian spectral data as intrinsic coordinates, but they differ substantially in operator choice, target space, topology sensitivity, regularity theory, and stability regime.