Abstract
We discuss the geometry of Laplacian eigenfunctions $-\Delta \phi = \lambda \phi$ on compact manifolds $(M,g)$ and combinatorial graphs $G=(V,E)$. The 'dual' geometry of Laplacian eigenfunctions is well understood on $\mathbb{T}^d$ (identified with $\mathbb{Z}^d$) and $\mathbb{R}^n$ (which is self-dual). The dual geometry is of tremendous role in various fields of pure and applied mathematics. The purpose of our paper is to point out a notion of similarity between eigenfunctions that allows to reconstruct that geometry. Our measure of 'similarity' $ \alpha(\phi_{\lambda}, \phi_{\mu})$ between eigenfunctions $\phi_{\lambda}$ and $\phi_{\mu}$ is given by a global average of local correlations:
$$\alpha(\phi_{\lambda}, \phi_{\mu})^2 = \| \phi_{\lambda} \phi_{\mu} \|_{L^2}^{-2}\int_{M}{ \left( \int_{M}{ p(t,x,y)( \phi_{\lambda}(y) - \phi_{\lambda}(x))( \phi_{\mu}(y) - \phi_{\mu}(x)) dy} \right)^2 dx},$$
where $p(t,x,y)$ is the classical heat kernel and $e^{-t \lambda} + e^{-t \mu} = 1$. This notion recovers all classical notions of duality but is equally applicable to other (rough) geometries and graphs; many numerical examples in different continuous and discrete settings illustrate the result.
Introduction
In classical Fourier analysis on a circle or flat Euclidean space, eigenfunctions (sines and cosines) possess a clear additive structure. They relate to one another through a well-defined "dual geometry" where multiplying two waves results in harmonics with frequencies that are the sum and difference of the originals. This structure is the backbone of harmonic analysis and signal processing.
However, when moving to complex domains—such as jagged manifolds or social network graphs—this intuitive structure vanishes. While we can compute the harmonics (eigenfunctions) and their pitch (eigenvalues), we lose the sense of "direction" or how these harmonics relate to each other. Eigenvalues provide only a one-dimensional ordering that fails to capture the multi-dimensional geometry of the underlying space. This paper introduces a similarity measure that recovers this hidden dual geometry for any arbitrary domain.
Methodology: Measuring Local Correlation
The core of the proposed method is a new measure of similarity, $\alpha(\phi_\lambda, \phi_\mu)$, based on local correlations weighted by the heat kernel. The heat kernel acts as a "local window," observing how functions behave over small distances. The authors demonstrate that this measure is equivalent to the $L^2$ norm of the heat-diffused product of the two eigenfunctions:
$$\alpha(\phi_\lambda, \phi_\mu) = \frac{\|e^{t\Delta}(\phi_\lambda \phi_\mu)\|_{L^2}}{\|\phi_\lambda \phi_\mu\|_{L^2}}$$
By calculating this similarity for every pair of eigenfunctions, we construct an $n \times n$ affinity matrix (Figure 4). This matrix reveals the structural relationships between eigenfunctions that are otherwise obscured by their eigenvalues.
Reconstructing the Dual Landscape
Once the affinity matrix is established, spectral embedding is used to map each eigenfunction into a low-dimensional Euclidean space (e.g., $\mathbb{R}^2$ or $\mathbb{R}^3$). This creates a visual "landscape" where eigenfunctions that play similar structural roles are placed close together.
As shown in Figure 5, this embedding can perfectly recover the linear structure of frequencies on a torus or the degree ($\ell$) and order ($m$) indices of spherical harmonics. This "dual landscape" allows for the discovery of product structures even in messy or random graphs, providing a multi-dimensional map of the spectrum that goes far beyond a simple 1D list of eigenvalues.
Key Results and Insights
- Product Structure Discovery: If a domain is a product of two spaces, the dual landscape reflects this Cartesian structure, even if one component is a structure-less random graph.
- Local to Global: Purely local information (correlations over tiny distances) is sufficient to reconstruct the global, large-scale dual geometry of the entire spectrum.
- Exploratory Power: The method serves as a tool for "Exploratory Spectral Graph Theory," allowing researchers to find hidden patterns in real-world networks where the dual geometry is not yet known.
Conclusion
By defining a similarity measure based on local correlations, we can reconstruct the hidden dual geometry of Laplacian eigenfunctions on arbitrary domains. This approach provides a powerful new lens for analyzing complex manifolds and graphs, moving beyond the limitations of one-dimensional eigenvalue ordering to reveal the multi-dimensional structural relationships inherent in the harmonics of a space.
Summary updated