p-Laplacian Equations on Point Clouds

• p-Laplacian equations on point clouds are a nonlinear regularization framework that generalizes classical Laplacian methods to discrete, high-dimensional data with applications in semi-supervised learning and clustering.
• The discrete-to-continuum analysis shows that as sample sizes grow, solutions of the discrete p-Laplacian converge to those of weighted continuum PDEs, ensuring methodological consistency.
• Algorithmic strategies like SPDHG and PDE-inspired methods enable scalable optimization and robust label propagation, outperforming traditional approaches in low-label regimes and image inpainting tasks.

The $p$-Laplacian on point clouds is a nonlinear operator and associated regularization framework that generalizes classical Laplacian-based methods to accommodate nonlinearity and adaptivity in semi-supervised learning, interpolation, clustering, and related computational tasks on high-dimensional data clouds. Point cloud $p$-Laplacian methods formalize the extension of $p$-Dirichlet functionals, traditionally defined on Euclidean domains, to discretized settings where only finite samples from an unknown manifold or density are available. The theory encompasses both pairwise (graph) and higher-order (hypergraph) connectivities, with rigorous connections to continuum $p$-Laplacian PDEs established via variational and viscosity-solution frameworks. Core results demonstrate that as the number of data points increases while the number of labeled points remains fixed and under appropriate growth conditions on neighborhood parameters, minimizers and solutions on finite point clouds converge to solutions of weighted continuum $p$-Laplacian equations with mixed Dirichlet and Neumann boundary conditions.

1. Discrete $p$-Laplacian Models on Point Clouds

Given a finite set of points $$\Omega_n = \{x_1,\ldots,x_n\} \subset \Omega \subset \mathbb{R}^d$$ drawn i.i.d. from a Borel probability measure $\mu=\rho(x)\,dx$, the standard approach is to define a neighborhood structure via either:

• An $\varepsilon_n$-ball relation, connecting points within distance $\varepsilon_n$;
• A $k_n$-nearest neighbor relation, connecting each point to its $k_n$ nearest neighbors.

Weights are typically assigned via a radial, compactly supported kernel $\eta$, possibly scaled to ensure locality and proper normalization. For graphs (pairwise relationships), the discrete $p$-Dirichlet energy is

$$E_p^d(u) = \frac{1}{\varepsilon_n^p n^2}\sum_{i,j=1}^n w_{ij}|u(x_i) - u(x_j)|^p,$$

with $w_{ij} = \eta_{\varepsilon_n}(|x_i-x_j|)$. For hypergraphs, the energy penalizes the maximal difference within each neighborhood: $$\mathcal{E}_{n,\varepsilon_n}(u) = \frac{1}{n\varepsilon_n^p} \sum_{k=1}^n \max_{x_i, x_j \in e_k} |u(x_i)-u(x_j)|^p,$$ where $e_k$ denotes the neighborhood of $x_k$ (edge or hyperedge) (Shi et al., 2024, Shi, 22 Jan 2026).

Boundary and labeling constraints are imposed via hard Dirichlet conditions $u(x_i)=y_i$ for a fixed label set $\mathcal{O}$.

2. Discrete-to-Continuum Limit and PDE Connections

A central result is the discrete-to-continuum consistency of such $p$-Laplacian energies and associated equations. Under the key assumptions:

• $p > d$ (to ensure coercivity and regularity in Sobolev spaces $W^{1,p}$),
• the number of labeled points $N$ is fixed as $n \to \infty$,
• the scale parameter $\varepsilon_n$ satisfies optimal-transport and connectivity lower bounds, e.g., $\varepsilon_n \gg (\log n)^{1/d}/n^{1/d}$,

the empirical solution $u_n$ converges (almost surely in appropriate topologies) to the unique viscosity or variational solution $u$ of the weighted $p$-Laplace equation: $$\mathrm{div}(\rho^2(x)|\nabla u|^{p-2}\nabla u)= 0 \quad \text{in } \Omega \setminus \mathcal{O},$$ subject to label Dirichlet data at $\mathcal{O}$ and homogeneous Neumann conditions on $\partial\Omega$ (Shi, 22 Jan 2026, Crook et al., 2019, Shi et al., 2024).

For hypergraph regularization, the corresponding continuum energy for the $\varepsilon_n$-ball case is

$$\mathcal{E}(u) = 2^p\int_\Omega |\nabla u|^p \rho(x)\,dx,$$

while $k_n$-NN constructions induce a density-weighted energy $\int_\Omega |\nabla u|^p \rho^{1-p/d}(x)\,dx$ (Shi et al., 2024).

3. Algorithmic Strategies and Numerical Schemes

Solving discrete $p$-Laplacian regularization problems on point clouds leads to large-scale, convex but often non-differentiable optimization. Two major classes of algorithms have been proposed:

• Stochastic primal-dual hybrid gradient (SPDHG): This approach solves the hypergraph $p$-Laplacian minimization by alternating updates of primal and dual variables, touching only a single hyperedge per iteration for scalability. The scheme exploits proximal mappings for the nonsmooth norms and enforces label constraints via projection (Shi et al., 2024).
• PDE-inspired methods: Alternatively, one may first estimate the sampling density via kernel or spline-based methods, then solve the continuum $p$-Laplacian PDE using spectral discretization (e.g., on Chebyshev grids), imposing Dirichlet constraints and updating via semi-implicit or gradient-flow schemes (Crook et al., 2019).

The table summarizes key algorithmic ingredients for point cloud $p$-Laplacian solvers:

Approach	Core Discretization	Label Handling
SPDHG (Shi et al., 2024)	Hypergraph, max norm	Projection/Prox
PDE-spectral (Crook et al., 2019)	Density + Chebyshev grid	Value clamping

4. Regularity, Stability, and Boundary Effects

For $p > d$, regularity results guarantee that minimizers are Hölder-continuous, ensuring stable propagation of labels without the formation of large spikes or artifacts near labeled points in the large-sample limit (Shi, 22 Jan 2026, Shi et al., 2024). Unlike lower-order ($p=2$) Laplacian regularization, which may form singularities or "spikes" around scarce labels as neighborhood size increases, hypergraph $p$-Laplacian penalization (based on the maximum difference in each neighborhood) enforces a more global smoothness, exhibiting Lipschitz-regularity properties inherited from the continuum theory.

Boundary treatment in these frameworks avoids artificial "ghost nodes" or padding; the Neumann condition arises naturally from the sampling geometry and analytic consistency arguments.

5. Empirical Performance and Applications

Numerical experiments on synthetic interpolation, image inpainting, and label propagation tasks demonstrate:

• For one-dimensional interpolation with few labels, the hypergraph $p$-Laplacian remains smooth and passes through labels as the neighborhood grows, in contrast to the graph $p$-Laplacian, which develops spikes at labeled nodes (Shi et al., 2024).
• In semi-supervised classification (such as MNIST), hypergraph $p$-Laplacian regularization significantly outperforms graph-based approaches at very low label rates, e.g., test accuracy of $40$-$70\%$ (hypergraph) vs. $15$-$30\%$ (graph) at $0.1\%$ label rates, with both converging for larger label proportions (Shi et al., 2024).
• For image inpainting on patch manifolds, hypergraph regularization improves peak signal-to-noise ratio (PSNR) by $0.3$-$1$ dB and structural similarity index (SSIM) by $0.05$-$0.1$ over graph approaches at all sampling rates (Shi et al., 2024).

These results point toward the advantage of higher-order, nonlocal regularization in data-scarce regimes, as well as the inheritability of continuum PDE regularity even in highly discrete settings.

6. Theoretical and Practical Implications

Recent advances provide rigorous justification for the use of $p$-Laplacian models—both graph-based and hypergraph-based—as discrete approximations to weighted $p$-Laplacian PDEs on the underlying data manifold. For $p \to \infty$, the schemes recover Lipschitz learning along with its optimality guarantees. The hypergraph construction enhances expressivity by encoding higher-order affinities, and the convergence theory (via $\Gamma$-convergence and viscosity solution arguments) extends to large-scale, sparse, and irregularly sampled data.

In manifold learning and semi-supervised contexts, these frameworks ensure well-posedness at very low sampling rates, resist label-spikes, and support scalable optimization. The choice of $p$ modulates the interpolation behavior: as $p$ increases, solutions become closer to piecewise-constant, and interfaces align with minimal-perimeter sets, as formalized for classification and clustering models incorporating $p$-Laplacian regularization (Cristoferi et al., 2018, Shi et al., 2024).

7. Extensions and Ongoing Directions

Extensions to anisotropic weights, phase transition models with nonlocal Ginzburg–Landau penalties, and density-weighted variants are well-established (Cristoferi et al., 2018). Optimal-transport-based frameworks provide additional flexibility in comparing discrete and continuum energies. Further research addresses fast solvers, adaptive neighborhood selection, and the role of $p$ in high-dimensional scaling. A plausible implication is that as sampling density increases and higher-order connectivity is exploited, the correspondence between point cloud $p$-Laplacian regularization and continuum geometric variational methods strengthens, supporting principled development of nonlinear, data-driven regularization schemes across learning and signal processing tasks (Shi, 22 Jan 2026, Shi et al., 2024, Crook et al., 2019).