Nonlocal Neighbor Extension Methods

• Nonlocal neighbor extension methods are techniques that generalize interaction domains in integro-differential models to improve stability and numerical accuracy.
• They integrate harmonic extension, partition-of-unity, and algorithmic neighbor expansion to relate nonlocal operators to local PDE frameworks.
• Applications include imposing nonlocal boundary conditions in PDEs, enhancing graph convolutional networks, and optimizing particle method simulations.

Non-local neighbor extension methods are a class of analytical and computational techniques designed to extend, inflate, or otherwise generalize the neighborhood system or interaction domain in models governed by nonlocal or integro-differential operators. These methods arise in a variety of contexts, from harmonic and variational extensions associated with fractional and non-symmetric kernels to algorithmic neighbor list management and nonlocal boundary data extension, and are foundational both in theoretical PDE analysis and in large-scale simulations. Representative applications include the realization of non-local operators as boundary-to-Neumann maps, the construction of robust trace and extension operators for nonlocal Sobolev-type spaces, and the adaptive enlargement of interaction lists in particle and graph-based numerical methods.

1. Mathematical Foundations of Non-local Neighbor Extension

The prototypical analytic manifestation of a non-local neighbor extension is the harmonic extension framework, which realizes a large class of non-local operators as Dirichlet-to-Neumann (DtN) maps for suitable second-order elliptic equations. Specifically, for operators of the form

$$K f(x) = a_0 f''(x) + B f'(x) - \gamma f(x) + \int_{\mathbb{R}} \left(f(x+z) - f(x) - f'(x)z\,\mathbf{1}_{|z|<1}\right) \nu(z)\,dz,$$

with completely monotone jump kernels $\nu$, there exists a unique elliptic operator

$$L u(x, y) = a(dy)\partial_{xx}u + 2 b(y)\partial_{xy}u + \partial_{yy}u,$$

in the half-plane $\mathbb{R} \times (0, \infty)$ such that the Neumann trace $Kf(x) = \lim_{y \to 0^+} \partial_y u(x, y)$ corresponds precisely to $K$ (Kwaśnicki, 2019).

The specification of the interaction neighborhood—whether through non-spherical truncation, as with half-sphere nonlocal gradients, or via dynamically enlarged neighbor sets in computational particle methods—plays a pivotal role in determining stability, coercivity, operator range, and numerical accuracy (Lee et al., 2019, Checkaraou et al., 2022). The treatment of nonlocal Dirichlet or boundary value problems likewise relies on extending data from a prescribed region (boundary or exterior) into the domain, minimizing an appropriately nonlocal energy (Kassmann et al., 2016, Grube et al., 2023, Du et al., 2021).

2. Extension Mechanisms: Analytical and Computational Schemes

Non-local neighbor extension is implemented through several advanced schemes:

• Harmonic extension and spectral correspondence: A bijective correspondence between a non-local operator $K$ and an extension PDE is established, leveraging spectral theory (notably Kreĭn’s theory of strings) to link the jump kernel $\nu$ to measures $a(dy),b(y)$. This approach generalizes the classic Caffarelli–Silvestre extension for symmetric kernels, admitting non-symmetric Lévy generators and Rogers jump kernels (Kwaśnicki, 2019).
• Whitney decomposition and partition-of-unity: In trace and extension for nonlocal Dirichlet problems, boundary data $g$ on $\Omega^c$ or a finite-thickness collar $\Omega_\delta$ is extended into $\Omega$ by reflecting data through Whitney cubes and patching with smooth functions. The energy of the extension is controlled via discrete or continuous averaging, ensuring stability in the nonlocal setting (Kassmann et al., 2016, Grube et al., 2023, Du et al., 2021).
• Neighbor truncation and sectoral domains: For nonlocal gradients, truncating the full spherical horizon to a half-sphere (or more generally, a non-symmetric cone) removes zero modes and yields stable, coercive operators without requiring singular kernels near the origin. This geometric manipulation of the neighbor set is central to robust nonlocal PDE models (Lee et al., 2019).
• Algorithmic neighbor expansion: In computational methods (e.g., DEM/MD, GCNs), a dynamic or hierarchical expansion of the neighbor list (e.g., via local Verlet buffers or higher-hop adjacency in graphs) enhances efficiency and faithfully captures complex multi-scale, multi-physics phenomena (Checkaraou et al., 2022, Hashemi et al., 2023, Liu et al., 2020).

3. Theoretical Properties: Coercivity, Well-posedness, and Approximation

Non-local neighbor extension methods are characterized by rigorous mathematical properties:

• Coercivity and energy preservation: Analytical extensions (harmonic/Poisson-type) minimize the nonlocal Dirichlet energy and provide isometries between trace and energy spaces. The extension is unique and optimal in the sense of energy minimization (Bogdan et al., 2017).
• Well-posedness: Under mild geometric and kernel regularity assumptions, nonlocal Dirichlet or Neumann problems with neighbor-extended boundary data admit unique solutions, with extension operators bounded uniformly in relevant norms (Grube et al., 2023, Du et al., 2021, Lee et al., 2021).
• Asymptotic consistency: As the nonlocality parameter ($\delta$ or $s$) tends to zero or one, nonlocal neighbor extension operators converge (in norm) to their local PDE analogues, ensuring compatibility with classical Sobolev extension results (Grube et al., 2023, Du et al., 2021).
• Spectral stability under neighborhood geometry: Breaking spherical symmetry in the interaction domain substantially improves the spectral properties—e.g., low-frequency coercivity—without necessitating strong kernel singularities (Lee et al., 2019).
• Parameter efficiency and computational scalability: Algorithmic extensions typically incur linear (in number of hops, or particle types) computational cost increases, with efficient scaling realized by appropriate optimization of update frequencies and neighbor set sizes (Checkaraou et al., 2022, Hashemi et al., 2023).

4. Representative Applications

Non-local neighbor extension frameworks underpin methodologies across several domains:

• Nonlocal boundary condition imposition and PDEs: Poisson-type or Whitney-type extensions enforce boundary conditions in fractional Laplacian and general nonlocal operators, enabling robust Dirichlet/Neumann problem solutions under finite or infinite interaction range (Bogdan et al., 2017, Kassmann et al., 2016, Grube et al., 2023, Du et al., 2021).
• Nonlocal elasticity and continuum mechanics: Truncated neighbor domains in nonlocal gradient operators provide well-posedness and strong spectral guarantees for peridynamic, elasticity, and incompressible Stokes models, including stable nonlocal Helmholtz-like decompositions (Lee et al., 2019).
• Isogeometric analysis and numerical spline spaces: Stable extension of trimmed spline spaces via neighbor-based polynomial continuation and projection yields robust, convergent cut finite element and isogeometric methods for elliptic PDEs (Burman et al., 2022).
• Graph convolutional learning: Higher-order neighbor visiting (n-hop GCN) realizes enhanced node representation learning and robustness to label sparsity by aggregating information from distant graph neighbors, outperforming single-hop GCNs in low-data regimes (Hashemi et al., 2023).
• Particle methods and neighbor list management: Adaptive non-local neighbor extension via local Verlet buffer strategies dramatically accelerates DEM and DEM–CFD simulations by reducing neighbor list rebuild frequencies and responding to local particle kinematics (Checkaraou et al., 2022).
• Manifold learning and embedding: LNPE generalizes LLE by incorporating multi-hop neighbor propagation, improving manifold unfolding and topological faithfulness in unsupervised dimensionality reduction (Liu et al., 2020).
• Video processing: Memory-augmented non-local attention exploits nonlocal neighbor extension in both physical (frame) and learned (memory) space, enhancing video super-resolution under large motion and sparse neighbor information (Yu et al., 2021).

5. Implementation Strategies and Parameter Optimization

Effective deployment of non-local neighbor extension methods involves careful selection of geometric, algorithmic, or variational parameters:

• Geometric optimization: Selecting non-symmetric or problem-tailored interaction domains (half-spheres, cones, sectors) can remove pathological modes and adapt models to anisotropy or flow direction (Lee et al., 2019).
• Partitioning and projection: In discrete settings (splines, lattices, meshes), extension combines local neighbor-based polynomial continuation with global $L^2$-projection or averaging, ensuring stability and optimal approximation rates (Burman et al., 2022).
• Adaptive update criteria: For computational particle methods, per-particle skin margins and local flow regime-based neighbor list resizing are tuned via empirical optimization (e.g., parameter $K$ in the Verlet buffer) to minimize simulation time (Checkaraou et al., 2022). In typical cases, $K=200$ yields performance within 2% of the empirical optimum.
• Spectral parameter correspondence: In nonlocal-to-local operator calculus, one leverages correspondence theorems (e.g., Nevanlinna–Pick functions for the Fourier symbol) to guarantee existence and uniqueness of the extension mapping (Kwaśnicki, 2019).
• Scalability and memory: Trade-offs between depth (propagation steps, neighbor hops), run time, and accuracy must be balanced, with extensions such as shared weights and scalar hop-coefficients ensuring parameter efficiency (Hashemi et al., 2023).

6. Limitations, Open Problems, and Directions

Despite significant advances, non-local neighbor extension methodologies present open topics and subtle challenges:

• Optimal shape of interaction domain: Determining the best truncation or extension geometry for specific PDEs remains an open problem, particularly under anisotropic physical constraints (Lee et al., 2019).
• Domain-general coercivity proofs: Extending uniform spectral guarantees and coercivity bounds to non-periodic, arbitrary Lipschitz, or unbounded domains is an active area of research (Lee et al., 2019, Grube et al., 2023).
• Consistent discretization and coupling: Achieving discrete consistency and asymptotic compatibility across blended local/nonlocal regions and hybrid spatial-temporal domains is nontrivial (Lee et al., 2021).
• Parameter selection in data-driven settings: For graph-based and manifold learning approaches, criteria for optimal hop number, propagation depth, and aggregation weighting are largely empirical, and theoretical convergence rates remain to be clearly established (Hashemi et al., 2023, Liu et al., 2020).
• Scalability for multi-element systems: In multi-component or chemically complex systems, explicit multi-element neighbor extensions introduce cubic scaling in descriptor count, though force evaluation can often be reduced to quadratic scaling by careful algorithmic design (Cusentino et al., 2020).
• Uniform operator-norm convergence as parameters vary: Rigorous justification of the convergence of nonlocal extension operators to local analogs as $\delta \to 0$ or $s \to 1$ must ensure independence of constants and sharp handling of boundary layers and measure rescalings (Grube et al., 2023, Du et al., 2021).

7. Synthesis and Broader Impact

Non-local neighbor extension methods provide a unifying framework for the rigorous mathematical analysis, stable numerical approximation, and efficient computational implementation of nonlocal operators across pure and applied mathematics, computational physics, data science, and engineering. By generalizing the notion of neighborhood, either geometrically, analytically, or algorithmically, these approaches yield improved stability, flexibility, and compatibility with both local and nonlocal regimes, paving the way for robust modeling and simulation of multi-scale, multi-physics, and high-dimensional systems (Kwaśnicki, 2019, Kassmann et al., 2016, Lee et al., 2021, Grube et al., 2023, Checkaraou et al., 2022).