Well-Posedness and Numerical Approximation of a Class of Nonlocal Elliptic Equations with Gaussian Kernels

This paper investigates the mathematical properties and numerical approximation of a class of nonlocal elliptic partial differential equations of the form \begin{equation*} -Δu + λ, G(u) = f, \end{equation*} where $Δ$ denotes the Laplacian, $λ> 0$ is a regularization parameter, and $G$ is a nonlocal operator defined by integral convolution with a kernel $K$. We establish the well-posedness of the problem in the Sobolev space $H_0^1(Ω)$ using the Lax–Milgram theorem, providing rigorous proofs for the existence, uniqueness, and positivity of the weak solution under standard assumptions on the kernel $K$ and the source term $f \in L^2(Ω)$. For the numerical treatment, we employ a finite difference discretization for the Laplacian and a Gaussian-based approximation for the nonlocal term. We analyze a fixed-point iterative scheme for solving the discrete system and derive explicit conditions for its convergence and stability. Numerical experiments validate the theoretical results, demonstrating the monotonic decay of the residual and the robustness of the approximation scheme on bounded domains with various padding strategies.

💡 Research Summary

This paper addresses the mathematical analysis and numerical solution of a class of nonlocal elliptic partial differential equations of the form
    −Δu + λ G(u) = f,  z ∈ Ω,
with homogeneous Dirichlet boundary conditions. The nonlocal operator G is defined by convolution with a kernel K, and the authors focus on the Gaussian kernel, which is positive, symmetric, integrable, and essentially normalized on a sufficiently large bounded domain.

The analytical part begins by setting the problem in the Sobolev space H₀¹(Ω). Under the standard assumptions that K∈L¹(Ω×Ω), K(z,θ)=K(θ,z) a.e., and sup₍z₎∫ΩK(z,θ)dθ≤M, the operator G maps L²(Ω) into itself, is bounded with norm ≤M, and is self‑adjoint. Using these properties, the bilinear form
  a(u,v)=∫Ω∇u·∇v dx + λ∫ΩG(u)v dx
is shown to be continuous and coercive on H₀¹(Ω). Continuity follows from the Cauchy–Schwarz inequality together with the Poincaré inequality, while coercivity is obtained because the nonlocal term is non‑negative and the Laplacian part dominates via the Poincaré constant. Consequently, the Lax–Milgram theorem guarantees a unique weak solution u∈H₀¹(Ω) for any f∈L²(Ω) and any λ>0.

A nonlocal positivity principle is proved: if the source term f is non‑negative almost everywhere, then the weak solution is also non‑negative. The proof uses the standard test function v=−u₋ (the negative part of u) and exploits the positivity of the kernel to show that the integral involving G(u)u₋ is non‑positive, forcing u₋ to vanish. This result mirrors the classical maximum principle for local elliptic equations.

The paper then turns to the numerical discretization. The Laplacian is approximated by a standard five‑point finite‑difference stencil on a uniform Cartesian grid. The nonlocal term is discretized by evaluating the discrete convolution of the grid function with the Gaussian kernel; this can be performed either by direct summation or efficiently via the Fast Fourier Transform when periodic padding is used.

A key difficulty is handling the convolution near the physical boundary of Ω. To this end the authors introduce an extended domain eΩ⊃Ω and define linear padding operators E:L²(Ω)→L²(eΩ). Four padding strategies are described: (i) zero padding (Dirichlet), (ii) replicate (clamp) padding, (iii) even reflection (Neumann‑like), and (iv) periodic wrapping. The padded nonlocal operator is then G_pad(u)=R∘G∘E(u), where R restricts back to Ω. The choice of padding directly determines the effective boundary condition imposed on the nonlocal term.

For solving the resulting discrete nonlinear system, a fixed‑point iteration is proposed:
  u^{k+1}=A^{-1}(f−λ G_pad(u^{k})),
where A is the discrete Laplacian matrix (symmetric positive definite). Using the boundedness of G (‖G(u)−G(v)‖≤M‖u−v‖) and the Poincaré constant α (‖∇w‖≥α‖w‖_{L²}), the iteration map T(u)=A^{-1}(f−λ G_pad(u)) satisfies
  ‖T(u)−T(v)‖≤(λM/α)‖u−v‖.
Thus, if λM<α, T is a contraction on the discrete space, guaranteeing global convergence of the fixed‑point scheme. The authors also discuss a relaxed iteration u^{k+1}=(1−θ)u^{k}+θ T(u^{k}) with relaxation parameter θ∈(0,1]; the same contraction condition ensures convergence for any θ.

Numerical experiments are conducted on the unit square with various Gaussian bandwidths σ and the four padding strategies. The residual norm ‖r^{k}‖=‖A u^{k}+λ G_pad(u^{k})−f‖ decays monotonically and exhibits geometric convergence, confirming the theoretical analysis. Even‑reflection and periodic padding produce the smallest overall L² errors (≈10⁻⁴), while zero padding shows slightly larger boundary errors, illustrating the impact of boundary handling on nonlocal terms. Parameter studies varying λ and σ demonstrate that the convergence condition λM<α is sharp: when violated, the iteration stagnates or diverges.

In conclusion, the paper provides a rigorous well‑posedness framework for Gaussian‑kernel nonlocal elliptic equations, establishes a positivity principle, and delivers a practical, provably convergent finite‑difference scheme with explicit guidelines for boundary padding. The authors suggest extensions to nonlinear kernels, adaptive meshes, higher‑order discretizations, and multiscale kernels as promising directions for future work.