Non-Local Integral Operators

• Non-local Integral Operators are defined via integration against kernels that capture extended spatial or spatio-temporal interactions, generalizing classical differential operators.
• They include models like fractional Laplacians and anisotropic operators, which are pivotal in PDE analysis, probability, and mathematical physics.
• Their study reveals rich spectral, regularity, and variational structures, providing insights for both theoretical analysis and numerical applications.

A non-local integral operator is an operator defined by integration against a kernel, typically representing spatial or spatio-temporal interactions that are not limited to infinitesimal neighborhoods but may have finite or even infinite range. Such operators generalize differential operators and arise naturally in analysis, PDE, probability, stochastic processes, mathematical physics, and numerical methods. Their foundational theory encompasses diverse classes—bounded or singular, symmetric or non-symmetric, linear or nonlinear, with anisotropic or variable coefficients—leading to rich functional-analytic, spectral, and variational structures.

1. Core Definitions and Classes of Non-Local Integral Operators

A prototypical non-local integral operator takes the form

$$\mathcal{L}u(x) = \int_{\Omega} (u(y)-u(x))\,K(x,y)\,dy$$

where $\Omega \subset \mathbb{R}^d$, $K(x,y)$ is the kernel, and the domain, regularity, and properties of $K$ determine the analytic properties and physical interpretation of $\mathcal{L}$ (Foss et al., 2024). Special cases include:

• Fractional Laplacians: $K(x,y) \sim |x-y|^{-d-\alpha}$ for $0<\alpha<2$, as the generator of symmetric $\alpha$-stable Lévy processes (Bogdan et al., 2011, Bañuelos et al., 2012).
• Anisotropic/fractional Laplacians: Sums or convex combinations of fractional powers along coordinate axes with possibly different orders (Chaker et al., 2018).
• General Lévy-type operators: Polynomials or mixtures of translation-invariant or spatially varying jump kernels (Kwaśnicki, 2019, Bogdan et al., 2017).
• Nonlinear nonlocal operators: Operators defined via variational energies with general convex or Bregman-type potentials, appearing e.g., in image processing, phase transitions, or metric measure geometry (Caffarelli et al., 2010, Bogdan et al., 2020).

Unlike classical differential operators, the lack of localization in the kernel may induce fundamentally new regularity, spectral, and boundary phenomena, especially when the kernel is only integrable, weakly singular, or highly anisotropic.

2. Operator-Theoretic Foundations and Nonlocal–to–Local Limits

For a class of operators with integrable kernels (compact support, possible asymmetry), the basic analytical properties are:

• Boundedness: The operator $\mathcal{L}$ maps $L^p$ to $L^p$ with

$$\|\mathcal{L}u\|_{L^p(\Omega)} \leq 2\|\mu\|_{L^1}\|u\|_{L^p(\Omega_\delta)},$$

where $\mu$ is the kernel density, for any $1\leq p \leq \infty$ (Foss et al., 2024).

• Compactness: $\mathcal{L}$ is compact on $L^p$ if and only if the kernel is mean-free, i.e., $\int \mu(z) dz = 0$. Otherwise, the operator includes a noncompact multiplication component (Foss et al., 2024).
• Integration by parts/adjoint: For arbitrary (not necessarily symmetric) $\mu$, the adjoint involves dual kernels and a possible singular part; for symmetric kernels the operator is self-adjoint and positive definite (Foss et al., 2024, Bogdan et al., 2017).
• Nonlocal-to-local convergence: For families of kernels with vanishing range and proper scaling of moments, the nonlocal operator $\mathcal{L}_\delta$ converges (in operator norm or weakly) to a first-order differential operator, e.g., a directional derivative, as $\delta \to 0$ (Foss et al., 2024, Shankar, 2015).

In the singular case, as for fractional Laplacians, the nonlocal operator recovers classical differential operators (Laplacian, divergence, curl) in the limiting sense under proper scaling and truncation (Bogdan et al., 2011, Ren et al., 2018).

3. Spectral Theory, Regularity, and Variational Structure

Non-local operators may be viewed as the generators of symmetric or non-symmetric jump processes, with deep links to Markov processes and Dirichlet forms. Fundamental results include:

• Spectral properties: For compact, self-adjoint kernel operators associated with phase-field models (e.g., linearizations of nonlocal Cahn–Hilliard equations), the spectrum contains:
  • A simple minimal eigenvalue $\lambda_0(L)$ decaying exponentially in domain size, governing slow modes (e.g., front translations).
  • Uniform spectral gap $\lambda_1(L)-\lambda_0(L) \geq D>0$ established via a nonlocal Cheeger-type inequality, crucial for scale separation and invariant manifold construction in interface problems (Orlandi et al., 2014).
• Regularity theory: Nonlocal equations with singular, anisotropic kernels satisfying mild comparability, symmetry, and integrability conditions yield:
  • Weak Harnack inequalities
  • Hölder regularity of weak solutions, by extensions of De Giorgi–Nash–Moser iteration to integral forms (Chaker et al., 2018, Caffarelli et al., 2010).
• Functional and variational setups:
  • Quadratic or Bregman-type energy forms define nonlocal Sobolev or Sobolev–Bregman spaces, supporting optimal extension, trace, and Douglas identities analogous to local theory (Bogdan et al., 2017, Bogdan et al., 2020).
  • Hard nonlocal Poincaré inequalities hold under kernel, domain, and zero set constraints, providing coercivity and well-posedness of the associated variational problems—even for integrable, possibly nonhomogeneous or lower-dimensional interactions (Foss, 2019, Foss et al., 2024, Scott et al., 2023).

4. Boundary Value Problems and Nonlocal–to–Local Correspondence

Well-posedness for nonlocal PDEs with local or nonlocal boundary conditions is achieved through structures that merge the variational and operator-theoretic frameworks:

• Green’s identities and fluxes: Precise nonlocal Green's identities relate the bilinear (energy) form of the operator to a boundary term that, in the vanishing-horizon limit, recovers the classical normal flux (Scott et al., 2023, Scott et al., 2023). This underpins the correct imposition of Dirichlet, Neumann, or Robin-type boundary data.
• **Trace and extension theorems