Nonlocal-to-local $L^p$-convergence of convolution operators with singular, anisotropic kernels

This paper establishes and quantifies the strong $L^p$-convergence of nonlocal convolution operators with singular, anisotropic kernels to local differential operators with natural boundary conditions, extending previous results to include stronger singularities and providing explicit convergence rates.

The Gist

Imagine you are trying to understand how a crowd of people moves through a city.

The Old Way (Local Models):
Traditionally, mathematicians model this by looking at one person and asking, "How are my immediate neighbors pushing or pulling me?" This is like looking at a single brick in a wall and only caring about the bricks touching it. This is called a local model. It's simple, but it assumes you only interact with the people right next to you.

The New Way (Nonlocal Models):
But in reality, people might react to someone three blocks away, or a whole group shouting from a stadium. This is a nonlocal model. You are influenced by everyone in the city, but the influence gets weaker the further away they are.

The Problem:
Nonlocal models are great for describing complex, messy real-world physics (like how crystals grow or how materials crack), but they are incredibly hard to solve with computers. Local models are easy to solve.

So, scientists have a big question: "Can we prove that if we make the 'reach' of the nonlocal influence very, very small, it behaves exactly like the simple local model?"

If the answer is yes, we can use the messy nonlocal model to justify why the simple local model works in the real world.

What This Paper Does

This paper by Abels, Hurm, and Knopf is like a master architect proving that two very different blueprints actually describe the same building, even under some very tricky conditions.

Here is the breakdown of their breakthrough using simple analogies:

1. The "Blurry Lens" Analogy (Convergence)

Imagine you have a camera with a very wide-angle, blurry lens (the nonlocal operator). You are taking a picture of a sharp, crisp object (the local differential operator).

• The Goal: As you zoom in and sharpen the lens (making the parameter $\epsilon$ go to zero), the blurry picture should eventually look exactly like the sharp one.
• The Paper's Achievement: They proved that no matter how you look at the picture (using different mathematical "lenses" called $L^p$ spaces), the blurry image converges to the sharp one. They didn't just say "it gets close"; they calculated exactly how fast it gets there. It's like saying, "If you zoom in by 10%, the picture gets 10% sharper."

2. The "Heavy Metal" Analogy (Singular Kernels)

In previous studies, the "influence" between points was like a gentle breeze. If you got too close to the source, the breeze just got a little stronger.

• The New Twist: This paper allows the influence to be like a black hole or a singularity. As you get closer to the source, the force becomes infinite (mathematically speaking). This is much harder to handle because the numbers blow up.
• The Achievement: They proved that even with these "infinite force" interactions (similar to fractional Laplacians used in quantum physics), the math still holds up and converges to the local model.

3. The "Anisotropic" Analogy (Direction Matters)

Imagine a crowd moving through a hallway.

• Isotropic (Old assumption): The crowd pushes equally in all directions. It's like a perfect circle of influence.
• Anisotropic (New assumption): The hallway is narrow. People can only push easily forward and backward, but it's very hard to push sideways. The influence is stretched out like a football, not a circle.
• The Achievement: Most previous math assumed the influence was a perfect circle. This paper handles the "football" shape. They showed that even if the interactions are stretched and directional (anisotropic), the system still settles into a predictable local pattern, just with a different "shape" of the resulting equation.

4. The "Curved Room" Analogy (Boundaries)

Imagine trying to model water flowing in a perfectly round pool versus a pool with a jagged, curved edge.

• The Challenge: When you are near the edge of a curved room, the "nonlocal" influence hits the wall and bounces back in weird ways. It's much harder to prove the math works there than in an infinite open field.
• The Achievement: They proved this works even for complex, curved boundaries (like the surface of a crystal or a biological cell), not just flat, infinite spaces.

Why Does This Matter?

Think of this as the physical justification for the models engineers use every day.

• Before: Engineers used simple local equations (like the heat equation) because they were easy to calculate. They hoped it matched reality, but they couldn't always prove why it worked for complex materials.
• Now: This paper says, "We have taken the complex, messy, long-range physics (nonlocal), added in the real-world complications (singularities, weird shapes, directional forces), and mathematically proven that as you zoom in, it becomes the simple local equation."

In a nutshell:
They took a very complex, messy, and difficult-to-solve mathematical problem (nonlocal operators with singular, directional kernels) and proved that it reliably turns into a simple, solvable problem (local differential operators) when you look at it closely enough. They did this with a high level of precision, telling us exactly how fast the transition happens, which is crucial for building accurate computer simulations of the real world.

Technical Summary

1. Problem Statement

The paper investigates the asymptotic behavior of nonlocal convolution-type operators as the interaction kernel concentrates at the origin. Specifically, the authors study operators of the form:
$$L^\Omega_\varepsilon u(x) = \text{P.V.} \int_\Omega J_\varepsilon(x-y) \big(u(x) - u(y)\big) \, dy$$
where the kernel is defined as $J_\varepsilon(x) = \varepsilon^{-n} \rho(x/\varepsilon) / |x|^2$.

The primary objective is to rigorously establish and quantify the nonlocal-to-local convergence of these operators to a local second-order differential operator:
$$L^\Omega u(x) = -\text{div}(M \nabla u) = -\sum_{i,j=1}^n M_{ij} \partial_{x_i} \partial_{x_j} u(x)$$
where $M$ is a symmetric "momentum matrix" determined by the kernel's second moments. The convergence is studied in the strong $L^p$ topology ($1 \le p < \infty$) for functions in Sobolev spaces $W^{k,p}(\Omega)$, including explicit convergence rates.

Key Challenges Addressed:

• Singular Kernels: The kernels are allowed to have singularities at the origin comparable to fractional Laplacians (order $s \in (0,1)$), which are stronger than previously studied $W^{1,1}_{loc}$ kernels.
• Anisotropy: The kernel $\rho$ is not required to be radially symmetric, allowing for direction-dependent interactions (anisotropy).
• General Domains: The results apply to unbounded domains and domains with compact, curved boundaries, not just flat tori or bounded domains with simple geometry.
• Strong Convergence & Rates: Moving beyond weak convergence of energies or operators to prove strong convergence in $L^p$ with explicit error bounds.

2. Methodology and Assumptions

Assumptions on the Kernel

The authors impose specific conditions on the scaling function $\rho: \mathbb{R}^n \to [0, \infty)$:

1. Integrability & Symmetry: $\rho$ is even, non-negative, and satisfies $\int \rho(x)(1+|x|)dx < \infty$.
2. Singularity Control: $\rho$ behaves like $|x|^{2-\alpha-n}$ near the origin with $\alpha \in (0, 2)$, allowing for singularities similar to the fractional Laplacian $(-\Delta)^s$ (where $\alpha = 2s$).
3. Anisotropy Condition: A crucial technical condition (Assumption A2, Eq. 2.7) ensures that the kernel's interaction with the boundary vanishes in a specific way. This involves the matrix $A = \sqrt{M}$ and requires:
   $$\int_{\mathbb{R}^{n-1}} J\left(AQ \begin{pmatrix} z' \\ z_n \end{pmatrix}\right) z' \, dz' = 0$$
   for all rotations $Q$. This condition is essential for handling curved boundaries and anisotropic effects.

Mathematical Framework

• Function Spaces: The analysis is conducted in Sobolev spaces $W^{k,p}(\Omega)$. The local operator $L^\Omega$ is equipped with natural boundary conditions: $M \nabla u \cdot n_{\partial \Omega} = 0$ on $\partial \Omega$.
• Localization Strategy: For general domains $\Omega$, the authors use a partition of unity to decompose the problem into local patches. These patches are mapped to a "curved half-space" or the full space $\mathbb{R}^n$ via diffeomorphisms.
• Change of Variables: A linear transformation $x \mapsto Ax$ (where $A = \sqrt{M}$) is used to normalize the momentum matrix to the identity, reducing the anisotropic problem to an isotropic one on a transformed domain.

3. Key Contributions

1. Generalization to Anisotropic Kernels: Unlike previous works that assumed radial symmetry, this paper handles anisotropic kernels. This is critical for physical models like crystal growth where interactions depend on direction. The resulting local operator has a non-diagonal diffusion matrix $M$.
2. Strong $L^p$ Convergence: The paper proves strong convergence ($L^\Omega_\varepsilon u \to L^\Omega u$ in $L^p$) rather than just weak convergence. This is a significant step up from earlier results that relied on $\Gamma$-convergence of energy functionals.
3. Relaxed Regularity Requirements: The authors remove the requirement for the kernel to be in $W^{1,1}_{loc}$. They allow for stronger singularities (fractional Laplacian type), broadening the applicability to nonlocal models with heavy-tailed interactions.
4. Explicit Convergence Rates: The paper establishes quantitative error bounds:
   $$\| L^\Omega_\varepsilon u - L^\Omega u \|_{L^p(\Omega)} \le C \varepsilon^{\gamma(p)} \| u \|_{W^{3,p}(\Omega)}$$
   where the rate $\gamma(p)$ depends on the domain:
   • $\gamma(p) = 1$ for $\Omega = \mathbb{R}^n$ (full space).
   • $\gamma(p) = 1/2$ for bounded domains with compact boundaries.
   • The rate $1/2$ for bounded domains is a known limitation due to boundary layer effects in nonlocal-to-local limits.
5. Unbounded Domains: The results extend to exterior domains (e.g., $\mathbb{R}^n \setminus B_1(0)$), provided the boundary is sufficiently regular and compact.

4. Main Results

Theorem 3.3 (Full Space $\mathbb{R}^n$):
For $u \in W^{3,p}(\mathbb{R}^n)$, the nonlocal operator converges strongly to the local operator with rate $O(\varepsilon)$:
$$\| L^\Omega_\varepsilon u - L^\Omega u \|_{L^p} \le C \varepsilon \| u \|_{W^{3,p}}$$

Theorem 3.4 (Curved Half-Space):
For a domain defined by a graph $x_n > \gamma(x')$, under the anisotropy condition (A2), the operator converges with rate $O(\sqrt{\varepsilon})$:
$$\| L^\Omega_\varepsilon u + \Delta u \|_{L^p} \le C \sqrt{\varepsilon} \| u \|_{W^{3,p}}$$
(Note: Here $M=I$ is assumed for the half-space case, but the logic extends via transformation).

Theorem 3.6 (General Domain with Compact Boundary):
For a general domain $\Omega$ with $C^3$ boundary and anisotropic kernel satisfying (A1) and (A2):

1. The nonlocal operator extends to a bounded linear operator $L^\Omega_\varepsilon: W^{2,p}_M(\Omega) \to L^p(\Omega)$.
2. Strong convergence holds: $L^\Omega_\varepsilon u \to L^\Omega u$ in $L^p(\Omega)$.
3. The convergence rate is $O(\sqrt{\varepsilon})$ for $u \in W^{3,p}_M(\Omega)$.

5. Significance and Applications

• Physical Justification of Local Models: The results provide a rigorous mathematical foundation for deriving local partial differential equations (PDEs) from nonlocal microscopic laws. This is particularly important for models like the Cahn-Hilliard equation and Allen-Cahn equation, where the local version cannot be directly derived from microscopic principles without a nonlocal limit.
• Anisotropic Materials: By accommodating non-radial kernels, the theory supports the modeling of materials with directional properties (e.g., crystals, composite materials) where diffusion or interaction varies by direction.
• Numerical Analysis: The explicit convergence rates ($\sqrt{\varepsilon}$) are vital for the analysis of nonlocal numerical schemes. They allow researchers to estimate the error introduced when approximating local PDEs with nonlocal models (which are often easier to handle numerically due to the lack of boundary condition complexities in some formulations).
• Robustness: The ability to handle singular kernels comparable to fractional Laplacians makes these results applicable to a wider class of anomalous diffusion processes and long-range interaction models.

In summary, this paper represents a substantial advancement in the theory of nonlocal operators by unifying the treatment of singularities, anisotropy, and general domains, while providing the first strong $L^p$ convergence rates for such general settings.