---
title: Fractional Laplacian Overview
url: https://www.emergentmind.com/topics/fractional-laplacian
type: topic
---

# Fractional Laplacian Overview

The fractional Laplacian is a canonical example of a nonlocal pseudo-differential operator that generalizes the classical Laplacian $-\Delta$ to fractional (and, more generally, non-integer) exponents. It arises naturally in the theory of stable Lévy processes, anomalous diffusion, nonlocal PDEs, and the analysis of function spaces. Multiple equivalent formulations exist in the whole space, while in bounded domains the choice of definition affects both analytical and numerical properties, reflecting the fundamentally nonlocal character of the operator.

## 1. Definitions and Equivalent Constructions

Several equivalent definitions exist for the fractional Laplacian $(-\Delta)^s$ on $\mathbb{R}^n$, $0<s<1$:

- **Fourier Multiplier Definition:** For $u\in\mathcal{S}(\mathbb{R}^n)$,
  $$
  (-\Delta)^s u(x) = \mathcal{F}^{-1} \bigl[ |\xi|^{2s} \hat{u}(\xi) \bigr](x),
  $$
  where $\mathcal{F}$ denotes the Fourier transform. The operator is thus a pseudo-differential operator with symbol $|\xi|^{2s}$ [2310.19118, 2305.04251, 1801.09767].

- **Singular-Integral (Riesz) Definition:**
  $$
  (-\Delta)^s u(x) = C_{n,s}~ \text{P.V.} \int_{\mathbb{R}^n} \frac{u(x)-u(y)}{|x-y|^{n+2s}}\,dy,
  $$
  with normalization $C_{n,s} = \frac{2^{2s}s\,\Gamma(\frac{n+2s}{2})}{\pi^{n/2}\Gamma(1-s)}$ [2310.19118, 1801.09767, 1501.01942, 1103.3387]. The singularity at $x=y$ is handled in the Cauchy principal value sense.

- **Semigroup (Bochner/Heat Kernel) Definition:**
  $$
  (-\Delta)^s u(x) = \frac{1}{\Gamma(-s)} \int_0^\infty \big(e^{t\Delta}u(x) - u(x)\big)\, t^{-1-s}\,dt,
  $$
  relating the fractional Laplacian to a subordinated Brownian motion [2310.19118, 2305.04251].

- **Spectral Definition (on bounded domains):** If $u$ admits an expansion $u = \sum_k u_k \phi_k$ in terms of Dirichlet eigenfunctions $\phi_k$ with $-\Delta \phi_k = \lambda_k \phi_k$, then
  $$
  (-\Delta)^s_{\Omega} u = \sum_k \lambda_k^s u_k \phi_k,
  $$
  which is the spectral fractional Laplacian [2310.19118, 2010.01383, 1801.09767].

- **Mellin Transform Definition (for radially symmetric $u$):** For $u(x)=f(|x|)$,
  $$
  \mathcal{M}\{(-\Delta)^{\alpha/2} f\}(s) = -2^{\alpha} \frac{\Gamma(s)\Gamma\left(\frac{n - (s-\alpha)}{2}\right)}{\Gamma\left(\frac{n-s}{2}\right)} \mathcal{M}\{f\}(s-\alpha),
  $$
  with inverse Mellin representation [2305.04251].

All these are equivalent on $\mathcal{S}(\mathbb{R}^n)$ or sufficiently decaying and regular functions [2310.19118, 2305.04251].

## 2. Analytical Properties and Functional Framework

Key analytical properties follow directly from the above definitions:

- **Linearity, Self-Adjointness, Positivity:** $(-\Delta)^s$ is linear and self-adjoint on $L^2(\mathbb{R}^n)$, with spectrum $[0,\infty)$ [2310.19118, 1801.09767].
- **Scaling:** For $u_\lambda(x) = u(\lambda x)$,
  $$
  (-\Delta)^s u_\lambda(x) = \lambda^{2s} \left[(-\Delta)^s u\right](\lambda x)
  $$
  [2304.12419, 2310.19118].
- **Nonlocality:** $(-\Delta)^s u(x)$ depends on values of $u$ at points $y$ arbitrarily far from $x$; it encodes effects of long-range interactions not present in the classical $\Delta$ [1801.09767, 2310.19118].
- **Maximum Principle:** For $s\in(0,1)$ and suitable exterior sign conditions, $(-\Delta)^s$ satisfies a nonlocal strong maximum principle [2310.19118, 1801.09767].
- **Regularity Theory:** Solutions to $(-\Delta)^s u = f$ inherit regularity depending on $f$ and $s$. For $f\in C^\alpha$ one gains $u \in C^{2s+\alpha}$ up to the loss due to boundary layers and nonlocality [2310.19118, 2010.01383, 1801.09767].

## 3. Boundary Value Problems and Distinction Between Notions

On bounded domains $\Omega$, at least two fundamentally different notions of the fractional Laplacian emerge; this reflects different nonlocal boundary constraints [1801.09767, 2010.01383]:

- **Riesz (Integral) Fractional Laplacian:**
  $$
  (-\Delta)^s u(x) = C_{n,s}~ \text{P.V.} \left[\int_{\Omega} \frac{u(x)-u(y)}{|x-y|^{n+2s}} dy + \int_{\mathbb{R}^n\setminus\Omega} \frac{u(x)-g(y)}{|x-y|^{n+2s}} dy \right], \quad x\in\Omega
  $$
  with "volume constraint" $u=g$ in $\mathbb{R}^n\setminus\Omega$. This is the generator of the killed $\alpha$-stable process [1801.09767, 1303.6934].

- **Spectral Fractional Laplacian:**
  Defined via the eigenfunction expansion with homogeneous Dirichlet (or Neumann) boundary conditions on $\partial \Omega$ alone. For inhomogeneous boundary data, a harmonic lifting (i.e., $u = v + w$ with $v$ harmonic and matching $g$ at the boundary, $w|_{\partial\Omega} = 0$) can be used [2010.01383, 1801.09767].

These two definitions are *not* equivalent: solutions exhibit distinct interior and boundary layer behaviors, and obey different stochastic interpretations (killed jump process vs. subordinate stopped Brownian motion) [1801.09767, 2010.01383].

## 4. Generalizations: Variable-Order, Anisotropic, and Generalized Operators

- **Variable-Order Fractional Laplacian (VOFL):** For $s : \mathbb{R}^n \rightarrow (0, n/2)$, especially radial $s(x)$, the VOFL is constructed as the inverse of a space-dependent Riesz potential,
  $$
  (-\Delta)^{s(\cdot)} = I_{2s(\cdot)}^{-1}
  $$
  with $I_{2s(\cdot)} f(x) = \int_{\mathbb{R}^n} K_{s(\cdot)}(x-y) f(y)\, dy$ and
  $$
  K_{s(\cdot)}(x) = \Gamma\left(\frac{n}{2} - s(|x|)\right) / \big[ 4^{s(|x|)} \pi^{n/2} \Gamma(s(|x|)) \big]\, |x|^{-n+2s(|x|)}
  $$
  [2109.01060]. Analytical properties like linearity, invertibility, and rotation-invariance remain, but scaling holds only locally.

- **Spatially Variant Fractional Laplacian:** For $s(\cdot)$ measurable (possibly in $[0,1]$), the operator can be characterized variationally as a Dirichlet-to-Neumann map of a variable-weight degenerate extension over $\Omega\times(0,\infty)$. This allows the well-posed definition and trace regularity in weighted Sobolev/Besov spaces [2106.11471].

- **Generalized Fractional Laplacian in Nonhomogeneous Medium:** In $\mathbb{R}^2$, write
  $$
  (-\Delta)^{s} u = \nabla \cdot I_{2-2s} \nabla u
  $$
  where $I_{2-2s}$ is the Riesz potential operator. Generalization to a variable positive-definite matrix field $K(x)$ yields
  $$
  L_{s,K} u = \nabla \cdot \left[ K(x) I_{2-2s} (K(x) \nabla u) \right]
  $$
  suitable for modeling anomalous diffusion in nonhomogeneous media [2304.12419].

## 5. Numerical Methods: Discretization and Computational Algorithms

Several classes of numerical methods are effective for approximating the fractional Laplacian:

| Scheme                    | Domain           | Main Feature                | Typical Accuracy                  |
|---------------------------|------------------|-----------------------------|-----------------------------------|
| Finite difference/gl      | $\mathbb{R}$, grids | Discrete convolution, explicit weights | $O(h^{2-\alpha})$ to $O(h^{3-\alpha})$ [1311.7691, 1611.00164]  |
| Truncated quadrature      | $\mathbb{R}$, $\mathbb{R}^d$ | Directly approximates PV integral   | 2$^{\text{nd}}$ order [2212.05143, 1802.03770]  |
| FFT-based convolution     | $\mathbb{R}^d$ / periodic | Exploits translation invariance     | Fast (FFT), 2$^{\text{nd}}$ order [1802.03770, 2212.05143]  |
| Galerkin FEM/adaptive AFEM| Bounded domains  | Variational, dense/stiff matrices    | AFEM $O(N^{-1/2})$ [1801.09767]  |
| Walk-on-spheres           | Bounded domains  | Stochastic (Feynman-Kac), mesh-free | $O(1/\sqrt{M})$ MC error [1801.09767, 2205.05229] |
| Deep neural networks      | Bounded domains, high-dim | Stochastic representation, overcomes curse of dimensionality | $O(\varepsilon)$ $L^2$ error realized [2205.05229] |

Key implementation considerations:

- Far-field truncation, tail treatment (e.g., asymptotic or exact integration of nonlocal influences) [1311.7691, 1611.00164].
- Discrete maximum principle and monotonicity for nonlinear and obstacle problems [1311.7691].
- Preconditioning by local Laplacians can improve the conditioning for Krylov solvers in higher dimensions [1802.03770].
- Efficient Gauss–Legendre-based quadrature schemes (Duffy transform) handle singular double integrals in 3D FEM [2208.05714].
- High-dimensional neural network solvers leverage stochastic Feynman–Kac representations to bypass the curse of dimensionality [2205.05229].

## 6. Special Cases, Extensions, and Applications

- **Explicit Action on Power Functions:** Closed-form evaluation for $u_p(x) = (1 - |x|^2)_+^p$ in $\mathbb{R}^d$ in terms of hypergeometric functions enables spectral and variational analysis on spheres/balls (e.g., for computing eigenvalues) [1103.3387].
- **Polynomial-Growth Functions:** For functions of polynomial growth at infinity, the classical PV definition diverges; an extension is defined up to a polynomial ambiguity, with optimal Schauder-type estimates and Liouville theorems in equivalence classes modulo polynomials [1610.04663].
- **Nonlocal Discrete Models and Periodic Kernel Limit:** Discrete periodic fractional Laplacians, when properly scaled as lattice spacing $h\to0$, yield continuum operators with explicit $L$-periodic Riesz kernels, relevant for anomalous diffusion and fractional quantum mechanics [1412.5904].

Fractional Laplacians are central in modeling jump processes, anomalous transport, phase transitions, image processing, finance (option pricing under Lévy models), peridynamics, and geometric analysis of nonlocal curvatures [2310.19118, 1801.09767]. Variable- and anisotropic-order extensions model complex materials with spatial or directional heterogeneity [2109.01060, 2304.12419].

## 7. Open Problems and Methodological Perspective

- The non-uniqueness of the fractional Laplacian in bounded domains (Riesz-integral vs. spectral vs. regional) and the matching of physical boundary data remain active areas of research [1801.09767, 2010.01383].
- Regularity theory, boundary Harnack inequalities, and adaptive numerical error control for solutions near the boundary depend delicately on the operator’s definition and the order $s$ [2010.01383, 1801.09767, 2310.19118].
- The design of efficient fast solvers (e.g., multigrid preconditioners for dense nonlocal FEM matrices) and scalable meshfree or deep-learning based approaches for high-dimensional problems present ongoing challenges [1802.03770, 2205.05229, 2212.05143].

The fractional Laplacian’s multiple guises (Fourier, Riesz, spectral, Mellin, variable order, matrix-valued) and its intricate nonlocal character make it a unifying theme in analysis, probability, and applied mathematics [2310.19118, 2305.04251, 2109.01060].

Source: https://www.emergentmind.com/topics/fractional-laplacian