Caffarelli–Silvestre Extension & Fractional Laplacians

• The Caffarelli–Silvestre extension is a technique that reinterprets nonlocal fractional Laplacians as boundary maps for local degenerate PDEs.
• It connects analytic, probabilistic, and functional calculus approaches through Dirichlet-to-Neumann and Dirichlet-to-Wentzell methods.
• Its applications include regularity theory, inverse problems, and embedding theorems, leveraging variational principles and semigroup subordination.

The Caffarelli–Silvestre extension is a central technique that realizes nonlocal fractional Laplacians and their functional calculus as boundary operators for degenerate elliptic problems in one higher dimension. It allows reinterpretation of fractional and Bernstein-function powers of a broad class of linear operators as Dirichlet-to-Neumann (D-to-N) or more general boundary maps for local PDEs with singular weights. This yields a powerful bridge between nonlocal analysis, probability (via excursion and hitting-time processes), PDE theory, and semigroup calculus, and has far-reaching applications in regularity, unique continuation, trace inequalities, theoretical physics, and inverse problems.

1. Analytic Foundation: Classical Extension and Dirichlet-to-Neumann Principle

The classical Caffarelli–Silvestre extension for $(-\Delta)^s$, $s \in (0,1)$, is formulated as follows: Given $f \in H^s(\mathbb{R}^n)$, find $U: \mathbb{R}^n \times [0, \infty) \to \mathbb{R}$ solving

$$\operatorname{div}(y^{1-2s} \nabla U) = 0,\quad (x, y) \in \mathbb{R}^{n} \times (0, \infty), \qquad U(x,0) = f(x).$$

The fractional Laplacian appears as the D-to-N boundary map,

$$(-\Delta)^s f(x) = -C_{n,s} \lim_{y \to 0^+} y^{1-2s} \partial_y U(x,y),$$

with $C_{n,s} = 2^{2s-1} \Gamma(s)/\Gamma(1-s)$ (Yang, 2013, Li et al., 2020, Li et al., 2022). The solution admits the Poisson kernel representation

$$U(x,y) = c_{n,s} \int_{\mathbb{R}^n} \frac{y^{2s}}{(|x-z|^2 + y^2)^{(n+2s)/2}}\, f(z)\, dz,$$

and a variational structure as the unique minimizer of the Dirichlet energy in the weighted Sobolev space $H^1(\mathbb{R}^{n+1}_+, y^{1-2s})$ (Yang, 2013, Petrosyan et al., 2014).

2. Generalized Framework: Banach Space, Abstract Operators, and Bernstein Calculus

The extension paradigm extends to closed, densely defined $m$-accretive operators $A: D(A) \subset X \to X$ on Banach spaces. Given a nondecreasing string measure $m(z)$, the abstract extension problem is

$(A + B_m) u = 0,\quad z > 0, \qquad u(0) = f,$

where $B_m u = \frac{d}{dm} \left( \frac{du}{dz} \right)$ (Hauer et al., 2021). The associated boundary operator—the Dirichlet-to-Wentzell map—is

$$A_m f = m(0^+) A(f) - \lim_{z\to0^+}\frac{du}{dm}(z).$$

Specializing $m_s(z) = z^{2s}/(2s)$ recovers the classical Caffarelli–Silvestre extension, and in full generality, the boundary operator $A_m$ coincides with $\psi(A)$ for any $\psi$ a complete Bernstein function, with the correspondence achieved via Kreĭn's string theory. This establishes a boundary-value realization of the full class of subordinate semigroups, connecting directly to Phillips' subordination theorem: $$e^{-t\psi(A)}f = \int_0^\infty e^{-sA}f\, \mu_t(ds),$$ where $\mu_t$ is the convolution semigroup associated with $\psi$ (Hauer et al., 2021, Galé et al., 2012).

3. Probabilistic Approach: First Hitting Times and Excursion Theory

The extension technique admits a deeply probabilistic interpretation through excursion theory and the analysis of first hitting times of Bessel-type processes. The solution of the extension problem for $A$ can be represented by

$u(z) = \mathbb{E}_x \left[ f(X_{T_z}) \right],$

where $T_z$ is the first hitting time of the $m$-diffusion to $z=0$ (Hauer et al., 2021, Cavina, 2023). The density of first hitting times translates into the Poisson kernel structure of solutions and enables explicit stochastic integral representations: $u(x,z) = \int_0^\infty e^{-tA} f(x)\, w(t,z)\, dt,$ where $w(t,z)$ encodes the law of the first hitting time of the Bessel process parameterized by the string $m$. For the fractional Laplacian, $w_s(t, y)$ is the density of the first hitting time of a $2(1-s)$-Bessel process. This framework also integrates local time inverses for subordinators and allows close connection between PDE, semigroup theory, and probabilistic processes (Hauer et al., 2021, Cavina, 2023).

4. Functional Calculus, Subordination, and Operator Characterization

The extension technique provides an explicit realization of the full Phillips-Bochner functional calculus for complete Bernstein functions. If $-\!A$ generates a $C_0$-contraction semigroup and $\psi$ is a complete Bernstein function, then

$$\psi(A)f = m(0^+)Af - \int_0^\infty (e^{-sA}f - f)\, \nu(ds),$$

where the measure $\nu$ is determined by the Lévy–Khintchine formula of $\psi$. The semigroup $e^{-t\psi(A)}$ admits the probabilistic representation

$e^{-t\psi(A)} f = \mathbb{E}[e^{-L_t^{-1} A}f],$

where $L_t^{-1}$ is the right-continuous inverse of the local time for the $m$-diffusion, thus linking excursion theory and operator calculus (Hauer et al., 2021). The Dirichlet-to-Wentzell boundary operator associated to the extension uniquely determines $\psi(A)$ and provides a PDE-based proof of Phillips' subordination result for the entire class of Bernstein functions.

5. Uniqueness, Regularity, and the Core of the Extension Problem

In the abstract setting, uniqueness of weak bounded solutions to the extension problem is established via convexity arguments in Banach spaces and the accretivity property of $A$. If $u(z)$ is bounded and $z \mapsto \|u(z)\|^2$ is convex, then the only bounded convex function is constant, implying that the solution is uniquely determined by its boundary value $u(0)=f$. The domain $D(A)$ forms an operator core for both $\psi(A)$ and the associated Dirichlet-to-Wentzell operator, with the semigroup generated by $\psi(A)$ matching that induced by the boundary operator (Hauer et al., 2021).

6. Applications and Connections: PDE, Nonlocal Operators, and Inverse Problems

The extension framework underpins analytic and numerical approaches in a wide array of applications:

• Fractional and nonlocal PDEs: Implementation of nonlocal equations as local degenerate PDEs, regularity theory, and unique continuation (Yang, 2013, Petrosyan et al., 2014).
• Spectral theory and inverse problems: Direct recovery of source-to-solution maps in the anisotropic Calderón problem, equivalence of Poisson, heat, and wave kernel representations, and reduction of the nonlocal inverse problem to a local one (Rüland, 2023, Covi et al., 2023).
• Stochastic processes: Subordination, local times, excursion theory, and representations of nonlocal fields via brane and bulk correlators (Frassino et al., 2019).
• Functional inequalities, embedding theorems: Use of the extension for trace inequalities, capacity characterization, and embedding of function spaces (Li et al., 2020, Li et al., 2022).

This extension technique simultaneously clarifies and unifies the analytic, functional-analytic, and probabilistic aspects underpinning nonlocal operators and their powers.

7. Summary Table: Extension Problem and Operator Calculus

Setting	Extension PDE	Boundary Map	Associated Operator
$(-\Delta)^s$, $s \in (0,1)$	$-\Delta u - z^{1-2s}\partial_z^2 u=0$	$-\lim_{z\to0^+} z^{1-2s}\partial_z u$	$(-\Delta)^s$
$A$ $m$-accretive	$(A + B_m)u = 0$	$A_m f= m(0^+)A(u_0) - \lim_{z\to0^+} \frac{du}{dm}$	$\psi(A)$ (CBF)
General $\psi(A)$ (CBF)	Extension with string $m$ linked to $\psi$	$A_m$	$\psi(A)$

These correspondences concretely realize a rich nonlocal functional calculus as boundary maps of weighted local PDEs, providing both analytic and probabilistic constructions (Hauer et al., 2021, Galé et al., 2012, Li et al., 2020).