on the fractional obstacle problem

Overview

Caffarelli, Salsa, and Silvestre developed the regularity theory for the obstacle problem associated with the fractional Laplacian (Caffarelli, Salsa & Silvestre, 2007). Their key move is to replace a nonlocal obstacle problem in ${\mathbb{R}}^n$ by a local degenerate elliptic thin obstacle problem in ${\mathbb{R}}^{n+1}$.

This post explains the fractional Laplacian, the Caffarelli-Silvestre extension, optimal regularity, blow-up analysis, and the mechanism that proves regularity of the free boundary near regular points.

Classical Obstacle Geometry

In the classical obstacle problem, one minimizes the Dirichlet energy subject to the constraint

$$u\ge \phi .$$

The solution is harmonic where it lies strictly above the obstacle and is pinned to the obstacle on the contact set. The free boundary is

$$\partial \{u>\phi \}.$$

The central questions are regularity of $u$ and regularity of this free boundary.

For the fractional problem, the Laplacian is replaced by

$$(-\Delta {)}^s,0<s<1.$$

The operator is nonlocal:

$$(-\Delta {)}^{s}u(x)=c_{n,s}p.v.{\int }_{{\mathbb{R}}^n}\frac{u(x)-u(y)}{|x-y{|}^{n+2s}}dy.$$

Thus the value at $x$ depends on all values of $u$. Local PDE arguments cannot be applied directly.

Variational Inequality

A standard formulation is

$$\min \{(-\Delta {)}^{s}u,u-\phi \}=0$$

in an appropriate domain, with exterior or boundary data prescribed. Equivalently,

$$u\ge \phi ,(-\Delta {)}^{s}u\ge 0,(u-\phi )(-\Delta {)}^{s}u=0.$$

The first equation is the obstacle constraint. The second says the solution is superharmonic for the fractional operator. The third says the operator vanishes in the non-contact region.

The free boundary separates

$$\Lambda =\{u=\phi \}$$

from

$$\Omega =\{u>\phi \}.$$

At points of $\Omega$, the equation is nonlocal harmonicity. At points of $\Lambda$, the obstacle constraint dominates.

Extension to a Thin Obstacle Problem

The Caffarelli-Silvestre extension realizes $(-\Delta {)}^s$ as a Dirichlet-to-Neumann operator. Given $u$ on ${\mathbb{R}}^n$, let $U(x,y)$ solve

$$\mathrm{div}(y^a\nabla U)=0\text{for }y>0,U(x,0)=u(x),$$

where

$$a=1-2s.$$

Then

$$(-\Delta {)}^{s}u(x)=-c_s\underset{y\downarrow 0}{\lim }{y}^a{\partial }_{y}U(x,y).$$

The nonlocal problem in ${\mathbb{R}}^n$ becomes a local degenerate elliptic problem in the upper half-space, with the obstacle imposed only on the boundary plane $\{y=0\}$.

Thin obstacle

The word thin means that the constraint lives on a codimension-one set in the extended space. The equation is local in ${\mathbb{R}}_{+}^{n+1}$, but the contact condition is imposed only on ${\mathbb{R}}^n\times \{0\}$.

After subtracting a smooth extension of the obstacle, one studies a function satisfying a degenerate elliptic equation and Signorini-type boundary conditions:

$$U\ge 0,-\underset{y\downarrow 0}{\lim }{y}^a{\partial }_{y}U\ge 0,U\underset{y\downarrow 0}{\lim }{y}^a{\partial }_{y}U=0$$

on the thin boundary.

Main Regularity Results

The paper proves sharp regularity for the solution and regularity of the free boundary near regular points. In simplified form:

Fractional obstacle regularity

If the obstacle is sufficiently smooth, then the solution to the fractional obstacle problem has the optimal Hölder differentiability predicted by the order of $(-\Delta {)}^s$. Moreover, at regular free-boundary points the free boundary is a $C^{1,\alpha }$ hypersurface.

For the model zero-obstacle problem, regular points are characterized by blow-ups of homogeneity

$$1+s.$$

This exponent is the fractional analogue of the quadratic blow-up in the classical obstacle problem. It records how fast the solution separates from the obstacle near a regular contact point.

Almgren Frequency and Blow-Ups

The blow-up method rescales around a free-boundary point $x_0$:

$$U_r(X)=\frac{U(x_0+rX)}{r^{\kappa }}.$$

The correct exponent $\kappa$ is identified by a monotonicity formula. The Almgren frequency has the schematic form

$$N(r)=\frac{r{\int }_{B_r}{y}^a|\nabla U{|}^2}{{\int }_{\partial B_r}{y}^a{U}^2}.$$

For solutions of the thin obstacle problem, $N(r)$ is monotone nondecreasing. Its limit as $r\downarrow 0$ gives the homogeneity of blow-up limits.

At regular points, the limiting homogeneity is $1+s$, and the blow-up profile is one-dimensional after rotation. A model profile is

$$U_0(x,y)=c\mathrm{Re}(x_n+i|y|{)}^{1+s}$$

with the appropriate branch and weight convention. This profile is positive on one side of the thin boundary and zero on the other.

Proof: From Blow-Ups to Free-Boundary Smoothness

After subtracting the obstacle and passing to the extension, the analysis is local near a thin free-boundary point. The first input is a modified Almgren frequency. If

$$F_U(r)={\int }_{\partial B_r}{U}^2|y{|}^{a}d\sigma ,$$

then the classical frequency is essentially

$$r\frac{d}{dr}\log F_U(r).$$

In the obstacle problem one must allow lower-order errors coming from the obstacle and from localization. Caffarelli, Salsa, and Silvestre therefore use a corrected expression of the form

$${\Phi }_U(r)=(r+C_0{r}^2)\frac{d}{dr}\log \max \{F_U(r),r^{n+a+4}\},$$

which is monotone for small $r$ after the normalization of the obstacle. Its limit gives the homogeneity of every blow-up.

Compactness of the rescalings then gives global homogeneous thin-obstacle solutions. Convexity properties of tangential derivatives and the Signorini conditions restrict the lowest possible homogeneity at a nontrivial contact point to $1+s$. When the contact set of a blow-up has positive measure, the degree-$1+s$ case is forced and the contact set is a half-space. Up to rotation and multiplication by a positive constant, the regular profile is the one-dimensional solution whose tangential derivative has the form

$${\partial }_{e}U=c{\left(\sqrt{(x\cdot e{)}^2+y^2}-x\cdot e\right)}^s.$$

This classification is the point at which the exponent $1+s$ becomes geometric rather than formal.

Once a free-boundary point has this profile, nearby rescalings are close to a half-space solution. Directional derivatives nearly parallel to the regular normal are positive in a smaller ball, because they are positive for the model profile and the equation supplies compactness plus a comparison principle. This positivity traps the contact set between two close half-spaces, so the free boundary has an approximate normal.

The improvement-of-flatness step repeats the same argument after rescaling: if the contact set is trapped in a narrow slab at one scale, positivity of the appropriate directional derivatives traps it in a narrower slab at the next scale, after a small rotation. Iterating this scale improvement gives Hölder continuity of the normal vector and hence $C^{1,\alpha }$ regularity of the regular free boundary.

Role of the Extension Method

The extension method does more than remove the nonlocal integral. It restores the geometric toolkit of elliptic PDE: energy estimates, monotonicity formulas, compactness of blow-ups, boundary Harnack principles, and improvement of flatness.

The price is degeneracy. The weight

$$y^a$$

may vanish or blow up at the thin boundary. The analysis must therefore use weighted Sobolev spaces and weighted elliptic estimates. These weights are still structured enough to support the regularity theory.

Singular points

Not every free-boundary point is regular. Higher-homogeneity blow-ups can occur, producing singular strata. The main achievement here is not that all free boundaries are smooth, but that the regular set has a robust local structure and the solution has the sharp differentiability expected from the fractional operator.

Weighted Energy and Natural Function Spaces

The extension problem is variational. The natural energy is

$${\int }_{{\mathbb{R}}_{+}^{n+1}}{y}^a|\nabla U{|}^{2}dxdy,a=1-2s.$$

The trace space of this weighted Sobolev energy is the fractional Sobolev space $H^s({\mathbb{R}}^n)$. Thus the extension does not merely produce the correct operator; it identifies the correct energy class for the obstacle problem.

The weight $y^a$ belongs to the Muckenhoupt $A_2$ class for $a\in (-1,1)$, which is exactly the range corresponding to $s\in (0,1)$. This permits weighted Poincare inequalities, Caccioppoli estimates, and compactness arguments. In graduate-level terms, most elliptic estimates survive, but every integration by parts must remember the degeneracy or singularity of $y^a$ at $y=0$.

Regular and Singular Free-Boundary Points

The Almgren frequency separates free-boundary points by the homogeneity of their blow-ups. Regular points are those whose limiting homogeneity is $1+s$. At such points the solution has a one-sided expansion

$$U(X)=c{U}_0(RX)+o(|X{|}^{1+s})$$

for some rotation $R$ and $c>0$. This expansion gives a normal direction to the free boundary.

Singular points have higher homogeneity. They are not errors in the proof; they represent genuinely different local geometry. The regularity theorem is therefore local and conditional: once a point is known to have the minimal homogeneity, the improvement-of-flatness scheme promotes first-order flatness to $C^{1,\alpha }$ regularity.

Proof: Boundary Harnack Input

Near a regular point, the blow-up classification gives a preferred normal direction $e_n$. Tangential derivatives $U_{\tau }$ for directions close to $e_n$ solve the same weighted equation

$$\mathrm{div}(|y{|}^a\nabla U_{\tau })=0$$

away from the thin contact set, and they vanish continuously on the slit portion of the boundary. The comparison estimates first show nondegeneracy:

$$U_{\tau }(X)\ge c\mathrm{dist}(X,\Lambda {)}^{2s}$$

in the non-contact region for directions in a cone around the normal. This prevents the solution from flattening too quickly as one approaches the free boundary from the positive side.

Boundary Harnack is then applied in the slit domain obtained by removing the contact set from the thin space. Ratios such as

$$\frac{U_{\tau }}{U_{e_n}}$$

are Hölder continuous up to the boundary, because numerator and denominator are positive weighted-harmonic functions that vanish on the same boundary portion. These ratios encode the slope of the free boundary: if the graph normal changed discontinuously, the ratios of directional derivatives would oscillate. Hölder control of the ratios is therefore exactly the analytic form of Hölder control of the normal vector.

The point of the blow-up classification is to put the solution in a geometric regime where this boundary Harnack argument applies. Without the one-dimensional regular blow-up, the slit-domain comparison need not describe the local geometry.

Model Profile and Optimal Regularity

The regular blow-up profile explains the optimal differentiability. In two variables $(x_n,y)$, the function

$$\mathrm{Re}(x_n+i|y|{)}^{1+s}$$

is homogeneous of degree $1+s$, is nonnegative on the thin boundary on one side, and has the correct degenerate Neumann behavior on the other. Its trace on $y=0$ grows like

$$(x_n{)}_{+}^{1+s}.$$

Thus one should not expect the solution to be better than $C^{1,s}$ in general.

This model also explains the geometry of the regular free boundary. Near a regular point, after rescaling, the contact set resembles a half-space and the non-contact set resembles the complementary half-space. The proof of free-boundary smoothness is the quantitative version of this statement: closeness to the model at one scale improves at smaller scales.

Transferable Mechanisms

The essential preliminary reference is the Caffarelli-Silvestre extension, which realizes $(-\Delta {)}^s$ as a boundary operator for a degenerate elliptic equation in one higher dimension (Caffarelli & Silvestre, 2007). The obstacle paper uses this not as a formal trick but as a way to import local PDE tools into a nonlocal free-boundary problem.

The same strategy plays a role in other problems involving fractional diffusion, nonlocal minimal surfaces, and boundary regularity: localize the operator by adding a dimension, prove weighted estimates in the extended space, and then translate boundary information back to the original nonlocal equation.

Links

• on Moser iteration --- both posts concern regularity mechanisms for elliptic or parabolic equations, though the fractional obstacle problem uses free-boundary blow-ups rather than iteration alone.
• on Hardy inequalities --- the extension problem involves weighted energies whose boundary behavior is governed by scale-sensitive inequalities.
• on interpolation theorems --- fractional powers of elliptic operators are naturally tied to interpolation scales and nonlocal regularity.

References

🐻  Caffarelli, L. & Silvestre, L. 2007. An Extension Problem Related to the Fractional Laplacian. Communications in Partial Differential Equations 32(8), 1245–1260.

🐻  Caffarelli, L.A., Salsa, S. & Silvestre, L. 2007. Regularity Estimates for the Solution and the Free Boundary of the Obstacle Problem for the Fractional Laplacian. Inventiones Mathematicae 171(2), 425–461.