The Dirichlet problem for nonlocal operators with singular kernels: convex and nonconvex domains

TL;DR

This paper investigates the interior regularity of solutions to a class of nonlocal Dirichlet problems involving anisotropic fractional operators, establishing new regularity results in convex and nonconvex domains with singular kernels.

Contribution

It proves that solutions are $C^{1+3s- ext{epsilon}}$ inside convex domains for general measures, extending known regularity results to more singular kernels and less regular domains.

Findings

Solutions are $C^{1+3s- ext{epsilon}}$ in convex domains for general measures.

Solutions are $C^{1,1}$ in all $C^{1,1}$ domains when $a ext{ is } L^ ext{infinity}.

Counterexamples show regularity fails in non-convex domains with singular measures.

Abstract

We study the interior regularity of solutions to the Dirichlet problem $Lu=g$ in $\Omega$, $u=0$ in $\R^n\setminus\Omega$, for anisotropic operators of fractional type $$ Lu(x)= \int_{0}^{+\infty}\,d\rho \int_{S^{n-1}}\,da(\omega)\, \frac{ 2u(x)-u(x+\rho\omega)-u(x-\rho\omega)}{\rho^{1+2s}}.$$ Here, $a$ is any measure on~$S^{n-1}$ (a prototype example for~$L$ is given by the sum of one-dimensional fractional Laplacians in fixed, given directions). When $a\in C^\infty(S^{n-1})$ and $g$ is $C^\infty(\Omega)$, solutions are known to be $C^\infty$ inside~$\Omega$ (but not up to the boundary). However, when $a$ is a general measure, or even when $a$ is $L^\infty(S^{n-1})$, solutions are only known to be $C^{3s}$ inside $\Omega$. We prove here that, for general measures $a$, solutions are $C^{1+3s-\epsilon}$ inside $\Omega$ for all $\epsilon>0$ whenever $\Omega$ is convex. When $a\in L^{\infty}(S^{n-1})$, we show that the same holds in all $C^{1,1}$ domains. In particular, solutions always possess a classical first derivative. The assumptions on the domain are sharp, since if the domain is not convex and the spectral measure is singular, we construct an explicit counterexample for which $u$ is \emph{not} $C^{3s+\epsilon}$ for any $\epsilon>0$ -- even if $g$ and $\Omega$ are $C^\infty$.