Nonlocal Evans-Krylov theorem

The classical Evans-Krylov theorem says that convex or concave fully nonlinear elliptic equations have $C^{2,\alpha}$ (therefore classical) solutions. This type of equations can be written as a Hamilton-Jacobi-Bellman equation. \[ \sup_\beta a_{ij}^\beta \partial_{ij} u = f \] for a family of uniformly elliptic coefficients $a_{ij}^\alpha$.

A purely integro-differential version of this theorem says that solutions of an integro-differential Bellman equation of the form \[ \sup_\beta \int_{\R^n} (u(x+y) - u(x)) K_\beta (y) \mathrm d y = 0 \qquad \text{in } B_1\] are $C^{s+\alpha}(B_{1/2})$ (which implies that they are classical) if the kernels satisfy the following assumptions \begin{align*} \frac{(2-s)\lambda}{|y|^{n+s}} \leq K(y) &\leq \frac{(2-s)\Lambda}{|y|^{n+s}} && \text{uniform ellipticity of order $s$} \\ D^2 K(y) &\leq \frac{(2-s)\Lambda}{|y|^{n+s+2}} && \text{Decay of the tails of $K$ in $C^2$} \\ K(y) &= K(-y) && \text{symmetry} \end{align*}

The $C^{s+\alpha}$ estimate does not blow up as $s \to 2$. Thus, the result is a true generalization of Evans-Krylov theorem.

Note that the result is relevant only if $s>1$, otherwise it is a weaker result compared to the $C^{1,\alpha}$ estimates.

The hypothesis above are most probably not optimal. Most likely a similar estimate would hold for kernels with $C^\alpha$ dependence respect $x$. Unlike the $C^{1,\alpha}$ estimates, no variation of this result is known.