Sharp decay of the spherical averages of the Fourier transform of open sets
For a bounded domain $Ω\subset\mathbb R^d$, the spherical average $E_Ω(R)=\int_{\mathbb S^{d-1}}|\widehat{\mathbb{1}_Ω}(Rω)|^2dω$ is never $O(R^{-a})$ with $a>d+1$, and for smooth boundaries it is $O(R^{-d-1})$, as it is for the ball, whether or not the boundary is curved. In 2003, Brandolini, Hofmann and Iosevich extended this to every convex domain and to every domain with $C^{1,\frac{1}{2}}$ boundary, and this regularity condition has not been improved since. How much regularity does the decay $R^{-d-1}$ require? We show that among Hölder classes the answer is $C^{1,\frac{1}{4}}$. Every bounded domain with $C^{1,\frac{1}{4}}$ boundary satisfies $E_Ω(R)\leq C_ΩR^{-d-1}$, while for every $0<s<\frac{1}{4}$ any smooth domain can be altered near a single boundary point to produce a $C^{1,s}$ domain for which this bound fails along a sequence of radii. More generally, when the boundary has fractional Sobolev regularity $W^{1+s,p}$, we determine the optimal decay exponent under a Dini condition on the normal, and in a large range of $s$ and $p$ without it. The upper bounds follow from $L^2$ estimates for nondegenerate oscillatory integral operators whose phases are distances between points of two graphs with Dini or fractional Sobolev gradients, and hence barely better than $C^1$. These are far too rough for the integrations by parts techniques from the smooth theory. We replace derivatives by finite differences adapted to the phase, a device that we expect to be useful in a variety of related problems.