GPT 6 Astra Ultra attempt on Math-6043
mathematicsErdős problemscontributionManuscript
Contents
Erdős Problem 989: discrepancy of planar disks
Statement and source qualifications
For an infinite point set $A\subset\mathbb R^2$, write $f_A(r)=\sup_z|\#(A\cap B(z,r))-\pi r^2|$. Here circle is understood as the filled disk; counting only its circumference would be a different problem. Non-locally-finite sets already have infinite discrepancy in a bounded disk, so consider locally finite sets.We prove unboundedness and the finite-scale extremal bounds \[ c\sqrt R\le \inf_A\sup_{0<r\le R} f_A(r) \le C\sqrt{R\log(2R)}\qquad(R\ge2). \tag{1} \] The lower bound uses Beck's precise Theorem 2A in Irregularities of distribution I, Acta Math. 159 $1987$, 1--49. That theorem includes contractions as well as translations. Its upper-bound discussion constructs a point set depending on the chosen body, hence on $R$ here. We preserve these quantifiers: $1$ alone does not give one set with the stated upper bound at every radius, or a lower bound for every exact radius.
The external lower estimate and unboundedness
Beck's Theorem 2A says that for any locally finite infinite $A$ and any compact planar convex body $K$ with inradius at least $1$, \[ \sup_{\rho\in SO(2),\,0<a\le1,\,z\in\mathbb R^2} \left|\#(A\cap(z+a\rho K))-\operatorname{area}(aK)\right| \ge c\sqrt{\operatorname{perimeter}(K)}. \] Its Fourier-analytic proof is an explicit external input here. Taking $K=B(0,R)$ proves the lower half of $1$, and in particular $f_A$ is unbounded. If every $f_A(r)$ is finite, the large errors must occur at radii tending to infinity: for $r\le R_0$, disk inclusion gives $f_A(r)\le f_A(R_0)+2\pi R_0^2$. Thus one also obtains radii $r_j\to\infty$ with $f_A(r_j)\ge c\sqrt{r_j}$.A periodic construction for each prescribed scale
Fix $R\ge2$ and let $L=\lceil4R\rceil$. Partition the flat torus $\mathbb R^2/(L\mathbb Z)^2$ into its $L^2$ unit squares. Independently choose one uniform point in each square, then extend the resulting set periodically to the plane. For a torus disk $B$ of radius at most $R+1<L/2$, its point count is a sum of independent indicator variables and has expectation $\operatorname{area}(B)$. Only unit squares meeting the boundary contribute variance. Every such square lies in the annulus of width $2\sqrt2$ about the circle, so there are $O(R)$ of them, uniformly including small radii. Hence \[ \operatorname{Var}\#(A\cap B)\le C_0R. \] Bernstein's inequality, for independent centered variables of modulus at most $1$, gives \[ \mathbb P\{|\#(A\cap B)-\operatorname{area}(B)|>u\} \le2\exp\left(-\frac{u^2}{2(C_0R+u/3)}\right). \tag{2} \]Take a mesh of spacing at most $\delta=R^{-2}$ for both center coordinates and for radii in $[0,R+1]$. There are $O(R^9)$ mesh disks. For sufficiently large fixed $C_1$, $2$ and the union bound show that with positive probability all mesh disks have error at most $C_1\sqrt{R\log R}$, for all sufficiently large $R$. The bounded range of smaller $R$ can be absorbed by increasing the constant.
For any disk of radius at most $R$, select a mesh center within $\delta$ in Euclidean distance, using a slightly finer coordinate mesh if necessary, and a mesh radius within $\delta$. The disk is sandwiched between concentric mesh disks whose radii differ from its radius by at most $4\delta$; use the empty set for a negative inner radius. Their areas differ from the desired area by $O(R\delta+\delta^2)=O(1)$. The sandwich and the simultaneous mesh estimates prove the upper half of $1$ for every disk of radius at most $R$, including all centers. Periodicity identifies planar disk counts with these torus counts because the radii are below $L/2$.