Late points for random walks in two dimensions

Let $\mathcal{T}_n(x)$ denote the time of first visit of a point $x$ on the lattice torus $\mathbb {Z}_n^2=\mathbb{Z}^2/n\mathbb{Z}^2$ by the simple random walk. The size of the set of $\alpha$, $n$-late points $\mathcal{L}_n(\alpha )=\{x\in \mathbb {Z}_n^2:\mathcal{T}_n(x)\geq \alpha \frac{4}{\pi}(n\log n)^2\}$ is approximately $n^{2(1-\alpha)}$, for $\alpha \in (0,1)$ [$\mathcal{L}_n(\alpha)$ is empty if $\alpha >1$ and $n$ is large enough]. These sets have interesting clustering and fractal properties: we show that for $\beta \in (0,1)$, a disc of radius $n^{\beta}$ centered at nonrandom $x$ typically contains about $n^{2\beta (1-\alpha /\beta ^2)}$ points from $\mathcal{L}_n(\alpha)$ (and is empty if $\beta <\sqrt{\alpha}$), whereas choosing the center $x$ of the disc uniformly in $\mathcal{L}_n(\alpha)$ boosts the typical number of $\alpha, n$-late points in it to $n^{2\beta (1-\alpha)}$. We also estimate the typical number of pairs of $\alpha$, $n$-late points within distance $n^{\beta}$ of each other; this typical number can be significantly smaller than the expected number of such pairs, calculated by Brummelhuis and Hilhorst [Phys. A 176 (1991) 387--408]. On the other hand, our results show that the number of ordered pairs of late points within distance $n^{\beta}$ of each other is larger than what one might predict by multiplying the total number of late points, by the number of late points in a disc of radius $n^{\beta}$ centered at a typical late point.