By bootstrap percolation we mean the following deterministic process on a graph $G$. Given a set $A$ of vertices infected at time 0, new vertices are subsequently infected, at each time step, if they have at least $rinmathbb{N}$ previously infected n
eighbors. When the set $A$ is chosen at random, the main aim is to determine the critical probability $p_c(G,r)$ at which percolation (infection of the entire graph) becomes likely to occur. This bootstrap process has been extensively studied on the $d$-dimensional grid $[n]^d$: with $2leq rleq d$ fixed, it was proved by Cerf and Cirillo (for $d=r=3$), and by Cerf and Manzo (in general), that [p_c([n]^d,r)=Thetabiggl(frac{1}{log_{(r-1)}n}biggr)^{d-r+1},] where $log_{(r)}$ is an $r$-times iterated logarithm. However, the exact threshold function is only known in the case $d=r=2$, where it was shown by Holroyd to be $(1+o(1))frac{pi^2}{18log n}$. In this paper we shall determine the exact threshold in the crucial case $d=r=3$, and lay the groundwork for solving the problem for all fixed $d$ and $r$.
