Threefold thresholds (Q1882775)

It is natural to attach numerical invariants to a singular variety \(D\) which measure the complexity of the singularities. For example we can associate the multiplicity to a hypersurface singularity, but this is rather a coarse invariant (for example consider the infinite family \(y^2+x^b=0\), \(b\geq 2\), of plane curves of multiplicity two). One such invariant is the log canonical threshold. Suppose that \(D\) is embedded as a hypersurface in the smooth variety \(X\). The log canonical threshold \(c\) is then the largest value of \(c\) such that the log pair \((X, cD)\) has log canonical singularities. A similar definition pertains when \(X\) is log canonical and \(D\) is an integral \(\mathbb{Q}\)-Cartier divisor: if one picks a log resolution \(\pi:Y\to X\) and one writes the numerical equivalences \(K_Y+E\equiv\pi^*K_X+\sum a_iE_i\) and \(\tilde{D}+\sum b_iE_i\equiv\pi^*D\) for appropriate rational numbers \(a_{k}\) and \(b_{k}\), where \(E\) is the sum over all exceptional divisors \(E_{k}\) and \(\tilde D\) is the strict transform of \(D\), then \(c=\min_i \frac {a_i}{b_i}\). If \(D\) is given as \(y^2+x^b=0\), then the log canonical threshold \(c=1/2+1/b\). More generally, if \(D\) is a hypersurface of dimension \(n\) and multiplicity \(a\geq n\), then the log canonical threshold \(1/a\leq c\leq (n+1)/a\), so that the log canonical threshold is a subtle refinement of the multiplicity. In fact the log canonical threshold is a very natural invariant: when \(D\) is a hypersurface, then the log canonical threshold is both the complex singular index of \(D\) (a number that measures the asymptotic behaviour of the integrals of vanishing cycles) and (minus) the largest root of the (reduced) Bernstein-Sato polynomial of \(D\). Besides being a very natural invariant of study when \(D\) is a hypersurface in \(X\), the log canonical threshold plays an important role in inductive approaches to higher dimensional geometry. For example, one of the most important remaining issues in S.Mori's program is to establish the existence and termination of flips. After V.V.Shokurov's work on the problem of existence of four-fold flips, some of the focus of interest has turned to the problem of termination of flips; the log canonical threshold and other closely related invariants should play a crucial role in this problem. V.V.Shokurov conjectured that the set of log canonical thresholds in dimension \(n\) satisfies ACC (the ascending chain condition). For example, if \(D\) is the plane curve \(y^a+x^b=0\) then the log canonical threshold is \(1/a+1/b\) and it was proved by J.Igusa that the log canonical threshold of any irreducible plane curve has this form. Now given any set that satisfies ACC, it is natural to ask for the set of accumulation points. J.Kollár conjectured that the set of accumulation points in dimension \(n\) is equal to the set of accumulation points in dimension \(n-1\). The main result of the paper under review is to prove this conjecture in dimension three, namely the authors prove the the following: the set of accumulation points of the log canonical threshold in dimension three is equal to the set of log canonical thresholds in dimension two (excluding the number one). In fact the authors prove a similar result in all dimensions, assuming the existence of the minimal model Program and a conjecture of Alexeev--Borisov.