0
$\begingroup$

I feel the following fact has been used in many argument in algebraic geometry, but I was not be able to prove it or find the precise reference:

Let $X$ be a $\mathbb{Q}$-factorial variety with log canonical singularities, and $Z \subseteq X$ be a subvariety. Then the set $$\{E \mid E {\rm{~is~ an~ exceptional~ divisor~ of~ some~ resolution~ }}Y \to X, {\rm ~such ~ that~} f(E)=Z {\rm~and~} discrepancy(K_X, E) \leq 1 \}$$ is finite.

Any suggestion for the proof or references is welcome!!

$\endgroup$
7
  • 2
    $\begingroup$ Since these $E$ aren't in $X$, but in various possibilities for $Y$, I'd say there's a proper class of such $E$. Are you really looking for something like the set of valuations on the structure sheaf of $X$, centered at $Z$? $\endgroup$ Mar 31, 2015 at 2:06
  • $\begingroup$ But those aren't finite either, as iterated blowing up of a point in $\mathbb P^2$ shows. $\endgroup$
    – Will Sawin
    Mar 31, 2015 at 3:24
  • $\begingroup$ Will Sawin is certainly right, there are infinitely many such $E$. But if you put stronger conditions on the singularities of $X$ (like log terminal and $\mathbb{Q}$-Gorenstein) and stronger conditions on the $E$ (like log discrepancy $\leq 1$) then the set is indeed finite... Or do you mean to consider $E$ that appear on every resolution (not just some resolution?) $\endgroup$ Mar 31, 2015 at 4:02
  • $\begingroup$ Dear@KarlSchwede, I had edited the condition: I do mean a variety $X$ with reasonable singularities, and the exceptional divisor with discrepancy $\leq 1$. Why in this case such $E$ is finite? $\endgroup$
    – Li Yutong
    Mar 31, 2015 at 13:58
  • $\begingroup$ You really want log discrepancy $\leq 1$ (log discrepancy is discrepancy + 1). There are only finitely many because all such divisors must be extracted on any resolution of $X$. $\endgroup$
    – user47305
    Mar 31, 2015 at 14:25

1 Answer 1

3
$\begingroup$

Ok, let me show the the following which is well known to experts (although I'm not sure where the right reference is):

Proposition: If $(X, \Delta)$ is KLT (or PLT), then there are at most finitely many divisorial valuations corresponding to divisors $E$ on birational models of $X$ such that the discrepancy of $K_X + \Delta$ along $E$ is $\leq 0$.

(Recall that if $X' \to X$ is birational then the coefficient of $K_{X'} - \pi^*(K_X + \Delta)$ along a prime divisor $E$ is called the discrepancy of $K_X + \Delta$ along $E$).

Proof: First let $\pi : Y \to X$ be a log resolution of $(X, \Delta)$. Consider then the simple normal crossings pair $(Y, -K_Y + \pi^*(K_X + \Delta))$. This pair is probably non-effective, but it is SNC.

Let $Y' \to Y$ be the composition of a sequence of blowups along the strata of $(Y, -K_Y + \pi^*(K_X + \Delta))$ such that in $(Y', -K_{Y'} + \pi'^*(K_X + \Delta) )$ every divisor on $Y'$ with discrepancy $\leq 0$ is disjoint from every other such divisor.

(This is not so hard to see how to do, note if I have two divisors with discrepancy $-1 < a \leq 0$ that intersect, then the blowup of their intersection separates the two divisors and creates a new divisor with discrepancy strictly bigger than that of the two original divisors. Keep doing this until separation is achieved).

At this point, we have reduced to the case of a pair $(Y', \Delta' = -K_{Y'} + \pi'^*(K_X + \Delta) )$ where $Y'$ is smooth, $\Delta'$ is SNC (not necessarily effective) and $\Delta'$ has finitely many prime divisors with coefficient $\geq 0$, all of which are disjoint from each other. In fact, it is easy to see that any exceptional divisor over $(Y', \Delta')$ has discrepancy $> 0$. Hence we are done. $\blacksquare$

Remark: This is false for log canonical singularities as mentioned in the comments (ie, consider $(\mathbb{A}^2, \text{Div}(xy))$).

$\endgroup$
1
  • $\begingroup$ Dear Karl, I understand it now. Thank you very very much for your help!!! $\endgroup$
    – Li Yutong
    Mar 31, 2015 at 18:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.