# when a section descends?

Let $C$ be a (reduced, possibly reducible, complex) projective singular curve. Let $\nu: C'\to C$ a finite surjective birational morphism. (For example the normalization, but could be some intermediate modification.) Let $F_C$ be a torsion free sheaf on $C$. Pull it back: $F_{C'}:=\nu^*(F_C)/Torsion$.

Suppose $F_{C'}$ has global sections. Let $0\neq s\in H^0(F_{C'})$. When does $s$ descend to $C$, i.e. when is $s$ the pullback of a section of $F_C$?

Or, what is the condition that no global section of $F_{C'}$ comes from a global section of $F_C$? (Of course $h^0(F_{C'})=0$ is enough but probably is not necessary.)

Here is some local-to-global issue. Let $0\neq s\in H^0(F_{C'})$, such that for each point $pt\in C'$ this $s$ belongs to the preimage (not the pullback!) of the stalk of $F_C$, i.e.: $\nu^{-1}(F_{C,\nu(pt)})$. So, at each point of $C'$ the section $s$ "descends locally". Does this imply that $s$ descends globally?

What are the tools to treat such questions? Some sort of relative C'/C cohomology?

upd. According to Sándor Kovács' reply, it seems from $0\to F_C\to\nu_*(F_{C'})\to Skyscraper\to0$ that a global section of $F_{C'}$ descends to $F_C$ iff it descends locally at each point. The main question remains:

How to characterize the torsion free sheaves $F_{C'}$, pullbacks from $F_C$, whose global sections do not descend to $F_C$. (I'd like the answer in terms of $h^i(F_{C'})$, the genera of $C,C'$, the invariants of $\nu_* \mathcal{O}_{C'}/\mathcal{O}_C$, e.g. the lengths.)

Another related issue. If the conclusion is true, it seems some cohomology here vanishes (i.e. the local descent implies the global one). "Could it not vanish?" i.e. is there some similar situation (higher dimensional, etc) where a section descends locally at each point but not globally?

-

First of all, I think that all of $\mathscr F_C$'s global sections appear as global sections of $\mathscr F_{C'}$, so the answer to your second question is that it only happens in that trivial case you're mentioning.

Observe that $\nu$ is an isomorphism outside of finitely many points, so the induced morphism $\mathscr F_C\to \nu_*\mathscr F_{C'}$ is injective with a torsion cokernel. In other words, we have a short exact sequence: $$0 \to \mathscr F_C \to \nu_*\mathscr F_{C'} \to \mathscr Q \to 0.$$ where $\mathscr Q$ is torsion. This shows which sections of $\nu_*\mathscr F_{C'}$ do not come from $\mathscr F_C$: the ones that map to something non-zero in $\mathscr Q$. It also shows that every global section of $\mathscr F_C$ appear as a global section of $\mathscr F_{C'}$. Of course, this is not very helpful since it is hard to compute $\mathscr Q$ or what sections map to non-zero sections there.

A somewhat different approach is this: Let $\mathscr T$ denote the torsion subsheaf of $\nu^*\mathscr F_C$, so one has a short exact sequence: $$0\to \mathscr T \to \nu^*\mathscr F_C \to \mathscr F_{C'} \to 0.$$ Notice that the sheaf on the left hand side is torsion so it has no $H^1$. Therefore one obtains a short exact sequence of global sections: $$0\to H^0(C',\mathscr T) \to H^0(C',\nu^*\mathscr F_C) \to H^0(C',\mathscr F_{C'}) \to 0.$$ In particular all the global sections of $\mathscr F_{C'}$ come from those of $\nu^*\mathscr F_C$.

Next apply $\nu_*$ to the first short exact sequence and notice that $\nu$ is finite and hence $\nu_*$ is exact: $$0\to \nu_*\mathscr T \to \nu_*\nu^*\mathscr F_C \to \nu_*\mathscr F_{C'} \to 0.$$ Also notice that the global sections of these sheaves are the same as the global sections of the ones they are the push forwards of.

Finally, notice that the natural morphism $\mathscr F_C\to \nu_*\nu^*\mathscr F_C$ is injective by the torsion-free assumption.

So, all the global sections of $\mathscr F_{C'}$ come from $\nu^*\mathscr F_C$. We throw out the torsion sections and then the ones you are looking for are the global sections of $\nu_*\nu^*\mathscr F_C$ that are not sections of the subsheaf generated by torsion plus the global sections of the original $\mathscr F_C$.

I guess this is still not a very good answer to your first question, but perhaps the moral is that you should look at all of these sheaves on $C$.

-
Dear Sándor, in the exact sequence containing $H^0$'s, I think the last group should be $H^0(C',\mathscr F_{C'})$, with a prime on the $C$. –  Georges Elencwajg Feb 8 '11 at 12:49
Thanks a lot, see the update. –  Dmitry Kerner Feb 8 '11 at 14:32
@Georges: nice catch. Thanks! –  Sándor Kovács Feb 8 '11 at 17:05