Let $pr:X\to Z$ be a line bundle of (possibly) singular varieties (that could be reducible) over a characteristic $p$ field ($p$ could be zero); $U$ is the complement to the zero section $Z\to X$. Then for any $l\neq p$, $n>0$, there should exist a Gysin long exact sequence for the (etale or singular) cohomology $\dots \to H^i(X,\mathbb{Z}/l^n\mathbb{Z}) \to H^i(U,\mathbb{Z}/l^n\mathbb{Z})\to H^{i-1}(Z,\mathbb{Z}/l^n\mathbb{Z}(-1))\to \dots$ that is functorial in $pr$. Is this true?

Note that one can compute the cohomology of $U$ as the hypercohomology $H^*(Z,Rpr'_\ast\mathbb{Z}/l^n\mathbb{Z}_U)$, where $pr': U\to Z$ is the corresponding prinicple $G_m$-bundle. Hence the problem is to verify that $R^1pr'_*\mathbb{Z}/l^n\mathbb{Z}_U\cong \mathbb{Z}/l^n\mathbb{Z}_Z(1)$. Certainly, there is such an isomorphism for the trivial $G_m$-bundle $U\cong G_m\times Z$. Since this isomorphism does not seem to depend on the choice of a trivialization, one can 'glue' 'global' functorial isomorphisms from these 'local' isomorphisms. Yet I would prefer to have a reference for this result (or for some similar one).

Cf. https://mathoverflow.net/questions/89171/on-the-cohomology-of-g-m-bundles-and-purity-for-singular-varieties

  • $\begingroup$ Do you mean $pr$ to be a projective bundle? $\endgroup$ Commented Mar 5, 2012 at 1:32
  • $\begingroup$ Actually, $pr$ is a line bundle. One can certainly compactify it to get a $P^1$-bundle. The problem is that the etale cohomology of a line bundle should be isomorphic to the one of the base (though I am not sure that I know the proof), and it seems to be more difficult to compute the cohomology of a $P^1$-bundle. $\endgroup$ Commented Mar 5, 2012 at 5:55
  • 1
    $\begingroup$ Right, but then wouldn't this fail in the simplest case where $Z$ is a separably closed point and $X=\mathbb{A}^1$? For then the cohomology of $X$ and $Z$ is trivial but $U$ has the cohomology of a circle. The cohomology of a projective bundle, on the other hand, satisfies the same projective bundle formula as in topology (in complete generality; this is in SGA5, exposé VII, Theorem 2.2.1), so that $H^\ast(X,\Lambda)=H^\ast(Z,\Lambda)\oplus H^{\ast-2}(Z,\Lambda(-1))$ (if I get the gradings right). $\endgroup$ Commented Mar 5, 2012 at 6:38
  • 2
    $\begingroup$ Ah, and I just notice now that the Gysin exact sequence is proved 2 pages earlier, in Corollaire 1.5. It works for any rank $r$ vector bundle $X\to Z$, $Z$ any scheme, and reads $\to H^i(Z,L)\to H^i(U,q^\ast L)\to H^{i-2r+1}(Z,L(-r))\to H^{i+1}(Z,L)\to$. Here $q: U\to Z$ is the projection and $L$ is any $\mathbb{Z}/n$ module where $n$ is coprime to the residual characteristics of $Z$. $\endgroup$ Commented Mar 5, 2012 at 6:52
  • $\begingroup$ And in my previous comment I was confused about reduced vs nonreduced cohomology, please disregard the "counterexample". The gradings in the sequence in your original post seem incorrect though. $\endgroup$ Commented Mar 5, 2012 at 6:59

1 Answer 1


Comment promoted to answer:

The following is Corollaire 1.5 in SGA5, Exposé VII:

Let $Z$ be any scheme, $p:X\to Z$ a rank $r$ vector bundle, $U=X-Z$, and $q:U\to Z$ the retsriction of $p$. Let $n$ be coprime to the residual characteristics of $Z$ and let $L$ be a sheaf of $\mathbb{Z}/n$-modules on $Z$. Then there is a long exact sequence

$$\to H^{\ast-2r}(Z,L(-r))\to H^\ast(Z,L)\to H^\ast(U,q^\ast L)\to H^{\ast+1-2r}(Z,L(-r))\to$$

That's how the corollary is stated but they also show that the sequence is naturally isomorphic to the more familar-looking sequence for cohomology with supports

$$\to H^{\ast}_Z(X,p^\ast L)\to H^\ast(X,p^\ast L)\to H^\ast(U,q^\ast L)\to H^{\ast+1}_Z(X,p^\ast L)\to$$

You should definitely have a look at the rest of exposé VII in SGA5, which proves a lot of related "expected" theorems for étale cohomology with very mild hypotheses (compared to what can be found elsewhere).

  • $\begingroup$ Yes, you are quite right: I found two papers where some weaker results are proved. $\endgroup$ Commented Mar 6, 2012 at 9:37

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.