Serre vanishing

For an ample line bundle $A$ on a projective scheme and a coherent sheaf $\mathcal{F}$, $H^i(X,\mathcal{F}\otimes A^m)=0$ for every $i>0$ and all sufficiently large $m$.

For ample line bundles $L,B$ and a coherent sheaf $\mathcal{F}$ on a projective scheme, one $M$ gives $H^i(X,\mathcal{F}\otimes L^m\otimes B^t)=0$ for every $i>0$, $m\ge M$ and $t\ge 0$. If $B$ is very ample, Serre vanishing for the finitely many fixed twists $\mathcal{F}(-iB)$ makes $\mathcal{F}\otimes L^m$ zero-regular. Persistence of Castelnuovo–Mumford regularity gives all nonnegative $B$-twists. For general ample $B$, use a very ample power and treat finitely many residues of $t$.

Let $H$ be a very ample effective Cartier divisor on a projective scheme $X$, and suppose $L{|}_H$ is ample. Then $H^i(X,L^m)=0$ for $i\ge 2$ and sufficiently large $m$. By uniform Serre vanishing for two ample twists, cohomology of $L^m(tH){|}_H$ vanishes uniformly for $t\ge 0$. The restriction sequence identifies the higher groups of $L^m(tH)$ and $L^m((t+1)H)$ for $i\ge 2$. For fixed $m$, large $t$ kills them by Serre vanishing for $H$; descend to $t=0$. No $H^1$ vanishing is claimed.

Given a coherent sheaf $\mathcal{F}$ and an ample Cartier divisor $A$ on a projective scheme, one integer $a_0$ ensures $H^i(X,\mathcal{F}(aA+N))=0$ for all $i>0$, all $a\ge a_0$, and every nef divisor $N$. The uniformity in $N$ is stronger than Serre vanishing. Fujino's note on Fujita vanishing states the theorem for arbitrary projective schemes over a field.

Let $N$ be a nef Cartier divisor and let the coherent sheaf $\mathcal{F}$ have support dimension $d$. Then $h^i(X,\mathcal{F}(mN))=O(m^{d-i})$ for $1\le i\le d$. This does not require reducedness or smoothness.

Choose a very ample $A$ and fix $a$ sufficiently large for Fujita vanishing. Choose a section of $aA$ avoiding the associated points of $\mathcal{F}$. Its multiplication gives $0\to \mathcal{F}\to \mathcal{F}(aA)\to \mathcal{Q}\to 0$, where $\mathcal{Q}$ has support dimension at most $d-1$. After tensoring with ${\mathcal{O}}_X(mN)$, the middle term has zero higher sheaf cohomology, uniformly in $m$. Hence $h^i(\mathcal{F}(mN))\le h^{i-1}(\mathcal{Q}(mN))$. Induct on support dimension: for $i=1$ use the polynomial bound for sections of a fixed divisor, and for $i>1$ use the induction hypothesis. Degree greater than $d$ vanishes by Grothendieck vanishing.

The top-degree case of cohomology growth for nef twists is bounded independently of $m$. It fills a gap when a dimension induction hypothesis only covers intermediate sheaf cohomology degrees.