---
title: Frobenius Cokernel in Algebraic Geometry
url: https://www.emergentmind.com/topics/frobenius-cokernel
type: topic
---

# Frobenius Cokernel in Algebraic Geometry

In algebraic geometry, arithmetic, and representation theory, the Frobenius cokernel denotes the cokernel of a Frobenius-induced map, but the concrete object depends on the ambient category. On a smooth projective variety in characteristic \(p>0\), it is the vector bundle \(\mathcal B_X:=\operatorname{coker}(\mathcal O_X\to F_*\mathcal O_X)\); on Witt vectors, it is the quotient \(W(R)/\operatorname{im}(F)\); in invariant theory, it appears as \(R/R^p\) for the Frobenius map on a graded ring; and in recent random-matrix models, cokernels of polynomial endomorphisms over \(\mathbb Z_p\) are treated as direct analogues of “Frobenius cokernels” [2110.15035] [1409.7530] [1705.01832] [2310.09491].

## 1. Basic definitions and exact sequences

Let \(X\) be a smooth projective variety over an algebraically closed field \(k\) of characteristic \(p>0\), and let
\[
F:X\to X,\qquad F^\sharp:\mathcal O_X\to\mathcal O_X,\ f\mapsto f^p
\]
be the absolute Frobenius morphism. Since \(X\) is reduced, there is a canonical injective map
\[
\mathcal O_X \longrightarrow F_*\mathcal O_X.
\]
The Frobenius cokernel is defined by
\[
\mathcal B_X:=\operatorname{coker}\bigl(\mathcal O_X\to F_*\mathcal O_X\bigr),
\]
so that one has a short exact sequence of locally free sheaves
\[
0\longrightarrow \mathcal O_X\longrightarrow F_*\mathcal O_X\longrightarrow \mathcal B_X\longrightarrow 0.
\]
Using Grothendieck duality for the finite morphism \(F\), one identifies
\[
\mathcal B_X^\vee \cong \ker\bigl(F_*(\omega_X^{1-p})\to \mathcal O_X\bigr),
\]
so the dual of the Frobenius cokernel is the kernel of the Frobenius trace map [2510.03193].

For the \(e\)-th iterate \(F^e\), with \(q=p^e\), one writes
\[
T_e:T_{e,X}:F^e_*\omega_X^{1-q}\to \mathcal O_X,
\qquad
\mathcal E_{e,X}:=\ker(T_e),
\]
and obtains the exact sequences
\[
0\to \mathcal E_{e,X}\to F^e_*\omega_X^{1-q}\xrightarrow{T_e}\mathcal O_X\to 0
\]
and
\[
0\to \mathcal O_X\xrightarrow{F^e_*}F^e_*\mathcal O_X\to \mathcal E_{e,X}^\vee\to 0.
\]
Thus the cokernel of the Frobenius endomorphism is exactly \(\mathcal E_{e,X}^\vee\), while \(\mathcal E_{e,X}\) is the Frobenius trace kernel. These sequences are locally split, and \(\mathcal E_{e,X}\) is locally free of rank \(q^d-1\) when \(d=\dim X\) [2110.15035].

The same formal pattern recurs in higher degree. If
\[
\mathcal B_X^i:=\operatorname{im}\bigl(F_*d:F_*\Omega_X^{i-1}\to F_*\Omega_X^i\bigr),
\qquad
\mathcal Z_X^i:=\ker\bigl(F_*d:F_*\Omega_X^i\to F_*\Omega_X^{i+1}\bigr),
\]
then the Cartier isomorphism yields short exact sequences
\[
0\to \mathcal B_X^i\to \mathcal Z_X^i\xrightarrow{\kappa_i}\Omega_X^i\to 0,
\]
and \((\mathcal B_X^i)^\vee\) is the kernel of the higher Cartier operator \(\kappa_i\) [2510.03193].

## 2. Positivity, negativity, and the geometry they detect

A central theme is that positivity of the Frobenius trace kernel is equivalent to negativity of the Frobenius cokernel. If \(\mathcal E_{e,X}\) is nef, then \(\omega_X^{-1}\) is nef; if \(\mathcal E_{e,X}\) is ample, then \(\omega_X^{-1}\) is ample, so \(X\) is Fano. If \(\mathcal E_{e,X}\) is globally generated and \(X\) is \(F\)-split, then \(\omega_X^{-1}\) is ample as well [2110.15035].

These implications are not merely numerical. An ample Frobenius trace kernel rules out nontrivial smooth fibrations: if \(f:X\to S\) is a fibration with smooth general fiber and \(\mathcal E_{e,X}\) is ample, then the general fibers are zero-dimensional, and all fibers are zero-dimensional if \(f\) is flat. Blow-ups in codimension \(>1\) likewise destroy ampleness of the dual Frobenius trace bundle, because the restriction of \(F^e_*\mathcal O_X\) to the exceptional divisor has a nontrivial trivial summand [2110.15035].

The opposite positivity problem concerns the Frobenius cokernel itself rather than its dual. For a smooth projective variety over an \(F\)-finite field of characteristic \(p>0\),
\[
\mathcal B_X \text{ is ample (resp. nef) } \iff \Omega_X^1 \text{ is ample (resp. nef).}
\]
This equivalence is proved by analyzing the Katz–Sun filtration on \(F^*F_*\mathcal O_X\), whose graded pieces are truncated symmetric powers \(T^i\Omega_X^1\). It gives a direct bridge between classical positivity of the cotangent bundle and Frobenius-based positivity [2510.03193].

This suggests two complementary regimes. When \(\mathcal B_X^\vee\) is ample, the variety is forced toward Fano-type geometry; when \(\mathcal B_X\) is ample, the cotangent bundle is ample, so the geometry is controlled by strong negativity of the tangent bundle. The literature treats these as distinct but formally parallel manifestations of Frobenius-controlled positivity [2110.15035] [2510.03193].

## 3. Projective space, quadrics, and classification results

Projective space is the basic model. On \(\mathbb P^d\),
\[
\mathcal E_{e,\mathbb P^d}\cong \bigoplus_{i=1}^d \mathcal O_{\mathbb P^d}(i)^{\oplus a(i,0;d,e)},
\qquad
\mathcal E_{e,\mathbb P^d}^\vee\cong \bigoplus_{i=1}^d \mathcal O_{\mathbb P^d}(-i)^{\oplus a(i,0;d,e)}.
\]
Hence \(\mathcal E_{e,\mathbb P^d}\) is ample, and the Frobenius cokernel is strongly negative. This behavior motivated the question whether ampleness of the Frobenius trace kernel characterizes projective space [2110.15035].

In low dimension, the answer is affirmative. If \(X\) is a smooth projective curve and \(\mathcal E_{e,X}\) is ample, then \(X\cong \mathbb P^1\). If \(X\) is a smooth projective surface, then \(\mathcal E_{e,X}\) is ample if and only if \(X\cong \mathbb P^2\). For smooth projective threefolds, ampleness of \(\mathcal E_{e,X}\) forces \(X\) to be a Fano threefold of Picard rank \(\rho(X)=1\) [2110.15035].

The threefold classification was sharpened later. If \(X\) is a smooth Fano threefold over an algebraically closed field of characteristic \(p>0\), then
\[
\mathcal B_X^\vee \text{ is ample} \iff X\cong \mathbb P^3 \text{ or } X \text{ is a smooth quadric threefold, with } p\ne 2.
\]
The key obstruction is geometric: if \(X\) contains a smooth positive-dimensional subvariety \(Z\) with normal bundle \(N_{Z/X}\) such that \((\det N_{Z/X})^{1-p}\) is effective, then \(\mathcal B_X^\vee\) is not ample. In particular, if \(X\) contains a smooth rational curve \(C\simeq \mathbb P^1\) with \(-K_X\cdot C\le 2\), then \(\mathcal B_X^\vee\) is not ample. This excludes Fano threefolds containing lines or conics with respect to \(-K_X\) [2510.03193].

The naive projective-space characterization fails in higher dimension. Quadrics provide the basic counterexample: for \(d\ge 3\), \(\mathcal E_{e,Q_d}\) is ample if and only if \(p\neq 2\), and \(\mathcal E_{e,Q_d}\otimes \omega_{Q_d}\) is ample for all \(p\) [2110.15035]. More broadly, for \(\mathbb P^n\) and quadric hypersurfaces, the full set of kernels of the higher Cartier operators does not characterize projective space [2510.03193]. This is the main correction to the optimistic expectation that Frobenius positivity should single out \(\mathbb P^d\).

Complete intersections furnish another obstruction class. If \(X\subset \mathbb P^n\) is a smooth complete intersection of degrees \(d_1,\dots,d_c\ge 2\) with \(\dim X\ge 2\) and \(\sum d_i=n-1\) or \(n\), then \(\mathcal B_X^\vee\) is not ample. The argument uses the existence of lines in \(X\) together with the inequality \(-K_X\cdot C\le 2\) for such lines [2510.03193].

## 4. Witt vector Frobenius and its cokernel

For a commutative ring \(R\) and a prime \(p\), the \(p\)-typical Witt vector Frobenius
\[
F:W(R)\to W(R)
\]
and its finite-level truncations
\[
F:W_{p^{n+1}}(R)\to W_{p^n}(R)
\]
provide another classical setting for the Frobenius cokernel. Here
\[
\ker(F)=\{\underline{x}:F(\underline{x})=0\},
\qquad
\operatorname{coker}(F)=W(R)/\operatorname{im}(F),
\]
and vanishing of the cokernel is equivalent to surjectivity of the Frobenius map [1409.7530].

The structure of the kernel is governed by a descending sequence of ideals
\[
I_0=R,\qquad I_i:=\{r\in R:r^p\in pI_{i-1}\}\quad (i>0),
\qquad
I_\infty:=\bigcap_{i=1}^\infty I_i.
\]
An element \(r\in R\) occurs as the first component of an element in
\(\ker(F:W_{p^n}(R)\to W_{p^{n-1}}(R))\) if and only if \(r\in I_n\) [1409.7530].

The finite-level cokernel vanishes precisely under a ring-theoretic \(p\)-root condition. Surjectivity of
\[
F:W_{p^n}(R)\to W_{p^{n-1}}(R)\quad \text{for all } n\ge 2
\]
is equivalent to surjectivity of the \(p\)-th power map on each quotient \(R/pI_n\). At infinite level, surjectivity is stricter:
\[
F:W(R)\to W(R)\text{ is surjective}
\iff
\text{finite-level surjectivity}+\text{spherical completeness}.
\]
Equivalently, full surjectivity is characterized by finite-level surjectivity together with Teichmüller density, or by surjectivity of the \(p\)-th power map on \(R/pI_\infty\) together with the same completeness condition [1409.7530].

Examples show that finite and infinite behavior diverge sharply. For \(R=\mathbb Z\), finite-level surjectivity fails, so the Frobenius cokernel is already nonzero at finite levels. For the ring of integers \(\mathcal O\) in an algebraic closure of \(\mathbb Q_p\), finite-level Frobenius is surjective for all \(n\), but spherical completeness fails, so the infinite-level cokernel is nonzero. For a spherically complete valuation ring of the kind constructed by Poonen, the infinite Witt Frobenius is surjective [1409.7530].

In the mixed-characteristic direction, the finite-level vanishing of the Frobenius cokernel is described as the integral perfectoid condition of Scholze, repackaged in Witt-vector language. The same condition is stable under certain integral extensions and underlies the almost-purity statement proved there [1409.7530].

## 5. Graded rings, invariant theory, and homogeneous varieties

In invariant theory, the Frobenius cokernel appears module-theoretically. Let
\[
R=S^G
\]
be the homogeneous coordinate ring of the Grassmannian \(\mathbb G=\operatorname{Gr}(2,n)\), with \(S=\operatorname{Sym}(F\otimes V)\) and \(G=\operatorname{SL}(V)\). The Frobenius on \(R\) is
\[
F_R:R\to R,\qquad r\mapsto r^p,
\]
with image \(R^p\). Viewing \(R\) as an \(R^p\)-module, the Frobenius cokernel is
\[
\operatorname{coker}(R^p\hookrightarrow R)\cong R/R^p.
\]
The decomposition of \(R\) as a graded \(R^p\)-module therefore controls the structure of the cokernel [1705.01832].

The indecomposable Frobenius summands are expressed in terms of modules of covariants \(T\{j\}\) and certain indecomposable Cohen–Macaulay modules \(K\{j\}\). Forgetting degrees, the Frobenius summands of \(R\) lie among
\[
\{T\{0\},\dots,T\{n-3\},K\{1\},\dots,K\{n-3\}\},
\]
and all of these appear when \(p\ge n-2\). This is the paper’s explicit instance of \(p\)-uniformity: for sufficiently large \(p\), the set of indecomposable summands does not depend on \(p\), although multiplicities and grading shifts still vary [1705.01832].

Geometrically, the same analysis yields a decomposition of the Frobenius pushforward \(\operatorname{Fr}_{\mathbb G*}\mathcal O_{\mathbb G}\) into indecomposable homogeneous bundles. Up to multiplicity, the summands are twists of \(\mathcal O_{\mathbb G}\), the bundles \(T_j\), and twists of \(\bigwedge^j\mathcal R\). This Frobenius pushforward is a tilting bundle only for \(n=4\) and \(p>3\); otherwise it is not tilting [1705.01832].

The module-theoretic consequence for the cokernel is that the Frobenius cokernel is far from projective in general. The nonfree indecomposable summands \(K\{j\}\) contribute to the nonfree part of \(R\) over \(R^p\), and hence to the nonfree part of the Frobenius cokernel. The same summands are then used to construct a noncommutative resolution \(\operatorname{End}_{R^p}(M)\) of finite global dimension for a suitable direct sum \(M\) of Frobenius summands [1705.01832].

## 6. Random and cohomological analogues

Recent random-matrix work treats polynomial cokernels over \(\mathbb Z_p\) as direct analogues of Frobenius cokernels. For
\[
X_n=A_n+pB_n\in M_n(\mathbb Z_p),
\]
with \(A_n\in M_n(\mathbb F_p)\) fixed and \(B_n\) random, and for a non-constant monic polynomial \(P(t)\in \mathbb Z_p[t]\), the central object is
\[
\operatorname{cok}(P(X_n)).
\]
More naturally, one equips it with the structure of an \(R\)-module for
\[
R:=\mathbb Z_p[t]/(P(t)),
\]
via letting the image of \(t\) act by \(X_n\). The key linearization is
\[
\operatorname{cok}(P(X))\cong_R \operatorname{cok}_R(X-tI_n),
\]
which replaces a polynomial in \(X\) by an \(R\)-linear map. The paper proves exact finite-\(n\) formulas for the distribution of \(\operatorname{cok}(P(X_n))\) when \(B_n\) is Haar-random, establishes universality for a broader concentrated-residue class, and derives a Cohen–Lenstra style limit law in the non-concentrated regime [2310.09491].

This random-matrix perspective is explicitly presented as directly analogous to “Frobenius cokernels.” Conceptually, one studies an endomorphism \(\varphi\) of a \(p\)-adic module and the modules
\[
\operatorname{coker}\bigl(P(\varphi):T\to T\bigr),
\]
while the random model replaces \(\varphi\) by \(X_n\). The resulting distribution depends on \(|\operatorname{Aut}_R(G)|\) and on the factorization of \(P\), and in the squarefree case becomes a direct generalization of the Cohen–Lenstra measure to modules over \(\mathbb Z_p[t]/(P)\) [2310.09491].

A complementary universality principle holds for random integral matrices. For an \(n\times(n+u)\) random integral matrix \(M\) with independent entries from a very broad class of distributions, the limiting probability that \(\operatorname{coker}(M)\cong B\) is
\[
\frac{1}{|B|^u|\operatorname{Aut}(B)|}\prod_{k=u+1}^\infty \zeta(k)^{-1}.
\]
This is a Cohen–Lenstra-type law, and it agrees with the distribution defined by the Haar model on \(\widehat{\mathbb Z}\). The paper interprets this as support for the principle that random endomorphisms modelling Frobenius produce canonical cokernel distributions independent of microscopic details of the entry distribution [1806.00596].

Representation-theoretic work on Frobenius kernels gives a different, but related, analogue. For the \(r\)-th Frobenius kernels \((SL_2)_r\), the cohomology ring \(H^\bullet((SL_2)_r,k)_{\mathrm{red}}\) is Cohen–Macaulay, and its maximal ideal spectrum is homeomorphic to \(G\times_B\mathfrak u^r\). That work is about kernels of Frobenius rather than cokernels, but it is formulated so that explicit knowledge of \(H^\bullet(G_r,k)\) constrains possible kernels and cokernels of Frobenius-induced maps on cohomology [1209.1662].

Across these settings, the Frobenius cokernel is best viewed not as a single invariant but as a family of formally parallel constructions. In geometry it is a vector bundle attached to \(F_*\mathcal O_X\); in Witt theory it measures the failure of Frobenius surjectivity on \(W(R)\); in invariant theory it records the nonfree part of \(R\) over \(R^p\); and in random-matrix models it becomes a probabilistic proxy for cokernels of polynomial functions of Frobenius-like endomorphisms [2110.15035] [1409.7530] [1705.01832] [2310.09491].

Source: https://www.emergentmind.com/topics/frobenius-cokernel