betaFeedback

WorksMathematics

Finite torsion in Hodge–Witt cohomology groups

Abstract

For bounded coherent complexes over the Raynaud ring, we establish natural Dieudonné isomorphisms between Ekedahl's biduality pieces Af2A_f^2 and A_{\SSpart}^1, the nilpotent and semi-simple torsion of the Illusie–Raynaud cœur, and the corresponding parts of the finite torsion t1t^1 of Ekedahl's diagonal cohomology. In bidegree (0,i+1)(0,i+1), this torsion measures the non-smoothness of the iith flat Artin–Mazur formal group; hence all such groups are smooth for Mazur–Ogus varieties. For a smooth projective surface SS, we describe the three possible dual pairs via the Galois module \NS(\bar{S})[p^\infty] and the local-local and local-étale parts of the finite Picard quotient.

The three descriptions and the comparison theorem

Let kk be a perfect field of characteristic p>0p>0, let W=W(k)W=W(k) and K=W[1/p]K=W[1/p], and let σ\sigma be the Frobenius of WW; we also write aσ=σ(a)a^\sigma=\sigma(a). As in [8 I, 1.1], R=R(k)=R0⊕R1R=R(k)=R^0\oplus R^1 denotes the Raynaud ring: the graded WW-algebra generated by FF and VV in degree 00 and by dd in degree $1, subject to the relations

Fa=aσF,aV=Vaσ,FV=VF=p,da=ad,d2=0,FdV=dFa=a^\sigma F,\qquad aV=Va^\sigma,\qquad FV=VF=p,\qquad da=ad,\qquad d^2=0,\qquad FdV=d

for a∈Wa\in W. Its degree-zero part R0R^0 is the Dieudonné ring, and we call R0R^0-modules Dieudonné modules. We write Db(R)D^b(R) for the bounded derived category of graded RR-modules with coherent cohomology [8 I, §3], [3]. Coherence allows dominos; it does not mean that the graded pieces are finitely generated over WW.

For M∈Db(R)M\in D^b(R), Hj(M)iH^j(M)^i denotes the part of internal (Raynaud) degree ii of the jjth cohomology module. Following [3 §0], M(a)[b]M(a)[b] is the shift by aa in the RR-grading and by bb in the cohomological grading:

Hn(M(a)[b])m=Hn+b(M)m+a.H^n(M(a)[b])^m=H^{n+b}(M)^{m+a}.

Thus (−i)(-i) moves a module concentrated in internal degree 00 to internal degree ii. “Finite torsion” always means finite WW-length, not finiteness of the underlying set. The letter TT with two indices is reserved for the domino numbers Ti,jT^{i,j} of Illusie–Raynaud, which appear only in Section 5; Ekedahl’s filtration Ta(L)T^a(L) in § 1.2 carries a single index.

The cœur and its finite torsion

Let LL be a coherent graded RR-module. Following [8 I, 1.4], put Zi=ker⁡(d:Li→Li+1)Z^i=\ker(d:L^i\to L^{i+1}), Bi=dLi−1B^i=dL^{i-1},

V−∞Zi=⋂r>0V−rZi,V−rZi={x∈Li:Vrx∈Zi},(1)V^{-\infty}Z^i=\bigcap_{r>0}V^{-r}Z^i,\qquad V^{-r}Z^i=\{x\in L^i:V^rx\in Z^i\}, \tag*{(1)}
F∞Bi=⋃s≥0FsBi,(2)F^\infty B^i=\bigcup_{s\geq0}F^sB^i, \tag*{(2)}

and define the cœur of LL in degree ii by

cœur⁡(Li)=V−∞Zi/F∞Bi.(3)\operatorname{cœur}(L^i)=V^{-\infty}Z^i/F^\infty B^i. \tag*{(3)}

When the module has to be specified we write V−∞Zi(L)V^{-\infty}Z^i(L) and F∞Bi(L)F^\infty B^i(L). The union in (2) is increasing because d=FdVd=FdV. Both V−∞ZiV^{-\infty}Z^i and F∞BiF^\infty B^i are R0R^0-submodules of LiL^i, and Bi⊂F∞Bi⊂V−∞Zi⊂ZiB^i\subset F^\infty B^i\subset V^{-\infty}Z^i\subset Z^i [8 I, 1.4]. By the structure theory of coherent modules, cœur(LiL^i) is a finitely generated WW-module on which VV is topologically nilpotent [8 I, Theorem 2.9]; see also [3 Lemma 0.8].

Every Dieudonné module DD of finite length has a canonical Fitting decomposition

D=Dnil⊕Dss,Dnil=ker⁡Fr,Dss=im⁡Fr(r≫0).(4)D=D_{\mathrm{nil}}\oplus D_{\mathrm{ss}},\qquad D_{\mathrm{nil}}=\ker F^r,\qquad D_{\mathrm{ss}}=\operatorname{im}F^r\quad(r\gg0). \tag*{(4)}

Here im⁡Fr\operatorname{im}F^r is a WW-submodule because kk is perfect; FF is nilpotent on DnilD_{\mathrm{nil}} and bijective on DssD_{\mathrm{ss}}; both summands are R0R^0-submodules because FF and VV commute; and the decomposition is functorial. As in the classical theory of pp-linear maps, we call DnilD_{\mathrm{nil}} and DssD_{\mathrm{ss}} the nilpotent and semi-simple parts of DD, and we call DD nilpotent or semi-simple if it equals the corresponding part. “Semi-simple” refers only to the bijectivity of FF, not to semisimplicity as a representation.

For M∈Dcb(R)M\in D^b_c(R) and i,j∈Zi,j\in\mathbb{Z} put

cœur⁡i,j(M)=cœur⁡(Hj(M)i),(5)\operatorname{cœur}^{i,j}(M)=\operatorname{cœur}\left(H^j(M)^i\right), \tag*{(5)}

and let cœur⁡i,j(M)tors\operatorname{cœur}^{i,j}(M)_{\mathrm{tors}} be its pp-power torsion submodule. This is a Dieudonné module of finite length, and we write

cœur⁡i,j(M)tors=cœur⁡i,j(M)nil⊕cœur⁡i,j(M)ss\operatorname{cœur}^{i,j}(M)_{\mathrm{tors}}=\operatorname{cœur}^{i,j}(M)_{\mathrm{nil}}\oplus\operatorname{cœur}^{i,j}(M)_{\mathrm{ss}}

for its decomposition (4). For a single coherent module LL we use the same notation cœur⁡(Li)tors\operatorname{cœur}(L^i)_{\mathrm{tors}}, cœur⁡(Li)nil\operatorname{cœur}(L^i)_{\mathrm{nil}} and cœur⁡(Li)ss\operatorname{cœur}(L^i)_{\mathrm{ss}}.

Ekedahl’s biduality filtration

Let DD be Ekedahl’s duality functor on Dcb(R)D^b_c(R), and put Da(L)=Ha(D(L))D^a(L)=H^a(D(L)) [3 IV, 2.1]. For a coherent graded RR-module LL, the biduality spectral sequence gives a filtration

0⊂T2(L)⊂T1(L)⊂T0(L)=L,Aa(L)=Ta(L)/Ta+1(L),0\subset T^2(L)\subset T^1(L)\subset T^0(L)=L,\qquad A^a(L)=T^a(L)/T^{a+1}(L),

with Aa(L)≃DaDa(L)A^a(L)\simeq D^aD^a(L) [3 IV, 2.3–2.4]. The filtration has length three, so A2(L)=T2(L)A^2(L)=T^2(L).

As in [3 IV, 3.2–3.3], the two finite pieces are

Af2(L)i=cœur⁡(A2(L)i),Ass1(L)=A1(L)tors.(6)A^2_f(L)^i=\operatorname{cœur}(A^2(L)^i),\qquad A^1_{\mathrm{ss}}(L)=A^1(L)_{\mathrm{tors}}. \tag*{(6)}

Both have finite WW-length; FF is nilpotent on Af2(L)A^2_f(L) and bijective on Ass1(L)A^1_{\mathrm{ss}}(L) [3 IV, Theorem 3.3]. Moreover,

T2(L)⊂Ltors,Ltors/T2(L)=A1(L)tors=Ass1(L),(7)T^2(L)\subset L_{\mathrm{tors}},\qquad L_{\mathrm{tors}}/T^2(L)=A^1(L)_{\mathrm{tors}}=A^1_{\mathrm{ss}}(L), \tag*{(7)}

where LtorsL_{\mathrm{tors}} is the pp-power torsion of LL. Indeed, T2(L)T^2(L) is pp-power torsion [3 IV, Proposition 2.4 and Lemma 2.4.1], and the filtration gives an exact sequence

0⟶A1(L)⟶L/T2(L)⟶A0(L)⟶00\longrightarrow A^1(L)\longrightarrow L/T^2(L)\longrightarrow A^0(L)\longrightarrow0

with A0(L)A^0(L) torsion-free [3 IV, Theorem 3.3]. Hence the torsion of L/T2(L)L/T^2(L), which is Ltors/T2(L)L_{\mathrm{tors}}/T^2(L), equals A1(L)torsA^1(L)_{\mathrm{tors}}. The differential vanishes on Ass1(L)A^1_{\mathrm{ss}}(L): VV is nilpotent there, so d=FrdVrd=F^r dV^r is zero for $r\gg 0.

Diagonal finite torsion

Write Δ\Delta for the heart of Ekedahl’s diagonal tt-structure on Dcb(R)D^b_c(R), and H~n\widetilde{\mathrm{H}}^n for its cohomology functors [5]. The ordinary and diagonal tt-structures commute. If A∈ΔA \in\Delta, its ordinary cohomology satisfies

Hq(A)a=0unless a∈{−q,1−q},F∞B1−q(Hq(A))=Hq(A)1−q.(8)\mathrm{H}^q(A)^a=0 \quad\text{unless } a\in\{-q,1-q\}, \qquad F^\infty B^{1-q}(\mathrm{H}^q(A))=\mathrm{H}^q(A)^{1-q}. \tag*{(8)}

Thus Hq(A)(−q)\mathrm{H}^q(A)(-q) is a normalized diagonal module: it is concentrated in degrees 0 and 1, and its degree-one term consists of saturated boundaries.

Ekedahl’s diagonal duality gives endofunctors t1t^1 and t0t^0 of Δ\Delta and, for every object AA of Δ\Delta, a natural exact sequence

0⟶t1A⟶A⟶t0A⟶0,(9)0 \longrightarrow t^1 A \longrightarrow A \longrightarrow t^0 A \longrightarrow0, \tag*{(9)}

in which t1At^1 A is the largest subobject of finite torsion and t0At^0 A has no finite torsion; in terms of the diagonal duality, t1=D~ 1D~ 1t^1=\widetilde{D}^{\,1}\widetilde{D}^{\,1} and t0=D~ 0D~ 0t^0=\widetilde{D}^{\,0}\widetilde{D}^{\,0} [5 I, Proposition 1.11]. The superscripts 1 and 0 belong to this notation; they are not cohomological degrees.

For M∈Dcb(R)M \in D^b_c(R) and i,j∈Zi,j \in\mathbb{Z}, the finite torsion of diagonal cohomology in bidegree (i,j)(i,j) is the finite-length Dieudonné module (Lemma Lemma 3.1)

H−i((t1H~i+j(M))i),(10)\mathrm{H}^{-i}\left(\left(t^1\widetilde{\mathrm{H}}^{i+j}(M)\right)^i\right), \tag*{(10)}

with nilpotent and semi-simple parts H−i((t1H~i+j(M))i)nil\mathrm{H}^{-i}\left(\left(t^1\widetilde{\mathrm{H}}^{i+j}(M)\right)^i\right)_{\mathrm{nil}} and H−i((t1H~i+j(M))i)ss\mathrm{H}^{-i}\left(\left(t^1\widetilde{\mathrm{H}}^{i+j}(M)\right)^i\right)_{\mathrm{ss}} in the sense of (4).

Theorem 1.1 (Comparison of finite torsion). Let M∈Dcb(R)M \in D^b_c(R) and i,j∈Zi,j \in\mathbb{Z}. There are natural isomorphisms of finite-length Dieudonné modules

Af2(Hj(M))i≅coeur⁡i,j(M)nil≅H−i((t1H~i+j(M))i)nil,(11)A_f^2(\mathrm{H}^j(M))^i \cong\operatorname{coeur}^{i,j}(M)_{\mathrm{nil}} \cong\mathrm{H}^{-i}\left(\left(t^1\widetilde{\mathrm{H}}^{i+j}(M)\right)^i\right)_{\mathrm{nil}}, \tag*{(11)}
Ass1(Hj(M))i≅coeur⁡i,j(M)ss≅H−i((t1H~i+j(M))i)ss.(12)A_{\mathrm{ss}}^1(\mathrm{H}^j(M))^i \cong\operatorname{coeur}^{i,j}(M)_{\mathrm{ss}} \cong\mathrm{H}^{-i}\left(\left(t^1\widetilde{\mathrm{H}}^{i+j}(M)\right)^i\right)_{\mathrm{ss}}. \tag*{(12)}

In particular, these isomorphisms commute with FF and VV; they are not merely equalities of lengths or isomorphisms of WW-modules.

Section § 2 proves the comparison with Ekedahl’s pieces directly from the structure of coherent modules and Ekedahl’s biduality filtration, and Section § 3 deduces the diagonal comparison from Ekedahl’s finite-torsion criterion, keeping track of both degrees. Section § 4 interprets the finite torsion in internal degree zero as the obstruction to smoothness of the flat Artin–Mazur formal groups, deduces the corresponding prorepresentability criterion for the étale Artin–Mazur functors, in particular for the formal Brauer group, and applies this to Mazur–Ogus varieties. Section § 5 specializes to surfaces: the Kummer sequence and logarithmic de Rham–Witt cohomology identify the Néron–Severi contribution, the Cartier-module comparison of Section § 4 identifies the connected Picard contributions, and geometric duality determines their partners. The two duality conventions are kept apart.

Comparison with Ekedahl’s pieces

Fix a coherent graded RR-module LL, and put

P=Ltors,Q=T2(L),so that Q⊂P⊂LP=L_{\mathrm{tors}}, \qquad Q=T^2(L), \qquad\text{so that } Q\subset P\subset L

by (7). By the structure theory of coherent modules [8 I, Theorem 2.9 and §3] (see also [3 IV, p. 223] for the torsion and torsion-free quotients), there is an integer e≥0e\geq0 with peP=0p^{e}P=0, and L/PL/P is a finitely generated torsion-free WW-module with zero differential on which VV is topologically nilpotent. In particular dL⊂PdL\subset P and pedL=0p^{e}dL=0. By (7), P/Q=Ass1(L)P/Q=A^{1}_{\mathrm{ss}}(L) has finite length and bijective FF, so V=pF−1V=pF^{-1} is nilpotent on it.

We also use the domino of LL in degrees (i,i+1)(i,i+1) [8 I, 2.D]:

Li/V−∞Zi→dF∞Bi+1L^{i}/V^{-\infty}Z^{i}\xrightarrow{d}F^{\infty}B^{i+1}

FF is nilpotent on its lower term. Indeed, FdV=dFdV=d gives dF=pFddF=pFd, hence dFm=pmFmddF^{m}=p^{m}F^{m}d, and for x∈Lix\in L^{i} and n≥en\ge e,

dVrFnx={pnFn−rdx=0,r≤n,pndVr−nx=0,r>n.dV^{r}F^{n}x= \begin{cases} p^{n}F^{n-r}dx=0, & r\le n,\\ p^{n}dV^{r-n}x=0, & r>n. \end{cases}

Thus FnLi⊂V−∞ZiF^{n}L^{i}\subset V^{-\infty}Z^{i} for n≥en\ge e. The argument only uses pedL=0p^{e}dL=0, so it applies equally to every graded RR-submodule of LL.

Lemma 2.1. The inclusions Q⊂P⊂LQ\subset P\subset L induce equalities of submodules

F∞Bi(Q)=F∞Bi(P)=F∞Bi(L)F^{\infty}B^{i}(Q)=F^{\infty}B^{i}(P)=F^{\infty}B^{i}(L)

and injective maps

\operatorname{coeur}(Q^{i})\lhook\joinrel\longrightarrow\operatorname{coeur}(P^{i})\lhook\joinrel\longrightarrow\operatorname{coeur}(L^{i}).

Proof. Let x∈Li−1x\in L^{i-1}. Since VV is topologically nilpotent on L/PL/P, there is an integer rr with

Vrx=pey+u,y∈Li−1,u∈Pi−1.V^{r}x=p^{e}y+u,\qquad y\in L^{i-1},\quad u\in P^{i-1}.

Iterating d=FdVd=FdV and using pedL=0p^{e}dL=0, we obtain

dx=FrdVrx=Fr(pedy+du)=Frdu∈F∞Bi(P).dx=F^{r}dV^{r}x=F^{r}(p^{e}dy+du)=F^{r}du\in F^{\infty}B^{i}(P).

Since F∞Bi(P)F^{\infty}B^{i}(P) is FF-stable, this proves F∞Bi(L)⊂F∞Bi(P)F^{\infty}B^{i}(L)\subset F^{\infty}B^{i}(P); the reverse inclusion is clear. Next, VV is nilpotent on P/QP/Q, so VrP⊂QV^{r}P\subset Q for some rr, and for x∈Pi−1x\in P^{i-1} we get dx=FrdVrx∈F∞Bi(Q)dx=F^{r}dV^{r}x\in F^{\infty}B^{i}(Q). This proves the other equality in the same way.

Finally, V−∞Zi(J)=Ji∩V−∞Zi(L)V^{-\infty}Z^{i}(J)=J^{i}\cap V^{-\infty}Z^{i}(L) for every graded RR-submodule J⊂LJ\subset L. Dividing these inclusions of stable cycles by their common submodule F∞BiF^{\infty}B^{i} gives the injections of cœurs.

Proposition 2.2. There are natural identifications

coeur⁡(Pi)=coeur⁡(Li)tors,coeur⁡(Qi)=coeur⁡(Li)nil,Pi/Qi≃coeur⁡(Li)ss.\operatorname{coeur}(P^{i})=\operatorname{coeur}(L^{i})_{\mathrm{tors}},\qquad \operatorname{coeur}(Q^{i})=\operatorname{coeur}(L^{i})_{\mathrm{nil}},\qquad P^{i}/Q^{i}\simeq\operatorname{coeur}(L^{i})_{\mathrm{ss}}.

Proof. By Lemma 2.1,

coeur⁡(Li)/coeur⁡(Pi)=V−∞Zi(L)/V−∞Zi(P).\operatorname{coeur}(L^{i})/\operatorname{coeur}(P^{i})=V^{-\infty}Z^{i}(L)/V^{-\infty}Z^{i}(P).

Since V−∞Zi(P)=V−∞Zi(L)∩PiV^{-\infty}Z^{i}(P)=V^{-\infty}Z^{i}(L)\cap P^{i}, this quotient embeds in Li/PiL^{i}/P^{i}, which is torsion-free. On the other hand, coeur⁡(Pi)\operatorname{coeur}(P^{i}) is a subquotient of PiP^{i}, hence torsion, and it is finitely generated over WW as a submodule of coeur⁡(Li)\operatorname{coeur}(L^{i}). It is therefore exactly the torsion submodule of coeur⁡(Li)\operatorname{coeur}(L^{i}).

Applying the same lemma to Q⊂PQ\subset P gives

coeur⁡(Pi)/coeur⁡(Qi)=V−∞Zi(P)/V−∞Zi(Q).(13)\operatorname{coeur}(P^{i})/\operatorname{coeur}(Q^{i})=V^{-\infty}Z^{i}(P)/V^{-\infty}Z^{i}(Q). \tag*{(13)}

The inclusions of stable cycles and their quotients give an exact sequence

0⟶V−∞Zi(P)V−∞Zi(Q)⟶PiQi⟶Pi/V−∞Zi(P)Qi/V−∞Zi(Q)⟶0.(14)0\longrightarrow \frac{V^{-\infty}Z^{i}(P)}{V^{-\infty}Z^{i}(Q)} \longrightarrow \frac{P^{i}}{Q^{i}} \longrightarrow \frac{P^{i}/V^{-\infty}Z^{i}(P)}{Q^{i}/V^{-\infty}Z^{i}(Q)} \longrightarrow0. \tag*{(14)}

The last term is a quotient of the lower term of the domino of PP, so FF is nilpotent on it. It is also a quotient of the finite-length module Pi/QiP^i/Q^i, on which FF is bijective; hence FF is surjective on it and, by finite length and the perfectness of kk, bijective. A module on which FF is both nilpotent and bijective is zero. By (6) and (7), cœur⁡(Qi)=Af2(L)i\operatorname{cœur}(Q^i)=A^2_{\mathrm{f}}(L)^i and Pi/Qi=Ass1(L)iP^i/Q^i=A^1_{\mathrm{ss}}(L)^i, so (13) and (14) give an exact sequence

0⟶Af2(L)i⟶cœur⁡(Li)tors⟶Ass1(L)i⟶0.(15)0 \longrightarrow A^2_{\mathrm{f}}(L)^i \longrightarrow\operatorname{cœur}(L^i)_{\mathrm{tors}} \longrightarrow A^1_{\mathrm{ss}}(L)^i \longrightarrow0. \tag*{(15)}

The first term has nilpotent FF and the last has bijective FF, so the canonical decomposition (4) identifies them with the nilpotent and semi-simple parts of cœur⁡(Li)tors\operatorname{cœur}(L^i)_{\mathrm{tors}}. Every map used is induced by the natural filtrations and is R0R^0-linear.

Remark 2.3. The proof does not use exactness of the cœur for arbitrary short exact sequences of coherent modules. The injections it needs come from the equality of saturated boundaries in Lemma 2.1.

Comparison with diagonal finite torsion

We first determine what t1t^1 does to ordinary cohomology, using Ekedahl’s description of finite torsion [5 I, Theorem 1.12]:

  1. A∈ΔA\in\Delta is of finite torsion if and only if every Hq(A)H^q(A) is a finite-length Dieudonné module concentrated in internal degree −q-q;

  1. A∈ΔA\in\Delta has no finite torsion if and only if no normalized diagonal module Hq(A)(−q)H^q(A)(-q) has finite torsion, and a normalized diagonal module has no finite torsion if and only if the cœur of its degree-zero term is torsion-free.

These criteria concern finite torsion only; the pp-torsion carried by dominos is not detected by t1t^1.

Lemma 3.1. Let A∈ΔA\in\Delta and q∈Zq\in\mathbb{Z}. Then Hq(t1A)H^q(t^1A) is concentrated in internal degree −q-q, and there is a natural R0R^0-linear identification

Hq(t1A)−q≅cœur⁡(Hq(A)−q)tors.(16)H^q(t^1A)^{-q}\cong\operatorname{cœur}(H^q(A)^{-q})_{\mathrm{tors}}. \tag*{(16)}

Proof. The first assertion is criterion (1). Apply ordinary cohomology to (9). The map Hq−1(t0A)⟶Hq(t1A)H^{q-1}(t^0A)\longrightarrow H^q(t^1A) is zero, because the source lives in internal degrees 1−q1-q and 2−q2-q and the target only in degree −q-q. For the same reason Hq(t0A)⟶Hq+1(t1A)H^q(t^0A)\longrightarrow H^{q+1}(t^1A) vanishes. Hence there are short exact sequences of graded RR-modules

0⟶Hq(t1A)⟶Hq(A)⟶Hq(t0A)⟶0.(17)0 \longrightarrow H^q(t^1A) \longrightarrow H^q(A) \longrightarrow H^q(t^0A) \longrightarrow0. \tag*{(17)}

The last two modules live in internal degrees −q-q and 1−q1-q. In particular nothing arrives in degree −q-q under dd, so their cœurs in that degree are their submodules V−∞Z−qV^{-\infty}Z^{-q} of stable cycles.

The first term of (17) has finite length and is concentrated in degree −q-q, so its image lies in cœur⁡(Hq(A)−q)tors\operatorname{cœur}(H^q(A)^{-q})_{\mathrm{tors}}. Conversely, a torsion stable cycle of Hq(A)−qH^q(A)^{-q} maps to a torsion element of cœur⁡(Hq(t0A)−q)\operatorname{cœur}(H^q(t^0A)^{-q}), which is torsion-free by criterion (2); so its image is zero. Exactness proves (16). All the maps involved commute with FF and VV.

We now relate the diagonal cohomology of a complex to its individual Hodge–Witt positions. For a coherent graded RR-module LL, the radical filtration defining the diagonal tt-structure is

(FiL)a={La,a≤i,F∞Bi+1(L),a=i+1,0,a>i+1.(18)(\mathcal{F}_iL)^a = \begin{cases} L^a, & a \le i,\\ F^\infty B^{i+1}(L), & a = i+1,\\ 0, & a > i+1. \end{cases} \tag*{(18)}

Its iith layer SiL=FiL/Fi−1LS^iL=\mathcal{F}_iL/\mathcal{F}_{i-1}L is the two-term graded RR-module

(SiL)i=Li/F∞Bi(L)→d(SiL)i+1=F∞Bi+1(L).(19)(S^iL)^i=L^i/F^\infty B^i(L)\xrightarrow{d}(S^iL)^{i+1}=F^\infty B^{i+1}(L). \tag*{(19)}

In particular,

cœur((SiL)i)=cœur(Li).(20)\mathrm{cœur}((S^iL)^i)=\mathrm{cœur}(L^i). \tag*{(20)}

Indeed, the stable cycles in the lower term of (19) are exactly V−∞Zi(L)/F∞Bi(L)V^{-\infty}Z^i(L)/F^\infty B^i(L), and SiLS^iL is zero in degree i−1i-1.

Commutation of the ordinary and diagonal truncations gives the canonical layer identity

H−i(H~i+j(M))≃Si(Hj(M)).(21)\mathrm{H}^{-i}(\widetilde{\mathrm{H}}^{i+j}(M))\simeq S^i(\mathrm{H}^j(M)). \tag*{(21)}

To check the indices, take LL in cohomological degree 00: its iith diagonal layer is SiL[i]S^iL[i], whose only ordinary cohomology lies in degree −i-i; a summand in cohomological degree jj therefore contributes to diagonal degree i+ji+j. The general identity follows from the commuting truncations, without assuming that MM splits as a sum of its cohomology modules.

Proposition 3.2. For every M∈Dcb(R)M\in D^b_c(R) and i,j∈Zi,j\in\mathbb{Z}, there is a natural R0R^0-linear isomorphism

H−i(t1H~i+j(M))i≃cœuri,j(M)tors.\mathrm{H}^{-i}(t^1\widetilde{\mathrm{H}}^{i+j}(M))^i \simeq\mathrm{cœur}^{i,j}(M)_{\mathrm{tors}}.

Proof. Apply Lemma 3.1 to A=H~i+j(M)A=\widetilde{\mathrm{H}}^{i+j}(M) and q=−iq=-i, then use (21) and (20):

H−i(t1H~i+j(M))i≃cœur(H−i(H~i+j(M))i)tors≃cœur(Si(Hj(M))i)tors≃cœur(Hj(M)i)tors.\begin{aligned} \mathrm{H}^{-i}(t^1\widetilde{\mathrm{H}}^{i+j}(M))^i \simeq\mathrm{cœur}\left(\mathrm{H}^{-i}(\widetilde{\mathrm{H}}^{i+j}(M))^i\right)_{\mathrm{tors}} \\ &\simeq\mathrm{cœur}\left(S^i(\mathrm{H}^j(M))^i\right)_{\mathrm{tors}} \\ &\simeq\mathrm{cœur}(\mathrm{H}^j(M)^i)_{\mathrm{tors}}. \end{aligned}

Taking nilpotent and semi-simple parts gives the comparison for each part separately.

Propositions 2.2 and 3.2 prove Theorem 1.1. They also give the comparison maps explicitly; no length count is needed.

Corollary 3.3. For every nn, the finite torsion of diagonal cohomology is

t1H~n(M)≃⨁i+j=ncœuri,j(M)tors(−i)[i],(22)t^1\widetilde{\mathrm{H}}^n(M)\simeq\bigoplus_{i+j=n}\mathrm{cœur}^{i,j}(M)_{\mathrm{tors}}(-i)[i], \tag*{(22)}

where each cœuri,j(M)tors\mathrm{cœur}^{i,j}(M)_{\mathrm{tors}} on the right is first placed in internal and cohomological degree zero. In particular, there is an isomorphism of WW-modules

H0Tot⁡(t1H~n(M))≃⨁i+j=ncœuri,j(M)tors,(23)\mathrm{H}^0\operatorname{Tot}(t^1\widetilde{\mathrm{H}}^n(M)) \simeq\bigoplus_{i+j=n}\mathrm{cœur}^{i,j}(M)_{\mathrm{tors}}, \tag*{(23)}

where Tot⁡\operatorname{Tot} is the associated simple complex.

Proof. Ekedahl’s finite-torsion criterion also decomposes a finite-torsion object of Δ\Delta into its ordinary cohomology pieces [5 I, Theorem 1.12]. Insert Proposition 3.2: a piece in cohomological degree −i-i has internal degree ii, which gives precisely the shift (−i)[i](-i)[i]. Each such piece has total degree zero, which proves (23).

Remark 3.4. The isomorphisms of Theorem 1.1 preserve the Raynaud operator FF in each fixed internal degree; no Frobenius twist enters (10). Geometrically, crystalline Frobenius in internal degree ii is piFp^{i}F. This is why (23) is stated only as an identity of WW-modules: it does not identify crystalline Frobenius with the unscaled FF on every summand.

Finite torsion and Artin–Mazur formal groups

In this section XX is a smooth proper scheme over the perfect field kk. Put

MX=RΓ(X,WΩX∙)∈Dcb(R),coeur⁡a,b(X)=coeur⁡a,b(MX),M_X = R\Gamma(X, W\Omega_X^\bullet) \in D_c^b(R), \qquad\operatorname{coeur}^{a,b}(X) = \operatorname{coeur}^{a,b}(M_X),

with torsion, nilpotent and semi-simple parts as in §1.1. Thus coeur⁡0,b(X)\operatorname{coeur}^{0,b}(X) is the coeur of Hb(MX)0=Hb(X,WOX)H^b(M_X)^0 = H^b(X, W\mathcal{O}_X).

For an Artinian kk-algebra BB, the Artin–Mazur functor [1] is

Φeˊti(X,Gm)(B)=ker⁡(Heˊti(XB,Gm)⟶Heˊti(XBred,Gm)).(24)\Phi_{\mathrm{\acute{e}t}}^i(X,\mathbb{G}_m)(B) = \ker\left(H_{\mathrm{\acute{e}t}}^i(X_B,\mathbb{G}_m) \longrightarrow H_{\mathrm{\acute{e}t}}^i(X_{B_{\mathrm{red}}},\mathbb{G}_m)\right). \tag*{(24)}

A functor of this kind is called prorepresentable if it is representable by a formal group. Let

Gi=Φfli(X,Gm)G_i = \Phi_{\mathrm{fl}}^i(X,\mathbb{G}_m)

be the fppf sheafification of Φeˊti(X,Gm)\Phi_{\mathrm{\acute{e}t}}^i(X,\mathbb{G}_m) on the category of Artinian kk-algebras. On each XBX_B, étale and fppf cohomology with coefficients in Gm\mathbb{G}_m agree, so the two functors differ only through this sheafification. The functor GiG_i is always prorepresentable, by a connected commutative formal group with finite-dimensional tangent space [2 Theorem 10.8]. In this section, a formal group is called smooth if it is formally smooth over kk.

The finite obstruction to smoothness

Since kk is perfect, the reduced subgroup (Gi)red(G_i)_{\mathrm{red}} is a smooth formal group, and

Qi=Gi/(Gi)redQ_i = G_i/(G_i)_{\mathrm{red}}

is a finite infinitesimal commutative group scheme; thus GiG_i is smooth if and only if Qi=0Q_i = 0. For instance, G0G_0 is smooth and Q0=0Q_0 = 0: the kk-algebra H0(X,OX)H^0(X,\mathcal{O}_X) is finite étale, and G0=Φeˊt0(X,Gm)G_0 = \Phi_{\mathrm{\acute{e}t}}^0(X,\mathbb{G}_m) is the formal completion at the identity of its group of units.

Let DMk\mathrm{DM}_k be the covariant Dieudonné functor over kk, namely the contravariant Dieudonné module of the Cartier dual. A finite commutative group scheme of pp-power order over kk is canonically the direct sum of three parts, indicated by the subscripts el\mathrm{el}, ll\mathrm{ll} and le\mathrm{le}: the first letter records whether the part is étale or local, and the second does the same for its Cartier dual. No étale part with étale Cartier dual occurs in pp-power order, and the decomposition is compatible with extension of perfect fields. Both FF and VV are nilpotent on DMk\mathrm{DM}_k of a local-local group scheme; FF is bijective and VV nilpotent on DMk\mathrm{DM}_k of a local-étale one; and VV is bijective and FF nilpotent on DMk\mathrm{DM}_k of an étale-local one. Being infinitesimal, QiQ_i is the sum (Qi)ll⊕(Qi)le(Q_i)_{\mathrm{ll}} \oplus(Q_i)_{\mathrm{le}} of its local-local and local-étale parts, and VV is nilpotent on DMk(Qi)\mathrm{DM}_k(Q_i).

Cartier theory identifies this Dieudonné module with the first derived functor of the Cartier module of typical curves TC\mathrm{TC}:

R1TC(Gi)≃DMk(Qi)citep‘[BO21,Remark11.10]‘.(25)R^1\mathrm{TC}(G_i) \simeq\mathrm{DM}_k(Q_i) \qquad{citep}`[BO21, Remark 11.10]`. \tag*{(25)}

In particular, GiG_i is smooth if and only if R1TC(Gi)=0R^1\mathrm{TC}(G_i) = 0.

Theorem 4.1 (Smoothness and finite torsion). For every i≥0i \ge0, there are natural isomorphisms of finite-length Dieudonné modules

cœur⁡0,i+1(X)tors=Hi+1(X,WOX)[V∞]≃R1TC⁡(Gi)≃DM⁡k(Qi).(26)\operatorname{cœur}^{0,i+1}(X)_{\mathrm{tors}} = H^{i+1}(X, W\mathcal{O}_X)[V^\infty] \simeq R^1\operatorname{TC}(G_i) \simeq\operatorname{DM}_k(Q_i). \tag*{(26)}

Consequently, the following conditions are equivalent:

  1. Φfli(X,Gm)\Phi_{\mathrm{fl}}^i(X,\mathbf{G}_m) is smooth;

  2. cœur⁡0,i+1(X)tors=0\operatorname{cœur}^{0,i+1}(X)_{\mathrm{tors}} = 0;

  3. Af2(Hi+1(MX))0A_f^2(H^{i+1}(M_X))^0 and Ass2(Hi+1(MX))0A_{\mathrm{ss}}^2(H^{i+1}(M_X))^0 both vanish;

  4. H0(t1H~i+1(MX))0=0H^0(t^1\widetilde{H}^{i+1}(M_X))^0 = 0.

More precisely,

cœur⁡0,i+1(X)nil≃DM⁡k((Qi)ll),cœur⁡0,i+1(X)ss≃DM⁡k((Qi)le).\operatorname{cœur}^{0,i+1}(X)_{\mathrm{nil}} \simeq\operatorname{DM}_k((Q_i)_{\mathrm{ll}}), \qquad\operatorname{cœur}^{0,i+1}(X)_{\mathrm{ss}} \simeq\operatorname{DM}_k((Q_i)_{\mathrm{le}}).

Proof. Ekedahl’s flat Artin–Mazur comparison [4 III, Proposition 8.1] gives a short exact sequence

0⟶R1TC⁡(Gi)⟶Hi+1(X,WOX)⟶TC⁡(Gi+1)⟶0.(27)0 \longrightarrow R^1\operatorname{TC}(G_i) \longrightarrow H^{i+1}(X,W\mathcal{O}_X) \longrightarrow\operatorname{TC}(G_{i+1}) \longrightarrow0. \tag*{(27)}

By (25), VV is nilpotent on the first term, and it is injective on the Cartier module TC⁡(Gi+1)\operatorname{TC}(G_{i+1}). Hence the first term is precisely the VV-power torsion of the middle term.

We identify this VV-power torsion with the cœur torsion. Nothing arrives in internal degree zero under dd, so the cœur in that degree is the module of stable cycles. If Vmx=0V^m x = 0, then pmx=FmVmx=0p^m x = F^mV^m x = 0 and, for every r≥0r \ge0 and n≥mn \ge m,

dVrx=FndVr+nx=0;dV^r x = F^n dV^{r+n}x = 0;

thus xx is a torsion stable cycle. Conversely, the torsion of the cœur has finite length and VV is topologically nilpotent on it, so VV is nilpotent on it. This proves the equality in (26), and the two isomorphisms follow from (27) and (25).

The covariant Dieudonné functor detects the zero group scheme, so DM⁡k(Qi)=0\operatorname{DM}_k(Q_i) = 0 if and only if Qi=0Q_i = 0, that is, if and only if GiG_i is smooth. The other formulations follow from Theorem 1.1 in bidegree (0,i+1)(0,i+1). Finally, FF is nilpotent on DM⁡k((Qi)ll)\operatorname{DM}_k((Q_i)_{\mathrm{ll}}) and bijective on DM⁡k((Qi)le)\operatorname{DM}_k((Q_i)_{\mathrm{le}}), which gives the last assertion.

Theorem 4.1 concerns only the finite torsion of the cœur: Hi+1(X,WOX)H^{i+1}(X,W\mathcal{O}_X) may contain further pp-torsion coming from dominos. Since G0G_0 is smooth, the case i=0i = 0 gives

cœur⁡0,1(X)tors=H1(X,WOX)[V∞]=0.(28)\operatorname{cœur}^{0,1}(X)_{\mathrm{tors}} = H^1(X,W\mathcal{O}_X)[V^\infty] = 0. \tag*{(28)}

The étale Artin–Mazur functors

The passage from the flat formal group GiG_i to the étale functor Φeˊti(X,Gm)\Phi_{\mathrm{\acute{e}t}}^i(X,\mathbf{G}_m) involves the preceding cohomological degree. The precise criterion is [2 Proposition 10.11]; the sufficient prorepresentability criterion goes back to [1 II, Proposition 1.8 and Corollary 2.11].

Corollary 4.2. Let i≥1i \ge1. The functor Φeˊti(X,Gm)\Phi_{\mathrm{\acute{e}t}}^i(X,\mathbf{G}_m) is prorepresentable if and only if cœur⁡0,i(X)tors=0\operatorname{cœur}^{0,i}(X)_{\mathrm{tors}} = 0, and it is prorepresentable and smooth if and only if

cœur⁡0,i(X)tors=cœur⁡0,i+1(X)tors=0.(29)\operatorname{cœur}^{0,i}(X)_{\mathrm{tors}} = \operatorname{cœur}^{0,i+1}(X)_{\mathrm{tors}} = 0. \tag*{(29)}

Whenever it is prorepresentable, the natural map

Φeˊti(X,Gm)⟶ΦfHi(X,Gm)\Phi_{\mathrm{\acute et}}^{i}(X,\mathbf{G}_{m}) \longrightarrow\Phi_{\mathrm{fH}}^{i}(X,\mathbf{G}_{m})

is an isomorphism. In particular, if all the flat Artin–Mazur formal groups are smooth, then all the étale Artin–Mazur functors are prorepresentable and smooth.

Proof. By [2 Propositions 10.10–10.11], Φeˊti(X,Gm)\Phi_{\mathrm{\acute et}}^{i}(X,\mathbf{G}_{m}) is prorepresentable if and only if Gi−1G_{i-1} is smooth, and this is also equivalent to the comparison map Φeˊti(X,Gm)⟶Gi\Phi_{\mathrm{\acute et}}^{i}(X,\mathbf{G}_{m}) \longrightarrow G_i being an isomorphism. Theorem 4.1, applied to i−1i-1 and to ii, proves the first two assertions. The last one follows because G0=Φeˊt0(X,Gm)G_0=\Phi_{\mathrm{\acute et}}^{0}(X,\mathbf{G}_{m}) is smooth.

Thus smoothness of ΦfHi(X,Gm)\Phi_{\mathrm{fH}}^{i}(X,\mathbf{G}_{m}) alone guarantees prorepresentability of Φeˊti+1(X,Gm)\Phi_{\mathrm{\acute et}}^{i+1}(X,\mathbf{G}_{m}), while prorepresentability and smoothness of Φeˊti(X,Gm)\Phi_{\mathrm{\acute et}}^{i}(X,\mathbf{G}_{m}) are governed by the two adjacent finite-torsion modules in (29). For i=1i=1 and i=2i=2, Φeˊti(X,Gm)\Phi_{\mathrm{\acute et}}^{i}(X,\mathbf{G}_{m}) is the formal Picard group and the formal Brauer group of [1]. By (28), the formal Picard group is always prorepresentable, and the formal Brauer group is prorepresentable if and only if the formal Picard group is smooth, that is, if and only if coeur⁡0,2(X)tors=0\operatorname{coeur}^{0,2}(X)_{\mathrm{tors}}=0.

Mazur–Ogus varieties

Recall that XX is a Mazur–Ogus variety if all the crystalline cohomology groups Hcrisn(X/W)H_{\mathrm{cris}}^{n}(X/W) are torsion-free and the Hodge–de Rham spectral sequence of XX degenerates at E1E_1.

Proposition 4.3. If XX is Mazur–Ogus, then

t~1H~n(MX)=0for all n,coeur⁡a,b(X)tors=0for all a,b.\widetilde{t}^{1}\widetilde{H}^{n}(M_X)=0 \quad\text{for all } n,\qquad\operatorname{coeur}^{a,b}(X)_{\mathrm{tors}}=0 \quad\text{for all } a,b.

Thus all three descriptions of finite torsion in Theorem 1.1 vanish in every bidegree.

Proof. Put An=H~n(MX)A_n=\widetilde{H}^{n}(M_X). By [5 IV, Theorem 1.2], AnA_n is a Mazur–Ogus object of Δ\Delta, so by [5 I, Lemma 4.4] the complex Tot⁡(An)\operatorname{Tot}(A_n) is concentrated in degree zero and H0Tot⁡(An)H^0\operatorname{Tot}(A_n) is torsion-free over WW. For every A∈ΔA\in\Delta, the module Hq(A)aH^q(A)^a contributes to Tot⁡(A)\operatorname{Tot}(A) in degree q+aq+a, so (8) shows that Tot⁡(A)\operatorname{Tot}(A) has cohomology only in degrees 0 and 1; in particular H−1Tot⁡(t0An)=0H^{-1}\operatorname{Tot}(t^0A_n)=0. Applying Tot⁡\operatorname{Tot} to the sequence (9) for AnA_n, we therefore obtain an injection

H0Tot⁡(t1An)⟶H0Tot⁡(An).H^0\operatorname{Tot}(t^1A_n)\longrightarrow H^0\operatorname{Tot}(A_n).

By (23), its source is the finite-length WW-module ⨁a+b=ncoeur⁡a,b(X)tors\bigoplus_{a+b=n}\operatorname{coeur}^{a,b}(X)_{\mathrm{tors}}, whereas its target is torsion-free. Hence coeur⁡a,b(X)tors=0\operatorname{coeur}^{a,b}(X)_{\mathrm{tors}}=0 whenever a+b=na+b=n, and (22) gives t1An=0t^1A_n=0.

Corollary 4.4. Let XX be Mazur–Ogus. For every i≥0i\geq0, the formal group ΦfHi(X,Gm)\Phi_{\mathrm{fH}}^{i}(X,\mathbf{G}_{m}) is smooth, and Φeˊti(X,Gm)\Phi_{\mathrm{\acute et}}^{i}(X,\mathbf{G}_{m}) is prorepresentable, smooth and isomorphic to it. In particular, this holds for every abelian variety over kk.

Proof. For i≥1i\geq1, combine Proposition 4.3, Theorem 4.1 and Corollary 4.2; for i=0i=0, recall that G0=Φeˊt0(X,Gm)G_0=\Phi_{\mathrm{\acute et}}^{0}(X,\mathbf{G}_{m}) is smooth. Abelian varieties have torsion-free crystalline cohomology and an E1E_1-degenerate Hodge–de Rham spectral sequence.

Remark 4.5. The Artin–Mazur conclusion holds under a weaker hypothesis. Over an algebraically closed field, Grammatica [6 Theorem 3.16] proves that torsion-freeness of Hcrisi+1(X/W)H_{\mathrm{cris}}^{i+1}(X/W) alone implies smoothness of ΦfHi(X,Gm)\Phi_{\mathrm{fH}}^{i}(X,\mathbf{G}_{m}). If all the crystalline cohomology groups are torsion-free, Corollary 4.2 therefore shows that every étale Artin–Mazur functor is prorepresentable and smooth, without assuming degeneration of the Hodge–de Rham spectral sequence. The argument of Proposition 4.3 gives in addition the vanishing of all three descriptions of finite torsion in every bidegree.

The three dual pairs for a surface

In this section kk is an arbitrary perfect field of characteristic p>0p > 0, and S/kS/k is a smooth projective geometrically connected surface. Fix an algebraic closure kˉ\bar{k} of kk, which is also a separable closure because kk is perfect, and put

Γk=Gal⁡(kˉ/k),Wˉ=W(kˉ),Sˉ=S×kkˉ.\Gamma_k = \operatorname{Gal}(\bar{k}/k), \qquad\bar{W} = W(\bar{k}), \qquad\bar{S} = S \times_k \bar{k}.

The results of § 1–§ 4 hold over every perfect field; we use them over kk and over kˉ\bar{k}. In particular, § 4 provides

MS∈Dcb(R),MSˉ∈Dcb(R(kˉ)),M_S \in D_c^b(R), \qquad M_{\bar{S}} \in D_c^b(R(\bar{k})),

with Hj(MS)i=Hi(S,WΩSj)H^j(M_S)^i = H^i(S, W\Omega_S^j), and the modules coeur⁡i,j(S)=coeur⁡i,j(MS)\operatorname{coeur}^{i,j}(S) = \operatorname{coeur}^{i,j}(M_S) and coeur⁡i,j(Sˉ)=coeur⁡i,j(MSˉ)\operatorname{coeur}^{i,j}(\bar{S}) = \operatorname{coeur}^{i,j}(M_{\bar{S}}). The group Γk\Gamma_k acts on Sˉ\bar{S} through the factor kˉ\bar{k}, hence by transport of structure on coeur⁡i,j(Sˉ)\operatorname{coeur}^{i,j}(\bar{S}); this action is semilinear with respect to Wˉ\bar{W} and commutes with FF and VV. The Néron–Severi term below is built from the geometric group NS⁡(Sˉ)[p∞]\operatorname{NS}(\bar{S})[p^\infty] together with its Galois action, whereas the two connected Picard terms are Dieudonné modules over kk itself.

The finite Picard quotient and the two duals

Let Pic⁡S/kτ\operatorname{Pic}^{\tau}_{S/k} be the union of the torsion components of the Picard scheme, and let (Pic⁡S/k0)red(\operatorname{Pic}^{0}_{S/k})_{\mathrm{red}} be its reduced identity component, which is an abelian variety because kk is perfect. Let

H=(Pic⁡S/kτ/(Pic⁡S/k0)red)[p∞](30)H = \left(\operatorname{Pic}^{\tau}_{S/k}/(\operatorname{Pic}^{0}_{S/k})_{\mathrm{red}}\right)[p^\infty] \tag*{(30)}

be the pp-primary part of the finite quotient. As in § 4.1, it decomposes canonically, compatibly with extension of perfect fields, as

H=Hel⊕Hll⊕Hle,Hel(kˉ)=NS⁡(Sˉ)[p∞],H = H_{\mathrm{el}} \oplus H_{\mathrm{ll}} \oplus H_{\mathrm{le}}, \qquad H_{\mathrm{el}}(\bar{k}) = \operatorname{NS}(\bar{S})[p^\infty],

where the second equality is an identification of finite Γk\Gamma_k-modules. The identity component of HH is

H∘=Hll⊕Hle=Pic⁡S/k0/(Pic⁡S/k0)red.H^\circ= H_{\mathrm{ll}} \oplus H_{\mathrm{le}} = \operatorname{Pic}^{0}_{S/k}/(\operatorname{Pic}^{0}_{S/k})_{\mathrm{red}}.

Put

Π=NS⁡(Sˉ)[p∞]=Hel(kˉ),Nˉ=Π⊗ZpWˉ,\Pi= \operatorname{NS}(\bar{S})[p^\infty] = H_{\mathrm{el}}(\bar{k}), \qquad\bar{N} = \Pi\otimes_{\mathbb{Z}_p} \bar{W},

where Nˉ\bar{N} carries the operators F=1⊗σF = 1 \otimes\sigma and V=pF−1V = pF^{-1}, and the semilinear action g(x⊗w)=g(x)⊗g(w)g(x \otimes w) = g(x) \otimes g(w) of g∈Γkg \in\Gamma_k, which commutes with FF and VV. With DM⁡k\operatorname{DM}_k as in § 4.1, define the finite-length Dieudonné modules over W=W(k)W = W(k)

N=NˉΓk,L=DM⁡k(Hll),E=DM⁡k(Hle).(31)N = \bar{N}^{\Gamma_k}, \qquad L = \operatorname{DM}_k(H_{\mathrm{ll}}), \qquad E = \operatorname{DM}_k(H_{\mathrm{le}}). \tag*{(31)}

If pnΠ=0p^n\Pi= 0, then Nˉ\bar{N} is a module over Wn(kˉ)W_n(\bar{k}), and the action of Γk\Gamma_k on Π\Pi factors through the Galois group of a finite Galois extension of kk. Finite Galois descent therefore gives a canonical isomorphism, compatible with FF, VV and Γk\Gamma_k,

N⊗WWˉ≃Nˉ.(32)N \otimes_W \bar{W} \simeq\bar{N}. \tag*{(32)}

In general NN is not ΠΓk⊗ZpW\Pi^{\Gamma_k} \otimes_{\mathbb{Z}_p} W: taking invariants of Π\Pi alone loses a nontrivial Galois action. When kk is algebraically closed, N=NS⁡(S)[p∞]⊗ZpWN = \operatorname{NS}(S)[p^\infty] \otimes_{\mathbb{Z}_p} W.

Both FF and VV are nilpotent on LL, while on EE the operator FF is bijective and VV is nilpotent. Since Dieudonné theory over kk is an equivalence of categories, LL and EE determine HllH_{\mathrm{ll}} and HleH_{\mathrm{le}} as group schemes over kk, including their descent data from kˉ\bar{k}. With the covariant convention, NN is not DMk(Het)\mathrm{DM}_k(H_{\mathrm{et}}), on which VV rather than FF is bijective: it is the semi-simple module attached to the Galois module Π\Pi of components.

Each dual below has underlying WW-module Hom⁡W(D,K/W)\operatorname{Hom}_W(D,K/W), but its operators depend on the type of DD. For a semi-simple Dieudonné module DD of finite length, the unit dual D∨uD^{\vee u} is defined by

(Fλ)(x)=σ(λ(F−1x)),V=pF−1.(33)(F\lambda)(x)=\sigma\bigl(\lambda(F^{-1}x)\bigr),\qquad V=pF^{-1}. \tag*{(33)}

For a Dieudonné module DD of finite length on which FF and VV are nilpotent, the Dieudonné dual D∨DD^{\vee D} is defined by

(Fλ)(x)=σ(λ(Vx)),(Vλ)(x)=σ−1(λ(Fx)).(34)(F\lambda)(x)=\sigma\bigl(\lambda(Vx)\bigr),\qquad(V\lambda)(x)=\sigma^{-1}\bigl(\lambda(Fx)\bigr). \tag*{(34)}

Both operations commute with extension of perfect fields. In particular,

N∨u≃(Π∨⊗ZpW‾)Γk,N^{\vee u}\simeq\left(\Pi^\vee\otimes_{\mathbb{Z}_p}\overline{W}\right)^{\Gamma_k},

where Π∨=Hom⁡Zp(Π,Qp/Zp)\Pi^\vee=\operatorname{Hom}_{\mathbb{Z}_p}(\Pi,\mathbb{Q}_p/\mathbb{Z}_p) carries the contragredient action (gλ)(x)=λ(g−1x)(g\lambda)(x)=\lambda(g^{-1}x) and FF is induced by 1⊗σ1\otimes\sigma. The two superscripts are essential: (33) and (34) are different duality operations.

The classification

Theorem 5.1 (Finite torsion of a surface over a perfect field). Let NN, LL and EE be as in (31). The only possibly nonzero finite-torsion pieces are the following three dual pairs of Dieudonné modules over W(k)W(k), where (i,j)(i,j) denotes the position Hj(S,WΩSi)H^j(S,W\Omega_S^i):

PartFirst position and moduleDual position and moduleOrigin
semi-simple(1,1)(1,1): NN(1,2)(1,2): N∨uN^{\vee u}NS⁡(Sˉ)[p∞]\operatorname{NS}(\bar{S})[p^\infty]
nilpotent(0,2)(0,2): LL(1,2)(1,2): L∨DL^{\vee D}HllH_{\mathrm{ll}}
semi-simple(0,2)(0,2): EE(2,1)(2,1): E∨uE^{\vee u}HleH_{\mathrm{le}}

Table 5.1.

All pieces not listed are zero. The same statements hold for Af2A_f^2 and Ass1A_{\mathrm{ss}}^1, for the nilpotent and semi-simple parts of cœur torsion, and for the modules H−i(t1Hi+j(MS))H^{-i}(t^1\mathop{H}\limits^{i+j}(M_S)) of Theorem 1.1.

The positions of the three pairs overlap. Equivalently, the complete table of cœur torsion is

cœur⁡i,j(S)tors\operatorname{cœur}^{i,j}(S)_{\mathrm{tors}}i=0i=0i=1i=1i=2i=2
j=2j=2L⊕EL\oplus EL∨D⊕N∨uL^{\vee D}\oplus N^{\vee u}00
j=1j=100NNE∨uE^{\vee u}
j=0j=0000000

Table 5.2.

Proof. We first compare semi-simple parts over kk and over kˉ\bar{k}, then identify the three initial modules, and finally apply duality over kk.

Base change and descent. Only the semi-simple parts need to be compared with their analogues over kˉ\bar{k}. By Theorem 1.1 and (6), cœur⁡i,j(M)ss≃A1(Hj(M))tors\operatorname{cœur}^{i,j}(M)_{\mathrm{ss}}\simeq A^1(H^j(M))_{\mathrm{tors}} for M=MSM=M_S and for M=MSˉM=M_{\bar{S}}. Ekedahl’s completed pullback along k→kˉk\to\bar{k}, from Dcb(R)D^b_c(R) to Dcb(R(kˉ))D^b_c(R(\bar{k})), is exact and conservative on coherent modules [4 III, Lemma 1.4]; it carries MSM_S to MSˉM_{\bar{S}}, and it commutes with internal derived Hom and with the duality functor DD [4 I, §9]. It therefore carries the biduality spectral sequence of Hj(MS)H^j(M_S) to that of Hj(MSˉ)H^j(M_{\bar{S}}), and hence A1(Hj(MS))A^1(H^j(M_S)) to A1(Hj(Mk‾))A^{1}(H^{j}(M_{\overline{k}})). The modules A1(Hj(M))iA^{1}(H^{j}(M))^{i} are finitely generated over WW [3 IV, Theorem 3.3], and on such modules completed pullback is ordinary extension of scalars, which preserves torsion submodules. Hence there are natural isomorphisms

cœuri,j(S)ss⊗WW‾≃cœuri,j(S‾)ss,(35)\mathrm{cœur}^{i,j}(S)_{\mathrm{ss}} \otimes_{W} \overline{W} \simeq\mathrm{cœur}^{i,j}(\overline{S})_{\mathrm{ss}}, \tag*{(35)}

compatible with FF, VV and the semilinear actions of Γk\Gamma_k. No base change statement is needed for the whole Hodge–Witt modules, which need not be finitely generated over WW. As (D⊗WW‾)Γk=D(D \otimes_{W} \overline{W})^{\Gamma_k}=D for every WW-module DD of finite length, it follows that cœuri,j(S)ss≃(cœuri,j(S‾)ss)Γk\mathrm{cœur}^{i,j}(S)_{\mathrm{ss}}\simeq(\mathrm{cœur}^{i,j}(\overline{S})_{\mathrm{ss}})^{\Gamma_k}.

The Néron–Severi module. We first work over k‾\overline{k}. For every n≥1n\geq1, the Kummer sequence gives an exact sequence

0⟶Pic⁡(S‾)/pn⟶Heˊt2(S‾,μpn)⟶Br⁡(S‾)[pn]⟶0.0 \longrightarrow\operatorname{Pic}(\overline{S})/p^{n} \longrightarrow H^{2}_{\mathrm{\acute{e}t}}(\overline{S},\mu_{p^{n}}) \longrightarrow\operatorname{Br}(\overline{S})[p^{n}] \longrightarrow0.

The group Pic⁡0(S‾)\operatorname{Pic}^{0}(\overline{S}) is the group of k‾\overline{k}-points of the abelian variety (Pic⁡S/k‾0)red(\operatorname{Pic}^{0}_{S/\overline{k}})_{\mathrm{red}}, so it is divisible and Pic⁡(S‾)/pn=NS⁡(S‾)/pn\operatorname{Pic}(\overline{S})/p^{n}=\operatorname{NS}(\overline{S})/p^{n}. The transition maps of this system are surjective, so passing to the limit gives an exact sequence

0⟶NS⁡(S‾)⊗ZZp⟶Heˊt2(S‾,Zp(1))⟶TpBr⁡(S‾)⟶0,(36)0 \longrightarrow\operatorname{NS}(\overline{S})\otimes_{\mathbb{Z}}\mathbb{Z}_{p} \longrightarrow H^{2}_{\mathrm{\acute{e}t}}(\overline{S},\mathbb{Z}_{p}(1)) \longrightarrow T_{p}\operatorname{Br}(\overline{S}) \longrightarrow0, \tag*{(36)}

where Heˊt2(S‾,Zp(1))=lim←⁡nHeˊt2(S‾,μpn)H^{2}_{\mathrm{\acute{e}t}}(\overline{S},\mathbb{Z}_{p}(1))=\varprojlim_{n}H^{2}_{\mathrm{\acute{e}t}}(\overline{S},\mu_{p^{n}}) and TpBr⁡(S‾)=lim←⁡nBr⁡(S‾)[pn]T_{p}\operatorname{Br}(\overline{S})=\varprojlim_{n}\operatorname{Br}(\overline{S})[p^{n}], the transition maps of the last system being multiplication by pp. A Tate module is torsion-free: if pr(xn)n=0p^{r}(x_{n})_{n}=0, then xn=prxn+r=0x_{n}=p^{r}x_{n+r}=0 for every nn. Since NS⁡(S‾)\operatorname{NS}(\overline{S}) is finitely generated, (36) gives

Heˊt2(S‾,Zp(1))tors≃NS⁡(S‾)[p∞]=Π.(37)H^{2}_{\mathrm{\acute{e}t}}(\overline{S},\mathbb{Z}_{p}(1))_{\mathrm{tors}}\simeq\operatorname{NS}(\overline{S})[p^{\infty}]=\Pi. \tag*{(37)}

The map dlog⁡d\log induces an isomorphism Heˊt2(S‾,Zp(1))≃H1(S‾,WΩS‾,log⁡1)H^{2}_{\mathrm{\acute{e}t}}(\overline{S},\mathbb{Z}_{p}(1))\simeq H^{1}(\overline{S},W\Omega^{1}_{\overline{S},\log}) [7 I, Corollary 3.27]. By the theorem of Illusie and Raynaud on logarithmic cohomology [8 IV, Theorem 3.3], the natural map

H1(S‾,WΩS‾,log⁡1)⟶cœur1,1(S‾)F=1H^{1}(\overline{S},W\Omega^{1}_{\overline{S},\log})\longrightarrow\mathrm{cœur}^{1,1}(\overline{S})^{F=1}

is surjective, and its kernel is the group of k‾\overline{k}-points of a connected unipotent group of dimension T0,1T^{0,1}, the domino number of the domino H1(S‾,WOS‾)/V−∞Z0→F∞B1H^{1}(\overline{S},W\mathcal{O}_{\overline{S}})/V^{-\infty}Z^{0}\to F^{\infty}B^{1} arriving in bidegree (1,1)(1,1). Ekedahl’s duality for domino numbers [3 IV, Corollary 3.5.1] gives T0,1=T0,3T^{0,1}=T^{0,3}, and T0,3=0T^{0,3}=0 because H3(MS‾)=0H^{3}(M_{\overline{S}})=0. Hence the map above is an isomorphism. Taking torsion subgroups and using (37), we obtain

Π≃(cœur1,1(S‾)tors)F=1=(cœur1,1(S‾)ss)F=1,\Pi\simeq\bigl(\mathrm{cœur}^{1,1}(\overline{S})_{\mathrm{tors}}\bigr)^{F=1}=\bigl(\mathrm{cœur}^{1,1}(\overline{S})_{\mathrm{ss}}\bigr)^{F=1},

because the nilpotent part has no nonzero FF-fixed elements. For a semi-simple Dieudonné module DD of finite length over W‾\overline{W}, the natural map DF=1⊗ZpW‾→DD^{F=1}\otimes_{\mathbb{Z}_{p}}\overline{W}\to D is an isomorphism: if pD=0pD=0 this is Lang’s theorem, and the general case follows by dévissage, since 1−F1-F is surjective on such modules, again by Lang’s theorem. Therefore

cœur1,1(S‾)ss≃Π⊗ZpW‾=N‾.\mathrm{cœur}^{1,1}(\overline{S})_{\mathrm{ss}}\simeq\Pi\otimes_{\mathbb{Z}_{p}}\overline{W}=\overline{N}.

All the maps involved are natural with respect to automorphisms of k‾\overline{k} over kk, so this isomorphism is Γk\Gamma_k-equivariant as well as compatible with FF and VV. Taking Γk\Gamma_k-invariants and using the base change step, we obtain

cœur1,1(S)ss≃N‾Γk=N.(38)\mathrm{cœur}^{1,1}(S)_{\mathrm{ss}}\simeq\overline{N}^{\Gamma_k}=N. \tag*{(38)}

The connected finite Picard module. By (28) and Corollary 4.2, the functor Φeˊt1(S,Gm)\Phi^{1}_{\mathrm{\acute{e}t}}(S,\mathbf{G}_{m}) is representable and equal to G1=Φfppfˊ1(S,Gm)G_{1}=\Phi^{1}_{\mathrm{\acute{fppf}}}(S,\mathbf{G}_{m}); it is the formal Picard group of SS, the formal completion of Pic⁡S/k\operatorname{Pic}_{S/k} at the origin [1 II, §4]. Its reduced subgroup is the formal completion of the abelian variety (Pic⁡S/k0)red(\operatorname{Pic}^{0}_{S/k})_{\mathrm{red}}, and the quotient is the infinitesimal group scheme Pic⁡S/k0/(Pic⁡S/k0)red\operatorname{Pic}^{0}_{S/k}/(\operatorname{Pic}^{0}_{S/k})_{\mathrm{red}}; thus Q1=H∘Q_{1}=H^{\circ}. Theorem 4.1 with X=SX=S and i=1i=1 therefore gives, over kk itself,

cœur⁡0,2(S)tors=H2(S,WOS)[V∞]≃DMk(H∘)=L⊕E,(39)\operatorname{cœur}^{0,2}(S)_{\mathrm{tors}}=\mathrm{H}^{2}(S,W\mathcal{O}_{S})[V^{\infty}]\simeq\mathrm{DM}_{k}(H^{\circ})=L\oplus E, \tag*{(39)}

with cœur⁡0,2(S)nil≃L\operatorname{cœur}^{0,2}(S)_{\mathrm{nil}}\simeq L and cœur⁡0,2(S)ss≃E\operatorname{cœur}^{0,2}(S)_{\mathrm{ss}}\simeq E.

Duality and vanishing. For a smooth proper geometrically connected variety X/kX/k of dimension NN, Ekedahl’s geometric duality over the perfect field kk [3 IV, Theorem 3.5] (see also [4 II, Theorem 1.1(ii)]), combined with the formulas of [3 IV, Theorem 3.3] and Theorem 1.1, reads

cœur⁡i,j(X)ss≃(cœur⁡N−i,N+1−j(X)ss)∨u,(40)\operatorname{cœur}^{i,j}(X)_{\mathrm{ss}}\simeq\left(\operatorname{cœur}^{N-i,N+1-j}(X)_{\mathrm{ss}}\right)^{\vee_{u}}, \tag*{(40)}
cœur⁡i,j(X)nil≃(cœur⁡N−1−i,N+2−j(X)nil)∨D;(41)\operatorname{cœur}^{i,j}(X)_{\mathrm{nil}}\simeq\left(\operatorname{cœur}^{N-1-i,N+2-j}(X)_{\mathrm{nil}}\right)^{\vee_{D}}; \tag*{(41)}

the internal shifts in Ekedahl’s formulas are already built into these indices, and the duals carry the operators (33) and (34). For N=2N=2, the first formula pairs (1,1)(1,1) with (1,2)(1,2) and (0,2)(0,2) with (2,1)(2,1), and the second pairs (0,2)(0,2) with (1,2)(1,2). Together with (38) and (39), this gives all the modules in the table.

It remains to see that no other position contributes. Cohomology vanishes outside 0≤i,j≤20\leq i,j\leq2. The groups H0(S,WΩSi)\mathrm{H}^{0}(S,W\Omega^{i}_{S}) are pp-torsion-free, because the sheaves WΩSiW\Omega^{i}_{S} are [7 I, §3]; by Proposition 2.2, their cœurs are torsion-free. Next, cœur⁡0,1(S)tors=0\operatorname{cœur}^{0,1}(S)_{\mathrm{tors}}=0. This is (28) for X=SX=S, but it can also be seen directly over kk. Since SS is proper and geometrically connected, the reduction map H0(S,WOS)=W(k)→H0(S,OS)=k\mathrm{H}^{0}(S,W\mathcal{O}_{S})=W(k)\to\mathrm{H}^{0}(S,\mathcal{O}_{S})=k is surjective, so the exact sequence

0⟶WOS→VWOS⟶OS⟶00\longrightarrow W\mathcal{O}_{S}\xrightarrow{V}W\mathcal{O}_{S}\longrightarrow\mathcal{O}_{S}\longrightarrow0

shows that VV is injective on H1(S,WOS)\mathrm{H}^{1}(S,W\mathcal{O}_{S}). As nothing arrives in internal degree zero under dd, the module cœur⁡0,1(S)tors\operatorname{cœur}^{0,1}(S)_{\mathrm{tors}} is a submodule of H1(S,WOS)\mathrm{H}^{1}(S,W\mathcal{O}_{S}); it has finite length and VV is topologically nilpotent on it, so it is killed by a power of VV and therefore vanishes. For the nilpotent part, (41) makes every position with j=1j=1 dual to one with j=3j=3, and every position with i=2i=2 dual to one with i=−1i=-1; all of these vanish, leaving only (0,2)(0,2) and (1,2)(1,2). For the semi-simple part, the vanishing at (0,1)(0,1) gives vanishing at (2,2)(2,2) by (40), leaving exactly the four positions listed.

Corollary 5.2. For a surface SS as above, tH~n(MS)=0{}^{t}\widetilde{\mathrm{H}}^{n}(M_{S})=0 unless n=2n=2 or 33, and

tH~2(MS)≃(L⊕E)⊕N(−1)[1],{}^{t}\widetilde{\mathrm{H}}^{2}(M_{S})\simeq(L\oplus E)\oplus N(-1)[1],
tH~3(MS)≃(L∨D⊕N∨u)(−1)[1]⊕E∨u(−2)[2].{}^{t}\widetilde{\mathrm{H}}^{3}(M_{S})\simeq(L^{\vee_{D}}\oplus N^{\vee_{u}})(-1)[1]\oplus E^{\vee_{u}}(-2)[2].

Thus the finite-torsion contributions in total degrees 2 and 3 are the WW-modules

L⊕E⊕NandL∨D⊕N∨u⊕E∨u.L\oplus E\oplus N\qquad\text{and}\qquad L^{\vee_{D}}\oplus N^{\vee_{u}}\oplus E^{\vee_{u}}.

These statements concern diagonal finite torsion only; they say nothing about the torsion carried by dominos.

Proof. Insert (35) into Corollary 3.3. The only nonzero entries have i+j=2i+j=2 or 33.

Remark 5.3. By Corollary 4.2 and (39), the formal Brauer group Φeˊt2(S,Gm)\Phi^{2}_{\mathrm{\acute{e}t}}(S,\mathbb{G}_{m}) of the surface SS is representable if and only if L=E=0L=E=0, that is, if and only if H∘=0H^{\circ}=0, i.e. Pic⁡S/k\operatorname{Pic}_{S/k} is reduced. It is then also smooth, since H3(S,WOS)=0\mathrm{H}^{3}(S,W\mathcal{O}_{S})=0.

Declarations

Relation to the author’s thesis. Part of this work appeared, over algebraically closed fields, in the author’s PhD thesis [9]; the present paper gives complete proofs and works over an arbitrary perfect field.

Funding. Y.Y. is partially supported by a grant from the National Natural Science Foundation of China (NSFC-12231001).

Conflict of interest. The author declares that there is no conflict of interest.

Use of AI tools. To turn part of the author’s thesis and proofs into a self-contained article, ChatGPT 6 Astra (OpenAI) was used to generate a first draft of this paper from this material. Claude Code (Anthropic), with the Claude Opus 5.5 model, was then used to revise the draft into the present manuscript, including the exposition and the written form of the proofs. The author has checked all arguments and takes full responsibility for the content of this paper.

References

  1. [1]M. Artin and B. Mazur, Formal groups arising from algebraic varieties, Ann. Sci. École Norm. Sup. (4) 10 (1977), 87–131. doi:10.24033/asens.1322.DOI
  2. [2]D. Bragg and M. Olsson, Representability of cohomology of finite flat abelian group schemes, preprint, 2021, arXiv:2107.11492v1. References to §§10–11 refer to this version.arxiv.org/abs/2107.11492
  3. [3]T. Ekedahl, On the multiplicative properties of the de Rham–Witt complex. I, Ark. Mat. 22 (1984), 185–239. doi:10.1007/BF02384380.DOI
  4. [4]T. Ekedahl, On the multiplicative properties of the de Rham–Witt complex. II, Ark. Mat. 23 (1985), 53–102. doi:10.1007/BF02384419.DOI
  5. [5]T. Ekedahl, Diagonal complexes and F-gauge structures, Travaux en Cours, vol. 18, Hermann, Paris, 1986.
  6. [6]L. Grammatica, Formal smoothness of the Artin–Mazur formal groups, preprint, 2025, arXiv:2510.03001.arxiv.org/abs/2510.03001
  7. [7]L. Illusie, Complexe de de Rham–Witt et cohomologie cristalline, Ann. Sci. École Norm. Sup. (4) 12 (1979), 501–661. doi:10.24033/asens.1374.DOI
  8. [8]L. Illusie and M. Raynaud, Les suites spectrales associées au complexe de de Rham–Witt, Publ. Math. IHÉS 57 (1983), 73–212. doi:10.1007/BF02698774.DOI
  9. [9]Y. Yang, p-primary torsion of the Brauer group in characteristic p, PhD thesis, Imperial College London, 2025, doi:10.25560/124274.DOI

Paper details

vlnc:W_3nrf2xvm
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

ChatGPT 6 Astra, Claude Code

View statement
Looking for related papers…
Contents