Section 10.4 · PDF pages 767–769
The integral polarized retract
The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.
10.4 The integral polarized retract
Lemma 10.19 (A polarization and an integral positive quasi-inverse). Let A/S be a projective abelian scheme with relatively ample line bundle L, and suppose S is connected. Its polarization λ : A →A^{∨} has
constant finite locally free degree n ≥1. There is a relatively ample line bundle N on A^{∨} whose polarization µ : A^{∨}→A satisfies
µλ = [n]_{A}, λµ = [n]_{A}∨. (AAG.4)
Proof. Take N = Norm_{λ} L. The norm for a finite locally free map is defined on functions by the deter-minant of multiplication. Applying it to the transition units of a line bundle defines its norm line bundle. Determinants show directly that it commutes with base change, tensor product, and pullback from the target, with Norm_{λ}(λ^{∗}Q) = Q^{n}. These are the canonical norm and induced line norm of [14, Tags 0BD2,0BCY]. The norm of an ample line is ample; the section proof is [14, Tag 0BD0]. Applied after arbitrary affine base restriction it gives relative ampleness of N.
On the dual abelian schemes this norm induces a homomorphism u : Pic^{0}(A/S) → Pic^{0}(A^{∨}/S). The preceding power identity gives uλ^{∨}= [n]. A polarization is symmetric, so λ^{∨}= λ under biduality. For any scheme-valued point x of A, translation by x above λ and translation by λx below it form a cartesian square. Base-change compatibility of the norm therefore gives
φ_{N}(λx) = Norm_{λ}(t^{∗}_{x}L ⊗L^{−1}) = u(λx) = [n]x.
Since λ is faithfully flat, this equality determines φ_{N} = u. Put µ = φ_{N}. We have µλ = [n]. Composing with λ and cancelling its faithfully flat right factor gives λµ = [n] as well. All arguments concern group schemes, not just geometric points, so include inseparable polarizations.
Proposition 10.20 (The full integral Zarhin input). Let C be algebraically closed, let R be a valuation ring in C, and let A/R be an abelian scheme with a relatively ample line bundle. Then A is a group-scheme retract of an abelian scheme equipped with a rigidified symmetric relatively ample line having trivial translation kernel. This supplies every field of the source’s ZarhinStatement.
Proof. The base Spec R is connected. Apply Lemma 10.19, and write n for the resulting positive integer. By Lagrange’s four-square theorem choose integers a, b, c, d with a^{2}+b^{2}+c^{2}+d^{2}= n−1; if n = 1 take all four equal to zero. The precise elementary theorem and its lattice proof are given in [15, Remark 4.20,pp. 76–77]. Let
Selectable PDF notation • transcription unchecked
(a −b −c −d)
((b a −d c ))
^{Q =}(c d a −b)^{.}
d −c b aMultiplying rows shows that each row has squared length n −1 and any two different rows have scalar product zero. Thus QQ^{t}= Q^{t}Q = (n −1)I_{4}. All its entries act by integer multiplication on the appropriate abelian schemes and hence commute with λ and µ.
Put B = A^{4}× (A^{∨})^{4} and identify B^{∨}= (A^{∨})^{4}× A^{4}. Consider the symmetric homomorphism
Selectable PDF notation • transcription unchecked
(λI Q) ( µI −Q)
Λ =_{t}^{4} , Λ^{−1}=^{4}_{t} . (AAAG.1)
Selectable PDF notation • transcription unchecked
Q µI_{4} −Q λI_{4}The two products are the identity: their diagonal blocks are nI_{4}−QQ^{t} and nI_{4}−Q^{t}Q, and their off-diagonal blocks cancel. In particular Λ is an isomorphism over R itself.
It is induced by a line bundle M. Take the exterior product of the four copies of L on the first factor and of N on the second, and tensor it with the Poincaré line pulled back by (x, y) →→(x, Qy). The defining translation law of the Poincaré biextension gives the off-diagonal blocks Q, Q^{t}; the other two lines give the diagonal blocks. Thus φ_{M} = Λ. To check positivity rather than infer it from symmetry, consider the finite faithfully flat isogeny
Selectable PDF notation • transcription unchecked
F : B −→(A^{∨})^{4}× (A^{∨})^{4}, F(x, y) = (λx + Qy, y).The polarization of the pullback of the product N^{⊠8} is
Selectable PDF notation • transcription unchecked
( λ 0) (µ 0) (λ Q) ( nλ nQ)
Q^{t} 1 0 µ 0 1^{=} nQ^{t} nµ^{= nΛ.}Here each displayed scalar block means four identical diagonal entries. Consequently on every geometric fiber M^{n} is algebraically equivalent to an ample line, the indicated finite pullback. Ampleness of a line on an abelian variety depends only on its algebraic-equivalence class, and a line is ample if a positive power is ample. Hence M is relatively ample. This is the quaternion argument with its positivity check made over the integral base; the usual field construction is recorded in [16, (11.28)–(11.29),pp. 170–172].
We still need a symmetric line representative, not merely the symmetric principal homomorphism. For an abelian scheme and a polarization Λ, the sheaf of symmetric representatives of its Néron–Severi class is a torsor under B^{∨}[2]. The exact sequence is [17, (3.7),p. 19]; it is fppf exact, so has no assumption that 2 be invertible. This torsor is a finite locally free, hence proper, algebraic space over R (in fact finite over the scheme base, hence a scheme). Its generic fiber has a C-point, since C is algebraically closed. The valuative criterion of properness extends this point to an R-point. Here Frac R = C: for every nonzero z ∈C either z or z^{−1} belongs to the valuation ring. The Picard functor of B/R has no line-bundle obstruction, because the zero section rigidifies it; see [17, §2.3,pp. 6–7]. Thus this point is represented by a rigidified symmetric line M^{′} on B. It induces the same Λ and is relatively ample, since its geometric fiber classes equal those of
Selectable PDF notation • transcription unchecked
M.
The group scheme B is proper and smooth with geometrically connected fibers. Its relative dimension is constant on the connected base, giving the source integer g. The above rigidification is exactly the required generator at the zero section, and the symmetry isomorphism supplies its locally trivial line difference. Rela-tive ampleness has the source’s explicit chart form: over an affine open of the base, a sufficiently high power gives a closed projective immersion; the coordinate-section nonzero loci are affine and cover, and their ratios generate their coordinate rings. This gives both the cover and the surjective polynomial-evaluation map in RelAmple.
Finally, if a scheme-valued point x has translation difference t^{∗}_{x}M^{′} ⊗(M^{′})^{−1} trivial locally over that scheme, its relative Picard class is zero. This class is φ_{M}′(x) = Λ(x). The inverse in (AAAG.1) forces x = 0. This is precisely KTriv, including its quantification over arbitrary affine test schemes. Inclusion of the first copy of A into B and projection onto that copy are group homomorphisms whose composite is the identity. Taking P = (B, M^{′}) and these two maps proves the stated retract. No characteristic or discrete-valuation restriction has been introduced.