Using a cohomological invariant of symplectic involutions, Garibaldi–Parimala–Tignol define in [14 §11] a cohomological invariant Δ \Delta Δ of central simple algebras of degree 8 and exponent 2 over arbitrary fields of characteristic different from 2. This invariant vanishes when the algebra decomposes into a tensor product of three quaternion algebras. Examples where it does not vanish were given by Barry in [7 ] ; Merkurjev–Neshitov–Zainoulline showed that it is nontrivial generically [22 §4a] . The purpose of the present paper is to describe a similar construction in characteristic 2, based on the discriminant Pfister form of symplectic involutions from [11 ] or [27 ] . Being rooted in quadratic forms, the invariant Δ \Delta Δ in Definition 1.3 below takes values in a quotient of the group I q 3 I_q^3 I q 3 from the fundamental filtration of the Witt group of the base field. Kato’s isomorphisms from [18 ] allow an equivalent formulation in terms of Kato–Milne cohomology.
While the invariant Δ \Delta Δ yields a necessary criterion for decomposability of central simple algebras of degree 8 and exponent 2, variants of the construction using discriminant Pfister forms yield necessary and sufficient criteria for the descent of biquaternion algebras over quadratic extensions, and for the descent of quaternion algebras over biquadratic extensions that are not
Key words and phrases. Incomposable division algebras, symplectic involutions, quadratic forms over discretely valued fields.
The authors acknowledge support from Wallonie–Bruxelles International and the French government in the framework of “Partenariats Hubert Curien” (Project 2-algèbres de degré 8 et d’exposant 2 ). Research leading to this work was initiated while the second author was visiting the Sorbonne Paris Nord university, whose hospitality is gratefully acknowledged. purely inseparable. These variants are an additional theme of the paper; they play a central role in the construction of examples showing that the invariant Δ \Delta Δ is not trivial.
The definition of the invariant Δ \Delta Δ is given in Section 1, see Definition 1.3. It relies on the proof that the discriminant 3-fold Pfister forms of all the symplectic involutions on a central simple algebra of degree 8 have the same image in a suitable quotient of the group I q 3 I_q^3 I q 3 of the center.
Section 2 discusses a necessary and sufficient condition for a 2-torsion Brauer class over a quadratic extension to descend to a 2-torsion Brauer class over the base field, i.e., to be obtained by scalar extension from a 2-torsion Brauer class over the base field. This condition is expressed by the vanishing of the corestriction when the extension is separable, and of a canonical map induced by the squaring map when it is not, see Theorem 2.1. The result is well-known when the characteristic is different from 2. For Brauer classes of quaternion algebras, a similar condition is given for the descent from a biquadratic extension that is not purely inseparable, see Theorem 2.3.
Even when the Brauer class of a biquaternion algebra over a quadratic extension descends, the algebra itself may not descend: it is Brauer-equivalent to the scalar extension of a central simple algebra defined over the base field, but this algebra may have larger degree. Given a quadratic extension K K K of a base field F F F of characteristic 2, section 3 introduces an invariant δ K / F \delta_{K/F} δ K / F of biquaternion algebras over K K K , whose vanishing characterizes biquaternion algebras that descend, i.e., which are isomorphic to the scalar extension of a biquaternion algebra over F F F , see Definition 3.6 and Proposition 3.7. The definition of δ K / F \delta_{K/F} δ K / F follows the same pattern as for the Δ \Delta Δ -invariant: it relies on the result that the given biquaternion algebra B B B over K K K is contained in a central simple F F F -algebra A A A of degree 8, and that all the discriminant Pfister forms of symplectic involutions on A A A that restrict to a symplectic involution on B B B have the same image in a suitable quotient of I q 3 ( F ) I_q^3(F) I q 3 ( F ) , independently of the choice of the algebra A A A , see Definition 3.6. This invariant turns out to be the same as the one defined in cohomological terms by Barry–Chapman–Laghribi [8 ] when K K K is separable over F F F , and by Laghribi–Lorenz [20 §6] when K K K is inseparable over F F F (and by Barry [7 ] when the characteristic is not 2), although the earlier constructions did not rely on the choice of an algebra A A A containing B B B .
A similar strategy is used in Section 4 to define an invariant δ L / F \delta_{L/F} δ L / F of quaternion algebras over a biquadratic extension L L L of a base field F F F of characteristic 2, assuming L L L is not purely inseparable over F F F : when the Brauer class of the quaternion algebra Q Q Q descends from L L L to F F F , the algebra Q Q Q can be embedded in a central simple F F F -algebra A A A of degree 8, and all the discriminant Pfister forms of involutions on A A A that extend the canonical conjugation involution on Q Q Q have the same image in a suitable quotient of I q 3 ( F ) I_q^3(F) I q 3 ( F ) , independently of the choice of A A A . This common image defines an invariant δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) , which vanishes if and only if Q Q Q descends to a quaternion algebra over F F F , see Definition 4.4 and Theorem 4.5. With a completely different method, this invariant has also been defined by Laghribi–Lorenz [20 ] (and by Sivatski [25 ] when the characteristic is not 2), although these authors consider it as an invariant of the central simple F F F -algebras that contain Q Q Q . The purely inseparable case is considered in the appendix. The situation there is quite different: a quaternion algebra over a purely inseparable biquadratic field extension descends if and only if its Brauer class descends, see Theorem A.1.
As this rough sketch makes clear, the various invariants defined in this paper are closely related: if a central simple F F F -algebra A A A of degree 8 and exponent 2 contains a biquadratic field extension L L L of F F F that is not purely inseparable, and K ⊂ L K \subset L K ⊂ L is a quadratic extension of F F F , then the centralizer of L L L in A A A is a quaternion L L L -algebra Q Q Q and the centralizer of K K K in A A A is a biquaternion K K K -algebra B B B . The following inclusions are clear from the definitions:
F ⊂ K ⊂ L ⊂ Q ⊂ B ⊂ A . F \subset K \subset L \subset Q \subset B \subset A. F ⊂ K ⊂ L ⊂ Q ⊂ B ⊂ A .
Canonical homomorphisms induced by the inclusions carry the invariants as follows:
δ L / F ( Q ) ↦ δ K / F ( B ) ↦ Δ ( A ) . \delta_{L/F}(Q) \mapsto\delta_{K/F}(B) \mapsto\Delta(A). δ L / F ( Q ) ↦ δ K / F ( B ) ↦ Δ ( A ) .
Therefore, examples where Δ ( A ) ≠ 0 \Delta(A) \ne0 Δ ( A ) = 0 yield examples of biquaternion algebras B B B such that δ K / F ( B ) ≠ 0 \delta_{K/F}(B) \ne0 δ K / F ( B ) = 0 and of quaternion algebras Q Q Q with δ L / F ( Q ) ≠ 0 \delta_{L/F}(Q) \ne0 δ L / F ( Q ) = 0 . We assume throughout that the characteristic is 2, but the reader may verify that all the results in Sections 1–4 are valid in arbitrary characteristic, with only minor alterations in the proofs involving separable field extensions.
Section 5 exhibits examples where the invariants do not vanish, using residue maps for quadratic forms as the main tool. A first construction takes as a starting ingredient any central simple F F F -algebra A A A of degree 8 and exponent 2 that does not decompose into a tensor product of quaternion algebras. The δ K / F \delta_{K/F} δ K / F -invariant of the centralizer of any separable quadratic extension K K K of F F F in A A A is shown to be nontrivial, even though it is not known whether Δ ( A ) \Delta(A) Δ ( A ) vanishes or not. A central simple algebra of degree 8 and exponent 2 with nontrivial Δ \Delta Δ -invariant over the rational function field F ( t ) F(t) F ( t ) in one indeterminate is built from A A A and the quadratic extension K K K , see Theorem 5.1.
A second construction is more explicit: it centers on the quaternion algebra [ a 1 a 2 , a 1 + a 2 ) F [a_1a_2,a_1+a_2)_F [ a 1 a 2 , a 1 + a 2 ) F over the rational function field F = k ( a 1 , a 2 ) F=k(a_1,a_2) F = k ( a 1 , a 2 ) in two indeterminates over an arbitrary field k k k of characteristic 2, and on the quadratic extension K = F ( a 1 a 2 ) K=F(\sqrt{a_1a_2}) K = F ( a 1 a 2 ) and the biquadratic extension L = K ( ℘ − 1 ( a 2 ) ) L=K(\wp^{-1}(a_2)) L = K ( ℘ − 1 ( a 2 )) . With these ingredients, and a sequence of rational function field extensions F ⊂ F ( t 1 ) ⊂ F ( t 1 , t 2 ) ⊂ F ( t 1 , t 2 , t 3 ) F \subset F(t_1) \subset F(t_1,t_2) \subset F(t_1,t_2,t_3) F ⊂ F ( t 1 ) ⊂ F ( t 1 , t 2 ) ⊂ F ( t 1 , t 2 , t 3 ) , the following are successively constructed:
a quaternion L ( t 1 ) L(t_1) L ( t 1 ) -algebra Q Q Q such that δ L ( t 1 ) / F ( t 1 ) ( Q ) ≠ 0 \delta_{L(t_1)/F(t_1)}(Q) \ne0 δ L ( t 1 ) / F ( t 1 ) ( Q ) = 0 (Theorem 5.2 );
a biquaternion K ( t 1 , t 2 ) K(t_1,t_2) K ( t 1 , t 2 ) -algebra B B B such that δ K ( t 1 , t 2 ) / F ( t 1 , t 2 ) ( B ) ≠ 0 \delta_{K(t_1,t_2)/F(t_1,t_2)}(B) \ne0 δ K ( t 1 , t 2 ) / F ( t 1 , t 2 ) ( B ) = 0 (Theorem 5.3);
a central simple F ( t 1 , t 2 , t 3 ) F(t_1,t_2,t_3) F ( t 1 , t 2 , t 3 ) -algebra A A A of degree 8 and exponent 2 such that Δ ( A ) ≠ 0 \Delta(A) \ne0 Δ ( A ) = 0 (Theorem 5.6 ).
In addition, a biquaternion K ( t 1 ) K(t_1) K ( t 1 ) -algebra B ′ B' B ′ containing Q Q Q such that δ K ( t 1 ) / F ( t 1 ) ( B ′ ) = 0 \delta_{K(t_1)/F(t_1)}(B')=0 δ K ( t 1 ) / F ( t 1 ) ( B ′ ) = 0 is given in Remark 5.4, and a central simple F ( t 1 , t 2 ) F(t_1,t_2) F ( t 1 , t 2 ) -algebra A ′ A' A ′ of degree 8 and exponent 2 containing B B B such that Δ ( A ′ ) = 0 \Delta(A')=0 Δ ( A ′ ) = 0 is given in Remark 5.7.
Since Δ ( A ) ≠ 0 \Delta(A) \ne0 Δ ( A ) = 0 , the algebra A A A in Theorem 5.6 does not decompose into a tensor product of quaternion algebras, but it is Brauer-equivalent to a tensor product of four quaternion algebras, which is particularly easy to describe explicitly.
Besides Rowen’s early examples in [23 ] , few methods to construct indecomposable division algebras of degree 8 and exponent 2 in characteristic 2 are known. A plausibly exhaustive list is given by the contributions of Karpenko [17 ] , McKinnie [21 ] and Aravire–Jacob–O’Ryan [5 ] . McKinnie’s examples have the noteworthy property of not containing any inseparable quadratic extension of the center. In [5 ] , Theorem 4.1(iii) asserts that any central division algebra of exponent 2 with a maximal subfield of the form F ( ℘ − 1 ( b 1 ) , ℘ − 1 ( b 2 ) , x ) F(\wp^{-1}(b_1),\wp^{-1}(b_2),\sqrt{x}) F ( ℘ − 1 ( b 1 ) , ℘ − 1 ( b 2 ) , x ) over a field F F F of characteristic 2 decomposes as a tensor product of three quaternion algebras. The algebra from Theorem 5.6 is a counterexample to this assertion. In addition, as explained in the appendix, [5 Theorem 4.1(ii)] fails for purely inseparable quartic field extensions of exponent 2: Remark A.3 provides an example of an indecomposable algebra of degree 8 and exponent 2 which contains a maximal subfield F ( ℘ − 1 ( a ) , x 4 ) F(\wp^{-1}(a),\sqrt[4]{x}) F ( ℘ − 1 ( a ) , 4 x ) .
Notation. Throughout the paper, F F F is a field of characteristic 2. For quadratic and bilinear forms over F F F , we use the same notation as [13 ] : the Witt group of quadratic forms and the Witt ring of bilinear forms over F F F are denoted by I q ( F ) I_q(F) I q ( F ) and W ( F ) W(F) W ( F ) respectively, and for every integer n ≥ 1 n \ge1 n ≥ 1 the n n n -th power of the fundamental ideal I ( F ) I(F) I ( F ) of W ( F ) W(F) W ( F ) is denoted by I n ( F ) I^n(F) I n ( F ) . We write I q n + 1 ( F ) I_q^{n+1}(F) I q n + 1 ( F ) for I n ( F ) ⋅ I q ( F ) I^n(F) \cdot I_q(F) I n ( F ) ⋅ I q ( F ) and let
I ‾ q n ( F ) = I q n ( F ) / I q n + 1 ( F ) . \overline{I}_q^n(F) = I_q^n(F)/I_q^{n+1}(F). I q n ( F ) = I q n ( F ) / I q n + 1 ( F ) .
For simplicity, we do not distinguish notationally between a form from its Witt-equivalence class. For φ ∈ I q n ( F ) \varphi\in I_q^n(F) φ ∈ I q n ( F ) , we write φ ‾ \overline{\varphi} φ for the image of φ \varphi φ in I ‾ q n ( F ) \overline{I}_q^n(F) I q n ( F ) .
Recall from [13 Th. 14.3] that the Clifford invariant defines a map e 2 e_2 e 2 from I q 2 ( F ) I_q^2(F) I q 2 ( F ) to Br 2 ( F ) \operatorname{Br}_2(F) Br 2 ( F ) , the 2-torsion subgroup of the Brauer group. A theorem of Sah [24 ] (see also [13 Th. 16.3] ) asserts that e 2 e_2 e 2 induces an isomorphism
e ‾ 2 : I ‾ q 2 ( F ) → ∼ Br 2 ( F ) . \overline{e}_2 : \overline{I}_q^2(F) \xrightarrow{\sim} \operatorname{Br}_2(F). e 2 : I q 2 ( F ) ∼ Br 2 ( F ) .
For any central simple F F F -algebra A A A such that 2 [ A ] = 0 2[A] = 0 2 [ A ] = 0 in Br ( F ) \operatorname{Br}(F) Br ( F ) , we write q A q_A q A for a quadratic form in I q 2 ( F ) I_q^2(F) I q 2 ( F ) such that e 2 ( q A ) = [ A ] e_2(q_A) = [A] e 2 ( q A ) = [ A ] . Its image q ‾ A ∈ I ‾ q 2 ( F ) \overline{q}_A \in\overline{I}_q^2(F) q A ∈ I q 2 ( F ) is thus uniquely determined by A A A . In particular, when A A A is a biquaternion algebra, the form q A q_A q A can be chosen to be 6-dimensional. The 6-dimensional quadratic forms q A q_A q A are known as Albert forms of the biquaternion algebra A A A , see [19 §16.A] .
For any central simple F F F -algebra A A A and any field extension L L L of F F F , we write A L A_L A L for the L L L -algebra A ⊗ F L A \otimes_F L A ⊗ F L obtained by scalar extension from F F F to L L L . We also write [ A ] L [A]_L [ A ] L for [ A L ] ∈ Br ( L ) [A_L] \in\operatorname{Br}(L) [ A L ] ∈ Br ( L ) . Similarly, for any quadratic space ( V , φ ) (V,\varphi) ( V , φ ) over F F F , we write ( V L , φ L ) (V_L,\varphi_L) ( V L , φ L ) for the quadratic space over L L L obtained by extending scalars.
For central simple algebras with involution, we use the notation of [19 ] . In particular, for every involution σ \sigma σ on a central simple F F F -algebra A A A , we let
Sym ( σ ) = { x ∈ A ∣ σ ( x ) = x } and Symd ( σ ) = { x + σ ( x ) ∣ x ∈ A } . \operatorname{Sym}(\sigma) = \{x \in A \mid\sigma(x) = x\}
\qquad\text{and}\qquad
\operatorname{Symd}(\sigma) = \{x + \sigma(x) \mid x \in A\}. Sym ( σ ) = { x ∈ A ∣ σ ( x ) = x } and Symd ( σ ) = { x + σ ( x ) ∣ x ∈ A } .
Recall that an F F F -linear involution σ \sigma σ is symplectic if and only if 1 ∈ Symd ( σ ) 1 \in\operatorname{Symd}(\sigma) 1 ∈ Symd ( σ ) , see [19 (2.6)] . The reduced characteristic polynomial of every x ∈ Symd ( σ ) x \in\operatorname{Symd}(\sigma) x ∈ Symd ( σ ) is then the square of the Pfaffian characteristic polynomial Prp σ , x \operatorname{Prp}_{\sigma,x} Prp σ , x , see [19 (2.10)] . If deg ( A ) = 2 m \deg(A) = 2m deg ( A ) = 2 m , then the forms Trp σ \operatorname{Trp}_{\sigma} Trp σ , Srp σ \operatorname{Srp}_{\sigma} Srp σ and Nrp σ \operatorname{Nrp}_{\sigma} Nrp σ of degree 1, 2, and m m m respectively are defined on Symd ( σ ) \operatorname{Symd}(\sigma) Symd ( σ ) as coefficients of the Pfaffian characteristic polynomial by the equation
Prp σ , x ( X ) = X m + Trp σ ( x ) X m − 1 + Srp σ ( x ) X m − 2 + ⋯ + Nrp σ ( x ) for x ∈ Symd ( σ ) . \operatorname{Prp}_{\sigma,x}(X) = X^m + \operatorname{Trp}_{\sigma}(x)X^{m-1} + \operatorname{Srp}_{\sigma}(x)X^{m-2} + \cdots+ \operatorname{Nrp}_{\sigma}(x)
\qquad\text{for } x \in\operatorname{Symd}(\sigma). Prp σ , x ( X ) = X m + Trp σ ( x ) X m − 1 + Srp σ ( x ) X m − 2 + ⋯ + Nrp σ ( x ) for x ∈ Symd ( σ ) .
Thus, for all x ∈ Symd ( σ ) x \in\operatorname{Symd}(\sigma) x ∈ Symd ( σ ) , Nrd A ( x ) \operatorname{Nrd}_A(x) Nrd A ( x ) is a square and Nrp σ ( x ) \operatorname{Nrp}_{\sigma}(x) Nrp σ ( x ) is its unique square root; it does not depend on the choice of the symplectic involution σ \sigma σ such that x ∈ Symd ( σ ) x \in\operatorname{Symd}(\sigma) x ∈ Symd ( σ ) .
The invariant Δ \Delta Δ
Throughout this section, A A A is a central simple F F F -algebra of degree 8 whose Brauer class [ A ] [A] [ A ] is 2-torsion. Recall from [11 ] that a 3-fold quadratic Pfister form π ( σ ) \pi(\sigma) π ( σ ) called the discriminant Pfister form is attached to each symplectic involution σ \sigma σ on A A A . We briefly recall the construction in characteristic 2, based on the existence of étale biquadratic F F F -algebras L ⊂ Symd ( σ ) L \subset\operatorname{Symd}(\sigma) L ⊂ Symd ( σ ) (i.e., F F F -algebras that carry a Galois structure with elementary abelian Galois group of order 4): let K ⊂ L K \subset L K ⊂ L be the quadratic étale subalgebra fixed under a nontrivial element α \alpha α in the Galois group and
W = { w ∈ Symd ( σ ) ∣ w ℓ = α ( ℓ ) w for all ℓ ∈ L } . W = \{w \in\operatorname{Symd}(\sigma) \mid w\ell= \alpha(\ell)w \text{ for all } \ell\in L\}. W = { w ∈ Symd ( σ ) ∣ w ℓ = α ( ℓ ) w for all ℓ ∈ L } .
It is shown in [11 Lemma 7.1] that dim F W = 8 \dim_F W=8 dim F W = 8 and w 2 ∈ K w^2\in K w 2 ∈ K for all w ∈ W w\in W w ∈ W . Moreover, since the trace map Tr K / F \operatorname{Tr}_{K/F} Tr K / F vanishes on F F F , [11 Th. 7.3] shows that the map w ↦ Tr K / F ( w 2 ) w\mapsto\operatorname{Tr}_{K/F}(w^2) w ↦ Tr K / F ( w 2 ) for w ∈ W w\in W w ∈ W is a scalar multiple of a 3-fold Pfister form. This form turns out to be independent up to isometry of the choices of L L L and α \alpha α , and to be hyperbolic if and only if A A A decomposes into a tensor product of three quaternion algebras stable under σ \sigma σ ; it is the discriminant Pfister form of σ \sigma σ according to the definition in [11 p. 269] or [27 Prop. 4, Cor. 6] . Since the characteristic of F F F is 2, it is shown in [27 ] that there is a 5-fold quadratic Pfister form π ^ ( σ ) \widehat{\pi}(\sigma) π ( σ ) such that the Witt-equivalence class of Srp σ \operatorname{Srp}_{\sigma} Srp σ decomposes as
Srp σ = [ 1 , 1 ] + π ( σ ) + π ^ ( σ ) in I q ( F ) . \operatorname{Srp}_{\sigma}=[1,1]+\pi(\sigma)+\widehat{\pi}(\sigma)\quad\text{in }I_q(F). Srp σ = [ 1 , 1 ] + π ( σ ) + π ( σ ) in I q ( F ) .
The discriminant Pfister form π ( σ ) \pi(\sigma) π ( σ ) is uniquely determined by this decomposition.
Our goal in this section is to show that the discriminant Pfister forms of the symplectic involutions on A A A all have the same image in the quotient of I q 3 ( F ) I_q^3(F) I q 3 ( F ) by the subgroup ⟨ ⟨ F × ⟩ ⟩ q A \langle\langle F^\times\rangle\rangle q_A ⟨⟨ F × ⟩⟩ q A , where q A ∈ I q 2 ( F ) q_A\in I_q^2(F) q A ∈ I q 2 ( F ) is any quadratic form such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] . This image is therefore an invariant of A A A , which we define to be the Δ \Delta Δ -invariant, see Definition 1.3.
As in characteristic different from 2 (see [14 §2] ), the proof proceeds by index reduction to the case where A A A is not a division algebra, i.e. A A A is Brauer-equivalent to a biquaternion algebra. In this case, we have the following:
Lemma 1.1. Suppose A A A is not a division algebra and let q A ∈ I q 2 ( F ) q_A\in I_q^2(F) q A ∈ I q 2 ( F ) be an Albert form such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] . Let σ 0 \sigma_0 σ 0 be a hyperbolic symplectic involution on A A A and let σ = Int ( u ) ∘ σ 0 \sigma=\operatorname{Int}(u)\circ\sigma_0 σ = Int ( u ) ∘ σ 0 for some unit u ∈ Symd ( σ 0 ) u\in\operatorname{Symd}(\sigma_0) u ∈ Symd ( σ 0 ) . Then π ( σ ) \pi(\sigma) π ( σ ) is Witt-equivalent to a scalar multiple of ⟨ ⟨ Nrp σ 0 ( u ) ⟩ ⟩ q A \langle\langle\operatorname{Nrp}_{\sigma_0}(u)\rangle\rangle q_A ⟨⟨ Nrp σ 0 ( u )⟩⟩ q A .
Proof. The proof has two steps. First, we prove that the quadratic form ⟨ ⟨ Nrp σ 0 ( u ) ⟩ ⟩ q A \langle\langle\operatorname{Nrp}_{\sigma_0}(u)\rangle\rangle q_A ⟨⟨ Nrp σ 0 ( u )⟩⟩ q A does not depend on the choices of σ 0 \sigma_0 σ 0 and u u u as in the statement. We then produce a particular choice of σ 0 \sigma_0 σ 0 and u u u for which π ( σ ) \pi(\sigma) π ( σ ) is Witt-equivalent to a scalar multiple of ⟨ ⟨ Nrp σ 0 ( u ) ⟩ ⟩ q A \langle\langle\operatorname{Nrp}_{\sigma_0}(u)\rangle\rangle q_A ⟨⟨ Nrp σ 0 ( u )⟩⟩ q A .
Assume σ = Int ( u ) ∘ σ 0 = Int ( u ′ ) ∘ σ 0 ′ \sigma=\operatorname{Int}(u)\circ\sigma_0=\operatorname{Int}(u')\circ\sigma'_0 σ = Int ( u ) ∘ σ 0 = Int ( u ′ ) ∘ σ 0 ′ for some hyperbolic involutions σ 0 \sigma_0 σ 0 and σ 0 ′ \sigma'_0 σ 0 ′ and some units u ∈ Symd ( σ 0 ) u\in\operatorname{Symd}(\sigma_0) u ∈ Symd ( σ 0 ) and u ′ ∈ Symd ( σ 0 ′ ) u'\in\operatorname{Symd}(\sigma'_0) u ′ ∈ Symd ( σ 0 ′ ) . Since all the hyperbolic symplectic involutions on A A A are conjugate, there exists a ∈ A × a\in A^\times a ∈ A × such that
σ 0 ′ = Int ( a ) ∘ σ 0 ∘ Int ( a ) − 1 = Int ( a σ 0 ( a ) ) ∘ σ 0 . \sigma'_0=\operatorname{Int}(a)\circ\sigma_0\circ\operatorname{Int}(a)^{-1}=\operatorname{Int}(a\sigma_0(a))\circ\sigma_0. σ 0 ′ = Int ( a ) ∘ σ 0 ∘ Int ( a ) − 1 = Int ( a σ 0 ( a )) ∘ σ 0 .
It follows that σ = Int ( u ) ∘ σ 0 = Int ( u ′ a σ 0 ( a ) ) ∘ σ 0 \sigma=\operatorname{Int}(u)\circ\sigma_0=\operatorname{Int}(u'a\sigma_0(a))\circ\sigma_0 σ = Int ( u ) ∘ σ 0 = Int ( u ′ a σ 0 ( a )) ∘ σ 0 , so
u = λ u ′ a σ 0 ( a ) for some λ ∈ F × . u=\lambda u'a\sigma_0(a)\qquad\text{for some }\lambda\in F^\times. u = λ u ′ a σ 0 ( a ) for some λ ∈ F × .
Computing the reduced norms, we get Nrd A ( u ) = λ 8 Nrd A ( u ′ ) Nrd A ( a ) 2 \operatorname{Nrd}_A(u)=\lambda^8\operatorname{Nrd}_A(u')\operatorname{Nrd}_A(a)^2 Nrd A ( u ) = λ 8 Nrd A ( u ′ ) Nrd A ( a ) 2 , hence
Nrp σ 0 ( u ) = λ 4 Nrd A ( a ) Nrp σ 0 ′ ( u ′ ) \operatorname{Nrp}_{\sigma_0}(u)=\lambda^4\operatorname{Nrd}_A(a)\operatorname{Nrp}_{\sigma'_0}(u') Nrp σ 0 ( u ) = λ 4 Nrd A ( a ) Nrp σ 0 ′ ( u ′ )
because Nrp σ 0 ( u ) \operatorname{Nrp}_{\sigma_0}(u) Nrp σ 0 ( u ) (resp. Nrp σ 0 ′ ( u ′ ) \operatorname{Nrp}_{\sigma'_0}(u') Nrp σ 0 ′ ( u ′ ) ) is the unique square root of Nrd A ( u ) \operatorname{Nrd}_A(u) Nrd A ( u ) (resp. Nrd A ( u ′ ) \operatorname{Nrd}_A(u') Nrd A ( u ′ ) ).
Now, Nrd A ( a ) \operatorname{Nrd}_A(a) Nrd A ( a ) is a reduced norm from any biquaternion algebra Brauer-equivalent to A A A , hence it is a similarity factor of q A q_A q A by [19 (16.6)] . Therefore, ⟨ λ 4 Nrd A ( a ) ⟩ q A ≃ q A \langle\lambda^4\operatorname{Nrd}_A(a)\rangle q_A\simeq q_A ⟨ λ 4 Nrd A ( a )⟩ q A ≃ q A , and we get ⟨ Nrp σ 0 ( u ) ⟩ q A ≃ ⟨ Nrp σ 0 ′ ( u ′ ) ⟩ q A \langle\operatorname{Nrp}_{\sigma_0}(u)\rangle q_A\simeq\langle\operatorname{Nrp}_{\sigma'_0}(u')\rangle q_A ⟨ Nrp σ 0 ( u )⟩ q A ≃ ⟨ Nrp σ 0 ′ ( u ′ )⟩ q A as expected.
To produce a suitable choice of σ 0 \sigma_0 σ 0 and u u u , choose a representation A = End D ( V ) A=\operatorname{End}_D(V) A = End D ( V ) for some division algebra D D D of degree 1, 2 or 4 and some right D D D -vector space V V V of dimension 8, 4 or 2 respectively. Then σ \sigma σ is adjoint to a form h h h on V V V that is alternating if D = F D=F D = F and hermitian for a symplectic involution on D D D if deg ( D ) = 2 \deg(D)=2 deg ( D ) = 2 or 4. In each case, we may consider an orthogonal decomposition
( V , h ) = ( V 1 , h 1 ) ⊥ ( V 2 , h 2 ) with dim V 1 = dim V 2 , (V,h)=(V_1,h_1)\perp(V_2,h_2)\qquad\text{with }\dim V_1=\dim V_2, ( V , h ) = ( V 1 , h 1 ) ⊥ ( V 2 , h 2 ) with dim V 1 = dim V 2 ,
for some nonsingular forms h 1 h_1 h 1 and h 2 h_2 h 2 . Using a D D D -linear isomorphism V 1 ≃ V 2 V_1 \simeq V_2 V 1 ≃ V 2 , we identify End D ( V 1 ) = End D ( V 2 ) \operatorname{End}_D(V_1)=\operatorname{End}_D(V_2) End D ( V 1 ) = End D ( V 2 ) , denoted by B B B , and A = M 2 ( B ) A=M_2(B) A = M 2 ( B ) . Let v ∈ B × v\in B^\times v ∈ B × be such that h 2 ( x , y ) = h 1 ( v ( x ) , y ) h_2(x,y)=h_1(v(x),y) h 2 ( x , y ) = h 1 ( v ( x ) , y ) for all x , y ∈ V 2 x,y\in V_2 x , y ∈ V 2 . Then the involutions σ 1 , σ 2 \sigma_1,\sigma_2 σ 1 , σ 2 on B B B adjoint to h 1 h_1 h 1 and h 2 h_2 h 2 are related by
σ 2 = Int ( v − 1 ) ∘ σ 1 . \sigma_2=\operatorname{Int}(v^{-1})\circ\sigma_1. σ 2 = Int ( v − 1 ) ∘ σ 1 .
These involutions are both symplectic since h 1 h_1 h 1 and h 2 h_2 h 2 have the same type as h h h , and the involution σ \sigma σ on M 2 ( B ) M_2(B) M 2 ( B ) is given by
σ ( ( a 11 a 12 a 21 a 22 ) ) = ( σ 1 ( a 11 ) σ 1 ( a 21 ) v v − 1 σ 1 ( a 12 ) v − 1 σ 1 ( a 22 ) v ) for a 11 , … , a 22 ∈ B . \sigma\left(\begin{pmatrix}a_{11}&a_{12}\\a_{21}&a_{22}\end{pmatrix}\right)=\begin{pmatrix}\sigma_1(a_{11})&\sigma_1(a_{21})v\\v^{-1}\sigma_1(a_{12})&v^{-1}\sigma_1(a_{22})v\end{pmatrix}
\quad\text{for }a_{11},\ldots,a_{22}\in B. σ ( ( a 11 a 21 a 12 a 22 ) ) = ( σ 1 ( a 11 ) v − 1 σ 1 ( a 12 ) σ 1 ( a 21 ) v v − 1 σ 1 ( a 22 ) v ) for a 11 , … , a 22 ∈ B .
Therefore, σ = Int ( u ) ∘ σ 0 \sigma=\operatorname{Int}(u)\circ\sigma_0 σ = Int ( u ) ∘ σ 0 , where u = ( 1 0 0 v − 1 ) u=\begin{pmatrix}1&0\\0&v^{-1}\end{pmatrix} u = ( 1 0 0 v − 1 ) and σ 0 \sigma_0 σ 0 is the involution given by
σ 0 ( ( a 11 a 12 a 21 a 22 ) ) = ( σ 1 ( a 11 ) σ 1 ( a 21 ) σ 1 ( a 12 ) σ 1 ( a 22 ) ) for a 11 , … , a 22 ∈ B . \sigma_0\left(\begin{pmatrix}a_{11}&a_{12}\\a_{21}&a_{22}\end{pmatrix}\right)=\begin{pmatrix}\sigma_1(a_{11})&\sigma_1(a_{21})\\\sigma_1(a_{12})&\sigma_1(a_{22})\end{pmatrix}
\quad\text{for }a_{11},\ldots,a_{22}\in B. σ 0 ( ( a 11 a 21 a 12 a 22 ) ) = ( σ 1 ( a 11 ) σ 1 ( a 12 ) σ 1 ( a 21 ) σ 1 ( a 22 ) ) for a 11 , … , a 22 ∈ B .
Since σ 1 \sigma_1 σ 1 is symplectic, there exists a ∈ B a\in B a ∈ B such that σ 1 ( a ) + a = 1 \sigma_1(a)+a=1 σ 1 ( a ) + a = 1 . Then the element e = ( a 1 + a a 1 + a ) e=\begin{pmatrix}a&1+a\\a&1+a\end{pmatrix} e = ( a a 1 + a 1 + a ) satisfies σ 0 ( e ) + e = 1 \sigma_0(e)+e=1 σ 0 ( e ) + e = 1 and e 2 = e e^2=e e 2 = e , which shows that σ 0 \sigma_0 σ 0 is a hyperbolic symplectic involution. Since σ 0 \sigma_0 σ 0 and σ \sigma σ are both symplectic involutions, it follows that u ∈ Symd ( σ 0 ) u\in\operatorname{Symd}(\sigma_0) u ∈ Symd ( σ 0 ) by [19 (2.7)] . To complete the proof, we compute π ( σ ) \pi(\sigma) π ( σ ) .
Observe that Symd ( σ 1 ) × Symd ( σ 2 ) \operatorname{Symd}(\sigma_1)\times\operatorname{Symd}(\sigma_2) Symd ( σ 1 ) × Symd ( σ 2 ) embeds diagonally in Symd ( σ ) \operatorname{Symd}(\sigma) Symd ( σ ) . Let K = F × F K=F\times F K = F × F diagonally embedded in Symd ( σ ) \operatorname{Symd}(\sigma) Symd ( σ ) . By [9 Th. 6.10] , there exists a biquadratic étale F F F -algebra L ⊂ Symd ( σ ) L\subset\operatorname{Symd}(\sigma) L ⊂ Symd ( σ ) containing K K K as the subalgebra fixed under a nontrivial element α \alpha α of its Galois group. This algebra has the form L = L 0 × L 0 L=L_0\times L_0 L = L 0 × L 0 for some quadratic étale F F F -algebra L 0 L_0 L 0 that embeds in Symd ( σ 1 ) \operatorname{Symd}(\sigma_1) Symd ( σ 1 ) and Symd ( σ 2 ) \operatorname{Symd}(\sigma_2) Symd ( σ 2 ) , and α = α 0 × α 0 \alpha=\alpha_0\times\alpha_0 α = α 0 × α 0 for α 0 \alpha_0 α 0 the nontrivial F F F -automorphism of L 0 L_0 L 0 . Letting
W = { w ∈ Symd ( σ ) ∣ w ℓ = α ( ℓ ) w for all ℓ ∈ L } W=\{w\in\operatorname{Symd}(\sigma)\mid w\ell=\alpha(\ell)w\text{ for all }\ell\in L\} W = { w ∈ Symd ( σ ) ∣ w ℓ = α ( ℓ ) w for all ℓ ∈ L }
and
W i = { w ∈ Symd ( σ i ) ∣ w ℓ = α ( ℓ ) w for all ℓ ∈ L 0 } for i = 1 , 2 , W_i=\{w\in\operatorname{Symd}(\sigma_i)\mid w\ell=\alpha(\ell)w\text{ for all }\ell\in L_0\}\quad\text{for }i=1,2, W i = { w ∈ Symd ( σ i ) ∣ w ℓ = α ( ℓ ) w for all ℓ ∈ L 0 } for i = 1 , 2 ,
we have for the squaring maps sq W : W → K \operatorname{sq}_W:W\to K sq W : W → K and sq W i : W i → F \operatorname{sq}_{W_i}:W_i\to F sq W i : W i → F
( W , sq W ) = ( W 1 , sq W 1 ) × ( W 2 , sq W 2 ) , (W,\operatorname{sq}_W)=(W_1,\operatorname{sq}_{W_1})\times(W_2,\operatorname{sq}_{W_2}), ( W , sq W ) = ( W 1 , sq W 1 ) × ( W 2 , sq W 2 ) ,
hence by definition π ( σ ) \pi(\sigma) π ( σ ) is Witt-equivalent to a scalar multiple of sq W 1 ⊥ sq W 2 \operatorname{sq}_{W_1}\perp\operatorname{sq}_{W_2} sq W 1 ⊥ sq W 2 . On the other hand, [9 Prop. 5.6 and 6.5] yields
( Symd ( σ i ) , Nrp σ i ) ≃ ( L 0 , N L 0 / F ) ⊥ ( W i , sq W i ) for i = 1 , 2 , (\operatorname{Symd}(\sigma_i),\operatorname{Nrp}_{\sigma_i})\simeq(L_0,N_{L_0/F})\perp(W_i,\operatorname{sq}_{W_i})\quad\text{for }i=1,2, ( Symd ( σ i ) , Nrp σ i ) ≃ ( L 0 , N L 0 / F ) ⊥ ( W i , sq W i ) for i = 1 , 2 ,
where N L 0 / F N_{L_0/F} N L 0 / F is the norm quadratic form. Therefore, sq W 1 ⊥ sq W 2 \operatorname{sq}_{W_1}\perp\operatorname{sq}_{W_2} sq W 1 ⊥ sq W 2 is Witt-equivalent to Nrp σ 1 ⊥ Nrp σ 2 \operatorname{Nrp}_{\sigma_1}\perp\operatorname{Nrp}_{\sigma_2} Nrp σ 1 ⊥ Nrp σ 2 . Since σ 1 = Int ( v ) ∘ σ 2 \sigma_1=\operatorname{Int}(v)\circ\sigma_2 σ 1 = Int ( v ) ∘ σ 2 , it follows from [19 (16.18)] that Nrp σ 1 ≃ ⟨ Nrp σ 2 ( v ) ⟩ Nrp σ 2 \operatorname{Nrp}_{\sigma_1}\simeq\langle\operatorname{Nrp}_{\sigma_2}(v)\rangle\operatorname{Nrp}_{\sigma_2} Nrp σ 1 ≃ ⟨ Nrp σ 2 ( v )⟩ Nrp σ 2 , hence π ( σ ) \pi(\sigma) π ( σ ) is Witt-equivalent to a scalar multiple of ⟨ ⟨ Nrp σ 2 ( v ) ⟩ ⟩ Nrp σ 2 \langle\langle\operatorname{Nrp}_{\sigma_2}(v)\rangle\rangle\operatorname{Nrp}_{\sigma_2} ⟨⟨ Nrp σ 2 ( v )⟩⟩ Nrp σ 2 . This completes the proof because Nrp σ ( u ) = Nrp σ 2 ( v ) \operatorname{Nrp}_\sigma(u)=\operatorname{Nrp}_{\sigma_2}(v) Nrp σ ( u ) = Nrp σ 2 ( v ) and Nrp σ 2 \operatorname{Nrp}_{\sigma_2} Nrp σ 2 is an Albert form of B B B by [19 (16.8)] , hence q A q_A q A is similar to Nrp σ 2 \operatorname{Nrp}_{\sigma_2} Nrp σ 2 by [19 (16.3)] . □
Without restriction on the index of A A A , let q A ∈ I q 2 ( F ) q_A\in I_q^2(F) q A ∈ I q 2 ( F ) be such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] . Note that q A q_A q A is uniquely determined modulo I q 3 ( F ) I_q^3(F) I q 3 ( F ) , hence q A ‾ ∈ I ‾ q 2 ( F ) \overline{q_A}\in\overline{I}_q^2(F) q A ∈ I q 2 ( F ) depends only on A A A .
Theorem 1.2. Let σ , σ ′ \sigma,\sigma' σ , σ ′ be symplectic involutions on A A A and let w ∈ Symd ( σ ) w\in\operatorname{Symd}(\sigma) w ∈ Symd ( σ ) be such that σ ′ = Int ( w ) ∘ σ \sigma'=\operatorname{Int}(w)\circ\sigma σ ′ = Int ( w ) ∘ σ . Then
π ( σ ′ ) ‾ = ⟨ ⟨ Nrp σ ( w ) ⟩ ⟩ q A + π ( σ ) ‾ in I ‾ q 3 ( F ) . \overline{\pi(\sigma')}=\langle\langle\operatorname{Nrp}_\sigma(w)\rangle\rangle q_A+\overline{\pi(\sigma)}
\quad\text{in }\overline{I}_q^3(F). π ( σ ′ ) = ⟨⟨ Nrp σ ( w )⟩⟩ q A + π ( σ ) in I q 3 ( F ) .
Proof. If A A A is not a division algebra, we may choose for q A q_A q A an Albert form and pick a hyperbolic symplectic involution σ 0 \sigma_0 σ 0 on A A A . If u ∈ Symd ( σ 0 ) u \in\operatorname{Symd}(\sigma_0) u ∈ Symd ( σ 0 ) satisfies σ = Int ( u ) ∘ σ 0 \sigma= \operatorname{Int}(u) \circ\sigma_0 σ = Int ( u ) ∘ σ 0 , then σ ′ = Int ( w u ) ∘ σ 0 \sigma' = \operatorname{Int}(wu) \circ\sigma_0 σ ′ = Int ( w u ) ∘ σ 0 and Lemma 1.1 shows that π ( σ ) \pi(\sigma) π ( σ ) and π ( σ ′ ) \pi(\sigma') π ( σ ′ ) are Witt-equivalent to scalar multiples of ⟨ ⟨ Nrp σ 0 ( u ) ⟩ ⟩ q A \langle\!\langle\operatorname{Nrp}_{\sigma_0}(u)\rangle\!\rangle_{q_A} ⟨ ⟨ Nrp σ 0 ( u )⟩ ⟩ q A and ⟨ ⟨ Nrp σ 0 ( w u ) ⟩ ⟩ q A \langle\!\langle\operatorname{Nrp}_{\sigma_0}(wu)\rangle\!\rangle_{q_A} ⟨ ⟨ Nrp σ 0 ( w u )⟩ ⟩ q A respectively, hence
π ( σ ) ‾ = ⟨ ⟨ Nrp σ 0 ( u ) ⟩ ⟩ q A and π ( σ ′ ) ‾ = ⟨ ⟨ Nrp σ 0 ( w u ) ⟩ ⟩ q A in I ‾ q 3 ( F ) . \overline{\pi(\sigma)}=\langle\!\langle\operatorname{Nrp}_{\sigma_0}(u)\rangle\!\rangle_{q_A}
\quad\text{and}\quad
\overline{\pi(\sigma')}=\langle\!\langle\operatorname{Nrp}_{\sigma_0}(wu)\rangle\!\rangle_{q_A}
\quad\text{in }\overline{I}^{3}_{q}(F). π ( σ ) = ⟨ ⟨ Nrp σ 0 ( u )⟩ ⟩ q A and π ( σ ′ ) = ⟨ ⟨ Nrp σ 0 ( w u )⟩ ⟩ q A in I q 3 ( F ) .
By taking the square root of each side in Nrd A ( w u ) = Nrd A ( w ) Nrd A ( u ) \operatorname{Nrd}_{A}(wu)=\operatorname{Nrd}_{A}(w)\operatorname{Nrd}_{A}(u) Nrd A ( w u ) = Nrd A ( w ) Nrd A ( u ) we obtain Nrp σ 0 ( w u ) = Nrp σ ( w ) Nrp σ 0 ( u ) \operatorname{Nrp}_{\sigma_0}(wu)=\operatorname{Nrp}_{\sigma}(w)\operatorname{Nrp}_{\sigma_0}(u) Nrp σ 0 ( w u ) = Nrp σ ( w ) Nrp σ 0 ( u ) , hence ⟨ ⟨ Nrp σ 0 ( w u ) ⟩ ⟩ ≡ ⟨ ⟨ Nrp σ ( w ) ⟩ ⟩ + ⟨ ⟨ Nrp σ 0 ( u ) ⟩ ⟩ ( m o d I 2 ( F ) ) \langle\!\langle\operatorname{Nrp}_{\sigma_0}(wu)\rangle\!\rangle\equiv\langle\!\langle\operatorname{Nrp}_{\sigma}(w)\rangle\!\rangle+\langle\!\langle\operatorname{Nrp}_{\sigma_0}(u)\rangle\!\rangle\pmod{I^2(F)} ⟨ ⟨ Nrp σ 0 ( w u )⟩ ⟩ ≡ ⟨ ⟨ Nrp σ ( w )⟩ ⟩ + ⟨ ⟨ Nrp σ 0 ( u )⟩ ⟩ ( mod I 2 ( F )) . Therefore,
π ( σ ′ ) ‾ = ⟨ ⟨ Nrp σ ( w ) ⟩ ⟩ q A + ⟨ ⟨ Nrp σ 0 ( u ) ⟩ ⟩ q A = ⟨ ⟨ Nrp σ ( w ) ⟩ ⟩ q A + π ( σ ) ‾ . \overline{\pi(\sigma')}=\langle\!\langle\operatorname{Nrp}_{\sigma}(w)\rangle\!\rangle_{q_A}+\langle\!\langle\operatorname{Nrp}_{\sigma_0}(u)\rangle\!\rangle_{q_A}
=\langle\!\langle\operatorname{Nrp}_{\sigma}(w)\rangle\!\rangle_{q_A}+\overline{\pi(\sigma)}. π ( σ ′ ) = ⟨ ⟨ Nrp σ ( w )⟩ ⟩ q A + ⟨ ⟨ Nrp σ 0 ( u )⟩ ⟩ q A = ⟨ ⟨ Nrp σ ( w )⟩ ⟩ q A + π ( σ ) .
Thus, the theorem is proved in the case where A A A is not a division algebra.
When A A A is a division algebra, we reduce the index by scalar extension to use the preceding case. Recall from [23 Th. 2] that A A A is Brauer-equivalent to a tensor product of four quaternion algebras; let
[ A ] = [ Q 1 ⊗ F Q 2 ⊗ F Q 3 ⊗ F Q 4 ] in Br ( F ) . [A]=[Q_1\otimes_F Q_2\otimes_F Q_3\otimes_F Q_4]\quad\text{in }\operatorname{Br}(F). [ A ] = [ Q 1 ⊗ F Q 2 ⊗ F Q 3 ⊗ F Q 4 ] in Br ( F ) .
Let F 1 F_1 F 1 be the function field of an Albert quadratic form of Q 3 ⊗ F Q 4 Q_3\otimes_F Q_4 Q 3 ⊗ F Q 4 . (If Q 3 ⊗ F Q 4 Q_3\otimes_F Q_4 Q 3 ⊗ F Q 4 is not a division algebra, then the Albert quadratic form is isotropic and F 1 F_1 F 1 is a rational function field over F F F .) After scalar extension from F F F to F 1 F_1 F 1 , Q 3 ⊗ F Q 4 Q_3\otimes_F Q_4 Q 3 ⊗ F Q 4 is Brauer-equivalent to a quaternion algebra H H H , hence
[ A F 1 ] = [ Q 1 ⊗ F Q 2 ⊗ F H ] in Br ( F 1 ) . [ A_{F_1} ]=[Q_1\otimes_F Q_2\otimes_F H]\quad\text{in }\operatorname{Br}(F_1). [ A F 1 ] = [ Q 1 ⊗ F Q 2 ⊗ F H ] in Br ( F 1 ) .
Extend scalars further from F 1 F_1 F 1 to the function field F 2 F_2 F 2 of an Albert quadric of Q 2 ⊗ F H Q_2\otimes_F H Q 2 ⊗ F H ; then A F 2 A_{F_2} A F 2 is not a division algebra, hence by the special case of the theorem proved above,
π ( σ ′ ) ‾ − ⟨ ⟨ Nrp σ ( w ) ⟩ ⟩ q A − π ( σ ) ‾ ∈ ker ( I ‾ q 3 ( F ) → I ‾ q 3 ( F 2 ) ) . \overline{\pi(\sigma')}-\langle\!\langle\operatorname{Nrp}_{\sigma}(w)\rangle\!\rangle_{q_A}-\overline{\pi(\sigma)}
\in\ker\bigl(\overline{I}^{3}_{q}(F)\to\overline{I}^{3}_{q}(F_2)\bigr). π ( σ ′ ) − ⟨ ⟨ Nrp σ ( w )⟩ ⟩ q A − π ( σ ) ∈ ker ( I q 3 ( F ) → I q 3 ( F 2 ) ) .
But by [15 Th. 5.6] scalar extension to the function field of an Albert quadratic form induces an injective map on I ‾ q 3 ( F ) \overline{I}^{3}_{q}(F) I q 3 ( F ) , hence ker ( I ‾ q 3 ( F ) → I ‾ q 3 ( F 2 ) ) = 0 \ker\bigl(\overline{I}^{3}_{q}(F)\to\overline{I}^{3}_{q}(F_2)\bigr)=0 ker ( I q 3 ( F ) → I q 3 ( F 2 ) ) = 0 and the proof is complete. □1
Theorem 1.2 allows us to set the following definition:
Definition 1.3
Definition 1.3. For a central simple F F F -algebra A A A of degree 8 with 2 [ A ] = 0 2[A]=0 2 [ A ] = 0 in Br ( F ) \operatorname{Br}(F) Br ( F ) and any quadratic form q A ∈ I q 2 ( F ) q_A\in I^2_q(F) q A ∈ I q 2 ( F ) such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] , the invariant Δ ( A ) \Delta(A) Δ ( A ) is the image in the quotient I ‾ q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q A \overline{I}^{3}_{q}(F)/\langle\!\langle F^\times\rangle\!\rangle_{q_A} I q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q A of the discriminant Pfister form of any symplectic involution σ \sigma σ on A A A ,
Δ ( A ) = π ( σ ) ‾ + ⟨ ⟨ F × ⟩ ⟩ q A ∈ I ‾ q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q A . \Delta(A)=\overline{\pi(\sigma)}+\langle\!\langle F^\times\rangle\!\rangle_{q_A}
\in\overline{I}^{3}_{q}(F)/\langle\!\langle F^\times\rangle\!\rangle_{q_A}. Δ ( A ) = π ( σ ) + ⟨ ⟨ F × ⟩ ⟩ q A ∈ I q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q A .
Vanishing of Δ ( A ) \Delta(A) Δ ( A ) is a necessary condition for the decomposability of A A A :
Proposition 1.4. If A = Q 1 ⊗ F Q 2 ⊗ F Q 3 A=Q_1\otimes_F Q_2\otimes_F Q_3 A = Q 1 ⊗ F Q 2 ⊗ F Q 3 for some quaternion F F F -algebras Q 1 Q_1 Q 1 , Q 2 Q_2 Q 2 , Q 3 Q_3 Q 3 , then Δ ( A ) = 0 \Delta(A)=0 Δ ( A ) = 0 .
Proof. The tensor product of the canonical involutions on Q 1 Q_1 Q 1 , Q 2 Q_2 Q 2 and Q 3 Q_3 Q 3 is a symplectic involution with hyperbolic discriminant Pfister form, by the main theorem in [11 ] or [27 Prop. 8] . □
Cohomological Δ \Delta Δ -invariant. Translating the construction above into cohomological terms is straightforward. Let Ω F 1 \Omega^1_F Ω F 1 be the F F F -vector space of absolute differentials on F F F and Ω F n = ⋀ n Ω F 1 \Omega^n_F=\bigwedge^n\Omega^1_F Ω F n = ⋀ n Ω F 1 for n ≥ 2 n\ge2 n ≥ 2 . Recall that the (Kato–Milne) cohomology group H n + 1 ( F ) H^{n+1}(F) H n + 1 ( F ) is defined as the cokernel of the map
℘ : Ω F n ⟶ Ω F n / d ( Ω F n − 1 ) \wp:\Omega^n_F\longrightarrow\Omega^n_F/d(\Omega^{n-1}_F) ℘ : Ω F n ⟶ Ω F n / d ( Ω F n − 1 )
such that
℘ ( x d y 1 y 1 ∧ ⋯ ∧ d y n y n ) = ( x 2 − x ) d y 1 y 1 ∧ ⋯ ∧ d y n y n \wp\left(x\frac{dy_1}{y_1}\wedge\cdots\wedge\frac{dy_n}{y_n}\right)=(x^2-x)\frac{dy_1}{y_1}\wedge\cdots\wedge\frac{dy_n}{y_n} ℘ ( x y 1 d y 1 ∧ ⋯ ∧ y n d y n ) = ( x 2 − x ) y 1 d y 1 ∧ ⋯ ∧ y n d y n
for x ∈ F x\in F x ∈ F and y 1 , … , y n ∈ F × y_1,\ldots,y_n\in F^\times y 1 , … , y n ∈ F × , see [18 ] . A different (but equivalent) definition is given in [13 §101] , where H n + 1 ( F ) H^{n+1}(F) H n + 1 ( F ) is denoted by H n + 1 , n ( F , Z / 2 Z ) H^{n+1,n}(F,\mathbb{Z}/2\mathbb{Z}) H n + 1 , n ( F , Z /2 Z ) . Kato proves in [18 ] that the map carrying x d y 1 y 1 ∧ ⋯ ∧ d y n y n ∈ Ω F n x\frac{dy_1}{y_1}\wedge\cdots\wedge\frac{dy_n}{y_n}\in\Omega^n_F x y 1 d y 1 ∧ ⋯ ∧ y n d y n ∈ Ω F n to ⟨ ⟨ y 1 , … , y n , x ] ] ∈ I ‾ q n + 1 ( F ) \langle\langle y_1,\ldots,y_n,x]]\in\overline{I}^{n+1}_q(F) ⟨⟨ y 1 , … , y n , x ]] ∈ I q n + 1 ( F ) for x ∈ F x\in F x ∈ F and y 1 , … , y n ∈ F × y_1,\ldots,y_n\in F^\times y 1 , … , y n ∈ F × induces a canonical isomorphism
κ n + 1 : H n + 1 ( F ) → ∼ I ‾ q n + 1 ( F ) . \kappa_{n+1}:H^{n+1}(F)\xrightarrow{\sim}\overline{I}^{n+1}_q(F). κ n + 1 : H n + 1 ( F ) ∼ I q n + 1 ( F ) .
In particular, the composition e ‾ 2 ∘ κ 2 : H 2 ( F ) → Br 2 ( F ) \overline{e}_2\circ\kappa_2:H^2(F)\to\operatorname{Br}_2(F) e 2 ∘ κ 2 : H 2 ( F ) → Br 2 ( F ) is an isomorphism, hence for every central simple F F F -algebra A A A such that 2 [ A ] = 0 2[A]=0 2 [ A ] = 0 we may find h A ∈ H 2 ( F ) h_A\in H^2(F) h A ∈ H 2 ( F ) such that e ‾ 2 ∘ κ 2 ( h A ) = [ A ] \overline{e}_2\circ\kappa_2(h_A)=[A] e 2 ∘ κ 2 ( h A ) = [ A ] . For x ∈ F × x\in F^\times x ∈ F × , let dlog ( x ) = d x x ∈ Ω F 1 \operatorname{dlog}(x)=\frac{dx}{x}\in\Omega^1_F dlog ( x ) = x d x ∈ Ω F 1 . The exterior product defines a map F × × H 2 ( F ) → H 3 ( F ) F^\times\times H^2(F)\to H^3(F) F × × H 2 ( F ) → H 3 ( F ) carrying ( x , h A ) (x,h_A) ( x , h A ) to dlog ( x ) ∧ h A \operatorname{dlog}(x)\wedge h_A dlog ( x ) ∧ h A . By definition, κ 3 ( dlog ( x ) ∧ h A ) = ⟨ ⟨ x ⟩ ⟩ κ 2 ( h A ) \kappa_3(\operatorname{dlog}(x)\wedge h_A)=\langle\langle x\rangle\rangle\kappa_2(h_A) κ 3 ( dlog ( x ) ∧ h A ) = ⟨⟨ x ⟩⟩ κ 2 ( h A ) , and κ 2 ( h A ) = q ‾ A \kappa_2(h_A)=\overline{q}_A κ 2 ( h A ) = q A for any quadratic form q A ∈ I q 2 ( F ) q_A\in I^2_q(F) q A ∈ I q 2 ( F ) such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] , hence κ 3 \kappa_3 κ 3 defines an isomorphism
H 3 ( F ) / dlog ( F × ) ∧ h A → ∼ I ‾ q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q ‾ A . (1) H^3(F)/\operatorname{dlog}(F^\times)\wedge h_A\xrightarrow{\sim}\overline{I}^3_q(F)/\langle\langle F^\times\rangle\rangle\overline{q}_A. \tag*{(1)} H 3 ( F ) / dlog ( F × ) ∧ h A ∼ I q 3 ( F ) / ⟨⟨ F × ⟩⟩ q A . (1)
If deg ( A ) = 8 \deg(A)=8 deg ( A ) = 8 and σ \sigma σ is a symplectic involution on A A A , let π h ( A , σ ) ∈ H 3 ( F ) \pi_h(A,\sigma)\in H^3(F) π h ( A , σ ) ∈ H 3 ( F ) be defined by the equation
κ 3 ( π h ( A , σ ) ) = π ( σ ) ‾ ∈ I ‾ q 3 ( F ) . \kappa_3\left(\pi_h(A,\sigma)\right)=\overline{\pi(\sigma)}\in\overline{I}^3_q(F). κ 3 ( π h ( A , σ ) ) = π ( σ ) ∈ I q 3 ( F ) .
Theorem 1.2 readily shows that if σ , σ ′ \sigma,\sigma' σ , σ ′ are symplectic involutions on A A A and w ∈ Symd ( σ ) w\in\operatorname{Symd}(\sigma) w ∈ Symd ( σ ) is such that σ ′ = Int ( w ) ∘ σ \sigma'=\operatorname{Int}(w)\circ\sigma σ ′ = Int ( w ) ∘ σ , then
π h ( A , σ ′ ) = dlog ( w ) ∧ h A + π h ( A , σ ) . (2) \pi_h(A,\sigma')=\operatorname{dlog}(w)\wedge h_A+\pi_h(A,\sigma). \tag*{(2)} π h ( A , σ ′ ) = dlog ( w ) ∧ h A + π h ( A , σ ) . (2)
Therefore, we can mimic Definition 1.3 by setting
Δ h ( A ) = π h ( A , σ ) + dlog ( F × ) ∧ h A ∈ H 3 ( F ) / dlog ( F × ) ∧ h A . \Delta_h(A)=\pi_h(A,\sigma)+\operatorname{dlog}(F^\times)\wedge h_A\in H^3(F)/\operatorname{dlog}(F^\times)\wedge h_A. Δ h ( A ) = π h ( A , σ ) + dlog ( F × ) ∧ h A ∈ H 3 ( F ) / dlog ( F × ) ∧ h A .
From the definitions, it is clear that the isomorphism (1) maps Δ h ( A ) ∈ H 3 ( F ) / dlog ( F × ) ∧ h A \Delta_h(A)\in H^3(F)/\operatorname{dlog}(F^\times)\wedge h_A Δ h ( A ) ∈ H 3 ( F ) / dlog ( F × ) ∧ h A to Δ ( A ) ∈ I ‾ q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q ‾ A \Delta(A)\in\overline{I}^3_q(F)/\langle\langle F^\times\rangle\rangle\overline{q}_A Δ ( A ) ∈ I q 3 ( F ) / ⟨⟨ F × ⟩⟩ q A .
The discussion in [22 §4a] can now be reproduced without any restriction on the characteristic: (2) shows that π h \pi_h π h defines a nondecomposable cohomological invariant of the adjoint symplectic group P G S p 8 \mathrm{PGSp}_8 PGSp 8 , and Δ h \Delta_h Δ h is nontrivial generically.
Descent of Brauer classes
Throughout this section, K K K denotes a quadratic extension of an arbitrary field F F F (of characteristic 2 2 2 ). We write N K / F N_{K/F} N K / F for the norm map N K / F : K × → F × N_{K/F}:K^\times\to F^\times N K / F : K × → F × (which is the squaring map sq K \operatorname{sq}_K sq K when K K K is inseparable over F F F ), and also for the map N K / F : Br 2 ( K ) → Br 2 ( F ) N_{K/F}:\operatorname{Br}_2(K)\to\operatorname{Br}_2(F) N K / F : Br 2 ( K ) → Br 2 ( F ) that is given by corestriction when K K K is separable over F F F , and that is functorially induced by the squaring field homomorphism sq K : K → F \operatorname{sq}_K:K\to F sq K : K → F when K K K is inseparable over F F F (see [16 §3.13, §4.1] or [12 §3] ).
Theorem 2.1. Let E E E be a central simple K K K -algebra such that 2 [ E ] = 0 2[E] = 0 2 [ E ] = 0 in Br ( K ) \operatorname{Br}(K) Br ( K ) . There exists a central simple F F F -algebra E ′ E' E ′ such that 2 [ E ′ ] = 0 2[E'] = 0 2 [ E ′ ] = 0 and [ E ′ ] K = [ E ] [E']_K = [E] [ E ′ ] K = [ E ] if and only if N K / F [ E ] = 0 N_{K/F}[E] = 0 N K / F [ E ] = 0 . When this condition holds, the algebra E ′ E' E ′ can be chosen of degree deg ( E ′ ) = 2 deg ( E ) \deg(E') = 2\deg(E) deg ( E ′ ) = 2 deg ( E ) .
The analogous statement in characteristic different from 2 is well-known, see [13 Th. 99.13] .
Proof. Suppose first that K K K is separable over F F F , and let ι \iota ι denote the nontrivial F F F -automorphism of K K K . If [ E ′ ] K = [ E ] [E']_K = [E] [ E ′ ] K = [ E ] for some central simple F F F -algebra E ′ E' E ′ , then N K / F [ E ] = 2 [ E ′ ] N_{K/F}[E] = 2[E'] N K / F [ E ] = 2 [ E ′ ] by [19 (3.13)(5)] . The “only if” part of the statement follows.
For the converse, recall from [19 (3.1)] that when 2 [ E ] = 0 2[E] = 0 2 [ E ] = 0 and N K / F [ E ] = 0 N_{K/F}[E] = 0 N K / F [ E ] = 0 , then E E E carries an involution σ \sigma σ of the first kind (i.e., whose restriction to K K K is the identity) and an involution τ \tau τ of the second kind, whose restriction to K K K is ι \iota ι . The composition τ ∘ σ ∘ τ \tau\circ\sigma\circ\tau τ ∘ σ ∘ τ is an involution of the first kind of E E E , therefore τ ∘ σ ∘ τ = Int ( u ) ∘ σ \tau\circ\sigma\circ\tau= \operatorname{Int}(u) \circ\sigma τ ∘ σ ∘ τ = Int ( u ) ∘ σ for some invertible element u ∈ E × u \in E^{\times} u ∈ E × satisfying σ ( u ) = u \sigma(u) = u σ ( u ) = u . It follows that the ι \iota ι -semilinear automorphism τ ∘ σ \tau\circ\sigma τ ∘ σ of E E E satisfies ( τ ∘ σ ) 2 = Int ( u ) (\tau\circ\sigma)^2 = \operatorname{Int}(u) ( τ ∘ σ ) 2 = Int ( u ) ; in particular, it commutes with Int ( u ) \operatorname{Int}(u) Int ( u ) . Hence, Int τ ∘ σ ( u ) = Int ( u ) \operatorname{Int}_{\tau\circ\sigma(u)} = \operatorname{Int}(u) Int τ ∘ σ ( u ) = Int ( u ) , so that τ ( u ) = τ ∘ σ ( u ) = λ u \tau(u) = \tau\circ\sigma(u) = \lambda u τ ( u ) = τ ∘ σ ( u ) = λ u for some λ ∈ K × \lambda\in K^{\times} λ ∈ K × . Applying τ \tau τ to the equality τ ( u ) = λ u \tau(u) = \lambda u τ ( u ) = λ u , we get that λ ι ( λ ) = 1 \lambda\iota(\lambda) = 1 λ ι ( λ ) = 1 . Hilbert’s Theorem 90 yields μ ∈ K × \mu\in K^{\times} μ ∈ K × such that λ = μ ι ( μ ) − 1 \lambda= \mu\iota(\mu)^{-1} λ = μ ι ( μ ) − 1 ; substituting μ u \mu u μu for u u u , we may assume σ ( u ) = τ ( u ) = u \sigma(u) = \tau(u) = u σ ( u ) = τ ( u ) = u . Define an F F F -algebra E ′ E' E ′ as a “generalized crossed product”
E ′ = E ⊕ E z E' = E \oplus Ez E ′ = E ⊕ E z
where the multiplication extends the multiplication in E E E by the following rules:
z 2 = u and z x = σ ∘ τ ( x ) z for x ∈ E . z^2 = u \qquad\text{and}\qquad zx = \sigma\circ\tau(x)z \qquad\text{for } x \in E. z 2 = u and z x = σ ∘ τ ( x ) z for x ∈ E .
By [2 Th. 11.11] , E ′ E' E ′ is a central simple F F F -algebra containing E E E as the centralizer of K K K , hence [ E ′ ] K = [ E ] [E']_K = [E] [ E ′ ] K = [ E ] . Moreover, setting σ ′ ( z ) = z \sigma'(z) = z σ ′ ( z ) = z defines an involution σ ′ \sigma' σ ′ on E ′ E' E ′ extending σ \sigma σ on E E E , hence 2 [ E ′ ] = 0 2[E'] = 0 2 [ E ′ ] = 0 . The theorem is thus proved in the case where K K K is separable over F F F .
Now, suppose K K K is inseparable over F F F . By a theorem of Hochschild (see [16 Th. 4.1.5] ), [ E ] = [ E ′ ] K [E] = [E']_K [ E ] = [ E ′ ] K for some central simple F F F -algebra E ′ E' E ′ . Since N K / F N_{K/F} N K / F is induced by the squaring map, we have N K / F [ E ] = 2 [ E ′ ] N_{K/F}[E] = 2[E'] N K / F [ E ] = 2 [ E ′ ] (see [16 Th. 4.1.2] ). This proves the first part of the statement.
Note that every field extension of K K K that splits E E E also splits E ′ E' E ′ , hence E ′ E' E ′ is split by field extensions of F F F of degree dividing 2 deg ( E ) 2\deg(E) 2 deg ( E ) . Therefore, the central simple F F F -algebra E ′ E' E ′ may be assumed to have deg ( E ′ ) = 2 deg ( E ) \deg(E') = 2\deg(E) deg ( E ′ ) = 2 deg ( E ) . (This is also clear from the proof above in the case where K K K is separable over F F F or from the proof of Hochschild’s theorem in [16 ] .) □
When E E E is a quaternion algebra, Theorem 2.1 can be improved as follows:
Proposition 2.2. Let Q Q Q be a quaternion algebra over K K K .
(a) If N K / F [ Q ] = 0 N_{K/F}[Q] = 0 N K / F [ Q ] = 0 , then there exists a quaternion F F F -algebra Q 0 Q_0 Q 0 such that Q ≃ Q 0 ⊗ F K Q \simeq Q_0 \otimes_F K Q ≃ Q 0 ⊗ F K .
(b) If Q ≃ ( α , β ) K Q \simeq(\alpha,\beta)_K Q ≃ ( α , β ) K with α ∈ F \alpha\in F α ∈ F and N K / F [ Q ] = 0 N_{K/F}[Q] = 0 N K / F [ Q ] = 0 , then there exists γ ∈ F × \gamma\in F^{\times} γ ∈ F × such that Q ≃ [ α , γ ) F ⊗ F K Q \simeq[\alpha,\gamma)_F \otimes_F K Q ≃ [ α , γ ) F ⊗ F K .
Proof. We first prove (b). The argument depends on whether K K K is separable over F F F or not. By assumption, the algebra [ α , N K / F ( β ) ] [\alpha,N_{K/F}(\beta)] [ α , N K / F ( β )] is split.
Assume first K K K is separable, and let ι \iota ι denote the nontrivial F F F -automorphism of K K K . If α ∈ ℘ ( K ) \alpha\in\wp(K) α ∈ ℘ ( K ) , then [ α , β ) K [\alpha,\beta)_K [ α , β ) K is split and we may choose γ = 1 \gamma= 1 γ = 1 . If α ∉ ℘ ( K ) \alpha\notin\wp(K) α ∈ / ℘ ( K ) , then K ( ℘ − 1 ( α ) ) K(\wp^{-1}(\alpha)) K ( ℘ − 1 ( α )) is a biquadratic field extension of F F F and N K / F ( β ) N_{K/F}(\beta) N K / F ( β ) is a norm for the quadratic separable extension F ( ℘ − 1 ( α ) ) / F F(\wp^{-1}(\alpha))/F F ( ℘ − 1 ( α )) / F . By the “biquadratic lemma” [26 Example 5.1] , there exist λ ∈ F × \lambda\in F^{\times} λ ∈ F × and μ ∈ K ( ℘ − 1 ( α ) ) \mu\in K(\wp^{-1}(\alpha)) μ ∈ K ( ℘ − 1 ( α )) such that
N K / F ( β ) = λ 2 N K ( ℘ − 1 ( α ) ) / F ( μ ) = N K / F ( λ N K ( ℘ − 1 ( α ) ) / K ( μ ) ) . N_{K/F}(\beta) = \lambda^2 N_{K(\wp^{-1}(\alpha))/F}(\mu) = N_{K/F}\left(\lambda N_{K(\wp^{-1}(\alpha))/K}(\mu)\right). N K / F ( β ) = λ 2 N K ( ℘ − 1 ( α )) / F ( μ ) = N K / F ( λ N K ( ℘ − 1 ( α )) / K ( μ ) ) .
Hilbert’s Theorem 90 then yields ν ∈ K × \nu\in K^{\times} ν ∈ K × such that
β = λ N K ( ℘ − 1 ( α ) ) / K ( μ ) ν ι ( ν ) − 1 = λ N K / F ( ν ) − 1 ν 2 N K ( ℘ − 1 ( α ) ) / K ( μ ) \beta=\lambda N_{K(\wp^{-1}(\alpha))/K}(\mu)\nu\iota(\nu)^{-1}=\lambda N_{K/F}(\nu)^{-1}\nu^{2}N_{K(\wp^{-1}(\alpha))/K}(\mu) β = λ N K ( ℘ − 1 ( α )) / K ( μ ) ν ι ( ν ) − 1 = λ N K / F ( ν ) − 1 ν 2 N K ( ℘ − 1 ( α )) / K ( μ )
and (b) holds with γ = λ N K / F ( ν ) − 1 \gamma=\lambda N_{K/F}(\nu)^{-1} γ = λ N K / F ( ν ) − 1 .
Now, assume K K K is inseparable over F F F . Since [ α , N K / F ( β ) ) F [\alpha,N_{K/F}(\beta))_F [ α , N K / F ( β ) ) F is split, there exist λ , μ ∈ F \lambda,\mu\in F λ , μ ∈ F such that
N K / F ( β ) = β 2 = λ 2 + λ μ + α μ 2 . N_{K/F}(\beta)=\beta^{2}=\lambda^{2}+\lambda\mu+\alpha\mu^{2}. N K / F ( β ) = β 2 = λ 2 + λ μ + α μ 2 .
If μ = 0 \mu=0 μ = 0 , then β = λ \beta=\lambda β = λ and (b) holds with γ = λ \gamma=\lambda γ = λ . If μ ≠ 0 \mu\ne0 μ = 0 , then
β 2 = μ − 2 ( ( λ + β ) 2 + ( λ + β ) μ + α μ 2 ) 2 , \beta^{2}=\mu^{-2}\left((\lambda+\beta)^{2}+(\lambda+\beta)\mu+\alpha\mu^{2}\right)^{2}, β 2 = μ − 2 ( ( λ + β ) 2 + ( λ + β ) μ + α μ 2 ) 2 ,
hence
β = μ − 1 ( ( λ + β ) 2 + ( λ + β ) μ + α μ 2 ) , \beta=\mu^{-1}\left((\lambda+\beta)^{2}+(\lambda+\beta)\mu+\alpha\mu^{2}\right), β = μ − 1 ( ( λ + β ) 2 + ( λ + β ) μ + α μ 2 ) ,
and (b) holds with γ = μ − 1 \gamma=\mu^{-1} γ = μ − 1 .
We deduce (a) from (b) when K K K is inseparable over F F F , as follows: we write Q = [ α 0 , β ) K Q=[\alpha_{0},\beta)_K Q = [ α 0 , β ) K for some α 0 ∈ K \alpha_{0}\in K α 0 ∈ K and β ∈ K × \beta\in K^{\times} β ∈ K × , and observe that Q = [ α 0 2 , β ) K Q=[\alpha_{0}^{2},\beta)_K Q = [ α 0 2 , β ) K with α 0 2 ∈ F \alpha_{0}^{2}\in F α 0 2 ∈ F because α 0 2 ≡ α 0 ( m o d ℘ ( K ) ) \alpha_{0}^{2}\equiv\alpha_{0}\pmod{\wp(K)} α 0 2 ≡ α 0 ( mod ℘ ( K )) and K 2 ⊂ F K^{2}\subset F K 2 ⊂ F . Since N K / F [ Q ] = [ [ α 0 2 , N K / F ( β ) ) F ] = 0 N_{K/F}[Q]=[[\alpha_{0}^{2},N_{K/F}(\beta))_F]=0 N K / F [ Q ] = [[ α 0 2 , N K / F ( β ) ) F ] = 0 , (b) yields γ ∈ F × \gamma\in F^{\times} γ ∈ F × such that Q ≃ [ α 0 2 , γ ) F ⊗ F K Q\simeq[\alpha_{0}^{2},\gamma)_F\otimes_F K Q ≃ [ α 0 2 , γ ) F ⊗ F K .
When K K K is separable over F F F , the hypothesis N K / F [ Q ] = 0 N_{K/F}[Q]=0 N K / F [ Q ] = 0 implies that Q Q Q carries an involution of unitary type by [19 (3.1)] . The existence of a quaternion F F F -algebra Q 0 Q_0 Q 0 such that Q ≃ Q 0 ⊗ F K Q\simeq Q_0\otimes_F K Q ≃ Q 0 ⊗ F K then follows from a theorem of Albert [19 (2.22)] .
The last result in this section characterizes the Brauer classes of quaternion algebras that satisfy a biquadratic descent property. By definition, the biquadratic field extensions we consider can be represented as tensor products of two quadratic extensions. We exclude the case of purely inseparable field extensions, whose features are significantly different: see the Appendix. Thus, the biquadratic field extensions L / F L/F L / F we consider are either separable—and then Galois with elementary abelian Galois group—or of the form L ≃ K 1 ⊗ F K 2 L\simeq K_1\otimes_F K_2 L ≃ K 1 ⊗ F K 2 where K 1 K_1 K 1 is a separable quadratic extension of F F F and K 2 K_2 K 2 is an inseparable quadratic extension of F F F . In the latter case, K 1 K_1 K 1 is uniquely determined as the separable closure of F F F in L L L , and it is easily verified that K 2 = { ℓ ∈ L ∣ ℓ 2 ∈ F } K_2=\{\ell\in L\mid\ell^{2}\in F\} K 2 = { ℓ ∈ L ∣ ℓ 2 ∈ F } , hence K 1 K_1 K 1 and K 2 K_2 K 2 are the only two quadratic extensions of F F F in L L L . By contrast, when L L L is separable over F F F it contains three quadratic extensions of F F F .
Theorem 2.3. Let L L L be a biquadratic field extension of F F F and let Q Q Q be a quaternion algebra over L L L . If L L L is not purely inseparable over F F F , the following are equivalent:
(i) there exists a central simple F F F -algebra A A A such that 2 [ A ] = 0 2[A]=0 2 [ A ] = 0 and [ A ] L = [ Q ] [A]_L=[Q] [ A ] L = [ Q ] ;
(ii) N L / K [ Q ] = 0 N_{L/K}[Q]=0 N L / K [ Q ] = 0 for every quadratic extension K K K of F F F in L L L .
When these conditions hold, the algebra A A A can be chosen of degree 8 8 8 , so that Q Q Q is isomorphic to the centralizer of L L L in A A A .
The proof below uses the following “inseparable common slot lemma” (a “separable common slot lemma” is given in [23 Lemma 3] ):
Lemma 2.4. Let α 1 , α 2 ∈ F \alpha_1,\alpha_2\in F α 1 , α 2 ∈ F and β 1 , β 2 ∈ F × \beta_1,\beta_2\in F^{\times} β 1 , β 2 ∈ F × be such that the quaternion F F F -algebras [ α 1 , β 1 ) F [\alpha_1,\beta_1)_F [ α 1 , β 1 ) F and [ α 2 , β 2 ) F [\alpha_2,\beta_2)_F [ α 2 , β 2 ) F are isomorphic. There exists γ ∈ F × \gamma\in F^{\times} γ ∈ F × such that
[ α 1 , β 1 ) F ≃ [ α 1 , γ ) F ≃ [ α 2 , γ ) F ≃ [ α 2 , β 2 ) F . [\alpha_1,\beta_1)_F\simeq[\alpha_1,\gamma)_F\simeq[\alpha_2,\gamma)_F\simeq[\alpha_2,\beta_2)_F. [ α 1 , β 1 ) F ≃ [ α 1 , γ ) F ≃ [ α 2 , γ ) F ≃ [ α 2 , β 2 ) F .
Proof. Let Q Q Q be a quaternion F F F -algebra isomorphic to [ α 1 , β 1 ) F [\alpha_1,\beta_1)_F [ α 1 , β 1 ) F and [ α 2 , β 2 ) F [\alpha_2,\beta_2)_F [ α 2 , β 2 ) F . We may assume Q Q Q is a division algebra, for otherwise we may readily choose γ = 1 \gamma=1 γ = 1 . By hypothesis, there exist i 1 , j 1 , i 2 , j 2 ∈ Q i_1,j_1,i_2,j_2\in Q i 1 , j 1 , i 2 , j 2 ∈ Q such that
i r 2 + i r = α r , j r i r + i r j r = j r and j r 2 = β r for r = 1 , 2. i_r^{2}+i_r=\alpha_r,\qquad j_ri_r+i_rj_r=j_r\quad\text{and}\quad j_r^{2}=\beta_r\qquad\text{for }r=1,2. i r 2 + i r = α r , j r i r + i r j r = j r and j r 2 = β r for r = 1 , 2.
The vector space of reduced trace zero elements in Q Q Q has dimension 3 and contains the 2-dimensional subspaces F ( i 1 ) j 1 F(i_1)j_1 F ( i 1 ) j 1 and F ( i 2 ) j 2 F(i_2)j_2 F ( i 2 ) j 2 . By dimension count, these subspaces have a nonzero intersection. If j ∈ F ( i 1 ) j 1 ∩ F ( i 2 ) j 2 j \in F(i_1)j_1 \cap F(i_2)j_2 j ∈ F ( i 1 ) j 1 ∩ F ( i 2 ) j 2 is nonzero, then the lemma holds with γ = j 2 \gamma= j^2 γ = j 2 .
Proof of Theorem 2.3. (i) ⇒ \Rightarrow ⇒ (ii) Assuming (i), we have [ A K ] L = [ Q ] [A_K]_L = [Q] [ A K ] L = [ Q ] for every quadratic extension K K K of F F F in L L L , hence Theorem 2.1 shows that (ii) holds.
(ii) ⇒ \Rightarrow ⇒ (i) The proof depends on whether L L L is separable over F F F or not. Assume first L L L is separable, and let L = F ( ℘ − 1 ( α 1 ) , ℘ − 1 ( α 2 ) ) L = F(\wp^{-1}(\alpha_1),\wp^{-1}(\alpha_2)) L = F ( ℘ − 1 ( α 1 ) , ℘ − 1 ( α 2 )) for some α 1 , α 2 ∈ F \alpha_1,\alpha_2 \in F α 1 , α 2 ∈ F . The quadratic extensions of F F F in L L L are then K 1 = F ( ℘ − 1 ( α 1 ) ) K_1 = F(\wp^{-1}(\alpha_1)) K 1 = F ( ℘ − 1 ( α 1 )) , K 2 = F ( ℘ − 1 ( α 2 ) ) K_2 = F(\wp^{-1}(\alpha_2)) K 2 = F ( ℘ − 1 ( α 2 )) and K 3 = F ( ℘ − 1 ( α 1 + α 2 ) ) K_3 = F(\wp^{-1}(\alpha_1+\alpha_2)) K 3 = F ( ℘ − 1 ( α 1 + α 2 )) . Assuming N L / K 3 [ Q ] = 0 N_{L/K_3}[Q] = 0 N L / K 3 [ Q ] = 0 , Proposition 2.2(a) yields a quaternion K 3 K_3 K 3 -algebra Q 3 Q_3 Q 3 such that Q = Q 3 ⊗ K 3 L Q = Q_3 \otimes_{K_3} L Q = Q 3 ⊗ K 3 L . Then N L / K i [ Q ] = ( N K 3 / F [ Q 3 ] ) K i N_{L/K_i}[Q] = (N_{K_3/F}[Q_3])_{K_i} N L / K i [ Q ] = ( N K 3 / F [ Q 3 ] ) K i for i = 1 , 2 i = 1,2 i = 1 , 2 , hence (ii) shows that N K 3 / F [ Q 3 ] N_{K_3/F}[Q_3] N K 3 / F [ Q 3 ] is split by K 1 K_1 K 1 and K 2 K_2 K 2 . By the inseparable common slot lemma 2.4, there exists γ ∈ F × \gamma\in F^\times γ ∈ F × such that
N K 3 / F [ Q 3 ] = [ [ α 1 , γ ) F ] = [ [ α 2 , γ ) F ] . (3) N_{K_3/F}[Q_3] = [[\alpha_1,\gamma)_F] = [[\alpha_2,\gamma)_F]. \tag*{(3)} N K 3 / F [ Q 3 ] = [[ α 1 , γ ) F ] = [[ α 2 , γ ) F ] . (3)
Then [ α 1 + α 2 , γ ) F [\alpha_1+\alpha_2,\gamma)_F [ α 1 + α 2 , γ ) F is split, hence there exists β ∈ K 3 × \beta\in K_3^\times β ∈ K 3 × such that γ = N K 3 / F ( β ) \gamma= N_{K_3/F}(\beta) γ = N K 3 / F ( β ) . From (3) it follows that N K 3 / F [ Q 3 ] = N K 3 / F [ [ α 1 , β ) K 3 ] N_{K_3/F}[Q_3] = N_{K_3/F}[[\alpha_1,\beta)_{K_3}] N K 3 / F [ Q 3 ] = N K 3 / F [[ α 1 , β ) K 3 ] , hence
N K 3 / F [ Q 3 ⊗ K 3 [ α 1 , β ) K 3 ] = 0. N_{K_3/F}[Q_3 \otimes_{K_3} [\alpha_1,\beta)_{K_3}] = 0. N K 3 / F [ Q 3 ⊗ K 3 [ α 1 , β ) K 3 ] = 0.
Theorem 2.1 yields a central simple F F F -algebra A A A of degree 8 such that 2 [ A ] = 0 2[A] = 0 2 [ A ] = 0 and [ A ] K 3 = [ Q 3 ⊗ K 3 [ α 1 , β ) K 3 ] [A]_{K_3} = [Q_3 \otimes_{K_3} [\alpha_1,\beta)_{K_3}] [ A ] K 3 = [ Q 3 ⊗ K 3 [ α 1 , β ) K 3 ] . We may then identify Q 3 ⊗ K 3 [ α 1 , β ) K 3 Q_3 \otimes_{K_3} [\alpha_1,\beta)_{K_3} Q 3 ⊗ K 3 [ α 1 , β ) K 3 with the centralizer of K 3 K_3 K 3 in A A A . Identifying L L L with K 3 ( ℘ − 1 ( α 1 ) ) ⊂ [ α 1 , β ) K 3 K_3(\wp^{-1}(\alpha_1)) \subset[\alpha_1,\beta)_{K_3} K 3 ( ℘ − 1 ( α 1 )) ⊂ [ α 1 , β ) K 3 , we see that the centralizer of L L L in A A A is Q 3 ⊗ K 3 L Q_3 \otimes_{K_3} L Q 3 ⊗ K 3 L . This concludes the proof in the case where L L L is separable over F F F since Q = Q 3 ⊗ K 3 L Q = Q_3 \otimes_{K_3} L Q = Q 3 ⊗ K 3 L .
Now, assume L = F ( ℘ − 1 ( α ) , ρ ) L = F(\wp^{-1}(\alpha),\sqrt{\rho}) L = F ( ℘ − 1 ( α ) , ρ ) for some α ∈ F \alpha\in F α ∈ F and ρ ∈ F × \rho\in F^\times ρ ∈ F × , hence the quadratic extensions of F F F in L L L are K 1 = F ( ℘ − 1 ( α ) ) K_1 = F(\wp^{-1}(\alpha)) K 1 = F ( ℘ − 1 ( α )) and K 2 = F ( ρ ) K_2 = F(\sqrt{\rho}) K 2 = F ( ρ ) . Since N L / K 2 [ Q ] = 0 N_{L/K_2}[Q] = 0 N L / K 2 [ Q ] = 0 , Proposition 2.2 yields a quaternion K 2 K_2 K 2 -algebra Q 2 Q_2 Q 2 such that Q = Q 2 ⊗ K 2 L Q = Q_2 \otimes_{K_2} L Q = Q 2 ⊗ K 2 L . Then N L / K 1 [ Q ] = ( N K 2 / F [ Q 2 ] ) K 1 N_{L/K_1}[Q] = (N_{K_2/F}[Q_2])_{K_1} N L / K 1 [ Q ] = ( N K 2 / F [ Q 2 ] ) K 1 , hence (ii) shows that N K 2 / F [ Q 2 ] N_{K_2/F}[Q_2] N K 2 / F [ Q 2 ] is split by K 1 K_1 K 1 . But ( N K 2 / F [ Q 2 ] ) K 2 = 2 [ Q 2 ] = 0 (N_{K_2/F}[Q_2])_{K_2} = 2[Q_2] = 0 ( N K 2 / F [ Q 2 ] ) K 2 = 2 [ Q 2 ] = 0 , hence N K 2 / F [ Q 2 ] N_{K_2/F}[Q_2] N K 2 / F [ Q 2 ] is also split by K 2 K_2 K 2 . Therefore, N K 2 / F [ Q 2 ] N_{K_2/F}[Q_2] N K 2 / F [ Q 2 ] is the Brauer class of a quaternion F F F -algebra Q 0 Q_0 Q 0 containing K 1 K_1 K 1 and K 2 K_2 K 2 as maximal subfields. Let i , r ∈ Q 0 i,r \in Q_0 i , r ∈ Q 0 be such that i 2 + i = α i^2+i=\alpha i 2 + i = α and r 2 = ρ r^2=\rho r 2 = ρ , and let j = i r + r i ∈ Q 0 j=ir+ri \in Q_0 j = i r + r i ∈ Q 0 . If j = 0 j=0 j = 0 , then i i i commutes with r r r hence i ∈ F ( r ) i \in F(r) i ∈ F ( r ) . This is impossible since i i i is separable over F F F whereas F ( r ) = K 2 F(r)=K_2 F ( r ) = K 2 is an inseparable extension; hence j ≠ 0 j \ne0 j = 0 . Since r 2 ∈ F × r^2 \in F^\times r 2 ∈ F × , it is readily seen that j j j commutes with r r r , hence j ∈ F ( r ) = K 2 j \in F(r)=K_2 j ∈ F ( r ) = K 2 . On the other hand, computation yields i j + j i = j ij+ji=j ij + j i = j , hence j ∉ F j \notin F j ∈ / F and i , j i,j i , j generate Q 0 Q_0 Q 0 as a quaternion F F F -algebra [ α , β ) F [\alpha,\beta)_F [ α , β ) F with β = j 2 ∈ F × \beta=j^2 \in F^\times β = j 2 ∈ F × and K 2 = F ( j ) K_2=F(j) K 2 = F ( j ) . Now, N K 2 / F [ [ α , j ) K 2 ] = [ [ α , β ) F ] = N K 2 / F [ Q 2 ] N_{K_2/F}[[\alpha,j)_{K_2}] = [[\alpha,\beta)_F] = N_{K_2/F}[Q_2] N K 2 / F [[ α , j ) K 2 ] = [[ α , β ) F ] = N K 2 / F [ Q 2 ] , hence
N K 2 / F [ Q 2 ⊗ K 2 [ α , j ) K 2 ] = 0. N_{K_2/F}[Q_2 \otimes_{K_2} [\alpha,j)_{K_2}] = 0. N K 2 / F [ Q 2 ⊗ K 2 [ α , j ) K 2 ] = 0.
Theorem 2.1 yields a central simple F F F -algebra A A A of degree 8 such that 2 [ A ] = 0 2[A] = 0 2 [ A ] = 0 and [ A ] K 2 = [ Q 2 ⊗ K 2 [ α , j ) K 2 ] [A]_{K_2} = [Q_2 \otimes_{K_2} [\alpha,j)_{K_2}] [ A ] K 2 = [ Q 2 ⊗ K 2 [ α , j ) K 2 ] . We may assume A A A contains K 2 K_2 K 2 , with centralizer Q 2 ⊗ K 2 [ α , j ) K 2 Q_2 \otimes_{K_2} [\alpha,j)_{K_2} Q 2 ⊗ K 2 [ α , j ) K 2 . Identifying L L L with K 2 ( ℘ − 1 ( α ) ) ⊂ [ α , j ) K 2 K_2(\wp^{-1}(\alpha)) \subset[\alpha,j)_{K_2} K 2 ( ℘ − 1 ( α )) ⊂ [ α , j ) K 2 , we see that the centralizer of L L L in A A A is Q 2 ⊗ K 2 L Q_2 \otimes_{K_2} L Q 2 ⊗ K 2 L . This completes the proof since Q 2 ⊗ K 2 L = Q Q_2 \otimes_{K_2} L = Q Q 2 ⊗ K 2 L = Q .
Quadratic descent of biquaternion algebras
Throughout this section, K K K denotes a quadratic extension of the field F F F of characteristic 2, and B B B is a biquaternion K K K -algebra. Theorem 2.1 yields a necessary and sufficient condition for the Brauer class [ B ] [B] [ B ] to descend to Br 2 ( F ) \operatorname{Br}_2(F) Br 2 ( F ) , but even when this condition holds the algebra B B B itself may not descend to a biquaternion F F F -algebra. Our goal in this section is to define an invariant δ K / F ( B ) \delta_{K/F}(B) δ K / F ( B ) that vanishes if and only if this stronger property holds. As the definition of the Δ \Delta Δ -invariant in Section 1, our definition of δ K / F ( B ) \delta_{K/F}(B) δ K / F ( B ) relies on discriminant Pfister forms of symplectic involutions on algebras of degree 8. It turns out to be equivalent to the definition based on Albert forms introduced by Barry–Chapman–Laghribi [8 §3] when K K K is separable over F F F and by Laghribi–Lorenz [20 Prop. 6.4] when K K K is inseparable.
We use the same notation N K / F N_{K/F} N K / F as in the preceding section, and fix a nonzero F F F -linear functional t K : K → F t_K: K \to F t K : K → F such that t K ( 1 ) = 0 t_K(1)=0 t K ( 1 ) = 0 . (If K K K is separable over F F F , we may take for t K t_K t K the trace map T K / F T_{K/F} T K / F .) Throughout, we assume N K / F [ B ] = 0 N_{K/F}[B]=0 N K / F [ B ] = 0 , so Theorem 2.1 shows that there exists a central simple F F F -algebra A A A of degree 8 such that [ B ] = [ A ] K [B]=[A]_K [ B ] = [ A ] K and 2 [ A ] = 0 2[A]=0 2 [ A ] = 0 . Let σ B \sigma_B σ B be a symplectic involution on B B B . By [19 (4.14)] , σ B \sigma_B σ B extends (not uniquely) to an involution σ A \sigma_A σ A on A A A , which is necessarily symplectic since 1 ∈ Symd ( σ B ) ⊂ Symd ( σ A ) 1 \in\operatorname{Symd}(\sigma_B) \subset\operatorname{Symd}(\sigma_A) 1 ∈ Symd ( σ B ) ⊂ Symd ( σ A ) . The following result shows that the discriminant Pfister form π ( σ A ) \pi(\sigma_A) π ( σ A ) can be determined from σ B \sigma_B σ B . In particular, it does not depend on the choices of A A A and σ A \sigma_A σ A .
Proposition 3.1. The transfer ( t K ) ∗ ( Nrp σ B ) (t_K)_*(\operatorname{Nrp}_{\sigma_B}) ( t K ) ∗ ( Nrp σ B ) is Witt-equivalent to a scalar multiple of π ( σ A ) \pi(\sigma_A) π ( σ A ) .
Proof. The argument branches into two parts, depending on whether K K K is separable over F F F or not. First, assume K K K is separable. By [9 Th. 6.10] , there exists a biquadratic étale F F F -algebra L ⊂ Symd ( σ A ) L \subset\operatorname{Symd}(\sigma_A) L ⊂ Symd ( σ A ) containing K K K as a subalgebra fixed under an element of order 2 in its Galois group. Let α \alpha α denote this order 2 element and define W W W as previously:
W = { w ∈ Symd ( σ A ) ∣ w ℓ = α ( ℓ ) w for all ℓ ∈ L } . W = \{w \in\operatorname{Symd}(\sigma_A) \mid w\ell= \alpha(\ell)w \text{ for all } \ell\in L\}. W = { w ∈ Symd ( σ A ) ∣ w ℓ = α ( ℓ ) w for all ℓ ∈ L } .
Note that W ⊂ Symd ( σ B ) W \subset\operatorname{Symd}(\sigma_B) W ⊂ Symd ( σ B ) because K K K is fixed under α \alpha α and Symd ( σ B ) = B ∩ Symd ( σ A ) \operatorname{Symd}(\sigma_B)=B \cap\operatorname{Symd}(\sigma_A) Symd ( σ B ) = B ∩ Symd ( σ A ) by [9 Prop. 6.1] . Letting sq W : W → K \operatorname{sq}_W: W \to K sq W : W → K denote the squaring map sq W ( w ) = w 2 \operatorname{sq}_W(w)=w^2 sq W ( w ) = w 2 , the transfer ( t K ) ∗ ( sq W ) (t_K)_*(\operatorname{sq}_W) ( t K ) ∗ ( sq W ) is a scalar multiple of π ( σ A ) \pi(\sigma_A) π ( σ A ) , by definition of the discriminant Pfister form. By [9 Prop. 5.6 and 6.5] , we have
( Symd ( σ B ) , Nrp σ B ) = ( L , N L / K ) ⊥ ( W , sq W ) , (\operatorname{Symd}(\sigma_B),\operatorname{Nrp}_{\sigma_B})=(L,N_{L/K})\perp(W,\operatorname{sq}_W), ( Symd ( σ B ) , Nrp σ B ) = ( L , N L / K ) ⊥ ( W , sq W ) ,
hence ( t K ) ∗ ( Nrp σ B ) = ( t K ) ∗ ( N L / K ) ⊥ ( t K ) ∗ ( sq W ) (t_K)_*(\operatorname{Nrp}_{\sigma_B})=(t_K)_*(N_{L/K})\perp(t_K)_*(\operatorname{sq}_W) ( t K ) ∗ ( Nrp σ B ) = ( t K ) ∗ ( N L / K ) ⊥ ( t K ) ∗ ( sq W ) . But L L L is a biquadratic F F F -algebra, hence the norm form N L / K N_{L/K} N L / K is defined over F F F , and ( t K ) ∗ ( N L / K ) (t_K)_*(N_{L/K}) ( t K ) ∗ ( N L / K ) is hyperbolic. Thus, the proof is complete in the case where K K K is separable over F F F .
In the case where K K K is inseparable over F F F , we need two lemmas.
Lemma 3.2. Assume K K K is an inseparable quadratic extension of F F F . There exists i 1 ∈ Symd ( σ B ) i_1 \in\operatorname{Symd}(\sigma_B) i 1 ∈ Symd ( σ B ) such that i 1 2 + i 1 ∈ F i_1^2+i_1 \in F i 1 2 + i 1 ∈ F . For any such i 1 i_1 i 1 , the K K K -vector space
V = { v ∈ Symd ( σ B ) ∣ v i 1 = ( i 1 + 1 ) v } V=\{v \in\operatorname{Symd}(\sigma_B) \mid vi_1=(i_1+1)v\} V = { v ∈ Symd ( σ B ) ∣ v i 1 = ( i 1 + 1 ) v } has dimension 4 and the squaring map sq V : v ↦ v 2 \operatorname{sq}_V: v \mapsto v^2 sq V : v ↦ v 2 is a nonsingular quadratic form V → K V \to K V → K whose transfer ( t K ) ∗ ( sq V ) (t_K)_*(\operatorname{sq}_V) ( t K ) ∗ ( sq V ) is Witt-equivalent to ( t K ) ∗ ( Nrp σ B ) (t_K)_*(\operatorname{Nrp}_{\sigma_B}) ( t K ) ∗ ( Nrp σ B ) .
Proof. Consider the 5-dimensional K K K -vector space Symd ( σ B ) 0 : = ker Trp σ B \operatorname{Symd}(\sigma_B)^0 := \ker\operatorname{Trp}_{\sigma_B} Symd ( σ B ) 0 := ker Trp σ B . By [19 (15.16)] , the restriction φ \varphi φ of Nrp σ B \operatorname{Nrp}_{\sigma_B} Nrp σ B to Symd ( σ B ) 0 \operatorname{Symd}(\sigma_B)^0 Symd ( σ B ) 0 is a nondegenerate quadratic form, and the inclusion Symd ( σ B ) 0 ↪ B \operatorname{Symd}(\sigma_B)^0 \hookrightarrow B Symd ( σ B ) 0 ↪ B induces a canonical isomorphism of algebras with involution from the even Clifford algebra C 0 ( φ ) C_0(\varphi) C 0 ( φ ) with its canonical involution γ \gamma γ to ( B , σ B ) (B,\sigma_B) ( B , σ B ) . If e 1 e_1 e 1 , f 1 f_1 f 1 and e 2 e_2 e 2 , f 2 f_2 f 2 are symplectic bases of two mutually orthogonal nonsingular planes in Symd ( σ B ) 0 \operatorname{Symd}(\sigma_B)^0 Symd ( σ B ) 0 , then in the even Clifford algebra e 1 f 1 + f 1 e 1 = e 2 f 2 + f 2 e 2 = 1 e_{1}f_{1}+f_{1}e_{1}=e_{2}f_{2}+f_{2}e_{2}=1 e 1 f 1 + f 1 e 1 = e 2 f 2 + f 2 e 2 = 1 , and e 1 f 1 e_{1}f_{1} e 1 f 1 and e 2 f 2 e_{2}f_{2} e 2 f 2 commute. Hence, we have
γ ( e 1 f 1 e 2 f 2 ) = f 2 e 2 f 1 e 1 = f 1 e 1 + e 2 f 2 + e 1 f 1 e 2 f 2 (4) \gamma(e_{1}f_{1}e_{2}f_{2})=f_{2}e_{2}f_{1}e_{1}=f_{1}e_{1}+e_{2}f_{2}+e_{1}f_{1}e_{2}f_{2} \tag*{(4)} γ ( e 1 f 1 e 2 f 2 ) = f 2 e 2 f 1 e 1 = f 1 e 1 + e 2 f 2 + e 1 f 1 e 2 f 2 (4)
and
( f 1 e 1 + e 2 f 2 ) 2 = f 1 e 1 + e 2 f 2 + φ ( e 1 ) φ ( f 1 ) + φ ( e 2 ) φ ( f 2 ) . (f_{1}e_{1}+e_{2}f_{2})^{2}=f_{1}e_{1}+e_{2}f_{2}+\varphi(e_{1})\varphi(f_{1})+\varphi(e_{2})\varphi(f_{2}). ( f 1 e 1 + e 2 f 2 ) 2 = f 1 e 1 + e 2 f 2 + φ ( e 1 ) φ ( f 1 ) + φ ( e 2 ) φ ( f 2 ) .
Let i 1 ∈ B ∖ K i_{1}\in B\setminus K i 1 ∈ B ∖ K be the image of f 1 e 1 + e 2 f 2 f_{1}e_{1}+e_{2}f_{2} f 1 e 1 + e 2 f 2 under the canonical isomorphism C 0 ( φ ) ≃ B C_{0}(\varphi)\simeq B C 0 ( φ ) ≃ B . From (4) it follows that i 1 ∈ Symd ( σ B ) i_{1}\in\operatorname{Symd}(\sigma_{B}) i 1 ∈ Symd ( σ B ) and i 1 2 + i 1 ∈ K i_{1}^{2}+i_{1}\in K i 1 2 + i 1 ∈ K . Substituting i 1 2 i_{1}^{2} i 1 2 for i 1 i_{1} i 1 , we may (and will) assume that i 1 2 + i 1 ∈ F i_{1}^{2}+i_{1}\in F i 1 2 + i 1 ∈ F , hence F ( i 1 ) ⊂ Symd ( σ B ) F(i_{1})\subset\operatorname{Symd}(\sigma_{B}) F ( i 1 ) ⊂ Symd ( σ B ) is a quadratic étale F F F -algebra. Since dim K K ( i 1 ) = 1 2 deg B \dim_{K}K(i_{1})=\frac{1}{2}\deg B dim K K ( i 1 ) = 2 1 deg B , Proposition 5.6 in [9 ] shows that K ( i 1 ) K(i_{1}) K ( i 1 ) is a neat subalgebra of ( B , σ B ) (B,\sigma_{B}) ( B , σ B ) , which means that K ( i 1 ) K(i_{1}) K ( i 1 ) is an étale K K K -algebra in Sym ( σ B ) \operatorname{Sym}(\sigma_{B}) Sym ( σ B ) over which B B B is a free left module. From [9 Prop. 6.5] it follows that
Symd ( σ B ) = K ( i 1 ) ⊕ V and Nrp σ B = N K ( i 1 ) / K ⊥ sq V . \operatorname{Symd}(\sigma_{B})=K(i_{1})\mathbin{\oplus}V
\qquad\text{and}\qquad
\operatorname{Nrp}_{\sigma_{B}}=N_{K(i_{1})/K}\mathbin{\perp}\operatorname{sq}_{V}. Symd ( σ B ) = K ( i 1 ) ⊕ V and Nrp σ B = N K ( i 1 ) / K ⊥ sq V .
The form N K ( i 1 ) / K N_{K(i_{1})/K} N K ( i 1 ) / K is defined over F F F since K ( i 1 ) ≃ K ⊗ F F ( i 1 ) K(i_{1})\simeq K\otimes_{F}F(i_{1}) K ( i 1 ) ≃ K ⊗ F F ( i 1 ) , hence ( t K ) ∗ ( N K ( i 1 ) / K ) (t_{K})_{*}(N_{K(i_{1})/K}) ( t K ) ∗ ( N K ( i 1 ) / K ) is hyperbolic and ( t K ) ∗ ( Nrp σ B ) (t_{K})_{*}(\operatorname{Nrp}_{\sigma_{B}}) ( t K ) ∗ ( Nrp σ B ) is Witt-equivalent to ( t K ) ∗ ( sq V ) (t_{K})_{*}(\operatorname{sq}_{V}) ( t K ) ∗ ( sq V ) . □
To proceed with the proof of Proposition 3.1, fix an element i 1 ∈ Symd ( σ B ) i_{1}\in\operatorname{Symd}(\sigma_{B}) i 1 ∈ Symd ( σ B ) as in Lemma 3.2. The F F F -algebra K 1 : = F ( i 1 ) K_{1}:=F(i_{1}) K 1 := F ( i 1 ) is quadratic étale and lies in Symd ( σ A ) \operatorname{Symd}(\sigma_{A}) Symd ( σ A ) since Symd ( σ B ) ⊂ Symd ( σ A ) \operatorname{Symd}(\sigma_{B})\subset\operatorname{Symd}(\sigma_{A}) Symd ( σ B ) ⊂ Symd ( σ A ) . It is neat in ( A , σ A ) (A,\sigma_{A}) ( A , σ A ) because K ( i 1 ) K(i_{1}) K ( i 1 ) is neat in ( B , σ B ) (B,\sigma_{B}) ( B , σ B ) , K ( i 1 ) K(i_{1}) K ( i 1 ) is a free left K 1 K_{1} K 1 -module and A A A is a free left B B B -module, hence A A A is a free left K 1 K_{1} K 1 -module. By [9 Th. 4.1 and 6.10] , there exists an étale biquadratic F F F -algebra L ⊂ Symd ( σ A ) L\subset\operatorname{Symd}(\sigma_{A}) L ⊂ Symd ( σ A ) containing K 1 K_{1} K 1 as the subalgebra fixed under an element α 1 \alpha_{1} α 1 of order 2 of its Galois group. Let α 2 \alpha_{2} α 2 and α 3 \alpha_{3} α 3 be the two other elements of order 2 of this Galois group, with fixed subfields respectively denoted by K 2 K_{2} K 2 and K 3 K_{3} K 3 . Consider
W i = { w ∈ Symd ( σ A ) ∣ w ℓ = α i ( ℓ ) w for all ℓ ∈ L } . W_{i}=\{w\in\operatorname{Symd}(\sigma_{A})\mid w\ell=\alpha_{i}(\ell)w\text{ for all }\ell\in L\}. W i = { w ∈ Symd ( σ A ) ∣ w ℓ = α i ( ℓ ) w for all ℓ ∈ L } .
Lemma 7.1 and Theorem 7.3 in [11 ] or Propositions 4 and 5 in [27 ] show that w i 2 ∈ K i w_{i}^{2}\in K_{i} w i 2 ∈ K i for i ∈ { 1 , 2 , 3 } i\in\{1,2,3\} i ∈ { 1 , 2 , 3 } and
Symd ( σ A ) = L ⊕ W 1 ⊕ W 2 ⊕ W 3 . (5) \operatorname{Symd}(\sigma_{A})=L\mathbin{\oplus}W_{1}\mathbin{\oplus}W_{2}\mathbin{\oplus}W_{3}. \tag*{(5)} Symd ( σ A ) = L ⊕ W 1 ⊕ W 2 ⊕ W 3 . (5)
Moreover, we have
L ⊕ W 1 = { w ∈ Symd ( σ A ) ∣ w i 1 = i 1 w } L\mathbin{\oplus}W_{1}=\{w\in\operatorname{Symd}(\sigma_{A})\mid wi_{1}=i_{1}w\} L ⊕ W 1 = { w ∈ Symd ( σ A ) ∣ w i 1 = i 1 w }
and
W 2 ⊕ W 3 = { w ∈ Symd ( σ A ) ∣ w i 1 = ( i 1 + 1 ) w } . W_{2}\mathbin{\oplus}W_{3}=\{w\in\operatorname{Symd}(\sigma_{A})\mid wi_{1}=(i_{1}+1)w\}. W 2 ⊕ W 3 = { w ∈ Symd ( σ A ) ∣ w i 1 = ( i 1 + 1 ) w } .
Writing x ∗ y = x y + y x x*y=xy+yx x ∗ y = x y + y x for x , y ∈ A x,y\in A x , y ∈ A , we also have
w i ∗ w j ∈ W k for w i ∈ W i , w j ∈ W j and { i , j , k } = { 1 , 2 , 3 } . w_{i}*w_{j}\in W_{k}\qquad\text{for }w_{i}\in W_{i},\ w_{j}\in W_{j}\text{ and }\{i,j,k\}=\{1,2,3\}. w i ∗ w j ∈ W k for w i ∈ W i , w j ∈ W j and { i , j , k } = { 1 , 2 , 3 } .
Lemma 3.3. Let r ∈ K r\in K r ∈ K be such that t K ( r ) = 1 t_{K}(r)=1 t K ( r ) = 1 . Then r = k 1 + w 1 r=k_{1}+w_{1} r = k 1 + w 1 for some k 1 ∈ K 1 k_{1}\in K_{1} k 1 ∈ K 1 and some nonzero w 1 ∈ W 1 w_{1}\in W_{1} w 1 ∈ W 1 . Moreover, V ⊂ W 2 ⊕ W 3 V\subset W_{2}\mathbin{\oplus}W_{3} V ⊂ W 2 ⊕ W 3 and for v = w 2 + w 3 ∈ V v=w_{2}+w_{3}\in V v = w 2 + w 3 ∈ V with w 2 ∈ W 2 w_{2}\in W_{2} w 2 ∈ W 2 and w 3 ∈ W 3 w_{3}\in W_{3} w 3 ∈ W 3 ,
w 1 ∗ w 2 = T K 1 / F ( k 1 ) w 3 , w 1 ∗ w 3 = T K 1 / F ( k 1 ) w 2 and w 2 ∗ w 3 = t K ( sq V ( v ) ) w 1 . (6) w_{1}*w_{2}=T_{K_{1}/F}(k_{1})w_{3},\qquad
w_{1}*w_{3}=T_{K_{1}/F}(k_{1})w_{2}
\qquad\text{and}\qquad
w_{2}*w_{3}=t_{K}(\operatorname{sq}_{V}(v))w_{1}. \tag*{(6)} w 1 ∗ w 2 = T K 1 / F ( k 1 ) w 3 , w 1 ∗ w 3 = T K 1 / F ( k 1 ) w 2 and w 2 ∗ w 3 = t K ( sq V ( v )) w 1 . (6)
Proof. As σ B \sigma_{B} σ B is symplectic, we have r ∈ Symd ( σ B ) ⊂ Symd ( σ A ) r\in\operatorname{Symd}(\sigma_{B})\subset\operatorname{Symd}(\sigma_{A}) r ∈ Symd ( σ B ) ⊂ Symd ( σ A ) , hence we may decompose r r r as a sum of terms from the direct sum (5). But r r r centralizes K 1 K_{1} K 1 since K 1 ⊂ B K_{1}\subset B K 1 ⊂ B , hence r ∈ L ⊕ W 1 r\in L\mathbin{\oplus}W_{1} r ∈ L ⊕ W 1 . Let r = k 1 + w 1 r=k_{1}+w_{1} r = k 1 + w 1 with k 1 ∈ L k_{1}\in L k 1 ∈ L and w 1 ∈ W 1 w_{1}\in W_{1} w 1 ∈ W 1 . If w 1 = 0 w_{1}=0 w 1 = 0 , then r r r lies in L L L . This is impossible since t K ( r ) = 1 t_{K}(r)=1 t K ( r ) = 1 implies r ∉ F r\notin F r ∈ / F , hence r r r is inseparable over F F F whereas L L L is separable over F F F . Therefore, w 1 ≠ 0 w_{1}\ne0 w 1 = 0 . As r 2 ∈ F × r^{2}\in F^{\times} r 2 ∈ F × , it follows that
k 1 2 + w 1 2 + k 1 ∗ w 1 ∈ F × . (7) k_{1}^{2}+w_{1}^{2}+k_{1}*w_{1}\in F^{\times}. \tag*{(7)} k 1 2 + w 1 2 + k 1 ∗ w 1 ∈ F × . (7)
We have k 1 2 , w 1 2 ∈ L k_1^2, w_1^2 \in L k 1 2 , w 1 2 ∈ L and k 1 ∗ w 1 ∈ W 1 k_1 * w_1 \in W_1 k 1 ∗ w 1 ∈ W 1 , hence (7) implies k 1 ∗ w 1 = 0 k_1 * w_1 = 0 k 1 ∗ w 1 = 0 . Since w 1 ≠ 0 w_1 \ne0 w 1 = 0 , the equations
0 = k 1 ∗ w 1 = k 1 w 1 + w 1 k 1 = ( k 1 + α 1 ( k 1 ) ) w 1 0 = k_1 * w_1 = k_1w_1 + w_1k_1 = (k_1 + \alpha_1(k_1))w_1 0 = k 1 ∗ w 1 = k 1 w 1 + w 1 k 1 = ( k 1 + α 1 ( k 1 )) w 1
show that k 1 ∈ K 1 k_1 \in K_1 k 1 ∈ K 1 .
By definition, V ⊂ Symd ( σ B ) ⊂ Symd ( σ A ) V \subset\operatorname{Symd}(\sigma_B) \subset\operatorname{Symd}(\sigma_A) V ⊂ Symd ( σ B ) ⊂ Symd ( σ A ) , and v x = α 1 ( x ) v vx = \alpha_1(x)v v x = α 1 ( x ) v for all x ∈ K 1 x \in K_1 x ∈ K 1 , hence V ⊂ W 2 ⊕ W 3 V \subset W_2 \oplus W_3 V ⊂ W 2 ⊕ W 3 . For v ∈ V v \in V v ∈ V , let v = w 2 + w 3 v = w_2 + w_3 v = w 2 + w 3 for some w 2 ∈ W 2 w_2 \in W_2 w 2 ∈ W 2 and w 3 ∈ W 3 w_3 \in W_3 w 3 ∈ W 3 . Since v v v commutes with r = k 1 + w 1 r = k_1 + w_1 r = k 1 + w 1 , it follows that
k 1 ∗ v + w 1 ∗ w 2 + w 1 ∗ w 3 = 0. (8) k_1 * v + w_1 * w_2 + w_1 * w_3 = 0. \tag*{(8)} k 1 ∗ v + w 1 ∗ w 2 + w 1 ∗ w 3 = 0. (8)
Note that k 1 ∗ v = k 1 v + v k 1 = T K 1 / F ( k 1 ) v = T K 1 / F ( k 1 ) w 2 + T K 1 / F ( k 1 ) w 3 k_1 * v = k_1v + vk_1 = T_{K_1/F}(k_1)v = T_{K_1/F}(k_1)w_2 + T_{K_1/F}(k_1)w_3 k 1 ∗ v = k 1 v + v k 1 = T K 1 / F ( k 1 ) v = T K 1 / F ( k 1 ) w 2 + T K 1 / F ( k 1 ) w 3 . Substituting in (8) we obtain
( T K 1 / F ( k 1 ) w 2 + w 1 ∗ w 3 ) + ( T K 1 / F ( k 1 ) w 3 + w 1 ∗ w 2 ) = 0. \left(T_{K_1/F}(k_1)w_2 + w_1 * w_3\right) + \left(T_{K_1/F}(k_1)w_3 + w_1 * w_2\right) = 0. ( T K 1 / F ( k 1 ) w 2 + w 1 ∗ w 3 ) + ( T K 1 / F ( k 1 ) w 3 + w 1 ∗ w 2 ) = 0.
But w 1 ∗ w 3 ∈ W 2 w_1 * w_3 \in W_2 w 1 ∗ w 3 ∈ W 2 and w 1 ∗ w 2 ∈ W 3 w_1 * w_2 \in W_3 w 1 ∗ w 2 ∈ W 3 , hence the first term on the left side lies in W 2 W_2 W 2 whereas the second term lies in W 3 W_3 W 3 . Each term must vanish since W 2 ⊕ W 3 W_2 \oplus W_3 W 2 ⊕ W 3 is a direct sum. This proves the first two equations in (6), and it only remains to compute w 2 ∗ w 3 w_2 * w_3 w 2 ∗ w 3 .
For this, observe that sq V ( v ) ∈ K \operatorname{sq}_V(v) \in K sq V ( v ) ∈ K , hence sq V ( v ) = λ 0 + λ 1 r \operatorname{sq}_V(v) = \lambda_0 + \lambda_1r sq V ( v ) = λ 0 + λ 1 r for some λ 0 , λ 1 ∈ F \lambda_0, \lambda_1 \in F λ 0 , λ 1 ∈ F . Taking images under t K t_K t K , we obtain λ 1 = t K ( sq V ( v ) ) \lambda_1 = t_K(\operatorname{sq}_V(v)) λ 1 = t K ( sq V ( v )) , hence
sq V ( v ) = ( λ 0 + λ 1 k 1 ) + t K ( sq V ( v ) ) w 1 with λ 0 + λ 1 k 1 ∈ L and t K ( sq V ( v ) ) w 1 ∈ W 1 . \operatorname{sq}_V(v) = (\lambda_0 + \lambda_1k_1) + t_K(\operatorname{sq}_V(v))w_1 \qquad\text{with } \lambda_0 + \lambda_1k_1 \in L \text{ and } t_K(\operatorname{sq}_V(v))w_1 \in W_1. sq V ( v ) = ( λ 0 + λ 1 k 1 ) + t K ( sq V ( v )) w 1 with λ 0 + λ 1 k 1 ∈ L and t K ( sq V ( v )) w 1 ∈ W 1 .
Comparing with
sq V ( v ) = ( w 2 + w 3 ) 2 = w 2 2 + w 3 2 + w 2 ∗ w 3 with w 2 2 , w 3 2 ∈ L and w 2 ∗ w 3 ∈ W 1 , \operatorname{sq}_V(v) = (w_2 + w_3)^2 = w_2^2 + w_3^2 + w_2 * w_3 \qquad\text{with } w_2^2, w_3^2 \in L \text{ and } w_2 * w_3 \in W_1, sq V ( v ) = ( w 2 + w 3 ) 2 = w 2 2 + w 3 2 + w 2 ∗ w 3 with w 2 2 , w 3 2 ∈ L and w 2 ∗ w 3 ∈ W 1 ,
we see that w 2 ∗ w 3 = t K ( sq V ( v ) ) w 1 w_2 * w_3 = t_K(\operatorname{sq}_V(v))w_1 w 2 ∗ w 3 = t K ( sq V ( v )) w 1 . This completes the proof of (6). □
Proof of Proposition 3.1, completed. We assume K K K is inseparable over F F F and we need to show that ( t K ) ∗ ( Nrp σ B ) (t_K)_*(\operatorname{Nrp}_{\sigma_B}) ( t K ) ∗ ( Nrp σ B ) is Witt-equivalent to a scalar multiple of π ( σ A ) \pi(\sigma_A) π ( σ A ) . Lemma 3.2 provides us with a quadratic space ( V , sq V ) (V,\operatorname{sq}_V) ( V , sq V ) over K K K such that ( t K ) ∗ ( Nrp σ B ) (t_K)_*(\operatorname{Nrp}_{\sigma_B}) ( t K ) ∗ ( Nrp σ B ) is Witt-equivalent to ( t K ) ∗ ( sq V ) (t_K)_*(\operatorname{sq}_V) ( t K ) ∗ ( sq V ) and with i 1 ∈ Symd ( σ B ) i_1 \in\operatorname{Symd}(\sigma_B) i 1 ∈ Symd ( σ B ) such that i 1 2 + i 1 ∈ F i_1^2 + i_1 \in F i 1 2 + i 1 ∈ F . Let K 1 = F ( i 1 ) K_1 = F(i_1) K 1 = F ( i 1 ) and choose a biquadratic étale F F F -algebra L ⊂ Symd ( σ A ) L \subset\operatorname{Symd}(\sigma_A) L ⊂ Symd ( σ A ) containing K 1 K_1 K 1 as the subalgebra fixed under an element α 1 \alpha_1 α 1 of order 2 in its Galois group. Using the same notation as in Lemma 3.3, we obtain a direct sum decomposition of Symd ( σ A ) \operatorname{Symd}(\sigma_A) Symd ( σ A ) as in (5). Fix r ∈ K r \in K r ∈ K such that t K ( r ) = 1 t_K(r) = 1 t K ( r ) = 1 and write r = k 1 + w 1 r = k_1 + w_1 r = k 1 + w 1 with k 1 ∈ K 1 k_1 \in K_1 k 1 ∈ K 1 and w 1 ∈ W 1 w_1 \in W_1 w 1 ∈ W 1 as in Lemma 3.3.
Suppose first k 1 ∉ F k_1 \notin F k 1 ∈ / F , hence T K 1 / F ( k 1 ) ≠ 0 T_{K_1/F}(k_1) \ne0 T K 1 / F ( k 1 ) = 0 . For v ∈ V v \in V v ∈ V , let v = w 2 + w 3 v = w_2 + w_3 v = w 2 + w 3 with w 2 ∈ W 2 w_2 \in W_2 w 2 ∈ W 2 and w 3 ∈ W 3 w_3 \in W_3 w 3 ∈ W 3 . The equations (6) yield w 2 ∗ w 3 = t K ( sq V ( v ) ) w 1 w_2 * w_3 = t_K(\operatorname{sq}_V(v))w_1 w 2 ∗ w 3 = t K ( sq V ( v )) w 1 and w 3 = T K 1 / F ( k 1 ) − 1 w 1 ∗ w 2 w_3 = T_{K_1/F}(k_1)^{-1}w_1 * w_2 w 3 = T K 1 / F ( k 1 ) − 1 w 1 ∗ w 2 . Eliminating w 3 w_3 w 3 from these equations yields
t K ( sq V ( v ) ) w 1 = T K 1 / F ( k 1 ) − 1 w 2 ∗ ( w 1 ∗ w 2 ) . (9) t_K(\operatorname{sq}_V(v))w_1 = T_{K_1/F}(k_1)^{-1}w_2 * (w_1 * w_2). \tag*{(9)} t K ( sq V ( v )) w 1 = T K 1 / F ( k 1 ) − 1 w 2 ∗ ( w 1 ∗ w 2 ) . (9)
Note that since w 2 2 ∈ K 2 w_2^2 \in K_2 w 2 2 ∈ K 2 ,
w 2 ∗ ( w 1 ∗ w 2 ) = w 2 w 1 w 2 + w 1 2 w 2 + w 1 w 2 2 + w 2 w 1 w 2 = ( w 2 2 + α 1 ( w 2 2 ) ) w 1 = T K 2 / F ( w 2 2 ) w 1 . w_2 * (w_1 * w_2) = w_2w_1w_2 + w_1^2w_2 + w_1w_2^2 + w_2w_1w_2 = (w_2^2 + \alpha_1(w_2^2))w_1 = T_{K_2/F}(w_2^2)w_1. w 2 ∗ ( w 1 ∗ w 2 ) = w 2 w 1 w 2 + w 1 2 w 2 + w 1 w 2 2 + w 2 w 1 w 2 = ( w 2 2 + α 1 ( w 2 2 )) w 1 = T K 2 / F ( w 2 2 ) w 1 .
Substituting in (9) and taking into account that w 1 w_1 w 1 is nonzero, we obtain
t K ( sq V ( v ) ) = T K 1 / F ( k 1 ) − 1 T K 2 / F ( w 2 2 ) = T K 1 / F ( k 1 ) − 1 T K 2 / F ( sq W 2 ( w 2 ) ) , t_K(\operatorname{sq}_V(v)) = T_{K_1/F}(k_1)^{-1}T_{K_2/F}(w_2^2) = T_{K_1/F}(k_1)^{-1}T_{K_2/F}(\operatorname{sq}_{W_2}(w_2)), t K ( sq V ( v )) = T K 1 / F ( k 1 ) − 1 T K 2 / F ( w 2 2 ) = T K 1 / F ( k 1 ) − 1 T K 2 / F ( sq W 2 ( w 2 )) ,
where sq W 2 : W 2 → K 2 \operatorname{sq}_{W_2}: W_2 \to K_2 sq W 2 : W 2 → K 2 is the squaring map. Therefore, the projection V → W 2 V \to W_2 V → W 2 parallel to W 3 W_3 W 3 , which maps v v v to w 2 w_2 w 2 , is a similitude from ( t K ) ∗ ( sq V ) (t_K)_*(\operatorname{sq}_V) ( t K ) ∗ ( sq V ) to ( T K 2 / F ) ∗ ( sq W 2 ) (T_{K_2/F})_*(\operatorname{sq}_{W_2}) ( T K 2 / F ) ∗ ( sq W 2 ) , with multiplier T K 1 / F ( k 1 ) − 1 T_{K_1/F}(k_1)^{-1} T K 1 / F ( k 1 ) − 1 . By definition, π ( σ A ) \pi(\sigma_A) π ( σ A ) is a scalar multiple of ( T K 2 / F ) ∗ ( sq W 2 ) (T_{K_2/F})_*(\operatorname{sq}_{W_2}) ( T K 2 / F ) ∗ ( sq W 2 ) . Therefore, π ( σ A ) \pi(\sigma_A) π ( σ A ) is isometric to a scalar multiple of ( t K ) ∗ ( s q V ) (t_K)_*(sq_V) ( t K ) ∗ ( s q V ) , which is Witt-equivalent to ( t K ) ∗ ( Nrp σ B ) (t_K)_*(\operatorname{Nrp}_{\sigma_B}) ( t K ) ∗ ( Nrp σ B ) , and the proof is complete when k 1 ∉ F k_1 \notin F k 1 ∈ / F .
For the rest of the proof, assume k 1 ∈ F k_1 \in F k 1 ∈ F . Then K = F ( r ) = F ( w 1 ) K = F(r) = F(w_1) K = F ( r ) = F ( w 1 ) , hence w 1 2 ∈ F × w_1^2 \in F^\times w 1 2 ∈ F × . Since conjugation by w 1 w_1 w 1 yields the nontrivial F F F -automorphism of K 2 K_2 K 2 , the F F F -algebra Q ⊂ A Q \subset A Q ⊂ A generated by K 2 K_2 K 2 and w 1 w_1 w 1 is a quaternion algebra stable under σ A \sigma_A σ A and containing K K K . The restriction of σ A \sigma_A σ A to Q Q Q is not the canonical symplectic involution on Q Q Q because K 2 ⊂ Symd ( σ A ) K_2 \subset\operatorname{Symd}(\sigma_A) K 2 ⊂ Symd ( σ A ) . Therefore, by [19 (2.23)] , σ A \sigma_A σ A restricts to a symplectic involution σ C \sigma_C σ C on the centralizer C C C of Q Q Q in A A A , which is a biquaternion algebra such that B ≃ C ⊗ F K B \simeq C \otimes_F K B ≃ C ⊗ F K since A = C ⊗ F Q A = C \otimes_F Q A = C ⊗ F Q . It follows that Nrp σ B = ( Nrp σ C ) K \operatorname{Nrp}_{\sigma_B} = (\operatorname{Nrp}_{\sigma_C})_K Nrp σ B = ( Nrp σ C ) K , hence ( t K ) ∗ ( Nrp σ B ) (t_K)_*(\operatorname{Nrp}_{\sigma_B}) ( t K ) ∗ ( Nrp σ B ) is hyperbolic. On the other hand, C C C decomposes into a tensor product of two quaternion F F F -algebras stable under σ C \sigma_C σ C by [19 (15.16)] , hence A A A decomposes into a tensor product of three quaternion F F F -algebras stable under σ A \sigma_A σ A . Theorem 7.11 of [11 ] or Proposition 8 of [27 ] then show that π ( σ A ) \pi(\sigma_A) π ( σ A ) is hyperbolic. Thus, ( t K ) ∗ ( Nrp σ B ) (t_K)_*(\operatorname{Nrp}_{\sigma_B}) ( t K ) ∗ ( Nrp σ B ) and π ( σ A ) \pi(\sigma_A) π ( σ A ) are Witt-equivalent. □
Proposition 3.1 shows that for a given symplectic involution σ B \sigma_B σ B on B B B , all the involutions σ A \sigma_A σ A on a central simple F F F -algebra A A A of degree 8 containing B B B that extend σ B \sigma_B σ B have the same discriminant Pfister form, and that form does not even depend on the choice of the algebra A A A but only on σ B \sigma_B σ B . We next describe how the form behaves under a change of involution on B B B .
Proposition 3.4. Let q B ∈ I q 2 ( K ) q_B \in I_q^2(K) q B ∈ I q 2 ( K ) be such that e 2 ( q B ) = [ B ] e_2(q_B) = [B] e 2 ( q B ) = [ B ] , and let t K : K → F t_K : K \to F t K : K → F be a nonzero F F F -linear functional such that t K ( 1 ) = 0 t_K(1) = 0 t K ( 1 ) = 0 . If σ A \sigma_A σ A and σ A ′ \sigma'_A σ A ′ are involutions on A A A that extend symplectic involutions σ B \sigma_B σ B , σ B ′ \sigma'_B σ B ′ on B B B , then in I q 3 ( F ) I_q^3(F) I q 3 ( F )
π ( σ A ′ ) ‾ ≡ π ( σ A ) ‾ ( m o d ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ) . \overline{\pi(\sigma'_A)} \equiv\overline{\pi(\sigma_A)} \pmod{(t_K)_*(\langle\langle K^\times\rangle\rangle q_B)}. π ( σ A ′ ) ≡ π ( σ A ) ( mod ( t K ) ∗ (⟨⟨ K × ⟩⟩ q B )) .
Proof. Let σ B ′ = Int ( u ) ∘ σ B \sigma'_B = \operatorname{Int}(u) \circ\sigma_B σ B ′ = Int ( u ) ∘ σ B for some unit u ∈ Symd ( σ B ) u \in\operatorname{Symd}(\sigma_B) u ∈ Symd ( σ B ) . By [19 (16.18)] we have
Nrp σ B − Nrp σ B ′ = ⟨ ⟨ Nrp σ B ( u ) ⟩ ⟩ Nrp σ B in I q 2 ( K ) , \operatorname{Nrp}_{\sigma_B} - \operatorname{Nrp}_{\sigma'_B} = \langle\langle\operatorname{Nrp}_{\sigma_B}(u)\rangle\rangle\operatorname{Nrp}_{\sigma_B} \quad\text{in } I_q^2(K), Nrp σ B − Nrp σ B ′ = ⟨⟨ Nrp σ B ( u )⟩⟩ Nrp σ B in I q 2 ( K ) ,
hence
( t K ) ∗ ( Nrp σ B ) − ( t K ) ∗ ( Nrp σ B ′ ) = ( t K ) ∗ ( ⟨ ⟨ Nrp σ B ( u ) ⟩ ⟩ Nrp σ B ) ∈ ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ Nrp σ B ) . (10) (t_K)_*(\operatorname{Nrp}_{\sigma_B}) - (t_K)_*(\operatorname{Nrp}_{\sigma'_B}) = (t_K)_*\bigl(\langle\langle\operatorname{Nrp}_{\sigma_B}(u)\rangle\rangle\operatorname{Nrp}_{\sigma_B}\bigr) \in(t_K)_*(\langle\langle K^\times\rangle\rangle\operatorname{Nrp}_{\sigma_B}). \tag*{(10)} ( t K ) ∗ ( Nrp σ B ) − ( t K ) ∗ ( Nrp σ B ′ ) = ( t K ) ∗ ( ⟨⟨ Nrp σ B ( u )⟩⟩ Nrp σ B ) ∈ ( t K ) ∗ (⟨⟨ K × ⟩⟩ Nrp σ B ) . (10)
By [19 (16.8)] , Nrp σ B \operatorname{Nrp}_{\sigma_B} Nrp σ B is an Albert form of B B B , hence Nrp σ B = q B \operatorname{Nrp}_{\sigma_B} = q_B Nrp σ B = q B and (10) readily yields
( t K ) ∗ ( Nrp σ B ) ‾ ≡ ( t K ) ∗ ( Nrp σ B ′ ) ‾ ( m o d ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ) . \overline{(t_K)_*(\operatorname{Nrp}_{\sigma_B})} \equiv\overline{(t_K)_*(\operatorname{Nrp}_{\sigma'_B})} \pmod{(t_K)_*(\langle\langle K^\times\rangle\rangle q_B)}. ( t K ) ∗ ( Nrp σ B ) ≡ ( t K ) ∗ ( Nrp σ B ′ ) ( mod ( t K ) ∗ (⟨⟨ K × ⟩⟩ q B )) .
This proves the proposition since ( t K ) ∗ ( Nrp σ B ) ‾ = π ( σ A ) ‾ \overline{(t_K)_*(\operatorname{Nrp}_{\sigma_B})} = \overline{\pi(\sigma_A)} ( t K ) ∗ ( Nrp σ B ) = π ( σ A ) and ( t K ) ∗ ( Nrp σ B ′ ) ‾ = π ( σ A ′ ) ‾ \overline{(t_K)_*(\operatorname{Nrp}_{\sigma'_B})} = \overline{\pi(\sigma'_A)} ( t K ) ∗ ( Nrp σ B ′ ) = π ( σ A ′ ) by Proposition 3.1. □
Remark 3.5. An alternative proof can be obtained from Theorem 1.2 instead of [19 (16.18)] .
We now define an invariant of the biquaternion algebra B B B , which also depends on the quadratic extension K / F K/F K / F .
Definition 3.6. Let B B B be a biquaternion algebra over a quadratic extension K K K of a field F F F , let q B ∈ I q 2 ( K ) q_B \in I_q^2(K) q B ∈ I q 2 ( K ) be such that e 2 ( q B ) = [ B ] e_2(q_B) = [B] e 2 ( q B ) = [ B ] and let s K : K → F s_K : K \to F s K : K → F be a nonzero F F F -linear functional. Assume N K / F [ B ] = 0 N_{K/F}[B] = 0 N K / F [ B ] = 0 . By Theorem 2.1 and [19 (4.14)] , there exists a central simple F F F -algebra A A A of degree 8 and an involution σ A \sigma_A σ A on A A A restricting to a symplectic involution on B B B . Taking s K = t K s_K = t_K s K = t K , with t K ( 1 ) = 0 t_K(1) = 0 t K ( 1 ) = 0 , Propositions 3.1 and 3.4 show that
π ( σ A ) ‾ + ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ∈ I q 3 ( F ) / ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) \overline{\pi(\sigma_A)} + (t_K)_*(\langle\langle K^\times\rangle\rangle q_B) \in I_q^3(F)/(t_K)_*(\langle\langle K^\times\rangle\rangle q_B) π ( σ A ) + ( t K ) ∗ (⟨⟨ K × ⟩⟩ q B ) ∈ I q 3 ( F ) / ( t K ) ∗ (⟨⟨ K × ⟩⟩ q B )
does not depend on the choices of A A A and σ A \sigma_A σ A . It coincides with the class
( t K ) ∗ ( Nrp σ B ) ‾ + ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ‾ ∈ I q 3 ( F ) ‾ / ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ‾ \overline{(t_K)_*(\operatorname{Nrp}_{\sigma_B})}+\overline{(t_K)_*(\langle\langle K^\times\rangle\rangle q_B)}\in\overline{I_q^3(F)}/\overline{(t_K)_*(\langle\langle K^\times\rangle\rangle q_B)} ( t K ) ∗ ( Nrp σ B ) + ( t K ) ∗ (⟨⟨ K × ⟩⟩ q B ) ∈ I q 3 ( F ) / ( t K ) ∗ (⟨⟨ K × ⟩⟩ q B )
for any symplectic involution σ B \sigma_B σ B of B B B . Moreover, let q A q_A q A be a quadratic form in I q 2 ( F ) I_q^2(F) I q 2 ( F ) such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] . We may take q B = ( q A ) K q_B=(q_A)_K q B = ( q A ) K and by [13 (20.3c) & (21.5)] , we have
( s K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ‾ = ( s K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ ) q A ‾ = ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ ) q A ‾ = ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ‾ ∈ I q 3 ( F ) ‾ . \overline{(s_K)_*(\langle\langle K^\times\rangle\rangle q_B)}=\overline{(s_K)_*(\langle\langle K^\times\rangle\rangle)q_A}=\overline{(t_K)_*(\langle\langle K^\times\rangle\rangle)q_A}=\overline{(t_K)_*(\langle\langle K^\times\rangle\rangle q_B)}\in\overline{I_q^3(F)}. ( s K ) ∗ (⟨⟨ K × ⟩⟩ q B ) = ( s K ) ∗ (⟨⟨ K × ⟩⟩) q A = ( t K ) ∗ (⟨⟨ K × ⟩⟩) q A = ( t K ) ∗ (⟨⟨ K × ⟩⟩ q B ) ∈ I q 3 ( F ) .
We therefore define
δ K / F ( B ) = π ( σ A ) ‾ + ( s K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ‾ ∈ I q 3 ( F ) ‾ / ( s K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ‾ . \delta_{K/F}(B)=\overline{\pi(\sigma_A)}+\overline{(s_K)_*(\langle\langle K^\times\rangle\rangle q_B)}\in\overline{I_q^3(F)}/\overline{(s_K)_*(\langle\langle K^\times\rangle\rangle q_B)}. δ K / F ( B ) = π ( σ A ) + ( s K ) ∗ (⟨⟨ K × ⟩⟩ q B ) ∈ I q 3 ( F ) / ( s K ) ∗ (⟨⟨ K × ⟩⟩ q B ) .
If φ B \varphi_B φ B is an Albert form of B B B , then φ B ≃ ⟨ λ ⟩ Nrp σ B \varphi_B\simeq\langle\lambda\rangle\operatorname{Nrp}_{\sigma_B} φ B ≃ ⟨ λ ⟩ Nrp σ B for some λ ∈ K × \lambda\in K^\times λ ∈ K × because all the Albert forms of B B B are similar by [19 (16.3) and (16.8)] . Hence,
( t K ) ∗ ( φ B ) ‾ = ( t K ) ∗ ( Nrp σ B ) ‾ + ( t K ) ∗ ( ⟨ ⟨ λ ⟩ ⟩ Nrp σ B ) ‾ , \overline{(t_K)_*(\varphi_B)}=\overline{(t_K)_*(\operatorname{Nrp}_{\sigma_B})}+\overline{(t_K)_*(\langle\langle\lambda\rangle\rangle\operatorname{Nrp}_{\sigma_B})}, ( t K ) ∗ ( φ B ) = ( t K ) ∗ ( Nrp σ B ) + ( t K ) ∗ (⟨⟨ λ ⟩⟩ Nrp σ B ) ,
so that ( t K ) ∗ ( φ B ) ‾ \overline{(t_K)_*(\varphi_B)} ( t K ) ∗ ( φ B ) and ( t K ) ∗ ( Nrp σ B ) ‾ \overline{(t_K)_*(\operatorname{Nrp}_{\sigma_B})} ( t K ) ∗ ( Nrp σ B ) have the same image in I q 3 ( F ) ‾ / ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ‾ \overline{I_q^3(F)}/\overline{(t_K)_*(\langle\langle K^\times\rangle\rangle q_B)} I q 3 ( F ) / ( t K ) ∗ (⟨⟨ K × ⟩⟩ q B ) . It follows that the definition of δ K / F ( B ) \delta_{K/F}(B) δ K / F ( B ) can also be stated as
δ K / F ( B ) = ( t K ) ∗ ( φ B ) ‾ + ( t K ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) ‾ \delta_{K/F}(B)=\overline{(t_K)_*(\varphi_B)}+\overline{(t_K)_*(\langle\langle K^\times\rangle\rangle q_B)} δ K / F ( B ) = ( t K ) ∗ ( φ B ) + ( t K ) ∗ (⟨⟨ K × ⟩⟩ q B )
for any Albert form φ B \varphi_B φ B of $B.
Thus, it is clear that the invariant δ K / F ( B ) \delta_{K/F}(B) δ K / F ( B ) is the same as the one defined in cohomological terms by Barry–Chapman–Laghribi in [8 §3] for the case where K K K is separable over F F F and by Laghribi–Lorenz in [20 §6] for the case where K K K is inseparable.
Vanishing of δ K / F ( B ) \delta_{K/F}(B) δ K / F ( B ) characterizes biquaternion K K K -algebras that descend to biquaternion algebras over F F F :
Proposition 3.7. With the notation of Definition 3.6, there exists a biquaternion F F F -algebra B 0 B_0 B 0 such that B ≃ B 0 ⊗ F K B\simeq B_0\otimes_F K B ≃ B 0 ⊗ F K if and only if δ K / F ( B ) = 0 \delta_{K/F}(B)=0 δ K / F ( B ) = 0 .
Proof. We sketch the proof given in [8 Th. 3.6] or [20 Prop. 6.4] for the reader’s convenience. If B ≃ B 0 ⊗ F K B\simeq B_0\otimes_F K B ≃ B 0 ⊗ F K , then we may choose φ B = ( φ B 0 ) K \varphi_B=(\varphi_{B_0})_K φ B = ( φ B 0 ) K with φ B 0 \varphi_{B_0} φ B 0 an Albert form of B 0 B_0 B 0 , and then ( t K ) ∗ ( φ B ) ‾ \overline{(t_K)_*(\varphi_B)} ( t K ) ∗ ( φ B ) is hyperbolic. Conversely, if φ B \varphi_B φ B is an Albert form of B B B and ( t K ) ∗ ( φ B ) ‾ = ( t K ) ∗ ( ⟨ ⟨ λ ⟩ ⟩ φ B ) ‾ \overline{(t_K)_*(\varphi_B)}=\overline{(t_K)_*(\langle\langle\lambda\rangle\rangle\varphi_B)} ( t K ) ∗ ( φ B ) = ( t K ) ∗ (⟨⟨ λ ⟩⟩ φ B ) for some λ ∈ K × \lambda\in K^\times λ ∈ K × , then ( t K ) ∗ ( ⟨ ⟨ λ ⟩ ⟩ φ B ) ‾ = 0 \overline{(t_K)_*(\langle\langle\lambda\rangle\rangle\varphi_B)}=0 ( t K ) ∗ (⟨⟨ λ ⟩⟩ φ B ) = 0 and we may find a quadratic form ψ ∈ I q 2 ( F ) \psi\in I_q^2(F) ψ ∈ I q 2 ( F ) of dimension 6 such that ⟨ λ ⟩ φ B ≃ ( ψ ) K \langle\lambda\rangle\varphi_B\simeq(\psi)_K ⟨ λ ⟩ φ B ≃ ( ψ ) K . Then ψ \psi ψ is an Albert form of a biquaternion F F F -algebra B 0 B_0 B 0 such that B ≃ B 0 ⊗ F K B\simeq B_0\otimes_F K B ≃ B 0 ⊗ F K . □
Biquadratic descent of quaternion algebras
Throughout this section, L L L is a biquadratic field extension of the field F F F of characteristic 2 and Q Q Q is a quaternion algebra over L L L . We assume that L L L is not purely inseparable2 over F F F and Q Q Q satisfies the norm conditions in Theorem 2.3 , so its Brauer class descends to F F F . More precisely, there exists a central simple F F F -algebra A A A , of degree 8 and exponent 2, containing L L L , and such that Q Q Q is isomorphic to the centralizer of L L L in A A A . Our goal is to define an invariant δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) whose vanishing characterizes quaternion algebras of the form Q 0 ⊗ F L Q_0\otimes_F L Q 0 ⊗ F L for some quaternion F F F -algebra Q 0 Q_0 Q 0 . As in the definitions of the Δ \Delta Δ -invariant in Section 1 and of the δ K / F \delta_{K/F} δ K / F -invariant in Section 3, we use discriminant Pfister forms of symplectic involutions on algebras of degree 8. This invariant is closely related to the one defined by Laghribi–Lorenz in [20 ] , as we also show. Its values lie in the quotient I q 3 ( F ) ‾ / ( s L ) ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) ‾ \overline{I_q^3(F)}/\overline{(s_L)_*(\langle\langle L^\times\rangle\rangle\operatorname{Nrd}_Q)} I q 3 ( F ) / ( s L ) ∗ (⟨⟨ L × ⟩⟩ Nrd Q ) , where Nrd Q \operatorname{Nrd}_Q Nrd Q is the reduced norm quadratic form and s L : L → F s_L:L\to F s L : L → F is any nonzero F F F -linear functional. Using a quadratic form q A ∈ I q 2 ( F ) q_A\in I_q^2(F) q A ∈ I q 2 ( F ) with e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] , the same kind of computations as in Definition 3.6 and [13 Ex. 21.6] show that this quotient does not depend on the choice of s L s_L s L .
The definition of δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) follows the same lines as for the invariant of biquaternion algebras in § 2 . By [19 (4.14)] , the canonical (symplectic) involution γ \gamma γ on Q Q Q extends to an involution σ A \sigma_A σ A on A A A , which is necessarily symplectic. We show in Proposition 4.3 below that the discriminant Pfister form π ( σ A ) \pi(\sigma_A) π ( σ A ) induces a class in I q 3 ( F ) / ( s L ) ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) I_q^3(F)/(s_L)_*(\langle\!\langle L^\times\rangle\!\rangle\operatorname{Nrd}_Q) I q 3 ( F ) / ( s L ) ∗ (⟨ ⟨ L × ⟩ ⟩ Nrd Q ) which does not depend on the choices of the algebra A A A and the involution σ A \sigma_A σ A . We define this class to be the invariant δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) , see Definition 4.4 below.
We first consider a change of algebra A A A containing Q Q Q :
Lemma 4.1. Let A A A and A ′ A' A ′ be central simple F F F -algebras of degree 8 8 8 containing Q Q Q such that 2 [ A ] = 2 [ A ′ ] = 0 2[A]=2[A']=0 2 [ A ] = 2 [ A ′ ] = 0 , and let σ A \sigma_A σ A be an involution on A A A extending the canonical involution γ \gamma γ on Q Q Q . There exists an involution σ A ′ \sigma_{A'} σ A ′ on A ′ A' A ′ extending γ \gamma γ on Q Q Q such that π ( σ A ′ ) ≃ π ( σ A ) \pi(\sigma_{A'})\simeq\pi(\sigma_A) π ( σ A ′ ) ≃ π ( σ A ) .
Proof. Let K 1 K_1 K 1 and K 2 K_2 K 2 be quadratic extensions of F F F such that L ≃ K 1 ⊗ F K 2 L\simeq K_1\otimes_F K_2 L ≃ K 1 ⊗ F K 2 . The Brauer class [ A ′ ] − [ A ] ∈ Br ( F ) [A']-[A]\in\operatorname{Br}(F) [ A ′ ] − [ A ] ∈ Br ( F ) is split by L L L since [ A ] L = [ Q ] = [ A ′ ] L [A]_L=[Q]=[A']_L [ A ] L = [ Q ] = [ A ′ ] L . Therefore, [ A ′ ] − [ A ] [A']-[A] [ A ′ ] − [ A ] is the Brauer class of a biquaternion F F F -algebra B B B . Let q B q_B q B be an Albert form of B B B . Since B B B is split by L L L , the form ( q B ) L (q_B)_L ( q B ) L is hyperbolic. Therefore, by [1 Th. 2.1] if one of K 1 , K 2 K_1,K_2 K 1 , K 2 is inseparable over F F F , or by [1 Th. 2.2] if K 1 K_1 K 1 and K 2 K_2 K 2 are separable over F F F , the form q B q_B q B is Witt-equivalent to an orthogonal sum φ 1 ⊥ φ 2 \varphi_1\perp\varphi_2 φ 1 ⊥ φ 2 , where φ i \varphi_i φ i is split by K i K_i K i for i = 1 , 2 i=1,2 i = 1 , 2 . Comparing Arf invariants shows that φ 1 , φ 2 ∈ I q 2 ( F ) \varphi_1,\varphi_2\in I_q^2(F) φ 1 , φ 2 ∈ I q 2 ( F ) . Letting e 2 ( φ i ) = [ Q i ] e_2(\varphi_i)=[Q_i] e 2 ( φ i ) = [ Q i ] we have
[ A ′ ] = [ A ] + [ Q 1 ] + [ Q 2 ] [A']=[A]+[Q_1]+[Q_2] [ A ′ ] = [ A ] + [ Q 1 ] + [ Q 2 ]
with Q i Q_i Q i a quaternion algebra split by K i K_i K i for i = 1 , 2 i=1,2 i = 1 , 2 .
Let B 1 ⊂ A B_1\subset A B 1 ⊂ A be the centralizer of K 1 K_1 K 1 ; thus B 1 ⊃ Q B_1\supset Q B 1 ⊃ Q and, since Q 1 Q_1 Q 1 is split by K 1 K_1 K 1 ,
[ B 1 ] = [ A ] K 1 = [ A ⊗ F Q 1 ] K 1 . [B_1]=[A]_{K_1}=[A\otimes_F Q_1]_{K_1}. [ B 1 ] = [ A ] K 1 = [ A ⊗ F Q 1 ] K 1 .
Every field extension of K 1 K_1 K 1 that splits A A A also splits A ⊗ F Q 1 A\otimes_F Q_1 A ⊗ F Q 1 , hence A ⊗ F Q 1 A\otimes_F Q_1 A ⊗ F Q 1 is Brauer-equivalent to a central simple F F F -algebra A 1 A_1 A 1 of degree 8 8 8 , and we may assume A 1 A_1 A 1 contains B 1 B_1 B 1 since [ A 1 ] K 1 = [ B 1 ] [A_1]_{K_1}=[B_1] [ A 1 ] K 1 = [ B 1 ] . Let σ B 1 \sigma_{B_1} σ B 1 be the restriction of σ A \sigma_A σ A to B 1 B_1 B 1 . By [19 (4.14)] there exists an involution σ A 1 \sigma_{A_1} σ A 1 on A 1 A_1 A 1 extending σ B 1 \sigma_{B_1} σ B 1 . Let s 1 : K 1 → F s_1:K_1\to F s 1 : K 1 → F be a nonzero F F F -linear functional such that s 1 ( 1 ) = 0 s_1(1)=0 s 1 ( 1 ) = 0 . Proposition 3.1 shows that π ( σ A ) \pi(\sigma_A) π ( σ A ) and π ( σ A 1 ) \pi(\sigma_{A_1}) π ( σ A 1 ) are both Witt-equivalent to a scalar multiple of ( s 1 ) ∗ ( Nrp σ B 1 ) (s_1)_*(\operatorname{Nrp}_{\sigma_{B_1}}) ( s 1 ) ∗ ( Nrp σ B 1 ) , hence π ( σ A ) ≃ π ( σ A 1 ) \pi(\sigma_A)\simeq\pi(\sigma_{A_1}) π ( σ A ) ≃ π ( σ A 1 ) .
Now, [ A ′ ] = [ A 1 ] + [ Q 2 ] [A']=[A_1]+[Q_2] [ A ′ ] = [ A 1 ] + [ Q 2 ] , hence [ A ′ ] K 2 = [ A 1 ] K 2 [A']_{K_2}=[A_1]_{K_2} [ A ′ ] K 2 = [ A 1 ] K 2 since K 2 K_2 K 2 splits Q 2 Q_2 Q 2 . Therefore, up to isomorphism we may assume that the centralizer B 2 B_2 B 2 of K 2 K_2 K 2 in A 1 A_1 A 1 is contained in A ′ A' A ′ . Let σ B 2 \sigma_{B_2} σ B 2 be the restriction of σ A 1 \sigma_{A_1} σ A 1 to B 2 B_2 B 2 . By [19 (4.14)] again, there exists an involution σ A ′ \sigma_{A'} σ A ′ on A ′ A' A ′ extending σ B 2 \sigma_{B_2} σ B 2 , and the same argument as for σ A 1 \sigma_{A_1} σ A 1 above using Proposition 3.1 shows that π ( σ A 1 ) ≃ π ( σ A ′ ) \pi(\sigma_{A_1})\simeq\pi(\sigma_{A'}) π ( σ A 1 ) ≃ π ( σ A ′ ) , hence also π ( σ A ) ≃ π ( σ A ′ ) \pi(\sigma_A)\simeq\pi(\sigma_{A'}) π ( σ A ) ≃ π ( σ A ′ ) . Note that the involutions σ A \sigma_A σ A , σ B 1 \sigma_{B_1} σ B 1 , σ A 1 \sigma_{A_1} σ A 1 , σ B 2 \sigma_{B_2} σ B 2 and σ A ′ \sigma_{A'} σ A ′ all coincide on Q Q Q , and therefore extend its canonical involution γ \gamma γ . Thus, σ A ′ \sigma_{A'} σ A ′ meets the requirements of the lemma. □
We also require the following technical result:
Lemma 4.2. Let A A A be a central simple F F F -algebra of degree 8 8 8 containing Q Q Q and σ A \sigma_A σ A be an involution on A A A extending γ \gamma γ on Q Q Q . Let also q A ∈ I q 2 ( F ) q_A\in I_q^2(F) q A ∈ I q 2 ( F ) be such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] , and let s L : L → F s_L:L\to F s L : L → F be a nonzero F F F -linear functional. For every ℓ ∈ L × \ell\in L^\times ℓ ∈ L × ,
N L / F ( ℓ ) = Nrp σ A ( ℓ ) and ( s L ) ∗ ( ⟨ ⟨ ℓ ⟩ ⟩ Nrd Q ) = ⟨ ⟨ N L / F ( ℓ ) ⟩ ⟩ q A in I q 3 ( F ) . (11) N_{L/F}(\ell)=\operatorname{Nrp}_{\sigma_A}(\ell)
\quad\text{and}\quad
(s_L)_*(\langle\!\langle\ell\rangle\!\rangle\operatorname{Nrd}_Q)=\langle\!\langle N_{L/F}(\ell)\rangle\!\rangle q_A
\quad\text{in }I_q^3(F). \tag*{(11)} N L / F ( ℓ ) = Nrp σ A ( ℓ ) and ( s L ) ∗ (⟨ ⟨ ℓ ⟩ ⟩ Nrd Q ) = ⟨ ⟨ N L / F ( ℓ )⟩ ⟩ q A in I q 3 ( F ) . (11)
Proof. First, observe that L = Symd ( γ ) ⊂ Symd ( σ A ) L=\operatorname{Symd}(\gamma)\subset\operatorname{Symd}(\sigma_A) L = Symd ( γ ) ⊂ Symd ( σ A ) , hence Nrp σ A ( ℓ ) \operatorname{Nrp}_{\sigma_A}(\ell) Nrp σ A ( ℓ ) is defined for all ℓ ∈ L \ell\in L ℓ ∈ L . Since L L L is a subfield of degree 4 in A A A of degree 8, we have Nrd A ( ℓ ) = N L / F ( ℓ ) 2 \operatorname{Nrd}_A(\ell)=N_{L/F}(\ell)^2 Nrd A ( ℓ ) = N L / F ( ℓ ) 2 , hence Nrp σ A ( ℓ ) = N L / F ( ℓ ) \operatorname{Nrp}_{\sigma_A}(\ell)=N_{L/F}(\ell) Nrp σ A ( ℓ ) = N L / F ( ℓ ) since Nrd A ( ℓ ) = Nrp σ A ( ℓ ) 2 \operatorname{Nrd}_A(\ell)=\operatorname{Nrp}_{\sigma_A}(\ell)^2 Nrd A ( ℓ ) = Nrp σ A ( ℓ ) 2 .
Next, note that e 2 ( Nrd Q ) = [ Q ] = [ A L ] e_2(\operatorname{Nrd}_Q)=[Q]=[A_L] e 2 ( Nrd Q ) = [ Q ] = [ A L ] , hence Nrd ‾ Q = ( q A ‾ ) L \overline{\operatorname{Nrd}}_Q=(\overline{q_A})_L Nrd Q = ( q A ) L in I q 2 ( L ) I_q^2(L) I q 2 ( L ) . By Frobenius reciprocity (see [13 Prop. 20.2] ), it follows that ( s L ) ∗ ( ⟨ ⟨ ℓ ⟩ ⟩ Nrd Q ) = ( s L ) ∗ ( ⟨ ⟨ ℓ ⟩ ⟩ ) q A (s_L)_*(\langle\langle\ell\rangle\rangle\operatorname{Nrd}_Q)=(s_L)_*(\langle\langle\ell\rangle\rangle)q_A ( s L ) ∗ (⟨⟨ ℓ ⟩⟩ Nrd Q ) = ( s L ) ∗ (⟨⟨ ℓ ⟩⟩) q A . If L = F ( ℓ ) L=F(\ell) L = F ( ℓ ) , then [13 Cor. 20.13 and 21.5] show that ( s L ) ∗ ( ⟨ ⟨ ℓ ⟩ ⟩ ) = ⟨ ⟨ N L / F ( ℓ ) ⟩ ⟩ (s_L)_*(\langle\langle\ell\rangle\rangle)=\langle\langle N_{L/F}(\ell)\rangle\rangle ( s L ) ∗ (⟨⟨ ℓ ⟩⟩) = ⟨⟨ N L / F ( ℓ )⟩⟩ in I ‾ ( F ) \overline{I}(F) I ( F ) , and (11) is proven. If F ( ℓ ) ⊊ L F(\ell)\subsetneq L F ( ℓ ) ⊊ L , then N L / F ( ℓ ) ∈ F × 2 N_{L/F}(\ell)\in F^{\times2} N L / F ( ℓ ) ∈ F × 2 , hence ⟨ ⟨ N L / F ( ℓ ) ⟩ ⟩ = 0 \langle\langle N_{L/F}(\ell)\rangle\rangle=0 ⟨⟨ N L / F ( ℓ )⟩⟩ = 0 . On the other hand, taking s L = s 1 ∘ s 2 s_L=s_1\circ s_2 s L = s 1 ∘ s 2 where s 2 : L → F ( ℓ ) s_2:L\to F(\ell) s 2 : L → F ( ℓ ) is a nonzero F ( ℓ ) F(\ell) F ( ℓ ) -linear functional that vanishes on F ( ℓ ) F(\ell) F ( ℓ ) and s 1 : F ( ℓ ) → F s_1:F(\ell)\to F s 1 : F ( ℓ ) → F is a nonzero F F F -linear functional, we obtain ( s L ) ∗ ( ⟨ ⟨ ℓ ⟩ ⟩ ) = 0 (s_L)_*(\langle\langle\ell\rangle\rangle)=0 ( s L ) ∗ (⟨⟨ ℓ ⟩⟩) = 0 . Therefore, (11) holds in all cases. □
Proposition 4.3. Let A A A and A ′ A' A ′ be central simple F F F -algebras of degree 8 containing Q Q Q and σ A \sigma_A σ A , σ A ′ \sigma_{A'} σ A ′ be involutions on A A A and A ′ A' A ′ respectively extending the canonical involution γ \gamma γ on Q Q Q . Then, for any nonzero F F F -linear functional s L : L → F s_L:L\to F s L : L → F ,
π ( σ A ) ‾ ≡ π ( σ A ′ ) ‾ ( m o d ( s L ) ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) ) in I ‾ q 3 ( F ) . \overline{\pi(\sigma_A)}\equiv\overline{\pi(\sigma_{A'})}\pmod{(s_L)_*(\langle\langle L^\times\rangle\rangle\operatorname{Nrd}_Q)}\quad\text{in } \overline{I}_q^3(F). π ( σ A ) ≡ π ( σ A ′ ) ( mod ( s L ) ∗ (⟨⟨ L × ⟩⟩ Nrd Q )) in I q 3 ( F ) .
Proof. Lemma 4.1 yields an involution σ A ′ ′ \sigma'_{A'} σ A ′ ′ on A ′ A' A ′ extending γ \gamma γ such that π ( σ A ′ ′ ) ≃ π ( σ A ) \pi(\sigma'_{A'})\simeq\pi(\sigma_A) π ( σ A ′ ′ ) ≃ π ( σ A ) . Let σ A ′ = Int ( ℓ ) ∘ σ A ′ ′ \sigma_{A'}=\operatorname{Int}(\ell)\circ\sigma'_{A'} σ A ′ = Int ( ℓ ) ∘ σ A ′ ′ for some ℓ ∈ Symd ( σ A ′ ′ ) \ell\in\operatorname{Symd}(\sigma'_{A'}) ℓ ∈ Symd ( σ A ′ ′ ) . Since the restrictions of σ A ′ ′ \sigma'_{A'} σ A ′ ′ and σ A ′ \sigma_{A'} σ A ′ to Q Q Q are the same, ℓ \ell ℓ centralizes Q Q Q hence ℓ ∈ L × \ell\in L^\times ℓ ∈ L × . Theorem 1.2 yields
π ( σ A ′ ) ‾ = ⟨ ⟨ Nrp σ A ′ ′ ( ℓ ) ⟩ ⟩ q A ′ + π ( σ A ′ ′ ) ‾ in I ‾ q 3 ( F ) , \overline{\pi(\sigma_{A'})}=\langle\langle\operatorname{Nrp}_{\sigma'_{A'}}(\ell)\rangle\rangle q_{A'}+\overline{\pi(\sigma'_{A'})}\quad\text{in }\overline{I}_q^3(F), π ( σ A ′ ) = ⟨⟨ Nrp σ A ′ ′ ( ℓ )⟩⟩ q A ′ + π ( σ A ′ ′ ) in I q 3 ( F ) ,
where q A ′ ∈ I q 2 ( F ) q_{A'}\in I_q^2(F) q A ′ ∈ I q 2 ( F ) is any quadratic form such that e 2 ( q A ′ ) = [ A ′ ] e_2(q_{A'})=[A'] e 2 ( q A ′ ) = [ A ′ ] and Lemma 4.2 shows that
⟨ ⟨ Nrp σ A ′ ( ℓ ) ⟩ ⟩ q A ′ = ( s L ) ∗ ( ⟨ ⟨ ℓ ⟩ ⟩ Nrd Q ) in I ‾ q 3 ( F ) . \langle\langle\operatorname{Nrp}_{\sigma_{A'}}(\ell)\rangle\rangle q_{A'}=(s_L)_*(\langle\langle\ell\rangle\rangle\operatorname{Nrd}_Q)\quad\text{in }\overline{I}_q^3(F). ⟨⟨ Nrp σ A ′ ( ℓ )⟩⟩ q A ′ = ( s L ) ∗ (⟨⟨ ℓ ⟩⟩ Nrd Q ) in I q 3 ( F ) .
Therefore, π ( σ A ) ‾ = π ( σ A ′ ′ ) ≡ π ( σ A ′ ) ‾ ( m o d ( s L ) ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) ) \overline{\pi(\sigma_A)}=\pi(\sigma'_{A'})\equiv\overline{\pi(\sigma_{A'})}\pmod{(s_L)_*(\langle\langle L^\times\rangle\rangle\operatorname{Nrd}_Q)} π ( σ A ) = π ( σ A ′ ′ ) ≡ π ( σ A ′ ) ( mod ( s L ) ∗ (⟨⟨ L × ⟩⟩ Nrd Q )) . □
Proposition 4.3 allows us to set the following definition:
Definition 4.4. Let Q Q Q be a quaternion algebra over a biquadratic field extension L L L of F F F . Assume L L L is not purely inseparable over F F F and N L / K [ Q ] = 0 N_{L/K}[Q]=0 N L / K [ Q ] = 0 for all quadratic extensions K K K of F F F in L L L , so that there exist central simple F F F -algebras A A A of degree 8 containing Q Q Q such that 2 [ A ] = 0 2[A]=0 2 [ A ] = 0 by Theorem 2.3. An invariant of Q Q Q is defined by
δ L / F ( Q ) = π ( σ A ) ‾ + ( s L ) ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) ∈ I ‾ q 3 ( F ) / ( s L ) ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) \delta_{L/F}(Q)=\overline{\pi(\sigma_A)}+(s_L)_*(\langle\langle L^\times\rangle\rangle\operatorname{Nrd}_Q)\in\overline{I}_q^3(F)/(s_L)_*(\langle\langle L^\times\rangle\rangle\operatorname{Nrd}_Q) δ L / F ( Q ) = π ( σ A ) + ( s L ) ∗ (⟨⟨ L × ⟩⟩ Nrd Q ) ∈ I q 3 ( F ) / ( s L ) ∗ (⟨⟨ L × ⟩⟩ Nrd Q ) for any involution σ A \sigma_A σ A on A A A extending the canonical involution γ \gamma γ on Q Q Q and any nonzero F F F -linear functional s L : L → F s_L:L\to F s L : L → F .
Vanishing of the invariant δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) characterizes quaternion L L L -algebras that descend to F F F :
Theorem 4.5. For a quaternion algebra Q Q Q as in Definition 4.4, δ L / F ( Q ) = 0 \delta_{L/F}(Q)=0 δ L / F ( Q ) = 0 if and only if there exists a quaternion F F F -algebra Q 0 Q_0 Q 0 such that Q ≃ Q 0 ⊗ F L Q\simeq Q_0\otimes_F L Q ≃ Q 0 ⊗ F L .
Proof. Assume Q = Q 0 ⊗ F L Q=Q_0\otimes_F L Q = Q 0 ⊗ F L for some quaternion F F F -algebra Q 0 Q_0 Q 0 , and let L ≃ K 1 ⊗ F K 2 L\simeq K_1\otimes_F K_2 L ≃ K 1 ⊗ F K 2 for some quadratic extensions K 1 K_1 K 1 , K 2 K_2 K 2 of F F F . Each End F ( K i ) \operatorname{End}_F(K_i) End F ( K i ) is a split quaternion F F F -algebra, and a canonical embedding K i ↪ End F ( K i ) K_i\hookrightarrow\operatorname{End}_F(K_i) K i ↪ End F ( K i ) is given by left multiplication. We may therefore take A = Q 0 ⊗ F End F ( K 1 ) ⊗ F End F ( K 2 ) A=Q_0\otimes_F\operatorname{End}_F(K_1)\otimes_F\operatorname{End}_F(K_2) A = Q 0 ⊗ F End F ( K 1 ) ⊗ F End F ( K 2 ) in Definition 4.4, since we can consider Q Q Q as a subalgebra of A A A . For i = 1 , 2 i=1,2 i = 1 , 2 , let θ i \theta_i θ i be an involution on End F ( K i ) \operatorname{End}_F(K_i) End F ( K i ) leaving K i K_i K i fixed. The involution σ A = γ 0 ⊗ θ 1 ⊗ θ 2 \sigma_A=\gamma_0\otimes\theta_1\otimes\theta_2 σ A = γ 0 ⊗ θ 1 ⊗ θ 2 on A A A , where γ 0 \gamma_0 γ 0 is the canonical involution on Q 0 Q_0 Q 0 , extends the canonical
involution γ \gamma γ on Q Q Q . Since σ A \sigma_A σ A is decomposable, π ( σ A ) \pi(\sigma_A) π ( σ A ) is hyperbolic by [11 Th. 7.11] or [27 Prop. 8] , so δ L / F ( Q ) = 0 \delta_{L/F}(Q)=0 δ L / F ( Q ) = 0 as required.
For the converse, assume δ L / F ( Q ) = 0 \delta_{L/F}(Q)=0 δ L / F ( Q ) = 0 and consider an involution σ A \sigma_A σ A extending γ \gamma γ on a central simple F F F -algebra A A A of degree 8 containing Q Q Q . By definition, the equation δ L / F ( Q ) = 0 \delta_{L/F}(Q)=0 δ L / F ( Q ) = 0 means that
π ( σ A ) ‾ = ( s L ) ∗ ( ⟨ ⟨ ℓ ⟩ ⟩ Nrd Q ) ‾ ∈ I q 3 ( F ) ‾ \overline{\pi(\sigma_A)}=\overline{(s_L)_*(\langle\!\langle\ell\rangle\!\rangle\operatorname{Nrd}_Q)}\in\overline{I_q^3(F)} π ( σ A ) = ( s L ) ∗ (⟨ ⟨ ℓ ⟩ ⟩ Nrd Q ) ∈ I q 3 ( F )
for some ℓ ∈ L × \ell\in L^\times ℓ ∈ L × .
Lemma 4.2 allows us to reformulate this equation as
π ( σ A ) ‾ = ⟨ ⟨ Nrp σ A ( ℓ ) ⟩ ⟩ q A ∈ I q 3 ( F ) ‾ \overline{\pi(\sigma_A)}=\langle\!\langle\operatorname{Nrp}_{\sigma_A}(\ell)\rangle\!\rangle q_A\in\overline{I_q^3(F)} π ( σ A ) = ⟨ ⟨ Nrp σ A ( ℓ )⟩ ⟩ q A ∈ I q 3 ( F )
for any quadratic form q A ∈ I q 2 ( F ) q_A\in I_q^2(F) q A ∈ I q 2 ( F ) such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] . For σ A ′ = Int ( ℓ ) ∘ σ A \sigma'_A=\operatorname{Int}(\ell)\circ\sigma_A σ A ′ = Int ( ℓ ) ∘ σ A we then have
π ( σ A ′ ) ‾ = 0 in I q 3 ( F ) ‾ \overline{\pi(\sigma'_A)}=0\text{ in }\overline{I_q^3(F)} π ( σ A ′ ) = 0 in I q 3 ( F )
by Theorem 1.2. Since π ( σ A ′ ) \pi(\sigma'_A) π ( σ A ′ ) is a 3-fold Pfister form, this implies that π ( σ A ′ ) \pi(\sigma'_A) π ( σ A ′ ) is hyperbolic. Note that ℓ ∈ L × \ell\in L^\times ℓ ∈ L × commutes with Q Q Q , so that σ A ′ \sigma'_A σ A ′ also extends the canonical involution γ \gamma γ of Q Q Q .
If L L L is separable over F F F , it follows from [11 Prop. 7.6] that σ A ′ \sigma'_A σ A ′ decomposes “along L L L ”, which implies that A ≃ Q 0 ⊗ F B A\simeq Q_0\otimes_F B A ≃ Q 0 ⊗ F B for some quaternion F F F -algebra Q 0 Q_0 Q 0 and some biquaternion F F F -algebra B B B containing L L L . As Q Q Q is the centralizer of L L L in A A A , we then have Q ≃ Q 0 ⊗ F L Q\simeq Q_0\otimes_F L Q ≃ Q 0 ⊗ F L , and the proof is complete when L L L is separable over F F F .
For the rest of the proof, assume L = K 1 ⊗ F K 2 L=K_1\otimes_F K_2 L = K 1 ⊗ F K 2 where K 1 K_1 K 1 (resp. K 2 K_2 K 2 ) is a separable (resp. inseparable) quadratic extension of F F F . Let B 2 ⊂ A B_2\subset A B 2 ⊂ A be the centralizer of K 2 K_2 K 2 in A A A and σ B 2 = σ A ′ ∣ B 2 \sigma_{B_2}=\sigma'_A|_{B_2} σ B 2 = σ A ′ ∣ B 2 be the restriction of σ A ′ \sigma'_A σ A ′ to B 2 B_2 B 2 . Let also i 1 ∈ K 1 i_1\in K_1 i 1 ∈ K 1 be such that T K 1 / F ( i 1 ) = 1 T_{K_1/F}(i_1)=1 T K 1 / F ( i 1 ) = 1 , and consider
V = { v ∈ Symd ( σ B 2 ) ∣ v i 1 = ( i 1 + 1 ) v } . V=\{v\in\operatorname{Symd}(\sigma_{B_2})\mid vi_1=(i_1+1)v\}. V = { v ∈ Symd ( σ B 2 ) ∣ v i 1 = ( i 1 + 1 ) v } .
Lemma 3.2 shows that the squaring map sq V : v ↦ v 2 \operatorname{sq}_V:v\mapsto v^2 sq V : v ↦ v 2 is a nonsingular quadratic form V → K 2 V\to K_2 V → K 2 , whose transfer ( t K 2 ) ∗ ( sq V ) (t_{K_2})_*(\operatorname{sq}_V) ( t K 2 ) ∗ ( sq V ) for any F F F -linear functional t K 2 : K 2 → F t_{K_2}:K_2\to F t K 2 : K 2 → F such that t K 2 ( 1 ) = 0 t_{K_2}(1)=0 t K 2 ( 1 ) = 0 is Witt-equivalent to ( t K 2 ) ∗ ( Nrp σ B 2 ) (t_{K_2})_*(\operatorname{Nrp}_{\sigma_{B_2}}) ( t K 2 ) ∗ ( Nrp σ B 2 ) . It is therefore also Witt-equivalent to π ( σ A ′ ) \pi(\sigma'_A) π ( σ A ′ ) by Proposition 3.1, hence it is hyperbolic. Theorem 2.2 in [10 ] then shows that sq V ≃ φ K 2 \operatorname{sq}_V\simeq\varphi_{K_2} sq V ≃ φ K 2 for some nonsingular quadratic form φ ∈ I q ( F ) \varphi\in I_q(F) φ ∈ I q ( F ) . In particular, we may find v ∈ V v\in V v ∈ V such that v 2 ∈ F × v^2\in F^\times v 2 ∈ F × .
Then i 1 i_1 i 1 and v v v generate a quaternion F F F -algebra Q 1 ⊂ B 2 Q_1\subset B_2 Q 1 ⊂ B 2 containing K 1 K_1 K 1 , and B 2 ≃ Q 1 ⊗ F Q 2 B_2\simeq Q_1\otimes_F Q_2 B 2 ≃ Q 1 ⊗ F Q 2 for Q 2 Q_2 Q 2 the centralizer of Q 1 Q_1 Q 1 in B 2 B_2 B 2 . As [ B 2 ] = [ A ] K 2 [B_2]=[A]_{K_2} [ B 2 ] = [ A ] K 2 , Theorem 2.1 shows that N K 2 / F [ B 2 ] = 0 N_{K_2/F}[B_2]=0 N K 2 / F [ B 2 ] = 0 , hence N K 2 / F [ Q 2 ] = 0 N_{K_2/F}[Q_2]=0 N K 2 / F [ Q 2 ] = 0 since N K 2 / F [ ( Q 1 ) K 2 ] = 0 N_{K_2/F}[(Q_1)_{K_2}]=0 N K 2 / F [( Q 1 ) K 2 ] = 0 . Therefore, Proposition 2.2 yields a quaternion F F F -algebra Q 0 Q_0 Q 0 such that Q 2 ≃ Q 0 ⊗ F K 2 Q_2\simeq Q_0\otimes_F K_2 Q 2 ≃ Q 0 ⊗ F K 2 . Thus, B 2 ≃ Q 1 ⊗ F Q 0 ⊗ F K 2 B_2\simeq Q_1\otimes_F Q_0\otimes_F K_2 B 2 ≃ Q 1 ⊗ F Q 0 ⊗ F K 2 . The algebra Q Q Q is the centralizer of K 1 ⊗ F K 2 K_1\otimes_F K_2 K 1 ⊗ F K 2 in A A A , hence also in B 2 B_2 B 2 . Since K 1 ⊂ Q 1 K_1\subset Q_1 K 1 ⊂ Q 1 , it follows that Q ≃ Q 0 ⊗ F K 1 ⊗ F K 2 = Q 0 ⊗ F L Q\simeq Q_0\otimes_F K_1\otimes_F K_2=Q_0\otimes_F L Q ≃ Q 0 ⊗ F K 1 ⊗ F K 2 = Q 0 ⊗ F L .
We complete this section with an explicit computation of the invariant δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) , which shows that this invariant is the same as the one defined by Laghribi–Lorenz in [20 ] (even though they view it as an invariant of a central simple F F F -algebra A A A of degree 8 containing Q Q Q ).
Proposition 4.6. Let Q Q Q be a quaternion algebra over a biquadratic field extension L L L of F F F . Assume L L L is not purely inseparable over F F F and let L ≃ F ( ℘ − 1 ( α ) ) ⊗ F K L\simeq F(\wp^{-1}(\alpha))\otimes_F K L ≃ F ( ℘ − 1 ( α )) ⊗ F K for some α ∈ F \alpha\in F α ∈ F and some quadratic field extension K K K of F F F (separable or not). Assume moreover that Q Q Q satisfies the conditions in Theorem 2.3; then
Q ≃ [ α 0 , β 0 ] L for some α 0 ∈ F and β 0 ∈ K × , Q\simeq[\alpha_0,\beta_0]_L\qquad\text{for some }\alpha_0\in F\text{ and }\beta_0\in K^\times, Q ≃ [ α 0 , β 0 ] L for some α 0 ∈ F and β 0 ∈ K × , and there exists β ∈ K × \beta\in K^\times β ∈ K × such that
[ α 0 , N K / F ( β 0 ) ] F ≃ [ α , N K / F ( β ) ] F . (12) [\alpha_0,N_{K/F}(\beta_0)]_F\simeq[\alpha,N_{K/F}(\beta)]_F. \tag*{(12)} [ α 0 , N K / F ( β 0 ) ] F ≃ [ α , N K / F ( β ) ] F . (12)
For any such β ∈ K × \beta\in K^{\times} β ∈ K × , the invariant δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) vanishes if β ∈ F \beta\in F β ∈ F or β 0 β ∈ F \beta_{0}\beta\in F β 0 β ∈ F ; if β ∉ F \beta\notin F β ∈ / F and β 0 β ∉ F \beta_{0}\beta\notin F β 0 β ∈ / F , then δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) is represented in I q 3 ( F ) / s ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) I_{q}^{3}(F)/s_{*}(\langle\langle L^{\times}\rangle\rangle\operatorname{Nrd}_{Q}) I q 3 ( F ) / s ∗ (⟨⟨ L × ⟩⟩ Nrd Q ) by the 3-fold Pfister form
⟨ ⟨ t K ( β ) t K ( β 0 β ) , N K / F ( β 0 β ) , α 0 ⟩ ⟩ , \langle\langle t_{K}(\beta)t_{K}(\beta_{0}\beta),N_{K/F}(\beta_{0}\beta),\alpha_{0}\rangle\rangle, ⟨⟨ t K ( β ) t K ( β 0 β ) , N K / F ( β 0 β ) , α 0 ⟩⟩ ,
where t K : K → F t_{K}:K\to F t K : K → F is any nonzero F F F -linear functional such that t K ( 1 ) = 0 t_{K}(1)=0 t K ( 1 ) = 0 .
Proof. Fix some central simple F F F -algebra A A A of degree 8 containing Q Q Q such that 2 [ A ] = 0 2[A]=0 2 [ A ] = 0 , and let B B B be the centralizer of K K K in A A A . Since N L / K [ Q ] = 0 N_{L/K}[Q]=0 N L / K [ Q ] = 0 , Proposition 2.2 yields a quaternion K K K -algebra Q 0 Q_{0} Q 0 such that Q ≃ Q 0 ⊗ K L Q\simeq Q_{0}\otimes_{K}L Q ≃ Q 0 ⊗ K L . The centralizer of Q 0 Q_{0} Q 0 in B B B is therefore a quaternion K K K -algebra Q 1 Q_{1} Q 1 containing L L L . Thus, B = Q 0 ⊗ K Q 1 B=Q_{0}\otimes_{K}Q_{1} B = Q 0 ⊗ K Q 1 and Q 1 = [ α , β ) K Q_{1}=[\alpha,\beta)_{K} Q 1 = [ α , β ) K for some β ∈ K × \beta\in K^{\times} β ∈ K × . Since B ⊂ A B\subset A B ⊂ A , Theorem 2.1 shows that N K / F [ B ] = 0 N_{K/F}[B]=0 N K / F [ B ] = 0 , hence
N K / F [ Q 0 ] = N K / F [ Q 1 ] = [ [ α , N K / F ( β ) ) F ] . (13) N_{K/F}[Q_{0}]=N_{K/F}[Q_{1}]=[[\alpha,N_{K/F}(\beta))_{F}]. \tag*{(13)} N K / F [ Q 0 ] = N K / F [ Q 1 ] = [[ α , N K / F ( β ) ) F ] . (13)
These equations show that the index of N K / F [ Q 0 ] N_{K/F}[Q_{0}] N K / F [ Q 0 ] is at most 2, hence by [10 Th. 1.1] Q 0 Q_{0} Q 0 is split by some quadratic étale K K K -algebra of the form K ( ℘ − 1 ( α 0 ) ) K(\wp^{-1}(\alpha_{0})) K ( ℘ − 1 ( α 0 )) with α 0 ∈ F \alpha_{0}\in F α 0 ∈ F . We may then find β 0 ∈ K × \beta_{0}\in K^{\times} β 0 ∈ K × such that Q 0 ≃ [ α 0 , β 0 ) K Q_{0}\simeq[\alpha_{0},\beta_{0})_{K} Q 0 ≃ [ α 0 , β 0 ) K . It follows that Q ≃ [ α 0 , β 0 ) L Q\simeq[\alpha_{0},\beta_{0})_{L} Q ≃ [ α 0 , β 0 ) L , and (13) yields (12).
To complete the proof, we calculate δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) . Let β ∈ K × \beta\in K^{\times} β ∈ K × be such that (12) holds, and let
B ′ = [ α 0 , β 0 ) K ⊗ K [ α , β ) K . B'=[\alpha_{0},\beta_{0})_{K}\otimes_{K}[\alpha,\beta)_{K}. B ′ = [ α 0 , β 0 ) K ⊗ K [ α , β ) K .
Since Q ≃ [ α 0 , β 0 ) K ⊗ K K ( ℘ − 1 ( α ) ) Q\simeq[\alpha_{0},\beta_{0})_{K}\otimes_{K}K(\wp^{-1}(\alpha)) Q ≃ [ α 0 , β 0 ) K ⊗ K K ( ℘ − 1 ( α )) , we may identify Q Q Q with a subalgebra of B ′ B' B ′ . From (12), it follows that N K / F [ B ′ ] = 0 N_{K/F}[B']=0 N K / F [ B ′ ] = 0 , hence by Theorem 2.1 there exists a central simple F F F -algebra A ′ A' A ′ of degree 8 containing B ′ B' B ′ , hence also Q Q Q , such that 2 [ A ′ ] = 0 2[A']=0 2 [ A ′ ] = 0 .
If β ∈ F × \beta\in F^{\times} β ∈ F × , then N K / F ( β ) ∈ F × 2 N_{K/F}(\beta)\in F^{\times2} N K / F ( β ) ∈ F × 2 hence (12) shows that [ α 0 , N K / F ( β 0 ) ) F [\alpha_{0},N_{K/F}(\beta_{0}))_{F} [ α 0 , N K / F ( β 0 ) ) F is split. By Proposition 2.2 it follows that Q 0 ≃ Q 0 ′ ⊗ F K Q_{0}\simeq Q'_{0}\otimes_{F}K Q 0 ≃ Q 0 ′ ⊗ F K for some quaternion F F F -algebra Q 0 ′ Q'_{0} Q 0 ′ . Then Q ≃ Q 0 ′ ⊗ F L Q\simeq Q'_{0}\otimes_{F}L Q ≃ Q 0 ′ ⊗ F L and Theorem 4.5 yields δ L / F ( Q ) = 0 \delta_{L/F}(Q)=0 δ L / F ( Q ) = 0 .
Similarly, if β 0 β ∈ F × \beta_{0}\beta\in F^{\times} β 0 β ∈ F × , then N K / F ( β 0 ) ≡ N K / F ( β ) ( m o d F × 2 ) N_{K/F}(\beta_{0})\equiv N_{K/F}(\beta)\pmod{F^{\times2}} N K / F ( β 0 ) ≡ N K / F ( β ) ( mod F × 2 ) , hence (12) shows that [ α 0 + α , N K / F ( β 0 ) ) F [\alpha_{0}+\alpha,N_{K/F}(\beta_{0}))_{F} [ α 0 + α , N K / F ( β 0 ) ) F is split. By Proposition 2.2 we may then find a quaternion F F F -algebra Q 0 ′ ′ Q''_{0} Q 0 ′′ such that [ α 0 + α , β 0 ) K ≃ Q 0 ′ ′ ⊗ F K [\alpha_{0}+\alpha,\beta_{0})_{K}\simeq Q''_{0}\otimes_{F}K [ α 0 + α , β 0 ) K ≃ Q 0 ′′ ⊗ F K . But since α ∈ ℘ ( L ) \alpha\in\wp(L) α ∈ ℘ ( L ) we have Q 0 ≃ [ α 0 + α , β 0 ) L Q_{0}\simeq[\alpha_{0}+\alpha,\beta_{0})_{L} Q 0 ≃ [ α 0 + α , β 0 ) L , hence Q ≃ Q 0 ′ ′ ⊗ F L Q\simeq Q''_{0}\otimes_{F}L Q ≃ Q 0 ′′ ⊗ F L and δ L / F ( Q ) = 0 \delta_{L/F}(Q)=0 δ L / F ( Q ) = 0 by Theorem 4.5.
For the rest of the proof, assume β ∉ F \beta\notin F β ∈ / F and β 0 β ∉ F \beta_{0}\beta\notin F β 0 β ∈ / F , and choose a nonzero F F F -linear functional t K : K → F t_{K}:K\to F t K : K → F such that t K ( 1 ) = 0 t_{K}(1)=0 t K ( 1 ) = 0 , hence t K ( β 0 β ) t K ( β ) ≠ 0 t_{K}(\beta_{0}\beta)t_{K}(\beta)\ne0 t K ( β 0 β ) t K ( β ) = 0 . Let ( 1 , i 0 , j 0 , i 0 j 0 ) (1,i_{0},j_{0},i_{0}j_{0}) ( 1 , i 0 , j 0 , i 0 j 0 ) and ( 1 , i , j , i j ) (1,i,j,ij) ( 1 , i , j , ij ) be the quaternion bases of [ α 0 , β 0 ) K [\alpha_{0},\beta_{0})_{K} [ α 0 , β 0 ) K and [ α , β ) K [\alpha,\beta)_{K} [ α , β ) K respectively, such that
i 0 2 + i 0 = α 0 , j 0 i 0 = ( i 0 + 1 ) j 0 , j 0 2 = β 0 i_{0}^{2}+i_{0}=\alpha_{0},\qquad j_{0}i_{0}=(i_{0}+1)j_{0},\qquad j_{0}^{2}=\beta_{0} i 0 2 + i 0 = α 0 , j 0 i 0 = ( i 0 + 1 ) j 0 , j 0 2 = β 0
and
i 2 + i = α , j i = ( i + 1 ) j , j 2 = β . i^{2}+i=\alpha,\qquad ji=(i+1)j,\qquad j^{2}=\beta. i 2 + i = α , j i = ( i + 1 ) j , j 2 = β .
Let γ 0 \gamma_{0} γ 0 and γ \gamma γ be the canonical involutions on [ α 0 , β 0 ) K [\alpha_{0},\beta_{0})_{K} [ α 0 , β 0 ) K and [ α , β ) K [\alpha,\beta)_{K} [ α , β ) K respectively, and consider the involution σ B ′ \sigma_{B'} σ B ′ on B ′ B' B ′ given by
σ B ′ = γ 0 ⊗ ( Int ( j ) ∘ γ ) . \sigma_{B'}=\gamma_{0}\otimes(\operatorname{Int}(j)\circ\gamma). σ B ′ = γ 0 ⊗ ( Int ( j ) ∘ γ ) .
This involution σ B ′ \sigma_{B'} σ B ′ extends the canonical involution on Q Q Q , and [19 (16.15)] yields
Nrd σ B ′ ≃ ⟨ β ⟩ [ 1 , α 0 + α ] ⊥ ⟨ β 0 ⟩ [ 1 , α 0 ] ⊥ ⟨ β ⟩ [ 1 , α ] , \operatorname{Nrd}_{\sigma_{B'}}\simeq\langle\beta\rangle[1,\alpha_{0}+\alpha]\perp\langle\beta_{0}\rangle[1,\alpha_{0}]\perp\langle\beta\rangle[1,\alpha], Nrd σ B ′ ≃ ⟨ β ⟩ [ 1 , α 0 + α ] ⊥ ⟨ β 0 ⟩ [ 1 , α 0 ] ⊥ ⟨ β ⟩ [ 1 , α ] ,
hence by Frobenius reciprocity (see [13 Prop. 20.2] ) ( t K ) ∗ ( Nrd σ B ′ ) (t_{K})_{*}(\operatorname{Nrd}_{\sigma_{B'}}) ( t K ) ∗ ( Nrd σ B ′ ) is Witt-equivalent to
( t K ) ∗ ( ⟨ β ⟩ [ 1 , α 0 + α ] ⊥ ⟨ β 0 ⟩ [ 1 , α 0 ] ) ≃ ⟨ t K ( β ) ⟩ ⟨ ⟨ N K / F ( β ) , α 0 + α ⟩ ⟩ ⊥ ⟨ t K ( β 0 β ) ⟩ ⟨ ⟨ N K / F ( β 0 β ) , α 0 ⟩ ⟩ . (14) (t_{K})_{*}(\langle\beta\rangle[1,\alpha_{0}+\alpha]\perp\langle\beta_{0}\rangle[1,\alpha_{0}])\simeq
\langle t_{K}(\beta)\rangle\langle\langle N_{K/F}(\beta),\alpha_{0}+\alpha\rangle\rangle\perp\langle t_{K}(\beta_{0}\beta)\rangle\langle\langle N_{K/F}(\beta_{0}\beta),\alpha_{0}\rangle\rangle. \tag*{(14)} ( t K ) ∗ (⟨ β ⟩ [ 1 , α 0 + α ] ⊥ ⟨ β 0 ⟩ [ 1 , α 0 ]) ≃ ⟨ t K ( β )⟩ ⟨⟨ N K / F ( β ) , α 0 + α ⟩⟩ ⊥ ⟨ t K ( β 0 β )⟩ ⟨⟨ N K / F ( β 0 β ) , α 0 ⟩⟩ . (14)
Observe that (12) implies [ α 0 + α , N K / F ( β 0 β ) ) F ≃ [ α 0 , N K / F ( β 0 β ) ) F [\alpha_{0}+\alpha,N_{K/F}(\beta_{0}\beta))_{F}\simeq[\alpha_{0},N_{K/F}(\beta_{0}\beta))_{F} [ α 0 + α , N K / F ( β 0 β ) ) F ≃ [ α 0 , N K / F ( β 0 β ) ) F , hence
⟨ ⟨ N K / F ( β ) , α 0 + α ⟩ ⟩ ≃ ⟨ ⟨ N K / F ( β 0 β ) , α 0 ⟩ ⟩ \langle\langle N_{K/F}(\beta),\alpha_{0}+\alpha\rangle\rangle\simeq\langle\langle N_{K/F}(\beta_{0}\beta),\alpha_{0}\rangle\rangle ⟨⟨ N K / F ( β ) , α 0 + α ⟩⟩ ≃ ⟨⟨ N K / F ( β 0 β ) , α 0 ⟩⟩
and we may rewrite the right side of (14) as
⟨ t K ( β ) , t K ( β 0 β ) ⟩ ⟨ ⟨ N K / F ( β 0 β ) , α 0 ⟩ ⟩ . \langle t_K(\beta),t_K(\beta_0\beta)\rangle\langle\!\langle N_{K/F}(\beta_0\beta),\alpha_0\rangle\!\rangle. ⟨ t K ( β ) , t K ( β 0 β )⟩ ⟨ ⟨ N K / F ( β 0 β ) , α 0 ⟩ ⟩ .
Therefore, ( t K ‾ ) ∗ ( Nrp σ B ′ ) = ⟨ ⟨ t K ( β ) t K ( β 0 β ) , N K / F ( β 0 β ) , α 0 ⟩ ⟩ (\overline{t_K})_*(\operatorname{Nrp}_{\sigma_{B'}})=\langle\!\langle t_K(\beta)t_K(\beta_0\beta),N_{K/F}(\beta_0\beta),\alpha_0\rangle\!\rangle ( t K ) ∗ ( Nrp σ B ′ ) = ⟨ ⟨ t K ( β ) t K ( β 0 β ) , N K / F ( β 0 β ) , α 0 ⟩ ⟩ in I ‾ q 3 ( F ) \overline{I}_q^3(F) I q 3 ( F ) . For any involution σ A ′ \sigma_{A'} σ A ′ on A ′ A' A ′ extending σ B ′ \sigma_{B'} σ B ′ we have π ( σ A ′ ) ‾ = ( t K ‾ ) ∗ ( Nrp σ B ′ ) \overline{\pi(\sigma_{A'})}=(\overline{t_K})_*(\operatorname{Nrp}_{\sigma_{B'}}) π ( σ A ′ ) = ( t K ) ∗ ( Nrp σ B ′ ) by Proposition 3.1. Since by definition δ L / F ( Q ) \delta_{L/F}(Q) δ L / F ( Q ) is represented by π ( σ A ′ ) ‾ \overline{\pi(\sigma_{A'})} π ( σ A ′ ) , the proof is complete. □
Nontriviality of the invariants
The purpose of this section is to give examples where the various invariants introduced in Sections 1, 3 and 4 do not vanish.
Let L L L be a biquadratic field extension of F F F contained in a central simple F F F -algebra A A A of degree 8 such that $2[A]=0$. Assume L L L is not purely inseparable over F F F , and let K ⊂ L K\subset L K ⊂ L be a quadratic extension of F F F in $L. Let also Q Q Q and B B B be the centralizers of L L L and K K K in A A A respectively, so
F ⊂ K ⊂ L ⊂ Q ⊂ B ⊂ A . F\subset K\subset L\subset Q\subset B\subset A. F ⊂ K ⊂ L ⊂ Q ⊂ B ⊂ A .
Fix nonzero F F F -linear functionals s L : L → F s_L:L\to F s L : L → F and s K : K → F s_K:K\to F s K : K → F , and let q B ∈ I q 2 ( K ) q_B\in I_q^2(K) q B ∈ I q 2 ( K ) , q A ∈ I q 2 ( F ) q_A\in I_q^2(F) q A ∈ I q 2 ( F ) be such that e 2 ( q B ) = [ B ] e_2(q_B)=[B] e 2 ( q B ) = [ B ] and e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] . As observed in Lemma 4.2, we have
( s L ‾ ) ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) = ⟨ ⟨ N L / F ( L × ) ⟩ ⟩ q A . (\overline{s_L})_*(\langle\!\langle L^\times\rangle\!\rangle\operatorname{Nrd}_Q)=\langle\!\langle N_{L/F}(L^\times)\rangle\!\rangle q_A. ( s L ) ∗ (⟨ ⟨ L × ⟩ ⟩ Nrd Q ) = ⟨ ⟨ N L / F ( L × )⟩ ⟩ q A .
Similarly, since ( q A ‾ ) K = q B ‾ (\overline{q_A})_K=\overline{q_B} ( q A ) K = q B , Frobenius reciprocity yields
( s K ‾ ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) = ⟨ ⟨ N K / F ( K × ) ⟩ ⟩ q A . (\overline{s_K})_*(\langle\!\langle K^\times\rangle\!\rangle q_B)=\langle\!\langle N_{K/F}(K^\times)\rangle\!\rangle q_A. ( s K ) ∗ (⟨ ⟨ K × ⟩ ⟩ q B ) = ⟨ ⟨ N K / F ( K × )⟩ ⟩ q A .
Therefore, there are canonical maps
I q 3 ( F ) / ( s L ‾ ) ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) → Φ I q 3 ( F ) / ( s K ‾ ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) → Ψ I q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q A (15) I_q^3(F)/(\overline{s_L})_*(\langle\!\langle L^\times\rangle\!\rangle\operatorname{Nrd}_Q)
\xrightarrow{\Phi}
I_q^3(F)/(\overline{s_K})_*(\langle\!\langle K^\times\rangle\!\rangle q_B)
\xrightarrow{\Psi}
I_q^3(F)/\langle\!\langle F^\times\rangle\!\rangle q_A \tag*{(15)} I q 3 ( F ) / ( s L ) ∗ (⟨ ⟨ L × ⟩ ⟩ Nrd Q ) Φ I q 3 ( F ) / ( s K ) ∗ (⟨ ⟨ K × ⟩ ⟩ q B ) Ψ I q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q A (15)
induced by the inclusions ⟨ ⟨ N L / F ( L × ) ⟩ ⟩ q A ⊂ ⟨ ⟨ N K / F ( K × ) ⟩ ⟩ q A ⊂ ⟨ ⟨ F × ⟩ ⟩ q A \langle\!\langle N_{L/F}(L^\times)\rangle\!\rangle q_A\subset\langle\!\langle N_{K/F}(K^\times)\rangle\!\rangle q_A\subset\langle\!\langle F^\times\rangle\!\rangle q_A ⟨ ⟨ N L / F ( L × )⟩ ⟩ q A ⊂ ⟨ ⟨ N K / F ( K × )⟩ ⟩ q A ⊂ ⟨ ⟨ F × ⟩ ⟩ q A .
If σ A \sigma_A σ A is an involution on A A A extending the canonical involution on Q Q Q , then π ( σ A ) \pi(\sigma_A) π ( σ A ) represents the invariant Δ ( A ) ∈ I q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q A \Delta(A)\in I_q^3(F)/\langle\!\langle F^\times\rangle\!\rangle q_A Δ ( A ) ∈ I q 3 ( F ) / ⟨ ⟨ F × ⟩ ⟩ q A , but also δ L / F ( Q ) ∈ I q 3 ( F ) / ( s L ‾ ) ∗ ( ⟨ ⟨ L × ⟩ ⟩ Nrd Q ) \delta_{L/F}(Q)\in I_q^3(F)/(\overline{s_L})_*(\langle\!\langle L^\times\rangle\!\rangle\operatorname{Nrd}_Q) δ L / F ( Q ) ∈ I q 3 ( F ) / ( s L ) ∗ (⟨ ⟨ L × ⟩ ⟩ Nrd Q ) , and δ K / F ( B ) ∈ I q 3 ( F ) / ( s K ‾ ) ∗ ( ⟨ ⟨ K × ⟩ ⟩ q B ) \delta_{K/F}(B)\in I_q^3(F)/(\overline{s_K})_*(\langle\!\langle K^\times\rangle\!\rangle q_B) δ K / F ( B ) ∈ I q 3 ( F ) / ( s K ) ∗ (⟨ ⟨ K × ⟩ ⟩ q B ) , see Definitions 1.3, 3.6 and 4.4. Therefore,
Φ ( δ L / F ( Q ) ) = δ K / F ( B ) and Ψ ( δ K / F ( B ) ) = Δ ( A ) . (16) \Phi(\delta_{L/F}(Q))=\delta_{K/F}(B)\qquad\text{and}\qquad\Psi(\delta_{K/F}(B))=\Delta(A). \tag*{(16)} Φ ( δ L / F ( Q )) = δ K / F ( B ) and Ψ ( δ K / F ( B )) = Δ ( A ) . (16)
Thus, any example of division algebra A A A such that Δ ( A ) ≠ 0 \Delta(A)\ne0 Δ ( A ) = 0 readily yields examples of biquaternion algebras B B B and quaternion algebras Q Q Q such that δ K / F ( B ) ≠ 0 \delta_{K/F}(B)\ne0 δ K / F ( B ) = 0 and δ L / F ( Q ) ≠ 0 \delta_{L/F}(Q)\ne0 δ L / F ( Q ) = 0 . The converse does not hold: Remark 5.4 yields an example where δ L / F ( Q ) ≠ 0 \delta_{L/F}(Q)\ne0 δ L / F ( Q ) = 0 and δ K / F ( B ) = 0 \delta_{K/F}(B)=0 δ K / F ( B ) = 0 , and Remark 5.7 an example where δ K / F ( B ) ≠ 0 \delta_{K/F}(B)\ne0 δ K / F ( B ) = 0 and Δ ( A ) = 0 \Delta(A)=0 Δ ( A ) = 0 .
We discuss below two types of construction: first, we show how to obtain examples where Δ ( A ) ≠ 0 \Delta(A)\ne0 Δ ( A ) = 0 (hence δ K / F ( B ) ≠ 0 \delta_{K/F}(B)\ne0 δ K / F ( B ) = 0 and δ L / F ( Q ) ≠ 0 \delta_{L/F}(Q)\ne0 δ L / F ( Q ) = 0 ) from any indecomposable division algebra of degree 8 and exponent 2. In these examples, the biquadratic extension L / F L/F L / F is separable. Then, we take the opposite approach: we start by giving an example of a quaternion algebra Q Q Q with δ L / F ( Q ) ≠ 0 \delta_{L/F}(Q)\ne0 δ L / F ( Q ) = 0 , and build from this a biquaternion algebra B B B such that δ K / F ( B ) ≠ 0 \delta_{K/F}(B)\ne0 δ K / F ( B ) = 0 and a central simple F F F -algebra A A A such that Δ ( A ) ≠ 0 \Delta(A)\ne0 Δ ( A ) = 0 . In contrast with the construction sketched above, the biquadratic field extension L / F L/F L / F and the quadratic field extension K / F K/F K / F are not separable.
In each case, the main tools for the proofs are residue maps for discrete valuations defined by Arason [3 ] and used by Barry–Chapman–Laghribi [8 ] and by Laghribi–Lorenz [20 ] .
We briefly recall the definitions from [3 ] and [8 §6] : if E E E is a field of characteristic 2 with a discrete valuation v v v and uniformizing element π \pi π , the tame subgroup I q ( E ) 0 ⊂ I q ( E ) I_q(E)_0 \subset I_q(E) I q ( E ) 0 ⊂ I q ( E ) is the W ( E ) W(E) W ( E ) -submodule generated by elements of the form [ 1 , u ] [1,u] [ 1 , u ] with v ( u ) ≥ 0 v(u) \ge0 v ( u ) ≥ 0 . Residue maps
∂ q 0 : I q ( E ) 0 → I q ( E ‾ ) and ∂ q 1 : I q ( E ) 0 → I q ( E ‾ ) , \partial_{q0}: I_q(E)_0 \to I_q(\overline{E}) \qquad\text{and}\qquad\partial_{q1}: I_q(E)_0 \to I_q(\overline{E}), ∂ q 0 : I q ( E ) 0 → I q ( E ) and ∂ q 1 : I q ( E ) 0 → I q ( E ) ,
where E ‾ \overline{E} E is the residue field of E E E , are defined as follows: for u 1 , u 2 ∈ E u_1,u_2 \in E u 1 , u 2 ∈ E with v ( u 1 ) = 0 v(u_1)=0 v ( u 1 ) = 0 and v ( u 2 ) ≥ 0 v(u_2)\ge0 v ( u 2 ) ≥ 0 ,
∂ q 0 ( ⟨ u 1 ⟩ [ 1 , u 2 ] ) = ⟨ u ‾ 1 ⟩ [ 1 , u ‾ 2 ] , ∂ q 0 ( ⟨ u 1 π ⟩ [ 1 , u 2 ] ) = 0 \partial_{q0}(\langle u_1\rangle[1,u_2])=\langle\overline{u}_1\rangle[1,\overline{u}_2], \qquad\partial_{q0}(\langle u_1\pi\rangle[1,u_2])=0 ∂ q 0 (⟨ u 1 ⟩ [ 1 , u 2 ]) = ⟨ u 1 ⟩ [ 1 , u 2 ] , ∂ q 0 (⟨ u 1 π ⟩ [ 1 , u 2 ]) = 0
and
∂ q 1 ( ⟨ u 1 ⟩ [ 1 , u 2 ] ) = 0 , ∂ q 1 ( ⟨ u 1 π ⟩ [ 1 , u 2 ] ) = ⟨ u ‾ 1 ⟩ [ 1 , u ‾ 2 ] . \partial_{q1}(\langle u_1\rangle[1,u_2])=0, \qquad\partial_{q1}(\langle u_1\pi\rangle[1,u_2])=\langle\overline{u}_1\rangle[1,\overline{u}_2]. ∂ q 1 (⟨ u 1 ⟩ [ 1 , u 2 ]) = 0 , ∂ q 1 (⟨ u 1 π ⟩ [ 1 , u 2 ]) = ⟨ u 1 ⟩ [ 1 , u 2 ] .
Thus, the residue maps ∂ q 0 \partial_{q0} ∂ q 0 and ∂ q 1 \partial_{q1} ∂ q 1 are related to the residue homomorphisms ∂ 0 , ∂ 1 : W ( E ) → W ( E ‾ ) \partial_0,\partial_1: W(E)\to W(\overline{E}) ∂ 0 , ∂ 1 : W ( E ) → W ( E ) of [13 §19.A] as follows: for i = 0 , 1 i=0,1 i = 0 , 1 ,
∂ q i ( φ ⋅ [ 1 , u ] ) = ∂ i ( φ ) ⋅ [ 1 , u ‾ ] \partial_{qi}(\varphi\cdot[1,u])=\partial_i(\varphi)\cdot[1,\overline{u}] ∂ q i ( φ ⋅ [ 1 , u ]) = ∂ i ( φ ) ⋅ [ 1 , u ]
for φ ∈ W ( E ) \varphi\in W(E) φ ∈ W ( E ) and u ∈ E × u\in E^\times u ∈ E × with v ( u ) ≥ 0 v(u)\ge0 v ( u ) ≥ 0 .
For n ≥ 2 n\ge2 n ≥ 2 , let I q n ( E ) 0 = I n − 1 ( E ) ⋅ I q ( E ) 0 I_q^n(E)_0=I^{n-1}(E)\cdot I_q(E)_0 I q n ( E ) 0 = I n − 1 ( E ) ⋅ I q ( E ) 0 . Arason’s theorem in [4 ] shows that I q n ( E ) 0 = I q n ( E ) ∩ I q ( E ) 0 I_q^n(E)_0=I_q^n(E)\cap I_q(E)_0 I q n ( E ) 0 = I q n ( E ) ∩ I q ( E ) 0 . Since ∂ i ( I n ( E ) ) ⊂ I n − 1 ( E ‾ ) \partial_i(I^n(E))\subset I^{n-1}(\overline{E}) ∂ i ( I n ( E )) ⊂ I n − 1 ( E ) for i = 0 , 1 i=0,1 i = 0 , 1 and n ≥ 1 n\ge1 n ≥ 1 (see [13 Lemma 19.14] ), it follows that ∂ q i ( I q n ( E ) 0 ) ⊂ I q n − 1 ( E ‾ ) \partial_{qi}(I_q^n(E)_0)\subset I_q^{n-1}(\overline{E}) ∂ q i ( I q n ( E ) 0 ) ⊂ I q n − 1 ( E ) for i = 0 , 1 i=0,1 i = 0 , 1 and n ≥ 2 n\ge2 n ≥ 2 . We also use the map
∂ q + : = ∂ q 0 + ∂ q 1 : I q ( E ) 0 → I q ( E ‾ ) \partial_{q+}:=\partial_{q0}+\partial_{q1}: I_q(E)_0\to I_q(\overline{E}) ∂ q + := ∂ q 0 + ∂ q 1 : I q ( E ) 0 → I q ( E )
such that for u 1 , u 2 ∈ E × u_1,u_2\in E^\times u 1 , u 2 ∈ E × with v ( u 1 ) = 0 v(u_1)=0 v ( u 1 ) = 0 and v ( u 2 ) ≥ 0 v(u_2)\ge0 v ( u 2 ) ≥ 0
∂ q + ( ⟨ u 1 ⟩ [ 1 , u 2 ] ) = ∂ q + ( ⟨ u 1 π ⟩ [ 1 , u 2 ] ) = ⟨ u ‾ 1 ⟩ [ 1 , u ‾ 2 ] . \partial_{q+}(\langle u_1\rangle[1,u_2])=\partial_{q+}(\langle u_1\pi\rangle[1,u_2])=\langle\overline{u}_1\rangle[1,\overline{u}_2]. ∂ q + (⟨ u 1 ⟩ [ 1 , u 2 ]) = ∂ q + (⟨ u 1 π ⟩ [ 1 , u 2 ]) = ⟨ u 1 ⟩ [ 1 , u 2 ] .
The calculations in [13 Ex. 19.13] show that ( ∂ 0 + ∂ 1 ) ( I n ( E ) ) ⊂ I n ( E ‾ ) (\partial_0+\partial_1)(I^n(E))\subset I^n(\overline{E}) ( ∂ 0 + ∂ 1 ) ( I n ( E )) ⊂ I n ( E ) for n ≥ 0 n\ge0 n ≥ 0 , hence ∂ q + ( I q n ( E ) 0 ) ⊂ I q n ( E ‾ ) \partial_{q+}(I_q^n(E)_0)\subset I_q^n(\overline{E}) ∂ q + ( I q n ( E ) 0 ) ⊂ I q n ( E ) .
First construction. Let A A A be a central simple F F F -algebra of degree 8 such that 2 [ A ] = 0 2[A]=0 2 [ A ] = 0 . If its index is 2 or 4, then clearly A A A contains a separable quadratic field extension of F F F since it is Brauer-equivalent to a quaternion algebra or to a biquaternion algebra. If A A A is a division algebra, then by a theorem of Rowen (see [16 Th. 5.6.10] ) it contains a separable triquadratic extension of F F F . Thus, assuming A A A is not split we may fix a separable quadratic field extension K = F ( ℘ − 1 ( a ) ) K=F(\wp^{-1}(a)) K = F ( ℘ − 1 ( a )) of F F F in A A A and consider its centralizer B ⊂ A B\subset A B ⊂ A . Let t t t be an indeterminate over F F F . The F ( t ) F(t) F ( t ) -algebra A ⊗ F [ a , t ) F ( t ) A\otimes_F[a,t)_{F(t)} A ⊗ F [ a , t ) F ( t ) is split by every extension of K K K that splits A A A , hence its index is at most 8 and we may fix a central simple F ( t ) F(t) F ( t ) -algebra A ′ A' A ′ of degree 8 such that
[ A ′ ] = [ A ⊗ F [ a , t ) F ( t ) ] in Br ( F ( t ) ) . (17) [A']=[A\otimes_F[a,t)_{F(t)}]\quad\text{in }\operatorname{Br}(F(t)). \tag*{(17)} [ A ′ ] = [ A ⊗ F [ a , t ) F ( t ) ] in Br ( F ( t )) . (17)
Theorem 5.1. The following conditions are equivalent:
(i) δ K / F ( B ) = 0 \delta_{K/F}(B)=0 δ K / F ( B ) = 0 ;
(ii) A ′ ≃ Q 1 ⊗ F ( t ) Q 2 ⊗ F ( t ) Q 3 A'\simeq Q_1\otimes_{F(t)}Q_2\otimes_{F(t)}Q_3 A ′ ≃ Q 1 ⊗ F ( t ) Q 2 ⊗ F ( t ) Q 3 for some quaternion F ( t ) F(t) F ( t ) -algebras Q 1 , Q 2 , Q 3 Q_1,Q_2,Q_3 Q 1 , Q 2 , Q 3 ;
(iii) Δ ( A ′ ) = 0 \Delta(A')=0 Δ ( A ′ ) = 0 .
They imply that A A A decomposes into a tensor product of quaternion F F F -algebras.
Proof. (i) ⇒ \Rightarrow ⇒ (ii): By Proposition 3.7, (i) implies that B ≃ H 1 ⊗ F H 2 ⊗ F K B\simeq H_1\otimes_F H_2\otimes_F K B ≃ H 1 ⊗ F H 2 ⊗ F K for some quaternion F F F -algebras H 1 , H 2 H_1,H_2 H 1 , H 2 . Since [ B ] = [ A ] K [B]=[A]_K [ B ] = [ A ] K , it follows that A ⊗ F H 1 ⊗ F H 2 A\otimes_F H_1\otimes_F H_2 A ⊗ F H 1 ⊗ F H 2 is split by K K K , hence
A ≃ H 1 ⊗ F H 2 ⊗ F [ a , u ) F A\simeq H_1\otimes_F H_2\otimes_F[a,u)_F A ≃ H 1 ⊗ F H 2 ⊗ F [ a , u ) F
for some u ∈ F × u\in F^\times u ∈ F × .
Then (ii) holds with
Q 1 = H 1 ⊗ F F ( t ) , Q 2 = H 2 ⊗ F F ( t ) and Q 3 = [ a , u t ) F ( t ) . Q_1=H_1\otimes_F F(t),\qquad Q_2=H_2\otimes_F F(t)\qquad\text{and}\qquad Q_3=[a,ut)_{F(t)}. Q 1 = H 1 ⊗ F F ( t ) , Q 2 = H 2 ⊗ F F ( t ) and Q 3 = [ a , u t ) F ( t ) .
(ii) ⇒ \Rightarrow ⇒ (iii): This readily follows from Proposition 1.4.
(iii) ⇒ \Rightarrow ⇒ (i): Let σ B \sigma_B σ B be a symplectic involution on B B B and t K : K → F t_K: K \to F t K : K → F a nonzero F F F -linear functional such that t K ( 1 ) = 0 t_K(1)=0 t K ( 1 ) = 0 . The algebra B ⊗ K K ( t ) B \otimes_K K(t) B ⊗ K K ( t ) is Brauer-equivalent to A K ( t ) ≃ A K ( t ) ′ A_{K(t)} \simeq A'_{K(t)} A K ( t ) ≃ A K ( t ) ′ , therefore, we may consider B ⊗ K K ( t ) B \otimes_K K(t) B ⊗ K K ( t ) as a subalgebra of A ′ A' A ′ . Hence Proposition 3.1 shows that Δ ( A ′ ) \Delta(A') Δ ( A ′ ) is represented by ( t K ) ∗ ( Nrp σ B ) F ( t ) (t_K)_*(\operatorname{Nrp}_{\sigma_B})_{F(t)} ( t K ) ∗ ( Nrp σ B ) F ( t ) . Let us write φ = ( t K ) ∗ ( Nrp σ B ) \varphi=(t_K)_*(\operatorname{Nrp}_{\sigma_B}) φ = ( t K ) ∗ ( Nrp σ B ) , so that φ \varphi φ represents δ K / F ( B ) \delta_{K/F}(B) δ K / F ( B ) and φ F ( t ) \varphi_{F(t)} φ F ( t ) represents Δ ( A ′ ) \Delta(A') Δ ( A ′ ) . We assume Δ ( A ′ ) = 0 \Delta(A')=0 Δ ( A ′ ) = 0 and we have to show that δ K / F ( B ) = 0 \delta_{K/F}(B)=0 δ K / F ( B ) = 0 .
Let q A ∈ I q 2 ( F ) q_A \in I_q^2(F) q A ∈ I q 2 ( F ) be such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] . Then e 2 ( q A + ⟨ ⟨ t , a ] ] = [ A ′ ] e_2(q_A+\langle\!\langle t,a]]=[A'] e 2 ( q A + ⟨ ⟨ t , a ]] = [ A ′ ] by (17), and Δ ( A ′ ) = 0 \Delta(A')=0 Δ ( A ′ ) = 0 means that there exists f ∈ F ( t ) × f\in F(t)^\times f ∈ F ( t ) × such that
φ F ( t ) ≡ ⟨ ⟨ f ⟩ ⟩ ( q A + ⟨ ⟨ t , a ] ] ) mod I q 4 ( F ( t ) ) . (18) \varphi_{F(t)} \equiv\langle\!\langle f\rangle\!\rangle(q_A+\langle\!\langle t,a]]) \mathbin{\operatorname{mod}} I_q^4(F(t)). \tag*{(18)} φ F ( t ) ≡ ⟨ ⟨ f ⟩ ⟩ ( q A + ⟨ ⟨ t , a ]]) mod I q 4 ( F ( t )) . (18)
We use the t t t -adic valuation v 0 v_0 v 0 on F ( t ) F(t) F ( t ) . Since φ \varphi φ , q A q_A q A and ⟨ ⟨ t , a ] ] \langle\!\langle t,a]] ⟨ ⟨ t , a ]] are tame quadratic forms, (18) is modulo I q 4 ( F ( t ) ) ∩ I q ( F ( t ) ) 0 = I q 4 ( F ( t ) ) 0 I_q^4(F(t))\cap I_q(F(t))_0=I_q^4(F(t))_0 I q 4 ( F ( t )) ∩ I q ( F ( t ) ) 0 = I q 4 ( F ( t ) ) 0 . Multiplying f f f by a suitable power of t 2 t^2 t 2 , we may assume v 0 ( f ) = 0 v_0(f)=0 v 0 ( f ) = 0 or 1 1 1 . If v 0 ( f ) = 0 v_0(f)=0 v 0 ( f ) = 0 , let f ( 0 ) ∈ F × f(0)\in F^\times f ( 0 ) ∈ F × denote the residue of f f f . Taking the image of each side of (18) under the residue maps ∂ q 1 \partial_{q1} ∂ q 1 and ∂ q + \partial_{q+} ∂ q + yields respectively
0 ≡ ⟨ ⟨ f ( 0 ) , a ] ] mod I q 3 ( F ) and φ ≡ ⟨ ⟨ f ( 0 ) ⟩ ⟩ q A mod I q 4 ( F ) . 0\equiv\langle\!\langle f(0),a]]\mathbin{\operatorname{mod}} I_q^3(F)
\qquad\text{and}\qquad
\varphi\equiv\langle\!\langle f(0)\rangle\!\rangle q_A\mathbin{\operatorname{mod}} I_q^4(F). 0 ≡ ⟨ ⟨ f ( 0 ) , a ]] mod I q 3 ( F ) and φ ≡ ⟨ ⟨ f ( 0 )⟩ ⟩ q A mod I q 4 ( F ) .
The first relation shows that f ( 0 ) f(0) f ( 0 ) is a norm from K / F K/F K / F ; the second relation then shows that φ \varphi φ vanishes in the quotient I q 3 ( F ) / ⟨ ⟨ N K / F ( K × ) ⟩ ⟩ q A I_q^3(F)/\langle\!\langle N_{K/F}(K^\times)\rangle\!\rangle q_A I q 3 ( F ) / ⟨ ⟨ N K / F ( K × )⟩ ⟩ q A , hence δ K / F ( B ) = 0 \delta_{K/F}(B)=0 δ K / F ( B ) = 0 . Assume next v 0 ( f ) = 1 v_0(f)=1 v 0 ( f ) = 1 and let f 0 = t − 1 f f_0=t^{-1}f f 0 = t − 1 f . Using ⟨ ⟨ f , t ⟩ ⟩ = ⟨ ⟨ f 0 , t ⟩ ⟩ \langle\!\langle f,t\rangle\!\rangle=\langle\!\langle f_0,t\rangle\!\rangle ⟨ ⟨ f , t ⟩ ⟩ = ⟨ ⟨ f 0 , t ⟩ ⟩ , we find by applying ∂ q 1 \partial_{q1} ∂ q 1 and ∂ q + \partial_{q+} ∂ q + to (18)
0 ≡ ⟨ f 0 ( 0 ) ⟩ q A + ⟨ ⟨ f 0 ( 0 ) , a ] ] mod I q 3 ( F ) and φ ≡ ⟨ ⟨ f 0 ( 0 ) ⟩ ⟩ q A mod I q 4 ( F ) . 0\equiv\langle f_0(0)\rangle q_A+\langle\!\langle f_0(0),a]]\mathbin{\operatorname{mod}} I_q^3(F)
\qquad\text{and}\qquad
\varphi\equiv\langle\!\langle f_0(0)\rangle\!\rangle q_A\mathbin{\operatorname{mod}} I_q^4(F). 0 ≡ ⟨ f 0 ( 0 )⟩ q A + ⟨ ⟨ f 0 ( 0 ) , a ]] mod I q 3 ( F ) and φ ≡ ⟨ ⟨ f 0 ( 0 )⟩ ⟩ q A mod I q 4 ( F ) .
Then φ ≡ ⟨ ⟨ f 0 ( 0 ) , f 0 ( 0 ) , a ] ] ≡ 0 mod I q 4 ( F ) \varphi\equiv\langle\!\langle f_0(0),f_0(0),a]]\equiv0\mathbin{\operatorname{mod}} I_q^4(F) φ ≡ ⟨ ⟨ f 0 ( 0 ) , f 0 ( 0 ) , a ]] ≡ 0 mod I q 4 ( F ) , so φ ˉ = 0 \bar{\varphi}=0 φ ˉ = 0 in I q 3 ( F ) I_q^3(F) I q 3 ( F ) . Thus, δ K / F ( B ) = 0 \delta_{K/F}(B)=0 δ K / F ( B ) = 0 in each case, proving that (iii) implies (i).
To complete the proof, observe that in the proof of (i) ⇒ \Rightarrow ⇒ (ii) above, we also saw that (i) implies that A A A decomposes into a tensor product of three quaternion algebras. □
As a consequence of Theorem 5.1, for any indecomposable division F F F -algebra A A A of degree 8 and exponent 2, such as those constructed by Rowen [23 ] , McKinnie [21 ] or Karpenko [17 ] , the corresponding F ( t ) F(t) F ( t ) -algebra A ′ A' A ′ has nontrivial Δ \Delta Δ -invariant. By Rowen’s theorem in [16 Th. 5.6.10] , A ′ A' A ′ contains a separable triquadratic extension of F ( t ) F(t) F ( t ) , hence also a separable biquadratic extension L ′ L' L ′ and a separable quadratic extension K ′ ⊂ L ′ K'\subset L' K ′ ⊂ L ′ . In view of (16), for Q ′ Q' Q ′ and B ′ B' B ′ the centralizers of L ′ L' L ′ and K ′ K' K ′ in A ′ A' A ′ , we have δ L ′ / F ( t ) ( Q ′ ) ≠ 0 \delta_{L'/F(t)}(Q')\ne0 δ L ′ / F ( t ) ( Q ′ ) = 0 and δ K ′ / F ( t ) ( B ′ ) ≠ 0 \delta_{K'/F(t)}(B')\ne0 δ K ′ / F ( t ) ( B ′ ) = 0 .
Remark. The proof of (iii) ⇒ \Rightarrow ⇒ (i) in Theorem 5.1 is directly inspired by the proof of [8, Prop. 6.1] by Barry–Chapman–Laghribi, which shows in cohomological terms that the map
I q 3 ( F ) ‾ / ⟨ ⟨ N K / F ( K × ) ⟩ ⟩ q A ⟶ I q 3 ( F ( t ) ) ‾ / ⟨ ⟨ F ( t ) × ⟩ ⟩ ( q A + ⟨ ⟨ t , a ] ] ) \overline{I_q^3(F)}/\langle\!\langle N_{K/F}(K^\times)\rangle\!\rangle q_A \longrightarrow\overline{I_q^3(F(t))}/\langle\!\langle F(t)^\times\rangle\!\rangle(q_A+\langle\!\langle t,a]]) I q 3 ( F ) / ⟨ ⟨ N K / F ( K × )⟩ ⟩ q A ⟶ I q 3 ( F ( t )) / ⟨ ⟨ F ( t ) × ⟩ ⟩ ( q A + ⟨ ⟨ t , a ]])
induced by scalar extension is injective. Since this map carries δ K / F ( B ) \delta_{K/F}(B) δ K / F ( B ) to Δ ( A ′ ) \Delta(A') Δ ( A ′ ) , it readily follows that (iii) ⇒ \Rightarrow ⇒ (i).
Note, however, that the proof of [8, Prop. 6.1] has a gap: it is claimed without proof in the first lines of p. 315 that Kato’s isomorphism induces an isomorphism H 2 m + 1 ( L ) 0 → I q m + 1 ( L ) 0 / I q m + 2 ( L ) 0 H_2^{m+1}(L)_0\to I_q^{m+1}(L)_0/I_q^{m+2}(L)_0 H 2 m + 1 ( L ) 0 → I q m + 1 ( L ) 0 / I q m + 2 ( L ) 0 (where the 0-subscript denotes the intersection with the unramified group), whereas the target group is the quotient of I q m + 1 ( L ) 0 I_q^{m+1}(L)_0 I q m + 1 ( L ) 0 by its intersection with I q m + 2 ( L ) I_q^{m+2}(L) I q m + 2 ( L ) . The claim holds, however, because Arason’s theorem [4 ] yields I q m + 2 ( L ) 0 = I q m + 1 ( L ) 0 ∩ I q m + 2 ( L ) I_q^{m+2}(L)_0=I_q^{m+1}(L)_0\cap I_q^{m+2}(L) I q m + 2 ( L ) 0 = I q m + 1 ( L ) 0 ∩ I q m + 2 ( L ) .
Second construction. The following notation is used until the end of the section: k k k is an arbitrary field of characteristic 2 2 2 and a 1 a_1 a 1 , a 2 a_2 a 2 , t 1 t_1 t 1 , t 2 t_2 t 2 , t 3 t_3 t 3 are independent indeterminates over k k k . Let
F = k ( a 1 , a 2 ) , F 1 = F ( t 1 ) , F 2 = F ( t 1 , t 2 ) , F 3 = F ( t 1 , t 2 , t 3 ) . F = k(a_1,a_2), \qquad F_1 = F(t_1), \qquad F_2 = F(t_1,t_2), \qquad F_3 = F(t_1,t_2,t_3). F = k ( a 1 , a 2 ) , F 1 = F ( t 1 ) , F 2 = F ( t 1 , t 2 ) , F 3 = F ( t 1 , t 2 , t 3 ) .
Let also K = F ( a 1 a 2 ) K = F(\sqrt{a_1a_2}) K = F ( a 1 a 2 ) and L = K ( ℘ − 1 ( a 2 ) ) L = K(\wp^{-1}(a_2)) L = K ( ℘ − 1 ( a 2 )) , and
K i = F i ( a 1 a 2 ) and L i = K i ( ℘ − 1 ( a 2 ) ) for i = 1 , 2 , 3. K_i = F_i(\sqrt{a_1a_2}) \qquad\text{and}\qquad L_i = K_i(\wp^{-1}(a_2)) \qquad\text{for } i = 1,2,3. K i = F i ( a 1 a 2 ) and L i = K i ( ℘ − 1 ( a 2 )) for i = 1 , 2 , 3.
Consider the following central simple algebras over L 1 L_1 L 1 , K 2 K_2 K 2 and F 3 F_3 F 3 respectively:
A 1 = [ a 1 a 2 , a 1 + a 2 ) F ⊗ F [ a 1 , t 1 ) F 1 ⊗ F 1 L 1 , (19) A_1 = [a_1a_2,a_1+a_2)_F \otimes_F [a_1,t_1)_{F_1} \otimes_{F_1} L_1, \tag*{(19)} A 1 = [ a 1 a 2 , a 1 + a 2 ) F ⊗ F [ a 1 , t 1 ) F 1 ⊗ F 1 L 1 , (19)
A 2 = [ a 1 a 2 , a 1 + a 2 ) F ⊗ F [ a 1 , t 1 ) F 1 ⊗ F 1 [ a 2 , t 2 ) F 2 ⊗ F 2 K 2 , (20) A_2 = [a_1a_2,a_1+a_2)_F \otimes_F [a_1,t_1)_{F_1} \otimes_{F_1} [a_2,t_2)_{F_2} \otimes_{F_2} K_2, \tag*{(20)} A 2 = [ a 1 a 2 , a 1 + a 2 ) F ⊗ F [ a 1 , t 1 ) F 1 ⊗ F 1 [ a 2 , t 2 ) F 2 ⊗ F 2 K 2 , (20)
A 3 = [ a 1 a 2 , a 1 + a 2 ) F ⊗ F [ a 1 , t 1 ) F 1 ⊗ F 1 [ a 2 , t 2 ) F 2 ⊗ F 2 [ t 3 , a 1 a 2 ) F 3 . (21) A_3 = [a_1a_2,a_1+a_2)_F \otimes_F [a_1,t_1)_{F_1} \otimes_{F_1} [a_2,t_2)_{F_2} \otimes_{F_2} [t_3,a_1a_2)_{F_3}. \tag*{(21)} A 3 = [ a 1 a 2 , a 1 + a 2 ) F ⊗ F [ a 1 , t 1 ) F 1 ⊗ F 1 [ a 2 , t 2 ) F 2 ⊗ F 2 [ t 3 , a 1 a 2 ) F 3 . (21)
We show in § 4.1 that A 1 A_1 A 1 is Brauer-equivalent to a quaternion L 1 L_1 L 1 -algebra Q Q Q such that δ L 1 / F 1 ( Q ) ≠ 0 \delta_{L_1/F_1}(Q) \ne0 δ L 1 / F 1 ( Q ) = 0 , in § 4.2 that A 2 A_2 A 2 is Brauer-equivalent to a biquaternion K 2 K_2 K 2 -algebra B B B such that δ K 2 / F 2 ( B ) ≠ 0 \delta_{K_2/F_2}(B) \ne0 δ K 2 / F 2 ( B ) = 0 and in §5.3 that A 3 A_3 A 3 is Brauer-equivalent to a central simple F 3 F_3 F 3 -algebra A A A of degree 8 8 8 such that Δ ( A ) ≠ 0 \Delta(A) \ne0 Δ ( A ) = 0 .
Nontriviality of the invariant of quaternion algebras
In order to show that the L 1 L_1 L 1 -algebra A 1 A_1 A 1 (defined in (19)) is Brauer-equivalent to a quaternion algebra, we use the following observation: for any field E E E of characteristic 2 2 2 and any α 1 , α 2 ∈ E \alpha_1,\alpha_2 \in E α 1 , α 2 ∈ E and β 1 , β 2 ∈ E × \beta_1,\beta_2 \in E^\times β 1 , β 2 ∈ E × with β 1 ≠ β 2 \beta_1 \ne\beta_2 β 1 = β 2 ,
[ α 1 , β 1 ) E ⊗ E [ α 2 , β 2 ) E ≃ [ α 1 + α 2 , β 1 + β 2 ) E ⊗ E [ α 1 β 2 + α 2 β 1 β 1 + β 2 , β 1 β 2 ) E . (22) [\alpha_1,\beta_1)_E \otimes_E [\alpha_2,\beta_2)_E \simeq[\alpha_1+\alpha_2,\beta_1+\beta_2)_E \otimes_E \left[\frac{\alpha_1\beta_2+\alpha_2\beta_1}{\beta_1+\beta_2},\beta_1\beta_2\right)_E. \tag*{(22)} [ α 1 , β 1 ) E ⊗ E [ α 2 , β 2 ) E ≃ [ α 1 + α 2 , β 1 + β 2 ) E ⊗ E [ β 1 + β 2 α 1 β 2 + α 2 β 1 , β 1 β 2 ) E . (22)
This can be seen as follows: if i 1 i_1 i 1 , i 2 i_2 i 2 , j 1 j_1 j 1 , j 2 ∈ [ α 1 , β 1 ) E ⊗ E [ α 2 , β 2 ) E j_2 \in[\alpha_1,\beta_1)_E \otimes_E [\alpha_2,\beta_2)_E j 2 ∈ [ α 1 , β 1 ) E ⊗ E [ α 2 , β 2 ) E are such that
i r 2 + i r = α r , j r 2 = β r , i r j r + j r i r = j r for r = 1 , 2 i_r^2+i_r=\alpha_r,\qquad j_r^2=\beta_r,\qquad i_rj_r+j_ri_r=j_r\qquad\text{for }r=1,2 i r 2 + i r = α r , j r 2 = β r , i r j r + j r i r = j r for r = 1 , 2
and i 1 i_1 i 1 , j 1 j_1 j 1 commute with i 2 i_2 i 2 and j 2 j_2 j 2 , then i 1 + i 2 i_1+i_2 i 1 + i 2 and j 1 + j 2 j_1+j_2 j 1 + j 2 generate a quaternion E E E -algebra isomorphic to [ α 1 + α 2 , β 1 + β 2 ) E [\alpha_1+\alpha_2,\beta_1+\beta_2)_E [ α 1 + α 2 , β 1 + β 2 ) E . Its centralizer contains ( i 1 j 2 + j 1 i 2 ) ( j 1 + j 2 ) β 1 + β 2 \frac{(i_1j_2+j_1i_2)(j_1+j_2)}{\beta_1+\beta_2} β 1 + β 2 ( i 1 j 2 + j 1 i 2 ) ( j 1 + j 2 ) and j 1 j 2 j_1j_2 j 1 j 2 , which generate a quaternion E E E -algebra [ α 1 β 2 + α 2 β 1 β 1 + β 2 , β 1 β 2 ) E \left[\frac{\alpha_1\beta_2+\alpha_2\beta_1}{\beta_1+\beta_2},\beta_1\beta_2\right)_E [ β 1 + β 2 α 1 β 2 + α 2 β 1 , β 1 β 2 ) E .
Applying (22) with E = K E=K E = K , α r = a r \alpha_r=a_r α r = a r and β r = a r + a 1 a 2 \beta_r=a_r+\sqrt{a_1a_2} β r = a r + a 1 a 2 shows that
[ a 1 , a 1 + a 1 a 2 ) K ⊗ K [ a 2 , a 2 + a 1 a 2 ) K ≃ [ a 1 + a 2 , a 1 + a 2 ) K ⊗ K [ a 1 a 2 , ( a 1 + a 2 ) a 1 a 2 ) K . [a_1,a_1+\sqrt{a_1a_2})_K \otimes_K [a_2,a_2+\sqrt{a_1a_2})_K \simeq[a_1+a_2,a_1+a_2)_K \otimes_K [\sqrt{a_1a_2},(a_1+a_2)\sqrt{a_1a_2})_K. [ a 1 , a 1 + a 1 a 2 ) K ⊗ K [ a 2 , a 2 + a 1 a 2 ) K ≃ [ a 1 + a 2 , a 1 + a 2 ) K ⊗ K [ a 1 a 2 , ( a 1 + a 2 ) a 1 a 2 ) K .
The first quaternion factor on the right side is split; since a 1 a 2 ≡ a 1 a 2 ( m o d ℘ ( K ) ) a_1a_2 \equiv\sqrt{a_1a_2} \pmod{\wp(K)} a 1 a 2 ≡ a 1 a 2 ( mod ℘ ( K )) , the second is isomorphic to
[ a 1 a 2 , a 1 + a 2 ) K ≃ [ a 1 a 2 , a 1 + a 2 ) K . [\sqrt{a_1a_2},a_1+a_2)_K \simeq[a_1a_2,a_1+a_2)_K. [ a 1 a 2 , a 1 + a 2 ) K ≃ [ a 1 a 2 , a 1 + a 2 ) K .
Therefore,
[ a 1 a 2 , a 1 + a 2 ) K = [ a 1 , a 1 + a 1 a 2 ) K ⊗ K [ a 2 , a 2 + a 1 a 2 ) K . (23) [a_1a_2,a_1+a_2)_K = [a_1,a_1+\sqrt{a_1a_2})_K \otimes_K [a_2,a_2+\sqrt{a_1a_2})_K. \tag*{(23)} [ a 1 a 2 , a 1 + a 2 ) K = [ a 1 , a 1 + a 1 a 2 ) K ⊗ K [ a 2 , a 2 + a 1 a 2 ) K . (23)
Extending scalars to L 1 L_1 L 1 splits the second factor on the right side since a 2 ∈ ℘ ( L 1 ) a_2 \in\wp(L_1) a 2 ∈ ℘ ( L 1 ) , hence [ a 1 a 2 , a 1 + a 2 ) L 1 ≃ [ a 1 , a 1 + a 1 a 2 ) L 1 [a_1a_2,a_1+a_2)_{L_1} \simeq[a_1,a_1+\sqrt{a_1a_2})_{L_1} [ a 1 a 2 , a 1 + a 2 ) L 1 ≃ [ a 1 , a 1 + a 1 a 2 ) L 1 . Substituting in (19), we see that A 1 A_1 A 1 is Brauer-equivalent to the quaternion algebra
Q = [ a 1 , ( a 1 + a 1 a 2 ) t 1 ) L 1 . Q = [a_1,(a_1+\sqrt{a_1a_2})t_1)_{L_1}. Q = [ a 1 , ( a 1 + a 1 a 2 ) t 1 ) L 1 .
The conditions in Theorem 2.3 are satisfied because A 1 A_1 A 1 has a descent to a tensor product of quaternion F 1 F_1 F 1 -algebras, hence the invariant δ L 1 / F 1 ( Q ) \delta_{L_1/F_1}(Q) δ L 1 / F 1 ( Q ) is defined. We may use Proposition 4.6 to compute δ L 1 / F 1 ( Q ) \delta_{L_1/F_1}(Q) δ L 1 / F 1 ( Q ) , since a 1 ∈ F 1 a_1 \in F_1 a 1 ∈ F 1 and ( a 1 + a 1 a 2 ) t 1 ∈ K 1 × (a_1+\sqrt{a_1a_2})t_1 \in K_1^\times ( a 1 + a 1 a 2 ) t 1 ∈ K 1 × . Note that
N K 1 / F 1 ( ( a 1 + a 1 a 2 ) t 1 ) = a 1 ( a 1 + a 2 ) t 1 2 . N_{K_1/F_1}\bigl((a_1+\sqrt{a_1a_2})t_1\bigr)=a_1(a_1+a_2)t_1^2. N K 1 / F 1 ( ( a 1 + a 1 a 2 ) t 1 ) = a 1 ( a 1 + a 2 ) t 1 2 .
hence
[ a 1 , N K 1 / F 1 ( ( a 1 + a 1 a 2 ) t 1 ) ] F 1 ≃ [ a 1 , a 1 + a 2 ) F 1 ≃ [ a 2 , a 1 + a 2 ) F 1 ≃ [ a 2 , N K 1 / F 1 ( a 2 + a 1 a 2 ) ] F 1 . [a_1,N_{K_1/F_1}((a_1+\sqrt{a_1a_2})t_1)]_{F_1}\simeq[a_1,a_1+a_2)_{F_1}\simeq[a_2,a_1+a_2)_{F_1}\simeq[a_2,N_{K_1/F_1}(a_2+\sqrt{a_1a_2})]_{F_1}. [ a 1 , N K 1 / F 1 (( a 1 + a 1 a 2 ) t 1 ) ] F 1 ≃ [ a 1 , a 1 + a 2 ) F 1 ≃ [ a 2 , a 1 + a 2 ) F 1 ≃ [ a 2 , N K 1 / F 1 ( a 2 + a 1 a 2 ) ] F 1 .
Therefore, Proposition 4.6 with α 0 = a 1 \alpha_0=a_1 α 0 = a 1 , β 0 = ( a 1 + a 1 a 2 ) t 1 \beta_0=(a_1+\sqrt{a_1a_2})t_1 β 0 = ( a 1 + a 1 a 2 ) t 1 , α = a 2 \alpha=a_2 α = a 2 and β = a 2 + a 1 a 2 \beta=a_2+\sqrt{a_1a_2} β = a 2 + a 1 a 2 , and with t K 1 : K 1 → F 1 t_{K_1}:K_1\to F_1 t K 1 : K 1 → F 1 given by t K 1 ( 1 ) = 0 t_{K_1}(1)=0 t K 1 ( 1 ) = 0 and t K 1 ( a 1 a 2 ) = 1 t_{K_1}(\sqrt{a_1a_2})=1 t K 1 ( a 1 a 2 ) = 1 , shows that δ L 1 / F 1 ( Q ) \delta_{L_1/F_1}(Q) δ L 1 / F 1 ( Q ) is represented by ⟨ ⟨ t 1 ( a 1 + a 2 ) , a 1 a 2 , a 1 ⟩ ⟩ \langle\langle t_1(a_1+a_2),a_1a_2,a_1\rangle\rangle ⟨⟨ t 1 ( a 1 + a 2 ) , a 1 a 2 , a 1 ⟩⟩ . Moreover, the Pfister form ⟨ ⟨ a 1 + a 2 , a 1 a 2 , a 1 ⟩ ⟩ \langle\langle a_1+a_2,a_1a_2,a_1\rangle\rangle ⟨⟨ a 1 + a 2 , a 1 a 2 , a 1 ⟩⟩ represents a 1 2 + a 1 a 2 + a 1 ( a 1 + a 2 ) = 0 a_1^2+a_1a_2+a_1(a_1+a_2)=0 a 1 2 + a 1 a 2 + a 1 ( a 1 + a 2 ) = 0 , hence it is hyperbolic. Since in addition ⟨ ⟨ a 1 a 2 , a 1 ⟩ ⟩ = ⟨ ⟨ a 2 , a 1 ⟩ ⟩ \langle\langle a_1a_2,a_1\rangle\rangle=\langle\langle a_2,a_1\rangle\rangle ⟨⟨ a 1 a 2 , a 1 ⟩⟩ = ⟨⟨ a 2 , a 1 ⟩⟩ , we finally get that δ L 1 / F 1 ( Q ) \delta_{L_1/F_1}(Q) δ L 1 / F 1 ( Q ) is represented by ⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ \langle\langle t_1,a_2,a_1\rangle\rangle ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ .
Theorem 5.2. δ L 1 / F 1 ( Q ) ≠ 0 \delta_{L_1/F_1}(Q)\ne0 δ L 1 / F 1 ( Q ) = 0 .
Proof. In view of the calculation of δ L 1 / F 1 ( Q ) \delta_{L_1/F_1}(Q) δ L 1 / F 1 ( Q ) above, we have to show
⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ ∉ ( s L 1 ) ∗ ( ⟨ ⟨ L 1 × ⟩ ⟩ Nrd Q ) . \langle\langle t_1,a_2,a_1\rangle\rangle\notin(s_{L_1})_*(\langle\langle L_1^\times\rangle\rangle\operatorname{Nrd}_Q). ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ ∈ / ( s L 1 ) ∗ (⟨⟨ L 1 × ⟩⟩ Nrd Q ) .
for s L 1 : L 1 → F 1 s_{L_1}:L_1\to F_1 s L 1 : L 1 → F 1 an arbitrary nonzero F 1 F_1 F 1 -linear functional. Since Q Q Q is Brauer-equivalent to A 1 A_1 A 1 , it follows from the definition of A 1 A_1 A 1 that
Nrd Q = ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ⟩ ⟩ + ⟨ ⟨ t 1 , a 1 ⟩ ⟩ ) L 1 \operatorname{Nrd}_Q=\bigl(\langle\langle a_1+a_2,a_1a_2\rangle\rangle+\langle\langle t_1,a_1\rangle\rangle\bigr)_{L_1} Nrd Q = ( ⟨⟨ a 1 + a 2 , a 1 a 2 ⟩⟩ + ⟨⟨ t 1 , a 1 ⟩⟩ ) L 1
in I q 2 ( L 1 ) I_q^2(L_1) I q 2 ( L 1 ) .
Therefore, by Frobenius reciprocity
( s L 1 ) ∗ ( ⟨ ⟨ L 1 × ⟩ ⟩ Nrd Q ) = ⟨ ⟨ N L 1 / F 1 ( L 1 × ) ⟩ ⟩ ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ⟩ ⟩ + ⟨ ⟨ t 1 , a 1 ⟩ ⟩ ) (s_{L_1})_*(\langle\langle L_1^\times\rangle\rangle\operatorname{Nrd}_Q)=\langle\langle N_{L_1/F_1}(L_1^\times)\rangle\rangle\bigl(\langle\langle a_1+a_2,a_1a_2\rangle\rangle+\langle\langle t_1,a_1\rangle\rangle\bigr) ( s L 1 ) ∗ (⟨⟨ L 1 × ⟩⟩ Nrd Q ) = ⟨⟨ N L 1 / F 1 ( L 1 × )⟩⟩ ( ⟨⟨ a 1 + a 2 , a 1 a 2 ⟩⟩ + ⟨⟨ t 1 , a 1 ⟩⟩ )
in I q 3 ( F 1 ) I_q^3(F_1) I q 3 ( F 1 ) .
By way of contradiction, suppose there exists ℓ ∈ L 1 × \ell\in L_1^\times ℓ ∈ L 1 × such that
⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ ≡ ⟨ ⟨ N L 1 / F 1 ( ℓ ) ⟩ ⟩ ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ⟩ ⟩ + ⟨ ⟨ t 1 , a 1 ⟩ ⟩ ) m o d I q 4 ( F 1 ) . (24) \langle\langle t_1,a_2,a_1\rangle\rangle\equiv\langle\langle N_{L_1/F_1}(\ell)\rangle\rangle\bigl(\langle\langle a_1+a_2,a_1a_2\rangle\rangle+\langle\langle t_1,a_1\rangle\rangle\bigr)\mod I_q^4(F_1). \tag*{(24)} ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ ≡ ⟨⟨ N L 1 / F 1 ( ℓ )⟩⟩ ( ⟨⟨ a 1 + a 2 , a 1 a 2 ⟩⟩ + ⟨⟨ t 1 , a 1 ⟩⟩ ) mod I q 4 ( F 1 ) . (24)
Consider the t 1 t_1 t 1 -adic valuation v 0 v_0 v 0 on L 1 L_1 L 1 (with constant field L L L ) and let ℓ = ℓ 0 t 1 n \ell=\ell_0t_1^n ℓ = ℓ 0 t 1 n with n = v 0 ( ℓ ) n=v_0(\ell) n = v 0 ( ℓ ) , so ℓ 0 \ell_0 ℓ 0 is a unit for the v 0 v_0 v 0 -valuation. Its residue is ℓ 0 ( 0 ) ∈ L × \ell_0(0)\in L^\times ℓ 0 ( 0 ) ∈ L × and the residue of N L 1 / F 1 ( ℓ 0 ) N_{L_1/F_1}(\ell_0) N L 1 / F 1 ( ℓ 0 ) is N L / F ( ℓ 0 ( 0 ) ) N_{L/F}(\ell_0(0)) N L / F ( ℓ 0 ( 0 )) . Since a 1 , a 2 ∈ F a_1,a_2\in F a 1 , a 2 ∈ F , all the terms in (24) are in the tame subgroup I q ( F 1 ) 0 I_q(F_1)_0 I q ( F 1 ) 0 , hence the equation is modulo I q 4 ( F 1 ) ∩ I q ( F 1 ) 0 = I q 4 ( F 1 ) 0 I_q^4(F_1)\cap I_q(F_1)_0=I_q^4(F_1)_0 I q 4 ( F 1 ) ∩ I q ( F 1 ) 0 = I q 4 ( F 1 ) 0 . Since N L 1 / F 1 ( ℓ ) = N L 1 / F 1 ( ℓ 0 ) t 1 4 n N_{L_1/F_1}(\ell)=N_{L_1/F_1}(\ell_0)t_1^{4n} N L 1 / F 1 ( ℓ ) = N L 1 / F 1 ( ℓ 0 ) t 1 4 n , we have ⟨ ⟨ N L 1 / F 1 ( ℓ ) ⟩ ⟩ ≃ ⟨ ⟨ N L 1 / F 1 ( ℓ 0 ) ⟩ ⟩ \langle\langle N_{L_1/F_1}(\ell)\rangle\rangle\simeq\langle\langle N_{L_1/F_1}(\ell_0)\rangle\rangle ⟨⟨ N L 1 / F 1 ( ℓ )⟩⟩ ≃ ⟨⟨ N L 1 / F 1 ( ℓ 0 )⟩⟩ , hence we may substitute N L 1 / F 1 ( ℓ 0 ) N_{L_1/F_1}(\ell_0) N L 1 / F 1 ( ℓ 0 ) for N L 1 / F 1 ( ℓ ) N_{L_1/F_1}(\ell) N L 1 / F 1 ( ℓ ) in (24). Applying the residue map ∂ q 1 \partial_{q_1} ∂ q 1 , we then obtain
⟨ ⟨ a 2 , a 1 ⟩ ⟩ ≡ ⟨ ⟨ N L / F ( ℓ 0 ( 0 ) ) , a 1 ⟩ ⟩ m o d I q 3 ( F ) , \langle\langle a_2,a_1\rangle\rangle\equiv\langle\langle N_{L/F}(\ell_0(0)),a_1\rangle\rangle\mod I_q^3(F), ⟨⟨ a 2 , a 1 ⟩⟩ ≡ ⟨⟨ N L / F ( ℓ 0 ( 0 )) , a 1 ⟩⟩ mod I q 3 ( F ) ,
hence ⟨ ⟨ a 2 N L / F ( ℓ 0 ( 0 ) ) , a 1 ⟩ ⟩ \langle\langle a_2N_{L/F}(\ell_0(0)),a_1\rangle\rangle ⟨⟨ a 2 N L / F ( ℓ 0 ( 0 )) , a 1 ⟩⟩ is hyperbolic, and therefore
a 2 N L / F ( ℓ 0 ( 0 ) ) = N F ( ℘ − 1 ( a 1 ) ) / F ( u ) (25) a_2N_{L/F}(\ell_0(0))=N_{F(\wp^{-1}(a_1))/F}(u) \tag*{(25)} a 2 N L / F ( ℓ 0 ( 0 )) = N F ( ℘ − 1 ( a 1 )) / F ( u ) (25)
for some u ∈ F ( ℘ − 1 ( a 1 ) ) × u\in F(\wp^{-1}(a_1))^\times u ∈ F ( ℘ − 1 ( a 1 ) ) × .
To complete the proof, we show that (25) is impossible. For this, we use the a 2 − 1 a_2^{-1} a 2 − 1 -adic valuation v v v on F = k ( a 1 ) ( a 2 − 1 ) F=k(a_1)(a_2^{-1}) F = k ( a 1 ) ( a 2 − 1 ) (with constant field k ( a 1 ) k(a_1) k ( a 1 ) ). This valuation extends uniquely and without ramification to F ( ℘ − 1 ( a 1 ) ) F(\wp^{-1}(a_1)) F ( ℘ − 1 ( a 1 )) with residue field k ( a 1 ) ( ℘ − 1 ( a 1 ) ) k(a_1)(\wp^{-1}(a_1)) k ( a 1 ) ( ℘ − 1 ( a 1 )) , hence, writing again v v v for the unique extension of v v v to F ( ℘ − 1 ( a 1 ) ) F(\wp^{-1}(a_1)) F ( ℘ − 1 ( a 1 )) ,
v ( N F ( ℘ − 1 ( a 1 ) ) / F ( u ) ) = 2 v ( u ) ∈ 2 Z . v\bigl(N_{F(\wp^{-1}(a_1))/F}(u)\bigr)=2v(u)\in2\mathbb{Z}. v ( N F ( ℘ − 1 ( a 1 )) / F ( u ) ) = 2 v ( u ) ∈ 2 Z .
To see that v v v also extends uniquely to L L L , observe that, writing w w w for an extension of v v v to L L L , we must have
w ( a 1 a 2 ) = 1 2 v ( a 2 ) = − 1 2 w\left(\sqrt{a_{1}a_{2}}\right)=\frac{1}{2}v(a_{2})=-\frac{1}{2} w ( a 1 a 2 ) = 2 1 v ( a 2 ) = − 2 1
and also
w ( ℘ − 1 ( a 2 ) ) = 1 2 v ( a 2 ) = − 1 2 w\left(\wp^{-1}(a_{2})\right)=\frac{1}{2}v(a_{2})=-\frac{1}{2} w ( ℘ − 1 ( a 2 ) ) = 2 1 v ( a 2 ) = − 2 1
because ℘ − 1 ( a 2 ) 2 + ℘ − 1 ( a 2 ) = a 2 \wp^{-1}(a_{2})^{2}+\wp^{-1}(a_{2})=a_{2} ℘ − 1 ( a 2 ) 2 + ℘ − 1 ( a 2 ) = a 2 . Moreover,
( ℘ − 1 ( a 2 ) a 1 a 2 − 1 ) 2 = ℘ − 1 ( a 2 ) ( a 1 a 2 ) − 1 + a 1 − 1 \left(\wp^{-1}(a_{2})\sqrt{a_{1}a_{2}}^{-1}\right)^{2}=\wp^{-1}(a_{2})(a_{1}a_{2})^{-1}+a_{1}^{-1} ( ℘ − 1 ( a 2 ) a 1 a 2 − 1 ) 2 = ℘ − 1 ( a 2 ) ( a 1 a 2 ) − 1 + a 1 − 1
with w ( ℘ − 1 ( a 2 ) ( a 1 a 2 ) − 1 ) = 1 2 w(\wp^{-1}(a_{2})(a_{1}a_{2})^{-1})=\frac{1}{2} w ( ℘ − 1 ( a 2 ) ( a 1 a 2 ) − 1 ) = 2 1 , hence the residue field L ‾ \overline{L} L contains a 1 \sqrt{a_{1}} a 1 . From the fundamental inequality
[ L : F ] ≥ [ L ‾ : F ‾ ] ( w ( L × ) : w ( F × ) ) [L:F]\geq[\overline{L}:\overline{F}]\bigl(w(L^{\times}):w(F^{\times})\bigr) [ L : F ] ≥ [ L : F ] ( w ( L × ) : w ( F × ) )
it then follows that w ( L × ) = 1 2 Z w(L^{\times})=\frac{1}{2}\mathbb{Z} w ( L × ) = 2 1 Z and L ‾ = k ( a 1 ) \overline{L}=k(\sqrt{a_{1}}) L = k ( a 1 ) , hence w w w is the unique extension of v v v to L L L . Therefore,
v ( N L / F ( ℓ 0 ( 0 ) ) ) = 4 w ( ℓ 0 ( 0 ) ) ∈ 2 Z . v\left(N_{L/F}\left(\ell_{0}(0)\right)\right)=4w(\ell_{0}(0))\in2\mathbb{Z}. v ( N L / F ( ℓ 0 ( 0 ) ) ) = 4 w ( ℓ 0 ( 0 )) ∈ 2 Z .
Equation (25) then yields v ( a 2 ) ∈ 2 Z v(a_{2})\in2\mathbb{Z} v ( a 2 ) ∈ 2 Z , a contradiction. □
Nontriviality of the invariant of biquaternion algebras
Using (23) to substitute [ a 1 a 2 , a 1 + a 2 ) K [a_{1}a_{2},a_{1}+a_{2})_{K} [ a 1 a 2 , a 1 + a 2 ) K in (20) we see that A 2 A_{2} A 2 is Brauer-equivalent to the following biquaternion K 2 K_{2} K 2 -algebra:
B = [ a 1 , ( a 1 + a 1 a 2 t 1 ) ] K 2 ⊗ K 2 [ a 2 , ( a 2 + a 1 a 2 t 2 ) ] K 2 . B=[a_{1},(a_{1}+\sqrt{a_{1}a_{2}}t_{1})]_{K_{2}}\otimes_{K_{2}}[a_{2},(a_{2}+\sqrt{a_{1}a_{2}}t_{2})]_{K_{2}}. B = [ a 1 , ( a 1 + a 1 a 2 t 1 ) ] K 2 ⊗ K 2 [ a 2 , ( a 2 + a 1 a 2 t 2 ) ] K 2 .
Theorem 2.1 shows that the condition N K 2 / F 2 [ B ] = 0 N_{K_{2}/F_{2}}[B]=0 N K 2 / F 2 [ B ] = 0 is satisfied, because A 2 A_{2} A 2 has a descent to a tensor product of quaternion F 2 F_{2} F 2 -algebras. We may therefore consider the invariant δ K 2 / F 2 ( B ) \delta_{K_{2}/F_{2}}(B) δ K 2 / F 2 ( B ) .
Since L 2 L_{2} L 2 splits the second factor in the definition of B B B , we have [ B ] L 2 = [ Q ] L 2 [B]_{L_{2}}=[Q]_{L_{2}} [ B ] L 2 = [ Q ] L 2 , hence we may identify Q K 2 Q_{K_{2}} Q K 2 with the centralizer of L 2 L_{2} L 2 in B B B . Therefore, δ K 2 / F 2 ( B ) = Φ ( δ L 2 / F 2 ( Q L 2 ) ) \delta_{K_{2}/F_{2}}(B)=\Phi(\delta_{L_{2}/F_{2}}(Q_{L_{2}})) δ K 2 / F 2 ( B ) = Φ ( δ L 2 / F 2 ( Q L 2 )) , and the computation of δ L 1 / F 1 ( Q ) \delta_{L_{1}/F_{1}}(Q) δ L 1 / F 1 ( Q ) in §5.1 shows that δ K 2 / F 2 ( B ) \delta_{K_{2}/F_{2}}(B) δ K 2 / F 2 ( B ) is represented by the form ⟨ ⟨ t 1 , a 2 , a 1 ] ⟩ \langle\!\langle t_{1},a_{2},a_{1}]\!\rangle ⟨ ⟨ t 1 , a 2 , a 1 ] ⟩ .
Theorem 5.3. δ K 2 / F 2 ( B ) ≠ 0 \delta_{K_{2}/F_{2}}(B)\ne0 δ K 2 / F 2 ( B ) = 0 .
Proof. Since B B B is Brauer-equivalent to A 2 A_{2} A 2 , it follows from the definition of A 2 A_{2} A 2 in (20) that a form q B ∈ I q 2 ( K 2 ) q_{B}\in I_{q}^{2}(K_{2}) q B ∈ I q 2 ( K 2 ) such that e 2 ( q B ) = [ B ] e_{2}(q_{B})=[B] e 2 ( q B ) = [ B ] is given by
q B = ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ] ⟩ + ⟨ ⟨ t 1 , a 1 ] ⟩ + ⟨ ⟨ t 2 , a 2 ] ⟩ ) K 2 . q_{B}=\left(\langle\!\langle a_{1}+a_{2},a_{1}a_{2}]\!\rangle+\langle\!\langle t_{1},a_{1}]\!\rangle+\langle\!\langle t_{2},a_{2}]\!\rangle\right)_{K_{2}}. q B = ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ] ⟩ + ⟨ ⟨ t 1 , a 1 ] ⟩ + ⟨ ⟨ t 2 , a 2 ] ⟩ ) K 2 .
By Frobenius reciprocity, we have
( s K 2 ) ∗ ( ⟨ ⟨ K 2 × ⟩ ⟩ q B ) = ⟨ ⟨ N K 2 / F 2 ( K 2 × ) ⟩ ⟩ ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ] ⟩ + ⟨ ⟨ t 1 , a 1 ] ⟩ + ⟨ ⟨ t 2 , a 2 ] ⟩ ) . (s_{K_{2}})_{*}\left(\langle\!\langle K_{2}^{\times}\rangle\!\rangle q_{B}\right)=\langle\!\langle N_{K_{2}/F_{2}}(K_{2}^{\times})\rangle\!\rangle\left(\langle\!\langle a_{1}+a_{2},a_{1}a_{2}]\!\rangle+\langle\!\langle t_{1},a_{1}]\!\rangle+\langle\!\langle t_{2},a_{2}]\!\rangle\right). ( s K 2 ) ∗ ( ⟨ ⟨ K 2 × ⟩ ⟩ q B ) = ⟨ ⟨ N K 2 / F 2 ( K 2 × )⟩ ⟩ ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ] ⟩ + ⟨ ⟨ t 1 , a 1 ] ⟩ + ⟨ ⟨ t 2 , a 2 ] ⟩ ) .
If δ K 2 / F 2 ( B ) = 0 \delta_{K_{2}/F_{2}}(B)=0 δ K 2 / F 2 ( B ) = 0 , then there exists g ∈ K 2 × g\in K_{2}^{\times} g ∈ K 2 × such that
⟨ ⟨ t 1 , a 2 , a 1 ] ⟩ ≡ ⟨ ⟨ N K 2 / F 2 ( g ) ⟩ ⟩ ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ] ⟩ + ⟨ ⟨ t 1 , a 1 ] ⟩ + ⟨ ⟨ t 2 , a 2 ] ⟩ ) ( m o d I q 4 ( F 2 ) ) . (26) \langle\!\langle t_{1},a_{2},a_{1}]\!\rangle\equiv\langle\!\langle N_{K_{2}/F_{2}}(g)\rangle\!\rangle\left(\langle\!\langle a_{1}+a_{2},a_{1}a_{2}]\!\rangle+\langle\!\langle t_{1},a_{1}]\!\rangle+\langle\!\langle t_{2},a_{2}]\!\rangle\right)\pmod{I_{q}^{4}(F_{2})}. \tag*{(26)} ⟨ ⟨ t 1 , a 2 , a 1 ] ⟩ ≡ ⟨ ⟨ N K 2 / F 2 ( g )⟩ ⟩ ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ] ⟩ + ⟨ ⟨ t 1 , a 1 ] ⟩ + ⟨ ⟨ t 2 , a 2 ] ⟩ ) ( mod I q 4 ( F 2 )) . (26)
Our goal is to show that this equation leads to a contradiction.
Consider the t 2 t_{2} t 2 -adic valuation v 0 v_{0} v 0 on F 2 F_{2} F 2 with constant field F 1 F_{1} F 1 . Multiplying g g g by a suitable power of t 2 t_{2} t 2 , which does not change the square class of N K 2 / F 2 ( g ) N_{K_{2}/F_{2}}(g) N K 2 / F 2 ( g ) , we may assume that v 0 ( g ) = 0 v_{0}(g)=0 v 0 ( g ) = 0 , which means that g ( 0 ) ≠ 0 g(0)\ne0 g ( 0 ) = 0 when g g g is viewed in F 2 = F 1 ( t 2 ) F_{2}=F_{1}(t_{2}) F 2 = F 1 ( t 2 ) . Since a 1 a_{1} a 1 , a 2 a_{2} a 2 and a 1 a 2 a_{1}a_{2} a 1 a 2 are constants, all the terms in (26) are in the tame subgroup I q ( F 2 ) 0 I_{q}(F_{2})_{0} I q ( F 2 ) 0 , hence the equation is modulo I q 4 ( F 2 ) ∩ I q ( F 2 ) 0 = I q 4 ( F 2 ) 0 I_{q}^{4}(F_{2})\cap I_{q}(F_{2})_{0}=I_{q}^{4}(F_{2})_{0} I q 4 ( F 2 ) ∩ I q ( F 2 ) 0 = I q 4 ( F 2 ) 0 . Applying to each side of (26) the residue map ∂ q 1 \partial_{q_{1}} ∂ q 1 , we obtain
0 ≡ ⟨ ⟨ N K 1 / F 1 ( g ( 0 ) ) , a 2 ] ⟩ ( m o d I q 3 ( F 1 ) ) , 0\equiv\langle\!\langle N_{K_{1}/F_{1}}(g(0)),a_{2}]\!\rangle\pmod{I_{q}^{3}(F_{1})}, 0 ≡ ⟨ ⟨ N K 1 / F 1 ( g ( 0 )) , a 2 ] ⟩ ( mod I q 3 ( F 1 )) ,
hence the quaternion algebra [ a 2 , N K 1 / F 1 ( g ( 0 ) ) ) F 1 [a_{2},N_{K_{1}/F_{1}}(g(0)) )_{F_{1}} [ a 2 , N K 1 / F 1 ( g ( 0 )) ) F 1 is split. By Proposition 2.2, we may find c ∈ F 1 × c\in F_{1}^{\times} c ∈ F 1 × such that [ a 2 , g ( 0 ) ) K 1 ≃ [ a 2 , c ) F 1 ⊗ F 1 K 1 [a_{2},g(0))_{K_{1}}\simeq[a_{2},c)_{F_{1}}\otimes_{F_{1}}K_{1} [ a 2 , g ( 0 ) ) K 1 ≃ [ a 2 , c ) F 1 ⊗ F 1 K 1 . The quaternion algebra [ a 2 , g ( 0 ) c − 1 ) K 1 [a_{2},g(0)c^{-1})_{K_{1}} [ a 2 , g ( 0 ) c − 1 ) K 1 is then split, so there exists ℓ ∈ L 1 × \ell\in L_1^\times ℓ ∈ L 1 × such that g ( 0 ) c − 1 = N L 1 / K 1 ( ℓ ) g(0)c^{-1}=N_{L_1/K_1}(\ell) g ( 0 ) c − 1 = N L 1 / K 1 ( ℓ ) . It follows that N K 1 / F 1 ( g ( 0 ) ) = c 2 N L 1 / F 1 ( ℓ ) N_{K_1/F_1}(g(0))=c^2N_{L_1/F_1}(\ell) N K 1 / F 1 ( g ( 0 )) = c 2 N L 1 / F 1 ( ℓ ) , hence
N K 1 / F 1 ( g ( 0 ) ) ≡ N L 1 / F 1 ( ℓ ) m o d F 1 × 2 . (27) N_{K_1/F_1}(g(0)) \equiv N_{L_1/F_1}(\ell) \mathbin{\mathrm{mod}} F_1^{\times2}. \tag*{(27)} N K 1 / F 1 ( g ( 0 )) ≡ N L 1 / F 1 ( ℓ ) mod F 1 × 2 . (27)
On the other hand, applying ∂ q + \partial_{q+} ∂ q + to each side of (26) yields
⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ ≡ ⟨ ⟨ N K 1 / F 1 ( g ( 0 ) ) ⟩ ⟩ ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ⟩ ⟩ + ⟨ ⟨ t 1 , a 1 ⟩ ⟩ ) m o d I q 4 ( F 1 ) . \langle\langle t_1,a_2,a_1\rangle\rangle\equiv\langle\langle N_{K_1/F_1}(g(0))\rangle\rangle\left(\langle\langle a_1+a_2,a_1a_2\rangle\rangle+\langle\langle t_1,a_1\rangle\rangle\right) \mathbin{\mathrm{mod}} I_q^4(F_1). ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ ≡ ⟨⟨ N K 1 / F 1 ( g ( 0 ))⟩⟩ ( ⟨⟨ a 1 + a 2 , a 1 a 2 ⟩⟩ + ⟨⟨ t 1 , a 1 ⟩⟩ ) mod I q 4 ( F 1 ) .
In view of (27), we may substitute N L 1 / F 1 ( ℓ ) N_{L_1/F_1}(\ell) N L 1 / F 1 ( ℓ ) for N K 1 / F 1 ( g ( 0 ) ) N_{K_1/F_1}(g(0)) N K 1 / F 1 ( g ( 0 )) ; we then see that (24) holds, a contradiction to Theorem 5.2 . □
Remark 5.4. Consider the biquaternion K 1 K_1 K 1 -algebra
B ′ = [ a 1 a 2 , a 1 + a 2 ) F 1 ⊗ F 1 [ a 1 , t 1 ) F 1 ⊗ F 1 K 1 . B'=[a_1a_2,a_1+a_2)_{F_1}\otimes_{F_1}[a_1,t_1)_{F_1}\otimes_{F_1}K_1. B ′ = [ a 1 a 2 , a 1 + a 2 ) F 1 ⊗ F 1 [ a 1 , t 1 ) F 1 ⊗ F 1 K 1 .
Using (23), we see that
B ′ ≃ [ a 1 , ( a 1 + a 1 a 2 ) t 1 ) K 1 ⊗ K 1 [ a 2 , a 2 + a 1 a 2 ) K 1 , B'\simeq[a_1,(a_1+\sqrt{a_1a_2})t_1)_{K_1}\otimes_{K_1}[a_2,a_2+\sqrt{a_1a_2})_{K_1}, B ′ ≃ [ a 1 , ( a 1 + a 1 a 2 ) t 1 ) K 1 ⊗ K 1 [ a 2 , a 2 + a 1 a 2 ) K 1 ,
hence [ B ′ ] L 1 = [ Q ] [B']_{L_1}=[Q] [ B ′ ] L 1 = [ Q ] . We may therefore identify Q Q Q with the centralizer of L 1 L_1 L 1 in B ′ B' B ′ , and from (16) we obtain
Φ ( δ L 1 / F 1 ( Q ) ) = δ K 1 / F 1 ( B ′ ) . \Phi\left(\delta_{L_1/F_1}(Q)\right)=\delta_{K_1/F_1}(B'). Φ ( δ L 1 / F 1 ( Q ) ) = δ K 1 / F 1 ( B ′ ) .
Now, δ L 1 / F 1 ( Q ) ≠ 0 \delta_{L_1/F_1}(Q)\ne0 δ L 1 / F 1 ( Q ) = 0 by Theorem 5.2 , but δ K 1 / F 1 ( B ′ ) = 0 \delta_{K_1/F_1}(B')=0 δ K 1 / F 1 ( B ′ ) = 0 by Proposition 3.7 since B ′ B' B ′ has a descent to F 1 F_1 F 1 .
Nontriviality of the Δ \Delta Δ -invariant
From (23), it follows that all the quaternion algebras in the definition of A 3 A_3 A 3 in (21) are split by the triquadratic extension F 3 ( ℘ − 1 ( a 1 ) , ℘ − 1 ( a 2 ) , a 1 a 2 ) F_3(\wp^{-1}(a_1),\wp^{-1}(a_2),\sqrt{a_1a_2}) F 3 ( ℘ − 1 ( a 1 ) , ℘ − 1 ( a 2 ) , a 1 a 2 ) of F 3 F_3 F 3 . Therefore, A 3 A_3 A 3 is Brauer-equivalent to a central simple F 3 F_3 F 3 -algebra A A A of degree 8, and the invariant Δ ( A ) \Delta(A) Δ ( A ) is defined. A quadratic form q A ∈ I q 2 ( F 3 ) q_A\in I_q^2(F_3) q A ∈ I q 2 ( F 3 ) such that e 2 ( q A ) = [ A ] e_2(q_A)=[A] e 2 ( q A ) = [ A ] is given by
q A = ⟨ ⟨ a 1 + a 2 , a 1 a 2 ⟩ ⟩ + ⟨ ⟨ t 1 , a 1 ⟩ ⟩ + ⟨ ⟨ t 2 , a 2 ⟩ ⟩ + ⟨ ⟨ a 1 a 2 , t 3 ⟩ ⟩ . (28) q_A=\langle\langle a_1+a_2,a_1a_2\rangle\rangle+\langle\langle t_1,a_1\rangle\rangle+\langle\langle t_2,a_2\rangle\rangle+\langle\langle a_1a_2,t_3\rangle\rangle. \tag*{(28)} q A = ⟨⟨ a 1 + a 2 , a 1 a 2 ⟩⟩ + ⟨⟨ t 1 , a 1 ⟩⟩ + ⟨⟨ t 2 , a 2 ⟩⟩ + ⟨⟨ a 1 a 2 , t 3 ⟩⟩ . (28)
Extending scalars from F 3 F_3 F 3 to K 3 K_3 K 3 splits the last quaternion factor on the right side of (21), hence
[ A ] K 3 = [ A 3 ] K 3 = [ A 2 ] K 3 = [ B ] K 3 . [A]_{K_3}=[A_3]_{K_3}=[A_2]_{K_3}=[B]_{K_3}. [ A ] K 3 = [ A 3 ] K 3 = [ A 2 ] K 3 = [ B ] K 3 .
Therefore, we may identify B K 3 B_{K_3} B K 3 with the centralizer of K 3 K_3 K 3 in A A A . It follows that Δ ( A ) = Ψ ( δ K 3 / F 3 ( B K 3 ) ) \Delta(A)=\Psi(\delta_{K_3/F_3}(B_{K_3})) Δ ( A ) = Ψ ( δ K 3 / F 3 ( B K 3 )) , and the computation of δ K 2 / F 2 ( B ) \delta_{K_2/F_2}(B) δ K 2 / F 2 ( B ) in § 4.2 shows that Δ ( A ) \Delta(A) Δ ( A ) is represented by ⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ \langle\langle t_1,a_2,a_1\rangle\rangle ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ .
The proof below that Δ ( A ) ≠ 0 \Delta(A)\ne0 Δ ( A ) = 0 follows the same pattern as the proofs of Theorem 5.2 and Theorem 5.3, using the t 3 − 1 t_3^{-1} t 3 − 1 -adic valuation v ∞ v_\infty v ∞ on F 3 = F 2 ( t 3 ) F_3=F_2(t_3) F 3 = F 2 ( t 3 ) with constant field F 2 F_2 F 2 . For f ∈ F 3 f\in F_3 f ∈ F 3 with v ∞ ( f ) ≥ 0 v_\infty(f)\ge0 v ∞ ( f ) ≥ 0 we write f ( ∞ ) f(\infty) f ( ∞ ) for the image of f f f in the residue field F 2 F_2 F 2 . In contrast with § 4.1 and § 4.2 , the use of residue maps is more delicate because the form q A q_A q A does not belong to the tame subgroup I q ( F 3 ) 0 I_q(F_3)_0 I q ( F 3 ) 0 . Since v ∞ ( 1 ) + v ∞ ( t 3 ) = − 1 v_\infty(1)+v_\infty(t_3)=-1 v ∞ ( 1 ) + v ∞ ( t 3 ) = − 1 , the form [ 1 , t 3 ] [1,t_3] [ 1 , t 3 ] lies in the first level I q ( F 3 ) 1 I_q(F_3)_1 I q ( F 3 ) 1 of Arason’s filtration [3 Th. 3] , hence also ⟨ ⟨ a 1 a 2 , t 3 ⟩ ⟩ ∈ I q ( F 3 ) 1 \langle\langle a_1a_2,t_3\rangle\rangle\in I_q(F_3)_1 ⟨⟨ a 1 a 2 , t 3 ⟩⟩ ∈ I q ( F 3 ) 1 because I q ( F 3 ) 1 I_q(F_3)_1 I q ( F 3 ) 1 is a module over W ( F 3 ) W(F_3) W ( F 3 ) . To prove that Δ ( A ) ≠ 0 \Delta(A)\ne0 Δ ( A ) = 0 , we need the following lemma:
Lemma 5.5. If f ∈ F 3 × f\in F_3^\times f ∈ F 3 × is such that ⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ ∈ I q 3 ( F 3 ) 0 + I q 4 ( F 3 ) \langle\langle f,a_1a_2,t_3\rangle\rangle\in I_q^3(F_3)_0+I_q^4(F_3) ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ ∈ I q 3 ( F 3 ) 0 + I q 4 ( F 3 ) , then there exists ψ ∈ I q 3 ( F 3 ) 0 \psi\in I_q^3(F_3)_0 ψ ∈ I q 3 ( F 3 ) 0 such that
⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ ≡ ψ m o d I q 4 ( F 3 ) and ∂ q + ( ψ ) = 0. \langle\langle f,a_1a_2,t_3\rangle\rangle\equiv\psi\mathbin{\mathrm{mod}} I_q^4(F_3)
\qquad\text{and}\qquad
\partial_{q+}(\psi)=0. ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ ≡ ψ mod I q 4 ( F 3 ) and ∂ q + ( ψ ) = 0. Moreover, if v ∞ ( f ) = n v_\infty(f)=n v ∞ ( f ) = n , then f 0 ( ∞ ) ∈ N K 2 / F 2 ( K 2 × ) f_0(\infty)\in N_{K_2/F_2}(K_2^\times) f 0 ( ∞ ) ∈ N K 2 / F 2 ( K 2 × ) for f 0 = f t 3 n f_0=ft_3^n f 0 = f t 3 n .
Proof. Since ⟨ ⟨ f , t 3 ⟩ ⟩ = ⟨ ⟨ f t 3 n , t 3 ⟩ ⟩ \langle\langle f,t_3\rangle\rangle=\langle\langle ft_3^n,t_3\rangle\rangle ⟨⟨ f , t 3 ⟩⟩ = ⟨⟨ f t 3 n , t 3 ⟩⟩ , we may substitute f 0 f_0 f 0 for f f f and we are thus reduced to consider the case where v ∞ ( f ) = 0 v_\infty(f)=0 v ∞ ( f ) = 0 . We assume this for the rest of the proof.
Since by definition the module I q ( F 3 ) 1 I_q(F_3)_1 I q ( F 3 ) 1 contains I q ( F 3 ) 0 I_q(F_3)_0 I q ( F 3 ) 0 , the hypothesis implies that ⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ \langle\langle f,a_1a_2,t_3\rangle\rangle ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ lies in I q 3 ( F 3 ) 0 + ( I q 4 ( F 3 ) ∩ I q ( F 3 ) 1 ) I_q^3(F_3)_0+(I_q^4(F_3)\cap I_q(F_3)_1) I q 3 ( F 3 ) 0 + ( I q 4 ( F 3 ) ∩ I q ( F 3 ) 1 ) . In [4 ] , Arason shows that I q 4 ( F 3 ) ∩ I q ( F 3 ) 1 = I q 4 ( F 3 ) 1 : = I 3 ( F 3 ) ⋅ I q ( F 3 ) 1 I_q^4(F_3)\cap I_q(F_3)_1=I_q^4(F_3)_1:=I^3(F_3)\cdot I_q(F_3)_1 I q 4 ( F 3 ) ∩ I q ( F 3 ) 1 = I q 4 ( F 3 ) 1 := I 3 ( F 3 ) ⋅ I q ( F 3 ) 1 . Moreover, he describes on [4 p. 79] a homomorphism from I q 4 ( F 3 ) 1 I_q^4(F_3)_1 I q 4 ( F 3 ) 1 to the vector space Ω F 2 2 \Omega^2_{F_2} Ω F 2 2 of absolute 2 2 2 -differential forms over F 2 F_2 F 2 . This homomorphism vanishes on I q 3 ( F 3 ) 0 I_q^3(F_3)_0 I q 3 ( F 3 ) 0 and on I q 4 ( F 3 ) 1 I_q^4(F_3)_1 I q 4 ( F 3 ) 1 , and maps ⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ \langle\langle f,a_1a_2,t_3\rangle\rangle ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ to d f ( ∞ ) f ( ∞ ) ∧ d ( a 1 a 2 ) a 1 a 2 \frac{df(\infty)}{f(\infty)}\wedge\frac{d(a_1a_2)}{a_1a_2} f ( ∞ ) df ( ∞ ) ∧ a 1 a 2 d ( a 1 a 2 ) , hence
d f ( ∞ ) f ( ∞ ) ∧ d ( a 1 a 2 ) a 1 a 2 = 0 in Ω F 2 2 . \frac{df(\infty)}{f(\infty)}\wedge\frac{d(a_1a_2)}{a_1a_2}=0 \quad\text{in }\Omega^2_{F_2}. f ( ∞ ) df ( ∞ ) ∧ a 1 a 2 d ( a 1 a 2 ) = 0 in Ω F 2 2 .
This equation does not hold if 1 1 1 , f ( ∞ ) f(\infty) f ( ∞ ) , a 1 a 2 a_1a_2 a 1 a 2 and f ( ∞ ) a 1 a 2 f(\infty)a_1a_2 f ( ∞ ) a 1 a 2 are F 2 2 F_2^2 F 2 2 -linearly independent in F 2 F_2 F 2 , for then d f ( ∞ ) f ( ∞ ) \frac{df(\infty)}{f(\infty)} f ( ∞ ) df ( ∞ ) and d ( a 1 a 2 ) a 1 a 2 \frac{d(a_1a_2)}{a_1a_2} a 1 a 2 d ( a 1 a 2 ) are part of an F 2 F_2 F 2 -base of Ω F 2 1 \Omega^1_{F_2} Ω F 2 1 . Therefore, there exist x , y ∈ F 2 x,y\in F_2 x , y ∈ F 2 such that
f ( ∞ ) = x 2 + a 1 a 2 y 2 = N K 2 / F 2 ( x + y a 1 a 2 ) . f(\infty)=x^2+a_1a_2y^2=N_{K_2/F_2}(x+y\sqrt{a_1a_2}). f ( ∞ ) = x 2 + a 1 a 2 y 2 = N K 2 / F 2 ( x + y a 1 a 2 ) .
To complete the proof, we construct a tame form ψ \psi ψ with ψ ≡ ⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ ( m o d I q 4 ( F 3 ) ) \psi\equiv\langle\langle f,a_1a_2,t_3\rangle\rangle\pmod{I_q^4(F_3)} ψ ≡ ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ ( mod I q 4 ( F 3 )) .
Let s = t 3 ( f + x 2 + a 1 a 2 y 2 ) ∈ F 3 s=t_3(f+x^2+a_1a_2y^2)\in F_3 s = t 3 ( f + x 2 + a 1 a 2 y 2 ) ∈ F 3 , so v ∞ ( s ) ≥ 0 v_\infty(s)\geq0 v ∞ ( s ) ≥ 0 . Then
t 3 f − 1 = ( x 2 + a 1 a 2 y 2 ) − 1 ( t 3 + f − 1 s ) . t_3f^{-1}=(x^2+a_1a_2y^2)^{-1}(t_3+f^{-1}s). t 3 f − 1 = ( x 2 + a 1 a 2 y 2 ) − 1 ( t 3 + f − 1 s ) .
Since ⟨ ⟨ f , t 3 ⟩ ⟩ = ⟨ ⟨ t 3 f − 1 , t 3 ⟩ ⟩ \langle\langle f,t_3\rangle\rangle=\langle\langle t_3f^{-1},t_3\rangle\rangle ⟨⟨ f , t 3 ⟩⟩ = ⟨⟨ t 3 f − 1 , t 3 ⟩⟩ , it follows that
⟨ ⟨ f , t 3 ⟩ ⟩ ≡ ⟨ ⟨ x 2 + a 1 a 2 y 2 , t 3 ⟩ ⟩ + ⟨ ⟨ t 3 + f − 1 s , t 3 ⟩ ⟩ ( m o d I q 3 ( F 3 ) ) . \langle\langle f,t_3\rangle\rangle\equiv\langle\langle x^2+a_1a_2y^2,t_3\rangle\rangle+\langle\langle t_3+f^{-1}s,t_3\rangle\rangle\pmod{I_q^3(F_3)}. ⟨⟨ f , t 3 ⟩⟩ ≡ ⟨⟨ x 2 + a 1 a 2 y 2 , t 3 ⟩⟩ + ⟨⟨ t 3 + f − 1 s , t 3 ⟩⟩ ( mod I q 3 ( F 3 )) .
Multiplying by ⟨ ⟨ a 1 a 2 ⟩ ⟩ \langle\langle a_1a_2\rangle\rangle ⟨⟨ a 1 a 2 ⟩⟩ and taking into account that ⟨ x 2 + a 1 a 2 y 2 ⟩ ⟨ ⟨ a 1 a 2 ⟩ ⟩ = ⟨ ⟨ a 1 a 2 ⟩ ⟩ \langle x^2+a_1a_2y^2\rangle\langle\langle a_1a_2\rangle\rangle=\langle\langle a_1a_2\rangle\rangle ⟨ x 2 + a 1 a 2 y 2 ⟩ ⟨⟨ a 1 a 2 ⟩⟩ = ⟨⟨ a 1 a 2 ⟩⟩ , we obtain
⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ ≡ ⟨ ⟨ a 1 a 2 , t 3 + f − 1 s , t 3 ⟩ ⟩ ( m o d I q 4 ( F 3 ) ) . \langle\langle f,a_1a_2,t_3\rangle\rangle\equiv\langle\langle a_1a_2,t_3+f^{-1}s,t_3\rangle\rangle\pmod{I_q^4(F_3)}. ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ ≡ ⟨⟨ a 1 a 2 , t 3 + f − 1 s , t 3 ⟩⟩ ( mod I q 4 ( F 3 )) .
Now, use ⟨ ⟨ t 3 + f − 1 s , t 3 + f − 1 s ⟩ ⟩ = 0 \langle\langle t_3+f^{-1}s,t_3+f^{-1}s\rangle\rangle=0 ⟨⟨ t 3 + f − 1 s , t 3 + f − 1 s ⟩⟩ = 0 to rewrite the right side:
⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ ≡ ⟨ ⟨ a 1 a 2 , t 3 + f − 1 s , f − 1 s ⟩ ⟩ ( m o d I q 4 ( F 3 ) ) . \langle\langle f,a_1a_2,t_3\rangle\rangle\equiv\langle\langle a_1a_2,t_3+f^{-1}s,f^{-1}s\rangle\rangle\pmod{I_q^4(F_3)}. ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ ≡ ⟨⟨ a 1 a 2 , t 3 + f − 1 s , f − 1 s ⟩⟩ ( mod I q 4 ( F 3 )) .
Letting ψ = ⟨ ⟨ a 1 a 2 , t 3 + f − 1 s , f − 1 s ⟩ ⟩ = ⟨ ⟨ a 1 a 2 , t 3 ( 1 + f − 1 s t 3 − 1 ) , f − 1 s ⟩ ⟩ \psi=\langle\langle a_1a_2,t_3+f^{-1}s,f^{-1}s\rangle\rangle=\langle\langle a_1a_2,t_3(1+f^{-1}st_3^{-1}),f^{-1}s\rangle\rangle ψ = ⟨⟨ a 1 a 2 , t 3 + f − 1 s , f − 1 s ⟩⟩ = ⟨⟨ a 1 a 2 , t 3 ( 1 + f − 1 s t 3 − 1 ) , f − 1 s ⟩⟩ , we have ψ ∈ I q 3 ( F 3 ) 0 \psi\in I_q^3(F_3)_0 ψ ∈ I q 3 ( F 3 ) 0 and ∂ q + ( ψ ) = 0 \partial_{q+}(\psi)=0 ∂ q + ( ψ ) = 0 because ( f − 1 s t 3 − 1 ) ( ∞ ) = 0 (f^{-1}st_3^{-1})(\infty)=0 ( f − 1 s t 3 − 1 ) ( ∞ ) = 0 . □
Theorem 5.6. Δ ( A ) ≠ 0 \Delta(A)\ne0 Δ ( A ) = 0 .
Proof. As observed at the beginning of this subsection, Δ ( A ) \Delta(A) Δ ( A ) is represented by ⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ \langle\langle t_1,a_2,a_1\rangle\rangle ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ . Therefore, Δ ( A ) = 0 \Delta(A)=0 Δ ( A ) = 0 means that there exists f ∈ F 3 × f\in F_3^\times f ∈ F 3 × such that
⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ ≡ ⟨ ⟨ f ⟩ ⟩ q A ( m o d I q 4 ( F 3 ) ) . (29) \langle\langle t_1,a_2,a_1\rangle\rangle\equiv\langle\langle f\rangle\rangle_{q_A}\pmod{I_q^4(F_3)}. \tag*{(29)} ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ ≡ ⟨⟨ f ⟩ ⟩ q A ( mod I q 4 ( F 3 )) . (29)
In view of the choice of q A q_A q A in (28), this equation implies
⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ ≡ ⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ − ⟨ ⟨ f , a 1 + a 2 , a 1 a 2 ⟩ ⟩ − ⟨ ⟨ f , t 1 , a 1 ⟩ ⟩ − ⟨ ⟨ f , t 2 , a 2 ⟩ ⟩ ( m o d I q 4 ( F 3 ) ) . \langle\langle f,a_1a_2,t_3\rangle\rangle\equiv\langle\langle t_1,a_2,a_1\rangle\rangle-\langle\langle f,a_1+a_2,a_1a_2\rangle\rangle-\langle\langle f,t_1,a_1\rangle\rangle-\langle\langle f,t_2,a_2\rangle\rangle\pmod{I_q^4(F_3)}. ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ ≡ ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ − ⟨⟨ f , a 1 + a 2 , a 1 a 2 ⟩⟩ − ⟨⟨ f , t 1 , a 1 ⟩⟩ − ⟨⟨ f , t 2 , a 2 ⟩⟩ ( mod I q 4 ( F 3 )) .
Since all the quadratic forms on the right side are in I q ( F 3 ) 0 ∩ I q 3 ( F 3 ) = I q 3 ( F 3 ) 0 I_q(F_3)_0\cap I_q^3(F_3)=I_q^3(F_3)_0 I q ( F 3 ) 0 ∩ I q 3 ( F 3 ) = I q 3 ( F 3 ) 0 , it follows that ⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ \langle\langle f,a_1a_2,t_3\rangle\rangle ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ belongs to I q 3 ( F 3 ) 0 + I q 4 ( F 3 ) I_q^3(F_3)_0+I_q^4(F_3) I q 3 ( F 3 ) 0 + I q 4 ( F 3 ) . Lemma 5.5 then yields ψ ∈ I q 3 ( F 3 ) 0 \psi\in I_q^3(F_3)_0 ψ ∈ I q 3 ( F 3 ) 0 such that ⟨ ⟨ f , a 1 a 2 , t 3 ⟩ ⟩ ≡ ψ ( m o d I q 4 ( F 3 ) ) \langle\langle f,a_1a_2,t_3\rangle\rangle\equiv\psi\pmod{I_q^4(F_3)} ⟨⟨ f , a 1 a 2 , t 3 ⟩⟩ ≡ ψ ( mod I q 4 ( F 3 )) and ∂ q + ( ψ ) = 0 \partial_{q+}(\psi)=0 ∂ q + ( ψ ) = 0 . Substituting in (29), we obtain
⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ ≡ ⟨ ⟨ f ⟩ ⟩ ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ⟩ ⟩ + ⟨ ⟨ t 1 , a 1 ⟩ ⟩ + ⟨ ⟨ t 2 , a 2 ⟩ ⟩ ) + ψ ( m o d I q 4 ( F 3 ) ) . \langle\langle t_1,a_2,a_1\rangle\rangle\equiv\langle\langle f\rangle\rangle\bigl(\langle\langle a_1+a_2,a_1a_2\rangle\rangle+\langle\langle t_1,a_1\rangle\rangle+\langle\langle t_2,a_2\rangle\rangle\bigr)+\psi\pmod{I_q^4(F_3)}. ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ ≡ ⟨⟨ f ⟩⟩ ( ⟨⟨ a 1 + a 2 , a 1 a 2 ⟩⟩ + ⟨⟨ t 1 , a 1 ⟩⟩ + ⟨⟨ t 2 , a 2 ⟩⟩ ) + ψ ( mod I q 4 ( F 3 )) .
Letting f 0 = f t 3 v ∞ ( f ) f_0=ft_3^{v_\infty(f)} f 0 = f t 3 v ∞ ( f ) and applying ∂ q + \partial_{q+} ∂ q + to each side of the last displayed equation yields
⟨ ⟨ t 1 , a 2 , a 1 ⟩ ⟩ ≡ ⟨ ⟨ f 0 ( ∞ ) ⟩ ⟩ ( ⟨ ⟨ a 1 + a 2 , a 1 a 2 ⟩ ⟩ + ⟨ ⟨ t 1 , a 1 ⟩ ⟩ + ⟨ ⟨ t 2 , a 2 ⟩ ⟩ ) ( m o d I q 4 ( F 2 ) ) , (30) \langle\langle t_1,a_2,a_1\rangle\rangle\equiv\langle\langle f_0(\infty)\rangle\rangle\bigl(\langle\langle a_1+a_2,a_1a_2\rangle\rangle+\langle\langle t_1,a_1\rangle\rangle+\langle\langle t_2,a_2\rangle\rangle\bigr)\pmod{I_q^4(F_2)}, \tag*{(30)} ⟨⟨ t 1 , a 2 , a 1 ⟩⟩ ≡ ⟨⟨ f 0 ( ∞ )⟩⟩ ( ⟨⟨ a 1 + a 2 , a 1 a 2 ⟩⟩ + ⟨⟨ t 1 , a 1 ⟩⟩ + ⟨⟨ t 2 , a 2 ⟩⟩ ) ( mod I q 4 ( F 2 )) , (30)
But Lemma 5.5 also yields f 0 ( ∞ ) = N K 2 / F 2 ( g ) f_0(\infty)=N_{K_2/F_2}(g) f 0 ( ∞ ) = N K 2 / F 2 ( g ) for some g ∈ K 2 × g\in K_2^\times g ∈ K 2 × , so (30) shows that (26) holds, contradicting Theorem 5.3. Thus, there is no f ∈ F 3 × f\in F_3^\times f ∈ F 3 × such that (29) holds, hence Δ ( A ) ≠ 0 \Delta(A)\ne0 Δ ( A ) = 0 .
Note that A A A is split by F 3 ( ℘ − 1 ( a 1 ) , ℘ − 1 ( a 2 ) , a 1 a 2 ) F_3(\wp^{-1}(a_1),\wp^{-1}(a_2),\sqrt{a_1a_2}) F 3 ( ℘ − 1 ( a 1 ) , ℘ − 1 ( a 2 ) , a 1 a 2 ) , and since Δ ( A ) ≠ 0 \Delta(A)\ne0 Δ ( A ) = 0 it does not decompose into a tensor product of quaternion algebras. Therefore it provides a counterexample to [5 Th. 4.1(iii)] , hence also to the “kernel splitting property” of the extension F 3 ( ℘ − 1 ( a 1 ) , ℘ − 1 ( a 2 ) , a 1 a 2 ) / F 3 F_3(\wp^{-1}(a_1),\wp^{-1}(a_2),\sqrt{a_1a_2})/F_3 F 3 ( ℘ − 1 ( a 1 ) , ℘ − 1 ( a 2 ) , a 1 a 2 ) / F 3 asserted in [6 Th. 1.2] , which is the sole ingredient in the proof of [5 Th. 4.1(iii)] .
Remark 5.7. Consider the F 2 F_2 F 2 -algebra
A ′ = [ a 1 a 2 , a 1 + a 2 ) F 2 ⊗ F 2 [ a 1 , a 2 t 1 ) F 2 ⊗ F 2 [ a 2 , t 2 ) F 2 . A'=[a_1a_2,a_1+a_2)_{F_2}\otimes_{F_2}[a_1,a_2t_1)_{F_2}\otimes_{F_2}[a_2,t_2)_{F_2}. A ′ = [ a 1 a 2 , a 1 + a 2 ) F 2 ⊗ F 2 [ a 1 , a 2 t 1 ) F 2 ⊗ F 2 [ a 2 , t 2 ) F 2 .
Extending scalars to K 2 K_2 K 2 splits [ a 1 , a 2 ) F 2 [a_1,a_2)_{F_2} [ a 1 , a 2 ) F 2 since [ a 1 , a 2 ) K 2 ≃ [ a 1 , a 1 ) K 2 [a_1,a_2)_{K_2}\simeq[a_1,a_1)_{K_2} [ a 1 , a 2 ) K 2 ≃ [ a 1 , a 1 ) K 2 , hence [ A ′ ] K 2 = [ B ] [A']_{K_2}=[B] [ A ′ ] K 2 = [ B ] and we may identify B B B with the centralizer of K 2 K_2 K 2 in A ′ A' A ′ . Then (16) yields
Ψ K 2 / F 2 ( δ K 2 / F 2 ( B ) ) = Δ ( A ′ ) . \Psi_{K_2/F_2}\left(\delta_{K_2/F_2}(B)\right)=\Delta(A'). Ψ K 2 / F 2 ( δ K 2 / F 2 ( B ) ) = Δ ( A ′ ) .
Now, δ K 2 / F 2 ( B ) ≠ 0 \delta_{K_2/F_2}(B)\ne0 δ K 2 / F 2 ( B ) = 0 by Theorem 5.3, but Δ ( A ′ ) = 0 \Delta(A')=0 Δ ( A ′ ) = 0 since A ′ A' A ′ is a tensor product of three quaternion algebras.
Quaternion algebras over purely inseparable biquadratic extensions
Purely inseparable biquadratic field extensions are excluded in Section 4. This appendix highlights how this case differs from those considered above. No invariant is needed here as a quaternion algebra always descends when its Brauer class does. More precisely, we have the following:
Theorem A.1. Let L L L be a purely inseparable biquadratic field extension (i.e., a purely inseparable quartic extension of exponent one) of an arbitrary field F F F . For a quaternion L L L -algebra Q Q Q , the following are equivalent:
(i) there exists a central simple F F F -algebra A A A such that 2 [ A ] = 0 2[A]=0 2 [ A ] = 0 and [ A ] L = [ Q ] [A]_L=[Q] [ A ] L = [ Q ] ;
(ii) there exists a quaternion F F F -algebra Q 0 Q_0 Q 0 such that Q ≃ Q 0 ⊗ F L Q\simeq Q_0\otimes_F L Q ≃ Q 0 ⊗ F L .
Moreover, when these conditions hold, for every central simple F F F -algebra A A A of degree 8 and exponent 2 containing Q Q Q , the following hold:
(a) A ≃ Q 0 ⊗ F Q 1 ⊗ F Q 2 A\simeq Q_0\otimes_F Q_1\otimes_F Q_2 A ≃ Q 0 ⊗ F Q 1 ⊗ F Q 2 for some quaternion F F F -algebras Q 0 Q_0 Q 0 , Q 1 Q_1 Q 1 , Q 2 Q_2 Q 2 such that L ⊂ Q 1 ⊗ F Q 2 L\subset Q_1\otimes_F Q_2 L ⊂ Q 1 ⊗ F Q 2 ;
(b) π ( σ ) \pi(\sigma) π ( σ ) is hyperbolic for every involution σ \sigma σ on A A A that extends the conjugation on Q Q Q .
The proof uses the following elementary ring-theoretic observation:
Lemma A.2. Let E E E be a central simple algebra over a field Z Z Z of characteristic 2 and let θ \theta θ be an involution (of any type) on E E E . Let also Z ( r ) ⊂ E Z(r)\subset E Z ( r ) ⊂ E be an inseparable quadratic extension of Z Z Z . If r = u + θ ( u ) r=u+\theta(u) r = u + θ ( u ) for some u ∈ E u\in E u ∈ E that commutes with r r r , then there exists e ∈ Symd ( θ ) e\in\operatorname{Symd}(\theta) e ∈ Symd ( θ ) such that r = r e + e r r=re+er r = r e + er .
Proof. Consider the Z Z Z -linear map r ∗ : E → E r*:E\to E r ∗ : E → E that maps x ∈ E x\in E x ∈ E to r x + x r rx+xr r x + x r . Its kernel is the centralizer C C C of r r r in E E E , and ( r ∗ ) 2 = 0 (r*)^2=0 ( r ∗ ) 2 = 0 because r 2 ∈ Z × r^2\in Z^\times r 2 ∈ Z × . Therefore, the image of r ∗ r* r ∗ lies in C C C . But dim Z C = 1 2 dim Z E \dim_Z C=\frac{1}{2}\dim_Z E dim Z C = 2 1 dim Z E since Z Z Z is a field, hence dimension count shows that C C C is also the image of r ∗ r* r ∗ . In particular, u = r x + x r u=rx+xr u = r x + x r for some x ∈ E x\in E x ∈ E . Then θ ( u ) = θ ( x ) r + r θ ( x ) \theta(u)=\theta(x)r+r\theta(x) θ ( u ) = θ ( x ) r + r θ ( x ) , hence
r = u + θ ( u ) = r ( x + θ ( x ) ) + ( x + θ ( x ) ) r . r=u+\theta(u)=r(x+\theta(x))+(x+\theta(x))r. r = u + θ ( u ) = r ( x + θ ( x )) + ( x + θ ( x )) r .
This yields r = r e + e r r=re+er r = r e + er with e = x + θ ( x ) ∈ Symd ( θ ) e=x+\theta(x)\in\operatorname{Symd}(\theta) e = x + θ ( x ) ∈ Symd ( θ ) .
Proof of Theorem A.1. (ii) ⇒ \Rightarrow ⇒ (i) If (ii) holds, then (i) holds with A = Q 0 A = Q_0 A = Q 0 .
(i) ⇒ \Rightarrow ⇒ (ii) If A A A satisfies (i), then A A A is split by the quadratic extensions of L L L that split Q Q Q , hence up to Brauer-equivalence we may assume that deg A = 8 \deg A = 8 deg A = 8 and A A A contains Q Q Q .
Let σ \sigma σ be an involution on A A A that extends the conjugation γ \gamma γ on Q Q Q . We show below that (a) holds with quaternion F F F -algebras Q 0 Q_0 Q 0 , Q 1 Q_1 Q 1 , Q 2 ⊂ A Q_2 \subset A Q 2 ⊂ A stable under σ \sigma σ . Besides proving (a), this proves (ii) since Q Q Q is the centralizer of L L L in A A A , and also (b) by the main theorem in [11 ] or by [27 Prop. 8] .
Pick r 1 r_1 r 1 , r 2 ∈ L r_2 \in L r 2 ∈ L such that L = F ( r 1 , r 2 ) L = F(r_1, r_2) L = F ( r 1 , r 2 ) and let B ⊂ A B \subset A B ⊂ A be the centralizer of r 2 r_2 r 2 . Thus,
F ⊂ F ( r 2 ) ⊂ L ⊂ Q ⊂ B ⊂ A . F \subset F(r_2) \subset L \subset Q \subset B \subset A. F ⊂ F ( r 2 ) ⊂ L ⊂ Q ⊂ B ⊂ A .
Let also σ B \sigma_B σ B be the restriction of σ \sigma σ to B B B , so σ B \sigma_B σ B also restricts to γ \gamma γ on Q Q Q . Since L = Symd ( γ ) L = \operatorname{Symd}(\gamma) L = Symd ( γ ) , there exists u 1 ∈ Q u_1 \in Q u 1 ∈ Q such that r 1 = u 1 + σ B ( u 1 ) r_1 = u_1 + \sigma_B(u_1) r 1 = u 1 + σ B ( u 1 ) . Lemma A.2 yields e 1 ∈ Symd ( σ B ) e_1 \in\operatorname{Symd}(\sigma_B) e 1 ∈ Symd ( σ B ) such that
r 1 = r 1 e 1 + e 1 r 1 , r_1 = r_1 e_1 + e_1 r_1, r 1 = r 1 e 1 + e 1 r 1 ,
hence
r 1 e 1 r 1 − 1 = e 1 + 1. r_1 e_1 r_1^{-1} = e_1 + 1. r 1 e 1 r 1 − 1 = e 1 + 1.
Then r 1 e 1 2 k r 1 − 1 = e 1 2 k + 1 r_1 e_1^{2^k} r_1^{-1} = e_1^{2^k} + 1 r 1 e 1 2 k r 1 − 1 = e 1 2 k + 1 for all k ≥ 0 k \ge0 k ≥ 0 ; in particular e 1 2 + e 1 e_1^2 + e_1 e 1 2 + e 1 commutes with r 1 r_1 r 1 . Since e 1 ∈ B e_1 \in B e 1 ∈ B , e 1 2 + e 1 e_1^2 + e_1 e 1 2 + e 1 also commutes with r 2 r_2 r 2 , and it follows that e 1 2 + e 1 ∈ Q e_1^2 + e_1 \in Q e 1 2 + e 1 ∈ Q . Moreover, e 1 2 + e 1 ∈ Sym ( γ ) e_1^2 + e_1 \in\operatorname{Sym}(\gamma) e 1 2 + e 1 ∈ Sym ( γ ) because e 1 ∈ Symd ( σ B ) e_1 \in\operatorname{Symd}(\sigma_B) e 1 ∈ Symd ( σ B ) . Therefore, ( e 1 2 + e 1 ) 2 ∈ L (e_1^2 + e_1)^2 \in L ( e 1 2 + e 1 ) 2 ∈ L , hence ( e 1 2 + e 1 ) 4 ∈ F (e_1^2 + e_1)^4 \in F ( e 1 2 + e 1 ) 4 ∈ F . Set i 1 = e 1 4 i_1 = e_1^4 i 1 = e 1 4 . Then i 1 2 + i 1 ∈ F i_1^2 + i_1 \in F i 1 2 + i 1 ∈ F and r 1 i 1 r 1 − 1 = i 1 + 1 r_1 i_1 r_1^{-1} = i_1 + 1 r 1 i 1 r 1 − 1 = i 1 + 1 , hence i 1 i_1 i 1 and r 1 r_1 r 1 generate a quaternion F F F -algebra Q 1 ⊂ A Q_1 \subset A Q 1 ⊂ A , which is stable under σ \sigma σ since i 1 i_1 i 1 , r 1 ∈ Sym ( σ ) r_1 \in\operatorname{Sym}(\sigma) r 1 ∈ Sym ( σ ) .
Let A ′ ⊂ A A' \subset A A ′ ⊂ A be the centralizer of Q 1 Q_1 Q 1 , and write σ 1 \sigma_1 σ 1 , σ ′ \sigma' σ ′ for the restrictions of σ \sigma σ to Q 1 Q_1 Q 1 and A ′ A' A ′ respectively. Under the map Q 1 × A ′ → A Q_1 \times A' \to A Q 1 × A ′ → A defined by multiplication in A A A , we thus have (31)
( Q 1 , σ 1 ) ⊗ F ( A ′ , σ ′ ) ≃ ( A , σ ) . (31) (Q_1,\sigma_1) \otimes_F (A',\sigma') \simeq(A,\sigma). \tag*{(31)} ( Q 1 , σ 1 ) ⊗ F ( A ′ , σ ′ ) ≃ ( A , σ ) . (31)
The involution σ 1 \sigma_1 σ 1 is orthogonal because i 1 ∈ Sym ( σ 1 ) i_1 \in\operatorname{Sym}(\sigma_1) i 1 ∈ Sym ( σ 1 ) while the reduced trace of i 1 i_1 i 1 in Q 1 Q_1 Q 1 is 1. Therefore, by [19 (2.23)] the involution σ ′ \sigma' σ ′ is symplectic.
Note that r 2 r_2 r 2 commutes with r 1 r_1 r 1 , but also with i 1 i_1 i 1 , since the latter is in B B B ; therefore, r 2 ∈ A ′ r_2 \in A' r 2 ∈ A ′ .
Just as for r 1 r_1 r 1 above, we may find u 2 ∈ Q u_2 \in Q u 2 ∈ Q such that r 2 = u 2 + σ ( u 2 ) r_2 = u_2 + \sigma(u_2) r 2 = u 2 + σ ( u 2 ) . Using (31) as an identification, write
u 2 = 1 ⊗ a 0 + r 1 ⊗ a 1 + i 1 ⊗ a 2 + i 1 r 1 ⊗ a 3 with a 0 , a 1 , a 2 , a 3 ∈ A ′ . u_2 = 1 \otimes a_0 + r_1 \otimes a_1 + i_1 \otimes a_2 + i_1 r_1 \otimes a_3
\qquad\text{with } a_0,a_1,a_2,a_3 \in A'. u 2 = 1 ⊗ a 0 + r 1 ⊗ a 1 + i 1 ⊗ a 2 + i 1 r 1 ⊗ a 3 with a 0 , a 1 , a 2 , a 3 ∈ A ′ .
Since u 2 u_2 u 2 commutes with r 1 r_1 r 1 , we have a 2 = a 3 = 0 a_2 = a_3 = 0 a 2 = a 3 = 0 . Moreover, a 0 a_0 a 0 and a 1 a_1 a 1 commute with r 2 r_2 r 2 because u 2 ∈ Q u_2 \in Q u 2 ∈ Q commutes with r 2 ∈ L r_2 \in L r 2 ∈ L . Now,
r 2 = 1 ⊗ r 2 = u 2 + σ ( u 2 ) = 1 ⊗ a 0 + r 1 ⊗ a 1 + 1 ⊗ σ ′ ( a 0 ) + r 1 ⊗ σ ′ ( a 1 ) , r_2 = 1 \otimes r_2 = u_2 + \sigma(u_2) = 1 \otimes a_0 + r_1 \otimes a_1 + 1 \otimes\sigma'(a_0) + r_1 \otimes\sigma'(a_1), r 2 = 1 ⊗ r 2 = u 2 + σ ( u 2 ) = 1 ⊗ a 0 + r 1 ⊗ a 1 + 1 ⊗ σ ′ ( a 0 ) + r 1 ⊗ σ ′ ( a 1 ) ,
hence r 2 = a 0 + σ ′ ( a 0 ) r_2 = a_0 + \sigma'(a_0) r 2 = a 0 + σ ′ ( a 0 ) . Lemma A.2 then yields e 2 ∈ Symd ( σ ′ ) e_2 \in\operatorname{Symd}(\sigma') e 2 ∈ Symd ( σ ′ ) such that
r 2 = r 2 e 2 + e 2 r 2 . r_2 = r_2 e_2 + e_2 r_2. r 2 = r 2 e 2 + e 2 r 2 .
As above, e 2 2 + e 2 e_2^2 + e_2 e 2 2 + e 2 commutes with r 2 r_2 r 2 ; since r 2 ≠ 0 r_2 \ne0 r 2 = 0 , e 2 e_2 e 2 does not commute with r 2 r_2 r 2 and it follows that F ( e 2 2 + e 2 ) ⊊ F ( e 2 ) F(e_2^2 + e_2) \subsetneq F(e_2) F ( e 2 2 + e 2 ) ⊊ F ( e 2 ) . But since σ ′ \sigma' σ ′ is a symplectic involution on a biquaternion algebra, the minimal polynomial of e 2 e_2 e 2 over F F F has degree at most 2, hence e 2 2 + e 2 ∈ F e_2^2 + e_2 \in F e 2 2 + e 2 ∈ F . Therefore, e 2 e_2 e 2 and r 2 r_2 r 2 generate a quaternion F F F -algebra Q 2 ⊂ A ′ Q_2 \subset A' Q 2 ⊂ A ′ , which is stable under σ ′ \sigma' σ ′ since e 2 e_2 e 2 , r 2 ∈ Sym ( σ ′ ) r_2 \in\operatorname{Sym}(\sigma') r 2 ∈ Sym ( σ ′ ) .
Let Q 0 ⊂ A ′ Q_0 \subset A' Q 0 ⊂ A ′ be the centralizer of Q 2 Q_2 Q 2 , and write σ 2 \sigma_2 σ 2 and σ 0 \sigma_0 σ 0 for the restrictions of σ ′ \sigma' σ ′ to Q 2 Q_2 Q 2 and Q 0 Q_0 Q 0 respectively. Then from (31) we obtain
( A , σ ) ≃ ( Q 0 , σ 0 ) ⊗ F ( Q 1 , σ 1 ) ⊗ F ( Q 2 , σ 2 ) . (A,\sigma) \simeq(Q_0,\sigma_0) \otimes_F (Q_1,\sigma_1) \otimes_F (Q_2,\sigma_2). ( A , σ ) ≃ ( Q 0 , σ 0 ) ⊗ F ( Q 1 , σ 1 ) ⊗ F ( Q 2 , σ 2 ) .
with r 1 ∈ Q 1 r_1 \in Q_1 r 1 ∈ Q 1 and r 2 ∈ Q 2 r_2 \in Q_2 r 2 ∈ Q 2 , hence L ⊂ Q 1 ⊗ F Q 2 L \subset Q_1 \otimes_F Q_2 L ⊂ Q 1 ⊗ F Q 2 .
Université Sorbonne Paris Nord, Institut Galilée, LAGA - CNRS (UMR 7539), F-93430 Villetaneuse, France
Email address: queguin@math.univ-paris13.fr
ICTEAM Institute, Box L4.05.01, Uclouvain, B-1348 Louvain-la-Neuve, Belgium
Email address: jean-pierre.tignol@uclouvain.be