arXiv:math/0110198v2 [math.RA] 24 Jun 2003 FIELDS OF DEFINITION FOR DIVISION ALGEBRAS M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN Abstract. Let A be a finite-dimensional division algebra containing a base field k in its center F . We say that A is defined over a subfield F0 if there exists an F0 -algebra A0 such that A = A0 ⊗F0 F . We show that: (i) In many cases A can be defined over a rational extension of k. (ii) If A has odd degree n ≥ 5, then A is defined over a field F0 of transcendence degree ≤ 21 (n − 1)(n − 2) over k. (iii) If A is a Z/m × Z/2-crossed product for some m ≥ 2 (and in particular, if A is any algebra of degree 4) then A is Brauer equivalent to a tensor product of two symbol algebras. Consequently, Mm (A) can be defined over a field F0 such that trdegk (F0 ) ≤ 4. (iv) If A has degree 4 then the trace form of A can be defined over a field F0 of transcendence degree ≤ 4. (In (i), (iii) and (iv) we assume that the center of A contains certain roots of unity.) Contents 1. Introduction 1.1. Parameter reduction 1.2. Rational fields of definition 1.3. Z/m × Z/2-crossed products 1.4. Fields of definition of quadratic forms 2. Preliminaries 2.1. G-lattices 2.2. Extension sequences 2.3. Twisted multiplicative G-fields 2.4. Rational specialization 3. G/H-crossed products 4. Proof of Theorem 1.1 2 2 3 4 4 5 5 5 6 8 9 12 1991 Mathematics Subject Classification. 16K20, 16W22, 20C10, 20J06, 11E81. Key words and phrases. division algebra, central simple algebra, crossed product, symbol algebra, cyclic algebra, biquaternion algebra, Brauer factor set, Brauer group, integral representation, G-lattice, quadratic form, trace form, Pfister form, essential dimension. M. Lorenz was supported in part by NSF grant DMS-9988756. Z. Reichstein was supported in part by NSF grant DMS-9801675 and an NSERC research grant. L. H. Rowen was supported by the Israel Science Foundation, founded by the Israel Academy of Sciences and Humanities - Center of Excellence Program no. 8007/99-3. D. J. Saltman was supported in part by NSF grant DMS-9970213. 1 2 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN 5. Proof of Theorem 1.2 5.1. Proof of part (a) 5.2. Proof of part(b) 6. Proof of Theorem 1.3 7. The field of definition of a quadratic form 7.1. Preliminaries 7.2. Proof of Theorem 1.5 References 14 14 16 17 19 19 21 22 1. Introduction Let F be a field and A a finite-dimensional F -algebra. We say that A is defined over a subfield F0 ⊂ F if there exists an F0 -algebra A0 such that A = A0 ⊗F0 F . Throughout this paper we will assume that A is a finite-dimensional central simple F -algebra, and F (and F0 ) contain a base field k. 1.1. Parameter reduction. If A is defined over F0 and trdegk (F0 ) < trdegk (F ) then the passage from A to A0 may be viewed as “parameter reduction” in A. This leads to the following natural question: What is the smallest value of trdegk (F0 ) such that A is defined over F0 ? This number is clearly finite; we shall denote it by τ (A). Of particular interest to us will be the case where A = UD(n) is the universal division algebra of degree n and F = Z(n) is the center of UD(n). Recall that UD(n) is the subalgebra of Mn (k(xij , yij )) generated, as a k-division algebra, by two generic n × n-matrices X = (xij ) and Y = (yij ), where xij and yij are 2n2 independent commuting variables over k; see, e.g., [Pr, Section II.1], [Row1 , Section 3.2] or [Sa3 , Section 14]. We shall write d(n) for τ (UD(n)). It is easy to show that d(n) ≥ τ (A) for any central simple algebra A of degree n whose center contains k; see Remark 2.8 (cf. also [Re2 , Lemma 9.2]). In other words, every central simple algebra of degree n can be “reduced to at most d(n) parameters”. To the best of our knowledge, the earliest attempt to determine the value of d(n) is due to Procesi, who showed that d(n) ≤ n2 ; see [Pr, Thm. 2.1]. If n = 2, 3 or 6 and k contains a primitive nth root of unity then d(n) = 2, because UD(n) is cyclic for these n and we can take A0 to be a symbol algebra; cf. [Re2 , Lemma 9.4]. Rost [Rost] recently proved that d(4) = 5 . (1.1) FIELDS OF DEFINITION 3 For other n the exact value of d(n) is not known. However, the following inequalities hold: d(n) ≤ n2 − 3n + 1 if n ≥ 4 [Le] , d(n) ≤ d(nm) ≤ d(n) + d(m) if (n, m) = 1 [Re2 , Sect. 9.4] , d(nr ) ≥ 2r [Re1 , Theorem 16.1] , (1.2) d(n) ≤ 12 (n − 1)(n − 2) + n for odd n [Row2 ]; cf. [Re2 , Sect. 9.3] . In this paper we will sharpen the last inequality by showing that 1 d(n) ≤ (n − 1)(n − 2) 2 for every odd n ≥ 5. Moreover, in UD(n), reduction to this number of parameters can be arranged in a particularly nice fashion: Theorem 1.1. Let n ≥ 5 be an odd integer, UD(n) the universal division algebra of degree n, and Z(n) its center. Then there exists a subfield F of Z(n) and a division algebra D of degree n with center F such that (a) UD(n) = D ⊗F Z(n), (b) trdegk (F ) = 12 (n − 1)(n − 2), and (c) Z(n) is a rational extension of F . In particular, d(n) ≤ 21 (n − 1)(n − 2). We remark that in the language of [Re2 ], the last assertion of Theorem 1.1 can be written as 1 (1.3) ed(PGLn ) ≤ (n − 1)(n − 2). 2 1.2. Rational fields of definition. Another natural question is whether or not a given central simple algebra A can be defined over a rational extension of k. We give the following partial answer to this question. Theorem 1.2. Let A be a finite-dimensional central simple algebra with center F and let t1 , t2 , . . . be algebraically independent central indeterminates over F . (a) Assume deg(A) = 2i p1 . . . pr , where i = 0, 1 or 2 and p1 , . . . , pr are distinct odd primes. Then for s ≫ 0 the algebra A(t1 , . . . , ts ) is defined over a rational extension of k. (b) Suppose the center of A contains a primitive eth root of unity, where e is the exponent of A. Then there exists an r ≥ 1 such that for s ≫ 0 the algebra Mr (A)(t1 , . . . , ts ) is defined over a rational extension of k. (Here we are imposing no restrictions on the degree of A.) Here A(t1 , . . . , ts ) stands for A⊗F F (t1 , . . . , ts ) and similarly for Mr (A)(t1 , . . . , ts ). Note that part (b) may be interpreted as saying that for s ≫ 0 the Brauer class of A(t1 , . . . , ts ) is defined over a rational extension of k. 4 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN 1.3. Z/m × Z/2-crossed products. As usual, we let (a, b)m denote the symbol algebra F {x, y}/(xm = a, y m = b, xy = ζm yx) , (1.4) ∗ th where a, b ∈ F and ζm is a (fixed) primitive m root of unity in F ; cf. [Row3 , pp. 194-197]. In Section 6 we will prove the following: Theorem 1.3. Let A be a Z/m × Z/2-crossed product central simple algebra whose center F contains a primitive 2mth root of unity ζ2m . Then A is Brauer equivalent (over F ) to (a, b)m ⊗F (c, d)2m for some a, b, c, d ∈ F ∗ . (In other words, Mm (A) is isomorphic to (a, b)m ⊗F (c, d)2m .) Note that Theorem 1.3 applies to any division algebra of degree 4, since, by a theorem of Albert, any such algebra is a Z/2 × Z/2-crossed product. In this setting our argument yields, in particular, an elementary proof of [S, Theorem 2, p. 288]. We also remark that Theorem 1.3 may be viewed as an explicit form of the Merkurjev-Suslin theorem for Z/m × Z/2-crossed products. Letting F0 = k(ζ2m , a, b, c, d), we note that   (a, b)m ⊗F (c, d)2m = (a, b)m ⊗F0 (c, d)2m ⊗F0 F . Thus Theorem 1.3 shows that Mm (A) is defined over F0 . Since trdegk (F0 ) ≤ 4, we obtain the following: Corollary 1.4. Let A be a Z/m × Z/2-crossed product central simple algebra whose center contains a primitive 2mth root of unity. Then τ (Mm (A)) ≤ 4. In particular, τ (M2 (A)) ≤ 4 for every central simple algebra A of degree 4 whose center contains a primitive 4th root of unity.  Note that the last assertion complements, in a somewhat surprising way, the above-mentioned result of Rost (1.1). Indeed, suppose the base field k contains a primitive 4th root of unity. Then for A = UD(4), Rost’s theorem says that τ (A) = 5, where as Corollary 1.4 says that τ (M2 (A)) ≤ 4. 1.4. Fields of definition of quadratic forms. A quadratic form q : V → F on an F -vector space V = F n is said to be defined over a subfield F0 of F if q = qF0 ⊗ F , where qF0 is a quadratic form on V0 = F0n . In the last section we discuss fields of definition of quadratic forms. Of particular interest to us will be trace forms of central simple algebras of degree 4, recently studied by Rost, Serre and Tignol [RST]. (Recall that the trace form of a central simple algebra A is the quadratic form x 7→ Tr(x2 ) defined over the center of A.) We will use a theorem of Serre [Se] (see our Proposition 7.3) to prove the following: Theorem 1.5. Let A be a central simple algebra of degree 4 whose center F contains a primitive 4th root of unity. Then the trace form of A is defined over a subfield F0 ⊂ F such that trdegk (F0 ) ≤ 4. FIELDS OF DEFINITION 5 Note that Theorem 1.5 may also be viewed as complementing (1.1). Acknowledgment. The authors would like to thank the referee for a number of helpful and constructive comments and for catching several mistakes in an earlier version of this paper. 2. Preliminaries 2.1. G-lattices. Throughout, G will denote a finite group and H will be a subgroup of G. Recall that a G-module is a (left) module over the integral group ring Z[G]. As usual, ExtG stands for ExtZ[G] . A G-lattice is a G-module that is free of finite rank over Z. Further, a G-lattice M is called • a permutation lattice if M has a Z-basis that is permuted by G; • permutation projective (or invertible) if M is a direct summand of some permutation G-lattice. A G-module M is called faithful if no 1 6= g ∈ G acts as the identity on M. The G/H-augmentation kernel is defined as the kernel of the natural augmentation map Z[G/H] = Z[G] ⊗Z[H] Z −→ Z , g = g ⊗ 1 7→ 1 (g ∈ G) . Thus there is a short exact sequence of G-lattices 0 → ω(G/H) → Z[G/H] → Z → 0 . (2.1) ω(G/{1}) will be written as ωG; this is the ordinary augmentation ideal of the group ring Z[G]; cf. [Pa, Chap. 3]. 2.2. Extension sequences. Exact sequences of G-lattices of the form 0 → M → P → ω(G/H) → 0 , with P permutation and M faithful (2.2) will play an important role in the sequel. In this subsection we introduce two such sequences, (2.3) and (2.4). Let dG (ω(G/H)) denote the minimum number of generators of ω(G/H) as a Z[G]-module. Then for any r ≥ dG (ω(G/H)) there exists an exact sequence f 0 → M → Z[G]r → ω(G/H) → 0 . (2.3) of G-lattices. Lemma 2.1. M is a faithful G-lattice if and only if r ≥ 2 or H = 6 {1}. Proof. It is enough to show that M ⊗ Q is G-faithful; thus we may work over the semisimple algebra Q[G]. Since f ⊗ Q splits, we have a Q[G]-isomorphism (ω(G/H) ⊗ Q) ⊕ (M ⊗ Q) ≃ Q[G]r . Similarly, the canonical exact sequence 6 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN Z[G]ωH ֌ Z[G] ։ Z[G/H] gives (ω(G/H) ⊗ Q) ⊕ Q ⊕ Q[G]ωH ≃ Q[G]. Therefore, M ⊗ Q ≃ Q[G]r−1 ⊕ Q ⊕ Q[G]ωH . If r ≥ 2 then Q[G]r−1 is G-faithful, and if H = 6 {1} then ωH ⊗ Q is H-faithful G and so Q[G]ωH ≃ (ωH ⊗ Q)↑H is G-faithful. In either case, M ⊗ Q is faithful, as desired. On the other hand, r = 1 and H = {1} leads to M ⊗ Q ≃ Q which is not faithful.  Lemma 2.2. There is an exact sequence m 0 → ω(G/H)⊗2 → P → ω(G/H) → 0 , (2.4) L where P is the (permutation) sublattice P = g1 6=g2 ∈G/H Z(g1 ⊗g2 ) of Z[G/H]⊗2 . The lattice ω(G/H)⊗2 is faithful if and only if H contains no normal subgroup 6= {1} of G and [G : H] ≥ 3. Proof. Tensoring sequence (2.1) with ω(G/H) and putting P ′ = ω(G/H) ⊗ Z[G/H], we obtain an exact sequence 0 → ω(G/H)⊗2 → P ′ → ω(G/H) → 0 , where ⊗ = ⊗Z . The elements (g1 − g2 ) ⊗ g2 with g1 6= g2 ∈ G/H form a Z-basis of P ′ , and the map m : (g1 − g2 ) ⊗ g2 7→ g1 ⊗ g2 ′ is a G-isomorphism P ≃ P . T For the faithfulness assertion, note that N = g∈G Hg acts trivially on Z[G/H] and hence on ω(G/H)⊗2; so N = {1} is surely required for faithfulness. Also, if [G : H] ≤ 2 then G acts trivially on ω(G/H)⊗2 . Conversely, if N = {1} and [G : H] ≥ 3 then it is easy to verify that ω(G/H)⊗2 is indeed faithful.  2.3. Twisted multiplicative G-fields. Recall that a G-field is a field F on which the finite group G acts by automorphisms, written f 7→ g(f ). Morphisms of G-fields are G-equivariant field homomorphisms. The G-field F is called faithful if every 1 6= g ∈ G acts non-trivially on F . If K ⊆ F is a field extension and V ⊆ F a subset of F (not necessarily algebraically independent over K) then we let K(V ) denote the subfield of F that is generated by K and V . Lemma 2.3. (cf. [Sh, Appendix 3]) Let K ⊆ F be an extension of G-fields with K faithful. Assume that F = K(V ) for some G-stable K-subspace V ⊆ F . Then (a) V = KV G , where V G denotes the G-invariants in V , (b) F = K(V G ), and (c) F G = K G (V G ). Proof. (a) Let S = K#G denote the skew group ring for the given G-action on K. The G-action on F and multiplication with K make F a (left) S-module, and V is a submodule. Moreover, since K is a faithful G-field, S is a simple ring; FIELDS OF DEFINITION 7 P see, e.g., [J, p. 473]. In particular, the element t = g∈G g ∈ S generates S as a 2-sided ideal. Thus, S = StS = KtK and consequently, V = KtKV = KV G . (b) is an immediate consequence of (a). (c) Let E = K G (V G ). We want to show that E = F G . Clearly E ⊆ F G . To prove equality, note that KE is a subring of F containing K and V G , and that dimE KE ≤ dimK G K = |G|. Thus, KE is a field, and hence (b) implies that KE = F . Therefore, dimE F ≤ |G| = dimF G F . Since E ⊆ F G , this is only possible if E = F G .  We recall a well-known construction of G-fields; cf. [Sa1 ]. Given a G-field E, a G-lattice M, and an extension class γ ∈ ExtG (M, E ∗ ), the twisted multiplicative G-field Eγ (M) is constructed as follows. Form the group algebra E[M] of M over E; this is a commutative integral domain with group of units U(E[M]) = E ∗ ×M. We shall use multiplicative notation for M in this setting. Let E(M) denote the field of fractions of E[M]. Choose an extension of G-modules 1 → E∗ → V → M → 1 (2.5) representing γ. So, as abelian groups, V ≃ U(E[M]). Using this identification, we obtain a G-action on U(E[M]) inducing the given action on E ∗ . The action of G on U(E[M]) extends uniquely to E[M], and to E(M); we will use Eγ [M] and Eγ (M) to denote E[M] and E(M) with the G-actions thus obtained. For γ = 1, we will simply write E[M] and E(M). We remark that the choice of the sequence (2.5) representing a given γ ∈ ExtG (M, E ∗ ) is insubstantial: a different choice leads to G-isomorphic results. For future reference, we record the following application of Lemma 2.3 essentially due to Masuda [Mas]; see also [Le, Proposition 1.6], [Sa3 , Lemma 12.8]. Proposition 2.4. Let E be a faithful G-field and let P be a permutation G-lattice. Then any twisted multiplicative G-field Eγ (P ) can be written as Eγ (P ) = E(t1 , . . . , tn ) with G-invariant transcendental (over E) elements ti . In particular, Eγ (P )G = E G (t1 , . . . , tn ) is rational over E G . Proof. We have an extension of G-modules 1 → E ∗ → U(Eγ [P ]) → P → 1 representing γ, as in (2.5). Fix a Z-basis, X, of P that is permuted by the action of G. For each x ∈ X, choose a preimage x′ ∈ U(Eγ [P ]). Then {x′ }x∈X is a collection of transcendental generators of Eγ (P ) over E, and G acts via g(x′ ) = g(x)′ y for some y = y(g, x) ∈ E ∗ . Letting V denote the E-subspace of Eγ (P ) that is generated by {x′ }x∈X , we conclude from Lemma 2.3 that V has a basis consisting of G-invariant elements, say t1 , . . . , tn , and Eγ (P ) = E(t1 , . . . , tn ), Eγ (P )G = E G (t1 , . . . , tn ). The ti are transcendental over E, since trdegE Eγ (P ) = rank(P ) = n.  8 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN 2.4. Rational specialization. Let A/F and B/K be central simple algebras. We will call B/K a rational specialization of A/F if there exists a field F ′ containing both F and K such that F ′ /K is rational and B ⊗K F ′ ≃ A ⊗F F ′ . In other words, B is a rational specialization of A if deg A = deg B and A embeds in some B(t1 , . . . , tn ), where t1 , t2 , . . . are independent variables over F ; cf. [Sa3 , p. 73], For the rest of this paper we fix an (arbitrary) base field k. All other fields are understood to contain a copy k and all maps (i.e., inclusions) between fields are understood to restrict to the identity map on k. Definition 2.5. Let Λ be a class of central simple algebras. We shall say that an algebra A ∈ Λ has the rational specialization property in the class Λ if every B ∈ Λ is a rational specialization of A. If Λ is the class of all central simple algebras of degree n = deg(A) then we will omit the reference to Λ and will simply say that A has the rational specialization property. Example 2.6. By [RV, Lemma 3.1], UD(n) has the rational specialization property. This is also implicit in [Sa2 ]. We remark that any central simple algebra A/F with the rational specialization property is a division algebra. To see this, specialize A to UD(n), where n = deg(A). Recall the definition of τ (A) given at the beginning of this paper. Lemma 2.7. Let A/F and B/K be a central simple algebras. (a) If A′ ≃ A⊗F F ′ for some rational field extension F ′ /F then τ (A) = τ (A′ ). (b) (cf. [Sa3 , Lemma 11.1]) If A is a rational specialization of B then τ (A) ≤ τ (B). Proof. (a) The inequality τ (A′ ) ≤ τ (A) is immediate from the definition of τ . To prove the opposite inequality, suppose A′ ≃ A0 ⊗F0 F ′ , where A0 is a central simple algebra over an intermediate field k ⊂ F0 ⊂ F ′ , and A0 is chosen so that trdegk (F0 ) = τ (A′ ). In particular, trdegk (F0 ) ≤ trdegk (F ). Then by [RV, Proposition 3.2], A0 embeds in A, i.e., A ≃ A0 ⊗F0 F for some embedding F0 ֒→ F . Consequently, τ (A) ≤ trdegk (K) = τ (A′ ), as desired. (b) We may assume B ⊗K F ′ = A′ , as in (a). Clearly τ (B) ≥ τ (A′ ), and part (a) tells us that τ (A′ ) = τ (A).  Remark 2.8. Combining Example 2.6 with Lemma 2.7(b), we see that τ (A) ≤ d(n) = τ (UD(n)) holds for every central simple algebra A of degree n; cf. [Re2 , Lemma 9.2]. FIELDS OF DEFINITION 9 3. G/H-crossed products We shall call a central simple algebra A/F an (E, G/H)-crossed product if A has a maximal subfield L whose Galois closure E over F has the property that Gal(E/F ) = G and Gal(E/L) = H. (We adopt the convention that a maximal subfield of A is a subfield L that is maximal as a commutative subring; so [L : F ] is equal to the degree of A.) We will say that A is a G/H-crossed product if it is an (E, G/H)-crossed product for some faithful G-field E. If H = {1} then a G/H-crossed product is just a G-crossed product in the usual sense (see, e.g., [Row1 , Definition 3.1.23]). Example 3.1. Consider the universal division algebra UD(n) generated by two generic matrices, X and Y , over k. Denote the center of this algebra by Z(n). Setting L = Z(n)(X), we see that UD(n) is an S n /S n−1 -crossed product [Pr, Theorem 1.9]; see also Section 4 below. Since the degree of a G/H-crossed product is equal to [G : H], we see that isomorphism classes of (E, G/H)-crossed products are in 1-1 correspondence with the relative Brauer group B(L/F ), which, in turn, is naturally identified with the kernel of the restriction homomorphism H 2 (G, E ∗ ) → H 2 (H, E ∗ ); cf. [Pi, 14.7]. A G-module M is called H 1 -trivial if H 1 (H, M) = 0 holds for every subgroup H ≤ G. Equivalently, M is H 1 -trivial if ExtG (P, M) = 0 for all permutation projective G-lattices P ; see, e.g., [Sa3 , Lemma 12.3]. Lemma 3.2. Given an exact sequence 0 → M → P → ω(G/H) → 0 , of G-lattices, with P -permutation, let N be an H 1 -trivial G-module. Denote the kernel of the restriction homomorphism H 2 (G, N) → H 2 (H, N) by K(G/H, N). Then there is a natural isomorphism ≃ φN : HomG (M, N)/ Im(HomG (P, N)) −→ K(G/H, N) . Here “natural” means that for every homomorphism N → N ′ of H 1 -trivial G-modules, the following diagram commutes HomG (M, N ′ )/ Im(HomG (P, N ′ )) φN ′ K(G/H, N ′ ) / HomG (M, N)/ Im(HomG (P, N)) (3.1) O O φN / K(G/H, N) Note that, other than in sequence (2.2), the G-lattice M need not be faithful. Proof. The lemma is a variant of [Sa3 , Theorem 12.10], where the same assertion is made for the sequence (2.4). The proof of [Sa3 , Theorem 12.10] goes through unchanged in our setting.  10 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN In subsequent applications we will always take N = E ∗ , where E is a faithful G-field. Note that E ∗ is H 1 -trivial by Hilbert’s Theorem 90. As we remarked above, K(G/H, E ∗ ) is naturally identified with B(L/F ), where L = E H , and elements of B(L/F ) are in 1-1 correspondence with (E, G/H)-crossed products. We shall denote the (E, G/H)-crossed product associated to a G-homomorphism f : M → E ∗ by Alg(f ). Lemma 3.3. Consider a sequence of G-lattices of the form (2.2). Let E be Gfield and f : M → E ∗ be a homomorphism of G-modules. If k(f (M)) is contained in a faithful G-subfield E0 of E then Alg(f ) is defined over E0G . Proof. Since f is the composition of f0 : M → E0∗ with the inclusion E0∗ ֒→ E ∗ , Lemma 3.2 tells us that A = Alg(f0 ) ⊗E0G E G .  Remark 3.4. In the special case where the sequence 0 → M → P → ω(G/H) → 0 is given by (2.4), M = ω(G/H)⊗2 has a particularly convenient set of generators yijh = (gi − gj ) ⊗ (gj − gh ) , where G/H = {g1 , . . . , gn } is the set of left cosets of H in G and i, j, h range from 1 to n = [G : H]; cf. [RS, Lemma 1.2]. If f : ω(G/H)⊗2 → E ∗ is a G-module homomorphism then the elements cijh = f (yijh) form a reduced Brauer factor set for Alg(f ) in the sense of [RS, p. 449]. Conversely, for any reduced Brauer factor set (cijh ) in E ∗ , there exists a homomorphism f : ω(G/H)⊗2 → E ∗ such that f (yijh) = cijh ; see [RS, Corollary 1.3]. Thus Lemma 3.3 takes the following form: Let A be an (E, G/H)-crossed product defined by a reduced Brauer factor set (cijh ). Suppose (cijh ) is contained in a faithful G-subfield E0 of E. Then A is defined over E0G .  This following theorem is a variant of [Sa3 , Theorem 12.11]. Theorem 3.5. Given the sequence (2.2), let µ : M ֒→ k(M)∗ be the natural inclusion. Then D = Alg(µ) has the rational specialization property in the class of G/H-crossed products containing a copy of k in their center. In particular, τ (A) ≤ rank(M) holds for any G/H-crossed product A/F with k ⊂ F . Proof. Write A = Alg(f ) for some G-homomorphism f : M → E ∗ , where E is a faithful G-field with E G = F ; see the remarks following Lemma 3.2. Furthermore, let E(P ) denote the fraction field of the group algebra E[P ], with the G-action induced from the G-actions on E and P . By Proposition 2.4, there exists an Eisomorphism j : E(P ) ≃ E(t) of G-fields, where t = (t1 , . . . , tr ), r = rank(P ), are indeterminates on which G acts trivially. Therefore, E(P )G ≃ E G (t) = F (t) is a rational extension of F . Let ft : M → E(t)∗ denote the composition of f with the FIELDS OF DEFINITION 11 natural inclusion E ∗ ֒→ E(t)∗ . Then Alg(ft ) = Alg(f ) ⊗F F (t) = A ⊗F F (t). By Lemma 3.2, Alg(ft ) ≃ Alg(ft + g|M ) for any g ∈ HomG (P, E(t)∗ ). Let g ∼ be the composite g : P ֒→ E(P )∗ −→ E(t)∗ and let ϕ be the G-module map ϕ : M → E(t)∗ , ϕ(m) = ft (m)g(m). We claim that ϕ lifts to an embedding of G-fields k(M) ֒→ E(t). Indeed, modulo E ∗ , ϕ(m) ≡ g(m) ∈ P ⊆ E(t)∗ . Hence, {ϕ(m)}m∈M isPan E-linearly Pindependent subset of E(t), and so the map k[ϕ] : k[M] → E(t), m km m 7→ m km ϕ(m) is a G-equivariant embedding of the group ring k[M] into E(t). This embedding lifts to an embedding of G-fields φ : k(M) = Q(k[M]) ֒→ E(t), as we have claimed. So φ ◦ µ = ϕ, and hence D ⊗k(M )G F (t) = Alg(φ ◦ µ) = Alg(ϕ) ≃ Alg(ft ) = A ⊗F F (t). This proves that A is a rational specialization of D. Lemmas 2.7(b) and 3.3 now imply that τ (A) ≤ τ (D) ≤ trdegk k(M)G = rank(M). This completes the proof of the theorem.  Remark 3.6. Continuing with the notation used in the above theorem, the rational specialization property of D = Alg(µ) implies that D is a division algebra of exponent [G : H]. Indeed, by [FSS, Appendix] there exists a G/Hcrossed product division algebra of exponent [G : H], and the above assertion can be proved by specializing D to this algebra. Alternatively, the fact that D is a division algebra of exponent [G : H] can be checked directly by showing that the image of µ in H 2 (G, k(M)∗ ) (see Lemma 3.2) has order [G : H]. Remark 3.7. The above construction applies in particular to sequences of the form (2.3). The following special type of sequence (2.3) has been particularly well-explored. Write G = hH, g1 , . . . , gr i for suitable gi ∈ G; the minimal such r is usually denoted by d(G/H). Then we can define an epimorphism of GP lattices f : Z[G]r ։ ω(G/H), f (α1 , . . . , αr ) = ri=1 αi (gi − 1), where : Z[G] ։ Z[G/H] is the canonical map; see [Pa, Lemma 3.1.1]. The kernel R(G/H) = Ker f is called a relative relation module; it has the following group theoretical description. Let Fr denote the free group on r generators and consider the presentation 1 → R → Fr ∗ H → G → 1 (3.2) where Fr ∗ H → G is the identity on H and sends the r generators of Fr to the elements g1 , . . . , gr . Then R(G/H) ≃ Rab = R/[R, R], with G acting by conjugation; see [Ki] and [Gr] (for H = {1}). Thus, we have the following version of sequence (2.3) with M = Rab : 0 → Rab → Z[G]r → ω(G/H) → 0 . (3.3) When H = {1} and r ≥ 2, the division algebra D constructed via (3.3) in Theorem 3.5 is identical with the generic G-crossed product of Snider [S]; see also Rosset [Ro]. Explicitly: Given a free presentation 1 → R → Fr → G → 1 of G with r ≥ 2, let M = Rab ≤ Fr = Fr /[R, R] and let a ∈ H 2 (G, k(M)∗ ) be the 12 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN image of the extension class [1 → M → Fr → G → 1] ∈ H 2 (G, M) under the natural inclusion µ : M ֒→ k(M)∗ . Then D = Alg(µ) is the G-crossed product (k(M), G, a) or, equivalently, the localization of the group algebra k[Fr ] at the nonzero elements of k[M]. Corollary 3.8. Let A be a G/H-crossed product and let dG (ω(G/H)) be the minimal number of generators of ω(G/H) as a G-module. Then τ (A) ≤ r|G| − [G : H] + 1 , where r= ( dG (ω(G/H)) if H = 6 {1} max{2, dG (ω(G/H))} if H = {1} Proof. Applying Theorem 3.5 to the exact sequence (2.3), we obtain τ (A) ≤ rank(M) = rank(Z[G]r ) − rank(ω(G/H)) = r|G| − [G : H] + 1 , as claimed. Note that for r as above, Lemma 2.1 tells us that M is faithful, so that Theorem 3.5 is, indeed, applicable.  Remark 3.9. As we pointed out in Remark 3.7, dG (ω(G/H)) ≤ d(G/H). The difference pr(G/H) = d(G/H) − dG (ω(G/H)) ≥ 0 can be arbitrarily large, even if H = {1}. In this case d(G) = d(G/{1}) is the minimal number of generators of G, and pr(G) = d(G) − dG (ωG) is usually called the presentation rank or generation gap of G. All solvable groups G have presentation rank pr(G) = 0; see [Gr, Lectures 6 and 7]. Moreover, if the derived subgroup [G, G] is nilpotent then pr(G/H) = 0 holds for every subgroup H of G; see [Ki]. Corollary 3.10. (a) Suppose a group G of order n can be generated by r ≥ 2 elements. Then τ (A) ≤ (r − 1)n + 1 for any G-crossed product central simple algebra A. (b) τ (A) ≤ (⌊log2 (n)⌋ − 1)n + 1 holds for any crossed product central simple algebra A of degree n ≥ 4. Here, as usual, ⌊x⌋ denotes the largest integer ≤ x. Proof. (a) is an immediate consequence Corollary 3.8. (b) follows from (a), because any group of order n can be generated by r ≤ log2 (n) elements. (Indeed, |hG0 , gi| ≥ 2|G0 | for any subgroup G0 of G and any g ∈ G \ G0 .) Note also that ⌊log2 (n)⌋ ≥ 2 for any n ≥ 4.  4. Proof of Theorem 1.1 For the next two sections we shall assume that G = S n and H = S n−1 . We will use the following standard notations for S n -lattices: Z[S n /S n−1 ] = Un and ω(S n /S n−1 ) = An−1 . (4.1) FIELDS OF DEFINITION 13 The natural generators of Un will be denoted by u1 , . . . , un ; the symmetric group S n permutes them via σ(ui) = uσ(i) . An−1 is the sublattice of Un generated by ui − u1 as i ranges from 2 to n. Recall that the universal division algebra UD(n) is generated, as a k-division algebra, by a pair of generic n × n-matrices X and Y . We may assume without loss of generality that X is diagonal. Following [Row2 ] we will denote the diagonal entries of X by ζii′ and the entries of Y by ζij , where ζii′ and ζij are algebraically independent variables over k. The group S n permutes these variables as follows: ′ σ(ζii′ ) = ζσ(i)σ(i) and σ(ζij ) = ζσ(i)σ(j) . We identify the multiplicative group generated by ζii′ with the S n -lattice Un (via ζii′ ↔ ui), and the multiplicative group generated by ζij with Un ⊗ Un (via ζij ↔ ui ⊗ uj ). Consider the exact sequence f 0 → Ker(f ) → Un ⊕ Un⊗2 → An−1 → 0 (4.2) of S n -lattices, where f (ui, uj ⊗uh ) = uj −uh . This sequence is the sequence (2.4) of Lemma 2.2 for G = S n and H = S n−1 , with two extra copies of Un added: the second copy of Un is the sublattice of Un⊗2 that is spanned by all elements ui ⊗ ui. Both copies of Un belong to Ker(f ); in fact, Ker(f ) = Un ⊕ Un ⊕ A⊗2 n−1 , ⊗2 where A⊗2 n−1 is identified with the sublattice of Un that is spanned by all elements (ui − uj ) ⊗ (ul − um ). Let E = k(Ker(f )) and F = E S n . By a theorem of Formanek and Procesi, F is naturally isomorphic to the center Z(n) of UD(n); see, e.g., [F1 , Theorem ′ ′ 3]. Note that E = F (ζ11 , . . . , ζnn ) is generated over F by the eigenvalues of the generic matrix X. Consequently, UD(n) is an (E, S n /S n−1 )-product, and E S n−1 is isomorphic to the maximal subfield Z(n)(X) of UD(n); see, [Pr, Section II.1]. Theorem 1.1 is now a consequence of the following: Proposition 4.1. Suppose n ≥ 5 is odd. Then V (a) UD(n) is defined over F0 = k( 2 An−1 )S n , V (b) Z(n) = k(Ker(f ))S n is rational over F0 = k( 2 An−1 )S n , V Here, we view 2 An−1 as the sublattice of antisymmetric tensors in A⊗2 n−1 , ′ ′ ′ ′ that is, the Z-span of all a ∧ a = a ⊗ a − a ⊗ a with a, a ∈ An−1 . Proof. We will deduce part (a) from Remark 3.4 by constructing a reduced V Brauer factor set contained in E0 = k( 2 An−1 ). First we note that the S n V2 action on E0 is faithful, because An−1 is a faithful S n -lattice for every n ≥ 4. V (Indeed, 2 An−1 ⊗ Q is the simple S n -representation corresponding to the partition (n − 2, 12) of n; cf. [FH, Exercise 4.6]). 14 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN We now proceed with the construction of the desired Brauer factor set. The computation in [Row2 , Section 2] shows that the elements −1 cijh = ζij ζjh ζih ∈ E∗ . form a Brauer factor set for UD(n). By [Row2 , Theorem 4], if n is odd, UD(n) has a normalized (and, in particular, reduced) Brauer factor set (c′ijh ) given by c′ijh = (cijh /chji) n+1 2 −1 −1 = (ζij ζji−1 ζjh ζhj ζhi ζih ) n+1 2 . −1 −1 Now observe that ζij ζji−1ζjh ζhj ζhi ζih is precisely the element of Un⊗2 we identified V with (ui − uj ) ∧ (uj − uh ). Thus every c′ijh lies in 2 An−1 ⊂ E0 , as desired. (b) The canonical exact sequence V 2 0 → 2 An−1 −→ A⊗2 n−1 −→ Sym An−1 → 0 of S n -lattices gives rise to an exact sequence V 0 → 2 An−1 → A⊗2 n−1 ⊕ Un ⊕ Z → Q → 0 , where we have put Q = Sym2 An−1 ⊕Un ⊕Z. The crucial fact here is that, by [LL, Section 3.5], if n is odd, Q is a permutation lattice. Applying Proposition 2.4 V2 ∼ to the extension of (faithful) S n -fields E0 = k( An−1 ) ⊆ k(A⊗2 n−1 ⊕ Un ⊕ Z) = ∗ (E0 )γ (Q), where γ is the image of the class of the above extension in ExtG (Q, E0 ), we conclude that V2 k(A⊗2 An−1 )(x1 , . . . , xm ) n−1 ⊕ Un ⊕ Z) ≃ k( as S n -fields, where m = n(n+1) + 1 and S n acts trivially on the xi ’s. Similarly, 2 ⊗2 2 putting Ln = An−1 ⊕ Un , the obvious sequence 0 → A⊗2 n−1 ⊕ Un → Ln → Un → 0 ⊗2 leads to k(Ln ) ≃ k(An−1 ⊕ Un )(t1 , . . . , tn ) as S n -fields. Therefore, k(Ln ) ≃ k(A⊗2 n−1 ⊕ Un )(t1 , . . . , tn ) = k(A⊗2 n−1 ⊕ Un ⊕ Z)(t1 , . . . , tn−1 ) V2 ≃ k( An−1 )(x1 , . . . , xm , t1 , . . . , tn−1 ) as S n -fields, which implies that V Z(n) ≃ k(Ln )S n ≃ k( 2 An−1 )S n (x1 , . . . , xm , t1 , . . . , tn−1 ) ; V so Z(n) is rational over F0 = k( 2 An−1 )S n .  5. Proof of Theorem 1.2 5.1. Proof of part (a). Reduction 5.1. Suppose an algebra A0 of degree n has the rational specialization property (see Section 2.4). If Theorem 1.2(a) holds for A0 then it holds for any central simple algebra A of degree n. FIELDS OF DEFINITION 15 Proof. Suppose the for some r ≥ 1, A0 (t1 , . . . , tr ) is defined over a rational extension F0 of k. Let A/F be an arbitrary central simple algebra of degree n. Then by the rational specialization property, A0 embeds in A(tr+1 , . . . , ts ) for some s ≫ 0; thus A0 (t1 , . . . , tr ) embeds in A(t1 , . . . , ts ). This shows that A(t1 , . . . , ts ) is defined over F0 , as desired.  In particular, in proving Theorem 1.2(a), we may assume that A is a division algebra of degree n; see Example 2.6. By primary decomposition (cf., e.g., [Pi, p. 261]), we only need to consider the cases where n = 2, n = 4 and n is an odd prime. This follows from the next reduction: Reduction 5.2. If the conclusion of Theorem 1.2(a) holds for central simple algebras A1 /F and A2 /F (for every choice of the base field k ⊂ F ) then it also holds for A = A1 ⊗F A2 . Proof. After replacing A1 and A2 by, respectively, A1 (t1 , . . . , ts ) and A2 (t1 , . . . , ts ), we may assume that A1 is defined over a subfield F1 ⊂ F such that k ⊂ F1 and F1 is rational over k. We will now think of F1 (rather than k) as our new base field. After adding more indeterminates, we may assume that A2 is defined over a subfield F2 ⊂ F , where F1 ⊂ F2 and F2 is rational over F1 . Now F2 is rational over k, and since A1 and A2 are both defined over F2 , so is A.  We are now ready to complete the proof of Theorem 1.2(a). First, suppose n = 2 or 4. Since UD(n) has the rational specialization property, we may assume A = UD(n); see Reduction 5.1. But since the center of UD(n) is known to be rational for n = 2 (see [Pr, Theorem 2.2]) and n = 4 (see [F2 ]), these algebras clearly satisfy the conclusion of Theorem 1.2(a). This completes the proof of the theorem for n = 2 and 4. We remark that the same argument goes through for n = 3 (because the center of UD(3) is known to be rational; see [F1 ]) and for n = 5, 7 (because the centers of UD(5) and UD(7) are known to be stably rational; see [BL]), but we shall not need it in these cases. From now on we will assume that n = p is an odd prime. Then the S n -lattice ⊗2 An−1 is faithful; see Lemma 2.2. Furthermore, by a theorem of Bessenrodt and LeBruyn [BL, Proposition 3] (see also [Be, Lemma 2.8] for a more explicit form of this result), A⊗2 n−1 is permutation projective, i.e., there exists an S n -lattice L ⊗2 such that P = An−1 ⊕ L is permutation. We can assume that k(P )S n is rational over k. Indeed, after adding a copy of Un if necessary, we have P = Un ⊕ Q for some permutation lattice Q, and so k(P ) ≃ k(Un )(Q). Proposition 2.4 implies that k(P )S n is rational over k(Un )S n , which in turn is rational over k. Let ⊗2 ∗ i : A⊗2 n−1 ֒→ k(An−1 ) and ⊗2 ∗ j : A⊗2 n−1 ֒→ k(An−1 ⊕ P ) 16 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN be the natural embeddings of S n -modules. (Here, j identifies A⊗2 n−1 with the ⊗2 first summand of An−1 ⊕ P .) Recall that by Lemma 3.2 these embeddings, in combination with the exact sequence (2.4) (for G = S n and H = S n−1 ; see (4.1)), give rise to central simple algebras Alg(i) and Alg(j). By Theorem 3.5, Alg(i) has the rational specialization property in the class of S n /S n−1 -crossed products. Thus, the universal division algebra UD(n), being an S n /S n−1 -crossed product (see Example 3.1), is a rational specialization of Alg(i). Since UD(n) has the rational specialization property in the class of all central simple algebras of degree n (see Example 2.6), so does Alg(i). We claim that Alg(j) also has the rational specialization property in the class of central simple algebras of degree n. Indeed, by Lemma 3.2, Alg(j) = Alg(i) ⊗F S n E S n , ⊗2 where F = k(A⊗2 n−1 ) and E = k(An−1 ⊕ P ). Now Proposition 2.4 tells us that Sn Sn E is a rational extension of F , and the claim follows. By Reduction 5.1 it now suffices to prove that Alg(j) is defined over a purely transcendental extension of k. Put E0 = k(A⊗2 n−1 ⊕ (0) ⊕ L) ⊆ E. Since the image of j is contained in E0∗ , Lemma 3.3 tells us that Alg(j) is defined over E0S n . But E0 ≃ k(P ) and so E0S n ≃ k(P )S n which is indeed rational over k. This completes the proof of Theorem 1.2(a).  5.2. Proof of part(b). By the Merkurjev-Suslin Theorem, Mr (A) = (a1 , b1 )n1 ⊗F · · · ⊗F (al , bl )nl , for some r, l ≥ 1, where (a, b)n denotes a symbol algebra; see (1.4). Let λ1 , . . . , λl , µ1 , . . . , µl be 2l central variables, algebraically independent over F . We will write λ in place of (λ1 , ..., λl ) and µ in place of (µ1 , ..., µl ). Then Mr (A)(λ, µ) = (a1 , b1 )n1 ⊗K(λ,µ) · · · ⊗K(λ,µ) (al , bl )nl = (a′1 , b′1 )n1 ⊗K(λ,µ) · · · ⊗K(λ,µ) (a′l , b′l )nl =   (a′1 , b′1 )n1 ⊗F0 · · · ⊗F0 (a′l , b′l )nl ⊗F0 K(λ, µ) , where a′i = ai λni i and b′i = bi µni i for i = 1, . . . , l and F0 = k(a′1 , b′1 , ..., a′l , b′l ). This shows that Mr (D)(λ, µ) is defined over F0 . It remains to prove that F0 is rational over k. The 2l elements a′1 , b′1 , . . . , a′l , b′l are clearly algebraically independent over F . Hence, they are algebraically independent over k, and consequently, F0 is rational over k, as claimed.  Remark 5.3. Our proof of Theorem 1.2 can be used to deduce explicit lower bounds on s in parts (a) and (b) from explicit lower bounds in Theorems of Bessenrodt-LeBruyn [BL, Proposition 3] (on rank(L)) and Merkurjev-Suslin (on r). The lowest possible value of r in part (b), called the Merkurjev-Suslin number, is of independent interest; see [Row3 , Section 7.2]. FIELDS OF DEFINITION 17 6. Proof of Theorem 1.3 Reduction 6.1. In the course of proving Theorem 1.3, we may assume without loss of generality that A is a division algebra. Indeed, let D = Alg(µ), as in Theorem 3.5, with G = Z/m×Z/2, and H = {1}. Then D is a division algebra (see Remark 3.6), and any other G-crossed product A/F is a rational specialization of D. Thus, if we know that Theorem 1.3 holds for D then it holds for A(t1 , . . . , ts ), where t1 , . . . , ts are independent variables over F . Using induction on s, we see that Reduction 6.1 is now a consequence of the following lemma (applied to B = Mm (A), with r = 2, m1 = m and m2 = 2m): Lemma 6.2. Let B/K be a central simple algebra of degree d = m1 . . . mr and let t be an independent variable over K. Assume K contains a primitive root of unity of degree lcm(m1 , . . . , mr ). If B(t) = (a1 (t), b1 (t))m1 ⊗ · · · ⊗ (ar (t), br (t))mr for some a1 (t), b1 (t), . . . , ar (t), br (t) ∈ K(t) then B = (a′1 , b′1 )m1 ⊗ · · · ⊗ (a′r , b′r )mr for some a′1 , b′1 , . . . , a′r , b′r ∈ K. Our proof is based on a standard specialization argument; for the sake of completeness, we supply the details below. Proof. We may assume that K is an infinite field; otherwise B is a matrix algebra over K, and we can take, e.g., a′i = 1, b′i = −1 for every i. Choose generators xi (t) and yi(t) for the cyclic subalgebra (ai (t), bi (t))mi of B(t) = B ⊗K K(t) such that xi (t)mi = ai (t), yi(t)mi = bi (t), and xi (t)yi (t) = ζmi yi(t)xi (t), where ζmi is a primitive root of unity of degree mi in K. Choose a K-basis b1 , . . . , bd2 of B and write 2 xi (t) = d X 2 αij (t)bi j=1 and yi (t) = d X βij (t)bi , (6.1) j=1 for some αij (t), βij (t) ∈ K(t). Since K is an infinite field, we can choose t0 ∈ K such that αij (t0 ) and βij (t0 ) are well-defined and 2 2 xi (t0 ) = d X j=1 αij (t0 )bi and yi (t0 ) = d X βij (t0 )bi j=1 are non-zero. Let Bi denote the subalgebra of B that is generated by xi (t0 ) and yi (t0 ). Then Bi = (a′i , b′i )mi , where a′i = xi (t0 )mi = ai (t0 ) and b′i = yi (t0 )mi = bi (t0 ), and B1 , . . . , Br are commuting subalgebras of B of degrees m1 , . . . , mr . 18 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN Hence, by the double centralizer theorem (cf., e.g., [Pi, Theorem 12.7]), B = B1 ⊗ · · · ⊗ Br . This completes the proof of Lemma 6.2 and thus of Reduction 6.1.  We now continue with the proof of Theorem 1.3. In the course of the proof we shall use the following notations. Write G = Z/m × Z/2 = hσ1 , σ2 i, where σ1m = σ22 = 1. Let K = F (α1 , α2 ), be a maximal G-Galois subfield of A, where α1m = a1 and α22 = a2 are elements of F , and σ1 (α1 ) = ζm α1 , σ1 (α2 ) = α2 , σ2 (α1 ) = α1 , σ2 (α2 ) = −α2 . (6.2) Here ζm ∈ F is a primitive mth root of unity, so that K is, indeed, a G-Galois extension of F . Note that the statement of Theorem 1.3 assumes that F contains not only a primitive mth root of unity ζm but also a primitive 2mth root of unity ζ2m ; we shall make use of ζ2m later in the proof. By the Skolem-Noether Theorem, there exist units z1 , z2 ∈ A such that zi xzi−1 = σi (x) for every x ∈ K (i = 1, 2). Set z1m = b1 ∈ F (α2 )∗ , z22 = b2 ∈ F (α1 )∗ , and u = z1 z2 z1−1 z2−1 ∈ K ∗ . (6.3) By [AS, Theorem 1.3], the algebra structure of A can be recovered from the G-field K and the elements u ∈ K ∗ , b1 ∈ F (α2 )∗ and b2 ∈ F (α1 )∗ . (These elements have to satisfy certain compatibility conditions; the exact form of these conditions shall not concern us in the sequel.) We will write A = (K, G, u, b1 , b2 ). Lemma 6.3. Let A = (K, G, u, b1, b2 ) and A′ = (K, G, u′, b′1 , b′2 ) be G-crossed products. Then A ⊗F A′ is Brauer equivalent to (K, G, uu′, b1 b′1 , b2 b′2 ). Proof. The class of A = (K, G, u, b1, b2 ) in the relative Brauer group B(K/F ) = H 2 (G, K ∗ ) is given by a normalized 2-cocycle a : G × G → K ∗ so that bi = a(σi , σi )a(σi2 , σi ) . . . a(σimi −1 , σi ) holds for i = 1, 2, where m1 = m and m2 = 2, and u = a(σ1 , σ2 )a(σs , σ1 )−1 . Similarly, the class of A′ is given by a 2-cocycle a′ . Then the class of A ⊗F A′ is given by the cocycle aa′ ; see, e.g., [Pi, Proposition 14.3]. This proves the lemma. The following alternative ring-theoretic argument was suggested by the referee: Choose z1 , z2 ∈ A, as in (6.3), and similarly for z1′ , z2′ in A′ . The subalgebra S of A ⊗F A′ generated by K ⊗ 1, z1 ⊗ z1′ and z2 ⊗ z2′ , is clearly isomorphic to (K, G, uu′, b1 b′1 , b2 b′2 ). Its centralizer CA⊗A′ (S) is an F -central simple algebra of degree 2m = [K : F ], containing (K ⊗ K)G , where G acts diagonally on K ⊗ K. Since G is abelian, (K ⊗ K)G ≃ F ⊕ · · · ⊕ F , as an F [G]-algebra. In particular, CA⊗A′ contains the idempotents of K ⊗ K and, hence, is split over F . We thus conclude that A ⊗F A′ ≃ S ⊗F CA⊗F A′ (S) ∼ S ≃ (K, G, uu′, b1 b′1 , b2 b′2 ) , FIELDS OF DEFINITION as claimed. (Here ∼ denotes Brauer equivalence over F .) 19  We now proceed with the proof of Theorem 1.3, using the notations of (6.2) and (6.3). Since b1 = z1m ∈ K σ1 = F (α2 ), we can write b1 = f1 + f2 α2 , (6.4) for some f1 , f2 ∈ F . Lemma 6.4. (a) If f1 = 0 then A is cyclic. (b) If f2 = 0 then A = (a, b)m ⊗ (c, d)2, for some a, b, c, d ∈ F ∗ . Proof. (a) If f1 = 0 then z12m = b21 = f22 a2 ∈ F ∗ but z1m = f2 α2 6∈ F . Since F contains a primitive root of unity of degree 2m, F (z1 ) is a cyclic maximal subfield of A of degree 2m; cf. [Lang, Theorem VIII.6.10(b)]. Thus A is a cyclic algebra, as claimed. (b) If f2 = 0, i.e., b1 ∈ F , then the F -subalgebra A0 of A generated by z1 and α1 is cyclic of degree m. By the double centralizer theorem, A = Am ⊗ Q, where Q is a quaternion algebra, as claimed.  We are now ready to finish the proof of Theorem 1.3. Lemma 6.4 tells us that Theorem 1.3 is immediate if f1 = 0 or f2 = 0. Thus from now on we shall assume f1 f2 6= 0. Now let A = (K, G, u, b1 , b2 ) and, for any f ∈ F ∗ , define Af = (K, G, u, f b1, b2 ). Since (a1 , f )m⊗F M2 (F ) ≃ (K, G, 1, f, 1), Lemma 6.3 tells us that Af ∼ (a1 , f )m ⊗F A, where ∼ denotes Brauer equivalence. In other words, A ∼ (f, a1 )m ⊗F Af . Thus it is enough to show that Af is cyclic, for some f ∈ F ∗ . To prove the last assertion, observe that if we expand (z1 + α1 )m then all terms, other than z1m and α1m , will cancel. (For a simple proof of this fact, due to Bergman, see [Row3 , p. 195]). Thus, if γ = z1 + α1 then γ m = z1m + α1m = f b1 + a1 in Af . Setting f = − af11 ∈ F ∗ , we obtain γ m = cα2 , where c = − af11f2 ∈ F ∗ ; see (6.4). Thus γ 2m = c2 a2 ∈ F ∗ but γ m 6∈ F . Since F contains a primitive 2mth root of unity, F (γ)/F is a cyclic field extension of degree 2m. In other words, F (γ) a cyclic maximal subfield of Af , and Af is a cyclic algebra of degree 2m, as claimed.  7. The field of definition of a quadratic form 7.1. Preliminaries. Let V = F n be an F -vector space, equipped with a quadratic form q : V → F . Recall that q is said to be defined over a subfield F0 of F if q = qF0 ⊗ F , where qF0 is a quadratic form on V0 = F0n . Is easy to see that q is defined over F0 if and only if V has an F -basis e1 , . . . , en such that b(ei , ej ) ∈ F0 , where b : V × V → F is the symmetric bilinear form associated to q (i.e., q(v) = b(v, v)). 20 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN We shall always assume that char(F ) 6= 2 and F (and F0 ) contain a base subfield k. As usual, we shall write <a1 , . . . , an > for the diagonal form an x21 + · · · + an x2n and ≪ a1 , . . . , an ≫ for the Pfister form <1, a1 > ⊗ · · · ⊗ <1, an >. Given a quadratic form q we shall ask: (a) What is the smallest value of trdegk (F0 ), where q is defined over F0 ? We shall denote this number by τ (q). (b) Can q be defined over a rational extension F0 of k? These are the same questions we asked for central simple algebras in the Introduction. In the case of quadratic forms our answers are more complete (and the proofs are easier). Proposition 7.1. Let V = F n and let q : V → F be a quadratic form on V . Then (a) τ (q) ≤ n. Moreover, if a1 , . . . , an are independent variables over k, F = k(a1 , . . . , an ), and q = <a1 , . . . , an > then τ (q) = n. (b) Let t1 , . . . , tn be independent variables over F . Then q ′ = q⊗F F (t1 , . . . , tn ) is defined over a rational extension F0 of k. Proof. Diagonalizing q, write q = <a1 , . . . , an > in the basis e1 , . . . , en . (a) To prove the first assertion, set F0 = k(a1 , . . . , an ). Then q is defined over F0 = k(a1 , . . . , an ) and trdegk (F0 ) ≤ n, as desired. For the proof of the second assertion see [Re2 , Proof of Theorem 10.3]. (b) Set a′i = t2i ai . Then q ′ = <a1 , . . . , an > = <a′1 , . . . , a′n > over F (t1 , . . . , tn ). Hence, q ′ is defined over F0 = k(a′1 , . . . , a′n ). We claim that F0 is rational over k. Indeed, since the nonzero elements of {a′1 , . . . , a′n } are algebraically independent over F , they are algebraically independent over k, and the claim follows.  In the sequel we shall need the following analogue of Lemma 2.7. Lemma 7.2. Let q be a quadratic form defined over F , t1 , . . . , tr be independent variables over F , and F ′ = F (t1 , . . . , tr ). Set q ′ = q ⊗F F ′ . Then τ (q) = τ (q ′ ). Proof. The inequality τ (q ′ ) ≤ τ (q) is obvious from the definition of τ (q). To prove the opposite inequality, we may assume F is an infinite field; otherwise τ (q) = 0, and there is nothing to prove. We may also assume r = 1; the general case will then follows by induction on r. Let b′ be the symmetric bilinear form associated to q ′ and choose a basis b1 (t), . . . , bn (t) of (F ′ )n so that trdegk k(αij (t)) = τ (q ′ ), where αij (t) = b′ (bi (t), bj (t)). Since F is an infinite field, we can find a c ∈ F such that (i) the vectors b1 (c), . . . , bd (c) are welldefined and form a basis of F d , and (ii) each αij (c) is well-defined. Now q is defined over k(αij (c)) and thus τ (q) ≤ trdegk (αij (c)) ≤ trdegk (αij (t)) = τ (q ′ ) , FIELDS OF DEFINITION as claimed. 21  7.2. Proof of Theorem 1.5. Let F be a field containing a primitive 4th root of unity. Note that for the purpose of proving Theorem 1.5, we may assume that A/F is a division algebra. Otherwise, A is isomorphic to M4 (F ) or to M2 (D), where D = (a, b)2 is a quaternion algebra. Thus, A is defined over k or over the field F0 = k(a, b), respectively, and so is the trace form of A. Alternatively, a simple direct computation shows that the trace form of M2 (E) is trivial (and thus is defined over k) for any central simple algebra E/F . From now on we will assume that A/F is a division algebra of degree 4. By a theorem of Albert [A], A is a G-crossed product, with G = Z/2 × Z/2. Let K be a G-Galois maximal subfield. Using the notations introduced in Section 6 (with m = 2), we write G = <σ1 , σ2 >, K = F (α1 , α2 ), αi2 = ai ∈ F and A = (K, G, u, b1 , b2 ) for some u ∈ K ∗ , b1 ∈ F (α2 ) = K σ1 , and b2 ∈ F (α1 ) = K σ2 . Set σ3 = σ1 σ2 ∈ Gal(K/F ); z3 = (z1 z2 )−1 , α3 = α1 α2 , a3 = α32 = a1 a2 , b3 = z32 (so that bi = zi2 for i = 1, 2, 3), and ti = 1 TrK σi /F (zi2 ) , 2 ni = NK σi /F (zi2 ) for i = 1, 2, 3. Our proof of Theorem 1.5 is based on the following result of Serre [Se], [RST]. Proposition 7.3. Suppose z1 and z2 are chosen so that ti 6= 0 and n2i − ti 6= 0 for any i = 1, 2, 3. Then the trace form q of A is Witt-equivalent (over F ) to q2 ⊕ q4 , where q2 =≪ n1 − t21 , n2 ≫ is a 2-fold Pfister form and q4 =≪ t1 − n21 , (n2 − t22 )n2 , t1 t2 , t2 t3 ≫ is a 4-fold Pfister form.  We claim that for the purpose of proving Theorem 1.5, we may assume without loss of generality that ti 6= 0 and n2i − ti 6= 0 for any i = 1, 2, 3. Indeed, in view of Lemma 7.2 it suffices to prove Theorem 1.5 for a single division algebra A which has the rational specialization property in the class of algebras of degree 4, e.g., for A = UD(4); see Remark 2.8. Thus we only need to show that in this algebra ti 6= 0 and ni − t2i 6= 0 for any choice of z1 , z2 . Indeed, we may assume without loss of generality that i = 1 (the cases where i = 2 and 3 will then follow by symmetry). Write b1 = f1 + f2 α2 for some f1 , f2 ∈ F , where t1 = f1 and n1 − t21 = f22 a2 . Lemma 6.4 shows that if t1 = 0 then A is cyclic and if n1 − t21 = 0 then A is biquaternion. But since our algebra A has the rational specialization property, it is neither cyclic nor biquaternion. We conclude that t1 (n1 − t21 ) 6= 0, as claimed. 22 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN We now proceed to simplify the form given by Proposition 7.3. After expand2 ing q2 and q4 , cancelling √ the common term <1, t1 − n1 > (which can be done, since we are assuming −1 ∈ F ) and dividing some of the entries by elements of (F ∗ )2 , we see that the trace form of A is Witt equivalent to the 16-dimensional form n1 n2 n2 n2 q = <1 , 1 − 2 > ⊗ (< 2 >⊕ ≪ (1 − 2 ) 2 , t1 t2 , t2 t3 ≫0 ) , (7.1) t1 t2 t2 t2 where ≪ λ1 , . . . , λr ≫0 is defined as a 2r − 1-dimensional form such that ≪ λ1 , λ2 , . . . , λr ≫0 ⊕<1> is the n-fold Pfister form ≪ λ1 , λ2 , . . . , λr ≫. Note that since q and the trace form of A are Witt equivalent 16-dimensional forms, the Witt Decomposition Theorem implies that they are, in fact, the same (i.e., isometric). We now observe that all entries of q lie in the subfield F0 = k( nt21 , nt22 , t1 t2 , t2 t3 ) of F . Thus the trace form of A is defined over F0 . Since F0 1 2 is generated by 4 elements over k, we have trdegk F0 ≤ 4. This completes the proof of Theorem 1.5.  References [A] [AS] [Be] [BL] [FSS] [F1 ] [F2 ] [FH] [Gr] [J] [Ki] [Lam] [Lang] [Le] [LL] A. A. Albert, Structure of Algebras, Amer. Math. Soc. Colloq. Publ. 24, 1939. S. A. Amitsur, S. J. Saltman, Generic abelian crossed products and p-algebras, J. Algebra, 51 (1978), 76–87. E. Beneish, Induction theorems on the stable rationality of the center of the ring of generic matrices, Trans. Amer. Math. Soc. 350 (1998), 3571–3585. C. Bessenrodt and L. Le Bruyn, Stable rationality of certain P GLn -quotients, Invent. Math. 104 (1991), 179–199. B. Fein, D. J. Saltman and M. Schacher, Embedding problems for finite dimensional division algebras, J. Algebra 167 (1994), 588–626. E. Formanek, The center of the ring of 3 × 3 generic matrices, Linear and Multilinear Algebra 7 (1979), 203–212. E. Formanek, The center of the ring of 4 × 4 generic matrices, J. Algebra 62 (1980), no. 2, 304–319. W. Fulton and J. Harris, Representation Theory, A First Course, Springer-Verlag, New York, 1991. K. W. Gruenberg, Relation Modules of Finite Groups, CBMS Regional Conference Series in Mathematics, No. 25, Amer. Math. Soc., Providence, 1976. N. Jacobson, Basic Algebra II, 2nd ed., W.H. Freeman and Co., New York, 1989. W. Kimmerle, Relative relation modules as generators for integral group rings of finite groups, Math. Z. 172 (1980), 143–156. T.-Y. Lam, The Algebraic Theory of Quadratic Forms, W. A. Benjamin Inc., 1973. S. Lang, Algebra, Addison-Wesley Publishing Co., Inc., 1965. N. Lemire, Essential dimension of algebraic groups and integral representations of Weyl groups, preprint. Available at http://www.math.uwo.ca/˜nlemire/preprints.html. N. Lemire and M. Lorenz, On certain lattices associated with generic division algebras, J. Group Theory 3 (2000), 385–405. FIELDS OF DEFINITION [Le] [Mas] [Pa] [Pi] [Pr] [Re1 ] [Re2 ] [RV] [Ro] [Rost] [RST] [Row1 ] [Row2 ] [Row3 ] [Row4 ] [RS] [Sa1 ] [Sa2 ] [Sa3 ] [Se] [Sh] [S] 23 H. W. Lenstra, Rational functions invariant under a finite abelian group, Invent. Math. 25 (1974), 299–325. K. Masuda, On a problem of Chevalley, Nagoya Math. J. 8 (1955), 59–63. D. S. Passman, The Algebraic Structure of Group Rings, John Wiley & Sons, New York, 1977. R. S. Pierce, Associative Algebras, Springer-Verlag, New York, 1982. C. Procesi, Non-commutative affine rings, Atti Acc. Naz. Lincei, S. VIII, v. VIII, fo. 6 (1967), 239–255. Z. Reichstein, On a theorem of Hermite and Joubert, Canadian J. Math., 51 (1) (1999), 69–95. Z. Reichstein, On the notion of essential dimension for algebraic groups, Transformation Groups, 5, no. 3 (2000), 265–304. Z. Reichstein and N. Vonessen, An embedding property of universal division algebras, J. Algebra 177 (1995), 451–462. S. Rosset, Group extensions and division algebras, J. Algebra 53 (1978), 297–303. M. Rost, Computation of some essential dimensions, preprint. Available at http://www.math.ohio-state.edu/˜rost/ed.html. M. Rost, J.-P. Serre, J.-P. Tignol, The trace form of a central simple algebra of degree four, in preparaton. L. H. Rowen, Polynomial Identities in Ring Theory, Academic Press, 1980. L. H. Rowen, Brauer factor sets and simple algebras, Trans. Amer. Math. Soc., 282, no. 2 (1984), 765–772. L. H. Rowen, Ring Theory, vol. 2, Academic Press, 1988. L. H. Rowen, Division algebras over C2 - and C3 -fields, Proc. Amer. Math. Soc. 130 (2002), no. 6, 1607–1610 L. H. Rowen and D. J. Saltman, Normalized Brauer factor sets, J. Algebra 198 (1997), 446–468. D. J. Saltman, Twisted multiplicative field invariants, Noether’s problem, and Galois extensions, J. Algebra 131 (1990), 535–558. D. J. Saltman, A note on generic division algebras, Contemp. Math. 130 (1992), 385–394. D. J. Saltman, Lectures on Division Algebras, CBMS Regional Conference Series in Mathematics, No. 94, Amer. Math. Soc., Providence, 1999. J.-P. Serre, letters of November 16 and November 24, 1998. I. R. Shafarevich, Basic Algebraic Geometry, vol. 1, second edition, Springer-Verlag, 1994. R. L. Snider, Is the Brauer group generated by cyclic algebras?, in “Ring Theory. Waterloo 1978”, Lect. Notes in Math. 734 (1979), Springer, 279–301. 24 M. LORENZ, Z. REICHSTEIN, L. H. ROWEN, AND D. J. SALTMAN Department of Mathematics, Temple University, Philadelphia, PA 19122-6094 E-mail address: lorenz@math.temple.edu URL: http://www.math.temple.edu/∼lorenz Department of Mathematics, University of British Columbia, Vancouver, BC, Canada, V6T 1Z2 E-mail address: reichst@math.ubc.ca URL: http://www.math.ubc.ca/∼reichst Department of Mathematics and Computer Science, Bar-Ilan University, Ramat-Gan 52900, Israel E-mail address: rowen@macs.biu.ac.il Department of Mathematics, University of Texas, Austin, TX 78712 E-mail address: saltman@math.utexas.edu