Bourbaki

§ 16. Differential criteria of separability

1. Extension of derivations : the case of rings

Let K be a commutative ring, A a commutative K-algebra and x=(xi)i∈Ix = (x_i)_{i \in I}, a family of elements of A. Further, let AA be a derivation of K into an A-module M, in other words (III, p. 553) a Z-linear mapping of K into M satisfying the relation Δ(cc′)=c⋅Δ(c′)+c′⋅Δ(c)\Delta(cc') = c \cdot \Delta(c') + c' \cdot \Delta(c) for c,c′c, c' in K. For each i∈Ii \in I let DiD_i be the partial derivation with respect to XiX_i in the polynomial ring K[Xi]i∈IK[X_i]_{i \in I}; this is the unique derivation of that ring into itself which is zero on K and on XjX_j for j∈I−{i}j \in I - \{i\}, and takes the value 1 on XiX_i (IV, p. 6). For every polynomial f=∑a∈N(I)ca⋅Xαf = \sum_{a \in N^{(I)}} c_a \cdot X^\alpha in K[Xi]i∈IK[X_i]_{i \in I}, we denote by fΔ(x)f^\Delta(x) the element ∑α∈N(I)xα⋅Δ(cα)\sum_{\alpha \in N^{(I)}} x^\alpha \cdot \Delta(c_\alpha) of M.

Proposition 1

— Suppose that the family x=(xi)i∈Ix = (x_i)_{i \in I}, generates the K-algebra A. Let (mi)i∈I(m_i)_{i \in I}, be a family of elements of M and (fλ)λ(f_\lambda)_\lambda, a family of polynomials generating the ideal aa of all polynomials f∈K[Xi]i∈If \in K[X_i]_{i \in I} such that f(x)=0f(x) = 0. For a derivation D of A into M to exist such that D(c⋅1)=Δ(c)D(c \cdot 1) = \Delta(c) for all c∈Kc \in K and D(xi)=miD(x_i) = m_i for all i∈Ii \in I it is necessary and sufficient that

fλΔ(x)+∑i∈IDifλ(x)⋅mi=0for allλ∈A. f_\lambda^\Delta(x) + \sum_{i \in I} D_i f_\lambda(x) \cdot m_i = 0 \quad \text{for all} \quad \lambda \in A .

If this is so, the derivation D is unique and satisfies

D(f(x))=fΔ(x)+∑i∈IDif(x)⋅mifor allf∈K[Xi]i∈I D(f(x)) = f^\Delta(x) + \sum_{i \in I} D_i f(x) \cdot m_i \quad \text{for all} \quad f \in K[X_i]_{i \in I}

Put E=K[Xi]i∈IE = K[X_i]_{i \in I} and write φ\varphi for the K-homomorphism of E onto A which maps XiX_i to xix_i for all i∈Ii \in I; we thus have φ(f)=f(x)\varphi(f) = f(x) for all f∈Ef \in E. We consider M as an E-module by means of the homomorphism φ:E→A\varphi : E \to A and define a mapping D′D' of E into M by D′(f)=fΔ(x)+∑i∈IDif(x)⋅miD'(f) = f^\Delta(x) + \sum_{i \in I} D_i f(x) \cdot m_i (note that the family (Dif)i∈I(D_i f)_{i \in I} has finite support for each f∈Ef \in E). It is clear that D′D' is the unique derivation of E into M extending A and mapping XiX_i to mim_i for each i∈Ii \in I.

Let D be a derivation of A into M such that D(c⋅1)=Δ(c)D(c \cdot 1) = \Delta(c) for all c∈Kc \in K and D(xi)=miD(x_i) = m_i for all i∈Ii \in I. Then D∘φD \circ \varphi is a derivation of E into M extending A and mapping XiX_i to mim_i for all i∈Ii \in I. We thus have D∘φ=D′D \circ \varphi = D', that is, the relation (2) holds. This proves the uniqueness of D ; moreover (1) is a consequence of f,(x)=0f, (x) = 0 and (2).

Conversely assume that (1) holds; in other words, we have D′(fλ)=0D'(f_\lambda) = 0 for all λ∈A\lambda \in A. Let f∈af \in a; there exists a family with finite support (qλ)λ∈Λ(q_\lambda)_{\lambda \in \Lambda} in EE such that f=∑λ∈Λqλ⋅fλf = \sum_{\lambda \in \Lambda} q_\lambda \cdot f_\lambda. We have

D′(f)=∑λ∈Λ[fλ(x)⋅D′(qλ)+qλ(x)⋅D′(fλ)] D'(f) = \sum_{\lambda \in \Lambda} [f_\lambda(x) \cdot D'(q_\lambda) + q_\lambda(x) \cdot D'(f_\lambda)]
whence D′(f)=0D'(f) = 0 because fλ(x)f_\lambda(x) and D′(fλ)D'(f_\lambda) are zero for all λ∈A\lambda \in A. Since D′D' vanishes on aa, there exists a ZZ-linear mapping DD of AA into MM such that D′=D∘φD' = D \circ \varphi; clearly DD is the required derivation of AA into MM.

2. Extension of derivations: the case of fields

Let K,LK, L and MM be fields such that K⊂L⊂MK \subset L \subset M. By Prop. 21 (III, p. 572) we have an exact sequence of vector MM-spaces

(EK,L,M)ΩK(L)⊗LM→αΩK(M)→βΩL(M)→0; (\mathrm{E}_{K,L,M}) \quad \Omega_K(L) \otimes_L M \xrightarrow{\alpha} \Omega_K(M) \xrightarrow{\beta} \Omega_L(M) \to 0 ;
the MM-linear mappings α\alpha and β\beta are characterized by the relations
(3)α(dL/Kx⊗1)=dM/Kxfor x∈L(4)β(dM/Ky)=dM/Lyfor y∈M (3) \quad \alpha(d_{L/K} x \otimes 1) = d_{M/K} x \quad \text{for } x \in L \\ (4) \quad \beta(d_{M/K} y) = d_{M/L} y \quad \text{for } y \in M

Let VV be a vector MM-space, denote by DK(M,V)D_K(M, V) the vector space of KK-derivations of MM with values in VV and similarly introduce DK(L,V)D_K(L, V) and DL(M,V)D_L(M, V). By III, p. 571, Diagram (42) and III, p. 572, Diagram (44), we have a commutative diagram of homomorphisms of vector spaces

0→HomM(ΩL(M),V)→Hom(β,1)HomM(ΩK(M),V)→Hom(α,1)HomM(ΩK(L)⊗LM,V)↓↓↓0→DerL(M,V)→iVDerK(M,V)→rVDerK(L,V), \begin{array}{ccccccccc} 0 \to \mathrm{Hom}_M(\Omega_L(M), V) & \xrightarrow{\mathrm{Hom}(\beta, 1)} & \mathrm{Hom}_M(\Omega_K(M), V) & \xrightarrow{\mathrm{Hom}(\alpha, 1)} & \mathrm{Hom}_M(\Omega_K(L) \otimes_L M, V) \\ \downarrow & & \downarrow & & \downarrow \\ 0 \to \mathrm{Der}_L(M, V) & \xrightarrow{i_V} & \mathrm{Der}_K(M, V) & \xrightarrow{r_V} & \mathrm{Der}_K(L, V), \end{array}
where the vertical arrows are isomorphisms, iVi_V is the canonical injection and rVr_V the restriction mapping.

We thus obtain from II, p. 299, Th. 5 and II, p. 301, Prop. 10 the following proposition:

Proposition 2

— The following conditions are equivalent: a) The mapping α\alpha is injective. b) Every KK-derivation of LL into MM extends to a KK-derivation of MM into MM. c) Every KK-derivation of LL into a vector MM-space VV extends to a KK-derivation of MM into VV.

Proposition 3

— Let KK be a field, LL a pure extension of KK and (xi)i∈I(x_i)_{i \in I} a pure basis of LL (V, p. 106, Def. 2).

a) Let VV be a vector space over LL, AA a derivation of KK into VV and (vi)i∈I(v_i)_{i \in I} a family of elements of VV. There exists a unique derivation DD of LL into VV, extending AA and such that D(xi)=viD(x_i) = v_i for all i∈Ii \in I.

b) The vector LL-space ΩK(L)\Omega_K(L) of KK-differentials of LL has the family (dxi)i∈I(dx_i)_{i \in I} as basis.

Assertion b) has been proved in IV, p. 23 and Assertion a) follows directly from this.

Corollary

— Let PP be a subfield of KK. The canonical mapping α\alpha of ΩP(K)⊗KL\Omega_P(K) \otimes_K L into ΩP(L)\Omega_P(L) is injective and the family (dL/Pxi)i∈I(d_{L/P} x_i)_{i \in I}, is a basis (over LL) of a subspace of ΩP(L)\Omega_P(L) supplementary to the image of α:ΩP(K)⊗KL→ΩP(L)\alpha : \Omega_P(K) \otimes_K L \to \Omega_P(L).

Prop. 3, a) shows that every PP-derivation of KK into a vector space VV over LL extends to a PP-derivation of LL into VV; the injectivity of α\alpha then follows by Prop. 2. The second assertion of the corollary follows from the exactness of the sequence (EP,K,L)(\mathbf{E}_{P,K,L}) and Prop. 3, b) (taking Formula (4) into account).

Proposition 4

— Let KK be a field, LL a separable algebraic extension of KK and VV a vector space over LL.

a) Every KK-derivation of LL into VV is zero.

b) If AA is a derivation of KK into VV, there exists a unique derivation DD of LL into VV which extends AA.

Let DD be a KK-derivation of LL into VV. If EE is a subextension of LL of finite degree over KK, then the KK-algebra EE is etale, and so ΩK(E)=0\Omega_K(E) = 0 (VV, p. 33, Th. 3), whence D∣E=0D|_E = 0 by the universal property of ΩK(E)\Omega_K(E) (III, p. 569). Since LL is the union of a family of subextensions of finite degree over KK, we have D=0D = 0, whence a).

Let AA be a derivation of KK into VV. If D′D' and D′′D'' are two extensions of AA to a derivation of LL into VV, then the difference D′−D′′D' - D'' is a KK-derivation of LL into VV; hence it is zero by a) and so D′=D′′D' = D''.

It remains to prove the existence of an extension of AA. Zorn's Lemma (E, III, p. 20) implies the existence of a maximal extension D0D_0 of AA to a derivation defined on a subfield L0L_0 of LL containing KK.

Let xx be in LL and gg the minimal polynomial of xx over L0L_0. Since LL is algebraic and separable over KK, xx is algebraic and separable over L0L_0 (VV, p. 39, Prop. 6 and p. 39, Cor. 2); therefore xx is a simple root of gg (VV, p. 39, Prop. 5), whence g′(x)≠0g'(x) \neq 0 (IV, p. 17, Prop. 7). If we define gD0(x)g^{D_0}(x) as in VV, p. 127, there exists thus an element uu of VV such that gD0(x)+g′(x)⋅u=0g^{D_0}(x) + g'(x) \cdot u = 0; by Prop. 1 (V, p. 127) there exists a derivation DD of L0(x)L_0(x) into VV extending D0D_0 and such that D(x)=uD(x) = u. In view of the maximal character of (L0,D0)(L_0, D_0) we thus have L0(x)=L0L_0(x) = L_0, whence x∈L0x \in L_0. Since xx was arbitrary we conclude that L0=LL_0 = L.

Corollary 1

— We have ΩK(L)=0\Omega_K(L) = 0 if LL is algebraic and separable over KK.

The corollary follows from Prop. 4, a) because the vector LL-space ΩK(L)\Omega_K(L) is generated by the image of the canonical KK-derivation dL/K:L→ΩK(L)d_{L/K} : L \to \Omega_K(L).

Corollary 2

— *If L is an algebraic and separable extension of a field K, the canonical mapping α:ΩP(K)⊗KL→ΩP(L)\alpha : \Omega_P(K) \otimes_K L \to \Omega_P(L) is an isomorphism for every subfield P of K.

The mapping α\alpha is injective by Prop. 2 (V, p. 128) and Prop. 4, b); since ΩK(L)\Omega_K(L) reduces to 0 (Cor. 1) the exactness the sequence (EP.K.L)(E_{P.K.L}) implies that α\alpha is surjective.

Corollary 3

— *Let E be an extension of a field K and D a derivation of E into E, mapping K into K. If L is a subextension of E which is algebraic and separable over K, then D(L)⊂LD(L) \subset L.

Let A be the derivation of K into L which agrees with D on K. By Prop. 4 (V, p. 129) there exists a derivation D' of L into L extending A. Now we may consider D' and the restriction D'' of D to L as derivations of LL into E ; since they agree on K, we have D′=D′′D' = D'' by Prop. 4, whence

D(L)=D′′(L)=D′(L)⊂L. D(L) = D''(L) = D'(L) \subset L .

Remark

— 1) Later (V, p. 131, Cor. 3 and p. 135, Cor. 2) we shall prove a converse to Cor. 1 of Prop. 4.

2) Every algebraic extension of a prime field is separable (V, p. 37, Cor.). Since every derivation of a prime field is zero, every derivation of an algebraic extension of a prime field is zero (Prop. 4).

3. Derivations in fields of characteristic zero

Theorem 1

— *Let K be a field of characteristic 0, L an extension of K and V a vector space over L. Let A be a derivation of K into V, (xi)i∈I(x_i)_{i \in I}, a transcendence basis of L over K and (ui)i∈I(u_i)_{i \in I}, a family of elements of V. Then there exists a unique derivation D of L into V extending A and such that D(xi)=uiD(x_i) = u_i for all i∈Ii \in I.

Put E=K(xi)i∈IE = K(x_i)_{i \in I}; Prop. 3 (V, p. 128) shows that A extends to a unique derivation D0D_0 of E into V such that D0(xi)=uiD_0(x_i) = u_i for all i∈Ii \in I. The field LL is an algebraic extension of E and since L is of characteristic 0, L is separable over E (V, p. 37, Cor.). Therefore (V, p. 129, Prop. 4) D0D_0 extends to a unique derivation D of LL into V.

Corollary 1

— *Every derivation of K into V extends to a derivation of L into V

Corollary 2

— *Let E be a subextension of L and U the vector sub-L-space of ΩK(L)\Omega_K(L) generated by the differentials of elements of E. For an element x to be algebraic over E it is necessary and sufficient that dx∈Udx \in U.

For each y∈Ly \in L let D(y) be the residue class of dydy mod U. Then D is an E-derivation of L into ΩK(L)/U\Omega_K(L)/U. Since K is of characteristic 0, every algebraic extension of E is separable (V, p. 37, Cor.) ; if x∈Lx \in L is algebraic over E, we have Dx=0Dx = 0 by Prop. 4 (V, p. 129), that is, dx∈Udx \in U.

If x∈Lx \in L is transcendental over EE, there exists an E-derivation AA of E(x)E(x) into LL such that Δ(x)=1\Delta(x) = 1 (VV, p. 128, Prop. 3); by Th. 1, AA extends to an E-derivation D of LL into L. Let φ\varphi be the linear form on ΩK(L)\Omega_K(L) such that D=φ∘dD = \varphi \circ d; we have φ(dy)=0\varphi(dy) = 0 for y∈Ey \in E and φ(x)=1\varphi(x) = 1, whence dx∉Udx \notin U.

Corollary 3

— *For an element x of L to be algebraic over K it is necessary and sufficient that dx=0dx = 0 in ΩK(L)\Omega_K(L). In particular, for L to be an algebraic extension of K it is necessary and sufficient that ΩK(L)=0\Omega_K(L) = 0. This follows immediately from Cor. 2 on taking E=KE = K.

Corollary 4

— *Let K, L and M be fields of characteristic 0 such that K⊂L⊂MK \subset L \subset M: then the canonical mapping a:ΩK(L)⊗LM→ΩK(M)a : \Omega_K(L) \otimes_L M \to \Omega_K(M) is injective. This follows at once from Cor. 1 and VV, p. 128, Prop. 2.

Theorem 2

— *Let K be a field of characteristic 0, L an extension of K and (xi)i∈I(x_i)_{i \in I}, a family of elements of L. a) For (xi)i∈I(x_i)_{i \in I}, to be algebraically free over K it is necessary and sufficient that the differentials dxidx_i (for i∈Ii \in I) in ΩK(L)\Omega_K(L) should be linearly free over L. b) For L to be algebraic over K(xi)i∈IK(x_i)_{i \in I} it is necessary and sufficient that the differentials dxidx_i, for i∈Ii \in I generate the vector space ΩK(L)\Omega_K(L) over L. c) For (xi)i∈I(x_i)_{i \in I} to be a transcendence basis of L over K, it is necessary and sufficient that the family (dxi)i∈I(dx_i)_{i \in I} should be a basis of ΩK(L)\Omega_K(L) over L. For every i∈Ii \in I let EiE_i be the subfield K(xj)j∈I,j≠iK(x_j)_{j \in I, j \neq i} of L. For the family (xi)i∈I(x_i)_{i \in I}, to be algebraically free over K it is necessary and sufficient (VV, p. 108, Prop. 5) that xix_i should be transcendental over EiE_i for each i∈Ii \in I. By Cor. 2 of Th. 1 this means that for each i∈Ii \in I the differential dxidx_i is not a linear combination with coefficients in L of the differentials dxjdx_j with j≠ij \neq i in I; this latter condition just means that the family (dxi)i∈I(dx_i)_{i \in I} is free over L, whence a). Assertion b) follows at once from Cor. 2 of Th. 1 and c) is a consequence of a) and b).

Corollary

— *We have [ΩK(L):L]=tr⁡⋅deg⁡KL[\Omega_K(L) : L] = \operatorname{tr} \cdot \deg_K L when K is of characteristic 0.

4. Derivations in separable extensions

We have seen (VV, p. 122, Th. 1) that every extension LL of a field KK of characteristic 0 is separable; moreover, every derivation of K into a vector space over LL extends then to a derivation of LL (VV, p. 130, Cor. 1). More generally we have the following statement:

Theorem 3

— *Let K be a field and L an extension of K. For L to be separable over K it is necessary and sufficient that every derivation of K into a vector L-space should extend to a derivation of L. We may assume that K has characteristic ≠0\neq 0. Suppose first that L is separable over K. Let VV be a vector space over L and AA a derivation of K into VV. By Mac

Lane's criterion (V, p. 124, Remark), the fields LpL^p and KK are linearly disjoint over KpK^p. Since AA is a KpK^p-linear mapping of KK into VV, it extends in a unique way to an LpL^p-linear mapping A′A' of K[Lp]=K(Lp)K[L^p] = K(L^p) into VV. It is clear that A′A' is a derivation of K(Lp)K(L^p) into VV which vanishes on LpL^p; thus it extends (V, p. 103, Cor. 3) to a derivation of LL into VV.

Conversely, suppose that every derivation of KK with values in LL extends to a derivation of LL into LL. Let BB be a p-basis of KK (V, p. 99) and let AA be the set of families (αh)h∈R(\alpha_h)_{h \in R} with finite support consisting of integers between ()() and p−1p - 1. For each b∈Bb \in B there exists a derivation Δb\Delta_b of KK into KK characterized by Δb(b′)=δbb′\Delta_b(b') = \delta_{bb'} (Kronecker symbol) for each b′∈Bb' \in B (V, p. 103). By hypothesis there exists for each b∈Bb \in B a derivation DbD_b of LL into LL extending Δb\Delta_b. We have Db(b′)=δbb′D_b(b') = \delta_{bb'} for b,b′b, b' in BB, which proves that in Ω1/p(L)\Omega_{1/p}(L) the differentials dbdb (for b∈Bb \in B) are linearly independent over LL. Therefore (V, p 102, Th. 1), BB is p-free over LpL^p. It follows (V, p. 98, Prop. 1 and p. 124, Remark) that the extension LL of KK is separable.

Corollary

— *If LL is a separable extension of KK, the canonical mapping αp:Ωp(K)⊗KL→Ωp(L)\alpha_p : \Omega_p(K) \otimes_K L \to \Omega_p(L) is injective for every subfield PP of KK. Conversely, if there exists a perfect subfield PP of KK (for example the prime subfield of KK) such that the mapping αp\alpha_p is injective, then LL is separable over KK.

The first assertion follows from Prop. 2 (V, p. 128) and Th. 3. Conversely, let PP be a perfect subfield of KK; we have P=Pp⊂KpP = P^p \subset K^p, hence every derivation of KK into a vector LL-space is a PP-derivation; the second assertion of the corollary now follows from Cor. 2 (V, p. 130) and Th. 3.

5. The index of a linear mapping

Let LL be a field 1^1, let UU and VV be two vector spaces 2^2 over LL and f:U→Vf : U \to V an LL-linear mapping; ff is said to have finite index if the kernel NN and cokernel CC of ff have finite dimension, and the integer

χ(f)=[C:L]−[N:L] \chi(f) = [C : L] - [N : L]

is called the index of ff.

Lemma 1

— *Let UU and VV be two vector spaces of finite dimension over a field LL. Every linear mapping f:U→Vf : U \to V possesses an index equal to [V:L]−[U:L][V : L] - [U : L].

Let NN be the kernel, II the image and CC the cokernel of ff. We have C=V/IC = V/I and II is isomorphic to U/NU/N; thus we have [U:L]=[N:L]+[I:L][U : L] = [N : L] + [I : L] and [V:L]=[C:L]+[I:L][V : L] = [C : L] + [I : L], whence the result.

1^1 Not necessarily commutative. 2^2 On the left.

Lemma 2

Let f:U→Vf : U \to V and g:V→Wg : V \to W be two L-linear mappings. Iff and g each have an index, the same is true of g∘fg \circ f and we have

χ(g∘f)=χ(f)+χ(g). \chi(g \circ f) = \chi(f) + \chi(g).

Put h=g∘fh = g \circ f, and denote by N, N', N'' the kernels off, g, h respectively and by C, C', C'' their cokernels. We have N⊂N′′⊂UN \subset N'' \subset U and f(N′′)=f(U)∩N′f(N'') = f(U) \cap N'; hence there exists a linear mapping fˉ:N′′→N′\bar{f} : N'' \to N' agreeing with f on N'' and with kernel N. The canonical mapping π\pi of V onto C=V/f(U)C = V / f(U) induces a mapping π′\pi' of N' into C whose kernel is f(U)∩N′=fˉ(N′′)f(U) \cap N' = \bar{f}(N''). By passage to quotients g defines a mapping gˉ\bar{g} of C=V/f(U)C = V/f(U) into C′′=W/g(f(U))C'' = W/g(f(U)) whose kernel is clearly (N′+f(U))/f(U)=π′(N′)(N' + f(U))/f(U) = \pi'(N'). Finally the canonical mapping ρ\rho of C′′=W/g(f(U))C'' = W/g(f(U)) onto C′=W/g(V)C' = W/g(V) has the kernel g(V)/g(f(U))=gˉ(C)g(V)/g(f(U)) = \bar{g}(C). To sum up, we have established the exactness of the sequence

0→N→iN′′→fˉN′→π′C→gˉC′′→ρC′→0 0 \to N \xrightarrow{i} N'' \xrightarrow{\bar{f}} N' \xrightarrow{\pi'} C \xrightarrow{\bar{g}} C'' \xrightarrow{\rho} C' \to 0

(where i is the canonical injection of N in N").

!Diagram showing the relationships between N, N', N'', U, V, W, C, C', C'', with arrows indicating mappings and kernels

Fig. 1

By hypothesis N, N', C and C' are of finite dimension; hence the same is true of N'' and C. By Cor. 2 (II, p. 295) we then have

[N:L]−[N′′:L]+[N′:L]−[C:L]+[C′′:L]−[C′:L]=0 [N : L] - [N'' : L] + [N' : L] - [C : L] + [C'' : L] - [C' : L] = 0

whence χ(h)=χ(f)+χ(g)\chi(h) = \chi(f) + \chi(g).

6. Differential properties of finitely generated extensions

Theorem 4

Let P be a perfect field, L an extension of P and K a subextension of L; we suppose that L is a finitely generated extension of K. Then the canonical L-linear mapping a:Ωp(K)⊗KL→Ωp(L)a : \Omega_p(K) \otimes_K L \to \Omega_p(L) has index equal to the transcendence degree of L over K.

If EE and FF are two subextensions of LL such that E⊂FE \subset F, we denote by α(F/E)\alpha(F/E) the canonical FF-linear mapping of Ωp(E)⊗EF\Omega_p(E) \otimes_E F into Ωp(F)\Omega_p(F) and when it is defined, the index of α(F/E)\alpha(F/E) will be written d(F/E)d(F/E). If E,FE, F and GG are three subextensions of LL such that E⊂F⊂GE \subset F \subset G, we have a commutative diagram

Ωp(F)⊗FG→α(G/F)Ωp(G)↑α(F/E)⊗FIdG↑α(G/E)(Ωp(E)⊗EF)⊗FG→βΩp(E)⊗EG \begin{array}{ccc} \Omega_p(F) \otimes_F G & \xrightarrow{\alpha(G/F)} & \Omega_p(G) \\ \uparrow \alpha(F/E) \otimes_F \mathrm{Id}_G & & \uparrow \alpha(G/E) \\ (\Omega_p(E) \otimes_E F) \otimes_F G & \xrightarrow{\beta} & \Omega_p(E) \otimes_E G \end{array}

where β\beta is the canonical isomorphism defined in Prop. 2 (II, p. 278). Since the index is clearly invariant under extension by scalars and the index of an isomorphism is zero, Lemma 2 (VV, p. 133) shows that the index d(G/E)d(G/E) is defined when d(F/E)d(F/E) and d(G/F)d(G/F) are and then

d(G/E)=d(G/F)+d(F/E). d(G/E) = d(G/F) + d(F/E).

Since LL is a finitely generated extension of KK, Formula (5) and the Cor. of VV, p. 111 shows that it is enough to consider the case where xx exists such that L=K(x)L = K(x); further if xx is algebraic over KK, there exists a power qq of the characteristic exponent of LL such that xqx^q is separable algebraic over KK (VV, p. 44, Prop. 13). So it is enough to establish the equality d(L/K)=tr⁡.deg⁡KLd(L/K) = \operatorname{tr} . \deg_K L in the three special cases below:

a) xx is transcendental over KK: then aa is injective and its cokernel is of rank 1 over LL (VV, p. 129, Cor.) ; so we have d(L/K)=1d(L/K) = 1 and also tr⁡.deg⁡KL=1\operatorname{tr} . \deg_K L = 1.

b) xx is separable algebraic over KK: then aa is bijective (VV, p. 130, Cor. 2), whence d(L/K)=0d(L/K) = 0; clearly also tr⁡.deg⁡KL=0\operatorname{tr} . \deg_K L = 0.

c) The field LL is of characteristic p≠0p \neq 0, x∉Kx \notin K and xp=ax^p = a is in KK: the cokernel CC of aa is isomorphic to ΩK(L)\Omega_K(L), and since {x}\{ x \} is a pp-basis of LL over KK, the space CC is of dimension 1 over LL (VV, p. 102, Th. 1). Since aa is a pp-th power in LL, we have dL/pa=0d_{L/pa} = 0 and the kernel of aa contains the subspace RR of U=Ωp(K)⊗KLU = \Omega_p(K) \otimes_K L generated by dK/pa⊗1d_{K/pa} \otimes 1. For each y∈Ky \in K let Δ(y)\Delta(y) be the residue class of dK/py⊗1d_{K/p}y \otimes 1 mod RR; then AA is a PP-derivation of KK into U/RU/R such that Δ(a)=0\Delta(a) = 0. Prop. 5 (VV, p. 101) shows that AA extends to a PP-derivation DD of LL into U/RU/R. Hence there exists an LL-linear mapping β:Ωp(L)→U/R\beta : \Omega_p(L) \to U/R such that D=β∘dL/PD = \beta \circ d_{L/P} and β∘a\beta \circ a is the canonical mapping of UU onto U/RU/R. This proves RR to be the kernel of aa. Since PP is perfect, we have P(KP)=KpP(K^P) = K^p, whence a∉P(KP)a \notin P(K^P) and finally dK/pa≠0d_{K/pa} \neq 0 (VV, p. 103, Prop. 6). The kernel and cokernel of aa are thus of dimension 1, whence d(L/K)=0d(L/K) = 0, and we also have tr⁡.deg⁡KL=0\operatorname{tr} . \deg_K L = 0.

Corollary 1

— *Let LL be a finitely generated extension of a field KK of transcendence degree ss. The vector space ΩK(L)\Omega_K(L) is of dimension ≥s\geq s over LL, with equality if and only if LL is separable over KK.

Let P be the prime subfield of K. Let N be the kernel of α, then by exactness of the sequence (EP,K,L)(\mathbf{E}_{P,K,L}) (V, p. 128) and Th. 4, we have [ΩK(L):L]=s+[N:L][ \Omega_K(L) : L ] = s + [N : L]; by V, p. 132, Cor., the extension L of K is separable if and only if N=0N = 0. The corollary now follows.

Corollary 2

— Let L be a finitely generated extension of a field K. For L to be algebraic and separable over K it is necessary and sufficient that ΩK(L)=0\Omega_K(L) = 0. This follows immediately from Cor. 1.

Corollary 3

— Let K be a field of characteristic p≠0p \neq 0 and L a finitely generated extension of K of transcendence degree s. If [K:Kp][K : K^p] is finite, the same is true of [L:Lp][L : L^p] and we have [L:Lp]=ps[L : L^p] = p^s. [K:Kp][K : K^p]. Let P be the prime subfield of K. If [K:Kp][K : K^p] is finite, then the vector space ΩP(K)=ΩKp(K)\Omega_P(K) = \Omega_{K^p}(K) is of finite dimension m over K, and we have [K:Kp]=pm[K : K^p] = p^m (V, p. 103, Th. 2); the vector space ΩP(K)⊗KL\Omega_P(K) \otimes_K L is then of finite dimension m over L. Further, since K is of finite degree over KpK^p, the field K(Lp)K(L^p) is of finite degree over Kp(Lp)=LpK^p(L^p) = L^p; since the field L is a finitely generated extension of K and is algebraic over K(Lp)K(L^p), it is an extension of finite degree of K(Lp)K(L^p) (V, p. 18, Th. 2); we thus conclude that L is of finite degree over LpL^p (V, p. 10, Th. 1). Then ΩP(L)\Omega_P(L) is a vector space of finite dimension n over L and we have [L:Lp]=pn[L : L^p] = p^n (V, p. 103, Th. 2). By Lemma 1 (V, p. 132) the L-linear mapping α:ΩP(K)⊗KL→ΩP(L)\alpha : \Omega_P(K) \otimes_K L \to \Omega_P(L) therefore has index n−mn - m, whence n−m=sn - m = s by Th. 4 (V, p. 133) and pn=ps⋅pmp^n = p^s \cdot p^m.

Remark 1

Let K be a field of characteristic p≠0p \neq 0 and L an extension of K. We have ΩK(L)=0\Omega_K(L) = 0 if and only L=K(Lp)L = K(L^p) (V, p. 103, Prop. 6). Therefore if L is finitely generated over K, then it is an algebraic and separable extension if and only if we have L=K(Lp)L = K(L^p). When L is not finitely generated over K, this result no longer holds generally as is shown by the case where L is the perfect closure of K.

Remark 2

Let K be a field, F1,…,FmF_1, \ldots, F_m polynomials in K[X1,…,Xn]K[X_1, \ldots, X_n] and L an extension of K generated by the elements x1,…,xnx_1, \ldots, x_n. Suppose that the polynomials F1,…,FmF_1, \ldots, F_m generate the ideal of K[X1,…,Xn]K[X_1, \ldots, X_n] consisting of all polynomials F such that F(x1,…,xn)=0F(x_1, \ldots, x_n) = 0. From Prop. 1 (V, p. 127) and the universal property of the module of differentials (III, p. 569) we easily deduce the following result: the vector space ΩK(L)\Omega_K(L) over L is generated by dx1,…,dxndx_1, \ldots, dx_n; we have the relations

∑i=1nDiFj(x1,…,xn)⋅dxi=0(for 1≤j≤m); \sum_{i=1}^n D_i F_j(x_1, \ldots, x_n) \cdot dx_i = 0 \quad \text{(for } 1 \leq j \leq m \text{)} ;

finally if u1,…,unu_1, \ldots, u_n are elements of L such that ∑i=1nui⋅dxi=0\sum_{i=1}^n u_i \cdot dx_i = 0, there exist elements v1,…,vmv_1, \ldots, v_m of L such that ui=∑j=1mDiFj(x1,…,xn) vju_i = \sum_{j=1}^m D_i F_j(x_1, \ldots, x_n) \ v_j for 1≤i≤n1 \leq i \leq n. Let us denote by r the rank of the matrix (DiFj(x1,…,xn))(D_i F_j(x_1, \ldots, x_n)) with n rows and m columns; let s be the transcendence degree of LL over KK. Then we have [ΩK(L):L]=n−r[ \Omega_K(L) : L ] = n - r. Therefore the extension LL of KK is separable if and only if r+s=nr + s = n (Cor. 1). and it is algebraic and separable if and only if r=nr = n (Cor. 2).

7. Separating transcendence bases

Definition 1

Let LL be an extension of a field KK. A transcendence basis BB of LL over KK is said to be separating if the algebraic extension LL of K(B)K(B) is separable.

If KK has characteristic 0, every transcendence basis of LL over KK is separating because every algebraic extension of a field of characteristic 0 is separable (VV, p. 37, Cor.). If an extension admits a separating transcendence basis, it is separable (VV, p. 121, Prop. 6 and p. 122, Prop. 9). The following theorem shows that every finitely generated separable extension admits a separating transcendence basis; this restriction is essential (VV, p. 177, Ex. 1).

Theorem 5

Let KK be a field, LL an extension of KK and (xi)i∈I(x_i)_{i \in I}, a family of elements of LL. If the family (xi)i∈I(x_i)_{i \in I}, is a separating transcendence basis of LL over KK, then the family (dxi)i∈I(dx_i)_{i \in I}, is a basis of the vector space ΩK(L)\Omega_K(L) over LL. The converse holds if LL is a finitely generated separable extension of KK.

Put M=K(xi)i∈IM = K(x_i)_{i \in I}, and denote by α\alpha the canonical mapping of ΩK(M)⊗ML\Omega_K(M) \otimes_M L into ΩK(L)\Omega_K(L). If (xi)i∈I(x_i)_{i \in I} is a separating transcendence basis of LL over KK, the family (dM/Kxi)i∈I(d_{M/K} x_i)_{i \in I}, is a basis of the vector MM-space ΩK(M)\Omega_K(M) (VV, p. 128, Prop. 3) and α\alpha is an isomorphism of vector LL-spaces since LL is algebraic and separable over MM (VV, p. 130, Cor. 2). Since α(dM/Kxi⊗1)=dL/Kxi\alpha(d_{M/K} x_i \otimes 1) = d_{L/K} x_i, the family (dL/Kxi)i∈I(d_{L/K} x_i)_{i \in I} is thus a basis of ΩK(L)\Omega_K(L) over LL.

Conversely, suppose that LL is a separable finitely generated extension of KK and that the family (dL/Kxi)i∈I(d_{L/K} x_i)_{i \in I} is a basis of the vector space ΩK(L)\Omega_K(L) over LL. By Cor. 1 of VV, p. 134, the transcendence degree of LL over KK is equal to the dimension of ΩK(L)\Omega_K(L) over LL, hence to the cardinal of II. From the exact sequence (EK,M,L)(E_{K,M,L}) (VV, p. 128) we have ΩM(L)=0\Omega_M(L) = 0; since LL is a finitely generated extension of MM, Cor. 2 of VV, p. 135 shows that LL is algebraic and separable over M=K(xi)i∈IM = K(x_i)_{i \in I}; since the transcendence degree of LL over KK is finite and equal to the cardinal of II, the family (xi)i∈I(x_i)_{i \in I} is a transcendence basis of LL over KK (VV, p. 110, Cor. 1).

Corollary

Let LL be a separable finitely generated extension of KK and let SS be a subset of LL such that L=K(S)L = K(S). Then there exists a separating transcendence basis BB of LL over KK, contained in SS.

Since ΩK(L)\Omega_K(L) is generated by the differentials of the elements of SS, there exist ss elements x1,…,xsx_1, \ldots, x_s of SS such that (dx1,…,dxs)(dx_1, \ldots, dx_s) is a basis of ΩK(L)\Omega_K(L) over LL. Now it suffices to apply Th. 5.

Remark

— Let LL be a separable finitely generated extension of a field KK of characteristic p≠0p \neq 0; there may exist transcendence bases of LL which are not separating. It is enough to observe that {Xp}\{X^p\} is a transcendence basis of K(X)K(X) but K(X)K(X) is a p-radical extension of degree pp of K(Xp)K(X^p).

Proposition 5

— *Let L and M be two extensions of a field K contained in a given extension and algebraically disjoint over K. If M is separable over K, then L(M) is separable over L.

It is enough to show that for every finite subset S of M, L(S)L(S) is separable over L (V, p. 122, Prop. 8). Let S be a finite subset of M. Since the field K(S)K(S) is separable over K, it has a separating transcendence basis B (Cor. of Th. 5). Since L and M are algebraically disjoint over K, B is a transcendence basis of L(B)L(B) over L (V, p. 113, Prop. 12). Further, every element of S is algebraic and separable over K(B)K(B), hence over L(B)L(B) (V, p. 39, Cor. 2). We deduce (V, p. 39, Prop. 6) that L(S)=L(B)(S)L(S) = L(B)(S) is algebraic and separable over L(B)L(B), hence L(S)L(S), is separable over L.

Corollary

— *If L and M are separable extensions, algebraically disjoint over K, then the field K(L∪M)K(L \cup M) is separable over K.

For K(L∪M)=L(M)K(L \cup M) = L(M) is separable over L by Prop. 5 (because M is separable over K) and L is separable over K, whence the corollary (V, p. 122, Prop. 9).

The hypothesis that the extensions L and M are algebraically disjoint is indispensable in Prop. 5 and its corollary. For let K be an imperfect field of characteristic p≠0p \neq 0 and E an extension of the form K(x,a)K(x, a) with x transcendental over K and a p-radical of height 1 over K ; put L=K(x)L = K(x) and M=K(x+a)M = K(x + a). Then x+ax + a is transcendental over K (if not, x=(x+a)−ax = (x + a) - a would be algebraic over K) and the fields L and M are separable over K. However, K(L∪M)=K(x,a)K(L \cup M) = K(x, a) is p-radical of degree pp over L=K(x)L = K(x) and is not separable over L, nor over K.

Exercises

See the exercises for § 16.

Transcribed from the printed edition, 2003, Springer. Not translated.