The Elementary Theory of the Frobenius Automorphisms

(1) If P is transformally prime, R/P is a difference domain, and \sigma extends naturally to the field of fractions: if b \neq 0 then \sigma(b) \neq 0, and one can define \sigma(a/b)=\sigma(a)/\sigma(b). Conversely, if L is a difference field, and R \to_h R/P \subset L is a homomorphism, if a \notin P then h(a) \neq 0 so h(\sigma(a)) = \sigma(h(a)) \neq 0, and thus \sigma(a) \notin P.

(2) If P is transformally prime, clearly P' is too. If P' is transformally prime, then the multiplicative set h(R' \setminus P') \subseteq K' \setminus (0) \subset K \setminus (0) is closed under \sigma, so \sigma extends to the localization of h(R) by this set, i.e. to M = K'(R/P). But since K' \subseteq M \subseteq K and K is algebraic over K', M is a field, hence M = K. \Box

The statement regarding \sqrt{0} is left to the reader. If a^{m \sigma^i} = 0, and m \leq m', i \leq i', then a^{m' \sigma^{i'}} = 0. It follows that I is closed under \sigma and under multiplication, and addition (if a^{m \sigma^i} = 0 and b^{m \sigma^i} = 0 then (a+b)^{2m \sigma^i} = 0.) Moreover if a a^\sigma \in I, then a^{(\sigma+1)m \sigma^i} = 0. So b b^\sigma = 0 where b =a^{m \sigma^i}. As R is well-mixed, b^{\sigma} b^\sigma = 0. Thus a^{2m \sigma^{i+1} } = 0, so a \in I. \Box

: Clear.

We write Ann(a/I) for \{b \in R: ab \in I \}, Ann(a) = \{b \in R: ab=0\}.

1. By ?, ?, Ann(b) and Ann(b/\sqrt{0}) are difference ideals of R. The statement about the intersection with S is therefore obvious.

2. Let S be a Noetherian subring of R. Then \sqrt{0}_S = q_1 \cap \ldots \cap q_l for some minimal primes q_1 \cap \ldots \cap q_l; and q=q_i for some i, say q=q_1. Let b_i \in q_i, b_i \notin q for i >1, and let b = b_2 \cdot \ldots \cdot b_l. Then q = Ann(b/\sqrt{0}) \cap S.

3. Let p be cofinally minimal, a \in p. Let F = \{a,\sigma(a)\}, and let S be a Noetherian subring of R containing F with p \cap S a minimal prime of S. By (2), \sigma(a) \in (p \cap S), so \sigma(a) \in p.

\Box

We may assume I=0. If a \neq 0, we must find an algebraically prime difference ideal q with a \notin q. By compactness (see below), we may assume here that R is finitely generated. Find using ? a cofinally minimal prime q with a \notin q. By ? it is an algebraically prime difference ideal. (Here is the model theoretic compactness argument. Assume for each finitely generated difference subring S of R we are given an algebraically prime difference ideal q_S with a \notin q_s. One can then choose an ultrafilter on the set of such subrings, concentrating for each finite F on \{S: F \subseteq S\}. Then R embeds into the ultraproduct of the S. But they come with an algebraically prime difference ideal q = the ultraproduct of the q_S, and the pullback is as required.) \Box

One direction is immediate. For the other, assume a^n \neq 0 for n \in \hbox{\msbmten N}[\sigma]. Let I be a maximal well-mixed ideal with a^n \notin I, n \in \hbox{\msbmten N}[\sigma]. By maximality, and using Lemma ?, \sqrt[\sigma] I = I. So I = \cap p_j, p_j transformally prime ideals. Since a \notin I, a \notin p_j for some j. \Box

Assume first that X is Noetherian (as a topological space). Say f(q) = \frac{c}{d} \in R_q. If ce = 0, e \notin q, then f(q)=0 \in R_q and we are done. If bc = 0, b \notin p, then bf(q) = 0 \in R_q and we are done again. Thus we may assume Ann(c) \subset p \cap q.

Since R is well-mixed, so is Ann(c); by ?, I=\sqrt[\sigma] Ann(c) is perfect. Thus I = p_1 \cap \ldots \cap p_n for some transformally prime ideals p_i.

If p_i \subset p \cap q, then f(p_i) = 0 \in R_{p_i} (since p_i \subset p) and f(p_i) = c/d \in R_{p_i} (since p_i \subset q), hence c=0 \in R_{p_i}, contradicting Ann(c) \subset p_i. Thus for each i, there exists a_i \in p_i, a_i \notin q or else b_i \in p_i, b_i \notin p. Let a be the product of the a_i, b the product of the b_i. So a \notin q, b \notin p, ab \in I; thus (ab)^n \in Ann(c), some n \in \hbox{\msbmten N}[\sigma]f. Replacing a,b by a^n,b^n, we obtain a \notin q, b \notin p, ab \in Ann(c). So abc=0, hence bc = 0 \in R_q, and thus b f(q) = 0 \in R_q. \Box

By definition of {\cal O}_X (X) for some a,b \in R, b \notin p, we have: f(p) = \frac{a}{b} \in R_p. If bf-a satisfies the lemma, then so does f. So we may assume f(p) = 0 \in R_p. By ?, for each q \in {\rm Spec^\sigma \, } R there exists b_q \in R, b_q \notin p, with b_q f (q) = 0 \in R_q. But by definition of {\cal O}_X (X) again, for some c,d,d' with d,d' \notin q we have f(q') = \frac{c}{d} \in R_{q'} whenever d' \notin q'. The equation b_q f(q) = 0 \in R_q means that there exists d'' \notin q with d''b_q c= 0. So b_q f(q') = 0 \in R_{q'} whenever dd'd'' \notin q' Let U_q = \{q': dd'd'' \notin q' \}. By compactness of X, finitely many open sets U_q cover X; say for q_1,\ldots,q_n. Let b = b_{q_1} \cdot \ldots b_{q_n}. Then bf(q' ) = 0 \in R_{q'} whenever q' \in X. \Box

By ?, i: R \to {\bar R} is an embedding. Let \bar{p} \in {\rm Spec^\sigma \, } \bar{R}, p = \bar{p} \cap R. By ?, there exist a,b \in R, b \notin p, with bf-a = 0 \in \bar{R}. So f \in \bar{p} iff a \in p. This shows that i^* is injective, and so induces a bijection of points; similarly ?, ? show that i induces an isomorphism R_p \to {\bar R}_{\bar p }. \Box

If p^{{\rm inv}} exists, then clearly a \in p^{{\rm inv}} iff \sigma^n(a) \in p for some n. So define p' in this way: p' = \cup_{n \in \hbox{\msbmten N}} \sigma^{-n}(p). It is easy to check that p' is prime, \sigma^{-1}(p')=p', and p' \cap D = p.

Any difference ring homomorphism of R into a difference ring without zero divisors factors through R/rad_{wm}(R), so the first point is clear. Now assume R is well-mixed, of total dimension n. Clearly n \leq t_2, while t_2 \leq t_4 since one can lift a a transcendence basis from the quotient ring to a set of elements of R, necessarily independent over k. We have t_4 \leq t_3 by standard commutative algebra. Finally, t_3 \leq t.dim(R): R contains a free polynomial ring R_1 of rank t_3. By ?, there exists a cofinally minimal (prime) ideal p of R with p \cap R_1 = (0); by ?, p is a difference ideal. The quotient R/p clearly has field of fractions with transcendence degree \geq n. \Box

Evidently (1) \Rightarrow (2) \Rightarrow (3) . To prove that (3) \Rightarrow (1) , we may assume X = {\rm Spec^\sigma \, } R, R a finitely generated difference k-algebra.

By ?, R does not contain a copy of the difference-polynomial ring in one variable k[t,t^\sigma,\ldots].

Let r_1,\ldots,r_n be generators for R, r_{ij} = \sigma^j(r_i). Let \Xi be the set of triples (h,D,L) as in the definition of total dimension. For each i, and each (h,D,L) \in \Xi, for some m, there exists a nontrivial polynomial relation F(h(r_i),\ldots,h(r_{i,m})) = 0, F \in k[X_0,\ldots,X_m].

By compactness, there exists m and finitely many F_1,\ldots,F_r \in k[X_0,\ldots,X_m] such that for all i and all (h,D,L) \in \Xi, for some j \leq r, F_j(h(r_i),\ldots,h(r_{i,m})) = 0. By the claim below, tr. deg._k(L) \leq mn, finishing the proof.

Claim Let k be a difference field, R=k[a,a^\sigma,\ldots] a difference k-algebra with no zero divisors, L the field of fractions of R. Write a_i = \sigma^i(a), and assume F(a_0,\ldots,a_m)=0, 0 \neq F \in k[X_0,\ldots,X_m]. Then tr. deg._k (L) \leq (m).

By the Claim of ?, we have \sigma^j(a_i) \in (L_0)^a for each i and all j. \Box

As R is a finitely generated k-algebra, the equality follows follows from Lemma ?. A finitely generated k-algebra of Krull dimension 0 is finite dimensional over k (explicitly: let I be the nil ideal of R; I = \cap_{i=1}^n p_i, with p_i prime. R/p_i is a finite field extension of k; so {\rm dim}_k R/p_i < \infty, hence also {\rm dim}_k R/I < \infty. Now R is Noetherian, so I is finitely generated; thus I^l/I^{l+1} is a finite-dimensional R/I-space for each k; so {\rm dim}_k I^l/I^{l+1} < \infty for each l \leq r. It follows that {\rm dim}_k R < \infty.) \Box

If \sum c_i x_i = \sum c_i y_i, x,y \in X distinct, then c = (c_1,\ldots,c_n) has transformal dimension <n over K(x,y). Thus \{c: (\exists x \neq y \in X) c \cdot x = c \cdot y \} has transformal dimension \leq (n-1) + 2 trans. dim (X) = n-1. So for any c, outside a proper difference subscheme of \hbox{\msbmten A}^n, H(x) = \sum c_i x_i is injective. \Box

Say X \subset \hbox{\msbmten P}^n. The set of hyperplanes passing through some point of X is contained in a difference scheme of transformal dimension n-1. So there exists a hyperplane defined over k and avoiding X. Thus we may assume X \subset \hbox{\msbmten A}^n. Now ? applies.

(Improve this to: X is isomorphic to a difference subscheme of \hbox{\msbmten P}^1.)

There exists a surjective homomorphism of transformal K-algebras h: R {\otimes}_D K \to D', D' has field of fractions K', tr. deg._K K' = m. So tr. deg_k K' \geq k+m, and h(R) generates K' as a field. \Box

Let a be the transformal dimension of {\rm Spec^\sigma \, } D over k; this equals the transformal dimension of {\rm Spec^\sigma \, } \sigma^n(D) over \sigma^n(k), for any n. Let b_n be the dimension growth degree of {\rm Spec^\sigma \, } \sigma^n(D) as a difference subscheme of {\rm Spec \, } \sigma^n(R), a variety over \sigma^n(k). In other words, let R_{n,m} be the subring of D generated by \sigma^n(R) \cup \ldots \cup \sigma^{n+m-1}(R), K_{n,m} the field of fractions of R_{n,m}, c(n,m) the transcendence degree of K_{n,m} over \sigma^n(k). Then for any n, for large enough m, c(n,m) = am+b_n. We have b=b_0 \geq b_1 \geq \ldots, so b_i=b_{i+1} for some i \leq b. The lemma now follows by applying the following claim to \sigma^i(D).

Claim If b=b_1, then D is a difference domain.

Applying ? in R/I, we find d,I_0 such that I R[d {^{-1}}] = I_0 R[d {^{-1}}]. But I_0 R[d {^{-1}}] = \{a / d^n: a \in I_0\}. So for any a \in I, for some b \in I_0, a=b/d^n; ad^n -b =0; ad^{n+m} -bd^m =0; ad^N \in I_0. \Box

Let K be the field of fractions of D, K ^{inv} the inversive hull of K. Let a_1,\ldots,a_r be generators of D, a = (a_1,\ldots,a_r). In this proof, for Y \subseteq K^{inv}, write k[Y] for the k-algebra of K ^{inv} generated by Y, k[Y]_\sigma = k[Y,\sigma(Y),\ldots]. Let D^- = k[\{\sigma^n(a): n < 0 \}]. By Corollary ?, there exists 0 \neq e \in D^- and a finite f \subset D^-[e {^{-1}}][X] such that the natural map induces an isomorphism

(*) D^-[e {^{-1}}, a] \cong D^-[e {^{-1}}][X]/(f),

and indeed R'[Y {^{-1}}][a] \cong R'[Y {^{-1}}][X]/(f) whenever R \subseteq R' \subseteq D and e \in Y \subseteq R' \setminus (0).

The element e as well as all coefficients of the polynomials in f lie in k[\sigma {^{-1}} (a),\ldots, \sigma ^{-l}(a)] for some l. Let d=\sigma^l(e). For n \geq 0, let R_{-n} = k[a^{\sigma^{-1}},\ldots, a^{\sigma ^{-l-n}},e {^{-1}}, \ldots, \sigma^{-n}(e {^{-1}})]. By (*),

Note that \sigma (e) \in R_{-l}[a]; say \sigma (e) = h(a) with h \in R_{-l}[T]. Let c = (\sigma (e))^{-1}, and let g = (f,Yh(X)-1). Then

Let

Applying \sigma^{n+l} to (**), we see that R_{n+1} \cong R_n[X,Y]/(\sigma^{n+l}(g)) naturally. Let H:k[X,Y]_\sigma \to D[d {^{-1}}]_\sigma be the difference ring homomorphism with X \mapsto a, Y \mapsto d {^{-1}}. Then the kernel of H is generated by the kernel of H | k[X,Y,\ldots,X^{\sigma^l},Y^{\sigma^l}] together with \sigma^l(g),\sigma (\sigma^l(g)), \ldots. As a difference ideal, this is finitely generated. \Box

This follows from ?, applied to R/P. \Box

Let H be a finite set of generators of R as a difference k-algebra. For a \in H, \sigma^m(a) \in k \cdot 1 for some m; let T(a) = \{a,\sigma(a),\ldots,\sigma^{m-1}(a)\}. Then H' = \cup_{a \in H} T(a) is a finite set of generators of R over as a k- algebra.

Now assume R is a difference domain. By Corollary ?, for some 0 \neq a \in R, S[a {^{-1}}]_\sigma is a finitely presented R[a {^{-1}}]_\sigma-algebra.

Now in general, if S'' is a difference R''-algebra, and a finitely presented R''-algebra, then it is a finitely presented R''-difference algebra: if S''=R''[b], b=(b_1,\ldots,b_n), then \sigma(b) = f(b) for some f \in R[X], and it suffices to add to the R''-algebra presentation the relation \sigma(b) = f(b). \Box

In such a descending chain, the transformal dimension must be non-increasing, and eventually stabilize; after that point, the dimension growth degree cannot increase, and eventually stabilizes too. By ?, if X \subset Y \subset V are algebraically integral, and have the same transformal dimension and dimension growth degree, then X=Y. \Box

(Affine case.) We may write X = {\rm Spec^\sigma \, } ([\sigma]_D D[X]/ J), where D[X] is a polynomial ring over D in finitely many variables, and J a prime ideal of [\sigma]_D D[X]. By Proposition ?, J is finitely generated as a well-mixed, algebraically reduced ideal. Adding variables, we may assume J is so generated by J_0 = J \cap D[X,\sigma(X)]. Let \widetilde{X} = {\rm Spec^\sigma \, } ([\sigma]_D D[X]/ J_0). \Box

Using Lemma ?, we can prove the statement by Noetherian induction on well-mixed, radical ideals. Let p be such an ideal. If p is algebraically prime, we are done; otherwise, say ab \in p, a,b \notin p; let p_1 = Ann(a/p), p_2 = Ann(p_1/p); then p_1,p_2 are well-mixed, radical ideals bigger than p, with p_1 p_2 \subseteq p \subseteq p_{1} \cap p_{2}. Let q = p_{1} \cap p_{2}; then q^{2 } \subseteq p_1 p_2 \subseteq p, so as p is radical, q \subseteq p; and thus p = p_{1} \cap p_{2}. Now by induction, each p_{i} is the intersection of finitely many algebraically prime difference ideals; hence so is p. \Box

We take the case: S \to V is smooth (the other being similar.)

Since S^{\sigma^n} \to {V^{\sigma^n}} is smooth, and smooth morphisms are preserved under base change, S[n+1] \to S[n] is also smooth. Compositions of smooth morphisms are also smooth, so for each n, S[n] \to V is a smooth morphism, and indeed S[n] is smooth over K.

It follows that each S[n] is a disjoint union of irreducible components. By the dimension theorem (in the smooth variety V \times V^\sigma \times \ldots \times V^{\sigma^n}), each component has dimension \geq d+ne. On the other hand by adding up relative dimensions we see that {\rm dim \, } S[n] \leq d+ne, so equality holds.

Since the components of the S[n] are disjoint, letting S[\infty] be the projective limit of the S[n], we see that the pro-algebraic variety S[\infty] is a disjoint union of irreducible pro-algebraic varieties C; C is an inverse limit of components C[n] of S[n]. Now by definition, C[n+1] projects to C[n]; but one can also project to V^\sigma \times \ldots \times V^{\sigma^n}; the image is a component which is either contained in, or disjoint from, C[n]^\sigma. In the former case, for any difference subvariety W of Z, W[n+1] \cap C[n+1] = \emptyset; call such a pro-component empty of difference points. If the latter case holds for each n, then C defines a perfect difference scheme, and so a component W(C) of Z. Thus Z is the disjoint union of the various (nonempty) W(C). It follows that Z is perfectly reduced.

Finally, assume e=0. The additivity of degree may be proved by induction on \deg_{cor} (S). If for all m, the S[m+1] is absolutely irreducible, then Z is transformally integral, and \deg_{lim}(Z) = \deg_{cor}(S). Otherwise, consider the minimal m such that S[m+1] is reducible. Let Y = S[m], and let

Let \{T_j\} be the components of T projecting dominantly to Y. They map finite-to-one to Y^\sigma, hence dominantly. Then

Since \deg_{cor} (T_j) < \deg_{lim} (Z) for each j,

where the Z' here range over the components of Z , Zariski dense in X, such that for a \in Z'

Since each Zariski dense Z' falls into a unique T_j in this sense, the required equation follows.

\Box

The uniqueness (2) is clear, since if Z,Z' are two candidates, each generic point of Z must lie on Z', and vice versa. To show existence (1), and (3), we use Noetherian induction on Y. We may assume Y is perfectly reduced. We may also assume Y is irreducible, since if it has a number of components Y_j, then we can let D_m(X/Y) = \cup_j D_m(f^{-1}Y_j/Y_j). If X/Y has generic total dimension \geq m, let D_m(X/Y) = Y. By ?, Y has total dimension \leq d-m in this case. Otherwise, Y has total dimension \leq m-1. By ?, there exists a proper closed difference subscheme Y' of Y such that for a \in Y \setminus Y', X_a has total dimension \leq m-1. Now D_m(f^{-1}Y'/Y') exists by Noetherian induction, and we can let D_m(X/Y)=D_m(f^{-1}Y'/Y').

Finally, (4) is clear since f^{-1}D_m(X/Y) is a difference subscheme of X of the same total dimension d (by the first part of (1).) \Box

F \subset \sigma(K) \subset K, and tr. deg._F (\sigma(K)) = tr.deg._F (K). \Box

By Lemma ?, we have:

Claim Let K be an inversive difference field, L a difference field extension of K of finite transcendence degree. Then L^{inv} is an algebraic extension of L.

(1) follows: since the transcendence degree of L over K is finite, so is that of LK^{inv} over K^{inv}, and the claim applies.

(2) The "if" direction follows from the transitivity properties of linear disjointness. Assume M/K is transformally separable. Then L/K is a fortiori transformally separable. We need to prove that M/L is transformally separable. M is linearly disjoint from the compositum K^{inv}L over L. We must show that K^{inv}M is linearly disjoint from L^{inv} over K^{inv}L. Since M/L is a regular field extension , and M is linearly disjoint from LK^{inv} over L, MK^{inv}/LK^{inv} is a regular extension. In other words MK^{inv} is linearly disjoint over LK^{inv} from the algebraic closure of LK^{inv}. By (1), L^{inv} \subset (LK^{inv})^{alg}.

(3) We may assume here that K is inversive. We have

Yet M is linearly disjoint from L^{\sigma^{-1}} over L, so M^\sigma is linearly disjoint over L^\sigma from L, and

Similarly

\Box

Let Y= B_{k+1}X, and let r=r_{k+1}: X \to Y be the reduction sequence map.

Let Y_k = D_k(r); cf. ?.

Let Mlt_kX = r^{-1}Y_k. (Or the underlying reduced scheme).

(1) By Lemma ? below, we have to show that if L' is an inversive difference ring and y is an L'-valued point of B_{k+1}(X \setminus Mlt_kX), then X_y has total dimension < k over L'. Now y \notin Y_k; by definition of Mlt_k(Y), and ?, the total dimension of X_y is <k.

(2),(3) come from ? (3), (4).

\Box

By ?, we may take L to be inversive.

We may assume X = {\rm Spec^\sigma \, } R, R a well-mixed difference ring. Let R_n = \sigma^n(R). Let y be the unique extension of y_{n'} to a difference ring morphism y: R \to L; let y_m = y | R_m. Let {\otimes}' denote the tensor product in the well-mixed category; A {\otimes} '_B C is the quotient of A {\otimes} _B C by the smallest well-mixed ideal of that ring. Let S = R {\otimes} ' _{R_{n'},y_{n'}} L, z: S \to L the induced map, S_n = \sigma^n(S), z_n = z | S_n.

We have to show that S_n has total dimension 0. Now S is a well-mixed L-algebra, and \sigma^{n'}(S) \subset L. By assumption, S has total dimension \leq n. We are reduced to showing:

It suffices to show that if n<n' and X \to B_{n'}(X) has total dimension \leq n, then so does X \to B_{n'+1}(X). Let k be a difference field, y an k-valued point of B_{n'+1}. In terms of local rings, we have R \supset \sigma^{n'}(R) \supset \sigma^{n'+1}(R), and z: \sigma^{n'+1}(R) \to k. We have to show that S = R {\otimes} _{\sigma^{n'+1}(R),z} k has total dimension \leq n. Let (h,D,L) be as in the definition of total dimension. Let S_m denote the image of \sigma^m(R) {\otimes} _{\sigma^{n'+1}(R),z} k in S.

Observe in general that if X \to_f Y has relative total dimension \leq n, then so does the induced map BX \to BY. Indeed, by Remark ?, the pullback f^{-1}BY \to BY has total dimension \leq n, and BX is a subscheme of f^{-1}BY, so ? applies again. In particular, as X \to B_{n+1}X has total dimension \leq n, so does B_{n'-n}X \to B_{n'+1}X .

By lemma ?, \sigma^n(\sigma^{n'-n}R {\otimes} _{\sigma^{n'+1}(R),z} k) has total dimension 0. Thus h(S_{n'}) is finite dimensional over k, so it is a difference subfield of the domain D, call it k'. So h(\sigma^{n'}(R)) \subset k'. As X \to B_{n'}(X) has total dimension \leq n, h(R) is contained in a field F of k'-transcendence degree \leq n. But [k':k]< \infty, so tr. deg._{k}(F) = tr. deg_{k'} (F) \leq n. \Box

This reduces to the affine case by taking an open affine cover; there it follows from Corollary ?. \Box

Since I is a well-mixed ideal, \sqrt[\sigma]{I} is a perfect ideal (Lemma ?.) If I is \sigma-primary, then P is a prime ideal, hence a transformal prime. Thus (1) implies (2). Assume (2). Let I' = \{a/b: a \in I, b \notin P\}. This is a well-mixed ideal of R': if (a/b)(a'/b') \in I', then (a/b)(a'/b') = a''/b'' with a'' \in I, b'' \notin P. So aa'b''=a''bb' in R'; upon multipliying denominators and numerators by an appropriate element outside P, we may assume aa'b''=a''bb' in R. As I is P-primary, and a'' \in I, we have aa' \in I; as I is well-mixed, a^\sigma a' \in I. Now (a/b)^\sigma (a'/b') = (a^\sigma a') / (b^\sigma b') \in I'. So I' is well-mixed. We have I \subseteq \iota {^{-1}} (I'), and if a/1 \in I', then a = a'/b for some a' \in I, b \notin P, so (ab-a')b'=0 for some b' \notin P, thus ab-a' \in I, so ab \in I, so a \in I. This proves (3). Now assume (3). If ab \in I, and b \notin P, then \iota(b) is invertible; so a/1 \in I', thus a \in I. \Box

Suppose I_1 \subset I_2 \subset \ldots \subset P is such a chain. Let P_k = \sqrt[\sigma]{I_k}. Then P_1 \subseteq P_2 \subseteq \ldots, so by Ritt-Raudenbush the P_k are eventually equal, and we may assume \sqrt[\sigma]{I_k}=P. Let I = \cup_k I_k. If b \in R \setminus P, R' = R[b {^{-1}}]_\sigma, I_k' = I_k R[b {^{-1}}]_\sigma, then the I_k' form a strictly ascending chain, by Lemma ? (3). Thus, replacing R by an appropriate R' using Lemma ?, we may assume I = \cup_k I_k is finitely generated; but it follows that I=I_k for large enough k, a contradiction. \Box

\sigma: L \to \sigma(L) is an isomorphism of valued fields; by the universal property of the Henselization, for any Henselian valued field M containing \sigma(L), \sigma extends uniquely to an embedding L^h \to M; in particular, with M=L^h, \sigma extends to \sigma: L^h \to L^h. The property of being \omega-increasing depends on the value group, which does not change. \Box

Pick t \in K, {\, \rm val \,}(t)>0. Since tr. deg. _{F(t)_\sigma} L < \infty, {\, \rm val \,}(L^a)/ {\, \rm val \,}({F(t)_\sigma}^a) is a finite-dimensional \hbox{\msbmten Q}-space. Thus {\, \rm val \,}(L^a) is a finitely generated \hbox{\msbmten Q}[\sigma]-module. Since \hbox{\msbmten Q}[\sigma] is a principal ideal domain, and {\, \rm val \,}(L^a) is torsion free, {\, \rm val \,}(L^a) is a free \hbox{\msbmten Q}[\sigma]-module. Now the quotient {\, \rm val \,}(L^a)/ {\, \rm val \,}({F(t)_\sigma}^a) is finite-dimensional, so the rank of {\, \rm val \,}(L^a) must equal one. Let h: \Q-si \to {\, \rm val \,} (L^a) be an isomorphism of \hbox{\msbmten Z}[\sigma]-modules. Replacing h by -h if necessary, we may assume h(1)>0. It follows that h is order preserving. Any nonzero \hbox{\msbmten Z}[\sigma] - submodule of \Q-si is cofinal, and co-torsion. \Box

The residue field of F(t)_\sigma is F itself, and L has finite transcendence degree over F(t)_\sigma. \Box

: Let {\cal O}_L be the valuation ring of L. (1) implies (2) since if {\rm res \,}(K(c))={\rm res \,}(K), and {\, \rm val \,} (K(c)) = {\, \rm val \,}(K), then {\rm res \,}(K(c)) is algebraically closed and {\, \rm val \,}(K(c)) is divisible; so they cannot change under algebraic field extensions. Now assume {\, \rm val \,}(K(c)) \neq {\, \rm val \,}(K). Then {\, \rm val \,} f(c) \notin {\, \rm val \,}(K) for some f \in K[X]. Splitting f into linear factors, wee see that we can take it to be linear. So {\, \rm val \,}(c-b) \notin {\, \rm val \,}(K) for some b \in K. It follows that {\, \rm val \,}(c-b') \leq {\, \rm val \,}(c-b) for all b' \in K. (Otherwise {\, \rm val \,}(c-b)={\, \rm val \,}(b-b').) Next suppose {\, \rm val \,}(K(c)) = {\, \rm val \,}(K), but {\rm res \,} f(c) \notin {\rm res \,} (K) for some f \in K[X] with f(c) \in {\cal O}_L. Taking into account {\, \rm val \,} (K(c)) = {\, \rm val \,}(K), we can split f into linear factors f_i such that each f_i(c) \in {\cal O}_L. Thus ac-b \in {\cal O}_L, {\rm res \,}(ac-b) \notin {\rm res \,}(K) for some a,b \in K. It follows that {\, \rm val \,}(ac-b') \leq 0 for all b' \in K. So {\, \rm val \,}(c-b/a) \geq {\, \rm val \,}(c-b'') for all b'' \in K. This proves that (2) implies (3).

Now assume (3): {\, \rm val \,}(c-b) \geq {\, \rm val \,}(c-b') for all b' \in K. If {\, \rm val \,}(c-b) \notin {\, \rm val \,}(K), (2) holds. If {\, \rm val \,}(c-b) = {\, \rm val \,}(d), d \in K, then (c-b)/d \in {\cal O}_L, and {\, \rm val \,}((c-b)/d-d') \leq 0 for all d'; so {\rm res \,}((c-b)/d ) \notin {\rm res \,}(K). Thus (1). \Box

If one of T(c/K),T(d/K) has a last element, then by ? so does the other, and E(c/K)=(0)=E(d/K). Thus we may assume T(c/K),T(d/K) have no last element, so that {\, \rm val \,} K = {\, \rm val \,} L =: \Gamma. Further, since E(c,K),E(d/K) are convex subgroups, one is contained in the other, so we may assume that E(c/K) \subseteq E :=E(d/K).

In case E = \Gamma, , by the last statement of ?, L embeds into \widehat{K}, and hence E(c/K)=\Gamma too.

In general, let \Gamma' = \Gamma / E, r: \Gamma \to \Gamma' the quotient map, and let {\, \rm val \,}': L \to \Gamma' be the valuation obtained by composition, {\, \rm val \,}' = r \circ {\, \rm val \,}. Note that {\, \rm val \,}'(L)={\, \rm val \,}'(K) = \Gamma'. Call the residue fields K',L'. Let T'(y/K) = \{val'(y-b): b \in K \}.

If E(c/K) \neq E(d/K), then T'(c/K) has a last element. (Indeed let \gamma \in E(d/K) with \gamma > 0 and \gamma \notin E(c/K). Then there exists \alpha \in T(c/K) with \alpha + \gamma \notin T(c/K). Let {\alpha}' be the common image of \alpha,\alpha+\gamma in \Gamma'. Then clearly \alpha' is the greatest element of T'(c/K).) By ?, T'(d/K) has a greatest element too. Effecting additive and multiplicative translations, we may assume these greatest elements are {\, \rm val \,}'(c)=0,{\, \rm val \,}'(d)=0. Let c',d' be the residues of c,d.

Since {\, \rm val \,}'(L)={\, \rm val \,}'(K) = \Gamma', by ? we must have K' \neq L'. We have an induced valuation {\, \rm val \,}'': L' \to E, {\, \rm val \,}'' ({\rm res \,}' (a) ) = {\, \rm val \,}(a) for a with {\, \rm val \,}(a) \in E. Clearly T''(c'/K') = \{{\, \rm val \,}''(c'-b': b' \in K') \} = E \cap T(c/K), and similarly for d; thus E''(c'/K') = E(c/K), E''(d'/K') = E(d/K). But now T''(d/K) is the entire value group E(d/K), so by the special case of E=\Gamma considered above we have E''(c'/K') = E''(d'/K'). \Box

Assume K',K'' are embedded in L = K'K''. Let a \in K', b \in K''. Then E(a/K) = E(c/K) \neq E(d/K) = E(b/K), so T(a/K) \neq T(b/K). One of T(a/K),T(b/K) is bigger, say T(a/K) \subset T(b/K). Find e \in K with {\, \rm val \,}(b-e) > T(a/K). Then {\, \rm val \,}(a-b) = {\, \rm val \,}(a-e). This shows that the values of elements a-b are all determined. In particular the values {\, \rm val \,}(c-b) for b \in K(d) are determined; this determines tp(c/K'') and hence tp(K'/K'').

Moreover we see that every element {\, \rm val \,}(c-a) of T(c/K'') either equals {\, \rm val \,}(c-e) for some e \in K, so that it lies in T(c/K), or else equals {\, \rm val \,}(a-e) where {\, \rm val \,}(c-e) > {\, \rm val \,}(a-e); so in either case it lies in {\cal H} T(c/K). \Box

By assumption, a_i \notin K^{alg}. Use induction on n. Some E_i is nonzero; say E_1 \neq (0). Let K' = K(a_1)^a. By ?, {\, \rm val \,} K' = {\, \rm val \,} K. By ?, a_i,a_1 are algebraically independent over K for i \neq 1. By ?, the type of (a_i,a_1)/K is determined, and E(a_i/K') = E(a_i/K). By induction, a_2,\ldots,a_n are independent over K', and their type over K' is determined. The conclusion follows. \Box

We may assume here that tr. deg. _{F }K_{1} < \infty. If F \leq F' \leq K_1 then the hypotheses hold for F,F',K_2 and then also for F',K_1,F'K_2; so using transitivity of almost orthogonality the lemma reduces to the case tr. deg. _{F }K_{1} =1. Fix some embedding of K_1,K_2 into a valued field L. Let a \in K_{1} \setminus F. Then a \notin K_{2}, by the linear disjointness assumption. Let B be a K_{2}-definable ball with a \in B. So B = B_{\rho}(b)= \{y: {\, \rm val \,}(y-b) \geq \rho \}, \rho \in {\, \rm val \,}(K_{2}), b \in K_{2}. As a \notin K_{2}, \rho < \infty. Since b \in \widehat{F} and \rho \in {\, \rm val \,}(\widehat{F}) = {\, \rm val \,}(F), there exists b' \in F with {\, \rm val \,}(b-b')> \rho. So B is F-definable. Thus tp(a/F) implies a \in B. Since B was arbitrary, and K_{2} is algebraically closed, the formulas x \in B of this kind generate tp(a/K_{2}). Thus tp(a/F) implies tp(a/K_{2}). So F(a),K_{2} are almost orthogonal over F. By Theorem 2.11 of [HHM2], K_{1} ,K_{2} are almost-orthogonal over F. \Box

Ł-L^- M-M^- (1) Let \L- =\widehat{K_{n-1}} , \, {\cal M}- = \L-{\bar{K}}. We may assume inductively that \bar{K},\L- are almost orthogonal over K, and {\cal M}- is an immediate extension of \L-, hence of {K_{n-1}}. We will show that {\cal M}-,L are are almost orthogonal over \L-, and M is an immediate extension of L. From this (1) follows by transitivity.

If K_n \subseteq \widehat{K_{n-1}}, then \L-=L and there is nothing to prove. Otherewise K_n/K_{n-1} is purely ramified, or purely inertial. Since {\cal M}- is an immediate extension of {K_{n-1}}, it follows that \L- K_n / \L- is purely ramified, or purely inertial; for instance in the purely inertial case, using (K_{n-1})_{{\rm res \,}} = (\L-)_{{\rm res \,}} we have

so the numbers are equal. This also shows that \L-K_n is almost orthogonal to {\cal M}- over \L-. By Lemma ?, L is almost orthogonal to {\cal M}- over \L-, and {\cal M}- L is an immediate extension of L. But M={\cal M}-L. \Box

Find an affine open subscheme containing the said n points q_1,\ldots,q_n, and thus reduce to X = {\rm Spec^\sigma \, } A, A a finitely generated R-algebra. Let B be the image of A in A {\otimes}_R K. As X is 0-dimensional, A {\otimes}_R K is a finite dimensional K-space, and so the finitely generated R-submodule B is a free R-module of finite rank. It follows that {\rm dim \, } _k (A {\otimes}_R k) \geq {\rm dim \, } _k ( B {\otimes} _R k) = {\rm dim \, } _K (B {\otimes} _R K) = {\rm dim \, } _K(A {\otimes}_R K).

If X is flat over R, then A is a flat R module, so A \to A {\otimes}_R K is an injective map of modules, and hence B \cong A.

\Box

Let a \in L; we will show that a \in \widehat{K}. We may assume {\, \rm val \,}(a)>0. Let a_n = \sigma^n(a), C_n = \{ \max(0,{\, \rm val \,}(a_n-b)): b \in K \}, C=C_0. Then C_n is closed downwards in the non-negative elements {\, \rm val \,}(K). C_n cannot have a maximal element, since L is an immediate extension of K.

If C is unbounded, then a \in \widehat{K}. Suppose for contradiction that C is bounded. We have \Q-si = \hbox{\msbmten Q} v \oplus \hbox{\msbmten Q} v^\sigma + \ldots; let E_m = \hbox{\msbmten Q} v \oplus \ldots \oplus \hbox{\msbmten Q} v^{\sigma^m}. Let m be least such that C \leq E_m + c_0 for some c_0 \in C (here we write X \leq Y if (\forall x \in X)(\exists y \in Y) (x \leq y).)

Since {\, \rm val \,}(L)={\, \rm val \,}(K), dividing a by an element of K with value c_0, we may assume C \leq E_m; as E_m is convex, we have C \subset E_m. Then \sigma(C) \subset E_{m+1}. Let H_1 be the convex hull of \sigma(C).

Claim E_m + H_1 = H_1.

Let e_0 = f(a). We will find a_1 \in {\cal O}_K, {\, \rm val \,}(a-a_1) \geq v-v', {\, \rm val \,} f(a_1) \geq \sigma(v). Since {\, \rm val \,} ( f_1(a) - f_1(a_1)) \geq v-v' > v' = {\, \rm val \,} f_1(a), it follows that {\, \rm val \,} f_1(a_1) = v'. Iterating this we get a sequence a_n, with v_n =_{def} {\, \rm val \,} f(a_n) \geq \sigma(v_{n-1}), and {\, \rm val \,} (a_{n-1} - a_n) \geq v_n - v'. As K is complete, there exists a unique b \in K such that the sequence a_n converges to b; by continuity, f(b)=0.

To obtain a_1, write a_1= a + re, with e=e_0 / f_1(a), and r \in {\cal O}_K to be found. Then

Here the f_i are transformal derivatives, cf. ? \nu runs over indices \geq \sigma, while i ranges over the nonzero finite indices. Note that {\, \rm val \,} f_\nu(a) (re)^\nu \geq {\, \rm val \,}(\sigma(e)) = \sigma(v). Thus it suffices to find r with e_0+\sum_{i=1}^m f_i(a) (re)^i =0, or with r a root of

(Divide through by e_0, noting f_1(a)(e/e_0) =1.) This (ordinary) polynomial has derivative

Now for k \geq 2, {\, \rm val \,} (e^k / e_0) = k (v-v') -v \geq 2(v-v') - v = v-2v' > 0. Thus {\rm res \,} g' is a nonzero constant polynomial. By the ordinary Hensel lemma, there exists r \in {\cal O}_K with g(r)=0, near any c \in {\cal O}_K with {\, \rm val \,} g(c) > 0. Letting c = -1 we see that there exists r \in {\cal O}_K with g(r)=0. \Box

Pick any a \in {\cal O}_L with {\rm res \,}(a)={\bar{a}}, and also lift f to {\cal O}_K[X]_\sigma. Then {\, \rm val \,} f(a) > 0, {\, \rm val \,} f'(a) = 0, so by the transformal Hensel property (cf. ?) there exists b \in {\cal O}_K with f(b)=0, {\rm res \,}(b)= {\bar{a}}. Clearly K'=K(b)_\sigma is purely inertial over K. \Box

Pick any a \in {\cal O}_K with {\rm res \,}(a)={\bar{a}}, and also lift f to {\cal O}_K[X]_\sigma. Then {\, \rm val \,} f(a) > 0, {\, \rm val \,} f'(a) = 0, so by ?, there exists b \in {\cal O}_K with f(b)=0, {\rm res \,}(b)= {\bar{a}}. The difference field F(b)_\sigma has finite transcendence degree m over F, and hence contained in F'; thus it is trivially valued, so the residue map is injective on F(b)_\sigma. since {\rm res \,}(b) = {\bar{a}}, it carries F(b)_\sigma into and onto F({\bar{a}})_\sigma. \Box

By Lemma ?, (F')^{inv} \subset (F')^a; and clearly F' is perfect. We may assume F'=F, and show that {\rm res \,}(K)=F. Suppose K_{{\rm res \,}} \neq F. In characteristic 0, choose a difference polynomial f \in F[X]_\sigma of smallest possible transformal order and degree, and {\bar{a}} \in K_{{\rm res \,}} with f({\bar{a}})=0. If f \in F[X^\sigma]_\sigma, then replacing {\bar{a}} by \sigma({\bar{a}}) we may lower the degree of f; unless \sigma({\bar{a}}) \in F. But in this case, {\bar{a}} \in F^a, so the order of f must be 0, so f \in F[X^\sigma] is impossible. Thus f \notin F[X^\sigma]. It follows that f' \neq 0 and f' has smaller order or degree. So the hypothesis of ? is met, and we can increase F'.

In positive characteristic, this may not be the case, because of polynomials such as x^\sigma - x^p. However, after replacing \sigma by \tau = \sigma(x^{1/p^m}) for large enough m, it is possible to find such an {\bar{a}} for \tau, lowering the degree of f further. Assuming the residue field of finite total dimension, this permits lifting the residue field; we may then return from \tau to \sigma. \Box

Assume(1), and let L' be as in (2). By ?, L has value group \hbox{\msbmten Q}_\sigma, and L_{{\rm res \,}} has transformal transcendence degree 0 over F. So L_{{\rm res \,}} is inversive.

Let t \in L' be such that {\, \rm val \,}(t) generates {\, \rm val \,}(L) as a \hbox{\msbmten Q}[\sigma]-module. By ?, L_{\rm res \,} lifts to a trivially valued difference subfield F' of L'. Let K = (F'(t)_\sigma)^a. Then K_{{\rm res \,}} = L_{{\rm res \,}}, and {\, \rm val \,}(K) = {\, \rm val \,}(L).

We have {\, \rm val \,}(K) \neq {\, \rm val \,}(K^\sigma), while tr. deg._{\sigma(K)} K = 1. By ?, the hypothesis (*) of ? holds for K. Thus by ?, L = \widehat{K}. Since \widehat{K} \leq L' \leq L, we have also L'=L. This proves (2).

The same proof shows that K embeds into L; so assuming (2), we conclude (3), using the assumption of (2) in place of ?.

(3) implies (1) trivially. \Box

Let K' be a maximal purely ramified difference field extension of K within \widehat{L}. Then K' = (K')^{a}.

By ?, \widehat{L} = \widehat { F(t)_\sigma^a} for some t. Let e = rk_{\hbox{\msbmten Q}} {\, \rm val \,}(L) / {\, \rm val \,}(K'). Then there exists s \in K', {\, \rm val \,}(s) >0, {\, \rm val \,}(s) \in H_{e+1}. Fix N \geq e+1 for a moment. Let t_n = \sigma^n(t), s_m = \sigma^m(s), \bar{t_n} =\, {\rm RES} \,_N(t_n), \bar{s_n} = \, {\rm RES} \,_N (s).

We saw that tr. deg._F \, {\rm RES} \,_{N} (L) = N. Thus \bar{t_0},\ldots,\bar{t_e},\bar{s_0},\ldots,\bar{s}_{N-e-1} are algebraically dependent over F. Since s \in H_{e+1}, \bar{s_0},\ldots,\bar{s}_{N-e-1} are algebraically independent over F, using the valuation {\, \rm val \,}^*_e . Let m be maximal such that \bar{t_m},\ldots,\bar{t_e},\bar{s_0},\ldots,\bar{s}_{N-e-1} are algebraically dependent over F.

Write f( \bar{t_m},\ldots,\bar{t_e},\bar{s_0},\ldots,\bar{s}_{N-e-1}) =0, f \in F[ X_m,\ldots,X_e,\bar{s_0},\ldots,\bar{s}_{N-e-1}] of minimal X_m- degree.

We now argue that m=e. Lift f to f( X_m,\ldots,X_e,s_0,\ldots,s_{N-e-1}) \in {\cal O}_L [X_m]_\sigma, viewed as a difference polynomial in X_m ( with coefficients involving the s_i, and X_{m+j} = (X_m)^{\sigma^j}.) In characteristic 0, let f' be the (first) derivative of f with respect to X_m; then f' \neq 0, and so by minimality of f, RES_N f'( \bar{t_m},\ldots,\bar{t_e}) \neq 0, i.e. {\, \rm val \,} f'( t_m,\ldots,t_e) \in H_N. By ?, applied to the element t_m, there exists b \in \widehat{L}, {\, \rm val \,}(t_m -b) > H_{N}, f( b,\ldots,\sigma^{e-m}(b),s_0,\ldots,s_{N-e-1}) = 0. Let L' = K'(b)_\sigma. Clearly tr. deg._K L' \leq e-m, while rk_{\hbox{\msbmten Q}} {\, \rm val \,}(L') / {\, \rm val \,}(K') \geq e-m (since {\, \rm val \,}(\sigma^i(b))={\, \rm val \,}(t_{m+i}), and {\, \rm val \,}(t_m) << {\, \rm val \,}(t_{m+1}) << \ldots << {\, \rm val \,}(t_{e-1}) << {\, \rm val \,}(c) for any c \in K', {\, \rm val \,}(c) > 0.) However rk_{\hbox{\msbmten Q}} {\, \rm val \,}(L') / {\, \rm val \,}(K') \leq tr. deg._{K'} L', so these numbers are equal, and L'/K' is purely ramified, of transcendence degree m-e. But K' is a maximal purely ramified extension; so K'=L', and m=e.

In characteristic p, we first replace \sigma by \sigma \circ \phi^{-l} where \phi(x)=x^{p}, reduce the equation if it is now a p’th power, till t_{m} occurs with an exponent that is not a power of p. We then use the above argument. So m=e in any characteristic.

Thus \bar{t_{e}} \in \, {\rm RES} \,_{N}(K') (since this residue field is algebraically closed). So for some b_N \in K', {\, \rm val \,}(t_{e }-b_N ) > H_{N }(L). Now letting N \to \infty, we see that t_e \in \widehat{K'}. Let K'' = F(t_e)^a_\sigma; then \sigma^e \widehat{L} = \widehat{K''} \subset \widehat{K'}, and it remains to show equality. Note that {\, \rm val \,}(K'') = \hbox{\msbmten Q}[\sigma] {\, \rm val \,}(t_e) = {\, \rm val \,} (K'), and {\rm res \,}(K'') = F = {\rm res \,}(K'). Thus K'/K'' is immediate, so for any c \in K', T(c) = \{{\, \rm val \,}(c-b): b \in K'' \} can have no maximum value (?). But by ?, for any c \in \widehat{L} \setminus \widehat{K''}, T(c) does have a maximal element. Thus K'=K''. \Box

By ?, we may assume L,K are transformally valued fields over F, F=K_{{\rm res \,}}=L_{{\rm res \,}}. Let K' be as in ?. Then K'/K is purely ramified, of finite transcendence degree r; there exist therefore K_0 \leq K,K'_0 \leq K', tr. deg._F (K_0) finite, tr. deg. _{K_0} (K'_0) = r = rk_{\hbox{\msbmten Q}} {\, \rm val \,}(K'_0)/{\, \rm val \,}(K_0), K'=K K'_0. We can choose {K_0} algebraically closed, and with value group equal to {\, \rm val \,}(K). So \widehat{K_0} \leq \widehat{K'_0} \leq \widehat{K'}. \widehat{K'} is isomorphic (by ? (1 \to 3)) to L, so it satisfies ? (1)-(3); by (2), since \widehat{K'_0} is a complete difference subfield with the same value group and residue field, \widehat{K'} = \widehat{K'_0}. Thus \widehat{K_0} \leq K \leq \widehat{K'} = \widehat{K'_0}. If c \in \widehat{K'} \setminus \widehat{K_0}, as K'_0/K_0 is purely ramified, and as in the proof of ?, K_0(c)/K_0 is also purely ramified. In particular if c \in K \setminus \widehat{K_0} then {\, \rm val \,} (c) \notin K_0; but we chose {\, \rm val \,}(K_0) = {\, \rm val \,}(K). Thus K=\widehat{K_0}. Now \widehat{K_0} satisfies ? (1). \Box

Since F is inversive, and L_{{\rm res \,}} is algebraically closed, by ? and ?, L_{{\rm res \,}} is also inversive. Thus for any d \geq 1, {\rm res \,} ( L^{\sigma^{d}}) = {\rm res \,} (L). By ?, {rk_{{\, \rm val \,}}}(\widehat{L} / \widehat{L}^{\sigma^{d}}) = rk_{\hbox{\msbmten Q}} (\Q-si / \sigma^d(\Q-si)) = d.

To show that (1) implies (2), we may assume K=K^a. By ?, there exists a field of representatives F_r \leq \widehat{K} for the residue map of \widehat{K}. We have tr. deg_F F_r = tr. deg._F K_{{\rm res \,}}. Let K_r = F_rK.

By ? again, applied to \widehat{L} over K_r, there exists a difference field F_r \leq L_1 such that {\rm res \,}: L_1 \to L_{{\rm res \,}} is an isomorphism. Let K_1 = K_r L_1. Then tr. deg._{K_r} K_1 \leq e_{{\rm res \,}}; equality holds by comparing the residue fields. Thus (b).

Now over K_1, ? applies, and gives K_2 with (c,d). We have K \leq \widehat{L}^ {\sigma^{e_2}}, and e_1 \leq rk_{\hbox{\msbmten Q}} {\, \rm val \,}( \widehat{L}^ {\sigma^{e_2}} / K); so e_1 + e_2 \leq rk_{\hbox{\msbmten Q}} ({\, \rm val \,}(\widehat{L} )/ {\, \rm val \,} K) = rk_{\hbox{\msbmten Q}} ({\, \rm val \,} L / {\, \rm val \,} K). Thus (e).

(2) obviously implies (3), with K'=K_2 \widehat{K}. To go from (3) to (4), let K' be as in (3). By ?, \widehat{L} \simeq \widehat{(L_{{\rm res \,}}(t)_\sigma)^a}, \widehat{K'} = \widehat{L}^{\sigma^{e_2}} = \widehat{(L_{{\rm res \,}}(t^{e_2})_\sigma)^a}. Let M = \widehat{(L_{{\rm res \,}}(t^{e_2})_\sigma)^a} (t). Then \widehat{M} = \widehat{L} and tr. deg._{\widehat{K'}}(M) = e_2.

(4) implies (1) since {rk_{{\, \rm val \,}}} is additive in towers, bounded by transcendence degree, and vanishes for completions ({rk_{{\, \rm val \,}}}_K \widehat{K} = 0.)

The final conclusion follows from (4) by taking e ={rk_{{\, \rm val \,}}}(L/K). We obtain M_0,\ldots,M_5 with M_{i+1} = \widehat{M_i} for even i<5, and tr.deg._{M_1}M_2 + tr. deg._{M_3} M_4 \leq e. But e = {rk_{{\, \rm val \,}}}(M_2/M_1) + {rk_{{\, \rm val \,}}}(M_4/M_3), and {rk_{{\, \rm val \,}}}(M_{i+1}/M_i) \leq tr.deg._{M_{i+1}}M_i; so all inequalities must be equalities. \Box

Note that taking algebraic closure or completion does not change {rk_{{\, \rm val \,}}}. We will thus assume all these fields are algebraically closed; and we will prove a more general statement, allowing L to be the completion of a transformal valued field of transformal dimension 1.

All residue fields are therefore also algebraically closed and, being extensions of K_{{\rm res \,}} of transformal transcendence degree 0, are inversive. Note (say by ?) that L' = \widehat{K'L} is also the completion of a transformal valued field of transformal dimension 1.

If K \leq M \leq \widehat{L}, then {rk_{{\, \rm val \,}}}(L/K) = {rk_{{\, \rm val \,}}}(L/M) + {rk_{{\, \rm val \,}}}(M/K), and {rk_{{\, \rm val \,}}}(L/K') = {rk_{{\, \rm val \,}}}(L/MK') + {rk_{{\, \rm val \,}}}(M/K'); so the lemma for M/K ; K' and for L/M; MK', implies the statement for L/K; K'. By ?, we may thus assume one of the following cases holds:

1. L \leq \widehat{K}

2. tr.deg._K (L) = {rk_{{\, \rm val \,}}} (L/K)

In case (1), LK' \leq \widehat{K'}, so {rk_{{\, \rm val \,}}}(L/K') = {rk_{{\, \rm val \,}}}(\widehat{K'L} / \widehat{K'}) = 0.

In case (2), {rk_{{\, \rm val \,}}}(L/K') \leq tr.deg._{K'}(K'L) \leq tr.deg. _K (L) = {rk_{{\, \rm val \,}}} (L/K).

\Box

The value group {\Gamma}(K)={\Gamma}(\bar{K}) is cofinal in {\Gamma}(L), since if a \in L has valuation > {\Gamma}(K) then it must be transformally transcendental over K.

It follows in particular that the completion \widehat{L} embeds uniquely into \widehat{M}. From this, once we show that L,\bar{K} are almost orthogonal over K, almost orthogonality of \widehat{L},\bar{K} over K follows.

By ?, there exists a tower K= K_0 \leq \cdots \leq K_6 = \widehat{L} of difference field extensions of K, with K_{i+1} purely ramified or purely inertial over K_i, or contained in \widehat{K_i}. (Namely K \leq K_r \leq K_1 \leq K_2 \leq \widehat{K_2} \leq M \leq \widehat{L}; with M as in ? (4).) The proposition follows from ?.

Discussion

This can be formulated in terms of domination for types of ACVF_K, or quantifier-free types of the language of valued fields over K; as in the Remark following Lemma ?.

This implies the existence of canonical base change of L/K to extensions K' of K as transformally valued fields, relative to a base change in {\rm res \,}(L) and in {\, \rm val \,}(L) making {\, \rm val \,}(L),{\, \rm val \,}(K') independent over {\, \rm val \,}(K); in particular this applies when {\, \rm val \,}(K') = {\, \rm val \,}(K), more generally when ({\, \rm val \,}(K') \cap \hbox{\msbmten Q}(\sigma) {\, \rm val \,}(K)) and ({\, \rm val \,}(L) \cap \hbox{\msbmten Q}(\sigma) {\, \rm val \,}(K)) intersect in {\, \rm val \,}(K) within \hbox{\msbmten Q}(\sigma){\, \rm val \,}(K). But note that in the example above, restricted to finitely generated approximations, this condition will not hold. In any case, this material will not be used in the present paper.

Let L_0 be a subfield of L, finitely generated over F as a field, generating L as a difference field.

Let L_n be the subfield of L generated by \cup_{k \leq n } \sigma^k(L_0). Then [L_{n+1}: L_n] \leq [L_1:L_0] < \infty. Let A_n = {\, \rm val \,} L_n. Then A_0 \subset A_1 \subset \ldots \subset {\, \rm val \,} L^a \simeq \hbox{\msbmten Q}[\sigma], A_0 is finitely generated (using Lemma ?), and A_{n+1} / A_n is bounded. By Lemma ? and lemma ? below, {\, \rm val \,} L is a finitely generated \hbox{\msbmten Z}[\sigma]-module. And by ?, every finitely generated \hbox{\msbmten Z}[\sigma] submodule of \hbox{\msbmten Q}[\sigma] is contained in a free \hbox{\msbmten Z}[\sigma]-module of rank one. \Box

Let A_i= (B_i + \hbox{\msbmten Z}[T]) / \hbox{\msbmten Z}[T], A= \cup_i A_i. We will show A_i is finitely generated as a \hbox{\msbmten Z}[T]-module. It follows that B+ \hbox{\msbmten Z}[T] is a finitely generated \hbox{\msbmten Z}[T]-module; and by Noetherianity of \hbox{\msbmten Z}[T], so is B. Note that A_{n+1} / A_n is finite and bounded, and TA_n \subseteq A_{n+1}.

We use induction on m = \limsup _n \left| A_{n+1} / A_n \right|. If A \neq 0, there exists c \in A, c \neq 0, lc = 0, l prime. Say c \in A_{n_0}.

Claim Let E=E_l = \{x: l x=0 \}. Then for n \geq n_0, E \cap A_{n+1} \not \subseteq A_n.

Suppose the claim is false; then any x \in E \cap A_n satisfies Tx \in E \cap A_{n+1} = E \cap A_n. Thus \hbox{\msbmten Z}[T] c \subset A_n. But A_n is a finitely generated group, while \hbox{\msbmten Z}[T] c is not: c,Tc,\ldots are \hbox{\msbmten Z}/l \hbox{\msbmten Z} - linearly independent. A contradiction.

Thus the natural surjective map e_l: A_{n+1}/A_n \to lA_{n+1} / lA_n, e_l(x)=lx, is not injective, for n \geq n_0. So \limsup _n \left| lA_{n+1} / lA_n \right| < m. By induction, lA is a finitely generated \hbox{\msbmten Z}[T]-module.

Now E_l / ( \hbox{\msbmten Z}[T] c) is finite; indeed E_l \simeq ( \hbox{\msbmten Z}/l \hbox{\msbmten Z})[T], so E_l / ( \hbox{\msbmten Z}[T] c) \simeq ( \hbox{\msbmten Z}/l \hbox{\msbmten Z})[T] / f for some nonzero polynomial f(T). Thus (A \cap E_l) / (\hbox{\msbmten Z}[T]c) is finite. Since \hbox{\msbmten Z}[T]c is a finitely generated \hbox{\msbmten Z}[T]-module, so is A \cap E_l = \ker (x \mapsto lx). We saw lA was finitely generated; hence so is A. \Box

If M is generated by a_1,\ldots,a_n, m_ia_i \in \hbox{\msbmten Z}_\sigma for some 0< m_i \in \hbox{\msbmten N}; so each a_i \in \hbox{\msbmten Z}_\sigma[1/m], where m = \Pi_i m_i. Thus M \subset N for some principal N \leq \Q-si.

Claim For any ideal J of \hbox{\msbmten Z}[\sigma], there exists a principal ideal J' containing J with J'/J finite.

By ?, {\, \rm val \,} L \subset {\widetilde{M}} for some free 1-generated \hbox{\msbmten Z}[\sigma]-module {\widetilde{M}} with {\widetilde{M}} / {\, \rm val \,} L finite; we have {\, \rm val \,} L = I {\widetilde{M}} for some ideal I, and necessarily n \in I for some n >0.

Since L is Henselian and F is closed under roots, the homorphism val: L^* {\to >} {\, \rm val \,} L is a surjection. It induces a surjection

Let H be the pullback of n {\widetilde{M}} \subset {\, \rm val \,} L. We have \sigma(H) \subset H; and

In particular, H/(L^*)^n is finite. By Kummer theory, there exists a unique Galois extension K of L, such that (K^*)^n \cap L = H, and [K:L] = [H: (L^*)^n] = {\widetilde{M}} / {\, \rm val \,} L. Now \sigma(K)L is also a Galois extension of L, generated by n’th roots of elements of H; hence it is contained in K; so K is a difference field. Tracing back the isomorphisms we see that {\widetilde{M}} \subset {\, \rm val \,}(K), so {\widetilde{M}} = {\, \rm val \,}(K). \Box

For each i, \sigma(a_i) is a root of a monic polynomial f_i = \sum_{l=0}^{d_i} c_{i,l}X^l \in D[a][X]. The coefficients c_{i,l} are themselves polynomials in a over D; take m bigger than the total degree of all these polynomials. Then if v is a valuation, with v(d) \geq 0 for d \in D and v(a_i) \geq \delta for each i, (where \delta < 0), then for each i,l, v(c_{i,l}) > m \delta.

Assume (K,R,{\, \rm val \,}) is m-increasing, and h(D) \subset R. We may replace D,D' by their images under h, and assume h is the inclusion. Say {\, \rm val \,}(ha_1) \leq {\, \rm val \,}(ha_2) \leq \ldots \leq val(ha_n). We have to show that \delta={\, \rm val \,}(a_1) \geq 0. Otherwise, as the (monic) leading monomial of f_i(\sigma(a_i)) cannot have valuation less than all other monomials, we have d_1 val(\sigma(a_1)) \geq {\, \rm val \,}(c_{1,l})+l val(\sigma(a_1)) for some l< d_1. Thus (d_1-l) val(\sigma(a_1)) \geq {\, \rm val \,}(c_{1,l}) > m \delta. So {\, \rm val \,} \sigma({a_1}^{-1}) \leq m {\, \rm val \,} {a_1}^{-1}. This contradicts the assumption that {\, \rm val \,} is m-increasing, and proves the lemma. \Box

Consider first the special case X = {\rm Spec^\sigma \, } R[x]/f(x), f(x) = x + g(x) + h(x) \in R[x]_\sigma; where g has coefficients in the maximal ideal M of R, and h is a sum of monomials x^{\nu}, \nu \in \hbox{\msbmten N}[\sigma], \nu > 1. Let a,b \in X(R) with {\rm res \,}(a)={\rm res \,}(b); we will show a=b. We may assume b=0. Then {\rm res \,}(a)=0 so {\, \rm val \,}(a) > 0; but then {\, \rm val \,} (h(a)) > {\, \rm val \,} (a), and also {\, \rm val \,} (g(a)) > {\, \rm val \,} (a), so f(a) = 0 implies a=0.

In general, we may assume X = {\rm Spec^\sigma \, } A; X_0 = {\rm Spec^\sigma \, } A_0, B_1X_0 = {\rm Spec^\sigma \, } {A_0}', with A_0= A {\otimes}_R {\bar{K}}, {A_0}' = \sigma(A_0), with X_0 \to B_1X_0 a standard unramified map (?): A_0 has generators y_1,\ldots,y_n over {A_0}', admitting relations F_1,\ldots,F_n, where (F_{i}) \in {A_0}'[y_1,\ldots,y_n] has invertible Jacobian matrix \det (\partial_{y_i}F_j) \in G_m({A_0}').

Complete y_1,\ldots, y_n to a system of generators of A_0 over \bar{K}; since the y_i already generate A over {A_0}', the additional generators y_j may be chosen from {A_0}'. Each new y_j solves the inhomogeneous linear polynomial Y_j-y_j \in {A_0}'[Y_j]; adding these generators and relations to the system clearly leaves the Jacobian invertible.

Now the coefficients of the F_i lie in \sigma(A_0), so they may themselves be expressed as difference polynomials in the y_i, indeed in the \sigma(y_i). Replacing the coefficients by these polynomials, we may assume that F_i \in {\bar K}[Y_1,\ldots,Y_n] has coefficients in \bar{K}, rather than in \sigma(A_0). If we convene that \partial \sigma(Y_j) / \partial Y_k = 0, the Jacobian matrix remains invertible.

Lift the generators y_i to x_i \in A. Then any element of A can be written as a difference polynomial over \sigma(A) in the x_i, up to an element of {\cal M} A = \ker A \to A {\otimes}_R {\bar{K}}. (Here {\cal M} is the maximal ideal of R. ) So A has generators x_1,\ldots,x_n,x_{n+1},\ldots,x_N, where for i>n the image of x_i in A_0 is y_i=0. We still have A_0= {A_0}'[y_1,\ldots,y_N]/ (F_1,\ldots,F_N), where F_i = y_i for i>n. Lift the F_i to f_i \in \sigma(A)[x_1,\ldots,x_N]. Then in A we have a relation f_i(x_1,\ldots,x_N) = g_i(x_1,\ldots,x_N), where g_i has coefficients in {\cal M}.

Consider two elements e=(e_1,\ldots,e_N), e'=(e'_1,\ldots,e'_N) of X(R) specializing to the same point of X_0. Replacing x_i by x_i-e'_i, we may assume e'=0. Suppose e \neq 0; let \min_{i} {\, \rm val \,} (e_i) = \beta > 0. In these coordinates, let L_i be the linear (monomials of order 1) part of f_i. The L_i are then the rows of an invertible matrix L over R, and L_i x - g_i(x) + h_i(x) = 0, where h_i is a sum of \sigma-monomials of degree >1. Since e \in X(R), we have Le - g(e) + h(e) = 0. (Where g,h are the matrices of \sigma-polynomials whose rows are g_i,h_i.)

Now \min_i {\, \rm val \,} (\sigma(e_i)) = \sigma(\beta) > \beta, so {\, \rm val \,} h_i(e) > \beta. Also {\, \rm val \,} g_i(e) > \beta. Writing e = - G^{-1} ( g(e) + h(e)), we see that \min_i {\, \rm val \,} (e_i) > \beta, a contradiction. \Box

First take X \subset K. Let \rho_0 < \rho_1 < \ldots \rho_n be the possible values of |x-y| for x,y \in X. Say \rho_i = |c_i|. Let E_i(x,y) \equiv |x-y| \leq \rho_i. Given Y, pick b \in Y, and let f_i(x) = {\rm res \,} {c_i}^{-1} (x-b).

When X \subset K^2, we let E_i(x,y) \equiv ( |pr_0(x)-pr_0(y)| \leq \rho_i \& pr_1(x)=pr_1(y)) for i \leq n, E_{j}(x,y) \equiv |pr_1(x)-pr_1(y)| \leq \rho_{j-n} for j>n; etc. \Box

This reduces to X \subset K, using projections. When X \subset K has finite total dimension, there exists a nonzero difference polynomial F over K with X \subset \{x: F(x)=0\}. Recall the transformal derivatives F_\nu. We have F(x+y) = \sum_\nu F_\nu(x) y^\nu. So if a,a+b \in X, then b is a root of \sum_{\nu >0} F_\nu(a) Y^\nu = 0. If F_\nu(a)=0 for all \nu > 0, then F is constant, so X = \emptyset. Otherwise, either b=0 or {\, \rm val \,} F_\nu(a) b^\nu = {\, \rm val \,} F_\mu(a) b^\mu for some \nu < \mu; so (\mu - \nu) {\, \rm val \,} (b) = {\, \rm val \,} F_\mu(a) - {\, \rm val \,} F_\nu(a). By the Claim below, \{ {\, \rm val \,} F_\nu (a) : a \in X \} is finite for each \nu; so \{ {\, \rm val \,}(b): b \neq 0, a,a+b \in X \} is also finite, and contained in the \hbox{\msbmten Z}[\sigma]-division points of {\Gamma}(K).

Claim Let G \in K[x]_\sigma. Then \{{\, \rm val \,} G(a): \ F(a)=0 \} is finite.

It suffices to see that \{{\, \rm val \,} G(a): \ F(a)=0 \} remains bounded in any elementary extension L of K. In fact if F(a)=0 then \lambda {\, \rm val \,} (a) \in {\, \rm val \,}(K) for some 0 \neq \lambda \in \hbox{\msbmten N}[\sigma]. Let a \in L, F(a)=0, c = G(a). Then K(a)_\sigma has transcendence degree \leq \deg_\sigma F over K; hence so does the subfield K(c)_\sigma. So H(c)=0 for some nonzero H \in K[x]_\sigma, H= \sum_\nu d_\nu x^\nu. Arguing as above, we see that (\mu-\nu) {\, \rm val \,}(c) = {\, \rm val \,} d_\nu - {\, \rm val \,} d_\mu for some \nu < \mu. \Box

We use induction on the number of steps of co-analyzability (the case of 0 steps being clear.) Suppose a/K is co-analyzable in h+1/2 steps, i.e. L \models P(a,b), b \in K, P co-analyzable in h+1/2 steps. Then there exists Q \in \Phi(x;u,v), co-analyzable in h steps, g \in {\Phi_{\rm fn}}(x,u;v), and d=g(a,b), such that M \models Q(a;b,d). By induction, there exists Q' \in \Phi(x;y',v) and b' \in K with M \models Q(a,b',d), {\rm dim}_V (Q') \leq Rk(L/K(d)). (Any parameters from K(d) can be written as terms in elements b' of K, and d.)

By (1), there exists R'(v;u) \in \Phi(v;u), d_V(R') \leq Rk(K(d)/K), and b'' \in K, with R'(d,b'').

Let P'(x;y,y',u) = Q' (x,y',g(x,y)) \wedge R (g(x,y);y''). Then P'(a;b,b',b'') shows that the V-dimension of a/K is \leq Rk(L/K(d)) + Rk(K(d)/K) = Rk(L/K).

Now suppose a/K is co-analyzable in h+1 steps. Then L \models P(a,b), b \in K, T \models P(x;u) \Rightarrow \bigvee_j (\exists y) Q_j(y,u), and for each j, (*) P(x;u) \wedge Q_j(y,u) is coanalyzable in h+1/2 steps. Let r= Rk(L/K),

and let \Xi' = \{ \neg S(x,b,b',y): S \in \Xi , b' \in K\}.

Suppose otherwise. Then in some M \models T_{\forall}, K \leq M, we can find a' \models tp _\Phi (a/K), and d with Q_j(d,b), such that (**) for any S \in \Xi, M \models \neg S (a,b,b',d). Let K'=K(d), L' = K(a',d). by (3), Rk(L'/K') \leq r. By (*), a'/K' is coanalyzable in h+1/2 steps. Thus by induction, {\rm dim}_V (a'/K') \leq r. So there exists S \in tp_\Phi (a'/K') with {\rm dim}_V (S) \leq r. We can take S = S(x,b,b',d), S \in \Phi (x, u,u',y). So S \in \Xi. But this contradicts (**). \Box

Note that the proof of ? uses Rk(K'/K'') only for finitely generated extensions K',K'' of K of transformal dimension 0 over K. For such extensions, take Rk(K'/K'') = {rk_{{\, \rm val \,}}}(K'/K''). Then hypothesis (1) of ? is clear since if {rk_{{\, \rm val \,}}}(K'/K'') \leq n then tr. deg. K'_{{\rm res \,}} / K''_{{\rm res \,}} \leq n. (2) follows from additivity of transcendence degree and vector space dimension. And the truth of (3) is the content of ?.

Now if K=(K_{field},K_{res}) is a two-sorted substructure of some model M of T, and c \in {\rm res \,}(K) has finite transformal dimension over {\rm res \,}(K_{field}), there exists c' \in M of finite transformal dimension over K with {\rm res \,}(c')=c; this follows from Lemma ?.

Hence by compactness we can find P(x,y) such that (i) P implies \phi(x,{\rm res \,}(y)); (ii) If \phi(a,b') then in some extension we can find b with P(a,b) and {\rm res \,}(b)=b'. By definition of {\rm dim}_V, we have {\rm dim}_V (\phi) \leq {\rm dim}_V (P).

For any M \models T, if a,b \in M, M \models P(a,b), letting K be the substructure of M generated by b, we see that there exists Q \in \Phi(x;y), M \models Q(a,b), {\rm dim}_V (Q) \leq r. By compactness, T \models P \Rightarrow \bigvee_j Q_j, with {\rm dim}_V (Q_j) \leq r. It follows that {\rm dim}_V (P) \leq r. Hence {\rm dim}_V (\phi) \leq r also.

\Box

If \sigma(u) \leq mu, then \sigma^2(u) \leq m \sigma(u), so neither u nor \sigma(u) are in {\Gamma}_I. Thus u << \sigma(u). Let 0 < u \leq v \leq \sigma(u). If mu \leq v for all m, applying \sigma we obtain mv < m \sigma(u) \leq \sigma(v) for all m, so v << \sigma(v). If v< mu for some m, then v \leq mu << \sigma(u) \leq \sigma(v).

In an ordered Abelian group, if 0< u_1 << u_2 << u_3 << \ldots then u_1,u_2,u_3,\ldots are linearly independent. Thus there is no infinite <<-chain of elements of \Gamma_n. So we may take u to be <<-maximal in \Gamma_n \cap {\Gamma}_I. But then u << \sigma(u), \sigma(u) \in {\Gamma}_I so we cannot have \sigma(u) \in \Gamma_n. At the same time \sigma(u) \in \Gamma_{n+1}. So \Gamma_n \neq \Gamma_{n+1}.

As val(t) \in \Gamma_1 \cap {\Gamma}_I, the previous claim applies for n \geq 1. It applies to n=0 too if \Gamma_0 has an element v \geq val(t). (Since v \leq val(t^{\sigma^m}) for some m, so v \in {\Gamma}_I.) Otherwise, \Gamma_0 < {\, \rm val \,}(t) \in \Gamma_1. This covers all the cases.

This follows from valuation theory, the previous claim, and the assumption tr. deg. _{K(n)} K(n+1) \leq 1.

This finishes the proof of the lemma. \Box

Note \Delta \subset \Gamma_\infty = \cup_n \Gamma_n. By Claim 3 of ?, \Gamma_{n+1} has an element greater than any element of \Gamma_n. But tr.deg. _{K(n)} K(n+1) \leq 1 implies rk_{\hbox{\msbmten Q}}(\Gamma_{n+1}/\Gamma_n) \leq 1, and it it follows that \Delta \cap \Gamma_n = \Delta \cap \Gamma_{n+1}. By induction, \Delta \cap \Gamma_n \subset \Gamma_0, and the lemma follows. \Box

Let \Delta be as in ?, and let P = \{a \in R: val(a) \notin \Delta \}. Then P /tR is the ideal of nilpotents of R/tR, and P is prime.

Let R' = \{a/b: a,b \in R, val(b) \in \Delta \}. R' is a valuation ring, with maximal ideal PR', and value group \Gamma/\Delta. R' is not in general invariant under \sigma, but R' \cap K_n has value group \Gamma_n / \Delta, and by ?, ? the rank of this value group grows with n. Thus as in ?, if K' is the field of fractions K' of R/P, then K' \subset K'(0)^a, where K'(0) = res R'(0). So tr. deg._F K' = tr. deg. _F K'(0) \leq tr. deg. K(0); the second inequality is strict unless the R'-valuation is trivial on K(0), and in this case R(0) \cap P = (0).

The total dimension of R/tR equals that of R/PR since P/tR = \sqrt (0) in R/tR, and so is bounded by the transcendence degree of the field of fractions of this domain. The total dimension of R/tR equals the reduced total dimension, r.dim(R/tR) = t.dim (R/ M) = tr. deg._F \bar{K}, plus the transformal multiplicity, or total dimension of {\rm Spec \, } R/t over {\rm Spec \, } K. As there is a chain of prime difference ideals of length rk_{ram}(K) between M and P, this total dimension is \geq rk_{ram}(K). Thus

So {rk_{{\, \rm val \,}}}(K) \leq d, and if equality holds then P \cap R(0) = (0).

The last statement is clear from the definition of P: if K is \omega-increasing, and a \in M, then {\, \rm val \,}(a) >0; we have {\, \rm val \,}(a) << {\, \rm val \,}(\sigma(a) ) << \ldots, so they are \hbox{\msbmten Q}- linearly independent, and hence cannot all be in \Delta; thus \sigma^m(a) \in P for some m \leq rk_{\hbox{\msbmten Q}} \Delta < \infty. \Box

View X[n] as a scheme over {\rm Spec \, } {k[\breve{t}]_\sigma}. Since {X_0}[n] \subset X[n]_0, it suffices to show that {\rm dim \, } X[n]_0 \leq d (where X[n]_0 is the fiber of X[n] above t_0=t_1=\ldots=0). X[n] arises by base extension from a scheme Y of finite type over {\rm Spec \, } S', where S' is a finitely generated k-subalgebra of S={k[\breve{t}]_\sigma}. Let K = k(t)_\sigma, K[m] = k(t_0,\ldots,t_m), S[m] = {k[\breve{t}]_\sigma}[m] = {k[\breve{t}]_\sigma} \cap K[m]; then for some m we can take S'={k[\breve{t}]_\sigma}[m]; and {\rm dim \, } _{K[m] } (Y {\otimes}_{S[m]} K[m]) =d.

Claim Let A be a finitely generated S[m]-algebra, such that x \mapsto t_mx is injective on A. Suppose {\rm dim \, } _{K[m] } (A {\otimes}_{S[m]} K[m]) = d. (This refers to Krull dimension.) Let A' = A/JA, where J is the ideal generated by (t_m,t_m t_{m-1}^{-1}, t_m t_{m-1}^{-2}, \ldots). Then A' is an S[m]/J = S[m-1]-algebra, x \mapsto t_{m-1}x is injective on A' , and {\rm dim \, } _{K[m-1] } (A' {\otimes}_{S[m-1]} K[m-1]) \leq d

In whatever order it is obtained, the ring is characterized by the universal property for difference rings W with maps R \to W, p=0 in W, such that the image of T is invertible, and \sigma(x)=x^q. \Box

Similar to that of ?

Let h: S \to K be a surjective homomorphism with kernel M, K a finite extension field of k. Let s_1,\ldots,s_m be generators for S as a k-algebra.

Consider first the case e=0, i.e. K/k separable. Using the primitive element theorem, K is generated over k by one element a=h(s_0); and one relation, the minimal monic polynomial P(a)=0, P \in k[X]. We have h(s_i) = Q_i(a) = h (Q_i(s_0)) (some Q_i \in k[X]), so s_i-Q_i(s_0) \in M; and clearly M is generated by the m+1 elements P(s_0),s_i - Q_i(s_0).

In general, K/k is generated by e+1 elements, and e+1 relations. (Let K \leq K_s \leq K, where K_s/k separable, K/K_s purely inseparable. Then K_s/k can be presented by one generator a_0 and one relation, as above. K/K_s is generated by \leq e generators a_1,\ldots,a_e and \leq e relations; these can be viewed as a relation over k between a_0 and the a_i.) As above, the maximal ideal M is generated by the preimages of the above relations. \Box

Let A be a k-subspace of S with {\rm dim}_k (A) = {\rm dim}_k (S/ MS) and A + M S = S. Let T be the k-span of RA. So T is an R-submodule of S, and {\rm dim}_k (T) \leq ({\rm dim}_k R) ({\rm dim}_k (S/ MS) ). Also T+MS = S. By Nakayama (applied to the R-module S/T), T=S.