Proof.
Let \pi \in A be a uniformizer. Let w \in B and P \in B[t] be as in Lemma 15.115.13 (for B). Set e_1 = \text{ord}_ B(w), so that w and \pi ^{e_1} are associates in B. Pick z \in M generating M over L with \xi = P(z) \in K and n such that \pi ^ n\xi \in B as in the definition of the level of M over L, i.e., l = n/e_1.
The proof of this lemma is completely similar to the proof of Lemma 15.115.12. To explain what is going on, observe that
15.115.16.1
\begin{equation} \label{more-algebra-equation-first-congruence} P(z) \equiv z^ p - z \bmod \pi ^{-n + e_1}B \end{equation}
for any z \in L such that \pi ^{-n} P(z) \in B (use that z has valuation at worst -n/p and the shape of the polynomial P). Moreover, we have
15.115.16.2
\begin{equation} \label{more-algebra-equation-second-congruence} \xi _1 + \xi _2 + w^ p \xi _1 \xi _2 \equiv \xi _1 + \xi _2 \bmod \pi ^{-2n + pe_1}B \end{equation}
for \xi _1, \xi _2 \in \pi ^{-n}B. Finally, observe that n - e_1 = (l - 1)/e_1 and -2n + pe_1 = -(2l - p)e_1. Write m = n - e_1 \max (0, l - 1, 2l - p). The above shows that doing calculations in \pi ^{-n}B / \pi ^{-n + m}B the polynomial P behaves exactly as the polynomial z^ p - z. This explains why the lemma is true but we also give the details below.
Assumption (4) implies that \kappa _ A is perfect. Observe that m \leq e_1 and hence A/\pi ^ m is annihilated by w and hence p. Thus we may choose compatible ring maps \overline{\sigma } : \kappa _ A \to A/\pi ^ mA and \overline{\sigma } : \kappa _ B \to B/\pi ^ mB as in Lemma 15.115.10. We lift the second of these to a map of sets \sigma : \kappa _ B \to B. Then we can write
\xi = \sum \nolimits _{i = n, \ldots , n - m + 1} \sigma (\lambda _ i) \pi ^{-i} + \pi ^{-n + m)} b
for some \lambda _ i \in \kappa _ B and b \in B. Let
I = \{ i \in \{ n, \ldots , n - m + 1\} \mid \lambda _ i \in \kappa _ A\}
and
J = \{ j \in \{ n, \ldots , n - m + 1\} \mid \lambda _ i \not\in \kappa _ A\}
We will argue by induction on the size of the finite set J.
The case J = \emptyset . Here for all i \in \{ n, \ldots , n - m + 1\} we have \sigma (\lambda _ i) = a_ i + \pi ^{n - m}b_ i for some a_ i \in A and b_ i \in B by our choice of \overline{\sigma }. Thus \xi = \pi ^{-n} a + \pi ^{-n + m} b for some a \in A and b \in B. If p | n, then we write a = a_0^ p + \pi a_1 for some a_0, a_1 \in A (as the residue field of A is perfect). Set z_1 = - \pi ^{-n/p} a_0. Note that P(z_1) \in \pi ^{-n}B and that z + z_1 + w z z_1 is an element generating M over L (note that wz_1 \not= -1 as n < pe_1). Moreover, by Lemma 15.115.13 we have
P(z + z_1 + w z z_1) = P(z) + P(z_1) + w^ p P(z) P(z_1) \in K
and by equations (15.115.16.1) and (15.115.16.2) we have
P(z) + P(z_1) + w^ p P(z) P(z_1) \equiv \xi + z_1^ p - z_1 \bmod \pi ^{-n + m}B
for some b' \in B. This contradict the minimality of n! Thus p does not divide n. Consider the degree p extension K_1 of K given by P(y) = -\pi ^{-n}a. By Lemma 15.115.14 this extension is separable and totally ramified with respect to A. Thus L_1 = L \otimes _ K K_1 is a field and A_1 \subset B_1 is weakly unramified (Lemma 15.115.3). By Lemma 15.115.14 the ring M_1 = M \otimes _ K K_1 is either a product of p copies of L_1 (in which case we are done) or a field extension of L_1 of degree p. Moreover, in the second case, either C_1 is weakly unramified over B_1 (in which case we are done) or M_1/L_1 is degree p, Galois, totally ramified with respect to B_1. In this last case the extension M_1/L_1 is generated by the element z + y + wzy and we see that P(z + y + wzy) \in L_1 and
\begin{align*} P(z + y + wzy) & = P(z) + P(y) + w^ p P(z) P(y) \\ & \equiv \xi - \pi ^{-n}a \bmod \pi ^{-n + m}B_1 \\ & \equiv 0 \bmod \pi ^{-n + m}B_1 \end{align*}
in exactly the same manner as above. By our choice of m this means exactly that M_1/L_1 has level at most \max (0, l - 1, 2l - p). From now on we assume that J \not= \emptyset .
Suppose that j', j \in J such that j' = p^ r j for some r > 0. Then we set
z_1 = - \sigma (\lambda _ j) \pi ^{-j} - \sigma (\lambda _ j^ p) \pi ^{-pj} - \ldots - \sigma (\lambda _ j^{p^{r - 1}}) \pi ^{-p^{r - 1}j}
and we change z into z' = z + z_1 + wzz_1. Observe that z' \in M generates M over L and that we have \xi ' = P(z') = P(z) + P(z_1) + wP(z)P(z_1) \in L with
\xi ' \equiv \xi - \sigma (\lambda _ j) \pi ^{-j} + \sigma (\lambda _ j^{p^ r}) \pi ^{-j'} \bmod \pi ^{-n + m}B
by using equations (15.115.16.1) and (15.115.16.2) as above. Writing
\xi ' = \sum \nolimits _{i = n, \ldots , n - m + 1} \sigma (\lambda '_ i) \pi ^{-i} + \pi ^{-n + m}b'
as before we find that \lambda '_ i = \lambda _ i for i \not= j, j' and \lambda '_ j = 0. Thus the set J has gotten smaller. By induction on the size of J we may assume there is no pair j, j' of J such that j'/j is a power of p. (Please observe that in this procedure we may get thrown back into the case that J = \emptyset we treated above.)
For j \in J write \lambda _ j = \mu _ j^{p^{r_ j}} for some r_ j \geq 0 and \mu _ j \in \kappa _ B which is not a pth power. This is possible by our assumption (4). Let j \in J be the unique index such that j p^{-r_ j} is maximal. (The index is unique by the result of the preceding paragraph.) Choose r > \max (r_ j + 1) and such that j p^{r - r_ j} > n for j \in J. Let K_1/K be the extension of degree p^ r, totally ramified with respect to A, defined by (\pi ')^{p^ r} = \pi . Observe that \pi ' is the uniformizer of the corresponding discrete valuation ring A_1 \subset K_1. Observe that L_1 = L \otimes _ K K_1 is a field and L_1/L is totally ramified with respect to B (Lemma 15.115.3). Computing in the integral closure B_1 we get
\xi = \sum \nolimits _{i \in I} \sigma (\lambda _ i) (\pi ')^{-i p^ r} + \sum \nolimits _{j \in J} \sigma (\mu _ j)^{p^{r_ j}} (\pi ')^{-j p^ r} + \pi ^{-n + m} b_1
for some b_1 \in B_1. Note that \sigma (\lambda _ i) for i \in I is a qth power modulo \pi ^ m, i.e., modulo (\pi ')^{m p^ r}. Hence we can rewrite the above as
\xi = \sum \nolimits _{i \in I} x_ i^{p^ r} (\pi ')^{-i p^ r} + \sum \nolimits _{j \in J} \sigma (\mu _ j)^{p^{r_ j}} (\pi ')^{- j p^ r} + \pi ^{-n + m}b_1
Similar to our choice in the previous paragraph we set
\begin{align*} z_1 & - \sum \nolimits _{i \in I} \left(x_ i (\pi ')^{-i} + \ldots + x_ i^{p^{r - 1}} (\pi ')^{-i p^{r - 1}}\right) \\ & - \sum \nolimits _{j \in J} \left( \sigma (\mu _ j) (\pi ')^{- j p^{r - r_ j}} + \ldots + \sigma (\mu _ j)^{p^{r_ j - 1}} (\pi ')^{- j p^{r - 1}} \right) \end{align*}
and we change our choice of z into z' = z + z_1 + wzz_1. Then z' generates M_1 over L_1 and \xi ' = P(z') = P(z) + P(z_1) + w^ p P(z) P(z_1) \in L_1 and a calculation shows that
\xi ' \equiv \sum \nolimits _{i \in I} x_ i (\pi ')^{-i} + \sum \nolimits _{j \in J} \sigma (\mu _ j) (\pi ')^{- j p^{r - r_ j}} + (\pi ')^{(-n + m)p^ r}b'_1
for some b'_1 \in B_1. There is a unique j such that j p^{r - r_ j} is maximal and j p^{r - r_ j} is bigger than i \in I. If j p^{r - r_ j} \leq (n - m)p^ r then the level of the extension M_1/L_1 is less than \max (0, l - 1, 2l - p). If not, then, as p divides j p^{r - r_ j}, we see that M_1 / L_1 falls into case (C) of Lemma 15.115.14. This finishes the proof.
\square
Comments (0)
There are also: