Let $K \subset L$ be a finite Galois extension, $M$ a field with $K \subset M \subset L$ and $G := \text{Aut}(L/K)$.

I want to show that if $\sigma \, \colon M \longrightarrow L$ is a $K$-homomorphism there exists a $\tau \in G$ with $\tau\vert_M=\sigma$.

$K$-homomorphism means that $\sigma\vert_K=\text{id}$.

$\\ \\ $

$\textbf{My ideas:}$

I thought to use the Isomorphism extension theorem, which I know in the following form:

Let $\varphi \, \colon K \longrightarrow K'$ be a field isomorphism, $f \in K[X]$ a polynomial with a root $x$ in a field extension of $K$ and $x'$ a root of $\varphi(f)$ in a field extension of $K'$.

Then there exists exactly one field isomorphism $\phi \, \colon K[x] \longrightarrow K'[x']$ with $\phi(x)=x'$ and $\phi\vert_K=\varphi$.

But I think this doesn't help in this situation.

  • $\begingroup$ What have you tried? Where do you struggle? Also what do you mean by $K$- homomorphism? I'm unfamiliar with this notation; does that mean it fixes $K$? $\endgroup$
    – DanZimm
    May 25 '14 at 0:30
  • $\begingroup$ Are the extensions finite? $\endgroup$ May 25 '14 at 9:13
  • $\begingroup$ @user148364 I'm fairly sure that there are infinite Galois extensions, although I don't know how they're categorised. $\endgroup$ May 25 '14 at 9:38

Suppose all degrees are finite.

We first prove that there are at most $\lvert M : K \rvert$ $K$-homomorphisms $M \to L$. Proceed by induction on $\lvert M : K \rvert$; suppose we know that there are $\le \lvert N : K \rvert$ $K$-homomorphisms $N \to L$. Let $\alpha \in L \setminus N$, with minimal polynomial $g \in N[X]$ of degree $k > 1$ over $N$. Then $\lvert N(\alpha) : N \rvert = k$, and any $\sigma : N \to L$ can be extended to at most $k$ $K$-homomorphisms $N(\alpha) \to L$, as $\alpha$ can only be mapped in one of the $k$ roots of $g$. Hence there are at most $\lvert N(\alpha) : N \rvert \cdot \lvert N : K \rvert = \lvert N(\alpha) : K \rvert$ $K$-homomorphisms $N(\alpha) \to L$.

Now consider the restriction of the elements $\tau \in Aut(L/K)$ to $M$. (Of course these are not necessarily automorphisms of $M$, but rather $K$-homomorphisms $M \to L$.) The restrictions of $\tau, \tau'$ coincide on $M$ iff $\tau Aut(L/M) = \tau' Aut(L/M)$. So by Lagrange there are $$ \frac{\lvert Aut(L/K) \rvert}{\lvert Aut(L/M) \rvert} = \frac{\lvert L:K \rvert}{\lvert L : M \rvert} = \lvert M : K \rvert $$ such restrictions.

So there are exactly $\lvert M : K \rvert$ $K$-homomorphisms $M \to L$, and each of these is the restriction to $M$ of an element of $Aut(L/K)$.

  • $\begingroup$ This is really nice, it is very reminiscent of the fact that $L/K$ is Galois iff $[L:K] = \#Aut(L/K)$, and I would never have thought to solve this by counting! $\endgroup$ May 25 '14 at 19:01
  • $\begingroup$ @TomOldfield, thanks! Actually, one can probably concoct a proof not dissimilar from yours from just the first part. $\endgroup$ May 26 '14 at 8:32

You can use the key theorem you state, if $L/M$ is finite then $L = M(x_1,\dots,x_n)$ for some $x_i$ in $L$. Let $M_0 = M$ and $M_{i+1} = M_i(x_{i+1})$.

The crucial part of this question is that $L/K$ is Galois, which for finite extensions is equivalent to being normal and separable. This means that the minimal polynomial of each of the $x_i/K$ splits completely into linear factors in $L$. Thus the minimal polynomial of each of the $x_i$ over $M_{i-1}$ also splits completely in $L$.

So, if we let $f_i(X)$ be the minimal polynomial for $x_i$/$M_{i-1}$, then for example $f_1(X)$ $\sigma(f_1)$ splits completely in $L$*. Thus we can extend $\sigma$ to some $\sigma_1:M_1\rightarrow L$ by the theorem, because we have shown that we have a root of $\sigma(f_1)$ in $L$. We can repeat this process iteratively until we have extended $\sigma$ up through each extension to $L$.

As pointed out in the comments, my original argument for this was invalid. We can prove this by taking $g_i(X)\in K[X]$ to be the minimal polynomial of $x_i$ over $K$, $f_i|g_i$, $g_i$ splits completely in $L$ and is fixed by $\sigma$ since it lies in $K[X]$. Thus $\sigma(f_i)|g_i$ also, so it too splits completely in $L$.

  • 1
    $\begingroup$ @user148364 For fields, any homomorphism is injective. (This is true since given any homomorphism of rings, the kernel is an ideal and the only ideals in a field are ${0}$ and the field itself, for example. There are other ways to show this too.) So $\sigma$ gives an isomorphism onto its image $\sigma(K)$ in $L$, and $L$ is a field extension of $\sigma(K)$ as required in the statement of the theorem. The theorem is almost always used when you have a homomorphism from one field into another one, since you can always apply what I have mentioned here. $\endgroup$ May 25 '14 at 10:49
  • 1
    $\begingroup$ @user148364 Yes, I did, sorry, I automatically use $K$ and $L$ for fields without thinking. To see that the image is all of $L$, we know $M(x_1,\dots,x_n)=L$ $\tau$ is a homomorphism from $L$ to $L$. It is injective (since all field homomorphisms are) and since $L$ is a finite dimensional $K$ vector space (this is what it means for the extension to be finite) $\tau$ is a linear map from $L$ to $L$ (thinking of $L$ as a $K$ vector space). For linear maps on vector spaces of the same dimension, being injective is equivalent to being surjective so the image of $\tau$ is all of $L$. $\endgroup$ May 25 '14 at 11:53
  • $\begingroup$ @user148364 No problem, happy to help! $\endgroup$ May 25 '14 at 12:00
  • $\begingroup$ @TomOldfield, could you please clarify the sentence $f_1(X)$ splits completely in $L$ and so $\sigma(f_1)$ does also? It is clear that if $\sigma(f_1)$ has a root in $L$, then it splits completely in $L$. But why does it have to have a root in $L$? $\endgroup$ May 25 '14 at 17:24
  • $\begingroup$ @AndreasCaranti Because if $f_1(X) = (X-\lambda_1)\dots(X-\lambda_m)$, $\sigma(f_1)(X)=(X-\sigma(\lambda_1))\dots(X-\sigma(\lambda_m))$. In fact, in general, if $f(X) \in K[X]$ has $\alpha$ as a root and $\sigma:K\rightarrow K'$ is a homomorphism, $\sigma(f)(X) \in k'[X]$ has $\sigma(\alpha)$ as a root, since $\sigma(f)(\sigma(\alpha)) = \sigma(f(\alpha))=\sigma(0)$. $\endgroup$ May 25 '14 at 18:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.