Definition 2.1.1 Let f\in F[x], F a field. f is solvable by radicals over F if there are elements a_i and n_i such that F=F_0\subseteq F(a_1)=F_1\subseteq F_1(a_2)=F_2\subseteq\cdots\subseteq F_{r-1}(a_r)=F_r, a_i^{n_i}\in F_{i-1}, and f splits in F_r. Each of the fields F_i is called a radical extension of F. \square
When f is solvable by radicals, we can choose the fields F_i with some nice additional properties. We see how to do this next.
An nth root of unity is any root of x^n-1. We say \omega is a primitive root of unity if \{1,\omega,\omega^2,\ldots, \omega^{n-1}\} is the set of all of the nth roots of unity. Let \omega_i be a primitive ith root of unity. Note that if \C\supseteq F\supseteq\Q, then F(\omega_i) is the splitting field of x^i-1 over F.
Suppose that f is solvable by radicals, and suppose that the fields J_i, elements b_i and exponents m_i, for i=1,\ldots,s, are as described by the definition. Let n be the maximum of all of the exponents m_i of the definition, and form the following chain of fields: \eqalign{F(\omega_3)=&L_1\subseteq L_1(\omega_4)=L_2\subseteq\cdots\subseteq L_{n-3}(\omega_n)=L_{n-2}=K_0\subseteq\cr &K_0(b_1)=K_1\subseteq K_1(b_2)=K_2\subseteq\cdots\subseteq K_{s-1}(b_s)=K_s.\cr} Rename the fields L_i and K_i as a single sequence F_i, i=1,\ldots,r, where r=n-2+s; rename the elements \omega_3,\ldots,\omega_n,b_1,\ldots,b_s as a_1,\ldots,a_r; and rename the integers 3,4,\ldots,n,m_1,\ldots,m_s as n_1,\ldots,n_r. Then f splits in F_r, a_i^{n_i}\in F_{i-1}, and F=F_0\subseteq F(a_1)=F_1\subseteq F_1(a_2)=F_2\subseteq\cdots\subseteq F_{r-1}(a_r)=F_r.
Lemma 2.1.2 Each of the fields F_i just described is a normal extension of the preceding field, F_{i-1}.
Proof. This is certainly true when the corresponding a_i is a primitive root of unity. Otherwise, note that a_i is a root of x^{n_i}-a_i^{n_i}\in F_{i-1}[x]. Let \omega=\omega_{n_i} and consider the set \{a_i,\omega a_i,\omega^2a_i,\ldots,\omega^{n_i-1}a_i\}. Each of these elements is in F_i, each is a root of x^{n_i}-a_i^{n_i}, and they are distinct. Thus, x^{n_i}-a_i^{n_i} splits in F_i=F_{i-1}(a_i), so F_i is the splitting field of x^{n_i}-a_i^{n_i} over F_{i-1} and by theorem 1.3.4, F_i is a normal extension of F_{i-1}. \qed
Proof. There are two cases, depending on whether a_i is a primitive root of unity or not. We do just the latter case; the former is similar. Remember that the roots of unity are all added first, so they are all in F_{i-1}.
Suppose that \sigma\in G(F_i/F_{i-1}). Since a_i^{n_i}\in F_{i-1}, \sigma(a_i^{n_i})=a_i^{n_i}, so (\sigma(a_i))^{n_i}-a_i^{n_i}= \sigma(a_i^{n_i}-a_i^{n_i})=0, or in other words, \sigma(a_i) is a root of x^{n_i}-a_i^{n_i}. By the proof of the previous lemma, \sigma(a_i)=\omega_{n_i}^{j(\sigma)}a_i for some integer j(\sigma). Now if \tau\in G(F_i/F_{i-1}), \eqalign{(\sigma\circ\tau)(a_i)=&\sigma(\omega_{n_i}^{j(\tau)}a_i)= \omega_{n_i}^{j(\tau)}\sigma(a_i)= \omega_{n_i}^{j(\tau)}\omega_{n_i}^{j(\sigma)}a_i=\cr &\omega_{n_i}^{j(\sigma)}\omega_{n_i}^{j(\tau)}a_i= \omega_{n_i}^{j(\sigma)}\tau(a_i)= \tau(\omega_{n_i}^{j(\sigma)}a_i)=(\tau\circ\sigma)(a_i).} Thus for all \sigma and \tau in G(F_i/F_{i-1}), \sigma\tau and \tau\sigma agree on a_i.
By theorem 19.3 in Gallian, every b\in F_i can be written as \sum c_ja_i^j, where c_j\in F_{i-1}. Then (\sigma\tau)(b)=\sum c_j((\sigma\tau)(a_i))^j= \sum c_j((\tau\sigma)(a_i))^j=(\tau\sigma)(b). Thus, \sigma\tau=\tau\sigma and so G(F_i/F_{i-1}) is abelian as desired. \qed
It need not be the case that F_r is a normal extension of F, but we can produce a new tower of fields with the properties we already have verified, and with the additional property that the ultimate field is a normal extension of F.
Proof. Let g_i be the minimal polynomial for a_i over F, and let b_i, 1\le i\le s be a list of all roots of all of the g_i, that is, s is the sum of the degrees of the g_i. For convenience, let b_1,\ldots, b_r be a_1,\ldots, a_r. Let N be the splitting field of the product g=\prod g_i, that is, N=F(b_1,\ldots,b_s)\supseteq F_r. We will extend the tower of fields: F=F_0\subseteq F_1\subseteq F_2\subseteq\cdots\subseteq F_r \subseteq F_{r+1}\subseteq\cdots\subseteq F_m=N.
For each root b_i of g_k, there is an isomorphism \sigma_i\colon F(a_k)\to F(b_i) that fixes F, by the lemma on page 372 in Gallian. By theorem 19.4, \sigma_i may be extended to an automorphism \overline\sigma_i on the splitting field N that fixes F(note that \sigma_i(g)=g). Now \overline\sigma_i(F_r) is a subfield of N that contains b_i and is isomorphic to F_r; in particular, \overline\sigma_i(F_r) may be viewed as arising from a tower of fields formed using the elements \overline\sigma_i(a_j), 1\le j\le r, and \overline\sigma_i(F_j)=\overline\sigma_i(F(a_1,\ldots,a_j))= F(\overline\sigma_i(a_1),\ldots,\overline\sigma_i(a_j)).
Thus, for each root b_i we have a corresponding sequence of elements \overline\sigma_i(a_j)=a_{i,j}, 1\le j\le r, one of which is b_i. Then we see that N=F(a_1,a_2,\ldots,a_r,a_{r+1,1},a_{r+1,2},\ldots,a_{s,r}), and \eqalign{F=&F_1\subseteq F_2\subseteq\cdots\subseteq F_r\subseteq N_{r+1,1}=F_r(a_{r+1,1})\subseteq N_{r+1,2}=N_{r+1,1}(a_{r+1,2}) \subseteq\cdots\cr &\subseteq N_{s,r-1}=N_{s,r-2}(a_{s,r-1})\subseteq N_{s,r}=N_{s,r-1}(a_{s,r})=N.\cr} Note that we add all of the elements \{a_{i,1}, a_{i,2},\ldots,a_{i,r}\} just to get b_i=a_{i,k}; we do this so that each field is obtained from the previous one by adding a root. Now we need to verify that each a_{i,j} is a root of some element in the previous field. There are two cases:
j=1: We need to show some power of a_{i,1} is in N_{i-1,r}: a_{i,1}^{n_1}=\left(\osigma_{i}(a_1)\right)^{n_1}=\osigma_{i}(a_1^{n_1}) \in \osigma_{i}(F)=F\subseteq N_{i-1,r}.
j>1: We need to show some power of a_{i,j} is in N_{i,j-1}: \eqalign{ a_{i,j}^{n_j}&=\left(\osigma_{i}(a_j)\right)^{n_j}=\osigma_{i}(a_j^{n_j}) \in \osigma_{i}(F_{j-1})=\osigma_i(F(a_1,\ldots,a_{j-1}))\cr &=F(\osigma_i(a_1),\ldots,\osigma_i(a_{j-1}))= F(a_{i,1},\ldots,a_{i,j-1}) \subseteq N_{i,j-1}.\cr} \qed
By the Fundamental Theorem, G(F_r/F_{i+1})\nsg G(F_r/F_i) and G(F_{i+1}/F_i)\cong \gmod{G(F_r/F_i)}/G(F_r/F_{i+1})., and by lemma 2.1.3, this group is abelian. So we have a tower of groups: \{\epsilon\}=G(F_r/F_r)\nsg G(F_r/F_{r-1})\nsg\cdots\nsg G(F_r/F_{i+1})\nsg G(F_r/F_i)\nsg\cdots\nsg G(F_r/F), in which each factor group G(F_r/F_i)/G(F_r/F_{i+1}) is abelian. It's not obvious that this property of G(F_r/F) is special, but in fact it is, and deserves a definition.
Definition 2.1.5 G is a solvable group if there are subgroups H_i of G such that \{e\}=H_0\nsg H_1\nsg H_2\nsg\cdots\nsg H_k=G and such that all factor groups H_{i+1}/H_i are abelian. \square
Remember that we are investigating what it means for a polynomial f to be solvable by radicals. Although f splits in F_r, F_r is likely to be much larger than the splitting field of f, and it is not obvious that a property held by F_r has much to say about f. Let E be the splitting field for f over F, so F\subseteq E\subseteq F_r. Since E is a normal extension of F, the Fundamental Theorem says G(E/F)\cong \gmod{G(F_r/F)}/G(F_r/E)., so G(E/F) is a factor group of a solvable group.
Lemma 2.1.6 A factor group of a solvable group is solvable.
Proof. Let H_0\nsg H_1\nsg H_2\nsg\cdots\nsg H_k=G be a tower of groups that illustrates that G is solvable, and let N\nsg G. Consider the groups \gmod H_0N/N. \le \gmod H_1N/N. \leq\cdots\leq \gmod H_kN/N. = \gmod G/N.. Suppose that x\in H_i, y\in H_{i+1}, n,m\in N, xnN=xN\in H_iN/N, ymN=yN\in H_{i+1}N/N. Then yNxN(yN)^{-1}=yxy^{-1}N. Since H_i\nsg H_{i+1}, yxy^{-1}\in H_i, so yxy^{-1}N\in H_iN/N and H_iN/N\nsg H_{i+1}N/N.
Now consider two elements of (H_{i+1}N/N)/(H_iN/N), (gnN)(H_iN\!/\!N)=(gN)(H_iN\!/\!N) and (hmN)(H_iN/N)=(hN)(H_iN/N). We want to show these elements commute, that is, (ghN)(H_iN/N)=(hgN)(H_iN/N). It suffices to show that (hgN)^{-1}(ghN)=g^{-1}h^{-1}ghN\in H_iN/N. Since H_{i+1}/H_i is abelian, g^{-1}h^{-1}gh\in H_i, this is true. \qed
To show that some polynomial f is not solvable by radicals, it is therefore sufficient to show that G(E/F) is not solvable, where E is the splitting field of f. We will show that for f=3x^5-15x+5, G(E/F)=S_5, and that S_n is not solvable when n\ge 5.