Galois Extensions

Given a polynomial equation, when can we understand all of its symmetries by counting a single number? This is the concrete question that Galois extensions answer. A field extension K/FK/F carries a group of automorphisms Aut(K/F)\mathrm{Aut}(K/F) fixing FF pointwise. In general this group can be far too small to see anything interesting. It may even be trivial. The extension is called Galois precisely when this group is as large as possible. This happens when ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)| = [K:F]. This equality holds exactly when K/FK/F is both normal and separable. This equality is not a curiosity. It is the hinge on which the entire structure of Galois theory turns. Once we know ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)| = [K:F], we can ask a sharper question. Which subfields of KK correspond to which subgroups of Aut(K/F)\mathrm{Aut}(K/F)? The fixed-field characterization of Galois extensions states this precisely. K/FK/F is Galois if and only if FF is exactly the fixed field of Aut(K/F)\mathrm{Aut}(K/F), no larger. This is the precise statement that makes this correspondence a genuine bijection, rather than a rough analogy. The big picture is a translation device. Field-theoretic questions become group-theoretic questions once we know we have a Galois extension. These questions include: does this polynomial factor, is this extension solvable by radicals, what are the intermediate fields. They translate into: subgroup structure, normal subgroups, solvability of a finite group. Every other theorem in the subject presupposes this equality and this fixed-field statement. This includes the Fundamental Theorem, the classification of cyclotomic and Kummer extensions, and the unsolvability of the quintic. This article proves, completely and from first principles, the equivalence of five conditions that define a Galois extension. The two headline results are proved in full: the counting equality ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)|=[K:F] and the fixed-field characterization. It assumes familiarity with Normal Extensions, Separable Extensions, and The Galois Group of an Extension. It also sets up, without yet proving, the correspondence developed in The Fundamental Theorem of Galois Theory.

Key takeaways
  • A finite extension K/FK/F is Galois iff it is normal and separable.
  • Equivalently: ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)| = [K:F] — the automorphism group is exactly as large as degree allows.
  • Equivalently: the fixed field of Aut(K/F)\mathrm{Aut}(K/F) is precisely FF, with no room to spare.
  • Equivalently: KK is a splitting field over FF of some separable polynomial.
  • When Galois, we write Gal(K/F):=Aut(K/F)\mathrm{Gal}(K/F) := \mathrm{Aut}(K/F), the Galois group of the extension.
  • Key idea: normality plus separability is exactly the condition that makes automorphism-counting saturate the degree, turning symmetry into a faithful mirror of the field extension.

Where this lives today

The equivalence proved here is the mechanism behind the Fundamental Theorem of Galois Theory — see The Fundamental Theorem of Galois Theory — which upgrades the counting equality into an order-reversing bijection between subgroups of Gal(K/F)\mathrm{Gal}(K/F) and intermediate fields. Galois extensions of Q\mathbb{Q} generated by roots of unity, the cyclotomic extensions, are a central family whose Galois groups are computed directly (see Cyclotomic Extensions) using exactly the separable-splitting-field criterion proved above. In algebraic number theory, Galois extensions of number fields organize ramification and decomposition of primes, with the Galois group replaced by decomposition and inertia subgroups. In cryptography, the Galois theory of finite fields — as in Example 2 — underlies the arithmetic of Fpn\mathbb F_{p^n} used in elliptic-curve and pairing-based cryptosystems, where Frobenius automorphisms provide fast field arithmetic. In coding theory, Galois groups of finite field extensions control the structure of BCH and Reed–Solomon codes. In computer algebra, algorithms for polynomial factorization and for computing Galois groups of specific integer polynomials rely on precisely detecting normality and separability as in this article. In differential algebra, an analogous "differential Galois theory" replaces finite automorphism groups by algebraic groups classifying symmetries of solutions to linear differential equations.

:::quiz[If K/FK/F is finite and ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)| = [K:F], which further conclusion is justified?]

  • K/FK/F need not be separable
  • K/FK/F is automatically both normal and separable ✓
  • K/FK/F is normal but its separability cannot be determined
  • FF must be algebraically closed

:::quiz[Why does K=Q(53)K=\mathbb{Q}(\sqrt[3]{5}) fail to be Galois over Q\mathbb{Q}?]

  • Because x3−5x^3-5 is inseparable over Q\mathbb{Q}
  • Because [K:Q][K:\mathbb{Q}] is not finite
  • Because the minimal polynomial x3−5x^3-5 does not split inside KK ✓
  • Because Q\mathbb Q is not a perfect field

:::faq[Why does ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)| = [K:F] require BOTH normality and separability, rather than just one?] Separability alone guarantees that the number of embeddings K↪F‾K \hookrightarrow \overline F over FF equals [K:F][K:F] (this is the content of the Lemma). But an embedding need not be an automorphism of KK — its image might land partly outside KK, inside the ambient algebraic closure. Normality is exactly the condition that forces every such embedding's image to be contained in (hence equal to, by dimension count) KK itself, converting embeddings into automorphisms. Drop normality and you may have many embeddings but few automorphisms (Example 1: 33 embeddings of Q(53)\mathbb Q(\sqrt[3]5), only 11 an automorphism). Drop separability and even the embedding count falls short of [K:F][K:F], so no repair is possible (the counterexample). Both hypotheses attack different halves of the same inequality chain ∣Aut(K/F)∣≤n(K/F)≤[K:F]|\mathrm{Aut}(K/F)| \le n(K/F) \le [K:F].

Ask the tutor: Can you walk through why the embedding-count lemma needs the extension-of-isomorphisms theorem, and what would go wrong if F‾\overline{F} weren't algebraically closed?

DefinitionGalois Extension

Let K/FK/F be a field extension, and assume it is algebraic. Let

Aut(K/F)={σ:K→K∣σ is a field automorphism and σ(a)=a ∀a∈F}.\mathrm{Aut}(K/F) = \{\sigma : K \to K \mid \sigma \text{ is a field automorphism and } \sigma(a) = a \ \forall a \in F\}.

For a subgroup (or any subset closed under the group operations is not needed — any subgroup) H≤Aut(K/F)H \le \mathrm{Aut}(K/F), define the fixed field

KH={x∈K∣σ(x)=x  for all σ∈H}.K^H = \{ x \in K \mid \sigma(x) = x \ \text{ for all } \sigma \in H\}.

The extension K/FK/F is called a Galois extension if it is finite and

KAut(K/F)=F.K^{\mathrm{Aut}(K/F)} = F.

When K/FK/F is Galois we write Gal(K/F)\mathrm{Gal}(K/F) in place of Aut(K/F)\mathrm{Aut}(K/F), and call it the Galois group of the extension. (This is the fixed-field definition; the theorem below shows it is equivalent to "K/FK/F finite, normal, and separable," and to "∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)| = [K:F].")

Intuition

Think of Aut(K/F)\mathrm{Aut}(K/F) as measuring how much "room to move" the elements of KK have while the ground field FF stays perfectly still. If K/FK/F is not normal, some roots of minimal polynomials get left behind in an extension field, so automorphisms of KK can't reach them — the group is too small. If K/FK/F is not separable, a minimal polynomial has a repeated root, and any automorphism is forced to send that root to itself (there's only "one root's worth" of image to hit even though the polynomial has higher degree) — again the group is too small. Galois-ness says: neither obstruction occurs, so every root of every minimal polynomial is both present in KK and distinct, giving the automorphism group the maximum conceivable size, [K:F][K:F]. The fixed-field statement is the same idea viewed from the other side. If FF were a proper subfield of KAut(K/F)K^{\mathrm{Aut}(K/F)}, there would be "extra" elements of KK that every automorphism is powerless to move — symmetries blind to structure that isn't really there in FF. Galois-ness rules this out: the group sees everything that distinguishes KK from FF, and fixes nothing more than FF itself.

Figure
Splitting a cubic with one real root vs. one with three: normal+separable saturates automorphisms
Case A: K = ℚ(∛2) over F = ℚ, degree 3 — NOT normal ∛2 ω∛2 (∉ K) ω²∛2 (∉ K) |Aut(K/F)| = 1 ≠ [K:F] = 3 Case B: K = splitting field of x³−2 over F = ℚ, degree 6 — Galois ∛2 ω∛2 ω²∛2 |Aut(K/F)| = 6 = [K:F] = 6
TheoremEquivalence of Galois Conditions

Let K/FK/F be a finite field extension. The following are equivalent: 1. K/FK/F is normal and separable. 2. ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)| = [K:F]. 3. KK is the splitting field over FF of some separable polynomial f∈F[x]f \in F[x]. 4. KAut(K/F)=FK^{\mathrm{Aut}(K/F)} = F (i.e. K/FK/F is Galois, per the definition above). Consequently, "Galois," "normal and separable," and "∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)|=[K:F]" are interchangeable descriptions of a finite extension.

LemmaCounting embeddings

Let F⊆K⊆LF \subseteq K \subseteq L be fields with K/FK/F finite, and let F‾\overline{F} be an algebraic closure of FF containing LL. Let ι:F↪F‾\iota: F \hookrightarrow \overline{F} be the inclusion. The number of field embeddings K↪F‾K \hookrightarrow \overline{F} extending ι\iota, call it n(K/F)n(K/F), satisfies: (a) If K=F(α)K = F(\alpha) for a single algebraic element α\alpha with minimal polynomial mα∈F[x]m_\alpha \in F[x], then n(K/F)n(K/F) equals the number of distinct roots of mαm_\alpha in F‾\overline{F}. (b) In general, n(K/F)≤[K:F]n(K/F) \le [K:F], with equality if and only if mαm_\alpha has no repeated roots for every α∈K\alpha \in K — equivalently, iff K/FK/F is separable. (c) (Multiplicativity) For a tower F⊆E⊆KF \subseteq E \subseteq K, n(K/F)=n(K/E)⋅n(E/F)n(K/F) = n(K/E)\cdot n(E/F), where each embedding of EE is first extended to F‾\overline F and then extended further to KK in n(K/E)n(K/E) ways (this count is independent of which extension of E↪F‾E \hookrightarrow \overline F is chosen, by an isomorphism-extension argument).

Proof

(a) An embedding σ:F(α)→F‾\sigma: F(\alpha) \to \overline{F} extending ι\iota is determined by σ(α)\sigma(\alpha), since every element of F(α)F(\alpha) is a polynomial in α\alpha with coefficients in FF. Since σ\sigma fixes FF and is a ring homomorphism, applying σ\sigma to mα(α)=0m_\alpha(\alpha) = 0 gives mα(σ(α))=0m_\alpha(\sigma(\alpha)) = 0 (coefficients of mαm_\alpha lie in FF, hence are fixed). So σ(α)\sigma(\alpha) must be a root of mαm_\alpha in F‾\overline F. Conversely, for any root β∈F‾\beta \in \overline F of mαm_\alpha, the map F[x]/(mα)→F‾F[x]/(m_\alpha) \to \overline F, x↦βx \mapsto \beta, is a well-defined field homomorphism (since mα(β)=0m_\alpha(\beta)=0, this factors through the quotient by the ideal (mα)(m_\alpha)), and composing with the isomorphism F(α)≅F[x]/(mα)F(\alpha) \cong F[x]/(m_\alpha) gives an embedding sending α↦β\alpha \mapsto \beta. This is a bijection between embeddings and roots, proving (a).

(b) Induct on [K:F][K:F]. If [K:F]=1[K:F]=1, K=FK=F, n=1=[K:F]n=1=[K:F], trivial. Otherwise pick α∈K∖F\alpha \in K \setminus F, so F⊊F(α)⊆KF \subsetneq F(\alpha) \subseteq K. By (a), n(F(α)/F)=n(F(\alpha)/F) = (number of distinct roots of mαm_\alpha) ≤deg⁡mα=[F(α):F]\le \deg m_\alpha = [F(\alpha):F], with equality iff mαm_\alpha is separable (has no repeated root in F‾\overline F; note mαm_\alpha irreducible over FF, and separability of mαm_\alpha doesn't depend on which algebraically closed field we count roots in, since a repeated root of mαm_\alpha is detected by gcd⁡(mα,mα′)≠1\gcd(m_\alpha,m_\alpha')\ne 1 over FF itself). By induction on the tower F(α)⊆KF(\alpha) \subseteq K, n(K/F(α))≤[K:F(α)]n(K/F(\alpha)) \le [K:F(\alpha)]. Each of the n(F(α)/F)n(F(\alpha)/F) embeddings of F(α)F(\alpha) extends to some embedding of F‾\overline F (since F‾\overline F is algebraically closed and K/F(α)K/F(\alpha) is algebraic, an isomorphism-extension theorem guarantees this), and for each such extension the number of further extensions to KK is n(K/F(α))n(K/F(\alpha)) by definition, independent of the chosen extension because any two embeddings of F(α)F(\alpha) into F‾\overline F over FF differ by an automorphism of F‾\overline F over FF (both being generated by roots of the same mαm_\alpha), and post-composing with such an automorphism permutes the extensions to KK bijectively. Hence

n(K/F)=n(F(α)/F)⋅n(K/F(α))≤[F(α):F]⋅[K:F(α)]=[K:F],n(K/F) = n(F(\alpha)/F)\cdot n(K/F(\alpha)) \le [F(\alpha):F]\cdot[K:F(\alpha)] = [K:F],

by the tower law. Equality holds iff both factors are equalities: mαm_\alpha separable for every choice of generator α\alpha of a simple subextension in the chain, and by induction this happens for the whole tower iff every element of KK has a separable minimal polynomial over the relevant intermediate field. Chasing through the induction (or invoking that separability of K/FK/F means every α∈K\alpha \in K has separable minimal polynomial over FF, and using that minimal polynomials over intermediate fields divide those over FF, hence share the separability property, since FF-separability of mα∈F[x]m_\alpha \in F[x] forces its irreducible factor over any extension field to remain separable) shows equality in the whole chain holds if and only if K/FK/F is separable. This proves (b).

(c) This is exactly the multiplicative bookkeeping described in the equality case of the argument in (b): every embedding of EE over FF extends to F‾\overline F, and the extensions of that embedding to KK number n(K/E)n(K/E) regardless of which extension to F‾\overline F was used, by the same automorphism-of-F‾\overline F argument. Summing over the n(E/F)n(E/F) embeddings of EE gives n(K/F)=n(K/E)⋅n(E/F)n(K/F) = n(K/E)\cdot n(E/F). ■\blacksquare

∎
Proof

Proof of the Theorem (equivalence of (1)–(4)). We prove (1)⇒(2)⇒(3)⇒(1)(1) \Rightarrow (2) \Rightarrow (3) \Rightarrow (1), and separately (1)∧(2)⇒(4)⇒(1)(1)\wedge(2) \Rightarrow (4) \Rightarrow (1), closing the cycle. Throughout, fix an algebraic closure F‾\overline F of FF containing KK (possible since K/FK/F is finite, hence algebraic).

(1)⇒(2)(1) \Rightarrow (2). Assume K/FK/F is normal and separable. By the Lemma (b), since K/FK/F is separable, n(K/F)=[K:F]n(K/F) = [K:F], where n(K/F)n(K/F) counts embeddings K↪F‾K \hookrightarrow \overline F over FF. We claim every such embedding is in fact an automorphism of KK, i.e. its image is KK itself; this converts the embedding count into an automorphism count. Let σ:K→F‾\sigma: K \to \overline F be an FF-embedding. For any α∈K\alpha \in K with minimal polynomial mα∈F[x]m_\alpha \in F[x], the argument in Lemma (a) shows σ(α)\sigma(\alpha) is a root of mαm_\alpha in F‾\overline F. Since K/FK/F is normal, mαm_\alpha splits completely in KK (this is the defining property of a normal extension applied to α∈K\alpha \in K: the minimal polynomial of every element of KK splits in KK). Hence every root of mαm_\alpha in F‾\overline F already lies in KK, so σ(α)∈K\sigma(\alpha) \in K. As α\alpha was arbitrary, σ(K)⊆K\sigma(K) \subseteq K. Since σ\sigma is an injective FF-linear map (it is a field homomorphism, automatically injective) on the finite-dimensional FF-vector space KK, and σ(K)⊆K\sigma(K) \subseteq K, dimension-counting forces σ(K)=K\sigma(K) = K: an injective linear self-map of a finite-dimensional vector space is surjective. Thus σ\sigma restricts to an automorphism of KK fixing FF, i.e. σ∈Aut(K/F)\sigma \in \mathrm{Aut}(K/F). Conversely every element of Aut(K/F)\mathrm{Aut}(K/F) is in particular an FF-embedding K→F‾K \to \overline F (compose with inclusion K↪F‾K \hookrightarrow \overline F), and distinct automorphisms give distinct embeddings. So the embeddings K→F‾K \to \overline F over FF are in bijection with Aut(K/F)\mathrm{Aut}(K/F), giving

∣Aut(K/F)∣=n(K/F)=[K:F].|\mathrm{Aut}(K/F)| = n(K/F) = [K:F].

(2)⇒(3)(2) \Rightarrow (3). Assume ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)| = [K:F]. First we show K/FK/F is separable. By Lemma (b), n(K/F)≤[K:F]n(K/F) \le [K:F] always, with equality iff separable. But ∣Aut(K/F)∣≤n(K/F)|\mathrm{Aut}(K/F)| \le n(K/F) always (every automorphism of KK over FF, followed by inclusion into F‾\overline F, is one of the n(K/F)n(K/F) embeddings, and distinct automorphisms give distinct embeddings — this uses no normality, just injectivity of the map "automorphism ↦\mapsto embedding"). Hence

[K:F]=∣Aut(K/F)∣≤n(K/F)≤[K:F],[K:F] = |\mathrm{Aut}(K/F)| \le n(K/F) \le [K:F],

forcing n(K/F)=[K:F]n(K/F) = [K:F], so by Lemma (b), K/FK/F is separable. Next we show K/FK/F is normal, and identify a defining separable polynomial. Since K/FK/F is finite, write K=F(α1,…,αr)K = F(\alpha_1,\dots,\alpha_r) (any finite extension is finitely generated by algebraic elements — e.g. an FF-basis of KK). Let mi∈F[x]m_i \in F[x] be the minimal polynomial of αi\alpha_i, and set f=m1m2⋯mrf = m_1 m_2 \cdots m_r after removing duplicate irreducible factors so that ff has no repeated irreducible factors; note each mim_i is separable (as a factor of the polynomial defining a separable extension — indeed αi∈K\alpha_i \in K and K/FK/F separable means every element, in particular αi\alpha_i, has separable minimal polynomial) and distinct irreducible polynomials share no roots, so ff is a separable polynomial (no repeated roots in F‾\overline F). We claim KK is the splitting field of ff over FF. Since each αi∈K\alpha_i \in K, and we must show all roots of each mim_i lie in KK (splitting), consider any automorphism σ∈Aut(K/F)\sigma \in \mathrm{Aut}(K/F); it permutes the roots of mim_i that lie in KK among themselves (as in the proof of (a) ⇒\Rightarrow (b), σ(αi)\sigma(\alpha_i) is again a root of mim_i, and σ(αi)∈K\sigma(\alpha_i) \in K automatically since σ\sigma maps K→KK \to K). We must further show all roots of mim_i (not just those already known in KK) lie in KK. Consider the orbit Oi={σ(αi):σ∈Aut(K/F)}⊆KO_i = \{\sigma(\alpha_i) : \sigma \in \mathrm{Aut}(K/F)\} \subseteq K; this is a finite subset of roots of mim_i lying in KK. We show OiO_i is all the roots of mim_i in F‾\overline F, by an embedding-counting argument: extend F⊆F(αi)F \subseteq F(\alpha_i) and use that ∣Aut(K/F)∣=[K:F]=[K:F(αi)]⋅[F(αi):F]|\mathrm{Aut}(K/F)| = [K:F] = [K:F(\alpha_i)]\cdot[F(\alpha_i):F]. Every root β\beta of mim_i in F‾\overline F gives an FF-embedding F(αi)→F‾F(\alpha_i) \to \overline F, αi↦β\alpha_i \mapsto \beta (Lemma (a)); since K/FK/F has n(K/F)=[K:F]n(K/F) = [K:F] embeddings into F‾\overline F total, and by the multiplicativity of Lemma (c) these decompose as extensions of the [F(αi):F][F(\alpha_i):F]-many embeddings of F(αi)F(\alpha_i) (one for each root β\beta of mim_i, since mim_i separable means all [F(αi):F]=deg⁡mi[F(\alpha_i):F] = \deg m_i roots are distinct and give distinct embeddings) — each extending in exactly [K:F(αi)][K:F(\alpha_i)] ways. In particular for every root β\beta of mim_i, there is at least one embedding τ:K→F‾\tau: K \to \overline F over FF with τ(αi)=β\tau(\alpha_i) = \beta. But we showed above (in (1)⇒(2)(1)\Rightarrow(2)'s argument, which only used normality — so we cannot yet invoke it circularly here) — instead, argue directly: ∣Aut(K/F)∣=n(K/F)|\mathrm{Aut}(K/F)| = n(K/F) was already established in this direction (both equal [K:F][K:F]), meaning every FF-embedding K→F‾K \to \overline F has image KK itself (else the images would be embeddings not coming from automorphisms, but we've shown the count of automorphisms already saturates the total embedding count n(K/F)n(K/F), so no embedding can have image other than KK, since automorphisms are automatically embeddings and there is no room for extra ones). Hence τ(K)=K\tau(K) = K, so τ(αi)=β∈K\tau(\alpha_i) = \beta \in K. Thus every root β\beta of mim_i lies in KK: mim_i splits in KK. As this holds for each i=1,…,ri=1,\dots,r, f=∏mif = \prod m_i splits in KK, and KK is generated over FF by the αi\alpha_i's, which are among the roots of ff; hence KK is generated over FF by roots of ff and ff splits in KK, so KK is precisely the splitting field of ff over FF. This proves (3), with ff separable as shown.

(3)⇒(1)(3) \Rightarrow (1). Assume KK is the splitting field over FF of a separable polynomial f∈F[x]f \in F[x]. That K/FK/F is normal is a standard fact about splitting fields (a splitting field of any polynomial is a normal extension of the base field — every splitting field is normal because any FF-embedding of KK into F‾\overline{F} must send roots of ff to roots of ff, hence stays within the subfield generated by those roots, which is KK itself, forcing every embedding to be an automorphism; equivalently, minimal polynomials of elements of KK, being factors of powers of ff up to the structure of KK, split in KK). We take this standard characterization of splitting fields as previously established under Normal Extensions and do not reprove it here. It remains to prove K/FK/F is separable, i.e. every α∈K\alpha \in K has separable minimal polynomial over FF. Since KK is generated over FF by the roots β1,…,βn\beta_1,\dots,\beta_n of ff (with n=deg⁡fn = \deg f, all distinct since ff is separable), it suffices to show that any subfield generated over FF by roots of a separable polynomial is separable over FF; this again is a standard closure property of separable extensions (separability is preserved under composita and subextensions generated by separable elements), which we invoke from Separable Extensions. Hence K/FK/F is separable, completing (1)(1). This closes the cycle (1)⇒(2)⇒(3)⇒(1)(1) \Rightarrow (2) \Rightarrow (3) \Rightarrow (1).

(2)⇒(4)(2) \Rightarrow (4). Assume ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)| = [K:F]; by the cycle just proved this is equivalent to (1), so K/FK/F is normal and separable. Let F′=KAut(K/F)F' = K^{\mathrm{Aut}(K/F)}, the fixed field; clearly F⊆F′F \subseteq F' since every σ∈Aut(K/F)\sigma \in \mathrm{Aut}(K/F) fixes FF by definition of Aut(K/F)\mathrm{Aut}(K/F). We must show F′⊆FF' \subseteq F, i.e. F′=FF' = F. Regard KK as an extension of the intermediate field F′F'; every automorphism in Aut(K/F)\mathrm{Aut}(K/F) fixes F′F' pointwise by definition of fixed field, so Aut(K/F)⊆Aut(K/F′)\mathrm{Aut}(K/F) \subseteq \mathrm{Aut}(K/F'). Also every automorphism fixing F′F' certainly fixes the smaller field FF, so Aut(K/F′)⊆Aut(K/F)\mathrm{Aut}(K/F') \subseteq \mathrm{Aut}(K/F). Hence Aut(K/F′)=Aut(K/F)\mathrm{Aut}(K/F') = \mathrm{Aut}(K/F), of size [K:F][K:F]. Now K/F′K/F' is normal (since K/FK/F is normal: every α∈K\alpha \in K has minimal polynomial over FF splitting in KK; its minimal polynomial over the larger field F′F' divides this splitting polynomial's factor, hence also splits in KK — a standard fact that normality persists over intermediate fields) and separable (a subextension of a separable extension is separable, since minimal polynomials over F′F' divide those over FF, and divisors of separable polynomials are separable). So by (1)⇒(2)(1)\Rightarrow(2) already proved, ∣Aut(K/F′)∣=[K:F′]|\mathrm{Aut}(K/F')| = [K:F']. Combining with ∣Aut(K/F′)∣=∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F')| = |\mathrm{Aut}(K/F)| = [K:F] gives

[K:F′]=[K:F].[K:F'] = [K:F].

Since F⊆F′⊆KF \subseteq F' \subseteq K and [K:F]=[K:F′]⋅[F′:F][K:F] = [K:F']\cdot[F':F] (tower law) with [K:F′]=[K:F][K:F']=[K:F] finite and nonzero, we get [F′:F]=1[F':F] = 1, i.e. F′=FF' = F. This proves (4).

(4)⇒(1)(4) \Rightarrow (1). Assume KAut(K/F)=FK^{\mathrm{Aut}(K/F)} = F. We show K/FK/F is separable, then normal. Separability: Let α∈K\alpha \in K be arbitrary, and let O={σ(α):σ∈Aut(K/F)}O = \{\sigma(\alpha) : \sigma \in \mathrm{Aut}(K/F)\} be its finite orbit under Aut(K/F)\mathrm{Aut}(K/F) (finite because Aut(K/F)\mathrm{Aut}(K/F) is a finite group — it injects into the set of embeddings K→F‾K \to \overline F, which is finite of size ≤[K:F]\le [K:F] by Lemma (b), as K/FK/F is finite). List the distinct elements of OO as β1=α,β2,…,βs\beta_1=\alpha, \beta_2,\dots,\beta_s, and form

g(x)=∏j=1s(x−βj)∈K[x].g(x) = \prod_{j=1}^s (x - \beta_j) \in K[x].

Every σ∈Aut(K/F)\sigma \in \mathrm{Aut}(K/F) permutes {β1,…,βs}\{\beta_1,\dots,\beta_s\} (since σ\sigma maps the orbit of α\alpha to the orbit of σ(α)\sigma(\alpha), which is the same orbit, as Aut(K/F)\mathrm{Aut}(K/F) is a group acting on itself's orbits — more precisely, composing σ\sigma with any τ\tau with τ(α)=βj\tau(\alpha)=\beta_j gives στ(α)=σ(βj)\sigma\tau(\alpha) = \sigma(\beta_j), still in the orbit of α\alpha). Hence σ\sigma permutes the roots βj\beta_j, so it permutes the linear factors of gg, and therefore fixes each coefficient of gg (symmetric functions of the βj\beta_j are permutation-invariant). This holds for every σ∈Aut(K/F)\sigma \in \mathrm{Aut}(K/F), so every coefficient of gg lies in KAut(K/F)=FK^{\mathrm{Aut}(K/F)} = F by hypothesis (4). Thus g∈F[x]g \in F[x], g(α)=0g(\alpha) = 0, and gg has distinct roots β1,…,βs\beta_1,\dots,\beta_s by construction (they were listed as distinct). Since the minimal polynomial mαm_\alpha of α\alpha over FF divides gg (as g∈F[x]g \in F[x] vanishes at α\alpha), and gg has no repeated roots, mαm_\alpha has no repeated roots either. Hence α\alpha is separable over FF. As α∈K\alpha \in K was arbitrary, K/FK/F is separable. Normality: With the same notation, we show mα=gm_\alpha = g, which shows mαm_\alpha splits in KK (its roots are the βj∈K\beta_j \in K), giving normality since α\alpha was arbitrary. We know mα∣gm_\alpha \mid g. Conversely, the identity map id∈Aut(K/F)\mathrm{id} \in \mathrm{Aut}(K/F) shows α=β1\alpha = \beta_1 is a root of mαm_\alpha; and for each jj, since βj=σ(α)\beta_j = \sigma(\alpha) for some σ∈Aut(K/F)\sigma \in \mathrm{Aut}(K/F), applying σ\sigma to mα(α)=0m_\alpha(\alpha)=0 (with mα∈F[x]m_\alpha \in F[x] fixed coefficient-wise by σ\sigma) gives mα(σ(α))=mα(βj)=0m_\alpha(\sigma(\alpha)) = m_\alpha(\beta_j) = 0. So every βj\beta_j is a root of mαm_\alpha, meaning g=∏(x−βj)g = \prod(x-\beta_j) divides mαm_\alpha (as mαm_\alpha has all ss distinct roots βj\beta_j, each simple since... to be careful: g∣mαg \mid m_\alpha requires the βj\beta_j to be roots of mαm_\alpha with at least their multiplicity in gg, which is 11 each; since all βj\beta_j are distinct roots of mαm_\alpha, ∏(x−βj)=g\prod(x-\beta_j) = g divides mαm_\alpha). Combined with mα∣gm_\alpha \mid g and both monic, mα=gm_\alpha = g. Since gg splits completely in KK by construction (its roots are β1,…,βs∈K\beta_1,\dots,\beta_s \in K), mαm_\alpha splits in KK. As α∈K\alpha \in K was arbitrary, every element of KK has minimal polynomial splitting in KK: K/FK/F is normal. This proves (4)⇒(1)(4) \Rightarrow (1), closing all implications and establishing the full equivalence of (1), (2), (3), (4). ■\blacksquare

∎

Worked examples

Example 1: A genuinely non-Galois extension, diagnosed and repaired. Let F=QF = \mathbb{Q} and let α=53∈R\alpha = \sqrt[3]{5} \in \mathbb{R}, and set K=Q(53)K = \mathbb{Q}(\sqrt[3]{5}). The minimal polynomial of α\alpha over Q\mathbb{Q} is mα(x)=x3−5m_\alpha(x) = x^3 - 5 (irreducible by Eisenstein at p=5p=5), so [K:F]=3[K:F] = 3. Is K/FK/F Galois? Check normality: the other roots of x3−5x^3-5 are ω53\omega\sqrt[3]{5} and ω253\omega^2\sqrt[3]{5} where ω=e2πi/3\omega = e^{2\pi i/3} is a primitive cube root of unity. These are complex (non-real) numbers, while K⊂RK \subset \mathbb{R}. So x3−5x^3-5 does not split in KK: K/FK/F is not normal. Directly counting automorphisms confirms this. Any σ∈Aut(K/F)\sigma \in \mathrm{Aut}(K/F) must send α↦\alpha \mapsto a root of x3−5x^3-5 lying in KK (by Lemma (a)'s argument, since σ(α)\sigma(\alpha) is a root of mαm_\alpha, and σ(K)⊆K\sigma(K)\subseteq K trivially as σ\sigma is an automorphism of KK). The only real root of x3−5x^3-5 is α\alpha itself (the other two are genuinely complex, since x3−5=(x−α)(x−ωα)(x−ω2α)x^3 - 5 = (x-\alpha)(x-\omega\alpha)(x - \omega^2\alpha) and ω∉R\omega \notin \mathbb{R}). Since K⊂RK \subset \mathbb{R}, the only possible image is σ(α)=α\sigma(\alpha) = \alpha, forcing σ=id\sigma = \mathrm{id} (as σ\sigma is determined by its value on the generator α\alpha). So Aut(K/F)={id}\mathrm{Aut}(K/F) = \{\mathrm{id}\}, giving ∣Aut(K/F)∣=1≠3=[K:F]|\mathrm{Aut}(K/F)| = 1 \ne 3 = [K:F]. Consistently, the fixed field of Aut(K/F)={id}\mathrm{Aut}(K/F) = \{\mathrm{id}\} is all of KK, not FF: KAut(K/F)=K≠QK^{\mathrm{Aut}(K/F)} = K \ne \mathbb{Q}. All four conditions of the theorem fail simultaneously, exactly as the equivalence predicts. To repair this, enlarge KK to its normal closure L=Q(53,ω)L = \mathbb{Q}(\sqrt[3]{5}, \omega), the splitting field of x3−5x^3-5 over Q\mathbb{Q}. Now [L:Q]=6[L:\mathbb{Q}] = 6 (since [Q(53):Q]=3[\mathbb{Q}(\sqrt[3]5):\mathbb{Q}]=3, and adjoining ω\omega, a root of the irreducible-over-Q(53)\mathbb{Q}(\sqrt[3]5) quadratic x2+x+1x^2+x+1 — irreducible because Q(53)⊂R\mathbb Q(\sqrt[3]5) \subset \mathbb R contains no non-real root — multiplies the degree by 22). Since x3−5x^3-5 has characteristic 00 coefficients, it is automatically separable (its formal derivative 3x23x^2 shares no root with x3−5x^3-5, as x=0x=0 is not a root of x3−5x^3-5). By condition (3) of the theorem, L/QL/\mathbb{Q} is Galois, so ∣Aut(L/Q)∣=6|\mathrm{Aut}(L/\mathbb{Q})| = 6, and one checks directly this group is isomorphic to S3S_3 (permuting the three roots α,ωα,ω2α\alpha,\omega\alpha,\omega^2\alpha), matching [L:Q]=6[L:\mathbb Q]=6 exactly.

Example 2: A Galois extension over a field of positive characteristic — finite fields. Let p=3p=3 and consider F=F3F = \mathbb{F}_3 and K=F27=F33K = \mathbb{F}_{27} = \mathbb{F}_{3^3} (see Finite Fields). We have [K:F]=3[K:F] = 3, since F27\mathbb F_{27} is the (unique, up to isomorphism) field of order 333^3 and F27\mathbb F_{27} is a 33-dimensional F3\mathbb F_3-vector space. Every finite field is a splitting field: K=F27K = \mathbb{F}_{27} is precisely the splitting field over F=F3F=\mathbb F_3 of f(x)=x27−xf(x) = x^{27} - x, since the roots of x27−xx^{27}-x in an algebraic closure of F3\mathbb F_3 are exactly the elements of F27\mathbb F_{27} (the multiplicative group F27×\mathbb F_{27}^\times has order 2626, and every nonzero element satisfies x26=1x^{26}=1, hence x27=xx^{27}=x; together with x=0x=0, all 2727 elements of F27\mathbb F_{27} satisfy x27=xx^{27}=x, and since deg⁡(x27−x)=27\deg(x^{27}-x) = 27 there can be no other roots). Moreover f(x)=x27−xf(x) = x^{27}-x is separable: its formal derivative is f′(x)=27x26−1=−1f'(x) = 27x^{26} - 1 = -1 in characteristic 33 (since 27≡0(mod3)27 \equiv 0 \pmod 3), which is a nonzero constant, so gcd⁡(f,f′)=1\gcd(f,f')=1 and ff has no repeated roots. By condition (3), K/FK/F is Galois, so ∣Aut(K/F)∣=[K:F]=3|\mathrm{Aut}(K/F)| = [K:F] = 3. We can identify Aut(K/F)\mathrm{Aut}(K/F) explicitly: the Frobenius map φ:K→K\varphi: K \to K, φ(x)=x3\varphi(x) = x^3, is a field automorphism (it is a ring homomorphism since (x+y)3=x3+y3(x+y)^3 = x^3+y^3 in characteristic 33 — the cross terms in the binomial expansion carry factors of 33 that vanish — and it is injective, hence bijective on the finite set KK), and it fixes F=F3F = \mathbb F_3 pointwise (Fermat's little theorem: a3=aa^3 = a for all a∈F3a \in \mathbb F_3). So φ∈Aut(K/F)\varphi \in \mathrm{Aut}(K/F). Its powers id,φ,φ2\mathrm{id}, \varphi, \varphi^2 are distinct: if φk=id\varphi^k = \mathrm{id} for k<3k<3, every element of KK would satisfy x3k=xx^{3^k}=x, giving at most 3k<273^k < 27 roots for the polynomial x3k−xx^{3^k}-x, contradicting that all 2727 elements of KK would need to be roots. Since ∣Aut(K/F)∣=3|\mathrm{Aut}(K/F)|=3 exactly, we conclude

Gal(K/F)={id,φ,φ2}≅Z/3Z,\mathrm{Gal}(K/F) = \{\mathrm{id}, \varphi, \varphi^2\} \cong \mathbb{Z}/3\mathbb{Z},

generated by the Frobenius automorphism — a fact that generalizes to Gal(Fpn/Fp)≅Z/nZ\mathrm{Gal}(\mathbb F_{p^n}/\mathbb F_p) \cong \mathbb Z/n\mathbb Z for every prime pp and every n≥1n \ge 1.

CounterexampleDropping separability breaks the equality — and the theorem itself needs it

The theorem requires K/FK/F finite with the further hypothesis, in condition (1), of both normal and separable. Drop separability and keep normality: let pp be a prime, let tt be an indeterminate, and consider

F=Fp(t),K=Fp(t1/p)=F(α),αp=t.F = \mathbb{F}_p(t), \qquad K = \mathbb{F}_p(t^{1/p}) = F(\alpha), \quad \alpha^p = t.

Here α\alpha is a root of f(x)=xp−t∈F[x]f(x) = x^p - t \in F[x]. This polynomial is irreducible over FF (an application of Eisenstein-type criteria in the ring Fp[t]\mathbb F_p[t], viewing tt as a "prime" of Fp[t]\mathbb F_p[t], localized appropriately — or directly: any factorization of xp−tx^p-t over F=Fp(t)F=\mathbb F_p(t) would force a root in FF by a valuation argument on the tt-adic valuation, and tt has no pp-th root in Fp(t)\mathbb F_p(t) by degree considerations), so [K:F]=p[K:F] = p. Over KK, since we are in characteristic pp, the Frobenius identity gives

xp−t=xp−αp=(x−α)px^p - t = x^p - \alpha^p = (x-\alpha)^p

(using (x−α)p=xp−αp(x-\alpha)^p = x^p - \alpha^p in characteristic pp, since all the intermediate binomial coefficients (pk)\binom{p}{k} for 0<k<p0<k<p are divisible by pp and vanish). So ff splits completely in KK — with a single root α\alpha of multiplicity pp. Since KK is generated by this one root, KK is the splitting field of ff over FF, hence K/FK/F is normal. But f=xp−tf = x^p - t has repeated roots (all pp of its roots coincide at α\alpha), so ff is not separable, and in fact α\alpha's minimal polynomial ff is inseparable: K/FK/F is not separable. Now check the automorphism count. Any σ∈Aut(K/F)\sigma \in \mathrm{Aut}(K/F) sends α\alpha to a root of ff in KK; but ff has only the single root α\alpha (with multiplicity pp, but as a set of roots it is {α}\{\alpha\}). So σ(α)=α\sigma(\alpha) = \alpha forcibly, giving σ=id\sigma = \mathrm{id}. Hence

∣Aut(K/F)∣=1≠p=[K:F].|\mathrm{Aut}(K/F)| = 1 \ne p = [K:F].

Normality alone was not enough: without separability, the automorphism count collapses to 11 regardless of the degree pp, and the fixed field KAut(K/F)=K≠FK^{\mathrm{Aut}(K/F)} = K \ne F as well. This shows separability cannot be dropped from condition (1), and correspondingly the equality in condition (2) and the fixed-field statement in (4) both fail the instant it is removed.

Scope, edge cases & common traps

SituationWhat holdsWhat can fail
K/FK/F finite, normal, separableAll four conditions equivalent; Gal(K/F)\mathrm{Gal}(K/F) well-defined, ∣Gal(K/F)∣=[K:F]|\mathrm{Gal}(K/F)|=[K:F]—
K/FK/F normal but not separable (char pp)Normality alone∣Aut(K/F)∣<[K:F]|\mathrm{Aut}(K/F)| < [K:F]; fixed field strictly bigger than FF
K/FK/F separable but not normalSeparability alone; n(K/F)=[K:F]n(K/F) = [K:F] (embedding count)∣Aut(K/F)∣<n(K/F)|\mathrm{Aut}(K/F)| < n(K/F); not every embedding is an automorphism
K/FK/F infinite algebraic (e.g. Q‾/Q\overline{\mathbb Q}/\mathbb Q)Analogous Galois theory exists (Krull topology)The finite-group counting argument here does not directly apply; requires profinite methods, outside this scope
FF perfect (e.g. Q\mathbb{Q}, finite fields, algebraically closed fields)Every algebraic extension is automatically separableNormality can still fail (Example 1) — perfectness only removes the separability obstruction

The mathematician behind it

Origins · 1 minÉvariste Galois and how this was discovered
ÉG
Évariste Galois
1811–1832 · Early 19th-century France, the last years of the classical algebra of equations

Galois developed the group-theoretic viewpoint on polynomial solvability as a teenager, submitting memoirs to the French Academy of Sciences that were repeatedly lost or rejected. He introduced the idea of associating a group of permutations of the roots of a polynomial to the polynomial itself, and showed that solvability of the equation by radicals corresponds to a solvability property of this group — the origin of the modern term "solvable group." He died at twenty from wounds sustained in a duel, having written down much of his mathematics the night before in a letter to a friend.

History
- 1830 — Galois submits his first memoir on the theory of equations to the Academy of Sciences.
- 1831 — A revised memoir is submitted and rejected by Poisson as "incomprehensible."
- 1832 — Galois dies on 31 May, after a duel; his papers, organized by Auguste Chevalier, survive.
- 1846 — Joseph Liouville publishes Galois's papers in the Journal de Mathématiques Pures et Appliquées.
- 1870s–1890s — Dedekind, Weber, and Artin recast Galois's ideas in the modern language of fields, automorphism groups, and fixed fields used in this article.
Is the fixed field of any finite group of automorphisms automatically giving a Galois extension?

Yes, in the following precise sense (a fact closely tied to, but distinct from, the theorem proved here): if GG is any finite group of automorphisms of a field KK, and F=KGF=K^G is its fixed field, then K/FK/F is automatically Galois with Gal(K/F)=G\mathrm{Gal}(K/F) = G. This is sometimes called Artin's theorem on fixed fields, and it is the reverse construction to the one in this article — instead of starting with an extension K/FK/F and asking whether it is Galois, one starts with a group GG acting on KK and manufactures a Galois extension K/KGK/K^G by definition. It is a companion result to, but not a restatement of, the equivalence proved above, and is used heavily in the proof of The Fundamental Theorem of Galois Theory.

Does every field extension have a "Galois closure"?

Yes, for any finite separable extension K/FK/F. One takes the splitting field LL over FF of the product of minimal polynomials of a generating set of KK; since K/FK/F is separable, this product can be chosen separable (after removing repeated irreducible factors), and LL is then a splitting field of a separable polynomial, hence Galois over FF by condition (3) of the theorem. This LL is called the Galois closure (or normal closure) of K/FK/F; it is the smallest Galois extension of FF containing KK. Example 1 illustrates this: Q(53)\mathbb Q(\sqrt[3]5) is not Galois over Q\mathbb Q, but its Galois closure is Q(53,ω)\mathbb Q(\sqrt[3]5,\omega), of degree 66.

Why is separability automatic in characteristic 00 and for finite fields, but not in general?

A field FF is called perfect if every algebraic extension of FF is separable; this holds automatically when char(F)=0\mathrm{char}(F)=0 (the derivative test for separability never produces the degenerate case, since gcd⁡(f,f′)\gcd(f,f') can only signal repeated roots via characteristic-pp phenomena in the derivative) and when FF is finite (every element of a finite field satisfies x∣F∣=xx^{|F|}=x, and the Frobenius map is an automorphism, which forces every irreducible polynomial to be separable). Function fields like Fp(t)\mathbb F_p(t) are not perfect: the element tt has no pp-th root in Fp(t)\mathbb F_p(t), and adjoining one, as in the counterexample, produces a genuinely inseparable extension. This is why the counterexample above needed a transcendental base field rather than Q\mathbb Q or a finite field.

Can an infinite extension be "Galois" in this sense?

The definition given here — via KAut(K/F)=FK^{\mathrm{Aut}(K/F)}=F — extends verbatim to infinite algebraic extensions, and conditions (1), (3), (4) (normal+separable, separable splitting field of a possibly infinite set of polynomials, fixed field equality) remain equivalent. However condition (2), the numeric equality ∣Aut(K/F)∣=[K:F]|\mathrm{Aut}(K/F)|=[K:F], needs reinterpretation since both sides can be infinite cardinals and a naive cardinality match is too weak to carry the same information; the correct infinite-dimensional replacement uses the Krull topology on Aut(K/F)\mathrm{Aut}(K/F) and profinite group theory. This article's counting proof is specific to the finite case and does not directly generalize; the infinite theory is a separate, more advanced topic.

How does this relate to separable but non-algebraic (e.g. purely transcendental) extensions?

It doesn't directly apply — the entire framework assumes K/FK/F is algebraic (indeed finite), since minimal polynomials and splitting fields are meaningless for transcendental elements. A purely transcendental extension like F(t)/FF(t)/F has an enormous automorphism group (all Möbius-type substitutions t↦at+bct+dt \mapsto \frac{at+b}{ct+d} with ad−bc≠0ad-bc\ne0, for instance) but is not algebraic at all, so the notion of "Galois" as defined here simply does not classify it; a different framework (transcendence degree, differential Galois theory for function fields) is needed for such extensions.

Before this

Where this leads