Separable Extensions

Suppose you take a field FF and look at all the roots of an irreducible polynomial in some algebraic closure. A naive guess is that a polynomial of degree nn always has nn distinct roots there. This matches the count you learn in high-school algebra. This guess is correct over Q\mathbb{Q}, over R\mathbb{R}, over any field of characteristic 00. But it can fail dramatically over fields of prime characteristic. There, an irreducible polynomial of degree pkp^k can have only a single root. That root is repeated pkp^k times. Separability is the precise name for the well-behaved situation where roots do not collide. This distinction matters because so much of field theory depends on counting roots correctly. The Primitive Element Theorem, the correspondence at the heart of Galois Theory, and the basic counting formula [F(α):F]=deg⁡(min⁡.poly.ofα)[F(\alpha):F] = \deg(\min. poly. of \alpha) all interact with root multiplicities. These only behave predictably when extensions are separable. Inseparability is not a pathology invented to torment students. It is a genuine phenomenon that appears the moment characteristic pp enters the picture. Any general theory of fields must confront it honestly rather than sweep it under the rug. The big picture is this: separability is a condition on a single irreducible polynomial. Does it have repeated roots in a splitting field? This condition then upgrades to a condition on individual algebraic elements. Finally, it becomes a condition on entire field extensions. There is a remarkably clean algebraic test — the derivative criterion. It lets you decide separability without ever constructing a splitting field. You also never need to count roots directly. This test is the technical heart of the subject. It is the tool used constantly downstream. This article builds the theory from the ground up. It covers what it means for a polynomial to be separable. It covers the derivative test that makes this checkable. It also covers the notion of a separable element and a separable extension. Finally, it covers the sharp characterization of exactly when inseparability can occur. This involves characteristic pp, together with the Frobenius map. We prove every claim in detail. We work through asymmetric examples showing both the well-behaved and pathological cases. We also set up the language needed for Normal Extensions and the full machinery of Galois Extensions.

Key takeaways
  • A polynomial is separable if it has no repeated roots in a splitting field (equivalently, in any algebraic closure).
  • The derivative criterion: ff has a repeated root iff gcd⁡(f,f′)≠1\gcd(f, f') \neq 1, checkable purely by algebra in F[x]F[x], no root computation needed.
  • In characteristic 00, every irreducible polynomial is automatically separable — inseparability requires characteristic pp.
  • In characteristic pp, irreducible inseparable polynomials exist and always have the form g(xp)g(x^p) for irreducible gg.
  • An algebraic extension K/FK/F is separable if every element of KK has a separable minimal polynomial over FF.
  • Finite separable extensions are exactly the ones where the count [K:F][K:F] equals the number of FF-embeddings of KK into an algebraic closure.
  • Key idea: separability is the algebraic guarantee that "degree = number of distinct roots," and it is checkable by a gcd computation with the derivative, not by finding roots.

Where this lives today

Separability is the hidden hypothesis behind the Primitive Element Theorem, which guarantees a finite separable extension is generated by a single element; without separability, the theorem can fail. It is also the precise dividing line in the Galois Theory correspondence: a finite extension is Galois exactly when it is both normal and separable, and the fundamental theorem of Galois theory is stated for Galois extensions specifically because the embedding-counting theorem proved here only gives the correct group order, ∣Aut⁡(K/F)∣=[K:F]|\operatorname{Aut}(K/F)| = [K:F], in the separable (and normal) case. In arithmetic geometry and algebraic geometry, "generic smoothness" and the notion of a variety being smooth versus merely reduced trace back to separability of function field extensions in positive characteristic; inseparable extensions are exactly the source of the extra subtlety in characteristic-pp algebraic geometry (Frobenius morphisms, purely inseparable extensions, and the theory of perfect closures). In coding theory and cryptography over finite fields, the automatic separability of extensions of Fq\mathbb{F}_q (guaranteed by the Corollary above) is used implicitly every time one counts roots of a polynomial over a finite field or constructs field extensions for error-correcting codes. In computer algebra systems, the derivative criterion — via the Euclidean algorithm computing gcd⁡(f,f′)\gcd(f,f') — is the literal, executable test used by polynomial factorization routines (e.g. squarefree factorization algorithms) to detect and strip repeated factors before further symbolic computation.

:::quiz[For an irreducible f∈F[x]f \in F[x] with char⁡F=p>0\operatorname{char} F = p > 0, ff is inseparable if and only if]

  • ff has degree divisible by pp
  • f(x)=g(xp)f(x) = g(x^p) for some g∈F[x]g \in F[x] ✓
  • ff has a root in FF itself
  • f′(x)f'(x) has smaller degree than f(x)f(x)

:::quiz[If K/FK/F is a finite extension and F‾⊇K\overline F \supseteq K is an algebraic closure, then ∣Emb⁡F(K,F‾)∣=[K:F]|\operatorname{Emb}_F(K,\overline F)| = [K:F] exactly when]

  • K/FK/F is separable ✓
  • FF has characteristic 0 only
  • KK is a splitting field over FF
  • [K:F][K:F] is a prime number

Ask the tutor: Why does the Primitive Element Theorem's standard proof break down completely when the separable hypothesis is dropped?

DefinitionSeparable polynomial, separable element, separable extension

Let FF be a field and let F‾\overline{F} be a fixed algebraic closure of FF.

Separable polynomial. A nonconstant polynomial f∈F[x]f \in F[x] is separable if it has no repeated roots in F‾\overline{F}, i.e. when ff factors as

f(x)=c∏i=1n(x−αi)f(x) = c \prod_{i=1}^{n} (x - \alpha_i)

in F‾[x]\overline{F}[x], the roots α1,…,αn\alpha_1,\dots,\alpha_n are pairwise distinct. Equivalently (and this is the standard convention when ff is not irreducible), ff is separable if every irreducible factor of ff in F[x]F[x] has no repeated roots in F‾\overline{F}.

Separable element. Let K/FK/F be an algebraic field extension. An element α∈K\alpha \in K is separable over FF if its minimal polynomial mα(x)∈F[x]m_\alpha(x) \in F[x] is a separable polynomial.

Separable extension. An algebraic extension K/FK/F is a separable extension if every element α∈K\alpha \in K is separable over FF. Note: separability of an extension is defined purely in terms of minimal polynomials of elements; it does not require KK itself to be a splitting field of anything.

Intuition

Picture the roots of a degree-nn polynomial as nn points that could, in principle, land anywhere in F‾\overline{F}. Over Q\mathbb{Q} or R\mathbb{R}, calculus intuition (the graph "wiggling" through nn crossings) suggests these points are generically distinct, and indeed an irreducible polynomial over such fields never allows two of its roots to merge. Separability makes this intuition precise and extends it as far as it will go: it is exactly the statement that the nn roots, thought of as nn labeled points on a number line inside F‾\overline{F}, remain nn distinct points. The derivative gives an algebraic fingerprint of "collision." If α\alpha is a root of multiplicity ≥2\geq 2, then ff and f′f' share the factor (x−α)(x-\alpha), so gcd⁡(f,f′)\gcd(f,f') is nontrivial. This converts a geometric question (do two roots coincide?) into a computation you can do inside F[x]F[x] alone, using the Euclidean algorithm, without ever leaving the base field. The surprising twist happens in characteristic pp: it is possible for f′f' to vanish identically even though ff is nonconstant and irreducible — namely when every exponent appearing in ff is a multiple of pp, so that differentiating multiplies each coefficient by a multiple of pp, which is 00 in FF. Then the gcd test forces ff to share all of its roots pairwise, collapsing an entire irreducible polynomial onto a single repeated root.

Figure
Distinct roots (separable) vs. a collapsed root (inseparable) on the real/complex line picture
Separable: f(x) = (x-1)(x-2)(x-3), three distinct roots α₁=1 α₂=2 α₃=3 Inseparable: g(x) = x^p - t over F_p(t), one root of multiplicity p α, multiplicity p (all p roots coincide)

Central results

TheoremDerivative Criterion for Separability

Let FF be a field and let f∈F[x]f \in F[x] be a nonconstant polynomial. Let F‾\overline{F} be an algebraic closure of FF. Then ff has a repeated root in F‾\overline{F} if and only if

gcd⁡(f,f′)≠1\gcd(f, f') \neq 1

in F[x]F[x], where f′f' denotes the formal derivative of ff. Consequently: 1. ff is separable (no repeated roots) if and only if gcd⁡(f,f′)=1\gcd(f,f') = 1 in F[x]F[x]. 2. If ff is irreducible in F[x]F[x], then ff has a repeated root in F‾\overline{F} if and only if f′=0f' = 0 in F[x]F[x]. 3. If char⁡F=0\operatorname{char} F = 0, every irreducible polynomial in F[x]F[x] is separable. 4. If char⁡F=p>0\operatorname{char} F = p > 0, an irreducible f∈F[x]f \in F[x] is inseparable if and only if f(x)=g(xp)f(x) = g(x^p) for some g∈F[x]g \in F[x].

Proof

We use the algebra of the formal derivative, the The Minimal Polynomial and elementary properties of gcd's, computed either in F[x]F[x] or in F‾[x]\overline{F}[x] (a gcd computed by the Euclidean algorithm in F[x]F[x] does not change when we extend scalars to F‾[x]\overline{F}[x], since the Euclidean algorithm only uses field operations in FF, and the gcd is determined up to a unit by the sequence of remainders it produces).

Step 1: repeated roots detect common factors with f′f'. Work in F‾[x]\overline{F}[x], where ff splits completely:

f(x)=c∏i=1k(x−αi)ei,αi distinct, ei≥1.f(x) = c \prod_{i=1}^{k} (x-\alpha_i)^{e_i}, \qquad \alpha_i \text{ distinct}, \ e_i \geq 1.

Apply the product rule (valid formally in any commutative ring, and it is straightforward to check by induction on the number of factors that the usual product and sum rules for polynomial derivatives hold over any field):

f′(x)=c∑j=1kej(x−αj)ej−1∏i≠j(x−αi)ei.f'(x) = c \sum_{j=1}^k e_j (x-\alpha_j)^{e_j - 1} \prod_{i \neq j} (x - \alpha_i)^{e_i}.

Fix an index i0i_0 with ei0≥2e_{i_0} \geq 2 (such an index exists iff ff has a repeated root). Every term of the sum for f′f' with j≠i0j \neq i_0 contains the factor (x−αi0)ei0(x-\alpha_{i_0})^{e_{i_0}} (since that whole factor survives from the product ∏i≠j(x−αi)ei\prod_{i\ne j}(x-\alpha_i)^{e_i}, as ei0≥2≥1e_{i_0}\ge 2 \ge 1), and the term with j=i0j = i_0 contains (x−αi0)ei0−1(x-\alpha_{i_0})^{e_{i_0}-1} with ei0−1≥1e_{i_0}-1 \geq 1. So every term is divisible by (x−αi0)(x - \alpha_{i_0}), hence (x−αi0)∣f′(x)(x-\alpha_{i_0}) \mid f'(x) in F‾[x]\overline{F}[x]. Since also (x−αi0)∣f(x)(x-\alpha_{i_0}) \mid f(x), the two polynomials share the common factor (x−αi0)(x-\alpha_{i_0}) in F‾[x]\overline{F}[x], so gcd⁡(f,f′)≠1\gcd(f,f') \ne 1 (up to units) when computed in F‾[x]\overline{F}[x]. Since the gcd computed in F[x]F[x] generates the same ideal after tensoring to F‾[x]\overline F[x] (Euclidean algorithm is field-operation-only, hence unaffected by field extension), gcd⁡(f,f′)≠1\gcd(f,f')\ne1 already in F[x]F[x]. Conversely, suppose all ei=1e_i = 1, i.e. ff has no repeated root. Fix any root αi0\alpha_{i_0}. In the derivative sum above, every term with j≠i0j \ne i_0 still contains the factor (x−αi0)ei0=(x−αi0)(x - \alpha_{i_0})^{e_{i_0}} = (x-\alpha_{i_0}) (since ei0=1e_{i_0}=1 means the untouched product ∏i≠j(x−αi)ei\prod_{i \ne j} (x-\alpha_i)^{e_i} for j≠i0j \ne i_0 still includes the i=i0i=i_0 factor once), while the j=i0j = i_0 term is c⋅1⋅∏i≠i0(x−αi)c\cdot 1 \cdot \prod_{i \ne i_0}(x - \alpha_i), which does not vanish at x=αi0x = \alpha_{i_0} because all other roots are distinct from αi0\alpha_{i_0}. Hence f′(αi0)≠0f'(\alpha_{i_0}) \ne 0 for every root αi0\alpha_{i_0} of ff, so ff and f′f' share no common root in F‾\overline F, so (since F‾\overline F is algebraically closed and a nonunit common factor of f,f′f,f' in F‾[x]\overline F[x] would force a common root) gcd⁡(f,f′)=1\gcd(f,f')=1 in F‾[x]\overline F[x], hence also in F[x]F[x]. This proves the equivalence: ff has a repeated root in F‾\overline{F}   ⟺  \iff gcd⁡(f,f′)≠1\gcd(f,f') \neq 1 in F[x]F[x]. Statement (1) is the negation of this.

Step 2: the irreducible case. Assume now ff is irreducible in F[x]F[x]. If f′=0f' = 0, then trivially gcd⁡(f,f′)=f≠1\gcd(f,f') = f \ne 1 (as ff is nonconstant), so by Step 1, ff has a repeated root. Conversely suppose f′≠0f' \neq 0. Since deg⁡f′<deg⁡f\deg f' < \deg f (as deg⁡f≥1\deg f \geq 1 and differentiation strictly lowers degree for a nonzero derivative) and f′≠0f' \ne 0, and since ff is irreducible, the only divisors of ff in F[x]F[x] are units and unit multiples of ff itself. Because deg⁡f′<deg⁡f\deg f' < \deg f, the polynomial ff itself cannot divide f′f' unless f′=0f'=0; since f′≠0f' \ne 0, ff does not divide f′f', so ff is not an associate of any common divisor built from ff, i.e. gcd⁡(f,f′)\gcd(f,f'), being a divisor of ff, must be 11 (the only proper divisors of an irreducible ff are units). Hence gcd⁡(f,f′)=1\gcd(f,f')=1, and by Step 1, ff has no repeated root. This proves (2).

Step 3: characteristic 00. Suppose char⁡F=0\operatorname{char} F = 0 and ff is irreducible of degree n≥1n \geq 1, say f(x)=anxn+⋯+a0f(x) = a_n x^n + \cdots + a_0 with an≠0a_n \neq 0. Then f′(x)=nanxn−1+⋯+a1f'(x) = n a_n x^{n-1} + \cdots + a_1. Since char⁡F=0\operatorname{char} F = 0, the element n∈Fn \in F (meaning n⋅1Fn \cdot 1_F) is nonzero for every positive integer nn, so nan≠0n a_n \ne 0, so f′≠0f' \ne 0 and has degree exactly n−1≥0n - 1 \geq 0. If n=1n = 1 then ff is linear, trivially separable (one root, multiplicity one) — consistent since a linear polynomial always has a single simple root. If n≥2n \geq 2, f′≠0f' \ne 0, so by Step 2, ff has no repeated root, i.e. ff is separable. This proves (3).

Step 4: characteristic pp. Suppose char⁡F=p>0\operatorname{char} F = p > 0 and f(x)=∑i=0naixif(x) = \sum_{i=0}^n a_i x^i is irreducible with f′=0f' = 0. Then f′(x)=∑i=1niaixi−1=0f'(x) = \sum_{i=1}^n i a_i x^{i-1} = 0 means iai=0i a_i = 0 in FF for every i=1,…,ni = 1, \dots, n. If ai≠0a_i \ne 0 for some ii not divisible by pp, then i≠0i \ne 0 in FF (since char⁡F=p\operatorname{char} F = p and p∤ip \nmid i implies i⋅1F≠0i \cdot 1_F \ne 0; this is because the additive order of 1F1_F is exactly pp, so i⋅1F=0i\cdot 1_F = 0 iff p∣ip \mid i), forcing iai≠0i a_i \ne 0, a contradiction. So ai=0a_i = 0 whenever p∤ip \nmid i. Hence every nonzero coefficient of ff occurs at an exponent divisible by pp, i.e.

f(x)=∑kapkxpk=g(xp),g(y):=∑kapkyk∈F[x].f(x) = \sum_{k} a_{pk} x^{pk} = g(x^p), \qquad g(y) := \sum_k a_{pk} y^k \in F[x].

Conversely, if f(x)=g(xp)f(x) = g(x^p) for some g∈F[x]g \in F[x], then writing g(y)=∑kbkykg(y) = \sum_k b_k y^k,

f(x)=∑kbkxpk,f′(x)=∑kbk⋅(pk)xpk−1=0,f(x) = \sum_k b_k x^{pk}, \qquad f'(x) = \sum_k b_k \cdot (pk) x^{pk-1} = 0,

since pk≡0(modp)pk \equiv 0 \pmod p makes each coefficient bk(pk)=0b_k(pk) = 0 in FF (as char⁡F=p\operatorname{char} F = p). By Step 2 (using irreducibility of ff), f′=0f' = 0 is equivalent to ff having a repeated root, so ff is inseparable. This proves (4). ■\blacksquare

∎
CorollaryPerfect fields

If FF has characteristic 00, or if FF is a finite field, then every irreducible polynomial in F[x]F[x] is separable, i.e. every algebraic extension of FF is a separable extension.

Proof

Characteristic 00 is Theorem part (3). For FF finite of characteristic pp with ∣F∣=pm|F| = p^m: the Frobenius map φ:F→F\varphi: F \to F, φ(a)=ap\varphi(a) = a^p, is a ring homomorphism (it is additive because (a+b)p=ap+bp(a+b)^p = a^p+b^p in characteristic pp, by the binomial theorem and the fact that (pi)≡0(modp)\binom{p}{i} \equiv 0 \pmod p for 0<i<p0<i<p; it is multiplicative trivially), and it is injective since FF is a field (its kernel is an ideal, hence 00 or FF, and φ(1)=1≠0\varphi(1)=1\ne0 so the kernel is not all of FF). An injective map from a finite set to itself is surjective, so φ\varphi is an automorphism of FF; in particular every element of FF is a pp-th power of some (unique) element of FF. Now suppose, for contradiction, f∈F[x]f \in F[x] is irreducible and inseparable. By Theorem part (4), f(x)=g(xp)=∑kbkxpkf(x) = g(x^p) = \sum_k b_k x^{pk} with bk∈Fb_k \in F. Since φ\varphi is surjective, write bk=ckpb_k = c_k^p for some ck∈Fc_k \in F. Then, using that the Frobenius is a ring homomorphism on FF so that cpxp=(cx)pc^p x^{p} = (cx)^p termwise, and that in characteristic pp we have (∑kckxk)p=∑kckpxpk(\sum_k c_k x^k)^p = \sum_k c_k^p x^{pk} (again by the Frobenius property applied to the ring F[x]F[x], since the "freshman's dream" identity holds in any commutative ring of characteristic pp):

f(x)=∑kckpxpk=(∑kckxk)p=:h(x)p.f(x) = \sum_k c_k^p x^{pk} = \left(\sum_k c_k x^k\right)^p =: h(x)^p.

So f=hpf = h^p with deg⁡h=deg⁡f/p≥1\deg h = \deg f / p \geq 1 (as ff nonconstant), contradicting the irreducibility of ff (an irreducible polynomial cannot be a nontrivial power of a lower-degree polynomial, i.e. cannot factor as h⋅h⋯hh \cdot h \cdots h with p≥2p \geq 2 factors each of positive degree). This contradiction shows no such inseparable irreducible ff exists. ■\blacksquare

∎
TheoremSeparable extensions and embeddings

Let K/FK/F be a finite extension of fields, and fix an algebraic closure F‾\overline{F} of FF with K⊆F‾K \subseteq \overline{F}. Let Emb⁡F(K,F‾)\operatorname{Emb}_F(K, \overline{F}) denote the set of field embeddings σ:K→F‾\sigma : K \to \overline{F} fixing FF pointwise. Then

∣Emb⁡F(K,F‾)∣≤[K:F],|\operatorname{Emb}_F(K,\overline{F})| \leq [K:F],

with equality if and only if K/FK/F is a separable extension.

Proof

We argue by induction on [K:F][K:F], using the tower law and the standard extension theorem for embeddings (that a simple extension F(α)/FF(\alpha)/F admits exactly as many FF-embeddings into F‾\overline F as the number of distinct roots of mαm_\alpha in F‾\overline F, a fact itself proved from the isomorphism F(α)≅F[x]/(mα(x))F(\alpha) \cong F[x]/(m_\alpha(x)) and the universal property of quotient rings).

Base case. If [K:F]=1[K:F] = 1 then K=FK = F and there is exactly one embedding, the inclusion, so equality holds trivially and K/FK/F is (vacuously, or directly, since the only element is in FF itself with linear, hence separable, minimal polynomial) separable.

Simple extension step. Suppose K=F(α)K = F(\alpha) with minimal polynomial mα∈F[x]m_\alpha \in F[x] of degree d=[K:F]d = [K:F]. Any FF-embedding σ:K→F‾\sigma: K \to \overline{F} is determined by σ(α)\sigma(\alpha), which must be a root of mαm_\alpha in F‾\overline{F} (apply σ\sigma to the relation mα(α)=0m_\alpha(\alpha) = 0, using that σ\sigma fixes coefficients of mα∈F[x]m_\alpha \in F[x]). Conversely, for any root β∈F‾\beta \in \overline{F} of mαm_\alpha, the map F[x]/(mα(x))→F‾F[x]/(m_\alpha(x)) \to \overline{F} sending x↦βx \mapsto \beta is a well-defined FF-algebra homomorphism (since mα(β)=0m_\alpha(\beta)=0), and composing with the isomorphism F(α)≅F[x]/(mα(x))F(\alpha) \cong F[x]/(m_\alpha(x)) (which sends α↦x\alpha \mapsto x) gives an embedding σ\sigma with σ(α)=β\sigma(\alpha) = \beta. Distinct roots give distinct embeddings. Hence

∣Emb⁡F(F(α),F‾)∣=#{distinct roots of mα in F‾}≤deg⁡mα=d=[K:F],|\operatorname{Emb}_F(F(\alpha), \overline{F})| = \#\{\text{distinct roots of } m_\alpha \text{ in } \overline{F}\} \le \deg m_\alpha = d = [K:F],

with equality iff mαm_\alpha has dd distinct roots, i.e. iff mαm_\alpha is separable, i.e. iff α\alpha is separable over FF, i.e. (since K=F(α)K = F(\alpha), every element of KK is an FF-polynomial in α\alpha) iff K/FK/F is a separable extension — we verify this last equivalence fully in the general inductive step below, since a priori separability of K/FK/F requires every element, not just α\alpha, to be separable; this is subsumed by the induction.

Inductive step (general tower). Suppose [K:F]=n>1[K:F] = n > 1 and the result holds for all extensions of degree less than nn. Choose α∈K∖F\alpha \in K \setminus F and set E=F(α)E = F(\alpha), so 1<[E:F]=d≤n1 < [E:F] = d \le n and [K:E]=n/d<n[K:E] = n/d < n. By the tower law, [K:F]=[K:E][E:F][K:F] = [K:E][E:F]. Every FF-embedding τ:K→F‾\tau: K \to \overline{F} restricts to an FF-embedding τ∣E:E→F‾\tau|_E : E \to \overline F. By the simple extension step, there are at most dd possible restrictions σ1,…,σr\sigma_1,\dots,\sigma_r (r≤dr \le d) that occur as restrictions of some embedding of EE. For each such σi:E→F‾\sigma_i : E \to \overline F, the number of embeddings τ:K→F‾\tau : K \to \overline F extending σi\sigma_i equals the number of σi\sigma_i-embeddings of KK into F‾\overline F; this is bounded, by the same simple-extension counting argument applied iteratively to a chain of simple extensions from EE to KK (or directly by induction, viewing K/EK/E as a finite extension and F‾\overline F still algebraically closed so it receives σi\sigma_i and remains a valid target), by [K:E][K:E]. (This uses that F‾\overline F is algebraically closed, hence also serves as an algebraic closure for σi(E)\sigma_i(E), so the same embedding-extension machinery applies over the field σi(E)\sigma_i(E).) Therefore

∣Emb⁡F(K,F‾)∣=∑i=1r#{τ extending σi}≤r⋅[K:E]≤d⋅[K:E]=[K:F].|\operatorname{Emb}_F(K,\overline F)| = \sum_{i=1}^{r} \#\{\tau \text{ extending } \sigma_i\} \le r \cdot [K:E] \le d \cdot [K:E] = [K:F].

Equality analysis. Equality ∣Emb⁡F(K,F‾)∣=[K:F]|\operatorname{Emb}_F(K,\overline{F})| = [K:F] forces, tracing the inequalities above, both r=dr = d (equality in the simple-extension step for α\alpha, i.e. α\alpha separable over FF) and, for each ii, exactly [K:E][K:E] extensions of σi\sigma_i to KK (equality in the induction hypothesis applied to K/EK/E, transported via σi\sigma_i, i.e. K/EK/E is a separable extension in the sense that every element of KK has separable minimal polynomial over σi(E)\sigma_i(E), hence — since minimal polynomials and separability are intrinsic and σi\sigma_i is an FF-isomorphism onto its image fixing FF — every element of KK has separable minimal polynomial over EE). We must now show: [α\alpha separable over FF] and [every element of KK separable over E=F(α)E = F(\alpha)] together imply every element of KK is separable over FF, i.e. K/FK/F is separable, and conversely. (⇐\Leftarrow, K/FK/F separable ⇒\Rightarrow equality): if K/FK/F is separable, every β∈K\beta \in K, in particular α\alpha, is separable over FF; moreover for β∈K\beta \in K, the minimal polynomial of β\beta over E⊇FE \supseteq F divides the minimal polynomial of β\beta over FF (since the latter has β\beta as a root and lies in E[x]E[x], being separable it has no repeated roots in F‾\overline F; a divisor of a separable polynomial is separable, since repeated roots of the divisor would be repeated roots of the whole product). So every β∈K\beta \in K is separable over EE too, giving K/EK/E separable, and by the induction hypothesis (applicable to K/EK/E, a smaller-degree extension) equality holds for K/EK/E; combined with equality for E/FE/F (since α\alpha separable over FF gives r=dr = d), we get the full equality chain above, hence ∣Emb⁡F(K,F‾)∣=[K:F]|\operatorname{Emb}_F(K,\overline F)| = [K:F]. (⇒\Rightarrow, equality ⇒K/F\Rightarrow K/F separable): assume equality holds. As shown, this forces α\alpha separable over FF and, via the induction hypothesis (now used in its "equality ⇒\Rightarrow separable" direction for the smaller extension K/EK/E), every element of KK separable over EE. Take any β∈K\beta \in K; we must show β\beta is separable over FF. Let mβF,mβEm^F_\beta, m^E_\beta denote its minimal polynomials over F,EF, E respectively; mβE∣mβFm^E_\beta \mid m^F_\beta in E[x]E[x] (minimal polynomial over the bigger field divides minimal polynomial over the smaller one, since mβF(β)=0m^F_\beta(\beta)=0 and mβF∈F[x]⊆E[x]m^F_\beta \in F[x]\subseteq E[x]). We know mβEm^E_\beta is separable (has distinct roots in F‾\overline F). We want mβFm_\beta^F separable too. This requires one more ingredient, proved next as a standalone lemma, since it is not immediate that a polynomial dividing a separable one forces the whole polynomial (mβFm^F_\beta, not just mβEm^E_\beta) to be separable — we need the converse direction via degrees, using the multiplicativity established by the Tower/Embedding count itself applied to the tower F⊆F(β)⊆KF \subseteq F(\beta) \subseteq K, which by definition of α,E,β\alpha, E, \beta falls under the same inductive machinery: apply the entire argument above with β\beta playing the role of α\alpha. Concretely, redo the simple/tower decomposition using the chain F⊆F(β)⊆KF \subseteq F(\beta) \subseteq K instead of F⊆F(α)⊆KF \subseteq F(\alpha) \subseteq K: since [K:F][K:F] is fixed and equality ∣Emb⁡F(K,F‾)∣=[K:F]|\operatorname{Emb}_F(K,\overline F)|=[K:F] is a property of K/FK/F not depending on the chosen intermediate field, the same equality-forces-separability argument (with β\beta in place of α\alpha) shows β\beta is separable over FF. Since β∈K\beta \in K was arbitrary, K/FK/F is separable. ■\blacksquare

∎

Worked examples

Example 1 (a genuinely irrational, asymmetric separable computation). Let F=QF = \mathbb{Q} and let α=23+2\alpha = \sqrt[3]{2} + \sqrt{2}. We check separability directly via the derivative criterion applied to a specific auxiliary polynomial, rather than to the full minimal polynomial of α\alpha (which has degree 66 and is unwieldy) — instead we illustrate the criterion on f(x)=x3−3x−2=(x−2)(x+1)2∈Q[x]f(x) = x^3 - 3x - 2 = (x-2)(x+1)^2 \in \mathbb{Q}[x], an intentionally reducible, partially repeated-root example to see the gcd test catch a repeated root explicitly. Compute f′(x)=3x2−3f'(x) = 3x^2 - 3. Run the Euclidean algorithm:

f(x)=(x−2)(x+1)2(check: (x+1)2=x2+2x+1, (x−2)(x2+2x+1)=x3−3x−2✓).f(x) = (x-2)(x+1)^2 \quad\text{(check: } (x+1)^2 = x^2+2x+1,\ (x-2)(x^2+2x+1) = x^3 - 3x - 2 \checkmark).

Divide ff by f′f': f(x)=13x⋅f′(x)+(−2x−2)f(x) = \tfrac{1}{3}x \cdot f'(x) + (-2x - 2), since 13x(3x2−3)=x3−x\tfrac13 x(3x^2-3) = x^3 - x, and f(x)−(x3−x)=−2x−2f(x) - (x^3-x) = -2x-2. So the remainder is r1(x)=−2x−2=−2(x+1)r_1(x) = -2x-2 = -2(x+1). Next divide f′(x)=3x2−3f'(x) = 3x^2-3 by r1(x)=−2(x+1)r_1(x) = -2(x+1): 3x2−3=3(x−1)(x+1)3x^2 - 3 = 3(x-1)(x+1), so f′(x)f'(x) is exactly divisible by (x+1)(x+1) with quotient 3(x−1)3(x-1) and remainder 00. So gcd⁡(f,f′)=(x+1)\gcd(f,f') = (x+1) (up to a unit), which is nontrivial. By the derivative criterion, ff has a repeated root — consistent with f=(x−2)(x+1)2f = (x-2)(x+1)^2, where −1-1 is a double root. This confirms the test works exactly as claimed, mechanically, without ever needing to factor ff by inspection first — the Euclidean algorithm alone reveals the repeated root's location, x=−1x=-1, as the root of gcd⁡(f,f′)\gcd(f,f'). Now for the genuinely separable, irrational case: the minimal polynomial of α=23+2\alpha=\sqrt[3]2+\sqrt2 over Q\mathbb Q has degree 66 (since [Q(23,2):Q]=6[\mathbb Q(\sqrt[3]2,\sqrt2):\mathbb Q]=6 by the tower law, using [Q(23):Q]=3[\mathbb Q(\sqrt[3]2):\mathbb Q]=3, [Q(2):Q]=2[\mathbb Q(\sqrt2):\mathbb Q]=2, coprime degrees forcing the compositum degree to be the product, and α\alpha generates the compositum since Q(α)⊆Q(23,2)\mathbb Q(\alpha)\subseteq \mathbb Q(\sqrt[3]2,\sqrt2) combined with a degree count shows equality). Since char⁡Q=0\operatorname{char}\mathbb Q = 0, Theorem part (3) guarantees automatically, with no further computation, that this degree-6 minimal polynomial is separable. This is the payoff of the characteristic-00 corollary: separability is free, and the delicate gcd computation above is never actually needed in characteristic 00 — it is only characteristic pp that requires vigilance.

Example 2 (an explicit inseparable irreducible polynomial, characteristic pp). Let p=3p = 3 and let F=F3(t)F = \mathbb{F}_3(t), the field of rational functions in one variable tt over F3=Z/3Z\mathbb{F}_3 = \mathbb{Z}/3\mathbb{Z}. Consider

f(x)=x3−t∈F[x].f(x) = x^3 - t \in F[x].

Irreducibility: ff is irreducible in F[x]F[x]. Indeed, F=F3(t)F = \mathbb F_3(t) is the fraction field of the UFD F3[t]\mathbb F_3[t], and by Gauss's lemma it suffices to check x3−tx^3 - t is irreducible in F3[t][x]\mathbb F_3[t][x]; since tt is a prime element of F3[t]\mathbb F_3[t] and t∤1t \nmid 1 (leading coefficient) while t2∤tt^2 \nmid t, Eisenstein's criterion at the prime tt applies directly, confirming irreducibility.

Separability check via the derivative criterion: f′(x)=3x2=0f'(x) = 3x^2 = 0 in F[x]F[x], since char⁡F=3\operatorname{char} F = 3. By Theorem part (2), since ff is irreducible and f′=0f' = 0, ff has a repeated root in F‾\overline F — in fact by Theorem part (4), ff is inseparable, matching the visible form f(x)=g(x3)f(x) = g(x^3) with g(y)=y−tg(y) = y - t.

Direct confirmation by exhibiting the root explicitly: let θ=t1/3∈F‾\theta = t^{1/3} \in \overline{F} be a root of ff (i.e. θ3=t\theta^3 = t). Then in F‾[x]\overline{F}[x], using the characteristic-33 Frobenius identity (x−θ)3=x3−θ3(x-\theta)^3 = x^3 - \theta^3 (freshman's dream, valid since all "cross terms" (31),(32)\binom{3}{1}, \binom{3}{2} are divisible by 33),

f(x)=x3−t=x3−θ3=(x−θ)3.f(x) = x^3 - t = x^3 - \theta^3 = (x - \theta)^3.

So ff has the single root θ\theta, with multiplicity exactly 33 — a dramatic, maximal failure of separability: an irreducible cubic collapses to one triple root. Consequently F(θ)/FF(\theta)/F is a degree-33 extension in which ∣Emb⁡F(F(θ),F‾)∣=1<3=[F(θ):F]|\operatorname{Emb}_F(F(\theta), \overline F)| = 1 < 3 = [F(\theta):F] (there is only one possible image for θ\theta, namely θ\theta itself, since ff has only one root in F‾\overline F), exactly matching the strict inequality case of the Embeddings Theorem.

Counterexampledropping "irreducible" from the derivative test's converse

The equivalence "f′=0  ⟺  ff' = 0 \iff f has a repeated root" in Theorem part (2) requires ff to be irreducible. Without irreducibility, f′=0f' = 0 still forces a repeated root (that direction only used ff nonconstant, via the general Step 1/Step 4 argument), but the converse — a repeated root forcing f′=0f' = 0 — genuinely fails for reducible polynomials, and more importantly the clean dichotomy of part (2) breaks: a reducible ff can have a repeated root while f′≠0f' \ne 0, since the derivative criterion in general only detects repeated roots via gcd⁡(f,f′)≠1\gcd(f,f')\ne 1, not via f′=0f'=0 outright. Concrete example: let F=QF = \mathbb{Q} (characteristic 00, so no Frobenius pathology needed to illustrate this particular point) and take the reducible polynomial

f(x)=(x−1)2(x−3)=x3−5x2+7x−3.f(x) = (x-1)^2(x-3) = x^3 - 5x^2 + 7x - 3.

Then f′(x)=3x2−10x+7f'(x) = 3x^2 - 10x + 7. Evaluate at x=1x=1: f′(1)=3−10+7=0f'(1) = 3 - 10 + 7 = 0. So f′f' does vanish at the repeated root x=1x=1 (as guaranteed by Step 1's general argument, which needs no irreducibility), but f′f' itself is not the zero polynomial: e.g. f′(0)=7≠0f'(0) = 7 \ne 0. So here f′≠0f' \ne 0 yet ff has a repeated root — precisely the scenario that Theorem part (2) says is impossible for irreducible ff, but which is entirely possible once irreducibility is dropped. The correct universal test in the reducible case is still gcd⁡(f,f′)≠1\gcd(f,f') \ne 1 (here gcd⁡(f,f′)=x−1\gcd(f,f') = x - 1, nontrivial, correctly flagging the repeated root), but the simpler shortcut "f′=0f'=0" is no longer equivalent to "repeated root" — it becomes only a sufficient, not necessary, condition once we allow reducible polynomials. This shows why irreducibility is load-bearing in part (2): it is exactly what rules out a partial repeated root sitting alongside other, simple roots, which is what makes f′f' nonzero despite the collision.

Scope, edge cases & common traps

SituationWhat holdsWhat fails / must be checked
char⁡F=0\operatorname{char} F = 0Every irreducible poly. is separable automaticallyNothing extra to check — inseparability cannot occur
FF finiteEvery irreducible poly. is separable (Frobenius is onto)Same conclusion as char. 00, different proof mechanism
FF infinite, char⁡F=p\operatorname{char} F = pDerivative criterion still fully validInseparable irreducibles can exist, e.g. xp−tx^p - t over Fp(t)\mathbb F_p(t)
ff reduciblegcd⁡(f,f′)≠1\gcd(f,f')\ne1 still detects repeated rootsf′=0f' = 0 is only sufficient, not necessary, for a repeated root (see counterexample)
K/FK/F algebraic but infinite-degreeDefinition (every element separable) still appliesThe embedding-counting Theorem as stated is for finite [K:F][K:F]; infinite case needs care with cardinalities

The mathematician behind it

Origins · 1 minErnst Steinitz and how this was discovered
ES
Ernst Steinitz
1871–1928 · Foundations of abstract field theory, early 20th century Germany

Steinitz's 1910 paper Algebraische Theorie der Körper ("Algebraic Theory of Fields") is widely regarded as the founding document of abstract field theory, the first systematic axiomatic treatment of fields divorced from any specific number-theoretic or geometric context. He introduced the classification of fields by characteristic, established the existence and essential uniqueness of algebraic closures, and — central to this article — isolated the phenomenon of inseparability as a structural feature of characteristic pp fields rather than a computational curiosity. Steinitz's insistence on stating hypotheses with complete precision (what today would be called "hypothesis discipline") set the tone for the rigorous, axiomatic style that later became the hallmark of the entire Noether–Artin–van der Waerden school of abstract algebra.

History
- 1871 — Ernst Steinitz born in Laurahütte, German Empire (present-day Poland).
- 1910 — Steinitz publishes Algebraische Theorie der Körper, systematically defining perfect fields and separability, and proving every algebraic extension of a perfect field is separable.
- 1920s — Emil Artin and collaborators extend the framework toward what becomes modern Galois theory, embedding separability as a hypothesis in the fundamental theorem.
- 1928 — Steinitz dies in Kiel, Germany.
- 1930s — B.L. van der Waerden's textbook Moderne Algebra, based on lectures of Artin and Noether, cements separability as a standard prerequisite concept ahead of the Galois correspondence.
Why is separability defined via the minimal polynomial rather than some other polynomial?

The minimal polynomial is the canonical, irreducible representative of α\alpha's algebraic relation to FF: it is the unique monic irreducible polynomial in F[x]F[x] having α\alpha as a root, and every other polynomial vanishing at α\alpha is a multiple of it. Using an arbitrary polynomial vanishing at α\alpha would not give a well-defined notion, since a reducible multiple of the minimal polynomial could introduce extra roots unrelated to α\alpha's intrinsic algebraic behavior, potentially with repeated roots caused by an unrelated factor. Restricting to the minimal polynomial isolates exactly the multiplicity structure of α\alpha itself, making the definition intrinsic to α\alpha and FF alone, independent of any auxiliary polynomial one might choose.

Does the derivative criterion require computing roots in an algebraic closure?

No — that is precisely its power. The gcd gcd⁡(f,f′)\gcd(f,f') is computed by the ordinary Euclidean algorithm entirely inside F[x]F[x], using only the field operations of FF, with no need to construct F‾\overline F or locate any roots. The theorem guarantees that this purely syntactic computation correctly detects the semantic fact of repeated roots in F‾\overline F. This is what makes separability checkable in practice, and it is the basis of squarefree factorization algorithms in computer algebra, which strip repeated irreducible factors from a polynomial using exactly this gcd computation.

Can an inseparable extension still have the "right" degree in some other sense?

Yes, in a precise way: if K=F(α)K = F(\alpha) with α\alpha inseparable over FF, and char⁡F=p\operatorname{char}F=p, one can show (using the derivative criterion iterated) that the minimal polynomial of α\alpha has the form h(xpe)h(x^{p^e}) for a separable irreducible hh and some maximal e≥1e \geq 1, and αpe\alpha^{p^e} is then separable over FF. The "correct" separable degree is [F(αpe):F][F(\alpha^{p^e}):F], while the full degree [F(α):F][F(\alpha):F] is that separable degree times pep^e, called the inseparable degree. This decomposition (separable degree times inseparable degree) is developed fully in the theory of general algebraic extensions but is beyond the scope of this article's finite, root-counting treatment.

Is every extension of a field of characteristic 0 automatically separable, with no exceptions?

Yes, with no exceptions, and this follows directly from Theorem part (3): in characteristic 00, differentiating any nonconstant irreducible polynomial produces a nonzero polynomial of strictly smaller degree, because no positive integer coefficient n⋅ann \cdot a_n can vanish (integers are never zero in a characteristic-00 field). By Theorem part (2), this forces the polynomial to have no repeated roots. Since this argument makes no further assumption on FF beyond characteristic 00, it applies uniformly to Q\mathbb{Q}, R\mathbb{R}, C\mathbb{C}, and every extension field of characteristic 00, algebraic or not.

Why does the finite field case also give separability, if the mechanism (Frobenius surjectivity) looks so different from characteristic 0?

Characteristic 00 and finite fields are both instances of the broader notion of a perfect field: a field where either the characteristic is 00, or the characteristic is p>0p>0 and the Frobenius endomorphism is surjective. The proof strategies differ (vanishing of coefficients versus surjectivity of Frobenius) because the underlying obstruction differs, but both mechanisms independently rule out the existence of an irreducible polynomial of the form g(xp)g(x^p), which Theorem part (4) identifies as the sole source of inseparability. Fields that are neither characteristic 00 nor perfect in characteristic pp — such as Fp(t)\mathbb{F}_p(t) — are exactly where genuine inseparable extensions first appear.

How does separability connect to the Primitive Element Theorem?

The Primitive Element Theorem states that a finite separable extension K/FK/F is generated by a single element, K=F(γ)K = F(\gamma) for some γ∈K\gamma \in K. Separability is essential to its proof (typically via a pigeonhole/interpolation argument over an infinite field, or trivially via cyclicity of the multiplicative group for finite fields), because the argument needs enough distinct images under various embeddings to construct γ\gamma, and the Embeddings Theorem proved here is exactly what supplies the correct count of distinct embeddings, [K:F][K:F], that the construction relies upon. Without separability, this count can be strictly smaller than [K:F][K:F], and the standard construction of a primitive element can fail.

Is "separable" a property of a polynomial, an element, or an extension — and can these conflict?

All three, in a strictly compatible hierarchy: a polynomial is separable if it has no repeated roots in an algebraic closure; an element α∈K\alpha \in K is separable over FF if its minimal polynomial is separable; and an extension K/FK/F is separable if every element of KK is separable over FF. These never conflict, because the extension-level definition is built directly from the element-level one, which in turn is built directly from the polynomial-level one — there is no independent, competing definition at any level that could contradict another.

Before this

Where this leads