A field containing one root of a polynomial may omit the others. This matters when we try to send that root to another one: the proposed image must actually lie in the target field. Splitting fields supply all roots at once, while separability ensures that distinct roots have not collapsed into repeated ones.
Motivation: roots that cannot be reached
Consider and the field . The polynomial has three complex roots, and one might hope that the field has symmetries exchanging them. It has none. Any field automorphism of fixes pointwise, so . Thus is a root of lying in , and the only real root is itself. Hence , and since is generated by over , the automorphism is the identity. The polynomial has three roots, but the field cannot move any of them, because the other two roots are not in the field.
This shows the first obstacle to field symmetry: the target of a proposed root-to-root map may not contain the required root. Normality and splitting fields are the tools that remove this obstacle. A second obstacle is quieter. Even in a field containing every root, a polynomial of degree may have fewer than distinct roots, and then there are fewer candidate images than the degree suggests. Separability removes this second obstacle. The unit below develops both notions, proves the facts needed to work with them, and ends by showing exactly how each obstacle limits the group of symmetries of a field extension.
Splitting fields
Definition 3.1.1
(Splitting and normality)
A splitting field of a nonconstant is an extension generated over by all the roots of , in which factors into linear polynomials. A finite algebraic extension is normal if every irreducible polynomial in with a root in splits completely over .
A splitting field exists: adjoin a root of an irreducible factor using a polynomial quotient, factor again, and repeat until every irreducible factor is linear; adjoining a root reduces the degree of a remaining nonsplit factor. The construction uses finitely many adjunctions. Any two splitting fields of the same polynomial are -isomorphic; this uniqueness theorem and the extension-of-embeddings results behind it are proved in the embeddings unit.
Proposition 3.1.2
(A finite normal extension is a splitting field)
A finite extension is normal if and only if it is the splitting field over of a polynomial in .
Proof
(Minimal polynomials of a basis and extension of embeddings)
If is normal, choose a finite -basis of . Each basis element is algebraic and its minimal polynomial splits in ; their product splits there and its roots generate . Conversely, suppose is a splitting field and an irreducible has a root . For any other root in an algebraic closure of , which exists by the unit on algebraic closures, the isomorphism sending to extends to an -embedding of into that closure. This is the extension-of-embeddings lemma. The image of is generated by the images of the roots of the defining polynomial, which permute its roots and hence all lie in . Finite-dimensionality makes the image equal to , so .
It is worth reading the definition and the proposition together. Normality is stated as a condition on every irreducible polynomial that meets , which looks like infinitely many conditions. The proposition reduces it to one condition: must be generated by the complete root set of a single polynomial. Conversely, to show that an extension is not normal, it suffices to exhibit one irreducible polynomial with a root in and a root outside , exactly as we did for above.
The size of a splitting field
Adjoining all the roots of a polynomial of degree might seem to require a very large extension. The next result gives a uniform bound.
Proposition 3.1.3
(Degree bound for splitting fields)
Let have degree and let be a splitting field of over . Then is finite and .
Proof
(Adjoin one root and induct)
We argue by induction on . If then is linear, its root lies in , and , so . Suppose and that the statement holds for all fields and all polynomials of degree . Because factors into linear polynomials in , it has a root . The minimal polynomial of over divides , so . Over we can write with of degree . The polynomial splits into linear factors in because does, and the roots of are together with the roots of . Since is generated over by the roots of , it is generated over by the roots of . Hence is a splitting field of over , and the inductive hypothesis gives . By the tower law,
The bound is attained in some cases, as the examples in the cubic and quintic continuation show, and is far from sharp in others. The proposition is useful mainly because it guarantees that all root-adjoining constructions stay finite-dimensional, which is the setting in which normality was defined.
Separable polynomials and the derivative test
Definition 3.1.4
(Separable polynomial and extension)
An irreducible polynomial over is separable if it has distinct roots in a splitting field. An algebraic element is separable if its minimal polynomial is separable; an algebraic extension is separable if every element is separable. A general polynomial is square-free if it has no repeated irreducible factor; here we call it separable when it has no repeated root in a splitting field. Authors use the term separable polynomial with different conventions.
Proposition 3.1.5
(Derivative test)
A nonconstant has no repeated root in a splitting field if and only if in . In characteristic zero every irreducible polynomial is separable.
Proof
(Common roots and the gcd)
In a splitting field write for a root . The product rule gives , so ; thus occurs at least twice precisely when it is a common root of and its formal derivative . If , Bézout's identity gives with , and evaluating at a common root would give . So a common root forces a nonconstant gcd over . Conversely, a nonconstant gcd divides , so it has a root in the splitting field of , and that root is common to and . In characteristic zero the derivative of a nonconstant irreducible polynomial is nonzero and has smaller degree, so its gcd with the irreducible polynomial is .
The practical value of the derivative test is that the gcd is computed inside by the Euclidean algorithm, with no need to find the roots or construct the splitting field. Whether is does not change when the base field is enlarged, which is why the criterion can be checked over even though the definition refers to a splitting field.
Repeated roots in positive characteristic
The last sentence of the derivative test is restricted to characteristic zero, and the restriction is essential. In characteristic a nonconstant polynomial can have derivative zero, for instance . The next result says exactly when this happens and what it means for irreducible polynomials.
Proposition 3.1.6
(Inseparable irreducibles in characteristic )
Let have characteristic and let be nonconstant. Then if and only if for some . If moreover is irreducible, then is inseparable if and only if .
Proof
(Coefficients killed by the derivative)
Write , so . The coefficient vanishes in exactly when or . Hence if and only if whenever , that is, if and only if is a polynomial in .
Now let be irreducible. By the derivative test, has a repeated root if and only if . If then is nonconstant, so is inseparable. If , then , and the gcd divides . The only monic divisors of the irreducible polynomial are and its monic associate, and the latter has degree larger than and so cannot divide . Therefore the gcd is and is separable.
Thus inseparability of an irreducible polynomial is a very rigid phenomenon: it occurs only for polynomials in . Over many familiar fields this never happens, as the next result shows.
Proposition 3.1.7
(Irreducibles over finite fields are separable)
Every irreducible polynomial over a finite field is separable.
Proof
(Frobenius makes a th power)
Let be a finite field of characteristic . The map is a ring homomorphism of , because in characteristic . It is injective since is a field, so it is bijective because is finite. Suppose an irreducible were inseparable. By the preceding proposition . Choose with . The -th power map is additive on the commutative ring of characteristic , so
Since , the polynomial is nonconstant, and is a product of two nonconstant polynomials (here ). This contradicts irreducibility.
Proof X-ray
Where finiteness is used
The argument has three steps, and only one of them needs the field to be finite. The preceding proposition turns inseparability of an irreducible into the shape ; this holds over every field of characteristic . The identity is the freshman's binomial rule in characteristic , again valid over every such field. Finiteness enters only to write each coefficient of as a -th power , through surjectivity of the injective map on a finite set. Over this step fails, since is not a -th power, and indeed is irreducible and inseparable there.
The argument used only that every element of is a -th power. Fields with this property are called perfect, and the field fails it because is not a -th power; this is precisely why is possible there.
Theorem 3.1.8
(Perfect fields)
A field is perfect if it has characteristic zero, or if it has characteristic and every element is a -th power. A field is perfect if and only if every irreducible polynomial in is separable. Consequently every algebraic extension of a perfect field is separable.
Proof
(One direction repeats the finite-field proof; the other uses )
In characteristic zero both conditions hold, by the derivative test. Let . If every element of is a -th power, the proof for finite fields applies word for word, since it used finiteness only to take -th roots of coefficients.
Conversely, suppose some is not a -th power, and let . In a splitting field choose with ; then . Let be a monic irreducible factor of in . By unique factorization in the splitting field, for some , and the coefficient of in is , which lies in . If , then is invertible in , so and is a -th power, contrary to the choice of . Hence and is irreducible. Its only root is , of multiplicity , so is an inseparable irreducible polynomial.
Finally, if is perfect and is algebraic, the minimal polynomial over of each element of is irreducible, hence separable.
Thus , , , every finite field and every algebraically closed field are perfect; an algebraically closed field contains a root of each . The field is not perfect, and the proof recovers as the standard inseparable irreducible polynomial over it.
Examples
Worked example 3.1.9
(Normality and separability are different)
Let . The irreducible polynomial has two nonreal roots, so is not normal. It is separable because the base has characteristic zero. In contrast, take and let . The polynomial is irreducible by Eisenstein at in and Gauss's lemma over its fraction field . In a splitting field . Thus is normal but inseparable. This example shows why characteristic zero must not be silently assumed.
Common mistake
(Square-free over the base is not always separable)
Over an imperfect field, an irreducible polynomial can have repeated roots after extension. The polynomial above has no repeated irreducible factor in but has one root of multiplicity in its splitting field.
Normality does not pass through towers
A natural hope is that normality is transitive: if and are normal then should be. The following example shows that it is not, which is one reason why normality has to be checked against the base field rather than an intermediate one.
Worked example 3.1.10
(Normal over a normal extension, but not normal over the base)
Let , let and . Since , we have . The polynomial is irreducible over by Eisenstein at , so and , hence .
First, is normal, being the splitting field of with roots . Second, is normal: is a root of , which has degree and is therefore the minimal polynomial of over , and its roots both lie in . By the exercise on quadratic extensions below, any extension of degree is normal, which agrees with this computation.
However, is not normal. The irreducible polynomial has the root , but its roots include . So an irreducible polynomial with a root in fails to split in .
The cubic and quartic splitting-field constructions appear in the companion note on cubic and quintic splitting fields.
Normal closures
The tower example shows that normality is not transitive. Two weaker statements do hold: normality of passes to the top step of a tower, though not to the bottom step, and every finite extension sits inside a smallest normal one.
Proposition 3.1.11
(Normality passes to the top of a tower)
If is finite normal and , then is normal. The extension need not be normal.
Proof
(The same polynomial over a larger base)
By the first proposition, is the splitting field over of some . Then , it splits in , and its roots generate over , hence also over the larger field . So is a splitting field of over , and the first proposition shows that is normal. For the second claim, let be a primitive cube root of unity. The field lies in , the splitting field of , which is normal over ; but is not normal, as the motivation showed.
Theorem 3.1.12
(Normality through embeddings)
Let be finite and let be an algebraically closed field containing . Then is normal if and only if every -embedding satisfies .
Proof
(Roots stay inside, and every root is reached)
Suppose is normal, and let be an -embedding. For , the element is a root of the minimal polynomial of over . That polynomial is irreducible with the root , so it splits in , and by unique factorization in all its roots in already lie in . Hence . Since is an injective -linear map of the finite-dimensional space into itself, .
Conversely, let be irreducible with a root , and let be any root of in . The one-root lemma gives an -embedding sending to , and the finite-tower theorem of the embeddings unit extends it to an -embedding . By hypothesis . The polynomial splits in the algebraically closed field , and all its roots there lie in , so it splits in . Hence is normal.
Definition 3.1.13
(Normal closure)
Let be finite and let be an algebraically closed field containing . A normal closure of in is a field with such that is finite and normal, while no field with is normal over .
Theorem 3.1.14
(Existence and uniqueness of the normal closure)
Let be finite over inside an algebraically closed field , let be the minimal polynomial of over , and let be the subfield of generated over by all roots in of . Then (a) every field with and normal contains ; (b) is a normal closure of in , and it is the only one; (c) is the compositum of the fields , where runs over the -embeddings . In particular does not depend on the chosen generators.
Proof
(A splitting field that every normal overfield must contain)
The field is a splitting field of over , so is finite and normal by the first proposition, and it contains every , hence .
For (a), each is irreducible over with the root , so it splits in ; by unique factorization in every root of in lies in . Thus contains all roots of in , and so .
For (b), a field with cannot be normal over , because by (a) it would contain . So is a normal closure. If is another one, (a) gives ; since contains and is normal over , the minimality in the definition of forces .
For (c), each , and is a root of , so . Conversely, let be a root of some . The one-root lemma and the finite-tower theorem give an -embedding with , so . Hence the compositum of the fields contains every root of , and so contains . By (a) and (b), is the smallest subfield of that contains and is normal over , a description that mentions no generators.
Worked example 3.1.15
(Normal closures of two radical fields)
Let inside , and let be a primitive cube root of unity. The generator has minimal polynomial , with roots . The normal closure is therefore , of degree six. The three -embeddings of have images , and , and the compositum of any two of them is already , since it contains the quotient or of the two generators.
For , the minimal polynomial has roots and , so the normal closure is . Its degree is computed in the practice section below. In the tower example both steps and are normal, yet the normal closure of over is strictly larger than .
The four combinations
The examples so far exhibit extensions with each combination of the two properties.
The extension is both normal and separable. The extension is separable but not normal. The extension with is normal but not separable. It remains to show that an extension can be neither.
Worked example 3.1.16
(An extension that is neither normal nor separable)
Let be an indeterminate over , put , and let . The polynomial is Eisenstein at the prime , hence irreducible over by Gauss's lemma, and is a root of it, so has minimal polynomial . Its derivative is in characteristic , so is inseparable and is not a separable extension.
In we have, in characteristic ,
using . We claim has no root in . A root would give a root of in . Write with coprime . Then . Reducing modulo shows , so is a nonzero constant since ; similarly forces to be constant. Then is the only possibility in , and , a contradiction. Hence the irreducible polynomial has a root in but does not split in , and is not normal.
Consequences for field symmetries
We can now state precisely how the two obstacles of the introduction restrict automorphisms.
Proposition 3.1.17
(Automorphisms move roots within their orbit)
Let be a finite normal extension and let be irreducible with roots . Then there is an -automorphism of with . If in addition is separable of degree , then has exactly roots in , the group permutes them transitively, and divides .
Proof
(Extend the root map, then count)
By the first proposition, is a splitting field of some polynomial over . As in its proof, the isomorphism sending to extends, by the finite-tower extension theorem of the embeddings unit, to an -embedding of into an algebraic closure, and because permutes the roots of the defining polynomial. An injective -linear map from a finite-dimensional -space to itself is bijective, so is an automorphism of with .
Now assume is separable of degree . Because is normal and has a root in , it splits in , and separability gives distinct roots there. Every -automorphism maps roots of to roots of , so acts on this set of roots, and the first part shows the action is transitive. By the orbit-stabilizer theorem , so divides the order of the group.
Both hypotheses are visible in this statement. Without normality, the element need not lie in , and the conclusion is meaningless. This is the situation of in the motivation. Without separability, there are fewer than roots to which can be sent.
Worked example 3.1.18
(Counting symmetries with and without separability)
For with , the polynomial is irreducible and separable of degree , and is irreducible and separable of degree with roots . The proposition gives and , so divides the order. On the other hand an automorphism is determined by and , giving at most possibilities. Therefore .
Now take the inseparable normal extension with over . The minimal polynomial has degree but only the single root , since . Any -automorphism sends to a root of , hence to , and so is the identity. Thus . The extension is normal, so no root is unreachable, but the missing distinct roots still cost us all the symmetry.
These two computations preview the central theme of Galois theory. The count of automorphisms equals the degree of the extension exactly when both obstacles are absent, namely for normal separable extensions. The section on separable degree below proves this equality.
Separable degree
The derivative test examines one polynomial at a time. Counting embeddings measures separability for a whole extension at once, and the count behaves well in towers.
Theorem 3.1.19
(The separable degree)
Let be finite, let be an embedding into an algebraically closed field, and let be the number of embeddings extending . Then (a) depends neither on nor on ; it is called the separable degree . (b) If , then is the number of distinct roots of the minimal polynomial of in a splitting field over . (c) For every intermediate field , . (d) .
Proof
(Settle simple extensions, then sort embeddings by restriction)
First let , with minimal polynomial . An extension of is determined by the image of , which must be a root of in , and by the one-root lemma each such root gives exactly one extension. So is the number of distinct roots of in . Let be the field generated over by these roots; it is a splitting field of over , because splits in . If is a splitting field of over , the uniqueness theorem of the embeddings unit gives an isomorphism extending . It maps the roots of in injectively to roots of , and transporting the factorization of over shows that every root of in is reached. Hence equals the number of distinct roots of in , which proves (a) and (b) for simple extensions, together with .
Now let . Every embedding extending restricts to an embedding extending , and conversely every extension of such a to extends . Sorting the extensions of by their restriction to gives
the sum running over the extensions of to .
We prove (a) by induction on , for all fields at once. If , then . Otherwise choose and put . If , the simple case applies. If not, , so by induction is a number that does not depend on or . The display gives , and both factors are independent of and . This proves (a). With (a) known for every finite extension, the display for an arbitrary intermediate field reads , which is (c). For (d), write as a tower of simple extensions; each step satisfies (d) by the simple case, and (c) together with the tower law gives (d) for .
The bound (d) is the embedding-count bound of the embeddings unit, and (b) with (c) shows that it is attained exactly when every simple step has a minimal polynomial with distinct roots. For example, if over , then while .
Theorem 3.1.20
(Separability through the separable degree)
For a finite extension the following are equivalent: (i) is separable; (ii) for some elements separable over ; (iii) .
Proof
(Generators are enough)
(i) implies (ii): take a finite -basis as generators. (ii) implies (iii): put . The minimal polynomial of over satisfies by the derivative test, so it has no repeated root in any field where it splits. The minimal polynomial of over divides , so it also has distinct roots in a splitting field, and (b) gives . Multiplying over with (c) and the tower law gives (iii). (iii) implies (i): let . By (c) and (d),
Equality holds by (iii), and both factors on the left are positive and bounded by the corresponding factors on the right, so . By (b), the minimal polynomial of has as many distinct roots as its degree; that is, is separable.
Proposition 3.1.21
(Consequences for towers and generators)
Let be finite. (a) For an intermediate field , is separable if and only if both and are separable. (b) If is the splitting field over of a polynomial whose irreducible factors are all separable, then is separable. (c) The elements of that are separable over form a subfield of . (d) If is separable, then its normal closure in an algebraically closed field is normal and separable over .
Proof
(Compare separable degrees with degrees)
For (a), (c) of the separable-degree theorem gives , and (d) bounds each factor by the corresponding degree. So if and only if and , and the criterion translates this into the statement. For (b), each root of the polynomial is a root of one of its irreducible factors, whose monic associate is the root's minimal polynomial; so is generated by separable elements, and criterion (ii) applies. For (c), if are separable over , then is separable by criterion (ii), so , and, for , are separable. For (d), write . The normal closure is the splitting field of the product of the minimal polynomials of the , which are irreducible and separable, so (b) applies.
Part (d) is the step used in the unit on simple extensions when a finite separable extension is placed inside a finite Galois extension.
Proposition 3.1.22
(Automorphisms and the separable degree)
For a finite extension , . The first inequality is an equality exactly when is normal. Hence if and only if is normal and separable.
Proof
(Automorphisms are the embeddings that land back in E)
Let be an algebraic closure of , which exists by the unit on algebraic closures. Composing an -automorphism of with the inclusion gives an -embedding , and distinct automorphisms give distinct embeddings. The embeddings obtained in this way are exactly those with . Hence is the number of -embeddings with , which is at most the number of all -embeddings, with equality exactly when every -embedding satisfies . By the embedding characterization, that happens exactly when is normal. The last sentence combines this with the separability criterion (iii).
This proves the equality anticipated after the symmetry computations. The fixed-field unit reaches the same count for normal separable extensions by a different route, through Artin's theorem.
Worked example 3.1.23
(Separable degree in a tower of characteristic two)
Return to with , from the example that is neither normal nor separable, and put . The element is a root of , which is Eisenstein at and so irreducible over ; thus and . The derivative of is , and with a nonzero constant in , so . Hence is separable and . Over , the element has minimal polynomial , of degree and with one distinct root, so . Therefore , while .
This agrees with a direct count: the minimal polynomial of over is , and has three distinct roots, since its derivative shares no root with it. The separable part of sits at the bottom, in , and the inseparability is concentrated in the top step .
Practice and solutions
Quick check
(Find the splitting field of over and explain why it is separable.)
Solution · Answer
Its roots are , both in , which they generate. Its derivative is coprime to , so its roots are distinct.
Quick check
(Is separable over ?)
Solution · Answer
No. In characteristic two, and the derivative is zero. It has one root with multiplicity four.
Quick check
(Show that is irreducible over , and find the degree of its splitting field over .)
Solution · Answer
Substitute : . All non-leading coefficients are divisible by and the constant term is not divisible by , so Eisenstein's criterion at shows this is irreducible, and therefore is irreducible over (a factorization of would give one of the shifted polynomial). Let . The element has order , since and ; thus its powers are distinct. They are roots of because for odd . These are all the roots by degree. Each is a power of , so the splitting field is , whose degree equals the degree of the minimal polynomial of . The bound is far from attained.
Quick check
(Let . Show that is irreducible over and inseparable, and determine its roots and their multiplicities in a splitting field.)
Solution · Answer
The polynomial is Eisenstein at the prime (the constant term is divisible by but not , other lower coefficients are zero), so it is irreducible over by Gauss's lemma. Its derivative is in characteristic , so by the proposition on inseparable irreducibles it is inseparable. In a splitting field let satisfy . In characteristic we have , so
Since and the characteristic is not , . The polynomial has exactly two distinct roots, and , each of multiplicity , and its splitting field is of degree .
Quick check
(Show that has no repeated root over , for any prime .)
Solution · Answer
Let . Then in characteristic , a nonzero constant. Therefore , and by the derivative test has no repeated root in a splitting field.
Quick check
(Find the degree over of the splitting field of , and say whether the bound is attained.)
Solution · Answer
Let . The roots of are , so the splitting field is generated by them, which is (it contains , and conversely and produce every root). The polynomial is irreducible by Eisenstein at , so . Since and , while is a root of , we get . Thus the splitting field has degree . The bound is , which is not attained.
Quick check
(Prove that every extension of degree is normal, in any characteristic.)
Solution · Answer
Let and let be irreducible with a root . Then is a scalar multiple of the minimal polynomial of over , whose degree divides . If the degree is , then is linear and splits. If it is , then and with up to the scalar . The polynomial has the root , so it equals with , hence . Thus splits over , and is normal.
Quick check
(Find the normal closure of over inside , and its degree.)
Solution · Answer
By the normal-closure theorem with the single generator , the normal closure is generated by the roots of , so it is the splitting field . The exercise on above shows that its degree is . It is twice the degree of .
Quick check
(Let be a perfect field of characteristic and a finite extension. Show that is perfect.)
Solution · Answer
The Frobenius map is an injective ring homomorphism , so is a subfield of . It contains . Let be an -basis of . Every element of has the form , so the span over . They are independent: if , write , using that is perfect; then , so and all . Hence , and the tower law gives . Thus .
Quick check
(Let and with . Compute directly and through the tower .)
Solution · Answer
The earlier exercise shows that is the minimal polynomial of , with two distinct roots, so by part (b). In the tower, is a root of , which is irreducible by Eisenstein at and separable because its derivative is coprime to it; so . Then , and is a root of in characteristic three, which is therefore its minimal polynomial over , with one distinct root. So and . The extension is not separable, although its lower step is.