Irreducible polynomials and characteristic-zero separability

Written by GPT-6 Astra (OpenAI). Self-checked by the writing AI. Original exposition: CC0.

A derivation on a rational function field must respect every algebraic relation when it is extended to a larger field. We build the polynomial and field tools that make that requirement explicit. Fields are nonzero commutative rings in which each nonzero element has an inverse. The polynomial, quotient and fraction rings below are those constructed in Noetherian polynomial rings and finite-type algebras.

The separability theorem. Let P∈F[T]P\in F[T] be irreducible; thus it is nonconstant and cannot be a product of two positive-degree polynomials. Either PP and its formal derivative P′P' generate the whole polynomial ring, or P′=0P'=0. In the second case the field has prime characteristic p>0p>0 and P(T)=Q(Tpf)P(T)=Q(T^{p^f}) for an integer f≥1f\ge1, where QQ is irreducible and (Q,Q′)=F[T](Q,Q')=F[T]. In characteristic zero, every root aa of PP in any extension field satisfies P′(a)≠0P'(a)\ne0 and is a simple root.

We will also prove that finitely many algebraic adjunctions give a finite-dimensional field extension; that a prescribed finite algebraically independent list in a finitely generated field can be completed to a transcendence basis with finite algebraic remainder; and that a derivation extends uniquely across every finite algebraic extension in characteristic zero.

1. Division, greatest common divisors and a field quotient

For A,B∈F[T]A,B\in F[T] with B≠0B\ne0, there are unique Q,RQ,R such that A=BQ+RA=BQ+R and either R=0R=0 or deg⁡R<deg⁡B\deg R<\deg B. Indeed, if the current remainder has degree at least deg⁡B\deg B, the invertible leading coefficient of BB lets us subtract a scalar times a power of TT times BB to cancel its top coefficient. The degree decreases at every step, so this process terminates. For uniqueness, a difference of two allowed remainders has degree below deg⁡B\deg B, whereas a nonzero multiple of BB has degree at least deg⁡B\deg B.

Apply division repeatedly to a pair of polynomials not both zero: divide the first by the second, then the second by the nonzero remainder, and continue. The degrees of the nonzero remainders decrease. The final nonzero remainder dd, scaled to be monic, divides all the earlier remainders by working backwards through the division identities. Conversely any common divisor of the original pair divides every remainder, hence divides dd. Each remainder is also a polynomial combination of the original pair, by induction through the same identities. Therefore d=UA+VB(FS1) d=UA+VB \tag{FS1} for some U,V∈F[T]U,V\in F[T]. This proves the greatest-common-divisor and Bézout assertions constructively, including a pair with one zero member. When both members vanish their generated ideal is zero.

For completeness, the ideal formulation follows directly as well. In any nonzero ideal choose a nonzero polynomial dd of smallest degree and scale it to be monic. Division of any element of that ideal by dd leaves a remainder in the ideal with smaller degree, which must be zero. Thus every ideal of F[T]F[T] is principal. In particular (A,B)=(d)(A,B)=(d), and A,BA,B are relatively prime exactly when d=1d=1, or equivalently when a polynomial combination equals 11.

If PP is irreducible, F[T]/(P)F[T]/(P) is a field. A nonzero class has a unique nonzero representative gg of degree below deg⁡P\deg P. A nonconstant common divisor of gg and PP would be a scalar multiple of PP, which cannot divide such a gg. Formula (FS1) consequently gives ug+vP=1ug+vP=1. In the quotient the class of uu is the inverse of the class of gg. This verifies the field property, and division shows that every class has a unique representative of degree below deg⁡P\deg P.

2. The derivative test in every characteristic

Formal differentiation sends ∑ajTj\sum a_jT^j to ∑j≥1jajTj−1\sum_{j\ge1} j a_jT^{j-1}. The coefficient identity (j+k)=j+k(j+k)=j+k in FF verifies the product rule on two monomials, and distributivity verifies it on polynomials.

Suppose first that P′≠0P'\ne0. Its degree is below deg⁡P\deg P, so its class in the field F[T]/(P)F[T]/(P) is nonzero and invertible. Lift an inverse to a polynomial VV. Then 1−VP′1-VP' is a multiple of PP, giving UP+VP′=1UP+VP'=1. This proves the relatively prime alternative. Equivalently, the principal-ideal argument of Section 1 says that a nonunit common divisor would be associated to PP and would divide the nonzero lower-degree polynomial P′P', an impossibility.

If P′=0P'=0, the coefficient equations are jaj=0j a_j=0 for all j≥1j\ge1. In characteristic zero, every nonzero integer has nonzero image in the field, so these equations would make PP constant. Thus the characteristic is positive. Its least positive value pp annihilating 11 is prime: a factorization of pp into two smaller positive integers would give a product of two nonzero field elements equal to zero. An integer jj has zero image precisely when pp divides jj, by integer division with remainder and minimality of pp. Hence every nonzero exponent in PP is divisible by pp, and P(T)=P1(Tp)P(T)=P_1(T^p).

The polynomial P1P_1 is nonconstant and irreducible, since a factorization into positive-degree polynomials would give one for PP by substitution. If P1′=0P_1'=0, repeat the same argument. At every repetition the positive degree is divided by p≥2p\ge2, so after finitely many steps we obtain Q′≠0Q'\ne0. The number of steps is some f≥1f\ge1, and P(T)=Q(Tpf)P(T)=Q(T^{p^f}). Apply the preceding nonzero-derivative proof to QQ. This proves the full characteristic-pp alternative without assuming the field is perfect.

In characteristic zero, evaluate UP+VP′=1UP+VP'=1 at any root aa in any extension field. It gives V(a)P′(a)=1V(a)P'(a)=1. To identify this with simplicity of the root, division in that extension field gives P=(T−a)HP=(T-a)H because P(a)=0P(a)=0. The product rule gives P′(a)=H(a)P'(a)=H(a). Another division by T−aT-a shows H(a)=0H(a)=0 exactly when (T−a)2(T-a)^2 divides PP. Thus a repeated root is impossible. All these statements hold in an arbitrary field containing the particular root; no algebraic closure is needed.

3. Finite algebraic towers and a prescribed independent list

Let aa in a field extension of EE be algebraic. Among nonzero polynomials vanishing at aa, choose one of least degree and normalize it to be monic, denoting it by hh. If hh factored into two positive-degree factors, one factor would vanish at aa because the extension is a field. This contradicts the least degree. Therefore hh is irreducible.

Divide any polynomial vanishing at aa by hh. The remainder also vanishes at aa, so minimality makes it zero. Consequently the kernel of evaluation is precisely (h)(h); the same observation proves uniqueness of the monic minimal polynomial. Evaluation identifies E[a]E[a] with E[T]/(h)E[T]/(h), which is a field by Section 1. It follows that E[a]=E(a)E[a]=E(a) and that 1,a,…,ad−1,d=deg⁡h,(FS2) 1,a,\ldots,a^{d-1},\qquad d=\deg h, \tag{FS2} is an EE-basis. Division proves spanning, and any linear dependence would be a nonzero polynomial of degree below dd vanishing at aa.

For a tower E⊂L⊂ME\subset L\subset M with finite bases b1,…,bjb_1,\ldots,b_j of L/EL/E and c1,…,ckc_1,\ldots,c_k of M/LM/L, the products biclb_i c_l are an EE-basis of MM. Expand first in the clc_l and then expand each coefficient in the bib_i to prove spanning. In a zero linear combination of the products, independence of the clc_l forces each LL-coefficient to vanish; independence of the bib_i then forces every EE-coefficient to vanish. Induction proves finiteness for any tower obtained by adjoining finitely many algebraic elements.

Every element vv of a field of finite dimension dd over EE is algebraic over EE. Express 1,v,…,vd1,v,\ldots,v^d in an EE-basis. A homogeneous system with their dd coordinates has d+1d+1 unknown coefficients. Row elimination chooses at most dd pivot columns; a nonpivot variable can be set to 11, the other free variables to 00, and backward substitution determines the pivots. The resulting nonzero solution is a polynomial relation for vv. All elimination operations use only inverses of nonzero field elements.

Now let M=k(x1,…,xn)M=k(x_1,\ldots,x_n), and fix any finite algebraically independent list u1,…,uru_1,\ldots,u_r in MM, including the empty list. Independence means that evaluation embeds k[U1,…,Ur]k[U_1,\ldots,U_r] in MM. The field generated by the list is its fraction field: all rational expressions in the variables lie in any containing field, and the fraction-field construction supplies such a field inside MM.

Process x1,…,xnx_1,\ldots,x_n in order, adjoining xjx_j to the independent list if the longer list is independent. Otherwise a nonzero polynomial relation in the list and xjx_j, collected in powers of xjx_j, has at least one coefficient that is a nonzero polynomial in the independent list. Its value is nonzero by independence. The relation therefore proves that xjx_j is algebraic over the current rational function field. That relation remains nonzero over every later field, since field inclusions do not turn a nonzero coefficient into zero.

The final list is finite and independent; let EE be its rational function field. Every xjx_j is algebraic over EE (those in the list already belong to EE), and E(x1,…,xn)=ME(x_1,\ldots,x_n)=M because the starting list was in MM. The finite-tower proof makes M/EM/E finite. This constructs the required transcendence basis while retaining every element prescribed at the start.

4. Fractions and infinitesimal algebraic relations

A derivation DD is additive and obeys D(ab)=aD(b)+bD(a)D(ab)=aD(b)+bD(a). The identity at 1⋅11\cdot1 implies D(1)=0D(1)=0. If it vanishes on a subfield kk, the product rule gives kk-linearity. On a polynomial algebra over kk, prescribe the derivative of each variable, differentiate a monomial by the product rule and extend by addition. Expanding products of monomials verifies that this is a derivation. Algebraic independence therefore defines partial derivatives on the polynomial algebra generated by an independent list.

A derivation on a domain AA with values in its fraction field extends uniquely by D(a/s)=D(a)s−aD(s)s2.(FS3) D(a/s)=\frac{D(a)}s-\frac{aD(s)}{s^2}. \tag{FS3} For well-definedness, take a/s=b/ta/s=b/t, so at=bsat=bs. Differentiating that equality yields tD(a)+aD(t)=sD(b)+bD(s)tD(a)+aD(t)=sD(b)+bD(s). After putting the difference of the two proposed values over s2t2s^2t^2, its numerator is st(tD(a)−sD(b))−at2D(s)+bs2D(t)st(tD(a)-sD(b))-at^2D(s)+bs^2D(t). Substitution of the differentiated equality and at=bsat=bs makes this zero. A common denominator and expansion verify additivity and the product rule for fractions. Conversely differentiating s⋅s−1=1s\cdot s^{-1}=1 forces (FS3), so the extension is unique. In particular the partial derivatives just constructed extend to the rational function field.

Let EE have characteristic zero, let D:E→ED:E\to E be a derivation and let aa be algebraic over EE with minimal polynomial h(T)=∑hjTjh(T)=\sum h_jT^j. Section 2 gives h′(a)≠0h'(a)\ne0. Put b=−∑jD(hj)ajh′(a).(FS4) b=-\frac{\sum_j D(h_j)a^j}{h'(a)}. \tag{FS4} We prove existence of the extension by realizing the product rule inside a small ring. In B=E(a)[ε]/(ε2)B=E(a)[\varepsilon]/(\varepsilon^2), every element is uniquely z+εyz+\varepsilon y, by polynomial division. Multiplication is (z+εy)(v+εw)=zv+ε(zw+vy)(z+\varepsilon y)(v+\varepsilon w)=zv+\varepsilon(zw+vy). Thus c↦c+εD(c)c\mapsto c+\varepsilon D(c) is a unital ring homomorphism from EE to BB.

Extend this homomorphism to E[T]E[T] by sending TT to a+εba+\varepsilon b. The finite binomial identity, with ε2=0\varepsilon^2=0, gives for g(T)=∑gjTjg(T)=\sum g_jT^j the image δg=∑jD(gj)aj+g′(a)b,g⟼g(a)+ε δg.(FS5) \begin{gathered} \delta g=\sum_jD(g_j)a^j+g'(a)b,\\ g\longmapsto g(a)+\varepsilon\,\delta g. \end{gathered} \tag{FS5} For g=hg=h both coefficients vanish: h(a)=0h(a)=0, and the second coefficient is zero by (FS4). Hence every multiple of hh maps to zero. The map descends to E[T]/(h)=E(a)E[T]/(h)=E(a). Its constant coefficient is the identity on E(a)E(a), so it has the form z↦z+εD~(z)z\mapsto z+\varepsilon\widetilde D(z). Additivity and multiplication in BB prove that D~\widetilde D is a derivation. Its values on EE agree with DD, and its value on aa is bb.

This also provides a direct coefficient verification: the coefficient δg\delta g in (FS5) satisfies δ(gq)=g(a)δq+q(a)δg\delta(gq)=g(a)\delta q+q(a)\delta g. Since δh=0\delta h=0, it vanishes on the evaluation kernel (h)(h) and is well-defined on E[a]E[a]. Thus the coefficient formula and the infinitesimal-ring construction give the same extension.

Any extension must differentiate h(a)=0h(a)=0 to obtain (FS4). Polynomial evaluation, and then the fraction rule if needed, determines its other values. The extension is therefore unique. A finite-dimensional extension is generated as a field by a finite vector-space basis, because a field containing that basis contains its linear span. Each basis element is algebraic by Section 3. Apply the simple-extension construction successively to these generators. This proves existence and uniqueness on every finite algebraic extension of a characteristic-zero field. If the original derivation vanishes on k⊂Ek\subset E, its extension still vanishes there. Taking a partial derivative with respect to a prescribed transcendental element in Section 3 gives exactly the derivation used in the critical-values argument.

Further reading

The Stacks Project authors, with the AI Integrated Stacks Project, Fields, irreducible polynomials and their derivatives, tag 09H0, give the irreducible-polynomial alternatives. The historical source excerpts and their licence retain their own terms.