Separable Elements and Split Minimal Polynomials
Introduction
Two conditions on the minimal polynomial of an element decide whether the element can be diagonalised inside its algebra. The polynomial must be split, meaning that all its roots lie in the base field, and it must be separable, meaning that its roots are distinct. The first condition produces eigenvalues in the field; the second produces idempotents that separate them. Together they are the hypothesis of the eigenvalue decomposition of The Eigenvalue Decomposition of an Element, and separately they explain the two ways in which that decomposition can fail.
The article assumes the algebra generated by an element and the Chinese remainder decomposition of The Algebra Generated by an Element and the Chinese Remainder Decomposition, the polynomial algebra of Polynomial Algebras, and the field theory of Splitting Fields and Algebraic Closure. The base is a field $k$ and the algebra $A$ is finite-dimensional, unital and associative. The relation to the separable algebras of Separable Algebras is treated in §Separable Elements and Separable Algebras: the two notions share a word but not a definition, and the difference matters over an imperfect field. The nilpotent part that survives when separability fails is Nilpotent Elements and the Jordan–Chevalley Decomposition.
Separable Polynomials
The definition
Let $f \in k[x]$ be a monic polynomial of degree $n \ge 1$. A splitting field of $f$ is a field extension $K/k$ in which $f$ factors into linear factors, $f(x) = \prod_{i=1}^{n}(x-\alpha_i)$ with $\alpha_i \in K$. The polynomial is separable if the roots $\alpha_1, \dots, \alpha_n$ are pairwise distinct in every splitting field, equivalently in one splitting field. The definition does not depend on the choice of the splitting field, because the multiset of roots is transported by any isomorphism of splitting fields.
For an irreducible polynomial, separability has a simple criterion in terms of the derivative: $f$ is separable if and only if $f' \neq 0$. The point is that an irreducible polynomial cannot have a repeated root unless it is a polynomial in $x^p$ in characteristic $p$, and then its derivative vanishes. This is the content of the next subsection.
The derivative criterion
Theorem. Let $f \in k[x]$ be monic of degree $n \ge 1$. The following are equivalent:
- $f$ is separable;
- $\gcd(f, f') = 1$ in $k[x]$;
- the discriminant $\Delta(f)$ is nonzero.
Proof. Over a splitting field $K$, write $f = \prod_i (x-\alpha_i)$. A repeated root $\alpha$ means that $(x-\alpha)^2$ divides $f$, so $f(\alpha) = 0$ and $f'(\alpha) = 0$; the linear factor $x-\alpha$ then divides both $f$ and $f'$ over $K$, and hence over $k$ because the gcd is defined over $k$ and is unchanged by field extension. Conversely, if a nonconstant $g$ divides both $f$ and $f'$, then a root of $g$ in a splitting field is a repeated root of $f$. This proves the equivalence of 1 and 2. The discriminant is
$$ \boxed{\;\Delta(f) = \prod_{1 \le i < j \le n}(\alpha_j - \alpha_i)^2 = (-1)^{n(n-1)/2}\operatorname{Res}(f,f')\;}, $$
a symmetric polynomial in the roots, hence an element of $k$; it vanishes exactly when two roots coincide. This proves the equivalence of 1 and 3.
The sign in the resultant formula is the standard one; the corpus uses the discriminant only through its vanishing, so the sign never matters. For a quadratic $x^2 + bx + c$ the discriminant is $b^2 - 4c$; for a cubic $x^3 + px + q$ it is $-4p^3 - 27q^2$.
Characteristic $p$
In characteristic zero every irreducible polynomial is separable, because a nonconstant polynomial has a nonzero derivative and an irreducible polynomial cannot share a nonconstant factor with its derivative unless that factor is the polynomial itself, which would force $f' = 0$. In characteristic $p$ the derivative can vanish identically: $f(x) = x^p - t$ has $f' = px^{p-1} = 0$.
- Over $k = \mathbb{F}_p(t)$, the polynomial $x^p - t$ is irreducible, because $t$ is not a $p$-th power in $\mathbb{F}_p(t)$, and it is inseparable: in a splitting field $f(x) = (x-\alpha)^p$ with $\alpha^p = t$. The extension $k(\alpha)/k$ is purely inseparable of degree $p$.
- The polynomial $x^p - x$ has derivative $-1$, so it is separable. Its roots are exactly the elements of the prime field $\mathbb{F}_p$, and over $\mathbb{F}_p$ it splits as $\prod_{\alpha \in \mathbb{F}_p}(x-\alpha)$.
- Over $\mathbb{F}_2$, the polynomial $x^2 - 1$ has derivative $2x = 0$ and is inseparable; indeed $x^2 - 1 = (x-1)^2$. The polynomial $x^2 - x = x(x-1)$ has derivative $1$ and is separable and split.
- Over $\mathbb{F}_2$, the polynomial $x^2 + x + 1$ has derivative $1$, so it is separable, but it has no root in $\mathbb{F}_2$ and is not split. The field $\mathbb{F}_4$ is its splitting field.
A field is perfect if every irreducible polynomial over it is separable. Fields of characteristic zero, finite fields and algebraically closed fields are perfect. Over a perfect field, every algebraic element is separable, and the distinction between separable and inseparable does not arise. The imperfection of a field is therefore a characteristic-$p$ phenomenon.
The discriminant as a determinant
For a monic $f$ of degree $n$ with roots $\alpha_1, \dots, \alpha_n$, the Vandermonde matrix $V$ with entries $V_{i\alpha} = \alpha^{i}$ for $0 \le i \le n-1$ satisfies
$$ \det V = \prod_{1 \le i < j \le n}(\alpha_j - \alpha_i), \qquad \det(VV^{\mathsf T}) = \Delta(f). $$
Consequently $\Delta(f) \neq 0$ exactly when the roots are distinct, which is the separability of $f$. This determinant form is used in §Separable Elements to identify the discriminant with the determinant of the trace form.
Separable Elements
The definition
Let $A$ be a finite-dimensional unital associative $k$-algebra and let $a \in A$, with minimal polynomial $m_a$ as in The Algebra Generated by an Element and the Chinese Remainder Decomposition. The element $a$ is separable if $m_a$ is separable, and split if $m_a$ splits over $k$. The element is split separable if both conditions hold; equivalently
$$ m_a(x) = \prod_{i=1}^{r}(x-\lambda_i), \qquad \lambda_i \in k \text{ pairwise distinct}. $$
The number $r$ of distinct roots is then $\deg m_a$, and the Chinese remainder decomposition of $k[a]$ has $r$ linear factors. The eigenvalues of $a$ are the $\lambda_i$, and they all lie in $k$.
The definition depends on the base field. The rotation matrix of The Algebra Generated by an Element and the Chinese Remainder Decomposition, §Examples, is separable over $\mathbb{R}$ (its minimal polynomial is $x^2+1$, of nonzero discriminant) but not split; over $\mathbb{C}$ it becomes split separable. The element itself does not change; the field does.
The trace form criterion
Let $L : k[a] \to \operatorname{End}_k(k[a])$ be the regular representation, $L_x(y) = xy$, and let
$$ \tau(x,y) = \operatorname{Tr}(L_{xy}) $$
be the trace form of the algebra $k[a]$. In the power basis $1, a, \dots, a^{d-1}$ with $d = \deg m_a$, the matrix of $\tau$ is
$$ T_{ij} = \operatorname{Tr}(L_{a^{i+j}}) = \operatorname{Tr}(a^{i+j}), \qquad 0 \le i,j \le d-1, $$
where the trace is that of the multiplication operator on $k[a]$.
Theorem. The element $a$ is separable if and only if the trace form $\tau$ on $k[a]$ is nondegenerate. In the power basis, $\det T = \Delta(m_a)$.
Proof. Over a splitting field, let $\alpha_1, \dots, \alpha_d$ be the roots of $m_a$ counted with multiplicity. The multiplication by $a$ on $k[a]$ has the companion matrix of $m_a$ as its matrix, and its eigenvalues are the $\alpha_i$. Hence $\operatorname{Tr}(a^k) = \sum_i \alpha_i^k$, and
$$ \boxed{\;T_{ij} = \sum_{\ell=1}^{d} \alpha_\ell^{i+j} = (V V^{\mathsf T})_{ij}\;}, $$
where $V_{i\ell} = \alpha_\ell^i$ is the Vandermonde matrix. Thus $\det T = (\det V)^2 = \Delta(m_a)$. The form is nondegenerate exactly when $\det T \neq 0$, which is exactly the separability of $m_a$.
The theorem is a criterion for the element, not for the ambient algebra: the trace form is taken on $k[a]$ with its own regular representation. The trace form of $A$ itself may be degenerate even when $a$ is separable, for instance when $A = M_n(k)$ and the characteristic divides $n$.
Examples
- The split quadratic. For $m = x^2 - 1$, the basis $1, a$ gives $T = \begin{pmatrix} 2 & 0 \\ 0 & 2 \end{pmatrix}$, of determinant $4 = \Delta(x^2-1)$. The element is separable and split over any field of characteristic not $2$.
- The repeated quadratic. For $m = x^2$, the basis $1, a$ gives $T = \begin{pmatrix} 2 & 0 \\ 0 & 0 \end{pmatrix}$, of determinant $0$. The element is not separable; the algebra $k[a]$ is the dual-number algebra.
- The cubic with three roots. For $m = x^3 - x$, the basis $1, a, a^2$ gives
$$ T = \begin{pmatrix} 3 & 0 & 2 \\ 0 & 2 & 0 \\ 2 & 0 & 2 \end{pmatrix}, $$
of determinant $4 = \Delta(x^3-x)$. The element is separable and split over a field of characteristic not $2$; over $\mathbb{F}_2$ the polynomial becomes $x(x-1)^2$ and is not separable. - The rotation. For $m = x^2+1$ over $\mathbb{R}$, the trace form is $T = \begin{pmatrix} 2 & 0 \\ 0 & -2 \end{pmatrix}$, of determinant $-4 = \Delta(x^2+1)$. The element is separable but not split. - A purely inseparable element. Over $k = \mathbb{F}_p(t)$ and $m = x^p - t$, the derivative vanishes, the discriminant vanishes and the trace form is degenerate. The element is not separable.
Separability and semisimplicity
Separability is stronger than the property of having a reduced subalgebra. The algebra $k[a]$ is reduced exactly when $m_a$ is squarefree, that is, when no irreducible factor of $m_a$ is repeated. A separable polynomial is squarefree, so a separable element generates a reduced algebra. The converse fails in characteristic $p$: the polynomial $x^p - t$ is irreducible over $\mathbb{F}_p(t)$, hence squarefree in the sense of having no repeated irreducible factor, so $k[a]$ is a field and is reduced, but it is inseparable: its single root over the algebraic closure has multiplicity $p$. In the terminology of Nilpotent Elements and the Jordan–Chevalley Decomposition, such an element is semisimple but not separable. Over a perfect field the two notions coincide.
The distinction is invisible in characteristic zero and over finite fields, and it is the reason the Jordan–Chevalley theorem is stated over a perfect field, or with a separability hypothesis, in the articles that use it.
Split Minimal Polynomials
The definition and the criterion
The element $a$ is split if $m_a$ splits over $k$,
$$ m_a(x) = \prod_{i=1}^{r}(x-\lambda_i)^{e_i}, \qquad \lambda_i \in k \text{ distinct}, \quad e_i \ge 1, $$
with $r$ the number of distinct eigenvalues and $\sum_i e_i = \deg m_a$. The split condition is exactly the condition that all eigenvalues of $a$ lie in $k$. It does not require the eigenvalues to be simple; the exponents $e_i$ may be greater than one.
The split condition is equivalent to the existence of a full set of eigenvalues in $k$, and it is also equivalent to the existence of the Chinese remainder decomposition of $k[a]$ into the linear factors $(x-\lambda_i)^{e_i}$. That decomposition always gives idempotents $\varepsilon_i$, the generalized eigenprojectors, and a decomposition
$$ a = \sum_i (\lambda_i \varepsilon_i + n_i), \qquad n_i = (a-\lambda_i)\varepsilon_i, \quad n_i^{e_i} = 0, $$
even when the eigenvalues are repeated. What fails without separability is not the existence of the idempotents but the vanishing of the nilpotent parts $n_i$. This is the precise sense in which split and separable play different roles.
The split squarefree case
If $m_a$ is split and squarefree, then $e_i = 1$ for every $i$, the nilpotent parts vanish, and
$$ \boxed{\;a = \sum_{i=1}^{r} \lambda_i \varepsilon_i, \qquad \varepsilon_i \varepsilon_j = \delta_{ij}\varepsilon_i, \quad \sum_i \varepsilon_i = 1\;} . $$
This is the eigenvalue decomposition. The idempotents are the Lagrange idempotents of The Lagrange–Sylvester Eigenprojector Formula, and the algebra $k[a]$ is isomorphic to $k^r$. The condition is also equivalent to the diagonalisability of the left multiplication $L_a$ and of the right multiplication $R_a$ on $A$, as proved in The Eigenvalue Decomposition of an Element.
The split non-separable case
If $m_a$ is split but not squarefree, the generalized eigenprojectors exist but the nilpotent parts do not vanish. The simplest case is $m_a = (x-\lambda)^d$, where there is a single idempotent $\varepsilon = 1$ and $a = \lambda + n$ with $n^d = 0$. The algebra $k[a]$ is local, and the element is a scalar plus a nilpotent. The general case is a sum of such local pieces, one for each eigenvalue, and it is the generalized eigenvalue decomposition of The Eigenvalue Decomposition of an Element and the Jordan–Chevalley decomposition of Nilpotent Elements and the Jordan–Chevalley Decomposition.
The Two Conditions Are Independent
The two conditions are logically independent. The following table gives one example of each combination; the examples are minimal polynomials, and the element is the class of $x$ in $k[x]/(m)$.
| Split over $k$ | Not split over $k$ | |
|---|---|---|
| Separable | $x(x-1)(x+1)$ over $\mathbb{Q}$; $x^2-x$ over $\mathbb{F}_2$ | $x^2+1$ over $\mathbb{R}$; $x^2-2$ over $\mathbb{Q}$; $x^2+x+1$ over $\mathbb{F}_2$ |
| Not separable | $(x-1)^2$ over any field; $x^2-1$ over $\mathbb{F}_2$ | $x^p-t$ over $\mathbb{F}_p(t)$; $x^2+t$ over $\mathbb{F}_2(t)$ |
The four entries are distinct phenomena.
- Separable and split. The element has distinct eigenvalues in $k$ and admits the eigenvalue decomposition. This is the regular case.
- Separable and not split. The element has distinct roots, but they lie in an extension. The algebra $k[a]$ is a product of separable field extensions; it has no eigenvalues in $k$ unless a factor is linear. The rotation matrix is the model case. Over a splitting field the decomposition exists, with idempotents in the scalar extension.
- Not separable and split. The element has eigenvalues in $k$, but some root is repeated. The generalized eigenprojectors exist, and the nilpotent parts are nonzero. The dual numbers and the Jordan blocks are the model cases.
- Not separable and not split. The element has neither distinct roots nor roots in $k$. The model case is a purely inseparable extension, and the failure is the most severe: no eigenvalues in $k$ and no separability over $k$.
The table also shows that neither condition implies the other, and that neither condition alone is enough for the eigenvalue decomposition. The decomposition needs both.
Separable Elements and Separable Algebras
The word "separable" is used for two different notions, and the difference is worth stating.
- A separable element is an element whose minimal polynomial is separable. This is a condition on one element, and it is the notion of this article.
- A separable algebra is a finite-dimensional $k$-algebra $A$ such that $A \otimes_k \bar k$ is semisimple; equivalently, $A$ is a product of matrix algebras over separable field extensions of $k$. This is a condition on the whole algebra, and it is the notion of Separable Algebras.
For a commutative finite-dimensional algebra, the two notions are compatible: $A$ is separable if and only if its trace form is nondegenerate, and this holds if and only if every element of $A$ is separable. The proof is the one indicated in §The trace form criterion together with the structure theorem for reduced commutative algebras: a separable commutative algebra is a product of separable fields, and every element of such a product has a separable minimal polynomial; conversely, if every element is separable then $A$ has no nilpotents and each of its residue fields is separable, because an inseparable residue field contains an element with inseparable minimal polynomial.
For a noncommutative algebra the implication fails in one direction. The matrix algebra $M_p(k)$ is separable over $k$ for every field $k$, because $M_p(k) \otimes_k \bar k \cong M_p(\bar k)$ is semisimple. Yet $M_p(k)$ contains an inseparable element when $k$ is imperfect: the companion matrix of $x^p - t$ over $k = \mathbb{F}_p(t)$ has minimal polynomial $x^p - t$, which is inseparable. So a separable algebra need not have all its elements separable.
This is why the eigenvalue decomposition is stated for an element, and not for an algebra: the algebra can be as separable as a matrix algebra and still contain an element that cannot be diagonalised. The separability of the element is the hypothesis that matters.
Summary
A monic polynomial is separable when its roots are distinct, equivalently when $\gcd(f,f') = 1$, equivalently when its discriminant is nonzero. Over a perfect field every irreducible polynomial is separable; in characteristic $p$ the polynomial $x^p - t$ over $\mathbb{F}_p(t)$ is the standard inseparable example. An element of a finite-dimensional algebra is separable when its minimal polynomial is, and split when its minimal polynomial splits over the base field.
For an element $a$, separability is equivalent to the nondegeneracy of the trace form $\tau(x,y) = \operatorname{Tr}(L_{xy})$ on $k[a]$, and the determinant of that form in the power basis is the discriminant of $m_a$. Separability is stronger than the reducedness of $k[a]$ in characteristic $p$: $x^p - t$ generates a field but is inseparable. Over a perfect field the two notions coincide.
The split condition gives eigenvalues in $k$ and the Chinese remainder decomposition into the linear factors $(x-\lambda_i)^{e_i}$, with generalized eigenprojectors and nilpotent parts. The separable condition kills the nilpotent parts. The two conditions are independent, and the eigenvalue decomposition needs both: split squarefree. Split non-separable gives the Jordan–Chevalley decomposition; separable non-split gives a product of field extensions and no eigenvalues in $k$. Separable elements and separable algebras are different notions: $M_n(k)$ is a separable algebra that may contain inseparable elements when $k$ is imperfect.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $k$ | A field, the base of the article |
| $f$, $m_a$ | A monic polynomial; the minimal polynomial of $a$ |
| $\alpha_1, \dots, \alpha_n$ | The roots of $f$ in a splitting field, counted with multiplicity |
| $\Delta(f)$ | The discriminant $\prod_{i |
| $f'$ | The formal derivative; $\gcd(f,f')=1$ exactly when $f$ is separable |
| perfect field | A field over which every irreducible polynomial is separable |
| separable element | An element whose minimal polynomial is separable |
| split element | An element whose minimal polynomial splits over $k$ |
| $\tau(x,y) = \operatorname{Tr}(L_{xy})$ | The trace form of $k[a]$, taken in the regular representation of $k[a]$ |
| $T_{ij} = \operatorname{Tr}(a^{i+j})$ | The matrix of the trace form in the power basis; $\det T = \Delta(m_a)$ |
| $m_a = \prod_i (x-\lambda_i)^{e_i}$ | The split form, with distinct $\lambda_i$ and exponents $e_i \ge 1$ |
| $x^p - t$ | The standard inseparable polynomial over $\mathbb{F}_p(t)$ |
Further Reading
- Nicolas Bourbaki, Algèbre, Chapitre V, for separable and inseparable polynomials, the discriminant and the trace form.
- Serge Lang, Algebra (Springer, 3rd ed. 2002), Chapters V and VI, for separability, perfect fields and the discriminant.
- Thomas W. Hungerford, Algebra (Springer, 1974), Chapter V, for separable extensions and the derivative criterion.
- Richard S. Pierce, Associative Algebras (Springer, 1982), for separable algebras, the trace form and the separability idempotent.
- Max-Albert Knus, Alexander Merkurjev, Markus Rost and Jean-Pierre Tignol, The Book of Involutions (American Mathematical Society, 1998), for separable algebras over rings and the trace form.
- Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge University Press, 2nd ed. 2013), Chapter 3, for the discriminant, the Vandermonde determinant and the trace form of a cyclic matrix.