Roots of the Identity, Split Algebras and Nontrivial Idempotents

Introduction

A root of the identity in an algebra $A$ is an element $u$ with $u^2=1$. The name distinguishes the equation from the roots of unity of higher order: a root of the identity is a solution of $x^2-1=0$, and the article never considers $u^n=1$ for $n>2$. The roots of the identity are the algebraic form of the reflections of a vector space, and their theory is the theory of the simplest spectral decomposition, the one with the two eigenvalues $+1$ and $-1$.

The link with idempotents is the halving correspondence. When $2$ is invertible, an idempotent $e$ gives the root $u=2e-1$, and a root $u$ gives the idempotent $e=(1+u)/2$; the two constructions are inverse. The root acts as $+1$ on the image of $e$ and as $-1$ on the image of $1-e$, so the pair $(e,1-e)$ is the spectral decomposition of $u$, and the halving is the passage between an idempotent and its reflection. The correspondence is a bijection of sets, not a ring homomorphism: it does not preserve products, even when the idempotents commute; it preserves instead the Boolean group structure of commuting idempotents, as §What the correspondence does not preserve shows. In characteristic two it collapses completely.

The article uses the correspondence to separate three classes of algebras. A division algebra has no nontrivial idempotent and no root of the identity other than $\pm1$. A split algebra, in the weak sense of this article, has a nontrivial idempotent, equivalently a root other than $\pm1$ when $2$ is invertible. The two classes do not exhaust the algebras: the dual numbers have no nontrivial idempotent and are not a division algebra, and they are the standard warning that "no nontrivial idempotent" is weaker than "division." In the semisimple case the warning disappears: a semisimple algebra has no nontrivial idempotent exactly when it is a division algebra, and it is split exactly when it is not a division algebra.

The article assumes the idempotent theory of Unital Algebras, §Idempotents and the Peirce Decomposition, the involutions of Involutive Algebras only as a contrast, and the primitive idempotents of Primitive Idempotents and Minimal Left Ideals. The matrix case is Matrix Algebras; the split-complex algebra is Split-Complex Algebra and The Number Systems as Clifford Algebras; the dual numbers are Dual-Numbers Algebra; the semisimple structure is Simple and Semisimple Modules, §Semisimple Algebras and Wedderburn–Artin; and the Jordan–Chevalley decomposition is Nilpotent Elements and the Jordan–Chevalley Decomposition. No form, norm or topology is used.

Throughout, $F$ is a field and $A$ is a unital associative $F$-algebra with $1\neq0$, not assumed commutative. The identity is $1_A$, written $1$ when no confusion is possible. The characteristic is arbitrary unless stated, and the halving correspondence is stated for $2$ invertible; the characteristic-two section records what replaces it. The idempotent is $e$ and its complement is $f=1-e$.

Roots of the Identity

Definition and first properties

Definition. A root of the identity of $A$ is an element $u\in A$ with $u^2=1$. The set of roots is

$$ R(A)=\{u\in A:u^2=1\}. $$

The trivial roots are $1$ and $-1$; they coincide in characteristic two. A root is invertible, with $u^{-1}=u$, so $R(A)\subseteq A^\times$.

Proposition. Let $u,v\in R(A)$.

(a) $u^{-1}=u$, and $u^{-1}\in R(A)$.

(b) If $u$ and $v$ commute, then $uv\in R(A)$; in particular the central roots $R(A)\cap Z(A)$ form a subgroup of $A^\times$.

(c) If $A$ is commutative, $R(A)$ is a subgroup of $A^\times$, the group of roots of the identity.

(d) If $A$ is not commutative, $R(A)$ need not be closed under multiplication.

Proof. (a) $u^2=1$ gives $u^{-1}=u$, and $(u^{-1})^2=u^2=1$. (b) $(uv)^2=uvuv=u^2v^2=1$ when $uv=vu$. The central roots are closed under multiplication, contain $1$, and are closed under inversion by (a), so they form a subgroup. (c) is (b) with every pair commuting. (d) In $M_2(F)$ with $F$ of characteristic not two, let

$$ u=\begin{pmatrix}0&1\\1&0\end{pmatrix}, \qquad v=\begin{pmatrix}1&0\\0&-1\end{pmatrix}. $$

Then $u^2=v^2=1$, but

$$ uv=\begin{pmatrix}0&-1\\1&0\end{pmatrix}, \qquad (uv)^2=\begin{pmatrix}-1&0\\0&-1\end{pmatrix}\neq1. $$

So the product of two roots need not be a root.

Remark. The roots of the identity are not the roots of $-1$. The quaternion algebra $\mathbb{H}$ has the two roots $\pm1$ of $x^2=1$ and the whole sphere of roots of $x^2=-1$; the two equations have different theories. The roots of $-1$ are the subject of the complex structures and the almost-complex geometry, and they are not used here.

Division algebras

Proposition. Let $D$ be a division algebra over $F$.

(a) If $2\neq0$ in $F$, then $R(D)=\{1,-1\}$.

(b) If $2=0$ in $F$, then $R(D)=\{1\}$.

(c) The only idempotents of $D$ are $0$ and $1$.

Proof. (a) $u^2-1=(u-1)(u+1)=0$. A division algebra has no zero divisors, so $u-1=0$ or $u+1=0$. (b) In characteristic two, $u^2-1=(u-1)^2$, so $u-1=0$. (c) $e^2=e$ gives $e(e-1)=0$, hence $e=0$ or $e=1$.

The proposition is the model of the division side of the article: no nontrivial idempotent, and no nontrivial root of the identity. The converse is the subject of §Division Algebras and Nontrivial Idempotents, where it is shown to fail without a hypothesis.

Matrix algebras

Proposition. Let $A=M_n(F)$ and suppose $2\neq0$ in $F$. A matrix $U\in A$ satisfies $U^2=I$ exactly when it is diagonalisable over $F$ with eigenvalues in $\{1,-1\}$. The idempotents of $A$ are the diagonalisable matrices with eigenvalues in $\{0,1\}$, and the halving correspondence of §The Halving Correspondence is a bijection between the two sets.

Proof. If $U^2=I$, the polynomial $x^2-1=(x-1)(x+1)$ has distinct roots and annihilates $U$, so the minimal polynomial divides $x^2-1$ and has distinct linear factors; hence $U$ is diagonalisable with eigenvalues $\pm1$. Conversely a diagonalisable matrix with eigenvalues $\pm1$ satisfies $U^2=I$. The idempotent statement is the same argument with $x^2-x=x(x-1)$, and the halving is the bijection $E\mapsto2E-I$ of §The Halving Correspondence.

Remark. In characteristic two the polynomial $x^2-1=(x-1)^2$ has a repeated root, and a root of the identity is a unipotent matrix $U=1+N$ with $N^2=0$; it need not be diagonalisable. The idempotents are still the diagonalisable matrices with eigenvalues $0,1$, and the halving correspondence is no longer available. The matrix case is therefore the cleanest illustration of both the correspondence and its characteristic-two failure.

Example (the rank). In $M_n(F)$ with $2\neq0$, an idempotent of rank $r$ corresponds to a root of the identity with eigenvalue $+1$ of multiplicity $r$ and eigenvalue $-1$ of multiplicity $n-r$. The trace of the root is $2r-n$, and the rank is recovered as $r=(n+\operatorname{tr}U)/2$. The correspondence is a bijection between the idempotents of rank $r$ and the involutions with $+1$-eigenspace of dimension $r$. For $n=2$ the nontrivial idempotents are the rank-one projections, and the nontrivial roots are the reflections $\operatorname{diag}(1,-1)$ and $\operatorname{diag}(-1,1)$ in a suitable basis, together with the off-diagonal reflections; the halving pairs each projection with its reflection.

The Halving Correspondence

The bijection

Theorem. Suppose $2$ is invertible in $F$. The maps

$$ \boxed{\;e\longmapsto u=2e-1, \qquad u\longmapsto e=\frac{1+u}{2}\;} $$

are inverse bijections between the set $\operatorname{Idem}(A)$ of idempotents and the set $R(A)$ of roots of the identity.

Proof. If $e^2=e$, then

$$ (2e-1)^2=4e^2-4e+1=1, $$

so $u=2e-1$ is a root. If $u^2=1$, then

$$ \left(\frac{1+u}{2}\right)^2=\frac{1+2u+u^2}{4}=\frac{2+2u}{4}=\frac{1+u}{2}, $$

so $e=(1+u)/2$ is idempotent. The two composites are

$$ \frac{1+(2e-1)}{2}=e, \qquad 2\left(\frac{1+u}{2}\right)-1=u, $$

so the maps are inverse. The use of $2$ is explicit in both formulas.

Corollary. Under the correspondence:

(a) $e=0$ corresponds to $u=-1$, and $e=1$ corresponds to $u=1$.

(b) The complement $1-e$ corresponds to $-u$.

(c) The idempotent $e$ is the projection onto the $+1$-eigenspace of $u$, and $1-e$ is the projection onto the $-1$-eigenspace; in formulas,

$$ ue=eu=e, \qquad u(1-e)=(1-e)u=-(1-e). $$

(d) The root $u$ is a reflection: $u^2=1$, and the algebra decomposes as $A=Ae\oplus A(1-e)$ with $u$ acting as $+1$ on the first summand and $-1$ on the second.

Proof. (a) and (b) are immediate from the formulas. (c) $ue=(2e-1)e=2e-e=e$ and $u(1-e)=(2e-1)(1-e)=-(1-e)$; the same on the other side. (d) is (c) read on the direct sum.

What the correspondence does not preserve

The halving is a bijection of sets, and it does not respect the multiplicative structures.

Proposition. Let $e,f$ be idempotents with roots $u=2e-1$, $v=2f-1$.

(a) $uv$ is a root of the identity exactly when $u$ and $v$ commute, equivalently when $e$ and $f$ commute.

(b) When $e$ and $f$ commute, the idempotent corresponding to $uv$ is

$$ 1-(e+f-2ef), $$

the complement of the symmetric difference of $e$ and $f$.

(c) In a commutative algebra, the idempotents form a Boolean ring under

$$ e\oplus f=e+f-2ef, \qquad e\wedge f=ef, $$

and the map $e\mapsto1-2e$ is an isomorphism of groups from $(\operatorname{Idem}(A),\oplus)$ onto $(R(A),\cdot)$. The map $e\mapsto2e-1$ is its negative; it is a bijection, but it is not a homomorphism, and it satisfies $w(e\oplus f)=-w(e)w(f)$.

Proof. (a) $(uv)^2=uvuv=u^2v^2=1$ exactly when $uv=vu$. Since $u=2e-1$ and $v=2f-1$ are polynomials in $e$ and $f$, $uv=vu$ exactly when $ef=fe$. (b) When $e$ and $f$ commute,

$$ uv=(2e-1)(2f-1)=4ef-2e-2f+1=1-2(e+f-2ef), $$

and $g=e+f-2ef$ is idempotent because $e$ and $f$ commute; the root $uv$ corresponds to $(1+uv)/2=1-g$. (c) The Boolean ring laws are the usual ones for commuting idempotents, and

$$ (1-2e)(1-2f)=1-2(e+f-2ef)=1-2(e\oplus f), $$

so $e\mapsto1-2e$ is multiplicative; it is bijective by the theorem and preserves $0\mapsto1$, hence is a group isomorphism. The map $w(e)=2e-1=-(1-2e)$ is the negative of this isomorphism; it is bijective, but $w(e\oplus f)=-(1-2(e\oplus f))=-w(e)w(f)$, so it is not a homomorphism.

Remark. In a noncommutative algebra the idempotents do not form a Boolean ring, and the halving is only a bijection. The product of two roots can fail to be a root, as the $2\times2$ example of §Definition and first properties shows; the product of two commuting roots is a root, but it corresponds to the complement of the symmetric difference of the two idempotents, not to their product. The correspondence is therefore a statement about the additive Boolean structure, not about the multiplicative structure of the algebra.

The spectral reading

The halving is the simplest case of the eigenvalue decomposition. The element $u$ has the two eigenvalues $+1$ and $-1$, and the idempotents $e$ and $1-e$ are the corresponding eigenprojectors:

$$ A=Ae\oplus A(1-e), \qquad u=\;1\cdot e+(-1)\cdot(1-e)=2e-1. $$

The decomposition is the two-term case of The Eigenvalue Decomposition of an Element, and the formula $e=(1+u)/2$ is the Lagrange–Sylvester formula for the two eigenvalues. The root $u$ is diagonalisable in the sense of that article exactly when $e$ exists, which is automatic when $2$ is invertible; in characteristic two the spectral decomposition is not available in this form, and the Jordan–Chevalley decomposition of Nilpotent Elements and the Jordan–Chevalley Decomposition is the substitute.

Division Algebras and Nontrivial Idempotents

The division implication

Proposition. A division algebra has no idempotent other than $0$ and $1$, and no root of the identity other than $\pm1$ when $2\neq0$, or other than $1$ when $2=0$.

Proof. This is the proposition of §Division Algebras.

The implication is one-way. The converse fails, and the dual numbers are the standard counterexample.

Example (the dual numbers). Let $A=F[\varepsilon]/(\varepsilon^2)$ with $2\neq0$. An idempotent $e=a+b\varepsilon$ satisfies $e^2=a^2+2ab\varepsilon$ and $e^2=e$, so $a^2=a$ and $2ab=b$. If $a=0$, then $b=0$; if $a=1$, then $2b=b$, hence $b=0$. Thus the only idempotents are $0$ and $1$. The algebra is not a division algebra, since $\varepsilon\neq0$ and $\varepsilon^2=0$. The roots of the identity are also only $\pm1$: $u=a+b\varepsilon$ gives $u^2=a^2+2ab\varepsilon=1$, so $a=\pm1$ and $b=0$. So the dual numbers have the same idempotents and the same roots as a division algebra, without being one.

Remark (the correct converse). The converse holds under a hypothesis that excludes the nilpotent direction. A finite-dimensional algebra over a field with no zero divisors is a division algebra; a finite-dimensional semisimple algebra with no nontrivial idempotent is a division algebra. The second statement follows from Wedderburn–Artin: a semisimple algebra is a product of matrix algebras $M_{n_i}(D_i)$, and a factor with $n_i\geq2$ has the nontrivial idempotent $E_{11}$, while a product of two or more factors has the central idempotents $(1,0,\dots,0)$; hence a semisimple algebra with no nontrivial idempotent is a single $M_1(D)=D$, a division algebra. The dual numbers fail both hypotheses: they have zero divisors and they are not semisimple.

The semisimple dichotomy

Theorem. Let $A$ be a finite-dimensional semisimple algebra over $F$. Then $A$ is a division algebra exactly when it has no nontrivial idempotent. Equivalently, $A$ has a nontrivial idempotent exactly when it is not a division algebra.

Proof. A division algebra has no nontrivial idempotent by the proposition. Conversely, if $A$ is semisimple and has no nontrivial idempotent, Wedderburn–Artin writes $A\cong\prod_i M_{n_i}(D_i)$; as in the remark, the product has only one factor and $n_1=1$, so $A\cong D_1$ is a division algebra. The contrapositive gives the second statement.

Corollary. In the semisimple case the three conditions

$$ \boxed{\;A\ \text{is a division algebra}, \qquad A\ \text{has no nontrivial idempotent}, \qquad R(A)=\{\pm1\}\ (2\neq0)\;} $$

are equivalent. The dual numbers show that none of the equivalences survives without semisimplicity.

Split Algebras

Definition

Definition (this article). An algebra $A$ is split when it contains a nontrivial idempotent. Equivalently, the regular left module splits as a direct sum $A=Ae\oplus A(1-e)$ of two nonzero left ideals, and the regular right module splits in the same way.

The word is used in the weak sense of the title: the algebra splits as a module over itself. It is not the stronger notion split semisimple, which requires $A\cong\prod_i M_{n_i}(F)$; the relation between the two is stated at the end of the section.

Proposition. Let $A$ be an algebra.

(a) $A$ is split exactly when it has a nontrivial idempotent.

(b) If $e$ is a nontrivial idempotent, then $A=Ae\oplus A(1-e)$ with both summands nonzero; if $e$ is central, the decomposition is a product of algebras, $A\cong Ae\times A(1-e)$.

(c) If $A$ is split and $2$ is invertible, then $A$ has a root of the identity other than $\pm1$.

Proof. (a) is the definition. (b) is the Peirce decomposition at the level of the regular module; the product statement is the central idempotent theorem of Unital Algebras. (c) If $e$ is nontrivial, $u=2e-1$ is a root by the halving theorem, and $u\neq\pm1$ because $e\neq0,1$.

The root criterion

Theorem. Suppose $2$ is invertible in $F$. An algebra $A$ is split exactly when it has a root of the identity other than $1$ and $-1$.

Proof. If $e$ is a nontrivial idempotent, $u=2e-1$ is a root and $u\neq\pm1$. Conversely, if $u\in R(A)$ and $u\neq\pm1$, then $e=(1+u)/2$ is an idempotent and $e\neq0,1$, because $u\neq-1,1$. The two constructions are inverse by the halving theorem.

Corollary. For an algebra over a field of characteristic not two, the following are equivalent:

$$ \boxed{\;A\ \text{is split}, \qquad A\ \text{has a nontrivial idempotent}, \qquad A\ \text{has a nontrivial root of the identity}\;}. $$

The equivalence fails in characteristic two, where the halving is unavailable and the roots of the identity can be plentiful while the idempotents are not; §Characteristic Two gives the examples.

Examples

Example (the split algebra $F\times F$). The idempotents are $(1,0)$ and $(0,1)$, both nontrivial; the roots of the identity are $(\pm1,\pm1)$, four elements when $2\neq0$, and the halving pairs

$$ (1,0)\longleftrightarrow(1,-1), \qquad (0,1)\longleftrightarrow(-1,1). $$

The algebra is the split-complex algebra $F[j]/(j^2-1)$ under the identification $j=(1,-1)$; the two nontrivial roots $\pm j$ correspond to the two idempotents $(1\pm j)/2$.

Example (matrix algebras). For $n\geq2$, $M_n(F)$ is split: the matrix units $E_{11}$ and $E_{22}$ are nontrivial idempotents, and the rank-one idempotents are the primitive ones. The roots of the identity are the involutions, and the halving pairs an idempotent of rank $r$ with the involution whose $+1$-eigenspace has dimension $r$. For $n=1$ the algebra is the field $F$, a division algebra, and it is not split.

Example (the split-complex algebra). $A=F[j]/(j^2-1)$ has the idempotents $e_\pm=(1\pm j)/2$ and the roots $\pm1,\pm j$. The two nontrivial roots $\pm j$ correspond to $e_\pm$; the algebra is isomorphic to $F\times F$ when $2\neq0$, and the isomorphism sends $e_+$ to $(1,0)$ and $e_-$ to $(0,1)$. In characteristic two, $j^2-1=(j-1)^2$, the algebra is the dual numbers $F[\varepsilon]/(\varepsilon^2)$ with $\varepsilon=j-1$, and it is not split: its only idempotents are $0,1$.

Example (the group algebra of the cyclic group of order two). Let $A=F[C_2]$ with $C_2=\{1,g\}$, $g^2=1$. If $2\neq0$, the elements $e_\pm=(1\pm g)/2$ are idempotents and $A\cong F\times F$; the roots are $\pm1,\pm g$, and the nontrivial ones correspond to $e_\pm$. If $2=0$, $A\cong F[\varepsilon]/(\varepsilon^2)$ with $\varepsilon=g-1$, and $A$ is not split.

Example (the upper triangular algebra). $T_2(F)$ is split for every field: $E_{11}$ and $E_{22}$ are nontrivial idempotents. When $2\neq0$, the roots of the identity include $\operatorname{diag}(1,-1)$ and $\operatorname{diag}(-1,1)$; in characteristic two the roots are the unipotent triangular matrices with $N^2=0$. The algebra is not semisimple, but it is split in the sense of the definition. This shows that split is not the same as semisimple.

Example (the dual numbers). $F[\varepsilon]/(\varepsilon^2)$ is not split when $2\neq0$: it has no nontrivial idempotent and no root other than $\pm1$. It is also not a division algebra. It is the standard example of an algebra that is neither division nor split, and it shows that the dichotomy of §The Semisimple Dichotomy is a semisimple phenomenon.

Example (the quaternions). $\mathbb{H}$ is a division algebra, so it is not split; its roots of $x^2=1$ are $\pm1$, while its roots of $x^2=-1$ are the unit imaginary quaternions. The example separates the two equations and shows that a division algebra can have many square roots of $-1$ while having no nontrivial idempotent.

Split versus split semisimple

The word split has a second, stronger meaning in the theory of algebras: a finite-dimensional algebra is split semisimple when it is isomorphic to a product $\prod_i M_{n_i}(F)$ of matrix algebras over the base field. A split semisimple algebra is semisimple, and every division algebra in its Wedderburn–Artin decomposition is the base field.

Proposition. Let $A$ be a finite-dimensional semisimple algebra.

(a) If $A$ is split semisimple, then $A$ is split in the weak sense exactly when $A$ is not a single copy of $F$; that is, when it has more than one factor or a factor $M_n(F)$ with $n\geq2$.

(b) A central division algebra $D\neq F$ is not split semisimple and is not split in the weak sense; it becomes a matrix algebra after a splitting field extension, but the extension, not the algebra, carries the idempotents.

Proof. (a) A factor $M_n(F)$ with $n\geq2$ has $E_{11}$, and a product of two or more factors has the central idempotents; the single factor $M_1(F)=F$ has none. (b) A division algebra has no nontrivial idempotent by §Division Algebras; for a central division algebra the splitting field extension $L/F$ with $D\otimes_F L\cong M_n(L)$ produces idempotents in the scalar extension, not in $D$.

Remark. The two senses of split are therefore different. The weak sense is a property of the algebra over the base field: it has an idempotent. The strong sense is a property of the semisimple algebra and its division algebra factors: all the factors are matrix algebras over the base field. A central division algebra $D\neq F$ is neither weakly split nor split semisimple, but it is split by a field extension; the phrase "split by $L$" refers to the extension and not to an idempotent of $D$.

Characteristic Two

In characteristic two the halving correspondence collapses: $2=0$, so the formula $e=(1+u)/2$ is undefined, and $u=2e-1=-1$ for every idempotent. The roots of the identity and the idempotents separate, and neither determines the other.

The roots

Proposition. Let $2=0$ in $F$.

(a) If $u^2=1$, then $(u-1)^2=0$; the root $u$ is a unipotent element $1+n$ with $n^2=0$.

(b) In a reduced algebra, the only root of the identity is $1$.

(c) In a division algebra, the only root of the identity is $1$.

Proof. (a) $u^2-1=(u-1)^2$ in characteristic two. (b) A reduced algebra has no nonzero nilpotent, so $n=0$. (c) A division algebra has no nonzero nilpotent, or directly $(u-1)^2=0$ gives $u=1$.

Example (the dual numbers in characteristic two). Let $A=F[\varepsilon]/(\varepsilon^2)$ with $2=0$. Every element $1+b\varepsilon$ is a root of the identity, since $(1+b\varepsilon)^2=1+2b\varepsilon+b^2\varepsilon^2=1$. The idempotents, however, are only $0$ and $1$: an idempotent $a+b\varepsilon$ satisfies $a^2=a$ and $2ab=0$, the second equation being automatic; if $a=0$ then $b=0$, and if $a=1$ then $b=0$. So the roots form the additive group $F$ under $1+b\varepsilon$, while the idempotents are the two trivial ones. The halving has no inverse and no meaning.

Example (the split algebra in characteristic two). Let $A=F\times F$ with $2=0$. The idempotents are $(1,0)$ and $(0,1)$, but the roots of the identity are only $(1,1)$: a root $(a,b)$ satisfies $a^2=1$ and $b^2=1$, and in characteristic two the only solution is $a=b=1$. So the algebra is split and has no nontrivial root. The example is the exact opposite of the dual numbers, where the roots are plentiful and the idempotents are not.

The table

Algebra (characteristic two) Nontrivial idempotents Nontrivial roots of $x^2=1$
$F\times F$ yes, $(1,0),(0,1)$ no, only $(1,1)$
$F[\varepsilon]/(\varepsilon^2)$ no, only $0,1$ yes, $1+b\varepsilon$ for $b\neq0$
a reduced algebra may have no, only $1$
a division algebra no no, only $1$

The table is the characteristic-two form of the separation: the halving correspondence is a characteristic-not-two theorem, and in characteristic two the two notions must be studied separately.

Summary

A root of the identity is a solution of $u^2=1$. In a division algebra the roots are $\pm1$ when $2\neq0$ and $1$ when $2=0$, and there is no nontrivial idempotent. In a matrix algebra over a field of characteristic not two, the roots are the diagonalisable involutions with eigenvalues $\pm1$, and the idempotents are the diagonalisable projections with eigenvalues $0,1$. When $2$ is invertible, the halving correspondence $e\mapsto2e-1$, $u\mapsto(1+u)/2$ is a bijection between idempotents and roots; it sends the complement $1-e$ to $-u$, reads $u$ as the reflection with $+1$-eigenspace the image of $e$ and $-1$-eigenspace the image of $1-e$, and is a Boolean group isomorphism $e\mapsto1-2e$ on a commutative algebra but only a bijection in general. An algebra is split in the weak sense of the article when it has a nontrivial idempotent, equivalently when it has a nontrivial root of the identity if $2$ is invertible; $F\times F$, $M_n(F)$ for $n\geq2$, the split-complex algebra, the group algebra $F[C_2]$ and the upper triangular algebra are split, while a division algebra and the dual numbers are not. The dual numbers are neither division nor split, so the division/split dichotomy is a semisimple phenomenon; in the semisimple case a finite-dimensional algebra is division exactly when it has no nontrivial idempotent, and split exactly when it is not division. In characteristic two the halving collapses, the roots of $x^2=1$ are the unipotent elements $1+n$ with $n^2=0$, and the roots and the idempotents separate: $F\times F$ has idempotents and only the trivial root, while the dual numbers have many roots and only the trivial idempotents. The article uses no form, norm or topology.

Summary of Notation

Symbol Meaning
$R(A)$ the roots of the identity, $\{u:u^2=1\}$
$e,f$ idempotent and its complement $f=1-e$
$u=2e-1$ the root of the identity attached to $e$
$e=(1+u)/2$ the idempotent attached to $u$
$e\oplus f=e+f-2ef$ the symmetric difference of commuting idempotents
split has a nontrivial idempotent
split semisimple isomorphic to a product of matrix algebras over the base field
division every nonzero element is invertible

Further Reading

  • Richard S. Pierce, Associative Algebras (Springer, 1982), for idempotents, Peirce decompositions and the matrix examples.
  • Nathan Jacobson, Basic Algebra II (Freeman, 1989), for the halving of an idempotent, the roots of the identity and the Wedderburn–Artin theory.
  • T. Y. Lam, A First Course in Noncommutative Rings (Springer, 2001), for the division, semisimple and local cases and the dual-numbers warning.
  • John Voight, Quaternion Algebras (Springer, 2021), for the division algebras, their splitting fields and the contrast between the roots of $+1$ and the roots of $-1$.