Clifford Algebras in Finite Dimensions

Introduction

This article introduces the properties of Clifford algebras in finite dimensions. The treatment is introductory and purely mathematical.

The goal is to explain what is common to all Clifford algebras in finite dimensions, regardless of the commutative ring of scalars.

We assume familiarity with modules, bilinear forms, quadratic forms, and Clifford algebras, as developed in the preceding articles.

Throughout this article, the module is finite-dimensional (that is, free of finite rank). The general definition of the Clifford algebra is treated in the preceding article. The article treats both the non-degenerate and the degenerate cases, and describes how the degenerate case reduces to the non-degenerate one.

The article treats only the properties that hold over any commutative ring (in which 2 is invertible) or over any field. The specialization to specific rings is not covered here.

A word on the base structure. The classical theory of Clifford algebras assumes that the scalars form a field. But the definition and many of the basic properties carry over to the more general setting where the scalars form a commutative ring. This is the setting we adopt here. The main differences from the field case are:

Modules over a commutative ring need not be free, so the Clifford algebra need not have a well-defined rank. We assume the module is free of finite rank.
Non-degeneracy of the quadratic form is defined via an isomorphism $M \to M^*$, which requires care over a general commutative ring.
The polarization identity requires that 2 is invertible in the base ring.
The classification of Clifford algebras over a general commutative ring is much more subtle than over a field.

We indicate where these differences matter. When we wish to specialize to a field, we say so explicitly.

Part I: The Definition and Universal Property

1. The Definition

Let $R$ be a commutative ring (in which 2 is invertible), and let $M$ be a free $R$-module of finite rank $n$ with a quadratic form $Q$.

The Clifford algebra $Cl(M, Q)$ is the quotient of the tensor algebra $T(M)$ by the two-sided ideal generated by the elements

$$ v \otimes v - Q(v) \cdot 1 $$

for all $v \in M$. Equivalently, it is the associative unital algebra over $R$ generated by $M$ subject to the relations

$$ v^2 = Q(v) \cdot 1. $$

This definition makes sense over any commutative ring.

2. The Universal Property

The construction has a universal property. Let $A$ be a unital associative algebra over $R$. Any $R$-linear map $f : M \to A$ such that

$$ f(v)^2 = Q(v) \cdot 1_A $$

for all $v \in M$ extends uniquely to an algebra homomorphism $Cl(M, Q) \to A$.

This property is what we mean when we say the Clifford algebra is the "largest" algebra in which the relation $v^2 = Q(v) \cdot 1$ holds. It holds over any commutative ring.

3. The Fundamental Relation

The fundamental relation

$$ uv + vu = 2B(u, v) \cdot 1 $$

holds over any commutative ring in which 2 is invertible, where $B$ is the bilinear form associated with $Q$. It is a consequence of the defining relation $v^2 = Q(v) \cdot 1$, not an independent assumption. The polarization identity used to derive it requires division by 2.

Key difference from the field case. Over a field of characteristic not equal to 2, 2 is automatically invertible. Over a general commutative ring, we must assume that 2 is a unit. Over $\mathbb{Z}$, this fails, and the theory of Clifford algebras over $\mathbb{Z}$ requires more care.

Part II: The Radical and the Reduction to the Non-Degenerate Case

4. The Radical

The radical of $Q$ is the submodule

$$ \mathrm{rad}(Q) = \{v \in M : B(v, w) = 0 \text{ for all } w \in M\}. $$

It is the kernel of the linear map $M \to M^*$, $v \mapsto B(v, \cdot)$. The restriction of $Q$ to $\mathrm{rad}(Q)$ is zero, and the quotient $M/\mathrm{rad}(Q)$ carries an induced quadratic form $\bar{Q}$ whose radical is zero.

Basic properties.

$\mathrm{rad}(Q)$ is the set of vectors orthogonal to every vector.
$Q$ is non-degenerate iff $\mathrm{rad}(Q) = 0$ and the map $M \to M^*$ is an isomorphism.
For $r, s \in \mathrm{rad}(Q)$, the fundamental relation gives $r s + s r = 2 B(r, s) \cdot 1 = 0$, so $r s = -s r$ and $r^2 = 0$.

The radical is the obstruction to non-degeneracy, and it is the source of the nilpotent factor in the decomposition.

5. The Reduction

Over a field, every subspace of a finite-dimensional vector space is a direct summand. So the radical $\mathrm{rad}(Q)$ is a direct summand of $M$, and there is a complement $W$ with

$$ M = W \oplus \mathrm{rad}(Q), $$

where the restriction of $Q$ to $W$ is non-degenerate. Over a general commutative ring, the radical need not be a direct summand, and the decomposition requires an additional hypothesis. We assume the radical is a direct summand when we state the decomposition over a general commutative ring.

6. The Decomposition of a Degenerate Clifford Algebra

Theorem. Let $M$ be a free $R$-module of finite rank with quadratic form $Q$, radical $\mathrm{rad}(Q)$, and induced non-degenerate form $\bar{Q}$ on $M/\mathrm{rad}(Q)$. Suppose that $M$ is the direct sum of $\mathrm{rad}(Q)$ and a complement $W$ on which $Q$ is non-degenerate. Then there is an algebra isomorphism

$$ Cl(M, Q) \cong Cl(M/\mathrm{rad}(Q), \bar{Q}) \hat{\otimes} \Lambda(\mathrm{rad}(Q)), $$

where $\hat{\otimes}$ is the graded tensor product, and $\Lambda(\mathrm{rad}(Q))$ is the exterior algebra on $\mathrm{rad}(Q)$.

Proof. Write $M = W \oplus \mathrm{rad}(Q)$ with $W$ non-degenerate. The Clifford relations split into three families:

For $v, w \in W$: $v w + w v = 2 B(v, w) \cdot 1$.
For $v \in W$, $r \in \mathrm{rad}(Q)$: $v r + r v = 2 B(v, r) \cdot 1 = 0$.
For $r, s \in \mathrm{rad}(Q)$: $r s + s r = 2 B(r, s) \cdot 1 = 0$.

The first family generates $Cl(W, Q|_W) \cong Cl(M/\mathrm{rad}(Q), \bar{Q})$. The third family generates $\Lambda(\mathrm{rad}(Q))$, since $r^2 = 0$ and $r s = -s r$. The second family says that the generators of $W$ and the generators of $\mathrm{rad}(Q)$ anticommute, which is the graded tensor product relation. So the algebra is the graded tensor product of the two.

The isomorphism is not canonical: it depends on the choice of complement $W$.

Key difference from the field case. Over a field, every module is free, so the complement always exists. Over a general commutative ring, the module need not be free, and the complement need not exist. The decomposition holds under the additional hypothesis that the radical is a direct summand.

7. The Non-Degenerate Factor

The factor $Cl(M/\mathrm{rad}(Q), \bar{Q})$ is the ordinary Clifford algebra of a non-degenerate form. Its structure depends on the ring $R$ and on the form $\bar{Q}$.

Over a general commutative ring, the non-degenerate Clifford algebra is a central simple algebra over $R$, or a product of two central simple algebras, when the relevant hypotheses hold. Its class in the Brauer group of $R$ is determined by the discriminant and the Hasse invariant of $\bar{Q}$. The algebra is $\mathbb{Z}/2$-graded, with even part $Cl^0$ and odd part $Cl^1$, and the even part is a central simple algebra over $R$ when the rank of $M/\mathrm{rad}(Q)$ is odd.

Key difference from the field case. Over a general commutative ring, the classification of non-degenerate Clifford algebras is much more subtle than over a field. It requires the theory of quadratic forms over rings, and the Brauer group is replaced by the Brauer group of Azumaya algebras over $R$. The classification is the subject of the article Clifford Algebras over Commutative Rings.

8. The Nilpotent Factor

The factor $\Lambda(\mathrm{rad}(Q))$ is the exterior algebra on the radical. If $(r_i)_{i \in I}$ is a basis of $\mathrm{rad}(Q)$, then

$$ \Lambda(\mathrm{rad}(Q)) = \bigoplus_{k \geq 0} \Lambda^k(\mathrm{rad}(Q)), $$

with basis the products $r_{i_1} \cdots r_{i_k}$ for $i_1 < \dots < i_k$ and $k$ ranging over the non-negative integers. The generators satisfy

$$ r_i^2 = 0, \qquad r_i r_j = -r_j r_i, \qquad i \neq j. $$

So $\Lambda(\mathrm{rad}(Q))$ is a graded-commutative algebra, and every element of positive degree is nilpotent.

Theorem. For a field $R$, the algebra $\Lambda(\mathrm{rad}(Q))$ is semisimple iff $\mathrm{rad}(Q) = 0$.

Proof. If $\mathrm{rad}(Q) \neq 0$, the ideal $\Lambda^{\geq 1}(\mathrm{rad}(Q))$ is a non-zero nilpotent ideal (the rank is finite here), so the algebra is not semisimple. If $\mathrm{rad}(Q) = 0$, the algebra is $R$, which is semisimple when $R$ is a field.

So the degenerate Clifford algebra is semisimple iff the quadratic form is non-degenerate, under the usual hypotheses on the base ring.

The structure of $\Lambda(\mathrm{rad}(Q))$. Since $M$ has finite rank, $\mathrm{rad}(Q)$ is free of finite rank $r$, and $\Lambda(\mathrm{rad}(Q))$ is free of rank $2^r$. It is local when the base ring is a field, with unique maximal ideal $\Lambda^{\geq 1}(\mathrm{rad}(Q))$; over a non-local ring it need not be local, since the base ring then contributes further maximal ideals.

9. The Rank of a Degenerate Clifford Algebra

Let $M$ be free of finite rank $n$, with radical $\mathrm{rad}(Q)$ of rank $r$ and non-degenerate quotient $M/\mathrm{rad}(Q)$ of rank $n - r$. Then

$$ \operatorname{rank} Cl(M, Q) = 2^{n-r} \cdot 2^r = 2^n. $$

The non-degenerate factor contributes $2^{n-r}$, and the radical factor contributes $2^r$. The total rank is $2^n$, as expected. The decomposition is compatible with the rank count.

10. The Grading and the Filtration in the Degenerate Case

The tensor product decomposition is compatible with the $\mathbb{Z}/2$-grading of the Clifford algebra and with the natural filtration.

The Clifford algebra $Cl(M, Q)$ is $\mathbb{Z}/2$-graded, with

$$ Cl(M, Q) = Cl^0(M, Q) \oplus Cl^1(M, Q), $$

where $Cl^0$ is spanned by products of an even number of elements of $M$, and $Cl^1$ by products of an odd number. The grading decomposes as

$$ Cl^0(M, Q) \cong Cl^0(M/\mathrm{rad}(Q), \bar{Q}) \otimes \Lambda^0(\mathrm{rad}(Q)) \oplus Cl^1(M/\mathrm{rad}(Q), \bar{Q}) \otimes \Lambda^1(\mathrm{rad}(Q)), $$

and similarly for the odd part, with the roles of $\Lambda^0$ and $\Lambda^1$ interchanged. The exterior algebra $\Lambda(\mathrm{rad}(Q))$ is itself $\mathbb{Z}$-graded, and the $\mathbb{Z}/2$-grading of the Clifford algebra is the reduction of the $\mathbb{Z}$-grading modulo $2$ on the radical factor.

The Clifford algebra also carries a natural filtration

$$ F^0 \subseteq F^1 \subseteq F^2 \subseteq \cdots \subseteq Cl(M, Q), $$

where $F^k$ is the span of products of at most $k$ elements of $M$. The associated graded algebra is the exterior algebra on $M$:

$$ \operatorname{gr} Cl(M, Q) \cong \Lambda(M). $$

The tensor product decomposition is compatible with the filtration: the filtration on $Cl(M/\mathrm{rad}(Q), \bar{Q})$ and the grading on $\Lambda(\mathrm{rad}(Q))$ combine to give the filtration on $Cl(M, Q)$.

11. The Lie Algebra Structure in the Degenerate Case

The Clifford algebra carries a Lie bracket, defined by the commutator

$$ [x, y] = xy - yx. $$

The submodule $M_2$ of bivectors is closed under the commutator:

$$ [M_2, M_2] \subseteq M_2. $$

So $M_2$, equipped with the commutator bracket, is a Lie subalgebra of $Cl(M, Q)$. In the non-degenerate case, $M_2$ is isomorphic to the orthogonal Lie algebra $\mathrm{SO}(M, Q)$:

$$ M_2 \cong \mathrm{SO}(M, Q). $$

The degenerate case. If $Q$ is degenerate, the bivectors involving radical elements are nilpotent, and the Lie algebra $\mathrm{SO}(M, Q)$ is replaced by a more complicated object. The spin group does not exist in the usual sense, because the radical elements do not have inverses. The bivectors involving radical elements generate a nilpotent ideal in $M_2$, and the quotient of $M_2$ by this ideal is the Lie algebra of the non-degenerate part.

Part III: Basis and Rank in the Non-Degenerate Case

12. Basis

Let $M$ be free of finite rank $n$, with basis $e_1, \ldots, e_n$. The Clifford algebra $Cl(M, Q)$ has a basis given by the products of generators with strictly increasing indices, including the empty product:

$$ 1, \quad e_i, \quad e_i e_j \ (i < j), \quad e_i e_j e_k \ (i < j < k), \quad \ldots, \quad e_1 e_2 \cdots e_n. $$

The proof is the same over any commutative ring. Every product of generators can be reordered into a product with increasing indices, using the anticommutation relations $e_i e_j = -e_j e_i$ for $i \neq j$. Each repeated factor $e_i^2$ can be replaced by the scalar $Q(e_i)$. So the basis is precisely the set of products with increasing indices.

13. Rank

The rank of $Cl(M, Q)$ as a free $R$-module is

$$ \operatorname{rank} Cl(M, Q) = 2^n, $$

where $n = \operatorname{rank} M$. The count is

$$ \sum_{k=0}^{n} \binom{n}{k} = 2^n, $$

since there is one basis element of degree 0 (the unit), $n$ of degree 1 (the generators), $\binom{n}{2}$ of degree 2, and so on, up to one basis element of degree $n$ (the product of all generators).

This is true over any commutative ring. The rank does not depend on the ring, only on $n = \operatorname{rank} M$.

Key difference from the field case. Over a field, the rank is the dimension, and it is well-defined. Over a general commutative ring, the Clifford algebra is free of rank $2^n$ only if the module $M$ is free of rank $n$. If $M$ is not free, the Clifford algebra need not be free, and its rank may not be well-defined.

14. $k$-Vectors and Multivectors

The elements of degree $k$ are called $k$-vectors. The elements of degree 0 are scalars, the elements of degree 1 are vectors, the elements of degree 2 are bivectors, the elements of degree 3 are trivectors, and so on. Elements that are sums of $k$-vectors of different degrees are called multivectors.

15. The Volume Element

Let $e_1, \ldots, e_n$ be an orthogonal basis, one with $B(e_i, e_j) = 0$ for $i \neq j$, so that the generators anticommute; such a basis exists whenever the form is diagonalizable (§26), in particular over a field, but not over a general commutative ring. The product of all basis generators

$$ \omega = e_1 e_2 \cdots e_n $$

is called the volume element (or pseudoscalar). It satisfies

$$ \omega^2 = (-1)^{n(n-1)/2} Q(e_1) Q(e_2) \cdots Q(e_n) \cdot 1. $$

For a basis that is not orthogonal the formula fails: if $B(e_1, e_2) = \tfrac{1}{2}$, so that $e_2 e_1 = 1 - e_1 e_2$, then $\omega = e_1 e_2$ satisfies $\omega^2 = e_1 (e_2 e_1) e_2 = e_1 e_2 - e_1^2 e_2^2 = \omega - 1$, which is not a scalar.

So the square of the volume element is a scalar, determined by the discriminant of the quadratic form. The volume element is central when $n$ is odd, and it is central up to sign when $n$ is even. More precisely:

If $n$ is odd, $\omega$ commutes with every element of $Cl(M, Q)$.
If $n$ is even, $\omega$ anticommutes with odd elements and commutes with even elements.

This is a finite-dimensional phenomenon: it depends on the existence of a basis and on the rank $n$.

16. The Center

The center of $Cl(M, Q)$ is the set of elements that commute with every element. In finite dimensions:

If $n$ is even, the center is $R$ (the scalars).
If $n$ is odd, the center is $R \oplus R\omega$, where $\omega$ is the volume element.

So the Clifford algebra is central exactly when $n$ is even. When $n$ is odd, the volume element provides a non-trivial central element, and the center is $R \oplus R\omega$, isomorphic to $R[t]/(t^2 - \omega^2)$. The algebra decomposes as a product of two subalgebras precisely when $\omega^2$ is a square in $R$, since such a splitting needs a non-trivial idempotent of the center; this holds when the discriminant is a square, but not otherwise — for $Cl(\mathbb{R}^3)$ one has $\omega^2 = -1$, the center is $\mathbb{C}$, and the algebra is the simple algebra $M_2(\mathbb{C})$. This is the source of the difference between the even and odd cases in the classification.

Part IV: The Even Subalgebra

17. The Decomposition

The Clifford algebra decomposes into an even part and an odd part:

$$ Cl(M, Q) = Cl^+(M, Q) \oplus Cl^-(M, Q), $$

where the even part is spanned by the basis elements of even degree (the unit, the bivectors, the four-vectors, etc.) and the odd part by the basis elements of odd degree (the generators, the trivectors, etc.).

18. Why the Even Part Is a Subalgebra

The even part is closed under multiplication. To see this, note that the grade of a product is determined by the grades of the factors modulo 2. If $u$ has grade $k$ and $v$ has grade $\ell$, then the product $uv$ has grade $k + \ell$ modulo 2. This is because each generator $e_i$ either appears in $u$, or in $v$, or in both. If it appears in both, it can be eliminated using $e_i^2 = Q(e_i)$, which reduces the count by 2. So the parity of the total count is the sum of the parities.

Therefore:

even $\times$ even $=$ even,
even $\times$ odd $=$ odd,
odd $\times$ even $=$ odd,
odd $\times$ odd $=$ even.

In particular, the even part is closed under multiplication: it is a subalgebra of $Cl(M, Q)$. The odd part is not closed under multiplication: the product of two odd elements is even. So the odd part is a module over the even part.

This is true over any commutative ring.

19. The Even Subalgebra as a Clifford Algebra

In finite dimensions, the even subalgebra is itself a Clifford algebra. If $M$ is free of finite rank $n$ with quadratic form $Q$, and if $\dim M \geq 1$, then

$$ Cl^+(M, Q) \cong Cl(M', Q'), $$

where $M'$ is a free module of rank $n - 1$ and $Q'$ is a quadratic form determined by $Q$. The precise form of $Q'$ depends on the choice of a vector $u$ with $Q(u) \neq 0$:

$$ Q'(v) = -Q(u) Q(v - \tfrac{B(u, v)}{Q(u)} u), \qquad v \in u^\perp / R u. $$

This is a finite-dimensional phenomenon: it requires the existence of a vector with non-zero value, which is guaranteed by non-degeneracy in finite dimensions.

The isomorphism $Cl^+(M, Q) \cong Cl(M', Q')$ is the basis of the inductive computation of Clifford algebras, because it reduces the rank by one at each step. It is the algebraic origin of the periodicity of the classification.

Key difference from the field case. The isomorphism holds over any commutative ring in which 2 is invertible, provided that the module is free of finite rank and the form is non-degenerate. Over a general commutative ring, the existence of a vector with $Q(u)$ invertible is not guaranteed, and the isomorphism may fail or require modification.

Part V: The Anti-Involutions

20. Reversion

The reversion anti-involution is the map

$$ x \mapsto x^{r} $$

that reverses the order of the factors in a multivector. On a product of generators, it acts as

$$ e_{i_1} e_{i_2} \cdots e_{i_k} \mapsto e_{i_k} \cdots e_{i_2} e_{i_1}. $$

Reversion is an anti-involution: it satisfies $(xy)^{r} = y^{r}x^{r}$ and $(x^{r})^{r} = x$.

21. Clifford Conjugation

The Clifford conjugation anti-involution is the map

$$ x \mapsto x^{\natural} $$

that reverses the order of the factors and multiplies each vector factor by $-1$. On a product of generators, it acts as

$$ e_{i_1} e_{i_2} \cdots e_{i_k} \mapsto (-1)^k e_{i_k} \cdots e_{i_2} e_{i_1}. $$

Clifford conjugation is also an anti-involution: it satisfies $(xy)^{\natural} = y^{\natural} x^{\natural}$ and $\overline{x^{\natural}} = x$.

22. The Grade Involution

By composing reversion and Clifford conjugation, we get the grade involution:

$$ x \mapsto \alpha(x) = \overline{x^{r}}. $$

On a product of generators, it acts as

$$ e_{i_1} e_{i_2} \cdots e_{i_k} \mapsto (-1)^k e_{i_1} e_{i_2} \cdots e_{i_k}. $$

So the grade involution multiplies an element of grade $k$ by $(-1)^k$. It fixes elements of even grade and changes the sign of elements of odd grade.

The grade involution is an involution, not an anti-involution: it satisfies $\alpha(xy) = \alpha(x)\alpha(y)$ and $\alpha(\alpha(x)) = x$.

23. The Behavior on Each Grade

The behavior of the three maps depends on the grade of the element. For an element of grade $k$:

Reversion multiplies by $(-1)^{k(k-1)/2}$.
Clifford conjugation multiplies by $(-1)^{k(k+1)/2}$.
The grade involution multiplies by $(-1)^k$.

These signs depend only on $k$ modulo 4, not on the ring.

Part VI: The Tensor Product Decomposition

24. The Decomposition

The Clifford algebra of a direct sum of two modules with quadratic forms is the graded tensor product of the Clifford algebras of the summands:

$$ Cl(M_1 \oplus M_2, Q_1 \oplus Q_2) \cong Cl(M_1, Q_1) \hat{\otimes} Cl(M_2, Q_2). $$

In the graded tensor product, odd elements from the two factors anticommute, rather than commute. This ensures that the images of $v_1 \in M_1$ and $v_2 \in M_2$ anticommute, as required by the Clifford relations.

25. Why This Matters

The decomposition reduces the classification of Clifford algebras to the one-dimensional cases. Over any commutative ring, a free module of finite rank can be written as a direct sum of free submodules of rank 1. So the Clifford algebra decomposes as a graded tensor product of rank-one Clifford algebras.

The classification of the rank-one cases, and the resulting periodic pattern, depend on the ring. This is not covered here.

Key difference from the field case. Over a field, the decomposition is always available because every vector space is free and can be decomposed into one-dimensional subspaces. Over a general commutative ring, the decomposition is available only if the module is free of finite rank. If the module is not free, the decomposition may not exist, and the classification is more subtle.

26. The Orthogonal Direct Sum and the Signature

In finite dimensions, the quadratic form can be diagonalized when the base ring is a field: there is a basis in which

$$ Q(x_1, \ldots, x_n) = a_1 x_1^2 + \cdots + a_n x_n^2 $$

for some scalars $a_i \in R$. The Clifford algebra then decomposes as a graded tensor product of rank-one Clifford algebras:

$$ Cl(M, Q) \cong \bigotimes_{i=1}^{n} Cl(R e_i, Q_i), \qquad Q_i(x e_i) = a_i x^2. $$

This is the form in which the classification is usually stated. The scalars $a_i$ determine the rank-one factors, and their product (up to squares) is the discriminant of the form. Over a field, the discriminant and the Hasse invariant are the complete invariants of the quadratic form, and they determine the Clifford algebra up to isomorphism. Over a general commutative ring, the situation is more subtle.

27. The Graded Tensor Product and the Parity

The graded tensor product has a subtlety: the parity of an element in the tensor product is the sum of the parities of its factors. This is compatible with the $\mathbb{Z}/2$-grading of the Clifford algebra:

$$ Cl^0(M_1 \oplus M_2) \cong Cl^0(M_1) \otimes Cl^0(M_2) \oplus Cl^1(M_1) \otimes Cl^1(M_2), $$

$$ Cl^1(M_1 \oplus M_2) \cong Cl^0(M_1) \otimes Cl^1(M_2) \oplus Cl^1(M_1) \otimes Cl^0(M_2). $$

So the even part of the tensor product is not the tensor product of the even parts: it also includes the tensor product of the odd parts. This is a consequence of the graded structure, and it is the reason the classification of Clifford algebras is periodic rather than simply multiplicative.

Summary

The Clifford algebra $Cl(M, Q)$ is the quotient of the tensor algebra $T(M)$ by the two-sided ideal generated by the elements $v \otimes v - Q(v) \cdot 1$. It is the largest associative unital algebra in which the relation $v^2 = Q(v) \cdot 1$ holds.

The universal property characterizes the Clifford algebra: any $R$-linear map $f : M \to A$ satisfying $f(v)^2 = Q(v) \cdot 1$ extends uniquely to an algebra homomorphism $Cl(M, Q) \to A$.

The fundamental relation is $uv + vu = 2B(u, v) \cdot 1$, where $B$ is the bilinear form associated with $Q$. It requires that 2 is invertible in $R$.

The radical is the submodule of vectors orthogonal to every vector. It is the obstruction to non-degeneracy, and it is the source of the nilpotent factor in the decomposition.

The reduction of the degenerate case. Over a field, the radical is a direct summand, and the degenerate Clifford algebra decomposes as

$$ Cl(M, Q) \cong Cl(M/\mathrm{rad}(Q), \bar{Q}) \hat{\otimes} \Lambda(\mathrm{rad}(Q)). $$

The non-degenerate factor is a Clifford algebra of rank $\dim(M/\mathrm{rad}(Q))$, and the nilpotent factor is an exterior algebra of rank $2^{\dim \mathrm{rad}(Q)}$. The total rank is $2^{\dim M}$, and the algebra is semisimple iff the form is non-degenerate. Over a general commutative ring, the decomposition requires the radical to be a direct summand.

The rank of $Cl(M, Q)$ over $R$ is $2^n$, where $n = \operatorname{rank} M$. A basis is given by the products of generators with strictly increasing indices.

The volume element $\omega = e_1 \cdots e_n$ satisfies $\omega^2 = (-1)^{n(n-1)/2} Q(e_1) \cdots Q(e_n) \cdot 1$. It is central when $n$ is odd, and central up to sign when $n$ is even. The center of the Clifford algebra is $R$ when $n$ is even and $R \oplus R\omega$ when $n$ is odd.

The algebra decomposes into an even part and an odd part. The even part is a subalgebra, and the odd part is a module over it. In finite dimensions, the even subalgebra is itself a Clifford algebra: $Cl^+(M, Q) \cong Cl(M', Q')$ with $\operatorname{rank} M' = \operatorname{rank} M - 1$. This reduction is the basis of the inductive computation of Clifford algebras.

The anti-involutions are reversion, Clifford conjugation, and the grade involution. Their behavior on each grade is determined by the grade alone, not by the ring.

The tensor product decomposition reduces the classification to the one-dimensional cases. In finite dimensions, the form can be diagonalized over a field, and the Clifford algebra decomposes as a graded tensor product of rank-one Clifford algebras. The even part of the tensor product is not the tensor product of the even parts: it also includes the tensor product of the odd parts.

Key differences from the field case.

Over a field, every module is free, so the Clifford algebra has a well-defined dimension. Over a general commutative ring, the module need not be free, so the Clifford algebra need not have a well-defined rank. We assume the module is free of finite rank.
Over a field, non-degeneracy is equivalent to the radical being zero. Over a general commutative ring, non-degeneracy requires the induced map $M \to M^*$ to be an isomorphism.
The polarization identity requires that 2 is invertible in $R$. Over a field of characteristic not equal to 2, this is automatic. Over a general commutative ring, it is an assumption.
The tensor product decomposition of a degenerate Clifford algebra requires the radical to be a direct summand. Over a field, this is automatic. Over a general commutative ring, it is an assumption.
The classification of Clifford algebras over a general commutative ring is much more subtle than over a field, and requires the theory of quadratic forms over rings.

The classification of the rank-one cases, and the structure of the Clifford algebras over specific rings, is not covered here.

Part VII: The Square Root of a Multivector in the Euclidean Algebra (Cl_{3,0})

28. The Problem and the Central Reduction

Fix the Euclidean Clifford algebra (Cl_{3,0}), with generators (e_1,e_2,e_3), (e_k^2=+1) for (k=1,2,3), and volume element (\omega=e_1e_2e_3). The volume element is central because (n=3) is odd (§15), and its square is (\omega^2=-1) (§15); it therefore supplies a central imaginary unit inside the algebra, and ({1,\omega}) spans a copy of the complex numbers in the centre (§16).

The algebra is eight-dimensional over (R=\mathbb{R}), so every element has a unique expression

$$ A = a + b\,\omega, $$

where (a) and (b) are paravectors of the underlying three-dimensional space, that is, sums of a scalar and a vector (§14). Because (\omega) is central with (\omega^2=-1), the product rule is

$$ (a+b\omega)(c+d\omega) = (ac-bd) + (ad+bc)\,\omega, $$

and therefore the square of (A) is

$$ A^2 = \left(a^2-b^2\right) + (ab+ba)\,\omega. $$

The square-root problem is: given (B\in Cl_{3,0}), find every (A) with (A^2=B). The reduction to the paravector decomposition is what makes the problem two-dimensional in appearance while remaining eight-dimensional in content, and it is the route by which the square roots of a quaternion or a complex quaternion are obtained once that algebra is identified with (Cl_{3,0}).

29. The Coupling Equations

Write the two paravectors in scalar–vector form,

$$ a = s + v, \qquad b = S + V, \qquad s,S\in\mathbb{R}, \qquad v = v_1e_1+v_2e_2+v_3e_3,\quad V = V_1e_1+V_2e_2+V_3e_3 . $$

A paravector squares as (a^2 = s^2 + 2sv + Q(v)), where (Q(v) = v_1^2+v_2^2+v_3^2) is the quadratic form, and a product of two paravectors is

$$ ab+ba = 2\left(sS + B(v,V)\right) + 2\left(sV+Sv\right), $$

where (B(v,V) = v_1V_1+v_2V_2+v_3V_3) is the polar form of (Q). Equating (A^2=(a^2-b^2)+(ab+ba)\omega) to a general element

$$ B = b_0 + b_1e_1+b_2e_2+b_3e_3 + b_{12}e_{12}+b_{13}e_{13}+b_{23}e_{23} + b_{123}\,\omega $$

gives four equations, one for each grade. The vector and bivector parts are linear in the unknowns and couple (v) to (V); the scalar and trivector parts are the scalar equations. To write them compactly, read the bivector coefficients through the duality (§Duality and the Hodge Star, §The Volume Element): put

$$ c_1 = b_{23}, \qquad c_2 = -b_{13}, \qquad c_3 = b_{12}, $$

so that the bivector part of (B) is (\omega\,c) with (c=c_1e_1+c_2e_2+c_3e_3). Then the graded equations are

$$ s\,v_k - S\,V_k = \tfrac12 b_k, \qquad S\,v_k + s\,V_k = \tfrac12 c_k \qquad (k=1,2,3), $$

together with

$$ s^2 + Q(v) - S^2 - Q(V) = b_0, \qquad 2sS + 2B(v,V) = b_{123}. $$

The first pair is a (2\times2) linear system for ((v_k,V_k)) whose determinant is (s^2+S^2); writing

$$ \sigma = s^2+S^2 $$

and solving gives the vectors explicitly in terms of the scalars and the data:

$$ v_k = \frac{s\,b_k + S\,c_k}{2\sigma}, \qquad V_k = \frac{s\,c_k - S\,b_k}{2\sigma}, \qquad (k=1,2,3), $$

valid whenever (\sigma\neq0). Substituting these into the two scalar equations and using the two combinations

$$ P = \sum_{k=1}^{3} b_kc_k = b_1b_{23}-b_2b_{13}+b_3b_{12}, \qquad Q = \sum_{k=1}^{3}\left(b_k^2-c_k^2\right) = b_1^2+b_2^2+b_3^2-b_{12}^2-b_{13}^2-b_{23}^2 $$

turns them into two equations for the pair ((\delta,\tau)) with

$$ \delta = s^2-S^2, \qquad \tau = 2sS, \qquad \sigma^2 = \delta^2+\tau^2 . $$

They are the two components of one complex equation.

Lemma. Let (\beta = b_0 - i\,b_{123}) and (\gamma = Q - 2iP), and let (z = \delta - i\tau). Then the two scalar equations are equivalent to

$$ 4z^2 - 4\beta z + \gamma = 0 . $$

Proof. With (\sigma = s^2+S^2) one has (s^2=( \sigma+\delta)/2) and (S^2=(\sigma-\delta)/2), and the identities

$$ Q(v)-Q(V) = \frac{\delta Q + 2\tau P}{4\sigma^2}, \qquad 2B(v,V) = \frac{2\delta P - \tau Q}{4\sigma^2} $$

follow from the solved (v_k,V_k) by expanding ((s b_k+S c_k)^2-(s c_k-S b_k)^2) and ((s b_k+S c_k)(s c_k-S b_k)). The two scalar equations therefore read

$$ \delta + \frac{\delta Q + 2\tau P}{4\sigma^2} = b_0, \qquad \tau + \frac{2\delta P - \tau Q}{4\sigma^2} = b_{123}. $$

Multiplying the first by (\delta) and the second by (\tau) and adding gives the real part, and multiplying the first by (\tau) and the second by (\delta) and subtracting gives the imaginary part, of the single equation in (\bar z = \delta+i\tau),

$$ 4|\bar z|^4 + \left(Q-2iP\right)\bar z^2 = 4\left(b_0-i b_{123}\right)|\bar z|^2 \bar z , $$

after using (\delta^2-\tau^2 = \mathrm{Re}\,\bar z^2), (2\delta\tau = \mathrm{Im}\,\bar z^2) and (|\bar z|^2=\sigma^2). Dividing by (\bar z^2|\bar z|^2) — legitimate because (\bar z\neq0) when (\sigma\neq0) — and putting (\gamma=Q-2iP), (\beta=b_0-ib_{123}), yields (4\bar z^2 - 4\beta\bar z + \gamma = 0). Taking the conjugate, and renaming, gives the stated quadratic in (z). (\square)

The discriminant of (4z^2-4\beta z+\gamma=0) is (16(\beta^2-\gamma)), and

$$ \beta^2 - \gamma = \left(b_0^2-b_{123}^2-Q\right) + 2i\left(P - b_0b_{123}\right) =: b_S + i\,b_I . $$

The real and imaginary parts (b_S), (b_I) are the two scalars that carry the existence of a root, and the modulus of (\beta^2-\gamma) is

$$ |\beta^2-\gamma|^2 = b_S^2 + b_I^2 =: \Delta . $$

30. The Isolated Roots

Theorem. A multivector (B) with (b_1,b_2,b_3,b_{12},b_{13},b_{23}) not all zero has an isolated square root in (Cl_{3,0}) for each complex number (z) satisfying (4z^2-4\beta z+\gamma=0) with (z\neq0), namely

$$ z = \frac{\beta \pm \sqrt{b_S + i\,b_I}}{2}, $$

(\sqrt{b_S+ib_I}) any square root of the complex number (b_S+ib_I). For each such (z), with (\delta=\mathrm{Re}\,z), (\tau=-\mathrm{Im}\,z), (\sigma=|z|), the two scalars are

$$ s = \pm\sqrt{\frac{\sigma+\delta}{2}}, \qquad S = \text{the choice of } \pm\sqrt{\frac{\sigma-\delta}{2}} \text{ with } 2sS=\tau, $$

and the vectors are (v_k=(s b_k+S c_k)/(2\sigma)), (V_k=(s c_k-S b_k)/(2\sigma)). The two signs of (s) give the two roots (\pm A), so that a nonzero (z) contributes a pair.

Proof. (\sigma=|z|=s^2+S^2>0) for (z\neq0), so the linear system is solvable and (v,V) are given by the formulas of §29. The scalars are recovered from (z) by (s^2=(\sigma+\delta)/2) and (S^2=(\sigma-\delta)/2), and the relative sign of (s) and (S) is fixed by (2sS=\tau); changing both signs gives (-A). The lemma of §29 shows that the quadratic is exactly the remaining pair of graded equations, so every solution of the quadratic yields a root and only those. (\square)

Because the quadratic is of degree two, the count of isolated roots is read off its discriminant:

if (b_S+ib_I\neq0), the two values (z=(\beta\pm\rho)/2), (\rho^2=b_S+ib_I), are distinct; a value with (|z|\neq0) contributes two roots, a value with (z=0) contributes none. The generic count is four roots, dropping to two when one of the two values of (z) vanishes;
if (b_S+ib_I=0), the two values coincide and there are two roots (or none, when the common value (\beta/2) also vanishes), so the count drops.

The vanishing of one value of (z) is exactly the case (b_S\ge0) and (b_I=0), which is the degeneracy that opens the continuous family of §31.

31. The Continuum of Roots

The reduction of §29 assumed (\sigma=s^2+S^2\neq0). The complementary case is genuine and produces a family rather than isolated points.

Theorem. If and only if (b_1=b_2=b_3=b_{12}=b_{13}=b_{23}=0), that is, if and only if (B=b_0+b_{123}\omega) is a complex scalar in the subalgebra spanned by (1) and (\omega), the roots of (B) include the whole four-parameter family

$$ A = v + V\omega, \qquad Q(v)-Q(V) = b_0, \qquad 2B(v,V) = b_{123}, $$

with (v) and (V) real vectors subject only to those two equations.

Proof. A root with (s=S=0) has the form (A=v+V\omega), and (A^2=(Q(v)-Q(V))+2B(v,V)\omega) by the product rule of §28. Conversely the graded equations with (s=S=0) force (b_k=c_k=0) for every (k). The two displayed equations are two affine equations in the six components of ((v,V)), so their solution set has real dimension four; it is nonempty over (\mathbb{R}) because the pair ((Q(v),B(v,V))) takes all prescribed values. (\square)

The family sits in the four cases where the data are a complex scalar. When (b_0b_{123}\neq0) the two values of (z) are distinct and nonzero, so the four isolated roots of §30 coexist with the family; when one value of (z) vanishes the family is present and two isolated roots remain; when both vanish the family is the whole root set. The root of (0) is the extreme instance: (b_0=b_{123}=0), and the roots of (0) are the nonzero elements of the family, the null elements (v+V\omega) with (Q(v)=Q(V)) and (B(v,V)=0) — the isotropic cone of the form.

32. The Classification

Collecting §30 and §31, every (B\in Cl_{3,0}) falls into exactly one of the following cases.

No root. (b_S+ib_I=0), (\beta=0), and (B) not a complex scalar: for instance (B=e_3+e_{23}). Then the quadratic has only the value (z=0), which the linear system cannot support. A necessary condition is (b_0=b_{123}=0) and (b_S=b_I=0).
Two roots. (b_S+ib_I=0) and (\beta\neq0), or one value of (z) vanishes: the quadratic supplies a single nonzero complex number (z), hence the pair (\pm A).
Four roots. (b_S+ib_I\neq0) and both values of (z) nonzero: two pairs (\pm A). This is the generic case.
A continuum, alone or alongside isolated roots. (B=b_0+b_{123}\omega): the four-parameter family of §31, together with the isolated roots supplied by the values of (z).

The three central values of the algebra illustrate the trichotomy: (B=e_0) has the two roots (\pm e_0) and a family; (B=-e_0) has the two roots (\pm\omega) and a family; (B=0) has the family alone, its nonzero members the null elements.

33. The Reading of the Data

The two data (b_0) and (b_{123}) have a direct reading in terms of the unknown root (A=v+V\omega) when it is a complex scalar, that is, on the continuum of §31: (b_0 = Q(v)-Q(V)) is the difference of the squared lengths of the two vectors, and (b_{123}=2B(v,V)) controls their mutual angle. So the scalar coefficient fixes a pair of concentric spheres on which (v) and (V) must lie, and the pseudoscalar coefficient fixes the angle between them; this is the geometric reading of the square-root problem given by Acus and Dargys, whose Clifford-layer algorithm is the closed form of §30–§31 written in the variables (b_S,b_I,\Delta).

For a general element the same two numbers are the real and imaginary parts of the complex number (\beta^2-\gamma) of §29, and the existence of roots is decided by its vanishing: (b_S+ib_I=0) is the boundary between the four roots and the two, and (\beta=0) on that boundary is the frontier with the rootless case.

Summary

The square-root problem in (Cl_{3,0}) reduces by the central volume element (\omega=e_1e_2e_3) ((\omega^2=-1)) to the paravector decomposition (A=a+b\omega=s+v+(S+V)\omega), with (A^2=(a^2-b^2)+(ab+ba)\omega).

The graded equations are the linear pair (s v_k-S V_k=b_k/2), (S v_k+s V_k=c_k/2) with (c=(b_{23},-b_{13},b_{12})), which solve for the vectors in terms of the scalars, and the two scalar equations, which combine into the single complex quadratic (4z^2-4\beta z+\gamma=0) for (z=\delta-i\tau), (\delta=s^2-S^2), (\tau=2sS), with (\beta=b_0-ib_{123}), (\gamma=Q-2iP) and (P=\sum b_kc_k), (Q=\sum(b_k^2-c_k^2)).

The existence scalars are the real and imaginary parts of (\beta^2-\gamma=b_S+ib_I), with (b_S=b_0^2-b_{123}^2-Q) and (b_I=2(P-b_0b_{123})); their squared modulus is (\Delta=b_S^2+b_I^2).

The isolated roots come from the two values (z=(\beta\pm\sqrt{b_S+ib_I})/2). A nonzero value gives a pair (\pm A) through (s=\pm\sqrt{(\sigma+\delta)/2}), (S) signed by (2sS=\tau), (\sigma=|z|), and the vector formulas. The generic count is four roots, falling to two when one value of (z) vanishes or when the two values coincide, and to none when both vanish at (\beta=0).

The continuum occurs exactly when (B=b_0+b_{123}\omega) is a complex scalar: the roots then include the four-parameter family (v+V\omega) with (Q(v)-Q(V)=b_0), (2B(v,V)=b_{123}), alongside the isolated roots. The roots of (0) are the null elements of this family.

The reading of the two data is (b_0=Q(v)-Q(V)) and (b_{123}=2B(v,V)) on the continuum: a difference of squared lengths and an angle. The closed form is due to Acus and Dargys; the quadratic of §29 is the form in which this article states and proves it.

Clifford Algebras in Finite Dimensions

Introduction

Part I: The Definition and Universal Property

1. The Definition

2. The Universal Property

3. The Fundamental Relation

Part II: The Radical and the Reduction to the Non-Degenerate Case

4. The Radical

5. The Reduction

6. The Decomposition of a Degenerate Clifford Algebra

7. The Non-Degenerate Factor

8. The Nilpotent Factor

9. The Rank of a Degenerate Clifford Algebra

10. The Grading and the Filtration in the Degenerate Case

11. The Lie Algebra Structure in the Degenerate Case

Part III: Basis and Rank in the Non-Degenerate Case

12. Basis

13. Rank

14. $k$-Vectors and Multivectors

15. The Volume Element

16. The Center

Part IV: The Even Subalgebra

17. The Decomposition

18. Why the Even Part Is a Subalgebra

19. The Even Subalgebra as a Clifford Algebra

Part V: The Anti-Involutions

20. Reversion

21. Clifford Conjugation

22. The Grade Involution

23. The Behavior on Each Grade

Part VI: The Tensor Product Decomposition

24. The Decomposition

25. Why This Matters

26. The Orthogonal Direct Sum and the Signature

27. The Graded Tensor Product and the Parity

Summary

Part VII: The Square Root of a Multivector in the Euclidean Algebra (Cl_{3,0})

28. The Problem and the Central Reduction

29. The Coupling Equations

30. The Isolated Roots

31. The Continuum of Roots

32. The Classification

33. The Reading of the Data

Summary

Further Reading