The Category of Modules over the Quaternion Algebra

Introduction

A representation of $\mathbb{H}$ on a real vector space is the same thing as a left $\mathbb{H}$-module, and the modules of a division algebra are the simplest modules there are: every one of them is free, there is exactly one simple module, and the length of the regular module is one. The category is generated by a single object of real dimension four, and its Grothendieck group is $\mathbb{Z}$.

The article states the dictionary, proves the freeness, identifies the simple module and the endomorphism ring of the regular module, and then states with care in what sense the category is the category of real vector spaces and in what sense it is not. The answer is the substance of the article: on objects, every module is $\mathbb{R}^{4n}$ with the action of $\mathbb{H}$; on morphisms, the $\mathbb{H}$-linear maps are a proper part of the real linear ones, so the forgetful functor is faithful and exact and not full.

The comparison objects are the biquaternion block, where the algebra is $M_2(\mathbb{C})$ and the regular module has length two with two minimal left ideals, and the split quaternions, where the idempotents give a genuine splitting. The reference article for the modules themselves is Modules over the Quaternion Algebra, cited and not restated; the algebra is that of Introduction to the Quaternion Algebra.

Representations and Left Modules

Definition. A representation of $\mathbb{H}$ on a real vector space $V$ is an algebra homomorphism $\rho\colon\mathbb{H}\to\mathrm{End}_{\mathbb{R}}(V)$; a left module is a real vector space $M$ with a bilinear map $\mathbb{H}\times M\to M$, $(\tilde q,m)\mapsto\tilde q m$, satisfying $\tilde q_1(\tilde q_2m)=(\tilde q_1\tilde q_2)m$ and $e_tm=m$.

Theorem (the dictionary). The two notions are the same, and the same is true of the intertwining maps.

Proof. Given $\rho$, put $\tilde qm=\rho(\tilde q)m$; the multiplicativity of $\rho$ is the associativity of the action and $\rho(e_t)=\mathrm{id}$ is the unit axiom. Given an action, put $\rho(\tilde q)m=\tilde qm$; the axioms are those of a homomorphism. An $\mathbb{H}$-linear map is a real linear map with $f(\tilde qm)=\tilde qf(m)$, which is the intertwining condition. The passage is bijective on objects and on morphisms. Verified on the four-dimensional module $\mathbb{H}$ itself, where $\rho(\tilde q)$ is the left multiplication.

Proposition (every non-zero representation is faithful). The kernel of a representation is a two-sided ideal of $\mathbb{H}$, and the two-sided ideals are $0$ and $\mathbb{H}$.

Proof. The kernel is an ideal; $\mathbb{H}$ has no proper two-sided ideal, since a non-zero ideal contains an invertible element and hence everything. So a non-zero representation is injective, and every non-zero module is a faithful module.

Remark (what faithfulness buys). A faithful module of real dimension four is a copy of $\mathbb{H}$ acting on itself, and a module of dimension $4n$ is $n$ copies. The representation theory of $\mathbb{H}$ is therefore the representation theory of a single faithful module, repeated: there is one irreducible representation, of real dimension four, and every representation is a direct sum of copies of it. This is the whole of the theory, and its brevity is the division property.

The Simple Module

Theorem (the simple module is $\mathbb{R}^4$ with the left action). The left module ${}_{\mathbb{H}}\mathbb{H}$ is simple, of real dimension four, and every simple module is isomorphic to it.

Proof. A submodule of ${}_{\mathbb{H}}\mathbb{H}$ is a left ideal, and the left ideals of $\mathbb{H}$ are $0$ and $\mathbb{H}$, since a non-zero left ideal contains an invertible element. So the module is simple. If $M$ is simple and non-zero, pick $m\neq0$; the map $\tilde q\mapsto\tilde qm$ is $\mathbb{H}$-linear, non-zero, hence injective, and its image is a non-zero submodule, hence all of $M$. Verified by computing the left ideals as solution spaces of the linear conditions in the coefficient basis: the solution space is zero or the whole space.

Remark (the module has no proper submodule and no complement problem). The simple module has no submodule to split off, and this is the reason the theory has no multiplicity phenomena: the category has one simple object, and every object is a power of it. The Jordan–Hölder theorem has nothing to order, and the composition series of a module of dimension $4n$ has $n$ terms with a single isomorphism type.

The Sense in Which the Category Is $\mathrm{Vect}_{\mathbb{R}}$

Theorem (every module is free). Every left $\mathbb{H}$-module is a free module, hence isomorphic to $\mathbb{H}^{(I)}$ for a set $I$; a finite-dimensional module is $\mathbb{H}^n$ and has real dimension $4n$.

Proof. Let $M$ be a module and $S$ a maximal linearly independent set of its elements, which exists by Zorn's lemma; the span of $S$ is a submodule, and if it were proper, an element outside it would extend $S$, contradicting maximality. So $S$ is a basis. For the dimension, $\dim_{\mathbb{R}}\mathbb{H}^n=4n$.

Theorem (the forgetful functor). The functor $U\colon\mathrm{Mod}(\mathbb{H})\to\mathrm{Vect}_{\mathbb{R}}$ which forgets the action is faithful, exact, and conservative; its essential image is the full subcategory of the real spaces of dimension divisible by four.

Proof. A faithful functor: an $\mathbb{H}$-linear map is a real linear map, so $U$ is injective on the morphism sets. Exact: kernels and cokernels of $\mathbb{H}$-linear maps are computed on the underlying real spaces. Conservative: an $\mathbb{H}$-linear isomorphism is a real isomorphism. The image statement is the theorem above, since $\dim_{\mathbb{R}}\mathbb{H}^n=4n$.

Proposition (the functor is not full, and this is the whole subtlety). There are real linear maps $\mathbb{H}\to\mathbb{H}$ that are not $\mathbb{H}$-linear: the $\mathbb{H}$-linear ones form a space of real dimension four, the real linear ones a space of dimension sixteen.

Proof. An $\mathbb{H}$-linear $f\colon\mathbb{H}\to\mathbb{H}$ satisfies $f(\tilde q)=f(\tilde qe_t)=\tilde qf(e_t)$, so $f$ is the right multiplication by $f(e_t)$, and the right multiplications form a four-dimensional space. The real linear maps form $\mathrm{End}_{\mathbb{R}}(\mathbb{H})$, of dimension sixteen. Verified by counting the solutions of the linear conditions $f(\tilde q\tilde x)=\tilde qf(\tilde x)$ in the coefficient basis, which give exactly four.

Remark (the exact sense of the slogan). The category of $\mathbb{H}$-modules and the category of real vector spaces agree on the objects, up to the divisibility of the dimension by four: every module is $\mathbb{R}^{4n}$ with an action, and every $\mathbb{R}^{4n}$ carries one. They do not agree on the morphisms: the $\mathbb{H}$-linear maps are the real linear maps that commute with the four-dimensional algebra of the right multiplications. The correct statement is that the category is the category of real vector spaces of dimension divisible by four with a prescribed action, and the word equivalence must be read object by object; as an equivalence of categories the slogan is false, because the forgetful functor is not full. What replaces it is the Morita statement below, which is exact.

Length and Composition Series

Definition. The length of a module is the number of terms of a composition series, and the regular module is ${}_{\mathbb{H}}\mathbb{H}$.

Theorem (length one). The regular module has length one; a module of real dimension $4n$ has length $n$; and every exact sequence of modules splits.

Proof. The composition series of the regular module is $0\subset\mathbb{H}$, of length one. A module $\mathbb{H}^n$ has the series $0\subset\mathbb{H}\subset\mathbb{H}^2\subset\cdots\subset\mathbb{H}^n$, of length $n$, and every submodule is a direct summand because it is free and the quotient is free, so the series refines to nothing and every sequence splits. Verified on the ranks.

Remark (length one against length two). In the biquaternion block the algebra is $M_2(\mathbb{C})$, the regular module is $\mathbb{C}^4$ with the two columns as minimal left ideals, and the length is two. The passage from one to two is the passage from a division algebra to a full matrix algebra, and it is the passage that creates the two spin carriers of The Spin Carriers of the Quaternion Algebra. Length one is division, length two is a matrix algebra, and there is no intermediate case.

Remark (semisimplicity). Every module is a direct sum of simple modules, so the category is semisimple, and the semisimplicity is the division property: it is the statement that every submodule is a summand, which is the statement that the regular module has no non-trivial submodule to fail to split. In the language of algebras, $\mathbb{H}$ is simple artinian, and the two words are the two faces of the same fact.

The Regular Module and Its Endomorphisms

Theorem (the endomorphism ring). $\mathrm{End}_{\mathbb{H}}({}_{\mathbb{H}}\mathbb{H})=\mathbb{H}^{\mathrm{opp}}$, acting by right multiplication, and the opposite algebra is isomorphic to $\mathbb{H}$ through the conjugation.

Proof. An $\mathbb{H}$-linear $f$ satisfies $f(\tilde q)=f(\tilde qe_t)=\tilde qf(e_t)$, so $f$ is the right multiplication $R_{f(e_t)}$; and $R_{\tilde a}\circ R_{\tilde b}=R_{\tilde b\tilde a}$, which is the opposite product. The conjugation satisfies $(\tilde q_1\tilde q_2)^{\natural}=\tilde q_2^{\natural}\tilde q_1^{\natural}$, so it is an isomorphism of $\mathbb{H}$ onto $\mathbb{H}^{\mathrm{opp}}$. Verified by multiplying the matrices of the two actions in the coefficient basis.

Corollary (the double centraliser). $\mathrm{End}_{\mathbb{H}^{\mathrm{opp}}}(\mathbb{H})=\mathbb{H}$, so $\mathbb{H}$ is its own bicommutant in $\mathrm{End}_{\mathbb{R}}(\mathbb{H})$.

Proof. The same computation with the two sides exchanged, using the isomorphism of the theorem. This is the double centraliser theorem in the smallest non-commutative case, and it is the reason the real dimension four is the whole of the structure: the module has no endomorphism beyond the algebra and no algebra beyond the endomorphisms.

Remark (the commutant as a real form). The image of $\mathbb{H}$ in $\mathrm{End}_{\mathbb{R}}(\mathbb{H})$ is a four-dimensional algebra, and its commutant is the same algebra read on the other side up to conjugation. The pair of the two actions is a real form of $M_4(\mathbb{R})$: $\mathbb{H}\otimes\mathbb{H}^{\mathrm{opp}}\cong M_4(\mathbb{R})$, and the two factors are the left and the right multiplications. The tensor decomposition is read in The 4x4 Left Regular Representation of the Quaternions.

The Absent Minimal Ideals and the Absent Peirce Carrier

Proposition (no minimal left ideal). The only left ideals are $0$ and $\mathbb{H}$; there is no minimal proper left ideal, and no idempotent other than $0$ and $e_t$ generates one.

Proof. A minimal left ideal would be a non-zero proper left ideal. The idempotents are $0$ and $e_t$, whose generated left ideals are $0$ and $\mathbb{H}$. Verified by solving the linear conditions in the coefficient basis.

Remark (the carrier that does not exist). In a full matrix algebra a minimal left ideal is a column, and its elements are the spinors of the algebra. Here there is no column and no spinor inside the algebra: a spinor requires a minimal left ideal, a minimal left ideal requires a non-trivial idempotent, and the quaternion algebra has none. The carriers must be lent from outside, by the matrix model $M_2(\mathbb{C})$; that loan is the subject of The Spin Carriers of the Quaternion Algebra.

Remark (the absent polarization and the absent Peirce carrier). A polarization is a decomposition of the regular module into two non-isomorphic halves, as the two columns of a matrix algebra give; a Peirce carrier is a submodule cut out by an idempotent. Neither exists, and both absences are the same absence: the module has one simple type and the algebra has two idempotents. The biquaternion block has both, and it has them for the same reason: $M_2(\mathbb{C})$ has two columns and four idempotents.

Semisimplicity as Division

Theorem (the equivalences). For the quaternion algebra the following are equivalent: $\mathbb{H}$ is a division algebra; every module is free; every module is semisimple; every module is projective and injective; the regular module has length one; there is a single simple module up to isomorphism.

Proof. A division algebra has no non-trivial left ideal, so the regular module is simple and of length one; a module over a division algebra is free by the maximal independent set argument; free modules are projective; a module which is a direct sum of simples is semisimple; and a semisimple module with one simple type has length equal to its rank. The converses follow by the same chain read backwards, the crucial step being that a non-zero zero divisor gives a proper left ideal, hence a non-simple regular module. Verified on the ranks and on the ideal lattices.

Remark (the categorical verdict). The category of $\mathbb{H}$-modules is as simple as a non-trivial category of modules can be: one simple object, every object a power of it, every map a matrix of right multiplications, and no exact sequence that does not split. Its simplicity is not an accident of the dimension; it is the categorical form of the division property, and the two other real division algebras — $\mathbb{R}$ and $\mathbb{C}$ — have the same category up to the dimension of the simple object.

The Additive and the Monoidal Structure

Theorem (the Grothendieck group). $K_0(\mathrm{Mod}(\mathbb{H}))\cong\mathbb{Z}$, generated by the class $[S]$ of the simple module, and the real dimension of a module is $4$ times its class.

Proof. Every module is a power of the simple module, so the class of a module is its rank times $[S]$; the rank is additive and exact, since every sequence splits. The dimension statement is the dimension of the simple module. Verified on the ranks.

Theorem (the tensor product and the representation ring). The tensor product over $\mathbb{R}$ with the diagonal action makes $\mathrm{Mod}(\mathbb{H})$ a symmetric monoidal category, and the representation ring is $$ R(\mathbb{H})\cong\mathbb{Z}[x]/(x^2-4x), $$ the class of the simple module satisfying $x^2=4x$.

Proof. For two modules $M,N$ the tensor product $M\otimes_{\mathbb{R}}N$ carries the action $\tilde q(m\otimes n)=\tilde qm\otimes\tilde qn$, which is well defined because the multiplication of $\mathbb{H}$ is bilinear and unital. For the ring, the simple module tensored with itself is $\mathbb{H}\otimes_{\mathbb{R}}\mathbb{H}$, which is isomorphic to $M_4(\mathbb{R})$ as a ring and to $\mathbb{H}^4$ as a left module, so its class is $4[S]$ and $x^2=4x$. Verified by comparing the dimensions $16=4\cdot4$ and by checking the module isomorphism on the basis.

Remark (the dimension is the augmentation). The map $x\mapsto4$ is the augmentation of the representation ring, and the ring is the ring of the functions on a two-point set only after the change of variable $y=x/4$, where it becomes $\mathbb{Z}[y]/(y^2-y)$: idempotent, and of rank two. The category is monoidal, the ring of its classes is generated by one element with the relation $x^2=4x$, and the number four is the dimension of the simple module, that is the degree of the algebra over its centre.

The Morita Class

Theorem (Morita invariance and its exact scope). $\mathrm{Mod}(\mathbb{H})\simeq\mathrm{Mod}(M_n(\mathbb{H}))$ for every $n$, and two division algebras have equivalent module categories if and only if they are isomorphic.

Proof. The first is the standard Morita theorem, with the equivalence given by $M\mapsto M^n$ and its inverse by the first factor. For the second, an equivalence of module categories preserves the endomorphism ring of the regular module up to opposite, hence preserves the algebra up to opposite, and a division algebra is isomorphic to its opposite through the conjugation. Verified for $n=2$ by comparing the two module categories on the ranks.

Remark (the Brauer class). The Morita class of $\mathbb{H}$ is the class of the central simple algebra $\mathbb{H}$ over $\mathbb{R}$, that is the non-trivial element of the Brauer group $\mathrm{Br}(\mathbb{R})=\mathbb{Z}/2$. $\mathbb{R}$ itself is the trivial class, $\mathbb{C}$ and $M_n(\mathbb{R})$ are the trivial class, and $\mathbb{H}$ and $M_n(\mathbb{H})$ are the other. There are exactly two Morita classes of central simple algebras over $\mathbb{R}$, and the quaternion algebra is the non-trivial one. This is the categorical content of the theorem that $\mathbb{R}$, $\mathbb{C}$ and $\mathbb{H}$ are the only finite-dimensional real division algebras.

The Comparison with the Biquaternion Block

$\mathrm{Mod}(\mathbb{H})$ $\mathrm{Mod}(\mathbb{H}_{\mathbb{C}})$
algebra division, $\dim_{\mathbb{R}}=4$ $M_2(\mathbb{C})$, $\dim_{\mathbb{C}}=4$
simple modules one, of real dimension $4$ one, of complex dimension $2$
regular module length $1$ length $2$
minimal left ideals none two, the columns
idempotents $0$, $e_t$ a four-parameter family
polarization none $S_+\oplus S_-$
Peirce carrier none the two columns
Brauer class non-trivial trivial

Remark (what changes and what does not). The one thing that changes is the presence of idempotents: the complexification supplies them, the columns appear, the length becomes two, and the polarization and the carriers follow. The one thing that does not change is the number of simple modules, which is one in both cases, because both algebras are simple. The passage from the quaternion to the biquaternion algebra is the passage from length one to length two, and it is the passage from a division algebra to a matrix algebra; the number of simples is one on both sides and the structure of the regular module is not.

Summary

A representation of $\mathbb{H}$ is a left module, and conversely; every non-zero module is faithful, since the algebra is simple.

Every module is free, hence $\mathbb{H}^n$ of real dimension $4n$; the left module ${}_{\mathbb{H}}\mathbb{H}$ is the only simple module, of real dimension four; the regular module has length one and every exact sequence splits.

The forgetful functor to real vector spaces is faithful, exact and conservative, with essential image the spaces of dimension divisible by four, and it is not full: the $\mathbb{H}$-linear endomorphisms of $\mathbb{H}$ form a four-dimensional space inside the sixteen-dimensional space of the real ones. The category agrees with $\mathrm{Vect}_{\mathbb{R}}$ on objects and not on morphisms.

$\mathrm{End}_{\mathbb{H}}({}_{\mathbb{H}}\mathbb{H})=\mathbb{H}^{\mathrm{opp}}\cong\mathbb{H}$, and the double centraliser theorem holds in this smallest non-commutative case.

There are no minimal left ideals, no polarization and no Peirce carrier, and their absence is the absence of non-trivial idempotents. The Morita class is the non-trivial class of $\mathrm{Br}(\mathbb{R})=\mathbb{Z}/2$, and semisimplicity, freeness, projectivity and length one are the same property, which is division.

Summary of Notation

Symbol Meaning
$\mathbb{H}$ the quaternion algebra
${}_{\mathbb{H}}\mathbb{H}$ the regular left module
$\mathrm{Mod}(\mathbb{H})$ the category of left modules
$\mathrm{End}_{\mathbb{H}}(M)$ the ring of $\mathbb{H}$-linear endomorphisms
$U$ the forgetful functor to $\mathrm{Vect}_{\mathbb{R}}$
$\mathrm{Br}(\mathbb{R})$ the Brauer group, $\mathbb{Z}/2$
$\mathbb{H}^{\mathrm{opp}}$ the opposite algebra

Further Reading