$M_2(R)$ Triangular Representation : Operators

Introduction

The same operator layer, on the only local algebra of the block. The carrier is $\mathbb{C}_{\mathrm{dln},R}$, of dimension two, and the ambient is

$$ \operatorname{End}_R(\mathbb{C}_{\mathrm{dln},R})\cong M_2(R), $$

of dimension four. The one-sided families coincide, the two-sided operator is the multiplication by the product, and the ambient is twice the size of the family. So the operator layer of this category has the same shape as the complex one, and it is not the same layer, because the algebra has a nilpotent.

The four objects of the general theory, and what becomes of them.

object general case here
the two one-sided families two copies, meeting in the scalars one copy, containing a square-zero member
the cone of products the image of the pair, cut out by determinantal equations a linear subspace of dimension two
the length a rank in $\{0,\dots,n\}$ in $\{0,1\}$, and only inside the span
the two ranks operator rank against coefficient rank $2,1,0$ against $1$, and the display in $\{0,1,2,4\}$

Three of the four collapse as they did in the complex category, and the fourth does not: the rank of the Kronecker display is the product of two ranks that can be $1$ as well as $2$ or $0$, the ladder $\{0,1,2,4\}$ already recorded for the split category. What differs from that category is the shape of the singular stratum, one line of nilpotents here against two lines of projectors there, and that is what the article is about.

The two facts the model article cannot state. The carrier has one invariant line, $\mathrm M$, so the one-sided operator drops rank there and its kernel is its own image: the family is $\{q\,\mathrm{id}+q'L_\varepsilon\}$ with $L_\varepsilon^{2}=0$, the smallest faithful operator algebra that is not semisimple. And the two-sided parametrisation attains the zero operator at a pair of nonzero parameters,

$$ L_{\varepsilon}R_{\varepsilon}(X)=\varepsilon X\varepsilon=\varepsilon^{2}X=0, $$

whereas in the complex and split categories the zero operator was in the family only through a zero parameter. Both facts say the same thing: the algebra is not reduced, and neither is its operator layer.

The one operator the family cannot produce. The derivation

$$ \partial_\varepsilon:\mathbb{C}_{\mathrm{dln},R}\longrightarrow\mathbb{C}_{\mathrm{dln},R},\qquad \partial_\varepsilon(q+\varepsilon q')=q'\varepsilon, $$

is $R$-linear and square-zero, and it is not a multiplication: $L_A(1)=A$ always lies in the subalgebra generated by $1$, while $\partial_\varepsilon(1)=0$ and $\partial_\varepsilon(\varepsilon)=\varepsilon$. So the dual numbers need one operator for their own calculus and do not contain it. The gap is one-dimensional, and the section on what the pair buys identifies it, together with the two other operators the ambient contains and the family does not.

Boundaries. The model, the frame, the trace, the determinant, the projectors, the adjugate and the subspaces are $M_2(R)$ Triangular Representation. The algebra and its product are Dual-Numbers Algebra; the norm, the unit criterion and the inverse are Dual-Numbers Norm and Invertibility; the spectrum, the minimal polynomial and the Jordan form are Spectral Regularity and Degeneracy in the Dual-Number Algebra; the representation theory, the classification and the non-split extension are Dual-Numbers Linear Representations; the group of invertibles, the shears and the parabolic angle are Shears and Parabolic Rotations and Dual-Numbers Exponential and Lie Group Structure; the automorphisms and the derivations are Dual-Numbers Automorphisms and Derivations; the polar decomposition is Dual-Numbers Polar Decomposition.

Conventions. The carrier has the natural basis $(1,\varepsilon)$, and the model is $\mathsf{M}_2(Q)=qI_2+q'E$ with $E=E_{12}$, the matrix of $L_Q$ in the reversed basis $(\varepsilon,1)$. The two-sided operator is $L_{Q}R_{P}(X)=QXP$. The coefficient array of an operator is its array of coordinates in the basis $\{T_1,T_\varepsilon\}=\{I_2,E\}$ of the family. The Frobenius form is $\langle X,Y\rangle_F=\operatorname{tr}(X^{\mathsf T}Y)$ and the plain form is $\langle X,Y\rangle_{\operatorname{tr}}=\operatorname{tr}(XY)$. The swap is $S$, the diagonal involution is $J=\operatorname{diag}(1,-1)$, and the residue is $\pi$.

The One-Sided Operators

The family. The left multiplications are the model, and the right multiplications are the same set:

$$ L(\mathbb{C}_{\mathrm{dln},R})=\{\mathsf{M}_2(Q)\}=\{qI_2+q'E\},\qquad \dim_R=2,\qquad L=R,\qquad L+R=L . $$

The generator is square-zero. The element $\varepsilon$ acts by $E$, and $E^{2}=0$, so

$$ L_\varepsilon^{2}=L_{\varepsilon^{2}}=0,\qquad L_\varepsilon(1)=\varepsilon,\qquad L_\varepsilon(\varepsilon)=0 . $$

The family is therefore the two-dimensional space spanned by the identity and a square-zero operator, and it is the smallest faithful operator algebra that is not semisimple: a one-dimensional module would kill $\varepsilon$, so two is the least possible dimension, and the algebra realises it. In the complex and split categories the same two dimensions carried a semisimple family, and the difference at equal size is the whole of the present section.

The rank ladder. For $Q=q+\varepsilon q'$,

$$ \ker L_Q=\{X:QX=0\},\qquad \operatorname{im}L_Q=Q\mathbb{C}_{\mathrm{dln},R}, $$

and the three strata are

$Q$ $\ker L_Q$ $\operatorname{im}L_Q$ $\operatorname{rank}L_Q$
$q\neq0$ $0$ $\mathbb{C}_{\mathrm{dln},R}$ $2$
$q=0$, $q'\neq0$ $\mathrm{M}$ $\mathrm{M}$ $1$
$Q=0$ $\mathbb{C}_{\mathrm{dln},R}$ $0$ $0$

The reading. On the complement of the maximal ideal the operator is a bijection, as in a field. On the maximal ideal minus the origin the kernel and the image are the same line, and this is the operator form of the statement that $L_\varepsilon$ is one Jordan block of size two: the single line is killed and the single line is hit, and the two are the same line because the block has one eigenvector. So the defect of the operator is visible in the matrix model as the identity $\ker L_{q'\varepsilon}=\operatorname{im}L_{q'\varepsilon}=\mathrm{M}$, a statement which no basis can remove.

The image is a line of operators. For $Q=q'\varepsilon$ the operator is $q'L_\varepsilon$, so the whole rank-one stratum is the single line $\mathbb{R}L_\varepsilon$ minus its origin: every singular member of the family is a scalar multiple of the same nilpotent. The family therefore has exactly one interesting point, the generator, and the rest of the line is its rescalings; in the split category the singular members filled two lines, one for each null direction, and in the complex category there were none.

The plain adjoint. For the plain form, $\langle L_QX,Y\rangle_{\operatorname{tr}}=\operatorname{tr}(QXY)=\operatorname{tr}(XYQ)=\langle X,L_QY\rangle_{\operatorname{tr}}$, so the operator is self-adjoint, and the reason is the commutativity of the algebra. The family is a commutative algebra of self-adjoint operators, and the whole involution theory of the complex category repeats with no change.

The Frobenius adjoint leaves the family. For the Frobenius form, $\langle L_QX,Y\rangle_F=\operatorname{tr}((QX)^{\mathsf T}Y)=\operatorname{tr}(X^{\mathsf T}Q^{\mathsf T}Y)=\langle X,\ \mathsf{M}_2(Q)^{\mathsf T}Y\rangle_F$, so the adjoint of $L_Q$ is the operator of left multiplication by the transposed matrix,

$$ L_Q^{*}:\ X\longmapsto \mathsf{M}_2(Q)^{\mathsf T}X=\begin{pmatrix}q&0\\ q'&q\end{pmatrix}X . $$

That matrix is the lower triangular Toeplitz matrix of the same element, and the section on the adjugate of the model article showed that it is not in the image. So

$$ L_Q^{*}=S\,L_Q\,S,\qquad S=\begin{pmatrix}0&1\\1&0\end{pmatrix},\qquad L_Q^{*}\notin L(\mathbb{C}_{\mathrm{dln},R})\ \ (q'\neq0), $$

and the family is carried by the adjoint onto its conjugate by the swap, which is the family of the other convention.

The reading. The model is a faithful representation of the algebra and it is not a $*$-representation of the involutive algebra, because the adjoint of a member is not a member. The failure belongs to the set and not to the basis: the image is an algebra of upper triangular matrices and its adjoint is the algebra of lower triangular ones, and the two are conjugate, so no choice of basis produces a family closed under the adjoint. The dual conjugation is an automorphism of the algebra and not an anti-automorphism, so a $*$-structure would have to be realised by an anti-linear map, and there is none in a real model. Compare the quaternion category, where the conjugation was an anti-automorphism and the model was a $*$-representation; here the algebra has an automorphism where the others had an involution, and the operator layer records the difference.

The Two-Sided Operator

Definition and collapse. For a pair $(Q,P)$,

$$ L_{Q}R_{P}(X)=QXP=L_{QP}=\mathsf{M}_2(QP), $$

because the algebra is commutative. The pair is not a pair; it is the product.

The zero operator is a product. The parametrisation is not injective even after the rescaling is quotiented out, because a product of two nonzero elements can vanish:

$$ L_{\varepsilon}R_{\varepsilon}=0,\qquad L_{\varepsilon}R_{1}=L_\varepsilon\neq0,\qquad L_{1}R_{\varepsilon}=L_\varepsilon . $$

So the fibre over the zero operator contains the pair $(\varepsilon,\varepsilon)$, and the zero operator has a description with both parameters nonzero: it is the product of two nilpotents, not merely a limit of such products. This is the operator form of the fact that $\mathrm{M}$ is a nilpotent ideal, and it has no analogue in the complex category, where the fibre over zero was the two axis pairs, and none in the quaternion category, where it was a single point.

The invariants. With $c=QP=qp+\varepsilon(qp'+q'p)$,

$$ \operatorname{rank}L_{Q}R_{P}=\operatorname{rank}\mathsf{M}_2(c),\qquad \operatorname{tr}L_{Q}R_{P}=2qp,\qquad \det L_{Q}R_{P}=(qp)^{2},\qquad \langle L_{Q}R_{P}\rangle_*=2(qp)^{2}+(qp'+q'p)^{2}. $$

Proof. $L_{Q}R_{P}$ is the matrix $\mathsf{M}_2(QP)$, so the four quantities are those of the model article read at the product.

The trace, and what it is not. The trace of the operator is $2qp$ and not $(2q)(2p)$: the trace of a product is not the product of the traces, and the two numbers belong to different operators. The trace of the display is $4qp$, and the difference between $2qp$ and $4qp$ is the difference between the two operators and not a discrepancy.

The inner automorphisms are trivial. Since the algebra is commutative, $L_{Q}R_{Q^{-1}}(X)=QXQ^{-1}=X$ for every invertible $Q$: the inner automorphism group of the algebra is trivial, and the two-sided family with equal inverse parameters is the identity alone. The automorphisms of the algebra are nevertheless a one-parameter group, and they are reached by conjugation in the ambient and not by multiplication in the algebra; the section on the pair as a semigroup computes them.

The two-sided operator, and the ray. The two-sided operator with the conjugate on the right is

$$ \mathrm{Sd}_Q(X)=QX\bar Q=Q\bar Q\,X=q^{2}X, $$

a scalar multiple of the identity, with scalar $q^{2}$. So the two-sided family is

$$ \{\mathrm{Sd}_Q:Q\in\mathbb{C}_{\mathrm{dln},\mathbb{R}}\}=\{q^{2}\,\mathrm{id}:q\in\mathbb{R}\}=[0,\infty)\cdot\mathrm{id}, $$

a ray of the identity, and the set of parameters for which it is the identity is

$$ \{Q:q^{2}=1\}=\{q=1\}\cup\{q=-1\}, $$

two affine lines in the plane, of real dimension one and two components. The comparison with the two previous categories is the whole point of the display: the complex two-sided operator gave the ray $[0,\infty)$ with kernel the circle $|Q|=1$, the split two-sided operator gave the line $\mathbb{R}$ with kernel the hyperbola $\langle Q\rangle_{*}=1$ of two branches, and the dual two-sided operator gives the ray with kernel two parallel lines. The conic has degenerated. The kernel is not compact in any of the three cases, and here it is the simplest disconnected set possible: two parallel lines, each a coset of the shear group $1+\mathrm{M}$.

The vanishing of the two-sided operator. The ray contains $0$, and $0$ is attained at a nonzero parameter, $\mathrm{Sd}_\varepsilon=0$. So the two-sided family is a monoid of scalings and not a group: it is closed under composition, it contains $0$, and its group of invertibles is the half-line $(0,\infty)\cdot\mathrm{id}$. In the split category the same family was all of $\mathbb{R}$, and the difference is the sign of the norm form.

The Kronecker Display

The identity. Vectorising by columns, $\operatorname{vec}(QXP)=(P^{\mathsf T}\otimes Q)\operatorname{vec}(X)$, so the operator $X\mapsto QXP$ on the four-dimensional space of matrices has the matrix

$$ K_{Q,P}=\mathsf{M}_2(P)^{\mathsf T}\otimes\mathsf{M}_2(Q), $$

of size $4\times4$.

The invariants. Since the two factors are triangular with diagonal $(p,p)$ and $(q,q)$,

$$ \operatorname{tr}K_{Q,P}=4pq,\qquad \det K_{Q,P}=p^{4}q^{4},\qquad \langle K_{Q,P}\rangle_*=(2q^{2}+q'^{2})(2p^{2}+p'^{2}),\qquad \operatorname{rank}K_{Q,P}=\operatorname{rank}\mathsf{M}_2(Q)\cdot\operatorname{rank}\mathsf{M}_2(P). $$

Proof. The trace of a Kronecker product is the product of the traces, $2p\cdot2q$; the determinant is $\det(P^{\mathsf T})^{2}\det(Q)^{2}=p^{4}q^{4}$; the Frobenius norm squared is the product of the two Frobenius norms squared; the rank is the product of the ranks.

The rank ladder. The two factors take the ranks $0,1,2$, so the display takes

$$ \operatorname{rank}K_{Q,P}\in\{0,1,2,4\}, $$

with the value $4$ on the pairs of invertibles, the value $2$ on the pairs with exactly one non-invertible of nonzero infinitesimal part, the value $1$ on the pairs of nilpotents, and the value $0$ on the pairs whose product vanishes. So the second rank is a genuine invariant, as it was in the split category: the ladder $\{0,1,2,4\}$ is complete and each value is attained. The difference from the split category is the shape of the rank-one factors, the two lines of projectors there and the single line of nilpotents here, and the difference is not visible in the trace, which sees only the product.

The reading. The display is block triangular, its two diagonal blocks being the same matrix $p\,\mathsf{M}_2(Q)$; hence

$$ \sigma(K_{Q,P})=\{pq\},\qquad \text{multiplicity }4, $$

and the nilpotent part is carried by the off-diagonal blocks. Its index is worth recording: with one infinitesimal part zero the nilpotent part is $K-pqI_4$ of square zero, and with both nonzero its square does not vanish, so the index is $2$ in the first case and $3$ in the second. Two nilpotent directions of the tensor square, and only the second-order cross term survives. So the display is degenerate exactly when the carrier operator is, and its defect is one step deeper on the open set where both parameters are non-invertibles.

The two operators. The same formula defines an operator of rank at most two on the carrier and an operator of rank up to four on the space of matrices: the carrier is two-dimensional and the space of matrices four-dimensional, so the display is the carrier operator acting on the tensor square, and its rank is the product of two ranks and not the rank of the product. The second rank is therefore a statement about the display and not about the carrier, and the two operators must be tabulated separately.

The Rank and the Invariants

The table. Two operators, one formula.

quantity on the carrier, $T=\mathsf{M}_2(QP)$ on the matrices, $K=\mathsf{M}_2(P)^{\mathsf T}\otimes\mathsf{M}_2(Q)$
size $2\times2$ $4\times4$
rank $2$, $1$ or $0$ by the stratum of $QP$ product of the two ranks, in $\{0,1,2,4\}$
trace $2qp$ $4pq$
the determinant of the image $(qp)^{2}$ $p^{4}q^{4}$
Frobenius norm squared $2(qp)^{2}+(qp'+q'p)^{2}$ $(2q^{2}+q'^{2})(2p^{2}+p'^{2})$
conformal? no, except on the scalars no, except on the scalars

The reading. Both operators are triangular and both are invertible exactly when both parameters are invertible. Neither is conformal, and the failure is the failure of the model to be closed under transposition: the display contains the transposed factor, and the transposed factor is the other convention. So the two operators are read in the two conventions at once, and this is why the display is the natural place to see that the algebra has two faithful triangular models and not one.

The determinant, and the blindness. The determinant of the carrier operator is $(qp)^{2}$, which vanishes on the whole set where $qp=0$; the determinant of the display is $p^{4}q^{4}$, which vanishes on the same set. Neither sees the nilpotent direction: two operators with the same product of real parts have the same determinant, whether the parameters are invertible or nilpotents. The invariants that see it are the ranks, and they are the subject of the penultimate section.

The Pair as a Semigroup

Composition. $L_{Q}R_{P}\circ L_{Q'}R_{P'}=L_{QQ'}R_{P'P}$, so the family is a monoid under composition, commutative, and the map $(Q,P)\mapsto L_{Q}R_{P}$ is a homomorphism of monoids onto the multiplicative monoid of the algebra.

The invertibles. The invertible members are the pairs of invertibles, and the group of invertible operators is

$$ \{L_{Q}R_{P}:Q,P\in(\mathbb{C}_{\mathrm{dln},R})^{\times}\}\cong(\mathbb{C}_{\mathrm{dln},R})^{\times}\cong R^{\times}\times(R,+), $$

the kernel of the map from the group of pairs onto it being the two-dimensional set $\{(Q,Q^{-1}):Q\in(\mathbb{C}_{\mathrm{dln},R})^{\times}\}$ of pairs with product one. So the group of the two-sided operators is the group of invertibles of the algebra, of dimension two, abelian, with two components over $\mathbb{R}$.

The monoid is not a group, for the right reason. The zero operator belongs to the family, as $L_{\varepsilon}R_{\varepsilon}$, and it is not invertible; so the monoid has a non-invertible member with both parameters nonzero, and it fails to be a group because of the zero divisors and not because of a sign. The temptation is to say that $-L_{Q}R_{P}$ is not in the family; it is, since $-L_{Q}R_{P}=L_{-Q}R_{P}$, the family being a subspace, and the correct reason is the one just given.

The automorphisms are inner in the ambient. The automorphisms of the algebra are $\varphi_c(\varepsilon)=c\varepsilon$ with $c$ an invertible element, so $\operatorname{Aut}_R(\mathbb{C}_{\mathrm{dln},R})\cong R^{\times}$, a one-parameter group. Each is realised by a conjugation in $M_2(R)$,

$$ \operatorname{diag}(c,1)\,\mathsf{M}_2(Q)\,\operatorname{diag}(c,1)^{-1}=\begin{pmatrix}q&cq'\\ 0&q\end{pmatrix}=\mathsf{M}_2\bigl(\varphi_c(Q)\bigr), $$

and the conjugating matrix is not in the image unless $c=1$. So the automorphism group is inner in the ambient and outer in the algebra, which is the opposite of the complex case, where the Galois conjugation was reachable by the transpose and not by a product at all. The reason is that the dual conjugation is not the Galois twist of a field but the rescaling of a nilpotent, and a rescaling of the off-diagonal entry is exactly what a diagonal conjugation does.

The derivations are not multiplications. The derivations of the algebra are the maps

$$ \partial_c(\varepsilon)=c\varepsilon,\qquad \partial_c(1)=0, $$

one for each $c\in R$, so the derivation module is one-dimensional, spanned by $\partial=\partial_1$, and the Leibniz rule is satisfied for every $c$ because $2\varepsilon\partial(\varepsilon)=0$ identically. No derivation is a multiplication: $L_A(1)=A$ while $\partial(1)=0$. In coordinates the derivation is

$$ \partial:\ (q,q')\longmapsto(0,q'),\qquad \text{matrix }\begin{pmatrix}0&0\\ 0&1\end{pmatrix}, $$

the projection onto the infinitesimal coordinate along the real one, and it is square-zero. So the calculus of the algebra is one operator, it is not in the family, and it is the operator the algebra is named after.

The Cone and the Coefficient Matrix

The parametrisation. The map

$$ \beta:\mathbb{C}_{\mathrm{dln},R}\times\mathbb{C}_{\mathrm{dln},R}\longrightarrow\operatorname{End}_R(\mathbb{C}_{\mathrm{dln},R}),\qquad \beta(Q,P)=L_{QP}, $$

has image the family itself, a linear subspace of dimension two, and not a cone.

The fibres. Over $L_c$ with $c\neq0$ the fibre is $\{(Q,P):QP=c\}$, of dimension two; over $0$ it is the union of three pieces,

$$ \{0\}\times\mathbb{C}_{\mathrm{dln},R}\ \ \cup\ \ \mathbb{C}_{\mathrm{dln},R}\times\{0\}\ \ \cup\ \ \mathrm{M}\times\mathrm{M}, $$

three linear pieces of dimension two. The third is the one the other categories of the block do not have: two nilpotents multiply to zero with neither factor zero, because the maximal ideal is square-zero.

The membership test. An ambient matrix $C=\begin{pmatrix}a&b\\ c&d\end{pmatrix}$ is a two-sided product exactly when it is upper triangular with equal diagonal entries,

$$ c=0,\qquad a=d, $$

two linear equations. The test is linear, as in the complex and split categories, and against the determinantal test of the quaternion ones; the image is a subspace and the word cone has no content for the image, though it has content for the fibre.

The coefficient array. In the basis $\{T_1,T_\varepsilon\}=\{I_2,E\}$ the coefficient array of a member is its pair of coordinates,

$$ c(T_Q)=(q,q'), $$

a vector of length two and not an array; the general theory's $n\times n$ array is here a $2\times1$ column, and the "rank" of the coefficient array of a single product is one, as it must be. The array of the pair is the matrix of products of coordinates,

$$ c(Q,P)=\begin{pmatrix}qp&qp'\\ q'p&q'p'\end{pmatrix},\qquad \operatorname{rank}c(Q,P)=1\ \text{when }Q\neq0\text{ and }P\neq0, $$

and it is the zero array when either factor vanishes. The pairs of nilpotents give the array $\begin{pmatrix}0&0\\ 0&q'p'\end{pmatrix}$, of rank one when both are nonzero, although the product they produce is the zero operator. This is the sharpest form of the difference between the pair and the operator: the pair always has a rank-one array, and the operator it produces can be zero.

The Length

The statement. For an operator of the span,

$$ \ell(L_{Q}R_{P})\in\{0,1\},\qquad \operatorname{span}\{L_{Q}R_{P}\}=L(\mathbb{C}_{\mathrm{dln},R})\cup\{0\}, $$

the value $0$ only for the zero operator written with a zero parameter, and the span equal to the family because the family is a subspace closed under products.

The degenerate case of the length theorem. The general theorem allows at most $n=2$ products and identifies the length with the rank of the coefficient array. Here the coefficient array of a single product has rank one and the span is closed, so no sum is ever needed and the bound is not attained. The nilpotent does not lengthen anything: $L_\varepsilon$ is one product, of length one, though its rank is one and its square is zero. So the length counts products and not ranks, exactly as in the other commutative categories, and the presence of a zero divisor changes nothing.

The zero operator has length zero and a nonzero description. Both statements are true and they are not in contradiction. The length of the zero operator is zero, because $L_{0}R_{0}=0$ is a product of a zero parameter. And the zero operator is also $L_{\varepsilon}R_{\varepsilon}$, a product of two nonzero parameters. The length is a minimum over descriptions and the minimum is zero; the fibre over zero is the set of all descriptions, and it has a component of nonzero pairs. The complex category knows only the first statement; the second is the operator form of a nonzero nilpotent ideal, and it is the precise sense in which this category is not reduced.

Why the span is closed. The products are the compositions $L_Q\circ R_P=L_{QP}$, so the multiplication map sends $L\times R$ into $L$; the image is the family and it is a subspace. In the quaternion category the same map is an isomorphism $A\otimes A_{\mathrm{opp}}\to\operatorname{End}_R(A)$ and fills the sixteen-dimensional ambient; the difference is central simplicity, and $\mathbb{C}_{\mathrm{dln},R}$ is commutative and local. The failure is the same one as in the two previous categories, and it is inherited from the algebra and not from the dimension.

What lies outside the span. Two dimensions: the traceless diagonal line $\mathbb{R}\operatorname{diag}(1,-1)$ and the lower corner $\mathbb{R}E_{21}$. Every ambient matrix is the sum of a member of the family and a combination of $\operatorname{diag}(1,-1)$ and $E_{21}$, so the ambient is the family, the traceless diagonal and the lower corner. The derivation lies in the first two, the residue too, and the transpose needs the third; this is the decomposition the last section uses.

The Two Ranks

The two numbers. The operator rank is the rank of the operator on the carrier, in $\{0,1,2\}$; the coefficient rank is the rank of its coefficient array, which is $1$ for every nonzero member. They differ by one on the complement of the maximal ideal and coincide on it, where both are one.

The table.

operator operator rank coefficient rank comment
$0=L_{0}R_{0}=L_{\varepsilon}R_{\varepsilon}$ $0$ $0$ the vertex, and a product of nonzero parameters
$I_2=L_{1}R_{0}$ $2$ $1$ the identity, of length one
$\mathsf{M}_2(1+s\varepsilon)=I_2+sE$ $2$ $1$ every shear, invertible
$q'L_\varepsilon$, $q'\neq0$ $1$ $1$ the whole rank-one stratum, one line
$\partial_\varepsilon$ $1$ — not in the span, no length
the residue $\pi$ $1$ — not in the span, no length
the transpose $4$ — not in the span, no length

The reading. The rank is one exactly on the line $\mathbb{R}L_\varepsilon$ minus the origin, and two on the rest of the family. The coefficient rank is one everywhere except at the zero operator, and the determinant is zero on the whole line and positive elsewhere. So the three invariants separate the family into two strata and disagree on one of them: on the rank-one line the determinant says "degenerate", the coefficient rank says "one product", and the operator rank says "singular". The rank is the one that names the stratum, and the determinant is the one that cannot.

The rank of the derivation. The derivation $\partial_\varepsilon$ has matrix $\operatorname{diag}(0,1)$, of rank one, and it is not in the span, so it has no coefficient array and no length. It is the clearest single example of the article's theme: an operator of the ambient, of rank one, of square zero, that is not a product of elements at any length. Its existence is not an accident; it is the derivative, and the algebra of the dual numbers is the algebra of the derivative precisely because it has an operator of this kind outside its own multiplication.

What the Pair Buys and Does Not

What it buys. Three statements, and they are the reasons the article exists.

  • The zero operator is a product of nonzero parameters, so the two-sided family is not the image of a group action and the fibre over zero has a component of nilpotent pairs.
  • The second rank is not vacuous: the ladder $\{0,1,2,4\}$ of the display is complete, and the value $2$ is taken exactly on the pairs with one nilpotent parameter, a stratum that the trace and the determinant cannot see.
  • The automorphism group is visible: $\operatorname{Aut}\cong R^{\times}$ is realised by the diagonal conjugations $\operatorname{diag}(c,1)$, so the automorphisms are inner in the ambient and outer in the algebra, the opposite of the complex case.

What it does not buy: the Peirce frame, the corners, the cone. The three structures of the other categories are absent here for the reasons of the model article, and the operator layer cannot supply them: no idempotent means no Peirce decomposition of the ambient, so there are no corners; and the commutativity means that the image is a subspace, so there is no cone to stratify. Each absence is inherited, and the article records that it is inherited rather than repeating the proof.

What it does not buy: a $*$-structure. The model is not a $*$-representation, as the section on the Frobenius adjoint showed, and the failure is one-dimensional: the transpose is the only involution the model needs and the only one it does not contain. So the operator layer of the dual numbers is the algebra, its calculus, and one transposition.

The residue. The ambient decomposes as

$$ M_2(R)=\underbrace{\mathsf{M}_2(\mathbb{C}_{\mathrm{dln},R})}_{\text{the family}}\ \oplus\ \underbrace{\mathbb{R}\operatorname{diag}(1,-1)}_{\text{the traceless diagonal}}\ \oplus\ \underbrace{\mathbb{R}E_{21}}_{\text{the lower corner}}, $$

of dimensions two, one and one. The family is the multiplication; with the scalars, the traceless diagonal carries the residue $\operatorname{diag}(1,0)$, the derivation $\operatorname{diag}(0,1)$ and every polynomial in them, so it carries the calculus of the algebra; the lower corner carries the transpose and with it the other triangular convention. So the operator layer of the dual numbers is the algebra plus its calculus plus one direction, and the one direction is the mirror in which the algebra's two faithful models see each other.

Summary

  1. The carrier is two-dimensional and the ambient four-dimensional; the one-sided families coincide, and the family is spanned by the identity and the square-zero operator $L_\varepsilon$, the smallest faithful operator algebra that is not semisimple.
  2. The rank ladder of the one-sided operator is $2,1,0$, with the rank-one stratum the single line $\mathbb{R}L_\varepsilon$ minus the origin, and there $\ker L_Q=\operatorname{im}L_Q=\mathrm{M}$, one Jordan block of size two.
  3. The plain adjoint is the operator itself, by commutativity; the Frobenius adjoint is the conjugate of the operator by the swap, and it leaves the family, so the model is a representation and not a $*$-representation.
  4. The two-sided operator is the multiplication by the product, and the zero operator is a product of nonzero parameters, $L_{\varepsilon}R_{\varepsilon}=0$: the fibre over zero has a component of nilpotent pairs.
  5. The two-sided operator with the conjugate is $q^{2}\operatorname{id}$, so the two-sided family is the ray $[0,\infty)$ of the identity and the parameters giving the identity are the two parallel lines $q=\pm1$: the conic of the previous categories has degenerated.
  6. The Kronecker display is $\mathsf{M}_2(P)^{\mathsf{T}}\otimes\mathsf{M}_2(Q)$, of trace $4pq$, determinant $p^{4}q^{4}$, and rank in $\{0,1,2,4\}$: the second rank is not vacuous, and the value $2$ marks the pairs with exactly one nilpotent parameter.
  7. The monoid of two-sided operators is commutative, its group of invertibles is the group of invertibles of the algebra, and it fails to be a group through the zero divisors and not through a sign.
  8. The automorphisms are $\varphi_c(\varepsilon)=c\varepsilon$, realised by $\operatorname{diag}(c,1)$: inner in the ambient, outer in the algebra, the opposite of the complex case; the derivations are one-dimensional and no derivation is a multiplication.
  9. The image of the parametrisation is a linear subspace of dimension two, the membership test is the two linear equations $c=0$, $a=d$, the coefficient array of a member is the vector $(q,q')$, and the length is in $\{0,1\}$.
  10. The operator rank takes $2,1,0$ while the coefficient rank is one, and the ambient decomposes as the family, the traceless diagonal carrying the residue and the derivation, and the lower corner carrying the transpose.

Summary of Notation

Symbol Meaning Article
$L_Q,R_Q$ left and right multiplication, equal by commutativity $M_2(R)$ Triangular Representation
$L_\varepsilon$ the square-zero generator $E$ of the family this article
$L_{Q}R_{P}$ $X\mapsto QXP$, equal to $L_{QP}$ this article
$K_{Q,P}$ the Kronecker display $\mathsf{M}_2(P)^{\mathsf{T}}\otimes\mathsf{M}_2(Q)$ this article
$\mathrm{Sd}_Q$ the two-sided operator $X\mapsto QX\bar Q=q^{2}X$ this article
$c(T_Q)=(q,q')$ the coefficient array of a member, a vector of length two this article
$c(Q,P)$ the array of the pair, of rank one for a pair with both factors nonzero this article
$\ell$ the length, in $\{0,1\}$ inside the span this article
$\varphi_c$ the automorphism $\varepsilon\mapsto c\varepsilon$, realised by $\operatorname{diag}(c,1)$ Dual-Numbers Automorphisms and Derivations
$\partial_\varepsilon$ the derivation $q+\varepsilon q'\mapsto q'\varepsilon$, not a multiplication Dual-Numbers Automorphisms and Derivations
$\pi$ the residue, the spectral projection of the model $M_2(R)$ Triangular Representation
$S$ the swap, the change between the two triangular conventions this article

Further Reading

  • Richard S. Pierce, Associative Algebras (Springer, 1982), for the map $A\otimes A_{\mathrm{opp}}\to\operatorname{End}(A)$ and the double centraliser theorem, whose failure for a commutative algebra is the collapse of this category.
  • T. Y. Lam, A First Course in Noncommutative Rings (Springer, 2001), for local rings, square-zero ideals and the trivial extension.
  • Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge University Press, 2013), for the Kronecker product, the vectorisation identity, and its rank, trace and determinant.
  • F. R. Gantmacher, The Theory of Matrices (Chelsea, 1959), for Jordan blocks and the centraliser of a nilpotent matrix.
  • Andreas Griewank and Andrea Walther, Evaluating Derivatives (SIAM, 2008), for the derivation of the truncated Taylor algebra and its matrix realisation.