$M_2(\mathbb{C})$ Pauli Representation : Operators

Introduction

The question. The $M_2(\mathbb{C})$ Pauli Representation article compresses the two regular matrices of an element to one block: $L(\tilde Q)=\mathrm I_2\otimes A$ compresses to $A$, and $R(\tilde Q)=A^{\mathsf T}\otimes\mathrm I_2$ compresses to $A^{\mathsf T}$. The compression says what one block of each regular matrix is. This article asks the question that follows: on the compressed side, which maps of $M_2(\mathbb{C})$ arise from multiplication, and how many $2\times2$ matrices does each need?

The count. The answer is a count of matrices.

  • A left multiplication is $Y\mapsto AY$, and one matrix $A$ describes it.
  • A right multiplication is $Y\mapsto YB$, and one matrix $B$ describes it.
  • A two-sided operator is $Y\mapsto AYB$, and a pair $(A,B)$ describes it.
  • A general endomorphism is not a single operator of these kinds, and four pairs are always enough.

The first two bullets are the one-sided operators; the third is the two-sided operator.

Why one matrix per side. A $2\times2$ matrix has four complex entries. A left multiplication is determined by its value on the unit, $L_A(\mathrm I_2)=A$, so the element and its left operator carry the same four numbers. A right multiplication is determined by $R_B(\mathrm I_2)=B$. A two-sided operator has seven effective complex parameters, more than the four of one $2\times2$ matrix, and no single $2\times2$ matrix records both sides: $A$ acts on the left and $B$ acts on the right.

The result in one line. On $M_2(\mathbb{C})$, a one-sided operator is one matrix, a two-sided operator is a pair of matrices, unique up to scaling when the operator is nonzero, and a general endomorphism is a sum of at most four pairs; the $4\times4$ matrix of the pair is the Kronecker product $B^{\mathsf T}\otimes A$.

Boundaries. The frame, the two blocks, the compression and the algebra model are $M_2(\mathbb{C})$ Pauli Representation. The uncompressed operators are $M_4(\mathbb{C})$ Pauli Representation : Operators; the same operators in the axis-free basis are $M_4(\mathbb{C})$ Axis-Free Representation : Operators. The matrix units and the two Peirce planes are The Polarized Module of the Biquaternion Algebra. The abstract family is Operators Acting on Two Sides. The Hermitian pair is Part V and is not treated here.

Conventions. The algebra is $\mathbb{B}=\mathbb{C}\otimes_{\mathbb{R}}\mathbb{H}$ on the basis $(e_t,\vec e_x,\vec e_y,\vec e_z)$, with $\vec e_k^2=-e_t$ and $\vec e_x\vec e_y=\vec e_z$, $\vec e_y\vec e_z=\vec e_x$, $\vec e_z\vec e_x=\vec e_y$. The frame is $$ \mathcal{P}=(\tilde\Pi_+,\vec\Upsilon_+,\vec\Upsilon_-,\tilde\Pi_-),\qquad \tilde\Pi_\pm=\tfrac12(e_t\pm i\vec e_z),\qquad \vec\Upsilon_\pm=\tfrac12(i\vec e_x\pm\vec e_y). $$ The frame order is the column-major matrix-unit order, with the matrix units $E_{ij}$ satisfying $E_{ij}E_{kl}=\delta_{jk}E_{il}$, $$ \tilde\Pi_+\leftrightarrow E_{11},\quad \vec\Upsilon_+\leftrightarrow E_{21},\quad \vec\Upsilon_-\leftrightarrow E_{12},\quad \tilde\Pi_-\leftrightarrow E_{22}, $$ so the coordinate column of a $2\times2$ matrix $Y$ is $(y_{11},y_{21},y_{12},y_{22})^{\mathsf T}$. The model is $A=\mathsf M_2(\tilde Q)$ with $$ \mathsf M_2(\tilde Q)=Q_t\mathrm I_2-i\bigl(Q_x\sigma_1+Q_y\sigma_2+Q_z\sigma_3\bigr), $$ where $\sigma_1,\sigma_2,\sigma_3$ are the Pauli matrices; the four units map to $\mathrm I_2,-i\sigma_1,-i\sigma_2,-i\sigma_3$. The trace is $\operatorname{Tr}A=2Q_t$, and the determinant is $\det A=\langle\tilde Q\rangle_{\natural}=Q_t^2+Q_x^2+Q_y^2+Q_z^2$, the quaternionic norm. A second element is $\tilde Q_2$, with $B=\mathsf M_2(\tilde Q_2)$. The $4\times4$ matrices $L,R$ are those of the Pauli Representation article, and $\otimes$ is the Kronecker product in the frame order.

The One-Sided Operators

The Element as One Matrix

The compression. The map $\mathsf M_2:\mathbb{B}\to M_2(\mathbb{C})$ is an algebra isomorphism: it is linear, bijective and multiplicative. It reads the element itself as one $2\times2$ matrix, with the four Peirce coordinates as entries. The unit is $\mathrm I_2$, the product is the matrix product, and the norm is the determinant.

The count for the element. One element costs one matrix. The question of the article is what an operator costs. The answer begins with the observation that each one-sided action is determined by its value on the unit.

The Left Operator

The operator. For $A\in M_2(\mathbb{C})$, the left multiplication is $$ L_A:M_2(\mathbb{C})\longrightarrow M_2(\mathbb{C}),\qquad L_A(Y)=AY. $$ It is $\mathbb{C}$-linear. The matrix $A$ is recovered from the operator by $A=L_A(\mathrm I_2)$, so one matrix describes the operator and the description is unique.

The $4\times4$ matrix. In the frame $\mathcal{P}$, $$ [L_A]_{\mathcal{P}}=\mathrm I_2\otimes A=\begin{pmatrix}A&0\\0&A\end{pmatrix}. $$ The second block is a copy of the first, which is the compression of the Pauli Representation article: the left regular matrix of $\tilde Q$ is this matrix, and its one block is $A$.

The law. The composition is $L_A L_C=L_{AC}$, and the map $A\mapsto L_A$ is injective. The left copy is the four-dimensional subalgebra $\mathbb{B}_L=\{\mathrm I_2\otimes A\}$.

The invariants. The rank is $2\operatorname{rank}A$, because the matrix is two copies of $A$. The trace is $2\operatorname{Tr}A=4Q_t$, and the determinant is $(\det A)^2=\langle\tilde Q\rangle_{\natural}^2$.

The Right Operator

The operator. For $B\in M_2(\mathbb{C})$, the right multiplication is $$ R_B:M_2(\mathbb{C})\longrightarrow M_2(\mathbb{C}),\qquad R_B(Y)=YB. $$ The matrix $B$ is recovered by $B=R_B(\mathrm I_2)$.

The $4\times4$ matrix. In the frame $\mathcal{P}$, $$ [R_B]_{\mathcal{P}}=B^{\mathsf T}\otimes\mathrm I_2= \begin{pmatrix} b_{11}\mathrm I_2&b_{21}\mathrm I_2\\ b_{12}\mathrm I_2&b_{22}\mathrm I_2 \end{pmatrix}. $$ The transpose is the price of the column-major frame: the block matrix records $B^{\mathsf T}$, while the operator uses $B$. The compressed block of the Pauli Representation article is $B^{\mathsf T}$ for the same reason, and the action itself is right multiplication by $B$.

The law. The composition is anti-multiplicative, $R_B R_D=R_{DB}$, and the map $B\mapsto R_B$ is injective. The right copy is the four-dimensional subalgebra $\mathbb{B}_R=\{B^{\mathsf T}\otimes\mathrm I_2\}$.

The invariants. The rank is $2\operatorname{rank}B$. The trace is $2\operatorname{Tr}B=4P_t$, and the determinant is $(\det B)^2=\langle\tilde Q_2\rangle_{\natural}^2$.

The Two Copies

The intersection. A matrix that is both block-diagonal with two equal blocks and block-scalar with scalar blocks is a scalar matrix: $$ \mathbb{B}_L\cap\mathbb{B}_R=\mathbb{C}\cdot\mathrm I_4. $$ In the $2\times2$ language, $L_A=R_B$ forces $A=B$ at the unit and then $AY=YA$ for all $Y$, so $A$ is scalar. The intersection is the scalar operators.

The commutation. The two actions commute, $L_A R_B=R_B L_A$, because $A(YB)=(AY)B$. The two copies are mutual centralisers; the proof is in the $4\times4$ operators article.

The one-sided level. The two copies are the one-sided level of the operator problem: two four-dimensional linear subspaces of the sixteen-dimensional ambient space. Each is described by one $2\times2$ matrix.

The Two-Sided Operator

The Pair

The operator. For $A,B\in M_2(\mathbb{C})$, the two-sided operator is $$ L_AR_B:M_2(\mathbb{C})\longrightarrow M_2(\mathbb{C}),\qquad L_AR_B(Y)=AYB. $$ It is $\mathbb{C}$-linear, and it is the composite $L_A R_B=R_B L_A$.

Two matrices are needed. The pair is the data. A single $2\times2$ matrix cannot describe the general operator: if $A$ and $B$ are nonzero, then $L_AR_B$ is a left multiplication exactly when $B$ is a scalar matrix, and a right multiplication exactly when $A$ is a scalar matrix. The proof is in §The One-Sided Members. A pair with neither slot scalar is genuinely two-sided, and it needs both matrices.

The $4\times4$ matrix. The display is computed in §The Kronecker Display: $$ [L_AR_B]_{\mathcal{P}}=B^{\mathsf T}\otimes A. $$ The operator is determined by the pair; for a nonzero operator, the pair is determined by the operator up to scaling.

The Scaling

The invariance. For $\lambda\in\mathbb{C}^{\times}$, $$ L_{\lambda A}R_{\lambda^{-1}B}=L_AR_B. $$ The scalar passes from one factor to the other, and the operator does not change.

The uniqueness. Conversely, if $L_AR_B=L_{A'}R_{B'}$ and the operator is nonzero, then $(A',B')=(\lambda A,\lambda^{-1}B)$ for some $\lambda\in\mathbb{C}^{\times}$.

Proof. For the matrix units, $$ L_AR_B(E_{ij})=A E_{ij}B=a_i b_j^{\mathsf T},\qquad L_{A'}R_{B'}(E_{ij})=a'_i (b'_j)^{\mathsf T}. $$ Equality for all $i,j$ gives $a'_i(b'_j)^{\mathsf T}=a_i b_j^{\mathsf T}$. Choose $i,j$ with $a_i\neq0$ and $b_j\neq0$; then $a'_i=\lambda a_i$ and $b'_j=\lambda^{-1}b_j$ for one nonzero $\lambda$. Comparing the pairs $(i,l)$ and $(k,j)$ gives $a'_k=\lambda a_k$ and $b'_l=\lambda^{-1}b_l$ for all $k,l$. Hence the pair is the scaling orbit.

The count. The pair space has complex dimension eight. The scaling group $\mathbb{C}^{\times}$ acts freely on the nonzero pairs, so the image has complex dimension $8-1=7$. The zero operator is the vertex of the cone, reached by every pair with $A=0$ or $B=0$.

The One-Sided Members

The criterion. Let $A$ and $B$ be nonzero. The operator $L_AR_B$ is a left multiplication exactly when $B$ is a scalar matrix, and a right multiplication exactly when $A$ is a scalar matrix.

Proof. If $B=\mu\mathrm I_2$, then $L_AR_B(Y)=\mu AY=L_{\mu A}(Y)$. If $A=\lambda\mathrm I_2$, then $L_AR_B(Y)=\lambda YB=R_{\lambda B}(Y)$.

Conversely, suppose $L_AR_B=L_C$. At $Y=\mathrm I_2$, $C=AB$. Hence $AYB=ABY$ for all $Y$, that is $A(YB-BY)=0$ for all $Y$. If $A$ is invertible, then $YB=BY$ for all $Y$, so $B$ is scalar. If $A$ is not invertible, then $\ker A$ is a line, and the image of $Y\mapsto YB-BY$ is the commutator space $[M_2(\mathbb{C}),B]$, of dimension two when $B$ is not scalar; it cannot lie in a line. Hence $B$ is scalar. The right case is symmetric.

The copies. The left copy is the set of pairs $(A,\mathrm I_2)$, and the right copy the set of pairs $(\mathrm I_2,B)$. A pair with a scalar slot is in the scaling orbit of one of these, and the scaling moves the scalar into the other slot without changing the operator. The intersection is the scalar operators, the pairs with both slots scalar.

The precise sense of one matrix per side. A one-sided operator is a pair with a scalar in one slot. The scalar can be moved into the other slot by the scaling, so the pair reduces to one matrix. A two-sided operator has no scalar slot; the two matrices are independent, and the pair is the minimal data.

The Table of Levels

The three levels and the general level are collected in one table. The complex dimension is the dimension of the set of operators of that level.

Level Operator $2\times2$ data $4\times4$ matrix Complex dimension
left $L_A(Y)=AY$ one matrix $A$ $\mathrm I_2\otimes A$ $4$
right $R_B(Y)=YB$ one matrix $B$ $B^{\mathsf T}\otimes\mathrm I_2$ $4$
two-sided $L_AR_B(Y)=AYB$ a pair $(A,B)$ modulo scaling $B^{\mathsf T}\otimes A$ $7$ (cone)
general an endomorphism up to four pairs an arbitrary $4\times4$ matrix $16$

The table is the answer to the operator question: one matrix, one matrix, a pair, four pairs.

The Kronecker Display

The Vectorisation Identity

The vectorisation. The coordinate column of $Y$ in the frame order is $$ \operatorname{vec}(Y)=(y_{11},y_{21},y_{12},y_{22})^{\mathsf T}. $$ It is the column-stacking of $Y$.

The identity. For all $A,B,Y\in M_2(\mathbb{C})$, $$ \operatorname{vec}(AYB)=(B^{\mathsf T}\otimes A)\operatorname{vec}(Y). $$

Proof. Both sides are linear in $Y$, so it is enough to check $Y=E_{ij}$. The unit has $\operatorname{vec}(E_{ij})=e_j\otimes e_i$, and $$ A E_{ij}B=a_i b_j^{\mathsf T},\qquad \operatorname{vec}(a_i b_j^{\mathsf T})=b_j\otimes a_i =(B^{\mathsf T}\otimes A)(e_j\otimes e_i). $$ The identity follows by linearity.

The three displays. The identity gives the three matrices of the article at once: $$ \operatorname{vec}(AY)=(\mathrm I_2\otimes A)\operatorname{vec}(Y),\qquad \operatorname{vec}(YB)=(B^{\mathsf T}\otimes\mathrm I_2)\operatorname{vec}(Y),\qquad \operatorname{vec}(AYB)=(B^{\mathsf T}\otimes A)\operatorname{vec}(Y). $$ The first is the left matrix, the second the right matrix, the third the pair. The order of the factors is fixed by the column-major frame: the right action occupies the first slot.

The Three Products

The ordinary product. The $4\times4$ matrix of the pair is the ordinary product of the left and right matrices: $$ (\mathrm I_2\otimes A)(B^{\mathsf T}\otimes\mathrm I_2)=B^{\mathsf T}\otimes A, $$ by the mixed-product property. It is the operator.

The Kronecker product of the $2\times2$ matrices. The same matrix is the Kronecker product $B^{\mathsf T}\otimes A$ of the two $2\times2$ matrices. This is the display that the Peirce frame makes possible: the element is already a matrix, and the operator is built from the two element matrices.

The Kronecker product of the $4\times4$ matrices. The Kronecker product $(\mathrm I_2\otimes A)\otimes(B^{\mathsf T}\otimes\mathrm I_2)$ is $16\times16$, acts on the tensor square of the carrier, and is not the operator. The distinction is the one drawn in $M_4(\mathbb{C})$ Pauli Representation : Operators, §The Kronecker Question: the ordinary product is the operator, the Kronecker product of the two $2\times2$ matrices is its display, and the Kronecker product of the two $4\times4$ matrices is a different operator on a different space.

The frame dependence. The display $B^{\mathsf T}\otimes A$ is a property of the Peirce frame. In the axis-free basis the same operator is a sum of sixteen signed-permutation basis operators, and no $2\times2$ block appears; the pair survives, but the frame does not display it. The two articles read the same operator in the two frames, and the pair is the invariant.

The Coefficient Matrix

The expansion. The sixteen products $L_{e_\mu}R_{e_\nu}$ are an orthogonal basis of $M_4(\mathbb{C})$ for the Hilbert–Schmidt inner product, and every matrix has a unique expansion $M=\sum c_{\mu\nu}L_{e_\mu}R_{e_\nu}$. The matrix $c(M)$ is the coefficient matrix. The definition and the length theorem are in the $4\times4$ operators article.

The coefficient matrix of a pair. If $A=\sum_\mu Q_\mu\mathsf M_2(e_\mu)$ and $B=\sum_\nu P_\nu\mathsf M_2(e_\nu)$, then $$ [L_AR_B]_{\mathcal{P}}=B^{\mathsf T}\otimes A =\sum_{\mu,\nu}Q_\mu P_\nu\,L_{e_\mu}R_{e_\nu}, $$ so $$ c(L_AR_B)=[Q][P]^{\mathsf T}, $$ the outer product of the two coefficient columns. It has rank at most one; conversely every rank-one coefficient matrix is such an outer product and gives a pair. The pair picture and the coefficient picture are the same statement, one in $2\times2$ matrices and one in the $4\times4$ coefficient matrix.

The two $4\times4$ matrices. The coefficient matrix $c(L_AR_B)$ and the operator matrix $[L_AR_B]_{\mathcal{P}}$ are both $4\times4$, and they are different objects. The first is the matrix of the tensor in the basis $L_{e_\mu}R_{e_\nu}$; the second is the matrix of the operator on the carrier. The first has rank at most one, and rank one when the operator is nonzero; the second has rank $\operatorname{rank}A\cdot\operatorname{rank}B$.

The length. By the length theorem, the least number of pairs summing to $M$ is $\operatorname{rank}c(M)$. The pair set is the cone $\operatorname{rank}c\le1$; its projectivisation is the Segre variety $\mathbb{P}^3\times\mathbb{P}^3$. The length is $\ell(M)=\operatorname{rank}c(M)$, and the general endomorphism needs up to four pairs, the subject of §The General Endomorphism.

The Rank and the Invariants

The Rank Formula

The theorem. For all $A,B$, $$ \operatorname{rank}L_AR_B=\operatorname{rank}A\cdot\operatorname{rank}B. $$ The possible values are $0,1,2,4$; a product never has rank three.

Proof. Choose invertible matrices $G,H,K,L$ with $$ A=G\begin{pmatrix}\mathrm I_r&0\\0&0\end{pmatrix}H,\qquad B=K\begin{pmatrix}\mathrm I_s&0\\0&0\end{pmatrix}L, $$ where $r=\operatorname{rank}A$ and $s=\operatorname{rank}B$. As $Y$ ranges over $M_2(\mathbb{C})$, the matrix $HYK$ ranges over $M_2(\mathbb{C})$, and the map $$ Z\longmapsto \begin{pmatrix}\mathrm I_r&0\\0&0\end{pmatrix}Z \begin{pmatrix}\mathrm I_s&0\\0&0\end{pmatrix} $$ has image the $r\times s$ corner, of dimension $rs$. Hence the rank is $rs$.

The Kronecker form. The same formula is the rank of the Kronecker product: $$ \operatorname{rank}(B^{\mathsf T}\otimes A)=\operatorname{rank}B\cdot\operatorname{rank}A. $$ The $4\times4$ matrix and the $2\times2$ pair have the same rank.

The Trace and the Determinant

The invariants. The trace and the determinant of the pair are $$ \operatorname{Tr}L_AR_B=4Q_tP_t,\qquad \det L_AR_B=\langle\tilde Q\rangle_{\natural}^2\langle\tilde Q_2\rangle_{\natural}^2. $$

Proof. The $4\times4$ matrix is $B^{\mathsf T}\otimes A$. The trace of a Kronecker product is the product of the traces, and $\operatorname{Tr}A=2Q_t$, $\operatorname{Tr}B=2P_t$. The determinant of a Kronecker product is $\det(B^{\mathsf T})^2\det(A)^2=(\det B)^2(\det A)^2$, and the determinants are the norms.

The inverse. If $A$ and $B$ are invertible, the inverse is $(L_AR_B)^{-1}=L_{A^{-1}}R_{B^{-1}}$. If one of them is singular, the rank is at most two and the operator is not invertible.

The Three Examples

The rank one. For $A=B=E_{11}$, the operator is $Y\mapsto y_{11}E_{11}$; its $4\times4$ matrix is $E_{11}\otimes E_{11}$, the matrix with a single $1$ at the first diagonal entry. The rank is $1\cdot1=1$.

The rank two. For $A=\mathrm I_2$ and $B=E_{11}$, the operator is $Y\mapsto YE_{11}$, the right multiplication by $E_{11}$; its matrix is $E_{11}\otimes\mathrm I_2=\operatorname{diag}(1,1,0,0)$. The rank is $2\cdot1=2$.

The rank four. For $A=B=\mathrm I_2$, the operator is the identity; its matrix is $\mathrm I_4$. The rank is $2\cdot2=4$.

The three examples are the values $1,2,4$ of the rank formula. The value $0$ is the zero operator.

The Pair as a Semigroup

Composition

The law. The composite of two pairs is $$ L_AR_BL_CR_D=L_{AC}R_{DB}. $$ Proof. For every $Y$, $$ L_AR_B\bigl(L_CR_D(Y)\bigr)=A(CYD)B=(AC)Y(DB). $$ The second slot reverses the order: the pair multiplies as $(A,B)(C,D)=(AC,DB)$.

The unit. The identity is $L_{\mathrm I_2}R_{\mathrm I_2}$. The pair multiplication is associative, and the scaling orbit is preserved: the product of the orbits of $(A,B)$ and $(C,D)$ is the orbit of $(AC,DB)$.

The algebra. The pair multiplication is the multiplication of $M_2(\mathbb{C})\times M_2(\mathbb{C})_{\mathrm{opp}}$: the first factor multiplies in the algebra, the second in the opposite algebra. This is the $2\times2$ form of the enveloping algebra $\mathbb{B}\otimes_{\mathbb{C}}\mathbb{B}_{\mathrm{opp}}$.

The Inverse

The inverse. If $A$ and $B$ are invertible, then $$ (L_AR_B)^{-1}=L_{A^{-1}}R_{B^{-1}}, $$ because $L_AR_BL_{A^{-1}}R_{B^{-1}}=L_{\mathrm I_2}R_{\mathrm I_2}$.

The group. The invertible pairs form $GL_2(\mathbb{C})\times GL_2(\mathbb{C})$, of complex dimension eight. The scaling $\mathbb{C}^{\times}$ is the kernel of the action on the carrier, so the image is a seven-dimensional subgroup of $GL_4(\mathbb{C})$. It is the group of invertible two-sided operators.

The Automorphisms

The criterion. An invertible pair $L_AR_B$ is an algebra automorphism exactly when $B=A^{-1}$.

Proof. For all $Y,W$, $$ L_AR_B(Y)L_AR_B(W)=AYB\,AWB,\qquad L_AR_B(YW)=AYWB. $$ The two are equal exactly when $A(Y(BA-\mathrm I_2)W)B=0$ for all $Y,W$, which for invertible $A,B$ is $BA=\mathrm I_2$, that is $B=A^{-1}$.

The conjugation. When $B=A^{-1}$, the operator is $$ L_AR_{A^{-1}}(Y)=AYA^{-1}, $$ conjugation by $A$. Every automorphism of $M_2(\mathbb{C})$ is inner, so the automorphism group is $PGL_2(\mathbb{C})$, of complex dimension three. The pair $(A,A^{-1})$ is determined by $A$ up to scalar, and the scaling orbit is the same automorphism.

The one-sided case. A left multiplication $L_A$ is an automorphism only when $A=\mathrm I_2$; the same holds for a right multiplication. The nontrivial automorphisms are genuinely two-sided, and they are the pairs $(A,A^{-1})$.

The example. For $A=\operatorname{diag}(2,1)$, $$ L_AR_{A^{-1}}(Y)=\begin{pmatrix}2&0\\0&1\end{pmatrix}Y\begin{pmatrix}\tfrac12&0\\0&1\end{pmatrix}, $$ which fixes the diagonal entries, doubles $y_{12}$ and halves $y_{21}$. The pair is $(A,A^{-1})$, and the scaling orbit $(\lambda A,\lambda^{-1}A^{-1})$ is the same automorphism.

The System of Matrix Units

The data. A system of matrix units in the Peirce frame is a family $E'_{ij}$ with $$ E'_{ij}E'_{kl}=\delta_{jk}E'_{il},\qquad E'_{11}+E'_{22}=\mathrm I_2,\qquad E'_{11}\neq0 . $$ The relations force the two diagonal elements to be complementary idempotents and the two off-diagonal elements to be square-zero.

The reciprocal scaling. Fix the two idempotents, $E'_{11}=E_{11}$ and $E'_{22}=E_{22}$. Each off-diagonal element then lies in one of the two mixed lines, so $E'_{12}=\mu E_{12}$ and $E'_{21}=\nu E_{21}$, and the product $E'_{12}E'_{21}=\mu\nu E_{11}$ forces $\mu\nu=1$. Hence $$ E'_{12}=\mu E_{12},\qquad E'_{21}=\mu^{-1}E_{21},\qquad \mu\in\mathbb{C}^{\times}, $$ and the scaling is reciprocal: one mixed element grows while the other shrinks, and every product of the system is unchanged.

It is a conjugation. The scaling is not an extra freedom of the data; it is conjugation by a diagonal matrix: $$ D_\mu=\operatorname{diag}(\mu,1),\qquad D_\mu E_{12}D_\mu^{-1}=\mu E_{12},\qquad D_\mu E_{21}D_\mu^{-1}=\mu^{-1}E_{21},\qquad D_\mu E_{11}D_\mu^{-1}=E_{11}. $$ The systems with the two idempotents fixed are therefore one orbit of the diagonal torus, of complex dimension one.

The theorem. Every system of matrix units is the image of the standard one under an inner automorphism: there is an invertible $G$ with $E'_{ij}=GE_{ij}G^{-1}$ for all $i,j$. A system is the standard system read in another basis: $E'_{11}$ is the projector onto a line along a complementary line, a basis of the two lines conjugates $E_{11}$ to $E'_{11}$, and the two mixed elements are then fixed up to the reciprocal scaling just computed. The operator $L_GR_{G^{-1}}$ is the automorphism, and every automorphism of $M_2(\mathbb{C})$ is of that form by §The Automorphisms, so the frame carries no modulus beyond the conjugation.

The consequence for a pair. A change of frame is a conjugation of the carrier, and it acts on a pair by conjugating the two slots: $$ L_{GAG^{-1}}R_{GBG^{-1}}\bigl(GYG^{-1}\bigr)=G\,L_AR_B(Y)\,G^{-1}. $$ The operator therefore has the same trace, the same determinant and the same rank in every system of matrix units, and only its matrix changes, by similarity. This is the reason the formulas of §The Rank and the Invariants may be stated in one frame.

Two reciprocal scalings, not one. The pair in §The Scaling and the system here carry the same $\mathbb{C}^{\times}$ in two different roles. On the pair, $(\lambda A,\lambda^{-1}B)$ leaves the operator unchanged: a symmetry of the data. On the frame, $\mu$ leaves the system of products unchanged: a change of basis. Neither is a symmetry of the other, and the formulas are invariant under both.

The General Endomorphism

Four Pairs

The isomorphism. The map $$ M_2(\mathbb{C})\otimes M_2(\mathbb{C})\longrightarrow\operatorname{End}_{\mathbb{C}}(M_2(\mathbb{C})),\qquad A\otimes B^{\mathsf T}\longmapsto\bigl[Y\mapsto AYB\bigr] $$ is an isomorphism of vector spaces. It sends a pure tensor to a pair, and the pair set is the image of the pure tensors.

The bound. Every endomorphism of $M_2(\mathbb{C})$ is a sum of at most four pairs.

Proof. An element of $U\otimes V$ with $\dim U=\dim V=4$ has tensor rank at most four: view it as a linear map $V^*\to U$, whose rank is at most four, and a rank decomposition gives at most four pure tensors. Under the isomorphism, the pure tensors are the pairs, so the sum has at most four terms.

The count. The ambient space has complex dimension sixteen. A pair has eight parameters and seven effective. Under the isomorphism, the tensor rank is the rank of the coefficient matrix $c(M)$; a generic $4\times4$ matrix has rank four, so a generic endomorphism needs four pairs.

The Transpose

The operator. The transpose $Y\mapsto Y^{\mathsf T}$ is an endomorphism of $M_2(\mathbb{C})$. It is a sum of four pairs: $$ Y^{\mathsf T}=\sum_{i,j=1}^{2}E_{ij}YE_{ij}. $$

Proof. The $(k,l)$ entry of $E_{ij}YE_{ij}$ is $\delta_{ki}\delta_{lj}y_{ji}$, and the sum over $i,j$ is $y_{lk}$.

It is not a pair. Suppose $Y^{\mathsf T}=AYB$ for all $Y$. Then $E_{ji}=AE_{ij}B=a_ib_j^{\mathsf T}$ for all $i,j$. The equality $E_{11}=a_1b_1^{\mathsf T}$ forces the support of $a_1$ to be the first coordinate, and $E_{21}=a_1b_2^{\mathsf T}$ forces it to be the second; the contradiction shows that no pair gives the transpose.

It needs four. The coefficient matrix of the transpose is $$ \tfrac12\operatorname{diag}(1,-1,1,-1) $$ in the coefficient order $(t,x,y,z)$, of rank four. By the length theorem, four pairs are the minimum. The transpose is therefore a natural operator that is a sum of four pairs and no fewer.

The Pair Set Is Not a Subspace

The example. The operator $L_{\mathrm I_2}R_{\mathrm I_2}+L_{E_{11}}R_{E_{11}}$ is $Y\mapsto Y+y_{11}E_{11}$. Its coefficient matrix is the sum of the two rank-one matrices $[e_t][e_t]^{\mathsf T}$ and $[\tilde\Pi_+][\tilde\Pi_+]^{\mathsf T}$, where the brackets are the coefficient columns, of rank two; the operator is not a pair.

The cone. The pair set is the image of the bilinear map $(A,B)\mapsto B^{\mathsf T}\otimes A$, the cone over the Segre variety $\mathbb{P}^3\times\mathbb{P}^3$. It is closed under composition, by §Composition, and not under addition. A sum of two pairs need not be a pair, and up to four pairs may be needed for a general endomorphism.

What the Pair Buys and Does Not

It gives.

  • The exact count: one matrix per side, a pair per product, four pairs for the general endomorphism.
  • The exact rank formula $\operatorname{rank}L_AR_B=\operatorname{rank}A\cdot\operatorname{rank}B$.
  • The composition $(A,B)(C,D)=(AC,DB)$, the inverse, and the automorphism criterion $B=A^{-1}$.
  • The Kronecker display $B^{\mathsf T}\otimes A$ and the coefficient matrix $[Q][P]^{\mathsf T}$.
  • The seven effective parameters and the scaling orbit, and the reciprocal scaling of a system of matrix units.

It does not give.

  • A single matrix for a two-sided operator: the pair is minimal.
  • Closure under addition: the pair set is a cone, not a subspace.
  • A short description of the general endomorphism: four pairs may be needed, and the tensor rank can be four.

Summary

On $M_2(\mathbb{C})$, an operator is described by a count of $2\times2$ matrices. A left multiplication $L_A(Y)=AY$ is one matrix, a right multiplication $R_B(Y)=YB$ is one matrix, and a two-sided operator $L_AR_B(Y)=AYB$ is a pair. For a nonzero operator, the pair is unique up to the scaling $(\lambda A,\lambda^{-1}B)$, so the two-sided operators have seven effective complex parameters; the zero operator is the vertex of the cone.

A pair with a scalar slot is one-sided, and a pair with neither slot scalar is genuinely two-sided. The rank is $\operatorname{rank}A\cdot\operatorname{rank}B$, with values $0,1,2,4$; the trace is $4Q_tP_t$ and the determinant is $\langle\tilde Q\rangle_{\natural}^2\langle\tilde Q_2\rangle_{\natural}^2$.

The $4\times4$ matrix of the pair is the Kronecker product $B^{\mathsf T}\otimes A$, by the vectorisation identity $\operatorname{vec}(AYB)=(B^{\mathsf T}\otimes A)\operatorname{vec}(Y)$. The same matrix is the ordinary product of the left and right matrices $\mathrm I_2\otimes A$ and $B^{\mathsf T}\otimes\mathrm I_2$. The coefficient matrix is the outer product $[Q][P]^{\mathsf T}$, of rank at most one; it is not the operator matrix, which has rank $\operatorname{rank}A\cdot\operatorname{rank}B$.

The pairs compose by $(A,B)(C,D)=(AC,DB)$, invert by $(A,B)^{-1}=(A^{-1},B^{-1})$, and an invertible pair is an automorphism exactly when $B=A^{-1}$, the inner automorphisms $PGL_2(\mathbb{C})$. The pair set is a cone, not a subspace. A general endomorphism is a sum of at most four pairs, and the transpose is a sum of four pairs and no fewer.

In one line: one matrix per side, a pair for the product, four pairs for the general endomorphism.

Summary of Notation

Symbol Meaning
$\mathcal{P}=(\tilde\Pi_+,\vec\Upsilon_+,\vec\Upsilon_-,\tilde\Pi_-)$ the standard Peirce frame
$A=\mathsf M_2(\tilde Q)$, $B=\mathsf M_2(\tilde Q_2)$ the $2\times2$ matrices of the two elements
$L_A(Y)=AY$ the left multiplication
$R_B(Y)=YB$ the right multiplication
$L_AR_B(Y)=AYB$ the two-sided operator
$[M]_{\mathcal{P}}$ the $4\times4$ matrix of the operator $M$ in the frame
$L(\tilde Q)=\mathrm I_2\otimes A$ the left regular matrix
$R(\tilde Q_2)=B^{\mathsf T}\otimes\mathrm I_2$ the right regular matrix
$B^{\mathsf T}\otimes A$ the $4\times4$ matrix of the pair
$\operatorname{vec}(Y)=(y_{11},y_{21},y_{12},y_{22})^{\mathsf T}$ the column-stacking in the frame order
$E_{ij}$ the matrix units, $E_{ij}E_{kl}=\delta_{jk}E_{il}$
$\sigma_1,\sigma_2,\sigma_3$ the Pauli matrices
$\langle\tilde Q\rangle_{\natural}=Q_t^2+Q_x^2+Q_y^2+Q_z^2$ the quaternionic norm, $\det A$
$L_{e_\mu}R_{e_\nu}$ the sixteen basis products
$[Q],[P]$ the coefficient columns $(Q_t,Q_x,Q_y,Q_z)^{\mathsf T}$ and $(P_t,P_x,P_y,P_z)^{\mathsf T}$
$c(M)$ the coefficient matrix, $c_{\mu\nu}=\frac14\operatorname{Tr}((L_{e_\mu}R_{e_\nu})^{\dagger}M)$
$\ell(M)=\operatorname{rank}c(M)$ the length, the least number of pairs summing to $M$
$\operatorname{rank}$ the matrix rank
$PGL_2(\mathbb{C})$ the inner automorphism group

Further Reading