The One-Sided Operators of $M_n(\mathbb{C})$

Introduction

The question. A matrix $A$ acts on the matrix algebra by left multiplication, $X\mapsto AX$, and a matrix $B$ acts by right multiplication, $X\mapsto XB$. Both are operators on the $n^{2}$ complex dimensions of $M_{n}(\mathbb{C})$, hence matrices of size $n^{2}$. This article computes those matrices, their rank, their trace and their determinant, and describes the two subsets of $\operatorname{End}(M_{n})$ that they fill.

The result in one line. Left multiplication is $\mathrm I_{n}\otimes A$, right multiplication is $B^{\mathsf T}\otimes\mathrm I_{n}$, the two subsets are copies of $M_{n}$ and of $M_{\mathrm{opp},n}$ that commute and meet in the scalars, and each is the commutant of the other.

Why the article is needed. The two-sided operators are built from the two one-sided families, and every count of the companion articles starts from the dimensions $n^{2}+n^{2}-1=2n^{2}-1$ and $n^{4}$. The one-sided case is also the only case in which the invariants of the operator are read off the parameter without any tensor rank computation: the rank of $L_{A}$ is $n$ times the rank of $A$, and the determinant is the $n$-th power.

Boundaries. The algebra and its trace and determinant are The Matrix Algebra $M_n(\mathbb{C})$. The general construction is The Regular Representation in Dimension $n$. The two-sided products are The Two-Sided Operators and the Segre Cone of $M_n(\mathbb{C})$. At $n=2$ the same matrices in the Peirce frame are the two four-dimensional copies of $M_4(\mathbb{C})$ Pauli Representation : Operators, and the pair description is $M_2(\mathbb{C})$ Pauli Representation : Operators.

Conventions. The algebra is $A=M_{n}(\mathbb{C})$, of complex dimension $N=n^{2}$. An element $X$ is written as a column $\operatorname{vec}X$ by stacking its columns, so that $\operatorname{vec}(AX)=(\mathrm I_{n}\otimes A)\operatorname{vec}X$ and $\operatorname{vec}(XB)=(B^{\mathsf T}\otimes\mathrm I_{n})\operatorname{vec}X$. An operator on the algebra is an $N\times N$ matrix; the identity operator is $\mathrm I_{N}$. The transpose of a matrix is ${}^{\mathsf T}$ and the conjugate transpose is ${}^{*}$.

The Left Multiplication

The Matrix

Proposition (the Kronecker form). In the column-stacking basis, $$ [L_{A}]=\mathrm I_{n}\otimes A, $$ the block matrix of $n^{2}$ blocks of size $n$, with $A$ on the diagonal and zeros elsewhere.

Proof. The column $\operatorname{vec}(AX)$ is the stack of the columns of $AX$. The $j$-th column of $AX$ is $A$ times the $j$-th column of $X$, so the operator acts as $A$ on each block of $n$ consecutive coordinates and does not mix the blocks. This is the block-diagonal matrix with $n$ copies of $A$.

The display. For $n=2$ and $A=\begin{pmatrix}a&b\\c&d\end{pmatrix}$, $$ [L_{A}]=\begin{pmatrix} a&b&0&0\\ c&d&0&0\\ 0&0&a&b\\ 0&0&c&d \end{pmatrix}, $$ which is the form used in the biquaternion articles, where $A=\mathsf M_{2}(\tilde Q)$ and the four entries are the coordinates of the element.

The Invariants

Theorem (rank, trace, determinant). For every $A\in M_{n}(\mathbb{C})$, $$ \operatorname{rank}L_{A}=n\operatorname{rank}A,\qquad \operatorname{tr}L_{A}=n\operatorname{tr}A,\qquad \det L_{A}=(\det A)^{n}. $$

Proof. The matrix $\mathrm I_{n}\otimes A$ is block diagonal with $n$ blocks equal to $A$, so its rank is the sum of the ranks, its trace the sum of the traces, and its determinant the product of the determinants.

The spectrum. If the eigenvalues of $A$ are $\lambda_{1},\dots,\lambda_{n}$ with algebraic multiplicities, then the eigenvalues of $L_{A}$ are the same numbers, each repeated $n$ times. The spectral projector of $L_{A}$ at $\lambda$ is the left multiplication $L_{P}$ by the spectral projector of $A$ at $\lambda$, because $\mathrm I_{n}\otimes P$ is the matrix of $L_{P}$.

The minimal polynomial does not grow. The minimal polynomial of $L_{A}$ is the minimal polynomial of $A$, of degree at most $n$, while $L_{A}$ has size $n^{2}$. For $n\ge2$ the degree is strictly smaller than the size, so the regular matrix is never nonderogatory: there is no vector $X$ whose images under the powers of $L_{A}$ fill the algebra. This is the operator form of the fact that the left multiplication sees the algebra through a single element, and it is why the two-sided operators are needed for a general endomorphism.

The Kernel and the Image

The kernel. The operator $L_{A}$ kills exactly the matrices whose columns lie in the kernel of $A$: $$ \ker L_{A}=\{X:AX=0\},\qquad \dim\ker L_{A}=n\,(n-\operatorname{rank}A)=N-n\operatorname{rank}A . $$ The dimension is the complement of the rank, as it must be.

The image. The image is the set of matrices whose columns lie in the column space of $A$: $$ \operatorname{im}L_{A}=A\cdot M_{n}(\mathbb{C})=\{AX\}, $$ of dimension $n\operatorname{rank}A$. In particular the image is the whole algebra exactly when $A$ is invertible, and the operator $L_{A}$ is invertible exactly then, with $L_{A}^{-1}=L_{A^{-1}}$.

The left ideal. The image is a left ideal: it is closed under left multiplication by any matrix, because $C(AX)=(CA)X$. The kernel is not an ideal: it is closed under right multiplication, $AX=0$ gives $AXC=0$. The two properties are the module-theoretic reading of the operator.

The Right Multiplication

The Matrix

Proposition (the Kronecker form). In the column-stacking basis, $$ [R_{B}]=B^{\mathsf T}\otimes\mathrm I_{n}. $$

Proof. The $j$-th column of $XB$ is $X$ times the $j$-th column of $B$, that is $\sum_{k}B_{kj}x_{k}$ with $x_{k}$ the $k$-th column of $X$. The $j$-th block of the output is therefore $\sum_{k}B_{kj}x_{k}$, whose coefficients in the input blocks are the $j$-th column of $B$; the block matrix of the operator is thus $B^{\mathsf T}$, and since each block is multiplied by $\mathrm I_{n}$ on the left, the matrix is $B^{\mathsf T}\otimes\mathrm I_{n}$.

The display. For $n=2$, $$ [R_{B}]=\begin{pmatrix} b_{11}&0&b_{21}&0\\ 0&b_{11}&0&b_{21}\\ b_{12}&0&b_{22}&0\\ 0&b_{12}&0&b_{22} \end{pmatrix}, $$ a block matrix whose four blocks are scalar multiples of $\mathrm I_{2}$. The transpose is not cosmetic: it is the difference between the left copy and the right copy, and it disappears only in the scalar case.

The transpose is the reason both letters are needed. The left copy is block diagonal with the element repeated; the right copy is block scalar with the transpose. A matrix that is both is a scalar matrix, and this is the computation of the next section.

The Invariants

Proposition. For every $B$, $$ \operatorname{rank}R_{B}=n\operatorname{rank}B,\qquad \operatorname{tr}R_{B}=n\operatorname{tr}B,\qquad \det R_{B}=(\det B)^{n}. $$

Proof. The matrix $B^{\mathsf T}\otimes\mathrm I_{n}$ has $n^{2}$ blocks, each a scalar multiple of $\mathrm I_{n}$. Its rank, trace and determinant are the same functions of $B^{\mathsf T}$ as those of $\mathrm I_{n}\otimes B$ are of $B$, and $\operatorname{rank}B^{\mathsf T}=\operatorname{rank}B$, $\operatorname{tr}B^{\mathsf T}=\operatorname{tr}B$, $\det B^{\mathsf T}=\det B$.

The right copy is an anti-representation. The composition is $R_{B}R_{C}=R_{CB}$, so the map $B\mapsto R_{B}$ reverses the order. The set $\{R_{B}\}$ is a subalgebra of $\operatorname{End}(M_{n})$, isomorphic to $M_{\mathrm{opp},n}$, and $M_{\mathrm{opp},n}\cong M_{n}$ through the transpose. The transpose is therefore the canonical comparison between the two copies, and it is an isomorphism of algebras and not an identification of subsets.

The Flip

Proposition (the two copies are exchanged by the flip). Let $\mathsf F$ be the flip of the two tensor slots, $\mathsf F(X\otimes Y)\mathsf F=Y\otimes X$, an involution of the $n^{2}$ dimensions with $\mathsf F=\mathsf F^{\mathsf T}=\mathsf F^{-1}$. Then $$ \mathsf F\,[L_{A}]\,\mathsf F=[R_{A^{\mathsf T}}],\qquad \mathsf F\,[R_{B}]\,\mathsf F=[L_{B^{\mathsf T}}]. $$

Proof. The flip sends $\mathrm I_{n}\otimes A$ to $A\otimes\mathrm I_{n}$, and $A\otimes\mathrm I_{n}=(A^{\mathsf T})^{\mathsf T}\otimes\mathrm I_{n}=[R_{A^{\mathsf T}}]$. The second identity is the first applied to $B^{\mathsf T}$.

The flip is an operator and not an element. It lives in $\operatorname{End}(M_{n}(\mathbb{C}))$ and not in the algebra, it exchanges the two copies, and it preserves neither the product of the algebra nor a single letter of the two families. In the biquaternion case the same role is played by the permutation $\mathsf S$ of the second and third coordinates of the frame, and the identity there reads $R_{\mathrm{pau}}(\tilde Q)=\mathsf S\,L_{\mathrm{pau}}(\tilde Q)^{\mathsf T}\mathsf S^{-1}$, which is the frame form of the proposition above.

The Two Copies

The Intersection

Theorem (the two copies meet in the scalars). For $n\ge2$, $$ \{L_{A}\}\cap\{R_{B}\}=\mathbb{C}\cdot\mathrm I_{N}. $$

Proof. Suppose $AX=XB$ for every $X$. With $X=I$ this gives $A=B$; then $AX=XA$ for every $X$, and the centre of $M_{n}(\mathbb{C})$ is $\mathbb{C}I$. Conversely a scalar matrix acts by the same scalar on the left and on the right.

The dimension count. Two subspaces of dimensions $n^{2}$ and $n^{2}$ with intersection of dimension one span a subspace of dimension $2n^{2}-1$. This is the third rung of the ladder, and it is also the dimension of the cone of two-sided products, although the two sets are different: the sum of the two copies is a subspace, the cone is not.

The Double Centraliser

Theorem. Each copy is the commutant of the other in $\operatorname{End}(M_{n}(\mathbb{C}))$.

Proof. A right multiplication commutes with every left multiplication, because $A(XB)=(AX)B$. Conversely let $F$ commute with every $L_{C}$ and put $B=F(I)$; then $F(X)=F(L_{X}I)=L_{X}F(I)=XB$, so $F=R_{B}$.

The reading. The theorem says that an operator that commutes with all left multiplications is a right multiplication, and this is the reason the two-sided family is the natural next object: a general endomorphism is a sum of products of one member of each copy, and the number of terms needed is the length of the companion article.

The Commutant of One Element

Proposition. Let $C(A)=\{Y\in M_{n}(\mathbb{C}):AY=YA\}$ be the centraliser of $A$ in the algebra. Then the commutant of $L_{A}$ in $\operatorname{End}(M_{n}(\mathbb{C}))$ is the set of operators $F(X)=\sum_{i}Y_{i}XZ_{i}$ with $Y_{i}\in C(A)$ and $Z_{i}$ arbitrary, of complex dimension $$ n^{2}\dim C(A). $$

Proof. Write an operator as $F=\sum_{i}Y_{i}\otimes Z_{i}$ in the tensor form $M_{n}\otimes M_{n}$, so that $F(X)=\sum_{i}Y_{i}XZ_{i}$. Then $$ F(AX)-AF(X)=\sum_{i}Y_{i}A X Z_{i}-\sum_{i}AY_{i}XZ_{i}=\sum_{i}(Y_{i}A-AY_{i})XZ_{i}, $$ and $F$ commutes with $L_{A}$ exactly when $\sum_{i}(Y_{i}A-AY_{i})\otimes Z_{i}=0$. With the $Z_{i}$ independent this forces $Y_{i}\in C(A)$, and the space of such $F$ is $C(A)\otimes M_{n}$; the two factors give the same dimension.

The two extremes. For a cyclic element, $C(A)=\mathbb{C}[A]$ of dimension $n$, and the commutant of $L_{A}$ has dimension $n^{3}$. For a scalar element, $C(A)=M_{n}(\mathbb{C})$ of dimension $n^{2}$, and the commutant is the whole ambient space of dimension $n^{4}$. The commutant of a single left multiplication is strictly larger than the right copy for every $n\ge2$ and every $A$, since $n^{2}\dim C(A)\ge n^{3}>n^{2}$, and its size is the size of the centraliser.

The table.

$A$ $\dim C(A)$ $\dim\{F:FC=CF\}$ for $C=L_{A}$ at $n=3$
all left multiplications taken together --- $n^{2}$ $9$
a scalar element $n^{2}$ $n^{4}$ $81$
a cyclic element, e.g. a single Jordan block $n$ $n^{3}$ $27$
$\operatorname{diag}(1,1,2)$ at $n=3$ $5$ $45$ $45$

The first row is the theorem and the other three are propositions about a single element. The commutant of the whole copy is the right copy of dimension $n^{2}$; the commutant of a single member is strictly larger for $n\ge2$, the least being $n^{3}$ at a cyclic element. The second row is the degenerate case in which $L_{A}$ is the identity operator and every operator commutes with it.

The Compression

The compression to the $2\times2$ model. At $n=2$ the algebra is the biquaternion algebra and the element is a $2\times2$ matrix $\mathsf M_{2}(\tilde Q)$. In the $2\times2$ model the left matrix compresses to $A$ and the right matrix compresses to $B^{\mathsf T}$: each is two identical blocks of the same $2\times2$ matrix. The two-sided product does not compress, because it is the $4\times4$ product $B^{\mathsf T}\otimes A$ and both tensor factors are needed. The two one-sided actions compress one at a time, and this is why the pair description of the biquaternion category is written with one matrix per side and a pair for the product.

Summary

Left multiplication by $A$ is the block-diagonal matrix $\mathrm I_{n}\otimes A$ with $n$ copies of $A$; right multiplication by $B$ is the block-scalar matrix $B^{\mathsf T}\otimes\mathrm I_{n}$. Both have rank $n\operatorname{rank}$, trace $n\operatorname{tr}$ and determinant $(\det)^{n}$ of the parameter, and both have the spectrum of the parameter repeated $n$ times.

The two families are copies of $M_{n}$ and of $M_{\mathrm{opp},n}$; they commute, they meet in the scalars, their sum has dimension $2n^{2}-1$, and each is the commutant of the other. The commutant of a single left multiplication is $C(A)\otimes M_{n}$, of dimension $n^{2}\dim C(A)$: $n^{3}$ for a cyclic element and $n^{4}$ for a scalar one.

The regular matrix is never nonderogatory for $n\ge2$, because its minimal polynomial has degree at most $n$ against a size of $n^{2}$. The one-sided operators are the two ends of the construction, and the two-sided products of the next article are the smallest family that reaches the rest of the ambient space.

Summary of Notation

Symbol Meaning
$M_{n}(\mathbb{C})$ the algebra, of complex dimension $N=n^{2}$
$L_{A}(X)=AX$ the left multiplication
$R_{B}(X)=XB$ the right multiplication
$[L_{A}]=\mathrm I_{n}\otimes A$ its matrix in the column-stacking basis
$[R_{B}]=B^{\mathsf T}\otimes\mathrm I_{n}$ its matrix in the same basis
$\mathrm I_{N}$ the identity operator
$\mathsf F$ the flip of the two tensor slots
$C(A)$ the centraliser of $A$ in $M_{n}(\mathbb{C})$
$\operatorname{End}_{L}$, $\operatorname{End}_{R}$ the two copies inside $\operatorname{End}(M_{n})$

Further Reading