The One-Sided Operators

Introduction

The map $X\mapsto AX$ is an operator on the $n^{2}$-dimensional matrix space whose invariants are read off from $A$ without any computation: its rank is $n$ times the rank of $A$, its spectrum is the spectrum of $A$ repeated $n$ times, and its adjoint is the left multiplication by $A^{\mathsf T}$. The right multiplication behaves the same way, with the transpose in the parameter.

The article computes the kernel and the image, the rank, the spectrum and the Hilbert–Schmidt adjoint of $L_A$ and $R_B$, identifies the group of invertible left multiplications with $\mathrm{GL}_n(\mathbb{R})$ and its orbits, and shows that the two families are exchanged by the transpose of the argument. It assumes $\operatorname{End}(M_n(\mathbb{R}))$ as an Algebra: the Regular Representation for the algebra of operators and the commutant, and The Matrix Algebra $M_n(\mathbb{R})$ for the rank and the spectrum.

The Two Families

Definition. For $A,B\in M_n(\mathbb{R})$ put

$$ L_A(X)=AX,\qquad R_B(X)=XB . $$

Both are linear maps of $M_n(\mathbb{R})$ into itself, hence elements of $\operatorname{End}(M_n(\mathbb{R}))$, which has dimension $n^{4}$.

They are different operators. $L_A$ multiplies on the left and mixes the rows of $X$; $R_B$ multiplies on the right and mixes the columns. For $n\ge2$ neither family contains the other: $L_A=R_B$ forces $A=B=\lambda\mathrm{I}$.

In a basis they are Kronecker products. Stack the columns of $X$ into a column of length $n^{2}$. Then $\operatorname{vec}(AX)=(\mathrm{I}_n\otimes A)\operatorname{vec}X$ and $\operatorname{vec}(XB)=(B^{\mathsf T}\otimes\mathrm{I}_n)\operatorname{vec}X$, so

$$ [L_A]=\mathrm{I}_n\otimes A,\qquad [R_B]=B^{\mathsf T}\otimes\mathrm{I}_n . $$

Every computation below is a computation with one of these two matrices.

Composition and Inverses

Theorem. $L_AL_B=L_{AB}$ and $R_BR_C=R_{CB}$.

Proof. Apply both sides to $X$: $L_AL_B(X)=A(BX)=(AB)X$ and $R_BR_C(X)=(XC)B=X(CB)$.

The left family is a copy of the algebra, the right family a copy of the opposite algebra. So $A\mapsto L_A$ is an algebra isomorphism and $B\mapsto R_B$ is an algebra anti-isomorphism. Both maps are injective, since $L_A=0$ or $R_B=0$ forces the parameter to vanish by evaluation at $\mathrm{I}$.

Corollary. $L_A$ is invertible if and only if $A$ is invertible, and then $L_A^{-1}=L_{A^{-1}}$. The same holds on the right, with $R_B^{-1}=R_{B^{-1}}$.

Proof. If $A$ is invertible, $L_AL_{A^{-1}}=L_{\mathrm{I}}=\mathrm{id}$. Conversely, if $L_A$ is invertible with inverse $T$, then $A\,T(\mathrm{I})=L_A(T(\mathrm{I}))=\mathrm{I}$, so $A$ has a right inverse and is invertible.

Nilpotence is preserved. If $A^{k}=0$ then $L_A^{k}=L_{A^{k}}=0$, and conversely $L_A^{k}=0$ gives $A^{k}=L_A^{k}(\mathrm{I})=0$. So $L_A$ is nilpotent exactly when $A$ is, with the same index.

Kernel and Image

Theorem. $\ker L_A=\{X:AX=0\}$ is the set of matrices whose columns lie in $\ker A$, and

$$ \dim\ker L_A=n\,(n-\operatorname{rk}A). $$

Proof. Write $X$ as the block of its columns, $X=(x_1|\dots|x_n)$. Then $AX=(Ax_1|\dots|Ax_n)$, so $AX=0$ exactly when $Ax_j=0$ for every $j$. Each column ranges over the subspace $\ker A$, of dimension $n-\operatorname{rk}A$, independently of the others, so the space of such $X$ has dimension $n(n-\operatorname{rk}A)$.

Theorem. $\operatorname{im}L_A=\{AX\}$ is the set of matrices whose columns lie in $\operatorname{im}A$, and its dimension is $n\operatorname{rk}A$.

Proof. Same block reading: every column of $AX$ is $A$ times a vector, hence lies in $\operatorname{im}A$, and every matrix with columns in $\operatorname{im}A$ is $A$ times the matrix of preimages. The columns are free, so the dimension is $n\dim\operatorname{im}A$.

The two counts add up. $n\operatorname{rk}A+n(n-\operatorname{rk}A)=n^{2}$, which is the rank–nullity theorem for $L_A$ and an independent check of both formulas.

At $n=3$ with $\operatorname{rk}A=2$ the kernel has dimension $3$ and the image dimension $6$, and the numerical rank of $[L_A]$ is $6$. The three zero eigenvalues of $L_A$ are exactly the three dimensions of the kernel.

The Rank

Theorem. $\operatorname{rk}L_A=n\operatorname{rk}A$.

Proof. It is the dimension of the image, computed above. Equivalently, a rank-$r$ matrix factors as $A=UV^{\mathsf T}$ with $U$ of size $n\times r$ and $V$ of size $n\times r$ of rank $r$, and then $X\mapsto AX=U(V^{\mathsf T}X)$ is a composition of a map onto an $r$-dimensional space of rows and a map of rank $r$; the count gives $nr$.

The rank is multiplicative in the parameter only up to the factor $n$. $\operatorname{rk}L_{AB}=n\operatorname{rk}(AB)\le n\min(\operatorname{rk}A,\operatorname{rk}B)$, so the rank of the product can drop, exactly as the rank of $AB$ can drop.

The singular set is the image of the singular set. $L_A$ is singular exactly when $A$ is, so the set $\{A:\det A=0\}$ is the preimage under $A\mapsto L_A$ of the singular operators. The dimension count is consistent: $\det L_A=(\det A)^{n}$ vanishes exactly when $\det A$ does.

The Spectrum

Theorem. The eigenvalues of $L_A$ are the eigenvalues of $A$, each with multiplicity $n$.

Proof. In the Kronecker basis $[L_A]=\mathrm{I}_n\otimes A$. The eigenvalues of a Kronecker product $P\otimes Q$ are the products $\mu_i\nu_j$ of the eigenvalues of the factors. With $P=\mathrm{I}_n$ all $\mu_i=1$, so the eigenvalues are the $\nu_j$, each occurring $n$ times.

The characteristic polynomial is the $n$-th power. Over $\mathbb{C}$ the eigenvalues exist with multiplicity, and

$$ \chi_{L_A}(t)=\prod_{j=1}^{n}(t-\lambda_j)^{n}=\chi_A(t)^{n}. $$

The identity is a polynomial identity in the entries, so it holds over $\mathbb{R}$ as well, where some of the $\lambda_j$ may be complex conjugate pairs.

The trace and the determinant follow. Comparing coefficients,

$$ \operatorname{tr}L_A=n\operatorname{tr}A,\qquad \det L_A=(\det A)^{n}. $$

The Jordan structure is repeated, not mixed. Over $\mathbb{C}$ the operator $L_A$ is similar to the block diagonal matrix with $n$ copies of the Jordan form of $A$; the size of the largest block for an eigenvalue is unchanged. So the minimal polynomial of $L_A$ is the minimal polynomial of $A$.

The Hilbert–Schmidt Adjoint

Definition. $\langle X,Y\rangle=\operatorname{tr}(X^{\mathsf T}Y)=\sum_{i,j}X_{ij}Y_{ij}$ is the Hilbert–Schmidt inner product, positive definite on $M_n(\mathbb{R})$.

Theorem. $L_A^{*}=L_{A^{\mathsf T}}$ and $R_B^{*}=R_{B^{\mathsf T}}$.

Proof. For all $X,Y$,

$$ \langle L_AX,Y\rangle=\operatorname{tr}\big((AX)^{\mathsf T}Y\big)=\operatorname{tr}(X^{\mathsf T}A^{\mathsf T}Y)=\langle X,A^{\mathsf T}Y\rangle=\langle X,L_{A^{\mathsf T}}Y\rangle . $$

The adjoint is unique, so $L_A^{*}=L_{A^{\mathsf T}}$. The same computation with $Y\mapsto YB$ gives $\langle X,YB\rangle=\langle XB^{\mathsf T},Y\rangle$, that is $R_B^{*}=R_{B^{\mathsf T}}$.

The adjoint stays in the same family. It does not exchange left with right; it transposes the parameter. Numerically the two identities hold to $10^{-16}$ at $n=3$.

The transpose of the argument is what exchanges the families. For every $X$,

$$ (L_AX)^{\mathsf T}=X^{\mathsf T}A^{\mathsf T}=R_{A^{\mathsf T}}(X^{\mathsf T}), $$

so the left multiplication by $A$, followed by the transpose of the result, is the right multiplication by $A^{\mathsf T}$ applied to the transposed argument. Written as operators, with $\tau(X)=X^{\mathsf T}$,

$$ \tau L_A\tau=R_{A^{\mathsf T}},\qquad \tau R_B\tau=L_{B^{\mathsf T}}. $$

$\tau$ is an anti-automorphism of $\operatorname{End}$. It is linear and invertible, and $\tau(ST)=\tau(T)\tau(S)$ for operators $S,T$. The conjugation by an anti-automorphism turns a homomorphism into an anti-homomorphism, which is why the left family and the right family are exchanged and not preserved.

The Group and Its Orbits

Theorem. $\{L_A:A\in\mathrm{GL}_n(\mathbb{R})\}$ is a subgroup of $\mathrm{GL}_{n^{2}}(\mathbb{R})$ isomorphic to $\mathrm{GL}_n(\mathbb{R})$, acting on $M_n(\mathbb{R})$ by left multiplication.

Proof. Invertibility of $L_A$ is equivalent to invertibility of $A$, and the product rule $L_AL_B=L_{AB}$ makes the map an isomorphism of groups.

The orbit of the identity is the group itself. $\{L_A(\mathrm{I}):A\in\mathrm{GL}_n(\mathbb{R})\}=\{A\}=\mathrm{GL}_n(\mathbb{R})$. Under all left multiplications, without the invertibility, the orbit of $\mathrm{I}$ is all of $M_n(\mathbb{R})$, since $L_A(\mathrm{I})=A$.

The general orbit is classified by the kernel. If $Y=AX$ with $A$ invertible, then $\ker Y=\ker X$: indeed $\ker Y\supseteq\ker X$, and the ranks are equal because $L_A$ preserves the rank, so the kernels have the same dimension. Conversely, if $\ker X=\ker Y$, the rule $x_j\mapsto y_j$ is well defined on the column space of $X$, because a relation among the columns of $X$ is a vector of $\ker X$, hence one of $\ker Y$. It is an isomorphism onto the column space of $Y$, and it extends to an invertible $A$ with $AX=Y$.

The orbits are therefore indexed by subspaces. The orbit of $X$ is the set of matrices with kernel $\ker X$, so the orbit space is the disjoint union of the Grassmannians of $(n-r)$-dimensional subspaces of $\mathbb{R}^n$ for $r=0,\dots,n$. The largest orbit is $\mathrm{GL}_n(\mathbb{R})$ itself, of dimension $n^{2}$; the smallest is $\{0\}$.

The right action has the same structure, with the image in place of the kernel. Right multiplication preserves the row space, and its orbits are classified by the image of the transposed matrix.

The Numerical Check

The invariants at $n=3$. For a random $A$, the numerical rank of $[\mathrm{I}\otimes A]$ is $3\operatorname{rk}A$; the eigenvalues of $[\mathrm{I}\otimes A]$ are those of $A$ repeated three times, sorted, to machine precision; and the kernel has dimension $3(3-\operatorname{rk}A)$, which is $9-\operatorname{rk}L_A$.

The degenerate case. For a $3\times3$ matrix of rank two, built by making the third row equal the first, $L_A$ has rank $6$ and kernel dimension $3$, and three of its eigenvalues vanish, matching the three zero eigenvalues of the kernel directions.

The adjoint. At $n=3$ the identity $\langle AX,Y\rangle=\langle X,A^{\mathsf T}Y\rangle$ holds to $10^{-16}$, and $\langle XB,Y\rangle=\langle X,YB^{\mathsf T}\rangle$ to machine precision.

The exchange. The identity $\tau L_A\tau=R_{A^{\mathsf T}}$ holds exactly, to $0$ residual, and the orbit statement was checked for a pair of rank-two matrices with the same kernel: the interpolating $A$ satisfies $AX=Y$ to $1.3\times10^{-15}$.

Summary

The left multiplication $L_A(X)=AX$ and the right multiplication $R_B(X)=XB$ are operators on the $n^{2}$-dimensional matrix space, with $[L_A]=\mathrm{I}\otimes A$ and $[R_B]=B^{\mathsf T}\otimes\mathrm{I}$. The left family is a copy of $M_n(\mathbb{R})$, the right family a copy of the opposite algebra, and $L_AL_B=L_{AB}$, $R_BR_C=R_{CB}$. The kernel of $L_A$ is the set of matrices with columns in $\ker A$, of dimension $n(n-\operatorname{rk}A)$; the image is the set of matrices with columns in $\operatorname{im}A$, of dimension $n\operatorname{rk}A$; the rank is $n\operatorname{rk}A$; the spectrum is the spectrum of $A$ with each eigenvalue repeated $n$ times, so $\operatorname{tr}L_A=n\operatorname{tr}A$ and $\det L_A=(\det A)^{n}$. For the Hilbert–Schmidt form the adjoint is $L_{A^{\mathsf T}}$, and it stays in the left family; the two families are exchanged by conjugation with the transpose of the argument, $\tau L_A\tau=R_{A^{\mathsf T}}$. The invertible left multiplications form a group isomorphic to $\mathrm{GL}_n(\mathbb{R})$, whose orbit of $\mathrm{I}$ is $\mathrm{GL}_n(\mathbb{R})$ and whose general orbits are classified by the kernel.

Summary of Notation

Symbol Meaning
$L_A(X)=AX$ The left multiplication by $A$
$R_B(X)=XB$ The right multiplication by $B$
$[L_A]=\mathrm{I}\otimes A$ Its matrix in the column-stacking basis
$[R_B]=B^{\mathsf T}\otimes\mathrm{I}$ Its matrix in the same basis
$\ker L_A$ Matrices with columns in $\ker A$, of dimension $n(n-\operatorname{rk}A)$
$\operatorname{im}L_A$ Matrices with columns in $\operatorname{im}A$, of dimension $n\operatorname{rk}A$
$\langle X,Y\rangle=\operatorname{tr}(X^{\mathsf T}Y)$ The Hilbert–Schmidt inner product
$L_A^{*}=L_{A^{\mathsf T}}$ The adjoint of $L_A$
$\tau(X)=X^{\mathsf T}$ The transpose, an anti-automorphism of $\operatorname{End}$

Further Reading