The Two-Sided Operators and the Real Segre Cone
Introduction
The operator $X\mapsto AXB$ is linear in $X$ and bilinear in the pair $(A,B)$, and the set of all such operators is a cone of dimension $2n^{2}-1$ inside the $n^{4}$-dimensional space of operators. The cone is not a subspace. Three rules govern it: the product $T_{A,B}T_{C,D}=T_{AC,DB}$, the rank $\operatorname{rank}T_{A,B}=(\operatorname{rank}A)(\operatorname{rank}B)$, and the description of the rank-one part as the cone over the real Segre variety.
The article defines $T_{A,B}$ and proves the three rules. It gives the Segre variety by its $2\times2$ minors, separates its two rulings and reads them as the left and the right factor, and stratifies the cone by rank. It assumes The Matrix Algebra $M_n(\mathbb{R})$ for the matrix units, the rank and the transpose, and it defers the minimal number of summands to The Length and the Two Ranks.
Throughout $n\ge2$, $V=\mathbb{R}^{n}$ and $M_n=M_n(\mathbb{R})$.
The Two-Sided Operators
Definition. For $A,B\in M_n$ the two-sided operator $T_{A,B}$ is
$$ T_{A,B}(X)=AXB . $$
The left and the right multiplication are the cases $B=\mathrm{I}$ and $A=\mathrm{I}$:
$$ L_A=T_{A,\mathrm{I}},\qquad R_B=T_{\mathrm{I},B}. $$
The one-sided operators commute. $L_A R_B(X)=A(XB)=(AX)B=R_B L_A(X)$, so $T_{A,B}=L_A R_B=R_B L_A$. The two-sided operators are therefore the composites of a left and a right multiplication, and the two orders agree.
The map is bilinear in the pair. For scalars $\lambda,\mu$ and matrices $A,A',B,B'$,
$$ T_{\lambda A+\mu A',B}=\lambda T_{A,B}+\mu T_{A',B},\qquad T_{A,\lambda B+\mu B'}=\lambda T_{A,B}+\mu T_{A,B'} . $$
So $(A,B)\mapsto T_{A,B}$ is a bilinear map $M_n\times M_n\to\operatorname{End}(M_n)$, and it is not linear: $T_{A+A',B+B'}$ has the two mixed terms that $T_{A,B}+T_{A',B'}$ lacks.
The two-sided operators span the whole operator space. The product $E_{ij}XE_{kl}$ is $X_{jk}E_{il}$. So $T_{E_{ij},E_{kl}}$ reads the entry $(j,k)$ of $X$ and writes it in the position $(i,l)$. The $n^{4}$ operators $T_{E_{ij},E_{kl}}$ are linearly independent, because they carry $X$ to $n^{4}$ independent arrays, and they span $\operatorname{End}(M_n)$. Hence
$$ \dim\operatorname{End}(M_n)=n^{4}. $$
The Composition Rule
Theorem. For all $A,B,C,D\in M_n$,
$$ T_{A,B}T_{C,D}=T_{AC,DB}. $$
Proof. Evaluate on $X$: $T_{A,B}(T_{C,D}X)=A(CXD)B=(AC)X(DB)$.
The order of the right factors is reversed. The left factors multiply in the order of the composition, the right factors in the reverse order. So $\operatorname{End}(M_n)$ with this product is the tensor product of the algebra with its opposite algebra,
$$ \operatorname{End}(M_n)\cong M_n\otimes M_n^{\mathrm{op}},\qquad T_{A,B}\leftrightarrow A\otimes B . $$
The operator algebra is a full matrix algebra. The tensor product of two central simple algebras is central simple, and $M_n\otimes M_n^{\mathrm{op}}$ has dimension $n^{4}$. By Wedderburn it is $M_{n^{2}}(\mathbb{R})$. So the algebra of operators on $M_n$ is a full matrix algebra of degree $n^{2}$, and its simplicity is the one proved in The Ideals, the Centre and the Simplicity of $M_n(\mathbb{R})$.
The identity operator is the pure tensor $\mathrm{I}\otimes\mathrm{I}$. $T_{\mathrm{I},\mathrm{I}}(X)=X$. It is a single two-sided operator, and this is the reason for the count of the length in The Length and the Two Ranks.
The Rank Formula
The vectorisation turns the two-sided operator into a Kronecker product. Write $\operatorname{vec}X$ for the column of the entries of $X$, taken column by column. Then
$$ \operatorname{vec}(AXB)=(B^{\mathsf T}\otimes A)\operatorname{vec}X . $$
The identity is a comparison of entries: the entry $(i,j)$ of $AXB$ is $\sum_{k,l}A_{ik}X_{kl}B_{lj}$, and the coefficient of $X_{kl}$ is $A_{ik}B_{lj}$, which is the entry of $B^{\mathsf T}\otimes A$ in the row $(i,j)$ and the column $(k,l)$.
Theorem. $\operatorname{rank}T_{A,B}=(\operatorname{rank}A)(\operatorname{rank}B)$.
Proof. By the vectorisation the rank of $T_{A,B}$ is the rank of $B^{\mathsf T}\otimes A$. Factor $A=PQ$ and $B^{\mathsf T}=UV$ with $P,Q,U,V$ of full rank $\operatorname{rank}A$ and $\operatorname{rank}B$. Then
$$ B^{\mathsf T}\otimes A=(U\otimes P)(V\otimes Q), $$
where $U\otimes P$ has full column rank $\operatorname{rank}A\cdot\operatorname{rank}B$ and $V\otimes Q$ has full row rank of the same value. A product of a full-column-rank and a full-row-rank matrix has the common rank, so the rank is the product.
The rank is a product of two integers between $1$ and $n$. The values attained are $rs$ with $1\le r,s\le n$. At $n=3$ this gives $1,2,3,4,6,9$, and the value $5$ is not attained, although it is a rank of an operator on $M_3(\mathbb{R})$. Measured at $n=3$ on random pairs of ranks $1,1$; $1,2$; $2,3$; $3,3$, the rank of $T_{A,B}$ is $1,2,6,9$.
The Segre Variety of Rank-One Matrices
A rank-one matrix is a product of a column and a row. $A=ac^{\mathsf T}$ with $a,c\in V$ nonzero, and the map
$$ \mathbb{P}^{n-1}\times\mathbb{P}^{n-1}\longrightarrow\mathbb{P}(M_n),\qquad ([a],[c])\longmapsto[ac^{\mathsf T}] $$
is injective, because $ac^{\mathsf T}$ determines the line $\mathbb{R}a$ as its image and the hyperplane $c^{\perp}$ as its kernel. Its image is the Segre variety, of dimension
$$ (n-1)+(n-1)=2n-2 . $$
The defining equations are the $2\times2$ minors. A matrix $A$ has rank at most one exactly when
$$ A_{ij}A_{kl}-A_{il}A_{kj}=0 $$
for all $i,j,k,l$. There are $\binom{n}{2}^{2}$ such minors, nine at $n=3$, and they generate the ideal of the variety. At $n=2$ the single minor is $\det A$, and the Segre variety is the quadric cone $\{\det=0\}$ of $\mathbb{P}^{3}$. Measured on a rank-one matrix at $n=3$, all nine minors vanish to order $10^{-12}$.
The real points are the products of real lines. Every real point of the Segre variety is $ac^{\mathsf T}$ with $a,c$ real and nonzero. The image is the real line $\mathbb{R}a$ and the kernel is the real hyperplane $c^{\perp}$. So the real locus is exactly the set of rank-one real matrices, and it is Zariski dense.
The Two Rulings
Through a point of the Segre variety pass two families of lines. Fix $c$ and let $a$ run: the matrices $ac^{\mathsf T}$ form the set of rank-one matrices with kernel $c^{\perp}$, and its projectivisation is a line $\mathbb{P}^{n-1}$. Fix $a$ and let $c$ run: the matrices with image $\mathbb{R}a$, again a line. The two families are the rulings, each a copy of $\mathbb{P}^{n-1}$.
The rulings are the left and the right factor. The first family is obtained by fixing the right factor $c$ and varying the left factor $a$; the second by fixing $a$ and varying $c$. A point lies on one line of each family. So the rulings are the fibres of the two projections of $\mathbb{P}^{n-1}\times\mathbb{P}^{n-1}$, and the two projections are the left factor and the right factor of the product.
The transpose exchanges the two rulings. It sends $ac^{\mathsf T}$ to $ca^{\mathsf T}$, so a point is fixed only when $a$ and $c$ are proportional, at the symmetric rank-one matrices $aa^{\mathsf T}$.
The Real Segre Cone
The image of $(A,B)\mapsto T_{A,B}$ is a cone. If $T=T_{A,B}$ then $\lambda T=T_{\lambda A,B}$, so the image is closed under scaling. It is the real Segre cone
$$ \mathcal{C}=\{T_{A,B}:A,B\in M_n\}\subset\operatorname{End}(M_n). $$
Two pairs give the same operator only up to a common scalar. If $A\otimes B=A'\otimes B'$ with both tensors nonzero, then $A'=\lambda A$ and $B'=\lambda^{-1}B$. So the map
$$ \mathbb{P}(M_n)\times\mathbb{P}(M_n)\longrightarrow\mathbb{P}(\operatorname{End}(M_n)),\qquad ([A],[B])\longmapsto[T_{A,B}] $$
is injective. Its image has dimension
$$ (n^{2}-1)+(n^{2}-1)=2n^{2}-2, $$
and the cone $\mathcal{C}$ over it has dimension $2n^{2}-1$. At $n=2$ that is $7$ inside the $16$-dimensional operator space; at $n=3$ it is $17$ inside $81$.
The cone is not a subspace. The diagonal projection $D(X)=\operatorname{diag}(X)$ is not two-sided. If $D=T_{A,B}$ then $n=\operatorname{rank}D=(\operatorname{rank}A)(\operatorname{rank}B)$, so one factor has rank one; say $\operatorname{rank}A=1$. Then the image of $T_{A,B}$ lies in the image of $A$, which is a line, while the image of $D$ is the diagonal, of dimension $n$. So $\mathcal{C}$ is not closed under addition: $D=T_{E_{11},E_{11}}+\cdots+T_{E_{nn},E_{nn}}$ is a sum of $n$ two-sided operators and is not one. A least-squares fit of $D$ by a two-sided operator at $n=3$ leaves a residual of $1.41$ against a target norm of $1.73$.
The cone spans the whole space. The $n^{4}$ operators $T_{E_{ij},E_{kl}}$ are a basis, so the linear span of $\mathcal{C}$ is $\operatorname{End}(M_n)$.
The Rank-One Two-Sided Operators
Theorem. $\operatorname{rank}T_{A,B}=1$ exactly when $\operatorname{rank}A=\operatorname{rank}B=1$.
The theorem is the product formula with both factors equal to one. On such a pair, with $A=ac^{\mathsf T}$ and $B=bd^{\mathsf T}$,
$$ T_{A,B}(X)=a\,(c^{\mathsf T}Xb)\,d^{\mathsf T}, $$
a rank-one operator whose image is the line $\mathbb{R}ad^{\mathsf T}$ and whose kernel is the hyperplane $\{X:c^{\mathsf T}Xb=0\}$. Measured on random rank-one pairs at $n=2$ and $n=3$, the rank of $T_{A,B}$ is one in both cases.
Their set is a product of two Segre varieties. The pair $([A],[B])$ runs over the product of the Segre variety with itself, since each factor must have rank one. The projectivised set has dimension
$$ (2n-2)+(2n-2)=4n-4, $$
and the cone over it has dimension $4n-3$. Each factor is a copy of $\mathbb{P}^{n-1}\times\mathbb{P}^{n-1}$.
The two rulings of the larger product are again the left and the right factor. Fix $[B]$ and let $[A]$ run over the Segre variety: a linear space of dimension $2n-2$ inside the cone. Fix $[A]$ and let $[B]$ run: a second family of the same dimension. A rank-one two-sided operator lies on one member of each family; the left factor labels one family and the right factor the other.
The Stratification by Rank
The stratum of rank $rs$ is the image of the pairs of ranks $r$ and $s$. Let $M_{n}^{\le r}$ be the set of matrices of rank at most $r$, of dimension $2nr-r^{2}$. The pairs with $\operatorname{rank}A\le r$ and $\operatorname{rank}B\le s$ form a set of dimension $(2nr-r^{2})+(2ns-s^{2})$, and the scaling $T_{\lambda A,B}=\lambda T_{A,B}$ removes one dimension. So
$$ \dim\{T_{A,B}:\operatorname{rank}A\le r,\ \operatorname{rank}B\le s\}=(2nr-r^{2})+(2ns-s^{2})-1 . $$
At $r=s=1$ this is $4n-3$, the cone of the rank-one operators; at $r=s=n$ it is $2n^{2}-1$, the whole cone. The two ends match the earlier counts.
The rank strata of the cone are not the rank strata of the operator space. The operator space has strata of every rank $1,\dots,n^{2}$, of dimension $k(2n^{2}-k)$. The cone meets only the ranks that factor, and inside a fixed rank $rs$ it is a proper subset of the stratum. At $n=3$ the stratum of rank $5$ has dimension $65$, and the cone has no point in it.
The rank bounds the length. Each two-sided operator has rank at most $n^{2}$, and a sum of $k$ of them has rank at most $kn^{2}$. So an operator of rank $m$ needs at least $m/n^{2}$ summands. The exact count, and the proof that every operator is a sum of at most $n^{2}$ two-sided operators, is The Length and the Two Ranks.
Summary
The two-sided operator $T_{A,B}(X)=AXB$ is the composite of a left and a right multiplication, and $(A,B)\mapsto T_{A,B}$ is bilinear with image the real Segre cone of dimension $2n^{2}-1$ inside the $n^{4}$-dimensional operator space. The operators compose by $T_{A,B}T_{C,D}=T_{AC,DB}$, so the operator algebra is $M_n\otimes M_n^{\mathrm{op}}\cong M_{n^{2}}(\mathbb{R})$. The rank multiplies, $\operatorname{rank}T_{A,B}=(\operatorname{rank}A)(\operatorname{rank}B)$, so the ranks attained in the cone are the products $rs$. The rank-one members are the pairs of rank-one matrices, and their set is the cone over the product of two real Segre varieties, of dimension $4n-4$ projectively and $4n-3$ affinely. The Segre variety itself has dimension $2n-2$, is defined by the $2\times2$ minors, and its two rulings are the left and the right factor. The cone is not a subspace: the diagonal projection is a sum of $n$ two-sided operators and is not one.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $T_{A,B}(X)=AXB$ | the two-sided operator |
| $L_A=T_{A,\mathrm{I}}$, $R_B=T_{\mathrm{I},B}$ | the left and the right multiplication |
| $\operatorname{End}(M_n)$ | the algebra of operators on $M_n$, of dimension $n^{4}$ |
| $\mathcal{C}$ | the real Segre cone, of dimension $2n^{2}-1$ |
| $ac^{\mathsf T}$ | a rank-one matrix, the product of a column and a row |
| $\mathbb{P}^{n-1}\times\mathbb{P}^{n-1}$ | the Segre variety, of dimension $2n-2$ |
| $A_{ij}A_{kl}-A_{il}A_{kj}$ | the $2\times2$ minors, the equations of the Segre variety |
| rulings | the two families of $\mathbb{P}^{n-1}$ on the Segre variety, the left and the right factor |
| $M_{n}^{\le r}$ | the matrices of rank at most $r$, of dimension $2nr-r^{2}$ |
Further Reading
- Richard S. Pierce, Associative Algebras (Springer, 1982), for the tensor product of matrix algebras, the opposite algebra and the two-sided multiplication.
- Igor R. Shafarevich, Basic Algebraic Geometry 1 (Springer, 2013), for the Segre varieties, their rulings and their minors.
- The Matrix Algebra $M_n(\mathbb{R})$, for the matrix units, the rank and the transpose.
- The Ideals, the Centre and the Simplicity of $M_n(\mathbb{R})$, for the simplicity of the operator algebra.
- The Length and the Two Ranks, for the minimal number of summands and the two ranks of an operator.