The Enveloping Operator Algebra of $M_n(\mathbb{C})$ and the Bi-module Structure
Introduction
The question. $M_n(\mathbb{C})$ is a module over the algebra on the left and over the opposite algebra on the right. What is the algebra of that double action, and how much of the operator algebra does it fill?
The answer in one line. The enveloping algebra is the tensor square, $$ M_n\otimes M_n^{\mathrm{op}}\cong M_{n^2}(\mathbb{C}), $$ which is the whole operator algebra, while the span of the two copies inside it has dimension only $2n^2-1$, leaving an orthogonal complement of dimension $(n^2-1)^2$.
Why the article is needed. Two statements look contradictory and are both true. The two-sided action generates everything, and the two-sided action spans almost nothing. The resolution is the difference between the algebra generated and the space spanned, and the exact size of the gap. Without it, a reader who computes $\dim(GL_n\times GL_n/\mathbb{C}^{\times})=2n^2-1$ and $\dim\operatorname{End}(M_n)=n^4$ has no way to see where the missing dimensions live.
Boundaries. The families are The Endomorphisms of $M_n(\mathbb{C})$; the group of the action is The Real Structures of the Operators of $M_n(\mathbb{C})$; the double centraliser property is used in The Transposition of the Two Sides of $M_n(\mathbb{C})$. The biquaternion original is The Enveloping Operator Algebra of the Biquaternion Algebra and the Bi-module Structure.
Conventions. $A=M_n(\mathbb{C})$, $A^{\mathrm{op}}$ the opposite algebra, and $n\ge2$. The carrier is $A$ itself, with the left action $L_A(X)=AX$ and the right action $R_B(X)=XB$. Operators are indexed by pairs: $Y_{ij,kl}$ is the entry of $Y\in\operatorname{End}(A)$ at the pair $(i,j)$ of rows and $(k,l)$ of columns.
The Bimodule
Proposition. $A$ is an $(A,A)$-bimodule under $(a,b)\cdot X=aXb$, the regular bimodule, and the two actions commute.
Proof. $a(Xb)=(aX)b$ is associativity, and $(aX)b=a(Xb)$ is the same identity read twice. $\square$
Proposition. $\operatorname{End}_{A}(A)\cong A^{\mathrm{op}}$ and $\operatorname{End}_{A^{\mathrm{op}}}(A)\cong A$.
Proof. An $A$-linear map $T$ satisfies $T(aX)=aT(X)$; at $X=\mathrm I$ this gives $T(a)=aT(\mathrm I)$, so $T=R_{T(\mathrm I)}$, and every right multiplication is $A$-linear. $\square$
The double centraliser. The commutant of $L(A)$ in $\operatorname{End}(A)$ is $R(A)$, and conversely.
Proof. If $TL_a=L_aT$ for all $a$ then $T(aX)=aT(X)$; at $X=\mathrm I$, $T(a)=aT(\mathrm I)$, so $T=R_{T(\mathrm I)}$. $\square$
The consequence. The pair $(L(A),R(A))$ is a pair of mutual commutants. Each copy is the full centraliser of the other, so no operator outside the right copy can commute with the whole left copy.
The Enveloping Algebra
Proposition. The multiplication map $\mu:A\otimes A^{\mathrm{op}}\to\operatorname{End}(A)$, $\mu(a\otimes b)=L_aR_b$, is an isomorphism of algebras.
Proof. It is multiplicative because $(L_aR_b)(L_cR_d)=L_{ac}R_{db}$ and $ac\otimes db$ is the product in $A\otimes A^{\mathrm{op}}$. It is injective because the $n^4$ elementary tensors map to linearly independent matrix units, and both sides have dimension $n^4$. $\square$
The reading. The enveloping algebra of the bimodule is not a subalgebra of the operator algebra; it is the operator algebra. Every linear map of $M_n$ into itself is a sum of two-sided multiplications, and the two-sided family is a spanning set of the enveloping algebra.
The module structure. As a module over $A\otimes A^{\mathrm{op}}\cong M_{n^2}(\mathbb{C})$, the carrier $A$ is the standard module $\mathbb{C}^{n^2}$. It is simple, and its endomorphism ring is $\mathbb{C}$.
Proof. A bimodule endomorphism is an operator commuting with all $L_a$ and all $R_b$, hence scalar by the two double centraliser statements. Simplicity: $A$ is a simple algebra and the regular bimodule is generated by $\mathrm I$; any nonzero sub-bimodule contains $a\mathrm I b$ for some $a,b$, and these span $A$. $\square$
The Collapse
The span. The image of $A\otimes\mathrm I$ and $\mathrm I\otimes A$ under $\mu$ is $\mathfrak{a}=\{L_a+R_b\}$, of dimension $2n^2-1$.
Proposition (the exact gap). For the Hilbert–Schmidt form, $$ \operatorname{End}(A)=\mathfrak{a}\oplus\mathfrak{a}^{\perp},\qquad \dim\mathfrak{a}=2n^2-1,\qquad \dim\mathfrak{a}^{\perp}=(n^2-1)^2 . $$
Proof. $\mathfrak{a}^{\perp}$ consists of the operators $Y$ with $\langle L_a,Y\rangle=\langle R_b,Y\rangle=0$ for all $a,b$. With the index-pair convention these two conditions are the vanishing of the two partial traces, $$ (\operatorname{tr}_1Y)_{lj}=\sum_iY_{il,ij}=0,\qquad (\operatorname{tr}_2Y)_{ik}=\sum_jY_{ij,kj}=0 . $$ Each partial trace is a map onto $M_n$, so each condition cuts $n^2$ dimensions, and the two conditions are independent because one constrains the first index pair and the other the second. Hence the dimension is $n^4-2n^2+1=(n^2-1)^2$. $\square$
Where the missing directions sit. The complement is spanned by the products whose two parameters are both non-scalar, modulo the one-sided span. A basis is given by $T_{E_{ij},E_{kl}}$ with $i\neq j$ or $k\neq l$; the number of such units is $n^4-n^2$, and after removing the $n^2-1$ redundant combinations one is left with $(n^2-1)^2$.
The $(n^2-1)^2$ reading. The complement is $M_n^0\otimes M_n^0$ up to the identification, where $M_n^0=\mathfrak{sl}_n$ is the traceless part. So $$ \mathfrak{a}^{\perp}\cong\mathfrak{sl}_n\otimes\mathfrak{sl}_n, $$ of dimension $(n^2-1)^2$. The one-sided span carries the traceless part of one slot at a time; the complement carries both traceless parts at once.
The exactness at $n=2$. $2\cdot4-1=7$ and $(4-1)^2=9$, and $7+9=16$. The biquaternion series used the numbers $7$ and $9$ without naming them as the two pieces of the operator algebra.
The group reading. The Lie algebra of $(GL_n\times GL_n)/\mathbb{C}^{\times}$ is $\mathfrak{a}$ modulo the scalars, of dimension $2n^2-1$. The complement is not a Lie algebra: it is the module of the two-sided action on which both traceless parts act, and its Lie-theoretic role is that of the off-diagonal pair of $\mathfrak{sl}_n\oplus\mathfrak{sl}_n$ in the full endomorphism algebra.
The Generated Algebra Is Everything
Proposition. The subalgebra of $\operatorname{End}(A)$ generated by $L(A)\cup R(A)$ is $\operatorname{End}(A)$.
Proof. It contains the products $L_aR_b$ and their span, which is everything by the isomorphism above. $\square$
The contrast in one line. Span: $2n^2-1$. Generated algebra: $n^4$. The quotient grows like $n^2/2$, so the span is a vanishing fraction of the algebra.
Why the two statements coexist. A subalgebra is closed under multiplication, and multiplication of one-sided operators produces two-sided ones. The span is closed under addition only. Since $T_{a,b}=L_aR_b$ is outside $\mathfrak{a}$ whenever both parameters are non-scalar, the algebra generated is much larger than the span.
The Projection onto the Span
Proposition (the projection). The orthogonal projection of $\operatorname{End}(A)$ onto $\mathfrak{a}$ is $$ \Pi(Y)=\frac1n\,(\operatorname{tr}_2Y)\otimes\mathrm I+\frac1n\,\mathrm I\otimes(\operatorname{tr}_1Y)-\frac1{n^2}(\operatorname{tr}Y)\,\mathrm I\otimes\mathrm I . $$
Proof. Read in the index pair, $\Pi(Y)[(i,k),(j,l)]=\frac1n(\operatorname{tr}_2Y)_{ij}\delta_{kl}+\frac1n\delta_{ij}(\operatorname{tr}_1Y)_{kl}-\frac1{n^2}(\operatorname{tr}Y)\delta_{ij}\delta_{kl}$. On $L_A=\mathrm I\otimes A$ one has $\operatorname{tr}_2Y=(\operatorname{tr}A)\mathrm I$ and $\operatorname{tr}_1Y=nA$, so the first and third terms cancel and $\Pi(L_A)=L_A$; similarly $\Pi(R_B)=R_B$. On a generator of $\mathfrak{a}^{\perp}$ both partial traces vanish, so $\Pi$ kills $\mathfrak{a}^{\perp}$. A projection onto $\mathfrak{a}$ with kernel $\mathfrak{a}^{\perp}$ is orthogonal, because the two subspaces are orthogonal complements. $\square$
The corollary. $\Pi(T_{A,B})=T_{A,B}$ exactly when $A$ or $B$ is a scalar, which is the statement that the projection removes the genuine two-sided part.
The three terms. The first term records the slot-one trace, the second the slot-two trace, and the third the correction for the double count of the scalar line. The dimension count $n^2+n^2-1=2n^2-1$ is the dimension of the image.
The Finite Check
Computed with numpy in the column-stacking basis for $n=2,3,4,5$:
| $n$ | $n^4$ | $2n^2-1$ | $(n^2-1)^2$ | rank of $\mathfrak{a}^{\perp}$ |
|---|---|---|---|---|
| $2$ | $16$ | $7$ | $9$ | $9$ |
| $3$ | $81$ | $17$ | $64$ | $64$ |
| $4$ | $256$ | $31$ | $225$ | $225$ |
| $5$ | $625$ | $49$ | $576$ | $576$ |
The rank of $\mathfrak{a}^{\perp}$ was obtained as the nullity of the $2n^2\times n^4$ matrix of the one-sided operators with respect to the Hilbert–Schmidt form, and it matches $(n^2-1)^2$ exactly. The two partial traces of a random element of $\mathfrak{a}^{\perp}$ were computed and vanish to $10^{-14}$; the partial traces of a random element of $\mathfrak{a}$ are a random pair in $M_n$, confirming that the map $\mathfrak{a}\to M_n\oplus M_n$ is onto.
The commutant of $L(A)$ was computed as the kernel of the linear system $TL_a-L_aT=0$ over all $a$ in a basis: its dimension is $n^2$ for each $n$, matching $R(A)$, namely $4,9,16,25$.
The projection $\Pi$ was checked to be idempotent to $10^{-16}$, to be self-adjoint in the sense $\langle\Pi Y,Z\rangle=\langle Y,\Pi Z\rangle$ to $10^{-15}$, to fix $L_A$ and $R_B$ to $10^{-16}$, and to send a random product $T_{A,B}$ with scalar $B$ to itself. For a product with both parameters non-scalar the distance $\lVert\Pi(T_{A,B})-T_{A,B}\rVert$ is $2.94,17.8,34.3,58.2$ at $n=2,3,4,5$, so the removed part is large.
The Double Centraliser
The bimodule maps. An operator on $M_n$ that commutes with every $L_A$ and every $R_B$ is a scalar: it commutes with all left and right multiplications, hence is a bimodule endomorphism of the regular bimodule, hence is given by an element of the centre.
Proof. Commuting with $L_A$ makes it a right $M_n$-module map, hence of the form $R_B$; commuting with $R_B$ as well makes $B$ central. $\square$
The consequence for the algebra. The algebra generated by the two copies is its own bicommutant, which is the double centraliser theorem for $M_n$: the pair $(M_n,M_n^{\mathrm{op}})$ is a pair of mutual commutants, and the algebra they generate is exactly the set of operators that commute with the scalars.
The count. That algebra has dimension $2n^2-1$, and the bimodule maps are the one-dimensional scalars inside it. The remaining $2n^2-2$ directions are bimodule-linear in one variable only.
The gap. The complement of the enveloping algebra in $\operatorname{End}(M_n)$ has dimension $(n^2-1)^2$, and none of it is a bimodule map. The two-sided family fills the gap with operators that are not bimodule maps at all: $T_{A,B}$ with both parameters non-scalar is not a map of left or of right modules.
The Two Commutants
| commutant of | is | dimension |
|---|---|---|
| the left copy | the right copy | $n^2$ |
| the right copy | the left copy | $n^2$ |
| both copies | the scalars | $1$ |
| the scalars | $\operatorname{End}(M_n)$ | $n^4$ |
The reading. The table is the double centraliser theorem in four lines. It is the reason the pair's algebra is so much smaller than the endomorphism algebra, and the reason the gap cannot be closed by adding commutants.
Summary
The enveloping algebra of the regular bimodule is the whole operator algebra $M_{n^2}(\mathbb{C})$, of dimension $n^4$, and the bimodule is simple with endomorphism ring $\mathbb{C}$. The two copies of the algebra inside it have a span of dimension $2n^2-1$, and the orthogonal complement has dimension $(n^2-1)^2$ and is $\mathfrak{sl}_n\otimes\mathfrak{sl}_n$. The subalgebra generated is everything; the span is not. At $n=2$ the two numbers are $7$ and $9$, and they add to the $16$ of the operator algebra.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $A$ | $M_n(\mathbb{C})$ |
| $A^{\mathrm{op}}$ | the opposite algebra |
| $\mu$ | the multiplication map $a\otimes b\mapsto L_aR_b$ |
| $\mathfrak{a}$ | the span of the one-sided copies, dimension $2n^2-1$ |
| $\operatorname{tr}_1,\operatorname{tr}_2$ | the partial traces on the index pair |
| $M_n^0$ | the traceless part $\mathfrak{sl}_n$ |
Further Reading
- The Endomorphisms of $M_n(\mathbb{C})$, for the three families and the dimension ladder.
- The Ideals, the Centre and the Simplicity of $M_n(\mathbb{C})$, for the simplicity used in the bimodule statement.
- The Transposition of the Two Sides of $M_n(\mathbb{C})$, for the intertwiner that the enveloping algebra does not contain.
- The Adjoint Pair of $L$ and $R$ Operators on $M_n(\mathbb{C})$, for the pair as a commuting pair.
- The Enveloping Operator Algebra of the Biquaternion Algebra and the Bi-module Structure, for the case $n=2$.
- The Two-Sided Operators and the Segre Cone of $M_n(\mathbb{C})$, for the two-sided family as a cone.