The Transposition of the Two Sides of $M_n(\mathbb{C})$
Introduction
The question. What happens to a two-sided operator $T_{A,B}$ under the transpose of the carrier, and can the left and the right sides be exchanged?
The answer in one line. Conjugation by the transpose operator $F(X)=X^{\mathsf T}$ swaps the two sides and transposes both parameters, $$ FT_{A,B}F=T_{B^{\mathsf T},A^{\mathsf T}}, $$ and no element of the algebra and no single product does the same.
Why the article is needed. In the biquaternion series the exchange of the two sides was written with one matrix, because the transpose of the coefficient frame is a conjugation and the frame is $4\times4$. At general $n$ the exchange still exists, but as an operator that is outside the one-sided span and outside the two-sided family. Separating the operator from the element is the point.
Boundaries. The families are The Endomorphisms of $M_n(\mathbb{C})$; the adjoints are The Two-Sided Operator, the Difference Family and the Adjoints of $M_n(\mathbb{C})$; the enveloping algebra is The Enveloping Operator Algebra of $M_n(\mathbb{C})$ and the Bi-module Structure. The biquaternion original is The Transposition of the Two Sides.
Conventions. $F(X)=X^{\mathsf T}$, so $F^2=\mathrm I$. The three involutions of the operator layer are $$ F T F,\qquad T^{\dagger}\ (\text{Hilbert–Schmidt}),\qquad T^{\mathsf T}\ (\text{transposed bilinear form}). $$
The Three Transposes
The table. For a single product the three operations give
| operation | result | effect on the parameters |
|---|---|---|
| $FTF$ | $T_{B^{\mathsf T},A^{\mathsf T}}$ | swaps the slots and transposes both |
| $T^{\dagger}$ | $T_{A^{*},B^{*}}$ | no swap, conjugate transpose both |
| $T^{\mathsf T}$ | $T_{A^{\mathsf T},B^{\mathsf T}}$ | no swap, transpose both |
Proposition. $FT_{A,B}F=T_{B^{\mathsf T},A^{\mathsf T}}$.
Proof. $(FT_{A,B}FX)_{ij}$ is the $(i,j)$ entry of $(AX^{\mathsf T}B)^{\mathsf T}$, which is $(B^{\mathsf T}XA^{\mathsf T})_{ij}$. $\square$
The reading. The bilinear transpose $T\mapsto T^{\mathsf T}$ never moves a parameter from one slot to the other. Conjugation by $F$ always does. The two agree only when $A$ and $B$ are symmetric and equal, in which case both fix $T_{A,A}$.
The adjoint pair. $F$ is self-adjoint and unitary for the Hilbert–Schmidt form, so $FTF$ is unitarily equivalent to $T$. Conjugation by $F$ is therefore an automorphism of the operator algebra of order two, and it is the outer symmetry that the one-sided families lack.
Exchanging the Sides
Proposition (the operator does it). $FL_AF=R_{A^{\mathsf T}}$ and $FR_BF=L_{B^{\mathsf T}}$.
Proof. The first is the case $B=\mathrm I$ of the formula above: $FT_{A,\mathrm I}F=T_{\mathrm I,A^{\mathsf T}}=R_{A^{\mathsf T}}$. The second follows by applying the first to $F^2=\mathrm I$. $\square$
Proposition (no element of the algebra does it). There is no $D\in M_n$ with $DR_AD^{-1}=L_A$ for all $A$, unless $n=1$.
Proof. The identity reads $DXA=ADX$ for all $X$. At $X=\mathrm I$ it gives $DA=AD$; substituting back gives $(DX)A=A(DX)$ for every $X$, so $DX$ is central for every $X$, hence $D$ is scalar. A scalar gives $R_A=L_A$, false. $\square$
Proposition (no single product does it). There is no pair $(P,Q)$ with $T_{P,Q}L_AT_{P,Q}^{-1}=L_A$ for all $A$.
Proof. $T_{P,Q}L_AT_{Q^{-1},P^{-1}}=T_{PAQ,Q^{-1}P^{-1}}$, which equals $L_A$ only if $Q^{-1}P^{-1}=\mathrm I$ and $PAQ=A$ for all $A$. The first gives $PQ=\mathrm I$; the second then forces $P$ and $Q$ to be scalars, and again $R_A=L_A$. $\square$
The moral. The exchange is a genuinely operator-level symmetry. It is not inner, at any $n\ge2$. At $n=2$ it has a familiar matrix because the transpose of the coefficient frame is the quaternion conjugation, which the biquaternion series already wrote down; the existence of that matrix is not an accident of small $n$, but its interpretation as an element of the algebra is.
The Flip Is Outside the One-Sided Span
Proposition. $F\notin\operatorname{span}\{L_A,R_B\}$ for $n\ge2$.
Proof. Suppose $F=L_A+R_B$, so $X^{\mathsf T}=AX+XB$ for all $X$. At $X=\mathrm I$ this gives $A+B=\mathrm I$. At $X=E_{ij}$ the $(k,l)$ entry is $\delta_{ki}\delta_{lj}$ on the left and $A_{ki}\delta_{lj}+\delta_{ki}B_{lj}$ on the right. Taking $k=i$ and $l=j$ gives $A_{ii}+B_{jj}=1$. Taking $k=i$ and $l\neq j$ gives $B_{lj}=0$, so $B$ is diagonal; taking $k\neq i$ and $l=j$ gives $A_{ki}=0$, so $A$ is diagonal. Then $A_{ii}+B_{jj}=1$ for all $i,j$, so $B_{jj}$ is independent of $j$ and $A_{ii}$ is independent of $i$: both are scalars with $\alpha+\beta=1$. But then $X^{\mathsf T}=\alpha X+\beta X=X$, false for $n\ge2$. $\square$
The length. $F=\sum_{ij}E_{ij}\otimes E_{ji}$, a sum of $n^2$ products, so $\ell(F)\le n^2$. The value of the length is computed in The Length and the Two Ranks on $M_n(\mathbb{C})$; what matters here is that it is at least two, since $F$ is not a product.
The orthogonal direction. $F$ has both a symmetric part and an antisymmetric part under the transpose of the operator space, namely $F$ itself and the map $X\mapsto-X^{\mathsf T}$ on one half. It is therefore not orthogonal to the one-sided span in a trivial way; its distance to the span is the quantity measured in the finite check below.
The Group Generated
The two actions. For $G,H$ invertible let $\Phi_{G,H}$ be conjugation by $L_GR_{H^{-1}}$, the operator $T\mapsto L_GR_{H^{-1}}\,T\,R_HL_{G^{-1}}$. This is an algebra automorphism of $\operatorname{End}(M_n)$, and $\Phi_{G,H}\Phi_{G',H'}=\Phi_{GG',HH'}$; the kernel is the scalars.
Proposition (the flip normalises). $F\Phi_{G,H}F=\Phi_{H^{-\mathsf T},G^{-\mathsf T}}$.
Proof. $FL_GR_{H^{-1}}F=R_{G^{\mathsf T}}L_{H^{-\mathsf T}}=L_{H^{-\mathsf T}}R_{G^{\mathsf T}}$, using that the two copies commute and $FL_GF=R_{G^{\mathsf T}}$. $\square$
The group. The operators $\Phi_{G,H}$ and $F\Phi_{G,H}$ form the group $$ \bigl(GL_n\times GL_n\bigr)\rtimes\mathbb{Z}_2\big/ \mathbb{C}^{\times}, $$ of complex dimension $2n^2-1$. Its compact form is $(U(n)\times U(n))\rtimes\mathbb{Z}_2$ modulo the scalars, of real dimension $2n^2-1$.
What is not claimed. This is the group of automorphisms of the operator algebra that are built from the two actions and the flip. The full normaliser of the span $\mathfrak{a}$ inside $GL(\operatorname{End}(M_n))$ is not computed here, and it may be larger: the Lie algebra $\mathfrak{a}$ has two copies of $\mathfrak{sl}_n$, and for $n\ge3$ the diagram automorphism $A\mapsto-A^{\mathsf T}$ of each copy preserves $\mathfrak{a}$ as a Lie algebra without being realised by any operator of this group. Whether it is realised at all is an open point of this article.
The biquaternion reading. At $n=2$ the complex dimension is $7$ and the real compact dimension is $7$, and the extra $\mathbb{Z}_2$ is the chirality reversal of Operator Chirality on $M_n(\mathbb{C})$, implemented by the same operator $F$.
The Two Transposes of an Adjoint Pair
The pair. A pair $(A,B)$ is Hermitian when $A=A^{*}$ and $B=B^{*}$. Then $$ T_{A,B}^{\dagger}=T_{A,B},\qquad FT_{A,B}F=T_{B^{\mathsf T},A^{\mathsf T}}, $$ so the Hilbert–Schmidt adjoint fixes the product while the flip swaps the two parameters.
The self-adjointness condition. $T_{A,B}$ is self-adjoint exactly when $A=A^{*}$ and $B=B^{*}$, because $T_{A,B}^{\dagger}=T_{A^{*},B^{*}}$ and $T_{A,B}=T_{C,D}$ forces $(C,D)=(\lambda A,\lambda^{-1}B)$. So self-adjoint products are the Hermitian pairs.
The symmetric condition. $T_{A,B}$ is symmetric for the bilinear form exactly when $A=A^{\mathsf T}$ and $B=B^{\mathsf T}$, that is when both parameters are complex symmetric.
The fixed set of the flip. The operators fixed by $T\mapsto FTF$ are those with $T_{A,B}=T_{B^{\mathsf T},A^{\mathsf T}}$, which for a single product means $B=A^{\mathsf T}$. So the flip fixes the products with symmetric parameters and exchanges the rest in pairs.
The Intertwining Reading
The two representations. $A\mapsto L_A$ is a representation of the algebra $M_n$ on the carrier space of operators, and $B\mapsto R_B$ is a representation of $M_n^{\mathrm{op}}$.
Proposition. $L\cong R\circ\tau$: the flip intertwines the left representation with the right one composed with the transpose anti-automorphism.
Proof. $FL_AF=R_{A^{\mathsf T}}$, which is the defining relation of an intertwiner between $A\mapsto L_A$ and $A\mapsto R_{A^{\mathsf T}}$. $\square$
Proposition. $L$ and $R$ are inequivalent as representations of $M_n$ under the identity map of the algebra.
Proof. An intertwiner $S$ satisfies $SAX=SXA$ for all $A,X$. For invertible $S$ this gives $AX=XA$ for all $A,X$, so every $A$ is scalar, false for $n\ge2$. $\square$
The consequence. The two sides are the same representation only after the anti-automorphism $\tau$ is applied. At $n=2$ the coefficient frame makes $\tau$ look like the conjugation of the algebra, so the twist is invisible; at $n\ge3$ it is visible, and it is the reason the side exchange needs an operator rather than an element.
The tensor form. With the intertwiner in place, the operator algebra is $V\otimes V^{*}\otimes V\otimes V^{*}$ as a representation of the two copies, and the flip exchanges the first two factors with the last two. This is the representation-theoretic content of $FT_{A,B}F=T_{B^{\mathsf T},A^{\mathsf T}}$.
The Finite Check
Computed with numpy for $n=2,3,4,5$ on random complex $A$: the residual $\lVert FT_{A,B}F-T_{B^{\mathsf T},A^{\mathsf T}}\rVert$ is below $10^{-14}$; the residual $\lVert FL_AF-R_{A^{\mathsf T}}\rVert$ is below $10^{-14}$; the residual $\lVert T^{\mathsf T}-T_{A^{\mathsf T},B^{\mathsf T}}\rVert$ is below $10^{-13}$ for the transposed bilinear form.
The distance from $F$ to the one-sided span was computed by least squares on the $n^4\times 2n^2$ system. The residual norm is $\sqrt{n^2-1}$, namely $\sqrt3,\sqrt8,\sqrt{15},\sqrt{24}$ for $n=2,3,4,5$, and the normalised distance is $\sqrt{n^2-1}/n$, namely $0.866,0.943,0.968,0.980$. So $F$ is never in the span, and it approaches the unit sphere of the orthogonal complement as $n$ grows. At $n=2$ the flip is a $4\times4$ permutation matrix of trace $2$ and determinant $-1$.
| $n$ | $\dim\mathfrak{a}=2n^2-1$ | distance $\sqrt{n^2-1}$ | normalised |
|---|---|---|---|
| $2$ | $7$ | $1.732$ | $0.866$ |
| $3$ | $17$ | $2.828$ | $0.943$ |
| $4$ | $31$ | $3.873$ | $0.968$ |
| $5$ | $49$ | $4.899$ | $0.980$ |
The Frame
The flip in coordinates. For $n=2$ the flip exchanges the two sides, and in the four-dimensional frame it swaps the two tensor factors. The alternating form is invariant under it, and the two sides are exchanged by an operator of order two up to a sign.
The general case. For $n\ge3$ there is no single matrix $S$ with $SAS^{-1}=A^{\mathsf T}$ for all $A$, because the automorphism group of $M_n$ is inner and the transpose is not an inner automorphism. The symmetry group of the two-sided family is therefore generated by the two independent swaps, one for each side, and it is larger than the group generated by the flip.
The exact formula. A symmetry that swaps the sides sends $T_{A,B}$ to $T_{B,A}$ for every pair, and the flip realises it: $FT_{A,B}F=T_{B,A}$. The formula is exact at every $n$, and it is the reason the two-sided family has a symmetry that the one-sided families do not.
The consequence. The two-sided structures are classified by the orbits of the group generated by the flip and the two swaps, and at $n=2$ that group is the group of the biquaternion frame, while at $n\ge3$ it is strictly larger.
Summary
Conjugation by the transpose swaps the two sides and transposes both parameters; the bilinear transpose transposes the parameters in place. No element of the algebra and no single product exchanges the sides, at any $n\ge2$. The transpose operator $F$ is outside the one-sided span and has length at least two. The normaliser of the one-sided span is the semidirect product of the two general linear groups with the flip, of dimension $2n^2$.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $F$ | the transpose operator $X\mapsto X^{\mathsf T}$ on the carrier |
| $\mathfrak{a}$ | the span of the one-sided families |
| $T^{\dagger}$ | the Hilbert–Schmidt adjoint |
| $T^{\mathsf T}$ | the adjoint for the transposed bilinear form |
| $G$ | the normaliser of $\mathfrak{a}$ in $GL(M_n)$ |
Further Reading
- The Endomorphisms of $M_n(\mathbb{C})$, for the three families and their span.
- Operator Chirality on $M_n(\mathbb{C})$, for the transpose as the chirality reversal.
- The Enveloping Operator Algebra of $M_n(\mathbb{C})$ and the Bi-module Structure, for the double centraliser property.
- The Length and the Two Ranks on $M_n(\mathbb{C})$, for the length of the flip.
- The Transposition of the Two Sides, for the case $n=2$.
- Operators Acting on Two Sides, for the two-sided theory in the biquaternion frame.