The Length and the Two Ranks on $M_n(\mathbb{C})$
Introduction
The question. A general endomorphism of $M_{n}(\mathbb{C})$ is not a single two-sided product, but it is a finite sum of them. Two numbers measure it: the length $\ell(M)$, the least number of products in a decomposition, and the rank $\operatorname{rank}M$, the ordinary rank of the $n^{2}\times n^{2}$ matrix. This article proves that the length is the rank of a certain $n^{2}\times n^{2}$ array of coefficients, computes the strata of each length, and shows that the two numbers are independent.
The result in one line. The length is the rank of the coefficient array $c(M)$, its values are $0,1,\dots,n^{2}$, the set of arrays of length at most $r$ is the determinantal variety of dimension $2rn^{2}-r^{2}$, and the generic matrix has length $n^{2}$.
The one identity that carries everything. With the $n^{4}$ basis products $L_{E_{\mu}}R_{E_{\nu}}$ and an orthonormal basis of the algebra, $$ c(L_{A}R_{B})=[A]\,[B]^{\mathsf T}, $$ so a product has a coefficient array of rank one and every array of rank one is a product. The product condition becomes a rank condition, and the length becomes the rank of the array; nothing else in the article is needed for the two theorems.
Boundaries. The products and their cone are The Two-Sided Operators and the Segre Cone of $M_n(\mathbb{C})$; the one-sided families are The One-Sided Operators of $M_n(\mathbb{C})$; the abstract setting is The Regular Representation in Dimension $n$. At $n=2$ the same theorem with the sixteen Peirce-frame operators is $M_4(\mathbb{C})$ Pauli Representation : Operators, and the coefficient matrix there is the $4\times4$ array of that article. The general statement for an algebra of dimension $N$ is the tensor form of The Regular Representation in Dimension $n$.
Conventions. The algebra is $A=M_{n}(\mathbb{C})$ of complex dimension $N=n^{2}$; the endomorphism algebra has complex dimension $N^{2}=n^{4}$. The matrix units are $E_{\mu}=E_{ij}$ with $\mu$ running over the $N$ pairs, and they are orthonormal for the Hilbert–Schmidt form, $\operatorname{tr}(E_{\mu}^{*}E_{\nu})=\delta_{\mu\nu}$. The basis products are $L_{E_{\mu}}R_{E_{\nu}}$. The coefficient array of $M\in\operatorname{End}(A)$ is $c(M)$, an $N\times N$ array. The length is $\ell(M)$.
The Coefficient Array
The Basis of Products
Theorem. The $N^{2}=n^{4}$ operators $L_{E_{\mu}}R_{E_{\nu}}$ are a basis of $\operatorname{End}(M_{n}(\mathbb{C}))$, and they are orthonormal for the Hilbert–Schmidt form: $$ \operatorname{tr}\bigl(L_{E_{\mu}}R_{E_{\nu}}^{*}L_{\rho}R_{\sigma}\bigr)=\delta_{\mu\rho}\delta_{\nu\sigma}. $$
Proof. The matrix of $L_{E_{\mu}}R_{E_{\nu}}$ is $E_{\nu}^{\mathsf T}\otimes E_{\mu}$, so $$ \operatorname{tr}\bigl(L_{E_{\mu}}R_{E_{\nu}}^{*}L_{\rho}R_{\sigma}\bigr) =\operatorname{tr}\bigl(E_{\nu}^{\mathsf T*}E_{\sigma}^{\mathsf T}\bigr)\operatorname{tr}\bigl(E_{\mu}^{*}E_{\rho}\bigr) =\delta_{\nu\sigma}\delta_{\mu\rho}, $$ using the orthonormality of the matrix units in each factor.
The consequence. Every operator has the unique expansion $$ M=\sum_{\mu,\nu}c(M)_{\mu\nu}L_{E_{\mu}}R_{E_{\nu}},\qquad c(M)_{\mu\nu}=\operatorname{tr}\bigl(L_{E_{\mu}}R_{E_{\nu}}^{*}M\bigr), $$ and the assignment $M\mapsto c(M)$ is a linear isomorphism of $\operatorname{End}(A)$ onto the space of $N\times N$ arrays.
The normalisation. The normalisation is the one of an orthonormal basis of the algebra. In the biquaternion articles the basis $e_{t},\vec e_{x},\vec e_{y},\vec e_{z}$ has $\operatorname{tr}(\mathsf M_{2}(e_{\mu})^{*}\mathsf M_{2}(e_{\nu}))=2\delta_{\mu\nu}$, so the products have Hilbert–Schmidt norm two, $\operatorname{tr}(L_{E_{\mu}}R_{E_{\nu}}^{*}L_{E_{\mu}}R_{E_{\nu}})=4$, and the coefficient is $\tfrac14$ of the scalar product. The array is the same object in both normalisations; only the factor of $4$ differs.
The Array of a Product
Theorem. For every pair, $$ c(L_{A}R_{B})=[A]\,[B]^{\mathsf T}, $$ where $[A]$ is the column of coordinates of $A$ in the basis $E_{\mu}$, $[A]_{\mu}=\operatorname{tr}(E_{\mu}^{*}A)$.
Proof. By definition $L_{A}R_{B}=\sum_{\mu,\nu}A_{\mu}B_{\nu}L_{E_{\mu}}R_{E_{\nu}}$ with $A=\sum A_{\mu}E_{\mu}$ and $B=\sum B_{\nu}E_{\nu}$, since the map $(A,B)\mapsto L_{A}R_{B}$ is bilinear and the basis products are the images of the basis pairs. The array of coefficients is therefore the outer product of the two coordinate columns.
Three readings of one identity. The rank of $[A][B]^{\mathsf T}$ is one when both vectors are nonzero; the array is zero when either is; and the array determines the operator by the basis expansion. The rank-one condition is therefore exactly the product condition, and it is a condition on an $N\times N$ array and not on the $N\times N$ matrix $M$.
The array of a one-sided operator. The array of $L_{A}=L_{A}R_{I}$ is the outer product of $[A]$ with the coordinate column of the identity, which is the column of the diagonal units. Its rank is one unless $A=0$, so every one-sided operator has length one, as it must.
The Length
The Definition and the Theorem
Definition. The length of $M$ is $$ \ell(M)=\min\{r\ge0: M=L_{A_{1}}R_{B_{1}}+\cdots+L_{A_{r}}R_{B_{r}}\}, $$ with $\ell(0)=0$.
Theorem (the length is the rank of the array). For every $M\in\operatorname{End}(M_{n}(\mathbb{C}))$, $$ \ell(M)=\operatorname{rank}c(M). $$
Proof. If $M$ is a sum of $r$ products, then $c(M)=\sum_{i\le r}[A_{i}][B_{i}]^{\mathsf T}$ is a sum of $r$ arrays of rank one, so its rank is at most $r$; hence $\ell(M)\ge\operatorname{rank}c(M)$. Conversely let $c=c(M)$ have rank $r$ and factor it as $c=UV$ with $U$ of size $N\times r$ and $V$ of size $r\times N$. The $i$-th column of $U$ is the coordinate column of a matrix $A_{i}$ and the $i$-th row of $V$ is the transpose of the coordinate column of a matrix $B_{i}$, so $c=\sum_{i\le r}[A_{i}][B_{i}]^{\mathsf T}$ and $M=\sum_{i\le r}L_{A_{i}}R_{B_{i}}$; hence $\ell(M)\le r$.
The values. The length takes the values $0,1,\dots,n^{2}$. The value $0$ is the zero operator; the value $1$ is the cone of products; the value $n^{2}$ is the complement of the determinantal variety of the arrays of rank at most $n^{2}-1$, which is open and dense. Every endomorphism of $M_{n}(\mathbb{C})$ is a sum of at most $n^{2}$ two-sided products, and a generic one needs exactly $n^{2}$.
The case $n=2$. The statement is the four of the Pauli Representation article: every $4\times4$ matrix is a sum of at most four products, and a generic matrix needs four.
The Strata
Theorem (the stratum of length at most $r$). The set $$ \mathcal{C}_{r}=\{M:\ell(M)\le r\}=\{M:\operatorname{rank}c(M)\le r\} $$ is an irreducible closed algebraic cone of complex dimension $$ 2rn^{2}-r^{2} $$ for $0\le r\le n^{2}$, and it is smooth away from the origin for $r=1$.
Proof. The set is the inverse image under the linear isomorphism $c$ of the determinantal variety of $N\times N$ arrays of rank at most $r$, whose dimension is the classical $r(2N-r)=2rN-r^{2}$. The cone property is the linearity of $c$. The smoothness for $r=1$ is the statement that the Segre cone is smooth off its vertex.
The table of dimensions. For $N=n^{2}=4$, that is $n=2$, the dimensions are
| $r$ | $\dim\mathcal{C}_{r}$ | the set |
|---|---|---|
| $0$ | $0$ | the zero operator |
| $1$ | $7$ | the cone of products |
| $2$ | $12$ | the sums of two products |
| $3$ | $15$ | the complement of the arrays of rank four |
| $4$ | $16$ | the whole space |
and for $N=9$, that is $n=3$,
| $r$ | $\dim\mathcal{C}_{r}$ |
|---|---|
| $0$ | $0$ |
| $1$ | $17$ |
| $2$ | $32$ |
| $3$ | $45$ |
| $4$ | $56$ |
| $5$ | $65$ |
| $6$ | $72$ |
| $7$ | $77$ |
| $8$ | $80$ |
| $9$ | $81$ |
Every $r$ from $0$ to $N$ is a length, and the table has no gaps. What has gaps is the table of the ordinary rank of a product, which skips every number that is not a product of two integers at most $n$. The two stratifications are different, and this is the subject of the next section.
The jump. The increments of the dimension are $2N-2r+1$: from $r-1$ to $r$ the stratum grows by that number. At $n=2$ the increments are $7,5,3,1$, the last being the single condition $\det c\ne0$; at $n=3$ they are $17,15,13,11,9,7,5,3,1$. The top stratum is the complement of a hypersurface in the arrays, while the bottom one is far from it: the products have codimension $(n^{2}-1)^{2}=n^{4}-2n^{2}+1$ in the operator space, which at $n=2$ is $9$ and at $n=3$ is $64$.
The Criterion of a Single Product
Theorem. For $M\neq0$, $$ \ell(M)=1\iff c(M)=[A][B]^{\mathsf T}\ \text{for some }A,B\neq0\iff\operatorname{rank}c(M)=1 . $$
Proof. The first equivalence is the array of a product, and the second is the criterion that an $N\times N$ array of rank one is exactly an outer product of two nonzero columns.
The test on $M$ alone. The criterion is a test on the coefficients and not on the matrix: the rank of $M$ tells nothing, because a product can have rank $1$ and a matrix of rank $1$ can have length $n^{2}$. The test is $$ M\ \text{is a product}\iff\text{the }2\times2\text{ minors of }c(M)\text{ vanish}, $$ that is $N^{2}(N-1)^{2}/4$ polynomial conditions, which define a variety of codimension $(N-1)^{2}$ in the arrays of size $N$.
How to read the parameters off the array. If $\operatorname{rank}c=1$, take $[A]$ to be a nonzero column of $c$ and $[B]^{\mathsf T}$ to be the corresponding row divided by the pivot entry. If $\operatorname{rank}c=r>1$, a rank factorisation $c=UV$ gives the $r$ products, one per column of $U$ against the corresponding row of $V$.
The Two Ranks
The Independence
Theorem. Neither number determines the other.
Proof. Two counterexamples at $n=2$ suffice, and both generalise.
A product of rank one and of length one. Take $A$ and $B$ of rank one; then $\operatorname{rank}L_{A}R_{B}=1$ and $\ell=1$.
A matrix of rank one and of length four. Take $M=uv^{\mathsf T}$ with $u,v$ generic columns of length $4$; the array $c(M)$ has rank $4$ for generic $u,v$, so $\ell(M)=4$ while $\operatorname{rank}M=1$. The set is not small: the rank-one matrices of length four form a dense open subset of the rank-one variety, and the computation is the one of the worked case $M_{1}$ of the Pauli Representation article. At $n=3$ the same construction gives a matrix of rank one and length $n^{2}=9$, and the phenomenon is not an accident of the small case.
A product of full rank and of length one. Take $A$ and $B$ invertible; then $\operatorname{rank}L_{A}R_{B}=n^{2}$ and $\ell=1$. So a matrix of the largest rank can have the smallest length.
The table at $n=2$.
| $M$ | $\operatorname{rank}M$ | $\operatorname{rank}c(M)$ | $\ell(M)$ |
|---|---|---|---|
| $L_{A}R_{B}$ with $A,B$ units | $4$ | $1$ | $1$ |
| $L_{A}R_{B}$ with $A,B$ of rank one | $1$ | $1$ | $1$ |
| a generic sum of three products | $4$ | $3$ | $3$ |
| a generic sum of two products | $4$ | $2$ | $2$ |
| a generic matrix of rank $1$ | $1$ | $4$ | $4$ |
| a generic matrix | $4$ | $4$ | $4$ |
The lesson. The ordinary rank of $M$ is a property of $M$ as a linear map of $\mathbb{C}^{n^{2}}$; the length is a property of $M$ as an element of $M_{n}\otimes M_{\mathrm{opp},n}$. The two are ranks of two different matrices attached to the same operator, and there is no inequality between them in either direction — beyond the trivial bounds $0\le\ell\le n^{2}$ and $0\le\operatorname{rank}M\le n^{2}$, which are the dimensions of two unrelated spaces.
The Rank of a Product
Theorem. For a product, $$ \operatorname{rank}L_{A}R_{B}=\operatorname{rank}A\cdot\operatorname{rank}B\in\{rs:0\le r,s\le n\}. $$
Proof. The matrix is $B^{\mathsf T}\otimes A$ and the rank of a Kronecker product is the product of the ranks.
The two ends of the same pair. At one end, $\ell=1$ for every product whatever its rank; at the other end, the products of full rank form the open dense part of the cone and the products of rank one form the closed part. The cone of products is stratified by the ordinary rank with strata of dimensions $r(2n-r)+s(2n-s)-1$, and by the length with a single stratum $\ell=1$.
The Generic Values
Theorem. On an open dense subset of $\operatorname{End}(M_{n}(\mathbb{C}))$, $$ \ell(M)=n^{2},\qquad \operatorname{rank}M=n^{2}. $$
Proof. The condition $\det c(M)\ne0$ is the nonvanishing of a polynomial, hence open and dense, and it gives $\ell=n^{2}$. The condition $\det M\ne0$ is likewise open and dense and gives the full rank. The two conditions are simultaneously generic, since the intersection of two dense open sets is dense and open.
The degenerate coincidence. The two generic values agree, and this is a coincidence of numbers and not a relation: they agree because both are the dimension $n^{2}$ of the two spaces, and the agreement fails on every stratum of positive codimension. A statement true generically and false on a dense-in-itself closed set is exactly what the length is for.
When the Theorem Fails
The Commutative Case
The collapse. Let $A$ be commutative. Then $A_{\mathrm{opp}}=A$ and $L_{A}R_{B}=L_{AB}$, so the product set is the one-sided family of dimension $N$, and every nonzero member has length one. The theorem $\ell=\operatorname{rank}c$ survives in the form "the length is $0$ or $1$", and it becomes empty.
The array is not well defined. In the commutative case the map $A\otimes A\to\operatorname{End}(A)$ has kernel of dimension $N^{2}-N$, so an operator has many coefficient arrays, of different ranks. The dual numbers give the example: $\varepsilon X\varepsilon=0$ for every $X$, so the tensor $\varepsilon\otimes\varepsilon$ lies in the kernel, and the operator $L_{1}$ has the array of $1\otimes1$, of rank one, and also the array of $1\otimes1+\varepsilon\otimes\varepsilon$, of rank two. The rank of the array is then a property of the chosen expansion and not of the operator.
The criterion. The identity $\ell(M)=\operatorname{rank}c(M)$ holds for an algebra $A$ exactly when the multiplication map $A\otimes A_{\mathrm{opp}}\to\operatorname{End}(A)$ is injective, that is exactly when $A$ is central simple. Over $\mathbb{C}$ these are the algebras $M_{n}(\mathbb{C})$, so the theorem is a theorem about matrix algebras and about nothing else: a split algebra, a commutative algebra or a product of two matrix algebras all fail it.
The reduced biquaternion case. The algebra $\mathbb{C}\times\mathbb{C}$ is commutative, of dimension two, and its operator space has dimension four. The products fill a two-dimensional subspace, every nonzero member has length one, and the length of the identity is one and not two. The algebra has the same complex dimension as the dual numbers and the same failure of the theorem; the failure follows from the commutativity and not from the size.
The Two-Theory Reading
The array is the tensor. The coefficient array is the tensor of $M$ in $A\otimes A_{\mathrm{opp}}$, read in the basis of the products. The length is the tensor rank. For a tensor with two factors the tensor rank is the rank of the flattening, and this is the whole content of the theorem. For a tensor with three factors the two notions part company, and this is why the statement is not a general fact about tensor rank.
The place of the theorem. The corpus uses the same identity three times: in the biquaternion Peirce frame with sixteen operators, in the canonical frame with the same sixteen in another basis, and here with $n^{4}$ operators. The three statements are one statement read in three frames, and the frame changes the array by a linear conjugation and not the rank.
Summary
The $n^{4}$ products $L_{E_{\mu}}R_{E_{\nu}}$ are an orthonormal basis of the endomorphism algebra, so every operator $M$ has a coefficient array $c(M)$ of size $n^{2}\times n^{2}$, and the array of a product is the outer product $[A][B]^{\mathsf T}$ of the two coordinate columns.
The length, the least number of products in a decomposition, equals the rank of $c(M)$. It takes the values $0,\dots,n^{2}$; the set of length at most $r$ is the determinantal variety of dimension $2rn^{2}-r^{2}$; a product is exactly a matrix of coefficient rank one; and a generic matrix has length $n^{2}$.
The length and the ordinary rank are independent: a product of full rank has length one, and a generic matrix of rank one has length $n^{2}$. The theorem holds exactly for the central simple algebras, that is for $M_{n}(\mathbb{C})$; for a commutative or split algebra the coefficient array is not unique and the length collapses to $\{0,1\}$. The length is the tensor rank in $M_{n}\otimes M_{\mathrm{opp},n}$, exact because the tensor has two factors, and it is the only invariant of the endomorphism algebra that the ordinary rank does not see.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $E_{\mu}$ | the matrix units, an orthonormal basis of the algebra |
| $N=n^{2}$ | the dimension of the algebra |
| $L_{E_{\mu}}R_{E_{\nu}}$ | the orthonormal basis of products |
| $c(M)$ | the coefficient array of $M$, of size $N\times N$ |
| $[A]$ | the coordinate column of $A$ in the invertibles |
| $\ell(M)$ | the length, the least number of products |
| $\mathcal{C}_{r}$ | the set of operators of length at most $r$ |
| $2rn^{2}-r^{2}$ | the complex dimension of $\mathcal{C}_{r}$ |
| $r(2n-r)+s(2n-s)-1$ | the dimension of the stratum of parameter ranks $(r,s)$ |
Further Reading
- Roger A. Horn and Charles R. Johnson, Topics in Matrix Analysis (Cambridge, 1991), for the Kronecker product, the rank of a product and the determinantal varieties of a matrix space.
- Joe Harris, Algebraic Geometry: A First Course (Springer, 1992), for the determinantal varieties and their dimensions.
- The Two-Sided Operators and the Segre Cone of $M_n(\mathbb{C})$, for the case $r=1$ and the cone of products.
- The Regular Representation in Dimension $n$, for the tensor form and the central simple criterion.
- The One-Sided Operators of $M_n(\mathbb{C})$, for the one-sided operators of length one.
- $M_4(\mathbb{C})$ Pauli Representation : Operators, for the same theorem with the sixteen Peirce-frame operators.
- Operators Acting on Two Sides, for the abstract family at $n=2$.