The Coefficient Matrix of a Two-Sided Operator: A Worked Computation
Introduction
The purpose. This article carries out one coefficient computation in full. It takes a concrete $4\times4$ matrix, computes its sixteen coefficients in the basis $L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu}$, assembles the coefficient matrix, finds its rank, factors it, and reconstructs the matrix. Every step is written out, with the matrices displayed.
The matrix. $$ M=\begin{pmatrix}1&2&0&0\\0&1&0&0\\0&0&1&2\\0&0&0&1\end{pmatrix}. $$ It is upper triangular, all diagonal entries are $1$, so $\det M=1$ and $\operatorname{rank}M=4$. Its only nonzero entries are $$ M_{11}=1,\quad M_{12}=2,\quad M_{22}=1,\quad M_{33}=1,\quad M_{34}=2,\quad M_{44}=1 . $$
The punchline, stated first. The ordinary rank is $4$, the maximal value. The length is $2$: the matrix is a sum of two such operators, and not of one. The two numbers measure different things, and the ordinary rank gives no information about the length. The coefficient rank does.
The method and its source. The basis, the orthogonality theorem and the length theorem are The Two-Sided Operators Method to Build $M_4(\mathbb{C})$. This article is the computation that accompanies it.
The Formula
The coefficient. For $A\in M_4(\mathbb{C})$ and $\mu,\nu\in\{t,x,y,z\}$, $$ c_{\mu\nu}(A)=\tfrac14\operatorname{tr}\bigl(L_{\mathrm{axf},e_\mu}{R_{\mathrm{axf},e_\nu}}^{*}A\bigr). $$
Why the factor four. The matrices $L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu}$ are orthogonal for the Hilbert–Schmidt form: $$ \operatorname{tr}\bigl(L_{\mathrm{axf},e_\mu}{R_{\mathrm{axf},e_\nu}}^{*}L_{\mathrm{axf},e_\rho}R_{\mathrm{axf},e_\sigma}\bigr)=4\,\delta_{\mu\rho}\delta_{\nu\sigma}. $$ The square norm of each is four, so the coefficient carries the factor $\tfrac14$.
The matrices are real. Every $L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu}$ is a real signed permutation matrix, so $L_{\mathrm{axf},e_\mu}{R_{\mathrm{axf},e_\nu}}^{*}=L_{\mathrm{axf},e_\mu}{R_{\mathrm{axf},e_\nu}}^{\mathsf T}$ and $$ \operatorname{tr}\bigl(L_{\mathrm{axf},e_\mu}{R_{\mathrm{axf},e_\nu}}^{*}A\bigr)=\sum_{i,j}(L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu})_{ij}A_{ij}. $$ In practice: read the four nonzero positions of $L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu}$, add the corresponding entries of $A$ with their signs, divide by four. No matrix product is needed, though two are shown below as a check.
The index order. The labels $t,x,y,z$ correspond to the rows and columns $1,2,3,4$. The first index of $c_{\mu\nu}$ is the left parameter, the second the right parameter. The order is part of the definition; swapping it transposes the coefficient matrix and returns the wrong operator.
The Sixteen Traces
The four diagonal matrices
$L_{\mathrm{axf},e_t}R_{\mathrm{axf},e_t}=\mathrm{I}_4$. The nonzero positions are the diagonal, all with sign $+1$: $$ \operatorname{tr}(L_{\mathrm{axf},e_t}{R_{\mathrm{axf},e_t}}^{\mathsf T}M)=M_{11}+M_{22}+M_{33}+M_{44}=1+1+1+1=4, \qquad c_{tt}=1 . $$
$L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},\vec e_x}=\operatorname{diag}(-1,-1,1,1)$. The signs are $-,-,+,+$: $$ \operatorname{tr}(L_{\mathrm{axf},\vec e_x}{R_{\mathrm{axf},\vec e_x}}^{\mathsf T}M)=-1-1+1+1=0, \qquad c_{xx}=0 . $$
$L_{\mathrm{axf},\vec e_y}R_{\mathrm{axf},\vec e_y}=\operatorname{diag}(-1,1,-1,1)$: $$ -1+1-1+1=0, \qquad c_{yy}=0 . $$
$L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_z}=\operatorname{diag}(-1,1,1,-1)$: $$ -1+1+1-1=0, \qquad c_{zz}=0 . $$
The reading. The four diagonal matrices see only the diagonal of $M$. Three of the four sums vanish because the diagonal of $M$ is constant.
Two full matrix products
The matrix $L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},e_t}$. Its nonzero positions are $(1,2)=-1$, $(2,1)=+1$, $(3,4)=-1$, $(4,3)=+1$: $$ L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},e_t}=\begin{pmatrix}0&-1&0&0\\1&0&0&0\\0&0&0&-1\\0&0&1&0\end{pmatrix}, \qquad L_{\mathrm{axf},\vec e_x}{R_{\mathrm{axf},e_t}}^{\mathsf T}=\begin{pmatrix}0&1&0&0\\-1&0&0&0\\0&0&0&1\\0&0&-1&0\end{pmatrix}. $$
The product, row by row. Row $1$ of $L_{\mathrm{axf},\vec e_x}{R_{\mathrm{axf},e_t}}^{\mathsf T}$ is $(0,1,0,0)$, so the first row of the product is row $2$ of $M$. Row $2$ is $(-1,0,0,0)$, so the second row is minus row $1$ of $M$. Row $3$ is $(0,0,0,1)$, so the third row is row $4$ of $M$. Row $4$ is $(0,0,-1,0)$, so the fourth row is minus row $3$ of $M$: $$ L_{\mathrm{axf},\vec e_x}{R_{\mathrm{axf},e_t}}^{\mathsf T}M= \begin{pmatrix} 0&1&0&0\\ -1&-2&0&0\\ 0&0&0&1\\ 0&0&-1&-2 \end{pmatrix}. $$ The trace is $0+(-2)+0+(-2)=-4$, hence $$ c_{xt}=\tfrac14(-4)=-1 . $$
The dot-product check. The signed positions of $L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},e_t}$ give $$ (-1)M_{12}+(+1)M_{21}+(-1)M_{34}+(+1)M_{43} =-2+0-2+0=-4 . $$
The matrix $L_{\mathrm{axf},e_t}R_{\mathrm{axf},\vec e_x}$. Its nonzero positions are $(1,2)=-1$, $(2,1)=+1$, $(3,4)=+1$, $(4,3)=-1$: $$ L_{\mathrm{axf},e_t}R_{\mathrm{axf},\vec e_x}=\begin{pmatrix}0&-1&0&0\\1&0&0&0\\0&0&0&1\\0&0&-1&0\end{pmatrix}, \qquad L_{\mathrm{axf},e_t}{R_{\mathrm{axf},\vec e_x}}^{\mathsf T}=\begin{pmatrix}0&1&0&0\\-1&0&0&0\\0&0&0&-1\\0&0&1&0\end{pmatrix}. $$ The product is $$ L_{\mathrm{axf},e_t}{R_{\mathrm{axf},\vec e_x}}^{\mathsf T}M= \begin{pmatrix} 0&1&0&0\\ -1&-2&0&0\\ 0&0&0&-1\\ 0&0&1&2 \end{pmatrix}, $$ whose trace is $0+(-2)+0+2=0$. Hence $$ c_{tx}=0 . $$
The trap. The two matrices agree in the first two rows and differ in the last two, and the two coefficients are different: $c_{xt}=-1$ and $c_{tx}=0$. A computation that prints $L_{\mathrm{axf},e_t}R_{\mathrm{axf},\vec e_x}$ and transposes $L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},e_t}$ returns $c_{tx}=-1$, transposes the coefficient matrix, and then fails at the reconstruction. The labels are not symmetric.
The ten remaining off-diagonal matrices
The table. Each row gives the signed positions of $L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu}$ and the resulting sum. Every corresponding entry of $M$ is zero except in the last row.
| Matrix | Signed positions | Sum | Coefficient |
|---|---|---|---|
| $L_{\mathrm{axf},e_t}R_{\mathrm{axf},\vec e_y}$ | $(1,3){-}1$, $(2,4){-}1$, $(3,1){+}1$, $(4,2){+}1$ | $0$ | $0$ |
| $L_{\mathrm{axf},e_t}R_{\mathrm{axf},\vec e_z}$ | $(1,4){-}1$, $(2,3){+}1$, $(3,2){-}1$, $(4,1){+}1$ | $0$ | $0$ |
| $L_{\mathrm{axf},\vec e_y}R_{\mathrm{axf},e_t}$ | $(1,3){-}1$, $(2,4){+}1$, $(3,1){+}1$, $(4,2){-}1$ | $0$ | $0$ |
| $L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},e_t}$ | $(1,4){-}1$, $(2,3){-}1$, $(3,2){+}1$, $(4,1){+}1$ | $0$ | $0$ |
| $L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},\vec e_y}$ | $(1,4){+}1$, $(2,3){-}1$, $(3,2){-}1$, $(4,1){+}1$ | $0$ | $0$ |
| $L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},\vec e_z}$ | $(1,3){-}1$, $(2,4){-}1$, $(3,1){-}1$, $(4,2){-}1$ | $0$ | $0$ |
| $L_{\mathrm{axf},\vec e_y}R_{\mathrm{axf},\vec e_x}$ | $(1,4){-}1$, $(2,3){-}1$, $(3,2){-}1$, $(4,1){-}1$ | $0$ | $0$ |
| $L_{\mathrm{axf},\vec e_y}R_{\mathrm{axf},\vec e_z}$ | $(1,2){+}1$, $(2,1){+}1$, $(3,4){-}1$, $(4,3){-}1$ | $+2+0-2+0=0$ | $0$ |
| $L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_x}$ | $(1,3){+}1$, $(2,4){-}1$, $(3,1){+}1$, $(4,2){-}1$ | $0$ | $0$ |
| $L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}$ | $(1,2){-}1$, $(2,1){-}1$, $(3,4){-}1$, $(4,3){-}1$ | $-2-0-2-0=-4$ | $-1$ |
The two cancellations. $L_{\mathrm{axf},\vec e_y}R_{\mathrm{axf},\vec e_z}$ and $L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}$ are the only off-diagonal matrices whose signed positions meet nonzero entries of $M$, and they meet the same two entries with opposite signs. The sum is $+2-2=0$ for $L_{\mathrm{axf},\vec e_y}R_{\mathrm{axf},\vec e_z}$ and $-2-2=-4$ for $L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}$. The cancellation in $L_{\mathrm{axf},\vec e_y}R_{\mathrm{axf},\vec e_z}$ is not an accident of this matrix: it is the sign difference between $L$ and $R$ on the same pair of indices.
The coefficient matrix
The result. $$ c(M)= \begin{pmatrix} 1&0&0&0\\ -1&0&0&0\\ 0&0&0&0\\ 0&0&-1&0 \end{pmatrix}. $$
The nonzero entries. Only three: $c_{tt}=1$, $c_{xt}=-1$, $c_{zy}=-1$. All the others vanish.
The Rank
The rows. Only rows $t$, $x$ and $z$ are nonzero, and rows $t$ and $x$ are proportional: row $x$ is minus row $t$. So the rank is at most $2$.
A minor. Take rows $\{t,z\}$ and columns $\{t,y\}$: $$ \det\begin{pmatrix}c_{tt}&c_{ty}\\c_{zt}&c_{zy}\end{pmatrix} =\det\begin{pmatrix}1&0\\0&-1\end{pmatrix}=-1\ne0 . $$ So the rank is exactly $2$.
The length. $$ \ell(M)=\operatorname{rank}c(M)=2 . $$
What the minor does not show. The minor shows that two rows and two columns are independent. It does not show which factorisation is simplest; that is the next section.
Reconstruction
The expansion. Since the only nonzero coefficients are $c_{tt}=1$, $c_{xt}=-1$ and $c_{zy}=-1$, $$ M=L_{\mathrm{axf},e_t}R_{\mathrm{axf},e_t}-L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},e_t}-L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}. $$
The three matrices. $$ L_{\mathrm{axf},e_t}R_{\mathrm{axf},e_t}=\begin{pmatrix}1&0&0&0\\0&1&0&0\\0&0&1&0\\0&0&0&1\end{pmatrix}, \qquad -L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},e_t}=\begin{pmatrix}0&1&0&0\\-1&0&0&0\\0&0&0&1\\0&0&-1&0\end{pmatrix}, \qquad -L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}=\begin{pmatrix}0&1&0&0\\1&0&0&0\\0&0&0&1\\0&0&1&0\end{pmatrix}. $$
The row sums. $$ \begin{aligned} \text{row }1&:\ (1,0,0,0)+(0,1,0,0)+(0,1,0,0)=(1,2,0,0),\\ \text{row }2&:\ (0,1,0,0)+(-1,0,0,0)+(1,0,0,0)=(0,1,0,0),\\ \text{row }3&:\ (0,0,1,0)+(0,0,0,1)+(0,0,0,1)=(0,0,1,2),\\ \text{row }4&:\ (0,0,0,1)+(0,0,-1,0)+(0,0,1,0)=(0,0,0,1). \end{aligned} $$ Each row is the row of $M$. The reconstruction is correct.
The wrong reconstruction. With the transposed coefficient matrix one would write $M=L_{\mathrm{axf},e_t}R_{\mathrm{axf},e_t}-L_{\mathrm{axf},e_t}R_{\mathrm{axf},\vec e_x}-L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}$. The first two rows still agree, but the third and fourth do not: $$ L_{\mathrm{axf},e_t}R_{\mathrm{axf},e_t}-L_{\mathrm{axf},e_t}R_{\mathrm{axf},\vec e_x}-L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y} =\begin{pmatrix}1&2&0&0\\0&1&0&0\\0&0&1&0\\0&0&2&1\end{pmatrix}, $$ which is not $M$. The two entries $M_{34}=2$ and $M_{43}=0$ are exchanged. This is the visible cost of the index slip.
The Factorisation
Split by rows. The coefficient matrix is $$ c(M)=e_t\,(1,0,0,0)+\vec e_x\,(-1,0,0,0)+\vec e_z\,(0,0,-1,0). $$ The first two terms share the right vector $(1,0,0,0)$, so $$ c(M)=(e_t-\vec e_x)\,(1,0,0,0)+\vec e_z\,(0,0,-1,0). $$
Read the two products. $$ a_1=e_t-\vec e_x,\quad b_1=e_t,\qquad a_2=\vec e_z,\quad b_2=-\vec e_y . $$ Hence $$ M=L_{\mathrm{axf},e_t-\vec e_x}R_{\mathrm{axf},e_t}+L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},-\vec e_y} =L_{\mathrm{axf},e_t-\vec e_x}-L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y} =L_{\mathrm{axf},e_t}R_{\mathrm{axf},e_t}-L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},e_t}-L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}. $$
The two matrices. $$ L_{\mathrm{axf},e_t-\vec e_x}=\mathrm{I}_4-L_{\mathrm{axf},\vec e_x}= \begin{pmatrix}1&1&0&0\\-1&1&0&0\\0&0&1&1\\0&0&-1&1\end{pmatrix}, \qquad -L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}=-L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}= \begin{pmatrix}0&1&0&0\\1&0&0&0\\0&0&0&1\\0&0&1&0\end{pmatrix}. $$ Their sum is $M$.
Why two is the minimum. The coefficient matrix has rank $2$. A single product would give a coefficient matrix of rank one. So no single two-sided operator equals $M$, and length $2$ is exact.
The two products are rank four. Each of $L_{\mathrm{axf},e_t-\vec e_x}$ and $L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}$ is an invertible matrix. The sum of two invertible matrices can be singular or not; here it is invertible. The length counts the terms, not the rank of the terms.
Ordinary Rank Is Not Length
The two numbers. $$ \operatorname{rank}M=4,\qquad \ell(M)=2 . $$
Why the ordinary rank is four. The matrix is upper triangular with unit diagonal, so $\det M=1$ and the four columns are independent.
Why the length is two. The coefficient matrix has rank two.
The matrix is not a left regular matrix. A left regular matrix $L_a$ has entries $(1,2)=-a_x$ and $(2,1)=+a_x$. The matrix $M$ has $M_{12}=2$ and $M_{21}=0$; it would need $a_x=-2$ and $a_x=0$. So $M$ is not of the form $L_a$: the right factor is needed. This is why the left regular model, of dimension four, cannot contain it.
The lesson. Ordinary rank measures the image of the matrix as a linear map. The length measures the least number of two-sided operators whose sum is the matrix. The two are independent: the identity has rank four and length one, the matrix unit $E_{11}$ has rank one and length four, and this matrix has rank four and length two.
The Recipe
The six steps.
- Compute the sixteen traces $$ c_{\mu\nu}=\tfrac14\sum_{i,j}(L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu})_{ij}M_{ij}. $$
- Assemble the $4\times4$ coefficient matrix $c(M)$.
- Compute its rank, by $2\times2$ minors or by row reduction. That rank is the length.
- If the rank is one, factor $c=a\,b^{\mathsf T}$ and read off $$ a=\sum_\mu a_\mu e_\mu,\qquad b=\sum_\nu b_\nu e_\nu,\qquad M=L_aR_b . $$
- If the rank is $r\ge2$, factor $c=UV$ with $U,V$ of size $4\times r$; the $r$ products are $L_{u_k}R_{v_k}$ for the columns $u_k$ of $U$ and $v_k$ of $V$.
- Reconstruct $M=\sum_{\mu,\nu}c_{\mu\nu}L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu}$ as a check.
The single-product test. It is step 3 with threshold one. It has nothing to do with the ordinary rank of $M$.
A remark on the factorisation. The factorisation of $c$ is not unique: any invertible $r\times r$ matrix $G$ gives $c=(UG)(G^{-1}V)$, and the $r$ products change while their sum does not. The row-split factorisation used above is the one that keeps each product as simple as possible.
Summary
| Quantity | Value |
|---|---|
| the matrix $M$ | $4\times4$ upper triangular, blocks $\begin{pmatrix}1&2\\0&1\end{pmatrix}$ |
| nonzero entries of $M$ | $M_{11}=1$, $M_{12}=2$, $M_{22}=1$, $M_{33}=1$, $M_{34}=2$, $M_{44}=1$ |
| nonzero coefficients | $c_{tt}=1$, $c_{xt}=-1$, $c_{zy}=-1$ |
| coefficient matrix | $\begin{pmatrix}1&0&0&0\\-1&0&0&0\\0&0&0&0\\0&0&-1&0\end{pmatrix}$ |
| rank of $c(M)$ | $2$, minor on rows $\{t,z\}$ and columns $\{t,y\}$ equal to $-1$ |
| length | $2$ |
| ordinary rank | $4$ |
| the determinant of the image | $1$ |
| minimal decomposition | $M=L_{\mathrm{axf},e_t-\vec e_x}R_{\mathrm{axf},e_t}-L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}=L_{\mathrm{axf},e_t}R_{\mathrm{axf},e_t}-L_{\mathrm{axf},\vec e_x}R_{\mathrm{axf},e_t}-L_{\mathrm{axf},\vec e_z}R_{\mathrm{axf},\vec e_y}$ |
The punchline. The ordinary rank is $4$, the maximal value, and the length is $2$. Ordinary rank gives no information about length; coefficient rank does.
The two traps. The index order is not symmetric: $c_{xt}=-1$ and $c_{tx}=0$, and a transposed coefficient matrix reconstructs a different matrix. The row split is not the column split: the terms combine as $(e_t-\vec e_x)(1,0,0,0)+\vec e_z(0,0,-1,0)$, not as $e_t(1,-1,0,0)+\vec e_z(0,0,-1,0)$.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $e_t,\vec e_x,\vec e_y,\vec e_z$ | the coefficient basis of $\mathbb{B}$, with $e_t=1$ and $\vec e_x^2=\vec e_y^2=\vec e_z^2=-e_t$ |
| $L_a$ | left multiplication $x\mapsto ax$ |
| $R_b$ | right multiplication $x\mapsto xb$ |
| $L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu}$ | $L_{\mathrm{axf},e_\mu}R_{\mathrm{axf},e_\nu}$, the sixteen basis matrices |
| $c_{\mu\nu}(A)$ | $\tfrac14\operatorname{tr}(L_{\mathrm{axf},e_\mu}{R_{\mathrm{axf},e_\nu}}^{*}A)$ |
| $c(A)$ | the $4\times4$ coefficient matrix |
| $\ell(A)$ | the length, $\operatorname{rank}c(A)$ |
| $L_{\mathrm{axf},\vec e_x}{R_{\mathrm{axf},e_t}}^{\mathsf T}$ | the transpose, equal to $L_{\mathrm{axf},\vec e_x}{R_{\mathrm{axf},e_t}}^{*}$ because the matrix is real |
Further Reading
- The Two-Sided Operators Method to Build $M_4(\mathbb{C})$, for the sixteen matrices, the multiplication tables, the orthogonality theorem and the length theorem.
- Operators Acting on Two Sides, for the construction.
- The Matrix Algebra $M_4(\mathbb{C})$, for the algebra and the matrix units.
- The Operators of the Peirce Decomposition of $M_4(\mathbb{C})$, for the same operators on the sixteen-dimensional algebra.