One-Sided Operators on the N-Degenerate Sector in $M_2(\mathbb{C})$
Introduction
The question. On the null cone $\det A=0$, $A\neq0$, the matrix is singular, so $L_A(Y)=AY$ and $R_A(Y)=YA$ drop rank. What are their ranks, kernels, spectra, Jordan types and one-parameter groups, on the two strata $\Sigma_{01}$ and $\Sigma_{00}$?
Why it is the same statement. This is the in-depth matrix reading of One-Sided Operators on the N-Degenerate Sector. The source's polar pair, its rank-one singular value decomposition and its two strata become, under $A=\mathsf M_2(\tilde Q)$, the rank-one matrix $A=2auv^{\dagger}$; the two slots become the two unit vectors $u$ and $v$; the overlap $v^{\dagger}u$ becomes the trace. The statements are the same statements, and only the writing is new.
What the article proves. In this order. The two strata as matrix sets. The rank-one decomposition $A=2auv^{\dagger}$ with $a=\lVert A\rVert_F/2$, and the two slots. The overlap $v^{\dagger}u=\operatorname{tr}A/(2a)$ and its criterion for the strata. The polar pair $P=uu^{\dagger}$, $A_H=2auu^{\dagger}$, $A_0=uv^{\dagger}$. The closed form on $\Sigma_{01}$ with $\kappa$ and $A_0^2=\frac{Q_t}{a}A_0$. The canonical Hermitian pair on $\Sigma_{00}$. The one-sided operators: rank $2$, two-dimensional kernel, minimal right ideal as image, spectra $c,c,0,0$ and $0,0,0,0$. Their exponentials.
What the article does not do. It does not treat the two-sided operator, which is Two-Sided Operators on the N-Degenerate Sector in $M_2(\mathbb{C})$. It does not re-derive the polar decomposition of a singular matrix, which is The Polar Decomposition from the Two Halves at $n=2$.
A note on the source. The source's text stops after the canonical Hermitian pair, although its plan lists further operator sections. The operator statements below are therefore computed from the display and verified numerically; they agree with every statement the source does make, and the place where the source is silent is marked in the body.
Conventions. The display is $\mathsf M_2(\tilde Q)=Q_t\mathrm I_2-i(Q_x\sigma_1+Q_y\sigma_2+Q_z\sigma_3)$, so $\operatorname{tr}A=2Q_t$ and $\det A=N=Q_t^2+B^2$. The null cone is $\det A=0$ with $A\neq0$. Its strata are $$ \Sigma_{01}=\{\det A=0,\ B^2\neq0\},\qquad \Sigma_{00}=\{\det A=0,\ B^2=0\}. $$ The Hermitian adjoint is $A^{\dagger}=\overline{A^{\mathsf T}}$. The Frobenius norm is $\lVert A\rVert_F$, and the scalar $a=\lVert A\rVert_F/2$ is the non-zero singular value.
Boundaries. The general rank trichotomy is Introduction to the Operators on the Four Sectors in $M_2(\mathbb{C})$. The invertible case is One-Sided Operators on the N-Regular Sector in $M_2(\mathbb{C})$. The rank-one projections are The Spin-Half Bispinors and N Norm Regularity in $M_2(\mathbb{C})$. The $n=2$ spine is The One-Sided Operators at $n=2$ and The Two-Sided Operators and the Segre Cone at $n=2$, and the display is $M_2(\mathbb{C})$ Pauli Representation : Operators.
Letters. $A$ is the parameter matrix, $u,v$ the two slots, $a$ the singular value, $P,P'$ the rank-one projections, $\Pi$ an idempotent, $\nu$ a nilpotent, $\kappa$ the obliqueness.
The Two Strata as Matrix Sets
The regular stratum. $\Sigma_{01}$ consists of the non-zero complex multiples of non-trivial idempotents: $$ A=c\Pi,\qquad \Pi^2=\Pi,\qquad \Pi\neq0,\mathrm I_2,\qquad c=2Q_t\neq0,\qquad \operatorname{tr}\Pi=1 . $$ A rank-one idempotent always has trace one, so $c=\operatorname{tr}A$ and $Q_t=\operatorname{tr}A/2$.
The degenerate stratum. $\Sigma_{00}$ consists of the non-zero nilpotents: $$ A=\nu,\qquad \nu^2=0,\qquad \nu\neq0 . $$ Every non-zero nilpotent matrix is similar to $\begin{pmatrix}0&1\\0&0\end{pmatrix}$, and its norm vanishes: $\det\nu=0$ and $\operatorname{tr}\nu=0$, so $Q_t=0$ and $B^2=0$.
The example. $A=\begin{pmatrix}0&1+2i\\0&0\end{pmatrix}$ has $\det A=0$, $\operatorname{tr}A=0$, and $A^2=0$; it is on $\Sigma_{00}$. The matrix $A=c\Pi$ with $\Pi=\frac{1}{v^{\dagger}u}uv^{\dagger}$ is on $\Sigma_{01}$ for every $c\neq0$.
The Rank-One Decomposition
The theorem. A non-zero $A$ has $\det A=0$ exactly when $$ A=2a\,uv^{\dagger},\qquad \lVert u\rVert=\lVert v\rVert=1,\qquad a>0, $$ and then $a=\lVert A\rVert_F/2$ and $A_0=uv^{\dagger}$ is a partial isometry, $$ A_0A_0^{\dagger}A_0=A_0,\qquad \operatorname{rank}A_0=1. $$
The proof. A non-zero matrix of determinant zero has rank one, so $A=\sigma uv^{\dagger}$ with $\sigma$ the non-zero singular value; setting $2a=\sigma$ gives the form. Conversely a rank-one matrix has determinant zero. The Frobenius norm is $\lVert A\rVert_F=2a$, and $(uv^{\dagger})(vu^{\dagger})(uv^{\dagger})=uv^{\dagger}$.
The two slots. The vectors $u$ and $v$ are the left and the right slot of the rank-one matrix: $u$ spans the image and $v^{\dagger}$ spans the row space. They are defined up to a common phase, and the relative phase is the phase of $A$.
The overlap. By the cyclic property of the trace, $$ v^{\dagger}u=\frac{\operatorname{tr}A}{2a}=\frac{2Q_t}{2a}=\frac{Q_t}{a},\qquad \lvert v^{\dagger}u\rvert\le1 . $$ Its modulus is the cosine of the angle between the two slots. The stratum criterion is the vanishing of the overlap: $$ \Sigma_{01}\iff v^{\dagger}u\neq0,\qquad \Sigma_{00}\iff v^{\dagger}u=0 . $$
The check. On $A=\begin{pmatrix}0&1+2i\\0&0\end{pmatrix}$ the decomposition is $u=(1,0)$, $v=(0,1)$, and $v^{\dagger}u=0$, as $\Sigma_{00}$ requires. On the idempotent $A=c\Pi$ with $\Pi=uv^{\dagger}/(v^{\dagger}u)$ the overlap is $v^{\dagger}u$, and $\operatorname{tr}\Pi=1$.
The Polar Pair
The factorisation. For non-zero $A$ on the null cone put $$ a=\frac{\lVert A\rVert_F}{2},\qquad P=\frac{AA^{\dagger}}{4a^2},\qquad A_H=2aP,\qquad A_0=\frac{A}{2a}. $$ Then $P$ is a rank-one Hermitian projection, $A_H$ is positive semidefinite of rank one with spectrum $\{0,2a\}$, $A_0$ is a partial isometry with $A_0A_0^{\dagger}=P$ and $A_0^{\dagger}A_0=P'$, and $$ A=A_HA_0=2aA_0 . $$ In the slots, $P=uu^{\dagger}$, $P'=vv^{\dagger}$, $A_H=2auu^{\dagger}$ and $A_0=uv^{\dagger}$; verified to $10^{-16}$.
What it is not. This is not the modulus–unit decomposition: $\det A=0$, so the complex modulus of the source vanishes, while $A_H$ is a non-zero rank-one matrix. The number $2a$ is the singular value, not an eigenvalue, except on the Hermitian part of $\Sigma_{01}$.
The comparison with the invertible case. On the N-regular sector the modulus is a scalar and the operator is a bijection; here the positive factor has rank one, and the collapse $$ L_A=2a\,L_{A_0} $$ is the whole content: the one-sided operator is a positive scalar times the operator of a partial isometry.
The Closed Form on $\Sigma_{01}$
The parameters. Let $A=c\Pi$ with $\Pi^2=\Pi$, $\operatorname{tr}\Pi=1$ and $c\neq0$. Put $$ \kappa=\frac{a^2}{\lvert Q_t\rvert^2}=\frac{1}{\lvert v^{\dagger}u\rvert^2}\ge1 . $$ Then $$ a=\lvert Q_t\rvert\sqrt{\kappa},\qquad P=\frac{\Pi\Pi^{\dagger}}{\kappa},\qquad A_H=\lvert Q_t\rvert\sqrt{\kappa}\,P,\qquad A_0=\frac{c}{\lvert c\rvert}\frac{\Pi}{\sqrt{\kappa}} . $$ Verified on a non-Hermitian idempotent: $a=1.80156$, $\kappa=2.07721$, $\lvert v^{\dagger}u\rvert=1/\sqrt{\kappa}=0.693841$.
The angle between the slots. $\cos\theta=\lvert v^{\dagger}u\rvert=1/\sqrt{\kappa}$. The idempotent is Hermitian exactly when the slots coincide, $\kappa=1$ and $\cos\theta=1$; the more oblique the idempotent, the larger $\kappa$.
The square of the normalized element. On $\Sigma_{01}$, $$ A_0^2=\frac{Q_t}{a}A_0, $$ verified to $3\cdot10^{-17}$. So $A_0/(Q_t/a)=\Pi$ is an idempotent: the normalized element is a scalar times an idempotent, not a nilpotent.
The failure at the wall. The scalar $Q_t/a$ vanishes exactly when the overlap does, that is on $\Sigma_{00}$; there the normalized element is nilpotent and no scalar multiple of it is an idempotent.
The Canonical Hermitian Pair on $\Sigma_{00}$
The nilpotent form. Write $\nu=\vec A+i\vec B$ with $\vec A,\vec B$ real, orthogonal and of equal norm $a>0$. Its matrix is $$ A=\vec B\cdot\sigma-i\,\vec A\cdot\sigma, $$ that is $B\cdot\sigma-iA\cdot\sigma$, and $A^2=(|B|^2-|A|^2)\mathrm I_2-2i(A\cdot B)\mathrm I_2=0$.
The pair. With $\vec c=\vec A\times\vec B/a^2$, a real unit vector, $$ \Pi_+=\frac{\mathrm I_2+\vec c\cdot\sigma}{2},\qquad \Pi_-=\frac{\mathrm I_2-\vec c\cdot\sigma}{2}, $$ both Hermitian rank-one projections with $\Pi_++\Pi_-=\mathrm I_2$, $\Pi_+\Pi_-=0$. They satisfy $$ \Pi_+A=A,\qquad A\Pi_+=0,\qquad \Pi_-A=0,\qquad A\Pi_-=A, $$ and $$ AA^{\dagger}=4a^2\Pi_+,\qquad A^{\dagger}A=4a^2\Pi_-,\qquad \operatorname{tr}(AA^{\dagger})=4a^2 . $$ All the identities are immediate from $(a\cdot\sigma)(b\cdot\sigma)=a\cdot b\,\mathrm I_2+i(a\times b)\cdot\sigma$.
The check. For $A=\begin{pmatrix}0&1+2i\\0&0\end{pmatrix}$ the singular value is $2a=\sqrt5$, $u=(1,0)$ and $v=(0,1)$, so $\Pi_+=uu^{\dagger}=E_{11}$ and $\Pi_-=vv^{\dagger}=E_{22}$; and indeed $E_{11}A=A$ and $AE_{22}=A$.
The One-Sided Operators
The rank. For $A$ of rank one, $L_A(Y)=A(Y)=2a\,u(v^{\dagger}Y)$ is zero exactly when $v^{\dagger}Y=0$. That is two independent linear conditions on $Y$, the columns of $Y$ lying in the hyperplane orthogonal to $v$. Hence $$ \ker L_A=\{Y:v^{\dagger}Y=0\},\qquad \dim\ker L_A=2,\qquad \operatorname{rank}L_A=2 . $$ Symmetrically $YA=2a(Yu)v^{\dagger}$ vanishes exactly when $Yu=0$, so $\ker R_A=\{Y:Yu=0\}$, also of dimension two. Verified on both strata.
The images. The image of $L_A$ is $$ A M_2(\mathbb{C})=\left\{u\,w^{\dagger}:w\in\mathbb{C}^2\right\}, $$ the minimal right ideal of matrices all of whose columns are proportional to $u$, of complex dimension two; it is $A_0M_2$, a spinor space. The image of $R_A$ is $M_2A=\{z\,v^{\dagger}\}$, the minimal left ideal of matrices whose columns are proportional to $u$. On $\Sigma_{01}$ the two ideals are distinct and meet in $\mathbb{C}A$.
The spectra. $L_A=\mathrm I_2\otimes A$ has the eigenvalues of $A$, each twice: $$ \Sigma_{01}:\ \operatorname{spec}L_A=\{c,c,0,0\},\qquad \Sigma_{00}:\ \operatorname{spec}L_A=\{0,0,0,0\}. $$ Verified on both strata with agreement to $10^{-15}$.
The Jordan types. $$ \Sigma_{01}:\ L_A\ \text{is diagonalizable, of type}\ \operatorname{diag}(c,c,0,0);\qquad \Sigma_{00}:\ L_A\ \text{is nilpotent, of type}\ J_2(0)\oplus J_2(0). $$ On $\Sigma_{01}$ the rank of $L_A-c\mathrm I_4$ is $2$, so the geometric multiplicity of $c$ is two. On $\Sigma_{00}$ the ranks of $L_A^k$ are $4,2,0$, so $L_A^2=0$ and the type is $2+2$.
The operator of a projection. For an idempotent $\Pi$, $L_\Pi$ is itself idempotent, $L_\Pi^2=L_{\Pi^2}=L_\Pi$, but not Hermitian unless $\Pi$ is. So $L_A$ is $2a$ times a projection of rank two, and the two-sided operator of the same $A$ is a projection of rank one: the one-sided reading keeps the rank of the slot, the two-sided one keeps the rank of the product.
The Exponential
The element identity survives. Since $L_A^n=L_{A^n}$ at every point, $$ \exp\left(tL_A\right)=L_{\exp(tA)} . $$
On $\Sigma_{01}$. With $A=c\Pi$ and $\Pi^2=\Pi$, $$ \exp(tA)=\mathrm I_2+\left(e^{tc}-1\right)\Pi,\qquad \exp\left(tL_A\right)=\mathrm I_4+\left(e^{tc}-1\right)L_\Pi, $$ verified to $4\cdot10^{-16}$. The group is the identity plus a multiple of a rank-two projection: a one-parameter family of affine maps along one direction of the operator space.
On $\Sigma_{00}$. With $A^2=0$, $$ \exp(tA)=\mathrm I_2+tA,\qquad \exp\left(tL_A\right)=\mathrm I_4+tL_A, $$ verified exactly. The group is a transvection, and the operator $L_A$ is its infinitesimal generator.
The contrast. On $\Sigma_{01}$ the group has the eigenvalue $e^{tc}$ twice and $1$ twice; on $\Sigma_{00}$ it is unipotent with a single eigenvalue $1$ and two blocks of size two.
Worked Examples
A Hermitian idempotent. $A=\operatorname{diag}(2,0)$: $\Pi=\operatorname{diag}(1,0)$, $c=2$, $a=1$, $\kappa=1$, $v^{\dagger}u=1$. Then $L_A=\operatorname{diag}(2,2,0,0)$ in the appropriate basis, of rank two, and $\exp(tL_A)=\operatorname{diag}(e^{2t},e^{2t},1,1)$.
A non-Hermitian idempotent. The idempotent of the closed-form section has $\kappa=2.07721>1$, so the two slots are at an angle $\arccos(1/\sqrt\kappa)\approx46.6^{\circ}$.
The standard nilpotent. $A=\begin{pmatrix}0&1\\0&0\end{pmatrix}$: $L_A$ has rank two and $L_A^2=0$, so its type is $J_2(0)\oplus J_2(0)$, and $\exp(tL_A)=\mathrm I_4+tL_A$.
A complex scalar multiple. $A=c\Pi$ with $\Pi$ Hermitian: $\kappa=1$, $a=|Q_t|$, and the exponential is $\mathrm I_4+(e^{tc}-1)L_\Pi$.
Where $n=2$ Differs
The rank. At $n=2$ a singular non-zero matrix has rank one, so the singular value decomposition has one term and the partial isometry is $uv^{\dagger}$. At order $n$ the rank can be anything from $1$ to $n-1$ and the decomposition has that many terms; the one-sided operator then has rank $2r$ and kernel of dimension $n^2-2r$.
The strata. At $n=2$ the null cone splits into exactly two strata, $\Sigma_{01}$ and $\Sigma_{00}$; there is no third. At order $n$ the discriminant is a form on a larger space and the Jordan types are partitions of $n^2$.
The canonical pair. At $n=2$ the nilpotent determines exactly two canonical rank-one projections $\Pi_\pm$, and they are the whole Peirce decomposition. At order $n$ a nilpotent of rank $r$ has a Peirce decomposition with more corners.
What does not differ. The scalar collapse $L_A=2a\,L_{A_0}$, the rank $2r$, the minimal ideals as images and the element identity for the exponential hold at every order.
Summary
On the null cone $\det A=0$, $A\neq0$, the matrix is rank one, $A=2auv^{\dagger}$ with $a=\lVert A\rVert_F/2$, and the overlap $v^{\dagger}u=\operatorname{tr}A/(2a)$ distinguishes the strata: $\Sigma_{01}$ has $v^{\dagger}u\neq0$ and $\Sigma_{00}$ has $v^{\dagger}u=0$. The polar pair is $P=uu^{\dagger}$, $A_H=2auu^{\dagger}$, $A_0=uv^{\dagger}$, and $A=A_HA_0=2aA_0$.
On $\Sigma_{01}$ the matrix is $c\Pi$ with $\Pi$ idempotent, $a=|Q_t|\sqrt\kappa$, $\cos\theta=1/\sqrt\kappa$, and $A_0^2=(Q_t/a)A_0$. On $\Sigma_{00}$ the canonical pair $\Pi_\pm=(\mathrm I_2\pm c\cdot\sigma)/2$ satisfies $\Pi_+A=A$ and $A\Pi_-=A$, with $AA^{\dagger}=4a^2\Pi_+$.
The one-sided operators have rank $2$, kernel of dimension two, and images the minimal right and left ideals $AM_2$ and $M_2A$. Their spectra are $c,c,0,0$ and $0,0,0,0$; their Jordan types are $\operatorname{diag}(c,c,0,0)$ and $J_2(0)\oplus J_2(0)$. The exponential is $\mathrm I_4+(e^{tc}-1)L_\Pi$ on $\Sigma_{01}$ and $\mathrm I_4+tL_A$ on $\Sigma_{00}$.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $A$ | the parameter matrix on the null cone |
| $u,v$ | the left and right slots |
| $a=\lVert A\rVert_F/2$ | the non-zero singular value |
| $P,P'$ | $uu^{\dagger}$ and $vv^{\dagger}$ |
| $\Pi$ | the idempotent with $A=c\Pi$ on $\Sigma_{01}$ |
| $\nu$ | the nilpotent on $\Sigma_{00}$ |
| $\kappa=a^2/|Q_t|^2$ | the obliqueness |
| $\Pi_\pm$ | the canonical Hermitian pair |
Further Reading
- One-Sided Operators on the N-Degenerate Sector, for the source's rank-one decomposition and its two strata.
- Two-Sided Operators on the N-Degenerate Sector in $M_2(\mathbb{C})$, for the rank-one two-sided operator and its single Jordan chain.
- One-Sided Operators on the N-Regular Sector in $M_2(\mathbb{C})$, for the invertible case that this article deforms.
- The Polar Decomposition from the Two Halves at $n=2$, for the polar decomposition of a singular matrix.
- Introduction to the Operators on the Four Sectors in $M_2(\mathbb{C})$, for the four sector table in one page.