The Hermitian Pair of Operators Acting on One Side in $M_n(\mathbb{C})$

Introduction

The question. What are the Hermitian and anti-Hermitian parts of a one-sided operator, and what real and complex structures do they define on the carrier?

The answer in one line. The decomposition of $L_A$ is read off from the decomposition of $A$, $$ L_A=L_H+iL_K,\qquad H=\tfrac12(A+A^{*}),\quad K=\tfrac1{2i}(A-A^{*}), $$ so the self-adjoint part of the left copy is the real space of Hermitian matrices, of real dimension $n^2$, and it is not a complex subspace.

Why the article is needed. The biquaternion article could describe the Hermitian part as the real part of a $4\times4$ matrix and the anti-Hermitian part as the imaginary part. At general $n$ the split is the split of the parameter into its Hermitian and skew-Hermitian parts, and the two parts have different real dimensions and different algebraic roles: one is a Jordan algebra, the other a Lie algebra.

Boundaries. The adjoints are The Adjoint Pair of $L$ and $R$ Operators on $M_n(\mathbb{C})$; the two-sided Hermitian pairs are The Hermitian Pair of Operators Acting on Two Sides in $M_n(\mathbb{C})$; the cone of positive elements is Positivity and the Positive Cone of $M_n(\mathbb{C})$. The biquaternion original is The Hermitian Pair of Operators Acting on One Side.

Conventions. $H=\tfrac12(A+A^{*})$ and $K=\tfrac1{2i}(A-A^{*})$ are Hermitian. The Hilbert–Schmidt form is $\langle S,T\rangle=\operatorname{tr}(S^{*}T)$. The real vector space of Hermitian matrices is written $M_n^{\mathrm{sa}}$.

The Two Parts

Proposition. Every $A$ has a unique decomposition $A=H+iK$ with $H,K$ Hermitian, and correspondingly $$ L_A=L_H+iL_K,\qquad L_H=L_H^{\dagger},\qquad (iL_K)^{\dagger}=-iL_K . $$

Proof. $H$ and $K$ are Hermitian by construction, and uniqueness follows from $A^{*}=H-iK$. The adjoint of $L_H$ is $L_{H^{*}}=L_H$, and the adjoint of $iL_K$ is $-iL_K$. $\square$

The real dimensions. $M_n^{\mathrm{sa}}$ has real dimension $n^2$, and the skew-Hermitian matrices $iM_n^{\mathrm{sa}}$ also have real dimension $n^2$. Together they give the $2n^2$ real dimensions of $M_n$.

The two are not complex subspaces. Multiplication by $i$ exchanges them, so neither is invariant under the complex structure of $M_n$. This is the whole difference from the transpose split of Operator Chirality on $M_n(\mathbb{C})$, where both halves are complex subspaces.

Self-Adjointness

Proposition. $L_A$ is self-adjoint if and only if $A$ is Hermitian; $L_A$ is anti-self-adjoint if and only if $A$ is skew-Hermitian.

Proof. $L_A^{\dagger}=L_{A^{*}}$, and $L$ is injective. $\square$

The real space. The self-adjoint operators of the left copy form the real space $L(M_n^{\mathrm{sa}})$, of real dimension $n^2$. It is a real subspace and not a complex one: $iL_A=L_{iA}$ is anti-self-adjoint for nonzero Hermitian $A$.

The spectral reading. A self-adjoint one-sided operator has real spectrum, namely the spectrum of the Hermitian parameter. The spectral theorem for the one-sided operator is the spectral theorem for the parameter, transported by $L$.

Positivity

Proposition. $L_A\ge0$ if and only if $A\ge0$.

Proof. $\langle X,L_AX\rangle=\operatorname{tr}(X^{*}AX)$, which is nonnegative for all $X$ exactly when $A$ is positive semidefinite. $\square$

The cone. The positive one-sided operators form a closed convex cone isomorphic to the cone of positive semidefinite matrices. Its interior is the set of $L_A$ with $A>0$, and its extreme rays are the $L_A$ with $A$ of rank one.

The order. $L_A\le L_B$ exactly when $A\le B$, so the Loewner order on $M_n$ is transported to the left copy without change.

The trace. $\operatorname{tr}L_A=n\operatorname{tr}A$, so the trace functional on the left copy is $n$ times the trace of the parameter.

The Real and the Complex Structure

The conjugation. $\Theta(X)=X^{*}$ is a conjugate-linear involution of the carrier. It does not preserve the left copy: it sends it to the right copy, $\Theta L_A=R_{A^{*}}\Theta$.

The consequence for the Hermitian pair. The Hermitian part of $L_A$ is the fixed part under the adjoint, and the adjoint is not a real structure of the left copy alone. What the left copy carries is a real form in the weaker sense of a real subspace, $L(M_n^{\mathrm{sa}})$, of real dimension $n^2$.

The complexification. The complexification of $L(M_n^{\mathrm{sa}})$ is $L(M_n)$, the whole left copy: every $L_A$ is $L_H+iL_K$ with both parts in the real space. So the left copy is the complexification of the Hermitian one-sided operators.

The unitary group. $L_A$ is unitary exactly when $A$ is unitary, so the left copy contains a copy of $U(n)$ acting on the carrier. Its Lie algebra is the anti-self-adjoint part $iL(M_n^{\mathrm{sa}})$, of real dimension $n^2$, and the Hermitian part is the complement.

The polar decomposition. For $A=U\lvert A\rvert$ with $U$ unitary, $L_A=L_UL_{\lvert A\rvert}$, and $L_{\lvert A\rvert}$ is the positive part. So the polar decomposition of the one-sided operator is the polar decomposition of the parameter, and it is unique exactly when $A$ is invertible.

The Jordan and the Lie Structure

The Hermitian part is a Jordan algebra. For Hermitian $H,K$ the product $HK$ need not be Hermitian, but the symmetrised product $\tfrac12(HK+KH)$ is, and $M_n^{\mathrm{sa}}$ with this product is a real Jordan algebra. The map $H\mapsto L_H$ transports the Jordan product to the symmetrised product of operators.

The skew part is a Lie algebra. For skew-Hermitian $K_1,K_2$ the commutator is skew-Hermitian, and $iM_n^{\mathrm{sa}}$ with the commutator is the real Lie algebra $\mathfrak{u}(n)$, of dimension $n^2$.

The two fit together. $M_n$ as a real algebra is the complexification of both: the Jordan algebra $M_n^{\mathrm{sa}}$ and the Lie algebra $\mathfrak{u}(n)$ are two real forms of the same complex object, and the Hermitian pair of a one-sided operator sees both.

The Finite Check

Computed with numpy for $n=2,3,4,5$: for random complex $A$, the residual $\lVert L_A-L_H-iL_K\rVert$ is below $10^{-15}$, and $\lVert L_H-L_H^{\dagger}\rVert$, $\lVert(iL_K)^{\dagger}+iL_K\rVert$ are below $10^{-15}$. The real dimension of the self-adjoint part, computed as the rank of the real linear system $L_A=L_A^{\dagger}$ over the $2n^2$ real parameters, is $n^2$: it is $4,9,16,25$ for $n=2,3,4,5$.

The positivity test agrees in both directions: for $A$ sampled as $B^{*}B$, the smallest eigenvalue of $L_A$ is nonnegative, and for $A$ sampled as $B^{*}B-\lambda\mathrm I$ with $\lambda$ larger than the top eigenvalue of $B^{*}B$ it is negative; in both cases the sign agrees with that of the smallest eigenvalue of $A$.

The Three Real Subspaces

The self-adjoint part. The one-sided operators that are self-adjoint for the Hilbert–Schmidt form are the $L_A$ with $A$ Hermitian, a real vector space of dimension $n^2$.

The skew part. The skew-adjoint one-sided operators are the $L_A$ with $A$ skew-Hermitian, also of dimension $n^2$.

The unitary part. The one-sided operators that are unitary are the $L_U$ with $U$ unitary, a group isomorphic to $U(n)$, of real dimension $n^2$.

The table.

$n$ $\dim\{\text{self-adjoint}\}$ $\dim\{\text{Hermitian }A\}$ $\dim U(n)$
$2$ $4$ $4$ $4$
$3$ $9$ $9$ $9$
$4$ $16$ $16$ $16$
$5$ $25$ $25$ $25$

The reading. The three real subspaces have the same dimension, because the Hermitian and skew-Hermitian parts are the two halves of the real space of parameters and the unitary group is the exponentiated skew part.

The Polar Decomposition

Proposition. Every $L_A$ factors as $L_UL_{\lvert A\rvert}$ with $U$ unitary and $\lvert A\rvert$ positive semidefinite, and the factorisation is unique when $A$ is invertible.

Proof. $A=U\lvert A\rvert$ and $L$ is multiplicative. $\square$

The consequence. The one-sided operators decompose into a unitary part and a positive part, and the two parts commute. The positive part carries the singular values, the unitary part the phases.

The reading at $n=2$. The four real dimensions of the parameter give a circle of unitary parts and a three-dimensional cone of positive parts; the biquaternion description of the positive cone is the case $n=2$ of the same decomposition.

The Commutation with the Conjugation

Proposition. With $\Theta(X)=X^{*}$ one has $\Theta L_A\Theta=R_{A^{*}}$ and $\Theta R_B\Theta=L_{B^{*}}$.

Proof. $\Theta L_A\Theta(X)=(AX^{*})^{*}=XA^{*}=R_{A^{*}}(X)$. $\square$

The consequence. The conjugation exchanges the two one-sided families, so the real structure pairs a left multiplication with a right one. A one-sided operator is $\Theta$-real only when the two coincide, which requires $A$ central.

The reading. The Hermitian pair of a left multiplication is a right multiplication by the adjoint parameter, and that is the one-sided shadow of the two-sided statement $\Theta T_{A,B}\Theta=T_{B^{*},A^{*}}$.

The self-adjoint case. For $A$ Hermitian, $\Theta L_A\Theta=R_A$, so the conjugation carries the self-adjoint left family to the self-adjoint right family, and the two families are mirror images.

The Real Space of Parameters

The decomposition. $M_n=\mathfrak{u}(n)\oplus i\mathfrak{u}(n)$ as real spaces, so the parameter space splits into a skew part and a Hermitian part, each of real dimension $n^2$.

The images. The skew part maps to the unitary directions of $L$ and the Hermitian part to the self-adjoint directions, and the two images are orthogonal for the Hilbert–Schmidt form.

The reading. A one-sided operator is unitary exactly when its parameter is skew-Hermitian, and self-adjoint exactly when its parameter is Hermitian; the two conditions are the two real halves of one complex condition.

Summary

The Hermitian and anti-Hermitian parts of a one-sided operator are the one-sided operators of the Hermitian and skew-Hermitian parts of the parameter. The self-adjoint part is a real space of dimension $n^2$, a Jordan algebra, and the anti-self-adjoint part is $\mathfrak{u}(n)$, a Lie algebra. Positivity is transported without change. The adjoint operation itself exchanges the two copies, so it is a real structure of the pair rather than of one side.

Summary of Notation

Symbol Meaning
$H,K$ the Hermitian and skew-Hermitian parts of $A$
$M_n^{\mathrm{sa}}$ the real space of Hermitian matrices, dimension $n^2$
$L(M_n^{\mathrm{sa}})$ the self-adjoint one-sided operators
$\Theta$ the conjugation $X\mapsto X^{*}$
$\mathfrak{u}(n)$ the skew-Hermitian matrices with the commutator

Further Reading