The Polar Decomposition of $M_6(\mathbb{C})$

Introduction

Every $A\in M_6(\mathbb{C})$ factors as

$$ A=UP, \qquad U\in U(6),\quad P=(A^*A)^{1/2}\ge0, $$

with $P$ positive semidefinite and Hermitian. The factor $P$ is unique; the $*$-unitary factor $U$ is unique when $A$ is invertible, and is free on the kernel when $A$ is singular. The decomposition separates the rotation $U$ from the scale $P$, and it is the matrix form of the singular value decomposition

$$ A=U\Sigma V^*, \qquad P=V\Sigma V^*, \qquad U_{\mathrm{polar}}=UV^* . $$

The quotient $GL_6/U(6)$ is the manifold of positive definite Hermitian matrices, of real dimension $36$; the polar decomposition is the diffeomorphism

$$ GL_6\cong U(6)\times\operatorname{Herm}_6^{>0}. $$

The exponential form is $A=Ue^H$ with $H$ Hermitian; the logarithm of the positive factor is the Hermitian exponent.

The article assumes the $*$-unitary group and the Hermitian form of The Involution and the Hermitian Structure of $M_6(\mathbb{C})$ and the singular values of The Forms, the Norms and the Singular Set of $M_6(\mathbb{C})$. It uses the exponential of Hermitian matrices, whose general theory is The Exponential and the Lie Group of $M_6(\mathbb{C})$. The mass block's polar decomposition is The Mass Block and the Chiral Polar Decomposition of $M_6(\mathbb{C})$; the biquaternion case is The Polar Decomposition of a Biquaternion.

The Polar Decomposition

Existence

Theorem (the right polar decomposition). Every $A\in M_6(\mathbb{C})$ has a factorization

$$ A=UP, \qquad U\in U(6),\quad P=P^*\ge0, $$

with $P=(A^*A)^{1/2}$.

Proof. $A^*A$ is positive semidefinite, so it has a unique positive semidefinite square root $P$. On the image of $P$, define $U(Px)=Ax$. It is well defined: if $Px=Py$, then $P(x-y)=0$ and

$$ \lVert A(x-y)\rVert^2=\langle x-y,A^*A(x-y)\rangle=\langle x-y,P^2(x-y)\rangle=0, $$

so $Ax=Ay$. It is isometric:

$$ \lVert Ax\rVert^2=\langle x,A^*Ax\rangle=\langle x,P^2x\rangle=\lVert Px\rVert^2 . $$

Extend $U$ from the image of $P$ to its orthogonal complement, the kernel of $P$, by any $*$-unitary isomorphism onto $(\operatorname{im}A)^\perp$; both complements have dimension $6-\operatorname{rank}A$. Then $U$ is $*$-unitary and $UPx=Ax$ for all $x$, so $A=UP$.

Corollary (the left polar decomposition). Every $A$ has a factorization

$$ A=P'U, \qquad P'=(AA^*)^{1/2}\ge0,\quad U\in U(6), $$

and $P'=UPU^*$.

Proof. Apply the right polar decomposition to $A^*$: $A^*=VQ$, so $A=QV^*$ with $Q=(AA^*)^{1/2}$ and $V^*\in U(6)$.

Uniqueness

Theorem. The positive factor $P=(A^*A)^{1/2}$ is unique. If $A$ is invertible, the $*$-unitary factor is unique as well.

Proof. $P$ is determined by $A^*A$. If $A=U_1P_1=U_2P_2$ with $P_i>0$, then $A^*A=P_1^2=P_2^2$, and the positive definite square root is unique, so $P_1=P_2$. Then $U_1P_1=U_2P_1$ and $P_1$ is invertible, so $U_1=U_2$.

Corollary (the singular case). If $A$ has rank $k<6$, the positive factor is unique, and the $*$-unitary factor is determined on the image of $P$ and arbitrary on the kernel:

$$ U\big|_{\ker P}:\ker P\longrightarrow(\operatorname{im}A)^{\perp} \ \text{any unitary isomorphism}. $$

Proof. $P$ is unique; on the image of $P$ the map $U$ is determined by $A$; on the kernel of $P$ any unitary isomorphism onto $(\operatorname{im}A)^{\perp}$ is allowed. The freedom is a torsor under the $*$-unitary group of the $(6-k)$-dimensional kernel.

The Singular Value Decomposition

Theorem (the SVD). Every $A$ has a factorization

$$ A=U\Sigma V^*, \qquad U,V\in U(6),\quad \Sigma=\operatorname{diag}(\sigma_1,\dots,\sigma_6)\ge0 . $$

The $\sigma_i$ are the singular values, and

$$ P=V\Sigma V^*, \qquad U_{\mathrm{polar}}=UV^* . $$

Proof. Diagonalize $A^*A=V\Sigma^2V^*$. The polar decomposition gives $A=U_{\mathrm{polar}}P$ with $P=V\Sigma V^*$; set $U=U_{\mathrm{polar}}V$. Then

$$ A=U_{\mathrm{polar}}V\Sigma V^*=U\Sigma V^*, $$

and the polar factor is $U_{\mathrm{polar}}=UV^*$ because

$$ A=(UV^*)(V\Sigma V^*). $$

Corollary (the polar decomposition is the SVD with phases separated). The SVD separates the two unitaries $U,V$; the polar decomposition combines them into $U_{\mathrm{polar}}=UV^*$ and keeps the positive factor $V\Sigma V^*$.

Proof. Multiply the two factors.

Corollary (the singular values are the eigenvalues of $P$). The eigenvalues of $P$ are $\sigma_1,\dots,\sigma_6$; the eigenvalues of $U_{\mathrm{polar}}$ are the phases of the SVD.

Proof. $P=V\Sigma V^*$; the $*$-unitary factor $UV^*$ has eigenvalues on the unit circle.

The Manifold $GL_6/U(6)$

The Positive Cone

Definition. The positive cone is

$$ \operatorname{Herm}_6^{>0}=\{P\in M_6(\mathbb{C}):P=P^*,\ P>0\}. $$

Theorem. $\operatorname{Herm}_6^{>0}$ is an open convex cone in the real vector space $\operatorname{Herm}_6$ of real dimension $36$.

Proof. The Hermitian matrices form a real vector space of dimension $36$; the positive definite ones are defined by the positivity of the six principal minors, which are open conditions; convexity follows from $\langle x,((1-t)P+tQ)x\rangle=(1-t)\langle x,Px\rangle+t\langle x,Qx\rangle>0$.

The Symmetric Space

Theorem (the polar diffeomorphism). The map

$$ U(6)\times\operatorname{Herm}_6^{>0}\longrightarrow GL_6, \qquad (U,P)\longmapsto UP $$

is a diffeomorphism. Hence

$$ GL_6/U(6)\cong\operatorname{Herm}_6^{>0}, \qquad \dim_{\mathbb{R}}GL_6/U(6)=36 . $$

Proof. The polar decomposition gives the map and its inverse $A\mapsto(U,P)$; it is smooth because $P=(A^*A)^{1/2}$ and $U=AP^{-1}$ are smooth on the invertible locus. The dimensions add: $36+36=72$.

Corollary (the principal bundle). $GL_6$ is a principal $U(6)$-bundle over the positive cone, with fibre $U(6)$.

Proof. The action of $U(6)$ on $GL_6$ by left multiplication is free, and the quotient is the positive cone by the polar decomposition.

Remark (the symmetric space). The quotient is the symmetric space of the Cartan involution $\theta(A)=-A^*$; the geodesics are the curves $P^{1/2}e^{tX}P^{1/2}$ with $X$ Hermitian, and through the identity they are the exponentials $e^{tX}$. The metric is the Hilbert–Schmidt metric.

The Singular Element

The Freedom on the Kernel

Theorem. Let $A$ have rank $k$ and polar decomposition $A=U_0P$. Then the polar factors of $A$ are exactly

$$ U=U_0V, \qquad V\in U(\ker P)\cong U(6-k), $$

with $V$ acting as the identity on the image of $P$. The set is a coset of $U(6-k)$, of real dimension $(6-k)^2$.

Proof. Any other $*$-unitary factor differs from $U_0$ by a $*$-unitary that commutes with $P$ on the image and is arbitrary on the kernel; the kernel is $\ker P=\ker A$, of dimension $6-k$.

Corollary (the image and the kernel). The polar factor maps

$$ \operatorname{im}P\longrightarrow\operatorname{im}A, \qquad \ker P\longrightarrow(\operatorname{im}A)^\perp, $$

and is arbitrary on the second factor.

Proof. $A=UP$; $U$ is $*$-unitary, so $U((\operatorname{im}P)^\perp)=(\operatorname{im}A)^\perp$; the kernel of $P$ is the orthogonal complement of the image of $P$.

The Limit at a Singular Element

Theorem. The positive factor $P(A)=(A^*A)^{1/2}$ is continuous everywhere. The $*$-unitary factor $U(A)$ is continuous on the invertible locus; at a singular element it may fail to have a limit, and when it has one, the limit depends on the direction of approach.

Proof. The square root is continuous; $U=AP^{-1}$ is continuous where $P$ is invertible. On the singular locus, the freedom on the kernel is a coset, and different approaches select different points of the coset.

Example. For $A(\varepsilon)=\operatorname{diag}(1,\varepsilon e^{\mathrm i\theta})$ with $\varepsilon>0$, the polar factor is $\operatorname{diag}(1,e^{\mathrm i\theta})$, which tends to $\operatorname{diag}(1,e^{\mathrm i\theta})$ as $\varepsilon\to0$; the limit depends on $\theta$. The positive factor tends to $\operatorname{diag}(1,0)$ independently of $\theta$.

Corollary (the polar map is not continuous at the singular set). The map $A\mapsto U(A)$ does not extend continuously to the singular matrices.

Proof. The example gives two approaches with different limits.

The Exponential Form

The Exponential of Hermitian Matrices

Theorem. The exponential map

$$ \exp:\operatorname{Herm}_6\longrightarrow\operatorname{Herm}_6^{>0}, \qquad H\longmapsto e^H $$

is a diffeomorphism.

Proof. For Hermitian $H$, $e^H$ is Hermitian positive definite (diagonalize $H$ unitarily: $H=U\operatorname{diag}(h_i)U^*$, $e^H=U\operatorname{diag}(e^{h_i})U^*$). The inverse is the logarithm $P\mapsto\log P$, defined by the spectral theorem; both maps are smooth.

The Polar Decomposition in Exponential Coordinates

Theorem. Every invertible $A$ has a unique factorization

$$ A=Ue^H, \qquad U\in U(6),\quad H\in\operatorname{Herm}_6 . $$

Proof. The polar decomposition $A=UP$ with $P>0$; write $P=e^H$ with $H=\log P$ Hermitian. Uniqueness follows from the uniqueness of the polar decomposition and of the logarithm.

Corollary (the Cartan decomposition). The map

$$ U(6)\times\operatorname{Herm}_6\longrightarrow GL_6, \qquad (U,H)\longmapsto Ue^H $$

is a diffeomorphism, and

$$ \mathfrak{gl}_6=\mathfrak{u}(6)\oplus\mathrm i\mathfrak{u}(6), \qquad \dim_{\mathbb{R}}GL_6=36+36=72 . $$

Proof. The polar diffeomorphism and the exponential diffeomorphism compose; the Lie algebra decomposition is the derivative at the identity.

The Exponential Form of the SVD

Theorem. With $A=U\Sigma V^*$ and $\Sigma>0$,

$$ A=U_{\mathrm{polar}}e^H, \qquad U_{\mathrm{polar}}=UV^*, \qquad H=V(\log\Sigma)V^* . $$

Proof. $V\Sigma V^*=e^{V(\log\Sigma)V^*}$; multiply by $UV^*$.

Corollary (the singular values are the exponents). The eigenvalues of $H$ are $\log\sigma_1,\dots,\log\sigma_6$; the singular values are the exponentials of the Hermitian spectrum.

Proof. $H$ is unitarily equivalent to $\operatorname{diag}(\log\sigma_i)$.

Summary

Every $A\in M_6(\mathbb{C})$ has the polar decomposition

$$ A=UP, \qquad U\in U(6),\quad P=(A^*A)^{1/2}\ge0 . $$

The positive factor is unique; the $*$-unitary factor is unique for invertible $A$ and free on the kernel for singular $A$, with freedom $U(6-k)$ of real dimension $(6-k)^2$. The SVD is

$$ A=U\Sigma V^*, \qquad P=V\Sigma V^*, \qquad U_{\mathrm{polar}}=UV^* . $$

$$ \boxed{\ A=UP,\qquad GL_6/U(6)\cong\operatorname{Herm}_6^{>0},\qquad \dim_{\mathbb{R}}=36,\qquad A=Ue^H .\ } $$

The polar decomposition is the diffeomorphism $GL_6\cong U(6)\times\operatorname{Herm}_6^{>0}$; the quotient is the positive cone, the symmetric space of the Cartan involution. The exponential form is $A=Ue^H$ with $H$ Hermitian; the Cartan decomposition is $\mathfrak{gl}_6=\mathfrak{u}(6)\oplus\mathrm i\mathfrak{u}(6)$.

Object Data
Polar decomposition $A=UP$, $U\in U(6)$, $P\ge0$
Positive factor $P=(A^*A)^{1/2}$, unique
$*$-Unitary factor unique if $A$ invertible
Singular freedom $U(6-k)$, dimension $(6-k)^2$
SVD $A=U\Sigma V^*$
Quotient $GL_6/U(6)\cong\operatorname{Herm}_6^{>0}$
Dimension $36$
Exponential form $A=Ue^H$, $H$ Hermitian

Summary of Notation

Symbol Meaning
$P=(A^*A)^{1/2}$ the positive factor
$U$ the $*$-unitary polar factor
$\sigma_1,\dots,\sigma_6$ the singular values
$\operatorname{Herm}_6^{>0}$ the positive definite Hermitian cone
$GL_6/U(6)$ the symmetric space of the polar decomposition
$H=\log P$ the Hermitian exponent

Further Reading