The Polar Decomposition of $M_5(\mathbb{C})$ in Exponential Coordinates
Introduction
The polar decomposition $A=UP$ of The Polar Decomposition of $M_5(\mathbb{C})$ has an exponential form:
$$ A=e^{iH}e^{S},\qquad H=H^*,\quad S=S^*, $$
where $e^{iH}$ is $*$-unitary and $e^S$ is positive definite. Indeed, the $*$-unitary part is $U=e^{iH}$ by the spectral theorem for $*$-unitary matrices, and the positive part is $P=e^S$ by the spectral theorem for positive definite matrices. The polar decomposition in exponential coordinates is the decomposition
$$ GL_5(\mathbb{C})=U(5)\cdot\exp(i\mathfrak{u}(5)), \qquad U(5)=\exp(\mathfrak{u}(5)), $$
written in exponential coordinates.
The decomposition is not unique in $H$: the $*$-unitary logarithm is defined modulo $2\pi$, because $e^{2\pi iI}=I$. The positive part $S$ is unique; the $*$-unitary part $e^{iH}$ is unique, but its logarithm $H$ is not. At the points where $e^{iH}$ has distinct eigenvalues, the exponential coordinates are coordinates on $GL_5$ modulo the lattice of the $*$-unitary periods; at a repeated eigenvalue the Hermitian logarithms need not commute with $H$, and the lattice description is only local.
The article develops the exponential form of the polar decomposition, the Cartan decomposition $G=K\exp(\mathfrak{p})$, the $KAK$ decomposition (the SVD in group form), and the Iwasawa decomposition $G=KAN$. It proves the decompositions, states the uniqueness and the Weyl-group ambiguity, and describes the geodesic polar coordinates on the symmetric space $GL_5/U(5)$. It assumes the polar decomposition, the exponential and the Lie group of the preceding articles.
The Exponential Form of the Polar Decomposition
The Right and Left Forms
Theorem. Every invertible $A\in GL_5(\mathbb{C})$ can be written as
$$ A=e^{iH}e^{S}, $$
with $H,S$ Hermitian. The positive part $e^S=|A|$ is unique, and the $*$-unitary part $e^{iH}=A|A|^{-1}$ is unique.
Proof. The polar decomposition gives $A=UP$ with $U$ $*$-unitary and $P$ positive definite. By the spectral theorem for $*$-unitary matrices, $U=V\operatorname{diag}(e^{i\theta_1},\dots,e^{i\theta_5})V^*=e^{iH}$ with $H=V\operatorname{diag}(\theta_1,\dots,\theta_5)V^*$ Hermitian. By the spectral theorem for positive definite matrices, $P=W\operatorname{diag}(e^{s_1},\dots,e^{s_5})W^*=e^S$ with $S=W\operatorname{diag}(s_1,\dots,s_5)W^*$ Hermitian, $s_k=\log\lambda_k$. Hence $A=e^{iH}e^S$. The uniqueness of $P$ and $U$ gives the uniqueness of $S$ and $e^{iH}$.
Corollary. The left form is
$$ A=e^{S'}e^{iH'},\qquad e^{S'}=|A^*|=(AA^*)^{1/2},\quad e^{iH'}=|A^*|^{-1}A . $$
The left and right positive parts are related by $S'=U S U^*$.
Proof. Apply the theorem to $A^*$ and take adjoints; or use the left polar decomposition $A=P'U$ with $P'=(AA^*)^{1/2}$; the relation follows from $AA^*=UP^2U^*$.
Corollary. The writing is the polar decomposition in exponential coordinates; the positive part is the exponential of a Hermitian matrix, the $*$-unitary part the exponential of $i$ times a Hermitian matrix.
Proof. The spectral theorem for $P$ and $U$; the exponentials are the spectral functions.
The Uniqueness
Theorem. The Hermitian part $S$ is unique. If $e^{iH}$ has distinct eigenvalues, the Hermitian part $H$ is unique modulo the lattice
$$ H\sim H+2\pi D,\qquad D=\operatorname{diag}(n_1,\dots,n_5),\quad n_k\in\mathbb{Z}, $$
in a basis diagonalizing $H$. The unitary part $e^{iH}$ is always unique.
Proof. $e^{2S}=A^*A$ determines $S$ uniquely because the positive definite square root is unique and the logarithm of a positive definite matrix is unique. The $*$-unitary part $e^{iH}=A|A|^{-1}$ is unique. If $e^{iH}$ has distinct eigenvalues, every Hermitian logarithm $H'$ of $e^{iH}$ commutes with $e^{iH}$, hence with $H$, because the commutant of a matrix with distinct eigenvalues is the commutative algebra of polynomials in it; then $e^{i(H'-H)}=I$ with $H'-H$ Hermitian, so $H'-H=2\pi D$ with $D$ diagonal and integral in the common eigenbasis. Without the distinctness, a Hermitian logarithm need not commute with $H$: for $e^{iH}=I$ and any Hermitian projection $P$, the matrix $H'=2\pi P$ is another Hermitian logarithm.
Corollary. The map
$$ i\mathfrak{u}(5)\times i\mathfrak{u}(5)\longrightarrow GL_5,\qquad (H,S)\longmapsto e^{iH}e^S $$
is surjective and is a local diffeomorphism; it is not injective, and at a point where $e^{iH}$ has distinct eigenvalues its fibre is the lattice translate described above.
Proof. Surjectivity is the theorem; the derivative at $(0,0)$ is $(H,S)\mapsto iH+S$, which is an isomorphism because $iH$ is anti-Hermitian, $S$ is Hermitian, and the two summands meet only at $0$. The map is the composition of the exponential on Hermitian matrices, the exponential on anti-Hermitian matrices and the group product, so it is a local diffeomorphism everywhere. The non-injectivity is $e^{2\pi iI}=I$.
Remark. The polar decomposition in exponential coordinates is unique in the group, not in the Lie algebra. The group elements $e^{iH}$ and $e^S$ are unique; the Lie algebra element $H$ is not, because the exponential is not injective on the $*$-unitary directions. The positive part $S$ is unique because the exponential is injective on the Hermitian matrices. The lattice ambiguity is the price of the exponential coordinates.
The Lattice Ambiguity
Proposition. Assume $e^{iH}$ has distinct eigenvalues. The fibre of the exponential coordinate map over $A=e^{iH}e^S$ is
$$ \{(H+2\pi D,\ S):D\text{ diagonal with integer entries in a basis diagonalizing }H\}. $$
Proof. $e^{i(H+2\pi D)}=e^{iH}$ for integer diagonal $D$; conversely, if $e^{iH'}=e^{iH}$ and $e^{iH}$ has distinct eigenvalues, then $H'$ and $H$ commute and differ by $2\pi$ times an integer diagonal in the common eigenbasis, by the theorem above.
Corollary. At the points where $e^{iH}$ has distinct eigenvalues, the exponential coordinates are coordinates on $GL_5$ modulo the lattice $\mathbb{Z}^5$ of the $*$-unitary periods; the positive part has no lattice.
Proof. The lattice is the one displayed; the positive part is unique and has no ambiguity.
The Cartan Decomposition
The Lie Algebra Decomposition
Theorem. The Lie algebra has the Cartan decomposition
$$ \mathfrak{gl}_5=\mathfrak{u}(5)\oplus\mathfrak{p},\qquad \mathfrak{p}=i\mathfrak{u}(5), $$
where $\mathfrak{p}$ is the space of Hermitian matrices. The decomposition is the $+1$ and $-1$ eigenspaces of the Cartan involution $\theta(X)=-X^*$.
Proof. Every $X$ is $X=\frac12(X-X^*)+\frac12(X+X^*)$, anti-Hermitian plus Hermitian; the intersection is zero. The involution $\theta(X)=-X^*$ has $+1$ eigenspace $\mathfrak{u}(5)$ and $-1$ eigenspace $i\mathfrak{u}(5)$; the two are orthogonal for the trace form.
Corollary. $[\mathfrak{u},\mathfrak{u}]\subseteq\mathfrak{u}$, $[\mathfrak{u},\mathfrak{p}]\subseteq\mathfrak{p}$, $[\mathfrak{p},\mathfrak{p}]\subseteq\mathfrak{u}$; the decomposition is a symmetric Lie algebra.
Proof. The bracket of anti-Hermitian matrices is anti-Hermitian; the bracket of an anti-Hermitian and a Hermitian is Hermitian; the bracket of two Hermitian is anti-Hermitian.
The Global Decomposition
Theorem (Cartan decomposition). Every $g\in GL_5(\mathbb{C})$ is uniquely
$$ g=Ue^S,\qquad U\in U(5),\quad S\in i\mathfrak{u}(5), $$
with $U$ $*$-unitary and $S$ Hermitian. This is the polar decomposition.
Proof. The polar decomposition $g=UP$ with $P$ positive definite; $P=e^S$ with $S$ Hermitian; the uniqueness is the polar uniqueness.
Corollary. The map $U(5)\times i\mathfrak{u}(5)\to GL_5$, $(U,S)\mapsto Ue^S$, is a diffeomorphism.
Proof. It is the polar diffeomorphism composed with the exponential diffeomorphism $i\mathfrak{u}(5)\to\mathcal{P}$.
The Symmetric Space
Theorem. The quotient $GL_5(\mathbb{C})/U(5)$ is diffeomorphic to $i\mathfrak{u}(5)\cong\mathbb{R}^{25}$; it is a symmetric space of noncompact type.
Proof. The quotient is the positive cone $\mathcal{P}$; the exponential of Hermitian matrices is a diffeomorphism onto $\mathcal{P}$; the symmetric space structure is the Cartan decomposition.
Corollary. The polar decomposition in exponential coordinates is the global form of the Cartan decomposition; the $*$-unitary group is the compact part and the Hermitian exponential the noncompact part.
Proof. The global Cartan decomposition is the polar decomposition; the exponential coordinates are the Hermitian part.
The $KAK$ Decomposition
The SVD in Group Form
Theorem ($KAK$). Every $g\in GL_5(\mathbb{C})$ can be written as
$$ g=K_1AK_2,\qquad K_1,K_2\in U(5),\quad A=\operatorname{diag}(a_1,\dots,a_5),\quad a_k>0 . $$
The diagonal entries $a_k$ are the singular values of $g$, and they are unique up to permutation; if ordered $a_1\ge\cdots\ge a_5$, they are unique.
Proof. The singular value decomposition $g=V\Sigma W^*$ with $V,W$ $*$-unitary and $\Sigma$ positive diagonal gives $g=K_1AK_2$ with $K_1=V$, $A=\Sigma$, $K_2=W^*$. The singular values are unique; the ordering fixes the diagonal.
Corollary. The $KAK$ decomposition is the SVD in group form; the $A$ part is the exponential of the real diagonal Lie algebra, $A=e^D$ with $D=\operatorname{diag}(\log a_k)$.
Proof. The positive diagonal matrix is the exponential of the real diagonal matrix of the logarithms.
The Weyl Group
Proposition. The diagonal part $A$ is unique up to the action of the Weyl group $S_5$, which permutes the diagonal entries. The $K$-parts are not unique: they are determined only up to the centralizer of $A$ in $U(5)$, which contains the diagonal torus of phases and is larger when singular values coincide.
Proof. The SVD gives $g=V\Sigma W^*$. If $g=V'\Sigma W'^*$ with the same ordered $\Sigma$ and distinct $\sigma_k$, then $V'=VD$ and $W'=WD$ for a diagonal unitary $D$; if some singular values coincide, $D$ may mix within the corresponding eigenspaces. A permutation of the diagonal entries is compensated by permuting the columns of $V$ and $W$, which is the Weyl group. At $g=I$, every $K_1$ with $K_2=K_1^{-1}$ gives $K_1IK_2=I$, so the $K$-parts are arbitrary there.
Corollary. The $KAK$ decomposition is unique up to the Weyl group and the centralizer of $A$; ordering the diagonal entries fixes $A$, not $K_1$ and $K_2$.
Proof. The ordering fixes the representative of the Weyl orbit; the centralizer ambiguity remains.
The Geodesic Polar Coordinates
Theorem. On the symmetric space $GL_5/U(5)\cong\mathcal{P}$, the $KAK$ decomposition gives the geodesic polar coordinates: every $P\in\mathcal{P}$ is $P=K e^D K^*$ with $K$ $*$-unitary and $D$ real diagonal; the geodesic from $I$ to $P$ is $K e^{tD}K^*$, and the distance is $\langle D\rangle_*^{1/2}$.
Proof. The spectral theorem for $P$ gives $P=K e^D K^*$ with $D$ the diagonal of the logarithms; the geodesic $t\mapsto K e^{tD}K^*$ starts at $I$ and ends at $P$; its length is $\langle D\rangle_*^{1/2}$ for the trace metric.
Corollary. The singular values of $g$ are the exponentials of the diagonal entries of $D$; the distance from $I$ to $P$ is the Euclidean norm of the logarithms of the singular values.
Proof. $a_k=e^{d_k}$; the distance is $\sqrt{\sum d_k^2}$.
Remark. The $KAK$ decomposition is the polar decomposition of the group, read as geodesic polar coordinates. The $K$ parts are the angular coordinates, the $A$ part the radial coordinate; the Weyl group is the ambiguity of the angles. The polar decomposition in exponential coordinates is the same picture with the $*$-unitary part written as $e^{iH}$.
The Iwasawa Decomposition
The $KAN$ Decomposition
Theorem (Iwasawa). Every $g\in GL_5(\mathbb{C})$ is uniquely
$$ g=KAN, $$
where $K=U(5)$, $A$ is the group of positive diagonal matrices, and $N$ is the unipotent group of upper triangular matrices with diagonal entries $1$.
Proof. Gram–Schmidt on the columns of $g$: $g=QR$ with $Q$ $*$-unitary and $R$ upper triangular with positive diagonal. Write $R=AN$ with $A$ the positive diagonal of $R$ and $N=A^{-1}R$ unipotent upper triangular. Then $g=QAN$; the uniqueness is the uniqueness of the Gram–Schmidt decomposition.
Corollary. The Iwasawa decomposition is a diffeomorphism $U(5)\times A\times N\to GL_5$; the real dimensions are $25+5+20=50$, which is the real dimension of $GL_5$.
Proof. The Gram–Schmidt map is a diffeomorphism. The dimensions are $25$ for $U(5)$, $5$ for the positive diagonal $A$, and $20$ for the unipotent $N$ (complex dimension $10$); the sum is $50$.
Corollary. The Iwasawa decomposition is not the polar decomposition: the polar positive part $P$ is positive definite, not diagonal, and the $*$-unitary part is $U$, while the Iwasawa $A$ is diagonal and $N$ is unipotent. The two decompositions are related by the Gram–Schmidt process.
Proof. The polar decomposition uses the positive square root of $g^*g$; the Iwasawa uses the QR decomposition. The $A$ part is diagonal, the $N$ part unipotent; the polar $P$ is general positive definite.
The Gram–Schmidt Proof
Proposition. The Gram–Schmidt decomposition of the columns $g_1,\dots,g_5$ gives
$$ g_k=\sum_{l\le k}r_{lk}q_l,\qquad r_{kk}>0, $$
with $q_1,\dots,q_5$ orthonormal; the matrix $Q=(q_1,\dots,q_5)$ is $*$-unitary and $R=(r_{lk})$ is upper triangular with positive diagonal.
Proof. The Gram–Schmidt process orthogonalizes the columns; the coefficients $r_{lk}$ are the projections; the diagonal entries are the norms of the orthogonalized columns, positive.
Corollary. The Iwasawa decomposition is the matrix form of the Gram–Schmidt process; the $A$ part is the diagonal of the norms, the $N$ part the unipotent upper triangular factor.
Proof. $R=AN$; $A$ is the diagonal of $R$; $N$ is unipotent upper triangular.
The Relation to the Polar Decomposition
Proposition. The polar decomposition and the Iwasawa decomposition are related by
$$ g=UP=QAN, $$
where $U,Q$ are $*$-unitary, $P$ is positive definite, $A$ is positive diagonal and $N$ is unipotent upper triangular. The two decompositions are distinct: $P=AN$ would require $AN$ to be positive definite, but $AN$ is Hermitian only when $N=I$. In general $U\ne Q$ and $P\ne AN$.
Proof. $AN$ is upper triangular with positive diagonal; it is Hermitian iff it is diagonal, i.e., $N=I$. The polar decomposition uses the positive square root of $g^*g$; the Iwasawa decomposition uses the Gram–Schmidt factorization. They are different coordinate systems.
Corollary. The Iwasawa decomposition is not the polar decomposition; the positive part of the polar decomposition is not the $A$ part of the Iwasawa decomposition. The $A$ part is diagonal, the $P$ part is positive definite.
Proof. The diagonal $A$ is a special positive definite matrix; the general $P$ is not diagonal.
Remark. The three decompositions are different coordinate systems on $GL_5$. The polar decomposition $G=K\exp(\mathfrak{p})$ uses the positive definite cone; the $KAK$ decomposition uses the maximal torus of the cone and the Weyl group; the Iwasawa decomposition $G=KAN$ uses the Borel subgroup. The polar decomposition in exponential coordinates is the polar one written with exponentials; the $KAK$ and $KAN$ are the geodesic and the Gram–Schmidt coordinates.
The Exponential Coordinates
The Map
Theorem. The map
$$ \Phi:i\mathfrak{u}(5)\times i\mathfrak{u}(5)\longrightarrow GL_5,\qquad (H,S)\longmapsto e^{iH}e^S $$
is surjective, smooth, and a local diffeomorphism at every point where the eigenvalues of $H$ have no differences in $2\pi\mathbb{Z}$; the fibre is the lattice of the $*$-unitary periods.
Proof. Surjectivity is the polar decomposition in exponential coordinates; smoothness is the exponential. The derivative at $(H,S)$ is the composition of the derivatives of the exponential and the product. The exponential $S\mapsto e^S$ is a diffeomorphism on the Hermitian matrices; the exponential $H\mapsto e^{iH}$ is a local diffeomorphism exactly when the eigenvalues of $\operatorname{ad}_{iH}$ avoid $2\pi i\mathbb{Z}\setminus\{0\}$, i.e., when no two eigenvalues of $H$ differ by a multiple of $2\pi$. The inverse function theorem then gives the local diffeomorphism; the global ambiguity is the lattice.
Corollary. The exponential coordinates are the polar coordinates of the group: the $*$-unitary directions are angular, the Hermitian directions radial.
Proof. The positive part $e^S$ is the radial coordinate; the $*$-unitary part $e^{iH}$ is the angular coordinate; the lattice is the periodicity of the angles.
The Jacobian
Proposition. The derivative of $\Phi$ at $(H,S)$ in the directions $(\dot H,\dot S)$ is
$$ \left(D\exp_{iH}(i\dot H)\right)e^S+e^{iH}\left(D\exp_S(\dot S)\right), $$
with the exponential derivatives of Analysis on $M_5$. At $(0,0)$ the derivative is $(\dot H,\dot S)\mapsto i\dot H+\dot S$, which is an isomorphism from the Hermitian product to $\mathfrak{gl}_5$.
Proof. The product rule and the chain rule; the derivative of $e^{iH}$ in the direction $\dot H$ is $D\exp_{iH}(i\dot H)$; the derivative of $e^S$ is $D\exp_S(\dot S)$; at $(0,0)$, $D\exp_0=I$.
Corollary. The Jacobian is singular exactly when the $*$-unitary exponential fails to be a local diffeomorphism, i.e., when two eigenvalues of $H$ differ by a multiple of $2\pi$. The Hermitian exponential is always a local diffeomorphism.
Proof. The derivative is the sum of the invertible Hermitian exponential derivative and the $*$-unitary exponential derivative; the latter is singular exactly at the period differences; the former is invertible everywhere.
The Geometry
Theorem. The polar decomposition in exponential coordinates gives the geodesic polar coordinates on the symmetric space $GL_5/U(5)$: the geodesic from $I$ to $e^S$ is $e^{tS}$, and the $*$-unitary part is the angular coordinate. The distance is $\langle S\rangle_*^{1/2}$.
Proof. The positive cone is the symmetric space; the geodesic from $I$ to $e^S$ is $e^{tS}$; the distance is the Frobenius norm of $S$.
Corollary. The $KAK$ decomposition refines the exponential coordinates by diagonalizing $S$: $S=KDK^*$; the radial coordinate is the diagonal $D$, and the angular coordinates are the two $*$-unitary factors.
Proof. The spectral theorem for $S$; the $KAK$ decomposition of $e^S$.
Remark. The polar decomposition in exponential coordinates is the group-level analogue of the polar form of a complex number. A complex number is $\varrho e^{i\theta}$; a matrix is $e^Se^{iH}$, with $S$ the logarithm of the modulus and $H$ the phase. The noncommutativity makes the order matter: the right and left forms differ, and the $*$-unitary logarithm is periodic. The $KAK$ and Iwasawa decompositions are the finer coordinate systems adapted to the torus and the Borel.
Summary
Every invertible matrix has the polar decomposition in exponential coordinates
$$ A=e^{iH}e^S,\qquad H,S\text{ Hermitian}, $$
with $e^S=|A|$ unique and $e^{iH}=A|A|^{-1}$ unique; when $e^{iH}$ has distinct eigenvalues, $H$ is unique modulo the lattice of $*$-unitary periods. The group decompositions are
$$ GL_5=U(5)\exp(i\mathfrak{u}(5))=KAK=KAN, $$
the Cartan (polar), the $KAK$ (SVD) and the Iwasawa (Gram–Schmidt) decompositions.
$$ \boxed{\ A=e^{iH}e^S;\quad GL_5=U(5)\exp(i\mathfrak{u}(5));\quad g=K_1AK_2;\quad g=KAN.\ } $$
The exponential coordinates are the polar coordinates of the group: the positive part is radial, the $*$-unitary part angular, the lattice is the periodicity. The $KAK$ decomposition gives the geodesic polar coordinates on the symmetric space $GL_5/U(5)$; the $KAN$ decomposition gives the Gram–Schmidt coordinates. The three decompositions are different coordinate systems, related by the polar decomposition, the SVD and the QR decomposition.
| Decomposition | Form | Factors |
|---|---|---|
| Polar | $A=UP$ | $U\in U(5)$, $P>0$ |
| Exponential coordinates | $A=e^{iH}e^S$ | $H,S$ Hermitian |
| Cartan | $G=K\exp(\mathfrak{p})$ | $K=U(5)$, $\mathfrak{p}=i\mathfrak{u}(5)$ |
| $KAK$ | $g=K_1AK_2$ | $A$ positive diagonal |
| Iwasawa | $g=KAN$ | $A$ positive diagonal, $N$ unipotent |
Summary of Notation
| Symbol | Meaning |
|---|---|
| $e^{iH}$ | The $*$-unitary part |
| $e^S$ | The positive part |
| $\mathfrak{p}=i\mathfrak{u}(5)$ | The Hermitian tangent directions |
| $K=U(5)$ | The $*$-unitary group |
| $A$ | The positive diagonal group |
| $N$ | The unipotent upper triangular group |
| $KAK$ | The SVD in group form |
| $KAN$ | The Iwasawa decomposition |
Further Reading
- Sigurdur Helgason, Differential Geometry, Lie Groups, and Symmetric Spaces (American Mathematical Society, 2001), for the Cartan, $KAK$ and Iwasawa decompositions and the symmetric spaces.
- Brian C. Hall, Lie Groups, Lie Algebras, and Representations (Springer, second edition, 2015), for the Cartan decomposition and the exponential map.
- Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge University Press, second edition, 2013), for the polar decomposition, the SVD and the QR decomposition.
- The Polar Decomposition of $M_5(\mathbb{C})$, for the polar decomposition and the positive cone.
- The Exponential and the Lie Group of $M_5(\mathbb{C})$, for the exponential, the logarithm and the Lie group.
- The Adjoint Action, the Orbits and the Flag Variety of $M_5(\mathbb{C})$, for the symmetric space and the group action.
- The Geometry of the Idempotents of $M_5(\mathbb{C})$, for the geometry of the idempotent varieties.