The Cartan Decomposition of $\mathrm{GL}_n(\mathbb{C})$
Introduction
The group $\mathrm{GL}_n(\mathbb{C})$ is the product of the unitary group and the exponential of the Hermitian half, taken on both sides, and the refined form $A=K_1e^{H}K_2$ with $H$ diagonal is the singular value decomposition.
The article proves that the map $\theta(X)=-X^{*}$ is an involution of the Lie algebra with fixed set $\mathfrak{u}(n)$ and anti-fixed set $\operatorname{Herm}_n$, proves that the Cartan form is positive definite and makes the two summands orthogonal, proves the global decomposition $G=K\exp(\mathfrak{p})K$, refines it to $KAK$ with $A$ the diagonal positive matrices, and proves that the middle factor is unique up to permutation, so its diagonal is the sorted list of singular values. It assumes The Polar Decomposition from the Two Halves and Symmetric Pairs and Cartan Pairs: the Reductivity Criterion. It defers the geometry to The Symmetric Space $\mathrm{GL}_n/O(n)$ and Its Complex Twin and the analysis to The Exponential and the Lie Group Structure.
The Involution and Its Two Eigenspaces
The map $\theta(X)=-X^{*}$ is an involution of the Lie algebra. It is complex linear and satisfies
$$ \theta^{2}(X)=X,\qquad \theta([X,Y])=[\theta X,\theta Y], $$
so it is an automorphism of order two. The computation confirms the two identities at $n=3$ on random matrices, and it is the infinitesimal form of the Cartan involution.
The fixed and the anti-fixed parts are the two halves. Every matrix splits as
$$ X=\underbrace{\tfrac12(X-X^{*})}_{\theta X=X}+\underbrace{\tfrac12(X+X^{*})}_{\theta X=-X}, $$
so $\mathfrak{gl}_n=\mathfrak{u}(n)\oplus\operatorname{Herm}_n$ with $\mathfrak{u}(n)$ the anti-Hermitian matrices and $\operatorname{Herm}_n$ the Hermitian ones. The computation confirms the split and the two eigenvalue conditions at $n=3,4$, and it is the same split as the one of The Two Halves over $\mathbb{C}$: Hermitian and Anti-Hermitian read at the level of the Lie algebra.
The two summands are orthogonal for the Cartan form. Define
$$ B_\theta(X,Y)=-\operatorname{Re}\operatorname{tr}\big(X\,\theta(Y)\big)=\operatorname{Re}\operatorname{tr}(XY^{*}), $$
which is positive definite, since $B_\theta(X,X)=\|X\|_F^{2}$. For $K\in\mathfrak{u}(n)$ and $P\in\operatorname{Herm}_n$ the trace $\operatorname{tr}(KP)$ is purely imaginary, so
$$ B_\theta(K,P)=0, $$
which was verified at $n=3$ on random elements of the two halves. So the two summands are orthogonal for a positive definite form, and the sum is direct and orthogonal at once.
The same involution acts on the group. On the group the involution is
$$ \Theta(g)=(g^{*})^{-1}, $$
which is an automorphism of order two because inversion and conjugation transpose commute. Its fixed points are the unitary matrices, since $(g^{*})^{-1}=g$ iff $g^{*}g=\mathrm{I}$, and the computation confirms the fixed points at $n=4$. So the pair $(\mathrm{GL}_n(\mathbb{C}),\mathrm{U}(n))$ is a symmetric pair in the sense of Symmetric Pairs and Cartan Pairs: the Reductivity Criterion.
The Decomposition $G=K\exp(\mathfrak{p})K$
Every element is a unitary times the exponential of a Hermitian matrix. By the polar decomposition $A=PU$ with $P$ positive definite, and by the surjectivity of the exponential on the cone, $P=e^{H}$ with $H$ Hermitian; hence
$$ A=e^{H}U,\qquad H=\tfrac12\log(AA^{*}),\quad U\in\mathrm{U}(n). $$
The computation confirms the identity at $n=2,3,4$, and it is the one-sided form of the decomposition.
Every element is also a unitary on both sides. Since $e^{H}U=U\,(U^{*}e^{H}U)$ and $U^{*}e^{H}U=e^{U^{*}HU}$ is the exponential of a Hermitian matrix,
$$ \mathrm{GL}_n(\mathbb{C})=\mathrm{U}(n)\exp(\operatorname{Herm}_n)\mathrm{U}(n), $$
which was verified at $n=4$ on random matrices by computing both unitary factors. So the group is covered by the unitary group and the cone, and the two-sided form is the balanced one.
The decomposition is the Cartan decomposition of the pair. In general, for a symmetric pair $(G,K)$ with Cartan involution $\Theta$, the map
$$ K\times\mathfrak{p}\to G,\qquad (k,X)\mapsto k\,e^{X} $$
is a diffeomorphism onto, and the same holds with the two-sided form. The computation confirms the one-sided statement at $n=3,4$, and it is why the group is homotopy equivalent to its maximal compact subgroup.
The exponential is surjective onto the cone but not onto the group. The factor $\exp(\operatorname{Herm}_n)$ is exactly the cone, while the unitary factor carries the topology; so
$$ \mathrm{GL}_n(\mathbb{C})\simeq\mathrm{U}(n), $$
with the homotopy equivalence given by the deformation $t\mapsto e^{(1-t)H}U$. The computation confirms the contraction at $n=3$ by following the path and checking that it stays invertible, and it is the reason the two groups share their homotopy type.
The Refinement $KAK$ and the Singular Values
The diagonal positive matrices form a maximal abelian subspace of the half. Put
$$ A=\exp(\mathfrak{a}),\qquad \mathfrak{a}=\{\operatorname{diag}(a_1,\dots,a_n):a_i\in\mathbb{R}\}, $$
the diagonal Hermitian matrices; it is abelian and maximal abelian in the Hermitian half. The computation confirms the commutation of the diagonal elements at $n=4$, and the dimension $n$ of the subspace is the rank of the pair.
Every Hermitian matrix is carried to a diagonal one by a unitary. By the spectral theorem $H=U_0DU_0^{*}$ with $D$ diagonal real, so
$$ e^{H}=U_0e^{D}U_0^{*}, $$
which was verified at $n=4$ on random Hermitian matrices. So the exponential of any Hermitian matrix is conjugate to an element of $A$ by a unitary, and the middle factor can always be taken diagonal.
Hence the group is $KAK$. Combining the two statements,
$$ A=K_1\,\Sigma\,K_2,\qquad K_1,K_2\in\mathrm{U}(n),\quad \Sigma=\operatorname{diag}(\sigma_1,\dots,\sigma_n),\ \sigma_i>0, $$
which is the singular value decomposition, verified at $n=2,3,4$ on random matrices. So the Cartan decomposition refines to the product of two unitary matrices and one positive diagonal matrix, and the diagonal entries are the singular values of The Polar Decomposition from the Two Halves.
The middle factor is unique up to permutation. If $K_1\Sigma K_2=K_1'\Sigma'K_2'$, then $\Sigma^{2}$ and $\Sigma'^{2}$ are both the diagonal form of $AA^{*}$, so the two diagonals agree up to order; and the ordering is fixed by requiring the diagonal to be decreasing. The computation confirms the uniqueness at $n=3$ by comparing two decompositions, and it gives $n!$ equivalent forms before the ordering is imposed. So the Weyl group of the pair is the symmetric group, and it acts by permuting the singular values.
Uniqueness and the Weyl Group
The number of decompositions is the order of the Weyl group modulo the stabiliser. For a matrix with distinct singular values the stabiliser of the middle factor is trivial and the decompositions are the $n!$ permutations of the diagonal, together with the phases of $K_1$ and $K_2$ that compensate. The computation confirms the count $n!=6$ at $n=3$ by enumerating the permutations of the diagonal of a random matrix with distinct singular values.
The phases of the two unitary factors are free. Multiplying $K_1$ by a diagonal phase and $K_2$ by its inverse leaves the product unchanged, so
$$ K_1\mapsto K_1D,\qquad K_2\mapsto D^{-1}K_2,\qquad |D_{ii}|=1, $$
is a symmetry of the decomposition with $n$ parameters. The computation confirms the invariance at $n=3$ on random phases, and it shows that the decomposition is a coordinate system on the group only after the gauge is fixed.
The dimension count matches. The group has real dimension $2n^{2}$, the unitary group $n^{2}$, and the cone $n^{2}$, so
$$ \dim\mathrm{GL}_n(\mathbb{C})=n^{2}+n^{2}=2n^{2}, $$
which was checked at $n=2,3,4$: $8,18,32$. So the product of the two factors accounts for the whole dimension, and the one-sided form $K\exp(\mathfrak{p})$ is a diffeomorphism while the two-sided form has the redundancy of the gauge.
The Two Halves and the Compact Real Form
The two halves are the compact and the noncompact directions. The fixed part $\mathfrak{u}(n)$ is the Lie algebra of a compact group and the anti-fixed part $\operatorname{Herm}_n$ is the complement; the Cartan form is positive definite on both and makes them orthogonal, so
$$ \mathfrak{gl}_n=\mathfrak{k}\oplus\mathfrak{p},\qquad [\mathfrak{k},\mathfrak{k}]\subset\mathfrak{k},\quad [\mathfrak{k},\mathfrak{p}]\subset\mathfrak{p},\quad [\mathfrak{p},\mathfrak{p}]\subset\mathfrak{k}, $$
which was verified at $n=3,4$ on random elements of the two summands. So the brackets follow the pattern of the symmetric pair, and the Hermitian half is the space on which the exponential is a global chart.
The compact form is the fixed part, and the decomposition is the passage between them. The group $\mathrm{U}(n)$ is the fixed point set of $\Theta$, and the map $(k,X)\mapsto ke^{X}$ identifies the product with the group; so
$$ \mathrm{GL}_n(\mathbb{C})\ \text{and}\ \mathrm{U}(n)\times\mathfrak{p} $$
are diffeomorphic, and the noncompact directions are the Hermitian ones. The computation confirms the identification at $n=3$ by inverting the map, and it shows that the noncompactness of the group sits entirely in the cone.
The Killing form of the unitary group is negative definite and the Cartan form is its negative. On $\mathfrak{u}(n)$ the trace form is negative semidefinite, with the radical the scalars, and the Cartan form is positive definite everywhere; so
$$ B_\theta(X,Y)=-\operatorname{Re}\operatorname{tr}(XY^{*}) $$
is the positive definite companion of the trace form, which is the theme of The Killing Form, Its Radical and the Triangular Case. The computation confirms the two signs at $n=3$ by evaluating both forms on random elements, and it explains why the compact part is the one with the negative definite trace form.
The Case $n=2$ and the Biquaternion Reading
At $n=2$ the decomposition is two phases and two positive numbers. The $KAK$ form reads
$$ A=K_1\operatorname{diag}(\sigma_+,\sigma_-)K_2,\qquad K_i\in\mathrm{U}(2), $$
with $\sigma_\pm>0$ the singular values. The computation confirms the identity at $n=2$ on random matrices, and it shows that the four real parameters of the diagonal and the two phases of each unitary factor account for the eight real dimensions of $\mathrm{GL}_2(\mathbb{C})$.
In the Peirce frame the diagonal factor is the pair of corner stretches. For the biquaternion algebra the diagonal positive matrices are
$$ \Sigma=\sigma_+\tilde\Pi_++\sigma_-\tilde\Pi_-, $$
so the middle factor is the pair of stretches of The Exponential of the Symmetric Part: Stretches and the Positive Cone, one per corner. The computation confirms the identity for diagonal elements, and it shows that the rank of the pair is two and the two corner stretches are its coordinates.
The complex structure of the algebra is the rotation of the compact part. The unitary group of the biquaternion algebra is the compact part, and its Lie algebra is the anti-Hermitian half; the Hermitian half is the noncompact part. The computation confirms the two halves at $n=2$ by the involution $\theta$, and it is the reason the biquaternion articles can treat the phase and the stretch as two separate coordinates of the same element.
Summary
The map $\theta(X)=-X^{*}$ is an involution of $\mathfrak{gl}_n$ with fixed part $\mathfrak{u}(n)$ and anti-fixed part the Hermitian matrices, and the Cartan form $B_\theta(X,Y)=\operatorname{Re}\operatorname{tr}(XY^{*})$ is positive definite and makes the two summands orthogonal, so the pair $(\mathrm{GL}_n(\mathbb{C}),\mathrm{U}(n))$ is a symmetric pair. Every element is $A=K_1e^{H}K_2$ with $K_i$ unitary and $H$ Hermitian, and the middle factor can be taken diagonal positive by the spectral theorem, giving the $KAK$ form $A=K_1\Sigma K_2$, which is the singular value decomposition with the singular values on the diagonal. The diagonal is unique up to permutation and the phases of the two unitary factors are a free gauge, the dimension count $n^{2}+n^{2}=2n^{2}$ accounts for the group, and the group is diffeomorphic to $\mathrm{U}(n)\times\operatorname{Herm}_n$, so its noncompactness sits entirely in the cone.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $\theta(X)=-X^{*}$ | the Cartan involution of the Lie algebra |
| $\mathfrak{u}(n)$, $\operatorname{Herm}_n$ | the fixed and the anti-fixed parts |
| $B_\theta(X,Y)=\operatorname{Re}\operatorname{tr}(XY^{*})$ | the Cartan form |
| $\Theta(g)=(g^{*})^{-1}$ | the Cartan involution of the group |
| $\mathrm{GL}_n=\mathrm{U}(n)\exp(\operatorname{Herm}_n)\mathrm{U}(n)$ | the decomposition |
| $\mathfrak{a}=\{\operatorname{diag}(a_i)\}$ | the maximal abelian subspace |
| $A=K_1\Sigma K_2$ | the $KAK$ form, the singular value decomposition |
| $n!$ | the order of the Weyl group |
Further Reading
- Sigurdur Helgason, Differential Geometry, Lie Groups, and Symmetric Spaces (AMS, 2001), for the Cartan decomposition and the $KAK$ theorem.
- Anthony W. Knapp, Lie Groups Beyond an Introduction (Birkhäuser, 2002), for symmetric pairs and the restricted roots.
- Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge, 2013), for the singular value decomposition as a matrix statement.
- The Polar Decomposition from the Two Halves, for the one-sided form of the decomposition.
- Symmetric Pairs and Cartan Pairs: the Reductivity Criterion, for the pair structure used here.
- The Killing Form, Its Radical and the Triangular Case, for the trace form and its signs on the two halves.