The Cartan Decomposition of $\mathrm{GL}_n(\mathbb{R})$
Introduction
The general linear group is the product of the orthogonal group and the exponentials of the symmetric matrices, and the product is a diffeomorphism. The statement is the polar decomposition read as a theorem about a group and a subspace: the group factor is compact, the subspace factor is a linear space, and the two are the eigenspaces of one involution.
The article states the decomposition, proves the three bracket relations, identifies the involution $\theta(X)=-X^{\mathsf T}$ and its group form, proves that the product map is a diffeomorphism, discusses the maximal compact subgroup and its two components, computes the restricted roots and the Weyl group, and identifies the quotient with the cone of positive definite matrices. It assumes The Polar Decomposition from the Two Halves and the Sign of the Determinant and The Exponential of the Symmetric Part: Stretches and the Positive Cone.
The Statement
Theorem (Cartan decomposition). Put $K=\mathrm{O}(n)$ and $\mathfrak{p}=\operatorname{Sym}_n$. Every $g\in\mathrm{GL}_n(\mathbb{R})$ is uniquely
$$ g=k\,\mathrm{e}^{X},\qquad k\in K,\ X\in\mathfrak{p}, $$
and the map $K\times\mathfrak{p}\to\mathrm{GL}_n(\mathbb{R})$, $(k,X)\mapsto k\mathrm{e}^{X}$, is a diffeomorphism.
The two factors are of different natures. $K$ is compact and bounded in $M_n(\mathbb{R})$; $\mathrm{e}^{\mathfrak{p}}$ is the cone $\operatorname{Sym}_n^{>0}$, open, convex and unbounded. A compact factor and a linear factor, multiplied once, give the group.
The count is exact. $\dim K=n(n-1)/2$ and $\dim\mathfrak{p}=n(n+1)/2$, and their sum is $n^{2}=\dim\mathrm{GL}_n(\mathbb{R})$.
The Lie Algebra Splits
The Lie algebra is $\mathfrak{gl}_n(\mathbb{R})=M_n(\mathbb{R})$ with the commutator. Every matrix is a symmetric plus an antisymmetric matrix in exactly one way:
$$ A=\underbrace{\tfrac12(A+A^{\mathsf T})}_{\text{symmetric}}+\underbrace{\tfrac12(A-A^{\mathsf T})}_{\text{antisymmetric}} . $$
So, as vector spaces,
$$ \mathfrak{gl}_n(\mathbb{R})=\mathfrak{k}\oplus\mathfrak{p},\qquad \mathfrak{k}=\mathfrak{so}(n)=\operatorname{Skew}_n,\quad \mathfrak{p}=\operatorname{Sym}_n, $$
and the sum is direct because a matrix that is both symmetric and antisymmetric is zero. The dimensions add, $\frac{n(n-1)}{2}+\frac{n(n+1)}{2}=n^{2}$.
$\mathfrak{k}$ is a subalgebra and $\mathfrak{p}$ is not. The antisymmetric matrices are closed under the commutator. The symmetric ones are not: the commutator of two symmetric matrices is antisymmetric whenever it is nonzero, so $\mathfrak{p}$ is only a subspace.
The Three Brackets
Theorem. The three relations hold:
$$ [\mathfrak{k},\mathfrak{k}]\subseteq\mathfrak{k},\qquad [\mathfrak{k},\mathfrak{p}]\subseteq\mathfrak{p},\qquad [\mathfrak{p},\mathfrak{p}]\subseteq\mathfrak{k}. $$
Proof. Let $X,Y$ be antisymmetric and $S,T$ symmetric. Using $(UV)^{\mathsf T}=V^{\mathsf T}U^{\mathsf T}$ and $X^{\mathsf T}=-X$, $S^{\mathsf T}=S$,
$$ [X,Y]^{\mathsf T}=Y^{\mathsf T}X^{\mathsf T}-X^{\mathsf T}Y^{\mathsf T}=YX-XY=-[X,Y], $$
so $[X,Y]$ is anti-Hermitian. Next,
$$ [X,S]^{\mathsf T}=S^{\mathsf T}X^{\mathsf T}-X^{\mathsf T}S^{\mathsf T}=-SX+XS=[X,S], $$
so $[X,S]$ is symmetric. Finally,
$$ [S,T]^{\mathsf T}=TS-ST=-[S,T], $$
so $[S,T]$ is anti-Hermitian. Numerically all three defects are $0$ at $n=3$ on random pairs.
The Involution
Definition. Put $\theta(X)=-X^{\mathsf T}$ for $X\in\mathfrak{gl}_n(\mathbb{R})$.
$\theta$ is an involutive automorphism. It is linear, $\theta^{2}=\mathrm{id}$, and
$$ \theta([X,Y])=-[X,Y]^{\mathsf T}=-[Y^{\mathsf T},X^{\mathsf T}]=[X^{\mathsf T},Y^{\mathsf T}]=[\theta X,\theta Y], $$
using $[U,V]^{\mathsf T}=[V^{\mathsf T},U^{\mathsf T}]$; numerically the defect is $0$ at $n=3$.
The eigenspaces are the two summands. A matrix with $\theta(X)=X$ satisfies $-X^{\mathsf T}=X$ and is antisymmetric; a matrix with $\theta(X)=-X$ satisfies $X^{\mathsf T}=X$ and is symmetric. Hence
$$ \mathfrak{k}=\{X:\theta(X)=X\},\qquad \mathfrak{p}=\{X:\theta(X)=-X\}. $$
Theorem (the group involution). The map $\Theta(g)=(g^{\mathsf T})^{-1}$ is an involutive automorphism of $\mathrm{GL}_n(\mathbb{R})$ whose fixed set is $K=\mathrm{O}(n)$, and its derivative at the identity is $\theta$.
Proof. $\Theta(gh)=(h^{\mathsf T}g^{\mathsf T})^{-1}=(g^{\mathsf T})^{-1}(h^{\mathsf T})^{-1}=\Theta(g)\Theta(h)$ and $\Theta^{2}(g)=g$. A fixed point satisfies $(g^{\mathsf T})^{-1}=g$, that is $g^{\mathsf T}g=\mathrm{I}$, that is $g\in\mathrm{O}(n)$. The derivative at $\mathrm{I}$ in the direction $X$ is $-X^{\mathsf T}$.
The two eigenspaces are what make the decomposition unique. $\mathfrak{k}$ and $\mathfrak{p}$ are subspaces of a linear space, so $\exp$ is injective on $\mathfrak{p}$, and the involution separates the two factors of the polar decomposition.
The Diffeomorphism
Theorem. The product map $K\times\mathfrak{p}\to\mathrm{GL}_n(\mathbb{R})$ is a diffeomorphism.
Proof. It is the polar decomposition. For $g$ put $P=(g^{\mathsf T}g)^{1/2}$, positive definite, and $k=gP^{-1}$, orthogonal. Then $k\in K$, $X=\log P\in\mathfrak{p}$, and $g=k\mathrm{e}^{X}$. Uniqueness is the uniqueness of the polar decomposition together with the bijection $\exp:\operatorname{Sym}_n\to\operatorname{Sym}_n^{>0}$. Both maps $g\mapsto(k,X)$ and $(k,X)\mapsto k\mathrm{e}^{X}$ are smooth.
The inverse is explicit. From $g=k\mathrm{e}^{X}$ one reads $\mathrm{e}^{2X}=g^{\mathsf T}g$, so $X=\tfrac12\log(g^{\mathsf T}g)$.
The two components. For $\det g>0$ the factor $k$ lies in $\mathrm{SO}(n)$, and $\mathrm{GL}_n(\mathbb{R})^{+}$ is diffeomorphic to $\mathrm{SO}(n)\times\operatorname{Sym}_n$. For $\det g<0$ it lies in the other component of $\mathrm{O}(n)$. The choice $K=\mathrm{O}(n)$ covers both components at the price of a disconnected $K$.
The Maximal Compact Subgroup
Theorem. $\mathrm{O}(n)$ is a maximal compact subgroup of $\mathrm{GL}_n(\mathbb{R})$, and every compact subgroup is conjugate to a subgroup of it.
Proof. Let $H$ be compact. Averaging the standard inner product over $H$ gives the invariant positive definite form $\langle x,y\rangle_H=\int_H\langle hx,hy\rangle\,\mathrm{d}h$. In an orthonormal basis of this form every $h\in H$ is orthogonal, so $H\subseteq\mathrm{O}(n)$ after a change of basis. If in addition $\mathrm{O}(n)\subseteq H$, the form is invariant under $\mathrm{O}(n)$, hence a positive multiple of the standard one, and the basis is orthonormal for the standard form too; then $H\subseteq\mathrm{O}(n)$ in the original basis, so $H=\mathrm{O}(n)$.
The two components. $\mathrm{O}(n)$ has two, $\mathrm{SO}(n)$ and $\mathrm{O}^{-}(n)=\{Q:\det Q=-1\}$, and $\mathrm{O}^{-}(n)=J\,\mathrm{SO}(n)$ for any $J\in\mathrm{O}^{-}(n)$. Only the first is a subgroup, and only the first is the identity component of the ambient group.
The maximal compact subgroup of the identity component is $\mathrm{SO}(n)$. For $\mathrm{GL}_n(\mathbb{R})^{+}$ the compact factor is connected, and the decomposition reads $\mathrm{GL}_n(\mathbb{R})^{+}=\mathrm{SO}(n)\exp(\operatorname{Sym}_n)$.
The Restricted Roots
The maximal abelian subspace of $\mathfrak{p}$ is the diagonal. Commuting symmetric matrices are simultaneously diagonalisable, so
$$ \mathfrak{a}=\{\operatorname{diag}(x_1,\dots,x_n):x_i\in\mathbb{R}\} $$
is a maximal abelian subspace of $\operatorname{Sym}_n$, of dimension $n$.
The joint eigenspaces are the coordinate lines. For $H=\operatorname{diag}(x)$ and the matrix unit $E_{ij}$,
$$ [H,E_{ij}]=(x_i-x_j)E_{ij}, $$
exactly at $n=3$ for all nine pairs. So the restricted roots are the functionals $\alpha_{ij}(x)=x_i-x_j$ with $i\ne j$, each with root space $\mathbb{R}E_{ij}$ of multiplicity one, and the centraliser of $\mathfrak{a}$ is $\mathfrak{a}$ itself.
The root system is that of $A_{n-1}$. The roots are the $n(n-1)$ ordered differences, the positive ones are the $\binom n2$ with $i The Weyl group is $S_n$. It is the normaliser of $\mathfrak{a}$ in $K$ modulo its centraliser, and the permutation matrices are orthogonal and conjugate $\mathfrak{a}$ to itself. So $W=S_n$, of order $n!$, acting by permuting the indices. Theorem. The quotient $\mathrm{GL}_n(\mathbb{R})/K$ is diffeomorphic to $\operatorname{Sym}_n^{>0}$, through $gK\mapsto gg^{\mathsf T}$. Proof. The map is well defined, since $(gk)(gk)^{\mathsf T}=gg^{\mathsf T}$ for $k\in K$. It is surjective, because every $P\in\operatorname{Sym}_n^{>0}$ is $gg^{\mathsf T}$ with $g=P^{1/2}$. It is injective, because $gg^{\mathsf T}=hh^{\mathsf T}$ gives $(h^{-1}g)(h^{-1}g)^{\mathsf T}=\mathrm{I}$, so $h^{-1}g\in K$. Both the map and its inverse are smooth. The dimension is $n(n+1)/2$, and the count closes. The quotient has the dimension of $\mathfrak{p}$, the cone has the dimension of $\operatorname{Sym}_n$, and $$
n^{2}=\frac{n(n-1)}{2}+\frac{n(n+1)}{2}.
$$ The space is a symmetric space. The involution $\Theta$ acts on the quotient, and the geodesic symmetry at the base point is $P\mapsto P^{-1}$ on the cone. The cone with the metric $\langle U,V\rangle_P=\operatorname{tr}(P^{-1}UP^{-1}V)$ is the Riemannian symmetric space of the pair $(\mathrm{GL}_n,\mathrm{O}(n))$. The flat factor. The scalar line $\{x\mathrm{I}\}$ lies in $\mathfrak{a}$ but is fixed by $K$, so the cone splits as $\mathbb{R}\times\{P:\det P=1\}$: a line of pure scaling and the stretches of unit determinant, which form the symmetric space of $\mathrm{SL}_n(\mathbb{R})$. The two summands explicitly. $\mathfrak{k}=\mathfrak{so}(2)$ is the line of matrices $\begin{pmatrix}0&-t\\t&0\end{pmatrix}$, and $\mathfrak{p}=\operatorname{Sym}_2$ is the three-dimensional space $\begin{pmatrix}x&y\\y&z\end{pmatrix}$. The dimensions are $1+3=4=\dim M_2(\mathbb{R})$. The stretch has a closed form. For $A=\begin{pmatrix}a&b\\c&d\end{pmatrix}$ invertible, put $M=A^{\mathsf T}A$ and $\delta=|\det A|$. Then $$
P=\frac{M+\delta\,\mathrm{I}}{\sqrt{\operatorname{tr}M+2\delta}},
$$ which is symmetric positive definite and agrees with the spectral square root to $2\times10^{-16}$. The rotation is $Q=AP^{-1}$, with $\det Q=\operatorname{sgn}\det A$. The roots at $n=2$. $\mathfrak{a}$ is the diagonal, of dimension two, and the only roots are $\pm(x_1-x_2)$: one positive root, of multiplicity one, plus the central line $x_1=x_2$. The Weyl group is $S_2=\{\pm1\}$, acting by swapping the diagonal entries. So $$
\mathfrak{gl}_2(\mathbb{R})=\mathfrak{a}\oplus\mathbb{R}E_{12}\oplus\mathbb{R}E_{21},
$$ of dimensions $2+1+1=4$, and the two root lines are exchanged by the Weyl group. The quotient at $n=2$. The cone is three-dimensional, the flat factor is the scaling line, and the remaining two dimensions are those of $\mathrm{SL}_2(\mathbb{R})/\mathrm{SO}(2)$. The decomposition is a diffeomorphism between the circle times a three-dimensional cone and the four-dimensional group; the reconstruction residual at $n=2$ is $2\times10^{-16}$. The Cartan decomposition of $\mathrm{GL}_n(\mathbb{R})$ is $G=K\exp(\mathfrak{p})$ with $K=\mathrm{O}(n)$ and $\mathfrak{p}=\operatorname{Sym}_n$, and the product map is a diffeomorphism from $K\times\mathfrak{p}$ onto $G$. At the level of the Lie algebra, $\mathfrak{gl}_n(\mathbb{R})=\mathfrak{k}\oplus\mathfrak{p}$ with $\mathfrak{k}=\mathfrak{so}(n)$, and the three brackets $[\mathfrak{k},\mathfrak{k}]\subseteq\mathfrak{k}$, $[\mathfrak{k},\mathfrak{p}]\subseteq\mathfrak{p}$, $[\mathfrak{p},\mathfrak{p}]\subseteq\mathfrak{k}$ hold. The splitting is the eigenspace splitting of the involution $\theta(X)=-X^{\mathsf T}$, whose group form $\Theta(g)=(g^{\mathsf T})^{-1}$ fixes exactly $\mathrm{O}(n)$. The maximal abelian subspace of $\mathfrak{p}$ is the diagonal, the restricted roots are the differences $x_i-x_j$ with multiplicity one, and the Weyl group is $S_n$. The quotient $G/K$ is the cone of positive definite matrices, of dimension $n(n+1)/2$, and the count $n^{2}=n(n-1)/2+n(n+1)/2$ closes the article. At $n=2$ the summands have dimensions $1$ and $3$, the positive root is $x_1-x_2$, and the stretch is $(A^{\mathsf T}A+|\det A|\mathrm{I})/\sqrt{\operatorname{tr}(A^{\mathsf T}A)+2|\det A|}$.The Symmetric Space
The Case at $n=2$
Summary
Summary of Notation
Symbol
Meaning
$K=\mathrm{O}(n)$
The maximal compact subgroup
$\mathfrak{k}=\mathfrak{so}(n)$
The antisymmetric matrices, the Lie algebra of $K$
$\mathfrak{p}=\operatorname{Sym}_n$
The symmetric matrices, the second summand
$\theta(X)=-X^{\mathsf T}$
The Cartan involution on the Lie algebra
$\Theta(g)=(g^{\mathsf T})^{-1}$
Its group form
$\mathfrak{a}$
The diagonal matrices in $\mathfrak{p}$
$\alpha_{ij}(x)=x_i-x_j$
The restricted roots
$W=S_n$
The Weyl group
$G/K\cong\operatorname{Sym}_n^{>0}$
The symmetric space, of dimension $n(n+1)/2$
Further Reading