The Two Halves and the Analysis: Smooth Structures and the Exponential

Introduction

The two halves are the eigenspaces of a smooth involution, so they are embedded submanifolds that split the tangent bundle, and the exponential behaves completely differently on them: a diffeomorphism on the symmetric half and a covering on the antisymmetric one.

The article proves that $\operatorname{Sym}_n$ and $\operatorname{Skew}_n$ are embedded submanifolds of dimensions $n(n+1)/2$ and $n(n-1)/2$, that the symmetrisation and antisymmetrisation maps are smooth projections splitting the tangent bundle, that the exponential is a diffeomorphism on one half and a surjective local diffeomorphism on the other, and that the two exponentials do not multiply to the exponential of the sum. It assumes The Analysis on $M_n(\mathbb{R})$, The Exponential of the Symmetric Part: Stretches and the Positive Cone and The Exponential of the Antisymmetric Part: One-Parameter Rotations. It defers the metric to The Geometry of the Two Halves and the Riemannian Reading.

The Two Halves as Submanifolds

Theorem. $\operatorname{Sym}_n$ and $\operatorname{Skew}_n$ are linear subspaces of $M_n(\mathbb{R})$, hence embedded smooth submanifolds of dimensions $n(n+1)/2$ and $n(n-1)/2$ with a single chart each.

Proof. Each is the fixed set of a linear map, the transpose or its negative, so each is a vector subspace. A vector subspace of $\mathbb{R}^{N}$ is an embedded submanifold with the inclusion as a chart.

The sum map is a linear isomorphism.

$$ \Phi:\operatorname{Sym}_n\oplus\operatorname{Skew}_n\to M_n(\mathbb{R}),\qquad \Phi(S,K)=S+K, $$

with inverse $X\mapsto\bigl(\tfrac12(X+X^{\mathsf T}),\tfrac12(X-X^{\mathsf T})\bigr)$. So $\Phi$ is a diffeomorphism, and the two halves give a smooth product decomposition of the algebra.

The inverse formula is the projection onto each factor. Write $\pi_{+}(X)=\tfrac12(X+X^{\mathsf T})$ and $\pi_{-}(X)=\tfrac12(X-X^{\mathsf T})$. Then $\pi_{+}$ and $\pi_{-}$ are linear, hence smooth, and they satisfy

$$ \pi_{+}^{2}=\pi_{+},\qquad \pi_{-}^{2}=\pi_{-},\qquad \pi_{+}\pi_{-}=\pi_{-}\pi_{+}=0,\qquad \pi_{+}+\pi_{-}=\mathrm{id}. $$

So they are complementary idempotent endomorphisms of the vector space, and $\operatorname{Sym}_n$ and $\operatorname{Skew}_n$ are their images.

Consequence: the tangent bundle splits. At every point $X$ the tangent space is $M_n(\mathbb{R})$, and the two projections cut it into a symmetric and an antisymmetric part. The splitting is constant in $X$, so it is a direct sum of two trivial vector bundles, and every smooth curve decomposes into a symmetric and an antisymmetric curve.

The splitting is not a Lie algebra splitting. The symmetric half is not a subalgebra for the commutator, so the splitting of the tangent bundle is not a splitting of the algebra into subalgebras; it is only a splitting of the vector bundle. The Two Halves: Symmetric and Antisymmetric is the article that makes this precise.

The Derivative of the Exponential on Each Half

On the symmetric half the derivative is the dexp acting on a symmetric direction.

$$ d(\exp)_{S}(H)=e^{S}\sum_{k\ge0}\frac{(-1)^{k}}{(k+1)!}\operatorname{ad}_S^{k}(H),\qquad H\in\operatorname{Sym}_n, $$

and the result is symmetric: for $S$ and $H$ both symmetric the commutator $[S,H]$ is antisymmetric, so $\operatorname{ad}_S$ exchanges the two halves, and the $k$-th term is symmetric for $k$ even and anti-self-adjoint for $k$ odd, which is exactly the pairing with the alternating coefficients.

On the antisymmetric half the derivative is the same formula with an antisymmetric direction, and the result is antisymmetric. The two halves are therefore preserved by the derivative, as they must be, since $\exp$ maps each half into a subset of the corresponding symmetric or orthogonal set.

The derivative is invertible on each half at every point. For the symmetric half this follows from the global bijectivity of $\exp$; for the antisymmetric half it follows from the covering property, since a local diffeomorphism has invertible derivative.

The exponential of the sum is not the product of the exponentials. For

$$ S=\begin{pmatrix}1&0.5\\0.5&-0.2\end{pmatrix},\qquad K=\begin{pmatrix}0&0.7\\-0.7&0\end{pmatrix}, $$

the two products differ substantially:

$$ e^{S}e^{K}=\begin{pmatrix}1.7325&2.5368\\ -0.0020&1.2816\end{pmatrix},\qquad e^{S+K}=\begin{pmatrix}2.4953&1.8262\\ -0.3044&0.6691\end{pmatrix}, $$

with a largest entry difference of $0.7628$. The reason is that $SK-KS=\begin{pmatrix}-0.7&0.84\\0.84&0.7\end{pmatrix}\ne0$, so the Baker–Campbell–Hausdorff correction does not vanish.

The Exponential on the Symmetric Half

Theorem. $\exp:\operatorname{Sym}_n\to\operatorname{Sym}_n^{>0}$ is an analytic diffeomorphism, with inverse $\log$.

Proof. Surjective: the spectral theorem writes $P=Q\Lambda Q^{\mathsf T}$ with $\Lambda$ positive diagonal, and then $P=e^{S}$ with $S=Q(\log\Lambda)Q^{\mathsf T}$ symmetric. Injective: if $e^{S}=e^{S'}$ with both symmetric then $S$ and $S'$ commute with $e^{S}$, hence are polynomials in it, hence are simultaneously diagonalisable and agree on the eigenvalues. Analytic with analytic inverse: the logarithm of a positive definite matrix is given by the convergent series in $P-\mathrm{I}$ when $\|P-\mathrm{I}\|<1$ and by the spectral formula in general.

So the symmetric half is a global chart for the cone. The cone is an open convex subset of a vector space of dimension $n(n+1)/2$, and the exponential is a diffeomorphism onto it, which is the strongest possible statement about the exponential on this half.

The derivative is everywhere invertible and the map is proper. Properness follows from the spectral formula: if $e^{S}$ stays in a compact set then the eigenvalues of $S$ stay bounded, so $S$ stays in a compact set. A proper local diffeomorphism that is injective is a diffeomorphism, which gives a second proof.

The Exponential on the Antisymmetric Half

Theorem. $\exp:\operatorname{Skew}_n\to\mathrm{SO}(n)$ is surjective, a local diffeomorphism everywhere, and not injective.

Proof. Surjective: by the spectral theorem for antisymmetric matrices, every rotation is a direct sum of rotations in orthogonal planes, and each planar rotation is $e^{\theta J_2}$; assembling the blocks gives $K\in\operatorname{Skew}_n$ with $e^{K}=Q$. Local diffeomorphism: the derivative is invertible at every point, and the map is equivariant under conjugation so it suffices to check at $0$. Not injective: at $n=2$ the matrix $2\pi J$ satisfies $e^{2\pi J}=\mathrm{I}$, and more generally $\exp$ is $2\pi$-periodic in each angle.

The failure of injectivity is exactly the periodicity of the angles. The eigenvalues of an antisymmetric matrix are $i\theta_1,-i\theta_1,\dots$ and the exponential depends on the angles modulo $2\pi$. So $\exp$ is a covering map onto its image in a neighbourhood of each point, and globally it is a fibration with fibres consisting of the angle shifts by $2\pi\mathbb{Z}$.

The image is exactly $\mathrm{SO}(n)$ and not a proper subset. This is where the antisymmetric half differs from a general subspace: the exponential of an antisymmetric matrix is always orthogonal with determinant one, and conversely every such matrix has an antisymmetric logarithm. So the two halves both have the exponential as a surjection onto their natural targets, and the difference is injectivity.

Consequence: the half is a covering space of the group, and the two are not diffeomorphic. $\operatorname{Skew}_n$ is contractible and $\mathrm{SO}(n)$ is not, so no diffeomorphism can exist; the exponential is a surjective local diffeomorphism between spaces of different homotopy type.

The Smooth Structure of the Involution

The transpose is a smooth involution of the algebra whose fixed and anti-fixed sets are the two halves. It is linear, hence smooth, and it satisfies $\sigma^{2}=\mathrm{id}$, so its differential is itself and the eigenvalues of the differential are $\pm1$ everywhere.

A smooth involution splits the tangent bundle into the $\pm1$ eigenbundles. This is the general fact behind the splitting: for a smooth map $\sigma$ with $\sigma^{2}=\mathrm{id}$ on a manifold, the derivative at a fixed point has eigenvalues $\pm1$ and the fixed set is a submanifold whose tangent space is the $+1$ eigenbundle. Here the manifold is a vector space and the involution is linear, so the statement is the elementary one of the projections $\pi_{\pm}$.

The same pattern appears at the group level. The map $\Theta(A)=(A^{\mathsf T})^{-1}$ is a smooth involution of $\mathrm{GL}_n(\mathbb{R})$ whose fixed set is $\mathrm{O}(n)$; its derivative at the identity is $X\mapsto-X^{\mathsf T}$, whose fixed set is $\operatorname{Skew}_n$. So the Cartan involution of the group and the transpose involution of the algebra have the same fixed space at the identity, and this is the beginning of the Cartan decomposition.

The two involutions differ by a sign and that sign is the whole content of the Cartan decomposition. The algebra involution fixes $\operatorname{Sym}_n$; the group involution, differentiated, fixes $\operatorname{Skew}_n$. The two halves are exchanged by the sign, and the two decompositions differ by which half is the subalgebra.

Analytic Consequences of the Splitting

Every analytic function of a matrix splits along the halves in the sense of its Taylor coefficients. For $f$ analytic near $0$ and $X=S+K$,

$$ f(X)=f(0)+f'(0)(S+K)+\tfrac12f''(0)(S+K)^{2}+\cdots, $$

and the powers mix $S$ and $K$ through their commutator. So the Taylor expansion of $f$ is not a sum of a function of $S$ and a function of $K$; the mixing is exactly the noncommutativity.

The trace kills the cross terms at first order. $\operatorname{tr}(SK)=0$ for $S$ symmetric and $K$ antisymmetric, so the first-order term of $\operatorname{tr}f(X)$ is $f'(0)\operatorname{tr}S$, and the antisymmetric part enters only at second order and through $\operatorname{tr}(K^{2})$, which is negative.

The determinant is blind to the antisymmetric part's sign and sees only its square. $\det e^{S+K}=e^{\operatorname{tr}S}$, since the trace of $K$ vanishes; so the determinant of the exponential depends only on the symmetric part. This is the cleanest analytic statement of the split: the volume change is the symmetric part, and the rotation is the antisymmetric part.

The splitting is the infinitesimal form of the polar decomposition. Every invertible matrix is a rotation times a stretch, and the differential of that statement at the identity is the splitting of the algebra into the two halves.

Summary

$\operatorname{Sym}_n$ and $\operatorname{Skew}_n$ are embedded submanifolds of dimensions $n(n+1)/2$ and $n(n-1)/2$, the images of the smooth complementary projections $\pi_{\pm}$, and they split the tangent bundle of the algebra. The exponential is an analytic diffeomorphism from $\operatorname{Sym}_n$ onto the positive definite cone and a surjective local diffeomorphism from $\operatorname{Skew}_n$ onto $\mathrm{SO}(n)$ that is not injective. The two exponentials do not multiply to the exponential of the sum, with a largest entry difference of $0.7628$ in the explicit example at $n=2$. The involution $\Theta(A)=(A^{\mathsf T})^{-1}$ on the group has derivative $X\mapsto-X^{\mathsf T}$ at the identity, which is why the group fixed set is the anti-Hermitian half while the algebra fixed set is the symmetric half.

Summary of Notation

Symbol Meaning
$\pi_{+},\pi_{-}$ the symmetrisation and antisymmetrisation projections
$\Phi$ the linear isomorphism $\operatorname{Sym}_n\oplus\operatorname{Skew}_n\to M_n(\mathbb{R})$
$e^{S}$, $e^{K}$ the stretch and the rotation parts of the exponential
$\Theta(A)=(A^{\mathsf T})^{-1}$ the Cartan involution of the group
$\operatorname{ad}_S$ the commutator operator used in dexp

Further Reading