The Geometry of the Two Halves and the Riemannian Reading

Introduction

The algebra is a pseudo-Euclidean vector space of inertia $(n(n+1)/2,\,n(n-1)/2)$, the two halves are its orthogonally complementary definite subspaces, and the group carries a bi-invariant pseudo-Riemannian metric whose curvature is the commutator; so the entire geometry of the two halves is read from one indefinite scalar product.

The article fixes the scalar product $\langle X,Y\rangle=\operatorname{tr}(XY)$, proves the inertia, shows that the two halves are flat totally geodesic subspaces of opposite definiteness, distinguishes the flat exponential from the matrix exponential, defines the bi-invariant metric on $\mathrm{GL}_n(\mathbb{R})$ and identifies its geodesics, and computes the curvature as a bilinear expression in the bracket. It assumes The Trace Form and Its Indefinite Inertia and The Exponential of the Symmetric Part: Stretches and the Positive Cone. It defers the symmetric space to The Symmetric Space $\mathrm{GL}_n(\mathbb{R})/\mathrm{O}(n)$ and Its Complex Twin.

The Indefinite Scalar Product

Definition. On $M_n(\mathbb{R})$ put $\langle X,Y\rangle=\operatorname{tr}(XY)$. It is symmetric and nondegenerate, being the trace of the product, and it is invariant under the transpose and under conjugation: $\langle AX,XB\rangle=\langle X,B\!AX\rangle$.

Theorem. The inertia is $\bigl(n(n+1)/2,\ n(n-1)/2\bigr)$, that is $p=n(n+1)/2$ positive and $q=n(n-1)/2$ negative directions.

Proof. The two halves are orthogonal, since $\operatorname{tr}(SK)=0$ for $S$ symmetric and $K$ antisymmetric, because $SK$ and $KS=(SK)^{\mathsf T}$ have the same trace. On $\operatorname{Sym}_n$ the form is positive definite, $\operatorname{tr}(S^{2})=\|S\|_{2}^{2}>0$ for $S\ne0$. On $\operatorname{Skew}_n$ it is negative definite, $\operatorname{tr}(K^{2})=-\|K\|_{2}^{2}<0$. So the inertia is the pair of dimensions, measured as $(3,1),(6,3),(10,6),(15,10)$ at $n=2,3,4,5$.

So the algebra is a pseudo-Euclidean space with a preferred orthogonal splitting into a definite positive and a definite negative part. The splitting is exactly the splitting into the two halves, and it is the only natural one.

The cone of isotropic vectors is worth naming. $X$ is isotropic when $\operatorname{tr}(X^{2})=0$. For a symmetric $X$ that happens only at $X=0$, since the form is definite there; for an antisymmetric $X$ likewise. So every isotropic vector has both halves nonzero, and the isotropic cone is the union of the lines on which the two quadratic forms cancel.

The form is not the Hilbert–Schmidt form. $\operatorname{tr}(X^{\mathsf T}X)$ is definite and equal to the Euclidean norm; $\operatorname{tr}(X^{2})$ is indefinite and equal to the difference of the norms of the two halves:

$$ \langle X,X\rangle=\|S\|_{2}^{2}-\|K\|_{2}^{2}\quad\text{for }X=S+K . $$

The Two Halves as Flat Subspaces

Each half is a linear subspace, hence flat and totally geodesic. In a pseudo-Euclidean vector space the geodesics are the straight lines, so a linear subspace is totally geodesic, and the induced metric on it is flat.

The two halves realise the two halves of the inertia. On $\operatorname{Sym}_n$ the form is positive definite of dimension $n(n+1)/2$, on $\operatorname{Skew}_n$ negative definite of dimension $n(n-1)/2$, exactly the two indices. They are not the only definite subspaces of those dimensions: a graph of a symmetric map over the antisymmetric half can be positive definite too. What is special about the halves is that they are the unique orthogonal pair of definite subspaces covering the whole space.

The intrinsic geometry of $\operatorname{Sym}_n$ is Euclidean. Its distance is $\|S-S'\|_{2}$, its geodesics are the affine lines $S+tH$, and its isometry group is the semidirect product of the translations and $\mathrm{O}(\operatorname{Sym}_n)$.

The intrinsic geometry of $\operatorname{Skew}_n$ is Euclidean with the opposite sign. Distances are the same numbers, $\|K-K'\|_{2}$, computed from $-\operatorname{tr}((K-K')^{2})$, but the squared length of a vector is negative, so the half is a negative definite pseudo-Riemannian manifold and not a Riemannian one. Its geodesics are still the affine lines.

Consequence: the algebra is the orthogonal direct sum of a Euclidean and an anti-Euclidean space. Every vector has a positive length, a negative length, or zero, and the sign is the sign of $\|S\|_{2}^{2}-\|K\|_{2}^{2}$. The three regions are the two parts of the isotropic cone and the cone itself.

The Two Exponentials

The flat exponential and the matrix exponential are different maps with the same source. The flat exponential at the origin of the vector space is the identity: $\operatorname{Exp}_0(V)=V$. The matrix exponential is the series, $\operatorname{Exp}(V)=e^{V}$. Both start at $\mathrm{I}$ when read as maps into the group, and their derivatives at $0$ agree, but their global behaviour is different.

The distance in the group metric is the length of the tangent vector, not the length of the image. For $S=\begin{pmatrix}1&0.5\\0.5&-0.2\end{pmatrix}$ the length of $S$ is $\sqrt{\operatorname{tr}S^{2}}=\sqrt{1.54}\approx1.241$, while the Euclidean distance from $\mathrm{I}$ to $e^{S}$ is $\|e^{S}-\mathrm{I}\|_{2}\approx2.2798$. So the exponential is far from an isometry.

The matrix exponential is a chart for the cone, and the chart is not the Euclidean one. The map $S\mapsto e^{S}$ is a diffeomorphism of the half onto the cone, and pulling back the flat metric of the half gives the cone a flat metric in which the exponential is an isometry by construction. That metric is not the restriction of the algebra metric to the cone.

The flat metric and the cone metric differ by the derivative of the exponential. The pullback is $\langle d\exp_S(H),d\exp_S(H')\rangle$, and it equals the flat one only where the dexp is orthogonal, that is on the centre.

The Bi-Invariant Metric on the Group

Definition. On the group put $g_A(H,H')=\operatorname{tr}(A^{-1}HA^{-1}H')$ for tangent vectors $H,H'\in T_A\mathrm{GL}_n(\mathbb{R})=M_n(\mathbb{R})$.

Theorem. $g$ is a pseudo-Riemannian metric of constant inertia $(p,q)$, invariant under both left and right translations, and its geodesics through $\mathrm{I}$ are exactly the one-parameter subgroups $t\mapsto e^{tX}$.

Proof. The formula is left-invariant by construction: translating $H$ by $A$ turns $A^{-1}H$ into $X$ and leaves the expression $\langle X,X'\rangle$ unchanged. Right translation by $B$ sends $A^{-1}H$ to $B^{-1}A^{-1}HB$, and cyclicity of the trace gives $\operatorname{tr}(B^{-1}MB\,B^{-1}NB)=\operatorname{tr}(MN)$, so the metric is bi-invariant. Inertia: at $A=\mathrm{I}$ the metric is the trace form, of inertia $(p,q)$. Geodesics: for a bi-invariant metric the Levi-Civita connection is $\nabla_XY=\tfrac12[X,Y]$ on left-invariant fields, and the curve $t\mapsto e^{tX}$ has velocity field right-invariant along itself, so $\nabla_{\dot\gamma}\dot\gamma=\tfrac12[\dot\gamma,\dot\gamma]=0$.

So the matrix exponential is the Riemannian exponential at the identity. This is the one place where the algebra exponential and the geometry agree, and it is the reason the exponential is the natural map in the Lie theory.

The exponential is not the Riemannian exponential at other points. At a point $A\ne\mathrm{I}$ the geodesics are the curves $Ae^{tX}$, not the curves $e^{A+tX}$; the metric is left-invariant, so it sees only the left-translated structure.

The inertia is indefinite, so the metric has null directions. The left-invariant vector fields with $X$ isotropic are null, and the light cone at each point is the left translate of the isotropic cone of the algebra.

Curvature

For a bi-invariant metric the curvature tensor is a bracket.

$$ R(X,Y)Z=-\tfrac14\bigl[[X,Y],Z\bigr],\qquad K(X,Y)=\tfrac14\langle[X,Y],[X,Y]\rangle . $$

The first formula is the standard one for a bi-invariant metric on a Lie group, and the second is the sectional curvature of the plane spanned by two vectors that are orthonormal and non-null.

Proof. For the bi-invariant metric the connection is $\nabla_XY=\tfrac12[X,Y]$, so $R(X,Y)Z=\nabla_X\nabla_YZ-\nabla_Y\nabla_XZ-\nabla_{[X,Y]}Z$ becomes $\tfrac14[Y,[X,Z]]-\tfrac14[X,[Y,Z]]-\tfrac12[[X,Y],Z]$, and the Jacobi identity turns this into $-\tfrac14[[X,Y],Z]$.

Consequence: the metric is flat exactly where the bracket vanishes. A bi-invariant metric on a connected Lie group is flat only when the group is abelian, so $\mathrm{GL}_n(\mathbb{R})$ with this metric is never flat for $n\ge2$. The directions that stay flat are the ones whose brackets vanish, that is the directions in a commutative subalgebra.

Consequence: the curvature on $\mathfrak{so}(3)$ is negative with the trace form. Take $X=E_{12}/\sqrt{2}$ and $Y=E_{23}/\sqrt{2}$, orthonormal for $-\operatorname{tr}$ on $\mathfrak{so}(3)$. Then $[X,Y]=E_{13}/2$ and $\operatorname{tr}([X,Y]^{2})=-1/2$, so

$$ K(X,Y)=\tfrac14(-0.5)=-0.125 . $$

So the sectional curvature is negative on the compact part, which is the opposite sign from the usual normalisation of a compact group and is the effect of the indefinite trace form.

Consequence: the symmetric half is flat while the group is not. The half is a vector space with the flat metric; the group is curved by the bracket. The passage from the half to the group is the exponential, and the curvature is the obstruction to the exponential's being a local isometry.

The two readings are related by the Cartan decomposition. $\mathfrak{gl}_n=\mathfrak{p}\oplus\mathfrak{k}$ with $\mathfrak{p}=\operatorname{Sym}_n$ and $\mathfrak{k}=\operatorname{Skew}_n$, and $[\mathfrak{p},\mathfrak{p}]\subseteq\mathfrak{k}$, $[\mathfrak{p},\mathfrak{k}]\subseteq\mathfrak{p}$, $[\mathfrak{k},\mathfrak{k}]\subseteq\mathfrak{k}$. So the curvature of the symmetric space is computed from $[\mathfrak{p},\mathfrak{p}]$, and the flat directions inside $\mathfrak{p}$ are the commuting ones.

Summary

The trace form gives $M_n(\mathbb{R})$ the structure of a pseudo-Euclidean space of inertia $(n(n+1)/2,\,n(n-1)/2)$, with the two halves as orthogonally complementary definite subspaces: $\operatorname{Sym}_n$ positive definite, $\operatorname{Skew}_n$ negative definite, each flat and totally geodesic. The flat exponential is the identity and the matrix exponential is a chart with a different metric; the length $\sqrt{1.54}\approx1.241$ of a symmetric matrix differs from the distance $\approx2.2798$ between $\mathrm{I}$ and its exponential. On the group the metric $g_A(H,H')=\operatorname{tr}(A^{-1}HA^{-1}H')$ is bi-invariant of the same inertia and its geodesics through $\mathrm{I}$ are the one-parameter subgroups, so the matrix exponential is the Riemannian exponential at $\mathrm{I}$. The curvature is $R(X,Y)Z=-\tfrac14[[X,Y],Z]$, negative on the compact part, with $K=-0.125$ for an orthonormal plane in $\mathfrak{so}(3)$.

Summary of Notation

Symbol Meaning
$\langle X,Y\rangle=\operatorname{tr}(XY)$ the indefinite scalar product of inertia $(p,q)$
$g_A(H,H')$ the bi-invariant metric on the group
$R(X,Y)Z$ the curvature tensor, $-\tfrac14[[X,Y],Z]$
$K(X,Y)$ the sectional curvature, $\tfrac14\langle[X,Y],[X,Y]\rangle$
$\mathfrak{p},\mathfrak{k}$ the symmetric and antisymmetric halves in the Cartan decomposition

Further Reading