The Positive Cone and Its Boundary
Introduction
The positive definite symmetric matrices form an open convex cone, and that cone is the whole of the noncompact geometry of the general linear group: it is a single orbit under congruence, it is self-dual for the trace form, and its boundary is stratified by the rank.
The article proves openness, convexity and the congruence transitivity of the cone, identifies its closure with the positive semidefinite matrices, proves self-duality, describes the face lattice as the lattice of subspaces, identifies the extreme rays with the Veronese variety, computes the dimension $r(n-r)+r(r+1)/2$ of the stratum of rank $r$, and proves that the cone is a symmetric cone of rank $n$ with the Jordan algebra $\operatorname{Sym}_n$. It assumes The Spectral Theorem: the Symmetric Case and The Exponential of the Symmetric Part: Stretches and the Positive Cone. It defers the global geometry to The Symmetric Space $\mathrm{GL}_n(\mathbb{R})/\mathrm{O}(n)$ and Its Complex Twin.
The Open Cone
Definition.
$$ \operatorname{Sym}_n^{>0}=\{X\in\operatorname{Sym}_n:v^{\mathsf T}Xv>0\ \text{for every}\ v\ne0\}. $$
Theorem. $\operatorname{Sym}_n^{>0}$ is open in $\operatorname{Sym}_n$, convex, and invariant under congruence $X\mapsto A^{\mathsf T}XA$ for every invertible $A$.
Proof. Openness: the smallest eigenvalue is a continuous function of the entries, by the min-max characterisation, and the set is $\{\lambda_{\min}>0\}$. Convexity: if $X,Y$ are positive definite and $t\in[0,1]$ then $v^{\mathsf T}(tX+(1-t)Y)v=t\,v^{\mathsf T}Xv+(1-t)v^{\mathsf T}Yv>0$ for $v\ne0$ as a convex combination of two positive numbers. Invariance: $v^{\mathsf T}A^{\mathsf T}XAv=(Av)^{\mathsf T}X(Av)>0$ for $v\ne0$, since $Av\ne0$.
Theorem. $\mathrm{GL}_n(\mathbb{R})$ acts transitively on $\operatorname{Sym}_n^{>0}$ by congruence, and the stabiliser of $\mathrm{I}$ is $\mathrm{O}(n)$.
Proof. The spectral theorem writes $X=Q\Lambda Q^{\mathsf T}$ with $Q$ orthogonal and $\Lambda$ diagonal positive; then $X=(Q\Lambda^{1/2})^{\mathsf T}(Q\Lambda^{1/2})$, so $X=A^{\mathsf T}\mathrm{I}A$ with $A=\Lambda^{1/2}Q^{\mathsf T}$ invertible. The stabiliser: $A^{\mathsf T}A=\mathrm{I}$ is the definition of orthogonality.
Consequence: the cone is a homogeneous space. $\operatorname{Sym}_n^{>0}\cong\mathrm{GL}_n(\mathbb{R})/\mathrm{O}(n)$, of dimension $n(n+1)/2$. The identification $A\mapsto A^{\mathsf T}A$ is a diffeomorphism, and it is the polar decomposition read as a quotient map.
The cone is the exponential image of the whole half. $\exp:\operatorname{Sym}_n\to\operatorname{Sym}_n^{>0}$ is a bijection with inverse $\log$, so the cone is also a vector space in the logarithm coordinates. The two readings, congruence and exponential, are the two structures the cone carries.
The Closure and the Boundary
The closure is the cone of positive semidefinite matrices.
$$ \overline{\operatorname{Sym}_n^{>0}}=\{X\in\operatorname{Sym}_n:v^{\mathsf T}Xv\ge0\ \text{for every}\ v\}=\{X\in\operatorname{Sym}_n:\lambda_{\min}\ge0\}. $$
The equality of the three descriptions is the spectral theorem plus the continuity of the smallest eigenvalue.
The boundary is the set of positive semidefinite matrices with a nonzero kernel. It is the union of the strata of rank $r The strata are manifolds. Let $\Sigma_r$ be the symmetric matrices of rank exactly $r$. Then $\Sigma_r$ is a smooth manifold of dimension $$
\dim\Sigma_r=r(n-r)+\frac{r(r+1)}{2},
$$ obtained by writing $X=USU^{\mathsf T}$ with $U$ an orthonormal $n\times r$ frame and $S$ an invertible symmetric $r\times r$ matrix: the frame contributes $r(n-r)$ after quotienting by $\mathrm{O}(r)$ and the matrix contributes $r(r+1)/2$. The dimension is verified as the rank of the tangent map at $\operatorname{diag}(\mathrm{I}_r,0)$, giving $2,3,3,5,6,4,7,9,10$ at $(n,r)=(2,1),(2,2),(3,1),(3,2),(3,3),(4,1),(4,2),(4,3),(4,4)$, matching the formula. The top stratum is the open cone and the bottom stratum is the origin. At $r=n$ the formula gives $n(n+1)/2$, the dimension of the whole space; at $r=0$ it gives $0$. The dimension increases with $r$, so the boundary has codimension at least one everywhere. The positive semidefinite cone with rank bound is the closure of the stratum. $\overline{\Sigma_{\le r}}=\{X\ge0:\operatorname{rk}X\le r\}$, and the interior of that set inside $\operatorname{Sym}_n$ is empty for $r Theorem. The faces of the closed cone $\overline{\operatorname{Sym}_n^{>0}}$ are exactly the sets $$
F_U=\{X\ge0:\operatorname{im}X\subseteq U\},
$$ one for each subspace $U\subseteq\mathbb{R}^{n}$, and $\dim F_U=(\dim U)(\dim U+1)/2$. Proof. A face is a convex subset $F$ with the property that $X+Y\in F$ and $X,Y\ge0$ imply $X,Y\in F$. If $X\in F$ and $\operatorname{im}X\subseteq U$ then every summand of $X$ in a decomposition into positive semidefinite parts also has image in $U$, since the images add. So $F_U$ is a face. Conversely, if $F$ is a face and $X\in F$ is in its relative interior, then every $Y\ge0$ with $\operatorname{im}Y\subseteq\operatorname{im}X$ lies in $F$, by a rank-one perturbation argument; hence $F=F_U$ with $U=\operatorname{im}X$. The face lattice is the lattice of subspaces. $F_U\subseteq F_W$ exactly when $U\subseteq W$, and $F_{U\cap W}=F_U\cap F_W$, $F_{U+W}=F_U+F_W$. So the face lattice of the cone is isomorphic to the lattice of subspaces of $\mathbb{R}^{n}$, with the cone itself for $U=\mathbb{R}^{n}$ and the origin for $U=0$. The faces come in continua. A face of dimension $d(d+1)/2$ corresponds to a $d$-dimensional subspace, so the faces of a given dimension are parametrised by the Grassmannian, which is a manifold of positive dimension for $0 The extreme rays are the rank-one matrices $vv^{\mathsf T}$. A ray is a one-dimensional face, and the one-dimensional faces correspond to the lines $U$, that is to $v$ up to sign; the positive semidefinite matrix with image in the line and trace one is $vv^{\mathsf T}/\|v\|^{2}$. So the extreme rays are parametrised by the projective space $\mathbb{RP}^{n-1}$, and the set of them is the Veronese variety, the image of $\mathbb{RP}^{n-1}$ under $[v]\mapsto[vv^{\mathsf T}]$. Consequence: the cone is the convex hull of the Veronese variety. Every positive semidefinite matrix is a sum of rank-one ones, $X=\sum_k\lambda_ku_ku_k^{\mathsf T}$, and the sum is a convex combination after normalising. So the cone is the convex hull of its extreme rays, as it must be for a closed convex cone with a compact base. Theorem. The cone is self-dual for the trace form: $$
X\in\overline{\operatorname{Sym}_n^{>0}}\iff\operatorname{tr}(XY)\ge0\ \text{for every}\ Y\in\overline{\operatorname{Sym}_n^{>0}}.
$$ Proof. If $X\ge0$ and $Y\ge0$ then $X=A^{\mathsf T}A$ and $Y=B^{\mathsf T}B$, so $\operatorname{tr}(XY)=\operatorname{tr}(A^{\mathsf T}AB^{\mathsf T}B)=\operatorname{tr}\bigl((AB^{\mathsf T})(AB^{\mathsf T})^{\mathsf T}\bigr)=\|AB^{\mathsf T}\|_{2}^{2}\ge0$. Conversely, if $X$ has a negative eigenvalue with eigenvector $v$, take $Y=vv^{\mathsf T}\ge0$; then $\operatorname{tr}(XY)=v^{\mathsf T}Xv<0$, so $X$ is not in the dual cone. The proof uses only the factorisation $X=A^{\mathsf T}A$, which is the spectral theorem. The self-duality is therefore a consequence of positivity of the Hilbert–Schmidt norm, and it is the reason the cone is the natural domain of semidefinite programming. The dual of the open cone is the closed cone, and the dual of a face is a face. The duality exchanges $F_U$ with $F_{U^{\perp}}$: a positive semidefinite $Y$ pairs nonnegatively with every element of $F_U$ exactly when its image lies in $U^{\perp}$. Self-duality is a real phenomenon. Over $\mathbb{C}$ the analogue is the cone of positive definite Hermitian matrices, self-dual for the Hermitian form $\operatorname{tr}(X^{*}Y)$; the statement survives, and what changes is the dimension, from $n(n+1)/2$ to $n^{2}$. Definition. A cone is symmetric when it is open, convex, self-dual, and homogeneous under its linear automorphism group. Theorem. $\operatorname{Sym}_n^{>0}$ is a symmetric cone, of rank $n$ and dimension $n(n+1)/2$. Proof. Open, convex and self-dual are proved above; homogeneous under $\mathrm{GL}_n(\mathbb{R})$ modulo the scalars, which act trivially on the cone. The rank of a symmetric cone is the length of a maximal chain of faces, minus one: the chain runs from the origin through the faces of dimensions $1,3,\dots,n(n+1)/2$, so it has $n+1$ elements and the rank is $n$. The cone is the cone of squares of the Jordan algebra $\operatorname{Sym}_n$. In the Jordan product $X\circ Y=\tfrac12(XY+YX)$ one has $X\circ X=X^{2}$, so the cone is $\{X^{2}:X\in\operatorname{Sym}_n\}$, and the interior is the set of squares of invertible elements. This is the definition of the symmetric cone of a formally real Jordan algebra. The determinant is the characteristic function of the cone. $\det$ is a polynomial that is positive exactly on the cone, homogeneous of degree $n$, and it satisfies $\det(A^{\mathsf T}XA)=\det(A)^{2}\det X$, which is the transformation law of the characteristic function. Its logarithm is strictly concave on the cone, and that concavity is the analytic expression of the self-duality. The automorphism group of the cone is $\mathrm{GL}_n(\mathbb{R})/\{\pm\mathrm{I}\}$. A linear map preserving the cone preserves the cone of squares, hence the Jordan algebra structure, hence is a congruence by an invertible matrix, by the fundamental theorem of the geometry of symmetric cones. So the cone determines the algebra up to isomorphism. The positive definite symmetric matrices form an open convex cone of dimension $n(n+1)/2$, a single orbit under congruence by $\mathrm{GL}_n(\mathbb{R})$ with stabiliser $\mathrm{O}(n)$, and the exponential image of $\operatorname{Sym}_n$. Its closure is the positive semidefinite cone, self-dual for the trace form, with faces the sets $F_U$ indexed by the subspaces of $\mathbb{R}^{n}$ and extreme rays the Veronese variety of rank-one matrices $vv^{\mathsf T}$. The stratum of rank $r$ is a manifold of dimension $r(n-r)+r(r+1)/2$, measured as $2,3$ at $n=2$ and $3,5,6$ at $n=3$. The cone is symmetric of rank $n$, it is the cone of squares of the Jordan algebra $\operatorname{Sym}_n$, and its characteristic function is the determinant.The Faces
Self-Duality
The Symmetric Cone and the Jordan Algebra
Summary
Summary of Notation
Symbol
Meaning
$\operatorname{Sym}_n^{>0}$
the open cone of positive definite matrices
$\Sigma_r$
the symmetric matrices of rank exactly $r$
$F_U$
the face of matrices with image in $U$
$vv^{\mathsf T}$
an extreme ray, a rank-one positive semidefinite matrix
$\det$
the characteristic function of the cone, homogeneous of degree $n$
$\operatorname{rank}$ of the cone
$n$, the length of a maximal chain of faces
Further Reading