Let $\def\S{\mathbf S}\S^n$ be the linear space of symmetric $n \times n$ matrices and $\S_+^n$ be the subset of positive semidefinite matrices. It is well-known that $\S_+^n$ is a convex cone in $\S^n$. In order to get a better geometric understanding of this object, I asked myself what the apex angle of this cone might be.

We use the inner product $\DeclareMathOperator{\tr}{tr}\def\<{\langle}\def\>{\rangle}\<A,B\>=\tr(AB)$, where $\tr(A)$ is the trace of $A$.

The apex angle $\theta$ of $\S_+^n$ is the biggest value of $\arccos\<A_1,A_2\>$ for $A_i\in\S_+^n$ with $\<A_i,A_i\>=1$.

My best result so far

Let $\def\E{\mathbf E}\E$ be some Euclidean space and $S\subset \E$ a proper subspace. Let $A_1\in\S_+^n$ be the orthogonal projection onto $S$ and $A_2\in\S_+^n$ the orthogonal projection onto the orthogonal complement $S^\bot$. Then $\<A_i,A_i\>=1$ but $A_1A_2=0$, hence $\<A_1,A_2\>=\tr(A_1A_2)=0$.

So we have that $\theta\ge 90^\circ$. Can we do better? Especially, can we have $\tr(A_1A_2)<0$?


It is impossible to have $\operatorname{tr}(A_1A_2) < 0$ if each $A_i$ is positive semidefinite.

A proof I like is as follows: note that (via the spectral theorem, for instance) we may decompose $A_2$ into $$ A_2 = \sum_{k=1}^n x_kx_k^T, \qquad x_k \in \Bbb R^n $$ With that, we find that $$ \operatorname{tr}(A_1A_2) = \operatorname{tr}\left(A_1\sum_{k=1}^n x_kx_k^T\right) = \sum_{k=1}^n \operatorname{tr}(A_1x_kx_k^T) = \sum_{k=1}^n \operatorname{tr}(x_k^TA_1x_k) = \\ \sum_{k=1}^n x_k^TA_1x_k \geq 0 $$

  • $\begingroup$ Another notable result is that the cone $S_+^n$ is self-dual $\endgroup$ – Omnomnomnom Jan 4 '18 at 16:06
  • $\begingroup$ Very nice proof! Thank you very much. $\endgroup$ – M. Winter Jan 4 '18 at 16:08
  • $\begingroup$ No problem. Off the top of my head, I believe that all self-dual (closed, convex) cones have a $90^\circ$ apex angle, but that's just a hunch. $\endgroup$ – Omnomnomnom Jan 4 '18 at 16:09
  • $\begingroup$ Yeah, I had the same thought. I will think about it. $\endgroup$ – M. Winter Jan 4 '18 at 16:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.