5
$\begingroup$

Let $PD_{n}$ be the cone of positive definite $n \times n$ real matrices and let $B$ be the unit sphere in $n \times n$ dimensions. What is the volume of $PD_{n} \cap B$?

EDIT: Let's assume that $B$ is the unit sphere w.r.t the operator norm: $||A||=\sup_{||x||=1}{||Ax||}$.

$\endgroup$
9
  • $\begingroup$ Please clarify your question. The symmetric $n \times n$ real matrices form a linear space of dimension $n(n+1)/2$. It's certainly not a subset of ${\mathbb R}^n$. $\endgroup$ Apr 27, 2014 at 8:15
  • $\begingroup$ @RobertIsrael Sorry, fixed that. <blush> $\endgroup$ Apr 27, 2014 at 8:21
  • 3
    $\begingroup$ Actually, do you mean to ask this in the symmetric $n$-by-$n$ matrices? Otherwise, I don't know what 'positive definite' means. Also, what (positive definite) norm are you assuming on the symmetric $n$-by-$n$ matrices? One of the ones that is invariant under $\mathrm{SO}(n)$? There is a $2$-parameter family of those, and the answer will depend on which one you choose. There is no such norm that is invariant under the full group $\mathrm{GL}(n,\mathbb{R})$. $\endgroup$ Apr 27, 2014 at 12:44
  • $\begingroup$ I'm presuming the cone is of "real symmetric positive definite" matrices, and the unit sphere that is meant is of the operator norm or perhaps Frobenius norm --- but needs to be clarified. $\endgroup$
    – Suvrit
    Apr 27, 2014 at 14:10
  • 1
    $\begingroup$ Have you taken a look at the following question? Looks relevant: mathoverflow.net/questions/118481/… $\endgroup$ May 1, 2014 at 16:26

1 Answer 1

15
$\begingroup$

Using the operator norm, as you have defined it, the fraction of the unit ball in real symmetric $n$-by-$n$ matrices that consists of positive definite matrices is $2^{-n(n+1)/2}$. Thus, this fraction shrinks very quickly as $n$ increases.

Added comment 1: Actually, I just realized that you don't need to do the calculation below; the ratio is obvious anyway: Consider the map $\Phi:B_n\to \overline{PD_n\cap B_n}$ defined by $\Phi(A) = \tfrac12(A+I_n)$. This affine invertible map identifies $B_n$ with $\overline{PD_n\cap B_n}$, which has the same volume as $PD_n\cap B_n$. Obviously, it multiplies volumes by a factor of $2^{-n(n+1)/2}$, since $\tfrac12n(n{+}1)$ is the dimension of the ambient space of symmetric $n$-by-$n$ matrices. QED.

Added comment 2: I note that the OP never did say that he wanted to concentrate on symmetric positive definite matrices, which seems bizarre to me, but maybe he really does mean this. I looked online and found that some sources say that a real $n$-by-$n$ matrix $A$ is positive definite if $x\cdot Ax \ge 0$ for all $x\in\mathbb{R}^n$, with equality if and only if $x = 0$. (Of course, this is a condition only on the symmetric part of $A$.) If one does adopt this convention and considers the problem as being posed on the set of all $n$-by-$n$ real matrices with the operator norm, then the problem becomes much harder, and the above scaling trick doesn't work. I don't know the answer in this case for general $n$, but, for $n=2$, there is a fortuitous simplification: If $$ A = \begin{pmatrix}a^1_1&a^1_2\\ a^2_1&a^2_2\end{pmatrix} = \begin{pmatrix}r+a&b+s\\ b-s& r-a\end{pmatrix}, $$ then, for the operator norm as defined by the OP, one has a nice formula $$ \| A\| = \sqrt{r^2+s^2} + \sqrt{a^2+b^2} $$ and $A$ is positive definite if and only if $r>\sqrt{a^2+b^2}$. (For higher values of $n$, these are not nice at all.)

Now, it's just a calculus problem. The volume of the unit ball turns out to be $V = 2\pi^2/3$, and the volume of the positive definite matrices turns out to be $$ V_+ = 4\pi(53-15\pi)/45 $$ with the resulting ratio being $$ V_+/V \approx 0.2493898621\ . $$ Thus, nearly one-quarter of the unit ball in this case consists of positive definite matrices.


Here is how the calculation goes: First, we need to parametrize this unit ball in some way. The easiest way to do this is to write $$ A = g R g^{-1} $$ where $g$ is an element of the orthogonal group $\mathrm{O}(n)$ and $R=\mathrm{diag}(r_1,\ldots,r_n)$ where the the $r_i$ belong either to the domain $$ B = \{(r_1,\ldots,r_n)\ |\ 1\ge r_1\ge r_2\ge\cdots\ge r_n\ge-1\} $$ (in which case $A$ belongs to the unit ball in the operator norm) or $$ B_+ = \{(r_1,\ldots,r_n)\ |\ 1\ge r_1\ge r_2\ge\cdots\ge r_n\ge0\} $$ (in which case $A$ belongs to the part of the unit ball that consists of positive semidefinite symmetric matrices). (Actually, this parametrization is generically $2^n$-to-$1$ from $B\times\mathrm{O}(n)$ to the unit ball in the operator norm and $2^n$-to-$1$ from $B_+\times\mathrm{O}(n)$ to the postive semi-definite matrices within the unit ball, but that won't matter to the calculation of the ratio of volumes.)

Now, a straightforward calculation using this parametrization shows that the two volumes we want to compute are $\Omega_n$ (the volume of $\mathrm{O}(n)$) divided by $2^n$ times the integrals over $B$ and $B_+$ of the following $n$-form: $$ \phi = \left(\prod_{i<j}(r_i-r_j)\right)\, \mathrm{d}r_1\wedge\mathrm{d}r_2\wedge\cdots\wedge \mathrm{d}r_n $$ Fortunately, it is not necessary to compute this integral over either domain because, if you look at the mapping $f:B_+\to B$ defined by $$ f(r_1,r_2,\ldots,r_n) = (2r_1{-}1,2r_2{-}1,\ldots,2r_n{-}1), $$ you see immediately that $$ f^\ast\phi = 2^{n(n+1)/2}\,\phi $$ Thus, the integral of $\phi$ over $B_+$ divided by the integral of $\phi$ over $B$ is $ 2^{-n(n+1)/2}$, as claimed.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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