1
$\begingroup$

Consider a real symmetric matrix $A\in\mathbb{R}^{n \times n}$. The associated quadratic form $x^T A x$ is a convex function on all of $\mathbb{R}^n$ iff $A$ is positive semidefinite, i.e., if $x^T A x \geq 0$ for all $x \in \mathbb{R}^n$.

Now suppose we have a convex subset $\Phi$ of $\mathbb{R}^n$ such that $x \in \Phi$ implies $x^T A x \geq 0$. Is $x^T A x$ a convex function on $\Phi$ (even if $A$ is not positive definite)? Of course, the answer in general is "no," but we can still ask about the most inclusive conditions under which convexity holds for a given $A$ and $\Phi$. In particular I'm interested in the question:

Suppose we have a quadratic form $Q:\mathbb{R}^{n \times n} \rightarrow \mathbb{R}$. What is the weakest condition on $Q$ that guarantees it will be convex when restricted to the set of positive semidefinite matrices?

Improve this question
$\endgroup$
1
  • $\begingroup$ $Q$ must be positive semidefinite on the set of all symmetric matrices. By the way, if you really need help with a mathematical question going beyond routine homeworks and think that it is too "specific" or "trivial" for MO, try College Playground on AoPS. We welcome there almost everything with mathematical content (no pure philosophy, please!) Just use your common sense when choosing the subforum and stick to some other easy to guess rules. You may not get as qualified answer there as on MO but you won't be told that your question is of no interest either. $\endgroup$
    – fedja
    Dec 31, 2009 at 5:06

2 Answers 2

Reset to default
8
$\begingroup$

$x^2-y^2$ is positive on $[2,3]\times [-1,1]$ but not convex there. This creates problems for any convex sets not containing the origin. You are, probably, after something else not so obviously false. Why don't you just tell us what it is?

Edit: Even then it is false: just take $B_{11}B_{22}$. By the way, for a pure quadratic form, convexity on an open set and convexity on the entire space are the same thing.

Improve this answer
$\endgroup$
5
  • $\begingroup$ And so did I :) $\endgroup$
    – fedja
    Dec 31, 2009 at 3:24
  • $\begingroup$ >>By the way, for a pure quadratic form, convexity on an open set and convexity on the entire space are the same thing. Thanks for pointing that out. Unfortunately, the cone of positive semidefinite matrices is not an open subset, so that doesn't give us an easy answer. $\endgroup$
    – TerronaBell
    Dec 31, 2009 at 4:50
  • $\begingroup$ This answer also covers Edit 2 of the question, since the set of positive definite matrices is open. $\endgroup$
    – Harald Hanche-Olsen
    Dec 31, 2009 at 5:07
  • 1
    $\begingroup$ Just an elaboration on fedja's answer. Let $\Phi\subset{\mathbb R}^n$ be any convex (nonempty) set. Then there is unique subspace $V\subset{\mathbb R}^n$ such that $\Phi$ contains an open subset of the coset $V+x$ for some $x\in{\mathbb R}^n$ ($V$ is spanned by the set $\Phi-x$ for any fixed vector $x$). In your example, $\Phi$ consists of positive semi-definite matrices, and $V$ consists of all symmetric matrices. Now if $Q$ is any quadratic form, it is convex on $\Phi$ if and only if it is convex (positive semi-definite) on $V$. This gives an explicit criterion. $\endgroup$
    – t3suji
    Dec 31, 2009 at 5:14
  • $\begingroup$ Hi t3suji -- just the answer I was looking for (thanks for being so explicit). If you make this an answer rather than a comment I will make this the accepted answer. $\endgroup$
    – TerronaBell
    Dec 31, 2009 at 16:18
4
$\begingroup$

The answer to the edited question is no. Let $Q\colon\Phi\to\mathbb{R}$ be the quadratic form $Q(B)=B_{11}B_{22}$. Clearly $Q\gt0$ on the set $\Phi$ of positive definite matrices. Equally clearly, this function is concave on the subset $\{B\in\Phi\colon B_{11}+B_{22}=1\}$.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Oops, only on coming back here did I realize that I had duplicated the result of the second paragraph of fedja's answer. Sorry about that. (Or did fedja's edit happen while I was typing in my answer?) $\endgroup$
    – Harald Hanche-Olsen
    Dec 31, 2009 at 5:10

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.