Let $f: D \to \mathbb{R}\ $ be a function on non-singular, convex domain $D \subseteq \mathbb{R}^d$ and let us assume the second-order derivatives of $f$ exist. It is well known that $f$ is convex if and only if its Hessian $\nabla^2 f(x)$ is positive semi-definite for all $x \in D$. It is also known that if $\nabla^2 f(x)$ is positive definite for all $x \in D$, we may conclude that $f$ is strictly convex (for a reference, see Boyd and Vandenberghe, 2004).

On the other hand, if $f$ is strictly convex, we still merely know that $\nabla^2 f(x)$ is positive semi-definite for all $x \in D$. That is, there may be $x \in D$ such that $y^T \nabla^2 f(x) y = 0\ $ for some $y\not=0$.

As an example, consider $f(x)=x^4$. In this case, $f$ is strictly convex, but $f''(x)=12x^2$ and, hence, $yf''(x)y=0$ for $x=0$ and $yf''(x)y>0$ for all $x\not=0$.

Yet, these points that ruin the complete positive definiteness seem to be very sparsely distributed within $D$. So, my question is as follows:

If $f$ is strictly convex, how can we characterize the set of points $X$ for which $\nabla^2f(x)$ is not positive definite for $x \in X$, and $\nabla^2f(x)$ is positive definite for $x \in D \setminus X$?

That is, has such a set $X$ been investigated before, what properties are known, and where can I learn more about it? Any reference is welcome.

In particular, my guess is that on can state the following:

Conjecture 1: The set $X$ is merely a discrete subset of $D$.

EDIT: Since Conjecture 1 has obviously been disproved by George Lowther below, allow me to restate my guess as the following (less bold) statement:

Conjecture 2: The set $X$ is does not contain a non-empty open ball.

or, even more cautious:

Conjecture 3: The set $D \setminus X$ does contain a non-empty open ball.

  • $\begingroup$ I'm no expert, but this reminds me of Sard's theorem. The right notion of smallness for $X$ is probably that it has measure $0$. $\endgroup$ – Greg Graviton Aug 26 '10 at 10:36
  • $\begingroup$ Thanks for the reference, Sard's theorem indeed seems kind of related. However, it seems that in a certain sense the set $X$ is even "smaller" than just a null set. For instance, $X$ may not contain any non-singular line segment (otherwise, $f$ is not strictly convex on this line segment). I edited my question to make this a little bit clearer. $\endgroup$ – MRA Aug 26 '10 at 12:52
  • 1
    $\begingroup$ You could have $\partial^2 f/\partial y^2=0$, but $\partial^2 f/\partial x^2>0$, on the line $y=0$. $\endgroup$ – George Lowther Aug 26 '10 at 23:40
  • $\begingroup$ A proof of Conjecture 2 in dimension d=1: Assume such an open interval exists, that is, there is a (closed, non-empty) interval $[a,b]$ such that $f''(t)=0$ for all $t \in [a,b]$. By Taylor expansion, we find $f(b) = f(a) + f'(a)\cdot(b-a)$. Let $c=(a+b)/2$. Again using Taylor expansion, we find $f(c) = f(a) + f'(a)\cdot(c-a) = f(a) + f'(a)\cdot(b-a)/2$. Hence, $(f(a)+f(b))/2 = f(c)$, which contradicts the strict convexity of $f$. $\endgroup$ – MRA Aug 27 '10 at 12:45

I can't fully characterize such sets, but can say one thing - the conjecture is false. For a counterexample on one dimension, choose a closed and nowhere dense set X (such as the Cantor middle thirds set, which is not discrete) and let $g(x)=\min\{\vert x-y\vert\colon y\in X\}$ be the distance of a point x from X. Integrate it twice, $f(x)=\int_0^x\int_0^yg(z)\,dz\,dy$ to get a strictly convex function whose second derivative $f^{\prime\prime}=g$ vanishes precisely on X.

In fact, you could take X to be the Smith-Volterra-Cantor set, which has positive Lebesgue measure.

This can be used to define a strictly convex function on $\mathbb{R}^2$, $\tilde f(x,y)=f(x)+f(y)$, whose Hessian vanishes completely on a set $X\times X$ of positive Lebesgue measure.

  • $\begingroup$ Thanks, this is a very convincing counter example. (I wasn't even aware that there are nowhere dense sets of positive Lebesgue measure.) I took the liberty of updating my question. $\endgroup$ – MRA Aug 27 '10 at 11:19

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.