0
$\begingroup$

Are there convex functions for which hessian is not defined, but the gradient is defined everywhere?

I was looking at projected gradient descent, as well as Newton's method for solving optimization problem involving linear constraint. In order to justify using projected gradient descent despite it's problem with ill-conditioning, I was looking at cases where function might be convex and differentiable, but not have a hessian defined.

$\endgroup$
  • 1
    $\begingroup$ $f(x)$ given by $0, x<0; x^2, x\ge 0$? $\endgroup$ – mfl Aug 17 '18 at 21:10
  • 2
    $\begingroup$ $\|x\|^{3/2}$ does not have a Hessian at $x={\bf 0}$. $\endgroup$ – Mike Earnest Aug 17 '18 at 21:13
1
$\begingroup$

Let $f$ be any nondecreasing and continuous function on $\mathbb R$ that has points of nondifferentiability. Then $F(x)=\int_0^x f$ is an example. You can get pretty wild with this $f;$ for example there is one whose corresponding $F$ fails to be twice differentiable at each rational.

$\endgroup$
  • $\begingroup$ This does not guarantee convexity I suppose? $\endgroup$ – Sridhar Thiagarajan Aug 18 '18 at 6:56
  • $\begingroup$ @SridharThiagarajan Sure it does. $F'=f$ is nondecreasing, therefore $F$ is convex. $\endgroup$ – zhw. Aug 18 '18 at 8:07
0
$\begingroup$

Say $f(x) = 0$ for $x<0$ and $x^2$ for $x\geq 0$. It has a derivative at all points, but is not twice differentiable at $0$.

$\endgroup$

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.