4
$\begingroup$

It is easy to see that if a function $f: \mathbb{R} \to \mathbb{R}$ is strictly convex, $C^1$ and $f'$ has bounded image, then as $t\to \infty$ the limit $$ \lim_{t\to\infty} f'(t) = \lim_{t\to\infty} \frac{f(t)}{t} $$ exists and is finite.

Suppose instead that we have a function $F : \mathbb{R}^n \to \mathbb{R}$ that is strictly convex, $C^1$ and the image of $\nabla F$ is bounded. Then do we know that for each $v \in \mathbb{R}^n$ the following limit exists $$ \lim_{t\to\infty} \nabla F(vt) $$ (and belongs to $\mathbb{R}^n$)? If not, is there a reasonable assumption we can add to $F$ to guarantee that these limits exist?

We know that the curve $t \mapsto F(vt)$ is $C^1$ and strictly convex and so the limit of it's derivatives $$ \lim_{t\to\infty} \langle \nabla F(vt), v \rangle $$ exists in $\mathbb{R}$ but can we say anything about convergence for $\nabla F$ in general?

Thanks in advance for any help/suggestions.

$\endgroup$
0

1 Answer 1

5
$\begingroup$

Here is a counterexample to the convergence of $\nabla F$ in the conditions you give.

Let $f(x)=\sqrt{1+x^2}$, let $s(x)=\cos(\ln(\ln(x^2+10)))$ and let $\phi:\mathbb{R}\to[0,1]$ be some $C^\infty$ bump function with supp$(\phi)\subseteq[-1,1]$ and $\phi(0)=1,\phi'(0)=0$. Then, for a small enough constant $\varepsilon>0$, the function $g:\mathbb{R}^2\to\mathbb{R}$;

$$g(x,y)=f(x)+f(y+\varepsilon\phi(y)s(x))$$

has $\frac{\partial}{\partial y}g(x,0)=f'(\varepsilon s(x))$ (which does not converge when $x\to\infty$). But $g$ is convex, as we will see below by just computing its hessian. We just need to check that the Hessian is positive definite for $|y|\leq1$, as otherwise $g(x,y)=f(x)+f(y)$, which is strictly convex. As $|y|\leq1$, we have $y+\varepsilon\phi(y)s(x)\in[-2,2]$ if $\varepsilon\leq1$.

$$\frac{\partial}{\partial x}g(x,y)=f'(x)+f'(y+\varepsilon\phi(y)s(x))\varepsilon\phi(y)s'(x). $$ $$\frac{\partial}{\partial y}g(x,y)=f'(y+\varepsilon\phi(y)s(x))(1+\varepsilon s(x)\phi'(y)).$$

Note that $\frac{\partial}{\partial x}g(x,y)$ and $\frac{\partial}{\partial y}g(x,y)$ are bounded.

$$\frac{\partial^2}{\partial x^2}g(x,y)=f''(x)+f''(y+\varepsilon\phi(y)s(x))(\varepsilon\phi(y)s'(x))^2+f'(y+\varepsilon\phi(y)s(x))\varepsilon\phi(y)s''(x)$$ $$\frac{\partial^2}{\partial x\partial y}g(x,y)=f''(y+\varepsilon\phi(y)s(x))\varepsilon\phi(y)s'(x)(1+\varepsilon s(x)\phi'(y))+f'(y+\varepsilon\phi(y)s(x))\varepsilon s'(x)\phi'(y)$$ $$\frac{\partial^2}{\partial y^2}g(x,y)=f''(y+\varepsilon\phi(y)s(x))(1+\varepsilon s(x)\phi'(y))^2+f'(y+\varepsilon\phi(y)s(x))\varepsilon s(x)\phi''(y)$$

Now, for some constant $C$ we have $f''(x)=\frac{1}{(1+x^2)^{3/2}}>C|s''(x)|$ for all $x\in\mathbb{R}$. So by taking $\varepsilon$ to be small enough, we have $$\frac{\partial^2}{\partial x^2}g(x,y)>\frac{1}{2}f''(x)=\frac{1}{2(1+x^2)^{3/2}}.$$ Also, as $|y|\leq1$ implies $f''(y+\varepsilon\phi(y)s(x))>0.01$, taking $\varepsilon$ small enough we have $\frac{\partial^2}{\partial y^2}g(x,y)>0.001$.

Meanwhile, for some constant $K$, the term $\frac{\partial^2}{\partial x\partial y}g(x,y)$ is bounded in norm by $K\varepsilon s'(x)\leq\frac{2K\varepsilon}{\sqrt{x^2+10}}$. From these estimations it follows that the trace and determinant of the Hessian of $g$ are positive, so the Hessian of $g$ is positive definite, so $g$ is convex.

$\endgroup$
1
  • 1
    $\begingroup$ Thanks for finding this example! $\endgroup$ Commented May 30, 2024 at 12:16

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.