10
$\begingroup$

Suppose we have a smooth function $f$ from $\mathbb R^n\to\mathbb R$ such that $\nabla f(0)=0$, and we want to check if $f$ has a local maximum at $0$ (as opposed to a local min or a saddle point).

For any vector $v$ in $\mathbb R^n$, we can form a function $g_v:\mathbb R\to\mathbb R$ defined by $g_v(x)=f(vx)$. Intuitively, we are looking at the behavior of $f$ along the $v$ direction. If $f$ has a local max at zero, then of course $g_v$ also has a local max at zero. My question is, is the converse true? If the single-variable function $g_v$ has a local max at $0$ for every possible direction $v$, does $f$ also have a local max at $0$?

If the Hessian of $f$ at zero has full rank, then this question is easily answered as yes. By taking $v$ to be in turn each of the eigenvectors of the Hessian, the fact that $g_v$ has a local max means that the corresponding eigenvalues are all negative, and so $f$ has a local max at $0$. But if some of the eigenvalues are $0$, then I'm not sure how to analyze it.

As a first step in a proof, I might think: well, for any fixed direction, there is some radius $r$ such that as long as $v$ is within distance $r$ of the origin, then $f(v)\le f(0)$. But since the $r$ depends on the direction, we can't necessarily find a single $r$ which works for all directions, so that might allow for a counter-example.

As a bonus, if it does hold for smooth functions, what about functions that are merely differentiable? Continuous? Or even all functions?

$\endgroup$
1
  • $\begingroup$ I just had the same question myself. We can possibly apply compactness to get a minimum $r$, as long as we can choose $r$ as a continuous function of the direction, but I don't quite see a way to make this work. $\endgroup$ Jan 3, 2019 at 18:48

1 Answer 1

3
+50
$\begingroup$

After thinking about it for a while I realized that the answer to this question is no, even for polynomial functions. Consider the following function, defined in polar coordinates: $$f(r,\theta)=r^4(r^2-\cos^2\theta)(r^2-2\cos^2\theta)$$ For any fixed value of $\theta$, we have a local min at $0$: if $\cos\theta=0$, then we just get $g(r)=r^8$, while for $\cos\theta\ne 0$, we get $g(r)=2r^4\cos^4\theta-3r^6\cos^2\theta+r^8$. The leading term for small $r$ is $2r^4\cos^4\theta$, which also has a local min at $r=0$.

However, can find arbitrarily small $r$ with $f(r,\theta)<0$. Given any (small) $r$, choose $\theta$ so that $\frac{r}{\sqrt 2}<\cos\theta<r$.

If we switch to rectangular coordinates, $x=r\cos\theta$, $y=r\sin\theta$, we get the degree $8$ polynomial: $$2x^4-3x^6+x^8-6x^4y^2+4x^6y^2-3x^2y^4+6x^4y^4+4x^2y^6+y^8$$ So this function is, in fact, a polynomial, the nicest type of function around!

$\endgroup$
3
  • 1
    $\begingroup$ Neat construction! To visualize what's going on, we can factor $f(x,y)$ as $(x^2+y^2+x)(x^2+y^2-x)(x^2+y^2+\sqrt2 x)(x^2+y^2-\sqrt2 x)$: the product of four circles passing through $(0,0)$, two small ones inside two larger ones. For this product to be negative, $(x,y)$ must be inside a large circle but outside a small one; here is a plot of this region. In any direction from $0$, we must go through a positive region before reaching a negative one. $\endgroup$ Jan 3, 2019 at 21:12
  • 1
    $\begingroup$ I guess on that basis, we could take half the picture for another counterexample: say, $f(x,y) = (x^2+y^2-x)(x^2+y^2-2x)$ would also work. $\endgroup$ Jan 3, 2019 at 21:18
  • $\begingroup$ @MishaLavrov Ah, good thinking! Looking at the regions where $f$ is positive and negative is a nice way to visualize it. $\endgroup$
    – Carmeister
    Jan 3, 2019 at 22:59

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.