22
$\begingroup$

Suppose that $V:\mathbb{R}^n \to \mathbb{R}$ is a smooth function such that $\nabla V : \mathbb{R}^n \to \mathbb{R}^n$ has no equilibria (i.e. $\forall x \in \mathbb{R}^n : \nabla V (x) \not = 0$). Under these hypotheses, is it possible that $\nabla V (x)$ can point in every direction?

To be more precise, under the above hypotheses the map $$\mathbb{R}^n \to \mathbb{S}^{n-1} $$ $$x \mapsto \frac{\nabla V(x)}{\|\nabla V (x)\|} $$ is well-defined. Is it impossible for such a map to be surjective? If not, what is a counterexample?

$\endgroup$
  • 3
    $\begingroup$ I believe this is possible from $R^2$ onwards. Can you write down a function for which one gradient flow line makes an $S$ shape? Can you make sure that at some distance from this $S$ the flowlines it straighten to vertical straight lines? If you can do that you can mirror this construction and obtain the right function. $\endgroup$ – Thomas Rot Mar 15 '16 at 14:18
  • $\begingroup$ That's an interesting idea. It seems plausible, but I'm having some trouble actually writing down such a function. $\endgroup$ – Matthew Kvalheim Mar 15 '16 at 16:55
  • $\begingroup$ The more I think about it, I'm not sure a function satisfying the requirements @ThomasRot suggested exists. $\endgroup$ – Matthew Kvalheim Mar 15 '16 at 23:57
  • $\begingroup$ A cool question! Also, it might be received well in MathOverflow, but given that you started a bounty I would think it better to let the bounty period expire before we consider migration. If you are in a hurry, respond to this and/or flag again. $\endgroup$ – Jyrki Lahtonen Mar 19 '16 at 6:53
12
+50
$\begingroup$

A quick solution for $n=2$, and an explanation of where it came from: $$\nabla (e^x \sin y) = (e^x \sin y, e^x \cos y).$$ The right hand side is never zero, but does assume every nonzero value in $\mathbb{R}^2$.

Motivation: If $f: \mathbb{C} \to \mathbb{C}$ is holomorphic, then $\nabla(\mathrm{Re}(f)) = (\mathrm{Re}(f'), -\mathrm{Im}(f'))$. I looked for an $f'$ (namely $e^z$) which takes $\mathbb{C}$ onto $\mathbb{C}_{\neq 0}$, and then integrated it to find $f$.


Inspired by this, let's try $$\nabla(e^{x_0} \cos(x_1^2+x_2^2+\cdots + x_n^2)) =$$ $$e^{x_0} (\cos(x_1^2+\cdots + x_n^2), - x_1 \sin(x_1^2+\cdots + x_n^2), \cdots, -x_n \sin(x_1^2+\cdots + x_n^2)).$$

First, we note that the gradient is not zero. The first coordinate only vanishes if $x_1^2+ \cdots + x_n^2$ is of the form $(2k+1) \pi$. But, in this case, at least one of $x_1$, $x_2$, ..., $x_n$ are nonzero; say $x_j$. Then $- x_j \sin(x_1^2+\cdots + x_n^2) = \pm x_j \neq 0$.

Now, let's take a nonzero vector $(v_0, \ldots, v_n)$. We need to go through cases.

If $v_0 = 0$, choose $(x_1, \ldots, x_n)$ proportional to $(v_1, \ldots, v_n)$ and such that $\sin(x_1^2+\cdots+x_n^2)$ has the right sign.

If $v_1 = \cdots = v_n = 0$ and $(-1)^k v_0>0$, take $x_1^2+\cdots+x_n^2 = k \pi$.

If neither of those cases holds, we'll take $(x_1, \ldots, x_n)$ of the form $(t v_1, \ldots, t v_n)$ for some $t$ to be determined soon. Set $s = x_1^2+ \cdots + x_n^2$. Then our vector points in direction $\pm (- \cot(t^2 s), v_1, v_2, \ldots, v_n)$. Since cotangent is surjective, we can choose $t$ such that $\cot (t^2 s) = v_0$, and we can get the sign right.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you for the insightful answer. Your counterexample for $\mathbb{R}^2$ and the motivation you gave are quite helpful. $\endgroup$ – Matthew Kvalheim Mar 24 '16 at 22:31
8
$\begingroup$

It is possible to have a vector field without equilibrium point and takes every possible directions in $\mathbb{R^n}$ for any $n \ge 2$.

The proof for the 2-dimension case is given below. From that, vector fields for higher dimension cases can be constructed.


Let $\rho : (-\epsilon,\infty) \to \mathbb{R}$ be any $C^\infty$ function such that

  1. $\rho(t)$ is even over $(-\epsilon,\epsilon)$ and $\rho(0) = \rho'(0) = 0$.
  2. $|r\rho'(r)| < 1$ for all $r \ge 0$.
  3. for some $r_c > 0$, $\rho(r_c) = 2\pi$.
  4. for some $r_m > 0$, $N \in \mathbb{Z}$, $\rho(r) = 2\pi N$ for all $r \ge r_m$.

For any point $(x,y) \in \mathbb{R}^2$, let $(r,\theta)$ be the corresponding polar coordinates.
Let $\hat{e}_x, \hat{e}_y, \hat{e}_r, \hat{e}_\theta$ be unit vectors along the $x$, $y$, $r$ and $\theta$ directions respectively.

Consider following function $$U(x,y) = r\cos(\theta - \rho(r)) = x\cos\rho(r) + y\sin\rho(r)$$ Its gradient is given by:

$$\begin{align} \vec{\nabla} U &= \hat{e}_x \left( \cos\rho + \frac{x\rho'}{r}( -x \sin\rho + y\cos\rho )\right) + \hat{e}_y \left( \sin\rho + \frac{y\rho'}{r}( -x \sin\rho + y\cos\rho )\right)\\ &= \hat{e}_r\left(\cos(\theta - \rho) + r\rho'\sin(\theta-\rho)\right) -\hat{e}_\theta\left(\sin(\theta-\rho)\right) \end{align} $$ From these two expressions, it is easy to deduce

  • $\vec{\nabla} U$ is well defined at $(0,0)$ because $\rho'(0) = 0$.
  • $\vec{\nabla} U(0,0) = \hat{e}_x$ because $\rho(0) = \rho'(0) = 0$.
  • $\vec{\nabla} U \ne 0$ for $r > 0$ because $$\begin{align} |\vec{\nabla} U| &= |\hat{e}_r\left(\cos(\theta - \rho) + r\rho'\sin(\theta-\rho)\right) -\hat{e}_\theta\left(\sin(\theta-\rho)\right)|\\ &\ge |\hat{e}_r\left(\cos(\theta - \rho)\right) -\hat{e}_\theta\left(\sin(\theta-\rho)\right)| - |\hat{e}_e \left(r\rho'\sin(\theta-\rho)\right)|\\ &\ge 1 - |r\rho'| > 0 \end{align} $$

Combine these, we find $\nabla U \ne 0$ over $\mathbb{R}^2$. This allow us to define following unit vector field globally on $\mathbb{R}^2$.

$$\hat{u}(x,y) = \frac{\vec{\nabla}U(x,y)}{\left|\vec{\nabla}U(x,y)\right|}$$

Over the curve $\displaystyle\;\gamma : [0,r_c] \ni t \quad\mapsto\quad (t\cos\rho(t),t\sin\rho(t)) \in \mathbb{R}^2\;$, we have $$\hat{u}(t) = \hat{e}_r = \hat{e}_x \cos\rho(t) + \hat{e}_y\sin\rho(t)$$ As we move from $\gamma(0)$ to $\gamma(r_c)$ along $\gamma$, $\hat{u}(\gamma(t))$ will cover the whole $S^1$ at least once. The map $\hat{u}$ is surjective on $\gamma$ and hence on $B(0,r_c)$ and $\mathbb{R}^2$. This means $U$ is a counter-example of $V$ for the 2-dimenisonal case.

As an illustration, following is a picture of what $U(x,y)$ will typically look like under this construction.

$\hspace0.75in$ A swirling potential

This particular $U(x,y)$ is computed using

$$\rho(t) = 2\pi + (f(t)-2\pi)g(t) \quad\text{ where }\quad \begin{align} f(t) &= \log\sqrt{(e^{4\pi}-1)t^2 + 1}\\ g(t) &= \begin{cases} 1, & t \le 1\\ 1 - e^{-1/(e^{t-1} -1)}, & t > 1\end{cases} \end{align} $$ and plotted over the region $|x|,|y| \le 2$. This $\rho(t)$ satisfies the first three requirement in the beginning of this answer with $r_c = 1$. For $r \gg r_c $, $\rho(t)$ tends to $2\pi$ exponentially as $t \to \infty$.

Back to the case of higher dimensions. Let's say we do have a $\rho$ that satisfies all four conditions. Consider following function:

$$V(x_1,x_2,\ldots,x_n) \stackrel{def}{=} U\left( x_1, \sqrt{x_2^2 + x_3^2 + \cdots + x_n^2} - (r_m+1)\right)$$

Since $B(0,r_m) \supset B(0,r_c)$, the map $\frac{\vec{\nabla}V}{|\vec{\nabla}V|}$ will be surjective over the hyper-torus $$ \bigg\{ (x_1,\ldots,x_n) : x_1^2 + \left(\sqrt{x_2^2 + x_3^2 + \cdots + x_n^2} - (r_m+1)\right)^2 \le r_m^2 \bigg\}$$ Since $U(x,y) = x$ whenever $x^2 + y^2 \ge r_m^2$, $V(x_1,\ldots,x_n)$ equals to $x_1$ over the complement of the hyper-torus. As that contains the $x_1$-axis, the square root in the definition of $V$ won't cause any problem. This makes $V$ smooth everywhere. Finally, it is easy to check $\nabla V$ never vanishes.

Update

About the question how this particular form of $\rho(t)$ is chosen.

First, $\rho(r)$ cannot change too rapidly or $\nabla U$ may become zero somewhere.
Second, we need $\rho(r)$ span across around a large enough interval to make $\hat{u}(r)$ surjective.
This suggest us to choose a $\rho(r)$ with $\rho'(r)$ as close to $\frac{1}{r}$ as possible. This means $$\rho(r) \sim \log r + constant.$$ In order for $U(x,y)$ to be smooth at the origin, I replace $\log r$ by $\log\sqrt{(e^{4\pi}-1)r^2 + 1}$. The constants are chosen so that when $r$ changes from $0$ to $1$, $\rho(r)$ pick up a change of $2\pi$ which we need.

Finally, we want $\rho(r)$ decay back to a constant multiple of $2\pi$ for large $r$ while keeping smooth all the way. We need this in order to extend the result to higher dimensions. This can be done using a smooth Partition of unity. The function $g(r)$ in the formula doesn't really have any specific meaning. It is simply something from trial and error which works.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ May I ask how you came up with the particular example of $\rho$ you gave? I'd like to understand the intuition behind your thought process. $\endgroup$ – Matthew Kvalheim Mar 21 '16 at 21:43
  • $\begingroup$ @MatthewKvalheim answer updated, see rationale of chosing such a $\rho$. $\endgroup$ – achille hui Mar 21 '16 at 22:25
  • $\begingroup$ You wrote an expression for $\nabla U$ that involves division by $r$. Am I correct in the interpretation that you actually mean that $\nabla U$ is defined by the expression you wrote on $\mathbb{R}^2\setminus \{0\}$, and $\hat{e}_x$ at the origin? $\endgroup$ – Matthew Kvalheim Mar 22 '16 at 3:49
  • $\begingroup$ Thank you very much for your help. $\endgroup$ – Matthew Kvalheim Mar 22 '16 at 3:58
  • $\begingroup$ @MatthewKvalheim $\nabla U$ is defined as the gradient of $U$. What I mean is it "equals to" the expression involves division by $r$ for $r \ne 0$ and $\hat{e}_x$. $\endgroup$ – achille hui Mar 22 '16 at 9:15

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.