Let $f:D\to\mathbb{R}$ where $D:=\{(x,y):x^2+y^2\leq 1\}$. Suppose all partial derivatives of $f$ are continuous and $|f|\leq 1$. Show that there exists $(a,b)\in Int_D$ s.t.:

$$\left[\frac{\partial f}{\partial x}(a,b)\right]^2+\left[\frac{\partial f}{\partial y}(a,b)\right]^2<4$$

Here $Int_D$ means the interior of $D$, that is , the open unit disk.

I try to reduce to the one dimension case, that is: there exists $a\in(-\sqrt{1-y^2},\sqrt{1-y^2})$ s.t $\left[\frac{\partial f}{\partial x}(a,y)\right]^2\leq 2$.


2 Answers 2


Proof by contradiction Assume $||\nabla f||\geq 2$ for each point in $D$ and define a vector field $v$ on $D$ such that $$\frac{dx}{dt}=\nabla f$$ Consider the solution curve $\gamma(t)$ passing through the origin at time 0. Since $||\nabla f||>0$ for each point in $D$, the curve cannot terminate in the region $D$. Then the arc length of $\gamma(t)$ from $t=t_1<0$ to $t=t_2>0$ must be no less than 2 (the diameter of $D$), where $\gamma(t_1),\gamma(t_2)\in\partial D$ belongs to the boundary of $D$. Then we have $$f(\gamma(t_2))-f(\gamma(t_1))=\int_{\gamma(t_1)}^{\gamma(t_2)}df=\int_{\gamma(t_1)}^{\gamma(t_2)}\nabla f\cdot dx=\int_{\gamma(t_1)}^{\gamma(t_2)}||\nabla f||~dx\geq4$$ On the other hand, $$f(\gamma(t_2))-f(\gamma(t_1))\leq||f(\gamma(t_2))-f(\gamma(t_1))||\leq||f(\gamma(t_2))||+||f(\gamma(t_1))||\leq 2$$ which leads to contradiction. Hence there must be some point $x\in D$ such that $||\nabla f(x)||^2<4$.


Suppose $\|\nabla f(z)\|\ge2$ for all $z\in D$. Note that $f$ cannot attain its minimum in the interior of $D$, since $\nabla f(z)\ne0$ for all $z\in D$. Hence $\epsilon := \frac14 (f(0)+1)\in(0,\frac12]$. We will reach a contradiction by following the direction field $\nabla f(z)$ starting from $z_0=0$.

Proof 1: (Like Shuchang's argument.) Suppose we can solve the ODE system $\dot z(t)=\nabla f(z(t))$, $z(0)=0$, on some maximal time interval $[0,T)$, with either $T=\infty$ or $\|z(t)\|\to1$ as $t\to T$. Then for any $t\in[0,T)$, since $\nabla f(z)\dot z = \|\nabla f(z)\|^2$, $$ 2-4\epsilon \ge f(z(t))-f(0) = \int_0^t \|\nabla f(z(s))\|^2\,ds \ge 2\int_0^t \|\dot z(s)\|\,ds \ge 2\|z(t)\|, $$ whence $\|z(t)\|\le 1-2\epsilon<1$. We infer $T=\infty$. However, $$ f(z(t))-f(0) = \int_0^t \|\nabla f(z(s))\|^2\,ds \ge 4t >2 $$ for $t>\frac12$, which yields a contradiction.

Proof 2: We can avoid invoking ODE existence and continuation theory by using a discrete argument, choosing a sufficiently small $h>0$ and defining $$ z_{j+1} = z_{j} + h\nabla f(z_{j}) $$ for $j=0,1,\ldots$, as long as $z_{j}\in D$.
We will find that the hypotheses imply that $|z_j|\le 1-\epsilon$ for all $j>0$ but also $f(z_j)-f(0)\ge 2hj$, which contradicts the bound $|f|\le1$ when $j>1/h$.

The gradient is uniformly continuous in $D$, hence there exists $\delta>0$ such that $\|\hat z\|<\delta$ implies $$ \| \nabla f(z+\hat z)-\nabla f(z)\| <\epsilon. $$ Let $h=\min(\frac12, \delta/M, \epsilon/M)$ where $M=\sup_D \|\nabla f(z)\|$. Note that as long as $\|z_j\|\le 1-\epsilon$ we have $$ 2h \le \| h\nabla f(z_j)\| = \|z_{j+1}-z_j\| \le hM\le \min(\delta, \epsilon), $$ hence $\|z_{j+1}\|\le1$. Now we can write $$ f(z_{j+1})-f(z_j) = \int_0^1 \nabla f(z_j+s(z_{j+1}-z_j)) (z_{j+1}-z_j)\,ds = \nabla f(z_j)(z_{j+1}-z_j) + E_j, $$ where $$ E_j = \int_0^1 (\nabla f(z_j+s(z_{j+1}-z_j))-\nabla f(z_j)) (z_{j+1}-z_j)\,ds. $$ This satisfies the bound $$ | E_j |\le \epsilon \|z_{j+1}-z_j\|. $$

Consequently, since $\nabla f(z_j)(z_{j+1}-z_j)=\|\nabla f(z_j)\|\|z_{j+1}-z_j\|\ge 2\|z_{j+1}-z_{j}\|$, $$ f(z_{j+1})-f(z_j) \ge \|z_{j+1}-z_j\| (2-\epsilon ) \ge 2h. $$ Summing, however (recall $z_0=0$), we infer that $$ \|z_{j+1}\| \le \sum_{k=0}^{j} \|z_{k+1}-z_k\| \le \frac{ f(z_{j+1})-f(0)}{2-\epsilon } \le \frac{ 2- 4\epsilon}{2-\epsilon }\le 1-\epsilon. $$ Thus $\|z_j\|\le 1-\epsilon$ for all $j$. However, $$ f(z_{j+1})-f(0) = \sum_{k=0}^{j} (f(z_{k+1})-f(z_k)) \ge 2h(j+1) $$ and this contradicts the bound $|f|\le 1$ for large $j$.


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.