# If all directional derivatives exist at $0\in \mathbb R^2$, why the gradient doesn't exist?

Let $f:\mathbb R^2\longrightarrow \mathbb R$ a function s.t. all directional derivative exist at $0$. Why the gradient doesn't exist ?

Don't we have that $\nabla f(0)=(\frac{\partial f}{\partial x}(0),\frac{\partial f}{\partial y}(0))$ ? Indeed, $$\frac{\partial f}{\partial x}(0)=\nabla f(0)\cdot (1,0)\quad \text{and}\quad \frac{\partial f}{\partial y}(0)=\nabla f(0)\cdot (0,1)$$ so the gradient should exist, no ?

• Is it possible you're conflating "$\nabla f(0)$ exists" with "$f$ is differentiable at $0$"? – Andrew D. Hwang Mar 12 '17 at 10:52
• @AndrewD.Hwang: How do you distinguish between those? (Are you assuming a coordinate-dependent definition of $\nabla f$?) – hmakholm left over Monica Mar 12 '17 at 10:58
• @HenningMakholm: Yes, I'm assuming OP means "$\nabla f(0)$ is the gradient vector in Cartesian coordinates at the origin". – Andrew D. Hwang Mar 12 '17 at 11:01
• Giving the definition of the gradient would help. – mvw Mar 12 '17 at 11:08
• @AndrewD.Hwang: Hmm, I consider a gradient to be "the thing whose inner product with every tangent vector is a directional derivative". I can see this is not exactly equivalent to $f$ being differentiable, but it's not clear the distinction is relevant for this question. – hmakholm left over Monica Mar 12 '17 at 11:08

As a counterexample, consider the function defined by $$f(r\cos\theta, r\sin\theta) = r\sin(3 \theta)$$ (in particular, $f(0,0)=0$ independently of which $\theta$ you choose).

This has all directional derivatives at $(0,0)$ -- actually $f$ restricted to any straight line through $0$ is linear -- but it cannot be approximated by a linear function near $0$. Therefore it doesn't have a gradient.

Your confusion may be that you think that the gradient is merely "whatever we get by making a vector out of the partial derivatives". However, a more abstract definition (which appears to be what your course is using) is

The gradient is the vector $w$ such that the directional derivative $D_v(f)$ equals $v\cdot w$ for all $v$.

We can easily see that if such a $w$ exists then its elements must be the partial derivatives -- but it is not guaranteed that $(\frac{\partial f}{\partial x},\frac{\partial f}{\partial y})$ has this property, and if it doesn't then it isn't a gradient.

In the example above, the partial derivatives are $(0,-1)$ -- but that doesn't work as $w$, because the directional derivative in the direction $(\cos \frac\pi6,\sin\frac\pi 6)$ is $1$, which does not equal $(0,-1)\cdot(\cos\frac\pi6,\sin\frac\pi6)=-\frac12$.

• I don't really understand your example. How can I get the gradient ? In my course they took as an example $u(x,y)=x^2\boldsymbol 1_{\mathbb Q}(x)(x)+y^2\boldsymbol 1_{\mathbb Q}(y)$ – user330587 Mar 12 '17 at 11:00
• but in this example, the gradient existe since $\lim_{t\to 0}\frac{u(t,0)-u(0,0)}{t}=0$ and $\lim_{t\to 0}\frac{u(0,t)-u(0,0)}{t}=0$. Therefore $\nabla f(0)=(0,0)$ – user330587 Mar 12 '17 at 11:12
• @user330587: I don't understand the notation in your first comment. Should $(x)(x)$ have been just $(x)$? In that case it is indeed not a counterexample -- $u$ is perfectly differentiable at $(0,0)$ since we can squeeze it between $0$ and $x^2+y^2$. – hmakholm left over Monica Mar 12 '17 at 11:26
• @user330587: I have edited the answer with more explanation. Does that help? – hmakholm left over Monica Mar 12 '17 at 11:33