Continuous function of multivariable which has directional derivative nowhere

89 Views Asked by At

I want to contruct a continuous function of multivariable such that the directional derivative exists nowhere. Let $W(x)$ be the Weierstrass function, which is continuous everywhere but differentiable nowhere. We can also prove that the left and right derivatives of $W(x)$ exists nowhere. Let $f (x,y)=W (x)\cdot\sin y$, then the function $f(x,y)$ is continuous and does has directional derivative in all direction $(\cos\theta, \sin\theta)\neq(0, \pm1)$. In fact, let $$\rho=\sqrt{(\Delta x)^2+(\Delta y)^2},\quad \Delta x=\rho\cdot\cos\theta,\quad \Delta y=\rho\cdot\sin\theta,$$ then the limit \begin{eqnarray*} % \nonumber to remove numbering (before each equation) & & \lim_{\rho\to0}\frac{W(x+\Delta x)\cdot\sin(y+\Delta y)-W(x)\cdot\sin y}{\rho} \\ &=& \lim_{\rho\to0}\frac{(W(x+\Delta x)-W(x))\cdot\sin(y+\Delta y)}{\rho}+ \lim_{\rho\to0}\frac{(\sin(y+\Delta y)-\sin (y))\cdot W(x)}{\rho} \\ &=& \sin y\cdot\cos\theta\cdot\fbox{$\lim\limits_{\Delta x\to0}\dfrac{W(x+\Delta x)-W(x)}{\Delta x}$}+W(x)\cdot\cos y\cdot\sin\theta \end{eqnarray*} does not exist. However, the directional derivative on the direction $(0, \pm1)$ exists.

I guess the function $f(x,y)=W(x)+W (y)$ or $W(x)\cdot W(y)$ is continuous everywhere, and the directional derivative does not exist everywhere. In fact, let $$f(x,y)=W(x)+W (y),$$ then for $(\cos\theta,\sin\theta)\neq(\pm1,0),(0,\pm1)$(for the directions $(\pm1,0)$ and $(0,\pm1)$, the result is obvious), we have that \begin{eqnarray*} % \nonumber to remove numbering (before each equation) & & \lim_{\rho\to0}\frac{W(x+\Delta x)+W(y+\Delta y)-W(x)-W(y)}{\rho} \\ &=& \lim_{\rho\to0}\frac{W(x+\Delta x)-W(x)}{\Delta x}\cdot\frac{\Delta x}{\rho}+ \lim_{\rho\to0}\frac{W(y+\Delta y)-W(y)}{\Delta y}\cdot\frac{\Delta y}{\rho} \\ &=& \lim_{\rho\to0}\frac{W(x+\Delta x)-W(x)}{\Delta x}\cdot\cos\theta+ \lim_{\rho\to0}\frac{W(y+\Delta y)-W(y)}{\Delta y}\cdot\sin\theta \end{eqnarray*}

We know that both the limit $\lim\limits_{\rho\to0}\dfrac{W(x+\Delta x)-W(x)}{\Delta x}$ and $\lim\limits_{\rho\to0}\dfrac{W(y+\Delta y)-W(y)}{\Delta y}$ do not exist, but how to prove that the plus of the two limit does not exist?

1

There are 1 best solutions below

4
On BEST ANSWER

You simply cannot infere that the directional derivative does not exist in any direction by simply knowing that the partial derivatives in the $x$ and $y$ directions do not exist. For instance, if you take $$ g(x,y)=W(x-y), $$ then the derivatives in the directions $(1,0),(0,1)$ do not exist, but the derivative in the direction $(1/\sqrt 2,1/\sqrt 2)$ exists and it is zero (as $g$ is constant in that direction).

Moreover, note that if $W$ was an odd function, then by a similar approach the function $f$ you wrote would have a vanishing directional derivative in the direction $(1,-1)$ at the origin. Clearly $W$ is not an odd function, but what I mean is that in order to prove what you want you really need to use the definition of the function $W$, not simply saying that $W$ is not differentiable anywhere (at least this is what it looks like).

I don’t know if the example you wrote has the property you want, or in general how to construct an explicit function in many variables with non existing partial derivatives.


Edit: I could give you a hint on how I would prove it. As I said, you cannot simply start from the fact that $W$ is not differentiable at any point. What would help is a property like the following one: there exists $0<\alpha<1$ such that $$ |W(x)-W(y)|\leq C|x-y|^\alpha, $$ and for every $x$ there exists constant $c$ and a sequence $y_n\to x$ such that $$ |W(x)-W(y_n)|>c|x-y_n|^\alpha. $$ With the first formula you ask for $W$ to be $\alpha$-Hölder continuous (this is true for the Weierstrass functions), while the second formula essentially says the condition is sharp (it cannot be improved); I don’t know if the second condition holds for $W$, but maybe that is known.

With the hypotheses above, consider a direction $(\cos(\theta),\sin(\theta))$ (assume $0<\theta<\pi/4$ for simplicity). Then, you simply want to prove that the function $$ g(x)=W(x)+W(\tan(\theta)x) $$ is not differentiable at a given point $x$. You can see that $$ |g(y_n)-g(x)|\geq|W(y_n)-W(x)|-|W(\tan(\theta)y_n)-W(\tan(\theta)x)|\geq $$ $$ \geq (c -C\tan(\theta)^\alpha)|x-y_n|^\alpha, $$ so that at least for $\theta$ small enough, one has that for some constant $\epsilon$ $$ |g(y_n)-g(x)|\geq \epsilon |x-y_n|^\alpha, $$ which implies that $g$ is not differentiable at $x$.


Edit 2: thinking about it, I can say for sure that you first example does not satisfy the property you are looking for. The Weierstrass function $W$ is odd around $x=1/2$, that is, $W(x)=-W(1-x)$. This implies that in your first example, $f(x,1-x)=0$ for all $x$, so that the dervative in the direction $(1/\sqrt 2,-1/\sqrt 2)$ at the point, say, $(1/2,1/2)$ exists and it is zero. Maybe this does not happen for $f(x,y)=W(x)\cdot W(y)$, I don’t know… I’ll think about it.