Given a differentiable function, proove that the directional derivative in the direction V (in point P) is < gradient . V >

If the function is a plane, kx+qy, the increment in the direction (a,b) is ka+qb by simple algebra, which in this case is < gradient . direction > . Now, if a function is differentiable, that means that locally, it is like a plane, and so the directional derivatives behave like that.

Is that enough proof?

  • $\begingroup$ That's a proof idea, try to write it out algebraically $\endgroup$ – kjetil b halvorsen Jul 31 '17 at 19:24

$\textbf{Comment}$: For all that is about to follow, I have better notes on here derivative as a linear map (pg 4).

I can start you off with $f: \mathbb{R}^n \to \mathbb{R}^m$ where $m=1$ which is the directional derivative. You can generalize the argument from this special case to the standard derivative.

$$f \in C^1(U \subset \mathbb{R}^n,\mathbb{R}) \Rightarrow Df(p) \in L( \mathbb{R}^n, \mathbb{R})$$

i.e $Df(p)$ is a linear map from $\mathbb{R}^n$ to $\mathbb{R}$. Given $(x_1,...,x_n)$ are the coordinate functions on $U$ and $\alpha:(- \epsilon, \epsilon) \to U$ is a differentiable curve (with $\alpha(0) = p$) then it is natural to have,

\begin{align} Df(p) \cdot \alpha'(0)&= \frac{d}{dt} \Bigr|_{t = 0} (f \circ \alpha)(t) \\ \\ & = \nabla f(p) \cdot \alpha'(0) \\ \\ & \Rightarrow Df(p) = \nabla f(p) \end{align}

The reason why this is natural to want is that $f$ is differentiable at $p$ i.e by the inverse function theorem, it is a local diffeomorphism i.e it maps curves to curves.

Hence, the curve $\alpha$ gets mapped to the curve $f \circ \alpha$ and so you would want the range vector to $\alpha$ at $0$ to be the tangent vector to $f \circ \alpha$ at $0$. We now recall the definition of differentiability. We say $g: U \subset \mathbb{R}^n \to V \subset \mathbb{R}^m$ is differentiable at $p \in U$ if there exists a linear map $\lambda: \mathbb{R}^n \to \mathbb{R}^m$ such that the following holds,

$$ \lim_{x \to p} \frac {\|f(x) - f(p) - \lambda(x-p)\|_m}{\|x-p\|_n} = 0$$

You can check that such a $\lambda$ must be unique. Here $x-p$ is just a tangent vector to $p$. Hence, you can just check that,

$$ \nabla f(p) \cdot \alpha'(0) = \sum_{j=1}^{n} \frac{\partial f}{\partial x_j}(p) \cdot \alpha_j'(0), \textbf{where} \ \ \alpha(t) = \begin{pmatrix} \alpha_1(t) \\ \vdots \\ \alpha_n(t) \end{pmatrix}$$

is $\lambda$. This is shown in any multivariable calculus book as the linearization of a function. They will typically write,

\begin{align} z(x,y)& \approx f_x(a,b)(x-a) + f_y(a,b)(y-b) + f(a,b) \\ \\ & = \nabla f(a,b) \cdot \begin{pmatrix} x-a \\ y-b \end{pmatrix} + f(a,b)\end{align}

This argument is done for $f: \mathbb{R}^2 \to \mathbb{R}$. Here $z(x,y)$ is an approximation of $f(x,y)$ for $(x,y)$ near $(a,b)$. Therefore we have the following,

$$f(x,y) - f(a,b) - \nabla f(a,b) \cdot\begin{pmatrix} x-a \\ y-b \end{pmatrix} = \epsilon(x,y)$$

If you accept this special case to be true, you see the extension almost immediately. This time you get a matrix for $Df(p)$. Well you had one last time, it was just $1 \times n$ or $n \times 1$. So similarly you guess that $Df(p) = \nabla f(p)$ as above, and now when checking, you'll see it amounts to this special case in each entry. Applying the triangle inequality (inductively) you get the result.


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.