So I know basically what a directional derivative is and how to calculate it using the gradient vector, but I'm a bit lost on the more advanced approach of looking at it as a linear transform.

I've read that multivariable calculus is about approximating nonlinear maps with linear ones, and I know that the Jacobian is the matrix associated with the directional derivative. However, I still don't really understand the directional derivative as a linear transform.

For example, what is the input of the transform? Also, what is meant by approximating nonlinear maps by linear maps? I only know a little linear algebra so that might be why I'm so confused.

  • $\begingroup$ The directional derivative is a linear transform of a direction if the gradient exists. Approximation = think in terms of Taylor decomposition where you get an approximation by neglecting the higher order terms (starting from the second derivative). $\endgroup$ – A.Γ. Aug 25 '15 at 21:07
  • $\begingroup$ Perhaps you mean the Gâteaux derivative? $\endgroup$ – Bernard Aug 25 '15 at 21:11

Given a function $f:\mathbb R^n \to \mathbb R^m$ and $h\in\mathbb R^n$. The directional derivative at $0$ in direction $h$ is defined by (if the limit exists) $$ f'(0; h) := \lim_{t\to 0} \frac{f(th) - f(0)}{t}. $$ The function, $h\mapsto f'(0; h)$ is called the Gâteaux differential or derivative. Notice that $f'(0;h)$ satisfies the linear approximation property in direction $h$: $$ \lim_{t\to 0} \frac{f(th) - f(0) - f'(0; th)}{t} = 0. $$

The function $f'(0; \cdot)$ is multiplicative by definition, that is $f'(0; th) = t f'(0; h)$. Some authors demand that $f'(0;\cdot)$ is also additive, that is $f'(0; h+k) = f'(0; h) + f'(0; k)$, thus linear. In that case we usually write $Df(0)$ or $df(0)$ for the linear map $f'(0; \cdot)$ and identify $Df(0)$ with its matrix representation, the Jacobi matrix.

If $f'(0; \cdot)$ is linear and satisfies following linear approximation property $$ \lim_{h\to 0} \frac{f(h) - f(0) - Df(0)h}{\|h\|} = 0, $$ then $Df(0)$ is called the Fréchet differential or derivative. $Df(0)$ is the unique linear map with that approximation property.


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.