0
$\begingroup$

Is there a property of gradient that allow me to compute $\partial f/\partial x$ with $f=f(x, y, z)$ and $y$ and $z$ are functions of $x$.Can we write that $\partial f/\partial x = \partial f/\partial y * \partial y / \partial x + \partial f/\partial z *\partial z / \partial x$ or something like that. I came upon this question when trying to understand the backpropagation equation of recurrent neural network and I don't understand while it sum all the partial gradient through time.

$\endgroup$
0
$\begingroup$

If $f$ is a function of $(x,y,z)$ and $(y,z)$ are themselves functions of $x$, then $f$ can be rewritten so that it is a function of $x$ alone. Let's call the single-variable function $\phi(x)$.

So there are two distinct quantities which one might call the derivative (gradient) of the function wrt $x$ $$\eqalign{ &\frac{\partial f}{\partial x} \cr {\rm -or-} \cr &\frac{d\phi}{dx} &= \frac{\partial f}{\partial x} + \frac{\partial f}{\partial y} \frac{dy}{dx} + \frac{\partial f}{\partial z} \frac{dz}{dx} }$$ where $\partial$ has been used for the partial derivative of multi-variable functions, and $d$ for the derivative of a single-variable function.

$\endgroup$

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.