0
$\begingroup$

I am struggling with the following derivative. Let $\pmb{x} \in \mathbb{R}^{n}$ be a vector, $\pmb{y} \in \mathbb{R}^{m}$ another vector that is a function of $\pmb{x}$, and $\pmb{g}$ and $\pmb{h}$ two functions returning vectors. I aim to obtain the following derivative:

$$ \frac{\partial}{\partial \pmb{x}} \left( \pmb{g}(\pmb{y})^{\top} \otimes \pmb{I} \right)\pmb{h}(\pmb{x}). $$ My thoughts were that this would involve a combination of the sum and chain rule. E.g., first take the sum-rule with the left and right-hand part. But then $\left( \pmb{g}(\pmb{y})^{\top} \otimes \pmb{I} \right)$ takes the form of a matrix and I am quite lost. I expect that taking a differential form is needed, but I don't really understand how to do these properly yet.

$\endgroup$
2
$\begingroup$

Define some new matrices $$\eqalign{ H &= {\rm Mat}(h) \implies h = {\rm vec}(H) \cr J &= \frac{\partial h}{\partial x},\quad K = \frac{\partial g}{\partial y},\quad L = \frac{\partial y}{\partial x} \cr }$$ Write the vector function and find its differential and gradient in terms of these new variables. $$\eqalign{ w &= (g^T\otimes I)h = Hg \cr dw &= (g^T\otimes I)\,dh + H\,dg \cr &= \Big((g^T\otimes I)J + HKL\Big)\,dx \cr \frac{\partial w}{\partial x} &= (g^T\otimes I)J + HKL \cr }$$ Knowing nothing about the nature of the vector functions $(g,h,y)$ this is as far as we can go.

$\endgroup$
  • $\begingroup$ Thank you so much for this answer and other answers you have given! I am writing a paper on maximum likelihood estimation on some multivariate models and this has all been really helpful for my work! $\endgroup$ – Sacha Epskamp Feb 17 at 20:02

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.