3
$\begingroup$

I have $\Sigma$ a symmetric $2 \times 2$ matrix, and $\Sigma^{-1}$ is its inverse.

Now, $\tilde{\Sigma}^{-1}=\Sigma^{-1} \otimes I_{n \times n}$ (Kronecker product).

I have a function $Y=f(\tilde{\Sigma}^{-1})$ that gives a value in $\mathbb R$.

Let's define $\Phi_{\Sigma}=vech(\Sigma)$

Now, I am trying to get $\frac{\partial Y}{\partial \Phi_{\Sigma}}$.

So far I have

$\frac{\partial Y}{\partial \Phi_{\Sigma}^T} = \Bigg( \frac{\partial vec(\tilde{\Sigma}^{-1})}{\partial \Phi_{\Sigma}^T} \Bigg)^T \Bigg( \frac{\partial vec(Y)}{\partial vec(\tilde{\Sigma}^{-1})^T} \Bigg)$

I've been working on this an got that $\Bigg( \frac{\partial vec(Y)}{\partial vec(\tilde{\Sigma}^{-1})^T} \Bigg)$ is a vector with $n \times n$ elements. Now, working with the first part of the derivative

$\Bigg( \frac{\partial vec(\tilde{\Sigma}^{-1})}{\partial \Phi_{\Sigma}^T} \Bigg)^T = \Bigg( \frac{\partial vec(\tilde{\Sigma}^{-1})}{\partial vec(\Sigma)^T} D_2 \Bigg)^T = \Bigg( vec \Big( \frac{\partial \tilde{\Sigma}^{-1}}{\partial \Sigma}\Big) D_2 \Bigg)^T = \Bigg( vec \Big( \frac{\partial \tilde{\Sigma}^{-1}}{\partial \Sigma^{-1}} \frac{\partial \Sigma^{-1}}{\partial \Sigma} \Big) D_2 \Bigg)^T$
$= \Bigg( vec \Big( (I_2 \otimes I_n) (-\Sigma^{-1} \Sigma^{-1}) \Big) D_2 \Bigg)^T$

where $D_2$ is the duplication matrix

However, the matrices $(I_2 \otimes I_n)$ and $-\Sigma^{-1} \Sigma^{-1}$ are not conformable. So it is wrong. Also, since $\Bigg( \frac{\partial vec(Y)}{\partial vec(\tilde{\Sigma}^{-1})^T} \Bigg)$ is a vector with $n \times n$ elements, and $\frac{\partial Y}{\partial \Phi_{\Sigma}}$ is $3 \times 1$, so $\Bigg( \frac{\partial vec(\tilde{\Sigma}^{-1})}{\partial \Phi_{\Sigma}^T} \Bigg)^T$ should be $3 \times (n \times n)$. May I ask for advice on solving this task?

$\endgroup$
1
  • $\begingroup$ Use \mathbb R for $\mathbb R$. $\endgroup$
    – Invisible
    Apr 4, 2020 at 11:33

1 Answer 1

2
$\begingroup$

For ease of typing, define $$\eqalign{ &M = \Sigma,\quad &N = \Sigma^{-1} \\ &R = M\otimes I,\quad &S = N\otimes I = R^{-1},\quad &f = f(S) \\ &h = {\rm vech}(M),\quad &v = {\rm vec}(M) \\ &D = D_2,\quad &v = Dh \\ }$$ You don't tell us anything about the function $f(S),\,$ so I'll assume you don't need help
calculating its gradient $G=\left(\frac{\partial f}{\partial S}\right)$

Before we begin, we need a few results from Wikipedia and this post which can be summarized $$\eqalign{ &A\in{\mathbb R}^{m\times n},\quad B\in{\mathbb R}^{p\times q} \\ &I_k\in{\mathbb R}^{k\times k}\qquad \big({\rm Identity\,Matrix}\big) \\ &a = {\rm vec}(A),\quad b={\rm vec}(B)\\ &x={\rm vec}(A^T) = K_{m,n}\,a\quad \big({\rm Commutation\,Matrix}\big) \\ &{\rm vec}(A\otimes B) = \left(I_n\otimes K_{q,m}\otimes I_p\right)(I_m\otimes I_n\otimes b)\,a \\ }$$ Using this, we can write $$\eqalign{ {\rm vec}(R) &= {\rm vec}\big(M\otimes I_n\big) \\ &= \Big(I_2\otimes K_{n,2}\otimes I_n\Big) \Big(I_2\otimes I_2\otimes{\rm vec}(I_n)\Big)\,v \\ &= Qv \\ }$$ Start by writing the differential of the function, then perform a sequence of changes of variables from $S\to R\to v\to h$. $$\eqalign{ df &= G:dS \\&= G:(-S\,dR\,S) \\&= -SGS:dR \\ &= -\operatorname{vec}\left(SGS\right):Q\,dv \\ &= -Q^T\operatorname{vec}\left(SGS\right):dv \\ &= -Q^T\operatorname{vec}\left(SGS\right):D\,dh \\ &= -D^TQ^T\operatorname{vec}\left(SGS\right):dh \\ \frac{\partial f}{\partial h} &= -D^TQ^T\operatorname{vec}\left(SGS\right) \\ \\ }$$ The trace/Frobenius product $\;A:B = {\rm Tr}\big(A^TB\big)\;$ is used in several steps.

The trace's cyclic property allows terms in such products to be rearranged in many ways, e.g. $$\eqalign{ A:B &= A^T:B^T &= B:A \\ A:BC &= B^TA:C &= AC^T:B \\ }$$ Several steps also made use of the fact that $(M,N)$ and therefore $(R,S)$ are symmetric matrices.

$\endgroup$

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.