# Chain rule in matrix derivation

I am having trouble differentiating $\frac {\partial(tr(AX^{-1}))}{\partial X}$. If I understand, how chain rule works in matrix derivation (if it works at all), then I'm getting this result: $$\frac {\partial(tr(AX^{-1}))}{\partial X} = \frac {\partial(tr(AX^{-1}))}{\partial (X^{-1})} \frac {\partial(X^{-1})}{\partial X}$$ which, using matrix cookbook, particularly equations (101) and (59), gives me $$\frac {\partial(tr(AX^{-1}))}{\partial (X^{-1})} \frac {\partial(X^{-1})}{\partial X} = A^TX^{-2}$$

This is the answer I get, but I'm not sure at all that that is how it works.

• a particular case is easier : if $\|Y\| < 1$ (the operator norm $\|Y\| = \max_{\|v\|=1} \|Y v\|$) then the series $\sum_{k=0}^\infty Y^k$ converges and is equal to $(I-Y)^{-1}$. hence $\frac{\partial (tr(A (I-Y)^{-1}))}{\partial Y} = ?$ – reuns Apr 17 '16 at 15:53
• @user1952009 I'm sorry, but I don't think I follow, what does this give. Also, I don't have any such guarantees. – Alex Apr 17 '16 at 16:02
• since $I$ is constant, $\frac{\partial}{\partial (I-Y)}$ and $-\frac{\partial}{\partial Y}$ is the same. hence if $X = I-Y$ with $\|Y\| < 1$ ... and what I wanted to show you is that even if you are treating matrices, you can write the Taylor expansion $((I-\epsilon Y)M)^{-1} = M^{-1} (I-\epsilon Y)^{-1} = M^{-1} \sum_{k=0}^\infty \epsilon^k Y^k = M^{-1} + \epsilon M^{-1} Y + \mathcal{O}(\epsilon^2)$ – reuns Apr 17 '16 at 16:12

Write the function in terms of the Frobenius (:) Inner Product and take its differential \eqalign{ f &= {\rm tr}(AX^{-1}) = A^T: X^{-1} \cr\cr df &= A^T: dX^{-1} \cr &= -A^T: X^{-1}\,dX\,X^{-1} \cr &= -X^{-T}A^TX^{-T}:dX \cr } Since $df=\big(\frac{\partial f}{\partial X}:dX\big),\,$ the gradient must be \eqalign{ \frac{\partial f}{\partial X} &= -X^{-T}A^TX^{-T} \cr }