0
$\begingroup$

Can somebody help me find the first of the following function $\mathcal{L}$ with respect to the elements $\phi_{mn}$ of the matrix $\mathbf{\Phi}$?

\begin{equation} \mathcal{L} = -\frac{1}{2}\big[N\log 2\pi + \log\left|\mathbf{C}\right| + \mathbf{\hat{t}}^{T}\mathbf{C}^{-1}\mathbf{\hat{t}} \big], \end{equation}

where \begin{equation} \mathbf{C} = \mathbf{B} + \mathbf{\Phi} \mathbf{A}^{-1} \mathbf{\Phi}^{T}, \\ \mathbf{\hat{t}} = \mathbf{\Phi}\boldsymbol{w} + \mathbf{B}^{-1}\mathbf{e}. \end{equation}

Note that bold small letters denote $N\times 1$ vectors, bold capital letters $N\times N$ matrices, and $\mathbf{A}$, $\mathbf{B}$ are diagonal.

Background information: $\mathcal{L}$ is a likelihood function and relates to the Relevance Vector Machine (RVM) as described in the following paper.

EDIT:

I came so far on my own for the first derivative of the last term (the first two terms were already answered in my previous question). I first expanded it to get

\begin{align} \mathbf{\hat{t}}^{T}\mathbf{C}^{-1}\mathbf{\hat{t}} ~=~&\mathbf{w}^{T}\mathbf{\Phi}^{T}\mathbf{B}\mathbf{\Phi}\mathbf{w} \\ ~+~&\mathbf{e}^{T}\mathbf{\Phi}\mathbf{w} \\ ~+~&\mathbf{w}^{T}\mathbf{\Phi}^{T}\mathbf{e} \\ ~+~&\mathbf{e}^{T}\mathbf{B}^{-T}\mathbf{e} \\ ~-~&\mathbf{w}^{T}\mathbf{\Phi}\mathbf{B}\mathbf{\Phi}\mathbf{\Sigma}\mathbf{\Phi}^{T}\mathbf{\Phi}\mathbf{w} \\ ~-~&\mathbf{e}^{T}\mathbf{\Phi}\mathbf{\Sigma}\mathbf{\Phi}^{T}\mathbf{B}\mathbf{\Phi}\mathbf{w} \\ ~-~&\mathbf{w}^{T}\mathbf{\Phi}\mathbf{B}\mathbf{\Phi}\mathbf{\Sigma}\mathbf{\Phi}^{T}\mathbf{e} \\ ~-~&\mathbf{e}^{T}\mathbf{\Phi}\mathbf{\Sigma}\mathbf{\Phi}^{T}\mathbf{e}. \end{align}

Then I took the derivatives for terms number 1,2,3, and 8 on the right hand side using formulas from the matrix cookbook

\begin{equation} \frac{\partial \left[ \mathbf{w}^{T}\mathbf{\Phi}^{T}\mathbf{B}\mathbf{\Phi}\mathbf{w} \right]}{\partial \mathbf{\Phi}} = 2 \mathbf{B}\mathbf{w}\mathbf{w}^{T} \\ \frac{\partial \left[ \mathbf{e}^{T}\mathbf{\Phi}\mathbf{w} \right]}{\partial \mathbf{\Phi}} = \mathbf{e}\mathbf{w}^{T} \\ \frac{\partial \left[ \mathbf{w}^{T}\mathbf{\Phi}^{T}\mathbf{e} \right]}{\partial \mathbf{\Phi}} = \mathbf{e}\mathbf{w}^{T} \\ \frac{\partial \left[ \mathbf{e}^{T}\mathbf{\Phi}\mathbf{\Sigma}\mathbf{\Phi}^{T}\mathbf{e} \right]}{\partial \mathbf{\Phi}} = 2 \mathbf{\Sigma}^{T}\mathbf{e}\mathbf{e}^{T} \end{equation}

The 4th term drops, but I don't know how to find the derivatives for the remaining higher order terms 5, 6, and 7. Any suggestions? Is what I did correct so far?

$\endgroup$
2
$\begingroup$

For ease of typing, I'm going to use $$\eqalign{ v &= \mathbf{\hat{t}}\cr P &= \mathbf{\Phi} \cr }$$ The term you are asking about can be written using the Frobenius product $$\eqalign{ f &= v^TC^{-1}v \cr &= C^{-1}:vv^T \cr }$$ Its differential is $$\eqalign{ df &= C^{-1}:d(vv^T) + dC^{-1}:vv^T \cr &= C^{-1}:2\,{\rm sym}(dv\,v^T) - C^{-1}\,dC\,C^{-1}:vv^T \cr &= 2\,C^{-1}v:dv - C^{-1}vv^TC^{-1}:dC \cr\cr }$$ Now we need expressions for $(dC, dv)$ in terms of $dP$. $$\eqalign{ v &= Pw +B^{-1}e \cr dv &= dP\,w \cr\cr C &= B + PA^{-1}P^T \cr dC &= 2\,{\rm sym}(dP\,A^{-1}P^T) \cr }$$ Substituting $$\eqalign{ df &= 2\,C^{-1}v:dP\,w - C^{-1}vv^TC^{-1}:2\,{\rm sym}(dP\,A^{-1}P^T) \cr &= 2\,C^{-1}vw^T:dP - 2\,C^{-1}vv^TC^{-1}PA^{-1}:dP \cr &= 2\,C^{-1}v(w^T - v^TC^{-1}PA^{-1}):dP \cr\cr \frac{\partial f}{\partial P} &= 2\,C^{-1}v(w^T - v^TC^{-1}PA^{-1}) \cr\cr }$$ In these manipulations, I've made use of the fact that $A$, $C$ (and their inverses) are symmetric.

I've also made extensive use of the algebraic properties of the Frobenius product, such as $$\eqalign{ A:BC &= AC^T:B \cr X:YZ &= Y^TX:Z \cr A:B &= B:A \cr A:B &= A^T:B^T \cr A:{\rm sym}(B) &= {\rm sym}(A):B \cr A:{\rm skew}(B) &= {\rm skew}(A):B \cr {\rm sym}(A):{\rm skew}(B) &= 0 \cr }$$ all of which can easily be verified from the definitions $$\eqalign{ A:B &= {\rm tr}(A^TB) \cr {\rm sym}(A) &= \frac{1}{2}(A+A^T) \cr {\rm skew}(A) &= \frac{1}{2}(A-A^T) \cr }$$

$\endgroup$
  • $\begingroup$ Thank you very much, and especially for listing the properties above. This was very helpful in understanding what you did! Now, for the full derivative, is the following correct? \begin{equation} \frac{\partial L}{\partial {\bf \Phi}} = \ \mathbf{C}^{-1}\left( \left[ \mathbf{\hat{t}}\mathbf{\hat{t}}^{T}\mathbf{C}^{-1} - \mathbf{I} \right]\boldsymbol{\Phi}\mathbf{A}^{-1} - \mathbf{\hat{t}}\mathbf{w}^{T} \right) \end{equation} $\endgroup$ – mwater Sep 14 '15 at 17:01
  • $\begingroup$ Btw I will totally upvote your nice and detailed answers, but can only do so once I gained 15 reputation. $\endgroup$ – mwater Sep 15 '15 at 12:27

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.