Take the 2-minute tour ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Given three matrices $A$ (broad), $B$ and $C$, I'd like to find the derivative of

\begin{align} f = \textrm{tr} \{BA^+\} + \textrm{tr} \{B(A^+)^TCA^+B^T\} \end{align}

with respect to $A$, where $A^+ = A^T(AA^T)^{-1}$ is the Moore–Penrose pseudo inverse. I know how to compute the derivative of $\frac{\partial \ \textrm{tr} \{ BA^+\}}{\partial A^+} = B^T$ , but cannot figure out a) how to compute the derivative of $A^+$ with respect to $A$ and b) how to use the matrix derivative chain rule to combine these two results.

share|improve this question

1 Answer 1

up vote 4 down vote accepted

You should think of $A$ as a time dependent matrix $A=A_t$, denote by $\dot{A}$ the $t$-derivative of $A$, and then think of $f$ as a function of $t$.

To compute the derivative of the inverse of $t$-dependent matrix $B_t$ proceed as follows

$$ 1=B_t B_t^{-1}\Rightarrow 0= \dot{B}_t B_t+B_t\frac{d}{dt}B_t^{-1} $$

so that

$$ \frac{d}{dt}B_t^{-1}=-B_t^{-1}\dot{B}_t B_t. $$

Now let $B_t= A_tA_t^*$. Then $\dot{B}_t=\dot{A}_tA_t^*+A_t\dot{A}_t^*$ and

$$A_t^+=A_t B_t^{-1},$$

$$\dot{A}_t^+= \dot{A}^*_t B_t^{-1}+A_t^*\frac{d}{dt} {B_t}^{-1}=\dot{A}^*_t B_t^{-1}-A_t^* B_t^{-1}\dot{B_t} B_t^{-1}. $$

The Frechet derivative of $A\mapsto A^+$ at the point $A_0$ is the linear map $H\to L(H)$ obtained from the above equality by setting $t=0$ and $\dot{A}_0= H$.

$$ H\mapsto L(H)=H^*B_0^{-1}-A_0^*B_0^{-1}\dot{B}_0 B_0^{-1} =H^*B_0^{-1}-A_0^*B_0^{-1}(HA_0^*+A_0H^*)B_0^{-1}, $$

$$ B_0= A_0A_0^*. $$

This follows from the fact that the Frechet derivative of $A\to A^+$ in the direction $H$ is the derivative at $t=0$ of the path $A_t^+$, $A_t=A_0+tH$.

Update. The $\DeclareMathOperator{\tr}{tr}$ trace defines an inner product onn the space of $n\times n$ matrice via the equality

$$(A,B):=\tr(A^* B). $$

The differential of the function $\tr(BA^+)$ is a linear functional on the space of $n\times n$ matrices given by

$$ H\mapsto L(H) =\tr B \bigl(H^*(A_0A_0^*)^{-1}-A_0^*(HA_0^*+A_0H^*)(A_0A_0^*)^{-1}\bigr). $$

To simplify the presentation let me set $S=(A_0A_0)^{-1}$ so that

$$ L(H)= \tr B(HS-A_0^*HA_0^*S-A_0^*A_0H S) $$

$$= \tr BHS-\tr B A_0^*HA_0^*S-\tr BA_0^*A_0 H S $$

$$ =\tr SB H-\tr A_0^*S B A_0^*H-\tr SBA_0^*A_0 H $$

$$= \tr \underbrace{(SB - A_0^*S B A_0^*- SBA_0^*A_0)}_{=: G} H =(G^*, H). $$

We deduce that the gradient at $A_0$ of the function $A\mapsto \tr(B A^+)$ is the matrix $G^*$.

share|improve this answer
    
Thanks @Liviu. One more question; how can I get the gradient from Frechet derivative? (I don't want the any direction $H$ in my final answer.) –  Mkl Apr 30 at 16:03
1  
I'll update the answer. –  Liviu Nicolaescu Apr 30 at 16:18

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.