Suppose $\mathbf{a}_{N\times1}$ is a complex-valued vector.

$B_{K\times N}$ is a complex-valued matrix.

$C_{K\times K}$ is a real-valued diagonal matrix.

Here is a function $f: \mathbb{C}^N \to \mathbb{R}$, $f = \frac{\mathbf{a}^H B^H C B \mathbf{a}}{\mathbf{a}^H B^H B \mathbf{a}}$, and whose numerator and denominator both are real-valued scalars (this can be verified). Here the operator $H$ means conjugate transpose.

Now I want to take the derivative of function $f$, but I only have the formula for real vectors like $\frac{\partial (\mathbf{x}^T B \mathbf{x})}{\partial \mathbf{x}} = (B+B^T)\mathbf{x}$, which seems not applicable in this case. How can I do this? And further, if I let the derivative of this function $f$ equal to zero, can I get the max or min value of the function $f$?

Thank you!

  • $\begingroup$ So, you want to take the derivative of $f$ with respect to $a$? $\endgroup$ – user550103 Mar 11 '20 at 16:56
  • $\begingroup$ Yes you’re right! Could you help me? $\endgroup$ – J L Mar 11 '20 at 16:58

HINT (well, it's nearly complete)

Some definitions with differentials.

$D := B^H B$ and $N := B^H C B$

$\alpha := a^H N a = \operatorname{Tr}(a^H N a) := a^* : Na,$ where notation $:$ corresponds to Frobenius product and the operator $(\cdot)^*$ is complex conjugate.

Differential of $\alpha$ is $\ d \alpha = Na : da^*$

Similarly, $\beta := a^H D a$ $\Rightarrow d\beta = Da : da^*$

$\gamma:= \beta^{-1}$ $\Rightarrow d\gamma = -\beta^{-2} \ d\beta$

Now, you can describe your function as $$f := \gamma \alpha \ .$$

Take the differential and then we obtain the gradient, i.e., \begin{align} df &= d\gamma \ \alpha + \gamma \ d \alpha\\ &= \left( -\beta^{-2} \ d\beta \right) \ \alpha + \gamma \ \left( Na : da^* \right)\\ &= -\beta^{-2} \alpha \ \left( Da : da^* \right) + \gamma \ \left( Na : da^* \right)\\ &= \left( -\beta^{-2} \alpha \ Da + \gamma Na \right): da^* \ . \end{align}

The gradient is \begin{align} \frac{\partial f}{\partial a^*} &= \left( -\beta^{-2} \alpha \ D + \gamma N \right)a \ . \end{align}

  • 1
    $\begingroup$ Since $f$ is presumed real, the other gradient (if needed) is $$\frac{\partial f}{\partial a} = \left(\frac{\partial f}{\partial a^*}\right)^*$$ $\endgroup$ – greg Mar 11 '20 at 19:18
  • $\begingroup$ @greg: Thank you. I missed that one. :). $\endgroup$ – user550103 Mar 11 '20 at 19:40
  • $\begingroup$ @user550103 Thank you for your kind help! $\endgroup$ – J L Mar 12 '20 at 3:21

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.