Application of KKT-Theorem Does Not Give Expected Result

66 Views Asked by At

I've got this exercise for uni that I've been absolutely racking my brain over for the past couple hours, the problem goes as follows:

Let $\theta_1, \cdots, \theta_d \in (0, \infty)$. Consider the optimization problem $$\max_{w \in \mathbb{R}^d}\sum_{i=1}^d \log{(\theta_i + w_i)} \ \ \text{ s.t. } \ w_i \geq 0, \text{for }i=1, \cdots, d, \ \sum_{i=1}^dw_i = 1$$ a) show that the problem is (equivalent to) a convex optimization problem

b) Show that the solution $w^*$ of this problem is given by $$w^*_i = \max\bigg\{0, \frac{1}{\lambda^* - \theta_i}\bigg\}$$ where $\lambda^*$ is uniquely determined by $$\sum_{i=1}^d \max\bigg\{0, \frac{1}{\lambda^* - \theta_i}\bigg\} = 1$$

The answer to (a) seemed quite simple to me, as $\log$ is concave, the new problem would be minimising the negative objective function from above. But (b) is where I am struggling severely.

It would seem logical to me to use the KKT-theorem here. Writing out the KKT-conditions for the problem resulting from (a) problem gives us, for all $i=1, \cdots, d$:

$$-\frac{1}{\theta_i + w^*_i} - \lambda^* - \mu_i^* = 0$$ $$\mu^*_iw^*_i = 0$$

Multiplying both sides of the first equation with $(\theta_i + w^*_i)$ gives $$-1 - (\theta_i + w^*_i)\lambda^* - (\theta_i + w^*_i)\mu_i^* = 0$$ $$\mu^*_iw^*_i = 0$$ Filling in our definition of $w_i^*$, and applying our second condition gives $$-1 - \max\Big\{\theta_i, \frac{1}{\lambda^*}\Big\}\lambda^* - \theta_i\mu_i^* = 0$$ $$\mu^*_i\max\Big\{0, \frac{1}{\lambda^* - \theta_i}\Big\} = 0$$

Which is in turn equivalent to: $$-1 - \max\Big\{\theta_i\lambda^*, 1\Big\} - \theta_i\mu_i^* = 0$$ $$\mu^*_i\max\Big\{0, \frac{1}{\lambda^* - \theta_i}\Big\} = 0$$

But now, looking at the first condition, we have two negative numbers being subtraced by a nonnegative number, which should combine to equal zero. I just don't see where I am mistaken. Any help with this would be greatly appreciated!! Thanks in advance.

1

There are 1 best solutions below

0
On BEST ANSWER

We have $\mu^*_i \geq 0$, but no restriction on $\lambda^*$. So if $\lambda^*$ is negative, then $\max\Big\{\theta_i, \frac{1}{\lambda^*}\Big\}= \theta_i$, and therefore $\max\Big\{\theta_i, \frac{1}{\lambda^*}\Big\}\lambda^*=\theta_i \lambda^*$